OLD | NEW |
(Empty) | |
| 1 /* |
| 2 * jsimd_mips.c |
| 3 * |
| 4 * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
| 5 * Copyright 2009-2011, 2014 D. R. Commander |
| 6 * Copyright (C) 2013-2014, MIPS Technologies, Inc., California |
| 7 * Copyright 2015 Matthieu Darbois |
| 8 * |
| 9 * Based on the x86 SIMD extension for IJG JPEG library, |
| 10 * Copyright (C) 1999-2006, MIYASAKA Masaru. |
| 11 * For conditions of distribution and use, see copyright notice in jsimdext.inc |
| 12 * |
| 13 * This file contains the interface between the "normal" portions |
| 14 * of the library and the SIMD implementations when running on a |
| 15 * MIPS architecture. |
| 16 */ |
| 17 |
| 18 #define JPEG_INTERNALS |
| 19 #include "../jinclude.h" |
| 20 #include "../jpeglib.h" |
| 21 #include "../jsimd.h" |
| 22 #include "../jdct.h" |
| 23 #include "../jsimddct.h" |
| 24 #include "jsimd.h" |
| 25 |
| 26 #include <stdio.h> |
| 27 #include <string.h> |
| 28 #include <ctype.h> |
| 29 |
| 30 static unsigned int simd_support = ~0; |
| 31 |
| 32 #if defined(__linux__) |
| 33 |
| 34 LOCAL(int) |
| 35 parse_proc_cpuinfo(const char* search_string) |
| 36 { |
| 37 const char* file_name = "/proc/cpuinfo"; |
| 38 char cpuinfo_line[256]; |
| 39 FILE* f = NULL; |
| 40 simd_support = 0; |
| 41 |
| 42 if ((f = fopen(file_name, "r")) != NULL) { |
| 43 while (fgets(cpuinfo_line, sizeof(cpuinfo_line), f) != NULL) { |
| 44 if (strstr(cpuinfo_line, search_string) != NULL) { |
| 45 fclose(f); |
| 46 simd_support |= JSIMD_MIPS_DSPR2; |
| 47 return 1; |
| 48 } |
| 49 } |
| 50 fclose(f); |
| 51 } |
| 52 /* Did not find string in the proc file, or not Linux ELF. */ |
| 53 return 0; |
| 54 } |
| 55 |
| 56 #endif |
| 57 |
| 58 /* |
| 59 * Check what SIMD accelerations are supported. |
| 60 * |
| 61 * FIXME: This code is racy under a multi-threaded environment. |
| 62 */ |
| 63 LOCAL(void) |
| 64 init_simd (void) |
| 65 { |
| 66 if (simd_support != ~0U) |
| 67 return; |
| 68 |
| 69 simd_support = 0; |
| 70 |
| 71 #if defined(__MIPSEL__) && defined(__mips_dsp) && (__mips_dsp_rev >= 2) |
| 72 simd_support |= JSIMD_MIPS_DSPR2; |
| 73 #elif defined(__linux__) |
| 74 /* We still have a chance to use MIPS DSPR2 regardless of globally used |
| 75 * -mdspr2 options passed to gcc by performing runtime detection via |
| 76 * /proc/cpuinfo parsing on linux */ |
| 77 if (!parse_proc_cpuinfo("MIPS 74K")) |
| 78 return; |
| 79 #endif |
| 80 } |
| 81 |
| 82 static const int mips_idct_ifast_coefs[4] = { |
| 83 0x45404540, // FIX( 1.082392200 / 2) = 17734 = 0x4546 |
| 84 0x5A805A80, // FIX( 1.414213562 / 2) = 23170 = 0x5A82 |
| 85 0x76407640, // FIX( 1.847759065 / 2) = 30274 = 0x7642 |
| 86 0xAC60AC60 // FIX(-2.613125930 / 4) = -21407 = 0xAC61 |
| 87 }; |
| 88 |
| 89 /* The following struct is borrowed from jdsample.c */ |
| 90 typedef void (*upsample1_ptr) (j_decompress_ptr cinfo, |
| 91 jpeg_component_info *compptr, |
| 92 JSAMPARRAY input_data, |
| 93 JSAMPARRAY *output_data_ptr); |
| 94 |
| 95 typedef struct { |
| 96 struct jpeg_upsampler pub; |
| 97 JSAMPARRAY color_buf[MAX_COMPONENTS]; |
| 98 upsample1_ptr methods[MAX_COMPONENTS]; |
| 99 int next_row_out; |
| 100 JDIMENSION rows_to_go; |
| 101 int rowgroup_height[MAX_COMPONENTS]; |
| 102 UINT8 h_expand[MAX_COMPONENTS]; |
| 103 UINT8 v_expand[MAX_COMPONENTS]; |
| 104 } my_upsampler; |
| 105 |
| 106 typedef my_upsampler *my_upsample_ptr; |
| 107 |
| 108 GLOBAL(int) |
| 109 jsimd_can_rgb_ycc (void) |
| 110 { |
| 111 init_simd(); |
| 112 |
| 113 /* The code is optimised for these values only */ |
| 114 if (BITS_IN_JSAMPLE != 8) |
| 115 return 0; |
| 116 if (sizeof(JDIMENSION) != 4) |
| 117 return 0; |
| 118 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
| 119 return 0; |
| 120 |
| 121 if (simd_support & JSIMD_MIPS_DSPR2) |
| 122 return 1; |
| 123 |
| 124 return 0; |
| 125 } |
| 126 |
| 127 GLOBAL(int) |
| 128 jsimd_can_rgb_gray (void) |
| 129 { |
| 130 init_simd(); |
| 131 |
