| Message ID | LbSrvEo--3-1@lynne.ee |
|---|---|
| State | New |
| Headers | show |
2019-04-02 13:47 GMT+02:00, Lynne <dev@lynne.ee>: > 153372 UNITS in postfilter_c, 65536 runs, 0 skips > 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup > > 80591 UNITS in deemphasis_c, 131072 runs, 0 skips > 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup > Total decoder speedup: ~15% on a Raspberry Pi 3 (from > 28.1x to 33.5x realtime) How can I reproduce this? On something called "X-Gene" Linux, I see a 2,5% overall speedup (126x to 129x) The issue is that there is some probability that your patch will break mac and Windows compilation, so I would at least like to understand why I can not reproduce your speedup. Carl Eugen
Apr 6, 2019, 2:42 PM by ceffmpeg@gmail.com: > 2019-04-02 13:47 GMT+02:00, Lynne <> dev@lynne.ee <mailto:dev@lynne.ee>> >: > >> 153372 UNITS in postfilter_c, 65536 runs, 0 skips >> 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup >> >> 80591 UNITS in deemphasis_c, 131072 runs, 0 skips >> 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup >> >> Total decoder speedup: ~15% on a Raspberry Pi 3 (from >> 28.1x to 33.5x realtime) >> > > How can I reproduce this? > On something called "X-Gene" Linux, I see a 2,5% overall > speedup (126x to 129x) > > The issue is that there is some probability that your patch will > break mac and Windows compilation, so I would at least > like to understand why I can not reproduce your speedup. > I tested this on real hardware. How can this break mac support at all when it doesn't support aarch64? And I can't find a windows aarch64 compiler either.
2019-04-06 16:08 GMT+02:00, Lynne <dev@lynne.ee>: > > > > Apr 6, 2019, 2:42 PM by ceffmpeg@gmail.com: > >> 2019-04-02 13:47 GMT+02:00, Lynne <> dev@lynne.ee <mailto:dev@lynne.ee>> >> >: >> >>> 153372 UNITS in postfilter_c, 65536 runs, 0 skips >>> 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x >>> speedup >>> >>> 80591 UNITS in deemphasis_c, 131072 runs, 0 skips >>> 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x >>> speedup >>> >>> Total decoder speedup: ~15% on a Raspberry Pi 3 (from >>> 28.1x to 33.5x realtime) >>> >> >> How can I reproduce this? Meaning: Where can I find the file you used? >> On something called "X-Gene" Linux, I see a 2,5% overall >> speedup (126x to 129x) >> >> The issue is that there is some probability that your patch will >> break mac and Windows compilation, so I would at least >> like to understand why I can not reproduce your speedup. >> > > I tested this on real hardware. Is X-Gene not real hardware? > How can this break mac support at all when it doesn't > support aarch64? Have you heard of iPhones? > And I can't find a windows aarch64 compiler either. There is one (I don't know more). Carl Eugen
Apr 6, 2019, 3:41 PM by ceffmpeg@gmail.com: > 2019-04-06 16:08 GMT+02:00, Lynne <> dev@lynne.ee <mailto:dev@lynne.ee>> >: > >> >> >> >> Apr 6, 2019, 2:42 PM by >> ceffmpeg@gmail.com <mailto:ceffmpeg@gmail.com>>> : >> >>> 2019-04-02 13:47 GMT+02:00, Lynne <> >>> dev@lynne.ee <mailto:dev@lynne.ee>>>> <mailto:>>> dev@lynne.ee <mailto:dev@lynne.ee>>>> >> >>> >: >>> >>>> 153372 UNITS in postfilter_c, 65536 runs, 0 skips >>>> 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x >>>> speedup >>>> >>>> 80591 UNITS in deemphasis_c, 131072 runs, 0 skips >>>> 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x >>>> speedup >>>> >>>> Total decoder speedup: ~15% on a Raspberry Pi 3 (from >>>> 28.1x to 33.5x realtime) >>>> >>> >>> How can I reproduce this? >>> > > Meaning: > Where can I find the file you used? > Sorry for the delay. Here's a test file: http://www.mediafire.com/file/awg5lfy4j2dv85k/testfile.opus/file <http://www.mediafire.com/file/awg5lfy4j2dv85k/testfile.opus/file> >>> On something called "X-Gene" Linux, I see a 2,5% overall >>> speedup (126x to 129x) >>> >>> The issue is that there is some probability that your patch will >>> break mac and Windows compilation, so I would at least >>> like to understand why I can not reproduce your speedup. >>> >> >> I tested this on real hardware. >> > > Is X-Gene not real hardware? > Didn't find much about it at all. If you don't get performance improvements I think someone with a more popular board should test it. >> How can this break mac support at all when it doesn't >> support aarch64? >> > > Have you heard of iPhones? > Those use iOS. >> And I can't find a windows aarch64 compiler either. >> > > There is one (I don't know more). > I still don't know how you can suggest this patch will break any operating systems when you haven't tested either.
