jcgryext-avx2.asm (20226B)
1 ; 2 ; jcgryext.asm - grayscale colorspace conversion (AVX2) 3 ; 4 ; Copyright (C) 2011, 2016, 2024, D. R. Commander. 5 ; Copyright (C) 2015, Intel Corporation. 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_gray_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 2 35 %define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 36 37 align 32 38 GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2) 39 40 EXTN(jsimd_rgb_gray_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 lea edi, [edi+ecx*SIZEOF_JSAMPROW] 68 69 pop ecx 70 71 mov esi, JSAMPARRAY [input_buf(eax)] 72 mov eax, INT [num_rows(eax)] 73 test eax, eax 74 jle near .return 75 ALIGNX 16, 7 76 .rowloop: 77 PUSHPIC eax 78 push edi 79 push esi 80 push ecx ; col 81 82 mov esi, JSAMPROW [esi] ; inptr 83 mov edi, JSAMPROW [edi] ; outptr0 84 MOVPIC eax, POINTER [gotptr] ; load GOT address (eax) 85 86 cmp ecx, byte SIZEOF_YMMWORD 87 jae near .columnloop 88 ALIGNX 16, 7 89 90 %if RGB_PIXELSIZE == 3 ; --------------- 91 92 .column_ld1: 93 push eax 94 push edx 95 lea ecx, [ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 96 test cl, SIZEOF_BYTE 97 jz short .column_ld2 98 sub ecx, byte SIZEOF_BYTE 99 movzx eax, byte [esi+ecx] 100 .column_ld2: 101 test cl, SIZEOF_WORD 102 jz short .column_ld4 103 sub ecx, byte SIZEOF_WORD 104 movzx edx, word [esi+ecx] 105 shl eax, WORD_BIT 106 or eax, edx 107 .column_ld4: 108 vmovd xmmA, eax 109 pop edx 110 pop eax 111 test cl, SIZEOF_DWORD 112 jz short .column_ld8 113 sub ecx, byte SIZEOF_DWORD 114 vmovd xmmF, XMM_DWORD [esi+ecx] 115 vpslldq xmmA, xmmA, SIZEOF_DWORD 116 vpor xmmA, xmmA, xmmF 117 .column_ld8: 118 test cl, SIZEOF_MMWORD 119 jz short .column_ld16 120 sub ecx, byte SIZEOF_MMWORD 121 vmovq xmmB, XMM_MMWORD [esi+ecx] 122 vpslldq xmmA, xmmA, SIZEOF_MMWORD 123 vpor xmmA, xmmA, xmmB 124 .column_ld16: 125 test cl, SIZEOF_XMMWORD 126 jz short .column_ld32 127 sub ecx, byte SIZEOF_XMMWORD 128 vmovdqu xmmB, XMM_MMWORD [esi+ecx] 129 vperm2i128 ymmA, ymmA, ymmA, 1 130 vpor ymmA, ymmB 131 .column_ld32: 132 test cl, SIZEOF_YMMWORD 133 jz short .column_ld64 134 sub ecx, byte SIZEOF_YMMWORD 135 vmovdqa ymmF, ymmA 136 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 137 .column_ld64: 138 test cl, 2*SIZEOF_YMMWORD 139 mov ecx, SIZEOF_YMMWORD 140 jz short .rgb_gray_cnv 141 vmovdqa ymmB, ymmA 142 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 143 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 144 jmp short .rgb_gray_cnv 145 ALIGNX 16, 7 146 147 .columnloop: 148 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 149 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 150 vmovdqu ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD] 151 152 .rgb_gray_cnv: 153 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 154 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 155 ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 156 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 157 ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 158 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 159 160 vmovdqu ymmC, ymmA 161 vinserti128 ymmA, ymmF, xmmA, 0 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 162 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 163 vinserti128 ymmC, ymmC, xmmB, 0 ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 164 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 165 vinserti128 ymmB, ymmB, xmmF, 0 ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 166 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 167 vperm2i128 ymmF, ymmC, ymmC, 1 ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A 168 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q) 169 170 vmovdqa ymmG, ymmA 171 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12 172 ; 22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I) 173 vpsrldq ymmG, ymmG, 8 ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I 174 ; 2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --) 175 176 vpunpckhbw ymmA, ymmA, ymmF ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A 177 ; 0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q) 178 vpslldq ymmF, ymmF, 8 ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27 179 ; 08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N) 180 181 vpunpcklbw ymmG, ymmG, ymmB ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D 182 ; 2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T) 183 vpunpckhbw ymmF, ymmF, ymmB ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F 184 ; 1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V) 185 186 vmovdqa ymmD, ymmA 187 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09 188 ; 11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P) 189 vpsrldq ymmD, ymmD, 8 ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P 190 ; 1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --) 191 192 vpunpckhbw ymmA, ymmA, ymmG ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D 193 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T) 194 vpslldq ymmG, ymmG, 8 ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B 195 ; 04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R) 