• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jdcolext.asm - colorspace conversion (64-bit SSE2)
3;
4; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 2009, 2012, D. R. Commander.
6;
7; Based on the x86 SIMD extension for IJG JPEG library
8; Copyright (C) 1999-2006, MIYASAKA Masaru.
9; For conditions of distribution and use, see copyright notice in jsimdext.inc
10;
11; This file should be assembled with NASM (Netwide Assembler),
12; can *not* be assembled with Microsoft's MASM or any compatible
13; assembler (including Borland's Turbo Assembler).
14; NASM is available from http://nasm.sourceforge.net/ or
15; http://sourceforge.net/project/showfiles.php?group_id=6208
16;
17; [TAB8]
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_sse2 (JDIMENSION out_width,
27;                             JSAMPIMAGE input_buf, JDIMENSION input_row,
28;                             JSAMPARRAY output_buf, int num_rows)
29;
30
31; r10 = JDIMENSION out_width
32; r11 = JSAMPIMAGE input_buf
33; r12 = JDIMENSION input_row
34; r13 = JSAMPARRAY output_buf
35; r14 = int num_rows
36
37%define wk(i)           rbp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
38%define WK_NUM          2
39
40        align   16
41        global  EXTN(jsimd_ycc_rgb_convert_sse2)
42
43EXTN(jsimd_ycc_rgb_convert_sse2):
44        push    rbp
45        mov     rax,rsp                         ; rax = original rbp
46        sub     rsp, byte 4
47        and     rsp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
48        mov     [rsp],rax
49        mov     rbp,rsp                         ; rbp = aligned rbp
50        lea     rsp, [wk(0)]
51        collect_args
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     rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
63        mov     rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
64        mov     rdx, 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     rsi, JSAMPROW [rsi]     ; inptr0
84        mov     rbx, JSAMPROW [rbx]     ; inptr1
85        mov     rdx, JSAMPROW [rdx]     ; inptr2
86        mov     rdi, JSAMPROW [rdi]     ; outptr
87.columnloop:
88
89        movdqa  xmm5, XMMWORD [rbx]     ; xmm5=Cb(0123456789ABCDEF)
90        movdqa  xmm1, XMMWORD [rdx]     ; xmm1=Cr(0123456789ABCDEF)
91
92        pcmpeqw xmm4,xmm4
93        pcmpeqw xmm7,xmm7
94        psrlw   xmm4,BYTE_BIT
95        psllw   xmm7,7                  ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
96        movdqa  xmm0,xmm4               ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
97
98        pand    xmm4,xmm5               ; xmm4=Cb(02468ACE)=CbE
99        psrlw   xmm5,BYTE_BIT           ; xmm5=Cb(13579BDF)=CbO
100        pand    xmm0,xmm1               ; xmm0=Cr(02468ACE)=CrE
101        psrlw   xmm1,BYTE_BIT           ; xmm1=Cr(13579BDF)=CrO
102
103        paddw   xmm4,xmm7
104        paddw   xmm5,xmm7
105        paddw   xmm0,xmm7
106        paddw   xmm1,xmm7
107
108        ; (Original)
109        ; R = Y                + 1.40200 * Cr
110        ; G = Y - 0.34414 * Cb - 0.71414 * Cr
111        ; B = Y + 1.77200 * Cb
112        ;
113        ; (This implementation)
114        ; R = Y                + 0.40200 * Cr + Cr
115        ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
116        ; B = Y - 0.22800 * Cb + Cb + Cb
117
118        movdqa  xmm2,xmm4               ; xmm2=CbE
119        movdqa  xmm3,xmm5               ; xmm3=CbO
120        paddw   xmm4,xmm4               ; xmm4=2*CbE
121        paddw   xmm5,xmm5               ; xmm5=2*CbO
122        movdqa  xmm6,xmm0               ; xmm6=CrE
123        movdqa  xmm7,xmm1               ; xmm7=CrO
124        paddw   xmm0,xmm0               ; xmm0=2*CrE
125        paddw   xmm1,xmm1               ; xmm1=2*CrO
126
127        pmulhw  xmm4,[rel PW_MF0228]    ; xmm4=(2*CbE * -FIX(0.22800))
128        pmulhw  xmm5,[rel PW_MF0228]    ; xmm5=(2*CbO * -FIX(0.22800))
129        pmulhw  xmm0,[rel PW_F0402]     ; xmm0=(2*CrE * FIX(0.40200))
130        pmulhw  xmm1,[rel PW_F0402]     ; xmm1=(2*CrO * FIX(0.40200))
131
132        paddw   xmm4,[rel PW_ONE]
133        paddw   xmm5,[rel PW_ONE]
134        psraw   xmm4,1                  ; xmm4=(CbE * -FIX(0.22800))
135        psraw   xmm5,1                  ; xmm5=(CbO * -FIX(0.22800))
136        paddw   xmm0,[rel PW_ONE]
137        paddw   xmm1,[rel PW_ONE]
