VP9 MC (ssse3) optimizations.

Decoding time of ped1080p.webm goes from 20.7sec to 11.3sec.

3 files changed, 394 insertions, 0 deletions

diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index c2d46586a5..0a6bab885c 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -34,6 +34,7 @@ OBJS-$(CONFIG_VORBIS_DECODER)          += x86/vorbisdsp_init.o
 OBJS-$(CONFIG_VP3DSP)                  += x86/vp3dsp_init.o
 OBJS-$(CONFIG_VP6_DECODER)             += x86/vp6dsp_init.o
 OBJS-$(CONFIG_VP8_DECODER)             += x86/vp8dsp_init.o
+OBJS-$(CONFIG_VP9_DECODER)             += x86/vp9dsp_init.o
 OBJS-$(CONFIG_WEBP_DECODER)            += x86/vp8dsp_init.o
 OBJS-$(CONFIG_XMM_CLOBBER_TEST)        += x86/w64xmmtest.o
 
@@ -98,4 +99,5 @@ YASM-OBJS-$(CONFIG_VORBIS_DECODER)     += x86/vorbisdsp.o
 YASM-OBJS-$(CONFIG_VP3DSP)             += x86/vp3dsp.o
 YASM-OBJS-$(CONFIG_VP6_DECODER)        += x86/vp6dsp.o
 YASM-OBJS-$(CONFIG_VP8_DECODER)        += x86/vp8dsp.o
+YASM-OBJS-$(CONFIG_VP9_DECODER)        += x86/vp9dsp.o
 YASM-OBJS-$(CONFIG_WEBP_DECODER)       += x86/vp8dsp.o
diff --git a/libavcodec/x86/vp9dsp.asm b/libavcodec/x86/vp9dsp.asm
new file mode 100644
index 0000000000..30db0ca3af
--- /dev/null
+++ b/libavcodec/x86/vp9dsp.asm
@@ -0,0 +1,221 @@
+;******************************************************************************
+;* VP9 SIMD optimizations
+;*
+;* Copyright (c) 2013 Ronald S. Bultje <rsbultje gmail com>
+;*
+;* 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
+
+; FIXME share with vp8dsp.asm
+pw_256:   times 8 dw 256
+
+%macro F8_TAPS 8
+times 8 db %1, %2
+times 8 db %3, %4
+times 8 db %5, %6
+times 8 db %7, %8
+%endmacro
+; int8_t ff_filters_ssse3[3][15][4][16]
+const filters_ssse3 ; smooth
+                    F8_TAPS -3, -1,  32,  64,  38,   1, -3,  0
+                    F8_TAPS -2, -2,  29,  63,  41,   2, -3,  0
+                    F8_TAPS -2, -2,  26,  63,  43,   4, -4,  0
+                    F8_TAPS -2, -3,  24,  62,  46,   5, -4,  0
+                    F8_TAPS -2, -3,  21,  60,  49,   7, -4,  0
+                    F8_TAPS -1, -4,  18,  59,  51,   9, -4,  0
+                    F8_TAPS -1, -4,  16,  57,  53,  12, -4, -1
+                    F8_TAPS -1, -4,  14,  55,  55,  14, -4, -1
+                    F8_TAPS -1, -4,  12,  53,  57,  16, -4, -1
+                    F8_TAPS  0, -4,   9,  51,  59,  18, -4, -1
+                    F8_TAPS  0, -4,   7,  49,  60,  21, -3, -2
+                    F8_TAPS  0, -4,   5,  46,  62,  24, -3, -2
+                    F8_TAPS  0, -4,   4,  43,  63,  26, -2, -2
+                    F8_TAPS  0, -3,   2,  41,  63,  29, -2, -2
+                    F8_TAPS  0, -3,   1,  38,  64,  32, -1, -3
+                    ; regular
+                    F8_TAPS  0,  1,  -5, 126,   8,  -3,  1,  0
+                    F8_TAPS -1,  3, -10, 122,  18,  -6,  2,  0
+                    F8_TAPS -1,  4, -13, 118,  27,  -9,  3, -1
+                    F8_TAPS -1,  4, -16, 112,  37, -11,  4, -1
+                    F8_TAPS -1,  5, -18, 105,  48, -14,  4, -1
+                    F8_TAPS -1,  5, -19,  97,  58, -16,  5, -1
+                    F8_TAPS -1,  6, -19,  88,  68, -18,  5, -1
+                    F8_TAPS -1,  6, -19,  78,  78, -19,  6, -1
+                    F8_TAPS -1,  5, -18,  68,  88, -19,  6, -1
+                    F8_TAPS -1,  5, -16,  58,  97, -19,  5, -1
+                    F8_TAPS -1,  4, -14,  48, 105, -18,  5, -1
