DRC | 246c3d9 | 2009-06-25 20:38:31 +0000 | [diff] [blame] | 1 | ; |
| 2 | ; jcclrss2.asm - colorspace conversion (64-bit SSE2) |
| 3 | ; |
| 4 | ; x86 SIMD extension for IJG JPEG library |
| 5 | ; Copyright (C) 1999-2006, MIYASAKA Masaru. |
| 6 | ; Copyright (C) 2009, D. R. Commander. |
| 7 | ; For conditions of distribution and use, see copyright notice in jsimdext.inc |
| 8 | ; |
| 9 | ; This file should be assembled with NASM (Netwide Assembler), |
| 10 | ; can *not* be assembled with Microsoft's MASM or any compatible |
| 11 | ; assembler (including Borland's Turbo Assembler). |
| 12 | ; NASM is available from http://nasm.sourceforge.net/ or |
| 13 | ; http://sourceforge.net/project/showfiles.php?group_id=6208 |
| 14 | ; |
| 15 | ; [TAB8] |
| 16 | |
| 17 | %include "jcolsamp.inc" |
| 18 | |
| 19 | ; -------------------------------------------------------------------------- |
| 20 | SECTION SEG_TEXT |
| 21 | BITS 64 |
| 22 | ; |
| 23 | ; Convert some rows of samples to the output colorspace. |
| 24 | ; |
| 25 | ; GLOBAL(void) |
| 26 | ; jsimd_rgb_ycc_convert_sse2 (JDIMENSION img_width, |
| 27 | ; JSAMPARRAY input_buf, JSAMPIMAGE output_buf, |
| 28 | ; JDIMENSION output_row, int num_rows); |
| 29 | ; |
| 30 | |
| 31 | ; r10 = JDIMENSION img_width |
| 32 | ; r11 = JSAMPARRAY input_buf |
| 33 | ; r12 = JSAMPIMAGE output_buf |
| 34 | ; r13 = JDIMENSION output_row |
| 35 | ; r14 = int num_rows |
| 36 | |
| 37 | %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM] |
| 38 | %define WK_NUM 8 |
| 39 | |
| 40 | align 16 |
| 41 | |
| 42 | global EXTN(jsimd_rgb_ycc_convert_sse2) |
| 43 | |
| 44 | EXTN(jsimd_rgb_ycc_convert_sse2): |
| 45 | push rbp |
| 46 | mov rax,rsp ; rax = original rbp |
| 47 | sub rsp, byte 4 |
| 48 | and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits |
| 49 | mov [rsp],rax |
| 50 | mov rbp,rsp ; rbp = aligned rbp |
| 51 | lea rsp, [wk(0)] |
| 52 | push rbx |
| 53 | collect_args |
| 54 | |
| 55 | mov rcx, r10 |
| 56 | test rcx,rcx |
| 57 | jz near .return |
| 58 | |
| 59 | push rcx |
| 60 | |
| 61 | mov rsi, r12 |
| 62 | mov rcx, r13 |
| 63 | mov rdi, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY] |
| 64 | mov rbx, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY] |
| 65 | mov rdx, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY] |
| 66 | lea rdi, [rdi+rcx*SIZEOF_JSAMPROW] |
| 67 | lea rbx, [rbx+rcx*SIZEOF_JSAMPROW] |
| 68 | lea rdx, [rdx+rcx*SIZEOF_JSAMPROW] |
| 69 | |
| 70 | pop rcx |
| 71 | |
| 72 | mov rsi, r11 |
| 73 | mov rax, r14 |
| 74 | test rax,rax |
| 75 | jle near .return |
| 76 | .rowloop: |
| 77 | push rdx |
| 78 | push rbx |
| 79 | push rdi |
