Adds a wrapper function for downmixing which detects channel count changes
and updates the selected downmix function accordingly.
---
libavcodec/ac3dec.c | 8 +-
libavcodec/ac3dsp.c | 72 +++++++++---
libavcodec/ac3dsp.h | 8 +-
libavcodec/x86/Makefile | 3 +-
libavcodec/x86/ac3dsp.h | 28 +++++
libavcodec/x86/ac3dsp_downmix.asm | 229 +++++++++++++++++++++++++++++++++++++
libavcodec/x86/ac3dsp_init.c | 185 +++++++++---------------------
7 files changed, 378 insertions(+), 155 deletions(-)
create mode 100644 libavcodec/x86/ac3dsp.h
create mode 100644 libavcodec/x86/ac3dsp_downmix.asm
diff --git a/libavcodec/ac3dec.c b/libavcodec/ac3dec.c
index b2a5eec..eeef3cd 100644
--- a/libavcodec/ac3dec.c
+++ b/libavcodec/ac3dec.c
@@ -1248,19 +1248,19 @@ static int decode_audio_block(AC3DecodeContext *s, int
blk)
do_imdct(s, s->channels);
if (downmix_output) {
- s->ac3dsp.downmix(s->outptr, s->downmix_coeffs,
+ ff_ac3dsp_downmix(&s->ac3dsp, s->outptr, s->downmix_coeffs,
s->out_channels, s->fbw_channels, 256);
}
} else {
if (downmix_output) {
- s->ac3dsp.downmix(s->xcfptr + 1, s->downmix_coeffs,
+ ff_ac3dsp_downmix(&s->ac3dsp, s->xcfptr + 1, s->downmix_coeffs,
s->out_channels, s->fbw_channels, 256);
}
if (downmix_output && !s->downmixed) {
s->downmixed = 1;
- s->ac3dsp.downmix(s->dlyptr, s->downmix_coeffs, s->out_channels,
- s->fbw_channels, 128);
+ ff_ac3dsp_downmix(&s->ac3dsp, s->dlyptr, s->downmix_coeffs,
+ s->out_channels, s->fbw_channels, 128);
}
do_imdct(s, s->out_channels);
diff --git a/libavcodec/ac3dsp.c b/libavcodec/ac3dsp.c
index 5b7e2e7..aa17a90 100644
--- a/libavcodec/ac3dsp.c
+++ b/libavcodec/ac3dsp.c
@@ -24,6 +24,10 @@
#include "ac3.h"
#include "ac3dsp.h"
+#if ARCH_X86
+#include "x86/ac3dsp.h"
+#endif
+
static void ac3_exponent_min_c(uint8_t *exp, int num_reuse_blocks, int
nb_coefs)
{
int blk, i;
@@ -171,18 +175,11 @@ static void ac3_extract_exponents_c(uint8_t *exp, int32_t
*coef, int nb_coefs)
}
}
-static void ac3_downmix_c(float **samples, float **matrix,
- int out_ch, int in_ch, int len)
+static void ac3_downmix_5_to_2_symmetric_c(float **samples, float **matrix,
+ int len)
{
- int **matrix_cmp = (int **)matrix;
- int i, j;
+ int i;
float v0, v1;
-
- if (in_ch == 5 && out_ch == 2 &&
- !(matrix_cmp[1][0] | matrix_cmp[0][2] |
- matrix_cmp[1][3] | matrix_cmp[0][4] |
- (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
- (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
float front_mix = matrix[0][0];
float center_mix = matrix[0][1];
float surround_mix = matrix[0][3];
@@ -199,9 +196,12 @@ static void ac3_downmix_c(float **samples, float **matrix,
samples[0][i] = v0;
samples[1][i] = v1;
}
- } else if (in_ch == 5 && out_ch == 1 &&
- matrix_cmp[0][0] == matrix_cmp[0][2] &&
- matrix_cmp[0][3] == matrix_cmp[0][4]) {
+}
+
+static void ac3_downmix_5_to_1_symmetric_c(float **samples, float **matrix,
+ int len)
+{
+ int i;
float front_mix = matrix[0][0];
float center_mix = matrix[0][1];
float surround_mix = matrix[0][3];
@@ -213,7 +213,15 @@ static void ac3_downmix_c(float **samples, float **matrix,
samples[3][i] * surround_mix +
samples[4][i] * surround_mix;
}
- } else if (out_ch == 2) {
+}
+
+static void ac3_downmix_c(float **samples, float **matrix,
