Add Neon implementation of high bitdepth 8x8 hadamard transform

Add Neon implementation of vpx_highbd_hadamard_8x8 as well as the
corresponding tests.

Change-Id: I3ef1ff199d76b6b010591ef15a81b0f36c9ded03

diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index f904e81..2062cbe 100644 (file)
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -324,5 +324,11 @@ INSTANTIATE_TEST_SUITE_P(
                                            32)));
 #endif  // HAVE_AVX2
 
+#if HAVE_NEON
+INSTANTIATE_TEST_SUITE_P(
+    NEON, HadamardHighbdTest,
+    ::testing::Values(HadamardFuncWithSize(&vpx_highbd_hadamard_8x8_neon, 8)));
+#endif
+
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 }  // namespace

diff --git a/vpx_dsp/arm/highbd_hadamard_neon.c b/vpx_dsp/arm/highbd_hadamard_neon.c
new file mode 100644 (file)
index 0000000..615de4b
--- /dev/null
+++ b/vpx_dsp/arm/highbd_hadamard_neon.c
@@ -0,0 +1,137 @@
+/*
+ *  Copyright (c) 2023 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+
+static INLINE void hadamard_highbd_col8_first_pass(int16x8_t *a0, int16x8_t *a1,
+                                                   int16x8_t *a2, int16x8_t *a3,
+                                                   int16x8_t *a4, int16x8_t *a5,
+                                                   int16x8_t *a6,
+                                                   int16x8_t *a7) {
+  int16x8_t b0 = vaddq_s16(*a0, *a1);
+  int16x8_t b1 = vsubq_s16(*a0, *a1);
+  int16x8_t b2 = vaddq_s16(*a2, *a3);
+  int16x8_t b3 = vsubq_s16(*a2, *a3);
+  int16x8_t b4 = vaddq_s16(*a4, *a5);
+  int16x8_t b5 = vsubq_s16(*a4, *a5);
+  int16x8_t b6 = vaddq_s16(*a6, *a7);
+  int16x8_t b7 = vsubq_s16(*a6, *a7);
+
+  int16x8_t c0 = vaddq_s16(b0, b2);
+  int16x8_t c2 = vsubq_s16(b0, b2);
+  int16x8_t c1 = vaddq_s16(b1, b3);
+  int16x8_t c3 = vsubq_s16(b1, b3);
+  int16x8_t c4 = vaddq_s16(b4, b6);
+  int16x8_t c6 = vsubq_s16(b4, b6);
+  int16x8_t c5 = vaddq_s16(b5, b7);
+  int16x8_t c7 = vsubq_s16(b5, b7);
+
+  *a0 = vaddq_s16(c0, c4);
+  *a2 = vsubq_s16(c0, c4);
+  *a7 = vaddq_s16(c1, c5);
+  *a6 = vsubq_s16(c1, c5);
+  *a3 = vaddq_s16(c2, c6);
+  *a1 = vsubq_s16(c2, c6);
+  *a4 = vaddq_s16(c3, c7);
+  *a5 = vsubq_s16(c3, c7);
+}
+
+static INLINE void hadamard_highbd_col4_second_pass(int16x4_t a0, int16x4_t a1,
+                                                    int16x4_t a2, int16x4_t a3,
+                                                    int16x4_t a4, int16x4_t a5,
+                                                    int16x4_t a6, int16x4_t a7,
+                                                    tran_low_t *coeff) {
+  int32x4_t b0 = vaddl_s16(a0, a1);
+  int32x4_t b1 = vsubl_s16(a0, a1);
+  int32x4_t b2 = vaddl_s16(a2, a3);
+  int32x4_t b3 = vsubl_s16(a2, a3);
+  int32x4_t b4 = vaddl_s16(a4, a5);
+  int32x4_t b5 = vsubl_s16(a4, a5);
+  int32x4_t b6 = vaddl_s16(a6, a7);
+  int32x4_t b7 = vsubl_s16(a6, a7);
+
+  int32x4_t c0 = vaddq_s32(b0, b2);
+  int32x4_t c2 = vsubq_s32(b0, b2);
+  int32x4_t c1 = vaddq_s32(b1, b3);
+  int32x4_t c3 = vsubq_s32(b1, b3);
+  int32x4_t c4 = vaddq_s32(b4, b6);
+  int32x4_t c6 = vsubq_s32(b4, b6);
+  int32x4_t c5 = vaddq_s32(b5, b7);
+  int32x4_t c7 = vsubq_s32(b5, b7);
+
+  int32x4_t d0 = vaddq_s32(c0, c4);
+  int32x4_t d2 = vsubq_s32(c0, c4);
+  int32x4_t d7 = vaddq_s32(c1, c5);
+  int32x4_t d6 = vsubq_s32(c1, c5);
+  int32x4_t d3 = vaddq_s32(c2, c6);
+  int32x4_t d1 = vsubq_s32(c2, c6);
+  int32x4_t d4 = vaddq_s32(c3, c7);
+  int32x4_t d5 = vsubq_s32(c3, c7);
+
+  store_s32q_to_tran_low(coeff + 0, d0);
+  store_s32q_to_tran_low(coeff + 4, d1);
+  store_s32q_to_tran_low(coeff + 8, d2);
+  store_s32q_to_tran_low(coeff + 12, d3);
+  store_s32q_to_tran_low(coeff + 16, d4);
+  store_s32q_to_tran_low(coeff + 20, d5);
+  store_s32q_to_tran_low(coeff + 24, d6);
+  store_s32q_to_tran_low(coeff + 28, d7);
+}
+
+void vpx_highbd_hadamard_8x8_neon(const int16_t *src_diff, ptrdiff_t src_stride,
+                                  tran_low_t *coeff) {
+  int16x4_t b0, b1, b2, b3, b4, b5, b6, b7;
+
+  int16x8_t s0 = vld1q_s16(src_diff + 0 * src_stride);
+  int16x8_t s1 = vld1q_s16(src_diff + 1 * src_stride);
+  int16x8_t s2 = vld1q_s16(src_diff + 2 * src_stride);
+  int16x8_t s3 = vld1q_s16(src_diff + 3 * src_stride);
+  int16x8_t s4 = vld1q_s16(src_diff + 4 * src_stride);
+  int16x8_t s5 = vld1q_s16(src_diff + 5 * src_stride);
+  int16x8_t s6 = vld1q_s16(src_diff + 6 * src_stride);
+  int16x8_t s7 = vld1q_s16(src_diff + 7 * src_stride);
+
+  // For the first pass we can stay in 16-bit elements (4095*8 = 32760).
+  hadamard_highbd_col8_first_pass(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
+
+  transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
+
+  // For the second pass we need to widen to 32-bit elements, so we're
+  // processing 4 columns at a time.
+  // Skip the second transpose because it is not required.
+
+  b0 = vget_low_s16(s0);
+  b1 = vget_low_s16(s1);
+  b2 = vget_low_s16(s2);
+  b3 = vget_low_s16(s3);
+  b4 = vget_low_s16(s4);
+  b5 = vget_low_s16(s5);
+  b6 = vget_low_s16(s6);
+  b7 = vget_low_s16(s7);
+
+  hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff);
+
+  b0 = vget_high_s16(s0);
+  b1 = vget_high_s16(s1);
+  b2 = vget_high_s16(s2);
+  b3 = vget_high_s16(s3);
+  b4 = vget_high_s16(s4);
+  b5 = vget_high_s16(s5);
+  b6 = vget_high_s16(s6);
+  b7 = vget_high_s16(s7);
+
+  hadamard_highbd_col4_second_pass(b0, b1, b2, b3, b4, b5, b6, b7, coeff + 32);
+}

