21 |
; * along with this program ; if not, write to the Free Software |
; * along with this program ; if not, write to the Free Software |
22 |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
; * |
; * |
24 |
; * $Id: quantize_h263_mmx.asm,v 1.8 2008-08-19 09:06:48 Isibaar Exp $ |
; * $Id: quantize_h263_mmx.asm,v 1.9 2008-11-11 20:46:24 Isibaar Exp $ |
25 |
; * |
; * |
26 |
; ****************************************************************************/ |
; ****************************************************************************/ |
27 |
|
|
35 |
%ifdef MARK_FUNCS |
%ifdef MARK_FUNCS |
36 |
global _%1:function %1.endfunc-%1 |
global _%1:function %1.endfunc-%1 |
37 |
%define %1 _%1:function %1.endfunc-%1 |
%define %1 _%1:function %1.endfunc-%1 |
38 |
|
%define ENDFUNC .endfunc |
39 |
%else |
%else |
40 |
global _%1 |
global _%1 |
41 |
%define %1 _%1 |
%define %1 _%1 |
42 |
|
%define ENDFUNC |
43 |
%endif |
%endif |
44 |
%else |
%else |
45 |
%ifdef MARK_FUNCS |
%ifdef MARK_FUNCS |
46 |
global %1:function %1.endfunc-%1 |
global %1:function %1.endfunc-%1 |
47 |
|
%define ENDFUNC .endfunc |
48 |
%else |
%else |
49 |
global %1 |
global %1 |
50 |
|
%define ENDFUNC |
51 |
%endif |
%endif |
52 |
%endif |
%endif |
53 |
%endmacro |
%endmacro |
165 |
movq mm7, [mmx_div+ecx * 8 - 8] |
movq mm7, [mmx_div+ecx * 8 - 8] |
166 |
mov ecx,4 |
mov ecx,4 |
167 |
|
|
168 |
.loop |
.loop: |
169 |
movq mm0, [esi] ; data |
movq mm0, [esi] ; data |
170 |
pxor mm4,mm4 |
pxor mm4,mm4 |
171 |
movq mm1, [esi + 8] |
movq mm1, [esi + 8] |
197 |
jne .loop |
jne .loop |
198 |
jmp .end |
jmp .end |
199 |
|
|
200 |
.low |
.low: |
201 |
movd mm7,ecx |
movd mm7,ecx |
202 |
mov ecx,4 |
mov ecx,4 |
203 |
.loop_low |
.loop_low: |
204 |
movq mm0, [esi] |
movq mm0, [esi] |
205 |
pxor mm4,mm4 |
pxor mm4,mm4 |
206 |
movq mm1, [esi + 8] |
movq mm1, [esi + 8] |
231 |
lea edx, [edx+32] |
lea edx, [edx+32] |
232 |
jne .loop_low |
jne .loop_low |
233 |
|
|
234 |
.end |
.end: |
235 |
mov edx, [esp + 4 + 4] ; coeff |
mov edx, [esp + 4 + 4] ; coeff |
236 |
mov [edx],ax |
mov [edx],ax |
237 |
xor eax,eax ; return 0 |
xor eax,eax ; return 0 |
238 |
|
|
239 |
pop esi |
pop esi |
240 |
ret |
ret |
241 |
.endfunc |
ENDFUNC |
242 |
|
|
243 |
|
|
244 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
277 |
mov ecx,2 |
mov ecx,2 |
278 |
movlhps xmm7,xmm7 |
movlhps xmm7,xmm7 |
279 |
|
|
280 |
.loop |
.loop: |
281 |
movdqa xmm0, [esi] |
movdqa xmm0, [esi] |
282 |
pxor xmm4,xmm4 |
pxor xmm4,xmm4 |
283 |
movdqa xmm1, [esi + 16] |
movdqa xmm1, [esi + 16] |
309 |
jne .loop |
jne .loop |
310 |
jmp .end |
jmp .end |
311 |
|
|
312 |
.low |
.low: |
313 |
movd xmm7,ecx |
movd xmm7,ecx |
314 |
mov ecx,2 |
mov ecx,2 |
315 |
.loop_low |
.loop_low: |
316 |
movdqa xmm0, [esi] |
movdqa xmm0, [esi] |
317 |
pxor xmm4,xmm4 |
pxor xmm4,xmm4 |
318 |
movdqa xmm1, [esi + 16] |
movdqa xmm1, [esi + 16] |
343 |
lea edx, [edx+64] |
lea edx, [edx+64] |
344 |
jne .loop_low |
jne .loop_low |
345 |
|
|
346 |
.end |
.end: |
347 |
mov edx, [esp + 4 + 4] ; coeff |
mov edx, [esp + 4 + 4] ; coeff |
348 |
mov [edx],ax |
mov [edx],ax |
349 |
xor eax,eax ; return 0 |
xor eax,eax ; return 0 |
350 |
|
|
351 |
pop esi |
pop esi |
352 |
ret |
ret |
353 |
.endfunc |
ENDFUNC |
354 |
|
|
355 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
356 |
; |
; |
383 |
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
384 |
|
|
385 |
ALIGN 8 |
ALIGN 8 |
386 |
.loop |
.loop: |
