annotate x86/h264_qpel_mmx.c @ 12492:58a960d6e34c libavcodec

Rename h264_idct_sse2.asm to h264_idct.asm; move inline IDCT asm from h264dsp_mmx.c to h264_idct.asm (as yasm code). Because the loops are now coded in asm instead of C, this is (depending on the function) up to 50% faster for cases where gcc didn't do a great job at looping. Since h264_idct_add8() is now faster than the manual loop setup in h264.c, in-asm idct calling can now be enabled for chroma as well (see r16207). For MMX, this is 5% faster. For SSE2 (which isn't done for chroma if h264.c does the looping), this makes it up to 50% faster. Speed gain overall is ~0.5-1.0%.
author rbultje
date Tue, 14 Sep 2010 13:36:26 +0000
parents 3941687b4fa9
children
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
12450
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1 /*
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
2 * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
3 *
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
4 * This file is part of FFmpeg.
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
5 *
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
6 * FFmpeg is free software; you can redistribute it and/or
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
7 * modify it under the terms of the GNU Lesser General Public
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
8 * License as published by the Free Software Foundation; either
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
9 * version 2.1 of the License, or (at your option) any later version.
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
10 *
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
11 * FFmpeg is distributed in the hope that it will be useful,
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
14 * Lesser General Public License for more details.
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
15 *
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
16 * You should have received a copy of the GNU Lesser General Public
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
17 * License along with FFmpeg; if not, write to the Free Software
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
19 */
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
20
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
21 #include "dsputil_mmx.h"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
22
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
23 /***********************************/
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
24 /* motion compensation */
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
25
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
26 #define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
27 "mov"#q" "#C", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
28 "mov"#d" (%0), "#F" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
29 "paddw "#D", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
30 "psllw $2, "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
31 "psubw "#B", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
32 "psubw "#E", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
33 "punpcklbw "#Z", "#F" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
34 "pmullw %4, "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
35 "paddw %5, "#A" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
36 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
37 "paddw "#F", "#A" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
38 "paddw "#A", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
39 "psraw $5, "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
40 "packuswb "#T", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
41 OP(T, (%1), A, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
42 "add %3, %1 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
43
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
44 #define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
45 "mov"#q" "#C", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
46 "mov"#d" (%0), "#F" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
47 "paddw "#D", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
48 "psllw $2, "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
49 "paddw %4, "#A" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
50 "psubw "#B", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
51 "psubw "#E", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
52 "punpcklbw "#Z", "#F" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
53 "pmullw %3, "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
54 "paddw "#F", "#A" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
55 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
56 "paddw "#A", "#T" \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
57 "mov"#q" "#T", "#OF"(%1) \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
58
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
59 #define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
60 #define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
61 #define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
62 #define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
63
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
64
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
65 #define QPEL_H264(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
66 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
67 int h=4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
68 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
69 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
70 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
71 "movq "MANGLE(ff_pw_5) ", %%mm4\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
72 "movq "MANGLE(ff_pw_16)", %%mm5\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
73 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
74 "movd -1(%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
75 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
76 "movd 1(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
77 "movd 2(%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
78 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
79 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
80 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
81 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
82 "paddw %%mm0, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
83 "paddw %%mm3, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
84 "movd -2(%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
85 "movd 3(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
86 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
87 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
88 "paddw %%mm3, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
89 "psllw $2, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
90 "psubw %%mm1, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
91 "pmullw %%mm4, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
92 "paddw %%mm5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
93 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
94 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
95 "packuswb %%mm0, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
96 OP(%%mm0, (%1),%%mm6, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
97 "add %3, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
98 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
99 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
100 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
101 : "+a"(src), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
102 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
103 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
104 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
105 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
106 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
107 int h=4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
108 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
109 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
110 "movq %0, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
111 "movq %1, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
112 :: "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
113 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
114 do{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
115 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
116 "movd -1(%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
117 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
118 "movd 1(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
119 "movd 2(%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
120 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
121 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
122 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
123 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
124 "paddw %%mm0, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
125 "paddw %%mm3, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
126 "movd -2(%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
127 "movd 3(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
128 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
129 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
130 "paddw %%mm3, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
131 "psllw $2, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
132 "psubw %%mm1, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
133 "pmullw %%mm4, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
134 "paddw %%mm5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
135 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
136 "movd (%2), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
137 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
138 "packuswb %%mm0, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
139 PAVGB" %%mm3, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
140 OP(%%mm0, (%1),%%mm6, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
141 "add %4, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
142 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
143 "add %3, %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
144 : "+a"(src), "+c"(dst), "+d"(src2)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
145 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
146 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
147 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
