1 |
|
;/************************************************************************** |
2 |
|
; * |
3 |
|
; * XVID MPEG-4 VIDEO CODEC |
4 |
|
; * - 3dne Quantization/Dequantization - |
5 |
|
; * |
6 |
|
; * Copyright(C) 2002-2003 Jaan Kalda |
7 |
|
; * |
8 |
|
; * This program is free software ; you can redistribute it and/or modify |
9 |
|
; * it under the terms of the GNU General Public License as published by |
10 |
|
; * the Free Software Foundation ; either version 2 of the License, or |
11 |
|
; * (at your option) any later version. |
12 |
|
; * |
13 |
|
; * This program is distributed in the hope that it will be useful, |
14 |
|
; * but WITHOUT ANY WARRANTY ; without even the implied warranty of |
15 |
|
; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
16 |
|
; * GNU General Public License for more details. |
17 |
|
; * |
18 |
|
; * You should have received a copy of the GNU General Public License |
19 |
|
; * along with this program ; if not, write to the Free Software |
20 |
|
; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA |
21 |
|
; * |
22 |
|
; * $Id$ |
23 |
|
; * |
24 |
|
; *************************************************************************/ |
25 |
|
; |
26 |
|
; these 3dne functions are compatible with iSSE, but are optimized specifically for |
27 |
|
; K7 pipelines |
28 |
|
|
29 |
|
; enable dequant saturate [-2048,2047], test purposes only. |
30 |
|
%define SATURATE |
31 |
|
|
32 |
|
BITS 32 |
33 |
|
|
34 |
|
%macro cglobal 1 |
35 |
|
%ifdef PREFIX |
36 |
|
global _%1 |
37 |
|
%define %1 _%1 |
38 |
|
%else |
39 |
|
global %1 |
40 |
|
%endif |
41 |
|
%endmacro |
42 |
|
|
43 |
|
;============================================================================= |
44 |
|
; Local data |
45 |
|
;============================================================================= |
46 |
|
|
47 |
|
SECTION .rodata |
48 |
|
|
49 |
|
align 4 |
50 |
|
int_div: |
51 |
|
dd 0 |
52 |
|
%assign i 1 |
53 |
|
%rep 255 |
54 |
|
dd (1 << 16) / (i) + 1 |
55 |
|
%assign i i+1 |
56 |
|
%endrep |
57 |
|
|
58 |
|
ALIGN 16 |
59 |
|
plus_one: |
60 |
|
times 8 dw 1 |
61 |
|
|
62 |
|
;----------------------------------------------------------------------------- |
63 |
|
; subtract by Q/2 table |
64 |
|
;----------------------------------------------------------------------------- |
65 |
|
|
66 |
|
ALIGN 16 |
67 |
|
mmx_sub: |
68 |
|
%assign i 1 |
69 |
|
%rep 31 |
70 |
|
times 4 dw i / 2 |
71 |
|
%assign i i+1 |
72 |
|
%endrep |
73 |
|
|
74 |
|
|
75 |
|
;----------------------------------------------------------------------------- |
76 |
|
; |
77 |
|
; divide by 2Q table |
78 |
|
; |
79 |
|
; use a shift of 16 to take full advantage of _pmulhw_ |
80 |
|
; for q=1, _pmulhw_ will overflow so it is treated seperately |
81 |
|
; (3dnow2 provides _pmulhuw_ which wont cause overflow) |
82 |
|
; |
83 |
|
;----------------------------------------------------------------------------- |
84 |
|
|
85 |
|
ALIGN 16 |
86 |
|
mmx_div: |
87 |
|
%assign i 1 |
88 |
|
%rep 31 |
89 |
|
