2
; jimmxred.asm - reduced-size IDCT (MMX)
4
; Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
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
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
17
; This file contains inverse-DCT routines that produce reduced-size
18
; output: either 4x4 or 2x2 pixels from an 8x8 DCT block.
19
; The following code is based directly on the IJG's original jidctred.c;
20
; see the jidctred.c for more details.
24
%include "jsimdext.inc"
27
; --------------------------------------------------------------------------
32
%define DESCALE_P1_4 (CONST_BITS-PASS1_BITS+1)
33
%define DESCALE_P2_4 (CONST_BITS+PASS1_BITS+3+1)
34
%define DESCALE_P1_2 (CONST_BITS-PASS1_BITS+2)
35
%define DESCALE_P2_2 (CONST_BITS+PASS1_BITS+3+2)
38
F_0_211 equ 1730 ; FIX(0.211164243)
39
F_0_509 equ 4176 ; FIX(0.509795579)
40
F_0_601 equ 4926 ; FIX(0.601344887)
41
F_0_720 equ 5906 ; FIX(0.720959822)
42
F_0_765 equ 6270 ; FIX(0.765366865)
43
F_0_850 equ 6967 ; FIX(0.850430095)
44
F_0_899 equ 7373 ; FIX(0.899976223)
45
F_1_061 equ 8697 ; FIX(1.061594337)
46
F_1_272 equ 10426 ; FIX(1.272758580)
47
F_1_451 equ 11893 ; FIX(1.451774981)
48
F_1_847 equ 15137 ; FIX(1.847759065)
49
F_2_172 equ 17799 ; FIX(2.172734803)
50
F_2_562 equ 20995 ; FIX(2.562915447)
51
F_3_624 equ 29692 ; FIX(3.624509785)
53
; NASM cannot do compile-time arithmetic on floating-point constants.
54
%define DESCALE(x,n) (((x)+(1<<((n)-1)))>>(n))
55
F_0_211 equ DESCALE( 226735879,30-CONST_BITS) ; FIX(0.211164243)
56
F_0_509 equ DESCALE( 547388834,30-CONST_BITS) ; FIX(0.509795579)
57
F_0_601 equ DESCALE( 645689155,30-CONST_BITS) ; FIX(0.601344887)
58
F_0_720 equ DESCALE( 774124714,30-CONST_BITS) ; FIX(0.720959822)
59
F_0_765 equ DESCALE( 821806413,30-CONST_BITS) ; FIX(0.765366865)
60
F_0_850 equ DESCALE( 913142361,30-CONST_BITS) ; FIX(0.850430095)
61
F_0_899 equ DESCALE( 966342111,30-CONST_BITS) ; FIX(0.899976223)
62
F_1_061 equ DESCALE(1139878239,30-CONST_BITS) ; FIX(1.061594337)
63
F_1_272 equ DESCALE(1366614119,30-CONST_BITS) ; FIX(1.272758580)
64
F_1_451 equ DESCALE(1558831516,30-CONST_BITS) ; FIX(1.451774981)
65
F_1_847 equ DESCALE(1984016188,30-CONST_BITS) ; FIX(1.847759065)
66
F_2_172 equ DESCALE(2332956230,30-CONST_BITS) ; FIX(2.172734803)
67
F_2_562 equ DESCALE(2751909506,30-CONST_BITS) ; FIX(2.562915447)
68
F_3_624 equ DESCALE(3891787747,30-CONST_BITS) ; FIX(3.624509785)
71
; --------------------------------------------------------------------------
75
global EXTN(jconst_idct_red_mmx)
77
EXTN(jconst_idct_red_mmx):
79
PW_F184_MF076 times 2 dw F_1_847,-F_0_765
80
PW_F256_F089 times 2 dw F_2_562, F_0_899
81
PW_F106_MF217 times 2 dw F_1_061,-F_2_172
82
PW_MF060_MF050 times 2 dw -F_0_601,-F_0_509
83
PW_F145_MF021 times 2 dw F_1_451,-F_0_211
84
PW_F362_MF127 times 2 dw F_3_624,-F_1_272
85
PW_F085_MF072 times 2 dw F_0_850,-F_0_720
86
PD_DESCALE_P1_4 times 2 dd 1 << (DESCALE_P1_4-1)
87
PD_DESCALE_P2_4 times 2 dd 1 << (DESCALE_P2_4-1)
88
PD_DESCALE_P1_2 times 2 dd 1 << (DESCALE_P1_2-1)
89
PD_DESCALE_P2_2 times 2 dd 1 << (DESCALE_P2_2-1)
90
PB_CENTERJSAMP times 8 db CENTERJSAMPLE
94
; --------------------------------------------------------------------------
98
; Perform dequantization and inverse DCT on one block of coefficients,
99
; producing a reduced-size 4x4 output block.
