/src/mozilla-central/media/libjpeg/simd/jsimd_x86_64.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* |
2 | | * jsimd_x86_64.c |
3 | | * |
4 | | * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
5 | | * Copyright (C) 2009-2011, 2014, 2016, D. R. Commander. |
6 | | * Copyright (C) 2015, Matthieu Darbois. |
7 | | * |
8 | | * Based on the x86 SIMD extension for IJG JPEG library, |
9 | | * Copyright (C) 1999-2006, MIYASAKA Masaru. |
10 | | * For conditions of distribution and use, see copyright notice in jsimdext.inc |
11 | | * |
12 | | * This file contains the interface between the "normal" portions |
13 | | * of the library and the SIMD implementations when running on a |
14 | | * 64-bit x86 architecture. |
15 | | */ |
16 | | |
17 | | #define JPEG_INTERNALS |
18 | | #include "../jinclude.h" |
19 | | #include "../jpeglib.h" |
20 | | #include "../jsimd.h" |
21 | | #include "../jdct.h" |
22 | | #include "../jsimddct.h" |
23 | | #include "jsimd.h" |
24 | | |
25 | | /* |
26 | | * In the PIC cases, we have no guarantee that constants will keep |
27 | | * their alignment. This macro allows us to verify it at runtime. |
28 | | */ |
29 | 0 | #define IS_ALIGNED(ptr, order) (((size_t)ptr & ((1 << order) - 1)) == 0) |
30 | | |
31 | 0 | #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */ |
32 | | |
33 | | static unsigned int simd_support = ~0; |
34 | | static unsigned int simd_huffman = 1; |
35 | | |
36 | | /* |
37 | | * Check what SIMD accelerations are supported. |
38 | | * |
39 | | * FIXME: This code is racy under a multi-threaded environment. |
40 | | */ |
41 | | LOCAL(void) |
42 | | init_simd (void) |
43 | | { |
44 | | char *env = NULL; |
45 | | |
46 | | if (simd_support != ~0U) |
47 | | return; |
48 | | |
49 | | simd_support = JSIMD_SSE2 | JSIMD_SSE; |
50 | | |
51 | | /* Force different settings through environment variables */ |
52 | | env = getenv("JSIMD_FORCENONE"); |
53 | | if ((env != NULL) && (strcmp(env, "1") == 0)) |
54 | | simd_support = 0; |
55 | | env = getenv("JSIMD_NOHUFFENC"); |
56 | | if ((env != NULL) && (strcmp(env, "1") == 0)) |
57 | | simd_huffman = 0; |
58 | | } |
59 | | |
60 | | GLOBAL(int) |
61 | | jsimd_can_rgb_ycc (void) |
62 | 0 | { |
63 | 0 | init_simd(); |
64 | 0 |
|
65 | 0 | /* The code is optimised for these values only */ |
66 | 0 | if (BITS_IN_JSAMPLE != 8) |
67 | 0 | return 0; |
68 | 0 | if (sizeof(JDIMENSION) != 4) |
69 | 0 | return 0; |
70 | 0 | if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
71 | 0 | return 0; |
72 | 0 | |
73 | 0 | if ((simd_support & JSIMD_SSE2) && |
74 | 0 | IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2)) |
75 | 0 | return 1; |
76 | 0 | |
77 | 0 | return 0; |
78 | 0 | } |
79 | | |
80 | | GLOBAL(int) |
81 | | jsimd_can_rgb_gray (void) |
82 | 0 | { |
83 | 0 | init_simd(); |
84 | 0 |
|
85 | 0 | /* The code is optimised for these values only */ |
86 | 0 | if (BITS_IN_JSAMPLE != 8) |
87 | 0 | return 0; |
88 | 0 | if (sizeof(JDIMENSION) != 4) |
89 | 0 | return 0; |
90 | 0 | if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
91 | 0 | return 0; |
92 | 0 | |
93 | 0 | if ((simd_support & JSIMD_SSE2) && |
94 | 0 | IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2)) |
95 | 0 | return 1; |
96 | 0 | |
97 | 0 | return 0; |
98 | 0 | } |
99 | | |
100 | | GLOBAL(int) |
101 | | jsimd_can_ycc_rgb (void) |
102 | 0 | { |
103 | 0 | init_simd(); |
