/src/FreeRDP/libfreerdp/codec/sse/nsc_sse2.c
Line | Count | Source (jump to first uncovered line) |
1 | | /** |
2 | | * FreeRDP: A Remote Desktop Protocol Implementation |
3 | | * NSCodec Library - SSE2 Optimizations |
4 | | * |
5 | | * Copyright 2012 Vic Lee |
6 | | * |
7 | | * Licensed under the Apache License, Version 2.0 (the "License"); |
8 | | * you may not use this file except in compliance with the License. |
9 | | * You may obtain a copy of the License at |
10 | | * |
11 | | * http://www.apache.org/licenses/LICENSE-2.0 |
12 | | * |
13 | | * Unless required by applicable law or agreed to in writing, software |
14 | | * distributed under the License is distributed on an "AS IS" BASIS, |
15 | | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
16 | | * See the License for the specific language governing permissions and |
17 | | * limitations under the License. |
18 | | */ |
19 | | |
20 | | #include <winpr/assert.h> |
21 | | #include <winpr/cast.h> |
22 | | #include <winpr/platform.h> |
23 | | #include <freerdp/config.h> |
24 | | |
25 | | #include "../nsc_types.h" |
26 | | #include "nsc_sse2.h" |
27 | | |
28 | | #include "../../core/simd.h" |
29 | | #include "../../primitives/sse/prim_avxsse.h" |
30 | | |
31 | | #if defined(SSE_AVX_INTRINSICS_ENABLED) |
32 | | #include <stdio.h> |
33 | | #include <stdlib.h> |
34 | | #include <string.h> |
35 | | |
36 | | #include <xmmintrin.h> |
37 | | #include <emmintrin.h> |
38 | | |
39 | | #include <freerdp/codec/color.h> |
40 | | #include <winpr/crt.h> |
41 | | #include <winpr/sysinfo.h> |
42 | | |
43 | | static inline size_t nsc_encode_next_rgba(UINT32 format, const BYTE* src, const BYTE* palette, |
44 | | __m128i* r_val, __m128i* g_val, __m128i* b_val, |
45 | | __m128i* a_val) |
46 | 0 | { |
47 | 0 | switch (format) |
48 | 0 | { |
49 | 0 | case PIXEL_FORMAT_BGRX32: |
50 | 0 | *b_val = _mm_set_epi16(*(src + 28), *(src + 24), *(src + 20), *(src + 16), *(src + 12), |
51 | 0 | *(src + 8), *(src + 4), *src); |
52 | 0 | *g_val = _mm_set_epi16(*(src + 29), *(src + 25), *(src + 21), *(src + 17), *(src + 13), |
53 | 0 | *(src + 9), *(src + 5), *(src + 1)); |
54 | 0 | *r_val = _mm_set_epi16(*(src + 30), *(src + 26), *(src + 22), *(src + 18), *(src + 14), |
55 | 0 | *(src + 10), *(src + 6), *(src + 2)); |
56 | 0 | *a_val = _mm_set1_epi16(0xFF); |
57 | 0 | return 32; |
58 | | |
59 | 0 | case PIXEL_FORMAT_BGRA32: |
60 | 0 | *b_val = _mm_set_epi16(*(src + 28), *(src + 24), *(src + 20), *(src + 16), *(src + 12), |
61 | 0 | *(src + 8), *(src + 4), *src); |
62 | 0 | *g_val = _mm_set_epi16(*(src + 29), *(src + 25), *(src + 21), *(src + 17), *(src + 13), |
63 | 0 | *(src + 9), *(src + 5), *(src + 1)); |
64 | 0 | *r_val = _mm_set_epi16(*(src + 30), *(src + 26), *(src + 22), *(src + 18), *(src + 14), |
65 | 0 | *(src + 10), *(src + 6), *(src + 2)); |
