/src/libavif/ext/libyuv/source/row_gcc.cc
Line | Count | Source |
1 | | /* |
2 | | * Copyright 2011 The LibYuv Project Authors. All rights reserved. |
3 | | * |
4 | | * Use of this source code is governed by a BSD-style license |
5 | | * that can be found in the LICENSE file in the root of the source |
6 | | * tree. An additional intellectual property rights grant can be found |
7 | | * in the file PATENTS. All contributing project authors may |
8 | | * be found in the AUTHORS file in the root of the source tree. |
9 | | */ |
10 | | |
11 | | #include "libyuv/convert_from_argb.h" // For ArgbConstants |
12 | | #include "libyuv/row.h" |
13 | | |
14 | | #ifdef __cplusplus |
15 | | namespace libyuv { |
16 | | extern "C" { |
17 | | #endif |
18 | | |
19 | | // This module is for GCC x86 and x64. |
20 | | #if !defined(LIBYUV_DISABLE_X86) && \ |
21 | | (defined(__x86_64__) || defined(__i386__)) && \ |
22 | | !defined(LIBYUV_ENABLE_ROWWIN) |
23 | | |
24 | | // Note: for avx and avx512 declare clobber as xmm registers due to |
25 | | // clang for windows needing to preserve xmm registers but not saving |
26 | | // them if declared as ymm or zmm. |
27 | | |
28 | | #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3) |
29 | | |
30 | | // Constants for ARGB |
31 | | |
32 | | // JPeg full range. |
33 | | static const uvec8 kARGBToYJ = {29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u, |
34 | | 29u, 150u, 77u, 0u, 29u, 150u, 77u, 0u}; |
35 | | |
36 | | #endif // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_ARGBGRAYROW_SSSE3) |
37 | | |
38 | | #if defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3) |
39 | | // Constants for BGRA |
40 | | |
41 | | // Constants for ABGR |
42 | | |
43 | | // Constants for RGBA. |
44 | | // 126 (7e) - (-109..110) = 16..235 |
45 | | |
46 | | static const uvec16 kSub128 = {0x8080u, 0x8080u, 0x8080u, 0x8080u, |
47 | | 0x8080u, 0x8080u, 0x8080u, 0x8080u}; |
48 | | |
49 | | #endif // defined(HAS_ARGBTOYROW_SSSE3) || defined(HAS_I422TOARGBROW_SSSE3) |
50 | | |
51 | | #ifdef HAS_RGB24TOARGBROW_SSSE3 |
52 | | |
53 | | // Shuffle table for converting RGB24 to ARGB. |
54 | | static const uvec8 kShuffleMaskRGB24ToARGB[2] = { |
55 | | {0u, 1u, 2u, 128u, 3u, 4u, 5u, 128u, 6u, 7u, 8u, 128u, 9u, 10u, 11u, 128u}, |
56 | | {4u, 5u, 6u, 128u, 7u, 8u, 9u, 128u, 10u, 11u, 12u, 128u, 13u, 14u, 15u, |
57 | | 128u}}; |
58 | | |
59 | | // Shuffle table for converting RAW to ARGB. |
60 | | static const uvec8 kShuffleMaskRAWToARGB = { |
61 | | 2u, 1u, 0u, 128u, 5u, 4u, 3u, 128u, 8u, 7u, 6u, 128u, 11u, 10u, 9u, 128u}; |
62 | | // Shuffle table for converting RAW to ARGB. Last 12 |
63 | | static const uvec8 kShuffleMaskRAWToARGB_0 = {6u, 5u, 4u, 128u, 9u, 8u, |
64 | | 7u, 128u, 12u, 11u, 10u, 128u, |
65 | | 15u, 14u, 13u, 128u}; |
66 | | |
67 | | // Shuffle table for converting RAW to RGBA. |
68 | | static const uvec8 kShuffleMaskRAWToRGBA = { |
69 | | 128u, 2u, 1u, 0u, 128u, 5u, 4u, 3u, 128u, 8u, 7u, 6u, 128u, 11u, 10u, 9u}; |
70 | | |
71 | | // Shuffle table for converting RAW to RGB24. First 8. |
72 | | static const uvec8 kShuffleMaskRAWToRGB24_0 = { |
73 | | 2u, 1u, 0u, 5u, 4u, 3u, 8u, 7u, |
74 | | 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u}; |
75 | | |
76 | | // Shuffle table for converting RAW to RGB24. Middle 8. |
77 | | static const uvec8 kShuffleMaskRAWToRGB24_1 = { |
78 | | 2u, 7u, 6u, 5u, 10u, 9u, 8u, 13u, |
79 | | 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u}; |
80 | | |
81 | | // Shuffle table for converting RAW to RGB24. Last 8. |
82 | | static const uvec8 kShuffleMaskRAWToRGB24_2 = { |
83 | | 8u, 7u, 12u, 11u, 10u, 15u, 14u, 13u, |
84 | | 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u}; |
85 | | |
86 | | // Shuffle table for converting ARGB to RGB24. |
87 | | static const uvec8 kShuffleMaskARGBToRGB24 = { |
88 | | 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, 14u, 128u, 128u, 128u, 128u}; |
89 | | |
90 | | // Shuffle table for converting ARGB to RAW. |
91 | | static const uvec8 kShuffleMaskARGBToRAW = { |
92 | | 2u, 1u, 0u, 6u, 5u, 4u, 10u, 9u, 8u, 14u, 13u, 12u, 128u, 128u, 128u, 128u}; |
93 | | |
94 | | // Shuffle table for converting ARGBToRGB24 for I422ToRGB24. First 8 + next 4 |
95 | | static const uvec8 kShuffleMaskARGBToRGB24_0 = { |
96 | | 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 128u, 128u, 128u, 128u, 10u, 12u, 13u, 14u}; |
97 | | |
98 | | // YUY2 shuf 16 Y to 32 Y. |
99 | | static const vec8 kShuffleYUY2Y = {0, 0, 2, 2, 4, 4, 6, 6, |
100 | | 8, 8, 10, 10, 12, 12, 14, 14}; |
101 | | |
102 | | // YUY2 shuf 8 UV to 16 UV. |
103 | | static const vec8 kShuffleYUY2UV = {1, 3, 1, 3, 5, 7, 5, 7, |
104 | | 9, 11, 9, 11, 13, 15, 13, 15}; |
105 | | |
106 | | // UYVY shuf 16 Y to 32 Y. |
107 | | static const vec8 kShuffleUYVYY = {1, 1, 3, 3, 5, 5, 7, 7, |
108 | | 9, 9, 11, 11, 13, 13, 15, 15}; |
109 | | |
110 | | // UYVY shuf 8 UV to 16 UV. |
111 | | static const vec8 kShuffleUYVYUV = {0, 2, 0, 2, 4, 6, 4, 6, |
112 | | 8, 10, 8, 10, 12, 14, 12, 14}; |
113 | | |
114 | | // NV21 shuf 8 VU to 16 UV. |
115 | | static const lvec8 kShuffleNV21 = { |
116 | | 1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6, |
117 | | 1, 0, 1, 0, 3, 2, 3, 2, 5, 4, 5, 4, 7, 6, 7, 6, |
118 | | }; |
119 | | #endif // HAS_RGB24TOARGBROW_SSSE3 |
120 | | |
121 | | #if defined(HAS_J400TOARGBROW_AVX2) || defined(HAS_J400TOARGBROW_AVX512BW) |
122 | | alignas(64) static const uint8_t kShuffleMaskJ400ToARGB[64] = { |
123 | | 0u, 0u, 0u, 128u, 1u, 1u, 1u, 128u, 2u, 2u, 2u, 128u, 3u, 3u, |
124 | | 3u, 128u, 4u, 4u, 4u, 128u, 5u, 5u, 5u, 128u, 6u, 6u, 6u, 128u, |
125 | | 7u, 7u, 7u, 128u, 8u, 8u, 8u, 128u, 9u, 9u, 9u, 128u, 10u, 10u, |
126 | | 10u, 128u, 11u, 11u, 11u, 128u, 12u, 12u, 12u, 128u, 13u, 13u, 13u, 128u, |
127 | | 14u, 14u, 14u, 128u, 15u, 15u, 15u, 128u}; |
128 | | #endif |
129 | | |
130 | | #ifdef HAS_J400TOARGBROW_AVX2 |
131 | 0 | void J400ToARGBRow_AVX2(const uint8_t* src_y, uint8_t* dst_argb, int width) { |
132 | 0 | asm volatile( |
133 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" |
134 | 0 | "vpslld $0x18,%%ymm7,%%ymm7 \n" |
135 | 0 | "vmovdqa (%3),%%ymm5 \n" |
136 | 0 | "vmovdqa 0x20(%3),%%ymm6 \n" |
137 | |
|
138 | 0 | LABELALIGN |
139 | 0 | "1: \n" |
140 | 0 | "vbroadcasti128 (%0),%%ymm0 \n" |
141 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm1 \n" |
142 | 0 | "vpshufb %%ymm6,%%ymm0,%%ymm2 \n" |
143 | 0 | "vpor %%ymm7,%%ymm1,%%ymm1 \n" |
144 | 0 | "vpor %%ymm7,%%ymm2,%%ymm2 \n" |
145 | 0 | "vmovdqu %%ymm1,(%1) \n" |
146 | 0 | "vmovdqu %%ymm2,0x20(%1) \n" |
147 | 0 | "lea 0x10(%0),%0 \n" |
148 | 0 | "lea 0x40(%1),%1 \n" |
149 | 0 | "sub $0x10,%2 \n" |
150 | 0 | "jg 1b \n" |
151 | 0 | "vzeroupper \n" |
152 | 0 | : "+r"(src_y), // %0 |
153 | 0 | "+r"(dst_argb), // %1 |
154 | 0 | "+r"(width) // %2 |
155 | 0 | : "r"(kShuffleMaskJ400ToARGB) // %3 |
156 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7"); |
157 | 0 | } |
158 | | #endif // HAS_J400TOARGBROW_AVX2 |
159 | | |
160 | | #ifdef HAS_J400TOARGBROW_AVX512BW |
161 | | void J400ToARGBRow_AVX512BW(const uint8_t* src_y, |
162 | | uint8_t* dst_argb, |
163 | 0 | int width) { |
164 | 0 | asm volatile( |
165 | 0 | "vpternlogd $0xff,%%zmm7,%%zmm7,%%zmm7 \n" // 0xffffffff |
166 | 0 | "vpslld $0x18,%%zmm7,%%zmm7 \n" // 0xff000000 |
167 | 0 | "vmovdqa64 %3,%%zmm5 \n" |
168 | |
|
169 | 0 | LABELALIGN |
170 | 0 | "1: \n" |
171 | 0 | "vbroadcasti32x4 (%0),%%zmm0 \n" |
172 | 0 | "vbroadcasti32x4 0x10(%0),%%zmm1 \n" |
173 | 0 | "vpshufb %%zmm5,%%zmm0,%%zmm0 \n" |
174 | 0 | "vpshufb %%zmm5,%%zmm1,%%zmm1 \n" |
175 | 0 | "vpord %%zmm7,%%zmm0,%%zmm0 \n" |
176 | 0 | "vpord %%zmm7,%%zmm1,%%zmm1 \n" |
177 | 0 | "vmovdqu64 %%zmm0,(%1) \n" |
178 | 0 | "vmovdqu64 %%zmm1,0x40(%1) \n" |
179 | 0 | "lea 0x20(%0),%0 \n" |
180 | 0 | "lea 0x80(%1),%1 \n" |
181 | 0 | "sub $0x20,%2 \n" |
182 | 0 | "jg 1b \n" |
183 | 0 | "vzeroupper \n" |
184 | 0 | : "+r"(src_y), // %0 |
185 | 0 | "+r"(dst_argb), // %1 |
186 | 0 | "+r"(width) // %2 |
187 | 0 | : "m"(kShuffleMaskJ400ToARGB) // %3 |
188 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5", "xmm7"); |
189 | 0 | } |
190 | | #endif // HAS_J400TOARGBROW_AVX512BW |
191 | | |
192 | | #ifdef HAS_RGB24TOARGBROW_SSSE3 |
193 | | void RGB24ToARGBRow_SSSE3(const uint8_t* src_rgb24, |
194 | | uint8_t* dst_argb, |
195 | 0 | int width) { |
196 | 0 | asm volatile( |
197 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" // 0xff000000 |
198 | 0 | "pslld $0x18,%%xmm5 \n" |
199 | 0 | "movdqa %3,%%xmm4 \n" |
200 | |
|
201 | 0 | LABELALIGN |
202 | 0 | "1: \n" |
203 | 0 | "movdqu (%0),%%xmm0 \n" |
204 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
205 | 0 | "movdqu 0x20(%0),%%xmm3 \n" |
206 | 0 | "lea 0x30(%0),%0 \n" |
207 | 0 | "movdqa %%xmm3,%%xmm2 \n" |
208 | 0 | "palignr $0x8,%%xmm1,%%xmm2 \n" |
209 | 0 | "pshufb %%xmm4,%%xmm2 \n" |
210 | 0 | "por %%xmm5,%%xmm2 \n" |
211 | 0 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
212 | 0 | "pshufb %%xmm4,%%xmm0 \n" |
213 | 0 | "movdqu %%xmm2,0x20(%1) \n" |
214 | 0 | "por %%xmm5,%%xmm0 \n" |
215 | 0 | "pshufb %%xmm4,%%xmm1 \n" |
216 | 0 | "movdqu %%xmm0,(%1) \n" |
217 | 0 | "por %%xmm5,%%xmm1 \n" |
218 | 0 | "palignr $0x4,%%xmm3,%%xmm3 \n" |
219 | 0 | "pshufb %%xmm4,%%xmm3 \n" |
220 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
221 | 0 | "por %%xmm5,%%xmm3 \n" |
222 | 0 | "movdqu %%xmm3,0x30(%1) \n" |
223 | 0 | "lea 0x40(%1),%1 \n" |
224 | 0 | "sub $0x10,%2 \n" |
225 | 0 | "jg 1b \n" |
226 | 0 | : "+r"(src_rgb24), // %0 |
227 | 0 | "+r"(dst_argb), // %1 |
228 | 0 | "+r"(width) // %2 |
229 | 0 | : "m"(kShuffleMaskRGB24ToARGB[0]) // %3 |
230 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
231 | 0 | } |
232 | | |
233 | | #ifdef HAS_RGB24TOARGBROW_AVX2 |
234 | | void RGB24ToARGBRow_AVX2(const uint8_t* src_rgb24, |
235 | | uint8_t* dst_argb, |
236 | 0 | int width) { |
237 | | // Reference to prevent discarding of kShuffleMaskRGB24ToARGB[1] which is |
238 | | // accessed via offset in assembly. |
239 | 0 | const uvec8* dummy = &kShuffleMaskRGB24ToARGB[1]; |
240 | 0 | (void)dummy; |
241 | 0 | asm volatile( |
242 | 0 | "vpcmpeqb %%ymm6,%%ymm6,%%ymm6 \n" // 0xff000000 |
243 | 0 | "vpslld $0x18,%%ymm6,%%ymm6 \n" |
244 | 0 | "vbroadcasti128 %3,%%ymm4 \n" |
245 | 0 | "vbroadcasti128 16+%3,%%ymm5 \n" |
246 | |
|
247 | 0 | LABELALIGN |
248 | 0 | "1: \n" |
249 | 0 | "vmovdqu (%0),%%xmm0 \n" // first 12 |
250 | 0 | "vinserti128 $1,12(%0),%%ymm0,%%ymm0 \n" // second 12 |
251 | 0 | "vmovdqu 24(%0),%%xmm1 \n" // third 12 |
252 | 0 | "vinserti128 $1,36(%0),%%ymm1,%%ymm1 \n" // forth 12 |
253 | 0 | "vmovdqu 48(%0),%%xmm2 \n" // fifth 12 |
254 | 0 | "vinserti128 $1,60(%0),%%ymm2,%%ymm2 \n" // sixth 12 |
255 | 0 | "vmovdqu 68(%0),%%xmm3 \n" // seventh 12 |
256 | 0 | "vinserti128 $1,80(%0),%%ymm3,%%ymm3 \n" // eighth 12 |
257 | 0 | "lea 96(%0),%0 \n" |
258 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
259 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
260 | 0 | "vpshufb %%ymm4,%%ymm2,%%ymm2 \n" |
261 | 0 | "vpshufb %%ymm5,%%ymm3,%%ymm3 \n" |
262 | 0 | "vpor %%ymm6,%%ymm0,%%ymm0 \n" |
263 | 0 | "vpor %%ymm6,%%ymm1,%%ymm1 \n" |
264 | 0 | "vpor %%ymm6,%%ymm2,%%ymm2 \n" |
265 | 0 | "vpor %%ymm6,%%ymm3,%%ymm3 \n" |
266 | 0 | "vmovdqu %%ymm0,(%1) \n" |
267 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
268 | 0 | "vmovdqu %%ymm2,0x40(%1) \n" |
269 | 0 | "vmovdqu %%ymm3,0x60(%1) \n" |
270 | 0 | "lea 0x80(%1),%1 \n" |
271 | 0 | "sub $0x20,%2 \n" |
272 | 0 | "jg 1b \n" |
273 | 0 | "vzeroupper \n" |
274 | 0 | : "+r"(src_rgb24), // %0 |
275 | 0 | "+r"(dst_argb), // %1 |
276 | 0 | "+r"(width) // %2 |
277 | 0 | : "m"(kShuffleMaskRGB24ToARGB[0]) // %3 |
278 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
279 | 0 | } |
280 | | #endif // HAS_RGB24TOARGBROW_AVX2 |
281 | | |
282 | 0 | void RAWToARGBRow_SSSE3(const uint8_t* src_raw, uint8_t* dst_argb, int width) { |
283 | 0 | asm volatile( |
284 | 0 | "pcmpeqb %%xmm6,%%xmm6 \n" // 0xff000000 |
285 | 0 | "pslld $0x18,%%xmm6 \n" |
286 | 0 | "movdqa %3,%%xmm4 \n" |
287 | 0 | "movdqa %4,%%xmm5 \n" |
288 | |
|
289 | 0 | LABELALIGN |
290 | 0 | "1: \n" |
291 | 0 | "movdqu (%0),%%xmm0 \n" |
292 | 0 | "movdqu 12(%0),%%xmm1 \n" |
293 | 0 | "movdqu 24(%0),%%xmm2 \n" |
294 | 0 | "movdqu 32(%0),%%xmm3 \n" |
295 | 0 | "lea 0x30(%0),%0 \n" |
296 | 0 | "pshufb %%xmm4,%%xmm0 \n" |
297 | 0 | "pshufb %%xmm4,%%xmm1 \n" |
298 | 0 | "pshufb %%xmm4,%%xmm2 \n" |
299 | 0 | "pshufb %%xmm5,%%xmm3 \n" |
300 | 0 | "por %%xmm6,%%xmm0 \n" |
301 | 0 | "por %%xmm6,%%xmm1 \n" |
302 | 0 | "por %%xmm6,%%xmm2 \n" |
303 | 0 | "por %%xmm6,%%xmm3 \n" |
304 | 0 | "movdqu %%xmm0,0x00(%1) \n" |
305 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
306 | 0 | "movdqu %%xmm2,0x20(%1) \n" |
307 | 0 | "movdqu %%xmm3,0x30(%1) \n" |
308 | 0 | "lea 0x40(%1),%1 \n" |
309 | 0 | "sub $0x10,%2 \n" |
310 | 0 | "jg 1b \n" |
311 | 0 | : "+r"(src_raw), // %0 |
312 | 0 | "+r"(dst_argb), // %1 |
313 | 0 | "+r"(width) // %2 |
314 | 0 | : "m"(kShuffleMaskRAWToARGB), // %3 |
315 | 0 | "m"(kShuffleMaskRAWToARGB_0) // %4 |
316 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
317 | 0 | } |
318 | | |
319 | 0 | void RAWToARGBRow_AVX2(const uint8_t* src_raw, uint8_t* dst_argb, int width) { |
320 | 0 | asm volatile( |
321 | 0 | "vpcmpeqb %%ymm6,%%ymm6,%%ymm6 \n" // 0xff000000 |
322 | 0 | "vpslld $0x18,%%ymm6,%%ymm6 \n" |
323 | 0 | "vbroadcasti128 %3,%%ymm4 \n" // |
324 | 0 | "vbroadcasti128 %4,%%ymm5 \n" // |
325 | | |
326 | | LABELALIGN // |
327 | 0 | "1: \n" |
328 | 0 | "vmovdqu (%0),%%xmm0 \n" // first 12 |
329 | 0 | "vinserti128 $1,12(%0),%%ymm0,%%ymm0 \n" // second 12 |
330 | 0 | "vmovdqu 24(%0),%%xmm1 \n" // third 12 |
331 | 0 | "vinserti128 $1,36(%0),%%ymm1,%%ymm1 \n" // forth 12 |
332 | 0 | "vmovdqu 48(%0),%%xmm2 \n" // fifth 12 |
333 | 0 | "vinserti128 $1,60(%0),%%ymm2,%%ymm2 \n" // sixth 12 |
334 | 0 | "vmovdqu 68(%0),%%xmm3 \n" // seventh 12 |
335 | 0 | "vinserti128 $1,80(%0),%%ymm3,%%ymm3 \n" // eighth 12 |
336 | 0 | "lea 96(%0),%0 \n" |
337 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
338 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
339 | 0 | "vpshufb %%ymm4,%%ymm2,%%ymm2 \n" |
340 | 0 | "vpshufb %%ymm5,%%ymm3,%%ymm3 \n" |
341 | 0 | "vpor %%ymm6,%%ymm0,%%ymm0 \n" |
342 | 0 | "vpor %%ymm6,%%ymm1,%%ymm1 \n" |
343 | 0 | "vpor %%ymm6,%%ymm2,%%ymm2 \n" |
344 | 0 | "vpor %%ymm6,%%ymm3,%%ymm3 \n" |
345 | 0 | "vmovdqu %%ymm0,(%1) \n" |
346 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
347 | 0 | "vmovdqu %%ymm2,0x40(%1) \n" |
348 | 0 | "vmovdqu %%ymm3,0x60(%1) \n" |
349 | 0 | "lea 0x80(%1),%1 \n" |
350 | 0 | "sub $0x20,%2 \n" |
351 | 0 | "jg 1b \n" |
352 | 0 | "vzeroupper \n" |
353 | 0 | : "+r"(src_raw), // %0 |
354 | 0 | "+r"(dst_argb), // %1 |
355 | 0 | "+r"(width) // %2 |
356 | 0 | : "m"(kShuffleMaskRAWToARGB), // %3 |
357 | 0 | "m"(kShuffleMaskRAWToARGB_0) // %4 |
358 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
359 | 0 | } |
360 | | |
361 | | #ifdef HAS_RAWTOARGBROW_AVX512BW |
362 | | static const uint32_t kPermdRAWToARGB_AVX512BW[16] = { |
363 | | 0, 1, 2, 3, 3, 4, 5, 6, 6, 7, 8, 9, 9, 10, 11, 12}; |
364 | | |
365 | | void RGBToARGBRow_AVX512BW(const uint8_t* src_raw, |
366 | | uint8_t* dst_argb, |
367 | | const uint32_t* shuffler, |
368 | 0 | int width) { |
369 | 0 | asm volatile( |
370 | 0 | "vpternlogd $0xff,%%zmm6,%%zmm6,%%zmm6 \n" // 0xffffffff |
371 | 0 | "vpslld $0x18,%%zmm6,%%zmm6 \n" // 0xff000000 |
372 | 0 | "movabs $0xffffffffffff,%%rax \n" // 48 bytes mask |
373 | 0 | "kmovq %%rax,%%k1 \n" |
374 | 0 | "vmovdqu32 %3,%%zmm5 \n" |
375 | 0 | "vbroadcasti32x4 %4,%%zmm4 \n" |
376 | | |
377 | | LABELALIGN // |
378 | 0 | "1: \n" |
379 | 0 | "vmovdqu8 (%0),%%zmm0%{%%k1%}%{z%} \n" |
380 | 0 | "vmovdqu8 48(%0),%%zmm1%{%%k1%}%{z%} \n" |
381 | 0 | "vmovdqu8 96(%0),%%zmm2%{%%k1%}%{z%} \n" |
382 | 0 | "vmovdqu8 144(%0),%%zmm3%{%%k1%}%{z%} \n" |
383 | 0 | "lea 192(%0),%0 \n" |
384 | 0 | "vpermd %%zmm0,%%zmm5,%%zmm0 \n" |
385 | 0 | "vpermd %%zmm1,%%zmm5,%%zmm1 \n" |
386 | 0 | "vpermd %%zmm2,%%zmm5,%%zmm2 \n" |
387 | 0 | "vpermd %%zmm3,%%zmm5,%%zmm3 \n" |
388 | 0 | "vpshufb %%zmm4,%%zmm0,%%zmm0 \n" |
389 | 0 | "vpshufb %%zmm4,%%zmm1,%%zmm1 \n" |
390 | 0 | "vpshufb %%zmm4,%%zmm2,%%zmm2 \n" |
391 | 0 | "vpshufb %%zmm4,%%zmm3,%%zmm3 \n" |
392 | 0 | "vpord %%zmm6,%%zmm0,%%zmm0 \n" |
393 | 0 | "vpord %%zmm6,%%zmm1,%%zmm1 \n" |
394 | 0 | "vpord %%zmm6,%%zmm2,%%zmm2 \n" |
395 | 0 | "vpord %%zmm6,%%zmm3,%%zmm3 \n" |
396 | 0 | "vmovdqu32 %%zmm0,(%1) \n" |
397 | 0 | "vmovdqu32 %%zmm1,0x40(%1) \n" |
398 | 0 | "vmovdqu32 %%zmm2,0x80(%1) \n" |
399 | 0 | "vmovdqu32 %%zmm3,0xc0(%1) \n" |
400 | 0 | "lea 0x100(%1),%1 \n" |
401 | 0 | "sub $0x40,%2 \n" |
402 | 0 | "jg 1b \n" |
403 | 0 | "vzeroupper \n" |
404 | 0 | : "+r"(src_raw), // %0 |
405 | 0 | "+r"(dst_argb), // %1 |
406 | 0 | "+r"(width) // %2 |
407 | 0 | : "m"(kPermdRAWToARGB_AVX512BW), // %3 |
408 | 0 | "m"(*shuffler) // %4 |
409 | 0 | : "memory", "cc", "rax", "k1", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", |
410 | 0 | "xmm5", "xmm6"); |
411 | 0 | } |
412 | | |
413 | | void RAWToARGBRow_AVX512BW(const uint8_t* src_raw, |
414 | | uint8_t* dst_argb, |
415 | 0 | int width) { |
416 | 0 | RGBToARGBRow_AVX512BW(src_raw, dst_argb, |
417 | 0 | (const uint32_t*)&kShuffleMaskRAWToARGB, width); |
418 | 0 | } |
419 | | |
420 | | void RGB24ToARGBRow_AVX512BW(const uint8_t* src_rgb24, |
421 | | uint8_t* dst_argb, |
422 | 0 | int width) { |
423 | 0 | RGBToARGBRow_AVX512BW(src_rgb24, dst_argb, |
424 | 0 | (const uint32_t*)&kShuffleMaskRGB24ToARGB[0], width); |
425 | 0 | } |
426 | | #endif |
427 | | |
428 | | // Same code as RAWToARGB with different shuffler and A in low bits |
429 | 0 | void RAWToRGBARow_SSSE3(const uint8_t* src_raw, uint8_t* dst_rgba, int width) { |
430 | 0 | asm volatile( |
431 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" // 0x000000ff |
432 | 0 | "psrld $24,%%xmm5 \n" |
433 | 0 | "movdqa %3,%%xmm4 \n" |
434 | |
|
435 | 0 | LABELALIGN |
436 | 0 | "1: \n" |
437 | 0 | "movdqu (%0),%%xmm0 \n" |
438 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
439 | 0 | "movdqu 0x20(%0),%%xmm3 \n" |
440 | 0 | "lea 0x30(%0),%0 \n" |
441 | 0 | "movdqa %%xmm3,%%xmm2 \n" |
442 | 0 | "palignr $0x8,%%xmm1,%%xmm2 \n" |
443 | 0 | "pshufb %%xmm4,%%xmm2 \n" |
444 | 0 | "por %%xmm5,%%xmm2 \n" |
445 | 0 | "palignr $0xc,%%xmm0,%%xmm1 \n" |
446 | 0 | "pshufb %%xmm4,%%xmm0 \n" |
447 | 0 | "movdqu %%xmm2,0x20(%1) \n" |
448 | 0 | "por %%xmm5,%%xmm0 \n" |
449 | 0 | "pshufb %%xmm4,%%xmm1 \n" |
450 | 0 | "movdqu %%xmm0,(%1) \n" |
451 | 0 | "por %%xmm5,%%xmm1 \n" |
452 | 0 | "palignr $0x4,%%xmm3,%%xmm3 \n" |
453 | 0 | "pshufb %%xmm4,%%xmm3 \n" |
454 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
455 | 0 | "por %%xmm5,%%xmm3 \n" |
456 | 0 | "movdqu %%xmm3,0x30(%1) \n" |
457 | 0 | "lea 0x40(%1),%1 \n" |
458 | 0 | "sub $0x10,%2 \n" |
459 | 0 | "jg 1b \n" |
460 | 0 | : "+r"(src_raw), // %0 |
461 | 0 | "+r"(dst_rgba), // %1 |
462 | 0 | "+r"(width) // %2 |
463 | 0 | : "m"(kShuffleMaskRAWToRGBA) // %3 |
464 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
465 | 0 | } |
466 | | |
467 | | void RAWToRGB24Row_SSSE3(const uint8_t* src_raw, |
468 | | uint8_t* dst_rgb24, |
469 | 0 | int width) { |
470 | 0 | asm volatile( |
471 | 0 | "movdqa %3,%%xmm3 \n" |
472 | 0 | "movdqa %4,%%xmm4 \n" |
473 | 0 | "movdqa %5,%%xmm5 \n" |
474 | |
|
475 | 0 | LABELALIGN |
476 | 0 | "1: \n" |
477 | 0 | "movdqu (%0),%%xmm0 \n" |
478 | 0 | "movdqu 0x4(%0),%%xmm1 \n" |
479 | 0 | "movdqu 0x8(%0),%%xmm2 \n" |
480 | 0 | "lea 0x18(%0),%0 \n" |
481 | 0 | "pshufb %%xmm3,%%xmm0 \n" |
482 | 0 | "pshufb %%xmm4,%%xmm1 \n" |
483 | 0 | "pshufb %%xmm5,%%xmm2 \n" |
484 | 0 | "movq %%xmm0,(%1) \n" |
485 | 0 | "movq %%xmm1,0x8(%1) \n" |
486 | 0 | "movq %%xmm2,0x10(%1) \n" |
487 | 0 | "lea 0x18(%1),%1 \n" |
488 | 0 | "sub $0x8,%2 \n" |
489 | 0 | "jg 1b \n" |
490 | 0 | : "+r"(src_raw), // %0 |
491 | 0 | "+r"(dst_rgb24), // %1 |
492 | 0 | "+r"(width) // %2 |
493 | 0 | : "m"(kShuffleMaskRAWToRGB24_0), // %3 |
494 | 0 | "m"(kShuffleMaskRAWToRGB24_1), // %4 |
495 | 0 | "m"(kShuffleMaskRAWToRGB24_2) // %5 |
496 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
497 | 0 | } |
498 | | |
499 | | #ifdef HAS_RGB565TOARGBROW_AVX2 |
500 | 0 | void RGB565ToARGBRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
501 | 0 | asm volatile( |
502 | 0 | "mov $0x1080108,%%eax \n" |
503 | 0 | "vmovd %%eax,%%xmm5 \n" |
504 | 0 | "vpbroadcastd %%xmm5,%%ymm5 \n" |
505 | 0 | "mov $0x20802080,%%eax \n" |
506 | 0 | "vmovd %%eax,%%xmm6 \n" |
507 | 0 | "vpbroadcastd %%xmm6,%%ymm6 \n" |
508 | 0 | "vpcmpeqb %%ymm3,%%ymm3,%%ymm3 \n" |
509 | 0 | "vpsllw $0xb,%%ymm3,%%ymm3 \n" |
510 | 0 | "vpcmpeqb %%ymm4,%%ymm4,%%ymm4 \n" |
511 | 0 | "vpsllw $10,%%ymm4,%%ymm4 \n" |
512 | 0 | "vpsrlw $5,%%ymm4,%%ymm4 \n" |
513 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" |
514 | 0 | "vpsllw $0x8,%%ymm7,%%ymm7 \n" |
515 | 0 | "sub %0,%1 \n" |
516 | 0 | "sub %0,%1 \n" |
517 | |
|
518 | 0 | LABELALIGN |
519 | 0 | "1: \n" |
520 | 0 | "vmovdqu (%0),%%ymm0 \n" |
521 | 0 | "vpand %%ymm3,%%ymm0,%%ymm1 \n" |
522 | 0 | "vpsllw $0xb,%%ymm0,%%ymm2 \n" |
523 | 0 | "vpmulhuw %%ymm5,%%ymm1,%%ymm1 \n" |
524 | 0 | "vpmulhuw %%ymm5,%%ymm2,%%ymm2 \n" |
525 | 0 | "vpsllw $0x8,%%ymm1,%%ymm1 \n" |
526 | 0 | "vpor %%ymm2,%%ymm1,%%ymm1 \n" |
527 | 0 | "vpand %%ymm4,%%ymm0,%%ymm0 \n" |
528 | 0 | "vpmulhuw %%ymm6,%%ymm0,%%ymm0 \n" |
529 | 0 | "vpor %%ymm7,%%ymm0,%%ymm0 \n" |
530 | 0 | "vpunpcklbw %%ymm0,%%ymm1,%%ymm2 \n" |
531 | 0 | "vpunpckhbw %%ymm0,%%ymm1,%%ymm1 \n" |
532 | 0 | "vperm2i128 $0x20,%%ymm1,%%ymm2,%%ymm0 \n" |
533 | 0 | "vperm2i128 $0x31,%%ymm1,%%ymm2,%%ymm1 \n" |
534 | 0 | "vmovdqu %%ymm0,(%1,%0,2) \n" |
535 | 0 | "vmovdqu %%ymm1,0x20(%1,%0,2) \n" |
536 | 0 | "lea 0x20(%0),%0 \n" |
537 | 0 | "sub $0x10,%2 \n" |
538 | 0 | "jg 1b \n" |
539 | 0 | "vzeroupper \n" |
540 | 0 | : "+r"(src), // %0 |
541 | 0 | "+r"(dst), // %1 |
542 | 0 | "+r"(width) // %2 |
543 | 0 | : |
544 | 0 | : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
545 | 0 | "xmm6", "xmm7"); |
546 | 0 | } |
547 | | #endif |
548 | | |
549 | | #ifdef HAS_ARGB1555TOARGBROW_AVX2 |
550 | 0 | void ARGB1555ToARGBRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
551 | 0 | asm volatile( |
552 | 0 | "mov $0x1080108,%%eax \n" |
553 | 0 | "vmovd %%eax,%%xmm5 \n" |
554 | 0 | "vpbroadcastd %%xmm5,%%ymm5 \n" |
555 | 0 | "mov $0x42004200,%%eax \n" |
556 | 0 | "vmovd %%eax,%%xmm6 \n" |
557 | 0 | "vpbroadcastd %%xmm6,%%ymm6 \n" |
558 | 0 | "vpcmpeqb %%ymm3,%%ymm3,%%ymm3 \n" |
559 | 0 | "vpsllw $0xb,%%ymm3,%%ymm3 \n" |
560 | 0 | "vpsrlw $0x6,%%ymm3,%%ymm4 \n" |
561 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" |
562 | 0 | "vpsllw $0x8,%%ymm7,%%ymm7 \n" |
563 | 0 | "sub %0,%1 \n" |
564 | 0 | "sub %0,%1 \n" |
565 | |
|
566 | 0 | LABELALIGN |
567 | 0 | "1: \n" |
568 | 0 | "vmovdqu (%0),%%ymm0 \n" |
569 | 0 | "vpsllw $0x1,%%ymm0,%%ymm1 \n" |
570 | 0 | "vpsllw $0xb,%%ymm0,%%ymm2 \n" |
571 | 0 | "vpand %%ymm3,%%ymm1,%%ymm1 \n" |
572 | 0 | "vpmulhuw %%ymm5,%%ymm2,%%ymm2 \n" |
573 | 0 | "vpmulhuw %%ymm5,%%ymm1,%%ymm1 \n" |
574 | 0 | "vpsllw $0x8,%%ymm1,%%ymm1 \n" |
575 | 0 | "vpor %%ymm2,%%ymm1,%%ymm1 \n" |
576 | 0 | "vpsraw $0x8,%%ymm0,%%ymm2 \n" |
577 | 0 | "vpand %%ymm4,%%ymm0,%%ymm0 \n" |
578 | 0 | "vpmulhuw %%ymm6,%%ymm0,%%ymm0 \n" |
579 | 0 | "vpand %%ymm7,%%ymm2,%%ymm2 \n" |
580 | 0 | "vpor %%ymm2,%%ymm0,%%ymm0 \n" |
581 | 0 | "vpunpcklbw %%ymm0,%%ymm1,%%ymm2 \n" |
582 | 0 | "vpunpckhbw %%ymm0,%%ymm1,%%ymm1 \n" |
583 | 0 | "vperm2i128 $0x20,%%ymm1,%%ymm2,%%ymm0 \n" |
584 | 0 | "vperm2i128 $0x31,%%ymm1,%%ymm2,%%ymm1 \n" |
585 | 0 | "vmovdqu %%ymm0,(%1,%0,2) \n" |
586 | 0 | "vmovdqu %%ymm1,0x20(%1,%0,2) \n" |
587 | 0 | "lea 0x20(%0),%0 \n" |
588 | 0 | "sub $0x10,%2 \n" |
589 | 0 | "jg 1b \n" |
590 | 0 | "vzeroupper \n" |
591 | 0 | : "+r"(src), // %0 |
592 | 0 | "+r"(dst), // %1 |
593 | 0 | "+r"(width) // %2 |
594 | 0 | : |
595 | 0 | : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
596 | 0 | "xmm6", "xmm7"); |
597 | 0 | } |
598 | | #endif |
599 | | |
600 | | #ifdef HAS_ARGB4444TOARGBROW_AVX2 |
601 | 0 | void ARGB4444ToARGBRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
602 | 0 | asm volatile( |
603 | 0 | "mov $0x0f0f0f0f,%%eax \n" |
604 | 0 | "vmovd %%eax,%%xmm4 \n" |
605 | 0 | "vpbroadcastd %%xmm4,%%ymm4 \n" |
606 | 0 | "vpslld $0x4,%%ymm4,%%ymm5 \n" |
607 | 0 | "sub %0,%1 \n" |
608 | 0 | "sub %0,%1 \n" |
609 | |
|
610 | 0 | LABELALIGN |
611 | 0 | "1: \n" |
612 | 0 | "vmovdqu (%0),%%ymm0 \n" |
613 | 0 | "vpand %%ymm5,%%ymm0,%%ymm2 \n" |
614 | 0 | "vpand %%ymm4,%%ymm0,%%ymm0 \n" |
615 | 0 | "vpsllw $0x4,%%ymm0,%%ymm1 \n" |
616 | 0 | "vpsrlw $0x4,%%ymm2,%%ymm3 \n" |
617 | 0 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" |
618 | 0 | "vpor %%ymm3,%%ymm2,%%ymm2 \n" |
619 | 0 | "vpunpckhbw %%ymm2,%%ymm0,%%ymm1 \n" |
620 | 0 | "vpunpcklbw %%ymm2,%%ymm0,%%ymm0 \n" |
621 | 0 | "vperm2i128 $0x20,%%ymm1,%%ymm0,%%ymm2 \n" |
622 | 0 | "vperm2i128 $0x31,%%ymm1,%%ymm0,%%ymm1 \n" |
623 | 0 | "vmovdqu %%ymm2,(%1,%0,2) \n" |
624 | 0 | "vmovdqu %%ymm1,0x20(%1,%0,2) \n" |
625 | 0 | "lea 0x20(%0),%0 \n" |
626 | 0 | "sub $0x10,%2 \n" |
627 | 0 | "jg 1b \n" |
628 | 0 | "vzeroupper \n" |
629 | 0 | : "+r"(src), // %0 |
630 | 0 | "+r"(dst), // %1 |
631 | 0 | "+r"(width) // %2 |
632 | 0 | : |
633 | 0 | : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
634 | 0 | } |
635 | | #endif |
636 | | |
637 | 0 | void ARGBToRGB24Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) { |
638 | 0 | asm volatile("movdqa %3,%%xmm6 \n" |
639 | |
|
640 | 0 | LABELALIGN |
641 | 0 | "1: \n" |
642 | 0 | "movdqu (%0),%%xmm0 \n" |
643 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
644 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
645 | 0 | "movdqu 0x30(%0),%%xmm3 \n" |
646 | 0 | "lea 0x40(%0),%0 \n" |
647 | 0 | "pshufb %%xmm6,%%xmm0 \n" |
648 | 0 | "pshufb %%xmm6,%%xmm1 \n" |
649 | 0 | "pshufb %%xmm6,%%xmm2 \n" |
650 | 0 | "pshufb %%xmm6,%%xmm3 \n" |
651 | 0 | "movdqa %%xmm1,%%xmm4 \n" |
652 | 0 | "psrldq $0x4,%%xmm1 \n" |
653 | 0 | "pslldq $0xc,%%xmm4 \n" |
654 | 0 | "movdqa %%xmm2,%%xmm5 \n" |
655 | 0 | "por %%xmm4,%%xmm0 \n" |
656 | 0 | "pslldq $0x8,%%xmm5 \n" |
657 | 0 | "movdqu %%xmm0,(%1) \n" |
658 | 0 | "por %%xmm5,%%xmm1 \n" |
659 | 0 | "psrldq $0x8,%%xmm2 \n" |
660 | 0 | "pslldq $0x4,%%xmm3 \n" |
661 | 0 | "por %%xmm3,%%xmm2 \n" |
662 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
663 | 0 | "movdqu %%xmm2,0x20(%1) \n" |
664 | 0 | "lea 0x30(%1),%1 \n" |
665 | 0 | "sub $0x10,%2 \n" |
666 | 0 | "jg 1b \n" |
667 | 0 | : "+r"(src), // %0 |
668 | 0 | "+r"(dst), // %1 |
669 | 0 | "+r"(width) // %2 |
670 | 0 | : "m"(kShuffleMaskARGBToRGB24) // %3 |
671 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
672 | 0 | "xmm6"); |
673 | 0 | } |
674 | | |
675 | 0 | void ARGBToRAWRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) { |
676 | 0 | asm volatile("movdqa %3,%%xmm6 \n" |
677 | |
|
678 | 0 | LABELALIGN |
679 | 0 | "1: \n" |
680 | 0 | "movdqu (%0),%%xmm0 \n" |
681 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
682 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
683 | 0 | "movdqu 0x30(%0),%%xmm3 \n" |
684 | 0 | "lea 0x40(%0),%0 \n" |
685 | 0 | "pshufb %%xmm6,%%xmm0 \n" |
686 | 0 | "pshufb %%xmm6,%%xmm1 \n" |
687 | 0 | "pshufb %%xmm6,%%xmm2 \n" |
688 | 0 | "pshufb %%xmm6,%%xmm3 \n" |
689 | 0 | "movdqa %%xmm1,%%xmm4 \n" |
690 | 0 | "psrldq $0x4,%%xmm1 \n" |
691 | 0 | "pslldq $0xc,%%xmm4 \n" |
692 | 0 | "movdqa %%xmm2,%%xmm5 \n" |
693 | 0 | "por %%xmm4,%%xmm0 \n" |
694 | 0 | "pslldq $0x8,%%xmm5 \n" |
695 | 0 | "movdqu %%xmm0,(%1) \n" |
696 | 0 | "por %%xmm5,%%xmm1 \n" |
697 | 0 | "psrldq $0x8,%%xmm2 \n" |
698 | 0 | "pslldq $0x4,%%xmm3 \n" |
699 | 0 | "por %%xmm3,%%xmm2 \n" |
700 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
701 | 0 | "movdqu %%xmm2,0x20(%1) \n" |
702 | 0 | "lea 0x30(%1),%1 \n" |
703 | 0 | "sub $0x10,%2 \n" |
704 | 0 | "jg 1b \n" |
705 | 0 | : "+r"(src), // %0 |
706 | 0 | "+r"(dst), // %1 |
707 | 0 | "+r"(width) // %2 |
708 | 0 | : "m"(kShuffleMaskARGBToRAW) // %3 |
709 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
710 | 0 | "xmm6"); |
711 | 0 | } |
712 | | |
713 | | #ifdef HAS_ARGBTORGB24ROW_AVX2 |
714 | | // vpermd for 12+12 to 24 |
715 | | static const lvec32 kPermdRGB24_AVX = {0, 1, 2, 4, 5, 6, 3, 7}; |
716 | | |
717 | 10.3k | void ARGBToRGB24Row_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
718 | 10.3k | asm volatile( |
719 | 10.3k | "vbroadcasti128 %3,%%ymm6 \n" |
720 | 10.3k | "vmovdqa %4,%%ymm7 \n" |
721 | | |
722 | 10.3k | LABELALIGN |
723 | 10.3k | "1: \n" |
724 | 10.3k | "vmovdqu (%0),%%ymm0 \n" |
725 | 10.3k | "vmovdqu 0x20(%0),%%ymm1 \n" |
726 | 10.3k | "vmovdqu 0x40(%0),%%ymm2 \n" |
727 | 10.3k | "vmovdqu 0x60(%0),%%ymm3 \n" |
728 | 10.3k | "lea 0x80(%0),%0 \n" |
729 | 10.3k | "vpshufb %%ymm6,%%ymm0,%%ymm0 \n" // xxx0yyy0 |
730 | 10.3k | "vpshufb %%ymm6,%%ymm1,%%ymm1 \n" |
731 | 10.3k | "vpshufb %%ymm6,%%ymm2,%%ymm2 \n" |
732 | 10.3k | "vpshufb %%ymm6,%%ymm3,%%ymm3 \n" |
733 | 10.3k | "vpermd %%ymm0,%%ymm7,%%ymm0 \n" // pack to 24 bytes |
734 | 10.3k | "vpermd %%ymm1,%%ymm7,%%ymm1 \n" |
735 | 10.3k | "vpermd %%ymm2,%%ymm7,%%ymm2 \n" |
736 | 10.3k | "vpermd %%ymm3,%%ymm7,%%ymm3 \n" |
737 | 10.3k | "vpermq $0x3f,%%ymm1,%%ymm4 \n" // combine 24 + 8 |
738 | 10.3k | "vpor %%ymm4,%%ymm0,%%ymm0 \n" |
739 | 10.3k | "vmovdqu %%ymm0,(%1) \n" |
740 | 10.3k | "vpermq $0xf9,%%ymm1,%%ymm1 \n" // combine 16 + 16 |
741 | 10.3k | "vpermq $0x4f,%%ymm2,%%ymm4 \n" |
742 | 10.3k | "vpor %%ymm4,%%ymm1,%%ymm1 \n" |
743 | 10.3k | "vmovdqu %%ymm1,0x20(%1) \n" |
744 | 10.3k | "vpermq $0xfe,%%ymm2,%%ymm2 \n" // combine 8 + 24 |
745 | 10.3k | "vpermq $0x93,%%ymm3,%%ymm3 \n" |
746 | 10.3k | "vpor %%ymm3,%%ymm2,%%ymm2 \n" |
747 | 10.3k | "vmovdqu %%ymm2,0x40(%1) \n" |
748 | 10.3k | "lea 0x60(%1),%1 \n" |
749 | 10.3k | "sub $0x20,%2 \n" |
750 | 10.3k | "jg 1b \n" |
751 | 10.3k | "vzeroupper \n" |
752 | 10.3k | : "+r"(src), // %0 |
753 | 10.3k | "+r"(dst), // %1 |
754 | 10.3k | "+r"(width) // %2 |
755 | 10.3k | : "m"(kShuffleMaskARGBToRGB24), // %3 |
756 | 10.3k | "m"(kPermdRGB24_AVX) // %4 |
757 | 10.3k | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
758 | 10.3k | "xmm7"); |
759 | 10.3k | } |
760 | | #endif |
761 | | |
762 | | #ifdef HAS_ARGBTORGB24ROW_AVX512VBMI |
763 | | // Shuffle table for converting ARGBToRGB24 |
764 | | static const ulvec8 kPermARGBToRGB24_0 = { |
765 | | 0u, 1u, 2u, 4u, 5u, 6u, 8u, 9u, 10u, 12u, 13u, |
766 | | 14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u, 25u, 26u, 28u, |
767 | | 29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u, 40u, 41u}; |
768 | | static const ulvec8 kPermARGBToRGB24_1 = { |
769 | | 10u, 12u, 13u, 14u, 16u, 17u, 18u, 20u, 21u, 22u, 24u, |
770 | | 25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u, 36u, 37u, 38u, |
771 | | 40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u, 50u, 52u}; |
772 | | static const ulvec8 kPermARGBToRGB24_2 = { |
773 | | 21u, 22u, 24u, 25u, 26u, 28u, 29u, 30u, 32u, 33u, 34u, |
774 | | 36u, 37u, 38u, 40u, 41u, 42u, 44u, 45u, 46u, 48u, 49u, |
775 | | 50u, 52u, 53u, 54u, 56u, 57u, 58u, 60u, 61u, 62u}; |
776 | | |
777 | 0 | void ARGBToRGB24Row_AVX512VBMI(const uint8_t* src, uint8_t* dst, int width) { |
778 | 0 | asm volatile( |
779 | 0 | "vmovdqa %3,%%ymm5 \n" |
780 | 0 | "vmovdqa %4,%%ymm6 \n" |
781 | 0 | "vmovdqa %5,%%ymm7 \n" |
782 | |
|
783 | 0 | LABELALIGN |
784 | 0 | "1: \n" |
785 | 0 | "vmovdqu (%0),%%ymm0 \n" |
786 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
787 | 0 | "vmovdqu 0x40(%0),%%ymm2 \n" |
788 | 0 | "vmovdqu 0x60(%0),%%ymm3 \n" |
789 | 0 | "lea 0x80(%0),%0 \n" |
790 | 0 | "vpermt2b %%ymm1,%%ymm5,%%ymm0 \n" |
791 | 0 | "vpermt2b %%ymm2,%%ymm6,%%ymm1 \n" |
792 | 0 | "vpermt2b %%ymm3,%%ymm7,%%ymm2 \n" |
793 | 0 | "vmovdqu %%ymm0,(%1) \n" |
794 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
795 | 0 | "vmovdqu %%ymm2,0x40(%1) \n" |
796 | 0 | "lea 0x60(%1),%1 \n" |
797 | 0 | "sub $0x20,%2 \n" |
798 | 0 | "jg 1b \n" |
799 | 0 | "vzeroupper \n" |
800 | 0 | : "+r"(src), // %0 |
801 | 0 | "+r"(dst), // %1 |
802 | 0 | "+r"(width) // %2 |
803 | 0 | : "m"(kPermARGBToRGB24_0), // %3 |
804 | 0 | "m"(kPermARGBToRGB24_1), // %4 |
805 | 0 | "m"(kPermARGBToRGB24_2) // %5 |
806 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5", "xmm6", "xmm7"); |
807 | 0 | } |
808 | | #endif |
809 | | |
810 | | #ifdef HAS_ARGBTORAWROW_AVX2 |
811 | 0 | void ARGBToRAWRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
812 | 0 | asm volatile( |
813 | 0 | "vbroadcasti128 %3,%%ymm6 \n" |
814 | 0 | "vmovdqa %4,%%ymm7 \n" |
815 | |
|
816 | 0 | LABELALIGN |
817 | 0 | "1: \n" |
818 | 0 | "vmovdqu (%0),%%ymm0 \n" |
819 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
820 | 0 | "vmovdqu 0x40(%0),%%ymm2 \n" |
821 | 0 | "vmovdqu 0x60(%0),%%ymm3 \n" |
822 | 0 | "lea 0x80(%0),%0 \n" |
823 | 0 | "vpshufb %%ymm6,%%ymm0,%%ymm0 \n" // xxx0yyy0 |
824 | 0 | "vpshufb %%ymm6,%%ymm1,%%ymm1 \n" |
825 | 0 | "vpshufb %%ymm6,%%ymm2,%%ymm2 \n" |
826 | 0 | "vpshufb %%ymm6,%%ymm3,%%ymm3 \n" |
827 | 0 | "vpermd %%ymm0,%%ymm7,%%ymm0 \n" // pack to 24 bytes |
828 | 0 | "vpermd %%ymm1,%%ymm7,%%ymm1 \n" |
829 | 0 | "vpermd %%ymm2,%%ymm7,%%ymm2 \n" |
830 | 0 | "vpermd %%ymm3,%%ymm7,%%ymm3 \n" |
831 | 0 | "vpermq $0x3f,%%ymm1,%%ymm4 \n" // combine 24 + 8 |
832 | 0 | "vpor %%ymm4,%%ymm0,%%ymm0 \n" |
833 | 0 | "vmovdqu %%ymm0,(%1) \n" |
834 | 0 | "vpermq $0xf9,%%ymm1,%%ymm1 \n" // combine 16 + 16 |
835 | 0 | "vpermq $0x4f,%%ymm2,%%ymm4 \n" |
836 | 0 | "vpor %%ymm4,%%ymm1,%%ymm1 \n" |
837 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
838 | 0 | "vpermq $0xfe,%%ymm2,%%ymm2 \n" // combine 8 + 24 |
839 | 0 | "vpermq $0x93,%%ymm3,%%ymm3 \n" |
840 | 0 | "vpor %%ymm3,%%ymm2,%%ymm2 \n" |
841 | 0 | "vmovdqu %%ymm2,0x40(%1) \n" |
842 | 0 | "lea 0x60(%1),%1 \n" |
843 | 0 | "sub $0x20,%2 \n" |
844 | 0 | "jg 1b \n" |
845 | 0 | "vzeroupper \n" |
846 | 0 | : "+r"(src), // %0 |
847 | 0 | "+r"(dst), // %1 |
848 | 0 | "+r"(width) // %2 |
849 | 0 | : "m"(kShuffleMaskARGBToRAW), // %3 |
850 | 0 | "m"(kPermdRGB24_AVX) // %4 |
851 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
852 | 0 | "xmm7"); |
853 | 0 | } |
854 | | #endif |
855 | | |
856 | | #ifdef HAS_ARGBTORGB565DITHERROW_AVX2 |
857 | | void ARGBToRGB565DitherRow_AVX2(const uint8_t* src, |
858 | | uint8_t* dst, |
859 | | uint32_t dither4, |
860 | 0 | int width) { |
861 | 0 | asm volatile( |
862 | 0 | "vbroadcastss %3,%%xmm6 \n" |
863 | 0 | "vpunpcklbw %%xmm6,%%xmm6,%%xmm6 \n" |
864 | 0 | "vpermq $0xd8,%%ymm6,%%ymm6 \n" |
865 | 0 | "vpunpcklwd %%ymm6,%%ymm6,%%ymm6 \n" |
866 | 0 | "vpcmpeqb %%ymm3,%%ymm3,%%ymm3 \n" |
867 | 0 | "vpsrld $0x1b,%%ymm3,%%ymm3 \n" |
868 | 0 | "vpcmpeqb %%ymm4,%%ymm4,%%ymm4 \n" |
869 | 0 | "vpsrld $0x1a,%%ymm4,%%ymm4 \n" |
870 | 0 | "vpslld $0x5,%%ymm4,%%ymm4 \n" |
871 | 0 | "vpslld $0xb,%%ymm3,%%ymm5 \n" |
872 | |
|
873 | 0 | LABELALIGN |
874 | 0 | "1: \n" |
875 | 0 | "vmovdqu (%0),%%ymm0 \n" |
876 | 0 | "vpaddusb %%ymm6,%%ymm0,%%ymm0 \n" |
877 | 0 | "vpsrld $0x5,%%ymm0,%%ymm2 \n" |
878 | 0 | "vpsrld $0x3,%%ymm0,%%ymm1 \n" |
879 | 0 | "vpsrld $0x8,%%ymm0,%%ymm0 \n" |
880 | 0 | "vpand %%ymm4,%%ymm2,%%ymm2 \n" |
881 | 0 | "vpand %%ymm3,%%ymm1,%%ymm1 \n" |
882 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" |
883 | 0 | "vpor %%ymm2,%%ymm1,%%ymm1 \n" |
884 | 0 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" |
885 | 0 | "vpackusdw %%ymm0,%%ymm0,%%ymm0 \n" |
886 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
887 | 0 | "lea 0x20(%0),%0 \n" |
888 | 0 | "vmovdqu %%xmm0,(%1) \n" |
889 | 0 | "lea 0x10(%1),%1 \n" |
890 | 0 | "sub $0x8,%2 \n" |
891 | 0 | "jg 1b \n" |
892 | 0 | "vzeroupper \n" |
893 | 0 | : "+r"(src), // %0 |
894 | 0 | "+r"(dst), // %1 |
895 | 0 | "+r"(width) // %2 |
896 | 0 | : "m"(dither4) // %3 |
897 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
898 | 0 | "xmm7"); |
899 | 0 | } |
900 | | #endif // HAS_ARGBTORGB565DITHERROW_AVX2 |
901 | | |
902 | | #endif // HAS_RGB24TOARGBROW_SSSE3 |
903 | | |
904 | | /* |
905 | | |
906 | | ARGBToAR30Row: |
907 | | |
908 | | Red Blue |
909 | | With the 8 bit value in the upper bits of a short, vpmulhuw by (1024+4) will |
910 | | produce a 10 bit value in the low 10 bits of each 16 bit value. This is whats |
911 | | wanted for the blue channel. The red needs to be shifted 4 left, so multiply by |
912 | | (1024+4)*16 for red. |
913 | | |
914 | | Alpha Green |
915 | | Alpha and Green are already in the high bits so vpand can zero out the other |
916 | | bits, keeping just 2 upper bits of alpha and 8 bit green. The same multiplier |
917 | | could be used for Green - (1024+4) putting the 10 bit green in the lsb. Alpha |
918 | | would be a simple multiplier to shift it into position. It wants a gap of 10 |
919 | | above the green. Green is 10 bits, so there are 6 bits in the low short. 4 |
920 | | more are needed, so a multiplier of 4 gets the 2 bits into the upper 16 bits, |
921 | | and then a shift of 4 is a multiply of 16, so (4*16) = 64. Then shift the |
922 | | result left 10 to position the A and G channels. |
923 | | */ |
924 | | |
925 | | // Shuffle table for converting RAW to RGB24. Last 8. |
926 | | static const uvec8 kShuffleRB30 = {128u, 0u, 128u, 2u, 128u, 4u, 128u, 6u, |
927 | | 128u, 8u, 128u, 10u, 128u, 12u, 128u, 14u}; |
928 | | |
929 | | static const uvec8 kShuffleBR30 = {128u, 2u, 128u, 0u, 128u, 6u, 128u, 4u, |
930 | | 128u, 10u, 128u, 8u, 128u, 14u, 128u, 12u}; |
931 | | |
932 | | static const uint32_t kMulRB10 = 1028 * 16 * 65536 + 1028; |
933 | | static const uint32_t kMaskRB10 = 0x3ff003ff; |
934 | | static const uint32_t kMaskAG10 = 0xc000ff00; |
935 | | static const uint32_t kMulAG10 = 64 * 65536 + 1028; |
936 | | |
937 | 0 | void ARGBToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) { |
938 | 0 | asm volatile( |
939 | 0 | "movdqa %3,%%xmm2 \n" // shuffler for RB |
940 | 0 | "movd %4,%%xmm3 \n" // multipler for RB |
941 | 0 | "movd %5,%%xmm4 \n" // mask for R10 B10 |
942 | 0 | "movd %6,%%xmm5 \n" // mask for AG |
943 | 0 | "movd %7,%%xmm6 \n" // multipler for AG |
944 | 0 | "pshufd $0x0,%%xmm3,%%xmm3 \n" |
945 | 0 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
946 | 0 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
947 | 0 | "pshufd $0x0,%%xmm6,%%xmm6 \n" |
948 | 0 | "sub %0,%1 \n" |
949 | |
|
950 | 0 | "1: \n" |
951 | 0 | "movdqu (%0),%%xmm0 \n" // fetch 4 ARGB pixels |
952 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
953 | 0 | "pshufb %%xmm2,%%xmm1 \n" // R0B0 |
954 | 0 | "pand %%xmm5,%%xmm0 \n" // A0G0 |
955 | 0 | "pmulhuw %%xmm3,%%xmm1 \n" // X2 R16 X4 B10 |
956 | 0 | "pmulhuw %%xmm6,%%xmm0 \n" // X10 A2 X10 G10 |
957 | 0 | "pand %%xmm4,%%xmm1 \n" // X2 R10 X10 B10 |
958 | 0 | "pslld $10,%%xmm0 \n" // A2 x10 G10 x10 |
959 | 0 | "por %%xmm1,%%xmm0 \n" // A2 R10 G10 B10 |
960 | 0 | "movdqu %%xmm0,(%1,%0) \n" // store 4 AR30 pixels |
961 | 0 | "add $0x10,%0 \n" |
962 | 0 | "sub $0x4,%2 \n" |
963 | 0 | "jg 1b \n" |
964 | |
|
965 | 0 | : "+r"(src), // %0 |
966 | 0 | "+r"(dst), // %1 |
967 | 0 | "+r"(width) // %2 |
968 | 0 | : "m"(kShuffleRB30), // %3 |
969 | 0 | "m"(kMulRB10), // %4 |
970 | 0 | "m"(kMaskRB10), // %5 |
971 | 0 | "m"(kMaskAG10), // %6 |
972 | 0 | "m"(kMulAG10) // %7 |
973 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
974 | 0 | } |
975 | | |
976 | 0 | void ABGRToAR30Row_SSSE3(const uint8_t* src, uint8_t* dst, int width) { |
977 | 0 | asm volatile( |
978 | 0 | "movdqa %3,%%xmm2 \n" // shuffler for RB |
979 | 0 | "movd %4,%%xmm3 \n" // multipler for RB |
980 | 0 | "movd %5,%%xmm4 \n" // mask for R10 B10 |
981 | 0 | "movd %6,%%xmm5 \n" // mask for AG |
982 | 0 | "movd %7,%%xmm6 \n" // multipler for AG |
983 | 0 | "pshufd $0x0,%%xmm3,%%xmm3 \n" |
984 | 0 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
985 | 0 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
986 | 0 | "pshufd $0x0,%%xmm6,%%xmm6 \n" |
987 | 0 | "sub %0,%1 \n" |
988 | |
|
989 | 0 | "1: \n" |
990 | 0 | "movdqu (%0),%%xmm0 \n" // fetch 4 ABGR pixels |
991 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
992 | 0 | "pshufb %%xmm2,%%xmm1 \n" // R0B0 |
993 | 0 | "pand %%xmm5,%%xmm0 \n" // A0G0 |
994 | 0 | "pmulhuw %%xmm3,%%xmm1 \n" // X2 R16 X4 B10 |
995 | 0 | "pmulhuw %%xmm6,%%xmm0 \n" // X10 A2 X10 G10 |
996 | 0 | "pand %%xmm4,%%xmm1 \n" // X2 R10 X10 B10 |
997 | 0 | "pslld $10,%%xmm0 \n" // A2 x10 G10 x10 |
998 | 0 | "por %%xmm1,%%xmm0 \n" // A2 R10 G10 B10 |
999 | 0 | "movdqu %%xmm0,(%1,%0) \n" // store 4 AR30 pixels |
1000 | 0 | "add $0x10,%0 \n" |
1001 | 0 | "sub $0x4,%2 \n" |
1002 | 0 | "jg 1b \n" |
1003 | |
|
1004 | 0 | : "+r"(src), // %0 |
1005 | 0 | "+r"(dst), // %1 |
1006 | 0 | "+r"(width) // %2 |
1007 | 0 | : "m"(kShuffleBR30), // %3 reversed shuffler |
1008 | 0 | "m"(kMulRB10), // %4 |
1009 | 0 | "m"(kMaskRB10), // %5 |
1010 | 0 | "m"(kMaskAG10), // %6 |
1011 | 0 | "m"(kMulAG10) // %7 |
1012 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
1013 | 0 | } |
1014 | | |
1015 | | #ifdef HAS_ARGBTOAR30ROW_AVX2 |
1016 | 0 | void ARGBToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
1017 | 0 | asm volatile( |
1018 | 0 | "vbroadcasti128 %3,%%ymm2 \n" // shuffler for RB |
1019 | 0 | "vbroadcastss %4,%%ymm3 \n" // multipler for RB |
1020 | 0 | "vbroadcastss %5,%%ymm4 \n" // mask for R10 B10 |
1021 | 0 | "vbroadcastss %6,%%ymm5 \n" // mask for AG |
1022 | 0 | "vbroadcastss %7,%%ymm6 \n" // multipler for AG |
1023 | 0 | "sub %0,%1 \n" |
1024 | |
|
1025 | 0 | "1: \n" |
1026 | 0 | "vmovdqu (%0),%%ymm0 \n" // fetch 8 ARGB pixels |
1027 | 0 | "vpshufb %%ymm2,%%ymm0,%%ymm1 \n" // R0B0 |
1028 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" // A0G0 |
1029 | 0 | "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" // X2 R16 X4 B10 |
1030 | 0 | "vpmulhuw %%ymm6,%%ymm0,%%ymm0 \n" // X10 A2 X10 G10 |
1031 | 0 | "vpand %%ymm4,%%ymm1,%%ymm1 \n" // X2 R10 X10 B10 |
1032 | 0 | "vpslld $10,%%ymm0,%%ymm0 \n" // A2 x10 G10 x10 |
1033 | 0 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" // A2 R10 G10 B10 |
1034 | 0 | "vmovdqu %%ymm0,(%1,%0) \n" // store 8 AR30 pixels |
1035 | 0 | "add $0x20,%0 \n" |
1036 | 0 | "sub $0x8,%2 \n" |
1037 | 0 | "jg 1b \n" |
1038 | 0 | "vzeroupper \n" |
1039 | |
|
1040 | 0 | : "+r"(src), // %0 |
1041 | 0 | "+r"(dst), // %1 |
1042 | 0 | "+r"(width) // %2 |
1043 | 0 | : "m"(kShuffleRB30), // %3 |
1044 | 0 | "m"(kMulRB10), // %4 |
1045 | 0 | "m"(kMaskRB10), // %5 |
1046 | 0 | "m"(kMaskAG10), // %6 |
1047 | 0 | "m"(kMulAG10) // %7 |
1048 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
1049 | 0 | } |
1050 | | #endif |
1051 | | |
1052 | | #ifdef HAS_ABGRTOAR30ROW_AVX2 |
1053 | 0 | void ABGRToAR30Row_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
1054 | 0 | asm volatile( |
1055 | 0 | "vbroadcasti128 %3,%%ymm2 \n" // shuffler for RB |
1056 | 0 | "vbroadcastss %4,%%ymm3 \n" // multipler for RB |
1057 | 0 | "vbroadcastss %5,%%ymm4 \n" // mask for R10 B10 |
1058 | 0 | "vbroadcastss %6,%%ymm5 \n" // mask for AG |
1059 | 0 | "vbroadcastss %7,%%ymm6 \n" // multipler for AG |
1060 | 0 | "sub %0,%1 \n" |
1061 | |
|
1062 | 0 | "1: \n" |
1063 | 0 | "vmovdqu (%0),%%ymm0 \n" // fetch 8 ABGR pixels |
1064 | 0 | "vpshufb %%ymm2,%%ymm0,%%ymm1 \n" // R0B0 |
1065 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" // A0G0 |
1066 | 0 | "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" // X2 R16 X4 B10 |
1067 | 0 | "vpmulhuw %%ymm6,%%ymm0,%%ymm0 \n" // X10 A2 X10 G10 |
1068 | 0 | "vpand %%ymm4,%%ymm1,%%ymm1 \n" // X2 R10 X10 B10 |
1069 | 0 | "vpslld $10,%%ymm0,%%ymm0 \n" // A2 x10 G10 x10 |
1070 | 0 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" // A2 R10 G10 B10 |
1071 | 0 | "vmovdqu %%ymm0,(%1,%0) \n" // store 8 AR30 pixels |
1072 | 0 | "add $0x20,%0 \n" |
1073 | 0 | "sub $0x8,%2 \n" |
1074 | 0 | "jg 1b \n" |
1075 | 0 | "vzeroupper \n" |
1076 | |
|
1077 | 0 | : "+r"(src), // %0 |
1078 | 0 | "+r"(dst), // %1 |
1079 | 0 | "+r"(width) // %2 |
1080 | 0 | : "m"(kShuffleBR30), // %3 reversed shuffler |
1081 | 0 | "m"(kMulRB10), // %4 |
1082 | 0 | "m"(kMaskRB10), // %5 |
1083 | 0 | "m"(kMaskAG10), // %6 |
1084 | 0 | "m"(kMulAG10) // %7 |
1085 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
1086 | 0 | } |
1087 | | #endif |
1088 | | |
1089 | | static const uvec8 kShuffleARGBToABGR = {2, 1, 0, 3, 6, 5, 4, 7, |
1090 | | 10, 9, 8, 11, 14, 13, 12, 15}; |
1091 | | |
1092 | | static const uvec8 kShuffleARGBToAB64Lo = {2, 2, 1, 1, 0, 0, 3, 3, |
1093 | | 6, 6, 5, 5, 4, 4, 7, 7}; |
1094 | | static const uvec8 kShuffleARGBToAB64Hi = {10, 10, 9, 9, 8, 8, 11, 11, |
1095 | | 14, 14, 13, 13, 12, 12, 15, 15}; |
1096 | | |
1097 | | void ARGBToAR64Row_SSSE3(const uint8_t* src_argb, |
1098 | | uint16_t* dst_ar64, |
1099 | 0 | int width) { |
1100 | 0 | asm volatile( |
1101 | 0 | "1: \n" |
1102 | 0 | "movdqu (%0),%%xmm0 \n" |
1103 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
1104 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
1105 | 0 | "punpckhbw %%xmm1,%%xmm1 \n" |
1106 | 0 | "movdqu %%xmm0,(%1) \n" |
1107 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
1108 | 0 | "lea 0x10(%0),%0 \n" |
1109 | 0 | "lea 0x20(%1),%1 \n" |
1110 | 0 | "sub $0x4,%2 \n" |
1111 | 0 | "jg 1b \n" |
1112 | 0 | : "+r"(src_argb), // %0 |
1113 | 0 | "+r"(dst_ar64), // %1 |
1114 | 0 | "+r"(width) // %2 |
1115 | 0 | ::"memory", |
1116 | 0 | "cc", "xmm0", "xmm1"); |
1117 | 0 | } |
1118 | | |
1119 | | void ARGBToAB64Row_SSSE3(const uint8_t* src_argb, |
1120 | | uint16_t* dst_ab64, |
1121 | 0 | int width) { |
1122 | 0 | asm volatile( |
1123 | 0 | "movdqa %3,%%xmm2 \n" |
1124 | 0 | "movdqa %4,%%xmm3 \n" LABELALIGN |
1125 | 0 | "1: \n" |
1126 | 0 | "movdqu (%0),%%xmm0 \n" |
1127 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
1128 | 0 | "pshufb %%xmm2,%%xmm0 \n" |
1129 | 0 | "pshufb %%xmm3,%%xmm1 \n" |
1130 | 0 | "movdqu %%xmm0,(%1) \n" |
1131 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
1132 | 0 | "lea 0x10(%0),%0 \n" |
1133 | 0 | "lea 0x20(%1),%1 \n" |
1134 | 0 | "sub $0x4,%2 \n" |
1135 | 0 | "jg 1b \n" |
1136 | 0 | : "+r"(src_argb), // %0 |
1137 | 0 | "+r"(dst_ab64), // %1 |
1138 | 0 | "+r"(width) // %2 |
1139 | 0 | : "m"(kShuffleARGBToAB64Lo), // %3 |
1140 | 0 | "m"(kShuffleARGBToAB64Hi) // %4 |
1141 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
1142 | 0 | } |
1143 | | |
1144 | | void AR64ToARGBRow_SSSE3(const uint16_t* src_ar64, |
1145 | | uint8_t* dst_argb, |
1146 | 0 | int width) { |
1147 | 0 | asm volatile( |
1148 | 0 | "1: \n" |
1149 | 0 | "movdqu (%0),%%xmm0 \n" |
1150 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
1151 | 0 | "psrlw $8,%%xmm0 \n" |
1152 | 0 | "psrlw $8,%%xmm1 \n" |
1153 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
1154 | 0 | "movdqu %%xmm0,(%1) \n" |
1155 | 0 | "lea 0x20(%0),%0 \n" |
1156 | 0 | "lea 0x10(%1),%1 \n" |
1157 | 0 | "sub $0x4,%2 \n" |
1158 | 0 | "jg 1b \n" |
1159 | 0 | : "+r"(src_ar64), // %0 |
1160 | 0 | "+r"(dst_argb), // %1 |
1161 | 0 | "+r"(width) // %2 |
1162 | 0 | ::"memory", |
1163 | 0 | "cc", "xmm0", "xmm1"); |
1164 | 0 | } |
1165 | | |
1166 | | void AB64ToARGBRow_SSSE3(const uint16_t* src_ab64, |
1167 | | uint8_t* dst_argb, |
1168 | 0 | int width) { |
1169 | 0 | asm volatile("movdqa %3,%%xmm2 \n" |
1170 | |
|
1171 | 0 | LABELALIGN |
1172 | 0 | "1: \n" |
1173 | 0 | "movdqu (%0),%%xmm0 \n" |
1174 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
1175 | 0 | "psrlw $8,%%xmm0 \n" |
1176 | 0 | "psrlw $8,%%xmm1 \n" |
1177 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
1178 | 0 | "pshufb %%xmm2,%%xmm0 \n" |
1179 | 0 | "movdqu %%xmm0,(%1) \n" |
1180 | 0 | "lea 0x20(%0),%0 \n" |
1181 | 0 | "lea 0x10(%1),%1 \n" |
1182 | 0 | "sub $0x4,%2 \n" |
1183 | 0 | "jg 1b \n" |
1184 | 0 | : "+r"(src_ab64), // %0 |
1185 | 0 | "+r"(dst_argb), // %1 |
1186 | 0 | "+r"(width) // %2 |
1187 | 0 | : "m"(kShuffleARGBToABGR) // %3 |
1188 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
1189 | 0 | } |
1190 | | |
1191 | | #ifdef HAS_ARGBTOAR64ROW_AVX2 |
1192 | | void ARGBToAR64Row_AVX2(const uint8_t* src_argb, |
1193 | | uint16_t* dst_ar64, |
1194 | 0 | int width) { |
1195 | 0 | asm volatile( |
1196 | 0 | "1: \n" |
1197 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1198 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
1199 | 0 | "vpunpckhbw %%ymm0,%%ymm0,%%ymm1 \n" |
1200 | 0 | "vpunpcklbw %%ymm0,%%ymm0,%%ymm0 \n" |
