Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 1 | /* |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 2 | * jsimd_i386.c |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 3 | * |
| 4 | * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 5 | * Copyright 2009 D. R. Commander |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 6 | * |
| 7 | * Based on the x86 SIMD extension for IJG JPEG library, |
| 8 | * Copyright (C) 1999-2006, MIYASAKA Masaru. |
DRC | a6da9f3 | 2011-02-02 05:45:43 +0000 | [diff] [blame] | 9 | * For conditions of distribution and use, see copyright notice in jsimdext.inc |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 10 | * |
| 11 | * This file contains the interface between the "normal" portions |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 12 | * of the library and the SIMD implementations when running on a |
| 13 | * 32-bit x86 architecture. |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 14 | */ |
| 15 | |
| 16 | #define JPEG_INTERNALS |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 17 | #include "../jinclude.h" |
| 18 | #include "../jpeglib.h" |
| 19 | #include "../jsimd.h" |
| 20 | #include "../jdct.h" |
| 21 | #include "../jsimddct.h" |
Peter Åstrand | 086f2a0 | 2009-08-19 13:53:48 +0000 | [diff] [blame] | 22 | #include "jsimd.h" |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 23 | |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 24 | /* |
| 25 | * In the PIC cases, we have no guarantee that constants will keep |
| 26 | * their alignment. This macro allows us to verify it at runtime. |
| 27 | */ |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 28 | #define IS_ALIGNED(ptr, order) (((unsigned)ptr & ((1 << order) - 1)) == 0) |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 29 | |
| 30 | #define IS_ALIGNED_SSE(ptr) (IS_ALIGNED(ptr, 4)) /* 16 byte alignment */ |
| 31 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 32 | static unsigned int simd_support = ~0; |
| 33 | |
| 34 | /* |
| 35 | * Check what SIMD accelerations are supported. |
| 36 | * |
| 37 | * FIXME: This code is racy under a multi-threaded environment. |
| 38 | */ |
| 39 | LOCAL(void) |
| 40 | init_simd (void) |
| 41 | { |
DRC | bec58d8 | 2009-04-03 11:27:17 +0000 | [diff] [blame] | 42 | char *env = NULL; |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 43 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 44 | if (simd_support != ~0) |
| 45 | return; |
| 46 | |
Pierre Ossman | 82c7f31 | 2009-03-09 13:21:27 +0000 | [diff] [blame] | 47 | simd_support = jpeg_simd_cpu_support(); |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 48 | |
| 49 | /* Force different settings through environment variables */ |
| 50 | env = getenv("JSIMD_FORCEMMX"); |
| 51 | if ((env != NULL) && (strcmp(env, "1") == 0)) |
DRC | f8ad8ff | 2010-03-03 08:48:44 +0000 | [diff] [blame] | 52 | simd_support &= JSIMD_MMX; |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 53 | env = getenv("JSIMD_FORCE3DNOW"); |
| 54 | if ((env != NULL) && (strcmp(env, "1") == 0)) |
DRC | f8ad8ff | 2010-03-03 08:48:44 +0000 | [diff] [blame] | 55 | simd_support &= JSIMD_3DNOW|JSIMD_MMX; |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 56 | env = getenv("JSIMD_FORCESSE"); |
| 57 | if ((env != NULL) && (strcmp(env, "1") == 0)) |
DRC | f8ad8ff | 2010-03-03 08:48:44 +0000 | [diff] [blame] | 58 | simd_support &= JSIMD_SSE|JSIMD_MMX; |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 59 | env = getenv("JSIMD_FORCESSE2"); |
| 60 | if ((env != NULL) && (strcmp(env, "1") == 0)) |
DRC | f8ad8ff | 2010-03-03 08:48:44 +0000 | [diff] [blame] | 61 | simd_support &= JSIMD_SSE2; |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 62 | } |
| 63 | |
| 64 | GLOBAL(int) |
| 65 | jsimd_can_rgb_ycc (void) |
| 66 | { |
| 67 | init_simd(); |
| 68 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 69 | /* The code is optimised for these values only */ |
| 70 | if (BITS_IN_JSAMPLE != 8) |
| 71 | return 0; |
| 72 | if (sizeof(JDIMENSION) != 4) |
| 73 | return 0; |
| 74 | if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
| 75 | return 0; |
| 76 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 77 | if ((simd_support & JSIMD_SSE2) && |
| 78 | IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2)) |
| 79 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 80 | if (simd_support & JSIMD_MMX) |
| 81 | return 1; |
| 82 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 83 | return 0; |
| 84 | } |
| 85 | |
| 86 | GLOBAL(int) |
| 87 | jsimd_can_ycc_rgb (void) |
| 88 | { |
| 89 | init_simd(); |
| 90 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 91 | /* The code is optimised for these values only */ |
| 92 | if (BITS_IN_JSAMPLE != 8) |
| 93 | return 0; |
| 94 | if (sizeof(JDIMENSION) != 4) |
| 95 | return 0; |
| 96 | if ((RGB_PIXELSIZE != 3) && (RGB_PIXELSIZE != 4)) |
| 97 | return 0; |
| 98 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 99 | if ((simd_support & JSIMD_SSE2) && |
| 100 | IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2)) |
| 101 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 102 | if (simd_support & JSIMD_MMX) |
