2 ; jdmrgext.asm - merged upsampling/color conversion (64-bit AVX2)
4 ; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
5 ; Copyright (C) 2009, 2012, 2016, D. R. Commander.
6 ; Copyright (C) 2015, Intel Corporation.
7 ; Copyright (C) 2018, Matthias Räncker.
8 ; Copyright (C) 2023, Aliaksiej Kandracienka.
10 ; Based on the x86 SIMD extension for IJG JPEG library
11 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
12 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
14 ; This file should be assembled with NASM (Netwide Assembler),
15 ; can *not* be assembled with Microsoft's MASM or any compatible
16 ; assembler (including Borland's Turbo Assembler).
17 ; NASM is available from http://nasm.sourceforge.net/ or
18 ; 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_avx2(JDIMENSION output_width,
28 ; JSAMPIMAGE input_buf,
29 ; JDIMENSION in_row_group_ctr,
30 ; JSAMPARRAY output_buf);
33 ; r10d = JDIMENSION output_width
34 ; r11 = JSAMPIMAGE input_buf
35 ; r12d = JDIMENSION in_row_group_ctr
36 ; r13 = JSAMPARRAY output_buf
38 %define wk(i) r15 - (WK_NUM - (i)) * SIZEOF_YMMWORD ; ymmword wk[WK_NUM]
42 GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_avx2)
44 EXTN(jsimd_h2v1_merged_upsample_avx2):
48 and rsp, byte (-SIZEOF_YMMWORD) ; align to 256 bits
49 ; Allocate stack space for wk array. r15 is used to access it.
51 sub rsp, SIZEOF_YMMWORD * WK_NUM
63 mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
64 mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
65 mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
67 mov rsip, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW] ; inptr0
68 mov rbxp, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW] ; inptr1
69 mov rdxp, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW] ; inptr2
70 mov rdip, JSAMPROW [rdi] ; outptr
76 vmovdqu ymm6, YMMWORD [rbx] ; ymm6=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
77 vmovdqu ymm7, YMMWORD [rdx] ; ymm7=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
79 vpxor ymm1, ymm1, ymm1 ; ymm1=(all 0's)
80 vpcmpeqw ymm3, ymm3, ymm3
81 vpsllw ymm3, ymm3, 7 ; ymm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
83 vpermq ymm6, ymm6, 0xd8 ; ymm6=Cb(01234567GHIJKLMN89ABCDEFOPQRSTUV)
84 vpermq ymm7, ymm7, 0xd8 ; ymm7=Cr(01234567GHIJKLMN89ABCDEFOPQRSTUV)
85 vpunpcklbw ymm4, ymm6, ymm1 ; ymm4=Cb(0123456789ABCDEF)=CbL
86 vpunpckhbw ymm6, ymm6, ymm1 ; ymm6=Cb(GHIJKLMNOPQRSTUV)=CbH
87 vpunpcklbw ymm0, ymm7, ymm1 ; ymm0=Cr(0123456789ABCDEF)=CrL
88 vpunpckhbw ymm7, ymm7, ymm1 ; ymm7=Cr(GHIJKLMNOPQRSTUV)=CrH
90 vpaddw ymm5, ymm6, ymm3
91 vpaddw ymm2, ymm4, ymm3
92 vpaddw ymm1, ymm7, ymm3
93 vpaddw ymm3, ymm0, ymm3
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 vpaddw ymm6, ymm5, ymm5 ; ymm6=2*CbH
106 vpaddw ymm4, ymm2, ymm2 ; ymm4=2*CbL
107 vpaddw ymm7, ymm1, ymm1 ; ymm7=2*CrH
108 vpaddw ymm0, ymm3, ymm3 ; ymm0=2*CrL
110 vpmulhw ymm6, ymm6, [rel PW_MF0228] ; ymm6=(2*CbH * -FIX(0.22800))
111 vpmulhw ymm4, ymm4, [rel PW_MF0228] ; ymm4=(2*CbL * -FIX(0.22800))
112 vpmulhw ymm7, ymm7, [rel PW_F0402] ; ymm7=(2*CrH * FIX(0.40200))
113 vpmulhw ymm0, ymm0, [rel PW_F0402] ; ymm0=(2*CrL * FIX(0.40200))
115 vpaddw ymm6, ymm6, [rel PW_ONE]
116 vpaddw ymm4, ymm4, [rel PW_ONE]
117 vpsraw ymm6, ymm6, 1 ; ymm6=(CbH * -FIX(0.22800))
118 vpsraw ymm4, ymm4, 1 ; ymm4=(CbL * -FIX(0.22800))
119 vpaddw ymm7, ymm7, [rel PW_ONE]
120 vpaddw ymm0, ymm0, [rel PW_ONE]
121 vpsraw ymm7, ymm7, 1 ; ymm7=(CrH * FIX(0.40200))
122 vpsraw ymm0, ymm0, 1 ; ymm0=(CrL * FIX(0.40200))
124 vpaddw ymm6, ymm6, ymm5
125 vpaddw ymm4, ymm4, ymm2
126 vpaddw ymm6, ymm6, ymm5 ; ymm6=(CbH * FIX(1.77200))=(B-Y)H
127 vpaddw ymm4, ymm4, ymm2 ; ymm4=(CbL * FIX(1.77200))=(B-Y)L
128 vpaddw ymm7, ymm7, ymm1 ; ymm7=(CrH * FIX(1.40200))=(R-Y)H
129 vpaddw ymm0, ymm0, ymm3 ; ymm0=(CrL * FIX(1.40200))=(R-Y)L
131 vmovdqa YMMWORD [wk(0)], ymm6 ; wk(0)=(B-Y)H
132 vmovdqa YMMWORD [wk(1)], ymm7 ; wk(1)=(R-Y)H
134 vpunpckhwd ymm6, ymm5, ymm1
135 vpunpcklwd ymm5, ymm5, ymm1
