2
* VC-1 and WMV3 - DSP functions MMX-optimized
3
* Copyright (c) 2007 Christophe GISQUET <christophe.gisquet@free.fr>
5
* Permission is hereby granted, free of charge, to any person
6
* obtaining a copy of this software and associated documentation
7
* files (the "Software"), to deal in the Software without
8
* restriction, including without limitation the rights to use,
9
* copy, modify, merge, publish, distribute, sublicense, and/or sell
10
* copies of the Software, and to permit persons to whom the
11
* Software is furnished to do so, subject to the following
14
* The above copyright notice and this permission notice shall be
15
* included in all copies or substantial portions of the Software.
17
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
18
* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
19
* OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
20
* NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT
21
* HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
22
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
23
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24
* OTHER DEALINGS IN THE SOFTWARE.
27
#include "libavutil/x86_cpu.h"
28
#include "libavcodec/dsputil.h"
29
#include "dsputil_mmx.h"
31
/** Add rounder from mm7 to mm3 and pack result at destination */
32
#define NORMALIZE_MMX(SHIFT) \
33
"paddw %%mm7, %%mm3 \n\t" /* +bias-r */ \
34
"paddw %%mm7, %%mm4 \n\t" /* +bias-r */ \
35
"psraw "SHIFT", %%mm3 \n\t" \
36
"psraw "SHIFT", %%mm4 \n\t"
38
#define TRANSFER_DO_PACK \
39
"packuswb %%mm4, %%mm3 \n\t" \
40
"movq %%mm3, (%2) \n\t"
42
#define TRANSFER_DONT_PACK \
43
"movq %%mm3, 0(%2) \n\t" \
44
"movq %%mm4, 8(%2) \n\t"
46
/** @see MSPEL_FILTER13_CORE for use as UNPACK macro */
47
#define DO_UNPACK(reg) "punpcklbw %%mm0, " reg "\n\t"
48
#define DONT_UNPACK(reg)
50
/** Compute the rounder 32-r or 8-r and unpacks it to mm7 */
51
#define LOAD_ROUNDER_MMX(ROUND) \
52
"movd "ROUND", %%mm7 \n\t" \
53
"punpcklwd %%mm7, %%mm7 \n\t" \
54
"punpckldq %%mm7, %%mm7 \n\t"
56
#define SHIFT2_LINE(OFF, R0,R1,R2,R3) \
57
"paddw %%mm"#R2", %%mm"#R1" \n\t" \
58
"movd (%0,%3), %%mm"#R0" \n\t" \
59
"pmullw %%mm6, %%mm"#R1" \n\t" \
60
"punpcklbw %%mm0, %%mm"#R0" \n\t" \
61
"movd (%0,%2), %%mm"#R3" \n\t" \
62
"psubw %%mm"#R0", %%mm"#R1" \n\t" \
63
"punpcklbw %%mm0, %%mm"#R3" \n\t" \
64
"paddw %%mm7, %%mm"#R1" \n\t" \
65
"psubw %%mm"#R3", %%mm"#R1" \n\t" \
66
"psraw %4, %%mm"#R1" \n\t" \
67
"movq %%mm"#R1", "#OFF"(%1) \n\t" \
70
DECLARE_ALIGNED_16(const uint64_t, ff_pw_9) = 0x0009000900090009ULL;
72
/** Sacrifying mm6 allows to pipeline loads from src */
73
static void vc1_put_ver_16b_shift2_mmx(int16_t *dst,
74
const uint8_t *src, x86_reg stride,
75
int rnd, int64_t shift)
78
"mov $3, %%"REG_c" \n\t"
79
LOAD_ROUNDER_MMX("%5")