| 103 | return 1; |
| 104 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 105 | return 0; |
| 106 | } |
| 107 | |
| 108 | GLOBAL(void) |
| 109 | jsimd_rgb_ycc_convert (j_compress_ptr cinfo, |
| 110 | JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
| 111 | JDIMENSION output_row, int num_rows) |
| 112 | { |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 113 | void (*sse2fct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
| 114 | void (*mmxfct)(JDIMENSION, JSAMPARRAY, JSAMPIMAGE, JDIMENSION, int); |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 115 | |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 116 | switch(cinfo->in_color_space) |
| 117 | { |
| 118 | case JCS_EXT_RGB: |
| 119 | sse2fct=jsimd_extrgb_ycc_convert_sse2; |
| 120 | mmxfct=jsimd_extrgb_ycc_convert_mmx; |
| 121 | break; |
| 122 | case JCS_EXT_RGBX: |
| 123 | sse2fct=jsimd_extrgbx_ycc_convert_sse2; |
| 124 | mmxfct=jsimd_extrgbx_ycc_convert_mmx; |
| 125 | break; |
| 126 | case JCS_EXT_BGR: |
| 127 | sse2fct=jsimd_extbgr_ycc_convert_sse2; |
| 128 | mmxfct=jsimd_extbgr_ycc_convert_mmx; |
| 129 | break; |
| 130 | case JCS_EXT_BGRX: |
| 131 | sse2fct=jsimd_extbgrx_ycc_convert_sse2; |
| 132 | mmxfct=jsimd_extbgrx_ycc_convert_mmx; |
| 133 | break; |
| 134 | case JCS_EXT_XBGR: |
| 135 | sse2fct=jsimd_extxbgr_ycc_convert_sse2; |
| 136 | mmxfct=jsimd_extxbgr_ycc_convert_mmx; |
| 137 | break; |
| 138 | case JCS_EXT_XRGB: |
| 139 | sse2fct=jsimd_extxrgb_ycc_convert_sse2; |
| 140 | mmxfct=jsimd_extxrgb_ycc_convert_mmx; |
| 141 | break; |
| 142 | default: |
| 143 | sse2fct=jsimd_rgb_ycc_convert_sse2; |
| 144 | mmxfct=jsimd_rgb_ycc_convert_mmx; |
| 145 | break; |
| 146 | } |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 147 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 148 | if ((simd_support & JSIMD_SSE2) && |
| 149 | IS_ALIGNED_SSE(jconst_rgb_ycc_convert_sse2)) |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 150 | sse2fct(cinfo->image_width, input_buf, |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 151 | output_buf, output_row, num_rows); |
| 152 | else if (simd_support & JSIMD_MMX) |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 153 | mmxfct(cinfo->image_width, input_buf, |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 154 | output_buf, output_row, num_rows); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 155 | } |
| 156 | |
| 157 | GLOBAL(void) |
| 158 | jsimd_ycc_rgb_convert (j_decompress_ptr cinfo, |
| 159 | JSAMPIMAGE input_buf, JDIMENSION input_row, |
| 160 | JSAMPARRAY output_buf, int num_rows) |
| 161 | { |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 162 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); |
| 163 | void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY, int); |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 164 | |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 165 | switch(cinfo->out_color_space) |
| 166 | { |
| 167 | case JCS_EXT_RGB: |
| 168 | sse2fct=jsimd_ycc_extrgb_convert_sse2; |
| 169 | mmxfct=jsimd_ycc_extrgb_convert_mmx; |
| 170 | break; |
| 171 | case JCS_EXT_RGBX: |
| 172 | sse2fct=jsimd_ycc_extrgbx_convert_sse2; |
| 173 | mmxfct=jsimd_ycc_extrgbx_convert_mmx; |
| 174 | break; |
| 175 | case JCS_EXT_BGR: |
| 176 | sse2fct=jsimd_ycc_extbgr_convert_sse2; |
| 177 | mmxfct=jsimd_ycc_extbgr_convert_mmx; |
| 178 | break; |
| 179 | case JCS_EXT_BGRX: |
| 180 | sse2fct=jsimd_ycc_extbgrx_convert_sse2; |
| 181 | mmxfct=jsimd_ycc_extbgrx_convert_mmx; |
| 182 | break; |
| 183 | case JCS_EXT_XBGR: |
| 184 | sse2fct=jsimd_ycc_extxbgr_convert_sse2; |
| 185 | mmxfct=jsimd_ycc_extxbgr_convert_mmx; |
| 186 | break; |
| 187 | case JCS_EXT_XRGB: |
| 188 | sse2fct=jsimd_ycc_extxrgb_convert_sse2; |
| 189 | mmxfct=jsimd_ycc_extxrgb_convert_mmx; |
| 190 | break; |
| 191 | default: |
| 192 | sse2fct=jsimd_ycc_rgb_convert_sse2; |
| 193 | mmxfct=jsimd_ycc_rgb_convert_mmx; |
| 194 | break; |
| 195 | } |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 196 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 197 | if ((simd_support & JSIMD_SSE2) && |
| 198 | IS_ALIGNED_SSE(jconst_ycc_rgb_convert_sse2)) |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 199 | sse2fct(cinfo->output_width, input_buf, |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 200 | input_row, output_buf, num_rows); |
| 201 | else if (simd_support & JSIMD_MMX) |
DRC | 8ca81ec | 2009-04-03 12:00:51 +0000 | [diff] [blame] | 202 | mmxfct(cinfo->output_width, input_buf, |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 203 | input_row, output_buf, num_rows); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 204 | } |
| 205 | |
| 206 | GLOBAL(int) |
| 207 | jsimd_can_h2v2_downsample (void) |
| 208 | { |
| 209 | init_simd(); |
| 210 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 211 | /* The code is optimised for these values only */ |
| 212 | if (BITS_IN_JSAMPLE != 8) |
| 213 | return 0; |
| 214 | if (sizeof(JDIMENSION) != 4) |
| 215 | return 0; |
| 216 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 217 | if (simd_support & JSIMD_SSE2) |
