1; 2; jcgryext.asm - grayscale colorspace conversion (AVX2) 3; 4; Copyright (C) 2011, 2016, 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), 12; can *not* be assembled with Microsoft's MASM or any compatible 13; assembler (including Borland's Turbo Assembler). 14; NASM is available from http://nasm.sourceforge.net/ or 15; http://sourceforge.net/project/showfiles.php?group_id=6208 16 17%include "jcolsamp.inc" 18 19; -------------------------------------------------------------------------- 20; 21; Convert some rows of samples to the output colorspace. 22; 23; GLOBAL(void) 24; jsimd_rgb_gray_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf, 25; JSAMPIMAGE output_buf, JDIMENSION output_row, 26; int num_rows); 27; 28 29%define img_width(b) (b) + 8 ; JDIMENSION img_width 30%define input_buf(b) (b) + 12 ; JSAMPARRAY input_buf 31%define output_buf(b) (b) + 16 ; JSAMPIMAGE output_buf 32%define output_row(b) (b) + 20 ; JDIMENSION output_row 33%define num_rows(b) (b) + 24 ; int num_rows 34 35%define original_ebp ebp + 0 36%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 37 ; ymmword wk[WK_NUM] 38%define WK_NUM 2 39%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 40 41 align 32 42 GLOBAL_FUNCTION(jsimd_rgb_gray_convert_avx2) 43 44EXTN(jsimd_rgb_gray_convert_avx2): 45 push ebp 46 mov eax, esp ; eax = original ebp 47 sub esp, byte 4 48 and esp, byte (-SIZEOF_YMMWORD) ; align to 256 bits 49 mov [esp], eax 50 mov ebp, esp ; ebp = aligned ebp 51 lea esp, [wk(0)] 52 pushpic eax ; make a room for GOT address 53 push ebx 54; push ecx ; need not be preserved 55; push edx ; need not be preserved 56 push esi 57 push edi 58 59 get_GOT ebx ; get GOT address 60 movpic POINTER [gotptr], ebx ; save GOT address 61 62 mov ecx, JDIMENSION [img_width(eax)] 63 test ecx, ecx 64 jz near .return 65 66 push ecx 67 68 mov esi, JSAMPIMAGE [output_buf(eax)] 69 mov ecx, JDIMENSION [output_row(eax)] 70 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY] 71 lea edi, [edi+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 edi 83 push esi 84 push ecx ; col 85 86 mov esi, JSAMPROW [esi] ; inptr 87 mov edi, JSAMPROW [edi] ; outptr0 88 movpic eax, POINTER [gotptr] ; load GOT address (eax) 89 90 cmp ecx, byte SIZEOF_YMMWORD 91 jae near .columnloop 92 alignx 16, 7 93 94%if RGB_PIXELSIZE == 3 ; --------------- 95 96.column_ld1: 97 push eax 98 push edx 99 lea ecx, [ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE 100 test cl, SIZEOF_BYTE 101 jz short .column_ld2 102 sub ecx, byte SIZEOF_BYTE 103 movzx eax, byte [esi+ecx] 104.column_ld2: 105 test cl, SIZEOF_WORD 106 jz short .column_ld4 107 sub ecx, byte SIZEOF_WORD 108 movzx edx, word [esi+ecx] 109 shl eax, WORD_BIT 110 or eax, edx 111.column_ld4: 112 vmovd xmmA, eax 113 pop edx 114 pop eax 115 test cl, SIZEOF_DWORD 116 jz short .column_ld8 117 sub ecx, byte SIZEOF_DWORD 118 vmovd xmmF, XMM_DWORD [esi+ecx] 119 vpslldq xmmA, xmmA, SIZEOF_DWORD 120 vpor xmmA, xmmA, xmmF 121.column_ld8: 122 test cl, SIZEOF_MMWORD 123 jz short .column_ld16 124 sub ecx, byte SIZEOF_MMWORD 125 vmovq xmmB, XMM_MMWORD [esi+ecx] 126 vpslldq xmmA, xmmA, SIZEOF_MMWORD 127 vpor xmmA, xmmA, xmmB 128.column_ld16: 129 test cl, SIZEOF_XMMWORD 130 jz short .column_ld32 131 sub ecx, byte SIZEOF_XMMWORD 132 vmovdqu xmmB, XMM_MMWORD [esi+ecx] 133 vperm2i128 ymmA, ymmA, ymmA, 1 134 vpor ymmA, ymmB 135.column_ld32: 136 test cl, SIZEOF_YMMWORD 137 jz short .column_ld64 138 sub ecx, byte SIZEOF_YMMWORD 139 vmovdqa ymmF, ymmA 140 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 141.column_ld64: 142 test cl, 2*SIZEOF_YMMWORD 143 mov ecx, SIZEOF_YMMWORD 144 jz short .rgb_gray_cnv 145 vmovdqa ymmB, ymmA 146 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 147 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 148 jmp short .rgb_gray_cnv 149 alignx 16, 7 150 151.columnloop: 152 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 153 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 154 vmovdqu ymmB, YMMWORD [esi+2*SIZEOF_YMMWORD] 155 156.rgb_gray_cnv: 157 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 158 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 159 ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 160 