• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jdmrgext.asm - merged upsampling/color 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; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
22;
23; GLOBAL(void)
24; jsimd_h2v1_merged_upsample_sse2(JDIMENSION output_width,
25;                                 JSAMPIMAGE input_buf,
26;                                 JDIMENSION in_row_group_ctr,
27;                                 JSAMPARRAY output_buf);
28;
29
30; r10d = JDIMENSION output_width
31; r11 = JSAMPIMAGE input_buf
32; r12d = JDIMENSION in_row_group_ctr
33; r13 = JSAMPARRAY output_buf
34
35%define wk(i)   rbp - (WK_NUM - (i)) * SIZEOF_XMMWORD  ; xmmword wk[WK_NUM]
36%define WK_NUM  3
37
38    align       32
39    GLOBAL_FUNCTION(jsimd_h2v1_merged_upsample_sse2)
40
41EXTN(jsimd_h2v1_merged_upsample_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 4
50    push        rbx
51
52    mov         ecx, r10d               ; col
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    mov         rdi, r13
64    mov         rsi, JSAMPROW [rsi+rcx*SIZEOF_JSAMPROW]  ; inptr0
65    mov         rbx, JSAMPROW [rbx+rcx*SIZEOF_JSAMPROW]  ; inptr1
66    mov         rdx, JSAMPROW [rdx+rcx*SIZEOF_JSAMPROW]  ; inptr2
67    mov         rdi, JSAMPROW [rdi]                      ; outptr
68
69    pop         rcx                     ; col
70
71.columnloop:
72
73    movdqa      xmm6, XMMWORD [rbx]     ; xmm6=Cb(0123456789ABCDEF)
74    movdqa      xmm7, XMMWORD [rdx]     ; xmm7=Cr(0123456789ABCDEF)
75
76    pxor        xmm1, xmm1              ; xmm1=(all 0's)
77    pcmpeqw     xmm3, xmm3
78    psllw       xmm3, 7                 ; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
79
80    movdqa      xmm4, xmm6
81    punpckhbw   xmm6, xmm1              ; xmm6=Cb(89ABCDEF)=CbH
82    punpcklbw   xmm4, xmm1              ; xmm4=Cb(01234567)=CbL
83    movdqa      xmm0, xmm7
84    punpckhbw   xmm7, xmm1              ; xmm7=Cr(89ABCDEF)=CrH
85    punpcklbw   xmm0, xmm1              ; xmm0=Cr(01234567)=CrL
86
87    paddw       xmm6, xmm3
88    paddw       xmm4, xmm3
89    paddw       xmm7, xmm3
90    paddw       xmm0, xmm3
91
92    ; (Original)
93    ; R = Y                + 1.40200 * Cr
94    ; G = Y - 0.34414 * Cb - 0.71414 * Cr
95    ; B = Y + 1.77200 * Cb
96    ;
97    ; (This implementation)
98    ; R = Y                + 0.40200 * Cr + Cr
99    ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
100    ; B = Y - 0.22800 * Cb + Cb + Cb
101
102    movdqa      xmm5, xmm6              ; xmm5=CbH
103    movdqa      xmm2, xmm4              ; xmm2=CbL
104    paddw       xmm6, xmm6              ; xmm6=2*CbH
105    paddw       xmm4, xmm4              ; xmm4=2*CbL
106    movdqa      xmm1, xmm7              ; xmm1=CrH
107    movdqa      xmm3, xmm0              ; xmm3=CrL
108    paddw       xmm7, xmm7              ; xmm7=2*CrH
109    paddw       xmm0, xmm0              ; xmm0=2*CrL
110
111    pmulhw      xmm6, [rel PW_MF0228]   ; xmm6=(2*CbH * -FIX(0.22800))
112    pmulhw      xmm4, [rel PW_MF0228]   ; xmm4=(2*CbL * -FIX(0.22800))
113    pmulhw      xmm7, [rel PW_F0402]    ; xmm7=(2*CrH * FIX(0.40200))
114    pmulhw      xmm0, [rel PW_F0402]    ; xmm0=(2*CrL * FIX(0.40200))
115
116    paddw       xmm6, [rel PW_ONE]
117    paddw       xmm4, [rel PW_ONE]
118    psraw       xmm6, 1                 ; xmm6=(CbH * -FIX(0.22800))
119    psraw       xmm4, 1                 ; xmm4=(CbL * -FIX(0.22800))
120    paddw       xmm7, [rel PW_ONE]
121    paddw       xmm0, [rel PW_ONE]
122    psraw       xmm7, 1                 ; xmm7=(CrH * FIX(0.40200))
123    psraw       xmm0, 1                 ; xmm0=(CrL * FIX(0.40200))
124
125    paddw       xmm6, xmm5
126    paddw       xmm4, xmm2
127    paddw       xmm6, xmm5              ; xmm6=(CbH * FIX(1.77200))=(B-Y)H
128    paddw       xmm4, xmm2              ; xmm4=(CbL * FIX(1.77200))=(B-Y)L
129    paddw       xmm7, xmm1              ; xmm7=(CrH * FIX(1.40200))=(R-Y)H
130    paddw       xmm0, xmm3              ; xmm0=(CrL * FIX(1.40200))=(R-Y)L
131
