x86/vp9: add AVX and AVX2 MC

Roughly 25% faster MC than ssse3 for blocksizes 32 and 64.

Reviewed-by: Ronald S. Bultje <rsbultje@gmail.com>
Signed-off-by: James Almer <jamrial@gmail.com>

3 files changed, 150 insertions, 88 deletions

diff --git a/libavcodec/x86/constants.c b/libavcodec/x86/constants.c
index 5d2c2372da..aa3df00633 100644
--- a/libavcodec/x86/constants.c
+++ b/libavcodec/x86/constants.c
@@ -41,7 +41,8 @@ DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_64)   = { 0x0040004000400040ULL, 0x004
 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_96)   =   0x0060006000600060ULL;
 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_128)  =   0x0080008000800080ULL;
 DECLARE_ALIGNED(8,  const uint64_t, ff_pw_255)  =   0x00ff00ff00ff00ffULL;
-DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_256)  = { 0x0100010001000100ULL, 0x0100010001000100ULL };
+DECLARE_ALIGNED(32, const ymm_reg,  ff_pw_256)  = { 0x0100010001000100ULL, 0x0100010001000100ULL,
+                                                    0x0100010001000100ULL, 0x0100010001000100ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_512)  = { 0x0200020002000200ULL, 0x0200020002000200ULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1019) = { 0x03FB03FB03FB03FBULL, 0x03FB03FB03FB03FBULL };
 DECLARE_ALIGNED(16, const xmm_reg,  ff_pw_1024) = { 0x0400040004000400ULL, 0x0400040004000400ULL };
diff --git a/libavcodec/x86/vp9dsp_init.c b/libavcodec/x86/vp9dsp_init.c
index b04e678118..2790024194 100644
--- a/libavcodec/x86/vp9dsp_init.c
+++ b/libavcodec/x86/vp9dsp_init.c
@@ -43,22 +43,27 @@ fpel_func(avg,  8, mmxext);
 fpel_func(avg, 16, sse2);
 fpel_func(avg, 32, sse2);
 fpel_func(avg, 64, sse2);
+fpel_func(put, 32, avx);
+fpel_func(put, 64, avx);
+fpel_func(avg, 32, avx2);
+fpel_func(avg, 64, avx2);
 #undef fpel_func
 
 #define mc_func(avg, sz, dir, opt) \
 void ff_vp9_##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);
+                                                 int h, const int8_t (*filter)[32])
+#define mc_funcs(sz, opt) \
+mc_func(put, sz, h, opt); \
+mc_func(avg, sz, h, opt); \
+mc_func(put, sz, v, opt); \
+mc_func(avg, sz, v, opt)
+
+mc_funcs(4, ssse3);
+mc_funcs(8, ssse3);
 #if ARCH_X86_64
-mc_funcs(16);
+mc_funcs(16, ssse3);
+mc_funcs(32, avx2);
 #endif
 
 #undef mc_funcs
@@ -68,7 +73,7 @@ mc_funcs(16);
 static av_always_inline void \
 ff_vp9_##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]) \
+                                            int h, const int8_t (*filter)[32]) \
 { \
     ff_vp9_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst,       dst_stride, src, \
                                                  src_stride, h, filter); \
@@ -76,81 +81,96 @@ ff_vp9_##avg##_8tap_1d_##dir##_##sz##_##opt(uint8_t *dst, ptrdiff_t dst_stride,
                                                  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)
+#define mc_rep_funcs(sz, hsz, opt) \
+mc_rep_func(put, sz, hsz, h, opt); \
+mc_rep_func(avg, sz, hsz, h, opt); \
+mc_rep_func(put, sz, hsz, v, opt); \
+mc_rep_func(avg, sz, hsz, v, opt)
 
 #if ARCH_X86_32
-mc_rep_funcs(16, 8);
+mc_rep_funcs(16, 8, ssse3);
+#endif
+mc_rep_funcs(32, 16, ssse3);
+mc_rep_funcs(64, 32, ssse3);
+#if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
+mc_rep_funcs(64, 32, avx2);
 #endif
-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];
+extern const int8_t ff_filters_ssse3[3][15][4][32];
 