148 }while(--h);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
149 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
150 static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
151 src -= 2*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
152 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
153 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
154 "movd (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
155 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
156 "movd (%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
157 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
158 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
159 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
160 "movd (%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
161 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
162 "movd (%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
163 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
164 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
165 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
166 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
167 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
168 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
169 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
170 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
171 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
172 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
173 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
174 : "+a"(src), "+c"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
175 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
176 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
177 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
178 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
179 static av_noinline void OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
180 int h=4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
181 int w=3;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
182 src -= 2*srcStride+2;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
183 while(w--){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
184 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
185 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
186 "movd (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
187 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
188 "movd (%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
189 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
190 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
191 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
192 "movd (%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
193 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
194 "movd (%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
195 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
196 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
197 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
198 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
199 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
200 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
201 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
202 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
203 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
204 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
205 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
206 : "+a"(src)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
207 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
208 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
209 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
210 tmp += 4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
211 src += 4 - 9*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
212 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
213 tmp -= 3*4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
214 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
215 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
216 "movq (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
217 "paddw 10(%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
218 "movq 2(%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
219 "paddw 8(%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
220 "movq 4(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
221 "paddw 6(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
222 "psubw %%mm1, %%mm0 \n\t"/*a-b (abccba)*/\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
223 "psraw $2, %%mm0 \n\t"/*(a-b)/4 */\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
224 "psubw %%mm1, %%mm0 \n\t"/*(a-b)/4-b */\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
225 "paddsw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
226 "psraw $2, %%mm0 \n\t"/*((a-b)/4-b+c)/4 */\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
227 "paddw %%mm2, %%mm0 \n\t"/*(a-5*b+20*c)/16 */\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
228 "psraw $6, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
229 "packuswb %%mm0, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
230 OP(%%mm0, (%1),%%mm7, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
231 "add $24, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
232 "add %3, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
233 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
234 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
235 : "+a"(tmp), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
236 : "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
237 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
238 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
239 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
240 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
241 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
242 int h=8;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
243 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
244 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
245 "movq "MANGLE(ff_pw_5)", %%mm6\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
246 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
247 "movq (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
248 "movq 1(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
249 "movq %%mm0, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
250 "movq %%mm2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
251 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
252 "punpckhbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
253 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
254 "punpckhbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
255 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
256 "paddw %%mm3, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
257 "psllw $2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
258 "psllw $2, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
259 "movq -1(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
260 "movq 2(%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
261 "movq %%mm2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
262 "movq %%mm4, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
263 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
264 "punpckhbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
265 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
266 "punpckhbw %%mm7, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
267 "paddw %%mm4, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
268 "paddw %%mm3, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
269 "psubw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
270 "psubw %%mm5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
271 "pmullw %%mm6, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
272 "pmullw %%mm6, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
273 "movd -2(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
274 "movd 7(%0), %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
275 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
276 "punpcklbw %%mm7, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
277 "paddw %%mm3, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
278 "paddw %%mm5, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
279 "movq "MANGLE(ff_pw_16)", %%mm5\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
280 "paddw %%mm5, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
281 "paddw %%mm5, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
282 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
283 "paddw %%mm4, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
284 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
285 "psraw $5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
286 "packuswb %%mm1, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
287 OP(%%mm0, (%1),%%mm5, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
288 "add %3, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
289 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
290 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
291 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
292 : "+a"(src), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
293 : "d"((x86_reg)srcStride), "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
294 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
295 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
296 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
297 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
298 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
299 int h=8;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
300 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
301 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
302 "movq %0, %%mm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
303 :: "m"(ff_pw_5)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
304 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
305 do{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
306 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
307 "movq (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
308 "movq 1(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
309 "movq %%mm0, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
310 "movq %%mm2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
311 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
312 "punpckhbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
313 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