102
; jsimd_idct_4x4_mmx (void * dct_table, JCOEFPTR coef_block,
103
; JSAMPARRAY output_buf, JDIMENSION output_col)
106
%define dct_table(b) (b)+8 ; void * dct_table
107
%define coef_block(b) (b)+12 ; JCOEFPTR coef_block
108
%define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
109
%define output_col(b) (b)+20 ; JDIMENSION output_col
111
%define original_ebp ebp+0
112
%define wk(i) ebp-(WK_NUM-(i))*SIZEOF_MMWORD ; mmword wk[WK_NUM]
114
%define workspace wk(0)-DCTSIZE2*SIZEOF_JCOEF
115
; JCOEF workspace[DCTSIZE2]
118
global EXTN(jsimd_idct_4x4_mmx)
120
EXTN(jsimd_idct_4x4_mmx):
122
mov eax,esp ; eax = original ebp
124
and esp, byte (-SIZEOF_MMWORD) ; align to 64 bits
126
mov ebp,esp ; ebp = aligned ebp
129
; push ecx ; need not be preserved
130
; push edx ; need not be preserved
134
get_GOT ebx ; get GOT address
136
; ---- Pass 1: process columns from input, store into work array.
138
; mov eax, [original_ebp]
139
mov edx, POINTER [dct_table(eax)] ; quantptr
140
mov esi, JCOEFPTR [coef_block(eax)] ; inptr
141
lea edi, [workspace] ; JCOEF * wsptr
142
mov ecx, DCTSIZE/4 ; ctr
145
%ifndef NO_ZERO_COLUMN_TEST_4X4_MMX
146
mov eax, DWORD [DWBLOCK(1,0,esi,SIZEOF_JCOEF)]
147
or eax, DWORD [DWBLOCK(2,0,esi,SIZEOF_JCOEF)]
150
movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
151
movq mm1, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
152
por mm0, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
153
por mm1, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
154
por mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
155
por mm1, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
162
; -- AC terms all zero
164
movq mm0, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
165
pmullw mm0, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
169
movq mm2,mm0 ; mm0=in0=(00 01 02 03)
170
punpcklwd mm0,mm0 ; mm0=(00 00 01 01)
171
punpckhwd mm2,mm2 ; mm2=(02 02 03 03)
174
punpckldq mm0,mm0 ; mm0=(00 00 00 00)
175
punpckhdq mm1,mm1 ; mm1=(01 01 01 01)
177
punpckldq mm2,mm2 ; mm2=(02 02 02 02)
178
punpckhdq mm3,mm3 ; mm3=(03 03 03 03)
180
movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm0
181
movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm1
182
movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm2
183
movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
191
movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
192
movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
193
pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
194
pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
195
movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
196
movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
197
pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
198
pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
206
pmaddwd mm4,[GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L)
207
pmaddwd mm5,[GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H)
208
pmaddwd mm0,[GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L)
209
pmaddwd mm1,[GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H)
217
pmaddwd mm6,[GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L)
218
pmaddwd mm7,[GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H)
