1; 2; jdmrgext.asm - merged upsampling/color conversion (AVX2) 3; 4; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB 5; Copyright (C) 2012, 2016, D. R. Commander. 6; Copyright (C) 2015, Intel Corporation. 7; 8; Based on the x86 SIMD extension for IJG JPEG library 9; Copyright (C) 1999-2006, MIYASAKA Masaru. 10; For conditions of distribution and use, see copyright notice in jsimdext.inc 11; 12; This file should be assembled with NASM (Netwide Assembler), 13; can *not* be assembled with Microsoft's MASM or any compatible 14; assembler (including Borland's Turbo Assembler). 15; NASM is available from http://nasm.sourceforge.net/ or 16; http://sourceforge.net/project/showfiles.php?group_id=6208 17 18%include "jcolsamp.inc" 19 20; -------------------------------------------------------------------------- 21; 22; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical. 23; 24; GLOBAL(void) 25; jsimd_h2v1_merged_upsample_avx2(JDIMENSION output_width, 26; JSAMPIMAGE input_buf, 27; JDIMENSION in_row_group_ctr, 28; JSAMPARRAY output_buf); 29; 30 31%define output_width(b) (b) + 8 ; JDIMENSION output_width 32%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 33%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 34%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 35 36%define original_ebp ebp + 0 37%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 38 ; ymmword wk[WK_NUM] 39%define WK_NUM 3 40%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 41 42 align 32 43 GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2) 44 45EXTN(jsimd_h2v1_merged_upsample_avx2): 46 push ebp 47 mov eax, esp ; eax = original ebp 48 sub esp, byte 4 49 and esp, byte (-SIZEOF_YMMWORD) ; align to 256 bits 50 mov [esp], eax 51 mov ebp, esp ; ebp = aligned ebp 52 lea esp, [wk(0)] 53 pushpic eax ; make a room for GOT address 54 push ebx 55; push ecx ; need not be preserved 56; push edx ; need not be preserved 57 push esi 58 push edi 59 60 get_GOT ebx ; get GOT address 61 movpic POINTER [gotptr], ebx ; save GOT address 62 63 mov ecx, JDIMENSION [output_width(eax)] ; col 64 test ecx, ecx 65 jz near .return 66 67 push ecx 68 69 mov edi, JSAMPIMAGE [input_buf(eax)] 70 mov ecx, JDIMENSION [in_row_group_ctr(eax)] 71 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 72 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 73 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 74 mov edi, JSAMPARRAY [output_buf(eax)] 75 mov esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW] ; inptr0 76 mov ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW] ; inptr1 77 mov edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW] ; inptr2 78 mov edi, JSAMPROW [edi] ; outptr 79 80 pop ecx ; col 81 82 alignx 16, 7 83.columnloop: 84 movpic eax, POINTER [gotptr] ; load GOT address (eax) 85 86 vmovdqu ymm6, YMMWORD [ebx] ; ymm6=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV) 87 vmovdqu ymm7, YMMWORD [edx] ; ymm7=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV) 88 89 vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's) 90 vpcmpeqw ymm3, ymm3, ymm3 91 vpsllw ymm3, ymm3, 7 ; ymm3={0xFF80 0xFF80 0xFF80 0xFF80 ..} 92 93 vpermq ymm6, ymm6, 0xd8 ; ymm6=Cb(01234567GHIJKLMN89ABCDEFOPQRSTUV) 94 vpermq ymm7, ymm7, 0xd8 ; ymm7=Cr(01234567GHIJKLMN89ABCDEFOPQRSTUV) 95 vpunpcklbw ymm4, ymm6, ymm1 ; ymm4=Cb(0123456789ABCDEF)=CbL 96 vpunpckhbw ymm6, ymm6, ymm1 ; ymm6=Cb(GHIJKLMNOPQRSTUV)=CbH 97 vpunpcklbw ymm0, ymm7, ymm1 ; ymm0=Cr(0123456789ABCDEF)=CrL 98 vpunpckhbw ymm7, ymm7, ymm1 ; ymm7=Cr(GHIJKLMNOPQRSTUV)=CrH 99 100 vpaddw ymm5, ymm6, ymm3 101 vpaddw ymm2, ymm4, ymm3 102 vpaddw ymm1, ymm7, ymm3 103 vpaddw ymm3, ymm0, ymm3 104 105 ; (Original) 106 ; R = Y + 1.40200 * Cr 107 ; G = Y - 0.34414 * Cb - 0.71414 * Cr 108 ; B = Y + 1.77200 * Cb 109 ; 110 ; (This implementation) 111 ; R = Y + 0.40200 * Cr + Cr 112 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 113 ; B = Y - 0.22800 * Cb + Cb + Cb 114 115 vpaddw ymm6, ymm5, ymm5 ; ymm6=2*CbH 116 vpaddw ymm4, ymm2, ymm2 ; ymm4=2*CbL 117 vpaddw ymm7, ymm1, ymm1 ; ymm7=2*CrH 118 vpaddw ymm0, ymm3, ymm3 ; ymm0=2*CrL 119 120 vpmulhw ymm6, ymm6, [GOTOFF(eax,PW_MF0228)] ; ymm6=(2*CbH * -FIX(0.22800)) 121 vpmulhw ymm4, ymm4, [GOTOFF(eax,PW_MF0228)] ; ymm4=(2*CbL * -FIX(0.22800)) 122 vpmulhw ymm7, ymm7, [GOTOFF(eax,PW_F0402)] ; ymm7=(2*CrH * FIX(0.40200)) 123 vpmulhw ymm0, ymm0, [GOTOFF(eax,PW_F0402)] ; ymm0=(2*CrL * FIX(0.40200)) 124 125 vpaddw ymm6, ymm6, [GOTOFF(eax,PW_ONE)] 126 vpaddw ymm4, ymm4, [GOTOFF(eax,PW_ONE)] 127 vpsraw ymm6, ymm6, 1 ; ymm6=(CbH * -FIX(0.22800)) 128 vpsraw ymm4, ymm4, 1 ; ymm4=(CbL * -FIX(0.22800)) 129 vpaddw ymm7, ymm7, [GOTOFF(eax,PW_ONE)] 130 vpaddw ymm0, ymm0, [GOTOFF(eax,PW_ONE)] 131 vpsraw ymm7, ymm7, 1 ; ymm7=(CrH * FIX(0.40200)) 132 vpsraw ymm0, ymm0, 1 ; ymm0=(CrL * FIX(0.40200)) 133 134 vpaddw ymm6, ymm6, ymm5 135 vpaddw ymm4, ymm4, ymm2 136 vpaddw ymm6, ymm6, ymm5 ; ymm6=(CbH * FIX(1.77200))=(B-Y)H 137 vpaddw ymm4, ymm4, ymm2 ; ymm4=(CbL * FIX(1.77200))=(B-Y)L 138 vpaddw ymm7, ymm7, ymm1 ; ymm7=(CrH * FIX(1.40200))=(R-Y)H 139 vpaddw ymm0, ymm0, ymm3 ; ymm0=(CrL * FIX(1.40200))=(R-Y)L 140 141 vmovdqa YMMWORD [wk(0)], ymm6 ; wk(0)=(B-Y)H 142 vmovdqa YMMWORD [wk(1)], ymm7 ; wk(1)=(R-Y)H 143 144 vpunpckhwd ymm6, ymm5, ymm1 145 vpunpcklwd ymm5, ymm5, ymm1 146 vpmaddwd ymm5, ymm5, [GOTOFF(eax,PW_MF0344_F0285)] 147 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_MF0344_F0285)] 148 vpunpckhwd ymm7, ymm2, ymm3 149 vpunpcklwd ymm2, ymm2, ymm3 150 vpmaddwd ymm2, ymm2, [GOTOFF(eax,PW_MF0344_F0285)] 151 vpmaddwd ymm7, ymm7, [GOTOFF(eax,PW_MF0344_F0285)] 152 153 vpaddd ymm5, ymm5, [GOTOFF(eax,PD_ONEHALF)] 154 vpaddd ymm6, ymm6, [GOTOFF(eax,PD_ONEHALF)] 155 vpsrad ymm5, ymm5, SCALEBITS 156 vpsrad ymm6, ymm6, SCALEBITS 157 vpaddd ymm2, ymm2, [GOTOFF(eax,PD_ONEHALF)] 158 vpaddd ymm7, ymm7, [GOTOFF(eax,PD_ONEHALF)] 159 vpsrad ymm2, ymm2, SCALEBITS 160 vpsrad ymm7, ymm7, SCALEBITS 161 162 vpackssdw ymm5, ymm5, ymm6 ; ymm5=CbH*-FIX(0.344)+CrH*FIX(0.285) 163 vpackssdw ymm2, ymm2, ymm7 ; ymm2=CbL*-FIX(0.344)+CrL*FIX(0.285) 164 vpsubw ymm5, ymm5, ymm1 ; ymm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H 165 vpsubw ymm2, ymm2, ymm3 ; ymm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L 166 167 vmovdqa YMMWORD [wk(2)], ymm5 ; wk(2)=(G-Y)H 168 169 mov al, 2 ; Yctr 170 jmp short .Yloop_1st 171 alignx 16, 7 172 173.Yloop_2nd: 174 vmovdqa ymm0, YMMWORD [wk(1)] ; ymm0=(R-Y)H 175 vmovdqa ymm2, YMMWORD [wk(2)] ; ymm2=(G-Y)H 176 vmovdqa ymm4, YMMWORD [wk(0)] ; ymm4=(B-Y)H 177 alignx 16, 7 178 179.Yloop_1st: 180 vmovdqu ymm7, YMMWORD [esi] ; ymm7=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV) 181 182 vpcmpeqw ymm6, ymm6, ymm6 183 vpsrlw ymm6, ymm6, BYTE_BIT ; ymm6={0xFF 0x00 0xFF 0x00 ..} 184 vpand ymm6, ymm6, ymm7 ; ymm6=Y(02468ACEGIKMOQSU)=YE 185 vpsrlw ymm7, ymm7, BYTE_BIT ; ymm7=Y(13579BDFHJLNPRTV)=YO 186 187 vmovdqa ymm1, ymm0 ; ymm1=ymm0=(R-Y)(L/H) 188 vmovdqa ymm3, ymm2 ; ymm3=ymm2=(G-Y)(L/H) 189 vmovdqa ymm5, ymm4 ; ymm5=ymm4=(B-Y)(L/H) 190 191 vpaddw ymm0, ymm0, ymm6 ; ymm0=((R-Y)+YE)=RE=R(02468ACEGIKMOQSU) 192 vpaddw ymm1, ymm1, ymm7 ; ymm1=((R-Y)+YO)=RO=R(13579BDFHJLNPRTV) 193 vpackuswb ymm0, ymm0, ymm0 ; ymm0=R(02468ACE********GIKMOQSU********) 194 vpackuswb ymm1, ymm1, ymm1 ; ymm1=R(13579BDF********HJLNPRTV********) 195 196 vpaddw ymm2, ymm2, ymm6 ; ymm2=((G-Y)+YE)=GE=G(02468ACEGIKMOQSU) 197 vpaddw ymm3, ymm3, ymm7 ; ymm3=((G-Y)+YO)=GO=G(13579BDFHJLNPRTV) 198 vpackuswb ymm2, ymm2, ymm2 ; ymm2=G(02468ACE********GIKMOQSU********) 199 vpackuswb ymm3, ymm3, ymm3 ; ymm3=G(13579BDF********HJLNPRTV********) 200 201 vpaddw ymm4, ymm4, ymm6 ; ymm4=((B-Y)+YE)=BE=B(02468ACEGIKMOQSU) 