136 vpmaddwd ymm5, ymm5, [rel PW_MF0344_F0285]
137 vpmaddwd ymm6, ymm6, [rel PW_MF0344_F0285]
138 vpunpckhwd ymm7, ymm2, ymm3
139 vpunpcklwd ymm2, ymm2, ymm3
140 vpmaddwd ymm2, ymm2, [rel PW_MF0344_F0285]
141 vpmaddwd ymm7, ymm7, [rel PW_MF0344_F0285]
143 vpaddd ymm5, ymm5, [rel PD_ONEHALF]
144 vpaddd ymm6, ymm6, [rel PD_ONEHALF]
145 vpsrad ymm5, ymm5, SCALEBITS
146 vpsrad ymm6, ymm6, SCALEBITS
147 vpaddd ymm2, ymm2, [rel PD_ONEHALF]
148 vpaddd ymm7, ymm7, [rel PD_ONEHALF]
149 vpsrad ymm2, ymm2, SCALEBITS
150 vpsrad ymm7, ymm7, SCALEBITS
152 vpackssdw ymm5, ymm5, ymm6 ; ymm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
153 vpackssdw ymm2, ymm2, ymm7 ; ymm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
154 vpsubw ymm5, ymm5, ymm1 ; ymm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
155 vpsubw ymm2, ymm2, ymm3 ; ymm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
157 vmovdqa YMMWORD [wk(2)], ymm5 ; wk(2)=(G-Y)H
163 vmovdqa ymm0, YMMWORD [wk(1)] ; ymm0=(R-Y)H
164 vmovdqa ymm2, YMMWORD [wk(2)] ; ymm2=(G-Y)H
165 vmovdqa ymm4, YMMWORD [wk(0)] ; ymm4=(B-Y)H
168 vmovdqu ymm7, YMMWORD [rsi] ; ymm7=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
170 vpcmpeqw ymm6, ymm6, ymm6
171 vpsrlw ymm6, ymm6, BYTE_BIT ; ymm6={0xFF 0x00 0xFF 0x00 ..}
172 vpand ymm6, ymm6, ymm7 ; ymm6=Y(02468ACEGIKMOQSU)=YE
173 vpsrlw ymm7, ymm7, BYTE_BIT ; ymm7=Y(13579BDFHJLNPRTV)=YO
175 vmovdqa ymm1, ymm0 ; ymm1=ymm0=(R-Y)(L/H)
176 vmovdqa ymm3, ymm2 ; ymm3=ymm2=(G-Y)(L/H)
177 vmovdqa ymm5, ymm4 ; ymm5=ymm4=(B-Y)(L/H)
179 vpaddw ymm0, ymm0, ymm6 ; ymm0=((R-Y)+YE)=RE=R(02468ACEGIKMOQSU)
180 vpaddw ymm1, ymm1, ymm7 ; ymm1=((R-Y)+YO)=RO=R(13579BDFHJLNPRTV)
181 vpackuswb ymm0, ymm0, ymm0 ; ymm0=R(02468ACE********GIKMOQSU********)
182 vpackuswb ymm1, ymm1, ymm1 ; ymm1=R(13579BDF********HJLNPRTV********)
184 vpaddw ymm2, ymm2, ymm6 ; ymm2=((G-Y)+YE)=GE=G(02468ACEGIKMOQSU)
185 vpaddw ymm3, ymm3, ymm7 ; ymm3=((G-Y)+YO)=GO=G(13579BDFHJLNPRTV)
186 vpackuswb ymm2, ymm2, ymm2 ; ymm2=G(02468ACE********GIKMOQSU********)
187 vpackuswb ymm3, ymm3, ymm3 ; ymm3=G(13579BDF********HJLNPRTV********)
189 vpaddw ymm4, ymm4, ymm6 ; ymm4=((B-Y)+YE)=BE=B(02468ACEGIKMOQSU)
190 vpaddw ymm5, ymm5, ymm7 ; ymm5=((B-Y)+YO)=BO=B(13579BDFHJLNPRTV)
191 vpackuswb ymm4, ymm4, ymm4 ; ymm4=B(02468ACE********GIKMOQSU********)
192 vpackuswb ymm5, ymm5, ymm5 ; ymm5=B(13579BDF********HJLNPRTV********)
194 %if RGB_PIXELSIZE == 3 ; ---------------
196 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
197 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
198 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
199 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
200 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
201 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
202 ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
203 ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
205 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
206 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
207 vpunpcklbw ymmE, ymmE, ymmB ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
208 ; 2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
209 vpunpcklbw ymmD, ymmD, ymmF ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
210 ; 1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
212 vpsrldq ymmH, ymmA, 2 ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
213 ; 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
214 vpunpckhwd ymmG, ymmA, ymmE ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
215 ; 0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
216 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
217 ; 0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
219 vpsrldq ymmE, ymmE, 2 ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
220 ; 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
222 vpsrldq ymmB, ymmD, 2 ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
223 ; 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
