1 |
;/************************************************************************** |
;/***************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * mmx quantization/dequantization |
; * - MPEG4 Quantization H263 implementation / MMX optimized - |
5 |
; * |
; * |
6 |
; * This program is an implementation of a part of one or more MPEG-4 |
; * Copyright(C) 2001-2003 Peter Ross <pross@xvid.org> |
7 |
; * Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
; * 2002-2003 Pascal Massimino <skal@planet-d.net> |
|
; * to use this software module in hardware or software products are |
|
|
; * advised that its use may infringe existing patents or copyrights, and |
|
|
; * any such use would be at such party's own risk. The original |
|
|
; * developer of this software module and his/her company, and subsequent |
|
|
; * editors and their companies, will have no liability for use of this |
|
|
; * software or modifications or derivatives thereof. |
|
8 |
; * |
; * |
9 |
; * This program is free software; you can redistribute it and/or modify |
; * This program is free software; you can redistribute it and/or modify |
10 |
; * it under the terms of the GNU General Public License as published by |
; * it under the terms of the GNU General Public License as published by |
18 |
; * |
; * |
19 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
20 |
; * along with this program; if not, write to the Free Software |
; * along with this program; if not, write to the Free Software |
21 |
; * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
22 |
; * |
; * |
23 |
; *************************************************************************/ |
; * $Id: quantize_h263_mmx.asm,v 1.1.2.2 2003-10-09 18:50:22 edgomez Exp $ |
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
24 |
; * |
; * |
25 |
; * 09.08.2002 sse2 dequant funcs revamped |
; ****************************************************************************/ |
|
; * 14.06.2002 mmx+xmm dequant_* funcs revamped -Skal- |
|
|
; * 24.02.2002 sse2 quant_intra / dequant_intra (have to use movdqu ???) |
|
|
; * 17.04.2002 sse2 quant_inter / dequant_inter |
|
|
; * 26.12.2001 minor bug fixes, dequant saturate, further optimization |
|
|
; * 19.11.2001 quant_inter_mmx now returns sum of abs. coefficient values |
|
|
; * 04.11.2001 nasm version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
26 |
|
|
27 |
; enable dequant saturate [-2048,2047], test purposes only. |
; enable dequant saturate [-2048,2047], test purposes only. |
28 |
%define SATURATE |
%define SATURATE |
32 |
|
|
33 |
bits 32 |
bits 32 |
34 |
|
|
|
section .data |
|
|
|
|
|
|
|
35 |
%macro cglobal 1 |
%macro cglobal 1 |
36 |
%ifdef PREFIX |
%ifdef PREFIX |
37 |
global _%1 |
global _%1 |
41 |
%endif |
%endif |
42 |
%endmacro |
%endmacro |
43 |
|
|
44 |
align 16 |
;*************************************************************************** |
45 |
|
; Local data |
46 |
|
;*************************************************************************** |
47 |
|
|
48 |
plus_one times 8 dw 1 |
%ifdef FORMAT_COFF |
49 |
|
section .data data |
50 |
|
%else |
51 |
|
section .data data align=16 |
52 |
|
%endif |
53 |
|
|
54 |
|
align 16 |
55 |
|
plus_one: |
56 |
|
times 8 dw 1 |
57 |
|
|
58 |
;=========================================================================== |