104 | 0 |
|
105 | 0 | /* The code is optimised for these values only */ |
106 | 0 | if (BITS_IN_JSAMPLE != 8) |
107 | 0 | return 0; |
108 | 0 | if (sizeof(JDIMENSION) != 4) |
109 | 0 | return 0; |
110 | 0 | if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
111 | 0 | return 0; |
112 | 0 | |
113 | 0 | if ((simd_support & JSIMD_SSE2) && |
114 | 0 | IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2)) |
115 | 0 | return 1; |
116 | 0 | |
117 | 0 | return 0; |
118 | 0 | } |
119 | | |
120 | | GLOBAL(int) |
121 | | jsimd_can_ycc_rgb565 (void) |
122 | 0 | { |
123 | 0 | return 0; |
124 | 0 | } |
125 | | |
126 | | GLOBAL(void) |
127 | | jsimd_rgb_ycc_convert (j_compress_ptr cinfo, |
128 | | JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
129 | | JDIMENSION output_row, int num_rows) |
130 | 0 | { |
131 | 0 | void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
132 | 0 |
|
133 | 0 | switch(cinfo->in_color_space) { |
134 | 0 | case JCS_EXT_RGB: |
135 | 0 | sse2fct=jsimd_extrgb_ycc_convert_sse2; |
136 | 0 | break; |
137 | 0 | case JCS_EXT_RGBX: |
138 | 0 | case JCS_EXT_RGBA: |
139 | 0 | sse2fct=jsimd_extrgbx_ycc_convert_sse2; |
140 | 0 | break; |
141 | 0 | case JCS_EXT_BGR: |
142 | 0 | sse2fct=jsimd_extbgr_ycc_convert_sse2; |
143 | 0 | break; |
144 | 0 | case JCS_EXT_BGRX: |
145 | 0 | case JCS_EXT_BGRA: |
146 | 0 | sse2fct=jsimd_extbgrx_ycc_convert_sse2; |
147 | 0 | break; |
148 | 0 | case JCS_EXT_XBGR: |
149 | 0 | case JCS_EXT_ABGR: |
150 | 0 | sse2fct=jsimd_extxbgr_ycc_convert_sse2; |
151 | 0 | break; |
152 | 0 | case JCS_EXT_XRGB: |
153 | 0 | case JCS_EXT_ARGB: |
154 | 0 | sse2fct=jsimd_extxrgb_ycc_convert_sse2; |
155 | 0 | break; |
156 | 0 | default: |
157 | 0 | sse2fct=jsimd_rgb_ycc_convert_sse2; |
158 | 0 | break; |
159 | 0 | } |
160 | 0 | |
161 | 0 | sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows); |
162 | 0 | } |
163 | | |
164 | | GLOBAL(void) |
165 | | jsimd_rgb_gray_convert (j_compress_ptr cinfo, |
166 | | JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
167 | | JDIMENSION output_row, int num_rows) |
168 | 0 | { |
169 | 0 | void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
170 | 0 |
|
171 | 0 | switch(cinfo->in_color_space) { |
172 | 0 | case JCS_EXT_RGB: |
173 | 0 | sse2fct=jsimd_extrgb_gray_convert_sse2; |
174 | 0 | break; |
175 | 0 | case JCS_EXT_RGBX: |
176 | 0 | case JCS_EXT_RGBA: |
177 | 0 | sse2fct=jsimd_extrgbx_gray_convert_sse2; |
178 | 0 | break; |
179 | 0 | case JCS_EXT_BGR: |
180 | 0 | sse2fct=jsimd_extbgr_gray_convert_sse2; |
181 | 0 | break; |
182 | 0 | case JCS_EXT_BGRX: |
183 | 0 | case JCS_EXT_BGRA: |
184 | 0 | sse2fct=jsimd_extbgrx_gray_convert_sse2; |
185 | 0 | break; |
186 | 0 | case JCS_EXT_XBGR: |
187 | 0 | case JCS_EXT_ABGR: |
188 | 0 | sse2fct=jsimd_extxbgr_gray_convert_sse2; |
189 | 0 | break; |
190 | 0 | case JCS_EXT_XRGB: |
191 | 0 | case JCS_EXT_ARGB: |
192 | 0 | sse2fct=jsimd_extxrgb_gray_convert_sse2; |
193 | 0 | break; |
194 | 0 | default: |
195 | 0 | sse2fct=jsimd_rgb_gray_convert_sse2; |
196 | 0 | break; |
197 | 0 | } |
198 | 0 | |
199 | 0 | sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows); |
200 | 0 | } |
201 | | |
202 | | GLOBAL(void) |
203 | | jsimd_ycc_rgb_convert (j_decompress_ptr cinfo, |
204 | | JSAMPIMAGE input_buf, JDIMENSION input_row, |
205 | | JSAMPARRAY output_buf, int num_rows) |
206 | 0 | { |
