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; [TAB8] 19 20%include "jcolsamp.inc" 21 22; -------------------------------------------------------------------------- 23; 24; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical. 25; 26; GLOBAL(void) 27; jsimd_h2v1_merged_upsample_avx2(JDIMENSION output_width, 28; JSAMPIMAGE input_buf, 29; JDIMENSION in_row_group_ctr, 30; JSAMPARRAY output_buf); 31; 32 33%define output_width(b) (b) + 8 ; JDIMENSION output_width 34%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 35%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 36%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 37 38%define original_ebp ebp + 0 39%define wk(i) ebp - (WK_NUM - (i)) * SIZEOF_YMMWORD 40 ; ymmword wk[WK_NUM] 41%define WK_NUM 3 42%define gotptr wk(0) - SIZEOF_POINTER ; void * gotptr 43 44 align 32 45 GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2) 46 47EXTN(jsimd_h2v1_merged_upsample_avx2): 48 push ebp 49 mov eax, esp ; eax = original ebp 50 sub esp, byte 4 51 and esp, byte (-SIZEOF_YMMWORD) ; align to 256 bits 52 mov [esp], eax 53 mov ebp, esp ; ebp = aligned ebp 54 lea esp, [wk(0)] 55 pushpic eax ; make a room for GOT address 56 push ebx 57; push ecx ; need not be preserved 58; push edx ; need not be preserved 59 push esi 60 push edi 61 62 get_GOT ebx ; get GOT address 63 movpic POINTER [gotptr], ebx ; save GOT address 64 65 mov ecx, JDIMENSION [output_width(eax)] ; col 66 test ecx, ecx 67 jz near .return 68 69 push ecx 70 71 mov edi, JSAMPIMAGE [input_buf(eax)] 72 mov ecx, JDIMENSION [in_row_group_ctr(eax)] 73 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 74 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 75 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 76 mov edi, JSAMPARRAY [output_buf(eax)] 77 mov esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW] ; inptr0 78 mov ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW] ; inptr1 79 mov edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW] ; inptr2 80 mov edi, JSAMPROW [edi] ; outptr 81 82 pop ecx ; col 83 84 alignx 16, 7 85.columnloop: 86 movpic eax, POINTER [gotptr] ; load GOT address (eax) 87 88 vmovdqu ymm6, YMMWORD [ebx] ; ymm6=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV) 89 vmovdqu ymm7, YMMWORD [edx] ; ymm7=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV) 90 91 vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's) 92 vpcmpeqw ymm3, ymm3, ymm3 93 vpsllw ymm3, ymm3, 7 ; ymm3={0xFF80 0xFF80 0xFF80 0xFF80 ..} 94 95 vpermq ymm6, ymm6, 0xd8 ; ymm6=Cb(01234567GHIJKLMN89ABCDEFOPQRSTUV) 96 vpermq ymm7, ymm7, 0xd8 ; ymm7=Cr(01234567GHIJKLMN89ABCDEFOPQRSTUV) 97 vpunpcklbw ymm4, ymm6, ymm1 ; ymm4=Cb(0123456789ABCDEF)=CbL 98 vpunpckhbw ymm6, ymm6, ymm1 ; ymm6=Cb(GHIJKLMNOPQRSTUV)=CbH 99 vpunpcklbw ymm0, ymm7, ymm1 ; ymm0=Cr(0123456789ABCDEF)=CrL 100 vpunpckhbw ymm7, ymm7, ymm1 ; ymm7=Cr(GHIJKLMNOPQRSTUV)=CrH 101 102 vpaddw ymm5, ymm6, ymm3 103 vpaddw ymm2, ymm4, ymm3 104 vpaddw ymm1, ymm7, ymm3 105 vpaddw ymm3, ymm0, ymm3 106 107 ; (Original) 108 ; R = Y + 1.40200 * Cr 109 ; G = Y - 0.34414 * Cb - 0.71414 * Cr 110 ; B = Y + 1.77200 * Cb 111 ; 112 ; (This implementation) 113 ; R = Y + 0.40200 * Cr + Cr 114 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr 115 ; B = Y - 0.22800 * Cb + Cb + Cb 116 117 vpaddw ymm6, ymm5, ymm5 ; ymm6=2*CbH 118 vpaddw ymm4, ymm2, ymm2 ; ymm4=2*CbL 119 vpaddw ymm7, ymm1, ymm1 ; ymm7=2*CrH 120 vpaddw ymm0, ymm3, ymm3 ; ymm0=2*CrL 121 122 vpmulhw ymm6, ymm6, [GOTOFF(eax,PW_MF0228)] ; ymm6=(2*CbH * -FIX(0.22800)) 123 vpmulhw ymm4, ymm4, [GOTOFF(eax,PW_MF0228)] ; ymm4=(2*CbL * -FIX(0.22800)) 124 vpmulhw ymm7, ymm7, [GOTOFF(eax,PW_F0402)] ; ymm7=(2*CrH * FIX(0.40200)) 125 vpmulhw