1201 | 0 | "vmovdqu %%ymm0,(%1) \n" |
1202 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
1203 | 0 | "lea 0x20(%0),%0 \n" |
1204 | 0 | "lea 0x40(%1),%1 \n" |
1205 | 0 | "sub $0x8,%2 \n" |
1206 | 0 | "jg 1b \n" |
1207 | 0 | "vzeroupper \n" |
1208 | 0 | : "+r"(src_argb), // %0 |
1209 | 0 | "+r"(dst_ar64), // %1 |
1210 | 0 | "+r"(width) // %2 |
1211 | 0 | ::"memory", |
1212 | 0 | "cc", "xmm0", "xmm1"); |
1213 | 0 | } |
1214 | | #endif |
1215 | | |
1216 | | #ifdef HAS_ARGBTOAB64ROW_AVX2 |
1217 | | void ARGBToAB64Row_AVX2(const uint8_t* src_argb, |
1218 | | uint16_t* dst_ab64, |
1219 | 0 | int width) { |
1220 | 0 | asm volatile( |
1221 | 0 | "vbroadcasti128 %3,%%ymm2 \n" |
1222 | 0 | "vbroadcasti128 %4,%%ymm3 \n" LABELALIGN |
1223 | 0 | "1: \n" |
1224 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1225 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
1226 | 0 | "vpshufb %%ymm3,%%ymm0,%%ymm1 \n" |
1227 | 0 | "vpshufb %%ymm2,%%ymm0,%%ymm0 \n" |
1228 | 0 | "vmovdqu %%ymm0,(%1) \n" |
1229 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
1230 | 0 | "lea 0x20(%0),%0 \n" |
1231 | 0 | "lea 0x40(%1),%1 \n" |
1232 | 0 | "sub $0x8,%2 \n" |
1233 | 0 | "jg 1b \n" |
1234 | 0 | "vzeroupper \n" |
1235 | 0 | : "+r"(src_argb), // %0 |
1236 | 0 | "+r"(dst_ab64), // %1 |
1237 | 0 | "+r"(width) // %2 |
1238 | 0 | : "m"(kShuffleARGBToAB64Lo), // %3 |
1239 | 0 | "m"(kShuffleARGBToAB64Hi) // %3 |
1240 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
1241 | 0 | } |
1242 | | #endif |
1243 | | |
1244 | | #ifdef HAS_AR64TOARGBROW_AVX2 |
1245 | | void AR64ToARGBRow_AVX2(const uint16_t* src_ar64, |
1246 | | uint8_t* dst_argb, |
1247 | 0 | int width) { |
1248 | 0 | asm volatile( |
1249 | 0 | "1: \n" |
1250 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1251 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
1252 | 0 | "vpsrlw $8,%%ymm0,%%ymm0 \n" |
1253 | 0 | "vpsrlw $8,%%ymm1,%%ymm1 \n" |
1254 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
1255 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
1256 | 0 | "vmovdqu %%ymm0,(%1) \n" |
1257 | 0 | "lea 0x40(%0),%0 \n" |
1258 | 0 | "lea 0x20(%1),%1 \n" |
1259 | 0 | "sub $0x8,%2 \n" |
1260 | 0 | "jg 1b \n" |
1261 | 0 | "vzeroupper \n" |
1262 | 0 | : "+r"(src_ar64), // %0 |
1263 | 0 | "+r"(dst_argb), // %1 |
1264 | 0 | "+r"(width) // %2 |
1265 | 0 | ::"memory", |
1266 | 0 | "cc", "xmm0", "xmm1"); |
1267 | 0 | } |
1268 | | #endif |
1269 | | |
1270 | | #ifdef HAS_AB64TOARGBROW_AVX2 |
1271 | | void AB64ToARGBRow_AVX2(const uint16_t* src_ab64, |
1272 | | uint8_t* dst_argb, |
1273 | 0 | int width) { |
1274 | 0 | asm volatile("vbroadcasti128 %3,%%ymm2 \n" LABELALIGN |
1275 | 0 | "1: \n" |
1276 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1277 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
1278 | 0 | "vpsrlw $8,%%ymm0,%%ymm0 \n" |
1279 | 0 | "vpsrlw $8,%%ymm1,%%ymm1 \n" |
1280 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
1281 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
1282 | 0 | "vpshufb %%ymm2,%%ymm0,%%ymm0 \n" |
1283 | 0 | "vmovdqu %%ymm0,(%1) \n" |
1284 | 0 | "lea 0x40(%0),%0 \n" |
1285 | 0 | "lea 0x20(%1),%1 \n" |
1286 | 0 | "sub $0x8,%2 \n" |
1287 | 0 | "jg 1b \n" |
1288 | 0 | "vzeroupper \n" |
1289 | 0 | : "+r"(src_ab64), // %0 |
1290 | 0 | "+r"(dst_argb), // %1 |
1291 | 0 | "+r"(width) // %2 |
1292 | 0 | : "m"(kShuffleARGBToABGR) // %3 |
1293 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
1294 | 0 | } |
1295 | | #endif |
1296 | | |
1297 | | // clang-format off |
1298 | | |
1299 | | // TODO(mraptis): Consider passing R, G, B multipliers as parameter. |
1300 | | // round parameter is register containing value to add before shift. |
1301 | | #define RGBTOY(round) \ |
1302 | | "1: \n" \ |
1303 | | "movdqu (%0),%%xmm0 \n" \ |
1304 | | "movdqu 0x10(%0),%%xmm1 \n" \ |
1305 | | "movdqu 0x20(%0),%%xmm2 \n" \ |
1306 | | "movdqu 0x30(%0),%%xmm3 \n" \ |
1307 | | "psubb %%xmm5,%%xmm0 \n" \ |
1308 | | "psubb %%xmm5,%%xmm1 \n" \ |
1309 | | "psubb %%xmm5,%%xmm2 \n" \ |
1310 | | "psubb %%xmm5,%%xmm3 \n" \ |
1311 | | "movdqu %%xmm4,%%xmm6 \n" \ |
1312 | | "pmaddubsw %%xmm0,%%xmm6 \n" \ |
1313 | | "movdqu %%xmm4,%%xmm0 \n" \ |
1314 | | "pmaddubsw %%xmm1,%%xmm0 \n" \ |
1315 | | "movdqu %%xmm4,%%xmm1 \n" \ |
1316 | | "pmaddubsw %%xmm2,%%xmm1 \n" \ |
1317 | | "movdqu %%xmm4,%%xmm2 \n" \ |
1318 | | "pmaddubsw %%xmm3,%%xmm2 \n" \ |
1319 | | "lea 0x40(%0),%0 \n" \ |
1320 | | "phaddw %%xmm0,%%xmm6 \n" \ |
1321 | | "phaddw %%xmm2,%%xmm1 \n" \ |
1322 | | "prefetcht0 1280(%0) \n" \ |
1323 | | "paddw %%" #round ",%%xmm6 \n" \ |
1324 | | "paddw %%" #round ",%%xmm1 \n" \ |
1325 | | "psrlw $0x8,%%xmm6 \n" \ |
1326 | | "psrlw $0x8,%%xmm1 \n" \ |
1327 | | "packuswb %%xmm1,%%xmm6 \n" \ |
1328 | | "movdqu %%xmm6,(%1) \n" \ |
1329 | | "lea 0x10(%1),%1 \n" \ |
1330 | | "sub $0x10,%2 \n" \ |
1331 | | "jg 1b \n" |
1332 | | |
1333 | | #define RGBTOY_AVX2(round) \ |
1334 | | "1: \n" \ |
1335 | | "vmovdqu (%0),%%ymm0 \n" \ |
1336 | | "vmovdqu 0x20(%0),%%ymm1 \n" \ |
1337 | | "vmovdqu 0x40(%0),%%ymm2 \n" \ |
1338 | | "vmovdqu 0x60(%0),%%ymm3 \n" \ |
1339 | | "vpsubb %%ymm5, %%ymm0, %%ymm0 \n" \ |
1340 | | "vpsubb %%ymm5, %%ymm1, %%ymm1 \n" \ |
1341 | | "vpsubb %%ymm5, %%ymm2, %%ymm2 \n" \ |
1342 | | "vpsubb %%ymm5, %%ymm3, %%ymm3 \n" \ |
1343 | | "vpmaddubsw %%ymm0,%%ymm4,%%ymm0 \n" \ |
1344 | | "vpmaddubsw %%ymm1,%%ymm4,%%ymm1 \n" \ |
1345 | | "vpmaddubsw %%ymm2,%%ymm4,%%ymm2 \n" \ |
1346 | | "vpmaddubsw %%ymm3,%%ymm4,%%ymm3 \n" \ |
1347 | | "lea 0x80(%0),%0 \n" \ |
1348 | | "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" /* mutates. */ \ |
1349 | | "vphaddw %%ymm3,%%ymm2,%%ymm2 \n" \ |
1350 | | "prefetcht0 1280(%0) \n" \ |
1351 | | "vpaddw %%" #round ",%%ymm0,%%ymm0 \n" /* Add 16 */ \ |
1352 | | "vpaddw %%" #round ",%%ymm2,%%ymm2 \n" \ |
1353 | | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" \ |
1354 | | "vpsrlw $0x8,%%ymm2,%%ymm2 \n" \ |
1355 | | "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" /* mutates. */ \ |
1356 | | "vpermd %%ymm0,%%ymm6,%%ymm0 \n" /* unmutate. */ \ |
1357 | | "vmovdqu %%ymm0,(%1) \n" \ |
1358 | | "lea 0x20(%1),%1 \n" \ |
1359 | | "sub $0x20,%2 \n" \ |
1360 | | "jg 1b \n" |
1361 | | |
1362 | | // clang-format on |
1363 | | |
1364 | | #ifdef HAS_ARGBTOYROW_SSSE3 |
1365 | | // Convert 16 ARGB pixels (64 bytes) to 16 Y values. |
1366 | | |
1367 | 0 | void ARGBToYRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1368 | 0 | ARGBToYMatrixRow_SSSE3(src_argb, dst_y, width, &kArgbI601Constants); |
1369 | 0 | } |
1370 | | #endif // HAS_ARGBTOYROW_SSSE3 |
1371 | | |
1372 | | #ifdef HAS_ARGBTOYJROW_SSSE3 |
1373 | | // Convert 16 ARGB pixels (64 bytes) to 16 YJ values. |
1374 | | // Same as ARGBToYRow but different coefficients, no add 16. |
1375 | 0 | void ARGBToYJRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1376 | 0 | ARGBToYMatrixRow_SSSE3(src_argb, dst_y, width, &kArgbJPEGConstants); |
1377 | 0 | } |
1378 | | #endif // HAS_ARGBTOYJROW_SSSE3 |
1379 | | |
1380 | | #ifdef HAS_ABGRTOYJROW_SSSE3 |
1381 | | // Convert 16 ABGR pixels (64 bytes) to 16 YJ values. |
1382 | | // Same as ABGRToYRow but different coefficients, no add 16. |
1383 | 0 | void ABGRToYJRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
1384 | 0 | ARGBToYMatrixRow_SSSE3(src_abgr, dst_y, width, &kAbgrJPEGConstants); |
1385 | 0 | } |
1386 | | #endif // HAS_ABGRTOYJROW_SSSE3 |
1387 | | |
1388 | | #ifdef HAS_RGBATOYJROW_SSSE3 |
1389 | | // Convert 16 RGBA pixels (64 bytes) to 16 YJ values. |
1390 | | // Same as ARGBToYRow but different coefficients, no add 16. |
1391 | 0 | void RGBAToYJRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
1392 | 0 | ARGBToYMatrixRow_SSSE3(src_rgba, dst_y, width, &kRgbaJPEGConstants); |
1393 | 0 | } |
1394 | | #endif // HAS_RGBATOYJROW_SSSE3 |
1395 | | |
1396 | | #if defined(HAS_ARGBTOYROW_AVX2) || defined(HAS_ABGRTOYROW_AVX2) || \ |
1397 | | defined(HAS_ARGBEXTRACTALPHAROW_AVX2) |
1398 | | // vpermd for vphaddw + vpackuswb vpermd. |
1399 | | #endif |
1400 | | |
1401 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1402 | | |
1403 | | // Convert 32 ARGB pixels (128 bytes) to 32 Y values. |
1404 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1405 | 0 | void ARGBToYRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1406 | 0 | ARGBToYMatrixRow_AVX2(src_argb, dst_y, width, &kArgbI601Constants); |
1407 | 0 | } |
1408 | | #endif |
1409 | | #endif // HAS_ARGBTOYROW_AVX2 |
1410 | | |
1411 | | #ifdef HAS_ABGRTOYROW_AVX2 |
1412 | | // Convert 32 ABGR pixels (128 bytes) to 32 Y values. |
1413 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1414 | 0 | void ABGRToYRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
1415 | 0 | ARGBToYMatrixRow_AVX2(src_abgr, dst_y, width, &kAbgrI601Constants); |
1416 | 0 | } |
1417 | | #endif |
1418 | | #endif // HAS_ABGRTOYROW_AVX2 |
1419 | | |
1420 | | #ifdef HAS_ARGBTOYJROW_AVX2 |
1421 | | // Convert 32 ARGB pixels (128 bytes) to 32 Y values. |
1422 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1423 | 0 | void ARGBToYJRow_AVX2(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1424 | 0 | ARGBToYMatrixRow_AVX2(src_argb, dst_y, width, &kArgbJPEGConstants); |
1425 | 0 | } |
1426 | | #endif |
1427 | | |
1428 | | #endif // HAS_ARGBTOYJROW_AVX2 |
1429 | | |
1430 | | #ifdef HAS_ABGRTOYJROW_AVX2 |
1431 | | // Convert 32 ABGR pixels (128 bytes) to 32 Y values. |
1432 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1433 | 0 | void ABGRToYJRow_AVX2(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
1434 | 0 | ARGBToYMatrixRow_AVX2(src_abgr, dst_y, width, &kAbgrJPEGConstants); |
1435 | 0 | } |
1436 | | #endif |
1437 | | #endif // HAS_ABGRTOYJROW_AVX2 |
1438 | | |
1439 | | #ifdef HAS_RGBATOYJROW_AVX2 |
1440 | | // Convert 32 ARGB pixels (128 bytes) to 32 Y values. |
1441 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1442 | 0 | void RGBAToYJRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
1443 | 0 | ARGBToYMatrixRow_AVX2(src_rgba, dst_y, width, &kRgbaJPEGConstants); |
1444 | 0 | } |
1445 | | #endif |
1446 | | #endif // HAS_RGBATOYJROW_AVX2 |
1447 | | |
1448 | | #if defined(HAS_ARGBTOYROW_AVX2) || defined(HAS_ARGBTOUV444ROW_AVX2) || \ |
1449 | | defined(HAS_ARGBEXTRACTALPHAROW_AVX2) |
1450 | | // vpermd for vphaddw + vpackuswb vpermd. |
1451 | | static const lvec32 kPermdARGBToY_AVX = {0, 4, 1, 5, 2, 6, 3, 7}; |
1452 | | #endif |
1453 | | |
1454 | | #ifdef HAS_ARGBTOYROW_SSSE3 |
1455 | | void ARGBToYMatrixRow_SSSE3(const uint8_t* src_argb, |
1456 | | uint8_t* dst_y, |
1457 | | int width, |
1458 | 0 | const struct ArgbConstants* c) { |
1459 | 0 | asm volatile( |
1460 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
1461 | 0 | "psllw $15,%%xmm5 \n" |
1462 | 0 | "packsswb %%xmm5,%%xmm5 \n" |
1463 | 0 | "movdqa 0(%3),%%xmm4 \n" |
1464 | 0 | "movdqa 0x60(%3),%%xmm7 \n" |
1465 | 0 | "movdqa %%xmm4,%%xmm6 \n" |
1466 | 0 | "pmaddubsw %%xmm5,%%xmm6 \n" |
1467 | 0 | "phaddw %%xmm6,%%xmm6 \n" |
1468 | 0 | "psubw %%xmm6,%%xmm7 \n" LABELALIGN "" RGBTOY(xmm7) |
1469 | 0 | : "+r"(src_argb), // %0 |
1470 | 0 | "+r"(dst_y), // %1 |
1471 | 0 | "+r"(width) // %2 |
1472 | 0 | : "r"(c) // %3 |
1473 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1474 | 0 | "xmm7"); |
1475 | 0 | } |
1476 | | #endif |
1477 | | |
1478 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1479 | | void ARGBToYMatrixRow_AVX2(const uint8_t* src_argb, |
1480 | | uint8_t* dst_y, |
1481 | | int width, |
1482 | 0 | const struct ArgbConstants* c) { |
1483 | 0 | asm volatile( |
1484 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
1485 | 0 | "vpsllw $15,%%ymm5,%%ymm5 \n" |
1486 | 0 | "vpacksswb %%ymm5,%%ymm5,%%ymm5 \n" |
1487 | 0 | "vbroadcasti128 0(%3),%%ymm4 \n" |
1488 | 0 | "vbroadcasti128 0x60(%3),%%ymm7 \n" |
1489 | 0 | "vpmaddubsw %%ymm5,%%ymm4,%%ymm6 \n" |
1490 | 0 | "vphaddw %%ymm6,%%ymm6,%%ymm6 \n" |
1491 | 0 | "vpsubw %%ymm6,%%ymm7,%%ymm7 \n" |
1492 | 0 | "vmovdqa %4,%%ymm6 \n" LABELALIGN |
1493 | 0 | "" RGBTOY_AVX2(ymm7) "vzeroupper \n" |
1494 | 0 | : "+r"(src_argb), // %0 |
1495 | 0 | "+r"(dst_y), // %1 |
1496 | 0 | "+r"(width) // %2 |
1497 | 0 | : "r"(c), // %3 |
1498 | 0 | "m"(kPermdARGBToY_AVX) // %4 |
1499 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1500 | 0 | "xmm7"); |
1501 | 0 | } |
1502 | | #endif |
1503 | | |
1504 | | #if defined(HAS_ARGBTOYROW_AVX512BW) || \ |
1505 | | defined(HAS_ARGBTOUV444ROW_AVX512BW) || defined(HAS_ARGBTOUVROW_AVX512BW) |
1506 | | static const uint32_t kPermdARGBToY_AVX512BW[16] = {0, 4, 8, 12, 1, 5, 9, 13, |
1507 | | 2, 6, 10, 14, 3, 7, 11, 15}; |
1508 | | #endif |
1509 | | |
1510 | | #if defined(HAS_ARGBTOUVROW_AVX512BW) || defined(HAS_ARGBTOUVJROW_AVX512BW) |
1511 | | static const uint32_t kPermdARGBToUV_AVX512BW[16] = { |
1512 | | 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15}; |
1513 | | #endif |
1514 | | |
1515 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
1516 | | void ARGBToYMatrixRow_AVX512BW(const uint8_t* src_argb, |
1517 | | uint8_t* dst_y, |
1518 | | int width, |
1519 | 0 | const struct ArgbConstants* c) { |
1520 | 0 | asm volatile( |
1521 | 0 | "vpternlogd $0xff,%%zmm16,%%zmm16,%%zmm16 \n" |
1522 | 0 | "vpsllw $15,%%zmm16,%%zmm5 \n" |
1523 | 0 | "vpacksswb %%zmm5,%%zmm5,%%zmm5 \n" |
1524 | 0 | "vpsrlw $15,%%zmm16,%%zmm16 \n" // zmm16 = 1 |
1525 | 0 | "vbroadcasti64x4 0(%3),%%zmm4 \n" |
1526 | 0 | "vbroadcasti64x4 0x60(%3),%%zmm7 \n" |
1527 | 0 | "vpmaddubsw %%zmm5,%%zmm4,%%zmm6 \n" |
1528 | 0 | "vpmaddwd %%zmm16,%%zmm6,%%zmm6 \n" |
1529 | 0 | "vpackssdw %%zmm6,%%zmm6,%%zmm6 \n" |
1530 | 0 | "vpsubw %%zmm6,%%zmm7,%%zmm7 \n" |
1531 | 0 | "vmovups %4,%%zmm6 \n" LABELALIGN |
1532 | 0 | "1: \n" |
1533 | 0 | "vmovups (%0),%%zmm0 \n" |
1534 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
1535 | 0 | "vmovups 0x80(%0),%%zmm2 \n" |
1536 | 0 | "vmovups 0xc0(%0),%%zmm3 \n" |
1537 | 0 | "vpsubb %%zmm5,%%zmm0,%%zmm0 \n" |
1538 | 0 | "vpsubb %%zmm5,%%zmm1,%%zmm1 \n" |
1539 | 0 | "vpsubb %%zmm5,%%zmm2,%%zmm2 \n" |
1540 | 0 | "vpsubb %%zmm5,%%zmm3,%%zmm3 \n" |
1541 | 0 | "vpmaddubsw %%zmm0,%%zmm4,%%zmm0 \n" |
1542 | 0 | "vpmaddubsw %%zmm1,%%zmm4,%%zmm1 \n" |
1543 | 0 | "vpmaddubsw %%zmm2,%%zmm4,%%zmm2 \n" |
1544 | 0 | "vpmaddubsw %%zmm3,%%zmm4,%%zmm3 \n" |
1545 | 0 | "lea 0x100(%0),%0 \n" |
1546 | 0 | "vpmaddwd %%zmm16,%%zmm0,%%zmm0 \n" |
1547 | 0 | "vpmaddwd %%zmm16,%%zmm1,%%zmm1 \n" |
1548 | 0 | "vpackssdw %%zmm1,%%zmm0,%%zmm0 \n" |
1549 | 0 | "vpmaddwd %%zmm16,%%zmm2,%%zmm2 \n" |
1550 | 0 | "vpmaddwd %%zmm16,%%zmm3,%%zmm3 \n" |
1551 | 0 | "vpackssdw %%zmm3,%%zmm2,%%zmm2 \n" |
1552 | 0 | "vpaddw %%zmm7,%%zmm0,%%zmm0 \n" |
1553 | 0 | "vpaddw %%zmm7,%%zmm2,%%zmm2 \n" |
1554 | 0 | "vpsrlw $0x8,%%zmm0,%%zmm0 \n" |
1555 | 0 | "vpsrlw $0x8,%%zmm2,%%zmm2 \n" |
1556 | 0 | "vpackuswb %%zmm2,%%zmm0,%%zmm0 \n" |
1557 | 0 | "vpermd %%zmm0,%%zmm6,%%zmm0 \n" |
1558 | 0 | "vmovups %%zmm0,(%1) \n" |
1559 | 0 | "lea 0x40(%1),%1 \n" |
1560 | 0 | "sub $0x40,%2 \n" |
1561 | 0 | "jg 1b \n" |
1562 | 0 | "vzeroupper \n" |
1563 | 0 | : "+r"(src_argb), // %0 |
1564 | 0 | "+r"(dst_y), // %1 |
1565 | 0 | "+r"(width) // %2 |
1566 | 0 | : "r"(c), // %3 |
1567 | 0 | "m"(kPermdARGBToY_AVX512BW) // %4 |
1568 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1569 | 0 | "xmm7", "xmm16"); |
1570 | 0 | } |
1571 | | #endif |
1572 | | |
1573 | | #ifdef HAS_ARGBTOUV444ROW_SSSE3 |
1574 | | void ARGBToUV444MatrixRow_SSSE3(const uint8_t* src_argb, |
1575 | | uint8_t* dst_u, |
1576 | | uint8_t* dst_v, |
1577 | | int width, |
1578 | 0 | const struct ArgbConstants* c) { |
1579 | 0 | asm volatile( |
1580 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" // 0x8000 |
1581 | 0 | "psllw $15,%%xmm5 \n" |
1582 | 0 | "movdqa 0x20(%4),%%xmm3 \n" // kRGBToU |
1583 | 0 | "movdqa 0x40(%4),%%xmm4 \n" // kRGBToV |
1584 | 0 | "sub %1,%2 \n" |
1585 | |
|
1586 | 0 | LABELALIGN |
1587 | 0 | "1: \n" |
1588 | 0 | "movdqu (%0),%%xmm0 \n" |
1589 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
1590 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
1591 | 0 | "movdqu 0x30(%0),%%xmm6 \n" |
1592 | 0 | "pmaddubsw %%xmm3,%%xmm0 \n" |
1593 | 0 | "pmaddubsw %%xmm3,%%xmm1 \n" |
1594 | 0 | "pmaddubsw %%xmm3,%%xmm2 \n" |
1595 | 0 | "pmaddubsw %%xmm3,%%xmm6 \n" |
1596 | 0 | "phaddw %%xmm1,%%xmm0 \n" |
1597 | 0 | "phaddw %%xmm6,%%xmm2 \n" |
1598 | 0 | "movdqa %%xmm5,%%xmm1 \n" |
1599 | 0 | "movdqa %%xmm5,%%xmm6 \n" |
1600 | 0 | "psubw %%xmm0,%%xmm1 \n" |
1601 | 0 | "psubw %%xmm2,%%xmm6 \n" |
1602 | 0 | "psrlw $0x8,%%xmm1 \n" |
1603 | 0 | "psrlw $0x8,%%xmm6 \n" |
1604 | 0 | "packuswb %%xmm6,%%xmm1 \n" |
1605 | 0 | "movdqu %%xmm1,(%1) \n" |
1606 | |
|
1607 | 0 | "movdqu (%0),%%xmm0 \n" |
1608 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
1609 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
1610 | 0 | "movdqu 0x30(%0),%%xmm6 \n" |
1611 | 0 | "pmaddubsw %%xmm4,%%xmm0 \n" |
1612 | 0 | "pmaddubsw %%xmm4,%%xmm1 \n" |
1613 | 0 | "pmaddubsw %%xmm4,%%xmm2 \n" |
1614 | 0 | "pmaddubsw %%xmm4,%%xmm6 \n" |
1615 | 0 | "phaddw %%xmm1,%%xmm0 \n" |
1616 | 0 | "phaddw %%xmm6,%%xmm2 \n" |
1617 | 0 | "movdqa %%xmm5,%%xmm1 \n" |
1618 | 0 | "movdqa %%xmm5,%%xmm6 \n" |
1619 | 0 | "psubw %%xmm0,%%xmm1 \n" |
1620 | 0 | "psubw %%xmm2,%%xmm6 \n" |
1621 | 0 | "psrlw $0x8,%%xmm1 \n" |
1622 | 0 | "psrlw $0x8,%%xmm6 \n" |
1623 | 0 | "packuswb %%xmm6,%%xmm1 \n" |
1624 | 0 | "movdqu %%xmm1,0x00(%1,%2,1) \n" |
1625 | |
|
1626 | 0 | "lea 0x40(%0),%0 \n" |
1627 | 0 | "lea 0x10(%1),%1 \n" |
1628 | 0 | "subl $0x10,%3 \n" |
1629 | 0 | "jg 1b \n" |
1630 | 0 | : "+r"(src_argb), // %0 |
1631 | 0 | "+r"(dst_u), // %1 |
1632 | 0 | "+r"(dst_v), // %2 |
1633 | | #if defined(__i386__) |
1634 | | "+m"(width) // %3 |
1635 | | #else |
1636 | 0 | "+rm"(width) // %3 |
1637 | 0 | #endif |
1638 | 0 | : "r"(c) // %4 |
1639 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
1640 | 0 | } |
1641 | | #endif // HAS_ARGBTOUV444ROW_SSSE3 |
1642 | | |
1643 | | #ifdef HAS_ARGBTOUV444ROW_AVX2 |
1644 | | |
1645 | | void ARGBToUV444MatrixRow_AVX2(const uint8_t* src_argb, |
1646 | | uint8_t* dst_u, |
1647 | | uint8_t* dst_v, |
1648 | | int width, |
1649 | 0 | const struct ArgbConstants* c) { |
1650 | 0 | asm volatile( |
1651 | 0 | "vmovdqa 0x20(%4),%%ymm3 \n" // kRGBToU |
1652 | 0 | "vmovdqa 0x40(%4),%%ymm4 \n" // kRGBToV |
1653 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // 0x8000 |
1654 | 0 | "vpsllw $15,%%ymm5,%%ymm5 \n" |
1655 | 0 | "vmovdqa %5,%%ymm7 \n" |
1656 | 0 | "sub %1,%2 \n" |
1657 | |
|
1658 | 0 | LABELALIGN |
1659 | 0 | "1: \n" |
1660 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1661 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
1662 | 0 | "vmovdqu 0x40(%0),%%ymm2 \n" |
1663 | 0 | "vmovdqu 0x60(%0),%%ymm6 \n" |
1664 | 0 | "vpmaddubsw %%ymm3,%%ymm0,%%ymm0 \n" |
1665 | 0 | "vpmaddubsw %%ymm3,%%ymm1,%%ymm1 \n" |
1666 | 0 | "vpmaddubsw %%ymm3,%%ymm2,%%ymm2 \n" |
1667 | 0 | "vpmaddubsw %%ymm3,%%ymm6,%%ymm6 \n" |
1668 | 0 | "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" // mutates |
1669 | 0 | "vphaddw %%ymm6,%%ymm2,%%ymm2 \n" |
1670 | 0 | "vpsubw %%ymm0,%%ymm5,%%ymm0 \n" |
1671 | 0 | "vpsubw %%ymm2,%%ymm5,%%ymm2 \n" |
1672 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
1673 | 0 | "vpsrlw $0x8,%%ymm2,%%ymm2 \n" |
1674 | 0 | "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" // mutates |
1675 | 0 | "vpermd %%ymm0,%%ymm7,%%ymm0 \n" // unmutate. |
1676 | 0 | "vmovdqu %%ymm0,(%1) \n" |
1677 | |
|
1678 | 0 | "vmovdqu (%0),%%ymm0 \n" |
1679 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
1680 | 0 | "vmovdqu 0x40(%0),%%ymm2 \n" |
1681 | 0 | "vmovdqu 0x60(%0),%%ymm6 \n" |
1682 | 0 | "vpmaddubsw %%ymm4,%%ymm0,%%ymm0 \n" |
1683 | 0 | "vpmaddubsw %%ymm4,%%ymm1,%%ymm1 \n" |
1684 | 0 | "vpmaddubsw %%ymm4,%%ymm2,%%ymm2 \n" |
1685 | 0 | "vpmaddubsw %%ymm4,%%ymm6,%%ymm6 \n" |
1686 | 0 | "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" // mutates |
1687 | 0 | "vphaddw %%ymm6,%%ymm2,%%ymm2 \n" |
1688 | 0 | "vpsubw %%ymm0,%%ymm5,%%ymm0 \n" |
1689 | 0 | "vpsubw %%ymm2,%%ymm5,%%ymm2 \n" |
1690 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
1691 | 0 | "vpsrlw $0x8,%%ymm2,%%ymm2 \n" |
1692 | 0 | "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" // mutates |
1693 | 0 | "vpermd %%ymm0,%%ymm7,%%ymm0 \n" // unmutate. |
1694 | 0 | "vmovdqu %%ymm0,(%1,%2,1) \n" |
1695 | 0 | "lea 0x80(%0),%0 \n" |
1696 | 0 | "lea 0x20(%1),%1 \n" |
1697 | 0 | "subl $0x20,%3 \n" |
1698 | 0 | "jg 1b \n" |
1699 | 0 | "vzeroupper \n" |
1700 | 0 | : "+r"(src_argb), // %0 |
1701 | 0 | "+r"(dst_u), // %1 |
1702 | 0 | "+r"(dst_v), // %2 |
1703 | | #if defined(__i386__) |
1704 | | "+m"(width) // %3 |
1705 | | #else |
1706 | 0 | "+rm"(width) // %3 |
1707 | 0 | #endif |
1708 | 0 | : "r"(c), // %4 |
1709 | 0 | "m"(kPermdARGBToY_AVX) // %5 |
1710 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1711 | 0 | "xmm7"); |
1712 | 0 | } |
1713 | | #endif // HAS_ARGBTOUV444ROW_AVX2 |
1714 | | |
1715 | | #ifdef HAS_ARGBTOUV444ROW_AVX512BW |
1716 | | |
1717 | | void ARGBToUV444MatrixRow_AVX512BW(const uint8_t* src_argb, |
1718 | | uint8_t* dst_u, |
1719 | | uint8_t* dst_v, |
1720 | | int width, |
1721 | 0 | const struct ArgbConstants* c) { |
1722 | 0 | asm volatile( |
1723 | 0 | "vbroadcasti64x4 0x20(%4),%%zmm3 \n" // kRGBToU |
1724 | 0 | "vbroadcasti64x4 0x40(%4),%%zmm4 \n" // kRGBToV |
1725 | 0 | "vpternlogd $0xff,%%zmm16,%%zmm16,%%zmm16 \n" // -1 |
1726 | 0 | "vpsllw $15,%%zmm16,%%zmm5 \n" // 0x8000 |
1727 | 0 | "vmovups %5,%%zmm7 \n" |
1728 | 0 | "sub %1,%2 \n" |
1729 | |
|
1730 | 0 | LABELALIGN |
1731 | 0 | "1: \n" |
1732 | 0 | "vmovups (%0),%%zmm0 \n" |
1733 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
1734 | 0 | "vmovups 0x80(%0),%%zmm2 \n" |
1735 | 0 | "vmovups 0xc0(%0),%%zmm6 \n" |
1736 | 0 | "vpmaddubsw %%zmm3,%%zmm0,%%zmm0 \n" |
1737 | 0 | "vpmaddubsw %%zmm3,%%zmm1,%%zmm1 \n" |
1738 | 0 | "vpmaddubsw %%zmm3,%%zmm2,%%zmm2 \n" |
1739 | 0 | "vpmaddubsw %%zmm3,%%zmm6,%%zmm6 \n" |
1740 | 0 | "vpmaddwd %%zmm16,%%zmm0,%%zmm0 \n" |
1741 | 0 | "vpmaddwd %%zmm16,%%zmm1,%%zmm1 \n" |
1742 | 0 | "vpmaddwd %%zmm16,%%zmm2,%%zmm2 \n" |
1743 | 0 | "vpmaddwd %%zmm16,%%zmm6,%%zmm6 \n" |
1744 | 0 | "vpackssdw %%zmm1,%%zmm0,%%zmm0 \n" // mutates |
1745 | 0 | "vpackssdw %%zmm6,%%zmm2,%%zmm2 \n" |
1746 | 0 | "vpsubw %%zmm5,%%zmm0,%%zmm0 \n" |
1747 | 0 | "vpsubw %%zmm5,%%zmm2,%%zmm2 \n" |
1748 | 0 | "vpsrlw $0x8,%%zmm0,%%zmm0 \n" |
1749 | 0 | "vpsrlw $0x8,%%zmm2,%%zmm2 \n" |
1750 | 0 | "vpackuswb %%zmm2,%%zmm0,%%zmm0 \n" // mutates |
1751 | 0 | "vpermd %%zmm0,%%zmm7,%%zmm0 \n" // unmutate. |
1752 | 0 | "vmovups %%zmm0,(%1) \n" |
1753 | |
|
1754 | 0 | "vmovups (%0),%%zmm0 \n" |
1755 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
1756 | 0 | "vmovups 0x80(%0),%%zmm2 \n" |
1757 | 0 | "vmovups 0xc0(%0),%%zmm6 \n" |
1758 | 0 | "vpmaddubsw %%zmm4,%%zmm0,%%zmm0 \n" |
1759 | 0 | "vpmaddubsw %%zmm4,%%zmm1,%%zmm1 \n" |
1760 | 0 | "vpmaddubsw %%zmm4,%%zmm2,%%zmm2 \n" |
1761 | 0 | "vpmaddubsw %%zmm4,%%zmm6,%%zmm6 \n" |
1762 | 0 | "vpmaddwd %%zmm16,%%zmm0,%%zmm0 \n" |
1763 | 0 | "vpmaddwd %%zmm16,%%zmm1,%%zmm1 \n" |
1764 | 0 | "vpmaddwd %%zmm16,%%zmm2,%%zmm2 \n" |
1765 | 0 | "vpmaddwd %%zmm16,%%zmm6,%%zmm6 \n" |
1766 | 0 | "vpackssdw %%zmm1,%%zmm0,%%zmm0 \n" // mutates |
1767 | 0 | "vpackssdw %%zmm6,%%zmm2,%%zmm2 \n" |
1768 | 0 | "vpsubw %%zmm5,%%zmm0,%%zmm0 \n" |
1769 | 0 | "vpsubw %%zmm5,%%zmm2,%%zmm2 \n" |
1770 | 0 | "vpsrlw $0x8,%%zmm0,%%zmm0 \n" |
1771 | 0 | "vpsrlw $0x8,%%zmm2,%%zmm2 \n" |
1772 | 0 | "vpackuswb %%zmm2,%%zmm0,%%zmm0 \n" // mutates |
1773 | 0 | "vpermd %%zmm0,%%zmm7,%%zmm0 \n" // unmutate. |
1774 | 0 | "vmovups %%zmm0,(%1,%2,1) \n" |
1775 | 0 | "lea 0x100(%0),%0 \n" |
1776 | 0 | "lea 0x40(%1),%1 \n" |
1777 | 0 | "subl $0x40,%3 \n" |
1778 | 0 | "jg 1b \n" |
1779 | 0 | "vzeroupper \n" |
1780 | 0 | : "+r"(src_argb), // %0 |
1781 | 0 | "+r"(dst_u), // %1 |
1782 | 0 | "+r"(dst_v), // %2 |
1783 | | #if defined(__i386__) |
1784 | | "+m"(width) // %3 |
1785 | | #else |
1786 | 0 | "+rm"(width) // %3 |
1787 | 0 | #endif |
1788 | 0 | : "r"(c), // %4 |
1789 | 0 | "m"(kPermdARGBToY_AVX512BW) // %5 |
1790 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1791 | 0 | "xmm7", "xmm16"); |
1792 | 0 | } |
1793 | | #endif // HAS_ARGBTOUV444ROW_AVX512BW |
1794 | | |
1795 | | #ifdef HAS_ARGBTOUVROW_SSSE3 |
1796 | | |
1797 | | // ARGBARGB to AARRGGBB shuffle |
1798 | | static const lvec8 kShuffleAARRGGBB = { |
1799 | | 0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15, |
1800 | | 0, 4, 1, 5, 2, 6, 3, 7, 8, 12, 9, 13, 10, 14, 11, 15, |
1801 | | }; |
1802 | | |
1803 | | // 8x2 -> 4x1 ARGB pixels converted to 4 U and 4 V |
1804 | | // ARGBToUV does rounding average of 4 ARGB pixels |
1805 | | void ARGBToUVMatrixRow_SSSE3(const uint8_t* src_argb, |
1806 | | int src_stride_argb, |
1807 | | uint8_t* dst_u, |
1808 | | uint8_t* dst_v, |
1809 | | int width, |
1810 | 0 | const struct ArgbConstants* c) { |
1811 | 0 | asm volatile( |
1812 | 0 | "movdqa 0x20(%5),%%xmm4 \n" // RGBToU |
1813 | 0 | "movdqa 0x40(%5),%%xmm5 \n" // RGBToV |
1814 | 0 | "pcmpeqb %%xmm6,%%xmm6 \n" // 0x0101 |
1815 | 0 | "pabsb %%xmm6,%%xmm6 \n" |
1816 | 0 | "movdqa %6,%%xmm7 \n" // kShuffleAARRGGBB |
1817 | 0 | "sub %1,%2 \n" |
1818 | |
|
1819 | 0 | "1: \n" |
1820 | 0 | "movdqu (%0),%%xmm0 \n" // Read 8x2 ARGB Pixels |
1821 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
1822 | 0 | "movdqu 0x00(%0,%4,1),%%xmm2 \n" |
1823 | 0 | "movdqu 0x10(%0,%4,1),%%xmm3 \n" |
1824 | 0 | "pshufb %%xmm7,%%xmm0 \n" // aarrggbb |
1825 | 0 | "pshufb %%xmm7,%%xmm1 \n" |
1826 | 0 | "pshufb %%xmm7,%%xmm2 \n" |
1827 | 0 | "pshufb %%xmm7,%%xmm3 \n" |
1828 | 0 | "pmaddubsw %%xmm6,%%xmm0 \n" // 8x2 -> 4x2 |
1829 | 0 | "pmaddubsw %%xmm6,%%xmm1 \n" |
1830 | 0 | "pmaddubsw %%xmm6,%%xmm2 \n" |
1831 | 0 | "pmaddubsw %%xmm6,%%xmm3 \n" |
1832 | 0 | "paddw %%xmm2,%%xmm0 \n" // 4x2 -> 4x1 |
1833 | 0 | "paddw %%xmm3,%%xmm1 \n" |
1834 | 0 | "pxor %%xmm2,%%xmm2 \n" // 0 for vpavgw |
1835 | 0 | "psrlw $1,%%xmm0 \n" |
1836 | 0 | "psrlw $1,%%xmm1 \n" |
1837 | 0 | "pavgw %%xmm2,%%xmm0 \n" |
1838 | 0 | "pavgw %%xmm2,%%xmm1 \n" |
1839 | 0 | "packuswb %%xmm1,%%xmm0 \n" // mutates |
1840 | |
|
1841 | 0 | "movdqa %%xmm6,%%xmm2 \n" |
1842 | 0 | "psllw $15,%%xmm2 \n" // 0x8000 |
1843 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
1844 | 0 | "pmaddubsw %%xmm5,%%xmm1 \n" // 4 V |
1845 | 0 | "pmaddubsw %%xmm4,%%xmm0 \n" // 4 U |
1846 | 0 | "phaddw %%xmm1,%%xmm0 \n" // uuuuvvvv |
1847 | 0 | "psubw %%xmm0,%%xmm2 \n" |
1848 | 0 | "psrlw $0x8,%%xmm2 \n" |
1849 | 0 | "packuswb %%xmm2,%%xmm2 \n" |
1850 | 0 | "movd %%xmm2,(%1) \n" // Write 4 U's |
1851 | 0 | "pshufd $0x55,%%xmm2,%%xmm2 \n" // Copy V to low 4 bytes |
1852 | 0 | "movd %%xmm2,0x00(%1,%2,1) \n" // Write 4 V's |
1853 | |
|
1854 | 0 | "lea 0x20(%0),%0 \n" |
1855 | 0 | "lea 0x4(%1),%1 \n" |
1856 | 0 | "subl $0x8,%3 \n" |
1857 | 0 | "jg 1b \n" |
1858 | 0 | : "+r"(src_argb), // %0 |
1859 | 0 | "+r"(dst_u), // %1 |
1860 | 0 | "+r"(dst_v), // %2 |
1861 | | #if defined(__i386__) |
1862 | | "+m"(width) // %3 |
1863 | | #else |
1864 | 0 | "+rm"(width) // %3 |
1865 | 0 | #endif |
1866 | 0 | : "r"((ptrdiff_t)(src_stride_argb)), // %4 |
1867 | 0 | "r"(c), // %5 |
1868 | 0 | "m"(kShuffleAARRGGBB) // %6 |
1869 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1870 | 0 | "xmm7"); |
1871 | 0 | } |
1872 | | |
1873 | | #endif // HAS_ARGBTOUVROW_SSSE3 |
1874 | | |
1875 | | #ifdef HAS_ARGBTOUVROW_AVX2 |
1876 | | |
1877 | | // 16x2 -> 8x1 ARGB pixels converted to 8 U and 8 V |
1878 | | // ARGBToUV does rounding average of 4 ARGB pixels |
1879 | | void ARGBToUVMatrixRow_AVX2(const uint8_t* src_argb, |
1880 | | int src_stride_argb, |
1881 | | uint8_t* dst_u, |
1882 | | uint8_t* dst_v, |
1883 | | int width, |
1884 | 0 | const struct ArgbConstants* c) { |
1885 | 0 | asm volatile( |
1886 | 0 | "vbroadcasti128 0x20(%5),%%ymm4 \n" // RGBToU |
1887 | 0 | "vbroadcasti128 0x40(%5),%%ymm5 \n" // RGBToV |
1888 | 0 | "vpcmpeqb %%ymm6,%%ymm6,%%ymm6 \n" // 0x0101 |
1889 | 0 | "vpabsb %%ymm6,%%ymm6 \n" |
1890 | 0 | "vmovdqa %6,%%ymm7 \n" // kShuffleAARRGGBB |
1891 | 0 | "sub %1,%2 \n" |
1892 | |
|
1893 | 0 | "1: \n" |
1894 | 0 | "vmovdqu (%0),%%ymm0 \n" // Read 16x2 ARGB Pixels |
1895 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
1896 | 0 | "vmovdqu 0x00(%0,%4,1),%%ymm2 \n" |
1897 | 0 | "vmovdqu 0x20(%0,%4,1),%%ymm3 \n" |
1898 | 0 | "vpshufb %%ymm7,%%ymm0,%%ymm0 \n" // aarrggbb |
1899 | 0 | "vpshufb %%ymm7,%%ymm1,%%ymm1 \n" |
1900 | 0 | "vpshufb %%ymm7,%%ymm2,%%ymm2 \n" |
1901 | 0 | "vpshufb %%ymm7,%%ymm3,%%ymm3 \n" |
1902 | 0 | "vpmaddubsw %%ymm6,%%ymm0,%%ymm0 \n" // 16x2 -> 8x2 |
1903 | 0 | "vpmaddubsw %%ymm6,%%ymm1,%%ymm1 \n" |
1904 | 0 | "vpmaddubsw %%ymm6,%%ymm2,%%ymm2 \n" |
1905 | 0 | "vpmaddubsw %%ymm6,%%ymm3,%%ymm3 \n" |
1906 | 0 | "vpaddw %%ymm0,%%ymm2,%%ymm0 \n" // 8x2 -> 8x1 |
1907 | 0 | "vpaddw %%ymm1,%%ymm3,%%ymm1 \n" |
1908 | 0 | "vpxor %%ymm2,%%ymm2,%%ymm2 \n" // 0 for vpavgw |
1909 | 0 | "vpsrlw $1,%%ymm0,%%ymm0 \n" |
1910 | 0 | "vpsrlw $1,%%ymm1,%%ymm1 \n" |
1911 | 0 | "vpavgw %%ymm2,%%ymm0,%%ymm0 \n" |
1912 | 0 | "vpavgw %%ymm2,%%ymm1,%%ymm1 \n" |
1913 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" // mutates |
1914 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" // 8 ARGB Pixels |
1915 | |
|
1916 | 0 | "vpsllw $15,%%ymm6,%%ymm2 \n" // 0x8000 |
1917 | 0 | "vpmaddubsw %%ymm5,%%ymm0,%%ymm1 \n" // 8 V |
1918 | 0 | "vpmaddubsw %%ymm4,%%ymm0,%%ymm0 \n" // 8 U |
1919 | 0 | "vphaddw %%ymm1,%%ymm0,%%ymm0 \n" // uuuuvvvv uuuuvvvv |
1920 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" // uuuuuuuu vvvvvvvv |
1921 | 0 | "vpsubw %%ymm0,%%ymm2,%%ymm0 \n" |
1922 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
1923 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" // mutates 8U8u- 8V8v |
1924 | 0 | "vmovq %%xmm0,(%1) \n" // Write 8 U's |
1925 | 0 | "vextractf128 $0x1,%%ymm0,%%xmm0 \n" // Copy V to low 8 bytes |
1926 | 0 | "vmovq %%xmm0,0x00(%1,%2,1) \n" // Write 8 V's |
1927 | |
|
1928 | 0 | "lea 0x40(%0),%0 \n" |
1929 | 0 | "lea 0x8(%1),%1 \n" |
1930 | 0 | "subl $0x10,%3 \n" |
1931 | 0 | "jg 1b \n" |
1932 | 0 | "vzeroupper \n" |
1933 | 0 | : "+r"(src_argb), // %0 |
1934 | 0 | "+r"(dst_u), // %1 |
1935 | 0 | "+r"(dst_v), // %2 |
1936 | | #if defined(__i386__) |
1937 | | "+m"(width) // %3 |
1938 | | #else |
1939 | 0 | "+rm"(width) // %3 |
1940 | 0 | #endif |
1941 | 0 | : "r"((ptrdiff_t)(src_stride_argb)), // %4 |
1942 | 0 | "r"(c), // %5 |
1943 | 0 | "m"(kShuffleAARRGGBB) // %6 |
1944 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
1945 | 0 | "xmm7"); |
1946 | 0 | } |
1947 | | #endif // HAS_ARGBTOUVROW_AVX2 |
1948 | | |
1949 | | // RGB to BT601 coefficients |
1950 | | // UB 0.875 coefficient = 112 |
1951 | | // UG -0.5781 coefficient = -74 |
1952 | | // UR -0.2969 coefficient = -38 |
1953 | | // VB -0.1406 coefficient = -18 |
1954 | | // VG -0.7344 coefficient = -94 |
1955 | | // VR 0.875 coefficient = 112 |
1956 | | |
1957 | | #ifdef HAS_ARGBTOUV444ROW_SSSE3 |
1958 | | void ARGBToUV444Row_SSSE3(const uint8_t* src_argb, |
1959 | | uint8_t* dst_u, |
1960 | | uint8_t* dst_v, |
1961 | 0 | int width) { |
1962 | 0 | ARGBToUV444MatrixRow_SSSE3(src_argb, dst_u, dst_v, width, |
1963 | 0 | &kArgbI601Constants); |
1964 | 0 | } |
1965 | | #endif // HAS_ARGBTOUV444ROW_SSSE3 |
1966 | | |
1967 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1968 | 0 | void RGBAToYRow_AVX2(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
1969 | 0 | ARGBToYMatrixRow_AVX2(src_rgba, dst_y, width, &kRgbaI601Constants); |
1970 | 0 | } |
1971 | | #endif |
1972 | | |
1973 | | #ifdef HAS_ARGBTOYROW_AVX2 |
1974 | 0 | void BGRAToYRow_AVX2(const uint8_t* src_bgra, uint8_t* dst_y, int width) { |
1975 | 0 | ARGBToYMatrixRow_AVX2(src_bgra, dst_y, width, &kBgraI601Constants); |
1976 | 0 | } |
1977 | | #endif |
1978 | | |
1979 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
1980 | 0 | void ARGBToYRow_AVX512BW(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1981 | 0 | ARGBToYMatrixRow_AVX512BW(src_argb, dst_y, width, &kArgbI601Constants); |
1982 | 0 | } |
1983 | | #endif |
1984 | | |
1985 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
1986 | 0 | void ARGBToYJRow_AVX512BW(const uint8_t* src_argb, uint8_t* dst_y, int width) { |
1987 | 0 | ARGBToYMatrixRow_AVX512BW(src_argb, dst_y, width, &kArgbJPEGConstants); |
1988 | 0 | } |
1989 | | #endif |
1990 | | |
1991 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
1992 | 0 | void ABGRToYRow_AVX512BW(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
1993 | 0 | ARGBToYMatrixRow_AVX512BW(src_abgr, dst_y, width, &kAbgrI601Constants); |
1994 | 0 | } |
1995 | | #endif |
1996 | | |
1997 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
1998 | 0 | void ABGRToYJRow_AVX512BW(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
1999 | 0 | ARGBToYMatrixRow_AVX512BW(src_abgr, dst_y, width, &kAbgrJPEGConstants); |
2000 | 0 | } |
2001 | | #endif |
2002 | | |
2003 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
2004 | 0 | void RGBAToYRow_AVX512BW(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
2005 | 0 | ARGBToYMatrixRow_AVX512BW(src_rgba, dst_y, width, &kRgbaI601Constants); |
2006 | 0 | } |
2007 | | #endif |
2008 | | |
2009 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
2010 | 0 | void RGBAToYJRow_AVX512BW(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
2011 | 0 | ARGBToYMatrixRow_AVX512BW(src_rgba, dst_y, width, &kRgbaJPEGConstants); |
2012 | 0 | } |
2013 | | #endif |
2014 | | |
2015 | | #ifdef HAS_ARGBTOYROW_AVX512BW |
2016 | 0 | void BGRAToYRow_AVX512BW(const uint8_t* src_bgra, uint8_t* dst_y, int width) { |
2017 | 0 | ARGBToYMatrixRow_AVX512BW(src_bgra, dst_y, width, &kBgraI601Constants); |
2018 | 0 | } |
2019 | | #endif |
2020 | | |
2021 | | #ifdef HAS_ARGBTOUV444ROW_AVX2 |
2022 | | void ARGBToUV444Row_AVX2(const uint8_t* src_argb, |
2023 | | uint8_t* dst_u, |
2024 | | uint8_t* dst_v, |
2025 | 0 | int width) { |
2026 | 0 | ARGBToUV444MatrixRow_AVX2(src_argb, dst_u, dst_v, width, &kArgbI601Constants); |
2027 | 0 | } |
2028 | | #endif // HAS_ARGBTOUV444ROW_AVX2 |
2029 | | |
2030 | | #ifdef HAS_ARGBTOUV444ROW_AVX512BW |
2031 | | void ARGBToUV444Row_AVX512BW(const uint8_t* src_argb, |
2032 | | uint8_t* dst_u, |
2033 | | uint8_t* dst_v, |
2034 | 0 | int width) { |
2035 | 0 | ARGBToUV444MatrixRow_AVX512BW(src_argb, dst_u, dst_v, width, |
2036 | 0 | &kArgbI601Constants); |
2037 | 0 | } |
2038 | | #endif // HAS_ARGBTOUV444ROW_AVX512BW |
2039 | | |
2040 | | #ifdef HAS_ARGBTOUVROW_SSSE3 |
2041 | | void ARGBToUVRow_SSSE3(const uint8_t* src_argb, |
2042 | | int src_stride_argb, |
2043 | | uint8_t* dst_u, |
2044 | | uint8_t* dst_v, |
2045 | 0 | int width) { |
2046 | 0 | ARGBToUVMatrixRow_SSSE3(src_argb, src_stride_argb, dst_u, dst_v, width, |
2047 | 0 | &kArgbI601Constants); |
2048 | 0 | } |
2049 | | |
2050 | | void ABGRToUVRow_SSSE3(const uint8_t* src_abgr, |
2051 | | int src_stride_abgr, |
2052 | | uint8_t* dst_u, |
2053 | | uint8_t* dst_v, |
2054 | 0 | int width) { |
2055 | 0 | ARGBToUVMatrixRow_SSSE3(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2056 | 0 | &kAbgrI601Constants); |
2057 | 0 | } |
2058 | | |
2059 | | void BGRAToUVRow_SSSE3(const uint8_t* src_bgra, |
2060 | | int src_stride_bgra, |
2061 | | uint8_t* dst_u, |
2062 | | uint8_t* dst_v, |
2063 | 0 | int width) { |
2064 | 0 | ARGBToUVMatrixRow_SSSE3(src_bgra, src_stride_bgra, dst_u, dst_v, width, |
2065 | 0 | &kBgraI601Constants); |
2066 | 0 | } |
2067 | | |
2068 | | void RGBAToUVRow_SSSE3(const uint8_t* src_rgba, |
2069 | | int src_stride_rgba, |
2070 | | uint8_t* dst_u, |
2071 | | uint8_t* dst_v, |
2072 | 0 | int width) { |
2073 | 0 | ARGBToUVMatrixRow_SSSE3(src_rgba, src_stride_rgba, dst_u, dst_v, width, |
2074 | 0 | &kRgbaI601Constants); |
2075 | 0 | } |
2076 | | #endif // HAS_ARGBTOUVROW_SSSE3 |
2077 | | |
2078 | | #ifdef HAS_ARGBTOUVROW_AVX2 |
2079 | | void ARGBToUVRow_AVX2(const uint8_t* src_argb, |
2080 | | int src_stride_argb, |
2081 | | uint8_t* dst_u, |
2082 | | uint8_t* dst_v, |
2083 | 0 | int width) { |
2084 | 0 | ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width, |
2085 | 0 | &kArgbI601Constants); |
2086 | 0 | } |
2087 | | |
2088 | | void ABGRToUVRow_AVX2(const uint8_t* src_abgr, |
2089 | | int src_stride_abgr, |
2090 | | uint8_t* dst_u, |
2091 | | uint8_t* dst_v, |
2092 | 0 | int width) { |
2093 | 0 | ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2094 | 0 | &kAbgrI601Constants); |
2095 | 0 | } |
2096 | | #endif // HAS_ARGBTOUVROW_AVX2 |
2097 | | |
2098 | | #ifdef HAS_ARGBTOUVJ444ROW_SSSE3 |
2099 | | void ARGBToUVJ444Row_SSSE3(const uint8_t* src_argb, |
2100 | | uint8_t* dst_u, |
2101 | | uint8_t* dst_v, |
2102 | 0 | int width) { |
2103 | 0 | ARGBToUV444MatrixRow_SSSE3(src_argb, dst_u, dst_v, width, |
2104 | 0 | &kArgbJPEGConstants); |
2105 | 0 | } |
2106 | | #endif // HAS_ARGBTOUVJ444ROW_SSSE3 |
2107 | | |
2108 | | #ifdef HAS_ARGBTOUVJ444ROW_AVX2 |
2109 | | void ARGBToUVJ444Row_AVX2(const uint8_t* src_argb, |
2110 | | uint8_t* dst_u, |
2111 | | uint8_t* dst_v, |
2112 | 0 | int width) { |
2113 | 0 | ARGBToUV444MatrixRow_AVX2(src_argb, dst_u, dst_v, width, &kArgbJPEGConstants); |
2114 | 0 | } |
2115 | | #endif // HAS_ARGBTOUVJ444ROW_AVX2 |
2116 | | |
2117 | | #ifdef HAS_ARGBTOUVJ444ROW_AVX512BW |
2118 | | void ARGBToUVJ444Row_AVX512BW(const uint8_t* src_argb, |
2119 | | uint8_t* dst_u, |
2120 | | uint8_t* dst_v, |
2121 | 0 | int width) { |
2122 | 0 | ARGBToUV444MatrixRow_AVX512BW(src_argb, dst_u, dst_v, width, |
2123 | 0 | &kArgbJPEGConstants); |
2124 | 0 | } |
2125 | | #endif // HAS_ARGBTOUVJ444ROW_AVX512BW |
2126 | | |
2127 | | #ifdef HAS_ARGBTOUVJROW_SSSE3 |
2128 | | void ARGBToUVJRow_SSSE3(const uint8_t* src_argb, |
2129 | | int src_stride_argb, |
2130 | | uint8_t* dst_u, |
2131 | | uint8_t* dst_v, |
2132 | 0 | int width) { |
2133 | 0 | ARGBToUVMatrixRow_SSSE3(src_argb, src_stride_argb, dst_u, dst_v, width, |
2134 | 0 | &kArgbJPEGConstants); |
2135 | 0 | } |
2136 | | #endif // HAS_ARGBTOUVJROW_SSSE3 |
2137 | | |
2138 | | #ifdef HAS_ABGRTOUVJROW_SSSE3 |
2139 | | void ABGRToUVJRow_SSSE3(const uint8_t* src_abgr, |
2140 | | int src_stride_abgr, |
2141 | | uint8_t* dst_u, |
2142 | | uint8_t* dst_v, |
2143 | 0 | int width) { |
2144 | 0 | ARGBToUVMatrixRow_SSSE3(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2145 | 0 | &kAbgrJPEGConstants); |
2146 | 0 | } |
2147 | | #endif // HAS_ABGRTOUVJROW_SSSE3 |
2148 | | |
2149 | | #ifdef HAS_ARGBTOUVJROW_AVX2 |
2150 | | void ARGBToUVJRow_AVX2(const uint8_t* src_argb, |
2151 | | int src_stride_argb, |
2152 | | uint8_t* dst_u, |
2153 | | uint8_t* dst_v, |
2154 | 0 | int width) { |
2155 | 0 | ARGBToUVMatrixRow_AVX2(src_argb, src_stride_argb, dst_u, dst_v, width, |
2156 | 0 | &kArgbJPEGConstants); |
2157 | 0 | } |
2158 | | #endif // HAS_ARGBTOUVJROW_AVX2 |
2159 | | |
2160 | | #ifdef HAS_ABGRTOUVJROW_AVX2 |
2161 | | void ABGRToUVJRow_AVX2(const uint8_t* src_abgr, |
2162 | | int src_stride_abgr, |
2163 | | uint8_t* dst_u, |
2164 | | uint8_t* dst_v, |
2165 | 0 | int width) { |
2166 | 0 | ARGBToUVMatrixRow_AVX2(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2167 | 0 | &kAbgrJPEGConstants); |
2168 | 0 | } |
2169 | | #endif // HAS_ABGRTOUVJROW_AVX2 |
2170 | | |
2171 | | #ifdef HAS_ARGBTOUVROW_AVX512BW |
2172 | | |
2173 | | // 32x2 -> 16x1 ARGB pixels converted to 16 U and 16 V |
2174 | | // ARGBToUV does rounding average of 4 ARGB pixels |
2175 | | |
2176 | | void ARGBToUVMatrixRow_AVX512BW(const uint8_t* src_argb, |
2177 | | int src_stride_argb, |
2178 | | uint8_t* dst_u, |
2179 | | uint8_t* dst_v, |
2180 | | int width, |
2181 | 0 | const struct ArgbConstants* c) { |
2182 | 0 | asm volatile( |
2183 | 0 | "vbroadcasti64x4 0x20(%5),%%zmm4 \n" // RGBToU |
2184 | 0 | "vbroadcasti64x4 0x40(%5),%%zmm5 \n" // RGBToV |
2185 | 0 | "vpternlogd $0xff,%%zmm16,%%zmm16,%%zmm16 \n" |
2186 | 0 | "vpabsb %%zmm16,%%zmm6 \n" // 0x0101 |
2187 | 0 | "vpsllw $15,%%zmm16,%%zmm17 \n" // 0x8000 |
2188 | 0 | "vbroadcasti64x4 %6,%%zmm7 \n" // kShuffleAARRGGBB |
2189 | 0 | "vmovups %7,%%zmm18 \n" // kPermdARGBToY_AVX512BW |
2190 | 0 | "vmovups %8,%%zmm19 \n" // kPermdARGBToUV_AVX512BW |
2191 | 0 | "sub %1,%2 \n" |
2192 | |
|
2193 | 0 | LABELALIGN |
2194 | 0 | "1: \n" |
2195 | 0 | "vmovups (%0),%%zmm0 \n" // Read 32x2 ARGB Pixels |
2196 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
2197 | 0 | "vmovups 0x00(%0,%4,1),%%zmm2 \n" |
2198 | 0 | "vmovups 0x40(%0,%4,1),%%zmm3 \n" |
2199 | 0 | "vpshufb %%zmm7,%%zmm0,%%zmm0 \n" // aarrggbb |
2200 | 0 | "vpshufb %%zmm7,%%zmm1,%%zmm1 \n" |
2201 | 0 | "vpshufb %%zmm7,%%zmm2,%%zmm2 \n" |
2202 | 0 | "vpshufb %%zmm7,%%zmm3,%%zmm3 \n" |
2203 | 0 | "vpmaddubsw %%zmm6,%%zmm0,%%zmm0 \n" // 32x2 -> 16x2 |
2204 | 0 | "vpmaddubsw %%zmm6,%%zmm1,%%zmm1 \n" |
2205 | 0 | "vpmaddubsw %%zmm6,%%zmm2,%%zmm2 \n" |
2206 | 0 | "vpmaddubsw %%zmm6,%%zmm3,%%zmm3 \n" |
2207 | 0 | "vpaddw %%zmm0,%%zmm2,%%zmm0 \n" // 16x2 -> 16x1 |
2208 | 0 | "vpaddw %%zmm1,%%zmm3,%%zmm1 \n" |
2209 | 0 | "vpxorq %%zmm2,%%zmm2,%%zmm2 \n" // 0 for vpavgw |
2210 | 0 | "vpsrlw $1,%%zmm0,%%zmm0 \n" |
2211 | 0 | "vpsrlw $1,%%zmm1,%%zmm1 \n" |
2212 | 0 | "vpavgw %%zmm2,%%zmm0,%%zmm0 \n" |
2213 | 0 | "vpavgw %%zmm2,%%zmm1,%%zmm1 \n" |
2214 | 0 | "vpackuswb %%zmm1,%%zmm0,%%zmm0 \n" // mutates |
2215 | 0 | "vpermd %%zmm0,%%zmm19,%%zmm0 \n" // unscramble pixels |
2216 | |
|
2217 | 0 | "vpmaddubsw %%zmm4,%%zmm0,%%zmm1 \n" // 16 U |
2218 | 0 | "vpmaddubsw %%zmm5,%%zmm0,%%zmm0 \n" // 16 V |
2219 | 0 | "vpmaddwd %%zmm16,%%zmm1,%%zmm1 \n" |
2220 | 0 | "vpmaddwd %%zmm16,%%zmm0,%%zmm0 \n" |
2221 | 0 | "vpackssdw %%zmm0,%%zmm1,%%zmm0 \n" // mutates (U in lower, V |
2222 | | // in upper) |
2223 | 0 | "vpaddw %%zmm17,%%zmm0,%%zmm0 \n" |
2224 | 0 | "vpsrlw $0x8,%%zmm0,%%zmm0 \n" |
2225 | 0 | "vpackuswb %%zmm0,%%zmm0,%%zmm0 \n" // mutates |
2226 | 0 | "vpermd %%zmm0,%%zmm18,%%zmm0 \n" // unmutate |
2227 | |
|
2228 | 0 | "vmovdqu %%xmm0,(%1) \n" // Write 16 U's |
2229 | 0 | "vextracti32x4 $0x1,%%zmm0,%%xmm0 \n" |
2230 | 0 | "vmovdqu %%xmm0,0x00(%1,%2,1) \n" // Write 16 V's |
2231 | |
|
2232 | 0 | "lea 0x80(%0),%0 \n" |
2233 | 0 | "lea 0x10(%1),%1 \n" |
2234 | 0 | "subl $0x20,%3 \n" |
2235 | 0 | "jg 1b \n" |
2236 | 0 | "vzeroupper \n" |
2237 | 0 | : "+r"(src_argb), // %0 |
2238 | 0 | "+r"(dst_u), // %1 |
2239 | 0 | "+r"(dst_v), // %2 |
2240 | | #if defined(__i386__) |
2241 | | "+m"(width) // %3 |
2242 | | #else |
2243 | 0 | "+rm"(width) // %3 |
2244 | 0 | #endif |
2245 | 0 | : "r"((ptrdiff_t)(src_stride_argb)), // %4 |
2246 | 0 | "r"(c), // %5 |
2247 | 0 | "m"(kShuffleAARRGGBB), // %6 |
2248 | 0 | "m"(kPermdARGBToY_AVX512BW), // %7 |
2249 | 0 | "m"(kPermdARGBToUV_AVX512BW) // %8 |
2250 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
2251 | 0 | "xmm7", "xmm16", "xmm17", "xmm18", "xmm19"); |
2252 | 0 | } |
2253 | | |
2254 | | void ARGBToUVRow_AVX512BW(const uint8_t* src_argb, |
2255 | | int src_stride_argb, |
2256 | | uint8_t* dst_u, |
2257 | | uint8_t* dst_v, |
2258 | 0 | int width) { |
2259 | 0 | ARGBToUVMatrixRow_AVX512BW(src_argb, src_stride_argb, dst_u, dst_v, width, |
2260 | 0 | &kArgbI601Constants); |
2261 | 0 | } |
2262 | | |
2263 | | void ABGRToUVRow_AVX512BW(const uint8_t* src_abgr, |
2264 | | int src_stride_abgr, |
2265 | | uint8_t* dst_u, |
2266 | | uint8_t* dst_v, |
2267 | 0 | int width) { |
2268 | 0 | ARGBToUVMatrixRow_AVX512BW(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2269 | 0 | &kAbgrI601Constants); |
2270 | 0 | } |
2271 | | |
2272 | | #ifdef HAS_ARGBTOUVJROW_AVX512BW |
2273 | | void ARGBToUVJRow_AVX512BW(const uint8_t* src_argb, |
2274 | | int src_stride_argb, |
2275 | | uint8_t* dst_u, |
2276 | | uint8_t* dst_v, |
2277 | 0 | int width) { |
2278 | 0 | ARGBToUVMatrixRow_AVX512BW(src_argb, src_stride_argb, dst_u, dst_v, width, |
2279 | 0 | &kArgbJPEGConstants); |
2280 | 0 | } |
2281 | | #endif // HAS_ARGBTOUVJROW_AVX512BW |
2282 | | |
2283 | | #ifdef HAS_ABGRTOUVJROW_AVX512BW |
2284 | | void ABGRToUVJRow_AVX512BW(const uint8_t* src_abgr, |
2285 | | int src_stride_abgr, |
2286 | | uint8_t* dst_u, |
2287 | | uint8_t* dst_v, |
2288 | | int width) { |
2289 | | ARGBToUVMatrixRow_AVX512BW(src_abgr, src_stride_abgr, dst_u, dst_v, width, |
2290 | | &kAbgrJPEGConstants); |
2291 | | } |
2292 | | #endif // HAS_ABGRTOUVJROW_AVX512BW |
2293 | | #endif // HAS_ARGBTOUVROW_AVX512BW |
2294 | | |
2295 | 0 | void BGRAToYRow_SSSE3(const uint8_t* src_bgra, uint8_t* dst_y, int width) { |
2296 | 0 | ARGBToYMatrixRow_SSSE3(src_bgra, dst_y, width, &kBgraI601Constants); |
2297 | 0 | } |
2298 | | |
2299 | 0 | void ABGRToYRow_SSSE3(const uint8_t* src_abgr, uint8_t* dst_y, int width) { |
2300 | 0 | ARGBToYMatrixRow_SSSE3(src_abgr, dst_y, width, &kAbgrI601Constants); |
2301 | 0 | } |
2302 | | |
2303 | 0 | void RGBAToYRow_SSSE3(const uint8_t* src_rgba, uint8_t* dst_y, int width) { |
2304 | 0 | ARGBToYMatrixRow_SSSE3(src_rgba, dst_y, width, &kRgbaI601Constants); |
2305 | 0 | } |
2306 | | |
2307 | | #if defined(HAS_I422TOARGBROW_SSSE3) || defined(HAS_I422TOARGBROW_AVX2) |
2308 | | |
2309 | | // Read 8 UV from 444 |
2310 | | #define READYUV444 \ |
2311 | | "movq (%[u_buf]),%%xmm3 \n" \ |
2312 | | "movq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2313 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
2314 | | "punpcklbw %%xmm1,%%xmm3 \n" \ |
2315 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2316 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2317 | | "lea 0x8(%[y_buf]),%[y_buf] \n" |
2318 | | |
2319 | | // Read 4 UV from 422, upsample to 8 UV |
2320 | | #define READYUV422 \ |
2321 | | "movd (%[u_buf]),%%xmm3 \n" \ |
2322 | | "movd 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2323 | | "lea 0x4(%[u_buf]),%[u_buf] \n" \ |
2324 | | "punpcklbw %%xmm1,%%xmm3 \n" \ |
2325 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2326 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2327 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2328 | | "lea 0x8(%[y_buf]),%[y_buf] \n" |
2329 | | |
2330 | | // Read 4 UV from 422 10 bit, upsample to 8 UV |
2331 | | #define READYUV210 \ |
2332 | | "movq (%[u_buf]),%%xmm3 \n" \ |
2333 | | "movq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2334 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
2335 | | "punpcklwd %%xmm1,%%xmm3 \n" \ |
2336 | | "psraw $2,%%xmm3 \n" \ |
2337 | | "packuswb %%xmm3,%%xmm3 \n" \ |
2338 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2339 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2340 | | "movdqa %%xmm4,%%xmm2 \n" \ |
2341 | | "psllw $6,%%xmm4 \n" \ |
2342 | | "psrlw $4,%%xmm2 \n" \ |
2343 | | "paddw %%xmm2,%%xmm4 \n" \ |
2344 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
2345 | | |
2346 | | #define READYUVA210 \ |
2347 | | "movq (%[u_buf]),%%xmm3 \n" \ |
2348 | | "movq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2349 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
2350 | | "punpcklwd %%xmm1,%%xmm3 \n" \ |
2351 | | "psraw $2,%%xmm3 \n" \ |
2352 | | "packuswb %%xmm3,%%xmm3 \n" \ |
2353 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2354 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2355 | | "movdqa %%xmm4,%%xmm2 \n" \ |
2356 | | "psllw $6,%%xmm4 \n" \ |
2357 | | "psrlw $4,%%xmm2 \n" \ |
2358 | | "paddw %%xmm2,%%xmm4 \n" \ |
2359 | | "lea 0x10(%[y_buf]),%[y_buf] \n" \ |
2360 | | "movdqu (%[a_buf]),%%xmm5 \n" \ |
2361 | | "psraw $2,%%xmm5 \n" \ |
2362 | | "packuswb %%xmm5,%%xmm5 \n" \ |
2363 | | "lea 0x10(%[a_buf]),%[a_buf] \n" |
2364 | | |
2365 | | // Read 8 UV from 444 10 bit |
2366 | | #define READYUV410 \ |
2367 | | "movdqu (%[u_buf]),%%xmm3 \n" \ |
2368 | | "movdqu 0x00(%[u_buf],%[v_buf],1),%%xmm2 \n" \ |
2369 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
2370 | | "psraw $2,%%xmm3 \n" \ |
2371 | | "psraw $2,%%xmm2 \n" \ |
2372 | | "movdqa %%xmm3,%%xmm1 \n" \ |
2373 | | "punpcklwd %%xmm2,%%xmm3 \n" \ |
2374 | | "punpckhwd %%xmm2,%%xmm1 \n" \ |
2375 | | "packuswb %%xmm1,%%xmm3 \n" \ |
2376 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2377 | | "movdqa %%xmm4,%%xmm2 \n" \ |
2378 | | "psllw $6,%%xmm4 \n" \ |
2379 | | "psrlw $4,%%xmm2 \n" \ |
2380 | | "paddw %%xmm2,%%xmm4 \n" \ |
2381 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
2382 | | |
2383 | | // Read 8 UV from 444 10 bit. With 8 Alpha. |
2384 | | #define READYUVA410 \ |
2385 | | "movdqu (%[u_buf]),%%xmm3 \n" \ |
2386 | | "movdqu 0x00(%[u_buf],%[v_buf],1),%%xmm2 \n" \ |
