vpx/vp9/encoder/x86/vp9_denoiser_sse2.c
JackyChen b29612fbbe vp9: Refactor vp9_denoiser_NxM_sse2.
Denoiser is ~1.5% faster in speed 6~8.

Change-Id: I7b350f3c50cce6773d9c4eded4c0c1b722d0a5fc
2016-05-02 13:19:40 -07:00

342 lines
14 KiB
C

/*
* Copyright (c) 2014 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 <emmintrin.h>
#include "./vpx_config.h"
#include "./vp9_rtcd.h"
#include "vpx_ports/emmintrin_compat.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 sum_diff_16x1(__m128i acc_diff) {
const __m128i k_1 = _mm_set1_epi16(1);
const __m128i acc_diff_lo =
_mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
const __m128i acc_diff_hi =
_mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
const __m128i hgfe_dcba =
_mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
const __m128i hgfedcba =
_mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
return _mm_cvtsi128_si32(hgfedcba);
}
// Denoise a 16x1 vector.
static INLINE __m128i vp9_denoiser_16x1_sse2(const uint8_t *sig,
const uint8_t *mc_running_avg_y,
uint8_t *running_avg_y,
const __m128i *k_0,
const __m128i *k_4,
const __m128i *k_8,
const __m128i *k_16,
const __m128i *l3,
const __m128i *l32,
const __m128i *l21,
__m128i acc_diff) {
// Calculate differences
const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y =
_mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
__m128i v_running_avg_y;
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
// Obtain the sign. FF if diff is negative.
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0);
// Clamp absolute difference to 16 to be used to get mask. Doing this
// allows us to use _mm_cmpgt_epi8, which operates on signed byte.
const __m128i clamped_absdiff =
_mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16);
// Get masks for l2 l1 and l0 adjustments.
const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff);
const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff);
const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff);
// Get adjustments for l2, l1, and l0.
__m128i adj2 = _mm_and_si128(mask2, *l32);
const __m128i adj1 = _mm_and_si128(mask1, *l21);
const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
__m128i adj, padj, nadj;
// Combine the adjustments and get absolute adjustments.
adj2 = _mm_add_epi8(adj2, adj1);
adj = _mm_sub_epi8(*l3, adj2);
adj = _mm_andnot_si128(mask0, adj);
adj = _mm_or_si128(adj, adj0);
// Restore the sign and get positive and negative adjustments.
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
// Calculate filtered value.
v_running_avg_y = _mm_adds_epu8(v_sig, padj);
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
// Adjustments <=7, and each element in acc_diff can fit in signed
// char.
acc_diff = _mm_adds_epi8(acc_diff, padj);
acc_diff = _mm_subs_epi8(acc_diff, nadj);
return acc_diff;
}
// Denoise a 16x1 vector with a weaker filter.
static INLINE __m128i vp9_denoiser_adj_16x1_sse2(
const uint8_t *sig, const uint8_t *mc_running_avg_y,
uint8_t *running_avg_y, const __m128i k_0,
const __m128i k_delta, __m128i acc_diff) {
__m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
// Calculate differences.
const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
const __m128i v_mc_running_avg_y =
_mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
// Obtain the sign. FF if diff is negative.
const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
// Clamp absolute difference to delta to get the adjustment.
const __m128i adj =
_mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
// Restore the sign and get positive and negative adjustments.
__m128i padj, nadj;
padj = _mm_andnot_si128(diff_sign, adj);
nadj = _mm_and_si128(diff_sign, adj);
// Calculate filtered value.
v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
_mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
// Accumulate the adjustments.
acc_diff = _mm_subs_epi8(acc_diff, padj);
acc_diff = _mm_adds_epi8(acc_diff, nadj);
return acc_diff;
}
// Denoise 8x8 and 8x16 blocks.
static int vp9_denoiser_NxM_sse2_small(
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];
__m128i acc_diff = _mm_setzero_si128();
const __m128i k_0 = _mm_setzero_si128();
const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
const __m128i k_8 = _mm_set1_epi8(8);
const __m128i k_16 = _mm_set1_epi8(16);
// Modify each level's adjustment according to motion_magnitude.
const __m128i l3 = _mm_set1_epi8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
// Difference between level 3 and level 2 is 2.
const __m128i l32 = _mm_set1_epi8(2);
// Difference between level 2 and level 1 is 1.
const __m128i l21 = _mm_set1_epi8(1);
const int b_height = (4 << b_height_log2_lookup[bs]) >> 1;
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);
acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
mc_running_buffer[r],
running_buffer[r],
&k_0, &k_4, &k_8, &k_16,
&l3, &l32, &l21, acc_diff);
memcpy(running_avg_y, running_buffer[r], width);
memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width);
// 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 = sum_diff_16x1(acc_diff);
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 __m128i k_delta = _mm_set1_epi8(delta);
running_avg_y -= avg_y_stride * (b_height << 1);
for (r = 0; r < b_height; ++r) {
acc_diff = vp9_denoiser_adj_16x1_sse2(
sig_buffer[r], mc_running_buffer[r], running_buffer[r],
k_0, k_delta, acc_diff);
memcpy(running_avg_y, running_buffer[r], width);
memcpy(running_avg_y + avg_y_stride,
running_buffer[r] + width, width);
// Update pointers for next iteration.
running_avg_y += (avg_y_stride << 1);
}
sum_diff = sum_diff_16x1(acc_diff);
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_sse2_big(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 sum_diff_thresh, r, c, sum_diff = 0;
const int shift_inc = (increase_denoising &&
motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
1 : 0;
__m128i acc_diff[4][4];
const __m128i k_0 = _mm_setzero_si128();
const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
const __m128i k_8 = _mm_set1_epi8(8);
const __m128i k_16 = _mm_set1_epi8(16);
// Modify each level's adjustment according to motion_magnitude.
const __m128i l3 = _mm_set1_epi8(
(motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
// Difference between level 3 and level 2 is 2.
const __m128i l32 = _mm_set1_epi8(2);
// Difference between level 2 and level 1 is 1.
const __m128i l21 = _mm_set1_epi8(1);
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;
for (r = 0; r < 4; ++r) {
for (c = 0; c < b_width_shift4; ++c) {
acc_diff[c][r] = _mm_setzero_si128();
}
}
for (r = 0; r < b_height; ++r) {
for (c = 0; c < b_width_shift4; ++c) {
acc_diff[c][r>>4] = vp9_denoiser_16x1_sse2(
sig, mc_running_avg_y, running_avg_y, &k_0, &k_4,
&k_8, &k_16, &l3, &l32, &l21, acc_diff[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 += sum_diff_16x1(acc_diff[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;
}
{
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 __m128i k_delta = _mm_set1_epi8(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) {
acc_diff[c][r>>4] = vp9_denoiser_adj_16x1_sse2(
sig, mc_running_avg_y, running_avg_y, k_0,
k_delta, acc_diff[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 += sum_diff_16x1(acc_diff[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_sse2(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_sse2_big(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_NxM_sse2_small(sig, sig_stride,
mc_avg, mc_avg_stride,
avg, avg_stride,
increase_denoising,
bs, motion_magnitude, 8);
} else {
return COPY_BLOCK;
}
}