xref: /aosp_15_r20/external/libjpeg-turbo/simd/x86_64/jdcolext-avx2.asm (revision dfc6aa5c1cfd4bc4e2018dc74aa96e29ee49c6da) 
1;
2; jdcolext.asm - colorspace conversion (64-bit AVX2)
3;
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;
9; Based on the x86 SIMD extension for IJG JPEG library
10; Copyright (C) 1999-2006, MIYASAKA Masaru.
11; For conditions of distribution and use, see copyright notice in jsimdext.inc
12;
13; This file should be assembled with NASM (Netwide Assembler),
14; can *not* be assembled with Microsoft's MASM or any compatible
15; assembler (including Borland's Turbo Assembler).
16; NASM is available from http://nasm.sourceforge.net/ or
17; http://sourceforge.net/project/showfiles.php?group_id=6208
18
19%include "jcolsamp.inc"
20
21; --------------------------------------------------------------------------
22;
23; Convert some rows of samples to the output colorspace.
24;
25; GLOBAL(void)
26; jsimd_ycc_rgb_convert_avx2(JDIMENSION out_width, JSAMPIMAGE input_buf,
27;                            JDIMENSION input_row, JSAMPARRAY output_buf,
28;                            int num_rows)
29;
30
31; r10d = JDIMENSION out_width
32; r11 = JSAMPIMAGE input_buf
33; r12d = JDIMENSION input_row
34; r13 = JSAMPARRAY output_buf
35; r14d = int num_rows
36
37%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_YMMWORD  ; ymmword wk[WK_NUM]
38%define WK_NUM  2
39
40    align       32
41    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_avx2)
42
43EXTN(jsimd_ycc_rgb_convert_avx2):
44    push        rbp
45    mov         rax, rsp                     ; rax = original rbp
46    sub         rsp, byte 4
47    and         rsp, byte (-SIZEOF_YMMWORD)  ; align to 256 bits
48    mov         [rsp], rax
49    mov         rbp, rsp                     ; rbp = aligned rbp
50    lea         rsp, [wk(0)]
51    collect_args 5
52    push        rbx
53
54    mov         ecx, r10d               ; num_cols
55    test        rcx, rcx
56    jz          near .return
57
58    push        rcx
59
60    mov         rdi, r11
61    mov         ecx, r12d
62    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
63    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
64    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
65    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
66    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
67    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
68
69    pop         rcx
70
71    mov         rdi, r13
72    mov         eax, r14d
73    test        rax, rax
74    jle         near .return
75.rowloop:
76    push        rax
77    push        rdi
78    push        rdx
79    push        rbx
80    push        rsi
81    push        rcx                     ; col
82
83    mov         rsip, JSAMPROW [rsi]    ; inptr0
84    mov         rbxp, JSAMPROW [rbx]    ; inptr1
85    mov         rdxp, JSAMPROW [rdx]    ; inptr2
86    mov         rdip, JSAMPROW [rdi]    ; outptr
87.columnloop:
88
89    vmovdqu     ymm5, YMMWORD [rbx]     ; ymm5=Cb(0123456789ABCDEFGHIJKLMNOPQRSTUV)
90    vmovdqu     ymm1, YMMWORD [rdx]     ; ymm1=Cr(0123456789ABCDEFGHIJKLMNOPQRSTUV)
91
92    vpcmpeqw    ymm0, ymm0, ymm0
93    vpcmpeqw    ymm7, ymm7, ymm7
94    vpsrlw      ymm0, ymm0, BYTE_BIT    ; ymm0={0xFF 0x00 0xFF 0x00 ..}
95    vpsllw      ymm7, ymm7, 7           ; ymm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
96
97    vpand       ymm4, ymm0, ymm5        ; ymm4=Cb(02468ACEGIKMOQSU)=CbE
98    vpsrlw      ymm5, ymm5, BYTE_BIT    ; ymm5=Cb(13579BDFHJLNPRTV)=CbO