2387 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
2388 | | "psraw $2,%%xmm3 \n" \ |
2389 | | "psraw $2,%%xmm2 \n" \ |
2390 | | "movdqa %%xmm3,%%xmm1 \n" \ |
2391 | | "punpcklwd %%xmm2,%%xmm3 \n" \ |
2392 | | "punpckhwd %%xmm2,%%xmm1 \n" \ |
2393 | | "packuswb %%xmm1,%%xmm3 \n" \ |
2394 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2395 | | "movdqa %%xmm4,%%xmm2 \n" \ |
2396 | | "psllw $6,%%xmm4 \n" \ |
2397 | | "psrlw $4,%%xmm2 \n" \ |
2398 | | "paddw %%xmm2,%%xmm4 \n" \ |
2399 | | "lea 0x10(%[y_buf]),%[y_buf] \n" \ |
2400 | | "movdqu (%[a_buf]),%%xmm5 \n" \ |
2401 | | "psraw $2,%%xmm5 \n" \ |
2402 | | "packuswb %%xmm5,%%xmm5 \n" \ |
2403 | | "lea 0x10(%[a_buf]),%[a_buf] \n" |
2404 | | |
2405 | | // Read 4 UV from 422 12 bit, upsample to 8 UV |
2406 | | #define READYUV212 \ |
2407 | | "movq (%[u_buf]),%%xmm3 \n" \ |
2408 | | "movq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2409 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
2410 | | "punpcklwd %%xmm1,%%xmm3 \n" \ |
2411 | | "psraw $0x4,%%xmm3 \n" \ |
2412 | | "packuswb %%xmm3,%%xmm3 \n" \ |
2413 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2414 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2415 | | "movdqa %%xmm4,%%xmm2 \n" \ |
2416 | | "psllw $4,%%xmm4 \n" \ |
2417 | | "psrlw $8,%%xmm2 \n" \ |
2418 | | "paddw %%xmm2,%%xmm4 \n" \ |
2419 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
2420 | | |
2421 | | // Read 4 UV from 422, upsample to 8 UV. With 8 Alpha. |
2422 | | #define READYUVA422 \ |
2423 | | "movd (%[u_buf]),%%xmm3 \n" \ |
2424 | | "movd 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2425 | | "lea 0x4(%[u_buf]),%[u_buf] \n" \ |
2426 | | "punpcklbw %%xmm1,%%xmm3 \n" \ |
2427 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2428 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2429 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2430 | | "lea 0x8(%[y_buf]),%[y_buf] \n" \ |
2431 | | "movq (%[a_buf]),%%xmm5 \n" \ |
2432 | | "lea 0x8(%[a_buf]),%[a_buf] \n" |
2433 | | |
2434 | | // Read 8 UV from 444. With 8 Alpha. |
2435 | | #define READYUVA444 \ |
2436 | | "movq (%[u_buf]),%%xmm3 \n" \ |
2437 | | "movq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
2438 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
2439 | | "punpcklbw %%xmm1,%%xmm3 \n" \ |
2440 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2441 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2442 | | "lea 0x8(%[y_buf]),%[y_buf] \n" \ |
2443 | | "movq (%[a_buf]),%%xmm5 \n" \ |
2444 | | "lea 0x8(%[a_buf]),%[a_buf] \n" |
2445 | | |
2446 | | // Read 4 UV from NV12, upsample to 8 UV |
2447 | | #define READNV12 \ |
2448 | | "movq (%[uv_buf]),%%xmm3 \n" \ |
2449 | | "lea 0x8(%[uv_buf]),%[uv_buf] \n" \ |
2450 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2451 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2452 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2453 | | "lea 0x8(%[y_buf]),%[y_buf] \n" |
2454 | | |
2455 | | // Read 4 VU from NV21, upsample to 8 UV |
2456 | | #define READNV21 \ |
2457 | | "movq (%[vu_buf]),%%xmm3 \n" \ |
2458 | | "lea 0x8(%[vu_buf]),%[vu_buf] \n" \ |
2459 | | "pshufb %[kShuffleNV21], %%xmm3 \n" \ |
2460 | | "movq (%[y_buf]),%%xmm4 \n" \ |
2461 | | "punpcklbw %%xmm4,%%xmm4 \n" \ |
2462 | | "lea 0x8(%[y_buf]),%[y_buf] \n" |
2463 | | |
2464 | | // Read 4 YUY2 with 8 Y and upsample 4 UV to 8 UV. |
2465 | | // xmm6 kShuffleYUY2Y, |
2466 | | // xmm7 kShuffleYUY2UV |
2467 | | #define READYUY2 \ |
2468 | | "movdqu (%[yuy2_buf]),%%xmm4 \n" \ |
2469 | | "lea 0x10(%[yuy2_buf]),%[yuy2_buf] \n" \ |
2470 | | "movdqa %%xmm4,%%xmm3 \n" \ |
2471 | | "pshufb %%xmm6,%%xmm4 \n" \ |
2472 | | "pshufb %%xmm7,%%xmm3 \n" |
2473 | | |
2474 | | // Read 4 UYVY with 8 Y and upsample 4 UV to 8 UV. |
2475 | | // xmm6 kShuffleUYVYY, |
2476 | | // xmm7 kShuffleUYVYUV |
2477 | | #define READUYVY \ |
2478 | | "movdqu (%[uyvy_buf]),%%xmm4 \n" \ |
2479 | | "lea 0x10(%[uyvy_buf]),%[uyvy_buf] \n" \ |
2480 | | "movdqa %%xmm4,%%xmm3 \n" \ |
2481 | | "pshufb %%xmm6,%%xmm4 \n" \ |
2482 | | "pshufb %%xmm7,%%xmm3 \n" |
2483 | | |
2484 | | // Read 4 UV from P210, upsample to 8 UV |
2485 | | #define READP210 \ |
2486 | | "movdqu (%[uv_buf]),%%xmm3 \n" \ |
2487 | | "lea 0x10(%[uv_buf]),%[uv_buf] \n" \ |
2488 | | "psrlw $0x8,%%xmm3 \n" \ |
2489 | | "packuswb %%xmm3,%%xmm3 \n" \ |
2490 | | "punpcklwd %%xmm3,%%xmm3 \n" \ |
2491 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2492 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
2493 | | |
2494 | | // Read 8 UV from P410 |
2495 | | #define READP410 \ |
2496 | | "movdqu (%[uv_buf]),%%xmm3 \n" \ |
2497 | | "movdqu 0x10(%[uv_buf]),%%xmm1 \n" \ |
2498 | | "lea 0x20(%[uv_buf]),%[uv_buf] \n" \ |
2499 | | "psrlw $0x8,%%xmm3 \n" \ |
2500 | | "psrlw $0x8,%%xmm1 \n" \ |
2501 | | "packuswb %%xmm1,%%xmm3 \n" \ |
2502 | | "movdqu (%[y_buf]),%%xmm4 \n" \ |
2503 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
2504 | | |
2505 | | #if defined(__x86_64__) |
2506 | | #define YUVTORGB_SETUP(yuvconstants) \ |
2507 | | "pcmpeqb %%xmm13,%%xmm13 \n" \ |
2508 | | "movdqa (%[yuvconstants]),%%xmm8 \n" \ |
2509 | | "pxor %%xmm12,%%xmm12 \n" \ |
2510 | | "movdqa 32(%[yuvconstants]),%%xmm9 \n" \ |
2511 | | "psllw $7,%%xmm13 \n" \ |
2512 | | "movdqa 64(%[yuvconstants]),%%xmm10 \n" \ |
2513 | | "pshufb %%xmm12,%%xmm13 \n" \ |
2514 | | "movdqa 96(%[yuvconstants]),%%xmm11 \n" \ |
2515 | | "movdqa 128(%[yuvconstants]),%%xmm12 \n" |
2516 | | |
2517 | | // Convert 8 pixels: 8 UV and 8 Y |
2518 | | #define YUVTORGB16(yuvconstants) \ |
2519 | | "psubb %%xmm13,%%xmm3 \n" \ |
2520 | | "pmulhuw %%xmm11,%%xmm4 \n" \ |
2521 | | "movdqa %%xmm8,%%xmm0 \n" \ |
2522 | | "movdqa %%xmm9,%%xmm1 \n" \ |
2523 | | "movdqa %%xmm10,%%xmm2 \n" \ |
2524 | | "paddw %%xmm12,%%xmm4 \n" \ |
2525 | | "pmaddubsw %%xmm3,%%xmm0 \n" \ |
2526 | | "pmaddubsw %%xmm3,%%xmm1 \n" \ |
2527 | | "pmaddubsw %%xmm3,%%xmm2 \n" \ |
2528 | | "paddsw %%xmm4,%%xmm0 \n" \ |
2529 | | "paddsw %%xmm4,%%xmm2 \n" \ |
2530 | | "psubsw %%xmm1,%%xmm4 \n" \ |
2531 | | "movdqa %%xmm4,%%xmm1 \n" |
2532 | | |
2533 | | #define YUVTORGB_REGS "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", |
2534 | | |
2535 | | #else |
2536 | | #define YUVTORGB_SETUP(yuvconstants) |
2537 | | |
2538 | | // Convert 8 pixels: 8 UV and 8 Y |
2539 | | #define YUVTORGB16(yuvconstants) \ |
2540 | | "pcmpeqb %%xmm0,%%xmm0 \n" \ |
2541 | | "pxor %%xmm1,%%xmm1 \n" \ |
2542 | | "psllw $7,%%xmm0 \n" \ |
2543 | | "pshufb %%xmm1,%%xmm0 \n" \ |
2544 | | "psubb %%xmm0,%%xmm3 \n" \ |
2545 | | "pmulhuw 96(%[yuvconstants]),%%xmm4 \n" \ |
2546 | | "movdqa (%[yuvconstants]),%%xmm0 \n" \ |
2547 | | "movdqa 32(%[yuvconstants]),%%xmm1 \n" \ |
2548 | | "movdqa 64(%[yuvconstants]),%%xmm2 \n" \ |
2549 | | "pmaddubsw %%xmm3,%%xmm0 \n" \ |
2550 | | "pmaddubsw %%xmm3,%%xmm1 \n" \ |
2551 | | "pmaddubsw %%xmm3,%%xmm2 \n" \ |
2552 | | "movdqa 128(%[yuvconstants]),%%xmm3 \n" \ |
2553 | | "paddw %%xmm3,%%xmm4 \n" \ |
2554 | | "paddsw %%xmm4,%%xmm0 \n" \ |
2555 | | "paddsw %%xmm4,%%xmm2 \n" \ |
2556 | | "psubsw %%xmm1,%%xmm4 \n" \ |
2557 | | "movdqa %%xmm4,%%xmm1 \n" |
2558 | | |
2559 | | #define YUVTORGB_REGS |
2560 | | #endif |
2561 | | |
2562 | | #define YUVTORGB(yuvconstants) \ |
2563 | | YUVTORGB16(yuvconstants) \ |
2564 | | "psraw $0x6,%%xmm0 \n" \ |
2565 | | "psraw $0x6,%%xmm1 \n" \ |
2566 | | "psraw $0x6,%%xmm2 \n" \ |
2567 | | "packuswb %%xmm0,%%xmm0 \n" \ |
2568 | | "packuswb %%xmm1,%%xmm1 \n" \ |
2569 | | "packuswb %%xmm2,%%xmm2 \n" |
2570 | | |
2571 | | // Store 8 ARGB values. |
2572 | | #define STOREARGB \ |
2573 | | "punpcklbw %%xmm1,%%xmm0 \n" \ |
2574 | | "punpcklbw %%xmm5,%%xmm2 \n" \ |
2575 | | "movdqa %%xmm0,%%xmm1 \n" \ |
2576 | | "punpcklwd %%xmm2,%%xmm0 \n" \ |
2577 | | "punpckhwd %%xmm2,%%xmm1 \n" \ |
2578 | | "movdqu %%xmm0,(%[dst_argb]) \n" \ |
2579 | | "movdqu %%xmm1,0x10(%[dst_argb]) \n" \ |
2580 | | "lea 0x20(%[dst_argb]), %[dst_argb] \n" |
2581 | | |
2582 | | // Store 8 RGBA values. |
2583 | | #define STORERGBA \ |
2584 | | "pcmpeqb %%xmm5,%%xmm5 \n" \ |
2585 | | "punpcklbw %%xmm2,%%xmm1 \n" \ |
2586 | | "punpcklbw %%xmm0,%%xmm5 \n" \ |
2587 | | "movdqa %%xmm5,%%xmm0 \n" \ |
2588 | | "punpcklwd %%xmm1,%%xmm5 \n" \ |
2589 | | "punpckhwd %%xmm1,%%xmm0 \n" \ |
2590 | | "movdqu %%xmm5,(%[dst_rgba]) \n" \ |
2591 | | "movdqu %%xmm0,0x10(%[dst_rgba]) \n" \ |
2592 | | "lea 0x20(%[dst_rgba]),%[dst_rgba] \n" |
2593 | | |
2594 | | // Store 8 RGB24 values. |
2595 | | #define STORERGB24 \ |
2596 | | "punpcklbw %%xmm1,%%xmm0 \n" \ |
2597 | | "punpcklbw %%xmm2,%%xmm2 \n" \ |
2598 | | "movdqa %%xmm0,%%xmm1 \n" \ |
2599 | | "punpcklwd %%xmm2,%%xmm0 \n" \ |
2600 | | "punpckhwd %%xmm2,%%xmm1 \n" \ |
2601 | | "pshufb %%xmm5,%%xmm0 \n" \ |
2602 | | "pshufb %%xmm6,%%xmm1 \n" \ |
2603 | | "palignr $0xc,%%xmm0,%%xmm1 \n" \ |
2604 | | "movq %%xmm0,(%[dst_rgb24]) \n" \ |
2605 | | "movdqu %%xmm1,0x8(%[dst_rgb24]) \n" \ |
2606 | | "lea 0x18(%[dst_rgb24]),%[dst_rgb24] \n" |
2607 | | |
2608 | | // Store 8 AR30 values. |
2609 | | #define STOREAR30 \ |
2610 | | "psraw $0x4,%%xmm0 \n" \ |
2611 | | "psraw $0x4,%%xmm1 \n" \ |
2612 | | "psraw $0x4,%%xmm2 \n" \ |
2613 | | "pminsw %%xmm7,%%xmm0 \n" \ |
2614 | | "pminsw %%xmm7,%%xmm1 \n" \ |
2615 | | "pminsw %%xmm7,%%xmm2 \n" \ |
2616 | | "pmaxsw %%xmm6,%%xmm0 \n" \ |
2617 | | "pmaxsw %%xmm6,%%xmm1 \n" \ |
2618 | | "pmaxsw %%xmm6,%%xmm2 \n" \ |
2619 | | "psllw $0x4,%%xmm2 \n" \ |
2620 | | "movdqa %%xmm0,%%xmm3 \n" \ |
2621 | | "punpcklwd %%xmm2,%%xmm0 \n" \ |
2622 | | "punpckhwd %%xmm2,%%xmm3 \n" \ |
2623 | | "movdqa %%xmm1,%%xmm2 \n" \ |
2624 | | "punpcklwd %%xmm5,%%xmm1 \n" \ |
2625 | | "punpckhwd %%xmm5,%%xmm2 \n" \ |
2626 | | "pslld $0xa,%%xmm1 \n" \ |
2627 | | "pslld $0xa,%%xmm2 \n" \ |
2628 | | "por %%xmm1,%%xmm0 \n" \ |
2629 | | "por %%xmm2,%%xmm3 \n" \ |
2630 | | "movdqu %%xmm0,(%[dst_ar30]) \n" \ |
2631 | | "movdqu %%xmm3,0x10(%[dst_ar30]) \n" \ |
2632 | | "lea 0x20(%[dst_ar30]), %[dst_ar30] \n" |
2633 | | |
2634 | | void OMITFP I444ToARGBRow_SSSE3(const uint8_t* y_buf, |
2635 | | const uint8_t* u_buf, |
2636 | | const uint8_t* v_buf, |
2637 | | uint8_t* dst_argb, |
2638 | | const struct YuvConstants* yuvconstants, |
2639 | 0 | int width) { |
2640 | 0 | asm volatile ( |
2641 | 0 | YUVTORGB_SETUP(yuvconstants) |
2642 | 0 | "sub %[u_buf],%[v_buf] \n" |
2643 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2644 | |
|
2645 | 0 | LABELALIGN |
2646 | 0 | "1: \n" |
2647 | 0 | READYUV444 |
2648 | 0 | YUVTORGB(yuvconstants) |
2649 | 0 | STOREARGB |
2650 | 0 | "sub $0x8,%[width] \n" |
2651 | 0 | "jg 1b \n" |
2652 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2653 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2654 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2655 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
2656 | 0 | [width]"+rm"(width) // %[width] |
2657 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2658 | 0 | : "memory", "cc", YUVTORGB_REGS |
2659 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
2660 | 0 | ); |
2661 | 0 | } |
2662 | | |
2663 | | #ifdef HAS_I444ALPHATOARGBROW_SSSE3 |
2664 | | void OMITFP I444AlphaToARGBRow_SSSE3(const uint8_t* y_buf, |
2665 | | const uint8_t* u_buf, |
2666 | | const uint8_t* v_buf, |
2667 | | const uint8_t* a_buf, |
2668 | | uint8_t* dst_argb, |
2669 | | const struct YuvConstants* yuvconstants, |
2670 | 0 | int width) { |
2671 | 0 | asm volatile(YUVTORGB_SETUP( |
2672 | 0 | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
2673 | |
|
2674 | 0 | LABELALIGN "1: \n" READYUVA444 YUVTORGB(yuvconstants) |
2675 | 0 | STOREARGB |
2676 | 0 | "subl $0x8,%[width] \n" |
2677 | 0 | "jg 1b \n" |
2678 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
2679 | 0 | [u_buf] "+r"(u_buf), // %[u_buf] |
2680 | 0 | [v_buf] "+r"(v_buf), // %[v_buf] |
2681 | 0 | [a_buf] "+r"(a_buf), // %[a_buf] |
2682 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
2683 | | #if defined(__i386__) |
2684 | | [width] "+m"(width) // %[width] |
2685 | | #else |
2686 | 0 | [width] "+rm"(width) // %[width] |
2687 | 0 | #endif |
2688 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
2689 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
2690 | 0 | "xmm4", "xmm5"); |
2691 | 0 | } |
2692 | | #endif // HAS_I444ALPHATOARGBROW_SSSE3 |
2693 | | |
2694 | | void OMITFP I422ToRGB24Row_SSSE3(const uint8_t* y_buf, |
2695 | | const uint8_t* u_buf, |
2696 | | const uint8_t* v_buf, |
2697 | | uint8_t* dst_rgb24, |
2698 | | const struct YuvConstants* yuvconstants, |
2699 | 0 | int width) { |
2700 | 0 | asm volatile ( |
2701 | 0 | YUVTORGB_SETUP(yuvconstants) |
2702 | 0 | "movdqa %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n" |
2703 | 0 | "movdqa %[kShuffleMaskARGBToRGB24],%%xmm6 \n" |
2704 | 0 | "sub %[u_buf],%[v_buf] \n" |
2705 | |
|
2706 | 0 | LABELALIGN |
2707 | 0 | "1: \n" |
2708 | 0 | READYUV422 |
2709 | 0 | YUVTORGB(yuvconstants) |
2710 | 0 | STORERGB24 |
2711 | 0 | "subl $0x8,%[width] \n" |
2712 | 0 | "jg 1b \n" |
2713 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2714 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2715 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2716 | 0 | [dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24] |
2717 | | #if defined(__i386__) |
2718 | | [width]"+m"(width) // %[width] |
2719 | | #else |
2720 | 0 | [width]"+rm"(width) // %[width] |
2721 | 0 | #endif |
2722 | 0 | : [yuvconstants]"r"(yuvconstants), // %[yuvconstants] |
2723 | 0 | [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0), |
2724 | 0 | [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24) |
2725 | 0 | : "memory", "cc", YUVTORGB_REGS |
2726 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
2727 | 0 | ); |
2728 | 0 | } |
2729 | | |
2730 | | void OMITFP I444ToRGB24Row_SSSE3(const uint8_t* y_buf, |
2731 | | const uint8_t* u_buf, |
2732 | | const uint8_t* v_buf, |
2733 | | uint8_t* dst_rgb24, |
2734 | | const struct YuvConstants* yuvconstants, |
2735 | 0 | int width) { |
2736 | 0 | asm volatile ( |
2737 | 0 | YUVTORGB_SETUP(yuvconstants) |
2738 | 0 | "movdqa %[kShuffleMaskARGBToRGB24_0],%%xmm5 \n" |
2739 | 0 | "movdqa %[kShuffleMaskARGBToRGB24],%%xmm6 \n" |
2740 | 0 | "sub %[u_buf],%[v_buf] \n" |
2741 | |
|
2742 | 0 | LABELALIGN |
2743 | 0 | "1: \n" |
2744 | 0 | READYUV444 |
2745 | 0 | YUVTORGB(yuvconstants) |
2746 | 0 | STORERGB24 |
2747 | 0 | "subl $0x8,%[width] \n" |
2748 | 0 | "jg 1b \n" |
2749 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2750 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2751 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2752 | 0 | [dst_rgb24]"+r"(dst_rgb24), // %[dst_rgb24] |
2753 | | #if defined(__i386__) |
2754 | | [width]"+m"(width) // %[width] |
2755 | | #else |
2756 | 0 | [width]"+rm"(width) // %[width] |
2757 | 0 | #endif |
2758 | 0 | : [yuvconstants]"r"(yuvconstants), // %[yuvconstants] |
2759 | 0 | [kShuffleMaskARGBToRGB24_0]"m"(kShuffleMaskARGBToRGB24_0), |
2760 | 0 | [kShuffleMaskARGBToRGB24]"m"(kShuffleMaskARGBToRGB24) |
2761 | 0 | : "memory", "cc", YUVTORGB_REGS |
2762 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6" |
2763 | 0 | ); |
2764 | 0 | } |
2765 | | |
2766 | | void OMITFP I422ToARGBRow_SSSE3(const uint8_t* y_buf, |
2767 | | const uint8_t* u_buf, |
2768 | | const uint8_t* v_buf, |
2769 | | uint8_t* dst_argb, |
2770 | | const struct YuvConstants* yuvconstants, |
2771 | 0 | int width) { |
2772 | 0 | asm volatile ( |
2773 | 0 | YUVTORGB_SETUP(yuvconstants) |
2774 | 0 | "sub %[u_buf],%[v_buf] \n" |
2775 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2776 | |
|
2777 | 0 | LABELALIGN |
2778 | 0 | "1: \n" |
2779 | 0 | READYUV422 |
2780 | 0 | YUVTORGB(yuvconstants) |
2781 | 0 | STOREARGB |
2782 | 0 | "sub $0x8,%[width] \n" |
2783 | 0 | "jg 1b \n" |
2784 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2785 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2786 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2787 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
2788 | 0 | [width]"+rm"(width) // %[width] |
2789 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2790 | 0 | : "memory", "cc", YUVTORGB_REGS |
2791 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
2792 | 0 | ); |
2793 | 0 | } |
2794 | | |
2795 | | void OMITFP I422ToAR30Row_SSSE3(const uint8_t* y_buf, |
2796 | | const uint8_t* u_buf, |
2797 | | const uint8_t* v_buf, |
2798 | | uint8_t* dst_ar30, |
2799 | | const struct YuvConstants* yuvconstants, |
2800 | 0 | int width) { |
2801 | 0 | asm volatile ( |
2802 | 0 | YUVTORGB_SETUP(yuvconstants) |
2803 | 0 | "sub %[u_buf],%[v_buf] \n" |
2804 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" // AR30 constants |
2805 | 0 | "psrlw $14,%%xmm5 \n" |
2806 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
2807 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
2808 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
2809 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
2810 | |
|
2811 | 0 | LABELALIGN |
2812 | 0 | "1: \n" |
2813 | 0 | READYUV422 |
2814 | 0 | YUVTORGB16(yuvconstants) |
2815 | 0 | STOREAR30 |
2816 | 0 | "sub $0x8,%[width] \n" |
2817 | 0 | "jg 1b \n" |
2818 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2819 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2820 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2821 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
2822 | 0 | [width]"+rm"(width) // %[width] |
2823 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2824 | 0 | : "memory", "cc", YUVTORGB_REGS |
2825 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
2826 | 0 | ); |
2827 | 0 | } |
2828 | | |
2829 | | // 10 bit YUV to ARGB |
2830 | | void OMITFP I210ToARGBRow_SSSE3(const uint16_t* y_buf, |
2831 | | const uint16_t* u_buf, |
2832 | | const uint16_t* v_buf, |
2833 | | uint8_t* dst_argb, |
2834 | | const struct YuvConstants* yuvconstants, |
2835 | 0 | int width) { |
2836 | 0 | asm volatile ( |
2837 | 0 | YUVTORGB_SETUP(yuvconstants) |
2838 | 0 | "sub %[u_buf],%[v_buf] \n" |
2839 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2840 | |
|
2841 | 0 | LABELALIGN |
2842 | 0 | "1: \n" |
2843 | 0 | READYUV210 |
2844 | 0 | YUVTORGB(yuvconstants) |
2845 | 0 | STOREARGB |
2846 | 0 | "sub $0x8,%[width] \n" |
2847 | 0 | "jg 1b \n" |
2848 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2849 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2850 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2851 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
2852 | 0 | [width]"+rm"(width) // %[width] |
2853 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2854 | 0 | : "memory", "cc", YUVTORGB_REGS |
2855 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
2856 | 0 | ); |
2857 | 0 | } |
2858 | | |
2859 | | // 12 bit YUV to ARGB |
2860 | | void OMITFP I212ToARGBRow_SSSE3(const uint16_t* y_buf, |
2861 | | const uint16_t* u_buf, |
2862 | | const uint16_t* v_buf, |
2863 | | uint8_t* dst_argb, |
2864 | | const struct YuvConstants* yuvconstants, |
2865 | 0 | int width) { |
2866 | 0 | asm volatile ( |
2867 | 0 | YUVTORGB_SETUP(yuvconstants) |
2868 | 0 | "sub %[u_buf],%[v_buf] \n" |
2869 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2870 | |
|
2871 | 0 | LABELALIGN |
2872 | 0 | "1: \n" |
2873 | 0 | READYUV212 |
2874 | 0 | YUVTORGB(yuvconstants) |
2875 | 0 | STOREARGB |
2876 | 0 | "sub $0x8,%[width] \n" |
2877 | 0 | "jg 1b \n" |
2878 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2879 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2880 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2881 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
2882 | 0 | [width]"+rm"(width) // %[width] |
2883 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2884 | 0 | : "memory", "cc", YUVTORGB_REGS |
2885 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
2886 | 0 | ); |
2887 | 0 | } |
2888 | | |
2889 | | // 10 bit YUV to AR30 |
2890 | | void OMITFP I210ToAR30Row_SSSE3(const uint16_t* y_buf, |
2891 | | const uint16_t* u_buf, |
2892 | | const uint16_t* v_buf, |
2893 | | uint8_t* dst_ar30, |
2894 | | const struct YuvConstants* yuvconstants, |
2895 | 0 | int width) { |
2896 | 0 | asm volatile ( |
2897 | 0 | YUVTORGB_SETUP(yuvconstants) |
2898 | 0 | "sub %[u_buf],%[v_buf] \n" |
2899 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2900 | 0 | "psrlw $14,%%xmm5 \n" |
2901 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
2902 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
2903 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
2904 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
2905 | |
|
2906 | 0 | LABELALIGN |
2907 | 0 | "1: \n" |
2908 | 0 | READYUV210 |
2909 | 0 | YUVTORGB16(yuvconstants) |
2910 | 0 | STOREAR30 |
2911 | 0 | "sub $0x8,%[width] \n" |
2912 | 0 | "jg 1b \n" |
2913 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2914 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2915 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2916 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
2917 | 0 | [width]"+rm"(width) // %[width] |
2918 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2919 | 0 | : "memory", "cc", YUVTORGB_REGS |
2920 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
2921 | 0 | ); |
2922 | 0 | } |
2923 | | |
2924 | | // 12 bit YUV to AR30 |
2925 | | void OMITFP I212ToAR30Row_SSSE3(const uint16_t* y_buf, |
2926 | | const uint16_t* u_buf, |
2927 | | const uint16_t* v_buf, |
2928 | | uint8_t* dst_ar30, |
2929 | | const struct YuvConstants* yuvconstants, |
2930 | 0 | int width) { |
2931 | 0 | asm volatile ( |
2932 | 0 | YUVTORGB_SETUP(yuvconstants) |
2933 | 0 | "sub %[u_buf],%[v_buf] \n" |
2934 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2935 | 0 | "psrlw $14,%%xmm5 \n" |
2936 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
2937 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
2938 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
2939 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
2940 | |
|
2941 | 0 | LABELALIGN |
2942 | 0 | "1: \n" |
2943 | 0 | READYUV212 |
2944 | 0 | YUVTORGB16(yuvconstants) |
2945 | 0 | STOREAR30 |
2946 | 0 | "sub $0x8,%[width] \n" |
2947 | 0 | "jg 1b \n" |
2948 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2949 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2950 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2951 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
2952 | 0 | [width]"+rm"(width) // %[width] |
2953 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2954 | 0 | : "memory", "cc", YUVTORGB_REGS |
2955 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
2956 | 0 | ); |
2957 | 0 | } |
2958 | | |
2959 | | // 10 bit YUV to ARGB |
2960 | | void OMITFP I410ToARGBRow_SSSE3(const uint16_t* y_buf, |
2961 | | const uint16_t* u_buf, |
2962 | | const uint16_t* v_buf, |
2963 | | uint8_t* dst_argb, |
2964 | | const struct YuvConstants* yuvconstants, |
2965 | 0 | int width) { |
2966 | 0 | asm volatile ( |
2967 | 0 | YUVTORGB_SETUP(yuvconstants) |
2968 | 0 | "sub %[u_buf],%[v_buf] \n" |
2969 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
2970 | |
|
2971 | 0 | LABELALIGN |
2972 | 0 | "1: \n" |
2973 | 0 | READYUV410 |
2974 | 0 | YUVTORGB(yuvconstants) |
2975 | 0 | STOREARGB |
2976 | 0 | "sub $0x8,%[width] \n" |
2977 | 0 | "jg 1b \n" |
2978 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
2979 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
2980 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
2981 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
2982 | 0 | [width]"+rm"(width) // %[width] |
2983 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
2984 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
2985 | 0 | ); |
2986 | 0 | } |
2987 | | |
2988 | | #ifdef HAS_I210ALPHATOARGBROW_SSSE3 |
2989 | | // 10 bit YUVA to ARGB |
2990 | | void OMITFP I210AlphaToARGBRow_SSSE3(const uint16_t* y_buf, |
2991 | | const uint16_t* u_buf, |
2992 | | const uint16_t* v_buf, |
2993 | | const uint16_t* a_buf, |
2994 | | uint8_t* dst_argb, |
2995 | | const struct YuvConstants* yuvconstants, |
2996 | 0 | int width) { |
2997 | 0 | asm volatile(YUVTORGB_SETUP( |
2998 | 0 | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
2999 | |
|
3000 | 0 | LABELALIGN "1: \n" READYUVA210 YUVTORGB(yuvconstants) |
3001 | 0 | STOREARGB |
3002 | 0 | "subl $0x8,%[width] \n" |
3003 | 0 | "jg 1b \n" |
3004 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3005 | 0 | [u_buf] "+r"(u_buf), // %[u_buf] |
3006 | 0 | [v_buf] "+r"(v_buf), // %[v_buf] |
3007 | 0 | [a_buf] "+r"(a_buf), |
3008 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3009 | | #if defined(__i386__) |
3010 | | [width] "+m"(width) // %[width] |
3011 | | #else |
3012 | 0 | [width] "+rm"(width) // %[width] |
3013 | 0 | #endif |
3014 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3015 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3016 | 0 | "xmm4", "xmm5"); |
3017 | 0 | } |
3018 | | #endif |
3019 | | |
3020 | | #ifdef HAS_I410ALPHATOARGBROW_SSSE3 |
3021 | | // 10 bit YUVA to ARGB |
3022 | | void OMITFP I410AlphaToARGBRow_SSSE3(const uint16_t* y_buf, |
3023 | | const uint16_t* u_buf, |
3024 | | const uint16_t* v_buf, |
3025 | | const uint16_t* a_buf, |
3026 | | uint8_t* dst_argb, |
3027 | | const struct YuvConstants* yuvconstants, |
3028 | 0 | int width) { |
3029 | 0 | asm volatile(YUVTORGB_SETUP( |
3030 | 0 | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
3031 | |
|
3032 | 0 | LABELALIGN "1: \n" READYUVA410 YUVTORGB(yuvconstants) |
3033 | 0 | STOREARGB |
3034 | 0 | "subl $0x8,%[width] \n" |
3035 | 0 | "jg 1b \n" |
3036 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3037 | 0 | [u_buf] "+r"(u_buf), // %[u_buf] |
3038 | 0 | [v_buf] "+r"(v_buf), // %[v_buf] |
3039 | 0 | [a_buf] "+r"(a_buf), |
3040 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3041 | | #if defined(__i386__) |
3042 | | [width] "+m"(width) // %[width] |
3043 | | #else |
3044 | 0 | [width] "+rm"(width) // %[width] |
3045 | 0 | #endif |
3046 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3047 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3048 | 0 | "xmm4", "xmm5"); |
3049 | 0 | } |
3050 | | #endif |
3051 | | |
3052 | | // 10 bit YUV to AR30 |
3053 | | void OMITFP I410ToAR30Row_SSSE3(const uint16_t* y_buf, |
3054 | | const uint16_t* u_buf, |
3055 | | const uint16_t* v_buf, |
3056 | | uint8_t* dst_ar30, |
3057 | | const struct YuvConstants* yuvconstants, |
3058 | 0 | int width) { |
3059 | 0 | asm volatile ( |
3060 | 0 | YUVTORGB_SETUP(yuvconstants) |
3061 | 0 | "sub %[u_buf],%[v_buf] \n" |
3062 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
3063 | 0 | "psrlw $14,%%xmm5 \n" |
3064 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
3065 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
3066 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
3067 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
3068 | |
|
3069 | 0 | LABELALIGN |
3070 | 0 | "1: \n" |
3071 | 0 | READYUV410 |
3072 | 0 | YUVTORGB16(yuvconstants) |
3073 | 0 | STOREAR30 |
3074 | 0 | "sub $0x8,%[width] \n" |
3075 | 0 | "jg 1b \n" |
3076 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3077 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
3078 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
3079 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
3080 | 0 | [width]"+rm"(width) // %[width] |
3081 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3082 | 0 | : "memory", "cc", YUVTORGB_REGS |
3083 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
3084 | 0 | ); |
3085 | 0 | } |
3086 | | |
3087 | | #ifdef HAS_I422ALPHATOARGBROW_SSSE3 |
3088 | | void OMITFP I422AlphaToARGBRow_SSSE3(const uint8_t* y_buf, |
3089 | | const uint8_t* u_buf, |
3090 | | const uint8_t* v_buf, |
3091 | | const uint8_t* a_buf, |
3092 | | uint8_t* dst_argb, |
3093 | | const struct YuvConstants* yuvconstants, |
3094 | 0 | int width) { |
3095 | 0 | asm volatile(YUVTORGB_SETUP( |
3096 | 0 | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
3097 | |
|
3098 | 0 | LABELALIGN "1: \n" READYUVA422 YUVTORGB(yuvconstants) |
3099 | 0 | STOREARGB |
3100 | 0 | "subl $0x8,%[width] \n" |
3101 | 0 | "jg 1b \n" |
3102 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3103 | 0 | [u_buf] "+r"(u_buf), // %[u_buf] |
3104 | 0 | [v_buf] "+r"(v_buf), // %[v_buf] |
3105 | 0 | [a_buf] "+r"(a_buf), // %[a_buf] |
3106 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3107 | | #if defined(__i386__) |
3108 | | [width] "+m"(width) // %[width] |
3109 | | #else |
3110 | 0 | [width] "+rm"(width) // %[width] |
3111 | 0 | #endif |
3112 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3113 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3114 | 0 | "xmm4", "xmm5"); |
3115 | 0 | } |
3116 | | #endif // HAS_I422ALPHATOARGBROW_SSSE3 |
3117 | | |
3118 | | void OMITFP NV12ToARGBRow_SSSE3(const uint8_t* y_buf, |
3119 | | const uint8_t* uv_buf, |
3120 | | uint8_t* dst_argb, |
3121 | | const struct YuvConstants* yuvconstants, |
3122 | 0 | int width) { |
3123 | 0 | asm volatile(YUVTORGB_SETUP( |
3124 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3125 | |
|
3126 | 0 | LABELALIGN "1: \n" READNV12 YUVTORGB(yuvconstants) |
3127 | 0 | STOREARGB |
3128 | 0 | "sub $0x8,%[width] \n" |
3129 | 0 | "jg 1b \n" |
3130 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3131 | 0 | [uv_buf] "+r"(uv_buf), // %[uv_buf] |
3132 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3133 | 0 | [width] "+rm"(width) // %[width] |
3134 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3135 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3136 | 0 | "xmm4", "xmm5"); |
3137 | 0 | } |
3138 | | |
3139 | | void OMITFP NV21ToARGBRow_SSSE3(const uint8_t* y_buf, |
3140 | | const uint8_t* vu_buf, |
3141 | | uint8_t* dst_argb, |
3142 | | const struct YuvConstants* yuvconstants, |
3143 | 0 | int width) { |
3144 | 0 | asm volatile(YUVTORGB_SETUP( |
3145 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3146 | |
|
3147 | 0 | LABELALIGN "1: \n" READNV21 YUVTORGB(yuvconstants) |
3148 | 0 | STOREARGB |
3149 | 0 | "sub $0x8,%[width] \n" |
3150 | 0 | "jg 1b \n" |
3151 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3152 | 0 | [vu_buf] "+r"(vu_buf), // %[vu_buf] |
3153 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3154 | 0 | [width] "+rm"(width) // %[width] |
3155 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
3156 | 0 | [kShuffleNV21] "m"(kShuffleNV21) |
3157 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3158 | 0 | "xmm4", "xmm5"); |
3159 | 0 | } |
3160 | | |
3161 | | void OMITFP YUY2ToARGBRow_SSSE3(const uint8_t* yuy2_buf, |
3162 | | uint8_t* dst_argb, |
3163 | | const struct YuvConstants* yuvconstants, |
3164 | 0 | int width) { |
3165 | 0 | asm volatile( |
3166 | 0 | "movdqa %[kShuffleYUY2Y],%%xmm6 \n" |
3167 | 0 | "movdqa %[kShuffleYUY2UV],%%xmm7 \n" YUVTORGB_SETUP( |
3168 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3169 | |
|
3170 | 0 | LABELALIGN "1: \n" READYUY2 YUVTORGB(yuvconstants) STOREARGB |
3171 | 0 | "sub $0x8,%[width] \n" |
3172 | 0 | "jg 1b \n" |
3173 | 0 | : [yuy2_buf] "+r"(yuy2_buf), // %[yuy2_buf] |
3174 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3175 | 0 | [width] "+rm"(width) // %[width] |
3176 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
3177 | 0 | [kShuffleYUY2Y] "m"(kShuffleYUY2Y), [kShuffleYUY2UV] "m"(kShuffleYUY2UV) |
3178 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", |
3179 | 0 | "xmm5", "xmm6", "xmm7"); |
3180 | 0 | } |
3181 | | |
3182 | | void OMITFP UYVYToARGBRow_SSSE3(const uint8_t* uyvy_buf, |
3183 | | uint8_t* dst_argb, |
3184 | | const struct YuvConstants* yuvconstants, |
3185 | 0 | int width) { |
3186 | 0 | asm volatile( |
3187 | 0 | "movdqa %[kShuffleUYVYY],%%xmm6 \n" |
3188 | 0 | "movdqa %[kShuffleUYVYUV],%%xmm7 \n" YUVTORGB_SETUP( |
3189 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3190 | |
|
3191 | 0 | LABELALIGN "1: \n" READUYVY YUVTORGB(yuvconstants) STOREARGB |
3192 | 0 | "sub $0x8,%[width] \n" |
3193 | 0 | "jg 1b \n" |
3194 | 0 | : [uyvy_buf] "+r"(uyvy_buf), // %[uyvy_buf] |
3195 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3196 | 0 | [width] "+rm"(width) // %[width] |
3197 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
3198 | 0 | [kShuffleUYVYY] "m"(kShuffleUYVYY), [kShuffleUYVYUV] "m"(kShuffleUYVYUV) |
3199 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", |
3200 | 0 | "xmm5"); |
3201 | 0 | } |
3202 | | |
3203 | | void OMITFP P210ToARGBRow_SSSE3(const uint16_t* y_buf, |
3204 | | const uint16_t* uv_buf, |
3205 | | uint8_t* dst_argb, |
3206 | | const struct YuvConstants* yuvconstants, |
3207 | 0 | int width) { |
3208 | 0 | asm volatile(YUVTORGB_SETUP( |
3209 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3210 | |
|
3211 | 0 | LABELALIGN "1: \n" READP210 YUVTORGB(yuvconstants) |
3212 | 0 | STOREARGB |
3213 | 0 | "sub $0x8,%[width] \n" |
3214 | 0 | "jg 1b \n" |
3215 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3216 | 0 | [uv_buf] "+r"(uv_buf), // %[u_buf] |
3217 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3218 | 0 | [width] "+rm"(width) // %[width] |
3219 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3220 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3221 | 0 | "xmm4", "xmm5"); |
3222 | 0 | } |
3223 | | |
3224 | | void OMITFP P410ToARGBRow_SSSE3(const uint16_t* y_buf, |
3225 | | const uint16_t* uv_buf, |
3226 | | uint8_t* dst_argb, |
3227 | | const struct YuvConstants* yuvconstants, |
3228 | 0 | int width) { |
3229 | 0 | asm volatile(YUVTORGB_SETUP( |
3230 | 0 | yuvconstants) "pcmpeqb %%xmm5,%%xmm5 \n" |
3231 | |
|
3232 | 0 | LABELALIGN "1: \n" READP410 YUVTORGB(yuvconstants) |
3233 | 0 | STOREARGB |
3234 | 0 | "sub $0x8,%[width] \n" |
3235 | 0 | "jg 1b \n" |
3236 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
3237 | 0 | [uv_buf] "+r"(uv_buf), // %[u_buf] |
3238 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
3239 | 0 | [width] "+rm"(width) // %[width] |
3240 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
3241 | 0 | : "memory", "cc", YUVTORGB_REGS "xmm0", "xmm1", "xmm2", "xmm3", |
3242 | 0 | "xmm4", "xmm5"); |
3243 | 0 | } |
3244 | | |
3245 | | void OMITFP P210ToAR30Row_SSSE3(const uint16_t* y_buf, |
3246 | | const uint16_t* uv_buf, |
3247 | | uint8_t* dst_ar30, |
3248 | | const struct YuvConstants* yuvconstants, |
3249 | 0 | int width) { |
3250 | 0 | asm volatile ( |
3251 | 0 | YUVTORGB_SETUP(yuvconstants) |
3252 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
3253 | 0 | "psrlw $14,%%xmm5 \n" |
3254 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
3255 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
3256 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
3257 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
3258 | |
|
3259 | 0 | LABELALIGN |
3260 | 0 | "1: \n" |
3261 | 0 | READP210 |
3262 | 0 | YUVTORGB16(yuvconstants) |
3263 | 0 | STOREAR30 |
3264 | 0 | "sub $0x8,%[width] \n" |
3265 | 0 | "jg 1b \n" |
3266 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3267 | 0 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
3268 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
3269 | 0 | [width]"+rm"(width) // %[width] |
3270 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3271 | 0 | : "memory", "cc", YUVTORGB_REGS |
3272 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
3273 | 0 | ); |
3274 | 0 | } |
3275 | | |
3276 | | void OMITFP P410ToAR30Row_SSSE3(const uint16_t* y_buf, |
3277 | | const uint16_t* uv_buf, |
3278 | | uint8_t* dst_ar30, |
3279 | | const struct YuvConstants* yuvconstants, |
3280 | 0 | int width) { |
3281 | 0 | asm volatile ( |
3282 | 0 | YUVTORGB_SETUP(yuvconstants) |
3283 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
3284 | 0 | "psrlw $14,%%xmm5 \n" |
3285 | 0 | "psllw $4,%%xmm5 \n" // 2 alpha bits |
3286 | 0 | "pxor %%xmm6,%%xmm6 \n" // 0 for min |
3287 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
3288 | 0 | "psrlw $6,%%xmm7 \n" // 1023 for max |
3289 | |
|
3290 | 0 | LABELALIGN |
3291 | 0 | "1: \n" |
3292 | 0 | READP410 |
3293 | 0 | YUVTORGB16(yuvconstants) |
3294 | 0 | STOREAR30 |
3295 | 0 | "sub $0x8,%[width] \n" |
3296 | 0 | "jg 1b \n" |
3297 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3298 | 0 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
3299 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
3300 | 0 | [width]"+rm"(width) // %[width] |
3301 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3302 | 0 | : "memory", "cc", YUVTORGB_REGS |
3303 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
3304 | 0 | ); |
3305 | 0 | } |
3306 | | |
3307 | | void OMITFP I422ToRGBARow_SSSE3(const uint8_t* y_buf, |
3308 | | const uint8_t* u_buf, |
3309 | | const uint8_t* v_buf, |
3310 | | uint8_t* dst_rgba, |
3311 | | const struct YuvConstants* yuvconstants, |
3312 | 0 | int width) { |
3313 | 0 | asm volatile ( |
3314 | 0 | YUVTORGB_SETUP(yuvconstants) |
3315 | 0 | "sub %[u_buf],%[v_buf] \n" |
3316 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
3317 | |
|
3318 | 0 | LABELALIGN |
3319 | 0 | "1: \n" |
3320 | 0 | READYUV422 |
3321 | 0 | YUVTORGB(yuvconstants) |
3322 | 0 | STORERGBA |
3323 | 0 | "sub $0x8,%[width] \n" |
3324 | 0 | "jg 1b \n" |
3325 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3326 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
3327 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
3328 | 0 | [dst_rgba]"+r"(dst_rgba), // %[dst_rgba] |
3329 | 0 | [width]"+rm"(width) // %[width] |
3330 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3331 | 0 | : "memory", "cc", YUVTORGB_REGS |
3332 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3333 | 0 | ); |
3334 | 0 | } |
3335 | | |
3336 | | #endif // HAS_I422TOARGBROW_SSSE3 |
3337 | | |
3338 | | // Read 16 UV from 444 |
3339 | | #define READYUV444_AVX2 \ |
3340 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3341 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3342 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3343 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3344 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3345 | | "vpunpcklbw %%ymm1,%%ymm3,%%ymm3 \n" \ |
3346 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3347 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3348 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3349 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
3350 | | |
3351 | | // Read 8 UV from 422, upsample to 16 UV. |
3352 | | #define READYUV422_AVX2 \ |
3353 | | "vmovq (%[u_buf]),%%xmm3 \n" \ |
3354 | | "vmovq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3355 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
3356 | | "vpunpcklbw %%ymm1,%%ymm3,%%ymm3 \n" \ |
3357 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3358 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3359 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3360 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3361 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3362 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
3363 | | |
3364 | | #define READYUV422_AVX512BW \ |
3365 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3366 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3367 | | "vpermq %%zmm3,%%zmm16,%%zmm3 \n" \ |
3368 | | "vpermq %%zmm1,%%zmm16,%%zmm1 \n" \ |
3369 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3370 | | "vpunpcklbw %%zmm1,%%zmm3,%%zmm3 \n" \ |
3371 | | "vpermq $0xd8,%%zmm3,%%zmm3 \n" \ |
3372 | | "vpunpcklwd %%zmm3,%%zmm3,%%zmm3 \n" \ |
3373 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3374 | | "vpermq %%zmm4,%%zmm17,%%zmm4 \n" \ |
3375 | | "vpermq $0xd8,%%zmm4,%%zmm4 \n" \ |
3376 | | "vpunpcklbw %%zmm4,%%zmm4,%%zmm4 \n" \ |
3377 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3378 | | |
3379 | | // Read 8 UV from 210, upsample to 16 UV |
3380 | | // TODO(fbarchard): Consider vpshufb to replace pack/unpack |
3381 | | // TODO(fbarchard): Consider vunpcklpd to combine the 2 registers into 1. |
3382 | | #define READYUV210_AVX2 \ |
3383 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3384 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3385 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3386 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3387 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3388 | | "vpunpcklwd %%ymm1,%%ymm3,%%ymm3 \n" \ |
3389 | | "vpsraw $2,%%ymm3,%%ymm3 \n" \ |
3390 | | "vpackuswb %%ymm3,%%ymm3,%%ymm3 \n" \ |
3391 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3392 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3393 | | "vpsllw $6,%%ymm4,%%ymm2 \n" \ |
3394 | | "vpsrlw $4,%%ymm4,%%ymm4 \n" \ |
3395 | | "vpaddw %%ymm2,%%ymm4,%%ymm4 \n" \ |
3396 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3397 | | |
3398 | | // Read 8 UV from 210, upsample to 16 UV. With 16 Alpha. |
3399 | | #define READYUVA210_AVX2 \ |
3400 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3401 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3402 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3403 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3404 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3405 | | "vpunpcklwd %%ymm1,%%ymm3,%%ymm3 \n" \ |
3406 | | "vpsraw $2,%%ymm3,%%ymm3 \n" \ |
3407 | | "vpackuswb %%ymm3,%%ymm3,%%ymm3 \n" \ |
3408 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3409 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3410 | | "vpsllw $6,%%ymm4,%%ymm2 \n" \ |
3411 | | "vpsrlw $4,%%ymm4,%%ymm4 \n" \ |
3412 | | "vpaddw %%ymm2,%%ymm4,%%ymm4 \n" \ |
3413 | | "lea 0x20(%[y_buf]),%[y_buf] \n" \ |
3414 | | "vmovdqu (%[a_buf]),%%ymm5 \n" \ |
3415 | | "vpsraw $2,%%ymm5,%%ymm5 \n" \ |
3416 | | "vpackuswb %%ymm5,%%ymm5,%%ymm5 \n" \ |
3417 | | "lea 0x20(%[a_buf]),%[a_buf] \n" |
3418 | | |
3419 | | // Read 16 UV from 410 |
3420 | | #define READYUV410_AVX2 \ |
3421 | | "vmovdqu (%[u_buf]),%%ymm3 \n" \ |
3422 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%ymm2 \n" \ |
3423 | | "lea 0x20(%[u_buf]),%[u_buf] \n" \ |
3424 | | "vpsraw $2,%%ymm3,%%ymm3 \n" \ |
3425 | | "vpsraw $2,%%ymm2,%%ymm2 \n" \ |
3426 | | "vpunpckhwd %%ymm2,%%ymm3,%%ymm1 \n" \ |
3427 | | "vpunpcklwd %%ymm2,%%ymm3,%%ymm3 \n" \ |
3428 | | "vpackuswb %%ymm1,%%ymm3,%%ymm3 \n" \ |
3429 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3430 | | "vpsllw $6,%%ymm4,%%ymm2 \n" \ |
3431 | | "vpsrlw $4,%%ymm4,%%ymm4 \n" \ |
3432 | | "vpaddw %%ymm2,%%ymm4,%%ymm4 \n" \ |
3433 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3434 | | |
3435 | | // Read 8 UV from 212 12 bit, upsample to 16 UV |
3436 | | #define READYUV212_AVX2 \ |
3437 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3438 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3439 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3440 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3441 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3442 | | "vpunpcklwd %%ymm1,%%ymm3,%%ymm3 \n" \ |
3443 | | "vpsraw $0x4,%%ymm3,%%ymm3 \n" \ |
3444 | | "vpackuswb %%ymm3,%%ymm3,%%ymm3 \n" \ |
3445 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3446 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3447 | | "vpsllw $4,%%ymm4,%%ymm2 \n" \ |
3448 | | "vpsrlw $8,%%ymm4,%%ymm4 \n" \ |
3449 | | "vpaddw %%ymm2,%%ymm4,%%ymm4 \n" \ |
3450 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3451 | | |
3452 | | // Read 16 UV from 410. With 16 Alpha. |
3453 | | #define READYUVA410_AVX2 \ |
3454 | | "vmovdqu (%[u_buf]),%%ymm3 \n" \ |
3455 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%ymm2 \n" \ |
3456 | | "lea 0x20(%[u_buf]),%[u_buf] \n" \ |
3457 | | "vpsraw $2,%%ymm3,%%ymm3 \n" \ |
3458 | | "vpsraw $2,%%ymm2,%%ymm2 \n" \ |
3459 | | "vpunpckhwd %%ymm2,%%ymm3,%%ymm1 \n" \ |
3460 | | "vpunpcklwd %%ymm2,%%ymm3,%%ymm3 \n" \ |
3461 | | "vpackuswb %%ymm1,%%ymm3,%%ymm3 \n" \ |
3462 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3463 | | "vpsllw $6,%%ymm4,%%ymm2 \n" \ |
3464 | | "vpsrlw $4,%%ymm4,%%ymm4 \n" \ |
3465 | | "vpaddw %%ymm2,%%ymm4,%%ymm4 \n" \ |
3466 | | "lea 0x20(%[y_buf]),%[y_buf] \n" \ |
3467 | | "vmovdqu (%[a_buf]),%%ymm5 \n" \ |
3468 | | "vpsraw $2,%%ymm5,%%ymm5 \n" \ |
3469 | | "vpackuswb %%ymm5,%%ymm5,%%ymm5 \n" \ |
3470 | | "lea 0x20(%[a_buf]),%[a_buf] \n" |
3471 | | |
3472 | | // Read 16 UV from 444. With 16 Alpha. |
3473 | | #define READYUVA444_AVX2 \ |
3474 | | "vmovdqu (%[u_buf]),%%xmm3 \n" \ |
3475 | | "vmovdqu 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3476 | | "lea 0x10(%[u_buf]),%[u_buf] \n" \ |
3477 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3478 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3479 | | "vpunpcklbw %%ymm1,%%ymm3,%%ymm3 \n" \ |
3480 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3481 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3482 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3483 | | "lea 0x10(%[y_buf]),%[y_buf] \n" \ |
3484 | | "vmovdqu (%[a_buf]),%%xmm5 \n" \ |
3485 | | "vpermq $0xd8,%%ymm5,%%ymm5 \n" \ |
3486 | | "lea 0x10(%[a_buf]),%[a_buf] \n" |
3487 | | |
3488 | | // Read 8 UV from 422, upsample to 16 UV. With 16 Alpha. |
3489 | | #define READYUVA422_AVX2 \ |
3490 | | "vmovq (%[u_buf]),%%xmm3 \n" \ |
3491 | | "vmovq 0x00(%[u_buf],%[v_buf],1),%%xmm1 \n" \ |
3492 | | "lea 0x8(%[u_buf]),%[u_buf] \n" \ |
3493 | | "vpunpcklbw %%ymm1,%%ymm3,%%ymm3 \n" \ |
3494 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3495 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3496 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3497 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3498 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3499 | | "lea 0x10(%[y_buf]),%[y_buf] \n" \ |
3500 | | "vmovdqu (%[a_buf]),%%xmm5 \n" \ |
3501 | | "vpermq $0xd8,%%ymm5,%%ymm5 \n" \ |
3502 | | "lea 0x10(%[a_buf]),%[a_buf] \n" |
3503 | | |
3504 | | // Read 8 UV from NV12, upsample to 16 UV. |
3505 | | #define READNV12_AVX2 \ |
3506 | | "vmovdqu (%[uv_buf]),%%xmm3 \n" \ |
3507 | | "lea 0x10(%[uv_buf]),%[uv_buf] \n" \ |
3508 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3509 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3510 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3511 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3512 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3513 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
3514 | | |
3515 | | // Read 8 VU from NV21, upsample to 16 UV. |
3516 | | #define READNV21_AVX2 \ |
3517 | | "vmovdqu (%[vu_buf]),%%xmm3 \n" \ |
3518 | | "lea 0x10(%[vu_buf]),%[vu_buf] \n" \ |
3519 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3520 | | "vpshufb %[kShuffleNV21], %%ymm3, %%ymm3 \n" \ |
3521 | | "vmovdqu (%[y_buf]),%%xmm4 \n" \ |
3522 | | "vpermq $0xd8,%%ymm4,%%ymm4 \n" \ |
3523 | | "vpunpcklbw %%ymm4,%%ymm4,%%ymm4 \n" \ |
3524 | | "lea 0x10(%[y_buf]),%[y_buf] \n" |
3525 | | |
3526 | | // Read 4 UV from P210, upsample to 8 UV |
3527 | | #define READP210_AVX2 \ |
3528 | | "vmovdqu (%[uv_buf]),%%ymm3 \n" \ |
3529 | | "lea 0x20(%[uv_buf]),%[uv_buf] \n" \ |
3530 | | "vpsrlw $0x8,%%ymm3,%%ymm3 \n" \ |
3531 | | "vpackuswb %%ymm3,%%ymm3,%%ymm3 \n" \ |
3532 | | "vpunpcklwd %%ymm3,%%ymm3,%%ymm3 \n" \ |
3533 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3534 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3535 | | |
3536 | | // Read 8 UV from P410 |
3537 | | #define READP410_AVX2 \ |
3538 | | "vmovdqu (%[uv_buf]),%%ymm3 \n" \ |
3539 | | "vmovdqu 0x20(%[uv_buf]),%%ymm1 \n" \ |
3540 | | "lea 0x40(%[uv_buf]),%[uv_buf] \n" \ |
3541 | | "vpsrlw $0x8,%%ymm3,%%ymm3 \n" \ |
3542 | | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" \ |
3543 | | "vpackuswb %%ymm1,%%ymm3,%%ymm3 \n" \ |
3544 | | "vpermq $0xd8,%%ymm3,%%ymm3 \n" \ |
3545 | | "vmovdqu (%[y_buf]),%%ymm4 \n" \ |
3546 | | "lea 0x20(%[y_buf]),%[y_buf] \n" |
3547 | | |
3548 | | // Read 8 YUY2 with 16 Y and upsample 8 UV to 16 UV. |
3549 | | // ymm6 kShuffleYUY2Y, |
3550 | | // ymm7 kShuffleYUY2UV |
3551 | | #define READYUY2_AVX2 \ |
3552 | | "vmovdqu (%[yuy2_buf]),%%ymm1 \n" \ |
3553 | | "vpshufb %%ymm6,%%ymm1,%%ymm4 \n" \ |
3554 | | "vpshufb %%ymm7,%%ymm1,%%ymm3 \n" \ |
3555 | | "lea 0x20(%[yuy2_buf]),%[yuy2_buf] \n" |
3556 | | |
3557 | | // Read 8 UYVY with 16 Y and upsample 8 UV to 16 UV. |
3558 | | // ymm6 kShuffleUYVYY, |
3559 | | // ymm7 kShuffleUYVYUV |
