Commit acf91215 authored by Diego Biurrun's avatar Diego Biurrun

x86: dsputil: Avoid pointless CONFIG_ENCODERS indirection

The remaining dsputil bits are encoding-specific anyway.
parent a8552ee3
......@@ -6,8 +6,7 @@ OBJS-$(CONFIG_AUDIODSP) += x86/audiodsp_init.o
OBJS-$(CONFIG_BLOCKDSP) += x86/blockdsp.o
OBJS-$(CONFIG_BSWAPDSP) += x86/bswapdsp_init.o
OBJS-$(CONFIG_DCT) += x86/dct_init.o
OBJS-$(CONFIG_DSPUTIL) += x86/dsputil_init.o
OBJS-$(CONFIG_ENCODERS) += x86/dsputilenc_mmx.o \
OBJS-$(CONFIG_DSPUTIL) += x86/dsputil_init.o \
x86/motion_est.o
OBJS-$(CONFIG_FDCTDSP) += x86/fdctdsp_init.o
OBJS-$(CONFIG_FFT) += x86/fft_init.o
......@@ -72,7 +71,7 @@ YASM-OBJS-$(CONFIG_AUDIODSP) += x86/audiodsp.o
YASM-OBJS-$(CONFIG_BSWAPDSP) += x86/bswapdsp.o
YASM-OBJS-$(CONFIG_DCT) += x86/dct32.o
YASM-OBJS-$(CONFIG_DNXHD_ENCODER) += x86/dnxhdenc.o
YASM-OBJS-$(CONFIG_ENCODERS) += x86/dsputilenc.o
YASM-OBJS-$(CONFIG_DSPUTIL) += x86/dsputilenc.o
YASM-OBJS-$(CONFIG_FFT) += x86/fft.o
YASM-OBJS-$(CONFIG_H263DSP) += x86/h263_loopfilter.o
YASM-OBJS-$(CONFIG_H264CHROMA) += x86/h264_chromamc.o \
......
/*
* MMX optimized DSP utils
* Copyright (c) 2000, 2001 Fabrice Bellard
* Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
*
* MMX optimization by Nick Kurshev <nickols_k@mail.ru>
*
* This file is part of Libav.
*
* Libav is free software; you can redistribute it and/or
......@@ -16,16 +22,859 @@
* 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/x86/asm.h"
#include "libavutil/x86/cpu.h"
#include "libavcodec/avcodec.h"
#include "libavcodec/dsputil.h"
#include "libavcodec/mpegvideo.h"
#include "dsputil_x86.h"
#if HAVE_INLINE_ASM
static int sse8_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int tmp;
__asm__ volatile (
"movl %4, %%ecx \n"
"shr $1, %%ecx \n"
"pxor %%mm0, %%mm0 \n" /* mm0 = 0 */
"pxor %%mm7, %%mm7 \n" /* mm7 holds the sum */
"1: \n"
"movq (%0), %%mm1 \n" /* mm1 = pix1[0][0 - 7] */
"movq (%1), %%mm2 \n" /* mm2 = pix2[0][0 - 7] */
"movq (%0, %3), %%mm3 \n" /* mm3 = pix1[1][0 - 7] */
"movq (%1, %3), %%mm4 \n" /* mm4 = pix2[1][0 - 7] */
/* todo: mm1-mm2, mm3-mm4 */
/* algo: subtract mm1 from mm2 with saturation and vice versa */
/* OR the results to get absolute difference */
"movq %%mm1, %%mm5 \n"
"movq %%mm3, %%mm6 \n"
"psubusb %%mm2, %%mm1 \n"
"psubusb %%mm4, %%mm3 \n"
"psubusb %%mm5, %%mm2 \n"
"psubusb %%mm6, %%mm4 \n"
"por %%mm1, %%mm2 \n"
"por %%mm3, %%mm4 \n"
/* now convert to 16-bit vectors so we can square them */
"movq %%mm2, %%mm1 \n"
"movq %%mm4, %%mm3 \n"
"punpckhbw %%mm0, %%mm2 \n"
"punpckhbw %%mm0, %%mm4 \n"
"punpcklbw %%mm0, %%mm1 \n" /* mm1 now spread over (mm1, mm2) */
"punpcklbw %%mm0, %%mm3 \n" /* mm4 now spread over (mm3, mm4) */
"pmaddwd %%mm2, %%mm2 \n"
"pmaddwd %%mm4, %%mm4 \n"
"pmaddwd %%mm1, %%mm1 \n"
"pmaddwd %%mm3, %%mm3 \n"