+ int out_ch, int in_ch, int len)
+{
+ int i, j;
+ float v0, v1;
+
+ if (out_ch == 2) {
for (i = 0; i < len; i++) {
v0 = v1 = 0.0f;
for (j = 0; j < in_ch; j++) {
@@ -233,6 +241,38 @@ static void ac3_downmix_c(float **samples, float **matrix,
}
}
+void ff_ac3dsp_downmix(AC3DSPContext *c, float **samples, float **matrix,
+ int out_ch, int in_ch, int len)
+{
+ if (c->in_channels != in_ch || c->out_channels != out_ch) {
+ int **matrix_cmp = (int **)matrix;
+
+ c->in_channels = in_ch;
+ c->out_channels = out_ch;
+ c->downmix = NULL;
+
+ if (in_ch == 5 && out_ch == 2 &&
+ !(matrix_cmp[1][0] | matrix_cmp[0][2] |
+ matrix_cmp[1][3] | matrix_cmp[0][4] |
+ (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
+ (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
+ c->downmix = ac3_downmix_5_to_2_symmetric_c;
+ } else if (in_ch == 5 && out_ch == 1 &&
+ matrix_cmp[0][0] == matrix_cmp[0][2] &&
+ matrix_cmp[0][3] == matrix_cmp[0][4]) {
+ c->downmix = ac3_downmix_5_to_1_symmetric_c;
+ }
+
+ if (ARCH_X86)
+ ff_ac3dsp_set_downmix_x86(c);
+ }
+
+ if (c->downmix)
+ c->downmix(samples, matrix, len);
+ else
+ ac3_downmix_c(samples, matrix, out_ch, in_ch, len);
+}
+
av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)
{
c->ac3_exponent_min = ac3_exponent_min_c;
@@ -244,7 +284,9 @@ av_cold void ff_ac3dsp_init(AC3DSPContext *c, int bit_exact)
c->update_bap_counts = ac3_update_bap_counts_c;
c->compute_mantissa_size = ac3_compute_mantissa_size_c;
c->extract_exponents = ac3_extract_exponents_c;
- c->downmix = ac3_downmix_c;
+ c->in_channels = 0;
+ c->out_channels = 0;
+ c->downmix = NULL;
if (ARCH_ARM)
ff_ac3dsp_init_arm(c, bit_exact);
diff --git a/libavcodec/ac3dsp.h b/libavcodec/ac3dsp.h
index 8f4b769..4fd5a1c 100644
--- a/libavcodec/ac3dsp.h
+++ b/libavcodec/ac3dsp.h
@@ -126,10 +126,14 @@ typedef struct AC3DSPContext {
void (*extract_exponents)(uint8_t *exp, int32_t *coef, int nb_coefs);
- void (*downmix)(float **samples, float **matrix, int out_ch,
- int in_ch, int len);
+ int out_channels;
+ int in_channels;
+ void (*downmix)(float **samples, float **matrix, int len);
} AC3DSPContext;
+void ff_ac3dsp_downmix(AC3DSPContext *c, float **samples, float **matrix,
+ int out_ch, int in_ch, int len);
+
void ff_ac3dsp_init (AC3DSPContext *c, int bit_exact);
void ff_ac3dsp_init_arm(AC3DSPContext *c, int bit_exact);
void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact);
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index 053290b..89db001 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -37,7 +37,8 @@ MMX-OBJS-$(CONFIG_ENCODERS) +=
x86/dsputilenc_mmx.o \
MMX-OBJS-$(CONFIG_VC1_DECODER) += x86/vc1dsp_mmx.o
YASM-OBJS-$(CONFIG_AAC_DECODER) += x86/sbrdsp.o
-YASM-OBJS-$(CONFIG_AC3DSP) += x86/ac3dsp.o
+YASM-OBJS-$(CONFIG_AC3DSP) += x86/ac3dsp.o \
+ x86/ac3dsp_downmix.o
YASM-OBJS-$(CONFIG_DCT) += x86/dct32.o
YASM-OBJS-$(CONFIG_ENCODERS) += x86/dsputilenc.o
YASM-OBJS-$(CONFIG_FFT) += x86/fft.o
diff --git a/libavcodec/x86/ac3dsp.h b/libavcodec/x86/ac3dsp.h
new file mode 100644
index 0000000..fb2a9b6
--- /dev/null
+++ b/libavcodec/x86/ac3dsp.h
@@ -0,0 +1,28 @@
+/*
+ * AC-3 DSP utils
+ *
+ * This file is part of Libav.
