Add Neon implementation of aom_convolve8_horiz
Add an Armv8.0 Neon path for aom_convolve8_horiz, and a corresponding
set of unit tests.
A Neon implementation of aom_convolve8_vert will be added in a
subsequent patch.
The Neon code added in this change is mostly a port of the Armv8.0
Neon implementation of the 'vpx_convolve8_horiz_neon' function in
libvpx.[1]
[1] https://github.com/webmproject/libvpx/blob/main/vpx_dsp/arm/vpx_convolve8_neon.c
Change-Id: I038508fb2a1f4db4eaa9119a210fc8c206665a4a
diff --git a/aom_dsp/aom_dsp.cmake b/aom_dsp/aom_dsp.cmake
index be041a2..af1ce4e 100644
--- a/aom_dsp/aom_dsp.cmake
+++ b/aom_dsp/aom_dsp.cmake
@@ -112,6 +112,7 @@
list(APPEND AOM_DSP_COMMON_INTRIN_NEON
"${AOM_ROOT}/aom_dsp/arm/aom_convolve_copy_neon.c"
+ "${AOM_ROOT}/aom_dsp/arm/aom_convolve8_neon.c"
"${AOM_ROOT}/aom_dsp/arm/fwd_txfm_neon.c"
"${AOM_ROOT}/aom_dsp/arm/loopfilter_neon.c"
"${AOM_ROOT}/aom_dsp/arm/highbd_intrapred_neon.c"
diff --git a/aom_dsp/aom_dsp_rtcd_defs.pl b/aom_dsp/aom_dsp_rtcd_defs.pl
index 16bb4a9..c936328 100755
--- a/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -472,7 +472,7 @@
add_proto qw/void aom_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h";
specialize qw/aom_convolve_copy neon sse2 avx2/;
-specialize qw/aom_convolve8_horiz sse2 ssse3/, "$avx2_ssse3";
+specialize qw/aom_convolve8_horiz neon sse2 ssse3/, "$avx2_ssse3";
specialize qw/aom_convolve8_vert sse2 ssse3/, "$avx2_ssse3";
add_proto qw/void aom_scaled_2d/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h";
diff --git a/aom_dsp/arm/aom_convolve8_neon.c b/aom_dsp/arm/aom_convolve8_neon.c
new file mode 100644
index 0000000..b5fb14d
--- /dev/null
+++ b/aom_dsp/arm/aom_convolve8_neon.c
@@ -0,0 +1,251 @@
+/*
+ * Copyright (c) 2014 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 <assert.h>
+#include <string.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/arm/mem_neon.h"
+#include "aom_dsp/arm/transpose_neon.h"
+#include "aom_ports/mem.h"
+
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+ // NOTE: This assumes that the filter table is 256-byte aligned.
+ return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
+}
+
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+ return (int)((const InterpKernel *)(intptr_t)f - base);
+}
+
+static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
+ const int16x4_t s2, const int16x4_t s3,
+ const int16x4_t s4, const int16x4_t s5,
+ const int16x4_t s6, const int16x4_t s7,
+ const int16x8_t filters) {
+ const int16x4_t filters_lo = vget_low_s16(filters);
+ const int16x4_t filters_hi = vget_high_s16(filters);
+ int16x4_t sum;
+
+ sum = vmul_lane_s16(s0, filters_lo, 0);
+ sum = vmla_lane_s16(sum, s1, filters_lo, 1);
+ sum = vmla_lane_s16(sum, s2, filters_lo, 2);
+ sum = vmla_lane_s16(sum, s5, filters_hi, 1);
+ sum = vmla_lane_s16(sum, s6, filters_hi, 2);
+ sum = vmla_lane_s16(sum, s7, filters_hi, 3);
+ sum = vqadd_s16(sum, vmul_lane_s16(s3, filters_lo, 3));
+ sum = vqadd_s16(sum, vmul_lane_s16(s4, filters_hi, 0));
+ return sum;
+}
+
+static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
+ const int16x8_t s2, const int16x8_t s3,
+ const int16x8_t s4, const int16x8_t s5,
+ const int16x8_t s6, const int16x8_t s7,
+ const int16x8_t filters) {
+ const int16x4_t filters_lo = vget_low_s16(filters);
+ const int16x4_t filters_hi = vget_high_s16(filters);
+ int16x8_t sum;
+
+ sum = vmulq_lane_s16(s0, filters_lo, 0);
+ sum = vmlaq_lane_s16(sum, s1, filters_lo, 1);
+ sum = vmlaq_lane_s16(sum, s2, filters_lo, 2);
+ sum = vmlaq_lane_s16(sum, s5, filters_hi, 1);
+ sum = vmlaq_lane_s16(sum, s6, filters_hi, 2);
+ sum = vmlaq_lane_s16(sum, s7, filters_hi, 3);
+ sum = vqaddq_s16(sum, vmulq_lane_s16(s3, filters_lo, 3));
+ sum = vqaddq_s16(sum, vmulq_lane_s16(s4, filters_hi, 0));
+ return vqrshrun_n_s16(sum, FILTER_BITS);
+}
+
+void aom_convolve8_horiz_neon(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4, int w,
+ int h) {
+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+ const int16x8_t filters = vld1q_s16(filters_x[x0_q4]);
+
+ assert((intptr_t)dst % 4 == 0);
+ assert(dst_stride % 4 == 0);
+
+ (void)x_step_q4;
+ (void)filter_y;
+ (void)y_step_q4;
+
+ src -= ((SUBPEL_TAPS / 2) - 1);
+
+ if (h == 4) {