387 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
388 |
movq mm3, [esi + 8*ecx + 8] |
movq mm3, [esi + 8*ecx + 8] |
389 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
411 |
cmp ecx, 16 |
cmp ecx, 16 |
412 |
jnz .loop |
jnz .loop |
413 |
|
|
414 |
.done |
.done: |
415 |
pmaddwd mm5, [plus_one] |
pmaddwd mm5, [plus_one] |
416 |
movq mm0, mm5 |
movq mm0, mm5 |
417 |
psrlq mm5, 32 |
psrlq mm5, 32 |
425 |
ret |
ret |
426 |
|
|
427 |
ALIGN 8 |
ALIGN 8 |
428 |
.q1loop |
.q1loop: |
429 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
430 |
movq mm3, [esi + 8*ecx+ 8] ; |
movq mm3, [esi + 8*ecx+ 8] ; |
431 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
454 |
jnz .q1loop |
jnz .q1loop |
455 |
|
|
456 |
jmp .done |
jmp .done |
457 |
.endfunc |
ENDFUNC |
458 |
|
|
459 |
|
|
460 |
|
|
488 |
cmp al, 1 |
cmp al, 1 |
489 |
jz near .qes2_q1loop |
jz near .qes2_q1loop |
490 |
|
|
491 |
.qes2_not1 |
.qes2_not1: |
492 |
movq mm0, [mmx_div + eax*8 - 8] ; divider |
movq mm0, [mmx_div + eax*8 - 8] ; divider |
493 |
movq2dq xmm7, mm0 |
movq2dq xmm7, mm0 |
494 |
movlhps xmm7, xmm7 |
movlhps xmm7, xmm7 |
495 |
|
|
496 |
ALIGN 16 |
ALIGN 16 |
497 |
.qes2_loop |
.qes2_loop: |
498 |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
499 |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
500 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
522 |
cmp ecx, 16 |
cmp ecx, 16 |
523 |
jnz .qes2_loop |
jnz .qes2_loop |
524 |
|
|
525 |
.qes2_done |
.qes2_done: |
526 |
movdqu xmm6, [plus_one] |
movdqu xmm6, [plus_one] |
527 |
pmaddwd xmm5, xmm6 |
pmaddwd xmm5, xmm6 |
528 |
movhlps xmm6, xmm5 |
movhlps xmm6, xmm5 |
541 |
ret |
ret |
542 |
|
|
543 |
ALIGN 16 |
ALIGN 16 |
544 |
.qes2_q1loop |
.qes2_q1loop: |
545 |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
movdqa xmm0, [esi + ecx*8] ; xmm0 = [1st] |
546 |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
movdqa xmm3, [esi + ecx*8 + 16] ; xmm3 = [2nd] |
547 |
pxor xmm1, xmm1 |
pxor xmm1, xmm1 |
569 |
cmp ecx, 16 |
cmp ecx, 16 |
570 |
jnz .qes2_q1loop |
jnz .qes2_q1loop |
571 |
jmp .qes2_done |
jmp .qes2_done |
572 |
.endfunc |
ENDFUNC |
573 |
|
|
574 |
|
|
575 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
655 |
|
|
656 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
657 |
ret |
ret |
658 |
.endfunc |
ENDFUNC |
659 |
|
|
660 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
661 |
; |
; |
740 |
|
|
741 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
742 |
ret |
ret |
743 |
.endfunc |
ENDFUNC |
744 |
|
|
745 |
|
|
746 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
829 |
|
|
830 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
831 |
ret |
ret |
832 |
.endfunc |
ENDFUNC |
833 |
|
|
834 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
835 |
; |
; |
900 |
|
|
901 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
902 |
ret |
ret |
903 |
.endfunc |
ENDFUNC |
904 |
|
|
905 |
|
|
906 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
969 |
|
|
970 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
971 |
ret |
ret |
972 |
.endfunc |
ENDFUNC |
973 |
|
|
974 |
|
|
975 |
;----------------------------------------------------------------------------- |
;----------------------------------------------------------------------------- |
1041 |
|
|
1042 |
xor eax, eax ; return 0 |
xor eax, eax ; return 0 |
1043 |
ret |
ret |
1044 |
.endfunc |
ENDFUNC |
1045 |
|
|
1046 |
|
|
1047 |
%ifidn __OUTPUT_FORMAT__,elf |
%ifidn __OUTPUT_FORMAT__,elf |