2 ; jccolext.asm - colorspace conversion (64-bit AVX2)
4 ; Copyright (C) 2009, 2016, D. R. Commander.
5 ; Copyright (C) 2015, Intel Corporation.
6 ; Copyright (C) 2018, Matthias Räncker.
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
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
18 %include "jcolsamp.inc"
20 ; --------------------------------------------------------------------------
22 ; Convert some rows of samples to the output colorspace.
25 ; jsimd_rgb_ycc_convert_avx2(JDIMENSION img_width, JSAMPARRAY input_buf,
26 ; JSAMPIMAGE output_buf, JDIMENSION output_row,
30 ; r10d = JDIMENSION img_width
31 ; r11 = JSAMPARRAY input_buf
32 ; r12 = JSAMPIMAGE output_buf
33 ; r13d = JDIMENSION output_row
36 %define wk(i) rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD ; ymmword wk[WK_NUM]
40 GLOBAL_FUNCTION(jsimd_rgb_ycc_convert_avx2)
42 EXTN(jsimd_rgb_ycc_convert_avx2):
44 mov rax, rsp ; rax = original rbp
46 and rsp, byte (-SIZEOF_YMMWORD) ; align to 256 bits
48 mov rbp, rsp ; rbp = aligned rbp
61 mov rdip, JSAMPARRAY [rsi+0*SIZEOF_JSAMPARRAY]
62 mov rbxp, JSAMPARRAY [rsi+1*SIZEOF_JSAMPARRAY]
63 mov rdxp, JSAMPARRAY [rsi+2*SIZEOF_JSAMPARRAY]
64 lea rdi, [rdi+rcx*SIZEOF_JSAMPROW]
65 lea rbx, [rbx+rcx*SIZEOF_JSAMPROW]
66 lea rdx, [rdx+rcx*SIZEOF_JSAMPROW]
81 mov rsip, JSAMPROW [rsi] ; inptr
82 mov rdip, JSAMPROW [rdi] ; outptr0
83 mov rbxp, JSAMPROW [rbx] ; outptr1
84 mov rdxp, JSAMPROW [rdx] ; outptr2
86 cmp rcx, byte SIZEOF_YMMWORD
89 %if RGB_PIXELSIZE == 3 ; ---------------
94 lea rcx, [rcx+rcx*2] ; imul ecx,RGB_PIXELSIZE
97 sub rcx, byte SIZEOF_BYTE
98 movzx rax, byte [rsi+rcx]
102 sub rcx, byte SIZEOF_WORD
103 movzx rdx, word [rsi+rcx]
110 test cl, SIZEOF_DWORD
112 sub rcx, byte SIZEOF_DWORD
113 vmovd xmmF, XMM_DWORD [rsi+rcx]
114 vpslldq xmmA, xmmA, SIZEOF_DWORD
115 vpor xmmA, xmmA, xmmF
117 test cl, SIZEOF_MMWORD
118 jz short .column_ld16
119 sub rcx, byte SIZEOF_MMWORD
120 vmovq xmmB, XMM_MMWORD [rsi+rcx]
121 vpslldq xmmA, xmmA, SIZEOF_MMWORD
122 vpor xmmA, xmmA, xmmB
124 test cl, SIZEOF_XMMWORD
125 jz short .column_ld32
126 sub rcx, byte SIZEOF_XMMWORD
127 vmovdqu xmmB, XMM_MMWORD [rsi+rcx]
128 vperm2i128 ymmA, ymmA, ymmA, 1
131 test cl, SIZEOF_YMMWORD
132 jz short .column_ld64
133 sub rcx, byte SIZEOF_YMMWORD
135 vmovdqu ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
137 test cl, 2*SIZEOF_YMMWORD
138 mov rcx, SIZEOF_YMMWORD
139 jz short .rgb_ycc_cnv
141 vmovdqu ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
142 vmovdqu ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
143 jmp short .rgb_ycc_cnv
146 vmovdqu ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
