• 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; Copyright (C) 2018, Matthias Räncker.
7;
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
11;
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
17
18%include "jcolsamp.inc"
19
20; --------------------------------------------------------------------------
21;
22; Convert some rows of samples to the output colorspace.
23;
24; GLOBAL(void)
25; jsimd_ycc_rgb_convert_sse2(JDIMENSION out_width, JSAMPIMAGE input_buf,
26;                            JDIMENSION input_row, JSAMPARRAY output_buf,
27;                            int num_rows)
28;
29
30; r10d = JDIMENSION out_width
31; r11 = JSAMPIMAGE input_buf
32; r12d = JDIMENSION input_row
33; r13 = JSAMPARRAY output_buf
34; r14d = int num_rows
35
36%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
37%define WK_NUM  2
38
39    align       32
40    GLOBAL_FUNCTION(jsimd_ycc_rgb_convert_sse2)
41
42EXTN(jsimd_ycc_rgb_convert_sse2):
43    push        rbp
44    mov         rax, rsp                     ; rax = original rbp
45    sub         rsp, byte 4
46    and         rsp, byte (-SIZEOF_XMMWORD)  ; align to 128 bits
47    mov         [rsp], rax
48    mov         rbp, rsp                     ; rbp = aligned rbp
49    lea         rsp, [wk(0)]
50    collect_args 5
51    push        rbx
52
53    mov         ecx, r10d               ; num_cols
54    test        rcx, rcx
55    jz          near .return
56
57    push        rcx
58
59    mov         rdi, r11
60    mov         ecx, r12d
61    mov         rsip, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
62    mov         rbxp, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
63    mov         rdxp, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
64    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
65    lea         rbx, [rbx+rcx*SIZEOF_JSAMPROW]
66    lea         rdx, [rdx+rcx*SIZEOF_JSAMPROW]
67
68    pop         rcx
69
70    mov         rdi, r13
71    mov         eax, r14d
72    test        rax, rax
73    jle         near .return
74.rowloop:
75    push        rax
76    push        rdi
77    push        rdx
78    push        rbx
79    push        rsi
80    push        rcx                     ; col
81
82    mov         rsip, JSAMPROW [rsi]    ; inptr0
83    mov         rbxp, JSAMPROW [rbx]    ; inptr1
84    mov         rdxp, JSAMPROW [rdx]    ; inptr2
85    mov         rdip, JSAMPROW [rdi]    ; outptr
86.columnloop:
87
88    movdqa      xmm5, XMMWORD [rbx]     ; xmm5=Cb(0123456789ABCDEF)
89    movdqa      xmm1, XMMWORD [rdx]     ; xmm1=Cr(0123456789ABCDEF)
90
91    pcmpeqw     xmm4, xmm4
92    pcmpeqw     xmm7, xmm7
93    psrlw       xmm4, BYTE_BIT
94    psllw       xmm7, 7                 ; xmm7={0xFF80 0xFF80 0xFF80 0xFF80 ..}
95    movdqa      xmm0, xmm4              ; xmm0=xmm4={0xFF 0x00 0xFF 0x00 ..}
96
97    pand        xmm4, xmm5              ; xmm4=Cb(02468ACE)=CbE
98    psrlw       xmm5, BYTE_BIT          ; xmm5=Cb(13579BDF)=CbO
99    pand        xmm0, xmm1              ; xmm0=Cr(02468ACE)=CrE
100    psrlw       xmm1, BYTE_BIT          ; xmm1=Cr(13579BDF)=CrO
101
102    paddw       xmm4, xmm7
103    paddw       xmm5, xmm7
104    paddw       xmm0, xmm7
105    paddw       xmm1, xmm7
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    movdqa      xmm2, xmm4              ; xmm2=CbE
118    movdqa      xmm3, xmm5              ; xmm3=CbO
119    paddw       xmm4, xmm4              ; xmm4=2*CbE
120    paddw       xmm5, xmm5              ; xmm5=2*CbO
121    movdqa      xmm6, xmm0              ; xmm6=CrE
122    movdqa      xmm7, xmm1              ; xmm7=CrO
123    paddw       xmm0, xmm0              ; xmm0=2*CrE
124    paddw       xmm1, xmm1              ; xmm1=2*CrO
125
126    pmulhw      xmm4, [rel PW_MF0228]   ; xmm4=(2*CbE * -FIX(0.22800))
127    pmulhw      xmm5, [rel PW_MF0228]   ; xmm5=(2*CbO * -FIX(0.22800))
128    pmulhw      xmm0, [rel PW_F0402]    ; xmm0=(2*CrE * FIX(0.40200))
129    pmulhw      xmm1, [rel PW_F0402]    ; xmm1=(2*CrO * FIX(0.40200))
130