+ *
+ * Libav 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.
+ *
+ * Libav 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 Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#ifndef AVCODEC_X86_AC3DSP_H
+#define AVCODEC_X86_AC3DSP_H
+
+#include "libavcodec/ac3dsp.h"
+
+void ff_ac3dsp_set_downmix_x86(AC3DSPContext *c);
+
+#endif /* AVCODEC_X86_AC3DSP_H */
diff --git a/libavcodec/x86/ac3dsp_downmix.asm
b/libavcodec/x86/ac3dsp_downmix.asm
new file mode 100644
index 0000000..913f595
--- /dev/null
+++ b/libavcodec/x86/ac3dsp_downmix.asm
@@ -0,0 +1,229 @@
+;*****************************************************************************
+;* x86-optimized AC-3 downmixing
+;* Copyright (c) 2012 Justin Ruggles
+;*
+;* This file is part of Libav.
+;*
+;* Libav 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.
+;*
+;* Libav 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 Libav; if not, write to the Free Software
+;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+;******************************************************************************
+
+;******************************************************************************
+;* This is based on the channel mixing asm in libavresample, but it is
+;* simplified for only float coefficients and only 3 to 6 channels.
+;******************************************************************************
+
+%include "libavutil/x86/x86util.asm"
+
+SECTION .text
+
+;-----------------------------------------------------------------------------
+; functions to downmix from 3 to 6 channels to mono or stereo
+; void ff_ac3_downmix_*(float **samples, float **matrix, int len);
+;-----------------------------------------------------------------------------
+
+%macro AC3_DOWNMIX 2 ; %1 = in channels, %2 = out channels
+; define some names to make the code clearer
+%assign in_channels %1
+%assign out_channels %2
+%assign stereo out_channels - 1
+
+; determine how many matrix elements must go on the stack vs. mmregs
+%assign matrix_elements in_channels * out_channels
+%if stereo
+ %assign needed_mmregs 4
+%else
+ %assign needed_mmregs 3
+%endif
+%assign matrix_elements_mm num_mmregs - needed_mmregs
+%if matrix_elements < matrix_elements_mm
+ %assign matrix_elements_mm matrix_elements
+%endif
+%assign total_mmregs needed_mmregs+matrix_elements_mm
+%if matrix_elements_mm < matrix_elements
+ %assign matrix_elements_stack matrix_elements - matrix_elements_mm
+%else
+ %assign matrix_elements_stack 0
+%endif
+
+; determine how many gp registers are needed
+%if matrix_elements_stack > 0 && mmsize == 32
+ ; need an extra gp reg to save stack pointer
+ %assign needed_gpregs in_channels+2
+%else
+ %assign needed_gpregs in_channels+1
+%endif
+
+cglobal ac3_downmix_%1_to_%2, 3,needed_gpregs,total_mmregs, src0, src1, len,
src2, src3, src4, src5
+
+; get aligned stack space if needed
+%if matrix_elements_stack > 0
+ %if mmsize == 32
+ %assign bkpreg %1 + 1
+ %define bkpq r %+ bkpreg %+ q
+ mov bkpq, rsp
+ and rsp, ~(mmsize-1)
