shithub: libvpx

Download patch

ref: a4f3751be5f012d66011ddc1c5f12bd12734a1d3
parent: c1e4f5a80d6d56f7227762bc76372b60fbedf46b
author: Scott LaVarnway <[email protected]>
date: Tue May 24 00:44:05 EDT 2016

Code clean of sub_pixel_variance4xh -- 2

Replace MMX with SSE2.

Change-Id: Id8482d2589131f9427e7f36bc64413f058caf31f

--- a/test/variance_test.cc
+++ b/test/variance_test.cc
@@ -1026,8 +1026,8 @@
                       make_tuple(3, 4, &vpx_sub_pixel_variance8x16_sse2, 0),
                       make_tuple(3, 3, &vpx_sub_pixel_variance8x8_sse2, 0),
                       make_tuple(3, 2, &vpx_sub_pixel_variance8x4_sse2, 0),
-                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_sse, 0),
-                      make_tuple(2, 2, &vpx_sub_pixel_variance4x4_sse, 0)));
+                      make_tuple(2, 3, &vpx_sub_pixel_variance4x8_sse2, 0),
+                      make_tuple(2, 2, &vpx_sub_pixel_variance4x4_sse2, 0)));
 
 INSTANTIATE_TEST_CASE_P(
     SSE2, VpxSubpelAvgVarianceTest,
@@ -1043,8 +1043,8 @@
         make_tuple(3, 4, &vpx_sub_pixel_avg_variance8x16_sse2, 0),
         make_tuple(3, 3, &vpx_sub_pixel_avg_variance8x8_sse2, 0),
         make_tuple(3, 2, &vpx_sub_pixel_avg_variance8x4_sse2, 0),
-        make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_sse, 0),
-        make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_sse, 0)));
+        make_tuple(2, 3, &vpx_sub_pixel_avg_variance4x8_sse2, 0),
+        make_tuple(2, 2, &vpx_sub_pixel_avg_variance4x4_sse2, 0)));
 #endif  // CONFIG_USE_X86INC
 
 #if CONFIG_VP9_HIGHBITDEPTH
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -1493,10 +1493,10 @@
   specialize qw/vpx_sub_pixel_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/uint32_t vpx_sub_pixel_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
-  specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
+  specialize qw/vpx_sub_pixel_variance4x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/uint32_t vpx_sub_pixel_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse";
-  specialize qw/vpx_sub_pixel_variance4x4 mmx msa/, "$sse_x86inc", "$ssse3_x86inc";
+  specialize qw/vpx_sub_pixel_variance4x4 mmx msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/uint32_t vpx_sub_pixel_avg_variance64x64/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
   specialize qw/vpx_sub_pixel_avg_variance64x64 avx2 msa/, "$sse2_x86inc", "$ssse3_x86inc";
@@ -1532,10 +1532,10 @@
   specialize qw/vpx_sub_pixel_avg_variance8x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x8/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
-  specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse_x86inc", "$ssse3_x86inc";
+  specialize qw/vpx_sub_pixel_avg_variance4x8 msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 add_proto qw/uint32_t vpx_sub_pixel_avg_variance4x4/, "const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred";
-  specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse_x86inc", "$ssse3_x86inc";
+  specialize qw/vpx_sub_pixel_avg_variance4x4 msa/, "$sse2_x86inc", "$ssse3_x86inc";
 
 #
 # Specialty Subpixel
--- a/vpx_dsp/x86/subpel_variance_sse2.asm
+++ b/vpx_dsp/x86/subpel_variance_sse2.asm
@@ -57,8 +57,8 @@
   paddd                %6, %1
 %endmacro
 