;=========================================================================== |
59 |
; |
; |
61 |
; |
; |
62 |
;=========================================================================== |
;=========================================================================== |
63 |
|
|
|
%macro MMX_SUB 1 |
|
|
times 4 dw %1 / 2 |
|
|
%endmacro |
|
|
|
|
64 |
align 16 |
align 16 |
65 |
mmx_sub |
mmx_sub: |
66 |
MMX_SUB 1 |
%assign quant 1 |
67 |
MMX_SUB 2 |
%rep 31 |
68 |
MMX_SUB 3 |
times 4 dw quant / 2 |
69 |
MMX_SUB 4 |
%assign quant quant+1 |
70 |
MMX_SUB 5 |
%endrep |
|
MMX_SUB 6 |
|
|
MMX_SUB 7 |
|
|
MMX_SUB 8 |
|
|
MMX_SUB 9 |
|
|
MMX_SUB 10 |
|
|
MMX_SUB 11 |
|
|
MMX_SUB 12 |
|
|
MMX_SUB 13 |
|
|
MMX_SUB 14 |
|
|
MMX_SUB 15 |
|
|
MMX_SUB 16 |
|
|
MMX_SUB 17 |
|
|
MMX_SUB 18 |
|
|
MMX_SUB 19 |
|
|
MMX_SUB 20 |
|
|
MMX_SUB 21 |
|
|
MMX_SUB 22 |
|
|
MMX_SUB 23 |
|
|
MMX_SUB 24 |
|
|
MMX_SUB 25 |
|
|
MMX_SUB 26 |
|
|
MMX_SUB 27 |
|
|
MMX_SUB 28 |
|
|
MMX_SUB 29 |
|
|
MMX_SUB 30 |
|
|
MMX_SUB 31 |
|
|
|
|
|
|
|
71 |
|
|
72 |
;=========================================================================== |
;=========================================================================== |
73 |
; |
; |
79 |
; |
; |
80 |
;=========================================================================== |
;=========================================================================== |
81 |
|
|
|
%macro MMX_DIV 1 |
|
|
times 4 dw (1 << 16) / (%1 * 2) + 1 |
|
|
%endmacro |
|
|
|
|
82 |
align 16 |
align 16 |
83 |
mmx_div |
mmx_div: |
84 |
MMX_DIV 1 |
%assign quant 1 |
85 |
MMX_DIV 2 |
%rep 31 |
86 |
MMX_DIV 3 |
times 4 dw (1<<16) / (quant*2) + 1 |
87 |
MMX_DIV 4 |
%assign quant quant+1 |
88 |
MMX_DIV 5 |
%endrep |
|
MMX_DIV 6 |
|
|
MMX_DIV 7 |
|
|
MMX_DIV 8 |
|
|
MMX_DIV 9 |
|
|
MMX_DIV 10 |
|
|
MMX_DIV 11 |
|
|
MMX_DIV 12 |
|
|
MMX_DIV 13 |
|
|
MMX_DIV 14 |
|
|
MMX_DIV 15 |
|
|
MMX_DIV 16 |
|
|
MMX_DIV 17 |
|
|
MMX_DIV 18 |
|
|
MMX_DIV 19 |
|
|
MMX_DIV 20 |
|
|
MMX_DIV 21 |
|
|
MMX_DIV 22 |
|
|
MMX_DIV 23 |
|
|
MMX_DIV 24 |
|
|
MMX_DIV 25 |
|
|
MMX_DIV 26 |
|
|
MMX_DIV 27 |
|
|
MMX_DIV 28 |
|
|
MMX_DIV 29 |
|
|
MMX_DIV 30 |
|
|
MMX_DIV 31 |
|
|
|
|
|
|
|
89 |
|
|
90 |
;=========================================================================== |
;=========================================================================== |
91 |
; |
; |
93 |
; |
; |
94 |
;=========================================================================== |
;=========================================================================== |
95 |
|
|
96 |
%macro MMX_ADD 1 |
align 16 |
97 |
%if %1 % 2 != 0 |
mmx_add: |
98 |
times 4 dw %1 |
%assign quant 1 |
99 |
|
%rep 31 |
100 |
|
%if quant % 2 != 0 |
101 |
|
times 4 dw quant |
102 |
%else |
%else |
103 |
times 4 dw %1 - 1 |
times 4 dw quant - 1 |
104 |
%endif |
%endif |
105 |
%endmacro |
%assign quant quant+1 |
106 |
|
%endrep |
|
align 16 |
|
|
mmx_add |
|
|
MMX_ADD 1 |
|
|
MMX_ADD 2 |
|
|
MMX_ADD 3 |
|
|
MMX_ADD 4 |
|
|
MMX_ADD 5 |
|
|
MMX_ADD 6 |
|
|
MMX_ADD 7 |
|
|
MMX_ADD 8 |
|
|
MMX_ADD 9 |
|
|
MMX_ADD 10 |
|
|
MMX_ADD 11 |
|
|
MMX_ADD 12 |
|
|
MMX_ADD 13 |
|
|
MMX_ADD 14 |
|
|
MMX_ADD 15 |
|
|
MMX_ADD 16 |
|
|
MMX_ADD 17 |
|
|
MMX_ADD 18 |
|
|
MMX_ADD 19 |
|
|
MMX_ADD 20 |
|
|
MMX_ADD 21 |
|
|
MMX_ADD 22 |
|
|
MMX_ADD 23 |
|
|
MMX_ADD 24 |
|
|
MMX_ADD 25 |
|
|