| 218 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 219 | if (simd_support & JSIMD_MMX) |
| 220 | return 1; |
| 221 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 222 | return 0; |
| 223 | } |
| 224 | |
| 225 | GLOBAL(int) |
| 226 | jsimd_can_h2v1_downsample (void) |
| 227 | { |
| 228 | init_simd(); |
| 229 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 230 | /* The code is optimised for these values only */ |
| 231 | if (BITS_IN_JSAMPLE != 8) |
| 232 | return 0; |
| 233 | if (sizeof(JDIMENSION) != 4) |
| 234 | return 0; |
| 235 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 236 | if (simd_support & JSIMD_SSE2) |
| 237 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 238 | if (simd_support & JSIMD_MMX) |
| 239 | return 1; |
| 240 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 241 | return 0; |
| 242 | } |
| 243 | |
| 244 | GLOBAL(void) |
| 245 | jsimd_h2v2_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, |
| 246 | JSAMPARRAY input_data, JSAMPARRAY output_data) |
| 247 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 248 | if (simd_support & JSIMD_SSE2) |
| 249 | jsimd_h2v2_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, |
| 250 | compptr->v_samp_factor, compptr->width_in_blocks, |
| 251 | input_data, output_data); |
| 252 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 253 | jsimd_h2v2_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor, |
| 254 | compptr->v_samp_factor, compptr->width_in_blocks, |
| 255 | input_data, output_data); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 256 | } |
| 257 | |
| 258 | GLOBAL(void) |
| 259 | jsimd_h2v1_downsample (j_compress_ptr cinfo, jpeg_component_info * compptr, |
| 260 | JSAMPARRAY input_data, JSAMPARRAY output_data) |
| 261 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 262 | if (simd_support & JSIMD_SSE2) |
| 263 | jsimd_h2v1_downsample_sse2(cinfo->image_width, cinfo->max_v_samp_factor, |
| 264 | compptr->v_samp_factor, compptr->width_in_blocks, |
| 265 | input_data, output_data); |
| 266 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 267 | jsimd_h2v1_downsample_mmx(cinfo->image_width, cinfo->max_v_samp_factor, |
| 268 | compptr->v_samp_factor, compptr->width_in_blocks, |
| 269 | input_data, output_data); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 270 | } |
| 271 | |
| 272 | GLOBAL(int) |
| 273 | jsimd_can_h2v2_upsample (void) |
| 274 | { |
| 275 | init_simd(); |
| 276 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 277 | /* The code is optimised for these values only */ |
| 278 | if (BITS_IN_JSAMPLE != 8) |
| 279 | return 0; |
| 280 | if (sizeof(JDIMENSION) != 4) |
| 281 | return 0; |
| 282 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 283 | if (simd_support & JSIMD_SSE2) |
| 284 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 285 | if (simd_support & JSIMD_MMX) |
| 286 | return 1; |
| 287 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 288 | return 0; |
| 289 | } |
| 290 | |
| 291 | GLOBAL(int) |
| 292 | jsimd_can_h2v1_upsample (void) |
| 293 | { |
| 294 | init_simd(); |
| 295 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 296 | /* The code is optimised for these values only */ |
| 297 | if (BITS_IN_JSAMPLE != 8) |
| 298 | return 0; |
| 299 | if (sizeof(JDIMENSION) != 4) |
| 300 | return 0; |
| 301 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 302 | if (simd_support & JSIMD_SSE2) |
| 303 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 304 | if (simd_support & JSIMD_MMX) |
| 305 | return 1; |
| 306 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 307 | return 0; |
| 308 | } |
| 309 | |
| 310 | GLOBAL(void) |
| 311 | jsimd_h2v2_upsample (j_decompress_ptr cinfo, |
| 312 | jpeg_component_info * compptr, |
| 313 | JSAMPARRAY input_data, |
| 314 | JSAMPARRAY * output_data_ptr) |
| 315 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 316 | if (simd_support & JSIMD_SSE2) |
| 317 | jsimd_h2v2_upsample_sse2(cinfo->max_v_samp_factor, |
| 318 | cinfo->output_width, input_data, output_data_ptr); |
| 319 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 320 | jsimd_h2v2_upsample_mmx(cinfo->max_v_samp_factor, |
| 321 | cinfo->output_width, input_data, output_data_ptr); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 322 | } |
| 323 | |
| 324 | GLOBAL(void) |
| 325 | jsimd_h2v1_upsample (j_decompress_ptr cinfo, |
| 326 | jpeg_component_info * compptr, |
| 327 | JSAMPARRAY input_data, |
| 328 | JSAMPARRAY * output_data_ptr) |
| 329 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 330 | if (simd_support & JSIMD_SSE2) |
| 331 | jsimd_h2v1_upsample_sse2(cinfo->max_v_samp_factor, |
| 332 | cinfo->output_width, input_data, output_data_ptr); |
| 333 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 334 | jsimd_h2v1_upsample_mmx(cinfo->max_v_samp_factor, |
| 335 | cinfo->output_width, input_data, output_data_ptr); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 336 | } |
| 337 | |
| 338 | GLOBAL(int) |
| 339 | jsimd_can_h2v2_fancy_upsample (void) |
| 340 | { |
| 341 | init_simd(); |
| 342 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 343 | /* The code is optimised for these values only */ |