66 | 0 | *a_val = _mm_set_epi16(*(src + 31), *(src + 27), *(src + 23), *(src + 19), *(src + 15), |
67 | 0 | *(src + 11), *(src + 7), *(src + 3)); |
68 | 0 | return 32; |
69 | | |
70 | 0 | case PIXEL_FORMAT_RGBX32: |
71 | 0 | *r_val = _mm_set_epi16(*(src + 28), *(src + 24), *(src + 20), *(src + 16), *(src + 12), |
72 | 0 | *(src + 8), *(src + 4), *src); |
73 | 0 | *g_val = _mm_set_epi16(*(src + 29), *(src + 25), *(src + 21), *(src + 17), *(src + 13), |
74 | 0 | *(src + 9), *(src + 5), *(src + 1)); |
75 | 0 | *b_val = _mm_set_epi16(*(src + 30), *(src + 26), *(src + 22), *(src + 18), *(src + 14), |
76 | 0 | *(src + 10), *(src + 6), *(src + 2)); |
77 | 0 | *a_val = _mm_set1_epi16(0xFF); |
78 | 0 | return 32; |
79 | | |
80 | 0 | case PIXEL_FORMAT_RGBA32: |
81 | 0 | *r_val = _mm_set_epi16(*(src + 28), *(src + 24), *(src + 20), *(src + 16), *(src + 12), |
82 | 0 | *(src + 8), *(src + 4), *src); |
83 | 0 | *g_val = _mm_set_epi16(*(src + 29), *(src + 25), *(src + 21), *(src + 17), *(src + 13), |
84 | 0 | *(src + 9), *(src + 5), *(src + 1)); |
85 | 0 | *b_val = _mm_set_epi16(*(src + 30), *(src + 26), *(src + 22), *(src + 18), *(src + 14), |
86 | 0 | *(src + 10), *(src + 6), *(src + 2)); |
87 | 0 | *a_val = _mm_set_epi16(*(src + 31), *(src + 27), *(src + 23), *(src + 19), *(src + 15), |
88 | 0 | *(src + 11), *(src + 7), *(src + 3)); |
89 | 0 | return 32; |
90 | | |
91 | 0 | case PIXEL_FORMAT_BGR24: |
92 | 0 | *b_val = _mm_set_epi16(*(src + 21), *(src + 18), *(src + 15), *(src + 12), *(src + 9), |
93 | 0 | *(src + 6), *(src + 3), *src); |
94 | 0 | *g_val = _mm_set_epi16(*(src + 22), *(src + 19), *(src + 16), *(src + 13), *(src + 10), |
95 | 0 | *(src + 7), *(src + 4), *(src + 1)); |
96 | 0 | *r_val = _mm_set_epi16(*(src + 23), *(src + 20), *(src + 17), *(src + 14), *(src + 11), |
97 | 0 | *(src + 8), *(src + 5), *(src + 2)); |
98 | 0 | *a_val = _mm_set1_epi16(0xFF); |
99 | 0 | return 24; |
100 | | |
101 | 0 | case PIXEL_FORMAT_RGB24: |
102 | 0 | *r_val = _mm_set_epi16(*(src + 21), *(src + 18), *(src + 15), *(src + 12), *(src + 9), |
103 | 0 | *(src + 6), *(src + 3), *src); |
104 | 0 | *g_val = _mm_set_epi16(*(src + 22), *(src + 19), *(src + 16), *(src + 13), *(src + 10), |
105 | 0 | *(src + 7), *(src + 4), *(src + 1)); |
106 | 0 | *b_val = _mm_set_epi16(*(src + 23), *(src + 20), *(src + 17), *(src + 14), *(src + 11), |
107 | 0 | *(src + 8), *(src + 5), *(src + 2)); |
108 | 0 | *a_val = _mm_set1_epi16(0xFF); |
109 | 0 | return 24; |
110 | | |
111 | 0 | case PIXEL_FORMAT_BGR16: |
112 | 0 | *b_val = _mm_set_epi16( |
113 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 15)) & 0xF8) | ((*(src + 15)) >> 5)), |
114 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 13)) & 0xF8) | ((*(src + 13)) >> 5)), |
115 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 11)) & 0xF8) | ((*(src + 11)) >> 5)), |