224 vpunpckhwd ymmC, ymmD, ymmH ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
225 ; 1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
226 vpunpcklwd ymmD, ymmD, ymmH ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
227 ; 1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
229 vpunpckhwd ymmF, ymmE, ymmB ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
230 ; 2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
231 vpunpcklwd ymmE, ymmE, ymmB ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
232 ; 2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
234 vpshufd ymmH, ymmA, 0x4E ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
235 ; 0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
236 vpunpckldq ymmA, ymmA, ymmD ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
237 ; 0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
238 vpunpckhdq ymmD, ymmD, ymmE ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
239 ; 1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
240 vpunpckldq ymmE, ymmE, ymmH ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
241 ; 2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
243 vpshufd ymmH, ymmG, 0x4E ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
244 ; 0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
245 vpunpckldq ymmG, ymmG, ymmC ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
246 ; 0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
247 vpunpckhdq ymmC, ymmC, ymmF ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
248 ; 1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
249 vpunpckldq ymmF, ymmF, ymmH ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
250 ; 2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
252 vpunpcklqdq ymmH, ymmA, ymmE ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
253 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
254 vpunpcklqdq ymmG, ymmD, ymmG ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
255 ; 1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
256 vpunpcklqdq ymmC, ymmF, ymmC ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
257 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
259 vperm2i128 ymmA, ymmH, ymmG, 0x20 ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
260 ; 15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
261 vperm2i128 ymmD, ymmC, ymmH, 0x30 ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
262 ; 0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
263 vperm2i128 ymmF, ymmG, ymmC, 0x31 ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
264 ; 2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
266 cmp rcx, byte SIZEOF_YMMWORD
267 jb short .column_st64
269 test rdi, SIZEOF_YMMWORD-1
271 ; --(aligned)-------------------
272 vmovntdq YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
273 vmovntdq YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
274 vmovntdq YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
276 .out1: ; --(unaligned)-----------------
277 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
278 vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
279 vmovdqu YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
281 add rdi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr
282 sub rcx, byte SIZEOF_YMMWORD
285 add rsi, byte SIZEOF_YMMWORD ; inptr0
289 add rbx, byte SIZEOF_YMMWORD ; inptr1
290 add rdx, byte SIZEOF_YMMWORD ; inptr2
294 lea rcx, [rcx+rcx*2] ; imul ecx, RGB_PIXELSIZE
295 cmp rcx, byte 2*SIZEOF_YMMWORD
296 jb short .column_st32
297 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
298 vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
299 add rdi, byte 2*SIZEOF_YMMWORD ; outptr
301 sub rcx, byte 2*SIZEOF_YMMWORD
302 jmp short .column_st31
304 cmp rcx, byte SIZEOF_YMMWORD
305 jb short .column_st31
306 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
307 add rdi, byte SIZEOF_YMMWORD ; outptr
309 sub rcx, byte SIZEOF_YMMWORD
310 jmp short .column_st31
312 cmp rcx, byte SIZEOF_XMMWORD
313 jb short .column_st15
