DRC | 720e161 | 2009-04-05 21:51:25 +0000 | [diff] [blame] | 1 | ; |
| 2 | ; jdmrgss2.asm - merged upsampling/color conversion (SSE2) |
| 3 | ; |
| 4 | ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB |
| 5 | ; |
| 6 | ; Based on |
| 7 | ; 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 | ; [TAB8] |
| 18 | |
| 19 | %include "jcolsamp.inc" |
| 20 | |
| 21 | ; -------------------------------------------------------------------------- |
| 22 | SECTION SEG_TEXT |
| 23 | BITS 32 |
| 24 | ; |
| 25 | ; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical. |
| 26 | ; |
| 27 | ; GLOBAL(void) |
| 28 | ; jsimd_h2v1_merged_upsample_sse2 (JDIMENSION output_width, |
| 29 | ; JSAMPIMAGE input_buf, |
| 30 | ; JDIMENSION in_row_group_ctr, |
| 31 | ; JSAMPARRAY output_buf); |
| 32 | ; |
| 33 | |
| 34 | %define output_width(b) (b)+8 ; JDIMENSION output_width |
| 35 | %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf |
| 36 | %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr |
| 37 | %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf |
| 38 | |
| 39 | %define original_ebp ebp+0 |
| 40 | %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] |
| 41 | %define WK_NUM 3 |
| 42 | %define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr |
| 43 | |
| 44 | align 16 |
| 45 | global EXTN(jsimd_h2v1_merged_upsample_sse2) |
| 46 | |
| 47 | EXTN(jsimd_h2v1_merged_upsample_sse2): |
| 48 | push ebp |
| 49 | mov eax,esp ; eax = original ebp |
| 50 | sub esp, byte 4 |
| 51 | and esp, byte (-SIZEOF_XMMWORD) ; align to 128 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 | movdqa xmm6, XMMWORD [ebx] ; xmm6=Cb(0123456789ABCDEF) |
| 89 | movdqa xmm7, XMMWORD [edx] ; xmm7=Cr(0123456789ABCDEF) |
| 90 | |
| 91 | pxor xmm1,xmm1 ; xmm1=(all 0's) |
| 92 | pcmpeqw xmm3,xmm3 |
| 93 | psllw xmm3,7 ; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..} |
| 94 | |
| 95 | movdqa xmm4,xmm6 |
| 96 | punpckhbw xmm6,xmm1 ; xmm6=Cb(89ABCDEF)=CbH |
| 97 | punpcklbw xmm4,xmm1 ; xmm4=Cb(01234567)=CbL |
| 98 | movdqa xmm0,xmm7 |
| 99 | punpckhbw xmm7,xmm1 ; xmm7=Cr(89ABCDEF)=CrH |
| 100 | punpcklbw xmm0,xmm1 ; xmm0=Cr(01234567)=CrL |
| 101 | |
| 102 | paddw xmm6,xmm3 |
| 103 | paddw xmm4,xmm3 |
| 104 | paddw xmm7,xmm3 |
| 105 | paddw xmm0,xmm3 |
| 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 | movdqa xmm5,xmm6 ; xmm5=CbH |
| 118 | movdqa xmm2,xmm4 ; xmm2=CbL |
| 119 | paddw xmm6,xmm6 ; xmm6=2*CbH |
| 120 | paddw xmm4,xmm4 ; xmm4=2*CbL |
| 121 | movdqa xmm1,xmm7 ; xmm1=CrH |
| 122 | movdqa xmm3,xmm0 ; xmm3=CrL |
| 123 | paddw xmm7,xmm7 ; xmm7=2*CrH |
| 124 | paddw xmm0,xmm0 ; xmm0=2*CrL |
| 125 | |
| 126 | pmulhw xmm6,[GOTOFF(eax,PW_MF0228)] ; xmm6=(2*CbH * -FIX(0.22800)) |
| 127 | pmulhw xmm4,[GOTOFF(eax,PW_MF0228)] ; xmm4=(2*CbL * -FIX(0.22800)) |
| 128 | pmulhw xmm7,[GOTOFF(eax,PW_F0402)] ; xmm7=(2*CrH * FIX(0.40200)) |
| 129 | pmulhw xmm0,[GOTOFF(eax,PW_F0402)] ; xmm0=(2*CrL * FIX(0.40200)) |
| 130 | |
| 131 | paddw xmm6,[GOTOFF(eax,PW_ONE)] |
| 132 | paddw xmm4,[GOTOFF(eax,PW_ONE)] |
| 133 | psraw xmm6,1 ; xmm6=(CbH * -FIX(0.22800)) |
| 134 | psraw xmm4,1 ; xmm4=(CbL * -FIX(0.22800)) |
| 135 | paddw xmm7,[GOTOFF(eax,PW_ONE)] |
| 136 | paddw xmm0,[GOTOFF(eax,PW_ONE)] |
| 137 | psraw xmm7,1 ; xmm7=(CrH * FIX(0.40200)) |
| 138 | psraw xmm0,1 ; xmm0=(CrL * FIX(0.40200)) |
| 139 | |
| 140 | paddw xmm6,xmm5 |
| 141 | paddw xmm4,xmm2 |
| 142 | paddw xmm6,xmm5 ; xmm6=(CbH * FIX(1.77200))=(B-Y)H |