99    vpand       ymm0, ymm0, ymm1        ; ymm0=Cr(02468ACEGIKMOQSU)=CrE
100    vpsrlw      ymm1, ymm1, BYTE_BIT    ; ymm1=Cr(13579BDFHJLNPRTV)=CrO
101
102    vpaddw      ymm2, ymm4, ymm7
103    vpaddw      ymm3, ymm5, ymm7
104    vpaddw      ymm6, ymm0, ymm7
105    vpaddw      ymm7, ymm1, ymm7
106
107    ; (Original)
108    ; R = Y                + 1.40200 * Cr
109    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
110    ; B = Y + 1.77200 * Cb
111    ;
112    ; (This implementation)
113    ; R = Y                + 0.40200 * Cr + Cr
114    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
115    ; B = Y - 0.22800 * Cb + Cb + Cb
116
117    vpaddw      ymm4, ymm2, ymm2             ; ymm4=2*CbE
118    vpaddw      ymm5, ymm3, ymm3             ; ymm5=2*CbO
119    vpaddw      ymm0, ymm6, ymm6             ; ymm0=2*CrE
120    vpaddw      ymm1, ymm7, ymm7             ; ymm1=2*CrO
121
122    vpmulhw     ymm4, ymm4, [rel PW_MF0228]  ; ymm4=(2*CbE * -FIX(0.22800))
123    vpmulhw     ymm5, ymm5, [rel PW_MF0228]  ; ymm5=(2*CbO * -FIX(0.22800))
124    vpmulhw     ymm0, ymm0, [rel PW_F0402]   ; ymm0=(2*CrE * FIX(0.40200))
125    vpmulhw     ymm1, ymm1, [rel PW_F0402]   ; ymm1=(2*CrO * FIX(0.40200))
126
127    vpaddw      ymm4, ymm4, [rel PW_ONE]
128    vpaddw      ymm5, ymm5, [rel PW_ONE]
129    vpsraw      ymm4, ymm4, 1                ; ymm4=(CbE * -FIX(0.22800))
130    vpsraw      ymm5, ymm5, 1                ; ymm5=(CbO * -FIX(0.22800))
131    vpaddw      ymm0, ymm0, [rel PW_ONE]
132    vpaddw      ymm1, ymm1, [rel PW_ONE]
133    vpsraw      ymm0, ymm0, 1                ; ymm0=(CrE * FIX(0.40200))
134    vpsraw      ymm1, ymm1, 1                ; ymm1=(CrO * FIX(0.40200))
135
136    vpaddw      ymm4, ymm4, ymm2
137    vpaddw      ymm5, ymm5, ymm3
138    vpaddw      ymm4, ymm4, ymm2             ; ymm4=(CbE * FIX(1.77200))=(B-Y)E
139    vpaddw      ymm5, ymm5, ymm3             ; ymm5=(CbO * FIX(1.77200))=(B-Y)O
140    vpaddw      ymm0, ymm0, ymm6             ; ymm0=(CrE * FIX(1.40200))=(R-Y)E
141    vpaddw      ymm1, ymm1, ymm7             ; ymm1=(CrO * FIX(1.40200))=(R-Y)O
142
143    vmovdqa     YMMWORD [wk(0)], ymm4        ; wk(0)=(B-Y)E
144    vmovdqa     YMMWORD [wk(1)], ymm5        ; wk(1)=(B-Y)O
145
146    vpunpckhwd  ymm4, ymm2, ymm6
147    vpunpcklwd  ymm2, ymm2, ymm6
148    vpmaddwd    ymm2, ymm2, [rel PW_MF0344_F0285]
149    vpmaddwd    ymm4, ymm4, [rel PW_MF0344_F0285]
150    vpunpckhwd  ymm5, ymm3, ymm7
151    vpunpcklwd  ymm3, ymm3, ymm7
152    vpmaddwd    ymm3, ymm3, [rel PW_MF0344_F0285]
153    vpmaddwd    ymm5, ymm5, [rel PW_MF0344_F0285]
154
155    vpaddd      ymm2, ymm2, [rel PD_ONEHALF]
156    vpaddd      ymm4, ymm4, [rel PD_ONEHALF]
157    vpsrad      ymm2, ymm2, SCALEBITS
158    vpsrad      ymm4, ymm4, SCALEBITS
159    vpaddd      ymm3, ymm3, [rel PD_ONEHALF]
160    vpaddd      ymm5, ymm5, [rel PD_ONEHALF]
161    vpsrad      ymm3, ymm3, SCALEBITS
162    vpsrad      ymm5, ymm5, SCALEBITS
163
164    vpackssdw   ymm2, ymm2, ymm4             ; ymm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
165    vpackssdw   ymm3, ymm3, ymm5             ; ymm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
166    vpsubw      ymm2, ymm2, ymm6             ; ymm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
167    vpsubw      ymm3, ymm3, ymm7             ; ymm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
168
169    vmovdqu     ymm5, YMMWORD [rsi]          ; ymm5=Y(0123456789ABCDEFGHIJKLMNOPQRSTUV)