314 vmovdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
315 add rdi, byte SIZEOF_XMMWORD ; outptr
316 vperm2i128 ymmA, ymmA, ymmA, 1
317 sub rcx, byte SIZEOF_XMMWORD
319 ; Store the lower 8 bytes of xmmA to the output when it has enough
321 cmp rcx, byte SIZEOF_MMWORD
323 vmovq XMM_MMWORD [rdi], xmmA
324 add rdi, byte SIZEOF_MMWORD
325 sub rcx, byte SIZEOF_MMWORD
326 vpsrldq xmmA, xmmA, SIZEOF_MMWORD
328 ; Store the lower 4 bytes of xmmA to the output when it has enough
330 cmp rcx, byte SIZEOF_DWORD
332 vmovd XMM_DWORD [rdi], xmmA
333 add rdi, byte SIZEOF_DWORD
334 sub rcx, byte SIZEOF_DWORD
335 vpsrldq xmmA, xmmA, SIZEOF_DWORD
337 ; Store the lower 2 bytes of rax to the output when it has enough
340 cmp rcx, byte SIZEOF_WORD
343 add rdi, byte SIZEOF_WORD
344 sub rcx, byte SIZEOF_WORD
347 ; Store the lower 1 byte of rax to the output when it has enough
353 %else ; RGB_PIXELSIZE == 4 ; -----------
355 %ifdef RGBX_FILLER_0XFF
356 vpcmpeqb ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********)
357 vpcmpeqb ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********)
359 vpxor ymm6, ymm6, ymm6 ; ymm6=XE=X(02468ACE********GIKMOQSU********)
360 vpxor ymm7, ymm7, ymm7 ; ymm7=XO=X(13579BDF********HJLNPRTV********)
362 ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
363 ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
364 ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
365 ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
366 ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
367 ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
368 ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
369 ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
371 vpunpcklbw ymmA, ymmA, ymmC ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
372 ; 0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
373 vpunpcklbw ymmE, ymmE, ymmG ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
374 ; 2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
375 vpunpcklbw ymmB, ymmB, ymmD ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
376 ; 0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
377 vpunpcklbw ymmF, ymmF, ymmH ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
378 ; 2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
380 vpunpckhwd ymmC, ymmA, ymmE ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
381 ; 0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
382 vpunpcklwd ymmA, ymmA, ymmE ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
383 ; 0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
384 vpunpckhwd ymmG, ymmB, ymmF ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
385 ; 0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
386 vpunpcklwd ymmB, ymmB, ymmF ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
387 ; 0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
389 vpunpckhdq ymmE, ymmA, ymmB ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
390 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
391 vpunpckldq ymmB, ymmA, ymmB ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
392 ; 0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
393 vpunpckhdq ymmF, ymmC, ymmG ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
394 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
395 vpunpckldq ymmG, ymmC, ymmG ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
396 ; 0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
398 vperm2i128 ymmA, ymmB, ymmE, 0x20 ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
399 ; 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
400 vperm2i128 ymmD, ymmG, ymmF, 0x20 ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
401 ; 0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
402 vperm2i128 ymmC, ymmB, ymmE, 0x31 ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
403 ; 0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
404 vperm2i128 ymmH, ymmG, ymmF, 0x31 ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