80
"movq "MANGLE(ff_pw_9)", %%mm6 \n\t"
82
"movd (%0), %%mm2 \n\t"
84
"movd (%0), %%mm3 \n\t"
85
"punpcklbw %%mm0, %%mm2 \n\t"
86
"punpcklbw %%mm0, %%mm3 \n\t"
87
SHIFT2_LINE( 0, 1, 2, 3, 4)
88
SHIFT2_LINE( 24, 2, 3, 4, 1)
89
SHIFT2_LINE( 48, 3, 4, 1, 2)
90
SHIFT2_LINE( 72, 4, 1, 2, 3)
91
SHIFT2_LINE( 96, 1, 2, 3, 4)
92
SHIFT2_LINE(120, 2, 3, 4, 1)
93
SHIFT2_LINE(144, 3, 4, 1, 2)
94
SHIFT2_LINE(168, 4, 1, 2, 3)
99
: "+r"(src), "+r"(dst)
100
: "r"(stride), "r"(-2*stride),
101
"m"(shift), "m"(rnd), "r"(9*stride-4)
107
* Data is already unpacked, so some operations can directly be made from
110
static void vc1_put_hor_16b_shift2_mmx(uint8_t *dst, x86_reg stride,
111
const int16_t *src, int rnd)
116
rnd -= (-1+9+9-1)*1024; /* Add -1024 bias */
118
LOAD_ROUNDER_MMX("%4")
119
"movq "MANGLE(ff_pw_128)", %%mm6\n\t"
120
"movq "MANGLE(ff_pw_9)", %%mm5 \n\t"
122
"movq 2*0+0(%1), %%mm1 \n\t"
123
"movq 2*0+8(%1), %%mm2 \n\t"
124
"movq 2*1+0(%1), %%mm3 \n\t"
125
"movq 2*1+8(%1), %%mm4 \n\t"
126
"paddw 2*3+0(%1), %%mm1 \n\t"
127
"paddw 2*3+8(%1), %%mm2 \n\t"
128
"paddw 2*2+0(%1), %%mm3 \n\t"
129
"paddw 2*2+8(%1), %%mm4 \n\t"
130
"pmullw %%mm5, %%mm3 \n\t"
131
"pmullw %%mm5, %%mm4 \n\t"
132
"psubw %%mm1, %%mm3 \n\t"
133
"psubw %%mm2, %%mm4 \n\t"
136
"paddw %%mm6, %%mm3 \n\t"
137
"paddw %%mm6, %%mm4 \n\t"
143
: "+r"(h), "+r" (src), "+r" (dst)
144
: "r"(stride), "m"(rnd)
151
* Purely vertical or horizontal 1/2 shift interpolation.
152
* Sacrify mm6 for *9 factor.
154
static void vc1_put_shift2_mmx(uint8_t *dst, const uint8_t *src,
155
x86_reg stride, int rnd, x86_reg offset)
159
"mov $8, %%"REG_c" \n\t"
160
LOAD_ROUNDER_MMX("%5")
161
"movq "MANGLE(ff_pw_9)", %%mm6\n\t"
163
"movd 0(%0 ), %%mm3 \n\t"
164
"movd 4(%0 ), %%mm4 \n\t"
165
"movd 0(%0,%2), %%mm1 \n\t"
166
"movd 4(%0,%2), %%mm2 \n\t"
168
"punpcklbw %%mm0, %%mm3 \n\t"
169
"punpcklbw %%mm0, %%mm4 \n\t"
170
"punpcklbw %%mm0, %%mm1 \n\t"
171
"punpcklbw %%mm0, %%mm2 \n\t"
172
"paddw %%mm1, %%mm3 \n\t"
173
"paddw %%mm2, %%mm4 \n\t"
174
"movd 0(%0,%3), %%mm1 \n\t"
175
"movd 4(%0,%3), %%mm2 \n\t"
176
"pmullw %%mm6, %%mm3 \n\t" /* 0,9,9,0*/
177
"pmullw %%mm6, %%mm4 \n\t" /* 0,9,9,0*/
178
"punpcklbw %%mm0, %%mm1 \n\t"
179
"punpcklbw %%mm0, %%mm2 \n\t"
180
"psubw %%mm1, %%mm3 \n\t" /*-1,9,9,0*/
181
"psubw %%mm2, %%mm4 \n\t" /*-1,9,9,0*/
182
"movd 0(%0,%2), %%mm1 \n\t"
183
"movd 4(%0,%2), %%mm2 \n\t"
184
"punpcklbw %%mm0, %%mm1 \n\t"
185
"punpcklbw %%mm0, %%mm2 \n\t"
186
"psubw %%mm1, %%mm3 \n\t" /*-1,9,9,-1*/
187
"psubw %%mm2, %%mm4 \n\t" /*-1,9,9,-1*/
189
"packuswb %%mm4, %%mm3 \n\t"
190
"movq %%mm3, (%1) \n\t"
195
: "+r"(src), "+r"(dst)
196
: "r"(offset), "r"(-2*offset), "g"(stride), "m"(rnd),
203
* Filter coefficients made global to allow access by all 1 or 3 quarter shift
204
* interpolation functions.