147 vmovdqu ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
148 vmovdqu ymmB, YMMWORD [rsi+2*SIZEOF_YMMWORD]
151 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
152 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
153 ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
154 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
155 ; ymmB=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
156 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
159 vinserti128 ymmA, ymmF, xmmA, 0 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
160 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
161 vinserti128 ymmC, ymmC, xmmB, 0 ; ymmC=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
162 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
163 vinserti128 ymmB, ymmB, xmmF, 0 ; ymmB=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
164 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
165 vperm2i128 ymmF, ymmC, ymmC, 1 ; ymmF=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
166 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
169 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 10 20 01 11 21 02 12
170 ; 22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I)
171 vpsrldq ymmG, ymmG, 8 ; ymmG=(22 03 13 23 04 14 24 05 0G 1G 2G 0H 1H 2H 0I 1I
172 ; 2I 0J 1J 2J 0K 1K 2K 0L -- -- -- -- -- -- -- --)
174 vpunpckhbw ymmA, ymmA, ymmF ; ymmA=(00 08 10 18 20 28 01 09 11 19 21 29 02 0A 12 1A
175 ; 0G 0O 1G 1O 2G 2O 0H 0P 1H 1P 2H 2P 0I 0Q 1I 1Q)
176 vpslldq ymmF, ymmF, 8 ; ymmF=(-- -- -- -- -- -- -- -- 15 25 06 16 26 07 17 27
177 ; 08 18 28 09 19 29 0A 1A 1L 2L 0M 1M 2M 0N 1N 2N)
179 vpunpcklbw ymmG, ymmG, ymmB ; ymmG=(22 2A 03 0B 13 1B 23 2B 04 0C 14 1C 24 2C 05 0D
180 ; 2I 2Q 0J 0R 1J 1R 2J 2R 0K 0S 1K 1S 2K 2S 0L 0T)
181 vpunpckhbw ymmF, ymmF, ymmB ; ymmF=(15 1D 25 2D 06 0E 16 1E 26 2E 07 0F 17 1F 27 2F
182 ; 1L 1T 2L 2T 0M 0U 1M 1U 2M 2U 0N 0V 1N 1V 2N 2V)
185 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 08 10 18 20 28 01 09
186 ; 11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P)
187 vpsrldq ymmD, ymmD, 8 ; ymmD=(11 19 21 29 02 0A 12 1A 0G 0O 1G 1O 2G 2O 0H 0P
188 ; 1H 1P 2H 2P 0I 0Q 1I 1Q -- -- -- -- -- -- -- --)
190 vpunpckhbw ymmA, ymmA, ymmG ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 01 05 09 0D
191 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 0H 0L 0P 0T)
192 vpslldq ymmG, ymmG, 8 ; ymmG=(-- -- -- -- -- -- -- -- 22 2A 03 0B 13 1B 23 2B
193 ; 04 0C 14 1C 24 2C 05 0D 2I 2Q 0J 0R 1J 1R 2J 2R)
195 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 15 19 1D 21 25 29 2D 02 06 0A 0E 12 16 1A 1E