3560 | | #define READUYVY_AVX2 \ |
3561 | | "vmovdqu (%[uyvy_buf]),%%ymm1 \n" \ |
3562 | | "vpshufb %%ymm6,%%ymm1,%%ymm4 \n" \ |
3563 | | "vpshufb %%ymm7,%%ymm1,%%ymm3 \n" \ |
3564 | | "lea 0x20(%[uyvy_buf]),%[uyvy_buf] \n" |
3565 | | |
3566 | | #if defined(__x86_64__) |
3567 | | #define YUVTORGB_SETUP_AVX2(yuvconstants) \ |
3568 | | "vpcmpeqb %%ymm13,%%ymm13,%%ymm13 \n" \ |
3569 | | "vmovdqa (%[yuvconstants]),%%ymm8 \n" \ |
3570 | | "vpabsb %%ymm13,%%ymm13 \n" \ |
3571 | | "vmovdqa 32(%[yuvconstants]),%%ymm9 \n" \ |
3572 | | "vpsllw $7,%%ymm13,%%ymm13 \n" \ |
3573 | | "vmovdqa 64(%[yuvconstants]),%%ymm10 \n" \ |
3574 | | "vmovdqa 96(%[yuvconstants]),%%ymm11 \n" \ |
3575 | | "vmovdqa 128(%[yuvconstants]),%%ymm12 \n" |
3576 | | |
3577 | | #define YUVTORGB_SETUP_AVX512BW(yuvconstants) \ |
3578 | | "vpternlogd $0xff,%%zmm13,%%zmm13,%%zmm13 \n" \ |
3579 | | "vpbroadcastq (%[yuvconstants]),%%zmm8 \n" \ |
3580 | | "vpabsb %%zmm13,%%zmm13 \n" \ |
3581 | | "vpsllw $7,%%zmm13,%%zmm13 \n" \ |
3582 | | "vpbroadcastq 32(%[yuvconstants]),%%zmm9 \n" \ |
3583 | | "vpbroadcastq 64(%[yuvconstants]),%%zmm10 \n" \ |
3584 | | "vpbroadcastq 96(%[yuvconstants]),%%zmm11 \n" \ |
3585 | | "vpbroadcastq 128(%[yuvconstants]),%%zmm12 \n" \ |
3586 | | "vmovups (%[quadsplitperm]),%%zmm16 \n" \ |
3587 | | "vmovups (%[dquadsplitperm]),%%zmm17 \n" \ |
3588 | | "vmovups (%[unperm]),%%zmm18 \n" |
3589 | | |
3590 | | #define YUVTORGB16_AVX2(yuvconstants) \ |
3591 | | "vpsubb %%ymm13,%%ymm3,%%ymm3 \n" \ |
3592 | | "vpmulhuw %%ymm11,%%ymm4,%%ymm4 \n" \ |
3593 | | "vpmaddubsw %%ymm3,%%ymm8,%%ymm0 \n" \ |
3594 | | "vpmaddubsw %%ymm3,%%ymm9,%%ymm1 \n" \ |
3595 | | "vpmaddubsw %%ymm3,%%ymm10,%%ymm2 \n" \ |
3596 | | "vpaddw %%ymm4,%%ymm12,%%ymm4 \n" \ |
3597 | | "vpaddsw %%ymm4,%%ymm0,%%ymm0 \n" \ |
3598 | | "vpsubsw %%ymm1,%%ymm4,%%ymm1 \n" \ |
3599 | | "vpaddsw %%ymm4,%%ymm2,%%ymm2 \n" |
3600 | | |
3601 | | #define YUVTORGB16_AVX512BW(yuvconstants) \ |
3602 | | "vpsubb %%zmm13,%%zmm3,%%zmm3 \n" \ |
3603 | | "vpmulhuw %%zmm11,%%zmm4,%%zmm4 \n" \ |
3604 | | "vpmaddubsw %%zmm3,%%zmm8,%%zmm0 \n" \ |
3605 | | "vpmaddubsw %%zmm3,%%zmm9,%%zmm1 \n" \ |
3606 | | "vpmaddubsw %%zmm3,%%zmm10,%%zmm2 \n" \ |
3607 | | "vpaddw %%zmm4,%%zmm12,%%zmm4 \n" \ |
3608 | | "vpaddsw %%zmm4,%%zmm0,%%zmm0 \n" \ |
3609 | | "vpsubsw %%zmm1,%%zmm4,%%zmm1 \n" \ |
3610 | | "vpaddsw %%zmm4,%%zmm2,%%zmm2 \n" |
3611 | | |
3612 | | #define YUVTORGB_REGS_AVX2 "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", |
3613 | | #define YUVTORGB_REGS_AVX512BW \ |
3614 | | "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13", "xmm16", "xmm17", "xmm18", |
3615 | | |
3616 | | #else // Convert 16 pixels: 16 UV and 16 Y. |
3617 | | |
3618 | | #define YUVTORGB_SETUP_AVX2(yuvconstants) |
3619 | | #define YUVTORGB16_AVX2(yuvconstants) \ |
3620 | | "vpcmpeqb %%xmm0,%%xmm0,%%xmm0 \n" \ |
3621 | | "vpsllw $7,%%xmm0,%%xmm0 \n" \ |
3622 | | "vpbroadcastb %%xmm0,%%ymm0 \n" \ |
3623 | | "vpsubb %%ymm0,%%ymm3,%%ymm3 \n" \ |
3624 | | "vpmulhuw 96(%[yuvconstants]),%%ymm4,%%ymm4 \n" \ |
3625 | | "vmovdqa (%[yuvconstants]),%%ymm0 \n" \ |
3626 | | "vmovdqa 32(%[yuvconstants]),%%ymm1 \n" \ |
3627 | | "vmovdqa 64(%[yuvconstants]),%%ymm2 \n" \ |
3628 | | "vpmaddubsw %%ymm3,%%ymm0,%%ymm0 \n" \ |
3629 | | "vpmaddubsw %%ymm3,%%ymm1,%%ymm1 \n" \ |
3630 | | "vpmaddubsw %%ymm3,%%ymm2,%%ymm2 \n" \ |
3631 | | "vmovdqa 128(%[yuvconstants]),%%ymm3 \n" \ |
3632 | | "vpaddw %%ymm4,%%ymm3,%%ymm4 \n" \ |
3633 | | "vpaddsw %%ymm4,%%ymm0,%%ymm0 \n" \ |
3634 | | "vpsubsw %%ymm1,%%ymm4,%%ymm1 \n" \ |
3635 | | "vpaddsw %%ymm4,%%ymm2,%%ymm2 \n" |
3636 | | |
3637 | | #define YUVTORGB_REGS_AVX2 |
3638 | | #endif |
3639 | | |
3640 | | #define YUVTORGB_AVX2(yuvconstants) \ |
3641 | | YUVTORGB16_AVX2(yuvconstants) \ |
3642 | | "vpsraw $0x6,%%ymm0,%%ymm0 \n" \ |
3643 | | "vpsraw $0x6,%%ymm1,%%ymm1 \n" \ |
3644 | | "vpsraw $0x6,%%ymm2,%%ymm2 \n" \ |
3645 | | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" \ |
3646 | | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" \ |
3647 | | "vpackuswb %%ymm2,%%ymm2,%%ymm2 \n" |
3648 | | |
3649 | | #define YUVTORGB_AVX512BW(yuvconstants) \ |
3650 | | YUVTORGB16_AVX512BW(yuvconstants) \ |
3651 | | "vpsraw $0x6,%%zmm0,%%zmm0 \n" \ |
3652 | | "vpsraw $0x6,%%zmm1,%%zmm1 \n" \ |
3653 | | "vpsraw $0x6,%%zmm2,%%zmm2 \n" \ |
3654 | | "vpackuswb %%zmm0,%%zmm0,%%zmm0 \n" \ |
3655 | | "vpackuswb %%zmm1,%%zmm1,%%zmm1 \n" \ |
3656 | | "vpackuswb %%zmm2,%%zmm2,%%zmm2 \n" |
3657 | | |
3658 | | // Store 16 ARGB values. |
3659 | | #define STOREARGB_AVX2 \ |
3660 | | "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" \ |
3661 | | "vpermq $0xd8,%%ymm0,%%ymm0 \n" \ |
3662 | | "vpunpcklbw %%ymm5,%%ymm2,%%ymm2 \n" \ |
3663 | | "vpermq $0xd8,%%ymm2,%%ymm2 \n" \ |
3664 | | "vpunpcklwd %%ymm2,%%ymm0,%%ymm1 \n" \ |
3665 | | "vpunpckhwd %%ymm2,%%ymm0,%%ymm0 \n" \ |
3666 | | "vmovdqu %%ymm1,(%[dst_argb]) \n" \ |
3667 | | "vmovdqu %%ymm0,0x20(%[dst_argb]) \n" \ |
3668 | | "lea 0x40(%[dst_argb]), %[dst_argb] \n" |
3669 | | |
3670 | | // Store 32 ARGB values. |
3671 | | #define STOREARGB_AVX512BW \ |
3672 | | "vpunpcklbw %%zmm1,%%zmm0,%%zmm0 \n" \ |
3673 | | "vpermq %%zmm0,%%zmm18,%%zmm0 \n" \ |
3674 | | "vpunpcklbw %%zmm5,%%zmm2,%%zmm2 \n" \ |
3675 | | "vpermq %%zmm2,%%zmm18,%%zmm2 \n" \ |
3676 | | "vpunpcklwd %%zmm2,%%zmm0,%%zmm1 \n" \ |
3677 | | "vpunpckhwd %%zmm2,%%zmm0,%%zmm0 \n" \ |
3678 | | "vmovups %%zmm1,(%[dst_argb]) \n" \ |
3679 | | "vmovups %%zmm0,0x40(%[dst_argb]) \n" \ |
3680 | | "lea 0x80(%[dst_argb]), %[dst_argb] \n" |
3681 | | |
3682 | | // Store 16 AR30 values. |
3683 | | #define STOREAR30_AVX2 \ |
3684 | | "vpsraw $0x4,%%ymm0,%%ymm0 \n" \ |
3685 | | "vpsraw $0x4,%%ymm1,%%ymm1 \n" \ |
3686 | | "vpsraw $0x4,%%ymm2,%%ymm2 \n" \ |
3687 | | "vpminsw %%ymm7,%%ymm0,%%ymm0 \n" \ |
3688 | | "vpminsw %%ymm7,%%ymm1,%%ymm1 \n" \ |
3689 | | "vpminsw %%ymm7,%%ymm2,%%ymm2 \n" \ |
3690 | | "vpmaxsw %%ymm6,%%ymm0,%%ymm0 \n" \ |
3691 | | "vpmaxsw %%ymm6,%%ymm1,%%ymm1 \n" \ |
3692 | | "vpmaxsw %%ymm6,%%ymm2,%%ymm2 \n" \ |
3693 | | "vpsllw $0x4,%%ymm2,%%ymm2 \n" \ |
3694 | | "vpermq $0xd8,%%ymm0,%%ymm0 \n" \ |
3695 | | "vpermq $0xd8,%%ymm1,%%ymm1 \n" \ |
3696 | | "vpermq $0xd8,%%ymm2,%%ymm2 \n" \ |
3697 | | "vpunpckhwd %%ymm2,%%ymm0,%%ymm3 \n" \ |
3698 | | "vpunpcklwd %%ymm2,%%ymm0,%%ymm0 \n" \ |
3699 | | "vpunpckhwd %%ymm5,%%ymm1,%%ymm2 \n" \ |
3700 | | "vpunpcklwd %%ymm5,%%ymm1,%%ymm1 \n" \ |
3701 | | "vpslld $0xa,%%ymm1,%%ymm1 \n" \ |
3702 | | "vpslld $0xa,%%ymm2,%%ymm2 \n" \ |
3703 | | "vpor %%ymm1,%%ymm0,%%ymm0 \n" \ |
3704 | | "vpor %%ymm2,%%ymm3,%%ymm3 \n" \ |
3705 | | "vmovdqu %%ymm0,(%[dst_ar30]) \n" \ |
3706 | | "vmovdqu %%ymm3,0x20(%[dst_ar30]) \n" \ |
3707 | | "lea 0x40(%[dst_ar30]), %[dst_ar30] \n" |
3708 | | |
3709 | | #ifdef HAS_I444TOARGBROW_AVX2 |
3710 | | // 16 pixels |
3711 | | // 16 UV values with 16 Y producing 16 ARGB (64 bytes). |
3712 | | void OMITFP I444ToARGBRow_AVX2(const uint8_t* y_buf, |
3713 | | const uint8_t* u_buf, |
3714 | | const uint8_t* v_buf, |
3715 | | uint8_t* dst_argb, |
3716 | | const struct YuvConstants* yuvconstants, |
3717 | 17.4k | int width) { |
3718 | 17.4k | asm volatile ( |
3719 | 17.4k | YUVTORGB_SETUP_AVX2(yuvconstants) |
3720 | 17.4k | "sub %[u_buf],%[v_buf] \n" |
3721 | 17.4k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
3722 | | |
3723 | 17.4k | LABELALIGN |
3724 | 17.4k | "1: \n" |
3725 | 17.4k | READYUV444_AVX2 |
3726 | 17.4k | YUVTORGB_AVX2(yuvconstants) |
3727 | 17.4k | STOREARGB_AVX2 |
3728 | 17.4k | "sub $0x10,%[width] \n" |
3729 | 17.4k | "jg 1b \n" |
3730 | 17.4k | "vzeroupper \n" |
3731 | 17.4k | : [y_buf]"+r"(y_buf), // %[y_buf] |
3732 | 17.4k | [u_buf]"+r"(u_buf), // %[u_buf] |
3733 | 17.4k | [v_buf]"+r"(v_buf), // %[v_buf] |
3734 | 17.4k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
3735 | 17.4k | [width]"+rm"(width) // %[width] |
3736 | 17.4k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3737 | 17.4k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3738 | 17.4k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3739 | 17.4k | ); |
3740 | 17.4k | } |
3741 | | #endif // HAS_I444TOARGBROW_AVX2 |
3742 | | |
3743 | | #if defined(HAS_I422TOARGBROW_AVX2) |
3744 | | // 16 pixels |
3745 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
3746 | | void OMITFP I422ToARGBRow_AVX2(const uint8_t* y_buf, |
3747 | | const uint8_t* u_buf, |
3748 | | const uint8_t* v_buf, |
3749 | | uint8_t* dst_argb, |
3750 | | const struct YuvConstants* yuvconstants, |
3751 | 8.36k | int width) { |
3752 | 8.36k | asm volatile ( |
3753 | 8.36k | YUVTORGB_SETUP_AVX2(yuvconstants) |
3754 | 8.36k | "sub %[u_buf],%[v_buf] \n" |
3755 | 8.36k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
3756 | | |
3757 | 8.36k | LABELALIGN |
3758 | 8.36k | "1: \n" |
3759 | 8.36k | READYUV422_AVX2 |
3760 | 8.36k | YUVTORGB_AVX2(yuvconstants) |
3761 | 8.36k | STOREARGB_AVX2 |
3762 | 8.36k | "sub $0x10,%[width] \n" |
3763 | 8.36k | "jg 1b \n" |
3764 | | |
3765 | 8.36k | "vzeroupper \n" |
3766 | 8.36k | : [y_buf]"+r"(y_buf), // %[y_buf] |
3767 | 8.36k | [u_buf]"+r"(u_buf), // %[u_buf] |
3768 | 8.36k | [v_buf]"+r"(v_buf), // %[v_buf] |
3769 | 8.36k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
3770 | 8.36k | [width]"+rm"(width) // %[width] |
3771 | 8.36k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3772 | 8.36k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3773 | 8.36k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3774 | 8.36k | ); |
3775 | 8.36k | } |
3776 | | #endif // HAS_I422TOARGBROW_AVX2 |
3777 | | |
3778 | | #if defined(HAS_I422TOARGBROW_AVX512BW) |
3779 | | static const uint64_t kSplitQuadWords[8] = {0, 2, 2, 2, 1, 2, 2, 2}; |
3780 | | static const uint64_t kSplitDoubleQuadWords[8] = {0, 1, 4, 4, 2, 3, 4, 4}; |
3781 | | static const uint64_t kUnpermuteAVX512[8] = {0, 4, 1, 5, 2, 6, 3, 7}; |
3782 | | |
3783 | | // 32 pixels |
3784 | | // 16 UV values upsampled to 32 UV, mixed with 32 Y producing 32 ARGB (128 |
3785 | | // bytes). |
3786 | | void OMITFP I422ToARGBRow_AVX512BW(const uint8_t* y_buf, |
3787 | | const uint8_t* u_buf, |
3788 | | const uint8_t* v_buf, |
3789 | | uint8_t* dst_argb, |
3790 | | const struct YuvConstants* yuvconstants, |
3791 | 0 | int width) { |
3792 | 0 | asm volatile ( |
3793 | 0 | YUVTORGB_SETUP_AVX512BW(yuvconstants) |
3794 | 0 | "sub %[u_buf],%[v_buf] \n" |
3795 | 0 | "vpcmpeqb %%xmm5,%%xmm5,%%xmm5 \n" |
3796 | 0 | "vpbroadcastq %%xmm5,%%zmm5 \n" |
3797 | |
|
3798 | 0 | LABELALIGN |
3799 | 0 | "1: \n" |
3800 | 0 | READYUV422_AVX512BW |
3801 | 0 | YUVTORGB_AVX512BW(yuvconstants) |
3802 | 0 | STOREARGB_AVX512BW |
3803 | 0 | "sub $0x20,%[width] \n" |
3804 | 0 | "jg 1b \n" |
3805 | |
|
3806 | 0 | "vzeroupper \n" |
3807 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3808 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
3809 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
3810 | 0 | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
3811 | 0 | [width]"+rm"(width) // %[width] |
3812 | 0 | : [yuvconstants]"r"(yuvconstants), // %[yuvconstants] |
3813 | 0 | [quadsplitperm]"r"(kSplitQuadWords), // %[quadsplitperm] |
3814 | 0 | [dquadsplitperm]"r"(kSplitDoubleQuadWords), // %[dquadsplitperm] |
3815 | 0 | [unperm]"r"(kUnpermuteAVX512) // %[unperm] |
3816 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX512BW |
3817 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3818 | 0 | ); |
3819 | 0 | } |
3820 | | #endif // HAS_I422TOARGBROW_AVX512BW |
3821 | | |
3822 | | #if defined(HAS_I422TOAR30ROW_AVX2) |
3823 | | // 16 pixels |
3824 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes). |
3825 | | void OMITFP I422ToAR30Row_AVX2(const uint8_t* y_buf, |
3826 | | const uint8_t* u_buf, |
3827 | | const uint8_t* v_buf, |
3828 | | uint8_t* dst_ar30, |
3829 | | const struct YuvConstants* yuvconstants, |
3830 | 0 | int width) { |
3831 | 0 | asm volatile ( |
3832 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
3833 | 0 | "sub %[u_buf],%[v_buf] \n" |
3834 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
3835 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
3836 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
3837 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
3838 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
3839 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
3840 | |
|
3841 | 0 | LABELALIGN |
3842 | 0 | "1: \n" |
3843 | 0 | READYUV422_AVX2 |
3844 | 0 | YUVTORGB16_AVX2(yuvconstants) |
3845 | 0 | STOREAR30_AVX2 |
3846 | 0 | "sub $0x10,%[width] \n" |
3847 | 0 | "jg 1b \n" |
3848 | |
|
3849 | 0 | "vzeroupper \n" |
3850 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3851 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
3852 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
3853 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
3854 | 0 | [width]"+rm"(width) // %[width] |
3855 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3856 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3857 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
3858 | 0 | ); |
3859 | 0 | } |
3860 | | #endif // HAS_I422TOAR30ROW_AVX2 |
3861 | | |
3862 | | #if defined(HAS_I210TOARGBROW_AVX2) |
3863 | | // 16 pixels |
3864 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
3865 | | void OMITFP I210ToARGBRow_AVX2(const uint16_t* y_buf, |
3866 | | const uint16_t* u_buf, |
3867 | | const uint16_t* v_buf, |
3868 | | uint8_t* dst_argb, |
3869 | | const struct YuvConstants* yuvconstants, |
3870 | 4.26k | int width) { |
3871 | 4.26k | asm volatile ( |
3872 | 4.26k | YUVTORGB_SETUP_AVX2(yuvconstants) |
3873 | 4.26k | "sub %[u_buf],%[v_buf] \n" |
3874 | 4.26k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
3875 | | |
3876 | 4.26k | LABELALIGN |
3877 | 4.26k | "1: \n" |
3878 | 4.26k | READYUV210_AVX2 |
3879 | 4.26k | YUVTORGB_AVX2(yuvconstants) |
3880 | 4.26k | STOREARGB_AVX2 |
3881 | 4.26k | "sub $0x10,%[width] \n" |
3882 | 4.26k | "jg 1b \n" |
3883 | | |
3884 | 4.26k | "vzeroupper \n" |
3885 | 4.26k | : [y_buf]"+r"(y_buf), // %[y_buf] |
3886 | 4.26k | [u_buf]"+r"(u_buf), // %[u_buf] |
3887 | 4.26k | [v_buf]"+r"(v_buf), // %[v_buf] |
3888 | 4.26k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
3889 | 4.26k | [width]"+rm"(width) // %[width] |
3890 | 4.26k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3891 | 4.26k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3892 | 4.26k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3893 | 4.26k | ); |
3894 | 4.26k | } |
3895 | | #endif // HAS_I210TOARGBROW_AVX2 |
3896 | | |
3897 | | #if defined(HAS_I212TOARGBROW_AVX2) |
3898 | | // 16 pixels |
3899 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
3900 | | void OMITFP I212ToARGBRow_AVX2(const uint16_t* y_buf, |
3901 | | const uint16_t* u_buf, |
3902 | | const uint16_t* v_buf, |
3903 | | uint8_t* dst_argb, |
3904 | | const struct YuvConstants* yuvconstants, |
3905 | 5.78k | int width) { |
3906 | 5.78k | asm volatile ( |
3907 | 5.78k | YUVTORGB_SETUP_AVX2(yuvconstants) |
3908 | 5.78k | "sub %[u_buf],%[v_buf] \n" |
3909 | 5.78k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
3910 | | |
3911 | 5.78k | LABELALIGN |
3912 | 5.78k | "1: \n" |
3913 | 5.78k | READYUV212_AVX2 |
3914 | 5.78k | YUVTORGB_AVX2(yuvconstants) |
3915 | 5.78k | STOREARGB_AVX2 |
3916 | 5.78k | "sub $0x10,%[width] \n" |
3917 | 5.78k | "jg 1b \n" |
3918 | | |
3919 | 5.78k | "vzeroupper \n" |
3920 | 5.78k | : [y_buf]"+r"(y_buf), // %[y_buf] |
3921 | 5.78k | [u_buf]"+r"(u_buf), // %[u_buf] |
3922 | 5.78k | [v_buf]"+r"(v_buf), // %[v_buf] |
3923 | 5.78k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
3924 | 5.78k | [width]"+rm"(width) // %[width] |
3925 | 5.78k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3926 | 5.78k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3927 | 5.78k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
3928 | 5.78k | ); |
3929 | 5.78k | } |
3930 | | #endif // HAS_I212TOARGBROW_AVX2 |
3931 | | |
3932 | | #if defined(HAS_I210TOAR30ROW_AVX2) |
3933 | | // 16 pixels |
3934 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes). |
3935 | | void OMITFP I210ToAR30Row_AVX2(const uint16_t* y_buf, |
3936 | | const uint16_t* u_buf, |
3937 | | const uint16_t* v_buf, |
3938 | | uint8_t* dst_ar30, |
3939 | | const struct YuvConstants* yuvconstants, |
3940 | 0 | int width) { |
3941 | 0 | asm volatile ( |
3942 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
3943 | 0 | "sub %[u_buf],%[v_buf] \n" |
3944 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
3945 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
3946 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
3947 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
3948 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
3949 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
3950 | |
|
3951 | 0 | LABELALIGN |
3952 | 0 | "1: \n" |
3953 | 0 | READYUV210_AVX2 |
3954 | 0 | YUVTORGB16_AVX2(yuvconstants) |
3955 | 0 | STOREAR30_AVX2 |
3956 | 0 | "sub $0x10,%[width] \n" |
3957 | 0 | "jg 1b \n" |
3958 | |
|
3959 | 0 | "vzeroupper \n" |
3960 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
3961 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
3962 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
3963 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
3964 | 0 | [width]"+rm"(width) // %[width] |
3965 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
3966 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
3967 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
3968 | 0 | ); |
3969 | 0 | } |
3970 | | #endif // HAS_I210TOAR30ROW_AVX2 |
3971 | | |
3972 | | #if defined(HAS_I212TOAR30ROW_AVX2) |
3973 | | // 16 pixels |
3974 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 AR30 (64 bytes). |
3975 | | void OMITFP I212ToAR30Row_AVX2(const uint16_t* y_buf, |
3976 | | const uint16_t* u_buf, |
3977 | | const uint16_t* v_buf, |
3978 | | uint8_t* dst_ar30, |
3979 | | const struct YuvConstants* yuvconstants, |
3980 | 0 | int width) { |
3981 | 0 | asm volatile ( |
3982 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
3983 | 0 | "sub %[u_buf],%[v_buf] \n" |
3984 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
3985 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
3986 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
3987 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
3988 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
3989 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
3990 | |
|
3991 | 0 | LABELALIGN |
3992 | 0 | "1: \n" |
3993 | 0 | READYUV212_AVX2 |
3994 | 0 | YUVTORGB16_AVX2(yuvconstants) |
3995 | 0 | STOREAR30_AVX2 |
3996 | 0 | "sub $0x10,%[width] \n" |
3997 | 0 | "jg 1b \n" |
3998 | |
|
3999 | 0 | "vzeroupper \n" |
4000 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
4001 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
4002 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
4003 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
4004 | 0 | [width]"+rm"(width) // %[width] |
4005 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4006 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4007 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
4008 | 0 | ); |
4009 | 0 | } |
4010 | | #endif // HAS_I212TOAR30ROW_AVX2 |
4011 | | |
4012 | | #if defined(HAS_I410TOARGBROW_AVX2) |
4013 | | // 16 pixels |
4014 | | // 16 UV values with 16 Y producing 16 ARGB (64 bytes). |
4015 | | void OMITFP I410ToARGBRow_AVX2(const uint16_t* y_buf, |
4016 | | const uint16_t* u_buf, |
4017 | | const uint16_t* v_buf, |
4018 | | uint8_t* dst_argb, |
4019 | | const struct YuvConstants* yuvconstants, |
4020 | 9.09k | int width) { |
4021 | 9.09k | asm volatile ( |
4022 | 9.09k | YUVTORGB_SETUP_AVX2(yuvconstants) |
4023 | 9.09k | "sub %[u_buf],%[v_buf] \n" |
4024 | 9.09k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4025 | | |
4026 | 9.09k | LABELALIGN |
4027 | 9.09k | "1: \n" |
4028 | 9.09k | READYUV410_AVX2 |
4029 | 9.09k | YUVTORGB_AVX2(yuvconstants) |
4030 | 9.09k | STOREARGB_AVX2 |
4031 | 9.09k | "sub $0x10,%[width] \n" |
4032 | 9.09k | "jg 1b \n" |
4033 | 9.09k | "vzeroupper \n" |
4034 | | |
4035 | 9.09k | : [y_buf]"+r"(y_buf), // %[y_buf] |
4036 | 9.09k | [u_buf]"+r"(u_buf), // %[u_buf] |
4037 | 9.09k | [v_buf]"+r"(v_buf), // %[v_buf] |
4038 | 9.09k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
4039 | 9.09k | [width]"+rm"(width) // %[width] |
4040 | 9.09k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4041 | 9.09k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4042 | 9.09k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
4043 | 9.09k | ); |
4044 | 9.09k | } |
4045 | | #endif // HAS_I410TOARGBROW_AVX2 |
4046 | | |
4047 | | #if defined(HAS_I210ALPHATOARGBROW_AVX2) |
4048 | | // 16 pixels |
4049 | | // 8 UV, 16 Y and 16 A producing 16 ARGB (64 bytes). |
4050 | | void OMITFP I210AlphaToARGBRow_AVX2(const uint16_t* y_buf, |
4051 | | const uint16_t* u_buf, |
4052 | | const uint16_t* v_buf, |
4053 | | const uint16_t* a_buf, |
4054 | | uint8_t* dst_argb, |
4055 | | const struct YuvConstants* yuvconstants, |
4056 | 2.93k | int width) { |
4057 | 2.93k | asm volatile(YUVTORGB_SETUP_AVX2( |
4058 | 2.93k | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
4059 | | |
4060 | 2.93k | LABELALIGN "1: \n" READYUVA210_AVX2 YUVTORGB_AVX2( |
4061 | 2.93k | yuvconstants) STOREARGB_AVX2 |
4062 | 2.93k | "subl $0x10,%[width] \n" |
4063 | 2.93k | "jg 1b \n" |
4064 | 2.93k | "vzeroupper \n" |
4065 | | |
4066 | 2.93k | : [y_buf] "+r"(y_buf), // %[y_buf] |
4067 | 2.93k | [u_buf] "+r"(u_buf), // %[u_buf] |
4068 | 2.93k | [v_buf] "+r"(v_buf), // %[v_buf] |
4069 | 2.93k | [a_buf] "+r"(a_buf), // %[a_buf] |
4070 | 2.93k | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4071 | | #if defined(__i386__) |
4072 | | [width] "+m"(width) // %[width] |
4073 | | #else |
4074 | 2.93k | [width] "+rm"(width) // %[width] |
4075 | 2.93k | #endif |
4076 | 2.93k | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4077 | 2.93k | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", |
4078 | 2.93k | "xmm3", "xmm4", "xmm5"); |
4079 | 2.93k | } |
4080 | | #endif // HAS_I210TOARGBROW_AVX2 |
4081 | | |
4082 | | #if defined(HAS_I410ALPHATOARGBROW_AVX2) |
4083 | | // 16 pixels |
4084 | | // 16 UV, 16 Y and 16 A producing 16 ARGB (64 bytes). |
4085 | | void OMITFP I410AlphaToARGBRow_AVX2(const uint16_t* y_buf, |
4086 | | const uint16_t* u_buf, |
4087 | | const uint16_t* v_buf, |
4088 | | const uint16_t* a_buf, |
4089 | | uint8_t* dst_argb, |
4090 | | const struct YuvConstants* yuvconstants, |
4091 | 4.14k | int width) { |
4092 | 4.14k | asm volatile(YUVTORGB_SETUP_AVX2( |
4093 | 4.14k | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
4094 | | |
4095 | 4.14k | LABELALIGN "1: \n" READYUVA410_AVX2 YUVTORGB_AVX2( |
4096 | 4.14k | yuvconstants) STOREARGB_AVX2 |
4097 | 4.14k | "subl $0x10,%[width] \n" |
4098 | 4.14k | "jg 1b \n" |
4099 | 4.14k | "vzeroupper \n" |
4100 | | |
4101 | 4.14k | : [y_buf] "+r"(y_buf), // %[y_buf] |
4102 | 4.14k | [u_buf] "+r"(u_buf), // %[u_buf] |
4103 | 4.14k | [v_buf] "+r"(v_buf), // %[v_buf] |
4104 | 4.14k | [a_buf] "+r"(a_buf), // %[a_buf] |
4105 | 4.14k | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4106 | | #if defined(__i386__) |
4107 | | [width] "+m"(width) // %[width] |
4108 | | #else |
4109 | 4.14k | [width] "+rm"(width) // %[width] |
4110 | 4.14k | #endif |
4111 | 4.14k | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4112 | 4.14k | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", |
4113 | 4.14k | "xmm3", "xmm4", "xmm5"); |
4114 | 4.14k | } |
4115 | | #endif // HAS_I410TOARGBROW_AVX2 |
4116 | | |
4117 | | #if defined(HAS_I410TOAR30ROW_AVX2) |
4118 | | // 16 pixels |
4119 | | // 16 UV values with 16 Y producing 16 AR30 (64 bytes). |
4120 | | void OMITFP I410ToAR30Row_AVX2(const uint16_t* y_buf, |
4121 | | const uint16_t* u_buf, |
4122 | | const uint16_t* v_buf, |
4123 | | uint8_t* dst_ar30, |
4124 | | const struct YuvConstants* yuvconstants, |
4125 | 0 | int width) { |
4126 | 0 | asm volatile ( |
4127 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
4128 | 0 | "sub %[u_buf],%[v_buf] \n" |
4129 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
4130 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
4131 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
4132 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
4133 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
4134 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
4135 | |
|
4136 | 0 | LABELALIGN |
4137 | 0 | "1: \n" |
4138 | 0 | READYUV410_AVX2 |
4139 | 0 | YUVTORGB16_AVX2(yuvconstants) |
4140 | 0 | STOREAR30_AVX2 |
4141 | 0 | "sub $0x10,%[width] \n" |
4142 | 0 | "jg 1b \n" |
4143 | |
|
4144 | 0 | "vzeroupper \n" |
4145 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
4146 | 0 | [u_buf]"+r"(u_buf), // %[u_buf] |
4147 | 0 | [v_buf]"+r"(v_buf), // %[v_buf] |
4148 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
4149 | 0 | [width]"+rm"(width) // %[width] |
4150 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4151 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4152 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
4153 | 0 | ); |
4154 | 0 | } |
4155 | | #endif // HAS_I410TOAR30ROW_AVX2 |
4156 | | |
4157 | | #if defined(HAS_I444ALPHATOARGBROW_AVX2) |
4158 | | // 16 pixels |
4159 | | // 16 UV values with 16 Y and 16 A producing 16 ARGB. |
4160 | | void OMITFP I444AlphaToARGBRow_AVX2(const uint8_t* y_buf, |
4161 | | const uint8_t* u_buf, |
4162 | | const uint8_t* v_buf, |
4163 | | const uint8_t* a_buf, |
4164 | | uint8_t* dst_argb, |
4165 | | const struct YuvConstants* yuvconstants, |
4166 | 11.3k | int width) { |
4167 | 11.3k | asm volatile(YUVTORGB_SETUP_AVX2( |
4168 | 11.3k | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
4169 | | |
4170 | 11.3k | LABELALIGN "1: \n" READYUVA444_AVX2 YUVTORGB_AVX2( |
4171 | 11.3k | yuvconstants) STOREARGB_AVX2 |
4172 | 11.3k | "subl $0x10,%[width] \n" |
4173 | 11.3k | "jg 1b \n" |
4174 | 11.3k | "vzeroupper \n" |
4175 | 11.3k | : [y_buf] "+r"(y_buf), // %[y_buf] |
4176 | 11.3k | [u_buf] "+r"(u_buf), // %[u_buf] |
4177 | 11.3k | [v_buf] "+r"(v_buf), // %[v_buf] |
4178 | 11.3k | [a_buf] "+r"(a_buf), // %[a_buf] |
4179 | 11.3k | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4180 | | #if defined(__i386__) |
4181 | | [width] "+m"(width) // %[width] |
4182 | | #else |
4183 | 11.3k | [width] "+rm"(width) // %[width] |
4184 | 11.3k | #endif |
4185 | 11.3k | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4186 | 11.3k | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", |
4187 | 11.3k | "xmm3", "xmm4", "xmm5"); |
4188 | 11.3k | } |
4189 | | #endif // HAS_I444ALPHATOARGBROW_AVX2 |
4190 | | |
4191 | | #if defined(HAS_I422ALPHATOARGBROW_AVX2) |
4192 | | // 16 pixels |
4193 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y and 16 A producing 16 ARGB. |
4194 | | void OMITFP I422AlphaToARGBRow_AVX2(const uint8_t* y_buf, |
4195 | | const uint8_t* u_buf, |
4196 | | const uint8_t* v_buf, |
4197 | | const uint8_t* a_buf, |
4198 | | uint8_t* dst_argb, |
4199 | | const struct YuvConstants* yuvconstants, |
4200 | 4.04k | int width) { |
4201 | 4.04k | asm volatile(YUVTORGB_SETUP_AVX2( |
4202 | 4.04k | yuvconstants) "sub %[u_buf],%[v_buf] \n" |
4203 | | |
4204 | 4.04k | LABELALIGN "1: \n" READYUVA422_AVX2 YUVTORGB_AVX2( |
4205 | 4.04k | yuvconstants) STOREARGB_AVX2 |
4206 | 4.04k | "subl $0x10,%[width] \n" |
4207 | 4.04k | "jg 1b \n" |
4208 | 4.04k | "vzeroupper \n" |
4209 | 4.04k | : [y_buf] "+r"(y_buf), // %[y_buf] |
4210 | 4.04k | [u_buf] "+r"(u_buf), // %[u_buf] |
4211 | 4.04k | [v_buf] "+r"(v_buf), // %[v_buf] |
4212 | 4.04k | [a_buf] "+r"(a_buf), // %[a_buf] |
4213 | 4.04k | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4214 | | #if defined(__i386__) |
4215 | | [width] "+m"(width) // %[width] |
4216 | | #else |
4217 | 4.04k | [width] "+rm"(width) // %[width] |
4218 | 4.04k | #endif |
4219 | 4.04k | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4220 | 4.04k | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", |
4221 | 4.04k | "xmm3", "xmm4", "xmm5"); |
4222 | 4.04k | } |
4223 | | #endif // HAS_I422ALPHATOARGBROW_AVX2 |
4224 | | |
4225 | | #if defined(HAS_I422TORGBAROW_AVX2) |
4226 | | // 16 pixels |
4227 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 RGBA (64 bytes). |
4228 | | void OMITFP I422ToRGBARow_AVX2(const uint8_t* y_buf, |
4229 | | const uint8_t* u_buf, |
4230 | | const uint8_t* v_buf, |
4231 | | uint8_t* dst_argb, |
4232 | | const struct YuvConstants* yuvconstants, |
4233 | 3.91k | int width) { |
4234 | 3.91k | asm volatile ( |
4235 | 3.91k | YUVTORGB_SETUP_AVX2(yuvconstants) |
4236 | 3.91k | "sub %[u_buf],%[v_buf] \n" |
4237 | 3.91k | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4238 | | |
4239 | 3.91k | LABELALIGN |
4240 | 3.91k | "1: \n" |
4241 | 3.91k | READYUV422_AVX2 |
4242 | 3.91k | YUVTORGB_AVX2(yuvconstants) |
4243 | | |
4244 | | // Step 3: Weave into RGBA |
4245 | 3.91k | "vpunpcklbw %%ymm2,%%ymm1,%%ymm1 \n" |
4246 | 3.91k | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
4247 | 3.91k | "vpunpcklbw %%ymm0,%%ymm5,%%ymm2 \n" |
4248 | 3.91k | "vpermq $0xd8,%%ymm2,%%ymm2 \n" |
4249 | 3.91k | "vpunpcklwd %%ymm1,%%ymm2,%%ymm0 \n" |
4250 | 3.91k | "vpunpckhwd %%ymm1,%%ymm2,%%ymm1 \n" |
4251 | 3.91k | "vmovdqu %%ymm0,(%[dst_argb]) \n" |
4252 | 3.91k | "vmovdqu %%ymm1,0x20(%[dst_argb]) \n" |
4253 | 3.91k | "lea 0x40(%[dst_argb]),%[dst_argb] \n" |
4254 | 3.91k | "sub $0x10,%[width] \n" |
4255 | 3.91k | "jg 1b \n" |
4256 | 3.91k | "vzeroupper \n" |
4257 | 3.91k | : [y_buf]"+r"(y_buf), // %[y_buf] |
4258 | 3.91k | [u_buf]"+r"(u_buf), // %[u_buf] |
4259 | 3.91k | [v_buf]"+r"(v_buf), // %[v_buf] |
4260 | 3.91k | [dst_argb]"+r"(dst_argb), // %[dst_argb] |
4261 | 3.91k | [width]"+rm"(width) // %[width] |
4262 | 3.91k | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4263 | 3.91k | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4264 | 3.91k | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5" |
4265 | 3.91k | ); |
4266 | 3.91k | } |
4267 | | #endif // HAS_I422TORGBAROW_AVX2 |
4268 | | |
4269 | | #if defined(HAS_NV12TOARGBROW_AVX2) |
4270 | | // 16 pixels. |
4271 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
4272 | | void OMITFP NV12ToARGBRow_AVX2(const uint8_t* y_buf, |
4273 | | const uint8_t* uv_buf, |
4274 | | uint8_t* dst_argb, |
4275 | | const struct YuvConstants* yuvconstants, |
4276 | 0 | int width) { |
4277 | 0 | asm volatile(YUVTORGB_SETUP_AVX2( |
4278 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4279 | |
|
4280 | 0 | LABELALIGN "1: \n" READNV12_AVX2 YUVTORGB_AVX2( |
4281 | 0 | yuvconstants) STOREARGB_AVX2 |
4282 | 0 | "sub $0x10,%[width] \n" |
4283 | 0 | "jg 1b \n" |
4284 | 0 | "vzeroupper \n" |
4285 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
4286 | 0 | [uv_buf] "+r"(uv_buf), // %[uv_buf] |
4287 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4288 | 0 | [width] "+rm"(width) // %[width] |
4289 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4290 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm0", "xmm1", |
4291 | 0 | "xmm2", "xmm3", "xmm4", "xmm5"); |
4292 | 0 | } |
4293 | | #endif // HAS_NV12TOARGBROW_AVX2 |
4294 | | |
4295 | | #if defined(HAS_NV21TOARGBROW_AVX2) |
4296 | | // 16 pixels. |
4297 | | // 8 VU values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
4298 | | void OMITFP NV21ToARGBRow_AVX2(const uint8_t* y_buf, |
4299 | | const uint8_t* vu_buf, |
4300 | | uint8_t* dst_argb, |
4301 | | const struct YuvConstants* yuvconstants, |
4302 | 0 | int width) { |
4303 | 0 | asm volatile(YUVTORGB_SETUP_AVX2( |
4304 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4305 | |
|
4306 | 0 | LABELALIGN "1: \n" READNV21_AVX2 YUVTORGB_AVX2( |
4307 | 0 | yuvconstants) STOREARGB_AVX2 |
4308 | 0 | "sub $0x10,%[width] \n" |
4309 | 0 | "jg 1b \n" |
4310 | 0 | "vzeroupper \n" |
4311 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
4312 | 0 | [vu_buf] "+r"(vu_buf), // %[vu_buf] |
4313 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4314 | 0 | [width] "+rm"(width) // %[width] |
4315 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
4316 | 0 | [kShuffleNV21] "m"(kShuffleNV21) |
4317 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm0", "xmm1", |
4318 | 0 | "xmm2", "xmm3", "xmm4", "xmm5"); |
4319 | 0 | } |
4320 | | #endif // HAS_NV21TOARGBROW_AVX2 |
4321 | | |
4322 | | #if defined(HAS_YUY2TOARGBROW_AVX2) |
4323 | | // 16 pixels. |
4324 | | // 8 YUY2 values with 16 Y and 8 UV producing 16 ARGB (64 bytes). |
4325 | | void OMITFP YUY2ToARGBRow_AVX2(const uint8_t* yuy2_buf, |
4326 | | uint8_t* dst_argb, |
4327 | | const struct YuvConstants* yuvconstants, |
4328 | 0 | int width) { |
4329 | 0 | asm volatile( |
4330 | 0 | "vbroadcasti128 %[kShuffleYUY2Y],%%ymm6 \n" |
4331 | 0 | "vbroadcasti128 %[kShuffleYUY2UV],%%ymm7 \n" YUVTORGB_SETUP_AVX2( |
4332 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4333 | |
|
4334 | 0 | LABELALIGN "1: \n" READYUY2_AVX2 YUVTORGB_AVX2(yuvconstants) |
4335 | 0 | STOREARGB_AVX2 |
4336 | 0 | "sub $0x10,%[width] \n" |
4337 | 0 | "jg 1b \n" |
4338 | 0 | "vzeroupper \n" |
4339 | 0 | : [yuy2_buf] "+r"(yuy2_buf), // %[yuy2_buf] |
4340 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4341 | 0 | [width] "+rm"(width) // %[width] |
4342 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
4343 | 0 | [kShuffleYUY2Y] "m"(kShuffleYUY2Y), [kShuffleYUY2UV] "m"(kShuffleYUY2UV) |
4344 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm0", "xmm1", "xmm2", |
4345 | 0 | "xmm3", "xmm4", "xmm5", "xmm6", "xmm7"); |
4346 | 0 | } |
4347 | | #endif // HAS_YUY2TOARGBROW_AVX2 |
4348 | | |
4349 | | #if defined(HAS_UYVYTOARGBROW_AVX2) |
4350 | | // 16 pixels. |
4351 | | // 8 UYVY values with 16 Y and 8 UV producing 16 ARGB (64 bytes). |
4352 | | void OMITFP UYVYToARGBRow_AVX2(const uint8_t* uyvy_buf, |
4353 | | uint8_t* dst_argb, |
4354 | | const struct YuvConstants* yuvconstants, |
4355 | 0 | int width) { |
4356 | 0 | asm volatile( |
4357 | 0 | "vbroadcasti128 %[kShuffleUYVYY],%%ymm6 \n" |
4358 | 0 | "vbroadcasti128 %[kShuffleUYVYUV],%%ymm7 \n" YUVTORGB_SETUP_AVX2( |
4359 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4360 | |
|
4361 | 0 | LABELALIGN "1: \n" READUYVY_AVX2 YUVTORGB_AVX2(yuvconstants) |
4362 | 0 | STOREARGB_AVX2 |
4363 | 0 | "sub $0x10,%[width] \n" |
4364 | 0 | "jg 1b \n" |
4365 | 0 | "vzeroupper \n" |
4366 | 0 | : [uyvy_buf] "+r"(uyvy_buf), // %[uyvy_buf] |
4367 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4368 | 0 | [width] "+rm"(width) // %[width] |
4369 | 0 | : [yuvconstants] "r"(yuvconstants), // %[yuvconstants] |
4370 | 0 | [kShuffleUYVYY] "m"(kShuffleUYVYY), [kShuffleUYVYUV] "m"(kShuffleUYVYUV) |
4371 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm1", "xmm2", "xmm3", |
4372 | 0 | "xmm4", "xmm5", "xmm6", "xmm7"); |
4373 | 0 | } |
4374 | | #endif // HAS_UYVYTOARGBROW_AVX2 |
4375 | | |
4376 | | #if defined(HAS_P210TOARGBROW_AVX2) |
4377 | | // 16 pixels. |
4378 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
4379 | | void OMITFP P210ToARGBRow_AVX2(const uint16_t* y_buf, |
4380 | | const uint16_t* uv_buf, |
4381 | | uint8_t* dst_argb, |
4382 | | const struct YuvConstants* yuvconstants, |
4383 | 0 | int width) { |
4384 | 0 | asm volatile(YUVTORGB_SETUP_AVX2( |
4385 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4386 | |
|
4387 | 0 | LABELALIGN "1: \n" READP210_AVX2 YUVTORGB_AVX2( |
4388 | 0 | yuvconstants) STOREARGB_AVX2 |
4389 | 0 | "sub $0x10,%[width] \n" |
4390 | 0 | "jg 1b \n" |
4391 | 0 | "vzeroupper \n" |
4392 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
4393 | 0 | [uv_buf] "+r"(uv_buf), // %[uv_buf] |
4394 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4395 | 0 | [width] "+rm"(width) // %[width] |
4396 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4397 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm0", "xmm1", |
4398 | 0 | "xmm2", "xmm3", "xmm4", "xmm5"); |
4399 | 0 | } |
4400 | | #endif // HAS_P210TOARGBROW_AVX2 |
4401 | | |
4402 | | #if defined(HAS_P410TOARGBROW_AVX2) |
4403 | | // 16 pixels. |
4404 | | // 8 UV values upsampled to 16 UV, mixed with 16 Y producing 16 ARGB (64 bytes). |
4405 | | void OMITFP P410ToARGBRow_AVX2(const uint16_t* y_buf, |
4406 | | const uint16_t* uv_buf, |
4407 | | uint8_t* dst_argb, |
4408 | | const struct YuvConstants* yuvconstants, |
4409 | 0 | int width) { |
4410 | 0 | asm volatile(YUVTORGB_SETUP_AVX2( |
4411 | 0 | yuvconstants) "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4412 | |
|
4413 | 0 | LABELALIGN "1: \n" READP410_AVX2 YUVTORGB_AVX2( |
4414 | 0 | yuvconstants) STOREARGB_AVX2 |
4415 | 0 | "sub $0x10,%[width] \n" |
4416 | 0 | "jg 1b \n" |
4417 | 0 | "vzeroupper \n" |
4418 | 0 | : [y_buf] "+r"(y_buf), // %[y_buf] |
4419 | 0 | [uv_buf] "+r"(uv_buf), // %[uv_buf] |
4420 | 0 | [dst_argb] "+r"(dst_argb), // %[dst_argb] |
4421 | 0 | [width] "+rm"(width) // %[width] |
4422 | 0 | : [yuvconstants] "r"(yuvconstants) // %[yuvconstants] |
4423 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 "xmm0", "xmm0", "xmm1", |
4424 | 0 | "xmm2", "xmm3", "xmm4", "xmm5"); |
4425 | 0 | } |
4426 | | #endif // HAS_P410TOARGBROW_AVX2 |
4427 | | |
4428 | | #if defined(HAS_P210TOAR30ROW_AVX2) |
4429 | | // 16 pixels |
4430 | | // 16 UV values with 16 Y producing 16 AR30 (64 bytes). |
4431 | | void OMITFP P210ToAR30Row_AVX2(const uint16_t* y_buf, |
4432 | | const uint16_t* uv_buf, |
4433 | | uint8_t* dst_ar30, |
4434 | | const struct YuvConstants* yuvconstants, |
4435 | 0 | int width) { |
4436 | 0 | asm volatile ( |
4437 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
4438 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
4439 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
4440 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
4441 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
4442 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
4443 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
4444 | |
|
4445 | 0 | LABELALIGN |
4446 | 0 | "1: \n" |
4447 | 0 | READP210_AVX2 |
4448 | 0 | YUVTORGB16_AVX2(yuvconstants) |
4449 | 0 | STOREAR30_AVX2 |
4450 | 0 | "sub $0x10,%[width] \n" |
4451 | 0 | "jg 1b \n" |
4452 | |
|
4453 | 0 | "vzeroupper \n" |
4454 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
4455 | 0 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
4456 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
4457 | 0 | [width]"+rm"(width) // %[width] |
4458 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4459 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4460 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
4461 | 0 | ); |
4462 | 0 | } |
4463 | | #endif // HAS_P210TOAR30ROW_AVX2 |
4464 | | |
4465 | | #if defined(HAS_P410TOAR30ROW_AVX2) |
4466 | | // 16 pixels |
4467 | | // 16 UV values with 16 Y producing 16 AR30 (64 bytes). |
4468 | | void OMITFP P410ToAR30Row_AVX2(const uint16_t* y_buf, |
4469 | | const uint16_t* uv_buf, |
4470 | | uint8_t* dst_ar30, |
4471 | | const struct YuvConstants* yuvconstants, |
4472 | 0 | int width) { |
4473 | 0 | asm volatile ( |
4474 | 0 | YUVTORGB_SETUP_AVX2(yuvconstants) |
4475 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
4476 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
4477 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
4478 | 0 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" // 0 for min |
4479 | 0 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" // 1023 for max |
4480 | 0 | "vpsrlw $6,%%ymm7,%%ymm7 \n" |
4481 | |
|
4482 | 0 | LABELALIGN |
4483 | 0 | "1: \n" |
4484 | 0 | READP410_AVX2 |
4485 | 0 | YUVTORGB16_AVX2(yuvconstants) |
4486 | 0 | STOREAR30_AVX2 |
4487 | 0 | "sub $0x10,%[width] \n" |
4488 | 0 | "jg 1b \n" |
4489 | |
|
4490 | 0 | "vzeroupper \n" |
4491 | 0 | : [y_buf]"+r"(y_buf), // %[y_buf] |
4492 | 0 | [uv_buf]"+r"(uv_buf), // %[uv_buf] |
4493 | 0 | [dst_ar30]"+r"(dst_ar30), // %[dst_ar30] |
4494 | 0 | [width]"+rm"(width) // %[width] |
4495 | 0 | : [yuvconstants]"r"(yuvconstants) // %[yuvconstants] |
4496 | 0 | : "memory", "cc", YUVTORGB_REGS_AVX2 |
4497 | 0 | "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7" |
4498 | 0 | ); |
4499 | 0 | } |
4500 | | #endif // HAS_P410TOAR30ROW_AVX2 |
4501 | | |
4502 | | #ifdef HAS_I400TOARGBROW_SSE2 |
4503 | | void I400ToARGBRow_SSE2(const uint8_t* y_buf, |
4504 | | uint8_t* dst_argb, |
4505 | | const struct YuvConstants* yuvconstants, |
4506 | 0 | int width) { |
4507 | 0 | asm volatile( |
4508 | 0 | "movdqa 96(%3),%%xmm2 \n" // yg = 18997 = 1.164 |
4509 | 0 | "movdqa 128(%3),%%xmm3 \n" // ygb = 1160 = 1.164 * 16 |
4510 | 0 | "pcmpeqb %%xmm4,%%xmm4 \n" // 0xff000000 |
4511 | 0 | "pslld $0x18,%%xmm4 \n" |
4512 | |
|
4513 | 0 | LABELALIGN |
4514 | 0 | "1: \n" |
4515 | | // Step 1: Scale Y contribution to 8 G values. G = (y - 16) * 1.164 |
4516 | 0 | "movq (%0),%%xmm0 \n" |
4517 | 0 | "lea 0x8(%0),%0 \n" |
4518 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
4519 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
4520 | 0 | "paddsw %%xmm3,%%xmm0 \n" |
4521 | 0 | "psraw $6, %%xmm0 \n" |
4522 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
4523 | | |
4524 | | // Step 2: Weave into ARGB |
4525 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
4526 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
4527 | 0 | "punpcklwd %%xmm0,%%xmm0 \n" |
4528 | 0 | "punpckhwd %%xmm1,%%xmm1 \n" |
4529 | 0 | "por %%xmm4,%%xmm0 \n" |
4530 | 0 | "por %%xmm4,%%xmm1 \n" |
4531 | 0 | "movdqu %%xmm0,(%1) \n" |
4532 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
4533 | 0 | "lea 0x20(%1),%1 \n" |
4534 | |
|
4535 | 0 | "sub $0x8,%2 \n" |
4536 | 0 | "jg 1b \n" |
4537 | 0 | : "+r"(y_buf), // %0 |
4538 | 0 | "+r"(dst_argb), // %1 |
4539 | 0 | "+rm"(width) // %2 |
4540 | 0 | : "r"(yuvconstants) // %3 |
4541 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
4542 | 0 | } |
4543 | | #endif // HAS_I400TOARGBROW_SSE2 |
4544 | | |
4545 | | #ifdef HAS_I400TOARGBROW_AVX2 |
4546 | | // 16 pixels of Y converted to 16 pixels of ARGB (64 bytes). |
4547 | | // note: vpunpcklbw mutates and vpackuswb unmutates. |
4548 | | void I400ToARGBRow_AVX2(const uint8_t* y_buf, |
4549 | | uint8_t* dst_argb, |
4550 | | const struct YuvConstants* yuvconstants, |
4551 | 439 | int width) { |
4552 | 439 | asm volatile( |
4553 | 439 | "vmovdqa 96(%3),%%ymm2 \n" // yg = 18997 = 1.164 |
4554 | 439 | "vmovdqa 128(%3),%%ymm3 \n" // ygb = -1160 = 1.164*16 |
4555 | 439 | "vpcmpeqb %%ymm4,%%ymm4,%%ymm4 \n" // 0xff000000 |
4556 | 439 | "vpslld $0x18,%%ymm4,%%ymm4 \n" |
4557 | | |
4558 | 439 | LABELALIGN |
4559 | 439 | "1: \n" |
4560 | | // Step 1: Scale Y contribution to 16 G values. G = (y - 16) * 1.164 |
4561 | 439 | "vmovdqu (%0),%%xmm0 \n" |
4562 | 439 | "lea 0x10(%0),%0 \n" |
4563 | 439 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
4564 | 439 | "vpunpcklbw %%ymm0,%%ymm0,%%ymm0 \n" |
4565 | 439 | "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" |
4566 | 439 | "vpaddsw %%ymm3,%%ymm0,%%ymm0 \n" |
4567 | 439 | "vpsraw $0x6,%%ymm0,%%ymm0 \n" |
4568 | 439 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
4569 | 439 | "vpunpcklbw %%ymm0,%%ymm0,%%ymm1 \n" |
4570 | 439 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
4571 | 439 | "vpunpcklwd %%ymm1,%%ymm1,%%ymm0 \n" |
4572 | 439 | "vpunpckhwd %%ymm1,%%ymm1,%%ymm1 \n" |
4573 | 439 | "vpor %%ymm4,%%ymm0,%%ymm0 \n" |
4574 | 439 | "vpor %%ymm4,%%ymm1,%%ymm1 \n" |
4575 | 439 | "vmovdqu %%ymm0,(%1) \n" |
4576 | 439 | "vmovdqu %%ymm1,0x20(%1) \n" |
4577 | 439 | "lea 0x40(%1),%1 \n" |
4578 | 439 | "sub $0x10,%2 \n" |
4579 | 439 | "jg 1b \n" |
4580 | 439 | "vzeroupper \n" |
4581 | 439 | : "+r"(y_buf), // %0 |
4582 | 439 | "+r"(dst_argb), // %1 |
4583 | 439 | "+rm"(width) // %2 |
4584 | 439 | : "r"(yuvconstants) // %3 |
4585 | 439 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
4586 | 439 | } |
4587 | | #endif // HAS_I400TOARGBROW_AVX2 |
4588 | | |
4589 | | #ifdef HAS_MIRRORROW_SSSE3 |
4590 | | // Shuffle table for reversing the bytes. |
4591 | | static const uvec8 kShuffleMirror = {15u, 14u, 13u, 12u, 11u, 10u, 9u, 8u, |
4592 | | 7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u}; |
4593 | | |
4594 | 0 | void MirrorRow_SSSE3(const uint8_t* src, uint8_t* dst, int width) { |
4595 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4596 | 0 | asm volatile("movdqa %3,%%xmm5 \n" |
4597 | |
|
4598 | 0 | LABELALIGN |
4599 | 0 | "1: \n" |
4600 | 0 | "movdqu -0x10(%0,%2,1),%%xmm0 \n" |
4601 | 0 | "pshufb %%xmm5,%%xmm0 \n" |
4602 | 0 | "movdqu %%xmm0,(%1) \n" |
4603 | 0 | "lea 0x10(%1),%1 \n" |
4604 | 0 | "sub $0x10,%2 \n" |
4605 | 0 | "jg 1b \n" |
4606 | 0 | : "+r"(src), // %0 |
4607 | 0 | "+r"(dst), // %1 |
4608 | 0 | "+r"(temp_width) // %2 |
4609 | 0 | : "m"(kShuffleMirror) // %3 |
4610 | 0 | : "memory", "cc", "xmm0", "xmm5"); |
4611 | 0 | } |
4612 | | #endif // HAS_MIRRORROW_SSSE3 |
4613 | | |
4614 | | #ifdef HAS_MIRRORROW_AVX512BW |
4615 | 0 | void MirrorRow_AVX512BW(const uint8_t* src, uint8_t* dst, int width) { |
4616 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4617 | 0 | asm volatile("vbroadcasti32x4 %3,%%zmm5 \n" |
4618 | |
|
4619 | 0 | LABELALIGN |
4620 | 0 | "1: \n" |
4621 | 0 | "vmovdqu8 -0x40(%0,%2,1),%%zmm0 \n" |
4622 | 0 | "vpshufb %%zmm5,%%zmm0,%%zmm0 \n" |
4623 | 0 | "vshufi64x2 $0x1b,%%zmm0,%%zmm0,%%zmm0 \n" |
4624 | 0 | "vmovdqu8 %%zmm0,(%1) \n" |
4625 | 0 | "lea 0x40(%1),%1 \n" |
4626 | 0 | "sub $0x40,%2 \n" |
4627 | 0 | "jg 1b \n" |
4628 | 0 | "vzeroupper \n" |
4629 | 0 | : "+r"(src), // %0 |
4630 | 0 | "+r"(dst), // %1 |
4631 | 0 | "+r"(temp_width) // %2 |
4632 | 0 | : "m"(kShuffleMirror) // %3 |
4633 | 0 | : "memory", "cc", "zmm0", "zmm5"); |
4634 | 0 | } |
4635 | | #endif // HAS_MIRRORROW_AVX512BW |
4636 | | |
4637 | | #ifdef HAS_MIRRORROW_AVX2 |
4638 | 0 | void MirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
4639 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4640 | 0 | asm volatile("vbroadcasti128 %3,%%ymm5 \n" |
4641 | |
|
4642 | 0 | LABELALIGN |
4643 | 0 | "1: \n" |
4644 | 0 | "vmovdqu -0x20(%0,%2,1),%%ymm0 \n" |
4645 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" |
4646 | 0 | "vpermq $0x4e,%%ymm0,%%ymm0 \n" |
4647 | 0 | "vmovdqu %%ymm0,(%1) \n" |
4648 | 0 | "lea 0x20(%1),%1 \n" |
4649 | 0 | "sub $0x20,%2 \n" |
4650 | 0 | "jg 1b \n" |
4651 | 0 | "vzeroupper \n" |
4652 | 0 | : "+r"(src), // %0 |
4653 | 0 | "+r"(dst), // %1 |
4654 | 0 | "+r"(temp_width) // %2 |
4655 | 0 | : "m"(kShuffleMirror) // %3 |
4656 | 0 | : "memory", "cc", "xmm0", "xmm5"); |
4657 | 0 | } |
4658 | | #endif // HAS_MIRRORROW_AVX2 |
4659 | | |
4660 | | #if defined(HAS_MIRRORSPLITUVROW_AVX2) || defined(HAS_MIRRORSPLITUVROW_AVX512BW) |
4661 | | // Shuffle table for reversing the bytes of UV channels. |
4662 | | static const uvec8 kShuffleMirrorSplitUV = {14u, 12u, 10u, 8u, 6u, 4u, 2u, 0u, |
4663 | | 15u, 13u, 11u, 9u, 7u, 5u, 3u, 1u}; |
4664 | | #endif |
4665 | | |
4666 | | #ifdef HAS_MIRRORSPLITUVROW_AVX512BW |
4667 | | static const uint64_t kMirrorSplitUVPermute[8] = {6, 4, 2, 0, 7, 5, 3, 1}; |
4668 | | |
4669 | | void MirrorSplitUVRow_AVX512BW(const uint8_t* src, |
4670 | | uint8_t* dst_u, |
4671 | | uint8_t* dst_v, |
4672 | 0 | int width) { |
4673 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4674 | 0 | asm volatile( |
4675 | 0 | "vbroadcasti32x4 %4,%%zmm1 \n" |
4676 | 0 | "lea -0x40(%0,%3,2),%0 \n" |
4677 | 0 | "sub %1,%2 \n" |
4678 | 0 | "vmovdqu64 %5,%%zmm3 \n" |
4679 | |
|
4680 | 0 | LABELALIGN |
4681 | 0 | "1: \n" |
4682 | 0 | "vmovdqu8 (%0),%%zmm0 \n" |
4683 | 0 | "lea -0x40(%0),%0 \n" |
4684 | 0 | "vpshufb %%zmm1,%%zmm0,%%zmm0 \n" |
4685 | 0 | "vpermq %%zmm0,%%zmm3,%%zmm0 \n" |
4686 | 0 | "vextracti64x4 $0x1,%%zmm0,%%ymm2 \n" |
4687 | 0 | "vmovdqu %%ymm0,(%1) \n" |
4688 | 0 | "vmovdqu %%ymm2,0x00(%1,%2,1) \n" |
4689 | 0 | "lea 0x20(%1),%1 \n" |
4690 | 0 | "sub $0x20,%3 \n" |
4691 | 0 | "jg 1b \n" |
4692 | 0 | "vzeroupper \n" |
4693 | 0 | : "+r"(src), // %0 |
4694 | 0 | "+r"(dst_u), // %1 |
4695 | 0 | "+r"(dst_v), // %2 |
4696 | 0 | "+r"(temp_width) // %3 |
4697 | 0 | : "m"(kShuffleMirrorSplitUV), // %4 |
4698 | 0 | "m"(kMirrorSplitUVPermute) // %5 |
4699 | 0 | : "memory", "cc", "zmm0", "zmm1", "zmm2", "zmm3"); |
4700 | 0 | } |
4701 | | #endif // HAS_MIRRORSPLITUVROW_AVX512BW |
4702 | | |
4703 | | #ifdef HAS_MIRRORSPLITUVROW_AVX2 |