+ uint8x8_t t0, t1, t2, t3, d01, d23;
+ int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, d0, d1, d2, d3;
+
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+ s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+
+ src += 7;
+
+ do {
+ load_u8_8x4(src, src_stride, &t0, &t1, &t2, &t3);
+ transpose_u8_8x4(&t0, &t1, &t2, &t3);
+ s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
+ s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
+ s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
+ s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
+
+ d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+ d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
+ d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
+
+ transpose_u8_4x4(&d01, &d23);
+
+ store_u8_4x1(dst + 0 * dst_stride, d01, 0);
+ store_u8_4x1(dst + 1 * dst_stride, d23, 0);
+ store_u8_4x1(dst + 2 * dst_stride, d01, 1);
+ store_u8_4x1(dst + 3 * dst_stride, d23, 1);
+
+ s0 = s4;
+ s1 = s5;
+ s2 = s6;
+ s3 = s7;
+ s4 = s8;
+ s5 = s9;
+ s6 = s10;
+ src += 4;
+ dst += 4;
+ w -= 4;
+ } while (w != 0);
+ } else {
+ uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7, d0, d1, d2, d3;
+ int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
+
+ if (w == 4) {
+ do {
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ load_u8_8x8(src + 7, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6,
+ &t7);
+ transpose_u8_4x8(&t0, &t1, &t2, &t3, t4, t5, t6, t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+
+ transpose_u8_8x4(&d0, &d1, &d2, &d3);
+
+ store_u8_4x1(dst + 0 * dst_stride, d0, 0);
+ store_u8_4x1(dst + 1 * dst_stride, d1, 0);
+ store_u8_4x1(dst + 2 * dst_stride, d2, 0);
+ store_u8_4x1(dst + 3 * dst_stride, d3, 0);
+ store_u8_4x1(dst + 4 * dst_stride, d0, 1);
+ store_u8_4x1(dst + 5 * dst_stride, d1, 1);
+ store_u8_4x1(dst + 6 * dst_stride, d2, 1);
+ store_u8_4x1(dst + 7 * dst_stride, d3, 1);
+
+ src += 8 * src_stride;
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h > 0);
+ } else {
+ uint8x8_t d4, d5, d6, d7;
+ int16x8_t s11, s12, s13, s14;
+ int width;
+ const uint8_t *s;
+ uint8_t *d;
+
+ do {
+ load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
+
+ width = w;
+ s = src + 7;
+ d = dst;
+
+ do {
+ load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ transpose_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
+ s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
+ s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
+ s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
+ s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
+ s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
+ s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
+ s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
+ s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
+
+ d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters);
+ d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters);
+ d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters);
+ d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters);
+ d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters);
+ d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters);
+ d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters);
+ d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters);
+
+ transpose_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
+
+ store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
+
+ s0 = s8;
+ s1 = s9;
+ s2 = s10;
+ s3 = s11;
+ s4 = s12;
+ s5 = s13;
+ s6 = s14;
+ s += 8;
+ d += 8;
+ width -= 8;
+ } while (width != 0);
+ src += 8 * src_stride;
+ dst += 8 * dst_stride;
+ h -= 8;
+ } while (h > 0);
+ }
+ }
+}
diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index a23cfcc..53cf1d6 100644
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -866,4 +866,13 @@
::testing::ValuesIn(kArray_Convolve8_avx2));
#endif // HAVE_AVX2
+#if HAVE_NEON
+const ConvolveFunctions convolve8_neon(aom_convolve8_horiz_neon,
+ aom_convolve8_vert_c, 0);
+const ConvolveParam kArray_Convolve8_neon[] = { ALL_SIZES(convolve8_neon) };
+
+INSTANTIATE_TEST_SUITE_P(NEON, LowbdConvolveTest,
+ ::testing::ValuesIn(kArray_Convolve8_neon));
+#endif // HAVE_NEON
+
} // namespace
