Merge "Intrinsic version of loopfilter now matches C code" into experimental

2 files changed, 298 insertions, 308 deletions

diff --git a/vp9/common/vp9_loopfilter.c b/vp9/common/vp9_loopfilter.c
index 418210408..9ce5a6378 100644
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -257,19 +257,19 @@ void vp9_loop_filter_frame(VP9_COMMON *cm,
                  tx_size >= TX_32X32))
               ) {
             if (tx_size >= TX_16X16)
-              vp9_lpf_mbv_w_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+              vp9_lpf_mbv_w(y_ptr, u_ptr, v_ptr, post->y_stride,
                             post->uv_stride, &lfi);
             else
-              vp9_loop_filter_mbv_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+              vp9_loop_filter_mbv(y_ptr, u_ptr, v_ptr, post->y_stride,
                                   post->uv_stride, &lfi);
           }
           if (!skip_lf) {
             if (tx_size >= TX_8X8) {
               if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
-                vp9_loop_filter_bv8x8_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+                vp9_loop_filter_bv8x8(y_ptr, u_ptr, v_ptr, post->y_stride,
                                       post->uv_stride, &lfi);
               else
-                vp9_loop_filter_bv8x8_c(y_ptr, NULL, NULL, post->y_stride,
+                vp9_loop_filter_bv8x8(y_ptr, NULL, NULL, post->y_stride,
                                       post->uv_stride, &lfi);
             } else {
               vp9_loop_filter_bv(y_ptr, u_ptr, v_ptr, post->y_stride,
@@ -283,19 +283,19 @@ void vp9_loop_filter_frame(VP9_COMMON *cm,
                 tx_size >= TX_32X32))
               ) {
             if (tx_size >= TX_16X16)
-              vp9_lpf_mbh_w_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+              vp9_lpf_mbh_w(y_ptr, u_ptr, v_ptr, post->y_stride,
                             post->uv_stride, &lfi);
             else
-              vp9_loop_filter_mbh_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+              vp9_loop_filter_mbh(y_ptr, u_ptr, v_ptr, post->y_stride,
                                   post->uv_stride, &lfi);
           }
           if (!skip_lf) {
             if (tx_size >= TX_8X8) {
               if (tx_size == TX_8X8 && (mode == I8X8_PRED || mode == SPLITMV))
-                vp9_loop_filter_bh8x8_c(y_ptr, u_ptr, v_ptr, post->y_stride,
+                vp9_loop_filter_bh8x8(y_ptr, u_ptr, v_ptr, post->y_stride,
                                       post->uv_stride, &lfi);
               else
-                vp9_loop_filter_bh8x8_c(y_ptr, NULL, NULL, post->y_stride,
+                vp9_loop_filter_bh8x8(y_ptr, NULL, NULL, post->y_stride,
                                       post->uv_stride, &lfi);
             } else {
               vp9_loop_filter_bh(y_ptr, u_ptr, v_ptr, post->y_stride,
diff --git a/vp9/common/x86/vp9_loopfilter_x86.c b/vp9/common/x86/vp9_loopfilter_x86.c
index e73850dd9..70499d94a 100644
--- a/vp9/common/x86/vp9_loopfilter_x86.c
+++ b/vp9/common/x86/vp9_loopfilter_x86.c
@@ -94,14 +94,16 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
   DECLARE_ALIGNED(16, unsigned char, flat2_op[7][16]);
   DECLARE_ALIGNED(16, unsigned char, flat2_oq[7][16]);
 
