42 |
; * * |
; * * |
43 |
; * Revision history: * |
; * Revision history: * |
44 |
; * * |
; * * |
45 |
|
; * 14.06.2002 mmx dequant4_* funcs revamped -Skal- * |
46 |
; * 22.01.2002 initial version * |
; * 22.01.2002 initial version * |
47 |
; * * |
; * * |
48 |
; ******************************************************************************/ |
; ******************************************************************************/ |
54 |
|
|
55 |
bits 32 |
bits 32 |
56 |
|
|
57 |
section .data |
%ifdef FORMAT_COFF |
58 |
|
SECTION .data data |
59 |
|
%else |
60 |
|
SECTION .data data align=8 |
61 |
|
%endif |
62 |
|
|
63 |
|
|
64 |
%macro cglobal 1 |
%macro cglobal 1 |
65 |
%ifdef PREFIX |
%ifdef PREFIX |
70 |
%endif |
%endif |
71 |
%endmacro |
%endmacro |
72 |
|
|
73 |
|
%macro cextern 1 |
74 |
|
%ifdef PREFIX |
75 |
|
extern _%1 |
76 |
|
%define %1 _%1 |
77 |
|
%else |
78 |
|
extern %1 |
79 |
|
%endif |
80 |
|
%endmacro |
81 |
|
|
82 |
mmx_one times 4 dw 1 |
mmx_one times 4 dw 1 |
83 |
|
|
84 |
;=========================================================================== |
;=========================================================================== |
128 |
|
|
129 |
;=========================================================================== |
;=========================================================================== |
130 |
; |
; |
131 |
; default intra matrix |
; intra matrix |
132 |
; |
; |
133 |
;=========================================================================== |
;=========================================================================== |
134 |
|
|
135 |
mmx_intra_matrix |
cextern intra_matrix |
136 |
dw 8, 17, 18, 19 |
cextern intra_matrix_fix |
|
dw 21, 23, 25, 27 |
|
|
dw 17, 18, 19, 21 |
|
|
dw 23, 25, 27, 28 |
|
|
dw 20, 21, 22, 23 |
|
|
dw 24, 26, 28, 30 |
|
|
dw 21, 22, 23, 24 |
|
|
dw 26, 28, 30, 32 |
|
|
dw 22, 23, 24, 26 |
|
|
dw 28, 30, 32, 35 |
|
|
dw 23, 24, 26, 28 |
|
|
dw 30, 32, 35, 38 |
|
|
dw 25, 26, 28, 30 |
|
|
dw 32, 35, 38, 41 |
|
|
dw 27, 28, 30, 32 |
|
|
dw 35, 38, 41, 45 |
|
|
|
|
|
%macro MMX_FIX 4 |
|
|
dw (1 << 16) / (%1) + 1, (1 << 16) / (%2) + 1, (1 << 16) / (%3) + 1, (1 << 16) / (%4) + 1 |
|
|
%endmacro |
|
|
|
|
|
mmx_intra_matrix_fix |
|
|
MMX_FIX 8, 17, 18, 19 |
|
|
MMX_FIX 21, 23, 25, 27 |
|
|
MMX_FIX 17, 18, 19, 21 |
|
|
MMX_FIX 23, 25, 27, 28 |
|
|
MMX_FIX 20, 21, 22, 23 |
|
|
MMX_FIX 24, 26, 28, 30 |
|
|
MMX_FIX 21, 22, 23, 24 |
|
|
MMX_FIX 26, 28, 30, 32 |
|
|
MMX_FIX 22, 23, 24, 26 |
|
|
MMX_FIX 28, 30, 32, 35 |
|
|
MMX_FIX 23, 24, 26, 28 |
|
|
MMX_FIX 30, 32, 35, 38 |
|
|
MMX_FIX 25, 26, 28, 30 |
|
|
MMX_FIX 32, 35, 38, 41 |
|
|
MMX_FIX 27, 28, 30, 32 |
|
|
MMX_FIX 35, 38, 41, 45 |
|
|
|
|
137 |
|
|
138 |
;=========================================================================== |
;=========================================================================== |
