1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jsimd_x86_64.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,753 @@
1.4 +/*
1.5 + * jsimd_x86_64.c
1.6 + *
1.7 + * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
1.8 + * Copyright 2009-2011 D. R. Commander
1.9 + *
1.10 + * Based on the x86 SIMD extension for IJG JPEG library,
1.11 + * Copyright (C) 1999-2006, MIYASAKA Masaru.
1.12 + * For conditions of distribution and use, see copyright notice in jsimdext.inc
1.13 + *
1.14 + * This file contains the interface between the "normal" portions
1.15 + * of the library and the SIMD implementations when running on a
1.16 + * x86_64 architecture.
1.17 + */
1.18 +
1.19 +#define JPEG_INTERNALS
1.20 +#include "../jinclude.h"
1.21 +#include "../jpeglib.h"
1.22 +#include "../jsimd.h"
1.23 +#include "../jdct.h"
1.24 +#include "../jsimddct.h"
1.25 +#include "jsimd.h"
1.26 +
1.27 +/*
1.28 + * In the PIC cases, we have no guarantee that constants will keep
1.29 + * their alignment. This macro allows us to verify it at runtime.
1.30 + */
1.31 +#define IS_ALIGNED(ptr, order) (((size_t)ptr & ((1 << order) - 1)) == 0)
1.32 +
1.33 +#define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */
1.34 +
1.35 +GLOBAL(int)
1.36 +jsimd_can_rgb_ycc (void)
1.37 +{
1.38 + /* The code is optimised for these values only */
1.39 + if (BITS_IN_JSAMPLE != 8)
1.40 + return 0;
1.41 + if (sizeof(JDIMENSION) != 4)
1.42 + return 0;
1.43 + if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
1.44 + return 0;
1.45 +
1.46 + if (!IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
1.47 + return 0;
1.48 +
1.49 + return 1;
1.50 +}
1.51 +
1.52 +GLOBAL(int)
1.53 +jsimd_can_rgb_gray (void)
1.54 +{
1.55 + /* The code is optimised for these values only */
1.56 + if (BITS_IN_JSAMPLE != 8)
1.57 + return 0;
1.58 + if (sizeof(JDIMENSION) != 4)
1.59 + return 0;
1.60 + if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
1.61 + return 0;
1.62 +
1.63 + if (!IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
1.64 + return 0;
1.65 +
1.66 + return 1;
1.67 +}
1.68 +
1.69 +GLOBAL(int)
1.70 +jsimd_can_ycc_rgb (void)
1.71 +{
1.72 + /* The code is optimised for these values only */
1.73 + if (BITS_IN_JSAMPLE != 8)
1.74 + return 0;
1.75 + if (sizeof(JDIMENSION) != 4)
1.76 + return 0;
1.77 + if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
1.78 + return 0;
1.79 +
1.80 + if (!IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
1.81 + return 0;
1.82 +
1.83 + return 1;
1.84 +}
1.85 +
1.86 +GLOBAL(void)
1.87 +jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
1.88 + JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
1.89 + JDIMENSION output_row, int num_rows)
1.90 +{
1.91 + void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.92 +
1.93 + switch(cinfo->in_color_space)
1.94 + {
1.95 + case JCS_EXT_RGB:
1.96 + sse2fct=jsimd_extrgb_ycc_convert_sse2;
1.97 + break;
1.98 + case JCS_EXT_RGBX:
1.99 + case JCS_EXT_RGBA:
1.100 + sse2fct=jsimd_extrgbx_ycc_convert_sse2;
1.101 + break;
1.102 + case JCS_EXT_BGR:
1.103 + sse2fct=jsimd_extbgr_ycc_convert_sse2;
1.104 + break;
1.105 + case JCS_EXT_BGRX:
1.106 + case JCS_EXT_BGRA:
1.107 + sse2fct=jsimd_extbgrx_ycc_convert_sse2;
1.108 + break;
1.109 + case JCS_EXT_XBGR:
1.110 + case JCS_EXT_ABGR:
1.111 + sse2fct=jsimd_extxbgr_ycc_convert_sse2;
1.112 + break;
1.113 + case JCS_EXT_XRGB:
1.114 + case JCS_EXT_ARGB:
1.115 + sse2fct=jsimd_extxrgb_ycc_convert_sse2;
