FFmpeg
mpeg4videodsp.c
Go to the documentation of this file.
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * FFmpeg is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with FFmpeg; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18 
19 #include "config.h"
20 #include "libavutil/attributes.h"
21 #include "libavutil/cpu.h"
22 #include "libavutil/x86/cpu.h"
24 #include "libavcodec/videodsp.h"
25 
26 #if HAVE_INLINE_ASM
27 
28 static void gmc_mmx(uint8_t *dst, const uint8_t *src,
29  int stride, int h, int ox, int oy,
30  int dxx, int dxy, int dyx, int dyy,
31  int shift, int r, int width, int height)
32 {
33  const int w = 8;
34  const int ix = ox >> (16 + shift);
35  const int iy = oy >> (16 + shift);
36  const int oxs = ox >> 4;
37  const int oys = oy >> 4;
38  const int dxxs = dxx >> 4;
39  const int dxys = dxy >> 4;
40  const int dyxs = dyx >> 4;
41  const int dyys = dyy >> 4;
42  const uint16_t r4[4] = { r, r, r, r };
43  const uint16_t dxy4[4] = { dxys, dxys, dxys, dxys };
44  const uint16_t dyy4[4] = { dyys, dyys, dyys, dyys };
45  const uint64_t shift2 = 2 * shift;
46 #define MAX_STRIDE 4096U
47 #define MAX_H 8U
48  uint8_t edge_buf[(MAX_H + 1) * MAX_STRIDE];
49  int x, y;
50 
51  const int dxw = (dxx - (1 << (16 + shift))) * (w - 1);
52  const int dyh = (dyy - (1 << (16 + shift))) * (h - 1);
53  const int dxh = dxy * (h - 1);
54  const int dyw = dyx * (w - 1);
55  int need_emu = (unsigned) ix >= width - w || width < w ||
56  (unsigned) iy >= height - h || height< h
57  ;
58 
59  if ( // non-constant fullpel offset (3% of blocks)
60  ((ox ^ (ox + dxw)) | (ox ^ (ox + dxh)) | (ox ^ (ox + dxw + dxh)) |
61  (oy ^ (oy + dyw)) | (oy ^ (oy + dyh)) | (oy ^ (oy + dyw + dyh))) >> (16 + shift) ||
62  // uses more than 16 bits of subpel mv (only at huge resolution)
63  (dxx | dxy | dyx | dyy) & 15 ||
64  (need_emu && (h > MAX_H || stride > MAX_STRIDE))) {
65  // FIXME could still use mmx for some of the rows
66  ff_gmc_c(dst, src, stride, h, ox, oy, dxx, dxy, dyx, dyy,
67  shift, r, width, height);
68  return;
69  }
70 
71  src += ix + iy * stride;
72  if (need_emu) {
73  ff_emulated_edge_mc_8(edge_buf, src, stride, stride, w + 1, h + 1, ix, iy, width, height);
74  src = edge_buf;
75  }
76 
77  __asm__ volatile (
78  "movd %0, %%mm6 \n\t"
79  "pxor %%mm7, %%mm7 \n\t"
80  "punpcklwd %%mm6, %%mm6 \n\t"
81  "punpcklwd %%mm6, %%mm6 \n\t"
82  :: "r" (1 << shift));
83 
84  for (x = 0; x < w; x += 4) {
85  uint16_t dx4[4] = { oxs - dxys + dxxs * (x + 0),
86  oxs - dxys + dxxs * (x + 1),
87  oxs - dxys + dxxs * (x + 2),
88  oxs - dxys + dxxs * (x + 3) };
89  uint16_t dy4[4] = { oys - dyys + dyxs * (x + 0),
90  oys - dyys + dyxs * (x + 1),
91  oys - dyys + dyxs * (x + 2),
92  oys - dyys + dyxs * (x + 3) };
93 
94  for (y = 0; y < h; y++) {
95  __asm__ volatile (
96  "movq %0, %%mm4 \n\t"
97  "movq %1, %%mm5 \n\t"
98  "paddw %2, %%mm4 \n\t"
99  "paddw %3, %%mm5 \n\t"
100  "movq %%mm4, %0 \n\t"
101  "movq %%mm5, %1 \n\t"
102  "psrlw $12, %%mm4 \n\t"
103  "psrlw $12, %%mm5 \n\t"
104  : "+m" (*dx4), "+m" (*dy4)
105  : "m" (*dxy4), "m" (*dyy4));
106 
107  __asm__ volatile (
108  "movq %%mm6, %%mm2 \n\t"
109  "movq %%mm6, %%mm1 \n\t"
110  "psubw %%mm4, %%mm2 \n\t"
111  "psubw %%mm5, %%mm1 \n\t"
112  "movq %%mm2, %%mm0 \n\t"
113  "movq %%mm4, %%mm3 \n\t"
114  "pmullw %%mm1, %%mm0 \n\t" // (s - dx) * (s - dy)
115  "pmullw %%mm5, %%mm3 \n\t" // dx * dy
116  "pmullw %%mm5, %%mm2 \n\t" // (s - dx) * dy
117  "pmullw %%mm4, %%mm1 \n\t" // dx * (s - dy)
118 
119  "movd %4, %%mm5 \n\t"
120  "movd %3, %%mm4 \n\t"
121  "punpcklbw %%mm7, %%mm5 \n\t"
