From: Johann <johannkoenig@google.com>
Date: Tue, 11 Jul 2017 14:22:26 +0000 (-0700)
Subject: sad4d neon: 16x[8,16,32]
X-Git-Tag: v1.7.0~314^2~1
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=807ce8fb1e97d35bc9743c4b02d721c044725fa5;p=libvpx

sad4d neon: 16x[8,16,32]

Rewrite 16x16. Use half the accumulator registers.

BUG=webm:1425

Change-Id: I44b48512b1e3629505d83c2645e800f53878ccc2
---

diff --git a/test/sad_test.cc b/test/sad_test.cc
index e53ba54a1..345178d4f 100644
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -677,7 +677,9 @@ INSTANTIATE_TEST_CASE_P(NEON, SADavgTest, ::testing::ValuesIn(avg_neon_tests));
 const SadMxNx4Param x4d_neon_tests[] = {
   SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon),
   SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon),
+  SadMxNx4Param(16, 32, &vpx_sad16x32x4d_neon),
   SadMxNx4Param(16, 16, &vpx_sad16x16x4d_neon),
+  SadMxNx4Param(16, 8, &vpx_sad16x8x4d_neon),
   SadMxNx4Param(8, 16, &vpx_sad8x16x4d_neon),
   SadMxNx4Param(8, 8, &vpx_sad8x8x4d_neon),
   SadMxNx4Param(8, 4, &vpx_sad8x4x4d_neon),
diff --git a/vpx_dsp/arm/sad4d_neon.c b/vpx_dsp/arm/sad4d_neon.c
index 420169ed3..6b7cc83a9 100644
--- a/vpx_dsp/arm/sad4d_neon.c
+++ b/vpx_dsp/arm/sad4d_neon.c
@@ -88,6 +88,48 @@ void vpx_sad8x16x4d_neon(const uint8_t *src, int src_stride,
   sad8x_4d(src, src_stride, ref, ref_stride, res, 16);
 }
 
+static INLINE void sad16x_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_u8 = vld1q_u8(a);
+    a += a_stride;
+    for (j = 0; j < 4; ++j) {
+      const uint8x16_t b_u8 = vld1q_u8(b_loop[j]);
+      b_loop[j] += b_stride;
+      sum[j] = vabal_u8(sum[j], vget_low_u8(a_u8), vget_low_u8(b_u8));
+      sum[j] = vabal_u8(sum[j], vget_high_u8(a_u8), vget_high_u8(b_u8));
+    }
+  }
+
+  for (j = 0; j < 4; ++j) {
+    result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
+  }
+}
+
+void vpx_sad16x8x4d_neon(const uint8_t *src, int src_stride,
+                         const uint8_t *const ref[4], int ref_stride,
+                         uint32_t *res) {
+  sad16x_4d(src, src_stride, ref, ref_stride, res, 8);
+}
+
+void vpx_sad16x16x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  sad16x_4d(src, src_stride, ref, ref_stride, res, 16);
+}
+
+void vpx_sad16x32x4d_neon(const uint8_t *src, int src_stride,
+                          const uint8_t *const ref[4], int ref_stride,
+                          uint32_t *res) {
+  sad16x_4d(src, src_stride, ref, ref_stride, res, 32);
+}
+
 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                                     const uint16x8_t vec_hi) {
   const uint32x4_t vec_l_lo =
@@ -241,58 +283,3 @@ void vpx_sad32x32x4d_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_sad16x16x4d_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 < 16; ++i) {
-    const uint8x16_t vec_src = vld1q_u8(src);
-    const uint8x16_t vec_ref0 = vld1q_u8(ref0);
-    const uint8x16_t vec_ref1 = vld1q_u8(ref1);
-    const uint8x16_t vec_ref2 = vld1q_u8(ref2);
-    const uint8x16_t vec_ref3 = vld1q_u8(ref3);
-
-    vec_sum_ref0_lo =
-        vabal_u8(vec_sum_ref0_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref0));
-    vec_sum_ref0_hi = vabal_u8(vec_sum_ref0_hi, vget_high_u8(vec_src),
-                               vget_high_u8(vec_ref0));
-    vec_sum_ref1_lo =
-        vabal_u8(vec_sum_ref1_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref1));
-    vec_sum_ref1_hi = vabal_u8(vec_sum_ref1_hi, vget_high_u8(vec_src),
-                               vget_high_u8(vec_ref1));
-    vec_sum_ref2_lo =
-        vabal_u8(vec_sum_ref2_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref2));
-    vec_sum_ref2_hi = vabal_u8(vec_sum_ref2_hi, vget_high_u8(vec_src),
-                               vget_high_u8(vec_ref2));
-    vec_sum_ref3_lo =
-        vabal_u8(vec_sum_ref3_lo, vget_low_u8(vec_src), vget_low_u8(vec_ref3));
-    vec_sum_ref3_hi = vabal_u8(vec_sum_ref3_hi, vget_high_u8(vec_src),
-                               vget_high_u8(vec_ref3));
-
-    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 8dfd2c0cc..4fe37b76b 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -872,13 +872,13 @@ add_proto qw/void vpx_sad32x16x4d/, "const uint8_t *src_ptr, int src_stride, con
 specialize qw/vpx_sad32x16x4d 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 msa sse2 vsx/;
+specialize qw/vpx_sad16x32x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad16x16x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad16x8x4d msa sse2 vsx/;
+specialize qw/vpx_sad16x8x4d neon msa sse2 vsx/;
 
 add_proto qw/void vpx_sad8x16x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad8x16x4d neon msa sse2/;