vp9/x86: sse2 MC assembly.

Also a slight change to the ssse3 code, which prevents a theoretical
overflow in the sharp filter.

Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

2 files changed, 323 insertions, 106 deletions

diff --git a/libavcodec/x86/vp9dsp_init.c b/libavcodec/x86/vp9dsp_init.c
index 06e77a83f8..37173fbec8 100644
--- a/libavcodec/x86/vp9dsp_init.c
+++ b/libavcodec/x86/vp9dsp_init.c
@@ -49,31 +49,33 @@ fpel_func(avg, 32, avx2);
 fpel_func(avg, 64, avx2);
 #undef fpel_func
 
-#define mc_func(avg, sz, dir, opt) \
+#define mc_func(avg, sz, dir, opt, type, f_sz) \
 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)[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);
+                                                 int h, const type (*filter)[f_sz])
+#define mc_funcs(sz, opt, type, fsz) \
+mc_func(put, sz, h, opt, type, fsz); \
+mc_func(avg, sz, h, opt, type, fsz); \
+mc_func(put, sz, v, opt, type, fsz); \
+mc_func(avg, sz, v, opt, type, fsz)
+
+mc_funcs(4, mmxext, int16_t, 8);
+mc_funcs(8, sse2, int16_t, 8);
+mc_funcs(4, ssse3, int8_t, 32);
+mc_funcs(8, ssse3, int8_t, 32);
 #if ARCH_X86_64
-mc_funcs(16, ssse3);
-mc_funcs(32, avx2);
+mc_funcs(16, ssse3, int8_t, 32);
+mc_funcs(32, avx2, int8_t, 32);
 #endif
 
 #undef mc_funcs
 #undef mc_func
 
-#define mc_rep_func(avg, sz, hsz, dir, opt) \
+#define mc_rep_func(avg, sz, hsz, dir, opt, type, f_sz) \
 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)[32]) \
+                                            int h, const type (*filter)[f_sz]) \
 { \
     ff_vp9_##avg##_8tap_1d_##dir##_##hsz##_##opt(dst,       dst_stride, src, \
                                                  src_stride, h, filter); \
@@ -81,95 +83,103 @@ 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, 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)
+#define mc_rep_funcs(sz, hsz, opt, type, fsz) \
+mc_rep_func(put, sz, hsz, h, opt, type, fsz); \
+mc_rep_func(avg, sz, hsz, h, opt, type, fsz); \
+mc_rep_func(put, sz, hsz, v, opt, type, fsz); \
+mc_rep_func(avg, sz, hsz, v, opt, type, fsz)
 
+mc_rep_funcs(16, 8, sse2, int16_t, 8);
 #if ARCH_X86_32
-mc_rep_funcs(16, 8, ssse3);
+mc_rep_funcs(16, 8, ssse3, int8_t, 32);
 #endif
-mc_rep_funcs(32, 16, ssse3);
-mc_rep_funcs(64, 32, ssse3);
+mc_rep_funcs(32, 16, sse2, int16_t, 8);
+mc_rep_funcs(32, 16, ssse3, int8_t, 32);
+mc_rep_funcs(64, 32, sse2, int16_t, 8);
+mc_rep_funcs(64, 32, ssse3, int8_t, 32);
 #if ARCH_X86_64 && HAVE_AVX2_EXTERNAL
-mc_rep_funcs(64, 32, avx2);
+mc_rep_funcs(64, 32, avx2, int8_t, 32);
 #endif
 
 #undef mc_rep_funcs
 #undef mc_rep_func
 
 extern const int8_t ff_filters_ssse3[3][15][4][32];
+extern const int16_t ff_filters_sse2[3][15][8][8];
 
