1;
2; jdcolext.asm - colorspace conversion (MMX)
3;
4; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
5;
6; Based on
7; 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_mmx (JDIMENSION out_width,
27;                            JSAMPIMAGE input_buf, JDIMENSION input_row,
28;                            JSAMPARRAY output_buf, int num_rows)
29;
30
31%define out_width(b)    (b)+8           ; JDIMENSION out_width
32%define input_buf(b)    (b)+12          ; JSAMPIMAGE input_buf
33%define input_row(b)    (b)+16          ; JDIMENSION input_row
34%define output_buf(b)   (b)+20          ; JSAMPARRAY output_buf
35%define num_rows(b)     (b)+24          ; int num_rows
36
37%define original_ebp    ebp+0
38%define wk(i)           ebp-(WK_NUM-(i))*SIZEOF_MMWORD  ; mmword wk[WK_NUM]
39%define WK_NUM          2
40%define gotptr          wk(0)-SIZEOF_POINTER    ; void * gotptr
41
42        align   16
43        global  EXTN(jsimd_ycc_rgb_convert_mmx)
44
45EXTN(jsimd_ycc_rgb_convert_mmx):
46        push    ebp
47        mov     eax,esp                         ; eax = original ebp
48        sub     esp, byte 4
49        and     esp, byte (-SIZEOF_MMWORD)      ; align to 64 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 [out_width(eax)]        ; num_cols
64        test    ecx,ecx
65        jz      near .return
66
67        push    ecx
68
69        mov     edi, JSAMPIMAGE [input_buf(eax)]
70        mov     ecx, JDIMENSION [input_row(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        lea     esi, [esi+ecx*SIZEOF_JSAMPROW]
75        lea     ebx, [ebx+ecx*SIZEOF_JSAMPROW]
76        lea     edx, [edx+ecx*SIZEOF_JSAMPROW]
77
78        pop     ecx
79
80        mov     edi, JSAMPARRAY [output_buf(eax)]
81        mov     eax, INT [num_rows(eax)]
82        test    eax,eax
83        jle     near .return
84        alignx  16,7
85.rowloop:
86        push    eax
87        push    edi
88        push    edx
89        push    ebx
90        push    esi
91        push    ecx                     ; col
92
93        mov     esi, JSAMPROW [esi]     ; inptr0
94        mov     ebx, JSAMPROW [ebx]     ; inptr1
95        mov     edx, JSAMPROW [edx]     ; inptr2
96        mov     edi, JSAMPROW [edi]     ; outptr
97        movpic  eax, POINTER [gotptr]   ; load GOT address (eax)
98        alignx  16,7
99.columnloop:
100
101        movq    mm5, MMWORD [ebx]       ; mm5=Cb(01234567)
102        movq    mm1, MMWORD [edx]       ; mm1=Cr(01234567)
103
104        pcmpeqw mm4,mm4
105        pcmpeqw mm7,mm7
106        psrlw   mm4,BYTE_BIT
107        psllw   mm7,7                   ; mm7={0xFF80 0xFF80 0xFF80 0xFF80}
108        movq    mm0,mm4                 ; mm0=mm4={0xFF 0x00 0xFF 0x00 ..}
109
110        pand    mm4,mm5                 ; mm4=Cb(0246)=CbE
111        psrlw   mm5,BYTE_BIT            ; mm5=Cb(1357)=CbO
112        pand    mm0,mm1                 ; mm0=Cr(0246)=CrE
113        psrlw   mm1,BYTE_BIT            ; mm1=Cr(1357)=CrO
114
115        paddw   mm4,mm7
116        paddw   mm5,mm7
117        paddw   mm0,mm7
118        paddw   mm1,mm7
119
120        ; (Original)
