jccolext-avx2.asm (25433B)
1 ; 2 ; jccolext.asm - colorspace conversion (AVX2) 3 ; 4 ; Copyright (C) 2015, Intel Corporation. 5 ; Copyright (C) 2016, 2024, D. R. Commander. 6 ; 7 ; Based on the x86 SIMD extension for IJG JPEG library 8 ; Copyright (C) 1999-2006, MIYASAKA Masaru. 9 ; For conditions of distribution and use, see copyright notice in jsimdext.inc 10 ; 11 ; This file should be assembled with NASM (Netwide Assembler) or Yasm. 12 13 %include "jcolsamp.inc" 14 15 ; -------------------------------------------------------------------------- 16 ; 17 ; Convert some rows of samples to the output colorspace. 18 ; 19 ; GLOBAL(void) 20 ; jsimd_rgb_ycc_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf, 21 ; JSAMPIMAGE output_buf, JDIMENSION output_row, 22 ; int num_rows); 23 ; 24 25 %define img_width(b) (b) + 8 ; JDIMENSION img_width 26 %define input_buf(b) (b) + 12 ; JSAMPARRAY input_buf 27 %define output_buf(b) (b) + 16 ; JSAMPIMAGE output_buf 28 %define output_row(b) (b) + 20 ; JDIMENSION output_row 29 %define num_rows(b) (b) + 24 ; int num_rows 30 31 %define original_ebp ebp + 0 32 %define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 33 ; ymmword wk[WK_NUM] 34 %define WK_NUM 8 35 %define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 36 37 align 32 38 GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2) 39 40 EXTN(jsimd_rgb_ycc_convert_avx2): 41 push ebp 42 mov eax, esp ; eax = original ebp 43 sub esp, byte 4 44 and esp, byte (-SIZEOF_YMMWORD) ; align to 256 bits 45 mov [esp], eax 46 mov ebp, esp ; ebp = aligned ebp 47 lea esp, [wk(0)] 48 PUSHPIC eax ; make a room for GOT address 49 push ebx 50 ; push ecx ; need not be preserved 51 ; push edx ; need not be preserved 52 push esi 53 push edi 54 55 GET_GOT ebx ; get GOT address 56 MOVPIC POINTER [gotptr], ebx ; save GOT address 57 58 mov ecx, JDIMENSION [img_width(eax)] 59 test ecx, ecx 60 jz near .return 61 62 push ecx 63 64 mov esi, JSAMPIMAGE [output_buf(eax)] 65 mov ecx, JDIMENSION [output_row(eax)] 66 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY] 67 mov ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY] 68 mov edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY] 69 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 70 lea ebx, [ebx+ecx*SIZEOF_JSAMPROW] 71 lea edx, [edx+ecx*SIZEOF_JSAMPROW] 72 73 pop ecx 74 75 mov esi, JSAMPARRAY [input_buf(eax)] 76 mov eax, INT [num_rows(eax)] 77 test eax, eax 78 jle near .return 79 ALIGNX 16, 7 80 .rowloop: 81 PUSHPIC eax 82 push edx 83 push ebx 84 push edi 85 push esi 86 push ecx ; col 87 88 mov esi, JSAMPROW [esi] ; inptr 89 mov edi, JSAMPROW [edi] ; outptr0 90 mov ebx, JSAMPROW [ebx] ; outptr1 91 mov edx, JSAMPROW [edx] ; outptr2 92 MOVPIC eax, POINTER [gotptr] ; load GOT address (eax) 93 94 cmp ecx, byte SIZEOF_YMMWORD 95 jae near .columnloop 96 ALIGNX 16, 7 97 98 %if RGB_PIXELSIZE == 3 ; --------------- 99 100 .column_ld1: 101 push eax 102 push edx 103 lea ecx, [ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 104 test cl, SIZEOF_BYTE 105 jz short .column_ld2 106 sub ecx, byte SIZEOF_BYTE 107 movzx eax, byte [esi+ecx] 108 .column_ld2: 109 test cl, SIZEOF_WORD 110 jz short .column_ld4 111 sub ecx, byte SIZEOF_WORD 112 movzx edx, word [esi+ecx] 113 shl eax, WORD_BIT 114 or eax, edx 115 .column_ld4: 116 vmovd xmmA, eax 117 pop edx 118 pop eax 119 test cl, SIZEOF_DWORD 120 jz short .column_ld8 121 sub ecx, byte SIZEOF_DWORD 122 vmovd xmmF, XMM_DWORD [esi+ecx] 123 vpslldq xmmA, xmmA, SIZEOF_DWORD 124 vpor xmmA, xmmA, xmmF 125 .column_ld8: 126 test cl, SIZEOF_MMWORD 127 jz short .column_ld16 128 sub ecx, byte SIZEOF_MMWORD 129 vmovq xmmB, XMM_MMWORD [esi+ecx] 130 vpslldq xmmA, xmmA, SIZEOF_MMWORD 131 vpor xmmA, xmmA, xmmB 132 .column_ld16: 133 test cl, SIZEOF_XMMWORD 134 jz short .column_ld32 135 sub ecx, byte SIZEOF_XMMWORD 136 vmovdqu xmmB, XMM_MMWORD [esi+ecx] 137 vperm2i128 ymmA, ymmA, ymmA, 1 138 vpor ymmA, ymmB 139 .column_ld32: 140 test cl, SIZEOF_YMMWORD 141 jz short .column_ld64 142 sub ecx, byte SIZEOF_YMMWORD 143 vmovdqa ymmF, ymmA 144 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 145 .column_ld64: 146 test cl, 2*SIZEOF_YMMWORD 147 mov ecx, SIZEOF_YMMWORD 148 jz short .rgb_ycc_cnv 149 vmovdqa ymmB, ymmA 150 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 151 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 152 jmp short .rgb_ycc_cnv 153 ALIGNX 16, 7 154 155 .columnloop: 156 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 157 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 158 vmovdqu ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD] 159 160 .rgb_ycc_cnv: 161 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 162 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 163 ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 164 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 165 ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 166 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 167 168 vmovdqu ymmC, ymmA 169 vinserti128 ymmA, ymmF, xmmA, 0 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 170 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 171 vinserti128 ymmC, ymmC, xmmB, 0 ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 172 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 173 vinserti128 ymmB, ymmB, xmmF, 0 ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 174 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 175 vperm2i128 ymmF, ymmC, ymmC, 1 ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A 176 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q) 177 178 vmovdqa ymmG, ymmA 179 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12 180 ; 22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I) 181 vpsrldq ymmG, ymmG, 8 ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I 182 ; 2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --) 183 184 vpunpckhbw ymmA, ymmA, ymmF ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A 185 ; 0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q) 186 vpslldq ymmF, ymmF, 8 ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27 187 ; 08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N) 188 189 vpunpcklbw ymmG, ymmG, ymmB ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D 190 ; 2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T) 191 vpunpckhbw ymmF, ymmF, ymmB ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F 192 ; 1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V) 193 194 vmovdqa ymmD, ymmA 195 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09 196 ; 11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P) 197 vpsrldq ymmD, ymmD, 8 ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P 198 ; 1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --) 199 200 vpunpckhbw ymmA, ymmA, ymmG ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D 201 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T) 202 vpslldq ymmG, ymmG, 8 ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B 203 ; 04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R) 204 205 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E 206 ; 1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U) 207 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 208 ; 2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V) 