Added AVX2 variant for av1_warp_affine_sse4_1
For speed = 1, 2, 3 and 4 presets observed encode time reduction of
0.8%, 0.95%, 0.2% and 0.2% (averaged across multiple test cases).
Module gains improved by factor of ~1.4x w.r.t SSE4_1 module
Change-Id: I5007f2a96347844e969a5cdcf9c7e3a2928bf8f3
diff --git a/av1/av1.cmake b/av1/av1.cmake
index 4e79437..b817fd0 100644
--- a/av1/av1.cmake
+++ b/av1/av1.cmake
@@ -266,6 +266,7 @@
"${AOM_ROOT}/av1/common/x86/jnt_convolve_avx2.c"
"${AOM_ROOT}/av1/common/x86/reconinter_avx2.c"
"${AOM_ROOT}/av1/common/x86/selfguided_avx2.c"
+ "${AOM_ROOT}/av1/common/x86/warp_plane_avx2.c"
"${AOM_ROOT}/av1/common/x86/wiener_convolve_avx2.c")
list(APPEND AOM_AV1_ENCODER_ASM_SSE2 "${AOM_ROOT}/av1/encoder/x86/dct_sse2.asm"
diff --git a/av1/common/av1_rtcd_defs.pl b/av1/common/av1_rtcd_defs.pl
index bb734ce..2d77bc1 100644
--- a/av1/common/av1_rtcd_defs.pl
+++ b/av1/common/av1_rtcd_defs.pl
@@ -332,7 +332,7 @@
# WARPED_MOTION / GLOBAL_MOTION functions
add_proto qw/void av1_warp_affine/, "const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta";
-specialize qw/av1_warp_affine sse4_1 neon/;
+specialize qw/av1_warp_affine sse4_1 avx2 neon/;
add_proto qw/void av1_highbd_warp_affine/, "const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta";
specialize qw/av1_highbd_warp_affine sse4_1/;
diff --git a/av1/common/warped_motion.h b/av1/common/warped_motion.h
index a1a4f06..39ee388 100644
--- a/av1/common/warped_motion.h
+++ b/av1/common/warped_motion.h
@@ -34,6 +34,9 @@
extern const int16_t warped_filter[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8];
+DECLARE_ALIGNED(8, extern const int8_t,
+ filter_8bit[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8]);
+
static const uint8_t warp_pad_left[14][16] = {
{ 1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 },
{ 2, 2, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 },
diff --git a/av1/common/x86/warp_plane_avx2.c b/av1/common/x86/warp_plane_avx2.c
new file mode 100644
index 0000000..8303e5e
--- /dev/null
+++ b/av1/common/x86/warp_plane_avx2.c
@@ -0,0 +1,1165 @@
+/*
+ * Copyright (c) 2019, 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 <immintrin.h>
+#include "config/av1_rtcd.h"
+#include "av1/common/warped_motion.h"
+
+static const uint8_t shuffle_alpha0_mask01_avx2[32] = {
+ 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+ 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1
+};
+
+static const uint8_t shuffle_alpha0_mask23_avx2[32] = {
+ 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+ 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3
+};
+
+static const uint8_t shuffle_alpha0_mask45_avx2[32] = {
+ 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+ 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5
+};
+
+static const uint8_t shuffle_alpha0_mask67_avx2[32] = {
+ 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+ 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7
+};
+
+static const uint8_t shuffle_gamma0_mask0_avx2[32] = { 0, 1, 2, 3, 0, 1, 2, 3,
+ 0, 1, 2, 3, 0, 1, 2, 3,
+ 0, 1, 2, 3, 0, 1, 2, 3,
+ 0, 1, 2, 3, 0, 1, 2, 3 };
+
+static const uint8_t shuffle_gamma0_mask1_avx2[32] = { 4, 5, 6, 7, 4, 5, 6, 7,
+ 4, 5, 6, 7, 4, 5, 6, 7,
+ 4, 5, 6, 7, 4, 5, 6, 7,
+ 4, 5, 6, 7, 4, 5, 6, 7 };
+
+static const uint8_t shuffle_gamma0_mask2_avx2[32] = {
+ 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11,
+ 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11
+};
+
+static const uint8_t shuffle_gamma0_mask3_avx2[32] = {
+ 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15,
+ 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15
+};
+
+static const uint8_t shuffle_src0[32] = { 0, 2, 2, 4, 4, 6, 6, 8, 1, 3, 3,
+ 5, 5, 7, 7, 9, 0, 2, 2, 4, 4, 6,
+ 6, 8, 1, 3, 3, 5, 5, 7, 7, 9 };
+
+static const uint8_t shuffle_src1[32] = { 4, 6, 6, 8, 8, 10, 10, 12,
+ 5, 7, 7, 9, 9, 11, 11, 13,
+ 4, 6, 6, 8, 8, 10, 10, 12,
+ 5, 7, 7, 9, 9, 11, 11, 13 };
+
+static const uint8_t shuffle_src2[32] = { 1, 3, 3, 5, 5, 7, 7, 9, 2, 4, 4,
+ 6, 6, 8, 8, 10, 1, 3, 3, 5, 5, 7,
+ 7, 9, 2, 4, 4, 6, 6, 8, 8, 10 };
+
+static const uint8_t shuffle_src3[32] = { 5, 7, 7, 9, 9, 11, 11, 13,
+ 6, 8, 8, 10, 10, 12, 12, 14,
+ 5, 7, 7, 9, 9, 11, 11, 13,
+ 6, 8, 8, 10, 10, 12, 12, 14 };
+
+static INLINE void filter_src_pixels_avx2(const __m256i src, __m256i *horz_out,
+ __m256i *coeff,
+ const __m256i *shuffle_src,
+ const __m256i *round_const,
+ const __m128i *shift, int row) {
+ const __m256i src_0 = _mm256_shuffle_epi8(src, shuffle_src[0]);
+ const __m256i src_1 = _mm256_shuffle_epi8(src, shuffle_src[1]);
+ const __m256i src_2 = _mm256_shuffle_epi8(src, shuffle_src[2]);
+ const __m256i src_3 = _mm256_shuffle_epi8(src, shuffle_src[3]);
+
+ const __m256i res_02 = _mm256_maddubs_epi16(src_0, coeff[0]);
+ const __m256i res_46 = _mm256_maddubs_epi16(src_1, coeff[1]);
+ const __m256i res_13 = _mm256_maddubs_epi16(src_2, coeff[2]);
+ const __m256i res_57 = _mm256_maddubs_epi16(src_3, coeff[3]);
+
+ const __m256i res_even = _mm256_add_epi16(res_02, res_46);
+ const __m256i res_odd = _mm256_add_epi16(res_13, res_57);
+ const __m256i res =
+ _mm256_add_epi16(_mm256_add_epi16(res_even, res_odd), *round_const);
+ horz_out[row] = _mm256_srl_epi16(res, *shift);
+}
+
+static INLINE void prepare_horizontal_filter_coeff_avx2(int alpha, int beta,
+ int sx,
+ __m256i *coeff) {
+ __m128i tmp_0 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_1 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_2 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_3 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_4 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_5 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_6 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_7 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
+
+ tmp_0 = _mm_unpacklo_epi16(tmp_0, tmp_2);
