ref: 069eedb3a03ea42dfd79991b6fece4c283c648d8
parent: 26aebd77b8bff98fd25cf089a5f94be699e19d51
author: Jerome Jiang <jianj@google.com>
date: Tue Apr 25 11:29:43 EDT 2017
vp9: Neon optimization for denoiser. Add unit tests. Denoiser on Neon is 5x faster than C code. BUG=webm:1420 Change-Id: I805ab64f809ff2137354116be6213e7ec29c1dcb
--- a/test/vp9_denoiser_test.cc
+++ b/test/vp9_denoiser_test.cc
@@ -116,4 +116,19 @@
make_tuple(&vp9_denoiser_filter_sse2, BLOCK_64X32),
make_tuple(&vp9_denoiser_filter_sse2, BLOCK_64X64)));
#endif // HAVE_SSE2
+
+#if HAVE_NEON
+INSTANTIATE_TEST_CASE_P(
+ NEON, VP9DenoiserTest,
+ ::testing::Values(make_tuple(&vp9_denoiser_filter_neon, BLOCK_8X8),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_8X16),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_16X8),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_16X16),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_16X32),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_32X16),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_32X32),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_32X64),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_64X32),
+ make_tuple(&vp9_denoiser_filter_neon, BLOCK_64X64)));
+#endif
} // namespace
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -120,7 +120,7 @@
#
if (vpx_config("CONFIG_VP9_TEMPORAL_DENOISING") eq "yes") {add_proto qw/int vp9_denoiser_filter/, "const uint8_t *sig, int sig_stride, const uint8_t *mc_avg, int mc_avg_stride, uint8_t *avg, int avg_stride, int increase_denoising, BLOCK_SIZE bs, int motion_magnitude";
- specialize qw/vp9_denoiser_filter sse2/;
+ specialize qw/vp9_denoiser_filter neon sse2/;
}
if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {--- /dev/null
+++ b/vp9/encoder/arm/neon/vp9_denoiser_neon.c
@@ -1,0 +1,352 @@
+/*
+ * Copyright (c) 2017 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_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vpx_mem/vpx_mem.h"
+
+// Compute the sum of all pixel differences of this MB.
+static INLINE int horizontal_add_s8x16(const int8x16_t v_sum_diff_total) {+ const int16x8_t fe_dc_ba_98_76_54_32_10 = vpaddlq_s8(v_sum_diff_total);
+ const int32x4_t fedc_ba98_7654_3210 = vpaddlq_s16(fe_dc_ba_98_76_54_32_10);
+ const int64x2_t fedcba98_76543210 = vpaddlq_s32(fedc_ba98_7654_3210);
+ const int64x1_t x = vqadd_s64(vget_high_s64(fedcba98_76543210),
+ vget_low_s64(fedcba98_76543210));
+ const int sum_diff = vget_lane_s32(vreinterpret_s32_s64(x), 0);
+ return sum_diff;
+}
+
+// Denoise a 16x1 vector.
+static INLINE int8x16_t denoiser_16x1_neon(
+ const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y,
+ const uint8x16_t v_level1_threshold, const uint8x16_t v_level2_threshold,
+ const uint8x16_t v_level3_threshold, const uint8x16_t v_level1_adjustment,
+ const uint8x16_t v_delta_level_1_and_2,
+ const uint8x16_t v_delta_level_2_and_3, int8x16_t v_sum_diff_total) {+ const uint8x16_t v_sig = vld1q_u8(sig);
+ const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+ /* Calculate absolute difference and sign masks. */
+ const uint8x16_t v_abs_diff = vabdq_u8(v_sig, v_mc_running_avg_y);
+ const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
+ const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
+
+ /* Figure out which level that put us in. */
+ const uint8x16_t v_level1_mask = vcleq_u8(v_level1_threshold, v_abs_diff);
+ const uint8x16_t v_level2_mask = vcleq_u8(v_level2_threshold, v_abs_diff);
+ const uint8x16_t v_level3_mask = vcleq_u8(v_level3_threshold, v_abs_diff);
+
+ /* Calculate absolute adjustments for level 1, 2 and 3. */
+ const uint8x16_t v_level2_adjustment =
+ vandq_u8(v_level2_mask, v_delta_level_1_and_2);
+ const uint8x16_t v_level3_adjustment =
+ vandq_u8(v_level3_mask, v_delta_level_2_and_3);
+ const uint8x16_t v_level1and2_adjustment =
+ vaddq_u8(v_level1_adjustment, v_level2_adjustment);
+ const uint8x16_t v_level1and2and3_adjustment =
+ vaddq_u8(v_level1and2_adjustment, v_level3_adjustment);
+
+ /* Figure adjustment absolute value by selecting between the absolute
+ * difference if in level0 or the value for level 1, 2 and 3.
