The Tor Browser / annotate

media/libjpeg/simd/jsimd_i386.c@b8a032363ba2 (annotated)

media/libjpeg/simd/jsimd_i386.c

Thu, 22 Jan 2015 13:21:57 +0100

author
Michael Schloh von Bennewitz <michael@schloh.com>
date
Thu, 22 Jan 2015 13:21:57 +0100
branch
TOR_BUG_9701
changeset 15
b8a032363ba2
permissions
-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0 1 /*
michael@0 2 * jsimd_i386.c
michael@0 3 *
michael@0 4 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
michael@0 5 * Copyright 2009-2011 D. R. Commander
michael@0 6 *
michael@0 7 * Based on the x86 SIMD extension for IJG JPEG library,
michael@0 8 * Copyright (C) 1999-2006, MIYASAKA Masaru.
michael@0 9 * For conditions of distribution and use, see copyright notice in jsimdext.inc
michael@0 10 *
michael@0 11 * This file contains the interface between the "normal" portions
michael@0 12 * of the library and the SIMD implementations when running on a
michael@0 13 * 32-bit x86 architecture.
michael@0 14 */
michael@0 15
michael@0 16 #define JPEG_INTERNALS
michael@0 17 #include "../jinclude.h"
michael@0 18 #include "../jpeglib.h"
michael@0 19 #include "../jsimd.h"
michael@0 20 #include "../jdct.h"
michael@0 21 #include "../jsimddct.h"
michael@0 22 #include "jsimd.h"
michael@0 23
michael@0 24 /*
michael@0 25 * In the PIC cases, we have no guarantee that constants will keep
michael@0 26 * their alignment. This macro allows us to verify it at runtime.
michael@0 27 */
michael@0 28 #define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0)
michael@0 29
michael@0 30 #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
michael@0 31
michael@0 32 static unsigned int simd_support = ~0;
michael@0 33
michael@0 34 /*
michael@0 35 * Check what SIMD accelerations are supported.
michael@0 36 *
michael@0 37 * FIXME: This code is racy under a multi-threaded environment.
michael@0 38 */
michael@0 39 LOCAL(void)
michael@0 40 init_simd (void)
michael@0 41 {
michael@0 42 char *env = NULL;
michael@0 43
michael@0 44 if (simd_support != ~0U)
michael@0 45 return;
michael@0 46
michael@0 47 simd_support = jpeg_simd_cpu_support();
michael@0 48
michael@0 49 /* Force different settings through environment variables */
michael@0 50 env = getenv("JSIMD_FORCEMMX");
michael@0 51 if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0 52 simd_support &= JSIMD_MMX;
michael@0 53 env = getenv("JSIMD_FORCE3DNOW");
michael@0 54 if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0 55 simd_support &= JSIMD_3DNOW|JSIMD_MMX;
michael@0 56 env = getenv("JSIMD_FORCESSE");
michael@0 57 if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0 58 simd_support &= JSIMD_SSE|JSIMD_MMX;
michael@0 59 env = getenv("JSIMD_FORCESSE2");
michael@0 60 if ((env != NULL) && (strcmp(env, "1") == 0))
michael@0 61 simd_support &= JSIMD_SSE2;
michael@0 62 }
michael@0 63
michael@0 64 GLOBAL(int)
michael@0 65 jsimd_can_rgb_ycc (void)
michael@0 66 {
michael@0 67 init_simd();
michael@0 68
michael@0 69 /* The code is optimised for these values only */
michael@0 70 if (BITS_IN_JSAMPLE != 8)
michael@0 71 return 0;
michael@0 72 if (sizeof(JDIMENSION) != 4)
michael@0 73 return 0;
michael@0 74 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0 75 return 0;
michael@0 76
michael@0 77 if ((simd_support & JSIMD_SSE2) &&
michael@0 78 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
michael@0 79 return 1;
michael@0 80 if (simd_support & JSIMD_MMX)
michael@0 81 return 1;
michael@0 82
michael@0 83 return 0;
michael@0 84 }
michael@0 85
michael@0 86 GLOBAL(int)
michael@0 87 jsimd_can_rgb_gray (void)
michael@0 88 {
michael@0 89 init_simd();
michael@0 90
michael@0 91 /* The code is optimised for these values only */
michael@0 92 if (BITS_IN_JSAMPLE != 8)
michael@0 93 return 0;
michael@0 94 if (sizeof(JDIMENSION) != 4)
michael@0 95 return 0;
michael@0 96 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0 97 return 0;
michael@0 98
michael@0 99 if ((simd_support & JSIMD_SSE2) &&
michael@0 100 IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
michael@0 101 return 1;
michael@0 102 if (simd_support & JSIMD_MMX)
michael@0 103 return 1;
michael@0 104
michael@0 105 return 0;
michael@0 106 }
michael@0 107
michael@0 108 GLOBAL(int)
michael@0 109 jsimd_can_ycc_rgb (void)
michael@0 110 {
michael@0 111 init_simd();
michael@0 112
michael@0 113 /* The code is optimised for these values only */
michael@0 114 if (BITS_IN_JSAMPLE != 8)
michael@0 115 return 0;
michael@0 116 if (sizeof(JDIMENSION) != 4)
michael@0 117 return 0;
michael@0 118 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
michael@0 119 return 0;
michael@0 120
michael@0 121 if ((simd_support & JSIMD_SSE2) &&
michael@0 122 IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
michael@0 123 return 1;
michael@0 124 if (simd_support & JSIMD_MMX)
