1 /*
2 * Copyright (c) 2015 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 <stdlib.h>
12 #include "./vp8_rtcd.h"
13 #include "vp8/common/mips/msa/vp8_macros_msa.h"
14 #include "vp8/encoder/denoising.h"
15
vp8_denoiser_filter_msa(uint8_t * mc_running_avg_y_ptr,int32_t mc_avg_y_stride,uint8_t * running_avg_y_ptr,int32_t avg_y_stride,uint8_t * sig_ptr,int32_t sig_stride,uint32_t motion_magnitude,int32_t increase_denoising)16 int32_t vp8_denoiser_filter_msa(uint8_t *mc_running_avg_y_ptr,
17 int32_t mc_avg_y_stride,
18 uint8_t *running_avg_y_ptr,
19 int32_t avg_y_stride,
20 uint8_t *sig_ptr, int32_t sig_stride,
21 uint32_t motion_magnitude,
22 int32_t increase_denoising)
23 {
24 uint8_t *running_avg_y_start = running_avg_y_ptr;
25 uint8_t *sig_start = sig_ptr;
26 int32_t cnt = 0;
27 int32_t sum_diff = 0;
28 int32_t shift_inc1 = 3;
29 int32_t delta = 0;
30 int32_t sum_diff_thresh;
31 v16u8 src0, src1, src2, src3, src4, src5, src6, src7;
32 v16u8 src8, src9, src10, src11, src12, src13, src14, src15;
33 v16u8 mc_running_avg_y0, running_avg_y, sig0;
34 v16u8 mc_running_avg_y1, running_avg_y1, sig1;
35 v16u8 coeff0, coeff1;
36 v8i16 diff0, diff1, abs_diff0, abs_diff1, abs_diff_neg0, abs_diff_neg1;
37 v8i16 adjust0, adjust1, adjust2, adjust3;
38 v8i16 shift_inc1_vec = { 0 };
39 v8i16 col_sum0 = { 0 };
40 v8i16 col_sum1 = { 0 };
41 v8i16 col_sum2 = { 0 };
42 v8i16 col_sum3 = { 0 };
43 v8i16 temp0_h, temp1_h, temp2_h, temp3_h, cmp, delta_vec;
44 v4i32 temp0_w;
45 v2i64 temp0_d, temp1_d;
46 v8i16 zero = { 0 };
47 v8i16 one = __msa_ldi_h(1);
48 v8i16 four = __msa_ldi_h(4);
49 v8i16 val_127 = __msa_ldi_h(127);
50 v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 };
51
52 if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
53 {
54 adj_val = __msa_add_a_h(adj_val, one);
55 if (increase_denoising)
56 {
57 adj_val = __msa_add_a_h(adj_val, one);
58 shift_inc1 = 4;
59 }
60
61 temp0_h = zero - adj_val;
62 adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val);
63 }
64
65 adj_val = __msa_insert_h(adj_val, 3, cnt);
66 adj_val = __msa_insert_h(adj_val, 7, cnt);
67 shift_inc1_vec = __msa_fill_h(shift_inc1);
68
69 for (cnt = 8; cnt--;)
70 {
71 v8i16 mask0 = { 0 };
72 v8i16 mask1 = { 0 };
73
74 mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
75 sig0 = LD_UB(sig_ptr);
76 sig_ptr += sig_stride;
77 mc_running_avg_y_ptr += mc_avg_y_stride;
78
79 mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
80 sig1 = LD_UB(sig_ptr);
81
82 ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1);
83 HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
84 abs_diff0 = __msa_add_a_h(diff0, zero);
85 abs_diff1 = __msa_add_a_h(diff1, zero);
86 cmp = __msa_clei_s_h(abs_diff0, 15);
87 cmp = cmp & one;
88 mask0 += cmp;
89 cmp = __msa_clei_s_h(abs_diff0, 7);
90 cmp = cmp & one;
91 mask0 += cmp;
92 cmp = abs_diff0 < shift_inc1_vec;
93 cmp = cmp & one;
94 mask0 += cmp;
95 cmp = __msa_clei_s_h(abs_diff1, 15);
96 cmp = cmp & one;
97 mask1 += cmp;
98 cmp = __msa_clei_s_h(abs_diff1, 7);
99 cmp = cmp & one;
100 mask1 += cmp;
101 cmp = abs_diff1 < shift_inc1_vec;
102 cmp = cmp & one;
103 mask1 += cmp;
