Parent Directory | Revision Log
Revision 1.12 - (view) (download)
1 : | edgomez | 1.2 | ;/************************************************************************** |
2 : | ; * | ||
3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
4 : | ; * - 3dne Quantization/Dequantization - | ||
5 : | ; * | ||
6 : | ; * Copyright(C) 2002-2003 Jaan Kalda | ||
7 : | ; * | ||
8 : | Isibaar | 1.9 | ; * This program is free software ; you can r_EDIstribute it and/or modify |
9 : | edgomez | 1.2 | ; * 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 : | Isibaar | 1.12 | ; * $Id: quantize_h263_3dne.asm,v 1.11 2008/12/04 18:30:36 Isibaar Exp $ |
23 : | edgomez | 1.2 | ; * |
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 : | Isibaar | 1.9 | %include "nasm.inc" |
33 : | edgomez | 1.2 | |
34 : | ;============================================================================= | ||
35 : | ; Local data | ||
36 : | ;============================================================================= | ||
37 : | |||
38 : | Isibaar | 1.9 | DATA |
39 : | edgomez | 1.2 | |
40 : | Isibaar | 1.9 | align SECTION_ALIGN |
41 : | edgomez | 1.2 | int_div: |
42 : | dd 0 | ||
43 : | %assign i 1 | ||
44 : | %rep 255 | ||
45 : | dd (1 << 16) / (i) + 1 | ||
46 : | %assign i i+1 | ||
47 : | %endrep | ||
48 : | |||
49 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
50 : | edgomez | 1.2 | plus_one: |
51 : | times 8 dw 1 | ||
52 : | |||
53 : | ;----------------------------------------------------------------------------- | ||
54 : | ; subtract by Q/2 table | ||
55 : | ;----------------------------------------------------------------------------- | ||
56 : | |||
57 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
58 : | edgomez | 1.2 | mmx_sub: |
59 : | %assign i 1 | ||
60 : | %rep 31 | ||
61 : | times 4 dw i / 2 | ||
62 : | %assign i i+1 | ||
63 : | %endrep | ||
64 : | |||
65 : | |||
66 : | ;----------------------------------------------------------------------------- | ||
67 : | ; | ||
68 : | ; divide by 2Q table | ||
69 : | ; | ||
70 : | ; use a shift of 16 to take full advantage of _pmulhw_ | ||
71 : | ; for q=1, _pmulhw_ will overflow so it is treated seperately | ||
72 : | ; (3dnow2 provides _pmulhuw_ which wont cause overflow) | ||
73 : | ; | ||
74 : | ;----------------------------------------------------------------------------- | ||
75 : | |||
76 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
77 : | edgomez | 1.2 | mmx_div: |
78 : | %assign i 1 | ||
79 : | %rep 31 | ||
80 : | times 4 dw (1 << 16) / (i * 2) + 1 | ||
81 : | %assign i i+1 | ||
82 : | %endrep | ||
83 : | |||
84 : | ;----------------------------------------------------------------------------- | ||
85 : | ; add by (odd(Q) ? Q : Q - 1) table | ||
86 : | ;----------------------------------------------------------------------------- | ||
87 : | |||
88 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
89 : | edgomez | 1.2 | mmx_add: |
90 : | %assign i 1 | ||
91 : | %rep 31 | ||
92 : | %if i % 2 != 0 | ||
93 : | times 4 dw i | ||
94 : | %else | ||
95 : | times 4 dw i - 1 | ||
96 : | %endif | ||
97 : | %assign i i+1 | ||
98 : | %endrep | ||
99 : | |||
100 : | ;----------------------------------------------------------------------------- | ||
101 : | ; multiple by 2Q table | ||
102 : | ;----------------------------------------------------------------------------- | ||
103 : | |||
104 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
105 : | edgomez | 1.2 | mmx_mul: |
106 : | %assign i 1 | ||
107 : | %rep 31 | ||
108 : | times 4 dw i * 2 | ||
109 : | %assign i i+1 | ||
110 : | %endrep | ||
111 : | |||
112 : | ;----------------------------------------------------------------------------- | ||
113 : | ; saturation limits | ||
114 : | ;----------------------------------------------------------------------------- | ||
115 : | |||
116 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
117 : | edgomez | 1.2 | mmx_32768_minus_2048: |
118 : | times 4 dw (32768-2048) | ||
119 : | mmx_32767_minus_2047: | ||
120 : | times 4 dw (32767-2047) | ||
121 : | |||
122 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
123 : | edgomez | 1.2 | mmx_2047: |
124 : | times 4 dw 2047 | ||
125 : | |||
126 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
127 : | edgomez | 1.2 | mmzero: |
128 : | dd 0, 0 | ||
129 : | int2047: | ||