314 "punpckhbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
315 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
316 "paddw %%mm3, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
317 "psllw $2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
318 "psllw $2, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
319 "movq -1(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
320 "movq 2(%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
321 "movq %%mm2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
322 "movq %%mm4, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
323 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
324 "punpckhbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
325 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
326 "punpckhbw %%mm7, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
327 "paddw %%mm4, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
328 "paddw %%mm3, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
329 "psubw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
330 "psubw %%mm5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
331 "pmullw %%mm6, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
332 "pmullw %%mm6, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
333 "movd -2(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
334 "movd 7(%0), %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
335 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
336 "punpcklbw %%mm7, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
337 "paddw %%mm3, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
338 "paddw %%mm5, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
339 "movq %5, %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
340 "paddw %%mm5, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
341 "paddw %%mm5, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
342 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
343 "paddw %%mm4, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
344 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
345 "psraw $5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
346 "movq (%2), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
347 "packuswb %%mm1, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
348 PAVGB" %%mm4, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
349 OP(%%mm0, (%1),%%mm5, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
350 "add %4, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
351 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
352 "add %3, %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
353 : "+a"(src), "+c"(dst), "+d"(src2)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
354 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
355 "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
356 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
357 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
358 }while(--h);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
359 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
360 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
361 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
362 int w= 2;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
363 src -= 2*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
364 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
365 while(w--){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
366 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
367 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
368 "movd (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
369 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
370 "movd (%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
371 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
372 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
373 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
374 "movd (%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
375 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
376 "movd (%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
377 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
378 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
379 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
380 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
381 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
382 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
383 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
384 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
385 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
386 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
387 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
388 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
389 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
390 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
391 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
392 : "+a"(src), "+c"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
393 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
394 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
395 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
396 if(h==16){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
397 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
398 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
399 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
400 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
401 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
402 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
403 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
404 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
405 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
406 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
407 : "+a"(src), "+c"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
408 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
409 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
410 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
411 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
412 src += 4-(h+5)*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
413 dst += 4-h*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
414 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
415 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
416 static av_always_inline void OPNAME ## h264_qpel8or16_hv1_lowpass_ ## MMX(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
417 int w = (size+8)>>2;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
418 src -= 2*srcStride+2;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
419 while(w--){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
420 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
421 "pxor %%mm7, %%mm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
422 "movd (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
423 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
424 "movd (%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
425 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
426 "movd (%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
427 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
428 "movd (%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
429 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
430 "movd (%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
431 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
432 "punpcklbw %%mm7, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
433 "punpcklbw %%mm7, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
434 "punpcklbw %%mm7, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
435 "punpcklbw %%mm7, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
436 "punpcklbw %%mm7, %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
437 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
438 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
439 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
440 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
441 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
442 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
443 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
444 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
445 : "+a"(src)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
446 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
447 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
448 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
449 if(size==16){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
450 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
451 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 8*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
452 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 9*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
453 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
454 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
455 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
456 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
457 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
458 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
459 : "+a"(src)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
460 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
461 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
462 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
463 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
464 tmp += 4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
465 src += 4 - (size+5)*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
466 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
467 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
468 static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
469 int w = size>>4;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
470 do{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
471 int h = size;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
472 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
473 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
474 "movq (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
475 "movq 8(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
476 "movq 2(%0), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
477 "movq 10(%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
478 "paddw %%mm4, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
479 "paddw %%mm3, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