| 344 | if (BITS_IN_JSAMPLE != 8) |
| 345 | return 0; |
| 346 | if (sizeof(JDIMENSION) != 4) |
| 347 | return 0; |
| 348 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 349 | if ((simd_support & JSIMD_SSE2) && |
| 350 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
| 351 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 352 | if (simd_support & JSIMD_MMX) |
| 353 | return 1; |
| 354 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 355 | return 0; |
| 356 | } |
| 357 | |
| 358 | GLOBAL(int) |
| 359 | jsimd_can_h2v1_fancy_upsample (void) |
| 360 | { |
| 361 | init_simd(); |
| 362 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 363 | /* The code is optimised for these values only */ |
| 364 | if (BITS_IN_JSAMPLE != 8) |
| 365 | return 0; |
| 366 | if (sizeof(JDIMENSION) != 4) |
| 367 | return 0; |
| 368 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 369 | if ((simd_support & JSIMD_SSE2) && |
| 370 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
| 371 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 372 | if (simd_support & JSIMD_MMX) |
| 373 | return 1; |
| 374 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 375 | return 0; |
| 376 | } |
| 377 | |
| 378 | GLOBAL(void) |
| 379 | jsimd_h2v2_fancy_upsample (j_decompress_ptr cinfo, |
| 380 | jpeg_component_info * compptr, |
| 381 | JSAMPARRAY input_data, |
| 382 | JSAMPARRAY * output_data_ptr) |
| 383 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 384 | if ((simd_support & JSIMD_SSE2) && |
| 385 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
DRC | bacbbaa | 2010-08-07 16:12:08 +0000 | [diff] [blame] | 386 | jsimd_h2v2_fancy_upsample_sse2(cinfo->max_v_samp_factor, |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 387 | compptr->downsampled_width, input_data, output_data_ptr); |
| 388 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 389 | jsimd_h2v2_fancy_upsample_mmx(cinfo->max_v_samp_factor, |
| 390 | compptr->downsampled_width, input_data, output_data_ptr); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 391 | } |
| 392 | |
| 393 | GLOBAL(void) |
| 394 | jsimd_h2v1_fancy_upsample (j_decompress_ptr cinfo, |
| 395 | jpeg_component_info * compptr, |
| 396 | JSAMPARRAY input_data, |
| 397 | JSAMPARRAY * output_data_ptr) |
| 398 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 399 | if ((simd_support & JSIMD_SSE2) && |
| 400 | IS_ALIGNED_SSE(jconst_fancy_upsample_sse2)) |
| 401 | jsimd_h2v1_fancy_upsample_sse2(cinfo->max_v_samp_factor, |
| 402 | compptr->downsampled_width, input_data, output_data_ptr); |
| 403 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 404 | jsimd_h2v1_fancy_upsample_mmx(cinfo->max_v_samp_factor, |
| 405 | compptr->downsampled_width, input_data, output_data_ptr); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 406 | } |
| 407 | |
| 408 | GLOBAL(int) |
| 409 | jsimd_can_h2v2_merged_upsample (void) |
| 410 | { |
| 411 | init_simd(); |
| 412 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 413 | /* The code is optimised for these values only */ |
| 414 | if (BITS_IN_JSAMPLE != 8) |
| 415 | return 0; |
| 416 | if (sizeof(JDIMENSION) != 4) |
| 417 | return 0; |
| 418 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 419 | if ((simd_support & JSIMD_SSE2) && |
| 420 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
| 421 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 422 | if (simd_support & JSIMD_MMX) |
| 423 | return 1; |
| 424 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 425 | return 0; |
| 426 | } |
| 427 | |
| 428 | GLOBAL(int) |
| 429 | jsimd_can_h2v1_merged_upsample (void) |
| 430 | { |
| 431 | init_simd(); |
| 432 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 433 | /* The code is optimised for these values only */ |
| 434 | if (BITS_IN_JSAMPLE != 8) |
| 435 | return 0; |
| 436 | if (sizeof(JDIMENSION) != 4) |
| 437 | return 0; |
| 438 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 439 | if ((simd_support & JSIMD_SSE2) && |
| 440 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
| 441 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 442 | if (simd_support & JSIMD_MMX) |
| 443 | return 1; |
| 444 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 445 | return 0; |
| 446 | } |
| 447 | |
| 448 | GLOBAL(void) |
| 449 | jsimd_h2v2_merged_upsample (j_decompress_ptr cinfo, |
| 450 | JSAMPIMAGE input_buf, |
| 451 | JDIMENSION in_row_group_ctr, |
| 452 | JSAMPARRAY output_buf) |
| 453 | { |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 454 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
| 455 | void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 456 | |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 457 | switch(cinfo->out_color_space) |
| 458 | { |
| 459 | case JCS_EXT_RGB: |
| 460 | sse2fct=jsimd_h2v2_extrgb_merged_upsample_sse2; |
| 461 | mmxfct=jsimd_h2v2_extrgb_merged_upsample_mmx; |
| 462 | break; |
| 463 | case JCS_EXT_RGBX: |
| 464 | sse2fct=jsimd_h2v2_extrgbx_merged_upsample_sse2; |
| 465 | mmxfct=jsimd_h2v2_extrgbx_merged_upsample_mmx; |
| 466 | break; |
| 467 | case JCS_EXT_BGR: |