+                    F8_TAPS -1,  4, -11,  37, 112, -16,  4, -1
+                    F8_TAPS -1,  3,  -9,  27, 118, -13,  4, -1
+                    F8_TAPS  0,  2,  -6,  18, 122, -10,  3, -1
+                    F8_TAPS  0,  1,  -3,   8, 126,  -5,  1,  0
+                    ; sharp
+                    F8_TAPS -1,  3,  -7, 127,   8,  -3,  1,  0
+                    F8_TAPS -2,  5, -13, 125,  17,  -6,  3, -1
+                    F8_TAPS -3,  7, -17, 121,  27, -10,  5, -2
+                    F8_TAPS -4,  9, -20, 115,  37, -13,  6, -2
+                    F8_TAPS -4, 10, -23, 108,  48, -16,  8, -3
+                    F8_TAPS -4, 10, -24, 100,  59, -19,  9, -3
+                    F8_TAPS -4, 11, -24,  90,  70, -21, 10, -4
+                    F8_TAPS -4, 11, -23,  80,  80, -23, 11, -4
+                    F8_TAPS -4, 10, -21,  70,  90, -24, 11, -4
+                    F8_TAPS -3,  9, -19,  59, 100, -24, 10, -4
+                    F8_TAPS -3,  8, -16,  48, 108, -23, 10, -4
+                    F8_TAPS -2,  6, -13,  37, 115, -20,  9, -4
+                    F8_TAPS -2,  5, -10,  27, 121, -17,  7, -3
+                    F8_TAPS -1,  3,  -6,  17, 125, -13,  5, -2
+                    F8_TAPS  0,  1,  -3,   8, 127,  -7,  3, -1
+
+SECTION .text
+
+%macro filter_h_fn 1
+%assign %%px mmsize/2
+cglobal %1_8tap_1d_h_%%px, 6, 6, 11, dst, dstride, src, sstride, h, filtery
+    mova        m6, [pw_256]
+    mova        m7, [filteryq+ 0]
+%if ARCH_X86_64 && mmsize > 8
+    mova        m8, [filteryq+16]
+    mova        m9, [filteryq+32]
+    mova       m10, [filteryq+48]
+%endif
+.loop:
+    movh        m0, [srcq-3]
+    movh        m1, [srcq-2]
+    movh        m2, [srcq-1]
+    movh        m3, [srcq+0]
+    movh        m4, [srcq+1]
+    movh        m5, [srcq+2]
+    punpcklbw   m0, m1
+    punpcklbw   m2, m3
+    movh        m1, [srcq+3]
+    movh        m3, [srcq+4]
+    add       srcq, sstrideq
+    punpcklbw   m4, m5
+    punpcklbw   m1, m3
+    pmaddubsw   m0, m7
+%if ARCH_X86_64 && mmsize > 8
+    pmaddubsw   m2, m8
+    pmaddubsw   m4, m9
+    pmaddubsw   m1, m10
+%else
+    pmaddubsw   m2, [filteryq+16]
+    pmaddubsw   m4, [filteryq+32]
+    pmaddubsw   m1, [filteryq+48]
+%endif
+    paddw       m0, m2
+    paddw       m4, m1
+    paddsw      m0, m4
+    pmulhrsw    m0, m6
+%ifidn %1, avg
+    movh        m1, [dstq]
+%endif
+    packuswb    m0, m0
+%ifidn %1, avg
+    pavgb       m0, m1
+%endif
+    movh    [dstq], m0
+    add       dstq, dstrideq
+    dec         hd
+    jg .loop
+    RET
+%endmacro
+
+INIT_MMX ssse3
+filter_h_fn put
+filter_h_fn avg
+
+INIT_XMM ssse3
+filter_h_fn put
+filter_h_fn avg
+
+%macro filter_v_fn 1
+%assign %%px mmsize/2
+%if ARCH_X86_64
+cglobal %1_8tap_1d_v_%%px, 6, 8, 11, dst, dstride, src, sstride, h, filtery, src4, sstride3
+%else
+cglobal %1_8tap_1d_v_%%px, 4, 7, 11, dst, dstride, src, sstride, filtery, src4, sstride3
+    mov   filteryq, r5mp
+%define hd r4mp
+%endif
+    sub       srcq, sstrideq
+    lea  sstride3q, [sstrideq*3]
+    sub       srcq, sstrideq
+    mova        m6, [pw_256]
+    sub       srcq, sstrideq
+    mova        m7, [filteryq+ 0]
+    lea      src4q, [srcq+sstrideq*4]
+%if ARCH_X86_64 && mmsize > 8
+    mova        m8, [filteryq+16]
+    mova        m9, [filteryq+32]
+    mova       m10, [filteryq+48]
+%endif
+.loop:
+    ; FIXME maybe reuse loads from previous rows, or just
+    ; more generally unroll this to prevent multiple loads of
+    ; the same data?
