1 |
;/************************************************************************** |
;/**************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * xmm sum of absolute difference |
; * - K7 optimized SAD operators - |
5 |
; * |
; * |
6 |
; * This program is an implementation of a part of one or more MPEG-4 |
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
; * Video tools as specified in ISO/IEC 14496-2 standard. Those intending |
; * 2001-2008 Michael Militzer <michael@xvid.org> |
8 |
; * to use this software module in hardware or software products are |
; * 2002 Pascal Massimino <skal@planet-d.net> |
|
; * 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. |
|
9 |
; * |
; * |
10 |
; * This program is free software; you can redistribute it and/or modify |
; * This program is free software; you can redistribute it and/or modify it |
11 |
; * it under the terms of the GNU General Public License as published by |
; * under the terms of the GNU General Public License as published by |
12 |
; * the Free Software Foundation; either version 2 of the License, or |
; * the Free Software Foundation; either version 2 of the License, or |
13 |
; * (at your option) any later version. |
; * (at your option) any later version. |
14 |
; * |
; * |
19 |
; * |
; * |
20 |
; * You should have received a copy of the GNU General Public License |
; * You should have received a copy of the GNU General Public License |
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., 675 Mass Ave, Cambridge, MA 02139, USA. |
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
23 |
; * |
; * |
24 |
; *************************************************************************/ |
; * $Id: sad_xmm.asm,v 1.13 2008-11-26 01:04:34 Isibaar Exp $ |
|
|
|
|
;/************************************************************************** |
|
|
; * |
|
|
; * History: |
|
25 |
; * |
; * |
26 |
; * 23.07.2002 sad8bi_xmm; <pross@xvid.org> |
; ***************************************************************************/ |
|
; * 04.06.2002 rewrote some funcs (XMM mainly) -Skal- |
|
|
; * 17.11.2001 bugfix and small improvement for dev16_xmm, |
|
|
; * removed terminate early in sad16_xmm (Isibaar) |
|
|
; * 12.11.2001 inital version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
27 |
|
|
28 |
bits 32 |
%include "nasm.inc" |
29 |
|
|
30 |
%macro cglobal 1 |
;============================================================================= |
31 |
%ifdef PREFIX |
; Read only data |
32 |
global _%1 |
;============================================================================= |
33 |
%define %1 _%1 |
|
34 |
%else |
DATA |
35 |
global %1 |
|
36 |
%endif |
ALIGN SECTION_ALIGN |
37 |
|
mmx_one: times 4 dw 1 |
38 |
|
|
39 |
|
;============================================================================= |
40 |
|
; Helper macros |
41 |
|
;============================================================================= |
42 |
|
|
43 |
|
%macro SAD_16x16_SSE 0 |
44 |
|
movq mm0, [_EAX] |
45 |
|
psadbw mm0, [TMP1] |
46 |
|
movq mm1, [_EAX+8] |
47 |
|
