Line data Source code
1 : /*
2 : * Copyright (c) 2012 The WebM 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 "vp8/encoder/denoising.h"
12 : #include "vp8/common/reconinter.h"
13 : #include "vpx/vpx_integer.h"
14 : #include "vpx_mem/vpx_mem.h"
15 : #include "vp8_rtcd.h"
16 :
17 : #include <emmintrin.h>
18 : #include "vpx_ports/emmintrin_compat.h"
19 :
20 : /* Compute the sum of all pixel differences of this MB. */
21 0 : static INLINE unsigned int abs_sum_diff_16x1(__m128i acc_diff) {
22 0 : const __m128i k_1 = _mm_set1_epi16(1);
23 0 : const __m128i acc_diff_lo =
24 0 : _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
25 0 : const __m128i acc_diff_hi =
26 0 : _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
27 0 : const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
28 0 : const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
29 0 : const __m128i hgfe_dcba =
30 0 : _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
31 0 : const __m128i hgfedcba =
32 0 : _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
33 0 : unsigned int sum_diff = abs(_mm_cvtsi128_si32(hgfedcba));
34 :
35 0 : return sum_diff;
36 : }
37 :
38 0 : int vp8_denoiser_filter_sse2(unsigned char *mc_running_avg_y,
39 : int mc_avg_y_stride, unsigned char *running_avg_y,
40 : int avg_y_stride, unsigned char *sig,
41 : int sig_stride, unsigned int motion_magnitude,
42 : int increase_denoising) {
43 0 : unsigned char *running_avg_y_start = running_avg_y;
44 0 : unsigned char *sig_start = sig;
45 : unsigned int sum_diff_thresh;
46 : int r;
47 0 : int shift_inc =
48 : (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
49 : ? 1
50 0 : : 0;
51 0 : __m128i acc_diff = _mm_setzero_si128();
52 0 : const __m128i k_0 = _mm_setzero_si128();
53 0 : const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
54 0 : const __m128i k_8 = _mm_set1_epi8(8);
55 0 : const __m128i k_16 = _mm_set1_epi8(16);
56 : /* Modify each level's adjustment according to motion_magnitude. */
57 0 : const __m128i l3 = _mm_set1_epi8(
58 0 : (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
59 : /* Difference between level 3 and level 2 is 2. */
60 0 : const __m128i l32 = _mm_set1_epi8(2);
61 : /* Difference between level 2 and level 1 is 1. */
62 0 : const __m128i l21 = _mm_set1_epi8(1);
63 :
64 0 : for (r = 0; r < 16; ++r) {
65 : /* Calculate differences */
66 0 : const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
67 0 : const __m128i v_mc_running_avg_y =
68 : _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
69 : __m128i v_running_avg_y;
70 0 : const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
71 0 : const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
72 : /* Obtain the sign. FF if diff is negative. */
73 0 : const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
74 : /* Clamp absolute difference to 16 to be used to get mask. Doing this
75 : * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
76 0 : const __m128i clamped_absdiff =
77 0 : _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
78 : /* Get masks for l2 l1 and l0 adjustments */
79 0 : const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
80 0 : const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
81 0 : const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
82 : /* Get adjustments for l2, l1, and l0 */
83 0 : __m128i adj2 = _mm_and_si128(mask2, l32);
84 0 : const __m128i adj1 = _mm_and_si128(mask1, l21);
85 0 : const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
86 : __m128i adj, padj, nadj;
87 :
88 : /* Combine the adjustments and get absolute adjustments. */
89 0 : adj2 = _mm_add_epi8(adj2, adj1);
90 0 : adj = _mm_sub_epi8(l3, adj2);
91 0 : adj = _mm_andnot_si128(mask0, adj);
92 0 : adj = _mm_or_si128(adj, adj0);
93 :
94 : /* Restore the sign and get positive and negative adjustments. */
95 0 : padj = _mm_andnot_si128(diff_sign, adj);
96 0 : nadj = _mm_and_si128(diff_sign, adj);
97 :
98 : /* Calculate filtered value. */
99 0 : v_running_avg_y = _mm_adds_epu8(v_sig, padj);
100 0 : v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
101 : _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
102 :
103 : /* Adjustments <=7, and each element in acc_diff can fit in signed
104 : * char.
105 : */
106 0 : acc_diff = _mm_adds_epi8(acc_diff, padj);
107 0 : acc_diff = _mm_subs_epi8(acc_diff, nadj);
108 :
109 : /* Update pointers for next iteration. */
110 0 : sig += sig_stride;
111 0 : mc_running_avg_y += mc_avg_y_stride;
112 0 : running_avg_y += avg_y_stride;
113 : }
114 :
115 : {
116 : /* Compute the sum of all pixel differences of this MB. */
117 0 : unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
118 0 : sum_diff_thresh = SUM_DIFF_THRESHOLD;
119 0 : if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
120 0 : if (abs_sum_diff > sum_diff_thresh) {
121 : // Before returning to copy the block (i.e., apply no denoising),
122 : // check if we can still apply some (weaker) temporal filtering to
123 : // this block, that would otherwise not be denoised at all. Simplest
124 : // is to apply an additional adjustment to running_avg_y to bring it
125 : // closer to sig. The adjustment is capped by a maximum delta, and
126 : // chosen such that in most cases the resulting sum_diff will be
127 : // within the acceptable range given by sum_diff_thresh.
128 :
129 : // The delta is set by the excess of absolute pixel diff over the
130 : // threshold.
131 0 : int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
132 : // Only apply the adjustment for max delta up to 3.
133 0 : if (delta < 4) {
134 0 : const __m128i k_delta = _mm_set1_epi8(delta);
135 0 : sig -= sig_stride * 16;
136 0 : mc_running_avg_y -= mc_avg_y_stride * 16;
137 0 : running_avg_y -= avg_y_stride * 16;
138 0 : for (r = 0; r < 16; ++r) {
139 0 : __m128i v_running_avg_y =
140 : _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
141 : // Calculate differences.
142 0 : const __m128i v_sig = _mm_loadu_si128((__m128i *)(&sig[0]));
143 0 : const __m128i v_mc_running_avg_y =
144 : _mm_loadu_si128((__m128i *)(&mc_running_avg_y[0]));
145 0 : const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
146 0 : const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
147 : // Obtain the sign. FF if diff is negative.
148 0 : const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
149 : // Clamp absolute difference to delta to get the adjustment.
150 0 : const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
151 : // Restore the sign and get positive and negative adjustments.
152 : __m128i padj, nadj;
153 0 : padj = _mm_andnot_si128(diff_sign, adj);
154 0 : nadj = _mm_and_si128(diff_sign, adj);
155 : // Calculate filtered value.
156 0 : v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
157 0 : v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
158 : _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
159 :
160 : // Accumulate the adjustments.
161 0 : acc_diff = _mm_subs_epi8(acc_diff, padj);
162 0 : acc_diff = _mm_adds_epi8(acc_diff, nadj);
163 :
164 : // Update pointers for next iteration.
165 0 : sig += sig_stride;
166 0 : mc_running_avg_y += mc_avg_y_stride;
167 0 : running_avg_y += avg_y_stride;
168 : }
169 0 : abs_sum_diff = abs_sum_diff_16x1(acc_diff);
170 0 : if (abs_sum_diff > sum_diff_thresh) {
171 0 : return COPY_BLOCK;
172 : }
173 : } else {
174 0 : return COPY_BLOCK;
175 : }
176 : }
177 : }
178 :
179 0 : vp8_copy_mem16x16(running_avg_y_start, avg_y_stride, sig_start, sig_stride);
180 0 : return FILTER_BLOCK;
181 : }
182 :
183 0 : int vp8_denoiser_filter_uv_sse2(unsigned char *mc_running_avg,
184 : int mc_avg_stride, unsigned char *running_avg,
185 : int avg_stride, unsigned char *sig,
186 : int sig_stride, unsigned int motion_magnitude,
187 : int increase_denoising) {
188 0 : unsigned char *running_avg_start = running_avg;
189 0 : unsigned char *sig_start = sig;
190 : unsigned int sum_diff_thresh;
191 : int r;
192 0 : int shift_inc =
193 : (increase_denoising && motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV)
194 : ? 1
195 0 : : 0;
196 0 : __m128i acc_diff = _mm_setzero_si128();
197 0 : const __m128i k_0 = _mm_setzero_si128();
198 0 : const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
199 0 : const __m128i k_8 = _mm_set1_epi8(8);
200 0 : const __m128i k_16 = _mm_set1_epi8(16);
201 : /* Modify each level's adjustment according to motion_magnitude. */
202 0 : const __m128i l3 = _mm_set1_epi8(
203 0 : (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD_UV) ? 7 + shift_inc : 6);
204 : /* Difference between level 3 and level 2 is 2. */
205 0 : const __m128i l32 = _mm_set1_epi8(2);
206 : /* Difference between level 2 and level 1 is 1. */
207 0 : const __m128i l21 = _mm_set1_epi8(1);
208 :
209 : {
210 0 : const __m128i k_1 = _mm_set1_epi16(1);
211 0 : __m128i vec_sum_block = _mm_setzero_si128();
212 :
213 : // Avoid denoising color signal if its close to average level.
214 0 : for (r = 0; r < 8; ++r) {
215 0 : const __m128i v_sig = _mm_loadl_epi64((__m128i *)(&sig[0]));
216 0 : const __m128i v_sig_unpack = _mm_unpacklo_epi8(v_sig, k_0);
217 0 : vec_sum_block = _mm_add_epi16(vec_sum_block, v_sig_unpack);
218 0 : sig += sig_stride;
219 : }
220 0 : sig -= sig_stride * 8;
221 : {
222 0 : const __m128i hg_fe_dc_ba = _mm_madd_epi16(vec_sum_block, k_1);
223 0 : const __m128i hgfe_dcba =
224 0 : _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
225 0 : const __m128i hgfedcba =
226 0 : _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
227 0 : const int sum_block = _mm_cvtsi128_si32(hgfedcba);
228 0 : if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV) {
229 0 : return COPY_BLOCK;
230 : }
231 : }
232 : }
233 :
234 0 : for (r = 0; r < 4; ++r) {
235 : /* Calculate differences */
236 0 : const __m128i v_sig_low =
237 0 : _mm_castpd_si128(_mm_load_sd((double *)(&sig[0])));
238 0 : const __m128i v_sig = _mm_castpd_si128(_mm_loadh_pd(
239 0 : _mm_castsi128_pd(v_sig_low), (double *)(&sig[sig_stride])));
240 0 : const __m128i v_mc_running_avg_low =
241 0 : _mm_castpd_si128(_mm_load_sd((double *)(&mc_running_avg[0])));
242 0 : const __m128i v_mc_running_avg = _mm_castpd_si128(
243 : _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
244 0 : (double *)(&mc_running_avg[mc_avg_stride])));
245 0 : const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
246 0 : const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
247 : /* Obtain the sign. FF if diff is negative. */
248 0 : const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
249 : /* Clamp absolute difference to 16 to be used to get mask. Doing this
250 : * allows us to use _mm_cmpgt_epi8, which operates on signed byte. */
251 0 : const __m128i clamped_absdiff =
252 0 : _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_16);
253 : /* Get masks for l2 l1 and l0 adjustments */
254 0 : const __m128i mask2 = _mm_cmpgt_epi8(k_16, clamped_absdiff);
255 0 : const __m128i mask1 = _mm_cmpgt_epi8(k_8, clamped_absdiff);
256 0 : const __m128i mask0 = _mm_cmpgt_epi8(k_4, clamped_absdiff);
257 : /* Get adjustments for l2, l1, and l0 */
258 0 : __m128i adj2 = _mm_and_si128(mask2, l32);
259 0 : const __m128i adj1 = _mm_and_si128(mask1, l21);
260 0 : const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
261 : __m128i adj, padj, nadj;
262 : __m128i v_running_avg;
263 :
264 : /* Combine the adjustments and get absolute adjustments. */
265 0 : adj2 = _mm_add_epi8(adj2, adj1);
266 0 : adj = _mm_sub_epi8(l3, adj2);
267 0 : adj = _mm_andnot_si128(mask0, adj);
268 0 : adj = _mm_or_si128(adj, adj0);
269 :
270 : /* Restore the sign and get positive and negative adjustments. */
271 0 : padj = _mm_andnot_si128(diff_sign, adj);
272 0 : nadj = _mm_and_si128(diff_sign, adj);
273 :
274 : /* Calculate filtered value. */
275 0 : v_running_avg = _mm_adds_epu8(v_sig, padj);
276 0 : v_running_avg = _mm_subs_epu8(v_running_avg, nadj);
277 :
278 0 : _mm_storel_pd((double *)&running_avg[0], _mm_castsi128_pd(v_running_avg));
279 0 : _mm_storeh_pd((double *)&running_avg[avg_stride],
280 : _mm_castsi128_pd(v_running_avg));
281 :
282 : /* Adjustments <=7, and each element in acc_diff can fit in signed
283 : * char.
284 : */
285 0 : acc_diff = _mm_adds_epi8(acc_diff, padj);
286 0 : acc_diff = _mm_subs_epi8(acc_diff, nadj);
287 :
288 : /* Update pointers for next iteration. */
289 0 : sig += sig_stride * 2;
290 0 : mc_running_avg += mc_avg_stride * 2;
291 0 : running_avg += avg_stride * 2;
292 : }
293 :
294 : {
295 0 : unsigned int abs_sum_diff = abs_sum_diff_16x1(acc_diff);
296 0 : sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
297 0 : if (increase_denoising) sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
298 0 : if (abs_sum_diff > sum_diff_thresh) {
299 : // Before returning to copy the block (i.e., apply no denoising),
300 : // check if we can still apply some (weaker) temporal filtering to
301 : // this block, that would otherwise not be denoised at all. Simplest
302 : // is to apply an additional adjustment to running_avg_y to bring it
303 : // closer to sig. The adjustment is capped by a maximum delta, and
304 : // chosen such that in most cases the resulting sum_diff will be
305 : // within the acceptable range given by sum_diff_thresh.
306 :
307 : // The delta is set by the excess of absolute pixel diff over the
308 : // threshold.
309 0 : int delta = ((abs_sum_diff - sum_diff_thresh) >> 8) + 1;
310 : // Only apply the adjustment for max delta up to 3.
311 0 : if (delta < 4) {
312 0 : const __m128i k_delta = _mm_set1_epi8(delta);
313 0 : sig -= sig_stride * 8;
314 0 : mc_running_avg -= mc_avg_stride * 8;
315 0 : running_avg -= avg_stride * 8;
316 0 : for (r = 0; r < 4; ++r) {
317 : // Calculate differences.
318 0 : const __m128i v_sig_low =
319 0 : _mm_castpd_si128(_mm_load_sd((double *)(&sig[0])));
320 0 : const __m128i v_sig = _mm_castpd_si128(_mm_loadh_pd(
321 0 : _mm_castsi128_pd(v_sig_low), (double *)(&sig[sig_stride])));
322 0 : const __m128i v_mc_running_avg_low =
323 0 : _mm_castpd_si128(_mm_load_sd((double *)(&mc_running_avg[0])));
324 0 : const __m128i v_mc_running_avg = _mm_castpd_si128(
325 : _mm_loadh_pd(_mm_castsi128_pd(v_mc_running_avg_low),
326 0 : (double *)(&mc_running_avg[mc_avg_stride])));
327 0 : const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg, v_sig);
328 0 : const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg);
329 : // Obtain the sign. FF if diff is negative.
330 0 : const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
331 : // Clamp absolute difference to delta to get the adjustment.
332 0 : const __m128i adj = _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
333 : // Restore the sign and get positive and negative adjustments.
334 : __m128i padj, nadj;
335 0 : const __m128i v_running_avg_low =
336 0 : _mm_castpd_si128(_mm_load_sd((double *)(&running_avg[0])));
337 0 : __m128i v_running_avg = _mm_castpd_si128(
338 : _mm_loadh_pd(_mm_castsi128_pd(v_running_avg_low),
339 0 : (double *)(&running_avg[avg_stride])));
340 0 : padj = _mm_andnot_si128(diff_sign, adj);
341 0 : nadj = _mm_and_si128(diff_sign, adj);
342 : // Calculate filtered value.
343 0 : v_running_avg = _mm_subs_epu8(v_running_avg, padj);
344 0 : v_running_avg = _mm_adds_epu8(v_running_avg, nadj);
345 :
346 0 : _mm_storel_pd((double *)&running_avg[0],
347 : _mm_castsi128_pd(v_running_avg));
348 0 : _mm_storeh_pd((double *)&running_avg[avg_stride],
349 : _mm_castsi128_pd(v_running_avg));
350 :
351 : // Accumulate the adjustments.
352 0 : acc_diff = _mm_subs_epi8(acc_diff, padj);
353 0 : acc_diff = _mm_adds_epi8(acc_diff, nadj);
354 :
355 : // Update pointers for next iteration.
356 0 : sig += sig_stride * 2;
357 0 : mc_running_avg += mc_avg_stride * 2;
358 0 : running_avg += avg_stride * 2;
359 : }
360 0 : abs_sum_diff = abs_sum_diff_16x1(acc_diff);
361 0 : if (abs_sum_diff > sum_diff_thresh) {
362 0 : return COPY_BLOCK;
363 : }
364 : } else {
365 0 : return COPY_BLOCK;
366 : }
367 : }
368 : }
369 :
370 0 : vp8_copy_mem8x8(running_avg_start, avg_stride, sig_start, sig_stride);
371 0 : return FILTER_BLOCK;
372 : }
|