138        psraw   xmm0,1                  ; xmm0=(CrE * FIX(0.40200))
139        psraw   xmm1,1                  ; xmm1=(CrO * FIX(0.40200))
140
141        paddw   xmm4,xmm2
142        paddw   xmm5,xmm3
143        paddw   xmm4,xmm2               ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
144        paddw   xmm5,xmm3               ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
145        paddw   xmm0,xmm6               ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
146        paddw   xmm1,xmm7               ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
147
148        movdqa  XMMWORD [wk(0)], xmm4   ; wk(0)=(B-Y)E
149        movdqa  XMMWORD [wk(1)], xmm5   ; wk(1)=(B-Y)O
150
151        movdqa    xmm4,xmm2
152        movdqa    xmm5,xmm3
153        punpcklwd xmm2,xmm6
154        punpckhwd xmm4,xmm6
155        pmaddwd   xmm2,[rel PW_MF0344_F0285]
156        pmaddwd   xmm4,[rel PW_MF0344_F0285]
157        punpcklwd xmm3,xmm7
158        punpckhwd xmm5,xmm7
159        pmaddwd   xmm3,[rel PW_MF0344_F0285]
160        pmaddwd   xmm5,[rel PW_MF0344_F0285]
161
162        paddd     xmm2,[rel PD_ONEHALF]
163        paddd     xmm4,[rel PD_ONEHALF]
164        psrad     xmm2,SCALEBITS
165        psrad     xmm4,SCALEBITS
166        paddd     xmm3,[rel PD_ONEHALF]
167        paddd     xmm5,[rel PD_ONEHALF]
168        psrad     xmm3,SCALEBITS
169        psrad     xmm5,SCALEBITS
170
171        packssdw  xmm2,xmm4     ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
172        packssdw  xmm3,xmm5     ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
173        psubw     xmm2,xmm6     ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
174        psubw     xmm3,xmm7     ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
175
176        movdqa    xmm5, XMMWORD [rsi]   ; xmm5=Y(0123456789ABCDEF)
177
178        pcmpeqw   xmm4,xmm4
179        psrlw     xmm4,BYTE_BIT         ; xmm4={0xFF 0x00 0xFF 0x00 ..}
180        pand      xmm4,xmm5             ; xmm4=Y(02468ACE)=YE
181        psrlw     xmm5,BYTE_BIT         ; xmm5=Y(13579BDF)=YO
182
183        paddw     xmm0,xmm4             ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
184        paddw     xmm1,xmm5             ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
185        packuswb  xmm0,xmm0             ; xmm0=R(02468ACE********)
186        packuswb  xmm1,xmm1             ; xmm1=R(13579BDF********)
187
188        paddw     xmm2,xmm4             ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
189        paddw     xmm3,xmm5             ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
190        packuswb  xmm2,xmm2             ; xmm2=G(02468ACE********)
191        packuswb  xmm3,xmm3             ; xmm3=G(13579BDF********)
192
193        paddw     xmm4, XMMWORD [wk(0)] ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
194        paddw     xmm5, XMMWORD [wk(1)] ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
195        packuswb  xmm4,xmm4             ; xmm4=B(02468ACE********)
196        packuswb  xmm5,xmm5             ; xmm5=B(13579BDF********)
197
198%if RGB_PIXELSIZE == 3 ; ---------------
199
200        ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
201        ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
202        ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
203        ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
204
205        punpcklbw xmmA,xmmC     ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
206        punpcklbw xmmE,xmmB     ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
207        punpcklbw xmmD,xmmF     ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
208
209        movdqa    xmmG,xmmA
210        movdqa    xmmH,xmmA
211        punpcklwd xmmA,xmmE     ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
212        punpckhwd xmmG,xmmE     ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
213
214        psrldq    xmmH,2        ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
215        psrldq    xmmE,2        ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
216
217        movdqa    xmmC,xmmD
218        movdqa    xmmB,xmmD
219        punpcklwd xmmD,xmmH     ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
220        punpckhwd xmmC,xmmH     ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
221
