2 ; jidctred.asm - reduced-size IDCT (MMX)
4 ; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
6 ; Based on the x86 SIMD extension for IJG JPEG library
7 ; Copyright (C) 1999-2006, MIYASAKA Masaru.
8 ; For conditions of distribution and use, see copyright notice in jsimdext.inc
10 ; This file should be assembled with NASM (Netwide Assembler),
11 ; can *not* be assembled with Microsoft's MASM or any compatible
12 ; assembler (including Borland's Turbo Assembler).
13 ; NASM is available from http://nasm.sourceforge.net/ or
14 ; http://sourceforge.net/project/showfiles.php?group_id=6208
16 ; This file contains inverse-DCT routines that produce reduced-size
17 ; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
18 ; The following code is based directly on the IJG's original jidctred.c;
19 ; see the jidctred.c for more details.
23 %include "jsimdext.inc"
26 ; --------------------------------------------------------------------------
31 %define DESCALE_P1_4 (CONST_BITS-PASS1_BITS+1)
32 %define DESCALE_P2_4 (CONST_BITS+PASS1_BITS+3+1)
33 %define DESCALE_P1_2 (CONST_BITS-PASS1_BITS+2)
34 %define DESCALE_P2_2 (CONST_BITS+PASS1_BITS+3+2)
37 F_0_211 equ 1730 ; FIX(0.211164243)
38 F_0_509 equ 4176 ; FIX(0.509795579)
39 F_0_601 equ 4926 ; FIX(0.601344887)
40 F_0_720 equ 5906 ; FIX(0.720959822)
41 F_0_765 equ 6270 ; FIX(0.765366865)
42 F_0_850 equ 6967 ; FIX(0.850430095)
43 F_0_899 equ 7373 ; FIX(0.899976223)
44 F_1_061 equ 8697 ; FIX(1.061594337)
45 F_1_272 equ 10426 ; FIX(1.272758580)
46 F_1_451 equ 11893 ; FIX(1.451774981)
47 F_1_847 equ 15137 ; FIX(1.847759065)
48 F_2_172 equ 17799 ; FIX(2.172734803)
49 F_2_562 equ 20995 ; FIX(2.562915447)
50 F_3_624 equ 29692 ; FIX(3.624509785)
52 ; NASM cannot do compile-time arithmetic on floating-point constants.
53 %define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
54 F_0_211 equ DESCALE( 226735879,30-CONST_BITS) ; FIX(0.211164243)
55 F_0_509 equ DESCALE( 547388834,30-CONST_BITS) ; FIX(0.509795579)
56 F_0_601 equ DESCALE( 645689155,30-CONST_BITS) ; FIX(0.601344887)
57 F_0_720 equ DESCALE( 774124714,30-CONST_BITS) ; FIX(0.720959822)
58 F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
59 F_0_850 equ DESCALE( 913142361,30-CONST_BITS) ; FIX(0.850430095)
60 F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
61 F_1_061 equ DESCALE(1139878239,30-CONST_BITS) ; FIX(1.061594337)
62 F_1_272 equ DESCALE(1366614119,30-CONST_BITS) ; FIX(1.272758580)
63 F_1_451 equ DESCALE(1558831516,30-CONST_BITS) ; FIX(1.451774981)
64 F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
65 F_2_172 equ DESCALE(2332956230,30-CONST_BITS) ; FIX(2.172734803)
66 F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
67 F_3_624 equ DESCALE(3891787747,30-CONST_BITS) ; FIX(3.624509785)
70 ; --------------------------------------------------------------------------
74 global EXTN(jconst_idct_red_mmx)
76 EXTN(jconst_idct_red_mmx):
78 PW_F184_MF076 times 2 dw F_1_847,-F_0_765
79 PW_F256_F089 times 2 dw F_2_562, F_0_899
80 PW_F106_MF217 times 2 dw F_1_061,-F_2_172
81 PW_MF060_MF050 times 2 dw -F_0_601,-F_0_509
82 PW_F145_MF021 times 2 dw F_1_451,-F_0_211
83 PW_F362_MF127 times 2 dw F_3_624,-F_1_272
84 PW_F085_MF072 times 2 dw F_0_850,-F_0_720
85 PD_DESCALE_P1_4 times 2 dd 1 << (DESCALE_P1_4-1)
86 PD_DESCALE_P2_4 times 2 dd 1 << (DESCALE_P2_4-1)
87 PD_DESCALE_P1_2 times 2 dd 1 << (DESCALE_P1_2-1)
88 PD_DESCALE_P2_2 times 2 dd 1 << (DESCALE_P2_2-1)
89 PB_CENTERJSAMP times 8 db CENTERJSAMPLE
93 ; --------------------------------------------------------------------------
97 ; Perform dequantization and inverse DCT on one block of coefficients,
98 ; producing a reduced-size 4x4 output block.