-#define filter_8tap_2d_fn(op, sz, f, fname, align, opt) \
+#define filter_8tap_2d_fn(op, sz, f, f_opt, 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_##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]); \
+                                      h + 7, ff_filters_##f_opt[f][mx - 1]); \
     ff_vp9_##op##_8tap_1d_v_##sz##_##opt(dst, dst_stride, temp + 3 * 64, 64, \
-                                         h, ff_filters_ssse3[f][my - 1]); \
+                                         h, ff_filters_##f_opt[f][my - 1]); \
 }
 
-#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)
+#define filters_8tap_2d_fn(op, sz, align, opt, f_opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_REGULAR, f_opt, regular, align, opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SHARP,   f_opt, sharp, align, opt) \
+filter_8tap_2d_fn(op, sz, FILTER_8TAP_SMOOTH,  f_opt, smooth, align, opt)
+
+#define filters_8tap_2d_fn2(op, align, opt4, opt8, f_opt) \
+filters_8tap_2d_fn(op, 64, align, opt8, f_opt) \
+filters_8tap_2d_fn(op, 32, align, opt8, f_opt) \
+filters_8tap_2d_fn(op, 16, align, opt8, f_opt) \
+filters_8tap_2d_fn(op, 8, align, opt8, f_opt) \
+filters_8tap_2d_fn(op, 4, align, opt4, f_opt)
+
+filters_8tap_2d_fn2(put, 16, mmxext, sse2, sse2)
+filters_8tap_2d_fn2(avg, 16, mmxext, sse2, sse2)
+filters_8tap_2d_fn2(put, 16, ssse3, ssse3, ssse3)
+filters_8tap_2d_fn2(avg, 16, ssse3, ssse3, 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)
+filters_8tap_2d_fn(put, 64, 32, avx2, ssse3)
+filters_8tap_2d_fn(put, 32, 32, avx2, ssse3)
+filters_8tap_2d_fn(avg, 64, 32, avx2, ssse3)
+filters_8tap_2d_fn(avg, 32, 32, avx2, ssse3)
 #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, opt) \
+#define filter_8tap_1d_fn(op, sz, f, f_opt, 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##_##opt(dst, dst_stride, src, src_stride, \
-                                               h, ff_filters_ssse3[f][dvar - 1]); \
+                                               h, ff_filters_##f_opt[f][dvar - 1]); \
 }
 
-#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)
+#define filters_8tap_1d_fn(op, sz, dir, dvar, opt, f_opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_REGULAR, f_opt, regular, dir, dvar, opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SHARP,   f_opt, sharp,   dir, dvar, opt) \
+filter_8tap_1d_fn(op, sz, FILTER_8TAP_SMOOTH,  f_opt, smooth,  dir, dvar, opt)
+
+#define filters_8tap_1d_fn2(op, sz, opt, f_opt) \
+filters_8tap_1d_fn(op, sz, h, mx, opt, f_opt) \
+filters_8tap_1d_fn(op, sz, v, my, opt, f_opt)
+
+#define filters_8tap_1d_fn3(op, opt4, opt8, f_opt) \
+filters_8tap_1d_fn2(op, 64, opt8, f_opt) \
+filters_8tap_1d_fn2(op, 32, opt8, f_opt) \
+filters_8tap_1d_fn2(op, 16, opt8, f_opt) \
+filters_8tap_1d_fn2(op, 8, opt8, f_opt) \
+filters_8tap_1d_fn2(op, 4, opt4, f_opt)
+
+filters_8tap_1d_fn3(put, mmxext, sse2, sse2)
+filters_8tap_1d_fn3(avg, mmxext, sse2, sse2)
+filters_8tap_1d_fn3(put, ssse3, ssse3, ssse3)
+filters_8tap_1d_fn3(avg, ssse3, ssse3, 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)
+filters_8tap_1d_fn2(put, 64, avx2, ssse3)
+filters_8tap_1d_fn2(put, 32, avx2, ssse3)
+filters_8tap_1d_fn2(avg, 64, avx2, ssse3)
+filters_8tap_1d_fn2(avg, 32, avx2, ssse3)
 #endif
 