-#define filter_8tap_2d_fn(op, sz, f, fname) \
-static void op##_8tap_##fname##_##sz##hv_ssse3(uint8_t *dst, ptrdiff_t dst_stride, \
+#define filter_8tap_2d_fn(op, sz, f, fname, align, opt) \
+static void op##_8tap_##fname##_##sz##hv_##opt(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_vp9_put_8tap_1d_h_##sz##_ssse3(temp, 64, src - 3 * src_stride, src_stride, \
+    LOCAL_ALIGNED_##align(uint8_t, temp, [71 * 64]); \
+    ff_vp9_put_8tap_1d_h_##sz##_##opt(temp, 64, src - 3 * src_stride, src_stride, \
                                       h + 7, ff_filters_ssse3[f][mx - 1]); \
-    ff_vp9_##op##_8tap_1d_v_##sz##_ssse3(dst, dst_stride, temp + 3 * 64, 64, \
+    ff_vp9_##op##_8tap_1d_v_##sz##_##opt(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)
+#define filters_8tap_2d_fn(op, sz, align, opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, regular, align, opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP,   sharp, align, opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH,  smooth, align, opt)
+
+#define filters_8tap_2d_fn2(op, align, opt) \
+filters_8tap_2d_fn(op, 64, align, opt) \
+filters_8tap_2d_fn(op, 32, align, opt) \
+filters_8tap_2d_fn(op, 16, align, opt) \
+filters_8tap_2d_fn(op, 8, align, opt) \
+filters_8tap_2d_fn(op, 4, align, opt)
+
+filters_8tap_2d_fn2(put, 16, ssse3)
+filters_8tap_2d_fn2(avg, 16, ssse3)
+#if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
+filters_8tap_2d_fn(put, 64, 32, avx2)
+filters_8tap_2d_fn(put, 32, 32, avx2)
+filters_8tap_2d_fn(avg, 64, 32, avx2)
+filters_8tap_2d_fn(avg, 32, 32, avx2)
+#endif
 
 #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, \
+#define filter_8tap_1d_fn(op, sz, f, fname, dir, dvar, opt) \
+static void op##_8tap_##fname##_##sz##dir##_##opt(uint8_t *dst, ptrdiff_t dst_stride, \
                                                   const uint8_t *src, ptrdiff_t src_stride, \
                                                   int h, int mx, int my) \
 { \
-    ff_vp9_##op##_8tap_1d_##dir##_##sz##_ssse3(dst, dst_stride, src, src_stride, \
+    ff_vp9_##op##_8tap_1d_##dir##_##sz##_##opt(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)
+#define filters_8tap_1d_fn(op, sz, dir, dvar, opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, regular, dir, dvar, opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP,   sharp,   dir, dvar, opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH,  smooth,  dir, dvar, opt)
+
+#define filters_8tap_1d_fn2(op, sz, opt) \
+filters_8tap_1d_fn(op, sz, h, mx, opt) \
+filters_8tap_1d_fn(op, sz, v, my, opt)
+
+#define filters_8tap_1d_fn3(op, opt) \
+filters_8tap_1d_fn2(op, 64, opt) \
+filters_8tap_1d_fn2(op, 32, opt) \
+filters_8tap_1d_fn2(op, 16, opt) \
+filters_8tap_1d_fn2(op, 8, opt) \
+filters_8tap_1d_fn2(op, 4, opt)
+
+filters_8tap_1d_fn3(put, ssse3)
+filters_8tap_1d_fn3(avg, ssse3)
+#if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
+filters_8tap_1d_fn2(put, 64, avx2)
+filters_8tap_1d_fn2(put, 32, avx2)
+filters_8tap_1d_fn2(avg, 64, avx2)
+filters_8tap_1d_fn2(avg, 32, avx2)
+#endif
 