116 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 9)) & 0xF8) | ((*(src + 9)) >> 5)), |
117 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 7)) & 0xF8) | ((*(src + 7)) >> 5)), |
118 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 5)) & 0xF8) | ((*(src + 5)) >> 5)), |
119 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 3)) & 0xF8) | ((*(src + 3)) >> 5)), |
120 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((*(src + 1)) & 0xF8) | ((*(src + 1)) >> 5))); |
121 | 0 | *g_val = |
122 | 0 | _mm_set_epi16(WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 15)) & 0x07) << 5) | |
123 | 0 | (((*(src + 14)) & 0xE0) >> 3)), |
124 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 13)) & 0x07) << 5) | |
125 | 0 | (((*(src + 12)) & 0xE0) >> 3)), |
126 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 11)) & 0x07) << 5) | |
127 | 0 | (((*(src + 10)) & 0xE0) >> 3)), |
128 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 9)) & 0x07) << 5) | |
129 | 0 | (((*(src + 8)) & 0xE0) >> 3)), |
130 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 7)) & 0x07) << 5) | |
131 | 0 | (((*(src + 6)) & 0xE0) >> 3)), |
132 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 5)) & 0x07) << 5) | |
133 | 0 | (((*(src + 4)) & 0xE0) >> 3)), |
134 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 3)) & 0x07) << 5) | |
135 | 0 | (((*(src + 2)) & 0xE0) >> 3)), |
136 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 1)) & 0x07) << 5) | |
137 | 0 | (((*src) & 0xE0) >> 3))); |
138 | 0 | *r_val = _mm_set_epi16( |
139 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 14)) & 0x1F) << 3) | |
140 | 0 | (((*(src + 14)) >> 2) & 0x07)), |
141 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 12)) & 0x1F) << 3) | |
142 | 0 | (((*(src + 12)) >> 2) & 0x07)), |
143 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 10)) & 0x1F) << 3) | |
144 | 0 | (((*(src + 10)) >> 2) & 0x07)), |
145 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 8)) & 0x1F) << 3) | |
146 | 0 | (((*(src + 8)) >> 2) & 0x07)), |
147 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 6)) & 0x1F) << 3) | |
148 | 0 | (((*(src + 6)) >> 2) & 0x07)), |
149 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 4)) & 0x1F) << 3) | |
150 | 0 | (((*(src + 4)) >> 2) & 0x07)), |
151 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 2)) & 0x1F) << 3) | |
152 | 0 | (((*(src + 2)) >> 2) & 0x07)), |
153 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*src) & 0x1F) << 3) | (((*src) >> 2) & 0x07))); |
154 | 0 | *a_val = _mm_set1_epi16(0xFF); |
155 | 0 | return 16; |
156 | | |
157 | 0 | case PIXEL_FORMAT_RGB16: |
158 | 0 | *r_val = |
159 | 0 | _mm_set_epi16(WINPR_ASSERTING_INT_CAST(INT16, ((src[15] & 0xF8) | (src[15] >> 5))), |
160 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[13] & 0xF8) | (src[13] >> 5))), |