michael@0 125 return 1;
michael@0 126
michael@0 127 return 0;
michael@0 128 }
michael@0 129
michael@0 130 GLOBAL(void)
michael@0 131 jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
michael@0 132 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
michael@0 133 JDIMENSION output_row, int num_rows)
michael@0 134 {
michael@0 135 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0 136 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0 137
michael@0 138 switch(cinfo->in_color_space)
michael@0 139 {
michael@0 140 case JCS_EXT_RGB:
michael@0 141 sse2fct=jsimd_extrgb_ycc_convert_sse2;
michael@0 142 mmxfct=jsimd_extrgb_ycc_convert_mmx;
michael@0 143 break;
michael@0 144 case JCS_EXT_RGBX:
michael@0 145 case JCS_EXT_RGBA:
michael@0 146 sse2fct=jsimd_extrgbx_ycc_convert_sse2;
michael@0 147 mmxfct=jsimd_extrgbx_ycc_convert_mmx;
michael@0 148 break;
michael@0 149 case JCS_EXT_BGR:
michael@0 150 sse2fct=jsimd_extbgr_ycc_convert_sse2;
michael@0 151 mmxfct=jsimd_extbgr_ycc_convert_mmx;
michael@0 152 break;
michael@0 153 case JCS_EXT_BGRX:
michael@0 154 case JCS_EXT_BGRA:
michael@0 155 sse2fct=jsimd_extbgrx_ycc_convert_sse2;
michael@0 156 mmxfct=jsimd_extbgrx_ycc_convert_mmx;
michael@0 157 break;
michael@0 158 case JCS_EXT_XBGR:
michael@0 159 case JCS_EXT_ABGR:
michael@0 160 sse2fct=jsimd_extxbgr_ycc_convert_sse2;
michael@0 161 mmxfct=jsimd_extxbgr_ycc_convert_mmx;
michael@0 162 break;
michael@0 163 case JCS_EXT_XRGB:
michael@0 164 case JCS_EXT_ARGB:
michael@0 165 sse2fct=jsimd_extxrgb_ycc_convert_sse2;
michael@0 166 mmxfct=jsimd_extxrgb_ycc_convert_mmx;
michael@0 167 break;
michael@0 168 default:
michael@0 169 sse2fct=jsimd_rgb_ycc_convert_sse2;
michael@0 170 mmxfct=jsimd_rgb_ycc_convert_mmx;
michael@0 171 break;
michael@0 172 }
michael@0 173
michael@0 174 if ((simd_support & JSIMD_SSE2) &&
michael@0 175 IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
michael@0 176 sse2fct(cinfo->image_width, input_buf,
michael@0 177 output_buf, output_row, num_rows);
michael@0 178 else if (simd_support & JSIMD_MMX)
michael@0 179 mmxfct(cinfo->image_width, input_buf,
michael@0 180 output_buf, output_row, num_rows);
michael@0 181 }
michael@0 182
michael@0 183 GLOBAL(void)
michael@0 184 jsimd_rgb_gray_convert (j_compress_ptr cinfo,
michael@0 185 JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
michael@0 186 JDIMENSION output_row, int num_rows)
michael@0 187 {
michael@0 188 void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0 189 void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
michael@0 190
michael@0 191 switch(cinfo->in_color_space)
michael@0 192 {
michael@0 193 case JCS_EXT_RGB:
michael@0 194 sse2fct=jsimd_extrgb_gray_convert_sse2;
michael@0 195 mmxfct=jsimd_extrgb_gray_convert_mmx;
michael@0 196 break;
michael@0 197 case JCS_EXT_RGBX:
michael@0 198 case JCS_EXT_RGBA:
michael@0 199 sse2fct=jsimd_extrgbx_gray_convert_sse2;
michael@0 200 mmxfct=jsimd_extrgbx_gray_convert_mmx;
michael@0 201 break;
michael@0 202 case JCS_EXT_BGR:
michael@0 203 sse2fct=jsimd_extbgr_gray_convert_sse2;
michael@0 204 mmxfct=jsimd_extbgr_gray_convert_mmx;
michael@0 205 break;
michael@0 206 case JCS_EXT_BGRX:
michael@0 207 case JCS_EXT_BGRA:
michael@0 208 sse2fct=jsimd_extbgrx_gray_convert_sse2;
michael@0 209 mmxfct=jsimd_extbgrx_gray_convert_mmx;
michael@0 210 break;
michael@0 211 case JCS_EXT_XBGR:
michael@0 212 case JCS_EXT_ABGR:
michael@0 213 sse2fct=jsimd_extxbgr_gray_convert_sse2;
michael@0 214 mmxfct=jsimd_extxbgr_gray_convert_mmx;
michael@0 215 break;
michael@0 216 case JCS_EXT_XRGB:
michael@0 217 case JCS_EXT_ARGB:
michael@0 218 sse2fct=jsimd_extxrgb_gray_convert_sse2;
michael@0 219 mmxfct=jsimd_extxrgb_gray_convert_mmx;
michael@0 220 break;
michael@0 221 default:
michael@0 222 sse2fct=jsimd_rgb_gray_convert_sse2;
michael@0 223 mmxfct=jsimd_rgb_gray_convert_mmx;
michael@0 224 break;
michael@0 225 }
michael@0 226
michael@0 227 if ((simd_support & JSIMD_SSE2) &&
michael@0 228 IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
michael@0 229 sse2fct(cinfo->image_width, input_buf,
michael@0 230 output_buf, output_row, num_rows);
michael@0 231 else if (simd_support & JSIMD_MMX)
michael@0 232 mmxfct(cinfo->image_width, input_buf,
michael@0 233 output_buf, output_row, num_rows);
michael@0 234 }
michael@0 235
michael@0 236 GLOBAL(void)
michael@0 237 jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
michael@0 238 JSAMPIMAGE input_buf, JDIMENSION input_row,
michael@0 239 JSAMPARRAY output_buf, int num_rows)
michael@0 240 {
michael@0 241 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
michael@0 242 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
michael@0 243
michael@0 244 switch(cinfo->out_color_space)
michael@0 245 {
michael@0 246 case JCS_EXT_RGB:
michael@0 247 sse2fct=jsimd_ycc_extrgb_convert_sse2;
michael@0 248 mmxfct=jsimd_ycc_extrgb_convert_mmx;
michael@0 249 break;
michael@0 250 case JCS_EXT_RGBX:
michael@0 251 case JCS_EXT_RGBA:
michael@0 252 sse2fct=jsimd_ycc_extrgbx_convert_sse2;
michael@0 253 mmxfct=jsimd_ycc_extrgbx_convert_mmx;
michael@0 254 break;
michael@0 255 case JCS_EXT_BGR:
michael@0 256 sse2fct=jsimd_ycc_extbgr_convert_sse2;
michael@0 257 mmxfct=jsimd_ycc_extbgr_convert_mmx;
michael@0 258 break;
michael@0 259 case JCS_EXT_BGRX:
michael@0 260 case JCS_EXT_BGRA:
michael@0 261 sse2fct=jsimd_ycc_extbgrx_convert_sse2;
michael@0 262 mmxfct=jsimd_ycc_extbgrx_convert_mmx;
michael@0 263 break;
michael@0 264 case JCS_EXT_XBGR:
michael@0 265 case JCS_EXT_ABGR:
michael@0 266 sse2fct=jsimd_ycc_extxbgr_convert_sse2;
michael@0 267 mmxfct=jsimd_ycc_extxbgr_convert_mmx;
michael@0 268 break;
michael@0 269 case JCS_EXT_XRGB:
michael@0 270 case JCS_EXT_ARGB:
michael@0 271 sse2fct=jsimd_ycc_extxrgb_convert_sse2;
michael@0 272 mmxfct=jsimd_ycc_extxrgb_convert_mmx;
michael@0 273 break;
michael@0 274 default:
michael@0 275 sse2fct=jsimd_ycc_rgb_convert_sse2;
michael@0 276 mmxfct=jsimd_ycc_rgb_convert_mmx;
michael@0 277 break;
michael@0 278 }
michael@0 279
michael@0 280 if ((simd_support & JSIMD_SSE2) &&
michael@0 281 IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
michael@0 282 sse2fct(cinfo->output_width, input_buf,
michael@0 283 input_row, output_buf, num_rows);
michael@0 284 else if (simd_support & JSIMD_MMX)
michael@0 285 mmxfct(cinfo->output_width, input_buf,
michael@0 286 input_row, output_buf, num_rows);
michael@0 287 }
michael@0 288
michael@0 289 GLOBAL(int)
michael@0 290 jsimd_can_h2v2_downsample (void)
michael@0 291 {
michael@0 292 init_simd();
michael@0 293
michael@0 294 /* The code is optimised for these values only */
michael@0 295 if (BITS_IN_JSAMPLE != 8)
michael@0 296 return 0;
michael@0 297 if (sizeof(JDIMENSION) != 4)
michael@0 298 return 0;
michael@0 299
michael@0 300 if (simd_support & JSIMD_SSE2)
michael@0 301 return 1;
michael@0 302 if (simd_support & JSIMD_MMX)
michael@0 303 return 1;
michael@0 304
michael@0 305 return 0;
michael@0 306 }
michael@0 307
michael@0 308 GLOBAL(int)
michael@0 309 jsimd_can_h2v1_downsample (void)
michael@0 310 {
michael@0 311 init_simd();
michael@0 312
michael@0 313 /* The code is optimised for these values only */
michael@0 314 if (BITS_IN_JSAMPLE != 8)
michael@0 315 return 0;
michael@0 316 if (sizeof(JDIMENSION) != 4)
michael@0 317 return 0;
michael@0 318
michael@0 319 if (simd_support & JSIMD_SSE2)
michael@0 320 return 1;
michael@0 321 if (simd_support & JSIMD_MMX)
michael@0 322 return 1;
michael@0 323
michael@0 324 return 0;
michael@0 325 }
michael@0 326
michael@0 327 GLOBAL(void)
michael@0 328 jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0 329 JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0 330 {
michael@0 331 if (simd_support & JSIMD_SSE2)
michael@0 332 jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0 333 compptr->v_samp_factor, compptr->width_in_blocks,
michael@0 334 input_data, output_data);
michael@0 335 else if (simd_support & JSIMD_MMX)
michael@0 336 jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0 337 compptr->v_samp_factor, compptr->width_in_blocks,
michael@0 338 input_data, output_data);
michael@0 339 }
michael@0 340
michael@0 341 GLOBAL(void)
michael@0 342 jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
michael@0 343 JSAMPARRAY input_data, JSAMPARRAY output_data)
michael@0 344 {
michael@0 345 if (simd_support & JSIMD_SSE2)
michael@0 346 jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0 347 compptr->v_samp_factor, compptr->width_in_blocks,
michael@0 348 input_data, output_data);
michael@0 349 else if (simd_support & JSIMD_MMX)
michael@0 350 jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
michael@0 351 compptr->v_samp_factor, compptr->width_in_blocks,
michael@0 352 input_data, output_data);
michael@0 353 }
michael@0 354
michael@0 355 GLOBAL(int)
michael@0 356 jsimd_can_h2v2_upsample (void)
michael@0 357 {
michael@0 358 init_simd();
michael@0 359
michael@0 360 /* The code is optimised for these values only */
michael@0 361 if (BITS_IN_JSAMPLE != 8)
michael@0 362 return 0;
michael@0 363 if (sizeof(JDIMENSION) != 4)
michael@0 364 return 0;
michael@0 365
michael@0 366 if (simd_support & JSIMD_SSE2)
michael@0 367 return 1;
michael@0 368 if (simd_support & JSIMD_MMX)
michael@0 369 return 1;
michael@0 370
michael@0 371 return 0;
michael@0 372 }
michael@0 373
michael@0 374 GLOBAL(int)
michael@0 375 jsimd_can_h2v1_upsample (void)
michael@0 376 {
michael@0 377 init_simd();
michael@0 378
michael@0 379 /* The code is optimised for these values only */
michael@0 380 if (BITS_IN_JSAMPLE != 8)
michael@0 381 return 0;
