1 |
;/************************************************************************** |
;/**************************************************************************** |
2 |
; * |
; * |
3 |
; * XVID MPEG-4 VIDEO CODEC |
; * XVID MPEG-4 VIDEO CODEC |
4 |
; * mmx sum of absolute difference |
; * - K7 optimized SAD operators - |
5 |
; * |
; * |
6 |
; * This program is free software; you can redistribute it and/or modify |
; * Copyright(C) 2001 Peter Ross <pross@xvid.org> |
7 |
; * it under the terms of the GNU General Public License as published by |
; * 2002 Pascal Massimino <skal@planet-d.net> |
8 |
|
; * |
9 |
|
; * This program is free software; you can redistribute it and/or modify it |
10 |
|
; * under the terms of the GNU General Public License as published by |
11 |
; * the Free Software Foundation; either version 2 of the License, or |
; * the Free Software Foundation; either version 2 of the License, or |
12 |
; * (at your option) any later version. |
; * (at your option) any later version. |
13 |
; * |
; * |
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 |
; * History: |
; * $Id: sad_mmx.asm,v 1.20 2008-11-26 01:04:34 Isibaar Exp $ |
24 |
; * |
; * |
25 |
; * 23.07.2002 sad[16,8]bi_xmm; <pross@xvid.org> |
; ***************************************************************************/ |
|
; * 04.06.2002 cleanup -Skal- |
|
|
; * 12.11.2001 inital version; (c)2001 peter ross <pross@cs.rmit.edu.au> |
|
|
; * |
|
|
; *************************************************************************/ |
|
|
|
|
|
bits 32 |
|
|
|
|
|
%macro cglobal 1 |
|
|
%ifdef PREFIX |
|
|
global _%1 |
|
|
%define %1 _%1 |
|
|
%else |
|
|
global %1 |
|
|
%endif |
|
|
%endmacro |
|
26 |
|
|
27 |
section .data |
%include "nasm.inc" |
28 |
|
|
29 |
align 16 |
;============================================================================= |
30 |
mmx_one times 4 dw 1 |
; Read only data |
31 |
|
;============================================================================= |
32 |
|
|
33 |
section .text |
DATA |
34 |
|
|
35 |
cglobal sad16_mmx |
ALIGN SECTION_ALIGN |
36 |
cglobal sad8_mmx |
mmx_one: |
37 |
cglobal sad16bi_mmx |
times 4 dw 1 |
|
cglobal sad8bi_mmx |
|
|
cglobal dev16_mmx |
|
38 |
|
|
39 |
;=========================================================================== |
;============================================================================= |
40 |
; |
; Helper macros |
41 |
; uint32_t sad16_mmx(const uint8_t * const cur, |
;============================================================================= |
|
; const uint8_t * const ref, |
|
|
; const uint32_t stride, |
|
|
; const uint32_t best_sad); |
|
|
; |
|
|
; (early termination ignore; slows this down) |
|
|
; |
|
|
;=========================================================================== |
|
42 |
|
|
43 |
%macro SAD_16x16_MMX 0 |