209 210 vmovdqa ymmE, ymmA 211 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C 212 ; 20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S) 213 vpsrldq ymmE, ymmE, 8 ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S 214 ; 2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --) 215 216 vpunpckhbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 217 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 218 vpslldq ymmD, ymmD, 8 ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D 219 ; 02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T) 220 221 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F 222 ; 2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V) 223 vpunpckhbw ymmD, ymmD, ymmG ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F 224 ; 1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V) 225 226 vpxor ymmH, ymmH, ymmH 227 228 vmovdqa ymmC, ymmA 229 vpunpcklbw ymmA, ymmA, ymmH ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 230 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 231 232 vmovdqa ymmB, ymmE 233 vpunpcklbw ymmE, ymmE, ymmH ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 234 vpunpckhbw ymmB, ymmB, ymmH ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 235 236 vmovdqa ymmF, ymmD 237 vpunpcklbw ymmD, ymmD, ymmH ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 238 vpunpckhbw ymmF, ymmF, ymmH ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 239 240 %else ; RGB_PIXELSIZE == 4 ; ----------- 241 242 .column_ld1: 243 test cl, SIZEOF_XMMWORD/16 244 jz short .column_ld2 245 sub ecx, byte SIZEOF_XMMWORD/16 246 vmovd xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE] 247 .column_ld2: 248 test cl, SIZEOF_XMMWORD/8 249 jz short .column_ld4 250 sub ecx, byte SIZEOF_XMMWORD/8 251 vmovq xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE] 252 vpslldq xmmA, xmmA, SIZEOF_MMWORD 253 vpor xmmA, xmmA, xmmF 254 .column_ld4: 255 test cl, SIZEOF_XMMWORD/4 256 jz short .column_ld8 257 sub ecx, byte SIZEOF_XMMWORD/4 258 vmovdqa xmmF, xmmA 259 vperm2i128 ymmF, ymmF, ymmF, 1 260 vmovdqu xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE] 261 vpor ymmA, ymmA, ymmF 262 .column_ld8: 263 test cl, SIZEOF_XMMWORD/2 264 jz short .column_ld16 265 sub ecx, byte SIZEOF_XMMWORD/2 266 vmovdqa ymmF, ymmA 267 vmovdqu ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE] 268 .column_ld16: 269 test cl, SIZEOF_XMMWORD 270 mov ecx, SIZEOF_YMMWORD 271 jz short .rgb_ycc_cnv 272 vmovdqa ymmE, ymmA 273 vmovdqa ymmH, ymmF 274 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 275 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 276 jmp short .rgb_ycc_cnv 277 ALIGNX 16, 7 278 279 .columnloop: 280 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 281 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 282 vmovdqu ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD] 283 vmovdqu ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD] 284 285 .rgb_ycc_cnv: 286 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 287 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) 288 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 289 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) 290 ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J 291 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 292 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R 293 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 294 295 vmovdqa ymmB, ymmA 296 vinserti128 ymmA, ymmA, xmmE, 1 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 297 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J) 298 vperm2i128 ymmE, ymmB, ymmE, 