196 197 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E 198 ; 1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U) 199 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 200 ; 2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V) 201 202 vmovdqa ymmE, ymmA 203 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C 204 ; 20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S) 205 vpsrldq ymmE, ymmE, 8 ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S 206 ; 2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --) 207 208 vpunpckhbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 209 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 210 vpslldq ymmD, ymmD, 8 ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D 211 ; 02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T) 212 213 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F 214 ; 2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V) 215 vpunpckhbw ymmD, ymmD, ymmG ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F 216 ; 1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V) 217 218 vpxor ymmH, ymmH, ymmH 219 220 vmovdqa ymmC, ymmA 221 vpunpcklbw ymmA, ymmA, ymmH ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 222 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 223 224 vmovdqa ymmB, ymmE 225 vpunpcklbw ymmE, ymmE, ymmH ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 226 vpunpckhbw ymmB, ymmB, ymmH ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 227 228 vmovdqa ymmF, ymmD 229 vpunpcklbw ymmD, ymmD, ymmH ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 230 vpunpckhbw ymmF, ymmF, ymmH ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 231 232 %else ; RGB_PIXELSIZE == 4 ; ----------- 233 234 .column_ld1: 235 test cl, SIZEOF_XMMWORD/16 236 jz short .column_ld2 237 sub ecx, byte SIZEOF_XMMWORD/16 238 vmovd xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE] 239 .column_ld2: 240 test cl, SIZEOF_XMMWORD/8 241 jz short .column_ld4 242 sub ecx, byte SIZEOF_XMMWORD/8 243 vmovq xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE] 244 vpslldq xmmA, xmmA, SIZEOF_MMWORD 245 vpor xmmA, xmmA, xmmF 246 .column_ld4: 247 test cl, SIZEOF_XMMWORD/4 248 jz short .column_ld8 249 sub ecx, byte SIZEOF_XMMWORD/4 250 vmovdqa xmmF, xmmA 251 vperm2i128 ymmF, ymmF, ymmF, 1 252 vmovdqu xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE] 253 vpor ymmA, ymmA, ymmF 254 .column_ld8: 255 test cl, SIZEOF_XMMWORD/2 256 jz short .column_ld16 257 sub ecx, byte SIZEOF_XMMWORD/2 258 vmovdqa ymmF, ymmA 259 vmovdqu ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE] 260 .column_ld16: 261 test cl, SIZEOF_XMMWORD 262 mov ecx, SIZEOF_YMMWORD 263 jz short .rgb_gray_cnv 264 vmovdqa ymmE, ymmA 265 vmovdqa ymmH, ymmF 266 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 267 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 268 jmp short .rgb_gray_cnv 269 ALIGNX 16, 7 270 271 .columnloop: 272 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 273 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 274 vmovdqu ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD] 275 vmovdqu ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD] 276 277 .rgb_gray_cnv: 278 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 279 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) 280 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 281 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) 282 ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J 283 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 284 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R 285 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 286 287 vmovdqa ymmB, ymmA 288 vinserti128 ymmA, ymmA, xmmE, 1 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 289 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J) 290 vperm2i128 ymmE, ymmB, ymmE, 0x31 ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 291 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 292 293 vmovdqa ymmB, ymmF 294 vinserti128 ymmF, ymmF, xmmH, 1 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 295 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R) 296 vperm2i128 ymmH, ymmB, ymmH, 0x31 ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F 297 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 298 299 vmovdqa ymmD, ymmA 300 vpunpcklbw ymmA, ymmA, ymmE ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35 301 ; 0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L) 302 vpunpckhbw ymmD, ymmD, ymmE ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37 303 ; 0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N) 304 305 vmovdqa ymmC, ymmF 306 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D 307 ; 0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T) 308 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F 309 ; 0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V) 310 311 vmovdqa ymmB, ymmA 312 vpunpcklwd ymmA, ymmA, ymmF ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C 313 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S) 314 vpunpckhwd ymmB, ymmB, ymmF ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D 