 #undef filters_8tap_1d_fn
 #undef filters_8tap_1d_fn2
@@ -270,9 +290,12 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
     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) \
+#define init_subpel2_32_64(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(1, idx, idxh, idxv, 32, dir, type, opt)
+
+#define init_subpel2(idx, idxh, idxv, dir, type, opt) \
+    init_subpel2_32_64(idx, idxh, idxv, 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)
@@ -389,6 +412,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
             dsp->itxfm_add[TX_32X32][DCT_ADST] =
             dsp->itxfm_add[TX_32X32][DCT_DCT] = ff_vp9_idct_idct_32x32_add_avx;
         }
+        init_fpel(1, 0, 32, put, avx);
+        init_fpel(0, 0, 64, put, avx);
         init_lpf(avx);
         init_ipred(TX_8X8,    8, avx);
         init_ipred(TX_16X16, 16, avx);
@@ -396,6 +421,18 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
     }
 
     if (EXTERNAL_AVX2(cpu_flags)) {
+        init_fpel(1, 1, 32, avg, avx2);
+        init_fpel(0, 1, 64, avg, avx2);
+        if (ARCH_X86_64) {
+#if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
+            init_subpel2_32_64(0, 1, 1, hv, put, avx2);
+            init_subpel2_32_64(0, 0, 1, v,  put, avx2);
+            init_subpel2_32_64(0, 1, 0, h,  put, avx2);
+            init_subpel2_32_64(1, 1, 1, hv, avg, avx2);
+            init_subpel2_32_64(1, 0, 1, v,  avg, avx2);
+            init_subpel2_32_64(1, 1, 0, h,  avg, avx2);
+#endif
+        }
         dsp->intra_pred[TX_32X32][DC_PRED] = ff_vp9_ipred_dc_32x32_avx2;
         dsp->intra_pred[TX_32X32][LEFT_DC_PRED] = ff_vp9_ipred_dc_left_32x32_avx2;
         dsp->intra_pred[TX_32X32][TOP_DC_PRED] = ff_vp9_ipred_dc_top_32x32_avx2;
diff --git a/libavcodec/x86/vp9mc.asm b/libavcodec/x86/vp9mc.asm
index aa10001be4..59e636da39 100644
--- a/libavcodec/x86/vp9mc.asm
+++ b/libavcodec/x86/vp9mc.asm
@@ -22,17 +22,17 @@
 
 %include "libavutil/x86/x86util.asm"
 
-SECTION_RODATA
+SECTION_RODATA 32
 
 cextern pw_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
+times 16 db %1, %2
+times 16 db %3, %4
+times 16 db %5, %6
+times 16 db %7, %8
 %endmacro
-; int8_t ff_filters_ssse3[3][15][4][16]
+; int8_t ff_filters_ssse3[3][15][4][32]
 const filters_ssse3 ; smooth
                     F8_TAPS -3, -1,  32,  64,  38,   1, -3,  0
                     F8_TAPS -2, -2,  29,  63,  41,   2, -3,  0
@@ -90,9 +90,9 @@ cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, dstride, src, sstride, h, filt
     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]
+    mova        m8, [filteryq+32]
+    mova        m9, [filteryq+64]
+    mova       m10, [filteryq+96]
 %endif
 .loop:
     movh        m0, [srcq-3]
@@ -114,9 +114,9 @@ cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, dstride, src, sstride, h, filt
     pmaddubsw   m4, m9
     pmaddubsw   m1, m10
 %else
-    pmaddubsw   m2, [filteryq+16]
-    pmaddubsw   m4, [filteryq+32]
-    pmaddubsw   m1, [filteryq+48]
+    pmaddubsw   m2, [filteryq+32]
+    pmaddubsw   m4, [filteryq+64]
+    pmaddubsw   m1, [filteryq+96]
 %endif
     paddw       m0, m2
     paddw       m4, m1
@@ -150,9 +150,9 @@ filter_h_fn avg
 cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 14, dst, dstride, src, sstride, h, filtery
     mova       m13, [pw_256]
     mova        m8, [filteryq+ 0]
-    mova        m9, [filteryq+16]
-    mova       m10, [filteryq+32]
-    mova       m11, [filteryq+48]
+    mova        m9, [filteryq+32]
+    mova       m10, [filteryq+64]
+    mova       m11, [filteryq+96]
 .loop:
     movu        m0, [srcq-3]
     movu        m1, [srcq-2]
@@ -198,6 +198,12 @@ INIT_XMM ssse3
 filter_hx2_fn put
 filter_hx2_fn avg
 