+ tmp_1 = _mm_unpacklo_epi16(tmp_1, tmp_3);
+ tmp_4 = _mm_unpacklo_epi16(tmp_4, tmp_6);
+ tmp_5 = _mm_unpacklo_epi16(tmp_5, tmp_7);
+
+ __m128i tmp_8 =
+ _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 0 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_9 =
+ _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 1 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_10 =
+ _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 2 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_11 =
+ _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 3 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ tmp_2 = _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 4 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ tmp_3 = _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 5 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ tmp_6 = _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 6 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+ tmp_7 = _mm_loadl_epi64((__m128i *)&filter_8bit[((sx + beta) + 7 * alpha) >>
+ WARPEDDIFF_PREC_BITS]);
+
+ tmp_8 = _mm_unpacklo_epi16(tmp_8, tmp_10);
+ tmp_2 = _mm_unpacklo_epi16(tmp_2, tmp_6);
+ tmp_9 = _mm_unpacklo_epi16(tmp_9, tmp_11);
+ tmp_3 = _mm_unpacklo_epi16(tmp_3, tmp_7);
+
+ const __m256i tmp_12 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_0), tmp_8, 0x1);
+ const __m256i tmp_13 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_1), tmp_9, 0x1);
+ const __m256i tmp_14 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_4), tmp_2, 0x1);
+ const __m256i tmp_15 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_5), tmp_3, 0x1);
+
+ const __m256i res_0 = _mm256_unpacklo_epi32(tmp_12, tmp_14);
+ const __m256i res_1 = _mm256_unpackhi_epi32(tmp_12, tmp_14);
+ const __m256i res_2 = _mm256_unpacklo_epi32(tmp_13, tmp_15);
+ const __m256i res_3 = _mm256_unpackhi_epi32(tmp_13, tmp_15);
+
+ coeff[0] = _mm256_unpacklo_epi64(res_0, res_2);
+ coeff[1] = _mm256_unpackhi_epi64(res_0, res_2);
+ coeff[2] = _mm256_unpacklo_epi64(res_1, res_3);
+ coeff[3] = _mm256_unpackhi_epi64(res_1, res_3);
+}
+
+static INLINE void prepare_horizontal_filter_coeff_beta0_avx2(int alpha, int sx,
+ __m256i *coeff) {
+ __m128i tmp_0 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_1 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_2 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_3 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_4 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_5 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_6 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ __m128i tmp_7 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
+
+ tmp_0 = _mm_unpacklo_epi16(tmp_0, tmp_2);
+ tmp_1 = _mm_unpacklo_epi16(tmp_1, tmp_3);
+ tmp_4 = _mm_unpacklo_epi16(tmp_4, tmp_6);
+ tmp_5 = _mm_unpacklo_epi16(tmp_5, tmp_7);
+
+ const __m256i tmp_12 = _mm256_broadcastsi128_si256(tmp_0);
+ const __m256i tmp_13 = _mm256_broadcastsi128_si256(tmp_1);
+ const __m256i tmp_14 = _mm256_broadcastsi128_si256(tmp_4);
+ const __m256i tmp_15 = _mm256_broadcastsi128_si256(tmp_5);
+
+ const __m256i res_0 = _mm256_unpacklo_epi32(tmp_12, tmp_14);
+ const __m256i res_1 = _mm256_unpackhi_epi32(tmp_12, tmp_14);
+ const __m256i res_2 = _mm256_unpacklo_epi32(tmp_13, tmp_15);
+ const __m256i res_3 = _mm256_unpackhi_epi32(tmp_13, tmp_15);
+
+ coeff[0] = _mm256_unpacklo_epi64(res_0, res_2);
+ coeff[1] = _mm256_unpackhi_epi64(res_0, res_2);
+ coeff[2] = _mm256_unpacklo_epi64(res_1, res_3);
+ coeff[3] = _mm256_unpackhi_epi64(res_1, res_3);
+}
+
+static INLINE void prepare_horizontal_filter_coeff_alpha0_avx2(int beta, int sx,
+ __m256i *coeff) {
+ const __m128i tmp_0 =
+ _mm_loadl_epi64((__m128i *)&filter_8bit[sx >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_1 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + beta) >> WARPEDDIFF_PREC_BITS]);
+
+ const __m256i res_0 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(tmp_0), tmp_1, 0x1);
+
+ coeff[0] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_alpha0_mask01_avx2));
+ coeff[1] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_alpha0_mask23_avx2));
+ coeff[2] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_alpha0_mask45_avx2));
+ coeff[3] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_alpha0_mask67_avx2));
+}
+
+static INLINE void horizontal_filter_avx2(const __m256i src, __m256i *horz_out,
+ int sx, int alpha, int beta, int row,
+ const __m256i *shuffle_src,
+ const __m256i *round_const,
+ const __m128i *shift) {
+ __m256i coeff[4];
+ prepare_horizontal_filter_coeff_avx2(alpha, beta, sx, coeff);
+ filter_src_pixels_avx2(src, horz_out, coeff, shuffle_src, round_const, shift,
+ row);
+}
+static INLINE void prepare_horizontal_filter_coeff(int alpha, int sx,
+ __m256i *coeff) {
+ const __m128i tmp_0 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_1 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_2 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_3 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_4 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_5 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_6 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
+ const __m128i tmp_7 = _mm_loadl_epi64(
+ (__m128i *)&filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
+
+ const __m128i tmp_8 = _mm_unpacklo_epi16(tmp_0, tmp_2);
+ const __m128i tmp_9 = _mm_unpacklo_epi16(tmp_1, tmp_3);
+ const __m128i tmp_10 = _mm_unpacklo_epi16(tmp_4, tmp_6);
+ const __m128i tmp_11 = _mm_unpacklo_epi16(tmp_5, tmp_7);
+
+ const __m128i tmp_12 = _mm_unpacklo_epi32(tmp_8, tmp_10);
+ const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_8, tmp_10);
+ const __m128i tmp_14 = _mm_unpacklo_epi32(tmp_9, tmp_11);
+ const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_9, tmp_11);