+    movh        m0, [srcq]
+    movh        m1, [srcq+sstrideq]
+    movh        m2, [srcq+sstrideq*2]
+    movh        m3, [srcq+sstride3q]
+    movh        m4, [src4q]
+    movh        m5, [src4q+sstrideq]
+    punpcklbw   m0, m1
+    punpcklbw   m2, m3
+    movh        m1, [src4q+sstrideq*2]
+    movh        m3, [src4q+sstride3q]
+    add       srcq, sstrideq
+    add      src4q, sstrideq
+    punpcklbw   m4, m5
+    punpcklbw   m1, m3
+    pmaddubsw   m0, m7
+%if ARCH_X86_64 && mmsize > 8
+    pmaddubsw   m2, m8
+    pmaddubsw   m4, m9
+    pmaddubsw   m1, m10
+%else
+    pmaddubsw   m2, [filteryq+16]
+    pmaddubsw   m4, [filteryq+32]
+    pmaddubsw   m1, [filteryq+48]
+%endif
+    paddw       m0, m2
+    paddw       m4, m1
+    paddsw      m0, m4
+    pmulhrsw    m0, m6
+%ifidn %1, avg
+    movh        m1, [dstq]
+%endif
+    packuswb    m0, m0
+%ifidn %1, avg
+    pavgb       m0, m1
+%endif
+    movh    [dstq], m0
+    add       dstq, dstrideq
+    dec         hd
+    jg .loop
+    RET
+%endmacro
+
+INIT_MMX ssse3
+filter_v_fn put
+filter_v_fn avg
+
+INIT_XMM ssse3
+filter_v_fn put
+filter_v_fn avg
diff --git a/libavcodec/x86/vp9dsp_init.c b/libavcodec/x86/vp9dsp_init.c
new file mode 100644
index 0000000000..d135cf4bae
--- /dev/null
+++ b/libavcodec/x86/vp9dsp_init.c
@@ -0,0 +1,171 @@
+/*
+ * VP9 SIMD optimizations
+ *
+ * Copyright (c) 2013 Ronald S. Bultje <rsbultje gmail com>
+ *
+ * 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/cpu.h"
+#include "libavutil/mem.h"
+#include "libavutil/x86/asm.h"
+#include "libavcodec/vp9dsp.h"
+
+#if HAVE_YASM
+
+#define mc_func(avg, sz, dir, opt) \
+void ff_##avg##_8tap_1d_##dir##_##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
+                                             const uint8_t *src, ptrdiff_t src_stride, \
+                                             int h, const int8_t (*filter)[16])
+#define mc_funcs(sz) \
+mc_func(put, sz, h, ssse3); \
+mc_func(avg, sz, h, ssse3); \
+mc_func(put, sz, v, ssse3); \
+mc_func(avg, sz, v, ssse3)
+
+mc_funcs(4);
+mc_funcs(8);
+
+#undef mc_funcs
+#undef mc_func
+
+#define mc_rep_func(avg, sz, hsz, dir, opt) \
+static av_always_inline void \
+ff_##avg##_8tap_1d_##dir##_##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
+                                        const uint8_t *src, ptrdiff_t src_stride, \
+                                        int h, const int8_t (*filter)[16]) \
+{ \
+    ff_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst,       dst_stride, src, \
+                                             src_stride, h, filter); \
+    ff_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst + hsz, dst_stride, src + hsz, \
+                                             src_stride, h, filter); \
+}
+
+#define mc_rep_funcs(sz, hsz) \
+mc_rep_func(put, sz, hsz, h, ssse3); \
+mc_rep_func(avg, sz, hsz, h, ssse3); \
+mc_rep_func(put, sz, hsz, v, ssse3); \
+mc_rep_func(avg, sz, hsz, v, ssse3)
+
+mc_rep_funcs(16, 8);
+mc_rep_funcs(32, 16);
+mc_rep_funcs(64, 32);
+
+#undef mc_rep_funcs
+#undef mc_rep_func
+
+extern const int8_t ff_filters_ssse3[3][15][4][16];
+
+#define filter_8tap_2d_fn(op, sz, f, fname) \
+static void op##_8tap_##fname##_##sz##hv_ssse3(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, [71 * 64]); \