0x31 ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 299 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 300 301 vmovdqa ymmB, ymmF 302 vinserti128 ymmF, ymmF, xmmH, 1 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 303 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R) 304 vperm2i128 ymmH, ymmB, ymmH, 0x31 ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F 305 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 306 307 vmovdqa ymmD, ymmA 308 vpunpcklbw ymmA, ymmA, ymmE ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35 309 ; 0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L) 310 vpunpckhbw ymmD, ymmD, ymmE ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37 311 ; 0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N) 312 313 vmovdqa ymmC, ymmF 314 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D 315 ; 0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T) 316 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F 317 ; 0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V) 318 319 vmovdqa ymmB, ymmA 320 vpunpcklwd ymmA, ymmA, ymmF ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C 321 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S) 322 vpunpckhwd ymmB, ymmB, ymmF ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D 323 ; 0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T) 324 325 vmovdqa ymmG, ymmD 326 vpunpcklwd ymmD, ymmD, ymmC ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E 327 ; 0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U) 328 vpunpckhwd ymmG, ymmG, ymmC ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F 329 ; 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V) 330 331 vmovdqa ymmE, ymmA 332 vpunpcklbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 333 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 334 vpunpckhbw ymmE, ymmE, ymmD ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E 335 ; 2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U) 336 337 vmovdqa ymmH, ymmB 338 vpunpcklbw ymmB, ymmB, ymmG ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F 339 ; 0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V) 340 vpunpckhbw ymmH, ymmH, ymmG ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F 341 ; 2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V) 342 343 vpxor ymmF, ymmF, ymmF 344 345 vmovdqa ymmC, ymmA 346 vpunpcklbw ymmA, ymmA, ymmF ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 347 vpunpckhbw ymmC, ymmC, ymmF ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 348 349 vmovdqa ymmD, ymmB 350 vpunpcklbw ymmB, ymmB, ymmF ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 351 vpunpckhbw ymmD, ymmD, ymmF ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 352 353 vmovdqa ymmG, ymmE 354 vpunpcklbw ymmE, ymmE, ymmF ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 355 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U) 356 357 vpunpcklbw ymmF, ymmF, ymmH 358 vpunpckhbw ymmH, ymmH, ymmH 359 vpsrlw ymmF, ymmF, BYTE_BIT ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 360 vpsrlw ymmH, ymmH, BYTE_BIT ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V) 361 362 %endif ; RGB_PIXELSIZE ; --------------- 363 364 ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE 365 ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO 366 367 ; (Original) 368 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 369 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 370 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 371 ; 372 ; (This implementation) 373 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 374 