1.116 + break;
1.117 + default:
1.118 + sse2fct=jsimd_rgb_ycc_convert_sse2;
1.119 + break;
1.120 + }
1.121 +
1.122 + sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
1.123 +}
1.124 +
1.125 +GLOBAL(void)
1.126 +jsimd_rgb_gray_convert (j_compress_ptr cinfo,
1.127 + JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
1.128 + JDIMENSION output_row, int num_rows)
1.129 +{
1.130 + void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.131 +
1.132 + switch(cinfo->in_color_space)
1.133 + {
1.134 + case JCS_EXT_RGB:
1.135 + sse2fct=jsimd_extrgb_gray_convert_sse2;
1.136 + break;
1.137 + case JCS_EXT_RGBX:
1.138 + case JCS_EXT_RGBA:
1.139 + sse2fct=jsimd_extrgbx_gray_convert_sse2;
1.140 + break;
1.141 + case JCS_EXT_BGR:
1.142 + sse2fct=jsimd_extbgr_gray_convert_sse2;
1.143 + break;
1.144 + case JCS_EXT_BGRX:
1.145 + case JCS_EXT_BGRA:
1.146 + sse2fct=jsimd_extbgrx_gray_convert_sse2;
1.147 + break;
1.148 + case JCS_EXT_XBGR:
1.149 + case JCS_EXT_ABGR:
1.150 + sse2fct=jsimd_extxbgr_gray_convert_sse2;
1.151 + break;
1.152 + case JCS_EXT_XRGB:
1.153 + case JCS_EXT_ARGB:
1.154 + sse2fct=jsimd_extxrgb_gray_convert_sse2;
1.155 + break;
1.156 + default:
1.157 + sse2fct=jsimd_rgb_gray_convert_sse2;
1.158 + break;
1.159 + }
1.160 +
1.161 + sse2fct(cinfo->image_width, input_buf, output_buf, output_row, num_rows);
1.162 +}
1.163 +
1.164 +GLOBAL(void)
1.165 +jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
1.166 + JSAMPIMAGE input_buf, JDIMENSION input_row,
1.167 + JSAMPARRAY output_buf, int num_rows)
1.168 +{
1.169 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
1.170 +
1.171 + switch(cinfo->out_color_space)
1.172 + {
1.173 + case JCS_EXT_RGB:
1.174 + sse2fct=jsimd_ycc_extrgb_convert_sse2;
1.175 + break;
1.176 + case JCS_EXT_RGBX:
1.177 + case JCS_EXT_RGBA:
1.178 + sse2fct=jsimd_ycc_extrgbx_convert_sse2;
1.179 + break;
1.180 + case JCS_EXT_BGR:
1.181 + sse2fct=jsimd_ycc_extbgr_convert_sse2;
1.182 + break;
1.183 + case JCS_EXT_BGRX:
1.184 + case JCS_EXT_BGRA:
1.185 + sse2fct=jsimd_ycc_extbgrx_convert_sse2;
1.186 + break;
1.187 + case JCS_EXT_XBGR:
1.188 + case JCS_EXT_ABGR:
1.189 + sse2fct=jsimd_ycc_extxbgr_convert_sse2;
1.190 + break;
1.191 + case JCS_EXT_XRGB:
1.192 + case JCS_EXT_ARGB:
1.193 + sse2fct=jsimd_ycc_extxrgb_convert_sse2;
1.194 + break;
1.195 + default:
1.196 + sse2fct=jsimd_ycc_rgb_convert_sse2;
1.197 + break;
1.198 + }
1.199 +
1.200 + sse2fct(cinfo->output_width, input_buf, input_row, output_buf, num_rows);
1.201 +}
1.202 +
1.203 +GLOBAL(int)
1.204 +jsimd_can_h2v2_downsample (void)
1.205 +{
1.206 + /* The code is optimised for these values only */
1.207 + if (BITS_IN_JSAMPLE != 8)
1.208 + return 0;
1.209 + if (sizeof(JDIMENSION) != 4)
1.210 + return 0;
1.211 +
1.212 + return 1;
1.213 +}
1.214 +
1.215 +GLOBAL(int)
1.216 +jsimd_can_h2v1_downsample (void)
1.217 +{
1.218 + /* The code is optimised for these values only */
1.219 + if (BITS_IN_JSAMPLE != 8)
1.220 + return 0;
1.221 + if (sizeof(JDIMENSION) != 4)
1.222 + return 0;
1.223 +
1.224 + return 1;
1.225 +}
1.226 +
1.227 +GLOBAL(void)
1.228 +jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
1.229 + JSAMPARRAY input_data, JSAMPARRAY output_data)
1.230 +{
1.231 + jsimd_h2v2_downsample_sse2(cinfo->image_width,
1.232 + cinfo->max_v_samp_factor,
1.233 + compptr->v_samp_factor,
1.234 + compptr->width_in_blocks,
1.235 + input_data, output_data);
1.236 +}
1.237 +
1.238 +GLOBAL(void)