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+filter_hx2_fn put
+filter_hx2_fn avg
+%endif
+
 %endif ; ARCH_X86_64
 
 %macro filter_v_fn 1
@@ -215,9 +221,9 @@ cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 11, dst, dstride, src, sstride, filtery
     sub       srcq, sstride3q
     mova        m7, [filteryq+ 0]
 %if ARCH_X86_64 && mmsize > 8
-    mova        m8, [filteryq+16]
-    mova        m9, [filteryq+32]
-    mova       m10, [filteryq+48]
+    mova        m8, [filteryq+32]
+    mova        m9, [filteryq+64]
+    mova       m10, [filteryq+96]
 %endif
 .loop:
     ; FIXME maybe reuse loads from previous rows, or just
@@ -243,9 +249,9 @@ cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 11, dst, dstride, src, sstride, filtery
     pmaddubsw   m4, m9
     pmaddubsw   m1, m10
 %else
-    pmaddubsw   m2, [filteryq+16]
-    pmaddubsw   m4, [filteryq+32]
-    pmaddubsw   m1, [filteryq+48]
+    pmaddubsw   m2, [filteryq+32]
+    pmaddubsw   m4, [filteryq+64]
+    pmaddubsw   m1, [filteryq+96]
 %endif
     paddw       m0, m2
     paddw       m4, m1
@@ -283,9 +289,9 @@ cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 14, dst, dstride, src, sstride, h, filt
     lea      src4q, [srcq+sstrideq]
     sub       srcq, sstride3q
     mova        m8, [filteryq+ 0]
-    mova        m9, [filteryq+16]
-    mova       m10, [filteryq+32]
-    mova       m11, [filteryq+48]
+    mova        m9, [filteryq+32]
+    mova       m10, [filteryq+64]
+    mova       m11, [filteryq+96]
 .loop:
     ; FIXME maybe reuse loads from previous rows, or just
     ; more generally unroll this to prevent multiple loads of
@@ -335,6 +341,12 @@ INIT_XMM ssse3
 filter_vx2_fn put
 filter_vx2_fn avg
 
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+filter_vx2_fn put
+filter_vx2_fn avg
+%endif
+
 %endif ; ARCH_X86_64
 
 %macro fpel_fn 6
@@ -346,7 +358,7 @@ filter_vx2_fn avg
 %define %%dstfn mova
 %endif
 
-%if %2 <= 16
+%if %2 <= mmsize
 cglobal vp9_%1%2, 5, 7, 4, dst, dstride, src, sstride, h, dstride3, sstride3
     lea  sstride3q, [sstrideq*3]
     lea  dstride3q, [dstrideq*3]
@@ -377,6 +389,8 @@ cglobal vp9_%1%2, 5, 5, 4, dst, dstride, src, sstride, h
 
 %define d16 16
 %define s16 16
+%define d32 32
+%define s32 32
 INIT_MMX mmx
 fpel_fn put, 4,  strideq, strideq*2, stride3q, 4
 fpel_fn put, 8,  strideq, strideq*2, stride3q, 4
@@ -391,5 +405,15 @@ INIT_XMM sse2
 fpel_fn avg, 16, strideq, strideq*2, stride3q, 4
 fpel_fn avg, 32, mmsize,  strideq,   strideq+mmsize, 2
 fpel_fn avg, 64, mmsize,  mmsize*2,  mmsize*3, 1
+INIT_YMM avx
+fpel_fn put, 32, strideq, strideq*2, stride3q, 4
+fpel_fn put, 64, mmsize,  strideq,   strideq+mmsize, 2
+%if HAVE_AVX2_EXTERNAL
+INIT_YMM avx2
+fpel_fn avg, 32, strideq, strideq*2, stride3q, 4
+fpel_fn avg, 64, mmsize,  strideq,   strideq+mmsize, 2
+%endif
 %undef s16
 %undef d16
+%undef s32
+%undef d32