• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1;
2; jdmrgext.asm - merged upsampling/color conversion (SSE2)
3;
4; Copyright 2009, 2012 Pierre Ossman <ossman@cendio.se> for Cendio AB
5; Copyright (C) 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; Upsample and color convert for the case of 2:1 horizontal and 1:1 vertical.
24;
25; GLOBAL(void)
26; jsimd_h2v1_merged_upsample_sse2 (JDIMENSION output_width,
27;                                  JSAMPIMAGE input_buf,
28;                                  JDIMENSION in_row_group_ctr,
29;                                  JSAMPARRAY output_buf);
30;
31
32%define output_width(b) (b)+8                   ; JDIMENSION output_width
33%define input_buf(b)            (b)+12          ; JSAMPIMAGE input_buf
34%define in_row_group_ctr(b)     (b)+16          ; JDIMENSION in_row_group_ctr
35%define output_buf(b)           (b)+20          ; JSAMPARRAY output_buf
36
37%define original_ebp    ebp+0
38%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_XMMWORD ; xmmword wk[WK_NUM]
39%define WK_NUM          3
40%define gotptr          wk(0)-SIZEOF_POINTER    ; void * gotptr
41
42        align   16
43        global  EXTN(jsimd_h2v1_merged_upsample_sse2)
44
45EXTN(jsimd_h2v1_merged_upsample_sse2):
46        push    ebp
47        mov     eax,esp                         ; eax = original ebp
48        sub     esp, byte 4
49        and     esp, byte (-SIZEOF_XMMWORD)     ; align to 128 bits
50        mov     [esp],eax
51        mov     ebp,esp                         ; ebp = aligned ebp
52        lea     esp, [wk(0)]
53        pushpic eax             ; make a room for GOT address
54        push    ebx
55;       push    ecx             ; need not be preserved
56;       push    edx             ; need not be preserved
57        push    esi
58        push    edi
59
60        get_GOT ebx                     ; get GOT address
61        movpic  POINTER [gotptr], ebx   ; save GOT address
62
63        mov     ecx, JDIMENSION [output_width(eax)]     ; col
64        test    ecx,ecx
65        jz      near .return
66
67        push    ecx
68
69        mov     edi, JSAMPIMAGE [input_buf(eax)]
70        mov     ecx, JDIMENSION [in_row_group_ctr(eax)]
71        mov     esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
72        mov     ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
73        mov     edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
74        mov     edi, JSAMPARRAY [output_buf(eax)]
75        mov     esi, JSAMPROW [esi+ecx*SIZEOF_JSAMPROW]         ; inptr0
76        mov     ebx, JSAMPROW [ebx+ecx*SIZEOF_JSAMPROW]         ; inptr1
77        mov     edx, JSAMPROW [edx+ecx*SIZEOF_JSAMPROW]         ; inptr2
78        mov     edi, JSAMPROW [edi]                             ; outptr
79
80        pop     ecx                     ; col
81
82        alignx  16,7
83.columnloop:
84        movpic  eax, POINTER [gotptr]   ; load GOT address (eax)
85
86        movdqa    xmm6, XMMWORD [ebx]   ; xmm6=Cb(0123456789ABCDEF)
87        movdqa    xmm7, XMMWORD [edx]   ; xmm7=Cr(0123456789ABCDEF)
88
89        pxor      xmm1,xmm1             ; xmm1=(all 0's)