170
171    vpcmpeqw    ymm4, ymm4, ymm4
172    vpsrlw      ymm4, ymm4, BYTE_BIT         ; ymm4={0xFF 0x00 0xFF 0x00 ..}
173    vpand       ymm4, ymm4, ymm5             ; ymm4=Y(02468ACEGIKMOQSU)=YE
174    vpsrlw      ymm5, ymm5, BYTE_BIT         ; ymm5=Y(13579BDFHJLNPRTV)=YO
175
176    vpaddw      ymm0, ymm0, ymm4             ; ymm0=((R-Y)E+YE)=RE=R(02468ACEGIKMOQSU)
177    vpaddw      ymm1, ymm1, ymm5             ; ymm1=((R-Y)O+YO)=RO=R(13579BDFHJLNPRTV)
178    vpackuswb   ymm0, ymm0, ymm0             ; ymm0=R(02468ACE********GIKMOQSU********)
179    vpackuswb   ymm1, ymm1, ymm1             ; ymm1=R(13579BDF********HJLNPRTV********)
180
181    vpaddw      ymm2, ymm2, ymm4             ; ymm2=((G-Y)E+YE)=GE=G(02468ACEGIKMOQSU)
182    vpaddw      ymm3, ymm3, ymm5             ; ymm3=((G-Y)O+YO)=GO=G(13579BDFHJLNPRTV)
183    vpackuswb   ymm2, ymm2, ymm2             ; ymm2=G(02468ACE********GIKMOQSU********)
184    vpackuswb   ymm3, ymm3, ymm3             ; ymm3=G(13579BDF********HJLNPRTV********)
185
186    vpaddw      ymm4, ymm4, YMMWORD [wk(0)]  ; ymm4=(YE+(B-Y)E)=BE=B(02468ACEGIKMOQSU)
187    vpaddw      ymm5, ymm5, YMMWORD [wk(1)]  ; ymm5=(YO+(B-Y)O)=BO=B(13579BDFHJLNPRTV)
188    vpackuswb   ymm4, ymm4, ymm4             ; ymm4=B(02468ACE********GIKMOQSU********)
189    vpackuswb   ymm5, ymm5, ymm5             ; ymm5=B(13579BDF********HJLNPRTV********)
190
191%if RGB_PIXELSIZE == 3  ; ---------------
192
193    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
194    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
195    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
196    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
197    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
198    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
199    ; ymmG=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
200    ; ymmH=(** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** ** **)
201
202    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
203                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
204    vpunpcklbw  ymmE, ymmE, ymmB        ; ymmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F
205                                        ;       2G 0H 2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V)
206    vpunpcklbw  ymmD, ymmD, ymmF        ; ymmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F
207                                        ;       1H 2H 1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V)
208
209    vpsrldq     ymmH, ymmA, 2           ; ymmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E 0G 1G
210                                        ;       0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U -- --)
211    vpunpckhwd  ymmG, ymmA, ymmE        ; ymmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F
212                                        ;       0O 1O 2O 0P 0Q 1Q 2Q 0R 0S 1S 2S 0T 0U 1U 2U 0V)
213    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07
214                                        ;       0G 1G 2G 0H 0I 1I 2I 0J 0K 1K 2K 0L 0M 1M 2M 0N)
215
216    vpsrldq     ymmE, ymmE, 2           ; ymmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F 2G 0H
217                                        ;       2I 0J 2K 0L 2M 0N 2O 0P 2Q 0R 2S 0T 2U 0V -- --)
218
219    vpsrldq     ymmB, ymmD, 2           ; ymmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F 1H 2H
220                                        ;       1J 2J 1L 2L 1N 2N 1P 2P 1R 2R 1T 2T 1V 2V -- --)