196 ; 1H 1L 1P 1T 2H 2L 2P 2T 0I 0M 0Q 0U 1I 1M 1Q 1U)
197 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(22 26 2A 2E 03 07 0B 0F 13 17 1B 1F 23 27 2B 2F
198 ; 2I 2M 2Q 2U 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V)
201 vpslldq ymmA, ymmA, 8 ; ymmA=(-- -- -- -- -- -- -- -- 00 04 08 0C 10 14 18 1C
202 ; 20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S)
203 vpsrldq ymmE, ymmE, 8 ; ymmE=(20 24 28 2C 01 05 09 0D 0G 0K 0O 0S 1G 1K 1O 1S
204 ; 2G 2K 2O 2S 0H 0L 0P 0T -- -- -- -- -- -- -- --)
206 vpunpckhbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
207 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
208 vpslldq ymmD, ymmD, 8 ; ymmD=(-- -- -- -- -- -- -- -- 11 15 19 1D 21 25 29 2D
209 ; 02 06 0A 0E 12 16 1A 1E 1H 1L 1P 1T 2H 2L 2P 2T)
211 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 22 24 26 28 2A 2C 2E 01 03 05 07 09 0B 0D 0F
212 ; 2G 2I 2K 2M 2O 2Q 2S 2U 0H 0J 0L 0N 0P 0R 0T 0V)
213 vpunpckhbw ymmD, ymmD, ymmG ; ymmD=(11 13 15 17 19 1B 1D 1F 21 23 25 27 29 2B 2D 2F
214 ; 1H 1J 1L 1N 1P 1R 1T 1V 2H 2J 2L 2N 2P 2R 2T 2V)
216 vpxor ymmH, ymmH, ymmH
219 vpunpcklbw ymmA, ymmA, ymmH ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
220 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
223 vpunpcklbw ymmE, ymmE, ymmH ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
224 vpunpckhbw ymmB, ymmB, ymmH ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
227 vpunpcklbw ymmD, ymmD, ymmH ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
228 vpunpckhbw ymmF, ymmF, ymmH ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
230 %else ; RGB_PIXELSIZE == 4 ; -----------
233 test cl, SIZEOF_XMMWORD/16
235 sub rcx, byte SIZEOF_XMMWORD/16
236 vmovd xmmA, XMM_DWORD [rsi+rcx*RGB_PIXELSIZE]
238 test cl, SIZEOF_XMMWORD/8
240 sub rcx, byte SIZEOF_XMMWORD/8
241 vmovq xmmF, XMM_MMWORD [rsi+rcx*RGB_PIXELSIZE]
242 vpslldq xmmA, xmmA, SIZEOF_MMWORD
243 vpor xmmA, xmmA, xmmF
245 test cl, SIZEOF_XMMWORD/4
247 sub rcx, byte SIZEOF_XMMWORD/4
249 vperm2i128 ymmF, ymmF, ymmF, 1
250 vmovdqu xmmA, XMMWORD [rsi+rcx*RGB_PIXELSIZE]
251 vpor ymmA, ymmA, ymmF
253 test cl, SIZEOF_XMMWORD/2
254 jz short .column_ld16
255 sub rcx, byte SIZEOF_XMMWORD/2
257 vmovdqu ymmA, YMMWORD [rsi+rcx*RGB_PIXELSIZE]
259 test cl, SIZEOF_XMMWORD
260 mov rcx, SIZEOF_YMMWORD
261 jz short .rgb_ycc_cnv
264 vmovdqu ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
265 vmovdqu ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
266 jmp short .rgb_ycc_cnv
269 vmovdqu ymmA, YMMWORD [rsi+0*SIZEOF_YMMWORD]
270 vmovdqu ymmF, YMMWORD [rsi+1*SIZEOF_YMMWORD]
271 vmovdqu ymmE, YMMWORD [rsi+2*SIZEOF_YMMWORD]