207 | 0 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); |
208 | 0 |
|
209 | 0 | switch(cinfo->out_color_space) { |
210 | 0 | case JCS_EXT_RGB: |
211 | 0 | sse2fct=jsimd_ycc_extrgb_convert_sse2; |
212 | 0 | break; |
213 | 0 | case JCS_EXT_RGBX: |
214 | 0 | case JCS_EXT_RGBA: |
215 | 0 | sse2fct=jsimd_ycc_extrgbx_convert_sse2; |
216 | 0 | break; |
217 | 0 | case JCS_EXT_BGR: |
218 | 0 | sse2fct=jsimd_ycc_extbgr_convert_sse2; |
219 | 0 | break; |
220 | 0 | case JCS_EXT_BGRX: |
221 | 0 | case JCS_EXT_BGRA: |
222 | 0 | sse2fct=jsimd_ycc_extbgrx_convert_sse2; |
223 | 0 | break; |
224 | 0 | case JCS_EXT_XBGR: |
225 | 0 | case JCS_EXT_ABGR: |
226 | 0 | sse2fct=jsimd_ycc_extxbgr_convert_sse2; |
227 | 0 | break; |
228 | 0 | case JCS_EXT_XRGB: |
229 | 0 | case JCS_EXT_ARGB: |
230 | 0 | sse2fct=jsimd_ycc_extxrgb_convert_sse2; |
231 | 0 | break; |
232 | 0 | default: |
233 | 0 | sse2fct=jsimd_ycc_rgb_convert_sse2; |
234 | 0 | break; |
235 | 0 | } |
236 | 0 | |
237 | 0 | sse2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows); |
238 | 0 | } |
239 | | |
240 | | GLOBAL(void) |
241 | | jsimd_ycc_rgb565_convert (j_decompress_ptr cinfo, |
242 | | JSAMPIMAGE input_buf, JDIMENSION input_row, |
243 | | JSAMPARRAY output_buf, int num_rows) |
244 | 0 | { |
245 | 0 | } |
246 | | |
247 | | GLOBAL(int) |
248 | | jsimd_can_h2v2_downsample (void) |
249 | 0 | { |
250 | 0 | init_simd(); |
251 | 0 |
|
252 | 0 | /* The code is optimised for these values only */ |
253 | 0 | if (BITS_IN_JSAMPLE != 8) |
254 | 0 | return 0; |
255 | 0 | if (sizeof(JDIMENSION) != 4) |
256 | 0 | return 0; |
257 | 0 | |
258 | 0 | if (simd_support & JSIMD_SSE2) |
259 | 0 | return 1; |
260 | 0 | |
261 | 0 | return 0; |
262 | 0 | } |
263 | | |
264 | | GLOBAL(int) |
265 | | jsimd_can_h2v1_downsample (void) |
266 | 0 | { |
267 | 0 | init_simd(); |
268 | 0 |
|
269 | 0 | /* The code is optimised for these values only */ |
270 | 0 | if (BITS_IN_JSAMPLE != 8) |
271 | 0 | return 0; |
272 | 0 | if (sizeof(JDIMENSION) != 4) |
273 | 0 | return 0; |
274 | 0 | |
275 | 0 | if (simd_support & JSIMD_SSE2) |
276 | 0 | return 1; |
277 | 0 | |
278 | 0 | return 0; |
279 | 0 | } |
280 | | |
281 | | GLOBAL(void) |
282 | | jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr, |
283 | | JSAMPARRAY input_data, JSAMPARRAY output_data) |
284 | 0 | { |
285 | 0 | jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, |
286 | 0 | compptr->v_samp_factor, compptr->width_in_blocks, |
287 | 0 | input_data, output_data); |
288 | 0 | } |
289 | | |
290 | | GLOBAL(void) |
291 | | jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr, |
292 | | JSAMPARRAY input_data, JSAMPARRAY output_data) |
293 | 0 | { |
294 | 0 | jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, |
295 | 0 | compptr->v_samp_factor, compptr->width_in_blocks, |
296 | 0 | input_data, output_data); |
297 | 0 | } |
298 | | |
299 | | GLOBAL(int) |
300 | | jsimd_can_h2v2_upsample (void) |
301 | 0 | { |
302 | 0 | init_simd(); |
303 | 0 |
|
304 | 0 | /* The code is optimised for these values only */ |
305 | 0 | if (BITS_IN_JSAMPLE != 8) |
306 | 0 | return 0; |
307 | 0 | if (sizeof(JDIMENSION) != 4) |
308 | 0 | return 0; |
309 | 0 | |
310 | 0 | if (simd_support & JSIMD_SSE2) |
311 | 0 | return 1; |
312 | 0 | |
313 | 0 | return 0; |
314 | 0 | } |
315 | | |
316 | | GLOBAL(int) |
317 | | jsimd_can_h2v1_upsample (void) |
318 | 0 | { |
319 | 0 | init_simd(); |
