Chromium Code Reviews
chromiumcodereview-hr@appspot.gserviceaccount.com (chromiumcodereview-hr) | Please choose your nickname with Settings | Help | Chromium Project | Gerrit Changes | Sign out
(565)

Side by Side Diff: simd/jsimd_i386.c

Issue 1953443002: Update to libjpeg_turbo 1.4.90 (Closed) Base URL: https://chromium.googlesource.com/chromium/deps/libjpeg_turbo.git@master
Patch Set: Created 4 years, 7 months ago
Use n/p to move between diff chunks; N/P to move between comments. Draft comments are only viewable by you.
Jump to:
View unified diff | Download patch
« no previous file with comments | « simd/jsimd_arm_neon.S ('k') | simd/jsimd_mips.c » ('j') | no next file with comments »
Toggle Intra-line Diffs ('i') | Expand Comments ('e') | Collapse Comments ('c') | Show Comments Hide Comments ('s')
OLDNEW
1 /* 1 /*
2 * jsimd_i386.c 2 * jsimd_i386.c
3 * 3 *
4 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB 4 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 * Copyright 2009-2011, 2013-2014 D. R. Commander 5 * Copyright 2009-2011, 2013-2014, 2016 D. R. Commander
6 * 6 * Copyright 2015 Matthieu Darbois
7 *
7 * Based on the x86 SIMD extension for IJG JPEG library, 8 * Based on the x86 SIMD extension for IJG JPEG library,
8 * Copyright (C) 1999-2006, MIYASAKA Masaru. 9 * Copyright (C) 1999-2006, MIYASAKA Masaru.
9 * For conditions of distribution and use, see copyright notice in jsimdext.inc 10 * For conditions of distribution and use, see copyright notice in jsimdext.inc
10 * 11 *
11 * This file contains the interface between the "normal" portions 12 * This file contains the interface between the "normal" portions
12 * of the library and the SIMD implementations when running on a 13 * of the library and the SIMD implementations when running on a
13 * 32-bit x86 architecture. 14 * 32-bit x86 architecture.
14 */ 15 */
15 16
16 #define JPEG_INTERNALS 17 #define JPEG_INTERNALS
17 #include "../jinclude.h" 18 #include "../jinclude.h"
18 #include "../jpeglib.h" 19 #include "../jpeglib.h"
19 #include "../jsimd.h" 20 #include "../jsimd.h"
20 #include "../jdct.h" 21 #include "../jdct.h"
21 #include "../jsimddct.h" 22 #include "../jsimddct.h"
22 #include "jsimd.h" 23 #include "jsimd.h"
23 24
24 /* 25 /*
25 * In the PIC cases, we have no guarantee that constants will keep 26 * In the PIC cases, we have no guarantee that constants will keep
26 * their alignment. This macro allows us to verify it at runtime. 27 * their alignment. This macro allows us to verify it at runtime.
27 */ 28 */
28 #define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0) 29 #define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0)
29 30
30 #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */ 31 #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
31 32
32 static unsigned int simd_support = ~0; 33 static unsigned int simd_support = ~0;
34 static unsigned int simd_huffman = 1;
33 35
34 /* 36 /*
35 * Check what SIMD accelerations are supported. 37 * Check what SIMD accelerations are supported.
36 * 38 *
37 * FIXME: This code is racy under a multi-threaded environment. 39 * FIXME: This code is racy under a multi-threaded environment.
38 */ 40 */
39 LOCAL(void) 41 LOCAL(void)
40 init_simd (void) 42 init_simd (void)
41 { 43 {
42 char *env = NULL; 44 char *env = NULL;
43 45
44 if (simd_support != ~0U) 46 if (simd_support != ~0U)
45 return; 47 return;
46 48
47 simd_support = jpeg_simd_cpu_support(); 49 simd_support = jpeg_simd_cpu_support();
48 50
49 /* Force different settings through environment variables */ 51 /* Force different settings through environment variables */
50 env = getenv("JSIMD_FORCEMMX"); 52 env = getenv("JSIMD_FORCEMMX");
51 if ((env != NULL) && (strcmp(env, "1") == 0)) 53 if ((env != NULL) && (strcmp(env, "1") == 0))
52 simd_support &= JSIMD_MMX; 54 simd_support &= JSIMD_MMX;
53 env = getenv("JSIMD_FORCE3DNOW"); 55 env = getenv("JSIMD_FORCE3DNOW");
54 if ((env != NULL) && (strcmp(env, "1") == 0)) 56 if ((env != NULL) && (strcmp(env, "1") == 0))
55 simd_support &= JSIMD_3DNOW|JSIMD_MMX; 57 simd_support &= JSIMD_3DNOW|JSIMD_MMX;
56 env = getenv("JSIMD_FORCESSE"); 58 env = getenv("JSIMD_FORCESSE");
57 if ((env != NULL) && (strcmp(env, "1") == 0)) 59 if ((env != NULL) && (strcmp(env, "1") == 0))
58 simd_support &= JSIMD_SSE|JSIMD_MMX; 60 simd_support &= JSIMD_SSE|JSIMD_MMX;
59 env = getenv("JSIMD_FORCESSE2"); 61 env = getenv("JSIMD_FORCESSE2");
60 if ((env != NULL) && (strcmp(env, "1") == 0)) 62 if ((env != NULL) && (strcmp(env, "1") == 0))
61 simd_support &= JSIMD_SSE2; 63 simd_support &= JSIMD_SSE2;
64 env = getenv("JSIMD_FORCENONE");
65 if ((env != NULL) && (strcmp(env, "1") == 0))
66 simd_support = 0;
67 env = getenv("JSIMD_NOHUFFENC");
68 if ((env != NULL) && (strcmp(env, "1") == 0))
69 simd_huffman = 0;
62 } 70 }
63 71
64 #ifndef JPEG_DECODE_ONLY
65 GLOBAL(int) 72 GLOBAL(int)
66 jsimd_can_rgb_ycc (void) 73 jsimd_can_rgb_ycc (void)
67 { 74 {
68 init_simd(); 75 init_simd();
69 76