+ sub rsp, matrix_elements_stack * mmsize
+ %else
+ %assign pad matrix_elements_stack * mmsize + (mmsize - gprsize) -
(stack_offset & (mmsize - gprsize))
+ SUB rsp, pad
+ %endif
+%endif
+
+; load matrix pointers
+%define matrix0q r1q
+%define matrix1q r3q
+%if stereo
+ mov matrix1q, [matrix0q+gprsize]
+%endif
+ mov matrix0q, [matrix0q]
+
+; define matrix coeff names
+%assign %%i 0
+%assign %%j needed_mmregs
+%rep in_channels
+ %if %%i >= matrix_elements_mm
+ CAT_XDEFINE mx_stack_0_, %%i, 1
+ CAT_XDEFINE mx_0_, %%i, [rsp+(%%i-matrix_elements_mm)*mmsize]
+ %else
+ CAT_XDEFINE mx_stack_0_, %%i, 0
+ CAT_XDEFINE mx_0_, %%i, m %+ %%j
+ %assign %%j %%j+1
+ %endif
+ %assign %%i %%i+1
+%endrep
+%if stereo
+%assign %%i 0
+%rep in_channels
+ %if in_channels + %%i >= matrix_elements_mm
+ CAT_XDEFINE mx_stack_1_, %%i, 1
+ CAT_XDEFINE mx_1_, %%i,
[rsp+(in_channels+%%i-matrix_elements_mm)*mmsize]
+ %else
+ CAT_XDEFINE mx_stack_1_, %%i, 0
+ CAT_XDEFINE mx_1_, %%i, m %+ %%j
+ %assign %%j %%j+1
+ %endif
+ %assign %%i %%i+1
+%endrep
+%endif
+
+; load/splat matrix coeffs
+%assign %%i 0
+%rep in_channels
+ %if mx_stack_0_ %+ %%i
+ VBROADCASTSS m0, [matrix0q+4*%%i]
+ mova mx_0_ %+ %%i, m0
+ %else
+ VBROADCASTSS mx_0_ %+ %%i, [matrix0q+4*%%i]
+ %endif
+ %if stereo
+ %if mx_stack_1_ %+ %%i
+ VBROADCASTSS m0, [matrix1q+4*%%i]
+ mova mx_1_ %+ %%i, m0
+ %else
+ VBROADCASTSS mx_1_ %+ %%i, [matrix1q+4*%%i]
+ %endif
+ %endif
+ %assign %%i %%i+1
+%endrep
+
+ lea lenq, [4*r2d]
+ ; load channel pointers to registers
+%assign %%i 1
+%rep (in_channels - 1)
+ mov src %+ %%i %+ q, [src0q+%%i*gprsize]
+ add src %+ %%i %+ q, lenq
+ %assign %%i %%i+1
+%endrep
+ mov src0q, [src0q]
+ add src0q, lenq
+ neg lenq
+.loop:
+ %if stereo || mx_stack_0_0
+ mova m0, [src0q+lenq]
+ %endif
+ %if stereo
+ mulps m1, m0, mx_1_0
+ %endif
+ %if stereo || mx_stack_0_0
+ mulps m0, m0, mx_0_0
+ %else
+ mulps m0, [src0q+lenq], mx_0_0
+ %endif
+%assign %%i 1
+%rep (in_channels - 1)
+ %define src_ptr src %+ %%i %+ q
+ ; avoid extra load for mono if matrix is in a mm register
+ %if stereo || mx_stack_0_ %+ %%i
+ mova m2, [src_ptr+lenq]
+ %endif
+ %if stereo
+ fmaddps m1, m2, mx_1_ %+ %%i, m1, m3
+ %endif
+ %if stereo || mx_stack_0_ %+ %%i
+ fmaddps m0, m2, mx_0_ %+ %%i, m0, m2
+ %else
+ fmaddps m0, mx_0_ %+ %%i, [src_ptr+lenq], m0, m1
+ %endif
+ %assign %%i %%i+1
+%endrep
+ mova [src0q+lenq], m0
+ %if stereo
+ mova [src1q+lenq], m1
+ %endif
+
+ add lenq, mmsize
+ jl .loop
+ ; restore stack pointer
+%if matrix_elements_stack > 0
+ %if mmsize == 32
+ mov rsp, bkpq
+ %else
+ ADD rsp, pad
+ %endif
+%endif
+ RET
+%endmacro
+
+%macro AC3_DOWNMIX_FUNCS 0
+%assign %%i 3
+%rep 4
+ INIT_XMM sse
+ AC3_DOWNMIX %%i, 1
+ AC3_DOWNMIX %%i, 2
+ ; Do not use ymm AVX or FMA4 on x86-32 for 6 channels due to stack
alignment
+ ; issues. We are not guaranteed a 32-byte aligned stack, so we have to use
+ ; an extra register to save/restore the stack pointer. For 6 channels on
+ ; x86-32 we do not have an extra register available.
