diff --git a/aom_dsp/x86/sad4d_sse2.asm b/aom_dsp/x86/sad4d_sse2.asm
index 9ab44c1..6de708b 100644
--- a/aom_dsp/x86/sad4d_sse2.asm
+++ b/aom_dsp/x86/sad4d_sse2.asm
@@ -22,114 +22,95 @@
   pavgb                 %2, m2
   lea                   second_predq, [second_predq+8]
 %endmacro
-; 'mflag' affect a lot how the code works.
+; 'spill_src_stride' affect a lot how the code works.
 ;
-; When 'mflag' is false, the 'src_strideq' resides in register,
-; [srcq + src_strideq + offset] is allowed, so we can simply
-; use such form to access src memory and don't bother to update
-; 'srcq' at each line. We only update 'srcq' each two-lines using
-; a compact LEA instruction like [srcq+src_strideq*2].
+; When 'spill_src_stride' is false, the 'src_strideq' resides in
+; register, [srcq + src_strideq + offset] is allowed, so we can simply
+; use such form to access src memory and don't bother to update 'srcq'
+; at each line. We only update 'srcq' each two-lines using a compact
+; LEA instruction like [srcq+src_strideq*2].
 ;
-; When 'mflag' is true, the 'src_strideq' resides in memory.
+; When 'spill_src_stride' is true, the 'src_strideq' resides in memory.
 ; we cannot use above form to access memory, we have to update
 ; 'srcq' at each line break. As we process two parts (first,second)
 ; together in each macro function, the second part may also sit
 ; in the next line, which means we also need to possibly add
 ; one 'src_strideq' to 'srcq' before processing second part.
 
-%macro HANDLE_FIRST_OFFSET 2
-  %define first_offset %2
-  %if mflag == 0 && %1 == 1
-    %define first_offset (src_strideq + %2)
-  %endif
-%endmacro
-
-; first_extraline, second_extraline, in_line_offset
-%macro HANDLE_SECOND_OFFSET 3
-  %define second_offset %3
-  %if mflag && %1 == 0 && %2 == 1
+%macro HANDLE_SECOND_OFFSET 0
+  %if spill_src_stride
+    %define second_offset 0
     add srcq, src_strideq
-  %endif
-  %if mflag == 0 && %2 == 1
-    %define second_offset (src_strideq + %3)
+  %else
+    %define second_offset (src_strideq)
   %endif
 %endmacro
 
-; Notes for line_ending:
-; 0 -- not a line ending
-; 1 -- line ending of a odd line [line numbers starts from one]
-; 2 -- line ending of a even line
 ; This is specically designed to handle when src_strideq is a
 ; memory position, under such case, we can not accomplish
 ; complex address calculation using LEA, and fall back to
 ; using simple ADD instruction at each line ending.
-%macro ADVANCE_END_OF_LINE 1
-  %if mflag
+%macro ADVANCE_END_OF_TWO_LINES 0
+  %if spill_src_stride
     add srcq, src_strideq
-  %endif
-  %if mflag == 0 && %1 == 2
-    lea                 srcq, [srcq +src_strideq*2]
+  %else
+    lea                 srcq, [srcq+src_strideq*2]
   %endif
 
-  %if %1 == 2
-    lea                ref1q, [ref1q+ref_strideq*2]
-    lea                ref2q, [ref2q+ref_strideq*2]
-    lea                ref3q, [ref3q+ref_strideq*2]
-    lea                ref4q, [ref4q+ref_strideq*2]
-  %endif
+; note: ref_stride is never spilled when processing two lines
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
 %endmacro
 