MMX_ADD 26 |
|
|
MMX_ADD 27 |
|
|
MMX_ADD 28 |
|
|
MMX_ADD 29 |
|
|
MMX_ADD 30 |
|
|
MMX_ADD 31 |
|
|
|
|
107 |
|
|
108 |
;=========================================================================== |
;=========================================================================== |
109 |
; |
; |
111 |
; |
; |
112 |
;=========================================================================== |
;=========================================================================== |
113 |
|
|
|
%macro MMX_MUL 1 |
|
|
times 4 dw %1 * 2 |
|
|
%endmacro |
|
|
|
|
114 |
align 16 |
align 16 |
115 |
mmx_mul |
mmx_mul: |
116 |
MMX_MUL 1 |
%assign quant 1 |
117 |
MMX_MUL 2 |
%rep 31 |
118 |
MMX_MUL 3 |
times 4 dw quant*2 |
119 |
MMX_MUL 4 |
%assign quant quant+1 |
120 |
MMX_MUL 5 |
%endrep |
|
MMX_MUL 6 |
|
|
MMX_MUL 7 |
|
|
MMX_MUL 8 |
|
|
MMX_MUL 9 |
|
|
MMX_MUL 10 |
|
|
MMX_MUL 11 |
|
|
MMX_MUL 12 |
|
|
MMX_MUL 13 |
|
|
MMX_MUL 14 |
|
|
MMX_MUL 15 |
|
|
MMX_MUL 16 |
|
|
MMX_MUL 17 |
|
|
MMX_MUL 18 |
|
|
MMX_MUL 19 |
|
|
MMX_MUL 20 |
|
|
MMX_MUL 21 |
|
|
MMX_MUL 22 |
|
|
MMX_MUL 23 |
|
|
MMX_MUL 24 |
|
|
MMX_MUL 25 |
|
|
MMX_MUL 26 |
|
|
MMX_MUL 27 |
|
|
MMX_MUL 28 |
|
|
MMX_MUL 29 |
|
|
MMX_MUL 30 |
|
|
MMX_MUL 31 |
|
|
|
|
121 |
|
|
122 |
;=========================================================================== |
;=========================================================================== |
123 |
; |
; |
126 |
;=========================================================================== |
;=========================================================================== |
127 |
|
|
128 |
align 16 |
align 16 |
129 |
sse2_2047 times 8 dw 2047 |
sse2_2047: |
130 |
|
times 8 dw 2047 |
131 |
|
|
132 |
align 16 |
align 16 |
133 |
mmx_2047 times 4 dw 2047 |
mmx_2047: |
134 |
|
times 4 dw 2047 |
135 |
|
|
136 |
align 8 |
align 8 |
137 |
mmx_32768_minus_2048 times 4 dw (32768-2048) |
mmx_32768_minus_2048: |
138 |
mmx_32767_minus_2047 times 4 dw (32767-2047) |
times 4 dw (32768-2048) |
139 |
|
|
140 |
|
mmx_32767_minus_2047: |
141 |
|
times 4 dw (32767-2047) |
142 |
|
|
143 |
|
|
144 |
|
;*************************************************************************** |
145 |
|
; Code |
146 |
|
;*************************************************************************** |
147 |
|
|
148 |
section .text |
section .text |
149 |
|
|
150 |
|
|
151 |
;=========================================================================== |
;=========================================================================== |
152 |
; |
; |
153 |
; void quant_intra_mmx(int16_t * coeff, |
; uint32_t quant_h263_intra_mmx(int16_t * coeff, |
154 |
; const int16_t const * data, |
; const int16_t const * data, |
155 |
; const uint32_t quant, |
; const uint32_t quant, |
156 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
174 |
jz .q1loop |
jz .q1loop |
175 |
|
|
176 |
movq mm7, [mmx_div + eax * 8 - 8] |
movq mm7, [mmx_div + eax * 8 - 8] |
177 |
|
|
178 |
align ALIGN |
align ALIGN |
179 |
.loop |
.loop |
180 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
181 |
movq mm3, [esi + 8*ecx + 8] ; |
movq mm3, [esi + 8*ecx + 8] |
182 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
183 |
pxor mm4, mm4 ; |
pxor mm4, mm4 ; |
184 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
201 |
jnz .loop |
jnz .loop |
202 |
|
|
203 |