add _EAX, TMP0 |
48 |
|
psadbw mm1, [TMP1+8] |
49 |
|
paddusw mm5, mm0 |
50 |
|
add TMP1, TMP0 |
51 |
|
paddusw mm6, mm1 |
52 |
%endmacro |
%endmacro |
53 |
|
|
54 |
section .data |
%macro SAD_8x8_SSE 0 |
55 |
|
movq mm0, [_EAX] |
56 |
|
movq mm1, [_EAX+TMP0] |
57 |
|
psadbw mm0, [TMP1] |
58 |
|
psadbw mm1, [TMP1+TMP0] |
59 |
|
add _EAX, _EBX |
60 |
|
add TMP1, _EBX |
61 |
|
paddusw mm5, mm0 |
62 |
|
paddusw mm6, mm1 |
63 |
|
%endmacro |
64 |
|
|
65 |
align 16 |
%macro SADBI_16x16_SSE 0 |
66 |
mmx_one times 4 dw 1 |
movq mm0, [_EAX] |
67 |
|
movq mm1, [_EAX+8] |
68 |
|
movq mm2, [TMP1] |
69 |
|
movq mm3, [TMP1+8] |
70 |
|
pavgb mm2, [_EBX] |
71 |
|
add TMP1, TMP0 |
72 |
|
pavgb mm3, [_EBX+8] |
73 |
|
add _EBX, TMP0 |
74 |
|
psadbw mm0, mm2 |
75 |
|
add _EAX, TMP0 |
76 |
|
psadbw mm1, mm3 |
77 |
|
paddusw mm5, mm0 |
78 |
|
paddusw mm6, mm1 |
79 |
|
%endmacro |
80 |
|
|
81 |
section .text |
%macro SADBI_8x8_XMM 0 |
82 |
|
movq mm0, [_EAX] |
83 |
|
movq mm1, [_EAX+TMP0] |
84 |
|
movq mm2, [TMP1] |
85 |
|
movq mm3, [TMP1+TMP0] |
86 |
|
pavgb mm2, [_EBX] |
87 |
|
lea TMP1, [TMP1+2*TMP0] |
88 |
|
pavgb mm3, [_EBX+TMP0] |
89 |
|
lea _EBX, [_EBX+2*TMP0] |
90 |
|
psadbw mm0, mm2 |
91 |
|
lea _EAX, [_EAX+2*TMP0] |
92 |
|
psadbw mm1, mm3 |
93 |
|
paddusw mm5, mm0 |
94 |
|
paddusw mm6, mm1 |
95 |
|
%endmacro |
96 |
|
|
97 |
|
%macro MEAN_16x16_SSE 0 |
98 |
|
movq mm0, [_EAX] |
99 |
|
movq mm1, [_EAX+8] |
100 |
|
psadbw mm0, mm7 |
101 |
|
psadbw mm1, mm7 |
102 |
|
add _EAX, TMP0 |
103 |
|
paddw mm5, mm0 |
104 |
|
paddw mm6, mm1 |
105 |
|
%endmacro |
106 |
|
|
107 |
|
%macro ABS_16x16_SSE 0 |
108 |
|
movq mm0, [_EAX] |
109 |
|
movq mm1, [_EAX+8] |
110 |
|
psadbw mm0, mm4 |
111 |
|
psadbw mm1, mm4 |
112 |
|
lea _EAX, [_EAX+TMP0] |
113 |
|
paddw mm5, mm0 |
114 |
|
paddw mm6, mm1 |
115 |
|
%endmacro |
116 |
|
|
117 |
|
;============================================================================= |
118 |
|
; Code |
119 |
|
;============================================================================= |
120 |
|
|
121 |
|
SECTION .rotext align=SECTION_ALIGN |
122 |
|
|
123 |
cglobal sad16_xmm |
cglobal sad16_xmm |
124 |
cglobal sad8_xmm |
cglobal sad8_xmm |
125 |
cglobal sad16bi_xmm |
cglobal sad16bi_xmm |
126 |
cglobal sad8bi_xmm |
cglobal sad8bi_xmm |
127 |
cglobal dev16_xmm |
cglobal dev16_xmm |
128 |
|
cglobal sad16v_xmm |
129 |
|
|
130 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
131 |
; |
; |
132 |
; uint32_t sad16_xmm(const uint8_t * const cur, |
; uint32_t sad16_xmm(const uint8_t * const cur, |
133 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
134 |
; const uint32_t stride, |
; const uint32_t stride, |
135 |
; const uint32_t best_sad); |
; const uint32_t best_sad); |
136 |
; |
; |
137 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro SAD_16x16_SSE 0 |
|
|
movq mm0, [eax] |
|
|
psadbw mm0, [edx] |
|
|
movq mm1, [eax+8] |
|
|
add eax, ecx |
|
|
psadbw mm1, [edx+8] |
|
|
paddusw mm5,mm0 |
|
|
add edx, ecx |
|
|
paddusw mm6,mm1 |
|
|
%endmacro |
|
138 |
|
|
139 |
align 16 |
ALIGN SECTION_ALIGN |