272 vmovdqu ymmH, YMMWORD [rsi+3*SIZEOF_YMMWORD]
275 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
276 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
277 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
278 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
279 ; ymmE=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
280 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
281 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
282 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
285 vinserti128 ymmA, ymmA, xmmE, 1 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
286 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
287 vperm2i128 ymmE, ymmB, ymmE, 0x31 ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
288 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
291 vinserti128 ymmF, ymmF, xmmH, 1 ; ymmF=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
292 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
293 vperm2i128 ymmH, ymmB, ymmH, 0x31 ; ymmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
294 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
297 vpunpcklbw ymmA, ymmA, ymmE ; ymmA=(00 04 10 14 20 24 30 34 01 05 11 15 21 25 31 35
298 ; 0G 0K 1G 1K 2G 2K 3G 3K 0H 0L 1H 1L 2H 2L 3H 3L)
299 vpunpckhbw ymmD, ymmD, ymmE ; ymmD=(02 06 12 16 22 26 32 36 03 07 13 17 23 27 33 37
300 ; 0I 0M 1I 1M 2I 2M 3I 3M 0J 0N 1J 1N 2J 2N 3J 3N)
303 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(08 0C 18 1C 28 2C 38 3C 09 0D 19 1D 29 2D 39 3D
304 ; 0O 0S 1O 1S 2O 2S 3O 3S 0P 0T 1P 1T 2P 2T 3P 3T)
305 vpunpckhbw ymmC, ymmC, ymmH ; ymmC=(0A 0E 1A 1E 2A 2E 3A 3E 0B 0F 1B 1F 2B 2F 3B 3F
306 ; 0Q 0U 1Q 1U 2Q 2U 3Q 3U 0R 0V 1R 1V 2R 2V 3R 3V)
309 vpunpcklwd ymmA, ymmA, ymmF ; ymmA=(00 04 08 0C 10 14 18 1C 20 24 28 2C 30 34 38 3C
310 ; 0G 0K 0O 0S 1G 1K 1O 1S 2G 2K 2O 2S 3G 3K 3O 3S)
311 vpunpckhwd ymmB, ymmB, ymmF ; ymmB=(01 05 09 0D 11 15 19 1D 21 25 29 2D 31 35 39 3D
312 ; 0H 0L 0P 0T 1H 1L 1P 1T 2H 2L 2P 2T 3H 3L 3P 3T)
315 vpunpcklwd ymmD, ymmD, ymmC ; ymmD=(02 06 0A 0E 12 16 1A 1E 22 26 2A 2E 32 36 3A 3E
316 ; 0I 0M 0Q 0U 1I 1M 1Q 1U 2I 2M 2Q 2U 3I 3M 3Q 3U)
317 vpunpckhwd ymmG, ymmG, ymmC ; ymmG=(03 07 0B 0F 13 17 1B 1F 23 27 2B 2F 33 37 3B 3F
318 ; 0J 0N 0R 0V 1J 1N 1R 1V 2J 2N 2R 2V 3J 3N 3R 3V)
321 vpunpcklbw ymmA, ymmA, ymmD ; ymmA=(00 02 04 06 08 0A 0C 0E 10 12 14 16 18 1A 1C 1E
322 ; 0G 0I 0K 0M 0O 0Q 0S 0U 1G 1I 1K 1M 1O 1Q 1S 1U)
323 vpunpckhbw ymmE, ymmE, ymmD ; ymmE=(20 22 24 26 28 2A 2C 2E 30 32 34 36 38 3A 3C 3E
324 ; 2G 2I 2K 2M 2O 2Q 2S 2U 3G 3I 3K 3M 3O 3Q 3S 3U)
327 vpunpcklbw ymmB, ymmB, ymmG ; ymmB=(01 03 05 07 09 0B 0D 0F 11 13 15 17 19 1B 1D 1F