.done |
.done |
|
; caclulate data[0] // (int32_t)dcscalar) |
|
204 |
|
|
205 |
|
; caclulate data[0] // (int32_t)dcscalar) |
206 |
mov ecx, [esp + 12 + 16] ; dcscalar |
mov ecx, [esp + 12 + 16] ; dcscalar |
207 |
mov edx, ecx |
mov edx, ecx |
208 |
movsx eax, word [esi] ; data[0] |
movsx eax, word [esi] ; data[0] |
212 |
|
|
213 |
sub eax, edx |
sub eax, edx |
214 |
jmp short .mul |
jmp short .mul |
215 |
|
|
216 |
.gtzero |
.gtzero |
217 |
add eax, edx |
add eax, edx |
218 |
.mul |
.mul |
219 |
cdq ; expand eax -> edx:eax |
cdq ; expand eax -> edx:eax |
220 |
idiv ecx ; eax = edx:eax / dcscalar |
idiv ecx ; eax = edx:eax / dcscalar |
|
|
|
221 |
mov [edi], ax ; coeff[0] = ax |
mov [edi], ax ; coeff[0] = ax |
222 |
|
|
223 |
|
xor eax, eax ; return(0); |
224 |
pop edi |
pop edi |
225 |
pop esi |
pop esi |
226 |
pop ecx |
pop ecx |
230 |
align ALIGN |
align ALIGN |
231 |
.q1loop |
.q1loop |
232 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
233 |
movq mm3, [esi + 8*ecx + 8] ; |
movq mm3, [esi + 8*ecx + 8] |
234 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
235 |
pxor mm4, mm4 ; |
pxor mm4, mm4 ; |
236 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
242 |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
psrlw mm0, 1 ; mm0 >>= 1 (/2) |
243 |
psrlw mm3, 1 ; |
psrlw mm3, 1 ; |
244 |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
pxor mm0, mm1 ; mm0 *= sign(mm0) |
245 |
pxor mm3, mm4 ; |
pxor mm3, mm4 |
246 |
psubw mm0, mm1 ; undisplace |
psubw mm0, mm1 ; undisplace |
247 |
psubw mm3, mm4 ; |
psubw mm3, mm4 ; |
248 |
movq [edi + 8*ecx], mm0 |
movq [edi + 8*ecx], mm0 |
257 |
|
|
258 |
;=========================================================================== |
;=========================================================================== |
259 |
; |
; |
260 |
; void quant_intra_sse2(int16_t * coeff, |
; uint32_t quant_h263_intra_sse2(int16_t * coeff, |
261 |
; const int16_t const * data, |
; const int16_t const * data, |
262 |
; const uint32_t quant, |
; const uint32_t quant, |
263 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
319 |
|
|
320 |
sub eax, edx |
sub eax, edx |
321 |
jmp short .qas2_mul |
jmp short .qas2_mul |
322 |
|
|
323 |
.qas2_gtzero |
.qas2_gtzero |
324 |
add eax, edx |
add eax, edx |
325 |
|
|
326 |
.qas2_mul |
.qas2_mul |
327 |
cdq |
cdq |
328 |
idiv ecx |
idiv ecx |
329 |
|
|
330 |
mov [edi], ax |
mov [edi], ax |
331 |
|
|
332 |
|
xor eax, eax ; return(0); |
333 |
pop edi |
pop edi |
334 |
pop esi |
pop esi |
335 |
|
|
365 |
|
|
366 |
;=========================================================================== |
;=========================================================================== |
367 |
; |
; |
368 |
; uint32_t quant_inter_mmx(int16_t * coeff, |
; uint32_t quant_h263_inter_mmx(int16_t * coeff, |
369 |
; const int16_t const * data, |
; const int16_t const * data, |
370 |
; const uint32_t quant); |
; const uint32_t quant); |
371 |
; |
; |
396 |
align ALIGN |
align ALIGN |
397 |
.loop |
.loop |
398 |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
movq mm0, [esi + 8*ecx] ; mm0 = [1st] |
399 |
movq mm3, [esi + 8*ecx + 8] ; |
movq mm3, [esi + 8*ecx + 8] |
400 |
pxor mm1, mm1 ; mm1 = 0 |
pxor mm1, mm1 ; mm1 = 0 |
401 |
pxor mm4, mm4 ; |
pxor mm4, mm4 ; |
402 |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