130 : | dd 2047 | ||
131 : | int_2048: | ||
132 : | dd -2048 | ||
133 : | |||
134 : | ;============================================================================= | ||
135 : | ; Code | ||
136 : | ;============================================================================= | ||
137 : | |||
138 : | Isibaar | 1.10 | TEXT |
139 : | edgomez | 1.2 | |
140 : | ;----------------------------------------------------------------------------- | ||
141 : | ; | ||
142 : | ; uint32_t quant_h263_intra_3dne(int16_t * coeff, | ||
143 : | ; const int16_t const * data, | ||
144 : | ; const uint32_t quant, | ||
145 : | ; const uint32_t dcscalar, | ||
146 : | ; const uint16_t *mpeg_matrices); | ||
147 : | ; | ||
148 : | ;----------------------------------------------------------------------------- | ||
149 : | ;This is Athlon-optimized code (ca 70 clk per call) | ||
150 : | |||
151 : | %macro quant_intra1 1 | ||
152 : | psubw mm1, mm0 ;A3 | ||
153 : | psubw mm3, mm2 ;B3 | ||
154 : | %if (%1) | ||
155 : | psubw mm5, mm4 ;C8 | ||
156 : | psubw mm7, mm6 ;D8 | ||
157 : | %endif | ||
158 : | |||
159 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
160 : | movq mm4, [_ECX + %1 * 32 +16] ;C1 | ||
161 : | edgomez | 1.2 | pmaxsw mm1, mm0 ;A4 |
162 : | Isibaar | 1.9 | movq mm6, [_ECX + %1 * 32 +24] ;D1 |
163 : | edgomez | 1.2 | pmaxsw mm3, mm2 ;B4 |
164 : | |||
165 : | |||
166 : | psraw mm0, 15 ;A5 | ||
167 : | psraw mm2, 15 ;B5 | ||
168 : | %if (%1) | ||
169 : | Isibaar | 1.9 | movq [_EDX + %1 * 32 + 16-32], mm5 ;C9 |
170 : | movq [_EDX + %1 * 32 + 24-32], mm7 ;D9 | ||
171 : | edgomez | 1.2 | %endif |
172 : | |||
173 : | psrlw mm1, 1 ;A6 | ||
174 : | psrlw mm3, 1 ;B6 | ||
175 : | Isibaar | 1.9 | movq mm5, [_EBX] ;C2 |
176 : | movq mm7, [_EBX] ;D2 | ||
177 : | edgomez | 1.2 | |
178 : | pxor mm1, mm0 ;A7 | ||
179 : | pxor mm3, mm2 ;B7 | ||
180 : | |||
181 : | psubw mm5, mm4 ;C3 | ||
182 : | psubw mm7, mm6 ;D3 | ||
183 : | psubw mm1, mm0 ;A8 | ||
184 : | psubw mm3, mm2 ;B8 | ||
185 : | |||
186 : | %if (%1 == 0) | ||
187 : | Isibaar | 1.9 | push _EBP |
188 : | movq mm0, [_ECX + %1 * 32 +32] | ||
189 : | edgomez | 1.2 | %elif (%1 < 3) |
190 : | Isibaar | 1.9 | movq mm0, [_ECX + %1 * 32 +32] ;A1 |
191 : | edgomez | 1.2 | %endif |
192 : | pmaxsw mm5, mm4 ;C4 | ||
193 : | %if (%1 < 3) | ||
194 : | Isibaar | 1.9 | movq mm2, [_ECX + %1 * 32 +8+32] ;B1 |
195 : | edgomez | 1.2 | %else |
196 : | Isibaar | 1.9 | cmp _ESP, _ESP |
197 : | edgomez | 1.2 | %endif |
198 : | pmaxsw mm7, mm6 ;D4 | ||
199 : | |||
200 : | psraw mm4, 15 ;C5 | ||
201 : | psraw mm6, 15 ;D5 | ||
202 : | Isibaar | 1.9 | movq [byte _EDX + %1 * 32], mm1 ;A9 |
203 : | movq [_EDX + %1 * 32+8], mm3 ;B9 | ||
204 : | edgomez | 1.2 | |
205 : | |||
206 : | psrlw mm5, 1 ;C6 | ||
207 : | psrlw mm7, 1 ;D6 | ||
208 : | %if (%1 < 3) | ||
209 : | Isibaar | 1.9 | movq mm1, [_EBX] ;A2 |
210 : | movq mm3, [_EBX] ;B2 | ||
211 : | edgomez | 1.2 | %endif |
212 : | %if (%1 == 3) | ||
213 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
214 : | lea r9, [int_div] | ||
215 : | imul eax, dword [r9+4*_EDI] | ||
216 : | %else | ||
217 : | imul _EAX, [int_div+4*_EDI] | ||
218 : | %endif | ||
219 : | edgomez | 1.2 | %endif |
220 : | pxor mm5, mm4 ;C7 | ||
221 : | pxor mm7, mm6 ;D7 | ||
222 : | %endm | ||
223 : | |||
224 : | |||
225 : | %macro quant_intra 1 | ||
226 : | ; Rules for athlon: | ||
227 : | ; 1) schedule latencies | ||
228 : | ; 2) add/mul and load/store in 2:1 proportion | ||
229 : | ; 3) avoid spliting >3byte instructions over 8byte boundaries | ||
230 : | |||
231 : | psubw mm1, mm0 ;A3 | ||
232 : | psubw mm3, mm2 ;B3 | ||
233 : | %if (%1) | ||
234 : | psubw mm5, mm4 ;C8 | ||
235 : | psubw mm7, mm6 ;D8 | ||
236 : | %endif | ||
237 : | |||
238 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
239 : | movq mm4, [_ECX + %1 * 32 +16] ;C1 | ||
240 : | edgomez | 1.2 | pmaxsw mm1, mm0 ;A4 |
241 : | Isibaar | 1.9 | movq mm6, [_ECX + %1 * 32 +24] ;D1 |
242 : | edgomez | 1.2 | pmaxsw mm3, mm2 ;B4 |
243 : | |||
244 : | |||
245 : | psraw mm0, 15 ;A5 | ||
246 : | psraw mm2, 15 ;B5 | ||
247 : | %if (%1) | ||
248 : | Isibaar | 1.9 | movq [_EDX + %1 * 32 + 16-32], mm5 ;C9 |
249 : | movq [_EDX + %1 * 32 + 24-32], mm7 ;D9 | ||
250 : | edgomez | 1.2 | %endif |
251 : | |||
252 : | Isibaar | 1.9 | pmulhw mm1, [_ESI] ;A6 |
253 : | pmulhw mm3, [_ESI] ;B6 | ||