90        pcmpeqw   xmm3,xmm3
91        psllw     xmm3,7                ; xmm3={0xFF80 0xFF80 0xFF80 0xFF80 ..}
92
93        movdqa    xmm4,xmm6
94        punpckhbw xmm6,xmm1             ; xmm6=Cb(89ABCDEF)=CbH
95        punpcklbw xmm4,xmm1             ; xmm4=Cb(01234567)=CbL
96        movdqa    xmm0,xmm7
97        punpckhbw xmm7,xmm1             ; xmm7=Cr(89ABCDEF)=CrH
98        punpcklbw xmm0,xmm1             ; xmm0=Cr(01234567)=CrL
99
100        paddw     xmm6,xmm3
101        paddw     xmm4,xmm3
102        paddw     xmm7,xmm3
103        paddw     xmm0,xmm3
104
105        ; (Original)
106        ; R = Y                + 1.40200 * Cr
107        ; G = Y - 0.34414 * Cb - 0.71414 * Cr
108        ; B = Y + 1.77200 * Cb
109        ;
110        ; (This implementation)
111        ; R = Y                + 0.40200 * Cr + Cr
112        ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
113        ; B = Y - 0.22800 * Cb + Cb + Cb
114
115        movdqa  xmm5,xmm6               ; xmm5=CbH
116        movdqa  xmm2,xmm4               ; xmm2=CbL
117        paddw   xmm6,xmm6               ; xmm6=2*CbH
118        paddw   xmm4,xmm4               ; xmm4=2*CbL
119        movdqa  xmm1,xmm7               ; xmm1=CrH
120        movdqa  xmm3,xmm0               ; xmm3=CrL
121        paddw   xmm7,xmm7               ; xmm7=2*CrH
122        paddw   xmm0,xmm0               ; xmm0=2*CrL
123
124        pmulhw  xmm6,[GOTOFF(eax,PW_MF0228)]    ; xmm6=(2*CbH * -FIX(0.22800))
125        pmulhw  xmm4,[GOTOFF(eax,PW_MF0228)]    ; xmm4=(2*CbL * -FIX(0.22800))
126        pmulhw  xmm7,[GOTOFF(eax,PW_F0402)]     ; xmm7=(2*CrH * FIX(0.40200))
127        pmulhw  xmm0,[GOTOFF(eax,PW_F0402)]     ; xmm0=(2*CrL * FIX(0.40200))
128
129        paddw   xmm6,[GOTOFF(eax,PW_ONE)]
130        paddw   xmm4,[GOTOFF(eax,PW_ONE)]
131        psraw   xmm6,1                  ; xmm6=(CbH * -FIX(0.22800))
132        psraw   xmm4,1                  ; xmm4=(CbL * -FIX(0.22800))
133        paddw   xmm7,[GOTOFF(eax,PW_ONE)]
134        paddw   xmm0,[GOTOFF(eax,PW_ONE)]
135        psraw   xmm7,1                  ; xmm7=(CrH * FIX(0.40200))
136        psraw   xmm0,1                  ; xmm0=(CrL * FIX(0.40200))
137
138        paddw   xmm6,xmm5
139        paddw   xmm4,xmm2
140        paddw   xmm6,xmm5               ; xmm6=(CbH * FIX(1.77200))=(B-Y)H
141        paddw   xmm4,xmm2               ; xmm4=(CbL * FIX(1.77200))=(B-Y)L
142        paddw   xmm7,xmm1               ; xmm7=(CrH * FIX(1.40200))=(R-Y)H
143        paddw   xmm0,xmm3               ; xmm0=(CrL * FIX(1.40200))=(R-Y)L
144
145        movdqa  XMMWORD [wk(0)], xmm6   ; wk(0)=(B-Y)H
146        movdqa  XMMWORD [wk(1)], xmm7   ; wk(1)=(R-Y)H
147
148        movdqa    xmm6,xmm5
149        movdqa    xmm7,xmm2
150        punpcklwd xmm5,xmm1
151        punpckhwd xmm6,xmm1
152        pmaddwd   xmm5,[GOTOFF(eax,PW_MF0344_F0285)]
153        pmaddwd   xmm6,[GOTOFF(eax,PW_MF0344_F0285)]
154        punpcklwd xmm2,xmm3
155        punpckhwd xmm7,xmm3
156        pmaddwd   xmm2,[GOTOFF(eax,PW_MF0344_F0285)]
157        pmaddwd   xmm7,[GOTOFF(eax,PW_MF0344_F0285)]
158
159        paddd     xmm5,[GOTOFF(eax,PD_ONEHALF)]
