lun_admin/android_multiviewwer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

...

Browse Source
This commit is contained in:
lun_admin
2025-08-27 15:09:05 +09:00
parent 57961d84de
commit a13f66d917
9896 changed files with 2193048 additions and 730 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ffmpeg/libavcodec/ppc/Makefile
+28
View File
@@ -0,0 +1,28 @@
# subsystems
OBJS-$(CONFIG_AUDIODSP) += ppc/audiodsp.o
OBJS-$(CONFIG_BLOCKDSP) += ppc/blockdsp.o
OBJS-$(CONFIG_FDCTDSP) += ppc/fdctdsp.o
OBJS-$(CONFIG_FMTCONVERT) += ppc/fmtconvert_altivec.o
OBJS-$(CONFIG_H264CHROMA) += ppc/h264chroma_init.o
OBJS-$(CONFIG_H264DSP) += ppc/h264dsp.o ppc/hpeldsp_altivec.o
OBJS-$(CONFIG_H264QPEL) += ppc/h264qpel.o
OBJS-$(CONFIG_HPELDSP) += ppc/hpeldsp_altivec.o
OBJS-$(CONFIG_IDCTDSP) += ppc/idctdsp.o
OBJS-$(CONFIG_LLVIDDSP) += ppc/lossless_videodsp_altivec.o
OBJS-$(CONFIG_ME_CMP) += ppc/me_cmp.o
OBJS-$(CONFIG_MPEGAUDIODSP) += ppc/mpegaudiodsp_altivec.o
OBJS-$(CONFIG_MPEGVIDEO) += ppc/mpegvideo_altivec.o
OBJS-$(CONFIG_MPEGVIDEOENCDSP) += ppc/mpegvideoencdsp.o
OBJS-$(CONFIG_PIXBLOCKDSP) += ppc/pixblockdsp.o
OBJS-$(CONFIG_VC1DSP) += ppc/vc1dsp_altivec.o
OBJS-$(CONFIG_VIDEODSP) += ppc/videodsp.o
OBJS-$(CONFIG_VP3DSP) += ppc/vp3dsp_altivec.o
OBJS-$(CONFIG_VP8DSP) += ppc/vp8dsp_altivec.o
# decoders/encoders
OBJS-$(CONFIG_HEVC_DECODER) += ppc/hevcdsp.o
OBJS-$(CONFIG_LLAUDDSP) += ppc/lossless_audiodsp_altivec.o
OBJS-$(CONFIG_MPEG4_DECODER) += ppc/mpeg4videodsp.o
OBJS-$(CONFIG_SVQ1_ENCODER) += ppc/svq1enc_altivec.o
OBJS-$(CONFIG_VORBIS_DECODER) += ppc/vorbisdsp_altivec.o
OBJS-$(CONFIG_VP7_DECODER) += ppc/vp8dsp_altivec.o
ffmpeg/libavcodec/ppc/asm.S
+147
View File
@@ -0,0 +1,147 @@
/*
* Copyright (c) 2009 Loren Merritt
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#define GLUE(a, b) a ## b
#define JOIN(a, b) GLUE(a, b)
#define X(s) JOIN(EXTERN_ASM, s)
#if __APPLE__
#define R(n) r ## n
#else
#define R(n) n
#endif
#if ARCH_PPC64
#define PTR .quad
#define lp ld
#define lpx ldx
#define stp std
#define stpu stdu
#define PS 8
#define L(s) JOIN(., s)
.macro extfunc name
.global X(\name)
#if _CALL_ELF == 2
.text
X(\name):
addis %r2, %r12, .TOC.-X(\name)@ha
addi %r2, %r2, .TOC.-X(\name)@l
.localentry X(\name), .-X(\name)
#else
.section .opd, "aw"
X(\name):
.quad L(\name), .TOC.@tocbase, 0
.previous
.type X(\name), STT_FUNC
L(\name):
#endif
.endm
.macro movrel rd, sym, gp
ld \rd, \sym@got(R(2))
.endm
.macro get_got rd
.endm
#else /* ARCH_PPC64 */
#define PTR .int
#define lp lwz
#define lpx lwzx
#define stp stw
#define stpu stwu
#define PS 4
#define L(s) s
.macro extfunc name
.global X(\name)
.type X(\name), STT_FUNC
X(\name):
\name:
.endm
.macro movrel rd, sym, gp
#if CONFIG_PIC
lwz \rd, \sym@got(\gp)
#else
lis \rd, \sym@ha
la \rd, \sym@l(\rd)
#endif
.endm
.macro get_got rd
#if CONFIG_PIC
bcl 20, 31, .Lgot\@
.Lgot\@:
mflr \rd
addis \rd, \rd, _GLOBAL_OFFSET_TABLE_ - .Lgot\@@ha
addi \rd, \rd, _GLOBAL_OFFSET_TABLE_ - .Lgot\@@l
#endif
.endm
#endif /* ARCH_PPC64 */
#if HAVE_IBM_ASM
.macro DEFINE_REG n
.equiv r\n, \n
.equiv f\n, \n
.equiv v\n, \n
.endm
DEFINE_REG 0
DEFINE_REG 1
DEFINE_REG 2
DEFINE_REG 3
DEFINE_REG 4
DEFINE_REG 5
DEFINE_REG 6
DEFINE_REG 7
DEFINE_REG 8
DEFINE_REG 9
DEFINE_REG 10
DEFINE_REG 11
DEFINE_REG 12
DEFINE_REG 13
DEFINE_REG 14
DEFINE_REG 15
DEFINE_REG 16
DEFINE_REG 17
DEFINE_REG 18
DEFINE_REG 19
DEFINE_REG 20
DEFINE_REG 21
DEFINE_REG 22
DEFINE_REG 23
DEFINE_REG 24
DEFINE_REG 25
DEFINE_REG 26
DEFINE_REG 27
DEFINE_REG 28
DEFINE_REG 29
DEFINE_REG 30
DEFINE_REG 31
#endif /* HAVE_IBM_ASM */
ffmpeg/libavcodec/ppc/audiodsp.c
+101
View File
@@ -0,0 +1,101 @@
/*
* Copyright (c) 2007 Luca Barbato <lu_zero@gentoo.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* miscellaneous audio operations
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/audiodsp.h"
#if HAVE_ALTIVEC
static int32_t scalarproduct_int16_altivec(const int16_t *v1, const int16_t *v2,
int order)
{
int i;
LOAD_ZERO;
register vec_s16 vec1;
register vec_s32 res = vec_splat_s32(0), t;
int32_t ires;
for (i = 0; i < order; i += 8) {
vec1 = vec_unaligned_load(v1);
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
res = vec_sums(t, res);
v1 += 8;
v2 += 8;
}
res = vec_splat(res, 3);
vec_ste(res, 0, &ires);
return ires;
}
#endif /* HAVE_ALTIVEC */
#if HAVE_VSX
static int32_t scalarproduct_int16_vsx(const int16_t *v1, const int16_t *v2, int order)
{
int i;
LOAD_ZERO;
register vec_s16 vec1;
register vec_s32 res = vec_splat_s32(0), t;
int32_t ires;
for (i = 0; i < order; i += 8) {
vec1 = vec_vsx_ld(0, v1);
t = vec_msum(vec1, vec_ld(0, v2), zero_s32v);
res = vec_sums(t, res);
v1 += 8;
v2 += 8;
}
res = vec_splat(res, 3);
vec_ste(res, 0, &ires);
return ires;
}
#endif /* HAVE_VSX */
av_cold void ff_audiodsp_init_ppc(AudioDSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->scalarproduct_int16 = scalarproduct_int16_altivec;
#endif /* HAVE_ALTIVEC */
#if HAVE_VSX
if (!PPC_VSX(av_get_cpu_flags()))
return;
c->scalarproduct_int16 = scalarproduct_int16_vsx;
#endif /* HAVE_VSX */
}
ffmpeg/libavcodec/ppc/blockdsp.c
+165
View File
@@ -0,0 +1,165 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include <string.h>
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/blockdsp.h"
/* ***** WARNING ***** WARNING ***** WARNING ***** */
/*
* clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with
* a cache line size not equal to 32 bytes. Fortunately all processors used
* by Apple up to at least the 7450 (AKA second generation G4) use 32-byte
* cache lines. This is due to the use of the 'dcbz' instruction. It simply
* clears a single cache line to zero, so you need to know the cache line
* size to use it! It's absurd, but it's fast...
*
* update 24/06/2003: Apple released the G5 yesterday, with a PPC970.
* cache line size: 128 bytes. Oups.
* The semantics of dcbz was changed, it always clears 32 bytes. So the function
* below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
* which is defined to clear a cache line (as dcbz before). So we can still
* distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
*
* see <http://developer.apple.com/technotes/tn/tn2087.html>
* and <http://developer.apple.com/technotes/tn/tn2086.html>
*/
static void clear_blocks_dcbz32_ppc(int16_t *blocks)
{
register int misal = (unsigned long) blocks & 0x00000010, i = 0;
if (misal) {
((unsigned long *) blocks)[0] = 0L;
((unsigned long *) blocks)[1] = 0L;
((unsigned long *) blocks)[2] = 0L;
((unsigned long *) blocks)[3] = 0L;
i += 16;
}
for (; i < sizeof(int16_t) * 6 * 64 - 31; i += 32)
__asm__ volatile ("dcbz %0,%1" :: "b" (blocks), "r" (i) : "memory");
if (misal) {
((unsigned long *) blocks)[188] = 0L;
((unsigned long *) blocks)[189] = 0L;
((unsigned long *) blocks)[190] = 0L;
((unsigned long *) blocks)[191] = 0L;
i += 16;
}
}
/* Same as above, when dcbzl clears a whole 128 bytes cache line
* i.e. the PPC970 AKA G5. */
static void clear_blocks_dcbz128_ppc(int16_t *blocks)
{
#if HAVE_DCBZL
register int misal = (unsigned long) blocks & 0x0000007f, i = 0;
if (misal) {
/* We could probably also optimize this case,
* but there's not much point as the machines
* aren't available yet (2003-06-26). */
memset(blocks, 0, sizeof(int16_t) * 6 * 64);
} else {
for (; i < sizeof(int16_t) * 6 * 64; i += 128)
__asm__ volatile ("dcbzl %0,%1" :: "b" (blocks), "r" (i) : "memory");
}
#else
memset(blocks, 0, sizeof(int16_t) * 6 * 64);
#endif
}
/* Check dcbz report how many bytes are set to 0 by dcbz. */
/* update 24/06/2003: Replace dcbz by dcbzl to get the intended effect
* (Apple "fixed" dcbz). Unfortunately this cannot be used unless the
* assembler knows about dcbzl ... */
static long check_dcbzl_effect(void)
{
long count = 0;
#if HAVE_DCBZL
register char *fakedata = av_malloc(1024);
register char *fakedata_middle;
register long zero = 0, i = 0;
if (!fakedata)
return 0L;
fakedata_middle = fakedata + 512;
memset(fakedata, 0xFF, 1024);
/* Below the constraint "b" seems to mean "address base register"
* in gcc-3.3 / RS/6000 speaks. Seems to avoid using r0, so.... */
__asm__ volatile ("dcbzl %0, %1" :: "b" (fakedata_middle), "r" (zero));
for (i = 0; i < 1024; i++)
if (fakedata[i] == (char) 0)
count++;
av_free(fakedata);
#endif
return count;
}
#if HAVE_ALTIVEC
static void clear_block_altivec(int16_t *block)
{
LOAD_ZERO;
vec_st(zero_s16v, 0, block);
vec_st(zero_s16v, 16, block);
vec_st(zero_s16v, 32, block);
vec_st(zero_s16v, 48, block);
vec_st(zero_s16v, 64, block);
vec_st(zero_s16v, 80, block);
vec_st(zero_s16v, 96, block);
vec_st(zero_s16v, 112, block);
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_blockdsp_init_ppc(BlockDSPContext *c)
{
// common optimizations whether AltiVec is available or not
switch (check_dcbzl_effect()) {
case 32:
c->clear_blocks = clear_blocks_dcbz32_ppc;
break;
case 128:
c->clear_blocks = clear_blocks_dcbz128_ppc;
break;
default:
break;
}
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->clear_block = clear_block_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/fdct.h
+26
View File
@@ -0,0 +1,26 @@
/*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVCODEC_PPC_FDCT_H
#define AVCODEC_PPC_FDCT_H
#include <stdint.h>
void ff_fdct_altivec(int16_t *block);
#endif /* AVCODEC_PPC_FDCT_H */
ffmpeg/libavcodec/ppc/fdctdsp.c
+479
View File
@@ -0,0 +1,479 @@
/*
* Copyright (C) 2003 James Klicman <james@klicman.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/fdctdsp.h"
#include "fdct.h"
#if HAVE_ALTIVEC
#define vs16(v) ((vector signed short) (v))
#define vs32(v) ((vector signed int) (v))
#define vu8(v) ((vector unsigned char) (v))
#define vu16(v) ((vector unsigned short) (v))
#define vu32(v) ((vector unsigned int) (v))
#define C1 0.98078528040323044912618224 /* cos(1 * PI / 16) */
#define C2 0.92387953251128675612818319 /* cos(2 * PI / 16) */
#define C3 0.83146961230254523707878838 /* cos(3 * PI / 16) */
#define C4 0.70710678118654752440084436 /* cos(4 * PI / 16) */
#define C5 0.55557023301960222474283081 /* cos(5 * PI / 16) */
#define C6 0.38268343236508977172845998 /* cos(6 * PI / 16) */
#define C7 0.19509032201612826784828487 /* cos(7 * PI / 16) */
#define W0 -(2 * C2)
#define W1 (2 * C6)
#define W2 (M_SQRT2 * C6)
#define W3 (M_SQRT2 * C3)
#define W4 (M_SQRT2 * (-C1 + C3 + C5 - C7))
#define W5 (M_SQRT2 * (C1 + C3 - C5 + C7))
#define W6 (M_SQRT2 * (C1 + C3 + C5 - C7))
#define W7 (M_SQRT2 * (C1 + C3 - C5 - C7))
#define W8 (M_SQRT2 * (C7 - C3))
#define W9 (M_SQRT2 * (-C1 - C3))
#define WA (M_SQRT2 * (-C3 - C5))
#define WB (M_SQRT2 * (C5 - C3))
static const vector float fdctconsts[3] = {
{ W0, W1, W2, W3 },
{ W4, W5, W6, W7 },
{ W8, W9, WA, WB }
};
#define LD_W0 vec_splat(cnsts0, 0)
#define LD_W1 vec_splat(cnsts0, 1)
#define LD_W2 vec_splat(cnsts0, 2)
#define LD_W3 vec_splat(cnsts0, 3)
#define LD_W4 vec_splat(cnsts1, 0)
#define LD_W5 vec_splat(cnsts1, 1)
#define LD_W6 vec_splat(cnsts1, 2)
#define LD_W7 vec_splat(cnsts1, 3)
#define LD_W8 vec_splat(cnsts2, 0)
#define LD_W9 vec_splat(cnsts2, 1)
#define LD_WA vec_splat(cnsts2, 2)
#define LD_WB vec_splat(cnsts2, 3)
#define FDCTROW(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
\
b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
\
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
cnst = LD_W2; \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
cnst = LD_W1; \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
cnst = LD_W0; \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
\
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
cnst = LD_W3; \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
\
cnst = LD_W8; \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
cnst = LD_W9; \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
cnst = LD_WA; \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
cnst = LD_WB; \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
\
cnst = LD_W4; \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
cnst = LD_W5; \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
cnst = LD_W6; \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
cnst = LD_W7; \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
\
b7 = vec_add(b7, x2); /* b7 = b7 + x2; */ \
b5 = vec_add(b5, x3); /* b5 = b5 + x3; */ \
b3 = vec_add(b3, x2); /* b3 = b3 + x2; */ \
b1 = vec_add(b1, x3) /* b1 = b1 + x3; */ \
/* }}} */
#define FDCTCOL(b0, b1, b2, b3, b4, b5, b6, b7) /* {{{ */ \
x0 = vec_add(b0, b7); /* x0 = b0 + b7; */ \
x7 = vec_sub(b0, b7); /* x7 = b0 - b7; */ \
x1 = vec_add(b1, b6); /* x1 = b1 + b6; */ \
x6 = vec_sub(b1, b6); /* x6 = b1 - b6; */ \
x2 = vec_add(b2, b5); /* x2 = b2 + b5; */ \
x5 = vec_sub(b2, b5); /* x5 = b2 - b5; */ \
x3 = vec_add(b3, b4); /* x3 = b3 + b4; */ \
x4 = vec_sub(b3, b4); /* x4 = b3 - b4; */ \
\
b7 = vec_add(x0, x3); /* b7 = x0 + x3; */ \
b1 = vec_add(x1, x2); /* b1 = x1 + x2; */ \
b0 = vec_add(b7, b1); /* b0 = b7 + b1; */ \
b4 = vec_sub(b7, b1); /* b4 = b7 - b1; */ \
\
b2 = vec_sub(x0, x3); /* b2 = x0 - x3; */ \
b6 = vec_sub(x1, x2); /* b6 = x1 - x2; */ \
b5 = vec_add(b6, b2); /* b5 = b6 + b2; */ \
cnst = LD_W2; \
b5 = vec_madd(cnst, b5, mzero); /* b5 = b5 * W2; */ \
cnst = LD_W1; \
b2 = vec_madd(cnst, b2, b5); /* b2 = b5 + b2 * W1; */ \
cnst = LD_W0; \
b6 = vec_madd(cnst, b6, b5); /* b6 = b5 + b6 * W0; */ \
\
x0 = vec_add(x4, x7); /* x0 = x4 + x7; */ \
x1 = vec_add(x5, x6); /* x1 = x5 + x6; */ \
x2 = vec_add(x4, x6); /* x2 = x4 + x6; */ \
x3 = vec_add(x5, x7); /* x3 = x5 + x7; */ \
x8 = vec_add(x2, x3); /* x8 = x2 + x3; */ \
cnst = LD_W3; \
x8 = vec_madd(cnst, x8, mzero); /* x8 = x8 * W3; */ \
\
cnst = LD_W8; \
x0 = vec_madd(cnst, x0, mzero); /* x0 *= W8; */ \
cnst = LD_W9; \
x1 = vec_madd(cnst, x1, mzero); /* x1 *= W9; */ \
cnst = LD_WA; \
x2 = vec_madd(cnst, x2, x8); /* x2 = x2 * WA + x8; */ \
cnst = LD_WB; \
x3 = vec_madd(cnst, x3, x8); /* x3 = x3 * WB + x8; */ \
\
cnst = LD_W4; \
b7 = vec_madd(cnst, x4, x0); /* b7 = x4 * W4 + x0; */ \
cnst = LD_W5; \
b5 = vec_madd(cnst, x5, x1); /* b5 = x5 * W5 + x1; */ \
cnst = LD_W6; \
b3 = vec_madd(cnst, x6, x1); /* b3 = x6 * W6 + x1; */ \
cnst = LD_W7; \
b1 = vec_madd(cnst, x7, x0); /* b1 = x7 * W7 + x0; */ \
\
b7 = vec_add(b7, x2); /* b7 += x2; */ \
b5 = vec_add(b5, x3); /* b5 += x3; */ \
b3 = vec_add(b3, x2); /* b3 += x2; */ \
b1 = vec_add(b1, x3) /* b1 += x3; */ \
/* }}} */
/* two dimensional discrete cosine transform */
void ff_fdct_altivec(int16_t *block)
{
vector signed short *bp;
const vector float *cp = fdctconsts;
vector float b00, b10, b20, b30, b40, b50, b60, b70;
vector float b01, b11, b21, b31, b41, b51, b61, b71;
vector float mzero, cnst, cnsts0, cnsts1, cnsts2;
vector float x0, x1, x2, x3, x4, x5, x6, x7, x8;
/* setup constants {{{ */
/* mzero = -0.0 */
mzero = ((vector float) vec_splat_u32(-1));
mzero = ((vector float) vec_sl(vu32(mzero), vu32(mzero)));
cnsts0 = vec_ld(0, cp);
cp++;
cnsts1 = vec_ld(0, cp);
cp++;
cnsts2 = vec_ld(0, cp);
/* }}} */
/* 8x8 matrix transpose (vector short[8]) {{{ */
#define MERGE_S16(hl, a, b) vec_merge ## hl(vs16(a), vs16(b))
bp = (vector signed short *) block;
b00 = ((vector float) vec_ld(0, bp));
b40 = ((vector float) vec_ld(16 * 4, bp));
b01 = ((vector float) MERGE_S16(h, b00, b40));
b11 = ((vector float) MERGE_S16(l, b00, b40));
bp++;
b10 = ((vector float) vec_ld(0, bp));
b50 = ((vector float) vec_ld(16 * 4, bp));
b21 = ((vector float) MERGE_S16(h, b10, b50));
b31 = ((vector float) MERGE_S16(l, b10, b50));
bp++;
b20 = ((vector float) vec_ld(0, bp));
b60 = ((vector float) vec_ld(16 * 4, bp));
b41 = ((vector float) MERGE_S16(h, b20, b60));
b51 = ((vector float) MERGE_S16(l, b20, b60));
bp++;
b30 = ((vector float) vec_ld(0, bp));
b70 = ((vector float) vec_ld(16 * 4, bp));
b61 = ((vector float) MERGE_S16(h, b30, b70));
b71 = ((vector float) MERGE_S16(l, b30, b70));
x0 = ((vector float) MERGE_S16(h, b01, b41));
x1 = ((vector float) MERGE_S16(l, b01, b41));
x2 = ((vector float) MERGE_S16(h, b11, b51));
x3 = ((vector float) MERGE_S16(l, b11, b51));
x4 = ((vector float) MERGE_S16(h, b21, b61));
x5 = ((vector float) MERGE_S16(l, b21, b61));
x6 = ((vector float) MERGE_S16(h, b31, b71));
x7 = ((vector float) MERGE_S16(l, b31, b71));
b00 = ((vector float) MERGE_S16(h, x0, x4));
b10 = ((vector float) MERGE_S16(l, x0, x4));
b20 = ((vector float) MERGE_S16(h, x1, x5));
b30 = ((vector float) MERGE_S16(l, x1, x5));
b40 = ((vector float) MERGE_S16(h, x2, x6));
b50 = ((vector float) MERGE_S16(l, x2, x6));
b60 = ((vector float) MERGE_S16(h, x3, x7));
b70 = ((vector float) MERGE_S16(l, x3, x7));
#undef MERGE_S16
/* }}} */
/* Some of the initial calculations can be done as vector short
* before conversion to vector float. The following code section
* takes advantage of this. */
/* fdct rows {{{ */
x0 = ((vector float) vec_add(vs16(b00), vs16(b70)));
x7 = ((vector float) vec_sub(vs16(b00), vs16(b70)));
x1 = ((vector float) vec_add(vs16(b10), vs16(b60)));
x6 = ((vector float) vec_sub(vs16(b10), vs16(b60)));
x2 = ((vector float) vec_add(vs16(b20), vs16(b50)));
x5 = ((vector float) vec_sub(vs16(b20), vs16(b50)));
x3 = ((vector float) vec_add(vs16(b30), vs16(b40)));
x4 = ((vector float) vec_sub(vs16(b30), vs16(b40)));
b70 = ((vector float) vec_add(vs16(x0), vs16(x3)));
b10 = ((vector float) vec_add(vs16(x1), vs16(x2)));
b00 = ((vector float) vec_add(vs16(b70), vs16(b10)));
b40 = ((vector float) vec_sub(vs16(b70), vs16(b10)));
#define CTF0(n) \
b ## n ## 1 = ((vector float) vec_unpackl(vs16(b ## n ## 0))); \
b ## n ## 0 = ((vector float) vec_unpackh(vs16(b ## n ## 0))); \
b ## n ## 1 = vec_ctf(vs32(b ## n ## 1), 0); \
b ## n ## 0 = vec_ctf(vs32(b ## n ## 0), 0)
CTF0(0);
CTF0(4);
b20 = ((vector float) vec_sub(vs16(x0), vs16(x3)));
b60 = ((vector float) vec_sub(vs16(x1), vs16(x2)));
CTF0(2);
CTF0(6);
#undef CTF0
x0 = vec_add(b60, b20);
x1 = vec_add(b61, b21);
cnst = LD_W2;
x0 = vec_madd(cnst, x0, mzero);
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_W1;
b20 = vec_madd(cnst, b20, x0);
b21 = vec_madd(cnst, b21, x1);
cnst = LD_W0;
b60 = vec_madd(cnst, b60, x0);
b61 = vec_madd(cnst, b61, x1);
#define CTFX(x, b) \
b ## 0 = ((vector float) vec_unpackh(vs16(x))); \
b ## 1 = ((vector float) vec_unpackl(vs16(x))); \
b ## 0 = vec_ctf(vs32(b ## 0), 0); \
b ## 1 = vec_ctf(vs32(b ## 1), 0)
CTFX(x4, b7);
CTFX(x5, b5);
CTFX(x6, b3);
CTFX(x7, b1);
#undef CTFX
x0 = vec_add(b70, b10);
x1 = vec_add(b50, b30);
x2 = vec_add(b70, b30);
x3 = vec_add(b50, b10);
x8 = vec_add(x2, x3);
cnst = LD_W3;
x8 = vec_madd(cnst, x8, mzero);
cnst = LD_W8;
x0 = vec_madd(cnst, x0, mzero);
cnst = LD_W9;
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_WA;
x2 = vec_madd(cnst, x2, x8);
cnst = LD_WB;
x3 = vec_madd(cnst, x3, x8);
cnst = LD_W4;
b70 = vec_madd(cnst, b70, x0);
cnst = LD_W5;
b50 = vec_madd(cnst, b50, x1);
cnst = LD_W6;
b30 = vec_madd(cnst, b30, x1);
cnst = LD_W7;
b10 = vec_madd(cnst, b10, x0);
b70 = vec_add(b70, x2);
b50 = vec_add(b50, x3);
b30 = vec_add(b30, x2);
b10 = vec_add(b10, x3);
x0 = vec_add(b71, b11);
x1 = vec_add(b51, b31);
x2 = vec_add(b71, b31);
x3 = vec_add(b51, b11);
x8 = vec_add(x2, x3);
cnst = LD_W3;
x8 = vec_madd(cnst, x8, mzero);
cnst = LD_W8;
x0 = vec_madd(cnst, x0, mzero);
cnst = LD_W9;
x1 = vec_madd(cnst, x1, mzero);
cnst = LD_WA;
x2 = vec_madd(cnst, x2, x8);
cnst = LD_WB;
x3 = vec_madd(cnst, x3, x8);
cnst = LD_W4;
b71 = vec_madd(cnst, b71, x0);
cnst = LD_W5;
b51 = vec_madd(cnst, b51, x1);
cnst = LD_W6;
b31 = vec_madd(cnst, b31, x1);
cnst = LD_W7;
b11 = vec_madd(cnst, b11, x0);