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 161 ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 162 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 163 164 vmovdqu ymmC, ymmA 165 vinserti128 ymmA, ymmF, xmmA, 0 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 166 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 167 vinserti128 ymmC, ymmC, xmmB, 0 ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 168 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 169 vinserti128 ymmB, ymmB, xmmF, 0 ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 170 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 171 vperm2i128 ymmF, ymmC, ymmC, 1 ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A 172 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q) 173 174 vmovdqa ymmG, ymmA 175 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12 176 ; 22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I) 177 vpsrldq ymmG, ymmG, 8 ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I 178 ; 2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --) 179 180 vpunpckhbw ymmA, ymmA, ymmF ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A 181 ; 0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q) 182 vpslldq ymmF, ymmF, 8 ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27 183 ; 08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N) 184 185 vpunpcklbw ymmG, ymmG, ymmB ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D 186 ; 2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T) 187 vpunpckhbw ymmF, ymmF, ymmB ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F 188 ; 1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V) 189 190 vmovdqa ymmD, ymmA 191 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09 192 ; 11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P) 193 vpsrldq ymmD, ymmD, 8 ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P 194 ; 1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --) 195 196 vpunpckhbw ymmA, ymmA, ymmG ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D 197 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T) 198 vpslldq ymmG, ymmG, 8 ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B 199 ; 04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R) 200 201 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E 202 ; 1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U) 203 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 204 ; 2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V) 205 206 vmovdqa ymmE, ymmA 207 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C 208 ; 20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S) 209 vpsrldq ymmE, ymmE, 8 ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S 210 ; 2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --) 211 212 vpunpckhbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 213 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 214 vpslldq ymmD, ymmD, 8 ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D 215 ; 02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T) 216 217 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F 218 ; 2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V) 219 vpunpckhbw ymmD, ymmD, ymmG ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F 220 ; 1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V) 221 222 vpxor ymmH, ymmH, ymmH 223 224 vmovdqa ymmC, ymmA 225 vpunpcklbw ymmA, ymmA, ymmH ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 226 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 227 228 vmovdqa ymmB, ymmE 229 vpunpcklbw ymmE, ymmE, ymmH ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 230 vpunpckhbw ymmB, ymmB, ymmH ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 231 232 vmovdqa ymmF, ymmD 233 vpunpcklbw ymmD, ymmD, ymmH ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 234 vpunpckhbw ymmF, ymmF, ymmH ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 235 236%else ; RGB_PIXELSIZE == 4 ; ----------- 237 238.column_ld1: 239 test cl, SIZEOF_XMMWORD/16 240 jz short .column_ld2 241 sub ecx, byte SIZEOF_XMMWORD/16 242 vmovd xmmA, XMM_DWORD [esi+ecx*RGB_PIXELSIZE] 243.column_ld2: 244 test cl, SIZEOF_XMMWORD/8 