70 /* The code is optimised for these values only */ 77 /* The code is optimised for these values only */
71 if (BITS_IN_JSAMPLE != 8) 78 if (BITS_IN_JSAMPLE != 8)
72 return 0; 79 return 0;
73 if (sizeof(JDIMENSION) != 4) 80 if (sizeof(JDIMENSION) != 4)
74 return 0; 81 return 0;
75 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) 82 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
76 return 0; 83 return 0;
77 84
78 if ((simd_support & JSIMD_SSE2) && 85 if ((simd_support & JSIMD_SSE2) &&
79 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2)) 86 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
80 return 1; 87 return 1;
81 if (simd_support & JSIMD_MMX) 88 if (simd_support & JSIMD_MMX)
82 return 1; 89 return 1;
83 90
84 return 0; 91 return 0;
85 } 92 }
86 #endif
87 93
88 GLOBAL(int) 94 GLOBAL(int)
89 jsimd_can_rgb_gray (void) 95 jsimd_can_rgb_gray (void)
90 { 96 {
91 init_simd(); 97 init_simd();
92 98
93 /* The code is optimised for these values only */ 99 /* The code is optimised for these values only */
94 if (BITS_IN_JSAMPLE != 8) 100 if (BITS_IN_JSAMPLE != 8)
95 return 0; 101 return 0;
96 if (sizeof(JDIMENSION) != 4) 102 if (sizeof(JDIMENSION) != 4)
(...skipping 31 matching lines...) Expand 10 before | Expand all | Expand 10 after
128 134
129 return 0; 135 return 0;
130 } 136 }
131 137
132 GLOBAL(int) 138 GLOBAL(int)
133 jsimd_can_ycc_rgb565 (void) 139 jsimd_can_ycc_rgb565 (void)
134 { 140 {
135 return 0; 141 return 0;
136 } 142 }
137 143
138 #ifndef JPEG_DECODE_ONLY
139 GLOBAL(void) 144 GLOBAL(void)
140 jsimd_rgb_ycc_convert (j_compress_ptr cinfo, 145 jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
141 JSAMPARRAY input_buf, JSAMPIMAGE output_buf, 146 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
142 JDIMENSION output_row, int num_rows) 147 JDIMENSION output_row, int num_rows)
143 { 148 {
144 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); 149 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
145 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); 150 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
146 151
147 switch(cinfo->in_color_space) 152 switch(cinfo->in_color_space) {
148 {
149 case JCS_EXT_RGB: 153 case JCS_EXT_RGB:
150 sse2fct=jsimd_extrgb_ycc_convert_sse2; 154 sse2fct=jsimd_extrgb_ycc_convert_sse2;
151 mmxfct=jsimd_extrgb_ycc_convert_mmx; 155 mmxfct=jsimd_extrgb_ycc_convert_mmx;
152 break; 156 break;
153 case JCS_EXT_RGBX: 157 case JCS_EXT_RGBX:
154 case JCS_EXT_RGBA: 158 case JCS_EXT_RGBA:
155 sse2fct=jsimd_extrgbx_ycc_convert_sse2; 159 sse2fct=jsimd_extrgbx_ycc_convert_sse2;
156 mmxfct=jsimd_extrgbx_ycc_convert_mmx; 160 mmxfct=jsimd_extrgbx_ycc_convert_mmx;
157 break; 161 break;
158 case JCS_EXT_BGR: 162 case JCS_EXT_BGR:
(...skipping 16 matching lines...) Expand all
175 mmxfct=jsimd_extxrgb_ycc_convert_mmx; 179 mmxfct=jsimd_extxrgb_ycc_convert_mmx;
176 break; 180 break;
177 default: 181 default:
178 sse2fct=jsimd_rgb_ycc_convert_sse2; 182 sse2fct=jsimd_rgb_ycc_convert_sse2;
179 mmxfct=jsimd_rgb_ycc_convert_mmx; 183 mmxfct=jsimd_rgb_ycc_convert_mmx;
180 break; 184 break;
181 } 185 }
182 186
183 if ((simd_support & JSIMD_SSE2) && 187 if ((simd_support & JSIMD_SSE2) &&
184 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2)) 188 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
185 sse2fct(cinfo->image_width, input_buf, 189 sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
186 output_buf, output_row, num_rows);
187 else if (simd_support & JSIMD_MMX) 190 else if (simd_support & JSIMD_MMX)
188 mmxfct(cinfo->image_width, input_buf, 191 mmxfct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
189 output_buf, output_row, num_rows);
190 }
191 #endif
192
193 GLOBAL(void)
194 jsimd_ycc_rgb565_convert (j_decompress_ptr cinfo,
195 JSAMPIMAGE input_buf, JDIMENSION input_row,
196 JSAMPARRAY output_buf, int num_rows)
197 {
198 } 192 }
199 193
200 GLOBAL(void) 194 GLOBAL(void)
201 jsimd_rgb_gray_convert (j_compress_ptr cinfo, 195 jsimd_rgb_gray_convert (j_compress_ptr cinfo,
202 JSAMPARRAY input_buf, JSAMPIMAGE output_buf, 196 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
203 JDIMENSION output_row, int num_rows) 197 JDIMENSION output_row, int num_rows)
204 { 198 {
205 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); 199 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
206 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); 200 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
207 201
208 switch(cinfo->in_color_space) 202 switch(cinfo->in_color_space) {
209 {
210 case JCS_EXT_RGB: 203 case JCS_EXT_RGB:
211 sse2fct=jsimd_extrgb_gray_convert_sse2; 204 sse2fct=jsimd_extrgb_gray_convert_sse2;
212 mmxfct=jsimd_extrgb_gray_convert_mmx; 205 mmxfct=jsimd_extrgb_gray_convert_mmx;
213 break; 206 break;
214 case JCS_EXT_RGBX: 207 case JCS_EXT_RGBX:
215 case JCS_EXT_RGBA: 208 case JCS_EXT_RGBA:
216 sse2fct=jsimd_extrgbx_gray_convert_sse2; 209 sse2fct=jsimd_extrgbx_gray_convert_sse2;
217 mmxfct=jsimd_extrgbx_gray_convert_mmx; 210 mmxfct=jsimd_extrgbx_gray_convert_mmx;