times 4 dw (1 << 16) / (i * 2) + 1 |
90 |
|
%assign i i+1 |
91 |
|
%endrep |
92 |
|
|
93 |
|
;----------------------------------------------------------------------------- |
94 |
|
; add by (odd(Q) ? Q : Q - 1) table |
95 |
|
;----------------------------------------------------------------------------- |
96 |
|
|
97 |
|
ALIGN 16 |
98 |
|
mmx_add: |
99 |
|
%assign i 1 |
100 |
|
%rep 31 |
101 |
|
%if i % 2 != 0 |
102 |
|
times 4 dw i |
103 |
|
%else |
104 |
|
times 4 dw i - 1 |
105 |
|
%endif |
106 |
|
%assign i i+1 |
107 |
|
%endrep |
108 |
|
|
109 |
|
;----------------------------------------------------------------------------- |
110 |
|
; multiple by 2Q table |
111 |
|
;----------------------------------------------------------------------------- |
112 |
|
|
113 |
|
ALIGN 16 |
114 |
|
mmx_mul: |
115 |
|
%assign i 1 |
116 |
|
%rep 31 |
117 |
|
times 4 dw i * 2 |
118 |
|
%assign i i+1 |
119 |
|
%endrep |
120 |
|
|
121 |
|
;----------------------------------------------------------------------------- |
122 |
|
; saturation limits |
123 |
|
;----------------------------------------------------------------------------- |
124 |
|
|
125 |
|
ALIGN 8 |
126 |
|
mmx_32768_minus_2048: |
127 |
|
times 4 dw (32768-2048) |
128 |
|
mmx_32767_minus_2047: |
129 |
|
times 4 dw (32767-2047) |
130 |
|
|
131 |
|
ALIGN 16 |
132 |
|
mmx_2047: |
133 |
|
times 4 dw 2047 |
134 |
|
|
135 |
|
ALIGN 8 |
136 |
|
mmzero: |
137 |
|
dd 0, 0 |
138 |
|
int2047: |
139 |
|
dd 2047 |
140 |
|
int_2048: |
141 |
|
dd -2048 |
142 |
|
|
143 |
|
;============================================================================= |
144 |
|
; Code |
145 |
|
;============================================================================= |
146 |
|
|
147 |
|
SECTION .text |
148 |
|
|
149 |
|
|
150 |
|
;----------------------------------------------------------------------------- |
151 |
|
; |
152 |
|
; uint32_t quant_h263_intra_3dne(int16_t * coeff, |
153 |
|
; const int16_t const * data, |
154 |
|
; const uint32_t quant, |
155 |
|
; const uint32_t dcscalar); |
156 |
|
; |
157 |
|
;----------------------------------------------------------------------------- |
158 |
|
;This is Athlon-optimized code (ca 70 clk per call) |
159 |
|
|
160 |
|
%macro quant_intra1 1 |
161 |
|
psubw mm1, mm0 ;A3 |
162 |
|
psubw mm3, mm2 ;B3 |
163 |
|
%if (%1) |
164 |
|
psubw mm5, mm4 ;C8 |
165 |
|
psubw mm7, mm6 ;D8 |
166 |
|
%endif |
167 |
|
|
168 |
|
ALIGN 8 |
169 |
|
db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 |
170 |
|
pmaxsw mm1, mm0 ;A4 |
171 |
|
db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 |
172 |
|
pmaxsw mm3, mm2 ;B4 |
173 |
|
|
174 |
|
|
175 |
|
psraw mm0, 15 ;A5 |
176 |
|
psraw mm2, 15 ;B5 |
177 |
|
%if (%1) |
178 |
|
movq [edx + %1 * 32 + 16-32], mm5 ;C9 |
179 |
|
movq [edx + %1 * 32 + 24-32], mm7 ;D9 |
180 |
|
%endif |
181 |
|
|
182 |
|
psrlw mm1, 1 ;A6 |
183 |
|
psrlw mm3, 1 ;B6 |
184 |
|
movq mm5, [ebx] ;C2 |
185 |
|
movq mm7, [ebx] ;D2 |
186 |
|
|
187 |
|
pxor mm1, mm0 ;A7 |
188 |
|
pxor mm3, mm2 ;B7 |
189 |
|
|
190 |
|