| 143 | paddw xmm4,xmm2 ; xmm4=(CbL * FIX(1.77200))=(B-Y)L |
| 144 | paddw xmm7,xmm1 ; xmm7=(CrH * FIX(1.40200))=(R-Y)H |
| 145 | paddw xmm0,xmm3 ; xmm0=(CrL * FIX(1.40200))=(R-Y)L |
| 146 | |
| 147 | movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=(B-Y)H |
| 148 | movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(R-Y)H |
| 149 | |
| 150 | movdqa xmm6,xmm5 |
| 151 | movdqa xmm7,xmm2 |
| 152 | punpcklwd xmm5,xmm1 |
| 153 | punpckhwd xmm6,xmm1 |
| 154 | pmaddwd xmm5,[GOTOFF(eax,PW_MF0344_F0285)] |
| 155 | pmaddwd xmm6,[GOTOFF(eax,PW_MF0344_F0285)] |
| 156 | punpcklwd xmm2,xmm3 |
| 157 | punpckhwd xmm7,xmm3 |
| 158 | pmaddwd xmm2,[GOTOFF(eax,PW_MF0344_F0285)] |
| 159 | pmaddwd xmm7,[GOTOFF(eax,PW_MF0344_F0285)] |
| 160 | |
| 161 | paddd xmm5,[GOTOFF(eax,PD_ONEHALF)] |
| 162 | paddd xmm6,[GOTOFF(eax,PD_ONEHALF)] |
| 163 | psrad xmm5,SCALEBITS |
| 164 | psrad xmm6,SCALEBITS |
| 165 | paddd xmm2,[GOTOFF(eax,PD_ONEHALF)] |
| 166 | paddd xmm7,[GOTOFF(eax,PD_ONEHALF)] |
| 167 | psrad xmm2,SCALEBITS |
| 168 | psrad xmm7,SCALEBITS |
| 169 | |
| 170 | packssdw xmm5,xmm6 ; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285) |
| 171 | packssdw xmm2,xmm7 ; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285) |
| 172 | psubw xmm5,xmm1 ; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H |
| 173 | psubw xmm2,xmm3 ; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L |
| 174 | |
| 175 | movdqa XMMWORD [wk(2)], xmm5 ; wk(2)=(G-Y)H |
| 176 | |
| 177 | mov al,2 ; Yctr |
| 178 | jmp short .Yloop_1st |
| 179 | alignx 16,7 |
| 180 | |
| 181 | .Yloop_2nd: |
| 182 | movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(R-Y)H |
| 183 | movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(G-Y)H |
| 184 | movdqa xmm4, XMMWORD [wk(0)] ; xmm4=(B-Y)H |
| 185 | alignx 16,7 |
| 186 | |
| 187 | .Yloop_1st: |
| 188 | movdqa xmm7, XMMWORD [esi] ; xmm7=Y(0123456789ABCDEF) |
| 189 | |
| 190 | pcmpeqw xmm6,xmm6 |
| 191 | psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..} |
| 192 | pand xmm6,xmm7 ; xmm6=Y(02468ACE)=YE |
| 193 | psrlw xmm7,BYTE_BIT ; xmm7=Y(13579BDF)=YO |
| 194 | |
| 195 | movdqa xmm1,xmm0 ; xmm1=xmm0=(R-Y)(L/H) |
| 196 | movdqa xmm3,xmm2 ; xmm3=xmm2=(G-Y)(L/H) |
| 197 | movdqa xmm5,xmm4 ; xmm5=xmm4=(B-Y)(L/H) |
| 198 | |
| 199 | paddw xmm0,xmm6 ; xmm0=((R-Y)+YE)=RE=R(02468ACE) |
| 200 | paddw xmm1,xmm7 ; xmm1=((R-Y)+YO)=RO=R(13579BDF) |
| 201 | packuswb xmm0,xmm0 ; xmm0=R(02468ACE********) |
| 202 | packuswb xmm1,xmm1 ; xmm1=R(13579BDF********) |
| 203 | |
| 204 | paddw xmm2,xmm6 ; xmm2=((G-Y)+YE)=GE=G(02468ACE) |
| 205 | paddw xmm3,xmm7 ; xmm3=((G-Y)+YO)=GO=G(13579BDF) |
| 206 | packuswb xmm2,xmm2 ; xmm2=G(02468ACE********) |
| 207 | packuswb xmm3,xmm3 ; xmm3=G(13579BDF********) |
| 208 | |
| 209 | paddw xmm4,xmm6 ; xmm4=((B-Y)+YE)=BE=B(02468ACE) |
| 210 | paddw xmm5,xmm7 ; xmm5=((B-Y)+YO)=BO=B(13579BDF) |
| 211 | packuswb xmm4,xmm4 ; xmm4=B(02468ACE********) |
| 212 | packuswb xmm5,xmm5 ; xmm5=B(13579BDF********) |
| 213 | |
| 214 | %if RGB_PIXELSIZE == 3 ; --------------- |
| 215 | |
| 216 | ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **) |
| 217 | ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **) |
| 218 | ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **) |
| 219 | ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **) |
| 220 | |
| 221 | punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E) |
| 222 | punpcklbw xmmE,xmmB ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F) |