206
DECLARE_ASM_CONST(16, uint64_t, ff_pw_53) = 0x0035003500350035ULL;
207
DECLARE_ASM_CONST(16, uint64_t, ff_pw_18) = 0x0012001200120012ULL;
210
* Core of the 1/4 and 3/4 shift bicubic interpolation.
212
* @param UNPACK Macro unpacking arguments from 8 to 16bits (can be empty).
213
* @param MOVQ "movd 1" or "movq 2", if data read is already unpacked.
214
* @param A1 Address of 1st tap (beware of unpacked/packed).
215
* @param A2 Address of 2nd tap
216
* @param A3 Address of 3rd tap
217
* @param A4 Address of 4th tap
219
#define MSPEL_FILTER13_CORE(UNPACK, MOVQ, A1, A2, A3, A4) \
220
MOVQ "*0+"A1", %%mm1 \n\t" \
221
MOVQ "*4+"A1", %%mm2 \n\t" \
224
"pmullw "MANGLE(ff_pw_3)", %%mm1\n\t" \
225
"pmullw "MANGLE(ff_pw_3)", %%mm2\n\t" \
226
MOVQ "*0+"A2", %%mm3 \n\t" \
227
MOVQ "*4+"A2", %%mm4 \n\t" \
230
"pmullw %%mm6, %%mm3 \n\t" /* *18 */ \
231
"pmullw %%mm6, %%mm4 \n\t" /* *18 */ \
232
"psubw %%mm1, %%mm3 \n\t" /* 18,-3 */ \
233
"psubw %%mm2, %%mm4 \n\t" /* 18,-3 */ \
234
MOVQ "*0+"A4", %%mm1 \n\t" \
235
MOVQ "*4+"A4", %%mm2 \n\t" \
238
"psllw $2, %%mm1 \n\t" /* 4* */ \
239
"psllw $2, %%mm2 \n\t" /* 4* */ \
240
"psubw %%mm1, %%mm3 \n\t" /* -4,18,-3 */ \
241
"psubw %%mm2, %%mm4 \n\t" /* -4,18,-3 */ \
242
MOVQ "*0+"A3", %%mm1 \n\t" \
243
MOVQ "*4+"A3", %%mm2 \n\t" \
246
"pmullw %%mm5, %%mm1 \n\t" /* *53 */ \
247
"pmullw %%mm5, %%mm2 \n\t" /* *53 */ \
248
"paddw %%mm1, %%mm3 \n\t" /* 4,53,18,-3 */ \
249
"paddw %%mm2, %%mm4 \n\t" /* 4,53,18,-3 */
252
* Macro to build the vertical 16bits version of vc1_put_shift[13].
253
* Here, offset=src_stride. Parameters passed A1 to A4 must use
254
* %3 (src_stride) and %4 (3*src_stride).