%macro SAD_16x16_MMX 0 |
44 |
movq mm0, [eax] |
movq mm0, [_EAX] |
45 |
movq mm1, [edx] |
movq mm1, [TMP1] |
46 |
|
|
47 |
movq mm2, [eax+8] |
movq mm2, [_EAX+8] |
48 |
movq mm3, [edx+8] |
movq mm3, [TMP1+8] |
49 |
|
|
50 |
movq mm4, mm0 |
movq mm4, mm0 |
51 |
psubusb mm0, mm1 |
psubusb mm0, mm1 |
52 |
lea eax,[eax+ecx] |
lea _EAX, [_EAX+TMP0] |
53 |
movq mm5, mm2 |
movq mm5, mm2 |
54 |
psubusb mm2, mm3 |
psubusb mm2, mm3 |
|
lea edx,[edx+ecx] |
|
55 |
|
|
56 |
psubusb mm1, mm4 |
psubusb mm1, mm4 |
|
por mm0, mm1 |
|
57 |
psubusb mm3, mm5 |
psubusb mm3, mm5 |
58 |
|
por mm0, mm1 |
59 |
por mm2, mm3 |
por mm2, mm3 |
60 |
|
|
61 |
movq mm1,mm0 |
movq mm1,mm0 |
62 |
|
punpcklbw mm0,mm7 |
63 |
movq mm3,mm2 |
movq mm3,mm2 |
64 |
|
punpckhbw mm1,mm7 |
65 |
|
lea TMP1, [TMP1+TMP0] |
66 |
|
punpcklbw mm2,mm7 |
67 |
|
paddusw mm0, mm1 |
68 |
|
punpckhbw mm3,mm7 |
69 |
|
paddusw mm6, mm0 |
70 |
|
paddusw mm2, mm3 |
71 |
|
paddusw mm6, mm2 |
72 |
|
|
73 |
|
%endmacro |
74 |
|
|
75 |
|
%macro SAD_8x8_MMX 0 |
76 |
|
movq mm0, [_EAX] |
77 |
|
movq mm1, [TMP1] |
78 |
|
|
79 |
|
movq mm2, [_EAX+TMP0] |
80 |
|
movq mm3, [TMP1+TMP0] |
81 |
|
|
82 |
|
lea _EAX,[_EAX+2*TMP0] |
83 |
|
lea TMP1,[TMP1+2*TMP0] |
84 |
|
|
85 |
|
movq mm4, mm0 |
86 |
|
psubusb mm0, mm1 |
87 |
|
movq mm5, mm2 |
88 |
|
psubusb mm2, mm3 |
89 |
|
|
90 |
|
psubusb mm1, mm4 |
91 |
|
psubusb mm3, mm5 |
92 |
|
por mm0, mm1 |
93 |
|
por mm2, mm3 |
94 |
|
|
95 |
|
movq mm1,mm0 |
96 |
punpcklbw mm0,mm7 |
punpcklbw mm0,mm7 |
97 |
|
movq mm3,mm2 |
98 |
punpckhbw mm1,mm7 |
punpckhbw mm1,mm7 |
99 |
punpcklbw mm2,mm7 |
punpcklbw mm2,mm7 |
100 |
|
paddusw mm0,mm1 |
101 |
punpckhbw mm3,mm7 |
punpckhbw mm3,mm7 |
102 |
|
paddusw mm6,mm0 |
103 |
|
paddusw mm2,mm3 |
104 |
|
paddusw mm6,mm2 |
105 |
|
%endmacro |
106 |
|
|
107 |
|
|
108 |
|
%macro SADV_16x16_MMX 0 |
109 |
|
movq mm0, [_EAX] |
110 |
|
movq mm1, [TMP1] |
111 |
|
|
112 |
|
movq mm2, [_EAX+8] |
113 |
|
movq mm4, mm0 |
114 |
|
movq mm3, [TMP1+8] |
115 |
|
psubusb mm0, mm1 |
116 |
|
|
117 |
|
psubusb mm1, mm4 |
118 |
|
lea _EAX,[_EAX+TMP0] |
119 |
|
por mm0, mm1 |
120 |
|
|
121 |
|
movq mm4, mm2 |
122 |
|
psubusb mm2, mm3 |
123 |
|
|
124 |
|
psubusb mm3, mm4 |
125 |
|
por mm2, mm3 |
126 |
|
|
127 |
|
movq mm1,mm0 |
128 |
|
punpcklbw mm0,mm7 |
129 |
|
movq mm3,mm2 |
130 |
|
punpckhbw mm1,mm7 |
131 |
|
punpcklbw mm2,mm7 |
132 |
paddusw mm0,mm1 |
paddusw mm0,mm1 |
133 |
paddusw mm6,mm0 |
punpckhbw mm3,mm7 |
134 |
|
paddusw mm5, mm0 |
135 |
paddusw mm2,mm3 |
paddusw mm2,mm3 |
136 |
|
lea TMP1,[TMP1+TMP0] |
137 |
paddusw mm6,mm2 |
paddusw mm6,mm2 |
138 |
%endmacro |
%endmacro |
139 |
|
|
140 |
align 16 |
%macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
141 |
|
|
142 |
|
movq mm0, [TMP1+%1] |
143 |
|
movq mm2, [_EBX+%1] |
144 |
|
movq mm1, mm0 |
145 |
|
movq mm3, mm2 |
146 |
|
|
147 |
|
%if %2 != 0 |