104 temp0_h = __msa_clei_s_h(diff0, 0);
105 temp0_h = temp0_h & four;
106 mask0 += temp0_h;
107 temp1_h = __msa_clei_s_h(diff1, 0);
108 temp1_h = temp1_h & four;
109 mask1 += temp1_h;
110 VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0,
111 adjust1);
112 temp2_h = __msa_ceqi_h(adjust0, 0);
113 temp3_h = __msa_ceqi_h(adjust1, 0);
114 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
115 (v16u8)temp2_h);
116 adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1,
117 (v16u8)temp3_h);
118 ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1);
119 UNPCK_UB_SH(sig0, temp0_h, temp1_h);
120 ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h);
121 MAXI_SH2_SH(temp0_h, temp1_h, 0);
122 SAT_UH2_SH(temp0_h, temp1_h, 7);
123 temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h);
124 running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h);
125 running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0,
126 (v16u8)temp2_h);
127 ST_UB(running_avg_y, running_avg_y_ptr);
128 running_avg_y_ptr += avg_y_stride;
129
130 mask0 = zero;
131 mask1 = zero;
132 ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1);
133 HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
134 abs_diff0 = __msa_add_a_h(diff0, zero);
135 abs_diff1 = __msa_add_a_h(diff1, zero);
136 cmp = __msa_clei_s_h(abs_diff0, 15);
137 cmp = cmp & one;
138 mask0 += cmp;
139 cmp = __msa_clei_s_h(abs_diff0, 7);
140 cmp = cmp & one;
141 mask0 += cmp;
142 cmp = abs_diff0 < shift_inc1_vec;
143 cmp = cmp & one;
144 mask0 += cmp;
145 cmp = __msa_clei_s_h(abs_diff1, 15);
146 cmp = cmp & one;
147 mask1 += cmp;
148 cmp = __msa_clei_s_h(abs_diff1, 7);
149 cmp = cmp & one;
150 mask1 += cmp;
151 cmp = abs_diff1 < shift_inc1_vec;
152 cmp = cmp & one;
153 mask1 += cmp;
154 temp0_h = __msa_clei_s_h(diff0, 0);
155 temp0_h = temp0_h & four;
156 mask0 += temp0_h;
157 temp1_h = __msa_clei_s_h(diff1, 0);
158 temp1_h = temp1_h & four;
159 mask1 += temp1_h;
160 VSHF_H2_SH(adj_val, adj_val, adj_val, adj_val, mask0, mask1, adjust0,
161 adjust1);
162 temp2_h = __msa_ceqi_h(adjust0, 0);
163 temp3_h = __msa_ceqi_h(adjust1, 0);
164 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
165 (v16u8)temp2_h);
166 adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)diff1,
167 (v16u8)temp3_h);
168 ADD2(col_sum0, adjust0, col_sum1, adjust1, col_sum0, col_sum1);
169 UNPCK_UB_SH(sig1, temp0_h, temp1_h);
170 ADD2(temp0_h, adjust0, temp1_h, adjust1, temp0_h, temp1_h);
171 MAXI_SH2_SH(temp0_h, temp1_h, 0);
172 SAT_UH2_SH(temp0_h, temp1_h, 7);
173 temp2_h = (v8i16)__msa_pckev_b((v16i8)temp3_h, (v16i8)temp2_h);
174 running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp1_h, (v16i8)temp0_h);
175 running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1,
176 (v16u8)temp2_h);
177 ST_UB(running_avg_y, running_avg_y_ptr);
178 sig_ptr += sig_stride;
179 mc_running_avg_y_ptr += mc_avg_y_stride;
180 running_avg_y_ptr += avg_y_stride;
181 }
182
183 col_sum0 = __msa_min_s_h(col_sum0, val_127);
184 col_sum1 = __msa_min_s_h(col_sum1, val_127);
185 temp0_h = col_sum0 + col_sum1;
186 temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