315 ; 0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T) 316 317 vmovdqa ymmG, ymmD 318 vpunpcklwd ymmD, ymmD, ymmC ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E 319 ; 0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U) 320 vpunpckhwd ymmG, ymmG, ymmC ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F 321 ; 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V) 322 323 vmovdqa ymmE, ymmA 324 vpunpcklbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 325 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 326 vpunpckhbw ymmE, ymmE, ymmD ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E 327 ; 2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U) 328 329 vmovdqa ymmH, ymmB 330 vpunpcklbw ymmB, ymmB, ymmG ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F 331 ; 0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V) 332 vpunpckhbw ymmH, ymmH, ymmG ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F 333 ; 2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V) 334 335 vpxor ymmF, ymmF, ymmF 336 337 vmovdqa ymmC, ymmA 338 vpunpcklbw ymmA, ymmA, ymmF ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 339 vpunpckhbw ymmC, ymmC, ymmF ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 340 341 vmovdqa ymmD, ymmB 342 vpunpcklbw ymmB, ymmB, ymmF ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 343 vpunpckhbw ymmD, ymmD, ymmF ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 344 345 vmovdqa ymmG, ymmE 346 vpunpcklbw ymmE, ymmE, ymmF ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 347 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U) 348 349 vpunpcklbw ymmF, ymmF, ymmH 350 vpunpckhbw ymmH, ymmH, ymmH 351 vpsrlw ymmF, ymmF, BYTE_BIT ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 352 vpsrlw ymmH, ymmH, BYTE_BIT ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V) 353 354 %endif ; RGB_PIXELSIZE ; --------------- 355 356 ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE 357 ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO 358 359 ; (Original) 360 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 361 ; 362 ; (This implementation) 363 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 364 365 vmovdqa ymm6, ymm1 366 vpunpcklwd ymm1, ymm1, ymm3 367 vpunpckhwd ymm6, ymm6, ymm3 368 vpmaddwd ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)] ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337) 369 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337) 370 371 vmovdqa ymm7, ymm6 ; ymm7=ROH*FIX(0.299)+GOH*FIX(0.337) 372 373 vmovdqa ymm6, ymm0 374 vpunpcklwd ymm0, ymm0, ymm2 375 vpunpckhwd ymm6, ymm6, ymm2 376 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)] ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337) 377 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337) 378 379 vmovdqa YMMWORD [wk(0)], ymm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) 380 vmovdqa YMMWORD [wk(1)], ymm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) 381 382 vmovdqa ymm0, ymm5 ; ymm0=BO 383 vmovdqa ymm6, ymm4 ; ymm6=BE 384 385 vmovdqa ymm4, ymm0 386 vpunpcklwd ymm0, ymm0, ymm3 387 vpunpckhwd ymm4, ymm4, ymm3 388 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)] ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250) 389 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250) 390 391 vmovdqa ymm3, [GOTOFF(eax,PD_ONEHALF)] ; ymm3=[PD_ONEHALF] 392 393 vpaddd ymm0, ymm0, ymm1 394 vpaddd ymm4, ymm4, ymm7 395 vpaddd ymm0, ymm0, ymm3 396 vpaddd ymm4, ymm4, ymm3 397 vpsrld ymm0, ymm0, SCALEBITS ; ymm0=YOL 398 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YOH 399 vpackssdw ymm0, ymm0, ymm4 ; ymm0=YO 400 401 vmovdqa ymm4, ymm6 402 vpunpcklwd ymm6, ymm6, ymm2 403 vpunpckhwd ymm4, ymm4, ymm2 404 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)] ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250) 405 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250) 406 407 vmovdqa ymm2, [GOTOFF(eax,PD_ONEHALF)] ; ymm2=[PD_ONEHALF] 408 409 vpaddd ymm6, ymm6, YMMWORD [wk(0)] 410 vpaddd ymm4, ymm4, YMMWORD [wk(1)] 411 vpaddd ymm6, ymm6, ymm2 412 vpaddd ymm4, ymm4, ymm2 413 vpsrld ymm6, ymm6, SCALEBITS ; ymm6=YEL 414 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YEH 415 vpackssdw ymm6, ymm6, ymm4 ; ymm6=YE 416 417 vpsllw ymm0, ymm0, BYTE_BIT 418 vpor ymm6, ymm6, ymm0 ; ymm6=Y 419 vmovdqu YMMWORD [edi], ymm6 ; Save Y 420 421 sub ecx, byte SIZEOF_YMMWORD 422 add esi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; inptr 423 add edi, byte SIZEOF_YMMWORD ; outptr0 424 cmp ecx, byte SIZEOF_YMMWORD 425 jae near .columnloop 426 test ecx, ecx 427 jnz near .column_ld1 428 429 pop ecx ; col 430 pop esi 431 pop edi 432 POPPIC eax 433 434 add esi, byte SIZEOF_JSAMPROW ; input_buf 435 add edi, byte SIZEOF_JSAMPROW 436 dec eax ; num_rows 437 jg near .rowloop 438 439 .return: 440 vzeroupper 441 pop edi 442 pop esi 443 ; pop edx ; need not be preserved 444 ; pop ecx ; need not be preserved 445 pop ebx 446 mov esp, ebp ; esp <- aligned ebp 447 pop esp ; esp <- original ebp 448 pop ebp 449 ret 450 451 ; For some reason, the OS X linker does not honor the request to align the 452 ; segment unless we do this. 453 align 32