+ */
+ const uint8x16_t v_abs_adjustment =
+ vbslq_u8(v_level1_mask, v_level1and2and3_adjustment, v_abs_diff);
+
+ /* Calculate positive and negative adjustments. Apply them to the signal
+ * and accumulate them. Adjustments are less than eight and the maximum
+ * sum of them (7 * 16) can fit in a signed char.
+ */
+ const uint8x16_t v_pos_adjustment =
+ vandq_u8(v_diff_pos_mask, v_abs_adjustment);
+ const uint8x16_t v_neg_adjustment =
+ vandq_u8(v_diff_neg_mask, v_abs_adjustment);
+
+ uint8x16_t v_running_avg_y = vqaddq_u8(v_sig, v_pos_adjustment);
+ v_running_avg_y = vqsubq_u8(v_running_avg_y, v_neg_adjustment);
+
+ /* Store results. */
+ vst1q_u8(running_avg_y, v_running_avg_y);
+
+ /* Sum all the accumulators to have the sum of all pixel differences
+ * for this macroblock.
+ */
+ {+ const int8x16_t v_sum_diff =
+ vqsubq_s8(vreinterpretq_s8_u8(v_pos_adjustment),
+ vreinterpretq_s8_u8(v_neg_adjustment));
+ v_sum_diff_total = vaddq_s8(v_sum_diff_total, v_sum_diff);
+ }
+ return v_sum_diff_total;
+}
+
+static INLINE int8x16_t denoiser_adjust_16x1_neon(
+ const uint8_t *sig, const uint8_t *mc_running_avg_y, uint8_t *running_avg_y,
+ const uint8x16_t k_delta, int8x16_t v_sum_diff_total) {+ uint8x16_t v_running_avg_y = vld1q_u8(running_avg_y);
+ const uint8x16_t v_sig = vld1q_u8(sig);
+ const uint8x16_t v_mc_running_avg_y = vld1q_u8(mc_running_avg_y);
+
+ /* Calculate absolute difference and sign masks. */
+ const uint8x16_t v_abs_diff = vabdq_u8(v_sig, v_mc_running_avg_y);
+ const uint8x16_t v_diff_pos_mask = vcltq_u8(v_sig, v_mc_running_avg_y);
+ const uint8x16_t v_diff_neg_mask = vcgtq_u8(v_sig, v_mc_running_avg_y);
+ // Clamp absolute difference to delta to get the adjustment.
+ const uint8x16_t v_abs_adjustment = vminq_u8(v_abs_diff, (k_delta));
+
+ const uint8x16_t v_pos_adjustment =
+ vandq_u8(v_diff_pos_mask, v_abs_adjustment);
+ const uint8x16_t v_neg_adjustment =
+ vandq_u8(v_diff_neg_mask, v_abs_adjustment);
+
+ v_running_avg_y = vqsubq_u8(v_running_avg_y, v_pos_adjustment);
+ v_running_avg_y = vqaddq_u8(v_running_avg_y, v_neg_adjustment);
+
+ /* Store results. */
+ vst1q_u8(running_avg_y, v_running_avg_y);
+
+ {+ const int8x16_t v_sum_diff =
+ vqsubq_s8(vreinterpretq_s8_u8(v_neg_adjustment),
+ vreinterpretq_s8_u8(v_pos_adjustment));
+ v_sum_diff_total = vaddq_s8(v_sum_diff_total, v_sum_diff);
+ }
+ return v_sum_diff_total;
+}
+
+// Denoise 8x8 and 8x16 blocks.