480 "paddw 18(%0), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
481 "paddw 16(%0), %%mm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
482 "movq 4(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
483 "movq 12(%0), %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
484 "paddw 6(%0), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
485 "paddw 14(%0), %%mm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
486 "psubw %%mm1, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
487 "psubw %%mm4, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
488 "psraw $2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
489 "psraw $2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
490 "psubw %%mm1, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
491 "psubw %%mm4, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
492 "paddsw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
493 "paddsw %%mm5, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
494 "psraw $2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
495 "psraw $2, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
496 "paddw %%mm2, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
497 "paddw %%mm5, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
498 "psraw $6, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
499 "psraw $6, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
500 "packuswb %%mm3, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
501 OP(%%mm0, (%1),%%mm7, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
502 "add $48, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
503 "add %3, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
504 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
505 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
506 : "+a"(tmp), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
507 : "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
508 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
509 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
510 tmp += 8 - size*24;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
511 dst += 8 - size*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
512 }while(w--);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
513 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
514 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
515 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
516 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
517 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
518 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
519 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
520 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
521 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
522 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
523 static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
524 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
525 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
526 src += 8*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
527 dst += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
528 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
529 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
530 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
531 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
532 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
533 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
534 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
535 src += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
536 dst += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
537 src2 += 8*src2Stride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
538 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
539 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
540 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
541 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
542 static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
543 put_h264_qpel8or16_hv1_lowpass_ ## MMX(tmp, src, tmpStride, srcStride, size);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
544 OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
545 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
546 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
547 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
548 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
549 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
550 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
551 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
552 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
553 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
554 static av_noinline void OPNAME ## pixels4_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
555 {\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
556 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
557 "movq (%1), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
558 "movq 24(%1), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
559 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
560 "psraw $5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
561 "packuswb %%mm0, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
562 "packuswb %%mm1, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
563 PAVGB" (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
564 PAVGB" (%0,%3), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
565 OP(%%mm0, (%2), %%mm4, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
566 OP(%%mm1, (%2,%4), %%mm5, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
567 "lea (%0,%3,2), %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
568 "lea (%2,%4,2), %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
569 "movq 48(%1), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
570 "movq 72(%1), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
571 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
572 "psraw $5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
573 "packuswb %%mm0, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
574 "packuswb %%mm1, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
575 PAVGB" (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
576 PAVGB" (%0,%3), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
577 OP(%%mm0, (%2), %%mm4, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
578 OP(%%mm1, (%2,%4), %%mm5, d)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
579 :"+a"(src8), "+c"(src16), "+d"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
580 :"S"((x86_reg)src8Stride), "D"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
581 :"memory");\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
582 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
583 static av_noinline void OPNAME ## pixels8_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
584 {\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
585 do{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
586 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
587 "movq (%1), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
588 "movq 8(%1), %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
589 "movq 48(%1), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
590 "movq 8+48(%1), %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
591 "psraw $5, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
592 "psraw $5, %%mm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
593 "psraw $5, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
594 "psraw $5, %%mm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
595 "packuswb %%mm1, %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
596 "packuswb %%mm3, %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
597 PAVGB" (%0), %%mm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
598 PAVGB" (%0,%3), %%mm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
599 OP(%%mm0, (%2), %%mm5, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
600 OP(%%mm2, (%2,%4), %%mm5, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
601 ::"a"(src8), "c"(src16), "d"(dst),\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
602 "r"((x86_reg)src8Stride), "r"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
603 :"memory");\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
604 src8 += 2L*src8Stride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
605 src16 += 48;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
606 dst += 2L*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
607 }while(h-=2);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
608 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
609 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
610 {\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
611 OPNAME ## pixels8_l2_shift5_ ## MMX(dst , src16 , src8 , dstStride, src8Stride, h);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
612 OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
613 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
614
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
615
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
616 #if ARCH_X86_64
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
617 #define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
618 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
619 int h=16;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
620 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
621 "pxor %%xmm15, %%xmm15 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
622 "movdqa %6, %%xmm14 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
623 "movdqa %7, %%xmm13 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
624 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
625 "lddqu 6(%0), %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
626 "lddqu -2(%0), %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
627 "movdqa %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
628 "punpckhbw %%xmm15, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
629 "punpcklbw %%xmm15, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
630 "punpcklbw %%xmm15, %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
631 "movdqa %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
632 "movdqa %%xmm0, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
633 "movdqa %%xmm1, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
634 "movdqa %%xmm0, %%xmm8 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
635 "movdqa %%xmm1, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
