aarch64: Add NEON optimizations for 10 and 12 bit vp9 MC

This work is sponsored by, and copyright, Google.

This has mostly got the same differences to the 8 bit version as
in the arm version. For the horizontal filters, we do 16 pixels
in parallel as well. For the 8 pixel wide vertical filters, we can
accumulate 4 rows before storing, just as in the 8 bit version.

Examples of runtimes vs the 32 bit version, on a Cortex A53:
                                           ARM   AArch64
vp9_avg4_10bpp_neon:                      35.7      30.7
vp9_avg8_10bpp_neon:                      93.5      84.7
vp9_avg16_10bpp_neon:                    324.4     296.6
vp9_avg32_10bpp_neon:                   1236.5    1148.2
vp9_avg64_10bpp_neon:                   4639.6    4571.1
vp9_avg_8tap_smooth_4h_10bpp_neon:       130.0     128.0
vp9_avg_8tap_smooth_4hv_10bpp_neon:      440.0     440.5
vp9_avg_8tap_smooth_4v_10bpp_neon:       114.0     105.5
vp9_avg_8tap_smooth_8h_10bpp_neon:       327.0     314.0
vp9_avg_8tap_smooth_8hv_10bpp_neon:      918.7     865.4
vp9_avg_8tap_smooth_8v_10bpp_neon:       330.0     300.2
vp9_avg_8tap_smooth_16h_10bpp_neon:     1187.5    1155.5
vp9_avg_8tap_smooth_16hv_10bpp_neon:    2663.1    2591.0
vp9_avg_8tap_smooth_16v_10bpp_neon:     1107.4    1078.3
vp9_avg_8tap_smooth_64h_10bpp_neon:    17754.6   17454.7
vp9_avg_8tap_smooth_64hv_10bpp_neon:   33285.2   33001.5
vp9_avg_8tap_smooth_64v_10bpp_neon:    16066.9   16048.6
vp9_put4_10bpp_neon:                      25.5      21.7
vp9_put8_10bpp_neon:                      56.0      52.0
vp9_put16_10bpp_neon/armv8:              183.0     163.1
vp9_put32_10bpp_neon/armv8:              678.6     563.1
vp9_put64_10bpp_neon/armv8:             2679.9    2195.8
vp9_put_8tap_smooth_4h_10bpp_neon:       120.0     118.0
vp9_put_8tap_smooth_4hv_10bpp_neon:      435.2     435.0
vp9_put_8tap_smooth_4v_10bpp_neon:       107.0      98.2
vp9_put_8tap_smooth_8h_10bpp_neon:       303.0     290.0
vp9_put_8tap_smooth_8hv_10bpp_neon:      893.7     828.7
vp9_put_8tap_smooth_8v_10bpp_neon:       305.5     263.5
vp9_put_8tap_smooth_16h_10bpp_neon:     1089.1    1059.2
vp9_put_8tap_smooth_16hv_10bpp_neon:    2578.8    2452.4
vp9_put_8tap_smooth_16v_10bpp_neon:     1009.5     933.5
vp9_put_8tap_smooth_64h_10bpp_neon:    16223.4   15918.6
vp9_put_8tap_smooth_64hv_10bpp_neon:   32153.0   31016.2
vp9_put_8tap_smooth_64v_10bpp_neon:    14516.5   13748.1

These are generally about as fast as the corresponding ARM
routines on the same CPU (at least on the A53), in most cases
marginally faster.

The speedup vs C code is around 4-9x.
Signed-off-by: Martin Storsjö <martin@martin.st>

libavcodec/aarch64/Makefile

@@ -15,7 +15,9 @@ OBJS-$(CONFIG_DCA_DECODER)              += aarch64/synth_filter_init.o
 OBJS-$(CONFIG_RV40_DECODER)             += aarch64/rv40dsp_init_aarch64.o
 OBJS-$(CONFIG_VC1DSP)                   += aarch64/vc1dsp_init_aarch64.o
 OBJS-$(CONFIG_VORBIS_DECODER)           += aarch64/vorbisdsp_init.o
-OBJS-$(CONFIG_VP9_DECODER)              += aarch64/vp9dsp_init_aarch64.o
+OBJS-$(CONFIG_VP9_DECODER)              += aarch64/vp9dsp_init_10bpp_aarch64.o \
+                                           aarch64/vp9dsp_init_12bpp_aarch64.o \
+                                           aarch64/vp9dsp_init_aarch64.o
 