b71 = vec_add(b71, x2);
b51 = vec_add(b51, x3);
b31 = vec_add(b31, x2);
b11 = vec_add(b11, x3);
/* }}} */
/* 8x8 matrix transpose (vector float[8][2]) {{{ */
x0 = vec_mergel(b00, b20);
x1 = vec_mergeh(b00, b20);
x2 = vec_mergel(b10, b30);
x3 = vec_mergeh(b10, b30);
b00 = vec_mergeh(x1, x3);
b10 = vec_mergel(x1, x3);
b20 = vec_mergeh(x0, x2);
b30 = vec_mergel(x0, x2);
x4 = vec_mergel(b41, b61);
x5 = vec_mergeh(b41, b61);
x6 = vec_mergel(b51, b71);
x7 = vec_mergeh(b51, b71);
b41 = vec_mergeh(x5, x7);
b51 = vec_mergel(x5, x7);
b61 = vec_mergeh(x4, x6);
b71 = vec_mergel(x4, x6);
x0 = vec_mergel(b01, b21);
x1 = vec_mergeh(b01, b21);
x2 = vec_mergel(b11, b31);
x3 = vec_mergeh(b11, b31);
x4 = vec_mergel(b40, b60);
x5 = vec_mergeh(b40, b60);
x6 = vec_mergel(b50, b70);
x7 = vec_mergeh(b50, b70);
b40 = vec_mergeh(x1, x3);
b50 = vec_mergel(x1, x3);
b60 = vec_mergeh(x0, x2);
b70 = vec_mergel(x0, x2);
b01 = vec_mergeh(x5, x7);
b11 = vec_mergel(x5, x7);
b21 = vec_mergeh(x4, x6);
b31 = vec_mergel(x4, x6);
/* }}} */
FDCTCOL(b00, b10, b20, b30, b40, b50, b60, b70);
FDCTCOL(b01, b11, b21, b31, b41, b51, b61, b71);
/* round, convert back to short {{{ */
#define CTS(n) \
b ## n ## 0 = vec_round(b ## n ## 0); \
b ## n ## 1 = vec_round(b ## n ## 1); \
b ## n ## 0 = ((vector float) vec_cts(b ## n ## 0, 0)); \
b ## n ## 1 = ((vector float) vec_cts(b ## n ## 1, 0)); \
b ## n ## 0 = ((vector float) vec_pack(vs32(b ## n ## 0), \
vs32(b ## n ## 1))); \
vec_st(vs16(b ## n ## 0), 0, bp)
bp = (vector signed short *) block;
CTS(0);
bp++;
CTS(1);
bp++;
CTS(2);
bp++;
CTS(3);
bp++;
CTS(4);
bp++;
CTS(5);
bp++;
CTS(6);
bp++;
CTS(7);
#undef CTS
/* }}} */
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_fdctdsp_init_ppc(FDCTDSPContext *c, AVCodecContext *avctx,
unsigned high_bit_depth)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (!high_bit_depth) {
if (avctx->dct_algo == FF_DCT_AUTO ||
avctx->dct_algo == FF_DCT_ALTIVEC) {
c->fdct = ff_fdct_altivec;
}
}
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/fmtconvert_altivec.c
+65
View File
@@ -0,0 +1,65 @@
/*
* Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/fmtconvert.h"
#if HAVE_ALTIVEC
static void int32_to_float_fmul_scalar_altivec(float *dst, const int32_t *src,
float mul, int len)
{
union {
vector float v;
float s[4];
} mul_u;
int i;
vector float src1, src2, dst1, dst2, mul_v, zero;
zero = (vector float)vec_splat_u32(0);
mul_u.s[0] = mul;
mul_v = vec_splat(mul_u.v, 0);
for (i = 0; i < len; i += 8) {
src1 = vec_ctf(vec_ld(0, src+i), 0);
src2 = vec_ctf(vec_ld(16, src+i), 0);
dst1 = vec_madd(src1, mul_v, zero);
dst2 = vec_madd(src2, mul_v, zero);
vec_st(dst1, 0, dst+i);
vec_st(dst2, 16, dst+i);
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_fmt_convert_init_ppc(FmtConvertContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/h264chroma_init.c
+65
View File
@@ -0,0 +1,65 @@
/*
* Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/intreadwrite.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/h264chroma.h"
#if HAVE_ALTIVEC
#define PUT_OP_U8_ALTIVEC(d, s, dst) d = s
#define AVG_OP_U8_ALTIVEC(d, s, dst) d = vec_avg(dst, s)
#define OP_U8_ALTIVEC PUT_OP_U8_ALTIVEC
#define PREFIX_h264_chroma_mc8_altivec put_h264_chroma_mc8_altivec
#define PREFIX_h264_chroma_mc8_num altivec_put_h264_chroma_mc8_num
#include "h264chroma_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_h264_chroma_mc8_altivec
#undef PREFIX_h264_chroma_mc8_num
#define OP_U8_ALTIVEC AVG_OP_U8_ALTIVEC
#define PREFIX_h264_chroma_mc8_altivec avg_h264_chroma_mc8_altivec
#define PREFIX_h264_chroma_mc8_num altivec_avg_h264_chroma_mc8_num
#include "h264chroma_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_h264_chroma_mc8_altivec
#undef PREFIX_h264_chroma_mc8_num
#endif /* HAVE_ALTIVEC */
av_cold void ff_h264chroma_init_ppc(H264ChromaContext *c, int bit_depth)
{
#if HAVE_ALTIVEC
const int high_bit_depth = bit_depth > 8;
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (!high_bit_depth) {
c->put_h264_chroma_pixels_tab[0] = put_h264_chroma_mc8_altivec;
c->avg_h264_chroma_pixels_tab[0] = avg_h264_chroma_mc8_altivec;
}
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/h264chroma_template.c
+246
View File
@@ -0,0 +1,246 @@
/*
* Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/util_altivec.h"
/* this code assume that stride % 16 == 0 */
#define CHROMA_MC8_ALTIVEC_CORE(BIAS1, BIAS2) \
vsrc2ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc2uc);\
vsrc3ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc3uc);\
\
psum = vec_mladd(vA, vsrc0ssH, BIAS1);\
psum = vec_mladd(vB, vsrc1ssH, psum);\
psum = vec_mladd(vC, vsrc2ssH, psum);\
psum = vec_mladd(vD, vsrc3ssH, psum);\
psum = BIAS2(psum);\
psum = vec_sr(psum, v6us);\
\
vdst = vec_ld(0, dst);\
ppsum = (vec_u8)vec_pack(psum, psum);\
vfdst = vec_perm(vdst, ppsum, fperm);\
\
OP_U8_ALTIVEC(fsum, vfdst, vdst);\
\
vec_st(fsum, 0, dst);\
\
vsrc0ssH = vsrc2ssH;\
vsrc1ssH = vsrc3ssH;\
\
dst += stride;\
src += stride;
#define CHROMA_MC8_ALTIVEC_CORE_SIMPLE \
\
vsrc0ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc0uc);\
vsrc1ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc1uc);\
\
psum = vec_mladd(vA, vsrc0ssH, v32ss);\
psum = vec_mladd(vE, vsrc1ssH, psum);\
psum = vec_sr(psum, v6us);\
\
vdst = vec_ld(0, dst);\
ppsum = (vec_u8)vec_pack(psum, psum);\
vfdst = vec_perm(vdst, ppsum, fperm);\
\
OP_U8_ALTIVEC(fsum, vfdst, vdst);\
\
vec_st(fsum, 0, dst);\
\
dst += stride;\
src += stride;
#define noop(a) a
#define add28(a) vec_add(v28ss, a)
#if HAVE_BIGENDIAN
#define GET_VSRC1(vs0, off, b, perm0, s){ \
vec_u8 vsrcCuc, vsrcDuc; \
vsrcCuc = vec_ld(off, s); \
if (loadSecond){ \
vsrcDuc = vec_ld(off + b, s); \
} else \
vsrcDuc = vsrcCuc; \
\
vs0 = vec_perm(vsrcCuc, vsrcDuc, perm0); \
}
#define GET_VSRC(vs0, vs1, off, b, perm0, perm1, s){ \
vec_u8 vsrcCuc, vsrcDuc; \
vsrcCuc = vec_ld(off, s); \
if (loadSecond){ \
vsrcDuc = vec_ld(off + b, s); \
} else \
vsrcDuc = vsrcCuc; \
\
vs0 = vec_perm(vsrcCuc, vsrcDuc, perm0); \
if (reallyBadAlign){ \
vs1 = vsrcDuc; \
} else \
vs1 = vec_perm(vsrcCuc, vsrcDuc, perm1); \
}
#else
#define GET_VSRC1(vs0, off, b, perm0, s){ \
vs0 = vec_vsx_ld(off, s); \
}
#define GET_VSRC(vs0, vs1, off, b, perm0, perm1, s){ \
vs0 = vec_vsx_ld(off, s); \
vs1 = vec_vsx_ld(off + 1, s); \
}
#endif /* HAVE_BIGENDIAN */
#ifdef PREFIX_h264_chroma_mc8_altivec
static void PREFIX_h264_chroma_mc8_altivec(uint8_t * dst, const uint8_t * src,
ptrdiff_t stride, int h,
int x, int y)
{
DECLARE_ALIGNED(16, signed int, ABCD)[4] =
{((8 - x) * (8 - y)),
(( x) * (8 - y)),
((8 - x) * ( y)),
(( x) * ( y))};
register int i;
vec_u8 fperm;
LOAD_ZERO;
const vec_s32 vABCD = vec_ld(0, ABCD);
const vec_s16 vA = VEC_SPLAT16(vABCD, 1);
const vec_s16 vB = VEC_SPLAT16(vABCD, 3);
const vec_s16 vC = VEC_SPLAT16(vABCD, 5);
const vec_s16 vD = VEC_SPLAT16(vABCD, 7);
const vec_s16 v32ss = vec_sl(vec_splat_s16(1),vec_splat_u16(5));
const vec_u16 v6us = vec_splat_u16(6);
vec_u8 vsrcperm0, vsrcperm1;
vec_u8 vsrc0uc, vsrc1uc;
vec_s16 vsrc0ssH, vsrc1ssH;
vec_u8 vsrc2uc, vsrc3uc;
vec_s16 vsrc2ssH, vsrc3ssH, psum;
vec_u8 vdst, ppsum, vfdst, fsum;
#if HAVE_BIGENDIAN
register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
vsrcperm0 = vec_lvsl(0, src);
vsrcperm1 = vec_lvsl(1, src);
#endif
if (((unsigned long)dst) % 16 == 0) {
fperm = (vec_u8){0x10, 0x11, 0x12, 0x13,
0x14, 0x15, 0x16, 0x17,
0x08, 0x09, 0x0A, 0x0B,
0x0C, 0x0D, 0x0E, 0x0F};
} else {
fperm = (vec_u8){0x00, 0x01, 0x02, 0x03,
0x04, 0x05, 0x06, 0x07,
0x18, 0x19, 0x1A, 0x1B,
0x1C, 0x1D, 0x1E, 0x1F};
}
GET_VSRC(vsrc0uc, vsrc1uc, 0, 16, vsrcperm0, vsrcperm1, src);
vsrc0ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc0uc);
vsrc1ssH = (vec_s16)VEC_MERGEH(zero_u8v,(vec_u8)vsrc1uc);
if (ABCD[3]) {
for (i = 0 ; i < h ; i++) {
GET_VSRC(vsrc2uc, vsrc3uc, stride, 16, vsrcperm0, vsrcperm1, src);
CHROMA_MC8_ALTIVEC_CORE(v32ss, noop);
}
} else {
const vec_s16 vE = vec_add(vB, vC);
if (ABCD[2]) { // x == 0 B == 0
for (i = 0 ; i < h ; i++) {
GET_VSRC1(vsrc1uc, stride, 15, vsrcperm0, src);
CHROMA_MC8_ALTIVEC_CORE_SIMPLE;
vsrc0uc = vsrc1uc;
}
} else { // y == 0 C == 0
for (i = 0 ; i < h ; i++) {
GET_VSRC(vsrc0uc, vsrc1uc, 0, 15, vsrcperm0, vsrcperm1, src);
CHROMA_MC8_ALTIVEC_CORE_SIMPLE;
}
}
}
}
#endif
/* this code assume that stride % 16 == 0 */
#ifdef PREFIX_no_rnd_vc1_chroma_mc8_altivec
static void PREFIX_no_rnd_vc1_chroma_mc8_altivec(uint8_t *dst, const uint8_t *src,
ptrdiff_t stride, int h,
int x, int y)
{
DECLARE_ALIGNED(16, signed int, ABCD)[4] =
{((8 - x) * (8 - y)),
(( x) * (8 - y)),
((8 - x) * ( y)),
(( x) * ( y))};
register int i;
vec_u8 fperm;
LOAD_ZERO;
const vec_s32 vABCD = vec_ld(0, ABCD);
const vec_s16 vA = VEC_SPLAT16(vABCD, 1);
const vec_s16 vB = VEC_SPLAT16(vABCD, 3);
const vec_s16 vC = VEC_SPLAT16(vABCD, 5);
const vec_s16 vD = VEC_SPLAT16(vABCD, 7);
const vec_s16 v28ss = vec_sub(vec_sl(vec_splat_s16(1),vec_splat_u16(5)),vec_splat_s16(4));
const vec_u16 v6us = vec_splat_u16(6);
vec_u8 vsrcperm0, vsrcperm1;
vec_u8 vsrc0uc, vsrc1uc;
vec_s16 vsrc0ssH, vsrc1ssH;
vec_u8 vsrc2uc, vsrc3uc;
vec_s16 vsrc2ssH, vsrc3ssH, psum;
vec_u8 vdst, ppsum, vfdst, fsum;
#if HAVE_BIGENDIAN
register int loadSecond = (((unsigned long)src) % 16) <= 7 ? 0 : 1;
register int reallyBadAlign = (((unsigned long)src) % 16) == 15 ? 1 : 0;
vsrcperm0 = vec_lvsl(0, src);
vsrcperm1 = vec_lvsl(1, src);
#endif
if (((unsigned long)dst) % 16 == 0) {
fperm = (vec_u8){0x10, 0x11, 0x12, 0x13,
0x14, 0x15, 0x16, 0x17,
0x08, 0x09, 0x0A, 0x0B,
0x0C, 0x0D, 0x0E, 0x0F};
} else {
fperm = (vec_u8){0x00, 0x01, 0x02, 0x03,
0x04, 0x05, 0x06, 0x07,
0x18, 0x19, 0x1A, 0x1B,
0x1C, 0x1D, 0x1E, 0x1F};
}
GET_VSRC(vsrc0uc, vsrc1uc, 0, 16, vsrcperm0, vsrcperm1, src);
vsrc0ssH = (vec_s16)VEC_MERGEH(zero_u8v, (vec_u8)vsrc0uc);
vsrc1ssH = (vec_s16)VEC_MERGEH(zero_u8v, (vec_u8)vsrc1uc);
for (i = 0 ; i < h ; i++) {
GET_VSRC(vsrc2uc, vsrc3uc, stride, 16, vsrcperm0, vsrcperm1, src);
CHROMA_MC8_ALTIVEC_CORE(vec_splat_s16(0), add28);
}
}
#endif
#undef noop
#undef add28
#undef CHROMA_MC8_ALTIVEC_CORE
ffmpeg/libavcodec/ppc/h264dsp.c
+814
View File
@@ -0,0 +1,814 @@
/*
* Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include <stdint.h>
#include <string.h>
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/intreadwrite.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/h264dec.h"
#include "libavcodec/h264dsp.h"
#if HAVE_ALTIVEC
/****************************************************************************
* IDCT transform:
****************************************************************************/
#define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3) \
/* 1st stage */ \
vz0 = vec_add(vb0,vb2); /* temp[0] = Y[0] + Y[2] */ \
vz1 = vec_sub(vb0,vb2); /* temp[1] = Y[0] - Y[2] */ \
vz2 = vec_sra(vb1,vec_splat_u16(1)); \
vz2 = vec_sub(vz2,vb3); /* temp[2] = Y[1].1/2 - Y[3] */ \
vz3 = vec_sra(vb3,vec_splat_u16(1)); \
vz3 = vec_add(vb1,vz3); /* temp[3] = Y[1] + Y[3].1/2 */ \
/* 2nd stage: output */ \
va0 = vec_add(vz0,vz3); /* x[0] = temp[0] + temp[3] */ \
va1 = vec_add(vz1,vz2); /* x[1] = temp[1] + temp[2] */ \
va2 = vec_sub(vz1,vz2); /* x[2] = temp[1] - temp[2] */ \
va3 = vec_sub(vz0,vz3) /* x[3] = temp[0] - temp[3] */
#define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
b0 = vec_mergeh( a0, a0 ); \
b1 = vec_mergeh( a1, a0 ); \
b2 = vec_mergeh( a2, a0 ); \
b3 = vec_mergeh( a3, a0 ); \
a0 = vec_mergeh( b0, b2 ); \
a1 = vec_mergel( b0, b2 ); \
a2 = vec_mergeh( b1, b3 ); \
a3 = vec_mergel( b1, b3 ); \
b0 = vec_mergeh( a0, a2 ); \
b1 = vec_mergel( a0, a2 ); \
b2 = vec_mergeh( a1, a3 ); \
b3 = vec_mergel( a1, a3 )
#if HAVE_BIGENDIAN
#define vdst_load(d) \
vdst_orig = vec_ld(0, dst); \
vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask);
#else
#define vdst_load(d) vdst = vec_vsx_ld(0, dst)
#endif
#define VEC_LOAD_U8_ADD_S16_STORE_U8(va) \
vdst_load(); \
vdst_ss = (vec_s16) VEC_MERGEH(zero_u8v, vdst); \
va = vec_add(va, vdst_ss); \
va_u8 = vec_packsu(va, zero_s16v); \
va_u32 = vec_splat((vec_u32)va_u8, 0); \
vec_ste(va_u32, element, (uint32_t*)dst);
static void h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
{
vec_s16 va0, va1, va2, va3;
vec_s16 vz0, vz1, vz2, vz3;
vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
vec_u8 va_u8;
vec_u32 va_u32;
vec_s16 vdst_ss;
const vec_u16 v6us = vec_splat_u16(6);
vec_u8 vdst, vdst_orig;
vec_u8 vdst_mask = vec_lvsl(0, dst);
int element = ((unsigned long)dst & 0xf) >> 2;
LOAD_ZERO;
block[0] += 32; /* add 32 as a DC-level for rounding */
vtmp0 = vec_ld(0,block);
vtmp1 = vec_sld(vtmp0, vtmp0, 8);
vtmp2 = vec_ld(16,block);
vtmp3 = vec_sld(vtmp2, vtmp2, 8);
memset(block, 0, 16 * sizeof(int16_t));
VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
va0 = vec_sra(va0,v6us);
va1 = vec_sra(va1,v6us);
va2 = vec_sra(va2,v6us);
va3 = vec_sra(va3,v6us);
VEC_LOAD_U8_ADD_S16_STORE_U8(va0);
dst += stride;
VEC_LOAD_U8_ADD_S16_STORE_U8(va1);
dst += stride;
VEC_LOAD_U8_ADD_S16_STORE_U8(va2);
dst += stride;
VEC_LOAD_U8_ADD_S16_STORE_U8(va3);
}
#define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7, d0, d1, d2, d3, d4, d5, d6, d7) {\
/* a0 = SRC(0) + SRC(4); */ \
vec_s16 a0v = vec_add(s0, s4); \
/* a2 = SRC(0) - SRC(4); */ \
vec_s16 a2v = vec_sub(s0, s4); \
/* a4 = (SRC(2)>>1) - SRC(6); */ \
vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6); \
/* a6 = (SRC(6)>>1) + SRC(2); */ \
vec_s16 a6v = vec_add(vec_sra(s6, onev), s2); \
/* b0 = a0 + a6; */ \
vec_s16 b0v = vec_add(a0v, a6v); \
/* b2 = a2 + a4; */ \
vec_s16 b2v = vec_add(a2v, a4v); \
/* b4 = a2 - a4; */ \
vec_s16 b4v = vec_sub(a2v, a4v); \
/* b6 = a0 - a6; */ \
vec_s16 b6v = vec_sub(a0v, a6v); \
/* a1 = SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
/* a1 = (SRC(5)-SRC(3)) - (SRC(7) + (SRC(7)>>1)); */ \
vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
/* a3 = SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
/* a3 = (SRC(7)+SRC(1)) - (SRC(3) + (SRC(3)>>1)); */ \
vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
/* a5 = SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
/* a5 = (SRC(7)-SRC(1)) + SRC(5) + (SRC(5)>>1); */ \
vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
/* a7 = SRC(5)+SRC(3) + SRC(1) + (SRC(1)>>1); */ \
vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
/* b1 = (a7>>2) + a1; */ \
vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
/* b3 = a3 + (a5>>2); */ \
vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
/* b5 = (a3>>2) - a5; */ \
vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
/* b7 = a7 - (a1>>2); */ \
vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
/* DST(0, b0 + b7); */ \
d0 = vec_add(b0v, b7v); \
/* DST(1, b2 + b5); */ \
d1 = vec_add(b2v, b5v); \
/* DST(2, b4 + b3); */ \
d2 = vec_add(b4v, b3v); \
/* DST(3, b6 + b1); */ \
d3 = vec_add(b6v, b1v); \
/* DST(4, b6 - b1); */ \
d4 = vec_sub(b6v, b1v); \
/* DST(5, b4 - b3); */ \
d5 = vec_sub(b4v, b3v); \
/* DST(6, b2 - b5); */ \
d6 = vec_sub(b2v, b5v); \
/* DST(7, b0 - b7); */ \
d7 = vec_sub(b0v, b7v); \
}
#if HAVE_BIGENDIAN
#define GET_2PERM(ldv, stv, d) \
ldv = vec_lvsl(0, d); \
stv = vec_lvsr(8, d);
#define dstv_load(d) \
vec_u8 hv = vec_ld( 0, d ); \
vec_u8 lv = vec_ld( 7, d); \
vec_u8 dstv = vec_perm( hv, lv, (vec_u8)perm_ldv );
#define dest_unligned_store(d) \
vec_u8 edgehv; \
vec_u8 bodyv = vec_perm( idstsum8, idstsum8, perm_stv ); \
vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv ); \
lv = vec_sel( lv, bodyv, edgelv ); \
vec_st( lv, 7, d ); \
hv = vec_ld( 0, d ); \
edgehv = vec_perm( zero_u8v, sel, perm_stv ); \
hv = vec_sel( hv, bodyv, edgehv ); \
vec_st( hv, 0, d );
#else
#define GET_2PERM(ldv, stv, d) {}
#define dstv_load(d) vec_u8 dstv = vec_vsx_ld(0, d)
#define dest_unligned_store(d)\
vec_u8 dst8 = vec_perm((vec_u8)idstsum8, dstv, vcprm(2,3,s2,s3));\
vec_vsx_st(dst8, 0, d)
#endif /* HAVE_BIGENDIAN */
#define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
/* unaligned load */ \
dstv_load(dest); \
vec_s16 idct_sh6 = vec_sra(idctv, sixv); \
vec_u16 dst16 = (vec_u16)VEC_MERGEH(zero_u8v, dstv); \
vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16); \
vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum); \
/* unaligned store */ \
dest_unligned_store(dest);\
}
static void h264_idct8_add_altivec(uint8_t *dst, int16_t *dct, int stride)
{
vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
vec_u8 perm_ldv, perm_stv;
GET_2PERM(perm_ldv, perm_stv, dst);
const vec_u16 onev = vec_splat_u16(1);
const vec_u16 twov = vec_splat_u16(2);
const vec_u16 sixv = vec_splat_u16(6);
const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
LOAD_ZERO;
dct[0] += 32; // rounding for the >>6 at the end
s0 = vec_ld(0x00, (int16_t*)dct);
s1 = vec_ld(0x10, (int16_t*)dct);
s2 = vec_ld(0x20, (int16_t*)dct);
s3 = vec_ld(0x30, (int16_t*)dct);
s4 = vec_ld(0x40, (int16_t*)dct);
s5 = vec_ld(0x50, (int16_t*)dct);
s6 = vec_ld(0x60, (int16_t*)dct);
s7 = vec_ld(0x70, (int16_t*)dct);
memset(dct, 0, 64 * sizeof(int16_t));
IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
d0, d1, d2, d3, d4, d5, d6, d7);
TRANSPOSE8( d0, d1, d2, d3, d4, d5, d6, d7 );
IDCT8_1D_ALTIVEC(d0, d1, d2, d3, d4, d5, d6, d7,
idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
}
#if HAVE_BIGENDIAN
#define DST_LD vec_ld
#else
#define DST_LD vec_vsx_ld
#endif
static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
{
vec_s16 dc16;
vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
vec_s32 v_dc32;
LOAD_ZERO;
DECLARE_ALIGNED(16, int, dc);
int i;
dc = (block[0] + 32) >> 6;
block[0] = 0;
v_dc32 = vec_lde(0, &dc);
dc16 = VEC_SPLAT16((vec_s16)v_dc32, 1);
if (size == 4)
dc16 = VEC_SLD16(dc16, zero_s16v, 8);
dcplus = vec_packsu(dc16, zero_s16v);
dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
#if HAVE_BIGENDIAN
aligner = vec_lvsr(0, dst);
dcplus = vec_perm(dcplus, dcplus, aligner);
dcminus = vec_perm(dcminus, dcminus, aligner);
#endif
for (i = 0; i < size; i += 4) {
v0 = DST_LD(0, dst+0*stride);
v1 = DST_LD(0, dst+1*stride);
v2 = DST_LD(0, dst+2*stride);
v3 = DST_LD(0, dst+3*stride);
v0 = vec_adds(v0, dcplus);
v1 = vec_adds(v1, dcplus);
v2 = vec_adds(v2, dcplus);
v3 = vec_adds(v3, dcplus);
v0 = vec_subs(v0, dcminus);
v1 = vec_subs(v1, dcminus);
v2 = vec_subs(v2, dcminus);
v3 = vec_subs(v3, dcminus);
VEC_ST(v0, 0, dst+0*stride);
VEC_ST(v1, 0, dst+1*stride);
VEC_ST(v2, 0, dst+2*stride);
VEC_ST(v3, 0, dst+3*stride);
dst += 4*stride;
}
}
static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
{
h264_idct_dc_add_internal(dst, block, stride, 4);
}
static void h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
{
h264_idct_dc_add_internal(dst, block, stride, 8);
}
static void h264_idct_add16_altivec(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
const uint8_t nnzc[5 * 8])
{
int i;
for(i=0; i<16; i++){
int nnz = nnzc[ scan8[i] ];
if(nnz){
if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
else h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
}
}
}
static void h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
const uint8_t nnzc[5 * 8])
{
int i;
for(i=0; i<16; i++){
if(nnzc[ scan8[i] ]) h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
}
}
static void h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset,
int16_t *block, int stride,
const uint8_t nnzc[5 * 8])
{
int i;
for(i=0; i<16; i+=4){
int nnz = nnzc[ scan8[i] ];
if(nnz){
if(nnz==1 && block[i*16]) h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