 #undef filters_8tap_1d_fn
@@ -297,20 +307,23 @@ 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_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)
+#define init_subpel2(idx1, idx2, sz, type, opt) \
+    init_subpel1(idx1, idx2, 1, 1, sz, hv, type, opt); \
+    init_subpel1(idx1, idx2, 0, 1, sz, v,  type, opt); \
+    init_subpel1(idx1, idx2, 1, 0, sz, h,  type, opt)
+
+#define init_subpel3_32_64(idx, type, opt) \
+    init_subpel2(0, idx, 64, type, opt); \
+    init_subpel2(1, idx, 32, 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)
+#define init_subpel3_8to64(idx, type, opt) \
+    init_subpel3_32_64(idx, type, opt); \
+    init_subpel2(2, idx, 16, type, opt); \
+    init_subpel2(3, idx,  8, 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)
+    init_subpel3_8to64(idx, type, opt); \
+    init_subpel2(4, idx,  4, type, opt)
 
 #define init_lpf(opt) do { \
     if (ARCH_X86_64) { \
@@ -357,6 +370,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
     }
 
     if (EXTERNAL_MMXEXT(cpu_flags)) {
+        init_subpel2(4, 0, 4, put, mmxext);
+        init_subpel2(4, 1, 4, avg, mmxext);
         init_fpel(4, 1,  4, avg, mmxext);
         init_fpel(3, 1,  8, avg, mmxext);
         dsp->itxfm_add[TX_4X4][DCT_DCT] = ff_vp9_idct_idct_4x4_add_mmxext;
@@ -369,6 +384,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
     }
 
     if (EXTERNAL_SSE2(cpu_flags)) {
+        init_subpel3_8to64(0, put, sse2);
+        init_subpel3_8to64(1, avg, sse2);
         init_fpel(2, 1, 16, avg, sse2);
         init_fpel(1, 1, 32, avg, sse2);
         init_fpel(0, 1, 64, avg, sse2);
@@ -444,12 +461,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp)
         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);
+            init_subpel3_32_64(0, put, avx2);
+            init_subpel3_32_64(1, avg, avx2);
 #endif
         }
         dsp->intra_pred[TX_32X32][DC_PRED] = ff_vp9_ipred_dc_32x32_avx2;
diff --git a/libavcodec/x86/vp9mc.asm b/libavcodec/x86/vp9mc.asm
index 59e636da39..53939579fc 100644
--- a/libavcodec/x86/vp9mc.asm
+++ b/libavcodec/x86/vp9mc.asm
@@ -25,15 +25,28 @@
 SECTION_RODATA 32
 
 cextern pw_256
+cextern pw_64
 
-%macro F8_TAPS 8
+%macro F8_SSSE3_TAPS 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][32]
-const filters_ssse3 ; smooth
+
+%macro F8_SSE2_TAPS 8
+times 8 dw %1
+times 8 dw %2
+times 8 dw %3
+times 8 dw %4
+times 8 dw %5
+times 8 dw %6
+times 8 dw %7
+times 8 dw %8
+%endmacro
+
+%macro FILTER 1
+const filters_%1 ; 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
@@ -81,9 +94,102 @@ const filters_ssse3 ; smooth
                     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
+%endmacro
+
+%define F8_TAPS F8_SSSE3_TAPS
+; int8_t ff_filters_ssse3[3][15][4][32]
+FILTER ssse3
+%define F8_TAPS F8_SSE2_TAPS
+; int16_t ff_filters_sse2[3][15][8][8]
+FILTER sse2
 
