Quelle  mpegaudiodsp.c   Sprache: C

/*
 * SIMD-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 <stddef.h>

#include "config.h"
#include "libavutil/attributes.h"
#include "libavutil/cpu.h"
#include "libavutil/mem_internal.h"
#include "libavutil/x86/asm.h"
#include "libavutil/x86/cpu.h"
#include "libavcodec/mpegaudiodsp.h"

#define DECL(CPU)\
static void imdct36_blocks_ ## CPU(float *out, float *buf, float *in, int count, int switch_point, int block_type);\
void ff_imdct36_float_ ## CPU(float *out, float *buf, float *in, float *win);

#if HAVE_X86ASM
DECL(sse2)
DECL(sse3)
DECL(ssse3)
DECL(avx)
#endif /* HAVE_X86ASM */

void ff_four_imdct36_float_sse(float *out, float *buf, float *in, float *win,
                               float *tmpbuf);
void ff_four_imdct36_float_avx(float *out, float *buf, float *in, float *win,
                               float *tmpbuf);

void ff_dct32_float_sse2(float *out, const float *in);
void ff_dct32_float_avx (float *out, const float *in);

DECLARE_ALIGNED(16, static float, mdct_win_sse)[2][4][4*40];

#if HAVE_6REGS && HAVE_SSE_INLINE

#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)
{
    x86_reg count = - 4*len;
    const float *win1a = win1+len;
    const float *win2a = win2+len;
    const float *bufa  = buf+len;
    float *sum1a = sum1+len;
    float *sum2a = sum2+len;


#define MULT(a, b)                                 \
    "movaps " #a "(%1,%0), %%xmm1 \n\t"  \
    "movaps " #a "(%3,%0), %%xmm2 \n\t"  \
    "mulps %%xmm2, %%xmm1 \n\t"  \
    "subps %%xmm1, %%xmm0 \n\t"  \
    "mulps " #b "(%2,%0), %%xmm2 \n\t"  \
    "subps %%xmm2, %%xmm4 \n\t"  \

    __asm__ volatile(
            "1: \n\t"
            "xorps %%xmm0, %%xmm0 \n\t"
            "xorps %%xmm4, %%xmm4 \n\t"

            MULT(   0,   0)
            MULT( 256,  64)
            MULT( 512, 128)
            MULT( 768, 192)
            MULT(1024, 256)
            MULT(1280, 320)
            MULT(1536, 384)
            MULT(1792, 448)

            "movaps %%xmm0, (%4,%0) \n\t"
            "movaps %%xmm4, (%5,%0) \n\t"
            "add $16, %0 \n\t"
            "jl 1b \n\t"
            :"+&r"(count)
            :"r"(win1a), "r"(win2a), "r"(bufa), "r"(sum1a), "r"(sum2a)
            );

#undef MULT
}

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;

    /* copy to avoid wrap */
    __asm__ volatile(
            "movaps 0(%0), %%xmm0 \n\t" \
            "movaps 16(%0), %%xmm1 \n\t" \
            "movaps 32(%0), %%xmm2 \n\t" \
            "movaps 48(%0), %%xmm3 \n\t" \
            "movaps %%xmm0, 0(%1) \n\t" \
            "movaps %%xmm1, 16(%1) \n\t" \
            "movaps %%xmm2, 32(%1) \n\t" \
            "movaps %%xmm3, 48(%1) \n\t" \
            "movaps 64(%0), %%xmm0 \n\t" \
            "movaps 80(%0), %%xmm1 \n\t" \
            "movaps 96(%0), %%xmm2 \n\t" \
            "movaps 112(%0), %%xmm3 \n\t" \
            "movaps %%xmm0, 64(%1) \n\t" \
            "movaps %%xmm1, 80(%1) \n\t" \
            "movaps %%xmm2, 96(%1) \n\t" \
            "movaps %%xmm3, 112(%1) \n\t"
            ::"r"(in), "r"(in+512)
            :"memory"
            );

    apply_window(in + 16, win     , win + 512, suma, sumc, 16);
    apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16);

    SUM8(MACS, suma[0], win + 32, in + 48);

    sumc[ 0] = 0;
    sumb[16] = 0;
    sumd[16] = 0;

#define SUMS(suma, sumb, sumc, sumd, out1, out2)               \
            "movups " #sumd "(%4), %%xmm0 \n\t" \
            "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \
            "subps " #suma "(%1), %%xmm0 \n\t" \
            "movaps %%xmm0," #out1 "(%0) \n\t" \
\
            "movups " #sumc "(%3), %%xmm0 \n\t" \
            "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \
            "addps " #sumb "(%2), %%xmm0 \n\t" \
            "movaps %%xmm0," #out2 "(%0) \n\t"

    if (incr == 1) {
        __asm__ volatile(
            SUMS( 0, 48,  4, 52,  0, 112)
            SUMS(16, 32, 20, 36, 16,  96)
            SUMS(32, 16, 36, 20, 32,  80)
            SUMS(48,  0, 52,  4, 48,  64)