+    ff_put_8tap_1d_h_##sz##_ssse3(temp, 64, src - 3 * src_stride, src_stride, \
+                                  h + 7, ff_filters_ssse3[f][mx - 1]); \
+    ff_##op##_8tap_1d_v_##sz##_ssse3(dst, dst_stride, temp + 3 * 64, 64, \
+                                     h, ff_filters_ssse3[f][my - 1]); \
+}
+
+#define filters_8tap_2d_fn(op, sz) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, regular) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP,   sharp) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH,  smooth)
+
+#define filters_8tap_2d_fn2(op) \
+filters_8tap_2d_fn(op, 64) \
+filters_8tap_2d_fn(op, 32) \
+filters_8tap_2d_fn(op, 16) \
+filters_8tap_2d_fn(op, 8) \
+filters_8tap_2d_fn(op, 4)
+
+filters_8tap_2d_fn2(put)
+filters_8tap_2d_fn2(avg)
+
+#undef filters_8tap_2d_fn2
+#undef filters_8tap_2d_fn
+#undef filter_8tap_2d_fn
+
+#define filter_8tap_1d_fn(op, sz, f, fname, dir, dvar) \
+static void op##_8tap_##fname##_##sz##dir##_ssse3(uint8_t *dst, ptrdiff_t dst_stride, \
+                                                  const uint8_t *src, ptrdiff_t src_stride, \
+                                                  int h, int mx, int my) \
+{ \
+    ff_##op##_8tap_1d_##dir##_##sz##_ssse3(dst, dst_stride, src, src_stride, \
+                                           h, ff_filters_ssse3[f][dvar - 1]); \
+}
+
+#define filters_8tap_1d_fn(op, sz, dir, dvar) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, regular, dir, dvar) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP,   sharp,   dir, dvar) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH,  smooth,  dir, dvar)
+
+#define filters_8tap_1d_fn2(op, sz) \
+filters_8tap_1d_fn(op, sz, h, mx) \
+filters_8tap_1d_fn(op, sz, v, my)
+
+#define filters_8tap_1d_fn3(op) \
+filters_8tap_1d_fn2(op, 64) \
+filters_8tap_1d_fn2(op, 32) \
+filters_8tap_1d_fn2(op, 16) \
+filters_8tap_1d_fn2(op, 8) \
+filters_8tap_1d_fn2(op, 4)
+
+filters_8tap_1d_fn3(put)
+filters_8tap_1d_fn3(avg)
+
+#undef filters_8tap_1d_fn
+#undef filters_8tap_1d_fn2
+#undef filters_8tap_1d_fn3
+#undef filter_8tap_1d_fn
+
+#endif /* HAVE_YASM */
+
+av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
+{
+#if HAVE_YASM
+    int cpu_flags = av_get_cpu_flags();
+
+#define init_subpel1(idx1, idx2, idxh, idxv, sz, dir, type, opt) \
+    dsp->mc[idx1][FILTER_8TAP_SMOOTH ][idx2][idxh][idxv] = type##_8tap_smooth_##sz##dir##_##opt; \
+    dsp->mc[idx1][FILTER_8TAP_REGULAR][idx2][idxh][idxv] = type##_8tap_regular_##sz##dir##_##opt; \
+    dsp->mc[idx1][FILTER_8TAP_SHARP  ][idx2][idxh][idxv] = type##_8tap_sharp_##sz##dir##_##opt
+
+#define init_subpel2(idx, idxh, idxv, dir, type, opt) \
+    init_subpel1(0, idx, idxh, idxv, 64, dir, type, opt); \
+    init_subpel1(1, idx, idxh, idxv, 32, dir, type, opt); \
+    init_subpel1(2, idx, idxh, idxv, 16, dir, type, opt); \
+    init_subpel1(3, idx, idxh, idxv,  8, dir, type, opt); \
+    init_subpel1(4, idx, idxh, idxv,  4, dir, type, opt)
+
+#define init_subpel3(idx, type, opt) \
+    init_subpel2(idx, 1, 1, hv, type, opt); \
+    init_subpel2(idx, 0, 1, v, type, opt); \
+    init_subpel2(idx, 1, 0, h, type, opt)
+
+    if (cpu_flags & AV_CPU_FLAG_SSSE3) {
+        init_subpel3(0, put, ssse3);
+        init_subpel3(1, avg, ssse3);
+    }
+
+#undef init_subpel1
+#undef init_subpel2
+#undef init_subpel3
+
+#endif /* HAVE_YASM */
+}