254 : | movq mm5, [_EBX] ;C2 | ||
255 : | movq mm7, [_EBX] ;D2 | ||
256 : | edgomez | 1.2 | |
257 : | nop | ||
258 : | nop | ||
259 : | pxor mm1, mm0 ;A7 | ||
260 : | pxor mm3, mm2 ;B7 | ||
261 : | |||
262 : | psubw mm5, mm4 ;C3 | ||
263 : | psubw mm7, mm6 ;D3 | ||
264 : | psubw mm1, mm0 ;A8 | ||
265 : | psubw mm3, mm2 ;B8 | ||
266 : | |||
267 : | |||
268 : | %if (%1 < 3) | ||
269 : | Isibaar | 1.9 | movq mm0, [_ECX + %1 * 32 +32] ;A1 |
270 : | edgomez | 1.2 | %endif |
271 : | pmaxsw mm5, mm4 ;C4 | ||
272 : | %if (%1 < 3) | ||
273 : | Isibaar | 1.9 | movq mm2, [_ECX + %1 * 32 +8+32] ;B1 |
274 : | edgomez | 1.2 | %else |
275 : | Isibaar | 1.9 | cmp _ESP, _ESP |
276 : | edgomez | 1.2 | %endif |
277 : | pmaxsw mm7,mm6 ;D4 | ||
278 : | |||
279 : | psraw mm4, 15 ;C5 | ||
280 : | psraw mm6, 15 ;D5 | ||
281 : | Isibaar | 1.9 | movq [byte _EDX + %1 * 32], mm1 ;A9 |
282 : | movq [_EDX + %1 * 32+8], mm3 ;B9 | ||
283 : | edgomez | 1.2 | |
284 : | |||
285 : | Isibaar | 1.9 | pmulhw mm5, [_ESI] ;C6 |
286 : | pmulhw mm7, [_ESI] ;D6 | ||
287 : | edgomez | 1.2 | %if (%1 < 3) |
288 : | Isibaar | 1.9 | movq mm1, [_EBX] ;A2 |
289 : | movq mm3, [_EBX] ;B2 | ||
290 : | edgomez | 1.2 | %endif |
291 : | %if (%1 == 0) | ||
292 : | Isibaar | 1.9 | push _EBP |
293 : | edgomez | 1.2 | %elif (%1 < 3) |
294 : | nop | ||
295 : | %endif | ||
296 : | nop | ||
297 : | %if (%1 == 3) | ||
298 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
299 : | lea r9, [int_div] | ||
300 : | imul eax, dword [r9+4*_EDI] | ||
301 : | %else | ||
302 : | imul _EAX, [int_div+4*_EDI] | ||
303 : | %endif | ||
304 : | edgomez | 1.2 | %endif |
305 : | pxor mm5, mm4 ;C7 | ||
306 : | pxor mm7, mm6 ;D7 | ||
307 : | %endmacro | ||
308 : | |||
309 : | |||
310 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
311 : | edgomez | 1.2 | cglobal quant_h263_intra_3dne |
312 : | quant_h263_intra_3dne: | ||
313 : | |||
314 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
315 : | mov TMP0, [_ESP] | ||
316 : | add _ESP, PTR_SIZE | ||
317 : | %ifndef WINDOWS | ||
318 : | push prm6 | ||
319 : | push prm5 | ||
320 : | %endif | ||
321 : | push prm4 | ||
322 : | push prm3 | ||
323 : | push prm2 | ||
324 : | push prm1 | ||
325 : | sub _ESP, PTR_SIZE | ||
326 : | mov [_ESP], TMP0 | ||
327 : | %endif | ||
328 : | |||
329 : | mov _EAX, [_ESP + 3*PTR_SIZE] ; quant | ||
330 : | mov _ECX, [_ESP + 2*PTR_SIZE] ; data | ||
331 : | mov _EDX, [_ESP + 1*PTR_SIZE] ; coeff | ||
332 : | edgomez | 1.2 | cmp al, 1 |
333 : | pxor mm1, mm1 | ||
334 : | pxor mm3, mm3 | ||
335 : | Isibaar | 1.9 | movq mm0, [_ECX] ; mm0 = [1st] |
336 : | movq mm2, [_ECX + 8] | ||
337 : | push _ESI | ||
338 : | %ifdef ARCH_IS_X86_64 | ||
339 : | lea _ESI, [mmx_div] | ||
340 : | lea _ESI, [_ESI + _EAX*8 - 8] | ||
341 : | %else | ||
342 : | lea _ESI, [mmx_div + _EAX*8 - 8] | ||
343 : | %endif | ||
344 : | |||
345 : | push _EBX | ||
346 : | Isibaar | 1.11 | lea _EBX, [mmzero] |
347 : | Isibaar | 1.9 | push _EDI |
348 : | edgomez | 1.2 | jz near .q1loop |
349 : | |||
350 : | quant_intra 0 | ||
351 : | Isibaar | 1.9 | mov _EBP, [_ESP + (4+4)*PTR_SIZE] ; dcscalar |
352 : | ; NB -- there are 3 pushes in the function preambule and one more | ||
353 : | ; in "quant_intra 0", thus an added offset of 16 bytes | ||
354 : | Isibaar | 1.11 | movsx _EAX, word [byte _ECX] ; DC |
355 : | edgomez | 1.2 | |
356 : | quant_intra 1 | ||
357 : | Isibaar | 1.11 | mov _EDI, _EAX |
358 : | sar _EDI, 31 ; sign(DC) | ||
359 : | shr _EBP, byte 1 ; _EBP = dcscalar/2 | ||
360 : | edgomez | 1.2 | |
361 : | quant_intra 2 | ||
362 : | Isibaar | 1.9 | sub _EAX, _EDI ; DC (+1) |
363 : | xor _EBP, _EDI ; sign(DC) dcscalar /2 (-1) | ||
364 : | mov _EDI, [_ESP + (4+4)*PTR_SIZE] ; dscalar | ||
365 : | Isibaar | 1.11 | lea _EAX, [byte _EAX + _EBP] ; DC + sign(DC) dcscalar/2 |
366 : | Isibaar | 1.9 | mov _EBP, [byte _ESP] |
367 : | edgomez | 1.2 | |
368 : | quant_intra 3 | ||
369 : | psubw mm5, mm4 ;C8 | ||
370 : | Isibaar | 1.9 | mov _ESI, [_ESP + 3*PTR_SIZE] ; pop back the register value |
371 : | mov _EDI, [_ESP + 1*PTR_SIZE] ; pop back the register value | ||
372 : | sar _EAX, 16 | ||
373 : | lea _EBX, [byte _EAX + 1] ; workaround for _EAX < 0 | ||
374 : | cmovs _EAX, _EBX ; conditionnaly move the corrected value | ||
375 : | mov [_EDX], ax ; coeff[0] = ax | ||