221    vpunpckhwd  ymmC, ymmD, ymmH        ; ymmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F 0G 1G
222                                        ;       1P 2P 0Q 1Q 1R 2R 0S 1S 1T 2T 0U 1U 1V 2V -- --)
223    vpunpcklwd  ymmD, ymmD, ymmH        ; ymmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18
224                                        ;       1H 2H 0I 1I 1J 2J 0K 1K 1L 2L 0M 1M 1N 2N 0O 1O)
225
226    vpunpckhwd  ymmF, ymmE, ymmB        ; ymmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F 2G 0H 1H 2H
227                                        ;       2Q 0R 1R 2R 2S 0T 1T 2T 2U 0V 1V 2V -- -- -- --)
228    vpunpcklwd  ymmE, ymmE, ymmB        ; ymmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29
229                                        ;       2I 0J 1J 2J 2K 0L 1L 2L 2M 0N 1N 2N 2O 0P 1P 2P)
230
231    vpshufd     ymmH, ymmA, 0x4E        ; ymmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03
232                                        ;       0K 1K 2K 0L 0M 1M 2M 0N 0G 1G 2G 0H 0I 1I 2I 0J)
233    vpunpckldq  ymmA, ymmA, ymmD        ; ymmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14
234                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 0I 1I 2I 0J 1J 2J 0K 1K)
235    vpunpckhdq  ymmD, ymmD, ymmE        ; ymmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29
236                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 1N 2N 0O 1O 2O 0P 1P 2P)
237    vpunpckldq  ymmE, ymmE, ymmH        ; ymmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07
238                                        ;       2I 0J 1J 2J 0K 1K 2K 0L 2K 0L 1L 2L 0M 1M 2M 0N)
239
240    vpshufd     ymmH, ymmG, 0x4E        ; ymmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B
241                                        ;       0S 1S 2S 0T 0U 1U 2U 0V 0O 1O 2O 0P 0Q 1Q 2Q 0R)
242    vpunpckldq  ymmG, ymmG, ymmC        ; ymmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C
243                                        ;       0O 1O 2O 0P 1P 2P 0Q 1Q 0Q 1Q 2Q 0R 1R 2R 0S 1S)
244    vpunpckhdq  ymmC, ymmC, ymmF        ; ymmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F 0G 1G 2G 0H 1H 2H
245                                        ;       1T 2T 0U 1U 2U 0V 1V 2V 1V 2V -- -- -- -- -- --)
246    vpunpckldq  ymmF, ymmF, ymmH        ; ymmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F
247                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 2S 0T 1T 2T 0U 1U 2U 0V)
248
249    vpunpcklqdq ymmH, ymmA, ymmE        ; ymmH=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
250                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
251    vpunpcklqdq ymmG, ymmD, ymmG        ; ymmG=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A
252                                        ;       1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q)
253    vpunpcklqdq ymmC, ymmF, ymmC        ; ymmC=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
254                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
255
256    vperm2i128  ymmA, ymmH, ymmG, 0x20  ; ymmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05
257                                        ;       15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
258    vperm2i128  ymmD, ymmC, ymmH, 0x30  ; ymmD=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F
259                                        ;       0G 1G 2G 0H 1H 2H 0I 1I 2I 0J 1J 2J 0K 1K 2K 0L)