-; Please note that the second_offset of src is for in_line_offset,
-; so it is less than src_stride.
-; PROCESS_4x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;               {first, second}_extraline, line_ending
-%macro PROCESS_4x2x4 9
-  HANDLE_FIRST_OFFSET   %7, %2
-  movd                  m0, [srcq + first_offset]
-  HANDLE_SECOND_OFFSET  %7, %8, %4
+; PROCESS_4x2x4 first, do_avg
+%macro PROCESS_4x2x4 2
+  movd                  m0, [srcq]
+  HANDLE_SECOND_OFFSET
 %if %1 == 1
-  movd                  m6, [ref1q+%3]
-  movd                  m4, [ref2q+%3]
-  movd                  m7, [ref3q+%3]
-  movd                  m5, [ref4q+%3]
+  movd                  m6, [ref1q]
+  movd                  m4, [ref2q]
+  movd                  m7, [ref3q]
+  movd                  m5, [ref4q]
 
   movd                  m1, [srcq + second_offset]
-  movd                  m2, [ref1q+%5]
+  movd                  m2, [ref1q+ref_strideq]
   punpckldq             m0, m1
   punpckldq             m6, m2
-  movd                  m1, [ref2q+%5]
-  movd                  m2, [ref3q+%5]
-  movd                  m3, [ref4q+%5]
+  movd                  m1, [ref2q+ref_strideq]
+  movd                  m2, [ref3q+ref_strideq]
+  movd                  m3, [ref4q+ref_strideq]
   punpckldq             m4, m1
   punpckldq             m7, m2
   punpckldq             m5, m3
   movlhps               m0, m0
   movlhps               m6, m4
   movlhps               m7, m5
-%if %6 == 1
+%if %2 == 1
   AVG_4x2x4             m6, m7
 %endif
   psadbw                m6, m0
   psadbw                m7, m0
 %else
-  movd                  m1, [ref1q+%3]
-  movd                  m5, [ref1q+%5]
-  movd                  m2, [ref2q+%3]
-  movd                  m4, [ref2q+%5]
+  movd                  m1, [ref1q]
+  movd                  m5, [ref1q+ref_strideq]
+  movd                  m2, [ref2q]
+  movd                  m4, [ref2q+ref_strideq]
   punpckldq             m1, m5
   punpckldq             m2, m4
-  movd                  m3, [ref3q+%3]
-  movd                  m5, [ref3q+%5]
+  movd                  m3, [ref3q]
+  movd                  m5, [ref3q+ref_strideq]
   punpckldq             m3, m5
-  movd                  m4, [ref4q+%3]
-  movd                  m5, [ref4q+%5]
+  movd                  m4, [ref4q]
+  movd                  m5, [ref4q+ref_strideq]
   punpckldq             m4, m5
   movd                  m5, [srcq + second_offset]
   punpckldq             m0, m5
   movlhps               m0, m0
   movlhps               m1, m2
   movlhps               m3, m4
-%if %6 == 1
+%if %2 == 1
   AVG_4x2x4             m1, m3
 %endif
   psadbw                m1, m0
@@ -137,28 +118,23 @@
   paddd                 m6, m1
   paddd                 m7, m3
 %endif
-%if %9 > 0
-  ADVANCE_END_OF_LINE %9
-%endif
 %endmacro
 
-; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;               {first,second}_extraline, line_ending
-%macro PROCESS_8x2x4 9
-  HANDLE_FIRST_OFFSET   %7, %2
-  movh                  m0, [srcq + first_offset]
-  HANDLE_SECOND_OFFSET  %7, %8, %4
+; PROCESS_8x2x4 first, do_avg
+%macro PROCESS_8x2x4 2
+  movh                  m0, [srcq]
+  HANDLE_SECOND_OFFSET
 %if %1 == 1
-  movh                  m4, [ref1q+%3]
-  movh                  m5, [ref2q+%3]
-  movh                  m6, [ref3q+%3]
-  movh                  m7, [ref4q+%3]
+  movh                  m4, [ref1q]
+  movh                  m5, [ref2q]
+  movh                  m6, [ref3q]
+  movh                  m7, [ref4q]
   movhps                m0, [srcq + second_offset]
-  movhps                m4, [ref1q+%5]
-  movhps                m5, [ref2q+%5]
-  movhps                m6, [ref3q+%5]
-  movhps                m7, [ref4q+%5]
-%if %6 == 1
+  movhps                m4, [ref1q+ref_strideq]
+  movhps                m5, [ref2q+ref_strideq]
+  movhps                m6, [ref3q+ref_strideq]
+  movhps                m7, [ref4q+ref_strideq]
+%if %2 == 1
   movu                  m3, [second_predq]
   pavgb                 m4, m3
   pavgb                 m5, m3
@@ -171,12 +147,12 @@
   psadbw                m6, m0
   psadbw                m7, m0
 %else
-  movh                  m1, [ref1q+%3]
-  movh                  m2, [ref2q+%3]
+  movh                  m1, [ref1q]
+  movh                  m2, [ref2q]
   movhps                m0, [srcq + second_offset]
-  movhps                m1, [ref1q+%5]
-  movhps                m2, [ref2q+%5]
-%if %6 == 1
+  movhps                m1, [ref1q+ref_strideq]
+  movhps                m2, [ref2q+ref_strideq]
+%if %2 == 1
   movu                  m3, [second_predq]
   pavgb                 m1, m3
   pavgb                 m2, m3
@@ -186,11 +162,11 @@
   paddd                 m4, m1
   paddd                 m5, m2
 
-  movh                  m1, [ref3q+%3]
-  movhps                m1, [ref3q+%5]
-  movh                  m2, [ref4q+%3]
-  movhps                m2, [ref4q+%5]
-%if %6 == 1
+  movh                  m1, [ref3q]
+  movhps                m1, [ref3q+ref_strideq]
+  movh                  m2, [ref4q]
+  movhps                m2, [ref4q+ref_strideq]
+%if %2 == 1
   pavgb                 m1, m3
   pavgb                 m2, m3
   lea                   second_predq, [second_predq+mmsize]
@@ -200,24 +176,16 @@
   paddd                 m6, m1
   paddd                 m7, m2
 %endif
-%if %9 > 0
-  ADVANCE_END_OF_LINE %9
-%endif
 %endmacro
 
-; PROCESS_16x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;                {first,second}_extraline, line_ending
-%macro PROCESS_16x2x4 9
-  ; 1st 16 px
-  HANDLE_FIRST_OFFSET   %7, %2
-  mova                  m0, [srcq + first_offset]
-  HANDLE_SECOND_OFFSET  %7, %8, %4
+; PROCESS_FIRST_MMSIZE do_avg
+%macro PROCESS_FIRST_MMSIZE 1
+  mova                  m0, [srcq]
+  movu                  m4, [ref1q]
+  movu                  m5, [ref2q]
+  movu                  m6, [ref3q]
+  movu                  m7, [ref4q]
 %if %1 == 1
-  movu                  m4, [ref1q+%3]
-  movu                  m5, [ref2q+%3]
-  movu                  m6, [ref3q+%3]
-  movu                  m7, [ref4q+%3]
-%if %6 == 1
   movu                  m3, [second_predq]
   pavgb                 m4, m3
   pavgb                 m5, m3
@@ -229,10 +197,14 @@
   psadbw                m5, m0
   psadbw                m6, m0
   psadbw                m7, m0
-%else ; %1 == 1
-  movu                  m1, [ref1q+%3]
-  movu                  m2, [ref2q+%3]
-%if %6 == 1
+%endmacro
+
+; PROCESS_16x1x4 offset, do_avg
+%macro PROCESS_16x1x4 2
+  mova                  m0, [srcq + %1]
+  movu                  m1, [ref1q + ref_offsetq + %1]
+  movu                  m2, [ref2q + ref_offsetq + %1]
+%if %2 == 1
   movu                  m3, [second_predq]
   pavgb                 m1, m3
   pavgb                 m2, m3
@@ -242,9 +214,9 @@
   paddd                 m4, m1
   paddd                 m5, m2
 