219
pmaddwd mm2,[GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L)
220
pmaddwd mm3,[GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H)
222
paddd mm6,mm4 ; mm6=tmp2L
223
paddd mm7,mm5 ; mm7=tmp2H
224
paddd mm2,mm0 ; mm2=tmp0L
225
paddd mm3,mm1 ; mm3=tmp0H
227
movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L
228
movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H
232
movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
233
movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
234
movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
235
pmullw mm4, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
236
pmullw mm5, MMWORD [MMBLOCK(2,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
237
pmullw mm0, MMWORD [MMBLOCK(6,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
241
punpcklwd mm1,mm4 ; mm1=tmp0L
242
punpckhwd mm2,mm4 ; mm2=tmp0H
243
psrad mm1,(16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1
244
psrad mm2,(16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1
246
movq mm3,mm5 ; mm5=in2=z2
247
punpcklwd mm5,mm0 ; mm0=in6=z3
249
pmaddwd mm5,[GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L
250
pmaddwd mm3,[GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H
254
paddd mm1,mm5 ; mm1=tmp10L
255
paddd mm2,mm3 ; mm2=tmp10H
256
psubd mm4,mm5 ; mm4=tmp12L
257
psubd mm0,mm3 ; mm0=tmp12H
259
; -- Final output stage
263
paddd mm1,mm6 ; mm1=data0L
264
paddd mm2,mm7 ; mm2=data0H
265
psubd mm5,mm6 ; mm5=data3L
266
psubd mm3,mm7 ; mm3=data3H
268
movq mm6,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm6=[PD_DESCALE_P1_4]
272
psrad mm1,DESCALE_P1_4
273
psrad mm2,DESCALE_P1_4
276
psrad mm5,DESCALE_P1_4
277
psrad mm3,DESCALE_P1_4
279
packssdw mm1,mm2 ; mm1=data0=(00 01 02 03)
280
packssdw mm5,mm3 ; mm5=data3=(30 31 32 33)
282
movq mm7, MMWORD [wk(0)] ; mm7=tmp0L
283
movq mm6, MMWORD [wk(1)] ; mm6=tmp0H
287
paddd mm4,mm7 ; mm4=data1L
288
paddd mm0,mm6 ; mm0=data1H
289
psubd mm2,mm7 ; mm2=data2L
290
psubd mm3,mm6 ; mm3=data2H
292
movq mm7,[GOTOFF(ebx,PD_DESCALE_P1_4)] ; mm7=[PD_DESCALE_P1_4]
296
psrad mm4,DESCALE_P1_4
297
psrad mm0,DESCALE_P1_4
300
psrad mm2,DESCALE_P1_4
301
psrad mm3,DESCALE_P1_4
303
packssdw mm4,mm0 ; mm4=data1=(10 11 12 13)
304
packssdw mm2,mm3 ; mm2=data2=(20 21 22 23)
306
movq mm6,mm1 ; transpose coefficients(phase 1)
307
punpcklwd mm1,mm4 ; mm1=(00 10 01 11)
308
punpckhwd mm6,mm4 ; mm6=(02 12 03 13)
309
movq mm7,mm2 ; transpose coefficients(phase 1)
310
punpcklwd mm2,mm5 ; mm2=(20 30 21 31)
311
punpckhwd mm7,mm5 ; mm7=(22 32 23 33)
313
movq mm0,mm1 ; transpose coefficients(phase 2)
314
punpckldq mm1,mm2 ; mm1=(00 10 20 30)
315
punpckhdq mm0,mm2 ; mm0=(01 11 21 31)
316
movq mm3,mm6 ; transpose coefficients(phase 2)
317
punpckldq mm6,mm7 ; mm6=(02 12 22 32)
318
punpckhdq mm3,mm7 ; mm3=(03 13 23 33)
320
movq MMWORD [MMBLOCK(0,0,edi,SIZEOF_JCOEF)], mm1
321
movq MMWORD [MMBLOCK(1,0,edi,SIZEOF_JCOEF)], mm0
322
movq MMWORD [MMBLOCK(2,0,edi,SIZEOF_JCOEF)], mm6
323
movq MMWORD [MMBLOCK(3,0,edi,SIZEOF_JCOEF)], mm3
326
add esi, byte 4*SIZEOF_JCOEF ; coef_block
327
add edx, byte 4*SIZEOF_ISLOW_MULT_TYPE ; quantptr
328
add edi, byte 4*DCTSIZE*SIZEOF_JCOEF ; wsptr
332
; ---- Pass 2: process rows from work array, store into output array.