| 132 /* The code is optimised for these values only */ |
| 133 if (BITS_IN_JSAMPLE != 8) |
| 134 return 0; |
| 135 if (sizeof(JDIMENSION) != 4) |
| 136 return 0; |
| 137 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
| 138 return 0; |
| 139 |
| 140 if (simd_support & JSIMD_MIPS_DSPR2) |
| 141 return 1; |
| 142 |
| 143 return 0; |
| 144 } |
| 145 |
| 146 GLOBAL(int) |
| 147 jsimd_can_ycc_rgb (void) |
| 148 { |
| 149 init_simd(); |
| 150 |
| 151 /* The code is optimised for these values only */ |
| 152 if (BITS_IN_JSAMPLE != 8) |
| 153 return 0; |
| 154 if (sizeof(JDIMENSION) != 4) |
| 155 return 0; |
| 156 if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
| 157 return 0; |
| 158 |
| 159 if (simd_support & JSIMD_MIPS_DSPR2) |
| 160 return 1; |
| 161 |
| 162 return 0; |
| 163 } |
| 164 |
| 165 GLOBAL(int) |
| 166 jsimd_can_ycc_rgb565 (void) |
| 167 { |
| 168 return 0; |
| 169 } |
| 170 |
| 171 GLOBAL(int) |
| 172 jsimd_c_can_null_convert (void) |
| 173 { |
| 174 init_simd(); |
| 175 |
| 176 /* The code is optimised for these values only */ |
| 177 if (BITS_IN_JSAMPLE != 8) |
| 178 return 0; |
| 179 if (sizeof(JDIMENSION) != 4) |
| 180 return 0; |
| 181 |
| 182 if (simd_support & JSIMD_MIPS_DSPR2) |
| 183 return 1; |
| 184 |
| 185 return 0; |
| 186 } |
| 187 |
| 188 GLOBAL(void) |
| 189 jsimd_rgb_ycc_convert (j_compress_ptr cinfo, |
| 190 JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
| 191 JDIMENSION output_row, int num_rows) |
| 192 { |
| 193 void (*mipsdspr2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
| 194 |
| 195 switch(cinfo->in_color_space) { |
| 196 case JCS_EXT_RGB: |
| 197 mipsdspr2fct=jsimd_extrgb_ycc_convert_mips_dspr2; |
| 198 break; |
| 199 case JCS_EXT_RGBX: |
| 200 case JCS_EXT_RGBA: |
| 201 mipsdspr2fct=jsimd_extrgbx_ycc_convert_mips_dspr2; |
| 202 break; |
| 203 case JCS_EXT_BGR: |
| 204 mipsdspr2fct=jsimd_extbgr_ycc_convert_mips_dspr2; |
| 205 break; |
| 206 case JCS_EXT_BGRX: |
| 207 case JCS_EXT_BGRA: |
| 208 mipsdspr2fct=jsimd_extbgrx_ycc_convert_mips_dspr2; |
| 209 break; |
| 210 case JCS_EXT_XBGR: |
| 211 case JCS_EXT_ABGR: |
| 212 mipsdspr2fct=jsimd_extxbgr_ycc_convert_mips_dspr2; |
| 213 |
| 214 break; |
| 215 case JCS_EXT_XRGB: |
| 216 case JCS_EXT_ARGB: |
| 217 mipsdspr2fct=jsimd_extxrgb_ycc_convert_mips_dspr2; |
| 218 break; |
| 219 default: |
| 220 mipsdspr2fct=jsimd_extrgb_ycc_convert_mips_dspr2; |
| 221 break; |
| 222 } |
| 223 |
| 224 if (simd_support & JSIMD_MIPS_DSPR2) |
| 225 mipsdspr2fct(cinfo->image_width, input_buf, output_buf, output_row, |
| 226 num_rows); |
| 227 } |
| 228 |
| 229 GLOBAL(void) |
| 230 jsimd_rgb_gray_convert (j_compress_ptr cinfo, |
| 231 JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
| 232 JDIMENSION output_row, int num_rows) |
| 233 { |
| 234 void (*mipsdspr2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
| 235 |
| 236 switch(cinfo->in_color_space) { |
| 237 case JCS_EXT_RGB: |
| 238 mipsdspr2fct=jsimd_extrgb_gray_convert_mips_dspr2; |
| 239 break; |
| 240 case JCS_EXT_RGBX: |
| 241 case JCS_EXT_RGBA: |
| 242 mipsdspr2fct=jsimd_extrgbx_gray_convert_mips_dspr2; |
| 243 break; |
| 244 case JCS_EXT_BGR: |
| 245 mipsdspr2fct=jsimd_extbgr_gray_convert_mips_dspr2; |
| 246 break; |
| 247 case JCS_EXT_BGRX: |
| 248 case JCS_EXT_BGRA: |
| 249 mipsdspr2fct=jsimd_extbgrx_gray_convert_mips_dspr2; |
| 250 break; |
| 251 case JCS_EXT_XBGR: |
| 252 case JCS_EXT_ABGR: |
| 253 mipsdspr2fct=jsimd_extxbgr_gray_convert_mips_dspr2; |
| 254 break; |
| 255 case JCS_EXT_XRGB: |
| 256 case JCS_EXT_ARGB: |
| 257 mipsdspr2fct=jsimd_extxrgb_gray_convert_mips_dspr2; |
| 258 break; |
| 259 default: |
| 260 mipsdspr2fct=jsimd_extrgb_gray_convert_mips_dspr2; |
| 261 break; |
| 262 } |
| 263 |
| 264 if (simd_support & JSIMD_MIPS_DSPR2) |
| 265 mipsdspr2fct(cinfo->image_width, input_buf, output_buf, output_row, |
| 266 num_rows); |
| 267 } |
| 268 |
| 269 GLOBAL(void) |
| 270 jsimd_ycc_rgb_convert (j_decompress_ptr cinfo, |
| 271 JSAMPIMAGE input_buf, JDIMENSION input_row, |