+ %if ARCH_X86_64 || %%i < 6
+ INIT_YMM avx
+ %else
+ INIT_XMM avx
+ %endif
+ AC3_DOWNMIX %%i, 1
+ AC3_DOWNMIX %%i, 2
+ %if HAVE_FMA4_EXTERNAL
+ %if ARCH_X86_64 || %%i < 6
+ INIT_YMM fma4
+ %else
+ INIT_XMM fma4
+ %endif
+ AC3_DOWNMIX %%i, 1
+ AC3_DOWNMIX %%i, 2
+ %endif
+ %assign %%i %%i+1
+%endrep
+%endmacro
+
+AC3_DOWNMIX_FUNCS
diff --git a/libavcodec/x86/ac3dsp_init.c b/libavcodec/x86/ac3dsp_init.c
index e76fa37..92a5f71 100644
--- a/libavcodec/x86/ac3dsp_init.c
+++ b/libavcodec/x86/ac3dsp_init.c
@@ -23,6 +23,7 @@
#include "libavutil/x86/asm.h"
#include "libavutil/x86/cpu.h"
#include "dsputil_mmx.h"
+#include "ac3dsp.h"
#include "libavcodec/ac3.h"
#include "libavcodec/ac3dsp.h"
@@ -51,135 +52,6 @@ extern void ff_ac3_extract_exponents_3dnow(uint8_t *exp,
int32_t *coef, int nb_c
extern void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int
nb_coefs);
extern void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int
nb_coefs);
-#if HAVE_SSE_INLINE && HAVE_7REGS
-
-#define IF1(x) x
-#define IF0(x)
-
-#define MIX5(mono, stereo) \
- __asm__ volatile ( \
- "movss 0(%1), %%xmm5 \n" \
- "movss 4(%1), %%xmm6 \n" \
- "movss 12(%1), %%xmm7 \n" \
- "shufps $0, %%xmm5, %%xmm5 \n" \
- "shufps $0, %%xmm6, %%xmm6 \n" \
- "shufps $0, %%xmm7, %%xmm7 \n" \
- "1: \n" \
- "movaps (%0, %2), %%xmm0 \n" \
- "movaps (%0, %3), %%xmm1 \n" \
- "movaps (%0, %4), %%xmm2 \n" \
- "movaps (%0, %5), %%xmm3 \n" \
- "movaps (%0, %6), %%xmm4 \n" \
- "mulps %%xmm5, %%xmm0 \n" \
- "mulps %%xmm6, %%xmm1 \n" \
- "mulps %%xmm5, %%xmm2 \n" \
- "mulps %%xmm7, %%xmm3 \n" \
- "mulps %%xmm7, %%xmm4 \n" \
- stereo("addps %%xmm1, %%xmm0 \n") \
- "addps %%xmm1, %%xmm2 \n" \
- "addps %%xmm3, %%xmm0 \n" \
- "addps %%xmm4, %%xmm2 \n" \
- mono("addps %%xmm2, %%xmm0 \n") \
- "movaps %%xmm0, (%0, %2) \n" \
- stereo("movaps %%xmm2, (%0, %3) \n") \
- "add $16, %0 \n" \
- "jl 1b \n" \
- : "+&r"(i) \
- : "r"(matrix[0]), \
- "r"(samples[0] + len), \
- "r"(samples[1] + len), \
- "r"(samples[2] + len), \
- "r"(samples[3] + len), \
- "r"(samples[4] + len) \
- : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3", \
- "%xmm4", "%xmm5", "%xmm6", "%xmm7",) \
- "memory" \
- );
-
-#define MIX_MISC(stereo) \
- __asm__ volatile ( \
- "mov %5, %2 \n" \
- "1: \n" \
- "mov -%c7(%6, %2, %c8), %3 \n" \
- "movaps (%3, %0), %%xmm0 \n" \
- stereo("movaps %%xmm0, %%xmm1 \n") \
- "mulps %%xmm4, %%xmm0 \n" \
- stereo("mulps %%xmm5, %%xmm1 \n") \
- "2: \n" \
- "mov (%6, %2, %c8), %1 \n" \
- "movaps (%1, %0), %%xmm2 \n" \
- stereo("movaps %%xmm2, %%xmm3 \n") \
- "mulps (%4, %2, 8), %%xmm2 \n" \
- stereo("mulps 16(%4, %2, 8), %%xmm3 \n") \
- "addps %%xmm2, %%xmm0 \n" \
- stereo("addps %%xmm3, %%xmm1 \n") \
- "add $4, %2 \n" \
- "jl 2b \n" \
- "mov %5, %2 \n" \
- stereo("mov (%6, %2, %c8), %1 \n") \
- "movaps %%xmm0, (%3, %0) \n" \
- stereo("movaps %%xmm1, (%1, %0) \n") \
- "add $16, %0 \n" \
- "jl 1b \n" \
- : "+&r"(i), "=&r"(j), "=&r"(k), "=&r"(m) \
- : "r"(matrix_simd + in_ch), \
- "g"((intptr_t) - 4 * (in_ch - 1)), \
- "r"(samp + in_ch), \
- "i"(sizeof(float *)), "i"(sizeof(float *)/4) \
- : "memory" \
- );
-