| 468 | sse2fct=jsimd_h2v2_extbgr_merged_upsample_sse2; |
| 469 | mmxfct=jsimd_h2v2_extbgr_merged_upsample_mmx; |
| 470 | break; |
| 471 | case JCS_EXT_BGRX: |
| 472 | sse2fct=jsimd_h2v2_extbgrx_merged_upsample_sse2; |
| 473 | mmxfct=jsimd_h2v2_extbgrx_merged_upsample_mmx; |
| 474 | break; |
| 475 | case JCS_EXT_XBGR: |
| 476 | sse2fct=jsimd_h2v2_extxbgr_merged_upsample_sse2; |
| 477 | mmxfct=jsimd_h2v2_extxbgr_merged_upsample_mmx; |
| 478 | break; |
| 479 | case JCS_EXT_XRGB: |
| 480 | sse2fct=jsimd_h2v2_extxrgb_merged_upsample_sse2; |
| 481 | mmxfct=jsimd_h2v2_extxrgb_merged_upsample_mmx; |
| 482 | break; |
| 483 | default: |
| 484 | sse2fct=jsimd_h2v2_merged_upsample_sse2; |
| 485 | mmxfct=jsimd_h2v2_merged_upsample_mmx; |
| 486 | break; |
| 487 | } |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 488 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 489 | if ((simd_support & JSIMD_SSE2) && |
| 490 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 491 | sse2fct(cinfo->output_width, input_buf, |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 492 | in_row_group_ctr, output_buf); |
| 493 | else if (simd_support & JSIMD_MMX) |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 494 | mmxfct(cinfo->output_width, input_buf, |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 495 | in_row_group_ctr, output_buf); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 496 | } |
| 497 | |
| 498 | GLOBAL(void) |
| 499 | jsimd_h2v1_merged_upsample (j_decompress_ptr cinfo, |
| 500 | JSAMPIMAGE input_buf, |
| 501 | JDIMENSION in_row_group_ctr, |
| 502 | JSAMPARRAY output_buf) |
| 503 | { |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 504 | void (*sse2fct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
| 505 | void (*mmxfct)(JDIMENSION, JSAMPIMAGE, JDIMENSION, JSAMPARRAY); |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 506 | |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 507 | switch(cinfo->out_color_space) |
| 508 | { |
| 509 | case JCS_EXT_RGB: |
| 510 | sse2fct=jsimd_h2v1_extrgb_merged_upsample_sse2; |
| 511 | mmxfct=jsimd_h2v1_extrgb_merged_upsample_mmx; |
| 512 | break; |
| 513 | case JCS_EXT_RGBX: |
| 514 | sse2fct=jsimd_h2v1_extrgbx_merged_upsample_sse2; |
| 515 | mmxfct=jsimd_h2v1_extrgbx_merged_upsample_mmx; |
| 516 | break; |
| 517 | case JCS_EXT_BGR: |
| 518 | sse2fct=jsimd_h2v1_extbgr_merged_upsample_sse2; |
| 519 | mmxfct=jsimd_h2v1_extbgr_merged_upsample_mmx; |
| 520 | break; |
| 521 | case JCS_EXT_BGRX: |
| 522 | sse2fct=jsimd_h2v1_extbgrx_merged_upsample_sse2; |
| 523 | mmxfct=jsimd_h2v1_extbgrx_merged_upsample_mmx; |
| 524 | break; |
| 525 | case JCS_EXT_XBGR: |
| 526 | sse2fct=jsimd_h2v1_extxbgr_merged_upsample_sse2; |
| 527 | mmxfct=jsimd_h2v1_extxbgr_merged_upsample_mmx; |
| 528 | break; |
| 529 | case JCS_EXT_XRGB: |
| 530 | sse2fct=jsimd_h2v1_extxrgb_merged_upsample_sse2; |
| 531 | mmxfct=jsimd_h2v1_extxrgb_merged_upsample_mmx; |
| 532 | break; |
| 533 | default: |
| 534 | sse2fct=jsimd_h2v1_merged_upsample_sse2; |
| 535 | mmxfct=jsimd_h2v1_merged_upsample_mmx; |
| 536 | break; |
| 537 | } |
Pierre Ossman | 0b7301e | 2009-06-29 11:20:42 +0000 | [diff] [blame] | 538 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 539 | if ((simd_support & JSIMD_SSE2) && |
| 540 | IS_ALIGNED_SSE(jconst_merged_upsample_sse2)) |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 541 | sse2fct(cinfo->output_width, input_buf, |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 542 | in_row_group_ctr, output_buf); |
| 543 | else if (simd_support & JSIMD_MMX) |
DRC | 80bc60b | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 544 | mmxfct(cinfo->output_width, input_buf, |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 545 | in_row_group_ctr, output_buf); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 546 | } |
| 547 | |
| 548 | GLOBAL(int) |
| 549 | jsimd_can_convsamp (void) |
| 550 | { |
| 551 | init_simd(); |
| 552 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 553 | /* The code is optimised for these values only */ |
| 554 | if (DCTSIZE != 8) |
| 555 | return 0; |
| 556 | if (BITS_IN_JSAMPLE != 8) |
| 557 | return 0; |
| 558 | if (sizeof(JDIMENSION) != 4) |
| 559 | return 0; |
| 560 | if (sizeof(DCTELEM) != 2) |
| 561 | return 0; |
| 562 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 563 | if (simd_support & JSIMD_SSE2) |
| 564 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 565 | if (simd_support & JSIMD_MMX) |
| 566 | return 1; |
| 567 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 568 | return 0; |
| 569 | } |
| 570 | |
| 571 | GLOBAL(int) |
| 572 | jsimd_can_convsamp_float (void) |
| 573 | { |
| 574 | init_simd(); |
| 575 | |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 576 | /* The code is optimised for these values only */ |
| 577 | if (DCTSIZE != 8) |
| 578 | return 0; |
| 579 | if (BITS_IN_JSAMPLE != 8) |
| 580 | return 0; |
| 581 | if (sizeof(JDIMENSION) != 4) |
| 582 | return 0; |
| 583 | if (sizeof(FAST_FLOAT) != 4) |
| 584 | return 0; |
| 585 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 586 | if (simd_support & JSIMD_SSE2) |