| 272 JSAMPARRAY output_buf, int num_rows) |
| 273 { |
| 274 void (*mipsdspr2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); |
| 275 |
| 276 switch(cinfo->out_color_space) { |
| 277 case JCS_EXT_RGB: |
| 278 mipsdspr2fct=jsimd_ycc_extrgb_convert_mips_dspr2; |
| 279 break; |
| 280 case JCS_EXT_RGBX: |
| 281 case JCS_EXT_RGBA: |
| 282 mipsdspr2fct=jsimd_ycc_extrgbx_convert_mips_dspr2; |
| 283 break; |
| 284 case JCS_EXT_BGR: |
| 285 mipsdspr2fct=jsimd_ycc_extbgr_convert_mips_dspr2; |
| 286 break; |
| 287 case JCS_EXT_BGRX: |
| 288 case JCS_EXT_BGRA: |
| 289 mipsdspr2fct=jsimd_ycc_extbgrx_convert_mips_dspr2; |
| 290 break; |
| 291 case JCS_EXT_XBGR: |
| 292 case JCS_EXT_ABGR: |
| 293 mipsdspr2fct=jsimd_ycc_extxbgr_convert_mips_dspr2; |
| 294 break; |
| 295 case JCS_EXT_XRGB: |
| 296 case JCS_EXT_ARGB: |
| 297 mipsdspr2fct=jsimd_ycc_extxrgb_convert_mips_dspr2; |
| 298 break; |
| 299 default: |
| 300 mipsdspr2fct=jsimd_ycc_extrgb_convert_mips_dspr2; |
| 301 break; |
| 302 } |
| 303 |
| 304 if (simd_support & JSIMD_MIPS_DSPR2) |
| 305 mipsdspr2fct(cinfo->output_width, input_buf, input_row, output_buf, |
| 306 num_rows); |
| 307 } |
| 308 |
| 309 GLOBAL(void) |
| 310 jsimd_ycc_rgb565_convert (j_decompress_ptr cinfo, |
| 311 JSAMPIMAGE input_buf, JDIMENSION input_row, |
| 312 JSAMPARRAY output_buf, int num_rows) |
| 313 { |
| 314 } |
| 315 |
| 316 GLOBAL(void) |
| 317 jsimd_c_null_convert (j_compress_ptr cinfo, |
| 318 JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
| 319 JDIMENSION output_row, int num_rows) |
| 320 { |
| 321 if (simd_support & JSIMD_MIPS_DSPR2) |
| 322 jsimd_c_null_convert_mips_dspr2(cinfo->image_width, input_buf, |
| 323 output_buf, output_row, num_rows, |
| 324 cinfo->num_components); |
| 325 } |
| 326 |
| 327 GLOBAL(int) |
| 328 jsimd_can_h2v2_downsample (void) |
| 329 { |
| 330 init_simd(); |
| 331 |
| 332 /* The code is optimised for these values only */ |
| 333 if (BITS_IN_JSAMPLE != 8) |
| 334 return 0; |
| 335 if (sizeof(JDIMENSION) != 4) |
| 336 return 0; |
| 337 |
| 338 if (simd_support & JSIMD_MIPS_DSPR2) |
| 339 return 1; |
| 340 |
| 341 return 0; |
| 342 } |
| 343 |
| 344 GLOBAL(int) |
| 345 jsimd_can_h2v2_smooth_downsample (void) |
| 346 { |
| 347 init_simd(); |
| 348 |
| 349 /* The code is optimised for these values only */ |
| 350 if (BITS_IN_JSAMPLE != 8) |
| 351 return 0; |
| 352 if (sizeof(JDIMENSION) != 4) |
| 353 return 0; |
| 354 if(DCTSIZE != 8) |
| 355 return 0; |
| 356 |
| 357 if (simd_support & JSIMD_MIPS_DSPR2) |
| 358 return 1; |
| 359 |
| 360 return 0; |
| 361 } |
| 362 |
| 363 GLOBAL(int) |
| 364 jsimd_can_h2v1_downsample (void) |
| 365 { |
| 366 init_simd(); |
| 367 |
| 368 /* The code is optimised for these values only */ |
| 369 if (BITS_IN_JSAMPLE != 8) |
| 370 return 0; |
| 371 if (sizeof(JDIMENSION) != 4) |
| 372 return 0; |
| 373 |
| 374 if (simd_support & JSIMD_MIPS_DSPR2) |
| 375 return 1; |
| 376 |
| 377 return 0; |
| 378 } |
| 379 |
| 380 GLOBAL(void) |
| 381 jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr, |
| 382 JSAMPARRAY input_data, JSAMPARRAY output_data) |
| 383 { |
| 384 if (simd_support & JSIMD_MIPS_DSPR2) |
| 385 jsimd_h2v2_downsample_mips_dspr2(cinfo->image_width, |
| 386 cinfo->max_v_samp_factor, |
| 387 compptr->v_samp_factor, |
| 388 compptr->width_in_blocks, input_data, |
| 389 output_data); |
| 390 } |
| 391 |
| 392 GLOBAL(void) |
| 393 jsimd_h2v2_smooth_downsample (j_compress_ptr cinfo, |
| 394 jpeg_component_info *compptr, |
| 395 JSAMPARRAY input_data, JSAMPARRAY output_data) |
| 396 { |
| 397 jsimd_h2v2_smooth_downsample_mips_dspr2(input_data, output_data, |
| 398 compptr->v_samp_factor, |
| 399 cinfo->max_v_samp_factor, |
| 400 cinfo->smoothing_factor, |
| 401 compptr->width_in_blocks, |
| 402 cinfo->image_width); |
| 403 } |
| 404 |
| 405 GLOBAL(void) |
| 406 jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info *compptr, |
| 407 JSAMPARRAY input_data, JSAMPARRAY output_data) |
| 408 { |