else h264_idct8_add_altivec(dst + block_offset[i], block + i*16, stride);
}
}
}
static void h264_idct_add8_altivec(uint8_t **dest, const int *block_offset,
int16_t *block, int stride,
const uint8_t nnzc[15 * 8])
{
int i, j;
for (j = 1; j < 3; j++) {
for(i = j * 16; i < j * 16 + 4; i++){
if(nnzc[ scan8[i] ])
h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
else if(block[i*16])
h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
}
}
}
#define transpose4x16(r0, r1, r2, r3) { \
register vec_u8 r4; \
register vec_u8 r5; \
register vec_u8 r6; \
register vec_u8 r7; \
\
r4 = vec_mergeh(r0, r2); /*0, 2 set 0*/ \
r5 = vec_mergel(r0, r2); /*0, 2 set 1*/ \
r6 = vec_mergeh(r1, r3); /*1, 3 set 0*/ \
r7 = vec_mergel(r1, r3); /*1, 3 set 1*/ \
\
r0 = vec_mergeh(r4, r6); /*all set 0*/ \
r1 = vec_mergel(r4, r6); /*all set 1*/ \
r2 = vec_mergeh(r5, r7); /*all set 2*/ \
r3 = vec_mergel(r5, r7); /*all set 3*/ \
}
static inline void write16x4(uint8_t *dst, int dst_stride,
register vec_u8 r0, register vec_u8 r1,
register vec_u8 r2, register vec_u8 r3) {
DECLARE_ALIGNED(16, unsigned char, result)[64];
uint32_t *src_int = (uint32_t *)result;
vec_st(r0, 0, result);
vec_st(r1, 16, result);
vec_st(r2, 32, result);
vec_st(r3, 48, result);
/* FIXME: there has to be a better way!!!! */
AV_WN32(dst, AV_RN32A(src_int + 0));
AV_WN32(dst + dst_stride, AV_RN32A(src_int + 1));
AV_WN32(dst + 2 * dst_stride, AV_RN32A(src_int + 2));
AV_WN32(dst + 3 * dst_stride, AV_RN32A(src_int + 3));
AV_WN32(dst + 4 * dst_stride, AV_RN32A(src_int + 4));
AV_WN32(dst + 5 * dst_stride, AV_RN32A(src_int + 5));
AV_WN32(dst + 6 * dst_stride, AV_RN32A(src_int + 6));
AV_WN32(dst + 7 * dst_stride, AV_RN32A(src_int + 7));
AV_WN32(dst + 8 * dst_stride, AV_RN32A(src_int + 8));
AV_WN32(dst + 9 * dst_stride, AV_RN32A(src_int + 9));
AV_WN32(dst + 10 * dst_stride, AV_RN32A(src_int + 10));
AV_WN32(dst + 11 * dst_stride, AV_RN32A(src_int + 11));
AV_WN32(dst + 12 * dst_stride, AV_RN32A(src_int + 12));
AV_WN32(dst + 13 * dst_stride, AV_RN32A(src_int + 13));
AV_WN32(dst + 14 * dst_stride, AV_RN32A(src_int + 14));
AV_WN32(dst + 15 * dst_stride, AV_RN32A(src_int + 15));
}
/** @brief performs a 6x16 transpose of data in src, and stores it to dst
@todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
out of unaligned_load() */
#define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
register vec_u8 r0 = unaligned_load(0, src); \
register vec_u8 r1 = unaligned_load( src_stride, src); \
register vec_u8 r2 = unaligned_load(2* src_stride, src); \
register vec_u8 r3 = unaligned_load(3* src_stride, src); \
register vec_u8 r4 = unaligned_load(4* src_stride, src); \
register vec_u8 r5 = unaligned_load(5* src_stride, src); \
register vec_u8 r6 = unaligned_load(6* src_stride, src); \
register vec_u8 r7 = unaligned_load(7* src_stride, src); \
register vec_u8 r14 = unaligned_load(14*src_stride, src); \
register vec_u8 r15 = unaligned_load(15*src_stride, src); \
\
r8 = unaligned_load( 8*src_stride, src); \
r9 = unaligned_load( 9*src_stride, src); \
r10 = unaligned_load(10*src_stride, src); \
r11 = unaligned_load(11*src_stride, src); \
r12 = unaligned_load(12*src_stride, src); \
r13 = unaligned_load(13*src_stride, src); \
\
/*Merge first pairs*/ \
r0 = vec_mergeh(r0, r8); /*0, 8*/ \
r1 = vec_mergeh(r1, r9); /*1, 9*/ \
r2 = vec_mergeh(r2, r10); /*2,10*/ \
r3 = vec_mergeh(r3, r11); /*3,11*/ \
r4 = vec_mergeh(r4, r12); /*4,12*/ \
r5 = vec_mergeh(r5, r13); /*5,13*/ \
r6 = vec_mergeh(r6, r14); /*6,14*/ \
r7 = vec_mergeh(r7, r15); /*7,15*/ \
\
/*Merge second pairs*/ \
r8 = vec_mergeh(r0, r4); /*0,4, 8,12 set 0*/ \
r9 = vec_mergel(r0, r4); /*0,4, 8,12 set 1*/ \
r10 = vec_mergeh(r1, r5); /*1,5, 9,13 set 0*/ \
r11 = vec_mergel(r1, r5); /*1,5, 9,13 set 1*/ \
r12 = vec_mergeh(r2, r6); /*2,6,10,14 set 0*/ \
r13 = vec_mergel(r2, r6); /*2,6,10,14 set 1*/ \
r14 = vec_mergeh(r3, r7); /*3,7,11,15 set 0*/ \
r15 = vec_mergel(r3, r7); /*3,7,11,15 set 1*/ \
\
/*Third merge*/ \
r0 = vec_mergeh(r8, r12); /*0,2,4,6,8,10,12,14 set 0*/ \
r1 = vec_mergel(r8, r12); /*0,2,4,6,8,10,12,14 set 1*/ \
r2 = vec_mergeh(r9, r13); /*0,2,4,6,8,10,12,14 set 2*/ \
r4 = vec_mergeh(r10, r14); /*1,3,5,7,9,11,13,15 set 0*/ \
r5 = vec_mergel(r10, r14); /*1,3,5,7,9,11,13,15 set 1*/ \
r6 = vec_mergeh(r11, r15); /*1,3,5,7,9,11,13,15 set 2*/ \
/* Don't need to compute 3 and 7*/ \
\
/*Final merge*/ \
r8 = vec_mergeh(r0, r4); /*all set 0*/ \
r9 = vec_mergel(r0, r4); /*all set 1*/ \
r10 = vec_mergeh(r1, r5); /*all set 2*/ \
r11 = vec_mergel(r1, r5); /*all set 3*/ \
r12 = vec_mergeh(r2, r6); /*all set 4*/ \
r13 = vec_mergel(r2, r6); /*all set 5*/ \
/* Don't need to compute 14 and 15*/ \
\
}
// out: o = |x-y| < a
static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
register vec_u8 y,
register vec_u8 a) {
register vec_u8 diff = vec_subs(x, y);
register vec_u8 diffneg = vec_subs(y, x);
register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
o = (vec_u8)vec_cmplt(o, a);
return o;
}
static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
register vec_u8 p1,
register vec_u8 q0,
register vec_u8 q1,
register vec_u8 alpha,
register vec_u8 beta) {
register vec_u8 mask;
register vec_u8 tempmask;
mask = diff_lt_altivec(p0, q0, alpha);
tempmask = diff_lt_altivec(p1, p0, beta);
mask = vec_and(mask, tempmask);
tempmask = diff_lt_altivec(q1, q0, beta);
mask = vec_and(mask, tempmask);
return mask;
}
// out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
register vec_u8 p1,
register vec_u8 p2,
register vec_u8 q0,
register vec_u8 tc0) {
register vec_u8 average = vec_avg(p0, q0);
register vec_u8 temp;
register vec_u8 unclipped;
register vec_u8 ones;
register vec_u8 max;
register vec_u8 min;
register vec_u8 newp1;
temp = vec_xor(average, p2);
average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */
ones = vec_splat_u8(1);
temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */
unclipped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
max = vec_adds(p1, tc0);
min = vec_subs(p1, tc0);
newp1 = vec_max(min, unclipped);
newp1 = vec_min(max, newp1);
return newp1;
}
#define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) { \
\
const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4)); \
\
register vec_u8 pq0bit = vec_xor(p0,q0); \
register vec_u8 q1minus; \
register vec_u8 p0minus; \
register vec_u8 stage1; \
register vec_u8 stage2; \
register vec_u8 vec160; \
register vec_u8 delta; \
register vec_u8 deltaneg; \
\
q1minus = vec_nor(q1, q1); /* 255 - q1 */ \
stage1 = vec_avg(p1, q1minus); /* (p1 - q1 + 256)>>1 */ \
stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */ \
p0minus = vec_nor(p0, p0); /* 255 - p0 */ \
stage1 = vec_avg(q0, p0minus); /* (q0 - p0 + 256)>>1 */ \
pq0bit = vec_and(pq0bit, vec_splat_u8(1)); \
stage2 = vec_avg(stage2, pq0bit); /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
stage2 = vec_adds(stage2, stage1); /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
vec160 = vec_ld(0, &A0v); \
deltaneg = vec_subs(vec160, stage2); /* -d */ \
delta = vec_subs(stage2, vec160); /* d */ \
deltaneg = vec_min(tc0masked, deltaneg); \
delta = vec_min(tc0masked, delta); \
p0 = vec_subs(p0, deltaneg); \
q0 = vec_subs(q0, delta); \
p0 = vec_adds(p0, delta); \
q0 = vec_adds(q0, deltaneg); \
}
#define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) { \
DECLARE_ALIGNED(16, unsigned char, temp)[16]; \
register vec_u8 alphavec; \
register vec_u8 betavec; \
register vec_u8 mask; \
register vec_u8 p1mask; \
register vec_u8 q1mask; \
register vector signed char tc0vec; \
register vec_u8 finaltc0; \
register vec_u8 tc0masked; \
register vec_u8 newp1; \
register vec_u8 newq1; \
\
temp[0] = alpha; \
temp[1] = beta; \
alphavec = vec_ld(0, temp); \
betavec = vec_splat(alphavec, 0x1); \
alphavec = vec_splat(alphavec, 0x0); \
mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */ \
\
AV_COPY32(temp, tc0); \
tc0vec = vec_ld(0, (signed char*)temp); \
tc0vec = vec_mergeh(tc0vec, tc0vec); \
tc0vec = vec_mergeh(tc0vec, tc0vec); \
mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1))); /* if tc0[i] >= 0 */ \
finaltc0 = vec_and((vec_u8)tc0vec, mask); /* tc = tc0 */ \
\
p1mask = diff_lt_altivec(p2, p0, betavec); \
p1mask = vec_and(p1mask, mask); /* if ( |p2 - p0| < beta) */ \
tc0masked = vec_and(p1mask, (vec_u8)tc0vec); \
finaltc0 = vec_sub(finaltc0, p1mask); /* tc++ */ \
newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked); \
/*end if*/ \
\
q1mask = diff_lt_altivec(q2, q0, betavec); \
q1mask = vec_and(q1mask, mask); /* if ( |q2 - q0| < beta ) */\
tc0masked = vec_and(q1mask, (vec_u8)tc0vec); \
finaltc0 = vec_sub(finaltc0, q1mask); /* tc++ */ \
newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked); \
/*end if*/ \
\
h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0); \
p1 = newp1; \
q1 = newq1; \
}
static void h264_v_loop_filter_luma_altivec(uint8_t *pix, ptrdiff_t stride, int alpha, int beta, int8_t *tc0) {
if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
register vec_u8 p2 = vec_ld(-3*stride, pix);
register vec_u8 p1 = vec_ld(-2*stride, pix);
register vec_u8 p0 = vec_ld(-1*stride, pix);
register vec_u8 q0 = vec_ld(0, pix);
register vec_u8 q1 = vec_ld(stride, pix);
register vec_u8 q2 = vec_ld(2*stride, pix);
h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
vec_st(p1, -2*stride, pix);
vec_st(p0, -1*stride, pix);
vec_st(q0, 0, pix);
vec_st(q1, stride, pix);
}
}
static void h264_h_loop_filter_luma_altivec(uint8_t *pix, ptrdiff_t stride, int alpha, int beta, int8_t *tc0) {
register vec_u8 line0, line1, line2, line3, line4, line5;
if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
return;
readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
transpose4x16(line1, line2, line3, line4);
write16x4(pix-2, stride, line1, line2, line3, line4);
}
static av_always_inline
void weight_h264_W_altivec(uint8_t *block, int stride, int height,
int log2_denom, int weight, int offset, int w)
{
int y, aligned;
vec_u8 vblock;
vec_s16 vtemp, vweight, voffset, v0, v1;
vec_u16 vlog2_denom;
DECLARE_ALIGNED(16, int32_t, temp)[4];
LOAD_ZERO;
offset *= 1 << log2_denom;
if(log2_denom) offset += 1<<(log2_denom-1);
temp[0] = log2_denom;
temp[1] = weight;
temp[2] = offset;
vtemp = (vec_s16)vec_ld(0, temp);
#if !HAVE_BIGENDIAN
vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
#endif
vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
vweight = vec_splat(vtemp, 3);
voffset = vec_splat(vtemp, 5);
aligned = !((unsigned long)block & 0xf);
for (y = 0; y < height; y++) {
vblock = vec_ld(0, block);
v0 = (vec_s16)VEC_MERGEH(zero_u8v, vblock);
v1 = (vec_s16)VEC_MERGEL(zero_u8v, vblock);
if (w == 16 || aligned) {
v0 = vec_mladd(v0, vweight, zero_s16v);
v0 = vec_adds(v0, voffset);
v0 = vec_sra(v0, vlog2_denom);
}
if (w == 16 || !aligned) {
v1 = vec_mladd(v1, vweight, zero_s16v);
v1 = vec_adds(v1, voffset);
v1 = vec_sra(v1, vlog2_denom);
}
vblock = vec_packsu(v0, v1);
vec_st(vblock, 0, block);
block += stride;
}
}
static av_always_inline
void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height,
int log2_denom, int weightd, int weights, int offset, int w)
{
int y, dst_aligned, src_aligned;
vec_u8 vsrc, vdst;
vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
vec_u16 vlog2_denom;
DECLARE_ALIGNED(16, int32_t, temp)[4];
LOAD_ZERO;
offset = ((offset + 1) | 1) * (1 << log2_denom);
temp[0] = log2_denom+1;
temp[1] = weights;
temp[2] = weightd;
temp[3] = offset;
vtemp = (vec_s16)vec_ld(0, temp);
#if !HAVE_BIGENDIAN
vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
#endif
vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
vweights = vec_splat(vtemp, 3);
vweightd = vec_splat(vtemp, 5);
voffset = vec_splat(vtemp, 7);
dst_aligned = !((unsigned long)dst & 0xf);
src_aligned = !((unsigned long)src & 0xf);
for (y = 0; y < height; y++) {
vdst = vec_ld(0, dst);
vsrc = vec_ld(0, src);
v0 = (vec_s16)VEC_MERGEH(zero_u8v, vdst);
v1 = (vec_s16)VEC_MERGEL(zero_u8v, vdst);
v2 = (vec_s16)VEC_MERGEH(zero_u8v, vsrc);
v3 = (vec_s16)VEC_MERGEL(zero_u8v, vsrc);
if (w == 8) {
if (src_aligned)
v3 = v2;
else
v2 = v3;
}
if (w == 16 || dst_aligned) {
v0 = vec_mladd(v0, vweightd, zero_s16v);
v2 = vec_mladd(v2, vweights, zero_s16v);
v0 = vec_adds(v0, voffset);
v0 = vec_adds(v0, v2);
v0 = vec_sra(v0, vlog2_denom);
}
if (w == 16 || !dst_aligned) {
v1 = vec_mladd(v1, vweightd, zero_s16v);
v3 = vec_mladd(v3, vweights, zero_s16v);
v1 = vec_adds(v1, voffset);
v1 = vec_adds(v1, v3);
v1 = vec_sra(v1, vlog2_denom);
}
vdst = vec_packsu(v0, v1);
vec_st(vdst, 0, dst);
dst += stride;
src += stride;
}
}
#define H264_WEIGHT(W) \
static void weight_h264_pixels ## W ## _altivec(uint8_t *block, ptrdiff_t stride, int height, \
int log2_denom, int weight, int offset) \
{ \
weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
}\
static void biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, ptrdiff_t stride, int height, \
int log2_denom, int weightd, int weights, int offset) \
{ \
biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
}
H264_WEIGHT(16)
H264_WEIGHT( 8)
#endif /* HAVE_ALTIVEC */
av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth,
const int chroma_format_idc)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (bit_depth == 8) {
c->h264_idct_add = h264_idct_add_altivec;
if (chroma_format_idc <= 1)
c->h264_idct_add8 = h264_idct_add8_altivec;
c->h264_idct_add16 = h264_idct_add16_altivec;
c->h264_idct_add16intra = h264_idct_add16intra_altivec;
c->h264_idct_dc_add= h264_idct_dc_add_altivec;
c->h264_idct8_dc_add = h264_idct8_dc_add_altivec;
c->h264_idct8_add = h264_idct8_add_altivec;
c->h264_idct8_add4 = h264_idct8_add4_altivec;
c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
c->weight_h264_pixels_tab[0] = weight_h264_pixels16_altivec;
c->weight_h264_pixels_tab[1] = weight_h264_pixels8_altivec;
c->biweight_h264_pixels_tab[0] = biweight_h264_pixels16_altivec;
c->biweight_h264_pixels_tab[1] = biweight_h264_pixels8_altivec;
}
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/h264qpel.c
+315
View File
@@ -0,0 +1,315 @@
/*
* Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/intreadwrite.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/h264qpel.h"
#include "hpeldsp_altivec.h"
#if HAVE_ALTIVEC
#define PUT_OP_U8_ALTIVEC(d, s, dst) d = s
#define AVG_OP_U8_ALTIVEC(d, s, dst) d = vec_avg(dst, s)
#define OP_U8_ALTIVEC PUT_OP_U8_ALTIVEC
#define PREFIX_h264_qpel16_h_lowpass_altivec put_h264_qpel16_h_lowpass_altivec
#define PREFIX_h264_qpel16_h_lowpass_num altivec_put_h264_qpel16_h_lowpass_num
#define PREFIX_h264_qpel16_v_lowpass_altivec put_h264_qpel16_v_lowpass_altivec
#define PREFIX_h264_qpel16_v_lowpass_num altivec_put_h264_qpel16_v_lowpass_num
#define PREFIX_h264_qpel16_hv_lowpass_altivec put_h264_qpel16_hv_lowpass_altivec
#define PREFIX_h264_qpel16_hv_lowpass_num altivec_put_h264_qpel16_hv_lowpass_num
#include "h264qpel_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_h264_qpel16_h_lowpass_altivec
#undef PREFIX_h264_qpel16_h_lowpass_num
#undef PREFIX_h264_qpel16_v_lowpass_altivec
#undef PREFIX_h264_qpel16_v_lowpass_num
#undef PREFIX_h264_qpel16_hv_lowpass_altivec
#undef PREFIX_h264_qpel16_hv_lowpass_num
#define OP_U8_ALTIVEC AVG_OP_U8_ALTIVEC
#define PREFIX_h264_qpel16_h_lowpass_altivec avg_h264_qpel16_h_lowpass_altivec
#define PREFIX_h264_qpel16_h_lowpass_num altivec_avg_h264_qpel16_h_lowpass_num
#define PREFIX_h264_qpel16_v_lowpass_altivec avg_h264_qpel16_v_lowpass_altivec
#define PREFIX_h264_qpel16_v_lowpass_num altivec_avg_h264_qpel16_v_lowpass_num
#define PREFIX_h264_qpel16_hv_lowpass_altivec avg_h264_qpel16_hv_lowpass_altivec
#define PREFIX_h264_qpel16_hv_lowpass_num altivec_avg_h264_qpel16_hv_lowpass_num
#include "h264qpel_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_h264_qpel16_h_lowpass_altivec
#undef PREFIX_h264_qpel16_h_lowpass_num
#undef PREFIX_h264_qpel16_v_lowpass_altivec
#undef PREFIX_h264_qpel16_v_lowpass_num
#undef PREFIX_h264_qpel16_hv_lowpass_altivec
#undef PREFIX_h264_qpel16_hv_lowpass_num
#define H264_MC(OPNAME, SIZE, CODETYPE) \
static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## CODETYPE (uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
ff_ ## OPNAME ## pixels ## SIZE ## _ ## CODETYPE(dst, src, stride, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{ \
DECLARE_ALIGNED(16, uint8_t, half)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(dst, src, stride, stride);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, half)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(half, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+1, half, stride, stride, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, half)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src, half, stride, stride, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(dst, src, stride, stride);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, half)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(half, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, src+stride, half, stride, stride, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, int16_t, tmp)[SIZE*(SIZE+8)];\
OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(dst, tmp, src, stride, SIZE, stride);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfHV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, int16_t, tmp)[SIZE*(SIZE+8)];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src, SIZE, stride);\
put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfH)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfHV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, int16_t, tmp)[SIZE*(SIZE+8)];\
put_h264_qpel ## SIZE ## _h_lowpass_ ## CODETYPE(halfH, src + stride, SIZE, stride);\
put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfH, halfHV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfHV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, int16_t, tmp)[SIZE*(SIZE+8)];\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src, SIZE, stride);\
put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\
}\
\
static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## CODETYPE(uint8_t *dst, const uint8_t *src, ptrdiff_t stride)\
{\
DECLARE_ALIGNED(16, uint8_t, halfV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, uint8_t, halfHV)[SIZE*SIZE];\
DECLARE_ALIGNED(16, int16_t, tmp)[SIZE*(SIZE+8)];\
put_h264_qpel ## SIZE ## _v_lowpass_ ## CODETYPE(halfV, src+1, SIZE, stride);\
put_h264_qpel ## SIZE ## _hv_lowpass_ ## CODETYPE(halfHV, tmp, src, SIZE, SIZE, stride);\
OPNAME ## pixels ## SIZE ## _l2_ ## CODETYPE(dst, halfV, halfHV, stride, SIZE, SIZE);\
}\
#if HAVE_BIGENDIAN
#define put_unligned_store(s, dest) { \
tmp1 = vec_ld(0, dest); \
mask = vec_lvsl(0, dest); \
tmp2 = vec_ld(15, dest); \
edges = vec_perm(tmp2, tmp1, mask); \
align = vec_lvsr(0, dest); \
tmp2 = vec_perm(s, edges, align); \
tmp1 = vec_perm(edges, s, align); \
vec_st(tmp2, 15, dest); \
vec_st(tmp1, 0 , dest); \
}
#else
#define put_unligned_store(s, dest) vec_vsx_st(s, 0, dest);
#endif /* HAVE_BIGENDIAN */
static inline void put_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1,
const uint8_t * src2, int dst_stride,
int src_stride1, int h)
{
int i;
vec_u8 a, b, d, mask_;
#if HAVE_BIGENDIAN
vec_u8 tmp1, tmp2, mask, edges, align;
mask_ = vec_lvsl(0, src2);
#endif
for (i = 0; i < h; i++) {
a = unaligned_load(i * src_stride1, src1);
b = load_with_perm_vec(i * 16, src2, mask_);
d = vec_avg(a, b);
