Quick modifications to wide loopfilter intrinsic functions

Modified to work with 8x8 blocks of memory.  Will revisit
later for further optimizations.  For the HD clip used, the
decoder improved my 20%.

Change-Id: Ia0057f55d66d1445882351ea6c43b595a5a980e5

5 files changed, 119 insertions, 139 deletions

diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index cad08bbf3..7b3f0be24 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -184,7 +184,7 @@ static void filter_selectively_vert(uint8_t *s, int pitch,
     if (mask & 1) {
       if (mask_16x16 & 1) {
         vp9_mb_lpf_vertical_edge_w(s, pitch, lfi->mblim, lfi->lim,
-                                   lfi->hev_thr, 1);
+                                   lfi->hev_thr);
         assert(!(mask_8x8 & 1));
         assert(!(mask_4x4 & 1));
         assert(!(mask_4x4_int & 1));
@@ -229,7 +229,7 @@ static void filter_selectively_horiz(uint8_t *s, int pitch,
       if (!only_4x4_1) {
         if (mask_16x16 & 1) {
           vp9_mb_lpf_horizontal_edge_w(s, pitch, lfi->mblim, lfi->lim,
-                                       lfi->hev_thr, 1);
+                                       lfi->hev_thr);
           assert(!(mask_8x8 & 1));
           assert(!(mask_4x4 & 1));
           assert(!(mask_4x4_int & 1));
diff --git a/vp9/common/vp9_loopfilter.h b/vp9/common/vp9_loopfilter.h
index 65f522bab..ce954c0c3 100644
--- a/vp9/common/vp9_loopfilter.h
+++ b/vp9/common/vp9_loopfilter.h
@@ -82,15 +82,4 @@ void vp9_loop_filter_partial_frame(struct VP9Common *cm,
 void vp9_loop_filter_update_sharpness(loop_filter_info_n *lfi,
                                       int sharpness_lvl);
 
-void vp9_mb_lpf_horizontal_edge_w(unsigned char *s, int p,
-                                  const unsigned char *blimit,
-                                  const unsigned char *limit,
-                                  const unsigned char *thresh,
-                                  int count);
-
-void vp9_mb_lpf_vertical_edge_w(unsigned char *s, int p,
-                                const unsigned char *blimit,
-                                const unsigned char *limit,
-                                const unsigned char *thresh,
-                                int count);
 #endif  // VP9_COMMON_VP9_LOOPFILTER_H_
diff --git a/vp9/common/vp9_loopfilter_filters.c b/vp9/common/vp9_loopfilter_filters.c
index bf03692a0..0efbcafe0 100644
--- a/vp9/common/vp9_loopfilter_filters.c
+++ b/vp9/common/vp9_loopfilter_filters.c
@@ -255,16 +255,15 @@ static INLINE void wide_mbfilter(int8_t mask, uint8_t hev,
   }
 }
 
-void vp9_mb_lpf_horizontal_edge_w(uint8_t *s, int p,
-                                 const uint8_t *blimit,
-                                 const uint8_t *limit,
-                                 const uint8_t *thresh,
-                                 int count) {
+void vp9_mb_lpf_horizontal_edge_w_c(uint8_t *s, int p,
+                                    const uint8_t *blimit,
+                                    const uint8_t *limit,
+                                    const uint8_t *thresh) {
   int i;
 
   // loop filter designed to work using chars so that we can make maximum use
   // of 8 bit simd instructions.
-  for (i = 0; i < 8 * count; ++i) {
+  for (i = 0; i < 8; ++i) {
     const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
     const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
     const int8_t mask = filter_mask(*limit, *blimit,
@@ -285,14 +284,13 @@ void vp9_mb_lpf_horizontal_edge_w(uint8_t *s, int p,
   }
 }
 
-void vp9_mb_lpf_vertical_edge_w(uint8_t *s, int p,
-                                const uint8_t *blimit,
-                                const uint8_t *limit,
-                                const uint8_t *thresh,
-                                int count) {
+void vp9_mb_lpf_vertical_edge_w_c(uint8_t *s, int p,
+                                  const uint8_t *blimit,
+                                  const uint8_t *limit,
+                                  const uint8_t *thresh) {
   int i;
 
