Add Neon implementation of aom_highbd_hadamard_<w>x<h>
Add Neon implementation of aom_highbd_hadamard_<w>x<h> as well as the
corresponding tests.
Change-Id: I33d307422d80076476a811bdf6471dcf16dc2969
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index 884dd14..be041a2 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -290,6 +290,7 @@
list(APPEND AOM_DSP_ENCODER_INTRIN_NEON
"${AOM_ROOT}/aom_dsp/arm/highbd_avg_neon.c"
+ "${AOM_ROOT}/aom_dsp/arm/highbd_hadamard_neon.c"
"${AOM_ROOT}/aom_dsp/arm/highbd_quantize_neon.c"
"${AOM_ROOT}/aom_dsp/arm/highbd_variance_neon.c")
endif()
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 8f087b5..d71ea2b 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -1243,13 +1243,13 @@
if (aom_config("CONFIG_AV1_HIGHBITDEPTH") eq "yes") {
add_proto qw/void aom_highbd_hadamard_8x8/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
- specialize qw/aom_highbd_hadamard_8x8 avx2/;
+ specialize qw/aom_highbd_hadamard_8x8 avx2 neon/;
add_proto qw/void aom_highbd_hadamard_16x16/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
- specialize qw/aom_highbd_hadamard_16x16 avx2/;
+ specialize qw/aom_highbd_hadamard_16x16 avx2 neon/;
add_proto qw/void aom_highbd_hadamard_32x32/, "const int16_t *src_diff, ptrdiff_t src_stride, tran_low_t *coeff";
- specialize qw/aom_highbd_hadamard_32x32 avx2/;
+ specialize qw/aom_highbd_hadamard_32x32 avx2 neon/;
}
add_proto qw/int aom_satd/, "const tran_low_t *coeff, int length";
specialize qw/aom_satd neon sse2 avx2/;
diff --git a/aom_dsp/arm/highbd_hadamard_neon.c b/aom_dsp/arm/highbd_hadamard_neon.c
new file mode 100644
index 0000000..2d0f8bc
--- /dev/null
+++ b/aom_dsp/arm/highbd_hadamard_neon.c
@@ -0,0 +1,213 @@
+/*
+ * Copyright (c) 2023 The WebM project authors. All Rights Reserved.
+ * Copyright (c) 2023, Alliance for Open Media. All Rights Reserved.
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <arm_neon.h>
+#include "config/aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/arm/mem_neon.h"
+#include "aom_dsp/arm/transpose_neon.h"
+#include "aom_dsp/arm/sum_neon.h"
+#include "aom_ports/mem.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);
+
+ vst1q_s32(coeff + 0, d0);
+ vst1q_s32(coeff + 4, d1);
+ vst1q_s32(coeff + 8, d2);
+ vst1q_s32(coeff + 12, d3);
+ vst1q_s32(coeff + 16, d4);
+ vst1q_s32(coeff + 20, d5);
+ vst1q_s32(coeff + 24, d6);
+ vst1q_s32(coeff + 28, d7);
+}
+
+void aom_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);
+}
+
+void aom_highbd_hadamard_16x16_neon(const int16_t *src_diff,
+ ptrdiff_t src_stride, tran_low_t *coeff) {
+ // Rearrange 16x16 to 8x32 and remove stride.
+ // Top left first.
+ aom_highbd_hadamard_8x8_neon(src_diff, src_stride, coeff);
+ // Top right.
+ aom_highbd_hadamard_8x8_neon(src_diff + 8, src_stride, coeff + 64);
+ // Bottom left.
+ aom_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride, src_stride,
+ coeff + 128);
+ // Bottom right.
+ aom_highbd_hadamard_8x8_neon(src_diff + 8 * src_stride + 8, src_stride,
+ coeff + 192);
+
+ for (int i = 0; i < 16; i++) {
+ int32x4_t a0 = vld1q_s32(coeff + 4 * i);
+ int32x4_t a1 = vld1q_s32(coeff + 4 * i + 64);
+ int32x4_t a2 = vld1q_s32(coeff + 4 * i + 128);
+ int32x4_t a3 = vld1q_s32(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);
+
+ vst1q_s32(coeff + 4 * i, c0);
+ vst1q_s32(coeff + 4 * i + 64, c1);
+ vst1q_s32(coeff + 4 * i + 128, c2);
+ vst1q_s32(coeff + 4 * i + 192, c3);
+ }
+}
+
+void aom_highbd_hadamard_32x32_neon(const int16_t *src_diff,
+ ptrdiff_t src_stride, tran_low_t *coeff) {
+ // Rearrange 32x32 to 16x64 and remove stride.
+ // Top left first.
+ aom_highbd_hadamard_16x16_neon(src_diff, src_stride, coeff);
+ // Top right.
+ aom_highbd_hadamard_16x16_neon(src_diff + 16, src_stride, coeff + 256);
+ // Bottom left.
+ aom_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride, src_stride,
+ coeff + 512);
+ // Bottom right.
+ aom_highbd_hadamard_16x16_neon(src_diff + 16 * src_stride + 16, src_stride,
+ coeff + 768);
+
+ for (int i = 0; i < 64; i++) {
+ int32x4_t a0 = vld1q_s32(coeff + 4 * i);
+ int32x4_t a1 = vld1q_s32(coeff + 4 * i + 256);
+ int32x4_t a2 = vld1q_s32(coeff + 4 * i + 512);
+ int32x4_t a3 = vld1q_s32(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);
+
+ vst1q_s32(coeff + 4 * i, c0);
+ vst1q_s32(coeff + 4 * i + 256, c1);
+ vst1q_s32(coeff + 4 * i + 512, c2);
+ vst1q_s32(coeff + 4 * i + 768, c3);
+ }
+}
diff --git a/test/hadamard_test.cc b/test/hadamard_test.cc
index 38c3b5c..e1061ed 100644
--- a/test/hadamard_test.cc
+++ b/test/hadamard_test.cc
@@ -387,6 +387,15 @@
HadamardFuncWithSize(&aom_highbd_hadamard_16x16_c, 16, 16),
HadamardFuncWithSize(&aom_highbd_hadamard_32x32_c, 32, 32)));
+#if HAVE_NEON
+INSTANTIATE_TEST_SUITE_P(
+ NEON, HadamardHighbdTest,
+ ::testing::Values(
+ HadamardFuncWithSize(&aom_highbd_hadamard_8x8_neon, 8, 8),
+ HadamardFuncWithSize(&aom_highbd_hadamard_16x16_neon, 16, 16),
+ HadamardFuncWithSize(&aom_highbd_hadamard_32x32_neon, 32, 32)));
+#endif // HAVE_NEON
+
#endif // CONFIG_AV1_HIGHBITDEPTH
// Tests for low precision