| 587 | return 1; |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 588 | if (simd_support & JSIMD_SSE) |
| 589 | return 1; |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 590 | if (simd_support & JSIMD_3DNOW) |
| 591 | return 1; |
| 592 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 593 | return 0; |
| 594 | } |
| 595 | |
| 596 | GLOBAL(void) |
| 597 | jsimd_convsamp (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 598 | DCTELEM * workspace) |
| 599 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 600 | if (simd_support & JSIMD_SSE2) |
| 601 | jsimd_convsamp_sse2(sample_data, start_col, workspace); |
| 602 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 603 | jsimd_convsamp_mmx(sample_data, start_col, workspace); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 604 | } |
| 605 | |
| 606 | GLOBAL(void) |
| 607 | jsimd_convsamp_float (JSAMPARRAY sample_data, JDIMENSION start_col, |
| 608 | FAST_FLOAT * workspace) |
| 609 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 610 | if (simd_support & JSIMD_SSE2) |
| 611 | jsimd_convsamp_float_sse2(sample_data, start_col, workspace); |
| 612 | else if (simd_support & JSIMD_SSE) |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 613 | jsimd_convsamp_float_sse(sample_data, start_col, workspace); |
| 614 | else if (simd_support & JSIMD_3DNOW) |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 615 | jsimd_convsamp_float_3dnow(sample_data, start_col, workspace); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 616 | } |
| 617 | |
| 618 | GLOBAL(int) |
| 619 | jsimd_can_fdct_islow (void) |
| 620 | { |
| 621 | init_simd(); |
| 622 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 623 | /* The code is optimised for these values only */ |
| 624 | if (DCTSIZE != 8) |
| 625 | return 0; |
| 626 | if (sizeof(DCTELEM) != 2) |
| 627 | return 0; |
| 628 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 629 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) |
| 630 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 631 | if (simd_support & JSIMD_MMX) |
| 632 | return 1; |
| 633 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 634 | return 0; |
| 635 | } |
| 636 | |
| 637 | GLOBAL(int) |
| 638 | jsimd_can_fdct_ifast (void) |
| 639 | { |
| 640 | init_simd(); |
| 641 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 642 | /* The code is optimised for these values only */ |
| 643 | if (DCTSIZE != 8) |
| 644 | return 0; |
| 645 | if (sizeof(DCTELEM) != 2) |
| 646 | return 0; |
| 647 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 648 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_ifast_sse2)) |
| 649 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 650 | if (simd_support & JSIMD_MMX) |
| 651 | return 1; |
| 652 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 653 | return 0; |
| 654 | } |
| 655 | |
| 656 | GLOBAL(int) |
| 657 | jsimd_can_fdct_float (void) |
| 658 | { |
| 659 | init_simd(); |
| 660 | |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 661 | /* The code is optimised for these values only */ |
| 662 | if (DCTSIZE != 8) |
| 663 | return 0; |
| 664 | if (sizeof(FAST_FLOAT) != 4) |
| 665 | return 0; |
| 666 | |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 667 | if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse)) |
| 668 | return 1; |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 669 | if (simd_support & JSIMD_3DNOW) |
| 670 | return 1; |
| 671 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 672 | return 0; |
| 673 | } |
| 674 | |
| 675 | GLOBAL(void) |
| 676 | jsimd_fdct_islow (DCTELEM * data) |
| 677 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 678 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) |
| 679 | jsimd_fdct_islow_sse2(data); |
| 680 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 681 | jsimd_fdct_islow_mmx(data); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 682 | } |
| 683 | |
| 684 | GLOBAL(void) |
| 685 | jsimd_fdct_ifast (DCTELEM * data) |
| 686 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 687 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_fdct_islow_sse2)) |
| 688 | jsimd_fdct_ifast_sse2(data); |
| 689 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 690 | jsimd_fdct_ifast_mmx(data); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 691 | } |
| 692 | |
| 693 | GLOBAL(void) |
| 694 | jsimd_fdct_float (FAST_FLOAT * data) |
| 695 | { |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 696 | if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_fdct_float_sse)) |
| 697 | jsimd_fdct_float_sse(data); |
| 698 | else if (simd_support & JSIMD_3DNOW) |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 699 | jsimd_fdct_float_3dnow(data); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 700 | } |
| 701 | |
| 702 | GLOBAL(int) |
| 703 | jsimd_can_quantize (void) |
| 704 | { |
| 705 | init_simd(); |
| 706 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 707 | /* The code is optimised for these values only */ |
| 708 | if (DCTSIZE != 8) |
| 709 | return 0; |
| 710 | if (sizeof(JCOEF) != 2) |
| 711 | return 0; |
| 712 | if (sizeof(DCTELEM) != 2) |
