blob: b6b89121b5fe3b5b0c8e5379377956db9a2198b2 [file] [log] [blame]
DRCf25c0712009-04-03 12:00:51 +00001;
2; jcclrmmx.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
DRCa2d19e12009-04-04 15:22:14 +000019%include "jcolsamp.inc"
20
DRCf25c0712009-04-03 12:00:51 +000021; --------------------------------------------------------------------------
22 SECTION SEG_TEXT
23 BITS 32
24;
25; Convert some rows of samples to the output colorspace.
26;
27; GLOBAL(void)
28; jsimd_rgb_ycc_convert_mmx (JDIMENSION img_width,
29; JSAMPARRAY input_buf, JSAMPIMAGE output_buf,
30; JDIMENSION output_row, int num_rows);
31;
32
33%define img_width(b) (b)+8 ; JDIMENSION img_width
34%define input_buf(b) (b)+12 ; JSAMPARRAY input_buf
35%define output_buf(b) (b)+16 ; JSAMPIMAGE output_buf
36%define output_row(b) (b)+20 ; JDIMENSION output_row
37%define num_rows(b) (b)+24 ; int num_rows
38
39%define original_ebp ebp+0
40%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
41%define WK_NUM 8
42%define gotptr wk(0)-SIZEOF_POINTER ; void * gotptr
43
44 align 16
45 global EXTN(jsimd_rgb_ycc_convert_mmx)
46
47EXTN(jsimd_rgb_ycc_convert_mmx):
48 push ebp
49 mov eax,esp ; eax = original ebp
50 sub esp, byte 4
51 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
52 mov [esp],eax
53 mov ebp,esp ; ebp = aligned ebp
54 lea esp, [wk(0)]
55 pushpic eax ; make a room for GOT address
56 push ebx
57; push ecx ; need not be preserved
58; push edx ; need not be preserved
59 push esi
60 push edi
61
62 get_GOT ebx ; get GOT address
63 movpic POINTER [gotptr], ebx ; save GOT address
64
65 mov ecx, JDIMENSION [img_width(eax)] ; num_cols
66 test ecx,ecx
67 jz near .return
68
69 push ecx
70
71 mov esi, JSAMPIMAGE [output_buf(eax)]
72 mov ecx, JDIMENSION [output_row(eax)]
73 mov edi, JSAMPARRAY [esi+0*SIZEOF_JSAMPARRAY]
74 mov ebx, JSAMPARRAY [esi+1*SIZEOF_JSAMPARRAY]
75 mov edx, JSAMPARRAY [esi+2*SIZEOF_JSAMPARRAY]
76 lea edi, [edi+ecx*SIZEOF_JSAMPROW]
77 lea ebx, [ebx+ecx*SIZEOF_JSAMPROW]
78 lea edx, [edx+ecx*SIZEOF_JSAMPROW]
79
80 pop ecx
81
82 mov esi, JSAMPARRAY [input_buf(eax)]
83 mov eax, INT [num_rows(eax)]
84 test eax,eax
85 jle near .return
86 alignx 16,7
87.rowloop:
88 pushpic eax
89 push edx
90 push ebx
91 push edi
92 push esi
93 push ecx ; col
94
95 mov esi, JSAMPROW [esi] ; inptr
96 mov edi, JSAMPROW [edi] ; outptr0
97 mov ebx, JSAMPROW [ebx] ; outptr1
98 mov edx, JSAMPROW [edx] ; outptr2
99 movpic eax, POINTER [gotptr] ; load GOT address (eax)
100
101 cmp ecx, byte SIZEOF_MMWORD
102 jae short .columnloop
103 alignx 16,7
104
105%if RGB_PIXELSIZE == 3 ; ---------------
106
107.column_ld1:
108 push eax
109 push edx
110 lea ecx,[ecx+ecx*2] ; imul ecx,RGB_PIXELSIZE
111 test cl, SIZEOF_BYTE
