Merge changes Id84d9780,Iaa6ea75b,I3362e0dd,I0020a49e,Ia42e4f36, ...

* changes:
  sad neon: avg for 64x[32,64]
  sad neon: macroize 64xN definitions
  sad neon: avg for 32x[16,32,64]
  sad neon: macroize 32xN definitions
  sad neon: avg for 16x[8,16,32]
  sad neon: macroize 16xN definitions

1 files changed, 134 insertions, 37 deletions

diff --git a/vpx_dsp/arm/sad_neon.c b/vpx_dsp/arm/sad_neon.c
index baaaf6378..e651f7e66 100644
--- a/vpx_dsp/arm/sad_neon.c
+++ b/vpx_dsp/arm/sad_neon.c
@@ -151,23 +151,44 @@ static INLINE uint16x8_t sad16x(const uint8_t *a, int a_stride,
   return abs;
 }
 
-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);
-}
+static INLINE uint16x8_t sad16x_avg(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    const uint8_t *c, const int height) {
+  int i;
+  uint16x8_t abs = vdupq_n_u16(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);
+  for (i = 0; i < height; ++i) {
+    const uint8x16_t a_u8 = vld1q_u8(a);
+    const uint8x16_t b_u8 = vld1q_u8(b);
+    const uint8x16_t c_u8 = vld1q_u8(c);
+    const uint8x16_t avg = vrhaddq_u8(b_u8, c_u8);
+    a += a_stride;
+    b += b_stride;
+    c += 16;
+    abs = vabal_u8(abs, vget_low_u8(a_u8), vget_low_u8(avg));
+    abs = vabal_u8(abs, vget_high_u8(a_u8), vget_high_u8(avg));
+  }
+  return 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);
-}
+#define sad16xN(n)                                                      \
+  uint32_t vpx_sad16x##n##_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, n); \
+    return horizontal_add_16x8(abs);                                    \
+  }                                                                     \
+                                                                        \
+  uint32_t vpx_sad16x##n##_avg_neon(const uint8_t *src, int src_stride, \
+                                    const uint8_t *ref, int ref_stride, \
+                                    const uint8_t *second_pred) {       \
+    const uint16x8_t abs =                                              \
+        sad16x_avg(src, src_stride, ref, ref_stride, second_pred, n);   \
+    return horizontal_add_16x8(abs);                                    \
+  }
+
+sad16xN(8);
+sad16xN(16);
+sad16xN(32);
 
 static INLINE uint16x8_t sad32x(const uint8_t *a, int a_stride,
                                 const uint8_t *b, int b_stride,
@@ -190,23 +211,50 @@ static INLINE uint16x8_t sad32x(const uint8_t *a, int a_stride,
   return abs;
 }
 
-uint32_t vpx_sad32x16_neon(const uint8_t *src, int src_stride,
-                           const uint8_t *ref, int ref_stride) {
-  const uint16x8_t abs = sad32x(src, src_stride, ref, ref_stride, 16);
-  return horizontal_add_16x8(abs);
-}
+static INLINE uint16x8_t sad32x_avg(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    const uint8_t *c, const int height) {
+  int i;
+  uint16x8_t abs = vdupq_n_u16(0);
 
-uint32_t vpx_sad32x32_neon(const uint8_t *src, int src_stride,
-                           const uint8_t *ref, int ref_stride) {
-  const uint16x8_t abs = sad32x(src, src_stride, ref, ref_stride, 32);
-  return horizontal_add_16x8(abs);
+  for (i = 0; i < height; ++i) {
+    const uint8x16_t a_lo = vld1q_u8(a);
+    const uint8x16_t a_hi = vld1q_u8(a + 16);
+    const uint8x16_t b_lo = vld1q_u8(b);
+    const uint8x16_t b_hi = vld1q_u8(b + 16);
+    const uint8x16_t c_lo = vld1q_u8(c);
+    const uint8x16_t c_hi = vld1q_u8(c + 16);
+    const uint8x16_t avg_lo = vrhaddq_u8(b_lo, c_lo);
+    const uint8x16_t avg_hi = vrhaddq_u8(b_hi, c_hi);
+    a += a_stride;
+    b += b_stride;
+    c += 32;
+    abs = vabal_u8(abs, vget_low_u8(a_lo), vget_low_u8(avg_lo));
+    abs = vabal_u8(abs, vget_high_u8(a_lo), vget_high_u8(avg_lo));
+    abs = vabal_u8(abs, vget_low_u8(a_hi), vget_low_u8(avg_hi));
+    abs = vabal_u8(abs, vget_high_u8(a_hi), vget_high_u8(avg_hi));
+  }
+  return abs;
 }
 