-  movu                  m1, [ref3q+%3]
-  movu                  m2, [ref4q+%3]
-%if %6 == 1
+  movu                  m1, [ref3q + ref_offsetq + %1]
+  movu                  m2, [ref4q + ref_offsetq + %1]
+%if %2 == 1
   pavgb                 m1, m3
   pavgb                 m2, m3
   lea                   second_predq, [second_predq+mmsize]
@@ -253,60 +225,6 @@
   psadbw                m2, m0
   paddd                 m6, m1
   paddd                 m7, m2
-%endif ; %1 == 1
-
-  ; 2nd 16 px
-  mova                  m0, [srcq + second_offset]
-  movu                  m1, [ref1q+%5]
-  movu                  m2, [ref2q+%5]
-
-%if %6 == 1
-  movu                  m3, [second_predq]
-  pavgb                 m1, m3
-  pavgb                 m2, m3
-%endif
-  psadbw                m1, m0
-  psadbw                m2, m0
-  paddd                 m4, m1
-  paddd                 m5, m2
-
-  movu                  m1, [ref3q+%5]
-  movu                  m2, [ref4q+%5]
-
-%if %9 > 0
-  ADVANCE_END_OF_LINE %9
-%endif
-
-%if %6 == 1
-  pavgb                 m1, m3
-  pavgb                 m2, m3
-  lea                   second_predq, [second_predq+mmsize]
-%endif
-  psadbw                m1, m0
-  psadbw                m2, m0
-  paddd                 m6, m1
-  paddd                 m7, m2
-%endmacro
-
-; PROCESS_32x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;                {first,second}_extraline, line_ending
-%macro PROCESS_32x2x4 9
-  PROCESS_16x2x4 %1, %2, %3, %2 + 16, %3 + 16, %6, %7, %7, %8 - %7
-  PROCESS_16x2x4  0, %4, %5, %4 + 16, %5 + 16, %6, %8, %8, %9
-%endmacro
-
-; PROCESS_64x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;                {first,second}_extraline, line_ending
-%macro PROCESS_64x2x4 9
-  PROCESS_32x2x4 %1, %2, %3, %2 + 32, %3 + 32, %6, %7, %7, %8 - %7
-  PROCESS_32x2x4  0, %4, %5, %4 + 32, %5 + 32, %6, %8, %8, %9
-%endmacro
-
-; PROCESS_128x2x4 first, off_{first,second}_{src,ref}, do_avg,
-;                 {first,second}_extraline, line_ending
-%macro PROCESS_128x2x4 9
-  PROCESS_64x2x4 %1, %2, %3, %2 + 64, %3 + 64, %6, %7, %7, %8 - %7
-  PROCESS_64x2x4  0, %4, %5, %4 + 64, %5 + 64, %6, %8, %8, %9
 %endmacro
 