-  DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
-  DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_op[3][16]);
+  DECLARE_ALIGNED(16, unsigned char, flat_oq[3][16]);
+
+  DECLARE_ALIGNED(16, unsigned char, ap[8][16]);
+  DECLARE_ALIGNED(16, unsigned char, aq[8][16]);
+
+
   __m128i mask, hev, flat, flat2;
   const __m128i zero = _mm_set1_epi16(0);
+  const __m128i one = _mm_set1_epi8(1);
   __m128i p7, p6, p5;
   __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
   __m128i q5, q6, q7;
@@ -126,12 +128,24 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
   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);
+
+
   {
     const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
                                           _mm_subs_epu8(p0, p1));
     const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
                                           _mm_subs_epu8(q0, q1));
-    const __m128i one = _mm_set1_epi8(1);
     const __m128i fe = _mm_set1_epi8(0xfe);
     const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
     __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
@@ -163,246 +177,8 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     mask = _mm_max_epu8(work, mask);
     mask = _mm_subs_epu8(mask, limit);
     mask = _mm_cmpeq_epi8(mask, zero);
-
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
-                                     _mm_subs_epu8(p0, p2)),
-                         _mm_or_si128(_mm_subs_epu8(q2, q0),
-                                      _mm_subs_epu8(q0, q2)));
-    flat = _mm_max_epu8(work, flat);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
-                                     _mm_subs_epu8(p0, p3)),
-                         _mm_or_si128(_mm_subs_epu8(q3, q0),
-                                      _mm_subs_epu8(q0, q3)));
-    flat = _mm_max_epu8(work, flat);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
-                                     _mm_subs_epu8(p0, p4)),
-                         _mm_or_si128(_mm_subs_epu8(q4, q0),
-                                      _mm_subs_epu8(q0, q4)));
-    flat = _mm_max_epu8(work, flat);
-    flat = _mm_subs_epu8(flat, one);
-    flat = _mm_cmpeq_epi8(flat, zero);
-    flat = _mm_and_si128(flat, mask);
   }
 
-  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-  // calculate flat2
-  p4 = _mm_loadu_si128((__m128i *)(s - 8 * p));
-  p3 = _mm_loadu_si128((__m128i *)(s - 7 * p));
-  p2 = _mm_loadu_si128((__m128i *)(s - 6 * p));
-  p1 = _mm_loadu_si128((__m128i *)(s - 5 * p));
-//  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
-//  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
-  q1 = _mm_loadu_si128((__m128i *)(s + 4 * p));
-  q2 = _mm_loadu_si128((__m128i *)(s + 5 * p));
-  q3 = _mm_loadu_si128((__m128i *)(s + 6 * p));
-  q4 = _mm_loadu_si128((__m128i *)(s + 7 * p));
-
-  {
-    const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
-                                          _mm_subs_epu8(p0, p1));
-    const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
-                                          _mm_subs_epu8(q0, q1));
-    const __m128i one = _mm_set1_epi8(1);
-    __m128i work;
-    flat2 = _mm_max_epu8(abs_p1p0, abs_q1q0);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
-                                     _mm_subs_epu8(p0, p2)),
-                         _mm_or_si128(_mm_subs_epu8(q2, q0),
-                                      _mm_subs_epu8(q0, q2)));
-    flat2 = _mm_max_epu8(work, flat2);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
-                                     _mm_subs_epu8(p0, p3)),
-                         _mm_or_si128(_mm_subs_epu8(q3, q0),
-                                      _mm_subs_epu8(q0, q3)));
-    flat2 = _mm_max_epu8(work, flat2);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
-                                     _mm_subs_epu8(p0, p4)),
-                         _mm_or_si128(_mm_subs_epu8(q4, q0),
-                                      _mm_subs_epu8(q0, q4)));
-    flat2 = _mm_max_epu8(work, flat2);
-    flat2 = _mm_subs_epu8(flat2, one);
-    flat2 = _mm_cmpeq_epi8(flat2, zero);
-    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
-  }
-  // calculate flat2
-  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-  {
-    const __m128i four = _mm_set1_epi16(4);
-    unsigned char *src = s;
-    i = 0;
-    do {
-      __m128i workp_a, workp_b, workp_shft;
-      p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 5 * p)), zero);
-      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
-      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
-      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
-      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
-      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
-      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
-      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
-      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
-      q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 4 * p)), zero);
-
-      workp_a = _mm_add_epi16(_mm_add_epi16(p4, p3), _mm_add_epi16(p2, p1));
-      workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
-      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p4);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_op2[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_op1[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p4), q2);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_op0[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_oq0[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q4);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_oq1[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q4);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
-      _mm_storel_epi64((__m128i *)&flat_oq2[i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      src += 8;
-    } while (++i < 2);
-  }
-  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-  // wide flat
-  // TODO(slavarnway): interleave with the flat pixel calculations (see above)
-  {
-    const __m128i eight = _mm_set1_epi16(8);
-    unsigned char *src = s;
-    int i = 0;
-    do {
-      __m128i workp_a, workp_b, workp_shft;
-      p7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 8 * p)), zero);
-      p6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 7 * p)), zero);
-      p5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 6 * p)), zero);
-      p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 5 * p)), zero);
-      p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
-      p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
-      p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
-      p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
-      q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
-      q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
-      q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
-      q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
-      q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 4 * p)), zero);
-      q5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 5 * p)), zero);
-      q6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 6 * p)), zero);
-      q7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 7 * p)), zero);
-
-
-      workp_a = _mm_sub_epi16(_mm_slli_epi16(p7, 3), p7);  // p7 * 7
-      workp_a = _mm_add_epi16(_mm_slli_epi16(p6, 1), workp_a);
-      workp_b = _mm_add_epi16(_mm_add_epi16(p5, p4), _mm_add_epi16(p3, p2));
-      workp_a = _mm_add_epi16(_mm_add_epi16(p1, p0), workp_a);
-      workp_b = _mm_add_epi16(_mm_add_epi16(q0, eight), workp_b);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[6][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p5);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p6), q1);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[5][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p4);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p5), q2);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[4][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p3);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p4), q3);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[3][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p2);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p3), q4);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[2][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p1);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p2), q5);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[1][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), p0);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), q6);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_op[0][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p7), q0);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[0][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p6), q1);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[1][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p5), q2);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[2][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p4), q3);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q2), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[3][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q4);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q3), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[4][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q5);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q4), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[5][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q6);
-      workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q5), q7);
-      workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 4);
-      _mm_storel_epi64((__m128i *)&flat2_oq[6][i*8],
-                       _mm_packus_epi16(workp_shft, workp_shft));
-
-      src += 8;
-    } while (++i < 2);
-  }
-  // wide flat
-  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
   // lp filter
   {
     const __m128i t4 = _mm_set1_epi8(4);
@@ -413,14 +189,10 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     const __m128i t1 = _mm_set1_epi8(0x1);
     const __m128i t7f = _mm_set1_epi8(0x7f);
 
