1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/media/libjpeg/simd/jsimd_i386.c Wed Dec 31 06:09:35 2014 +0100
1.3 @@ -0,0 +1,1048 @@
1.4 +/*
1.5 + * jsimd_i386.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 + * 32-bit x86 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) (((unsigned)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 +static unsigned int simd_support = ~0;
1.36 +
1.37 +/*
1.38 + * Check what SIMD accelerations are supported.
1.39 + *
1.40 + * FIXME: This code is racy under a multi-threaded environment.
1.41 + */
1.42 +LOCAL(void)
1.43 +init_simd (void)
1.44 +{
1.45 + char *env = NULL;
1.46 +
1.47 + if (simd_support != ~0U)
1.48 + return;
1.49 +
1.50 + simd_support = jpeg_simd_cpu_support();
1.51 +
1.52 + /* Force different settings through environment variables */
1.53 + env = getenv("JSIMD_FORCEMMX");
1.54 + if ((env != NULL) && (strcmp(env, "1") == 0))
1.55 + simd_support &= JSIMD_MMX;
1.56 + env = getenv("JSIMD_FORCE3DNOW");
1.57 + if ((env != NULL) && (strcmp(env, "1") == 0))
1.58 + simd_support &= JSIMD_3DNOW|JSIMD_MMX;
1.59 + env = getenv("JSIMD_FORCESSE");
1.60 + if ((env != NULL) && (strcmp(env, "1") == 0))
1.61 + simd_support &= JSIMD_SSE|JSIMD_MMX;
1.62 + env = getenv("JSIMD_FORCESSE2");
1.63 + if ((env != NULL) && (strcmp(env, "1") == 0))
1.64 + simd_support &= JSIMD_SSE2;
1.65 +}
1.66 +
1.67 +GLOBAL(int)
1.68 +jsimd_can_rgb_ycc (void)
1.69 +{
1.70 + init_simd();
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 ((simd_support & JSIMD_SSE2) &&
1.81 + IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
1.82 + return 1;
1.83 + if (simd_support & JSIMD_MMX)
1.84 + return 1;
1.85 +
1.86 + return 0;
1.87 +}
1.88 +
1.89 +GLOBAL(int)
1.90 +jsimd_can_rgb_gray (void)
1.91 +{
1.92 + init_simd();
1.93 +
1.94 + /* The code is optimised for these values only */
1.95 + if (BITS_IN_JSAMPLE != 8)
1.96 + return 0;
1.97 + if (sizeof(JDIMENSION) != 4)
1.98 + return 0;
1.99 + if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
1.100 + return 0;
1.101 +
1.102 + if ((simd_support & JSIMD_SSE2) &&
1.103 + IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
1.104 + return 1;
1.105 + if (simd_support & JSIMD_MMX)
1.106 + return 1;
1.107 +
1.108 + return 0;
1.109 +}
1.110 +
1.111 +GLOBAL(int)
1.112 +jsimd_can_ycc_rgb (void)
1.113 +{
1.114 + init_simd();
1.115 +
1.116 + /* The code is optimised for these values only */
1.117 + if (BITS_IN_JSAMPLE != 8)
1.118 + return 0;
1.119 + if (sizeof(JDIMENSION) != 4)
1.120 + return 0;
1.121 + if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4))
1.122 + return 0;
1.123 +
1.124 + if ((simd_support & JSIMD_SSE2) &&
1.125 + IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
1.126 + return 1;
1.127 + if (simd_support & JSIMD_MMX)
1.128 + return 1;
1.129 +
1.130 + return 0;
1.131 +}
1.132 +
1.133 +GLOBAL(void)
1.134 +jsimd_rgb_ycc_convert (j_compress_ptr cinfo,
1.135 + JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
1.136 + JDIMENSION output_row, int num_rows)
1.137 +{
1.138 + void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.139 + void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.140 +
1.141 + switch(cinfo->in_color_space)
1.142 + {
1.143 + case JCS_EXT_RGB:
1.144 + sse2fct=jsimd_extrgb_ycc_convert_sse2;
1.145 + mmxfct=jsimd_extrgb_ycc_convert_mmx;
1.146 + break;
1.147 + case JCS_EXT_RGBX:
1.148 + case JCS_EXT_RGBA:
1.149 + sse2fct=jsimd_extrgbx_ycc_convert_sse2;
1.150 + mmxfct=jsimd_extrgbx_ycc_convert_mmx;
1.151 + break;
1.152 + case JCS_EXT_BGR:
1.153 + sse2fct=jsimd_extbgr_ycc_convert_sse2;
1.154 + mmxfct=jsimd_extbgr_ycc_convert_mmx;
1.155 + break;
1.156 + case JCS_EXT_BGRX:
1.157 + case JCS_EXT_BGRA:
1.158 + sse2fct=jsimd_extbgrx_ycc_convert_sse2;
1.159 + mmxfct=jsimd_extbgrx_ycc_convert_mmx;
1.160 + break;
1.161 + case JCS_EXT_XBGR:
1.162 + case JCS_EXT_ABGR:
1.163 + sse2fct=jsimd_extxbgr_ycc_convert_sse2;
1.164 + mmxfct=jsimd_extxbgr_ycc_convert_mmx;
1.165 + break;
1.166 + case JCS_EXT_XRGB:
1.167 + case JCS_EXT_ARGB:
1.168 + sse2fct=jsimd_extxrgb_ycc_convert_sse2;
1.169 + mmxfct=jsimd_extxrgb_ycc_convert_mmx;
1.170 + break;
1.171 + default:
1.172 + sse2fct=jsimd_rgb_ycc_convert_sse2;
1.173 + mmxfct=jsimd_rgb_ycc_convert_mmx;
1.174 + break;
1.175 + }
1.176 +
1.177 + if ((simd_support & JSIMD_SSE2) &&
1.178 + IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2))
1.179 + sse2fct(cinfo->image_width, input_buf,
1.180 + output_buf, output_row, num_rows);
1.181 + else if (simd_support & JSIMD_MMX)
1.182 + mmxfct(cinfo->image_width, input_buf,
1.183 + output_buf, output_row, num_rows);
1.184 +}
1.185 +
1.186 +GLOBAL(void)
1.187 +jsimd_rgb_gray_convert (j_compress_ptr cinfo,
1.188 + JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
1.189 + JDIMENSION output_row, int num_rows)
1.190 +{
1.191 + void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.192 + void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int);
1.193 +
1.194 + switch(cinfo->in_color_space)
1.195 + {
1.196 + case JCS_EXT_RGB:
1.197 + sse2fct=jsimd_extrgb_gray_convert_sse2;
1.198 + mmxfct=jsimd_extrgb_gray_convert_mmx;
1.199 + break;
1.200 + case JCS_EXT_RGBX:
1.201 + case JCS_EXT_RGBA:
1.202 + sse2fct=jsimd_extrgbx_gray_convert_sse2;
1.203 + mmxfct=jsimd_extrgbx_gray_convert_mmx;
1.204 + break;
1.205 + case JCS_EXT_BGR:
1.206 + sse2fct=jsimd_extbgr_gray_convert_sse2;
1.207 + mmxfct=jsimd_extbgr_gray_convert_mmx;
1.208 + break;
1.209 + case JCS_EXT_BGRX:
1.210 + case JCS_EXT_BGRA:
1.211 + sse2fct=jsimd_extbgrx_gray_convert_sse2;
1.212 + mmxfct=jsimd_extbgrx_gray_convert_mmx;
1.213 + break;
1.214 + case JCS_EXT_XBGR:
1.215 + case JCS_EXT_ABGR:
1.216 + sse2fct=jsimd_extxbgr_gray_convert_sse2;
1.217 + mmxfct=jsimd_extxbgr_gray_convert_mmx;
1.218 + break;
1.219 + case JCS_EXT_XRGB:
1.220 + case JCS_EXT_ARGB:
1.221 + sse2fct=jsimd_extxrgb_gray_convert_sse2;
1.222 + mmxfct=jsimd_extxrgb_gray_convert_mmx;
1.223 + break;
1.224 + default:
1.225 + sse2fct=jsimd_rgb_gray_convert_sse2;
1.226 + mmxfct=jsimd_rgb_gray_convert_mmx;
1.227 + break;
1.228 + }
1.229 +
1.230 + if ((simd_support & JSIMD_SSE2) &&
1.231 + IS_ALIGNED_SSE(jconst_rgb_gray_convert_sse2))
1.232 + sse2fct(cinfo->image_width, input_buf,
1.233 + output_buf, output_row, num_rows);
1.234 + else if (simd_support & JSIMD_MMX)
1.235 + mmxfct(cinfo->image_width, input_buf,
1.236 + output_buf, output_row, num_rows);
1.237 +}
1.238 +
1.239 +GLOBAL(void)
1.240 +jsimd_ycc_rgb_convert (j_decompress_ptr cinfo,
1.241 + JSAMPIMAGE input_buf, JDIMENSION input_row,
1.242 + JSAMPARRAY output_buf, int num_rows)
1.243 +{
1.244 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
1.245 + void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int);
1.246 +
1.247 + switch(cinfo->out_color_space)
1.248 + {
1.249 + case JCS_EXT_RGB:
1.250 + sse2fct=jsimd_ycc_extrgb_convert_sse2;
1.251 + mmxfct=jsimd_ycc_extrgb_convert_mmx;
1.252 + break;
1.253 + case JCS_EXT_RGBX:
1.254 + case JCS_EXT_RGBA:
1.255 + sse2fct=jsimd_ycc_extrgbx_convert_sse2;
1.256 + mmxfct=jsimd_ycc_extrgbx_convert_mmx;
1.257 + break;
1.258 + case JCS_EXT_BGR:
1.259 + sse2fct=jsimd_ycc_extbgr_convert_sse2;
1.260 + mmxfct=jsimd_ycc_extbgr_convert_mmx;
1.261 + break;
1.262 + case JCS_EXT_BGRX:
1.263 + case JCS_EXT_BGRA:
1.264 + sse2fct=jsimd_ycc_extbgrx_convert_sse2;
1.265 + mmxfct=jsimd_ycc_extbgrx_convert_mmx;
1.266 + break;
1.267 + case JCS_EXT_XBGR:
1.268 + case JCS_EXT_ABGR:
1.269 + sse2fct=jsimd_ycc_extxbgr_convert_sse2;
1.270 + mmxfct=jsimd_ycc_extxbgr_convert_mmx;
1.271 + break;
1.272 + case JCS_EXT_XRGB:
1.273 + case JCS_EXT_ARGB:
1.274 + sse2fct=jsimd_ycc_extxrgb_convert_sse2;
1.275 + mmxfct=jsimd_ycc_extxrgb_convert_mmx;
1.276 + break;