148 |
|
add TMP1, TMP0 |
149 |
|
%endif |
150 |
|
|
151 |
|
punpcklbw mm0, mm7 |
152 |
|
punpckhbw mm1, mm7 |
153 |
|
punpcklbw mm2, mm7 |
154 |
|
punpckhbw mm3, mm7 |
155 |
|
|
156 |
|
%if %2 != 0 |
157 |
|
add _EBX, TMP0 |
158 |
|
%endif |
159 |
|
|
160 |
|
paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
161 |
|
paddusw mm1, mm3 |
162 |
|
paddusw mm0, [mmx_one] ; mm01 += 1 |
163 |
|
paddusw mm1, [mmx_one] |
164 |
|
psrlw mm0, 1 ; mm01 >>= 1 |
165 |
|
psrlw mm1, 1 |
166 |
|
|
167 |
|
movq mm2, [_EAX+%1] |
168 |
|
movq mm3, mm2 |
169 |
|
punpcklbw mm2, mm7 ; mm23 = src |
170 |
|
punpckhbw mm3, mm7 |
171 |
|
|
172 |
|
%if %2 != 0 |
173 |
|
add _EAX, TMP0 |
174 |
|
%endif |
175 |
|
|
176 |
|
movq mm4, mm0 |
177 |
|
movq mm5, mm1 |
178 |
|
psubusw mm0, mm2 |
179 |
|
psubusw mm1, mm3 |
180 |
|
psubusw mm2, mm4 |
181 |
|
psubusw mm3, mm5 |
182 |
|
por mm0, mm2 ; mm01 = ABS(mm01 - mm23) |
183 |
|
por mm1, mm3 |
184 |
|
|
185 |
|
paddusw mm6, mm0 ; mm6 += mm01 |
186 |
|
paddusw mm6, mm1 |
187 |
|
|
188 |
|
%endmacro |
189 |
|
|
190 |
|
%macro MEAN_16x16_MMX 0 |
191 |
|
movq mm0, [_EAX] |
192 |
|
movq mm2, [_EAX+8] |
193 |
|
lea _EAX, [_EAX+TMP0] |
194 |
|
movq mm1, mm0 |
195 |
|
punpcklbw mm0, mm7 |
196 |
|
movq mm3, mm2 |
197 |
|
punpckhbw mm1, mm7 |
198 |
|
paddw mm5, mm0 |
199 |
|
punpcklbw mm2, mm7 |
200 |
|
paddw mm6, mm1 |
201 |
|
punpckhbw mm3, mm7 |
202 |
|
paddw mm5, mm2 |
203 |
|
paddw mm6, mm3 |
204 |
|
%endmacro |
205 |
|
|
206 |
|
%macro ABS_16x16_MMX 0 |
207 |
|
movq mm0, [_EAX] |
208 |
|
movq mm2, [_EAX+8] |
209 |
|
lea _EAX, [_EAX+TMP0] |
210 |
|
movq mm1, mm0 |
211 |
|
movq mm3, mm2 |
212 |
|
punpcklbw mm0, mm7 |
213 |
|
punpcklbw mm2, mm7 |
214 |
|
punpckhbw mm1, mm7 |
215 |
|
punpckhbw mm3, mm7 |
216 |
|
movq mm4, mm6 |
217 |
|
psubusw mm4, mm0 |
218 |
|
|
219 |
|
psubusw mm0, mm6 |
220 |
|
por mm0, mm4 |
221 |
|
movq mm4, mm6 |
222 |
|
psubusw mm4, mm1 |
223 |
|
psubusw mm1, mm6 |
224 |
|
por mm1, mm4 |
225 |
|
|
226 |
|
movq mm4, mm6 |
227 |
|
psubusw mm4, mm2 |
228 |
|
psubusw mm2, mm6 |
229 |
|
por mm2, mm4 |
230 |
|
movq mm4, mm6 |
231 |
|
psubusw mm4, mm3 |
232 |
|
psubusw mm3, mm6 |
233 |
|
por mm3, mm4 |
234 |
|
|
235 |
|
paddw mm0, mm1 |
236 |
|
paddw mm2, mm3 |
237 |
|
paddw mm5, mm0 |
238 |
|
paddw mm5, mm2 |
239 |
|
%endmacro |
240 |
|
|
241 |
|
;============================================================================= |
242 |
|
; Code |
243 |
|
;============================================================================= |
244 |
|
|
245 |
|
SECTION .rotext align=SECTION_ALIGN |
246 |
|
|
247 |
|
cglobal sad16_mmx |
248 |
|
cglobal sad16v_mmx |
249 |
|
cglobal sad8_mmx |
250 |
|
cglobal sad16bi_mmx |
251 |
|
cglobal sad8bi_mmx |
252 |
|
cglobal dev16_mmx |
253 |
|
cglobal sse8_16bit_mmx |
254 |
|
cglobal sse8_8bit_mmx |
255 |
|
|
256 |
|