 SECTION .text
 
+%macro filter_sse2_h_fn 1
+%assign %%px mmsize/2
+cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 15, dst, dstride, src, sstride, h, filtery
+    pxor        m5, m5
+    mova        m6, [pw_64]
+    mova        m7, [filteryq+  0]
+%if ARCH_X86_64 && mmsize > 8
+    mova        m8, [filteryq+ 16]
+    mova        m9, [filteryq+ 32]
+    mova       m10, [filteryq+ 48]
+    mova       m11, [filteryq+ 64]
+    mova       m12, [filteryq+ 80]
+    mova       m13, [filteryq+ 96]
+    mova       m14, [filteryq+112]
+%endif
+.loop:
+    movh        m0, [srcq-3]
+    movh        m1, [srcq-2]
+    movh        m2, [srcq-1]
+    movh        m3, [srcq+0]
+    movh        m4, [srcq+1]
+    punpcklbw   m0, m5
+    punpcklbw   m1, m5
+    punpcklbw   m2, m5
+    punpcklbw   m3, m5
+    punpcklbw   m4, m5
+    pmullw      m0, m7
+%if ARCH_X86_64 && mmsize > 8
+    pmullw      m1, m8
+    pmullw      m2, m9
+    pmullw      m3, m10
+    pmullw      m4, m11
+%else
+    pmullw      m1, [filteryq+ 16]
+    pmullw      m2, [filteryq+ 32]
+    pmullw      m3, [filteryq+ 48]
+    pmullw      m4, [filteryq+ 64]
+%endif
+    paddw       m0, m1
+    paddw       m2, m3
+    paddw       m0, m4
+    movh        m1, [srcq+2]
+    movh        m3, [srcq+3]
+    movh        m4, [srcq+4]
+    add       srcq, sstrideq
+    punpcklbw   m1, m5
+    punpcklbw   m3, m5
+    punpcklbw   m4, m5
+%if ARCH_X86_64 && mmsize > 8
+    pmullw      m1, m12
+    pmullw      m3, m13
+    pmullw      m4, m14
+%else
+    pmullw      m1, [filteryq+ 80]
+    pmullw      m3, [filteryq+ 96]
+    pmullw      m4, [filteryq+112]
+%endif
+    paddw       m0, m1
+    paddw       m3, m4
+    paddw       m0, m6
+    paddw       m2, m3
+    paddsw      m0, m2
+    psraw       m0, 7
+%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 mmxext
+filter_sse2_h_fn put
+filter_sse2_h_fn avg
+
+INIT_XMM sse2
+filter_sse2_h_fn put
+filter_sse2_h_fn avg
+
 %macro filter_h_fn 1
 %assign %%px mmsize/2
 cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, dstride, src, sstride, h, filtery
@@ -118,9 +224,9 @@ cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 11, dst, dstride, src, sstride, h, filt
     pmaddubsw   m4, [filteryq+64]
     pmaddubsw   m1, [filteryq+96]
 %endif
-    paddw       m0, m2
-    paddw       m4, m1
-    paddsw      m0, m4
+    paddw       m0, m4
+    paddw       m2, m1
+    paddsw      m0, m2
     pmulhrsw    m0, m6
 %ifidn %1, avg
     movh        m1, [dstq]
@@ -175,12 +281,12 @@ cglobal vp9_%1_8tap_1d_h_ %+ %%px, 6, 6, 14, dst, dstride, src, sstride, h, filt
     pmaddubsw   m5, m10
     pmaddubsw   m6, m11
     pmaddubsw   m7, m11
-    paddw       m0, m2
-    paddw       m1, m3
-    paddw       m4, m6
-    paddw       m5, m7
-    paddsw      m0, m4
-    paddsw      m1, m5
+    paddw       m0, m4
+    paddw       m1, m5
+    paddw       m2, m6
+    paddw       m3, m7
+    paddsw      m0, m2
+    paddsw      m1, m3
     pmulhrsw    m0, m13
     pmulhrsw    m1, m13
     packuswb    m0, m1
@@ -206,6 +312,104 @@ filter_hx2_fn avg
 