1.239 +jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
1.240 + JSAMPARRAY input_data, JSAMPARRAY output_data)
1.241 +{
1.242 + jsimd_h2v1_downsample_sse2(cinfo->image_width,
1.243 + cinfo->max_v_samp_factor,
1.244 + compptr->v_samp_factor,
1.245 + compptr->width_in_blocks,
1.246 + input_data, output_data);
1.247 +}
1.248 +
1.249 +GLOBAL(int)
1.250 +jsimd_can_h2v2_upsample (void)
1.251 +{
1.252 + /* The code is optimised for these values only */
1.253 + if (BITS_IN_JSAMPLE != 8)
1.254 + return 0;
1.255 + if (sizeof(JDIMENSION) != 4)
1.256 + return 0;
1.257 +
1.258 + return 1;
1.259 +}
1.260 +
1.261 +GLOBAL(int)
1.262 +jsimd_can_h2v1_upsample (void)
1.263 +{
1.264 + /* The code is optimised for these values only */
1.265 + if (BITS_IN_JSAMPLE != 8)
1.266 + return 0;
1.267 + if (sizeof(JDIMENSION) != 4)
1.268 + return 0;
1.269 +
1.270 + return 1;
1.271 +}
1.272 +
1.273 +GLOBAL(void)
1.274 +jsimd_h2v2_upsample (j_decompress_ptr cinfo,
1.275 + jpeg_component_info * compptr,
1.276 + JSAMPARRAY input_data,
1.277 + JSAMPARRAY * output_data_ptr)
1.278 +{
1.279 + jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
1.280 + cinfo->output_width,
1.281 + input_data, output_data_ptr);
1.282 +}
1.283 +
1.284 +GLOBAL(void)
1.285 +jsimd_h2v1_upsample (j_decompress_ptr cinfo,
1.286 + jpeg_component_info * compptr,
1.287 + JSAMPARRAY input_data,
1.288 + JSAMPARRAY * output_data_ptr)
1.289 +{
1.290 + jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
1.291 + cinfo->output_width,
1.292 + input_data, output_data_ptr);
1.293 +}
1.294 +
1.295 +GLOBAL(int)
1.296 +jsimd_can_h2v2_fancy_upsample (void)
1.297 +{
1.298 + /* The code is optimised for these values only */
1.299 + if (BITS_IN_JSAMPLE != 8)
1.300 + return 0;
1.301 + if (sizeof(JDIMENSION) != 4)
1.302 + return 0;
1.303 +
1.304 + if (!IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.305 + return 0;
1.306 +
1.307 + return 1;
1.308 +}
1.309 +
1.310 +GLOBAL(int)
1.311 +jsimd_can_h2v1_fancy_upsample (void)
1.312 +{
1.313 + /* The code is optimised for these values only */
1.314 + if (BITS_IN_JSAMPLE != 8)
1.315 + return 0;
1.316 + if (sizeof(JDIMENSION) != 4)
1.317 + return 0;
1.318 +
1.319 + if (!IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.320 + return 0;
1.321 +
1.322 + return 1;
1.323 +}
1.324 +
1.325 +GLOBAL(void)
1.326 +jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
1.327 + jpeg_component_info * compptr,
1.328 + JSAMPARRAY input_data,
1.329 + JSAMPARRAY * output_data_ptr)
1.330 +{
1.331 + jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
1.332 + compptr->downsampled_width,
1.333 + input_data, output_data_ptr);
1.334 +}
1.335 +
1.336 +GLOBAL(void)
1.337 +jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
1.338 + jpeg_component_info * compptr,
1.339 + JSAMPARRAY input_data,
1.340 + JSAMPARRAY * output_data_ptr)
1.341 +{
1.342 + jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
1.343 + compptr->downsampled_width,
1.344 + input_data, output_data_ptr);
1.345 +}
1.346 +
1.347 +GLOBAL(int)
1.348 +jsimd_can_h2v2_merged_upsample (void)
1.349 +{
1.350 + /* The code is optimised for these values only */
1.351 + if (BITS_IN_JSAMPLE != 8)
1.352 + return 0;
1.353 + if (sizeof(JDIMENSION) != 4)
1.354 + return 0;
1.355 +
1.356 + if (!IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.357 + return 0;
1.358 +
1.359 + return 1;
1.360 +}
1.361 +
1.362 +GLOBAL(int)
1.363 +jsimd_can_h2v1_merged_upsample (void)
1.364 +{