4704 | | void MirrorSplitUVRow_AVX2(const uint8_t* src, |
4705 | | uint8_t* dst_u, |
4706 | | uint8_t* dst_v, |
4707 | 0 | int width) { |
4708 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4709 | 0 | asm volatile( |
4710 | 0 | "vbroadcasti128 %4,%%ymm1 \n" |
4711 | 0 | "lea -0x20(%0,%3,2),%0 \n" |
4712 | 0 | "sub %1,%2 \n" |
4713 | |
|
4714 | 0 | LABELALIGN |
4715 | 0 | "1: \n" |
4716 | 0 | "vmovdqu (%0),%%ymm0 \n" |
4717 | 0 | "lea -0x20(%0),%0 \n" |
4718 | 0 | "vpshufb %%ymm1,%%ymm0,%%ymm0 \n" |
4719 | 0 | "vpermq $0x72,%%ymm0,%%ymm0 \n" |
4720 | 0 | "vextracti128 $0x1,%%ymm0,%%xmm2 \n" |
4721 | 0 | "vmovdqu %%xmm0,(%1) \n" |
4722 | 0 | "vmovdqu %%xmm2,0x00(%1,%2,1) \n" |
4723 | 0 | "lea 0x10(%1),%1 \n" |
4724 | 0 | "sub $0x10,%3 \n" |
4725 | 0 | "jg 1b \n" |
4726 | 0 | "vzeroupper \n" |
4727 | 0 | : "+r"(src), // %0 |
4728 | 0 | "+r"(dst_u), // %1 |
4729 | 0 | "+r"(dst_v), // %2 |
4730 | 0 | "+r"(temp_width) // %3 |
4731 | 0 | : "m"(kShuffleMirrorSplitUV) // %4 |
4732 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
4733 | 0 | } |
4734 | | #endif // HAS_MIRRORSPLITUVROW_AVX2 |
4735 | | |
4736 | | #ifdef HAS_MIRRORUVROW_SSSE3 |
4737 | | // Shuffle table for reversing the UV. |
4738 | | static const uvec8 kShuffleMirrorUV = {14u, 15u, 12u, 13u, 10u, 11u, 8u, 9u, |
4739 | | 6u, 7u, 4u, 5u, 2u, 3u, 0u, 1u}; |
4740 | | |
4741 | 0 | void MirrorUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_uv, int width) { |
4742 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4743 | 0 | asm volatile("movdqa %3,%%xmm5 \n" |
4744 | |
|
4745 | 0 | LABELALIGN |
4746 | 0 | "1: \n" |
4747 | 0 | "movdqu -0x10(%0,%2,2),%%xmm0 \n" |
4748 | 0 | "pshufb %%xmm5,%%xmm0 \n" |
4749 | 0 | "movdqu %%xmm0,(%1) \n" |
4750 | 0 | "lea 0x10(%1),%1 \n" |
4751 | 0 | "sub $0x8,%2 \n" |
4752 | 0 | "jg 1b \n" |
4753 | 0 | : "+r"(src_uv), // %0 |
4754 | 0 | "+r"(dst_uv), // %1 |
4755 | 0 | "+r"(temp_width) // %2 |
4756 | 0 | : "m"(kShuffleMirrorUV) // %3 |
4757 | 0 | : "memory", "cc", "xmm0", "xmm5"); |
4758 | 0 | } |
4759 | | #endif // HAS_MIRRORUVROW_SSSE3 |
4760 | | |
4761 | | #ifdef HAS_MIRRORUVROW_AVX2 |
4762 | 0 | void MirrorUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_uv, int width) { |
4763 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4764 | 0 | asm volatile("vbroadcasti128 %3,%%ymm5 \n" |
4765 | |
|
4766 | 0 | LABELALIGN |
4767 | 0 | "1: \n" |
4768 | 0 | "vmovdqu -0x20(%0,%2,2),%%ymm0 \n" |
4769 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" |
4770 | 0 | "vpermq $0x4e,%%ymm0,%%ymm0 \n" |
4771 | 0 | "vmovdqu %%ymm0,(%1) \n" |
4772 | 0 | "lea 0x20(%1),%1 \n" |
4773 | 0 | "sub $0x10,%2 \n" |
4774 | 0 | "jg 1b \n" |
4775 | 0 | "vzeroupper \n" |
4776 | 0 | : "+r"(src_uv), // %0 |
4777 | 0 | "+r"(dst_uv), // %1 |
4778 | 0 | "+r"(temp_width) // %2 |
4779 | 0 | : "m"(kShuffleMirrorUV) // %3 |
4780 | 0 | : "memory", "cc", "xmm0", "xmm5"); |
4781 | 0 | } |
4782 | | #endif // HAS_MIRRORUVROW_AVX2 |
4783 | | |
4784 | | #ifdef HAS_RGB24MIRRORROW_SSSE3 |
4785 | | |
4786 | | // Shuffle first 5 pixels to last 5 mirrored. first byte zero |
4787 | | static const uvec8 kShuffleMirrorRGB0 = {128u, 12u, 13u, 14u, 9u, 10u, 11u, 6u, |
4788 | | 7u, 8u, 3u, 4u, 5u, 0u, 1u, 2u}; |
4789 | | |
4790 | | // Shuffle last 5 pixels to first 5 mirrored. last byte zero |
4791 | | static const uvec8 kShuffleMirrorRGB1 = { |
4792 | | 13u, 14u, 15u, 10u, 11u, 12u, 7u, 8u, 9u, 4u, 5u, 6u, 1u, 2u, 3u, 128u}; |
4793 | | |
4794 | | // Shuffle 5 pixels at a time (15 bytes) |
4795 | | void RGB24MirrorRow_SSSE3(const uint8_t* src_rgb24, |
4796 | | uint8_t* dst_rgb24, |
4797 | | int width) { |
4798 | | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4799 | | src_rgb24 += width * 3 - 48; |
4800 | | asm volatile( |
4801 | | "movdqa %3,%%xmm4 \n" |
4802 | | "movdqa %4,%%xmm5 \n" |
4803 | | |
4804 | | LABELALIGN |
4805 | | "1: \n" |
4806 | | "movdqu (%0),%%xmm0 \n" // first 5 |
4807 | | "movdqu 15(%0),%%xmm1 \n" // next 5 |
4808 | | "movdqu 30(%0),%%xmm2 \n" // next 5 |
4809 | | "movdqu 32(%0),%%xmm3 \n" // last 1 special |
4810 | | "pshufb %%xmm4,%%xmm0 \n" |
4811 | | "pshufb %%xmm4,%%xmm1 \n" |
4812 | | "pshufb %%xmm4,%%xmm2 \n" |
4813 | | "pshufb %%xmm5,%%xmm3 \n" |
4814 | | "lea -0x30(%0),%0 \n" |
4815 | | "movdqu %%xmm0,32(%1) \n" // last 5 |
4816 | | "movdqu %%xmm1,17(%1) \n" // next 5 |
4817 | | "movdqu %%xmm2,2(%1) \n" // next 5 |
4818 | | "movlpd %%xmm3,0(%1) \n" // first 1 |
4819 | | "lea 0x30(%1),%1 \n" |
4820 | | "sub $0x10,%2 \n" |
4821 | | "jg 1b \n" |
4822 | | : "+r"(src_rgb24), // %0 |
4823 | | "+r"(dst_rgb24), // %1 |
4824 | | "+r"(temp_width) // %2 |
4825 | | : "m"(kShuffleMirrorRGB0), // %3 |
4826 | | "m"(kShuffleMirrorRGB1) // %4 |
4827 | | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
4828 | | } |
4829 | | #endif // HAS_RGB24MIRRORROW_SSSE3 |
4830 | | |
4831 | | #ifdef HAS_RGB24MIRRORROW_AVX2 |
4832 | | // Shuffle first 10 pixels to last 10 mirrored. first byte zero |
4833 | | static const uvec8 kShuffleMirrorRGB0_AVX = { |
4834 | | 128u, 12u, 13u, 14u, 9u, 10u, 11u, 6u, 7u, 8u, 3u, 4u, 5u, 0u, 1u, 2u}; |
4835 | | |
4836 | | // Shuffle last 2 pixels to first 2 mirrored. last byte zero |
4837 | | static const uvec8 kShuffleMirrorRGB1_AVX = { |
4838 | | 13u, 14u, 15u, 10u, 11u, 12u, 7u, 8u, 9u, 4u, 5u, 6u, 1u, 2u, 3u, 128u}; |
4839 | | |
4840 | | void RGB24MirrorRow_AVX2(const uint8_t* src_rgb24, |
4841 | | uint8_t* dst_rgb24, |
4842 | 0 | int width) { |
4843 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4844 | 0 | src_rgb24 += width * 3 - 96; |
4845 | 0 | asm volatile( |
4846 | 0 | "vbroadcasti128 %3,%%ymm4 \n" |
4847 | 0 | "vmovdqa %4,%%xmm5 \n" |
4848 | |
|
4849 | 0 | LABELALIGN |
4850 | 0 | "1: \n" |
4851 | 0 | "vmovdqu (%0),%%xmm0 \n" // first 10 |
4852 | 0 | "vinserti128 $1,15(%0),%%ymm0,%%ymm0 \n" |
4853 | 0 | "vmovdqu 30(%0),%%xmm1 \n" // next 10 |
4854 | 0 | "vinserti128 $1,45(%0),%%ymm1,%%ymm1 \n" |
4855 | 0 | "vmovdqu 60(%0),%%xmm2 \n" // next 10 |
4856 | 0 | "vinserti128 $1,75(%0),%%ymm2,%%ymm2 \n" |
4857 | 0 | "vmovdqu 80(%0),%%xmm3 \n" // last 2 special |
4858 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
4859 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
4860 | 0 | "vpshufb %%ymm4,%%ymm2,%%ymm2 \n" |
4861 | 0 | "vpshufb %%xmm5,%%xmm3,%%xmm3 \n" |
4862 | 0 | "lea -0x60(%0),%0 \n" |
4863 | 0 | "vmovdqu %%xmm0,80(%1) \n" |
4864 | 0 | "vextracti128 $1,%%ymm0,65(%1) \n" |
4865 | 0 | "vmovdqu %%xmm1,50(%1) \n" |
4866 | 0 | "vextracti128 $1,%%ymm1,35(%1) \n" |
4867 | 0 | "vmovdqu %%xmm2,20(%1) \n" |
4868 | 0 | "vextracti128 $1,%%ymm2,5(%1) \n" |
4869 | 0 | "vmovq %%xmm3,0(%1) \n" |
4870 | 0 | "lea 0x60(%1),%1 \n" |
4871 | 0 | "sub $0x20,%2 \n" |
4872 | 0 | "jg 1b \n" |
4873 | 0 | "vzeroupper \n" |
4874 | 0 | : "+r"(src_rgb24), // %0 |
4875 | 0 | "+r"(dst_rgb24), // %1 |
4876 | 0 | "+r"(temp_width) // %2 |
4877 | 0 | : "m"(kShuffleMirrorRGB0_AVX), // %3 |
4878 | 0 | "m"(kShuffleMirrorRGB1_AVX) // %4 |
4879 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
4880 | 0 | } |
4881 | | #endif // HAS_RGB24MIRRORROW_AVX2 |
4882 | | |
4883 | | #ifdef HAS_ARGBMIRRORROW_SSE2 |
4884 | | |
4885 | 0 | void ARGBMirrorRow_SSE2(const uint8_t* src, uint8_t* dst, int width) { |
4886 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4887 | 0 | asm volatile("lea -0x10(%0,%2,4),%0 \n" |
4888 | |
|
4889 | 0 | LABELALIGN |
4890 | 0 | "1: \n" |
4891 | 0 | "movdqu (%0),%%xmm0 \n" |
4892 | 0 | "pshufd $0x1b,%%xmm0,%%xmm0 \n" |
4893 | 0 | "lea -0x10(%0),%0 \n" |
4894 | 0 | "movdqu %%xmm0,(%1) \n" |
4895 | 0 | "lea 0x10(%1),%1 \n" |
4896 | 0 | "sub $0x4,%2 \n" |
4897 | 0 | "jg 1b \n" |
4898 | 0 | : "+r"(src), // %0 |
4899 | 0 | "+r"(dst), // %1 |
4900 | 0 | "+r"(temp_width) // %2 |
4901 | 0 | : |
4902 | 0 | : "memory", "cc", "xmm0"); |
4903 | 0 | } |
4904 | | #endif // HAS_ARGBMIRRORROW_SSE2 |
4905 | | |
4906 | | #ifdef HAS_ARGBMIRRORROW_AVX2 |
4907 | | // Shuffle table for reversing the bytes. |
4908 | | static const ulvec32 kARGBShuffleMirror_AVX2 = {7u, 6u, 5u, 4u, 3u, 2u, 1u, 0u}; |
4909 | 0 | void ARGBMirrorRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
4910 | 0 | ptrdiff_t temp_width = (ptrdiff_t)(width); |
4911 | 0 | asm volatile("vmovdqu %3,%%ymm5 \n" |
4912 | |
|
4913 | 0 | LABELALIGN |
4914 | 0 | "1: \n" |
4915 | 0 | "vpermd -0x20(%0,%2,4),%%ymm5,%%ymm0 \n" |
4916 | 0 | "vmovdqu %%ymm0,(%1) \n" |
4917 | 0 | "lea 0x20(%1),%1 \n" |
4918 | 0 | "sub $0x8,%2 \n" |
4919 | 0 | "jg 1b \n" |
4920 | 0 | "vzeroupper \n" |
4921 | 0 | : "+r"(src), // %0 |
4922 | 0 | "+r"(dst), // %1 |
4923 | 0 | "+r"(temp_width) // %2 |
4924 | 0 | : "m"(kARGBShuffleMirror_AVX2) // %3 |
4925 | 0 | : "memory", "cc", "xmm0", "xmm5"); |
4926 | 0 | } |
4927 | | #endif // HAS_ARGBMIRRORROW_AVX2 |
4928 | | |
4929 | | #ifdef HAS_SPLITUVROW_AVX2 |
4930 | | void SplitUVRow_AVX2(const uint8_t* src_uv, |
4931 | | uint8_t* dst_u, |
4932 | | uint8_t* dst_v, |
4933 | 0 | int width) { |
4934 | 0 | asm volatile( |
4935 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
4936 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
4937 | 0 | "sub %1,%2 \n" |
4938 | |
|
4939 | 0 | LABELALIGN |
4940 | 0 | "1: \n" |
4941 | 0 | "vmovdqu (%0),%%ymm0 \n" |
4942 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
4943 | 0 | "lea 0x40(%0),%0 \n" |
4944 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm2 \n" |
4945 | 0 | "vpsrlw $0x8,%%ymm1,%%ymm3 \n" |
4946 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" |
4947 | 0 | "vpand %%ymm5,%%ymm1,%%ymm1 \n" |
4948 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
4949 | 0 | "vpackuswb %%ymm3,%%ymm2,%%ymm2 \n" |
4950 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
4951 | 0 | "vpermq $0xd8,%%ymm2,%%ymm2 \n" |
4952 | 0 | "vmovdqu %%ymm0,(%1) \n" |
4953 | 0 | "vmovdqu %%ymm2,0x00(%1,%2,1) \n" |
4954 | 0 | "lea 0x20(%1),%1 \n" |
4955 | 0 | "sub $0x20,%3 \n" |
4956 | 0 | "jg 1b \n" |
4957 | 0 | "vzeroupper \n" |
4958 | 0 | : "+r"(src_uv), // %0 |
4959 | 0 | "+r"(dst_u), // %1 |
4960 | 0 | "+r"(dst_v), // %2 |
4961 | 0 | "+r"(width) // %3 |
4962 | 0 | : |
4963 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
4964 | 0 | } |
4965 | | #endif // HAS_SPLITUVROW_AVX2 |
4966 | | |
4967 | | #ifdef HAS_SPLITUVROW_AVX512BW |
4968 | | static const uint64_t kSplitUVPermute[8] = {0, 2, 4, 6, 1, 3, 5, 7}; |
4969 | | |
4970 | | void SplitUVRow_AVX512BW(const uint8_t* src_uv, |
4971 | | uint8_t* dst_u, |
4972 | | uint8_t* dst_v, |
4973 | 0 | int width) { |
4974 | 0 | asm volatile( |
4975 | 0 | "vpternlogd $0xff,%%zmm5,%%zmm5,%%zmm5 \n" |
4976 | 0 | "vpsrlw $0x8,%%zmm5,%%zmm5 \n" |
4977 | 0 | "vmovdqu64 %4,%%zmm4 \n" |
4978 | 0 | "sub %1,%2 \n" |
4979 | |
|
4980 | 0 | LABELALIGN |
4981 | 0 | "1: \n" |
4982 | 0 | "vmovdqu8 (%0),%%zmm0 \n" |
4983 | 0 | "vmovdqu8 0x40(%0),%%zmm1 \n" |
4984 | 0 | "lea 0x80(%0),%0 \n" |
4985 | 0 | "vpsrlw $0x8,%%zmm0,%%zmm2 \n" |
4986 | 0 | "vpsrlw $0x8,%%zmm1,%%zmm3 \n" |
4987 | 0 | "vpandd %%zmm5,%%zmm0,%%zmm0 \n" |
4988 | 0 | "vpandd %%zmm5,%%zmm1,%%zmm1 \n" |
4989 | 0 | "vpackuswb %%zmm1,%%zmm0,%%zmm0 \n" |
4990 | 0 | "vpackuswb %%zmm3,%%zmm2,%%zmm2 \n" |
4991 | 0 | "vpermq %%zmm0,%%zmm4,%%zmm0 \n" |
4992 | 0 | "vpermq %%zmm2,%%zmm4,%%zmm2 \n" |
4993 | 0 | "vmovdqu8 %%zmm0,(%1) \n" |
4994 | 0 | "vmovdqu8 %%zmm2,0x00(%1,%2,1) \n" |
4995 | 0 | "lea 0x40(%1),%1 \n" |
4996 | 0 | "sub $0x40,%3 \n" |
4997 | 0 | "jg 1b \n" |
4998 | 0 | "vzeroupper \n" |
4999 | 0 | : "+r"(src_uv), // %0 |
5000 | 0 | "+r"(dst_u), // %1 |
5001 | 0 | "+r"(dst_v), // %2 |
5002 | 0 | "+r"(width) // %3 |
5003 | 0 | : "m"(kSplitUVPermute) // %4 |
5004 | 0 | : "memory", "cc", "zmm0", "zmm1", "zmm2", "zmm3", "zmm4", "zmm5"); |
5005 | 0 | } |
5006 | | #endif // HAS_SPLITUVROW_AVX512BW |
5007 | | |
5008 | | #ifdef HAS_SPLITUVROW_SSE2 |
5009 | | void SplitUVRow_SSE2(const uint8_t* src_uv, |
5010 | | uint8_t* dst_u, |
5011 | | uint8_t* dst_v, |
5012 | 0 | int width) { |
5013 | 0 | asm volatile( |
5014 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
5015 | 0 | "psrlw $0x8,%%xmm5 \n" |
5016 | 0 | "sub %1,%2 \n" |
5017 | |
|
5018 | 0 | LABELALIGN |
5019 | 0 | "1: \n" |
5020 | 0 | "movdqu (%0),%%xmm0 \n" |
5021 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5022 | 0 | "lea 0x20(%0),%0 \n" |
5023 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
5024 | 0 | "movdqa %%xmm1,%%xmm3 \n" |
5025 | 0 | "pand %%xmm5,%%xmm0 \n" |
5026 | 0 | "pand %%xmm5,%%xmm1 \n" |
5027 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
5028 | 0 | "psrlw $0x8,%%xmm2 \n" |
5029 | 0 | "psrlw $0x8,%%xmm3 \n" |
5030 | 0 | "packuswb %%xmm3,%%xmm2 \n" |
5031 | 0 | "movdqu %%xmm0,(%1) \n" |
5032 | 0 | "movdqu %%xmm2,0x00(%1,%2,1) \n" |
5033 | 0 | "lea 0x10(%1),%1 \n" |
5034 | 0 | "sub $0x10,%3 \n" |
5035 | 0 | "jg 1b \n" |
5036 | 0 | : "+r"(src_uv), // %0 |
5037 | 0 | "+r"(dst_u), // %1 |
5038 | 0 | "+r"(dst_v), // %2 |
5039 | 0 | "+r"(width) // %3 |
5040 | 0 | : |
5041 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
5042 | 0 | } |
5043 | | #endif // HAS_SPLITUVROW_SSE2 |
5044 | | |
5045 | | #ifdef HAS_DETILEROW_SSE2 |
5046 | | void DetileRow_SSE2(const uint8_t* src, |
5047 | | ptrdiff_t src_tile_stride, |
5048 | | uint8_t* dst, |
5049 | 0 | int width) { |
5050 | 0 | asm volatile( |
5051 | 0 | "1: \n" |
5052 | 0 | "movdqu (%0),%%xmm0 \n" |
5053 | 0 | "sub $0x10,%2 \n" |
5054 | 0 | "lea (%0,%3),%0 \n" |
5055 | 0 | "movdqu %%xmm0,(%1) \n" |
5056 | 0 | "lea 0x10(%1),%1 \n" |
5057 | 0 | "jg 1b \n" |
5058 | 0 | : "+r"(src), // %0 |
5059 | 0 | "+r"(dst), // %1 |
5060 | 0 | "+r"(width) // %2 |
5061 | 0 | : "r"(src_tile_stride) // %3 |
5062 | 0 | : "cc", "memory", "xmm0"); |
5063 | 0 | } |
5064 | | #endif // HAS_DETILEROW_SSE2 |
5065 | | |
5066 | | #ifdef HAS_DETILEROW_16_SSE2 |
5067 | | void DetileRow_16_SSE2(const uint16_t* src, |
5068 | | ptrdiff_t src_tile_stride, |
5069 | | uint16_t* dst, |
5070 | 0 | int width) { |
5071 | 0 | asm volatile( |
5072 | 0 | "1: \n" |
5073 | 0 | "movdqu (%0),%%xmm0 \n" |
5074 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5075 | 0 | "lea (%0,%3,2),%0 \n" |
5076 | 0 | "movdqu %%xmm0,(%1) \n" |
5077 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
5078 | 0 | "lea 0x20(%1),%1 \n" |
5079 | 0 | "sub $0x10,%2 \n" |
5080 | 0 | "jg 1b \n" |
5081 | 0 | : "+r"(src), // %0 |
5082 | 0 | "+r"(dst), // %1 |
5083 | 0 | "+r"(width) // %2 |
5084 | 0 | : "r"(src_tile_stride) // %3 |
5085 | 0 | : "cc", "memory", "xmm0", "xmm1"); |
5086 | 0 | } |
5087 | | #endif // HAS_DETILEROW_SSE2 |
5088 | | |
5089 | | #ifdef HAS_DETILEROW_16_AVX |
5090 | | void DetileRow_16_AVX(const uint16_t* src, |
5091 | | ptrdiff_t src_tile_stride, |
5092 | | uint16_t* dst, |
5093 | 0 | int width) { |
5094 | 0 | asm volatile( |
5095 | 0 | "1: \n" |
5096 | 0 | "vmovdqu (%0),%%ymm0 \n" |
5097 | 0 | "lea (%0,%3,2),%0 \n" |
5098 | 0 | "vmovdqu %%ymm0,(%1) \n" |
5099 | 0 | "lea 0x20(%1),%1 \n" |
5100 | 0 | "sub $0x10,%2 \n" |
5101 | 0 | "jg 1b \n" |
5102 | 0 | "vzeroupper \n" |
5103 | 0 | : "+r"(src), // %0 |
5104 | 0 | "+r"(dst), // %1 |
5105 | 0 | "+r"(width) // %2 |
5106 | 0 | : "r"(src_tile_stride) // %3 |
5107 | 0 | : "cc", "memory", "xmm0"); |
5108 | 0 | } |
5109 | | #endif // HAS_DETILEROW_AVX |
5110 | | |
5111 | | #ifdef HAS_DETILETOYUY2_SSE2 |
5112 | | // Read 16 Y, 8 UV, and write 8 YUYV. |
5113 | | void DetileToYUY2_SSE2(const uint8_t* src_y, |
5114 | | ptrdiff_t src_y_tile_stride, |
5115 | | const uint8_t* src_uv, |
5116 | | ptrdiff_t src_uv_tile_stride, |
5117 | | uint8_t* dst_yuy2, |
5118 | 0 | int width) { |
5119 | 0 | asm volatile( |
5120 | 0 | "1: \n" |
5121 | 0 | "movdqu (%0),%%xmm0 \n" // Load 16 Y |
5122 | 0 | "sub $0x10,%3 \n" |
5123 | 0 | "lea (%0,%4),%0 \n" |
5124 | 0 | "movdqu (%1),%%xmm1 \n" // Load 8 UV |
5125 | 0 | "lea (%1,%5),%1 \n" |
5126 | 0 | "movdqu %%xmm0,%%xmm2 \n" |
5127 | 0 | "punpcklbw %%xmm1,%%xmm0 \n" |
5128 | 0 | "punpckhbw %%xmm1,%%xmm2 \n" |
5129 | 0 | "movdqu %%xmm0,(%2) \n" |
5130 | 0 | "movdqu %%xmm2,0x10(%2) \n" |
5131 | 0 | "lea 0x20(%2),%2 \n" |
5132 | 0 | "jg 1b \n" |
5133 | 0 | : "+r"(src_y), // %0 |
5134 | 0 | "+r"(src_uv), // %1 |
5135 | 0 | "+r"(dst_yuy2), // %2 |
5136 | 0 | "+r"(width) // %3 |
5137 | 0 | : "r"(src_y_tile_stride), // %4 |
5138 | 0 | "r"(src_uv_tile_stride) // %5 |
5139 | 0 | : "cc", "memory", "xmm0", "xmm1", "xmm2" // Clobber list |
5140 | 0 | ); |
5141 | 0 | } |
5142 | | #endif |
5143 | | |
5144 | | #ifdef HAS_DETILESPLITUVROW_SSSE3 |
5145 | | // TODO(greenjustin): Look into generating these constants instead of loading |
5146 | | // them since this can cause branch mispredicts for fPIC code on 32-bit |
5147 | | // machines. |
5148 | | static const uvec8 kDeinterlaceUV = {0, 2, 4, 6, 8, 10, 12, 14, |
5149 | | 1, 3, 5, 7, 9, 11, 13, 15}; |
5150 | | |
5151 | | // TODO(greenjustin): Research alternatives to pshufb, since pshufb can be very |
5152 | | // slow on older SSE2 processors. |
5153 | | void DetileSplitUVRow_SSSE3(const uint8_t* src_uv, |
5154 | | ptrdiff_t src_tile_stride, |
5155 | | uint8_t* dst_u, |
5156 | | uint8_t* dst_v, |
5157 | 0 | int width) { |
5158 | 0 | asm volatile( |
5159 | 0 | "movdqu %4,%%xmm1 \n" |
5160 | 0 | "1: \n" |
5161 | 0 | "movdqu (%0),%%xmm0 \n" |
5162 | 0 | "lea (%0, %5),%0 \n" |
5163 | 0 | "pshufb %%xmm1,%%xmm0 \n" |
5164 | 0 | "movq %%xmm0,(%1) \n" |
5165 | 0 | "lea 0x8(%1),%1 \n" |
5166 | 0 | "movhps %%xmm0,(%2) \n" |
5167 | 0 | "lea 0x8(%2),%2 \n" |
5168 | 0 | "sub $0x10,%3 \n" |
5169 | 0 | "jg 1b \n" |
5170 | 0 | : "+r"(src_uv), // %0 |
5171 | 0 | "+r"(dst_u), // %1 |
5172 | 0 | "+r"(dst_v), // %2 |
5173 | 0 | "+r"(width) // %3 |
5174 | 0 | : "m"(kDeinterlaceUV), // %4 |
5175 | 0 | "r"(src_tile_stride) // %5 |
5176 | 0 | : "cc", "memory", "xmm0", "xmm1"); |
5177 | 0 | } |
5178 | | #endif // HAS_DETILESPLITUVROW_SSSE3 |
5179 | | |
5180 | | #ifdef HAS_MERGEUVROW_AVX512BW |
5181 | | void MergeUVRow_AVX512BW(const uint8_t* src_u, |
5182 | | const uint8_t* src_v, |
5183 | | uint8_t* dst_uv, |
5184 | 0 | int width) { |
5185 | 0 | asm volatile("sub %0,%1 \n" |
5186 | |
|
5187 | 0 | LABELALIGN |
5188 | 0 | "1: \n" |
5189 | 0 | "vpmovzxbw (%0),%%zmm0 \n" |
5190 | 0 | "vpmovzxbw 0x00(%0,%1,1),%%zmm1 \n" |
5191 | 0 | "lea 0x20(%0),%0 \n" |
5192 | 0 | "vpsllw $0x8,%%zmm1,%%zmm1 \n" |
5193 | 0 | "vporq %%zmm0,%%zmm1,%%zmm2 \n" |
5194 | 0 | "vmovdqu64 %%zmm2,(%2) \n" |
5195 | 0 | "lea 0x40(%2),%2 \n" |
5196 | 0 | "sub $0x20,%3 \n" |
5197 | 0 | "jg 1b \n" |
5198 | 0 | "vzeroupper \n" |
5199 | 0 | : "+r"(src_u), // %0 |
5200 | 0 | "+r"(src_v), // %1 |
5201 | 0 | "+r"(dst_uv), // %2 |
5202 | 0 | "+r"(width) // %3 |
5203 | 0 | : |
5204 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5205 | 0 | } |
5206 | | #endif // HAS_MERGEUVROW_AVX512BW |
5207 | | |
5208 | | #ifdef HAS_MERGEUVROW_AVX2 |
5209 | | void MergeUVRow_AVX2(const uint8_t* src_u, |
5210 | | const uint8_t* src_v, |
5211 | | uint8_t* dst_uv, |
5212 | 0 | int width) { |
5213 | 0 | asm volatile("sub %0,%1 \n" |
5214 | |
|
5215 | 0 | LABELALIGN |
5216 | 0 | "1: \n" |
5217 | 0 | "vpmovzxbw (%0),%%ymm0 \n" |
5218 | 0 | "vpmovzxbw 0x00(%0,%1,1),%%ymm1 \n" |
5219 | 0 | "lea 0x10(%0),%0 \n" |
5220 | 0 | "vpsllw $0x8,%%ymm1,%%ymm1 \n" |
5221 | 0 | "vpor %%ymm0,%%ymm1,%%ymm2 \n" |
5222 | 0 | "vmovdqu %%ymm2,(%2) \n" |
5223 | 0 | "lea 0x20(%2),%2 \n" |
5224 | 0 | "sub $0x10,%3 \n" |
5225 | 0 | "jg 1b \n" |
5226 | 0 | "vzeroupper \n" |
5227 | 0 | : "+r"(src_u), // %0 |
5228 | 0 | "+r"(src_v), // %1 |
5229 | 0 | "+r"(dst_uv), // %2 |
5230 | 0 | "+r"(width) // %3 |
5231 | 0 | : |
5232 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5233 | 0 | } |
5234 | | #endif // HAS_MERGEUVROW_AVX2 |
5235 | | |
5236 | | #ifdef HAS_MERGEUVROW_SSE2 |
5237 | | void MergeUVRow_SSE2(const uint8_t* src_u, |
5238 | | const uint8_t* src_v, |
5239 | | uint8_t* dst_uv, |
5240 | 0 | int width) { |
5241 | 0 | asm volatile("sub %0,%1 \n" |
5242 | |
|
5243 | 0 | LABELALIGN |
5244 | 0 | "1: \n" |
5245 | 0 | "movdqu (%0),%%xmm0 \n" |
5246 | 0 | "movdqu 0x00(%0,%1,1),%%xmm1 \n" |
5247 | 0 | "lea 0x10(%0),%0 \n" |
5248 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
5249 | 0 | "punpcklbw %%xmm1,%%xmm0 \n" |
5250 | 0 | "punpckhbw %%xmm1,%%xmm2 \n" |
5251 | 0 | "movdqu %%xmm0,(%2) \n" |
5252 | 0 | "movdqu %%xmm2,0x10(%2) \n" |
5253 | 0 | "lea 0x20(%2),%2 \n" |
5254 | 0 | "sub $0x10,%3 \n" |
5255 | 0 | "jg 1b \n" |
5256 | 0 | : "+r"(src_u), // %0 |
5257 | 0 | "+r"(src_v), // %1 |
5258 | 0 | "+r"(dst_uv), // %2 |
5259 | 0 | "+r"(width) // %3 |
5260 | 0 | : |
5261 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5262 | 0 | } |
5263 | | #endif // HAS_MERGEUVROW_SSE2 |
5264 | | |
5265 | | #ifdef HAS_MERGEUVROW_16_AVX2 |
5266 | | void MergeUVRow_16_AVX2(const uint16_t* src_u, |
5267 | | const uint16_t* src_v, |
5268 | | uint16_t* dst_uv, |
5269 | | int depth, |
5270 | 0 | int width) { |
5271 | 0 | asm volatile( |
5272 | 0 | "vmovd %4,%%xmm3 \n" |
5273 | 0 | "vmovd %5,%%xmm4 \n" |
5274 | |
|
5275 | 0 | "sub %0,%1 \n" |
5276 | | // 8 pixels per loop. |
5277 | |
|
5278 | 0 | LABELALIGN |
5279 | 0 | "1: \n" |
5280 | 0 | "vpmovzxwd (%0),%%ymm0 \n" |
5281 | 0 | "vpmovzxwd 0x00(%0,%1,1),%%ymm1 \n" |
5282 | 0 | "lea 0x10(%0),%0 \n" |
5283 | 0 | "vpsllw %%xmm3,%%ymm0,%%ymm0 \n" |
5284 | 0 | "vpslld %%xmm4,%%ymm1,%%ymm1 \n" |
5285 | 0 | "vpor %%ymm0,%%ymm1,%%ymm2 \n" |
5286 | 0 | "vmovdqu %%ymm2,(%2) \n" |
5287 | 0 | "lea 0x20(%2),%2 \n" |
5288 | 0 | "sub $0x8,%3 \n" |
5289 | 0 | "jg 1b \n" |
5290 | 0 | "vzeroupper \n" |
5291 | 0 | : "+r"(src_u), // %0 |
5292 | 0 | "+r"(src_v), // %1 |
5293 | 0 | "+r"(dst_uv), // %2 |
5294 | 0 | "+r"(width) // %3 |
5295 | 0 | : "r"(16 - depth), // %4 |
5296 | 0 | "r"(32 - depth) // %5 |
5297 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
5298 | 0 | } |
5299 | | #endif // HAS_MERGEUVROW_AVX2 |
5300 | | |
5301 | | #ifdef HAS_SPLITUVROW_16_AVX2 |
5302 | | const uvec8 kSplitUVShuffle16 = {0, 1, 4, 5, 8, 9, 12, 13, |
5303 | | 2, 3, 6, 7, 10, 11, 14, 15}; |
5304 | | void SplitUVRow_16_AVX2(const uint16_t* src_uv, |
5305 | | uint16_t* dst_u, |
5306 | | uint16_t* dst_v, |
5307 | | int depth, |
5308 | 0 | int width) { |
5309 | 0 | depth = 16 - depth; |
5310 | 0 | asm volatile( |
5311 | 0 | "vmovd %4,%%xmm3 \n" |
5312 | 0 | "vbroadcasti128 %5,%%ymm4 \n" |
5313 | 0 | "sub %1,%2 \n" |
5314 | | |
5315 | | // 16 pixels per loop. |
5316 | 0 | LABELALIGN |
5317 | 0 | "1: \n" |
5318 | 0 | "vmovdqu (%0),%%ymm0 \n" |
5319 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
5320 | 0 | "add $0x40,%0 \n" |
5321 | |
|
5322 | 0 | "vpsrlw %%xmm3,%%ymm0,%%ymm0 \n" |
5323 | 0 | "vpsrlw %%xmm3,%%ymm1,%%ymm1 \n" |
5324 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
5325 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
5326 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
5327 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
5328 | 0 | "vextractf128 $0x0,%%ymm0,(%1) \n" |
5329 | 0 | "vextractf128 $0x0,%%ymm1,0x10(%1) \n" |
5330 | 0 | "vextractf128 $0x1,%%ymm0,(%1,%2) \n" |
5331 | 0 | "vextractf128 $0x1,%%ymm1,0x10(%1,%2) \n" |
5332 | 0 | "add $0x20,%1 \n" |
5333 | 0 | "sub $0x10,%3 \n" |
5334 | 0 | "jg 1b \n" |
5335 | 0 | "vzeroupper \n" |
5336 | 0 | : "+r"(src_uv), // %0 |
5337 | 0 | "+r"(dst_u), // %1 |
5338 | 0 | "+r"(dst_v), // %2 |
5339 | 0 | "+r"(width) // %3 |
5340 | 0 | : "r"(depth), // %4 |
5341 | 0 | "m"(kSplitUVShuffle16) // %5 |
5342 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
5343 | 0 | } |
5344 | | #endif // HAS_SPLITUVROW_16_AVX2 |
5345 | | |
5346 | | // Use scale to convert lsb formats to msb, depending how many bits there are: |
5347 | | // 128 = 9 bits |
5348 | | // 64 = 10 bits |
5349 | | // 16 = 12 bits |
5350 | | // 1 = 16 bits |
5351 | | #ifdef HAS_MULTIPLYROW_16_AVX2 |
5352 | | void MultiplyRow_16_AVX2(const uint16_t* src_y, |
5353 | | uint16_t* dst_y, |
5354 | | int scale, |
5355 | 0 | int width) { |
5356 | 0 | asm volatile( |
5357 | 0 | "vmovd %3,%%xmm3 \n" |
5358 | 0 | "vpbroadcastw %%xmm3,%%ymm3 \n" |
5359 | 0 | "sub %0,%1 \n" |
5360 | | |
5361 | | // 32 pixels per loop. |
5362 | 0 | LABELALIGN |
5363 | 0 | "1: \n" |
5364 | 0 | "vmovdqu (%0),%%ymm0 \n" |
5365 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
5366 | 0 | "vpmullw %%ymm3,%%ymm0,%%ymm0 \n" |
5367 | 0 | "vpmullw %%ymm3,%%ymm1,%%ymm1 \n" |
5368 | 0 | "vmovdqu %%ymm0,(%0,%1) \n" |
5369 | 0 | "vmovdqu %%ymm1,0x20(%0,%1) \n" |
5370 | 0 | "add $0x40,%0 \n" |
5371 | 0 | "sub $0x20,%2 \n" |
5372 | 0 | "jg 1b \n" |
5373 | 0 | "vzeroupper \n" |
5374 | 0 | : "+r"(src_y), // %0 |
5375 | 0 | "+r"(dst_y), // %1 |
5376 | 0 | "+r"(width) // %2 |
5377 | 0 | : "r"(scale) // %3 |
5378 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm3"); |
5379 | 0 | } |
5380 | | #endif // HAS_MULTIPLYROW_16_AVX2 |
5381 | | |
5382 | | // Use scale to convert msb formats to lsb, depending how many bits there are: |
5383 | | // 512 = 9 bits |
5384 | | // 1024 = 10 bits |
5385 | | // 4096 = 12 bits |
5386 | | // 65536 = 16 bits |
5387 | | #ifdef HAS_DIVIDEROW_16_AVX2 |
5388 | | void DivideRow_16_AVX2(const uint16_t* src_y, |
5389 | | uint16_t* dst_y, |
5390 | | int scale, |
5391 | 0 | int width) { |
5392 | 0 | asm volatile( |
5393 | 0 | "vmovd %3,%%xmm3 \n" |
5394 | 0 | "vpbroadcastw %%xmm3,%%ymm3 \n" |
5395 | 0 | "sub %0,%1 \n" |
5396 | | |
5397 | | // 32 pixels per loop. |
5398 | 0 | LABELALIGN |
5399 | 0 | "1: \n" |
5400 | 0 | "vmovdqu (%0),%%ymm0 \n" |
5401 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
5402 | 0 | "vpmulhuw %%ymm3,%%ymm0,%%ymm0 \n" |
5403 | 0 | "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" |
5404 | 0 | "vmovdqu %%ymm0,(%0,%1) \n" |
5405 | 0 | "vmovdqu %%ymm1,0x20(%0,%1) \n" |
5406 | 0 | "add $0x40,%0 \n" |
5407 | 0 | "sub $0x20,%2 \n" |
5408 | 0 | "jg 1b \n" |
5409 | 0 | "vzeroupper \n" |
5410 | 0 | : "+r"(src_y), // %0 |
5411 | 0 | "+r"(dst_y), // %1 |
5412 | 0 | "+r"(width), // %2 |
5413 | 0 | "+r"(scale) // %3 |
5414 | 0 | : |
5415 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm3"); |
5416 | 0 | } |
5417 | | #endif // HAS_MULTIPLYROW_16_AVX2 |
5418 | | |
5419 | | // Use scale to convert lsb formats to msb, depending how many bits there are: |
5420 | | // 32768 = 9 bits |
5421 | | // 16384 = 10 bits |
5422 | | // 4096 = 12 bits |
5423 | | // 256 = 16 bits |
5424 | | void Convert16To8Row_SSSE3(const uint16_t* src_y, |
5425 | | uint8_t* dst_y, |
5426 | | int scale, |
5427 | 0 | int width) { |
5428 | 0 | asm volatile( |
5429 | 0 | "movd %3,%%xmm2 \n" |
5430 | 0 | "punpcklwd %%xmm2,%%xmm2 \n" |
5431 | 0 | "pshufd $0x0,%%xmm2,%%xmm2 \n" |
5432 | | |
5433 | | // 32 pixels per loop. |
5434 | 0 | LABELALIGN |
5435 | 0 | "1: \n" |
5436 | 0 | "movdqu (%0),%%xmm0 \n" |
5437 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5438 | 0 | "add $0x20,%0 \n" |
5439 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
5440 | 0 | "pmulhuw %%xmm2,%%xmm1 \n" |
5441 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
5442 | 0 | "movdqu %%xmm0,(%1) \n" |
5443 | 0 | "add $0x10,%1 \n" |
5444 | 0 | "sub $0x10,%2 \n" |
5445 | 0 | "jg 1b \n" |
5446 | 0 | : "+r"(src_y), // %0 |
5447 | 0 | "+r"(dst_y), // %1 |
5448 | 0 | "+r"(width) // %2 |
5449 | 0 | : "r"(scale) // %3 |
5450 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5451 | 0 | } |
5452 | | |
5453 | | #ifdef HAS_CONVERT16TO8ROW_AVX2 |
5454 | | void Convert16To8Row_AVX2(const uint16_t* src_y, |
5455 | | uint8_t* dst_y, |
5456 | | int scale, |
5457 | 724 | int width) { |
5458 | 724 | asm volatile( |
5459 | 724 | "vmovd %3,%%xmm2 \n" |
5460 | 724 | "vpbroadcastw %%xmm2,%%ymm2 \n" |
5461 | | |
5462 | | // 32 pixels per loop. |
5463 | 724 | LABELALIGN |
5464 | 724 | "1: \n" |
5465 | 724 | "vmovdqu (%0),%%ymm0 \n" |
5466 | 724 | "vmovdqu 0x20(%0),%%ymm1 \n" |
5467 | 724 | "add $0x40,%0 \n" |
5468 | 724 | "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" |
5469 | 724 | "vpmulhuw %%ymm2,%%ymm1,%%ymm1 \n" |
5470 | 724 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" // mutates |
5471 | 724 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
5472 | 724 | "vmovdqu %%ymm0,(%1) \n" |
5473 | 724 | "add $0x20,%1 \n" |
5474 | 724 | "sub $0x20,%2 \n" |
5475 | 724 | "jg 1b \n" |
5476 | 724 | "vzeroupper \n" |
5477 | 724 | : "+r"(src_y), // %0 |
5478 | 724 | "+r"(dst_y), // %1 |
5479 | 724 | "+r"(width) // %2 |
5480 | 724 | : "r"(scale) // %3 |
5481 | 724 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5482 | 724 | } |
5483 | | #endif // HAS_CONVERT16TO8ROW_AVX2 |
5484 | | |
5485 | | #ifdef HAS_CONVERT16TO8ROW_AVX512BW |
5486 | | void Convert16To8Row_AVX512BW(const uint16_t* src_y, |
5487 | | uint8_t* dst_y, |
5488 | | int scale, |
5489 | 0 | int width) { |
5490 | 0 | asm volatile("vpbroadcastw %3,%%zmm2 \n" |
5491 | | |
5492 | | // 64 pixels per loop. |
5493 | 0 | LABELALIGN |
5494 | 0 | "1: \n" |
5495 | 0 | "vmovups (%0),%%zmm0 \n" |
5496 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
5497 | 0 | "add $0x80,%0 \n" |
5498 | 0 | "vpmulhuw %%zmm2,%%zmm0,%%zmm0 \n" |
5499 | 0 | "vpmulhuw %%zmm2,%%zmm1,%%zmm1 \n" |
5500 | 0 | "vpmovuswb %%zmm0,%%ymm0 \n" |
5501 | 0 | "vpmovuswb %%zmm1,%%ymm1 \n" |
5502 | 0 | "vmovups %%ymm0,(%1) \n" |
5503 | 0 | "vmovups %%ymm1,0x20(%1) \n" |
5504 | 0 | "add $0x40,%1 \n" |
5505 | 0 | "sub $0x40,%2 \n" |
5506 | 0 | "jg 1b \n" |
5507 | 0 | "vzeroupper \n" |
5508 | 0 | : "+r"(src_y), // %0 |
5509 | 0 | "+r"(dst_y), // %1 |
5510 | 0 | "+r"(width) // %2 |
5511 | 0 | : "r"(scale) // %3 |
5512 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5513 | 0 | } |
5514 | | #endif // HAS_CONVERT16TO8ROW_AVX2 |
5515 | | |
5516 | | // Use scale to convert to lsb formats depending how many bits there are: |
5517 | | // 512 = 9 bits |
5518 | | // 1024 = 10 bits |
5519 | | // 4096 = 12 bits |
5520 | | void Convert8To16Row_SSE2(const uint8_t* src_y, |
5521 | | uint16_t* dst_y, |
5522 | | int scale, |
5523 | 0 | int width) { |
5524 | 0 | asm volatile( |
5525 | 0 | "movd %3,%%xmm2 \n" |
5526 | 0 | "punpcklwd %%xmm2,%%xmm2 \n" |
5527 | 0 | "pshufd $0x0,%%xmm2,%%xmm2 \n" |
5528 | | |
5529 | | // 32 pixels per loop. |
5530 | 0 | LABELALIGN |
5531 | 0 | "1: \n" |
5532 | 0 | "movdqu (%0),%%xmm0 \n" |
5533 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
5534 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
5535 | 0 | "punpckhbw %%xmm1,%%xmm1 \n" |
5536 | 0 | "add $0x10,%0 \n" |
5537 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
5538 | 0 | "pmulhuw %%xmm2,%%xmm1 \n" |
5539 | 0 | "movdqu %%xmm0,(%1) \n" |
5540 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
5541 | 0 | "add $0x20,%1 \n" |
5542 | 0 | "sub $0x10,%2 \n" |
5543 | 0 | "jg 1b \n" |
5544 | 0 | : "+r"(src_y), // %0 |
5545 | 0 | "+r"(dst_y), // %1 |
5546 | 0 | "+r"(width) // %2 |
5547 | 0 | : "r"(scale) // %3 |
5548 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5549 | 0 | } |
5550 | | |
5551 | | #ifdef HAS_CONVERT8TO16ROW_AVX2 |
5552 | | void Convert8To16Row_AVX2(const uint8_t* src_y, |
5553 | | uint16_t* dst_y, |
5554 | | int scale, |
5555 | 0 | int width) { |
5556 | 0 | const int shift = __builtin_clz(scale) - 15; |
5557 | 0 | asm volatile("vmovd %3,%%xmm2 \n" |
5558 | | |
5559 | | // 32 pixels per loop. |
5560 | 0 | LABELALIGN |
5561 | 0 | "1: \n" |
5562 | 0 | "vmovdqu (%0),%%ymm0 \n" |
5563 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
5564 | 0 | "add $0x20,%0 \n" |
5565 | 0 | "vpunpckhbw %%ymm0,%%ymm0,%%ymm1 \n" |
5566 | 0 | "vpunpcklbw %%ymm0,%%ymm0,%%ymm0 \n" |
5567 | 0 | "vpsrlw %%xmm2,%%ymm0,%%ymm0 \n" |
5568 | 0 | "vpsrlw %%xmm2,%%ymm1,%%ymm1 \n" |
5569 | 0 | "vmovdqu %%ymm0,(%1) \n" |
5570 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
5571 | 0 | "add $0x40,%1 \n" |
5572 | 0 | "sub $0x20,%2 \n" |
5573 | 0 | "jg 1b \n" |
5574 | 0 | "vzeroupper \n" |
5575 | 0 | : "+r"(src_y), // %0 |
5576 | 0 | "+r"(dst_y), // %1 |
5577 | 0 | "+r"(width) // %2 |
5578 | 0 | : "r"(shift) // %3 |
5579 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5580 | 0 | } |
5581 | | #endif // HAS_CONVERT8TO16ROW_AVX2 |
5582 | | |
5583 | | #ifdef HAS_SPLITRGBROW_SSSE3 |
5584 | | // Shuffle table for converting RGB to Planar. |
5585 | | static const uvec8 kSplitRGBShuffle[9] = { |
5586 | | {0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, |
5587 | | 128u, 128u}, |
5588 | | {128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u, |
5589 | | 128u, 128u}, |
5590 | | {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 1u, 4u, |
5591 | | 7u, 10u, 13u}, |
5592 | | {1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, |
5593 | | 128u, 128u}, |
5594 | | {128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u, 12u, 15u, 128u, 128u, 128u, |
5595 | | 128u, 128u}, |
5596 | | {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 2u, 5u, |
5597 | | 8u, 11u, 14u}, |
5598 | | {2u, 5u, 8u, 11u, 14u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, |
5599 | | 128u, 128u}, |
5600 | | {128u, 128u, 128u, 128u, 128u, 1u, 4u, 7u, 10u, 13u, 128u, 128u, 128u, 128u, |
5601 | | 128u, 128u}, |
5602 | | {128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 128u, 0u, 3u, 6u, 9u, |
5603 | | 12u, 15u}}; |
5604 | | |
5605 | | void SplitRGBRow_SSSE3(const uint8_t* src_rgb, |
5606 | | uint8_t* dst_r, |
5607 | | uint8_t* dst_g, |
5608 | | uint8_t* dst_b, |
5609 | 0 | int width) { |
5610 | 0 | asm volatile( |
5611 | 0 | "1: \n" |
5612 | 0 | "movdqu (%0),%%xmm0 \n" |
5613 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5614 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
5615 | 0 | "pshufb 0(%5), %%xmm0 \n" |
5616 | 0 | "pshufb 16(%5), %%xmm1 \n" |
5617 | 0 | "pshufb 32(%5), %%xmm2 \n" |
5618 | 0 | "por %%xmm1,%%xmm0 \n" |
5619 | 0 | "por %%xmm2,%%xmm0 \n" |
5620 | 0 | "movdqu %%xmm0,(%1) \n" |
5621 | 0 | "lea 0x10(%1),%1 \n" |
5622 | |
|
5623 | 0 | "movdqu (%0),%%xmm0 \n" |
5624 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5625 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
5626 | 0 | "pshufb 48(%5),%%xmm0 \n" |
5627 | 0 | "pshufb 64(%5),%%xmm1 \n" |
5628 | 0 | "pshufb 80(%5), %%xmm2 \n" |
5629 | 0 | "por %%xmm1,%%xmm0 \n" |
5630 | 0 | "por %%xmm2,%%xmm0 \n" |
5631 | 0 | "movdqu %%xmm0,(%2) \n" |
5632 | 0 | "lea 0x10(%2),%2 \n" |
5633 | |
|
5634 | 0 | "movdqu (%0),%%xmm0 \n" |
5635 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
5636 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
5637 | 0 | "pshufb 96(%5), %%xmm0 \n" |
5638 | 0 | "pshufb 112(%5), %%xmm1 \n" |
5639 | 0 | "pshufb 128(%5), %%xmm2 \n" |
5640 | 0 | "por %%xmm1,%%xmm0 \n" |
5641 | 0 | "por %%xmm2,%%xmm0 \n" |
5642 | 0 | "movdqu %%xmm0,(%3) \n" |
5643 | 0 | "lea 0x10(%3),%3 \n" |
5644 | 0 | "lea 0x30(%0),%0 \n" |
5645 | 0 | "sub $0x10,%4 \n" |
5646 | 0 | "jg 1b \n" |
5647 | 0 | : "+r"(src_rgb), // %0 |
5648 | 0 | "+r"(dst_r), // %1 |
5649 | 0 | "+r"(dst_g), // %2 |
5650 | 0 | "+r"(dst_b), // %3 |
5651 | 0 | "+r"(width) // %4 |
5652 | 0 | : "r"(&kSplitRGBShuffle[0]) // %5 |
5653 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5654 | 0 | } |
5655 | | #endif // HAS_SPLITRGBROW_SSSE3 |
5656 | | |
5657 | | #ifdef HAS_SPLITRGBROW_SSE41 |
5658 | | // Shuffle table for converting RGB to Planar, SSE4.1. Note: these are used for |
5659 | | // the AVX2 implementation as well. |
5660 | | static const uvec8 kSplitRGBShuffleSSE41[5] = { |
5661 | | {0u, 3u, 6u, 9u, 12u, 15u, 2u, 5u, 8u, 11u, 14u, 1u, 4u, 7u, 10u, 13u}, |
5662 | | {1u, 4u, 7u, 10u, 13u, 0u, 3u, 6u, 9u, 12u, 15u, 2u, 5u, 8u, 11u, 14u}, |
5663 | | {2u, 5u, 8u, 11u, 14u, 1u, 4u, 7u, 10u, 13u, 0u, 3u, 6u, 9u, 12u, 15u}, |
5664 | | {0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u}, |
5665 | | {0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u, 0u, 128u, 0u}, |
5666 | | }; |
5667 | | |
5668 | | void SplitRGBRow_SSE41(const uint8_t* src_rgb, |
5669 | | uint8_t* dst_r, |
5670 | | uint8_t* dst_g, |
5671 | | uint8_t* dst_b, |
5672 | 0 | int width) { |
5673 | 0 | asm volatile( |
5674 | 0 | "movdqa 48(%5), %%xmm0 \n" |
5675 | 0 | "1: \n" |
5676 | 0 | "movdqu (%0),%%xmm1 \n" |
5677 | 0 | "movdqu 0x10(%0),%%xmm2 \n" |
5678 | 0 | "movdqu 0x20(%0),%%xmm3 \n" |
5679 | 0 | "lea 0x30(%0),%0 \n" |
5680 | 0 | "movdqa %%xmm1, %%xmm4 \n" |
5681 | 0 | "pblendvb %%xmm3, %%xmm1 \n" |
5682 | 0 | "pblendvb %%xmm2, %%xmm3 \n" |
5683 | 0 | "pblendvb %%xmm4, %%xmm2 \n" |
5684 | 0 | "palignr $0xF, %%xmm0, %%xmm0 \n" |
5685 | 0 | "pblendvb %%xmm2, %%xmm1 \n" |
5686 | 0 | "pblendvb %%xmm3, %%xmm2 \n" |
5687 | 0 | "pblendvb %%xmm4, %%xmm3 \n" |
5688 | 0 | "palignr $0x1, %%xmm0, %%xmm0 \n" |
5689 | 0 | "pshufb 0(%5), %%xmm1 \n" |
5690 | 0 | "pshufb 16(%5), %%xmm2 \n" |
5691 | 0 | "pshufb 32(%5), %%xmm3 \n" |
5692 | 0 | "movdqu %%xmm1,(%1) \n" |
5693 | 0 | "lea 0x10(%1),%1 \n" |
5694 | 0 | "movdqu %%xmm2,(%2) \n" |
5695 | 0 | "lea 0x10(%2),%2 \n" |
5696 | 0 | "movdqu %%xmm3,(%3) \n" |
5697 | 0 | "lea 0x10(%3),%3 \n" |
5698 | 0 | "sub $0x10,%4 \n" |
5699 | 0 | "jg 1b \n" |
5700 | 0 | : "+r"(src_rgb), // %0 |
5701 | 0 | "+r"(dst_r), // %1 |
5702 | 0 | "+r"(dst_g), // %2 |
5703 | 0 | "+r"(dst_b), // %3 |
5704 | 0 | "+r"(width) // %4 |
5705 | 0 | : "r"(&kSplitRGBShuffleSSE41[0]) // %5 |
5706 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
5707 | 0 | } |
5708 | | #endif // HAS_SPLITRGBROW_SSE41 |
5709 | | |
5710 | | #ifdef HAS_SPLITRGBROW_AVX2 |
5711 | | void SplitRGBRow_AVX2(const uint8_t* src_rgb, |
5712 | | uint8_t* dst_r, |
5713 | | uint8_t* dst_g, |
5714 | | uint8_t* dst_b, |
5715 | 0 | int width) { |
5716 | 0 | asm volatile( |
5717 | 0 | "vbroadcasti128 48(%5), %%ymm0 \n" |
5718 | 0 | "vbroadcasti128 64(%5), %%ymm7 \n" |
5719 | 0 | #if defined(__x86_64__) |
5720 | 0 | "vbroadcasti128 0(%5), %%ymm8 \n" |
5721 | 0 | "vbroadcasti128 16(%5), %%ymm9 \n" |
5722 | 0 | "vbroadcasti128 32(%5), %%ymm10 \n" |
5723 | 0 | #endif |
5724 | 0 | "1: \n" |
5725 | 0 | "vmovdqu (%0),%%ymm4 \n" |
5726 | 0 | "vmovdqu 0x20(%0),%%ymm5 \n" |
5727 | 0 | "vmovdqu 0x40(%0),%%ymm6 \n" |
5728 | 0 | "lea 0x60(%0),%0 \n" |
5729 | 0 | "vpblendd $240, %%ymm5, %%ymm4, %%ymm1 \n" |
5730 | 0 | "vperm2i128 $33, %%ymm6, %%ymm4, %%ymm2 \n" |
5731 | 0 | "vpblendd $240, %%ymm6, %%ymm5, %%ymm3 \n" |
5732 | 0 | "vpblendvb %%ymm0, %%ymm3, %%ymm1, %%ymm4 \n" |
5733 | 0 | "vpblendvb %%ymm0, %%ymm1, %%ymm2, %%ymm5 \n" |
5734 | 0 | "vpblendvb %%ymm0, %%ymm2, %%ymm3, %%ymm6 \n" |
5735 | 0 | "vpblendvb %%ymm7, %%ymm5, %%ymm4, %%ymm1 \n" |
5736 | 0 | "vpblendvb %%ymm7, %%ymm6, %%ymm5, %%ymm2 \n" |
5737 | 0 | "vpblendvb %%ymm7, %%ymm4, %%ymm6, %%ymm3 \n" |
5738 | 0 | #if defined(__x86_64__) |
5739 | 0 | "vpshufb %%ymm8, %%ymm1, %%ymm1 \n" |
5740 | 0 | "vpshufb %%ymm9, %%ymm2, %%ymm2 \n" |
5741 | 0 | "vpshufb %%ymm10, %%ymm3, %%ymm3 \n" |
5742 | | #else |
5743 | | "vbroadcasti128 0(%5), %%ymm4 \n" |
5744 | | "vbroadcasti128 16(%5), %%ymm5 \n" |
5745 | | "vbroadcasti128 32(%5), %%ymm6 \n" |
5746 | | "vpshufb %%ymm4, %%ymm1, %%ymm1 \n" |
5747 | | "vpshufb %%ymm5, %%ymm2, %%ymm2 \n" |
5748 | | "vpshufb %%ymm6, %%ymm3, %%ymm3 \n" |
5749 | | #endif |
5750 | 0 | "vmovdqu %%ymm1,(%1) \n" |
5751 | 0 | "lea 0x20(%1),%1 \n" |
5752 | 0 | "vmovdqu %%ymm2,(%2) \n" |
5753 | 0 | "lea 0x20(%2),%2 \n" |
5754 | 0 | "vmovdqu %%ymm3,(%3) \n" |
5755 | 0 | "lea 0x20(%3),%3 \n" |
5756 | 0 | "sub $0x20,%4 \n" |
5757 | 0 | "jg 1b \n" |
5758 | 0 | : "+r"(src_rgb), // %0 |
5759 | 0 | "+r"(dst_r), // %1 |
5760 | 0 | "+r"(dst_g), // %2 |
5761 | 0 | "+r"(dst_b), // %3 |
5762 | 0 | "+r"(width) // %4 |
5763 | 0 | : "r"(&kSplitRGBShuffleSSE41[0]) // %5 |
5764 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
5765 | 0 | "xmm7" |
5766 | 0 | #if defined(__x86_64__) |
5767 | 0 | , |
5768 | 0 | "xmm8", "xmm9", "xmm10" |
5769 | 0 | #endif |
5770 | 0 | ); |
5771 | 0 | } |
5772 | | #endif // HAS_SPLITRGBROW_AVX2 |
5773 | | |
5774 | | #ifdef HAS_MERGERGBROW_SSSE3 |
5775 | | // Shuffle table for converting Planar to RGB. |
5776 | | static const uvec8 kMergeRGBShuffle[9] = { |
5777 | | {0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u, 128u, |
5778 | | 128u, 5u}, |
5779 | | {128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, 4u, |
5780 | | 128u, 128u}, |
5781 | | {128u, 128u, 0u, 128u, 128u, 1u, 128u, 128u, 2u, 128u, 128u, 3u, 128u, 128u, |
5782 | | 4u, 128u}, |
5783 | | {128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u, 128u, |
5784 | | 10u, 128u}, |
5785 | | {5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, 128u, |
5786 | | 128u, 10u}, |
5787 | | {128u, 5u, 128u, 128u, 6u, 128u, 128u, 7u, 128u, 128u, 8u, 128u, 128u, 9u, |
5788 | | 128u, 128u}, |
5789 | | {128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u, 128u, |
5790 | | 15u, 128u, 128u}, |
5791 | | {128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, 128u, |
5792 | | 128u, 15u, 128u}, |
5793 | | {10u, 128u, 128u, 11u, 128u, 128u, 12u, 128u, 128u, 13u, 128u, 128u, 14u, |
5794 | | 128u, 128u, 15u}}; |
5795 | | |
5796 | | void MergeRGBRow_SSSE3(const uint8_t* src_r, |
5797 | | const uint8_t* src_g, |
5798 | | const uint8_t* src_b, |
5799 | | uint8_t* dst_rgb, |
5800 | 0 | int width) { |
5801 | 0 | asm volatile( |
5802 | 0 | "1: \n" |
5803 | 0 | "movdqu (%0),%%xmm0 \n" |
5804 | 0 | "movdqu (%1),%%xmm1 \n" |
5805 | 0 | "movdqu (%2),%%xmm2 \n" |
5806 | 0 | "pshufb (%5), %%xmm0 \n" |
5807 | 0 | "pshufb 16(%5), %%xmm1 \n" |
5808 | 0 | "pshufb 32(%5), %%xmm2 \n" |
5809 | 0 | "por %%xmm1,%%xmm0 \n" |
5810 | 0 | "por %%xmm2,%%xmm0 \n" |
5811 | 0 | "movdqu %%xmm0,(%3) \n" |
5812 | |
|
5813 | 0 | "movdqu (%0),%%xmm0 \n" |
5814 | 0 | "movdqu (%1),%%xmm1 \n" |
5815 | 0 | "movdqu (%2),%%xmm2 \n" |
5816 | 0 | "pshufb 48(%5), %%xmm0 \n" |
5817 | 0 | "pshufb 64(%5), %%xmm1 \n" |
5818 | 0 | "pshufb 80(%5), %%xmm2 \n" |
5819 | 0 | "por %%xmm1,%%xmm0 \n" |
5820 | 0 | "por %%xmm2,%%xmm0 \n" |
5821 | 0 | "movdqu %%xmm0,16(%3) \n" |
5822 | |
|
5823 | 0 | "movdqu (%0),%%xmm0 \n" |
5824 | 0 | "movdqu (%1),%%xmm1 \n" |
5825 | 0 | "movdqu (%2),%%xmm2 \n" |
5826 | 0 | "pshufb 96(%5), %%xmm0 \n" |
5827 | 0 | "pshufb 112(%5), %%xmm1 \n" |
5828 | 0 | "pshufb 128(%5), %%xmm2 \n" |
5829 | 0 | "por %%xmm1,%%xmm0 \n" |
5830 | 0 | "por %%xmm2,%%xmm0 \n" |
5831 | 0 | "movdqu %%xmm0,32(%3) \n" |
5832 | |
|
5833 | 0 | "lea 0x10(%0),%0 \n" |
5834 | 0 | "lea 0x10(%1),%1 \n" |
5835 | 0 | "lea 0x10(%2),%2 \n" |
5836 | 0 | "lea 0x30(%3),%3 \n" |
5837 | 0 | "sub $0x10,%4 \n" |
5838 | 0 | "jg 1b \n" |
5839 | 0 | : "+r"(src_r), // %0 |
5840 | 0 | "+r"(src_g), // %1 |
5841 | 0 | "+r"(src_b), // %2 |
5842 | 0 | "+r"(dst_rgb), // %3 |
5843 | 0 | "+r"(width) // %4 |
5844 | 0 | : "r"(&kMergeRGBShuffle[0]) // %5 |
5845 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5846 | 0 | } |
5847 | | #endif // HAS_MERGERGBROW_SSSE3 |
5848 | | |
5849 | | #ifdef HAS_MERGEARGBROW_SSE2 |
5850 | | void MergeARGBRow_SSE2(const uint8_t* src_r, |
5851 | | const uint8_t* src_g, |
5852 | | const uint8_t* src_b, |
5853 | | const uint8_t* src_a, |
5854 | | uint8_t* dst_argb, |
5855 | 0 | int width) { |
5856 | 0 | asm volatile( |
5857 | 0 | "sub %0,%1 \n" |
5858 | 0 | "sub %0,%2 \n" |
5859 | 0 | "sub %0,%3 \n" |
5860 | |
|
5861 | 0 | LABELALIGN |
5862 | 0 | "1: \n" |
5863 | |
|
5864 | 0 | "movq (%0,%2),%%xmm0 \n" // B |
5865 | 0 | "movq (%0),%%xmm1 \n" // R |
5866 | 0 | "movq (%0,%1),%%xmm2 \n" // G |
5867 | 0 | "punpcklbw %%xmm1,%%xmm0 \n" // BR |
5868 | 0 | "movq (%0,%3),%%xmm1 \n" // A |
5869 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" // GA |
5870 | 0 | "movdqa %%xmm0,%%xmm1 \n" // BR |
5871 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // BGRA (hi) |
5872 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // BGRA (lo) |
5873 | 0 | "movdqu %%xmm0,(%4) \n" |
5874 | 0 | "movdqu %%xmm1,16(%4) \n" |
5875 | |
|
5876 | 0 | "lea 8(%0),%0 \n" |
5877 | 0 | "lea 32(%4),%4 \n" |
5878 | 0 | "sub $0x8,%5 \n" |
5879 | 0 | "jg 1b \n" |
5880 | 0 | : "+r"(src_r), // %0 |
5881 | 0 | "+r"(src_g), // %1 |
5882 | 0 | "+r"(src_b), // %2 |
5883 | 0 | "+r"(src_a), // %3 |
5884 | 0 | "+r"(dst_argb), // %4 |
5885 | 0 | "+r"(width) // %5 |
5886 | 0 | : |
5887 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5888 | 0 | } |
5889 | | #endif |
5890 | | |
5891 | | #ifdef HAS_MERGEXRGBROW_SSE2 |
5892 | | void MergeXRGBRow_SSE2(const uint8_t* src_r, |
5893 | | const uint8_t* src_g, |
5894 | | const uint8_t* src_b, |
5895 | | uint8_t* dst_argb, |
5896 | 0 | int width) { |
5897 | 0 | asm volatile( |
5898 | 0 | "1: \n" |
5899 | |
|
5900 | 0 | "movq (%2),%%xmm0 \n" // B |
5901 | 0 | "movq (%0),%%xmm1 \n" // R |
5902 | 0 | "movq (%1),%%xmm2 \n" // G |
5903 | 0 | "punpcklbw %%xmm1,%%xmm0 \n" // BR |
5904 | 0 | "pcmpeqd %%xmm1,%%xmm1 \n" // A(255) |
5905 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" // GA |
5906 | 0 | "movdqa %%xmm0,%%xmm1 \n" // BR |
5907 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // BGRA (hi) |
5908 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // BGRA (lo) |
5909 | 0 | "movdqu %%xmm0,(%3) \n" |
5910 | 0 | "movdqu %%xmm1,16(%3) \n" |
5911 | |
|
5912 | 0 | "lea 8(%0),%0 \n" |
5913 | 0 | "lea 8(%1),%1 \n" |
5914 | 0 | "lea 8(%2),%2 \n" |
5915 | 0 | "lea 32(%3),%3 \n" |
5916 | 0 | "sub $0x8,%4 \n" |
5917 | 0 | "jg 1b \n" |
5918 | 0 | : "+r"(src_r), // %0 |
5919 | 0 | "+r"(src_g), // %1 |
5920 | 0 | "+r"(src_b), // %2 |
5921 | 0 | "+r"(dst_argb), // %3 |
5922 | 0 | "+r"(width) // %4 |
5923 | 0 | : |
5924 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5925 | 0 | } |
5926 | | #endif // HAS_MERGEARGBROW_SSE2 |
5927 | | |
5928 | | #ifdef HAS_MERGEARGBROW_AVX2 |
5929 | | void MergeARGBRow_AVX2(const uint8_t* src_r, |
5930 | | const uint8_t* src_g, |
5931 | | const uint8_t* src_b, |
5932 | | const uint8_t* src_a, |
5933 | | uint8_t* dst_argb, |
5934 | 0 | int width) { |
5935 | 0 | asm volatile( |
5936 | 0 | "sub %0,%1 \n" |
5937 | 0 | "sub %0,%2 \n" |
5938 | 0 | "sub %0,%3 \n" |
5939 | |
|
5940 | 0 | LABELALIGN |
5941 | 0 | "1: \n" |
5942 | |
|
5943 | 0 | "vmovdqu (%0,%2),%%xmm0 \n" // B |
5944 | 0 | "vmovdqu (%0,%1),%%xmm1 \n" // R |
5945 | 0 | "vinserti128 $1,(%0),%%ymm0,%%ymm0 \n" // G |
5946 | 0 | "vinserti128 $1,(%0,%3),%%ymm1,%%ymm1 \n" // A |
5947 | 0 | "vpunpckhbw %%ymm1,%%ymm0,%%ymm2 \n" |
5948 | 0 | "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" |
5949 | 0 | "vperm2i128 $0x31,%%ymm2,%%ymm0,%%ymm1 \n" |
5950 | 0 | "vperm2i128 $0x20,%%ymm2,%%ymm0,%%ymm0 \n" |
5951 | 0 | "vpunpckhwd %%ymm1,%%ymm0,%%ymm2 \n" |
5952 | 0 | "vpunpcklwd %%ymm1,%%ymm0,%%ymm0 \n" |
5953 | 0 | "vperm2i128 $0x31,%%ymm2,%%ymm0,%%ymm1 \n" |
5954 | 0 | "vperm2i128 $0x20,%%ymm2,%%ymm0,%%ymm0 \n" |
5955 | 0 | "vmovdqu %%ymm0,(%4) \n" // First 8 |
5956 | 0 | "vmovdqu %%ymm1,32(%4) \n" // Next 8 |
5957 | |
|
5958 | 0 | "lea 16(%0),%0 \n" |
5959 | 0 | "lea 64(%4),%4 \n" |
5960 | 0 | "sub $0x10,%5 \n" |
5961 | 0 | "jg 1b \n" |
5962 | 0 | "vzeroupper \n" |
5963 | 0 | : "+r"(src_r), // %0 |
5964 | 0 | "+r"(src_g), // %1 |
5965 | 0 | "+r"(src_b), // %2 |
5966 | 0 | "+r"(src_a), // %3 |
5967 | 0 | "+r"(dst_argb), // %4 |
5968 | 0 | "+r"(width) // %5 |
5969 | 0 | : |
5970 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
5971 | 0 | } |
5972 | | #endif |
5973 | | |
5974 | | #ifdef HAS_MERGEXRGBROW_AVX2 |
5975 | | void MergeXRGBRow_AVX2(const uint8_t* src_r, |
5976 | | const uint8_t* src_g, |
5977 | | const uint8_t* src_b, |
5978 | | uint8_t* dst_argb, |
5979 | 0 | int width) { |
5980 | 0 | asm volatile( |
5981 | 0 | "1: \n" |
5982 | |
|
5983 | 0 | "vmovdqu (%2),%%xmm0 \n" // B |
5984 | 0 | "vpcmpeqb %%ymm1,%%ymm1,%%ymm1 \n" // A(255) |
5985 | 0 | "vinserti128 $0,(%1),%%ymm1,%%ymm1 \n" // R |
5986 | 0 | "vinserti128 $1,(%0),%%ymm0,%%ymm0 \n" // G |
5987 | 0 | "vpunpckhbw %%ymm1,%%ymm0,%%ymm2 \n" |
5988 | 0 | "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" |
5989 | 0 | "vperm2i128 $0x31,%%ymm2,%%ymm0,%%ymm1 \n" |
5990 | 0 | "vperm2i128 $0x20,%%ymm2,%%ymm0,%%ymm0 \n" |
5991 | 0 | "vpunpckhwd %%ymm1,%%ymm0,%%ymm2 \n" |
5992 | 0 | "vpunpcklwd %%ymm1,%%ymm0,%%ymm0 \n" |
5993 | 0 | "vperm2i128 $0x31,%%ymm2,%%ymm0,%%ymm1 \n" |
5994 | 0 | "vperm2i128 $0x20,%%ymm2,%%ymm0,%%ymm0 \n" |
5995 | 0 | "vmovdqu %%ymm0,(%3) \n" // First 8 |
5996 | 0 | "vmovdqu %%ymm1,32(%3) \n" // Next 8 |
5997 | |
|
5998 | 0 | "lea 16(%0),%0 \n" |
5999 | 0 | "lea 16(%1),%1 \n" |
6000 | 0 | "lea 16(%2),%2 \n" |
6001 | 0 | "lea 64(%3),%3 \n" |
6002 | 0 | "sub $0x10,%4 \n" |
6003 | 0 | "jg 1b \n" |
6004 | 0 | "vzeroupper \n" |
6005 | 0 | : "+r"(src_r), // %0 |
6006 | 0 | "+r"(src_g), // %1 |
6007 | 0 | "+r"(src_b), // %2 |
6008 | 0 | "+r"(dst_argb), // %3 |
6009 | 0 | "+rm"(width) // %4 |