636 "movdqa %%xmm0, %%xmm9 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
637 "movdqa %%xmm0, %%xmm12 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
638 "movdqa %%xmm1, %%xmm11 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
639 "palignr $10,%%xmm0, %%xmm11\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
640 "palignr $10,%%xmm7, %%xmm12\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
641 "palignr $2, %%xmm0, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
642 "palignr $2, %%xmm7, %%xmm9 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
643 "palignr $4, %%xmm0, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
644 "palignr $4, %%xmm7, %%xmm8 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
645 "palignr $6, %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
646 "palignr $6, %%xmm7, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
647 "paddw %%xmm0 ,%%xmm11 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
648 "palignr $8, %%xmm0, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
649 "palignr $8, %%xmm7, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
650 "paddw %%xmm12,%%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
651 "paddw %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
652 "paddw %%xmm8, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
653 "paddw %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
654 "paddw %%xmm9, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
655 "psllw $2, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
656 "psllw $2, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
657 "psubw %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
658 "psubw %%xmm0, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
659 "paddw %%xmm13,%%xmm11 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
660 "paddw %%xmm13,%%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
661 "pmullw %%xmm14,%%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
662 "pmullw %%xmm14,%%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
663 "lddqu (%2), %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
664 "paddw %%xmm11,%%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
665 "paddw %%xmm7, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
666 "psraw $5, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
667 "psraw $5, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
668 "packuswb %%xmm2,%%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
669 "pavgb %%xmm3, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
670 OP(%%xmm6, (%1), %%xmm4, dqa)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
671 "add %5, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
672 "add %5, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
673 "add %4, %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
674 "decl %3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
675 "jg 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
676 : "+a"(src), "+c"(dst), "+d"(src2), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
677 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
678 "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
679 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
680 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
681 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
682 #else // ARCH_X86_64
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
683 #define QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
684 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
685 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
686 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
687 src += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
688 dst += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
689 src2 += 8*src2Stride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
690 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
691 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
692 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
693 #endif // ARCH_X86_64
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
694
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
695 #define QPEL_H264_H_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
696 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
697 int h=8;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
698 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
699 "pxor %%xmm7, %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
700 "movdqa %0, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
701 :: "m"(ff_pw_5)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
702 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
703 do{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
704 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
705 "lddqu -2(%0), %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
706 "movdqa %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
707 "punpckhbw %%xmm7, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
708 "punpcklbw %%xmm7, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
709 "movdqa %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
710 "movdqa %%xmm1, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
711 "movdqa %%xmm1, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
712 "movdqa %%xmm1, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
713 "palignr $2, %%xmm0, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
714 "palignr $4, %%xmm0, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
715 "palignr $6, %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
716 "palignr $8, %%xmm0, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
717 "palignr $10,%%xmm0, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
718 "paddw %%xmm5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
719 "paddw %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
720 "paddw %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
721 "psllw $2, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
722 "movq (%2), %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
723 "psubw %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
724 "paddw %5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
725 "pmullw %%xmm6, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
726 "paddw %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
727 "psraw $5, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
728 "packuswb %%xmm2, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
729 "pavgb %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
730 OP(%%xmm2, (%1), %%xmm4, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
731 "add %4, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
732 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
733 "add %3, %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
734 : "+a"(src), "+c"(dst), "+d"(src2)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
735 : "D"((x86_reg)src2Stride), "S"((x86_reg)dstStride),\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
736 "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
737 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
738 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
739 }while(--h);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
740 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
741 QPEL_H264_H16_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
742 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
743 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
744 int h=8;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
745 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
746 "pxor %%xmm7, %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
747 "movdqa "MANGLE(ff_pw_5)", %%xmm6\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
748 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
749 "lddqu -2(%0), %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
750 "movdqa %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
751 "punpckhbw %%xmm7, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
752 "punpcklbw %%xmm7, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
753 "movdqa %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
754 "movdqa %%xmm1, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
755 "movdqa %%xmm1, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
756 "movdqa %%xmm1, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
757 "palignr $2, %%xmm0, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
758 "palignr $4, %%xmm0, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
759 "palignr $6, %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
760 "palignr $8, %%xmm0, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
761 "palignr $10,%%xmm0, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
762 "paddw %%xmm5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
763 "paddw %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
764 "paddw %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
765 "psllw $2, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
766 "psubw %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
767 "paddw "MANGLE(ff_pw_16)", %%xmm0\n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
768 "pmullw %%xmm6, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
769 "paddw %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
770 "psraw $5, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
771 "packuswb %%xmm2, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
772 OP(%%xmm2, (%1), %%xmm4, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
773 "add %3, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
774 "add %4, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
775 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
776 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
777 : "+a"(src), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
778 : "D"((x86_reg)srcStride), "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
779 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
780 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
781 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
782 static void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
783 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
784 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
785 src += 8*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
786 dst += 8*dstStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
787 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