michael@0 382 if (sizeof(JDIMENSION) != 4)
michael@0 383 return 0;
michael@0 384
michael@0 385 if (simd_support & JSIMD_SSE2)
michael@0 386 return 1;
michael@0 387 if (simd_support & JSIMD_MMX)
michael@0 388 return 1;
michael@0 389
michael@0 390 return 0;
michael@0 391 }
michael@0 392
michael@0 393 GLOBAL(void)
michael@0 394 jsimd_h2v2_upsample (j_decompress_ptr cinfo,
michael@0 395 jpeg_component_info * compptr,
michael@0 396 JSAMPARRAY input_data,
michael@0 397 JSAMPARRAY * output_data_ptr)
michael@0 398 {
michael@0 399 if (simd_support & JSIMD_SSE2)
michael@0 400 jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
michael@0 401 cinfo->output_width, input_data, output_data_ptr);
michael@0 402 else if (simd_support & JSIMD_MMX)
michael@0 403 jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor,
michael@0 404 cinfo->output_width, input_data, output_data_ptr);
michael@0 405 }
michael@0 406
michael@0 407 GLOBAL(void)
michael@0 408 jsimd_h2v1_upsample (j_decompress_ptr cinfo,
michael@0 409 jpeg_component_info * compptr,
michael@0 410 JSAMPARRAY input_data,
michael@0 411 JSAMPARRAY * output_data_ptr)
michael@0 412 {
michael@0 413 if (simd_support & JSIMD_SSE2)
michael@0 414 jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
michael@0 415 cinfo->output_width, input_data, output_data_ptr);
michael@0 416 else if (simd_support & JSIMD_MMX)
michael@0 417 jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor,
michael@0 418 cinfo->output_width, input_data, output_data_ptr);
michael@0 419 }
michael@0 420
michael@0 421 GLOBAL(int)
michael@0 422 jsimd_can_h2v2_fancy_upsample (void)
michael@0 423 {
michael@0 424 init_simd();
michael@0 425
michael@0 426 /* The code is optimised for these values only */
michael@0 427 if (BITS_IN_JSAMPLE != 8)
michael@0 428 return 0;
michael@0 429 if (sizeof(JDIMENSION) != 4)
michael@0 430 return 0;
michael@0 431
michael@0 432 if ((simd_support & JSIMD_SSE2) &&
michael@0 433 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0 434 return 1;
michael@0 435 if (simd_support & JSIMD_MMX)
michael@0 436 return 1;
michael@0 437
michael@0 438 return 0;
michael@0 439 }
michael@0 440
michael@0 441 GLOBAL(int)
michael@0 442 jsimd_can_h2v1_fancy_upsample (void)
michael@0 443 {
michael@0 444 init_simd();
michael@0 445
michael@0 446 /* The code is optimised for these values only */
michael@0 447 if (BITS_IN_JSAMPLE != 8)
michael@0 448 return 0;
michael@0 449 if (sizeof(JDIMENSION) != 4)
michael@0 450 return 0;
michael@0 451
michael@0 452 if ((simd_support & JSIMD_SSE2) &&
michael@0 453 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0 454 return 1;
michael@0 455 if (simd_support & JSIMD_MMX)
michael@0 456 return 1;
michael@0 457
michael@0 458 return 0;
michael@0 459 }
michael@0 460
michael@0 461 GLOBAL(void)
michael@0 462 jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
michael@0 463 jpeg_component_info * compptr,
michael@0 464 JSAMPARRAY input_data,
michael@0 465 JSAMPARRAY * output_data_ptr)
michael@0 466 {
michael@0 467 if ((simd_support & JSIMD_SSE2) &&
michael@0 468 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0 469 jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
michael@0 470 compptr->downsampled_width, input_data, output_data_ptr);
michael@0 471 else if (simd_support & JSIMD_MMX)
michael@0 472 jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor,
michael@0 473 compptr->downsampled_width, input_data, output_data_ptr);
michael@0 474 }
michael@0 475
michael@0 476 GLOBAL(void)
michael@0 477 jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
michael@0 478 jpeg_component_info * compptr,
michael@0 479 JSAMPARRAY input_data,
michael@0 480 JSAMPARRAY * output_data_ptr)
michael@0 481 {
michael@0 482 if ((simd_support & JSIMD_SSE2) &&
michael@0 483 IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
michael@0 484 jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
michael@0 485 compptr->downsampled_width, input_data, output_data_ptr);
michael@0 486 else if (simd_support & JSIMD_MMX)
michael@0 487 jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor,
michael@0 488 compptr->downsampled_width, input_data, output_data_ptr);
michael@0 489 }
michael@0 490
michael@0 491 GLOBAL(int)
michael@0 492 jsimd_can_h2v2_merged_upsample (void)
michael@0 493 {
michael@0 494 init_simd();
michael@0 495
michael@0 496 /* The code is optimised for these values only */
michael@0 497 if (BITS_IN_JSAMPLE != 8)
michael@0 498 return 0;
michael@0 499 if (sizeof(JDIMENSION) != 4)
michael@0 500 return 0;
michael@0 501
michael@0 502 if ((simd_support & JSIMD_SSE2) &&
michael@0 503 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0 504 return 1;
michael@0 505 if (simd_support & JSIMD_MMX)
michael@0 506 return 1;
michael@0 507
michael@0 508 return 0;