 # ARMv8 optimizations
 
@@ -42,4 +44,5 @@ NEON-OBJS-$(CONFIG_DCA_DECODER)         += aarch64/synth_filter_neon.o
 NEON-OBJS-$(CONFIG_VORBIS_DECODER)      += aarch64/vorbisdsp_neon.o
 NEON-OBJS-$(CONFIG_VP9_DECODER)         += aarch64/vp9itxfm_neon.o             \
                                            aarch64/vp9lpf_neon.o               \
+                                           aarch64/vp9mc_16bpp_neon.o          \
                                            aarch64/vp9mc_neon.o

libavcodec/aarch64/vp9dsp_init.h

+/*
+ * Copyright (c) 2017 Google Inc.
+ *
+ * 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
+ */
+
+#ifndef AVCODEC_AARCH64_VP9DSP_INIT_H
+#define AVCODEC_AARCH64_VP9DSP_INIT_H
+
+#include "libavcodec/vp9dsp.h"
+
+void ff_vp9dsp_init_10bpp_aarch64(VP9DSPContext *dsp);
+void ff_vp9dsp_init_12bpp_aarch64(VP9DSPContext *dsp);
+
+#endif /* AVCODEC_AARCH64_VP9DSP_INIT_H */

libavcodec/aarch64/vp9dsp_init_10bpp_aarch64.c

+/*
+ * Copyright (c) 2017 Google Inc.
+ *
+ * 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
+ */
+
+#define BPP 10
+#define INIT_FUNC ff_vp9dsp_init_10bpp_aarch64
+#include "vp9dsp_init_16bpp_aarch64_template.c"

libavcodec/aarch64/vp9dsp_init_12bpp_aarch64.c

+/*
+ * Copyright (c) 2017 Google Inc.
+ *
+ * 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
+ */
+
+#define BPP 12
+#define INIT_FUNC ff_vp9dsp_init_12bpp_aarch64
+#include "vp9dsp_init_16bpp_aarch64_template.c"