1.277 + default:
1.278 + sse2fct=jsimd_ycc_rgb_convert_sse2;
1.279 + mmxfct=jsimd_ycc_rgb_convert_mmx;
1.280 + break;
1.281 + }
1.282 +
1.283 + if ((simd_support & JSIMD_SSE2) &&
1.284 + IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2))
1.285 + sse2fct(cinfo->output_width, input_buf,
1.286 + input_row, output_buf, num_rows);
1.287 + else if (simd_support & JSIMD_MMX)
1.288 + mmxfct(cinfo->output_width, input_buf,
1.289 + input_row, output_buf, num_rows);
1.290 +}
1.291 +
1.292 +GLOBAL(int)
1.293 +jsimd_can_h2v2_downsample (void)
1.294 +{
1.295 + init_simd();
1.296 +
1.297 + /* The code is optimised for these values only */
1.298 + if (BITS_IN_JSAMPLE != 8)
1.299 + return 0;
1.300 + if (sizeof(JDIMENSION) != 4)
1.301 + return 0;
1.302 +
1.303 + if (simd_support & JSIMD_SSE2)
1.304 + return 1;
1.305 + if (simd_support & JSIMD_MMX)
1.306 + return 1;
1.307 +
1.308 + return 0;
1.309 +}
1.310 +
1.311 +GLOBAL(int)
1.312 +jsimd_can_h2v1_downsample (void)
1.313 +{
1.314 + init_simd();
1.315 +
1.316 + /* The code is optimised for these values only */
1.317 + if (BITS_IN_JSAMPLE != 8)
1.318 + return 0;
1.319 + if (sizeof(JDIMENSION) != 4)
1.320 + return 0;
1.321 +
1.322 + if (simd_support & JSIMD_SSE2)
1.323 + return 1;
1.324 + if (simd_support & JSIMD_MMX)
1.325 + return 1;
1.326 +
1.327 + return 0;
1.328 +}
1.329 +
1.330 +GLOBAL(void)
1.331 +jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
1.332 + JSAMPARRAY input_data, JSAMPARRAY output_data)
1.333 +{
1.334 + if (simd_support & JSIMD_SSE2)
1.335 + jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
1.336 + compptr->v_samp_factor, compptr->width_in_blocks,
1.337 + input_data, output_data);
1.338 + else if (simd_support & JSIMD_MMX)
1.339 + jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
1.340 + compptr->v_samp_factor, compptr->width_in_blocks,
1.341 + input_data, output_data);
1.342 +}
1.343 +
1.344 +GLOBAL(void)
1.345 +jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr,
1.346 + JSAMPARRAY input_data, JSAMPARRAY output_data)
1.347 +{
1.348 + if (simd_support & JSIMD_SSE2)
1.349 + jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor,
1.350 + compptr->v_samp_factor, compptr->width_in_blocks,
1.351 + input_data, output_data);
1.352 + else if (simd_support & JSIMD_MMX)
1.353 + jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor,
1.354 + compptr->v_samp_factor, compptr->width_in_blocks,
1.355 + input_data, output_data);
1.356 +}
1.357 +
1.358 +GLOBAL(int)
1.359 +jsimd_can_h2v2_upsample (void)
1.360 +{
1.361 + init_simd();
1.362 +
1.363 + /* The code is optimised for these values only */
1.364 + if (BITS_IN_JSAMPLE != 8)
1.365 + return 0;
1.366 + if (sizeof(JDIMENSION) != 4)
1.367 + return 0;
1.368 +
1.369 + if (simd_support & JSIMD_SSE2)
1.370 + return 1;
1.371 + if (simd_support & JSIMD_MMX)
1.372 + return 1;
1.373 +
1.374 + return 0;
1.375 +}
1.376 +
1.377 +GLOBAL(int)
1.378 +jsimd_can_h2v1_upsample (void)
1.379 +{
1.380 + init_simd();
1.381 +
1.382 + /* The code is optimised for these values only */
1.383 + if (BITS_IN_JSAMPLE != 8)
1.384 + return 0;
1.385 + if (sizeof(JDIMENSION) != 4)
1.386 + return 0;
1.387 +
1.388 + if (simd_support & JSIMD_SSE2)
1.389 + return 1;
1.390 + if (simd_support & JSIMD_MMX)
1.391 + return 1;
1.392 +
1.393 + return 0;
1.394 +}
1.395 +
1.396 +GLOBAL(void)
1.397 +jsimd_h2v2_upsample (j_decompress_ptr cinfo,
1.398 + jpeg_component_info * compptr,
1.399 + JSAMPARRAY input_data,
1.400 + JSAMPARRAY * output_data_ptr)
1.401 +{
1.402 + if (simd_support & JSIMD_SSE2)
1.403 + jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor,
1.404 + cinfo->output_width, input_data, output_data_ptr);
1.405 + else if (simd_support & JSIMD_MMX)
1.406 + jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor,
1.407 + cinfo->output_width, input_data, output_data_ptr);
1.408 +}