256
* @param NAME Either 1 or 3
257
* @see MSPEL_FILTER13_CORE for information on A1->A4
259
#define MSPEL_FILTER13_VER_16B(NAME, A1, A2, A3, A4) \
261
vc1_put_ver_16b_ ## NAME ## _mmx(int16_t *dst, const uint8_t *src, \
262
x86_reg src_stride, \
263
int rnd, int64_t shift) \
268
LOAD_ROUNDER_MMX("%5") \
269
"movq "MANGLE(ff_pw_53)", %%mm5\n\t" \
270
"movq "MANGLE(ff_pw_18)", %%mm6\n\t" \
273
MSPEL_FILTER13_CORE(DO_UNPACK, "movd 1", A1, A2, A3, A4) \
274
NORMALIZE_MMX("%6") \
276
/* Last 3 (in fact 4) bytes on the line */ \
277
"movd 8+"A1", %%mm1 \n\t" \
279
"movq %%mm1, %%mm3 \n\t" \
280
"paddw %%mm1, %%mm1 \n\t" \
281
"paddw %%mm3, %%mm1 \n\t" /* 3* */ \
282
"movd 8+"A2", %%mm3 \n\t" \
284
"pmullw %%mm6, %%mm3 \n\t" /* *18 */ \
285
"psubw %%mm1, %%mm3 \n\t" /*18,-3 */ \
286
"movd 8+"A3", %%mm1 \n\t" \
288
"pmullw %%mm5, %%mm1 \n\t" /* *53 */ \
289
"paddw %%mm1, %%mm3 \n\t" /*53,18,-3 */ \
290
"movd 8+"A4", %%mm1 \n\t" \
292
"psllw $2, %%mm1 \n\t" /* 4* */ \
293
"psubw %%mm1, %%mm3 \n\t" \
294
"paddw %%mm7, %%mm3 \n\t" \
295
"psraw %6, %%mm3 \n\t" \
296
"movq %%mm3, 16(%2) \n\t" \
301
: "+r"(h), "+r" (src), "+r" (dst) \
302
: "r"(src_stride), "r"(3*src_stride), \
303
"m"(rnd), "m"(shift) \
309
* Macro to build the horizontal 16bits version of vc1_put_shift[13].
310
* Here, offset=16bits, so parameters passed A1 to A4 should be simple.
312
* @param NAME Either 1 or 3
313
* @see MSPEL_FILTER13_CORE for information on A1->A4
315
#define MSPEL_FILTER13_HOR_16B(NAME, A1, A2, A3, A4) \
317
vc1_put_hor_16b_ ## NAME ## _mmx(uint8_t *dst, x86_reg stride, \
318
const int16_t *src, int rnd) \
322
rnd -= (-4+58+13-3)*256; /* Add -256 bias */ \
324
LOAD_ROUNDER_MMX("%4") \
325
"movq "MANGLE(ff_pw_18)", %%mm6 \n\t" \
326
"movq "MANGLE(ff_pw_53)", %%mm5 \n\t" \
329
MSPEL_FILTER13_CORE(DONT_UNPACK, "movq 2", A1, A2, A3, A4) \
330
NORMALIZE_MMX("$7") \
332
"paddw "MANGLE(ff_pw_128)", %%mm3 \n\t" \
333
"paddw "MANGLE(ff_pw_128)", %%mm4 \n\t" \
339
: "+r"(h), "+r" (src), "+r" (dst) \
340
: "r"(stride), "m"(rnd) \
346
* Macro to build the 8bits, any direction, version of vc1_put_shift[13].
347
* Here, offset=src_stride. Parameters passed A1 to A4 must use
348
* %3 (offset) and %4 (3*offset).