| 80 | push rsi |
| 81 | push rcx ; col |
| 82 | |
| 83 | mov rsi, JSAMPROW [rsi] ; inptr |
| 84 | mov rdi, JSAMPROW [rdi] ; outptr0 |
| 85 | mov rbx, JSAMPROW [rbx] ; outptr1 |
| 86 | mov rdx, JSAMPROW [rdx] ; outptr2 |
| 87 | |
| 88 | cmp rcx, byte SIZEOF_XMMWORD |
| 89 | jae near .columnloop |
| 90 | |
| 91 | %if RGB_PIXELSIZE == 3 ; --------------- |
| 92 | |
| 93 | .column_ld1: |
| 94 | push rax |
| 95 | push rdx |
| 96 | lea rcx,[rcx+rcx*2] ; imul ecx,RGB_PIXELSIZE |
| 97 | test cl, SIZEOF_BYTE |
| 98 | jz short .column_ld2 |
| 99 | sub rcx, byte SIZEOF_BYTE |
| 100 | movzx rax, BYTE [rsi+rcx] |
| 101 | .column_ld2: |
| 102 | test cl, SIZEOF_WORD |
| 103 | jz short .column_ld4 |
| 104 | sub rcx, byte SIZEOF_WORD |
| 105 | movzx rdx, WORD [rsi+rcx] |
| 106 | shl rax, WORD_BIT |
| 107 | or rax,rdx |
| 108 | .column_ld4: |
| 109 | movd xmmA,eax |
| 110 | pop rdx |
| 111 | pop rax |
| 112 | test cl, SIZEOF_DWORD |
| 113 | jz short .column_ld8 |
| 114 | sub rcx, byte SIZEOF_DWORD |
| 115 | movd xmmF, XMM_DWORD [rsi+rcx] |
| 116 | pslldq xmmA, SIZEOF_DWORD |
| 117 | por xmmA,xmmF |
| 118 | .column_ld8: |
| 119 | test cl, SIZEOF_MMWORD |
| 120 | jz short .column_ld16 |
| 121 | sub rcx, byte SIZEOF_MMWORD |
| 122 | movq xmmB, XMM_MMWORD [rsi+rcx] |
| 123 | pslldq xmmA, SIZEOF_MMWORD |
| 124 | por xmmA,xmmB |
| 125 | .column_ld16: |
| 126 | test cl, SIZEOF_XMMWORD |
| 127 | jz short .column_ld32 |
| 128 | movdqa xmmF,xmmA |
| 129 | movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 130 | mov rcx, SIZEOF_XMMWORD |
| 131 | jmp short .rgb_ycc_cnv |
| 132 | .column_ld32: |
| 133 | test cl, 2*SIZEOF_XMMWORD |
| 134 | mov rcx, SIZEOF_XMMWORD |
| 135 | jz short .rgb_ycc_cnv |
| 136 | movdqa xmmB,xmmA |
| 137 | movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 138 | movdqu xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 139 | jmp short .rgb_ycc_cnv |
| 140 | |
| 141 | .columnloop: |
| 142 | movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 143 | movdqu xmmF, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 144 | movdqu xmmB, XMMWORD [rsi+2*SIZEOF_XMMWORD] |
| 145 | |
| 146 | .rgb_ycc_cnv: |
| 147 | ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05) |
| 148 | ; xmmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A) |
| 149 | ; xmmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F) |
| 150 | |
| 151 | movdqa xmmG,xmmA |
| 152 | pslldq xmmA,8 ; xmmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12) |
| 153 | psrldq xmmG,8 ; xmmG=(22 03 13 23 04 14 24 05 -- -- -- -- -- -- -- --) |
| 154 | |
| 155 | punpckhbw xmmA,xmmF ; xmmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A) |