131    paddw       xmm4, [rel PW_ONE]
132    paddw       xmm5, [rel PW_ONE]
133    psraw       xmm4, 1                 ; xmm4=(CbE * -FIX(0.22800))
134    psraw       xmm5, 1                 ; xmm5=(CbO * -FIX(0.22800))
135    paddw       xmm0, [rel PW_ONE]
136    paddw       xmm1, [rel PW_ONE]
137    psraw       xmm0, 1                 ; xmm0=(CrE * FIX(0.40200))
138    psraw       xmm1, 1                 ; xmm1=(CrO * FIX(0.40200))
139
140    paddw       xmm4, xmm2
141    paddw       xmm5, xmm3
142    paddw       xmm4, xmm2              ; xmm4=(CbE * FIX(1.77200))=(B-Y)E
143    paddw       xmm5, xmm3              ; xmm5=(CbO * FIX(1.77200))=(B-Y)O
144    paddw       xmm0, xmm6              ; xmm0=(CrE * FIX(1.40200))=(R-Y)E
145    paddw       xmm1, xmm7              ; xmm1=(CrO * FIX(1.40200))=(R-Y)O
146
147    movdqa      XMMWORD [wk(0)], xmm4   ; wk(0)=(B-Y)E
148    movdqa      XMMWORD [wk(1)], xmm5   ; wk(1)=(B-Y)O
149
150    movdqa      xmm4, xmm2
151    movdqa      xmm5, xmm3
152    punpcklwd   xmm2, xmm6
153    punpckhwd   xmm4, xmm6
154    pmaddwd     xmm2, [rel PW_MF0344_F0285]
155    pmaddwd     xmm4, [rel PW_MF0344_F0285]
156    punpcklwd   xmm3, xmm7
157    punpckhwd   xmm5, xmm7
158    pmaddwd     xmm3, [rel PW_MF0344_F0285]
159    pmaddwd     xmm5, [rel PW_MF0344_F0285]
160
161    paddd       xmm2, [rel PD_ONEHALF]
162    paddd       xmm4, [rel PD_ONEHALF]
163    psrad       xmm2, SCALEBITS
164    psrad       xmm4, SCALEBITS
165    paddd       xmm3, [rel PD_ONEHALF]
166    paddd       xmm5, [rel PD_ONEHALF]
167    psrad       xmm3, SCALEBITS
168    psrad       xmm5, SCALEBITS
169
170    packssdw    xmm2, xmm4              ; xmm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
171    packssdw    xmm3, xmm5              ; xmm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
172    psubw       xmm2, xmm6              ; xmm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
173    psubw       xmm3, xmm7              ; xmm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
174
175    movdqa      xmm5, XMMWORD [rsi]     ; xmm5=Y(0123456789ABCDEF)
176
177    pcmpeqw     xmm4, xmm4
178    psrlw       xmm4, BYTE_BIT          ; xmm4={0xFF 0x00 0xFF 0x00 ..}
179    pand        xmm4, xmm5              ; xmm4=Y(02468ACE)=YE
180    psrlw       xmm5, BYTE_BIT          ; xmm5=Y(13579BDF)=YO
181
182    paddw       xmm0, xmm4              ; xmm0=((R-Y)E+YE)=RE=R(02468ACE)
183    paddw       xmm1, xmm5              ; xmm1=((R-Y)O+YO)=RO=R(13579BDF)
184    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
185    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
186
187    paddw       xmm2, xmm4              ; xmm2=((G-Y)E+YE)=GE=G(02468ACE)
188    paddw       xmm3, xmm5              ; xmm3=((G-Y)O+YO)=GO=G(13579BDF)
189    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
190    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
191
192    paddw       xmm4, XMMWORD [wk(0)]   ; xmm4=(YE+(B-Y)E)=BE=B(02468ACE)
193    paddw       xmm5, XMMWORD [wk(1)]   ; xmm5=(YO+(B-Y)O)=BO=B(13579BDF)
194    packuswb    xmm4, xmm4              ; xmm4=B(02468ACE********)
195    packuswb    xmm5, xmm5              ; xmm5=B(13579BDF********)
196
197%if RGB_PIXELSIZE == 3  ; ---------------
198
199    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
200    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
201    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
202    ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
203
204    punpcklbw   xmmA, xmmC        ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
205    punpcklbw   xmmE, xmmB        ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
206    punpcklbw   xmmD, xmmF        ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
207
208    movdqa      xmmG, xmmA
209    movdqa      xmmH, xmmA
210    punpcklwd   xmmA, xmmE        ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
211    punpckhwd   xmmG, xmmE        ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
212