6010 | 0 | ::"memory", |
6011 | 0 | "cc", "xmm0", "xmm1", "xmm2"); |
6012 | 0 | } |
6013 | | #endif // HAS_MERGEARGBROW_AVX2 |
6014 | | |
6015 | | #ifdef HAS_SPLITARGBROW_SSE2 |
6016 | | void SplitARGBRow_SSE2(const uint8_t* src_argb, |
6017 | | uint8_t* dst_r, |
6018 | | uint8_t* dst_g, |
6019 | | uint8_t* dst_b, |
6020 | | uint8_t* dst_a, |
6021 | 0 | int width) { |
6022 | 0 | asm volatile( |
6023 | 0 | "sub %1,%2 \n" |
6024 | 0 | "sub %1,%3 \n" |
6025 | 0 | "sub %1,%4 \n" |
6026 | |
|
6027 | 0 | LABELALIGN |
6028 | 0 | "1: \n" |
6029 | |
|
6030 | 0 | "movdqu (%0),%%xmm0 \n" // 00-0F |
6031 | 0 | "movdqu 16(%0),%%xmm1 \n" // 10-1F |
6032 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6033 | 0 | "punpcklqdq %%xmm1,%%xmm0 \n" // 00-07 10-17 |
6034 | 0 | "punpckhqdq %%xmm1,%%xmm2 \n" // 08-0F 18-1F |
6035 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
6036 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // 08192A3B4C5D6E7F (lo) |
6037 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // 08192A3B4C5D6E7F (hi) |
6038 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6039 | 0 | "punpcklqdq %%xmm1,%%xmm0 \n" // 08192A3B08192A3B |
6040 | 0 | "punpckhqdq %%xmm1,%%xmm2 \n" // 4C5D6E7F4C5D6E7F |
6041 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
6042 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // 048C159D26AE37BF (lo) |
6043 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // 048C159D26AE37BF (hi) |
6044 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6045 | 0 | "punpckldq %%xmm1,%%xmm0 \n" // 048C048C159D159D (BG) |
6046 | 0 | "punpckhdq %%xmm1,%%xmm2 \n" // 26AE26AE37BF37BF (RA) |
6047 | 0 | "movlps %%xmm0,(%1,%3) \n" // B |
6048 | 0 | "movhps %%xmm0,(%1,%2) \n" // G |
6049 | 0 | "movlps %%xmm2,(%1) \n" // R |
6050 | 0 | "movhps %%xmm2,(%1,%4) \n" // A |
6051 | |
|
6052 | 0 | "lea 32(%0),%0 \n" |
6053 | 0 | "lea 8(%1),%1 \n" |
6054 | 0 | "sub $0x8,%5 \n" |
6055 | 0 | "jg 1b \n" |
6056 | 0 | : "+r"(src_argb), // %0 |
6057 | 0 | "+r"(dst_r), // %1 |
6058 | 0 | "+r"(dst_g), // %2 |
6059 | 0 | "+r"(dst_b), // %3 |
6060 | 0 | "+r"(dst_a), // %4 |
6061 | 0 | "+rm"(width) // %5 |
6062 | 0 | : |
6063 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
6064 | 0 | } |
6065 | | #endif |
6066 | | |
6067 | | #ifdef HAS_SPLITXRGBROW_SSE2 |
6068 | | void SplitXRGBRow_SSE2(const uint8_t* src_argb, |
6069 | | uint8_t* dst_r, |
6070 | | uint8_t* dst_g, |
6071 | | uint8_t* dst_b, |
6072 | 0 | int width) { |
6073 | 0 | asm volatile( |
6074 | 0 | "1: \n" |
6075 | |
|
6076 | 0 | "movdqu (%0),%%xmm0 \n" // 00-0F |
6077 | 0 | "movdqu 16(%0),%%xmm1 \n" // 10-1F |
6078 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6079 | 0 | "punpcklqdq %%xmm1,%%xmm0 \n" // 00-07 10-17 |
6080 | 0 | "punpckhqdq %%xmm1,%%xmm2 \n" // 08-0F 18-1F |
6081 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
6082 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // 08192A3B4C5D6E7F (lo) |
6083 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // 08192A3B4C5D6E7F (hi) |
6084 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6085 | 0 | "punpcklqdq %%xmm1,%%xmm0 \n" // 08192A3B08192A3B |
6086 | 0 | "punpckhqdq %%xmm1,%%xmm2 \n" // 4C5D6E7F4C5D6E7F |
6087 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
6088 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" // 048C159D26AE37BF (lo) |
6089 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" // 048C159D26AE37BF (hi) |
6090 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6091 | 0 | "punpckldq %%xmm1,%%xmm0 \n" // 048C048C159D159D (BG) |
6092 | 0 | "punpckhdq %%xmm1,%%xmm2 \n" // 26AE26AE37BF37BF (RA) |
6093 | 0 | "movlps %%xmm0,(%3) \n" // B |
6094 | 0 | "movhps %%xmm0,(%2) \n" // G |
6095 | 0 | "movlps %%xmm2,(%1) \n" // R |
6096 | |
|
6097 | 0 | "lea 32(%0),%0 \n" |
6098 | 0 | "lea 8(%1),%1 \n" |
6099 | 0 | "lea 8(%2),%2 \n" |
6100 | 0 | "lea 8(%3),%3 \n" |
6101 | 0 | "sub $0x8,%4 \n" |
6102 | 0 | "jg 1b \n" |
6103 | 0 | : "+r"(src_argb), // %0 |
6104 | 0 | "+r"(dst_r), // %1 |
6105 | 0 | "+r"(dst_g), // %2 |
6106 | 0 | "+r"(dst_b), // %3 |
6107 | 0 | "+rm"(width) // %4 |
6108 | 0 | : |
6109 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
6110 | 0 | } |
6111 | | #endif |
6112 | | |
6113 | | static const uvec8 kShuffleMaskARGBSplit = {0, 4, 8, 12, 1, 5, 9, 13, |
6114 | | 2, 6, 10, 14, 3, 7, 11, 15}; |
6115 | | #ifdef HAS_SPLITARGBROW_SSSE3 |
6116 | | void SplitARGBRow_SSSE3(const uint8_t* src_argb, |
6117 | | uint8_t* dst_r, |
6118 | | uint8_t* dst_g, |
6119 | | uint8_t* dst_b, |
6120 | | uint8_t* dst_a, |
6121 | 0 | int width) { |
6122 | 0 | asm volatile( |
6123 | 0 | "movdqa %6,%%xmm3 \n" |
6124 | 0 | "sub %1,%2 \n" |
6125 | 0 | "sub %1,%3 \n" |
6126 | 0 | "sub %1,%4 \n" |
6127 | |
|
6128 | 0 | LABELALIGN |
6129 | 0 | "1: \n" |
6130 | |
|
6131 | 0 | "movdqu (%0),%%xmm0 \n" // 00-0F |
6132 | 0 | "movdqu 16(%0),%%xmm1 \n" // 10-1F |
6133 | 0 | "pshufb %%xmm3,%%xmm0 \n" // 048C159D26AE37BF (lo) |
6134 | 0 | "pshufb %%xmm3,%%xmm1 \n" // 048C159D26AE37BF (hi) |
6135 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6136 | 0 | "punpckldq %%xmm1,%%xmm0 \n" // 048C048C159D159D (BG) |
6137 | 0 | "punpckhdq %%xmm1,%%xmm2 \n" // 26AE26AE37BF37BF (RA) |
6138 | 0 | "movlps %%xmm0,(%1,%3) \n" // B |
6139 | 0 | "movhps %%xmm0,(%1,%2) \n" // G |
6140 | 0 | "movlps %%xmm2,(%1) \n" // R |
6141 | 0 | "movhps %%xmm2,(%1,%4) \n" // A |
6142 | |
|
6143 | 0 | "lea 32(%0),%0 \n" |
6144 | 0 | "lea 8(%1),%1 \n" |
6145 | 0 | "subl $0x8,%5 \n" |
6146 | 0 | "jg 1b \n" |
6147 | 0 | : "+r"(src_argb), // %0 |
6148 | 0 | "+r"(dst_r), // %1 |
6149 | 0 | "+r"(dst_g), // %2 |
6150 | 0 | "+r"(dst_b), // %3 |
6151 | 0 | "+r"(dst_a), // %4 |
6152 | | #if defined(__i386__) |
6153 | | "+m"(width) // %5 |
6154 | | #else |
6155 | 0 | "+rm"(width) // %5 |
6156 | 0 | #endif |
6157 | 0 | : "m"(kShuffleMaskARGBSplit) // %6 |
6158 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3"); |
6159 | 0 | } |
6160 | | #endif |
6161 | | |
6162 | | #ifdef HAS_SPLITXRGBROW_SSSE3 |
6163 | | void SplitXRGBRow_SSSE3(const uint8_t* src_argb, |
6164 | | uint8_t* dst_r, |
6165 | | uint8_t* dst_g, |
6166 | | uint8_t* dst_b, |
6167 | 0 | int width) { |
6168 | 0 | asm volatile( |
6169 | 0 | "movdqa %5,%%xmm3 \n" |
6170 | |
|
6171 | 0 | LABELALIGN |
6172 | 0 | "1: \n" |
6173 | |
|
6174 | 0 | "movdqu (%0),%%xmm0 \n" // 00-0F |
6175 | 0 | "movdqu 16(%0),%%xmm1 \n" // 10-1F |
6176 | 0 | "pshufb %%xmm3,%%xmm0 \n" // 048C159D26AE37BF (lo) |
6177 | 0 | "pshufb %%xmm3,%%xmm1 \n" // 048C159D26AE37BF (hi) |
6178 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
6179 | 0 | "punpckldq %%xmm1,%%xmm0 \n" // 048C048C159D159D (BG) |
6180 | 0 | "punpckhdq %%xmm1,%%xmm2 \n" // 26AE26AE37BF37BF (RA) |
6181 | 0 | "movlps %%xmm0,(%3) \n" // B |
6182 | 0 | "movhps %%xmm0,(%2) \n" // G |
6183 | 0 | "movlps %%xmm2,(%1) \n" // R |
6184 | |
|
6185 | 0 | "lea 32(%0),%0 \n" |
6186 | 0 | "lea 8(%1),%1 \n" |
6187 | 0 | "lea 8(%2),%2 \n" |
6188 | 0 | "lea 8(%3),%3 \n" |
6189 | 0 | "sub $0x8,%4 \n" |
6190 | 0 | "jg 1b \n" |
6191 | 0 | : "+r"(src_argb), // %0 |
6192 | 0 | "+r"(dst_r), // %1 |
6193 | 0 | "+r"(dst_g), // %2 |
6194 | 0 | "+r"(dst_b), // %3 |
6195 | 0 | "+r"(width) // %4 |
6196 | 0 | : "m"(kShuffleMaskARGBSplit) // %5 |
6197 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3"); |
6198 | 0 | } |
6199 | | #endif |
6200 | | |
6201 | | #ifdef HAS_SPLITARGBROW_AVX2 |
6202 | | static const ulvec32 kShuffleMaskARGBPermute = {0, 4, 1, 5, 2, 6, 3, 7}; |
6203 | | void SplitARGBRow_AVX2(const uint8_t* src_argb, |
6204 | | uint8_t* dst_r, |
6205 | | uint8_t* dst_g, |
6206 | | uint8_t* dst_b, |
6207 | | uint8_t* dst_a, |
6208 | 0 | int width) { |
6209 | 0 | asm volatile( |
6210 | 0 | "sub %1,%2 \n" |
6211 | 0 | "sub %1,%3 \n" |
6212 | 0 | "sub %1,%4 \n" |
6213 | 0 | "vmovdqa %7,%%ymm3 \n" |
6214 | 0 | "vbroadcasti128 %6,%%ymm4 \n" |
6215 | |
|
6216 | 0 | LABELALIGN |
6217 | 0 | "1: \n" |
6218 | |
|
6219 | 0 | "vmovdqu (%0),%%xmm0 \n" // 00-0F |
6220 | 0 | "vmovdqu 16(%0),%%xmm1 \n" // 10-1F |
6221 | 0 | "vinserti128 $1,32(%0),%%ymm0,%%ymm0 \n" // 00-0F 20-2F |
6222 | 0 | "vinserti128 $1,48(%0),%%ymm1,%%ymm1 \n" // 10-1F 30-3F |
6223 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
6224 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
6225 | 0 | "vpermd %%ymm0,%%ymm3,%%ymm0 \n" |
6226 | 0 | "vpermd %%ymm1,%%ymm3,%%ymm1 \n" |
6227 | 0 | "vpunpckhdq %%ymm1,%%ymm0,%%ymm2 \n" // GA |
6228 | 0 | "vpunpckldq %%ymm1,%%ymm0,%%ymm0 \n" // BR |
6229 | 0 | "vmovdqu %%xmm0,(%1,%3) \n" // B |
6230 | 0 | "vextracti128 $1,%%ymm0,(%1) \n" // R |
6231 | 0 | "vmovdqu %%xmm2,(%1,%2) \n" // G |
6232 | 0 | "vextracti128 $1,%%ymm2,(%1,%4) \n" // A |
6233 | 0 | "lea 64(%0),%0 \n" |
6234 | 0 | "lea 16(%1),%1 \n" |
6235 | 0 | "subl $0x10,%5 \n" |
6236 | 0 | "jg 1b \n" |
6237 | 0 | "vzeroupper \n" |
6238 | 0 | : "+r"(src_argb), // %0 |
6239 | 0 | "+r"(dst_r), // %1 |
6240 | 0 | "+r"(dst_g), // %2 |
6241 | 0 | "+r"(dst_b), // %3 |
6242 | 0 | "+r"(dst_a), // %4 |
6243 | | #if defined(__i386__) |
6244 | | "+m"(width) // %5 |
6245 | | #else |
6246 | 0 | "+rm"(width) // %5 |
6247 | 0 | #endif |
6248 | 0 | : "m"(kShuffleMaskARGBSplit), // %6 |
6249 | 0 | "m"(kShuffleMaskARGBPermute) // %7 |
6250 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
6251 | 0 | } |
6252 | | #endif |
6253 | | |
6254 | | #ifdef HAS_SPLITXRGBROW_AVX2 |
6255 | | void SplitXRGBRow_AVX2(const uint8_t* src_argb, |
6256 | | uint8_t* dst_r, |
6257 | | uint8_t* dst_g, |
6258 | | uint8_t* dst_b, |
6259 | 0 | int width) { |
6260 | 0 | asm volatile( |
6261 | 0 | "vmovdqa %6,%%ymm3 \n" |
6262 | 0 | "vbroadcasti128 %5,%%ymm4 \n" |
6263 | |
|
6264 | 0 | LABELALIGN |
6265 | 0 | "1: \n" |
6266 | |
|
6267 | 0 | "vmovdqu (%0),%%xmm0 \n" // 00-0F |
6268 | 0 | "vmovdqu 16(%0),%%xmm1 \n" // 10-1F |
6269 | 0 | "vinserti128 $1,32(%0),%%ymm0,%%ymm0 \n" // 00-0F 20-2F |
6270 | 0 | "vinserti128 $1,48(%0),%%ymm1,%%ymm1 \n" // 10-1F 30-3F |
6271 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" |
6272 | 0 | "vpshufb %%ymm4,%%ymm1,%%ymm1 \n" |
6273 | 0 | "vpermd %%ymm0,%%ymm3,%%ymm0 \n" |
6274 | 0 | "vpermd %%ymm1,%%ymm3,%%ymm1 \n" |
6275 | 0 | "vpunpckhdq %%ymm1,%%ymm0,%%ymm2 \n" // GA |
6276 | 0 | "vpunpckldq %%ymm1,%%ymm0,%%ymm0 \n" // BR |
6277 | 0 | "vmovdqu %%xmm0,(%3) \n" // B |
6278 | 0 | "vextracti128 $1,%%ymm0,(%1) \n" // R |
6279 | 0 | "vmovdqu %%xmm2,(%2) \n" // G |
6280 | |
|
6281 | 0 | "lea 64(%0),%0 \n" |
6282 | 0 | "lea 16(%1),%1 \n" |
6283 | 0 | "lea 16(%2),%2 \n" |
6284 | 0 | "lea 16(%3),%3 \n" |
6285 | 0 | "sub $0x10,%4 \n" |
6286 | 0 | "jg 1b \n" |
6287 | 0 | "vzeroupper \n" |
6288 | 0 | : "+r"(src_argb), // %0 |
6289 | 0 | "+r"(dst_r), // %1 |
6290 | 0 | "+r"(dst_g), // %2 |
6291 | 0 | "+r"(dst_b), // %3 |
6292 | 0 | "+r"(width) // %4 |
6293 | 0 | : "m"(kShuffleMaskARGBSplit), // %5 |
6294 | 0 | "m"(kShuffleMaskARGBPermute) // %6 |
6295 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
6296 | 0 | } |
6297 | | #endif |
6298 | | |
6299 | | #ifdef HAS_MERGEXR30ROW_AVX2 |
6300 | | void MergeXR30Row_AVX2(const uint16_t* src_r, |
6301 | | const uint16_t* src_g, |
6302 | | const uint16_t* src_b, |
6303 | | uint8_t* dst_ar30, |
6304 | | int depth, |
6305 | 0 | int width) { |
6306 | 0 | int shift = depth - 10; |
6307 | 0 | asm volatile( |
6308 | 0 | "sub %0,%1 \n" |
6309 | 0 | "sub %0,%2 \n" |
6310 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" // AR30 constants |
6311 | 0 | "vpsrlw $14,%%ymm5,%%ymm5 \n" |
6312 | 0 | "vpsllw $4,%%ymm5,%%ymm5 \n" // 2 alpha bits |
6313 | 0 | "vpcmpeqb %%ymm6,%%ymm6,%%ymm6 \n" |
6314 | 0 | "vpsrlw $6,%%ymm6,%%ymm6 \n" |
6315 | 0 | "vmovd %5,%%xmm4 \n" |
6316 | |
|
6317 | 0 | LABELALIGN |
6318 | 0 | "1: \n" |
6319 | 0 | "vmovdqu (%0),%%ymm0 \n" |
6320 | 0 | "vmovdqu (%0,%1),%%ymm1 \n" |
6321 | 0 | "vmovdqu (%0,%2),%%ymm2 \n" |
6322 | 0 | "vpsrlw %%xmm4,%%ymm0,%%ymm0 \n" |
6323 | 0 | "vpsrlw %%xmm4,%%ymm1,%%ymm1 \n" |
6324 | 0 | "vpsrlw %%xmm4,%%ymm2,%%ymm2 \n" |
6325 | 0 | "vpminuw %%ymm0,%%ymm6,%%ymm0 \n" |
6326 | 0 | "vpminuw %%ymm1,%%ymm6,%%ymm1 \n" |
6327 | 0 | "vpminuw %%ymm2,%%ymm6,%%ymm2 \n" |
6328 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
6329 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
6330 | 0 | "vpermq $0xd8,%%ymm2,%%ymm2 \n" |
6331 | 0 | "vpsllw $0x4,%%ymm0,%%ymm0 \n" // Shift R to target bit |
6332 | 0 | "vpunpckhwd %%ymm0,%%ymm2,%%ymm3 \n" // RB |
6333 | 0 | "vpunpcklwd %%ymm0,%%ymm2,%%ymm0 \n" |
6334 | 0 | "vpunpckhwd %%ymm5,%%ymm1,%%ymm2 \n" // AG |
6335 | 0 | "vpunpcklwd %%ymm5,%%ymm1,%%ymm1 \n" |
6336 | 0 | "vpslld $0xa,%%ymm1,%%ymm1 \n" // Shift AG to target bit |
6337 | 0 | "vpslld $0xa,%%ymm2,%%ymm2 \n" |
6338 | 0 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" // Combine |
6339 | 0 | "vpor %%ymm2,%%ymm3,%%ymm3 \n" |
6340 | 0 | "vmovdqu %%ymm0,(%3) \n" |
6341 | 0 | "vmovdqu %%ymm3,0x20(%3) \n" |
6342 | 0 | "lea 0x20(%0),%0 \n" |
6343 | 0 | "lea 0x40(%3),%3 \n" |
6344 | 0 | "sub $0x10,%4 \n" |
6345 | 0 | "jg 1b \n" |
6346 | 0 | "vzeroupper \n" |
6347 | 0 | : "+r"(src_r), // %0 |
6348 | 0 | "+r"(src_g), // %1 |
6349 | 0 | "+r"(src_b), // %2 |
6350 | 0 | "+r"(dst_ar30), // %3 |
6351 | 0 | "+r"(width) // %4 |
6352 | | #if defined(__i386__) |
6353 | | : "m"(shift) // %5 |
6354 | | #else |
6355 | 0 | : "rm"(shift) // %5 |
6356 | 0 | #endif |
6357 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
6358 | 0 | } |
6359 | | #endif |
6360 | | |
6361 | | #ifdef HAS_MERGEAR64ROW_AVX2 |
6362 | | static const lvec32 MergeAR64Permute = {0, 4, 2, 6, 1, 5, 3, 7}; |
6363 | | void MergeAR64Row_AVX2(const uint16_t* src_r, |
6364 | | const uint16_t* src_g, |
6365 | | const uint16_t* src_b, |
6366 | | const uint16_t* src_a, |
6367 | | uint16_t* dst_ar64, |
6368 | | int depth, |
6369 | 0 | int width) { |
6370 | 0 | int shift = 16 - depth; |
6371 | 0 | int mask = (1 << depth) - 1; |
6372 | 0 | mask = (mask << 16) + mask; |
6373 | 0 | asm volatile( |
6374 | 0 | "sub %0,%1 \n" |
6375 | 0 | "sub %0,%2 \n" |
6376 | 0 | "sub %0,%3 \n" |
6377 | 0 | "vmovdqa %8,%%ymm5 \n" |
6378 | 0 | "vmovd %6,%%xmm6 \n" |
6379 | 0 | "vbroadcastss %7,%%ymm7 \n" |
6380 | |
|
6381 | 0 | LABELALIGN |
6382 | 0 | "1: \n" |
6383 | 0 | "vmovdqu (%0),%%ymm0 \n" // R |
6384 | 0 | "vmovdqu (%0,%1),%%ymm1 \n" // G |
6385 | 0 | "vmovdqu (%0,%2),%%ymm2 \n" // B |
6386 | 0 | "vmovdqu (%0,%3),%%ymm3 \n" // A |
6387 | 0 | "vpminuw %%ymm0,%%ymm7,%%ymm0 \n" |
6388 | 0 | "vpminuw %%ymm1,%%ymm7,%%ymm1 \n" |
6389 | 0 | "vpminuw %%ymm2,%%ymm7,%%ymm2 \n" |
6390 | 0 | "vpminuw %%ymm3,%%ymm7,%%ymm3 \n" |
6391 | 0 | "vpsllw %%xmm6,%%ymm0,%%ymm0 \n" |
6392 | 0 | "vpsllw %%xmm6,%%ymm1,%%ymm1 \n" |
6393 | 0 | "vpsllw %%xmm6,%%ymm2,%%ymm2 \n" |
6394 | 0 | "vpsllw %%xmm6,%%ymm3,%%ymm3 \n" |
6395 | 0 | "vpermd %%ymm0,%%ymm5,%%ymm0 \n" |
6396 | 0 | "vpermd %%ymm1,%%ymm5,%%ymm1 \n" |
6397 | 0 | "vpermd %%ymm2,%%ymm5,%%ymm2 \n" |
6398 | 0 | "vpermd %%ymm3,%%ymm5,%%ymm3 \n" |
6399 | 0 | "vpunpcklwd %%ymm1,%%ymm2,%%ymm4 \n" // BG(low) |
6400 | 0 | "vpunpckhwd %%ymm1,%%ymm2,%%ymm1 \n" // BG(hi) |
6401 | 0 | "vpunpcklwd %%ymm3,%%ymm0,%%ymm2 \n" // RA(low) |
6402 | 0 | "vpunpckhwd %%ymm3,%%ymm0,%%ymm0 \n" // RA(hi) |
6403 | 0 | "vpunpckldq %%ymm2,%%ymm4,%%ymm3 \n" // BGRA(1) |
6404 | 0 | "vpunpckhdq %%ymm2,%%ymm4,%%ymm4 \n" // BGRA(3) |
6405 | 0 | "vpunpckldq %%ymm0,%%ymm1,%%ymm2 \n" // BGRA(2) |
6406 | 0 | "vpunpckhdq %%ymm0,%%ymm1,%%ymm1 \n" // BGRA(4) |
6407 | 0 | "vmovdqu %%ymm3,(%4) \n" |
6408 | 0 | "vmovdqu %%ymm2,0x20(%4) \n" |
6409 | 0 | "vmovdqu %%ymm4,0x40(%4) \n" |
6410 | 0 | "vmovdqu %%ymm1,0x60(%4) \n" |
6411 | 0 | "lea 0x20(%0),%0 \n" |
6412 | 0 | "lea 0x80(%4),%4 \n" |
6413 | 0 | "subl $0x10,%5 \n" |
6414 | 0 | "jg 1b \n" |
6415 | 0 | "vzeroupper \n" |
6416 | 0 | : "+r"(src_r), // %0 |
6417 | 0 | "+r"(src_g), // %1 |
6418 | 0 | "+r"(src_b), // %2 |
6419 | 0 | "+r"(src_a), // %3 |
6420 | 0 | "+r"(dst_ar64), // %4 |
6421 | | #if defined(__i386__) |
6422 | | "+m"(width) // %5 |
6423 | | #else |
6424 | 0 | "+rm"(width) // %5 |
6425 | 0 | #endif |
6426 | 0 | : "m"(shift), // %6 |
6427 | 0 | "m"(mask), // %7 |
6428 | 0 | "m"(MergeAR64Permute) // %8 |
6429 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
6430 | 0 | "xmm7"); |
6431 | 0 | } |
6432 | | #endif |
6433 | | |
6434 | | #ifdef HAS_MERGEXR64ROW_AVX2 |
6435 | | void MergeXR64Row_AVX2(const uint16_t* src_r, |
6436 | | const uint16_t* src_g, |
6437 | | const uint16_t* src_b, |
6438 | | uint16_t* dst_ar64, |
6439 | | int depth, |
6440 | 0 | int width) { |
6441 | 0 | int shift = 16 - depth; |
6442 | 0 | int mask = (1 << depth) - 1; |
6443 | 0 | mask = (mask << 16) + mask; |
6444 | 0 | asm volatile( |
6445 | 0 | "sub %0,%1 \n" |
6446 | 0 | "sub %0,%2 \n" |
6447 | 0 | "vmovdqa %7,%%ymm5 \n" |
6448 | 0 | "vmovd %5,%%xmm6 \n" |
6449 | 0 | "vbroadcastss %6,%%ymm7 \n" |
6450 | |
|
6451 | 0 | LABELALIGN |
6452 | 0 | "1: \n" |
6453 | 0 | "vmovdqu (%0),%%ymm0 \n" // R |
6454 | 0 | "vmovdqu (%0,%1),%%ymm1 \n" // G |
6455 | 0 | "vmovdqu (%0,%2),%%ymm2 \n" // B |
6456 | 0 | "vpminuw %%ymm0,%%ymm7,%%ymm0 \n" |
6457 | 0 | "vpminuw %%ymm1,%%ymm7,%%ymm1 \n" |
6458 | 0 | "vpminuw %%ymm2,%%ymm7,%%ymm2 \n" |
6459 | 0 | "vpsllw %%xmm6,%%ymm0,%%ymm0 \n" |
6460 | 0 | "vpsllw %%xmm6,%%ymm1,%%ymm1 \n" |
6461 | 0 | "vpsllw %%xmm6,%%ymm2,%%ymm2 \n" |
6462 | 0 | "vpermd %%ymm0,%%ymm5,%%ymm0 \n" |
6463 | 0 | "vpermd %%ymm1,%%ymm5,%%ymm1 \n" |
6464 | 0 | "vpermd %%ymm2,%%ymm5,%%ymm2 \n" |
6465 | 0 | "vpcmpeqb %%ymm3,%%ymm3,%%ymm3 \n" // A (0xffff) |
6466 | 0 | "vpunpcklwd %%ymm1,%%ymm2,%%ymm4 \n" // BG(low) |
6467 | 0 | "vpunpckhwd %%ymm1,%%ymm2,%%ymm1 \n" // BG(hi) |
6468 | 0 | "vpunpcklwd %%ymm3,%%ymm0,%%ymm2 \n" // RA(low) |
6469 | 0 | "vpunpckhwd %%ymm3,%%ymm0,%%ymm0 \n" // RA(hi) |
6470 | 0 | "vpunpckldq %%ymm2,%%ymm4,%%ymm3 \n" // BGRA(1) |
6471 | 0 | "vpunpckhdq %%ymm2,%%ymm4,%%ymm4 \n" // BGRA(3) |
6472 | 0 | "vpunpckldq %%ymm0,%%ymm1,%%ymm2 \n" // BGRA(2) |
6473 | 0 | "vpunpckhdq %%ymm0,%%ymm1,%%ymm1 \n" // BGRA(4) |
6474 | 0 | "vmovdqu %%ymm3,(%3) \n" |
6475 | 0 | "vmovdqu %%ymm2,0x20(%3) \n" |
6476 | 0 | "vmovdqu %%ymm4,0x40(%3) \n" |
6477 | 0 | "vmovdqu %%ymm1,0x60(%3) \n" |
6478 | 0 | "lea 0x20(%0),%0 \n" |
6479 | 0 | "lea 0x80(%3),%3 \n" |
6480 | 0 | "subl $0x10,%4 \n" |
6481 | 0 | "jg 1b \n" |
6482 | 0 | "vzeroupper \n" |
6483 | 0 | : "+r"(src_r), // %0 |
6484 | 0 | "+r"(src_g), // %1 |
6485 | 0 | "+r"(src_b), // %2 |
6486 | 0 | "+r"(dst_ar64), // %3 |
6487 | 0 | "+r"(width) // %4 |
6488 | 0 | : "m"(shift), // %5 |
6489 | 0 | "m"(mask), // %6 |
6490 | 0 | "m"(MergeAR64Permute) // %7 |
6491 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
6492 | 0 | "xmm7"); |
6493 | 0 | } |
6494 | | #endif |
6495 | | |
6496 | | #ifdef HAS_MERGEARGB16TO8ROW_AVX2 |
6497 | | static const uvec8 MergeARGB16To8Shuffle = {0, 8, 1, 9, 2, 10, 3, 11, |
6498 | | 4, 12, 5, 13, 6, 14, 7, 15}; |
6499 | | void MergeARGB16To8Row_AVX2(const uint16_t* src_r, |
6500 | | const uint16_t* src_g, |
6501 | | const uint16_t* src_b, |
6502 | | const uint16_t* src_a, |
6503 | | uint8_t* dst_argb, |
6504 | | int depth, |
6505 | 0 | int width) { |
6506 | 0 | int shift = depth - 8; |
6507 | 0 | asm volatile( |
6508 | 0 | "sub %0,%1 \n" |
6509 | 0 | "sub %0,%2 \n" |
6510 | 0 | "sub %0,%3 \n" |
6511 | 0 | "vbroadcasti128 %7,%%ymm5 \n" |
6512 | 0 | "vmovd %6,%%xmm6 \n" |
6513 | |
|
6514 | 0 | LABELALIGN |
6515 | 0 | "1: \n" |
6516 | 0 | "vmovdqu (%0),%%ymm0 \n" // R |
6517 | 0 | "vmovdqu (%0,%1),%%ymm1 \n" // G |
6518 | 0 | "vmovdqu (%0,%2),%%ymm2 \n" // B |
6519 | 0 | "vmovdqu (%0,%3),%%ymm3 \n" // A |
6520 | 0 | "vpsrlw %%xmm6,%%ymm0,%%ymm0 \n" |
6521 | 0 | "vpsrlw %%xmm6,%%ymm1,%%ymm1 \n" |
6522 | 0 | "vpsrlw %%xmm6,%%ymm2,%%ymm2 \n" |
6523 | 0 | "vpsrlw %%xmm6,%%ymm3,%%ymm3 \n" |
6524 | 0 | "vpackuswb %%ymm1,%%ymm2,%%ymm1 \n" // BG (planar) |
6525 | 0 | "vpackuswb %%ymm3,%%ymm0,%%ymm0 \n" // RA (planar) |
6526 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" // BG (interleave) |
6527 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" // RA (interleave) |
6528 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
6529 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
6530 | 0 | "vpunpcklwd %%ymm0,%%ymm1,%%ymm2 \n" // BGRA (low) |
6531 | 0 | "vpunpckhwd %%ymm0,%%ymm1,%%ymm0 \n" // BGRA (hi) |
6532 | 0 | "vmovdqu %%ymm2,(%4) \n" |
6533 | 0 | "vmovdqu %%ymm0,0x20(%4) \n" |
6534 | 0 | "lea 0x20(%0),%0 \n" |
6535 | 0 | "lea 0x40(%4),%4 \n" |
6536 | 0 | "subl $0x10,%5 \n" |
6537 | 0 | "jg 1b \n" |
6538 | 0 | "vzeroupper \n" |
6539 | 0 | : "+r"(src_r), // %0 |
6540 | 0 | "+r"(src_g), // %1 |
6541 | 0 | "+r"(src_b), // %2 |
6542 | 0 | "+r"(src_a), // %3 |
6543 | 0 | "+r"(dst_argb), // %4 |
6544 | | #if defined(__i386__) |
6545 | | "+m"(width) // %5 |
6546 | | #else |
6547 | 0 | "+rm"(width) // %5 |
6548 | 0 | #endif |
6549 | 0 | : "m"(shift), // %6 |
6550 | 0 | "m"(MergeARGB16To8Shuffle) // %7 |
6551 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
6552 | 0 | } |
6553 | | #endif |
6554 | | |
6555 | | #ifdef HAS_MERGEXRGB16TO8ROW_AVX2 |
6556 | | void MergeXRGB16To8Row_AVX2(const uint16_t* src_r, |
6557 | | const uint16_t* src_g, |
6558 | | const uint16_t* src_b, |
6559 | | uint8_t* dst_argb, |
6560 | | int depth, |
6561 | 0 | int width) { |
6562 | 0 | int shift = depth - 8; |
6563 | 0 | asm volatile( |
6564 | 0 | "sub %0,%1 \n" |
6565 | 0 | "sub %0,%2 \n" |
6566 | 0 | "vbroadcasti128 %6,%%ymm5 \n" |
6567 | 0 | "vmovd %5,%%xmm6 \n" |
6568 | 0 | "vpcmpeqb %%ymm3,%%ymm3,%%ymm3 \n" |
6569 | 0 | "vpsrlw $8,%%ymm3,%%ymm3 \n" // A (0xff) |
6570 | |
|
6571 | 0 | LABELALIGN |
6572 | 0 | "1: \n" |
6573 | 0 | "vmovdqu (%0),%%ymm0 \n" // R |
6574 | 0 | "vmovdqu (%0,%1),%%ymm1 \n" // G |
6575 | 0 | "vmovdqu (%0,%2),%%ymm2 \n" // B |
6576 | 0 | "vpsrlw %%xmm6,%%ymm0,%%ymm0 \n" |
6577 | 0 | "vpsrlw %%xmm6,%%ymm1,%%ymm1 \n" |
6578 | 0 | "vpsrlw %%xmm6,%%ymm2,%%ymm2 \n" |
6579 | 0 | "vpackuswb %%ymm1,%%ymm2,%%ymm1 \n" // BG (planar) |
6580 | 0 | "vpackuswb %%ymm3,%%ymm0,%%ymm0 \n" // RA (planar) |
6581 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" // BG (interleave) |
6582 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" // RA (interleave) |
6583 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
6584 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
6585 | 0 | "vpunpcklwd %%ymm0,%%ymm1,%%ymm2 \n" // BGRA (low) |
6586 | 0 | "vpunpckhwd %%ymm0,%%ymm1,%%ymm0 \n" // BGRA (hi) |
6587 | 0 | "vmovdqu %%ymm2,(%3) \n" |
6588 | 0 | "vmovdqu %%ymm0,0x20(%3) \n" |
6589 | 0 | "lea 0x20(%0),%0 \n" |
6590 | 0 | "lea 0x40(%3),%3 \n" |
6591 | 0 | "subl $0x10,%4 \n" |
6592 | 0 | "jg 1b \n" |
6593 | 0 | "vzeroupper \n" |
6594 | 0 | : "+r"(src_r), // %0 |
6595 | 0 | "+r"(src_g), // %1 |
6596 | 0 | "+r"(src_b), // %2 |
6597 | 0 | "+r"(dst_argb), // %3 |
6598 | 0 | "+r"(width) // %4 |
6599 | 0 | : "m"(shift), // %5 |
6600 | 0 | "m"(MergeARGB16To8Shuffle) // %6 |
6601 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
6602 | 0 | } |
6603 | | #endif |
6604 | | |
6605 | | #ifdef HAS_COPYROW_SSE2 |
6606 | 0 | void CopyRow_SSE2(const uint8_t* src, uint8_t* dst, int width) { |
6607 | 0 | asm volatile( |
6608 | 0 | "test $0xf,%0 \n" |
6609 | 0 | "jne 2f \n" |
6610 | 0 | "test $0xf,%1 \n" |
6611 | 0 | "jne 2f \n" |
6612 | |
|
6613 | 0 | LABELALIGN |
6614 | 0 | "1: \n" |
6615 | 0 | "movdqa (%0),%%xmm0 \n" |
6616 | 0 | "movdqa 0x10(%0),%%xmm1 \n" |
6617 | 0 | "lea 0x20(%0),%0 \n" |
6618 | 0 | "movdqa %%xmm0,(%1) \n" |
6619 | 0 | "movdqa %%xmm1,0x10(%1) \n" |
6620 | 0 | "lea 0x20(%1),%1 \n" |
6621 | 0 | "sub $0x20,%2 \n" |
6622 | 0 | "jg 1b \n" |
6623 | 0 | "jmp 9f \n" |
6624 | |
|
6625 | 0 | LABELALIGN |
6626 | 0 | "2: \n" |
6627 | 0 | "movdqu (%0),%%xmm0 \n" |
6628 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
6629 | 0 | "lea 0x20(%0),%0 \n" |
6630 | 0 | "movdqu %%xmm0,(%1) \n" |
6631 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
6632 | 0 | "lea 0x20(%1),%1 \n" |
6633 | 0 | "sub $0x20,%2 \n" |
6634 | 0 | "jg 2b \n" |
6635 | |
|
6636 | 0 | LABELALIGN "9: \n" |
6637 | 0 | : "+r"(src), // %0 |
6638 | 0 | "+r"(dst), // %1 |
6639 | 0 | "+r"(width) // %2 |
6640 | 0 | : |
6641 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
6642 | 0 | } |
6643 | | #endif // HAS_COPYROW_SSE2 |
6644 | | |
6645 | | #ifdef HAS_COPYROW_AVX |
6646 | 0 | void CopyRow_AVX(const uint8_t* src, uint8_t* dst, int width) { |
6647 | 0 | asm volatile( |
6648 | 0 | "1: \n" |
6649 | 0 | "vmovdqu (%0),%%ymm0 \n" |
6650 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
6651 | 0 | "lea 0x40(%0),%0 \n" |
6652 | 0 | "vmovdqu %%ymm0,(%1) \n" |
6653 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
6654 | 0 | "lea 0x40(%1),%1 \n" |
6655 | 0 | "sub $0x40,%2 \n" |
6656 | 0 | "jg 1b \n" |
6657 | 0 | "vzeroupper \n" |
6658 | 0 | : "+r"(src), // %0 |
6659 | 0 | "+r"(dst), // %1 |
6660 | 0 | "+r"(width) // %2 |
6661 | 0 | : |
6662 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
6663 | 0 | } |
6664 | | #endif // HAS_COPYROW_AVX |
6665 | | |
6666 | | #ifdef HAS_COPYROW_AVX512BW |
6667 | 0 | void CopyRow_AVX512BW(const uint8_t* src, uint8_t* dst, int width) { |
6668 | 0 | asm volatile( |
6669 | 0 | "1: \n" |
6670 | 0 | "vmovups (%0),%%zmm0 \n" |
6671 | 0 | "vmovups 0x40(%0),%%zmm1 \n" |
6672 | 0 | "lea 0x80(%0),%0 \n" |
6673 | 0 | "vmovups %%zmm0,(%1) \n" |
6674 | 0 | "vmovups %%zmm1,0x40(%1) \n" |
6675 | 0 | "lea 0x80(%1),%1 \n" |
6676 | 0 | "sub $0x80,%2 \n" |
6677 | 0 | "jg 1b \n" |
6678 | 0 | "vzeroupper \n" |
6679 | 0 | : "+r"(src), // %0 |
6680 | 0 | "+r"(dst), // %1 |
6681 | 0 | "+r"(width) // %2 |
6682 | 0 | : |
6683 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
6684 | 0 | } |
6685 | | #endif // HAS_COPYROW_AVX512 |
6686 | | |
6687 | | #ifdef HAS_COPYROW_ERMS |
6688 | | // Multiple of 1. |
6689 | 22.6k | void CopyRow_ERMS(const uint8_t* src, uint8_t* dst, int width) { |
6690 | 22.6k | size_t width_tmp = (size_t)(width); |
6691 | 22.6k | asm volatile("rep movsb \n" |
6692 | 22.6k | : "+S"(src), // %0 |
6693 | 22.6k | "+D"(dst), // %1 |
6694 | 22.6k | "+c"(width_tmp) // %2 |
6695 | 22.6k | : |
6696 | 22.6k | : "memory", "cc"); |
6697 | 22.6k | } |
6698 | | #endif // HAS_COPYROW_ERMS |
6699 | | |
6700 | | #ifdef HAS_ARGBCOPYALPHAROW_SSE2 |
6701 | | // width in pixels |
6702 | 0 | void ARGBCopyAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) { |
6703 | 0 | asm volatile( |
6704 | 0 | "pcmpeqb %%xmm0,%%xmm0 \n" |
6705 | 0 | "pslld $0x18,%%xmm0 \n" |
6706 | 0 | "pcmpeqb %%xmm1,%%xmm1 \n" |
6707 | 0 | "psrld $0x8,%%xmm1 \n" |
6708 | |
|
6709 | 0 | LABELALIGN |
6710 | 0 | "1: \n" |
6711 | 0 | "movdqu (%0),%%xmm2 \n" |
6712 | 0 | "movdqu 0x10(%0),%%xmm3 \n" |
6713 | 0 | "lea 0x20(%0),%0 \n" |
6714 | 0 | "movdqu (%1),%%xmm4 \n" |
6715 | 0 | "movdqu 0x10(%1),%%xmm5 \n" |
6716 | 0 | "pand %%xmm0,%%xmm2 \n" |
6717 | 0 | "pand %%xmm0,%%xmm3 \n" |
6718 | 0 | "pand %%xmm1,%%xmm4 \n" |
6719 | 0 | "pand %%xmm1,%%xmm5 \n" |
6720 | 0 | "por %%xmm4,%%xmm2 \n" |
6721 | 0 | "por %%xmm5,%%xmm3 \n" |
6722 | 0 | "movdqu %%xmm2,(%1) \n" |
6723 | 0 | "movdqu %%xmm3,0x10(%1) \n" |
6724 | 0 | "lea 0x20(%1),%1 \n" |
6725 | 0 | "sub $0x8,%2 \n" |
6726 | 0 | "jg 1b \n" |
6727 | 0 | : "+r"(src), // %0 |
6728 | 0 | "+r"(dst), // %1 |
6729 | 0 | "+r"(width) // %2 |
6730 | 0 | : |
6731 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
6732 | 0 | } |
6733 | | #endif // HAS_ARGBCOPYALPHAROW_SSE2 |
6734 | | |
6735 | | #ifdef HAS_ARGBCOPYALPHAROW_AVX2 |
6736 | | // width in pixels |
6737 | 0 | void ARGBCopyAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
6738 | 0 | asm volatile( |
6739 | 0 | "vpcmpeqb %%ymm0,%%ymm0,%%ymm0 \n" |
6740 | 0 | "vpsrld $0x8,%%ymm0,%%ymm0 \n" |
6741 | |
|
6742 | 0 | LABELALIGN |
6743 | 0 | "1: \n" |
6744 | 0 | "vmovdqu (%0),%%ymm1 \n" |
6745 | 0 | "vmovdqu 0x20(%0),%%ymm2 \n" |
6746 | 0 | "lea 0x40(%0),%0 \n" |
6747 | 0 | "vpblendvb %%ymm0,(%1),%%ymm1,%%ymm1 \n" |
6748 | 0 | "vpblendvb %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n" |
6749 | 0 | "vmovdqu %%ymm1,(%1) \n" |
6750 | 0 | "vmovdqu %%ymm2,0x20(%1) \n" |
6751 | 0 | "lea 0x40(%1),%1 \n" |
6752 | 0 | "sub $0x10,%2 \n" |
6753 | 0 | "jg 1b \n" |
6754 | 0 | "vzeroupper \n" |
6755 | 0 | : "+r"(src), // %0 |
6756 | 0 | "+r"(dst), // %1 |
6757 | 0 | "+r"(width) // %2 |
6758 | 0 | : |
6759 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
6760 | 0 | } |
6761 | | #endif // HAS_ARGBCOPYALPHAROW_AVX2 |
6762 | | |
6763 | | #ifdef HAS_ARGBEXTRACTALPHAROW_SSE2 |
6764 | | // width in pixels |
6765 | | void ARGBExtractAlphaRow_SSE2(const uint8_t* src_argb, |
6766 | | uint8_t* dst_a, |
6767 | 0 | int width) { |
6768 | 0 | asm volatile( |
6769 | 0 | "1: \n" |
6770 | 0 | "movdqu (%0), %%xmm0 \n" |
6771 | 0 | "movdqu 0x10(%0), %%xmm1 \n" |
6772 | 0 | "lea 0x20(%0), %0 \n" |
6773 | 0 | "psrld $0x18, %%xmm0 \n" |
6774 | 0 | "psrld $0x18, %%xmm1 \n" |
6775 | 0 | "packssdw %%xmm1, %%xmm0 \n" |
6776 | 0 | "packuswb %%xmm0, %%xmm0 \n" |
6777 | 0 | "movq %%xmm0,(%1) \n" |
6778 | 0 | "lea 0x8(%1), %1 \n" |
6779 | 0 | "sub $0x8, %2 \n" |
6780 | 0 | "jg 1b \n" |
6781 | 0 | : "+r"(src_argb), // %0 |
6782 | 0 | "+r"(dst_a), // %1 |
6783 | 0 | "+rm"(width) // %2 |
6784 | 0 | : |
6785 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
6786 | 0 | } |
6787 | | #endif // HAS_ARGBEXTRACTALPHAROW_SSE2 |
6788 | | |
6789 | | #ifdef HAS_ARGBEXTRACTALPHAROW_AVX2 |
6790 | | static const uvec8 kShuffleAlphaShort_AVX2 = { |
6791 | | 3u, 128u, 128u, 128u, 7u, 128u, 128u, 128u, |
6792 | | 11u, 128u, 128u, 128u, 15u, 128u, 128u, 128u}; |
6793 | | |
6794 | | void ARGBExtractAlphaRow_AVX2(const uint8_t* src_argb, |
6795 | | uint8_t* dst_a, |
6796 | 0 | int width) { |
6797 | 0 | asm volatile( |
6798 | 0 | "vmovdqa %3,%%ymm4 \n" |
6799 | 0 | "vbroadcasti128 %4,%%ymm5 \n" |
6800 | |
|
6801 | 0 | LABELALIGN |
6802 | 0 | "1: \n" |
6803 | 0 | "vmovdqu (%0), %%ymm0 \n" |
6804 | 0 | "vmovdqu 0x20(%0), %%ymm1 \n" |
6805 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" // vpsrld $0x18, %%ymm0 |
6806 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" |
6807 | 0 | "vmovdqu 0x40(%0), %%ymm2 \n" |
6808 | 0 | "vmovdqu 0x60(%0), %%ymm3 \n" |
6809 | 0 | "lea 0x80(%0), %0 \n" |
6810 | 0 | "vpackssdw %%ymm1, %%ymm0, %%ymm0 \n" // mutates |
6811 | 0 | "vpshufb %%ymm5,%%ymm2,%%ymm2 \n" |
6812 | 0 | "vpshufb %%ymm5,%%ymm3,%%ymm3 \n" |
6813 | 0 | "vpackssdw %%ymm3, %%ymm2, %%ymm2 \n" // mutates |
6814 | 0 | "vpackuswb %%ymm2,%%ymm0,%%ymm0 \n" // mutates. |
6815 | 0 | "vpermd %%ymm0,%%ymm4,%%ymm0 \n" // unmutate. |
6816 | 0 | "vmovdqu %%ymm0,(%1) \n" |
6817 | 0 | "lea 0x20(%1),%1 \n" |
6818 | 0 | "sub $0x20, %2 \n" |
6819 | 0 | "jg 1b \n" |
6820 | 0 | "vzeroupper \n" |
6821 | 0 | : "+r"(src_argb), // %0 |
6822 | 0 | "+r"(dst_a), // %1 |
6823 | 0 | "+rm"(width) // %2 |
6824 | 0 | : "m"(kPermdARGBToY_AVX), // %3 |
6825 | 0 | "m"(kShuffleAlphaShort_AVX2) // %4 |
6826 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
6827 | 0 | } |
6828 | | #endif // HAS_ARGBEXTRACTALPHAROW_AVX2 |
6829 | | |
6830 | | #ifdef HAS_ARGBCOPYYTOALPHAROW_SSE2 |
6831 | | // width in pixels |
6832 | 0 | void ARGBCopyYToAlphaRow_SSE2(const uint8_t* src, uint8_t* dst, int width) { |
6833 | 0 | asm volatile( |
6834 | 0 | "pcmpeqb %%xmm0,%%xmm0 \n" |
6835 | 0 | "pslld $0x18,%%xmm0 \n" |
6836 | 0 | "pcmpeqb %%xmm1,%%xmm1 \n" |
6837 | 0 | "psrld $0x8,%%xmm1 \n" |
6838 | |
|
6839 | 0 | LABELALIGN |
6840 | 0 | "1: \n" |
6841 | 0 | "movq (%0),%%xmm2 \n" |
6842 | 0 | "lea 0x8(%0),%0 \n" |
6843 | 0 | "punpcklbw %%xmm2,%%xmm2 \n" |
6844 | 0 | "punpckhwd %%xmm2,%%xmm3 \n" |
6845 | 0 | "punpcklwd %%xmm2,%%xmm2 \n" |
6846 | 0 | "movdqu (%1),%%xmm4 \n" |
6847 | 0 | "movdqu 0x10(%1),%%xmm5 \n" |
6848 | 0 | "pand %%xmm0,%%xmm2 \n" |
6849 | 0 | "pand %%xmm0,%%xmm3 \n" |
6850 | 0 | "pand %%xmm1,%%xmm4 \n" |
6851 | 0 | "pand %%xmm1,%%xmm5 \n" |
6852 | 0 | "por %%xmm4,%%xmm2 \n" |
6853 | 0 | "por %%xmm5,%%xmm3 \n" |
6854 | 0 | "movdqu %%xmm2,(%1) \n" |
6855 | 0 | "movdqu %%xmm3,0x10(%1) \n" |
6856 | 0 | "lea 0x20(%1),%1 \n" |
6857 | 0 | "sub $0x8,%2 \n" |
6858 | 0 | "jg 1b \n" |
6859 | 0 | : "+r"(src), // %0 |
6860 | 0 | "+r"(dst), // %1 |
6861 | 0 | "+r"(width) // %2 |
6862 | 0 | : |
6863 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
6864 | 0 | } |
6865 | | #endif // HAS_ARGBCOPYYTOALPHAROW_SSE2 |
6866 | | |
6867 | | #ifdef HAS_ARGBCOPYYTOALPHAROW_AVX2 |
6868 | | // width in pixels |
6869 | 0 | void ARGBCopyYToAlphaRow_AVX2(const uint8_t* src, uint8_t* dst, int width) { |
6870 | 0 | asm volatile( |
6871 | 0 | "vpcmpeqb %%ymm0,%%ymm0,%%ymm0 \n" |
6872 | 0 | "vpsrld $0x8,%%ymm0,%%ymm0 \n" |
6873 | |
|
6874 | 0 | LABELALIGN |
6875 | 0 | "1: \n" |
6876 | 0 | "vpmovzxbd (%0),%%ymm1 \n" |
6877 | 0 | "vpmovzxbd 0x8(%0),%%ymm2 \n" |
6878 | 0 | "lea 0x10(%0),%0 \n" |
6879 | 0 | "vpslld $0x18,%%ymm1,%%ymm1 \n" |
6880 | 0 | "vpslld $0x18,%%ymm2,%%ymm2 \n" |
6881 | 0 | "vpblendvb %%ymm0,(%1),%%ymm1,%%ymm1 \n" |
6882 | 0 | "vpblendvb %%ymm0,0x20(%1),%%ymm2,%%ymm2 \n" |
6883 | 0 | "vmovdqu %%ymm1,(%1) \n" |
6884 | 0 | "vmovdqu %%ymm2,0x20(%1) \n" |
6885 | 0 | "lea 0x40(%1),%1 \n" |
6886 | 0 | "sub $0x10,%2 \n" |
6887 | 0 | "jg 1b \n" |
6888 | 0 | "vzeroupper \n" |
6889 | 0 | : "+r"(src), // %0 |
6890 | 0 | "+r"(dst), // %1 |
6891 | 0 | "+r"(width) // %2 |
6892 | 0 | : |
6893 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
6894 | 0 | } |
6895 | | #endif // HAS_ARGBCOPYYTOALPHAROW_AVX2 |
6896 | | |
6897 | | #ifdef HAS_SETROW_X86 |
6898 | 0 | void SetRow_X86(uint8_t* dst, uint8_t v8, int width) { |
6899 | 0 | size_t width_tmp = (size_t)(width >> 2); |
6900 | 0 | const uint32_t v32 = v8 * 0x01010101u; // Duplicate byte to all bytes. |
6901 | 0 | asm volatile("rep stosl \n" |
6902 | 0 | : "+D"(dst), // %0 |
6903 | 0 | "+c"(width_tmp) // %1 |
6904 | 0 | : "a"(v32) // %2 |
6905 | 0 | : "memory", "cc"); |
6906 | 0 | } |
6907 | | |
6908 | 0 | void SetRow_ERMS(uint8_t* dst, uint8_t v8, int width) { |
6909 | 0 | size_t width_tmp = (size_t)(width); |
6910 | 0 | asm volatile("rep stosb \n" |
6911 | 0 | : "+D"(dst), // %0 |
6912 | 0 | "+c"(width_tmp) // %1 |
6913 | 0 | : "a"(v8) // %2 |
6914 | 0 | : "memory", "cc"); |
6915 | 0 | } |
6916 | | |
6917 | 0 | void ARGBSetRow_X86(uint8_t* dst_argb, uint32_t v32, int width) { |
6918 | 0 | size_t width_tmp = (size_t)(width); |
6919 | 0 | asm volatile("rep stosl \n" |
6920 | 0 | : "+D"(dst_argb), // %0 |
6921 | 0 | "+c"(width_tmp) // %1 |
6922 | 0 | : "a"(v32) // %2 |
6923 | 0 | : "memory", "cc"); |
6924 | 0 | } |
6925 | | #endif // HAS_SETROW_X86 |
6926 | | |
6927 | | #ifdef HAS_YUY2TOYROW_SSE2 |
6928 | 0 | void YUY2ToYRow_SSE2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) { |
6929 | 0 | asm volatile( |
6930 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
6931 | 0 | "psrlw $0x8,%%xmm5 \n" |
6932 | |
|
6933 | 0 | LABELALIGN |
6934 | 0 | "1: \n" |
6935 | 0 | "movdqu (%0),%%xmm0 \n" |
6936 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
6937 | 0 | "lea 0x20(%0),%0 \n" |
6938 | 0 | "pand %%xmm5,%%xmm0 \n" |
6939 | 0 | "pand %%xmm5,%%xmm1 \n" |
6940 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
6941 | 0 | "movdqu %%xmm0,(%1) \n" |
6942 | 0 | "lea 0x10(%1),%1 \n" |
6943 | 0 | "sub $0x10,%2 \n" |
6944 | 0 | "jg 1b \n" |
6945 | 0 | : "+r"(src_yuy2), // %0 |
6946 | 0 | "+r"(dst_y), // %1 |
6947 | 0 | "+r"(width) // %2 |
6948 | 0 | : |
6949 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
6950 | 0 | } |
6951 | | |
6952 | | void YUY2ToNVUVRow_SSE2(const uint8_t* src_yuy2, |
6953 | | int stride_yuy2, |
6954 | | uint8_t* dst_uv, |
6955 | 0 | int width) { |
6956 | 0 | asm volatile( |
6957 | 0 | "1: \n" |
6958 | 0 | "movdqu (%0),%%xmm0 \n" |
6959 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
6960 | 0 | "movdqu 0x00(%0,%3,1),%%xmm2 \n" |
6961 | 0 | "movdqu 0x10(%0,%3,1),%%xmm3 \n" |
6962 | 0 | "lea 0x20(%0),%0 \n" |
6963 | 0 | "pavgb %%xmm2,%%xmm0 \n" |
6964 | 0 | "pavgb %%xmm3,%%xmm1 \n" |
6965 | 0 | "psrlw $0x8,%%xmm0 \n" |
6966 | 0 | "psrlw $0x8,%%xmm1 \n" |
6967 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
6968 | 0 | "movdqu %%xmm0,(%1) \n" |
6969 | 0 | "lea 0x10(%1),%1 \n" |
6970 | 0 | "sub $0x10,%2 \n" |
6971 | 0 | "jg 1b \n" |
6972 | 0 | : "+r"(src_yuy2), // %0 |
6973 | 0 | "+r"(dst_uv), // %1 |
6974 | 0 | "+r"(width) // %2 |
6975 | 0 | : "r"((ptrdiff_t)(stride_yuy2)) // %3 |
6976 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3"); |
6977 | 0 | } |
6978 | | |
6979 | | void YUY2ToUVRow_SSE2(const uint8_t* src_yuy2, |
6980 | | int stride_yuy2, |
6981 | | uint8_t* dst_u, |
6982 | | uint8_t* dst_v, |
6983 | 0 | int width) { |
6984 | 0 | asm volatile( |
6985 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
6986 | 0 | "psrlw $0x8,%%xmm5 \n" |
6987 | 0 | "sub %1,%2 \n" |
6988 | |
|
6989 | 0 | LABELALIGN |
6990 | 0 | "1: \n" |
6991 | 0 | "movdqu (%0),%%xmm0 \n" |
6992 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
6993 | 0 | "movdqu 0x00(%0,%4,1),%%xmm2 \n" |
6994 | 0 | "movdqu 0x10(%0,%4,1),%%xmm3 \n" |
6995 | 0 | "lea 0x20(%0),%0 \n" |
6996 | 0 | "pavgb %%xmm2,%%xmm0 \n" |
6997 | 0 | "pavgb %%xmm3,%%xmm1 \n" |
6998 | 0 | "psrlw $0x8,%%xmm0 \n" |
6999 | 0 | "psrlw $0x8,%%xmm1 \n" |
7000 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7001 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
7002 | 0 | "pand %%xmm5,%%xmm0 \n" |
7003 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7004 | 0 | "psrlw $0x8,%%xmm1 \n" |
7005 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
7006 | 0 | "movq %%xmm0,(%1) \n" |
7007 | 0 | "movq %%xmm1,0x00(%1,%2,1) \n" |
7008 | 0 | "lea 0x8(%1),%1 \n" |
7009 | 0 | "sub $0x10,%3 \n" |
7010 | 0 | "jg 1b \n" |
7011 | 0 | : "+r"(src_yuy2), // %0 |
7012 | 0 | "+r"(dst_u), // %1 |
7013 | 0 | "+r"(dst_v), // %2 |
7014 | 0 | "+r"(width) // %3 |
7015 | 0 | : "r"((ptrdiff_t)(stride_yuy2)) // %4 |
7016 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
7017 | 0 | } |
7018 | | |
7019 | | void YUY2ToUV422Row_SSE2(const uint8_t* src_yuy2, |
7020 | | uint8_t* dst_u, |
7021 | | uint8_t* dst_v, |
7022 | 0 | int width) { |
7023 | 0 | asm volatile( |
7024 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7025 | 0 | "psrlw $0x8,%%xmm5 \n" |
7026 | 0 | "sub %1,%2 \n" |
7027 | |
|
7028 | 0 | LABELALIGN |
7029 | 0 | "1: \n" |
7030 | 0 | "movdqu (%0),%%xmm0 \n" |
7031 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7032 | 0 | "lea 0x20(%0),%0 \n" |
7033 | 0 | "psrlw $0x8,%%xmm0 \n" |
7034 | 0 | "psrlw $0x8,%%xmm1 \n" |
7035 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7036 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
7037 | 0 | "pand %%xmm5,%%xmm0 \n" |
7038 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7039 | 0 | "psrlw $0x8,%%xmm1 \n" |
7040 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
7041 | 0 | "movq %%xmm0,(%1) \n" |
7042 | 0 | "movq %%xmm1,0x00(%1,%2,1) \n" |
7043 | 0 | "lea 0x8(%1),%1 \n" |
7044 | 0 | "sub $0x10,%3 \n" |
7045 | 0 | "jg 1b \n" |
7046 | 0 | : "+r"(src_yuy2), // %0 |
7047 | 0 | "+r"(dst_u), // %1 |
7048 | 0 | "+r"(dst_v), // %2 |
7049 | 0 | "+r"(width) // %3 |
7050 | 0 | : |
7051 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7052 | 0 | } |
7053 | | |
7054 | 0 | void UYVYToYRow_SSE2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) { |
7055 | 0 | asm volatile( |
7056 | 0 | "1: \n" |
7057 | 0 | "movdqu (%0),%%xmm0 \n" |
7058 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7059 | 0 | "lea 0x20(%0),%0 \n" |
7060 | 0 | "psrlw $0x8,%%xmm0 \n" |
7061 | 0 | "psrlw $0x8,%%xmm1 \n" |
7062 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7063 | 0 | "movdqu %%xmm0,(%1) \n" |
7064 | 0 | "lea 0x10(%1),%1 \n" |
7065 | 0 | "sub $0x10,%2 \n" |
7066 | 0 | "jg 1b \n" |
7067 | 0 | : "+r"(src_uyvy), // %0 |
7068 | 0 | "+r"(dst_y), // %1 |
7069 | 0 | "+r"(width) // %2 |
7070 | 0 | : |
7071 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
7072 | 0 | } |
7073 | | |
7074 | | void UYVYToUVRow_SSE2(const uint8_t* src_uyvy, |
7075 | | int stride_uyvy, |
7076 | | uint8_t* dst_u, |
7077 | | uint8_t* dst_v, |
7078 | 0 | int width) { |
7079 | 0 | asm volatile( |
7080 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7081 | 0 | "psrlw $0x8,%%xmm5 \n" |
7082 | 0 | "sub %1,%2 \n" |
7083 | |
|
7084 | 0 | LABELALIGN |
7085 | 0 | "1: \n" |
7086 | 0 | "movdqu (%0),%%xmm0 \n" |
7087 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7088 | 0 | "movdqu 0x00(%0,%4,1),%%xmm2 \n" |
7089 | 0 | "movdqu 0x10(%0,%4,1),%%xmm3 \n" |
7090 | 0 | "lea 0x20(%0),%0 \n" |
7091 | 0 | "pavgb %%xmm2,%%xmm0 \n" |
7092 | 0 | "pavgb %%xmm3,%%xmm1 \n" |
7093 | 0 | "pand %%xmm5,%%xmm0 \n" |
7094 | 0 | "pand %%xmm5,%%xmm1 \n" |
7095 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7096 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
7097 | 0 | "pand %%xmm5,%%xmm0 \n" |
7098 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7099 | 0 | "psrlw $0x8,%%xmm1 \n" |
7100 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
7101 | 0 | "movq %%xmm0,(%1) \n" |
7102 | 0 | "movq %%xmm1,0x00(%1,%2,1) \n" |
7103 | 0 | "lea 0x8(%1),%1 \n" |
7104 | 0 | "sub $0x10,%3 \n" |
7105 | 0 | "jg 1b \n" |
7106 | 0 | : "+r"(src_uyvy), // %0 |
7107 | 0 | "+r"(dst_u), // %1 |
7108 | 0 | "+r"(dst_v), // %2 |
7109 | 0 | "+r"(width) // %3 |
7110 | 0 | : "r"((ptrdiff_t)(stride_uyvy)) // %4 |
7111 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
7112 | 0 | } |
7113 | | |
7114 | | void UYVYToUV422Row_SSE2(const uint8_t* src_uyvy, |
7115 | | uint8_t* dst_u, |
7116 | | uint8_t* dst_v, |
7117 | 0 | int width) { |
7118 | 0 | asm volatile( |
7119 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7120 | 0 | "psrlw $0x8,%%xmm5 \n" |
7121 | 0 | "sub %1,%2 \n" |
7122 | |
|
7123 | 0 | LABELALIGN |
7124 | 0 | "1: \n" |
7125 | 0 | "movdqu (%0),%%xmm0 \n" |
7126 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7127 | 0 | "lea 0x20(%0),%0 \n" |
7128 | 0 | "pand %%xmm5,%%xmm0 \n" |
7129 | 0 | "pand %%xmm5,%%xmm1 \n" |
7130 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7131 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
7132 | 0 | "pand %%xmm5,%%xmm0 \n" |
7133 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7134 | 0 | "psrlw $0x8,%%xmm1 \n" |
7135 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
7136 | 0 | "movq %%xmm0,(%1) \n" |
7137 | 0 | "movq %%xmm1,0x00(%1,%2,1) \n" |
7138 | 0 | "lea 0x8(%1),%1 \n" |
7139 | 0 | "sub $0x10,%3 \n" |
7140 | 0 | "jg 1b \n" |
7141 | 0 | : "+r"(src_uyvy), // %0 |
7142 | 0 | "+r"(dst_u), // %1 |
7143 | 0 | "+r"(dst_v), // %2 |
7144 | 0 | "+r"(width) // %3 |
7145 | 0 | : |
7146 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7147 | 0 | } |
7148 | | #endif // HAS_YUY2TOYROW_SSE2 |
7149 | | |
7150 | | #ifdef HAS_YUY2TOYROW_AVX2 |
7151 | 0 | void YUY2ToYRow_AVX2(const uint8_t* src_yuy2, uint8_t* dst_y, int width) { |
7152 | 0 | asm volatile( |
7153 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7154 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
7155 | |
|
7156 | 0 | LABELALIGN |
7157 | 0 | "1: \n" |
7158 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7159 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7160 | 0 | "lea 0x40(%0),%0 \n" |
7161 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" |
7162 | 0 | "vpand %%ymm5,%%ymm1,%%ymm1 \n" |
7163 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7164 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7165 | 0 | "vmovdqu %%ymm0,(%1) \n" |
7166 | 0 | "lea 0x20(%1),%1 \n" |
7167 | 0 | "sub $0x20,%2 \n" |
7168 | 0 | "jg 1b \n" |
7169 | 0 | "vzeroupper \n" |
7170 | 0 | : "+r"(src_yuy2), // %0 |
7171 | 0 | "+r"(dst_y), // %1 |
7172 | 0 | "+r"(width) // %2 |
7173 | 0 | : |
7174 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7175 | 0 | } |
7176 | | |
7177 | | void YUY2ToNVUVRow_AVX2(const uint8_t* src_yuy2, |
7178 | | int stride_yuy2, |
7179 | | uint8_t* dst_uv, |
7180 | 0 | int width) { |
7181 | 0 | asm volatile( |
7182 | 0 | "1: \n" |
7183 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7184 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7185 | 0 | "vpavgb 0x00(%0,%3,1),%%ymm0,%%ymm0 \n" |
7186 | 0 | "vpavgb 0x20(%0,%3,1),%%ymm1,%%ymm1 \n" |
7187 | 0 | "lea 0x40(%0),%0 \n" |
7188 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7189 | 0 | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
7190 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7191 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7192 | 0 | "vmovdqu %%ymm0,(%1) \n" |
7193 | 0 | "lea 0x20(%1),%1 \n" |
7194 | 0 | "sub $0x20,%2 \n" |
7195 | 0 | "jg 1b \n" |
7196 | 0 | "vzeroupper \n" |
7197 | 0 | : "+r"(src_yuy2), // %0 |
7198 | 0 | "+r"(dst_uv), // %1 |
7199 | 0 | "+r"(width) // %2 |
7200 | 0 | : "r"((ptrdiff_t)(stride_yuy2)) // %3 |
7201 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
7202 | 0 | } |
7203 | | |
7204 | | void YUY2ToUVRow_AVX2(const uint8_t* src_yuy2, |
7205 | | int stride_yuy2, |
7206 | | uint8_t* dst_u, |
7207 | | uint8_t* dst_v, |
7208 | 0 | int width) { |
7209 | 0 | asm volatile( |
7210 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7211 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
7212 | 0 | "sub %1,%2 \n" |
7213 | |
|
7214 | 0 | LABELALIGN |
7215 | 0 | "1: \n" |
7216 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7217 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7218 | 0 | "vpavgb 0x00(%0,%4,1),%%ymm0,%%ymm0 \n" |
7219 | 0 | "vpavgb 0x20(%0,%4,1),%%ymm1,%%ymm1 \n" |
7220 | 0 | "lea 0x40(%0),%0 \n" |
7221 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7222 | 0 | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
7223 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7224 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7225 | 0 | "vpand %%ymm5,%%ymm0,%%ymm1 \n" |
7226 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7227 | 0 | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" |
7228 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
7229 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
7230 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7231 | 0 | "vextractf128 $0x0,%%ymm1,(%1) \n" |
7232 | 0 | "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1) \n" |
7233 | 0 | "lea 0x10(%1),%1 \n" |
7234 | 0 | "sub $0x20,%3 \n" |
7235 | 0 | "jg 1b \n" |
7236 | 0 | "vzeroupper \n" |
7237 | 0 | : "+r"(src_yuy2), // %0 |
7238 | 0 | "+r"(dst_u), // %1 |
7239 | 0 | "+r"(dst_v), // %2 |
7240 | 0 | "+r"(width) // %3 |
7241 | 0 | : "r"((ptrdiff_t)(stride_yuy2)) // %4 |
7242 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7243 | 0 | } |
7244 | | |
7245 | | void YUY2ToUV422Row_AVX2(const uint8_t* src_yuy2, |
7246 | | uint8_t* dst_u, |
7247 | | uint8_t* dst_v, |
7248 | 0 | int width) { |
7249 | 0 | asm volatile( |
7250 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7251 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
7252 | 0 | "sub %1,%2 \n" |
7253 | |
|
7254 | 0 | LABELALIGN |
7255 | 0 | "1: \n" |
7256 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7257 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7258 | 0 | "lea 0x40(%0),%0 \n" |
7259 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7260 | 0 | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
7261 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7262 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7263 | 0 | "vpand %%ymm5,%%ymm0,%%ymm1 \n" |
7264 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7265 | 0 | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" |
7266 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
7267 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
7268 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7269 | 0 | "vextractf128 $0x0,%%ymm1,(%1) \n" |
7270 | 0 | "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1) \n" |
7271 | 0 | "lea 0x10(%1),%1 \n" |
7272 | 0 | "sub $0x20,%3 \n" |
7273 | 0 | "jg 1b \n" |
7274 | 0 | "vzeroupper \n" |
7275 | 0 | : "+r"(src_yuy2), // %0 |
7276 | 0 | "+r"(dst_u), // %1 |
7277 | 0 | "+r"(dst_v), // %2 |
7278 | 0 | "+r"(width) // %3 |
7279 | 0 | : |
7280 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7281 | 0 | } |
7282 | | |
7283 | 0 | void UYVYToYRow_AVX2(const uint8_t* src_uyvy, uint8_t* dst_y, int width) { |
7284 | 0 | asm volatile( |
7285 | 0 | "1: \n" |
7286 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7287 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7288 | 0 | "lea 0x40(%0),%0 \n" |