+
+ coeff[0] = _mm256_castsi128_si256(_mm_unpacklo_epi64(tmp_12, tmp_14));
+ coeff[1] = _mm256_castsi128_si256(_mm_unpackhi_epi64(tmp_12, tmp_14));
+ coeff[2] = _mm256_castsi128_si256(_mm_unpacklo_epi64(tmp_13, tmp_15));
+ coeff[3] = _mm256_castsi128_si256(_mm_unpackhi_epi64(tmp_13, tmp_15));
+}
+
+static INLINE void warp_horizontal_filter_avx2(
+ const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
+ int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+ const __m256i *round_const, const __m128i *shift,
+ const __m256i *shuffle_src) {
+ int k, iy, sx, row = 0;
+ __m256i coeff[4];
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_0 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_1 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ const __m256i src_01 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
+ sx = sx4 + beta * (k + 4);
+ horizontal_filter_avx2(src_01, horz_out, sx, alpha, beta, row, shuffle_src,
+ round_const, shift);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i src_01 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
+ sx = sx4 + beta * (k + 4);
+ prepare_horizontal_filter_coeff(alpha, sx, coeff);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+}
+
+static INLINE void warp_horizontal_filter_alpha0_avx2(
+ const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
+ int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+ const __m256i *round_const, const __m128i *shift,
+ const __m256i *shuffle_src) {
+ (void)alpha;
+ int k, iy, sx, row = 0;
+ __m256i coeff[4];
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_0 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_1 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ const __m256i src_01 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
+ sx = sx4 + beta * (k + 4);
+ prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx, coeff);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i src_01 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
+ sx = sx4 + beta * (k + 4);
+ prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx, coeff);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+}
+
+static INLINE void warp_horizontal_filter_beta0_avx2(
+ const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
+ int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+ const __m256i *round_const, const __m128i *shift,
+ const __m256i *shuffle_src) {
+ (void)beta;
+ int k, iy, row = 0;
+ __m256i coeff[4];
+ prepare_horizontal_filter_coeff_beta0_avx2(alpha, sx4, coeff);
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_0 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src_1 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ const __m256i src_01 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(src_0), src_1, 0x1);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i src_01 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+}
+
+static INLINE void warp_horizontal_filter_alpha0_beta0_avx2(
+ const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
+ int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+ const __m256i *round_const, const __m128i *shift,
+ const __m256i *shuffle_src) {
+ (void)alpha;
+ int k, iy, row = 0;
+ __m256i coeff[4];
+ prepare_horizontal_filter_coeff_alpha0_avx2(beta, sx4, coeff);
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src0 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m128i src1 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ const __m256i src_01 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(src0), src1, 0x1);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i src_01 = _mm256_castsi128_si256(
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7)));
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src, round_const,
+ shift, row);
+}
+
+static INLINE void unpack_weights_and_set_round_const_avx2(
+ ConvolveParams *conv_params, const int round_bits, const int offset_bits,
+ __m256i *res_sub_const, __m256i *round_bits_const, __m256i *wt) {
+ *res_sub_const =
+ _mm256_set1_epi16(-(1 << (offset_bits - conv_params->round_1)) -
+ (1 << (offset_bits - conv_params->round_1 - 1)));
+ *round_bits_const = _mm256_set1_epi16(((1 << round_bits) >> 1));
+
+ const int w0 = conv_params->fwd_offset;
+ const int w1 = conv_params->bck_offset;
+ const __m256i wt0 = _mm256_set1_epi16(w0);
+ const __m256i wt1 = _mm256_set1_epi16(w1);
+ *wt = _mm256_unpacklo_epi16(wt0, wt1);
+}
+
+static INLINE void prepare_vertical_filter_coeffs_avx2(int gamma, int delta,
+ int sy,
+ __m256i *coeffs) {
+ __m128i filt_00 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_01 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_02 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_03 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+ __m128i filt_10 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 0 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ __m128i filt_11 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 2 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ __m128i filt_12 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 4 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ __m128i filt_13 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 6 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+
+ __m256i filt_0 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_00), filt_10, 0x1);
+ __m256i filt_1 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_01), filt_11, 0x1);
+ __m256i filt_2 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_02), filt_12, 0x1);
+ __m256i filt_3 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_03), filt_13, 0x1);
+
+ __m256i res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
+ __m256i res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
+ __m256i res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