psubw mm5, mm4 ;C3 |
191 |
|
psubw mm7, mm6 ;D3 |
192 |
|
psubw mm1, mm0 ;A8 |
193 |
|
psubw mm3, mm2 ;B8 |
194 |
|
|
195 |
|
%if (%1 == 0) |
196 |
|
push ebp |
197 |
|
movq mm0, [ecx + %1 * 32 +32] |
198 |
|
%elif (%1 < 3) |
199 |
|
db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 |
200 |
|
%endif |
201 |
|
pmaxsw mm5, mm4 ;C4 |
202 |
|
%if (%1 < 3) |
203 |
|
db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 |
204 |
|
%else |
205 |
|
cmp esp, esp |
206 |
|
%endif |
207 |
|
pmaxsw mm7, mm6 ;D4 |
208 |
|
|
209 |
|
psraw mm4, 15 ;C5 |
210 |
|
psraw mm6, 15 ;D5 |
211 |
|
movq [byte edx + %1 * 32], mm1 ;A9 |
212 |
|
movq [edx + %1 * 32+8], mm3 ;B9 |
213 |
|
|
214 |
|
|
215 |
|
psrlw mm5, 1 ;C6 |
216 |
|
psrlw mm7, 1 ;D6 |
217 |
|
%if (%1 < 3) |
218 |
|
movq mm1, [ebx] ;A2 |
219 |
|
movq mm3, [ebx] ;B2 |
220 |
|
%endif |
221 |
|
%if (%1 == 3) |
222 |
|
imul eax, [int_div+4*edi] |
223 |
|
%endif |
224 |
|
pxor mm5, mm4 ;C7 |
225 |
|
pxor mm7, mm6 ;D7 |
226 |
|
%endm |
227 |
|
|
228 |
|
|
229 |
|
%macro quant_intra 1 |
230 |
|
; Rules for athlon: |
231 |
|
; 1) schedule latencies |
232 |
|
; 2) add/mul and load/store in 2:1 proportion |
233 |
|
; 3) avoid spliting >3byte instructions over 8byte boundaries |
234 |
|
|
235 |
|
psubw mm1, mm0 ;A3 |
236 |
|
psubw mm3, mm2 ;B3 |
237 |
|
%if (%1) |
238 |
|
psubw mm5, mm4 ;C8 |
239 |
|
psubw mm7, mm6 ;D8 |
240 |
|
%endif |
241 |
|
|
242 |
|
ALIGN 8 |
243 |
|
db 0Fh, 6Fh, 64h, 21h, (%1 * 32 +16) ;movq mm4, [ecx + %1 * 32 +16+32] ;C1 |
244 |
|
pmaxsw mm1, mm0 ;A4 |
245 |
|
db 0Fh, 6Fh, 74h, 21h, (%1 * 32 +24) ;movq mm6, [ecx + %1 * 32 +24+32] ;D1 |
246 |
|
pmaxsw mm3, mm2 ;B4 |
247 |
|
|
248 |
|
|
249 |
|
psraw mm0, 15 ;A5 |
250 |
|
psraw mm2, 15 ;B5 |
251 |
|
%if (%1) |
252 |
|
movq [edx + %1 * 32 + 16-32], mm5 ;C9 |
253 |
|
movq [edx + %1 * 32 + 24-32], mm7 ;D9 |
254 |
|
%endif |
255 |
|
|
256 |
|
pmulhw mm1, [esi] ;A6 |
257 |
|
pmulhw mm3, [esi] ;B6 |
258 |
|
movq mm5, [ebx] ;C2 |
259 |
|
movq mm7, [ebx] ;D2 |
260 |
|
|
261 |
|
nop |
262 |
|
nop |
263 |
|
pxor mm1, mm0 ;A7 |
264 |
|
pxor mm3, mm2 ;B7 |
265 |
|
|
266 |
|
psubw mm5, mm4 ;C3 |
267 |
|
psubw mm7, mm6 ;D3 |
268 |
|
psubw mm1, mm0 ;A8 |
269 |
|
psubw mm3, mm2 ;B8 |
270 |
|
|
271 |
|
|
272 |
|
%if (%1 < 3) |
273 |
|
db 0Fh, 6Fh, 44h, 21h, (%1 * 32 +32) ;movq mm0, [ecx + %1 * 32 +32] ;A1 |
274 |
|
%endif |
275 |
|
pmaxsw mm5, mm4 ;C4 |
276 |
|
%if (%1 < 3) |
277 |
|
db 0Fh, 6Fh, 54h, 21h, ( %1 * 32 +8+32) ;movq mm2, [ecx + %1 * 32 +8+32] ;B1 |
278 |
|
%else |
279 |
|
cmp esp, esp |
280 |
|
%endif |
281 |
|
pmaxsw mm7,mm6 ;D4 |
282 |
|
|
283 |
|
psraw mm4, 15 ;C5 |
284 |
|
psraw mm6, 15 ;D5 |
285 |
|
movq [byte edx + %1 * 32], mm1 ;A9 |
286 |
|
movq [edx + %1 * 32+8], mm3 ;B9 |
287 |
|
|
288 |
|
|
289 |
|
pmulhw mm5, [esi] ;C6 |
290 |
|
pmulhw mm7, [esi] ;D6 |
291 |
|
%if (%1 < 3) |
292 |
|
movq mm1, [ebx] ;A2 |
293 |
|
movq mm3, [ebx] ;B2 |
294 |
|
%endif |
295 |
|
%if (%1 == 0) |
296 |
|
push ebp |
297 |
|
%elif (%1 < 3) |
298 |
|
nop |
299 |
|
%endif |
300 |
|
nop |
301 |