-    __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
-                                      t80);
-    __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
-                                      t80);
-    __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
-                                      t80);
-    __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
-                                      t80);
+    __m128i ps1 = _mm_xor_si128(p1, t80);
+    __m128i ps0 = _mm_xor_si128(p0, t80);
+    __m128i qs0 = _mm_xor_si128(q0, t80);
+    __m128i qs1 = _mm_xor_si128(q1, t80);
     __m128i filt;
     __m128i work_a;
     __m128i filter1, filter2;
@@ -442,6 +214,7 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     work_a = _mm_and_si128(work_a, te0);
     filter1 = _mm_and_si128(filter1, t1f);
     filter1 = _mm_or_si128(filter1, work_a);
+    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
 
     /* Filter2 >> 3 */
     work_a = _mm_cmpgt_epi8(zero, filter2);
@@ -449,6 +222,7 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     work_a = _mm_and_si128(work_a, te0);
     filter2 = _mm_and_si128(filter2, t1f);
     filter2 = _mm_or_si128(filter2, work_a);
+    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
 
     /* filt >> 1 */
     filt = _mm_adds_epi8(filter1, t1);
@@ -457,20 +231,265 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
     work_a = _mm_and_si128(work_a, t80);
     filt = _mm_and_si128(filt, t7f);
     filt = _mm_or_si128(filt, work_a);
-
     filt = _mm_andnot_si128(hev, filt);
-
-    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
     ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
-    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
     qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+    // loopfilter done
+
+    {
+      __m128i work;
+      work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
+                                       _mm_subs_epu8(p0, p2)),
+                           _mm_or_si128(_mm_subs_epu8(q2, q0),
+                                        _mm_subs_epu8(q0, q2)));
+      flat = _mm_max_epu8(work, flat);
+      work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
+                                       _mm_subs_epu8(p0, p3)),
+                           _mm_or_si128(_mm_subs_epu8(q3, q0),
+                                        _mm_subs_epu8(q0, q3)));
+      flat = _mm_max_epu8(work, flat);
+      work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
+                                       _mm_subs_epu8(p0, p4)),
+                           _mm_or_si128(_mm_subs_epu8(q4, q0),
+                                        _mm_subs_epu8(q0, q4)));
+      flat = _mm_subs_epu8(flat, one);
+      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));
+      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);
+      flat2 = _mm_max_epu8(work, flat2);
+      p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+      q6 = _mm_loadu_si128((__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);
+      flat2 = _mm_max_epu8(work, flat2);
+
+      p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+      q7 = _mm_loadu_si128((__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);
+      flat2 = _mm_max_epu8(work, flat2);
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    }
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    {
+      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);
+
+        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));
+
+        b = _mm_add_epi16(_mm_add_epi16(p3, four), _mm_add_epi16(p3, p2));
+        a = _mm_add_epi16(p3, _mm_add_epi16(p2, p1));
+        a = _mm_add_epi16(_mm_add_epi16(p0, q0), a);
+
+        _mm_storel_epi64((__m128i *)&flat_op[2][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+
+        c = _mm_add_epi16(_mm_add_epi16(p5, eight), c);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[6][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q1, a);
+        b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p2)), p1);
+        _mm_storel_epi64((__m128i *)&flat_op[1][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p6)), p5);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[5][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q2, a);
+        b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p1)), p0);
+        _mm_storel_epi64((__m128i *)&flat_op[0][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p5)), p4);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[4][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q3, a);
+        b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p0)), q0);