;----------------------------------------------------------------------------- |
257 |
|
; |
258 |
|
; uint32_t sad16_mmx(const uint8_t * const cur, |
259 |
|
; const uint8_t * const ref, |
260 |
|
; const uint32_t stride, |
261 |
|
; const uint32_t best_sad); |
262 |
|
; |
263 |
|
; (early termination ignore; slows this down) |
264 |
|
; |
265 |
|
;----------------------------------------------------------------------------- |
266 |
|
|
267 |
|
ALIGN SECTION_ALIGN |
268 |
sad16_mmx: |
sad16_mmx: |
269 |
|
|
270 |
mov eax, [esp+ 4] ; Src1 |
mov _EAX, prm1 ; Src1 |
271 |
mov edx, [esp+ 8] ; Src2 |
mov TMP1, prm2 ; Src2 |
272 |
mov ecx, [esp+12] ; Stride |
mov TMP0, prm3 ; Stride |
273 |
|
|
274 |
pxor mm6, mm6 ; accum |
pxor mm6, mm6 ; accum |
275 |
pxor mm7, mm7 ; zero |
pxor mm7, mm7 ; zero |
300 |
movd eax, mm6 |
movd eax, mm6 |
301 |
|
|
302 |
ret |
ret |
303 |
|
ENDFUNC |
304 |
|
|
305 |
|
;----------------------------------------------------------------------------- |
|
;=========================================================================== |
|
306 |
; |
; |
307 |
; uint32_t sad8_mmx(const uint8_t * const cur, |
; uint32_t sad8_mmx(const uint8_t * const cur, |
308 |
; const uint8_t * const ref, |
; const uint8_t * const ref, |
309 |
; const uint32_t stride); |
; const uint32_t stride); |
310 |
; |
; |
311 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro SAD_8x8_MMX 0 |
|
|
movq mm0, [eax] |
|
|
movq mm1, [edx] |
|
312 |
|
|
313 |
movq mm2, [eax+ecx] |
ALIGN SECTION_ALIGN |
|
movq mm3, [edx+ecx] |
|
|
|
|
|
lea eax,[eax+2*ecx] |
|
|
lea edx,[edx+2*ecx] |
|
|
|
|
|
movq mm4, mm0 |
|
|
psubusb mm0, mm1 |
|
|
movq mm5, mm2 |
|
|
psubusb mm2, mm3 |
|
|
|
|
|
psubusb mm1, mm4 |
|
|
por mm0, mm1 |
|
|
psubusb mm3, mm5 |
|
|
por mm2, mm3 |
|
|
|
|
|
movq mm1,mm0 |
|
|
movq mm3,mm2 |
|
|
|
|
|
punpcklbw mm0,mm7 |
|
|
punpckhbw mm1,mm7 |
|
|
punpcklbw mm2,mm7 |
|
|
punpckhbw mm3,mm7 |
|
|
|
|
|
paddusw mm0,mm1 |
|
|
paddusw mm6,mm0 |
|
|
paddusw mm2,mm3 |
|
|
paddusw mm6,mm2 |
|
|
%endmacro |
|
|
|
|
|
align 16 |
|
314 |
sad8_mmx: |
sad8_mmx: |
315 |
|
|
316 |
mov eax, [esp+ 4] ; Src1 |
mov _EAX, prm1 ; Src1 |
317 |
mov edx, [esp+ 8] ; Src2 |
mov TMP1, prm2 ; Src2 |
318 |
mov ecx, [esp+12] ; Stride |
mov TMP0, prm3 ; Stride |
319 |
|
|
320 |
pxor mm6, mm6 ; accum |
pxor mm6, mm6 ; accum |
321 |
pxor mm7, mm7 ; zero |
pxor mm7, mm7 ; zero |
333 |
movd eax, mm6 |
movd eax, mm6 |
334 |
|
|
335 |
ret |
ret |
336 |
|
ENDFUNC |
337 |
|
|
338 |
|
;----------------------------------------------------------------------------- |
339 |
|
; |
340 |
|
; uint32_t sad16v_mmx(const uint8_t * const cur, |
341 |
|
; const uint8_t * const ref, |
342 |
|
; const uint32_t stride, |
343 |
|
; int32_t *sad); |
344 |
|
; |
345 |
|
;----------------------------------------------------------------------------- |