-  for (i = 0; i < 8 * count; ++i) {
+  for (i = 0; i < 8; ++i) {
     const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
     const uint8_t q0 = s[0], q1 = s[1],  q2 = s[2], q3 = s[3];
     const int8_t mask = filter_mask(*limit, *blimit,
diff --git a/vp9/common/vp9_rtcd_defs.sh b/vp9/common/vp9_rtcd_defs.sh
index 74d3ab57f..ddfc85aa0 100644
--- a/vp9/common/vp9_rtcd_defs.sh
+++ b/vp9/common/vp9_rtcd_defs.sh
@@ -86,8 +86,8 @@ fi
 #
 # Loopfilter
 #
-prototype void vp9_mb_lpf_vertical_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count"
-specialize vp9_mb_lpf_vertical_edge_w
+prototype void vp9_mb_lpf_vertical_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"
+specialize vp9_mb_lpf_vertical_edge_w sse2
 
 prototype void vp9_mbloop_filter_vertical_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count"
 specialize vp9_mbloop_filter_vertical_edge
@@ -95,8 +95,8 @@ specialize vp9_mbloop_filter_vertical_edge
 prototype void vp9_loop_filter_vertical_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count"
 specialize vp9_loop_filter_vertical_edge mmx
 
-prototype void vp9_mb_lpf_horizontal_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count"
-specialize vp9_mb_lpf_horizontal_edge_w
+prototype void vp9_mb_lpf_horizontal_edge_w "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh"
+specialize vp9_mb_lpf_horizontal_edge_w sse2
 
 prototype void vp9_mbloop_filter_horizontal_edge "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count"
 specialize vp9_mbloop_filter_horizontal_edge
diff --git a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c b/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
index 7982ca6a2..9ab51208d 100644
--- a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
+++ b/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
@@ -23,14 +23,14 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
                                        const unsigned char *_blimit,
                                        const unsigned char *_limit,
                                        const unsigned char *_thresh) {
-  DECLARE_ALIGNED(16, unsigned char, flat2_op[7][16]);
-  DECLARE_ALIGNED(16, unsigned char, flat2_oq[7][16]);
+  DECLARE_ALIGNED(16, unsigned char, flat2_op[7][8]);
+  DECLARE_ALIGNED(16, unsigned char, flat2_oq[7][8]);
 
-  DECLARE_ALIGNED(16, unsigned char, flat_op[3][16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_oq[3][16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op[3][8]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq[3][8]);
 
-  DECLARE_ALIGNED(16, unsigned char, ap[8][16]);
-  DECLARE_ALIGNED(16, unsigned char, aq[8][16]);
+  DECLARE_ALIGNED(16, unsigned char, ap[8][8]);
+  DECLARE_ALIGNED(16, unsigned char, aq[8][8]);
 
 
   __m128i mask, hev, flat, flat2;
@@ -50,27 +50,27 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
   const __m128i blimit =
       _mm_shuffle_epi32(_mm_cvtsi32_si128((int)extended_blimit), 0);
 
-  p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
-  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
-  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
-  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
-  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
-  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
-  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
-  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
-  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
-  q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
-
-  _mm_store_si128((__m128i *)ap[4], p4);
-  _mm_store_si128((__m128i *)ap[3], p3);
-  _mm_store_si128((__m128i *)ap[2], p2);
-  _mm_store_si128((__m128i *)ap[1], p1);
-  _mm_store_si128((__m128i *)ap[0], p0);
-  _mm_store_si128((__m128i *)aq[4], q4);
-  _mm_store_si128((__m128i *)aq[3], q3);
-  _mm_store_si128((__m128i *)aq[2], q2);
-  _mm_store_si128((__m128i *)aq[1], q1);
-  _mm_store_si128((__m128i *)aq[0], q0);
+  p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+  p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+  q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
+  q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p));
+  q4 = _mm_loadl_epi64((__m128i *)(s + 4 * p));
+
+  _mm_storel_epi64((__m128i *)ap[4], p4);
+  _mm_storel_epi64((__m128i *)ap[3], p3);
+  _mm_storel_epi64((__m128i *)ap[2], p2);
+  _mm_storel_epi64((__m128i *)ap[1], p1);
+  _mm_storel_epi64((__m128i *)ap[0], p0);
+  _mm_storel_epi64((__m128i *)aq[4], q4);
+  _mm_storel_epi64((__m128i *)aq[3], q3);
+  _mm_storel_epi64((__m128i *)aq[2], q2);
+  _mm_storel_epi64((__m128i *)aq[1], q1);
+  _mm_storel_epi64((__m128i *)aq[0], q0);
 
 
   {
@@ -188,33 +188,33 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
       flat = _mm_cmpeq_epi8(flat, zero);
       flat = _mm_and_si128(flat, mask);
 