-%macro STORE_AND_RET 0
-%if mmsize == 16
+%macro STORE_AND_RET 1
+%if %1 > 4
   ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
   ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
   ; We have to sign-extend it before adding the words within the register
@@ -78,9 +78,9 @@
   movd               [r1], m7           ; store sse
   paddd                m6, m4
   movd               raxd, m6           ; store sum as return value
-%else ; mmsize == 8
-  pshufw               m4, m6, 0xe
-  pshufw               m3, m7, 0xe
+%else ; 4xh
+  pshuflw              m4, m6, 0xe
+  pshuflw              m3, m7, 0xe
   paddw                m6, m4
   paddd                m7, m3
   pcmpgtw              m5, m6           ; mask for 0 > x
@@ -87,7 +87,7 @@
   mov                  r1, ssem         ; r1 = unsigned int *sse
   punpcklwd            m6, m5           ; sign-extend m6 word->dword
   movd               [r1], m7           ; store sse
-  pshufw               m4, m6, 0xe
+  pshuflw              m4, m6, 0xe
   paddd                m6, m4
   movd               raxd, m6           ; store sum as return value
 %endif
@@ -196,6 +196,12 @@
   %endif
 %endif
 
+%if %1 == 4
+  %define movx movd
+%else
+  %define movx movh
+%endif
+
   ASSERT               %1 <= 16         ; m6 overflows if w > 16
   pxor                 m6, m6           ; sum
   pxor                 m7, m7           ; sse
@@ -228,6 +234,7 @@
 %endif
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+
 %if %2 == 0 ; !avg
   punpckhbw            m3, m1, m5
   punpcklbw            m1, m5
@@ -237,24 +244,37 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
+  movx                 m0, [srcq]
 %if %2 == 1 ; avg
-%if mmsize == 16
+%if %1 > 4
   movhps               m0, [srcq+src_strideq]
-%else ; mmsize == 8
-  punpckldq            m0, [srcq+src_strideq]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
 %endif
 %else ; !avg
-  movh                 m2, [srcq+src_strideq]
+  movx                 m2, [srcq+src_strideq]
 %endif
-  movh                 m1, [dstq]
-  movh                 m3, [dstq+dst_strideq]
+
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+
 %if %2 == 1 ; avg
+%if %1 > 4
   pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
   punpcklbw            m3, m5
   punpcklbw            m1, m5
+%if %1 > 4
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else ; 4xh
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
 %else ; !avg
   punpcklbw            m0, m5
   punpcklbw            m2, m5
@@ -271,7 +291,7 @@
 %endif
   dec                   block_height
   jg .x_zero_y_zero_loop
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_zero_y_nonzero:
   cmp           y_offsetd, 4
@@ -296,37 +316,41 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m2, [srcq+src_strideq]
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
 %if %2 == 1 ; avg
-%if mmsize == 16
+%if %1 > 4
   movhps               m2, [srcq+src_strideq*2]
-%else ; mmsize == 8
-%if %1 == 4
-  movh                 m1, [srcq+src_strideq*2]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq*2]
   punpckldq            m2, m1
-%else
-  punpckldq            m2, [srcq+src_strideq*2]
 %endif
-%endif
-  movh                 m1, [dstq]
-%if mmsize == 16
+  movx                 m1, [dstq]
+%if %1 > 4
   movlhps              m0, m2
-%else ; mmsize == 8
+%else ; 4xh
   punpckldq            m0, m2
 %endif
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
   pavgb                m0, m2
   punpcklbw            m1, m5
+%if %1 > 4
   pavgb                m0, [secq]
   punpcklbw            m3, m5
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m4, [secq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
 %else ; !avg
-  movh                 m4, [srcq+src_strideq*2]
-  movh                 m1, [dstq]
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m1, [dstq]
   pavgb                m0, m2
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
   pavgb                m2, m4
   punpcklbw            m0, m5
   punpcklbw            m2, m5
@@ -343,7 +367,7 @@
 %endif
   dec                   block_height
   jg .x_zero_y_half_loop
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_zero_y_nonhalf:
   ; x_offset == 0 && y_offset == bilin interpolation
@@ -351,7 +375,7 @@
   lea        bilin_filter, [bilin_filter_m]
 %endif
   shl           y_offsetd, filter_idx_shift
-%if ARCH_X86_64 && mmsize == 16
+%if ARCH_X86_64 && %1 > 4
   mova                 m8, [bilin_filter+y_offsetq]
 %if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
   mova                 m9, [bilin_filter+y_offsetq+16]
@@ -424,12 +448,12 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m2, [srcq+src_strideq]
-  movh                 m4, [srcq+src_strideq*2]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m3, [dstq+dst_strideq]
 %if cpuflag(ssse3)
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   punpcklbw            m0, m2
   punpcklbw            m2, m4
   pmaddubsw            m0, filter_y_a
@@ -449,7 +473,7 @@
   pmullw               m4, filter_y_b
   paddw                m0, m1
   paddw                m2, filter_rnd
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   paddw                m2, m4
 %endif
   psraw                m0, 4
@@ -456,11 +480,21 @@
   psraw                m2, 4
 %if %2 == 1 ; avg
   ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
   packuswb             m0, m2
+%if %1 > 4
   pavgb                m0, [secq]
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
 %endif
+%endif
   punpcklbw            m1, m5
   SUM_SSE              m0, m1, m2, m3, m6, m7
 