218 break; 211 break;
219 case JCS_EXT_BGR: 212 case JCS_EXT_BGR:
(...skipping 16 matching lines...) Expand all
236 mmxfct=jsimd_extxrgb_gray_convert_mmx; 229 mmxfct=jsimd_extxrgb_gray_convert_mmx;
237 break; 230 break;
238 default: 231 default:
239 sse2fct=jsimd_rgb_gray_convert_sse2; 232 sse2fct=jsimd_rgb_gray_convert_sse2;
240 mmxfct=jsimd_rgb_gray_convert_mmx; 233 mmxfct=jsimd_rgb_gray_convert_mmx;
241 break; 234 break;
242 } 235 }
243 236
244 if ((simd_support & JSIMD_SSE2) && 237 if ((simd_support & JSIMD_SSE2) &&
245 IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2)) 238 IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
246 sse2fct(cinfo->image_width, input_buf, 239 sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
247 output_buf, output_row, num_rows);
248 else if (simd_support & JSIMD_MMX) 240 else if (simd_support & JSIMD_MMX)
249 mmxfct(cinfo->image_width, input_buf, 241 mmxfct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
250 output_buf, output_row, num_rows);
251 } 242 }
252 243
253 GLOBAL(void) 244 GLOBAL(void)
254 jsimd_ycc_rgb_convert (j_decompress_ptr cinfo, 245 jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
255 JSAMPIMAGE input_buf, JDIMENSION input_row, 246 JSAMPIMAGE input_buf, JDIMENSION input_row,
256 JSAMPARRAY output_buf, int num_rows) 247 JSAMPARRAY output_buf, int num_rows)
257 { 248 {
258 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); 249 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
259 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); 250 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
260 251
261 switch(cinfo->out_color_space) 252 switch(cinfo->out_color_space) {
262 {
263 case JCS_EXT_RGB: 253 case JCS_EXT_RGB:
264 sse2fct=jsimd_ycc_extrgb_convert_sse2; 254 sse2fct=jsimd_ycc_extrgb_convert_sse2;
265 mmxfct=jsimd_ycc_extrgb_convert_mmx; 255 mmxfct=jsimd_ycc_extrgb_convert_mmx;
266 break; 256 break;
267 case JCS_EXT_RGBX: 257 case JCS_EXT_RGBX:
268 case JCS_EXT_RGBA: 258 case JCS_EXT_RGBA:
269 sse2fct=jsimd_ycc_extrgbx_convert_sse2; 259 sse2fct=jsimd_ycc_extrgbx_convert_sse2;
270 mmxfct=jsimd_ycc_extrgbx_convert_mmx; 260 mmxfct=jsimd_ycc_extrgbx_convert_mmx;
271 break; 261 break;
272 case JCS_EXT_BGR: 262 case JCS_EXT_BGR:
(...skipping 16 matching lines...) Expand all
289 mmxfct=jsimd_ycc_extxrgb_convert_mmx; 279 mmxfct=jsimd_ycc_extxrgb_convert_mmx;
290 break; 280 break;
291 default: 281 default:
292 sse2fct=jsimd_ycc_rgb_convert_sse2; 282 sse2fct=jsimd_ycc_rgb_convert_sse2;
293 mmxfct=jsimd_ycc_rgb_convert_mmx; 283 mmxfct=jsimd_ycc_rgb_convert_mmx;
294 break; 284 break;
295 } 285 }
296 286
297 if ((simd_support & JSIMD_SSE2) && 287 if ((simd_support & JSIMD_SSE2) &&
298 IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2)) 288 IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
299 sse2fct(cinfo->output_width, input_buf, 289 sse2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
300 input_row, output_buf, num_rows);
301 else if (simd_support & JSIMD_MMX) 290 else if (simd_support & JSIMD_MMX)
302 mmxfct(cinfo->output_width, input_buf, 291 mmxfct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
303 input_row, output_buf, num_rows);
304 } 292 }
305 293
306 #ifndef JPEG_DECODE_ONLY 294 GLOBAL(void)
295 jsimd_ycc_rgb565_convert (j_decompress_ptr cinfo,
296 JSAMPIMAGE input_buf, JDIMENSION input_row,
297 JSAMPARRAY output_buf, int num_rows)
298 {
299 }
300
307 GLOBAL(int) 301 GLOBAL(int)
308 jsimd_can_h2v2_downsample (void) 302 jsimd_can_h2v2_downsample (void)
309 { 303 {
310 init_simd(); 304 init_simd();
311 305
312 /* The code is optimised for these values only */ 306 /* The code is optimised for these values only */
313 if (BITS_IN_JSAMPLE != 8) 307 if (BITS_IN_JSAMPLE != 8)
314 return 0; 308 return 0;
315 if (sizeof(JDIMENSION) != 4) 309 if (sizeof(JDIMENSION) != 4)
316 return 0; 310 return 0;
(...skipping 19 matching lines...) Expand all
336 330
337 if (simd_support & JSIMD_SSE2) 331 if (simd_support & JSIMD_SSE2)
338 return 1; 332 return 1;
339 if (simd_support & JSIMD_MMX) 333 if (simd_support & JSIMD_MMX)
340 return 1; 334 return 1;
341 335
342 return 0; 336 return 0;
343 } 337 }
344 338
345 GLOBAL(void) 339 GLOBAL(void)
346 jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 340 jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
347 JSAMPARRAY input_data, JSAMPARRAY output_data) 341 JSAMPARRAY input_data, JSAMPARRAY output_data)
348 { 342 {
349 if (simd_support & JSIMD_SSE2) 343 if (simd_support & JSIMD_SSE2)
350 jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, 344 jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
351 compptr->v_samp_factor, compptr->width_in_blocks, 345 compptr->v_samp_factor,
352 input_data, output_data); 346 compptr->width_in_blocks, input_data,
347 output_data);
353 else if (simd_support & JSIMD_MMX) 348 else if (simd_support & JSIMD_MMX)
354 jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor, 349 jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