320 | 0 |
|
321 | 0 | /* The code is optimised for these values only */ |
322 | 0 | if (BITS_IN_JSAMPLE != 8) |
323 | 0 | return 0; |
324 | 0 | if (sizeof(JDIMENSION) != 4) |
325 | 0 | return 0; |
326 | 0 | |
327 | 0 | if (simd_support & JSIMD_SSE2) |
328 | 0 | return 1; |
329 | 0 | |
330 | 0 | return 0; |
331 | 0 | } |
332 | | |
333 | | GLOBAL(void) |
334 | | jsimd_h2v2_upsample (j_decompress_ptr cinfo, |
335 | | jpeg_component_info *compptr, |
336 | | JSAMPARRAY input_data, |
337 | | JSAMPARRAY *output_data_ptr) |
338 | 0 | { |
339 | 0 | jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width, |
340 | 0 | input_data, output_data_ptr); |
341 | 0 | } |
342 | | |
343 | | GLOBAL(void) |
344 | | jsimd_h2v1_upsample (j_decompress_ptr cinfo, |
345 | | jpeg_component_info *compptr, |
346 | | JSAMPARRAY input_data, |
347 | | JSAMPARRAY *output_data_ptr) |
348 | 0 | { |
349 | 0 | jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width, |
350 | 0 | input_data, output_data_ptr); |
351 | 0 | } |
352 | | |
353 | | GLOBAL(int) |
354 | | jsimd_can_h2v2_fancy_upsample (void) |
355 | 0 | { |
356 | 0 | init_simd(); |
357 | 0 |
|
358 | 0 | /* The code is optimised for these values only */ |
359 | 0 | if (BITS_IN_JSAMPLE != 8) |
360 | 0 | return 0; |
361 | 0 | if (sizeof(JDIMENSION) != 4) |
362 | 0 | return 0; |
363 | 0 | |
364 | 0 | if ((simd_support & JSIMD_SSE2) && |
365 | 0 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
366 | 0 | return 1; |
367 | 0 | |
368 | 0 | return 0; |
369 | 0 | } |
370 | | |
371 | | GLOBAL(int) |
372 | | jsimd_can_h2v1_fancy_upsample (void) |
373 | 0 | { |
374 | 0 | init_simd(); |
375 | 0 |
|
376 | 0 | /* The code is optimised for these values only */ |
377 | 0 | if (BITS_IN_JSAMPLE != 8) |
378 | 0 | return 0; |
379 | 0 | if (sizeof(JDIMENSION) != 4) |
380 | 0 | return 0; |
381 | 0 | |
382 | 0 | if ((simd_support & JSIMD_SSE2) && |
383 | 0 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
384 | 0 | return 1; |
385 | 0 | |
386 | 0 | return 0; |
387 | 0 | } |
388 | | |
389 | | GLOBAL(void) |
390 | | jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo, |
391 | | jpeg_component_info *compptr, |
392 | | JSAMPARRAY input_data, |
393 | | JSAMPARRAY *output_data_ptr) |
394 | 0 | { |
395 | 0 | jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor, |
396 | 0 | compptr->downsampled_width, input_data, |
397 | 0 | output_data_ptr); |
398 | 0 | } |
399 | | |
400 | | GLOBAL(void) |
401 | | jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo, |
402 | | jpeg_component_info *compptr, |
403 | | JSAMPARRAY input_data, |
404 | | JSAMPARRAY *output_data_ptr) |
405 | 0 | { |
406 | 0 | jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor, |
407 | 0 | compptr->downsampled_width, input_data, |
408 | 0 | output_data_ptr); |
409 | 0 | } |
410 | | |
411 | | GLOBAL(int) |
412 | | jsimd_can_h2v2_merged_upsample (void) |
413 | 0 | { |
414 | 0 | init_simd(); |
415 | 0 |
|
416 | 0 | /* The code is optimised for these values only */ |
417 | 0 | if (BITS_IN_JSAMPLE != 8) |
418 | 0 | return 0; |
419 | 0 | if (sizeof(JDIMENSION) != 4) |
420 | 0 | return 0; |
421 | 0 | |
422 | 0 | if ((simd_support & JSIMD_SSE2) && |
423 | 0 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
424 | 0 | return 1; |
425 | 0 | |
426 | 0 | return 0; |
427 | 0 | } |
428 | | |
429 | | GLOBAL(int) |