 %endif ; ARCH_X86_64
 
+%macro filter_sse2_v_fn 1
+%assign %%px mmsize/2
+%if ARCH_X86_64
+cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 15, dst, dstride, src, sstride, h, filtery, src4, sstride3
+%else
+cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 15, dst, dstride, src, sstride, filtery, src4, sstride3
+    mov   filteryq, r5mp
+%define hd r4mp
+%endif
+    pxor        m5, m5
+    mova        m6, [pw_64]
+    lea  sstride3q, [sstrideq*3]
+    lea      src4q, [srcq+sstrideq]
+    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       m11, [filteryq+ 64]
+    mova       m12, [filteryq+ 80]
+    mova       m13, [filteryq+ 96]
+    mova       m14, [filteryq+112]
+%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]
+    add       srcq, sstrideq
+    movh        m4, [src4q]
+    punpcklbw   m0, m5
+    punpcklbw   m1, m5
+    punpcklbw   m2, m5
+    punpcklbw   m3, m5
+    punpcklbw   m4, m5
+    pmullw      m0, m7
+%if ARCH_X86_64 && mmsize > 8
+    pmullw      m1, m8
+    pmullw      m2, m9
+    pmullw      m3, m10
+    pmullw      m4, m11
+%else
+    pmullw      m1, [filteryq+ 16]
+    pmullw      m2, [filteryq+ 32]
+    pmullw      m3, [filteryq+ 48]
+    pmullw      m4, [filteryq+ 64]
+%endif
+    paddw       m0, m1
+    paddw       m2, m3
+    paddw       m0, m4
+    movh        m1, [src4q+sstrideq]
+    movh        m3, [src4q+sstrideq*2]
+    movh        m4, [src4q+sstride3q]
+    add      src4q, sstrideq
+    punpcklbw   m1, m5
+    punpcklbw   m3, m5
+    punpcklbw   m4, m5
+%if ARCH_X86_64 && mmsize > 8
+    pmullw      m1, m12
+    pmullw      m3, m13
+    pmullw      m4, m14
+%else
+    pmullw      m1, [filteryq+ 80]
+    pmullw      m3, [filteryq+ 96]
+    pmullw      m4, [filteryq+112]
+%endif
+    paddw       m0, m1
+    paddw       m3, m4
+    paddw       m0, m6
+    paddw       m2, m3
+    paddsw      m0, m2
+    psraw       m0, 7
+%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 mmxext
+filter_sse2_v_fn put
+filter_sse2_v_fn avg
+
+INIT_XMM sse2
+filter_sse2_v_fn put
+filter_sse2_v_fn avg
+
 %macro filter_v_fn 1
 %assign %%px mmsize/2
 %if ARCH_X86_64
@@ -253,9 +457,9 @@ cglobal vp9_%1_8tap_1d_v_ %+ %%px, 4, 7, 11, dst, dstride, src, sstride, filtery
     pmaddubsw   m4, [filteryq+64]
     pmaddubsw   m1, [filteryq+96]
 %endif
-    paddw       m0, m2
-    paddw       m4, m1
-    paddsw      m0, m4
+    paddw       m0, m4
+    paddw       m2, m1
+    paddsw      m0, m2
     pmulhrsw    m0, m6
 %ifidn %1, avg
     movh        m1, [dstq]
@@ -318,12 +522,12 @@ cglobal vp9_%1_8tap_1d_v_ %+ %%px, 6, 8, 14, dst, dstride, src, sstride, h, filt
     pmaddubsw   m5, m10
     pmaddubsw   m6, m11
     pmaddubsw   m7, m11
-    paddw       m0, m2
-    paddw       m1, m3
-    paddw       m4, m6
-    paddw       m5, m7
-    paddsw      m0, m4
-    paddsw      m1, m5
+    paddw       m0, m4
+    paddw       m1, m5
+    paddw       m2, m6
+    paddw       m3, m7
+    paddsw      m0, m2
+    paddsw      m1, m3
     pmulhrsw    m0, m13
     pmulhrsw    m1, m13
     packuswb    m0, m1