355 compptr->v_samp_factor, compptr->width_in_blocks, 350 compptr->v_samp_factor, compptr->width_in_blocks,
356 input_data, output_data); 351 input_data, output_data);
357 } 352 }
358 353
359 GLOBAL(void) 354 GLOBAL(void)
360 jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, 355 jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr,
361 JSAMPARRAY input_data, JSAMPARRAY output_data) 356 JSAMPARRAY input_data, JSAMPARRAY output_data)
362 { 357 {
363 if (simd_support & JSIMD_SSE2) 358 if (simd_support & JSIMD_SSE2)
364 jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, 359 jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
365 compptr->v_samp_factor, compptr->width_in_blocks, 360 compptr->v_samp_factor,
366 input_data, output_data); 361 compptr->width_in_blocks, input_data,
362 output_data);
367 else if (simd_support & JSIMD_MMX) 363 else if (simd_support & JSIMD_MMX)
368 jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor, 364 jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
369 compptr->v_samp_factor, compptr->width_in_blocks, 365 compptr->v_samp_factor, compptr->width_in_blocks,
370 input_data, output_data); 366 input_data, output_data);
371 } 367 }
372 #endif
373 368
374 GLOBAL(int) 369 GLOBAL(int)
375 jsimd_can_h2v2_upsample (void) 370 jsimd_can_h2v2_upsample (void)
376 { 371 {
377 init_simd(); 372 init_simd();
378 373
379 /* The code is optimised for these values only */ 374 /* The code is optimised for these values only */
380 if (BITS_IN_JSAMPLE != 8) 375 if (BITS_IN_JSAMPLE != 8)
381 return 0; 376 return 0;
382 if (sizeof(JDIMENSION) != 4) 377 if (sizeof(JDIMENSION) != 4)
(...skipping 21 matching lines...) Expand all
404 if (simd_support & JSIMD_SSE2) 399 if (simd_support & JSIMD_SSE2)
405 return 1; 400 return 1;
406 if (simd_support & JSIMD_MMX) 401 if (simd_support & JSIMD_MMX)
407 return 1; 402 return 1;
408 403
409 return 0; 404 return 0;
410 } 405 }
411 406
412 GLOBAL(void) 407 GLOBAL(void)
413 jsimd_h2v2_upsample (j_decompress_ptr cinfo, 408 jsimd_h2v2_upsample (j_decompress_ptr cinfo,
414 jpeg_component_info * compptr, 409 jpeg_component_info *compptr,
415 JSAMPARRAY input_data, 410 JSAMPARRAY input_data,
416 JSAMPARRAY * output_data_ptr) 411 JSAMPARRAY *output_data_ptr)
417 { 412 {
418 if (simd_support & JSIMD_SSE2) 413 if (simd_support & JSIMD_SSE2)
419 jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor, 414 jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width,
420 cinfo->output_width, input_data, output_data_ptr); 415 input_data, output_data_ptr);
421 else if (simd_support & JSIMD_MMX) 416 else if (simd_support & JSIMD_MMX)
422 jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor, 417 jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor, cinfo->output_width,
423 cinfo->output_width, input_data, output_data_ptr); 418 input_data, output_data_ptr);
424 } 419 }
425 420
426 GLOBAL(void) 421 GLOBAL(void)
427 jsimd_h2v1_upsample (j_decompress_ptr cinfo, 422 jsimd_h2v1_upsample (j_decompress_ptr cinfo,
428 jpeg_component_info * compptr, 423 jpeg_component_info *compptr,
429 JSAMPARRAY input_data, 424 JSAMPARRAY input_data,
430 JSAMPARRAY * output_data_ptr) 425 JSAMPARRAY *output_data_ptr)
431 { 426 {
432 if (simd_support & JSIMD_SSE2) 427 if (simd_support & JSIMD_SSE2)
433 jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor, 428 jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor, cinfo->output_width,
434 cinfo->output_width, input_data, output_data_ptr); 429 input_data, output_data_ptr);
435 else if (simd_support & JSIMD_MMX) 430 else if (simd_support & JSIMD_MMX)
436 jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor, 431 jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor, cinfo->output_width,
437 cinfo->output_width, input_data, output_data_ptr); 432 input_data, output_data_ptr);
438 } 433 }
439 434
440 GLOBAL(int) 435 GLOBAL(int)
441 jsimd_can_h2v2_fancy_upsample (void) 436 jsimd_can_h2v2_fancy_upsample (void)
442 { 437 {
443 init_simd(); 438 init_simd();
444 439
445 /* The code is optimised for these values only */ 440 /* The code is optimised for these values only */
446 if (BITS_IN_JSAMPLE != 8) 441 if (BITS_IN_JSAMPLE != 8)
447 return 0; 442 return 0;
(...skipping 24 matching lines...) Expand all
472 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) 467 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
473 return 1; 468 return 1;
474 if (simd_support & JSIMD_MMX) 469 if (simd_support & JSIMD_MMX)
475 return 1; 470 return 1;
476 471
477 return 0; 472 return 0;
478 } 473 }
479 474
480 GLOBAL(void) 475 GLOBAL(void)
481 jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo, 476 jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
482 jpeg_component_info * compptr, 477 jpeg_component_info *compptr,
483 JSAMPARRAY input_data, 478 JSAMPARRAY input_data,
484 JSAMPARRAY * output_data_ptr) 479 JSAMPARRAY *output_data_ptr)
485 { 480 {