+ __m256i res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
+
+ coeffs[0] = _mm256_unpacklo_epi64(res_0, res_1);
+ coeffs[1] = _mm256_unpackhi_epi64(res_0, res_1);
+ coeffs[2] = _mm256_unpacklo_epi64(res_2, res_3);
+ coeffs[3] = _mm256_unpackhi_epi64(res_2, res_3);
+
+ filt_00 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_01 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_02 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_03 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+ filt_10 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 1 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ filt_11 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 3 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ filt_12 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 5 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+ filt_13 =
+ _mm_loadu_si128((__m128i *)(warped_filter + (((sy + delta) + 7 * gamma) >>
+ WARPEDDIFF_PREC_BITS)));
+
+ filt_0 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_00), filt_10, 0x1);
+ filt_1 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_01), filt_11, 0x1);
+ filt_2 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_02), filt_12, 0x1);
+ filt_3 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_03), filt_13, 0x1);
+
+ res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
+ res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
+ res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
+ res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
+
+ coeffs[4] = _mm256_unpacklo_epi64(res_0, res_1);
+ coeffs[5] = _mm256_unpackhi_epi64(res_0, res_1);
+ coeffs[6] = _mm256_unpacklo_epi64(res_2, res_3);
+ coeffs[7] = _mm256_unpackhi_epi64(res_2, res_3);
+}
+
+static INLINE void prepare_vertical_filter_coeffs_delta0_avx2(int gamma, int sy,
+ __m256i *coeffs) {
+ __m128i filt_00 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_01 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_02 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ __m128i filt_03 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+ __m256i filt_0 = _mm256_broadcastsi128_si256(filt_00);
+ __m256i filt_1 = _mm256_broadcastsi128_si256(filt_01);
+ __m256i filt_2 = _mm256_broadcastsi128_si256(filt_02);
+ __m256i filt_3 = _mm256_broadcastsi128_si256(filt_03);
+
+ __m256i res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
+ __m256i res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
+ __m256i res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
+ __m256i res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
+
+ coeffs[0] = _mm256_unpacklo_epi64(res_0, res_1);
+ coeffs[1] = _mm256_unpackhi_epi64(res_0, res_1);
+ coeffs[2] = _mm256_unpacklo_epi64(res_2, res_3);
+ coeffs[3] = _mm256_unpackhi_epi64(res_2, res_3);
+
+ filt_00 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_01 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_02 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
+ filt_03 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+ filt_0 = _mm256_broadcastsi128_si256(filt_00);
+ filt_1 = _mm256_broadcastsi128_si256(filt_01);
+ filt_2 = _mm256_broadcastsi128_si256(filt_02);
+ filt_3 = _mm256_broadcastsi128_si256(filt_03);
+
+ res_0 = _mm256_unpacklo_epi32(filt_0, filt_1);
+ res_1 = _mm256_unpacklo_epi32(filt_2, filt_3);
+ res_2 = _mm256_unpackhi_epi32(filt_0, filt_1);
+ res_3 = _mm256_unpackhi_epi32(filt_2, filt_3);
+
+ coeffs[4] = _mm256_unpacklo_epi64(res_0, res_1);
+ coeffs[5] = _mm256_unpackhi_epi64(res_0, res_1);
+ coeffs[6] = _mm256_unpacklo_epi64(res_2, res_3);
+ coeffs[7] = _mm256_unpackhi_epi64(res_2, res_3);
+}
+
+static INLINE void prepare_vertical_filter_coeffs_gamma0_avx2(int delta, int sy,
+ __m256i *coeffs) {
+ const __m128i filt_0 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
+ const __m128i filt_1 = _mm_loadu_si128(
+ (__m128i *)(warped_filter + ((sy + delta) >> WARPEDDIFF_PREC_BITS)));
+
+ __m256i res_0 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(filt_0), filt_1, 0x1);
+
+ coeffs[0] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_gamma0_mask0_avx2));
+ coeffs[1] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_gamma0_mask1_avx2));
+ coeffs[2] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_gamma0_mask2_avx2));
+ coeffs[3] = _mm256_shuffle_epi8(
+ res_0, _mm256_loadu_si256((__m256i *)shuffle_gamma0_mask3_avx2));
+
+ coeffs[4] = coeffs[0];
+ coeffs[5] = coeffs[1];
+ coeffs[6] = coeffs[2];
+ coeffs[7] = coeffs[3];
+}
+
+static INLINE void filter_src_pixels_vertical_avx2(__m256i *horz_out,
+ __m256i *src,
+ __m256i *coeffs,
+ __m256i *res_lo,
+ __m256i *res_hi, int row) {
+ const __m256i src_6 = horz_out[row + 3];
+ const __m256i src_7 =
+ _mm256_permute2x128_si256(horz_out[row + 3], horz_out[row + 4], 0x21);
+
+ src[6] = _mm256_unpacklo_epi16(src_6, src_7);
+
+ const __m256i res_0 = _mm256_madd_epi16(src[0], coeffs[0]);
+ const __m256i res_2 = _mm256_madd_epi16(src[2], coeffs[1]);
+ const __m256i res_4 = _mm256_madd_epi16(src[4], coeffs[2]);
+ const __m256i res_6 = _mm256_madd_epi16(src[6], coeffs[3]);
+
+ const __m256i res_even = _mm256_add_epi32(_mm256_add_epi32(res_0, res_2),
+ _mm256_add_epi32(res_4, res_6));
+
+ src[7] = _mm256_unpackhi_epi16(src_6, src_7);
+
+ const __m256i res_1 = _mm256_madd_epi16(src[1], coeffs[4]);
+ const __m256i res_3 = _mm256_madd_epi16(src[3], coeffs[5]);
+ const __m256i res_5 = _mm256_madd_epi16(src[5], coeffs[6]);
+ const __m256i res_7 = _mm256_madd_epi16(src[7], coeffs[7]);
+
+ const __m256i res_odd = _mm256_add_epi32(_mm256_add_epi32(res_1, res_3),
+ _mm256_add_epi32(res_5, res_7));
+
+ // Rearrange pixels back into the order 0 ... 7
+ *res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
+ *res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