161 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[11] & 0xF8) | (src[11] >> 5))), |
162 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[9] & 0xF8) | (src[9] >> 5))), |
163 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[7] & 0xF8) | (src[7] >> 5))), |
164 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[5] & 0xF8) | (src[5] >> 5))), |
165 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[3] & 0xF8) | (src[3] >> 5))), |
166 | 0 | WINPR_ASSERTING_INT_CAST(INT16, ((src[1] & 0xF8) | (src[1] >> 5)))); |
167 | 0 | *g_val = |
168 | 0 | _mm_set_epi16(WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 15)) & 0x07) << 5) | |
169 | 0 | (((*(src + 14)) & 0xE0) >> 3)), |
170 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 13)) & 0x07) << 5) | |
171 | 0 | (((*(src + 12)) & 0xE0) >> 3)), |
172 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 11)) & 0x07) << 5) | |
173 | 0 | (((*(src + 10)) & 0xE0) >> 3)), |
174 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 9)) & 0x07) << 5) | |
175 | 0 | (((*(src + 8)) & 0xE0) >> 3)), |
176 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 7)) & 0x07) << 5) | |
177 | 0 | (((*(src + 6)) & 0xE0) >> 3)), |
178 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 5)) & 0x07) << 5) | |
179 | 0 | (((*(src + 4)) & 0xE0) >> 3)), |
180 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 3)) & 0x07) << 5) | |
181 | 0 | (((*(src + 2)) & 0xE0) >> 3)), |
182 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 1)) & 0x07) << 5) | |
183 | 0 | (((*src) & 0xE0) >> 3))); |
184 | 0 | *b_val = _mm_set_epi16( |
185 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 14)) & 0x1F) << 3) | |
186 | 0 | (((*(src + 14)) >> 2) & 0x07)), |
187 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 12)) & 0x1F) << 3) | |
188 | 0 | (((*(src + 12)) >> 2) & 0x07)), |
189 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 10)) & 0x1F) << 3) | |
190 | 0 | (((*(src + 10)) >> 2) & 0x07)), |
191 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 8)) & 0x1F) << 3) | |
192 | 0 | (((*(src + 8)) >> 2) & 0x07)), |
193 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 6)) & 0x1F) << 3) | |
194 | 0 | (((*(src + 6)) >> 2) & 0x07)), |
195 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 4)) & 0x1F) << 3) | |
196 | 0 | (((*(src + 4)) >> 2) & 0x07)), |
197 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*(src + 2)) & 0x1F) << 3) | |
198 | 0 | (((*(src + 2)) >> 2) & 0x07)), |
199 | 0 | WINPR_ASSERTING_INT_CAST(INT16, (((*src) & 0x1F) << 3) | (((*src) >> 2) & 0x07))); |
200 | 0 | *a_val = _mm_set1_epi16(0xFF); |
201 | 0 | return 16; |
202 | | |
203 | 0 | case PIXEL_FORMAT_A4: |
204 | 0 | { |
205 | 0 | BYTE idx[8] = { 0 }; |
206 | |
|
207 | 0 | for (int shift = 7; shift >= 0; shift--) |
208 | 0 | { |
209 | 0 | idx[shift] = ((*src) >> shift) & 1; |
210 | 0 | idx[shift] |= (((*(src + 1)) >> shift) & 1) << 1; |
