shithub: libvpx

Info  •  Files  •  Log  •  Branches

Switch remaining Neon variance functions to shifts

Saves 5 instructions on 8x8 and 16x16 and 8 instructions
on 32x32, when compiled with 4.9.

Change-Id: Id3da613a36a9d27d8c5169c59ba45d247c920c6c

--- a/vp9/encoder/arm/neon/vp9_variance_neon.c

+++ b/vp9/encoder/arm/neon/vp9_variance_neon.c

@@ -91,7 +91,7 @@

                                   unsigned int *sse) {

   int sum;

   variance_neon_w8(a, a_stride, b, b_stride, kWidth8, kHeight8, sse, &sum);

-  return *sse - (((int64_t)sum * sum) / (kWidth8 * kHeight8));

+  return *sse - (((int64_t)sum * sum) >> 6);  //  >> 6 = / 8 * 8

 }

 void vp9_get16x16var_neon(const uint8_t *src_ptr, int source_stride,

@@ -106,7 +106,7 @@

                                     unsigned int *sse) {

   int sum;

   variance_neon_w8(a, a_stride, b, b_stride, kWidth16, kHeight16, sse, &sum);

-  return *sse - (((int64_t)sum * sum) / (kWidth16 * kHeight16));

+  return *sse - (((int64_t)sum * sum) >> 8);  //  >> 8 = / 16 * 16

 }

 static void var_filter_block2d_bil_w8(const uint8_t *src_ptr,

@@ -208,7 +208,7 @@

                                     unsigned int *sse) {

   int sum;

   variance_neon_w8(a, a_stride, b, b_stride, kWidth32, kHeight32, sse, &sum);

-  return *sse - (((int64_t)sum * sum) / (kWidth32 * kHeight32));

+  return *sse - (((int64_t)sum * sum) >> 10);  // >> 10 = / 32 * 32

 }

 unsigned int vp9_variance32x64_neon(const uint8_t *a, int a_stride,