213    psrldq      xmmH, 2           ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
214    psrldq      xmmE, 2           ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
215
216    movdqa      xmmC, xmmD
217    movdqa      xmmB, xmmD
218    punpcklwd   xmmD, xmmH        ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
219    punpckhwd   xmmC, xmmH        ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
220
221    psrldq      xmmB, 2           ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
222
223    movdqa      xmmF, xmmE
224    punpcklwd   xmmE, xmmB        ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
225    punpckhwd   xmmF, xmmB        ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
226
227    pshufd      xmmH, xmmA, 0x4E  ; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
228    movdqa      xmmB, xmmE
229    punpckldq   xmmA, xmmD        ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
230    punpckldq   xmmE, xmmH        ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
231    punpckhdq   xmmD, xmmB        ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
232
233    pshufd      xmmH, xmmG, 0x4E  ; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
234    movdqa      xmmB, xmmF
235    punpckldq   xmmG, xmmC        ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
236    punpckldq   xmmF, xmmH        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
237    punpckhdq   xmmC, xmmB        ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
238
239    punpcklqdq  xmmA, xmmE        ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
240    punpcklqdq  xmmD, xmmG        ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
241    punpcklqdq  xmmF, xmmC        ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
242
243    cmp         rcx, byte SIZEOF_XMMWORD
244    jb          short .column_st32
245
246    test        rdi, SIZEOF_XMMWORD-1
247    jnz         short .out1
248    ; --(aligned)-------------------
249    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
250    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
251    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
252    jmp         short .out0
253.out1:  ; --(unaligned)-----------------
254    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
255    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
256    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmF
257.out0:
258    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
259    sub         rcx, byte SIZEOF_XMMWORD
260    jz          near .nextrow
261
262    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
263    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
264    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
265    jmp         near .columnloop
266
267.column_st32:
268    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
269    cmp         rcx, byte 2*SIZEOF_XMMWORD
270    jb          short .column_st16
271    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
272    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
273    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
274    movdqa      xmmA, xmmF
275    sub         rcx, byte 2*SIZEOF_XMMWORD
276    jmp         short .column_st15
277.column_st16:
278    cmp         rcx, byte SIZEOF_XMMWORD
279    jb          short .column_st15
280    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
281    add         rdi, byte SIZEOF_XMMWORD    ; outptr
282    movdqa      xmmA, xmmD
283    sub         rcx, byte SIZEOF_XMMWORD
284.column_st15:
285    ; Store the lower 8 bytes of xmmA to the output when it has enough
286    ; space.
287    cmp         rcx, byte SIZEOF_MMWORD
288    jb          short .column_st7
289    movq        XMM_MMWORD [rdi], xmmA
290    add         rdi, byte SIZEOF_MMWORD
291    sub         rcx, byte SIZEOF_MMWORD
292    psrldq      xmmA, SIZEOF_MMWORD
293.column_st7:
294    ; Store the lower 4 bytes of xmmA to the output when it has enough
295    ; space.