101 ; jsimd_idct_4x4_mmx (void *dct_table, JCOEFPTR coef_block,
102 ; JSAMPARRAY output_buf, JDIMENSION output_col)
105 %define dct_table(b) (b)+8 ; void *dct_table
106 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
107 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
108 %define output_col(b) (b)+20 ; JDIMENSION output_col
110 %define original_ebp ebp+0
111 %define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
113 %define workspace wk(0)-DCTSIZE2*SIZEOF_JCOEF
114 ; JCOEF workspace[DCTSIZE2]
117 global EXTN(jsimd_idct_4x4_mmx)
119 EXTN(jsimd_idct_4x4_mmx):
121 mov eax,esp ; eax = original ebp
123 and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
125 mov ebp,esp ; ebp = aligned ebp
128 ; push ecx ; need not be preserved
129 ; push edx ; need not be preserved
133 get_GOT ebx ; get GOT address
135 ; ---- Pass 1: process columns from input, store into work array.
137 ; mov eax, [original_ebp]
138 mov edx, POINTER [dct_table(eax)] ; quantptr
139 mov esi, JCOEFPTR [coef_block(eax)] ; inptr
140 lea edi, [workspace] ; JCOEF *wsptr
141 mov ecx, DCTSIZE/4 ; ctr
144 %ifndef NO_ZERO_COLUMN_TEST_4X4_MMX
145 mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
146 or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
149 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
150 movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
151 por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
152 por mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
153 por mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
154 por mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
161 ; -- AC terms all zero
163 movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
164 pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
168 movq mm2,mm0 ; mm0=in0=(00 01 02 03)
169 punpcklwd mm0,mm0 ; mm0=(00 00 01 01)
170 punpckhwd mm2,mm2 ; mm2=(02 02 03 03)
173 punpckldq mm0,mm0 ; mm0=(00 00 00 00)
174 punpckhdq mm1,mm1 ; mm1=(01 01 01 01)
176 punpckldq mm2,mm2 ; mm2=(02 02 02 02)
177 punpckhdq mm3,mm3 ; mm3=(03 03 03 03)
179 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
180 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
181 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
182 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
190 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
191 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
192 pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
193 pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
194 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
195 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
196 pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
197 pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
205 pmaddwd mm4,[GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L)
206 pmaddwd mm5,[GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H)
207 pmaddwd mm0,[GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L)
208 pmaddwd mm1,[GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H)
216 pmaddwd mm6,[GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L)
217 pmaddwd mm7,[GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H)