486 if ((simd_support & JSIMD_SSE2) && 481 if ((simd_support & JSIMD_SSE2) &&
487 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) 482 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
488 jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor, 483 jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
489 compptr->downsampled_width, input_data, output_data_ptr); 484 compptr->downsampled_width, input_data,
485 output_data_ptr);
490 else if (simd_support & JSIMD_MMX) 486 else if (simd_support & JSIMD_MMX)
491 jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor, 487 jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor,
492 compptr->downsampled_width, input_data, output_data_ptr); 488 compptr->downsampled_width, input_data,
489 output_data_ptr);
493 } 490 }
494 491
495 GLOBAL(void) 492 GLOBAL(void)
496 jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo, 493 jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
497 jpeg_component_info * compptr, 494 jpeg_component_info *compptr,
498 JSAMPARRAY input_data, 495 JSAMPARRAY input_data,
499 JSAMPARRAY * output_data_ptr) 496 JSAMPARRAY *output_data_ptr)
500 { 497 {
501 if ((simd_support & JSIMD_SSE2) && 498 if ((simd_support & JSIMD_SSE2) &&
502 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) 499 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
503 jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor, 500 jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
504 compptr->downsampled_width, input_data, output_data_ptr); 501 compptr->downsampled_width, input_data,
502 output_data_ptr);
505 else if (simd_support & JSIMD_MMX) 503 else if (simd_support & JSIMD_MMX)
506 jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor, 504 jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor,
507 compptr->downsampled_width, input_data, output_data_ptr); 505 compptr->downsampled_width, input_data,
506 output_data_ptr);
508 } 507 }
509 508
510 GLOBAL(int) 509 GLOBAL(int)
511 jsimd_can_h2v2_merged_upsample (void) 510 jsimd_can_h2v2_merged_upsample (void)
512 { 511 {
513 init_simd(); 512 init_simd();
514 513
515 /* The code is optimised for these values only */ 514 /* The code is optimised for these values only */
516 if (BITS_IN_JSAMPLE != 8) 515 if (BITS_IN_JSAMPLE != 8)
517 return 0; 516 return 0;
(...skipping 31 matching lines...) Expand 10 before | Expand all | Expand 10 after
549 548
550 GLOBAL(void) 549 GLOBAL(void)
551 jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo, 550 jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
552 JSAMPIMAGE input_buf, 551 JSAMPIMAGE input_buf,
553 JDIMENSION in_row_group_ctr, 552 JDIMENSION in_row_group_ctr,
554 JSAMPARRAY output_buf) 553 JSAMPARRAY output_buf)
555 { 554 {
556 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); 555 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
557 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); 556 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
558 557
559 switch(cinfo->out_color_space) 558 switch(cinfo->out_color_space) {
560 {
561 case JCS_EXT_RGB: 559 case JCS_EXT_RGB:
562 sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2; 560 sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
563 mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx; 561 mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx;
564 break; 562 break;
565 case JCS_EXT_RGBX: 563 case JCS_EXT_RGBX:
566 case JCS_EXT_RGBA: 564 case JCS_EXT_RGBA:
567 sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2; 565 sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
568 mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx; 566 mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx;
569 break; 567 break;
570 case JCS_EXT_BGR: 568 case JCS_EXT_BGR:
(...skipping 16 matching lines...) Expand all
587 mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx; 585 mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx;
588 break; 586 break;
589 default: 587 default:
590 sse2fct=jsimd_h2v2_merged_upsample_sse2; 588 sse2fct=jsimd_h2v2_merged_upsample_sse2;
591 mmxfct=jsimd_h2v2_merged_upsample_mmx; 589 mmxfct=jsimd_h2v2_merged_upsample_mmx;
592 break; 590 break;
593 } 591 }
594 592
595 if ((simd_support & JSIMD_SSE2) && 593 if ((simd_support & JSIMD_SSE2) &&
596 IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) 594 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
597 sse2fct(cinfo->output_width, input_buf, 595 sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
598 in_row_group_ctr, output_buf);
599 else if (simd_support & JSIMD_MMX) 596 else if (simd_support & JSIMD_MMX)
600 mmxfct(cinfo->output_width, input_buf, 597 mmxfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
601 in_row_group_ctr, output_buf);
602 } 598 }
603 599
604 GLOBAL(void) 600 GLOBAL(void)
605 jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo, 601 jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
606 JSAMPIMAGE input_buf, 602 JSAMPIMAGE input_buf,
607 JDIMENSION in_row_group_ctr, 603 JDIMENSION in_row_group_ctr,