296    cmp         rcx, byte SIZEOF_DWORD
297    jb          short .column_st3
298    movd        XMM_DWORD [rdi], xmmA
299    add         rdi, byte SIZEOF_DWORD
300    sub         rcx, byte SIZEOF_DWORD
301    psrldq      xmmA, SIZEOF_DWORD
302.column_st3:
303    ; Store the lower 2 bytes of rax to the output when it has enough
304    ; space.
305    movd        eax, xmmA
306    cmp         rcx, byte SIZEOF_WORD
307    jb          short .column_st1
308    mov         word [rdi], ax
309    add         rdi, byte SIZEOF_WORD
310    sub         rcx, byte SIZEOF_WORD
311    shr         rax, 16
312.column_st1:
313    ; Store the lower 1 byte of rax to the output when it has enough
314    ; space.
315    test        rcx, rcx
316    jz          short .nextrow
317    mov         byte [rdi], al
318
319%else  ; RGB_PIXELSIZE == 4 ; -----------
320
321%ifdef RGBX_FILLER_0XFF
322    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
323    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
324%else
325    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
326    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
327%endif
328    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
329    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
330    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
331    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
332
333    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
334    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
335    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
336    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
337
338    movdqa      xmmC, xmmA
339    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
340    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
341    movdqa      xmmG, xmmB
342    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
343    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
344
345    movdqa      xmmD, xmmA
346    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
347    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
348    movdqa      xmmH, xmmC
349    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
350    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
351
352    cmp         rcx, byte SIZEOF_XMMWORD
353    jb          short .column_st32
354
355    test        rdi, SIZEOF_XMMWORD-1
356    jnz         short .out1
357    ; --(aligned)-------------------
358    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
359    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
360    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
361    movntdq     XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
362    jmp         short .out0
363.out1:  ; --(unaligned)-----------------
364    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
365    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
366    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
367    movdqu      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
368.out0:
369    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
370    sub         rcx, byte SIZEOF_XMMWORD
371    jz          near .nextrow
372
373    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
374    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
375    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
376    jmp         near .columnloop
377
378.column_st32:
379    cmp         rcx, byte SIZEOF_XMMWORD/2
380    jb          short .column_st16
381    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
382    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
383    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
384    movdqa      xmmA, xmmC
385    movdqa      xmmD, xmmH
386    sub         rcx, byte SIZEOF_XMMWORD/2
387.column_st16:
388    cmp         rcx, byte SIZEOF_XMMWORD/4
389    jb          short .column_st15
390    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
391    add         rdi, byte SIZEOF_XMMWORD    ; outptr
392    movdqa      xmmA, xmmD
393    sub         rcx, byte SIZEOF_XMMWORD/4
394.column_st15:
395    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
396    ; space.
397    cmp         rcx, byte SIZEOF_XMMWORD/8
398    jb          short .column_st7
399    movq        MMWORD [rdi], xmmA
400    add         rdi, byte SIZEOF_XMMWORD/8*4
401    sub         rcx, byte SIZEOF_XMMWORD/8
402    psrldq      xmmA, SIZEOF_XMMWORD/8*4
403.column_st7:
404    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
405    ; space.
406    test        rcx, rcx
407    jz          short .nextrow
408    movd        XMM_DWORD [rdi], xmmA
409
410%endif  ; RGB_PIXELSIZE ; ---------------
411
412.nextrow:
413    pop         rcx
414    pop         rsi
415    pop         rbx
416    pop         rdx
417    pop         rdi
418    pop         rax
419
420    add         rsi, byte SIZEOF_JSAMPROW
421    add         rbx, byte SIZEOF_JSAMPROW
422    add         rdx, byte SIZEOF_JSAMPROW
423    add         rdi, byte SIZEOF_JSAMPROW  ; output_buf
424    dec         rax                        ; num_rows
425    jg          near .rowloop
426
427    sfence                              ; flush the write buffer
428
429.return:
430    pop         rbx
431    uncollect_args 5
432    mov         rsp, rbp                ; rsp <- aligned rbp
433    pop         rsp                     ; rsp <- original rbp
434    pop         rbp
435    ret
436
437; For some reason, the OS X linker does not honor the request to align the
438; segment unless we do this.
439    align       32
440