|
%if (%1 == 3) |
302 |
|
imul eax, [int_div+4*edi] |
303 |
|
%endif |
304 |
|
pxor mm5, mm4 ;C7 |
305 |
|
pxor mm7, mm6 ;D7 |
306 |
|
%endmacro |
307 |
|
|
308 |
|
|
309 |
|
ALIGN 16 |
310 |
|
cglobal quant_h263_intra_3dne |
311 |
|
quant_h263_intra_3dne: |
312 |
|
|
313 |
|
mov eax, [esp + 12] ; quant |
314 |
|
mov ecx, [esp + 8] ; data |
315 |
|
mov edx, [esp + 4] ; coeff |
316 |
|
cmp al, 1 |
317 |
|
pxor mm1, mm1 |
318 |
|
pxor mm3, mm3 |
319 |
|
movq mm0, [ecx] ; mm0 = [1st] |
320 |
|
movq mm2, [ecx + 8] |
321 |
|
push esi |
322 |
|
lea esi, [mmx_div + eax*8 - 8] |
323 |
|
|
324 |
|
push ebx |
325 |
|
mov ebx, mmzero |
326 |
|
push edi |
327 |
|
jz near .q1loop |
328 |
|
|
329 |
|
quant_intra 0 |
330 |
|
mov ebp, [esp + 16 + 16] ; dcscalar |
331 |
|
; NB -- there are 3 pushes in the function preambule and one more |
332 |
|
; in "quant_intra 0", thus an added offset of 16 bytes |
333 |
|
movsx eax, word [byte ecx] ; DC |
334 |
|
|
335 |
|
quant_intra 1 |
336 |
|
mov edi, eax |
337 |
|
sar edi, 31 ; sign(DC) |
338 |
|
shr ebp, byte 1 ; ebp = dcscalar/2 |
339 |
|
|
340 |
|
quant_intra 2 |
341 |
|
sub eax, edi ; DC (+1) |
342 |
|
xor ebp, edi ; sign(DC) dcscalar /2 (-1) |
343 |
|
mov edi, [esp + 16 + 16] ; dscalar |
344 |
|
lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar/2 |
345 |
|
mov ebp, [byte esp] |
346 |
|
|
347 |
|
quant_intra 3 |
348 |
|
psubw mm5, mm4 ;C8 |
349 |
|
mov esi, [esp + 12] ; pop back the register value |
350 |
|
mov edi, [esp + 4] ; pop back the register value |
351 |
|
sar eax, 16 |
352 |
|
lea ebx, [byte eax + 1] ; workaround for eax < 0 |
353 |
|
cmovs eax, ebx ; conditionnaly move the corrected value |
354 |
|
mov [edx], ax ; coeff[0] = ax |
355 |
|
mov ebx, [esp + 8] ; pop back the register value |
356 |
|
add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 |
357 |
|
psubw mm7, mm6 ;D8 |
358 |
|
movq [edx + 3 * 32 + 16], mm5 ;C9 |
359 |
|
movq [edx + 3 * 32 + 24], mm7 ;D9 |
360 |
|
|
361 |
|
xor eax, eax |
362 |
|
ret |
363 |
|
|
364 |
|
ALIGN 16 |
365 |
|
|
366 |
|
.q1loop |
367 |
|
quant_intra1 0 |
368 |
|
mov ebp, [esp + 16 + 16] ; dcscalar |
369 |
|
movsx eax, word [byte ecx] ; DC |
370 |
|
|
371 |
|
quant_intra1 1 |
372 |
|
mov edi, eax |
373 |
|
sar edi, 31 ; sign(DC) |
374 |
|
shr ebp, byte 1 ; ebp = dcscalar /2 |
375 |
|
|
376 |
|
quant_intra1 2 |
377 |
|
sub eax, edi ; DC (+1) |
378 |
|
xor ebp, edi ; sign(DC) dcscalar /2 (-1) |
379 |
|
mov edi, [esp + 16 + 16] ; dcscalar |
380 |
|
lea eax, [byte eax + ebp] ; DC + sign(DC) dcscalar /2 |
381 |
|
mov ebp, [byte esp] |
382 |
|
|
383 |
|
quant_intra1 3 |
384 |
|
psubw mm5, mm4 ;C8 |
385 |
|
mov esi, [dword esp + 12] ; pop back the register value |
386 |
|
mov edi, [esp + 4] ; pop back the register value |
387 |
|
sar eax, 16 |
388 |
|
lea ebx, [byte eax + 1] ; workaround for eax < 0 |
389 |
|
cmovs eax, ebx ; conditionnaly move the corrected value |
390 |
|
mov [edx], ax ; coeff[0] = ax |
391 |
|
mov ebx, [esp + 8] ; pop back the register value |
392 |
|
add esp, byte 16 ; "quant_intra 0" pushed ebp, but we don't restore that one, just correct the stack offset by 16 |
393 |
|
psubw mm7, mm6 ;D8 |
394 |
|
movq [edx + 3 * 32 + 16], mm5 ;C9 |
395 |
|
movq [edx + 3 * 32 + 24], mm7 ;D9 |
396 |
|
|
397 |
|
xor eax, eax |
398 |
|
ret |
399 |
|
|
400 |
|
|
401 |
|
|
402 |
|
|
403 |
|
;----------------------------------------------------------------------------- |
404 |
|
; |
405 |
|
; uint32_t quant_h263_inter_3dne(int16_t * coeff, |
406 |
|
; const int16_t const * data, |
407 |
|
; const uint32_t quant); |
408 |
|
; |
409 |
|
;----------------------------------------------------------------------------- |
410 |
|
;This is Athlon-optimized code (ca 90 clk per call) |
411 |
|
;Optimized by Jaan, 30 Nov 2002 |
412 |
|
|
413 |
|
|
414 |
|
%macro quantinter 1 |
415 |
|
movq mm1, [eax] ;A2 |
416 |
|
psraw mm3, 15 ;B6 |
417 |
|
%if (%1) |
418 |
|
psubw mm2, mm6 ;C10 |
419 |
|
%endif |
420 |
|
psubw mm1, mm0 ;A3 |
421 |
|
pmulhw mm4, mm7 ;B7 |
422 |
|
movq mm6, [ecx + %1*24+16] ;C1 |
423 |
|
pmaxsw mm1, mm0 ;A4 |
424 |
|
paddw mm5, mm4 ;B8 |
425 |
|
%if (%1) |
426 |
|
movq [edx + %1*24+16-24], mm2 ;C11 |
427 |
|
%endif |
428 |
|
psubusw mm1, [ebx] ;A5 mm0 -= sub (unsigned, dont go < 0) |
429 |
|
pxor mm4, mm3 ;B9 |
430 |
|
movq mm2, [eax] ;C2 |
431 |
|
psraw mm0, 15 ;A6 |
432 |
|
psubw mm4, mm3 ;B10 |
433 |
|
psubw mm2, mm6 ;C3 |
434 |
|
pmulhw mm1, mm7 ;A7 mm0 = (mm0 / 2Q) >> 24 |
435 |
|
movq mm3, [ecx + %1*24+8] ;B1 |
436 |
|
pmaxsw mm2, mm6 ;C4 |
437 |
|
paddw mm5, mm1 ;A8 sum += mm0 |
438 |
|
%if (%1) |
439 |
|
movq [edx + %1*24+8-24], mm4 ;B11 |
440 |
|
%else |
441 |
|
movq [edx + 120], mm4 ;B11 |
442 |
|
%endif |
443 |
|
psubusw mm2, [ebx] ;C5 |
444 |
|
pxor mm1, mm0 ;A9 mm0 *= sign(mm0) |
445 |
|
movq mm4, [eax] ;B2 |
446 |
|
psraw mm6, 15 ;C6 |
447 |
|
psubw mm1, mm0 ;A10 undisplace |
448 |
|
psubw mm4, mm3 ;B3 |
449 |
|
pmulhw mm2, mm7 ;C7 |
450 |
|
movq mm0, [ecx + %1*24+24] ;A1 mm0 = [1st] |
451 |
|
pmaxsw mm4, mm3 ;B4 |
452 |
|
paddw mm5, mm2 ;C8 |
453 |
|
movq [byte edx + %1*24], mm1 ;A11 |
454 |
|
psubusw mm4, [ebx] ;B5 |
455 |
|
pxor mm2, mm6 ;C9 |
456 |
|
%endmacro |
457 |
|
|
458 |
|
%macro quantinter1 1 |
459 |
|
movq mm0, [byte ecx + %1*16] ;mm0 = [1st] |
460 |
|
movq mm3, [ecx + %1*16+8] ; |
461 |
|
movq mm1, [eax] |
462 |
|
movq mm4, [eax] |
463 |
|
psubw mm1, mm0 |
464 |
|
psubw mm4, mm3 |
465 |
|
pmaxsw mm1, mm0 |
466 |
|
pmaxsw mm4, mm3 |
467 |
|
psubusw mm1, mm6 ; mm0 -= sub (unsigned, dont go < 0) |
468 |
|
psubusw mm4, mm6 ; |
469 |
|
psraw mm0, 15 |
470 |
|
psraw mm3, 15 |
471 |
|
psrlw mm1, 1 ; mm0 = (mm0 / 2Q) >> 16 |
472 |
|
psrlw mm4, 1 ; |
473 |
|
paddw mm5, mm1 ; sum += mm0 |
474 |
|
pxor mm1, mm0 ; mm0 *= sign(mm0) |
475 |
|
paddw mm5, mm4 |
476 |
|
pxor mm4, mm3 ; |
477 |
|
psubw mm1, mm0 ; undisplace |
478 |
|
psubw mm4, mm3 |
479 |
|
cmp esp, esp |
480 |
|
movq [byte edx + %1*16], mm1 |
481 |
|
movq [edx + %1*16+8], mm4 |
482 |
|
%endmacro |
483 |
|
|
484 |
|
ALIGN 16 |
485 |
|
cglobal quant_h263_inter_3dne |
486 |
|