| 409 if (simd_support & JSIMD_MIPS_DSPR2) |
| 410 jsimd_h2v1_downsample_mips_dspr2(cinfo->image_width, |
| 411 cinfo->max_v_samp_factor, |
| 412 compptr->v_samp_factor, |
| 413 compptr->width_in_blocks, |
| 414 input_data, output_data); |
| 415 } |
| 416 |
| 417 GLOBAL(int) |
| 418 jsimd_can_h2v2_upsample (void) |
| 419 { |
| 420 init_simd(); |
| 421 |
| 422 /* The code is optimised for these values only */ |
| 423 if (BITS_IN_JSAMPLE != 8) |
| 424 return 0; |
| 425 if (sizeof(JDIMENSION) != 4) |
| 426 return 0; |
| 427 |
| 428 if (simd_support & JSIMD_MIPS_DSPR2) |
| 429 return 1; |
| 430 |
| 431 return 0; |
| 432 } |
| 433 |
| 434 GLOBAL(int) |
| 435 jsimd_can_h2v1_upsample (void) |
| 436 { |
| 437 init_simd(); |
| 438 |
| 439 /* The code is optimised for these values only */ |
| 440 if (BITS_IN_JSAMPLE != 8) |
| 441 return 0; |
| 442 if (sizeof(JDIMENSION) != 4) |
| 443 return 0; |
| 444 |
| 445 if (simd_support & JSIMD_MIPS_DSPR2) |
| 446 return 1; |
| 447 |
| 448 return 0; |
| 449 } |
| 450 |
| 451 GLOBAL(int) |
| 452 jsimd_can_int_upsample (void) |
| 453 { |
| 454 init_simd(); |
| 455 |
| 456 /* The code is optimised for these values only */ |
| 457 if (BITS_IN_JSAMPLE != 8) |
| 458 return 0; |
| 459 if (sizeof(JDIMENSION) != 4) |
| 460 return 0; |
| 461 |
| 462 if (simd_support & JSIMD_MIPS_DSPR2) |
| 463 return 1; |
| 464 |
| 465 return 0; |
| 466 } |
| 467 |
| 468 GLOBAL(void) |
| 469 jsimd_h2v2_upsample (j_decompress_ptr cinfo, |
| 470 jpeg_component_info *compptr, |
| 471 JSAMPARRAY input_data, |
| 472 JSAMPARRAY *output_data_ptr) |
| 473 { |
| 474 if (simd_support & JSIMD_MIPS_DSPR2) |
| 475 jsimd_h2v2_upsample_mips_dspr2(cinfo->max_v_samp_factor, |
| 476 cinfo->output_width, input_data, |
| 477 output_data_ptr); |
| 478 } |
| 479 |
| 480 GLOBAL(void) |
| 481 jsimd_h2v1_upsample (j_decompress_ptr cinfo, |
| 482 jpeg_component_info *compptr, |
| 483 JSAMPARRAY input_data, |
| 484 JSAMPARRAY *output_data_ptr) |
| 485 { |
| 486 if (simd_support & JSIMD_MIPS_DSPR2) |
| 487 jsimd_h2v1_upsample_mips_dspr2(cinfo->max_v_samp_factor, |
| 488 cinfo->output_width, input_data, |
| 489 output_data_ptr); |
| 490 } |
| 491 |
| 492 GLOBAL(void) |
| 493 jsimd_int_upsample (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 494 JSAMPARRAY input_data, JSAMPARRAY *output_data_ptr) |
| 495 { |
| 496 my_upsample_ptr upsample = (my_upsample_ptr) cinfo->upsample; |
| 497 |
| 498 jsimd_int_upsample_mips_dspr2(upsample->h_expand[compptr->component_index], |
| 499 upsample->v_expand[compptr->component_index], |
| 500 input_data, output_data_ptr, |
| 501 cinfo->output_width, |
| 502 cinfo->max_v_samp_factor); |
| 503 } |
| 504 |
| 505 GLOBAL(int) |
| 506 jsimd_can_h2v2_fancy_upsample (void) |
| 507 { |
| 508 init_simd(); |
| 509 |
| 510 /* The code is optimised for these values only */ |
| 511 if (BITS_IN_JSAMPLE != 8) |
| 512 return 0; |
| 513 if (sizeof(JDIMENSION) != 4) |
| 514 return 0; |
| 515 |
| 516 if (simd_support & JSIMD_MIPS_DSPR2) |
| 517 return 1; |
| 518 |
| 519 return 0; |
| 520 } |
| 521 |
| 522 GLOBAL(int) |
| 523 jsimd_can_h2v1_fancy_upsample (void) |
| 524 { |
| 525 init_simd(); |
| 526 |
| 527 /* The code is optimised for these values only */ |
| 528 if (BITS_IN_JSAMPLE != 8) |
| 529 return 0; |
| 530 if (sizeof(JDIMENSION) != 4) |
| 531 return 0; |
| 532 |
| 533 if (simd_support & JSIMD_MIPS_DSPR2) |
| 534 return 1; |
| 535 |
| 536 return 0; |
| 537 } |
| 538 |
| 539 GLOBAL(void) |
| 540 jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo, |
| 541 jpeg_component_info *compptr, |
| 542 JSAMPARRAY input_data, |
| 543 JSAMPARRAY *output_data_ptr) |
| 544 { |
| 545 if (simd_support & JSIMD_MIPS_DSPR2) |
| 546 jsimd_h2v2_fancy_upsample_mips_dspr2(cinfo->max_v_samp_factor, |
| 547 compptr->downsampled_width, |
| 548 input_data, output_data_ptr); |
| 549 } |
| 550 |
| 551 GLOBAL(void) |
| 552 jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo, |
| 553 jpeg_component_info *compptr, |
| 554 JSAMPARRAY input_data, |