1.409 +
1.410 +GLOBAL(void)
1.411 +jsimd_h2v1_upsample (j_decompress_ptr cinfo,
1.412 + jpeg_component_info * compptr,
1.413 + JSAMPARRAY input_data,
1.414 + JSAMPARRAY * output_data_ptr)
1.415 +{
1.416 + if (simd_support & JSIMD_SSE2)
1.417 + jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor,
1.418 + cinfo->output_width, input_data, output_data_ptr);
1.419 + else if (simd_support & JSIMD_MMX)
1.420 + jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor,
1.421 + cinfo->output_width, input_data, output_data_ptr);
1.422 +}
1.423 +
1.424 +GLOBAL(int)
1.425 +jsimd_can_h2v2_fancy_upsample (void)
1.426 +{
1.427 + init_simd();
1.428 +
1.429 + /* The code is optimised for these values only */
1.430 + if (BITS_IN_JSAMPLE != 8)
1.431 + return 0;
1.432 + if (sizeof(JDIMENSION) != 4)
1.433 + return 0;
1.434 +
1.435 + if ((simd_support & JSIMD_SSE2) &&
1.436 + IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.437 + return 1;
1.438 + if (simd_support & JSIMD_MMX)
1.439 + return 1;
1.440 +
1.441 + return 0;
1.442 +}
1.443 +
1.444 +GLOBAL(int)
1.445 +jsimd_can_h2v1_fancy_upsample (void)
1.446 +{
1.447 + init_simd();
1.448 +
1.449 + /* The code is optimised for these values only */
1.450 + if (BITS_IN_JSAMPLE != 8)
1.451 + return 0;
1.452 + if (sizeof(JDIMENSION) != 4)
1.453 + return 0;
1.454 +
1.455 + if ((simd_support & JSIMD_SSE2) &&
1.456 + IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.457 + return 1;
1.458 + if (simd_support & JSIMD_MMX)
1.459 + return 1;
1.460 +
1.461 + return 0;
1.462 +}
1.463 +
1.464 +GLOBAL(void)
1.465 +jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo,
1.466 + jpeg_component_info * compptr,
1.467 + JSAMPARRAY input_data,
1.468 + JSAMPARRAY * output_data_ptr)
1.469 +{
1.470 + if ((simd_support & JSIMD_SSE2) &&
1.471 + IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.472 + jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor,
1.473 + compptr->downsampled_width, input_data, output_data_ptr);
1.474 + else if (simd_support & JSIMD_MMX)
1.475 + jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor,
1.476 + compptr->downsampled_width, input_data, output_data_ptr);
1.477 +}
1.478 +
1.479 +GLOBAL(void)
1.480 +jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo,
1.481 + jpeg_component_info * compptr,
1.482 + JSAMPARRAY input_data,
1.483 + JSAMPARRAY * output_data_ptr)
1.484 +{
1.485 + if ((simd_support & JSIMD_SSE2) &&
1.486 + IS_ALIGNED_SSE(jconst_fancy_upsample_sse2))
1.487 + jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor,
1.488 + compptr->downsampled_width, input_data, output_data_ptr);
1.489 + else if (simd_support & JSIMD_MMX)
1.490 + jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor,
1.491 + compptr->downsampled_width, input_data, output_data_ptr);
1.492 +}
1.493 +
1.494 +GLOBAL(int)
1.495 +jsimd_can_h2v2_merged_upsample (void)
1.496 +{
1.497 + init_simd();
1.498 +
1.499 + /* The code is optimised for these values only */
1.500 + if (BITS_IN_JSAMPLE != 8)
1.501 + return 0;
1.502 + if (sizeof(JDIMENSION) != 4)
1.503 + return 0;
1.504 +
1.505 + if ((simd_support & JSIMD_SSE2) &&
1.506 + IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.507 + return 1;
1.508 + if (simd_support & JSIMD_MMX)
1.509 + return 1;
1.510 +
1.511 + return 0;
1.512 +}
1.513 +
1.514 +GLOBAL(int)
1.515 +jsimd_can_h2v1_merged_upsample (void)
1.516 +{
1.517 + init_simd();
1.518 +
1.519 + /* The code is optimised for these values only */
1.520 + if (BITS_IN_JSAMPLE != 8)
1.521 + return 0;
1.522 + if (sizeof(JDIMENSION) != 4)
1.523 + return 0;
1.524 +
1.525 + if ((simd_support & JSIMD_SSE2) &&
1.526 + IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.527 + return 1;
1.528 + if (simd_support & JSIMD_MMX)
1.529 + return 1;
1.530 +
1.531 + return 0;
1.532 +}
1.533 +
1.534 +GLOBAL(void)
