[FFmpeg-devel] [PATCH] x86/aacpsdsp: add SSE and SSE3 optimized functions
James Almer
jamrial at gmail.com
Sat Jul 25 18:12:06 CEST 2015
Between 1.5 and 2.5 times faster
Signed-off-by: James Almer <jamrial at gmail.com>
---
libavcodec/aacps.c | 4 +-
libavcodec/aacpsdsp.h | 1 +
libavcodec/aacpsdsp_template.c | 2 +
libavcodec/x86/Makefile | 6 +-
libavcodec/x86/aacpsdsp.asm | 215 +++++++++++++++++++++++++++++++++++++++++
libavcodec/x86/aacpsdsp_init.c | 55 +++++++++++
6 files changed, 279 insertions(+), 4 deletions(-)
create mode 100644 libavcodec/x86/aacpsdsp.asm
create mode 100644 libavcodec/x86/aacpsdsp_init.c
diff --git a/libavcodec/aacps.c b/libavcodec/aacps.c
index 664330d..1165d9b 100644
--- a/libavcodec/aacps.c
+++ b/libavcodec/aacps.c
@@ -936,8 +936,8 @@ static void stereo_processing(PSContext *ps, INTFLOAT (*l)[32][2], INTFLOAT (*r)
H22[0][e+1][b] = h22;
}
for (k = 0; k < NR_BANDS[is34]; k++) {
- INTFLOAT h[2][4];
- INTFLOAT h_step[2][4];
+ LOCAL_ALIGNED_16(INTFLOAT, h, [2], [4]);
+ LOCAL_ALIGNED_16(INTFLOAT, h_step, [2], [4]);
int start = ps->border_position[e];
int stop = ps->border_position[e+1];
INTFLOAT width = Q30(1.f) / (stop - start);
diff --git a/libavcodec/aacpsdsp.h b/libavcodec/aacpsdsp.h
index 9e3c5aa..c194bbe 100644
--- a/libavcodec/aacpsdsp.h
+++ b/libavcodec/aacpsdsp.h
@@ -52,5 +52,6 @@ typedef struct PSDSPContext {
void AAC_RENAME(ff_psdsp_init)(PSDSPContext *s);
void ff_psdsp_init_arm(PSDSPContext *s);
void ff_psdsp_init_mips(PSDSPContext *s);
+void ff_psdsp_init_x86(PSDSPContext *s);
#endif /* LIBAVCODEC_AACPSDSP_H */
diff --git a/libavcodec/aacpsdsp_template.c b/libavcodec/aacpsdsp_template.c
index bfec828..3049ce8 100644
--- a/libavcodec/aacpsdsp_template.c
+++ b/libavcodec/aacpsdsp_template.c
@@ -224,5 +224,7 @@ av_cold void AAC_RENAME(ff_psdsp_init)(PSDSPContext *s)
ff_psdsp_init_arm(s);
if (ARCH_MIPS)
ff_psdsp_init_mips(s);
+ if (ARCH_X86)
+ ff_psdsp_init_x86(s);
#endif /* !USE_FIXED */
}
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index a515ebd..c403770 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -38,7 +38,8 @@ OBJS-$(CONFIG_VP8DSP) += x86/vp8dsp_init.o
OBJS-$(CONFIG_XMM_CLOBBER_TEST) += x86/w64xmmtest.o
# decoders/encoders
-OBJS-$(CONFIG_AAC_DECODER) += x86/sbrdsp_init.o
+OBJS-$(CONFIG_AAC_DECODER) += x86/aacpsdsp_init.o \
+ x86/sbrdsp_init.o
OBJS-$(CONFIG_ADPCM_G722_DECODER) += x86/g722dsp_init.o
OBJS-$(CONFIG_ADPCM_G722_ENCODER) += x86/g722dsp_init.o
OBJS-$(CONFIG_APNG_DECODER) += x86/pngdsp_init.o
@@ -130,7 +131,8 @@ YASM-OBJS-$(CONFIG_VP8DSP) += x86/vp8dsp.o \
x86/vp8dsp_loopfilter.o
# decoders/encoders
-YASM-OBJS-$(CONFIG_AAC_DECODER) += x86/sbrdsp.o
+YASM-OBJS-$(CONFIG_AAC_DECODER) += x86/aacpsdsp.o \
+ x86/sbrdsp.o
YASM-OBJS-$(CONFIG_ADPCM_G722_DECODER) += x86/g722dsp.o
YASM-OBJS-$(CONFIG_ADPCM_G722_ENCODER) += x86/g722dsp.o
YASM-OBJS-$(CONFIG_APNG_DECODER) += x86/pngdsp.o
diff --git a/libavcodec/x86/aacpsdsp.asm b/libavcodec/x86/aacpsdsp.asm
new file mode 100644