pcmpgtw mm1, mm0 ; mm1 = (0 > mm0) |
427 |
movq mm0, mm5 |
movq mm0, mm5 |
428 |
psrlq mm5, 32 |
psrlq mm5, 32 |
429 |
paddd mm0, mm5 |
paddd mm0, mm5 |
|
movd eax, mm0 ; return sum |
|
430 |
|
|
431 |
|
movd eax, mm0 ; return sum |
432 |
pop edi |
pop edi |
433 |
pop esi |
pop esi |
434 |
pop ecx |
pop ecx |
470 |
|
|
471 |
;=========================================================================== |
;=========================================================================== |
472 |
; |
; |
473 |
; uint32_t quant_inter_sse2(int16_t * coeff, |
; uint32_t quant_h263_inter_sse2(int16_t * coeff, |
474 |
; const int16_t const * data, |
; const int16_t const * data, |
475 |
; const uint32_t quant); |
; const uint32_t quant); |
476 |
; |
; |
542 |
movq mm5, mm0 |
movq mm5, mm0 |
543 |
psrlq mm5, 32 |
psrlq mm5, 32 |
544 |
paddd mm0, mm5 |
paddd mm0, mm5 |
545 |
|
|
546 |
movd eax, mm0 ; return sum |
movd eax, mm0 ; return sum |
547 |
|
|
548 |
pop edi |
pop edi |
583 |
|
|
584 |
;=========================================================================== |
;=========================================================================== |
585 |
; |
; |
586 |
; void dequant_intra_mmx(int16_t *data, |
; uint32_t dequant_h263_intra_mmx(int16_t *data, |
587 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
588 |
; const uint32_t quant, |
; const uint32_t quant, |
589 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
644 |
jnz near .loop |
jnz near .loop |
645 |
|
|
646 |
; deal with DC |
; deal with DC |
|
|
|
647 |
movd mm0, [ecx] |
movd mm0, [ecx] |
648 |
pmullw mm0, [esp+16] ; dcscalar |
pmullw mm0, [esp+16] ; dcscalar |
649 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
655 |
movd eax, mm0 |
movd eax, mm0 |
656 |
mov [edx], ax |
mov [edx], ax |
657 |
|
|
658 |
|
xor eax, eax ; return(0); |
659 |
ret |
ret |
660 |
|
|
661 |
;=========================================================================== |
;=========================================================================== |
662 |
; |
; |
663 |
; void dequant_intra_xmm(int16_t *data, |
; uint32_t dequant_h263_intra_xmm(int16_t *data, |
664 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
665 |
; const uint32_t quant, |
; const uint32_t quant, |
666 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
719 |
jnz near .loop |
jnz near .loop |
720 |
|
|
721 |
; deal with DC |
; deal with DC |
|
|
|
722 |
movd mm0, [ecx] |
movd mm0, [ecx] |
723 |
pmullw mm0, [esp+16] ; dcscalar |
pmullw mm0, [esp+16] ; dcscalar |
724 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
730 |
movd eax, mm0 |
movd eax, mm0 |
731 |
mov [edx], ax |
mov [edx], ax |
732 |
|
|
733 |
|
xor eax, eax |
734 |
ret |
ret |
735 |
|
|
736 |
|
|
737 |
;=========================================================================== |
;=========================================================================== |
738 |
; |
; |
739 |
; void dequant_intra_sse2(int16_t *data, |
; uint32_t dequant_h263_intra_sse2(int16_t *data, |
740 |
; const int16_t const *coeff, |
; const int16_t const *coeff, |
741 |
; const uint32_t quant, |
; const uint32_t quant, |
742 |
; const uint32_t dcscalar); |
; const uint32_t dcscalar); |
743 |
; |
; |
744 |
;=========================================================================== |