michael@0 509 }
michael@0 510
michael@0 511 GLOBAL(int)
michael@0 512 jsimd_can_h2v1_merged_upsample (void)
michael@0 513 {
michael@0 514 init_simd();
michael@0 515
michael@0 516 /* The code is optimised for these values only */
michael@0 517 if (BITS_IN_JSAMPLE != 8)
michael@0 518 return 0;
michael@0 519 if (sizeof(JDIMENSION) != 4)
michael@0 520 return 0;
michael@0 521
michael@0 522 if ((simd_support & JSIMD_SSE2) &&
michael@0 523 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0 524 return 1;
michael@0 525 if (simd_support & JSIMD_MMX)
michael@0 526 return 1;
michael@0 527
michael@0 528 return 0;
michael@0 529 }
michael@0 530
michael@0 531 GLOBAL(void)
michael@0 532 jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
michael@0 533 JSAMPIMAGE input_buf,
michael@0 534 JDIMENSION in_row_group_ctr,
michael@0 535 JSAMPARRAY output_buf)
michael@0 536 {
michael@0 537 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0 538 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0 539
michael@0 540 switch(cinfo->out_color_space)
michael@0 541 {
michael@0 542 case JCS_EXT_RGB:
michael@0 543 sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
michael@0 544 mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx;
michael@0 545 break;
michael@0 546 case JCS_EXT_RGBX:
michael@0 547 case JCS_EXT_RGBA:
michael@0 548 sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
michael@0 549 mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx;
michael@0 550 break;
michael@0 551 case JCS_EXT_BGR:
michael@0 552 sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
michael@0 553 mmxfct=jsimd_h2v2_extbgr_merged_upsample_mmx;
michael@0 554 break;
michael@0 555 case JCS_EXT_BGRX:
michael@0 556 case JCS_EXT_BGRA:
michael@0 557 sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
michael@0 558 mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx;
michael@0 559 break;
michael@0 560 case JCS_EXT_XBGR:
michael@0 561 case JCS_EXT_ABGR:
michael@0 562 sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
michael@0 563 mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx;
michael@0 564 break;
michael@0 565 case JCS_EXT_XRGB:
michael@0 566 case JCS_EXT_ARGB:
michael@0 567 sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
michael@0 568 mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx;
michael@0 569 break;
michael@0 570 default:
michael@0 571 sse2fct=jsimd_h2v2_merged_upsample_sse2;
michael@0 572 mmxfct=jsimd_h2v2_merged_upsample_mmx;
michael@0 573 break;
michael@0 574 }
michael@0 575
michael@0 576 if ((simd_support & JSIMD_SSE2) &&
michael@0 577 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0 578 sse2fct(cinfo->output_width, input_buf,
michael@0 579 in_row_group_ctr, output_buf);
michael@0 580 else if (simd_support & JSIMD_MMX)
michael@0 581 mmxfct(cinfo->output_width, input_buf,
michael@0 582 in_row_group_ctr, output_buf);
michael@0 583 }
michael@0 584
michael@0 585 GLOBAL(void)
michael@0 586 jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
michael@0 587 JSAMPIMAGE input_buf,
michael@0 588 JDIMENSION in_row_group_ctr,
michael@0 589 JSAMPARRAY output_buf)
michael@0 590 {
michael@0 591 void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0 592 void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
michael@0 593
michael@0 594 switch(cinfo->out_color_space)
michael@0 595 {
michael@0 596 case JCS_EXT_RGB:
michael@0 597 sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
michael@0 598 mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx;
michael@0 599 break;
michael@0 600 case JCS_EXT_RGBX:
michael@0 601 case JCS_EXT_RGBA:
michael@0 602 sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
michael@0 603 mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx;
michael@0 604 break;
michael@0 605 case JCS_EXT_BGR:
michael@0 606 sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
michael@0 607 mmxfct=jsimd_h2v1_extbgr_merged_upsample_mmx;
michael@0 608 break;
michael@0 609 case JCS_EXT_BGRX:
michael@0 610 case JCS_EXT_BGRA:
michael@0 611 sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
michael@0 612 mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx;
michael@0 613 break;
michael@0 614 case JCS_EXT_XBGR:
michael@0 615 case JCS_EXT_ABGR:
michael@0 616 sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
michael@0 617 mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx;
michael@0 618 break;
michael@0 619 case JCS_EXT_XRGB:
michael@0 620 case JCS_EXT_ARGB:
michael@0 621 sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
michael@0 622 mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx;
michael@0 623 break;
michael@0 624 default:
michael@0 625 sse2fct=jsimd_h2v1_merged_upsample_sse2;
michael@0 626 mmxfct=jsimd_h2v1_merged_upsample_mmx;