788 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
789 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
790
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
791 #define QPEL_H264_V_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
792 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
793 src -= 2*srcStride;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
794 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
795 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
796 "pxor %%xmm7, %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
797 "movq (%0), %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
798 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
799 "movq (%0), %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
800 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
801 "movq (%0), %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
802 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
803 "movq (%0), %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
804 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
805 "movq (%0), %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
806 "add %2, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
807 "punpcklbw %%xmm7, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
808 "punpcklbw %%xmm7, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
809 "punpcklbw %%xmm7, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
810 "punpcklbw %%xmm7, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
811 "punpcklbw %%xmm7, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
812 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
813 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
814 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
815 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
816 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
817 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
818 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
819 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
820 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
821 : "+a"(src), "+c"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
822 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
823 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
824 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
825 if(h==16){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
826 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
827 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
828 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
829 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
830 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
831 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
832 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
833 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
834 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
835 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
836 : "+a"(src), "+c"(dst)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
837 : "S"((x86_reg)srcStride), "D"((x86_reg)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
838 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
839 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
840 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
841 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
842 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
843 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
844 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
845 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
846 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
847 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
848 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
849
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
850 static av_always_inline void put_h264_qpel8or16_hv1_lowpass_sse2(int16_t *tmp, uint8_t *src, int tmpStride, int srcStride, int size){
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
851 int w = (size+8)>>3;
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
852 src -= 2*srcStride+2;
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
853 while(w--){
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
854 __asm__ volatile(
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
855 "pxor %%xmm7, %%xmm7 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
856 "movq (%0), %%xmm0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
857 "add %2, %0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
858 "movq (%0), %%xmm1 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
859 "add %2, %0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
860 "movq (%0), %%xmm2 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
861 "add %2, %0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
862 "movq (%0), %%xmm3 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
863 "add %2, %0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
864 "movq (%0), %%xmm4 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
865 "add %2, %0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
866 "punpcklbw %%xmm7, %%xmm0 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
867 "punpcklbw %%xmm7, %%xmm1 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
868 "punpcklbw %%xmm7, %%xmm2 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
869 "punpcklbw %%xmm7, %%xmm3 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
870 "punpcklbw %%xmm7, %%xmm4 \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
871 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 0*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
872 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 1*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
873 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 2*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
874 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 3*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
875 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 4*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
876 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 5*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
877 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 6*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
878 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 7*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
879 : "+a"(src)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
880 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
881 : "memory"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
882 );
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
883 if(size==16){
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
884 __asm__ volatile(
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
885 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 8*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
886 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 9*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
887 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 10*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
888 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 11*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
889 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 12*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
890 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 13*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
891 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 14*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
892 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 15*48)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
893 : "+a"(src)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
894 : "c"(tmp), "S"((x86_reg)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
895 : "memory"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
896 );
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
897 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
898 tmp += 8;
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
899 src += 8 - (size+5)*srcStride;
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
900 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
901 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
902
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
903 #define QPEL_H264_HV2_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
904 static av_always_inline void OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, int dstStride, int tmpStride, int size){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
905 int h = size;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
906 if(size == 16){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
907 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
908 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
909 "movdqa 32(%0), %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
910 "movdqa 16(%0), %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
911 "movdqa (%0), %%xmm7 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
912 "movdqa %%xmm4, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
913 "movdqa %%xmm4, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
914 "movdqa %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
915 "movdqa %%xmm4, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
916 "palignr $10, %%xmm5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
917 "palignr $8, %%xmm5, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
918 "palignr $6, %%xmm5, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
919 "palignr $4, %%xmm5, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
920 "palignr $2, %%xmm5, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
921 "paddw %%xmm5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
922 "paddw %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
923 "paddw %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
924 "movdqa %%xmm5, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
925 "movdqa %%xmm5, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
926 "movdqa %%xmm5, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
927 "palignr $8, %%xmm7, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
928 "palignr $2, %%xmm7, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
929 "palignr $10, %%xmm7, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
930 "paddw %%xmm6, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
931 "movdqa %%xmm5, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