121        ; R = Y                + 1.40200 * Cr
122        ; G = Y - 0.34414 * Cb - 0.71414 * Cr
123        ; B = Y + 1.77200 * Cb
124        ;
125        ; (This implementation)
126        ; R = Y                + 0.40200 * Cr + Cr
127        ; G = Y - 0.34414 * Cb + 0.28586 * Cr - Cr
128        ; B = Y - 0.22800 * Cb + Cb + Cb
129
130        movq    mm2,mm4                 ; mm2=CbE
131        movq    mm3,mm5                 ; mm3=CbO
132        paddw   mm4,mm4                 ; mm4=2*CbE
133        paddw   mm5,mm5                 ; mm5=2*CbO
134        movq    mm6,mm0                 ; mm6=CrE
135        movq    mm7,mm1                 ; mm7=CrO
136        paddw   mm0,mm0                 ; mm0=2*CrE
137        paddw   mm1,mm1                 ; mm1=2*CrO
138
139        pmulhw  mm4,[GOTOFF(eax,PW_MF0228)]     ; mm4=(2*CbE * -FIX(0.22800))
140        pmulhw  mm5,[GOTOFF(eax,PW_MF0228)]     ; mm5=(2*CbO * -FIX(0.22800))
141        pmulhw  mm0,[GOTOFF(eax,PW_F0402)]      ; mm0=(2*CrE * FIX(0.40200))
142        pmulhw  mm1,[GOTOFF(eax,PW_F0402)]      ; mm1=(2*CrO * FIX(0.40200))
143
144        paddw   mm4,[GOTOFF(eax,PW_ONE)]
145        paddw   mm5,[GOTOFF(eax,PW_ONE)]
146        psraw   mm4,1                   ; mm4=(CbE * -FIX(0.22800))
147        psraw   mm5,1                   ; mm5=(CbO * -FIX(0.22800))
148        paddw   mm0,[GOTOFF(eax,PW_ONE)]
149        paddw   mm1,[GOTOFF(eax,PW_ONE)]
150        psraw   mm0,1                   ; mm0=(CrE * FIX(0.40200))
151        psraw   mm1,1                   ; mm1=(CrO * FIX(0.40200))
152
153        paddw   mm4,mm2
154        paddw   mm5,mm3
155        paddw   mm4,mm2                 ; mm4=(CbE * FIX(1.77200))=(B-Y)E
156        paddw   mm5,mm3                 ; mm5=(CbO * FIX(1.77200))=(B-Y)O
157        paddw   mm0,mm6                 ; mm0=(CrE * FIX(1.40200))=(R-Y)E
158        paddw   mm1,mm7                 ; mm1=(CrO * FIX(1.40200))=(R-Y)O
159
160        movq    MMWORD [wk(0)], mm4     ; wk(0)=(B-Y)E
161        movq    MMWORD [wk(1)], mm5     ; wk(1)=(B-Y)O
162
163        movq      mm4,mm2
164        movq      mm5,mm3
165        punpcklwd mm2,mm6
166        punpckhwd mm4,mm6
167        pmaddwd   mm2,[GOTOFF(eax,PW_MF0344_F0285)]
168        pmaddwd   mm4,[GOTOFF(eax,PW_MF0344_F0285)]
169        punpcklwd mm3,mm7
170        punpckhwd mm5,mm7
171        pmaddwd   mm3,[GOTOFF(eax,PW_MF0344_F0285)]
172        pmaddwd   mm5,[GOTOFF(eax,PW_MF0344_F0285)]
173
174        paddd     mm2,[GOTOFF(eax,PD_ONEHALF)]
175        paddd     mm4,[GOTOFF(eax,PD_ONEHALF)]
176        psrad     mm2,SCALEBITS
177        psrad     mm4,SCALEBITS
178        paddd     mm3,[GOTOFF(eax,PD_ONEHALF)]
179        paddd     mm5,[GOTOFF(eax,PD_ONEHALF)]
180        psrad     mm3,SCALEBITS
181        psrad     mm5,SCALEBITS
182
183        packssdw  mm2,mm4       ; mm2=CbE*-FIX(0.344)+CrE*FIX(0.285)
184        packssdw  mm3,mm5       ; mm3=CbO*-FIX(0.344)+CrO*FIX(0.285)
185        psubw     mm2,mm6       ; mm2=CbE*-FIX(0.344)+CrE*-FIX(0.714)=(G-Y)E
186        psubw     mm3,mm7       ; mm3=CbO*-FIX(0.344)+CrO*-FIX(0.714)=(G-Y)O
187
