shithub: libvpx

Download patch

ref: 469643757f3b3616acd0157b670a910fc5b78577
parent: e40e78be246b2521749ce1fa6bdcd02e1f729a37
author: Johann <johannkoenig@google.com>
date: Tue Jun 27 10:15:58 EDT 2017

sad neon: rewrite 16x8, 16x16, add 16x32

BUG=webm:1425

Change-Id: Ie126553e5fffcdfaf3d82a85b368ac10ce9ab082

--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -645,6 +645,7 @@
 const SadMxNParam neon_tests[] = {
   SadMxNParam(64, 64, &vpx_sad64x64_neon),
   SadMxNParam(32, 32, &vpx_sad32x32_neon),
+  SadMxNParam(16, 32, &vpx_sad16x32_neon),
   SadMxNParam(16, 16, &vpx_sad16x16_neon),
   SadMxNParam(16, 8, &vpx_sad16x8_neon),
   SadMxNParam(8, 16, &vpx_sad8x16_neon),
--- a/vpx_dsp/arm/sad_neon.c
+++ b/vpx_dsp/arm/sad_neon.c
@@ -82,40 +82,41 @@
   return horizontal_add_16x8(abs);
 }
 
-unsigned int vpx_sad16x8_neon(unsigned char *src_ptr, int src_stride,
-                              unsigned char *ref_ptr, int ref_stride) {
-  uint8x16_t q0, q4;
-  uint16x8_t q12, q13;
-  uint32x4_t q1;
-  uint64x2_t q3;
-  uint32x2_t d5;
+static INLINE uint16x8_t sad16x(const uint8_t *a, int a_stride,
+                                const uint8_t *b, int b_stride,
+                                const int height) {
   int i;
+  uint16x8_t abs = vdupq_n_u16(0);
 
-  q0 = vld1q_u8(src_ptr);
-  src_ptr += src_stride;
-  q4 = vld1q_u8(ref_ptr);
-  ref_ptr += ref_stride;
-  q12 = vabdl_u8(vget_low_u8(q0), vget_low_u8(q4));
-  q13 = vabdl_u8(vget_high_u8(q0), vget_high_u8(q4));
-
-  for (i = 0; i < 7; i++) {
-    q0 = vld1q_u8(src_ptr);
-    src_ptr += src_stride;
-    q4 = vld1q_u8(ref_ptr);
-    ref_ptr += ref_stride;
-    q12 = vabal_u8(q12, vget_low_u8(q0), vget_low_u8(q4));
-    q13 = vabal_u8(q13, vget_high_u8(q0), vget_high_u8(q4));
+  for (i = 0; i < height; ++i) {
+    const uint8x16_t a_u8 = vld1q_u8(a);
+    const uint8x16_t b_u8 = vld1q_u8(b);
+    a += a_stride;
+    b += b_stride;
+    abs = vabal_u8(abs, vget_low_u8(a_u8), vget_low_u8(b_u8));
+    abs = vabal_u8(abs, vget_high_u8(a_u8), vget_high_u8(b_u8));
   }
+  return abs;
+}
 
-  q12 = vaddq_u16(q12, q13);
-  q1 = vpaddlq_u16(q12);
-  q3 = vpaddlq_u32(q1);
-  d5 = vadd_u32(vreinterpret_u32_u64(vget_low_u64(q3)),
-                vreinterpret_u32_u64(vget_high_u64(q3)));
+uint32_t vpx_sad16x8_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *ref, int ref_stride) {
+  const uint16x8_t abs = sad16x(src, src_stride, ref, ref_stride, 8);
+  return horizontal_add_16x8(abs);
+}
 
-  return vget_lane_u32(d5, 0);
+uint32_t vpx_sad16x16_neon(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride) {
+  const uint16x8_t abs = sad16x(src, src_stride, ref, ref_stride, 16);
+  return horizontal_add_16x8(abs);
 }
 
+uint32_t vpx_sad16x32_neon(const uint8_t *src, int src_stride,
+                           const uint8_t *ref, int ref_stride) {
+  const uint16x8_t abs = sad16x(src, src_stride, ref, ref_stride, 32);
+  return horizontal_add_16x8(abs);
+}
+
 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                                     const uint16x8_t vec_hi) {
   const uint32x4_t vec_l_lo =
@@ -186,25 +187,6 @@
                             vget_low_u8(vec_ref_16));
     vec_accum_hi = vabal_u8(vec_accum_hi, vget_high_u8(vec_src_16),
                             vget_high_u8(vec_ref_16));
-  }
-  return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
-}
-
-unsigned int vpx_sad16x16_neon(const uint8_t *src, int src_stride,
-                               const uint8_t *ref, int ref_stride) {
-  int i;
-  uint16x8_t vec_accum_lo = vdupq_n_u16(0);
-  uint16x8_t vec_accum_hi = vdupq_n_u16(0);
-
-  for (i = 0; i < 16; ++i) {
-    const uint8x16_t vec_src = vld1q_u8(src);
-    const uint8x16_t vec_ref = vld1q_u8(ref);
-    src += src_stride;
-    ref += ref_stride;
-    vec_accum_lo =
-        vabal_u8(vec_accum_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref));
-    vec_accum_hi =
-        vabal_u8(vec_accum_hi, vget_high_u8(vec_src), vget_high_u8(vec_ref));
   }
   return horizontal_add_16x8(vaddq_u16(vec_accum_lo, vec_accum_hi));
 }
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -711,7 +711,7 @@
 specialize qw/vpx_sad32x16 avx2 msa sse2 vsx/;
 
 add_proto qw/unsigned int vpx_sad16x32/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
-specialize qw/vpx_sad16x32 msa sse2 vsx/;
+specialize qw/vpx_sad16x32 neon msa sse2 vsx/;
 
 add_proto qw/unsigned int vpx_sad16x16/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride";
 specialize qw/vpx_sad16x16 neon msa sse2 vsx/;