-static void ac3_downmix_sse(float **samples, float **matrix,
- int out_ch, int in_ch, int len)
-{
- int **matrix_cmp = (int **)matrix;
- intptr_t i, j, k, m;
-
- i = -len * sizeof(float);
- if (in_ch == 5 && out_ch == 2 &&
- !(matrix_cmp[1][0] | matrix_cmp[0][2] |
- matrix_cmp[1][3] | matrix_cmp[0][4] |
- (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
- (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
- MIX5(IF0, IF1);
- } else if (in_ch == 5 && out_ch == 1 &&
- matrix_cmp[0][0] == matrix_cmp[0][2] &&
- matrix_cmp[0][3] == matrix_cmp[0][4]) {
- MIX5(IF1, IF0);
- } else {
- DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];
- float *samp[AC3_MAX_CHANNELS];
-
- for (j = 0; j < in_ch; j++)
- samp[j] = samples[j] + len;
-
- j = 2 * in_ch * sizeof(float);
- k = in_ch * sizeof(float);
- __asm__ volatile (
- "1: \n"
- "sub $4, %1 \n"
- "sub $8, %0 \n"
- "movss (%3, %1), %%xmm4 \n"
- "movss (%4, %1), %%xmm5 \n"
- "shufps $0, %%xmm4, %%xmm4 \n"
- "shufps $0, %%xmm5, %%xmm5 \n"
- "movaps %%xmm4, (%2, %0, 4) \n"
- "movaps %%xmm5, 16(%2, %0, 4) \n"
- "jg 1b \n"
- : "+&r"(j), "+&r"(k)
- : "r"(matrix_simd), "r"(matrix[0]), "r"(matrix[1])
- : "memory"
- );
- if (out_ch == 2) {
- MIX_MISC(IF1);
- } else {
- MIX_MISC(IF0);
- }
- }
-}
-
-#endif /* HAVE_SSE_INLINE && HAVE_7REGS */
-
av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)
{
int mm_flags = av_get_cpu_flags();
@@ -220,10 +92,57 @@ av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int
bit_exact)
c->extract_exponents = ff_ac3_extract_exponents_ssse3;
}
}
+}
-#if HAVE_SSE_INLINE && HAVE_7REGS
- if (INLINE_SSE(mm_flags)) {
- c->downmix = ac3_downmix_sse;
+#define DEFINE_DOWNMIX(ch) \
+void ff_ac3_downmix_ ## ch ## _to_1_sse(float **samples, float **matrix, \
+ int len); \
+void ff_ac3_downmix_ ## ch ## _to_2_sse(float **samples, float **matrix, \
+ int len); \
+ \
+void ff_ac3_downmix_ ## ch ## _to_1_avx(float **samples, float **matrix, \
+ int len); \
+void ff_ac3_downmix_ ## ch ## _to_2_avx(float **samples, float **matrix, \
+ int len); \
+ \
+void ff_ac3_downmix_ ## ch ## _to_1_fma4(float **samples, float **matrix, \
+ int len); \
+void ff_ac3_downmix_ ## ch ## _to_2_fma4(float **samples, float **matrix, \
+ int len); \
+
+DEFINE_DOWNMIX(3)
+DEFINE_DOWNMIX(4)
+DEFINE_DOWNMIX(5)
+DEFINE_DOWNMIX(6)
+
+void ff_ac3dsp_set_downmix_x86(AC3DSPContext *c)
+{
+ int mm_flags = av_get_cpu_flags();
+
+#define SET_DOWNMIX(ch) \
+ if (ch == c->in_channels) { \
+ if (EXTERNAL_SSE(mm_flags)) { \
+ if (c->out_channels == 1) \
+ c->downmix = ff_ac3_downmix_## ch ##_to_1_sse; \
+ else \
+ c->downmix = ff_ac3_downmix_## ch ##_to_2_sse; \
+ } \
+ if (EXTERNAL_AVX(mm_flags)) { \
+ if (c->out_channels == 1) \
+ c->downmix = ff_ac3_downmix_## ch ##_to_1_avx; \
+ else \
+ c->downmix = ff_ac3_downmix_## ch ##_to_2_avx; \
+ } \
+ if (EXTERNAL_FMA4(mm_flags)) { \
+ if (c->out_channels == 1) \
+ c->downmix = ff_ac3_downmix_## ch ##_to_1_fma4; \
+ else \
+ c->downmix = ff_ac3_downmix_## ch ##_to_2_fma4; \
+ } \
}
-#endif
+
+ SET_DOWNMIX(3)
+ SET_DOWNMIX(4)
+ SET_DOWNMIX(5)
+ SET_DOWNMIX(6)
}
--
1.7.1
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel
