2 ; jdmrgss2-64.asm - merged upsampling/color conversion (64-bit SSE2)
4 ; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright 2009, 2012 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 ; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
27 ; jsimd_h2v1_merged_upsample_sse2 (JDIMENSION output_width,
28 ; JSAMPIMAGE input_buf,
29 ; JDIMENSION in_row_group_ctr,
30 ; JSAMPARRAY output_buf);
33 ; r10 = JDIMENSION output_width
34 ; r11 = JSAMPIMAGE input_buf
35 ; r12 = JDIMENSION in_row_group_ctr
36 ; r13 = JSAMPARRAY output_buf
38 %define wk(i) rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
42 global EXTN(jsimd_h2v1_merged_upsample_sse2) PRIVATE
44 EXTN(jsimd_h2v1_merged_upsample_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
63 mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
64 mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
65 mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
67 mov rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
68 mov rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
69 mov rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
70 mov rdi, JSAMPROW [rdi] ; outptr
76 movdqa xmm6, XMMWORD [rbx] ; xmm6=Cb(0123456789ABCDEF)
77 movdqa xmm7, XMMWORD [rdx] ; xmm7=Cr(0123456789ABCDEF)
79 pxor xmm1,xmm1 ; xmm1=(all 0's)
81 psllw xmm3,7 ; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
84 punpckhbw xmm6,xmm1 ; xmm6=Cb(89ABCDEF)=CbH
85 punpcklbw xmm4,xmm1 ; xmm4=Cb(01234567)=CbL
87 punpckhbw xmm7,xmm1 ; xmm7=Cr(89ABCDEF)=CrH
88 punpcklbw xmm0,xmm1 ; xmm0=Cr(01234567)=CrL
96 ; R = Y + 1.40200 * Cr
97 ; G = Y - 0.34414 * Cb - 0.71414 * Cr
98 ; B = Y + 1.77200 * Cb
100 ; (This implementation)
101 ; R = Y + 0.40200 * Cr + Cr
102 ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
103 ; B = Y - 0.22800 * Cb + Cb + Cb
105 movdqa xmm5,xmm6 ; xmm5=CbH
106 movdqa xmm2,xmm4 ; xmm2=CbL
107 paddw xmm6,xmm6 ; xmm6=2*CbH
108 paddw xmm4,xmm4 ; xmm4=2*CbL
109 movdqa xmm1,xmm7 ; xmm1=CrH
110 movdqa xmm3,xmm0 ; xmm3=CrL
111 paddw xmm7,xmm7 ; xmm7=2*CrH
112 paddw xmm0,xmm0 ; xmm0=2*CrL
114 pmulhw xmm6,[rel PW_MF0228] ; xmm6=(2*CbH * -FIX(0.22800))
115 pmulhw xmm4,[rel PW_MF0228] ; xmm4=(2*CbL * -FIX(0.22800))
116 pmulhw xmm7,[rel PW_F0402] ; xmm7=(2*CrH * FIX(0.40200))
117 pmulhw xmm0,[rel PW_F0402] ; xmm0=(2*CrL * FIX(0.40200))
119 paddw xmm6,[rel PW_ONE]
120 paddw xmm4,[rel PW_ONE]
121 psraw xmm6,1 ; xmm6=(CbH * -FIX(0.22800))
122 psraw xmm4,1 ; xmm4=(CbL * -FIX(0.22800))
123 paddw xmm7,[rel PW_ONE]
124 paddw xmm0,[rel PW_ONE]
125 psraw xmm7,1 ; xmm7=(CrH * FIX(0.40200))
126 psraw xmm0,1 ; xmm0=(CrL * FIX(0.40200))
130 paddw xmm6,xmm5 ; xmm6=(CbH * FIX(1.77200))=(B-Y)H
131 paddw xmm4,xmm2 ; xmm4=(CbL * FIX(1.77200))=(B-Y)L
132 paddw xmm7,xmm1 ; xmm7=(CrH * FIX(1.40200))=(R-Y)H
133 paddw xmm0,xmm3 ; xmm0=(CrL * FIX(1.40200))=(R-Y)L
135 movdqa XMMWORD [wk(0)], xmm6 ; wk(0)=(B-Y)H
136 movdqa XMMWORD [wk(1)], xmm7 ; wk(1)=(R-Y)H
142 pmaddwd xmm5,[rel PW_MF0344_F0285]
143 pmaddwd xmm6,[rel PW_MF0344_F0285]
146 pmaddwd xmm2,[rel PW_MF0344_F0285]
147 pmaddwd xmm7,[rel PW_MF0344_F0285]
149 paddd xmm5,[rel PD_ONEHALF]
150 paddd xmm6,[rel PD_ONEHALF]
153 paddd xmm2,[rel PD_ONEHALF]
154 paddd xmm7,[rel PD_ONEHALF]
158 packssdw xmm5,xmm6 ; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
159 packssdw xmm2,xmm7 ; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
160 psubw xmm5,xmm1 ; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
161 psubw xmm2,xmm3 ; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
163 movdqa XMMWORD [wk(2)], xmm5 ; wk(2)=(G-Y)H
169 movdqa xmm0, XMMWORD [wk(1)] ; xmm0=(R-Y)H
170 movdqa xmm2, XMMWORD [wk(2)] ; xmm2=(G-Y)H
171 movdqa xmm4, XMMWORD [wk(0)] ; xmm4=(B-Y)H
174 movdqa xmm7, XMMWORD [rsi] ; xmm7=Y(0123456789ABCDEF)
177 psrlw xmm6,BYTE_BIT ; xmm6={0xFF 0x00 0xFF 0x00 ..}
178 pand xmm6,xmm7 ; xmm6=Y(02468ACE)=YE
179 psrlw xmm7,BYTE_BIT ; xmm7=Y(13579BDF)=YO
181 movdqa xmm1,xmm0 ; xmm1=xmm0=(R-Y)(L/H)
182 movdqa xmm3,xmm2 ; xmm3=xmm2=(G-Y)(L/H)
183 movdqa xmm5,xmm4 ; xmm5=xmm4=(B-Y)(L/H)
185 paddw xmm0,xmm6 ; xmm0=((R-Y)+YE)=RE=R(02468ACE)
186 paddw xmm1,xmm7 ; xmm1=((R-Y)+YO)=RO=R(13579BDF)
187 packuswb xmm0,xmm0 ; xmm0=R(02468ACE********)
188 packuswb xmm1,xmm1 ; xmm1=R(13579BDF********)
190 paddw xmm2,xmm6 ; xmm2=((G-Y)+YE)=GE=G(02468ACE)
191 paddw xmm3,xmm7 ; xmm3=((G-Y)+YO)=GO=G(13579BDF)