218 pmaddwd mm2,[GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L)
219 pmaddwd mm3,[GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H)
221 paddd mm6,mm4 ; mm6=tmp2L
222 paddd mm7,mm5 ; mm7=tmp2H
223 paddd mm2,mm0 ; mm2=tmp0L
224 paddd mm3,mm1 ; mm3=tmp0H
226 movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L
227 movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H
231 movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
232 movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
233 movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
234 pmullw mm4, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
235 pmullw mm5, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
236 pmullw mm0, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
240 punpcklwd mm1,mm4 ; mm1=tmp0L
241 punpckhwd mm2,mm4 ; mm2=tmp0H
242 psrad mm1,(16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1
243 psrad mm2,(16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1
245 movq mm3,mm5 ; mm5=in2=z2
246 punpcklwd mm5,mm0 ; mm0=in6=z3
248 pmaddwd mm5,[GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L
249 pmaddwd mm3,[GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H
253 paddd mm1,mm5 ; mm1=tmp10L
254 paddd mm2,mm3 ; mm2=tmp10H
255 psubd mm4,mm5 ; mm4=tmp12L
256 psubd mm0,mm3 ; mm0=tmp12H
258 ; -- Final output stage
262 paddd mm1,mm6 ; mm1=data0L
263 paddd mm2,mm7 ; mm2=data0H
264 psubd mm5,mm6 ; mm5=data3L
265 psubd mm3,mm7 ; mm3=data3H
267 movq mm6,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm6=[PD_DESCALE_P1_4]
271 psrad mm1,DESCALE_P1_4
272 psrad mm2,DESCALE_P1_4
275 psrad mm5,DESCALE_P1_4
276 psrad mm3,DESCALE_P1_4
278 packssdw mm1,mm2 ; mm1=data0=(00 01 02 03)
279 packssdw mm5,mm3 ; mm5=data3=(30 31 32 33)
281 movq mm7, MMWORD [wk(0)] ; mm7=tmp0L
282 movq mm6, MMWORD [wk(1)] ; mm6=tmp0H
286 paddd mm4,mm7 ; mm4=data1L
287 paddd mm0,mm6 ; mm0=data1H
288 psubd mm2,mm7 ; mm2=data2L
289 psubd mm3,mm6 ; mm3=data2H
291 movq mm7,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm7=[PD_DESCALE_P1_4]
295 psrad mm4,DESCALE_P1_4
296 psrad mm0,DESCALE_P1_4
299 psrad mm2,DESCALE_P1_4
300 psrad mm3,DESCALE_P1_4
302 packssdw mm4,mm0 ; mm4=data1=(10 11 12 13)
303 packssdw mm2,mm3 ; mm2=data2=(20 21 22 23)
305 movq mm6,mm1 ; transpose coefficients(phase 1)
306 punpcklwd mm1,mm4 ; mm1=(00 10 01 11)
307 punpckhwd mm6,mm4 ; mm6=(02 12 03 13)
308 movq mm7,mm2 ; transpose coefficients(phase 1)
309 punpcklwd mm2,mm5 ; mm2=(20 30 21 31)
310 punpckhwd mm7,mm5 ; mm7=(22 32 23 33)
312 movq mm0,mm1 ; transpose coefficients(phase 2)
313 punpckldq mm1,mm2 ; mm1=(00 10 20 30)
314 punpckhdq mm0,mm2 ; mm0=(01 11 21 31)
315 movq mm3,mm6 ; transpose coefficients(phase 2)
316 punpckldq mm6,mm7 ; mm6=(02 12 22 32)
317 punpckhdq mm3,mm7 ; mm3=(03 13 23 33)
319 movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm1
320 movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
321 movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm6
322 movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
325 add esi, byte 4*SIZEOF_JCOEF ; coef_block
326 add edx, byte 4*SIZEOF_ISLOW_MULT_TYPE ; quantptr
327 add edi, byte 4*DCTSIZE*SIZEOF_JCOEF ; wsptr
331 ; ---- Pass 2: process rows from work array, store into output array.