260    vperm2i128  ymmF, ymmG, ymmC, 0x31  ; ymmF=(1L 2L 0M 1M 2M 0N 1N 2N 0O 1O 2O 0P 1P 2P 0Q 1Q
261                                        ;       2Q 0R 1R 2R 0S 1S 2S 0T 1T 2T 0U 1U 2U 0V 1V 2V)
262
263    cmp         rcx, byte SIZEOF_YMMWORD
264    jb          short .column_st64
265
266    test        rdi, SIZEOF_YMMWORD-1
267    jnz         short .out1
268    ; --(aligned)-------------------
269    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
270    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
271    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
272    jmp         short .out0
273.out1:  ; --(unaligned)-----------------
274    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
275    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
276    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmF
277.out0:
278    add         rdi, byte RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
279    sub         rcx, byte SIZEOF_YMMWORD
280    jz          near .nextrow
281
282    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
283    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
284    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
285    jmp         near .columnloop
286
287.column_st64:
288    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
289    cmp         rcx, byte 2*SIZEOF_YMMWORD
290    jb          short .column_st32
291    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
292    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
293    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
294    vmovdqa     ymmA, ymmF
295    sub         rcx, byte 2*SIZEOF_YMMWORD
296    jmp         short .column_st31
297.column_st32:
298    cmp         rcx, byte SIZEOF_YMMWORD
299    jb          short .column_st31
300    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
301    add         rdi, byte SIZEOF_YMMWORD    ; outptr
302    vmovdqa     ymmA, ymmD
303    sub         rcx, byte SIZEOF_YMMWORD
304    jmp         short .column_st31
305.column_st31:
306    cmp         rcx, byte SIZEOF_XMMWORD
307    jb          short .column_st15
308    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
309    add         rdi, byte SIZEOF_XMMWORD    ; outptr
310    vperm2i128  ymmA, ymmA, ymmA, 1
311    sub         rcx, byte SIZEOF_XMMWORD
312.column_st15:
313    ; Store the lower 8 bytes of xmmA to the output when it has enough
314    ; space.
315    cmp         rcx, byte SIZEOF_MMWORD
316    jb          short .column_st7
317    vmovq       XMM_MMWORD [rdi], xmmA
318    add         rdi, byte SIZEOF_MMWORD
319    sub         rcx, byte SIZEOF_MMWORD
320    vpsrldq     xmmA, xmmA, SIZEOF_MMWORD
321.column_st7:
322    ; Store the lower 4 bytes of xmmA to the output when it has enough
323    ; space.
324    cmp         rcx, byte SIZEOF_DWORD
325    jb          short .column_st3
326    vmovd       XMM_DWORD [rdi], xmmA
327    add         rdi, byte SIZEOF_DWORD
328    sub         rcx, byte SIZEOF_DWORD
329    vpsrldq     xmmA, xmmA, SIZEOF_DWORD
330.column_st3:
331    ; Store the lower 2 bytes of rax to the output when it has enough
332    ; space.
333    vmovd       eax, xmmA
334    cmp         rcx, byte SIZEOF_WORD
335    jb          short .column_st1
336    mov         word [rdi], ax
337    add         rdi, byte SIZEOF_WORD
338    sub         rcx, byte SIZEOF_WORD
339    shr         rax, 16
340.column_st1:
341    ; Store the lower 1 byte of rax to the output when it has enough
342    ; space.
