Use lane-referencing intrinsics in Neon convolution kernels

The Neon convolution helper functions take a pointer to a filter and
load the 8 values into a single Neon register. For some reason,
filter values 3 and 4 are then duplicated into their own separate
registers.

This patch modifies these helper functions so that they access filter
values 3 and 4 via the lane-referencing versions of the various Neon
multiply instructions. This reduces register pressure and tidies up
the source code quite a bit.

Change-Id: Ia4aeee8b46fe218658fb8577dc07ff04a9324b3e

diff --git a/vpx_dsp/arm/vpx_convolve8_neon.c b/vpx_dsp/arm/vpx_convolve8_neon.c
index ca5222fa073acaa9aef3c49a5f799ae52cd16050..28018398a5bb7ee073b72f56e97b4e3355e8663e 100644 (file)
--- a/vpx_dsp/arm/vpx_convolve8_neon.c
+++ b/vpx_dsp/arm/vpx_convolve8_neon.c
@@ -807,16 +807,13 @@ void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
 
   if (h == 4) {
     uint8x8_t d01, d23;
-    int16x4_t filter3, filter4, s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0,
-        d1, d2, d3;
+    int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
     int16x8_t tt0, tt1, tt2, tt3;
 
     __builtin_prefetch(src + 0 * src_stride);
     __builtin_prefetch(src + 1 * src_stride);
     __builtin_prefetch(src + 2 * src_stride);
     __builtin_prefetch(src + 3 * src_stride);
-    filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-    filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
     load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
     transpose_u8_8x4(&t0, &t1, &t2, &t3);
     tt0 = vreinterpretq_s16_u16(vmovl_u8(t0));
@@ -848,14 +845,10 @@ void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
       s9 = vget_low_s16(tt2);
       s10 = vget_low_s16(tt3);
 
-      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                       filter4);
-      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                       filter4);
-      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                       filter4);
-      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                       filter4);
+      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
       d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), 7);
       d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), 7);
@@ -882,8 +875,6 @@ void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
       w -= 4;
     } while (w != 0);
   } else {
-    const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
-    const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
     int width;
     const uint8_t *s;
     uint8x8_t t4, t5, t6, t7;
@@ -926,14 +917,10 @@ void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
         __builtin_prefetch(src + 5 * src_stride);
         __builtin_prefetch(src + 6 * src_stride);
         __builtin_prefetch(src + 7 * src_stride);
-        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                         filter4);
-        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                         filter4);
-        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                         filter4);
-        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                         filter4);
+        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
         transpose_u8_8x4(&t0, &t1, &t2, &t3);
         vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(t0), 0);
@@ -1001,22 +988,14 @@ void vpx_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
           s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
           s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
 
-          t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                           filter4);
-          t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                           filter4);
-          t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                           filter4);
-          t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                           filter4);
-          t4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, filter3,
-                           filter4);
-          t5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, filter3,
-                           filter4);
-          t6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, filter3,
-                           filter4);
-          t7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters,
-                           filter3, filter4);
+          t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+          t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+          t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+          t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+          t4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters);
+          t5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters);
+          t6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters);
+          t7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters);
 
           transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
           store_u8_8x8(d, dst_stride, t0, t1, t2, t3, t4, t5, t6, t7);
@@ -1060,8 +1039,7 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
 
   if (h == 4) {
     uint8x8_t d01, d23;
-    int16x4_t filter3, filter4, s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0,
-        d1, d2, d3;
+    int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
     int16x8_t tt0, tt1, tt2, tt3;
     uint32x4_t d0123 = vdupq_n_u32(0);
 
@@ -1069,8 +1047,6 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
     __builtin_prefetch(src + 1 * src_stride);
     __builtin_prefetch(src + 2 * src_stride);
     __builtin_prefetch(src + 3 * src_stride);
-    filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-    filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
     load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
     transpose_u8_8x4(&t0, &t1, &t2, &t3);
     tt0 = vreinterpretq_s16_u16(vmovl_u8(t0));
@@ -1102,14 +1078,10 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
       s9 = vget_low_s16(tt2);
       s10 = vget_low_s16(tt3);
 
-      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                       filter4);
-      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                       filter4);
-      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                       filter4);
-      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                       filter4);
+      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
       d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), 7);
       d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), 7);
@@ -1139,8 +1111,6 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
       w -= 4;
     } while (w != 0);
   } else {
-    const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
-    const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
     int width;
     const uint8_t *s;
     uint8x8_t t4, t5, t6, t7;
@@ -1185,14 +1155,10 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
         __builtin_prefetch(src + 5 * src_stride);
         __builtin_prefetch(src + 6 * src_stride);
         __builtin_prefetch(src + 7 * src_stride);
-        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                         filter4);
-        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                         filter4);
-        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                         filter4);
-        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                         filter4);
+        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
         transpose_u8_8x4(&t0, &t1, &t2, &t3);
 