187 temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
188 temp1_d = __msa_splati_d(temp0_d, 1);
189 temp0_d += temp1_d;
190 sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
191 sig_ptr -= sig_stride * 16;
192 mc_running_avg_y_ptr -= mc_avg_y_stride * 16;
193 running_avg_y_ptr -= avg_y_stride * 16;
194
195 if (increase_denoising)
196 {
197 sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH;
198 }
199
200 if (abs(sum_diff) > sum_diff_thresh)
201 {
202 delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
203 delta_vec = __msa_fill_h(delta);
204 if (delta < 4)
205 {
206 for (cnt = 8; cnt--;)
207 {
208 running_avg_y = LD_UB(running_avg_y_ptr);
209 mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
210 sig0 = LD_UB(sig_ptr);
211 sig_ptr += sig_stride;
212 mc_running_avg_y_ptr += mc_avg_y_stride;
213 running_avg_y_ptr += avg_y_stride;
214 mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
215 sig1 = LD_UB(sig_ptr);
216 running_avg_y1 = LD_UB(running_avg_y_ptr);
217 ILVRL_B2_UB(mc_running_avg_y0, sig0, coeff0, coeff1);
218 HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
219 abs_diff0 = __msa_add_a_h(diff0, zero);
220 abs_diff1 = __msa_add_a_h(diff1, zero);
221 temp0_h = abs_diff0 < delta_vec;
222 temp1_h = abs_diff1 < delta_vec;
223 abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
224 (v16u8)delta_vec,
225 (v16u8)temp0_h);
226 abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1,
227 (v16u8)delta_vec,
228 (v16u8)temp1_h);
229 SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0,
230 abs_diff_neg1);
231 abs_diff_neg0 = zero - abs_diff0;
232 abs_diff_neg1 = zero - abs_diff1;
233 temp0_h = __msa_clei_s_h(diff0, 0);
234 temp1_h = __msa_clei_s_h(diff1, 0);
235 adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
236 (v16u8)abs_diff_neg0,
237 (v16u8)temp0_h);
238 adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1,
239 (v16u8)abs_diff_neg1,
240 (v16u8)temp1_h);
241 ILVRL_B2_SH(zero, running_avg_y, temp2_h, temp3_h);
242 ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3);
243 MAXI_SH2_SH(adjust2, adjust3, 0);
244 SAT_UH2_SH(adjust2, adjust3, 7);
245 temp0_h = __msa_ceqi_h(diff0, 0);
246 temp1_h = __msa_ceqi_h(diff1, 0);
247 adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h,
248 (v16u8)temp0_h);
249 adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h,
250 (v16u8)temp1_h);
251 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
252 (v16u8)temp0_h);
253 adjust1 = (v8i16)__msa_bmnz_v((v16u8)adjust1, (v16u8)zero,
254 (v16u8)temp1_h);
255 ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3);
256 running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3,
257 (v16i8)adjust2);
258 ST_UB(running_avg_y, running_avg_y_ptr - avg_y_stride);
259 ILVRL_B2_UB(mc_running_avg_y1, sig1, coeff0, coeff1);
260 HSUB_UB2_SH(coeff0, coeff1, diff0, diff1);
261 abs_diff0 = __msa_add_a_h(diff0, zero);
262 abs_diff1 = __msa_add_a_h(diff1, zero);
263 temp0_h = abs_diff0 < delta_vec;
264 temp1_h = abs_diff1 < delta_vec;
265 abs_diff0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
266 (v16u8)delta_vec,
267 (v16u8)temp0_h);