@@ -475,7 +509,7 @@
 %undef filter_y_a
 %undef filter_y_b
 %undef filter_rnd
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_nonzero:
   cmp           x_offsetd, 4
@@ -503,30 +537,40 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m4, [srcq+1]
+  movx                 m0, [srcq]
+  movx                 m4, [srcq+1]
 %if %2 == 1 ; avg
-%if mmsize == 16
+%if %1 > 4
   movhps               m0, [srcq+src_strideq]
   movhps               m4, [srcq+src_strideq+1]
-%else ; mmsize == 8
-  punpckldq            m0, [srcq+src_strideq]
-  punpckldq            m4, [srcq+src_strideq+1]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+  movx                 m2, [srcq+src_strideq+1]
+  punpckldq            m4, m2
 %endif
-  movh                 m1, [dstq]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
   pavgb                m0, m4
   punpcklbw            m3, m5
+%if %1 > 4
   pavgb                m0, [secq]
   punpcklbw            m1, m5
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
 %else ; !avg
-  movh                 m2, [srcq+src_strideq]
-  movh                 m1, [dstq]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m1, [dstq]
   pavgb                m0, m4
-  movh                 m4, [srcq+src_strideq+1]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
   pavgb                m2, m4
   punpcklbw            m0, m5
   punpcklbw            m2, m5
@@ -543,7 +587,7 @@
 %endif
   dec                   block_height
   jg .x_half_y_zero_loop
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_half_y_nonzero:
   cmp           y_offsetd, 4
@@ -578,53 +622,58 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m3, [srcq+1]
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
   add                srcq, src_strideq
   pavgb                m0, m3
 .x_half_y_half_loop:
-  movh                 m2, [srcq]
-  movh                 m3, [srcq+1]
+  movx                 m2, [srcq]
+  movx                 m3, [srcq+1]
 %if %2 == 1 ; avg
-%if mmsize == 16
+%if %1 > 4
   movhps               m2, [srcq+src_strideq]
   movhps               m3, [srcq+src_strideq+1]
 %else
-%if %1 == 4
-  movh                 m1, [srcq+src_strideq]
+  movx                 m1, [srcq+src_strideq]
   punpckldq            m2, m1
-  movh                 m1, [srcq+src_strideq+1]
+  movx                 m1, [srcq+src_strideq+1]
   punpckldq            m3, m1
-%else
-  punpckldq            m2, [srcq+src_strideq]
-  punpckldq            m3, [srcq+src_strideq+1]
 %endif
-%endif
   pavgb                m2, m3
-%if mmsize == 16
+%if %1 > 4
   movlhps              m0, m2
   movhlps              m4, m2
-%else ; mmsize == 8
+%else ; 4xh
   punpckldq            m0, m2
-  pshufw               m4, m2, 0xe
+  pshuflw              m4, m2, 0xe
 %endif
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   pavgb                m0, m2
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
+%if %1 > 4
   pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
   punpcklbw            m3, m5
   punpcklbw            m1, m5
+%if %1 > 4
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
 %else ; !avg
-  movh                 m4, [srcq+src_strideq]
-  movh                 m1, [srcq+src_strideq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m1, [srcq+src_strideq+1]
   pavgb                m2, m3
   pavgb                m4, m1
   pavgb                m0, m2
   pavgb                m2, m4
-  movh                 m1, [dstq]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
   punpcklbw            m0, m5
   punpcklbw            m2, m5
   punpcklbw            m3, m5
@@ -641,7 +690,7 @@
 %endif
   dec                   block_height
   jg .x_half_y_half_loop
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_half_y_nonhalf:
   ; x_offset == 0.5 && y_offset == bilin interpolation
@@ -649,7 +698,7 @@
   lea        bilin_filter, [bilin_filter_m]
 %endif
   shl           y_offsetd, filter_idx_shift
-%if ARCH_X86_64 && mmsize == 16
+%if ARCH_X86_64 && %1 > 4
   mova                 m8, [bilin_filter+y_offsetq]
 %if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
   mova                 m9, [bilin_filter+y_offsetq+16]
@@ -724,8 +773,8 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m3, [srcq+1]
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
   add                srcq, src_strideq
   pavgb                m0, m3
 %if notcpuflag(ssse3)
@@ -732,15 +781,15 @@
   punpcklbw            m0, m5
 %endif
 .x_half_y_other_loop:
-  movh                 m2, [srcq]
-  movh                 m1, [srcq+1]
-  movh                 m4, [srcq+src_strideq]
-  movh                 m3, [srcq+src_strideq+1]
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
   pavgb                m2, m1
   pavgb                m4, m3
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
 %if cpuflag(ssse3)
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   punpcklbw            m0, m2
   punpcklbw            m2, m4
   pmaddubsw            m0, filter_y_a
@@ -760,17 +809,27 @@
   pmullw               m1, m4, filter_y_b
   paddw                m2, filter_rnd
   paddw                m2, m1
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
 %endif
   psraw                m0, 4
   psraw                m2, 4
 %if %2 == 1 ; avg
   ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
   packuswb             m0, m2
+%if %1 > 4
   pavgb                m0, [secq]
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
 %endif
+%endif
   punpcklbw            m1, m5
   SUM_SSE              m0, m1, m2, m3, m6, m7
   mova                 m0, m4
@@ -786,7 +845,7 @@
 %undef filter_y_a
 %undef filter_y_b
 %undef filter_rnd
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_nonhalf:
   test          y_offsetd, y_offsetd
@@ -797,7 +856,7 @@
   lea        bilin_filter, [bilin_filter_m]
 %endif
   shl           x_offsetd, filter_idx_shift
-%if ARCH_X86_64 && mmsize == 16
+%if ARCH_X86_64 && %1 > 4
   mova                 m8, [bilin_filter+x_offsetq]
 %if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
   mova                 m9, [bilin_filter+x_offsetq+16]
@@ -865,14 +924,14 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m1, [srcq+1]
-  movh                 m2, [srcq+src_strideq]
-  movh                 m4, [srcq+src_strideq+1]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
 %if cpuflag(ssse3)
   punpcklbw            m0, m1
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   punpcklbw            m2, m4
   pmaddubsw            m0, filter_x_a
   pmaddubsw            m2, filter_x_a
@@ -892,7 +951,7 @@
   pmullw               m4, filter_x_b
   paddw                m0, m1
   paddw                m2, filter_rnd
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   paddw                m2, m4
 %endif
   psraw                m0, 4
@@ -899,11 +958,21 @@
   psraw                m2, 4
 %if %2 == 1 ; avg
   ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
   packuswb             m0, m2
+%if %1 > 4
   pavgb                m0, [secq]
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
 %endif
+%endif
   punpcklbw            m1, m5
   SUM_SSE              m0, m1, m2, m3, m6, m7
 