put_unligned_store(d, dst);
dst += dst_stride;
}
}
#if HAVE_BIGENDIAN
#define avg_unligned_store(s, dest){ \
tmp1 = vec_ld(0, dest); \
mask = vec_lvsl(0, dest); \
tmp2 = vec_ld(15, dest); \
a = vec_avg(vec_perm(tmp1, tmp2, mask), s); \
edges = vec_perm(tmp2, tmp1, mask); \
align = vec_lvsr(0, dest); \
tmp2 = vec_perm(a, edges, align); \
tmp1 = vec_perm(edges, a, align); \
vec_st(tmp2, 15, dest); \
vec_st(tmp1, 0 , dest); \
}
#else
#define avg_unligned_store(s, dest){ \
a = vec_avg(vec_vsx_ld(0, dst), s); \
vec_vsx_st(a, 0, dst); \
}
#endif /* HAVE_BIGENDIAN */
static inline void avg_pixels16_l2_altivec( uint8_t * dst, const uint8_t * src1,
const uint8_t * src2, int dst_stride,
int src_stride1, int h)
{
int i;
vec_u8 a, b, d, mask_;
#if HAVE_BIGENDIAN
vec_u8 tmp1, tmp2, mask, edges, align;
mask_ = vec_lvsl(0, src2);
#endif
for (i = 0; i < h; i++) {
a = unaligned_load(i * src_stride1, src1);
b = load_with_perm_vec(i * 16, src2, mask_);
d = vec_avg(a, b);
avg_unligned_store(d, dst);
dst += dst_stride;
}
}
/* Implemented but could be faster
#define put_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) put_pixels16_l2(d,s1,s2,ds,s1s,16,h)
#define avg_pixels16_l2_altivec(d,s1,s2,ds,s1s,h) avg_pixels16_l2(d,s1,s2,ds,s1s,16,h)
*/
H264_MC(put_, 16, altivec)
H264_MC(avg_, 16, altivec)
#endif /* HAVE_ALTIVEC */
av_cold void ff_h264qpel_init_ppc(H264QpelContext *c, int bit_depth)
{
#if HAVE_ALTIVEC
const int high_bit_depth = bit_depth > 8;
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (!high_bit_depth) {
#define dspfunc(PFX, IDX, NUM) \
c->PFX ## _pixels_tab[IDX][ 0] = PFX ## NUM ## _mc00_altivec; \
c->PFX ## _pixels_tab[IDX][ 1] = PFX ## NUM ## _mc10_altivec; \
c->PFX ## _pixels_tab[IDX][ 2] = PFX ## NUM ## _mc20_altivec; \
c->PFX ## _pixels_tab[IDX][ 3] = PFX ## NUM ## _mc30_altivec; \
c->PFX ## _pixels_tab[IDX][ 4] = PFX ## NUM ## _mc01_altivec; \
c->PFX ## _pixels_tab[IDX][ 5] = PFX ## NUM ## _mc11_altivec; \
c->PFX ## _pixels_tab[IDX][ 6] = PFX ## NUM ## _mc21_altivec; \
c->PFX ## _pixels_tab[IDX][ 7] = PFX ## NUM ## _mc31_altivec; \
c->PFX ## _pixels_tab[IDX][ 8] = PFX ## NUM ## _mc02_altivec; \
c->PFX ## _pixels_tab[IDX][ 9] = PFX ## NUM ## _mc12_altivec; \
c->PFX ## _pixels_tab[IDX][10] = PFX ## NUM ## _mc22_altivec; \
c->PFX ## _pixels_tab[IDX][11] = PFX ## NUM ## _mc32_altivec; \
c->PFX ## _pixels_tab[IDX][12] = PFX ## NUM ## _mc03_altivec; \
c->PFX ## _pixels_tab[IDX][13] = PFX ## NUM ## _mc13_altivec; \
c->PFX ## _pixels_tab[IDX][14] = PFX ## NUM ## _mc23_altivec; \
c->PFX ## _pixels_tab[IDX][15] = PFX ## NUM ## _mc33_altivec
dspfunc(put_h264_qpel, 0, 16);
dspfunc(avg_h264_qpel, 0, 16);
#undef dspfunc
}
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/h264qpel_template.c
+483
View File
@@ -0,0 +1,483 @@
/*
* Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#if HAVE_UNISTD_H
#include <unistd.h>
#endif
#include "libavutil/avassert.h"
#include "libavutil/mem.h"
#include "libavutil/ppc/util_altivec.h"
#define ASSERT_ALIGNED(ptr) av_assert2(!((uintptr_t)ptr&0x0000000F));
#if HAVE_BIGENDIAN
#define load_alignment(s, ali, pm2, pm1, pp0, pp1, pp2, pp3){\
vec_u8 srcR1 = vec_ld(-2, s);\
vec_u8 srcR2 = vec_ld(14, s);\
switch (ali) {\
default: {\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = vec_perm(srcR1, srcR2, pm1);\
srcP0 = vec_perm(srcR1, srcR2, pp0);\
srcP1 = vec_perm(srcR1, srcR2, pp1);\
srcP2 = vec_perm(srcR1, srcR2, pp2);\
srcP3 = vec_perm(srcR1, srcR2, pp3);\
} break;\
case 11: {\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = vec_perm(srcR1, srcR2, pm1);\
srcP0 = vec_perm(srcR1, srcR2, pp0);\
srcP1 = vec_perm(srcR1, srcR2, pp1);\
srcP2 = vec_perm(srcR1, srcR2, pp2);\
srcP3 = srcR2;\
} break;\
case 12: {\
vec_u8 srcR3 = vec_ld(30, s);\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = vec_perm(srcR1, srcR2, pm1);\
srcP0 = vec_perm(srcR1, srcR2, pp0);\
srcP1 = vec_perm(srcR1, srcR2, pp1);\
srcP2 = srcR2;\
srcP3 = vec_perm(srcR2, srcR3, pp3);\
} break;\
case 13: {\
vec_u8 srcR3 = vec_ld(30, s);\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = vec_perm(srcR1, srcR2, pm1);\
srcP0 = vec_perm(srcR1, srcR2, pp0);\
srcP1 = srcR2;\
srcP2 = vec_perm(srcR2, srcR3, pp2);\
srcP3 = vec_perm(srcR2, srcR3, pp3);\
} break;\
case 14: {\
vec_u8 srcR3 = vec_ld(30, s);\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = vec_perm(srcR1, srcR2, pm1);\
srcP0 = srcR2;\
srcP1 = vec_perm(srcR2, srcR3, pp1);\
srcP2 = vec_perm(srcR2, srcR3, pp2);\
srcP3 = vec_perm(srcR2, srcR3, pp3);\
} break;\
case 15: {\
vec_u8 srcR3 = vec_ld(30, s);\
srcM2 = vec_perm(srcR1, srcR2, pm2);\
srcM1 = srcR2;\
srcP0 = vec_perm(srcR2, srcR3, pp0);\
srcP1 = vec_perm(srcR2, srcR3, pp1);\
srcP2 = vec_perm(srcR2, srcR3, pp2);\
srcP3 = vec_perm(srcR2, srcR3, pp3);\
} break;\
}\
}
#else
#define load_alignment(s, ali, pm2, pm1, pp0, pp1, pp2, pp3){\
srcM2 = vec_vsx_ld(-2, s);\
srcM1 = vec_vsx_ld(-1, s);\
srcP0 = vec_vsx_ld(0, s);\
srcP1 = vec_vsx_ld(1, s);\
srcP2 = vec_vsx_ld(2, s);\
srcP3 = vec_vsx_ld(3, s);\
}
#endif /* HAVE_BIGENDIAN */
/* this code assume stride % 16 == 0 */
#ifdef PREFIX_h264_qpel16_h_lowpass_altivec
static void PREFIX_h264_qpel16_h_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 permM2, permM1, permP0, permP1, permP2, permP3;
const vec_s16 v5ss = vec_splat_s16(5);
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 srcP0A, srcP0B, srcP1A, srcP1B,
srcP2A, srcP2B, srcP3A, srcP3B,
srcM1A, srcM1B, srcM2A, srcM2B,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB;
vec_u8 sum, fsum;
#if HAVE_BIGENDIAN
permM2 = vec_lvsl(-2, src);
permM1 = vec_lvsl(-1, src);
permP0 = vec_lvsl(+0, src);
permP1 = vec_lvsl(+1, src);
permP2 = vec_lvsl(+2, src);
permP3 = vec_lvsl(+3, src);
#endif /* HAVE_BIGENDIAN */
for (i = 0 ; i < 16 ; i ++) {
load_alignment(src, align, permM2, permM1, permP0, permP1, permP2, permP3);
srcP0A = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
srcP0B = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
srcP1A = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
srcP1B = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
srcP2A = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
srcP2B = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
srcP3A = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3B = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
srcM1A = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
srcM1B = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
srcM2A = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
srcM2B = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
sum1A = vec_adds(srcP0A, srcP1A);
sum1B = vec_adds(srcP0B, srcP1B);
sum2A = vec_adds(srcM1A, srcP2A);
sum2B = vec_adds(srcM1B, srcP2B);
sum3A = vec_adds(srcM2A, srcP3A);
sum3B = vec_adds(srcM2B, srcP3B);
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
src += srcStride;
dst += dstStride;
}
}
#endif /* PREFIX_h264_qpel16_h_lowpass_altivec */
/* this code assume stride % 16 == 0 */
#ifdef PREFIX_h264_qpel16_v_lowpass_altivec
static void PREFIX_h264_qpel16_v_lowpass_altivec(uint8_t *dst,
const uint8_t *src,
int dstStride, int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 perm;
#if HAVE_BIGENDIAN
perm = vec_lvsl(0, src);
#endif
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u16 v5us = vec_splat_u16(5);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v16ss = vec_sl(vec_splat_s16(1),vec_splat_u16(4));
const uint8_t *srcbis = src - (srcStride * 2);
const vec_u8 srcM2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcM1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP0 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP1 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
const vec_u8 srcP2 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
vec_s16 srcM2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
vec_s16 srcM2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
vec_s16 srcM1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
vec_s16 srcM1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
vec_s16 srcP0ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
vec_s16 srcP0ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
vec_s16 srcP1ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
vec_s16 srcP1ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
vec_s16 srcP2ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
vec_s16 srcP2ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
vec_s16 pp1A, pp1B, pp2A, pp2B, pp3A, pp3B,
psumA, psumB, sumA, sumB,
srcP3ssA, srcP3ssB,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B;
vec_u8 sum, fsum, srcP3;
for (i = 0 ; i < 16 ; i++) {
srcP3 = load_with_perm_vec(0, srcbis, perm);
srcbis += srcStride;
srcP3ssA = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3ssB = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
sum1A = vec_adds(srcP0ssA, srcP1ssA);
sum1B = vec_adds(srcP0ssB, srcP1ssB);
sum2A = vec_adds(srcM1ssA, srcP2ssA);
sum2B = vec_adds(srcM1ssB, srcP2ssB);
sum3A = vec_adds(srcM2ssA, srcP3ssA);
sum3B = vec_adds(srcM2ssB, srcP3ssB);
srcM2ssA = srcM1ssA;
srcM2ssB = srcM1ssB;
srcM1ssA = srcP0ssA;
srcM1ssB = srcP0ssB;
srcP0ssA = srcP1ssA;
srcP0ssB = srcP1ssB;
srcP1ssA = srcP2ssA;
srcP1ssB = srcP2ssB;
srcP2ssA = srcP3ssA;
srcP2ssB = srcP3ssB;
pp1A = vec_mladd(sum1A, v20ss, v16ss);
pp1B = vec_mladd(sum1B, v20ss, v16ss);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
pp3A = vec_add(sum3A, pp1A);
pp3B = vec_add(sum3B, pp1B);
psumA = vec_sub(pp3A, pp2A);
psumB = vec_sub(pp3B, pp2B);
sumA = vec_sra(psumA, v5us);
sumB = vec_sra(psumB, v5us);
sum = vec_packsu(sumA, sumB);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
#endif /* PREFIX_h264_qpel16_v_lowpass_altivec */
/* this code assume stride % 16 == 0 *and* tmp is properly aligned */
#ifdef PREFIX_h264_qpel16_hv_lowpass_altivec
static void PREFIX_h264_qpel16_hv_lowpass_altivec(uint8_t *dst, int16_t *tmp,
const uint8_t *src,
int dstStride, int tmpStride,
int srcStride)
{
register int i;
LOAD_ZERO;
vec_u8 permM2, permM1, permP0, permP1, permP2, permP3;
const vec_s16 v20ss = vec_sl(vec_splat_s16(5),vec_splat_u16(2));
const vec_u32 v10ui = vec_splat_u32(10);
const vec_s16 v5ss = vec_splat_s16(5);
const vec_s16 v1ss = vec_splat_s16(1);
const vec_s32 v512si = vec_sl(vec_splat_s32(1),vec_splat_u32(9));
const vec_u32 v16ui = vec_sl(vec_splat_u32(1),vec_splat_u32(4));
register int align = ((((unsigned long)src) - 2) % 16);
vec_s16 srcP0A, srcP0B, srcP1A, srcP1B,
srcP2A, srcP2B, srcP3A, srcP3B,
srcM1A, srcM1B, srcM2A, srcM2B,
sum1A, sum1B, sum2A, sum2B, sum3A, sum3B,
pp1A, pp1B, pp2A, pp2B, psumA, psumB;
const vec_u8 mperm = (const vec_u8)
{0x00, 0x08, 0x01, 0x09, 0x02, 0x0A, 0x03, 0x0B,
0x04, 0x0C, 0x05, 0x0D, 0x06, 0x0E, 0x07, 0x0F};
int16_t *tmpbis = tmp;
vec_s16 tmpM1ssA, tmpM1ssB, tmpM2ssA, tmpM2ssB,
tmpP0ssA, tmpP0ssB, tmpP1ssA, tmpP1ssB,
tmpP2ssA, tmpP2ssB;
vec_s32 pp1Ae, pp1Ao, pp1Be, pp1Bo, pp2Ae, pp2Ao, pp2Be, pp2Bo,
pp3Ae, pp3Ao, pp3Be, pp3Bo, pp1cAe, pp1cAo, pp1cBe, pp1cBo,
pp32Ae, pp32Ao, pp32Be, pp32Bo, sumAe, sumAo, sumBe, sumBo,
ssumAe, ssumAo, ssumBe, ssumBo;
vec_u8 fsum, sumv, sum;
vec_s16 ssume, ssumo;
#if HAVE_BIGENDIAN
permM2 = vec_lvsl(-2, src);
permM1 = vec_lvsl(-1, src);
permP0 = vec_lvsl(+0, src);
permP1 = vec_lvsl(+1, src);
permP2 = vec_lvsl(+2, src);
permP3 = vec_lvsl(+3, src);
#endif /* HAVE_BIGENDIAN */
src -= (2 * srcStride);
for (i = 0 ; i < 21 ; i ++) {
vec_u8 srcM2, srcM1, srcP0, srcP1, srcP2, srcP3;
load_alignment(src, align, permM2, permM1, permP0, permP1, permP2, permP3);
srcP0A = (vec_s16) VEC_MERGEH(zero_u8v, srcP0);
srcP0B = (vec_s16) VEC_MERGEL(zero_u8v, srcP0);
srcP1A = (vec_s16) VEC_MERGEH(zero_u8v, srcP1);
srcP1B = (vec_s16) VEC_MERGEL(zero_u8v, srcP1);
srcP2A = (vec_s16) VEC_MERGEH(zero_u8v, srcP2);
srcP2B = (vec_s16) VEC_MERGEL(zero_u8v, srcP2);
srcP3A = (vec_s16) VEC_MERGEH(zero_u8v, srcP3);
srcP3B = (vec_s16) VEC_MERGEL(zero_u8v, srcP3);
srcM1A = (vec_s16) VEC_MERGEH(zero_u8v, srcM1);
srcM1B = (vec_s16) VEC_MERGEL(zero_u8v, srcM1);
srcM2A = (vec_s16) VEC_MERGEH(zero_u8v, srcM2);
srcM2B = (vec_s16) VEC_MERGEL(zero_u8v, srcM2);
sum1A = vec_adds(srcP0A, srcP1A);
sum1B = vec_adds(srcP0B, srcP1B);
sum2A = vec_adds(srcM1A, srcP2A);
sum2B = vec_adds(srcM1B, srcP2B);
sum3A = vec_adds(srcM2A, srcP3A);
sum3B = vec_adds(srcM2B, srcP3B);
pp1A = vec_mladd(sum1A, v20ss, sum3A);
pp1B = vec_mladd(sum1B, v20ss, sum3B);
pp2A = vec_mladd(sum2A, v5ss, zero_s16v);
pp2B = vec_mladd(sum2B, v5ss, zero_s16v);
psumA = vec_sub(pp1A, pp2A);
psumB = vec_sub(pp1B, pp2B);
vec_st(psumA, 0, tmp);
vec_st(psumB, 16, tmp);
src += srcStride;
tmp += tmpStride; /* int16_t*, and stride is 16, so it's OK here */
}
tmpM2ssA = vec_ld(0, tmpbis);
tmpM2ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpM1ssA = vec_ld(0, tmpbis);
tmpM1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP0ssA = vec_ld(0, tmpbis);
tmpP0ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP1ssA = vec_ld(0, tmpbis);
tmpP1ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
tmpP2ssA = vec_ld(0, tmpbis);
tmpP2ssB = vec_ld(16, tmpbis);
tmpbis += tmpStride;
for (i = 0 ; i < 16 ; i++) {
const vec_s16 tmpP3ssA = vec_ld(0, tmpbis);
const vec_s16 tmpP3ssB = vec_ld(16, tmpbis);
const vec_s16 sum1A = vec_adds(tmpP0ssA, tmpP1ssA);
const vec_s16 sum1B = vec_adds(tmpP0ssB, tmpP1ssB);
const vec_s16 sum2A = vec_adds(tmpM1ssA, tmpP2ssA);
const vec_s16 sum2B = vec_adds(tmpM1ssB, tmpP2ssB);
vec_s16 sum3A = vec_adds(tmpM2ssA, tmpP3ssA);
vec_s16 sum3B = vec_adds(tmpM2ssB, tmpP3ssB);
tmpbis += tmpStride;
tmpM2ssA = tmpM1ssA;
tmpM2ssB = tmpM1ssB;
tmpM1ssA = tmpP0ssA;
tmpM1ssB = tmpP0ssB;
tmpP0ssA = tmpP1ssA;
tmpP0ssB = tmpP1ssB;
tmpP1ssA = tmpP2ssA;
tmpP1ssB = tmpP2ssB;
tmpP2ssA = tmpP3ssA;
tmpP2ssB = tmpP3ssB;
pp1Ae = vec_mule(sum1A, v20ss);
pp1Ao = vec_mulo(sum1A, v20ss);
pp1Be = vec_mule(sum1B, v20ss);
pp1Bo = vec_mulo(sum1B, v20ss);
pp2Ae = vec_mule(sum2A, v5ss);
pp2Ao = vec_mulo(sum2A, v5ss);
pp2Be = vec_mule(sum2B, v5ss);
pp2Bo = vec_mulo(sum2B, v5ss);
pp3Ao = vec_mulo(sum3A, v1ss);
pp3Bo = vec_mulo(sum3B, v1ss);
#if !HAVE_BIGENDIAN
sum3A = (vec_s16)vec_perm(sum3A, sum3A,vcswapi2s(0,1,2,3));
sum3B = (vec_s16)vec_perm(sum3B, sum3B,vcswapi2s(0,1,2,3));
#endif
pp3Ae = vec_sra((vec_s32)sum3A, v16ui);
pp3Be = vec_sra((vec_s32)sum3B, v16ui);
pp1cAe = vec_add(pp1Ae, v512si);
pp1cAo = vec_add(pp1Ao, v512si);
pp1cBe = vec_add(pp1Be, v512si);
pp1cBo = vec_add(pp1Bo, v512si);
pp32Ae = vec_sub(pp3Ae, pp2Ae);
pp32Ao = vec_sub(pp3Ao, pp2Ao);
pp32Be = vec_sub(pp3Be, pp2Be);
pp32Bo = vec_sub(pp3Bo, pp2Bo);
sumAe = vec_add(pp1cAe, pp32Ae);
sumAo = vec_add(pp1cAo, pp32Ao);
sumBe = vec_add(pp1cBe, pp32Be);
sumBo = vec_add(pp1cBo, pp32Bo);
ssumAe = vec_sra(sumAe, v10ui);
ssumAo = vec_sra(sumAo, v10ui);
ssumBe = vec_sra(sumBe, v10ui);
ssumBo = vec_sra(sumBo, v10ui);
ssume = vec_packs(ssumAe, ssumBe);
ssumo = vec_packs(ssumAo, ssumBo);
sumv = vec_packsu(ssume, ssumo);
sum = vec_perm(sumv, sumv, mperm);
ASSERT_ALIGNED(dst);
OP_U8_ALTIVEC(fsum, sum, vec_ld(0, dst));
vec_st(fsum, 0, dst);
dst += dstStride;
}
}
#endif /* PREFIX_h264_qpel16_hv_lowpass_altivec */
ffmpeg/libavcodec/ppc/hevcdsp.c
+121
View File
@@ -0,0 +1,121 @@
/*
* SIMD-optimized IDCT functions for HEVC decoding
* Copyright (c) Alexandra Hajkova
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/hevc/dsp.h"
#if HAVE_ALTIVEC
static const vec_s16 trans4[4] = {
{ 64, 64, 64, 64, 64, 64, 64, 64 },
{ 83, 36, 83, 36, 83, 36, 83, 36 },
{ 64, -64, 64, -64, 64, -64, 64, -64 },
{ 36, -83, 36, -83, 36, -83, 36, -83 },
};
static const vec_u8 mask[2] = {
{ 0x00, 0x01, 0x08, 0x09, 0x10, 0x11, 0x18, 0x19, 0x02, 0x03, 0x0A, 0x0B, 0x12, 0x13, 0x1A, 0x1B },
{ 0x04, 0x05, 0x0C, 0x0D, 0x14, 0x15, 0x1C, 0x1D, 0x06, 0x07, 0x0E, 0x0F, 0x16, 0x17, 0x1E, 0x1F },
};
static av_always_inline void transform4x4(vec_s16 src_01, vec_s16 src_23,
vec_s32 res[4], const int shift,
int16_t *coeffs)
{
vec_s16 src_02, src_13;
vec_s32 zero = vec_splat_s32(0);
vec_s32 e0, o0, e1, o1;
vec_s32 add;
src_13 = vec_mergel(src_01, src_23);
src_02 = vec_mergeh(src_01, src_23);
e0 = vec_msums(src_02, trans4[0], zero);
o0 = vec_msums(src_13, trans4[1], zero);
e1 = vec_msums(src_02, trans4[2], zero);
o1 = vec_msums(src_13, trans4[3], zero);
switch(shift) {
case 7: add = vec_sl(vec_splat_s32(1), vec_splat_u32( 7 - 1)); break;
case 10: add = vec_sl(vec_splat_s32(1), vec_splat_u32(10 - 1)); break;
case 12: add = vec_sl(vec_splat_s32(1), vec_splat_u32(12 - 1)); break;
default: abort();
}
e0 = vec_add(e0, add);
e1 = vec_add(e1, add);
res[0] = vec_add(e0, o0);
res[1] = vec_add(e1, o1);
res[2] = vec_sub(e1, o1);
res[3] = vec_sub(e0, o0);
}
static av_always_inline void scale(vec_s32 res[4], vec_s16 res_packed[2],
const int shift)
{
int i;
vec_u32 v_shift;
switch(shift) {
case 7: v_shift = vec_splat_u32(7) ; break;
case 10: v_shift = vec_splat_u32(10); break;
case 12: v_shift = vec_splat_u32(12); break;
default: abort();
}
for (i = 0; i < 4; i++)
res[i] = vec_sra(res[i], v_shift);
// clip16
res_packed[0] = vec_packs(res[0], res[1]);
res_packed[1] = vec_packs(res[2], res[3]);
}
#define FUNCDECL(a, depth) a ## _ ## depth ## _altivec
#define FUNC(a, b) FUNCDECL(a, b)
#define BIT_DEPTH 8
#include "hevcdsp_template.c"
#undef BIT_DEPTH
#define BIT_DEPTH 10
#include "hevcdsp_template.c"
#undef BIT_DEPTH
#endif /* HAVE_ALTIVEC */
av_cold void ff_hevc_dsp_init_ppc(HEVCDSPContext *c, const int bit_depth)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (bit_depth == 8)
c->idct[0] = ff_hevc_idct_4x4_8_altivec;
if (bit_depth == 10)
c->idct[0] = ff_hevc_idct_4x4_10_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/hevcdsp_template.c
+48
View File
@@ -0,0 +1,48 @@
/*
* Copyright (c) Alexandra Hajkova
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
static void FUNC(ff_hevc_idct_4x4, BIT_DEPTH)(int16_t *coeffs, int col_limit)
{
const int shift = 7;
const int shift2 = 20 - BIT_DEPTH;
vec_s16 src_01, src_23;
vec_s32 res[4];
vec_s16 res_packed[2];
src_01 = vec_ld(0, coeffs);
src_23 = vec_ld(16, coeffs);
transform4x4(src_01, src_23, res, shift, coeffs);
src_01 = vec_packs(res[0], res[1]);
src_23 = vec_packs(res[2], res[3]);
scale(res, res_packed, shift);
// transpose
src_01 = vec_perm(res_packed[0], res_packed[1], mask[0]);
src_23 = vec_perm(res_packed[0], res_packed[1], mask[1]);
transform4x4(src_01, src_23, res, shift2, coeffs);
scale(res, res_packed, shift2);
// transpose
src_01 = vec_perm(res_packed[0], res_packed[1], mask[0]);
src_23 = vec_perm(res_packed[0], res_packed[1], mask[1]);
vec_st(src_01, 0, coeffs);
vec_st(src_23, 16, coeffs);
}
ffmpeg/libavcodec/ppc/hpeldsp_altivec.c
+386
View File
@@ -0,0 +1,386 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/hpeldsp.h"
#include "hpeldsp_altivec.h"
#if HAVE_ALTIVEC
/* next one assumes that ((line_size % 16) == 0) */
void ff_put_pixels16_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register vector unsigned char pixelsv1;
register vector unsigned char pixelsv1B;
register vector unsigned char pixelsv1C;
register vector unsigned char pixelsv1D;
int i;
register ptrdiff_t line_size_2 = line_size * (1 << 1);
register ptrdiff_t line_size_3 = line_size + line_size_2;
register ptrdiff_t line_size_4 = line_size * (1 << 2);
// hand-unrolling the loop by 4 gains about 15%
// minimum execution time goes from 74 to 60 cycles
// it's faster than -funroll-loops, but using
// -funroll-loops w/ this is bad - 74 cycles again.