ymm0, ymm0, [GOTOFF(eax,PW_F0402)] ; ymm0=(2*CrL * FIX(0.40200)) 126 127 vpaddw ymm6, ymm6, [GOTOFF(eax,PW_ONE)] 128 vpaddw ymm4, ymm4, [GOTOFF(eax,PW_ONE)] 129 vpsraw ymm6, ymm6, 1 ; ymm6=(CbH * -FIX(0.22800)) 130 vpsraw ymm4, ymm4, 1 ; ymm4=(CbL * -FIX(0.22800)) 131 vpaddw ymm7, ymm7, [GOTOFF(eax,PW_ONE)] 132 vpaddw ymm0, ymm0, [GOTOFF(eax,PW_ONE)] 133 vpsraw ymm7, ymm7, 1 ; ymm7=(CrH * FIX(0.40200)) 134 vpsraw ymm0, ymm0, 1 ; ymm0=(CrL * FIX(0.40200)) 135 136 vpaddw ymm6, ymm6, ymm5 137 vpaddw ymm4, ymm4, ymm2 138 vpaddw ymm6, ymm6, ymm5 ; ymm6=(CbH * FIX(1.77200))=(B-Y)H 139 vpaddw ymm4, ymm4, ymm2 ; ymm4=(CbL * FIX(1.77200))=(B-Y)L 140 vpaddw ymm7, ymm7, ymm1 ; ymm7=(CrH * FIX(1.40200))=(R-Y)H 141 vpaddw ymm0, ymm0, ymm3 ; ymm0=(CrL * FIX(1.40200))=(R-Y)L 142 143 vmovdqa YMMWORD [wk(0)], ymm6 ; wk(0)=(B-Y)H 144 vmovdqa YMMWORD [wk(1)], ymm7 ; wk(1)=(R-Y)H 145 146 vpunpckhwd ymm6, ymm5, ymm1 147 vpunpcklwd ymm5, ymm5, ymm1 148 vpmaddwd ymm5, ymm5, [GOTOFF(eax,PW_MF0344_F0285)] 149 vpmaddwd ymm6, ymm6, [GOTOFF(eax,PW_MF0344_F0285)] 150 vpunpckhwd ymm7, ymm2, ymm3 151 vpunpcklwd ymm2, ymm2, ymm3 152 vpmaddwd ymm2, ymm2, [GOTOFF(eax,PW_MF0344_F0285)] 153 vpmaddwd ymm7, ymm7, [GOTOFF(eax,PW_MF0344_F0285)] 154 155 vpaddd ymm5, ymm5, [GOTOFF(eax,PD_ONEHALF)] 156 vpaddd ymm6, ymm6, [GOTOFF(eax,PD_ONEHALF)] 157 vpsrad ymm5, ymm5, SCALEBITS 158 vpsrad ymm6, ymm6, SCALEBITS 159 vpaddd ymm2, ymm2, [GOTOFF(eax,PD_ONEHALF)] 160 vpaddd ymm7, ymm7, [GOTOFF(eax,PD_ONEHALF)] 161 vpsrad ymm2, ymm2, SCALEBITS 162 vpsrad ymm7, ymm7, SCALEBITS 163 164 vpackssdw ymm5, ymm5, ymm6 ; ymm5=CbH*-FIX(0.344)+CrH*FIX(0.285) 165 vpackssdw ymm2, ymm2, ymm7 ; ymm2=CbL*-FIX(0.344)+CrL*FIX(0.285) 166 vpsubw ymm5, ymm5, ymm1 ; ymm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H 167 vpsubw ymm2, ymm2, ymm3 ; ymm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L 168 169 vmovdqa YMMWORD [wk(2)], ymm5 ; wk(2)=(G-Y)H 170 171 mov al, 2 ; Yctr 172 jmp short .Yloop_1st 173 alignx 16, 7 174 175.Yloop_2nd: 176 vmovdqa ymm0, YMMWORD [wk(1)] ; ymm0=(R-Y)H 177 vmovdqa ymm2, YMMWORD [wk(2)] ; ymm2=(G-Y)H 178 vmovdqa ymm4, YMMWORD [wk(0)] ; ymm4=(B-Y)H 179 alignx 16, 7 180 181.Yloop_1st: 182 vmovdqu ymm7, YMMWORD [esi] ; ymm7=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV) 183 184 vpcmpeqw ymm6, ymm6, ymm6 185 vpsrlw ymm6, ymm6, BYTE_BIT ; ymm6={0xFF 0x00 0xFF 0x00 ..} 186 vpand ymm6, ymm6, ymm7 ; ymm6=Y(02468ACEGIKMOQSU)=YE 187 vpsrlw ymm7, ymm7, BYTE_BIT ; ymm7=Y(13579BDFHJLNPRTV)=YO 188 189 vmovdqa ymm1, ymm0 ; ymm1=ymm0=(R-Y)(L/H) 190 vmovdqa ymm3, ymm2 ; ymm3=ymm2=(G-Y)(L/H) 191 vmovdqa ymm5, ymm4 ; ymm5=ymm4=(B-Y)(L/H) 192 193 vpaddw ymm0, ymm0, ymm6 ; ymm0=((R-Y)+YE)=RE=R(02468ACEGIKMOQSU) 194 vpaddw ymm1, ymm1, ymm7 ; ymm1=((R-Y)+YO)=RO=R(13579BDFHJLNPRTV) 195 vpackuswb ymm0, ymm0, ymm0 ; ymm0=R(02468ACE********GIKMOQSU********) 196 vpackuswb ymm1, ymm1, ymm1 ; ymm1=R(13579BDF********HJLNPRTV********) 197 198 vpaddw ymm2, ymm2, ymm6 ; ymm2=((G-Y)+YE)=GE=G(02468ACEGIKMOQSU) 199 vpaddw ymm3, ymm3, ymm7 ; ymm3=((G-Y)+YO)=GO=G(13579BDFHJLNPRTV) 200 vpackuswb ymm2, ymm2, ymm2 ; ymm2=G(02468ACE********GIKMOQSU********) 201 vpackuswb ymm3, ymm3, ymm3 ; ymm3=G(13579BDF********HJLNPRTV********) 202 203 vpaddw ymm4, ymm4, ymm6 ; ymm4=((B-Y)+YE)=BE=B(02468ACEGIKMOQSU) 204 vpaddw ymm5, ymm5, ymm7 ; ymm5=((B-Y)+YO)=BO=B(13579BDFHJLNPRTV) 205 vpackuswb ymm4, ymm4, ymm4 ; ymm4=B(02468ACE********GIKMOQSU********) 206 vpackuswb ymm5, ymm5, ymm5 ; ymm5=B(13579BDF********HJLNPRTV********) 207 208%if RGB_PIXELSIZE == 3 ; --------------- 209 210 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **) 211 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **) 212 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **) 