334
mov eax, [original_ebp]
335
lea esi, [workspace] ; JCOEF * wsptr
336
mov edi, JSAMPARRAY [output_buf(eax)] ; (JSAMPROW *)
337
mov eax, JDIMENSION [output_col(eax)]
341
movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
342
movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
343
movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
344
movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
352
pmaddwd mm4,[GOTOFF(ebx,PW_F256_F089)] ; mm4=(tmp2L)
353
pmaddwd mm5,[GOTOFF(ebx,PW_F256_F089)] ; mm5=(tmp2H)
354
pmaddwd mm0,[GOTOFF(ebx,PW_F106_MF217)] ; mm0=(tmp0L)
355
pmaddwd mm1,[GOTOFF(ebx,PW_F106_MF217)] ; mm1=(tmp0H)
363
pmaddwd mm6,[GOTOFF(ebx,PW_MF060_MF050)] ; mm6=(tmp2L)
364
pmaddwd mm7,[GOTOFF(ebx,PW_MF060_MF050)] ; mm7=(tmp2H)
365
pmaddwd mm2,[GOTOFF(ebx,PW_F145_MF021)] ; mm2=(tmp0L)
366
pmaddwd mm3,[GOTOFF(ebx,PW_F145_MF021)] ; mm3=(tmp0H)
368
paddd mm6,mm4 ; mm6=tmp2L
369
paddd mm7,mm5 ; mm7=tmp2H
370
paddd mm2,mm0 ; mm2=tmp0L
371
paddd mm3,mm1 ; mm3=tmp0H
373
movq MMWORD [wk(0)], mm2 ; wk(0)=tmp0L
374
movq MMWORD [wk(1)], mm3 ; wk(1)=tmp0H
378
movq mm4, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
379
movq mm5, MMWORD [MMBLOCK(2,0,esi,SIZEOF_JCOEF)]
380
movq mm0, MMWORD [MMBLOCK(6,0,esi,SIZEOF_JCOEF)]
384
punpcklwd mm1,mm4 ; mm1=tmp0L
385
punpckhwd mm2,mm4 ; mm2=tmp0H
386
psrad mm1,(16-CONST_BITS-1) ; psrad mm1,16 & pslld mm1,CONST_BITS+1
387
psrad mm2,(16-CONST_BITS-1) ; psrad mm2,16 & pslld mm2,CONST_BITS+1
389
movq mm3,mm5 ; mm5=in2=z2
390
punpcklwd mm5,mm0 ; mm0=in6=z3
392
pmaddwd mm5,[GOTOFF(ebx,PW_F184_MF076)] ; mm5=tmp2L
393
pmaddwd mm3,[GOTOFF(ebx,PW_F184_MF076)] ; mm3=tmp2H
397
paddd mm1,mm5 ; mm1=tmp10L
398
paddd mm2,mm3 ; mm2=tmp10H
399
psubd mm4,mm5 ; mm4=tmp12L
400
psubd mm0,mm3 ; mm0=tmp12H
402
; -- Final output stage
406
paddd mm1,mm6 ; mm1=data0L
407
paddd mm2,mm7 ; mm2=data0H
408
psubd mm5,mm6 ; mm5=data3L
409
psubd mm3,mm7 ; mm3=data3H
411
movq mm6,[GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm6=[PD_DESCALE_P2_4]
415
psrad mm1,DESCALE_P2_4
416
psrad mm2,DESCALE_P2_4
419
psrad mm5,DESCALE_P2_4
420
psrad mm3,DESCALE_P2_4
422
packssdw mm1,mm2 ; mm1=data0=(00 10 20 30)
423
packssdw mm5,mm3 ; mm5=data3=(03 13 23 33)
425
movq mm7, MMWORD [wk(0)] ; mm7=tmp0L
426
movq mm6, MMWORD [wk(1)] ; mm6=tmp0H