 ; void aom_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
@@ -318,38 +236,118 @@
 ;   3: If 0, then normal sad, else avg
 ;   4: If 0, then normal sad, else skip rows
 %macro SADNXN4D 2-4 0,0
+
+%define spill_src_stride 0
+%define spill_ref_stride 0
+%define spill_cnt 0
+
+; Whether a shared offset should be used instead of adding strides to
+; each reference array. With this option, only one line will be processed
+; per loop iteration.
+%define use_ref_offset (%1 >= mmsize)
+
+; Remove loops in the 4x4 and 8x4 case
+%define use_loop (use_ref_offset || %2 > 4)
+
 %if %4 == 1  ; skip rows
 %if ARCH_X86_64
-cglobal sad_skip_%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
-                              res, ref2, ref3, ref4
+%if use_ref_offset
+cglobal sad_skip_%1x%2x4d, 5, 10, 8, src, src_stride, ref1, ref_stride, res, \
+                                     ref2, ref3, ref4, cnt, ref_offset
+%elif use_loop
+cglobal sad_skip_%1x%2x4d, 5, 9, 8, src, src_stride, ref1, ref_stride, res, \
+                                    ref2, ref3, ref4, cnt
 %else
-cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
-                              ref2, ref3, ref4
+cglobal sad_skip_%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, res, \
+                                    ref2, ref3, ref4
+%endif
+%else
+%if use_ref_offset
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, ref_offset, ref1, cnt, ref2, ref3, \
+                                    ref4
+%define spill_src_stride 1
+%define spill_ref_stride 1
+%elif use_loop
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, cnt, ref1, ref_stride, ref2, \
+                                    ref3, ref4
+%define spill_src_stride 1
+%else
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, ref2, \
+                                    ref3, ref4
+%endif
 %endif
 %elif %3 == 0  ; normal sad
 %if ARCH_X86_64
-cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
-                              res, ref2, ref3, ref4
+%if use_ref_offset
+cglobal sad%1x%2x4d, 5, 10, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                               ref3, ref4, cnt, ref_offset
+%elif use_loop
+cglobal sad%1x%2x4d, 5, 9, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                              ref3, ref4, cnt
 %else
-cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
-                              ref2, ref3, ref4
+cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                              ref3, ref4
+%endif
+%else
+%if use_ref_offset
+cglobal sad%1x%2x4d, 4, 7, 8, src, ref_offset, ref1, cnt, ref2, ref3, ref4
+  %define spill_src_stride 1
+  %define spill_ref_stride 1
+%elif use_loop
+cglobal sad%1x%2x4d, 4, 7, 8, src, cnt, ref1, ref_stride, ref2, ref3, ref4
+  %define spill_src_stride 1
+%else
+cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, ref2, ref3, \
+                              ref4
+%endif
 %endif
 %else ; avg
 %if ARCH_X86_64
+%if use_ref_offset
+cglobal sad%1x%2x4d_avg, 6, 11, 8, src, src_stride, ref1, ref_stride, \
+                                   second_pred, res, ref2, ref3, ref4, cnt, \
+                                   ref_offset
+%elif use_loop
 cglobal sad%1x%2x4d_avg, 6, 10, 8, src, src_stride, ref1, ref_stride, \
-                                  second_pred, res, ref2, ref3, ref4
+                                   second_pred, res, ref2, ref3, ref4, cnt
 %else
-cglobal sad%1x%2x4d_avg, 5, 7, 8, src, ref4, ref1, ref_stride, \
-                                  second_pred, ref2, ref3
-  %define src_strideq r1mp
-  %define src_strided r1mp
+cglobal sad%1x%2x4d_avg, 6, 9, 8, src, src_stride, ref1, ref_stride, \
+                                   second_pred, res, ref2, ref3, ref4
+%endif
+%else
+%if use_ref_offset
+cglobal sad%1x%2x4d_avg, 5, 7, 8, src, ref4, ref1, ref_offset, second_pred, ref2, ref3
+  %define spill_src_stride 1
+  %define spill_ref_stride 1
+  %define spill_cnt 1
+%elif use_loop
+cglobal sad%1x%2x4d_avg, 5, 7, 8, src, ref4, ref1, ref_stride, second_pred, ref2, ref3
+  %define spill_src_stride 1
+  %define spill_cnt 1
+%else
+cglobal sad%1x%2x4d_avg, 5, 7, 8, src, ref4, ref1, ref_stride, second_pred, ref2, ref3
+  %define spill_src_stride 1
+%endif
 %endif
 %endif
 