7289 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7290 | 0 | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
7291 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7292 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7293 | 0 | "vmovdqu %%ymm0,(%1) \n" |
7294 | 0 | "lea 0x20(%1),%1 \n" |
7295 | 0 | "sub $0x20,%2 \n" |
7296 | 0 | "jg 1b \n" |
7297 | 0 | "vzeroupper \n" |
7298 | 0 | : "+r"(src_uyvy), // %0 |
7299 | 0 | "+r"(dst_y), // %1 |
7300 | 0 | "+r"(width) // %2 |
7301 | 0 | : |
7302 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7303 | 0 | } |
7304 | | void UYVYToUVRow_AVX2(const uint8_t* src_uyvy, |
7305 | | int stride_uyvy, |
7306 | | uint8_t* dst_u, |
7307 | | uint8_t* dst_v, |
7308 | 0 | int width) { |
7309 | 0 | asm volatile( |
7310 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7311 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
7312 | 0 | "sub %1,%2 \n" |
7313 | |
|
7314 | 0 | LABELALIGN |
7315 | 0 | "1: \n" |
7316 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7317 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7318 | 0 | "vpavgb 0x00(%0,%4,1),%%ymm0,%%ymm0 \n" |
7319 | 0 | "vpavgb 0x20(%0,%4,1),%%ymm1,%%ymm1 \n" |
7320 | 0 | "lea 0x40(%0),%0 \n" |
7321 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" |
7322 | 0 | "vpand %%ymm5,%%ymm1,%%ymm1 \n" |
7323 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7324 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7325 | 0 | "vpand %%ymm5,%%ymm0,%%ymm1 \n" |
7326 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7327 | 0 | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" |
7328 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
7329 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
7330 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7331 | 0 | "vextractf128 $0x0,%%ymm1,(%1) \n" |
7332 | 0 | "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1) \n" |
7333 | 0 | "lea 0x10(%1),%1 \n" |
7334 | 0 | "sub $0x20,%3 \n" |
7335 | 0 | "jg 1b \n" |
7336 | 0 | "vzeroupper \n" |
7337 | 0 | : "+r"(src_uyvy), // %0 |
7338 | 0 | "+r"(dst_u), // %1 |
7339 | 0 | "+r"(dst_v), // %2 |
7340 | 0 | "+r"(width) // %3 |
7341 | 0 | : "r"((ptrdiff_t)(stride_uyvy)) // %4 |
7342 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7343 | 0 | } |
7344 | | |
7345 | | void UYVYToUV422Row_AVX2(const uint8_t* src_uyvy, |
7346 | | uint8_t* dst_u, |
7347 | | uint8_t* dst_v, |
7348 | 0 | int width) { |
7349 | 0 | asm volatile( |
7350 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7351 | 0 | "vpsrlw $0x8,%%ymm5,%%ymm5 \n" |
7352 | 0 | "sub %1,%2 \n" |
7353 | |
|
7354 | 0 | LABELALIGN |
7355 | 0 | "1: \n" |
7356 | 0 | "vmovdqu (%0),%%ymm0 \n" |
7357 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
7358 | 0 | "lea 0x40(%0),%0 \n" |
7359 | 0 | "vpand %%ymm5,%%ymm0,%%ymm0 \n" |
7360 | 0 | "vpand %%ymm5,%%ymm1,%%ymm1 \n" |
7361 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7362 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7363 | 0 | "vpand %%ymm5,%%ymm0,%%ymm1 \n" |
7364 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7365 | 0 | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" |
7366 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
7367 | 0 | "vpermq $0xd8,%%ymm1,%%ymm1 \n" |
7368 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
7369 | 0 | "vextractf128 $0x0,%%ymm1,(%1) \n" |
7370 | 0 | "vextractf128 $0x0,%%ymm0,0x00(%1,%2,1) \n" |
7371 | 0 | "lea 0x10(%1),%1 \n" |
7372 | 0 | "sub $0x20,%3 \n" |
7373 | 0 | "jg 1b \n" |
7374 | 0 | "vzeroupper \n" |
7375 | 0 | : "+r"(src_uyvy), // %0 |
7376 | 0 | "+r"(dst_u), // %1 |
7377 | 0 | "+r"(dst_v), // %2 |
7378 | 0 | "+r"(width) // %3 |
7379 | 0 | : |
7380 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
7381 | 0 | } |
7382 | | #endif // HAS_YUY2TOYROW_AVX2 |
7383 | | |
7384 | | #ifdef HAS_ARGBBLENDROW_SSSE3 |
7385 | | // Shuffle table for isolating alpha. |
7386 | | static const uvec8 kShuffleAlpha = {3u, 0x80, 3u, 0x80, 7u, 0x80, 7u, 0x80, |
7387 | | 11u, 0x80, 11u, 0x80, 15u, 0x80, 15u, 0x80}; |
7388 | | |
7389 | | // Blend 8 pixels at a time |
7390 | | void ARGBBlendRow_SSSE3(const uint8_t* src_argb, |
7391 | | const uint8_t* src_argb1, |
7392 | | uint8_t* dst_argb, |
7393 | 0 | int width) { |
7394 | 0 | asm volatile( |
7395 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
7396 | 0 | "psrlw $0xf,%%xmm7 \n" |
7397 | 0 | "pcmpeqb %%xmm6,%%xmm6 \n" |
7398 | 0 | "psrlw $0x8,%%xmm6 \n" |
7399 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7400 | 0 | "psllw $0x8,%%xmm5 \n" |
7401 | 0 | "pcmpeqb %%xmm4,%%xmm4 \n" |
7402 | 0 | "pslld $0x18,%%xmm4 \n" |
7403 | 0 | "sub $0x4,%3 \n" |
7404 | 0 | "jl 49f \n" |
7405 | | |
7406 | | // 4 pixel loop. |
7407 | 0 | LABELALIGN |
7408 | 0 | "40: \n" |
7409 | 0 | "movdqu (%0),%%xmm3 \n" |
7410 | 0 | "lea 0x10(%0),%0 \n" |
7411 | 0 | "movdqa %%xmm3,%%xmm0 \n" |
7412 | 0 | "pxor %%xmm4,%%xmm3 \n" |
7413 | 0 | "movdqu (%1),%%xmm2 \n" |
7414 | 0 | "pshufb %4,%%xmm3 \n" |
7415 | 0 | "pand %%xmm6,%%xmm2 \n" |
7416 | 0 | "paddw %%xmm7,%%xmm3 \n" |
7417 | 0 | "pmullw %%xmm3,%%xmm2 \n" |
7418 | 0 | "movdqu (%1),%%xmm1 \n" |
7419 | 0 | "lea 0x10(%1),%1 \n" |
7420 | 0 | "psrlw $0x8,%%xmm1 \n" |
7421 | 0 | "por %%xmm4,%%xmm0 \n" |
7422 | 0 | "pmullw %%xmm3,%%xmm1 \n" |
7423 | 0 | "psrlw $0x8,%%xmm2 \n" |
7424 | 0 | "paddusb %%xmm2,%%xmm0 \n" |
7425 | 0 | "pand %%xmm5,%%xmm1 \n" |
7426 | 0 | "paddusb %%xmm1,%%xmm0 \n" |
7427 | 0 | "movdqu %%xmm0,(%2) \n" |
7428 | 0 | "lea 0x10(%2),%2 \n" |
7429 | 0 | "sub $0x4,%3 \n" |
7430 | 0 | "jge 40b \n" |
7431 | |
|
7432 | 0 | "49: \n" |
7433 | 0 | "add $0x3,%3 \n" |
7434 | 0 | "jl 99f \n" |
7435 | | |
7436 | | // 1 pixel loop. |
7437 | 0 | "91: \n" |
7438 | 0 | "movd (%0),%%xmm3 \n" |
7439 | 0 | "lea 0x4(%0),%0 \n" |
7440 | 0 | "movdqa %%xmm3,%%xmm0 \n" |
7441 | 0 | "pxor %%xmm4,%%xmm3 \n" |
7442 | 0 | "movd (%1),%%xmm2 \n" |
7443 | 0 | "pshufb %4,%%xmm3 \n" |
7444 | 0 | "pand %%xmm6,%%xmm2 \n" |
7445 | 0 | "paddw %%xmm7,%%xmm3 \n" |
7446 | 0 | "pmullw %%xmm3,%%xmm2 \n" |
7447 | 0 | "movd (%1),%%xmm1 \n" |
7448 | 0 | "lea 0x4(%1),%1 \n" |
7449 | 0 | "psrlw $0x8,%%xmm1 \n" |
7450 | 0 | "por %%xmm4,%%xmm0 \n" |
7451 | 0 | "pmullw %%xmm3,%%xmm1 \n" |
7452 | 0 | "psrlw $0x8,%%xmm2 \n" |
7453 | 0 | "paddusb %%xmm2,%%xmm0 \n" |
7454 | 0 | "pand %%xmm5,%%xmm1 \n" |
7455 | 0 | "paddusb %%xmm1,%%xmm0 \n" |
7456 | 0 | "movd %%xmm0,(%2) \n" |
7457 | 0 | "lea 0x4(%2),%2 \n" |
7458 | 0 | "sub $0x1,%3 \n" |
7459 | 0 | "jge 91b \n" |
7460 | 0 | "99: \n" |
7461 | 0 | : "+r"(src_argb), // %0 |
7462 | 0 | "+r"(src_argb1), // %1 |
7463 | 0 | "+r"(dst_argb), // %2 |
7464 | 0 | "+r"(width) // %3 |
7465 | 0 | : "m"(kShuffleAlpha) // %4 |
7466 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7467 | 0 | "xmm7"); |
7468 | 0 | } |
7469 | | #endif // HAS_ARGBBLENDROW_SSSE3 |
7470 | | |
7471 | | #ifdef HAS_BLENDPLANEROW_SSSE3 |
7472 | | // Blend 8 pixels at a time. |
7473 | | // unsigned version of math |
7474 | | // =((A2*C2)+(B2*(255-C2))+255)/256 |
7475 | | // signed version of math |
7476 | | // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256 |
7477 | | void BlendPlaneRow_SSSE3(const uint8_t* src0, |
7478 | | const uint8_t* src1, |
7479 | | const uint8_t* alpha, |
7480 | | uint8_t* dst, |
7481 | 0 | int width) { |
7482 | 0 | asm volatile( |
7483 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7484 | 0 | "psllw $0x8,%%xmm5 \n" |
7485 | 0 | "mov $0x80808080,%%eax \n" |
7486 | 0 | "movd %%eax,%%xmm6 \n" |
7487 | 0 | "pshufd $0x0,%%xmm6,%%xmm6 \n" |
7488 | 0 | "mov $0x807f807f,%%eax \n" |
7489 | 0 | "movd %%eax,%%xmm7 \n" |
7490 | 0 | "pshufd $0x0,%%xmm7,%%xmm7 \n" |
7491 | 0 | "sub %2,%0 \n" |
7492 | 0 | "sub %2,%1 \n" |
7493 | 0 | "sub %2,%3 \n" |
7494 | | |
7495 | | // 8 pixel loop. |
7496 | 0 | LABELALIGN |
7497 | 0 | "1: \n" |
7498 | 0 | "movq (%2),%%xmm0 \n" |
7499 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
7500 | 0 | "pxor %%xmm5,%%xmm0 \n" |
7501 | 0 | "movq (%0,%2,1),%%xmm1 \n" |
7502 | 0 | "movq (%1,%2,1),%%xmm2 \n" |
7503 | 0 | "punpcklbw %%xmm2,%%xmm1 \n" |
7504 | 0 | "psubb %%xmm6,%%xmm1 \n" |
7505 | 0 | "pmaddubsw %%xmm1,%%xmm0 \n" |
7506 | 0 | "paddw %%xmm7,%%xmm0 \n" |
7507 | 0 | "psrlw $0x8,%%xmm0 \n" |
7508 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7509 | 0 | "movq %%xmm0,(%3,%2,1) \n" |
7510 | 0 | "lea 0x8(%2),%2 \n" |
7511 | 0 | "sub $0x8,%4 \n" |
7512 | 0 | "jg 1b \n" |
7513 | 0 | : "+r"(src0), // %0 |
7514 | 0 | "+r"(src1), // %1 |
7515 | 0 | "+r"(alpha), // %2 |
7516 | 0 | "+r"(dst), // %3 |
7517 | 0 | "+rm"(width) // %4 |
7518 | 0 | ::"memory", |
7519 | 0 | "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm5", "xmm6", "xmm7"); |
7520 | 0 | } |
7521 | | #endif // HAS_BLENDPLANEROW_SSSE3 |
7522 | | |
7523 | | #ifdef HAS_BLENDPLANEROW_AVX2 |
7524 | | // Blend 32 pixels at a time. |
7525 | | // unsigned version of math |
7526 | | // =((A2*C2)+(B2*(255-C2))+255)/256 |
7527 | | // signed version of math |
7528 | | // =(((A2-128)*C2)+((B2-128)*(255-C2))+32768+127)/256 |
7529 | | void BlendPlaneRow_AVX2(const uint8_t* src0, |
7530 | | const uint8_t* src1, |
7531 | | const uint8_t* alpha, |
7532 | | uint8_t* dst, |
7533 | 0 | int width) { |
7534 | 0 | asm volatile( |
7535 | 0 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7536 | 0 | "vpsllw $0x8,%%ymm5,%%ymm5 \n" |
7537 | 0 | "mov $0x80808080,%%eax \n" |
7538 | 0 | "vmovd %%eax,%%xmm6 \n" |
7539 | 0 | "vbroadcastss %%xmm6,%%ymm6 \n" |
7540 | 0 | "mov $0x807f807f,%%eax \n" |
7541 | 0 | "vmovd %%eax,%%xmm7 \n" |
7542 | 0 | "vbroadcastss %%xmm7,%%ymm7 \n" |
7543 | 0 | "sub %2,%0 \n" |
7544 | 0 | "sub %2,%1 \n" |
7545 | 0 | "sub %2,%3 \n" |
7546 | | |
7547 | | // 32 pixel loop. |
7548 | 0 | LABELALIGN |
7549 | 0 | "1: \n" |
7550 | 0 | "vmovdqu (%2),%%ymm0 \n" |
7551 | 0 | "vpunpckhbw %%ymm0,%%ymm0,%%ymm3 \n" |
7552 | 0 | "vpunpcklbw %%ymm0,%%ymm0,%%ymm0 \n" |
7553 | 0 | "vpxor %%ymm5,%%ymm3,%%ymm3 \n" |
7554 | 0 | "vpxor %%ymm5,%%ymm0,%%ymm0 \n" |
7555 | 0 | "vmovdqu (%0,%2,1),%%ymm1 \n" |
7556 | 0 | "vmovdqu (%1,%2,1),%%ymm2 \n" |
7557 | 0 | "vpunpckhbw %%ymm2,%%ymm1,%%ymm4 \n" |
7558 | 0 | "vpunpcklbw %%ymm2,%%ymm1,%%ymm1 \n" |
7559 | 0 | "vpsubb %%ymm6,%%ymm4,%%ymm4 \n" |
7560 | 0 | "vpsubb %%ymm6,%%ymm1,%%ymm1 \n" |
7561 | 0 | "vpmaddubsw %%ymm4,%%ymm3,%%ymm3 \n" |
7562 | 0 | "vpmaddubsw %%ymm1,%%ymm0,%%ymm0 \n" |
7563 | 0 | "vpaddw %%ymm7,%%ymm3,%%ymm3 \n" |
7564 | 0 | "vpaddw %%ymm7,%%ymm0,%%ymm0 \n" |
7565 | 0 | "vpsrlw $0x8,%%ymm3,%%ymm3 \n" |
7566 | 0 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7567 | 0 | "vpackuswb %%ymm3,%%ymm0,%%ymm0 \n" |
7568 | 0 | "vmovdqu %%ymm0,(%3,%2,1) \n" |
7569 | 0 | "lea 0x20(%2),%2 \n" |
7570 | 0 | "sub $0x20,%4 \n" |
7571 | 0 | "jg 1b \n" |
7572 | 0 | "vzeroupper \n" |
7573 | 0 | : "+r"(src0), // %0 |
7574 | 0 | "+r"(src1), // %1 |
7575 | 0 | "+r"(alpha), // %2 |
7576 | 0 | "+r"(dst), // %3 |
7577 | 0 | "+rm"(width) // %4 |
7578 | 0 | ::"memory", |
7579 | 0 | "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7580 | 0 | "xmm7"); |
7581 | 0 | } |
7582 | | #endif // HAS_BLENDPLANEROW_AVX2 |
7583 | | |
7584 | | #ifdef HAS_ARGBATTENUATEROW_SSSE3 |
7585 | | // Shuffle table duplicating alpha. |
7586 | | static const vec8 kAttenuateShuffle = {6, -128, 6, -128, 6, -128, |
7587 | | -128, -128, 14, -128, 14, -128, |
7588 | | 14, -128, -128, -128}; |
7589 | | |
7590 | | // Attenuate 4 pixels at a time. |
7591 | | void ARGBAttenuateRow_SSSE3(const uint8_t* src_argb, |
7592 | | uint8_t* dst_argb, |
7593 | 0 | int width) { |
7594 | 0 | asm volatile( |
7595 | 0 | "movdqa %3,%%xmm4 \n" |
7596 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
7597 | 0 | "pslld $0x18,%%xmm5 \n" |
7598 | 0 | "pxor %%xmm6,%%xmm6 \n" |
7599 | 0 | "pcmpeqb %%xmm7,%%xmm7 \n" |
7600 | 0 | "punpcklbw %%xmm6,%%xmm7 \n" |
7601 | 0 | "sub %0,%1 \n" |
7602 | | |
7603 | | // 4 pixel loop. |
7604 | 0 | LABELALIGN |
7605 | 0 | "1: \n" |
7606 | 0 | "movdqu (%0),%%xmm6 \n" |
7607 | 0 | "movdqa %%xmm6,%%xmm0 \n" |
7608 | 0 | "movdqa %%xmm6,%%xmm1 \n" |
7609 | 0 | "punpcklbw %%xmm5,%%xmm0 \n" |
7610 | 0 | "punpckhbw %%xmm5,%%xmm1 \n" |
7611 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
7612 | 0 | "movdqa %%xmm1,%%xmm3 \n" |
7613 | 0 | "pshufb %%xmm4,%%xmm2 \n" // a,a,a,0 |
7614 | 0 | "pshufb %%xmm4,%%xmm3 \n" |
7615 | 0 | "pmullw %%xmm2,%%xmm0 \n" // rgb * alpha |
7616 | 0 | "pmullw %%xmm3,%%xmm1 \n" |
7617 | 0 | "paddw %%xmm7,%%xmm0 \n" // + 255 |
7618 | 0 | "paddw %%xmm7,%%xmm1 \n" |
7619 | 0 | "psrlw $0x8,%%xmm0 \n" |
7620 | 0 | "psrlw $0x8,%%xmm1 \n" |
7621 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7622 | 0 | "pand %%xmm5,%%xmm6 \n" |
7623 | 0 | "por %%xmm6,%%xmm0 \n" |
7624 | 0 | "movdqu %%xmm0,(%0,%1) \n" |
7625 | 0 | "lea 0x10(%0),%0 \n" |
7626 | 0 | "sub $0x4,%2 \n" |
7627 | 0 | "jg 1b \n" |
7628 | 0 | : "+r"(src_argb), // %0 |
7629 | 0 | "+r"(dst_argb), // %1 |
7630 | 0 | "+r"(width) // %2 |
7631 | 0 | : "m"(kAttenuateShuffle) // %3 |
7632 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7633 | 0 | "xmm7"); |
7634 | 0 | } |
7635 | | #endif // HAS_ARGBATTENUATEROW_SSSE3 |
7636 | | |
7637 | | #ifdef HAS_ARGBATTENUATEROW_AVX2 |
7638 | | |
7639 | | // Shuffle table duplicating alpha. |
7640 | | static const lvec8 kAttenuateShuffle_AVX2 = { |
7641 | | 6, -128, 6, -128, 6, -128, -128, -128, 14, -128, 14, |
7642 | | -128, 14, -128, -128, -128, 22, -128, 22, -128, 22, -128, |
7643 | | -128, -128, 30, -128, 30, -128, 30, -128, -128, -128}; |
7644 | | |
7645 | | // Attenuate 8 pixels at a time. |
7646 | | void ARGBAttenuateRow_AVX2(const uint8_t* src_argb, |
7647 | | uint8_t* dst_argb, |
7648 | 913 | int width) { |
7649 | 913 | asm volatile( |
7650 | 913 | "vmovdqa %3,%%ymm4 \n" |
7651 | 913 | "vpcmpeqb %%ymm5,%%ymm5,%%ymm5 \n" |
7652 | 913 | "vpslld $0x18,%%ymm5,%%ymm5 \n" |
7653 | 913 | "vpxor %%ymm6,%%ymm6,%%ymm6 \n" |
7654 | 913 | "vpcmpeqb %%ymm7,%%ymm7,%%ymm7 \n" |
7655 | 913 | "vpunpcklbw %%ymm6,%%ymm7,%%ymm7 \n" |
7656 | 913 | "sub %0,%1 \n" |
7657 | | |
7658 | | // 8 pixel loop. |
7659 | 913 | LABELALIGN |
7660 | 913 | "1: \n" |
7661 | 913 | "vmovdqu (%0),%%ymm6 \n" |
7662 | 913 | "vpunpcklbw %%ymm5,%%ymm6,%%ymm0 \n" |
7663 | 913 | "vpunpckhbw %%ymm5,%%ymm6,%%ymm1 \n" |
7664 | 913 | "vpshufb %%ymm4,%%ymm0,%%ymm2 \n" |
7665 | 913 | "vpshufb %%ymm4,%%ymm1,%%ymm3 \n" |
7666 | 913 | "vpmullw %%ymm2,%%ymm0,%%ymm0 \n" |
7667 | 913 | "vpmullw %%ymm3,%%ymm1,%%ymm1 \n" |
7668 | 913 | "vpaddw %%ymm7,%%ymm0,%%ymm0 \n" |
7669 | 913 | "vpaddw %%ymm7,%%ymm1,%%ymm1 \n" |
7670 | 913 | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
7671 | 913 | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
7672 | 913 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7673 | 913 | "vpand %%ymm5,%%ymm6,%%ymm1 \n" |
7674 | 913 | "vpor %%ymm1,%%ymm0,%%ymm0 \n" |
7675 | 913 | "vmovdqu %%ymm0,0x00(%0,%1,1) \n" |
7676 | 913 | "lea 0x20(%0),%0 \n" |
7677 | 913 | "sub $0x8,%2 \n" |
7678 | 913 | "jg 1b \n" |
7679 | 913 | "vzeroupper \n" |
7680 | 913 | : "+r"(src_argb), // %0 |
7681 | 913 | "+r"(dst_argb), // %1 |
7682 | 913 | "+r"(width) // %2 |
7683 | 913 | : "m"(kAttenuateShuffle_AVX2) // %3 |
7684 | 913 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7685 | 913 | "xmm7"); |
7686 | 913 | } |
7687 | | #endif // HAS_ARGBATTENUATEROW_AVX2 |
7688 | | |
7689 | | #ifdef HAS_ARGBUNATTENUATEROW_SSE2 |
7690 | | // Unattenuate 4 pixels at a time. |
7691 | | void ARGBUnattenuateRow_SSE2(const uint8_t* src_argb, |
7692 | | uint8_t* dst_argb, |
7693 | 0 | int width) { |
7694 | 0 | uintptr_t alpha; |
7695 | 0 | asm volatile( |
7696 | | // 4 pixel loop. |
7697 | 0 | LABELALIGN |
7698 | 0 | "1: \n" |
7699 | 0 | "movdqu (%0),%%xmm0 \n" |
7700 | 0 | "movzb 0x03(%0),%3 \n" |
7701 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
7702 | 0 | "movd 0x00(%4,%3,4),%%xmm2 \n" |
7703 | 0 | "movzb 0x07(%0),%3 \n" |
7704 | 0 | "movd 0x00(%4,%3,4),%%xmm3 \n" |
7705 | 0 | "pshuflw $0x40,%%xmm2,%%xmm2 \n" |
7706 | 0 | "pshuflw $0x40,%%xmm3,%%xmm3 \n" |
7707 | 0 | "movlhps %%xmm3,%%xmm2 \n" |
7708 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
7709 | 0 | "movdqu (%0),%%xmm1 \n" |
7710 | 0 | "movzb 0x0b(%0),%3 \n" |
7711 | 0 | "punpckhbw %%xmm1,%%xmm1 \n" |
7712 | 0 | "movd 0x00(%4,%3,4),%%xmm2 \n" |
7713 | 0 | "movzb 0x0f(%0),%3 \n" |
7714 | 0 | "movd 0x00(%4,%3,4),%%xmm3 \n" |
7715 | 0 | "pshuflw $0x40,%%xmm2,%%xmm2 \n" |
7716 | 0 | "pshuflw $0x40,%%xmm3,%%xmm3 \n" |
7717 | 0 | "movlhps %%xmm3,%%xmm2 \n" |
7718 | 0 | "pmulhuw %%xmm2,%%xmm1 \n" |
7719 | 0 | "lea 0x10(%0),%0 \n" |
7720 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
7721 | 0 | "movdqu %%xmm0,(%1) \n" |
7722 | 0 | "lea 0x10(%1),%1 \n" |
7723 | 0 | "sub $0x4,%2 \n" |
7724 | 0 | "jg 1b \n" |
7725 | 0 | : "+r"(src_argb), // %0 |
7726 | 0 | "+r"(dst_argb), // %1 |
7727 | 0 | "+r"(width), // %2 |
7728 | 0 | "=&r"(alpha) // %3 |
7729 | 0 | : "r"(fixed_invtbl8) // %4 |
7730 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
7731 | 0 | } |
7732 | | #endif // HAS_ARGBUNATTENUATEROW_SSE2 |
7733 | | |
7734 | | #ifdef HAS_ARGBUNATTENUATEROW_AVX2 |
7735 | | // Shuffle table duplicating alpha. |
7736 | | static const uvec8 kUnattenShuffleAlpha_AVX2 = { |
7737 | | 0u, 1u, 0u, 1u, 0u, 1u, 6u, 7u, 8u, 9u, 8u, 9u, 8u, 9u, 14u, 15u}; |
7738 | | // Unattenuate 8 pixels at a time. |
7739 | | void ARGBUnattenuateRow_AVX2(const uint8_t* src_argb, |
7740 | | uint8_t* dst_argb, |
7741 | 0 | int width) { |
7742 | 0 | uintptr_t alpha; |
7743 | 0 | asm volatile( |
7744 | 0 | "sub %0,%1 \n" |
7745 | 0 | "vbroadcasti128 %5,%%ymm5 \n" |
7746 | | |
7747 | | // 8 pixel loop. |
7748 | 0 | LABELALIGN |
7749 | 0 | "1: \n" |
7750 | | // replace VPGATHER |
7751 | 0 | "movzb 0x03(%0),%3 \n" |
7752 | 0 | "vmovd 0x00(%4,%3,4),%%xmm0 \n" |
7753 | 0 | "movzb 0x07(%0),%3 \n" |
7754 | 0 | "vmovd 0x00(%4,%3,4),%%xmm1 \n" |
7755 | 0 | "movzb 0x0b(%0),%3 \n" |
7756 | 0 | "vpunpckldq %%xmm1,%%xmm0,%%xmm6 \n" |
7757 | 0 | "vmovd 0x00(%4,%3,4),%%xmm2 \n" |
7758 | 0 | "movzb 0x0f(%0),%3 \n" |
7759 | 0 | "vmovd 0x00(%4,%3,4),%%xmm3 \n" |
7760 | 0 | "movzb 0x13(%0),%3 \n" |
7761 | 0 | "vpunpckldq %%xmm3,%%xmm2,%%xmm7 \n" |
7762 | 0 | "vmovd 0x00(%4,%3,4),%%xmm0 \n" |
7763 | 0 | "movzb 0x17(%0),%3 \n" |
7764 | 0 | "vmovd 0x00(%4,%3,4),%%xmm1 \n" |
7765 | 0 | "movzb 0x1b(%0),%3 \n" |
7766 | 0 | "vpunpckldq %%xmm1,%%xmm0,%%xmm0 \n" |
7767 | 0 | "vmovd 0x00(%4,%3,4),%%xmm2 \n" |
7768 | 0 | "movzb 0x1f(%0),%3 \n" |
7769 | 0 | "vmovd 0x00(%4,%3,4),%%xmm3 \n" |
7770 | 0 | "vpunpckldq %%xmm3,%%xmm2,%%xmm2 \n" |
7771 | 0 | "vpunpcklqdq %%xmm7,%%xmm6,%%xmm3 \n" |
7772 | 0 | "vpunpcklqdq %%xmm2,%%xmm0,%%xmm0 \n" |
7773 | 0 | "vinserti128 $0x1,%%xmm0,%%ymm3,%%ymm3 \n" |
7774 | | // end of VPGATHER |
7775 | |
|
7776 | 0 | "vmovdqu (%0),%%ymm6 \n" |
7777 | 0 | "vpunpcklbw %%ymm6,%%ymm6,%%ymm0 \n" |
7778 | 0 | "vpunpckhbw %%ymm6,%%ymm6,%%ymm1 \n" |
7779 | 0 | "vpunpcklwd %%ymm3,%%ymm3,%%ymm2 \n" |
7780 | 0 | "vpunpckhwd %%ymm3,%%ymm3,%%ymm3 \n" |
7781 | 0 | "vpshufb %%ymm5,%%ymm2,%%ymm2 \n" |
7782 | 0 | "vpshufb %%ymm5,%%ymm3,%%ymm3 \n" |
7783 | 0 | "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" |
7784 | 0 | "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" |
7785 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
7786 | 0 | "vmovdqu %%ymm0,0x00(%0,%1,1) \n" |
7787 | 0 | "lea 0x20(%0),%0 \n" |
7788 | 0 | "sub $0x8,%2 \n" |
7789 | 0 | "jg 1b \n" |
7790 | 0 | "vzeroupper \n" |
7791 | 0 | : "+r"(src_argb), // %0 |
7792 | 0 | "+r"(dst_argb), // %1 |
7793 | 0 | "+r"(width), // %2 |
7794 | 0 | "=&r"(alpha) // %3 |
7795 | 0 | : "r"(fixed_invtbl8), // %4 |
7796 | 0 | "m"(kUnattenShuffleAlpha_AVX2) // %5 |
7797 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7798 | 0 | "xmm7"); |
7799 | 0 | } |
7800 | | #endif // HAS_ARGBUNATTENUATEROW_AVX2 |
7801 | | |
7802 | | #ifdef HAS_ARGBGRAYROW_SSSE3 |
7803 | | // Convert 8 ARGB pixels (64 bytes) to 8 Gray ARGB pixels |
7804 | 0 | void ARGBGrayRow_SSSE3(const uint8_t* src_argb, uint8_t* dst_argb, int width) { |
7805 | 0 | asm volatile( |
7806 | 0 | "movdqa %3,%%xmm4 \n" |
7807 | 0 | "movdqa %4,%%xmm5 \n" |
7808 | | |
7809 | | // 8 pixel loop. |
7810 | 0 | LABELALIGN |
7811 | 0 | "1: \n" |
7812 | 0 | "movdqu (%0),%%xmm0 \n" |
7813 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7814 | 0 | "psubb %%xmm5,%%xmm0 \n" |
7815 | 0 | "psubb %%xmm5,%%xmm1 \n" |
7816 | 0 | "movdqu %%xmm4,%%xmm6 \n" |
7817 | 0 | "pmaddubsw %%xmm0,%%xmm6 \n" |
7818 | 0 | "movdqu %%xmm4,%%xmm0 \n" |
7819 | 0 | "pmaddubsw %%xmm1,%%xmm0 \n" |
7820 | 0 | "phaddw %%xmm0,%%xmm6 \n" |
7821 | 0 | "paddw %%xmm5,%%xmm6 \n" |
7822 | 0 | "psrlw $0x8,%%xmm6 \n" |
7823 | 0 | "packuswb %%xmm6,%%xmm6 \n" |
7824 | 0 | "movdqu (%0),%%xmm2 \n" |
7825 | 0 | "movdqu 0x10(%0),%%xmm3 \n" |
7826 | 0 | "lea 0x20(%0),%0 \n" |
7827 | 0 | "psrld $0x18,%%xmm2 \n" |
7828 | 0 | "psrld $0x18,%%xmm3 \n" |
7829 | 0 | "packuswb %%xmm3,%%xmm2 \n" |
7830 | 0 | "packuswb %%xmm2,%%xmm2 \n" |
7831 | 0 | "movdqa %%xmm6,%%xmm3 \n" |
7832 | 0 | "punpcklbw %%xmm6,%%xmm6 \n" |
7833 | 0 | "punpcklbw %%xmm2,%%xmm3 \n" |
7834 | 0 | "movdqa %%xmm6,%%xmm1 \n" |
7835 | 0 | "punpcklwd %%xmm3,%%xmm6 \n" |
7836 | 0 | "punpckhwd %%xmm3,%%xmm1 \n" |
7837 | 0 | "movdqu %%xmm6,(%1) \n" |
7838 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
7839 | 0 | "lea 0x20(%1),%1 \n" |
7840 | 0 | "sub $0x8,%2 \n" |
7841 | 0 | "jg 1b \n" |
7842 | 0 | : "+r"(src_argb), // %0 |
7843 | 0 | "+r"(dst_argb), // %1 |
7844 | 0 | "+r"(width) // %2 |
7845 | 0 | : "m"(kARGBToYJ), // %3 |
7846 | 0 | "m"(kSub128) // %4 |
7847 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
7848 | 0 | } |
7849 | | #endif // HAS_ARGBGRAYROW_SSSE3 |
7850 | | |
7851 | | #ifdef HAS_ARGBSEPIAROW_SSSE3 |
7852 | | // b = (r * 35 + g * 68 + b * 17) >> 7 |
7853 | | // g = (r * 45 + g * 88 + b * 22) >> 7 |
7854 | | // r = (r * 50 + g * 98 + b * 24) >> 7 |
7855 | | // Constant for ARGB color to sepia tone |
7856 | | static const vec8 kARGBToSepiaB = {17, 68, 35, 0, 17, 68, 35, 0, |
7857 | | 17, 68, 35, 0, 17, 68, 35, 0}; |
7858 | | |
7859 | | static const vec8 kARGBToSepiaG = {22, 88, 45, 0, 22, 88, 45, 0, |
7860 | | 22, 88, 45, 0, 22, 88, 45, 0}; |
7861 | | |
7862 | | static const vec8 kARGBToSepiaR = {24, 98, 50, 0, 24, 98, 50, 0, |
7863 | | 24, 98, 50, 0, 24, 98, 50, 0}; |
7864 | | |
7865 | | // Convert 8 ARGB pixels (32 bytes) to 8 Sepia ARGB pixels. |
7866 | 0 | void ARGBSepiaRow_SSSE3(uint8_t* dst_argb, int width) { |
7867 | 0 | asm volatile( |
7868 | 0 | "movdqa %2,%%xmm2 \n" |
7869 | 0 | "movdqa %3,%%xmm3 \n" |
7870 | 0 | "movdqa %4,%%xmm4 \n" |
7871 | | |
7872 | | // 8 pixel loop. |
7873 | 0 | LABELALIGN |
7874 | 0 | "1: \n" |
7875 | 0 | "movdqu (%0),%%xmm0 \n" |
7876 | 0 | "movdqu 0x10(%0),%%xmm6 \n" |
7877 | 0 | "pmaddubsw %%xmm2,%%xmm0 \n" |
7878 | 0 | "pmaddubsw %%xmm2,%%xmm6 \n" |
7879 | 0 | "phaddw %%xmm6,%%xmm0 \n" |
7880 | 0 | "psrlw $0x7,%%xmm0 \n" |
7881 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7882 | 0 | "movdqu (%0),%%xmm5 \n" |
7883 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7884 | 0 | "pmaddubsw %%xmm3,%%xmm5 \n" |
7885 | 0 | "pmaddubsw %%xmm3,%%xmm1 \n" |
7886 | 0 | "phaddw %%xmm1,%%xmm5 \n" |
7887 | 0 | "psrlw $0x7,%%xmm5 \n" |
7888 | 0 | "packuswb %%xmm5,%%xmm5 \n" |
7889 | 0 | "punpcklbw %%xmm5,%%xmm0 \n" |
7890 | 0 | "movdqu (%0),%%xmm5 \n" |
7891 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7892 | 0 | "pmaddubsw %%xmm4,%%xmm5 \n" |
7893 | 0 | "pmaddubsw %%xmm4,%%xmm1 \n" |
7894 | 0 | "phaddw %%xmm1,%%xmm5 \n" |
7895 | 0 | "psrlw $0x7,%%xmm5 \n" |
7896 | 0 | "packuswb %%xmm5,%%xmm5 \n" |
7897 | 0 | "movdqu (%0),%%xmm6 \n" |
7898 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7899 | 0 | "psrld $0x18,%%xmm6 \n" |
7900 | 0 | "psrld $0x18,%%xmm1 \n" |
7901 | 0 | "packuswb %%xmm1,%%xmm6 \n" |
7902 | 0 | "packuswb %%xmm6,%%xmm6 \n" |
7903 | 0 | "punpcklbw %%xmm6,%%xmm5 \n" |
7904 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
7905 | 0 | "punpcklwd %%xmm5,%%xmm0 \n" |
7906 | 0 | "punpckhwd %%xmm5,%%xmm1 \n" |
7907 | 0 | "movdqu %%xmm0,(%0) \n" |
7908 | 0 | "movdqu %%xmm1,0x10(%0) \n" |
7909 | 0 | "lea 0x20(%0),%0 \n" |
7910 | 0 | "sub $0x8,%1 \n" |
7911 | 0 | "jg 1b \n" |
7912 | 0 | : "+r"(dst_argb), // %0 |
7913 | 0 | "+r"(width) // %1 |
7914 | 0 | : "m"(kARGBToSepiaB), // %2 |
7915 | 0 | "m"(kARGBToSepiaG), // %3 |
7916 | 0 | "m"(kARGBToSepiaR) // %4 |
7917 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
7918 | 0 | } |
7919 | | #endif // HAS_ARGBSEPIAROW_SSSE3 |
7920 | | |
7921 | | #ifdef HAS_ARGBCOLORMATRIXROW_SSSE3 |
7922 | | // Tranform 8 ARGB pixels (32 bytes) with color matrix. |
7923 | | // Same as Sepia except matrix is provided. |
7924 | | void ARGBColorMatrixRow_SSSE3(const uint8_t* src_argb, |
7925 | | uint8_t* dst_argb, |
7926 | | const int8_t* matrix_argb, |
7927 | 0 | int width) { |
7928 | 0 | asm volatile( |
7929 | 0 | "movdqu (%3),%%xmm5 \n" |
7930 | 0 | "pshufd $0x00,%%xmm5,%%xmm2 \n" |
7931 | 0 | "pshufd $0x55,%%xmm5,%%xmm3 \n" |
7932 | 0 | "pshufd $0xaa,%%xmm5,%%xmm4 \n" |
7933 | 0 | "pshufd $0xff,%%xmm5,%%xmm5 \n" |
7934 | | |
7935 | | // 8 pixel loop. |
7936 | 0 | LABELALIGN |
7937 | 0 | "1: \n" |
7938 | 0 | "movdqu (%0),%%xmm0 \n" |
7939 | 0 | "movdqu 0x10(%0),%%xmm7 \n" |
7940 | 0 | "pmaddubsw %%xmm2,%%xmm0 \n" |
7941 | 0 | "pmaddubsw %%xmm2,%%xmm7 \n" |
7942 | 0 | "movdqu (%0),%%xmm6 \n" |
7943 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
7944 | 0 | "pmaddubsw %%xmm3,%%xmm6 \n" |
7945 | 0 | "pmaddubsw %%xmm3,%%xmm1 \n" |
7946 | 0 | "phaddsw %%xmm7,%%xmm0 \n" |
7947 | 0 | "phaddsw %%xmm1,%%xmm6 \n" |
7948 | 0 | "psraw $0x6,%%xmm0 \n" |
7949 | 0 | "psraw $0x6,%%xmm6 \n" |
7950 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
7951 | 0 | "packuswb %%xmm6,%%xmm6 \n" |
7952 | 0 | "punpcklbw %%xmm6,%%xmm0 \n" |
7953 | 0 | "movdqu (%0),%%xmm1 \n" |
7954 | 0 | "movdqu 0x10(%0),%%xmm7 \n" |
7955 | 0 | "pmaddubsw %%xmm4,%%xmm1 \n" |
7956 | 0 | "pmaddubsw %%xmm4,%%xmm7 \n" |
7957 | 0 | "phaddsw %%xmm7,%%xmm1 \n" |
7958 | 0 | "movdqu (%0),%%xmm6 \n" |
7959 | 0 | "movdqu 0x10(%0),%%xmm7 \n" |
7960 | 0 | "pmaddubsw %%xmm5,%%xmm6 \n" |
7961 | 0 | "pmaddubsw %%xmm5,%%xmm7 \n" |
7962 | 0 | "phaddsw %%xmm7,%%xmm6 \n" |
7963 | 0 | "psraw $0x6,%%xmm1 \n" |
7964 | 0 | "psraw $0x6,%%xmm6 \n" |
7965 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
7966 | 0 | "packuswb %%xmm6,%%xmm6 \n" |
7967 | 0 | "punpcklbw %%xmm6,%%xmm1 \n" |
7968 | 0 | "movdqa %%xmm0,%%xmm6 \n" |
7969 | 0 | "punpcklwd %%xmm1,%%xmm0 \n" |
7970 | 0 | "punpckhwd %%xmm1,%%xmm6 \n" |
7971 | 0 | "movdqu %%xmm0,(%1) \n" |
7972 | 0 | "movdqu %%xmm6,0x10(%1) \n" |
7973 | 0 | "lea 0x20(%0),%0 \n" |
7974 | 0 | "lea 0x20(%1),%1 \n" |
7975 | 0 | "sub $0x8,%2 \n" |
7976 | 0 | "jg 1b \n" |
7977 | 0 | : "+r"(src_argb), // %0 |
7978 | 0 | "+r"(dst_argb), // %1 |
7979 | 0 | "+r"(width) // %2 |
7980 | 0 | : "r"(matrix_argb) // %3 |
7981 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
7982 | 0 | "xmm7"); |
7983 | 0 | } |
7984 | | #endif // HAS_ARGBCOLORMATRIXROW_SSSE3 |
7985 | | |
7986 | | #ifdef HAS_ARGBQUANTIZEROW_SSE2 |
7987 | | // Quantize 4 ARGB pixels (16 bytes). |
7988 | | void ARGBQuantizeRow_SSE2(uint8_t* dst_argb, |
7989 | | int scale, |
7990 | | int interval_size, |
7991 | | int interval_offset, |
7992 | 0 | int width) { |
7993 | 0 | asm volatile( |
7994 | 0 | "movd %2,%%xmm2 \n" |
7995 | 0 | "movd %3,%%xmm3 \n" |
7996 | 0 | "movd %4,%%xmm4 \n" |
7997 | 0 | "pshuflw $0x40,%%xmm2,%%xmm2 \n" |
7998 | 0 | "pshufd $0x44,%%xmm2,%%xmm2 \n" |
7999 | 0 | "pshuflw $0x40,%%xmm3,%%xmm3 \n" |
8000 | 0 | "pshufd $0x44,%%xmm3,%%xmm3 \n" |
8001 | 0 | "pshuflw $0x40,%%xmm4,%%xmm4 \n" |
8002 | 0 | "pshufd $0x44,%%xmm4,%%xmm4 \n" |
8003 | 0 | "pxor %%xmm5,%%xmm5 \n" |
8004 | 0 | "pcmpeqb %%xmm6,%%xmm6 \n" |
8005 | 0 | "pslld $0x18,%%xmm6 \n" |
8006 | | |
8007 | | // 4 pixel loop. |
8008 | 0 | LABELALIGN |
8009 | 0 | "1: \n" |
8010 | 0 | "movdqu (%0),%%xmm0 \n" |
8011 | 0 | "punpcklbw %%xmm5,%%xmm0 \n" |
8012 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
8013 | 0 | "movdqu (%0),%%xmm1 \n" |
8014 | 0 | "punpckhbw %%xmm5,%%xmm1 \n" |
8015 | 0 | "pmulhuw %%xmm2,%%xmm1 \n" |
8016 | 0 | "pmullw %%xmm3,%%xmm0 \n" |
8017 | 0 | "movdqu (%0),%%xmm7 \n" |
8018 | 0 | "pmullw %%xmm3,%%xmm1 \n" |
8019 | 0 | "pand %%xmm6,%%xmm7 \n" |
8020 | 0 | "paddw %%xmm4,%%xmm0 \n" |
8021 | 0 | "paddw %%xmm4,%%xmm1 \n" |
8022 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
8023 | 0 | "por %%xmm7,%%xmm0 \n" |
8024 | 0 | "movdqu %%xmm0,(%0) \n" |
8025 | 0 | "lea 0x10(%0),%0 \n" |
8026 | 0 | "sub $0x4,%1 \n" |
8027 | 0 | "jg 1b \n" |
8028 | 0 | : "+r"(dst_argb), // %0 |
8029 | 0 | "+r"(width) // %1 |
8030 | 0 | : "r"(scale), // %2 |
8031 | 0 | "r"(interval_size), // %3 |
8032 | 0 | "r"(interval_offset) // %4 |
8033 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
8034 | 0 | "xmm7"); |
8035 | 0 | } |
8036 | | #endif // HAS_ARGBQUANTIZEROW_SSE2 |
8037 | | |
8038 | | #ifdef HAS_ARGBSHADEROW_SSE2 |
8039 | | // Shade 4 pixels at a time by specified value. |
8040 | | void ARGBShadeRow_SSE2(const uint8_t* src_argb, |
8041 | | uint8_t* dst_argb, |
8042 | | int width, |
8043 | 0 | uint32_t value) { |
8044 | 0 | asm volatile( |
8045 | 0 | "movd %3,%%xmm2 \n" |
8046 | 0 | "punpcklbw %%xmm2,%%xmm2 \n" |
8047 | 0 | "punpcklqdq %%xmm2,%%xmm2 \n" |
8048 | | |
8049 | | // 4 pixel loop. |
8050 | 0 | LABELALIGN |
8051 | 0 | "1: \n" |
8052 | 0 | "movdqu (%0),%%xmm0 \n" |
8053 | 0 | "lea 0x10(%0),%0 \n" |
8054 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
8055 | 0 | "punpcklbw %%xmm0,%%xmm0 \n" |
8056 | 0 | "punpckhbw %%xmm1,%%xmm1 \n" |
8057 | 0 | "pmulhuw %%xmm2,%%xmm0 \n" |
8058 | 0 | "pmulhuw %%xmm2,%%xmm1 \n" |
8059 | 0 | "psrlw $0x8,%%xmm0 \n" |
8060 | 0 | "psrlw $0x8,%%xmm1 \n" |
8061 | 0 | "packuswb %%xmm1,%%xmm0 \n" |
8062 | 0 | "movdqu %%xmm0,(%1) \n" |
8063 | 0 | "lea 0x10(%1),%1 \n" |
8064 | 0 | "sub $0x4,%2 \n" |
8065 | 0 | "jg 1b \n" |
8066 | 0 | : "+r"(src_argb), // %0 |
8067 | 0 | "+r"(dst_argb), // %1 |
8068 | 0 | "+r"(width) // %2 |
8069 | 0 | : "r"(value) // %3 |
8070 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
8071 | 0 | } |
8072 | | #endif // HAS_ARGBSHADEROW_SSE2 |
8073 | | |
8074 | | #ifdef HAS_ARGBMULTIPLYROW_SSE2 |
8075 | | // Multiply 2 rows of ARGB pixels together, 4 pixels at a time. |
8076 | | void ARGBMultiplyRow_SSE2(const uint8_t* src_argb, |
8077 | | const uint8_t* src_argb1, |
8078 | | uint8_t* dst_argb, |
8079 | | int width) { |
8080 | | asm volatile("pxor %%xmm5,%%xmm5 \n" |
8081 | | |
8082 | | // 4 pixel loop. |
8083 | | LABELALIGN |
8084 | | "1: \n" |
8085 | | "movdqu (%0),%%xmm0 \n" |
8086 | | "lea 0x10(%0),%0 \n" |
8087 | | "movdqu (%1),%%xmm2 \n" |
8088 | | "lea 0x10(%1),%1 \n" |
8089 | | "movdqu %%xmm0,%%xmm1 \n" |
8090 | | "movdqu %%xmm2,%%xmm3 \n" |
8091 | | "punpcklbw %%xmm0,%%xmm0 \n" |
8092 | | "punpckhbw %%xmm1,%%xmm1 \n" |
8093 | | "punpcklbw %%xmm5,%%xmm2 \n" |
8094 | | "punpckhbw %%xmm5,%%xmm3 \n" |
8095 | | "pmulhuw %%xmm2,%%xmm0 \n" |
8096 | | "pmulhuw %%xmm3,%%xmm1 \n" |
8097 | | "packuswb %%xmm1,%%xmm0 \n" |
8098 | | "movdqu %%xmm0,(%2) \n" |
8099 | | "lea 0x10(%2),%2 \n" |
8100 | | "sub $0x4,%3 \n" |
8101 | | "jg 1b \n" |
8102 | | : "+r"(src_argb), // %0 |
8103 | | "+r"(src_argb1), // %1 |
8104 | | "+r"(dst_argb), // %2 |
8105 | | "+r"(width) // %3 |
8106 | | : |
8107 | | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
8108 | | } |
8109 | | #endif // HAS_ARGBMULTIPLYROW_SSE2 |
8110 | | |
8111 | | #ifdef HAS_ARGBMULTIPLYROW_AVX2 |
8112 | | // Multiply 2 rows of ARGB pixels together, 8 pixels at a time. |
8113 | | void ARGBMultiplyRow_AVX2(const uint8_t* src_argb, |
8114 | | const uint8_t* src_argb1, |
8115 | | uint8_t* dst_argb, |
8116 | | int width) { |
8117 | | asm volatile("vpxor %%ymm5,%%ymm5,%%ymm5 \n" |
8118 | | |
8119 | | // 4 pixel loop. |
8120 | | LABELALIGN |
8121 | | "1: \n" |
8122 | | "vmovdqu (%0),%%ymm1 \n" |
8123 | | "lea 0x20(%0),%0 \n" |
8124 | | "vmovdqu (%1),%%ymm3 \n" |
8125 | | "lea 0x20(%1),%1 \n" |
8126 | | "vpunpcklbw %%ymm1,%%ymm1,%%ymm0 \n" |
8127 | | "vpunpckhbw %%ymm1,%%ymm1,%%ymm1 \n" |
8128 | | "vpunpcklbw %%ymm5,%%ymm3,%%ymm2 \n" |
8129 | | "vpunpckhbw %%ymm5,%%ymm3,%%ymm3 \n" |
8130 | | "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" |
8131 | | "vpmulhuw %%ymm3,%%ymm1,%%ymm1 \n" |
8132 | | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
8133 | | "vmovdqu %%ymm0,(%2) \n" |
8134 | | "lea 0x20(%2),%2 \n" |
8135 | | "sub $0x8,%3 \n" |
8136 | | "jg 1b \n" |
8137 | | "vzeroupper \n" |
8138 | | : "+r"(src_argb), // %0 |
8139 | | "+r"(src_argb1), // %1 |
8140 | | "+r"(dst_argb), // %2 |
8141 | | "+r"(width) // %3 |
8142 | | : |
8143 | | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
8144 | | } |
8145 | | #endif // HAS_ARGBMULTIPLYROW_AVX2 |
8146 | | |
8147 | | #ifdef HAS_ARGBADDROW_SSE2 |
8148 | | // Add 2 rows of ARGB pixels together, 4 pixels at a time. |
8149 | | void ARGBAddRow_SSE2(const uint8_t* src_argb, |
8150 | | const uint8_t* src_argb1, |
8151 | | uint8_t* dst_argb, |
8152 | 0 | int width) { |
8153 | 0 | asm volatile( |
8154 | | // 4 pixel loop. |
8155 | 0 | LABELALIGN |
8156 | 0 | "1: \n" |
8157 | 0 | "movdqu (%0),%%xmm0 \n" |
8158 | 0 | "lea 0x10(%0),%0 \n" |
8159 | 0 | "movdqu (%1),%%xmm1 \n" |
8160 | 0 | "lea 0x10(%1),%1 \n" |
8161 | 0 | "paddusb %%xmm1,%%xmm0 \n" |
8162 | 0 | "movdqu %%xmm0,(%2) \n" |
8163 | 0 | "lea 0x10(%2),%2 \n" |
8164 | 0 | "sub $0x4,%3 \n" |
8165 | 0 | "jg 1b \n" |
8166 | 0 | : "+r"(src_argb), // %0 |
8167 | 0 | "+r"(src_argb1), // %1 |
8168 | 0 | "+r"(dst_argb), // %2 |
8169 | 0 | "+r"(width) // %3 |
8170 | 0 | : |
8171 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
8172 | 0 | } |
8173 | | #endif // HAS_ARGBADDROW_SSE2 |
8174 | | |
8175 | | #ifdef HAS_ARGBADDROW_AVX2 |
8176 | | // Add 2 rows of ARGB pixels together, 4 pixels at a time. |
8177 | | void ARGBAddRow_AVX2(const uint8_t* src_argb, |
8178 | | const uint8_t* src_argb1, |
8179 | | uint8_t* dst_argb, |
8180 | 0 | int width) { |
8181 | 0 | asm volatile( |
8182 | | // 4 pixel loop. |
8183 | 0 | LABELALIGN |
8184 | 0 | "1: \n" |
8185 | 0 | "vmovdqu (%0),%%ymm0 \n" |
8186 | 0 | "lea 0x20(%0),%0 \n" |
8187 | 0 | "vpaddusb (%1),%%ymm0,%%ymm0 \n" |
8188 | 0 | "lea 0x20(%1),%1 \n" |
8189 | 0 | "vmovdqu %%ymm0,(%2) \n" |
8190 | 0 | "lea 0x20(%2),%2 \n" |
8191 | 0 | "sub $0x8,%3 \n" |
8192 | 0 | "jg 1b \n" |
8193 | 0 | "vzeroupper \n" |
8194 | 0 | : "+r"(src_argb), // %0 |
8195 | 0 | "+r"(src_argb1), // %1 |
8196 | 0 | "+r"(dst_argb), // %2 |
8197 | 0 | "+r"(width) // %3 |
8198 | 0 | : |
8199 | 0 | : "memory", "cc", "xmm0"); |
8200 | 0 | } |
8201 | | #endif // HAS_ARGBADDROW_AVX2 |
8202 | | |
8203 | | #ifdef HAS_ARGBSUBTRACTROW_SSE2 |
8204 | | // Subtract 2 rows of ARGB pixels, 4 pixels at a time. |
8205 | | void ARGBSubtractRow_SSE2(const uint8_t* src_argb, |
8206 | | const uint8_t* src_argb1, |
8207 | | uint8_t* dst_argb, |
8208 | 0 | int width) { |
8209 | 0 | asm volatile( |
8210 | | // 4 pixel loop. |
8211 | 0 | LABELALIGN |
8212 | 0 | "1: \n" |
8213 | 0 | "movdqu (%0),%%xmm0 \n" |
8214 | 0 | "lea 0x10(%0),%0 \n" |
8215 | 0 | "movdqu (%1),%%xmm1 \n" |
8216 | 0 | "lea 0x10(%1),%1 \n" |
8217 | 0 | "psubusb %%xmm1,%%xmm0 \n" |
8218 | 0 | "movdqu %%xmm0,(%2) \n" |
8219 | 0 | "lea 0x10(%2),%2 \n" |
8220 | 0 | "sub $0x4,%3 \n" |
8221 | 0 | "jg 1b \n" |
8222 | 0 | : "+r"(src_argb), // %0 |
8223 | 0 | "+r"(src_argb1), // %1 |
8224 | 0 | "+r"(dst_argb), // %2 |
8225 | 0 | "+r"(width) // %3 |
8226 | 0 | : |
8227 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
8228 | 0 | } |
8229 | | #endif // HAS_ARGBSUBTRACTROW_SSE2 |
8230 | | |
8231 | | #ifdef HAS_ARGBSUBTRACTROW_AVX2 |
8232 | | // Subtract 2 rows of ARGB pixels, 8 pixels at a time. |
8233 | | void ARGBSubtractRow_AVX2(const uint8_t* src_argb, |
8234 | | const uint8_t* src_argb1, |
8235 | | uint8_t* dst_argb, |
8236 | 0 | int width) { |
8237 | 0 | asm volatile( |
8238 | | // 4 pixel loop. |
8239 | 0 | LABELALIGN |
8240 | 0 | "1: \n" |
8241 | 0 | "vmovdqu (%0),%%ymm0 \n" |
8242 | 0 | "lea 0x20(%0),%0 \n" |
8243 | 0 | "vpsubusb (%1),%%ymm0,%%ymm0 \n" |
8244 | 0 | "lea 0x20(%1),%1 \n" |
8245 | 0 | "vmovdqu %%ymm0,(%2) \n" |
8246 | 0 | "lea 0x20(%2),%2 \n" |
8247 | 0 | "sub $0x8,%3 \n" |
8248 | 0 | "jg 1b \n" |
8249 | 0 | "vzeroupper \n" |
8250 | 0 | : "+r"(src_argb), // %0 |
8251 | 0 | "+r"(src_argb1), // %1 |
8252 | 0 | "+r"(dst_argb), // %2 |
8253 | 0 | "+r"(width) // %3 |
8254 | 0 | : |
8255 | 0 | : "memory", "cc", "xmm0"); |
8256 | 0 | } |
8257 | | #endif // HAS_ARGBSUBTRACTROW_AVX2 |
8258 | | |
8259 | | #ifdef HAS_SOBELXROW_SSE2 |
8260 | | // SobelX as a matrix is |
8261 | | // -1 0 1 |
8262 | | // -2 0 2 |
8263 | | // -1 0 1 |
8264 | | void SobelXRow_SSE2(const uint8_t* src_y0, |
8265 | | const uint8_t* src_y1, |
8266 | | const uint8_t* src_y2, |
8267 | | uint8_t* dst_sobelx, |
8268 | 0 | int width) { |
8269 | 0 | asm volatile( |
8270 | 0 | "sub %0,%1 \n" |
8271 | 0 | "sub %0,%2 \n" |
8272 | 0 | "sub %0,%3 \n" |
8273 | 0 | "pxor %%xmm5,%%xmm5 \n" |
8274 | | |
8275 | | // 8 pixel loop. |
8276 | 0 | LABELALIGN |
8277 | 0 | "1: \n" |
8278 | 0 | "movq (%0),%%xmm0 \n" |
8279 | 0 | "movq 0x2(%0),%%xmm1 \n" |
8280 | 0 | "punpcklbw %%xmm5,%%xmm0 \n" |
8281 | 0 | "punpcklbw %%xmm5,%%xmm1 \n" |
8282 | 0 | "psubw %%xmm1,%%xmm0 \n" |
8283 | 0 | "movq 0x00(%0,%1,1),%%xmm1 \n" |
8284 | 0 | "movq 0x02(%0,%1,1),%%xmm2 \n" |
8285 | 0 | "punpcklbw %%xmm5,%%xmm1 \n" |
8286 | 0 | "punpcklbw %%xmm5,%%xmm2 \n" |
8287 | 0 | "psubw %%xmm2,%%xmm1 \n" |
8288 | 0 | "movq 0x00(%0,%2,1),%%xmm2 \n" |
8289 | 0 | "movq 0x02(%0,%2,1),%%xmm3 \n" |
8290 | 0 | "punpcklbw %%xmm5,%%xmm2 \n" |
8291 | 0 | "punpcklbw %%xmm5,%%xmm3 \n" |
8292 | 0 | "psubw %%xmm3,%%xmm2 \n" |
8293 | 0 | "paddw %%xmm2,%%xmm0 \n" |
8294 | 0 | "paddw %%xmm1,%%xmm0 \n" |
8295 | 0 | "paddw %%xmm1,%%xmm0 \n" |
8296 | 0 | "pxor %%xmm1,%%xmm1 \n" |
8297 | 0 | "psubw %%xmm0,%%xmm1 \n" |
8298 | 0 | "pmaxsw %%xmm1,%%xmm0 \n" |
8299 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
8300 | 0 | "movq %%xmm0,0x00(%0,%3,1) \n" |
8301 | 0 | "lea 0x8(%0),%0 \n" |
8302 | 0 | "sub $0x8,%4 \n" |
8303 | 0 | "jg 1b \n" |
8304 | 0 | : "+r"(src_y0), // %0 |
8305 | 0 | "+r"(src_y1), // %1 |
8306 | 0 | "+r"(src_y2), // %2 |
8307 | 0 | "+r"(dst_sobelx), // %3 |
8308 | 0 | "+r"(width) // %4 |
8309 | 0 | : |
8310 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
8311 | 0 | } |
8312 | | #endif // HAS_SOBELXROW_SSE2 |
8313 | | |
8314 | | #ifdef HAS_SOBELYROW_SSE2 |
8315 | | // SobelY as a matrix is |
8316 | | // -1 -2 -1 |
8317 | | // 0 0 0 |
8318 | | // 1 2 1 |
8319 | | void SobelYRow_SSE2(const uint8_t* src_y0, |
8320 | | const uint8_t* src_y1, |
8321 | | uint8_t* dst_sobely, |
8322 | 0 | int width) { |
8323 | 0 | asm volatile( |
8324 | 0 | "sub %0,%1 \n" |
8325 | 0 | "sub %0,%2 \n" |
8326 | 0 | "pxor %%xmm5,%%xmm5 \n" |
8327 | | |
8328 | | // 8 pixel loop. |
8329 | 0 | LABELALIGN |
8330 | 0 | "1: \n" |
8331 | 0 | "movq (%0),%%xmm0 \n" |
8332 | 0 | "movq 0x00(%0,%1,1),%%xmm1 \n" |
8333 | 0 | "punpcklbw %%xmm5,%%xmm0 \n" |
8334 | 0 | "punpcklbw %%xmm5,%%xmm1 \n" |
8335 | 0 | "psubw %%xmm1,%%xmm0 \n" |
8336 | 0 | "movq 0x1(%0),%%xmm1 \n" |
8337 | 0 | "movq 0x01(%0,%1,1),%%xmm2 \n" |
8338 | 0 | "punpcklbw %%xmm5,%%xmm1 \n" |
8339 | 0 | "punpcklbw %%xmm5,%%xmm2 \n" |
8340 | 0 | "psubw %%xmm2,%%xmm1 \n" |
8341 | 0 | "movq 0x2(%0),%%xmm2 \n" |
8342 | 0 | "movq 0x02(%0,%1,1),%%xmm3 \n" |
8343 | 0 | "punpcklbw %%xmm5,%%xmm2 \n" |
8344 | 0 | "punpcklbw %%xmm5,%%xmm3 \n" |
8345 | 0 | "psubw %%xmm3,%%xmm2 \n" |
8346 | 0 | "paddw %%xmm2,%%xmm0 \n" |
8347 | 0 | "paddw %%xmm1,%%xmm0 \n" |
8348 | 0 | "paddw %%xmm1,%%xmm0 \n" |
8349 | 0 | "pxor %%xmm1,%%xmm1 \n" |
8350 | 0 | "psubw %%xmm0,%%xmm1 \n" |
8351 | 0 | "pmaxsw %%xmm1,%%xmm0 \n" |
8352 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
8353 | 0 | "movq %%xmm0,0x00(%0,%2,1) \n" |
8354 | 0 | "lea 0x8(%0),%0 \n" |
8355 | 0 | "sub $0x8,%3 \n" |
8356 | 0 | "jg 1b \n" |
8357 | 0 | : "+r"(src_y0), // %0 |
8358 | 0 | "+r"(src_y1), // %1 |
8359 | 0 | "+r"(dst_sobely), // %2 |
8360 | 0 | "+r"(width) // %3 |
8361 | 0 | : |
8362 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
8363 | 0 | } |
8364 | | #endif // HAS_SOBELYROW_SSE2 |
8365 | | |
8366 | | #ifdef HAS_SOBELROW_SSE2 |
8367 | | // Adds Sobel X and Sobel Y and stores Sobel into ARGB. |
8368 | | // A = 255 |
8369 | | // R = Sobel |
8370 | | // G = Sobel |
8371 | | // B = Sobel |
8372 | | void SobelRow_SSE2(const uint8_t* src_sobelx, |
8373 | | const uint8_t* src_sobely, |
8374 | | uint8_t* dst_argb, |
8375 | 0 | int width) { |
8376 | 0 | asm volatile( |
8377 | 0 | "sub %0,%1 \n" |
8378 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
8379 | 0 | "pslld $0x18,%%xmm5 \n" |
8380 | | |
8381 | | // 8 pixel loop. |
8382 | 0 | LABELALIGN |
8383 | 0 | "1: \n" |
8384 | 0 | "movdqu (%0),%%xmm0 \n" |
8385 | 0 | "movdqu 0x00(%0,%1,1),%%xmm1 \n" |
8386 | 0 | "lea 0x10(%0),%0 \n" |
8387 | 0 | "paddusb %%xmm1,%%xmm0 \n" |
8388 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
8389 | 0 | "punpcklbw %%xmm0,%%xmm2 \n" |
8390 | 0 | "punpckhbw %%xmm0,%%xmm0 \n" |
8391 | 0 | "movdqa %%xmm2,%%xmm1 \n" |
8392 | 0 | "punpcklwd %%xmm2,%%xmm1 \n" |
8393 | 0 | "punpckhwd %%xmm2,%%xmm2 \n" |
8394 | 0 | "por %%xmm5,%%xmm1 \n" |
8395 | 0 | "por %%xmm5,%%xmm2 \n" |
8396 | 0 | "movdqa %%xmm0,%%xmm3 \n" |
8397 | 0 | "punpcklwd %%xmm0,%%xmm3 \n" |
8398 | 0 | "punpckhwd %%xmm0,%%xmm0 \n" |
8399 | 0 | "por %%xmm5,%%xmm3 \n" |
8400 | 0 | "por %%xmm5,%%xmm0 \n" |
8401 | 0 | "movdqu %%xmm1,(%2) \n" |
8402 | 0 | "movdqu %%xmm2,0x10(%2) \n" |
8403 | 0 | "movdqu %%xmm3,0x20(%2) \n" |
8404 | 0 | "movdqu %%xmm0,0x30(%2) \n" |
8405 | 0 | "lea 0x40(%2),%2 \n" |
8406 | 0 | "sub $0x10,%3 \n" |
8407 | 0 | "jg 1b \n" |
8408 | 0 | : "+r"(src_sobelx), // %0 |
8409 | 0 | "+r"(src_sobely), // %1 |
8410 | 0 | "+r"(dst_argb), // %2 |
8411 | 0 | "+r"(width) // %3 |
8412 | 0 | : |
8413 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm5"); |
8414 | 0 | } |
8415 | | #endif // HAS_SOBELROW_SSE2 |
8416 | | |
8417 | | #ifdef HAS_SOBELTOPLANEROW_SSE2 |
8418 | | // Adds Sobel X and Sobel Y and stores Sobel into a plane. |
8419 | | void SobelToPlaneRow_SSE2(const uint8_t* src_sobelx, |
8420 | | const uint8_t* src_sobely, |
8421 | | uint8_t* dst_y, |
8422 | 0 | int width) { |
8423 | 0 | asm volatile( |
8424 | 0 | "sub %0,%1 \n" |
8425 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
8426 | 0 | "pslld $0x18,%%xmm5 \n" |
8427 | | |
8428 | | // 8 pixel loop. |
8429 | 0 | LABELALIGN |
8430 | 0 | "1: \n" |
8431 | 0 | "movdqu (%0),%%xmm0 \n" |
8432 | 0 | "movdqu 0x00(%0,%1,1),%%xmm1 \n" |
8433 | 0 | "lea 0x10(%0),%0 \n" |
8434 | 0 | "paddusb %%xmm1,%%xmm0 \n" |
8435 | 0 | "movdqu %%xmm0,(%2) \n" |
8436 | 0 | "lea 0x10(%2),%2 \n" |
8437 | 0 | "sub $0x10,%3 \n" |
8438 | 0 | "jg 1b \n" |
8439 | 0 | : "+r"(src_sobelx), // %0 |
8440 | 0 | "+r"(src_sobely), // %1 |
8441 | 0 | "+r"(dst_y), // %2 |
8442 | 0 | "+r"(width) // %3 |
8443 | 0 | : |
8444 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
8445 | 0 | } |
8446 | | #endif // HAS_SOBELTOPLANEROW_SSE2 |
8447 | | |
8448 | | #ifdef HAS_SOBELXYROW_SSE2 |
8449 | | // Mixes Sobel X, Sobel Y and Sobel into ARGB. |
8450 | | // A = 255 |
8451 | | // R = Sobel X |
8452 | | // G = Sobel |
8453 | | // B = Sobel Y |
8454 | | void SobelXYRow_SSE2(const uint8_t* src_sobelx, |
8455 | | const uint8_t* src_sobely, |
8456 | | uint8_t* dst_argb, |
8457 | 0 | int width) { |
8458 | 0 | asm volatile( |
8459 | 0 | "sub %0,%1 \n" |
8460 | 0 | "pcmpeqb %%xmm5,%%xmm5 \n" |
8461 | | |
8462 | | // 8 pixel loop. |
8463 | 0 | LABELALIGN |
8464 | 0 | "1: \n" |
8465 | 0 | "movdqu (%0),%%xmm0 \n" |
8466 | 0 | "movdqu 0x00(%0,%1,1),%%xmm1 \n" |
8467 | 0 | "lea 0x10(%0),%0 \n" |
8468 | 0 | "movdqa %%xmm0,%%xmm2 \n" |
8469 | 0 | "paddusb %%xmm1,%%xmm2 \n" |
8470 | 0 | "movdqa %%xmm0,%%xmm3 \n" |
8471 | 0 | "punpcklbw %%xmm5,%%xmm3 \n" |
8472 | 0 | "punpckhbw %%xmm5,%%xmm0 \n" |
8473 | 0 | "movdqa %%xmm1,%%xmm4 \n" |
8474 | 0 | "punpcklbw %%xmm2,%%xmm4 \n" |
8475 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" |
8476 | 0 | "movdqa %%xmm4,%%xmm6 \n" |
8477 | 0 | "punpcklwd %%xmm3,%%xmm6 \n" |
8478 | 0 | "punpckhwd %%xmm3,%%xmm4 \n" |
8479 | 0 | "movdqa %%xmm1,%%xmm7 \n" |
8480 | 0 | "punpcklwd %%xmm0,%%xmm7 \n" |
8481 | 0 | "punpckhwd %%xmm0,%%xmm1 \n" |
8482 | 0 | "movdqu %%xmm6,(%2) \n" |
8483 | 0 | "movdqu %%xmm4,0x10(%2) \n" |
8484 | 0 | "movdqu %%xmm7,0x20(%2) \n" |
8485 | 0 | "movdqu %%xmm1,0x30(%2) \n" |
8486 | 0 | "lea 0x40(%2),%2 \n" |
8487 | 0 | "sub $0x10,%3 \n" |
8488 | 0 | "jg 1b \n" |
8489 | 0 | : "+r"(src_sobelx), // %0 |
8490 | 0 | "+r"(src_sobely), // %1 |
8491 | 0 | "+r"(dst_argb), // %2 |
8492 | 0 | "+r"(width) // %3 |
8493 | 0 | : |
8494 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
8495 | 0 | "xmm7"); |
8496 | 0 | } |
8497 | | #endif // HAS_SOBELXYROW_SSE2 |
8498 | | |
8499 | | #ifdef HAS_COMPUTECUMULATIVESUMROW_SSE2 |
8500 | | // Creates a table of cumulative sums where each value is a sum of all values |
8501 | | // above and to the left of the value, inclusive of the value. |
8502 | | void ComputeCumulativeSumRow_SSE2(const uint8_t* row, |
8503 | | int32_t* cumsum, |
8504 | | const int32_t* previous_cumsum, |
8505 | 0 | int width) { |
8506 | 0 | asm volatile( |
8507 | 0 | "pxor %%xmm0,%%xmm0 \n" |
8508 | 0 | "pxor %%xmm1,%%xmm1 \n" |
8509 | 0 | "sub $0x4,%3 \n" |
8510 | 0 | "jl 49f \n" |
8511 | 0 | "test $0xf,%1 \n" |
8512 | 0 | "jne 49f \n" |
8513 | | |
8514 | | // 4 pixel loop. |
8515 | 0 | LABELALIGN |
8516 | 0 | "40: \n" |
8517 | 0 | "movdqu (%0),%%xmm2 \n" |
8518 | 0 | "lea 0x10(%0),%0 \n" |
8519 | 0 | "movdqa %%xmm2,%%xmm4 \n" |
8520 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" |
8521 | 0 | "movdqa %%xmm2,%%xmm3 \n" |
8522 | 0 | "punpcklwd %%xmm1,%%xmm2 \n" |
8523 | 0 | "punpckhwd %%xmm1,%%xmm3 \n" |
8524 | 0 | "punpckhbw %%xmm1,%%xmm4 \n" |
8525 | 0 | "movdqa %%xmm4,%%xmm5 \n" |
8526 | 0 | "punpcklwd %%xmm1,%%xmm4 \n" |
8527 | 0 | "punpckhwd %%xmm1,%%xmm5 \n" |
8528 | 0 | "paddd %%xmm2,%%xmm0 \n" |
8529 | 0 | "movdqu (%2),%%xmm2 \n" |
8530 | 0 | "paddd %%xmm0,%%xmm2 \n" |
8531 | 0 | "paddd %%xmm3,%%xmm0 \n" |
8532 | 0 | "movdqu 0x10(%2),%%xmm3 \n" |
8533 | 0 | "paddd %%xmm0,%%xmm3 \n" |
8534 | 0 | "paddd %%xmm4,%%xmm0 \n" |
8535 | 0 | "movdqu 0x20(%2),%%xmm4 \n" |
8536 | 0 | "paddd %%xmm0,%%xmm4 \n" |
8537 | 0 | "paddd %%xmm5,%%xmm0 \n" |
8538 | 0 | "movdqu 0x30(%2),%%xmm5 \n" |
8539 | 0 | "lea 0x40(%2),%2 \n" |
8540 | 0 | "paddd %%xmm0,%%xmm5 \n" |
8541 | 0 | "movdqu %%xmm2,(%1) \n" |
8542 | 0 | "movdqu %%xmm3,0x10(%1) \n" |
8543 | 0 | "movdqu %%xmm4,0x20(%1) \n" |
8544 | 0 | "movdqu %%xmm5,0x30(%1) \n" |
8545 | 0 | "lea 0x40(%1),%1 \n" |
8546 | 0 | "sub $0x4,%3 \n" |
8547 | 0 | "jge 40b \n" |
8548 | |
|
8549 | 0 | "49: \n" |
8550 | 0 | "add $0x3,%3 \n" |
8551 | 0 | "jl 19f \n" |
8552 | | |
8553 | | // 1 pixel loop. |
8554 | 0 | LABELALIGN |
8555 | 0 | "10: \n" |
8556 | 0 | "movd (%0),%%xmm2 \n" |
8557 | 0 | "lea 0x4(%0),%0 \n" |
8558 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" |
8559 | 0 | "punpcklwd %%xmm1,%%xmm2 \n" |
8560 | 0 | "paddd %%xmm2,%%xmm0 \n" |
8561 | 0 | "movdqu (%2),%%xmm2 \n" |
8562 | 0 | "lea 0x10(%2),%2 \n" |
8563 | 0 | "paddd %%xmm0,%%xmm2 \n" |
8564 | 0 | "movdqu %%xmm2,(%1) \n" |
8565 | 0 | "lea 0x10(%1),%1 \n" |
8566 | 0 | "sub $0x1,%3 \n" |
8567 | 0 | "jge 10b \n" |
8568 | |
|
8569 | 0 | "19: \n" |
8570 | 0 | : "+r"(row), // %0 |
8571 | 0 | "+r"(cumsum), // %1 |
8572 | 0 | "+r"(previous_cumsum), // %2 |
8573 | 0 | "+r"(width) // %3 |
8574 | 0 | : |
8575 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
8576 | 0 | } |
8577 | | #endif // HAS_COMPUTECUMULATIVESUMROW_SSE2 |
8578 | | |
8579 | | #ifdef HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 |
8580 | | void CumulativeSumToAverageRow_SSE2(const int32_t* topleft, |
8581 | | const int32_t* botleft, |
8582 | | int width, |
8583 | | int area, |
8584 | | uint8_t* dst, |
8585 | 0 | int count) { |
8586 | 0 | asm volatile( |
8587 | 0 | "movd %5,%%xmm5 \n" |
8588 | 0 | "cvtdq2ps %%xmm5,%%xmm5 \n" |
8589 | 0 | "rcpss %%xmm5,%%xmm4 \n" |
8590 | 0 | "pshufd $0x0,%%xmm4,%%xmm4 \n" |
8591 | 0 | "sub $0x4,%3 \n" |
8592 | 0 | "jl 49f \n" |
8593 | 0 | "cmpl $0x80,%5 \n" |
8594 | 0 | "ja 40f \n" |
8595 | |
|
8596 | 0 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