2019-04-02 13:47 GMT+02:00, Lynne <dev@lynne.ee>: > 153372 UNITS in postfilter_c, 65536 runs, 0 skips > 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup > > 80591 UNITS in deemphasis_c, 131072 runs, 0 skips > 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup > Total decoder speedup: ~15% on a Raspberry Pi 3 (from 28.1x to 33.5x > realtime) Reproduced 7%, tested on compilation on macos and pushed. Carl Eugen
From afcab2dbeac26ae07689c7e58d0acc94a1c1c3e3 Mon Sep 17 00:00:00 2001
From: Lynne <dev@lynne.ee>
Date: Fri, 15 Mar 2019 14:37:31 +0000
Subject: [PATCH v2] aarch64/opusdsp: implement NEON accelerated postfilter and
deemphasis
153372 UNITS in postfilter_c, 65536 runs, 0 skips
73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup
80591 UNITS in deemphasis_c, 131072 runs, 0 skips
43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup
Total decoder speedup: ~15% on a Raspberry Pi 3 (from 28.1x to 33.5x realtime)
Deemphasis SIMD based on the following unrolling:
const float c1 = CELT_EMPH_COEFF, c2 = c1*c1, c3 = c2*c1, c4 = c3*c1;
float state = coeff;
for (int i = 0; i < len; i += 4) {
y[0] = x[0] + c1*state;
y[1] = x[1] + c2*state + c1*x[0];
y[2] = x[2] + c3*state + c1*x[1] + c2*x[0];
y[3] = x[3] + c4*state + c1*x[2] + c2*x[1] + c3*x[0];
state = y[3];
y += 4;
x += 4;
}
Unlike the x86 version, duplication is used instead of pslldq so
the structure and tables are different.
---
libavcodec/aarch64/Makefile | 2 +
libavcodec/aarch64/opusdsp_init.c | 35 +++++++++
libavcodec/aarch64/opusdsp_neon.S | 113 ++++++++++++++++++++++++++++++
libavcodec/opusdsp.c | 3 +
libavcodec/opusdsp.h | 1 +
5 files changed, 154 insertions(+)
create mode 100644 libavcodec/aarch64/opusdsp_init.c
create mode 100644 libavcodec/aarch64/opusdsp_neon.S
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
index 8bc8bc528c..00f93bf59f 100644
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -15,6 +15,7 @@ OBJS-$(CONFIG_VP8DSP) += aarch64/vp8dsp_init_aarch64.o
OBJS-$(CONFIG_AAC_DECODER) += aarch64/aacpsdsp_init_aarch64.o \
aarch64/sbrdsp_init_aarch64.o
OBJS-$(CONFIG_DCA_DECODER) += aarch64/synth_filter_init.o
+OBJS-$(CONFIG_OPUS_DECODER) += aarch64/opusdsp_init.o
OBJS-$(CONFIG_RV40_DECODER) += aarch64/rv40dsp_init_aarch64.o
OBJS-$(CONFIG_VC1DSP) += aarch64/vc1dsp_init_aarch64.o