-  %define mflag ((1 - ARCH_X86_64) & %3)
+%if spill_src_stride
+  %define src_strideq r1mp
+  %define src_strided r1mp
+%endif
+%if spill_ref_stride
+  %define ref_strideq r3mp
+  %define ref_strided r3mp
+%endif
+
+%if spill_cnt
+  SUB                  rsp, 4
+  %define cntd word [rsp]
+%endif
+
 %if %4 == 1
-  lea          src_strided, [2*src_strided]
-  lea          ref_strided, [2*ref_strided]
+  sal          src_strided, 1
+  sal          ref_strided, 1
 %endif
   movsxdifnidn src_strideq, src_strided
   movsxdifnidn ref_strideq, ref_strided
@@ -359,18 +357,67 @@
   mov                ref4q, [ref1q+gprsize*3]
   mov                ref1q, [ref1q+gprsize*0]
 
-  PROCESS_%1x2x4 1, 0, 0, 0, ref_strideq, %3, 0, 1, 2
-%if %4 == 1  ; downsample number of rows by 2
-%define num_rep (%2-8)/4
-%else
-%define num_rep (%2-4)/2
-%endif
-%rep num_rep
-  PROCESS_%1x2x4 0, 0, 0, 0, ref_strideq, %3, 0, 1, 2
-%endrep
-%undef num_rep
-  PROCESS_%1x2x4 0, 0, 0, 0, ref_strideq, %3, 0, 1, 2
+; Is the loop for this wxh in another function?
+; If so, we jump into that function for the loop and returning
+%define external_loop (use_ref_offset && %1 > mmsize && %1 != %2)
 
+%if use_ref_offset
+  PROCESS_FIRST_MMSIZE %3
+%if %1 > mmsize
+  mov          ref_offsetq, 0
+  mov                 cntd, %2 >> %4
+; Jump part way into the loop for the square version of this width
+%if %3 == 1
+  jmp mangle(private_prefix %+ _sad%1x%1x4d_avg %+ SUFFIX).midloop
+%elif %4 == 1
+  jmp mangle(private_prefix %+ _sad_skip_%1x%1x4d %+ SUFFIX).midloop
+%else
+  jmp mangle(private_prefix %+ _sad%1x%1x4d %+ SUFFIX).midloop
+%endif
+%else
+  mov          ref_offsetq, ref_strideq
+  add                 srcq, src_strideq
+  mov                 cntd, (%2 >> %4) - 1
+%endif
+%if external_loop == 0
+.loop:
+; Unrolled horizontal loop
+%assign h_offset 0
+%rep %1/mmsize
+  PROCESS_16x1x4 h_offset, %3
+%if h_offset == 0
+; The first row of the first column is done outside the loop and jumps here
+.midloop:
+%endif
+%assign h_offset h_offset+mmsize
+%endrep
+
+  add                 srcq, src_strideq
+  add          ref_offsetq, ref_strideq
+  sub                 cntd, 1
+  jnz .loop
+%endif
+%else
+  PROCESS_%1x2x4 1, %3
+  ADVANCE_END_OF_TWO_LINES
+%if use_loop
+  mov                 cntd, (%2/2 >> %4) - 1
+.loop:
+%endif
+  PROCESS_%1x2x4 0, %3
+%if use_loop
+  ADVANCE_END_OF_TWO_LINES
+  sub                 cntd, 1
+  jnz .loop
+%endif
+%endif
+
+%if spill_cnt
+; Undo stack allocation for cnt
+  ADD                  rsp, 4
+%endif
+
+%if external_loop == 0
 %if %3 == 0
   %define resultq r4
   %define resultmp r4mp
@@ -379,6 +426,16 @@
   %define resultmp r5mp
 %endif
 
+; Undo modifications on parameters on the stack
+%if %4 == 1
+%if spill_src_stride
+  shr          src_strided, 1
+%endif
+%if spill_ref_stride
+  shr          ref_strided, 1
+%endif
+%endif
+
 %if %1 > 4
   pslldq                m5, 4
   pslldq                m7, 4
@@ -407,6 +464,7 @@
   movq              [resultq+8], m7
   RET
 %endif
+%endif ; external_loop == 0
 %endmacro
 
 INIT_XMM sse2