quant_h263_inter_3dne: |
487 |
|
mov edx, [esp + 4] ; coeff |
488 |
|
mov ecx, [esp + 8] ; data |
489 |
|
mov eax, [esp + 12] ; quant |
490 |
|
push ebx |
491 |
|
|
492 |
|
pxor mm5, mm5 ; sum |
493 |
|
nop |
494 |
|
lea ebx,[mmx_sub + eax * 8 - 8] ; sub |
495 |
|
movq mm7, [mmx_div + eax * 8 - 8] ; divider |
496 |
|
|
497 |
|
cmp al, 1 |
498 |
|
lea eax, [mmzero] |
499 |
|
jz near .q1loop |
500 |
|
cmp esp, esp |
501 |
|
ALIGN 8 |
502 |
|
movq mm3, [ecx + 120] ;B1 |
503 |
|
pxor mm4, mm4 ;B2 |
504 |
|
psubw mm4, mm3 ;B3 |
505 |
|
movq mm0, [ecx] ;A1 mm0 = [1st] |
506 |
|
pmaxsw mm4, mm3 ;B4 |
507 |
|
psubusw mm4, [ebx] ;B5 |
508 |
|
|
509 |
|
quantinter 0 |
510 |
|
quantinter 1 |
511 |
|
quantinter 2 |
512 |
|
quantinter 3 |
513 |
|
quantinter 4 |
514 |
|
|
515 |
|
psraw mm3, 15 ;B6 |
516 |
|
psubw mm2, mm6 ;C10 |
517 |
|
pmulhw mm4, mm7 ;B7 |
518 |
|
paddw mm5, mm4 ;B8 |
519 |
|
pxor mm4, mm3 ;B9 |
520 |
|
psubw mm4, mm3 ;B10 |
521 |
|
movq [edx + 4*24+16], mm2 ;C11 |
522 |
|
pop ebx |
523 |
|
movq [edx + 4*24+8], mm4 ;B11 |
524 |
|
pmaddwd mm5, [plus_one] |
525 |
|
movq mm0, mm5 |
526 |
|
punpckhdq mm5, mm5 |
527 |
|
paddd mm0, mm5 |
528 |
|
movd eax, mm0 ; return sum |
529 |
|
|
530 |
|
ret |
531 |
|
|
532 |
|
ALIGN 16 |
533 |
|
.q1loop |
534 |
|
movq mm6, [byte ebx] |
535 |
|
|
536 |
|
quantinter1 0 |
537 |
|
quantinter1 1 |
538 |
|
quantinter1 2 |
539 |
|
quantinter1 3 |
540 |
|
quantinter1 4 |
541 |
|
quantinter1 5 |
542 |
|
quantinter1 6 |
543 |
|
quantinter1 7 |
544 |
|
|
545 |
|
pmaddwd mm5, [plus_one] |
546 |
|
movq mm0, mm5 |
547 |
|
psrlq mm5, 32 |
548 |
|
paddd mm0, mm5 |
549 |
|
movd eax, mm0 ; return sum |
550 |
|
|
551 |
|
pop ebx |
552 |
|
|
553 |
|
ret |
554 |
|
|
555 |
|
;----------------------------------------------------------------------------- |
556 |
|
; |
557 |
|
; uint32_t dequant_h263_intra_3dne(int16_t *data, |
558 |
|
; const int16_t const *coeff, |
559 |
|
; const uint32_t quant, |
560 |
|
; const uint32_t dcscalar); |
561 |
|
; |
562 |
|
;----------------------------------------------------------------------------- |
563 |
|
|
564 |
|
; this is the same as dequant_inter_3dne, except that we're |
565 |
|
; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) |
566 |
|
|
567 |
|
;This is Athlon-optimized code (ca 106 clk per call) |
568 |
|
|
569 |
|
%macro dequant 1 |
570 |
|
movq mm1, [ecx+%1*24] ; c = coeff[i] ;A2 |
571 |
|
psubw mm0, mm1 ;-c ;A3 (1st dep) |
572 |
|
%if (%1) |
573 |
|
paddw mm4, mm6 ;C11 mm6 free (4th+) |
574 |
|
%endif |
575 |
|
pmaxsw mm0, mm1 ;|c| ;A4 (2nd) |
576 |
|
%if (%1) |
577 |
|
mov ebp, ebp |
578 |
|
pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) later |
579 |
|
%endif |
580 |
|
movq mm6, [esi] ;0 ;A5 mm6 in use |
581 |
|
pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) |
582 |
|
%if (%1) |
583 |
|
pxor mm5, mm4 ;C13 (6th+) 1later |
584 |
|
%endif |
585 |
|
movq mm4, [esi] ;C1 ;0 |
586 |
|
mov esp, esp |
587 |
|
pcmpeqw mm6, [ecx+%1*24] ;A6 (c ==0) ? -1 : 0 (1st) |
588 |
|
ALIGN 4 |
589 |
|