211 | 0 | idx[shift] |= (((*(src + 2)) >> shift) & 1) << 2; |
212 | 0 | idx[shift] |= (((*(src + 3)) >> shift) & 1) << 3; |
213 | 0 | idx[shift] *= 3; |
214 | 0 | } |
215 | |
|
216 | 0 | *r_val = |
217 | 0 | _mm_set_epi16(palette[idx[0]], palette[idx[1]], palette[idx[2]], palette[idx[3]], |
218 | 0 | palette[idx[4]], palette[idx[5]], palette[idx[6]], palette[idx[7]]); |
219 | 0 | *g_val = _mm_set_epi16(palette[idx[0] + 1], palette[idx[1] + 1], palette[idx[2] + 1], |
220 | 0 | palette[idx[3] + 1], palette[idx[4] + 1], palette[idx[5] + 1], |
221 | 0 | palette[idx[6] + 1], palette[idx[7] + 1]); |
222 | 0 | *b_val = _mm_set_epi16(palette[idx[0] + 2], palette[idx[1] + 2], palette[idx[2] + 2], |
223 | 0 | palette[idx[3] + 2], palette[idx[4] + 2], palette[idx[5] + 2], |
224 | 0 | palette[idx[6] + 2], palette[idx[7] + 2]); |
225 | 0 | *a_val = _mm_set1_epi16(0xFF); |
226 | 0 | return 4; |
227 | 0 | } |
228 | | |
229 | 0 | case PIXEL_FORMAT_RGB8: |
230 | 0 | { |
231 | 0 | *r_val = _mm_set_epi16(palette[(*(src + 7ULL)) * 3ULL], palette[(*(src + 6ULL)) * 3ULL], |
232 | 0 | palette[(*(src + 5ULL)) * 3ULL], palette[(*(src + 4ULL)) * 3ULL], |
233 | 0 | palette[(*(src + 3ULL)) * 3ULL], palette[(*(src + 2ULL)) * 3ULL], |
234 | 0 | palette[(*(src + 1ULL)) * 3ULL], palette[(*src) * 3ULL]); |
235 | 0 | *g_val = _mm_set_epi16( |
236 | 0 | palette[(*(src + 7ULL)) * 3ULL + 1ULL], palette[(*(src + 6ULL)) * 3ULL + 1ULL], |
237 | 0 | palette[(*(src + 5ULL)) * 3ULL + 1ULL], palette[(*(src + 4ULL)) * 3ULL + 1ULL], |
238 | 0 | palette[(*(src + 3ULL)) * 3ULL + 1ULL], palette[(*(src + 2ULL)) * 3ULL + 1ULL], |
239 | 0 | palette[(*(src + 1ULL)) * 3ULL + 1ULL], palette[(*src) * 3ULL + 1ULL]); |
240 | 0 | *b_val = _mm_set_epi16( |
241 | 0 | palette[(*(src + 7ULL)) * 3ULL + 2ULL], palette[(*(src + 6ULL)) * 3ULL + 2ULL], |
242 | 0 | palette[(*(src + 5ULL)) * 3ULL + 2ULL], palette[(*(src + 4ULL)) * 3ULL + 2ULL], |
243 | 0 | palette[(*(src + 3ULL)) * 3ULL + 2ULL], palette[(*(src + 2ULL)) * 3ULL + 2ULL], |
244 | 0 | palette[(*(src + 1ULL)) * 3ULL + 2ULL], palette[(*src) * 3ULL + 2ULL]); |
245 | 0 | *a_val = _mm_set1_epi16(0xFF); |
246 | 0 | return 8; |
247 | 0 | } |
248 | | |
249 | 0 | default: |
250 | 0 | return 0; |
251 | 0 | } |
252 | 0 | } |
253 | | |
254 | | static BOOL nsc_encode_argb_to_aycocg_sse2(NSC_CONTEXT* context, const BYTE* data, UINT32 scanline) |
255 | 0 | { |
256 | 0 | size_t y = 0; |
257 | |
|
258 | 0 | if (!context || !data || (scanline == 0)) |
259 | 0 | return FALSE; |
260 | | |
261 | 0 | const UINT16 tempWidth = ROUND_UP_TO(context->width, 8); |
262 | 0 | const UINT16 rw = (context->ChromaSubsamplingLevel > 0 ? tempWidth : context->width); |
263 | |
|
264 | 0 | const BYTE ccl = WINPR_ASSERTING_INT_CAST(BYTE, context->ColorLossLevel); |
265 | | |
266 | 0 | for (; y < context->height; y++) |
267 | 0 | { |