213 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **) 214 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **) 215 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **) 216 ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **) 217 ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **) 218 219 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 220 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U) 221 vpunpcklbw ymmE, ymmE, ymmB ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 222 ; 2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V) 223 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 224 ; 1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V) 225 226 vpsrldq ymmH, ymmA, 2 ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G 227 ; 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --) 228 vpunpckhwd ymmG, ymmA, ymmE ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F 229 ; 0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V) 230 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07 231 ; 0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N) 232 233 vpsrldq ymmE, ymmE, 2 ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H 234 ; 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --) 235 236 vpsrldq ymmB, ymmD, 2 ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H 237 ; 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --) 238 vpunpckhwd ymmC, ymmD, ymmH ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G 239 ; 1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --) 240 vpunpcklwd ymmD, ymmD, ymmH ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18 241 ; 1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O) 242 243 vpunpckhwd ymmF, ymmE, ymmB ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H 244 ; 2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --) 245 vpunpcklwd ymmE, ymmE, ymmB ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29 246 ; 2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P) 247 248 vpshufd ymmH, ymmA, 0x4E ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03 249 ; 0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J) 250 vpunpckldq ymmA, ymmA, ymmD ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14 251 ; 0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K) 252 vpunpckhdq ymmD, ymmD, ymmE ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29 253 ; 1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P) 254 vpunpckldq ymmE, ymmE, ymmH ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07 255 ; 2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N) 256 257 vpshufd ymmH, ymmG, 0x4E ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B 258 ; 0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R) 259 vpunpckldq ymmG, ymmG, ymmC ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C 260 ; 0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S) 261 vpunpckhdq ymmC, ymmC, ymmF ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H 262 ; 1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --) 263 vpunpckldq ymmF, ymmF, ymmH ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F 264 ; 2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V) 265 266 vpunpcklqdq ymmH, ymmA, ymmE ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 267 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 268 vpunpcklqdq ymmG, ymmD, ymmG ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A 269 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q) 270 vpunpcklqdq ymmC, ymmF, ymmC ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 271 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 272 273 vperm2i128 ymmA, ymmH, ymmG, 