"lea (%0, %3, 2), %0 \n" /* pix1 += 2 * line_size */
"lea (%1, %3, 2), %1 \n" /* pix2 += 2 * line_size */
"paddd %%mm2, %%mm1 \n"
"paddd %%mm4, %%mm3 \n"
"paddd %%mm1, %%mm7 \n"
"paddd %%mm3, %%mm7 \n"
"decl %%ecx \n"
"jnz 1b \n"
"movq %%mm7, %%mm1 \n"
"psrlq $32, %%mm7 \n" /* shift hi dword to lo */
"paddd %%mm7, %%mm1 \n"
"movd %%mm1, %2 \n"
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp;
}
static int sse16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int tmp;
__asm__ volatile (
"movl %4, %%ecx\n"
"pxor %%mm0, %%mm0\n" /* mm0 = 0 */
"pxor %%mm7, %%mm7\n" /* mm7 holds the sum */
"1:\n"
"movq (%0), %%mm1\n" /* mm1 = pix1[0 - 7] */
"movq (%1), %%mm2\n" /* mm2 = pix2[0 - 7] */
"movq 8(%0), %%mm3\n" /* mm3 = pix1[8 - 15] */
"movq 8(%1), %%mm4\n" /* mm4 = pix2[8 - 15] */
/* todo: mm1-mm2, mm3-mm4 */
/* algo: subtract mm1 from mm2 with saturation and vice versa */
/* OR the results to get absolute difference */
"movq %%mm1, %%mm5\n"
"movq %%mm3, %%mm6\n"
"psubusb %%mm2, %%mm1\n"
"psubusb %%mm4, %%mm3\n"
"psubusb %%mm5, %%mm2\n"
"psubusb %%mm6, %%mm4\n"
"por %%mm1, %%mm2\n"
"por %%mm3, %%mm4\n"
/* now convert to 16-bit vectors so we can square them */
"movq %%mm2, %%mm1\n"
"movq %%mm4, %%mm3\n"
"punpckhbw %%mm0, %%mm2\n"
"punpckhbw %%mm0, %%mm4\n"
"punpcklbw %%mm0, %%mm1\n" /* mm1 now spread over (mm1, mm2) */
"punpcklbw %%mm0, %%mm3\n" /* mm4 now spread over (mm3, mm4) */
"pmaddwd %%mm2, %%mm2\n"
"pmaddwd %%mm4, %%mm4\n"
"pmaddwd %%mm1, %%mm1\n"
"pmaddwd %%mm3, %%mm3\n"
"add %3, %0\n"
"add %3, %1\n"
"paddd %%mm2, %%mm1\n"
"paddd %%mm4, %%mm3\n"
"paddd %%mm1, %%mm7\n"
"paddd %%mm3, %%mm7\n"
"decl %%ecx\n"
"jnz 1b\n"
"movq %%mm7, %%mm1\n"
"psrlq $32, %%mm7\n" /* shift hi dword to lo */
"paddd %%mm7, %%mm1\n"
"movd %%mm1, %2\n"
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp;
}
static int hf_noise8_mmx(uint8_t *pix1, int line_size, int h)
{
int tmp;
__asm__ volatile (
"movl %3, %%ecx\n"
"pxor %%mm7, %%mm7\n"
"pxor %%mm6, %%mm6\n"
"movq (%0), %%mm0\n"
"movq %%mm0, %%mm1\n"
"psllq $8, %%mm0\n"
"psrlq $8, %%mm1\n"
"psrlq $8, %%mm0\n"
"movq %%mm0, %%mm2\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm0\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm2\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm0\n"
"psubw %%mm3, %%mm2\n"
"add %2, %0\n"
"movq (%0), %%mm4\n"
"movq %%mm4, %%mm1\n"
"psllq $8, %%mm4\n"
"psrlq $8, %%mm1\n"
"psrlq $8, %%mm4\n"
"movq %%mm4, %%mm5\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm4\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm5\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm4\n"
"psubw %%mm3, %%mm5\n"
"psubw %%mm4, %%mm0\n"
"psubw %%mm5, %%mm2\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm0, %%mm3\n\t"
"pcmpgtw %%mm2, %%mm1\n\t"
"pxor %%mm3, %%mm0\n"
"pxor %%mm1, %%mm2\n"
"psubw %%mm3, %%mm0\n"