608 JSAMPARRAY output_buf) 604 JSAMPARRAY output_buf)
609 { 605 {
610 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); 606 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
611 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); 607 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
612 608
613 switch(cinfo->out_color_space) 609 switch(cinfo->out_color_space) {
614 {
615 case JCS_EXT_RGB: 610 case JCS_EXT_RGB:
616 sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2; 611 sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
617 mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx; 612 mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx;
618 break; 613 break;
619 case JCS_EXT_RGBX: 614 case JCS_EXT_RGBX:
620 case JCS_EXT_RGBA: 615 case JCS_EXT_RGBA:
621 sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2; 616 sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
622 mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx; 617 mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx;
623 break; 618 break;
624 case JCS_EXT_BGR: 619 case JCS_EXT_BGR:
(...skipping 16 matching lines...) Expand all
641 mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx; 636 mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx;
642 break; 637 break;
643 default: 638 default:
644 sse2fct=jsimd_h2v1_merged_upsample_sse2; 639 sse2fct=jsimd_h2v1_merged_upsample_sse2;
645 mmxfct=jsimd_h2v1_merged_upsample_mmx; 640 mmxfct=jsimd_h2v1_merged_upsample_mmx;
646 break; 641 break;
647 } 642 }
648 643
649 if ((simd_support & JSIMD_SSE2) && 644 if ((simd_support & JSIMD_SSE2) &&
650 IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) 645 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
651 sse2fct(cinfo->output_width, input_buf, 646 sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
652 in_row_group_ctr, output_buf);
653 else if (simd_support & JSIMD_MMX) 647 else if (simd_support & JSIMD_MMX)
654 mmxfct(cinfo->output_width, input_buf, 648 mmxfct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
655 in_row_group_ctr, output_buf);
656 } 649 }
657 650
658 #ifndef JPEG_DECODE_ONLY
659 GLOBAL(int) 651 GLOBAL(int)
660 jsimd_can_convsamp (void) 652 jsimd_can_convsamp (void)
661 { 653 {
662 init_simd(); 654 init_simd();
663 655
664 /* The code is optimised for these values only */ 656 /* The code is optimised for these values only */
665 if (DCTSIZE != 8) 657 if (DCTSIZE != 8)
666 return 0; 658 return 0;
667 if (BITS_IN_JSAMPLE != 8) 659 if (BITS_IN_JSAMPLE != 8)
668 return 0; 660 return 0;
(...skipping 30 matching lines...) Expand all
699 if (simd_support & JSIMD_SSE) 691 if (simd_support & JSIMD_SSE)
700 return 1; 692 return 1;
701 if (simd_support & JSIMD_3DNOW) 693 if (simd_support & JSIMD_3DNOW)
702 return 1; 694 return 1;
703 695
704 return 0; 696 return 0;
705 } 697 }
706 698
707 GLOBAL(void) 699 GLOBAL(void)
708 jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, 700 jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
709 DCTELEM * workspace) 701 DCTELEM *workspace)
710 { 702 {
711 if (simd_support & JSIMD_SSE2) 703 if (simd_support & JSIMD_SSE2)
712 jsimd_convsamp_sse2(sample_data, start_col, workspace); 704 jsimd_convsamp_sse2(sample_data, start_col, workspace);
713 else if (simd_support & JSIMD_MMX) 705 else if (simd_support & JSIMD_MMX)
714 jsimd_convsamp_mmx(sample_data, start_col, workspace); 706 jsimd_convsamp_mmx(sample_data, start_col, workspace);
715 } 707 }
716 708
717 GLOBAL(void) 709 GLOBAL(void)
718 jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, 710 jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
719 FAST_FLOAT * workspace) 711 FAST_FLOAT *workspace)
720 { 712 {
721 if (simd_support & JSIMD_SSE2) 713 if (simd_support & JSIMD_SSE2)
722 jsimd_convsamp_float_sse2(sample_data, start_col, workspace); 714 jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
723 else if (simd_support & JSIMD_SSE) 715 else if (simd_support & JSIMD_SSE)
724 jsimd_convsamp_float_sse(sample_data, start_col, workspace); 716 jsimd_convsamp_float_sse(sample_data, start_col, workspace);
725 else if (simd_support & JSIMD_3DNOW) 717 else if (simd_support & JSIMD_3DNOW)
726 jsimd_convsamp_float_3dnow(sample_data, start_col, workspace); 718 jsimd_convsamp_float_3dnow(sample_data, start_col, workspace);
727 } 719 }
728 720
729 GLOBAL(int) 721 GLOBAL(int)
(...skipping 47 matching lines...) Expand 10 before | Expand all | Expand 10 after
777 769
778 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse)) 770 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
779 return 1; 771 return 1;
780 if (simd_support & JSIMD_3DNOW) 772 if (simd_support & JSIMD_3DNOW)
781 return 1; 773 return 1;
782 774
783 return 0; 775 return 0;
784 } 776 }
785 777
786 GLOBAL(void) 778 GLOBAL(void)
787 jsimd_fdct_islow (DCTELEM * data) 779 jsimd_fdct_islow (DCTELEM *data)
788 { 780 {
789 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) 781 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