430 | | jsimd_can_h2v1_merged_upsample (void) |
431 | 0 | { |
432 | 0 | init_simd(); |
433 | 0 |
|
434 | 0 | /* The code is optimised for these values only */ |
435 | 0 | if (BITS_IN_JSAMPLE != 8) |
436 | 0 | return 0; |
437 | 0 | if (sizeof(JDIMENSION) != 4) |
438 | 0 | return 0; |
439 | 0 | |
440 | 0 | if ((simd_support & JSIMD_SSE2) && |
441 | 0 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
442 | 0 | return 1; |
443 | 0 | |
444 | 0 | return 0; |
445 | 0 | } |
446 | | |
447 | | GLOBAL(void) |
448 | | jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo, |
449 | | JSAMPIMAGE input_buf, |
450 | | JDIMENSION in_row_group_ctr, |
451 | | JSAMPARRAY output_buf) |
452 | 0 | { |
453 | 0 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
454 | 0 |
|
455 | 0 | switch(cinfo->out_color_space) { |
456 | 0 | case JCS_EXT_RGB: |
457 | 0 | sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2; |
458 | 0 | break; |
459 | 0 | case JCS_EXT_RGBX: |
460 | 0 | case JCS_EXT_RGBA: |
461 | 0 | sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2; |
462 | 0 | break; |
463 | 0 | case JCS_EXT_BGR: |
464 | 0 | sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2; |
465 | 0 | break; |
466 | 0 | case JCS_EXT_BGRX: |
467 | 0 | case JCS_EXT_BGRA: |
468 | 0 | sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2; |
469 | 0 | break; |
470 | 0 | case JCS_EXT_XBGR: |
471 | 0 | case JCS_EXT_ABGR: |
472 | 0 | sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2; |
473 | 0 | break; |
474 | 0 | case JCS_EXT_XRGB: |
475 | 0 | case JCS_EXT_ARGB: |
476 | 0 | sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2; |
477 | 0 | break; |
478 | 0 | default: |
479 | 0 | sse2fct=jsimd_h2v2_merged_upsample_sse2; |
480 | 0 | break; |
481 | 0 | } |
482 | 0 | |
483 | 0 | sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf); |
484 | 0 | } |
485 | | |
486 | | GLOBAL(void) |
487 | | jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo, |
488 | | JSAMPIMAGE input_buf, |
489 | | JDIMENSION in_row_group_ctr, |
490 | | JSAMPARRAY output_buf) |
491 | 0 | { |
492 | 0 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
493 | 0 |
|
494 | 0 | switch(cinfo->out_color_space) { |
495 | 0 | case JCS_EXT_RGB: |
496 | 0 | sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2; |
497 | 0 | break; |
498 | 0 | case JCS_EXT_RGBX: |
499 | 0 | case JCS_EXT_RGBA: |
500 | 0 | sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2; |
501 | 0 | break; |
502 | 0 | case JCS_EXT_BGR: |
503 | 0 | sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2; |
504 | 0 | break; |
505 | 0 | case JCS_EXT_BGRX: |
506 | 0 | case JCS_EXT_BGRA: |
507 | 0 | sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2; |
508 | 0 | break; |
509 | 0 | case JCS_EXT_XBGR: |
510 | 0 | case JCS_EXT_ABGR: |
511 | 0 | sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2; |
512 | 0 | break; |
513 | 0 | case JCS_EXT_XRGB: |
514 | 0 | case JCS_EXT_ARGB: |
515 | 0 | sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2; |
516 | 0 | break; |
517 | 0 | default: |
518 | 0 | sse2fct=jsimd_h2v1_merged_upsample_sse2; |
519 | 0 | break; |
520 | 0 | } |
521 | 0 | |
522 | 0 | sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf); |
523 | 0 | } |
524 | | |
525 | | GLOBAL(int) |
526 | | jsimd_can_convsamp (void) |
527 | 0 | { |
528 | 0 | init_simd(); |
529 | 0 |
|
530 | 0 | /* The code is optimised for these values only */ |