188        movq      mm5, MMWORD [esi]     ; mm5=Y(01234567)
189
190        pcmpeqw   mm4,mm4
191        psrlw     mm4,BYTE_BIT          ; mm4={0xFF 0x00 0xFF 0x00 ..}
192        pand      mm4,mm5               ; mm4=Y(0246)=YE
193        psrlw     mm5,BYTE_BIT          ; mm5=Y(1357)=YO
194
195        paddw     mm0,mm4               ; mm0=((R-Y)E+YE)=RE=(R0 R2 R4 R6)
196        paddw     mm1,mm5               ; mm1=((R-Y)O+YO)=RO=(R1 R3 R5 R7)
197        packuswb  mm0,mm0               ; mm0=(R0 R2 R4 R6 ** ** ** **)
198        packuswb  mm1,mm1               ; mm1=(R1 R3 R5 R7 ** ** ** **)
199
200        paddw     mm2,mm4               ; mm2=((G-Y)E+YE)=GE=(G0 G2 G4 G6)
201        paddw     mm3,mm5               ; mm3=((G-Y)O+YO)=GO=(G1 G3 G5 G7)
202        packuswb  mm2,mm2               ; mm2=(G0 G2 G4 G6 ** ** ** **)
203        packuswb  mm3,mm3               ; mm3=(G1 G3 G5 G7 ** ** ** **)
204
205        paddw     mm4, MMWORD [wk(0)]   ; mm4=(YE+(B-Y)E)=BE=(B0 B2 B4 B6)
206        paddw     mm5, MMWORD [wk(1)]   ; mm5=(YO+(B-Y)O)=BO=(B1 B3 B5 B7)
207        packuswb  mm4,mm4               ; mm4=(B0 B2 B4 B6 ** ** ** **)
208        packuswb  mm5,mm5               ; mm5=(B1 B3 B5 B7 ** ** ** **)
209
210%if RGB_PIXELSIZE == 3 ; ---------------
211
212        ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
213        ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
214        ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
215        ; mmG=(** ** ** ** ** ** ** **), mmH=(** ** ** ** ** ** ** **)
216
217        punpcklbw mmA,mmC               ; mmA=(00 10 02 12 04 14 06 16)
218        punpcklbw mmE,mmB               ; mmE=(20 01 22 03 24 05 26 07)
219        punpcklbw mmD,mmF               ; mmD=(11 21 13 23 15 25 17 27)
220
221        movq      mmG,mmA
222        movq      mmH,mmA
223        punpcklwd mmA,mmE               ; mmA=(00 10 20 01 02 12 22 03)
224        punpckhwd mmG,mmE               ; mmG=(04 14 24 05 06 16 26 07)
225
226        psrlq     mmH,2*BYTE_BIT        ; mmH=(02 12 04 14 06 16 -- --)
227        psrlq     mmE,2*BYTE_BIT        ; mmE=(22 03 24 05 26 07 -- --)
228
229        movq      mmC,mmD
230        movq      mmB,mmD
231        punpcklwd mmD,mmH               ; mmD=(11 21 02 12 13 23 04 14)
232        punpckhwd mmC,mmH               ; mmC=(15 25 06 16 17 27 -- --)
233
234        psrlq     mmB,2*BYTE_BIT        ; mmB=(13 23 15 25 17 27 -- --)
235
236        movq      mmF,mmE
237        punpcklwd mmE,mmB               ; mmE=(22 03 13 23 24 05 15 25)
238        punpckhwd mmF,mmB               ; mmF=(26 07 17 27 -- -- -- --)
239
240        punpckldq mmA,mmD               ; mmA=(00 10 20 01 11 21 02 12)
241        punpckldq mmE,mmG               ; mmE=(22 03 13 23 04 14 24 05)
242        punpckldq mmC,mmF               ; mmC=(15 25 06 16 26 07 17 27)
243
244        cmp     ecx, byte SIZEOF_MMWORD
245        jb      short .column_st16
246
247        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
248        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmE
249        movq    MMWORD [edi+2*SIZEOF_MMWORD], mmC