405 ; 0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
407 cmp rcx, byte SIZEOF_YMMWORD
408 jb short .column_st64
410 test rdi, SIZEOF_YMMWORD-1
412 ; --(aligned)-------------------
413 vmovntdq YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
414 vmovntdq YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
415 vmovntdq YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
416 vmovntdq YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
418 .out1: ; --(unaligned)-----------------
419 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
420 vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
421 vmovdqu YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
422 vmovdqu YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
424 add rdi, RGB_PIXELSIZE*SIZEOF_YMMWORD ; outptr
425 sub rcx, byte SIZEOF_YMMWORD
428 add rsi, byte SIZEOF_YMMWORD ; inptr0
432 add rbx, byte SIZEOF_YMMWORD ; inptr1
433 add rdx, byte SIZEOF_YMMWORD ; inptr2
437 cmp rcx, byte SIZEOF_YMMWORD/2
438 jb short .column_st32
439 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
440 vmovdqu YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
441 add rdi, byte 2*SIZEOF_YMMWORD ; outptr
444 sub rcx, byte SIZEOF_YMMWORD/2
446 cmp rcx, byte SIZEOF_YMMWORD/4
447 jb short .column_st16
448 vmovdqu YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
449 add rdi, byte SIZEOF_YMMWORD ; outptr
451 sub rcx, byte SIZEOF_YMMWORD/4
453 cmp rcx, byte SIZEOF_YMMWORD/8
454 jb short .column_st15
455 vmovdqu XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
456 add rdi, byte SIZEOF_XMMWORD ; outptr
457 vperm2i128 ymmA, ymmA, ymmA, 1
458 sub rcx, byte SIZEOF_YMMWORD/8
460 ; Store two pixels (8 bytes) of ymmA to the output when it has enough
462 cmp rcx, byte SIZEOF_YMMWORD/16
464 vmovq MMWORD [rdi], xmmA
465 add rdi, byte SIZEOF_YMMWORD/16*4
466 sub rcx, byte SIZEOF_YMMWORD/16
467 vpsrldq xmmA, SIZEOF_YMMWORD/16*4
469 ; Store one pixel (4 bytes) of ymmA to the output when it has enough
473 vmovd XMM_DWORD [rdi], xmmA
475 %endif ; RGB_PIXELSIZE ; ---------------
478 sfence ; flush the write buffer
489 ; --------------------------------------------------------------------------
491 ; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
494 ; jsimd_h2v2_merged_upsample_avx2(JDIMENSION output_width,
495 ; JSAMPIMAGE input_buf,
496 ; JDIMENSION in_row_group_ctr,
497 ; JSAMPARRAY output_buf);
500 ; r10d = JDIMENSION output_width
501 ; r11 = JSAMPIMAGE input_buf
502 ; r12d = JDIMENSION in_row_group_ctr
503 ; r13 = JSAMPARRAY output_buf
506 GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_avx2)
508 EXTN(jsimd_h2v2_merged_upsample_avx2):
518 mov rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
519 mov rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
520 mov rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
522 lea rsi, [rsi+rcx*SIZEOF_JSAMPROW]
524 sub rsp, SIZEOF_JSAMPARRAY*4
525 mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
526 mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
527 mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
546 call EXTN(jsimd_h2v1_merged_upsample_avx2)
551 mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
552 mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
553 mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
555 add rdi, byte SIZEOF_JSAMPROW ; outptr1
556 add rsi, byte SIZEOF_JSAMPROW ; inptr01
558 mov JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY], rsip ; intpr00
559 mov JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY], rbxp ; intpr1
560 mov JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY], rdxp ; intpr2
579 call EXTN(jsimd_h2v1_merged_upsample_avx2)
584 mov rsip, JSAMPARRAY [rsp+0*SIZEOF_JSAMPARRAY]
585 mov rbxp, JSAMPARRAY [rsp+1*SIZEOF_JSAMPARRAY]
586 mov rdxp, JSAMPARRAY [rsp+2*SIZEOF_JSAMPARRAY]
587 add rsp, SIZEOF_JSAMPARRAY*4
594 ; For some reason, the OS X linker does not honor the request to align the
595 ; segment unless we do this.