160        paddd     xmm6,[GOTOFF(eax,PD_ONEHALF)]
161        psrad     xmm5,SCALEBITS
162        psrad     xmm6,SCALEBITS
163        paddd     xmm2,[GOTOFF(eax,PD_ONEHALF)]
164        paddd     xmm7,[GOTOFF(eax,PD_ONEHALF)]
165        psrad     xmm2,SCALEBITS
166        psrad     xmm7,SCALEBITS
167
168        packssdw  xmm5,xmm6     ; xmm5=CbH*-FIX(0.344)+CrH*FIX(0.285)
169        packssdw  xmm2,xmm7     ; xmm2=CbL*-FIX(0.344)+CrL*FIX(0.285)
170        psubw     xmm5,xmm1     ; xmm5=CbH*-FIX(0.344)+CrH*-FIX(0.714)=(G-Y)H
171        psubw     xmm2,xmm3     ; xmm2=CbL*-FIX(0.344)+CrL*-FIX(0.714)=(G-Y)L
172
173        movdqa  XMMWORD [wk(2)], xmm5   ; wk(2)=(G-Y)H
174
175        mov     al,2                    ; Yctr
176        jmp     short .Yloop_1st
177        alignx  16,7
178
179.Yloop_2nd:
180        movdqa  xmm0, XMMWORD [wk(1)]   ; xmm0=(R-Y)H
181        movdqa  xmm2, XMMWORD [wk(2)]   ; xmm2=(G-Y)H
182        movdqa  xmm4, XMMWORD [wk(0)]   ; xmm4=(B-Y)H
183        alignx  16,7
184
185.Yloop_1st:
186        movdqa  xmm7, XMMWORD [esi]     ; xmm7=Y(0123456789ABCDEF)
187
188        pcmpeqw xmm6,xmm6
189        psrlw   xmm6,BYTE_BIT           ; xmm6={0xFF 0x00 0xFF 0x00 ..}
190        pand    xmm6,xmm7               ; xmm6=Y(02468ACE)=YE
191        psrlw   xmm7,BYTE_BIT           ; xmm7=Y(13579BDF)=YO
192
193        movdqa  xmm1,xmm0               ; xmm1=xmm0=(R-Y)(L/H)
194        movdqa  xmm3,xmm2               ; xmm3=xmm2=(G-Y)(L/H)
195        movdqa  xmm5,xmm4               ; xmm5=xmm4=(B-Y)(L/H)
196
197        paddw     xmm0,xmm6             ; xmm0=((R-Y)+YE)=RE=R(02468ACE)
198        paddw     xmm1,xmm7             ; xmm1=((R-Y)+YO)=RO=R(13579BDF)
199        packuswb  xmm0,xmm0             ; xmm0=R(02468ACE********)
200        packuswb  xmm1,xmm1             ; xmm1=R(13579BDF********)
201
202        paddw     xmm2,xmm6             ; xmm2=((G-Y)+YE)=GE=G(02468ACE)
203        paddw     xmm3,xmm7             ; xmm3=((G-Y)+YO)=GO=G(13579BDF)
204        packuswb  xmm2,xmm2             ; xmm2=G(02468ACE********)
205        packuswb  xmm3,xmm3             ; xmm3=G(13579BDF********)
206
207        paddw     xmm4,xmm6             ; xmm4=((B-Y)+YE)=BE=B(02468ACE)
208        paddw     xmm5,xmm7             ; xmm5=((B-Y)+YO)=BO=B(13579BDF)
209        packuswb  xmm4,xmm4             ; xmm4=B(02468ACE********)
210        packuswb  xmm5,xmm5             ; xmm5=B(13579BDF********)
211
212%if RGB_PIXELSIZE == 3 ; ---------------
213
214        ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
215        ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
216        ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
217        ; xmmG=(** ** ** ** ** ** ** ** **), xmmH=(** ** ** ** ** ** ** ** **)
218
219        punpcklbw xmmA,xmmC     ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
220        punpcklbw xmmE,xmmB     ; xmmE=(20 01 22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F)
221        punpcklbw xmmD,xmmF     ; xmmD=(11 21 13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F)
222
223        movdqa    xmmG,xmmA
224        movdqa    xmmH,xmmA