222        psrldq    xmmB,2        ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
223
224        movdqa    xmmF,xmmE
225        punpcklwd xmmE,xmmB     ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
226        punpckhwd xmmF,xmmB     ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
227
228        pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
229        movdqa    xmmB,xmmE
230        punpckldq xmmA,xmmD     ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
231        punpckldq xmmE,xmmH     ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
232        punpckhdq xmmD,xmmB     ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
233
234        pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
235        movdqa    xmmB,xmmF
236        punpckldq xmmG,xmmC     ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
237        punpckldq xmmF,xmmH     ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
238        punpckhdq xmmC,xmmB     ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
239
240        punpcklqdq xmmA,xmmE    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
241        punpcklqdq xmmD,xmmG    ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
242        punpcklqdq xmmF,xmmC    ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
243
244        cmp     rcx, byte SIZEOF_XMMWORD
245        jb      short .column_st32
246
247        test    rdi, SIZEOF_XMMWORD-1
248        jnz     short .out1
249        ; --(aligned)-------------------
250        movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
251        movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
252        movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
253        jmp     short .out0
254.out1:  ; --(unaligned)-----------------
255        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
256        movdqu  XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
257        movdqu  XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
258.out0:
259        add     rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
260        sub     rcx, byte SIZEOF_XMMWORD
261        jz      near .nextrow
262
263        add     rsi, byte SIZEOF_XMMWORD        ; inptr0
264        add     rbx, byte SIZEOF_XMMWORD        ; inptr1
265        add     rdx, byte SIZEOF_XMMWORD        ; inptr2
266        jmp     near .columnloop
267
268.column_st32:
269        lea     rcx, [rcx+rcx*2]                ; imul ecx, RGB_PIXELSIZE
270        cmp     rcx, byte 2*SIZEOF_XMMWORD
271        jb      short .column_st16
272        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
273        movdqu  XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
274        add     rdi, byte 2*SIZEOF_XMMWORD      ; outptr
275        movdqa  xmmA,xmmF
276        sub     rcx, byte 2*SIZEOF_XMMWORD
277        jmp     short .column_st15
278.column_st16:
279        cmp     rcx, byte SIZEOF_XMMWORD
280        jb      short .column_st15
281        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
282        add     rdi, byte SIZEOF_XMMWORD        ; outptr
283        movdqa  xmmA,xmmD
284        sub     rcx, byte SIZEOF_XMMWORD
285.column_st15:
286        ; Store the lower 8 bytes of xmmA to the output when it has enough
287        ; space.
288        cmp     rcx, byte SIZEOF_MMWORD
289        jb      short .column_st7
290        movq    XMM_MMWORD [rdi], xmmA
291        add     rdi, byte SIZEOF_MMWORD
292        sub     rcx, byte SIZEOF_MMWORD
293        psrldq  xmmA, SIZEOF_MMWORD
294.column_st7:
295        ; Store the lower 4 bytes of xmmA to the output when it has enough
296        ; space.
297        cmp     rcx, byte SIZEOF_DWORD
298        jb      short .column_st3
299        movd    XMM_DWORD [rdi], xmmA
300        add     rdi, byte SIZEOF_DWORD
301        sub     rcx, byte SIZEOF_DWORD
302        psrldq  xmmA, SIZEOF_DWORD
303.column_st3:
304        ; Store the lower 2 bytes of rax to the output when it has enough
305        ; space.
306        movd    eax, xmmA
307        cmp     rcx, byte SIZEOF_WORD
308        jb      short .column_st1
309        mov     WORD [rdi], ax
310        add     rdi, byte SIZEOF_WORD
311        sub     rcx, byte SIZEOF_WORD
312        shr     rax, 16
313.column_st1:
314        ; Store the lower 1 byte of rax to the output when it has enough
315        ; space.