139 |
; |
; |
140 |
; default inter matrix |
; inter matrix |
141 |
; |
; |
142 |
;=========================================================================== |
;=========================================================================== |
143 |
|
|
144 |
mmx_inter_matrix |
cextern inter_matrix |
145 |
dw 16,17,18,19 |
cextern inter_matrix_fix |
146 |
dw 20,21,22,23 |
|
|
dw 17,18,19,20 |
|
|
dw 21,22,23,24 |
|
|
dw 18,19,20,21 |
|
|
dw 22,23,24,25 |
|
|
dw 19,20,21,22 |
|
|
dw 23,24,26,27 |
|
|
dw 20,21,22,23 |
|
|
dw 25,26,27,28 |
|
|
dw 21,22,23,24 |
|
|
dw 26,27,28,30 |
|
|
dw 22,23,24,26 |
|
|
dw 27,28,30,31 |
|
|
dw 23,24,25,27 |
|
|
dw 28,30,31,33 |
|
|
|
|
|
|
|
|
mmx_inter_matrix_fix |
|
|
MMX_FIX 16,17,18,19 |
|
|
MMX_FIX 20,21,22,23 |
|
|
MMX_FIX 17,18,19,20 |
|
|
MMX_FIX 21,22,23,24 |
|
|
MMX_FIX 18,19,20,21 |
|
|
MMX_FIX 22,23,24,25 |
|
|
MMX_FIX 19,20,21,22 |
|
|
MMX_FIX 23,24,26,27 |
|
|
MMX_FIX 20,21,22,23 |
|
|
MMX_FIX 25,26,27,28 |
|
|
MMX_FIX 21,22,23,24 |
|
|
MMX_FIX 26,27,28,30 |
|
|
MMX_FIX 22,23,24,26 |
|
|
MMX_FIX 27,28,30,31 |
|
|
MMX_FIX 23,24,25,27 |
|
|
MMX_FIX 28,30,31,33 |
|
147 |
|
|
148 |
%define VM18P 3 |
%define VM18P 3 |
149 |
%define VM18Q 4 |
%define VM18Q 4 |
150 |
|
|
151 |
|
|
152 |
;=========================================================================== |
;=========================================================================== |
153 |
; |
; |
154 |
; quantd table |
; quantd table |
195 |
|
|
196 |
;=========================================================================== |
;=========================================================================== |
197 |
; |
; |
|
; multiple by matrix table |
|
|
; |
|
|
;=========================================================================== |
|
|
|
|
|
%macro MMX_MUL 4 |
|
|
dw %1 |
|
|
dw %2 |
|
|
dw %3 |
|
|
dw %4 |
|
|
%endmacro |
|
|
|
|
|
default_inter_matrix_mul |
|
|
MMX_MUL 16,17,18,19 |
|
|
MMX_MUL 20,21,22,23 |
|
|
MMX_MUL 17,18,19,20 |
|
|
MMX_MUL 21,22,23,24 |
|
|
MMX_MUL 18,19,20,21 |
|
|
MMX_MUL 22,23,24,25 |
|
|
MMX_MUL 19,20,21,22 |
|
|
MMX_MUL 23,24,26,27 |
|
|
MMX_MUL 20,21,22,23 |
|
|
MMX_MUL 25,26,27,28 |
|
|
MMX_MUL 21,22,23,24 |
|
|
MMX_MUL 26,27,28,30 |
|
|
MMX_MUL 22,23,24,26 |
|
|
MMX_MUL 27,28,30,31 |
|
|
MMX_MUL 23,24,25,27 |
|
|
MMX_MUL 28,30,31,33 |
|
|
|
|
|
|
|
|
default_intra_matrix_mul |
|
|
MMX_MUL 8,17,18,19 |
|
|
MMX_MUL 21,23,25,27 |
|
|
MMX_MUL 17,18,19,21 |
|
|
MMX_MUL 23,25,27,28 |
|
|
MMX_MUL 20,21,22,23 |
|
|
MMX_MUL 24,26,28,30 |
|
|
MMX_MUL 21,22,23,24 |
|
|
MMX_MUL 26,28,30,32 |
|
|
MMX_MUL 22,23,24,26 |
|
|
MMX_MUL 28,30,32,35 |
|
|
MMX_MUL 23,24,26,28 |
|
|
MMX_MUL 30,32,35,38 |
|
|
MMX_MUL 25,26,28,30 |
|
|
MMX_MUL 32,35,38,41 |
|
|
MMX_MUL 27,28,30,32 |
|
|
MMX_MUL 35,38,41,45 |
|
|
|
|
|
|
|
|