225        punpcklwd xmmA,xmmE     ; xmmA=(00 10 20 01 02 12 22 03 04 14 24 05 06 16 26 07)
226        punpckhwd xmmG,xmmE     ; xmmG=(08 18 28 09 0A 1A 2A 0B 0C 1C 2C 0D 0E 1E 2E 0F)
227
228        psrldq    xmmH,2        ; xmmH=(02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E -- --)
229        psrldq    xmmE,2        ; xmmE=(22 03 24 05 26 07 28 09 2A 0B 2C 0D 2E 0F -- --)
230
231        movdqa    xmmC,xmmD
232        movdqa    xmmB,xmmD
233        punpcklwd xmmD,xmmH     ; xmmD=(11 21 02 12 13 23 04 14 15 25 06 16 17 27 08 18)
234        punpckhwd xmmC,xmmH     ; xmmC=(19 29 0A 1A 1B 2B 0C 1C 1D 2D 0E 1E 1F 2F -- --)
235
236        psrldq    xmmB,2        ; xmmB=(13 23 15 25 17 27 19 29 1B 2B 1D 2D 1F 2F -- --)
237
238        movdqa    xmmF,xmmE
239        punpcklwd xmmE,xmmB     ; xmmE=(22 03 13 23 24 05 15 25 26 07 17 27 28 09 19 29)
240        punpckhwd xmmF,xmmB     ; xmmF=(2A 0B 1B 2B 2C 0D 1D 2D 2E 0F 1F 2F -- -- -- --)
241
242        pshufd    xmmH,xmmA,0x4E; xmmH=(04 14 24 05 06 16 26 07 00 10 20 01 02 12 22 03)
243        movdqa    xmmB,xmmE
244        punpckldq xmmA,xmmD     ; xmmA=(00 10 20 01 11 21 02 12 02 12 22 03 13 23 04 14)
245        punpckldq xmmE,xmmH     ; xmmE=(22 03 13 23 04 14 24 05 24 05 15 25 06 16 26 07)
246        punpckhdq xmmD,xmmB     ; xmmD=(15 25 06 16 26 07 17 27 17 27 08 18 28 09 19 29)
247
248        pshufd    xmmH,xmmG,0x4E; xmmH=(0C 1C 2C 0D 0E 1E 2E 0F 08 18 28 09 0A 1A 2A 0B)
249        movdqa    xmmB,xmmF
250        punpckldq xmmG,xmmC     ; xmmG=(08 18 28 09 19 29 0A 1A 0A 1A 2A 0B 1B 2B 0C 1C)
251        punpckldq xmmF,xmmH     ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 2C 0D 1D 2D 0E 1E 2E 0F)
252        punpckhdq xmmC,xmmB     ; xmmC=(1D 2D 0E 1E 2E 0F 1F 2F 1F 2F -- -- -- -- -- --)
253
254        punpcklqdq xmmA,xmmE    ; xmmA=(00 10 20 01 11 21 02 12 22 03 13 23 04 14 24 05)
255        punpcklqdq xmmD,xmmG    ; xmmD=(15 25 06 16 26 07 17 27 08 18 28 09 19 29 0A 1A)
256        punpcklqdq xmmF,xmmC    ; xmmF=(2A 0B 1B 2B 0C 1C 2C 0D 1D 2D 0E 1E 2E 0F 1F 2F)
257
258        cmp     ecx, byte SIZEOF_XMMWORD
259        jb      short .column_st32
260
261        test    edi, SIZEOF_XMMWORD-1
262        jnz     short .out1
263        ; --(aligned)-------------------
264        movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
265        movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
266        movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
267        jmp     short .out0
268.out1:  ; --(unaligned)-----------------
269        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
270        movdqu  XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
271        movdqu  XMMWORD [edi+2*SIZEOF_XMMWORD], xmmF
272.out0:
273        add     edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
274        sub     ecx, byte SIZEOF_XMMWORD
275        jz      near .endcolumn
276
277        add     esi, byte SIZEOF_XMMWORD        ; inptr0
278        dec     al                      ; Yctr
279        jnz     near .Yloop_2nd
280
281        add     ebx, byte SIZEOF_XMMWORD        ; inptr1