8597 | 0 | "pcmpeqb %%xmm6,%%xmm6 \n" |
8598 | 0 | "psrld $0x10,%%xmm6 \n" |
8599 | 0 | "cvtdq2ps %%xmm6,%%xmm6 \n" |
8600 | 0 | "addps %%xmm6,%%xmm5 \n" |
8601 | 0 | "mulps %%xmm4,%%xmm5 \n" |
8602 | 0 | "cvtps2dq %%xmm5,%%xmm5 \n" |
8603 | 0 | "packssdw %%xmm5,%%xmm5 \n" |
8604 | | |
8605 | | // 4 pixel small loop. |
8606 | 0 | LABELALIGN |
8607 | 0 | "4: \n" |
8608 | 0 | "movdqu (%0),%%xmm0 \n" |
8609 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
8610 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
8611 | 0 | "movdqu 0x30(%0),%%xmm3 \n" |
8612 | 0 | "psubd 0x00(%0,%4,4),%%xmm0 \n" |
8613 | 0 | "psubd 0x10(%0,%4,4),%%xmm1 \n" |
8614 | 0 | "psubd 0x20(%0,%4,4),%%xmm2 \n" |
8615 | 0 | "psubd 0x30(%0,%4,4),%%xmm3 \n" |
8616 | 0 | "lea 0x40(%0),%0 \n" |
8617 | 0 | "psubd (%1),%%xmm0 \n" |
8618 | 0 | "psubd 0x10(%1),%%xmm1 \n" |
8619 | 0 | "psubd 0x20(%1),%%xmm2 \n" |
8620 | 0 | "psubd 0x30(%1),%%xmm3 \n" |
8621 | 0 | "paddd 0x00(%1,%4,4),%%xmm0 \n" |
8622 | 0 | "paddd 0x10(%1,%4,4),%%xmm1 \n" |
8623 | 0 | "paddd 0x20(%1,%4,4),%%xmm2 \n" |
8624 | 0 | "paddd 0x30(%1,%4,4),%%xmm3 \n" |
8625 | 0 | "lea 0x40(%1),%1 \n" |
8626 | 0 | "packssdw %%xmm1,%%xmm0 \n" |
8627 | 0 | "packssdw %%xmm3,%%xmm2 \n" |
8628 | 0 | "pmulhuw %%xmm5,%%xmm0 \n" |
8629 | 0 | "pmulhuw %%xmm5,%%xmm2 \n" |
8630 | 0 | "packuswb %%xmm2,%%xmm0 \n" |
8631 | 0 | "movdqu %%xmm0,(%2) \n" |
8632 | 0 | "lea 0x10(%2),%2 \n" |
8633 | 0 | "sub $0x4,%3 \n" |
8634 | 0 | "jge 4b \n" |
8635 | 0 | "jmp 49f \n" |
8636 | | |
8637 | | // 4 pixel loop |
8638 | 0 | LABELALIGN |
8639 | 0 | "40: \n" |
8640 | 0 | "movdqu (%0),%%xmm0 \n" |
8641 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
8642 | 0 | "movdqu 0x20(%0),%%xmm2 \n" |
8643 | 0 | "movdqu 0x30(%0),%%xmm3 \n" |
8644 | 0 | "psubd 0x00(%0,%4,4),%%xmm0 \n" |
8645 | 0 | "psubd 0x10(%0,%4,4),%%xmm1 \n" |
8646 | 0 | "psubd 0x20(%0,%4,4),%%xmm2 \n" |
8647 | 0 | "psubd 0x30(%0,%4,4),%%xmm3 \n" |
8648 | 0 | "lea 0x40(%0),%0 \n" |
8649 | 0 | "psubd (%1),%%xmm0 \n" |
8650 | 0 | "psubd 0x10(%1),%%xmm1 \n" |
8651 | 0 | "psubd 0x20(%1),%%xmm2 \n" |
8652 | 0 | "psubd 0x30(%1),%%xmm3 \n" |
8653 | 0 | "paddd 0x00(%1,%4,4),%%xmm0 \n" |
8654 | 0 | "paddd 0x10(%1,%4,4),%%xmm1 \n" |
8655 | 0 | "paddd 0x20(%1,%4,4),%%xmm2 \n" |
8656 | 0 | "paddd 0x30(%1,%4,4),%%xmm3 \n" |
8657 | 0 | "lea 0x40(%1),%1 \n" |
8658 | 0 | "cvtdq2ps %%xmm0,%%xmm0 \n" |
8659 | 0 | "cvtdq2ps %%xmm1,%%xmm1 \n" |
8660 | 0 | "mulps %%xmm4,%%xmm0 \n" |
8661 | 0 | "mulps %%xmm4,%%xmm1 \n" |
8662 | 0 | "cvtdq2ps %%xmm2,%%xmm2 \n" |
8663 | 0 | "cvtdq2ps %%xmm3,%%xmm3 \n" |
8664 | 0 | "mulps %%xmm4,%%xmm2 \n" |
8665 | 0 | "mulps %%xmm4,%%xmm3 \n" |
8666 | 0 | "cvtps2dq %%xmm0,%%xmm0 \n" |
8667 | 0 | "cvtps2dq %%xmm1,%%xmm1 \n" |
8668 | 0 | "cvtps2dq %%xmm2,%%xmm2 \n" |
8669 | 0 | "cvtps2dq %%xmm3,%%xmm3 \n" |
8670 | 0 | "packssdw %%xmm1,%%xmm0 \n" |
8671 | 0 | "packssdw %%xmm3,%%xmm2 \n" |
8672 | 0 | "packuswb %%xmm2,%%xmm0 \n" |
8673 | 0 | "movdqu %%xmm0,(%2) \n" |
8674 | 0 | "lea 0x10(%2),%2 \n" |
8675 | 0 | "sub $0x4,%3 \n" |
8676 | 0 | "jge 40b \n" |
8677 | |
|
8678 | 0 | "49: \n" |
8679 | 0 | "add $0x3,%3 \n" |
8680 | 0 | "jl 19f \n" |
8681 | | |
8682 | | // 1 pixel loop |
8683 | 0 | LABELALIGN |
8684 | 0 | "10: \n" |
8685 | 0 | "movdqu (%0),%%xmm0 \n" |
8686 | 0 | "psubd 0x00(%0,%4,4),%%xmm0 \n" |
8687 | 0 | "lea 0x10(%0),%0 \n" |
8688 | 0 | "psubd (%1),%%xmm0 \n" |
8689 | 0 | "paddd 0x00(%1,%4,4),%%xmm0 \n" |
8690 | 0 | "lea 0x10(%1),%1 \n" |
8691 | 0 | "cvtdq2ps %%xmm0,%%xmm0 \n" |
8692 | 0 | "mulps %%xmm4,%%xmm0 \n" |
8693 | 0 | "cvtps2dq %%xmm0,%%xmm0 \n" |
8694 | 0 | "packssdw %%xmm0,%%xmm0 \n" |
8695 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
8696 | 0 | "movd %%xmm0,(%2) \n" |
8697 | 0 | "lea 0x4(%2),%2 \n" |
8698 | 0 | "sub $0x1,%3 \n" |
8699 | 0 | "jge 10b \n" |
8700 | 0 | "19: \n" |
8701 | 0 | : "+r"(topleft), // %0 |
8702 | 0 | "+r"(botleft), // %1 |
8703 | 0 | "+r"(dst), // %2 |
8704 | 0 | "+rm"(count) // %3 |
8705 | 0 | : "r"((ptrdiff_t)(width)), // %4 |
8706 | 0 | "rm"(area) // %5 |
8707 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
8708 | 0 | } |
8709 | | #endif // HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 |
8710 | | |
8711 | | #ifdef HAS_ARGBAFFINEROW_SSE2 |
8712 | | // Copy ARGB pixels from source image with slope to a row of destination. |
8713 | | LIBYUV_API |
8714 | | void ARGBAffineRow_SSE2(const uint8_t* src_argb, |
8715 | | int src_argb_stride, |
8716 | | uint8_t* dst_argb, |
8717 | | const float* src_dudv, |
8718 | 0 | int width) { |
8719 | 0 | ptrdiff_t src_argb_stride_temp = src_argb_stride; |
8720 | 0 | intptr_t temp; |
8721 | 0 | asm volatile( |
8722 | 0 | "movq (%3),%%xmm2 \n" |
8723 | 0 | "movq 0x08(%3),%%xmm7 \n" |
8724 | 0 | "shl $0x10,%1 \n" |
8725 | 0 | "add $0x4,%1 \n" |
8726 | 0 | "movd %1,%%xmm5 \n" |
8727 | 0 | "sub $0x4,%4 \n" |
8728 | 0 | "jl 49f \n" |
8729 | |
|
8730 | 0 | "pshufd $0x44,%%xmm7,%%xmm7 \n" |
8731 | 0 | "pshufd $0x0,%%xmm5,%%xmm5 \n" |
8732 | 0 | "movdqa %%xmm2,%%xmm0 \n" |
8733 | 0 | "addps %%xmm7,%%xmm0 \n" |
8734 | 0 | "movlhps %%xmm0,%%xmm2 \n" |
8735 | 0 | "movdqa %%xmm7,%%xmm4 \n" |
8736 | 0 | "addps %%xmm4,%%xmm4 \n" |
8737 | 0 | "movdqa %%xmm2,%%xmm3 \n" |
8738 | 0 | "addps %%xmm4,%%xmm3 \n" |
8739 | 0 | "addps %%xmm4,%%xmm4 \n" |
8740 | | |
8741 | | // 4 pixel loop |
8742 | 0 | LABELALIGN |
8743 | 0 | "40: \n" |
8744 | 0 | "cvttps2dq %%xmm2,%%xmm0 \n" // x,y float->int first 2 |
8745 | 0 | "cvttps2dq %%xmm3,%%xmm1 \n" // x,y float->int next 2 |
8746 | 0 | "packssdw %%xmm1,%%xmm0 \n" // x, y as 8 shorts |
8747 | 0 | "pmaddwd %%xmm5,%%xmm0 \n" // off = x*4 + y*stride |
8748 | 0 | "movd %%xmm0,%k1 \n" |
8749 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
8750 | 0 | "movd %%xmm0,%k5 \n" |
8751 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
8752 | 0 | "movd 0x00(%0,%1,1),%%xmm1 \n" |
8753 | 0 | "movd 0x00(%0,%5,1),%%xmm6 \n" |
8754 | 0 | "punpckldq %%xmm6,%%xmm1 \n" |
8755 | 0 | "addps %%xmm4,%%xmm2 \n" |
8756 | 0 | "movq %%xmm1,(%2) \n" |
8757 | 0 | "movd %%xmm0,%k1 \n" |
8758 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
8759 | 0 | "movd %%xmm0,%k5 \n" |
8760 | 0 | "movd 0x00(%0,%1,1),%%xmm0 \n" |
8761 | 0 | "movd 0x00(%0,%5,1),%%xmm6 \n" |
8762 | 0 | "punpckldq %%xmm6,%%xmm0 \n" |
8763 | 0 | "addps %%xmm4,%%xmm3 \n" |
8764 | 0 | "movq %%xmm0,0x08(%2) \n" |
8765 | 0 | "lea 0x10(%2),%2 \n" |
8766 | 0 | "sub $0x4,%4 \n" |
8767 | 0 | "jge 40b \n" |
8768 | |
|
8769 | 0 | "49: \n" |
8770 | 0 | "add $0x3,%4 \n" |
8771 | 0 | "jl 19f \n" |
8772 | | |
8773 | | // 1 pixel loop |
8774 | 0 | LABELALIGN |
8775 | 0 | "10: \n" |
8776 | 0 | "cvttps2dq %%xmm2,%%xmm0 \n" |
8777 | 0 | "packssdw %%xmm0,%%xmm0 \n" |
8778 | 0 | "pmaddwd %%xmm5,%%xmm0 \n" |
8779 | 0 | "addps %%xmm7,%%xmm2 \n" |
8780 | 0 | "movd %%xmm0,%k1 \n" |
8781 | 0 | "movd 0x00(%0,%1,1),%%xmm0 \n" |
8782 | 0 | "movd %%xmm0,(%2) \n" |
8783 | 0 | "lea 0x04(%2),%2 \n" |
8784 | 0 | "sub $0x1,%4 \n" |
8785 | 0 | "jge 10b \n" |
8786 | 0 | "19: \n" |
8787 | 0 | : "+r"(src_argb), // %0 |
8788 | 0 | "+r"(src_argb_stride_temp), // %1 |
8789 | 0 | "+r"(dst_argb), // %2 |
8790 | 0 | "+r"(src_dudv), // %3 |
8791 | 0 | "+rm"(width), // %4 |
8792 | 0 | "=&r"(temp) // %5 |
8793 | 0 | : |
8794 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
8795 | 0 | "xmm7"); |
8796 | 0 | } |
8797 | | #endif // HAS_ARGBAFFINEROW_SSE2 |
8798 | | |
8799 | | #ifdef HAS_INTERPOLATEROW_AVX2 |
8800 | | // Bilinear filter 32x2 -> 32x1 |
8801 | | void InterpolateRow_AVX2(uint8_t* dst_ptr, |
8802 | | const uint8_t* src_ptr, |
8803 | | ptrdiff_t src_stride, |
8804 | | int width, |
8805 | 4.84M | int source_y_fraction) { |
8806 | 4.84M | asm volatile( |
8807 | 4.84M | "sub %1,%0 \n" |
8808 | 4.84M | "cmp $0x0,%3 \n" |
8809 | 4.84M | "je 100f \n" |
8810 | 4.84M | "cmp $0x80,%3 \n" |
8811 | 4.84M | "je 50f \n" |
8812 | | |
8813 | 4.84M | "vmovd %3,%%xmm0 \n" |
8814 | 4.84M | "neg %3 \n" |
8815 | 4.84M | "add $0x100,%3 \n" |
8816 | 4.84M | "vmovd %3,%%xmm5 \n" |
8817 | 4.84M | "vpunpcklbw %%xmm0,%%xmm5,%%xmm5 \n" |
8818 | 4.84M | "vpunpcklwd %%xmm5,%%xmm5,%%xmm5 \n" |
8819 | 4.84M | "vbroadcastss %%xmm5,%%ymm5 \n" |
8820 | 4.84M | "mov $0x80808080,%%eax \n" |
8821 | 4.84M | "vmovd %%eax,%%xmm4 \n" |
8822 | 4.84M | "vbroadcastss %%xmm4,%%ymm4 \n" |
8823 | | |
8824 | | // General purpose row blend. |
8825 | 4.84M | LABELALIGN |
8826 | 4.84M | "1: \n" |
8827 | 4.84M | "vmovdqu (%1),%%ymm0 \n" |
8828 | 4.84M | "vmovdqu 0x00(%1,%4,1),%%ymm2 \n" |
8829 | 4.84M | "vpunpckhbw %%ymm2,%%ymm0,%%ymm1 \n" |
8830 | 4.84M | "vpunpcklbw %%ymm2,%%ymm0,%%ymm0 \n" |
8831 | 4.84M | "vpsubb %%ymm4,%%ymm1,%%ymm1 \n" |
8832 | 4.84M | "vpsubb %%ymm4,%%ymm0,%%ymm0 \n" |
8833 | 4.84M | "vpmaddubsw %%ymm1,%%ymm5,%%ymm1 \n" |
8834 | 4.84M | "vpmaddubsw %%ymm0,%%ymm5,%%ymm0 \n" |
8835 | 4.84M | "vpaddw %%ymm4,%%ymm1,%%ymm1 \n" |
8836 | 4.84M | "vpaddw %%ymm4,%%ymm0,%%ymm0 \n" |
8837 | 4.84M | "vpsrlw $0x8,%%ymm1,%%ymm1 \n" |
8838 | 4.84M | "vpsrlw $0x8,%%ymm0,%%ymm0 \n" |
8839 | 4.84M | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" |
8840 | 4.84M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8841 | 4.84M | "lea 0x20(%1),%1 \n" |
8842 | 4.84M | "sub $0x20,%2 \n" |
8843 | 4.84M | "jg 1b \n" |
8844 | 4.84M | "jmp 99f \n" |
8845 | | |
8846 | | // Blend 50 / 50. |
8847 | 4.84M | LABELALIGN |
8848 | 4.84M | "50: \n" |
8849 | 4.84M | "vmovdqu (%1),%%ymm0 \n" |
8850 | 4.84M | "vpavgb 0x00(%1,%4,1),%%ymm0,%%ymm0 \n" |
8851 | 4.84M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8852 | 4.84M | "lea 0x20(%1),%1 \n" |
8853 | 4.84M | "sub $0x20,%2 \n" |
8854 | 4.84M | "jg 50b \n" |
8855 | 4.84M | "jmp 99f \n" |
8856 | | |
8857 | | // Blend 100 / 0 - Copy row unchanged. |
8858 | 4.84M | LABELALIGN |
8859 | 4.84M | "100: \n" |
8860 | 4.84M | "vmovdqu (%1),%%ymm0 \n" |
8861 | 4.84M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8862 | 4.84M | "lea 0x20(%1),%1 \n" |
8863 | 4.84M | "sub $0x20,%2 \n" |
8864 | 4.84M | "jg 100b \n" |
8865 | | |
8866 | 4.84M | "99: \n" |
8867 | 4.84M | "vzeroupper \n" |
8868 | 4.84M | : "+r"(dst_ptr), // %0 |
8869 | 4.84M | "+r"(src_ptr), // %1 |
8870 | 4.84M | "+r"(width), // %2 |
8871 | 4.84M | "+r"(source_y_fraction) // %3 |
8872 | 4.84M | : "r"(src_stride) // %4 |
8873 | 4.84M | : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm4", "xmm5"); |
8874 | 4.84M | } |
8875 | | #endif // HAS_INTERPOLATEROW_AVX2 |
8876 | | |
8877 | | #ifdef HAS_INTERPOLATEROW_16_AVX2 |
8878 | | // Bilinear filter 16x2 -> 16x1 |
8879 | | void InterpolateRow_16_AVX2(uint16_t* dst_ptr, |
8880 | | const uint16_t* src_ptr, |
8881 | | ptrdiff_t src_stride, |
8882 | | int width, |
8883 | 7.49M | int source_y_fraction) { |
8884 | 7.49M | asm volatile( |
8885 | 7.49M | "sub %1,%0 \n" |
8886 | 7.49M | "cmp $0x0,%3 \n" |
8887 | 7.49M | "je 100f \n" |
8888 | 7.49M | "cmp $0x80,%3 \n" |
8889 | 7.49M | "je 50f \n" |
8890 | | |
8891 | 7.49M | "vmovd %3,%%xmm0 \n" |
8892 | 7.49M | "neg %3 \n" |
8893 | 7.49M | "add $0x100,%3 \n" |
8894 | 7.49M | "vmovd %3,%%xmm5 \n" |
8895 | 7.49M | "vpunpcklwd %%xmm0,%%xmm5,%%xmm5 \n" |
8896 | 7.49M | "vpbroadcastd %%xmm5,%%ymm5 \n" |
8897 | 7.49M | "mov $0x80008000,%%eax \n" // 0x80008000 used to bias |
8898 | | // unsigned words to |
8899 | | // signed range for |
8900 | | // vpmaddwd. |
8901 | 7.49M | "vmovd %%eax,%%xmm4 \n" |
8902 | 7.49M | "vbroadcastss %%xmm4,%%ymm4 \n" |
8903 | 7.49M | "mov $8388736,%%eax \n" // 32768 * 256 + 128 |
8904 | | // rounding constant. |
8905 | 7.49M | "vmovd %%eax,%%xmm3 \n" |
8906 | 7.49M | "vbroadcastss %%xmm3,%%ymm3 \n" |
8907 | | |
8908 | 7.49M | LABELALIGN |
8909 | 7.49M | "1: \n" |
8910 | 7.49M | "vmovdqu (%1),%%ymm0 \n" |
8911 | 7.49M | "vmovdqu (%1,%4,2),%%ymm1 \n" |
8912 | 7.49M | "vpunpckhwd %%ymm1,%%ymm0,%%ymm2 \n" |
8913 | 7.49M | "vpunpcklwd %%ymm1,%%ymm0,%%ymm0 \n" |
8914 | 7.49M | "vpsubw %%ymm4,%%ymm2,%%ymm2 \n" |
8915 | 7.49M | "vpsubw %%ymm4,%%ymm0,%%ymm0 \n" |
8916 | 7.49M | "vpmaddwd %%ymm5,%%ymm2,%%ymm2 \n" |
8917 | 7.49M | "vpmaddwd %%ymm5,%%ymm0,%%ymm0 \n" |
8918 | 7.49M | "vpaddd %%ymm3,%%ymm2,%%ymm2 \n" |
8919 | 7.49M | "vpaddd %%ymm3,%%ymm0,%%ymm0 \n" |
8920 | 7.49M | "vpsrad $0x8,%%ymm2,%%ymm2 \n" |
8921 | 7.49M | "vpsrad $0x8,%%ymm0,%%ymm0 \n" |
8922 | 7.49M | "vpackusdw %%ymm2,%%ymm0,%%ymm0 \n" |
8923 | 7.49M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8924 | 7.49M | "lea 0x20(%1),%1 \n" |
8925 | 7.49M | "sub $0x10,%2 \n" |
8926 | 7.49M | "jg 1b \n" |
8927 | 7.49M | "jmp 99f \n" |
8928 | | |
8929 | 7.49M | "50: \n" LABELALIGN |
8930 | 7.49M | "2: \n" |
8931 | 7.49M | "vmovdqu (%1),%%ymm0 \n" |
8932 | 7.49M | "vpavgw (%1,%4,2),%%ymm0,%%ymm0 \n" |
8933 | 7.49M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8934 | 7.49M | "lea 0x20(%1),%1 \n" |
8935 | 7.49M | "sub $0x10,%2 \n" |
8936 | 7.49M | "jg 2b \n" |
8937 | 7.49M | "jmp 99f \n" |
8938 | | |
8939 | 7.49M | "100: \n" LABELALIGN |
8940 | 7.49M | "3: \n" |
8941 | 7.49M | "vmovdqu (%1),%%ymm0 \n" |
8942 | 7.49M | "vmovdqu %%ymm0,0x00(%1,%0,1) \n" |
8943 | 7.49M | "lea 0x20(%1),%1 \n" |
8944 | 7.49M | "sub $0x10,%2 \n" |
8945 | 7.49M | "jg 3b \n" |
8946 | | |
8947 | 7.49M | "99: \n" |
8948 | 7.49M | "vzeroupper \n" |
8949 | 7.49M | : "+r"(dst_ptr), // %0 |
8950 | 7.49M | "+r"(src_ptr), // %1 |
8951 | 7.49M | "+r"(width), // %2 |
8952 | 7.49M | "+r"(source_y_fraction) // %3 |
8953 | 7.49M | : "r"(src_stride) // %4 |
8954 | 7.49M | : "memory", "cc", "eax", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
8955 | 7.49M | } |
8956 | | #endif // HAS_INTERPOLATEROW_16_AVX2 |
8957 | | |
8958 | | #ifdef HAS_ARGBSHUFFLEROW_SSSE3 |
8959 | | // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. |
8960 | | void ARGBShuffleRow_SSSE3(const uint8_t* src_argb, |
8961 | | uint8_t* dst_argb, |
8962 | | const uint8_t* shuffler, |
8963 | 0 | int width) { |
8964 | 0 | asm volatile("movdqu (%3),%%xmm5 \n" |
8965 | |
|
8966 | 0 | LABELALIGN |
8967 | 0 | "1: \n" |
8968 | 0 | "movdqu (%0),%%xmm0 \n" |
8969 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
8970 | 0 | "lea 0x20(%0),%0 \n" |
8971 | 0 | "pshufb %%xmm5,%%xmm0 \n" |
8972 | 0 | "pshufb %%xmm5,%%xmm1 \n" |
8973 | 0 | "movdqu %%xmm0,(%1) \n" |
8974 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
8975 | 0 | "lea 0x20(%1),%1 \n" |
8976 | 0 | "sub $0x8,%2 \n" |
8977 | 0 | "jg 1b \n" |
8978 | 0 | : "+r"(src_argb), // %0 |
8979 | 0 | "+r"(dst_argb), // %1 |
8980 | 0 | "+r"(width) // %2 |
8981 | 0 | : "r"(shuffler) // %3 |
8982 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
8983 | 0 | } |
8984 | | #endif // HAS_ARGBSHUFFLEROW_SSSE3 |
8985 | | |
8986 | | #ifdef HAS_ARGBSHUFFLEROW_AVX2 |
8987 | | // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. |
8988 | | void ARGBShuffleRow_AVX2(const uint8_t* src_argb, |
8989 | | uint8_t* dst_argb, |
8990 | | const uint8_t* shuffler, |
8991 | 0 | int width) { |
8992 | 0 | asm volatile("vbroadcasti128 (%3),%%ymm5 \n" |
8993 | |
|
8994 | 0 | LABELALIGN |
8995 | 0 | "1: \n" |
8996 | 0 | "vmovdqu (%0),%%ymm0 \n" |
8997 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
8998 | 0 | "lea 0x40(%0),%0 \n" |
8999 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" |
9000 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" |
9001 | 0 | "vmovdqu %%ymm0,(%1) \n" |
9002 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
9003 | 0 | "lea 0x40(%1),%1 \n" |
9004 | 0 | "sub $0x10,%2 \n" |
9005 | 0 | "jg 1b \n" |
9006 | 0 | "vzeroupper \n" |
9007 | 0 | : "+r"(src_argb), // %0 |
9008 | 0 | "+r"(dst_argb), // %1 |
9009 | 0 | "+r"(width) // %2 |
9010 | 0 | : "r"(shuffler) // %3 |
9011 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
9012 | 0 | } |
9013 | | #endif // HAS_ARGBSHUFFLEROW_AVX2 |
9014 | | |
9015 | | #ifdef HAS_ARGBSHUFFLEROW_AVX512BW |
9016 | | // For BGRAToARGB, ABGRToARGB, RGBAToARGB, and ARGBToRGBA. |
9017 | | void ARGBShuffleRow_AVX512BW(const uint8_t* src_argb, |
9018 | | uint8_t* dst_argb, |
9019 | | const uint8_t* shuffler, |
9020 | 0 | int width) { |
9021 | 0 | asm volatile("vbroadcasti32x4 (%3),%%zmm5 \n" |
9022 | |
|
9023 | 0 | LABELALIGN |
9024 | 0 | "1: \n" |
9025 | 0 | "vmovdqu8 (%0),%%zmm0 \n" |
9026 | 0 | "vmovdqu8 0x40(%0),%%zmm1 \n" |
9027 | 0 | "lea 0x80(%0),%0 \n" |
9028 | 0 | "vpshufb %%zmm5,%%zmm0,%%zmm0 \n" |
9029 | 0 | "vpshufb %%zmm5,%%zmm1,%%zmm1 \n" |
9030 | 0 | "vmovdqu8 %%zmm0,(%1) \n" |
9031 | 0 | "vmovdqu8 %%zmm1,0x40(%1) \n" |
9032 | 0 | "lea 0x80(%1),%1 \n" |
9033 | 0 | "sub $0x20,%2 \n" |
9034 | 0 | "jg 1b \n" |
9035 | 0 | "vzeroupper \n" |
9036 | 0 | : "+r"(src_argb), // %0 |
9037 | 0 | "+r"(dst_argb), // %1 |
9038 | 0 | "+r"(width) // %2 |
9039 | 0 | : "r"(shuffler) // %3 |
9040 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
9041 | 0 | } |
9042 | | #endif // HAS_ARGBSHUFFLEROW_AVX512BW |
9043 | | |
9044 | | #ifdef HAS_I422TOYUY2ROW_SSE2 |
9045 | | void I422ToYUY2Row_SSE2(const uint8_t* src_y, |
9046 | | const uint8_t* src_u, |
9047 | | const uint8_t* src_v, |
9048 | | uint8_t* dst_yuy2, |
9049 | 0 | int width) { |
9050 | 0 | asm volatile("sub %1,%2 \n" |
9051 | |
|
9052 | 0 | LABELALIGN |
9053 | 0 | "1: \n" |
9054 | 0 | "movq (%1),%%xmm2 \n" |
9055 | 0 | "movq 0x00(%1,%2,1),%%xmm1 \n" |
9056 | 0 | "add $0x8,%1 \n" |
9057 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" |
9058 | 0 | "movdqu (%0),%%xmm0 \n" |
9059 | 0 | "add $0x10,%0 \n" |
9060 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
9061 | 0 | "punpcklbw %%xmm2,%%xmm0 \n" |
9062 | 0 | "punpckhbw %%xmm2,%%xmm1 \n" |
9063 | 0 | "movdqu %%xmm0,(%3) \n" |
9064 | 0 | "movdqu %%xmm1,0x10(%3) \n" |
9065 | 0 | "lea 0x20(%3),%3 \n" |
9066 | 0 | "sub $0x10,%4 \n" |
9067 | 0 | "jg 1b \n" |
9068 | 0 | : "+r"(src_y), // %0 |
9069 | 0 | "+r"(src_u), // %1 |
9070 | 0 | "+r"(src_v), // %2 |
9071 | 0 | "+r"(dst_yuy2), // %3 |
9072 | 0 | "+rm"(width) // %4 |
9073 | 0 | : |
9074 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
9075 | 0 | } |
9076 | | #endif // HAS_I422TOYUY2ROW_SSE2 |
9077 | | |
9078 | | #ifdef HAS_I422TOUYVYROW_SSE2 |
9079 | | void I422ToUYVYRow_SSE2(const uint8_t* src_y, |
9080 | | const uint8_t* src_u, |
9081 | | const uint8_t* src_v, |
9082 | | uint8_t* dst_uyvy, |
9083 | 0 | int width) { |
9084 | 0 | asm volatile("sub %1,%2 \n" |
9085 | |
|
9086 | 0 | LABELALIGN |
9087 | 0 | "1: \n" |
9088 | 0 | "movq (%1),%%xmm2 \n" |
9089 | 0 | "movq 0x00(%1,%2,1),%%xmm1 \n" |
9090 | 0 | "add $0x8,%1 \n" |
9091 | 0 | "punpcklbw %%xmm1,%%xmm2 \n" |
9092 | 0 | "movdqu (%0),%%xmm0 \n" |
9093 | 0 | "movdqa %%xmm2,%%xmm1 \n" |
9094 | 0 | "add $0x10,%0 \n" |
9095 | 0 | "punpcklbw %%xmm0,%%xmm1 \n" |
9096 | 0 | "punpckhbw %%xmm0,%%xmm2 \n" |
9097 | 0 | "movdqu %%xmm1,(%3) \n" |
9098 | 0 | "movdqu %%xmm2,0x10(%3) \n" |
9099 | 0 | "lea 0x20(%3),%3 \n" |
9100 | 0 | "sub $0x10,%4 \n" |
9101 | 0 | "jg 1b \n" |
9102 | 0 | : "+r"(src_y), // %0 |
9103 | 0 | "+r"(src_u), // %1 |
9104 | 0 | "+r"(src_v), // %2 |
9105 | 0 | "+r"(dst_uyvy), // %3 |
9106 | 0 | "+rm"(width) // %4 |
9107 | 0 | : |
9108 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
9109 | 0 | } |
9110 | | #endif // HAS_I422TOUYVYROW_SSE2 |
9111 | | |
9112 | | #ifdef HAS_I422TOYUY2ROW_AVX2 |
9113 | | void I422ToYUY2Row_AVX2(const uint8_t* src_y, |
9114 | | const uint8_t* src_u, |
9115 | | const uint8_t* src_v, |
9116 | | uint8_t* dst_yuy2, |
9117 | 0 | int width) { |
9118 | 0 | asm volatile("sub %1,%2 \n" |
9119 | |
|
9120 | 0 | LABELALIGN |
9121 | 0 | "1: \n" |
9122 | 0 | "vpmovzxbw (%1),%%ymm1 \n" |
9123 | 0 | "vpmovzxbw 0x00(%1,%2,1),%%ymm2 \n" |
9124 | 0 | "add $0x10,%1 \n" |
9125 | 0 | "vpsllw $0x8,%%ymm2,%%ymm2 \n" |
9126 | 0 | "vpor %%ymm1,%%ymm2,%%ymm2 \n" |
9127 | 0 | "vmovdqu (%0),%%ymm0 \n" |
9128 | 0 | "add $0x20,%0 \n" |
9129 | 0 | "vpunpcklbw %%ymm2,%%ymm0,%%ymm1 \n" |
9130 | 0 | "vpunpckhbw %%ymm2,%%ymm0,%%ymm2 \n" |
9131 | 0 | "vextractf128 $0x0,%%ymm1,(%3) \n" |
9132 | 0 | "vextractf128 $0x0,%%ymm2,0x10(%3) \n" |
9133 | 0 | "vextractf128 $0x1,%%ymm1,0x20(%3) \n" |
9134 | 0 | "vextractf128 $0x1,%%ymm2,0x30(%3) \n" |
9135 | 0 | "lea 0x40(%3),%3 \n" |
9136 | 0 | "sub $0x20,%4 \n" |
9137 | 0 | "jg 1b \n" |
9138 | 0 | "vzeroupper \n" |
9139 | 0 | : "+r"(src_y), // %0 |
9140 | 0 | "+r"(src_u), // %1 |
9141 | 0 | "+r"(src_v), // %2 |
9142 | 0 | "+r"(dst_yuy2), // %3 |
9143 | 0 | "+rm"(width) // %4 |
9144 | 0 | : |
9145 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
9146 | 0 | } |
9147 | | #endif // HAS_I422TOYUY2ROW_AVX2 |
9148 | | |
9149 | | #ifdef HAS_I422TOUYVYROW_AVX2 |
9150 | | void I422ToUYVYRow_AVX2(const uint8_t* src_y, |
9151 | | const uint8_t* src_u, |
9152 | | const uint8_t* src_v, |
9153 | | uint8_t* dst_uyvy, |
9154 | 0 | int width) { |
9155 | 0 | asm volatile("sub %1,%2 \n" |
9156 | |
|
9157 | 0 | LABELALIGN |
9158 | 0 | "1: \n" |
9159 | 0 | "vpmovzxbw (%1),%%ymm1 \n" |
9160 | 0 | "vpmovzxbw 0x00(%1,%2,1),%%ymm2 \n" |
9161 | 0 | "add $0x10,%1 \n" |
9162 | 0 | "vpsllw $0x8,%%ymm2,%%ymm2 \n" |
9163 | 0 | "vpor %%ymm1,%%ymm2,%%ymm2 \n" |
9164 | 0 | "vmovdqu (%0),%%ymm0 \n" |
9165 | 0 | "add $0x20,%0 \n" |
9166 | 0 | "vpunpcklbw %%ymm0,%%ymm2,%%ymm1 \n" |
9167 | 0 | "vpunpckhbw %%ymm0,%%ymm2,%%ymm2 \n" |
9168 | 0 | "vextractf128 $0x0,%%ymm1,(%3) \n" |
9169 | 0 | "vextractf128 $0x0,%%ymm2,0x10(%3) \n" |
9170 | 0 | "vextractf128 $0x1,%%ymm1,0x20(%3) \n" |
9171 | 0 | "vextractf128 $0x1,%%ymm2,0x30(%3) \n" |
9172 | 0 | "lea 0x40(%3),%3 \n" |
9173 | 0 | "sub $0x20,%4 \n" |
9174 | 0 | "jg 1b \n" |
9175 | 0 | "vzeroupper \n" |
9176 | 0 | : "+r"(src_y), // %0 |
9177 | 0 | "+r"(src_u), // %1 |
9178 | 0 | "+r"(src_v), // %2 |
9179 | 0 | "+r"(dst_uyvy), // %3 |
9180 | 0 | "+rm"(width) // %4 |
9181 | 0 | : |
9182 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2"); |
9183 | 0 | } |
9184 | | #endif // HAS_I422TOUYVYROW_AVX2 |
9185 | | |
9186 | | #ifdef HAS_ARGBPOLYNOMIALROW_SSE2 |
9187 | | void ARGBPolynomialRow_SSE2(const uint8_t* src_argb, |
9188 | | uint8_t* dst_argb, |
9189 | | const float* poly, |
9190 | 0 | int width) { |
9191 | 0 | asm volatile("pxor %%xmm3,%%xmm3 \n" |
9192 | | |
9193 | | // 2 pixel loop. |
9194 | 0 | LABELALIGN |
9195 | 0 | "1: \n" |
9196 | 0 | "movq (%0),%%xmm0 \n" |
9197 | 0 | "lea 0x8(%0),%0 \n" |
9198 | 0 | "punpcklbw %%xmm3,%%xmm0 \n" |
9199 | 0 | "movdqa %%xmm0,%%xmm4 \n" |
9200 | 0 | "punpcklwd %%xmm3,%%xmm0 \n" |
9201 | 0 | "punpckhwd %%xmm3,%%xmm4 \n" |
9202 | 0 | "cvtdq2ps %%xmm0,%%xmm0 \n" |
9203 | 0 | "cvtdq2ps %%xmm4,%%xmm4 \n" |
9204 | 0 | "movdqa %%xmm0,%%xmm1 \n" |
9205 | 0 | "movdqa %%xmm4,%%xmm5 \n" |
9206 | 0 | "mulps 0x10(%3),%%xmm0 \n" |
9207 | 0 | "mulps 0x10(%3),%%xmm4 \n" |
9208 | 0 | "addps (%3),%%xmm0 \n" |
9209 | 0 | "addps (%3),%%xmm4 \n" |
9210 | 0 | "movdqa %%xmm1,%%xmm2 \n" |
9211 | 0 | "movdqa %%xmm5,%%xmm6 \n" |
9212 | 0 | "mulps %%xmm1,%%xmm2 \n" |
9213 | 0 | "mulps %%xmm5,%%xmm6 \n" |
9214 | 0 | "mulps %%xmm2,%%xmm1 \n" |
9215 | 0 | "mulps %%xmm6,%%xmm5 \n" |
9216 | 0 | "mulps 0x20(%3),%%xmm2 \n" |
9217 | 0 | "mulps 0x20(%3),%%xmm6 \n" |
9218 | 0 | "mulps 0x30(%3),%%xmm1 \n" |
9219 | 0 | "mulps 0x30(%3),%%xmm5 \n" |
9220 | 0 | "addps %%xmm2,%%xmm0 \n" |
9221 | 0 | "addps %%xmm6,%%xmm4 \n" |
9222 | 0 | "addps %%xmm1,%%xmm0 \n" |
9223 | 0 | "addps %%xmm5,%%xmm4 \n" |
9224 | 0 | "cvttps2dq %%xmm0,%%xmm0 \n" |
9225 | 0 | "cvttps2dq %%xmm4,%%xmm4 \n" |
9226 | 0 | "packuswb %%xmm4,%%xmm0 \n" |
9227 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
9228 | 0 | "movq %%xmm0,(%1) \n" |
9229 | 0 | "lea 0x8(%1),%1 \n" |
9230 | 0 | "sub $0x2,%2 \n" |
9231 | 0 | "jg 1b \n" |
9232 | 0 | : "+r"(src_argb), // %0 |
9233 | 0 | "+r"(dst_argb), // %1 |
9234 | 0 | "+r"(width) // %2 |
9235 | 0 | : "r"(poly) // %3 |
9236 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", |
9237 | 0 | "xmm6"); |
9238 | 0 | } |
9239 | | #endif // HAS_ARGBPOLYNOMIALROW_SSE2 |
9240 | | |
9241 | | #ifdef HAS_ARGBPOLYNOMIALROW_AVX2 |
9242 | | void ARGBPolynomialRow_AVX2(const uint8_t* src_argb, |
9243 | | uint8_t* dst_argb, |
9244 | | const float* poly, |
9245 | 0 | int width) { |
9246 | 0 | asm volatile( |
9247 | 0 | "vbroadcasti128 (%3),%%ymm4 \n" |
9248 | 0 | "vbroadcasti128 0x10(%3),%%ymm5 \n" |
9249 | 0 | "vbroadcasti128 0x20(%3),%%ymm6 \n" |
9250 | 0 | "vbroadcasti128 0x30(%3),%%ymm7 \n" |
9251 | | |
9252 | | // 2 pixel loop. |
9253 | 0 | LABELALIGN |
9254 | 0 | "1: \n" |
9255 | 0 | "vpmovzxbd (%0),%%ymm0 \n" // 2 ARGB pixels |
9256 | 0 | "lea 0x8(%0),%0 \n" |
9257 | 0 | "vcvtdq2ps %%ymm0,%%ymm0 \n" // X 8 floats |
9258 | 0 | "vmulps %%ymm0,%%ymm0,%%ymm2 \n" // X * X |
9259 | 0 | "vmulps %%ymm7,%%ymm0,%%ymm3 \n" // C3 * X |
9260 | 0 | "vfmadd132ps %%ymm5,%%ymm4,%%ymm0 \n" // result = C0 + C1 * X |
9261 | 0 | "vfmadd231ps %%ymm6,%%ymm2,%%ymm0 \n" // result += C2 * X * X |
9262 | 0 | "vfmadd231ps %%ymm3,%%ymm2,%%ymm0 \n" // result += C3 * X * X * |
9263 | | // X |
9264 | 0 | "vcvttps2dq %%ymm0,%%ymm0 \n" |
9265 | 0 | "vpackusdw %%ymm0,%%ymm0,%%ymm0 \n" |
9266 | 0 | "vpermq $0xd8,%%ymm0,%%ymm0 \n" |
9267 | 0 | "vpackuswb %%xmm0,%%xmm0,%%xmm0 \n" |
9268 | 0 | "vmovq %%xmm0,(%1) \n" |
9269 | 0 | "lea 0x8(%1),%1 \n" |
9270 | 0 | "sub $0x2,%2 \n" |
9271 | 0 | "jg 1b \n" |
9272 | 0 | "vzeroupper \n" |
9273 | 0 | : "+r"(src_argb), // %0 |
9274 | 0 | "+r"(dst_argb), // %1 |
9275 | 0 | "+r"(width) // %2 |
9276 | 0 | : "r"(poly) // %3 |
9277 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", |
9278 | 0 | "xmm7"); |
9279 | 0 | } |
9280 | | #endif // HAS_ARGBPOLYNOMIALROW_AVX2 |
9281 | | |
9282 | | #ifdef HAS_HALFFLOATROW_AVX2 |
9283 | | static float kScaleBias = 1.9259299444e-34f; |
9284 | | void HalfFloatRow_AVX2(const uint16_t* src, |
9285 | | uint16_t* dst, |
9286 | | float scale, |
9287 | 382 | int width) { |
9288 | 382 | scale *= kScaleBias; |
9289 | 382 | asm volatile( |
9290 | 382 | "vbroadcastss %3, %%ymm4 \n" |
9291 | 382 | "vpxor %%ymm5,%%ymm5,%%ymm5 \n" |
9292 | 382 | "sub %0,%1 \n" |
9293 | | |
9294 | | // 16 pixel loop. |
9295 | 382 | LABELALIGN |
9296 | 382 | "1: \n" |
9297 | 382 | "vmovdqu (%0),%%ymm2 \n" // 16 shorts |
9298 | 382 | "add $0x20,%0 \n" |
9299 | 382 | "vpunpckhwd %%ymm5,%%ymm2,%%ymm3 \n" // mutates |
9300 | 382 | "vpunpcklwd %%ymm5,%%ymm2,%%ymm2 \n" |
9301 | 382 | "vcvtdq2ps %%ymm3,%%ymm3 \n" |
9302 | 382 | "vcvtdq2ps %%ymm2,%%ymm2 \n" |
9303 | 382 | "vmulps %%ymm3,%%ymm4,%%ymm3 \n" |
9304 | 382 | "vmulps %%ymm2,%%ymm4,%%ymm2 \n" |
9305 | 382 | "vpsrld $0xd,%%ymm3,%%ymm3 \n" |
9306 | 382 | "vpsrld $0xd,%%ymm2,%%ymm2 \n" |
9307 | 382 | "vpackssdw %%ymm3, %%ymm2, %%ymm2 \n" // unmutates |
9308 | 382 | "vmovdqu %%ymm2,-0x20(%0,%1,1) \n" |
9309 | 382 | "sub $0x10,%2 \n" |
9310 | 382 | "jg 1b \n" |
9311 | | |
9312 | 382 | "vzeroupper \n" |
9313 | 382 | : "+r"(src), // %0 |
9314 | 382 | "+r"(dst), // %1 |
9315 | 382 | "+r"(width) // %2 |
9316 | 382 | #if defined(__x86_64__) |
9317 | 382 | : "x"(scale) // %3 |
9318 | | #else |
9319 | | : "m"(scale) // %3 |
9320 | | #endif |
9321 | 382 | : "memory", "cc", "xmm2", "xmm3", "xmm4", "xmm5"); |
9322 | 382 | } |
9323 | | #endif // HAS_HALFFLOATROW_AVX2 |
9324 | | |
9325 | | #ifdef HAS_HALFFLOATROW_F16C |
9326 | | void HalfFloatRow_F16C(const uint16_t* src, |
9327 | | uint16_t* dst, |
9328 | | float scale, |
9329 | | int width) { |
9330 | | asm volatile( |
9331 | | "vbroadcastss %3, %%ymm4 \n" |
9332 | | "sub %0,%1 \n" |
9333 | | |
9334 | | // 16 pixel loop. |
9335 | | LABELALIGN |
9336 | | "1: \n" |
9337 | | "vpmovzxwd (%0),%%ymm2 \n" // 16 shorts -> 16 ints |
9338 | | "vpmovzxwd 0x10(%0),%%ymm3 \n" |
9339 | | "vcvtdq2ps %%ymm2,%%ymm2 \n" |
9340 | | "vcvtdq2ps %%ymm3,%%ymm3 \n" |
9341 | | "vmulps %%ymm2,%%ymm4,%%ymm2 \n" |
9342 | | "vmulps %%ymm3,%%ymm4,%%ymm3 \n" |
9343 | | "vcvtps2ph $3, %%ymm2, %%xmm2 \n" |
9344 | | "vcvtps2ph $3, %%ymm3, %%xmm3 \n" |
9345 | | "vmovdqu %%xmm2,0x00(%0,%1,1) \n" |
9346 | | "vmovdqu %%xmm3,0x10(%0,%1,1) \n" |
9347 | | "add $0x20,%0 \n" |
9348 | | "sub $0x10,%2 \n" |
9349 | | "jg 1b \n" |
9350 | | "vzeroupper \n" |
9351 | | : "+r"(src), // %0 |
9352 | | "+r"(dst), // %1 |
9353 | | "+r"(width) // %2 |
9354 | | #if defined(__x86_64__) |
9355 | | : "x"(scale) // %3 |
9356 | | #else |
9357 | | : "m"(scale) // %3 |
9358 | | #endif |
9359 | | : "memory", "cc", "xmm2", "xmm3", "xmm4"); |
9360 | | } |
9361 | | #endif // HAS_HALFFLOATROW_F16C |
9362 | | |
9363 | | #ifdef HAS_HALFFLOATROW_F16C |
9364 | | void HalfFloat1Row_F16C(const uint16_t* src, uint16_t* dst, float, int width) { |
9365 | | asm volatile( |
9366 | | "sub %0,%1 \n" |
9367 | | // 16 pixel loop. |
9368 | | LABELALIGN |
9369 | | "1: \n" |
9370 | | "vpmovzxwd (%0),%%ymm2 \n" // 16 shorts -> 16 ints |
9371 | | "vpmovzxwd 0x10(%0),%%ymm3 \n" |
9372 | | "vcvtdq2ps %%ymm2,%%ymm2 \n" |
9373 | | "vcvtdq2ps %%ymm3,%%ymm3 \n" |
9374 | | "vcvtps2ph $3, %%ymm2, %%xmm2 \n" |
9375 | | "vcvtps2ph $3, %%ymm3, %%xmm3 \n" |
9376 | | "vmovdqu %%xmm2,0x00(%0,%1,1) \n" |
9377 | | "vmovdqu %%xmm3,0x10(%0,%1,1) \n" |
9378 | | "add $0x20,%0 \n" |
9379 | | "sub $0x10,%2 \n" |
9380 | | "jg 1b \n" |
9381 | | "vzeroupper \n" |
9382 | | : "+r"(src), // %0 |
9383 | | "+r"(dst), // %1 |
9384 | | "+r"(width) // %2 |
9385 | | : |
9386 | | : "memory", "cc", "xmm2", "xmm3"); |
9387 | | } |
9388 | | #endif // HAS_HALFFLOATROW_F16C |
9389 | | |
9390 | | #ifdef HAS_ARGBCOLORTABLEROW_X86 |
9391 | | // Tranform ARGB pixels with color table. |
9392 | | void ARGBColorTableRow_X86(uint8_t* dst_argb, |
9393 | | const uint8_t* table_argb, |
9394 | 0 | int width) { |
9395 | 0 | uintptr_t pixel_temp; |
9396 | 0 | asm volatile( |
9397 | | // 1 pixel loop. |
9398 | 0 | LABELALIGN |
9399 | 0 | "1: \n" |
9400 | 0 | "movzb (%0),%1 \n" |
9401 | 0 | "lea 0x4(%0),%0 \n" |
9402 | 0 | "movzb 0x00(%3,%1,4),%1 \n" |
9403 | 0 | "mov %b1,-0x4(%0) \n" |
9404 | 0 | "movzb -0x3(%0),%1 \n" |
9405 | 0 | "movzb 0x01(%3,%1,4),%1 \n" |
9406 | 0 | "mov %b1,-0x3(%0) \n" |
9407 | 0 | "movzb -0x2(%0),%1 \n" |
9408 | 0 | "movzb 0x02(%3,%1,4),%1 \n" |
9409 | 0 | "mov %b1,-0x2(%0) \n" |
9410 | 0 | "movzb -0x1(%0),%1 \n" |
9411 | 0 | "movzb 0x03(%3,%1,4),%1 \n" |
9412 | 0 | "mov %b1,-0x1(%0) \n" |
9413 | 0 | "dec %2 \n" |
9414 | 0 | "jg 1b \n" |
9415 | 0 | : "+r"(dst_argb), // %0 |
9416 | 0 | "=&d"(pixel_temp), // %1 |
9417 | 0 | "+r"(width) // %2 |
9418 | 0 | : "r"(table_argb) // %3 |
9419 | 0 | : "memory", "cc"); |
9420 | 0 | } |
9421 | | #endif // HAS_ARGBCOLORTABLEROW_X86 |
9422 | | |
9423 | | #ifdef HAS_RGBCOLORTABLEROW_X86 |
9424 | | // Tranform RGB pixels with color table. |
9425 | | void RGBColorTableRow_X86(uint8_t* dst_argb, |
9426 | | const uint8_t* table_argb, |
9427 | 0 | int width) { |
9428 | 0 | uintptr_t pixel_temp; |
9429 | 0 | asm volatile( |
9430 | | // 1 pixel loop. |
9431 | 0 | LABELALIGN |
9432 | 0 | "1: \n" |
9433 | 0 | "movzb (%0),%1 \n" |
9434 | 0 | "lea 0x4(%0),%0 \n" |
9435 | 0 | "movzb 0x00(%3,%1,4),%1 \n" |
9436 | 0 | "mov %b1,-0x4(%0) \n" |
9437 | 0 | "movzb -0x3(%0),%1 \n" |
9438 | 0 | "movzb 0x01(%3,%1,4),%1 \n" |
9439 | 0 | "mov %b1,-0x3(%0) \n" |
9440 | 0 | "movzb -0x2(%0),%1 \n" |
9441 | 0 | "movzb 0x02(%3,%1,4),%1 \n" |
9442 | 0 | "mov %b1,-0x2(%0) \n" |
9443 | 0 | "dec %2 \n" |
9444 | 0 | "jg 1b \n" |
9445 | 0 | : "+r"(dst_argb), // %0 |
9446 | 0 | "=&d"(pixel_temp), // %1 |
9447 | 0 | "+r"(width) // %2 |
9448 | 0 | : "r"(table_argb) // %3 |
9449 | 0 | : "memory", "cc"); |
9450 | 0 | } |
9451 | | #endif // HAS_RGBCOLORTABLEROW_X86 |
9452 | | |
9453 | | #ifdef HAS_ARGBLUMACOLORTABLEROW_SSSE3 |
9454 | | // Tranform RGB pixels with luma table. |
9455 | | void ARGBLumaColorTableRow_SSSE3(const uint8_t* src_argb, |
9456 | | uint8_t* dst_argb, |
9457 | | int width, |
9458 | | const uint8_t* luma, |
9459 | 0 | uint32_t lumacoeff) { |
9460 | 0 | uintptr_t pixel_temp; |
9461 | 0 | uintptr_t table_temp; |
9462 | 0 | asm volatile( |
9463 | 0 | "movd %6,%%xmm3 \n" |
9464 | 0 | "pshufd $0x0,%%xmm3,%%xmm3 \n" |
9465 | 0 | "pcmpeqb %%xmm4,%%xmm4 \n" |
9466 | 0 | "psllw $0x8,%%xmm4 \n" |
9467 | 0 | "pxor %%xmm5,%%xmm5 \n" |
9468 | | |
9469 | | // 4 pixel loop. |
9470 | 0 | LABELALIGN |
9471 | 0 | "1: \n" |
9472 | 0 | "movdqu (%2),%%xmm0 \n" |
9473 | 0 | "pmaddubsw %%xmm3,%%xmm0 \n" |
9474 | 0 | "phaddw %%xmm0,%%xmm0 \n" |
9475 | 0 | "pand %%xmm4,%%xmm0 \n" |
9476 | 0 | "punpcklwd %%xmm5,%%xmm0 \n" |
9477 | 0 | "movd %%xmm0,%k1 \n" // 32 bit offset |
9478 | 0 | "add %5,%1 \n" |
9479 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
9480 | |
|
9481 | 0 | "movzb (%2),%0 \n" |
9482 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9483 | 0 | "mov %b0,(%3) \n" |
9484 | 0 | "movzb 0x1(%2),%0 \n" |
9485 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9486 | 0 | "mov %b0,0x1(%3) \n" |
9487 | 0 | "movzb 0x2(%2),%0 \n" |
9488 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9489 | 0 | "mov %b0,0x2(%3) \n" |
9490 | 0 | "movzb 0x3(%2),%0 \n" |
9491 | 0 | "mov %b0,0x3(%3) \n" |
9492 | |
|
9493 | 0 | "movd %%xmm0,%k1 \n" // 32 bit offset |
9494 | 0 | "add %5,%1 \n" |
9495 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
9496 | |
|
9497 | 0 | "movzb 0x4(%2),%0 \n" |
9498 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9499 | 0 | "mov %b0,0x4(%3) \n" |
9500 | 0 | "movzb 0x5(%2),%0 \n" |
9501 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9502 | 0 | "mov %b0,0x5(%3) \n" |
9503 | 0 | "movzb 0x6(%2),%0 \n" |
9504 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9505 | 0 | "mov %b0,0x6(%3) \n" |
9506 | 0 | "movzb 0x7(%2),%0 \n" |
9507 | 0 | "mov %b0,0x7(%3) \n" |
9508 | |
|
9509 | 0 | "movd %%xmm0,%k1 \n" // 32 bit offset |
9510 | 0 | "add %5,%1 \n" |
9511 | 0 | "pshufd $0x39,%%xmm0,%%xmm0 \n" |
9512 | |
|
9513 | 0 | "movzb 0x8(%2),%0 \n" |
9514 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9515 | 0 | "mov %b0,0x8(%3) \n" |
9516 | 0 | "movzb 0x9(%2),%0 \n" |
9517 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9518 | 0 | "mov %b0,0x9(%3) \n" |
9519 | 0 | "movzb 0xa(%2),%0 \n" |
9520 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9521 | 0 | "mov %b0,0xa(%3) \n" |
9522 | 0 | "movzb 0xb(%2),%0 \n" |
9523 | 0 | "mov %b0,0xb(%3) \n" |
9524 | |
|
9525 | 0 | "movd %%xmm0,%k1 \n" // 32 bit offset |
9526 | 0 | "add %5,%1 \n" |
9527 | |
|
9528 | 0 | "movzb 0xc(%2),%0 \n" |
9529 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9530 | 0 | "mov %b0,0xc(%3) \n" |
9531 | 0 | "movzb 0xd(%2),%0 \n" |
9532 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9533 | 0 | "mov %b0,0xd(%3) \n" |
9534 | 0 | "movzb 0xe(%2),%0 \n" |
9535 | 0 | "movzb 0x00(%1,%0,1),%0 \n" |
9536 | 0 | "mov %b0,0xe(%3) \n" |
9537 | 0 | "movzb 0xf(%2),%0 \n" |
9538 | 0 | "mov %b0,0xf(%3) \n" |
9539 | 0 | "lea 0x10(%2),%2 \n" |
9540 | 0 | "lea 0x10(%3),%3 \n" |
9541 | 0 | "sub $0x4,%4 \n" |
9542 | 0 | "jg 1b \n" |
9543 | 0 | : "=&d"(pixel_temp), // %0 |
9544 | 0 | "=&a"(table_temp), // %1 |
9545 | 0 | "+r"(src_argb), // %2 |
9546 | 0 | "+r"(dst_argb), // %3 |
9547 | 0 | "+rm"(width) // %4 |
9548 | 0 | : "r"(luma), // %5 |
9549 | 0 | "rm"(lumacoeff) // %6 |
9550 | 0 | : "memory", "cc", "xmm0", "xmm3", "xmm4", "xmm5"); |
9551 | 0 | } |
9552 | | #endif // HAS_ARGBLUMACOLORTABLEROW_SSSE3 |
9553 | | |
9554 | | static const uvec8 kYUV24Shuffle[3] = { |
9555 | | {8, 9, 0, 8, 9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12}, |
9556 | | {9, 1, 10, 11, 2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15}, |
9557 | | {2, 10, 11, 3, 12, 13, 4, 12, 13, 5, 14, 15, 6, 14, 15, 7}}; |
9558 | | |
9559 | | // Convert biplanar NV21 to packed YUV24 |
9560 | | // NV21 has VU in memory for chroma. |
9561 | | // YUV24 is VUY in memory |
9562 | | void NV21ToYUV24Row_SSSE3(const uint8_t* src_y, |
9563 | | const uint8_t* src_vu, |
9564 | | uint8_t* dst_yuv24, |
9565 | 0 | int width) { |
9566 | 0 | asm volatile( |
9567 | 0 | "sub %0,%1 \n" |
9568 | 0 | "movdqa (%4),%%xmm4 \n" // 3 shuffler constants |
9569 | 0 | "movdqa 16(%4),%%xmm5 \n" |
9570 | 0 | "movdqa 32(%4),%%xmm6 \n" |
9571 | 0 | "1: \n" |
9572 | 0 | "movdqu (%0),%%xmm2 \n" // load 16 Y values |
9573 | 0 | "movdqu (%0,%1),%%xmm3 \n" // load 8 VU values |
9574 | 0 | "lea 16(%0),%0 \n" |
9575 | 0 | "movdqa %%xmm2,%%xmm0 \n" |
9576 | 0 | "movdqa %%xmm2,%%xmm1 \n" |
9577 | 0 | "shufps $0x44,%%xmm3,%%xmm0 \n" // Y 0..7, UV 0..3 |
9578 | 0 | "shufps $0x99,%%xmm3,%%xmm1 \n" // Y 4..11, UV 2..5 |
9579 | 0 | "shufps $0xee,%%xmm3,%%xmm2 \n" // Y 8..15, UV 4..7 |
9580 | 0 | "pshufb %%xmm4, %%xmm0 \n" // weave into YUV24 |
9581 | 0 | "pshufb %%xmm5, %%xmm1 \n" |
9582 | 0 | "pshufb %%xmm6, %%xmm2 \n" |
9583 | 0 | "movdqu %%xmm0,(%2) \n" |
9584 | 0 | "movdqu %%xmm1,16(%2) \n" |
9585 | 0 | "movdqu %%xmm2,32(%2) \n" |
9586 | 0 | "lea 48(%2),%2 \n" |
9587 | 0 | "sub $16,%3 \n" // 16 pixels per loop |
9588 | 0 | "jg 1b \n" |
9589 | 0 | : "+r"(src_y), // %0 |
9590 | 0 | "+r"(src_vu), // %1 |
9591 | 0 | "+r"(dst_yuv24), // %2 |
9592 | 0 | "+r"(width) // %3 |
9593 | 0 | : "r"(&kYUV24Shuffle[0]) // %4 |
9594 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
9595 | 0 | } |
9596 | | |
9597 | | // Convert biplanar NV21 to packed YUV24 |
9598 | | // NV21 has VU in memory for chroma. |
9599 | | // YUV24 is VUY in memory |
9600 | | void NV21ToYUV24Row_AVX2(const uint8_t* src_y, |
9601 | | const uint8_t* src_vu, |
9602 | | uint8_t* dst_yuv24, |
9603 | 0 | int width) { |
9604 | 0 | asm volatile( |
9605 | 0 | "sub %0,%1 \n" |
9606 | 0 | "vbroadcasti128 (%4),%%ymm4 \n" // 3 shuffler constants |
9607 | 0 | "vbroadcasti128 16(%4),%%ymm5 \n" |
9608 | 0 | "vbroadcasti128 32(%4),%%ymm6 \n" |
9609 | |
|
9610 | 0 | "1: \n" |
9611 | 0 | "vmovdqu (%0),%%ymm2 \n" // load 32 Y values |
9612 | 0 | "vmovdqu (%0,%1),%%ymm3 \n" // load 16 VU values |
9613 | 0 | "lea 32(%0),%0 \n" |
9614 | 0 | "vshufps $0x44,%%ymm3,%%ymm2,%%ymm0 \n" // Y 0..7, UV 0..3 |
9615 | 0 | "vshufps $0x99,%%ymm3,%%ymm2,%%ymm1 \n" // Y 4..11, UV 2..5 |
9616 | 0 | "vshufps $0xee,%%ymm3,%%ymm2,%%ymm2 \n" // Y 8..15, UV 4..7 |
9617 | 0 | "vpshufb %%ymm4,%%ymm0,%%ymm0 \n" // weave into YUV24 |
9618 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" |
9619 | 0 | "vpshufb %%ymm6,%%ymm2,%%ymm2 \n" |
9620 | 0 | "vperm2i128 $0x20,%%ymm1,%%ymm0,%%ymm3 \n" |
9621 | 0 | "vperm2i128 $0x30,%%ymm0,%%ymm2,%%ymm0 \n" |
9622 | 0 | "vperm2i128 $0x31,%%ymm2,%%ymm1,%%ymm1 \n" |
9623 | 0 | "vmovdqu %%ymm3,(%2) \n" |
9624 | 0 | "vmovdqu %%ymm0,32(%2) \n" |
9625 | 0 | "vmovdqu %%ymm1,64(%2) \n" |
9626 | 0 | "lea 96(%2),%2 \n" |
9627 | 0 | "sub $32,%3 \n" // 32 pixels per loop |
9628 | 0 | "jg 1b \n" |
9629 | 0 | "vzeroupper \n" |
9630 | 0 | : "+r"(src_y), // %0 |
9631 | 0 | "+r"(src_vu), // %1 |
9632 | 0 | "+r"(dst_yuv24), // %2 |
9633 | 0 | "+r"(width) // %3 |
9634 | 0 | : "r"(&kYUV24Shuffle[0]) // %4 |
9635 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
9636 | 0 | } |
9637 | | |
9638 | | #ifdef HAS_NV21ToYUV24ROW_AVX512 |
9639 | | // The following VMBI VEX256 code tests okay with the intelsde emulator. |
9640 | | static const lvec8 kYUV24Perm[3] = { |
9641 | | {32, 33, 0, 32, 33, 1, 34, 35, 2, 34, 35, 3, 36, 37, 4, 36, |
9642 | | 37, 5, 38, 39, 6, 38, 39, 7, 40, 41, 8, 40, 41, 9, 42, 43}, |
9643 | | {10, 42, 43, 11, 44, 45, 12, 44, 45, 13, 46, 47, 14, 46, 47, 15, |
9644 | | 48, 49, 16, 48, 49, 17, 50, 51, 18, 50, 51, 19, 52, 53, 20, 52}, |
9645 | | {53, 21, 54, 55, 22, 54, 55, 23, 56, 57, 24, 56, 57, 25, 58, 59, |
9646 | | 26, 58, 59, 27, 60, 61, 28, 60, 61, 29, 62, 63, 30, 62, 63, 31}}; |
9647 | | |
9648 | | void NV21ToYUV24Row_AVX512(const uint8_t* src_y, |
9649 | | const uint8_t* src_vu, |
9650 | | uint8_t* dst_yuv24, |
9651 | | int width) { |
9652 | | asm volatile( |
9653 | | "sub %0,%1 \n" |
9654 | | "vmovdqa (%4),%%ymm4 \n" // 3 shuffler constants |
9655 | | "vmovdqa 32(%4),%%ymm5 \n" |
9656 | | "vmovdqa 64(%4),%%ymm6 \n" LABELALIGN |
9657 | | "1: \n" |
9658 | | "vmovdqu (%0),%%ymm2 \n" // load 32 Y values |
9659 | | "vmovdqu (%0,%1),%%ymm3 \n" // load 16 VU values |
9660 | | "lea 32(%0),%0 \n" |
9661 | | "vmovdqa %%ymm2, %%ymm0 \n" |
9662 | | "vmovdqa %%ymm2, %%ymm1 \n" |
9663 | | "vpermt2b %%ymm3,%%ymm4,%%ymm0 \n" |
9664 | | "vpermt2b %%ymm3,%%ymm5,%%ymm1 \n" |
9665 | | "vpermt2b %%ymm3,%%ymm6,%%ymm2 \n" |
9666 | | "vmovdqu %%ymm0,(%2) \n" |
9667 | | "vmovdqu %%ymm1,32(%2) \n" |
9668 | | "vmovdqu %%ymm2,64(%2) \n" |
9669 | | "lea 96(%2),%2 \n" |
9670 | | "sub $32,%3 \n" |
9671 | | "jg 1b \n" |
9672 | | "vzeroupper \n" |
9673 | | : "+r"(src_y), // %0 |
9674 | | "+r"(src_vu), // %1 |
9675 | | "+r"(dst_yuv24), // %2 |
9676 | | "+r"(width) // %3 |
9677 | | : "r"(&kYUV24Perm[0]) // %4 |
9678 | | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6"); |
9679 | | } |
9680 | | |
9681 | | #endif // HAS_NV21ToYUV24ROW_AVX512 |
9682 | | |
9683 | | #ifdef HAS_SWAPUVROW_SSSE3 |
9684 | | |
9685 | | // Shuffle table for reversing the bytes. |
9686 | | static const uvec8 kShuffleUVToVU = {1u, 0u, 3u, 2u, 5u, 4u, 7u, 6u, |
9687 | | 9u, 8u, 11u, 10u, 13u, 12u, 15u, 14u}; |
9688 | | |
9689 | | // Convert UV plane of NV12 to VU of NV21. |
9690 | 0 | void SwapUVRow_SSSE3(const uint8_t* src_uv, uint8_t* dst_vu, int width) { |
9691 | 0 | asm volatile("movdqu %3,%%xmm5 \n" |
9692 | |
|
9693 | 0 | LABELALIGN |
9694 | 0 | "1: \n" |
9695 | 0 | "movdqu (%0),%%xmm0 \n" |
9696 | 0 | "movdqu 0x10(%0),%%xmm1 \n" |
9697 | 0 | "lea 0x20(%0),%0 \n" |
9698 | 0 | "pshufb %%xmm5,%%xmm0 \n" |
9699 | 0 | "pshufb %%xmm5,%%xmm1 \n" |
9700 | 0 | "movdqu %%xmm0,(%1) \n" |
9701 | 0 | "movdqu %%xmm1,0x10(%1) \n" |
9702 | 0 | "lea 0x20(%1),%1 \n" |
9703 | 0 | "sub $0x10,%2 \n" |
9704 | 0 | "jg 1b \n" |
9705 | 0 | : "+r"(src_uv), // %0 |
9706 | 0 | "+r"(dst_vu), // %1 |
9707 | 0 | "+r"(width) // %2 |
9708 | 0 | : "m"(kShuffleUVToVU) // %3 |
9709 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
9710 | 0 | } |
9711 | | #endif // HAS_SWAPUVROW_SSSE3 |
9712 | | |
9713 | | #ifdef HAS_SWAPUVROW_AVX2 |
9714 | 0 | void SwapUVRow_AVX2(const uint8_t* src_uv, uint8_t* dst_vu, int width) { |
9715 | 0 | asm volatile("vbroadcasti128 %3,%%ymm5 \n" |
9716 | |
|
9717 | 0 | LABELALIGN |
9718 | 0 | "1: \n" |
9719 | 0 | "vmovdqu (%0),%%ymm0 \n" |
9720 | 0 | "vmovdqu 0x20(%0),%%ymm1 \n" |
9721 | 0 | "lea 0x40(%0),%0 \n" |
9722 | 0 | "vpshufb %%ymm5,%%ymm0,%%ymm0 \n" |
9723 | 0 | "vpshufb %%ymm5,%%ymm1,%%ymm1 \n" |
9724 | 0 | "vmovdqu %%ymm0,(%1) \n" |
9725 | 0 | "vmovdqu %%ymm1,0x20(%1) \n" |
9726 | 0 | "lea 0x40(%1),%1 \n" |
9727 | 0 | "sub $0x20,%2 \n" |
9728 | 0 | "jg 1b \n" |
9729 | 0 | "vzeroupper \n" |
9730 | 0 | : "+r"(src_uv), // %0 |
9731 | 0 | "+r"(dst_vu), // %1 |
9732 | 0 | "+r"(width) // %2 |
9733 | 0 | : "m"(kShuffleUVToVU) // %3 |
9734 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm5"); |
9735 | 0 | } |
9736 | | #endif // HAS_SWAPUVROW_AVX2 |
9737 | | |
9738 | | void HalfMergeUVRow_SSSE3(const uint8_t* src_u, |
9739 | | int src_stride_u, |
9740 | | const uint8_t* src_v, |
9741 | | int src_stride_v, |
9742 | | uint8_t* dst_uv, |
9743 | 0 | int width) { |
9744 | 0 | asm volatile( |
9745 | 0 | "pcmpeqb %%xmm4,%%xmm4 \n" // 0x0101 |
9746 | 0 | "pabsb %%xmm4,%%xmm4 \n" |
9747 | 0 | "pxor %%xmm5,%%xmm5 \n" |
9748 | |
|
9749 | 0 | LABELALIGN |
9750 | 0 | "1: \n" |
9751 | 0 | "movdqu (%0),%%xmm0 \n" // load 16 U values |
9752 | 0 | "movdqu (%1),%%xmm1 \n" // load 16 V values |
9753 | 0 | "movdqu 0(%0,%4,1),%%xmm2 \n" // 16 from next row |
9754 | 0 | "movdqu 0(%1,%5,1),%%xmm3 \n" |
9755 | 0 | "lea 0x10(%0),%0 \n" |
9756 | 0 | "pmaddubsw %%xmm4,%%xmm0 \n" // half size |
9757 | 0 | "pmaddubsw %%xmm4,%%xmm1 \n" |
9758 | 0 | "pmaddubsw %%xmm4,%%xmm2 \n" |
9759 | 0 | "pmaddubsw %%xmm4,%%xmm3 \n" |
9760 | 0 | "lea 0x10(%1),%1 \n" |
9761 | 0 | "paddw %%xmm2,%%xmm0 \n" |
9762 | 0 | "paddw %%xmm3,%%xmm1 \n" |
9763 | 0 | "psrlw $0x1,%%xmm0 \n" |
9764 | 0 | "psrlw $0x1,%%xmm1 \n" |
9765 | 0 | "pavgw %%xmm5,%%xmm0 \n" |
9766 | 0 | "pavgw %%xmm5,%%xmm1 \n" |
9767 | 0 | "packuswb %%xmm0,%%xmm0 \n" |
9768 | 0 | "packuswb %%xmm1,%%xmm1 \n" |
9769 | 0 | "punpcklbw %%xmm1,%%xmm0 \n" |
9770 | 0 | "movdqu %%xmm0,(%2) \n" // store 8 UV pixels |
9771 | 0 | "lea 0x10(%2),%2 \n" |
9772 | 0 | "sub $0x10,%3 \n" // 16 src pixels per loop |
9773 | 0 | "jg 1b \n" |
9774 | 0 | : "+r"(src_u), // %0 |
9775 | 0 | "+r"(src_v), // %1 |
9776 | 0 | "+r"(dst_uv), // %2 |
9777 | 0 | "+r"(width) // %3 |
9778 | 0 | : "r"((ptrdiff_t)(src_stride_u)), // %4 |
9779 | 0 | "r"((ptrdiff_t)(src_stride_v)) // %5 |
9780 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
9781 | 0 | } |
9782 | | |
9783 | | void HalfMergeUVRow_AVX2(const uint8_t* src_u, |
9784 | | int src_stride_u, |
9785 | | const uint8_t* src_v, |
9786 | | int src_stride_v, |
9787 | | uint8_t* dst_uv, |
9788 | 0 | int width) { |
9789 | 0 | asm volatile( |
9790 | 0 | "vpcmpeqb %%ymm4,%%ymm4,%%ymm4 \n" |
9791 | 0 | "vpabsb %%ymm4,%%ymm4 \n" |
9792 | 0 | "vpxor %%ymm5,%%ymm5,%%ymm5 \n" |
9793 | |
|
9794 | 0 | LABELALIGN |
9795 | 0 | "1: \n" |
9796 | 0 | "vmovdqu (%0),%%ymm0 \n" // load 32 U values |
9797 | 0 | "vmovdqu (%1),%%ymm1 \n" // load 32 V values |
9798 | 0 | "vmovdqu 0(%0,%4,1),%%ymm2 \n" // 32 from next row |
9799 | 0 | "vmovdqu 0(%1,%5,1),%%ymm3 \n" |
9800 | 0 | "lea 0x20(%0),%0 \n" |
9801 | 0 | "vpmaddubsw %%ymm4,%%ymm0,%%ymm0 \n" // half size |
9802 | 0 | "vpmaddubsw %%ymm4,%%ymm1,%%ymm1 \n" |
9803 | 0 | "vpmaddubsw %%ymm4,%%ymm2,%%ymm2 \n" |
9804 | 0 | "vpmaddubsw %%ymm4,%%ymm3,%%ymm3 \n" |
9805 | 0 | "lea 0x20(%1),%1 \n" |
9806 | 0 | "vpaddw %%ymm2,%%ymm0,%%ymm0 \n" |
9807 | 0 | "vpaddw %%ymm3,%%ymm1,%%ymm1 \n" |
9808 | 0 | "vpsrlw $0x1,%%ymm0,%%ymm0 \n" |
9809 | 0 | "vpsrlw $0x1,%%ymm1,%%ymm1 \n" |
9810 | 0 | "vpavgw %%ymm5,%%ymm0,%%ymm0 \n" |
9811 | 0 | "vpavgw %%ymm5,%%ymm1,%%ymm1 \n" |
9812 | 0 | "vpackuswb %%ymm0,%%ymm0,%%ymm0 \n" |
9813 | 0 | "vpackuswb %%ymm1,%%ymm1,%%ymm1 \n" |
9814 | 0 | "vpunpcklbw %%ymm1,%%ymm0,%%ymm0 \n" |
9815 | 0 | "vmovdqu %%ymm0,(%2) \n" // store 16 UV pixels |
9816 | 0 | "lea 0x20(%2),%2 \n" |
9817 | 0 | "sub $0x20,%3 \n" // 32 src pixels per loop |
9818 | 0 | "jg 1b \n" |
9819 | 0 | "vzeroupper \n" |
9820 | 0 | : "+r"(src_u), // %0 |
9821 | 0 | "+r"(src_v), // %1 |
9822 | 0 | "+r"(dst_uv), // %2 |
9823 | 0 | "+r"(width) // %3 |
9824 | 0 | : "r"((ptrdiff_t)(src_stride_u)), // %4 |
9825 | 0 | "r"((ptrdiff_t)(src_stride_v)) // %5 |
9826 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5"); |
9827 | 0 | } |
9828 | | |
9829 | 0 | void ClampFloatToZero_SSE2(const float* src_x, float* dst_y, int width) { |
9830 | 0 | asm volatile( |
9831 | 0 | "pxor %%xmm1,%%xmm1 \n" |
9832 | |
|
9833 | 0 | LABELALIGN |
9834 | 0 | "1: \n" |
9835 | 0 | "movd (%0),%%xmm0 \n" // load float |
9836 | 0 | "maxss %%xmm1, %%xmm0 \n" // clamp to zero |
9837 | 0 | "add 4, %0 \n" |
9838 | 0 | "movd %%xmm0, (%1) \n" // store float |
9839 | 0 | "add 4, %1 \n" |
9840 | 0 | "sub $0x4,%2 \n" // 1 float per loop |
9841 | 0 | "jg 1b \n" |
9842 | 0 | : "+r"(src_x), // %0 |
9843 | 0 | "+r"(dst_y), // %1 |
9844 | 0 | "+r"(width) // %2 |
9845 | 0 | : |
9846 | 0 | : "memory", "cc", "xmm0", "xmm1"); |
9847 | 0 | } |
9848 | | |
9849 | | #ifdef HAS_CONVERT16TO8ROW_AVX2 |
9850 | | void Convert8To8Row_AVX2(const uint8_t* src_y, |
9851 | | uint8_t* dst_y, |
9852 | | int scale, |
9853 | | int bias, |
9854 | 0 | int width) { |
9855 | 0 | asm volatile( |
9856 | 0 | "sub %0,%1 \n" |
9857 | 0 | "vmovd %3,%%xmm2 \n" |
9858 | 0 | "vmovd %4,%%xmm3 \n" |
9859 | 0 | "vpbroadcastw %%xmm2,%%ymm2 \n" |
9860 | 0 | "vpbroadcastb %%xmm3,%%ymm3 \n" |
9861 | 0 | "vpxor %%ymm4,%%ymm4,%%ymm4 \n" |
9862 | 0 | "vpsllw $8,%%ymm2,%%ymm2 \n" |
9863 | | |
9864 | | // 32 pixels per loop. |
9865 | 0 | LABELALIGN |
9866 | 0 | "1: \n" |
9867 | 0 | "vmovdqu (%0),%%ymm0 \n" |
9868 | 0 | "vpunpckhbw %%ymm4,%%ymm0,%%ymm1 \n" // mutates |
9869 | 0 | "vpunpcklbw %%ymm4,%%ymm0,%%ymm0 \n" |
9870 | 0 | "vpmulhuw %%ymm2,%%ymm0,%%ymm0 \n" |
9871 | 0 | "vpmulhuw %%ymm2,%%ymm1,%%ymm1 \n" |
9872 | 0 | "vpackuswb %%ymm1,%%ymm0,%%ymm0 \n" // unmutates |
9873 | 0 | "vpaddb %%ymm3,%%ymm0,%%ymm0 \n" |
9874 | 0 | "vmovdqu %%ymm0,(%0,%1) \n" |
9875 | 0 | "add $0x20,%0 \n" |
9876 | 0 | "sub $0x20,%2 \n" |
9877 | 0 | "jg 1b \n" |
9878 | 0 | "vzeroupper \n" |
9879 | 0 | : "+r"(src_y), // %0 |
9880 | 0 | "+r"(dst_y), // %1 |
9881 | 0 | "+r"(width) // %2 |
9882 | 0 | : "r"(scale), // %3 |
9883 | 0 | "r"(bias) // %4 |
9884 | 0 | : "memory", "cc", "xmm0", "xmm1", "xmm2", "xmm3", "xmm4"); |
9885 | 0 | } |
9886 | | #endif // HAS_CONVERT16TO8ROW_AVX2 |
9887 | | |
9888 | | #endif // defined(__x86_64__) || defined(__i386__) |
9889 | | |
9890 | | #ifdef __cplusplus |
9891 | | } // extern "C" |
9892 | | } // namespace libyuv |
9893 | | #endif |