Add Neon implementation of high bitdepth 16x16 hadamard transform

Add Neon implementation of vpx_highbd_hadamard_16x16 as well as the
corresponding tests.

Change-Id: If3299fe556351dfe3db994ac171d83a95ea1504b

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index 2062cbe34074f8b18f93298ffbee8b6b57694921..2482e87cb6efa1525257da4e8900f08fccf94dee 100644 (file)
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -327,7 +327,9 @@ INSTANTIATE_TEST_SUITE_P(
 #if HAVE_NEON
 INSTANTIATE_TEST_SUITE_P(
     NEON, HadamardHighbdTest,
-    ::testing::Values(HadamardFuncWithSize(&vpx_highbd_hadamard_8x8_neon, 8)));
+    ::testing::Values(HadamardFuncWithSize(&vpx_highbd_hadamard_8x8_neon, 8),
+                      HadamardFuncWithSize(&vpx_highbd_hadamard_16x16_neon,
+                                           16)));
 #endif
 
 #endif  // CONFIG_VP9_HIGHBITDEPTH

diff --git a/vpx_dsp/arm/highbd_hadamard_neon.c b/vpx_dsp/arm/highbd_hadamard_neon.c
index 615de4b0cef3637ece608eefd4a514fb7c597ec6..013f7148f4ae6c79f7b4d987fb42a2fee29fc128 100644 (file)
--- a/vpx_dsp/arm/highbd_hadamard_neon.c
+++ b/vpx_dsp/arm/highbd_hadamard_neon.c
@@ -135,3 +135,42 @@ void vpx_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride,
 
   hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff + 32);
 }
+
+void vpx_highbd_hadamard_16x16_neon(const int16_t *src_diff,
+                                    ptrdiff_t src_stride, tran_low_t *coeff) {
+  int i = 0;
+
+  // Rearrange 16x16 to 8x32 and remove stride.
+  // Top left first.
+  vpx_highbd_hadamard_8x8_neon(src_diff, src_stride, coeff);
+  // Top right.
+  vpx_highbd_hadamard_8x8_neon(src_diff + 8, src_stride, coeff + 64);
+  // Bottom left.
+  vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride, src_stride,
+                               coeff + 128);
+  // Bottom right.
+  vpx_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride + 8, src_stride,
+                               coeff + 192);
+
+  do {
+    int32x4_t a0 = load_tran_low_to_s32q(coeff + 4 * i);
+    int32x4_t a1 = load_tran_low_to_s32q(coeff + 4 * i + 64);
+    int32x4_t a2 = load_tran_low_to_s32q(coeff + 4 * i + 128);
+    int32x4_t a3 = load_tran_low_to_s32q(coeff + 4 * i + 192);
+
+    int32x4_t b0 = vhaddq_s32(a0, a1);
+    int32x4_t b1 = vhsubq_s32(a0, a1);
+    int32x4_t b2 = vhaddq_s32(a2, a3);
+    int32x4_t b3 = vhsubq_s32(a2, a3);
+
+    int32x4_t c0 = vaddq_s32(b0, b2);
+    int32x4_t c1 = vaddq_s32(b1, b3);
+    int32x4_t c2 = vsubq_s32(b0, b2);
+    int32x4_t c3 = vsubq_s32(b1, b3);
+
+    store_s32q_to_tran_low(coeff + 4 * i, c0);
+    store_s32q_to_tran_low(coeff + 4 * i + 64, c1);
+    store_s32q_to_tran_low(coeff + 4 * i + 128, c2);
+    store_s32q_to_tran_low(coeff + 4 * i + 192, c3);
+  } while (++i < 16);
+}

diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h
index b7a363891e29ab1b02737bbb367a98173335391a..2122956dc67246be0d980779a66edb1e7b5daf94 100644 (file)
--- a/vpx_dsp/arm/mem_neon.h
+++ b/vpx_dsp/arm/mem_neon.h
@@ -106,6 +106,10 @@ static INLINE void store_s16q_to_tran_low(tran_low_t *buf, const int16x8_t a) {
 static INLINE void store_s32q_to_tran_low(tran_low_t *buf, const int32x4_t a) {
   vst1q_s32(buf, a);
 }
+
+static INLINE int32x4_t load_tran_low_to_s32q(const tran_low_t *buf) {
+  return vld1q_s32(buf);
+}
 #endif
 
 // Propagate type information to the compiler. Without this the compiler may

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index dc3cdc414571ef0a8fe212792d4d4b643fdac98a..276d55baff10fc93d8a563026d7668d07f6c5e1c 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -805,7 +805,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_highbd_hadamard_8x8 avx2 neon/;
 
     add_proto qw/void vpx_highbd_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
-    specialize qw/vpx_highbd_hadamard_16x16 avx2/;
+    specialize qw/vpx_highbd_hadamard_16x16 avx2 neon/;
 
     add_proto qw/void vpx_highbd_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
     specialize qw/vpx_highbd_hadamard_32x32 avx2/;