333 mov eax, [original_ebp]
334 lea esi, [workspace] ; JCOEF *wsptr
335 mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
336 mov eax, JDIMENSION [output_col(eax)]
340 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
341 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
342 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
343 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
351 pmaddwd mm4,[GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L)
352 pmaddwd mm5,[GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H)
353 pmaddwd mm0,[GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L)
354 pmaddwd mm1,[GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H)
362 pmaddwd mm6,[GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L)
363 pmaddwd mm7,[GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H)
364 pmaddwd mm2,[GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L)
365 pmaddwd mm3,[GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H)
367 paddd mm6,mm4 ; mm6=tmp2L
368 paddd mm7,mm5 ; mm7=tmp2H
369 paddd mm2,mm0 ; mm2=tmp0L
370 paddd mm3,mm1 ; mm3=tmp0H
372 movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L
373 movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H
377 movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
378 movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
379 movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
383 punpcklwd mm1,mm4 ; mm1=tmp0L
384 punpckhwd mm2,mm4 ; mm2=tmp0H
385 psrad mm1,(16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1
386 psrad mm2,(16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1
388 movq mm3,mm5 ; mm5=in2=z2
389 punpcklwd mm5,mm0 ; mm0=in6=z3
391 pmaddwd mm5,[GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L
392 pmaddwd mm3,[GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H
396 paddd mm1,mm5 ; mm1=tmp10L
397 paddd mm2,mm3 ; mm2=tmp10H
398 psubd mm4,mm5 ; mm4=tmp12L
399 psubd mm0,mm3 ; mm0=tmp12H
401 ; -- Final output stage
405 paddd mm1,mm6 ; mm1=data0L
406 paddd mm2,mm7 ; mm2=data0H
407 psubd mm5,mm6 ; mm5=data3L
408 psubd mm3,mm7 ; mm3=data3H
410 movq mm6,[GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm6=[PD_DESCALE_P2_4]
414 psrad mm1,DESCALE_P2_4
415 psrad mm2,DESCALE_P2_4
418 psrad mm5,DESCALE_P2_4
419 psrad mm3,DESCALE_P2_4
421 packssdw mm1,mm2 ; mm1=data0=(00 10 20 30)
422 packssdw mm5,mm3 ; mm5=data3=(03 13 23 33)
424 movq mm7, MMWORD [wk(0)] ; mm7=tmp0L
425 movq mm6, MMWORD [wk(1)] ; mm6=tmp0H
429 paddd mm4,mm7 ; mm4=data1L
430 paddd mm0,mm6 ; mm0=data1H
431 psubd mm2,mm7 ; mm2=data2L
432 psubd mm3,mm6 ; mm3=data2H
434 movq mm7,[GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm7=[PD_DESCALE_P2_4]
438 psrad mm4,DESCALE_P2_4
439 psrad mm0,DESCALE_P2_4
442 psrad mm2,DESCALE_P2_4
443 psrad mm3,DESCALE_P2_4
445 packssdw mm4,mm0 ; mm4=data1=(01 11 21 31)
446 packssdw mm2,mm3 ; mm2=data2=(02 12 22 32)
448 movq mm6,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm6=[PB_CENTERJSAMP]
450 packsswb mm1,mm2 ; mm1=(00 10 20 30 02 12 22 32)
451 packsswb mm4,mm5 ; mm4=(01 11 21 31 03 13 23 33)
455 movq mm7,mm1 ; transpose coefficients(phase 1)
456 punpcklbw mm1,mm4 ; mm1=(00 01 10 11 20 21 30 31)
457 punpckhbw mm7,mm4 ; mm7=(02 03 12 13 22 23 32 33)
459 movq mm0,mm1 ; transpose coefficients(phase 2)
460 punpcklwd mm1,mm7 ; mm1=(00 01 02 03 10 11 12 13)
461 punpckhwd mm0,mm7 ; mm0=(20 21 22 23 30 31 32 33)
463 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
464 mov esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
465 movd DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
466 movd DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
471 mov edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
472 mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
473 movd DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
474 movd DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
476 emms ; empty MMX state
480 ; pop edx ; need not be preserved
481 ; pop ecx ; need not be preserved
483 mov esp,ebp ; esp <- aligned ebp
484 pop esp ; esp <- original ebp
489 ; --------------------------------------------------------------------------
491 ; Perform dequantization and inverse DCT on one block of coefficients,
492 ; producing a reduced-size 2x2 output block.
495 ; jsimd_idct_2x2_mmx (void *dct_table, JCOEFPTR coef_block,
496 ; JSAMPARRAY output_buf, JDIMENSION output_col)
499 %define dct_table(b) (b)+8 ; void *dct_table
500 %define coef_block(b) (b)+12 ; JCOEFPTR coef_block
501 %define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
502 %define output_col(b) (b)+20 ; JDIMENSION output_col
505 global EXTN(jsimd_idct_2x2_mmx)
507 EXTN(jsimd_idct_2x2_mmx):
511 ; push ecx ; need not be preserved
512 ; push edx ; need not be preserved
516 get_GOT ebx ; get GOT address
518 ; ---- Pass 1: process columns from input.