| 713 | return 0; |
| 714 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 715 | if (simd_support & JSIMD_SSE2) |
| 716 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 717 | if (simd_support & JSIMD_MMX) |
| 718 | return 1; |
| 719 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 720 | return 0; |
| 721 | } |
| 722 | |
| 723 | GLOBAL(int) |
| 724 | jsimd_can_quantize_float (void) |
| 725 | { |
| 726 | init_simd(); |
| 727 | |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 728 | /* The code is optimised for these values only */ |
| 729 | if (DCTSIZE != 8) |
| 730 | return 0; |
| 731 | if (sizeof(JCOEF) != 2) |
| 732 | return 0; |
| 733 | if (sizeof(FAST_FLOAT) != 4) |
| 734 | return 0; |
| 735 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 736 | if (simd_support & JSIMD_SSE2) |
| 737 | return 1; |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 738 | if (simd_support & JSIMD_SSE) |
| 739 | return 1; |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 740 | if (simd_support & JSIMD_3DNOW) |
| 741 | return 1; |
| 742 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 743 | return 0; |
| 744 | } |
| 745 | |
| 746 | GLOBAL(void) |
| 747 | jsimd_quantize (JCOEFPTR coef_block, DCTELEM * divisors, |
| 748 | DCTELEM * workspace) |
| 749 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 750 | if (simd_support & JSIMD_SSE2) |
| 751 | jsimd_quantize_sse2(coef_block, divisors, workspace); |
| 752 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 753 | jsimd_quantize_mmx(coef_block, divisors, workspace); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 754 | } |
| 755 | |
| 756 | GLOBAL(void) |
| 757 | jsimd_quantize_float (JCOEFPTR coef_block, FAST_FLOAT * divisors, |
| 758 | FAST_FLOAT * workspace) |
| 759 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 760 | if (simd_support & JSIMD_SSE2) |
| 761 | jsimd_quantize_float_sse2(coef_block, divisors, workspace); |
| 762 | else if (simd_support & JSIMD_SSE) |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 763 | jsimd_quantize_float_sse(coef_block, divisors, workspace); |
| 764 | else if (simd_support & JSIMD_3DNOW) |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 765 | jsimd_quantize_float_3dnow(coef_block, divisors, workspace); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 766 | } |
| 767 | |
| 768 | GLOBAL(int) |
| 769 | jsimd_can_idct_2x2 (void) |
| 770 | { |
| 771 | init_simd(); |
| 772 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 773 | /* The code is optimised for these values only */ |
| 774 | if (DCTSIZE != 8) |
| 775 | return 0; |
| 776 | if (sizeof(JCOEF) != 2) |
| 777 | return 0; |
| 778 | if (BITS_IN_JSAMPLE != 8) |
| 779 | return 0; |
| 780 | if (sizeof(JDIMENSION) != 4) |
| 781 | return 0; |
| 782 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 783 | return 0; |
| 784 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 785 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
| 786 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 787 | if (simd_support & JSIMD_MMX) |
| 788 | return 1; |
| 789 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 790 | return 0; |
| 791 | } |
| 792 | |
| 793 | GLOBAL(int) |
| 794 | jsimd_can_idct_4x4 (void) |
| 795 | { |
| 796 | init_simd(); |
| 797 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 798 | /* The code is optimised for these values only */ |
| 799 | if (DCTSIZE != 8) |
| 800 | return 0; |
| 801 | if (sizeof(JCOEF) != 2) |
| 802 | return 0; |
| 803 | if (BITS_IN_JSAMPLE != 8) |
| 804 | return 0; |
| 805 | if (sizeof(JDIMENSION) != 4) |
| 806 | return 0; |
| 807 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 808 | return 0; |
| 809 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 810 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
| 811 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 812 | if (simd_support & JSIMD_MMX) |
| 813 | return 1; |
| 814 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 815 | return 0; |
| 816 | } |
| 817 | |
| 818 | GLOBAL(void) |
| 819 | jsimd_idct_2x2 (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 820 | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 821 | JDIMENSION output_col) |
| 822 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 823 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
| 824 | jsimd_idct_2x2_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
| 825 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 826 | jsimd_idct_2x2_mmx(compptr->dct_table, coef_block, output_buf, output_col); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 827 | } |
| 828 | |
| 829 | GLOBAL(void) |
| 830 | jsimd_idct_4x4 (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 831 | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 832 | JDIMENSION output_col) |
| 833 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 834 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_red_sse2)) |
| 835 | jsimd_idct_4x4_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
| 836 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 837 | jsimd_idct_4x4_mmx(compptr->dct_table, coef_block, output_buf, output_col); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 838 | } |