140 |
sad16_xmm: |
sad16_xmm: |
141 |
|
|
142 |
mov eax, [esp+ 4] ; Src1 |
mov _EAX, prm1 ; Src1 |
143 |
mov edx, [esp+ 8] ; Src2 |
mov TMP1, prm2 ; Src2 |
144 |
mov ecx, [esp+12] ; Stride |
mov TMP0, prm3 ; Stride |
145 |
|
|
146 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
147 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
167 |
paddusw mm6,mm5 |
paddusw mm6,mm5 |
168 |
movd eax, mm6 |
movd eax, mm6 |
169 |
ret |
ret |
170 |
|
ENDFUNC |
171 |
|
|
172 |
|
|
173 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
174 |
; |
; |
175 |
; uint32_t sad8_xmm(const uint8_t * const cur, |
; uint32_t sad8_xmm(const uint8_t * const cur, |
176 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
177 |
; const uint32_t stride); |
; const uint32_t stride); |
178 |
; |
; |
179 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro SAD_8x8_SSE 0 |
|
|
movq mm0, [eax] |
|
|
movq mm1, [eax+ecx] |
|
|
|
|
|
psadbw mm0, [edx] |
|
|
psadbw mm1, [edx+ecx] |
|
|
add eax, ebx |
|
|
add edx, ebx |
|
|
|
|
|
paddusw mm5,mm0 |
|
|
paddusw mm6,mm1 |
|
|
%endmacro |
|
180 |
|
|
181 |
align 16 |
ALIGN SECTION_ALIGN |
182 |
sad8_xmm: |
sad8_xmm: |
183 |
|
|
184 |
mov eax, [esp+ 4] ; Src1 |
mov _EAX, prm1 ; Src1 |
185 |
mov edx, [esp+ 8] ; Src2 |
mov TMP1, prm2 ; Src2 |
186 |
mov ecx, [esp+12] ; Stride |
mov TMP0, prm3 ; Stride |
187 |
push ebx |
push _EBX |
188 |
lea ebx, [ecx+ecx] |
lea _EBX, [TMP0+TMP0] |
189 |
|
|
190 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
191 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
194 |
SAD_8x8_SSE |
SAD_8x8_SSE |
195 |
SAD_8x8_SSE |
SAD_8x8_SSE |
196 |
|
|
197 |
movq mm0, [eax] |
movq mm0, [_EAX] |
198 |
movq mm1, [eax+ecx] |
movq mm1, [_EAX+TMP0] |
199 |
psadbw mm0, [edx] |
psadbw mm0, [TMP1] |
200 |
psadbw mm1, [edx+ecx] |
psadbw mm1, [TMP1+TMP0] |
201 |
|
|
202 |
pop ebx |
pop _EBX |
203 |
|
|
204 |
paddusw mm5,mm0 |
paddusw mm5,mm0 |
205 |
paddusw mm6,mm1 |
paddusw mm6,mm1 |
208 |
movd eax, mm6 |
movd eax, mm6 |
209 |
|
|
210 |
ret |
ret |
211 |
|
ENDFUNC |
212 |
|
|
213 |
|
|
214 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
215 |
; |
; |
216 |
; uint32_t sad16bi_xmm(const uint8_t * const cur, |
; uint32_t sad16bi_xmm(const uint8_t * const cur, |
217 |
; const uint8_t * const ref1, |
; const uint8_t * const ref1, |
218 |
; const uint8_t * const ref2, |
; const uint8_t * const ref2, |
219 |
; const uint32_t stride); |
; const uint32_t stride); |
220 |
; |
; |
221 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
222 |
|
|
223 |
%macro SADBI_16x16_SSE 0 |
ALIGN SECTION_ALIGN |
|
movq mm0, [eax] |
|
|
movq mm1, [eax+8] |
|
|
|
|
|
movq mm2, [edx] |
|
|
movq mm3, [edx+8] |
|
|
|
|
|
pavgb mm2, [ebx] |
|
|
add edx, ecx |
|
|
|
|
|
pavgb mm3, [ebx+8] |
|
|
add ebx, ecx |
|
|
|
|
|
psadbw mm0, mm2 |
|
|
add eax, ecx |
|
|
|
|
|
psadbw mm1, mm3 |
|
|
paddusw mm5,mm0 |
|
|
|
|
|
paddusw mm6,mm1 |
|
|
%endmacro |
|
|
|
|
|
align 16 |
|
224 |
sad16bi_xmm: |
sad16bi_xmm: |
225 |
push ebx |
mov _EAX, prm1 ; Src |
226 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