@@ -1275,22 +1241,14 @@ void vpx_convolve8_avg_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
           s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
           s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
 
-          t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                           filter4);
-          t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                           filter4);
-          t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                           filter4);
-          t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                           filter4);
-          t4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, filter3,
-                           filter4);
-          t5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, filter3,
-                           filter4);
-          t6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, filter3,
-                           filter4);
-          t7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters,
-                           filter3, filter4);
+          t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+          t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+          t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+          t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+          t4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters);
+          t5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters);
+          t6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters);
+          t7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters);
 
           transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
 
@@ -1348,8 +1306,6 @@ void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
   src -= 3 * src_stride;
 
   if (w == 4) {
-    const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-    const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
     uint8x8_t d01, d23;
     int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
 
@@ -1386,14 +1342,10 @@ void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
       __builtin_prefetch(src + 1 * src_stride);
       __builtin_prefetch(src + 2 * src_stride);
       __builtin_prefetch(src + 3 * src_stride);
-      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                       filter4);
-      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                       filter4);
-      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                       filter4);
-      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                       filter4);
+      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
       d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), 7);
       d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), 7);
@@ -1416,8 +1368,6 @@ void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
       h -= 4;
     } while (h != 0);
   } else {
-    const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
-    const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
     int height;
     const uint8_t *s;
     uint8_t *d;
@@ -1468,14 +1418,10 @@ void vpx_convolve8_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
         __builtin_prefetch(s + 1 * src_stride);
         __builtin_prefetch(s + 2 * src_stride);
         __builtin_prefetch(s + 3 * src_stride);
-        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                         filter4);
-        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                         filter4);
-        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                         filter4);
-        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                         filter4);
+        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
         vst1_u8(d, t0);
         d += dst_stride;
@@ -1520,8 +1466,6 @@ void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
   src -= 3 * src_stride;
 
   if (w == 4) {
-    const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-    const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
     uint8x8_t d01, d23;
     int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
     uint32x4_t d0123 = vdupq_n_u32(0);
@@ -1559,14 +1503,10 @@ void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
       __builtin_prefetch(src + 1 * src_stride);
       __builtin_prefetch(src + 2 * src_stride);
       __builtin_prefetch(src + 3 * src_stride);
-      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                       filter4);
-      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                       filter4);
-      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                       filter4);
-      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                       filter4);
+      d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+      d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+      d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+      d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
       d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), 7);
       d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), 7);
@@ -1597,8 +1537,6 @@ void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
       h -= 4;
     } while (h != 0);
   } else {
-    const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
-    const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
     int height;
     const uint8_t *s;
     uint8_t *d;
@@ -1650,14 +1588,10 @@ void vpx_convolve8_avg_vert_neon(const uint8_t *src, ptrdiff_t src_stride,
         __builtin_prefetch(s + 1 * src_stride);
         __builtin_prefetch(s + 2 * src_stride);
         __builtin_prefetch(s + 3 * src_stride);
-        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, filter3,
-                         filter4);
-        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, filter3,
-                         filter4);
-        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, filter3,
-                         filter4);
-        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, filter3,
-                         filter4);
+        t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+        t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+        t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+        t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
 
         d01 = vcombine_u8(t0, t1);
         d23 = vcombine_u8(t2, t3);

diff --git a/vpx_dsp/arm/vpx_convolve8_neon.h b/vpx_dsp/arm/vpx_convolve8_neon.h
index b112cb249a3e321cca792099337dcfc270b8b63f..b8dfce71ea1e955d9cc384af318c1229a31bf4a3 100644 (file)
--- a/vpx_dsp/arm/vpx_convolve8_neon.h
+++ b/vpx_dsp/arm/vpx_convolve8_neon.h
@@ -176,9 +176,7 @@ static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
                                     const int16x4_t s2, const int16x4_t s3,
                                     const int16x4_t s4, const int16x4_t s5,
                                     const int16x4_t s6, const int16x4_t s7,
-                                    const int16x8_t filters,
-                                    const int16x4_t filter3,
-                                    const int16x4_t filter4) {
+                                    const int16x8_t filters) {
   const int16x4_t filters_lo = vget_low_s16(filters);
   const int16x4_t filters_hi = vget_high_s16(filters);
   int16x4_t sum;
@@ -189,8 +187,8 @@ static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
   sum = vmla_lane_s16(sum, s5, filters_hi, 1);
   sum = vmla_lane_s16(sum, s6, filters_hi, 2);
   sum = vmla_lane_s16(sum, s7, filters_hi, 3);
-  sum = vqadd_s16(sum, vmul_s16(s3, filter3));
-  sum = vqadd_s16(sum, vmul_s16(s4, filter4));
+  sum = vqadd_s16(sum, vmul_lane_s16(s3, filters_lo, 3));
+  sum = vqadd_s16(sum, vmul_lane_s16(s4, filters_hi, 0));
   return sum;
 }
 