OBJS-$(CONFIG_VORBIS_DECODER) += aarch64/vorbisdsp_init.o
@@ -49,6 +50,7 @@ NEON-OBJS-$(CONFIG_VP8DSP) += aarch64/vp8dsp_neon.o
# decoders/encoders
NEON-OBJS-$(CONFIG_AAC_DECODER) += aarch64/aacpsdsp_neon.o
NEON-OBJS-$(CONFIG_DCA_DECODER) += aarch64/synth_filter_neon.o
+NEON-OBJS-$(CONFIG_OPUS_DECODER) += aarch64/opusdsp_neon.o
NEON-OBJS-$(CONFIG_VORBIS_DECODER) += aarch64/vorbisdsp_neon.o
NEON-OBJS-$(CONFIG_VP9_DECODER) += aarch64/vp9itxfm_16bpp_neon.o \
aarch64/vp9itxfm_neon.o \
diff --git a/libavcodec/aarch64/opusdsp_init.c b/libavcodec/aarch64/opusdsp_init.c
new file mode 100644
index 0000000000..cc6a1b672d
--- /dev/null
+++ b/libavcodec/aarch64/opusdsp_init.c
@@ -0,0 +1,35 @@
+/*
+ * 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/aarch64/cpu.h"
+#include "libavcodec/opusdsp.h"
+
+void ff_opus_postfilter_neon(float *data, int period, float *gains, int len);
+float ff_opus_deemphasis_neon(float *out, float *in, float coeff, int len);
+
+av_cold void ff_opus_dsp_init_aarch64(OpusDSP *ctx)
+{
+ int cpu_flags = av_get_cpu_flags();
+
+ if (have_neon(cpu_flags)) {
+ ctx->postfilter = ff_opus_postfilter_neon;
+ ctx->deemphasis = ff_opus_deemphasis_neon;
+ }
+}
diff --git a/libavcodec/aarch64/opusdsp_neon.S b/libavcodec/aarch64/opusdsp_neon.S
new file mode 100644
index 0000000000..3d805e2c9c
--- /dev/null
+++ b/libavcodec/aarch64/opusdsp_neon.S
@@ -0,0 +1,113 @@
+/*
+ * 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/aarch64/asm.S"
+
+ // 0.85..^1 0.85..^2 0.85..^3 0.85..^4
+const tab_st, align=4
+ .word 0x3f599a00, 0x3f38f671, 0x3f1d382a, 0x3f05a32f
+endconst
+const tab_x0, align=4
+ .word 0x0, 0x3f599a00, 0x3f38f671, 0x3f1d382a
+endconst
+const tab_x1, align=4
+ .word 0x0, 0x0, 0x3f599a00, 0x3f38f671
+endconst
+const tab_x2, align=4
+ .word 0x0, 0x0, 0x0, 0x3f599a00
+endconst
+
+function ff_opus_deemphasis_neon, export=1
+ movrel x4, tab_st
+ ld1 {v4.4s}, [x4]
+ movrel x4, tab_x0
+ ld1 {v5.4s}, [x4]
+ movrel x4, tab_x1
+ ld1 {v6.4s}, [x4]
+ movrel x4, tab_x2
+ ld1 {v7.4s}, [x4]
+
+ fmul v0.4s, v4.4s, v0.s[0]
+
+1: ld1 {v1.4s, v2.4s}, [x1], #32
+
+ fmla v0.4s, v5.4s, v1.s[0]
+ fmul v3.4s, v7.4s, v2.s[2]
+
+ fmla v0.4s, v6.4s, v1.s[1]
+ fmla v3.4s, v6.4s, v2.s[1]
+
+ fmla v0.4s, v7.4s, v1.s[2]
+ fmla v3.4s, v5.4s, v2.s[0]
+
+ fadd v1.4s, v1.4s, v0.4s