| 156 | pslldq xmmF,8 ; xmmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27) |
| 157 | |
| 158 | punpcklbw xmmG,xmmB ; xmmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D) |
| 159 | punpckhbw xmmF,xmmB ; xmmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F) |
| 160 | |
| 161 | movdqa xmmD,xmmA |
| 162 | pslldq xmmA,8 ; xmmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09) |
| 163 | psrldq xmmD,8 ; xmmD=(11 19 21 29 02 0A 12 1A -- -- -- -- -- -- -- --) |
| 164 | |
| 165 | punpckhbw xmmA,xmmG ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D) |
| 166 | pslldq xmmG,8 ; xmmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B) |
| 167 | |
| 168 | punpcklbw xmmD,xmmF ; xmmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E) |
| 169 | punpckhbw xmmG,xmmF ; xmmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F) |
| 170 | |
| 171 | movdqa xmmE,xmmA |
| 172 | pslldq xmmA,8 ; xmmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C) |
| 173 | psrldq xmmE,8 ; xmmE=(20 24 28 2C 01 05 09 0D -- -- -- -- -- -- -- --) |
| 174 | |
| 175 | punpckhbw xmmA,xmmD ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E) |
| 176 | pslldq xmmD,8 ; xmmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D) |
| 177 | |
| 178 | punpcklbw xmmE,xmmG ; xmmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F) |
| 179 | punpckhbw xmmD,xmmG ; xmmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F) |
| 180 | |
| 181 | pxor xmmH,xmmH |
| 182 | |
| 183 | movdqa xmmC,xmmA |
| 184 | punpcklbw xmmA,xmmH ; xmmA=(00 02 04 06 08 0A 0C 0E) |
| 185 | punpckhbw xmmC,xmmH ; xmmC=(10 12 14 16 18 1A 1C 1E) |
| 186 | |
| 187 | movdqa xmmB,xmmE |
| 188 | punpcklbw xmmE,xmmH ; xmmE=(20 22 24 26 28 2A 2C 2E) |
| 189 | punpckhbw xmmB,xmmH ; xmmB=(01 03 05 07 09 0B 0D 0F) |
| 190 | |
| 191 | movdqa xmmF,xmmD |
| 192 | punpcklbw xmmD,xmmH ; xmmD=(11 13 15 17 19 1B 1D 1F) |
| 193 | punpckhbw xmmF,xmmH ; xmmF=(21 23 25 27 29 2B 2D 2F) |
| 194 | |
| 195 | %else ; RGB_PIXELSIZE == 4 ; ----------- |
| 196 | |
| 197 | .column_ld1: |
| 198 | test cl, SIZEOF_XMMWORD/16 |
| 199 | jz short .column_ld2 |
| 200 | sub rcx, byte SIZEOF_XMMWORD/16 |
| 201 | movd xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE] |
| 202 | .column_ld2: |
| 203 | test cl, SIZEOF_XMMWORD/8 |
| 204 | jz short .column_ld4 |
| 205 | sub rcx, byte SIZEOF_XMMWORD/8 |
| 206 | movq xmmE, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE] |
| 207 | pslldq xmmA, SIZEOF_MMWORD |
| 208 | por xmmA,xmmE |
| 209 | .column_ld4: |
| 210 | test cl, SIZEOF_XMMWORD/4 |
| 211 | jz short .column_ld8 |
| 212 | sub rcx, byte SIZEOF_XMMWORD/4 |