531 | 0 | if (DCTSIZE != 8) |
532 | 0 | return 0; |
533 | 0 | if (BITS_IN_JSAMPLE != 8) |
534 | 0 | return 0; |
535 | 0 | if (sizeof(JDIMENSION) != 4) |
536 | 0 | return 0; |
537 | 0 | if (sizeof(DCTELEM) != 2) |
538 | 0 | return 0; |
539 | 0 | |
540 | 0 | if (simd_support & JSIMD_SSE2) |
541 | 0 | return 1; |
542 | 0 | |
543 | 0 | return 0; |
544 | 0 | } |
545 | | |
546 | | GLOBAL(int) |
547 | | jsimd_can_convsamp_float (void) |
548 | 0 | { |
549 | 0 | init_simd(); |
550 | 0 |
|
551 | 0 | /* The code is optimised for these values only */ |
552 | 0 | if (DCTSIZE != 8) |
553 | 0 | return 0; |
554 | 0 | if (BITS_IN_JSAMPLE != 8) |
555 | 0 | return 0; |
556 | 0 | if (sizeof(JDIMENSION) != 4) |
557 | 0 | return 0; |
558 | 0 | if (sizeof(FAST_FLOAT) != 4) |
559 | 0 | return 0; |
560 | 0 | |
561 | 0 | if (simd_support & JSIMD_SSE2) |
562 | 0 | return 1; |
563 | 0 | |
564 | 0 | return 0; |
565 | 0 | } |
566 | | |
567 | | GLOBAL(void) |
568 | | jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, |
569 | | DCTELEM *workspace) |
570 | 0 | { |
571 | 0 | jsimd_convsamp_sse2(sample_data, start_col, workspace); |
572 | 0 | } |
573 | | |
574 | | GLOBAL(void) |
575 | | jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, |
576 | | FAST_FLOAT *workspace) |
577 | 0 | { |
578 | 0 | jsimd_convsamp_float_sse2(sample_data, start_col, workspace); |
579 | 0 | } |
580 | | |
581 | | GLOBAL(int) |
582 | | jsimd_can_fdct_islow (void) |
583 | 0 | { |
584 | 0 | init_simd(); |
585 | 0 |
|
586 | 0 | /* The code is optimised for these values only */ |
587 | 0 | if (DCTSIZE != 8) |
588 | 0 | return 0; |
589 | 0 | if (sizeof(DCTELEM) != 2) |
590 | 0 | return 0; |
591 | 0 | |
592 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) |
593 | 0 | return 1; |
594 | 0 | |
595 | 0 | return 0; |
596 | 0 | } |
597 | | |
598 | | GLOBAL(int) |
599 | | jsimd_can_fdct_ifast (void) |
600 | 0 | { |
601 | 0 | init_simd(); |
602 | 0 |
|
603 | 0 | /* The code is optimised for these values only */ |
604 | 0 | if (DCTSIZE != 8) |
605 | 0 | return 0; |
606 | 0 | if (sizeof(DCTELEM) != 2) |
607 | 0 | return 0; |
608 | 0 | |
609 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2)) |
610 | 0 | return 1; |
611 | 0 | |
612 | 0 | return 0; |
613 | 0 | } |
614 | | |
615 | | GLOBAL(int) |
616 | | jsimd_can_fdct_float (void) |
617 | 0 | { |
618 | 0 | init_simd(); |
619 | 0 |
|
620 | 0 | /* The code is optimised for these values only */ |
621 | 0 | if (DCTSIZE != 8) |
622 | 0 | return 0; |
623 | 0 | if (sizeof(FAST_FLOAT) != 4) |
624 | 0 | return 0; |
625 | 0 | |
626 | 0 | if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse)) |
627 | 0 | return 1; |
628 | 0 | |
629 | 0 | return 0; |
630 | 0 | } |
631 | | |
632 | | GLOBAL(void) |
633 | | jsimd_fdct_islow (DCTELEM *data) |
634 | 0 | { |
635 | 0 | jsimd_fdct_islow_sse2(data); |
636 | 0 | } |
637 | | |
638 | | GLOBAL(void) |
639 | | jsimd_fdct_ifast (DCTELEM *data) |
640 | 0 | { |
641 | 0 | jsimd_fdct_ifast_sse2(data); |
642 | 0 | } |
643 | | |
644 | | GLOBAL(void) |
645 | | jsimd_fdct_float (FAST_FLOAT *data) |
646 | 0 | { |
647 | 0 | jsimd_fdct_float_sse(data); |
648 | 0 | } |
649 | | |
650 | | GLOBAL(int) |
651 | | jsimd_can_quantize (void) |
652 | 0 | { |
653 | 0 | init_simd(); |
654 | 0 |
|
655 | 0 | /* The code is optimised for these values only */ |
656 | 0 | if (DCTSIZE != 8) |