790 jsimd_fdct_islow_sse2(data); 782 jsimd_fdct_islow_sse2(data);
791 else if (simd_support & JSIMD_MMX) 783 else if (simd_support & JSIMD_MMX)
792 jsimd_fdct_islow_mmx(data); 784 jsimd_fdct_islow_mmx(data);
793 } 785 }
794 786
795 GLOBAL(void) 787 GLOBAL(void)
796 jsimd_fdct_ifast (DCTELEM * data) 788 jsimd_fdct_ifast (DCTELEM *data)
797 { 789 {
798 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) 790 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
799 jsimd_fdct_ifast_sse2(data); 791 jsimd_fdct_ifast_sse2(data);
800 else if (simd_support & JSIMD_MMX) 792 else if (simd_support & JSIMD_MMX)
801 jsimd_fdct_ifast_mmx(data); 793 jsimd_fdct_ifast_mmx(data);
802 } 794 }
803 795
804 GLOBAL(void) 796 GLOBAL(void)
805 jsimd_fdct_float (FAST_FLOAT * data) 797 jsimd_fdct_float (FAST_FLOAT *data)
806 { 798 {
807 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse)) 799 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
808 jsimd_fdct_float_sse(data); 800 jsimd_fdct_float_sse(data);
809 else if (simd_support & JSIMD_3DNOW) 801 else if (simd_support & JSIMD_3DNOW)
810 jsimd_fdct_float_3dnow(data); 802 jsimd_fdct_float_3dnow(data);
811 } 803 }
812 804
813 GLOBAL(int) 805 GLOBAL(int)
814 jsimd_can_quantize (void) 806 jsimd_can_quantize (void)
815 { 807 {
(...skipping 32 matching lines...) Expand 10 before | Expand all | Expand 10 after
848 return 1; 840 return 1;
849 if (simd_support & JSIMD_SSE) 841 if (simd_support & JSIMD_SSE)
850 return 1; 842 return 1;
851 if (simd_support & JSIMD_3DNOW) 843 if (simd_support & JSIMD_3DNOW)
852 return 1; 844 return 1;
853 845
854 return 0; 846 return 0;
855 } 847 }
856 848
857 GLOBAL(void) 849 GLOBAL(void)
858 jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors, 850 jsimd_quantize (JCOEFPTR coef_block, DCTELEM *divisors,
859 DCTELEM * workspace) 851 DCTELEM *workspace)
860 { 852 {
861 if (simd_support & JSIMD_SSE2) 853 if (simd_support & JSIMD_SSE2)
862 jsimd_quantize_sse2(coef_block, divisors, workspace); 854 jsimd_quantize_sse2(coef_block, divisors, workspace);
863 else if (simd_support & JSIMD_MMX) 855 else if (simd_support & JSIMD_MMX)
864 jsimd_quantize_mmx(coef_block, divisors, workspace); 856 jsimd_quantize_mmx(coef_block, divisors, workspace);
865 } 857 }
866 858
867 GLOBAL(void) 859 GLOBAL(void)
868 jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, 860 jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT *divisors,
869 FAST_FLOAT * workspace) 861 FAST_FLOAT *workspace)
870 { 862 {
871 if (simd_support & JSIMD_SSE2) 863 if (simd_support & JSIMD_SSE2)
872 jsimd_quantize_float_sse2(coef_block, divisors, workspace); 864 jsimd_quantize_float_sse2(coef_block, divisors, workspace);
873 else if (simd_support & JSIMD_SSE) 865 else if (simd_support & JSIMD_SSE)
874 jsimd_quantize_float_sse(coef_block, divisors, workspace); 866 jsimd_quantize_float_sse(coef_block, divisors, workspace);
875 else if (simd_support & JSIMD_3DNOW) 867 else if (simd_support & JSIMD_3DNOW)
876 jsimd_quantize_float_3dnow(coef_block, divisors, workspace); 868 jsimd_quantize_float_3dnow(coef_block, divisors, workspace);
877 } 869 }
878 #endif
879 870
880 GLOBAL(int) 871 GLOBAL(int)
881 jsimd_can_idct_2x2 (void) 872 jsimd_can_idct_2x2 (void)
882 { 873 {
883 init_simd(); 874 init_simd();
884 875
885 /* The code is optimised for these values only */ 876 /* The code is optimised for these values only */
886 if (DCTSIZE != 8) 877 if (DCTSIZE != 8)
887 return 0; 878 return 0;
888 if (sizeof(JCOEF) != 2) 879 if (sizeof(JCOEF) != 2)
(...skipping 32 matching lines...) Expand 10 before | Expand all | Expand 10 after
921 912
922 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) 913 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
923 return 1; 914 return 1;
924 if (simd_support & JSIMD_MMX) 915 if (simd_support & JSIMD_MMX)
925 return 1; 916 return 1;
926 917
927 return 0; 918 return 0;
928 } 919 }
929 920
930 GLOBAL(void) 921 GLOBAL(void)
931 jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr, 922 jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info *compptr,
932 JCOEFPTR coef_block, JSAMPARRAY output_buf, 923 JCOEFPTR coef_block, JSAMPARRAY output_buf,
933 JDIMENSION output_col) 924 JDIMENSION output_col)
934 { 925 {
935 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) 926 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
936 jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col); 927 jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf,
928 output_col);
937 else if (simd_support & JSIMD_MMX) 929 else if (simd_support & JSIMD_MMX)
938 jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col); 930 jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col);
939 } 931 }
940 932
941 GLOBAL(void) 933 GLOBAL(void)
942 jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr, 934 jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info *compptr,
943 JCOEFPTR coef_block, JSAMPARRAY output_buf, 935 JCOEFPTR coef_block, JSAMPARRAY output_buf,