+        _mm_storel_epi64((__m128i *)&flat_oq[0][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p4)), p3);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[3][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        b = _mm_add_epi16(q3, b);
+        b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p2, q0)), q1);
+        _mm_storel_epi64((__m128i *)&flat_oq[1][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+
+        c = _mm_add_epi16(q4, c);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p3)), p2);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[2][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        b = _mm_add_epi16(q3, b);
+        b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p1, q1)), q2);
+        _mm_storel_epi64((__m128i *)&flat_oq[2][i*8],
+                         _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
+                                          , b));
+        a = _mm_add_epi16(q5, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p2)), p1);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[1][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q6, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p1)), p0);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_op[0][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p0)), q0);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[0][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p6, q0)), q1);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[1][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p5, q1)), q2);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[2][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p4, q2)), q3);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[3][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p3, q3)), q4);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[4][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p2, q4)), q5);
+        workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
+        _mm_storel_epi64((__m128i *)&flat2_oq[5][i*8],
+                         _mm_packus_epi16(workp_shft, workp_shft));
+
+        a = _mm_add_epi16(q7, a);
+        c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p1, q5)), q6);
+        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_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);
+
+    p1 = _mm_load_si128((__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);
+
+    p0 = _mm_load_si128((__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);
+
+    q0 = _mm_load_si128((__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);
+
+    q1 = _mm_load_si128((__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);
+
+    work_a = _mm_load_si128((__m128i *)aq[2]);
+    q2 = _mm_load_si128((__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);
 
     // write out op6 - op3
     {
       unsigned char *dst = (s - 7 * p);
       for (i = 6; i > 2; i--) {
         __m128i flat2_output;
-        work_a = _mm_loadu_si128((__m128i *)dst);
+        work_a = _mm_load_si128((__m128i *)ap[i]);
         flat2_output = _mm_load_si128((__m128i *)flat2_op[i]);
         work_a = _mm_andnot_si128(flat2, work_a);
         flat2_output = _mm_and_si128(flat2, flat2_output);
@@ -480,62 +499,42 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
       }
     }
 
-    work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
-    p2 = _mm_load_si128((__m128i *)flat_op2);
-    work_a = _mm_andnot_si128(flat, work_a);
-    p2 = _mm_and_si128(flat, p2);
-    work_a = _mm_or_si128(work_a, p2);
+    work_a = _mm_load_si128((__m128i *)flat_op[2]);
     p2 = _mm_load_si128((__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);
 
-    p1 = _mm_load_si128((__m128i *)flat_op1);
-    work_a = _mm_andnot_si128(flat, ps1);
-    p1 = _mm_and_si128(flat, p1);
-    work_a = _mm_or_si128(work_a, p1);
+    work_a = _mm_load_si128((__m128i *)flat_op[1]);
     p1 = _mm_load_si128((__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);
 
-    p0 = _mm_load_si128((__m128i *)flat_op0);
-    work_a = _mm_andnot_si128(flat, ps0);
-    p0 = _mm_and_si128(flat, p0);
-    work_a = _mm_or_si128(work_a, p0);
+    work_a = _mm_load_si128((__m128i *)flat_op[0]);
     p0 = _mm_load_si128((__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);
 
-    q0 = _mm_load_si128((__m128i *)flat_oq0);
-    work_a = _mm_andnot_si128(flat, qs0);
-    q0 = _mm_and_si128(flat, q0);
-    work_a = _mm_or_si128(work_a, q0);
+    work_a = _mm_load_si128((__m128i *)flat_oq[0]);
     q0 = _mm_load_si128((__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);
 