112 jz short .column_ld2
113 sub ecx, byte SIZEOF_BYTE
114 xor eax,eax
115 mov al, BYTE [esi+ecx]
116.column_ld2:
117 test cl, SIZEOF_WORD
118 jz short .column_ld4
119 sub ecx, byte SIZEOF_WORD
120 xor edx,edx
121 mov dx, WORD [esi+ecx]
122 shl eax, WORD_BIT
123 or eax,edx
124.column_ld4:
125 movd mmA,eax
126 pop edx
127 pop eax
128 test cl, SIZEOF_DWORD
129 jz short .column_ld8
130 sub ecx, byte SIZEOF_DWORD
131 movd mmG, DWORD [esi+ecx]
132 psllq mmA, DWORD_BIT
133 por mmA,mmG
134.column_ld8:
135 test cl, SIZEOF_MMWORD
136 jz short .column_ld16
137 movq mmG,mmA
138 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD]
139 mov ecx, SIZEOF_MMWORD
140 jmp short .rgb_ycc_cnv
141.column_ld16:
142 test cl, 2*SIZEOF_MMWORD
143 mov ecx, SIZEOF_MMWORD
144 jz short .rgb_ycc_cnv
145 movq mmF,mmA
146 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD]
147 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD]
148 jmp short .rgb_ycc_cnv
149 alignx 16,7
150
151.columnloop:
152 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD]
153 movq mmG, MMWORD [esi+1*SIZEOF_MMWORD]
154 movq mmF, MMWORD [esi+2*SIZEOF_MMWORD]
155
156.rgb_ycc_cnv:
157 ; mmA=(00 10 20 01 11 21 02 12)
158 ; mmG=(22 03 13 23 04 14 24 05)
159 ; mmF=(15 25 06 16 26 07 17 27)
160
161 movq mmD,mmA
162 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 10 20 01)
163 psrlq mmD,4*BYTE_BIT ; mmD=(11 21 02 12 -- -- -- --)
164
165 punpckhbw mmA,mmG ; mmA=(00 04 10 14 20 24 01 05)
166 psllq mmG,4*BYTE_BIT ; mmG=(-- -- -- -- 22 03 13 23)
167
168 punpcklbw mmD,mmF ; mmD=(11 15 21 25 02 06 12 16)
169 punpckhbw mmG,mmF ; mmG=(22 26 03 07 13 17 23 27)
170
171 movq mmE,mmA
172 psllq mmA,4*BYTE_BIT ; mmA=(-- -- -- -- 00 04 10 14)
173 psrlq mmE,4*BYTE_BIT ; mmE=(20 24 01 05 -- -- -- --)
174
175 punpckhbw mmA,mmD ; mmA=(00 02 04 06 10 12 14 16)
176 psllq mmD,4*BYTE_BIT ; mmD=(-- -- -- -- 11 15 21 25)
177
178 punpcklbw mmE,mmG ; mmE=(20 22 24 26 01 03 05 07)
179 punpckhbw mmD,mmG ; mmD=(11 13 15 17 21 23 25 27)
180
181 pxor mmH,mmH
182
183 movq mmC,mmA
184 punpcklbw mmA,mmH ; mmA=(00 02 04 06)
185 punpckhbw mmC,mmH ; mmC=(10 12 14 16)
186
187 movq mmB,mmE
188 punpcklbw mmE,mmH ; mmE=(20 22 24 26)
189 punpckhbw mmB,mmH ; mmB=(01 03 05 07)
190
191 movq mmF,mmD
192 punpcklbw mmD,mmH ; mmD=(11 13 15 17)
193 punpckhbw mmF,mmH ; mmF=(21 23 25 27)
194
195%else ; RGB_PIXELSIZE == 4 ; -----------
196
197.column_ld1:
198 test cl, SIZEOF_MMWORD/8
199 jz short .column_ld2
200 sub ecx, byte SIZEOF_MMWORD/8
201 movd mmA, DWORD [esi+ecx*RGB_PIXELSIZE]
202.column_ld2:
203 test cl, SIZEOF_MMWORD/4
204 jz short .column_ld4
205 sub ecx, byte SIZEOF_MMWORD/4
206 movq mmF,mmA
207 movq mmA, MMWORD [esi+ecx*RGB_PIXELSIZE]
208.column_ld4:
209 test cl, SIZEOF_MMWORD/2
210 mov ecx, SIZEOF_MMWORD
211 jz short .rgb_ycc_cnv
212 movq mmD,mmA
213 movq mmC,mmF
214 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD]
215 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD]
216 jmp short .rgb_ycc_cnv
217 alignx 16,7
218
219.columnloop:
220 movq mmA, MMWORD [esi+0*SIZEOF_MMWORD]
221 movq mmF, MMWORD [esi+1*SIZEOF_MMWORD]
222 movq mmD, MMWORD [esi+2*SIZEOF_MMWORD]
223 movq mmC, MMWORD [esi+3*SIZEOF_MMWORD]
224
225.rgb_ycc_cnv:
226 ; mmA=(00 10 20 30 01 11 21 31)
227 ; mmF=(02 12 22 32 03 13 23 33)
228 ; mmD=(04 14 24 34 05 15 25 35)
229 ; mmC=(06 16 26 36 07 17 27 37)
230
231 movq mmB,mmA
232 punpcklbw mmA,mmF ; mmA=(00 02 10 12 20 22 30 32)
233 punpckhbw mmB,mmF ; mmB=(01 03 11 13 21 23 31 33)
234
235 movq mmG,mmD
236 punpcklbw mmD,mmC ; mmD=(04 06 14 16 24 26 34 36)
237 punpckhbw mmG,mmC ; mmG=(05 07 15 17 25 27 35 37)
238
239 movq mmE,mmA
240 punpcklwd mmA,mmD ; mmA=(00 02 04 06 10 12 14 16)
241 punpckhwd mmE,mmD ; mmE=(20 22 24 26 30 32 34 36)
242
243 movq mmH,mmB
244 punpcklwd mmB,mmG ; mmB=(01 03 05 07 11 13 15 17)
245 punpckhwd mmH,mmG ; mmH=(21 23 25 27 31 33 35 37)
246
247 pxor mmF,mmF
248
249 movq mmC,mmA
250 punpcklbw mmA,mmF ; mmA=(00 02 04 06)
251 punpckhbw mmC,mmF ; mmC=(10 12 14 16)
252
253 movq mmD,mmB
254 punpcklbw mmB,mmF ; mmB=(01 03 05 07)
255 punpckhbw mmD,mmF ; mmD=(11 13 15 17)
256
257 movq mmG,mmE
258 punpcklbw mmE,mmF ; mmE=(20 22 24 26)
259 punpckhbw mmG,mmF ; mmG=(30 32 34 36)
260
261 punpcklbw mmF,mmH
262 punpckhbw mmH,mmH
263 psrlw mmF,BYTE_BIT ; mmF=(21 23 25 27)
264 psrlw mmH,BYTE_BIT ; mmH=(31 33 35 37)
265
266%endif ; RGB_PIXELSIZE ; ---------------
267
268 ; mm0=(R0 R2 R4 R6)=RE, mm2=(G0 G2 G4 G6)=GE, mm4=(B0 B2 B4 B6)=BE
269 ; mm1=(R1 R3 R5 R7)=RO, mm3=(G1 G3 G5 G7)=GO, mm5=(B1 B3 B5 B7)=BO
270
271 ; (Original)
272 ; Y = 0.29900 * R + 0.58700 * G + 0.11400 * B
273 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
274 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
275 ;
276 ; (This implementation)
277 ; Y = 0.29900 * R + 0.33700 * G + 0.11400 * B + 0.25000 * G
278 ; Cb = -0.16874 * R - 0.33126 * G + 0.50000 * B + CENTERJSAMPLE
279 ; Cr = 0.50000 * R - 0.41869 * G - 0.08131 * B + CENTERJSAMPLE
280
281 movq MMWORD [wk(0)], mm0 ; wk(0)=RE
282 movq MMWORD [wk(1)], mm1 ; wk(1)=RO
283 movq MMWORD [wk(2)], mm4 ; wk(2)=BE
284 movq MMWORD [wk(3)], mm5 ; wk(3)=BO
285
286 movq mm6,mm1
287 punpcklwd mm1,mm3
288 punpckhwd mm6,mm3
289 movq mm7,mm1
290 movq mm4,mm6
291 pmaddwd mm1,[GOTOFF(eax,PW_F0299_F0337)] ; mm1=ROL*FIX(0.299)+GOL*FIX(0.337)
292 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=ROH*FIX(0.299)+GOH*FIX(0.337)
293 pmaddwd mm7,[GOTOFF(eax,PW_MF016_MF033)] ; mm7=ROL*-FIX(0.168)+GOL*-FIX(0.331)