932 "palignr $6, %%xmm7, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
933 "palignr $4, %%xmm7, %%xmm6 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
934 "paddw %%xmm7, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
935 "paddw %%xmm6, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
936 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
937 "psubw %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
938 "psubw %%xmm4, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
939 "psraw $2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
940 "psraw $2, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
941 "psubw %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
942 "psubw %%xmm4, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
943 "paddw %%xmm2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
944 "paddw %%xmm5, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
945 "psraw $2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
946 "psraw $2, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
947 "paddw %%xmm2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
948 "paddw %%xmm5, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
949 "psraw $6, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
950 "psraw $6, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
951 "packuswb %%xmm0, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
952 OP(%%xmm3, (%1), %%xmm7, dqa)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
953 "add $48, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
954 "add %3, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
955 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
956 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
957 : "+a"(tmp), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
958 : "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
959 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
960 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
961 }else{\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
962 __asm__ volatile(\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
963 "1: \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
964 "movdqa 16(%0), %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
965 "movdqa (%0), %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
966 "movdqa %%xmm1, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
967 "movdqa %%xmm1, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
968 "movdqa %%xmm1, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
969 "movdqa %%xmm1, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
970 "palignr $10, %%xmm0, %%xmm5 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
971 "palignr $8, %%xmm0, %%xmm4 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
972 "palignr $6, %%xmm0, %%xmm3 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
973 "palignr $4, %%xmm0, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
974 "palignr $2, %%xmm0, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
975 "paddw %%xmm5, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
976 "paddw %%xmm4, %%xmm1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
977 "paddw %%xmm3, %%xmm2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
978 "psubw %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
979 "psraw $2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
980 "psubw %%xmm1, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
981 "paddw %%xmm2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
982 "psraw $2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
983 "paddw %%xmm2, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
984 "psraw $6, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
985 "packuswb %%xmm0, %%xmm0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
986 OP(%%xmm0, (%1), %%xmm7, q)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
987 "add $48, %0 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
988 "add %3, %1 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
989 "decl %2 \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
990 " jnz 1b \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
991 : "+a"(tmp), "+c"(dst), "+g"(h)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
992 : "S"((x86_reg)dstStride)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
993 : "memory"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
994 );\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
995 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
996 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
997
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
998 #define QPEL_H264_HV_XMM(OPNAME, OP, MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
999 static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1000 put_h264_qpel8or16_hv1_lowpass_sse2(tmp, src, tmpStride, srcStride, size);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1001 OPNAME ## h264_qpel8or16_hv2_lowpass_ ## MMX(dst, tmp, dstStride, tmpStride, size);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1002 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1003 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1004 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 8);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1005 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1006 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1007 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst, tmp, src, dstStride, tmpStride, srcStride, 16);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1008 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1009
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1010 #define put_pixels8_l2_sse2 put_pixels8_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1011 #define avg_pixels8_l2_sse2 avg_pixels8_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1012 #define put_pixels16_l2_sse2 put_pixels16_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1013 #define avg_pixels16_l2_sse2 avg_pixels16_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1014 #define put_pixels8_l2_ssse3 put_pixels8_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1015 #define avg_pixels8_l2_ssse3 avg_pixels8_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1016 #define put_pixels16_l2_ssse3 put_pixels16_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1017 #define avg_pixels16_l2_ssse3 avg_pixels16_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1018
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1019 #define put_pixels8_l2_shift5_sse2 put_pixels8_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1020 #define avg_pixels8_l2_shift5_sse2 avg_pixels8_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1021 #define put_pixels16_l2_shift5_sse2 put_pixels16_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1022 #define avg_pixels16_l2_shift5_sse2 avg_pixels16_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1023 #define put_pixels8_l2_shift5_ssse3 put_pixels8_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1024 #define avg_pixels8_l2_shift5_ssse3 avg_pixels8_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1025 #define put_pixels16_l2_shift5_ssse3 put_pixels16_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1026 #define avg_pixels16_l2_shift5_ssse3 avg_pixels16_l2_shift5_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1027
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1028 #define put_h264_qpel8_h_lowpass_l2_sse2 put_h264_qpel8_h_lowpass_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1029 #define avg_h264_qpel8_h_lowpass_l2_sse2 avg_h264_qpel8_h_lowpass_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1030 #define put_h264_qpel16_h_lowpass_l2_sse2 put_h264_qpel16_h_lowpass_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1031 #define avg_h264_qpel16_h_lowpass_l2_sse2 avg_h264_qpel16_h_lowpass_l2_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1032
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1033 #define put_h264_qpel8_v_lowpass_ssse3 put_h264_qpel8_v_lowpass_sse2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1034 #define avg_h264_qpel8_v_lowpass_ssse3 avg_h264_qpel8_v_lowpass_sse2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1035 #define put_h264_qpel16_v_lowpass_ssse3 put_h264_qpel16_v_lowpass_sse2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1036 #define avg_h264_qpel16_v_lowpass_ssse3 avg_h264_qpel16_v_lowpass_sse2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1037
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1038 #define put_h264_qpel8or16_hv2_lowpass_sse2 put_h264_qpel8or16_hv2_lowpass_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1039 #define avg_h264_qpel8or16_hv2_lowpass_sse2 avg_h264_qpel8or16_hv2_lowpass_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1040
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1041 #define H264_MC(OPNAME, SIZE, MMX, ALIGN) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1042 H264_MC_C(OPNAME, SIZE, MMX, ALIGN)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1043 H264_MC_V(OPNAME, SIZE, MMX, ALIGN)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1044 H264_MC_H(OPNAME, SIZE, MMX, ALIGN)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1045 H264_MC_HV(OPNAME, SIZE, MMX, ALIGN)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1046
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1047 static void put_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1048 put_pixels16_sse2(dst, src, stride, 16);
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1049 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1050 static void avg_h264_qpel16_mc00_sse2 (uint8_t *dst, uint8_t *src, int stride){
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1051 avg_pixels16_sse2(dst, src, stride, 16);
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1052 }
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1053 #define put_h264_qpel8_mc00_sse2 put_h264_qpel8_mc00_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1054 #define avg_h264_qpel8_mc00_sse2 avg_h264_qpel8_mc00_mmx2
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1055
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1056 #define H264_MC_C(OPNAME, SIZE, MMX, ALIGN) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1057 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1058 OPNAME ## pixels ## SIZE ## _ ## MMX(dst, src, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1059 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1060