+ fadd v2.4s, v2.4s, v3.4s
+
+ fmla v2.4s, v4.4s, v1.s[3]
+
+ st1 {v1.4s, v2.4s}, [x0], #32
+ fmul v0.4s, v4.4s, v2.s[3]
+
+ subs w2, w2, #8
+ b.gt 1b
+
+ mov s0, v2.s[3]
+
+ ret
+endfunc
+
+function ff_opus_postfilter_neon, export=1
+ ld1 {v0.4s}, [x2]
+ dup v1.4s, v0.s[1]
+ dup v2.4s, v0.s[2]
+ dup v0.4s, v0.s[0]
+
+ add w1, w1, #2
+ sub x1, x0, x1, lsl #2
+
+ ld1 {v3.4s}, [x1]
+ fmul v3.4s, v3.4s, v2.4s
+
+1: add x1, x1, #4
+ ld1 {v4.4s}, [x1]
+ add x1, x1, #4
+ ld1 {v5.4s}, [x1]
+ add x1, x1, #4
+ ld1 {v6.4s}, [x1]
+ add x1, x1, #4
+ ld1 {v7.4s}, [x1]
+
+ fmla v3.4s, v7.4s, v2.4s
+ fadd v6.4s, v6.4s, v4.4s
+
+ ld1 {v8.4s}, [x0]
+ fmla v8.4s, v5.4s, v0.4s
+
+ fmul v6.4s, v6.4s, v1.4s
+ fadd v6.4s, v6.4s, v3.4s
+
+ fadd v8.4s, v8.4s, v6.4s
+ fmul v3.4s, v7.4s, v2.4s
+
+ st1 {v8.4s}, [x0], #16
+
+ subs w3, w3, #4
+ b.gt 1b
+
+ ret
+endfunc
diff --git a/libavcodec/opusdsp.c b/libavcodec/opusdsp.c
index 17e819f977..0e179c98c9 100644
--- a/libavcodec/opusdsp.c
+++ b/libavcodec/opusdsp.c
@@ -61,4 +61,7 @@ av_cold void ff_opus_dsp_init(OpusDSP *ctx)
if (ARCH_X86)
ff_opus_dsp_init_x86(ctx);
+
+ if (ARCH_AARCH64)
+ ff_opus_dsp_init_aarch64(ctx);
}
diff --git a/libavcodec/opusdsp.h b/libavcodec/opusdsp.h
index e3eb457f1b..ee3186dd44 100644
--- a/libavcodec/opusdsp.h
+++ b/libavcodec/opusdsp.h
@@ -31,5 +31,6 @@ typedef struct OpusDSP {
void ff_opus_dsp_init(OpusDSP *ctx);
void ff_opus_dsp_init_x86(OpusDSP *ctx);
+void ff_opus_dsp_init_aarch64(OpusDSP *ctx);
#endif /* AVCODEC_OPUSDSP_H */
--
2.20.1
153372 UNITS in postfilter_c, 65536 runs, 0 skips 73164 UNITS in postfilter_neon, 65536 runs, 0 skips -> 2.1x speedup 80591 UNITS in deemphasis_c, 131072 runs, 0 skips 43969 UNITS in deemphasis_neon, 131072 runs, 0 skips -> 1.83x speedup Total decoder speedup: ~15% on a Raspberry Pi 3 (from 28.1x to 33.5x realtime) Deemphasis SIMD based on the following unrolling: const float c1 = CELT_EMPH_COEFF, c2 = c1*c1, c3 = c2*c1, c4 = c3*c1; float state = coeff; for (int i = 0; i < len; i += 4) { y[0] = x[0] + c1*state; y[1] = x[1] + c2*state + c1*x[0]; y[2] = x[2] + c3*state + c1*x[1] + c2*x[0]; y[3] = x[3] + c4*state + c1*x[2] + c2*x[1] + c3*x[0]; state = y[3]; y += 4; x += 4; } Unlike the x86 version, duplication is used instead of pslldq so the structure and tables are different.