;=========================================================================== |
|
|
; |
|
198 |
; multiple by 2Q table |
; multiple by 2Q table |
199 |
; |
; |
200 |
;=========================================================================== |
;=========================================================================== |
243 |
;=========================================================================== |
;=========================================================================== |
244 |
|
|
245 |
align 16 |
align 16 |
246 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
|
247 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
248 |
|
mmx_32768_minus_2048 times 4 dw (32768-2048) |
249 |
|
mmx_2047 times 4 dw 2047 |
250 |
|
mmx_minus_2048 times 4 dw (-2048) |
251 |
|
zero times 4 dw 0 |
252 |
|
|
253 |
section .text |
section .text |
254 |
|
|
303 |
psllw mm0, 4 ; level << 4 |
psllw mm0, 4 ; level << 4 |
304 |
psllw mm3, 4 ; |
psllw mm3, 4 ; |
305 |
|
|
306 |
movq mm2, [mmx_intra_matrix + 8*ecx] |
movq mm2, [intra_matrix + 8*ecx] |
307 |
psrlw mm2, 1 ; intra_matrix[i]>>1 |
psrlw mm2, 1 ; intra_matrix[i]>>1 |
308 |
paddw mm0, mm2 |
paddw mm0, mm2 |
309 |
|
|
310 |
movq mm2, [mmx_intra_matrix_fix + ecx*8] |
movq mm2, [intra_matrix_fix + ecx*8] |
311 |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
312 |
|
|
313 |
movq mm2, [mmx_intra_matrix + 8*ecx + 8] |
movq mm2, [intra_matrix + 8*ecx + 8] |
314 |
psrlw mm2, 1 |
psrlw mm2, 1 |
315 |
paddw mm3, mm2 |
paddw mm3, mm2 |
316 |
|
|
317 |
movq mm2, [mmx_intra_matrix_fix + ecx*8 + 8] |
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
318 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
319 |
|
|
320 |
paddw mm0, mm5 ; + quantd |
paddw mm0, mm5 ; + quantd |
382 |
psllw mm0, 4 |
psllw mm0, 4 |
383 |
psllw mm3, 4 |
psllw mm3, 4 |
384 |
|
|
385 |
movq mm2, [mmx_intra_matrix + 8*ecx] |
movq mm2, [intra_matrix + 8*ecx] |
386 |
psrlw mm2, 1 |
psrlw mm2, 1 |
387 |
paddw mm0, mm2 |
paddw mm0, mm2 |
388 |
|
|
389 |
movq mm2, [mmx_intra_matrix_fix + ecx*8] |
movq mm2, [intra_matrix_fix + ecx*8] |
390 |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
391 |
|
|
392 |
movq mm2, [mmx_intra_matrix + 8*ecx + 8] |
movq mm2, [intra_matrix + 8*ecx + 8] |
393 |
psrlw mm2, 1 |
psrlw mm2, 1 |
394 |
paddw mm3, mm2 |
paddw mm3, mm2 |
395 |
|
|
396 |
movq mm2, [mmx_intra_matrix_fix + ecx*8 + 8] |
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
397 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
398 |
|
|
399 |
paddw mm0, mm5 |
paddw mm0, mm5 |
435 |
psllw mm0, 4 |
psllw mm0, 4 |
436 |
psllw mm3, 4 |
psllw mm3, 4 |
437 |
|
|
438 |
movq mm2, [mmx_intra_matrix + 8*ecx] |
movq mm2, [intra_matrix + 8*ecx] |
439 |
psrlw mm2, 1 |
psrlw mm2, 1 |
440 |
paddw mm0, mm2 |
paddw mm0, mm2 |
441 |
|
|
442 |
movq mm2, [mmx_intra_matrix_fix + ecx*8] |
movq mm2, [intra_matrix_fix + ecx*8] |