;=========================================================================== |
745 |
|
|
746 |
align ALIGN |
align ALIGN |
747 |
cglobal dequant_h263_intra_sse2 |
cglobal dequant_h263_intra_sse2 |
748 |
dequant_h263_intra_sse2: |
dequant_h263_intra_sse2: |
806 |
movd eax, mm0 |
movd eax, mm0 |
807 |
mov [edx], ax |
mov [edx], ax |
808 |
|
|
809 |
|
xor eax, eax |
810 |
ret |
ret |
811 |
|
|
|
|
|
|
|
|
|
|
|
812 |
;=========================================================================== |
;=========================================================================== |
813 |
; |
; |
814 |
; void dequant_inter_mmx(int16_t * data, |
; uint32t dequant_h263_inter_mmx(int16_t * data, |
815 |
; const int16_t * const coeff, |
; const int16_t * const coeff, |
816 |
; const uint32_t quant); |
; const uint32_t quant); |
817 |
; |
; |
852 |
paddw mm3, mm5 ; + offset |
paddw mm3, mm5 ; + offset |
853 |
paddw mm0, mm1 ; negate back |
paddw mm0, mm1 ; negate back |
854 |
paddw mm3, mm4 ; negate back |
paddw mm3, mm4 ; negate back |
|
|
|
855 |
; saturates to +2047 |
; saturates to +2047 |
856 |
movq mm2, [mmx_32767_minus_2047] |
movq mm2, [mmx_32767_minus_2047] |
857 |
add eax, 2 |
add eax, 2 |
866 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
867 |
jnz near .loop |
jnz near .loop |
868 |
|
|
869 |
|
xor eax, eax |
870 |
ret |
ret |
871 |
|
|
872 |
;=========================================================================== |
;=========================================================================== |
873 |
; |
; |
874 |
; void dequant_inter_xmm(int16_t * data, |
; uint32_t dequant_h263_inter_xmm(int16_t * data, |
875 |
; const int16_t * const coeff, |
; const int16_t * const coeff, |
876 |
; const uint32_t quant); |
; const uint32_t quant); |
877 |
; |
; |
915 |
paddw mm3, mm5 ; + offset |
paddw mm3, mm5 ; + offset |
916 |
paddw mm0, mm1 ; start restoring sign |
paddw mm0, mm1 ; start restoring sign |
917 |
paddw mm3, mm4 ; start restoring sign |
paddw mm3, mm4 ; start restoring sign |
|
|
|
918 |
; saturates to +2047 |
; saturates to +2047 |
919 |
movq mm2, [mmx_2047] |
movq mm2, [mmx_2047] |
920 |
pminsw mm0, mm2 |
pminsw mm0, mm2 |
927 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
movq [edx + 8*eax + 8*16+8 - 2*8], mm3 |
928 |
jnz near .loop |
jnz near .loop |
929 |
|
|
930 |
|
xor eax, eax |
931 |
ret |
ret |
932 |
|
|
933 |
;=========================================================================== |
;=========================================================================== |
934 |
; |
; |
935 |
; void dequant_inter_sse2(int16_t * data, |
; uint32_t dequant_h263_inter_sse2(int16_t * data, |
936 |
; const int16_t * const coeff, |
; const int16_t * const coeff, |
937 |
; const uint32_t quant); |
; const uint32_t quant); |
938 |
; |
; |
939 |
;=========================================================================== |
;=========================================================================== |
940 |
|
|
941 |
align ALIGN |
align ALIGN |
942 |
cglobal dequant_h263_inter_sse2 |
cglobal dequant_h263_inter_sse2 |
943 |
dequant_h263_inter_sse2: |
dequant_h263_inter_sse2: |
991 |
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
movdqa [edx + 8*16 - 8*4 + 8*eax + 16], xmm3 |
992 |
jnz near .loop |
jnz near .loop |
993 |
|
|
994 |
|
xor eax, eax |
995 |
ret |
ret |