268 | 0 | const BYTE* src = data + (context->height - 1 - y) * scanline; |
269 | 0 | BYTE* yplane = context->priv->PlaneBuffers[0] + y * rw; |
270 | 0 | BYTE* coplane = context->priv->PlaneBuffers[1] + y * rw; |
271 | 0 | BYTE* cgplane = context->priv->PlaneBuffers[2] + y * rw; |
272 | 0 | BYTE* aplane = context->priv->PlaneBuffers[3] + y * context->width; |
273 | |
|
274 | 0 | for (UINT16 x = 0; x < context->width; x += 8) |
275 | 0 | { |
276 | 0 | __m128i r_val = { 0 }; |
277 | 0 | __m128i g_val = { 0 }; |
278 | 0 | __m128i b_val = { 0 }; |
279 | 0 | __m128i a_val = { 0 }; |
280 | |
|
281 | 0 | const size_t rc = nsc_encode_next_rgba(context->format, src, context->palette, &r_val, |
282 | 0 | &g_val, &b_val, &a_val); |
283 | 0 | src += rc; |
284 | |
|
285 | 0 | __m128i y_val = _mm_srai_epi16(r_val, 2); |
286 | 0 | y_val = _mm_add_epi16(y_val, _mm_srai_epi16(g_val, 1)); |
287 | 0 | y_val = _mm_add_epi16(y_val, _mm_srai_epi16(b_val, 2)); |
288 | 0 | __m128i co_val = _mm_sub_epi16(r_val, b_val); |
289 | 0 | co_val = _mm_srai_epi16(co_val, ccl); |
290 | 0 | __m128i cg_val = _mm_sub_epi16(g_val, _mm_srai_epi16(r_val, 1)); |
291 | 0 | cg_val = _mm_sub_epi16(cg_val, _mm_srai_epi16(b_val, 1)); |
292 | 0 | cg_val = _mm_srai_epi16(cg_val, ccl); |
293 | 0 | y_val = _mm_packus_epi16(y_val, y_val); |
294 | 0 | STORE_SI128(yplane, y_val); |
295 | 0 | co_val = _mm_packs_epi16(co_val, co_val); |
296 | 0 | STORE_SI128(coplane, co_val); |
297 | 0 | cg_val = _mm_packs_epi16(cg_val, cg_val); |
298 | 0 | STORE_SI128(cgplane, cg_val); |
299 | 0 | a_val = _mm_packus_epi16(a_val, a_val); |
300 | 0 | STORE_SI128(aplane, a_val); |
301 | 0 | yplane += 8; |
302 | 0 | coplane += 8; |
303 | 0 | cgplane += 8; |
304 | 0 | aplane += 8; |
305 | 0 | } |
306 | |
|
307 | 0 | if (context->ChromaSubsamplingLevel > 0 && (context->width % 2) == 1) |
308 | 0 | { |
309 | 0 | context->priv->PlaneBuffers[0][y * rw + context->width] = |
310 | 0 | context->priv->PlaneBuffers[0][y * rw + context->width - 1]; |
311 | 0 | context->priv->PlaneBuffers[1][y * rw + context->width] = |
312 | 0 | context->priv->PlaneBuffers[1][y * rw + context->width - 1]; |
313 | 0 | context->priv->PlaneBuffers[2][y * rw + context->width] = |
314 | 0 | context->priv->PlaneBuffers[2][y * rw + context->width - 1]; |
315 | 0 | } |
316 | 0 | } |
317 | |
|
318 | 0 | if (context->ChromaSubsamplingLevel > 0 && (y % 2) == 1) |
319 | 0 | { |
320 | 0 | BYTE* yplane = context->priv->PlaneBuffers[0] + y * rw; |
321 | 0 | BYTE* coplane = context->priv->PlaneBuffers[1] + y * rw; |
322 | 0 | BYTE* cgplane = context->priv->PlaneBuffers[2] + y * rw; |
323 | 0 | CopyMemory(yplane, yplane - rw, rw); |
324 | 0 | CopyMemory(coplane, coplane - rw, rw); |
325 | 0 | CopyMemory(cgplane, cgplane - rw, rw); |
326 | 0 | } |
327 | |
|
328 | 0 | return TRUE; |
329 | 0 | } |
330 | | |