1.365 + /* The code is optimised for these values only */
1.366 + if (BITS_IN_JSAMPLE != 8)
1.367 + return 0;
1.368 + if (sizeof(JDIMENSION) != 4)
1.369 + return 0;
1.370 +
1.371 + if (!IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.372 + return 0;
1.373 +
1.374 + return 1;
1.375 +}
1.376 +
1.377 +GLOBAL(void)
1.378 +jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
1.379 + JSAMPIMAGE input_buf,
1.380 + JDIMENSION in_row_group_ctr,
1.381 + JSAMPARRAY output_buf)
1.382 +{
1.383 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.384 +
1.385 + switch(cinfo->out_color_space)
1.386 + {
1.387 + case JCS_EXT_RGB:
1.388 + sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
1.389 + break;
1.390 + case JCS_EXT_RGBX:
1.391 + case JCS_EXT_RGBA:
1.392 + sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
1.393 + break;
1.394 + case JCS_EXT_BGR:
1.395 + sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
1.396 + break;
1.397 + case JCS_EXT_BGRX:
1.398 + case JCS_EXT_BGRA:
1.399 + sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
1.400 + break;
1.401 + case JCS_EXT_XBGR:
1.402 + case JCS_EXT_ABGR:
1.403 + sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
1.404 + break;
1.405 + case JCS_EXT_XRGB:
1.406 + case JCS_EXT_ARGB:
1.407 + sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
1.408 + break;
1.409 + default:
1.410 + sse2fct=jsimd_h2v2_merged_upsample_sse2;
1.411 + break;
1.412 + }
1.413 +
1.414 + sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
1.415 +}
1.416 +
1.417 +GLOBAL(void)
1.418 +jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
1.419 + JSAMPIMAGE input_buf,
1.420 + JDIMENSION in_row_group_ctr,
1.421 + JSAMPARRAY output_buf)
1.422 +{
1.423 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.424 +
1.425 + switch(cinfo->out_color_space)
1.426 + {
1.427 + case JCS_EXT_RGB:
1.428 + sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
1.429 + break;
1.430 + case JCS_EXT_RGBX:
1.431 + case JCS_EXT_RGBA:
1.432 + sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
1.433 + break;
1.434 + case JCS_EXT_BGR:
1.435 + sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
1.436 + break;
1.437 + case JCS_EXT_BGRX:
1.438 + case JCS_EXT_BGRA:
1.439 + sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
1.440 + break;
1.441 + case JCS_EXT_XBGR:
1.442 + case JCS_EXT_ABGR:
1.443 + sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
1.444 + break;
1.445 + case JCS_EXT_XRGB:
1.446 + case JCS_EXT_ARGB:
1.447 + sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
1.448 + break;
1.449 + default:
1.450 + sse2fct=jsimd_h2v1_merged_upsample_sse2;
1.451 + break;
1.452 + }
1.453 +
1.454 + sse2fct(cinfo->output_width, input_buf, in_row_group_ctr, output_buf);
1.455 +}
1.456 +
1.457 +GLOBAL(int)
1.458 +jsimd_can_convsamp (void)
1.459 +{
1.460 + /* The code is optimised for these values only */
1.461 + if (DCTSIZE != 8)
1.462 + return 0;
1.463 + if (BITS_IN_JSAMPLE != 8)
1.464 + return 0;
1.465 + if (sizeof(JDIMENSION) != 4)
1.466 + return 0;
1.467 + if (sizeof(DCTELEM) != 2)
1.468 + return 0;
1.469 +
1.470 + return 1;
1.471 +}
1.472 +
1.473 +GLOBAL(int)
1.474 +jsimd_can_convsamp_float (void)
1.475 +{
1.476 + /* The code is optimised for these values only */
1.477 + if (DCTSIZE != 8)
1.478 + return 0;
1.479 + if (BITS_IN_JSAMPLE != 8)
1.480 + return 0;
1.481 + if (sizeof(JDIMENSION) != 4)
1.482 + return 0;
1.483 + if (sizeof(FAST_FLOAT) != 4)
1.484 + return 0;