346 |
|
|
347 |
|
ALIGN SECTION_ALIGN |
348 |
|
sad16v_mmx: |
349 |
|
|
350 |
|
mov _EAX, prm1 ; Src1 |
351 |
|
mov TMP1, prm2 ; Src2 |
352 |
|
mov TMP0, prm3 ; Stride |
353 |
|
|
354 |
|
push _EBX |
355 |
|
push _EDI |
356 |
|
%ifdef ARCH_IS_X86_64 |
357 |
|
mov _EBX, prm4 |
358 |
|
%else |
359 |
|
mov _EBX, [_ESP + 8 + 16] ; sad ptr |
360 |
|
%endif |
361 |
|
|
362 |
;=========================================================================== |
pxor mm5, mm5 ; accum |
363 |
; |
pxor mm6, mm6 ; accum |
364 |
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
pxor mm7, mm7 ; zero |
|
; const uint8_t * const ref1, |
|
|
; const uint8_t * const ref2, |
|
|
; const uint32_t stride); |
|
|
; |
|
|
;=========================================================================== |
|
|
%macro SADBI_16x16_MMX 2 ; SADBI_16x16_MMX( int_ptr_offset, bool_increment_ptr ); |
|
365 |
|
|
366 |
movq mm0, [edx+%1] |
SADV_16x16_MMX |
367 |
movq mm2, [ebx+%1] |
SADV_16x16_MMX |
368 |
movq mm1, mm0 |
SADV_16x16_MMX |
369 |
movq mm3, mm2 |
SADV_16x16_MMX |
370 |
|
SADV_16x16_MMX |
371 |
|
SADV_16x16_MMX |
372 |
|
SADV_16x16_MMX |
373 |
|
SADV_16x16_MMX |
374 |
|
|
375 |
%if %2 != 0 |
pmaddwd mm5, [mmx_one] ; collapse |
376 |
add edx, ecx |
pmaddwd mm6, [mmx_one] ; collapse |
|
%endif |
|
377 |
|
|
378 |
punpcklbw mm0, mm7 |
movq mm2, mm5 |
379 |
punpckhbw mm1, mm7 |
movq mm3, mm6 |
|
punpcklbw mm2, mm7 |
|
|
punpckhbw mm3, mm7 |
|
380 |
|
|
381 |
%if %2 != 0 |
psrlq mm2, 32 |
382 |
add ebx, ecx |
psrlq mm3, 32 |
|
%endif |
|
383 |
|
|
384 |
paddusw mm0, mm2 ; mm01 = ref1 + ref2 |
paddd mm5, mm2 |
385 |
paddusw mm1, mm3 |
paddd mm6, mm3 |
|
paddusw mm0, [mmx_one] ; mm01 += 1 |
|
|
paddusw mm1, [mmx_one] |
|
|
psrlw mm0, 1 ; mm01 >>= 1 |
|
|
psrlw mm1, 1 |
|
386 |
|
|
387 |
movq mm2, [eax+%1] |
movd [_EBX], mm5 |
388 |
movq mm3, mm2 |
movd [_EBX + 4], mm6 |
|
punpcklbw mm2, mm7 ; mm23 = src |
|
|
punpckhbw mm3, mm7 |
|
389 |
|
|
390 |
%if %2 != 0 |
paddd mm5, mm6 |
|
add eax, ecx |
|
|
%endif |
|
391 |
|
|
392 |
movq mm4, mm0 |
movd edi, mm5 |
|
movq mm5, mm1 |
|
|
psubusw mm0, mm2 |
|
|
psubusw mm1, mm3 |
|
|
psubusw mm2, mm4 |
|
|
psubusw mm3, mm5 |
|
|
por mm0, mm2 ; mm01 = ABS(mm01 - mm23) |
|
|
por mm1, mm3 |
|
393 |
|
|
394 |
paddusw mm6,mm0 ; mm6 += mm01 |
pxor mm5, mm5 |
395 |
paddusw mm6,mm1 |
pxor mm6, mm6 |
396 |
|
|
397 |
%endmacro |
SADV_16x16_MMX |
398 |
|
SADV_16x16_MMX |
399 |
|
SADV_16x16_MMX |
400 |
|
SADV_16x16_MMX |
401 |
|
SADV_16x16_MMX |
402 |
|
SADV_16x16_MMX |
403 |
|
SADV_16x16_MMX |
404 |
|
SADV_16x16_MMX |
405 |
|
|
406 |
align 16 |
pmaddwd mm5, [mmx_one] ; collapse |
407 |
|
pmaddwd mm6, [mmx_one] ; collapse |
408 |
|
|
409 |
|
movq mm2, mm5 |
410 |
|
movq mm3, mm6 |
411 |
|
|
412 |
|
psrlq mm2, 32 |
413 |
|
psrlq mm3, 32 |
414 |
|
|
415 |
|
paddd mm5, mm2 |
416 |
|
paddd mm6, mm3 |