122  "punpcklbw %%mm7, %%mm4 \n\t"
123  "pmullw %%mm5, %%mm3 \n\t" // src[1, 1] * dx * dy
124  "pmullw %%mm4, %%mm2 \n\t" // src[0, 1] * (s - dx) * dy
125 
126  "movd %2, %%mm5 \n\t"
127  "movd %1, %%mm4 \n\t"
128  "punpcklbw %%mm7, %%mm5 \n\t"
129  "punpcklbw %%mm7, %%mm4 \n\t"
130  "pmullw %%mm5, %%mm1 \n\t" // src[1, 0] * dx * (s - dy)
131  "pmullw %%mm4, %%mm0 \n\t" // src[0, 0] * (s - dx) * (s - dy)
132  "paddw %5, %%mm1 \n\t"
133  "paddw %%mm3, %%mm2 \n\t"
134  "paddw %%mm1, %%mm0 \n\t"
135  "paddw %%mm2, %%mm0 \n\t"
136 
137  "psrlw %6, %%mm0 \n\t"
138  "packuswb %%mm0, %%mm0 \n\t"
139  "movd %%mm0, %0 \n\t"
140 
141  : "=m" (dst[x + y * stride])
142  : "m" (src[0]), "m" (src[1]),
143  "m" (src[stride]), "m" (src[stride + 1]),
144  "m" (*r4), "m" (shift2));
145  src += stride;
146  }
147  src += 4 - h * stride;
148  }
149 }
150 
151 #endif /* HAVE_INLINE_ASM */
152 
154 {
155 #if HAVE_INLINE_ASM
156  int cpu_flags = av_get_cpu_flags();
157 
158  if (INLINE_MMX(cpu_flags))
159  c->gmc = gmc_mmx;
160 #endif /* HAVE_INLINE_ASM */
161 }
INLINE_MMX
#define INLINE_MMX(flags)
Definition: cpu.h:87
MAX_STRIDE
#define MAX_STRIDE
Definition: llviddspenc.c:42
ff_mpeg4videodsp_init_x86
av_cold void ff_mpeg4videodsp_init_x86(Mpeg4VideoDSPContext *c)
Definition: mpeg4videodsp.c:153
cpu.h
r
const char * r
Definition: vf_curves.c:127
w
uint8_t w
Definition: llviddspenc.c:38
Mpeg4VideoDSPContext
Definition: mpeg4videodsp.h:28
av_get_cpu_flags
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
Definition: cpu.c:109
cpu_flags
static atomic_int cpu_flags
Definition: cpu.c:56
av_cold
#define av_cold
Definition: attributes.h:90
ff_gmc_c
void ff_gmc_c(uint8_t *dst, const uint8_t *src, int stride, int h, int ox, int oy, int dxx, int dxy, int dyx, int dyy, int shift, int r, int width, int height)
Definition: mpeg4videodsp.c:47
c
Undefined Behavior In the C some operations are like signed integer dereferencing freed accessing outside allocated Undefined Behavior must not occur in a C it is not safe even if the output of undefined operations is unused The unsafety may seem nit picking but Optimizing compilers have in fact optimized code on the assumption that no undefined Behavior occurs Optimizing code based on wrong assumptions can and has in some cases lead to effects beyond the output of computations The signed integer overflow problem in speed critical code Code which is highly optimized and works with signed integers sometimes has the problem that often the output of the computation does not c
Definition: undefined.txt:32
height
#define height
Definition: dsp.h:85
shift
static int shift(int a, int b)
Definition: bonk.c:261
dst
uint8_t ptrdiff_t const uint8_t ptrdiff_t int intptr_t intptr_t int int16_t * dst
Definition: dsp.h:83
cpu.h
attributes.h
shift2
static const uint8_t shift2[6]
Definition: dxa.c:49
stride
#define stride
Definition: h264pred_template.c:537
__asm__
__asm__(".macro parse_r var r\n\t" "\\var = -1\n\t" _IFC_REG(0) _IFC_REG(1) _IFC_REG(2) _IFC_REG(3) _IFC_REG(4) _IFC_REG(5) _IFC_REG(6) _IFC_REG(7) _IFC_REG(8) _IFC_REG(9) _IFC_REG(10) _IFC_REG(11) _IFC_REG(12) _IFC_REG(13) _IFC_REG(14) _IFC_REG(15) _IFC_REG(16) _IFC_REG(17) _IFC_REG(18) _IFC_REG(19) _IFC_REG(20) _IFC_REG(21) _IFC_REG(22) _IFC_REG(23) _IFC_REG(24) _IFC_REG(25) _IFC_REG(26) _IFC_REG(27) _IFC_REG(28) _IFC_REG(29) _IFC_REG(30) _IFC_REG(31) ".iflt \\var\n\t" ".error \"Unable to parse register name \\r\"\n\t" ".endif\n\t" ".endm")
mpeg4videodsp.h
videodsp.h
h
h
Definition: vp9dsp_template.c:2070
width
#define width
Definition: dsp.h:85
src
#define src
Definition: vp8dsp.c:248