132    movdqa      XMMWORD [wk(0)], xmm6   ; wk(0)=(B-Y)H
133    movdqa      XMMWORD [wk(1)], xmm7   ; wk(1)=(R-Y)H
134
135    movdqa      xmm6, xmm5
136    movdqa      xmm7, xmm2
137    punpcklwd   xmm5, xmm1
138    punpckhwd   xmm6, xmm1
139    pmaddwd     xmm5, [rel PW_MF0344_F0285]
140    pmaddwd     xmm6, [rel PW_MF0344_F0285]
141    punpcklwd   xmm2, xmm3
142    punpckhwd   xmm7, xmm3
143    pmaddwd     xmm2, [rel PW_MF0344_F0285]
144    pmaddwd     xmm7, [rel PW_MF0344_F0285]
145
146    paddd       xmm5, [rel PD_ONEHALF]
147    paddd       xmm6, [rel PD_ONEHALF]
148    psrad       xmm5, SCALEBITS
149    psrad       xmm6, SCALEBITS
150    paddd       xmm2, [rel PD_ONEHALF]
151    paddd       xmm7, [rel PD_ONEHALF]
152    psrad       xmm2, SCALEBITS
153    psrad       xmm7, SCALEBITS
154
155    packssdw    xmm5, xmm6              ; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
156    packssdw    xmm2, xmm7              ; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
157    psubw       xmm5, xmm1              ; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
158    psubw       xmm2, xmm3              ; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
159
160    movdqa      XMMWORD [wk(2)], xmm5   ; wk(2)=(G-Y)H
161
162    mov         al, 2                   ; Yctr
163    jmp         short .Yloop_1st
164
165.Yloop_2nd:
166    movdqa      xmm0, XMMWORD [wk(1)]   ; xmm0=(R-Y)H
167    movdqa      xmm2, XMMWORD [wk(2)]   ; xmm2=(G-Y)H
168    movdqa      xmm4, XMMWORD [wk(0)]   ; xmm4=(B-Y)H
169
170.Yloop_1st:
171    movdqa      xmm7, XMMWORD [rsi]     ; xmm7=Y(0123456789ABCDEF)
172
173    pcmpeqw     xmm6, xmm6
174    psrlw       xmm6, BYTE_BIT          ; xmm6={0xFF 0x00 0xFF 0x00 ..}
175    pand        xmm6, xmm7              ; xmm6=Y(02468ACE)=YE
176    psrlw       xmm7, BYTE_BIT          ; xmm7=Y(13579BDF)=YO
177
178    movdqa      xmm1, xmm0              ; xmm1=xmm0=(R-Y)(L/H)
179    movdqa      xmm3, xmm2              ; xmm3=xmm2=(G-Y)(L/H)
180    movdqa      xmm5, xmm4              ; xmm5=xmm4=(B-Y)(L/H)
181
182    paddw       xmm0, xmm6              ; xmm0=((R-Y)+YE)=RE=R(02468ACE)
183    paddw       xmm1, xmm7              ; xmm1=((R-Y)+YO)=RO=R(13579BDF)
184    packuswb    xmm0, xmm0              ; xmm0=R(02468ACE********)
185    packuswb    xmm1, xmm1              ; xmm1=R(13579BDF********)
186
187    paddw       xmm2, xmm6              ; xmm2=((G-Y)+YE)=GE=G(02468ACE)
188    paddw       xmm3, xmm7              ; xmm3=((G-Y)+YO)=GO=G(13579BDF)
189    packuswb    xmm2, xmm2              ; xmm2=G(02468ACE********)
190    packuswb    xmm3, xmm3              ; xmm3=G(13579BDF********)
191
192    paddw       xmm4, xmm6              ; xmm4=((B-Y)+YE)=BE=B(02468ACE)
193    paddw       xmm5, xmm7              ; xmm5=((B-Y)+YO)=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 .endcolumn
261
262    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
263    dec         al                        ; Yctr
264    jnz         near .Yloop_2nd
265
266    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
267    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
268    jmp         near .columnloop
269
270.column_st32:
271    lea         rcx, [rcx+rcx*2]            ; imul ecx, RGB_PIXELSIZE
272    cmp         rcx, byte 2*SIZEOF_XMMWORD
273    jb          short .column_st16
274    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
275    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
276    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
277    movdqa      xmmA, xmmF
278    sub         rcx, byte 2*SIZEOF_XMMWORD
279    jmp         short .column_st15
280.column_st16:
281    cmp         rcx, byte SIZEOF_XMMWORD
282    jb          short .column_st15
283    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
284    add         rdi, byte SIZEOF_XMMWORD    ; outptr
285    movdqa      xmmA, xmmD
286    sub         rcx, byte SIZEOF_XMMWORD
287.column_st15:
288    ; Store the lower 8 bytes of xmmA to the output when it has enough
289    ; space.