250
251        sub     ecx, byte SIZEOF_MMWORD
252        jz      short .nextrow
253
254        add     esi, byte SIZEOF_MMWORD                 ; inptr0
255        add     ebx, byte SIZEOF_MMWORD                 ; inptr1
256        add     edx, byte SIZEOF_MMWORD                 ; inptr2
257        add     edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD   ; outptr
258        jmp     near .columnloop
259        alignx  16,7
260
261.column_st16:
262        lea     ecx, [ecx+ecx*2]        ; imul ecx, RGB_PIXELSIZE
263        cmp     ecx, byte 2*SIZEOF_MMWORD
264        jb      short .column_st8
265        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
266        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmE
267        movq    mmA,mmC
268        sub     ecx, byte 2*SIZEOF_MMWORD
269        add     edi, byte 2*SIZEOF_MMWORD
270        jmp     short .column_st4
271.column_st8:
272        cmp     ecx, byte SIZEOF_MMWORD
273        jb      short .column_st4
274        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
275        movq    mmA,mmE
276        sub     ecx, byte SIZEOF_MMWORD
277        add     edi, byte SIZEOF_MMWORD
278.column_st4:
279        movd    eax,mmA
280        cmp     ecx, byte SIZEOF_DWORD
281        jb      short .column_st2
282        mov     DWORD [edi+0*SIZEOF_DWORD], eax
283        psrlq   mmA,DWORD_BIT
284        movd    eax,mmA
285        sub     ecx, byte SIZEOF_DWORD
286        add     edi, byte SIZEOF_DWORD
287.column_st2:
288        cmp     ecx, byte SIZEOF_WORD
289        jb      short .column_st1
290        mov     WORD [edi+0*SIZEOF_WORD], ax
291        shr     eax,WORD_BIT
292        sub     ecx, byte SIZEOF_WORD
293        add     edi, byte SIZEOF_WORD
294.column_st1:
295        cmp     ecx, byte SIZEOF_BYTE
296        jb      short .nextrow
297        mov     BYTE [edi+0*SIZEOF_BYTE], al
298
299%else ; RGB_PIXELSIZE == 4 ; -----------
300
301%ifdef RGBX_FILLER_0XFF
302        pcmpeqb   mm6,mm6               ; mm6=(X0 X2 X4 X6 ** ** ** **)
303        pcmpeqb   mm7,mm7               ; mm7=(X1 X3 X5 X7 ** ** ** **)
304%else
305        pxor      mm6,mm6               ; mm6=(X0 X2 X4 X6 ** ** ** **)
306        pxor      mm7,mm7               ; mm7=(X1 X3 X5 X7 ** ** ** **)
307%endif
308        ; mmA=(00 02 04 06 ** ** ** **), mmB=(01 03 05 07 ** ** ** **)
309        ; mmC=(10 12 14 16 ** ** ** **), mmD=(11 13 15 17 ** ** ** **)
310        ; mmE=(20 22 24 26 ** ** ** **), mmF=(21 23 25 27 ** ** ** **)
311        ; mmG=(30 32 34 36 ** ** ** **), mmH=(31 33 35 37 ** ** ** **)
312
313        punpcklbw mmA,mmC               ; mmA=(00 10 02 12 04 14 06 16)
314        punpcklbw mmE,mmG               ; mmE=(20 30 22 32 24 34 26 36)
315        punpcklbw mmB,mmD               ; mmB=(01 11 03 13 05 15 07 17)
316        punpcklbw mmF,mmH               ; mmF=(21 31 23 33 25 35 27 37)
317
318        movq      mmC,mmA
319        punpcklwd mmA,mmE               ; mmA=(00 10 20 30 02 12 22 32)
320        punpckhwd mmC,mmE               ; mmC=(04 14 24 34 06 16 26 36)
321        movq      mmG,mmB