0x20 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05 274 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) 275 vperm2i128 ymmD, ymmC, ymmH, 0x30 ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F 276 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L) 277 vperm2i128 ymmF, ymmG, ymmC, 0x31 ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q 278 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V) 279 280 cmp ecx, byte SIZEOF_YMMWORD 281 jb short .column_st64 282 283 test edi, SIZEOF_YMMWORD-1 284 jnz short .out1 285 ; --(aligned)------------------- 286 vmovntdq YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 287 vmovntdq YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 288 vmovntdq YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF 289 jmp short .out0 290.out1: ; --(unaligned)----------------- 291 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 292 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 293 vmovdqu YMMWORD [edi+2*SIZEOF_YMMWORD], ymmF 294.out0: 295 add edi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr 296 sub ecx, byte SIZEOF_YMMWORD 297 jz near .endcolumn 298 299 add esi, byte SIZEOF_YMMWORD ; inptr0 300 dec al ; Yctr 301 jnz near .Yloop_2nd 302 303 add ebx, byte SIZEOF_YMMWORD ; inptr1 304 add edx, byte SIZEOF_YMMWORD ; inptr2 305 jmp near .columnloop 306 alignx 16, 7 307 308.column_st64: 309 lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE 310 cmp ecx, byte 2*SIZEOF_YMMWORD 311 jb short .column_st32 312 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 313 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 314 add edi, byte 2*SIZEOF_YMMWORD ; outptr 315 vmovdqa ymmA, ymmF 316 sub ecx, byte 2*SIZEOF_YMMWORD 317 jmp short .column_st31 318.column_st32: 319 cmp ecx, byte SIZEOF_YMMWORD 320 jb short .column_st31 321 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 322 add edi, byte SIZEOF_YMMWORD ; outptr 323 vmovdqa ymmA, ymmD 324 sub ecx, byte SIZEOF_YMMWORD 325 jmp short .column_st31 326.column_st31: 327 cmp ecx, byte SIZEOF_XMMWORD 328 jb short .column_st15 329 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA 330 add edi, byte SIZEOF_XMMWORD ; outptr 331 vperm2i128 ymmA, ymmA, ymmA, 1 332 sub ecx, byte SIZEOF_XMMWORD 333.column_st15: 334 ; Store the lower 8 bytes of xmmA to the output when it has enough 335 ; space. 336 cmp ecx, byte SIZEOF_MMWORD 337 jb short .column_st7 338 vmovq XMM_MMWORD [edi], xmmA 339 add edi, byte SIZEOF_MMWORD 340 sub ecx, byte SIZEOF_MMWORD 341 vpsrldq xmmA, xmmA, SIZEOF_MMWORD 342.column_st7: 343 ; Store the lower 4 bytes of xmmA to the output when it has enough 344 ; space. 345 cmp ecx, byte SIZEOF_DWORD 346 jb short .column_st3 347 vmovd XMM_DWORD [edi], xmmA 348 add edi, byte SIZEOF_DWORD 349 sub ecx, byte SIZEOF_DWORD 350 vpsrldq xmmA, xmmA, SIZEOF_DWORD 351.column_st3: 352 ; Store the lower 2 bytes of eax to the output when it has enough 353 ; space. 354 vmovd eax, xmmA 355 cmp ecx, byte SIZEOF_WORD 356 jb short .column_st1 357 mov WORD [edi], ax 358 add edi, byte SIZEOF_WORD 359 sub ecx, byte SIZEOF_WORD 360 shr eax, 16 361.column_st1: 362 ; Store the lower 1 byte of eax to the output when it has enough 363 ; space. 