268 abs_diff1 = (v8i16)__msa_bmz_v((v16u8)abs_diff1,
269 (v16u8)delta_vec,
270 (v16u8)temp1_h);
271 SUB2(zero, abs_diff0, zero, abs_diff1, abs_diff_neg0,
272 abs_diff_neg1);
273 temp0_h = __msa_clei_s_h(diff0, 0);
274 temp1_h = __msa_clei_s_h(diff1, 0);
275 adjust0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
276 (v16u8)abs_diff_neg0,
277 (v16u8)temp0_h);
278 adjust1 = (v8i16)__msa_bmnz_v((v16u8)abs_diff1,
279 (v16u8)abs_diff_neg1,
280 (v16u8)temp1_h);
281 ILVRL_H2_SH(zero, running_avg_y1, temp2_h, temp3_h);
282 ADD2(temp2_h, adjust0, temp3_h, adjust1, adjust2, adjust3);
283 MAXI_SH2_SH(adjust2, adjust3, 0);
284 SAT_UH2_SH(adjust2, adjust3, 7);
285 temp0_h = __msa_ceqi_h(diff0, 0);
286 temp1_h = __msa_ceqi_h(diff1, 0);
287 adjust2 = (v8i16)__msa_bmz_v((v16u8)adjust2, (v16u8)temp2_h,
288 (v16u8)temp0_h);
289 adjust3 = (v8i16)__msa_bmz_v((v16u8)adjust3, (v16u8)temp3_h,
290 (v16u8)temp1_h);
291 adjust0 = (v8i16)__msa_bmz_v((v16u8)adjust0, (v16u8)zero,
292 (v16u8)temp0_h);
293 adjust1 = (v8i16)__msa_bmz_v((v16u8)adjust1, (v16u8)zero,
294 (v16u8)temp1_h);
295 ADD2(col_sum2, adjust0, col_sum3, adjust1, col_sum2, col_sum3);
296 running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust3,
297 (v16i8)adjust2);
298 ST_UB(running_avg_y, running_avg_y_ptr);
299 running_avg_y_ptr += avg_y_stride;
300 }
301
302 col_sum2 = __msa_min_s_h(col_sum2, val_127);
303 col_sum3 = __msa_min_s_h(col_sum3, val_127);
304 temp0_h = col_sum2 + col_sum3;
305 temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
306 temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
307 temp1_d = __msa_splati_d(temp0_d, 1);
308 temp0_d += (v2i64)temp1_d;
309 sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
310 if (abs(sum_diff) > SUM_DIFF_THRESHOLD)
311 {
312 return COPY_BLOCK;
313 }
314 }
315 else
316 {
317 return COPY_BLOCK;
318 }
319 }
320
321 LD_UB8(sig_start, sig_stride, src0, src1, src2, src3, src4, src5, src6,
322 src7);
323 sig_start += (8 * sig_stride);
324 LD_UB8(sig_start, sig_stride, src8, src9, src10, src11, src12, src13,
325 src14, src15);
326
327 ST_UB8(src0, src1, src2, src3, src4, src5, src6, src7, running_avg_y_start,
328 avg_y_stride);
329 running_avg_y_start += (8 * avg_y_stride);
330 ST_UB8(src8, src9, src10, src11, src12, src13, src14, src15,
331 running_avg_y_start, avg_y_stride);
332
333 return FILTER_BLOCK;
334 }
335
vp8_denoiser_filter_uv_msa(uint8_t * mc_running_avg_y_ptr,int32_t mc_avg_y_stride,uint8_t * running_avg_y_ptr,int32_t avg_y_stride,uint8_t * sig_ptr,int32_t sig_stride,uint32_t motion_magnitude,int32_t increase_denoising)336 int32_t vp8_denoiser_filter_uv_msa(uint8_t *mc_running_avg_y_ptr,
337 int32_t mc_avg_y_stride,
338 uint8_t *running_avg_y_ptr,
339 int32_t avg_y_stride,
340 uint8_t *sig_ptr,
341 int32_t sig_stride,
342 uint32_t motion_magnitude,
343 int32_t increase_denoising)
344 {
345 uint8_t *running_avg_y_start = running_avg_y_ptr;
346 uint8_t *sig_start = sig_ptr;
347 int32_t cnt = 0;
348 int32_t sum_diff = 0;
349 int32_t shift_inc1 = 3;
350 int32_t delta = 0;
351 int32_t sum_block = 0;
352 int32_t sum_diff_thresh;
353 int64_t dst0, dst1, src0, src1, src2, src3;