331 | | static void nsc_encode_subsampling_sse2(NSC_CONTEXT* context) |
332 | 0 | { |
333 | 0 | BYTE* co_dst = NULL; |
334 | 0 | BYTE* cg_dst = NULL; |
335 | 0 | INT8* co_src0 = NULL; |
336 | 0 | INT8* co_src1 = NULL; |
337 | 0 | INT8* cg_src0 = NULL; |
338 | 0 | INT8* cg_src1 = NULL; |
339 | 0 | UINT32 tempWidth = 0; |
340 | 0 | UINT32 tempHeight = 0; |
341 | 0 | __m128i t; |
342 | 0 | __m128i val; |
343 | 0 | __m128i mask = _mm_set1_epi16(0xFF); |
344 | 0 | tempWidth = ROUND_UP_TO(context->width, 8); |
345 | 0 | tempHeight = ROUND_UP_TO(context->height, 2); |
346 | |
|
347 | 0 | for (size_t y = 0; y < tempHeight >> 1; y++) |
348 | 0 | { |
349 | 0 | co_dst = context->priv->PlaneBuffers[1] + y * (tempWidth >> 1); |
350 | 0 | cg_dst = context->priv->PlaneBuffers[2] + y * (tempWidth >> 1); |
351 | 0 | co_src0 = (INT8*)context->priv->PlaneBuffers[1] + (y << 1) * tempWidth; |
352 | 0 | co_src1 = co_src0 + tempWidth; |
353 | 0 | cg_src0 = (INT8*)context->priv->PlaneBuffers[2] + (y << 1) * tempWidth; |
354 | 0 | cg_src1 = cg_src0 + tempWidth; |
355 | |
|
356 | 0 | for (UINT32 x = 0; x < tempWidth >> 1; x += 8) |
357 | 0 | { |
358 | 0 | t = LOAD_SI128(co_src0); |
359 | 0 | t = _mm_avg_epu8(t, LOAD_SI128(co_src1)); |
360 | 0 | val = _mm_and_si128(_mm_srli_si128(t, 1), mask); |
361 | 0 | val = _mm_avg_epu16(val, _mm_and_si128(t, mask)); |
362 | 0 | val = _mm_packus_epi16(val, val); |
363 | 0 | STORE_SI128(co_dst, val); |
364 | 0 | co_dst += 8; |
365 | 0 | co_src0 += 16; |
366 | 0 | co_src1 += 16; |
367 | 0 | t = LOAD_SI128(cg_src0); |
368 | 0 | t = _mm_avg_epu8(t, LOAD_SI128(cg_src1)); |
369 | 0 | val = _mm_and_si128(_mm_srli_si128(t, 1), mask); |
370 | 0 | val = _mm_avg_epu16(val, _mm_and_si128(t, mask)); |
371 | 0 | val = _mm_packus_epi16(val, val); |
372 | 0 | STORE_SI128(cg_dst, val); |
373 | 0 | cg_dst += 8; |
374 | 0 | cg_src0 += 16; |
375 | 0 | cg_src1 += 16; |
376 | 0 | } |
377 | 0 | } |
378 | 0 | } |
379 | | |
380 | | static BOOL nsc_encode_sse2(NSC_CONTEXT* WINPR_RESTRICT context, const BYTE* WINPR_RESTRICT data, |
381 | | UINT32 scanline) |
382 | 0 | { |
383 | 0 | if (!nsc_encode_argb_to_aycocg_sse2(context, data, scanline)) |
384 | 0 | return FALSE; |
385 | | |
386 | 0 | if (context->ChromaSubsamplingLevel > 0) |
387 | 0 | nsc_encode_subsampling_sse2(context); |
388 | |
|
389 | 0 | return TRUE; |
390 | 0 | } |
391 | | #endif |
392 | | |
393 | | void nsc_init_sse2_int(NSC_CONTEXT* WINPR_RESTRICT context) |
394 | 0 | { |
395 | 0 | #if defined(SSE_AVX_INTRINSICS_ENABLED) |
396 | 0 | WLog_VRB(PRIM_TAG, "SSE2/SSE3 optimizations"); |
397 | 0 | PROFILER_RENAME(context->priv->prof_nsc_encode, "nsc_encode_sse2") |
398 | 0 | context->encode = nsc_encode_sse2; |
399 | | #else |
400 | | WLog_VRB(PRIM_TAG, "undefined WITH_SIMD or SSE2 intrinsics not available"); |
401 | | WINPR_UNUSED(context); |
402 | | #endif |
403 | 0 | } |