443 |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + intra_matrix[i]>>1) / intra_matrix[i] |
444 |
|
|
445 |
movq mm2, [mmx_intra_matrix + 8*ecx + 8] |
movq mm2, [intra_matrix + 8*ecx + 8] |
446 |
psrlw mm2, 1 |
psrlw mm2, 1 |
447 |
paddw mm3, mm2 |
paddw mm3, mm2 |
448 |
|
|
449 |
movq mm2, [mmx_intra_matrix_fix + ecx*8 + 8] |
movq mm2, [intra_matrix_fix + ecx*8 + 8] |
450 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
451 |
|
|
452 |
paddw mm0, mm5 |
paddw mm0, mm5 |
517 |
psllw mm0, 4 |
psllw mm0, 4 |
518 |
psllw mm3, 4 |
psllw mm3, 4 |
519 |
|
|
520 |
movq mm2, [mmx_inter_matrix + 8*ecx] |
movq mm2, [inter_matrix + 8*ecx] |
521 |
psrlw mm2, 1 |
psrlw mm2, 1 |
522 |
paddw mm0, mm2 |
paddw mm0, mm2 |
523 |
|
|
524 |
movq mm2, [mmx_inter_matrix_fix + ecx*8] |
movq mm2, [inter_matrix_fix + ecx*8] |
525 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
526 |
|
|
527 |
movq mm2, [mmx_inter_matrix + 8*ecx + 8] |
movq mm2, [inter_matrix + 8*ecx + 8] |
528 |
psrlw mm2, 1 |
psrlw mm2, 1 |
529 |
paddw mm3, mm2 |
paddw mm3, mm2 |
530 |
|
|
531 |
movq mm2, [mmx_inter_matrix_fix + ecx*8 + 8] |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
532 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
533 |
|
|
534 |
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
pmulhw mm0, mm7 ; mm0 = (mm0 / 2Q) >> 16 |
581 |
psllw mm0, 4 |
psllw mm0, 4 |
582 |
psllw mm3, 4 |
psllw mm3, 4 |
583 |
|
|
584 |
movq mm2, [mmx_inter_matrix + 8*ecx] |
movq mm2, [inter_matrix + 8*ecx] |
585 |
psrlw mm2, 1 |
psrlw mm2, 1 |
586 |
paddw mm0, mm2 |
paddw mm0, mm2 |
587 |
|
|
588 |
movq mm2, [mmx_inter_matrix_fix + ecx*8] |
movq mm2, [inter_matrix_fix + ecx*8] |
589 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
590 |
|
|
591 |
movq mm2, [mmx_inter_matrix + 8*ecx + 8] |
movq mm2, [inter_matrix + 8*ecx + 8] |
592 |
psrlw mm2, 1 |
psrlw mm2, 1 |
593 |
paddw mm3, mm2 |
paddw mm3, mm2 |
594 |
|
|
595 |
movq mm2, [mmx_inter_matrix_fix + ecx*8 + 8] |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
596 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
597 |
|
|
598 |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
634 |
psllw mm0, 4 |
psllw mm0, 4 |
635 |
psllw mm3, 4 |
psllw mm3, 4 |
636 |
|
|
637 |
movq mm2, [mmx_inter_matrix + 8*ecx] |
movq mm2, [inter_matrix + 8*ecx] |
638 |
psrlw mm2, 1 |
psrlw mm2, 1 |
639 |
paddw mm0, mm2 |
paddw mm0, mm2 |
640 |
|
|
641 |
movq mm2, [mmx_inter_matrix_fix + ecx*8] |
movq mm2, [inter_matrix_fix + ecx*8] |
642 |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
pmulhw mm0, mm2 ; (level<<4 + inter_matrix[i]>>1) / inter_matrix[i] |
643 |
|
|
644 |
movq mm2, [mmx_inter_matrix + 8*ecx + 8] |
movq mm2, [inter_matrix + 8*ecx + 8] |
645 |
psrlw mm2, 1 |
psrlw mm2, 1 |
646 |
paddw mm3, mm2 |
paddw mm3, mm2 |
647 |
|
|
648 |
movq mm2, [mmx_inter_matrix_fix + ecx*8 + 8] |