-      p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
-      q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+      p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+      q5 = _mm_loadl_epi64((__m128i *)(s + 5 * p));
       flat2 = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p5, p0),
                                        _mm_subs_epu8(p0, p5)),
                            _mm_or_si128(_mm_subs_epu8(q5, q0),
                                         _mm_subs_epu8(q0, q5)));
-      _mm_store_si128((__m128i *)ap[5], p5);
-      _mm_store_si128((__m128i *)aq[5], q5);
+      _mm_storel_epi64((__m128i *)ap[5], p5);
+      _mm_storel_epi64((__m128i *)aq[5], q5);
       flat2 = _mm_max_epu8(work, flat2);
-      p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
-      q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+      p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+      q6 = _mm_loadl_epi64((__m128i *)(s + 6 * p));
       work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p6, p0),
                                        _mm_subs_epu8(p0, p6)),
                            _mm_or_si128(_mm_subs_epu8(q6, q0),
                                         _mm_subs_epu8(q0, q6)));
-      _mm_store_si128((__m128i *)ap[6], p6);
-      _mm_store_si128((__m128i *)aq[6], q6);
+      _mm_storel_epi64((__m128i *)ap[6], p6);
+      _mm_storel_epi64((__m128i *)aq[6], q6);
       flat2 = _mm_max_epu8(work, flat2);
 
-      p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
-      q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
+      p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+      q7 = _mm_loadl_epi64((__m128i *)(s + 7 * p));
       work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p7, p0),
                                        _mm_subs_epu8(p0, p7)),
                            _mm_or_si128(_mm_subs_epu8(q7, q0),
                                         _mm_subs_epu8(q0, q7)));
-      _mm_store_si128((__m128i *)ap[7], p7);
-      _mm_store_si128((__m128i *)aq[7], q7);
+      _mm_storel_epi64((__m128i *)ap[7], p7);
+      _mm_storel_epi64((__m128i *)aq[7], q7);
       flat2 = _mm_max_epu8(work, flat2);
       flat2 = _mm_subs_epu8(flat2, one);
       flat2 = _mm_cmpeq_epi8(flat2, zero);
@@ -226,30 +226,26 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     {
       const __m128i eight = _mm_set1_epi16(8);
       const __m128i four = _mm_set1_epi16(4);
-      __m128i temp_flat2 = flat2;
-      unsigned char *src = s;
-      int i = 0;
-      do {
+      {
         __m128i workp_shft;
         __m128i a, b, c;
 
-        unsigned int off = i * 8;
-        p7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[7] + off)), zero);
-        p6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[6] + off)), zero);
-        p5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[5] + off)), zero);
-        p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[4] + off)), zero);
-        p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[3] + off)), zero);
-        p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[2] + off)), zero);
-        p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[1] + off)), zero);
-        p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[0] + off)), zero);
-        q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[0] + off)), zero);
-        q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[1] + off)), zero);
-        q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[2] + off)), zero);
-        q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[3] + off)), zero);
-        q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[4] + off)), zero);
-        q5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[5] + off)), zero);
-        q6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[6] + off)), zero);
-        q7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[7] + off)), zero);
+        p7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[7])), zero);
+        p6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[6])), zero);
+        p5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[5])), zero);
+        p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[4])), zero);
+        p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[3])), zero);
+        p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[2])), zero);
+        p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[1])), zero);
+        p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(ap[0])), zero);
+        q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[0])), zero);
+        q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[1])), zero);
+        q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[2])), zero);
+        q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[3])), zero);
+        q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[4])), zero);
+        q5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[5])), zero);
+        q6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[6])), zero);
+        q7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(aq[7])), zero);
 
         c = _mm_sub_epi16(_mm_slli_epi16(p7, 3), p7);  // p7 * 7
         c = _mm_add_epi16(_mm_slli_epi16(p6, 1), _mm_add_epi16(p4, c));
@@ -370,120 +366,117 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
         workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
         _mm_storel_epi64((__m128i *)&flat2_oq[6][i*8],
                          _mm_packus_epi16(workp_shft, workp_shft));
-
-        temp_flat2 = _mm_srli_si128(temp_flat2, 8);
-        src += 8;
-      } while (++i < 2);
+      }
     }
     // wide flat
     // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-    work_a = _mm_load_si128((__m128i *)ap[2]);
-    p2 = _mm_load_si128((__m128i *)flat_op[2]);
+    work_a = _mm_loadl_epi64((__m128i *)ap[2]);
+    p2 = _mm_loadl_epi64((__m128i *)flat_op[2]);
     work_a = _mm_andnot_si128(flat, work_a);
     p2 = _mm_and_si128(flat, p2);
     p2 = _mm_or_si128(work_a, p2);
-    _mm_store_si128((__m128i *)flat_op[2], p2);
+    _mm_storel_epi64((__m128i *)flat_op[2], p2);
 