+static int vp9_denoiser_8xN_neon(const uint8_t *sig, int sig_stride,
+ const uint8_t *mc_running_avg_y,
+ int mc_avg_y_stride, uint8_t *running_avg_y,
+ int avg_y_stride, int increase_denoising,
+ BLOCK_SIZE bs, int motion_magnitude,
+ int width) {+ int sum_diff_thresh, r, sum_diff = 0;
+ const int shift_inc =
+ (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
+ ? 1
+ : 0;
+ uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16];
+
+ const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+ (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 + shift_inc : 3);
+ const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+ const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+ const uint8x16_t v_level1_threshold = vdupq_n_u8(4 + shift_inc);
+ const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+ const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+
+ const int b_height = (4 << b_height_log2_lookup[bs]) >> 1;
+
+ int8x16_t v_sum_diff_total = vdupq_n_s8(0);
+
+ for (r = 0; r < b_height; ++r) {+ memcpy(sig_buffer[r], sig, width);
+ memcpy(sig_buffer[r] + width, sig + sig_stride, width);
+ memcpy(mc_running_buffer[r], mc_running_avg_y, width);
+ memcpy(mc_running_buffer[r] + width, mc_running_avg_y + mc_avg_y_stride,
+ width);
+ memcpy(running_buffer[r], running_avg_y, width);
+ memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width);
+ v_sum_diff_total = denoiser_16x1_neon(
+ sig_buffer[r], mc_running_buffer[r], running_buffer[r],
+ v_level1_threshold, v_level2_threshold, v_level3_threshold,
+ v_level1_adjustment, v_delta_level_1_and_2, v_delta_level_2_and_3,
+ v_sum_diff_total);
+ {+ const uint8x16_t v_running_buffer = vld1q_u8(running_buffer[r]);
+ const uint8x8_t v_running_buffer_high = vget_high_u8(v_running_buffer);
+ const uint8x8_t v_running_buffer_low = vget_low_u8(v_running_buffer);
+ vst1_u8(running_avg_y, v_running_buffer_low);
+ vst1_u8(running_avg_y + avg_y_stride, v_running_buffer_high);
+ }
+ // Update pointers for next iteration.
+ sig += (sig_stride << 1);
+ mc_running_avg_y += (mc_avg_y_stride << 1);
+ running_avg_y += (avg_y_stride << 1);
+ }
+
+ {+ sum_diff = horizontal_add_s8x16(v_sum_diff_total);
+ sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+ if (abs(sum_diff) > sum_diff_thresh) {+ // Before returning to copy the block (i.e., apply no denoising),
+ // check if we can still apply some (weaker) temporal filtering to
+ // this block, that would otherwise not be denoised at all. Simplest
+ // is to apply an additional adjustment to running_avg_y to bring it
+ // closer to sig. The adjustment is capped by a maximum delta, and
+ // chosen such that in most cases the resulting sum_diff will be
+ // within the acceptable range given by sum_diff_thresh.
+
+ // The delta is set by the excess of absolute pixel diff over the
+ // threshold.
+ const int delta =
+ ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1;
+ // Only apply the adjustment for max delta up to 3.
+ if (delta < 4) {+ const uint8x16_t k_delta = vmovq_n_u8(delta);
+ running_avg_y -= avg_y_stride * (b_height << 1);
+ for (r = 0; r < b_height; ++r) {+ v_sum_diff_total = denoiser_adjust_16x1_neon(
+ sig_buffer[r], mc_running_buffer[r], running_buffer[r], k_delta,
+ v_sum_diff_total);
+ {+ const uint8x16_t v_running_buffer = vld1q_u8(running_buffer[r]);
+ const uint8x8_t v_running_buffer_high =
+ vget_high_u8(v_running_buffer);
+ const uint8x8_t v_running_buffer_low =
+ vget_low_u8(v_running_buffer);
+ vst1_u8(running_avg_y, v_running_buffer_low);
+ vst1_u8(running_avg_y + avg_y_stride, v_running_buffer_high);
+ }
+ // Update pointers for next iteration.
+ running_avg_y += (avg_y_stride << 1);
+ }
+ sum_diff = horizontal_add_s8x16(v_sum_diff_total);
+ if (abs(sum_diff) > sum_diff_thresh) {+ return COPY_BLOCK;
+ }
+ } else {+ return COPY_BLOCK;
+ }
+ }
+ }
+
+ return FILTER_BLOCK;
+}
+
+// Denoise 16x16, 16x32, 32x16, 32x32, 32x64, 64x32 and 64x64 blocks.
+static int vp9_denoiser_NxM_neon(const uint8_t *sig, int sig_stride,
+ const uint8_t *mc_running_avg_y,
+ int mc_avg_y_stride, uint8_t *running_avg_y,
+ int avg_y_stride, int increase_denoising,
+ BLOCK_SIZE bs, int motion_magnitude) {+ const int shift_inc =
+ (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
+ ? 1
+ : 0;
+ const uint8x16_t v_level1_adjustment = vmovq_n_u8(
+ (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 4 + shift_inc : 3);
+ const uint8x16_t v_delta_level_1_and_2 = vdupq_n_u8(1);
+ const uint8x16_t v_delta_level_2_and_3 = vdupq_n_u8(2);
+ const uint8x16_t v_level1_threshold = vmovq_n_u8(4 + shift_inc);
+ const uint8x16_t v_level2_threshold = vdupq_n_u8(8);
+ const uint8x16_t v_level3_threshold = vdupq_n_u8(16);
+
+ const int b_width = (4 << b_width_log2_lookup[bs]);
+ const int b_height = (4 << b_height_log2_lookup[bs]);
+ const int b_width_shift4 = b_width >> 4;
+
+ int8x16_t v_sum_diff_total[4][4];
+ int r, c, sum_diff = 0;
+
+ for (r = 0; r < 4; ++r) {+ for (c = 0; c < b_width_shift4; ++c) {+ v_sum_diff_total[c][r] = vdupq_n_s8(0);
+ }
+ }
+
+ for (r = 0; r < b_height; ++r) {+ for (c = 0; c < b_width_shift4; ++c) {+ v_sum_diff_total[c][r >> 4] = denoiser_16x1_neon(
+ sig, mc_running_avg_y, running_avg_y, v_level1_threshold,
+ v_level2_threshold, v_level3_threshold, v_level1_adjustment,
+ v_delta_level_1_and_2, v_delta_level_2_and_3,
+ v_sum_diff_total[c][r >> 4]);
+
+ // Update pointers for next iteration.