520 mov edx, POINTER [dct_table(ebp)] ; quantptr
521 mov esi, JCOEFPTR [coef_block(ebp)] ; inptr
523 ; | input: | result: |
524 ; | 00 01 ** 03 ** 05 ** 07 | |
525 ; | 10 11 ** 13 ** 15 ** 17 | |
526 ; | ** ** ** ** ** ** ** ** | |
527 ; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
528 ; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
529 ; | 50 51 ** 53 ** 55 ** 57 | |
530 ; | ** ** ** ** ** ** ** ** | |
531 ; | 70 71 ** 73 ** 75 ** 77 | |
535 movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
536 movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
537 pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
538 pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
539 movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
540 movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
541 pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
542 pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
544 ; mm0=(10 11 ** 13), mm1=(30 31 ** 33)
545 ; mm2=(50 51 ** 53), mm3=(70 71 ** 73)
548 pslld mm7,WORD_BIT ; mm7={0x0000 0xFFFF 0x0000 0xFFFF}
550 movq mm4,mm0 ; mm4=(10 11 ** 13)
551 movq mm5,mm2 ; mm5=(50 51 ** 53)
552 punpcklwd mm4,mm1 ; mm4=(10 30 11 31)
553 punpcklwd mm5,mm3 ; mm5=(50 70 51 71)
554 pmaddwd mm4,[GOTOFF(ebx,PW_F362_MF127)]
555 pmaddwd mm5,[GOTOFF(ebx,PW_F085_MF072)]
557 psrld mm0,WORD_BIT ; mm0=(11 -- 13 --)
558 pand mm1,mm7 ; mm1=(-- 31 -- 33)
559 psrld mm2,WORD_BIT ; mm2=(51 -- 53 --)
560 pand mm3,mm7 ; mm3=(-- 71 -- 73)
561 por mm0,mm1 ; mm0=(11 31 13 33)
562 por mm2,mm3 ; mm2=(51 71 53 73)
563 pmaddwd mm0,[GOTOFF(ebx,PW_F362_MF127)]
564 pmaddwd mm2,[GOTOFF(ebx,PW_F085_MF072)]
566 paddd mm4,mm5 ; mm4=tmp0[col0 col1]
568 movq mm6, MMWORD [MMBLOCK(1,1,esi,SIZEOF_JCOEF)]
569 movq mm1, MMWORD [MMBLOCK(3,1,esi,SIZEOF_JCOEF)]
570 pmullw mm6, MMWORD [MMBLOCK(1,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
571 pmullw mm1, MMWORD [MMBLOCK(3,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
572 movq mm3, MMWORD [MMBLOCK(5,1,esi,SIZEOF_JCOEF)]
573 movq mm5, MMWORD [MMBLOCK(7,1,esi,SIZEOF_JCOEF)]
574 pmullw mm3, MMWORD [MMBLOCK(5,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
575 pmullw mm5, MMWORD [MMBLOCK(7,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
577 ; mm6=(** 15 ** 17), mm1=(** 35 ** 37)
578 ; mm3=(** 55 ** 57), mm5=(** 75 ** 77)
580 psrld mm6,WORD_BIT ; mm6=(15 -- 17 --)
581 pand mm1,mm7 ; mm1=(-- 35 -- 37)
582 psrld mm3,WORD_BIT ; mm3=(55 -- 57 --)
583 pand mm5,mm7 ; mm5=(-- 75 -- 77)
584 por mm6,mm1 ; mm6=(15 35 17 37)
585 por mm3,mm5 ; mm3=(55 75 57 77)
586 pmaddwd mm6,[GOTOFF(ebx,PW_F362_MF127)]
587 pmaddwd mm3,[GOTOFF(ebx,PW_F085_MF072)]
589 paddd mm0,mm2 ; mm0=tmp0[col1 col3]