417 |
|
|
418 |
|
movd [_EBX + 8], mm5 |
419 |
|
movd [_EBX + 12], mm6 |
420 |
|
|
421 |
|
paddd mm5, mm6 |
422 |
|
|
423 |
|
movd eax, mm5 |
424 |
|
|
425 |
|
add _EAX, _EDI |
426 |
|
|
427 |
|
pop _EDI |
428 |
|
pop _EBX |
429 |
|
|
430 |
|
ret |
431 |
|
ENDFUNC |
432 |
|
|
433 |
|
;----------------------------------------------------------------------------- |
434 |
|
; |
435 |
|
; uint32_t sad16bi_mmx(const uint8_t * const cur, |
436 |
|
; const uint8_t * const ref1, |
437 |
|
; const uint8_t * const ref2, |
438 |
|
; const uint32_t stride); |
439 |
|
; |
440 |
|
;----------------------------------------------------------------------------- |
441 |
|
|
442 |
|
ALIGN SECTION_ALIGN |
443 |
sad16bi_mmx: |
sad16bi_mmx: |
444 |
push ebx |
mov _EAX, prm1 ; Src |
445 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
446 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
447 |
mov ebx, [esp+4+12] ; Ref2 |
|
448 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
449 |
|
%ifdef ARCH_IS_X86_64 |
450 |
|
mov _EBX, prm3 ; Ref2 |
451 |
|
%else |
452 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
453 |
|
%endif |
454 |
|
|
455 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
456 |
pxor mm7, mm7 |
pxor mm7, mm7 |
457 |
.Loop |
.Loop: |
458 |
SADBI_16x16_MMX 0, 0 |
SADBI_16x16_MMX 0, 0 |
459 |
SADBI_16x16_MMX 8, 1 |
SADBI_16x16_MMX 8, 1 |
460 |
SADBI_16x16_MMX 0, 0 |
SADBI_16x16_MMX 0, 0 |
495 |
paddd mm6, mm7 |
paddd mm6, mm7 |
496 |
|
|
497 |
movd eax, mm6 |
movd eax, mm6 |
498 |
pop ebx |
pop _EBX |
499 |
|
|
500 |
ret |
ret |
501 |
|
ENDFUNC |
502 |
|
|
503 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
504 |
; |
; |
505 |
; uint32_t sad8bi_mmx(const uint8_t * const cur, |
; uint32_t sad8bi_mmx(const uint8_t * const cur, |
506 |
; const uint8_t * const ref1, |
; const uint8_t * const ref1, |
507 |
; const uint8_t * const ref2, |
; const uint8_t * const ref2, |
508 |
; const uint32_t stride); |
; const uint32_t stride); |
509 |
; |
; |
510 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
511 |
align 16 |
|
512 |
|
ALIGN SECTION_ALIGN |
513 |
sad8bi_mmx: |
sad8bi_mmx: |
514 |
push ebx |
mov _EAX, prm1 ; Src |
515 |
mov eax, [esp+4+ 4] ; Src |
mov TMP1, prm2 ; Ref1 |
516 |
mov edx, [esp+4+ 8] ; Ref1 |
mov TMP0, prm4 ; Stride |
517 |
mov ebx, [esp+4+12] ; Ref2 |
|
518 |
mov ecx, [esp+4+16] ; Stride |
push _EBX |
519 |
|
%ifdef ARCH_IS_X86_64 |
520 |
|
mov _EBX, prm3 |
521 |
|
%else |
522 |
|
mov _EBX, [_ESP+4+12] ; Ref2 |
523 |
|
%endif |
524 |
|
|
525 |
pxor mm6, mm6 ; accum2 |
pxor mm6, mm6 ; accum2 |
526 |
pxor mm7, mm7 |
pxor mm7, mm7 |
527 |
.Loop |
.Loop: |
528 |
SADBI_16x16_MMX 0, 1 |
SADBI_16x16_MMX 0, 1 |
529 |
SADBI_16x16_MMX 0, 1 |
SADBI_16x16_MMX 0, 1 |
530 |
SADBI_16x16_MMX 0, 1 |
SADBI_16x16_MMX 0, 1 |
540 |
paddd mm6, mm7 |
paddd mm6, mm7 |
541 |
|
|