657 | 0 | return 0; |
658 | 0 | if (sizeof(JCOEF) != 2) |
659 | 0 | return 0; |
660 | 0 | if (sizeof(DCTELEM) != 2) |
661 | 0 | return 0; |
662 | 0 | |
663 | 0 | if (simd_support & JSIMD_SSE2) |
664 | 0 | return 1; |
665 | 0 | |
666 | 0 | return 0; |
667 | 0 | } |
668 | | |
669 | | GLOBAL(int) |
670 | | jsimd_can_quantize_float (void) |
671 | 0 | { |
672 | 0 | init_simd(); |
673 | 0 |
|
674 | 0 | /* The code is optimised for these values only */ |
675 | 0 | if (DCTSIZE != 8) |
676 | 0 | return 0; |
677 | 0 | if (sizeof(JCOEF) != 2) |
678 | 0 | return 0; |
679 | 0 | if (sizeof(FAST_FLOAT) != 4) |
680 | 0 | return 0; |
681 | 0 | |
682 | 0 | if (simd_support & JSIMD_SSE2) |
683 | 0 | return 1; |
684 | 0 | |
685 | 0 | return 0; |
686 | 0 | } |
687 | | |
688 | | GLOBAL(void) |
689 | | jsimd_quantize (JCOEFPTR coef_block, DCTELEM *divisors, |
690 | | DCTELEM *workspace) |
691 | 0 | { |
692 | 0 | jsimd_quantize_sse2(coef_block, divisors, workspace); |
693 | 0 | } |
694 | | |
695 | | GLOBAL(void) |
696 | | jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT *divisors, |
697 | | FAST_FLOAT *workspace) |
698 | 0 | { |
699 | 0 | jsimd_quantize_float_sse2(coef_block, divisors, workspace); |
700 | 0 | } |
701 | | |
702 | | GLOBAL(int) |
703 | | jsimd_can_idct_2x2 (void) |
704 | 0 | { |
705 | 0 | init_simd(); |
706 | 0 |
|
707 | 0 | /* The code is optimised for these values only */ |
708 | 0 | if (DCTSIZE != 8) |
709 | 0 | return 0; |
710 | 0 | if (sizeof(JCOEF) != 2) |
711 | 0 | return 0; |
712 | 0 | if (BITS_IN_JSAMPLE != 8) |
713 | 0 | return 0; |
714 | 0 | if (sizeof(JDIMENSION) != 4) |
715 | 0 | return 0; |
716 | 0 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
717 | 0 | return 0; |
718 | 0 | |
719 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
720 | 0 | return 1; |
721 | 0 | |
722 | 0 | return 0; |
723 | 0 | } |
724 | | |
725 | | GLOBAL(int) |
726 | | jsimd_can_idct_4x4 (void) |
727 | 0 | { |
728 | 0 | init_simd(); |
729 | 0 |
|
730 | 0 | /* The code is optimised for these values only */ |
731 | 0 | if (DCTSIZE != 8) |
732 | 0 | return 0; |
733 | 0 | if (sizeof(JCOEF) != 2) |
734 | 0 | return 0; |
735 | 0 | if (BITS_IN_JSAMPLE != 8) |
736 | 0 | return 0; |
737 | 0 | if (sizeof(JDIMENSION) != 4) |
738 | 0 | return 0; |
739 | 0 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
740 | 0 | return 0; |
741 | 0 | |
742 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
743 | 0 | return 1; |
744 | 0 | |
745 | 0 | return 0; |
746 | 0 | } |
747 | | |
748 | | GLOBAL(void) |
749 | | jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
750 | | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
751 | | JDIMENSION output_col) |
752 | 0 | { |
753 | 0 | jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
754 | 0 | } |
755 | | |
756 | | GLOBAL(void) |
757 | | jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
758 | | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
759 | | JDIMENSION output_col) |
760 | 0 | { |
761 | 0 | jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
762 | 0 | } |
763 | | |
764 | | GLOBAL(int) |
765 | | jsimd_can_idct_islow (void) |
766 | 0 | { |
767 | 0 | init_simd(); |
768 | 0 |
|
769 | 0 | /* The code is optimised for these values only */ |
770 | 0 | if (DCTSIZE != 8) |
771 | 0 | return 0; |
772 | 0 | if (sizeof(JCOEF) != 2) |
773 | 0 | return 0; |
774 | 0 | if (BITS_IN_JSAMPLE != 8) |