libavcodec/aarch64/vp9dsp_init_16bpp_aarch64_template.c

+/*
+ * Copyright (c) 2017 Google Inc.
+ *
+ * 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 <stdint.h>
+
+#include "libavutil/attributes.h"
+#include "libavutil/aarch64/cpu.h"
+#include "vp9dsp_init.h"
+
+#define declare_fpel(type, sz, suffix)                                          \
+void ff_vp9_##type##sz##suffix##_neon(uint8_t *dst, ptrdiff_t dst_stride,       \
+                                      const uint8_t *src, ptrdiff_t src_stride, \
+                                      int h, int mx, int my)
+
+#define decl_mc_func(op, filter, dir, sz, bpp)                                                   \
+void ff_vp9_##op##_##filter##sz##_##dir##_##bpp##_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, bpp)                                      \
+static void op##_##filter##sz##_hv_##bpp##_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, [((1 + (sz < 64)) * sz + 8) * sz * 2]);         \
+    /* 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_##bpp##_neon(temp, 2 * sz,                          \
+                                             src - 3 * src_stride, src_stride,      \
+                                             h + 8, mx, 0);                         \
+    ff_vp9_##op##_##filter##sz##_v_##bpp##_neon(dst, dst_stride,                    \
+                                                temp + 3 * 2 * sz, 2 * sz,          \
+                                                h, 0, my);                          \
+}
+
+#define decl_filter_funcs(op, dir, sz, bpp)  \
+    decl_mc_func(op, regular, dir, sz, bpp); \
+    decl_mc_func(op, sharp,   dir, sz, bpp); \
+    decl_mc_func(op, smooth,  dir, sz, bpp)
+
+#define decl_mc_funcs(sz, bpp)           \
+    decl_filter_funcs(put, h,  sz, bpp); \
+    decl_filter_funcs(avg, h,  sz, bpp); \
+    decl_filter_funcs(put, v,  sz, bpp); \
+    decl_filter_funcs(avg, v,  sz, bpp); \
+    decl_filter_funcs(put, hv, sz, bpp); \
+    decl_filter_funcs(avg, hv, sz, bpp)
+
+#define ff_vp9_copy32_neon  ff_vp9_copy32_aarch64
+#define ff_vp9_copy64_neon  ff_vp9_copy64_aarch64
+#define ff_vp9_copy128_neon ff_vp9_copy128_aarch64
+
+declare_fpel(copy, 128, );
+declare_fpel(copy, 64,  );
+declare_fpel(copy, 32,  );
+declare_fpel(copy, 16,  );
+declare_fpel(copy, 8,   );
+declare_fpel(avg, 64, _16);
+declare_fpel(avg, 32, _16);
+declare_fpel(avg, 16, _16);
+declare_fpel(avg, 8,  _16);
+declare_fpel(avg, 4,  _16);
+
+decl_mc_funcs(64, BPP);
+decl_mc_funcs(32, BPP);
+decl_mc_funcs(16, BPP);
+decl_mc_funcs(8, BPP);
+decl_mc_funcs(4, BPP);
+
+#define define_8tap_2d_funcs(sz, bpp)        \
+    define_8tap_2d_fn(put, regular, sz, bpp) \
+    define_8tap_2d_fn(put, sharp,   sz, bpp) \
+    define_8tap_2d_fn(put, smooth,  sz, bpp) \
+    define_8tap_2d_fn(avg, regular, sz, bpp) \
+    define_8tap_2d_fn(avg, sharp,   sz, bpp) \
+    define_8tap_2d_fn(avg, smooth,  sz, bpp)
+
+define_8tap_2d_funcs(64, BPP)
+define_8tap_2d_funcs(32, BPP)
+define_8tap_2d_funcs(16, BPP)
+define_8tap_2d_funcs(8,  BPP)
+define_8tap_2d_funcs(4,  BPP)
+
+static av_cold void vp9dsp_mc_init_aarch64(VP9DSPContext *dsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+#define init_fpel(idx1, idx2, sz, type, suffix)      \
+    dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][0][0] = \
+    dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][0][0] = \
+    dsp->mc[idx1][FILTER_8TAP_SHARP  ][idx2][0][0] = \
+    dsp->mc[idx1][FILTER_BILINEAR    ][idx2][0][0] = ff_vp9_##type##sz##suffix
+
+#define init_copy(idx, sz, suffix) \
+    init_fpel(idx, 0, sz, copy, suffix)
+
+#define init_avg(idx, sz, suffix) \
+    init_fpel(idx, 1, sz, avg,  suffix)
+
+#define init_copy_avg(idx, sz1, sz2) \
+    init_copy(idx, sz2, _neon);      \
+    init_avg (idx, sz1, _16_neon)
+
+    if (have_armv8(cpu_flags)) {
+        init_copy(0, 128, _aarch64);
+        init_copy(1, 64,  _aarch64);
+        init_copy(2, 32,  _aarch64);
+    }
+
+    if (have_neon(cpu_flags)) {
+#define init_mc_func(idx1, idx2, op, filter, fname, dir, mx, my, sz, pfx, bpp) \
+    dsp->mc[idx1][filter][idx2][mx][my] = pfx##op##_##fname##sz##_##dir##_##bpp##_neon
+
+#define init_mc_funcs(idx, dir, mx, my, sz, pfx, bpp)                                   \
+    init_mc_func(idx, 0, put, FILTER_8TAP_REGULAR, regular, dir, mx, my, sz, pfx, bpp); \
+    init_mc_func(idx, 0, put, FILTER_8TAP_SHARP,   sharp,   dir, mx, my, sz, pfx, bpp); \
+    init_mc_func(idx, 0, put, FILTER_8TAP_SMOOTH,  smooth,  dir, mx, my, sz, pfx, bpp); \
+    init_mc_func(idx, 1, avg, FILTER_8TAP_REGULAR, regular, dir, mx, my, sz, pfx, bpp); \
+    init_mc_func(idx, 1, avg, FILTER_8TAP_SHARP,   sharp,   dir, mx, my, sz, pfx, bpp); \
+    init_mc_func(idx, 1, avg, FILTER_8TAP_SMOOTH,  smooth,  dir, mx, my, sz, pfx, bpp)
+
+#define init_mc_funcs_dirs(idx, sz, bpp)            \
+    init_mc_funcs(idx, v,  0, 1, sz, ff_vp9_, bpp); \
+    init_mc_funcs(idx, h,  1, 0, sz, ff_vp9_, bpp); \
+    init_mc_funcs(idx, hv, 1, 1, sz,        , bpp)
+
+
+        init_avg(0, 64, _16_neon);
+        init_avg(1, 32, _16_neon);
+        init_avg(2, 16, _16_neon);
+        init_copy_avg(3, 8, 16);
+        init_copy_avg(4, 4, 8);
+
+        init_mc_funcs_dirs(0, 64, BPP);
+        init_mc_funcs_dirs(1, 32, BPP);
+        init_mc_funcs_dirs(2, 16, BPP);
+        init_mc_funcs_dirs(3, 8,  BPP);
+        init_mc_funcs_dirs(4, 4,  BPP);
+    }
+}
+
+av_cold void INIT_FUNC(VP9DSPContext *dsp)
+{
+    vp9dsp_mc_init_aarch64(dsp);
+}

libavcodec/aarch64/vp9dsp_init_aarch64.c

@@ -23,6 +23,7 @@
 #include "libavutil/attributes.h"
 #include "libavutil/aarch64/cpu.h"
 #include "libavcodec/vp9dsp.h"
+#include "vp9dsp_init.h"
 
 #define declare_fpel(type, sz)                                          \
 void ff_vp9_##type##sz##_neon(uint8_t *dst, ptrdiff_t dst_stride,       \
@@ -241,7 +242,13 @@ static av_cold void vp9dsp_loopfilter_init_aarch64(VP9DSPContext *dsp)
 
 av_cold void ff_vp9dsp_init_aarch64(VP9DSPContext *dsp, int bpp)
 {
-    if (bpp != 8)
+    if (bpp == 10) {
+        ff_vp9dsp_init_10bpp_aarch64(dsp);
+        return;
+    } else if (bpp == 12) {
+        ff_vp9dsp_init_12bpp_aarch64(dsp);
+        return;
+    } else if (bpp != 8)
         return;
 
     vp9dsp_mc_init_aarch64(dsp);