944 JDIMENSION output_col) 936 JDIMENSION output_col)
945 { 937 {
946 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) 938 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
947 jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col); 939 jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf,
940 output_col);
948 else if (simd_support & JSIMD_MMX) 941 else if (simd_support & JSIMD_MMX)
949 jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col); 942 jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col);
950 } 943 }
951 944
952 GLOBAL(int) 945 GLOBAL(int)
953 jsimd_can_idct_islow (void) 946 jsimd_can_idct_islow (void)
954 { 947 {
955 init_simd(); 948 init_simd();
956 949
957 /* The code is optimised for these values only */ 950 /* The code is optimised for these values only */
(...skipping 65 matching lines...) Expand 10 before | Expand all | Expand 10 after
1023 return 1; 1016 return 1;
1024 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse)) 1017 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
1025 return 1; 1018 return 1;
1026 if (simd_support & JSIMD_3DNOW) 1019 if (simd_support & JSIMD_3DNOW)
1027 return 1; 1020 return 1;
1028 1021
1029 return 0; 1022 return 0;
1030 } 1023 }
1031 1024
1032 GLOBAL(void) 1025 GLOBAL(void)
1033 jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr, 1026 jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info *compptr,
1034 JCOEFPTR coef_block, JSAMPARRAY output_buf, 1027 JCOEFPTR coef_block, JSAMPARRAY output_buf,
1035 JDIMENSION output_col) 1028 JDIMENSION output_col)
1036 { 1029 {
1037 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2)) 1030 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
1038 jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col ); 1031 jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf,
1032 output_col);
1039 else if (simd_support & JSIMD_MMX) 1033 else if (simd_support & JSIMD_MMX)
1040 jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col) ; 1034 jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf,
1035 output_col);
1041 } 1036 }
1042 1037
1043 GLOBAL(void) 1038 GLOBAL(void)
1044 jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr, 1039 jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info *compptr,
1045 JCOEFPTR coef_block, JSAMPARRAY output_buf, 1040 JCOEFPTR coef_block, JSAMPARRAY output_buf,
1046 JDIMENSION output_col) 1041 JDIMENSION output_col)
1047 { 1042 {
1048 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2)) 1043 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
1049 jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col ); 1044 jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf,
1045 output_col);
1050 else if (simd_support & JSIMD_MMX) 1046 else if (simd_support & JSIMD_MMX)
1051 jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col) ; 1047 jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf,
1048 output_col);
1052 } 1049 }
1053 1050
1054 GLOBAL(void) 1051 GLOBAL(void)
1055 jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, 1052 jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info *compptr,
1056 JCOEFPTR coef_block, JSAMPARRAY output_buf, 1053 JCOEFPTR coef_block, JSAMPARRAY output_buf,
1057 JDIMENSION output_col) 1054 JDIMENSION output_col)
1058 { 1055 {
1059 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2)) 1056 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
1060 jsimd_idct_float_sse2(compptr->dct_table, coef_block, 1057 jsimd_idct_float_sse2(compptr->dct_table, coef_block, output_buf,
1061 output_buf, output_col); 1058 output_col);
1062 else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse)) 1059 else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
1063 jsimd_idct_float_sse(compptr->dct_table, coef_block, 1060 jsimd_idct_float_sse(compptr->dct_table, coef_block, output_buf,
1064 output_buf, output_col); 1061 output_col);
1065 else if (simd_support & JSIMD_3DNOW) 1062 else if (simd_support & JSIMD_3DNOW)
1066 jsimd_idct_float_3dnow(compptr->dct_table, coef_block, 1063 jsimd_idct_float_3dnow(compptr->dct_table, coef_block, output_buf,
1067 output_buf, output_col); 1064 output_col);
1068 } 1065 }
1066
1067 GLOBAL(int)
1068 jsimd_can_huff_encode_one_block (void)
1069 {
1070 init_simd();
1071
1072 if (DCTSIZE != 8)
1073 return 0;
1074 if (sizeof(JCOEF) != 2)
1075 return 0;
1076
1077 if ((simd_support & JSIMD_SSE2) && simd_huffman &&
1078 IS_ALIGNED_SSE(jconst_huff_encode_one_block))
1079 return 1;
1080
1081 return 0;
1082 }
1083
1084 GLOBAL(JOCTET*)
1085 jsimd_huff_encode_one_block (void *state, JOCTET *buffer, JCOEFPTR block,
1086 int last_dc_val, c_derived_tbl *dctbl,
1087 c_derived_tbl *actbl)
1088 {
1089 return jsimd_huff_encode_one_block_sse2(state, buffer, block, last_dc_val,
1090 dctbl, actbl);
1091 }
OLDNEW
« no previous file with comments | « simd/jsimd_arm_neon.S ('k') | simd/jsimd_mips.c » ('j') | no next file with comments »

Powered by Google App Engine
This is Rietveld 408576698