-    q1 = _mm_load_si128((__m128i *)flat_oq1);
-    work_a = _mm_andnot_si128(flat, qs1);
-    q1 = _mm_and_si128(flat, q1);
-    work_a = _mm_or_si128(work_a, q1);
+    work_a = _mm_load_si128((__m128i *)flat_oq[1]);
     q1 = _mm_load_si128((__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);
 
-    work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
-    q2 = _mm_load_si128((__m128i *)flat_oq2);
-    work_a = _mm_andnot_si128(flat, work_a);
-    q2 = _mm_and_si128(flat, q2);
-    work_a = _mm_or_si128(work_a, q2);
+    work_a = _mm_load_si128((__m128i *)flat_oq[2]);
     q2 = _mm_load_si128((__m128i *)flat2_oq[2]);
     work_a = _mm_andnot_si128(flat2, work_a);
     q2 = _mm_and_si128(flat2, q2);
@@ -547,7 +546,7 @@ void vp9_mb_lpf_horizontal_edge_w_sse2(unsigned char *s,
       unsigned char *dst = (s + 3 * p);
       for (i = 3; i < 7; i++) {
         __m128i flat2_output;
-        work_a = _mm_loadu_si128((__m128i *)dst);
+        work_a = _mm_load_si128((__m128i *)aq[i]);
         flat2_output = _mm_load_si128((__m128i *)flat2_oq[i]);
         work_a = _mm_andnot_si128(flat2, work_a);
         flat2_output = _mm_and_si128(flat2, flat2_output);
@@ -572,7 +571,7 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
   DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
   __m128i mask, hev, flat;
   const __m128i zero = _mm_set1_epi16(0);
-  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
   const unsigned int extended_thresh = _thresh[0] * 0x01010101u;
   const unsigned int extended_limit  = _limit[0]  * 0x01010101u;
   const unsigned int extended_blimit = _blimit[0] * 0x01010101u;
@@ -583,7 +582,6 @@ void vp9_mbloop_filter_horizontal_edge_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));
@@ -592,7 +590,6 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
   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));
   {
     const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
                                           _mm_subs_epu8(p0, p1));
@@ -641,11 +638,6 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
                          _mm_or_si128(_mm_subs_epu8(q3, q0),
                                       _mm_subs_epu8(q0, q3)));
     flat = _mm_max_epu8(work, flat);
-    work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
-                                     _mm_subs_epu8(p0, p4)),
-                         _mm_or_si128(_mm_subs_epu8(q4, q0),
-                                      _mm_subs_epu8(q0, q4)));
-    flat = _mm_max_epu8(work, flat);
     flat = _mm_subs_epu8(flat, one);
     flat = _mm_cmpeq_epi8(flat, zero);
     flat = _mm_and_si128(flat, mask);
@@ -656,7 +648,6 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
     int i = 0;
     do {
       __m128i workp_a, workp_b, workp_shft;
-      p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 5 * p)), zero);
       p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
       p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
       p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
@@ -665,11 +656,10 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
       q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
       q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
       q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
-      q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 4 * p)), zero);
 
-      workp_a = _mm_add_epi16(_mm_add_epi16(p4, p3), _mm_add_epi16(p2, p1));
+      workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
       workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
-      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p4);
+      workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
       workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
       _mm_storel_epi64((__m128i *)&flat_op2[i*8],
                        _mm_packus_epi16(workp_shft, workp_shft));
@@ -679,7 +669,7 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
       _mm_storel_epi64((__m128i *)&flat_op1[i*8],
                        _mm_packus_epi16(workp_shft, workp_shft));
 
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p4), q2);
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
       workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
       workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
       _mm_storel_epi64((__m128i *)&flat_op0[i*8],
@@ -691,13 +681,13 @@ void vp9_mbloop_filter_horizontal_edge_sse2(unsigned char *s,
       _mm_storel_epi64((__m128i *)&flat_oq0[i*8],
                        _mm_packus_epi16(workp_shft, workp_shft));
 
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q4);
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
       workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
       workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
       _mm_storel_epi64((__m128i *)&flat_oq1[i*8],
                        _mm_packus_epi16(workp_shft, workp_shft));
 
-      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q4);
+      workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
       workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
       workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
       _mm_storel_epi64((__m128i *)&flat_oq2[i*8],