354 v16u8 mc_running_avg_y0, running_avg_y, sig0;
355 v16u8 mc_running_avg_y1, running_avg_y1, sig1;
356 v16u8 sig2, sig3, sig4, sig5, sig6, sig7;
357 v16u8 coeff0;
358 v8i16 diff0, abs_diff0, abs_diff_neg0;
359 v8i16 adjust0, adjust2;
360 v8i16 shift_inc1_vec = { 0 };
361 v8i16 col_sum0 = { 0 };
362 v8i16 temp0_h, temp2_h, cmp, delta_vec;
363 v4i32 temp0_w;
364 v2i64 temp0_d, temp1_d;
365 v16i8 zero = { 0 };
366 v8i16 one = __msa_ldi_h(1);
367 v8i16 four = __msa_ldi_h(4);
368 v8i16 adj_val = { 6, 4, 3, 0, -6, -4, -3, 0 };
369
370
371 sig0 = LD_UB(sig_ptr);
372 sig_ptr += sig_stride;
373 temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0);
374 sig1 = LD_UB(sig_ptr);
375 sig_ptr += sig_stride;
376 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig1);
377 sig2 = LD_UB(sig_ptr);
378 sig_ptr += sig_stride;
379 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig2);
380 sig3 = LD_UB(sig_ptr);
381 sig_ptr += sig_stride;
382 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig3);
383 sig4 = LD_UB(sig_ptr);
384 sig_ptr += sig_stride;
385 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig4);
386 sig5 = LD_UB(sig_ptr);
387 sig_ptr += sig_stride;
388 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig5);
389 sig6 = LD_UB(sig_ptr);
390 sig_ptr += sig_stride;
391 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig6);
392 sig7 = LD_UB(sig_ptr);
393 sig_ptr += sig_stride;
394 temp0_h += (v8i16)__msa_ilvr_b(zero, (v16i8)sig7);
395 temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
396 temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
397 temp1_d = __msa_splati_d(temp0_d, 1);
398 temp0_d += temp1_d;
399 sum_block = __msa_copy_s_w((v4i32)temp0_d, 0);
400 sig_ptr -= sig_stride * 8;
401
402 if (abs(sum_block - (128 * 8 * 8)) < SUM_DIFF_FROM_AVG_THRESH_UV)
403 {
404 return COPY_BLOCK;
405 }
406
407 if (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD)
408 {
409 adj_val = __msa_add_a_h(adj_val, one);
410
411 if (increase_denoising)
412 {
413 adj_val = __msa_add_a_h(adj_val, one);
414 shift_inc1 = 4;
415 }
416
417 temp0_h = (v8i16)zero - adj_val;
418 adj_val = (v8i16)__msa_ilvev_d((v2i64)temp0_h, (v2i64)adj_val);
419 }
420
421 adj_val = __msa_insert_h(adj_val, 3, cnt);
422 adj_val = __msa_insert_h(adj_val, 7, cnt);
423 shift_inc1_vec = __msa_fill_h(shift_inc1);
424 for (cnt = 4; cnt--;)
425 {
426 v8i16 mask0 = { 0 };
427 mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
428 sig0 = LD_UB(sig_ptr);
429 sig_ptr += sig_stride;
430 mc_running_avg_y_ptr += mc_avg_y_stride;
431 mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
432 sig1 = LD_UB(sig_ptr);
433 coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0, (v16i8)sig0);
434 diff0 = __msa_hsub_u_h(coeff0, coeff0);
435 abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
436 cmp = __msa_clei_s_h(abs_diff0, 15);
437 cmp = cmp & one;
438 mask0 += cmp;
439 cmp = __msa_clei_s_h(abs_diff0, 7);
440 cmp = cmp & one;
441 mask0 += cmp;
442 cmp = abs_diff0 < shift_inc1_vec;
443 cmp = cmp & one;
444 mask0 += cmp;
445 temp0_h = __msa_clei_s_h(diff0, 0);
446 temp0_h = temp0_h & four;
447 mask0 += temp0_h;