430
paddd mm4,mm7 ; mm4=data1L
431
paddd mm0,mm6 ; mm0=data1H
432
psubd mm2,mm7 ; mm2=data2L
433
psubd mm3,mm6 ; mm3=data2H
435
movq mm7,[GOTOFF(ebx,PD_DESCALE_P2_4)] ; mm7=[PD_DESCALE_P2_4]
439
psrad mm4,DESCALE_P2_4
440
psrad mm0,DESCALE_P2_4
443
psrad mm2,DESCALE_P2_4
444
psrad mm3,DESCALE_P2_4
446
packssdw mm4,mm0 ; mm4=data1=(01 11 21 31)
447
packssdw mm2,mm3 ; mm2=data2=(02 12 22 32)
449
movq mm6,[GOTOFF(ebx,PB_CENTERJSAMP)] ; mm6=[PB_CENTERJSAMP]
451
packsswb mm1,mm2 ; mm1=(00 10 20 30 02 12 22 32)
452
packsswb mm4,mm5 ; mm4=(01 11 21 31 03 13 23 33)
456
movq mm7,mm1 ; transpose coefficients(phase 1)
457
punpcklbw mm1,mm4 ; mm1=(00 01 10 11 20 21 30 31)
458
punpckhbw mm7,mm4 ; mm7=(02 03 12 13 22 23 32 33)
460
movq mm0,mm1 ; transpose coefficients(phase 2)
461
punpcklwd mm1,mm7 ; mm1=(00 01 02 03 10 11 12 13)
462
punpckhwd mm0,mm7 ; mm0=(20 21 22 23 30 31 32 33)
464
mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
465
mov esi, JSAMPROW [edi+2*SIZEOF_JSAMPROW]
466
movd DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
467
movd DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
472
mov edx, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
473
mov esi, JSAMPROW [edi+3*SIZEOF_JSAMPROW]
474
movd DWORD [edx+eax*SIZEOF_JSAMPLE], mm1
475
movd DWORD [esi+eax*SIZEOF_JSAMPLE], mm0
477
emms ; empty MMX state
481
; pop edx ; need not be preserved
482
; pop ecx ; need not be preserved
484
mov esp,ebp ; esp <- aligned ebp
485
pop esp ; esp <- original ebp
490
; --------------------------------------------------------------------------
492
; Perform dequantization and inverse DCT on one block of coefficients,
493
; producing a reduced-size 2x2 output block.
496
; jsimd_idct_2x2_mmx (void * dct_table, JCOEFPTR coef_block,
497
; JSAMPARRAY output_buf, JDIMENSION output_col)
500
%define dct_table(b) (b)+8 ; void * dct_table
501
%define coef_block(b) (b)+12 ; JCOEFPTR coef_block
502
%define output_buf(b) (b)+16 ; JSAMPARRAY output_buf
503
%define output_col(b) (b)+20 ; JDIMENSION output_col
506
global EXTN(jsimd_idct_2x2_mmx)
508
EXTN(jsimd_idct_2x2_mmx):
512
; push ecx ; need not be preserved
513
; push edx ; need not be preserved
517
get_GOT ebx ; get GOT address
519
; ---- Pass 1: process columns from input.