psraw mm1, 15 ; sign(c) ;A7 (2nd) |
590 |
|
%if (%1) |
591 |
|
movq [edx+%1*24+16-24], mm5 ; C14 (7th) 2later |
592 |
|
%endif |
593 |
|
paddw mm7, mm3 ;B10 offset +negate back (3rd) |
594 |
|
pmullw mm0, [edi] ;*= 2Q ;A8 (3rd+) |
595 |
|
paddw mm2, mm7 ;B11 mm7 free (4th+) |
596 |
|
lea ebp, [byte ebp] |
597 |
|
movq mm5, [ecx+%1*24+16] ;C2 ; c = coeff[i] |
598 |
|
psubw mm4, mm5 ;-c ;C3 (1st dep) |
599 |
|
pandn mm6, [eax] ;A9 offset = isZero ? 0 : quant_add (2nd) |
600 |
|
pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) |
601 |
|
pxor mm3, mm2 ;B13 (6th+) |
602 |
|
movq mm2, [byte esi] ;B1 ;0 |
603 |
|
%if (%1) |
604 |
|
movq [edx+%1*24+8-24], mm3 ;B14 (7th) |
605 |
|
%else |
606 |
|
movq [edx+120], mm3 |
607 |
|
%endif |
608 |
|
pmaxsw mm4, mm5 ;|c| ;C4 (2nd) |
609 |
|
paddw mm6, mm1 ;A10 offset +negate back (3rd) |
610 |
|
movq mm3, [ecx+%1*24 + 8] ;B2 ; c = coeff[i] |
611 |
|
psubw mm2, mm3 ;-c ;B3 (1st dep) |
612 |
|
paddw mm0, mm6 ;A11 mm6 free (4th+) |
613 |
|
movq mm6, [byte esi] ;0 ;C5 mm6 in use |
614 |
|
pcmpeqw mm6, [ecx+%1*24+16] ;C6 (c ==0) ? -1 : 0 (1st) |
615 |
|
pminsw mm0, [ebx] ;A12 saturates to +2047 (5th+) |
616 |
|
pmaxsw mm2, mm3 ;|c| ;B4 (2nd) |
617 |
|
pxor mm1, mm0 ;A13 (6th+) |
618 |
|
pmullw mm4, [edi] ;*= 2Q ;C8 (3rd+) |
619 |
|
psraw mm5, 15 ; sign(c) ;C7 (2nd) |
620 |
|
movq mm7, [byte esi] ;0 ;B5 mm7 in use |
621 |
|
pcmpeqw mm7, [ecx+%1*24 + 8] ;B6 (c ==0) ? -1 : 0 (1st) |
622 |
|
%if (%1 < 4) |
623 |
|
movq mm0, [byte esi] ;A1 ;0 |
624 |
|
%endif |
625 |
|
pandn mm6, [byte eax] ;C9 offset = isZero ? 0 : quant_add (2nd) |
626 |
|
psraw mm3, 15 ;sign(c) ;B7 (2nd) |
627 |
|
movq [byte edx+%1*24], mm1 ;A14 (7th) |
628 |
|
paddw mm6, mm5 ;C10 offset +negate back (3rd) |
629 |
|
pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) |
630 |
|
mov esp, esp |
631 |
|
%endmacro |
632 |
|
|
633 |
|
|
634 |
|
ALIGN 16 |
635 |
|
cglobal dequant_h263_intra_3dne |
636 |
|
dequant_h263_intra_3dne: |
637 |
|
mov ecx, [esp+ 8] ; coeff |
638 |
|
mov eax, [esp+12] ; quant |
639 |
|
pxor mm0, mm0 |
640 |
|
pxor mm2, mm2 |
641 |
|
push edi |
642 |
|
push ebx |
643 |
|
lea edi, [mmx_mul + eax*8 - 8] ; 2*quant |
644 |
|
push ebp |
645 |
|
mov ebx, mmx_2047 |
646 |
|
movsx ebp, word [ecx] |
647 |
|
lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 |
648 |
|
push esi |
649 |
|
mov esi, mmzero |
650 |
|
pxor mm7, mm7 |
651 |
|
movq mm3, [ecx+120] ;B2 ; c = coeff[i] |
652 |
|
pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) |
653 |
|
|
654 |
|
imul ebp, [esp+16+16] ; dcscalar |
655 |
|
psubw mm2, mm3 ;-c ;B3 (1st dep) |
656 |
|
pmaxsw mm2, mm3 ;|c| ;B4 (2nd) |
657 |
|
pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) |
658 |
|
psraw mm3, 15 ; sign(c) ;B7 (2nd) |
659 |
|
mov edx, [esp+ 4+16] ; data |
660 |
|
|
661 |
|
ALIGN 8 |
662 |
|
dequant 0 |
663 |
|
|
664 |
|
cmp ebp, -2048 |
665 |
|
mov esp, esp |
666 |
|
|
667 |
|
dequant 1 |
668 |
|
|
669 |
|
cmovl ebp, [int_2048] |
670 |
|
nop |
671 |
|
|
672 |
|
dequant 2 |
673 |
|
|
674 |
|
cmp ebp, 2047 |