"psubw %%mm1, %%mm2\n"
"paddw %%mm0, %%mm2\n"
"paddw %%mm2, %%mm6\n"
"add %2, %0\n"
"1:\n"
"movq (%0), %%mm0\n"
"movq %%mm0, %%mm1\n"
"psllq $8, %%mm0\n"
"psrlq $8, %%mm1\n"
"psrlq $8, %%mm0\n"
"movq %%mm0, %%mm2\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm0\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm2\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm0\n"
"psubw %%mm3, %%mm2\n"
"psubw %%mm0, %%mm4\n"
"psubw %%mm2, %%mm5\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm4, %%mm3\n\t"
"pcmpgtw %%mm5, %%mm1\n\t"
"pxor %%mm3, %%mm4\n"
"pxor %%mm1, %%mm5\n"
"psubw %%mm3, %%mm4\n"
"psubw %%mm1, %%mm5\n"
"paddw %%mm4, %%mm5\n"
"paddw %%mm5, %%mm6\n"
"add %2, %0\n"
"movq (%0), %%mm4\n"
"movq %%mm4, %%mm1\n"
"psllq $8, %%mm4\n"
"psrlq $8, %%mm1\n"
"psrlq $8, %%mm4\n"
"movq %%mm4, %%mm5\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm4\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm5\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm4\n"
"psubw %%mm3, %%mm5\n"
"psubw %%mm4, %%mm0\n"
"psubw %%mm5, %%mm2\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm0, %%mm3\n\t"
"pcmpgtw %%mm2, %%mm1\n\t"
"pxor %%mm3, %%mm0\n"
"pxor %%mm1, %%mm2\n"
"psubw %%mm3, %%mm0\n"
"psubw %%mm1, %%mm2\n"
"paddw %%mm0, %%mm2\n"
"paddw %%mm2, %%mm6\n"
"add %2, %0\n"
"subl $2, %%ecx\n"
" jnz 1b\n"
"movq %%mm6, %%mm0\n"
"punpcklwd %%mm7, %%mm0\n"
"punpckhwd %%mm7, %%mm6\n"
"paddd %%mm0, %%mm6\n"
"movq %%mm6, %%mm0\n"
"psrlq $32, %%mm6\n"
"paddd %%mm6, %%mm0\n"
"movd %%mm0, %1\n"
: "+r" (pix1), "=r" (tmp)
: "r" ((x86_reg) line_size), "g" (h - 2)
: "%ecx");
return tmp;
}
static int hf_noise16_mmx(uint8_t *pix1, int line_size, int h)
{
int tmp;
uint8_t *pix = pix1;
__asm__ volatile (
"movl %3, %%ecx\n"
"pxor %%mm7, %%mm7\n"
"pxor %%mm6, %%mm6\n"
"movq (%0), %%mm0\n"
"movq 1(%0), %%mm1\n"
"movq %%mm0, %%mm2\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm0\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm2\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm0\n"
"psubw %%mm3, %%mm2\n"
"add %2, %0\n"
"movq (%0), %%mm4\n"
"movq 1(%0), %%mm1\n"
"movq %%mm4, %%mm5\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm4\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm5\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm4\n"
"psubw %%mm3, %%mm5\n"
"psubw %%mm4, %%mm0\n"
"psubw %%mm5, %%mm2\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm0, %%mm3\n\t"
"pcmpgtw %%mm2, %%mm1\n\t"
"pxor %%mm3, %%mm0\n"
"pxor %%mm1, %%mm2\n"
"psubw %%mm3, %%mm0\n"
"psubw %%mm1, %%mm2\n"
"paddw %%mm0, %%mm2\n"
"paddw %%mm2, %%mm6\n"
"add %2, %0\n"
"1:\n"
"movq (%0), %%mm0\n"
"movq 1(%0), %%mm1\n"
"movq %%mm0, %%mm2\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm0\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm2\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm0\n"