-    p1 = _mm_load_si128((__m128i *)flat_op[1]);
+    p1 = _mm_loadl_epi64((__m128i *)flat_op[1]);
     work_a = _mm_andnot_si128(flat, ps1);
     p1 = _mm_and_si128(flat, p1);
     p1 = _mm_or_si128(work_a, p1);
-    _mm_store_si128((__m128i *)flat_op[1], p1);
+    _mm_storel_epi64((__m128i *)flat_op[1], p1);
 
-    p0 = _mm_load_si128((__m128i *)flat_op[0]);
+    p0 = _mm_loadl_epi64((__m128i *)flat_op[0]);
     work_a = _mm_andnot_si128(flat, ps0);
     p0 = _mm_and_si128(flat, p0);
     p0 = _mm_or_si128(work_a, p0);
-    _mm_store_si128((__m128i *)flat_op[0], p0);
+    _mm_storel_epi64((__m128i *)flat_op[0], p0);
 
-    q0 = _mm_load_si128((__m128i *)flat_oq[0]);
+    q0 = _mm_loadl_epi64((__m128i *)flat_oq[0]);
     work_a = _mm_andnot_si128(flat, qs0);
     q0 = _mm_and_si128(flat, q0);
     q0 = _mm_or_si128(work_a, q0);
-    _mm_store_si128((__m128i *)flat_oq[0], q0);
+    _mm_storel_epi64((__m128i *)flat_oq[0], q0);
 
-    q1 = _mm_load_si128((__m128i *)flat_oq[1]);
+    q1 = _mm_loadl_epi64((__m128i *)flat_oq[1]);
     work_a = _mm_andnot_si128(flat, qs1);
     q1 = _mm_and_si128(flat, q1);
     q1 = _mm_or_si128(work_a, q1);
-    _mm_store_si128((__m128i *)flat_oq[1], q1);
+    _mm_storel_epi64((__m128i *)flat_oq[1], q1);
 
-    work_a = _mm_load_si128((__m128i *)aq[2]);
-    q2 = _mm_load_si128((__m128i *)flat_oq[2]);
+    work_a = _mm_loadl_epi64((__m128i *)aq[2]);
+    q2 = _mm_loadl_epi64((__m128i *)flat_oq[2]);
     work_a = _mm_andnot_si128(flat, work_a);
     q2 = _mm_and_si128(flat, q2);
     q2 = _mm_or_si128(work_a, q2);
-    _mm_store_si128((__m128i *)flat_oq[2], q2);
+    _mm_storel_epi64((__m128i *)flat_oq[2], q2);
 
     // write out op6 - op3
     {
       unsigned char *dst = (s - 7 * p);
       for (i = 6; i > 2; i--) {
         __m128i flat2_output;
-        work_a = _mm_load_si128((__m128i *)ap[i]);
-        flat2_output = _mm_load_si128((__m128i *)flat2_op[i]);
+        work_a = _mm_loadl_epi64((__m128i *)ap[i]);
+        flat2_output = _mm_loadl_epi64((__m128i *)flat2_op[i]);
         work_a = _mm_andnot_si128(flat2, work_a);
         flat2_output = _mm_and_si128(flat2, flat2_output);
         work_a = _mm_or_si128(work_a, flat2_output);
-        _mm_storeu_si128((__m128i *)dst, work_a);
+        _mm_storel_epi64((__m128i *)dst, work_a);
         dst += p;
       }
     }
 
-    work_a = _mm_load_si128((__m128i *)flat_op[2]);
-    p2 = _mm_load_si128((__m128i *)flat2_op[2]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_op[2]);
+    p2 = _mm_loadl_epi64((__m128i *)flat2_op[2]);
     work_a = _mm_andnot_si128(flat2, work_a);
     p2 = _mm_and_si128(flat2, p2);
     p2 = _mm_or_si128(work_a, p2);
-    _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+    _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
 
-    work_a = _mm_load_si128((__m128i *)flat_op[1]);
-    p1 = _mm_load_si128((__m128i *)flat2_op[1]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_op[1]);
+    p1 = _mm_loadl_epi64((__m128i *)flat2_op[1]);
     work_a = _mm_andnot_si128(flat2, work_a);
     p1 = _mm_and_si128(flat2, p1);
     p1 = _mm_or_si128(work_a, p1);
-    _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+    _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
 