282        add     edx, byte SIZEOF_XMMWORD        ; inptr2
283        jmp     near .columnloop
284        alignx  16,7
285
286.column_st32:
287        lea     ecx, [ecx+ecx*2]                ; imul ecx, RGB_PIXELSIZE
288        cmp     ecx, byte 2*SIZEOF_XMMWORD
289        jb      short .column_st16
290        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
291        movdqu  XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
292        add     edi, byte 2*SIZEOF_XMMWORD      ; outptr
293        movdqa  xmmA,xmmF
294        sub     ecx, byte 2*SIZEOF_XMMWORD
295        jmp     short .column_st15
296.column_st16:
297        cmp     ecx, byte SIZEOF_XMMWORD
298        jb      short .column_st15
299        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
300        add     edi, byte SIZEOF_XMMWORD        ; outptr
301        movdqa  xmmA,xmmD
302        sub     ecx, byte SIZEOF_XMMWORD
303.column_st15:
304        ; Store the lower 8 bytes of xmmA to the output when it has enough
305        ; space.
306        cmp     ecx, byte SIZEOF_MMWORD
307        jb      short .column_st7
308        movq    XMM_MMWORD [edi], xmmA
309        add     edi, byte SIZEOF_MMWORD
310        sub     ecx, byte SIZEOF_MMWORD
311        psrldq  xmmA, SIZEOF_MMWORD
312.column_st7:
313        ; Store the lower 4 bytes of xmmA to the output when it has enough
314        ; space.
315        cmp     ecx, byte SIZEOF_DWORD
316        jb      short .column_st3
317        movd    XMM_DWORD [edi], xmmA
318        add     edi, byte SIZEOF_DWORD
319        sub     ecx, byte SIZEOF_DWORD
320        psrldq  xmmA, SIZEOF_DWORD
321.column_st3:
322        ; Store the lower 2 bytes of eax to the output when it has enough
323        ; space.
324        movd    eax, xmmA
325        cmp     ecx, byte SIZEOF_WORD
326        jb      short .column_st1
327        mov     WORD [edi], ax
328        add     edi, byte SIZEOF_WORD
329        sub     ecx, byte SIZEOF_WORD
330        shr     eax, 16
331.column_st1:
332        ; Store the lower 1 byte of eax to the output when it has enough
333        ; space.
334        test    ecx, ecx
335        jz      short .endcolumn
336        mov     BYTE [edi], al
337
338%else ; RGB_PIXELSIZE == 4 ; -----------
339
340%ifdef RGBX_FILLER_0XFF
341        pcmpeqb   xmm6,xmm6             ; xmm6=XE=X(02468ACE********)
342        pcmpeqb   xmm7,xmm7             ; xmm7=XO=X(13579BDF********)
343%else
344        pxor      xmm6,xmm6             ; xmm6=XE=X(02468ACE********)
345        pxor      xmm7,xmm7             ; xmm7=XO=X(13579BDF********)
346%endif
347        ; xmmA=(00 02 04 06 08 0A 0C 0E **), xmmB=(01 03 05 07 09 0B 0D 0F **)
348        ; xmmC=(10 12 14 16 18 1A 1C 1E **), xmmD=(11 13 15 17 19 1B 1D 1F **)
349        ; xmmE=(20 22 24 26 28 2A 2C 2E **), xmmF=(21 23 25 27 29 2B 2D 2F **)
350        ; xmmG=(30 32 34 36 38 3A 3C 3E **), xmmH=(31 33 35 37 39 3B 3D 3F **)
351
352        punpcklbw xmmA,xmmC     ; xmmA=(00 10 02 12 04 14 06 16 08 18 0A 1A 0C 1C 0E 1E)
353        punpcklbw xmmE,xmmG     ; xmmE=(20 30 22 32 24 34 26 36 28 38 2A 3A 2C 3C 2E 3E)
354        punpcklbw xmmB,xmmD     ; xmmB=(01 11 03 13 05 15 07 17 09 19 0B 1B 0D 1D 0F 1F)