index 0000000..d4dfd5b
--- /dev/null
+++ b/libavcodec/x86/aacpsdsp.asm
@@ -0,0 +1,215 @@
+;******************************************************************************
+;* SIMD optimized MPEG-4 Parametric Stereo decoding functions
+;*
+;* Copyright (C) 2015 James Almer
+;*
+;* 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 "libavutil/x86/x86util.asm"
+
+SECTION_RODATA
+
+ps_p1m1p1m1: dd 0, 0x80000000, 0, 0x80000000
+
+SECTION_TEXT
+
+;*************************************************************************
+;void ff_ps_add_squares_<opt>(float *dst, const float (*src)[2], int n);
+;*************************************************************************
+%macro PS_ADD_SQUARES 1
+cglobal ps_add_squares, 3, 3, %1, dst, src, n
+.loop:
+ movaps m0, [srcq]
+ movaps m1, [srcq+mmsize]
+ mulps m0, m0
+ mulps m1, m1
+%if cpuflag(sse3)
+ haddps m0, m1
+%else
+ movaps m3, m0
+ movaps m4, m1
+ shufps m3, m3, q0301
+ shufps m4, m4, q0301
+ addps m0, m3
+ addps m1, m4
+ shufps m0, m1, q2020
+%endif
+ addps m0, [dstq]
+ movaps [dstq], m0
+ add dstq, mmsize
+ add srcq, mmsize*2
+ sub nd, mmsize/4
+ jg .loop
+ REP_RET
+%endmacro
+
+INIT_XMM sse
+PS_ADD_SQUARES 3
+INIT_XMM sse3
+PS_ADD_SQUARES 5
+
+;*******************************************************************
+;void ff_ps_mul_pair_single_sse(float (*dst)[2], float (*src0)[2],
+; float *src1, int n);
+;*******************************************************************
+INIT_XMM sse
+cglobal ps_mul_pair_single, 4, 5, 4, dst, src1, src2, n
+ xor r4q, r4q
+
+.loop:
+ movu m0, [src1q+r4q]
+ movu m1, [src1q+r4q+mmsize]
+ mova m2, [src2q]
+ mova m3, m2
+ unpcklps m2, m2
+ unpckhps m3, m3
+ mulps m0, m2
+ mulps m1, m3
+ mova [dstq+r4q], m0
+ mova [dstq+r4q+mmsize], m1
+ add src2q, mmsize
+ add r4q, mmsize*2
+ sub nd, mmsize/4
+ jg .loop
+ REP_RET
+
+;***********************************************************************
+;void ff_ps_stereo_interpolate_sse3(float (*l)[2], float (*r)[2],
+; float h[2][4], float h_step[2][4],
+; int len);
+;***********************************************************************
+INIT_XMM sse3
+cglobal ps_stereo_interpolate, 5, 5, 6, l, r, h, h_step, n
+ movaps m0, [hq]
+ movaps m1, [h_stepq]
+ cmp nd, 0
+ jle .ret
+ shl nd, 3
+ add lq, nq
+ add rq, nq
+ neg nq
+
+align 16
+.loop:
+ addps m0, m1
+ movddup m2, [lq+nq]
+ movddup m3, [rq+nq]
+ movaps m4, m0
+ movaps m5, m0
+ unpcklps m4, m4
+ unpckhps m5, m5
+ mulps m2, m4
+ mulps m3, m5
+ addps m2, m3
+ movsd [lq+nq], m2
+ movhps [rq+nq], m2
+ add nq, 8
+ jl .loop
+.ret:
+ REP_RET
+
+;*******************************************************************
+;void ff_ps_hybrid_analysis_<opt>(float (*out)[2], float (*in)[2],
+; const float (*filter)[8][2],
+; int stride, int n);
+;*******************************************************************
+%macro PS_HYBRID_ANALYSIS_LOOP 3
+ movu %1, [inq+mmsize*%3]
+ movu m1, [inq+mmsize*(5-%3)+8]
+%if cpuflag(sse3)
+ pshufd %2, %1, q2301
+ pshufd m4, m1, q0123