| 213 | movdqa xmmE,xmmA |
| 214 | movdqu xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE] |
| 215 | .column_ld8: |
| 216 | test cl, SIZEOF_XMMWORD/2 |
| 217 | mov rcx, SIZEOF_XMMWORD |
| 218 | jz short .rgb_ycc_cnv |
| 219 | movdqa xmmF,xmmA |
| 220 | movdqa xmmH,xmmE |
| 221 | movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 222 | movdqu xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 223 | jmp short .rgb_ycc_cnv |
| 224 | |
| 225 | .columnloop: |
| 226 | movdqu xmmA, XMMWORD [rsi+0*SIZEOF_XMMWORD] |
| 227 | movdqu xmmE, XMMWORD [rsi+1*SIZEOF_XMMWORD] |
| 228 | movdqu xmmF, XMMWORD [rsi+2*SIZEOF_XMMWORD] |
| 229 | movdqu xmmH, XMMWORD [rsi+3*SIZEOF_XMMWORD] |
| 230 | |
| 231 | .rgb_ycc_cnv: |
| 232 | ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33) |
| 233 | ; xmmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37) |
| 234 | ; xmmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B) |
| 235 | ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F) |
| 236 | |
| 237 | movdqa xmmD,xmmA |
| 238 | punpcklbw xmmA,xmmE ; xmmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35) |
| 239 | punpckhbw xmmD,xmmE ; xmmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37) |
| 240 | |
| 241 | movdqa xmmC,xmmF |
| 242 | punpcklbw xmmF,xmmH ; xmmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D) |
| 243 | punpckhbw xmmC,xmmH ; xmmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F) |
| 244 | |
| 245 | movdqa xmmB,xmmA |
| 246 | punpcklwd xmmA,xmmF ; xmmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C) |
| 247 | punpckhwd xmmB,xmmF ; xmmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D) |
| 248 | |
| 249 | movdqa xmmG,xmmD |
| 250 | punpcklwd xmmD,xmmC ; xmmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E) |
| 251 | punpckhwd xmmG,xmmC ; xmmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F) |
| 252 | |
| 253 | movdqa xmmE,xmmA |
| 254 | punpcklbw xmmA,xmmD ; xmmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E) |
| 255 | punpckhbw xmmE,xmmD ; xmmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E) |
| 256 | |
| 257 | movdqa xmmH,xmmB |
| 258 | punpcklbw xmmB,xmmG ; xmmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F) |
| 259 | punpckhbw xmmH,xmmG ; xmmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F) |
| 260 | |
| 261 | pxor xmmF,xmmF |
| 262 | |
| 263 | movdqa xmmC,xmmA |
| 264 | punpcklbw xmmA,xmmF ; xmmA=(00 02 04 06 08 0A 0C 0E) |
| 265 | punpckhbw xmmC,xmmF ; xmmC=(10 12 14 16 18 1A 1C 1E) |
| 266 | |
| 267 | movdqa xmmD,xmmB |
| 268 | punpcklbw xmmB,xmmF ; xmmB=(01 03 05 07 09 0B 0D 0F) |
| 269 | punpckhbw xmmD,xmmF ; xmmD=(11 13 15 17 19 1B 1D 1F) |