675 |
|
mov esp, esp |
676 |
|
|
677 |
|
dequant 3 |
678 |
|
|
679 |
|
cmovg ebp, [int2047] |
680 |
|
nop |
681 |
|
|
682 |
|
dequant 4 |
683 |
|
|
684 |
|
paddw mm4, mm6 ;C11 mm6 free (4th+) |
685 |
|
pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) |
686 |
|
pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) |
687 |
|
mov eax, ebp |
688 |
|
mov esi, [esp] |
689 |
|
mov ebp, [esp+4] |
690 |
|
pxor mm5, mm4 ;C13 (6th+) |
691 |
|
paddw mm7, mm3 ;B10 offset +negate back (3rd) |
692 |
|
movq [edx+4*24+16], mm5 ;C14 (7th) |
693 |
|
paddw mm2, mm7 ;B11 mm7 free (4th+) |
694 |
|
pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) |
695 |
|
mov ebx, [esp+8] |
696 |
|
mov edi, [esp+12] |
697 |
|
add esp, byte 16 |
698 |
|
pxor mm3, mm2 ;B13 (6th+) |
699 |
|
movq [edx+4*24+8], mm3 ;B14 (7th) |
700 |
|
mov [edx], ax |
701 |
|
|
702 |
|
xor eax, eax |
703 |
|
ret |
704 |
|
|
705 |
|
;----------------------------------------------------------------------------- |
706 |
|
; |
707 |
|
; uint32_t dequant_h263_inter_3dne(int16_t * data, |
708 |
|
; const int16_t * const coeff, |
709 |
|
; const uint32_t quant); |
710 |
|
; |
711 |
|
;----------------------------------------------------------------------------- |
712 |
|
|
713 |
|
; this is the same as dequant_inter_3dne, |
714 |
|
; except that we're saturating using 'pminsw' (saves 2 cycles/loop) |
715 |
|
; This is Athlon-optimized code (ca 100 clk per call) |
716 |
|
|
717 |
|
ALIGN 16 |
718 |
|
cglobal dequant_h263_inter_3dne |
719 |
|
dequant_h263_inter_3dne: |
720 |
|
mov ecx, [esp+ 8] ; coeff |
721 |
|
mov eax, [esp+12] ; quant |
722 |
|
pxor mm0, mm0 |
723 |
|
pxor mm2, mm2 |
724 |
|
push edi |
725 |
|
push ebx |
726 |
|
push esi |
727 |
|
lea edi, [mmx_mul + eax*8 - 8] ; 2*quant |
728 |
|
mov ebx, mmx_2047 |
729 |
|
pxor mm7, mm7 |
730 |
|
movq mm3, [ecx+120] ;B2 ; c = coeff[i] |
731 |
|
pcmpeqw mm7, [ecx+120] ;B6 (c ==0) ? -1 : 0 (1st) |
732 |
|
lea eax, [mmx_add + eax*8 - 8] ; quant or quant-1 |
733 |
|
psubw mm2, mm3 ;-c ;B3 (1st dep) |
734 |
|
mov esi, mmzero |
735 |
|
pmaxsw mm2, mm3 ;|c| ;B4 (2nd) |
736 |
|
pmullw mm2, [edi] ;*= 2Q ;B8 (3rd+) |
737 |
|
psraw mm3, 15 ; sign(c) ;B7 (2nd) |
738 |
|
mov edx, [dword esp+ 4+12] ; data |
739 |
|
|
740 |
|
ALIGN 8 |
741 |
|
|
742 |
|
dequant 0 |
743 |
|
dequant 1 |
744 |
|
dequant 2 |
745 |
|
dequant 3 |
746 |
|
dequant 4 |
747 |
|
|
748 |
|
paddw mm4, mm6 ;C11 mm6 free (4th+) |
749 |
|
pminsw mm4, [ebx] ;C12 saturates to +2047 (5th+) |
750 |
|
pandn mm7, [eax] ;B9 offset = isZero ? 0 : quant_add (2nd) |
751 |
|
mov esi, [esp] |
752 |
|
pxor mm5, mm4 ;C13 (6th+) |
753 |
|
paddw mm7, mm3 ;B10 offset +negate back (3rd) |
754 |
|
movq [edx+4*24+16], mm5 ;C14 (7th) |
755 |
|
paddw mm2, mm7 ;B11 mm7 free (4th+) |
756 |
|
pminsw mm2, [ebx] ;B12 saturates to +2047 (5th+) |
757 |
|
mov ebx, [esp+4] |
758 |
|
mov edi, [esp+8] |
759 |
|
add esp, byte 12 |
760 |
|
pxor mm3, mm2 ;B13 (6th+) |
761 |
|
movq [edx+4*24+8], mm3 ;B14 (7th) |
762 |
|
|
763 |
|
xor eax, eax |
764 |
|
ret |