| 223 | punpcklbw xmmD,xmmF ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F) |
| 224 | |
| 225 | movdqa xmmG,xmmA |
| 226 | movdqa xmmH,xmmA |
| 227 | punpcklwd xmmA,xmmE ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07) |
| 228 | punpckhwd xmmG,xmmE ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F) |
| 229 | |
| 230 | psrldq xmmH,2 ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --) |
| 231 | psrldq xmmE,2 ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --) |
| 232 | |
| 233 | movdqa xmmC,xmmD |
| 234 | movdqa xmmB,xmmD |
| 235 | punpcklwd xmmD,xmmH ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18) |
| 236 | punpckhwd xmmC,xmmH ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --) |
| 237 | |
| 238 | psrldq xmmB,2 ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --) |
| 239 | |
| 240 | movdqa xmmF,xmmE |
| 241 | punpcklwd xmmE,xmmB ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29) |
| 242 | punpckhwd xmmF,xmmB ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --) |
| 243 | |
| 244 | pshufd xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03) |
| 245 | movdqa xmmB,xmmE |
| 246 | punpckldq xmmA,xmmD ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14) |
| 247 | punpckldq xmmE,xmmH ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07) |
| 248 | punpckhdq xmmD,xmmB ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29) |
| 249 | |
| 250 | pshufd xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B) |
| 251 | movdqa xmmB,xmmF |
| 252 | punpckldq xmmG,xmmC ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C) |
| 253 | punpckldq xmmF,xmmH ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F) |
| 254 | punpckhdq xmmC,xmmB ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --) |
| 255 | |
| 256 | punpcklqdq xmmA,xmmE ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05) |
| 257 | punpcklqdq xmmD,xmmG ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) |
| 258 | punpcklqdq xmmF,xmmC ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F) |
| 259 | |
| 260 | cmp ecx, byte SIZEOF_XMMWORD |
| 261 | jb short .column_st32 |
| 262 | |
| 263 | test edi, SIZEOF_XMMWORD-1 |
| 264 | jnz short .out1 |
| 265 | ; --(aligned)------------------- |
| 266 | movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA |
| 267 | movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD |
| 268 | movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF |
| 269 | add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr |
| 270 | jmp short .out0 |
| 271 | .out1: ; --(unaligned)----------------- |
| 272 | pcmpeqb xmmH,xmmH ; xmmH=(all 1's) |
| 273 | maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA |
| 274 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 275 | maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [edi], xmmD |
| 276 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 277 | maskmovdqu xmmF,xmmH ; movntdqu XMMWORD [edi], xmmF |
| 278 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 279 | .out0: |
| 280 | sub ecx, byte SIZEOF_XMMWORD |
| 281 | jz near .endcolumn |
| 282 | |
| 283 | add esi, byte SIZEOF_XMMWORD ; inptr0 |
| 284 | dec al ; Yctr |
| 285 | jnz near .Yloop_2nd |
| 286 | |
| 287 | add ebx, byte SIZEOF_XMMWORD ; inptr1 |
| 288 | add edx, byte SIZEOF_XMMWORD ; inptr2 |
| 289 | jmp near .columnloop |
| 290 | alignx 16,7 |
| 291 | |
| 292 | .column_st32: |
| 293 | pcmpeqb xmmH,xmmH ; xmmH=(all 1's) |