202 vpaddw ymm5, ymm5, ymm7 ; ymm5=((B-Y)+YO)=BO=B(13579BDFHJLNPRTV) 203 vpackuswb ymm4, ymm4, ymm4 ; ymm4=B(02468ACE********GIKMOQSU********) 204 vpackuswb ymm5, ymm5, ymm5 ; ymm5=B(13579BDF********HJLNPRTV********) 205 206%if RGB_PIXELSIZE == 3 ; --------------- 207 208 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **) 209 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **) 210 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **) 211 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **) 212 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **) 213 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **) 214 ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **) 215 ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **) 216 217 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 218 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U) 219 vpunpcklbw ymmE, ymmE, ymmB ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 220 ; 2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V) 221 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 222 ; 1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V) 223 224 vpsrldq ymmH, ymmA, 2 ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G 225 ; 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --) 226 vpunpckhwd ymmG, ymmA, ymmE ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F 227 ; 0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V) 228 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07 229 ; 0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N) 230 231 vpsrldq ymmE, ymmE, 2 ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H 232 ; 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --) 233 234 vpsrldq ymmB, ymmD, 2 ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H 235 ; 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --) 236 vpunpckhwd ymmC, ymmD, ymmH ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G 237 ; 1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --) 238 vpunpcklwd ymmD, ymmD, ymmH ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18 239 ; 1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O) 240 241 vpunpckhwd ymmF, ymmE, ymmB ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H 242 ; 2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --) 243 vpunpcklwd ymmE, ymmE, ymmB ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29 244 ; 2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P) 245 246 vpshufd ymmH, ymmA, 0x4E ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03 247 ; 0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J) 248 vpunpckldq ymmA, ymmA, ymmD ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14 249 ; 0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K) 250 vpunpckhdq ymmD, ymmD, ymmE ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29 251 ; 1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P) 252 vpunpckldq ymmE, ymmE, ymmH ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07 253 ; 2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N) 254 255 vpshufd ymmH, ymmG, 0x4E ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B 256 ; 0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R) 257 vpunpckldq ymmG, ymmG, ymmC ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C 258 ; 0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S) 259 vpunpckhdq ymmC, ymmC, ymmF ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H 260 ; 1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --) 261 vpunpckldq ymmF, ymmF, ymmH ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F 262 ; 2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V) 263 264 vpunpcklqdq ymmH, ymmA, ymmE ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 265 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 266 