@@ -198,9 +196,7 @@ static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
                                     const int16x8_t s2, const int16x8_t s3,
                                     const int16x8_t s4, const int16x8_t s5,
                                     const int16x8_t s6, const int16x8_t s7,
-                                    const int16x8_t filters,
-                                    const int16x8_t filter3,
-                                    const int16x8_t filter4) {
+                                    const int16x8_t filters) {
   const int16x4_t filters_lo = vget_low_s16(filters);
   const int16x4_t filters_hi = vget_high_s16(filters);
   int16x8_t sum;
@@ -211,15 +207,13 @@ static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
   sum = vmlaq_lane_s16(sum, s5, filters_hi, 1);
   sum = vmlaq_lane_s16(sum, s6, filters_hi, 2);
   sum = vmlaq_lane_s16(sum, s7, filters_hi, 3);
-  sum = vqaddq_s16(sum, vmulq_s16(s3, filter3));
-  sum = vqaddq_s16(sum, vmulq_s16(s4, filter4));
+  sum = vqaddq_s16(sum, vmulq_lane_s16(s3, filters_lo, 3));
+  sum = vqaddq_s16(sum, vmulq_lane_s16(s4, filters_hi, 0));
   return vqrshrun_n_s16(sum, 7);
 }
 
 static INLINE uint8x8_t scale_filter_8(const uint8x8_t *const s,
                                        const int16x8_t filters) {
-  const int16x8_t filter3 = vdupq_lane_s16(vget_low_s16(filters), 3);
-  const int16x8_t filter4 = vdupq_lane_s16(vget_high_s16(filters), 0);
   int16x8_t ss[8];
 
   ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
@@ -232,7 +226,7 @@ static INLINE uint8x8_t scale_filter_8(const uint8x8_t *const s,
   ss[7] = vreinterpretq_s16_u16(vmovl_u8(s[7]));
 
   return convolve8_8(ss[0], ss[1], ss[2], ss[3], ss[4], ss[5], ss[6], ss[7],
-                     filters, filter3, filter4);
+                     filters);
 }
 
 #endif  // VPX_VPX_DSP_ARM_VPX_CONVOLVE8_NEON_H_

diff --git a/vpx_dsp/arm/vpx_scaled_convolve8_neon.c b/vpx_dsp/arm/vpx_scaled_convolve8_neon.c
index 8edf8a66e6efbfa7d664d37aacaf2d7cb2a65cc1..8491ca7ac5e9d69b23394cd4ec65a1f1bc134a4f 100644 (file)
--- a/vpx_dsp/arm/vpx_scaled_convolve8_neon.c
+++ b/vpx_dsp/arm/vpx_scaled_convolve8_neon.c
@@ -38,8 +38,6 @@ static INLINE void scaledconvolve_horiz_w4(
         const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
         if (x_q4 & SUBPEL_MASK) {
           const int16x8_t filters = vld1q_s16(x_filters[x_q4 & SUBPEL_MASK]);
-          const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-          const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
           uint8x8_t s[8], d;
           int16x8_t ss[4];
           int16x4_t t[8], tt;
@@ -61,7 +59,7 @@ static INLINE void scaledconvolve_horiz_w4(
           t[7] = vget_high_s16(ss[3]);
 
           tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7],
-                           filters, filter3, filter4);
+                           filters);
           d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
           vst1_lane_u32((uint32_t *)&temp[4 * z], vreinterpret_u32_u8(d), 0);
         } else {
@@ -167,8 +165,6 @@ static INLINE void scaledconvolve_vert_w4(
 
     if (y_q4 & SUBPEL_MASK) {
       const int16x8_t filters = vld1q_s16(y_filters[y_q4 & SUBPEL_MASK]);
-      const int16x4_t filter3 = vdup_lane_s16(vget_low_s16(filters), 3);
-      const int16x4_t filter4 = vdup_lane_s16(vget_high_s16(filters), 0);
       uint8x8_t s[8], d;
       int16x4_t t[8], tt;
 
@@ -183,8 +179,7 @@ static INLINE void scaledconvolve_vert_w4(
       t[6] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[6])));
       t[7] = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(s[7])));
 
-      tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7], filters,
-                       filter3, filter4);
+      tt = convolve8_4(t[0], t[1], t[2], t[3], t[4], t[5], t[6], t[7], filters);
       d = vqrshrun_n_s16(vcombine_s16(tt, tt), 7);
       vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
     } else {