// all this is on a 7450, tuning for the 7450
for (i = 0; i < h; i += 4) {
pixelsv1 = unaligned_load( 0, pixels);
pixelsv1B = unaligned_load(line_size, pixels);
pixelsv1C = unaligned_load(line_size_2, pixels);
pixelsv1D = unaligned_load(line_size_3, pixels);
VEC_ST(pixelsv1, 0, (unsigned char*)block);
VEC_ST(pixelsv1B, line_size, (unsigned char*)block);
VEC_ST(pixelsv1C, line_size_2, (unsigned char*)block);
VEC_ST(pixelsv1D, line_size_3, (unsigned char*)block);
pixels+=line_size_4;
block +=line_size_4;
}
}
/* next one assumes that ((line_size % 16) == 0) */
#define op_avg(a,b) a = ( ((a)|(b)) - ((((a)^(b))&0xFEFEFEFEUL)>>1) )
void ff_avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register vector unsigned char pixelsv, blockv;
int i;
for (i = 0; i < h; i++) {
blockv = vec_ld(0, block);
pixelsv = VEC_LD( 0, pixels);
blockv = vec_avg(blockv,pixelsv);
vec_st(blockv, 0, (unsigned char*)block);
pixels+=line_size;
block +=line_size;
}
}
/* next one assumes that ((line_size % 8) == 0) */
static void avg_pixels8_altivec(uint8_t * block, const uint8_t * pixels, ptrdiff_t line_size, int h)
{
register vector unsigned char pixelsv, blockv;
int i;
for (i = 0; i < h; i++) {
/* block is 8 bytes-aligned, so we're either in the
left block (16 bytes-aligned) or in the right block (not) */
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
pixelsv = VEC_LD( 0, pixels);
if (rightside) {
pixelsv = vec_perm(blockv, pixelsv, vcprm(0,1,s0,s1));
} else {
pixelsv = vec_perm(blockv, pixelsv, vcprm(s0,s1,2,3));
}
blockv = vec_avg(blockv, pixelsv);
vec_st(blockv, 0, block);
pixels += line_size;
block += line_size;
}
}
/* next one assumes that ((line_size % 8) == 0) */
static void put_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
pixelsv1 = VEC_LD(0, pixels);
pixelsv2 = VEC_LD(1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vctwo);
for (i = 0; i < h ; i++) {
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
pixelsv1 = unaligned_load(line_size, pixels);
pixelsv2 = unaligned_load(line_size+1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum1 = vec_add(pixelssum2, vctwo);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside) {
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
} else {
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
vec_st(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
/* next one assumes that ((line_size % 8) == 0) */
static void put_no_rnd_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
pixelsv1 = VEC_LD(0, pixels);
pixelsv2 = VEC_LD(1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vcone);
for (i = 0; i < h ; i++) {
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
pixelsv1 = unaligned_load(line_size, pixels);
pixelsv2 = unaligned_load(line_size+1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum1 = vec_add(pixelssum2, vcone);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside) {
blockv = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
} else {
blockv = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
vec_st(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
/* next one assumes that ((line_size % 16) == 0) */
static void put_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char blockv;
register vector unsigned short temp3, temp4,
pixelssum1, pixelssum2, pixelssum3, pixelssum4;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
pixelsv1 = VEC_LD(0, pixels);
pixelsv2 = VEC_LD(1, pixels);
pixelsv3 = VEC_MERGEL(vczero, pixelsv1);
pixelsv4 = VEC_MERGEL(vczero, pixelsv2);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum3 = vec_add((vector unsigned short)pixelsv3,
(vector unsigned short)pixelsv4);
pixelssum3 = vec_add(pixelssum3, vctwo);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vctwo);
for (i = 0; i < h ; i++) {
blockv = vec_ld(0, block);
pixelsv1 = unaligned_load(line_size, pixels);
pixelsv2 = unaligned_load(line_size+1, pixels);
pixelsv3 = VEC_MERGEL(vczero, pixelsv1);
pixelsv4 = VEC_MERGEL(vczero, pixelsv2);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum4 = vec_add((vector unsigned short)pixelsv3,
(vector unsigned short)pixelsv4);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp4 = vec_add(pixelssum3, pixelssum4);
temp4 = vec_sra(temp4, vctwo);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum3 = vec_add(pixelssum4, vctwo);
pixelssum1 = vec_add(pixelssum2, vctwo);
blockv = vec_packsu(temp3, temp4);
vec_st(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
/* next one assumes that ((line_size % 16) == 0) */
static void put_no_rnd_pixels16_xy2_altivec(uint8_t * block, const uint8_t * pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsv3, pixelsv4;
register vector unsigned char blockv;
register vector unsigned short temp3, temp4,
pixelssum1, pixelssum2, pixelssum3, pixelssum4;
register const vector unsigned char vczero = (const vector unsigned char)vec_splat_u8(0);
register const vector unsigned short vcone = (const vector unsigned short)vec_splat_u16(1);
register const vector unsigned short vctwo = (const vector unsigned short)vec_splat_u16(2);
pixelsv1 = VEC_LD(0, pixels);
pixelsv2 = VEC_LD(1, pixels);
pixelsv3 = VEC_MERGEL(vczero, pixelsv1);
pixelsv4 = VEC_MERGEL(vczero, pixelsv2);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum3 = vec_add((vector unsigned short)pixelsv3,
(vector unsigned short)pixelsv4);
pixelssum3 = vec_add(pixelssum3, vcone);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vcone);
for (i = 0; i < h ; i++) {
pixelsv1 = unaligned_load(line_size, pixels);
pixelsv2 = unaligned_load(line_size+1, pixels);
pixelsv3 = VEC_MERGEL(vczero, pixelsv1);
pixelsv4 = VEC_MERGEL(vczero, pixelsv2);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum4 = vec_add((vector unsigned short)pixelsv3,
(vector unsigned short)pixelsv4);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp4 = vec_add(pixelssum3, pixelssum4);
temp4 = vec_sra(temp4, vctwo);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum3 = vec_add(pixelssum4, vcone);
pixelssum1 = vec_add(pixelssum2, vcone);
blockv = vec_packsu(temp3, temp4);
VEC_ST(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
/* next one assumes that ((line_size % 8) == 0) */
static void avg_pixels8_xy2_altivec(uint8_t *block, const uint8_t *pixels, ptrdiff_t line_size, int h)
{
register int i;
register vector unsigned char pixelsv1, pixelsv2, pixelsavg;
register vector unsigned char blockv, blocktemp;
register vector unsigned short pixelssum1, pixelssum2, temp3;
register const vector unsigned char vczero = (const vector unsigned char)
vec_splat_u8(0);
register const vector unsigned short vctwo = (const vector unsigned short)
vec_splat_u16(2);
pixelsv1 = VEC_LD(0, pixels);
pixelsv2 = VEC_LD(1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum1 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
pixelssum1 = vec_add(pixelssum1, vctwo);
for (i = 0; i < h ; i++) {
int rightside = ((unsigned long)block & 0x0000000F);
blockv = vec_ld(0, block);
pixelsv1 = unaligned_load(line_size, pixels);
pixelsv2 = unaligned_load(line_size+1, pixels);
pixelsv1 = VEC_MERGEH(vczero, pixelsv1);
pixelsv2 = VEC_MERGEH(vczero, pixelsv2);
pixelssum2 = vec_add((vector unsigned short)pixelsv1,
(vector unsigned short)pixelsv2);
temp3 = vec_add(pixelssum1, pixelssum2);
temp3 = vec_sra(temp3, vctwo);
pixelssum1 = vec_add(pixelssum2, vctwo);
pixelsavg = vec_packsu(temp3, (vector unsigned short) vczero);
if (rightside) {
blocktemp = vec_perm(blockv, pixelsavg, vcprm(0, 1, s0, s1));
} else {
blocktemp = vec_perm(blockv, pixelsavg, vcprm(s0, s1, 2, 3));
}
blockv = vec_avg(blocktemp, blockv);
vec_st(blockv, 0, block);
block += line_size;
pixels += line_size;
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_hpeldsp_init_ppc(HpelDSPContext *c, int flags)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->avg_pixels_tab[0][0] = ff_avg_pixels16_altivec;
c->avg_pixels_tab[1][0] = avg_pixels8_altivec;
c->avg_pixels_tab[1][3] = avg_pixels8_xy2_altivec;
c->put_pixels_tab[0][0] = ff_put_pixels16_altivec;
c->put_pixels_tab[1][3] = put_pixels8_xy2_altivec;
c->put_pixels_tab[0][3] = put_pixels16_xy2_altivec;
c->put_no_rnd_pixels_tab[0][0] = ff_put_pixels16_altivec;
c->put_no_rnd_pixels_tab[1][3] = put_no_rnd_pixels8_xy2_altivec;
c->put_no_rnd_pixels_tab[0][3] = put_no_rnd_pixels16_xy2_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/hpeldsp_altivec.h
+34
View File
@@ -0,0 +1,34 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVCODEC_PPC_HPELDSP_ALTIVEC_H
#define AVCODEC_PPC_HPELDSP_ALTIVEC_H
#include <stddef.h>
#include <stdint.h>
void ff_avg_pixels16_altivec(uint8_t *block, const uint8_t *pixels,
ptrdiff_t line_size, int h);
void ff_put_pixels16_altivec(uint8_t *block, const uint8_t *pixels,
ptrdiff_t line_size, int h);
#endif /* AVCODEC_PPC_HPELDSP_ALTIVEC_H */
ffmpeg/libavcodec/ppc/idctdsp.c
+273
View File
@@ -0,0 +1,273 @@
/*
* Copyright (c) 2001 Michel Lespinasse
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/* NOTE: This code is based on GPL code from the libmpeg2 project. The
* author, Michel Lespinasses, has given explicit permission to release
* under LGPL as part of FFmpeg.
*
* FFmpeg integration by Dieter Shirley
*
* This file is a direct copy of the AltiVec IDCT module from the libmpeg2
* project. I've deleted all of the libmpeg2-specific code, renamed the
* functions and reordered the function parameters. The only change to the
* IDCT function itself was to factor out the partial transposition, and to
* perform a full transpose at the end of the function. */
#include "config.h"
#include <stdlib.h>
#include <string.h>
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/idctdsp.h"
#if HAVE_ALTIVEC
#define IDCT_HALF \
/* 1st stage */ \
t1 = vec_mradds(a1, vx7, vx1); \
t8 = vec_mradds(a1, vx1, vec_subs(zero, vx7)); \
t7 = vec_mradds(a2, vx5, vx3); \
t3 = vec_mradds(ma2, vx3, vx5); \
\
/* 2nd stage */ \
t5 = vec_adds(vx0, vx4); \
t0 = vec_subs(vx0, vx4); \
t2 = vec_mradds(a0, vx6, vx2); \
t4 = vec_mradds(a0, vx2, vec_subs(zero, vx6)); \
t6 = vec_adds(t8, t3); \
t3 = vec_subs(t8, t3); \
t8 = vec_subs(t1, t7); \
t1 = vec_adds(t1, t7); \
\
/* 3rd stage */ \
t7 = vec_adds(t5, t2); \
t2 = vec_subs(t5, t2); \
t5 = vec_adds(t0, t4); \
t0 = vec_subs(t0, t4); \
t4 = vec_subs(t8, t3); \
t3 = vec_adds(t8, t3); \
\
/* 4th stage */ \
vy0 = vec_adds(t7, t1); \
vy7 = vec_subs(t7, t1); \
vy1 = vec_mradds(c4, t3, t5); \
vy6 = vec_mradds(mc4, t3, t5); \
vy2 = vec_mradds(c4, t4, t0); \
vy5 = vec_mradds(mc4, t4, t0); \
vy3 = vec_adds(t2, t6); \
vy4 = vec_subs(t2, t6)
#define IDCT \
vec_s16 vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7; \
vec_s16 t0, t1, t2, t3, t4, t5, t6, t7, t8; \
\
vec_s16 c4 = vec_splat(constants[0], 0); \
vec_s16 a0 = vec_splat(constants[0], 1); \
vec_s16 a1 = vec_splat(constants[0], 2); \
vec_s16 a2 = vec_splat(constants[0], 3); \
vec_s16 mc4 = vec_splat(constants[0], 4); \
vec_s16 ma2 = vec_splat(constants[0], 5); \
vec_s16 bias = (vec_s16) vec_splat((vec_s32) constants[0], 3); \
\
vec_s16 zero = vec_splat_s16(0); \
vec_u16 shift = vec_splat_u16(4); \
\
vec_s16 vx0 = vec_mradds(vec_sl(block[0], shift), constants[1], zero); \
vec_s16 vx1 = vec_mradds(vec_sl(block[1], shift), constants[2], zero); \
vec_s16 vx2 = vec_mradds(vec_sl(block[2], shift), constants[3], zero); \
vec_s16 vx3 = vec_mradds(vec_sl(block[3], shift), constants[4], zero); \
vec_s16 vx4 = vec_mradds(vec_sl(block[4], shift), constants[1], zero); \
vec_s16 vx5 = vec_mradds(vec_sl(block[5], shift), constants[4], zero); \
vec_s16 vx6 = vec_mradds(vec_sl(block[6], shift), constants[3], zero); \
vec_s16 vx7 = vec_mradds(vec_sl(block[7], shift), constants[2], zero); \
\
IDCT_HALF; \
\
vx0 = vec_mergeh(vy0, vy4); \
vx1 = vec_mergel(vy0, vy4); \
vx2 = vec_mergeh(vy1, vy5); \
vx3 = vec_mergel(vy1, vy5); \
vx4 = vec_mergeh(vy2, vy6); \
vx5 = vec_mergel(vy2, vy6); \
vx6 = vec_mergeh(vy3, vy7); \
vx7 = vec_mergel(vy3, vy7); \
\
vy0 = vec_mergeh(vx0, vx4); \
vy1 = vec_mergel(vx0, vx4); \
vy2 = vec_mergeh(vx1, vx5); \
vy3 = vec_mergel(vx1, vx5); \
vy4 = vec_mergeh(vx2, vx6); \
vy5 = vec_mergel(vx2, vx6); \
vy6 = vec_mergeh(vx3, vx7); \
vy7 = vec_mergel(vx3, vx7); \
\
vx0 = vec_adds(vec_mergeh(vy0, vy4), bias); \
vx1 = vec_mergel(vy0, vy4); \
vx2 = vec_mergeh(vy1, vy5); \
vx3 = vec_mergel(vy1, vy5); \
vx4 = vec_mergeh(vy2, vy6); \
vx5 = vec_mergel(vy2, vy6); \
vx6 = vec_mergeh(vy3, vy7); \
vx7 = vec_mergel(vy3, vy7); \
\
IDCT_HALF; \
\
shift = vec_splat_u16(6); \
vx0 = vec_sra(vy0, shift); \
vx1 = vec_sra(vy1, shift); \
vx2 = vec_sra(vy2, shift); \
vx3 = vec_sra(vy3, shift); \
vx4 = vec_sra(vy4, shift); \
vx5 = vec_sra(vy5, shift); \
vx6 = vec_sra(vy6, shift); \
vx7 = vec_sra(vy7, shift)
static const vec_s16 constants[5] = {
{ 23170, 13573, 6518, 21895, -23170, -21895, 32, 31 },
{ 16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725 },
{ 22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521 },
{ 21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692 },
{ 19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722 }
};
static void idct_altivec(int16_t *blk)
{
vec_s16 *block = (vec_s16 *) blk;
IDCT;
block[0] = vx0;
block[1] = vx1;
block[2] = vx2;
block[3] = vx3;
block[4] = vx4;
block[5] = vx5;
block[6] = vx6;
block[7] = vx7;
}
static void idct_put_altivec(uint8_t *dest, ptrdiff_t stride, int16_t *blk)
{
vec_s16 *block = (vec_s16 *) blk;
vec_u8 tmp;
IDCT;
#define COPY(dest, src) \
tmp = vec_packsu(src, src); \
vec_ste((vec_u32) tmp, 0, (unsigned int *) dest); \
vec_ste((vec_u32) tmp, 4, (unsigned int *) dest)
COPY(dest, vx0);
dest += stride;
COPY(dest, vx1);
dest += stride;
COPY(dest, vx2);
dest += stride;
COPY(dest, vx3);
dest += stride;
COPY(dest, vx4);
dest += stride;
COPY(dest, vx5);
dest += stride;
COPY(dest, vx6);
dest += stride;
COPY(dest, vx7);
}
static void idct_add_altivec(uint8_t *dest, ptrdiff_t stride, int16_t *blk)
{
vec_s16 *block = (vec_s16 *) blk;
vec_u8 tmp;
vec_s16 tmp2, tmp3;
vec_u8 perm0;
vec_u8 perm1;
vec_u8 p0, p1, p;
IDCT;
#if HAVE_BIGENDIAN
p0 = vec_lvsl(0, dest);
p1 = vec_lvsl(stride, dest);
p = vec_splat_u8(-1);
perm0 = vec_mergeh(p, p0);
perm1 = vec_mergeh(p, p1);
#endif
#if HAVE_BIGENDIAN
#define GET_TMP2(dest, prm) \
tmp = vec_ld(0, dest); \
tmp2 = (vec_s16) vec_perm(tmp, (vec_u8) zero, prm);
#else
#define GET_TMP2(dest, prm) \
tmp = vec_vsx_ld(0, dest); \
tmp2 = (vec_s16) vec_mergeh(tmp, (vec_u8) zero)
#endif
#define ADD(dest, src, perm) \
GET_TMP2(dest, perm); \
tmp3 = vec_adds(tmp2, src); \
tmp = vec_packsu(tmp3, tmp3); \
vec_ste((vec_u32) tmp, 0, (unsigned int *) dest); \
vec_ste((vec_u32) tmp, 4, (unsigned int *) dest)
ADD(dest, vx0, perm0);
dest += stride;
ADD(dest, vx1, perm1);
dest += stride;
ADD(dest, vx2, perm0);
dest += stride;
ADD(dest, vx3, perm1);
dest += stride;
ADD(dest, vx4, perm0);
dest += stride;
ADD(dest, vx5, perm1);
dest += stride;
ADD(dest, vx6, perm0);
dest += stride;
ADD(dest, vx7, perm1);
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_idctdsp_init_ppc(IDCTDSPContext *c, AVCodecContext *avctx,
unsigned high_bit_depth)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
if (!high_bit_depth && avctx->lowres == 0) {
if ((avctx->idct_algo == FF_IDCT_AUTO && !(avctx->flags & AV_CODEC_FLAG_BITEXACT)) ||
(avctx->idct_algo == FF_IDCT_ALTIVEC)) {
c->idct = idct_altivec;
c->idct_add = idct_add_altivec;
c->idct_put = idct_put_altivec;
c->perm_type = FF_IDCT_PERM_TRANSPOSE;
}
}
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/lossless_audiodsp_altivec.c
+91
View File
@@ -0,0 +1,91 @@
/*
* Copyright (c) 2007 Luca Barbato <lu_zero@gentoo.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/lossless_audiodsp.h"
#if HAVE_BIGENDIAN
#define GET_T(tt0,tt1,src,a,b){ \
a = vec_ld(16, src); \
tt0 = vec_perm(b, a, align); \
b = vec_ld(32, src); \
tt1 = vec_perm(a, b, align); \
}
#else
#define GET_T(tt0,tt1,src,a,b){ \
tt0 = vec_vsx_ld(0, src); \
tt1 = vec_vsx_ld(16, src); \
}
#endif
#if HAVE_ALTIVEC
static int32_t scalarproduct_and_madd_int16_altivec(int16_t *v1,
const int16_t *v2,
const int16_t *v3,
int order, int mul)
{
LOAD_ZERO;
vec_s16 *pv1 = (vec_s16 *) v1;
register vec_s16 muls = { mul, mul, mul, mul, mul, mul, mul, mul };
register vec_s16 t0, t1, i0, i1, i4, i2, i3;
register vec_s32 res = zero_s32v;
#if HAVE_BIGENDIAN
register vec_u8 align = vec_lvsl(0, v2);
i2 = vec_ld(0, v2);
i3 = vec_ld(0, v3);
#endif
int32_t ires;
order >>= 4;
do {
GET_T(t0,t1,v2,i1,i2);
i0 = pv1[0];
i1 = pv1[1];
res = vec_msum(t0, i0, res);
res = vec_msum(t1, i1, res);
GET_T(t0,t1,v3,i4,i3);
pv1[0] = vec_mladd(t0, muls, i0);
pv1[1] = vec_mladd(t1, muls, i1);
pv1 += 2;
v2 += 16;
v3 += 16;
} while (--order);
res = vec_splat(vec_sums(res, zero_s32v), 3);
vec_ste(res, 0, &ires);
return ires;
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_llauddsp_init_ppc(LLAudDSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->scalarproduct_and_madd_int16 = scalarproduct_and_madd_int16_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/lossless_videodsp_altivec.c
+59
View File
@@ -0,0 +1,59 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/lossless_videodsp.h"
#if HAVE_ALTIVEC
static void add_bytes_altivec(uint8_t *dst, uint8_t *src, ptrdiff_t w)
{
register int i;
register vector unsigned char vdst, vsrc;
/* dst and src are 16 bytes-aligned (guaranteed). */
for (i = 0; i + 15 < w; i += 16) {
vdst = vec_ld(i, (unsigned char *) dst);
vsrc = vec_ld(i, (unsigned char *) src);
vdst = vec_add(vsrc, vdst);
vec_st(vdst, i, (unsigned char *) dst);
}
/* If w is not a multiple of 16. */
for (; i < w; i++)
dst[i] = src[i];
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_llviddsp_init_ppc(LLVidDSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->add_bytes = add_bytes_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/mathops.h
+79
View File
@@ -0,0 +1,79 @@
/*
* simple math operations
* Copyright (c) 2001, 2002 Fabrice Bellard
* Copyright (c) 2006 Michael Niedermayer <michaelni@gmx.at> et al
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#ifndef AVCODEC_PPC_MATHOPS_H
#define AVCODEC_PPC_MATHOPS_H
#include <stdint.h>
#include "config.h"
#include "libavutil/common.h"
#if HAVE_PPC4XX
/* signed 16x16 -> 32 multiply add accumulate */
#define MAC16(rt, ra, rb) \
__asm__ ("maclhw %0, %2, %3" : "=r" (rt) : "0" (rt), "r" (ra), "r" (rb));
/* signed 16x16 -> 32 multiply */
#define MUL16(ra, rb) \
({ int __rt; \
__asm__ ("mullhw %0, %1, %2" : "=r" (__rt) : "r" (ra), "r" (rb)); \
__rt; })
#endif
#define MULH MULH
static inline av_const int MULH(int a, int b){
int r;
__asm__ ("mulhw %0, %1, %2" : "=r"(r) : "r"(a), "r"(b));
return r;
}
#if !ARCH_PPC64
static inline av_const int64_t MAC64(int64_t d, int a, int b)
{
union { uint64_t x; unsigned hl[2]; } x = { d };
int h, l;
__asm__ ("mullw %3, %4, %5 \n\t"
"mulhw %2, %4, %5 \n\t"
"addc %1, %1, %3 \n\t"
"adde %0, %0, %2 \n\t"
: "+r"(x.hl[0]), "+r"(x.hl[1]), "=&r"(h), "=&r"(l)
: "r"(a), "r"(b));
return x.x;
}
#define MAC64(d, a, b) ((d) = MAC64(d, a, b))
static inline av_const int64_t MLS64(int64_t d, int a, int b)
{
union { uint64_t x; unsigned hl[2]; } x = { d };
int h, l;
__asm__ ("mullw %3, %4, %5 \n\t"
"mulhw %2, %4, %5 \n\t"
"subfc %1, %3, %1 \n\t"
"subfe %0, %2, %0 \n\t"
: "+r"(x.hl[0]), "+r"(x.hl[1]), "=&r"(h), "=&r"(l)
: "r"(a), "r"(b));
return x.x;
}
#define MLS64(d, a, b) ((d) = MLS64(d, a, b))
#endif
#endif /* AVCODEC_PPC_MATHOPS_H */
ffmpeg/libavcodec/ppc/me_cmp.c
+746
View File
@@ -0,0 +1,746 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/mpegvideo.h"
#include "libavcodec/me_cmp.h"
#if HAVE_ALTIVEC
#if HAVE_BIGENDIAN
#define GET_PERM(per1, per2, pix) {\
per1 = vec_lvsl(0, pix);\
per2 = vec_add(per1, vec_splat_u8(1));\
}
#define LOAD_PIX(v, iv, pix, per1, per2) {\
vector unsigned char pix2l = vec_ld(0, pix);\
vector unsigned char pix2r = vec_ld(16, pix);\
v = vec_perm(pix2l, pix2r, per1);\
iv = vec_perm(pix2l, pix2r, per2);\
}
#else
#define GET_PERM(per1, per2, pix) {}
#define LOAD_PIX(v, iv, pix, per1, per2) {\
v = vec_vsx_ld(0, pix);\
iv = vec_vsx_ld(1, pix);\
}
#endif
static int sad16_x2_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s = 0;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
vector unsigned char perm1, perm2, pix2v, pix2iv;
GET_PERM(perm1, perm2, pix2);
for (i = 0; i < h; i++) {
/* Read unaligned pixels into our vectors. The vectors are as follows:
* pix1v: pix1[0] - pix1[15]
* pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16] */
vector unsigned char pix1v = vec_ld(0, pix1);
LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
/* Calculate the average vector. */
vector unsigned char avgv = vec_avg(pix2v, pix2iv);
/* Calculate a sum of abs differences vector. */
vector unsigned char t5 = vec_sub(vec_max(pix1v, avgv),
vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
pix1 += stride;
pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad16_y2_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s = 0;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
vector unsigned char pix1v, pix3v, avgv, t5;
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
const uint8_t *pix3 = pix2 + stride;
/* Due to the fact that pix3 = pix2 + stride, the pix3 of one
* iteration becomes pix2 in the next iteration. We can use this
* fact to avoid a potentially expensive unaligned read, each
* time around the loop.
* Read unaligned pixels into our vectors. The vectors are as follows:
* pix2v: pix2[0] - pix2[15]
* Split the pixel vectors into shorts. */
vector unsigned char pix2v = VEC_LD(0, pix2);
for (i = 0; i < h; i++) {
/* Read unaligned pixels into our vectors. The vectors are as follows:
* pix1v: pix1[0] - pix1[15]
* pix3v: pix3[0] - pix3[15] */
pix1v = vec_ld(0, pix1);
pix3v = VEC_LD(0, pix3);
/* Calculate the average vector. */
avgv = vec_avg(pix2v, pix3v);
/* Calculate a sum of abs differences vector. */
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
pix1 += stride;
pix2v = pix3v;
pix3 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad16_xy2_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s = 0;
const uint8_t *pix3 = pix2 + stride;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
const vector unsigned short two =
(const vector unsigned short) vec_splat_u16(2);
vector unsigned char avgv, t5;
vector unsigned char pix1v, pix3v, pix3iv;
vector unsigned short pix3lv, pix3hv, pix3ilv, pix3ihv;
vector unsigned short avghv, avglv;
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
vector unsigned char perm1, perm2, pix2v, pix2iv;
GET_PERM(perm1, perm2, pix2);
/* Due to the fact that pix3 = pix2 + stride, the pix3 of one
* iteration becomes pix2 in the next iteration. We can use this
* fact to avoid a potentially expensive unaligned read, as well
* as some splitting, and vector addition each time around the loop.
* Read unaligned pixels into our vectors. The vectors are as follows:
* pix2v: pix2[0] - pix2[15] pix2iv: pix2[1] - pix2[16]
* Split the pixel vectors into shorts. */
LOAD_PIX(pix2v, pix2iv, pix2, perm1, perm2);
vector unsigned short pix2hv =
(vector unsigned short) VEC_MERGEH(zero, pix2v);
vector unsigned short pix2lv =
(vector unsigned short) VEC_MERGEL(zero, pix2v);
vector unsigned short pix2ihv =
(vector unsigned short) VEC_MERGEH(zero, pix2iv);
vector unsigned short pix2ilv =
(vector unsigned short) VEC_MERGEL(zero, pix2iv);
vector unsigned short t1 = vec_add(pix2hv, pix2ihv);
vector unsigned short t2 = vec_add(pix2lv, pix2ilv);
vector unsigned short t3, t4;
for (i = 0; i < h; i++) {
/* Read unaligned pixels into our vectors. The vectors are as follows:
* pix1v: pix1[0] - pix1[15]
* pix3v: pix3[0] - pix3[15] pix3iv: pix3[1] - pix3[16] */
pix1v = vec_ld(0, pix1);
LOAD_PIX(pix3v, pix3iv, pix3, perm1, perm2);
/* Note that AltiVec does have vec_avg, but this works on vector pairs
* and rounds up. We could do avg(avg(a, b), avg(c, d)), but the
* rounding would mean that, for example, avg(3, 0, 0, 1) = 2, when
* it should be 1. Instead, we have to split the pixel vectors into
* vectors of shorts and do the averaging by hand. */
/* Split the pixel vectors into shorts. */
pix3hv = (vector unsigned short) VEC_MERGEH(zero, pix3v);
pix3lv = (vector unsigned short) VEC_MERGEL(zero, pix3v);
pix3ihv = (vector unsigned short) VEC_MERGEH(zero, pix3iv);
pix3ilv = (vector unsigned short) VEC_MERGEL(zero, pix3iv);
/* Do the averaging on them. */
t3 = vec_add(pix3hv, pix3ihv);
t4 = vec_add(pix3lv, pix3ilv);
avghv = vec_sr(vec_add(vec_add(t1, t3), two), two);
avglv = vec_sr(vec_add(vec_add(t2, t4), two), two);
/* Pack the shorts back into a result. */
avgv = vec_pack(avghv, avglv);
/* Calculate a sum of abs differences vector. */
t5 = vec_sub(vec_max(pix1v, avgv), vec_min(pix1v, avgv));
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
pix1 += stride;
pix3 += stride;
/* Transfer the calculated values for pix3 into pix2. */
t1 = t3;
t2 = t4;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad16_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2. */
vector unsigned char t1 =vec_ld(0, pix1);
vector unsigned char t2 = VEC_LD(0, pix2);
/* Calculate a sum of abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
vector unsigned char t4 = vec_min(t1, t2);
vector unsigned char t5 = vec_sub(t3, t4);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
pix1 += stride;
pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
static int sad8_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
const vector unsigned char permclear =
(vector unsigned char)
{ 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2.