227 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
228 |
mov ebx, [esp+4+12] ; Ref2 |
|
229 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
230 |
|
%ifdef ARCH_IS_X86_64 |
231 |
|
mov _EBX, prm3 |
232 |
|
%else |
233 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
234 |
|
%endif |
235 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
236 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
237 |
|
|
255 |
|
|
256 |
paddusw mm6,mm5 |
paddusw mm6,mm5 |
257 |
movd eax, mm6 |
movd eax, mm6 |
258 |
pop ebx |
pop _EBX |
259 |
ret |
ret |
260 |
|
ENDFUNC |
261 |
|
|
262 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
263 |
; |
; |
264 |
; uint32_t sad8bi_xmm(const uint8_t * const cur, |
; uint32_t sad8bi_xmm(const uint8_t * const cur, |
265 |
; const uint8_t * const ref1, |
; const uint8_t * const ref1, |
266 |
; const uint8_t * const ref2, |
; const uint8_t * const ref2, |
267 |
; const uint32_t stride); |
; const uint32_t stride); |
268 |
; |
; |
269 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro SADBI_8x8_XMM 0 |
|
|
movq mm0, [eax] |
|
|
movq mm1, [eax+ecx] |
|
|
|
|
|
movq mm2, [edx] |
|
|
movq mm3, [edx+ecx] |
|
|
|
|
|
pavgb mm2, [ebx] |
|
|
lea edx, [edx+2*ecx] |
|
|
|
|
|
pavgb mm3, [ebx+ecx] |
|
|
lea ebx, [ebx+2*ecx] |
|
|
|
|
|
psadbw mm0, mm2 |
|
|
lea eax, [eax+2*ecx] |
|
|
|
|
|
psadbw mm1, mm3 |
|
|
paddusw mm5,mm0 |
|
|
|
|
|
paddusw mm6,mm1 |
|
|
%endmacro |
|
270 |
|
|
271 |
align 16 |
ALIGN SECTION_ALIGN |
272 |
sad8bi_xmm: |
sad8bi_xmm: |
273 |
push ebx |
mov _EAX, prm1 ; Src |
274 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
275 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
276 |
mov ebx, [esp+4+12] ; Ref2 |
|
277 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
278 |
|
%ifdef ARCH_IS_X86_64 |
279 |
|
mov _EBX, prm3 |
280 |
|
%else |
281 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
282 |
|
%endif |
283 |
|
|
284 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
285 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
286 |
.Loop |
.Loop: |
287 |
SADBI_8x8_XMM |
SADBI_8x8_XMM |
288 |
SADBI_8x8_XMM |
SADBI_8x8_XMM |
289 |
SADBI_8x8_XMM |
SADBI_8x8_XMM |
291 |
|
|
292 |
paddusw mm6,mm5 |
paddusw mm6,mm5 |
293 |
movd eax, mm6 |
movd eax, mm6 |
294 |
pop ebx |
pop _EBX |
295 |
ret |
ret |
296 |
|
ENDFUNC |
297 |
|
|
298 |
|
|
299 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
300 |
; |
; |
301 |
; uint32_t dev16_xmm(const uint8_t * const cur, |
; uint32_t dev16_xmm(const uint8_t * const cur, |
302 |
; const uint32_t stride); |
; const uint32_t stride); |
303 |
; |
; |
304 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro MEAN_16x16_SSE 0 |
|
|
movq mm0, [eax] |
|
|
movq mm1, [eax+8] |
|
|
psadbw mm0, mm7 |
|
|
psadbw mm1, mm7 |
|
|
add eax, ecx |
|
|
paddw mm5, mm0 |
|
|
paddw mm6, mm1 |
|
|
%endmacro |
|
|
|
|
|
%macro ABS_16x16_SSE 0 |
|
|
movq mm0, [eax] |
|
|
movq mm1, [eax+8] |
|
|
psadbw mm0, mm4 |
|
|
psadbw mm1, mm4 |
|
|
lea eax,[eax+ecx] |
|
|
paddw mm5, mm0 |
|
|
paddw mm6, mm1 |