322        punpcklwd mmB,mmF               ; mmB=(01 11 21 31 03 13 23 33)
323        punpckhwd mmG,mmF               ; mmG=(05 15 25 35 07 17 27 37)
324
325        movq      mmD,mmA
326        punpckldq mmA,mmB               ; mmA=(00 10 20 30 01 11 21 31)
327        punpckhdq mmD,mmB               ; mmD=(02 12 22 32 03 13 23 33)
328        movq      mmH,mmC
329        punpckldq mmC,mmG               ; mmC=(04 14 24 34 05 15 25 35)
330        punpckhdq mmH,mmG               ; mmH=(06 16 26 36 07 17 27 37)
331
332        cmp     ecx, byte SIZEOF_MMWORD
333        jb      short .column_st16
334
335        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
336        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmD
337        movq    MMWORD [edi+2*SIZEOF_MMWORD], mmC
338        movq    MMWORD [edi+3*SIZEOF_MMWORD], mmH
339
340        sub     ecx, byte SIZEOF_MMWORD
341        jz      short .nextrow
342
343        add     esi, byte SIZEOF_MMWORD                 ; inptr0
344        add     ebx, byte SIZEOF_MMWORD                 ; inptr1
345        add     edx, byte SIZEOF_MMWORD                 ; inptr2
346        add     edi, byte RGB_PIXELSIZE*SIZEOF_MMWORD   ; outptr
347        jmp     near .columnloop
348        alignx  16,7
349
350.column_st16:
351        cmp     ecx, byte SIZEOF_MMWORD/2
352        jb      short .column_st8
353        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
354        movq    MMWORD [edi+1*SIZEOF_MMWORD], mmD
355        movq    mmA,mmC
356        movq    mmD,mmH
357        sub     ecx, byte SIZEOF_MMWORD/2
358        add     edi, byte 2*SIZEOF_MMWORD
359.column_st8:
360        cmp     ecx, byte SIZEOF_MMWORD/4
361        jb      short .column_st4
362        movq    MMWORD [edi+0*SIZEOF_MMWORD], mmA
363        movq    mmA,mmD
364        sub     ecx, byte SIZEOF_MMWORD/4
365        add     edi, byte 1*SIZEOF_MMWORD
366.column_st4:
367        cmp     ecx, byte SIZEOF_MMWORD/8
368        jb      short .nextrow
369        movd    DWORD [edi+0*SIZEOF_DWORD], mmA
370
371%endif ; RGB_PIXELSIZE ; ---------------
372
373        alignx  16,7
374
375.nextrow:
376        pop     ecx
377        pop     esi
378        pop     ebx
379        pop     edx
380        pop     edi
381        pop     eax
382
383        add     esi, byte SIZEOF_JSAMPROW
384        add     ebx, byte SIZEOF_JSAMPROW
385        add     edx, byte SIZEOF_JSAMPROW
386        add     edi, byte SIZEOF_JSAMPROW       ; output_buf
387        dec     eax                             ; num_rows
388        jg      near .rowloop
389
390        emms            ; empty MMX state
391
392.return:
393        pop     edi
394        pop     esi
395;       pop     edx             ; need not be preserved
396;       pop     ecx             ; need not be preserved
397        pop     ebx
398        mov     esp,ebp         ; esp <- aligned ebp
399        pop     esp             ; esp <- original ebp
400        pop     ebp
401        ret
402
403; For some reason, the OS X linker does not honor the request to align the
404; segment unless we do this.
405        align   16
406