diff --git a/vpx_dsp/arm/mem_neon.h b/vpx_dsp/arm/mem_neon.h
index 866be74..b7a3638 100644 (file)
--- a/vpx_dsp/arm/mem_neon.h
+++ b/vpx_dsp/arm/mem_neon.h
@@ -102,6 +102,12 @@ static INLINE void store_s16q_to_tran_low(tran_low_t *buf, const int16x8_t a) {
 #endif
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE void store_s32q_to_tran_low(tran_low_t *buf, const int32x4_t a) {
+  vst1q_s32(buf, a);
+}
+#endif
+
 // Propagate type information to the compiler. Without this the compiler may
 // assume the required alignment of uint32_t (4 bytes) and add alignment hints
 // to the memory access.

diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index 5535f82..ab8e5bd 100644 (file)
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -342,6 +342,9 @@ DSP_SRCS-$(HAVE_SSE2)  += x86/avg_intrin_sse2.c
 DSP_SRCS-$(HAVE_AVX2)  += x86/avg_intrin_avx2.c
 DSP_SRCS-$(HAVE_NEON)  += arm/avg_neon.c
 DSP_SRCS-$(HAVE_NEON)  += arm/hadamard_neon.c
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_NEON)  += arm/highbd_hadamard_neon.c
+endif
 DSP_SRCS-$(HAVE_MSA)   += mips/avg_msa.c
 DSP_SRCS-$(HAVE_LSX)   += loongarch/avg_lsx.c
 ifeq ($(VPX_ARCH_X86_64),yes)

diff --git a/vpx_dsp/vpx_dsp_rtcd_defs.pl b/vpx_dsp/vpx_dsp_rtcd_defs.pl
index 8725821..dc3cdc4 100644 (file)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -802,7 +802,7 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
     specialize qw/vpx_hadamard_32x32 sse2 avx2 neon/;
 
     add_proto qw/void vpx_highbd_hadamard_8x8/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
-    specialize qw/vpx_highbd_hadamard_8x8 avx2/;
+    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/;