@@ -918,7 +987,7 @@
 %undef filter_x_a
 %undef filter_x_b
 %undef filter_rnd
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_nonhalf_y_nonzero:
   cmp           y_offsetd, 4
@@ -929,7 +998,7 @@
   lea        bilin_filter, [bilin_filter_m]
 %endif
   shl           x_offsetd, filter_idx_shift
-%if ARCH_X86_64 && mmsize == 16
+%if ARCH_X86_64 && %1 > 4
   mova                 m8, [bilin_filter+x_offsetq]
 %if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
   mova                 m9, [bilin_filter+x_offsetq+16]
@@ -1037,8 +1106,8 @@
   add                srcq, src_strideq
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m1, [srcq+1]
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
 %if cpuflag(ssse3)
   punpcklbw            m0, m1
   pmaddubsw            m0, filter_x_a
@@ -1054,17 +1123,17 @@
   add                srcq, src_strideq
   psraw                m0, 4
 .x_other_y_half_loop:
-  movh                 m2, [srcq]
-  movh                 m1, [srcq+1]
-  movh                 m4, [srcq+src_strideq]
-  movh                 m3, [srcq+src_strideq+1]
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
 %if cpuflag(ssse3)
   punpcklbw            m2, m1
   punpcklbw            m4, m3
   pmaddubsw            m2, filter_x_a
   pmaddubsw            m4, filter_x_a
-  movh                 m1, [dstq]
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
   paddw                m2, filter_rnd
   paddw                m4, filter_rnd
 %else
@@ -1079,9 +1148,9 @@
   pmullw               m3, filter_x_b
   paddw                m4, filter_rnd
   paddw                m2, m1
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   paddw                m4, m3
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
 %endif
   psraw                m2, 4
   psraw                m4, 4
@@ -1089,11 +1158,21 @@
   pavgw                m2, m4
 %if %2 == 1 ; avg
   ; FIXME(rbultje) pipeline - also consider going to bytes here
+%if %1 == 4
+  movlhps              m0, m2
+%endif
   packuswb             m0, m2
+%if %1 > 4
   pavgb                m0, [secq]
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
 %endif
+%endif
   punpcklbw            m3, m5
   punpcklbw            m1, m5
   SUM_SSE              m0, m1, m2, m3, m6, m7
@@ -1110,7 +1189,7 @@
 %undef filter_x_a
 %undef filter_x_b
 %undef filter_rnd
-  STORE_AND_RET
+  STORE_AND_RET %1
 