| 555 JSAMPARRAY *output_data_ptr) |
| 556 { |
| 557 if (simd_support & JSIMD_MIPS_DSPR2) |
| 558 jsimd_h2v1_fancy_upsample_mips_dspr2(cinfo->max_v_samp_factor, |
| 559 compptr->downsampled_width, |
| 560 input_data, output_data_ptr); |
| 561 } |
| 562 |
| 563 GLOBAL(int) |
| 564 jsimd_can_h2v2_merged_upsample (void) |
| 565 { |
| 566 init_simd(); |
| 567 |
| 568 if (BITS_IN_JSAMPLE != 8) |
| 569 return 0; |
| 570 if (sizeof(JDIMENSION) != 4) |
| 571 return 0; |
| 572 |
| 573 if (simd_support & JSIMD_MIPS_DSPR2) |
| 574 return 1; |
| 575 |
| 576 return 0; |
| 577 } |
| 578 |
| 579 GLOBAL(int) |
| 580 jsimd_can_h2v1_merged_upsample (void) |
| 581 { |
| 582 init_simd(); |
| 583 |
| 584 if (BITS_IN_JSAMPLE != 8) |
| 585 return 0; |
| 586 if (sizeof(JDIMENSION) != 4) |
| 587 return 0; |
| 588 |
| 589 if (simd_support & JSIMD_MIPS_DSPR2) |
| 590 return 1; |
| 591 |
| 592 return 0; |
| 593 } |
| 594 |
| 595 GLOBAL(void) |
| 596 jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo, |
| 597 JSAMPIMAGE input_buf, |
| 598 JDIMENSION in_row_group_ctr, |
| 599 JSAMPARRAY output_buf) |
| 600 { |
| 601 void (*mipsdspr2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, |
| 602 JSAMPLE *); |
| 603 |
| 604 switch(cinfo->out_color_space) { |
| 605 case JCS_EXT_RGB: |
| 606 mipsdspr2fct=jsimd_h2v2_extrgb_merged_upsample_mips_dspr2; |
| 607 break; |
| 608 case JCS_EXT_RGBX: |
| 609 case JCS_EXT_RGBA: |
| 610 mipsdspr2fct=jsimd_h2v2_extrgbx_merged_upsample_mips_dspr2; |
| 611 break; |
| 612 case JCS_EXT_BGR: |
| 613 mipsdspr2fct=jsimd_h2v2_extbgr_merged_upsample_mips_dspr2; |
| 614 break; |
| 615 case JCS_EXT_BGRX: |
| 616 case JCS_EXT_BGRA: |
| 617 mipsdspr2fct=jsimd_h2v2_extbgrx_merged_upsample_mips_dspr2; |
| 618 break; |
| 619 case JCS_EXT_XBGR: |
| 620 case JCS_EXT_ABGR: |
| 621 mipsdspr2fct=jsimd_h2v2_extxbgr_merged_upsample_mips_dspr2; |
| 622 break; |
| 623 case JCS_EXT_XRGB: |
| 624 case JCS_EXT_ARGB: |
| 625 mipsdspr2fct=jsimd_h2v2_extxrgb_merged_upsample_mips_dspr2; |
| 626 break; |
| 627 default: |
| 628 mipsdspr2fct=jsimd_h2v2_extrgb_merged_upsample_mips_dspr2; |
| 629 break; |
| 630 } |
| 631 |
| 632 mipsdspr2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf, |
| 633 cinfo->sample_range_limit); |
| 634 } |
| 635 |
| 636 GLOBAL(void) |
| 637 jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo, |
| 638 JSAMPIMAGE input_buf, |
| 639 JDIMENSION in_row_group_ctr, |
| 640 JSAMPARRAY output_buf) |
| 641 { |
| 642 void (*mipsdspr2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, |
| 643 JSAMPLE *); |
| 644 |
| 645 switch(cinfo->out_color_space) { |
| 646 case JCS_EXT_RGB: |
| 647 mipsdspr2fct=jsimd_h2v1_extrgb_merged_upsample_mips_dspr2; |
| 648 break; |
| 649 case JCS_EXT_RGBX: |
| 650 case JCS_EXT_RGBA: |
| 651 mipsdspr2fct=jsimd_h2v1_extrgbx_merged_upsample_mips_dspr2; |
| 652 break; |
| 653 case JCS_EXT_BGR: |
| 654 mipsdspr2fct=jsimd_h2v1_extbgr_merged_upsample_mips_dspr2; |
| 655 break; |
| 656 case JCS_EXT_BGRX: |
| 657 case JCS_EXT_BGRA: |
| 658 mipsdspr2fct=jsimd_h2v1_extbgrx_merged_upsample_mips_dspr2; |
| 659 break; |
| 660 case JCS_EXT_XBGR: |
| 661 case JCS_EXT_ABGR: |
| 662 mipsdspr2fct=jsimd_h2v1_extxbgr_merged_upsample_mips_dspr2; |
| 663 break; |
| 664 case JCS_EXT_XRGB: |
| 665 case JCS_EXT_ARGB: |
| 666 mipsdspr2fct=jsimd_h2v1_extxrgb_merged_upsample_mips_dspr2; |
| 667 break; |
| 668 default: |
| 669 mipsdspr2fct=jsimd_h2v1_extrgb_merged_upsample_mips_dspr2; |
| 670 break; |
| 671 } |
| 672 |
| 673 mipsdspr2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf, |
| 674 cinfo->sample_range_limit); |
| 675 } |
| 676 |
| 677 GLOBAL(int) |
| 678 jsimd_can_convsamp (void) |
| 679 { |
| 680 init_simd(); |
| 681 |
| 682 /* The code is optimised for these values only */ |
| 683 if (DCTSIZE != 8) |
| 684 return 0; |
| 685 if (BITS_IN_JSAMPLE != 8) |
| 686 return 0; |
| 687 if (sizeof(JDIMENSION) != 4) |
| 688 return 0; |
| 689 if (sizeof(DCTELEM) != 2) |
| 690 return 0; |