364 test ecx, ecx 365 jz short .endcolumn 366 mov BYTE [edi], al 367 368%else ; RGB_PIXELSIZE == 4 ; ----------- 369 370%ifdef RGBX_FILLER_0XFF 371 vpcmpeqb ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********) 372 vpcmpeqb ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********) 373%else 374 vpxor ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********) 375 vpxor ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********) 376%endif 377 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **) 378 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **) 379 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **) 380 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **) 381 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **) 382 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **) 383 ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **) 384 ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **) 385 386 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 387 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U) 388 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E 389 ; 2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U) 390 vpunpcklbw ymmB, ymmB, ymmD ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F 391 ; 0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V) 392 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F 393 ; 2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V) 394 395 vpunpckhwd ymmC, ymmA, ymmE ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E 396 ; 0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U) 397 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36 398 ; 0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M) 399 vpunpckhwd ymmG, ymmB, ymmF ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F 400 ; 0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V) 401 vpunpcklwd ymmB, ymmB, ymmF ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37 402 ; 0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N) 403 404 vpunpckhdq ymmE, ymmA, ymmB ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37 405 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 406 vpunpckldq ymmB, ymmA, ymmB ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 407 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J) 408 vpunpckhdq ymmF, ymmC, ymmG ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F 409 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 410 vpunpckldq ymmG, ymmC, ymmG ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 411 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R) 412 413 vperm2i128 ymmA, ymmB, ymmE, 0x20 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33 414 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) 415 vperm2i128 ymmD, ymmG, ymmF, 0x20 ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B 416 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) 417 vperm2i128 ymmC, ymmB, ymmE, 0x31 ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J 418 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N) 419 vperm2i128 ymmH, ymmG, ymmF, 0x31 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R 420 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V) 421 422 cmp ecx, byte SIZEOF_YMMWORD 423 jb short .column_st64 424 425 test edi, SIZEOF_YMMWORD-1 426 jnz short .out1 427 ; --(aligned)------------------- 428 vmovntdq YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 429 vmovntdq YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 430 vmovntdq YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC 431 vmovntdq YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH 432 jmp short .out0 433.out1: ; --(unaligned)----------------- 434 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 435 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 436 vmovdqu YMMWORD [edi+2*SIZEOF_YMMWORD], ymmC 437 vmovdqu YMMWORD [edi+3*SIZEOF_YMMWORD], ymmH 438.out0: 439 add edi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr 440 sub ecx, byte SIZEOF_YMMWORD 441 jz near .endcolumn 442 443 add esi, byte SIZEOF_YMMWORD ; inptr0 444 dec al 445 jnz near .Yloop_2nd 446 447 add ebx, byte SIZEOF_YMMWORD ; inptr1 448 add edx, byte SIZEOF_YMMWORD ; inptr2 449 jmp near .columnloop 450 alignx 16, 7 451 452.column_st64: 453 cmp ecx, byte SIZEOF_YMMWORD/2 454 jb short .column_st32 455 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 456 vmovdqu YMMWORD [edi+1*SIZEOF_YMMWORD], ymmD 457 add edi, byte 2*SIZEOF_YMMWORD ; outptr 458 vmovdqa ymmA, ymmC 459 vmovdqa ymmD, ymmH 460 sub ecx, byte SIZEOF_YMMWORD/2 461.column_st32: 462 cmp ecx, byte SIZEOF_YMMWORD/4 463 jb short .column_st16 464 vmovdqu YMMWORD [edi+0*SIZEOF_YMMWORD], ymmA 465 add edi, byte SIZEOF_YMMWORD ; outptr 466 vmovdqa ymmA, ymmD 467 sub ecx, byte SIZEOF_YMMWORD/4 468.column_st16: 469 cmp ecx, byte SIZEOF_YMMWORD/8 470 jb short .column_st15 471 vmovdqu XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA 472 add edi, byte SIZEOF_XMMWORD ; outptr 473 vperm2i128 ymmA, ymmA, ymmA, 1 474 sub ecx, byte SIZEOF_YMMWORD/8 475.column_st15: 476 ; Store two pixels (8 bytes) of ymmA to the output when it has enough 477 ; space. 478 cmp ecx, byte SIZEOF_YMMWORD/16 479 jb short .column_st7 480 vmovq MMWORD [edi], xmmA 481 add edi, byte SIZEOF_YMMWORD/16*4 482 sub ecx, byte SIZEOF_YMMWORD/16 483 vpsrldq xmmA, SIZEOF_YMMWORD/16*4 484.column_st7: 485 ; Store one pixel (4 bytes) of ymmA to the output when it has enough 486 ; space. 487 test ecx, ecx 488 jz short .endcolumn 489 vmovd XMM_DWORD [edi], xmmA 490 491%endif ; RGB_PIXELSIZE ; --------------- 492 493.endcolumn: 494 sfence ; flush the write buffer 495 496.return: 497 vzeroupper 498 pop edi 499 pop esi 500; pop edx ; need not be preserved 501; pop ecx ; need not be preserved 502 pop ebx 503 mov esp, ebp ; esp <- aligned ebp 504 pop esp ; esp <- original ebp 505 pop ebp 506 ret 507 508; -------------------------------------------------------------------------- 509; 510; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical. 511; 512; GLOBAL(void) 513; jsimd_h2v2_merged_upsample_avx2(JDIMENSION output_width, 514; JSAMPIMAGE input_buf, 515; JDIMENSION in_row_group_ctr, 516; JSAMPARRAY output_buf); 517; 518 519%define output_width(b) (b) + 8 ; JDIMENSION output_width 520%define input_buf(b) (b) + 12 ; JSAMPIMAGE input_buf 521%define in_row_group_ctr(b) (b) + 16 ; JDIMENSION in_row_group_ctr 522%define output_buf(b) (b) + 20 ; JSAMPARRAY output_buf 523 524 align 32 525 GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2) 526 527EXTN(jsimd_h2v2_merged_upsample_avx2): 528 push ebp 529 mov ebp, esp 530 push ebx 531; push ecx ; need not be preserved 532; push edx ; need not be preserved 533 push esi 534 push edi 535 536 mov eax, POINTER [output_width(ebp)] 537 538 mov edi, JSAMPIMAGE [input_buf(ebp)] 539 mov ecx, JDIMENSION [in_row_group_ctr(ebp)] 540 mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] 541 mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] 542 mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] 543 mov edi, JSAMPARRAY [output_buf(ebp)] 544 lea esi, [esi+ecx*SIZEOF_JSAMPROW] 545 546 push edx ; inptr2 547 push ebx ; inptr1 548 push esi ; inptr00 549 mov ebx, esp 550 551 push edi ; output_buf (outptr0) 552 push ecx ; in_row_group_ctr 553 push ebx ; input_buf 554 push eax ; output_width 555 556 call near EXTN(jsimd_h2v1_merged_upsample_avx2) 557 558 add esi, byte SIZEOF_JSAMPROW ; inptr01 559 add edi, byte SIZEOF_JSAMPROW ; outptr1 560 mov POINTER [ebx+0*SIZEOF_POINTER], esi 561 mov POINTER [ebx-1*SIZEOF_POINTER], edi 562 563 call near EXTN(jsimd_h2v1_merged_upsample_avx2) 564 565 add esp, byte 7*SIZEOF_DWORD 566 567 pop edi 568 pop esi 569; pop edx ; need not be preserved 570; pop ecx ; need not be preserved 571 pop ebx 572 pop ebp 573 ret 574 575; For some reason, the OS X linker does not honor the request to align the 576; segment unless we do this. 577 align 32 578