+}
+
+static INLINE void store_vertical_filter_output_avx2(
+ const __m256i *res_lo, const __m256i *res_hi, const __m256i *res_add_const,
+ const __m256i *wt, const __m256i *res_sub_const,
+ const __m256i *round_bits_const, uint8_t *pred, ConvolveParams *conv_params,
+ int i, int j, int k, const int reduce_bits_vert, int p_stride, int p_width,
+ const int round_bits) {
+ __m256i res_lo_1 = *res_lo;
+ __m256i res_hi_1 = *res_hi;
+
+ if (conv_params->is_compound) {
+ __m128i *const p_0 =
+ (__m128i *)&conv_params->dst[(i + k + 4) * conv_params->dst_stride + j];
+ __m128i *const p_1 =
+ (__m128i *)&conv_params
+ ->dst[(i + (k + 1) + 4) * conv_params->dst_stride + j];
+
+ res_lo_1 = _mm256_srai_epi32(_mm256_add_epi32(res_lo_1, *res_add_const),
+ reduce_bits_vert);
+
+ const __m256i temp_lo_16 = _mm256_packus_epi32(res_lo_1, res_lo_1);
+ __m256i res_lo_16;
+ if (conv_params->do_average) {
+ __m128i *const dst8_0 = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+ __m128i *const dst8_1 =
+ (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j];
+ const __m128i p_16_0 = _mm_loadl_epi64(p_0);
+ const __m128i p_16_1 = _mm_loadl_epi64(p_1);
+ const __m256i p_16 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(p_16_0), p_16_1, 1);
+ if (conv_params->use_dist_wtd_comp_avg) {
+ const __m256i p_16_lo = _mm256_unpacklo_epi16(p_16, temp_lo_16);
+ const __m256i wt_res_lo = _mm256_madd_epi16(p_16_lo, *wt);
+ const __m256i shifted_32 =
+ _mm256_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
+ res_lo_16 = _mm256_packus_epi32(shifted_32, shifted_32);
+ } else {
+ res_lo_16 = _mm256_srai_epi16(_mm256_add_epi16(p_16, temp_lo_16), 1);
+ }
+ res_lo_16 = _mm256_add_epi16(res_lo_16, *res_sub_const);
+ res_lo_16 = _mm256_srai_epi16(
+ _mm256_add_epi16(res_lo_16, *round_bits_const), round_bits);
+ const __m256i res_8_lo = _mm256_packus_epi16(res_lo_16, res_lo_16);
+ const __m128i res_8_lo_0 = _mm256_castsi256_si128(res_8_lo);
+ const __m128i res_8_lo_1 = _mm256_extracti128_si256(res_8_lo, 1);
+ *(uint32_t *)dst8_0 = _mm_cvtsi128_si32(res_8_lo_0);
+ *(uint32_t *)dst8_1 = _mm_cvtsi128_si32(res_8_lo_1);
+ } else {
+ const __m128i temp_lo_16_0 = _mm256_castsi256_si128(temp_lo_16);
+ const __m128i temp_lo_16_1 = _mm256_extracti128_si256(temp_lo_16, 1);
+ _mm_storel_epi64(p_0, temp_lo_16_0);
+ _mm_storel_epi64(p_1, temp_lo_16_1);
+ }
+ if (p_width > 4) {
+ __m128i *const p4_0 =
+ (__m128i *)&conv_params
+ ->dst[(i + k + 4) * conv_params->dst_stride + j + 4];
+ __m128i *const p4_1 =
+ (__m128i *)&conv_params
+ ->dst[(i + (k + 1) + 4) * conv_params->dst_stride + j + 4];
+ res_hi_1 = _mm256_srai_epi32(_mm256_add_epi32(res_hi_1, *res_add_const),
+ reduce_bits_vert);
+ const __m256i temp_hi_16 = _mm256_packus_epi32(res_hi_1, res_hi_1);
+ __m256i res_hi_16;
+ if (conv_params->do_average) {
+ __m128i *const dst8_4_0 =
+ (__m128i *)&pred[(i + k + 4) * p_stride + j + 4];
+ __m128i *const dst8_4_1 =
+ (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j + 4];
+ const __m128i p4_16_0 = _mm_loadl_epi64(p4_0);
+ const __m128i p4_16_1 = _mm_loadl_epi64(p4_1);
+ const __m256i p4_16 = _mm256_inserti128_si256(
+ _mm256_castsi128_si256(p4_16_0), p4_16_1, 1);
+ if (conv_params->use_dist_wtd_comp_avg) {
+ const __m256i p_16_hi = _mm256_unpacklo_epi16(p4_16, temp_hi_16);
+ const __m256i wt_res_hi = _mm256_madd_epi16(p_16_hi, *wt);
+ const __m256i shifted_32 =
+ _mm256_srai_epi32(wt_res_hi, DIST_PRECISION_BITS);
+ res_hi_16 = _mm256_packus_epi32(shifted_32, shifted_32);
+ } else {
+ res_hi_16 = _mm256_srai_epi16(_mm256_add_epi16(p4_16, temp_hi_16), 1);
+ }
+ res_hi_16 = _mm256_add_epi16(res_hi_16, *res_sub_const);
+ res_hi_16 = _mm256_srai_epi16(
+ _mm256_add_epi16(res_hi_16, *round_bits_const), round_bits);
+ __m256i res_8_hi = _mm256_packus_epi16(res_hi_16, res_hi_16);
+ const __m128i res_8_hi_0 = _mm256_castsi256_si128(res_8_hi);
+ const __m128i res_8_hi_1 = _mm256_extracti128_si256(res_8_hi, 1);
+ *(uint32_t *)dst8_4_0 = _mm_cvtsi128_si32(res_8_hi_0);
+ *(uint32_t *)dst8_4_1 = _mm_cvtsi128_si32(res_8_hi_1);
+ } else {
+ const __m128i temp_hi_16_0 = _mm256_castsi256_si128(temp_hi_16);
+ const __m128i temp_hi_16_1 = _mm256_extracti128_si256(temp_hi_16, 1);
+ _mm_storel_epi64(p4_0, temp_hi_16_0);
+ _mm_storel_epi64(p4_1, temp_hi_16_1);
+ }
+ }
+ } else {
+ const __m256i res_lo_round = _mm256_srai_epi32(
+ _mm256_add_epi32(res_lo_1, *res_add_const), reduce_bits_vert);
+ const __m256i res_hi_round = _mm256_srai_epi32(
+ _mm256_add_epi32(res_hi_1, *res_add_const), reduce_bits_vert);
+
+ const __m256i res_16bit = _mm256_packs_epi32(res_lo_round, res_hi_round);
+ const __m256i res_8bit = _mm256_packus_epi16(res_16bit, res_16bit);
+ const __m128i res_8bit0 = _mm256_castsi256_si128(res_8bit);
+ const __m128i res_8bit1 = _mm256_extracti128_si256(res_8bit, 1);
+
+ // Store, blending with 'pred' if needed
+ __m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+ __m128i *const p1 = (__m128i *)&pred[(i + (k + 1) + 4) * p_stride + j];
+
+ if (p_width == 4) {
+ *(uint32_t *)p = _mm_cvtsi128_si32(res_8bit0);
+ *(uint32_t *)p1 = _mm_cvtsi128_si32(res_8bit1);
+ } else {
+ _mm_storel_epi64(p, res_8bit0);
+ _mm_storel_epi64(p1, res_8bit1);
+ }
+ }
+}
+
+static INLINE void warp_vertical_filter_avx2(
+ uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
+ int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
+ int i, int j, int sy4, const int reduce_bits_vert,
+ const __m256i *res_add_const, const int round_bits,
+ const __m256i *res_sub_const, const __m256i *round_bits_const,
+ const __m256i *wt) {
+ int k, row = 0;
+ __m256i src[8];
+ const __m256i src_0 = horz_out[0];
+ const __m256i src_1 =
+ _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
+ const __m256i src_2 = horz_out[1];
+ const __m256i src_3 =
+ _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