290    cmp         rcx, byte SIZEOF_MMWORD
291    jb          short .column_st7
292    movq        XMM_MMWORD [rdi], xmmA
293    add         rdi, byte SIZEOF_MMWORD
294    sub         rcx, byte SIZEOF_MMWORD
295    psrldq      xmmA, SIZEOF_MMWORD
296.column_st7:
297    ; Store the lower 4 bytes of xmmA to the output when it has enough
298    ; space.
299    cmp         rcx, byte SIZEOF_DWORD
300    jb          short .column_st3
301    movd        XMM_DWORD [rdi], xmmA
302    add         rdi, byte SIZEOF_DWORD
303    sub         rcx, byte SIZEOF_DWORD
304    psrldq      xmmA, SIZEOF_DWORD
305.column_st3:
306    ; Store the lower 2 bytes of rax to the output when it has enough
307    ; space.
308    movd        eax, xmmA
309    cmp         rcx, byte SIZEOF_WORD
310    jb          short .column_st1
311    mov         word [rdi], ax
312    add         rdi, byte SIZEOF_WORD
313    sub         rcx, byte SIZEOF_WORD
314    shr         rax, 16
315.column_st1:
316    ; Store the lower 1 byte of rax to the output when it has enough
317    ; space.
318    test        rcx, rcx
319    jz          short .endcolumn
320    mov         byte [rdi], al
321
322%else  ; RGB_PIXELSIZE == 4 ; -----------
323
324%ifdef RGBX_FILLER_0XFF
325    pcmpeqb     xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
326    pcmpeqb     xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
327%else
328    pxor        xmm6, xmm6              ; xmm6=XE=X(02468ACE********)
329    pxor        xmm7, xmm7              ; xmm7=XO=X(13579BDF********)
330%endif
331    ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
332    ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
333    ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
334    ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
335
336    punpcklbw   xmmA, xmmC  ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
337    punpcklbw   xmmE, xmmG  ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
338    punpcklbw   xmmB, xmmD  ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
339    punpcklbw   xmmF, xmmH  ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
340
341    movdqa      xmmC, xmmA
342    punpcklwd   xmmA, xmmE  ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
343    punpckhwd   xmmC, xmmE  ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
344    movdqa      xmmG, xmmB
345    punpcklwd   xmmB, xmmF  ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
346    punpckhwd   xmmG, xmmF  ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
347
348    movdqa      xmmD, xmmA
349    punpckldq   xmmA, xmmB  ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
350    punpckhdq   xmmD, xmmB  ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
351    movdqa      xmmH, xmmC
352    punpckldq   xmmC, xmmG  ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
353    punpckhdq   xmmH, xmmG  ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
354
355    cmp         rcx, byte SIZEOF_XMMWORD
356    jb          short .column_st32
357
358    test        rdi, SIZEOF_XMMWORD-1
359    jnz         short .out1
360    ; --(aligned)-------------------
361    movntdq     XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
362    movntdq     XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
363    movntdq     XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
364    movntdq     XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
365    jmp         short .out0
366.out1:  ; --(unaligned)-----------------
367    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
368    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
369    movdqu      XMMWORD [rdi+2*SIZEOF_XMMWORD], xmmC
370    movdqu      XMMWORD [rdi+3*SIZEOF_XMMWORD], xmmH
371.out0:
372    add         rdi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
373    sub         rcx, byte SIZEOF_XMMWORD
374    jz          near .endcolumn
375
376    add         rsi, byte SIZEOF_XMMWORD  ; inptr0
377    dec         al                        ; Yctr
378    jnz         near .Yloop_2nd
379
380    add         rbx, byte SIZEOF_XMMWORD  ; inptr1
381    add         rdx, byte SIZEOF_XMMWORD  ; inptr2
382    jmp         near .columnloop
383
384.column_st32:
385    cmp         rcx, byte SIZEOF_XMMWORD/2
386    jb          short .column_st16
387    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
388    movdqu      XMMWORD [rdi+1*SIZEOF_XMMWORD], xmmD
389    add         rdi, byte 2*SIZEOF_XMMWORD  ; outptr
390    movdqa      xmmA, xmmC
391    movdqa      xmmD, xmmH
392    sub         rcx, byte SIZEOF_XMMWORD/2
393.column_st16:
394    cmp         rcx, byte SIZEOF_XMMWORD/4
395    jb          short .column_st15
396    movdqu      XMMWORD [rdi+0*SIZEOF_XMMWORD], xmmA
397    add         rdi, byte SIZEOF_XMMWORD    ; outptr
398    movdqa      xmmA, xmmD
399    sub         rcx, byte SIZEOF_XMMWORD/4
400.column_st15:
401    ; Store two pixels (8 bytes) of xmmA to the output when it has enough
402    ; space.