 .x_nonhalf_y_nonhalf:
 %ifdef PIC
@@ -1118,7 +1197,7 @@
 %endif
   shl           x_offsetd, filter_idx_shift
   shl           y_offsetd, filter_idx_shift
-%if ARCH_X86_64 && mmsize == 16
+%if ARCH_X86_64 && %1 > 4
   mova                 m8, [bilin_filter+x_offsetq]
 %if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
   mova                 m9, [bilin_filter+x_offsetq+16]
@@ -1261,8 +1340,8 @@
   INC_SRC_BY_SRC_STRIDE
   add                dstq, dst_strideq
 %else ; %1 < 16
-  movh                 m0, [srcq]
-  movh                 m1, [srcq+1]
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
 %if cpuflag(ssse3)
   punpcklbw            m0, m1
   pmaddubsw            m0, filter_x_a
@@ -1283,12 +1362,12 @@
   INC_SRC_BY_SRC_STRIDE
 
 .x_other_y_other_loop:
-  movh                 m2, [srcq]
-  movh                 m1, [srcq+1]
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
 
   INC_SRC_BY_SRC_STRIDE
-  movh                 m4, [srcq]
-  movh                 m3, [srcq+1]
+  movx                 m4, [srcq]
+  movx                 m3, [srcq+1]
 
 %if cpuflag(ssse3)
   punpcklbw            m2, m1
@@ -1295,8 +1374,8 @@
   punpcklbw            m4, m3
   pmaddubsw            m2, filter_x_a
   pmaddubsw            m4, filter_x_a
-  movh                 m3, [dstq+dst_strideq]
-  movh                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
   paddw                m2, filter_rnd
   paddw                m4, filter_rnd
   psraw                m2, 4
@@ -1335,9 +1414,9 @@
   pmullw               m1, m4, filter_y_b
   paddw                m2, filter_rnd
   paddw                m0, m3
-  movh                 m3, [dstq+dst_strideq]
+  movx                 m3, [dstq+dst_strideq]
   paddw                m2, m1
-  movh                 m1, [dstq]
+  movx                 m1, [dstq]
   psraw                m0, 4
   psraw                m2, 4
   punpcklbw            m3, m5
@@ -1345,11 +1424,21 @@
 %endif
 %if %2 == 1 ; avg
   ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
   packuswb             m0, m2
+%if %1 > 4
   pavgb                m0, [secq]
   punpckhbw            m2, m0, m5
   punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
 %endif
+%endif
   SUM_SSE              m0, m1, m2, m3, m6, m7
   mova                 m0, m4
 
@@ -1366,7 +1455,8 @@
 %undef filter_y_a
 %undef filter_y_b
 %undef filter_rnd
-  STORE_AND_RET
+%undef movx
+  STORE_AND_RET %1
 %endmacro
 
 ; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical
@@ -1375,26 +1465,22 @@
 ; location in the sse/2 version, rather than duplicating that code in the
 ; binary.
 