| 691 |
| 692 if (simd_support & JSIMD_MIPS_DSPR2) |
| 693 return 1; |
| 694 |
| 695 return 0; |
| 696 } |
| 697 |
| 698 GLOBAL(int) |
| 699 jsimd_can_convsamp_float (void) |
| 700 { |
| 701 init_simd(); |
| 702 |
| 703 /* The code is optimised for these values only */ |
| 704 if (DCTSIZE != 8) |
| 705 return 0; |
| 706 if (sizeof(JCOEF) != 2) |
| 707 return 0; |
| 708 if (BITS_IN_JSAMPLE != 8) |
| 709 return 0; |
| 710 if (sizeof(JDIMENSION) != 4) |
| 711 return 0; |
| 712 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 713 return 0; |
| 714 |
| 715 if (simd_support & JSIMD_MIPS_DSPR2) |
| 716 return 1; |
| 717 |
| 718 return 0; |
| 719 } |
| 720 |
| 721 GLOBAL(void) |
| 722 jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 723 DCTELEM *workspace) |
| 724 { |
| 725 if (simd_support & JSIMD_MIPS_DSPR2) |
| 726 jsimd_convsamp_mips_dspr2(sample_data, start_col, workspace); |
| 727 } |
| 728 |
| 729 GLOBAL(void) |
| 730 jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 731 FAST_FLOAT *workspace) |
| 732 { |
| 733 if ((simd_support & JSIMD_MIPS_DSPR2)) |
| 734 jsimd_convsamp_float_mips_dspr2(sample_data, start_col, workspace); |
| 735 } |
| 736 |
| 737 GLOBAL(int) |
| 738 jsimd_can_fdct_islow (void) |
| 739 { |
| 740 init_simd(); |
| 741 |
| 742 /* The code is optimised for these values only */ |
| 743 if (DCTSIZE != 8) |
| 744 return 0; |
| 745 if (sizeof(DCTELEM) != 2) |
| 746 return 0; |
| 747 |
| 748 if (simd_support & JSIMD_MIPS_DSPR2) |
| 749 return 1; |
| 750 |
| 751 return 0; |
| 752 } |
| 753 |
| 754 GLOBAL(int) |
| 755 jsimd_can_fdct_ifast (void) |
| 756 { |
| 757 init_simd(); |
| 758 |
| 759 /* The code is optimised for these values only */ |
| 760 if (DCTSIZE != 8) |
| 761 return 0; |
| 762 if (sizeof(DCTELEM) != 2) |
| 763 return 0; |
| 764 |
| 765 if (simd_support & JSIMD_MIPS_DSPR2) |
| 766 return 1; |
| 767 |
| 768 return 0; |
| 769 } |
| 770 |
| 771 GLOBAL(int) |
| 772 jsimd_can_fdct_float (void) |
| 773 { |
| 774 init_simd(); |
| 775 |
| 776 return 0; |
| 777 } |
| 778 |
| 779 GLOBAL(void) |
| 780 jsimd_fdct_islow (DCTELEM *data) |
| 781 { |
| 782 if (simd_support & JSIMD_MIPS_DSPR2) |
| 783 jsimd_fdct_islow_mips_dspr2(data); |
| 784 } |
| 785 |
| 786 GLOBAL(void) |
| 787 jsimd_fdct_ifast (DCTELEM *data) |
| 788 { |
| 789 if (simd_support & JSIMD_MIPS_DSPR2) |
| 790 jsimd_fdct_ifast_mips_dspr2(data); |
| 791 } |
| 792 |
| 793 GLOBAL(void) |
| 794 jsimd_fdct_float (FAST_FLOAT *data) |
| 795 { |
| 796 } |
| 797 |
| 798 GLOBAL(int) |
| 799 jsimd_can_quantize (void) |
| 800 { |
| 801 init_simd(); |
| 802 |
| 803 /* The code is optimised for these values only */ |
| 804 if (DCTSIZE != 8) |
| 805 return 0; |
| 806 if (sizeof(JCOEF) != 2) |
| 807 return 0; |
| 808 if (sizeof(DCTELEM) != 2) |
| 809 return 0; |
| 810 |
| 811 if (simd_support & JSIMD_MIPS_DSPR2) |
| 812 return 1; |
| 813 |
| 814 return 0; |
| 815 } |
| 816 |
| 817 GLOBAL(int) |
| 818 jsimd_can_quantize_float (void) |
| 819 { |
| 820 init_simd(); |
| 821 |
| 822 /* The code is optimised for these values only */ |
| 823 if (DCTSIZE != 8) |
| 824 return 0; |
| 825 if (sizeof(JCOEF) != 2) |
| 826 return 0; |
| 827 if (BITS_IN_JSAMPLE != 8) |
| 828 return 0; |
| 829 if (sizeof(JDIMENSION) != 4) |
| 830 return 0; |
| 831 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 832 return 0; |
| 833 |
| 834 if (simd_support & JSIMD_MIPS_DSPR2) |
| 835 return 1; |
| 836 |
| 837 return 0; |
| 838 } |
| 839 |
| 840 GLOBAL(void) |
| 841 jsimd_quantize (JCOEFPTR coef_block, DCTELEM *divisors, |
| 842 DCTELEM *workspace) |
| 843 { |
| 844 if (simd_support & JSIMD_MIPS_DSPR2) |
| 845 jsimd_quantize_mips_dspr2(coef_block, divisors, workspace); |
| 846 } |
| 847 |
| 848 GLOBAL(void) |
| 849 jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT *divisors, |
| 850 FAST_FLOAT *workspace) |
| 851 { |
| 852 if (simd_support & JSIMD_MIPS_DSPR2) |
| 853 jsimd_quantize_float_mips_dspr2(coef_block, divisors, workspace); |
| 854 } |
| 855 |