+ sig += 16;
+ mc_running_avg_y += 16;
+ running_avg_y += 16;
+ }
+
+ if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) {+ for (c = 0; c < b_width_shift4; ++c) {+ sum_diff += horizontal_add_s8x16(v_sum_diff_total[c][r >> 4]);
+ }
+ }
+
+ // Update pointers for next iteration.
+ sig = sig - b_width + sig_stride;
+ mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride;
+ running_avg_y = running_avg_y - b_width + avg_y_stride;
+ }
+
+ {+ const int sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+ if (abs(sum_diff) > sum_diff_thresh) {+ const int delta =
+ ((abs(sum_diff) - sum_diff_thresh) >> num_pels_log2_lookup[bs]) + 1;
+ // Only apply the adjustment for max delta up to 3.
+ if (delta < 4) {+ const uint8x16_t k_delta = vdupq_n_u8(delta);
+ sig -= sig_stride * b_height;
+ mc_running_avg_y -= mc_avg_y_stride * b_height;
+ running_avg_y -= avg_y_stride * b_height;
+ sum_diff = 0;
+
+ for (r = 0; r < b_height; ++r) {+ for (c = 0; c < b_width_shift4; ++c) {+ v_sum_diff_total[c][r >> 4] =
+ denoiser_adjust_16x1_neon(sig, mc_running_avg_y, running_avg_y,
+ k_delta, v_sum_diff_total[c][r >> 4]);
+
+ // Update pointers for next iteration.
+ sig += 16;
+ mc_running_avg_y += 16;
+ running_avg_y += 16;
+ }
+ if ((r & 0xf) == 0xf || (bs == BLOCK_16X8 && r == 7)) {+ for (c = 0; c < b_width_shift4; ++c) {+ sum_diff += horizontal_add_s8x16(v_sum_diff_total[c][r >> 4]);
+ }
+ }
+
+ sig = sig - b_width + sig_stride;
+ mc_running_avg_y = mc_running_avg_y - b_width + mc_avg_y_stride;
+ running_avg_y = running_avg_y - b_width + avg_y_stride;
+ }
+
+ if (abs(sum_diff) > sum_diff_thresh) {+ return COPY_BLOCK;
+ }
+ } else {+ return COPY_BLOCK;
+ }
+ }
+ }
+ return FILTER_BLOCK;
+}
+
+int vp9_denoiser_filter_neon(const uint8_t *sig, int sig_stride,
+ const uint8_t *mc_avg, int mc_avg_stride,
+ uint8_t *avg, int avg_stride,
+ int increase_denoising, BLOCK_SIZE bs,
+ int motion_magnitude) {+ // Rank by frequency of the block type to have an early termination.
+ if (bs == BLOCK_16X16 || bs == BLOCK_32X32 || bs == BLOCK_64X64 ||
+ bs == BLOCK_16X32 || bs == BLOCK_16X8 || bs == BLOCK_32X16 ||
+ bs == BLOCK_32X64 || bs == BLOCK_64X32) {+ return vp9_denoiser_NxM_neon(sig, sig_stride, mc_avg, mc_avg_stride, avg,
+ avg_stride, increase_denoising, bs,
+ motion_magnitude);
+ } else if (bs == BLOCK_8X8 || bs == BLOCK_8X16) {+ return vp9_denoiser_8xN_neon(sig, sig_stride, mc_avg, mc_avg_stride, avg,
+ avg_stride, increase_denoising, bs,
+ motion_magnitude, 8);
+ }
+ return COPY_BLOCK;
+}
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -120,6 +120,7 @@
ifeq ($(CONFIG_VP9_TEMPORAL_DENOISING),yes)
VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_denoiser_sse2.c
+VP9_CX_SRCS-$(HAVE_NEON) += encoder/arm/neon/vp9_denoiser_neon.c
endif
VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_error_intrin_avx2.c