movq mm2, [inter_matrix_fix + ecx*8 + 8] |
649 |
pmulhw mm3, mm2 |
pmulhw mm3, mm2 |
650 |
|
|
651 |
psrlw mm0, 2 ; mm0 >>= 1 (/2) |
psrlw mm0, 2 ; mm0 >>= 1 (/2) |
677 |
; |
; |
678 |
;=========================================================================== |
;=========================================================================== |
679 |
|
|
680 |
align 16 |
; Note: in order to saturate 'easily', we pre-shift the quantifier |
681 |
cglobal dequant4_intra_mmx |
; by 4. Then, the high-word of (coeff[]*matrix[i]*quant) are used to |
682 |
dequant4_intra_mmx |
; build a saturating mask. It is non-zero only when an overflow occured. |
683 |
|
; We thus avoid packing/unpacking toward double-word. |
684 |
|
; Moreover, we perform the mult (matrix[i]*quant) first, instead of, e.g., |
685 |
|
; (coeff[i]*matrix[i]). This is less prone to overflow if coeff[] are not |
686 |
|
; checked. Input ranges are: coeff in [-127,127], inter_matrix in [1..255],a |
687 |
|
; and quant in [1..31]. |
688 |
|
; |
689 |
|
; The original loop is: |
690 |
|
; |
691 |
|
%if 0 |
692 |
|
movq mm0, [ecx+8*eax + 8*16] ; mm0 = coeff[i] |
693 |
|
pxor mm1, mm1 |
694 |
|
pcmpgtw mm1, mm0 |
695 |
|
pxor mm0, mm1 ; change sign if negative |
696 |
|
psubw mm0, mm1 ; -> mm0 = abs(coeff[i]), mm1 = sign of coeff[i] |
697 |
|
|
698 |
|
movq mm2, mm7 ; mm2 = quant |
699 |
|
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant. |
700 |
|
|
701 |
|
movq mm6, mm2 |
702 |
|
pmulhw mm2, mm0 ; high of coeff*(matrix*quant) (should be 0 if no overflow) |
703 |
|
pmullw mm0, mm6 ; low of coeff*(matrix*quant) |
704 |
|
|
705 |
|
pxor mm5, mm5 |
706 |
|
pcmpgtw mm2, mm5 ; otherflow? |
707 |
|
psrlw mm2, 5 ; =0 if no clamp, 2047 otherwise |
708 |
|
psrlw mm0, 5 |
709 |
|
paddw mm0, mm1 ; start restoring sign |
710 |
|
por mm0, mm2 ; saturate to 2047 if needed |
711 |
|
pxor mm0, mm1 ; finish negating back |
712 |
|
|
713 |
push esi |
movq [edx + 8*eax + 8*16], mm0 ; data[i] |
714 |
push edi |
add eax, 1 |
715 |
|
%endif |
716 |
|
|
717 |
mov edi, [esp + 8 + 4] ; data |
;******************************************************************** |
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
718 |
|
|
719 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
align 16 |
720 |
|
cglobal dequant4_intra_mmx |
721 |
|
dequant4_intra_mmx: |
722 |
|
|
723 |
xor eax, eax |
mov edx, [esp+4] ; data |
724 |
|
mov ecx, [esp+8] ; coeff |
725 |
|
mov eax, [esp+12] ; quant |
726 |
|
|
727 |
|
movq mm7, [mmx_mul_quant + eax*8 - 8] |
728 |
|
mov eax, -16 ; to keep aligned, we regularly process coeff[0] |
729 |
|
psllw mm7, 2 ; << 2. See comment. |
730 |
|
pxor mm6, mm6 ; this is a NOP |
731 |
|
|
732 |
align 16 |
align 16 |
733 |
.loop |
.loop |
734 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax + 8*16] ; mm0 = c = coeff[i] |