"psubw %%mm3, %%mm2\n"
"psubw %%mm0, %%mm4\n"
"psubw %%mm2, %%mm5\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm4, %%mm3\n\t"
"pcmpgtw %%mm5, %%mm1\n\t"
"pxor %%mm3, %%mm4\n"
"pxor %%mm1, %%mm5\n"
"psubw %%mm3, %%mm4\n"
"psubw %%mm1, %%mm5\n"
"paddw %%mm4, %%mm5\n"
"paddw %%mm5, %%mm6\n"
"add %2, %0\n"
"movq (%0), %%mm4\n"
"movq 1(%0), %%mm1\n"
"movq %%mm4, %%mm5\n"
"movq %%mm1, %%mm3\n"
"punpcklbw %%mm7, %%mm4\n"
"punpcklbw %%mm7, %%mm1\n"
"punpckhbw %%mm7, %%mm5\n"
"punpckhbw %%mm7, %%mm3\n"
"psubw %%mm1, %%mm4\n"
"psubw %%mm3, %%mm5\n"
"psubw %%mm4, %%mm0\n"
"psubw %%mm5, %%mm2\n"
"pxor %%mm3, %%mm3\n"
"pxor %%mm1, %%mm1\n"
"pcmpgtw %%mm0, %%mm3\n\t"
"pcmpgtw %%mm2, %%mm1\n\t"
"pxor %%mm3, %%mm0\n"
"pxor %%mm1, %%mm2\n"
"psubw %%mm3, %%mm0\n"
"psubw %%mm1, %%mm2\n"
"paddw %%mm0, %%mm2\n"
"paddw %%mm2, %%mm6\n"
"add %2, %0\n"
"subl $2, %%ecx\n"
" jnz 1b\n"
"movq %%mm6, %%mm0\n"
"punpcklwd %%mm7, %%mm0\n"
"punpckhwd %%mm7, %%mm6\n"
"paddd %%mm0, %%mm6\n"
"movq %%mm6, %%mm0\n"
"psrlq $32, %%mm6\n"
"paddd %%mm6, %%mm0\n"
"movd %%mm0, %1\n"
: "+r" (pix1), "=r" (tmp)
: "r" ((x86_reg) line_size), "g" (h - 2)
: "%ecx");
return tmp + hf_noise8_mmx(pix + 8, line_size, h);
}
static int nsse16_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int score1, score2;
if (c)
score1 = c->dsp.sse[0](c, pix1, pix2, line_size, h);
else
score1 = sse16_mmx(c, pix1, pix2, line_size, h);
score2 = hf_noise16_mmx(pix1, line_size, h) -
hf_noise16_mmx(pix2, line_size, h);
if (c)
return score1 + FFABS(score2) * c->avctx->nsse_weight;
else
return score1 + FFABS(score2) * 8;
}
static int nsse8_mmx(MpegEncContext *c, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int score1 = sse8_mmx(c, pix1, pix2, line_size, h);
int score2 = hf_noise8_mmx(pix1, line_size, h) -
hf_noise8_mmx(pix2, line_size, h);
if (c)
return score1 + FFABS(score2) * c->avctx->nsse_weight;
else
return score1 + FFABS(score2) * 8;
}
static int vsad_intra16_mmx(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
int line_size, int h)
{
int tmp;
assert((((int) pix) & 7) == 0);
assert((line_size & 7) == 0);
#define SUM(in0, in1, out0, out1) \
"movq (%0), %%mm2\n" \
"movq 8(%0), %%mm3\n" \
"add %2,%0\n" \
"movq %%mm2, " #out0 "\n" \
"movq %%mm3, " #out1 "\n" \
"psubusb " #in0 ", %%mm2\n" \
"psubusb " #in1 ", %%mm3\n" \
"psubusb " #out0 ", " #in0 "\n" \
"psubusb " #out1 ", " #in1 "\n" \
"por %%mm2, " #in0 "\n" \
"por %%mm3, " #in1 "\n" \
"movq " #in0 ", %%mm2\n" \
"movq " #in1 ", %%mm3\n" \
"punpcklbw %%mm7, " #in0 "\n" \
"punpcklbw %%mm7, " #in1 "\n" \
"punpckhbw %%mm7, %%mm2\n" \
"punpckhbw %%mm7, %%mm3\n" \
"paddw " #in1 ", " #in0 "\n" \
"paddw %%mm3, %%mm2\n" \
"paddw %%mm2, " #in0 "\n" \
"paddw " #in0 ", %%mm6\n"
__asm__ volatile (
"movl %3, %%ecx\n"
"pxor %%mm6, %%mm6\n"