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE 375 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE 376 377 vmovdqa YMMWORD [wk(0)], ymm0 ; wk(0)=RE 378 vmovdqa YMMWORD [wk(1)], ymm1 ; wk(1)=RO 379 vmovdqa YMMWORD [wk(2)], ymm4 ; wk(2)=BE 380 vmovdqa YMMWORD [wk(3)], ymm5 ; wk(3)=BO 381 382 vmovdqa ymm6, ymm1 383 vpunpcklwd ymm1, ymm1, ymm3 384 vpunpckhwd ymm6, ymm6, ymm3 385 vmovdqa ymm7, ymm1 386 vmovdqa ymm4, ymm6 387 vpmaddwd ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)] ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337) 388 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337) 389 vpmaddwd ymm7, ymm7, [GOTOFF(eax,PW_MF016_MF033)] ; ymm7=ROL*-FIX(0.168)+GOL*-FIX(0.331) 390 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_MF016_MF033)] ; ymm4=ROH*-FIX(0.168)+GOH*-FIX(0.331) 391 392 vmovdqa YMMWORD [wk(4)], ymm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337) 393 vmovdqa YMMWORD [wk(5)], ymm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337) 394 395 vpxor ymm1, ymm1, ymm1 396 vpxor ymm6, ymm6, ymm6 397 vpunpcklwd ymm1, ymm1, ymm5 ; ymm1=BOL 398 vpunpckhwd ymm6, ymm6, ymm5 ; ymm6=BOH 399 vpsrld ymm1, ymm1, 1 ; ymm1=BOL*FIX(0.500) 400 vpsrld ymm6, ymm6, 1 ; ymm6=BOH*FIX(0.500) 401 402 vmovdqa ymm5, [GOTOFF(eax,PD_ONEHALFM1_CJ)] ; ymm5=[PD_ONEHALFM1_CJ] 403 404 vpaddd ymm7, ymm7, ymm1 405 vpaddd ymm4, ymm4, ymm6 406 vpaddd ymm7, ymm7, ymm5 407 vpaddd ymm4, ymm4, ymm5 408 vpsrld ymm7, ymm7, SCALEBITS ; ymm7=CbOL 409 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=CbOH 410 vpackssdw ymm7, ymm7, ymm4 ; ymm7=CbO 411 412 vmovdqa ymm1, YMMWORD [wk(2)] ; ymm1=BE 413 414 vmovdqa ymm6, ymm0 415 vpunpcklwd ymm0, ymm0, ymm2 416 vpunpckhwd ymm6, ymm6, ymm2 417 vmovdqa ymm5, ymm0 418 vmovdqa ymm4, ymm6 419 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)] ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337) 420 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337) 421 vpmaddwd ymm5, ymm5, [GOTOFF(eax,PW_MF016_MF033)] ; ymm5=REL*-FIX(0.168)+GEL*-FIX(0.331) 422 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_MF016_MF033)] ; ymm4=REH*-FIX(0.168)+GEH*-FIX(0.331) 423 424 vmovdqa YMMWORD [wk(6)], ymm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337) 425 vmovdqa YMMWORD [wk(7)], ymm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337) 426 427 vpxor ymm0, ymm0, ymm0 428 vpxor ymm6, ymm6, ymm6 429 vpunpcklwd ymm0, ymm0, ymm1 ; ymm0=BEL 430 vpunpckhwd ymm6, ymm6, ymm1 ; ymm6=BEH 431 vpsrld ymm0, ymm0, 1 ; ymm0=BEL*FIX(0.500) 432 vpsrld ymm6, ymm6, 1 ; ymm6=BEH*FIX(0.500) 433 434 vmovdqa ymm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)] ; ymm1=[PD_ONEHALFM1_CJ] 435 436 vpaddd ymm5, ymm5, ymm0 437 vpaddd ymm4, ymm4, ymm6 438 vpaddd ymm5, ymm5, ymm1 439 vpaddd ymm4, ymm4, ymm1 440 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CbEL 441 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=CbEH 442 vpackssdw ymm5, ymm5, ymm4 ; ymm5=CbE 443 444 vpsllw ymm7, ymm7, BYTE_BIT 445 vpor ymm5, ymm5, ymm7 ; ymm5=Cb 446 vmovdqu YMMWORD [ebx], ymm5 ; Save Cb 447 448 vmovdqa ymm0, YMMWORD [wk(3)] ; ymm0=BO 449 vmovdqa ymm6, YMMWORD [wk(2)] ; ymm6=BE 450 vmovdqa ymm1, YMMWORD [wk(1)] ; ymm1=RO 451 452 vmovdqa ymm4, ymm0 453 vpunpcklwd ymm0, ymm0, ymm3 454 vpunpckhwd ymm4, ymm4, ymm3 455 vmovdqa ymm7, ymm0 456 vmovdqa ymm5, ymm4 457 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)] ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250) 458 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250) 459 vpmaddwd ymm7, ymm7, [GOTOFF(eax,PW_MF008_MF041)] ; ymm7=BOL*-FIX(0.081)+GOL*-FIX(0.418) 460 vpmaddwd ymm5, ymm5, [GOTOFF(eax,PW_MF008_MF041)] ; ymm5=BOH*-FIX(0.081)+GOH*-FIX(0.418) 