yading@10: /* yading@10: * x86-optimized AC-3 DSP utils yading@10: * Copyright (c) 2011 Justin Ruggles 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/mem.h" yading@10: #include "libavutil/x86/asm.h" yading@10: #include "libavutil/x86/cpu.h" yading@10: #include "dsputil_mmx.h" yading@10: #include "libavcodec/ac3.h" yading@10: #include "libavcodec/ac3dsp.h" yading@10: yading@10: void ff_ac3_exponent_min_mmx (uint8_t *exp, int num_reuse_blocks, int nb_coefs); yading@10: void ff_ac3_exponent_min_mmxext(uint8_t *exp, int num_reuse_blocks, int nb_coefs); yading@10: void ff_ac3_exponent_min_sse2 (uint8_t *exp, int num_reuse_blocks, int nb_coefs); yading@10: yading@10: int ff_ac3_max_msb_abs_int16_mmx (const int16_t *src, int len); yading@10: int ff_ac3_max_msb_abs_int16_mmxext(const int16_t *src, int len); yading@10: int ff_ac3_max_msb_abs_int16_sse2 (const int16_t *src, int len); yading@10: int ff_ac3_max_msb_abs_int16_ssse3(const int16_t *src, int len); yading@10: yading@10: void ff_ac3_lshift_int16_mmx (int16_t *src, unsigned int len, unsigned int shift); yading@10: void ff_ac3_lshift_int16_sse2(int16_t *src, unsigned int len, unsigned int shift); yading@10: yading@10: void ff_ac3_rshift_int32_mmx (int32_t *src, unsigned int len, unsigned int shift); yading@10: void ff_ac3_rshift_int32_sse2(int32_t *src, unsigned int len, unsigned int shift); yading@10: yading@10: void ff_float_to_fixed24_3dnow(int32_t *dst, const float *src, unsigned int len); yading@10: void ff_float_to_fixed24_sse (int32_t *dst, const float *src, unsigned int len); yading@10: void ff_float_to_fixed24_sse2 (int32_t *dst, const float *src, unsigned int len); yading@10: yading@10: int ff_ac3_compute_mantissa_size_sse2(uint16_t mant_cnt[6][16]); yading@10: yading@10: void ff_ac3_extract_exponents_3dnow(uint8_t *exp, int32_t *coef, int nb_coefs); yading@10: void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int nb_coefs); yading@10: void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int nb_coefs); yading@10: yading@10: #if ARCH_X86_32 && defined(__INTEL_COMPILER) yading@10: # undef HAVE_7REGS yading@10: # define HAVE_7REGS 0 yading@10: #endif yading@10: yading@10: #if HAVE_SSE_INLINE && HAVE_7REGS yading@10: yading@10: #define IF1(x) x yading@10: #define IF0(x) yading@10: yading@10: #define MIX5(mono, stereo) \ yading@10: __asm__ volatile ( \ yading@10: "movss 0(%1), %%xmm5 \n" \ yading@10: "movss 8(%1), %%xmm6 \n" \ yading@10: "movss 24(%1), %%xmm7 \n" \ yading@10: "shufps $0, %%xmm5, %%xmm5 \n" \ yading@10: "shufps $0, %%xmm6, %%xmm6 \n" \ yading@10: "shufps $0, %%xmm7, %%xmm7 \n" \ yading@10: "1: \n" \ yading@10: "movaps (%0, %2), %%xmm0 \n" \ yading@10: "movaps (%0, %3), %%xmm1 \n" \ yading@10: "movaps (%0, %4), %%xmm2 \n" \ yading@10: "movaps (%0, %5), %%xmm3 \n" \ yading@10: "movaps (%0, %6), %%xmm4 \n" \ yading@10: "mulps %%xmm5, %%xmm0 \n" \ yading@10: "mulps %%xmm6, %%xmm1 \n" \ yading@10: "mulps %%xmm5, %%xmm2 \n" \ yading@10: "mulps %%xmm7, %%xmm3 \n" \ yading@10: "mulps %%xmm7, %%xmm4 \n" \ yading@10: stereo("addps %%xmm1, %%xmm0 \n") \ yading@10: "addps %%xmm1, %%xmm2 \n" \ yading@10: "addps %%xmm3, %%xmm0 \n" \ yading@10: "addps %%xmm4, %%xmm2 \n" \ yading@10: mono("addps %%xmm2, %%xmm0 \n") \ yading@10: "movaps %%xmm0, (%0, %2) \n" \ yading@10: stereo("movaps %%xmm2, (%0, %3) \n") \ yading@10: "add $16, %0 \n" \ yading@10: "jl 1b \n" \ yading@10: : "+&r"(i) \ yading@10: : "r"(matrix), \ yading@10: "r"(samples[0] + len), \ yading@10: "r"(samples[1] + len), \ yading@10: "r"(samples[2] + len), \ yading@10: "r"(samples[3] + len), \ yading@10: "r"(samples[4] + len) \ yading@10: : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", \ yading@10: "%xmm4", "%xmm5", "%xmm6", "%xmm7",) \ yading@10: "memory" \ yading@10: ); yading@10: yading@10: #define MIX_MISC(stereo) \ yading@10: __asm__ volatile ( \ yading@10: "mov %5, %2 \n" \ yading@10: "1: \n" \ yading@10: "mov -%c7(%6, %2, %c8), %3 \n" \ yading@10: "movaps (%3, %0), %%xmm0 \n" \ yading@10: stereo("movaps %%xmm0, %%xmm1 \n") \ yading@10: "mulps %%xmm4, %%xmm0 \n" \ yading@10: stereo("mulps %%xmm5, %%xmm1 \n") \ yading@10: "2: \n" \ yading@10: "mov (%6, %2, %c8), %1 \n" \ yading@10: "movaps (%1, %0), %%xmm2 \n" \ yading@10: stereo("movaps %%xmm2, %%xmm3 \n") \ yading@10: "mulps (%4, %2, 8), %%xmm2 \n" \ yading@10: stereo("mulps 16(%4, %2, 8), %%xmm3 \n") \ yading@10: "addps %%xmm2, %%xmm0 \n" \ yading@10: stereo("addps %%xmm3, %%xmm1 \n") \ yading@10: "add $4, %2 \n" \ yading@10: "jl 2b \n" \ yading@10: "mov %5, %2 \n" \ yading@10: stereo("mov (%6, %2, %c8), %1 \n") \ yading@10: "movaps %%xmm0, (%3, %0) \n" \ yading@10: stereo("movaps %%xmm1, (%1, %0) \n") \ yading@10: "add $16, %0 \n" \ yading@10: "jl 1b \n" \ yading@10: : "+&r"(i), "=&r"(j), "=&r"(k), "=&r"(m) \ yading@10: : "r"(matrix_simd + in_ch), \ yading@10: "g"((intptr_t) - 4 * (in_ch - 1)), \ yading@10: "r"(samp + in_ch), \ yading@10: "i"(sizeof(float *)), "i"(sizeof(float *)/4) \ yading@10: : "memory" \ yading@10: ); yading@10: yading@10: static void ac3_downmix_sse(float **samples, float (*matrix)[2], yading@10: int out_ch, int in_ch, int len) yading@10: { yading@10: int (*matrix_cmp)[2] = (int(*)[2])matrix; yading@10: intptr_t i, j, k, m; yading@10: yading@10: i = -len * sizeof(float); yading@10: if (in_ch == 5 && out_ch == 2 && yading@10: !