1.535 +jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo,
1.536 + JSAMPIMAGE input_buf,
1.537 + JDIMENSION in_row_group_ctr,
1.538 + JSAMPARRAY output_buf)
1.539 +{
1.540 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.541 + void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.542 +
1.543 + switch(cinfo->out_color_space)
1.544 + {
1.545 + case JCS_EXT_RGB:
1.546 + sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2;
1.547 + mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx;
1.548 + break;
1.549 + case JCS_EXT_RGBX:
1.550 + case JCS_EXT_RGBA:
1.551 + sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2;
1.552 + mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx;
1.553 + break;
1.554 + case JCS_EXT_BGR:
1.555 + sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2;
1.556 + mmxfct=jsimd_h2v2_extbgr_merged_upsample_mmx;
1.557 + break;
1.558 + case JCS_EXT_BGRX:
1.559 + case JCS_EXT_BGRA:
1.560 + sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2;
1.561 + mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx;
1.562 + break;
1.563 + case JCS_EXT_XBGR:
1.564 + case JCS_EXT_ABGR:
1.565 + sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2;
1.566 + mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx;
1.567 + break;
1.568 + case JCS_EXT_XRGB:
1.569 + case JCS_EXT_ARGB:
1.570 + sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2;
1.571 + mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx;
1.572 + break;
1.573 + default:
1.574 + sse2fct=jsimd_h2v2_merged_upsample_sse2;
1.575 + mmxfct=jsimd_h2v2_merged_upsample_mmx;
1.576 + break;
1.577 + }
1.578 +
1.579 + if ((simd_support & JSIMD_SSE2) &&
1.580 + IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.581 + sse2fct(cinfo->output_width, input_buf,
1.582 + in_row_group_ctr, output_buf);
1.583 + else if (simd_support & JSIMD_MMX)
1.584 + mmxfct(cinfo->output_width, input_buf,
1.585 + in_row_group_ctr, output_buf);
1.586 +}
1.587 +
1.588 +GLOBAL(void)
1.589 +jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo,
1.590 + JSAMPIMAGE input_buf,
1.591 + JDIMENSION in_row_group_ctr,
1.592 + JSAMPARRAY output_buf)
1.593 +{
1.594 + void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.595 + void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY);
1.596 +
1.597 + switch(cinfo->out_color_space)
1.598 + {
1.599 + case JCS_EXT_RGB:
1.600 + sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2;
1.601 + mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx;
1.602 + break;
1.603 + case JCS_EXT_RGBX:
1.604 + case JCS_EXT_RGBA:
1.605 + sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2;
1.606 + mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx;
1.607 + break;
1.608 + case JCS_EXT_BGR:
1.609 + sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2;
1.610 + mmxfct=jsimd_h2v1_extbgr_merged_upsample_mmx;
1.611 + break;
1.612 + case JCS_EXT_BGRX:
1.613 + case JCS_EXT_BGRA:
1.614 + sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2;
1.615 + mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx;
1.616 + break;
1.617 + case JCS_EXT_XBGR:
1.618 + case JCS_EXT_ABGR:
1.619 + sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2;
1.620 + mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx;
1.621 + break;
1.622 + case JCS_EXT_XRGB:
1.623 + case JCS_EXT_ARGB:
1.624 + sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2;
1.625 + mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx;
1.626 + break;
1.627 + default:
1.628 + sse2fct=jsimd_h2v1_merged_upsample_sse2;
1.629 + mmxfct=jsimd_h2v1_merged_upsample_mmx;
1.630 + break;
1.631 + }
1.632 +
1.633 + if ((simd_support & JSIMD_SSE2) &&
1.634 + IS_ALIGNED_SSE(jconst_merged_upsample_sse2))
1.635 + sse2fct(cinfo->output_width, input_buf,
1.636 + in_row_group_ctr, output_buf);
1.637 + else if (simd_support & JSIMD_MMX)