775 | 0 | return 0; |
776 | 0 | if (sizeof(JDIMENSION) != 4) |
777 | 0 | return 0; |
778 | 0 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
779 | 0 | return 0; |
780 | 0 | |
781 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2)) |
782 | 0 | return 1; |
783 | 0 | |
784 | 0 | return 0; |
785 | 0 | } |
786 | | |
787 | | GLOBAL(int) |
788 | | jsimd_can_idct_ifast (void) |
789 | 0 | { |
790 | 0 | init_simd(); |
791 | 0 |
|
792 | 0 | /* The code is optimised for these values only */ |
793 | 0 | if (DCTSIZE != 8) |
794 | 0 | return 0; |
795 | 0 | if (sizeof(JCOEF) != 2) |
796 | 0 | return 0; |
797 | 0 | if (BITS_IN_JSAMPLE != 8) |
798 | 0 | return 0; |
799 | 0 | if (sizeof(JDIMENSION) != 4) |
800 | 0 | return 0; |
801 | 0 | if (sizeof(IFAST_MULT_TYPE) != 2) |
802 | 0 | return 0; |
803 | 0 | if (IFAST_SCALE_BITS != 2) |
804 | 0 | return 0; |
805 | 0 | |
806 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2)) |
807 | 0 | return 1; |
808 | 0 | |
809 | 0 | return 0; |
810 | 0 | } |
811 | | |
812 | | GLOBAL(int) |
813 | | jsimd_can_idct_float (void) |
814 | 0 | { |
815 | 0 | init_simd(); |
816 | 0 |
|
817 | 0 | if (DCTSIZE != 8) |
818 | 0 | return 0; |
819 | 0 | if (sizeof(JCOEF) != 2) |
820 | 0 | return 0; |
821 | 0 | if (BITS_IN_JSAMPLE != 8) |
822 | 0 | return 0; |
823 | 0 | if (sizeof(JDIMENSION) != 4) |
824 | 0 | return 0; |
825 | 0 | if (sizeof(FAST_FLOAT) != 4) |
826 | 0 | return 0; |
827 | 0 | if (sizeof(FLOAT_MULT_TYPE) != 4) |
828 | 0 | return 0; |
829 | 0 | |
830 | 0 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2)) |
831 | 0 | return 1; |
832 | 0 | |
833 | 0 | return 0; |
834 | 0 | } |
835 | | |
836 | | GLOBAL(void) |
837 | | jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
838 | | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
839 | | JDIMENSION output_col) |
840 | 0 | { |
841 | 0 | jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, |
842 | 0 | output_col); |
843 | 0 | } |
844 | | |
845 | | GLOBAL(void) |
846 | | jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
847 | | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
848 | | JDIMENSION output_col) |
849 | 0 | { |
850 | 0 | jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, |
851 | 0 | output_col); |
852 | 0 | } |
853 | | |
854 | | GLOBAL(void) |
855 | | jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
856 | | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
857 | | JDIMENSION output_col) |
858 | 0 | { |
859 | 0 | jsimd_idct_float_sse2(compptr->dct_table, coef_block, output_buf, |
860 | 0 | output_col); |
861 | 0 | } |
862 | | |
863 | | GLOBAL(int) |
864 | | jsimd_can_huff_encode_one_block (void) |
865 | 0 | { |
866 | 0 | init_simd(); |
867 | 0 |
|
868 | 0 | if (DCTSIZE != 8) |
869 | 0 | return 0; |
870 | 0 | if (sizeof(JCOEF) != 2) |
871 | 0 | return 0; |
872 | 0 | |
873 | 0 | if ((simd_support & JSIMD_SSE2) && simd_huffman && |
874 | 0 | IS_ALIGNED_SSE(jconst_huff_encode_one_block)) |
875 | 0 | return 1; |
876 | 0 | |
877 | 0 | return 0; |
878 | 0 | } |
879 | | |
880 | | GLOBAL(JOCTET*) |
881 | | jsimd_huff_encode_one_block (void *state, JOCTET *buffer, JCOEFPTR block, |
882 | | int last_dc_val, c_derived_tbl *dctbl, |
883 | | c_derived_tbl *actbl) |
884 | 0 | { |
885 | 0 | return jsimd_huff_encode_one_block_sse2(state, buffer, block, last_dc_val, |
886 | 0 | dctbl, actbl); |
887 | 0 | } |