350
* @param NAME Either 1 or 3
351
* @see MSPEL_FILTER13_CORE for information on A1->A4
353
#define MSPEL_FILTER13_8B(NAME, A1, A2, A3, A4) \
355
vc1_put_## NAME ## _mmx(uint8_t *dst, const uint8_t *src, \
356
x86_reg stride, int rnd, x86_reg offset) \
362
LOAD_ROUNDER_MMX("%6") \
363
"movq "MANGLE(ff_pw_53)", %%mm5 \n\t" \
364
"movq "MANGLE(ff_pw_18)", %%mm6 \n\t" \
367
MSPEL_FILTER13_CORE(DO_UNPACK, "movd 1", A1, A2, A3, A4) \
368
NORMALIZE_MMX("$6") \
374
: "+r"(h), "+r" (src), "+r" (dst) \
375
: "r"(offset), "r"(3*offset), "g"(stride), "m"(rnd) \
380
/** 1/4 shift bicubic interpolation */
381
MSPEL_FILTER13_8B (shift1, "0(%1,%4 )", "0(%1,%3,2)", "0(%1,%3 )", "0(%1 )")
382
MSPEL_FILTER13_VER_16B(shift1, "0(%1,%4 )", "0(%1,%3,2)", "0(%1,%3 )", "0(%1 )")
383
MSPEL_FILTER13_HOR_16B(shift1, "2*3(%1)", "2*2(%1)", "2*1(%1)", "2*0(%1)")
385
/** 3/4 shift bicubic interpolation */
386
MSPEL_FILTER13_8B (shift3, "0(%1 )", "0(%1,%3 )", "0(%1,%3,2)", "0(%1,%4 )")
387
MSPEL_FILTER13_VER_16B(shift3, "0(%1 )", "0(%1,%3 )", "0(%1,%3,2)", "0(%1,%4 )")
388
MSPEL_FILTER13_HOR_16B(shift3, "2*0(%1)", "2*1(%1)", "2*2(%1)", "2*3(%1)")
390
typedef void (*vc1_mspel_mc_filter_ver_16bits)(int16_t *dst, const uint8_t *src, x86_reg src_stride, int rnd, int64_t shift);
391
typedef void (*vc1_mspel_mc_filter_hor_16bits)(uint8_t *dst, x86_reg dst_stride, const int16_t *src, int rnd);
392
typedef void (*vc1_mspel_mc_filter_8bits)(uint8_t *dst, const uint8_t *src, x86_reg stride, int rnd, x86_reg offset);
395
* Interpolates fractional pel values by applying proper vertical then
398
* @param dst Destination buffer for interpolated pels.
399
* @param src Source buffer.
400
* @param stride Stride for both src and dst buffers.
401
* @param hmode Horizontal filter (expressed in quarter pixels shift).
402
* @param hmode Vertical filter.
403
* @param rnd Rounding bias.
405
static void vc1_mspel_mc(uint8_t *dst, const uint8_t *src, int stride,
406
int hmode, int vmode, int rnd)
408
static const vc1_mspel_mc_filter_ver_16bits vc1_put_shift_ver_16bits[] =
409
{ NULL, vc1_put_ver_16b_shift1_mmx, vc1_put_ver_16b_shift2_mmx, vc1_put_ver_16b_shift3_mmx };
410
static const vc1_mspel_mc_filter_hor_16bits vc1_put_shift_hor_16bits[] =
411
{ NULL, vc1_put_hor_16b_shift1_mmx, vc1_put_hor_16b_shift2_mmx, vc1_put_hor_16b_shift3_mmx };
412
static const vc1_mspel_mc_filter_8bits vc1_put_shift_8bits[] =
413
{ NULL, vc1_put_shift1_mmx, vc1_put_shift2_mmx, vc1_put_shift3_mmx };
416
"pxor %%mm0, %%mm0 \n\t"
420
if (vmode) { /* Vertical filter to apply */
