• 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, 2016, 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%include "jcolsamp.inc"
18
19; --------------------------------------------------------------------------
20;
21; Convert some rows of samples to the output colorspace.
22;
23; GLOBAL(void)
24; jsimd_ycc_rgb_convert_sse2(JDIMENSION out_width, JSAMPIMAGE input_buf,
25;                            JDIMENSION input_row, JSAMPARRAY output_buf,
26;                            int num_rows)
27;
28
29; r10d = JDIMENSION out_width
30; r11 = JSAMPIMAGE input_buf
31; r12d = JDIMENSION input_row
32; r13 = JSAMPARRAY output_buf
33; r14d = int num_rows
34
35%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
36%define WK_NUM  2
37
38    align       32
39    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
40
41EXTN(jsimd_ycc_rgb_convert_sse2):
42    push        rbp
43    mov         rax, rsp                     ; rax = original rbp
44    sub         rsp, byte 4
45    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
46    mov         [rsp], rax
47    mov         rbp, rsp                     ; rbp = aligned rbp
48    lea         rsp, [wk(0)]
49    collect_args 5
50    push        rbx
51
52    mov         ecx, r10d               ; num_cols
53    test        rcx, rcx
54    jz          near .return
55
56    push        rcx
57
58    mov         rdi, r11
59    mov         ecx, r12d
60    mov         rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
61    mov         rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
62    mov         rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
63    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
64    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
65    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
66
67    pop         rcx
68
69    mov         rdi, r13
70    mov         eax, r14d
71    test        rax, rax
72    jle         near .return
73.rowloop:
74    push        rax
75    push        rdi
76    push        rdx
77    push        rbx
78    push        rsi
79    push        rcx                     ; col
80
81    mov         rsi, JSAMPROW [rsi]     ; inptr0
82    mov         rbx, JSAMPROW [rbx]     ; inptr1
83    mov         rdx, JSAMPROW [rdx]     ; inptr2
84    mov         rdi, JSAMPROW [rdi]     ; outptr
85.columnloop:
86
87    movdqa      xmm5, XMMWORD [rbx]     ; xmm5=Cb(0123456789ABCDEF)
88    movdqa      xmm1, XMMWORD [rdx]     ; xmm1=Cr(0123456789ABCDEF)
89
90    pcmpeqw     xmm4, xmm4
91    pcmpeqw     xmm7, xmm7
92    psrlw       xmm4, BYTE_BIT
93    psllw       xmm7, 7                 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
94    movdqa      xmm0, xmm4              ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
95
96    pand        xmm4, xmm5              ; xmm4=Cb(02468ACE)=CbE
97    psrlw       xmm5, BYTE_BIT          ; xmm5=Cb(13579BDF)=CbO
98    pand        xmm0, xmm1              ; xmm0=Cr(02468ACE)=CrE
99    psrlw       xmm1, BYTE_BIT          ; xmm1=Cr(13579BDF)=CrO
100
101    paddw       xmm4, xmm7
102    paddw       xmm5, xmm7
103    paddw       xmm0, xmm7
104    paddw       xmm1, xmm7
105
106    ; (Original)
107    ; R = Y                + 1.40200 * Cr
108    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
109    ; B = Y + 1.77200 * Cb
110    ;
111    ; (This implementation)
112    ; R = Y                + 0.40200 * Cr + Cr
113    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
114    ; B = Y - 0.22800 * Cb + Cb + Cb
115
116    movdqa      xmm2, xmm4              ; xmm2=CbE
117    movdqa      xmm3, xmm5              ; xmm3=CbO
118    paddw       xmm4, xmm4              ; xmm4=2*CbE
119    paddw       xmm5, xmm5              ; xmm5=2*CbO
120    movdqa      xmm6, xmm0              ; xmm6=CrE
121    movdqa      xmm7, xmm1              ; xmm7=CrO
122    paddw       xmm0, xmm0              ; xmm0=2*CrE
123    paddw       xmm1, xmm1              ; xmm1=2*CrO
124
125    pmulhw      xmm4, [rel PW_MF0228]   ; xmm4=(2*CbE * -FIX(0.22800))
126    pmulhw      xmm5, [rel PW_MF0228]   ; xmm5=(2*CbO * -FIX(0.22800))
127    pmulhw      xmm0, [rel PW_F0402]    ; xmm0=(2*CrE * FIX(0.40200))
128    pmulhw      xmm1, [rel PW_F0402]    ; xmm1=(2*CrO * FIX(0.40200))
129
130    paddw       xmm4, [rel PW_ONE]
131    paddw       xmm5, [rel PW_ONE]
132    psraw       xmm4, 1                 ; xmm4=(CbE * -FIX(0.22800))
133    psraw       xmm5, 1                 ; xmm5=(CbO * -FIX(0.22800))
134    paddw       xmm0, [rel PW_ONE]
135    paddw       xmm1, [rel PW_ONE]
136    psraw       xmm0, 1                 ; xmm0=(CrE * FIX(0.40200))
137    psraw       xmm1, 1                 ; xmm1=(CrO * FIX(0.40200))
138
139    paddw       xmm4, xmm2
140    paddw       xmm5, xmm3
141    paddw       xmm4, xmm2              ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
142    paddw       xmm5, xmm3              ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
143    paddw       xmm0, xmm6              ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
144    paddw       xmm1, xmm7              ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
145
146    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=(B-Y)E
147    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(B-Y)O
148
149    movdqa      xmm4, xmm2
150    movdqa      xmm5, xmm3
151    punpcklwd   xmm2, xmm6
152    punpckhwd   xmm4, xmm6
153    pmaddwd     xmm2, [rel PW_MF0344_F0285]
154    pmaddwd     xmm4, [rel PW_MF0344_F0285]
155    punpcklwd   xmm3, xmm7
156    punpckhwd   xmm5, xmm7
157    pmaddwd     xmm3, [rel PW_MF0344_F0285]
158    pmaddwd     xmm5, [rel PW_MF0344_F0285]
159
160    paddd       xmm2, [rel PD_ONEHALF]
161    paddd       xmm4, [rel PD_ONEHALF]
162    psrad       xmm2, SCALEBITS
163    psrad       xmm4, SCALEBITS
164    paddd       xmm3, [rel PD_ONEHALF]
165    paddd       xmm5, [rel PD_ONEHALF]
166    psrad       xmm3, SCALEBITS
167    psrad       xmm5, SCALEBITS
168
169    packssdw    xmm2, xmm4              ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