294 pmaddwd mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=ROH*-FIX(0.168)+GOH*-FIX(0.331)
295
296 movq MMWORD [wk(4)], mm1 ; wk(4)=ROL*FIX(0.299)+GOL*FIX(0.337)
297 movq MMWORD [wk(5)], mm6 ; wk(5)=ROH*FIX(0.299)+GOH*FIX(0.337)
298
299 pxor mm1,mm1
300 pxor mm6,mm6
301 punpcklwd mm1,mm5 ; mm1=BOL
302 punpckhwd mm6,mm5 ; mm6=BOH
303 psrld mm1,1 ; mm1=BOL*FIX(0.500)
304 psrld mm6,1 ; mm6=BOH*FIX(0.500)
305
306 movq mm5,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm5=[PD_ONEHALFM1_CJ]
307
308 paddd mm7,mm1
309 paddd mm4,mm6
310 paddd mm7,mm5
311 paddd mm4,mm5
312 psrld mm7,SCALEBITS ; mm7=CbOL
313 psrld mm4,SCALEBITS ; mm4=CbOH
314 packssdw mm7,mm4 ; mm7=CbO
315
316 movq mm1, MMWORD [wk(2)] ; mm1=BE
317
318 movq mm6,mm0
319 punpcklwd mm0,mm2
320 punpckhwd mm6,mm2
321 movq mm5,mm0
322 movq mm4,mm6
323 pmaddwd mm0,[GOTOFF(eax,PW_F0299_F0337)] ; mm0=REL*FIX(0.299)+GEL*FIX(0.337)
324 pmaddwd mm6,[GOTOFF(eax,PW_F0299_F0337)] ; mm6=REH*FIX(0.299)+GEH*FIX(0.337)
325 pmaddwd mm5,[GOTOFF(eax,PW_MF016_MF033)] ; mm5=REL*-FIX(0.168)+GEL*-FIX(0.331)
326 pmaddwd mm4,[GOTOFF(eax,PW_MF016_MF033)] ; mm4=REH*-FIX(0.168)+GEH*-FIX(0.331)
327
328 movq MMWORD [wk(6)], mm0 ; wk(6)=REL*FIX(0.299)+GEL*FIX(0.337)
329 movq MMWORD [wk(7)], mm6 ; wk(7)=REH*FIX(0.299)+GEH*FIX(0.337)
330
331 pxor mm0,mm0
332 pxor mm6,mm6
333 punpcklwd mm0,mm1 ; mm0=BEL
334 punpckhwd mm6,mm1 ; mm6=BEH
335 psrld mm0,1 ; mm0=BEL*FIX(0.500)
336 psrld mm6,1 ; mm6=BEH*FIX(0.500)
337
338 movq mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ]
339
340 paddd mm5,mm0
341 paddd mm4,mm6
342 paddd mm5,mm1
343 paddd mm4,mm1
344 psrld mm5,SCALEBITS ; mm5=CbEL
345 psrld mm4,SCALEBITS ; mm4=CbEH
346 packssdw mm5,mm4 ; mm5=CbE
347
348 psllw mm7,BYTE_BIT
349 por mm5,mm7 ; mm5=Cb
350 movq MMWORD [ebx], mm5 ; Save Cb
351
352 movq mm0, MMWORD [wk(3)] ; mm0=BO
353 movq mm6, MMWORD [wk(2)] ; mm6=BE
354 movq mm1, MMWORD [wk(1)] ; mm1=RO
355
356 movq mm4,mm0
357 punpcklwd mm0,mm3
358 punpckhwd mm4,mm3
359 movq mm7,mm0
360 movq mm5,mm4
361 pmaddwd mm0,[GOTOFF(eax,PW_F0114_F0250)] ; mm0=BOL*FIX(0.114)+GOL*FIX(0.250)
362 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BOH*FIX(0.114)+GOH*FIX(0.250)
363 pmaddwd mm7,[GOTOFF(eax,PW_MF008_MF041)] ; mm7=BOL*-FIX(0.081)+GOL*-FIX(0.418)
364 pmaddwd mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BOH*-FIX(0.081)+GOH*-FIX(0.418)
365
366 movq mm3,[GOTOFF(eax,PD_ONEHALF)] ; mm3=[PD_ONEHALF]
367
368 paddd mm0, MMWORD [wk(4)]
369 paddd mm4, MMWORD [wk(5)]
370 paddd mm0,mm3
371 paddd mm4,mm3
372 psrld mm0,SCALEBITS ; mm0=YOL
373 psrld mm4,SCALEBITS ; mm4=YOH
374 packssdw mm0,mm4 ; mm0=YO
375
376 pxor mm3,mm3
377 pxor mm4,mm4
378 punpcklwd mm3,mm1 ; mm3=ROL