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1061 #define H264_MC_H(OPNAME, SIZE, MMX, ALIGN) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1062 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1063 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1064 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1065 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1066 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1067 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1068 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1069 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1070 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1071 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1072 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1073
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1074 #define H264_MC_V(OPNAME, SIZE, MMX, ALIGN) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1075 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1076 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1077 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1078 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, temp, stride, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1079 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1080 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1081 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1082 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1083 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1084 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1085 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1086 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1087 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1088 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, temp, stride, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1089 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1090
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1091 #define H264_MC_HV(OPNAME, SIZE, MMX, ALIGN) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1092 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1093 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1094 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1095 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1096 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1097 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1098 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1099 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1100 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1101 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, temp, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1102 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1103 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1104 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1105 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1106 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1107 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1108 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1109 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1110 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1111 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1112 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(temp, src+1, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1113 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, temp, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1114 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1115 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1116 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1117 DECLARE_ALIGNED(ALIGN, uint16_t, temp)[SIZE*(SIZE<8?12:24)];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1118 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, temp, src, stride, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1119 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1120 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1121 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1122 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1123 uint8_t * const halfHV= temp;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1124 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1125 assert(((int)temp & 7) == 0);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1126 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1127 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1128 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1129 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1130 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1131 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1132 uint8_t * const halfHV= temp;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1133 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1134 assert(((int)temp & 7) == 0);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1135 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1136 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1137 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1138 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1139 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1140 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1141 uint8_t * const halfHV= temp;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1142 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1143 assert(((int)temp & 7) == 0);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1144 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1145 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1146 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1147 \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1148 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1149 DECLARE_ALIGNED(ALIGN, uint8_t, temp)[SIZE*(SIZE<8?12:24)*2 + SIZE*SIZE];\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1150 uint8_t * const halfHV= temp;\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1151 int16_t * const halfV= (int16_t*)(temp + SIZE*SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1152 assert(((int)temp & 7) == 0);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1153 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1154 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1155 }\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1156
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1157 #define H264_MC_4816(MMX)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1158 H264_MC(put_, 4, MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1159 H264_MC(put_, 8, MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1160 H264_MC(put_, 16,MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1161 H264_MC(avg_, 4, MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1162 H264_MC(avg_, 8, MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1163 H264_MC(avg_, 16,MMX, 8)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1164
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1165 #define H264_MC_816(QPEL, XMM)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1166 QPEL(put_, 8, XMM, 16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1167 QPEL(put_, 16,XMM, 16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1168 QPEL(avg_, 8, XMM, 16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1169 QPEL(avg_, 16,XMM, 16)\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1170
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1171
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1172 #define AVG_3DNOW_OP(a,b,temp, size) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1173 "mov" #size " " #b ", " #temp " \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1174 "pavgusb " #temp ", " #a " \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1175 "mov" #size " " #a ", " #b " \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1176 #define AVG_MMX2_OP(a,b,temp, size) \
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1177 "mov" #size " " #b ", " #temp " \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1178 "pavgb " #temp ", " #a " \n\t"\
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1179 "mov" #size " " #a ", " #b " \n\t"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1180
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1181 #define PAVGB "pavgusb"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1182 QPEL_H264(put_, PUT_OP, 3dnow)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1183 QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1184 #undef PAVGB
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1185 #define PAVGB "pavgb"
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1186 QPEL_H264(put_, PUT_OP, mmx2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1187 QPEL_H264(avg_, AVG_MMX2_OP, mmx2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1188 QPEL_H264_V_XMM(put_, PUT_OP, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1189 QPEL_H264_V_XMM(avg_, AVG_MMX2_OP, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1190 QPEL_H264_HV_XMM(put_, PUT_OP, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1191 QPEL_H264_HV_XMM(avg_, AVG_MMX2_OP, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1192 #if HAVE_SSSE3
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1193 QPEL_H264_H_XMM(put_, PUT_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1194 QPEL_H264_H_XMM(avg_, AVG_MMX2_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1195 QPEL_H264_HV2_XMM(put_, PUT_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1196 QPEL_H264_HV2_XMM(avg_, AVG_MMX2_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1197 QPEL_H264_HV_XMM(put_, PUT_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1198 QPEL_H264_HV_XMM(avg_, AVG_MMX2_OP, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1199 #endif
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1200 #undef PAVGB
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1201
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1202 H264_MC_4816(3dnow)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1203 H264_MC_4816(mmx2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1204 H264_MC_816(H264_MC_V, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1205 H264_MC_816(H264_MC_HV, sse2)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1206 #if HAVE_SSSE3
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1207 H264_MC_816(H264_MC_H, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1208 H264_MC_816(H264_MC_HV, ssse3)
3941687b4fa9 Split h264dsp_mmx.c (which was #included in dsputil_mmx.c) in h264_qpel_mmx.c,
rbultje
parents:
diff changeset
1209 #endif