316        test    rcx, rcx
317        jz      short .nextrow
318        mov     BYTE [rdi], al
319
320%else ; RGB_PIXELSIZE == 4 ; -----------
321
322%ifdef RGBX_FILLER_0XFF
323        pcmpeqb   xmm6,xmm6             ; xmm6=XE=X(02468ACE********)
324        pcmpeqb   xmm7,xmm7             ; xmm7=XO=X(13579BDF********)
325%else
326        pxor      xmm6,xmm6             ; xmm6=XE=X(02468ACE********)
327        pxor      xmm7,xmm7             ; xmm7=XO=X(13579BDF********)
328%endif
329        ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
330        ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
331        ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
332        ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
333
334        punpcklbw xmmA,xmmC     ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
335        punpcklbw xmmE,xmmG     ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
336        punpcklbw xmmB,xmmD     ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
337        punpcklbw xmmF,xmmH     ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
338
339        movdqa    xmmC,xmmA
340        punpcklwd xmmA,xmmE     ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
341        punpckhwd xmmC,xmmE     ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
342        movdqa    xmmG,xmmB
343        punpcklwd xmmB,xmmF     ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
344        punpckhwd xmmG,xmmF     ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
345
346        movdqa    xmmD,xmmA
347        punpckldq xmmA,xmmB     ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
348        punpckhdq xmmD,xmmB     ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
349        movdqa    xmmH,xmmC
350        punpckldq xmmC,xmmG     ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
351        punpckhdq xmmH,xmmG     ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
352
353        cmp     rcx, byte SIZEOF_XMMWORD
354        jb      short .column_st32
355
356        test    rdi, SIZEOF_XMMWORD-1
357        jnz     short .out1
358        ; --(aligned)-------------------
359        movntdq XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
360        movntdq XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
361        movntdq XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
362        movntdq XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
363        jmp     short .out0
364.out1:  ; --(unaligned)-----------------
365        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
366        movdqu  XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
367        movdqu  XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
368        movdqu  XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
369.out0:
370        add     rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
371        sub     rcx, byte SIZEOF_XMMWORD
372        jz      near .nextrow
373
374        add     rsi, byte SIZEOF_XMMWORD        ; inptr0
375        add     rbx, byte SIZEOF_XMMWORD        ; inptr1
376        add     rdx, byte SIZEOF_XMMWORD        ; inptr2
377        jmp     near .columnloop
378
379.column_st32:
380        cmp     rcx, byte SIZEOF_XMMWORD/2
381        jb      short .column_st16
382        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
383        movdqu  XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
384        add     rdi, byte 2*SIZEOF_XMMWORD      ; outptr
385        movdqa  xmmA,xmmC
386        movdqa  xmmD,xmmH
387        sub     rcx, byte SIZEOF_XMMWORD/2
388.column_st16:
389        cmp     rcx, byte SIZEOF_XMMWORD/4
390        jb      short .column_st15
391        movdqu  XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
392        add     rdi, byte SIZEOF_XMMWORD        ; outptr
393        movdqa  xmmA,xmmD
394        sub     rcx, byte SIZEOF_XMMWORD/4
395.column_st15:
396        ; Store two pixels (8 bytes) of xmmA to the output when it has enough
397        ; space.
398        cmp     rcx, byte SIZEOF_XMMWORD/8
399        jb      short .column_st7
400        movq    MMWORD [rdi], xmmA
401        add     rdi, byte SIZEOF_XMMWORD/8*4
402        sub     rcx, byte SIZEOF_XMMWORD/8
403        psrldq  xmmA, SIZEOF_XMMWORD/8*4
404.column_st7:
405        ; Store one pixel (4 bytes) of xmmA to the output when it has enough
406        ; space.
407        test    rcx, rcx
408        jz      short .nextrow
409        movd    XMM_DWORD [rdi], xmmA
410
411%endif ; RGB_PIXELSIZE ; ---------------
412
413.nextrow:
414        pop     rcx
415        pop     rsi
416        pop     rbx
417        pop     rdx
418        pop     rdi
419        pop     rax
420
421        add     rsi, byte SIZEOF_JSAMPROW
422        add     rbx, byte SIZEOF_JSAMPROW
423        add     rdx, byte SIZEOF_JSAMPROW
424        add     rdi, byte SIZEOF_JSAMPROW       ; output_buf
425        dec     rax                             ; num_rows
426        jg      near .rowloop
427
428        sfence          ; flush the write buffer
429
430.return:
431        pop     rbx
432        uncollect_args
433        mov     rsp,rbp         ; rsp <- aligned rbp
434        pop     rsp             ; rsp <- original rbp
435        pop     rbp
436        ret
437
438; For some reason, the OS X linker does not honor the request to align the
439; segment unless we do this.
440        align   16
441