1.638 + mmxfct(cinfo->output_width, input_buf,
1.639 + in_row_group_ctr, output_buf);
1.640 +}
1.641 +
1.642 +GLOBAL(int)
1.643 +jsimd_can_convsamp (void)
1.644 +{
1.645 + init_simd();
1.646 +
1.647 + /* The code is optimised for these values only */
1.648 + if (DCTSIZE != 8)
1.649 + return 0;
1.650 + if (BITS_IN_JSAMPLE != 8)
1.651 + return 0;
1.652 + if (sizeof(JDIMENSION) != 4)
1.653 + return 0;
1.654 + if (sizeof(DCTELEM) != 2)
1.655 + return 0;
1.656 +
1.657 + if (simd_support & JSIMD_SSE2)
1.658 + return 1;
1.659 + if (simd_support & JSIMD_MMX)
1.660 + return 1;
1.661 +
1.662 + return 0;
1.663 +}
1.664 +
1.665 +GLOBAL(int)
1.666 +jsimd_can_convsamp_float (void)
1.667 +{
1.668 + init_simd();
1.669 +
1.670 + /* The code is optimised for these values only */
1.671 + if (DCTSIZE != 8)
1.672 + return 0;
1.673 + if (BITS_IN_JSAMPLE != 8)
1.674 + return 0;
1.675 + if (sizeof(JDIMENSION) != 4)
1.676 + return 0;
1.677 + if (sizeof(FAST_FLOAT) != 4)
1.678 + return 0;
1.679 +
1.680 + if (simd_support & JSIMD_SSE2)
1.681 + return 1;
1.682 + if (simd_support & JSIMD_SSE)
1.683 + return 1;
1.684 + if (simd_support & JSIMD_3DNOW)
1.685 + return 1;
1.686 +
1.687 + return 0;
1.688 +}
1.689 +
1.690 +GLOBAL(void)
1.691 +jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col,
1.692 + DCTELEM * workspace)
1.693 +{
1.694 + if (simd_support & JSIMD_SSE2)
1.695 + jsimd_convsamp_sse2(sample_data, start_col, workspace);
1.696 + else if (simd_support & JSIMD_MMX)
1.697 + jsimd_convsamp_mmx(sample_data, start_col, workspace);
1.698 +}
1.699 +
1.700 +GLOBAL(void)
1.701 +jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col,
1.702 + FAST_FLOAT * workspace)
1.703 +{
1.704 + if (simd_support & JSIMD_SSE2)
1.705 + jsimd_convsamp_float_sse2(sample_data, start_col, workspace);
1.706 + else if (simd_support & JSIMD_SSE)
1.707 + jsimd_convsamp_float_sse(sample_data, start_col, workspace);
1.708 + else if (simd_support & JSIMD_3DNOW)
1.709 + jsimd_convsamp_float_3dnow(sample_data, start_col, workspace);
1.710 +}
1.711 +
1.712 +GLOBAL(int)
1.713 +jsimd_can_fdct_islow (void)
1.714 +{
1.715 + init_simd();
1.716 +
1.717 + /* The code is optimised for these values only */
1.718 + if (DCTSIZE != 8)
1.719 + return 0;
1.720 + if (sizeof(DCTELEM) != 2)
1.721 + return 0;
1.722 +
1.723 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
1.724 + return 1;
1.725 + if (simd_support & JSIMD_MMX)
1.726 + return 1;
1.727 +
1.728 + return 0;
1.729 +}
1.730 +
1.731 +GLOBAL(int)
1.732 +jsimd_can_fdct_ifast (void)
1.733 +{
1.734 + init_simd();
1.735 +
1.736 + /* The code is optimised for these values only */
1.737 + if (DCTSIZE != 8)
1.738 + return 0;
1.739 + if (sizeof(DCTELEM) != 2)
1.740 + return 0;
1.741 +
1.742 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2))
1.743 + return 1;
1.744 + if (simd_support & JSIMD_MMX)
1.745 + return 1;
1.746 +
1.747 + return 0;
1.748 +}
1.749 +
1.750 +GLOBAL(int)
1.751 +jsimd_can_fdct_float (void)
1.752 +{
1.753 + init_simd();
1.754 +
1.755 + /* The code is optimised for these values only */
1.756 + if (DCTSIZE != 8)
1.757 + return 0;
1.758 + if (sizeof(FAST_FLOAT) != 4)
1.759 + return 0;
1.760 +
1.761 + if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
1.762 + return 1;
1.763 + if (simd_support & JSIMD_3DNOW)
1.764 + return 1;
1.765 +
1.766 + return 0;
1.767 +}
1.768 +
1.769 +GLOBAL(void)
1.770 +jsimd_fdct_islow (DCTELEM * data)
1.771 +{
1.772 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
1.773 + jsimd_fdct_islow_sse2(data);
1.774 + else if (simd_support & JSIMD_MMX)
1.775 + jsimd_fdct_islow_mmx(data);
1.776 +}
1.777 +
1.778 +GLOBAL(void)
1.779 +jsimd_fdct_ifast (DCTELEM * data)
1.780 +{
1.781 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2))
1.782 + jsimd_fdct_ifast_sse2(data);
1.783 + else if (simd_support & JSIMD_MMX)
1.784 + jsimd_fdct_ifast_mmx(data);
1.785 +}