735 |
|
movq mm3, [ecx+8*eax + 8*16 +8]; mm3 = c' = coeff[i+1] |
736 |
|
pxor mm1, mm1 |
737 |
|
pxor mm4, mm4 |
738 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) |
739 |
|
movq mm2, mm7 ; mm2 = quant |
740 |
|
|
741 |
pxor mm1, mm1 ; mm1 = 0 |
pcmpgtw mm4, mm3 ; mm4 = sgn(c') |
742 |
pcmpeqw mm1, mm0 ; mm1 = (0 == mm0) |
pmullw mm2, [intra_matrix + 8*eax + 8*16 ] ; matrix[i]*quant |
743 |
|
|
744 |
pxor mm2, mm2 ; mm2 = 0 |
pxor mm0, mm1 ; negate if negative |
745 |
pcmpgtw mm2, mm0 ; mm2 = (0 > mm0) |
pxor mm3, mm4 ; negate if negative |
|
pxor mm0, mm2 ; mm0 = |mm0| |
|
|
psubw mm0, mm2 ; displace |
|
746 |
|
|
747 |
pmullw mm0, mm7 ; mm0 *= quant |
psubw mm0, mm1 |
748 |
|
psubw mm3, mm4 |
749 |
|
|
750 |
movq mm3, [default_intra_matrix_mul + 8*eax] |
; we're short on register, here. Poor pairing... |
751 |
|
|
752 |
movq mm4, mm0 ; |
movq mm5, mm2 |
753 |
pmullw mm0, mm3 ; mm0 = low(mm0 * mm3) |
pmullw mm2, mm0 ; low of coeff*(matrix*quant) |
|
pmulhw mm3, mm4 ; mm3 = high(mm0 * mm3) |
|
754 |
|
|
755 |
movq mm4, mm0 ; mm0,mm4 = unpack(mm3, mm0) |
pmulhw mm0, mm5 ; high of coeff*(matrix*quant) |
756 |
punpcklwd mm0, mm3 ; |
movq mm5, mm7 ; mm2 = quant |
|
punpckhwd mm4, mm3 ; |
|
|
psrld mm0, 3 ; mm0,mm4 /= 8 |
|
|
psrld mm4, 3 ; |
|
|
packssdw mm0, mm4 ; mm0 = pack(mm4, mm0) |
|
757 |
|
|
758 |
pxor mm0, mm2 ; mm0 *= sign(mm0) |
pmullw mm5, [intra_matrix + 8*eax + 8*16 +8] ; matrix[i+1]*quant |
|
psubw mm0, mm2 ; undisplace |
|
|
pandn mm1, mm0 ; mm1 = ~(iszero) & mm0 |
|
759 |
|
|
760 |
%ifdef SATURATE |
movq mm6, mm5 |
761 |
movq mm2, [mmx_32767_minus_2047] |
add eax,2 ; z-flag will be tested later |
|
movq mm6, [mmx_32768_minus_2048] |
|
|
paddsw mm1, mm2 |
|
|
psubsw mm1, mm2 |
|
|
psubsw mm1, mm6 |
|
|
paddsw mm1, mm6 |
|
|
%endif |
|
762 |
|
|
763 |
movq [edi + 8*eax], mm1 ; [data] = mm0 |
pmullw mm6, mm3 ; low of coeff*(matrix*quant) |
764 |
|
pmulhw mm3, mm5 ; high of coeff*(matrix*quant) |
765 |
|
|
766 |
add eax, 1 |
pcmpgtw mm0, [zero] |
767 |
cmp eax, 16 |
paddusw mm2, mm0 |
768 |
jnz near .loop |
psrlw mm2, 5 |
769 |
|
|
770 |
mov ax, [esi] ; ax = data[0] |
pcmpgtw mm3, [zero] |
771 |
imul ax, [esp + 8 + 16] ; eax = data[0] * dcscalar |
paddusw mm6, mm3 |
772 |
mov [edi], ax ; data[0] = ax |
psrlw mm6, 5 |
|
|
|
|
%ifdef SATURATE |
|
|
cmp ax, -2048 |
|
|
jl .set_n2048 |
|
|
cmp ax, 2047 |
|
|
jg .set_2047 |
|
|
%endif |
|
773 |
|
|
774 |
pop edi |
pxor mm2, mm1 ; start negating back |
775 |
pop esi |
pxor mm6, mm4 ; start negating back |
|
ret |
|
776 |
|
|
777 |
%ifdef SATURATE |
psubusw mm1, mm0 |
778 |
.set_n2048 |
psubusw mm4, mm3 |
|
mov word [edi], -2048 |
|
|
pop edi |
|
|
pop esi |
|
|
ret |
|
779 |
|
|
780 |
.set_2047 |
psubw mm2, mm1 ; finish negating back |
781 |
mov word [edi], 2047 |
psubw mm6, mm4 ; finish negating back |
|
pop edi |
|
|
pop esi |
|