245 jz short .column_ld4 246 sub ecx, byte SIZEOF_XMMWORD/8 247 vmovq xmmF, XMM_MMWORD [esi+ecx*RGB_PIXELSIZE] 248 vpslldq xmmA, xmmA, SIZEOF_MMWORD 249 vpor xmmA, xmmA, xmmF 250.column_ld4: 251 test cl, SIZEOF_XMMWORD/4 252 jz short .column_ld8 253 sub ecx, byte SIZEOF_XMMWORD/4 254 vmovdqa xmmF, xmmA 255 vperm2i128 ymmF, ymmF, ymmF, 1 256 vmovdqu xmmA, XMMWORD [esi+ecx*RGB_PIXELSIZE] 257 vpor ymmA, ymmA, ymmF 258.column_ld8: 259 test cl, SIZEOF_XMMWORD/2 260 jz short .column_ld16 261 sub ecx, byte SIZEOF_XMMWORD/2 262 vmovdqa ymmF, ymmA 263 vmovdqu ymmA, YMMWORD [esi+ecx*RGB_PIXELSIZE] 264.column_ld16: 265 test cl, SIZEOF_XMMWORD 266 mov ecx, SIZEOF_YMMWORD 267 jz short .rgb_gray_cnv 268 vmovdqa ymmE, ymmA 269 vmovdqa ymmH, ymmF 270 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 271 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 272 jmp short .rgb_gray_cnv 273 alignx 16, 7 274 275.columnloop: 276 vmovdqu ymmA, YMMWORD [esi+0*SIZEOF_YMMWORD] 277 vmovdqu ymmF, YMMWORD [esi+1*SIZEOF_YMMWORD] 278 vmovdqu ymmE, YMMWORD [esi+2*SIZEOF_YMMWORD] 279 vmovdqu ymmH, YMMWORD [esi+3*SIZEOF_YMMWORD] 280 281.rgb_gray_cnv: 282 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 283 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) 284 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 285 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) 286 ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J 287 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 288 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R 289 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 290 291 vmovdqa ymmB, ymmA 292 vinserti128 ymmA, ymmA, xmmE, 1 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 293 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J) 294 vperm2i128 ymmE, ymmB, ymmE, 0x31 ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 295 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 296 297 vmovdqa ymmB, ymmF 298 vinserti128 ymmF, ymmF, xmmH, 1 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 299 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R) 300 vperm2i128 ymmH, ymmB, ymmH, 0x31 ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F 301 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 302 303 vmovdqa ymmD, ymmA 304 vpunpcklbw ymmA, ymmA, ymmE ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35 305 ; 0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L) 306 vpunpckhbw ymmD, ymmD, ymmE ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37 307 ; 0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N) 308 309 vmovdqa ymmC, ymmF 310 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D 311 ; 0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T) 312 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F 313 ; 0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V) 314 315 vmovdqa ymmB, ymmA 316 vpunpcklwd ymmA, ymmA, ymmF ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C 317 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S) 318 vpunpckhwd ymmB, ymmB, ymmF ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D 319 ; 0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T) 320 321 vmovdqa ymmG, ymmD 322 vpunpcklwd ymmD, ymmD, ymmC ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E 323 ; 0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U) 324 vpunpckhwd ymmG, ymmG, ymmC ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F 325 ; 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V) 326 327 vmovdqa ymmE, ymmA 328 vpunpcklbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E 329 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U) 330 vpunpckhbw ymmE, ymmE, ymmD ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E 331 ; 2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U) 332 333 vmovdqa ymmH, ymmB 334 vpunpcklbw ymmB, ymmB, ymmG ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F 335 ; 0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V) 336 vpunpckhbw ymmH, ymmH, ymmG ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F 337 ; 2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V) 338 339 vpxor ymmF, ymmF, ymmF 340 341 vmovdqa ymmC, ymmA 342 vpunpcklbw