376 : | mov _EBX, [_ESP + 2*PTR_SIZE] ; pop back the register value | ||
377 : | add _ESP, byte 4*PTR_SIZE ; "quant_intra 0" pushed _EBP, but we don't restore that one, just correct the stack offset by 16 | ||
378 : | edgomez | 1.2 | psubw mm7, mm6 ;D8 |
379 : | Isibaar | 1.9 | movq [_EDX + 3 * 32 + 16], mm5 ;C9 |
380 : | movq [_EDX + 3 * 32 + 24], mm7 ;D9 | ||
381 : | |||
382 : | xor _EAX, _EAX | ||
383 : | |||
384 : | %ifdef ARCH_IS_X86_64 | ||
385 : | mov TMP0, [_ESP] | ||
386 : | %ifndef WINDOWS | ||
387 : | add _ESP, 6*PTR_SIZE | ||
388 : | %else | ||
389 : | add _ESP, 4*PTR_SIZE | ||
390 : | %endif | ||
391 : | mov [_ESP], TMP0 | ||
392 : | %endif | ||
393 : | edgomez | 1.2 | |
394 : | ret | ||
395 : | |||
396 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
397 : | edgomez | 1.2 | |
398 : | Isibaar | 1.8 | .q1loop: |
399 : | edgomez | 1.2 | quant_intra1 0 |
400 : | Isibaar | 1.9 | mov _EBP, [_ESP + (4+4)*PTR_SIZE] ; dcscalar |
401 : | Isibaar | 1.11 | movsx _EAX, word [byte _ECX] ; DC |
402 : | edgomez | 1.2 | |
403 : | quant_intra1 1 | ||
404 : | Isibaar | 1.11 | mov _EDI, _EAX |
405 : | sar _EDI, 31 ; sign(DC) | ||
406 : | shr _EBP, byte 1 ; _EBP = dcscalar /2 | ||
407 : | edgomez | 1.2 | |
408 : | quant_intra1 2 | ||
409 : | Isibaar | 1.9 | sub _EAX, _EDI ; DC (+1) |
410 : | xor _EBP, _EDI ; sign(DC) dcscalar /2 (-1) | ||
411 : | mov _EDI, [_ESP + (4+4)*PTR_SIZE] ; dcscalar | ||
412 : | Isibaar | 1.11 | lea _EAX, [byte _EAX + _EBP] ; DC + sign(DC) dcscalar /2 |
413 : | Isibaar | 1.9 | mov _EBP, [byte _ESP] |
414 : | edgomez | 1.2 | |
415 : | quant_intra1 3 | ||
416 : | psubw mm5, mm4 ;C8 | ||
417 : | Isibaar | 1.9 | mov _ESI, [_ESP + 3*PTR_SIZE] ; pop back the register value |
418 : | mov _EDI, [_ESP + 1*PTR_SIZE] ; pop back the register value | ||
419 : | sar _EAX, 16 | ||
420 : | lea _EBX, [byte _EAX + 1] ; workaround for _EAX < 0 | ||
421 : | cmovs _EAX, _EBX ; conditionnaly move the corrected value | ||
422 : | mov [_EDX], ax ; coeff[0] = ax | ||
423 : | mov _EBX, [_ESP + 2*PTR_SIZE] ; pop back the register value | ||
424 : | add _ESP, byte 4*PTR_SIZE ; "quant_intra 0" pushed _EBP, but we don't restore that one, just correct the stack offset by 16 | ||
425 : | edgomez | 1.2 | psubw mm7, mm6 ;D8 |
426 : | Isibaar | 1.9 | movq [_EDX + 3 * 32 + 16], mm5 ;C9 |
427 : | movq [_EDX + 3 * 32 + 24], mm7 ;D9 | ||
428 : | |||
429 : | xor _EAX, _EAX | ||
430 : | |||
431 : | %ifdef ARCH_IS_X86_64 | ||
432 : | mov TMP0, [_ESP] | ||
433 : | %ifndef WINDOWS | ||
434 : | add _ESP, 6*PTR_SIZE | ||
435 : | %else | ||
436 : | add _ESP, 4*PTR_SIZE | ||
437 : | %endif | ||
438 : | mov [_ESP], TMP0 | ||
439 : | %endif | ||
440 : | edgomez | 1.2 | |
441 : | ret | ||
442 : | Isibaar | 1.8 | ENDFUNC |
443 : | edgomez | 1.2 | |
444 : | |||
445 : | ;----------------------------------------------------------------------------- | ||
446 : | ; | ||
447 : | ; uint32_t quant_h263_inter_3dne(int16_t * coeff, | ||
448 : | ; const int16_t const * data, | ||
449 : | ; const uint32_t quant, | ||
450 : | ; const uint16_t *mpeg_matrices); | ||
451 : | ; | ||
452 : | ;----------------------------------------------------------------------------- | ||
453 : | ;This is Athlon-optimized code (ca 90 clk per call) | ||
454 : | ;Optimized by Jaan, 30 Nov 2002 | ||
455 : | |||
456 : | |||
457 : | %macro quantinter 1 | ||
458 : | Isibaar | 1.9 | movq mm1, [_EAX] ;A2 |
459 : | edgomez | 1.2 | psraw mm3, 15 ;B6 |
460 : | %if (%1) | ||
461 : | psubw mm2, mm6 ;C10 | ||
462 : | %endif | ||
463 : | psubw mm1, mm0 ;A3 | ||
464 : | pmulhw mm4, mm7 ;B7 | ||
465 : | Isibaar | 1.9 | movq mm6, [_ECX + %1*24+16] ;C1 |
466 : | edgomez | 1.2 | pmaxsw mm1, mm0 ;A4 |
467 : | paddw mm5, mm4 ;B8 | ||
468 : | %if (%1) | ||
469 : | Isibaar | 1.9 | movq [_EDX + %1*24+16-24], mm2 ;C11 |
470 : | edgomez | 1.2 | %endif |
471 : | Isibaar | 1.9 | psubusw mm1, [_EBX] ;A5 mm0 -= sub (unsigned, dont go < 0) |
472 : | edgomez | 1.2 | pxor mm4, mm3 ;B9 |
473 : | Isibaar | 1.9 | movq mm2, [_EAX] ;C2 |
474 : | edgomez | 1.2 | psraw mm0, 15 ;A6 |
475 : | psubw mm4, mm3 ;B10 | ||
476 : | psubw mm2, mm6 ;C3 | ||
477 : | pmulhw mm1, mm7 ;A7 mm0 = (mm0 / 2Q) >> 24 | ||
478 : | Isibaar | 1.9 | movq mm3, [_ECX + %1*24+8] ;B1 |
479 : | edgomez | 1.2 | pmaxsw mm2, mm6 ;C4 |
480 : | paddw mm5, mm1 ;A8 sum += mm0 | ||
481 : | %if (%1) | ||