+ pshufd m1, m1, q1032
+ pshufd m2, [filterq+nq+mmsize*%3], q2301
+ addsubps %2, m4
+ addsubps %1, m1
+%else
+ mova m2, [filterq+nq+mmsize*%3]
+ mova %2, %1
+ mova m4, m1
+ shufps %2, %2, q2301
+ shufps m4, m4, q0123
+ shufps m1, m1, q1032
+ shufps m2, m2, q2301
+ xorps m4, m7
+ xorps m1, m7
+ subps %2, m4
+ subps %1, m1
+%endif
+ mulps %2, m2
+ mulps %1, m2
+%if %3
+ addps m3, %2
+ addps m0, %1
+%endif
+%endmacro
+
+%macro PS_HYBRID_ANALYSIS 0
+cglobal ps_hybrid_analysis, 5, 5, 8, out, in, filter, stride, n
+%if cpuflag(sse3)
+%define MOVH movsd
+%else
+%define MOVH movlps
+%endif
+ shl strided, 3
+ shl nd, 6
+ add filterq, nq
+ neg nq
+ mova m7, [ps_p1m1p1m1]
+
+align 16
+.loop:
+ PS_HYBRID_ANALYSIS_LOOP m0, m3, 0
+ PS_HYBRID_ANALYSIS_LOOP m5, m6, 1
+ PS_HYBRID_ANALYSIS_LOOP m5, m6, 2
+
+%if cpuflag(sse3)
+ pshufd m3, m3, q2301
+ xorps m0, m7
+ hsubps m3, m0
+ pshufd m1, m3, q0020
+ pshufd m3, m3, q0031
+ addps m1, m3
+ movsd m2, [inq+6*8]
+%else
+ mova m1, m3
+ mova m2, m0
+ shufps m1, m1, q2301
+ shufps m2, m2, q2301
+ subps m1, m3
+ addps m2, m0
+ unpcklps m3, m1, m2
+ unpckhps m1, m2
+ addps m1, m3
+ movu m2, [inq+6*8] ; faster than movlps and no risk of overread
+%endif
+ movss m3, [filterq+nq+8*6]
+ SPLATD m3
+ mulps m2, m3
+ addps m1, m2
+ MOVH [outq], m1
+ add outq, strideq
+ add nq, 64
+ jl .loop
+ REP_RET
+%endmacro
+
+INIT_XMM sse
+PS_HYBRID_ANALYSIS
+INIT_XMM sse3
+PS_HYBRID_ANALYSIS
diff --git a/libavcodec/x86/aacpsdsp_init.c b/libavcodec/x86/aacpsdsp_init.c
new file mode 100644
index 0000000..f6d6c03
--- /dev/null
+++ b/libavcodec/x86/aacpsdsp_init.c
@@ -0,0 +1,55 @@
+/*
+ * SIMD optimized MPEG-4 Parametric Stereo decoding functions
+ *
+ * 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 "config.h"
+
+#include "libavutil/x86/cpu.h"
+#include "libavutil/attributes.h"
+#include "libavcodec/aacpsdsp.h"
+
+void ff_ps_add_squares_sse (float *dst, const float (*src)[2], int n);
+void ff_ps_add_squares_sse3 (float *dst, const float (*src)[2], int n);
+void ff_ps_mul_pair_single_sse (float (*dst)[2], float (*src0)[2],
+ float *src1, int n);
+void ff_ps_hybrid_analysis_sse (float (*out)[2], float (*in)[2],
+ const float (*filter)[8][2],
+ int stride, int n);
+void ff_ps_hybrid_analysis_sse3(float (*out)[2], float (*in)[2],
+ const float (*filter)[8][2],
+ int stride, int n);
+void ff_ps_stereo_interpolate_sse3(float (*l)[2], float (*r)[2],
+ float h[2][4], float h_step[2][4],
+ int len);
+
+av_cold void ff_psdsp_init_x86(PSDSPContext *s)
+{
+ int cpu_flags = av_get_cpu_flags();
+
+ if (EXTERNAL_SSE(cpu_flags)) {
+ s->add_squares = ff_ps_add_squares_sse;
+ s->mul_pair_single = ff_ps_mul_pair_single_sse;
+ s->hybrid_analysis = ff_ps_hybrid_analysis_sse;
+ }
+ if (EXTERNAL_SSE3(cpu_flags)) {
+ s->add_squares = ff_ps_add_squares_sse3;
+ s->stereo_interpolate[0] = ff_ps_stereo_interpolate_sse3;
+ s->hybrid_analysis = ff_ps_hybrid_analysis_sse3;
+ }
+}
--
2.4.5
More information about the ffmpeg-devel
mailing list