Line data Source code
1 : /*
2 : * Copyright (c) 2015 The WebRTC project authors. All Rights Reserved.
3 : *
4 : * Use of this source code is governed by a BSD-style license
5 : * that can be found in the LICENSE file in the root of the source
6 : * tree. An additional intellectual property rights grant can be found
7 : * in the file PATENTS. All contributing project authors may
8 : * be found in the AUTHORS file in the root of the source tree.
9 : */
10 :
11 : #include <emmintrin.h>
12 : #include "webrtc/modules/video_processing/util/denoiser_filter_sse2.h"
13 :
14 : namespace webrtc {
15 :
16 0 : static void Get8x8varSse2(const uint8_t* src,
17 : int src_stride,
18 : const uint8_t* ref,
19 : int ref_stride,
20 : unsigned int* sse,
21 : int* sum) {
22 0 : const __m128i zero = _mm_setzero_si128();
23 0 : __m128i vsum = _mm_setzero_si128();
24 0 : __m128i vsse = _mm_setzero_si128();
25 :
26 0 : for (int i = 0; i < 8; i += 2) {
27 0 : const __m128i src0 = _mm_unpacklo_epi8(
28 0 : _mm_loadl_epi64((const __m128i*)(src + i * src_stride)), zero);
29 0 : const __m128i ref0 = _mm_unpacklo_epi8(
30 0 : _mm_loadl_epi64((const __m128i*)(ref + i * ref_stride)), zero);
31 0 : const __m128i diff0 = _mm_sub_epi16(src0, ref0);
32 :
33 0 : const __m128i src1 = _mm_unpacklo_epi8(
34 0 : _mm_loadl_epi64((const __m128i*)(src + (i + 1) * src_stride)), zero);
35 0 : const __m128i ref1 = _mm_unpacklo_epi8(
36 0 : _mm_loadl_epi64((const __m128i*)(ref + (i + 1) * ref_stride)), zero);
37 0 : const __m128i diff1 = _mm_sub_epi16(src1, ref1);
38 :
39 0 : vsum = _mm_add_epi16(vsum, diff0);
40 0 : vsum = _mm_add_epi16(vsum, diff1);
41 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
42 0 : vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
43 : }
44 :
45 : // sum
46 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
47 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
48 0 : vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
49 0 : *sum = static_cast<int16_t>(_mm_extract_epi16(vsum, 0));
50 :
51 : // sse
52 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8));
53 0 : vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4));
54 0 : *sse = _mm_cvtsi128_si32(vsse);
55 0 : }
56 :
57 0 : static void VarianceSSE2(const unsigned char* src,
58 : int src_stride,
59 : const unsigned char* ref,
60 : int ref_stride,
61 : int w,
62 : int h,
63 : uint32_t* sse,
64 : int64_t* sum,
65 : int block_size) {
66 0 : *sse = 0;
67 0 : *sum = 0;
68 :
69 0 : for (int i = 0; i < h; i += block_size) {
70 0 : for (int j = 0; j < w; j += block_size) {
71 0 : uint32_t sse0 = 0;
72 0 : int32_t sum0 = 0;
73 :
74 0 : Get8x8varSse2(src + src_stride * i + j, src_stride,
75 0 : ref + ref_stride * i + j, ref_stride, &sse0, &sum0);
76 0 : *sse += sse0;
77 0 : *sum += sum0;
78 : }
79 : }
80 0 : }
81 :
82 : // Compute the sum of all pixel differences of this MB.
83 0 : static uint32_t AbsSumDiff16x1(__m128i acc_diff) {
84 0 : const __m128i k_1 = _mm_set1_epi16(1);
85 : const __m128i acc_diff_lo =
86 0 : _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
87 : const __m128i acc_diff_hi =
88 0 : _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
89 0 : const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
90 0 : const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
91 : const __m128i hgfe_dcba =
92 0 : _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
93 : const __m128i hgfedcba =
94 0 : _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
95 0 : unsigned int sum_diff = abs(_mm_cvtsi128_si32(hgfedcba));
96 :
97 0 : return sum_diff;
98 : }
99 :
100 : // TODO(jackychen): Optimize this function using SSE2.
101 0 : void DenoiserFilterSSE2::CopyMem16x16(const uint8_t* src,
102 : int src_stride,
103 : uint8_t* dst,
104 : int dst_stride) {
105 0 : for (int i = 0; i < 16; i++) {
106 0 : memcpy(dst, src, 16);
107 0 : src += src_stride;
108 0 : dst += dst_stride;
109 : }
110 0 : }
111 :
112 0 : uint32_t DenoiserFilterSSE2::Variance16x8(const uint8_t* src,
113 : int src_stride,
114 : const uint8_t* ref,
115 : int ref_stride,
116 : uint32_t* sse) {
117 0 : int64_t sum = 0;
118 0 : VarianceSSE2(src, src_stride << 1, ref, ref_stride << 1, 16, 8, sse, &sum, 8);
119 0 : return *sse - ((sum * sum) >> 7);
120 : }
121 :
122 0 : DenoiserDecision DenoiserFilterSSE2::MbDenoise(const uint8_t* mc_running_avg_y,
123 : int mc_avg_y_stride,
124 : uint8_t* running_avg_y,
125 : int avg_y_stride,
126 : const uint8_t* sig,
127 : int sig_stride,
128 : uint8_t motion_magnitude,
129 : int increase_denoising) {
130 0 : DenoiserDecision decision = FILTER_BLOCK;
131 0 : unsigned int sum_diff_thresh = 0;
132 : int shift_inc =
133 0 : (increase_denoising && motion_magnitude <= kMotionMagnitudeThreshold) ? 1
134 0 : : 0;
135 0 : __m128i acc_diff = _mm_setzero_si128();
136 0 : const __m128i k_0 = _mm_setzero_si128();
137 0 : const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
138 0 : const __m128i k_8 = _mm_set1_epi8(8);
139 0 : const __m128i k_16 = _mm_set1_epi8(16);
140 : // Modify each level's adjustment according to motion_magnitude.
141 0 : const __m128i l3 = _mm_set1_epi8(
142 0 : (motion_magnitude <= kMotionMagnitudeThreshold) ? 7 + shift_inc : 6);
143 : // Difference between level 3 and level 2 is 2.
144 0 : const __m128i l32 = _mm_set1_epi8(2);
145 : // Difference between level 2 and level 1 is 1.
146 0 : const __m128i l21 = _mm_set1_epi8(1);
147 :
148 0 : for (int r = 0; r < 16; ++r) {
149 : // Calculate differences.
150 : const __m128i v_sig =
151 0 : _mm_loadu_si128(reinterpret_cast<const __m128i*>(&sig[0]));
152 : const __m128i v_mc_running_avg_y =
153 0 : _mm_loadu_si128(reinterpret_cast<const __m128i*>(&mc_running_avg_y[0]));
154 : __m128i v_running_avg_y;
155 0 : const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
156 0 : const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
157 : // Obtain the sign. FF if diff is negative.
158 0 : const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
159 : // Clamp absolute difference to 16 to be used to get mask. Doing this
160 : // allows us to use _mm_cmpgt_epi8, which operates on signed byte.
161 : const __m128i clamped_absdiff =
162 0 : _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
163 : // Get masks for l2 l1 and l0 adjustments.
164 0 : const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
165 0 : const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
166 0 : const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
167 : // Get adjustments for l2, l1, and l0.
168 0 : __m128i adj2 = _mm_and_si128(mask2, l32);
169 0 : const __m128i adj1 = _mm_and_si128(mask1, l21);
170 0 : const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
171 : __m128i adj, padj, nadj;
172 :
173 : // Combine the adjustments and get absolute adjustments.
174 0 : adj2 = _mm_add_epi8(adj2, adj1);
175 0 : adj = _mm_sub_epi8(l3, adj2);
176 0 : adj = _mm_andnot_si128(mask0, adj);
177 0 : adj = _mm_or_si128(adj, adj0);
178 :
179 : // Restore the sign and get positive and negative adjustments.
180 0 : padj = _mm_andnot_si128(diff_sign, adj);
181 0 : nadj = _mm_and_si128(diff_sign, adj);
182 :
183 : // Calculate filtered value.
184 0 : v_running_avg_y = _mm_adds_epu8(v_sig, padj);
185 0 : v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
186 : _mm_storeu_si128(reinterpret_cast<__m128i*>(running_avg_y),
187 : v_running_avg_y);
188 :
189 : // Adjustments <=7, and each element in acc_diff can fit in signed
190 : // char.
191 0 : acc_diff = _mm_adds_epi8(acc_diff, padj);
192 0 : acc_diff = _mm_subs_epi8(acc_diff, nadj);
193 :
194 : // Update pointers for next iteration.
195 0 : sig += sig_stride;
196 0 : mc_running_avg_y += mc_avg_y_stride;
197 0 : running_avg_y += avg_y_stride;
198 : }
199 :
200 : // Compute the sum of all pixel differences of this MB.
201 0 : unsigned int abs_sum_diff = AbsSumDiff16x1(acc_diff);
202 0 : sum_diff_thresh =
203 0 : increase_denoising ? kSumDiffThresholdHigh : kSumDiffThreshold;
204 0 : if (abs_sum_diff > sum_diff_thresh)
205 0 : decision = COPY_BLOCK;
206 0 : return decision;
207 : }
208 :
209 : } // namespace webrtc
|