| 294 | lea ecx, [ecx+ecx*2] ; imul ecx, RGB_PIXELSIZE |
| 295 | cmp ecx, byte 2*SIZEOF_XMMWORD |
| 296 | jb short .column_st16 |
| 297 | maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA |
| 298 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 299 | maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [edi], xmmD |
| 300 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 301 | movdqa xmmA,xmmF |
| 302 | sub ecx, byte 2*SIZEOF_XMMWORD |
| 303 | jmp short .column_st15 |
| 304 | .column_st16: |
| 305 | cmp ecx, byte SIZEOF_XMMWORD |
| 306 | jb short .column_st15 |
| 307 | maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [edi], xmmA |
| 308 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 309 | movdqa xmmA,xmmD |
| 310 | sub ecx, byte SIZEOF_XMMWORD |
| 311 | .column_st15: |
| 312 | mov eax,ecx |
| 313 | xor ecx, byte 0x0F |
| 314 | shl ecx, 2 |
| 315 | movd xmmB,ecx |
| 316 | psrlq xmmH,4 |
| 317 | pcmpeqb xmmE,xmmE |
| 318 | psrlq xmmH,xmmB |
| 319 | psrlq xmmE,xmmB |
| 320 | punpcklbw xmmE,xmmH |
| 321 | ; ---------------- |
| 322 | mov ecx,edi |
| 323 | and ecx, byte SIZEOF_XMMWORD-1 |
| 324 | jz short .adj0 |
| 325 | add eax,ecx |
| 326 | cmp eax, byte SIZEOF_XMMWORD |
| 327 | ja short .adj0 |
| 328 | and edi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary |
| 329 | shl ecx, 3 ; pslldq xmmA,ecx & pslldq xmmE,ecx |
| 330 | movdqa xmmG,xmmA |
| 331 | movdqa xmmC,xmmE |
| 332 | pslldq xmmA, SIZEOF_XMMWORD/2 |
| 333 | pslldq xmmE, SIZEOF_XMMWORD/2 |
| 334 | movd xmmD,ecx |
| 335 | sub ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT |
| 336 | jb short .adj1 |
| 337 | movd xmmF,ecx |
| 338 | psllq xmmA,xmmF |
| 339 | psllq xmmE,xmmF |
| 340 | jmp short .adj0 |
| 341 | .adj1: neg ecx |
| 342 | movd xmmF,ecx |
| 343 | psrlq xmmA,xmmF |
| 344 | psrlq xmmE,xmmF |
| 345 | psllq xmmG,xmmD |
| 346 | psllq xmmC,xmmD |
| 347 | por xmmA,xmmG |
| 348 | por xmmE,xmmC |
| 349 | .adj0: ; ---------------- |
| 350 | maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA |
| 351 | |
| 352 | %else ; RGB_PIXELSIZE == 4 ; ----------- |
| 353 | |
| 354 | %ifdef RGBX_FILLER_0XFF |
| 355 | pcmpeqb xmm6,xmm6 ; xmm6=XE=X(02468ACE********) |
| 356 | pcmpeqb xmm7,xmm7 ; xmm7=XO=X(13579BDF********) |
| 357 | %else |
| 358 | pxor xmm6,xmm6 ; xmm6=XE=X(02468ACE********) |
| 359 | pxor xmm7,xmm7 ; xmm7=XO=X(13579BDF********) |
| 360 | %endif |
| 361 | ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **) |
| 362 | ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **) |
| 363 | ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **) |
| 364 | ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **) |
| 365 | |
| 366 | punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E) |
| 367 | punpcklbw xmmE,xmmG ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E) |
| 368 | punpcklbw xmmB,xmmD ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F) |
| 369 | punpcklbw xmmF,xmmH ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F) |
| 370 | |
| 371 | movdqa xmmC,xmmA |
| 372 | punpcklwd xmmA,xmmE ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36) |
| 373 | punpckhwd xmmC,xmmE ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E) |
| 374 | movdqa xmmG,xmmB |
| 375 | punpcklwd xmmB,xmmF ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37) |
| 376 | punpckhwd xmmG,xmmF ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F) |
| 377 | |
| 378 | movdqa xmmD,xmmA |