| 270 | |
| 271 | movdqa xmmG,xmmE |
| 272 | punpcklbw xmmE,xmmF ; xmmE=(20 22 24 26 28 2A 2C 2E) |
| 273 | punpckhbw xmmG,xmmF ; xmmG=(30 32 34 36 38 3A 3C 3E) |
| 274 | |
| 275 | punpcklbw xmmF,xmmH |
| 276 | punpckhbw xmmH,xmmH |
| 277 | psrlw xmmF,BYTE_BIT ; xmmF=(21 23 25 27 29 2B 2D 2F) |
| 278 | psrlw xmmH,BYTE_BIT ; xmmH=(31 33 35 37 39 3B 3D 3F) |
| 279 | |
| 280 | %endif ; RGB_PIXELSIZE ; --------------- |
| 281 | |
| 282 | ; xmm0=R(02468ACE)=RE, xmm2=G(02468ACE)=GE, xmm4=B(02468ACE)=BE |
| 283 | ; xmm1=R(13579BDF)=RO, xmm3=G(13579BDF)=GO, xmm5=B(13579BDF)=BO |
| 284 | |
| 285 | ; (Original) |
| 286 | ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B |
| 287 | ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE |
| 288 | ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE |
| 289 | ; |
| 290 | ; (This implementation) |
| 291 | ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G |
| 292 | ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE |
| 293 | ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE |
| 294 | |
| 295 | movdqa XMMWORD [wk(0)], xmm0 ; wk(0)=RE |
| 296 | movdqa XMMWORD [wk(1)], xmm1 ; wk(1)=RO |
| 297 | movdqa XMMWORD [wk(2)], xmm4 ; wk(2)=BE |
| 298 | movdqa XMMWORD [wk(3)], xmm5 ; wk(3)=BO |
| 299 | |
| 300 | movdqa xmm6,xmm1 |
| 301 | punpcklwd xmm1,xmm3 |
| 302 | punpckhwd xmm6,xmm3 |
| 303 | movdqa xmm7,xmm1 |
| 304 | movdqa xmm4,xmm6 |
| 305 | pmaddwd xmm1,[PW_F0299_F0337] ; xmm1=ROL*FIX(0.299)+GOL*FIX(0.337) |
| 306 | pmaddwd xmm6,[PW_F0299_F0337] ; xmm6=ROH*FIX(0.299)+GOH*FIX(0.337) |
| 307 | pmaddwd xmm7,[PW_MF016_MF033] ; xmm7=ROL*-FIX(0.168)+GOL*-FIX(0.331) |
| 308 | pmaddwd xmm4,[PW_MF016_MF033] ; xmm4=ROH*-FIX(0.168)+GOH*-FIX(0.331) |
| 309 | |
| 310 | movdqa XMMWORD [wk(4)], xmm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337) |
| 311 | movdqa XMMWORD [wk(5)], xmm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337) |
| 312 | |
| 313 | pxor xmm1,xmm1 |
| 314 | pxor xmm6,xmm6 |
| 315 | punpcklwd xmm1,xmm5 ; xmm1=BOL |
| 316 | punpckhwd xmm6,xmm5 ; xmm6=BOH |
| 317 | psrld xmm1,1 ; xmm1=BOL*FIX(0.500) |
| 318 | psrld xmm6,1 ; xmm6=BOH*FIX(0.500) |
| 319 | |
| 320 | movdqa xmm5,[PD_ONEHALFM1_CJ] ; xmm5=[PD_ONEHALFM1_CJ] |
| 321 | |
| 322 | paddd xmm7,xmm1 |
| 323 | paddd xmm4,xmm6 |
| 324 | paddd xmm7,xmm5 |
| 325 | paddd xmm4,xmm5 |
| 326 | psrld xmm7,SCALEBITS ; xmm7=CbOL |
| 327 | psrld xmm4,SCALEBITS ; xmm4=CbOH |
| 328 | packssdw xmm7,xmm4 ; xmm7=CbO |
| 329 | |
| 330 | movdqa xmm1, XMMWORD [wk(2)] ; xmm1=BE |
| 331 | |
| 332 | movdqa xmm6,xmm0 |
| 333 | punpcklwd xmm0,xmm2 |
| 334 | punpckhwd xmm6,xmm2 |