355        punpcklbw xmmF,xmmH     ; xmmF=(21 31 23 33 25 35 27 37 29 39 2B 3B 2D 3D 2F 3F)
356
357        movdqa    xmmC,xmmA
358        punpcklwd xmmA,xmmE     ; xmmA=(00 10 20 30 02 12 22 32 04 14 24 34 06 16 26 36)
359        punpckhwd xmmC,xmmE     ; xmmC=(08 18 28 38 0A 1A 2A 3A 0C 1C 2C 3C 0E 1E 2E 3E)
360        movdqa    xmmG,xmmB
361        punpcklwd xmmB,xmmF     ; xmmB=(01 11 21 31 03 13 23 33 05 15 25 35 07 17 27 37)
362        punpckhwd xmmG,xmmF     ; xmmG=(09 19 29 39 0B 1B 2B 3B 0D 1D 2D 3D 0F 1F 2F 3F)
363
364        movdqa    xmmD,xmmA
365        punpckldq xmmA,xmmB     ; xmmA=(00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33)
366        punpckhdq xmmD,xmmB     ; xmmD=(04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37)
367        movdqa    xmmH,xmmC
368        punpckldq xmmC,xmmG     ; xmmC=(08 18 28 38 09 19 29 39 0A 1A 2A 3A 0B 1B 2B 3B)
369        punpckhdq xmmH,xmmG     ; xmmH=(0C 1C 2C 3C 0D 1D 2D 3D 0E 1E 2E 3E 0F 1F 2F 3F)
370
371        cmp     ecx, byte SIZEOF_XMMWORD
372        jb      short .column_st32
373
374        test    edi, SIZEOF_XMMWORD-1
375        jnz     short .out1
376        ; --(aligned)-------------------
377        movntdq XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
378        movntdq XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
379        movntdq XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
380        movntdq XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
381        jmp     short .out0
382.out1:  ; --(unaligned)-----------------
383        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
384        movdqu  XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
385        movdqu  XMMWORD [edi+2*SIZEOF_XMMWORD], xmmC
386        movdqu  XMMWORD [edi+3*SIZEOF_XMMWORD], xmmH
387.out0:
388        add     edi, byte RGB_PIXELSIZE*SIZEOF_XMMWORD  ; outptr
389        sub     ecx, byte SIZEOF_XMMWORD
390        jz      near .endcolumn
391
392        add     esi, byte SIZEOF_XMMWORD        ; inptr0
393        dec     al                      ; Yctr
394        jnz     near .Yloop_2nd
395
396        add     ebx, byte SIZEOF_XMMWORD        ; inptr1
397        add     edx, byte SIZEOF_XMMWORD        ; inptr2
398        jmp     near .columnloop
399        alignx  16,7
400
401.column_st32:
402        cmp     ecx, byte SIZEOF_XMMWORD/2
403        jb      short .column_st16
404        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
405        movdqu  XMMWORD [edi+1*SIZEOF_XMMWORD], xmmD
406        add     edi, byte 2*SIZEOF_XMMWORD      ; outptr
407        movdqa  xmmA,xmmC
408        movdqa  xmmD,xmmH
409        sub     ecx, byte SIZEOF_XMMWORD/2
410.column_st16:
411        cmp     ecx, byte SIZEOF_XMMWORD/4
412        jb      short .column_st15
413        movdqu  XMMWORD [edi+0*SIZEOF_XMMWORD], xmmA
414        add     edi, byte SIZEOF_XMMWORD        ; outptr
415        movdqa  xmmA,xmmD
416        sub     ecx, byte SIZEOF_XMMWORD/4
417.column_st15:
418        ; Store two pixels (8 bytes) of xmmA to the output when it has enough
419        ; space.