343    test        rcx, rcx
344    jz          short .nextrow
345    mov         byte [rdi], al
346
347%else  ; RGB_PIXELSIZE == 4 ; -----------
348
349%ifdef RGBX_FILLER_0XFF
350    vpcmpeqb    ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
351    vpcmpeqb    ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
352%else
353    vpxor       ymm6, ymm6, ymm6        ; ymm6=XE=X(02468ACE********GIKMOQSU********)
354    vpxor       ymm7, ymm7, ymm7        ; ymm7=XO=X(13579BDF********HJLNPRTV********)
355%endif
356    ; ymmA=(00 02 04 06 08 0A 0C 0E ** 0G 0I 0K 0M 0O 0Q 0S 0U **)
357    ; ymmB=(01 03 05 07 09 0B 0D 0F ** 0H 0J 0L 0N 0P 0R 0T 0V **)
358    ; ymmC=(10 12 14 16 18 1A 1C 1E ** 1G 1I 1K 1M 1O 1Q 1S 1U **)
359    ; ymmD=(11 13 15 17 19 1B 1D 1F ** 1H 1J 1L 1N 1P 1R 1T 1V **)
360    ; ymmE=(20 22 24 26 28 2A 2C 2E ** 2G 2I 2K 2M 2O 2Q 2S 2U **)
361    ; ymmF=(21 23 25 27 29 2B 2D 2F ** 2H 2J 2L 2N 2P 2R 2T 2V **)
362    ; ymmG=(30 32 34 36 38 3A 3C 3E ** 3G 3I 3K 3M 3O 3Q 3S 3U **)
363    ; ymmH=(31 33 35 37 39 3B 3D 3F ** 3H 3J 3L 3N 3P 3R 3T 3V **)
364
365    vpunpcklbw  ymmA, ymmA, ymmC        ; ymmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E
366                                        ;       0G 1G 0I 1I 0K 1K 0M 1M 0O 1O 0Q 1Q 0S 1S 0U 1U)
367    vpunpcklbw  ymmE, ymmE, ymmG        ; ymmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E
368                                        ;       2G 3G 2I 3I 2K 3K 2M 3M 2O 3O 2Q 3Q 2S 3S 2U 3U)
369    vpunpcklbw  ymmB, ymmB, ymmD        ; ymmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F
370                                        ;       0H 1H 0J 1J 0L 1L 0N 1N 0P 1P 0R 1R 0T 1T 0V 1V)
371    vpunpcklbw  ymmF, ymmF, ymmH        ; ymmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F
372                                        ;       2H 3H 2J 3J 2L 3L 2N 3N 2P 3P 2R 3R 2T 3T 2V 3V)
373
374    vpunpckhwd  ymmC, ymmA, ymmE        ; ymmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E
375                                        ;       0O 1O 2O 3O 0Q 1Q 2Q 3Q 0S 1S 2S 3S 0U 1U 2U 3U)
376    vpunpcklwd  ymmA, ymmA, ymmE        ; ymmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36
377                                        ;       0G 1G 2G 3G 0I 1I 2I 3I 0K 1K 2K 3K 0M 1M 2M 3M)
378    vpunpckhwd  ymmG, ymmB, ymmF        ; ymmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F
379                                        ;       0P 1P 2P 3P 0R 1R 2R 3R 0T 1T 2T 3T 0V 1V 2V 3V)
380    vpunpcklwd  ymmB, ymmB, ymmF        ; ymmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37
381                                        ;       0H 1H 2H 3H 0J 1J 2J 3J 0L 1L 2L 3L 0N 1N 2N 3N)
382
383    vpunpckhdq  ymmE, ymmA, ymmB        ; ymmE=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
384                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
385    vpunpckldq  ymmB, ymmA, ymmB        ; ymmB=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
386                                        ;       0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J)
387    vpunpckhdq  ymmF, ymmC, ymmG        ; ymmF=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F
388                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
389    vpunpckldq  ymmG, ymmC, ymmG        ; ymmG=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
390                                        ;       0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R)
391
392    vperm2i128  ymmA, ymmB, ymmE, 0x20  ; ymmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
393                                        ;       04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
394    vperm2i128  ymmD, ymmG, ymmF, 0x20  ; ymmD=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B
395                                        ;       0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
396    vperm2i128  ymmC, ymmB, ymmE, 0x31  ; ymmC=(0G 1G 2G 3G 0H 1H 2H 3H 0I 1I 2I 3I 0J 1J 2J 3J