+ const __m256i src_4 = horz_out[2];
+ const __m256i src_5 =
+ _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
+
+ src[0] = _mm256_unpacklo_epi16(src_0, src_1);
+ src[2] = _mm256_unpacklo_epi16(src_2, src_3);
+ src[4] = _mm256_unpacklo_epi16(src_4, src_5);
+
+ src[1] = _mm256_unpackhi_epi16(src_0, src_1);
+ src[3] = _mm256_unpackhi_epi16(src_2, src_3);
+ src[5] = _mm256_unpackhi_epi16(src_4, src_5);
+
+ for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
+ int sy = sy4 + delta * (k + 4);
+ __m256i coeffs[8];
+ prepare_vertical_filter_coeffs_avx2(gamma, delta, sy, coeffs);
+ __m256i res_lo, res_hi;
+ filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
+ row);
+ store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
+ res_sub_const, round_bits_const, pred,
+ conv_params, i, j, k, reduce_bits_vert,
+ p_stride, p_width, round_bits);
+ src[0] = src[2];
+ src[2] = src[4];
+ src[4] = src[6];
+ src[1] = src[3];
+ src[3] = src[5];
+ src[5] = src[7];
+
+ row += 1;
+ }
+}
+
+static INLINE void warp_vertical_filter_gamma0_avx2(
+ uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
+ int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
+ int i, int j, int sy4, const int reduce_bits_vert,
+ const __m256i *res_add_const, const int round_bits,
+ const __m256i *res_sub_const, const __m256i *round_bits_const,
+ const __m256i *wt) {
+ (void)gamma;
+ int k, row = 0;
+ __m256i src[8];
+ const __m256i src_0 = horz_out[0];
+ const __m256i src_1 =
+ _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
+ const __m256i src_2 = horz_out[1];
+ const __m256i src_3 =
+ _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
+ const __m256i src_4 = horz_out[2];
+ const __m256i src_5 =
+ _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
+
+ src[0] = _mm256_unpacklo_epi16(src_0, src_1);
+ src[2] = _mm256_unpacklo_epi16(src_2, src_3);
+ src[4] = _mm256_unpacklo_epi16(src_4, src_5);
+
+ src[1] = _mm256_unpackhi_epi16(src_0, src_1);
+ src[3] = _mm256_unpackhi_epi16(src_2, src_3);
+ src[5] = _mm256_unpackhi_epi16(src_4, src_5);
+
+ for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
+ int sy = sy4 + delta * (k + 4);
+ __m256i coeffs[8];
+ prepare_vertical_filter_coeffs_gamma0_avx2(delta, sy, coeffs);
+ __m256i res_lo, res_hi;
+ filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
+ row);
+ store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
+ res_sub_const, round_bits_const, pred,
+ conv_params, i, j, k, reduce_bits_vert,
+ p_stride, p_width, round_bits);
+ src[0] = src[2];
+ src[2] = src[4];
+ src[4] = src[6];
+ src[1] = src[3];
+ src[3] = src[5];
+ src[5] = src[7];
+ row += 1;
+ }
+}
+
+static INLINE void warp_vertical_filter_delta0_avx2(
+ uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
+ int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
+ int i, int j, int sy4, const int reduce_bits_vert,
+ const __m256i *res_add_const, const int round_bits,
+ const __m256i *res_sub_const, const __m256i *round_bits_const,
+ const __m256i *wt) {
+ (void)delta;
+ int k, row = 0;
+ __m256i src[8], coeffs[8];
+ const __m256i src_0 = horz_out[0];
+ const __m256i src_1 =
+ _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
+ const __m256i src_2 = horz_out[1];
+ const __m256i src_3 =
+ _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
+ const __m256i src_4 = horz_out[2];
+ const __m256i src_5 =
+ _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
+
+ src[0] = _mm256_unpacklo_epi16(src_0, src_1);
+ src[2] = _mm256_unpacklo_epi16(src_2, src_3);
+ src[4] = _mm256_unpacklo_epi16(src_4, src_5);
+
+ src[1] = _mm256_unpackhi_epi16(src_0, src_1);
+ src[3] = _mm256_unpackhi_epi16(src_2, src_3);
+ src[5] = _mm256_unpackhi_epi16(src_4, src_5);
+
+ prepare_vertical_filter_coeffs_delta0_avx2(gamma, sy4, coeffs);
+
+ for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
+ __m256i res_lo, res_hi;
+ filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
+ row);
+ store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
+ res_sub_const, round_bits_const, pred,
+ conv_params, i, j, k, reduce_bits_vert,
+ p_stride, p_width, round_bits);
+ src[0] = src[2];
+ src[2] = src[4];
+ src[4] = src[6];
+ src[1] = src[3];
+ src[3] = src[5];
+ src[5] = src[7];
+ row += 1;
+ }
+}
+
+static INLINE void warp_vertical_filter_gamma0_delta0_avx2(
+ uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
+ int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
+ int i, int j, int sy4, const int reduce_bits_vert,
+ const __m256i *res_add_const, const int round_bits,
+ const __m256i *res_sub_const, const __m256i *round_bits_const,
+ const __m256i *wt) {
+ (void)gamma;
+ int k, row = 0;
+ __m256i src[8], coeffs[8];
+ const __m256i src_0 = horz_out[0];
+ const __m256i src_1 =
+ _mm256_permute2x128_si256(horz_out[0], horz_out[1], 0x21);
+ const __m256i src_2 = horz_out[1];
+ const __m256i src_3 =
+ _mm256_permute2x128_si256(horz_out[1], horz_out[2], 0x21);
+ const __m256i src_4 = horz_out[2];
+ const __m256i src_5 =
+ _mm256_permute2x128_si256(horz_out[2], horz_out[3], 0x21);
+
+ src[0] = _mm256_unpacklo_epi16(src_0, src_1);
+ src[2] = _mm256_unpacklo_epi16(src_2, src_3);
+ src[4] = _mm256_unpacklo_epi16(src_4, src_5);
+
+ src[1] = _mm256_unpackhi_epi16(src_0, src_1);
+ src[3] = _mm256_unpackhi_epi16(src_2, src_3);
+ src[5] = _mm256_unpackhi_epi16(src_4, src_5);
+
+ prepare_vertical_filter_coeffs_gamma0_avx2(delta, sy4, coeffs);
+
+ for (k = -4; k < AOMMIN(4, p_height - i - 4); k += 2) {
+ __m256i res_lo, res_hi;
+ filter_src_pixels_vertical_avx2(horz_out, src, coeffs, &res_lo, &res_hi,
+ row);
+ store_vertical_filter_output_avx2(&res_lo, &res_hi, res_add_const, wt,
+ res_sub_const, round_bits_const, pred,
+ conv_params, i, j, k, reduce_bits_vert,
+ p_stride, p_width, round_bits);
+ src[0] = src[2];
+ src[2] = src[4];
+ src[4] = src[6];
+ src[1] = src[3];
