yading@10: /* yading@10: * MMX optimized MP3 decoding functions yading@10: * Copyright (c) 2010 Vitor Sessak yading@10: * yading@10: * This file is part of FFmpeg. yading@10: * yading@10: * FFmpeg is free software; you can redistribute it and/or yading@10: * modify it under the terms of the GNU Lesser General Public yading@10: * License as published by the Free Software Foundation; either yading@10: * version 2.1 of the License, or (at your option) any later version. yading@10: * yading@10: * FFmpeg is distributed in the hope that it will be useful, yading@10: * but WITHOUT ANY WARRANTY; without even the implied warranty of yading@10: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU yading@10: * Lesser General Public License for more details. yading@10: * yading@10: * You should have received a copy of the GNU Lesser General Public yading@10: * License along with FFmpeg; if not, write to the Free Software yading@10: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA yading@10: */ yading@10: yading@10: #include "libavutil/attributes.h" yading@10: #include "libavutil/cpu.h" yading@10: #include "libavutil/internal.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavutil/x86/cpu.h" yading@10: #include "libavcodec/mpegaudiodsp.h" yading@10: yading@10: #define DECL(CPU)\ yading@10: static void imdct36_blocks_ ## CPU(float *out, float *buf, float *in, int count, int switch_point, int block_type);\ yading@10: void ff_imdct36_float_ ## CPU(float *out, float *buf, float *in, float *win); yading@10: yading@10: DECL(sse) yading@10: DECL(sse2) yading@10: DECL(sse3) yading@10: DECL(ssse3) yading@10: DECL(avx) yading@10: yading@10: void ff_four_imdct36_float_sse(float *out, float *buf, float *in, float *win, yading@10: float *tmpbuf); yading@10: void ff_four_imdct36_float_avx(float *out, float *buf, float *in, float *win, yading@10: float *tmpbuf); yading@10: yading@10: DECLARE_ALIGNED(16, static float, mdct_win_sse)[2][4][4*40]; yading@10: yading@10: #if HAVE_SSE2_INLINE yading@10: yading@10: #define MACS(rt, ra, rb) rt+=(ra)*(rb) yading@10: #define MLSS(rt, ra, rb) rt-=(ra)*(rb) yading@10: yading@10: #define SUM8(op, sum, w, p) \ yading@10: { \ yading@10: op(sum, (w)[0 * 64], (p)[0 * 64]); \ yading@10: op(sum, (w)[1 * 64], (p)[1 * 64]); \ yading@10: op(sum, (w)[2 * 64], (p)[2 * 64]); \ yading@10: op(sum, (w)[3 * 64], (p)[3 * 64]); \ yading@10: op(sum, (w)[4 * 64], (p)[4 * 64]); \ yading@10: op(sum, (w)[5 * 64], (p)[5 * 64]); \ yading@10: op(sum, (w)[6 * 64], (p)[6 * 64]); \ yading@10: op(sum, (w)[7 * 64], (p)[7 * 64]); \ yading@10: } yading@10: yading@10: static void apply_window(const float *buf, const float *win1, yading@10: const float *win2, float *sum1, float *sum2, int len) yading@10: { yading@10: x86_reg count = - 4*len; yading@10: const float *win1a = win1+len; yading@10: const float *win2a = win2+len; yading@10: const float *bufa = buf+len; yading@10: float *sum1a = sum1+len; yading@10: float *sum2a = sum2+len; yading@10: yading@10: yading@10: #define MULT(a, b) \ yading@10: "movaps " #a "(%1,%0), %%xmm1 \n\t" \ yading@10: "movaps " #a "(%3,%0), %%xmm2 \n\t" \ yading@10: "mulps %%xmm2, %%xmm1 \n\t" \ yading@10: "subps %%xmm1, %%xmm0 \n\t" \ yading@10: "mulps " #b "(%2,%0), %%xmm2 \n\t" \ yading@10: "subps %%xmm2, %%xmm4 \n\t" \ yading@10: yading@10: __asm__ volatile( yading@10: "1: \n\t" yading@10: "xorps %%xmm0, %%xmm0 \n\t" yading@10: "xorps %%xmm4, %%xmm4 \n\t" yading@10: yading@10: MULT( 0, 0) yading@10: MULT( 256, 64) yading@10: MULT( 512, 128) yading@10: MULT( 768, 192) yading@10: MULT(1024, 256) yading@10: MULT(1280, 320) yading@10: MULT(1536, 384) yading@10: MULT(1792, 448) yading@10: yading@10: "movaps %%xmm0, (%4,%0) \n\t" yading@10: "movaps %%xmm4, (%5,%0) \n\t" yading@10: "add $16, %0 \n\t" yading@10: "jl 1b \n\t" yading@10: :"+&r"(count) yading@10: :"r"(win1a), "r"(win2a), "r"(bufa), "r"(sum1a), "r"(sum2a) yading@10: ); yading@10: yading@10: #undef MULT yading@10: } yading@10: yading@10: static void apply_window_mp3(float *in, float *win, int *unused, float *out, yading@10: int incr) yading@10: { yading@10: LOCAL_ALIGNED_16(float, suma, [17]); yading@10: LOCAL_ALIGNED_16(float, sumb, [17]); yading@10: LOCAL_ALIGNED_16(float, sumc, [17]); yading@10: LOCAL_ALIGNED_16(float, sumd, [17]); yading@10: yading@10: float sum; yading@10: yading@10: /* copy to avoid wrap */ yading@10: __asm__ volatile( yading@10: "movaps 0(%0), %%xmm0 \n\t" \ yading@10: "movaps 16(%0), %%xmm1 \n\t" \ yading@10: "movaps 32(%0), %%xmm2 \n\t" \ yading@10: "movaps 48(%0), %%xmm3 \n\t" \ yading@10: "movaps %%xmm0, 0(%1) \n\t" \ yading@10: "movaps %%xmm1, 16(%1) \n\t" \ yading@10: "movaps %%xmm2, 32(%1) \n\t" \ yading@10: "movaps %%xmm3, 48(%1) \n\t" \ yading@10: "movaps 64(%0), %%xmm0 \n\t" \ yading@10: "movaps 80(%0), %%xmm1 \n\t" \ yading@10: "movaps 96(%0), %%xmm2 \n\t" \ yading@10: "movaps 112(%0), %%xmm3 \n\t" \ yading@10: "movaps %%xmm0, 64(%1) \n\t" \ yading@10: "movaps %%xmm1, 80(%1) \n\t" \ yading@10: "movaps %%xmm2, 96(%1) \n\t" \ yading@10: "movaps %%xmm3, 112(%1) \n\t" yading@10: ::"r"(in), "r"(in+512) yading@10: :"memory" yading@10: ); yading@10: yading@10: apply_window(in + 16, win , win + 512, suma, sumc, 16); yading@10: apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16); yading@10: yading@10: SUM8(MACS, suma[0], win + 32, in + 48); yading@10: yading@10: sumc[ 0] = 0; yading@10: sumb[16] = 0; yading@10: sumd[16] = 0; yading@10: yading@10: #define SUMS(suma, sumb, sumc, sumd, out1, out2) \ yading@10: "movups " #sumd "(%4), %%xmm0 \n\t" \ yading@10: "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \ yading@10: "subps " #suma "(%1), %%xmm0 \n\t" \ yading@10: "movaps %%xmm0," #out1 "(%0) \n\t" \ yading@10: \ yading@10: "movups " #sumc "(%3), %%xmm0 \n\t" \ yading@10: "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \ yading@10: "addps " #sumb "(%2), %%xmm0 \n\t" \ yading@10: "movaps %%xmm0," #out2 "(%0) \n\t" yading@10: yading@10: if (incr == 1) { yading@10: __asm__ volatile( yading@10: SUMS( 0, 48, 4, 52, 0, 112) yading@10: SUMS(16, 32, 20, 36, 16, 96) yading@10: SUMS(32, 16, 36, 20, 32, 80) yading@10: SUMS(48, 0, 52, 4, 48, 64) yading@10: yading@10: :"+&r"(out) yading@10: :"r"(&suma[0]), "r"(&sumb[0]), "r"(&sumc[0]), "r"(&sumd[0]) yading@10: :"memory" yading@10: ); yading@10: out += 16*incr; yading@10: } else { yading@10: int j; yading@10: float *out2 = out + 32 * incr; yading@10: out[0 ] = -suma[ 0]; yading@10: out += incr; yading@10: out2 -= incr; yading@10: for(j=1;j<16;j++) { yading@10: *out = -suma[ j] + sumd[16-j]; yading@10: *out2 = sumb[16-j] + sumc[ j]; yading@10: out += incr; yading@10: out2 -= incr; yading@10: } yading@10: } yading@10: yading@10: sum = 0; yading@10: SUM8(MLSS, sum, win + 16 + 32, in + 32); yading@10: *out = sum; yading@10: } yading@10: yading@10: #endif /* HAVE_SSE2_INLINE */ yading@10: yading@10: #if HAVE_YASM yading@10: #define DECL_IMDCT_BLOCKS(CPU1, CPU2) \ yading@10: static void imdct36_blocks_ ## CPU1(float *out, float *buf, float *in, \ yading@10: int count, int switch_point, int block_type) \ yading@10: { \ yading@10: int align_end = count - (count & 3); \ yading@10: int j; \ yading@10: for (j = 0; j < align_end; j+= 4) { \ yading@10: LOCAL_ALIGNED_16(float, tmpbuf, [1024]); \ yading@10: float *win = mdct_win_sse[switch_point && j < 4][block_type]; \ yading@10: /* apply window & overlap with previous buffer */ \ yading@10: \ yading@10: /* select window */ \ yading@10: ff_four_imdct36_float_ ## CPU2(out, buf, in, win, tmpbuf); \ yading@10: in += 4*18; \ yading@10: buf += 4*18; \ yading@10: out += 4; \ yading@10: } \ yading@10: for (; j < count; j++) { \ yading@10: /* apply window & overlap with previous buffer */ \ yading@10: \ yading@10: /* select window */ \ yading@10: int win_idx = (switch_point && j < 2) ? 0 : block_type; \ yading@10: float *win = ff_mdct_win_float[win_idx + (4 & -(j & 1))]; \ yading@10: \ yading@10: ff_imdct36_float_ ## CPU1(out, buf, in, win); \ yading@10: \ yading@10: in += 18; \ yading@10: buf++; \ yading@10: out++; \ yading@10: } \ yading@10: } yading@10: yading@10: #if HAVE_SSE yading@10: DECL_IMDCT_BLOCKS(sse,sse) yading@10: DECL_IMDCT_BLOCKS(sse2,sse) yading@10: DECL_IMDCT_BLOCKS(sse3,sse) yading@10: DECL_IMDCT_BLOCKS(ssse3,sse) yading@10: #endif yading@10: #if HAVE_AVX_EXTERNAL yading@10: DECL_IMDCT_BLOCKS(avx,avx) yading@10: #endif yading@10: #endif /* HAVE_YASM */ yading@10: yading@10: av_cold void ff_mpadsp_init_x86(MPADSPContext *s) yading@10: { yading@10: int mm_flags = av_get_cpu_flags(); yading@10: yading@10: int i, j; yading@10: for (j = 0; j < 4; j++) { yading@10: for (i = 0; i < 40; i ++) { yading@10: mdct_win_sse[0][j][4*i ] = ff_mdct_win_float[j ][i]; yading@10: mdct_win_sse[0][j][4*i + 1] = ff_mdct_win_float[j + 4][i]; yading@10: mdct_win_sse[0][j][4*i + 2] = ff_mdct_win_float[j ][i]; yading@10: mdct_win_sse[0][j][4*i + 3] = ff_mdct_win_float[j + 4][i]; yading@10: mdct_win_sse[1][j][4*i ] = ff_mdct_win_float[0 ][i]; yading@10: mdct_win_sse[1][j][4*i + 1] = ff_mdct_win_float[4 ][i]; yading@10: mdct_win_sse[1][j][4*i + 2] = ff_mdct_win_float[j ][i]; yading@10: mdct_win_sse[1][j][4*i + 3] = ff_mdct_win_float[j + 4][i]; yading@10: } yading@10: } yading@10: yading@10: #if HAVE_SSE2_INLINE yading@10: if (mm_flags & AV_CPU_FLAG_SSE2) { yading@10: s->apply_window_float = apply_window_mp3; yading@10: } yading@10: #endif /* HAVE_SSE2_INLINE */ yading@10: yading@10: #if HAVE_YASM yading@10: if (EXTERNAL_AVX(mm_flags)) { yading@10: s->imdct36_blocks_float = imdct36_blocks_avx; yading@10: } else if (EXTERNAL_SSSE3(mm_flags)) { yading@10: s->imdct36_blocks_float = imdct36_blocks_ssse3; yading@10: } else if (EXTERNAL_SSE3(mm_flags)) { yading@10: s->imdct36_blocks_float = imdct36_blocks_sse3; yading@10: } else if (EXTERNAL_SSE2(mm_flags)) { yading@10: s->imdct36_blocks_float = imdct36_blocks_sse2; yading@10: } else if (EXTERNAL_SSE(mm_flags)) { yading@10: s->imdct36_blocks_float = imdct36_blocks_sse; yading@10: } yading@10: #endif /* HAVE_YASM */ yading@10: }