421
if (hmode) { /* Horizontal filter to apply, output to tmp */
422
static const int shift_value[] = { 0, 5, 1, 5 };
423
int shift = (shift_value[hmode]+shift_value[vmode])>>1;
425
DECLARE_ALIGNED_16(int16_t, tmp[12*8]);
427
r = (1<<(shift-1)) + rnd-1;
428
vc1_put_shift_ver_16bits[vmode](tmp, src-1, stride, r, shift);
430
vc1_put_shift_hor_16bits[hmode](dst, stride, tmp+1, 64-rnd);
433
else { /* No horizontal filter, output 8 lines to dst */
434
vc1_put_shift_8bits[vmode](dst, src, stride, 1-rnd, stride);
439
/* Horizontal mode with no vertical mode */
440
vc1_put_shift_8bits[hmode](dst, src, stride, rnd, 1);
443
void ff_put_vc1_mspel_mc00_mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd);
445
/** Macro to ease bicubic filter interpolation functions declarations */
446
#define DECLARE_FUNCTION(a, b) \
447
static void put_vc1_mspel_mc ## a ## b ## _mmx(uint8_t *dst, const uint8_t *src, int stride, int rnd) { \
448
vc1_mspel_mc(dst, src, stride, a, b, rnd); \
451
DECLARE_FUNCTION(0, 1)
452
DECLARE_FUNCTION(0, 2)
453
DECLARE_FUNCTION(0, 3)
455
DECLARE_FUNCTION(1, 0)
456
DECLARE_FUNCTION(1, 1)
457
DECLARE_FUNCTION(1, 2)
458
DECLARE_FUNCTION(1, 3)
460
DECLARE_FUNCTION(2, 0)
461
DECLARE_FUNCTION(2, 1)
462
DECLARE_FUNCTION(2, 2)
463
DECLARE_FUNCTION(2, 3)
465
DECLARE_FUNCTION(3, 0)
466
DECLARE_FUNCTION(3, 1)
467
DECLARE_FUNCTION(3, 2)
468
DECLARE_FUNCTION(3, 3)
470
void ff_vc1dsp_init_mmx(DSPContext* dsp, AVCodecContext *avctx) {
471
dsp->put_vc1_mspel_pixels_tab[ 0] = ff_put_vc1_mspel_mc00_mmx;
472
dsp->put_vc1_mspel_pixels_tab[ 4] = put_vc1_mspel_mc01_mmx;
473
dsp->put_vc1_mspel_pixels_tab[ 8] = put_vc1_mspel_mc02_mmx;
474
dsp->put_vc1_mspel_pixels_tab[12] = put_vc1_mspel_mc03_mmx;
476
dsp->put_vc1_mspel_pixels_tab[ 1] = put_vc1_mspel_mc10_mmx;
477
dsp->put_vc1_mspel_pixels_tab[ 5] = put_vc1_mspel_mc11_mmx;
478
dsp->put_vc1_mspel_pixels_tab[ 9] = put_vc1_mspel_mc12_mmx;
479
dsp->put_vc1_mspel_pixels_tab[13] = put_vc1_mspel_mc13_mmx;
481
dsp->put_vc1_mspel_pixels_tab[ 2] = put_vc1_mspel_mc20_mmx;
482
dsp->put_vc1_mspel_pixels_tab[ 6] = put_vc1_mspel_mc21_mmx;
483
dsp->put_vc1_mspel_pixels_tab[10] = put_vc1_mspel_mc22_mmx;
484
dsp->put_vc1_mspel_pixels_tab[14] = put_vc1_mspel_mc23_mmx;
486
dsp->put_vc1_mspel_pixels_tab[ 3] = put_vc1_mspel_mc30_mmx;
487
dsp->put_vc1_mspel_pixels_tab[ 7] = put_vc1_mspel_mc31_mmx;
488
dsp->put_vc1_mspel_pixels_tab[11] = put_vc1_mspel_mc32_mmx;
489
dsp->put_vc1_mspel_pixels_tab[15] = put_vc1_mspel_mc33_mmx;
added
removed
libavcodec/x86/vc1dsp_mmx.c