590 paddd mm6,mm3 ; mm6=tmp0[col5 col7]
594 movq mm1, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
595 movq mm5, MMWORD [MMBLOCK(0,1,esi,SIZEOF_JCOEF)]
596 pmullw mm1, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
597 pmullw mm5, MMWORD [MMBLOCK(0,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
599 ; mm1=(00 01 ** 03), mm5=(** 05 ** 07)
601 movq mm2,mm1 ; mm2=(00 01 ** 03)
602 pslld mm1,WORD_BIT ; mm1=(-- 00 -- **)
603 psrad mm1,(WORD_BIT-CONST_BITS-2) ; mm1=tmp10[col0 ****]
605 pand mm2,mm7 ; mm2=(-- 01 -- 03)
606 pand mm5,mm7 ; mm5=(-- 05 -- 07)
607 psrad mm2,(WORD_BIT-CONST_BITS-2) ; mm2=tmp10[col1 col3]
608 psrad mm5,(WORD_BIT-CONST_BITS-2) ; mm5=tmp10[col5 col7]
610 ; -- Final output stage
613 paddd mm1,mm4 ; mm1=data0[col0 ****]=(A0 **)
614 psubd mm3,mm4 ; mm3=data1[col0 ****]=(B0 **)
615 punpckldq mm1,mm3 ; mm1=(A0 B0)
617 movq mm7,[GOTOFF(ebx,PD_DESCALE_P1_2)] ; mm7=[PD_DESCALE_P1_2]
621 paddd mm2,mm0 ; mm2=data0[col1 col3]=(A1 A3)
622 paddd mm5,mm6 ; mm5=data0[col5 col7]=(A5 A7)
623 psubd mm4,mm0 ; mm4=data1[col1 col3]=(B1 B3)
624 psubd mm3,mm6 ; mm3=data1[col5 col7]=(B5 B7)
627 psrad mm1,DESCALE_P1_2
631 psrad mm2,DESCALE_P1_2
632 psrad mm5,DESCALE_P1_2
635 psrad mm4,DESCALE_P1_2
636 psrad mm3,DESCALE_P1_2
638 ; ---- Pass 2: process rows, store into output array.
640 mov edi, JSAMPARRAY [output_buf(ebp)] ; (JSAMPROW *)
641 mov eax, JDIMENSION [output_col(ebp)]
652 packssdw mm2,mm4 ; mm2=(A1 A3 B1 B3)
653 packssdw mm5,mm3 ; mm5=(A5 A7 B5 B7)
654 pmaddwd mm2,[GOTOFF(ebx,PW_F362_MF127)]
655 pmaddwd mm5,[GOTOFF(ebx,PW_F085_MF072)]
657 paddd mm2,mm5 ; mm2=tmp0[row0 row1]
661 pslld mm1,(CONST_BITS+2) ; mm1=tmp10[row0 row1]
663 ; -- Final output stage
665 movq mm0,[GOTOFF(ebx,PD_DESCALE_P2_2)] ; mm0=[PD_DESCALE_P2_2]
668 paddd mm1,mm2 ; mm1=data0[row0 row1]=(C0 C1)
669 psubd mm6,mm2 ; mm6=data1[row0 row1]=(D0 D1)
673 psrad mm1,DESCALE_P2_2
674 psrad mm6,DESCALE_P2_2
676 movq mm7,mm1 ; transpose coefficients
677 punpckldq mm1,mm6 ; mm1=(C0 D0)
678 punpckhdq mm7,mm6 ; mm7=(C1 D1)
680 packssdw mm1,mm7 ; mm1=(C0 D0 C1 D1)
681 packsswb mm1,mm1 ; mm1=(C0 D0 C1 D1 C0 D0 C1 D1)
682 paddb mm1,[GOTOFF(ebx,PB_CENTERJSAMP)]
685 movd ebx,mm1 ; ebx=(C0 D0 C1 D1)
686 shr ecx,2*BYTE_BIT ; ecx=(C1 D1 -- --)
688 mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
689 mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
690 mov WORD [edx+eax*SIZEOF_JSAMPLE], bx
691 mov WORD [esi+eax*SIZEOF_JSAMPLE], cx
693 emms ; empty MMX state
697 ; pop edx ; need not be preserved
698 ; pop ecx ; need not be preserved
703 ; For some reason, the OS X linker does not honor the request to align the
704 ; segment unless we do this.