| 856 GLOBAL(int) |
| 857 jsimd_can_idct_2x2 (void) |
| 858 { |
| 859 init_simd(); |
| 860 |
| 861 /* The code is optimised for these values only */ |
| 862 if (DCTSIZE != 8) |
| 863 return 0; |
| 864 if (sizeof(JCOEF) != 2) |
| 865 return 0; |
| 866 if (BITS_IN_JSAMPLE != 8) |
| 867 return 0; |
| 868 if (sizeof(JDIMENSION) != 4) |
| 869 return 0; |
| 870 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 871 return 0; |
| 872 |
| 873 if (simd_support & JSIMD_MIPS_DSPR2) |
| 874 return 1; |
| 875 |
| 876 return 0; |
| 877 } |
| 878 |
| 879 GLOBAL(int) |
| 880 jsimd_can_idct_4x4 (void) |
| 881 { |
| 882 init_simd(); |
| 883 |
| 884 /* The code is optimised for these values only */ |
| 885 if (DCTSIZE != 8) |
| 886 return 0; |
| 887 if (sizeof(JCOEF) != 2) |
| 888 return 0; |
| 889 if (BITS_IN_JSAMPLE != 8) |
| 890 return 0; |
| 891 if (sizeof(JDIMENSION) != 4) |
| 892 return 0; |
| 893 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 894 return 0; |
| 895 |
| 896 if (simd_support & JSIMD_MIPS_DSPR2) |
| 897 return 1; |
| 898 |
| 899 return 0; |
| 900 } |
| 901 |
| 902 GLOBAL(int) |
| 903 jsimd_can_idct_6x6 (void) |
| 904 { |
| 905 init_simd(); |
| 906 |
| 907 /* The code is optimised for these values only */ |
| 908 if (DCTSIZE != 8) |
| 909 return 0; |
| 910 if (sizeof(JCOEF) != 2) |
| 911 return 0; |
| 912 if (BITS_IN_JSAMPLE != 8) |
| 913 return 0; |
| 914 if (sizeof(JDIMENSION) != 4) |
| 915 return 0; |
| 916 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 917 return 0; |
| 918 |
| 919 if (simd_support & JSIMD_MIPS_DSPR2) |
| 920 return 1; |
| 921 |
| 922 return 0; |
| 923 } |
| 924 |
| 925 GLOBAL(int) |
| 926 jsimd_can_idct_12x12 (void) |
| 927 { |
| 928 init_simd(); |
| 929 |
| 930 if (BITS_IN_JSAMPLE != 8) |
| 931 return 0; |
| 932 if (DCTSIZE != 8) |
| 933 return 0; |
| 934 if (sizeof(JCOEF) != 2) |
| 935 return 0; |
| 936 if (sizeof(JDIMENSION) != 4) |
| 937 return 0; |
| 938 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 939 return 0; |
| 940 |
| 941 if (simd_support & JSIMD_MIPS_DSPR2) |
| 942 return 1; |
| 943 |
| 944 return 0; |
| 945 } |
| 946 |
| 947 GLOBAL(void) |
| 948 jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 949 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 950 JDIMENSION output_col) |
| 951 { |
| 952 if (simd_support & JSIMD_MIPS_DSPR2) |
| 953 jsimd_idct_2x2_mips_dspr2(compptr->dct_table, coef_block, output_buf, |
| 954 output_col); |
| 955 } |
| 956 |
| 957 GLOBAL(void) |
| 958 jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 959 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 960 JDIMENSION output_col) |
| 961 { |
| 962 if (simd_support & JSIMD_MIPS_DSPR2) { |
| 963 int workspace[DCTSIZE*4]; /* buffers data between passes */ |
| 964 jsimd_idct_4x4_mips_dspr2(compptr->dct_table, coef_block, output_buf, |
| 965 output_col, workspace); |
| 966 } |
| 967 } |
| 968 |
| 969 GLOBAL(void) |
| 970 jsimd_idct_6x6 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 971 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 972 JDIMENSION output_col) |
| 973 { |
| 974 if (simd_support & JSIMD_MIPS_DSPR2) |
| 975 jsimd_idct_6x6_mips_dspr2(compptr->dct_table, coef_block, output_buf, |
| 976 output_col); |
| 977 } |
| 978 |
| 979 GLOBAL(void) |
| 980 jsimd_idct_12x12 (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 981 JCOEFPTR coef_block, |
| 982 JSAMPARRAY output_buf, JDIMENSION output_col) |
| 983 { |
| 984 if (simd_support & JSIMD_MIPS_DSPR2) { |
| 985 int workspace[96]; |
| 986 int output[12] = { |
| 987 (int)(output_buf[0] + output_col), |
| 988 (int)(output_buf[1] + output_col), |
| 989 (int)(output_buf[2] + output_col), |
| 990 (int)(output_buf[3] + output_col), |
| 991 (int)(output_buf[4] + output_col), |
| 992 (int)(output_buf[5] + output_col), |
| 993 (int)(output_buf[6] + output_col), |
| 994 (int)(output_buf[7] + output_col), |
| 995 (int)(output_buf[8] + output_col), |
| 996 (int)(output_buf[9] + output_col), |