* Since we're reading 16 pixels, and actually only want 8,
* mask out the last 8 pixels. The 0s don't change the sum. */
vector unsigned char pix1l = VEC_LD(0, pix1);
vector unsigned char pix2l = VEC_LD(0, pix2);
vector unsigned char t1 = vec_and(pix1l, permclear);
vector unsigned char t2 = vec_and(pix2l, permclear);
/* Calculate a sum of abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
vector unsigned char t4 = vec_min(t1, t2);
vector unsigned char t5 = vec_sub(t3, t4);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t5, sad);
pix1 += stride;
pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
/* Sum of Squared Errors for an 8x8 block, AltiVec-enhanced.
* It's the sad8_altivec code above w/ squaring added. */
static int sse8_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
const vector unsigned char permclear =
(vector unsigned char)
{ 255, 255, 255, 255, 255, 255, 255, 255, 0, 0, 0, 0, 0, 0, 0, 0 };
vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
vector signed int sumsqr;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2.
* Since we're reading 16 pixels, and actually only want 8,
* mask out the last 8 pixels. The 0s don't change the sum. */
vector unsigned char t1 = vec_and(VEC_LD(0, pix1), permclear);
vector unsigned char t2 = vec_and(VEC_LD(0, pix2), permclear);
/* Since we want to use unsigned chars, we can take advantage
* of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
/* Calculate abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
vector unsigned char t4 = vec_min(t1, t2);
vector unsigned char t5 = vec_sub(t3, t4);
/* Square the values and add them to our sum. */
sum = vec_msum(t5, t5, sum);
pix1 += stride;
pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
sumsqr = vec_splat(sumsqr, 3);
vec_ste(sumsqr, 0, &s);
return s;
}
/* Sum of Squared Errors for a 16x16 block, AltiVec-enhanced.
* It's the sad16_altivec code above w/ squaring added. */
static int sse16_altivec(MPVEncContext *v, const uint8_t *pix1, const uint8_t *pix2,
ptrdiff_t stride, int h)
{
int i;
int __attribute__((aligned(16))) s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned int sum = (vector unsigned int) vec_splat_u32(0);
vector signed int sumsqr;
for (i = 0; i < h; i++) {
/* Read potentially unaligned pixels into t1 and t2. */
vector unsigned char t1 = vec_ld(0, pix1);
vector unsigned char t2 = VEC_LD(0, pix2);
/* Since we want to use unsigned chars, we can take advantage
* of the fact that abs(a - b) ^ 2 = (a - b) ^ 2. */
/* Calculate abs differences vector. */
vector unsigned char t3 = vec_max(t1, t2);
vector unsigned char t4 = vec_min(t1, t2);
vector unsigned char t5 = vec_sub(t3, t4);
/* Square the values and add them to our sum. */
sum = vec_msum(t5, t5, sum);
pix1 += stride;
pix2 += stride;
}
/* Sum up the four partial sums, and put the result into s. */
sumsqr = vec_sums((vector signed int) sum, (vector signed int) zero);
sumsqr = vec_splat(sumsqr, 3);
vec_ste(sumsqr, 0, &s);
return s;
}
static int hadamard8_diff8x8_altivec(MPVEncContext *s, const uint8_t *dst,
const uint8_t *src, ptrdiff_t stride, int h)
{
int __attribute__((aligned(16))) sum;
register const vector unsigned char vzero =
(const vector unsigned char) vec_splat_u8(0);
register vector signed short temp0, temp1, temp2, temp3, temp4,
temp5, temp6, temp7;
{
register const vector signed short vprod1 =
(const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
register const vector signed short vprod2 =
(const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
register const vector signed short vprod3 =
(const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
register const vector unsigned char perm1 =
(const vector unsigned char)
{ 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
register const vector unsigned char perm2 =
(const vector unsigned char)
{ 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
register const vector unsigned char perm3 =
(const vector unsigned char)
{ 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
#define ONEITERBUTTERFLY(i, res) \
{ \
register vector unsigned char srcO = unaligned_load(stride * i, src); \
register vector unsigned char dstO = unaligned_load(stride * i, dst);\
\
/* Promote the unsigned chars to signed shorts. */ \
/* We're in the 8x8 function, we only care for the first 8. */ \
register vector signed short srcV = \
(vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstV = \
(vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) dstO); \
\
/* subtractions inside the first butterfly */ \
register vector signed short but0 = vec_sub(srcV, dstV); \
register vector signed short op1 = vec_perm(but0, but0, perm1); \
register vector signed short but1 = vec_mladd(but0, vprod1, op1); \
register vector signed short op2 = vec_perm(but1, but1, perm2); \
register vector signed short but2 = vec_mladd(but1, vprod2, op2); \
register vector signed short op3 = vec_perm(but2, but2, perm3); \
res = vec_mladd(but2, vprod3, op3); \
}
ONEITERBUTTERFLY(0, temp0);
ONEITERBUTTERFLY(1, temp1);
ONEITERBUTTERFLY(2, temp2);
ONEITERBUTTERFLY(3, temp3);
ONEITERBUTTERFLY(4, temp4);
ONEITERBUTTERFLY(5, temp5);
ONEITERBUTTERFLY(6, temp6);
ONEITERBUTTERFLY(7, temp7);
}
#undef ONEITERBUTTERFLY
{
register vector signed int vsum;
register vector signed short line0 = vec_add(temp0, temp1);
register vector signed short line1 = vec_sub(temp0, temp1);
register vector signed short line2 = vec_add(temp2, temp3);
register vector signed short line3 = vec_sub(temp2, temp3);
register vector signed short line4 = vec_add(temp4, temp5);
register vector signed short line5 = vec_sub(temp4, temp5);
register vector signed short line6 = vec_add(temp6, temp7);
register vector signed short line7 = vec_sub(temp6, temp7);
register vector signed short line0B = vec_add(line0, line2);
register vector signed short line2B = vec_sub(line0, line2);
register vector signed short line1B = vec_add(line1, line3);
register vector signed short line3B = vec_sub(line1, line3);
register vector signed short line4B = vec_add(line4, line6);
register vector signed short line6B = vec_sub(line4, line6);
register vector signed short line5B = vec_add(line5, line7);
register vector signed short line7B = vec_sub(line5, line7);
register vector signed short line0C = vec_add(line0B, line4B);
register vector signed short line4C = vec_sub(line0B, line4B);
register vector signed short line1C = vec_add(line1B, line5B);
register vector signed short line5C = vec_sub(line1B, line5B);
register vector signed short line2C = vec_add(line2B, line6B);
register vector signed short line6C = vec_sub(line2B, line6B);
register vector signed short line3C = vec_add(line3B, line7B);
register vector signed short line7C = vec_sub(line3B, line7B);
vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
vsum = vec_sum4s(vec_abs(line1C), vsum);
vsum = vec_sum4s(vec_abs(line2C), vsum);
vsum = vec_sum4s(vec_abs(line3C), vsum);
vsum = vec_sum4s(vec_abs(line4C), vsum);
vsum = vec_sum4s(vec_abs(line5C), vsum);
vsum = vec_sum4s(vec_abs(line6C), vsum);
vsum = vec_sum4s(vec_abs(line7C), vsum);
vsum = vec_sums(vsum, (vector signed int) vzero);
vsum = vec_splat(vsum, 3);
vec_ste(vsum, 0, &sum);
}
return sum;
}
/*
* 16x8 works with 16 elements; it can avoid replicating loads, and
* gives the compiler more room for scheduling. It's only used from
* inside hadamard8_diff16_altivec.
*
* Unfortunately, it seems gcc-3.3 is a bit dumb, and the compiled code has
* a LOT of spill code, it seems gcc (unlike xlc) cannot keep everything in
* registers by itself. The following code includes hand-made register
* allocation. It's not clean, but on a 7450 the resulting code is much faster
* (best case falls from 700+ cycles to 550).
*
* xlc doesn't add spill code, but it doesn't know how to schedule for the
* 7450, and its code isn't much faster than gcc-3.3 on the 7450 (but uses
* 25% fewer instructions...)
*
* On the 970, the hand-made RA is still a win (around 690 vs. around 780),
* but xlc goes to around 660 on the regular C code...
*/
static int hadamard8_diff16x8_altivec(MPVEncContext *s, const uint8_t *dst,
const uint8_t *src, ptrdiff_t stride, int h)
{
int __attribute__((aligned(16))) sum;
register vector signed short
temp0 __asm__ ("v0"),
temp1 __asm__ ("v1"),
temp2 __asm__ ("v2"),
temp3 __asm__ ("v3"),
temp4 __asm__ ("v4"),
temp5 __asm__ ("v5"),
temp6 __asm__ ("v6"),
temp7 __asm__ ("v7");
register vector signed short
temp0S __asm__ ("v8"),
temp1S __asm__ ("v9"),
temp2S __asm__ ("v10"),
temp3S __asm__ ("v11"),
temp4S __asm__ ("v12"),
temp5S __asm__ ("v13"),
temp6S __asm__ ("v14"),
temp7S __asm__ ("v15");
register const vector unsigned char vzero __asm__ ("v31") =
(const vector unsigned char) vec_splat_u8(0);
{
register const vector signed short vprod1 __asm__ ("v16") =
(const vector signed short) { 1, -1, 1, -1, 1, -1, 1, -1 };
register const vector signed short vprod2 __asm__ ("v17") =
(const vector signed short) { 1, 1, -1, -1, 1, 1, -1, -1 };
register const vector signed short vprod3 __asm__ ("v18") =
(const vector signed short) { 1, 1, 1, 1, -1, -1, -1, -1 };
register const vector unsigned char perm1 __asm__ ("v19") =
(const vector unsigned char)
{ 0x02, 0x03, 0x00, 0x01, 0x06, 0x07, 0x04, 0x05,
0x0A, 0x0B, 0x08, 0x09, 0x0E, 0x0F, 0x0C, 0x0D };
register const vector unsigned char perm2 __asm__ ("v20") =
(const vector unsigned char)
{ 0x04, 0x05, 0x06, 0x07, 0x00, 0x01, 0x02, 0x03,
0x0C, 0x0D, 0x0E, 0x0F, 0x08, 0x09, 0x0A, 0x0B };
register const vector unsigned char perm3 __asm__ ("v21") =
(const vector unsigned char)
{ 0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F,
0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07 };
#define ONEITERBUTTERFLY(i, res1, res2) \
{ \
register vector unsigned char srcO __asm__ ("v22") = \
unaligned_load(stride * i, src); \
register vector unsigned char dstO __asm__ ("v23") = \
unaligned_load(stride * i, dst);\
\
/* Promote the unsigned chars to signed shorts. */ \
register vector signed short srcV __asm__ ("v24") = \
(vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstV __asm__ ("v25") = \
(vector signed short) VEC_MERGEH((vector signed char) vzero, \
(vector signed char) dstO); \
register vector signed short srcW __asm__ ("v26") = \
(vector signed short) VEC_MERGEL((vector signed char) vzero, \
(vector signed char) srcO); \
register vector signed short dstW __asm__ ("v27") = \
(vector signed short) VEC_MERGEL((vector signed char) vzero, \
(vector signed char) dstO); \
\
/* subtractions inside the first butterfly */ \
register vector signed short but0 __asm__ ("v28") = \
vec_sub(srcV, dstV); \
register vector signed short but0S __asm__ ("v29") = \
vec_sub(srcW, dstW); \
register vector signed short op1 __asm__ ("v30") = \
vec_perm(but0, but0, perm1); \
register vector signed short but1 __asm__ ("v22") = \
vec_mladd(but0, vprod1, op1); \
register vector signed short op1S __asm__ ("v23") = \
vec_perm(but0S, but0S, perm1); \
register vector signed short but1S __asm__ ("v24") = \
vec_mladd(but0S, vprod1, op1S); \
register vector signed short op2 __asm__ ("v25") = \
vec_perm(but1, but1, perm2); \
register vector signed short but2 __asm__ ("v26") = \
vec_mladd(but1, vprod2, op2); \
register vector signed short op2S __asm__ ("v27") = \
vec_perm(but1S, but1S, perm2); \
register vector signed short but2S __asm__ ("v28") = \
vec_mladd(but1S, vprod2, op2S); \
register vector signed short op3 __asm__ ("v29") = \
vec_perm(but2, but2, perm3); \
register vector signed short op3S __asm__ ("v30") = \
vec_perm(but2S, but2S, perm3); \
res1 = vec_mladd(but2, vprod3, op3); \
res2 = vec_mladd(but2S, vprod3, op3S); \
}
ONEITERBUTTERFLY(0, temp0, temp0S);
ONEITERBUTTERFLY(1, temp1, temp1S);
ONEITERBUTTERFLY(2, temp2, temp2S);
ONEITERBUTTERFLY(3, temp3, temp3S);
ONEITERBUTTERFLY(4, temp4, temp4S);
ONEITERBUTTERFLY(5, temp5, temp5S);
ONEITERBUTTERFLY(6, temp6, temp6S);
ONEITERBUTTERFLY(7, temp7, temp7S);
}
#undef ONEITERBUTTERFLY
{
register vector signed int vsum;
register vector signed short line0 = vec_add(temp0, temp1);
register vector signed short line1 = vec_sub(temp0, temp1);
register vector signed short line2 = vec_add(temp2, temp3);
register vector signed short line3 = vec_sub(temp2, temp3);
register vector signed short line4 = vec_add(temp4, temp5);
register vector signed short line5 = vec_sub(temp4, temp5);
register vector signed short line6 = vec_add(temp6, temp7);
register vector signed short line7 = vec_sub(temp6, temp7);
register vector signed short line0B = vec_add(line0, line2);
register vector signed short line2B = vec_sub(line0, line2);
register vector signed short line1B = vec_add(line1, line3);
register vector signed short line3B = vec_sub(line1, line3);
register vector signed short line4B = vec_add(line4, line6);
register vector signed short line6B = vec_sub(line4, line6);
register vector signed short line5B = vec_add(line5, line7);
register vector signed short line7B = vec_sub(line5, line7);
register vector signed short line0C = vec_add(line0B, line4B);
register vector signed short line4C = vec_sub(line0B, line4B);
register vector signed short line1C = vec_add(line1B, line5B);
register vector signed short line5C = vec_sub(line1B, line5B);
register vector signed short line2C = vec_add(line2B, line6B);
register vector signed short line6C = vec_sub(line2B, line6B);
register vector signed short line3C = vec_add(line3B, line7B);
register vector signed short line7C = vec_sub(line3B, line7B);
register vector signed short line0S = vec_add(temp0S, temp1S);
register vector signed short line1S = vec_sub(temp0S, temp1S);
register vector signed short line2S = vec_add(temp2S, temp3S);
register vector signed short line3S = vec_sub(temp2S, temp3S);
register vector signed short line4S = vec_add(temp4S, temp5S);
register vector signed short line5S = vec_sub(temp4S, temp5S);
register vector signed short line6S = vec_add(temp6S, temp7S);
register vector signed short line7S = vec_sub(temp6S, temp7S);
register vector signed short line0BS = vec_add(line0S, line2S);
register vector signed short line2BS = vec_sub(line0S, line2S);
register vector signed short line1BS = vec_add(line1S, line3S);
register vector signed short line3BS = vec_sub(line1S, line3S);
register vector signed short line4BS = vec_add(line4S, line6S);
register vector signed short line6BS = vec_sub(line4S, line6S);
register vector signed short line5BS = vec_add(line5S, line7S);
register vector signed short line7BS = vec_sub(line5S, line7S);
register vector signed short line0CS = vec_add(line0BS, line4BS);
register vector signed short line4CS = vec_sub(line0BS, line4BS);
register vector signed short line1CS = vec_add(line1BS, line5BS);
register vector signed short line5CS = vec_sub(line1BS, line5BS);
register vector signed short line2CS = vec_add(line2BS, line6BS);
register vector signed short line6CS = vec_sub(line2BS, line6BS);
register vector signed short line3CS = vec_add(line3BS, line7BS);
register vector signed short line7CS = vec_sub(line3BS, line7BS);
vsum = vec_sum4s(vec_abs(line0C), vec_splat_s32(0));
vsum = vec_sum4s(vec_abs(line1C), vsum);
vsum = vec_sum4s(vec_abs(line2C), vsum);
vsum = vec_sum4s(vec_abs(line3C), vsum);
vsum = vec_sum4s(vec_abs(line4C), vsum);
vsum = vec_sum4s(vec_abs(line5C), vsum);
vsum = vec_sum4s(vec_abs(line6C), vsum);
vsum = vec_sum4s(vec_abs(line7C), vsum);
vsum = vec_sum4s(vec_abs(line0CS), vsum);
vsum = vec_sum4s(vec_abs(line1CS), vsum);
vsum = vec_sum4s(vec_abs(line2CS), vsum);
vsum = vec_sum4s(vec_abs(line3CS), vsum);
vsum = vec_sum4s(vec_abs(line4CS), vsum);
vsum = vec_sum4s(vec_abs(line5CS), vsum);
vsum = vec_sum4s(vec_abs(line6CS), vsum);
vsum = vec_sum4s(vec_abs(line7CS), vsum);
vsum = vec_sums(vsum, (vector signed int) vzero);
vsum = vec_splat(vsum, 3);
vec_ste(vsum, 0, &sum);
}
return sum;
}
static int hadamard8_diff16_altivec(MPVEncContext *s, const uint8_t *dst,
const uint8_t *src, ptrdiff_t stride, int h)
{
int score = hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
if (h == 16) {
dst += 8 * stride;
src += 8 * stride;
score += hadamard8_diff16x8_altivec(s, dst, src, stride, 8);
}
return score;
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_me_cmp_init_ppc(MECmpContext *c, AVCodecContext *avctx)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->pix_abs[0][1] = sad16_x2_altivec;
c->pix_abs[0][2] = sad16_y2_altivec;
c->pix_abs[0][3] = sad16_xy2_altivec;
c->pix_abs[0][0] = sad16_altivec;
c->pix_abs[1][0] = sad8_altivec;
c->sad[0] = sad16_altivec;
c->sad[1] = sad8_altivec;
c->sse[0] = sse16_altivec;
c->sse[1] = sse8_altivec;
c->hadamard8_diff[0] = hadamard8_diff16_altivec;
c->hadamard8_diff[1] = hadamard8_diff8x8_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/mpeg4videodsp.c
+139
View File
@@ -0,0 +1,139 @@
/*
* GMC (Global Motion Compensation), AltiVec-enabled
*
* Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/mpeg4videodsp.h"
#if HAVE_ALTIVEC
/* AltiVec-enhanced gmc1. ATM this code assumes stride is a multiple of 8
* to preserve proper dst alignment. */
static void gmc1_altivec(uint8_t *dst /* align 8 */, const uint8_t *src /* align1 */,
int stride, int h, int x16, int y16, int rounder)
{
int i;
const DECLARE_ALIGNED(16, unsigned short, rounder_a) = rounder;
const DECLARE_ALIGNED(16, unsigned short, ABCD)[8] = {
(16 - x16) * (16 - y16), /* A */
(x16) * (16 - y16), /* B */
(16 - x16) * (y16), /* C */
(x16) * (y16), /* D */
0, 0, 0, 0 /* padding */
};
register const vector unsigned char vczero =
(const vector unsigned char) vec_splat_u8(0);
register const vector unsigned short vcsr8 =
(const vector unsigned short) vec_splat_u16(8);
register vector unsigned char dstv, dstv2, srcvB, srcvC, srcvD;
register vector unsigned short tempB, tempC, tempD;
unsigned long dst_odd = (unsigned long) dst & 0x0000000F;
unsigned long src_really_odd = (unsigned long) src & 0x0000000F;
register vector unsigned short tempA =
vec_ld(0, (const unsigned short *) ABCD);
register vector unsigned short Av = vec_splat(tempA, 0);
register vector unsigned short Bv = vec_splat(tempA, 1);
register vector unsigned short Cv = vec_splat(tempA, 2);
register vector unsigned short Dv = vec_splat(tempA, 3);
register vector unsigned short rounderV =
vec_splat((vec_u16) vec_lde(0, &rounder_a), 0);
/* we'll be able to pick-up our 9 char elements at src from those
* 32 bytes we load the first batch here, as inside the loop we can
* reuse 'src + stride' from one iteration as the 'src' of the next. */
register vector unsigned char src_0 = vec_ld(0, src);
register vector unsigned char src_1 = vec_ld(16, src);
register vector unsigned char srcvA = vec_perm(src_0, src_1,
vec_lvsl(0, src));
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvB = vec_perm(src_0, src_1, vec_lvsl(1, src));
else
srcvB = src_1;
srcvA = vec_mergeh(vczero, srcvA);
srcvB = vec_mergeh(vczero, srcvB);
for (i = 0; i < h; i++) {
dst_odd = (unsigned long) dst & 0x0000000F;
src_really_odd = (((unsigned long) src) + stride) & 0x0000000F;
dstv = vec_ld(0, dst);
/* We'll be able to pick-up our 9 char elements at src + stride from
* those 32 bytes then reuse the resulting 2 vectors srvcC and srcvD
* as the next srcvA and srcvB. */
src_0 = vec_ld(stride + 0, src);
src_1 = vec_ld(stride + 16, src);
srcvC = vec_perm(src_0, src_1, vec_lvsl(stride + 0, src));
if (src_really_odd != 0x0000000F)
/* If (src & 0xF) == 0xF, then (src + 1) is properly aligned
* on the second vector. */
srcvD = vec_perm(src_0, src_1, vec_lvsl(stride + 1, src));
else
srcvD = src_1;
srcvC = vec_mergeh(vczero, srcvC);
srcvD = vec_mergeh(vczero, srcvD);
/* OK, now we (finally) do the math :-)
* Those four instructions replace 32 int muls & 32 int adds.