(matrix_cmp[0][1] | matrix_cmp[2][0] | yading@10: matrix_cmp[3][1] | matrix_cmp[4][0] | yading@10: (matrix_cmp[1][0] ^ matrix_cmp[1][1]) | yading@10: (matrix_cmp[0][0] ^ matrix_cmp[2][1]))) { yading@10: MIX5(IF0, IF1); yading@10: } else if (in_ch == 5 && out_ch == 1 && yading@10: matrix_cmp[0][0] == matrix_cmp[2][0] && yading@10: matrix_cmp[3][0] == matrix_cmp[4][0]) { yading@10: MIX5(IF1, IF0); yading@10: } else { yading@10: DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4]; yading@10: float *samp[AC3_MAX_CHANNELS]; yading@10: yading@10: for (j = 0; j < in_ch; j++) yading@10: samp[j] = samples[j] + len; yading@10: yading@10: j = 2 * in_ch * sizeof(float); yading@10: __asm__ volatile ( yading@10: "1: \n" yading@10: "sub $8, %0 \n" yading@10: "movss (%2, %0), %%xmm4 \n" yading@10: "movss 4(%2, %0), %%xmm5 \n" yading@10: "shufps $0, %%xmm4, %%xmm4 \n" yading@10: "shufps $0, %%xmm5, %%xmm5 \n" yading@10: "movaps %%xmm4, (%1, %0, 4) \n" yading@10: "movaps %%xmm5, 16(%1, %0, 4) \n" yading@10: "jg 1b \n" yading@10: : "+&r"(j) yading@10: : "r"(matrix_simd), "r"(matrix) yading@10: : "memory" yading@10: ); yading@10: if (out_ch == 2) { yading@10: MIX_MISC(IF1); yading@10: } else { yading@10: MIX_MISC(IF0); yading@10: } yading@10: } yading@10: } yading@10: yading@10: #endif /* HAVE_SSE_INLINE && HAVE_7REGS */ yading@10: yading@10: av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact) yading@10: { yading@10: int mm_flags = av_get_cpu_flags(); yading@10: yading@10: if (EXTERNAL_MMX(mm_flags)) { yading@10: c->ac3_exponent_min = ff_ac3_exponent_min_mmx; yading@10: c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmx; yading@10: c->ac3_lshift_int16 = ff_ac3_lshift_int16_mmx; yading@10: c->ac3_rshift_int32 = ff_ac3_rshift_int32_mmx; yading@10: } yading@10: if (EXTERNAL_AMD3DNOW(mm_flags)) { yading@10: c->extract_exponents = ff_ac3_extract_exponents_3dnow; yading@10: if (!bit_exact) { yading@10: c->float_to_fixed24 = ff_float_to_fixed24_3dnow; yading@10: } yading@10: } yading@10: if (EXTERNAL_MMXEXT(mm_flags)) { yading@10: c->ac3_exponent_min = ff_ac3_exponent_min_mmxext; yading@10: c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmxext; yading@10: } yading@10: if (EXTERNAL_SSE(mm_flags)) { yading@10: c->float_to_fixed24 = ff_float_to_fixed24_sse; yading@10: } yading@10: if (EXTERNAL_SSE2(mm_flags)) { yading@10: c->ac3_exponent_min = ff_ac3_exponent_min_sse2; yading@10: c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_sse2; yading@10: c->float_to_fixed24 = ff_float_to_fixed24_sse2; yading@10: c->compute_mantissa_size = ff_ac3_compute_mantissa_size_sse2; yading@10: c->extract_exponents = ff_ac3_extract_exponents_sse2; yading@10: if (!(mm_flags & AV_CPU_FLAG_SSE2SLOW)) { yading@10: c->ac3_lshift_int16 = ff_ac3_lshift_int16_sse2; yading@10: c->ac3_rshift_int32 = ff_ac3_rshift_int32_sse2; yading@10: } yading@10: } yading@10: if (EXTERNAL_SSSE3(mm_flags)) { yading@10: c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_ssse3; yading@10: if (!(mm_flags & AV_CPU_FLAG_ATOM)) { yading@10: c->extract_exponents = ff_ac3_extract_exponents_ssse3; yading@10: } yading@10: } yading@10: yading@10: #if HAVE_SSE_INLINE && HAVE_7REGS yading@10: if (INLINE_SSE(mm_flags)) { yading@10: c->downmix = ac3_downmix_sse; yading@10: } yading@10: #endif yading@10: }