vpunpcklqdq ymmG, ymmD, ymmG ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A 267 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q) 268 vpunpcklqdq ymmC, ymmF, ymmC ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 269 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 270 271 vperm2i128 ymmA, ymmH, ymmG, 0x20 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 272 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 273 vperm2i128 ymmD, ymmC, ymmH, 0x30 ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 274 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 275 vperm2i128 ymmF, ymmG, ymmC, 0x31 ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 276 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 277 278 cmp ecx, byte SIZEOF_YMMWORD 279 jb short .column_st64 280 281 test edi, SIZEOF_YMMWORD-1 282 jnz short .out1 283 ; --(aligned)------------------- 284 vmovntdq YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 285 vmovntdq YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 286 vmovntdq YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF 287 jmp short .out0 288.out1: ; --(unaligned)----------------- 289 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 290 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 291 vmovdqu YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF 292.out0: 293 add edi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr 294 sub ecx, byte SIZEOF_YMMWORD 295 jz near .endcolumn 296 297 add esi, byte SIZEOF_YMMWORD ; inptr0 298 dec al ; Yctr 299 jnz near .Yloop_2nd 300 301 add ebx, byte SIZEOF_YMMWORD ; inptr1 302 add edx, byte SIZEOF_YMMWORD ; inptr2 303 jmp near .columnloop 304 alignx 16, 7 305 306.column_st64: 307 lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE 308 cmp ecx, byte 2*SIZEOF_YMMWORD 309 jb short .column_st32 310 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 311 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 312 add edi, byte 2*SIZEOF_YMMWORD ; outptr 313 vmovdqa ymmA, ymmF 314 sub ecx, byte 2*SIZEOF_YMMWORD 315 jmp short .column_st31 316.column_st32: 317 cmp ecx, byte SIZEOF_YMMWORD 318 jb short .column_st31 319 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 320 add edi, byte SIZEOF_YMMWORD ; outptr 321 vmovdqa ymmA, ymmD 322 sub ecx, byte SIZEOF_YMMWORD 323 jmp short .column_st31 324.column_st31: 325 cmp ecx, byte SIZEOF_XMMWORD 326 jb short .column_st15 327 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA 328 add edi, byte SIZEOF_XMMWORD ; outptr 329 vperm2i128 ymmA, ymmA, ymmA, 1 330 sub ecx, byte SIZEOF_XMMWORD 331.column_st15: 332 ; Store the lower 8 bytes of xmmA to the output when it has enough 333 ; space. 334 cmp ecx, byte SIZEOF_MMWORD 335 jb short .column_st7 336 vmovq XMM_MMWORD [edi], xmmA 337 add edi, byte SIZEOF_MMWORD 338 sub ecx, byte SIZEOF_MMWORD 339 vpsrldq xmmA, xmmA, SIZEOF_MMWORD 340.column_st7: 341 ; Store the lower 4 bytes of xmmA to the output when it has enough 342 ; space. 343 cmp ecx, byte SIZEOF_DWORD 344 jb short .column_st3 345 vmovd XMM_DWORD [edi], xmmA 346 add edi, byte SIZEOF_DWORD 347 sub ecx, byte SIZEOF_DWORD 348 vpsrldq xmmA, xmmA, SIZEOF_DWORD 349.column_st3: 350 ; Store the lower 2 bytes of eax to the output when it has enough 351 ; space. 352 vmovd eax, xmmA 353 cmp ecx, byte SIZEOF_WORD 354 jb short .column_st1 355 mov word [edi], ax 356 add edi, byte SIZEOF_WORD 357 sub ecx, byte SIZEOF_WORD 358 shr eax, 16 359.column_st1: 360 ; Store the lower 1 byte of eax to the output when it has enough 361 ; space. 