* Isn't AltiVec nice? */
tempA = vec_mladd((vector unsigned short) srcvA, Av, rounderV);
tempB = vec_mladd((vector unsigned short) srcvB, Bv, tempA);
tempC = vec_mladd((vector unsigned short) srcvC, Cv, tempB);
tempD = vec_mladd((vector unsigned short) srcvD, Dv, tempC);
srcvA = srcvC;
srcvB = srcvD;
tempD = vec_sr(tempD, vcsr8);
dstv2 = vec_pack(tempD, (vector unsigned short) vczero);
if (dst_odd)
dstv2 = vec_perm(dstv, dstv2, vcprm(0, 1, s0, s1));
else
dstv2 = vec_perm(dstv, dstv2, vcprm(s0, s1, 2, 3));
vec_st(dstv2, 0, dst);
dst += stride;
src += stride;
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_mpeg4videodsp_init_ppc(Mpeg4VideoDSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->gmc1 = gmc1_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/mpegaudiodsp_altivec.c
+144
View File
@@ -0,0 +1,144 @@
/*
* Altivec optimized MP3 decoding functions
* Copyright (c) 2010 Vitor Sessak
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <string.h>
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/internal.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/mpegaudiodsp.h"
#if HAVE_ALTIVEC
#define MACS(rt, ra, rb) rt+=(ra)*(rb)
#define MLSS(rt, ra, rb) rt-=(ra)*(rb)
#define SUM8(op, sum, w, p) \
{ \
op(sum, (w)[0 * 64], (p)[0 * 64]); \
op(sum, (w)[1 * 64], (p)[1 * 64]); \
op(sum, (w)[2 * 64], (p)[2 * 64]); \
op(sum, (w)[3 * 64], (p)[3 * 64]); \
op(sum, (w)[4 * 64], (p)[4 * 64]); \
op(sum, (w)[5 * 64], (p)[5 * 64]); \
op(sum, (w)[6 * 64], (p)[6 * 64]); \
op(sum, (w)[7 * 64], (p)[7 * 64]); \
}
static void apply_window(const float *buf, const float *win1,
const float *win2, float *sum1, float *sum2, int len)
{
const vector float *win1a = (const vector float *) win1;
const vector float *win2a = (const vector float *) win2;
const vector float *bufa = (const vector float *) buf;
vector float *sum1a = (vector float *) sum1;
vector float *sum2a = (vector float *) sum2;
vector float av_uninit(v0), av_uninit(v4);
vector float v1, v2, v3;
len = len >> 2;
#define MULT(a, b) \
{ \
v1 = vec_ld(a, win1a); \
v2 = vec_ld(b, win2a); \
v3 = vec_ld(a, bufa); \
v0 = vec_madd(v3, v1, v0); \
v4 = vec_madd(v2, v3, v4); \
}
while (len--) {
v0 = vec_xor(v0, v0);
v4 = vec_xor(v4, v4);
MULT( 0, 0);
MULT( 256, 64);
MULT( 512, 128);
MULT( 768, 192);
MULT(1024, 256);
MULT(1280, 320);
MULT(1536, 384);
MULT(1792, 448);
vec_st(v0, 0, sum1a);
vec_st(v4, 0, sum2a);
sum1a++;
sum2a++;
win1a++;
win2a++;
bufa++;
}
}
static void apply_window_mp3(float *in, float *win, int *unused, float *out,
ptrdiff_t incr)
{
LOCAL_ALIGNED_16(float, suma, [17]);
LOCAL_ALIGNED_16(float, sumb, [17]);
LOCAL_ALIGNED_16(float, sumc, [17]);
LOCAL_ALIGNED_16(float, sumd, [17]);
float sum;
int j;
float *out2 = out + 32 * incr;
/* copy to avoid wrap */
memcpy(in + 512, in, 32 * sizeof(*in));
apply_window(in + 16, win , win + 512, suma, sumc, 16);
apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16);
SUM8(MLSS, suma[0], win + 32, in + 48);
sumc[ 0] = 0;
sumb[16] = 0;
sumd[16] = 0;
out[0 ] = suma[ 0];
out += incr;
out2 -= incr;
for(j=1;j<16;j++) {
*out = suma[ j] - sumd[16-j];
*out2 = -sumb[16-j] - sumc[ j];
out += incr;
out2 -= incr;
}
sum = 0;
SUM8(MLSS, sum, win + 16 + 32, in + 32);
*out = sum;
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_mpadsp_init_ppc(MPADSPContext *s)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
s->apply_window_float = apply_window_mp3;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/mpegvideo_altivec.c
+116
View File
@@ -0,0 +1,116 @@
/*
* Copyright (c) 2002 Dieter Shirley
*
* dct_unquantize_h263_altivec:
* Copyright (c) 2003 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <stdlib.h>
#include <stdio.h>
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/avassert.h"
#include "libavutil/cpu.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/mpegvideo.h"
#include "libavcodec/mpegvideo_unquantize.h"
#if HAVE_ALTIVEC
/* AltiVec version of dct_unquantize_h263
this code assumes `block' is 16 bytes-aligned */
static void dct_unquantize_h263_altivec(MpegEncContext *s,
int16_t *block, int n, int qscale)
{
int i, qmul, qadd;
int nCoeffs;
qadd = (qscale - 1) | 1;
qmul = qscale << 1;
if (s->mb_intra) {
if (!s->h263_aic) {
if (n < 4)
block[0] = block[0] * s->y_dc_scale;
else
block[0] = block[0] * s->c_dc_scale;
}else
qadd = 0;
i = 1;
if (s->ac_pred)
nCoeffs = 63;
else
nCoeffs = s->intra_scantable.raster_end[s->block_last_index[n]];
} else {
i = 0;
av_assert2(s->block_last_index[n]>=0);
nCoeffs = s->inter_scantable.raster_end[s->block_last_index[n]];
}
{
register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
DECLARE_ALIGNED(16, short, qmul8) = qmul;
DECLARE_ALIGNED(16, short, qadd8) = qadd;
register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
register vector bool short blockv_null, blockv_neg;
register short backup_0 = block[0];
qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
nqaddv = vec_sub(vczero, qaddv);
// vectorize all the 16 bytes-aligned blocks
// of 8 elements
for (register int j = 0; j <= nCoeffs ; j += 8) {
blockv = vec_ld(j << 1, block);
blockv_neg = vec_cmplt(blockv, vczero);
blockv_null = vec_cmpeq(blockv, vczero);
// choose between +qadd or -qadd as the third operand
temp1 = vec_sel(qaddv, nqaddv, blockv_neg);
// multiply & add (block{i,i+7} * qmul [+-] qadd)
temp1 = vec_mladd(blockv, qmulv, temp1);
// put 0 where block[{i,i+7} used to have 0
blockv = vec_sel(temp1, blockv, blockv_null);
vec_st(blockv, j << 1, block);
}
if (i == 1) {
// cheat. this avoid special-casing the first iteration
block[0] = backup_0;
}
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_mpv_unquantize_init_ppc(MPVUnquantDSPContext *s, int bitexact)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
s->dct_unquantize_h263_intra = dct_unquantize_h263_altivec;
s->dct_unquantize_h263_inter = dct_unquantize_h263_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/mpegvideoencdsp.c
+161
View File
@@ -0,0 +1,161 @@
/*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include <stdint.h>
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/mpegvideoencdsp.h"
#if HAVE_ALTIVEC
#if HAVE_VSX
static int pix_norm1_altivec(const uint8_t *pix, ptrdiff_t line_size)
{
int i, s = 0;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned int sv = (vector unsigned int) vec_splat_u32(0);
vector signed int sum;
for (i = 0; i < 16; i++) {
/* Read the potentially unaligned pixels. */
//vector unsigned char pixl = vec_ld(0, pix);
//vector unsigned char pixr = vec_ld(15, pix);
//vector unsigned char pixv = vec_perm(pixl, pixr, perm);
vector unsigned char pixv = vec_vsx_ld(0, pix);
/* Square the values, and add them to our sum. */
sv = vec_msum(pixv, pixv, sv);
pix += line_size;
}
/* Sum up the four partial sums, and put the result into s. */
sum = vec_sums((vector signed int) sv, (vector signed int) zero);
sum = vec_splat(sum, 3);
vec_ste(sum, 0, &s);
return s;
}
#else
static int pix_norm1_altivec(const uint8_t *pix, ptrdiff_t line_size)
{
int i, s = 0;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned char perm = vec_lvsl(0, pix);
vector unsigned int sv = (vector unsigned int) vec_splat_u32(0);
vector signed int sum;
for (i = 0; i < 16; i++) {
/* Read the potentially unaligned pixels. */
vector unsigned char pixl = vec_ld(0, pix);
vector unsigned char pixr = vec_ld(15, pix);
vector unsigned char pixv = vec_perm(pixl, pixr, perm);
/* Square the values, and add them to our sum. */
sv = vec_msum(pixv, pixv, sv);
pix += line_size;
}
/* Sum up the four partial sums, and put the result into s. */
sum = vec_sums((vector signed int) sv, (vector signed int) zero);
sum = vec_splat(sum, 3);
vec_ste(sum, 0, &s);
return s;
}
#endif /* HAVE_VSX */
#if HAVE_VSX
static int pix_sum_altivec(const uint8_t *pix, ptrdiff_t line_size)
{
int i, s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
for (i = 0; i < 16; i++) {
/* Read the potentially unaligned 16 pixels into t1. */
//vector unsigned char pixl = vec_ld(0, pix);
//vector unsigned char pixr = vec_ld(15, pix);
//vector unsigned char t1 = vec_perm(pixl, pixr, perm);
vector unsigned char t1 = vec_vsx_ld(0, pix);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t1, sad);
pix += line_size;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
#else
static int pix_sum_altivec(const uint8_t *pix, ptrdiff_t line_size)
{
int i, s;
const vector unsigned int zero =
(const vector unsigned int) vec_splat_u32(0);
vector unsigned char perm = vec_lvsl(0, pix);
vector unsigned int sad = (vector unsigned int) vec_splat_u32(0);
vector signed int sumdiffs;
for (i = 0; i < 16; i++) {
/* Read the potentially unaligned 16 pixels into t1. */
vector unsigned char pixl = vec_ld(0, pix);
vector unsigned char pixr = vec_ld(15, pix);
vector unsigned char t1 = vec_perm(pixl, pixr, perm);
/* Add each 4 pixel group together and put 4 results into sad. */
sad = vec_sum4s(t1, sad);
pix += line_size;
}
/* Sum up the four partial sums, and put the result into s. */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) zero);
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, &s);
return s;
}
#endif /* HAVE_VSX */
#endif /* HAVE_ALTIVEC */
av_cold void ff_mpegvideoencdsp_init_ppc(MpegvideoEncDSPContext *c,
AVCodecContext *avctx)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->pix_norm1 = pix_norm1_altivec;
c->pix_sum = pix_sum_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/pixblockdsp.c
+287
View File
@@ -0,0 +1,287 @@
/*
* Copyright (c) 2002 Brian Foley
* Copyright (c) 2002 Dieter Shirley
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/pixblockdsp.h"
#if HAVE_ALTIVEC
#if HAVE_VSX
static void get_pixels_altivec(int16_t *restrict block, const uint8_t *pixels,
ptrdiff_t stride)
{
int i;
vector unsigned char perm =
(vector unsigned char) {0x00,0x10, 0x01,0x11,0x02,0x12,0x03,0x13,\
0x04,0x14,0x05,0x15,0x06,0x16,0x07,0x17};
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
for (i = 0; i < 8; i++) {
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
vector unsigned char bytes = vec_vsx_ld(0, pixels);
// Convert the bytes into shorts.
//vector signed short shorts = (vector signed short) vec_perm(zero, bytes, perm);
vector signed short shorts = (vector signed short) vec_perm(bytes, zero, perm);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_vsx_st(shorts, i * 16, (vector signed short *) block);
pixels += stride;
}
}
#else
static void get_pixels_altivec(int16_t *restrict block, const uint8_t *pixels,
ptrdiff_t stride)
{
int i;
const vec_u8 zero = (const vec_u8)vec_splat_u8(0);
for (i = 0; i < 8; i++) {
vec_u8 perm = vec_lvsl(0, pixels);
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
vec_u8 pixl = vec_ld(0, pixels);
vec_u8 pixr = vec_ld(7, pixels);
vec_u8 bytes = vec_perm(pixl, pixr, perm);
// Convert the bytes into shorts.
vec_s16 shorts = (vec_s16)vec_mergeh(zero, bytes);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_st(shorts, i * 16, (vec_s16 *)block);
pixels += stride;
}
}
#endif /* HAVE_VSX */
#if HAVE_VSX
static void diff_pixels_altivec(int16_t *restrict block, const uint8_t *s1,
const uint8_t *s2, ptrdiff_t stride)
{
int i;
const vector unsigned char zero =
(const vector unsigned char) vec_splat_u8(0);
vector signed short shorts1, shorts2;
for (i = 0; i < 4; i++) {
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
vector unsigned char bytes = vec_vsx_ld(0, s1);
// Convert the bytes into shorts.
shorts1 = (vector signed short) vec_mergeh(bytes, zero);
// Do the same for the second block of pixels.
bytes =vec_vsx_ld(0, s2);
// Convert the bytes into shorts.
shorts2 = (vector signed short) vec_mergeh(bytes, zero);
// Do the subtraction.
shorts1 = vec_sub(shorts1, shorts2);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_vsx_st(shorts1, 0, (vector signed short *) block);
s1 += stride;
s2 += stride;
block += 8;
/* The code below is a copy of the code above...
* This is a manual unroll. */
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
bytes = vec_vsx_ld(0, s1);
// Convert the bytes into shorts.
shorts1 = (vector signed short) vec_mergeh(bytes, zero);
// Do the same for the second block of pixels.
bytes = vec_vsx_ld(0, s2);
// Convert the bytes into shorts.
shorts2 = (vector signed short) vec_mergeh(bytes, zero);
// Do the subtraction.
shorts1 = vec_sub(shorts1, shorts2);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_vsx_st(shorts1, 0, (vector signed short *) block);
s1 += stride;
s2 += stride;
block += 8;
}
}
#else
static void diff_pixels_altivec(int16_t *restrict block, const uint8_t *s1,
const uint8_t *s2, ptrdiff_t stride)
{
int i;
vec_u8 perm;
const vec_u8 zero = (const vec_u8)vec_splat_u8(0);
vec_s16 shorts1, shorts2;
for (i = 0; i < 4; i++) {
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
perm = vec_lvsl(0, s1);
vec_u8 pixl = vec_ld(0, s1);
vec_u8 pixr = vec_ld(15, s1);
vec_u8 bytes = vec_perm(pixl, pixr, perm);
// Convert the bytes into shorts.
shorts1 = (vec_s16)vec_mergeh(zero, bytes);
// Do the same for the second block of pixels.
perm = vec_lvsl(0, s2);
pixl = vec_ld(0, s2);
pixr = vec_ld(15, s2);
bytes = vec_perm(pixl, pixr, perm);
// Convert the bytes into shorts.
shorts2 = (vec_s16)vec_mergeh(zero, bytes);
// Do the subtraction.
shorts1 = vec_sub(shorts1, shorts2);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_st(shorts1, 0, (vec_s16 *)block);
s1 += stride;
s2 += stride;
block += 8;
/* The code below is a copy of the code above...
* This is a manual unroll. */
/* Read potentially unaligned pixels.
* We're reading 16 pixels, and actually only want 8,
* but we simply ignore the extras. */
perm = vec_lvsl(0, s1);
pixl = vec_ld(0, s1);
pixr = vec_ld(15, s1);
bytes = vec_perm(pixl, pixr, perm);
// Convert the bytes into shorts.
shorts1 = (vec_s16)vec_mergeh(zero, bytes);
// Do the same for the second block of pixels.
perm = vec_lvsl(0, s2);
pixl = vec_ld(0, s2);
pixr = vec_ld(15, s2);
bytes = vec_perm(pixl, pixr, perm);
// Convert the bytes into shorts.
shorts2 = (vec_s16)vec_mergeh(zero, bytes);
// Do the subtraction.
shorts1 = vec_sub(shorts1, shorts2);
// Save the data to the block, we assume the block is 16-byte aligned.
vec_st(shorts1, 0, (vec_s16 *)block);
s1 += stride;
s2 += stride;
block += 8;
}
}
#endif /* HAVE_VSX */
#endif /* HAVE_ALTIVEC */
#if HAVE_VSX
static void get_pixels_vsx(int16_t *restrict block, const uint8_t *pixels,
ptrdiff_t stride)
{
int i;
for (i = 0; i < 8; i++) {
vec_s16 shorts = vsx_ld_u8_s16(0, pixels);
vec_vsx_st(shorts, i * 16, block);
pixels += stride;
}
}
static void diff_pixels_vsx(int16_t *restrict block, const uint8_t *s1,
const uint8_t *s2, ptrdiff_t stride)
{
int i;
vec_s16 shorts1, shorts2;
for (i = 0; i < 8; i++) {
shorts1 = vsx_ld_u8_s16(0, s1);
shorts2 = vsx_ld_u8_s16(0, s2);
shorts1 = vec_sub(shorts1, shorts2);
vec_vsx_st(shorts1, 0, block);
s1 += stride;
s2 += stride;
block += 8;
}
}
#endif /* HAVE_VSX */
av_cold void ff_pixblockdsp_init_ppc(PixblockDSPContext *c,
unsigned high_bit_depth)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->diff_pixels = diff_pixels_altivec;
if (!high_bit_depth) {
c->get_pixels = get_pixels_altivec;
}
#endif /* HAVE_ALTIVEC */
#if HAVE_VSX
if (!PPC_VSX(av_get_cpu_flags()))
return;
c->diff_pixels = diff_pixels_vsx;
if (!high_bit_depth)
c->get_pixels = get_pixels_vsx;
#endif /* HAVE_VSX */
}
ffmpeg/libavcodec/ppc/svq1enc_altivec.c
+82
View File
@@ -0,0 +1,82 @@
/*
* Copyright (c) 2007 Luca Barbato <lu_zero@gentoo.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include <stdint.h>
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/svq1encdsp.h"
#if HAVE_ALTIVEC && HAVE_BIGENDIAN
static int ssd_int8_vs_int16_altivec(const int8_t *pix1, const int16_t *pix2,
intptr_t size)
{
int i, size16 = size >> 4;
vector signed char vpix1;
vector signed short vpix2, vdiff, vpix1l, vpix1h;
union {
vector signed int vscore;
int32_t score[4];
} u = { .vscore = vec_splat_s32(0) };
while (size16) {
// score += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
// load pix1 and the first batch of pix2
vpix1 = vec_unaligned_load(pix1);
vpix2 = vec_unaligned_load(pix2);
pix2 += 8;
// unpack
vpix1h = vec_unpackh(vpix1);
vdiff = vec_sub(vpix1h, vpix2);
vpix1l = vec_unpackl(vpix1);
// load another batch from pix2
vpix2 = vec_unaligned_load(pix2);
u.vscore = vec_msum(vdiff, vdiff, u.vscore);
vdiff = vec_sub(vpix1l, vpix2);
u.vscore = vec_msum(vdiff, vdiff, u.vscore);
pix1 += 16;
pix2 += 8;
size16--;
}
u.vscore = vec_sums(u.vscore, vec_splat_s32(0));
size %= 16;
for (i = 0; i < size; i++)
u.score[3] += (pix1[i] - pix2[i]) * (pix1[i] - pix2[i]);
return u.score[3];
}
#endif /* HAVE_ALTIVEC && HAVE_BIGENDIAN */
av_cold void ff_svq1enc_init_ppc(SVQ1EncDSPContext *c)
{
#if HAVE_ALTIVEC && HAVE_BIGENDIAN
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->ssd_int8_vs_int16 = ssd_int8_vs_int16_altivec;
#endif /* HAVE_ALTIVEC && HAVE_BIGENDIAN */
}
ffmpeg/libavcodec/ppc/vc1dsp_altivec.c
+365
View File
@@ -0,0 +1,365 @@
/*
* VC-1 and WMV3 decoder - DSP functions AltiVec-optimized
* Copyright (c) 2006 Konstantin Shishkov
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/vc1dsp.h"
#if HAVE_ALTIVEC
// main steps of 8x8 transform
#define STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_rnd) \
do { \
t0 = vec_sl(vec_add(s0, s4), vec_2); \
t0 = vec_add(vec_sl(t0, vec_1), t0); \
t0 = vec_add(t0, vec_rnd); \
t1 = vec_sl(vec_sub(s0, s4), vec_2); \
t1 = vec_add(vec_sl(t1, vec_1), t1); \
t1 = vec_add(t1, vec_rnd); \
t2 = vec_add(vec_sl(s6, vec_2), vec_sl(s6, vec_1)); \
t2 = vec_add(t2, vec_sl(s2, vec_4)); \
t3 = vec_add(vec_sl(s2, vec_2), vec_sl(s2, vec_1)); \
t3 = vec_sub(t3, vec_sl(s6, vec_4)); \
t4 = vec_add(t0, t2); \
t5 = vec_add(t1, t3); \
t6 = vec_sub(t1, t3); \
t7 = vec_sub(t0, t2); \
\
t0 = vec_sl(vec_add(s1, s3), vec_4); \
t0 = vec_add(t0, vec_sl(s5, vec_3)); \
t0 = vec_add(t0, vec_sl(s7, vec_2)); \
t0 = vec_add(t0, vec_sub(s5, s3)); \
\
t1 = vec_sl(vec_sub(s1, s5), vec_4); \
t1 = vec_sub(t1, vec_sl(s7, vec_3)); \
t1 = vec_sub(t1, vec_sl(s3, vec_2)); \
t1 = vec_sub(t1, vec_add(s1, s7)); \
\
t2 = vec_sl(vec_sub(s7, s3), vec_4); \
t2 = vec_add(t2, vec_sl(s1, vec_3)); \
t2 = vec_add(t2, vec_sl(s5, vec_2)); \
t2 = vec_add(t2, vec_sub(s1, s7)); \
\
t3 = vec_sl(vec_sub(s5, s7), vec_4); \
t3 = vec_sub(t3, vec_sl(s3, vec_3)); \
t3 = vec_add(t3, vec_sl(s1, vec_2)); \
t3 = vec_sub(t3, vec_add(s3, s5)); \
\
s0 = vec_add(t4, t0); \
s1 = vec_add(t5, t1); \
s2 = vec_add(t6, t2); \
s3 = vec_add(t7, t3); \
s4 = vec_sub(t7, t3); \
s5 = vec_sub(t6, t2); \
s6 = vec_sub(t5, t1); \
s7 = vec_sub(t4, t0); \
}while(0)
#define SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7) \
do { \
s0 = vec_sra(s0, vec_3); \
s1 = vec_sra(s1, vec_3); \
s2 = vec_sra(s2, vec_3); \
s3 = vec_sra(s3, vec_3); \
s4 = vec_sra(s4, vec_3); \
s5 = vec_sra(s5, vec_3); \
s6 = vec_sra(s6, vec_3); \
s7 = vec_sra(s7, vec_3); \
}while(0)
#define SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7) \
do { \
s0 = vec_sra(s0, vec_7); \
s1 = vec_sra(s1, vec_7); \
s2 = vec_sra(s2, vec_7); \
s3 = vec_sra(s3, vec_7); \
s4 = vec_sra(vec_add(s4, vec_1s), vec_7); \
s5 = vec_sra(vec_add(s5, vec_1s), vec_7); \
s6 = vec_sra(vec_add(s6, vec_1s), vec_7); \
s7 = vec_sra(vec_add(s7, vec_1s), vec_7); \
}while(0)
/* main steps of 4x4 transform */
#define STEP4(s0, s1, s2, s3, vec_rnd) \
do { \
t1 = vec_add(vec_sl(s0, vec_4), s0); \
t1 = vec_add(t1, vec_rnd); \
t2 = vec_add(vec_sl(s2, vec_4), s2); \
t0 = vec_add(t1, t2); \
t1 = vec_sub(t1, t2); \
t3 = vec_sl(vec_sub(s3, s1), vec_1); \
t3 = vec_add(t3, vec_sl(t3, vec_2)); \
t2 = vec_add(t3, vec_sl(s1, vec_5)); \
t3 = vec_add(t3, vec_sl(s3, vec_3)); \
t3 = vec_add(t3, vec_sl(s3, vec_2)); \
s0 = vec_add(t0, t2); \
s1 = vec_sub(t1, t3); \
s2 = vec_add(t1, t3); \
s3 = vec_sub(t0, t2); \
}while (0)
#define SHIFT_HOR4(s0, s1, s2, s3) \
s0 = vec_sra(s0, vec_3); \
s1 = vec_sra(s1, vec_3); \
s2 = vec_sra(s2, vec_3); \
s3 = vec_sra(s3, vec_3);
#define SHIFT_VERT4(s0, s1, s2, s3) \
s0 = vec_sra(s0, vec_7); \
s1 = vec_sra(s1, vec_7); \
s2 = vec_sra(s2, vec_7); \
s3 = vec_sra(s3, vec_7);
/** Do inverse transform on 8x8 block
*/
static void vc1_inv_trans_8x8_altivec(int16_t block[64])
{
vector signed short src0, src1, src2, src3, src4, src5, src6, src7;
vector signed int s0, s1, s2, s3, s4, s5, s6, s7;
vector signed int s8, s9, sA, sB, sC, sD, sE, sF;
vector signed int t0, t1, t2, t3, t4, t5, t6, t7;
const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4));
const vector unsigned int vec_7 = vec_splat_u32(7);
const vector unsigned int vec_4 = vec_splat_u32(4);
const vector signed int vec_4s = vec_splat_s32(4);
const vector unsigned int vec_3 = vec_splat_u32(3);
const vector unsigned int vec_2 = vec_splat_u32(2);
const vector signed int vec_1s = vec_splat_s32(1);
const vector unsigned int vec_1 = vec_splat_u32(1);
src0 = vec_ld( 0, block);
src1 = vec_ld( 16, block);
src2 = vec_ld( 32, block);
src3 = vec_ld( 48, block);
src4 = vec_ld( 64, block);
src5 = vec_ld( 80, block);
src6 = vec_ld( 96, block);
src7 = vec_ld(112, block);
s0 = vec_unpackl(src0);
s1 = vec_unpackl(src1);
s2 = vec_unpackl(src2);
s3 = vec_unpackl(src3);
s4 = vec_unpackl(src4);
s5 = vec_unpackl(src5);
s6 = vec_unpackl(src6);
s7 = vec_unpackl(src7);
s8 = vec_unpackh(src0);
s9 = vec_unpackh(src1);
sA = vec_unpackh(src2);
sB = vec_unpackh(src3);
sC = vec_unpackh(src4);
sD = vec_unpackh(src5);
sE = vec_unpackh(src6);
sF = vec_unpackh(src7);
STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s);
SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7);
STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s);
SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF);
src0 = vec_pack(s8, s0);
src1 = vec_pack(s9, s1);
src2 = vec_pack(sA, s2);
src3 = vec_pack(sB, s3);
src4 = vec_pack(sC, s4);
src5 = vec_pack(sD, s5);
src6 = vec_pack(sE, s6);
src7 = vec_pack(sF, s7);
TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);
s0 = vec_unpackl(src0);
s1 = vec_unpackl(src1);
s2 = vec_unpackl(src2);
s3 = vec_unpackl(src3);
s4 = vec_unpackl(src4);
s5 = vec_unpackl(src5);
s6 = vec_unpackl(src6);
s7 = vec_unpackl(src7);
s8 = vec_unpackh(src0);
s9 = vec_unpackh(src1);
sA = vec_unpackh(src2);
sB = vec_unpackh(src3);
sC = vec_unpackh(src4);
sD = vec_unpackh(src5);
sE = vec_unpackh(src6);
sF = vec_unpackh(src7);
STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_64);
SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7);
STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_64);
SHIFT_VERT8(s8, s9, sA, sB, sC, sD, sE, sF);
src0 = vec_pack(s8, s0);
src1 = vec_pack(s9, s1);
src2 = vec_pack(sA, s2);
src3 = vec_pack(sB, s3);
src4 = vec_pack(sC, s4);
src5 = vec_pack(sD, s5);
src6 = vec_pack(sE, s6);
src7 = vec_pack(sF, s7);
vec_st(src0, 0, block);
vec_st(src1, 16, block);
vec_st(src2, 32, block);
vec_st(src3, 48, block);
vec_st(src4, 64, block);
vec_st(src5, 80, block);
vec_st(src6, 96, block);
vec_st(src7,112, block);
}
/** Do inverse transform on 8x4 part of block
*/
static void vc1_inv_trans_8x4_altivec(uint8_t *dest, ptrdiff_t stride,
int16_t *block)
{
vector signed short src0, src1, src2, src3, src4, src5, src6, src7;
vector signed int s0, s1, s2, s3, s4, s5, s6, s7;
vector signed int s8, s9, sA, sB, sC, sD, sE, sF;
vector signed int t0, t1, t2, t3, t4, t5, t6, t7;
const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4));
const vector unsigned int vec_7 = vec_splat_u32(7);
const vector unsigned int vec_5 = vec_splat_u32(5);
const vector unsigned int vec_4 = vec_splat_u32(4);
const vector signed int vec_4s = vec_splat_s32(4);
const vector unsigned int vec_3 = vec_splat_u32(3);
const vector unsigned int vec_2 = vec_splat_u32(2);
const vector unsigned int vec_1 = vec_splat_u32(1);
vector unsigned char tmp;
vector signed short tmp2, tmp3;
vector unsigned char perm0, perm1, p0, p1, p;
src0 = vec_ld( 0, block);
src1 = vec_ld( 16, block);
src2 = vec_ld( 32, block);
src3 = vec_ld( 48, block);
src4 = vec_ld( 64, block);
src5 = vec_ld( 80, block);
src6 = vec_ld( 96, block);
src7 = vec_ld(112, block);
TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);
s0 = vec_unpackl(src0);
s1 = vec_unpackl(src1);
s2 = vec_unpackl(src2);
s3 = vec_unpackl(src3);
s4 = vec_unpackl(src4);
s5 = vec_unpackl(src5);
s6 = vec_unpackl(src6);
s7 = vec_unpackl(src7);
s8 = vec_unpackh(src0);
s9 = vec_unpackh(src1);
sA = vec_unpackh(src2);
sB = vec_unpackh(src3);
sC = vec_unpackh(src4);
sD = vec_unpackh(src5);
sE = vec_unpackh(src6);
sF = vec_unpackh(src7);
STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s);
SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7);
STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s);
SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF);
src0 = vec_pack(s8, s0);
src1 = vec_pack(s9, s1);
src2 = vec_pack(sA, s2);
src3 = vec_pack(sB, s3);
src4 = vec_pack(sC, s4);
src5 = vec_pack(sD, s5);
src6 = vec_pack(sE, s6);
src7 = vec_pack(sF, s7);
TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);
s0 = vec_unpackh(src0);
s1 = vec_unpackh(src1);
s2 = vec_unpackh(src2);
s3 = vec_unpackh(src3);
s8 = vec_unpackl(src0);
s9 = vec_unpackl(src1);
sA = vec_unpackl(src2);
sB = vec_unpackl(src3);
STEP4(s0, s1, s2, s3, vec_64);
SHIFT_VERT4(s0, s1, s2, s3);
STEP4(s8, s9, sA, sB, vec_64);
SHIFT_VERT4(s8, s9, sA, sB);
src0 = vec_pack(s0, s8);
src1 = vec_pack(s1, s9);
src2 = vec_pack(s2, sA);
src3 = vec_pack(s3, sB);
#if HAVE_BIGENDIAN
p0 = vec_lvsl (0, dest);
p1 = vec_lvsl (stride, dest);
p = vec_splat_u8 (-1);
perm0 = vec_mergeh (p, p0);
perm1 = vec_mergeh (p, p1);
#define GET_TMP2(dst, p) \
tmp = vec_ld (0, dest); \
tmp2 = (vector signed short)vec_perm (tmp, vec_splat_u8(0), p);
#else
#define GET_TMP2(dst,p) \
tmp = vec_vsx_ld (0, dst); \
tmp2 = (vector signed short)vec_mergeh (tmp, vec_splat_u8(0));
#endif
#define ADD(dest,src,perm) \
GET_TMP2(dest, perm); \
tmp3 = vec_adds (tmp2, src); \
tmp = vec_packsu (tmp3, tmp3); \
vec_ste ((vector unsigned int)tmp, 0, (unsigned int *)dest); \
vec_ste ((vector unsigned int)tmp, 4, (unsigned int *)dest);
ADD (dest, src0, perm0) dest += stride;
ADD (dest, src1, perm1) dest += stride;
ADD (dest, src2, perm0) dest += stride;
ADD (dest, src3, perm1)
}
#define PUT_OP_U8_ALTIVEC(d, s, dst) d = s
#define AVG_OP_U8_ALTIVEC(d, s, dst) d = vec_avg(dst, s)
#define OP_U8_ALTIVEC PUT_OP_U8_ALTIVEC
#define PREFIX_no_rnd_vc1_chroma_mc8_altivec put_no_rnd_vc1_chroma_mc8_altivec
#include "h264chroma_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_no_rnd_vc1_chroma_mc8_altivec
#define OP_U8_ALTIVEC AVG_OP_U8_ALTIVEC
#define PREFIX_no_rnd_vc1_chroma_mc8_altivec avg_no_rnd_vc1_chroma_mc8_altivec
#include "h264chroma_template.c"
#undef OP_U8_ALTIVEC
#undef PREFIX_no_rnd_vc1_chroma_mc8_altivec
#endif /* HAVE_ALTIVEC */
av_cold void ff_vc1dsp_init_ppc(VC1DSPContext *dsp)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_altivec;
dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_altivec;
dsp->put_no_rnd_vc1_chroma_pixels_tab[0] = put_no_rnd_vc1_chroma_mc8_altivec;
dsp->avg_no_rnd_vc1_chroma_pixels_tab[0] = avg_no_rnd_vc1_chroma_mc8_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/videodsp.c
+36
View File
@@ -0,0 +1,36 @@
/*
* Copyright (c) 2003-2004 Romain Dolbeau
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "libavutil/attributes.h"
#include "libavcodec/videodsp.h"
static void prefetch_ppc(const uint8_t *mem, ptrdiff_t stride, int h)
{
register const uint8_t *p = mem;
do {
__asm__ volatile ("dcbt 0,%0" : : "r" (p));
p += stride;
} while(--h);
}
av_cold void ff_videodsp_init_ppc(VideoDSPContext *ctx, int bpc)
{
ctx->prefetch = prefetch_ppc;
}
ffmpeg/libavcodec/ppc/vorbisdsp_altivec.c
+63
View File
@@ -0,0 +1,63 @@
/*
* Copyright (c) 2006 Luca Barbato <lu_zero@gentoo.org>
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/vorbisdsp.h"
#if HAVE_ALTIVEC
static void vorbis_inverse_coupling_altivec(float *mag, float *ang,
ptrdiff_t blocksize)
{
vector float m, a;
vector bool int t0, t1;
const vector unsigned int v_31 = //XXX
vec_add(vec_add(vec_splat_u32(15),vec_splat_u32(15)),vec_splat_u32(1));
for (ptrdiff_t i = 0; i < blocksize; i += 4) {
m = vec_ld(0, mag+i);
a = vec_ld(0, ang+i);
t0 = vec_cmple(m, (vector float)vec_splat_u32(0));
t1 = vec_cmple(a, (vector float)vec_splat_u32(0));
a = vec_xor(a, (vector float) vec_sl((vector unsigned int)t0, v_31));
t0 = (vector bool int)vec_and(a, t1);
t1 = (vector bool int)vec_andc(a, t1);
a = vec_sub(m, (vector float)t1);
m = vec_add(m, (vector float)t0);
vec_stl(a, 0, ang+i);
vec_stl(m, 0, mag+i);
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_vorbisdsp_init_ppc(VorbisDSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->vorbis_inverse_coupling = vorbis_inverse_coupling_altivec;
#endif /* HAVE_ALTIVEC */
}
ffmpeg/libavcodec/ppc/vp3dsp_altivec.c
+204
View File
@@ -0,0 +1,204 @@
/*
* Copyright (C) 2009 David Conrad
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <string.h>
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/vp3dsp.h"
#if HAVE_ALTIVEC
static const vec_s16 constants =
{0, 64277, 60547, 54491, 46341, 36410, 25080, 12785};
#if HAVE_BIGENDIAN
static const vec_u8 interleave_high =
{0, 1, 16, 17, 4, 5, 20, 21, 8, 9, 24, 25, 12, 13, 28, 29};
#else
static const vec_u8 interleave_high =
{2, 3, 18, 19, 6, 7, 22, 23, 10, 11, 26, 27, 14, 15, 30, 31};
#endif
#define IDCT_START \
vec_s16 A, B, C, D, Ad, Bd, Cd, Dd, E, F, G, H;\
vec_s16 Ed, Gd, Add, Bdd, Fd, Hd;\
vec_s16 eight = vec_splat_s16(8);\
vec_u16 four = vec_splat_u16(4);\
\
vec_s16 C1 = vec_splat(constants, 1);\
vec_s16 C2 = vec_splat(constants, 2);\
vec_s16 C3 = vec_splat(constants, 3);\
vec_s16 C4 = vec_splat(constants, 4);\
vec_s16 C5 = vec_splat(constants, 5);\
vec_s16 C6 = vec_splat(constants, 6);\
vec_s16 C7 = vec_splat(constants, 7);\
\
vec_s16 b0 = vec_ld(0x00, block);\
vec_s16 b1 = vec_ld(0x10, block);\
vec_s16 b2 = vec_ld(0x20, block);\
vec_s16 b3 = vec_ld(0x30, block);\
vec_s16 b4 = vec_ld(0x40, block);\
vec_s16 b5 = vec_ld(0x50, block);\
vec_s16 b6 = vec_ld(0x60, block);\
vec_s16 b7 = vec_ld(0x70, block);
// these functions do (a*C)>>16
// things are tricky because a is signed, but C unsigned.