420        cmp     ecx, byte SIZEOF_XMMWORD/8
421        jb      short .column_st7
422        movq    XMM_MMWORD [edi], xmmA
423        add     edi, byte SIZEOF_XMMWORD/8*4
424        sub     ecx, byte SIZEOF_XMMWORD/8
425        psrldq  xmmA, SIZEOF_XMMWORD/8*4
426.column_st7:
427        ; Store one pixel (4 bytes) of xmmA to the output when it has enough
428        ; space.
429        test    ecx, ecx
430        jz      short .endcolumn
431        movd    XMM_DWORD [edi], xmmA
432
433%endif ; RGB_PIXELSIZE ; ---------------
434
435.endcolumn:
436        sfence          ; flush the write buffer
437
438.return:
439        pop     edi
440        pop     esi
441;       pop     edx             ; need not be preserved
442;       pop     ecx             ; need not be preserved
443        pop     ebx
444        mov     esp,ebp         ; esp <- aligned ebp
445        pop     esp             ; esp <- original ebp
446        pop     ebp
447        ret
448
449; --------------------------------------------------------------------------
450;
451; Upsample and color convert for the case of 2:1 horizontal and 2:1 vertical.
452;
453; GLOBAL(void)
454; jsimd_h2v2_merged_upsample_sse2 (JDIMENSION output_width,
455;                                  JSAMPIMAGE input_buf,
456;                                  JDIMENSION in_row_group_ctr,
457;                                  JSAMPARRAY output_buf);
458;
459
460%define output_width(b) (b)+8                   ; JDIMENSION output_width
461%define input_buf(b)            (b)+12          ; JSAMPIMAGE input_buf
462%define in_row_group_ctr(b)     (b)+16          ; JDIMENSION in_row_group_ctr
463%define output_buf(b)           (b)+20          ; JSAMPARRAY output_buf
464
465        align   16
466        global  EXTN(jsimd_h2v2_merged_upsample_sse2)
467
468EXTN(jsimd_h2v2_merged_upsample_sse2):
469        push    ebp
470        mov     ebp,esp
471        push    ebx
472;       push    ecx             ; need not be preserved
473;       push    edx             ; need not be preserved
474        push    esi
475        push    edi
476
477        mov     eax, POINTER [output_width(ebp)]
478
479        mov     edi, JSAMPIMAGE [input_buf(ebp)]
480        mov     ecx, JDIMENSION [in_row_group_ctr(ebp)]
481        mov     esi, JSAMPARRAY [edi+0*SIZEOF_JSAMPARRAY]
482        mov     ebx, JSAMPARRAY [edi+1*SIZEOF_JSAMPARRAY]
483        mov     edx, JSAMPARRAY [edi+2*SIZEOF_JSAMPARRAY]
484        mov     edi, JSAMPARRAY [output_buf(ebp)]
485        lea     esi, [esi+ecx*SIZEOF_JSAMPROW]
486
487        push    edx                     ; inptr2
488        push    ebx                     ; inptr1
489        push    esi                     ; inptr00
490        mov     ebx,esp
491
492        push    edi                     ; output_buf (outptr0)
493        push    ecx                     ; in_row_group_ctr
494        push    ebx                     ; input_buf
495        push    eax                     ; output_width
496
497        call    near EXTN(jsimd_h2v1_merged_upsample_sse2)
498
499        add     esi, byte SIZEOF_JSAMPROW       ; inptr01
500        add     edi, byte SIZEOF_JSAMPROW       ; outptr1
501        mov     POINTER [ebx+0*SIZEOF_POINTER], esi
502        mov     POINTER [ebx-1*SIZEOF_POINTER], edi
503
504        call    near EXTN(jsimd_h2v1_merged_upsample_sse2)
505
506        add     esp, byte 7*SIZEOF_DWORD
507
508        pop     edi
509        pop     esi
510;       pop     edx             ; need not be preserved
511;       pop     ecx             ; need not be preserved
512        pop     ebx
513        pop     ebp
514        ret
515
516; For some reason, the OS X linker does not honor the request to align the
517; segment unless we do this.
518        align   16
519