170    packssdw    xmm3, xmm5              ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
171    psubw       xmm2, xmm6              ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
172    psubw       xmm3, xmm7              ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
173
174    movdqa      xmm5, XMMWORD [rsi]     ; xmm5=Y(0123456789ABCDEF)
175
176    pcmpeqw     xmm4, xmm4
177    psrlw       xmm4, BYTE_BIT          ; xmm4={0xFF 0x00 0xFF 0x00 ..}
178    pand        xmm4, xmm5              ; xmm4=Y(02468ACE)=YE
179    psrlw       xmm5, BYTE_BIT          ; xmm5=Y(13579BDF)=YO
180
181    paddw       xmm0, xmm4              ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
182    paddw       xmm1, xmm5              ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
183    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
184    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
185
186    paddw       xmm2, xmm4              ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
187    paddw       xmm3, xmm5              ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
188    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
189    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
190
191    paddw       xmm4, XMMWORD [wk(0)]   ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
192    paddw       xmm5, XMMWORD [wk(1)]   ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
193    packuswb    xmm4, xmm4              ; xmm4=B(02468ACE********)
194    packuswb    xmm5, xmm5              ; xmm5=B(13579BDF********)
195
196%if RGB_PIXELSIZE == 3  ; ---------------
197
198    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
199    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
200    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
201    ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
202
203    punpcklbw   xmmA, xmmC        ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
204    punpcklbw   xmmE, xmmB        ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
205    punpcklbw   xmmD, xmmF        ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
206
207    movdqa      xmmG, xmmA
208    movdqa      xmmH, xmmA
209    punpcklwd   xmmA, xmmE        ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
210    punpckhwd   xmmG, xmmE        ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
211
212    psrldq      xmmH, 2           ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
213    psrldq      xmmE, 2           ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
214
215    movdqa      xmmC, xmmD
216    movdqa      xmmB, xmmD
217    punpcklwd   xmmD, xmmH        ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
218    punpckhwd   xmmC, xmmH        ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
219
220    psrldq      xmmB, 2           ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
221
222    movdqa      xmmF, xmmE
223    punpcklwd   xmmE, xmmB        ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
224    punpckhwd   xmmF, xmmB        ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
225
226    pshufd      xmmH, xmmA, 0x4E  ; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
227    movdqa      xmmB, xmmE
228    punpckldq   xmmA, xmmD        ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
229    punpckldq   xmmE, xmmH        ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
230    punpckhdq   xmmD, xmmB        ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
231
232    pshufd      xmmH, xmmG, 0x4E  ; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
233    movdqa      xmmB, xmmF
234    punpckldq   xmmG, xmmC        ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
235    punpckldq   xmmF, xmmH        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
236    punpckhdq   xmmC, xmmB        ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
237
238    punpcklqdq  xmmA, xmmE        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
239    punpcklqdq  xmmD, xmmG        ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
240    punpcklqdq  xmmF, xmmC        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
241
242    cmp         rcx, byte SIZEOF_XMMWORD
243    jb          short .column_st32
244
245    test        rdi, SIZEOF_XMMWORD-1
246    jnz         short .out1
247    ; --(aligned)-------------------
248    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
249    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
250    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
251    jmp         short .out0
252.out1:  ; --(unaligned)-----------------
253    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
254    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
255    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
256.out0:
257    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
258    sub         rcx, byte SIZEOF_XMMWORD
259    jz          near .nextrow
260
261    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
262    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
263    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
264    jmp         near .columnloop
265
266.column_st32:
267    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
268    cmp         rcx, byte 2*SIZEOF_XMMWORD
269    jb          short .column_st16
270    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
271    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
272    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
273    movdqa      xmmA, xmmF
274    sub         rcx, byte 2*SIZEOF_XMMWORD
275    jmp         short .column_st15
276.column_st16:
277    cmp         rcx, byte SIZEOF_XMMWORD
278    jb          short .column_st15
279    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
280    add         rdi, byte SIZEOF_XMMWORD    ; outptr
281    movdqa      xmmA, xmmD
282    sub         rcx, byte SIZEOF_XMMWORD
283.column_st15:
284    ; Store the lower 8 bytes of xmmA to the output when it has enough
285    ; space.