1.485 +
1.486 + return 1;
1.487 +}
1.488 +
1.489 +GLOBAL(void)
1.490 +jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
1.491 + DCTELEM * workspace)
1.492 +{
1.493 + jsimd_convsamp_sse2(sample_data, start_col, workspace);
1.494 +}
1.495 +
1.496 +GLOBAL(void)
1.497 +jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
1.498 + FAST_FLOAT * workspace)
1.499 +{
1.500 + jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
1.501 +}
1.502 +
1.503 +GLOBAL(int)
1.504 +jsimd_can_fdct_islow (void)
1.505 +{
1.506 + /* The code is optimised for these values only */
1.507 + if (DCTSIZE != 8)
1.508 + return 0;
1.509 + if (sizeof(DCTELEM) != 2)
1.510 + return 0;
1.511 +
1.512 + if (!IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
1.513 + return 0;
1.514 +
1.515 + return 1;
1.516 +}
1.517 +
1.518 +GLOBAL(int)
1.519 +jsimd_can_fdct_ifast (void)
1.520 +{
1.521 + /* The code is optimised for these values only */
1.522 + if (DCTSIZE != 8)
1.523 + return 0;
1.524 + if (sizeof(DCTELEM) != 2)
1.525 + return 0;
1.526 +
1.527 + if (!IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
1.528 + return 0;
1.529 +
1.530 + return 1;
1.531 +}
1.532 +
1.533 +GLOBAL(int)
1.534 +jsimd_can_fdct_float (void)
1.535 +{
1.536 + /* The code is optimised for these values only */
1.537 + if (DCTSIZE != 8)
1.538 + return 0;
1.539 + if (sizeof(FAST_FLOAT) != 4)
1.540 + return 0;
1.541 +
1.542 + if (!IS_ALIGNED_SSE(jconst_fdct_float_sse))
1.543 + return 0;
1.544 +
1.545 + return 1;
1.546 +}
1.547 +
1.548 +GLOBAL(void)
1.549 +jsimd_fdct_islow (DCTELEM * data)
1.550 +{
1.551 + jsimd_fdct_islow_sse2(data);
1.552 +}
1.553 +
1.554 +GLOBAL(void)
1.555 +jsimd_fdct_ifast (DCTELEM * data)
1.556 +{
1.557 + jsimd_fdct_ifast_sse2(data);
1.558 +}
1.559 +
1.560 +GLOBAL(void)
1.561 +jsimd_fdct_float (FAST_FLOAT * data)
1.562 +{
1.563 + jsimd_fdct_float_sse(data);
1.564 +}
1.565 +
1.566 +GLOBAL(int)
1.567 +jsimd_can_quantize (void)
1.568 +{
1.569 + /* The code is optimised for these values only */
1.570 + if (DCTSIZE != 8)
1.571 + return 0;
1.572 + if (sizeof(JCOEF) != 2)
1.573 + return 0;
1.574 + if (sizeof(DCTELEM) != 2)
1.575 + return 0;
1.576 +
1.577 + return 1;
1.578 +}
1.579 +
1.580 +GLOBAL(int)
1.581 +jsimd_can_quantize_float (void)
1.582 +{
1.583 + /* The code is optimised for these values only */
1.584 + if (DCTSIZE != 8)
1.585 + return 0;
1.586 + if (sizeof(JCOEF) != 2)
1.587 + return 0;
1.588 + if (sizeof(FAST_FLOAT) != 4)
1.589 + return 0;
1.590 +
1.591 + return 1;
1.592 +}
1.593 +
1.594 +GLOBAL(void)
1.595 +jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
1.596 + DCTELEM * workspace)
1.597 +{
1.598 + jsimd_quantize_sse2(coef_block, divisors, workspace);
1.599 +}
1.600 +
1.601 +GLOBAL(void)
1.602 +jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
1.603 + FAST_FLOAT * workspace)
1.604 +{
1.605 + jsimd_quantize_float_sse2(coef_block, divisors, workspace);
1.606 +}
1.607 +
1.608 +GLOBAL(int)
1.609 +jsimd_can_idct_2x2 (void)
1.610 +{
1.611 + /* The code is optimised for these values only */
1.612 + if (DCTSIZE != 8)
1.613 + return 0;
1.614 + if (sizeof(JCOEF) != 2)
1.615 + return 0;
1.616 + if (BITS_IN_JSAMPLE != 8)
1.617 + return 0;
1.618 + if (sizeof(JDIMENSION) != 4)
1.619 + return 0;
1.620 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.621 + return 0;
1.622 +
1.623 + if (!IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.624 + return 0;
1.625 +
1.626 + return 1;
1.627 +}
1.628 +