michael@0 627 break;
michael@0 628 }
michael@0 629
michael@0 630 if ((simd_support & JSIMD_SSE2) &&
michael@0 631 IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
michael@0 632 sse2fct(cinfo->output_width, input_buf,
michael@0 633 in_row_group_ctr, output_buf);
michael@0 634 else if (simd_support & JSIMD_MMX)
michael@0 635 mmxfct(cinfo->output_width, input_buf,
michael@0 636 in_row_group_ctr, output_buf);
michael@0 637 }
michael@0 638
michael@0 639 GLOBAL(int)
michael@0 640 jsimd_can_convsamp (void)
michael@0 641 {
michael@0 642 init_simd();
michael@0 643
michael@0 644 /* The code is optimised for these values only */
michael@0 645 if (DCTSIZE != 8)
michael@0 646 return 0;
michael@0 647 if (BITS_IN_JSAMPLE != 8)
michael@0 648 return 0;
michael@0 649 if (sizeof(JDIMENSION) != 4)
michael@0 650 return 0;
michael@0 651 if (sizeof(DCTELEM) != 2)
michael@0 652 return 0;
michael@0 653
michael@0 654 if (simd_support & JSIMD_SSE2)
michael@0 655 return 1;
michael@0 656 if (simd_support & JSIMD_MMX)
michael@0 657 return 1;
michael@0 658
michael@0 659 return 0;
michael@0 660 }
michael@0 661
michael@0 662 GLOBAL(int)
michael@0 663 jsimd_can_convsamp_float (void)
michael@0 664 {
michael@0 665 init_simd();
michael@0 666
michael@0 667 /* The code is optimised for these values only */
michael@0 668 if (DCTSIZE != 8)
michael@0 669 return 0;
michael@0 670 if (BITS_IN_JSAMPLE != 8)
michael@0 671 return 0;
michael@0 672 if (sizeof(JDIMENSION) != 4)
michael@0 673 return 0;
michael@0 674 if (sizeof(FAST_FLOAT) != 4)
michael@0 675 return 0;
michael@0 676
michael@0 677 if (simd_support & JSIMD_SSE2)
michael@0 678 return 1;
michael@0 679 if (simd_support & JSIMD_SSE)
michael@0 680 return 1;
michael@0 681 if (simd_support & JSIMD_3DNOW)
michael@0 682 return 1;
michael@0 683
michael@0 684 return 0;
michael@0 685 }
michael@0 686
michael@0 687 GLOBAL(void)
michael@0 688 jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0 689 DCTELEM * workspace)
michael@0 690 {
michael@0 691 if (simd_support & JSIMD_SSE2)
michael@0 692 jsimd_convsamp_sse2(sample_data, start_col, workspace);
michael@0 693 else if (simd_support & JSIMD_MMX)
michael@0 694 jsimd_convsamp_mmx(sample_data, start_col, workspace);
michael@0 695 }
michael@0 696
michael@0 697 GLOBAL(void)
michael@0 698 jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
michael@0 699 FAST_FLOAT * workspace)
michael@0 700 {
michael@0 701 if (simd_support & JSIMD_SSE2)
michael@0 702 jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
michael@0 703 else if (simd_support & JSIMD_SSE)
michael@0 704 jsimd_convsamp_float_sse(sample_data, start_col, workspace);
michael@0 705 else if (simd_support & JSIMD_3DNOW)
michael@0 706 jsimd_convsamp_float_3dnow(sample_data, start_col, workspace);
michael@0 707 }
michael@0 708
michael@0 709 GLOBAL(int)
michael@0 710 jsimd_can_fdct_islow (void)
michael@0 711 {
michael@0 712 init_simd();
michael@0 713
michael@0 714 /* The code is optimised for these values only */
michael@0 715 if (DCTSIZE != 8)
michael@0 716 return 0;
michael@0 717 if (sizeof(DCTELEM) != 2)
michael@0 718 return 0;
michael@0 719
michael@0 720 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0 721 return 1;
michael@0 722 if (simd_support & JSIMD_MMX)
michael@0 723 return 1;
michael@0 724
michael@0 725 return 0;
michael@0 726 }
michael@0 727
michael@0 728 GLOBAL(int)
michael@0 729 jsimd_can_fdct_ifast (void)
michael@0 730 {
michael@0 731 init_simd();
michael@0 732
michael@0 733 /* The code is optimised for these values only */
michael@0 734 if (DCTSIZE != 8)
michael@0 735 return 0;
michael@0 736 if (sizeof(DCTELEM) != 2)
michael@0 737 return 0;
michael@0 738
michael@0 739 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
michael@0 740 return 1;
michael@0 741 if (simd_support & JSIMD_MMX)
michael@0 742 return 1;
michael@0 743
michael@0 744 return 0;
michael@0 745 }
michael@0 746
michael@0 747 GLOBAL(int)
michael@0 748 jsimd_can_fdct_float (void)
michael@0 749 {
michael@0 750 init_simd();
michael@0 751
michael@0 752 /* The code is optimised for these values only */
michael@0 753 if (DCTSIZE != 8)
michael@0 754 return 0;
michael@0 755 if (sizeof(FAST_FLOAT) != 4)
michael@0 756 return 0;
michael@0 757
michael@0 758 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
michael@0 759 return 1;
michael@0 760 if (simd_support & JSIMD_3DNOW)
michael@0 761 return 1;
michael@0 762
michael@0 763 return 0;
michael@0 764 }
michael@0 765
michael@0 766 GLOBAL(void)
michael@0 767 jsimd_fdct_islow (DCTELEM * data)
michael@0 768 {
michael@0 769 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0 770 jsimd_fdct_islow_sse2(data);
michael@0 771 else if (simd_support & JSIMD_MMX)
michael@0 772 jsimd_fdct_islow_mmx(data);
michael@0 773 }
michael@0 774
