sad4d neon: 32x[16,32,64]

Rewrite 32x32. Use half the accumulator registers.

BUG=webm:1425

Change-Id: Ibf5e61dc4ba15056102aef8495f4a02c668c5d13

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 345178d4fc69566625cbdeca4a7bf4671307f9e8..e37356a687e2af9dd82bbfc4c19cd000a15431cc 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -676,7 +676,9 @@ INSTANTIATE_TEST_CASE_P(NEON, SADavgTest, ::testing::ValuesIn(avg_neon_tests));
 
 const SadMxNx4Param x4d_neon_tests[] = {
   SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon),
+  SadMxNx4Param(32, 64, &vpx_sad32x64x4d_neon),
   SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon),
+  SadMxNx4Param(32, 16, &vpx_sad32x16x4d_neon),
   SadMxNx4Param(16, 32, &vpx_sad16x32x4d_neon),
   SadMxNx4Param(16, 16, &vpx_sad16x16x4d_neon),
   SadMxNx4Param(16, 8, &vpx_sad16x8x4d_neon),

diff --git a/vpx_dsp/arm/sad4d_neon.c b/vpx_dsp/arm/sad4d_neon.c
index 6b7cc83a969456616fc56e70ca7381caf53f91aa..afb320aca49e4eb9ec1d7996ed233975ad68ac20 100644 (file)
--- a/vpx_dsp/arm/sad4d_neon.c
+++ b/vpx_dsp/arm/sad4d_neon.c
@@ -130,6 +130,52 @@ void vpx_sad16x32x4d_neon(const uint8_t *src, int src_stride,
   sad16x_4d(src, src_stride, ref, ref_stride, res, 32);
 }
 
+static INLINE void sad32x_4d(const uint8_t *a, int a_stride,
+                             const uint8_t *const b[4], int b_stride,
+                             uint32_t *result, const int height) {
+  int i, j;
+  uint16x8_t sum[4] = { vdupq_n_u16(0), vdupq_n_u16(0), vdupq_n_u16(0),
+                        vdupq_n_u16(0) };
+  const uint8_t *b_loop[4] = { b[0], b[1], b[2], b[3] };
+
+  for (i = 0; i < height; ++i) {
+    const uint8x16_t a_0 = vld1q_u8(a);
+    const uint8x16_t a_1 = vld1q_u8(a + 16);
+    a += a_stride;
+    for (j = 0; j < 4; ++j) {
+      const uint8x16_t b_0 = vld1q_u8(b_loop[j]);
+      const uint8x16_t b_1 = vld1q_u8(b_loop[j] + 16);
+      b_loop[j] += b_stride;
+      sum[j] = vabal_u8(sum[j], vget_low_u8(a_0), vget_low_u8(b_0));
+      sum[j] = vabal_u8(sum[j], vget_high_u8(a_0), vget_high_u8(b_0));
+      sum[j] = vabal_u8(sum[j], vget_low_u8(a_1), vget_low_u8(b_1));
+      sum[j] = vabal_u8(sum[j], vget_high_u8(a_1), vget_high_u8(b_1));
+    }
+  }
+
+  for (j = 0; j < 4; ++j) {
+    result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
+  }
+}
+
+void vpx_sad32x16x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  sad32x_4d(src, src_stride, ref, ref_stride, res, 16);
+}
+
+void vpx_sad32x32x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  sad32x_4d(src, src_stride, ref, ref_stride, res, 32);
+}
+
+void vpx_sad32x64x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  sad32x_4d(src, src_stride, ref, ref_stride, res, 64);
+}
+
 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                                     const uint16x8_t vec_hi) {
   const uint32x4_t vec_l_lo =
@@ -175,25 +221,6 @@ static void sad_neon_64(const uint8x16_t vec_src_00,
                              vget_high_u8(vec_ref_48));
 }
 