-uint32_t vpx_sad32x64_neon(const uint8_t *src, int src_stride,
-                           const uint8_t *ref, int ref_stride) {
-  const uint16x8_t abs = sad32x(src, src_stride, ref, ref_stride, 64);
-  return horizontal_add_16x8(abs);
-}
+#define sad32xN(n)                                                      \
+  uint32_t vpx_sad32x##n##_neon(const uint8_t *src, int src_stride,     \
+                                const uint8_t *ref, int ref_stride) {   \
+    const uint16x8_t abs = sad32x(src, src_stride, ref, ref_stride, n); \
+    return horizontal_add_16x8(abs);                                    \
+  }                                                                     \
+                                                                        \
+  uint32_t vpx_sad32x##n##_avg_neon(const uint8_t *src, int src_stride, \
+                                    const uint8_t *ref, int ref_stride, \
+                                    const uint8_t *second_pred) {       \
+    const uint16x8_t abs =                                              \
+        sad32x_avg(src, src_stride, ref, ref_stride, second_pred, n);   \
+    return horizontal_add_16x8(abs);                                    \
+  }
+
+sad32xN(16);
+sad32xN(32);
+sad32xN(64);
 
 static INLINE uint32_t horizontal_add_32x4(const uint32x4_t a) {
   const uint64x2_t b = vpaddlq_u32(a);
@@ -249,14 +297,63 @@ static INLINE uint32x4_t sad64x(const uint8_t *a, int a_stride,
   }
 }
 
-uint32_t vpx_sad64x32_neon(const uint8_t *src, int src_stride,
-                           const uint8_t *ref, int ref_stride) {
-  const uint32x4_t abs = sad64x(src, src_stride, ref, ref_stride, 32);
-  return horizontal_add_32x4(abs);
-}
+static INLINE uint32x4_t sad64x_avg(const uint8_t *a, int a_stride,
+                                    const uint8_t *b, int b_stride,
+                                    const uint8_t *c, const int height) {
+  int i;
+  uint16x8_t abs_0 = vdupq_n_u16(0);
+  uint16x8_t abs_1 = vdupq_n_u16(0);
 
-uint32_t vpx_sad64x64_neon(const uint8_t *src, int src_stride,
-                           const uint8_t *ref, int ref_stride) {
-  const uint32x4_t abs = sad64x(src, src_stride, ref, ref_stride, 64);
-  return horizontal_add_32x4(abs);
+  for (i = 0; i < height; ++i) {
+    const uint8x16_t a_0 = vld1q_u8(a);
+    const uint8x16_t a_1 = vld1q_u8(a + 16);
+    const uint8x16_t a_2 = vld1q_u8(a + 32);
+    const uint8x16_t a_3 = vld1q_u8(a + 48);
+    const uint8x16_t b_0 = vld1q_u8(b);
+    const uint8x16_t b_1 = vld1q_u8(b + 16);
+    const uint8x16_t b_2 = vld1q_u8(b + 32);
+    const uint8x16_t b_3 = vld1q_u8(b + 48);
+    const uint8x16_t c_0 = vld1q_u8(c);
+    const uint8x16_t c_1 = vld1q_u8(c + 16);
+    const uint8x16_t c_2 = vld1q_u8(c + 32);
+    const uint8x16_t c_3 = vld1q_u8(c + 48);
+    const uint8x16_t avg_0 = vrhaddq_u8(b_0, c_0);
+    const uint8x16_t avg_1 = vrhaddq_u8(b_1, c_1);
+    const uint8x16_t avg_2 = vrhaddq_u8(b_2, c_2);
+    const uint8x16_t avg_3 = vrhaddq_u8(b_3, c_3);
+    a += a_stride;
+    b += b_stride;
+    c += 64;
+    abs_0 = vabal_u8(abs_0, vget_low_u8(a_0), vget_low_u8(avg_0));
+    abs_0 = vabal_u8(abs_0, vget_high_u8(a_0), vget_high_u8(avg_0));
+    abs_0 = vabal_u8(abs_0, vget_low_u8(a_1), vget_low_u8(avg_1));
+    abs_0 = vabal_u8(abs_0, vget_high_u8(a_1), vget_high_u8(avg_1));
+    abs_1 = vabal_u8(abs_1, vget_low_u8(a_2), vget_low_u8(avg_2));
+    abs_1 = vabal_u8(abs_1, vget_high_u8(a_2), vget_high_u8(avg_2));
+    abs_1 = vabal_u8(abs_1, vget_low_u8(a_3), vget_low_u8(avg_3));
+    abs_1 = vabal_u8(abs_1, vget_high_u8(a_3), vget_high_u8(avg_3));
+  }
+
+  {
+    const uint32x4_t sum = vpaddlq_u16(abs_0);
+    return vpadalq_u16(sum, abs_1);
+  }
 }
+
+#define sad64xN(n)                                                      \
+  uint32_t vpx_sad64x##n##_neon(const uint8_t *src, int src_stride,     \
+                                const uint8_t *ref, int ref_stride) {   \
+    const uint32x4_t abs = sad64x(src, src_stride, ref, ref_stride, n); \
+    return horizontal_add_32x4(abs);                                    \
+  }                                                                     \
+                                                                        \
+  uint32_t vpx_sad64x##n##_avg_neon(const uint8_t *src, int src_stride, \
+                                    const uint8_t *ref, int ref_stride, \
+                                    const uint8_t *second_pred) {       \
+    const uint32x4_t abs =                                              \
+        sad64x_avg(src, src_stride, ref, ref_stride, second_pred, n);   \
+    return horizontal_add_32x4(abs);                                    \
+  }
+
+sad64xN(32);
+sad64xN(64);