On Tue, 11 Oct 2016, Martin Storsjö wrote:
This work is sponsored by, and copyright, Google.
The filter coefficients are signed values, where the product of the
multiplication with one individual filter coefficient doesn't
overflow a 16 bit signed value (the largest filter coefficient is
127). But when the products are accumulated, the resulting sum can
overflow the 16 bit signed range. Instead of accumulating in 32 bit,
we accumulate all filter taps but the largest one in one register, and
the largest one (either index 3 or 4) in a separate one, added with
saturation afterwards.
(The VP8 MC asm does something similar, but slightly simpler, by
accumulating each half of the filter separately. In the VP9 MC
filters, each half of the filter can also overflow though, so the
largest component has to be handled individually.)
Examples of relative speedup compared with the C version, from checkasm:
Cortex A7 A8 A9 A53
vp9_avg4_neon: 1.63 1.19 1.37 1.54
vp9_avg8_neon: 2.24 3.65 3.30 2.56
vp9_avg16_neon: 2.64 6.72 2.92 2.80
vp9_avg32_neon: 2.50 5.45 2.76 2.45
vp9_avg64_neon: 2.69 5.81 2.72 2.79
vp9_avg_8tap_smooth_4h_neon: 3.29 4.71 2.90 4.78
vp9_avg_8tap_smooth_4hv_neon: 3.78 4.54 3.33 4.51
vp9_avg_8tap_smooth_4v_neon: 5.18 6.52 4.29 5.50
vp9_avg_8tap_smooth_8h_neon: 6.47 9.07 5.47 9.66
vp9_avg_8tap_smooth_8hv_neon: 6.60 8.16 5.98 7.91
vp9_avg_8tap_smooth_8v_neon: 9.25 12.67 8.07 9.61
vp9_avg_8tap_smooth_64h_neon: 6.98 10.36 5.91 11.59
vp9_avg_8tap_smooth_64hv_neon: 6.39 9.15 6.05 8.48
vp9_avg_8tap_smooth_64v_neon: 10.64 14.12 9.39 11.05
vp9_put4_neon: 1.30 1.15 0.89 1.35
vp9_put8_neon: 1.28 2.07 1.80 1.62
vp9_put16_neon: 1.64 4.08 1.71 1.93
vp9_put32_neon: 1.53 3.68 2.12 1.68
vp9_put64_neon: 2.01 3.98 1.91 1.91
vp9_put_8tap_smooth_4h_neon: 3.05 4.47 2.68 4.53
vp9_put_8tap_smooth_4hv_neon: 3.74 4.46 3.31 4.50
vp9_put_8tap_smooth_4v_neon: 5.23 6.19 4.28 5.87
vp9_put_8tap_smooth_8h_neon: 5.89 8.40 4.97 8.99
vp9_put_8tap_smooth_8hv_neon: 6.84 8.05 5.90 7.91
vp9_put_8tap_smooth_8v_neon: 9.41 11.97 7.99 10.49
vp9_put_8tap_smooth_64h_neon: 6.61 9.79 5.06 11.35
vp9_put_8tap_smooth_64hv_neon: 7.01 9.13 6.37 9.17
vp9_put_8tap_smooth_64v_neon: 11.09 13.31 9.32 12.52
For the larger 8tap filters, the speedup vs C code is around 6-14x.
This is significantly faster than libvpx's implementation of the same
functions, at least when comparing the put_8tap_smooth_64 functions
(compared to vpx_convolve8_horiz_neon and vpx_convolve8_vert_neon from
libvpx).
Absolute runtimes from checkasm:
Cortex A7 A8 A9 A53
vp9_put_8tap_smooth_64h_neon: 21229.8 14474.5 19790.1 10885.1
libvpx vpx_convolve8_horiz_neon: 52623.3 19736.4 21907.7 25027.7
vp9_put_8tap_smooth_64v_neon: 14966.6 12297.2 13786.5 11679.4
libvpx vpx_convolve8_vert_neon: 42090.0 17706.2 17659.9 16941.2
Thus, on the A9, the horizontal filter is only marginally faster than
libvpx, while our version is significantly faster on the other cores,
and the vertical filter is significantly faster on all cores. The
difference is especially large on the A7.
The libvpx implementation does the accumulation in 32 bit, which
probably explains most the differences.
---
Since the previous version, I tuned the avg4 and put4 a bit further,
making avg4 faster than C on the A8, and improving put4 a little on A7
and A53. put4 is still marginally slower than the C version on A9 though,
but I think it isn't worth the trouble to try to workaround it.
Rewrapped some paragraphs in the commit message, that were unnecessarily
narrow. Changed the big benchmark table into one with relative speedups.