| 335 | movdqa xmm5,xmm0 |
| 336 | movdqa xmm4,xmm6 |
| 337 | pmaddwd xmm0,[PW_F0299_F0337] ; xmm0=REL*FIX(0.299)+GEL*FIX(0.337) |
| 338 | pmaddwd xmm6,[PW_F0299_F0337] ; xmm6=REH*FIX(0.299)+GEH*FIX(0.337) |
| 339 | pmaddwd xmm5,[PW_MF016_MF033] ; xmm5=REL*-FIX(0.168)+GEL*-FIX(0.331) |
| 340 | pmaddwd xmm4,[PW_MF016_MF033] ; xmm4=REH*-FIX(0.168)+GEH*-FIX(0.331) |
| 341 | |
| 342 | movdqa XMMWORD [wk(6)], xmm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337) |
| 343 | movdqa XMMWORD [wk(7)], xmm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337) |
| 344 | |
| 345 | pxor xmm0,xmm0 |
| 346 | pxor xmm6,xmm6 |
| 347 | punpcklwd xmm0,xmm1 ; xmm0=BEL |
| 348 | punpckhwd xmm6,xmm1 ; xmm6=BEH |
| 349 | psrld xmm0,1 ; xmm0=BEL*FIX(0.500) |
| 350 | psrld xmm6,1 ; xmm6=BEH*FIX(0.500) |
| 351 | |
| 352 | movdqa xmm1,[PD_ONEHALFM1_CJ] ; xmm1=[PD_ONEHALFM1_CJ] |
| 353 | |
| 354 | paddd xmm5,xmm0 |
| 355 | paddd xmm4,xmm6 |
| 356 | paddd xmm5,xmm1 |
| 357 | paddd xmm4,xmm1 |
| 358 | psrld xmm5,SCALEBITS ; xmm5=CbEL |
| 359 | psrld xmm4,SCALEBITS ; xmm4=CbEH |
| 360 | packssdw xmm5,xmm4 ; xmm5=CbE |
| 361 | |
| 362 | psllw xmm7,BYTE_BIT |
| 363 | por xmm5,xmm7 ; xmm5=Cb |
| 364 | movdqa XMMWORD [ebx], xmm5 ; Save Cb |
| 365 | |
| 366 | movdqa xmm0, XMMWORD [wk(3)] ; xmm0=BO |
| 367 | movdqa xmm6, XMMWORD [wk(2)] ; xmm6=BE |
| 368 | movdqa xmm1, XMMWORD [wk(1)] ; xmm1=RO |
| 369 | |
| 370 | movdqa xmm4,xmm0 |
| 371 | punpcklwd xmm0,xmm3 |
| 372 | punpckhwd xmm4,xmm3 |
| 373 | movdqa xmm7,xmm0 |
| 374 | movdqa xmm5,xmm4 |
| 375 | pmaddwd xmm0,[PW_F0114_F0250] ; xmm0=BOL*FIX(0.114)+GOL*FIX(0.250) |
| 376 | pmaddwd xmm4,[PW_F0114_F0250] ; xmm4=BOH*FIX(0.114)+GOH*FIX(0.250) |
| 377 | pmaddwd xmm7,[PW_MF008_MF041] ; xmm7=BOL*-FIX(0.081)+GOL*-FIX(0.418) |
| 378 | pmaddwd xmm5,[PW_MF008_MF041] ; xmm5=BOH*-FIX(0.081)+GOH*-FIX(0.418) |
| 379 | |
| 380 | movdqa xmm3,[PD_ONEHALF] ; xmm3=[PD_ONEHALF] |
| 381 | |
| 382 | paddd xmm0, XMMWORD [wk(4)] |
| 383 | paddd xmm4, XMMWORD [wk(5)] |
| 384 | paddd xmm0,xmm3 |
| 385 | paddd xmm4,xmm3 |
| 386 | psrld xmm0,SCALEBITS ; xmm0=YOL |
| 387 | psrld xmm4,SCALEBITS ; xmm4=YOH |
| 388 | packssdw xmm0,xmm4 ; xmm0=YO |
| 389 | |
| 390 | pxor xmm3,xmm3 |
| 391 | pxor xmm4,xmm4 |
| 392 | punpcklwd xmm3,xmm1 ; xmm3=ROL |
| 393 | punpckhwd xmm4,xmm1 ; xmm4=ROH |
| 394 | psrld xmm3,1 ; xmm3=ROL*FIX(0.500) |
| 395 | psrld xmm4,1 ; xmm4=ROH*FIX(0.500) |
| 396 | |
| 397 | movdqa xmm1,[PD_ONEHALFM1_CJ] ; xmm1=[PD_ONEHALFM1_CJ] |
| 398 | |
| 399 | paddd xmm7,xmm3 |
| 400 | paddd xmm5,xmm4 |
| 401 | paddd xmm7,xmm1 |
| 402 | paddd xmm5,xmm1 |
| 403 | psrld xmm7,SCALEBITS ; xmm7=CrOL |
