sad4d neon: 8x[4,8,16]

BUG=webm:1425

Change-Id: I7de2500cca4b621f21478c4b0333c56d76dbc9a4

diff --git a/test/sad_test.cc b/test/sad_test.cc
index 13eae70..e53ba54 100644 (file)
--- a/test/sad_test.cc
+++ b/test/sad_test.cc
@@ -678,6 +678,9 @@ const SadMxNx4Param x4d_neon_tests[] = {
   SadMxNx4Param(64, 64, &vpx_sad64x64x4d_neon),
   SadMxNx4Param(32, 32, &vpx_sad32x32x4d_neon),
   SadMxNx4Param(16, 16, &vpx_sad16x16x4d_neon),
+  SadMxNx4Param(8, 16, &vpx_sad8x16x4d_neon),
+  SadMxNx4Param(8, 8, &vpx_sad8x8x4d_neon),
+  SadMxNx4Param(8, 4, &vpx_sad8x4x4d_neon),
   SadMxNx4Param(4, 8, &vpx_sad4x8x4d_neon),
   SadMxNx4Param(4, 4, &vpx_sad4x4x4d_neon),
 };

diff --git a/vpx_dsp/arm/sad4d_neon.c b/vpx_dsp/arm/sad4d_neon.c
index 9c45ae1..420169e 100644 (file)
--- a/vpx_dsp/arm/sad4d_neon.c
+++ b/vpx_dsp/arm/sad4d_neon.c
@@ -47,6 +47,47 @@ void vpx_sad4x8x4d_neon(const uint8_t *src, int src_stride,
   }
 }
 
+static INLINE void sad8x_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 uint8x8_t a_u8 = vld1_u8(a);
+    a += a_stride;
+    for (j = 0; j < 4; ++j) {
+      const uint8x8_t b_u8 = vld1_u8(b_loop[j]);
+      b_loop[j] += b_stride;
+      sum[j] = vabal_u8(sum[j], a_u8, b_u8);
+    }
+  }
+
+  for (j = 0; j < 4; ++j) {
+    result[j] = vget_lane_u32(horizontal_add_uint16x8(sum[j]), 0);
+  }
+}
+
+void vpx_sad8x4x4d_neon(const uint8_t *src, int src_stride,
+                        const uint8_t *const ref[4], int ref_stride,
+                        uint32_t *res) {
+  sad8x_4d(src, src_stride, ref, ref_stride, res, 4);
+}
+
+void vpx_sad8x8x4d_neon(const uint8_t *src, int src_stride,
+                        const uint8_t *const ref[4], int ref_stride,
+                        uint32_t *res) {
+  sad8x_4d(src, src_stride, ref, ref_stride, res, 8);
+}
+
+void vpx_sad8x16x4d_neon(const uint8_t *src, int src_stride,
+                         const uint8_t *const ref[4], int ref_stride,
+                         uint32_t *res) {
+  sad8x_4d(src, src_stride, ref, ref_stride, res, 16);
+}
+
 static INLINE unsigned int horizontal_long_add_16x8(const uint16x8_t vec_lo,
                                                     const uint16x8_t vec_hi) {
   const uint32x4_t vec_l_lo =

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 57c6a1e..8dfd2c0 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -881,13 +881,13 @@ add_proto qw/void vpx_sad16x8x4d/, "const uint8_t *src_ptr, int src_stride, cons
 specialize qw/vpx_sad16x8x4d 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 msa sse2/;
+specialize qw/vpx_sad8x16x4d neon msa sse2/;
 
 add_proto qw/void vpx_sad8x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad8x8x4d msa sse2/;
+specialize qw/vpx_sad8x8x4d neon msa sse2/;
 
 add_proto qw/void vpx_sad8x4x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
-specialize qw/vpx_sad8x4x4d msa sse2/;
+specialize qw/vpx_sad8x4x4d neon msa sse2/;
 
 add_proto qw/void vpx_sad4x8x4d/, "const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[], int ref_stride, uint32_t *sad_array";
 specialize qw/vpx_sad4x8x4d neon msa sse2/;