Add Neon implementation of aom_sum_squares_i16

Add Neon implementation of aom_sum_squares_i16 as well as the
corresponding tests.

Change-Id: I813e2bd89124e0d9cf48fcf933e5cd0e379aecfc

aom_dsp/aom_dsp_rtcd_defs.pl

diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 76827cd..de6573d 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -747,7 +747,7 @@
     specialize qw/aom_sum_squares_2d_i16 sse2 avx2 neon/;
 
     add_proto qw/uint64_t aom_sum_squares_i16/, "const int16_t *src, uint32_t N";
-    specialize qw/aom_sum_squares_i16 sse2/;
+    specialize qw/aom_sum_squares_i16 sse2 neon/;
 
     add_proto qw/uint64_t aom_var_2d_u8/, "uint8_t *src, int src_stride, int width, int height";
     specialize qw/aom_var_2d_u8 sse2 avx2/;

aom_dsp/arm/sum_squares_neon.c

diff --git a/aom_dsp/arm/sum_squares_neon.c b/aom_dsp/arm/sum_squares_neon.c
index 095a2c6..20c9c0d 100644
--- a/aom_dsp/arm/sum_squares_neon.c
+++ b/aom_dsp/arm/sum_squares_neon.c
@@ -223,3 +223,66 @@
 
   return sse;
 }
+
+static INLINE uint64_t aom_sum_squares_i16_4xn_neon(const int16_t *src,
+                                                    uint32_t n) {
+  uint64x2_t sum_u64 = vdupq_n_u64(0);
+
+  int i = n;
+  do {
+    uint32x4_t sum;
+    int16x4_t s0 = vld1_s16(src);
+
+    sum = vreinterpretq_u32_s32(vmull_s16(s0, s0));
+
+    sum_u64 = vpadalq_u32(sum_u64, sum);
+
+    src += 4;
+    i -= 4;
+  } while (i >= 4);
+
+  if (i > 0) {
+    return horizontal_add_u64x2(sum_u64) + aom_sum_squares_i16_c(src, i);
+  }
+  return horizontal_add_u64x2(sum_u64);
+}
+
+static INLINE uint64_t aom_sum_squares_i16_8xn_neon(const int16_t *src,
+                                                    uint32_t n) {
+  uint64x2_t sum_u64[2] = { vdupq_n_u64(0), vdupq_n_u64(0) };
+
+  int i = n;
+  do {
+    uint32x4_t sum[2];
+    int16x8_t s0 = vld1q_s16(src);
+
+    sum[0] =
+        vreinterpretq_u32_s32(vmull_s16(vget_low_s16(s0), vget_low_s16(s0)));
+    sum[1] =
+        vreinterpretq_u32_s32(vmull_s16(vget_high_s16(s0), vget_high_s16(s0)));
+
+    sum_u64[0] = vpadalq_u32(sum_u64[0], sum[0]);
+    sum_u64[1] = vpadalq_u32(sum_u64[1], sum[1]);
+
+    src += 8;
+    i -= 8;
+  } while (i >= 8);
+
+  if (i > 0) {
+    return horizontal_add_u64x2(vaddq_u64(sum_u64[0], sum_u64[1])) +
+           aom_sum_squares_i16_c(src, i);
+  }
+  return horizontal_add_u64x2(vaddq_u64(sum_u64[0], sum_u64[1]));
+}
+
+uint64_t aom_sum_squares_i16_neon(const int16_t *src, uint32_t n) {
+  // This function seems to be called only for values of N >= 64. See
+  // av1/encoder/compound_type.c.
+  if (LIKELY(n >= 8)) {
+    return aom_sum_squares_i16_8xn_neon(src, n);
+  }
+  if (n >= 4) {
+    return aom_sum_squares_i16_4xn_neon(src, n);
+  }
+  return aom_sum_squares_i16_c(src, n);
+}

test/sum_squares_test.cc

diff --git a/test/sum_squares_test.cc b/test/sum_squares_test.cc
index 5c049a5..a89e58c 100644
--- a/test/sum_squares_test.cc
+++ b/test/sum_squares_test.cc
@@ -238,6 +238,13 @@
 
 #endif  // HAVE_SSE2
 
+#if HAVE_NEON
+INSTANTIATE_TEST_SUITE_P(NEON, SumSquares1DTest,
+                         ::testing::Values(TestFuncs1D(
+                             aom_sum_squares_i16_c, aom_sum_squares_i16_neon)));
+
+#endif  // HAVE_NEON
+
 typedef int64_t (*sse_func)(const uint8_t *a, int a_stride, const uint8_t *b,
                             int b_stride, int width, int height);
 typedef libaom_test::FuncParam<sse_func> TestSSEFuncs;