2 ; jdclrss2-64.asm - colorspace conversion (64-bit SSE2)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright 2009 D. R. Commander
8 ; 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
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
20 %include "jcolsamp.inc"
22 ; --------------------------------------------------------------------------
24 ; Convert some rows of samples to the output colorspace.
27 ; jsimd_ycc_rgb_convert_sse2 (JDIMENSION out_width,
28 ; JSAMPIMAGE input_buf, JDIMENSION input_row,
29 ; JSAMPARRAY output_buf, int num_rows)
32 ; r10 = JDIMENSION out_width
33 ; r11 = JSAMPIMAGE input_buf
34 ; r12 = JDIMENSION input_row
35 ; r13 = JSAMPARRAY output_buf
38 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
42 global EXTN(jsimd_ycc_rgb_convert_sse2)
44 EXTN(jsimd_ycc_rgb_convert_sse2):
46 mov rax,rsp ; rax = original rbp
48 and rsp, byte (-SIZEOF_XMMWORD) ; align to 128 bits
50 mov rbp,rsp ; rbp = aligned rbp
55 mov rcx, r10 ; num_cols
63 mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
64 mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
65 mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
66 lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
67 lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
68 lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
84 mov rsi, JSAMPROW [rsi] ; inptr0
85 mov rbx, JSAMPROW [rbx] ; inptr1
86 mov rdx, JSAMPROW [rdx] ; inptr2
87 mov rdi, JSAMPROW [rdi] ; outptr
90 movdqa xmm5, XMMWORD [rbx] ; xmm5=Cb(0123456789ABCDEF)
91 movdqa xmm1, XMMWORD [rdx] ; xmm1=Cr(0123456789ABCDEF)
96 psllw xmm7,7 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
97 movdqa xmm0,xmm4 ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
99 pand xmm4,xmm5 ; xmm4=Cb(02468ACE)=CbE
100 psrlw xmm5,BYTE_BIT ; xmm5=Cb(13579BDF)=CbO
101 pand xmm0,xmm1 ; xmm0=Cr(02468ACE)=CrE
102 psrlw xmm1,BYTE_BIT ; xmm1=Cr(13579BDF)=CrO
110 ; R = Y + 1.40200 * Cr
111 ; G = Y - 0.34414 * Cb - 0.71414 * Cr
112 ; B = Y + 1.77200 * Cb
114 ; (This implementation)
115 ; R = Y + 0.40200 * Cr + Cr
116 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
117 ; B = Y - 0.22800 * Cb + Cb + Cb
119 movdqa xmm2,xmm4 ; xmm2=CbE
120 movdqa xmm3,xmm5 ; xmm3=CbO
121 paddw xmm4,xmm4 ; xmm4=2*CbE
122 paddw xmm5,xmm5 ; xmm5=2*CbO
123 movdqa xmm6,xmm0 ; xmm6=CrE
124 movdqa xmm7,xmm1 ; xmm7=CrO
125 paddw xmm0,xmm0 ; xmm0=2*CrE
126 paddw xmm1,xmm1 ; xmm1=2*CrO
128 pmulhw xmm4,[rel PW_MF0228] ; xmm4=(2*CbE * -FIX(0.22800))
129 pmulhw xmm5,[rel PW_MF0228] ; xmm5=(2*CbO * -FIX(0.22800))
130 pmulhw xmm0,[rel PW_F0402] ; xmm0=(2*CrE * FIX(0.40200))
131 pmulhw xmm1,[rel PW_F0402] ; xmm1=(2*CrO * FIX(0.40200))
133 paddw xmm4,[rel PW_ONE]
134 paddw xmm5,[rel PW_ONE]
135 psraw xmm4,1 ; xmm4=(CbE * -FIX(0.22800))
136 psraw xmm5,1 ; xmm5=(CbO * -FIX(0.22800))
137 paddw xmm0,[rel PW_ONE]
138 paddw xmm1,[rel PW_ONE]
139 psraw xmm0,1 ; xmm0=(CrE * FIX(0.40200))
140 psraw xmm1,1 ; xmm1=(CrO * FIX(0.40200))
144 paddw xmm4,xmm2 ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
145 paddw xmm5,xmm3 ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
146 paddw xmm0,xmm6 ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
147 paddw xmm1,xmm7 ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
149 movdqa XMMWORD [wk(0)], xmm4 ; wk(0)=(B-Y)E
150 movdqa XMMWORD [wk(1)], xmm5 ; wk(1)=(B-Y)O
156 pmaddwd xmm2,[rel PW_MF0344_F0285]
157 pmaddwd xmm4,[rel PW_MF0344_F0285]
160 pmaddwd xmm3,[rel PW_MF0344_F0285]
161 pmaddwd xmm5,[rel PW_MF0344_F0285]
163 paddd xmm2,[rel PD_ONEHALF]
164 paddd xmm4,[rel PD_ONEHALF]
167 paddd xmm3,[rel PD_ONEHALF]
168 paddd xmm5,[rel PD_ONEHALF]
172 packssdw xmm2,xmm4 ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
173 packssdw xmm3,xmm5 ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
174 psubw xmm2,xmm6 ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
175 psubw xmm3,xmm7 ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
177 movdqa xmm5, XMMWORD [rsi] ; xmm5=Y(0123456789ABCDEF)
180 psrlw xmm4,BYTE_BIT ; xmm4={0xFF 0x00 0xFF 0x00 ..}
181 pand xmm4,xmm5 ; xmm4=Y(02468ACE)=YE
182 psrlw xmm5,BYTE_BIT ; xmm5=Y(13579BDF)=YO
184 paddw xmm0,xmm4 ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
185 paddw xmm1,xmm5 ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