            :"+&r"(out)
            :"r"(&suma[0]), "r"(&sumb[0]), "r"(&sumc[0]), "r"(&sumd[0])
            :"memory"
            );
        out += 16*incr;
    } else {
        int j;
        float *out2 = out + 32 * incr;
        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_6REGS && HAVE_SSE_INLINE */

#if HAVE_X86ASM
#define DECL_IMDCT_BLOCKS(CPU1, CPU2)                                       \
static void imdct36_blocks_ ## CPU1(float *out, float *buf, float *in,      \
                               int count, int switch_point, int block_type) \
{                                                                           \
    int align_end = count - (count & 3);                                \
    int j;                                                              \
    for (j = 0; j < align_end; j+= 4) {                                 \
        LOCAL_ALIGNED_16(float, tmpbuf, [1024]);                        \
        float *win = mdct_win_sse[switch_point && j < 4][block_type];   \
        /* apply window & overlap with previous buffer */               \
                                                                        \
        /* select window */                                             \
        ff_four_imdct36_float_ ## CPU2(out, buf, in, win, tmpbuf);      \
        in      += 4*18;                                                \
        buf     += 4*18;                                                \
        out     += 4;                                                   \
    }                                                                   \
    for (; j < count; j++) {                                            \
        /* apply window & overlap with previous buffer */               \
                                                                        \
        /* select window */                                             \
        int win_idx = (switch_point && j < 2) ? 0 : block_type;         \
        float *win = ff_mdct_win_float[win_idx + (4 & -(j & 1))];       \
                                                                        \
        ff_imdct36_float_ ## CPU1(out, buf, in, win);                   \
                                                                        \
        in  += 18;                                                      \
        buf++;                                                          \
        out++;                                                          \
    }                                                                   \
}

#if HAVE_SSE
DECL_IMDCT_BLOCKS(sse2,sse)
DECL_IMDCT_BLOCKS(sse3,sse)
DECL_IMDCT_BLOCKS(ssse3,sse)
#endif
#if HAVE_AVX_EXTERNAL
DECL_IMDCT_BLOCKS(avx,avx)
#endif
#endif /* HAVE_X86ASM */

av_cold void ff_mpadsp_init_x86_tabs(void)
{
    int i, j;
    for (j = 0; j < 4; j++) {
        for (i = 0; i < 40; i ++) {
            mdct_win_sse[0][j][4*i    ] = ff_mdct_win_float[j    ][i];
            mdct_win_sse[0][j][4*i + 1] = ff_mdct_win_float[j + 4][i];
            mdct_win_sse[0][j][4*i + 2] = ff_mdct_win_float[j    ][i];
            mdct_win_sse[0][j][4*i + 3] = ff_mdct_win_float[j + 4][i];
            mdct_win_sse[1][j][4*i    ] = ff_mdct_win_float[0    ][i];
            mdct_win_sse[1][j][4*i + 1] = ff_mdct_win_float[4    ][i];
            mdct_win_sse[1][j][4*i + 2] = ff_mdct_win_float[j    ][i];
            mdct_win_sse[1][j][4*i + 3] = ff_mdct_win_float[j + 4][i];
        }
    }
}

av_cold void ff_mpadsp_init_x86(MPADSPContext *s)
{
    av_unused int cpu_flags = av_get_cpu_flags();

#if HAVE_6REGS && HAVE_SSE_INLINE
    if (INLINE_SSE(cpu_flags)) {
        s->apply_window_float = apply_window_mp3;
    }
#endif /* HAVE_SSE_INLINE */

#if HAVE_X86ASM
#if HAVE_SSE
    if (EXTERNAL_SSE2(cpu_flags)) {
        s->imdct36_blocks_float = imdct36_blocks_sse2;
        s->dct32_float          = ff_dct32_float_sse2;
    }
    if (EXTERNAL_SSE3(cpu_flags)) {
        s->imdct36_blocks_float = imdct36_blocks_sse3;
    }
    if (EXTERNAL_SSSE3(cpu_flags)) {
        s->imdct36_blocks_float = imdct36_blocks_ssse3;
    }
#endif
#if HAVE_AVX_EXTERNAL
    if (EXTERNAL_AVX(cpu_flags)) {
        s->imdct36_blocks_float = imdct36_blocks_avx;
    }
    if (EXTERNAL_AVX_FAST(cpu_flags))
        s->dct32_float          = ff_dct32_float_avx;
#endif
#endif /* HAVE_X86ASM */
}