542 |
movd eax, mm6 |
movd eax, mm6 |
543 |
pop ebx |
pop _EBX |
544 |
ret |
ret |
545 |
|
ENDFUNC |
546 |
|
|
547 |
|
;----------------------------------------------------------------------------- |
|
|
|
|
|
|
|
;=========================================================================== |
|
548 |
; |
; |
549 |
; uint32_t dev16_mmx(const uint8_t * const cur, |
; uint32_t dev16_mmx(const uint8_t * const cur, |
550 |
; const uint32_t stride); |
; const uint32_t stride); |
551 |
; |
; |
552 |
;=========================================================================== |
;----------------------------------------------------------------------------- |
|
|
|
|
%macro MEAN_16x16_MMX 0 |
|
|
movq mm0, [eax] |
|
|
movq mm2, [eax+8] |
|
|
lea eax,[eax+ecx] |
|
|
movq mm1, mm0 |
|
|
movq mm3, mm2 |
|
|
punpcklbw mm0,mm7 |
|
|
punpcklbw mm2,mm7 |
|
|
punpckhbw mm1,mm7 |
|
|
punpckhbw mm3,mm7 |
|
|
paddw mm5, mm0 |
|
|
paddw mm6, mm1 |
|
|
paddw mm5, mm2 |
|
|
paddw mm6, mm3 |
|
|
%endmacro |
|
553 |
|
|
554 |
%macro ABS_16x16_MMX 0 |
ALIGN SECTION_ALIGN |
|
movq mm0, [eax] |
|
|
movq mm2, [eax+8] |
|
|
lea eax,[eax+ecx] |
|
|
movq mm1, mm0 |
|
|
movq mm3, mm2 |
|
|
punpcklbw mm0, mm7 |
|
|
punpcklbw mm2, mm7 |
|
|
punpckhbw mm1, mm7 |
|
|
punpckhbw mm3, mm7 |
|
|
movq mm4, mm6 |
|
|
psubusw mm4, mm0 |
|
|
|
|
|
psubusw mm0, mm6 |
|
|
por mm0, mm4 |
|
|
movq mm4, mm6 |
|
|
psubusw mm4, mm1 |
|
|
psubusw mm1, mm6 |
|
|
por mm1, mm4 |
|
|
|
|
|
movq mm4, mm6 |
|
|
psubusw mm4, mm2 |
|
|
psubusw mm2, mm6 |
|
|
por mm2, mm4 |
|
|
movq mm4, mm6 |
|
|
psubusw mm4, mm3 |
|
|
psubusw mm3, mm6 |
|
|
por mm3, mm4 |
|
|
|
|
|
paddw mm0, mm1 |
|
|
paddw mm2, mm3 |
|
|
paddw mm5, mm0 |
|
|
paddw mm5, mm2 |
|
|
%endmacro |
|
|
|
|
|
align 16 |
|
555 |
dev16_mmx: |
dev16_mmx: |
556 |
mov eax, [esp+ 4] ; Src |
mov _EAX, prm1 ; Src |
557 |
mov ecx, [esp+ 8] ; Stride |
mov TMP0, prm2 ; Stride |
558 |
|
|
559 |
pxor mm7, mm7 ; zero |
pxor mm7, mm7 ; zero |
560 |
pxor mm5, mm5 ; accum1 |
pxor mm5, mm5 ; accum1 |
594 |
; mm5 is the new accum |
; mm5 is the new accum |
595 |
|
|
596 |
pxor mm5, mm5 |
pxor mm5, mm5 |
597 |
mov eax, [esp+ 4] ; Src |
mov _EAX, prm1 ; Src |
598 |
|
|
599 |
ABS_16x16_MMX |
ABS_16x16_MMX |
600 |
ABS_16x16_MMX |
ABS_16x16_MMX |
620 |
paddd mm6, mm5 |
paddd mm6, mm5 |
621 |
|
|
622 |
movd eax, mm6 |
movd eax, mm6 |
623 |
|
|
624 |
|
ret |
625 |
|
ENDFUNC |
626 |
|
|
627 |
|
;----------------------------------------------------------------------------- |
628 |
|
; |
629 |
|
; uint32_t sse8_16bit_mmx(const int16_t *b1, |
630 |
|
; const int16_t *b2, |
631 |
|
; const uint32_t stride); |
632 |
|
; |
633 |
|
;----------------------------------------------------------------------------- |
634 |
|
|
635 |
|
%macro ROW_SSE_16bit_MMX 2 |
636 |
|
movq mm0, [%1] |
637 |
|
movq mm1, [%1+8] |
638 |
|