186 packuswb xmm0,xmm0 ; xmm0=R(02468ACE********)
187 packuswb xmm1,xmm1 ; xmm1=R(13579BDF********)
189 paddw xmm2,xmm4 ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
190 paddw xmm3,xmm5 ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
191 packuswb xmm2,xmm2 ; xmm2=G(02468ACE********)
192 packuswb xmm3,xmm3 ; xmm3=G(13579BDF********)
194 paddw xmm4, XMMWORD [wk(0)] ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
195 paddw xmm5, XMMWORD [wk(1)] ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
196 packuswb xmm4,xmm4 ; xmm4=B(02468ACE********)
197 packuswb xmm5,xmm5 ; xmm5=B(13579BDF********)
199 %if RGB_PIXELSIZE == 3 ; ---------------
201 ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
202 ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
203 ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
204 ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
206 punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
207 punpcklbw xmmE,xmmB ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
208 punpcklbw xmmD,xmmF ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
212 punpcklwd xmmA,xmmE ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
213 punpckhwd xmmG,xmmE ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
215 psrldq xmmH,2 ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
216 psrldq xmmE,2 ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
220 punpcklwd xmmD,xmmH ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
221 punpckhwd xmmC,xmmH ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
223 psrldq xmmB,2 ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
226 punpcklwd xmmE,xmmB ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
227 punpckhwd xmmF,xmmB ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
229 pshufd xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
231 punpckldq xmmA,xmmD ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
232 punpckldq xmmE,xmmH ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
233 punpckhdq xmmD,xmmB ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
235 pshufd xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
237 punpckldq xmmG,xmmC ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
238 punpckldq xmmF,xmmH ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
239 punpckhdq xmmC,xmmB ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
241 punpcklqdq xmmA,xmmE ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
242 punpcklqdq xmmD,xmmG ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
243 punpcklqdq xmmF,xmmC ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
245 cmp rcx, byte SIZEOF_XMMWORD
246 jb short .column_st32
248 test rdi, SIZEOF_XMMWORD-1
250 ; --(aligned)-------------------
251 movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
252 movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
253 movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
254 add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
256 .out1: ; --(unaligned)-----------------
257 pcmpeqb xmmH,xmmH ; xmmH=(all 1's)
258 maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [rdi], xmmA
259 add rdi, byte SIZEOF_XMMWORD ; outptr
260 maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [rdi], xmmD
261 add rdi, byte SIZEOF_XMMWORD ; outptr
262 maskmovdqu xmmF,xmmH ; movntdqu XMMWORD [rdi], xmmF
263 add rdi, byte SIZEOF_XMMWORD ; outptr
265 sub rcx, byte SIZEOF_XMMWORD
268 add rsi, byte SIZEOF_XMMWORD ; inptr0
269 add rbx, byte SIZEOF_XMMWORD ; inptr1
270 add rdx, byte SIZEOF_XMMWORD ; inptr2
274 pcmpeqb xmmH,xmmH ; xmmH=(all 1's)
275 lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE
276 cmp rcx, byte 2*SIZEOF_XMMWORD
277 jb short .column_st16
278 maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [rdi], xmmA
279 add rdi, byte SIZEOF_XMMWORD ; outptr
280 maskmovdqu xmmD,xmmH ; movntdqu XMMWORD [rdi], xmmD
281 add rdi, byte SIZEOF_XMMWORD ; outptr
283 sub rcx, byte 2*SIZEOF_XMMWORD
284 jmp short .column_st15
286 cmp rcx, byte SIZEOF_XMMWORD
287 jb short .column_st15
288 maskmovdqu xmmA,xmmH ; movntdqu XMMWORD [rdi], xmmA
289 add rdi, byte SIZEOF_XMMWORD ; outptr
291 sub rcx, byte SIZEOF_XMMWORD
304 and rcx, byte SIZEOF_XMMWORD-1
307 cmp rax, byte SIZEOF_XMMWORD
309 and rdi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary
310 shl rcx, 3 ; pslldq xmmA,ecx & pslldq xmmE,rcx
313 pslldq xmmA, SIZEOF_XMMWORD/2
314 pslldq xmmE, SIZEOF_XMMWORD/2
316 sub rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
330 .adj0: ; ----------------
331 maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [rdi], xmmA
333 %else ; RGB_PIXELSIZE == 4 ; -----------
335 %ifdef RGBX_FILLER_0XFF
336 pcmpeqb xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
337 pcmpeqb xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
339 pxor xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
340 pxor xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
342 ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
343 ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
344 ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
345 ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
347 punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
348 punpcklbw xmmE,xmmG ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
349 punpcklbw xmmB,xmmD ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
350 punpcklbw xmmF,xmmH ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
353 punpcklwd xmmA,xmmE ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
354 punpckhwd xmmC,xmmE ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
356 punpcklwd xmmB,xmmF ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
357 punpckhwd xmmG,xmmF ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
360 punpckldq xmmA,xmmB ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
361 punpckhdq xmmD,xmmB ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
363 punpckldq xmmC,xmmG ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
364 punpckhdq xmmH,xmmG ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
366 cmp rcx, byte SIZEOF_XMMWORD
367 jb short .column_st32
369 test rdi, SIZEOF_XMMWORD-1
371 ; --(aligned)-------------------
372 movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
373 movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
374 movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
375 movntdq XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
376 add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
378 .out1: ; --(unaligned)-----------------
379 pcmpeqb xmmE,xmmE ; xmmE=(all 1's)
380 maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [rdi], xmmA
381 add rdi, byte SIZEOF_XMMWORD ; outptr
382 maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [rdi], xmmD
383 add rdi, byte SIZEOF_XMMWORD ; outptr
384 maskmovdqu xmmC,xmmE ; movntdqu XMMWORD [rdi], xmmC
385 add rdi, byte SIZEOF_XMMWORD ; outptr
386 maskmovdqu xmmH,xmmE ; movntdqu XMMWORD [rdi], xmmH
387 add rdi, byte SIZEOF_XMMWORD ; outptr
389 sub rcx, byte SIZEOF_XMMWORD
392 add rsi, byte SIZEOF_XMMWORD ; inptr0
393 add rbx, byte SIZEOF_XMMWORD ; inptr1
394 add rdx, byte SIZEOF_XMMWORD ; inptr2
398 pcmpeqb xmmE,xmmE ; xmmE=(all 1's)
399 cmp rcx, byte SIZEOF_XMMWORD/2
400 jb short .column_st16
401 maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [rdi], xmmA
402 add rdi, byte SIZEOF_XMMWORD ; outptr
403 maskmovdqu xmmD,xmmE ; movntdqu XMMWORD [rdi], xmmD
404 add rdi, byte SIZEOF_XMMWORD ; outptr
407 sub rcx, byte SIZEOF_XMMWORD/2
409 cmp rcx, byte SIZEOF_XMMWORD/4
410 jb short .column_st15
411 maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [rdi], xmmA
412 add rdi, byte SIZEOF_XMMWORD ; outptr
414 sub rcx, byte SIZEOF_XMMWORD/4
416 cmp rcx, byte SIZEOF_XMMWORD/16
427 and rcx, byte SIZEOF_XMMWORD-1
429 lea rax, [rcx+rax*4] ; RGB_PIXELSIZE
430 cmp rax, byte SIZEOF_XMMWORD
432 and rdi, byte (-SIZEOF_XMMWORD) ; align to 16-byte boundary
433 shl rcx, 3 ; pslldq xmmA,ecx & pslldq xmmE,ecx
436 pslldq xmmA, SIZEOF_XMMWORD/2
437 pslldq xmmE, SIZEOF_XMMWORD/2
439 sub rcx, byte (SIZEOF_XMMWORD/2)*BYTE_BIT
453 .adj0: ; ----------------
454 maskmovdqu xmmA,xmmE ; movntdqu XMMWORD [rdi], xmmA
456 %endif ; RGB_PIXELSIZE ; ---------------
466 add rsi, byte SIZEOF_JSAMPROW
467 add rbx, byte SIZEOF_JSAMPROW
468 add rdx, byte SIZEOF_JSAMPROW
469 add rdi, byte SIZEOF_JSAMPROW ; output_buf
473 sfence ; flush the write buffer
478 mov rsp,rbp ; rsp <- aligned rbp
479 pop rsp ; rsp <- original rbp
483 ; For some reason, the OS X linker does not honor the request to align the
484 ; segment unless we do this.