379 punpckhwd mm4,mm1 ; mm4=ROH
380 psrld mm3,1 ; mm3=ROL*FIX(0.500)
381 psrld mm4,1 ; mm4=ROH*FIX(0.500)
382
383 movq mm1,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm1=[PD_ONEHALFM1_CJ]
384
385 paddd mm7,mm3
386 paddd mm5,mm4
387 paddd mm7,mm1
388 paddd mm5,mm1
389 psrld mm7,SCALEBITS ; mm7=CrOL
390 psrld mm5,SCALEBITS ; mm5=CrOH
391 packssdw mm7,mm5 ; mm7=CrO
392
393 movq mm3, MMWORD [wk(0)] ; mm3=RE
394
395 movq mm4,mm6
396 punpcklwd mm6,mm2
397 punpckhwd mm4,mm2
398 movq mm1,mm6
399 movq mm5,mm4
400 pmaddwd mm6,[GOTOFF(eax,PW_F0114_F0250)] ; mm6=BEL*FIX(0.114)+GEL*FIX(0.250)
401 pmaddwd mm4,[GOTOFF(eax,PW_F0114_F0250)] ; mm4=BEH*FIX(0.114)+GEH*FIX(0.250)
402 pmaddwd mm1,[GOTOFF(eax,PW_MF008_MF041)] ; mm1=BEL*-FIX(0.081)+GEL*-FIX(0.418)
403 pmaddwd mm5,[GOTOFF(eax,PW_MF008_MF041)] ; mm5=BEH*-FIX(0.081)+GEH*-FIX(0.418)
404
405 movq mm2,[GOTOFF(eax,PD_ONEHALF)] ; mm2=[PD_ONEHALF]
406
407 paddd mm6, MMWORD [wk(6)]
408 paddd mm4, MMWORD [wk(7)]
409 paddd mm6,mm2
410 paddd mm4,mm2
411 psrld mm6,SCALEBITS ; mm6=YEL
412 psrld mm4,SCALEBITS ; mm4=YEH
413 packssdw mm6,mm4 ; mm6=YE
414
415 psllw mm0,BYTE_BIT
416 por mm6,mm0 ; mm6=Y
417 movq MMWORD [edi], mm6 ; Save Y
418
419 pxor mm2,mm2
420 pxor mm4,mm4
421 punpcklwd mm2,mm3 ; mm2=REL
422 punpckhwd mm4,mm3 ; mm4=REH
423 psrld mm2,1 ; mm2=REL*FIX(0.500)
424 psrld mm4,1 ; mm4=REH*FIX(0.500)
425
426 movq mm0,[GOTOFF(eax,PD_ONEHALFM1_CJ)] ; mm0=[PD_ONEHALFM1_CJ]
427
428 paddd mm1,mm2
429 paddd mm5,mm4
430 paddd mm1,mm0
431 paddd mm5,mm0
432 psrld mm1,SCALEBITS ; mm1=CrEL
433 psrld mm5,SCALEBITS ; mm5=CrEH
434 packssdw mm1,mm5 ; mm1=CrE
435
436 psllw mm7,BYTE_BIT
437 por mm1,mm7 ; mm1=Cr
438 movq MMWORD [edx], mm1 ; Save Cr
439
440 sub ecx, byte SIZEOF_MMWORD
441 add esi, byte RGB_PIXELSIZE*SIZEOF_MMWORD ; inptr
442 add edi, byte SIZEOF_MMWORD ; outptr0
443 add ebx, byte SIZEOF_MMWORD ; outptr1
444 add edx, byte SIZEOF_MMWORD ; outptr2
445 cmp ecx, byte SIZEOF_MMWORD
446 jae near .columnloop
447 test ecx,ecx
448 jnz near .column_ld1
449
450 pop ecx ; col
451 pop esi
452 pop edi
453 pop ebx
454 pop edx
455 poppic eax
456
457 add esi, byte SIZEOF_JSAMPROW ; input_buf
458 add edi, byte SIZEOF_JSAMPROW
459 add ebx, byte SIZEOF_JSAMPROW
460 add edx, byte SIZEOF_JSAMPROW
461 dec eax ; num_rows
462 jg near .rowloop
463
464 emms ; empty MMX state
465
466.return:
467 pop edi
468 pop esi
469; pop edx ; need not be preserved
470; pop ecx ; need not be preserved
471 pop ebx
472 mov esp,ebp ; esp <- aligned ebp
473 pop esp ; esp <- original ebp
474 pop ebp
475 ret
476
DRC132b5fd2009-10-08 09:04:56 +0000477; For some reason, the OS X linker does not honor the request to align the
478; segment unless we do this.
479 align 16