448 adjust0 = __msa_vshf_h(mask0, adj_val, adj_val);
449 temp2_h = __msa_ceqi_h(adjust0, 0);
450 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
451 (v16u8)temp2_h);
452 col_sum0 += adjust0;
453 temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig0);
454 temp0_h += adjust0;
455 temp0_h = __msa_maxi_s_h(temp0_h, 0);
456 temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7);
457 temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h);
458 running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h);
459 running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y0,
460 (v16u8)temp2_h);
461 dst0 = __msa_copy_s_d((v2i64)running_avg_y, 0);
462 SD(dst0, running_avg_y_ptr);
463 running_avg_y_ptr += avg_y_stride;
464
465 mask0 = __msa_ldi_h(0);
466 coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1, (v16i8)sig1);
467 diff0 = __msa_hsub_u_h(coeff0, coeff0);
468 abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
469 cmp = __msa_clei_s_h(abs_diff0, 15);
470 cmp = cmp & one;
471 mask0 += cmp;
472 cmp = __msa_clei_s_h(abs_diff0, 7);
473 cmp = cmp & one;
474 mask0 += cmp;
475 cmp = abs_diff0 < shift_inc1_vec;
476 cmp = cmp & one;
477 mask0 += cmp;
478 temp0_h = __msa_clei_s_h(diff0, 0);
479 temp0_h = temp0_h & four;
480 mask0 += temp0_h;
481 adjust0 = __msa_vshf_h(mask0, adj_val, adj_val);
482 temp2_h = __msa_ceqi_h(adjust0, 0);
483 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)diff0,
484 (v16u8)temp2_h);
485 col_sum0 += adjust0;
486 temp0_h = (v8i16)__msa_ilvr_b(zero, (v16i8)sig1);
487 temp0_h += adjust0;
488 temp0_h = __msa_maxi_s_h(temp0_h, 0);
489 temp0_h = (v8i16)__msa_sat_u_h((v8u16)temp0_h, 7);
490
491 temp2_h = (v8i16)__msa_pckev_b((v16i8)temp2_h, (v16i8)temp2_h);
492 running_avg_y = (v16u8)__msa_pckev_b((v16i8)temp0_h, (v16i8)temp0_h);
493 running_avg_y = __msa_bmnz_v(running_avg_y, mc_running_avg_y1,
494 (v16u8)temp2_h);
495 dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0);
496 SD(dst1, running_avg_y_ptr);
497
498 sig_ptr += sig_stride;
499 mc_running_avg_y_ptr += mc_avg_y_stride;
500 running_avg_y_ptr += avg_y_stride;
501 }
502
503 temp0_h = col_sum0;
504 temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
505 temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
506 temp1_d = __msa_splati_d(temp0_d, 1);
507 temp0_d += temp1_d;
508 sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
509 sig_ptr -= sig_stride * 8;
510 mc_running_avg_y_ptr -= mc_avg_y_stride * 8;
511 running_avg_y_ptr -= avg_y_stride * 8;
512 sum_diff_thresh = SUM_DIFF_THRESHOLD_UV;
513
514 if (increase_denoising)
515 {
516 sum_diff_thresh = SUM_DIFF_THRESHOLD_HIGH_UV;
517 }
518
519 if (abs(sum_diff) > sum_diff_thresh)
520 {
521 delta = ((abs(sum_diff) - sum_diff_thresh) >> 8) + 1;
522 delta_vec = __msa_fill_h(delta);
523 if (delta < 4)
524 {
525 for (cnt = 4; cnt--;)
526 {
527 running_avg_y = LD_UB(running_avg_y_ptr);
528 mc_running_avg_y0 = LD_UB(mc_running_avg_y_ptr);
529 sig0 = LD_UB(sig_ptr);
530 /* Update pointers for next iteration. */
531 sig_ptr += sig_stride;
532 mc_running_avg_y_ptr += mc_avg_y_stride;
533 running_avg_y_ptr += avg_y_stride;
534