| 404 | psrld xmm5,SCALEBITS ; xmm5=CrOH |
| 405 | packssdw xmm7,xmm5 ; xmm7=CrO |
| 406 | |
| 407 | movdqa xmm3, XMMWORD [wk(0)] ; xmm3=RE |
| 408 | |
| 409 | movdqa xmm4,xmm6 |
| 410 | punpcklwd xmm6,xmm2 |
| 411 | punpckhwd xmm4,xmm2 |
| 412 | movdqa xmm1,xmm6 |
| 413 | movdqa xmm5,xmm4 |
| 414 | pmaddwd xmm6,[PW_F0114_F0250] ; xmm6=BEL*FIX(0.114)+GEL*FIX(0.250) |
| 415 | pmaddwd xmm4,[PW_F0114_F0250] ; xmm4=BEH*FIX(0.114)+GEH*FIX(0.250) |
| 416 | pmaddwd xmm1,[PW_MF008_MF041] ; xmm1=BEL*-FIX(0.081)+GEL*-FIX(0.418) |
| 417 | pmaddwd xmm5,[PW_MF008_MF041] ; xmm5=BEH*-FIX(0.081)+GEH*-FIX(0.418) |
| 418 | |
| 419 | movdqa xmm2,[PD_ONEHALF] ; xmm2=[PD_ONEHALF] |
| 420 | |
| 421 | paddd xmm6, XMMWORD [wk(6)] |
| 422 | paddd xmm4, XMMWORD [wk(7)] |
| 423 | paddd xmm6,xmm2 |
| 424 | paddd xmm4,xmm2 |
| 425 | psrld xmm6,SCALEBITS ; xmm6=YEL |
| 426 | psrld xmm4,SCALEBITS ; xmm4=YEH |
| 427 | packssdw xmm6,xmm4 ; xmm6=YE |
| 428 | |
| 429 | psllw xmm0,BYTE_BIT |
| 430 | por xmm6,xmm0 ; xmm6=Y |
| 431 | movdqa XMMWORD [rdi], xmm6 ; Save Y |
| 432 | |
| 433 | pxor xmm2,xmm2 |
| 434 | pxor xmm4,xmm4 |
| 435 | punpcklwd xmm2,xmm3 ; xmm2=REL |
| 436 | punpckhwd xmm4,xmm3 ; xmm4=REH |
| 437 | psrld xmm2,1 ; xmm2=REL*FIX(0.500) |
| 438 | psrld xmm4,1 ; xmm4=REH*FIX(0.500) |
| 439 | |
| 440 | movdqa xmm0,[PD_ONEHALFM1_CJ] ; xmm0=[PD_ONEHALFM1_CJ] |
| 441 | |
| 442 | paddd xmm1,xmm2 |
| 443 | paddd xmm5,xmm4 |
| 444 | paddd xmm1,xmm0 |
| 445 | paddd xmm5,xmm0 |
| 446 | psrld xmm1,SCALEBITS ; xmm1=CrEL |
| 447 | psrld xmm5,SCALEBITS ; xmm5=CrEH |
| 448 | packssdw xmm1,xmm5 ; xmm1=CrE |
| 449 | |
| 450 | psllw xmm7,BYTE_BIT |
| 451 | por xmm1,xmm7 ; xmm1=Cr |
| 452 | movdqa XMMWORD [rdx], xmm1 ; Save Cr |
| 453 | |
| 454 | sub rcx, byte SIZEOF_XMMWORD |
| 455 | add rsi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; inptr |
| 456 | add rdi, byte SIZEOF_XMMWORD ; outptr0 |
| 457 | add rbx, byte SIZEOF_XMMWORD ; outptr1 |
| 458 | add rdx, byte SIZEOF_XMMWORD ; outptr2 |
| 459 | cmp rcx, byte SIZEOF_XMMWORD |
| 460 | jae near .columnloop |
| 461 | test rcx,rcx |
| 462 | jnz near .column_ld1 |
| 463 | |
| 464 | pop rcx ; col |
| 465 | pop rsi |
| 466 | pop rdi |
| 467 | pop rbx |
| 468 | pop rdx |
| 469 | |
| 470 | add rsi, byte SIZEOF_JSAMPROW ; input_buf |
| 471 | add rdi, byte SIZEOF_JSAMPROW |
| 472 | add rbx, byte SIZEOF_JSAMPROW |
| 473 | add rdx, byte SIZEOF_JSAMPROW |
| 474 | dec rax ; num_rows |
| 475 | jg near .rowloop |
| 476 | |
| 477 | .return: |
| 478 | uncollect_args |
| 479 | pop rbx |
| 480 | mov rsp,rbp ; rsp <- aligned rbp |
| 481 | pop rsp ; rsp <- original rbp |
| 482 | pop rbp |
| 483 | ret |
| 484 | |