1.786 +
1.787 +GLOBAL(void)
1.788 +jsimd_fdct_float (FAST_FLOAT * data)
1.789 +{
1.790 + if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse))
1.791 + jsimd_fdct_float_sse(data);
1.792 + else if (simd_support & JSIMD_3DNOW)
1.793 + jsimd_fdct_float_3dnow(data);
1.794 +}
1.795 +
1.796 +GLOBAL(int)
1.797 +jsimd_can_quantize (void)
1.798 +{
1.799 + init_simd();
1.800 +
1.801 + /* The code is optimised for these values only */
1.802 + if (DCTSIZE != 8)
1.803 + return 0;
1.804 + if (sizeof(JCOEF) != 2)
1.805 + return 0;
1.806 + if (sizeof(DCTELEM) != 2)
1.807 + return 0;
1.808 +
1.809 + if (simd_support & JSIMD_SSE2)
1.810 + return 1;
1.811 + if (simd_support & JSIMD_MMX)
1.812 + return 1;
1.813 +
1.814 + return 0;
1.815 +}
1.816 +
1.817 +GLOBAL(int)
1.818 +jsimd_can_quantize_float (void)
1.819 +{
1.820 + init_simd();
1.821 +
1.822 + /* The code is optimised for these values only */
1.823 + if (DCTSIZE != 8)
1.824 + return 0;
1.825 + if (sizeof(JCOEF) != 2)
1.826 + return 0;
1.827 + if (sizeof(FAST_FLOAT) != 4)
1.828 + return 0;
1.829 +
1.830 + if (simd_support & JSIMD_SSE2)
1.831 + return 1;
1.832 + if (simd_support & JSIMD_SSE)
1.833 + return 1;
1.834 + if (simd_support & JSIMD_3DNOW)
1.835 + return 1;
1.836 +
1.837 + return 0;
1.838 +}
1.839 +
1.840 +GLOBAL(void)
1.841 +jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors,
1.842 + DCTELEM * workspace)
1.843 +{
1.844 + if (simd_support & JSIMD_SSE2)
1.845 + jsimd_quantize_sse2(coef_block, divisors, workspace);
1.846 + else if (simd_support & JSIMD_MMX)
1.847 + jsimd_quantize_mmx(coef_block, divisors, workspace);
1.848 +}
1.849 +
1.850 +GLOBAL(void)
1.851 +jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors,
1.852 + FAST_FLOAT * workspace)
1.853 +{
1.854 + if (simd_support & JSIMD_SSE2)
1.855 + jsimd_quantize_float_sse2(coef_block, divisors, workspace);
1.856 + else if (simd_support & JSIMD_SSE)
1.857 + jsimd_quantize_float_sse(coef_block, divisors, workspace);
1.858 + else if (simd_support & JSIMD_3DNOW)
1.859 + jsimd_quantize_float_3dnow(coef_block, divisors, workspace);
1.860 +}
1.861 +
1.862 +GLOBAL(int)
1.863 +jsimd_can_idct_2x2 (void)
1.864 +{
1.865 + init_simd();
1.866 +
1.867 + /* The code is optimised for these values only */
1.868 + if (DCTSIZE != 8)
1.869 + return 0;
1.870 + if (sizeof(JCOEF) != 2)
1.871 + return 0;
1.872 + if (BITS_IN_JSAMPLE != 8)
1.873 + return 0;
1.874 + if (sizeof(JDIMENSION) != 4)
1.875 + return 0;
1.876 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.877 + return 0;
1.878 +
1.879 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.880 + return 1;
1.881 + if (simd_support & JSIMD_MMX)
1.882 + return 1;
1.883 +
1.884 + return 0;
1.885 +}
1.886 +
1.887 +GLOBAL(int)
1.888 +jsimd_can_idct_4x4 (void)
1.889 +{
1.890 + init_simd();
1.891 +
1.892 + /* The code is optimised for these values only */
1.893 + if (DCTSIZE != 8)
1.894 + return 0;
1.895 + if (sizeof(JCOEF) != 2)
1.896 + return 0;
1.897 + if (BITS_IN_JSAMPLE != 8)
1.898 + return 0;
1.899 + if (sizeof(JDIMENSION) != 4)
1.900 + return 0;
1.901 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.902 + return 0;
1.903 +
1.904 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.905 + return 1;
1.906 + if (simd_support & JSIMD_MMX)
1.907 + return 1;
1.908 +
1.909 + return 0;
1.910 +}
1.911 +
1.912 +GLOBAL(void)
1.913 +jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.914 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.915 + JDIMENSION output_col)
1.916 +{
1.917 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.918 + jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.919 + else if (simd_support & JSIMD_MMX)
1.920 + jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col);
1.921 +}
1.922 +
1.923 +GLOBAL(void)
1.924 +jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.925 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.926 + JDIMENSION output_col)