// M15 is used if C fits in 15 bit unsigned (C6,C7)
// M16 is used if C requires 16 bits unsigned
static inline vec_s16 M15(vec_s16 a, vec_s16 C)
{
return (vec_s16)vec_perm(vec_mule(a,C), vec_mulo(a,C), interleave_high);
}
static inline vec_s16 M16(vec_s16 a, vec_s16 C)
{
return vec_add(a, M15(a, C));
}
#define IDCT_1D(ADD, SHIFT)\
A = vec_add(M16(b1, C1), M15(b7, C7));\
B = vec_sub(M15(b1, C7), M16(b7, C1));\
C = vec_add(M16(b3, C3), M16(b5, C5));\
D = vec_sub(M16(b5, C3), M16(b3, C5));\
\
Ad = M16(vec_sub(A, C), C4);\
Bd = M16(vec_sub(B, D), C4);\
\
Cd = vec_add(A, C);\
Dd = vec_add(B, D);\
\
E = ADD(M16(vec_add(b0, b4), C4));\
F = ADD(M16(vec_sub(b0, b4), C4));\
\
G = vec_add(M16(b2, C2), M15(b6, C6));\
H = vec_sub(M15(b2, C6), M16(b6, C2));\
\
Ed = vec_sub(E, G);\
Gd = vec_add(E, G);\
\
Add = vec_add(F, Ad);\
Bdd = vec_sub(Bd, H);\
\
Fd = vec_sub(F, Ad);\
Hd = vec_add(Bd, H);\
\
b0 = SHIFT(vec_add(Gd, Cd));\
b7 = SHIFT(vec_sub(Gd, Cd));\
\
b1 = SHIFT(vec_add(Add, Hd));\
b2 = SHIFT(vec_sub(Add, Hd));\
\
b3 = SHIFT(vec_add(Ed, Dd));\
b4 = SHIFT(vec_sub(Ed, Dd));\
\
b5 = SHIFT(vec_add(Fd, Bdd));\
b6 = SHIFT(vec_sub(Fd, Bdd));
#define NOP(a) a
#define ADD8(a) vec_add(a, eight)
#define SHIFT4(a) vec_sra(a, four)
static void vp3_idct_put_altivec(uint8_t *dst, ptrdiff_t stride, int16_t block[64])
{
vec_u8 t;
IDCT_START
// pixels are signed; so add 128*16 in addition to the normal 8
vec_s16 v2048 = vec_sl(vec_splat_s16(1), vec_splat_u16(11));
eight = vec_add(eight, v2048);
IDCT_1D(NOP, NOP)
TRANSPOSE8(b0, b1, b2, b3, b4, b5, b6, b7);
IDCT_1D(ADD8, SHIFT4)
#define PUT(a)\
t = vec_packsu(a, a);\
vec_ste((vec_u32)t, 0, (unsigned int *)dst);\
vec_ste((vec_u32)t, 4, (unsigned int *)dst);
PUT(b0) dst += stride;
PUT(b1) dst += stride;
PUT(b2) dst += stride;
PUT(b3) dst += stride;
PUT(b4) dst += stride;
PUT(b5) dst += stride;
PUT(b6) dst += stride;
PUT(b7)
memset(block, 0, sizeof(*block) * 64);
}
static void vp3_idct_add_altivec(uint8_t *dst, ptrdiff_t stride, int16_t block[64])
{
LOAD_ZERO;
vec_u8 t, vdst;
vec_s16 vdst_16;
vec_u8 vdst_mask = vec_mergeh(vec_splat_u8(-1), vec_lvsl(0, dst));
IDCT_START
IDCT_1D(NOP, NOP)
TRANSPOSE8(b0, b1, b2, b3, b4, b5, b6, b7);
IDCT_1D(ADD8, SHIFT4)
#if HAVE_BIGENDIAN
#define GET_VDST16\
vdst = vec_ld(0, dst);\
vdst_16 = (vec_s16)vec_perm(vdst, zero_u8v, vdst_mask);
#else
#define GET_VDST16\
vdst = vec_vsx_ld(0,dst);\
vdst_16 = (vec_s16)vec_mergeh(vdst, zero_u8v);
#endif
#define ADD(a)\
GET_VDST16;\
vdst_16 = vec_adds(a, vdst_16);\
t = vec_packsu(vdst_16, vdst_16);\
vec_ste((vec_u32)t, 0, (unsigned int *)dst);\
vec_ste((vec_u32)t, 4, (unsigned int *)dst);
ADD(b0) dst += stride;
ADD(b1) dst += stride;
ADD(b2) dst += stride;
ADD(b3) dst += stride;
ADD(b4) dst += stride;
ADD(b5) dst += stride;
ADD(b6) dst += stride;
ADD(b7)
memset(block, 0, sizeof(*block) * 64);
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_vp3dsp_init_ppc(VP3DSPContext *c, int flags)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->idct_put = vp3_idct_put_altivec;
c->idct_add = vp3_idct_add_altivec;
#endif
}
ffmpeg/libavcodec/ppc/vp8dsp_altivec.c
+363
View File
@@ -0,0 +1,363 @@
/*
* VP8 compatible video decoder
*
* Copyright (C) 2010 David Conrad
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem_internal.h"
#include "libavutil/ppc/cpu.h"
#include "libavutil/ppc/util_altivec.h"
#include "libavcodec/vp8dsp.h"
#include "hpeldsp_altivec.h"
#if HAVE_ALTIVEC
#define REPT4(...) { __VA_ARGS__, __VA_ARGS__, __VA_ARGS__, __VA_ARGS__ }
// h subpel filter uses msum to multiply+add 4 pixel taps at once
static const vec_s8 h_subpel_filters_inner[7] =
{
REPT4( -6, 123, 12, -1),
REPT4(-11, 108, 36, -8),
REPT4( -9, 93, 50, -6),
REPT4(-16, 77, 77, -16),
REPT4( -6, 50, 93, -9),
REPT4( -8, 36, 108, -11),
REPT4( -1, 12, 123, -6),
};
// for 6tap filters, these are the outer two taps
// The zeros mask off pixels 4-7 when filtering 0-3
// and vice-versa
static const vec_s8 h_subpel_filters_outer[4] =
{
REPT4(0, 0, 2, 1),
REPT4(0, 0, 3, 3),
REPT4(0, 0, 1, 2),
REPT4(0, 0, 0, 0),
};
#define LOAD_H_SUBPEL_FILTER(i) \
vec_s8 filter_inner = h_subpel_filters_inner[i]; \
vec_s8 filter_outerh = h_subpel_filters_outer[(i)>>1]; \
vec_s8 filter_outerl = vec_sld(filter_outerh, filter_outerh, 2)
#if HAVE_BIGENDIAN
#define GET_PIXHL(offset) \
a = vec_ld((offset)-is6tap-1, src); \
b = vec_ld((offset)-is6tap-1+15, src); \
pixh = vec_perm(a, b, permh##offset); \
pixl = vec_perm(a, b, perml##offset)
#define GET_OUTER(offset) outer = vec_perm(a, b, perm_6tap##offset)
#else
#define GET_PIXHL(offset) \
a = vec_vsx_ld((offset)-is6tap-1, src); \
pixh = vec_perm(a, a, perm_inner); \
pixl = vec_perm(a, a, vec_add(perm_inner, vec_splat_u8(4)))
#define GET_OUTER(offset) outer = vec_perm(a, a, perm_outer)
#endif
#define FILTER_H(dstv, off) \
GET_PIXHL(off); \
filth = vec_msum(filter_inner, pixh, c64); \
filtl = vec_msum(filter_inner, pixl, c64); \
\
if (is6tap) { \
GET_OUTER(off); \
filth = vec_msum(filter_outerh, outer, filth); \
filtl = vec_msum(filter_outerl, outer, filtl); \
} \
if (w == 4) \
filtl = filth; /* discard pixels 4-7 */ \
dstv = vec_packs(filth, filtl); \
dstv = vec_sra(dstv, c7)
static av_always_inline
void put_vp8_epel_h_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
const uint8_t *src, ptrdiff_t src_stride,
int h, int mx, int w, int is6tap)
{
LOAD_H_SUBPEL_FILTER(mx-1);
#if HAVE_BIGENDIAN
vec_u8 align_vec0, align_vec8, permh0, permh8;
vec_u8 perm_6tap0, perm_6tap8, perml0, perml8;
vec_u8 b;
#endif
vec_u8 filt, a, pixh, pixl, outer;
vec_s16 f16h, f16l;
vec_s32 filth, filtl;
vec_u8 perm_inner6 = { 1,2,3,4, 2,3,4,5, 3,4,5,6, 4,5,6,7 };
vec_u8 perm_inner4 = { 0,1,2,3, 1,2,3,4, 2,3,4,5, 3,4,5,6 };
vec_u8 perm_inner = is6tap ? perm_inner6 : perm_inner4;
vec_u8 perm_outer = { 4,9, 0,5, 5,10, 1,6, 6,11, 2,7, 7,12, 3,8 };
vec_s32 c64 = vec_sl(vec_splat_s32(1), vec_splat_u32(6));
vec_u16 c7 = vec_splat_u16(7);
#if HAVE_BIGENDIAN
align_vec0 = vec_lvsl( -is6tap-1, src);
align_vec8 = vec_lvsl(8-is6tap-1, src);
permh0 = vec_perm(align_vec0, align_vec0, perm_inner);
permh8 = vec_perm(align_vec8, align_vec8, perm_inner);
perm_inner = vec_add(perm_inner, vec_splat_u8(4));
perml0 = vec_perm(align_vec0, align_vec0, perm_inner);
perml8 = vec_perm(align_vec8, align_vec8, perm_inner);
perm_6tap0 = vec_perm(align_vec0, align_vec0, perm_outer);
perm_6tap8 = vec_perm(align_vec8, align_vec8, perm_outer);
#endif
while (h --> 0) {
FILTER_H(f16h, 0);
if (w == 16) {
FILTER_H(f16l, 8);
filt = vec_packsu(f16h, f16l);
vec_st(filt, 0, dst);
} else {
filt = vec_packsu(f16h, f16h);
vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
if (w == 8)
vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
}
src += src_stride;
dst += dst_stride;
}
}
// v subpel filter does a simple vertical multiply + add
static const vec_u8 v_subpel_filters[7] =
{
{ 0, 6, 123, 12, 1, 0 },
{ 2, 11, 108, 36, 8, 1 },
{ 0, 9, 93, 50, 6, 0 },
{ 3, 16, 77, 77, 16, 3 },
{ 0, 6, 50, 93, 9, 0 },
{ 1, 8, 36, 108, 11, 2 },
{ 0, 1, 12, 123, 6, 0 },
};
#define LOAD_V_SUBPEL_FILTER(i) \
vec_u8 subpel_filter = v_subpel_filters[i]; \
vec_u8 f0 = vec_splat(subpel_filter, 0); \
vec_u8 f1 = vec_splat(subpel_filter, 1); \
vec_u8 f2 = vec_splat(subpel_filter, 2); \
vec_u8 f3 = vec_splat(subpel_filter, 3); \
vec_u8 f4 = vec_splat(subpel_filter, 4); \
vec_u8 f5 = vec_splat(subpel_filter, 5)
#define FILTER_V(dstv, vec_mul) \
s1f = (vec_s16)vec_mul(s1, f1); \
s2f = (vec_s16)vec_mul(s2, f2); \
s3f = (vec_s16)vec_mul(s3, f3); \
s4f = (vec_s16)vec_mul(s4, f4); \
s2f = vec_subs(s2f, s1f); \
s3f = vec_subs(s3f, s4f); \
if (is6tap) { \
s0f = (vec_s16)vec_mul(s0, f0); \
s5f = (vec_s16)vec_mul(s5, f5); \
s2f = vec_adds(s2f, s0f); \
s3f = vec_adds(s3f, s5f); \
} \
dstv = vec_adds(s2f, s3f); \
dstv = vec_adds(dstv, c64); \
dstv = vec_sra(dstv, c7)
#if HAVE_BIGENDIAN
#define LOAD_HL(off, s, perm) load_with_perm_vec(off, s, perm)
#else
#define LOAD_HL(off, s, perm) vec_mergeh(vec_vsx_ld(off,s), vec_vsx_ld(off+8,s))
#endif
static av_always_inline
void put_vp8_epel_v_altivec_core(uint8_t *dst, ptrdiff_t dst_stride,
const uint8_t *src, ptrdiff_t src_stride,
int h, int my, int w, int is6tap)
{
LOAD_V_SUBPEL_FILTER(my-1);
vec_u8 s0, s1, s2, s3, s4, s5, filt, align_vech, perm_vec, align_vecl;
vec_s16 s0f, s1f, s2f, s3f, s4f, s5f, f16h, f16l;
vec_s16 c64 = vec_sl(vec_splat_s16(1), vec_splat_u16(6));
vec_u16 c7 = vec_splat_u16(7);
#if HAVE_BIGENDIAN
// we want pixels 0-7 to be in the even positions and 8-15 in the odd,
// so combine this permute with the alignment permute vector
align_vech = vec_lvsl(0, src);
align_vecl = vec_sld(align_vech, align_vech, 8);
if (w ==16)
perm_vec = vec_mergeh(align_vech, align_vecl);
else
perm_vec = vec_mergeh(align_vech, align_vech);
#endif
if (is6tap)
s0 = LOAD_HL(-2*src_stride, src, perm_vec);
s1 = LOAD_HL(-1*src_stride, src, perm_vec);
s2 = LOAD_HL( 0*src_stride, src, perm_vec);
s3 = LOAD_HL( 1*src_stride, src, perm_vec);
if (is6tap)
s4 = LOAD_HL( 2*src_stride, src, perm_vec);
src += (2+is6tap)*src_stride;
while (h --> 0) {
if (is6tap)
s5 = LOAD_HL(0, src, perm_vec);
else
s4 = LOAD_HL(0, src, perm_vec);
FILTER_V(f16h, vec_mule);
if (w == 16) {
FILTER_V(f16l, vec_mulo);
filt = vec_packsu(f16h, f16l);
vec_st(filt, 0, dst);
} else {
filt = vec_packsu(f16h, f16h);
if (w == 4)
filt = (vec_u8)vec_splat((vec_u32)filt, 0);
else
vec_ste((vec_u32)filt, 4, (uint32_t*)dst);
vec_ste((vec_u32)filt, 0, (uint32_t*)dst);
}
if (is6tap)
s0 = s1;
s1 = s2;
s2 = s3;
s3 = s4;
if (is6tap)
s4 = s5;
dst += dst_stride;
src += src_stride;
}
}
#define EPEL_FUNCS(WIDTH, TAPS) \
static av_noinline \
void put_vp8_epel ## WIDTH ## _h ## TAPS ## _altivec(uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *src, ptrdiff_t src_stride, int h, int mx, int my) \
{ \
put_vp8_epel_h_altivec_core(dst, dst_stride, src, src_stride, h, mx, WIDTH, TAPS == 6); \
} \
\
static av_noinline \
void put_vp8_epel ## WIDTH ## _v ## TAPS ## _altivec(uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *src, ptrdiff_t src_stride, int h, int mx, int my) \
{ \
put_vp8_epel_v_altivec_core(dst, dst_stride, src, src_stride, h, my, WIDTH, TAPS == 6); \
}
#define EPEL_HV(WIDTH, HTAPS, VTAPS) \
static void put_vp8_epel ## WIDTH ## _h ## HTAPS ## v ## VTAPS ## _altivec(uint8_t *dst, ptrdiff_t dstride, const uint8_t *src, ptrdiff_t sstride, int h, int mx, int my) \
{ \
DECLARE_ALIGNED(16, uint8_t, tmp)[(2*WIDTH+5)*16]; \
if (VTAPS == 6) { \
put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16, src-2*sstride, sstride, h+5, mx, my); \
put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+2*16, 16, h, mx, my); \
} else { \
put_vp8_epel ## WIDTH ## _h ## HTAPS ## _altivec(tmp, 16, src-sstride, sstride, h+4, mx, my); \
put_vp8_epel ## WIDTH ## _v ## VTAPS ## _altivec(dst, dstride, tmp+16, 16, h, mx, my); \
} \
}
EPEL_FUNCS(16,6)
EPEL_FUNCS(8, 6)
EPEL_FUNCS(8, 4)
EPEL_FUNCS(4, 6)
EPEL_FUNCS(4, 4)
EPEL_HV(16, 6,6)
EPEL_HV(8, 6,6)
EPEL_HV(8, 4,6)
EPEL_HV(8, 6,4)
EPEL_HV(8, 4,4)
EPEL_HV(4, 6,6)
EPEL_HV(4, 4,6)
EPEL_HV(4, 6,4)
EPEL_HV(4, 4,4)
static void put_vp8_pixels16_altivec(uint8_t *dst, ptrdiff_t dstride, const uint8_t *src, ptrdiff_t sstride, int h, int mx, int my)
{
register vector unsigned char perm;
int i;
register ptrdiff_t dstride2 = dstride << 1, sstride2 = sstride << 1;
register ptrdiff_t dstride3 = dstride2 + dstride, sstride3 = sstride + sstride2;
register ptrdiff_t dstride4 = dstride << 2, sstride4 = sstride << 2;
#if HAVE_BIGENDIAN
perm = vec_lvsl(0, src);
#endif
// hand-unrolling the loop by 4 gains about 15%
// minimum execution time goes from 74 to 60 cycles
// it's faster than -funroll-loops, but using
// -funroll-loops w/ this is bad - 74 cycles again.
// all this is on a 7450, tuning for the 7450
for (i = 0; i < h; i += 4) {
vec_st(load_with_perm_vec(0, src, perm), 0, dst);
vec_st(load_with_perm_vec(sstride, src, perm), dstride, dst);
vec_st(load_with_perm_vec(sstride2, src, perm), dstride2, dst);
vec_st(load_with_perm_vec(sstride3, src, perm), dstride3, dst);
src += sstride4;
dst += dstride4;
}
}
#endif /* HAVE_ALTIVEC */
av_cold void ff_vp78dsp_init_ppc(VP8DSPContext *c)
{
#if HAVE_ALTIVEC
if (!PPC_ALTIVEC(av_get_cpu_flags()))
return;
c->put_vp8_epel_pixels_tab[0][0][0] = put_vp8_pixels16_altivec;
c->put_vp8_epel_pixels_tab[0][0][2] = put_vp8_epel16_h6_altivec;
c->put_vp8_epel_pixels_tab[0][2][0] = put_vp8_epel16_v6_altivec;
c->put_vp8_epel_pixels_tab[0][2][2] = put_vp8_epel16_h6v6_altivec;
c->put_vp8_epel_pixels_tab[1][0][2] = put_vp8_epel8_h6_altivec;
c->put_vp8_epel_pixels_tab[1][2][0] = put_vp8_epel8_v6_altivec;
c->put_vp8_epel_pixels_tab[1][0][1] = put_vp8_epel8_h4_altivec;
c->put_vp8_epel_pixels_tab[1][1][0] = put_vp8_epel8_v4_altivec;
c->put_vp8_epel_pixels_tab[1][2][2] = put_vp8_epel8_h6v6_altivec;
c->put_vp8_epel_pixels_tab[1][1][1] = put_vp8_epel8_h4v4_altivec;
c->put_vp8_epel_pixels_tab[1][1][2] = put_vp8_epel8_h6v4_altivec;
c->put_vp8_epel_pixels_tab[1][2][1] = put_vp8_epel8_h4v6_altivec;
c->put_vp8_epel_pixels_tab[2][0][2] = put_vp8_epel4_h6_altivec;
c->put_vp8_epel_pixels_tab[2][2][0] = put_vp8_epel4_v6_altivec;
c->put_vp8_epel_pixels_tab[2][0][1] = put_vp8_epel4_h4_altivec;
c->put_vp8_epel_pixels_tab[2][1][0] = put_vp8_epel4_v4_altivec;
c->put_vp8_epel_pixels_tab[2][2][2] = put_vp8_epel4_h6v6_altivec;
c->put_vp8_epel_pixels_tab[2][1][1] = put_vp8_epel4_h4v4_altivec;
c->put_vp8_epel_pixels_tab[2][1][2] = put_vp8_epel4_h6v4_altivec;
c->put_vp8_epel_pixels_tab[2][2][1] = put_vp8_epel4_h4v6_altivec;
#endif /* HAVE_ALTIVEC */
}