-// Calculate the absolute difference of 32 bytes from vec_src_00, vec_src_16,
-// and ref. Accumulate partial sums in vec_sum_ref_lo and vec_sum_ref_hi.
-static void sad_neon_32(const uint8x16_t vec_src_00,
-                        const uint8x16_t vec_src_16, const uint8_t *ref,
-                        uint16x8_t *vec_sum_ref_lo,
-                        uint16x8_t *vec_sum_ref_hi) {
-  const uint8x16_t vec_ref_00 = vld1q_u8(ref);
-  const uint8x16_t vec_ref_16 = vld1q_u8(ref + 16);
-
-  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_00),
-                             vget_low_u8(vec_ref_00));
-  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_00),
-                             vget_high_u8(vec_ref_00));
-  *vec_sum_ref_lo = vabal_u8(*vec_sum_ref_lo, vget_low_u8(vec_src_16),
-                             vget_low_u8(vec_ref_16));
-  *vec_sum_ref_hi = vabal_u8(*vec_sum_ref_hi, vget_high_u8(vec_src_16),
-                             vget_high_u8(vec_ref_16));
-}
-
 void vpx_sad64x64x4d_neon(const uint8_t *src, int src_stride,
                           const uint8_t *const ref[4], int ref_stride,
                           uint32_t *res) {
@@ -239,47 +266,3 @@ void vpx_sad64x64x4d_neon(const uint8_t *src, int src_stride,
   res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
   res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
 }
-
-void vpx_sad32x32x4d_neon(const uint8_t *src, int src_stride,
-                          const uint8_t *const ref[4], int ref_stride,
-                          uint32_t *res) {
-  int i;
-  uint16x8_t vec_sum_ref0_lo = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref0_hi = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref1_lo = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref1_hi = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref2_lo = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref2_hi = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref3_lo = vdupq_n_u16(0);
-  uint16x8_t vec_sum_ref3_hi = vdupq_n_u16(0);
-  const uint8_t *ref0, *ref1, *ref2, *ref3;
-  ref0 = ref[0];
-  ref1 = ref[1];
-  ref2 = ref[2];
-  ref3 = ref[3];
-
-  for (i = 0; i < 32; ++i) {
-    const uint8x16_t vec_src_00 = vld1q_u8(src);
-    const uint8x16_t vec_src_16 = vld1q_u8(src + 16);
-
-    sad_neon_32(vec_src_00, vec_src_16, ref0, &vec_sum_ref0_lo,
-                &vec_sum_ref0_hi);
-    sad_neon_32(vec_src_00, vec_src_16, ref1, &vec_sum_ref1_lo,
-                &vec_sum_ref1_hi);
-    sad_neon_32(vec_src_00, vec_src_16, ref2, &vec_sum_ref2_lo,
-                &vec_sum_ref2_hi);
-    sad_neon_32(vec_src_00, vec_src_16, ref3, &vec_sum_ref3_lo,
-                &vec_sum_ref3_hi);
-
-    src += src_stride;
-    ref0 += ref_stride;
-    ref1 += ref_stride;
-    ref2 += ref_stride;
-    ref3 += ref_stride;
-  }
-
-  res[0] = horizontal_long_add_16x8(vec_sum_ref0_lo, vec_sum_ref0_hi);
-  res[1] = horizontal_long_add_16x8(vec_sum_ref1_lo, vec_sum_ref1_hi);
-  res[2] = horizontal_long_add_16x8(vec_sum_ref2_lo, vec_sum_ref2_hi);
-  res[3] = horizontal_long_add_16x8(vec_sum_ref3_lo, vec_sum_ref3_hi);
-}

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 4fe37b76b495453071df810e0cec84a1dbb77780..44d6e4c7155f839e0f63d7a6fbb9cf90de15a103 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -863,13 +863,13 @@ add_proto qw/void vpx_sad64x32x4d/, "const uint8_t *src_ptr, int src_stride, con
 specialize qw/vpx_sad64x32x4d msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x64x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad32x64x4d msa sse2 vsx/;
+specialize qw/vpx_sad32x64x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad32x32x4d avx2 neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad32x16x4d msa sse2 vsx/;
+specialize qw/vpx_sad32x16x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x32x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad16x32x4d neon msa sse2 vsx/;