|
|
%endmacro |
|
305 |
|
|
306 |
align 16 |
ALIGN SECTION_ALIGN |
307 |
dev16_xmm: |
dev16_xmm: |
308 |
|
|
309 |
mov eax, [esp+ 4] ; Src |
mov _EAX, prm1 ; Src |
310 |
mov ecx, [esp+ 8] ; Stride |
mov TMP0, prm2 ; Stride |
311 |
|
|
312 |
pxor mm7, mm7 ; zero |
pxor mm7, mm7 ; zero |
313 |
pxor mm5, mm5 ; mean accums |
pxor mm5, mm5 ; mean accums |
343 |
|
|
344 |
; mm4 contains the mean |
; mm4 contains the mean |
345 |
|
|
346 |
mov eax, [esp+ 4] ; Src |
mov _EAX, prm1 ; Src |
347 |
|
|
348 |
|
|
349 |
pxor mm5, mm5 ; sums |
pxor mm5, mm5 ; sums |
374 |
|
|
375 |
movd eax, mm6 |
movd eax, mm6 |
376 |
ret |
ret |
377 |
|
ENDFUNC |
378 |
|
|
379 |
cglobal sad16v_xmm |
;----------------------------------------------------------------------------- |
|
|
|
|
;=========================================================================== |
|
380 |
;int sad16v_xmm(const uint8_t * const cur, |
;int sad16v_xmm(const uint8_t * const cur, |
381 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
382 |
; const uint32_t stride, |
; const uint32_t stride, |
383 |
; int* sad8); |
; int* sad8); |
384 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
385 |
align 16 |
|
386 |
|
ALIGN SECTION_ALIGN |
387 |
sad16v_xmm: |
sad16v_xmm: |
388 |
push ebx |
mov _EAX, prm1 ; Src1 |
389 |
mov eax, [esp+4+ 4] ; Src1 |
mov TMP1, prm2 ; Src2 |
390 |
mov edx, [esp+4+ 8] ; Src2 |
mov TMP0, prm3 ; Stride |
391 |
mov ecx, [esp+4+12] ; Stride |
|
392 |
mov ebx, [esp+4+16] ; sad ptr |
push _EBX |
393 |
|
%ifdef ARCH_IS_X86_64 |
394 |
|
mov _EBX, prm4 |
395 |
|
%else |
396 |
|
mov _EBX, [_ESP+4+16] ; sad ptr |
397 |
|
%endif |
398 |
|
|
399 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
400 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
401 |
pxor mm7, mm7 ; total |
pxor mm7, mm7 ; total |
402 |
|
|
403 |
SAD_16x16_SSE |
SAD_16x16_SSE |
404 |
SAD_16x16_SSE |
SAD_16x16_SSE |
405 |
SAD_16x16_SSE |
SAD_16x16_SSE |
408 |
SAD_16x16_SSE |
SAD_16x16_SSE |
409 |
SAD_16x16_SSE |
SAD_16x16_SSE |
410 |
SAD_16x16_SSE |
SAD_16x16_SSE |
411 |
|
|
412 |
paddusw mm7, mm5 |
paddusw mm7, mm5 |
413 |
paddusw mm7, mm6 |
paddusw mm7, mm6 |
414 |
movd [ebx], mm5 |
movd [_EBX], mm5 |
415 |
movd [ebx+4], mm6 |
movd [_EBX+4], mm6 |
416 |
|
|
417 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
418 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
419 |
|
|
420 |
SAD_16x16_SSE |
SAD_16x16_SSE |
421 |
SAD_16x16_SSE |
SAD_16x16_SSE |
422 |
SAD_16x16_SSE |
SAD_16x16_SSE |
425 |
SAD_16x16_SSE |
SAD_16x16_SSE |
426 |
SAD_16x16_SSE |
SAD_16x16_SSE |
427 |
SAD_16x16_SSE |
SAD_16x16_SSE |
428 |
|
|
429 |
paddusw mm7, mm5 |
paddusw mm7, mm5 |
430 |
paddusw mm7, mm6 |
paddusw mm7, mm6 |
431 |
movd [ebx+8], mm5 |
movd [_EBX+8], mm5 |
432 |
movd [ebx+12], mm6 |
movd [_EBX+12], mm6 |
433 |
|
|
434 |
movd eax, mm7 |
movd eax, mm7 |
435 |
pop ebx |
pop _EBX |
436 |
ret |
ret |
437 |
;-------- |
ENDFUNC |
438 |
|
|
439 |
|
|
440 |
|
%ifidn __OUTPUT_FORMAT__,elf |
441 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
442 |
|
%endif |
443 |
|
|