1.629 +GLOBAL(int)
1.630 +jsimd_can_idct_4x4 (void)
1.631 +{
1.632 + /* The code is optimised for these values only */
1.633 + if (DCTSIZE != 8)
1.634 + return 0;
1.635 + if (sizeof(JCOEF) != 2)
1.636 + return 0;
1.637 + if (BITS_IN_JSAMPLE != 8)
1.638 + return 0;
1.639 + if (sizeof(JDIMENSION) != 4)
1.640 + return 0;
1.641 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.642 + return 0;
1.643 +
1.644 + if (!IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.645 + return 0;
1.646 +
1.647 + return 1;
1.648 +}
1.649 +
1.650 +GLOBAL(void)
1.651 +jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.652 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.653 + JDIMENSION output_col)
1.654 +{
1.655 + jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.656 +}
1.657 +
1.658 +GLOBAL(void)
1.659 +jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.660 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.661 + JDIMENSION output_col)
1.662 +{
1.663 + jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.664 +}
1.665 +
1.666 +GLOBAL(int)
1.667 +jsimd_can_idct_islow (void)
1.668 +{
1.669 + /* The code is optimised for these values only */
1.670 + if (DCTSIZE != 8)
1.671 + return 0;
1.672 + if (sizeof(JCOEF) != 2)
1.673 + return 0;
1.674 + if (BITS_IN_JSAMPLE != 8)
1.675 + return 0;
1.676 + if (sizeof(JDIMENSION) != 4)
1.677 + return 0;
1.678 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.679 + return 0;
1.680 +
1.681 + if (!IS_ALIGNED_SSE(jconst_idct_islow_sse2))
1.682 + return 0;
1.683 +
1.684 + return 1;
1.685 +}
1.686 +
1.687 +GLOBAL(int)
1.688 +jsimd_can_idct_ifast (void)
1.689 +{
1.690 + /* The code is optimised for these values only */
1.691 + if (DCTSIZE != 8)
1.692 + return 0;
1.693 + if (sizeof(JCOEF) != 2)
1.694 + return 0;
1.695 + if (BITS_IN_JSAMPLE != 8)
1.696 + return 0;
1.697 + if (sizeof(JDIMENSION) != 4)
1.698 + return 0;
1.699 + if (sizeof(IFAST_MULT_TYPE) != 2)
1.700 + return 0;
1.701 + if (IFAST_SCALE_BITS != 2)
1.702 + return 0;
1.703 +
1.704 + if (!IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
1.705 + return 0;
1.706 +
1.707 + return 1;
1.708 +}
1.709 +
1.710 +GLOBAL(int)
1.711 +jsimd_can_idct_float (void)
1.712 +{
1.713 + if (DCTSIZE != 8)
1.714 + return 0;
1.715 + if (sizeof(JCOEF) != 2)
1.716 + return 0;
1.717 + if (BITS_IN_JSAMPLE != 8)
1.718 + return 0;
1.719 + if (sizeof(JDIMENSION) != 4)
1.720 + return 0;
1.721 + if (sizeof(FAST_FLOAT) != 4)
1.722 + return 0;
1.723 + if (sizeof(FLOAT_MULT_TYPE) != 4)
1.724 + return 0;
1.725 +
1.726 + if (!IS_ALIGNED_SSE(jconst_idct_float_sse2))
1.727 + return 0;
1.728 +
1.729 + return 1;
1.730 +}
1.731 +
1.732 +GLOBAL(void)
1.733 +jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.734 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.735 + JDIMENSION output_col)
1.736 +{
1.737 + jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.738 +}
1.739 +
1.740 +GLOBAL(void)
1.741 +jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.742 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.743 + JDIMENSION output_col)
1.744 +{
1.745 + jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.746 +}
1.747 +
1.748 +GLOBAL(void)
1.749 +jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.750 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.751 + JDIMENSION output_col)
1.752 +{
1.753 + jsimd_idct_float_sse2(compptr->dct_table, coef_block,
1.754 + output_buf, output_col);
1.755 +}
1.756 +