"pxor %%mm7, %%mm7\n"
"movq (%0), %%mm0\n"
"movq 8(%0), %%mm1\n"
"add %2, %0\n"
"jmp 2f\n"
"1:\n"
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
"2:\n"
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
"subl $2, %%ecx\n"
"jnz 1b\n"
"movq %%mm6, %%mm0\n"
"psrlq $32, %%mm6\n"
"paddw %%mm6, %%mm0\n"
"movq %%mm0, %%mm6\n"
"psrlq $16, %%mm0\n"
"paddw %%mm6, %%mm0\n"
"movd %%mm0, %1\n"
: "+r" (pix), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp & 0xFFFF;
}
#undef SUM
static int vsad_intra16_mmxext(MpegEncContext *v, uint8_t *pix, uint8_t *dummy,
int line_size, int h)
{
int tmp;
assert((((int) pix) & 7) == 0);
assert((line_size & 7) == 0);
#define SUM(in0, in1, out0, out1) \
"movq (%0), " #out0 "\n" \
"movq 8(%0), " #out1 "\n" \
"add %2, %0\n" \
"psadbw " #out0 ", " #in0 "\n" \
"psadbw " #out1 ", " #in1 "\n" \
"paddw " #in1 ", " #in0 "\n" \
"paddw " #in0 ", %%mm6\n"
__asm__ volatile (
"movl %3, %%ecx\n"
"pxor %%mm6, %%mm6\n"
"pxor %%mm7, %%mm7\n"
"movq (%0), %%mm0\n"
"movq 8(%0), %%mm1\n"
"add %2, %0\n"
"jmp 2f\n"
"1:\n"
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
"2:\n"
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
"subl $2, %%ecx\n"
"jnz 1b\n"
"movd %%mm6, %1\n"
: "+r" (pix), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp;
}
#undef SUM
static int vsad16_mmx(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int tmp;
assert((((int) pix1) & 7) == 0);
assert((((int) pix2) & 7) == 0);
assert((line_size & 7) == 0);
#define SUM(in0, in1, out0, out1) \
"movq (%0), %%mm2\n" \
"movq (%1), " #out0 "\n" \
"movq 8(%0), %%mm3\n" \
"movq 8(%1), " #out1 "\n" \
"add %3, %0\n" \
"add %3, %1\n" \
"psubb " #out0 ", %%mm2\n" \
"psubb " #out1 ", %%mm3\n" \
"pxor %%mm7, %%mm2\n" \
"pxor %%mm7, %%mm3\n" \
"movq %%mm2, " #out0 "\n" \
"movq %%mm3, " #out1 "\n" \
"psubusb " #in0 ", %%mm2\n" \
"psubusb " #in1 ", %%mm3\n" \
"psubusb " #out0 ", " #in0 "\n" \
"psubusb " #out1 ", " #in1 "\n" \
"por %%mm2, " #in0 "\n" \
"por %%mm3, " #in1 "\n" \
"movq " #in0 ", %%mm2\n" \
"movq " #in1 ", %%mm3\n" \
"punpcklbw %%mm7, " #in0 "\n" \
"punpcklbw %%mm7, " #in1 "\n" \
"punpckhbw %%mm7, %%mm2\n" \
"punpckhbw %%mm7, %%mm3\n" \
"paddw " #in1 ", " #in0 "\n" \
"paddw %%mm3, %%mm2\n" \
"paddw %%mm2, " #in0 "\n" \
"paddw " #in0 ", %%mm6\n"
__asm__ volatile (
"movl %4, %%ecx\n"
"pxor %%mm6, %%mm6\n"
"pcmpeqw %%mm7, %%mm7\n"
"psllw $15, %%mm7\n"
"packsswb %%mm7, %%mm7\n"
"movq (%0), %%mm0\n"
"movq (%1), %%mm2\n"
"movq 8(%0), %%mm1\n"
"movq 8(%1), %%mm3\n"
"add %3, %0\n"
"add %3, %1\n"
"psubb %%mm2, %%mm0\n"
"psubb %%mm3, %%mm1\n"
"pxor %%mm7, %%mm0\n"
"pxor %%mm7, %%mm1\n"
"jmp 2f\n"
"1:\n"
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
"2:\n"
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
"subl $2, %%ecx\n"
"jnz 1b\n"
"movq %%mm6, %%mm0\n"