286    cmp         rcx, byte SIZEOF_MMWORD
287    jb          short .column_st7
288    movq        XMM_MMWORD [rdi], xmmA
289    add         rdi, byte SIZEOF_MMWORD
290    sub         rcx, byte SIZEOF_MMWORD
291    psrldq      xmmA, SIZEOF_MMWORD
292.column_st7:
293    ; Store the lower 4 bytes of xmmA to the output when it has enough
294    ; space.
295    cmp         rcx, byte SIZEOF_DWORD
296    jb          short .column_st3
297    movd        XMM_DWORD [rdi], xmmA
298    add         rdi, byte SIZEOF_DWORD
299    sub         rcx, byte SIZEOF_DWORD
300    psrldq      xmmA, SIZEOF_DWORD
301.column_st3:
302    ; Store the lower 2 bytes of rax to the output when it has enough
303    ; space.
304    movd        eax, xmmA
305    cmp         rcx, byte SIZEOF_WORD
306    jb          short .column_st1
307    mov         word [rdi], ax
308    add         rdi, byte SIZEOF_WORD
309    sub         rcx, byte SIZEOF_WORD
310    shr         rax, 16
311.column_st1:
312    ; Store the lower 1 byte of rax to the output when it has enough
313    ; space.
314    test        rcx, rcx
315    jz          short .nextrow
316    mov         byte [rdi], al
317
318%else  ; RGB_PIXELSIZE == 4 ; -----------
319
320%ifdef RGBX_FILLER_0XFF
321    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
322    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
323%else
324    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
325    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
326%endif
327    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
328    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
329    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
330    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
331
332    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
333    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
334    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
335    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
336
337    movdqa      xmmC, xmmA
338    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
339    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
340    movdqa      xmmG, xmmB
341    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
342    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
343
344    movdqa      xmmD, xmmA
345    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
346    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
347    movdqa      xmmH, xmmC
348    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
349    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
350
351    cmp         rcx, byte SIZEOF_XMMWORD
352    jb          short .column_st32
353
354    test        rdi, SIZEOF_XMMWORD-1
355    jnz         short .out1
356    ; --(aligned)-------------------
357    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
358    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
359    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
360    movntdq     XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
361    jmp         short .out0
362.out1:  ; --(unaligned)-----------------
363    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
364    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
365    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
366    movdqu      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
367.out0:
368    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
369    sub         rcx, byte SIZEOF_XMMWORD
370    jz          near .nextrow
371
372    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
373    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
374    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
375    jmp         near .columnloop
376
377.column_st32:
378    cmp         rcx, byte SIZEOF_XMMWORD/2
379    jb          short .column_st16
380    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
381    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
382    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
383    movdqa      xmmA, xmmC
384    movdqa      xmmD, xmmH
385    sub         rcx, byte SIZEOF_XMMWORD/2
386.column_st16:
387    cmp         rcx, byte SIZEOF_XMMWORD/4
388    jb          short .column_st15
389    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
390    add         rdi, byte SIZEOF_XMMWORD    ; outptr
391    movdqa      xmmA, xmmD
392    sub         rcx, byte SIZEOF_XMMWORD/4
393.column_st15:
394    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
395    ; space.
396    cmp         rcx, byte SIZEOF_XMMWORD/8
397    jb          short .column_st7
398    movq        MMWORD [rdi], xmmA
399    add         rdi, byte SIZEOF_XMMWORD/8*4
400    sub         rcx, byte SIZEOF_XMMWORD/8
401    psrldq      xmmA, SIZEOF_XMMWORD/8*4
402.column_st7:
403    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
404    ; space.
405    test        rcx, rcx
406    jz          short .nextrow
407    movd        XMM_DWORD [rdi], xmmA
408
409%endif  ; RGB_PIXELSIZE ; ---------------
410
411.nextrow:
412    pop         rcx
413    pop         rsi
414    pop         rbx
415    pop         rdx
416    pop         rdi
417    pop         rax
418
419    add         rsi, byte SIZEOF_JSAMPROW
420    add         rbx, byte SIZEOF_JSAMPROW
421    add         rdx, byte SIZEOF_JSAMPROW
422    add         rdi, byte SIZEOF_JSAMPROW  ; output_buf
423    dec         rax                        ; num_rows
424    jg          near .rowloop
425
426    sfence                              ; flush the write buffer
427
428.return:
429    pop         rbx
430    uncollect_args 5
431    mov         rsp, rbp                ; rsp <- aligned rbp
432    pop         rsp                     ; rsp <- original rbp
433    pop         rbp
434    ret
435
436; For some reason, the OS X linker does not honor the request to align the
437; segment unless we do this.
438    align       32
439