vp9: neon: use aligned stores in convolve functions

The destination is block-aligned so it is safe to use aligned
stores.

Change-Id: I38261e4fa40bc60e6472edffece59e372908da7e

3 files changed, 23 insertions, 22 deletions

diff --git a/vp9/common/arm/neon/vp9_convolve8_avg_neon.asm b/vp9/common/arm/neon/vp9_convolve8_avg_neon.asm
index 437295245..6b20cb9bf 100644
--- a/vp9/common/arm/neon/vp9_convolve8_avg_neon.asm
+++ b/vp9/common/arm/neon/vp9_convolve8_avg_neon.asm
@@ -159,10 +159,10 @@ loop_horiz
     ; average the new value and the dst value
     vrhadd.u8       q1, q1, q3
 
-    vst1.u32        {d2[0]}, [r2], r3
-    vst1.u32        {d3[0]}, [r2], r3
-    vst1.u32        {d2[1]}, [r2], r3
-    vst1.u32        {d3[1]}, [r2], r4
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
 
     vmov            q8,  q9
     vmov            d20, d23
@@ -234,10 +234,10 @@ loop_vert
     vmovl.u8        q12, d24
     vmovl.u8        q13, d26
 
-    vld1.u32        {d6[0]}, [r5], r3
-    vld1.u32        {d6[1]}, [r8], r3
-    vld1.u32        {d7[0]}, [r5], r3
-    vld1.u32        {d7[1]}, [r8], r3
+    vld1.u32        {d6[0]}, [r5@32], r3
+    vld1.u32        {d6[1]}, [r8@32], r3
+    vld1.u32        {d7[0]}, [r5@32], r3
+    vld1.u32        {d7[1]}, [r8@32], r3
 
     pld             [r7]
     pld             [r4]
@@ -276,10 +276,10 @@ loop_vert
     sub             r5, r5, r3, lsl #1      ; reset for store
     sub             r8, r8, r3, lsl #1
 
-    vst1.u32        {d2[0]}, [r5], r3
-    vst1.u32        {d2[1]}, [r8], r3
-    vst1.u32        {d3[0]}, [r5], r3
-    vst1.u32        {d3[1]}, [r8], r3
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
 
     vmov            q8, q10
     vmov            d18, d22
diff --git a/vp9/common/arm/neon/vp9_convolve8_neon.asm b/vp9/common/arm/neon/vp9_convolve8_neon.asm
index 667cddb36..45258454c 100644
--- a/vp9/common/arm/neon/vp9_convolve8_neon.asm
+++ b/vp9/common/arm/neon/vp9_convolve8_neon.asm
@@ -148,10 +148,10 @@ loop_horiz
     vtrn.32         d2, d3
     vtrn.8          d2, d3
 
-    vst1.u32        {d2[0]}, [r2], r3
-    vst1.u32        {d3[0]}, [r2], r3
-    vst1.u32        {d2[1]}, [r2], r3
-    vst1.u32        {d3[1]}, [r2], r4
+    vst1.u32        {d2[0]}, [r2@32], r3
+    vst1.u32        {d3[0]}, [r2@32], r3
+    vst1.u32        {d2[1]}, [r2@32], r3
+    vst1.u32        {d3[1]}, [r2@32], r4
 
     vmov            q8,  q9
     vmov            d20, d23
@@ -254,10 +254,10 @@ loop_vert
     vqmovn.u16      d2, q1
     vqmovn.u16      d3, q2
 
-    vst1.u32        {d2[0]}, [r5], r3
-    vst1.u32        {d2[1]}, [r8], r3
-    vst1.u32        {d3[0]}, [r5], r3
-    vst1.u32        {d3[1]}, [r8], r3
+    vst1.u32        {d2[0]}, [r5@32], r3
+    vst1.u32        {d2[1]}, [r8@32], r3
+    vst1.u32        {d3[0]}, [r5@32], r3
+    vst1.u32        {d3[1]}, [r8@32], r3
 
     vmov            q8, q10
     vmov            d18, d22
diff --git a/vp9/common/arm/neon/vp9_convolve_neon.c b/vp9/common/arm/neon/vp9_convolve_neon.c
index 6e37ff6f9..d8b24bfaf 100644
--- a/vp9/common/arm/neon/vp9_convolve_neon.c
+++ b/vp9/common/arm/neon/vp9_convolve_neon.c
@@ -10,6 +10,7 @@
 
 #include "./vp9_rtcd.h"
 #include "vp9/common/vp9_common.h"
+#include "vpx_ports/mem.h"
 
 void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
                         uint8_t *dst, ptrdiff_t dst_stride,
@@ -19,7 +20,7 @@ void vp9_convolve8_neon(const uint8_t *src, ptrdiff_t src_stride,
   /* Given our constraints: w <= 64, h <= 64, taps == 8 we can reduce the
    * maximum buffer size to 64 * 64 + 7 (+ 1 to make it divisible by 4).
    */
-  uint8_t temp[64 * 72];
+  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
 
   // Account for the vertical phase needing 3 lines prior and 4 lines post
   int intermediate_height = h + 7;
@@ -53,7 +54,7 @@ void vp9_convolve8_avg_neon(const uint8_t *src, ptrdiff_t src_stride,
                             const int16_t *filter_x, int x_step_q4,
                             const int16_t *filter_y, int y_step_q4,
                             int w, int h) {
-  uint8_t temp[64 * 72];
+  DECLARE_ALIGNED_ARRAY(8, uint8_t, temp, 64 * 72);
   int intermediate_height = h + 7;
 
   if (x_step_q4 != 16 || y_step_q4 != 16)