328 ; 0H 0J 0L 0N 0P 0R 0T 0V 1H 1J 1L 1N 1P 1R 1T 1V)
329 vpunpckhbw ymmH, ymmH, ymmG ; ymmH=(21 23 25 27 29 2B 2D 2F 31 33 35 37 39 3B 3D 3F
330 ; 2H 2J 2L 2N 2P 2R 2T 2V 3H 3J 3L 3N 3P 3R 3T 3V)
332 vpxor ymmF, ymmF, ymmF
335 vpunpcklbw ymmA, ymmA, ymmF ; ymmA=(00 02 04 06 08 0A 0C 0E 0G 0I 0K 0M 0O 0Q 0S 0U)
336 vpunpckhbw ymmC, ymmC, ymmF ; ymmC=(10 12 14 16 18 1A 1C 1E 1G 1I 1K 1M 1O 1Q 1S 1U)
339 vpunpcklbw ymmB, ymmB, ymmF ; ymmB=(01 03 05 07 09 0B 0D 0F 0H 0J 0L 0N 0P 0R 0T 0V)
340 vpunpckhbw ymmD, ymmD, ymmF ; ymmD=(11 13 15 17 19 1B 1D 1F 1H 1J 1L 1N 1P 1R 1T 1V)
343 vpunpcklbw ymmE, ymmE, ymmF ; ymmE=(20 22 24 26 28 2A 2C 2E 2G 2I 2K 2M 2O 2Q 2S 2U)
344 vpunpckhbw ymmG, ymmG, ymmF ; ymmG=(30 32 34 36 38 3A 3C 3E 3G 3I 3K 3M 3O 3Q 3S 3U)
346 vpunpcklbw ymmF, ymmF, ymmH
347 vpunpckhbw ymmH, ymmH, ymmH
348 vpsrlw ymmF, ymmF, BYTE_BIT ; ymmF=(21 23 25 27 29 2B 2D 2F 2H 2J 2L 2N 2P 2R 2T 2V)
349 vpsrlw ymmH, ymmH, BYTE_BIT ; ymmH=(31 33 35 37 39 3B 3D 3F 3H 3J 3L 3N 3P 3R 3T 3V)
351 %endif ; RGB_PIXELSIZE ; ---------------
353 ; ymm0=R(02468ACEGIKMOQSU)=RE, ymm2=G(02468ACEGIKMOQSU)=GE, ymm4=B(02468ACEGIKMOQSU)=BE
354 ; ymm1=R(13579BDFHJLNPRTV)=RO, ymm3=G(13579BDFHJLNPRTV)=GO, ymm5=B(13579BDFHJLNPRTV)=BO
357 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
358 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
359 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
361 ; (This implementation)
362 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
363 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
364 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
366 vmovdqa YMMWORD [wk(0)], ymm0 ; wk(0)=RE
367 vmovdqa YMMWORD [wk(1)], ymm1 ; wk(1)=RO
368 vmovdqa YMMWORD [wk(2)], ymm4 ; wk(2)=BE
369 vmovdqa YMMWORD [wk(3)], ymm5 ; wk(3)=BO
372 vpunpcklwd ymm1, ymm1, ymm3
373 vpunpckhwd ymm6, ymm6, ymm3
376 vpmaddwd ymm1, ymm1, [rel PW_F0299_F0337] ; ymm1=ROL*FIX(0.299)+GOL*FIX(0.337)
377 vpmaddwd ymm6, ymm6, [rel PW_F0299_F0337] ; ymm6=ROH*FIX(0.299)+GOH*FIX(0.337)
378 vpmaddwd ymm7, ymm7, [rel PW_MF016_MF033] ; ymm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
379 vpmaddwd ymm4, ymm4, [rel PW_MF016_MF033] ; ymm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
381 vmovdqa YMMWORD [wk(4)], ymm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
382 vmovdqa YMMWORD [wk(5)], ymm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
384 vpxor ymm1, ymm1, ymm1
385 vpxor ymm6, ymm6, ymm6
386 vpunpcklwd ymm1, ymm1, ymm5 ; ymm1=BOL
387 vpunpckhwd ymm6, ymm6, ymm5 ; ymm6=BOH
388 vpsrld ymm1, ymm1, 1 ; ymm1=BOL*FIX(0.500)
389 vpsrld ymm6, ymm6, 1 ; ymm6=BOH*FIX(0.500)