michael@0 775 GLOBAL(void)
michael@0 776 jsimd_fdct_ifast (DCTELEM * data)
michael@0 777 {
michael@0 778 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
michael@0 779 jsimd_fdct_ifast_sse2(data);
michael@0 780 else if (simd_support & JSIMD_MMX)
michael@0 781 jsimd_fdct_ifast_mmx(data);
michael@0 782 }
michael@0 783
michael@0 784 GLOBAL(void)
michael@0 785 jsimd_fdct_float (FAST_FLOAT * data)
michael@0 786 {
michael@0 787 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
michael@0 788 jsimd_fdct_float_sse(data);
michael@0 789 else if (simd_support & JSIMD_3DNOW)
michael@0 790 jsimd_fdct_float_3dnow(data);
michael@0 791 }
michael@0 792
michael@0 793 GLOBAL(int)
michael@0 794 jsimd_can_quantize (void)
michael@0 795 {
michael@0 796 init_simd();
michael@0 797
michael@0 798 /* The code is optimised for these values only */
michael@0 799 if (DCTSIZE != 8)
michael@0 800 return 0;
michael@0 801 if (sizeof(JCOEF) != 2)
michael@0 802 return 0;
michael@0 803 if (sizeof(DCTELEM) != 2)
michael@0 804 return 0;
michael@0 805
michael@0 806 if (simd_support & JSIMD_SSE2)
michael@0 807 return 1;
michael@0 808 if (simd_support & JSIMD_MMX)
michael@0 809 return 1;
michael@0 810
michael@0 811 return 0;
michael@0 812 }
michael@0 813
michael@0 814 GLOBAL(int)
michael@0 815 jsimd_can_quantize_float (void)
michael@0 816 {
michael@0 817 init_simd();
michael@0 818
michael@0 819 /* The code is optimised for these values only */
michael@0 820 if (DCTSIZE != 8)
michael@0 821 return 0;
michael@0 822 if (sizeof(JCOEF) != 2)
michael@0 823 return 0;
michael@0 824 if (sizeof(FAST_FLOAT) != 4)
michael@0 825 return 0;
michael@0 826
michael@0 827 if (simd_support & JSIMD_SSE2)
michael@0 828 return 1;
michael@0 829 if (simd_support & JSIMD_SSE)
michael@0 830 return 1;
michael@0 831 if (simd_support & JSIMD_3DNOW)
michael@0 832 return 1;
michael@0 833
michael@0 834 return 0;
michael@0 835 }
michael@0 836
michael@0 837 GLOBAL(void)
michael@0 838 jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
michael@0 839 DCTELEM * workspace)
michael@0 840 {
michael@0 841 if (simd_support & JSIMD_SSE2)
michael@0 842 jsimd_quantize_sse2(coef_block, divisors, workspace);
michael@0 843 else if (simd_support & JSIMD_MMX)
michael@0 844 jsimd_quantize_mmx(coef_block, divisors, workspace);
michael@0 845 }
michael@0 846
michael@0 847 GLOBAL(void)
michael@0 848 jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
michael@0 849 FAST_FLOAT * workspace)
michael@0 850 {
michael@0 851 if (simd_support & JSIMD_SSE2)
michael@0 852 jsimd_quantize_float_sse2(coef_block, divisors, workspace);
michael@0 853 else if (simd_support & JSIMD_SSE)
michael@0 854 jsimd_quantize_float_sse(coef_block, divisors, workspace);
michael@0 855 else if (simd_support & JSIMD_3DNOW)
michael@0 856 jsimd_quantize_float_3dnow(coef_block, divisors, workspace);
michael@0 857 }
michael@0 858
michael@0 859 GLOBAL(int)
michael@0 860 jsimd_can_idct_2x2 (void)
michael@0 861 {
michael@0 862 init_simd();
michael@0 863
michael@0 864 /* The code is optimised for these values only */
michael@0 865 if (DCTSIZE != 8)
michael@0 866 return 0;
michael@0 867 if (sizeof(JCOEF) != 2)
michael@0 868 return 0;
michael@0 869 if (BITS_IN_JSAMPLE != 8)
michael@0 870 return 0;
michael@0 871 if (sizeof(JDIMENSION) != 4)
michael@0 872 return 0;
michael@0 873 if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0 874 return 0;
michael@0 875
michael@0 876 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0 877 return 1;
michael@0 878 if (simd_support & JSIMD_MMX)
michael@0 879 return 1;
michael@0 880
michael@0 881 return 0;
michael@0 882 }
michael@0 883
michael@0 884 GLOBAL(int)
michael@0 885 jsimd_can_idct_4x4 (void)
michael@0 886 {
michael@0 887 init_simd();
michael@0 888
michael@0 889 /* The code is optimised for these values only */
michael@0 890 if (DCTSIZE != 8)
michael@0 891 return 0;
michael@0 892 if (sizeof(JCOEF) != 2)
michael@0 893 return 0;
michael@0 894 if (BITS_IN_JSAMPLE != 8)
michael@0 895 return 0;
michael@0 896 if (sizeof(JDIMENSION) != 4)
michael@0 897 return 0;
michael@0 898 if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0 899 return 0;
michael@0 900
michael@0 901 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0 902 return 1;
michael@0 903 if (simd_support & JSIMD_MMX)
michael@0 904 return 1;
michael@0 905
michael@0 906 return 0;
michael@0 907 }
michael@0 908
michael@0 909 GLOBAL(void)
michael@0 910 jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0 911 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0 912 JDIMENSION output_col)
michael@0 913 {
michael@0 914 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0 915 jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 916 else if (simd_support & JSIMD_MMX)