-    work_a = _mm_load_si128((__m128i *)flat_op[0]);
-    p0 = _mm_load_si128((__m128i *)flat2_op[0]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_op[0]);
+    p0 = _mm_loadl_epi64((__m128i *)flat2_op[0]);
     work_a = _mm_andnot_si128(flat2, work_a);
     p0 = _mm_and_si128(flat2, p0);
     p0 = _mm_or_si128(work_a, p0);
-    _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+    _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
 
-    work_a = _mm_load_si128((__m128i *)flat_oq[0]);
-    q0 = _mm_load_si128((__m128i *)flat2_oq[0]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_oq[0]);
+    q0 = _mm_loadl_epi64((__m128i *)flat2_oq[0]);
     work_a = _mm_andnot_si128(flat2, work_a);
     q0 = _mm_and_si128(flat2, q0);
     q0 = _mm_or_si128(work_a, q0);
-    _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+    _mm_storel_epi64((__m128i *)(s - 0 * p), q0);
 
-    work_a = _mm_load_si128((__m128i *)flat_oq[1]);
-    q1 = _mm_load_si128((__m128i *)flat2_oq[1]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_oq[1]);
+    q1 = _mm_loadl_epi64((__m128i *)flat2_oq[1]);
     work_a = _mm_andnot_si128(flat2, work_a);
     q1 = _mm_and_si128(flat2, q1);
     q1 = _mm_or_si128(work_a, q1);
-    _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+    _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
 
-    work_a = _mm_load_si128((__m128i *)flat_oq[2]);
-    q2 = _mm_load_si128((__m128i *)flat2_oq[2]);
+    work_a = _mm_loadl_epi64((__m128i *)flat_oq[2]);
+    q2 = _mm_loadl_epi64((__m128i *)flat2_oq[2]);
     work_a = _mm_andnot_si128(flat2, work_a);
     q2 = _mm_and_si128(flat2, q2);
     q2 = _mm_or_si128(work_a, q2);
-    _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+    _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
 
     // write out oq3 - oq7
     {
       unsigned char *dst = (s + 3 * p);
       for (i = 3; i < 7; i++) {
         __m128i flat2_output;
-        work_a = _mm_load_si128((__m128i *)aq[i]);
-        flat2_output = _mm_load_si128((__m128i *)flat2_oq[i]);
+        work_a = _mm_loadl_epi64((__m128i *)aq[i]);
+        flat2_output = _mm_loadl_epi64((__m128i *)flat2_oq[i]);
         work_a = _mm_andnot_si128(flat2, work_a);
         flat2_output = _mm_and_si128(flat2, flat2_output);
         work_a = _mm_or_si128(work_a, flat2_output);
-        _mm_storeu_si128((__m128i *)dst, work_a);
+        _mm_storel_epi64((__m128i *)dst, work_a);
         dst += p;
       }
     }
@@ -967,9 +960,14 @@ void vp9_mb_lpf_vertical_edge_w_sse2(unsigned char *s,
   unsigned char *src[4];
   unsigned char *dst[4];
 
+  dst[0] = t_dst;
+  dst[1] = t_dst + 8 * 16;
+
+  src[0] = s - 8;
+  src[1] = s - 8 + 8;
+
   /* Transpose 16x16 */
-  transpose8x16(s - 8, s - 8 + p * 8, p, t_dst, 16);
-  transpose8x16(s, s + p * 8, p, t_dst + 16 * 8, 16);
+  transpose(src, p, dst, 16, 2);
 
   /* Loop filtering */
   vp9_mb_lpf_horizontal_edge_w_sse2(t_dst + 8 * 16, 16, blimit, limit,
@@ -977,16 +975,11 @@ void vp9_mb_lpf_vertical_edge_w_sse2(unsigned char *s,
 
   src[0] = t_dst;
   src[1] = t_dst + 8 * 16;
-  src[2] = t_dst + 8;
-  src[3] = t_dst + 8 * 16 + 8;
 
   dst[0] = s - 8;
   dst[1] = s - 8 + 8;
-  dst[2] = s - 8 + p * 8;
-  dst[3] = s - 8 + p * 8 + 8;
 
-  /* Transpose 16x16 */
-  transpose(src, 16, dst, p, 4);
+  transpose(src, 16, dst, p, 2);
 }