535 mc_running_avg_y1 = LD_UB(mc_running_avg_y_ptr);
536 sig1 = LD_UB(sig_ptr);
537 running_avg_y1 = LD_UB(running_avg_y_ptr);
538
539 coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y0,
540 (v16i8)sig0);
541 diff0 = __msa_hsub_u_h(coeff0, coeff0);
542 abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
543 temp0_h = delta_vec < abs_diff0;
544 abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
545 (v16u8)delta_vec,
546 (v16u8)temp0_h);
547 abs_diff_neg0 = (v8i16)zero - abs_diff0;
548 temp0_h = __msa_clei_s_h(diff0, 0);
549 adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
550 (v16u8)abs_diff_neg0,
551 (v16u8)temp0_h);
552 temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y);
553 adjust2 = temp2_h + adjust0;
554 adjust2 = __msa_maxi_s_h(adjust2, 0);
555 adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7);
556 temp0_h = __msa_ceqi_h(diff0, 0);
557 adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h,
558 (v16u8)temp0_h);
559 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
560 (v16u8)temp0_h);
561 col_sum0 += adjust0;
562 running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2,
563 (v16i8)adjust2);
564 dst0 = __msa_copy_s_d((v2i64)running_avg_y, 0);
565 SD(dst0, running_avg_y_ptr - avg_y_stride);
566
567 coeff0 = (v16u8)__msa_ilvr_b((v16i8)mc_running_avg_y1,
568 (v16i8)sig1);
569 diff0 = __msa_hsub_u_h(coeff0, coeff0);
570 abs_diff0 = __msa_add_a_h(diff0, (v8i16)zero);
571 temp0_h = delta_vec < abs_diff0;
572 abs_diff0 = (v8i16)__msa_bmnz_v((v16u8)abs_diff0,
573 (v16u8)delta_vec,
574 (v16u8)temp0_h);
575 abs_diff_neg0 = (v8i16)zero - abs_diff0;
576 temp0_h = __msa_clei_s_h(diff0, 0);
577 adjust0 = (v8i16)__msa_bmz_v((v16u8)abs_diff0,
578 (v16u8)abs_diff_neg0,
579 (v16u8)temp0_h);
580 temp2_h = (v8i16)__msa_ilvr_b(zero, (v16i8)running_avg_y1);
581 adjust2 = temp2_h + adjust0;
582 adjust2 = __msa_maxi_s_h(adjust2, 0);
583 adjust2 = (v8i16)__msa_sat_u_h((v8u16)adjust2, 7);
584 temp0_h = __msa_ceqi_h(diff0, 0);
585 adjust2 = (v8i16)__msa_bmnz_v((v16u8)adjust2, (v16u8)temp2_h,
586 (v16u8)temp0_h);
587 adjust0 = (v8i16)__msa_bmnz_v((v16u8)adjust0, (v16u8)zero,
588 (v16u8)temp0_h);
589 col_sum0 += adjust0;
590 running_avg_y = (v16u8)__msa_pckev_b((v16i8)adjust2,
591 (v16i8)adjust2);
592 dst1 = __msa_copy_s_d((v2i64)running_avg_y, 0);
593 SD(dst1, running_avg_y_ptr);
594 running_avg_y_ptr += avg_y_stride;
595 }
596
597 temp0_h = col_sum0;
598 temp0_w = __msa_hadd_s_w(temp0_h, temp0_h);
599 temp0_d = __msa_hadd_s_d(temp0_w, temp0_w);
600 temp1_d = __msa_splati_d(temp0_d, 1);
601 temp0_d += temp1_d;
602 sum_diff = __msa_copy_s_w((v4i32)temp0_d, 0);
603
604 if (abs(sum_diff) > sum_diff_thresh)
605 {
606 return COPY_BLOCK;
607 }
608 }
609 else
610 {
611 return COPY_BLOCK;
612 }
613 }
614
615 LD4(sig_start, sig_stride, src0, src1, src2, src3);
616 sig_start += (4 * sig_stride);
617 SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride);
618 running_avg_y_start += (4 * avg_y_stride);
619
620 LD4(sig_start, sig_stride, src0, src1, src2, src3);
621 SD4(src0, src1, src2, src3, running_avg_y_start, avg_y_stride);
622
623 return FILTER_BLOCK;
624 }
625