michael@0 917 jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 918 }
michael@0 919
michael@0 920 GLOBAL(void)
michael@0 921 jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0 922 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0 923 JDIMENSION output_col)
michael@0 924 {
michael@0 925 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
michael@0 926 jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 927 else if (simd_support & JSIMD_MMX)
michael@0 928 jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 929 }
michael@0 930
michael@0 931 GLOBAL(int)
michael@0 932 jsimd_can_idct_islow (void)
michael@0 933 {
michael@0 934 init_simd();
michael@0 935
michael@0 936 /* The code is optimised for these values only */
michael@0 937 if (DCTSIZE != 8)
michael@0 938 return 0;
michael@0 939 if (sizeof(JCOEF) != 2)
michael@0 940 return 0;
michael@0 941 if (BITS_IN_JSAMPLE != 8)
michael@0 942 return 0;
michael@0 943 if (sizeof(JDIMENSION) != 4)
michael@0 944 return 0;
michael@0 945 if (sizeof(ISLOW_MULT_TYPE) != 2)
michael@0 946 return 0;
michael@0 947
michael@0 948 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
michael@0 949 return 1;
michael@0 950 if (simd_support & JSIMD_MMX)
michael@0 951 return 1;
michael@0 952
michael@0 953 return 0;
michael@0 954 }
michael@0 955
michael@0 956 GLOBAL(int)
michael@0 957 jsimd_can_idct_ifast (void)
michael@0 958 {
michael@0 959 init_simd();
michael@0 960
michael@0 961 /* The code is optimised for these values only */
michael@0 962 if (DCTSIZE != 8)
michael@0 963 return 0;
michael@0 964 if (sizeof(JCOEF) != 2)
michael@0 965 return 0;
michael@0 966 if (BITS_IN_JSAMPLE != 8)
michael@0 967 return 0;
michael@0 968 if (sizeof(JDIMENSION) != 4)
michael@0 969 return 0;
michael@0 970 if (sizeof(IFAST_MULT_TYPE) != 2)
michael@0 971 return 0;
michael@0 972 if (IFAST_SCALE_BITS != 2)
michael@0 973 return 0;
michael@0 974
michael@0 975 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
michael@0 976 return 1;
michael@0 977 if (simd_support & JSIMD_MMX)
michael@0 978 return 1;
michael@0 979
michael@0 980 return 0;
michael@0 981 }
michael@0 982
michael@0 983 GLOBAL(int)
michael@0 984 jsimd_can_idct_float (void)
michael@0 985 {
michael@0 986 init_simd();
michael@0 987
michael@0 988 if (DCTSIZE != 8)
michael@0 989 return 0;
michael@0 990 if (sizeof(JCOEF) != 2)
michael@0 991 return 0;
michael@0 992 if (BITS_IN_JSAMPLE != 8)
michael@0 993 return 0;
michael@0 994 if (sizeof(JDIMENSION) != 4)
michael@0 995 return 0;
michael@0 996 if (sizeof(FAST_FLOAT) != 4)
michael@0 997 return 0;
michael@0 998 if (sizeof(FLOAT_MULT_TYPE) != 4)
michael@0 999 return 0;
michael@0 1000
michael@0 1001 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
michael@0 1002 return 1;
michael@0 1003 if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
michael@0 1004 return 1;
michael@0 1005 if (simd_support & JSIMD_3DNOW)
michael@0 1006 return 1;
michael@0 1007
michael@0 1008 return 0;
michael@0 1009 }
michael@0 1010
michael@0 1011 GLOBAL(void)
michael@0 1012 jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0 1013 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0 1014 JDIMENSION output_col)
michael@0 1015 {
michael@0 1016 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
michael@0 1017 jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 1018 else if (simd_support & JSIMD_MMX)
michael@0 1019 jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 1020 }
michael@0 1021
michael@0 1022 GLOBAL(void)
michael@0 1023 jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0 1024 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0 1025 JDIMENSION output_col)
michael@0 1026 {
michael@0 1027 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
michael@0 1028 jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 1029 else if (simd_support & JSIMD_MMX)
michael@0 1030 jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col);
michael@0 1031 }
michael@0 1032
michael@0 1033 GLOBAL(void)
michael@0 1034 jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
michael@0 1035 JCOEFPTR coef_block, JSAMPARRAY output_buf,
michael@0 1036 JDIMENSION output_col)
michael@0 1037 {
michael@0 1038 if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
michael@0 1039 jsimd_idct_float_sse2(compptr->dct_table, coef_block,
michael@0 1040 output_buf, output_col);
michael@0 1041 else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
michael@0 1042 jsimd_idct_float_sse(compptr->dct_table, coef_block,
michael@0 1043 output_buf, output_col);
michael@0 1044 else if (simd_support & JSIMD_3DNOW)
michael@0 1045 jsimd_idct_float_3dnow(compptr->dct_table, coef_block,
michael@0 1046 output_buf, output_col);
michael@0 1047 }
michael@0 1048

Mercurial Repository: The Tor Browser

Europalab SCM Repository Server

mercurial