391 vmovdqa ymm5, [rel PD_ONEHALFM1_CJ] ; ymm5=[PD_ONEHALFM1_CJ]
393 vpaddd ymm7, ymm7, ymm1
394 vpaddd ymm4, ymm4, ymm6
395 vpaddd ymm7, ymm7, ymm5
396 vpaddd ymm4, ymm4, ymm5
397 vpsrld ymm7, ymm7, SCALEBITS ; ymm7=CbOL
398 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=CbOH
399 vpackssdw ymm7, ymm7, ymm4 ; ymm7=CbO
401 vmovdqa ymm1, YMMWORD [wk(2)] ; ymm1=BE
404 vpunpcklwd ymm0, ymm0, ymm2
405 vpunpckhwd ymm6, ymm6, ymm2
408 vpmaddwd ymm0, ymm0, [rel PW_F0299_F0337] ; ymm0=REL*FIX(0.299)+GEL*FIX(0.337)
409 vpmaddwd ymm6, ymm6, [rel PW_F0299_F0337] ; ymm6=REH*FIX(0.299)+GEH*FIX(0.337)
410 vpmaddwd ymm5, ymm5, [rel PW_MF016_MF033] ; ymm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
411 vpmaddwd ymm4, ymm4, [rel PW_MF016_MF033] ; ymm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
413 vmovdqa YMMWORD [wk(6)], ymm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
414 vmovdqa YMMWORD [wk(7)], ymm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
416 vpxor ymm0, ymm0, ymm0
417 vpxor ymm6, ymm6, ymm6
418 vpunpcklwd ymm0, ymm0, ymm1 ; ymm0=BEL
419 vpunpckhwd ymm6, ymm6, ymm1 ; ymm6=BEH
420 vpsrld ymm0, ymm0, 1 ; ymm0=BEL*FIX(0.500)
421 vpsrld ymm6, ymm6, 1 ; ymm6=BEH*FIX(0.500)
423 vmovdqa ymm1, [rel PD_ONEHALFM1_CJ] ; ymm1=[PD_ONEHALFM1_CJ]
425 vpaddd ymm5, ymm5, ymm0
426 vpaddd ymm4, ymm4, ymm6
427 vpaddd ymm5, ymm5, ymm1
428 vpaddd ymm4, ymm4, ymm1
429 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CbEL
430 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=CbEH
431 vpackssdw ymm5, ymm5, ymm4 ; ymm5=CbE
433 vpsllw ymm7, ymm7, BYTE_BIT
434 vpor ymm5, ymm5, ymm7 ; ymm5=Cb
435 vmovdqu YMMWORD [rbx], ymm5 ; Save Cb
437 vmovdqa ymm0, YMMWORD [wk(3)] ; ymm0=BO
438 vmovdqa ymm6, YMMWORD [wk(2)] ; ymm6=BE
439 vmovdqa ymm1, YMMWORD [wk(1)] ; ymm1=RO
442 vpunpcklwd ymm0, ymm0, ymm3
443 vpunpckhwd ymm4, ymm4, ymm3
446 vpmaddwd ymm0, ymm0, [rel PW_F0114_F0250] ; ymm0=BOL*FIX(0.114)+GOL*FIX(0.250)
447 vpmaddwd ymm4, ymm4, [rel PW_F0114_F0250] ; ymm4=BOH*FIX(0.114)+GOH*FIX(0.250)
448 vpmaddwd ymm7, ymm7, [rel PW_MF008_MF041] ; ymm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
449 vpmaddwd ymm5, ymm5, [rel PW_MF008_MF041] ; ymm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
451 vmovdqa ymm3, [rel PD_ONEHALF] ; ymm3=[PD_ONEHALF]
453 vpaddd ymm0, ymm0, YMMWORD [wk(4)]
454 vpaddd ymm4, ymm4, YMMWORD [wk(5)]
455 vpaddd ymm0, ymm0, ymm3
456 vpaddd ymm4, ymm4, ymm3
457 vpsrld ymm0, ymm0, SCALEBITS ; ymm0=YOL
458 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YOH
459 vpackssdw ymm0, ymm0, ymm4 ; ymm0=YO
461 vpxor ymm3, ymm3, ymm3
462 vpxor ymm4, ymm4, ymm4
463 vpunpcklwd ymm3, ymm3, ymm1 ; ymm3=ROL