461 462 vmovdqa ymm3, [GOTOFF(eax,PD_ONEHALF)] ; ymm3=[PD_ONEHALF] 463 464 vpaddd ymm0, ymm0, YMMWORD [wk(4)] 465 vpaddd ymm4, ymm4, YMMWORD [wk(5)] 466 vpaddd ymm0, ymm0, ymm3 467 vpaddd ymm4, ymm4, ymm3 468 vpsrld ymm0, ymm0, SCALEBITS ; ymm0=YOL 469 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YOH 470 vpackssdw ymm0, ymm0, ymm4 ; ymm0=YO 471 472 vpxor ymm3, ymm3, ymm3 473 vpxor ymm4, ymm4, ymm4 474 vpunpcklwd ymm3, ymm3, ymm1 ; ymm3=ROL 475 vpunpckhwd ymm4, ymm4, ymm1 ; ymm4=ROH 476 vpsrld ymm3, ymm3, 1 ; ymm3=ROL*FIX(0.500) 477 vpsrld ymm4, ymm4, 1 ; ymm4=ROH*FIX(0.500) 478 479 vmovdqa ymm1, [GOTOFF(eax,PD_ONEHALFM1_CJ)] ; ymm1=[PD_ONEHALFM1_CJ] 480 481 vpaddd ymm7, ymm7, ymm3 482 vpaddd ymm5, ymm5, ymm4 483 vpaddd ymm7, ymm7, ymm1 484 vpaddd ymm5, ymm5, ymm1 485 vpsrld ymm7, ymm7, SCALEBITS ; ymm7=CrOL 486 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CrOH 487 vpackssdw ymm7, ymm7, ymm5 ; ymm7=CrO 488 489 vmovdqa ymm3, YMMWORD [wk(0)] ; ymm3=RE 490 491 vmovdqa ymm4, ymm6 492 vpunpcklwd ymm6, ymm6, ymm2 493 vpunpckhwd ymm4, ymm4, ymm2 494 vmovdqa ymm1, ymm6 495 vmovdqa ymm5, ymm4 496 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)] ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250) 497 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250) 498 vpmaddwd ymm1, ymm1, [GOTOFF(eax,PW_MF008_MF041)] ; ymm1=BEL*-FIX(0.081)+GEL*-FIX(0.418) 499 vpmaddwd ymm5, ymm5, [GOTOFF(eax,PW_MF008_MF041)] ; ymm5=BEH*-FIX(0.081)+GEH*-FIX(0.418) 500 501 vmovdqa ymm2, [GOTOFF(eax,PD_ONEHALF)] ; ymm2=[PD_ONEHALF] 502 503 vpaddd ymm6, ymm6, YMMWORD [wk(6)] 504 vpaddd ymm4, ymm4, YMMWORD [wk(7)] 505 vpaddd ymm6, ymm6, ymm2 506 vpaddd ymm4, ymm4, ymm2 507 vpsrld ymm6, ymm6, SCALEBITS ; ymm6=YEL 508 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YEH 509 vpackssdw ymm6, ymm6, ymm4 ; ymm6=YE 510 511 vpsllw ymm0, ymm0, BYTE_BIT 512 vpor ymm6, ymm6, ymm0 ; ymm6=Y 513 vmovdqu YMMWORD [edi], ymm6 ; Save Y 514 515 vpxor ymm2, ymm2, ymm2 516 vpxor ymm4, ymm4, ymm4 517 vpunpcklwd ymm2, ymm2, ymm3 ; ymm2=REL 518 vpunpckhwd ymm4, ymm4, ymm3 ; ymm4=REH 519 vpsrld ymm2, ymm2, 1 ; ymm2=REL*FIX(0.500) 520 vpsrld ymm4, ymm4, 1 ; ymm4=REH*FIX(0.500) 521 522 vmovdqa ymm0, [GOTOFF(eax,PD_ONEHALFM1_CJ)] ; ymm0=[PD_ONEHALFM1_CJ] 523 524 vpaddd ymm1, ymm1, ymm2 525 vpaddd ymm5, ymm5, ymm4 526 vpaddd ymm1, ymm1, ymm0 527 vpaddd ymm5, ymm5, ymm0 528 vpsrld ymm1, ymm1, SCALEBITS ; ymm1=CrEL 529 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CrEH 530 vpackssdw ymm1, ymm1, ymm5 ; ymm1=CrE 531 532 vpsllw ymm7, ymm7, BYTE_BIT 533 vpor ymm1, ymm1, ymm7 ; ymm1=Cr 534 vmovdqu YMMWORD [edx], ymm1 ; Save Cr 535 536 sub ecx, byte SIZEOF_YMMWORD 537 add esi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; inptr 538 add edi, byte SIZEOF_YMMWORD ; outptr0 539 add ebx, byte SIZEOF_YMMWORD ; outptr1 540 add edx, byte SIZEOF_YMMWORD ; outptr2 541 cmp ecx, byte SIZEOF_YMMWORD 542 jae near .columnloop 543 test ecx, ecx 544 jnz near .column_ld1 545 546 pop ecx ; col 547 pop esi 548 pop edi 549 pop ebx 550 pop edx 551 POPPIC eax 552 553 add esi, byte SIZEOF_JSAMPROW ; input_buf 554 add edi, byte SIZEOF_JSAMPROW 555 add ebx, byte SIZEOF_JSAMPROW 556 add edx, byte SIZEOF_JSAMPROW 557 dec eax ; num_rows 558 jg near .rowloop 559 560 .return: 561 vzeroupper 562 pop edi 563 pop esi 564 ; pop edx ; need not be preserved 565 ; pop ecx ; need not be preserved 566 pop ebx 567 mov esp, ebp ; esp <- aligned ebp 568 pop esp ; esp <- original ebp 569 pop ebp 570 ret 571 572 ; For some reason, the OS X linker does not honor the request to align the 573 ; segment unless we do this. 574 align 32