521
mov edx, POINTER [dct_table(ebp)] ; quantptr
522
mov esi, JCOEFPTR [coef_block(ebp)] ; inptr
524
; | input: | result: |
525
; | 00 01 ** 03 ** 05 ** 07 | |
526
; | 10 11 ** 13 ** 15 ** 17 | |
527
; | ** ** ** ** ** ** ** ** | |
528
; | 30 31 ** 33 ** 35 ** 37 | A0 A1 A3 A5 A7 |
529
; | ** ** ** ** ** ** ** ** | B0 B1 B3 B5 B7 |
530
; | 50 51 ** 53 ** 55 ** 57 | |
531
; | ** ** ** ** ** ** ** ** | |
532
; | 70 71 ** 73 ** 75 ** 77 | |
536
movq mm0, MMWORD [MMBLOCK(1,0,esi,SIZEOF_JCOEF)]
537
movq mm1, MMWORD [MMBLOCK(3,0,esi,SIZEOF_JCOEF)]
538
pmullw mm0, MMWORD [MMBLOCK(1,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
539
pmullw mm1, MMWORD [MMBLOCK(3,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
540
movq mm2, MMWORD [MMBLOCK(5,0,esi,SIZEOF_JCOEF)]
541
movq mm3, MMWORD [MMBLOCK(7,0,esi,SIZEOF_JCOEF)]
542
pmullw mm2, MMWORD [MMBLOCK(5,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
543
pmullw mm3, MMWORD [MMBLOCK(7,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
545
; mm0=(10 11 ** 13), mm1=(30 31 ** 33)
546
; mm2=(50 51 ** 53), mm3=(70 71 ** 73)
549
pslld mm7,WORD_BIT ; mm7={0x0000 0xFFFF 0x0000 0xFFFF}
551
movq mm4,mm0 ; mm4=(10 11 ** 13)
552
movq mm5,mm2 ; mm5=(50 51 ** 53)
553
punpcklwd mm4,mm1 ; mm4=(10 30 11 31)
554
punpcklwd mm5,mm3 ; mm5=(50 70 51 71)
555
pmaddwd mm4,[GOTOFF(ebx,PW_F362_MF127)]
556
pmaddwd mm5,[GOTOFF(ebx,PW_F085_MF072)]
558
psrld mm0,WORD_BIT ; mm0=(11 -- 13 --)
559
pand mm1,mm7 ; mm1=(-- 31 -- 33)
560
psrld mm2,WORD_BIT ; mm2=(51 -- 53 --)
561
pand mm3,mm7 ; mm3=(-- 71 -- 73)
562
por mm0,mm1 ; mm0=(11 31 13 33)
563
por mm2,mm3 ; mm2=(51 71 53 73)
564
pmaddwd mm0,[GOTOFF(ebx,PW_F362_MF127)]
565
pmaddwd mm2,[GOTOFF(ebx,PW_F085_MF072)]
567
paddd mm4,mm5 ; mm4=tmp0[col0 col1]
569
movq mm6, MMWORD [MMBLOCK(1,1,esi,SIZEOF_JCOEF)]
570
movq mm1, MMWORD [MMBLOCK(3,1,esi,SIZEOF_JCOEF)]
571
pmullw mm6, MMWORD [MMBLOCK(1,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
572
pmullw mm1, MMWORD [MMBLOCK(3,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
573
movq mm3, MMWORD [MMBLOCK(5,1,esi,SIZEOF_JCOEF)]
574
movq mm5, MMWORD [MMBLOCK(7,1,esi,SIZEOF_JCOEF)]
575
pmullw mm3, MMWORD [MMBLOCK(5,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
576
pmullw mm5, MMWORD [MMBLOCK(7,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
578
; mm6=(** 15 ** 17), mm1=(** 35 ** 37)
579
; mm3=(** 55 ** 57), mm5=(** 75 ** 77)
581
psrld mm6,WORD_BIT ; mm6=(15 -- 17 --)
582
pand mm1,mm7 ; mm1=(-- 35 -- 37)
583
psrld mm3,WORD_BIT ; mm3=(55 -- 57 --)
584
pand mm5,mm7 ; mm5=(-- 75 -- 77)
585
por mm6,mm1 ; mm6=(15 35 17 37)
586
por mm3,mm5 ; mm3=(55 75 57 77)
587
pmaddwd mm6,[GOTOFF(ebx,PW_F362_MF127)]