482 : | Isibaar | 1.9 | movq [_EDX + %1*24+8-24], mm4 ;B11 |
483 : | edgomez | 1.2 | %else |
484 : | Isibaar | 1.9 | movq [_EDX + 120], mm4 ;B11 |
485 : | edgomez | 1.2 | %endif |
486 : | Isibaar | 1.9 | psubusw mm2, [_EBX] ;C5 |
487 : | edgomez | 1.2 | pxor mm1, mm0 ;A9 mm0 *= sign(mm0) |
488 : | Isibaar | 1.9 | movq mm4, [_EAX] ;B2 |
489 : | edgomez | 1.2 | psraw mm6, 15 ;C6 |
490 : | psubw mm1, mm0 ;A10 undisplace | ||
491 : | psubw mm4, mm3 ;B3 | ||
492 : | pmulhw mm2, mm7 ;C7 | ||
493 : | Isibaar | 1.9 | movq mm0, [_ECX + %1*24+24] ;A1 mm0 = [1st] |
494 : | edgomez | 1.2 | pmaxsw mm4, mm3 ;B4 |
495 : | paddw mm5, mm2 ;C8 | ||
496 : | Isibaar | 1.9 | movq [byte _EDX + %1*24], mm1 ;A11 |
497 : | psubusw mm4, [_EBX] ;B5 | ||
498 : | edgomez | 1.2 | pxor mm2, mm6 ;C9 |
499 : | %endmacro | ||
500 : | |||
501 : | %macro quantinter1 1 | ||
502 : | Isibaar | 1.9 | movq mm0, [byte _ECX + %1*16] ;mm0 = [1st] |
503 : | movq mm3, [_ECX + %1*16+8] ; | ||
504 : | movq mm1, [_EAX] | ||
505 : | movq mm4, [_EAX] | ||
506 : | edgomez | 1.2 | psubw mm1, mm0 |
507 : | psubw mm4, mm3 | ||
508 : | pmaxsw mm1, mm0 | ||
509 : | pmaxsw mm4, mm3 | ||
510 : | psubusw mm1, mm6 ; mm0 -= sub (unsigned, dont go < 0) | ||
511 : | psubusw mm4, mm6 ; | ||
512 : | psraw mm0, 15 | ||
513 : | psraw mm3, 15 | ||
514 : | psrlw mm1, 1 ; mm0 = (mm0 / 2Q) >> 16 | ||
515 : | psrlw mm4, 1 ; | ||
516 : | paddw mm5, mm1 ; sum += mm0 | ||
517 : | pxor mm1, mm0 ; mm0 *= sign(mm0) | ||
518 : | paddw mm5, mm4 | ||
519 : | pxor mm4, mm3 ; | ||
520 : | psubw mm1, mm0 ; undisplace | ||
521 : | psubw mm4, mm3 | ||
522 : | Isibaar | 1.9 | cmp _ESP, _ESP |
523 : | movq [byte _EDX + %1*16], mm1 | ||
524 : | movq [_EDX + %1*16+8], mm4 | ||
525 : | edgomez | 1.2 | %endmacro |
526 : | |||
527 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
528 : | edgomez | 1.2 | cglobal quant_h263_inter_3dne |
529 : | quant_h263_inter_3dne: | ||
530 : | Isibaar | 1.9 | |
531 : | %ifdef ARCH_IS_X86_64 | ||
532 : | mov TMP0, [_ESP] | ||
533 : | add _ESP, PTR_SIZE | ||
534 : | %ifndef WINDOWS | ||
535 : | push prm6 | ||
536 : | push prm5 | ||
537 : | %endif | ||
538 : | push prm4 | ||
539 : | push prm3 | ||
540 : | push prm2 | ||
541 : | push prm1 | ||
542 : | sub _ESP, PTR_SIZE | ||
543 : | mov [_ESP], TMP0 | ||
544 : | %endif | ||
545 : | |||
546 : | mov _EDX, [_ESP + 1*PTR_SIZE] ; coeff | ||
547 : | mov _ECX, [_ESP + 2*PTR_SIZE] ; data | ||
548 : | mov _EAX, [_ESP + 3*PTR_SIZE] ; quant | ||
549 : | push _EBX | ||
550 : | edgomez | 1.2 | |
551 : | pxor mm5, mm5 ; sum | ||
552 : | nop | ||
553 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
554 : | lea _EBX, [mmx_div] | ||
555 : | movq mm7, [_EBX + _EAX * 8 - 8] | ||
556 : | lea _EBX, [mmx_sub] | ||
557 : | lea _EBX, [_EBX + _EAX * 8 - 8] | ||
558 : | %else | ||
559 : | lea _EBX,[mmx_sub + _EAX * 8 - 8] ; sub | ||
560 : | movq mm7, [mmx_div + _EAX * 8 - 8] ; divider | ||
561 : | %endif | ||
562 : | edgomez | 1.2 | |
563 : | cmp al, 1 | ||
564 : | Isibaar | 1.9 | lea _EAX, [mmzero] |
565 : | edgomez | 1.2 | jz near .q1loop |
566 : | Isibaar | 1.9 | cmp _ESP, _ESP |
567 : | ALIGN SECTION_ALIGN | ||
568 : | movq mm3, [_ECX + 120] ;B1 | ||
569 : | edgomez | 1.2 | pxor mm4, mm4 ;B2 |
570 : | psubw mm4, mm3 ;B3 | ||
571 : | Isibaar | 1.9 | movq mm0, [_ECX] ;A1 mm0 = [1st] |
572 : | edgomez | 1.2 | pmaxsw mm4, mm3 ;B4 |
573 : | Isibaar | 1.9 | psubusw mm4, [_EBX] ;B5 |
574 : | edgomez | 1.2 | |
575 : | quantinter 0 | ||
576 : | quantinter 1 | ||
577 : | quantinter 2 | ||
578 : | quantinter 3 | ||
579 : | quantinter 4 | ||
580 : | |||
581 : | psraw mm3, 15 ;B6 | ||
582 : | psubw mm2, mm6 ;C10 | ||
583 : | pmulhw mm4, mm7 ;B7 | ||
584 : | paddw mm5, mm4 ;B8 | ||
585 : | pxor mm4, mm3 ;B9 | ||
586 : | psubw mm4, mm3 ;B10 | ||
587 : | Isibaar | 1.9 | movq [_EDX + 4*24+16], mm2 ;C11 |
588 : | pop _EBX | ||
589 : | movq [_EDX + 4*24+8], mm4 ;B11 | ||
590 : | edgomez | 1.2 | pmaddwd mm5, [plus_one] |
591 : | movq mm0, mm5 | ||
592 : | punpckhdq mm5, mm5 | ||
593 : | paddd mm0, mm5 | ||
594 : | movd eax, mm0 ; return sum | ||
595 : | |||
596 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
597 : | mov TMP0, [_ESP] | ||
598 : | %ifndef WINDOWS | ||
599 : | add _ESP, 6*PTR_SIZE | ||
600 : | %else | ||
601 : | add _ESP, 4*PTR_SIZE | ||
602 : | %endif | ||
603 : | mov [_ESP], TMP0 | ||