+ src[3] = src[5];
+ src[5] = src[7];
+ row += 1;
+ }
+}
+
+static INLINE void prepare_warp_vertical_filter_avx2(
+ uint8_t *pred, __m256i *horz_out, ConvolveParams *conv_params,
+ int16_t gamma, int16_t delta, int p_height, int p_stride, int p_width,
+ int i, int j, int sy4, const int reduce_bits_vert,
+ const __m256i *res_add_const, const int round_bits,
+ const __m256i *res_sub_const, const __m256i *round_bits_const,
+ const __m256i *wt) {
+ if (gamma == 0 && delta == 0)
+ warp_vertical_filter_gamma0_delta0_avx2(
+ pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
+ i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
+ round_bits_const, wt);
+ else if (gamma == 0 && delta != 0)
+ warp_vertical_filter_gamma0_avx2(
+ pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
+ i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
+ round_bits_const, wt);
+ else if (gamma != 0 && delta == 0)
+ warp_vertical_filter_delta0_avx2(
+ pred, horz_out, conv_params, gamma, delta, p_height, p_stride, p_width,
+ i, j, sy4, reduce_bits_vert, res_add_const, round_bits, res_sub_const,
+ round_bits_const, wt);
+ else
+ warp_vertical_filter_avx2(pred, horz_out, conv_params, gamma, delta,
+ p_height, p_stride, p_width, i, j, sy4,
+ reduce_bits_vert, res_add_const, round_bits,
+ res_sub_const, round_bits_const, wt);
+}
+
+static INLINE void prepare_warp_horizontal_filter_avx2(
+ const uint8_t *ref, __m256i *horz_out, int stride, int32_t ix4, int32_t iy4,
+ int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+ const __m256i *round_const, const __m128i *shift,
+ const __m256i *shuffle_src) {
+ if (alpha == 0 && beta == 0)
+ warp_horizontal_filter_alpha0_beta0_avx2(
+ ref, horz_out, stride, ix4, iy4, sx4, alpha, beta, p_height, height, i,
+ round_const, shift, shuffle_src);
+ else if (alpha == 0 && beta != 0)
+ warp_horizontal_filter_alpha0_avx2(ref, horz_out, stride, ix4, iy4, sx4,
+ alpha, beta, p_height, height, i,
+ round_const, shift, shuffle_src);
+ else if (alpha != 0 && beta == 0)
+ warp_horizontal_filter_beta0_avx2(ref, horz_out, stride, ix4, iy4, sx4,
+ alpha, beta, p_height, height, i,
+ round_const, shift, shuffle_src);
+ else
+ warp_horizontal_filter_avx2(ref, horz_out, stride, ix4, iy4, sx4, alpha,
+ beta, p_height, height, i, round_const, shift,
+ shuffle_src);
+}
+
+void av1_warp_affine_avx2(const int32_t *mat, const uint8_t *ref, int width,
+ int height, int stride, uint8_t *pred, int p_col,
+ int p_row, int p_width, int p_height, int p_stride,
+ int subsampling_x, int subsampling_y,
+ ConvolveParams *conv_params, int16_t alpha,
+ int16_t beta, int16_t gamma, int16_t delta) {
+ __m256i horz_out[8];
+ int i, j, k;
+ const int bd = 8;
+ const int reduce_bits_horiz = conv_params->round_0;
+ const int reduce_bits_vert = conv_params->is_compound
+ ? conv_params->round_1
+ : 2 * FILTER_BITS - reduce_bits_horiz;
+ const int offset_bits_horiz = bd + FILTER_BITS - 1;
+ assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
+
+ const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
+ const __m256i reduce_bits_vert_const =
+ _mm256_set1_epi32(((1 << reduce_bits_vert) >> 1));
+ const __m256i res_add_const = _mm256_set1_epi32(1 << offset_bits_vert);
+ const int round_bits =
+ 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+ const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+ assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
+
+ const __m256i round_const = _mm256_set1_epi16(
+ (1 << offset_bits_horiz) + ((1 << reduce_bits_horiz) >> 1));
+ const __m128i shift = _mm_cvtsi32_si128(reduce_bits_horiz);
+
+ __m256i res_sub_const, round_bits_const, wt;
+ unpack_weights_and_set_round_const_avx2(conv_params, round_bits, offset_bits,
+ &res_sub_const, &round_bits_const,
+ &wt);
+
+ __m256i res_add_const_1;
+ if (conv_params->is_compound == 1) {
+ res_add_const_1 = _mm256_add_epi32(reduce_bits_vert_const, res_add_const);
+ } else {
+ res_add_const_1 = _mm256_set1_epi32(-(1 << (bd + reduce_bits_vert - 1)) +
+ ((1 << reduce_bits_vert) >> 1));
+ }
+ const int32_t const1 = alpha * (-4) + beta * (-4) +
+ (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+ (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+ const int32_t const2 = gamma * (-4) + delta * (-4) +
+ (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+ (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+ const int32_t const3 = ((1 << WARP_PARAM_REDUCE_BITS) - 1);
+ const int16_t const4 = (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1));
+ const int16_t const5 = (1 << (FILTER_BITS - reduce_bits_horiz));
+
+ __m256i shuffle_src[4];
+ shuffle_src[0] = _mm256_loadu_si256((__m256i *)shuffle_src0);
+ shuffle_src[1] = _mm256_loadu_si256((__m256i *)shuffle_src1);
+ shuffle_src[2] = _mm256_loadu_si256((__m256i *)shuffle_src2);
+ shuffle_src[3] = _mm256_loadu_si256((__m256i *)shuffle_src3);
+
+ for (i = 0; i < p_height; i += 8) {
+ for (j = 0; j < p_width; j += 8) {
+ const int32_t src_x = (p_col + j + 4) << subsampling_x;
+ const int32_t src_y = (p_row + i + 4) << subsampling_y;
+ const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
+ const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
+ const int32_t x4 = dst_x >> subsampling_x;
+ const int32_t y4 = dst_y >> subsampling_y;
+
+ int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+ int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+ int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+ int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+ // Add in all the constant terms, including rounding and offset
+ sx4 += const1;
+ sy4 += const2;
+
+ sx4 &= ~const3;
+ sy4 &= ~const3;
+
+ // Horizontal filter
+ // If the block is aligned such that, after clamping, every sample
+ // would be taken from the leftmost/rightmost column, then we can
+ // skip the expensive horizontal filter.