ymmA, ymmA, ymmF ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U) 343 vpunpckhbw ymmC, ymmC, ymmF ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U) 344 345 vmovdqa ymmD, ymmB 346 vpunpcklbw ymmB, ymmB, ymmF ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V) 347 vpunpckhbw ymmD, ymmD, ymmF ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V) 348 349 vmovdqa ymmG, ymmE 350 vpunpcklbw ymmE, ymmE, ymmF ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U) 351 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U) 352 353 vpunpcklbw ymmF, ymmF, ymmH 354 vpunpckhbw ymmH, ymmH, ymmH 355 vpsrlw ymmF, ymmF, BYTE_BIT ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V) 356 vpsrlw ymmH, ymmH, BYTE_BIT ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V) 357 358%endif ; RGB_PIXELSIZE ; --------------- 359 360 ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE 361 ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO 362 363 ; (Original) 364 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B 365 ; 366 ; (This implementation) 367 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G 368 369 vmovdqa ymm6, ymm1 370 vpunpcklwd ymm1, ymm1, ymm3 371 vpunpckhwd ymm6, ymm6, ymm3 372 vpmaddwd ymm1, ymm1, [GOTOFF(eax,PW_F0299_F0337)] ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337) 373 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337) 374 375 vmovdqa ymm7, ymm6 ; ymm7=ROH*FIX(0.299)+GOH*FIX(0.337) 376 377 vmovdqa ymm6, ymm0 378 vpunpcklwd ymm0, ymm0, ymm2 379 vpunpckhwd ymm6, ymm6, ymm2 380 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0299_F0337)] ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337) 381 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0299_F0337)] ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337) 382 383 vmovdqa YMMWORD [wk(0)], ymm0 ; wk(0)=REL*FIX(0.299)+GEL*FIX(0.337) 384 vmovdqa YMMWORD [wk(1)], ymm6 ; wk(1)=REH*FIX(0.299)+GEH*FIX(0.337) 385 386 vmovdqa ymm0, ymm5 ; ymm0=BO 387 vmovdqa ymm6, ymm4 ; ymm6=BE 388 389 vmovdqa ymm4, ymm0 390 vpunpcklwd ymm0, ymm0, ymm3 391 vpunpckhwd ymm4, ymm4, ymm3 392 vpmaddwd ymm0, ymm0, [GOTOFF(eax,PW_F0114_F0250)] ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250) 393 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250) 394 395 vmovdqa ymm3, [GOTOFF(eax,PD_ONEHALF)] ; ymm3=[PD_ONEHALF] 396 397 vpaddd ymm0, ymm0, ymm1 398 vpaddd ymm4, ymm4, ymm7 399 vpaddd ymm0, ymm0, ymm3 400 vpaddd ymm4, ymm4, ymm3 401 vpsrld ymm0, ymm0, SCALEBITS ; ymm0=YOL 402 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YOH 403 vpackssdw ymm0, ymm0, ymm4 ; ymm0=YO 404 405 vmovdqa ymm4, ymm6 406 vpunpcklwd ymm6, ymm6, ymm2 407 vpunpckhwd ymm4, ymm4, ymm2 408 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_F0114_F0250)] ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250) 409 vpmaddwd ymm4, ymm4, [GOTOFF(eax,PW_F0114_F0250)] ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250) 410 411 vmovdqa ymm2, [GOTOFF(eax,PD_ONEHALF)] ; ymm2=[PD_ONEHALF] 412 413 vpaddd ymm6, ymm6, YMMWORD [wk(0)] 414 vpaddd ymm4, ymm4, YMMWORD [wk(1)] 415 vpaddd ymm6, ymm6, ymm2 416 vpaddd ymm4, ymm4, ymm2 417 vpsrld ymm6, ymm6, SCALEBITS ; ymm6=YEL 418 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YEH 419 vpackssdw ymm6, ymm6, ymm4 ; ymm6=YE 420 421 vpsllw ymm0, ymm0, BYTE_BIT 422 vpor ymm6, ymm6, ymm0 ; ymm6=Y 423 vmovdqu YMMWORD [edi], ymm6 ; Save Y 424 425 sub ecx, byte SIZEOF_YMMWORD 426 add esi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; inptr 427 add edi, byte SIZEOF_YMMWORD ; outptr0 428 cmp ecx, byte SIZEOF_YMMWORD 429 jae near .columnloop 430 test ecx, ecx 431 jnz near .column_ld1 432 433 pop ecx ; col 434 pop esi 435 pop edi 436 poppic eax 437 438 add esi, byte SIZEOF_JSAMPROW ; input_buf 439 add edi, byte SIZEOF_JSAMPROW 440 dec eax ; num_rows 441 jg near .rowloop 442 443.return: 444 vzeroupper 445 pop edi 446 pop esi 447; pop edx ; need not be preserved 448; pop ecx ; need not be preserved 449 pop ebx 450 mov esp, ebp ; esp <- aligned ebp 451 pop esp ; esp <- original ebp 452 pop ebp 453 ret 454 455; For some reason, the OS X linker does not honor the request to align the 456; segment unless we do this. 457 align 32 458