1.927 +{
1.928 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2))
1.929 + jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.930 + else if (simd_support & JSIMD_MMX)
1.931 + jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col);
1.932 +}
1.933 +
1.934 +GLOBAL(int)
1.935 +jsimd_can_idct_islow (void)
1.936 +{
1.937 + init_simd();
1.938 +
1.939 + /* The code is optimised for these values only */
1.940 + if (DCTSIZE != 8)
1.941 + return 0;
1.942 + if (sizeof(JCOEF) != 2)
1.943 + return 0;
1.944 + if (BITS_IN_JSAMPLE != 8)
1.945 + return 0;
1.946 + if (sizeof(JDIMENSION) != 4)
1.947 + return 0;
1.948 + if (sizeof(ISLOW_MULT_TYPE) != 2)
1.949 + return 0;
1.950 +
1.951 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
1.952 + return 1;
1.953 + if (simd_support & JSIMD_MMX)
1.954 + return 1;
1.955 +
1.956 + return 0;
1.957 +}
1.958 +
1.959 +GLOBAL(int)
1.960 +jsimd_can_idct_ifast (void)
1.961 +{
1.962 + init_simd();
1.963 +
1.964 + /* The code is optimised for these values only */
1.965 + if (DCTSIZE != 8)
1.966 + return 0;
1.967 + if (sizeof(JCOEF) != 2)
1.968 + return 0;
1.969 + if (BITS_IN_JSAMPLE != 8)
1.970 + return 0;
1.971 + if (sizeof(JDIMENSION) != 4)
1.972 + return 0;
1.973 + if (sizeof(IFAST_MULT_TYPE) != 2)
1.974 + return 0;
1.975 + if (IFAST_SCALE_BITS != 2)
1.976 + return 0;
1.977 +
1.978 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
1.979 + return 1;
1.980 + if (simd_support & JSIMD_MMX)
1.981 + return 1;
1.982 +
1.983 + return 0;
1.984 +}
1.985 +
1.986 +GLOBAL(int)
1.987 +jsimd_can_idct_float (void)
1.988 +{
1.989 + init_simd();
1.990 +
1.991 + if (DCTSIZE != 8)
1.992 + return 0;
1.993 + if (sizeof(JCOEF) != 2)
1.994 + return 0;
1.995 + if (BITS_IN_JSAMPLE != 8)
1.996 + return 0;
1.997 + if (sizeof(JDIMENSION) != 4)
1.998 + return 0;
1.999 + if (sizeof(FAST_FLOAT) != 4)
1.1000 + return 0;
1.1001 + if (sizeof(FLOAT_MULT_TYPE) != 4)
1.1002 + return 0;
1.1003 +
1.1004 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
1.1005 + return 1;
1.1006 + if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
1.1007 + return 1;
1.1008 + if (simd_support & JSIMD_3DNOW)
1.1009 + return 1;
1.1010 +
1.1011 + return 0;
1.1012 +}
1.1013 +
1.1014 +GLOBAL(void)
1.1015 +jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.1016 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.1017 + JDIMENSION output_col)
1.1018 +{
1.1019 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2))
1.1020 + jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.1021 + else if (simd_support & JSIMD_MMX)
1.1022 + jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col);
1.1023 +}
1.1024 +
1.1025 +GLOBAL(void)
1.1026 +jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.1027 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.1028 + JDIMENSION output_col)
1.1029 +{
1.1030 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2))
1.1031 + jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col);
1.1032 + else if (simd_support & JSIMD_MMX)
1.1033 + jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col);
1.1034 +}
1.1035 +
1.1036 +GLOBAL(void)
1.1037 +jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr,
1.1038 + JCOEFPTR coef_block, JSAMPARRAY output_buf,
1.1039 + JDIMENSION output_col)
1.1040 +{
1.1041 + if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2))
1.1042 + jsimd_idct_float_sse2(compptr->dct_table, coef_block,
1.1043 + output_buf, output_col);
1.1044 + else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse))
1.1045 + jsimd_idct_float_sse(compptr->dct_table, coef_block,
1.1046 + output_buf, output_col);
1.1047 + else if (simd_support & JSIMD_3DNOW)
1.1048 + jsimd_idct_float_3dnow(compptr->dct_table, coef_block,
1.1049 + output_buf, output_col);
1.1050 +}
1.1051 +