397                                        ;       0K 1K 2K 3K 0L 1L 2L 3L 0M 1M 2M 3M 0N 1N 2N 3N)
398    vperm2i128  ymmH, ymmG, ymmF, 0x31  ; ymmH=(0O 1O 2O 3O 0P 1P 2P 3P 0Q 1Q 2Q 3Q 0R 1R 2R 3R
399                                        ;       0S 1S 2S 3S 0T 1T 2T 3T 0U 1U 2U 3U 0V 1V 2V 3V)
400
401    cmp         rcx, byte SIZEOF_YMMWORD
402    jb          short .column_st64
403
404    test        rdi, SIZEOF_YMMWORD-1
405    jnz         short .out1
406    ; --(aligned)-------------------
407    vmovntdq    YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
408    vmovntdq    YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
409    vmovntdq    YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
410    vmovntdq    YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
411    jmp         short .out0
412.out1:  ; --(unaligned)-----------------
413    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
414    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
415    vmovdqu     YMMWORD [rdi+2*SIZEOF_YMMWORD], ymmC
416    vmovdqu     YMMWORD [rdi+3*SIZEOF_YMMWORD], ymmH
417.out0:
418    add         rdi, RGB_PIXELSIZE*SIZEOF_YMMWORD  ; outptr
419    sub         rcx, byte SIZEOF_YMMWORD
420    jz          near .nextrow
421
422    add         rsi, byte SIZEOF_YMMWORD  ; inptr0
423    add         rbx, byte SIZEOF_YMMWORD  ; inptr1
424    add         rdx, byte SIZEOF_YMMWORD  ; inptr2
425    jmp         near .columnloop
426
427.column_st64:
428    cmp         rcx, byte SIZEOF_YMMWORD/2
429    jb          short .column_st32
430    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
431    vmovdqu     YMMWORD [rdi+1*SIZEOF_YMMWORD], ymmD
432    add         rdi, byte 2*SIZEOF_YMMWORD  ; outptr
433    vmovdqa     ymmA, ymmC
434    vmovdqa     ymmD, ymmH
435    sub         rcx, byte SIZEOF_YMMWORD/2
436.column_st32:
437    cmp         rcx, byte SIZEOF_YMMWORD/4
438    jb          short .column_st16
439    vmovdqu     YMMWORD [rdi+0*SIZEOF_YMMWORD], ymmA
440    add         rdi, byte SIZEOF_YMMWORD    ; outptr
441    vmovdqa     ymmA, ymmD
442    sub         rcx, byte SIZEOF_YMMWORD/4
443.column_st16:
444    cmp         rcx, byte SIZEOF_YMMWORD/8
445    jb          short .column_st15
446    vmovdqu     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
447    vperm2i128  ymmA, ymmA, ymmA, 1
448    add         rdi, byte SIZEOF_XMMWORD    ; outptr
449    sub         rcx, byte SIZEOF_YMMWORD/8
450.column_st15:
451    ; Store two pixels (8 bytes) of ymmA to the output when it has enough
452    ; space.
453    cmp         rcx, byte SIZEOF_YMMWORD/16
454    jb          short .column_st7
455    vmovq       MMWORD [rdi], xmmA
456    add         rdi, byte SIZEOF_YMMWORD/16*4
457    sub         rcx, byte SIZEOF_YMMWORD/16
458    vpsrldq     xmmA, SIZEOF_YMMWORD/16*4
459.column_st7:
460    ; Store one pixel (4 bytes) of ymmA to the output when it has enough
461    ; space.
462    test        rcx, rcx
463    jz          short .nextrow
464    vmovd       XMM_DWORD [rdi], xmmA
465
466%endif  ; RGB_PIXELSIZE ; ---------------
467
468.nextrow:
469    pop         rcx
470    pop         rsi
471    pop         rbx
472    pop         rdx
473    pop         rdi
474    pop         rax
475
476    add         rsi, byte SIZEOF_JSAMPROW
477    add         rbx, byte SIZEOF_JSAMPROW
478    add         rdx, byte SIZEOF_JSAMPROW
479    add         rdi, byte SIZEOF_JSAMPROW  ; output_buf
480    dec         rax                        ; num_rows
481    jg          near .rowloop
482
483    sfence                              ; flush the write buffer
484
485.return:
486    pop         rbx
487    vzeroupper
488    uncollect_args 5
489    mov         rsp, rbp                ; rsp <- aligned rbp
490    pop         rsp                     ; rsp <- original rbp
491    pop         rbp
492    ret
493
494; For some reason, the OS X linker does not honor the request to align the
495; segment unless we do this.
496    align       32
497