[FFmpeg-cvslog] arm64: int32_to_float_fmul neon asm
Janne Grunau
git at videolan.org
Sat Jan 2 11:21:27 CET 2016
ffmpeg | branch: master | Janne Grunau <janne-libav at jannau.net> | Thu Dec 3 11:04:29 2015 +0100| [a0fc780a2093784e8664f88205ee1b215e109cee] | committer: Janne Grunau
arm64: int32_to_float_fmul neon asm
3% faster dts decoding on a cortex-a57.
cortex-a57 cortex-a53
int32_to_float_fmul_array8_c: 1270.9 4475.6
int32_to_float_fmul_array8_neon: 328.6 569.2
int32_to_float_fmul_scalar_c: 928.5 4119.6
int32_to_float_fmul_scalar_neon: 309.1 524.1
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=a0fc780a2093784e8664f88205ee1b215e109cee
---
libavcodec/aarch64/Makefile | 2 +
libavcodec/aarch64/fmtconvert_init.c | 43 +++++++++++++++++++
libavcodec/aarch64/fmtconvert_neon.S | 76 ++++++++++++++++++++++++++++++++++
libavcodec/fmtconvert.c | 11 +++--
libavcodec/fmtconvert.h | 1 +
5 files changed, 130 insertions(+), 3 deletions(-)
diff --git a/libavcodec/aarch64/Makefile b/libavcodec/aarch64/Makefile
index 2175578..022ed84 100644
--- a/libavcodec/aarch64/Makefile
+++ b/libavcodec/aarch64/Makefile
@@ -1,5 +1,6 @@
OBJS-$(CONFIG_DCA_DECODER) += aarch64/dcadsp_init.o
OBJS-$(CONFIG_FFT) += aarch64/fft_init_aarch64.o
+OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_init.o
OBJS-$(CONFIG_H264CHROMA) += aarch64/h264chroma_init_aarch64.o
OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_init_aarch64.o
OBJS-$(CONFIG_H264PRED) += aarch64/h264pred_init.o
@@ -19,6 +20,7 @@ ARMV8-OBJS-$(CONFIG_VIDEODSP) += aarch64/videodsp.o
NEON-OBJS-$(CONFIG_DCA_DECODER) += aarch64/dcadsp_neon.o \
aarch64/synth_filter_neon.o
NEON-OBJS-$(CONFIG_FFT) += aarch64/fft_neon.o
+NEON-OBJS-$(CONFIG_FMTCONVERT) += aarch64/fmtconvert_neon.o
NEON-OBJS-$(CONFIG_H264CHROMA) += aarch64/h264cmc_neon.o
NEON-OBJS-$(CONFIG_H264DSP) += aarch64/h264dsp_neon.o \
aarch64/h264idct_neon.o
diff --git a/libavcodec/aarch64/fmtconvert_init.c b/libavcodec/aarch64/fmtconvert_init.c
new file mode 100644
index 0000000..0a55a1b
--- /dev/null
+++ b/libavcodec/aarch64/fmtconvert_init.c
@@ -0,0 +1,43 @@
+/*
+ * ARM optimized Format Conversion 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
+ */
+
+#include <stdint.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/aarch64/cpu.h"
+#include "libavcodec/avcodec.h"
+#include "libavcodec/fmtconvert.h"
+
+void ff_int32_to_float_fmul_array8_neon(FmtConvertContext *c, float *dst,
+ const int32_t *src, const float *mul,
+ int len);
+void ff_int32_to_float_fmul_scalar_neon(float *dst, const int32_t *src,
+ float mul, int len);
+
+av_cold void ff_fmt_convert_init_aarch64(FmtConvertContext *c,
+ AVCodecContext *avctx)
+{
+ int cpu_flags = av_get_cpu_flags();
+
+ if (have_neon(cpu_flags)) {
+ c->int32_to_float_fmul_array8 = ff_int32_to_float_fmul_array8_neon;
+ c->int32_to_float_fmul_scalar = ff_int32_to_float_fmul_scalar_neon;
+ }
+}