604 : | %endif | ||
605 : | |||
606 : | edgomez | 1.2 | ret |
607 : | |||
608 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
609 : | Isibaar | 1.8 | .q1loop: |
610 : | Isibaar | 1.9 | movq mm6, [byte _EBX] |
611 : | edgomez | 1.2 | |
612 : | quantinter1 0 | ||
613 : | quantinter1 1 | ||
614 : | quantinter1 2 | ||
615 : | quantinter1 3 | ||
616 : | quantinter1 4 | ||
617 : | quantinter1 5 | ||
618 : | quantinter1 6 | ||
619 : | quantinter1 7 | ||
620 : | |||
621 : | pmaddwd mm5, [plus_one] | ||
622 : | movq mm0, mm5 | ||
623 : | psrlq mm5, 32 | ||
624 : | paddd mm0, mm5 | ||
625 : | movd eax, mm0 ; return sum | ||
626 : | |||
627 : | Isibaar | 1.9 | pop _EBX |
628 : | |||
629 : | %ifdef ARCH_IS_X86_64 | ||
630 : | mov TMP0, [_ESP] | ||
631 : | %ifndef WINDOWS | ||
632 : | add _ESP, 6*PTR_SIZE | ||
633 : | %else | ||
634 : | add _ESP, 4*PTR_SIZE | ||
635 : | %endif | ||
636 : | mov [_ESP], TMP0 | ||
637 : | %endif | ||
638 : | edgomez | 1.2 | |
639 : | ret | ||
640 : | Isibaar | 1.8 | ENDFUNC |
641 : | edgomez | 1.2 | |
642 : | Isibaar | 1.9 | |
643 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
644 : | ; | ||
645 : | ; uint32_t dequant_h263_intra_3dne(int16_t *data, | ||
646 : | ; const int16_t const *coeff, | ||
647 : | ; const uint32_t quant, | ||
648 : | ; const uint32_t dcscalar, | ||
649 : | ; const uint16_t *mpeg_matrices); | ||
650 : | ; | ||
651 : | ;----------------------------------------------------------------------------- | ||
652 : | |||
653 : | ; this is the same as dequant_inter_3dne, except that we're | ||
654 : | ; saturating using 'pminsw' (saves 2 cycles/loop => ~5% faster) | ||
655 : | |||
656 : | ;This is Athlon-optimized code (ca 106 clk per call) | ||
657 : | |||
658 : | %macro dequant 1 | ||
659 : | Isibaar | 1.9 | movq mm1, [_ECX+%1*24] ; c = coeff[i] ;A2 |
660 : | edgomez | 1.2 | psubw mm0, mm1 ;-c ;A3 (1st dep) |
661 : | %if (%1) | ||
662 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
663 : | %endif | ||
664 : | pmaxsw mm0, mm1 ;|c| ;A4 (2nd) | ||
665 : | %if (%1) | ||
666 : | Isibaar | 1.9 | mov _EBP, _EBP |
667 : | pminsw mm4, [_EBX] ;C12 saturates to +2047 (5th+) later | ||
668 : | edgomez | 1.2 | %endif |
669 : | Isibaar | 1.9 | movq mm6, [_ESI] ;0 ;A5 mm6 in use |
670 : | pandn mm7, [_EAX] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
671 : | edgomez | 1.2 | %if (%1) |
672 : | pxor mm5, mm4 ;C13 (6th+) 1later | ||
673 : | %endif | ||
674 : | Isibaar | 1.9 | movq mm4, [_ESI] ;C1 ;0 |
675 : | mov _ESP, _ESP | ||
676 : | pcmpeqw mm6, [_ECX+%1*24] ;A6 (c ==0) ? -1 : 0 (1st) | ||
677 : | ALIGN SECTION_ALIGN | ||
678 : | edgomez | 1.2 | psraw mm1, 15 ; sign(c) ;A7 (2nd) |
679 : | %if (%1) | ||
680 : | Isibaar | 1.9 | movq [_EDX+%1*24+16-24], mm5 ; C14 (7th) 2later |
681 : | edgomez | 1.2 | %endif |
682 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
683 : | Isibaar | 1.9 | pmullw mm0, [_EDI] ;*= 2Q ;A8 (3rd+) |
684 : | edgomez | 1.2 | paddw mm2, mm7 ;B11 mm7 free (4th+) |
685 : | Isibaar | 1.9 | lea _EBP, [byte _EBP] |
686 : | movq mm5, [_ECX+%1*24+16] ;C2 ; c = coeff[i] | ||
687 : | edgomez | 1.2 | psubw mm4, mm5 ;-c ;C3 (1st dep) |
688 : | Isibaar | 1.9 | pandn mm6, [_EAX] ;A9 offset = isZero ? 0 : quant_add (2nd) |
689 : | pminsw mm2, [_EBX] ;B12 saturates to +2047 (5th+) | ||
690 : | edgomez | 1.2 | pxor mm3, mm2 ;B13 (6th+) |
691 : | Isibaar | 1.9 | movq mm2, [byte _ESI] ;B1 ;0 |
692 : | edgomez | 1.2 | %if (%1) |
693 : | Isibaar | 1.9 | movq [_EDX+%1*24+8-24], mm3 ;B14 (7th) |
694 : | edgomez | 1.2 | %else |
695 : | Isibaar | 1.9 | movq [_EDX+120], mm3 |
696 : | edgomez | 1.2 | %endif |
697 : | pmaxsw mm4, mm5 ;|c| ;C4 (2nd) | ||
698 : | paddw mm6, mm1 ;A10 offset +negate back (3rd) | ||
699 : | Isibaar | 1.9 | movq mm3, [_ECX+%1*24 + 8] ;B2 ; c = coeff[i] |
700 : | edgomez | 1.2 | psubw mm2, mm3 ;-c ;B3 (1st dep) |
701 : | paddw mm0, mm6 ;A11 mm6 free (4th+) | ||
702 : | Isibaar | 1.9 | movq mm6, [byte _ESI] ;0 ;C5 mm6 in use |
703 : | pcmpeqw mm6, [_ECX+%1*24+16] ;C6 (c ==0) ? -1 : 0 (1st) | ||
704 : | pminsw mm0, [_EBX] ;A12 saturates to +2047 (5th+) | ||
705 : | edgomez | 1.2 | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) |
706 : | pxor mm1, mm0 ;A13 (6th+) | ||
707 : | Isibaar | 1.9 | pmullw mm4, [_EDI] ;*= 2Q ;C8 (3rd+) |
708 : | edgomez | 1.2 | psraw mm5, 15 ; sign(c) ;C7 (2nd) |