psubw mm0, [%2] |
639 |
|
psubw mm1, [%2+8] |
640 |
|
pmaddwd mm0, mm0 |
641 |
|
pmaddwd mm1, mm1 |
642 |
|
paddd mm2, mm0 |
643 |
|
paddd mm2, mm1 |
644 |
|
%endmacro |
645 |
|
|
646 |
|
sse8_16bit_mmx: |
647 |
|
|
648 |
|
;; Load the function params |
649 |
|
mov _EAX, prm1 |
650 |
|
mov TMP0, prm2 |
651 |
|
mov TMP1, prm3 |
652 |
|
|
653 |
|
;; Reset the sse accumulator |
654 |
|
pxor mm2, mm2 |
655 |
|
|
656 |
|
;; Let's go |
657 |
|
%rep 8 |
658 |
|
ROW_SSE_16bit_MMX _EAX, TMP0 |
659 |
|
lea _EAX, [_EAX+TMP1] |
660 |
|
lea TMP0, [TMP0+TMP1] |
661 |
|
%endrep |
662 |
|
|
663 |
|
;; Finish adding each dword of the accumulator |
664 |
|
movq mm3, mm2 |
665 |
|
psrlq mm2, 32 |
666 |
|
paddd mm2, mm3 |
667 |
|
movd eax, mm2 |
668 |
|
|
669 |
|
;; All done |
670 |
|
ret |
671 |
|
ENDFUNC |
672 |
|
|
673 |
|
;----------------------------------------------------------------------------- |
674 |
|
; |
675 |
|
; uint32_t sse8_8bit_mmx(const int8_t *b1, |
676 |
|
; const int8_t *b2, |
677 |
|
; const uint32_t stride); |
678 |
|
; |
679 |
|
;----------------------------------------------------------------------------- |
680 |
|
|
681 |
|
%macro ROW_SSE_8bit_MMX 2 |
682 |
|
movq mm0, [%1] ; load a row |
683 |
|
movq mm2, [%2] ; load a row |
684 |
|
|
685 |
|
movq mm1, mm0 ; copy row |
686 |
|
movq mm3, mm2 ; copy row |
687 |
|
|
688 |
|
punpcklbw mm0, mm7 ; turn the 4low elements into 16bit |
689 |
|
punpckhbw mm1, mm7 ; turn the 4high elements into 16bit |
690 |
|
|
691 |
|
punpcklbw mm2, mm7 ; turn the 4low elements into 16bit |
692 |
|
punpckhbw mm3, mm7 ; turn the 4high elements into 16bit |
693 |
|
|
694 |
|
psubw mm0, mm2 ; low part of src-dst |
695 |
|
psubw mm1, mm3 ; high part of src-dst |
696 |
|
|
697 |
|
pmaddwd mm0, mm0 ; compute the square sum |
698 |
|
pmaddwd mm1, mm1 ; compute the square sum |
699 |
|
|
700 |
|
paddd mm6, mm0 ; add to the accumulator |
701 |
|
paddd mm6, mm1 ; add to the accumulator |
702 |
|
%endmacro |
703 |
|
|
704 |
|
sse8_8bit_mmx: |
705 |
|
|
706 |
|
;; Load the function params |
707 |
|
mov _EAX, prm1 |
708 |
|
mov TMP0, prm2 |
709 |
|
mov TMP1, prm3 |
710 |
|
|
711 |
|
;; Reset the sse accumulator |
712 |
|
pxor mm6, mm6 |
713 |
|
|
714 |
|
;; Used to interleave 8bit data with 0x00 values |
715 |
|
pxor mm7, mm7 |
716 |
|
|
717 |
|
;; Let's go |
718 |
|
%rep 8 |
719 |
|
ROW_SSE_8bit_MMX _EAX, TMP0 |
720 |
|
lea _EAX, [_EAX+TMP1] |
721 |
|
lea TMP0, [TMP0+TMP1] |
722 |
|
%endrep |
723 |
|
|
724 |
|
;; Finish adding each dword of the accumulator |
725 |
|
movq mm7, mm6 |
726 |
|
psrlq mm6, 32 |
727 |
|
paddd mm6, mm7 |
728 |
|
movd eax, mm6 |
729 |
|
|
730 |
|
;; All done |
731 |
ret |
ret |
732 |
|
ENDFUNC |
733 |
|
|
734 |
|
|
735 |
|
%ifidn __OUTPUT_FORMAT__,elf |
736 |
|
section ".note.GNU-stack" noalloc noexec nowrite progbits |
737 |
|
%endif |
738 |
|
|