| 997 (int)(output_buf[10] + output_col), |
| 998 (int)(output_buf[11] + output_col), |
| 999 }; |
| 1000 jsimd_idct_12x12_pass1_mips_dspr2(coef_block, compptr->dct_table, |
| 1001 workspace); |
| 1002 jsimd_idct_12x12_pass2_mips_dspr2(workspace, output); |
| 1003 } |
| 1004 } |
| 1005 |
| 1006 GLOBAL(int) |
| 1007 jsimd_can_idct_islow (void) |
| 1008 { |
| 1009 init_simd(); |
| 1010 |
| 1011 /* The code is optimised for these values only */ |
| 1012 if (DCTSIZE != 8) |
| 1013 return 0; |
| 1014 if (sizeof(JCOEF) != 2) |
| 1015 return 0; |
| 1016 if (BITS_IN_JSAMPLE != 8) |
| 1017 return 0; |
| 1018 if (sizeof(JDIMENSION) != 4) |
| 1019 return 0; |
| 1020 if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 1021 return 0; |
| 1022 |
| 1023 if (simd_support & JSIMD_MIPS_DSPR2) |
| 1024 return 1; |
| 1025 |
| 1026 return 0; |
| 1027 } |
| 1028 |
| 1029 GLOBAL(int) |
| 1030 jsimd_can_idct_ifast (void) |
| 1031 { |
| 1032 init_simd(); |
| 1033 |
| 1034 /* The code is optimised for these values only */ |
| 1035 if (DCTSIZE != 8) |
| 1036 return 0; |
| 1037 if (sizeof(JCOEF) != 2) |
| 1038 return 0; |
| 1039 if (BITS_IN_JSAMPLE != 8) |
| 1040 return 0; |
| 1041 if (sizeof(JDIMENSION) != 4) |
| 1042 return 0; |
| 1043 if (sizeof(IFAST_MULT_TYPE) != 2) |
| 1044 return 0; |
| 1045 if (IFAST_SCALE_BITS != 2) |
| 1046 return 0; |
| 1047 |
| 1048 if (simd_support & JSIMD_MIPS_DSPR2) |
| 1049 return 1; |
| 1050 |
| 1051 return 0; |
| 1052 } |
| 1053 |
| 1054 GLOBAL(int) |
| 1055 jsimd_can_idct_float (void) |
| 1056 { |
| 1057 init_simd(); |
| 1058 |
| 1059 return 0; |
| 1060 } |
| 1061 |
| 1062 GLOBAL(void) |
| 1063 jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 1064 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 1065 JDIMENSION output_col) |
| 1066 { |
| 1067 if (simd_support & JSIMD_MIPS_DSPR2) { |
| 1068 int output[8] = { |
| 1069 (int)(output_buf[0] + output_col), |
| 1070 (int)(output_buf[1] + output_col), |
| 1071 (int)(output_buf[2] + output_col), |
| 1072 (int)(output_buf[3] + output_col), |
| 1073 (int)(output_buf[4] + output_col), |
| 1074 (int)(output_buf[5] + output_col), |
| 1075 (int)(output_buf[6] + output_col), |
| 1076 (int)(output_buf[7] + output_col), |
| 1077 }; |
| 1078 |
| 1079 jsimd_idct_islow_mips_dspr2(coef_block, compptr->dct_table, |
| 1080 output, IDCT_range_limit(cinfo)); |
| 1081 } |
| 1082 } |
| 1083 |
| 1084 GLOBAL(void) |
| 1085 jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 1086 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 1087 JDIMENSION output_col) |
| 1088 { |
| 1089 if (simd_support & JSIMD_MIPS_DSPR2) { |
| 1090 JCOEFPTR inptr; |
| 1091 IFAST_MULT_TYPE *quantptr; |
| 1092 DCTELEM workspace[DCTSIZE2]; /* buffers data between passes */ |
| 1093 |
| 1094 /* Pass 1: process columns from input, store into work array. */ |
| 1095 |
| 1096 inptr = coef_block; |
| 1097 quantptr = (IFAST_MULT_TYPE *) compptr->dct_table; |
| 1098 |
| 1099 jsimd_idct_ifast_cols_mips_dspr2(inptr, quantptr, |
| 1100 workspace, mips_idct_ifast_coefs); |
| 1101 |
| 1102 /* Pass 2: process rows from work array, store into output array. */ |
| 1103 /* Note that we must descale the results by a factor of 8 == 2**3, */ |
| 1104 /* and also undo the PASS1_BITS scaling. */ |
| 1105 |
| 1106 jsimd_idct_ifast_rows_mips_dspr2(workspace, output_buf, |
| 1107 output_col, mips_idct_ifast_coefs); |
| 1108 } |
| 1109 } |
| 1110 |
| 1111 GLOBAL(void) |
| 1112 jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info *compptr, |
| 1113 JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 1114 JDIMENSION output_col) |
| 1115 { |
| 1116 } |
| 1117 |
| 1118 GLOBAL(int) |
| 1119 jsimd_can_huff_encode_one_block (void) |
| 1120 { |
| 1121 return 0; |
| 1122 } |
| 1123 |
| 1124 GLOBAL(JOCTET*) |
| 1125 jsimd_huff_encode_one_block (void *state, JOCTET *buffer, JCOEFPTR block, |
| 1126 int last_dc_val, c_derived_tbl *dctbl, |
| 1127 c_derived_tbl *actbl) |
| 1128 { |
| 1129 return NULL; |
| 1130 } |
OLD | NEW |