709 : | Isibaar | 1.9 | movq mm7, [byte _ESI] ;0 ;B5 mm7 in use |
710 : | pcmpeqw mm7, [_ECX+%1*24 + 8] ;B6 (c ==0) ? -1 : 0 (1st) | ||
711 : | edgomez | 1.2 | %if (%1 < 4) |
712 : | Isibaar | 1.9 | movq mm0, [byte _ESI] ;A1 ;0 |
713 : | edgomez | 1.2 | %endif |
714 : | Isibaar | 1.9 | pandn mm6, [byte _EAX] ;C9 offset = isZero ? 0 : quant_add (2nd) |
715 : | edgomez | 1.2 | psraw mm3, 15 ;sign(c) ;B7 (2nd) |
716 : | Isibaar | 1.9 | movq [byte _EDX+%1*24], mm1 ;A14 (7th) |
717 : | edgomez | 1.2 | paddw mm6, mm5 ;C10 offset +negate back (3rd) |
718 : | Isibaar | 1.9 | pmullw mm2, [_EDI] ;*= 2Q ;B8 (3rd+) |
719 : | mov _ESP, _ESP | ||
720 : | edgomez | 1.2 | %endmacro |
721 : | |||
722 : | |||
723 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
724 : | edgomez | 1.2 | cglobal dequant_h263_intra_3dne |
725 : | dequant_h263_intra_3dne: | ||
726 : | Isibaar | 1.9 | |
727 : | %ifdef ARCH_IS_X86_64 | ||
728 : | mov TMP0, [_ESP] | ||
729 : | add _ESP, PTR_SIZE | ||
730 : | %ifndef WINDOWS | ||
731 : | push prm6 | ||
732 : | push prm5 | ||
733 : | %endif | ||
734 : | push prm4 | ||
735 : | push prm3 | ||
736 : | push prm2 | ||
737 : | push prm1 | ||
738 : | sub _ESP, PTR_SIZE | ||
739 : | mov [_ESP], TMP0 | ||
740 : | %endif | ||
741 : | |||
742 : | mov _ECX, [_ESP+ 2*PTR_SIZE] ; coeff | ||
743 : | mov _EAX, [_ESP+ 3*PTR_SIZE] ; quant | ||
744 : | edgomez | 1.2 | pxor mm0, mm0 |
745 : | pxor mm2, mm2 | ||
746 : | Isibaar | 1.9 | push _EDI |
747 : | push _EBX | ||
748 : | %ifdef ARCH_IS_X86_64 | ||
749 : | lea _EDI, [mmx_mul] | ||
750 : | lea _EDI, [_EDI + _EAX*8 - 8] ; 2*quant | ||
751 : | %else | ||
752 : | lea _EDI, [mmx_mul + _EAX*8 - 8] ; 2*quant | ||
753 : | %endif | ||
754 : | push _EBP | ||
755 : | Isibaar | 1.11 | lea _EBX, [mmx_2047] |
756 : | movsx _EBP, word [_ECX] | ||
757 : | Isibaar | 1.9 | %ifdef ARCH_IS_X86_64 |
758 : | lea r9, [mmx_add] | ||
759 : | lea _EAX, [r9 + _EAX*8 - 8] ; quant or quant-1 | ||
760 : | %else | ||
761 : | lea _EAX, [mmx_add + _EAX*8 - 8] ; quant or quant-1 | ||
762 : | %endif | ||
763 : | push _ESI | ||
764 : | Isibaar | 1.11 | lea _ESI, [mmzero] |
765 : | edgomez | 1.2 | pxor mm7, mm7 |
766 : | Isibaar | 1.9 | movq mm3, [_ECX+120] ;B2 ; c = coeff[i] |
767 : | pcmpeqw mm7, [_ECX+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
768 : | edgomez | 1.2 | |
769 : | Isibaar | 1.9 | imul _EBP, [_ESP+(4+4)*PTR_SIZE] ; dcscalar |
770 : | edgomez | 1.2 | psubw mm2, mm3 ;-c ;B3 (1st dep) |
771 : | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) | ||
772 : | Isibaar | 1.9 | pmullw mm2, [_EDI] ;*= 2Q ;B8 (3rd+) |
773 : | edgomez | 1.2 | psraw mm3, 15 ; sign(c) ;B7 (2nd) |
774 : | Isibaar | 1.9 | mov _EDX, [_ESP+ (1+4)*PTR_SIZE] ; data |
775 : | edgomez | 1.2 | |
776 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
777 : | edgomez | 1.2 | dequant 0 |
778 : | |||
779 : | Isibaar | 1.9 | cmp _EBP, -2048 |
780 : | mov _ESP, _ESP | ||
781 : | edgomez | 1.2 | |
782 : | dequant 1 | ||
783 : | |||
784 : | Isibaar | 1.9 | cmovl _EBP, [int_2048] |
785 : | edgomez | 1.2 | nop |
786 : | |||
787 : | dequant 2 | ||
788 : | |||
789 : | Isibaar | 1.9 | cmp _EBP, 2047 |
790 : | mov _ESP, _ESP | ||
791 : | edgomez | 1.2 | |
792 : | dequant 3 | ||
793 : | |||
794 : | Isibaar | 1.9 | cmovg _EBP, [int2047] |
795 : | edgomez | 1.2 | nop |
796 : | |||
797 : | dequant 4 | ||
798 : | |||
799 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
800 : | Isibaar | 1.9 | pminsw mm4, [_EBX] ;C12 saturates to +2047 (5th+) |
801 : | pandn mm7, [_EAX] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
802 : | mov _EAX, _EBP | ||
803 : | mov _ESI, [_ESP] | ||
804 : | mov _EBP, [_ESP+PTR_SIZE] | ||
805 : | edgomez | 1.2 | pxor mm5, mm4 ;C13 (6th+) |
806 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
807 : | Isibaar | 1.9 | movq [_EDX+4*24+16], mm5 ;C14 (7th) |
808 : | edgomez | 1.2 | paddw mm2, mm7 ;B11 mm7 free (4th+) |
809 : | Isibaar | 1.9 | pminsw mm2, [_EBX] ;B12 saturates to +2047 (5th+) |
810 : | mov _EBX, [_ESP+2*PTR_SIZE] | ||
811 : | mov _EDI, [_ESP+3*PTR_SIZE] | ||
812 : | add _ESP, byte 4*PTR_SIZE | ||
813 : | edgomez | 1.2 | pxor mm3, mm2 ;B13 (6th+) |
814 : | Isibaar | 1.9 | movq [_EDX+4*24+8], mm3 ;B14 (7th) |
815 : | mov [_EDX], ax | ||
816 : | |||
817 : | xor _EAX, _EAX | ||
818 : | |||
819 : | %ifdef ARCH_IS_X86_64 | ||