-INIT_MMX sse
-SUBPEL_VARIANCE  4
 INIT_XMM sse2
+SUBPEL_VARIANCE  4
 SUBPEL_VARIANCE  8
 SUBPEL_VARIANCE 16
 
-INIT_MMX ssse3
-SUBPEL_VARIANCE  4
 INIT_XMM ssse3
+SUBPEL_VARIANCE  4
 SUBPEL_VARIANCE  8
 SUBPEL_VARIANCE 16
 
-INIT_MMX sse
-SUBPEL_VARIANCE  4, 1
 INIT_XMM sse2
+SUBPEL_VARIANCE  4, 1
 SUBPEL_VARIANCE  8, 1
 SUBPEL_VARIANCE 16, 1
 
-INIT_MMX ssse3
-SUBPEL_VARIANCE  4, 1
 INIT_XMM ssse3
+SUBPEL_VARIANCE  4, 1
 SUBPEL_VARIANCE  8, 1
 SUBPEL_VARIANCE 16, 1
--- a/vpx_dsp/x86/variance_sse2.c
+++ b/vpx_dsp/x86/variance_sse2.c
@@ -320,11 +320,11 @@
                                           int height, unsigned int *sse, \
                                           void *unused0, void *unused)
 #define DECLS(opt1, opt2) \
-  DECL(4, opt2); \
+  DECL(4, opt1); \
   DECL(8, opt1); \
   DECL(16, opt1)
 
-DECLS(sse2, sse);
+DECLS(sse2, sse2);
 DECLS(ssse3, ssse3);
 #undef DECLS
 #undef DECL
@@ -380,10 +380,10 @@
 FN(8,  16,  8, 3, 4, opt1, (int32_t), (int32_t)); \
 FN(8,   8,  8, 3, 3, opt1, (int32_t), (int32_t)); \
 FN(8,   4,  8, 3, 2, opt1, (int32_t), (int32_t)); \
-FN(4,   8,  4, 2, 3, opt2, (int32_t), (int32_t)); \
-FN(4,   4,  4, 2, 2, opt2, (int32_t), (int32_t))
+FN(4,   8,  4, 2, 3, opt1, (int32_t), (int32_t)); \
+FN(4,   4,  4, 2, 2, opt1, (int32_t), (int32_t))
 
-FNS(sse2, sse);
+FNS(sse2, sse2);
 FNS(ssse3, ssse3);
 
 #undef FNS
@@ -401,11 +401,11 @@
                                             int height, unsigned int *sse, \
                                             void *unused0, void *unused)
 #define DECLS(opt1, opt2) \
-DECL(4, opt2); \
+DECL(4, opt1); \
 DECL(8, opt1); \
 DECL(16, opt1)
 
-DECLS(sse2, sse);
+DECLS(sse2, sse2);
 DECLS(ssse3, ssse3);
 #undef DECL
 #undef DECLS
@@ -466,8 +466,8 @@
 FN(8,  16,  8, 3, 4, opt1, (uint32_t), (int32_t)); \
 FN(8,   8,  8, 3, 3, opt1, (uint32_t), (int32_t)); \
 FN(8,   4,  8, 3, 2, opt1, (uint32_t), (int32_t)); \
-FN(4,   8,  4, 2, 3, opt2, (uint32_t), (int32_t)); \
-FN(4,   4,  4, 2, 2, opt2, (uint32_t), (int32_t))
+FN(4,   8,  4, 2, 3, opt1, (uint32_t), (int32_t)); \
+FN(4,   4,  4, 2, 2, opt1, (uint32_t), (int32_t))
 
 FNS(sse2, sse);
 FNS(ssse3, ssse3);