588
pmaddwd mm3,[GOTOFF(ebx,PW_F085_MF072)]
590
paddd mm0,mm2 ; mm0=tmp0[col1 col3]
591
paddd mm6,mm3 ; mm6=tmp0[col5 col7]
595
movq mm1, MMWORD [MMBLOCK(0,0,esi,SIZEOF_JCOEF)]
596
movq mm5, MMWORD [MMBLOCK(0,1,esi,SIZEOF_JCOEF)]
597
pmullw mm1, MMWORD [MMBLOCK(0,0,edx,SIZEOF_ISLOW_MULT_TYPE)]
598
pmullw mm5, MMWORD [MMBLOCK(0,1,edx,SIZEOF_ISLOW_MULT_TYPE)]
600
; mm1=(00 01 ** 03), mm5=(** 05 ** 07)
602
movq mm2,mm1 ; mm2=(00 01 ** 03)
603
pslld mm1,WORD_BIT ; mm1=(-- 00 -- **)
604
psrad mm1,(WORD_BIT-CONST_BITS-2) ; mm1=tmp10[col0 ****]
606
pand mm2,mm7 ; mm2=(-- 01 -- 03)
607
pand mm5,mm7 ; mm5=(-- 05 -- 07)
608
psrad mm2,(WORD_BIT-CONST_BITS-2) ; mm2=tmp10[col1 col3]
609
psrad mm5,(WORD_BIT-CONST_BITS-2) ; mm5=tmp10[col5 col7]
611
; -- Final output stage
614
paddd mm1,mm4 ; mm1=data0[col0 ****]=(A0 **)
615
psubd mm3,mm4 ; mm3=data1[col0 ****]=(B0 **)
616
punpckldq mm1,mm3 ; mm1=(A0 B0)
618
movq mm7,[GOTOFF(ebx,PD_DESCALE_P1_2)] ; mm7=[PD_DESCALE_P1_2]
622
paddd mm2,mm0 ; mm2=data0[col1 col3]=(A1 A3)
623
paddd mm5,mm6 ; mm5=data0[col5 col7]=(A5 A7)
624
psubd mm4,mm0 ; mm4=data1[col1 col3]=(B1 B3)
625
psubd mm3,mm6 ; mm3=data1[col5 col7]=(B5 B7)
628
psrad mm1,DESCALE_P1_2
632
psrad mm2,DESCALE_P1_2
633
psrad mm5,DESCALE_P1_2
636
psrad mm4,DESCALE_P1_2
637
psrad mm3,DESCALE_P1_2
639
; ---- Pass 2: process rows, store into output array.
641
mov edi, JSAMPARRAY [output_buf(ebp)] ; (JSAMPROW *)
642
mov eax, JDIMENSION [output_col(ebp)]
653
packssdw mm2,mm4 ; mm2=(A1 A3 B1 B3)
654
packssdw mm5,mm3 ; mm5=(A5 A7 B5 B7)
655
pmaddwd mm2,[GOTOFF(ebx,PW_F362_MF127)]
656
pmaddwd mm5,[GOTOFF(ebx,PW_F085_MF072)]
658
paddd mm2,mm5 ; mm2=tmp0[row0 row1]
662
pslld mm1,(CONST_BITS+2) ; mm1=tmp10[row0 row1]
664
; -- Final output stage
666
movq mm0,[GOTOFF(ebx,PD_DESCALE_P2_2)] ; mm0=[PD_DESCALE_P2_2]
669
paddd mm1,mm2 ; mm1=data0[row0 row1]=(C0 C1)
670
psubd mm6,mm2 ; mm6=data1[row0 row1]=(D0 D1)
674
psrad mm1,DESCALE_P2_2
675
psrad mm6,DESCALE_P2_2
677
movq mm7,mm1 ; transpose coefficients
678
punpckldq mm1,mm6 ; mm1=(C0 D0)
679
punpckhdq mm7,mm6 ; mm7=(C1 D1)
681
packssdw mm1,mm7 ; mm1=(C0 D0 C1 D1)
682
packsswb mm1,mm1 ; mm1=(C0 D0 C1 D1 C0 D0 C1 D1)
683
paddb mm1,[GOTOFF(ebx,PB_CENTERJSAMP)]
686
movd ebx,mm1 ; ebx=(C0 D0 C1 D1)
687
shr ecx,2*BYTE_BIT ; ecx=(C1 D1 -- --)
689
mov edx, JSAMPROW [edi+0*SIZEOF_JSAMPROW]
690
mov esi, JSAMPROW [edi+1*SIZEOF_JSAMPROW]
691
mov WORD [edx+eax*SIZEOF_JSAMPLE], bx
692
mov WORD [esi+eax*SIZEOF_JSAMPLE], cx
694
emms ; empty MMX state
698
; pop edx ; need not be preserved
699
; pop ecx ; need not be preserved
704
; For some reason, the OS X linker does not honor the request to align the
705
; segment unless we do this.