+
+ if (ix4 <= -7) {
+ int iy, row = 0;
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i temp_0 =
+ _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i temp_1 =
+ _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
+ horz_out[row] = _mm256_blend_epi32(temp_0, temp_1, 0xf0);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ horz_out[row] = _mm256_set1_epi16(const4 + ref[iy * stride] * const5);
+ } else if (ix4 >= width + 6) {
+ int iy, row = 0;
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i temp_0 = _mm256_set1_epi16(
+ const4 + ref[iy * stride + (width - 1)] * const5);
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ const __m256i temp_1 = _mm256_set1_epi16(
+ const4 + ref[iy * stride + (width - 1)] * const5);
+ horz_out[row] = _mm256_blend_epi32(temp_0, temp_1, 0xf0);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ horz_out[row] =
+ _mm256_set1_epi16(const4 + ref[iy * stride + (width - 1)] * const5);
+ } else if (((ix4 - 7) < 0) || ((ix4 + 9) > width)) {
+ const int out_of_boundary_left = -(ix4 - 6);
+ const int out_of_boundary_right = (ix4 + 8) - width;
+ int iy, sx, row = 0;
+ for (k = -7; k <= (AOMMIN(8, p_height - i) - 2); k += 2) {
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ __m128i src0 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ iy = iy4 + k + 1;
+ iy = clamp(iy, 0, height - 1);
+ __m128i src1 =
+ _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+
+ if (out_of_boundary_left >= 0) {
+ const __m128i shuffle_reg_left =
+ _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
+ src0 = _mm_shuffle_epi8(src0, shuffle_reg_left);
+ src1 = _mm_shuffle_epi8(src1, shuffle_reg_left);
+ }
+ if (out_of_boundary_right >= 0) {
+ const __m128i shuffle_reg_right = _mm_loadu_si128(
+ (__m128i *)warp_pad_right[out_of_boundary_right]);
+ src0 = _mm_shuffle_epi8(src0, shuffle_reg_right);
+ src1 = _mm_shuffle_epi8(src1, shuffle_reg_right);
+ }
+ sx = sx4 + beta * (k + 4);
+ const __m256i src_01 =
+ _mm256_inserti128_si256(_mm256_castsi128_si256(src0), src1, 0x1);
+ horizontal_filter_avx2(src_01, horz_out, sx, alpha, beta, row,
+ shuffle_src, &round_const, &shift);
+ row += 1;
+ }
+ iy = iy4 + k;
+ iy = clamp(iy, 0, height - 1);
+ __m128i src = _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+ if (out_of_boundary_left >= 0) {
+ const __m128i shuffle_reg_left =
+ _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
+ src = _mm_shuffle_epi8(src, shuffle_reg_left);
+ }
+ if (out_of_boundary_right >= 0) {
+ const __m128i shuffle_reg_right =
+ _mm_loadu_si128((__m128i *)warp_pad_right[out_of_boundary_right]);
+ src = _mm_shuffle_epi8(src, shuffle_reg_right);
+ }
+ sx = sx4 + beta * (k + 4);
+ const __m256i src_01 = _mm256_castsi128_si256(src);
+ __m256i coeff[4];
+ prepare_horizontal_filter_coeff(alpha, sx, coeff);
+ filter_src_pixels_avx2(src_01, horz_out, coeff, shuffle_src,
+ &round_const, &shift, row);
+ } else {
+ prepare_warp_horizontal_filter_avx2(
+ ref, horz_out, stride, ix4, iy4, sx4, alpha, beta, p_height, height,
+ i, &round_const, &shift, shuffle_src);
+ }
+
+ // Vertical filter
+ prepare_warp_vertical_filter_avx2(
+ pred, horz_out, conv_params, gamma, delta, p_height, p_stride,
+ p_width, i, j, sy4, reduce_bits_vert, &res_add_const_1, round_bits,
+ &res_sub_const, &round_bits_const, &wt);
+ }
+ }
+}
diff --git a/av1/common/x86/warp_plane_sse4.c b/av1/common/x86/warp_plane_sse4.c
index e847370..9530525 100644
--- a/av1/common/x86/warp_plane_sse4.c
+++ b/av1/common/x86/warp_plane_sse4.c
@@ -31,7 +31,7 @@
coefficients into the correct order more quickly.
*/
/* clang-format off */
-DECLARE_ALIGNED(8, static const int8_t,
+DECLARE_ALIGNED(8, const int8_t,
filter_8bit[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8]) = {
#if WARPEDPIXEL_PREC_BITS == 6
// [-1, 0)
diff --git a/test/warp_filter_test.cc b/test/warp_filter_test.cc
index d7b3ec9..cafb344 100644
--- a/test/warp_filter_test.cc
+++ b/test/warp_filter_test.cc
@@ -47,6 +47,12 @@
#endif // HAVE_SSE4_1
+#if HAVE_AVX2
+INSTANTIATE_TEST_CASE_P(
+ AVX2, AV1WarpFilterTest,
+ libaom_test::AV1WarpFilter::BuildParams(av1_warp_affine_avx2));
+#endif // HAVE_AVX2
+
#if HAVE_NEON
INSTANTIATE_TEST_CASE_P(
NEON, AV1WarpFilterTest,