403    cmp         rcx, byte SIZEOF_XMMWORD/8
404    jb          short .column_st7
405    movq        XMM_MMWORD [rdi], xmmA
406    add         rdi, byte SIZEOF_XMMWORD/8*4
407    sub         rcx, byte SIZEOF_XMMWORD/8
408    psrldq      xmmA, SIZEOF_XMMWORD/8*4
409.column_st7:
410    ; Store one pixel (4 bytes) of xmmA to the output when it has enough
411    ; space.
412    test        rcx, rcx
413    jz          short .endcolumn
414    movd        XMM_DWORD [rdi], xmmA
415
416%endif  ; RGB_PIXELSIZE ; ---------------
417
418.endcolumn:
419    sfence                              ; flush the write buffer
420
421.return:
422    pop         rbx
423    uncollect_args 4
424    mov         rsp, rbp                ; rsp <- aligned rbp
425    pop         rsp                     ; rsp <- original rbp
426    pop         rbp
427    ret
428
429; --------------------------------------------------------------------------
430;
431; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
432;
433; GLOBAL(void)
434; jsimd_h2v2_merged_upsample_sse2(JDIMENSION output_width,
435;                                 JSAMPIMAGE input_buf,
436;                                 JDIMENSION in_row_group_ctr,
437;                                 JSAMPARRAY output_buf);
438;
439
440; r10d = JDIMENSION output_width
441; r11 = JSAMPIMAGE input_buf
442; r12d = JDIMENSION in_row_group_ctr
443; r13 = JSAMPARRAY output_buf
444
445    align       32
446    GLOBAL_FUNCTION(jsimd_h2v2_merged_upsample_sse2)
447
448EXTN(jsimd_h2v2_merged_upsample_sse2):
449    push        rbp
450    mov         rax, rsp
451    mov         rbp, rsp
452    collect_args 4
453    push        rbx
454
455    mov         eax, r10d
456
457    mov         rdi, r11
458    mov         ecx, r12d
459    mov         rsi, JSAMPARRAY [rdi+0*SIZEOF_JSAMPARRAY]
460    mov         rbx, JSAMPARRAY [rdi+1*SIZEOF_JSAMPARRAY]
461    mov         rdx, JSAMPARRAY [rdi+2*SIZEOF_JSAMPARRAY]
462    mov         rdi, r13
463    lea         rsi, [rsi+rcx*SIZEOF_JSAMPROW]
464
465    push        rdx                     ; inptr2
466    push        rbx                     ; inptr1
467    push        rsi                     ; inptr00
468    mov         rbx, rsp
469
470    push        rdi
471    push        rcx
472    push        rax
473
474    %ifdef WIN64
475    mov         r8, rcx
476    mov         r9, rdi
477    mov         rcx, rax
478    mov         rdx, rbx
479    %else
480    mov         rdx, rcx
481    mov         rcx, rdi
482    mov         rdi, rax
483    mov         rsi, rbx
484    %endif
485
486    call        EXTN(jsimd_h2v1_merged_upsample_sse2)
487
488    pop         rax
489    pop         rcx
490    pop         rdi
491    pop         rsi
492    pop         rbx
493    pop         rdx
494
495    add         rdi, byte SIZEOF_JSAMPROW  ; outptr1
496    add         rsi, byte SIZEOF_JSAMPROW  ; inptr01
497
498    push        rdx                     ; inptr2
499    push        rbx                     ; inptr1
500    push        rsi                     ; inptr00
501    mov         rbx, rsp
502
503    push        rdi
504    push        rcx
505    push        rax
506
507    %ifdef WIN64
508    mov         r8, rcx
509    mov         r9, rdi
510    mov         rcx, rax
511    mov         rdx, rbx
512    %else
513    mov         rdx, rcx
514    mov         rcx, rdi
515    mov         rdi, rax
516    mov         rsi, rbx
517    %endif
518
519    call        EXTN(jsimd_h2v1_merged_upsample_sse2)
520
521    pop         rax
522    pop         rcx
523    pop         rdi
524    pop         rsi
525    pop         rbx
526    pop         rdx
527
528    pop         rbx
529    uncollect_args 4
530    pop         rbp
531    ret
532
533; For some reason, the OS X linker does not honor the request to align the
534; segment unless we do this.
535    align       32
536