| 839 | |
| 840 | GLOBAL(int) |
| 841 | jsimd_can_idct_islow (void) |
| 842 | { |
| 843 | init_simd(); |
| 844 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 845 | /* The code is optimised for these values only */ |
| 846 | if (DCTSIZE != 8) |
| 847 | return 0; |
| 848 | if (sizeof(JCOEF) != 2) |
| 849 | return 0; |
| 850 | if (BITS_IN_JSAMPLE != 8) |
| 851 | return 0; |
| 852 | if (sizeof(JDIMENSION) != 4) |
| 853 | return 0; |
| 854 | if (sizeof(ISLOW_MULT_TYPE) != 2) |
| 855 | return 0; |
| 856 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 857 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2)) |
| 858 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 859 | if (simd_support & JSIMD_MMX) |
| 860 | return 1; |
| 861 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 862 | return 0; |
| 863 | } |
| 864 | |
| 865 | GLOBAL(int) |
| 866 | jsimd_can_idct_ifast (void) |
| 867 | { |
| 868 | init_simd(); |
| 869 | |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 870 | /* The code is optimised for these values only */ |
| 871 | if (DCTSIZE != 8) |
| 872 | return 0; |
| 873 | if (sizeof(JCOEF) != 2) |
| 874 | return 0; |
| 875 | if (BITS_IN_JSAMPLE != 8) |
| 876 | return 0; |
| 877 | if (sizeof(JDIMENSION) != 4) |
| 878 | return 0; |
| 879 | if (sizeof(IFAST_MULT_TYPE) != 2) |
| 880 | return 0; |
| 881 | if (IFAST_SCALE_BITS != 2) |
| 882 | return 0; |
| 883 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 884 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2)) |
| 885 | return 1; |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 886 | if (simd_support & JSIMD_MMX) |
| 887 | return 1; |
| 888 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 889 | return 0; |
| 890 | } |
| 891 | |
| 892 | GLOBAL(int) |
| 893 | jsimd_can_idct_float (void) |
| 894 | { |
| 895 | init_simd(); |
| 896 | |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 897 | if (DCTSIZE != 8) |
| 898 | return 0; |
| 899 | if (sizeof(JCOEF) != 2) |
| 900 | return 0; |
| 901 | if (BITS_IN_JSAMPLE != 8) |
| 902 | return 0; |
| 903 | if (sizeof(JDIMENSION) != 4) |
| 904 | return 0; |
| 905 | if (sizeof(FAST_FLOAT) != 4) |
| 906 | return 0; |
| 907 | if (sizeof(FLOAT_MULT_TYPE) != 4) |
| 908 | return 0; |
| 909 | |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 910 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2)) |
| 911 | return 1; |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 912 | if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse)) |
| 913 | return 1; |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 914 | if (simd_support & JSIMD_3DNOW) |
| 915 | return 1; |
| 916 | |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 917 | return 0; |
| 918 | } |
| 919 | |
| 920 | GLOBAL(void) |
| 921 | jsimd_idct_islow (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 922 | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 923 | JDIMENSION output_col) |
| 924 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 925 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_islow_sse2)) |
| 926 | jsimd_idct_islow_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
| 927 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 928 | jsimd_idct_islow_mmx(compptr->dct_table, coef_block, output_buf, output_col); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 929 | } |
| 930 | |
| 931 | GLOBAL(void) |
| 932 | jsimd_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 933 | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 934 | JDIMENSION output_col) |
| 935 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 936 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_ifast_sse2)) |
| 937 | jsimd_idct_ifast_sse2(compptr->dct_table, coef_block, output_buf, output_col); |
| 938 | else if (simd_support & JSIMD_MMX) |
Pierre Ossman | 3e0e2de | 2009-03-09 13:25:30 +0000 | [diff] [blame] | 939 | jsimd_idct_ifast_mmx(compptr->dct_table, coef_block, output_buf, output_col); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 940 | } |
| 941 | |
| 942 | GLOBAL(void) |
| 943 | jsimd_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 944 | JCOEFPTR coef_block, JSAMPARRAY output_buf, |
| 945 | JDIMENSION output_col) |
| 946 | { |
Pierre Ossman | 7469386 | 2009-03-09 13:34:17 +0000 | [diff] [blame] | 947 | if ((simd_support & JSIMD_SSE2) && IS_ALIGNED_SSE(jconst_idct_float_sse2)) |
| 948 | jsimd_idct_float_sse2(compptr->dct_table, coef_block, |
| 949 | output_buf, output_col); |
| 950 | else if ((simd_support & JSIMD_SSE) && IS_ALIGNED_SSE(jconst_idct_float_sse)) |
Pierre Ossman | 0d37c57 | 2009-03-09 13:31:56 +0000 | [diff] [blame] | 951 | jsimd_idct_float_sse(compptr->dct_table, coef_block, |
| 952 | output_buf, output_col); |
| 953 | else if (simd_support & JSIMD_3DNOW) |
Pierre Ossman | 2c2e54b | 2009-03-09 13:28:10 +0000 | [diff] [blame] | 954 | jsimd_idct_float_3dnow(compptr->dct_table, coef_block, |
| 955 | output_buf, output_col); |
Pierre Ossman | 9ad5234 | 2009-03-09 13:15:56 +0000 | [diff] [blame] | 956 | } |
| 957 | |