192 packuswb xmm2,xmm2 ; xmm2=G(02468ACE********)
193 packuswb xmm3,xmm3 ; xmm3=G(13579BDF********)
195 paddw xmm4,xmm6 ; xmm4=((B-Y)+YE)=BE=B(02468ACE)
196 paddw xmm5,xmm7 ; xmm5=((B-Y)+YO)=BO=B(13579BDF)
197 packuswb xmm4,xmm4 ; xmm4=B(02468ACE********)
198 packuswb xmm5,xmm5 ; xmm5=B(13579BDF********)
200 %if RGB_PIXELSIZE == 3 ; ---------------
202 ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
203 ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
204 ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
205 ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
207 punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
208 punpcklbw xmmE,xmmB ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
209 punpcklbw xmmD,xmmF ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
213 punpcklwd xmmA,xmmE ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
214 punpckhwd xmmG,xmmE ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
216 psrldq xmmH,2 ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
217 psrldq xmmE,2 ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
221 punpcklwd xmmD,xmmH ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
222 punpckhwd xmmC,xmmH ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
224 psrldq xmmB,2 ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
227 punpcklwd xmmE,xmmB ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
228 punpckhwd xmmF,xmmB ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
230 pshufd xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
232 punpckldq xmmA,xmmD ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
233 punpckldq xmmE,xmmH ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
234 punpckhdq xmmD,xmmB ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
236 pshufd xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
238 punpckldq xmmG,xmmC ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
239 punpckldq xmmF,xmmH ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
240 punpckhdq xmmC,xmmB ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
242 punpcklqdq xmmA,xmmE ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
243 punpcklqdq xmmD,xmmG ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
244 punpcklqdq xmmF,xmmC ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
246 cmp rcx, byte SIZEOF_XMMWORD
247 jb short .column_st32
249 test rdi, SIZEOF_XMMWORD-1
251 ; --(aligned)-------------------
252 movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
253 movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
254 movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
256 .out1: ; --(unaligned)-----------------
257 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
258 movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
259 movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
261 add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
262 sub rcx, byte SIZEOF_XMMWORD
265 add rsi, byte SIZEOF_XMMWORD ; inptr0
269 add rbx, byte SIZEOF_XMMWORD ; inptr1
270 add rdx, byte SIZEOF_XMMWORD ; inptr2
274 lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE
275 cmp rcx, byte 2*SIZEOF_XMMWORD
276 jb short .column_st16
277 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
278 movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
279 add rdi, byte 2*SIZEOF_XMMWORD ; outptr
281 sub rcx, byte 2*SIZEOF_XMMWORD
282 jmp short .column_st15
284 cmp rcx, byte SIZEOF_XMMWORD
285 jb short .column_st15
286 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
287 add rdi, byte SIZEOF_XMMWORD ; outptr
289 sub rcx, byte SIZEOF_XMMWORD
291 ; Store the lower 8 bytes of xmmA to the output when it has enough
293 cmp rcx, byte SIZEOF_MMWORD
295 movq XMM_MMWORD [rdi], xmmA
296 add rdi, byte SIZEOF_MMWORD
297 sub rcx, byte SIZEOF_MMWORD
298 psrldq xmmA, SIZEOF_MMWORD
300 ; Store the lower 4 bytes of xmmA to the output when it has enough
302 cmp rcx, byte SIZEOF_DWORD
304 movd XMM_DWORD [rdi], xmmA
305 add rdi, byte SIZEOF_DWORD
306 sub rcx, byte SIZEOF_DWORD
307 psrldq xmmA, SIZEOF_DWORD
309 ; Store the lower 2 bytes of rax to the output when it has enough
312 cmp rcx, byte SIZEOF_WORD
315 add rdi, byte SIZEOF_WORD
316 sub rcx, byte SIZEOF_WORD
319 ; Store the lower 1 byte of rax to the output when it has enough
325 %else ; RGB_PIXELSIZE == 4 ; -----------
327 %ifdef RGBX_FILLER_0XFF
328 pcmpeqb xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
329 pcmpeqb xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
331 pxor xmm6,xmm6 ; xmm6=XE=X(02468ACE********)
332 pxor xmm7,xmm7 ; xmm7=XO=X(13579BDF********)
334 ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
335 ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
336 ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
337 ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
339 punpcklbw xmmA,xmmC ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
340 punpcklbw xmmE,xmmG ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
341 punpcklbw xmmB,xmmD ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
342 punpcklbw xmmF,xmmH ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
345 punpcklwd xmmA,xmmE ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
346 punpckhwd xmmC,xmmE ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
348 punpcklwd xmmB,xmmF ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
349 punpckhwd xmmG,xmmF ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
352 punpckldq xmmA,xmmB ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
353 punpckhdq xmmD,xmmB ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
355 punpckldq xmmC,xmmG ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
356 punpckhdq xmmH,xmmG ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
358 cmp rcx, byte SIZEOF_XMMWORD
359 jb short .column_st32
361 test rdi, SIZEOF_XMMWORD-1
363 ; --(aligned)-------------------
364 movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
365 movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
366 movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
367 movntdq XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
369 .out1: ; --(unaligned)-----------------
370 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
371 movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
372 movdqu XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
373 movdqu XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
375 add rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD ; outptr
376 sub rcx, byte SIZEOF_XMMWORD
379 add rsi, byte SIZEOF_XMMWORD ; inptr0
383 add rbx, byte SIZEOF_XMMWORD ; inptr1
384 add rdx, byte SIZEOF_XMMWORD ; inptr2
388 cmp rcx, byte SIZEOF_XMMWORD/2
389 jb short .column_st16
390 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
391 movdqu XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
392 add rdi, byte 2*SIZEOF_XMMWORD ; outptr
395 sub rcx, byte SIZEOF_XMMWORD/2
397 cmp rcx, byte SIZEOF_XMMWORD/4
398 jb short .column_st15
399 movdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
400 add rdi, byte SIZEOF_XMMWORD ; outptr
402 sub rcx, byte SIZEOF_XMMWORD/4
404 ; Store two pixels (8 bytes) of xmmA to the output when it has enough
406 cmp rcx, byte SIZEOF_XMMWORD/8
408 movq XMM_MMWORD [rdi], xmmA
409 add rdi, byte SIZEOF_XMMWORD/8*4
410 sub rcx, byte SIZEOF_XMMWORD/8
411 psrldq xmmA, SIZEOF_XMMWORD/8*4
413 ; Store one pixel (4 bytes) of xmmA to the output when it has enough
417 movd XMM_DWORD [rdi], xmmA
419 %endif ; RGB_PIXELSIZE ; ---------------
422 sfence ; flush the write buffer
427 mov rsp,rbp ; rsp <- aligned rbp
428 pop rsp ; rsp <- original rbp
432 ; --------------------------------------------------------------------------
434 ; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
437 ; jsimd_h2v2_merged_upsample_sse2 (JDIMENSION output_width,
438 ; JSAMPIMAGE input_buf,
439 ; JDIMENSION in_row_group_ctr,
440 ; JSAMPARRAY output_buf);
443 ; r10 = JDIMENSION output_width
444 ; r11 = JSAMPIMAGE input_buf
445 ; r12 = JDIMENSION in_row_group_ctr
446 ; r13 = JSAMPARRAY output_buf
449 global EXTN(jsimd_h2v2_merged_upsample_sse2) PRIVATE
451 EXTN(jsimd_h2v2_merged_upsample_sse2):
462 mov rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
463 mov rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
464 mov rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
466 lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
489 call EXTN(jsimd_h2v1_merged_upsample_sse2)
498 add rdi, byte SIZEOF_JSAMPROW ; outptr1
499 add rsi, byte SIZEOF_JSAMPROW ; inptr01
522 call EXTN(jsimd_h2v1_merged_upsample_sse2)
536 ; For some reason, the OS X linker does not honor the request to align the
537 ; segment unless we do this.