820 : | mov TMP0, [_ESP] | ||
821 : | %ifndef WINDOWS | ||
822 : | add _ESP, 6*PTR_SIZE | ||
823 : | %else | ||
824 : | add _ESP, 4*PTR_SIZE | ||
825 : | %endif | ||
826 : | mov [_ESP], TMP0 | ||
827 : | %endif | ||
828 : | edgomez | 1.2 | |
829 : | ret | ||
830 : | Isibaar | 1.8 | ENDFUNC |
831 : | edgomez | 1.2 | |
832 : | Isibaar | 1.9 | |
833 : | edgomez | 1.2 | ;----------------------------------------------------------------------------- |
834 : | ; | ||
835 : | ; uint32_t dequant_h263_inter_3dne(int16_t * data, | ||
836 : | ; const int16_t * const coeff, | ||
837 : | ; const uint32_t quant, | ||
838 : | ; const uint16_t *mpeg_matrices); | ||
839 : | ; | ||
840 : | ;----------------------------------------------------------------------------- | ||
841 : | |||
842 : | ; this is the same as dequant_inter_3dne, | ||
843 : | ; except that we're saturating using 'pminsw' (saves 2 cycles/loop) | ||
844 : | ; This is Athlon-optimized code (ca 100 clk per call) | ||
845 : | |||
846 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
847 : | edgomez | 1.2 | cglobal dequant_h263_inter_3dne |
848 : | dequant_h263_inter_3dne: | ||
849 : | Isibaar | 1.9 | |
850 : | %ifdef ARCH_IS_X86_64 | ||
851 : | mov TMP0, [_ESP] | ||
852 : | add _ESP, PTR_SIZE | ||
853 : | %ifndef WINDOWS | ||
854 : | push prm6 | ||
855 : | push prm5 | ||
856 : | %endif | ||
857 : | push prm4 | ||
858 : | push prm3 | ||
859 : | push prm2 | ||
860 : | push prm1 | ||
861 : | sub _ESP, PTR_SIZE | ||
862 : | mov [_ESP], TMP0 | ||
863 : | %endif | ||
864 : | |||
865 : | mov _ECX, [_ESP+ 2*PTR_SIZE] ; coeff | ||
866 : | mov _EAX, [_ESP+ 3*PTR_SIZE] ; quant | ||
867 : | edgomez | 1.2 | pxor mm0, mm0 |
868 : | pxor mm2, mm2 | ||
869 : | Isibaar | 1.9 | push _EDI |
870 : | push _EBX | ||
871 : | push _ESI | ||
872 : | %ifdef ARCH_IS_X86_64 | ||
873 : | lea _EDI, [mmx_mul] | ||
874 : | lea _EDI, [_EDI + _EAX*8 - 8] ; 2*quant | ||
875 : | %else | ||
876 : | lea _EDI, [mmx_mul + _EAX*8 - 8] ; 2*quant | ||
877 : | %endif | ||
878 : | Isibaar | 1.11 | lea _EBX, [mmx_2047] |
879 : | edgomez | 1.2 | pxor mm7, mm7 |
880 : | Isibaar | 1.9 | movq mm3, [_ECX+120] ;B2 ; c = coeff[i] |
881 : | pcmpeqw mm7, [_ECX+120] ;B6 (c ==0) ? -1 : 0 (1st) | ||
882 : | %ifdef ARCH_IS_X86_64 | ||
883 : | lea r9, [mmx_add] | ||
884 : | lea _EAX, [r9 + _EAX*8 - 8] ; quant or quant-1 | ||
885 : | %else | ||
886 : | lea _EAX, [mmx_add + _EAX*8 - 8] ; quant or quant-1 | ||
887 : | %endif | ||
888 : | edgomez | 1.2 | psubw mm2, mm3 ;-c ;B3 (1st dep) |
889 : | Isibaar | 1.11 | lea _ESI, [mmzero] |
890 : | edgomez | 1.2 | pmaxsw mm2, mm3 ;|c| ;B4 (2nd) |
891 : | Isibaar | 1.9 | pmullw mm2, [_EDI] ;*= 2Q ;B8 (3rd+) |
892 : | edgomez | 1.2 | psraw mm3, 15 ; sign(c) ;B7 (2nd) |
893 : | Isibaar | 1.9 | mov _EDX, [_ESP+ (1+3)*PTR_SIZE] ; data |
894 : | edgomez | 1.2 | |
895 : | Isibaar | 1.9 | ALIGN SECTION_ALIGN |
896 : | edgomez | 1.2 | |
897 : | dequant 0 | ||
898 : | dequant 1 | ||
899 : | dequant 2 | ||
900 : | dequant 3 | ||
901 : | dequant 4 | ||
902 : | |||
903 : | paddw mm4, mm6 ;C11 mm6 free (4th+) | ||
904 : | Isibaar | 1.9 | pminsw mm4, [_EBX] ;C12 saturates to +2047 (5th+) |
905 : | pandn mm7, [_EAX] ;B9 offset = isZero ? 0 : quant_add (2nd) | ||
906 : | mov _ESI, [_ESP] | ||
907 : | edgomez | 1.2 | pxor mm5, mm4 ;C13 (6th+) |
908 : | paddw mm7, mm3 ;B10 offset +negate back (3rd) | ||
909 : | Isibaar | 1.9 | movq [_EDX+4*24+16], mm5 ;C14 (7th) |
910 : | edgomez | 1.2 | paddw mm2, mm7 ;B11 mm7 free (4th+) |
911 : | Isibaar | 1.9 | pminsw mm2, [_EBX] ;B12 saturates to +2047 (5th+) |
912 : | mov _EBX, [_ESP+PTR_SIZE] | ||
913 : | mov _EDI, [_ESP+2*PTR_SIZE] | ||
914 : | add _ESP, byte 3*PTR_SIZE | ||
915 : | edgomez | 1.2 | pxor mm3, mm2 ;B13 (6th+) |
916 : | Isibaar | 1.9 | movq [_EDX+4*24+8], mm3 ;B14 (7th) |
917 : | |||
918 : | xor _EAX, _EAX | ||
919 : | |||
920 : | %ifdef ARCH_IS_X86_64 | ||
921 : | mov TMP0, [_ESP] | ||
922 : | %ifndef WINDOWS | ||
923 : | add _ESP, 6*PTR_SIZE | ||
924 : | %else | ||
925 : | add _ESP, 4*PTR_SIZE | ||
926 : | %endif | ||
927 : | mov [_ESP], TMP0 | ||
928 : | %endif | ||
929 : | edgomez | 1.2 | |
930 : | ret | ||
931 : | Isibaar | 1.8 | ENDFUNC |
932 : | edgomez | 1.5 | |
933 : | Isibaar | 1.12 | NON_EXEC_STACK |
No admin address has been configured | ViewVC Help |
Powered by ViewVC 1.0.4 |