782 |
|
|
783 |
ret |
movq [edx + 8*eax + 8*16 -2*8 ], mm2 ; data[i] |
784 |
%endif |
movq [edx + 8*eax + 8*16 -2*8 +8], mm6 ; data[i+1] |
785 |
|
|
786 |
|
jnz near .loop |
787 |
|
|
788 |
|
; deal with DC |
789 |
|
|
790 |
|
movd mm0, [ecx] |
791 |
|
pmullw mm0, [esp+16] ; dcscalar |
792 |
|
movq mm2, [mmx_32767_minus_2047] |
793 |
|
paddsw mm0, mm2 |
794 |
|
psubsw mm0, mm2 |
795 |
|
movq mm2, [mmx_32768_minus_2048] |
796 |
|
psubsw mm0, mm2 |
797 |
|
paddsw mm0, mm2 |
798 |
|
movd eax, mm0 |
799 |
|
mov [edx], ax |
800 |
|
|
801 |
|
ret |
802 |
|
|
803 |
;=========================================================================== |
;=========================================================================== |
804 |
; |
; |
808 |
; |
; |
809 |
;=========================================================================== |
;=========================================================================== |
810 |
|
|
811 |
|
; Note: We use (2*c + sgn(c) - sgn(-c)) as multiplier |
812 |
|
; so we handle the 3 cases: c<0, c==0, and c>0 in one shot. |
813 |
|
; sgn(x) is the result of 'pcmpgtw 0,x': 0 if x>=0, -1 if x<0. |
814 |
|
; It's mixed with the extraction of the absolute value. |
815 |
|
|
816 |
align 16 |
align 16 |
817 |
cglobal dequant4_inter_mmx |
cglobal dequant4_inter_mmx |
818 |
dequant4_inter_mmx |
dequant4_inter_mmx: |
819 |
|
|
820 |
push esi |
mov edx, [esp+ 4] ; data |
821 |
push edi |
mov ecx, [esp+ 8] ; coeff |
822 |
|
mov eax, [esp+12] ; quant |
|
mov edi, [esp + 8 + 4] ; data |
|
|
mov esi, [esp + 8 + 8] ; coeff |
|
|
mov eax, [esp + 8 + 12] ; quant |
|
823 |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
movq mm7, [mmx_mul_quant + eax*8 - 8] |
824 |
movq mm6, [mmx_one] |
mov eax, -16 |
825 |
xor eax, eax |
paddw mm7, mm7 ; << 1 |
826 |
pxor mm5, mm5 ; mismatch sum |
pxor mm6, mm6 ; mismatch sum |
|
|
|
827 |
|
|
828 |
align 16 |
align 16 |
829 |
.loop |
.loop |
830 |
movq mm0, [esi + 8*eax] ; mm0 = [coeff] |
movq mm0, [ecx+8*eax + 8*16 ] ; mm0 = coeff[i] |
831 |
|
movq mm2, [ecx+8*eax + 8*16 +8] ; mm2 = coeff[i+1] |
832 |
|
add eax,2 |
833 |
|
|
834 |
|
pxor mm1, mm1 |
835 |
|
pxor mm3, mm3 |
836 |
|
pcmpgtw mm1, mm0 ; mm1 = sgn(c) (preserved) |
837 |
|
pcmpgtw mm3, mm2 ; mm3 = sgn(c') (preserved) |
838 |
|
paddsw mm0, mm1 ; c += sgn(c) |
839 |
|
paddsw mm2, mm3 ; c += sgn(c') |
840 |
|
paddw mm0, mm0 ; c *= 2 |
841 |
|
paddw mm2, mm2 ; c'*= 2 |
842 |
|
|
843 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm4, mm4 |
844 |
pcmpeqw mm1, mm0 ; mm1 = (0 == mm0) |
pxor mm5, mm5 |
845 |
|
psubw mm4, mm0 ; -c |
846 |
pxor mm2, mm2 ; mm2 = 0 |
psubw mm5, mm2 ; -c' |
847 |
pcmpgtw mm2, mm0 ; mm2 = (0 > mm0) |
psraw mm4, 16 ; mm4 = sgn(-c) |
848 |
pxor mm0, mm2 ; mm0 = |mm0| |
psraw mm5, 16 ; mm5 = sgn(-c') |
849 |
psubw mm0, mm2 ; displace |
psubsw mm0, mm4 ; c -= sgn(-c) |