| 379 | punpckldq xmmA,xmmB ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33) |
| 380 | punpckhdq xmmD,xmmB ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) |
| 381 | movdqa xmmH,xmmC |
| 382 | punpckldq xmmC,xmmG ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B) |
| 383 | punpckhdq xmmH,xmmG ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) |
| 384 | |
| 385 | cmp ecx, byte SIZEOF_XMMWORD |
| 386 | jb short .column_st32 |
| 387 | |
| 388 | test edi, SIZEOF_XMMWORD-1 |
| 389 | jnz short .out1 |
| 390 | ; --(aligned)------------------- |
| 391 | movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA |
| 392 | movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD |
| 393 | movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC |
| 394 | movntdq XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH |
| 395 | add edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr |
| 396 | jmp short .out0 |
| 397 | .out1: ; --(unaligned)----------------- |
| 398 | pcmpeqb xmmE,xmmE ; xmmE=(all 1's) |
| 399 | maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA |
| 400 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 401 | maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [edi], xmmD |
| 402 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 403 | maskmovdqu xmmC,xmmE ; movntdqu XMMWORD [edi], xmmC |
| 404 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 405 | maskmovdqu xmmH,xmmE ; movntdqu XMMWORD [edi], xmmH |
| 406 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 407 | .out0: |
| 408 | sub ecx, byte SIZEOF_XMMWORD |
| 409 | jz near .endcolumn |
| 410 | |
| 411 | add esi, byte SIZEOF_XMMWORD ; inptr0 |
| 412 | dec al ; Yctr |
| 413 | jnz near .Yloop_2nd |
| 414 | |
| 415 | add ebx, byte SIZEOF_XMMWORD ; inptr1 |
| 416 | add edx, byte SIZEOF_XMMWORD ; inptr2 |
| 417 | jmp near .columnloop |
| 418 | alignx 16,7 |
| 419 | |
| 420 | .column_st32: |
| 421 | pcmpeqb xmmE,xmmE ; xmmE=(all 1's) |
| 422 | cmp ecx, byte SIZEOF_XMMWORD/2 |
| 423 | jb short .column_st16 |
| 424 | maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA |
| 425 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 426 | maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [edi], xmmD |
| 427 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 428 | movdqa xmmA,xmmC |
| 429 | movdqa xmmD,xmmH |
| 430 | sub ecx, byte SIZEOF_XMMWORD/2 |
| 431 | .column_st16: |
| 432 | cmp ecx, byte SIZEOF_XMMWORD/4 |
| 433 | jb short .column_st15 |
| 434 | maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA |
| 435 | add edi, byte SIZEOF_XMMWORD ; outptr |
| 436 | movdqa xmmA,xmmD |
| 437 | sub ecx, byte SIZEOF_XMMWORD/4 |
| 438 | .column_st15: |
| 439 | cmp ecx, byte SIZEOF_XMMWORD/16 |
| 440 | jb short .endcolumn |
| 441 | mov eax,ecx |
| 442 | xor ecx, byte 0x03 |
| 443 | inc ecx |
| 444 | shl ecx, 4 |
| 445 | movd xmmF,ecx |
| 446 | psrlq xmmE,xmmF |
| 447 | punpcklbw xmmE,xmmE |
| 448 | ; ---------------- |
| 449 | mov ecx,edi |
| 450 | and ecx, byte SIZEOF_XMMWORD-1 |
| 451 | jz short .adj0 |
| 452 | lea eax, [ecx+eax*4] ; RGB_PIXELSIZE |
| 453 | cmp eax, byte SIZEOF_XMMWORD |
| 454 | ja short .adj0 |
| 455 | and edi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary |
| 456 | shl ecx, 3 ; pslldq xmmA,ecx & pslldq xmmE,ecx |
| 457 | movdqa xmmB,xmmA |
| 458 | movdqa xmmG,xmmE |
| 459 | pslldq xmmA, SIZEOF_XMMWORD/2 |
| 460 | pslldq xmmE, SIZEOF_XMMWORD/2 |
| 461 | movd xmmC,ecx |