362 test ecx, ecx 363 jz short .endcolumn 364 mov byte [edi], al 365 366%else ; RGB_PIXELSIZE == 4 ; ----------- 367 368%ifdef RGBX_FILLER_0XFF 369 vpcmpeqb ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********) 370 vpcmpeqb ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********) 371%else 372 vpxor ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********) 373 vpxor ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********) 374%endif 375 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **) 376 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **) 377 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **) 378 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **) 379 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **) 380 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **) 381 ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **) 382 ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **) 383 384 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 385 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U) 386 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E 387 ; 2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U) 388 vpunpcklbw ymmB, ymmB, ymmD ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F 389 ; 0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V) 390 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F 391 ; 2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V) 392 393 vpunpckhwd ymmC, ymmA, ymmE ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E 394 ; 0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U) 395 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36 396 ; 0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M) 397 vpunpckhwd ymmG, ymmB, ymmF ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F 398 ; 0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V) 399 vpunpcklwd ymmB, ymmB, ymmF ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37 400 ; 0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N) 401 402 vpunpckhdq ymmE, ymmA, ymmB ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 403 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 404 vpunpckldq ymmB, ymmA, ymmB ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 405 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J) 406 vpunpckhdq ymmF, ymmC, ymmG ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F 407 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 408 vpunpckldq ymmG, ymmC, ymmG ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 409 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R) 410 411 vperm2i128 ymmA, ymmB, ymmE, 0x20 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 412 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) 413 vperm2i128 ymmD, ymmG, ymmF, 0x20 ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 414 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) 415 vperm2i128 ymmC, ymmB, ymmE, 0x31 ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J 416 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 417 vperm2i128 ymmH, ymmG, ymmF, 0x31 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R 418 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 419 420 cmp ecx, byte SIZEOF_YMMWORD 421 jb short .column_st64 422 423 test edi, SIZEOF_YMMWORD-1 424 jnz short .out1 425 ; --(aligned)------------------- 426 vmovntdq YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 427 vmovntdq YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 428 vmovntdq YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC 429 vmovntdq YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH 430 jmp short .out0 431.out1: ; --(unaligned)----------------- 432 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 433 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 434 vmovdqu YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC 435 vmovdqu YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH 436.out0: 437 add edi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr 438 sub ecx, byte SIZEOF_YMMWORD 439 jz near .endcolumn 440 441 add esi, byte SIZEOF_YMMWORD ; inptr0 442 dec al 443 jnz near .Yloop_2nd 444 445 add ebx, byte SIZEOF_YMMWORD ; inptr1 446 add edx, byte SIZEOF_YMMWORD ; inptr2 447 jmp near .columnloop 448 alignx 16, 7 449 450.column_st64: 451 cmp ecx, byte SIZEOF_YMMWORD/2 452 jb short .column_st32 453 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 454 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 455 add edi, byte 2*SIZEOF_YMMWORD ; outptr 456 vmovdqa ymmA, ymmC 457 vmovdqa ymmD, ymmH 458 sub ecx, byte SIZEOF_YMMWORD/2 459.column_st32: 460 cmp ecx, byte SIZEOF_YMMWORD/4 461 jb short .column_st16 462 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 463 add edi, byte SIZEOF_YMMWORD ; outptr 464 vmovdqa ymmA, ymmD 465 sub ecx, byte SIZEOF_YMMWORD/4 466.column_st16: 467 cmp ecx, byte SIZEOF_YMMWORD/8 468 jb short .column_st15 469 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA 470 add edi, byte SIZEOF_XMMWORD ; outptr 471 vperm2i128 ymmA, ymmA, ymmA, 1 472 sub ecx, byte SIZEOF_YMMWORD/8 473.column_st15: 474 ; Store two pixels (8 bytes) of ymmA to the output when it has enough 475 ; space. 476 cmp ecx, byte SIZEOF_YMMWORD/16 477 jb short .column_st7 478 vmovq MMWORD [edi], xmmA 479 add edi, byte SIZEOF_YMMWORD/16*4 480 sub ecx, byte SIZEOF_YMMWORD/16 481 vpsrldq xmmA, SIZEOF_YMMWORD/16*4 482.column_st7: 483 ; Store one pixel (4 bytes) of ymmA to the output when it has enough 484 ; space. 485 test ecx, ecx 486 jz short .endcolumn 487 vmovd XMM_DWORD [edi], xmmA 488 489%endif ; RGB_PIXELSIZE ; --------------- 490 491.endcolumn: 492 sfence ; flush the write buffer 493 494.return: 495 vzeroupper 496 pop edi 497 pop esi 498; pop edx ; need not be preserved 499; pop ecx ; need not be preserved 500 pop ebx 501 mov esp, ebp ; esp <- aligned ebp 502 pop esp ; esp <- original ebp 503 pop ebp 504 ret 505 506; -------------------------------------------------------------------------- 507; 508; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical. 509; 510; GLOBAL(void) 511; jsimd_h2v2_merged_upsample_avx2(JDIMENSION output_width, 512; JSAMPIMAGE input_buf, 513; JDIMENSION in_row_group_ctr, 514; JSAMPARRAY output_buf); 515; 516 517%define output_width(b) (b) + 8 ; JDIMENSION output_width 518%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 519%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 520%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 521 522 align 32 523 GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2) 524 525EXTN(jsimd_h2v2_merged_upsample_avx2): 526 push ebp 527 mov ebp, esp 528 push ebx 529; push ecx ; need not be preserved 530; push edx ; need not be preserved 531 push esi 532 push edi 533 534 mov eax, POINTER [output_width(ebp)] 535 536 mov edi, JSAMPIMAGE [input_buf(ebp)] 537 mov ecx, JDIMENSION [in_row_group_ctr(ebp)] 538 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 539 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 540 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 541 mov edi, JSAMPARRAY [output_buf(ebp)] 542 lea esi, [esi+ecx*SIZEOF_JSAMPROW] 543 544 push edx ; inptr2 545 push ebx ; inptr1 546 push esi ; inptr00 547 mov ebx, esp 548 549 push edi ; output_buf (outptr0) 550 push ecx ; in_row_group_ctr 551 push ebx ; input_buf 552 push eax ; output_width 553 554 call near EXTN(jsimd_h2v1_merged_upsample_avx2) 555 556 add esi, byte SIZEOF_JSAMPROW ; inptr01 557 add edi, byte SIZEOF_JSAMPROW ; outptr1 558 mov POINTER [ebx+0*SIZEOF_POINTER], esi 559 mov POINTER [ebx-1*SIZEOF_POINTER], edi 560 561 call near EXTN(jsimd_h2v1_merged_upsample_avx2) 562 563 add esp, byte 7*SIZEOF_DWORD 564 565 pop edi 566 pop esi 567; pop edx ; need not be preserved 568; pop ecx ; need not be preserved 569 pop ebx 570 pop ebp 571 ret 572 573; For some reason, the OS X linker does not honor the request to align the 574; segment unless we do this. 575 align 32 576