"psrlq $32, %%mm6\n"
"paddw %%mm6, %%mm0\n"
"movq %%mm0, %%mm6\n"
"psrlq $16, %%mm0\n"
"paddw %%mm6, %%mm0\n"
"movd %%mm0, %2\n"
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp & 0x7FFF;
}
#undef SUM
static int vsad16_mmxext(MpegEncContext *v, uint8_t *pix1, uint8_t *pix2,
int line_size, int h)
{
int tmp;
assert((((int) pix1) & 7) == 0);
assert((((int) pix2) & 7) == 0);
assert((line_size & 7) == 0);
#define SUM(in0, in1, out0, out1) \
"movq (%0), " #out0 "\n" \
"movq (%1), %%mm2\n" \
"movq 8(%0), " #out1 "\n" \
"movq 8(%1), %%mm3\n" \
"add %3, %0\n" \
"add %3, %1\n" \
"psubb %%mm2, " #out0 "\n" \
"psubb %%mm3, " #out1 "\n" \
"pxor %%mm7, " #out0 "\n" \
"pxor %%mm7, " #out1 "\n" \
"psadbw " #out0 ", " #in0 "\n" \
"psadbw " #out1 ", " #in1 "\n" \
"paddw " #in1 ", " #in0 "\n" \
"paddw " #in0 ", %%mm6\n "
__asm__ volatile (
"movl %4, %%ecx\n"
"pxor %%mm6, %%mm6\n"
"pcmpeqw %%mm7, %%mm7\n"
"psllw $15, %%mm7\n"
"packsswb %%mm7, %%mm7\n"
"movq (%0), %%mm0\n"
"movq (%1), %%mm2\n"
"movq 8(%0), %%mm1\n"
"movq 8(%1), %%mm3\n"
"add %3, %0\n"
"add %3, %1\n"
"psubb %%mm2, %%mm0\n"
"psubb %%mm3, %%mm1\n"
"pxor %%mm7, %%mm0\n"
"pxor %%mm7, %%mm1\n"
"jmp 2f\n"
"1:\n"
SUM(%%mm4, %%mm5, %%mm0, %%mm1)
"2:\n"
SUM(%%mm0, %%mm1, %%mm4, %%mm5)
"subl $2, %%ecx\n"
"jnz 1b\n"
"movd %%mm6, %2\n"
: "+r" (pix1), "+r" (pix2), "=r" (tmp)
: "r" ((x86_reg) line_size), "m" (h)
: "%ecx");
return tmp;
}
#undef SUM
#define MMABS_MMX(a,z) \
"pxor " #z ", " #z " \n\t" \
"pcmpgtw " #a ", " #z " \n\t" \
"pxor " #z ", " #a " \n\t" \
"psubw " #z ", " #a " \n\t"
#define MMABS_MMXEXT(a, z) \
"pxor " #z ", " #z " \n\t" \
"psubw " #a ", " #z " \n\t" \
"pmaxsw " #z ", " #a " \n\t"
#define MMABS_SSSE3(a,z) \
"pabsw " #a ", " #a " \n\t"
#define MMABS_SUM(a,z, sum) \
MMABS(a,z) \
"paddusw " #a ", " #sum " \n\t"
/* FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get
* up to about 100k on extreme inputs. But that's very unlikely to occur in
* natural video, and it's even more unlikely to not have any alternative
* mvs/modes with lower cost. */
#define HSUM_MMX(a, t, dst) \
"movq " #a ", " #t " \n\t" \
"psrlq $32, " #a " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"movq " #a ", " #t " \n\t" \
"psrlq $16, " #a " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"movd " #a ", " #dst " \n\t" \
#define HSUM_MMXEXT(a, t, dst) \
"pshufw $0x0E, " #a ", " #t " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"pshufw $0x01, " #a ", " #t " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"movd " #a ", " #dst " \n\t" \
#define HSUM_SSE2(a, t, dst) \
"movhlps " #a ", " #t " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"pshuflw $0x0E, " #a ", " #t " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"pshuflw $0x01, " #a ", " #t " \n\t" \
"paddusw " #t ", " #a " \n\t" \
"movd " #a ", " #dst " \n\t" \