| 462 | sub ecx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT |
| 463 | jb short .adj1 |
| 464 | movd xmmH,ecx |
| 465 | psllq xmmA,xmmH |
| 466 | psllq xmmE,xmmH |
| 467 | jmp short .adj0 |
| 468 | .adj1: neg ecx |
| 469 | movd xmmH,ecx |
| 470 | psrlq xmmA,xmmH |
| 471 | psrlq xmmE,xmmH |
| 472 | psllq xmmB,xmmC |
| 473 | psllq xmmG,xmmC |
| 474 | por xmmA,xmmB |
| 475 | por xmmE,xmmG |
| 476 | .adj0: ; ---------------- |
| 477 | maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [edi], xmmA |
| 478 | |
| 479 | %endif ; RGB_PIXELSIZE ; --------------- |
| 480 | |
| 481 | .endcolumn: |
| 482 | sfence ; flush the write buffer |
| 483 | |
| 484 | .return: |
| 485 | pop edi |
| 486 | pop esi |
| 487 | ; pop edx ; need not be preserved |
| 488 | ; pop ecx ; need not be preserved |
| 489 | pop ebx |
| 490 | mov esp,ebp ; esp <- aligned ebp |
| 491 | pop esp ; esp <- original ebp |
| 492 | pop ebp |
| 493 | ret |
| 494 | |
| 495 | ; -------------------------------------------------------------------------- |
| 496 | ; |
| 497 | ; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical. |
| 498 | ; |
| 499 | ; GLOBAL(void) |
| 500 | ; jsimd_h2v2_merged_upsample_sse2 (JDIMENSION output_width, |
| 501 | ; JSAMPIMAGE input_buf, |
| 502 | ; JDIMENSION in_row_group_ctr, |
| 503 | ; JSAMPARRAY output_buf); |
| 504 | ; |
| 505 | |
| 506 | %define output_width(b) (b)+8 ; JDIMENSION output_width |
| 507 | %define input_buf(b) (b)+12 ; JSAMPIMAGE input_buf |
| 508 | %define in_row_group_ctr(b) (b)+16 ; JDIMENSION in_row_group_ctr |
| 509 | %define output_buf(b) (b)+20 ; JSAMPARRAY output_buf |
| 510 | |
| 511 | align 16 |
| 512 | global EXTN(jsimd_h2v2_merged_upsample_sse2) |
| 513 | |
| 514 | EXTN(jsimd_h2v2_merged_upsample_sse2): |
| 515 | push ebp |
| 516 | mov ebp,esp |
| 517 | push ebx |
| 518 | ; push ecx ; need not be preserved |
| 519 | ; push edx ; need not be preserved |
| 520 | push esi |
| 521 | push edi |
| 522 | |
| 523 | mov eax, POINTER [output_width(ebp)] |
| 524 | |
| 525 | mov edi, JSAMPIMAGE [input_buf(ebp)] |
| 526 | mov ecx, JDIMENSION [in_row_group_ctr(ebp)] |
| 527 | mov esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY] |
| 528 | mov ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY] |
| 529 | mov edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY] |
| 530 | mov edi, JSAMPARRAY [output_buf(ebp)] |
| 531 | lea esi, [esi+ecx*SIZEOF_JSAMPROW] |
| 532 | |
| 533 | push edx ; inptr2 |
| 534 | push ebx ; inptr1 |
| 535 | push esi ; inptr00 |
| 536 | mov ebx,esp |
| 537 | |
| 538 | push edi ; output_buf (outptr0) |
| 539 | push ecx ; in_row_group_ctr |
| 540 | push ebx ; input_buf |
| 541 | push eax ; output_width |
| 542 | |
| 543 | call near EXTN(jsimd_h2v1_merged_upsample_sse2) |
| 544 | |
| 545 | add esi, byte SIZEOF_JSAMPROW ; inptr01 |
| 546 | add edi, byte SIZEOF_JSAMPROW ; outptr1 |
| 547 | mov POINTER [ebx+0*SIZEOF_POINTER], esi |
| 548 | mov POINTER [ebx-1*SIZEOF_POINTER], edi |
| 549 | |
| 550 | call near EXTN(jsimd_h2v1_merged_upsample_sse2) |
| 551 | |
| 552 | add esp, byte 7*SIZEOF_DWORD |
| 553 | |
| 554 | pop edi |
| 555 | pop esi |
| 556 | ; pop edx ; need not be preserved |
| 557 | ; pop ecx ; need not be preserved |
| 558 | pop ebx |
| 559 | pop ebp |
| 560 | ret |
| 561 | |
DRC | 132b5fd | 2009-10-08 09:04:56 +0000 | [diff] [blame] | 562 | ; For some reason, the OS X linker does not honor the request to align the |
| 563 | ; segment unless we do this. |
| 564 | align 16 |