850 |
|
psubsw mm2, mm5 ; c' -= sgn(-c') |
851 |
psllw mm0, 1 ; |
pxor mm0, mm1 ; finish changing sign if needed |
852 |
paddsw mm0, mm6 ; mm0 = 2*mm0 + 1 |
pxor mm2, mm3 ; finish changing sign if needed |
853 |
pmullw mm0, mm7 ; mm0 *= quant |
|
854 |
|
; we're short on register, here. Poor pairing... |
855 |
movq mm3, [default_inter_matrix_mul + 8*eax] |
|
856 |
|
movq mm4, mm7 ; (matrix*quant) |
857 |
movq mm4, mm0 |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8] |
858 |
pmullw mm0, mm3 ; mm0 = low(mm0 * mm3) |
movq mm5, mm4 |
859 |
pmulhw mm3, mm4 ; mm3 = high(mm0 * mm3) |
pmulhw mm5, mm0 ; high of c*(matrix*quant) |
860 |
|
pmullw mm0, mm4 ; low of c*(matrix*quant) |
861 |
movq mm4, mm0 ; mm0,mm4 = unpack(mm3, mm0) |
|
862 |
punpcklwd mm0, mm3 ; |
movq mm4, mm7 ; (matrix*quant) |
863 |
punpckhwd mm4, mm3 ; |
pmullw mm4, [inter_matrix + 8*eax + 8*16 -2*8 + 8] |
864 |
|
|
865 |
psrad mm0, 4 ; mm0,mm4 /= 16 |
pcmpgtw mm5, [zero] |
866 |
psrad mm4, 4 ; |
paddusw mm0, mm5 |
867 |
packssdw mm0, mm4 ; mm0 = pack(mm4, mm0) |
psrlw mm0, 5 |
868 |
|
pxor mm0, mm1 ; start restoring sign |
869 |
pxor mm0, mm2 ; mm0 *= sign(mm0) |
psubusw mm1, mm5 |
870 |
psubw mm0, mm2 ; undisplace |
|
871 |
pandn mm1, mm0 ; mm1 = ~(iszero) & mm0 |
movq mm5, mm4 |
872 |
|
pmulhw mm5, mm2 ; high of c*(matrix*quant) |
873 |
|
pmullw mm2, mm4 ; low of c*(matrix*quant) |
874 |
;%ifdef SATURATE |
psubw mm0, mm1 ; finish restoring sign |
875 |
movq mm2, [mmx_32767_minus_2047] |
|
876 |
movq mm4, [mmx_32768_minus_2048] |
pcmpgtw mm5, [zero] |
877 |
paddsw mm1, mm2 |
paddusw mm2, mm5 |
878 |
psubsw mm1, mm2 |
psrlw mm2, 5 |
879 |
psubsw mm1, mm4 |
pxor mm2, mm3 ; start restoring sign |
880 |
paddsw mm1, mm4 |
psubusw mm3, mm5 |
881 |
;%endif |
psubw mm2, mm3 ; finish restoring sign |
882 |
|
|
883 |
pxor mm5, mm1 ; mismatch |
pxor mm6, mm0 ; mismatch control |
884 |
|
movq [edx + 8*eax + 8*16 -2*8 ], mm0 ; data[i] |
885 |
movq [edi + 8*eax], mm1 ; [data] = mm0 |
pxor mm6, mm2 ; mismatch control |
886 |
|
movq [edx + 8*eax + 8*16 -2*8 +8], mm2 ; data[i+1] |
887 |
|
|
|
add eax, 1 |
|
|
cmp eax, 16 |
|
888 |
jnz near .loop |
jnz near .loop |
889 |
|
|
890 |
; mismatch control |
; mismatch control |
891 |
|
|
892 |
movq mm0, mm5 |
movq mm0, mm6 |
|
movq mm1, mm5 |
|
|
movq mm2, mm5 |
|
893 |
psrlq mm0, 48 |
psrlq mm0, 48 |
894 |
|
movq mm1, mm6 |
895 |
|
movq mm2, mm6 |
896 |
psrlq mm1, 32 |
psrlq mm1, 32 |
897 |
|
pxor mm6, mm0 |
898 |
psrlq mm2, 16 |
psrlq mm2, 16 |
899 |
pxor mm5, mm0 |
pxor mm6, mm1 |
900 |
pxor mm5, mm1 |
pxor mm6, mm2 |
901 |
pxor mm5, mm2 |
movd eax, mm6 |
902 |
|
and eax, 1 |
903 |
movd eax, mm5 |
xor eax, 1 |
904 |
test eax, 0x1 |
xor word [edx + 2*63], ax |
|
jnz .done |
|
|
|
|
|
xor word [edi + 2*63], 1 |
|
|
|
|
|
.done |
|
|
pop edi |
|
|
pop esi |
|
905 |
|
|
906 |
ret |
ret |
907 |
|
|