---
libavcodec/arm/Makefile | 2 +
libavcodec/arm/vp9dsp_init_arm.c | 140 +++++++
libavcodec/arm/vp9mc_neon.S | 787 +++++++++++++++++++++++++++++++++++++++
libavcodec/vp9.h | 1 +
libavcodec/vp9dsp.c | 2 +
5 files changed, 932 insertions(+)
create mode 100644 libavcodec/arm/vp9dsp_init_arm.c
create mode 100644 libavcodec/arm/vp9mc_neon.S
diff --git a/libavcodec/arm/Makefile b/libavcodec/arm/Makefile
index bd4dd4e..2638230 100644
--- a/libavcodec/arm/Makefile
+++ b/libavcodec/arm/Makefile
@@ -45,6 +45,7 @@ OBJS-$(CONFIG_MLP_DECODER) +=
arm/mlpdsp_init_arm.o
OBJS-$(CONFIG_RV40_DECODER) += arm/rv40dsp_init_arm.o
OBJS-$(CONFIG_VORBIS_DECODER) += arm/vorbisdsp_init_arm.o
OBJS-$(CONFIG_VP6_DECODER) += arm/vp6dsp_init_arm.o
+OBJS-$(CONFIG_VP9_DECODER) += arm/vp9dsp_init_arm.o
# ARMv5 optimizations
@@ -138,3 +139,4 @@ NEON-OBJS-$(CONFIG_RV40_DECODER) +=
arm/rv34dsp_neon.o \
arm/rv40dsp_neon.o
NEON-OBJS-$(CONFIG_VORBIS_DECODER) += arm/vorbisdsp_neon.o
NEON-OBJS-$(CONFIG_VP6_DECODER) += arm/vp6dsp_neon.o
+NEON-OBJS-$(CONFIG_VP9_DECODER) += arm/vp9mc_neon.o
diff --git a/libavcodec/arm/vp9dsp_init_arm.c b/libavcodec/arm/vp9dsp_init_arm.c
new file mode 100644
index 0000000..db8c683
--- /dev/null
+++ b/libavcodec/arm/vp9dsp_init_arm.c
@@ -0,0 +1,140 @@
+/*
+ * Copyright (c) 2016 Google Inc.
+ *
+ * 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/arm/cpu.h"
+#include "libavcodec/vp9.h"
+
+#define declare_fpel(type, sz) \
+void ff_vp9_##type##sz##_neon(uint8_t *dst, ptrdiff_t dst_stride, \
+ const uint8_t *src, ptrdiff_t src_stride, \
+ int h, int mx, int my)
+
+#define declare_copy_avg(sz) \
+ declare_fpel(copy, sz); \
+ declare_fpel(avg , sz)
+
+#define decl_mc_func(op, filter, dir, sz)
\
+void ff_vp9_##op##_##filter##sz##_##dir##_neon(uint8_t *dst, ptrdiff_t
dst_stride, \
+ const uint8_t *src, ptrdiff_t
src_stride, \
+ int h, int mx, int my)
+
+#define define_8tap_2d_fn(op, filter, sz)
\
+static void op##_##filter##sz##_hv_neon(uint8_t *dst, ptrdiff_t dst_stride,
\
+ const uint8_t *src, ptrdiff_t
src_stride, \
+ int h, int mx, int my)
\
+{
\
+ LOCAL_ALIGNED_16(uint8_t, temp, [72 * 64]);
\
+ /* We only need h + 7 lines, but the horizontal filter assumes an
\
+ * even number of rows, so filter h + 8 lines here. */
\
+ ff_vp9_put_##filter##sz##_h_neon(temp, 64,
\
+ src - 3 * src_stride, src_stride,
\
+ h + 8, mx, 0);
\
+ ff_vp9_##op##_##filter##sz##_v_neon(dst, dst_stride,
\
+ temp + 3 * 64, 64,
\
+ h, 0, my);
\
+}
Since this reads one more line of input than necessary (h + 8 instead of h
+ 7), I had to squash in this change locally:
diff --git a/libavcodec/vp9.h b/libavcodec/vp9.h
index 84bed6d..ba622b1 100644
--- a/libavcodec/vp9.h
+++ b/libavcodec/vp9.h
@@ -420,7 +420,7 @@ typedef struct VP9Context {
// whole-frame cache
uint8_t *intra_pred_data[3];
VP9Filter *lflvl;
- DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[71 * 80];
+ DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[72 * 80];
// block reconstruction intermediates
int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];
Nothing else is needed, since that part of the buffer doesn't need to be
initialized; it just needs to be allocated and not out of bounds.
(In the hv filter, the last uninitialized line is filtered horizontally
when the filter does 2 lines at a time, but the vertical half doesn't read
it.)
// Martin
_______________________________________________
libav-devel mailing list
libav-devel@libav.org
https://lists.libav.org/mailman/listinfo/libav-devel
