From 111068923b4ca778a680330d00161d7ee93f61e1 Mon Sep 17 00:00:00 2001
From: Salome Thirot <salome.thirot@arm.com>
Date: Thu, 23 Feb 2023 12:05:30 +0000
Subject: [PATCH] Add Neon implementation of high bitdepth 32x32 hadamard
 transform

Add Neon implementation of vpx_highbd_hadamard_32x32 as well as the
corresponding tests.

Change-Id: I65d8603896649de1996b353aa79eee54824b4708
---
 test/hadamard_test.cc              |  5 +++--
 vpx_dsp/arm/highbd_hadamard_neon.c | 39 ++++++++++++++++++++++++++++++++++++++
 vpx_dsp/vpx_dsp_rtcd_defs.pl       |  2 +-
 3 files changed, 43 insertions(+), 3 deletions(-)

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index 2482e87..9f6c99f 100644
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -328,8 +328,9 @@ INSTANTIATE_TEST_SUITE_P(
 INSTANTIATE_TEST_SUITE_P(
     NEON, HadamardHighbdTest,
     ::testing::Values(HadamardFuncWithSize(&vpx_highbd_hadamard_8x8_neon, 8),
-                      HadamardFuncWithSize(&vpx_highbd_hadamard_16x16_neon,
-                                           16)));
+                      HadamardFuncWithSize(&vpx_highbd_hadamard_16x16_neon, 16),
+                      HadamardFuncWithSize(&vpx_highbd_hadamard_32x32_neon,
+                                           32)));
 #endif
 
 #endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/vpx_dsp/arm/highbd_hadamard_neon.c b/vpx_dsp/arm/highbd_hadamard_neon.c
index 013f714..499eb65 100644
--- a/vpx_dsp/arm/highbd_hadamard_neon.c
+++ b/vpx_dsp/arm/highbd_hadamard_neon.c
@@ -174,3 +174,42 @@ void vpx_highbd_hadamard_16x16_neon(const int16_t *src_diff,
     store_s32q_to_tran_low(coeff + 4 * i + 192, c3);
   } while (++i < 16);
 }
+
+void vpx_highbd_hadamard_32x32_neon(const int16_t *src_diff,
+                                    ptrdiff_t src_stride, tran_low_t *coeff) {
+  int i = 0;
+
+  // Rearrange 32x32 to 16x64 and remove stride.
+  // Top left first.
+  vpx_highbd_hadamard_16x16_neon(src_diff, src_stride, coeff);
+  // Top right.
+  vpx_highbd_hadamard_16x16_neon(src_diff + 16, src_stride, coeff + 256);
+  // Bottom left.
+  vpx_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride, src_stride,
+                                 coeff + 512);
+  // Bottom right.
+  vpx_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride + 16, src_stride,
+                                 coeff + 768);
+
+  do {
+    int32x4_t a0 = load_tran_low_to_s32q(coeff + 4 * i);
+    int32x4_t a1 = load_tran_low_to_s32q(coeff + 4 * i + 256);
+    int32x4_t a2 = load_tran_low_to_s32q(coeff + 4 * i + 512);
+    int32x4_t a3 = load_tran_low_to_s32q(coeff + 4 * i + 768);
+
+    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 = vhaddq_s32(b0, b2);
+    int32x4_t c1 = vhaddq_s32(b1, b3);
+    int32x4_t c2 = vhsubq_s32(b0, b2);
+    int32x4_t c3 = vhsubq_s32(b1, b3);
+
+    store_s32q_to_tran_low(coeff + 4 * i, c0);
+    store_s32q_to_tran_low(coeff + 4 * i + 256, c1);
+    store_s32q_to_tran_low(coeff + 4 * i + 512, c2);
+    store_s32q_to_tran_low(coeff + 4 * i + 768, c3);
+  } while (++i < 64);
+}
diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 276d55b..eef7224 100644
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -808,7 +808,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     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/;
+    specialize qw/vpx_highbd_hadamard_32x32 avx2 neon/;
 
     add_proto qw/int vpx_satd/, "const tran_low_t *coeff, int length";
     specialize qw/vpx_satd avx2 sse2 neon/;
-- 
2.7.4