diff --git a/libavcodec/aarch64/fmtconvert_neon.S b/libavcodec/aarch64/fmtconvert_neon.S
new file mode 100644
index 0000000..3b33c87
--- /dev/null
+++ b/libavcodec/aarch64/fmtconvert_neon.S
@@ -0,0 +1,76 @@
+/*
+ * ARM NEON optimised Format Conversion Utils
+ * Copyright (c) 2008 Mans Rullgard <mans at mansr.com>
+ * Copyright (c) 2015 Janne Grunau <janne-libav at jannau.net>
+ *
+ * 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
+ */
+
+#include "config.h"
+#include "libavutil/aarch64/asm.S"
+
+function ff_int32_to_float_fmul_scalar_neon, export=1
+ ld1 {v1.4s,v2.4s}, [x1], #32
+ scvtf v1.4s, v1.4s
+ scvtf v2.4s, v2.4s
+1:
+ subs w2, w2, #8
+ fmul v3.4s, v1.4s, v0.s[0]
+ fmul v4.4s, v2.4s, v0.s[0]
+ b.le 2f
+ ld1 {v1.4s,v2.4s}, [x1], #32
+ st1 {v3.4s,v4.4s}, [x0], #32
+ scvtf v1.4s, v1.4s
+ scvtf v2.4s, v2.4s
+ b 1b
+2:
+ st1 {v3.4s,v4.4s}, [x0]
+ ret
+endfunc
+
+function ff_int32_to_float_fmul_array8_neon, export=1
+ lsr w4, w4, #3
+ subs w5, w4, #1
+ b.eq 1f
+2:
+ ld1 {v0.4s,v1.4s}, [x2], #32
+ ld1 {v2.4s,v3.4s}, [x2], #32
+ scvtf v0.4s, v0.4s
+ scvtf v1.4s, v1.4s
+ ld1 {v16.2s}, [x3], #8
+ scvtf v2.4s, v2.4s
+ scvtf v3.4s, v3.4s
+ fmul v4.4s, v0.4s, v16.s[0]
+ fmul v5.4s, v1.4s, v16.s[0]
+ fmul v6.4s, v2.4s, v16.s[1]
+ fmul v7.4s, v3.4s, v16.s[1]
+ st1 {v4.4s,v5.4s}, [x1], #32
+ st1 {v6.4s,v7.4s}, [x1], #32
+ subs w5, w5, #2
+ b.gt 2b
+ b.eq 1f
+ ret
+1:
+ ld1 {v0.4s,v1.4s}, [x2]
+ ld1 {v16.s}[0], [x3]
+ scvtf v0.4s, v0.4s
+ scvtf v1.4s, v1.4s
+ fmul v4.4s, v0.4s, v16.s[0]
+ fmul v5.4s, v1.4s, v16.s[0]
+ st1 {v4.4s,v5.4s}, [x1]
+ ret
+endfunc
diff --git a/libavcodec/fmtconvert.c b/libavcodec/fmtconvert.c
index 2dff704..5e29bfa 100644
--- a/libavcodec/fmtconvert.c
+++ b/libavcodec/fmtconvert.c
@@ -46,7 +46,12 @@ av_cold void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx)
c->int32_to_float_fmul_scalar = int32_to_float_fmul_scalar_c;
c->int32_to_float_fmul_array8 = int32_to_float_fmul_array8_c;
- if (ARCH_ARM) ff_fmt_convert_init_arm(c, avctx);
- if (ARCH_PPC) ff_fmt_convert_init_ppc(c, avctx);
- if (ARCH_X86) ff_fmt_convert_init_x86(c, avctx);
+ if (ARCH_AARCH64)
+ ff_fmt_convert_init_aarch64(c, avctx);
+ if (ARCH_ARM)
+ ff_fmt_convert_init_arm(c, avctx);
+ if (ARCH_PPC)
+ ff_fmt_convert_init_ppc(c, avctx);
+ if (ARCH_X86)
+ ff_fmt_convert_init_x86(c, avctx);
}
diff --git a/libavcodec/fmtconvert.h b/libavcodec/fmtconvert.h
index 7de890b..dd603f9 100644
--- a/libavcodec/fmtconvert.h
+++ b/libavcodec/fmtconvert.h
@@ -58,6 +58,7 @@ typedef struct FmtConvertContext {
void ff_fmt_convert_init(FmtConvertContext *c, AVCodecContext *avctx);
+void ff_fmt_convert_init_aarch64(FmtConvertContext *c, AVCodecContext *avctx);
void ff_fmt_convert_init_arm(FmtConvertContext *c, AVCodecContext *avctx);
void ff_fmt_convert_init_ppc(FmtConvertContext *c, AVCodecContext *avctx);
void ff_fmt_convert_init_x86(FmtConvertContext *c, AVCodecContext *avctx);
More information about the ffmpeg-cvslog
mailing list