464 vpunpckhwd ymm4, ymm4, ymm1 ; ymm4=ROH
465 vpsrld ymm3, ymm3, 1 ; ymm3=ROL*FIX(0.500)
466 vpsrld ymm4, ymm4, 1 ; ymm4=ROH*FIX(0.500)
468 vmovdqa ymm1, [rel PD_ONEHALFM1_CJ] ; ymm1=[PD_ONEHALFM1_CJ]
470 vpaddd ymm7, ymm7, ymm3
471 vpaddd ymm5, ymm5, ymm4
472 vpaddd ymm7, ymm7, ymm1
473 vpaddd ymm5, ymm5, ymm1
474 vpsrld ymm7, ymm7, SCALEBITS ; ymm7=CrOL
475 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CrOH
476 vpackssdw ymm7, ymm7, ymm5 ; ymm7=CrO
478 vmovdqa ymm3, YMMWORD [wk(0)] ; ymm3=RE
481 vpunpcklwd ymm6, ymm6, ymm2
482 vpunpckhwd ymm4, ymm4, ymm2
485 vpmaddwd ymm6, ymm6, [rel PW_F0114_F0250] ; ymm6=BEL*FIX(0.114)+GEL*FIX(0.250)
486 vpmaddwd ymm4, ymm4, [rel PW_F0114_F0250] ; ymm4=BEH*FIX(0.114)+GEH*FIX(0.250)
487 vpmaddwd ymm1, ymm1, [rel PW_MF008_MF041] ; ymm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
488 vpmaddwd ymm5, ymm5, [rel PW_MF008_MF041] ; ymm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
490 vmovdqa ymm2, [rel PD_ONEHALF] ; ymm2=[PD_ONEHALF]
492 vpaddd ymm6, ymm6, YMMWORD [wk(6)]
493 vpaddd ymm4, ymm4, YMMWORD [wk(7)]
494 vpaddd ymm6, ymm6, ymm2
495 vpaddd ymm4, ymm4, ymm2
496 vpsrld ymm6, ymm6, SCALEBITS ; ymm6=YEL
497 vpsrld ymm4, ymm4, SCALEBITS ; ymm4=YEH
498 vpackssdw ymm6, ymm6, ymm4 ; ymm6=YE
500 vpsllw ymm0, ymm0, BYTE_BIT
501 vpor ymm6, ymm6, ymm0 ; ymm6=Y
502 vmovdqu YMMWORD [rdi], ymm6 ; Save Y
504 vpxor ymm2, ymm2, ymm2
505 vpxor ymm4, ymm4, ymm4
506 vpunpcklwd ymm2, ymm2, ymm3 ; ymm2=REL
507 vpunpckhwd ymm4, ymm4, ymm3 ; ymm4=REH
508 vpsrld ymm2, ymm2, 1 ; ymm2=REL*FIX(0.500)
509 vpsrld ymm4, ymm4, 1 ; ymm4=REH*FIX(0.500)
511 vmovdqa ymm0, [rel PD_ONEHALFM1_CJ] ; ymm0=[PD_ONEHALFM1_CJ]
513 vpaddd ymm1, ymm1, ymm2
514 vpaddd ymm5, ymm5, ymm4
515 vpaddd ymm1, ymm1, ymm0
516 vpaddd ymm5, ymm5, ymm0
517 vpsrld ymm1, ymm1, SCALEBITS ; ymm1=CrEL
518 vpsrld ymm5, ymm5, SCALEBITS ; ymm5=CrEH
519 vpackssdw ymm1, ymm1, ymm5 ; ymm1=CrE
521 vpsllw ymm7, ymm7, BYTE_BIT
522 vpor ymm1, ymm1, ymm7 ; ymm1=Cr
523 vmovdqu YMMWORD [rdx], ymm1 ; Save Cr
525 sub rcx, byte SIZEOF_YMMWORD
526 add rsi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; inptr
527 add rdi, byte SIZEOF_YMMWORD ; outptr0
528 add rbx, byte SIZEOF_YMMWORD ; outptr1
529 add rdx, byte SIZEOF_YMMWORD ; outptr2
530 cmp rcx, byte SIZEOF_YMMWORD
541 add rsi, byte SIZEOF_JSAMPROW ; input_buf
542 add rdi, byte SIZEOF_JSAMPROW
543 add rbx, byte SIZEOF_JSAMPROW
544 add rdx, byte SIZEOF_JSAMPROW
552 mov rsp, rbp ; rsp <- aligned rbp
553 pop rsp ; rsp <- original rbp
557 ; For some reason, the OS X linker does not honor the request to align the
558 ; segment unless we do this.