1 /*
2 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <emmintrin.h>
13
14 #include "config/av1_rtcd.h"
15
16 #include "aom_dsp/aom_dsp_common.h"
17 #include "aom_dsp/aom_filter.h"
18 #include "aom_dsp/x86/convolve_sse2.h"
19 #include "aom_dsp/x86/convolve_common_intrin.h"
20 #include "av1/common/convolve.h"
21
av1_convolve_2d_sr_12tap_sse2(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_qn,const int subpel_y_qn,ConvolveParams * conv_params)22 void av1_convolve_2d_sr_12tap_sse2(const uint8_t *src, int src_stride,
23 uint8_t *dst, int dst_stride, int w, int h,
24 const InterpFilterParams *filter_params_x,
25 const InterpFilterParams *filter_params_y,
26 const int subpel_x_qn, const int subpel_y_qn,
27 ConvolveParams *conv_params) {
28 const int bd = 8;
29
30 DECLARE_ALIGNED(16, int16_t,
31 im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
32 int im_h = h + filter_params_y->taps - 1;
33 int im_stride = w;
34 int i, j;
35 const int fo_vert = filter_params_y->taps / 2 - 1;
36 const int fo_horiz = filter_params_x->taps / 2 - 1;
37 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
38
39 const __m128i zero = _mm_setzero_si128();
40 const int bits =
41 FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
42 const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
43
44 assert(conv_params->round_0 > 0);
45 __m128i coeffs[6];
46
47 /* Horizontal filter */
48 {
49 prepare_coeffs_12tap(filter_params_x, subpel_x_qn, coeffs);
50
51 const __m128i round_const = _mm_set1_epi32(
52 (1 << (bd + FILTER_BITS - 1)) + ((1 << conv_params->round_0) >> 1));
53 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
54
55 for (i = 0; i < im_h; ++i) {
56 for (j = 0; j < w; j += 8) {
57 const __m128i data =
58 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
59 const __m128i data_2 =
60 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 4)]);
61
62 // Filter even-index pixels
63 const __m128i src_0 = _mm_unpacklo_epi8(data, zero);
64 const __m128i res_0 = _mm_madd_epi16(src_0, coeffs[0]);
65 const __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero);
66 const __m128i res_2 = _mm_madd_epi16(src_2, coeffs[1]);
67 const __m128i src_4 = _mm_unpacklo_epi8(data_2, zero);
68 const __m128i res_4 = _mm_madd_epi16(src_4, coeffs[2]);
69 const __m128i src_6 =
70 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 2), zero);
71 const __m128i res_6 = _mm_madd_epi16(src_6, coeffs[3]);
72 const __m128i src_8 =
73 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 4), zero);
74 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]);
75 const __m128i src_10 =
76 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 6), zero);
77 const __m128i res_10 = _mm_madd_epi16(src_10, coeffs[5]);
78
79 const __m128i res_0246 = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
80 _mm_add_epi32(res_2, res_6));
81 __m128i res_even =
82 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246);
83 res_even =
84 _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
85
86 // Filter odd-index pixels
87 const __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero);
88 const __m128i res_1 = _mm_madd_epi16(src_1, coeffs[0]);
89 const __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero);
90 const __m128i res_3 = _mm_madd_epi16(src_3, coeffs[1]);
91 const __m128i src_5 =
92 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 1), zero);
93 const __m128i res_5 = _mm_madd_epi16(src_5, coeffs[2]);
94 const __m128i src_7 =
95 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 3), zero);
96 const __m128i res_7 = _mm_madd_epi16(src_7, coeffs[3]);
97 const __m128i src_9 =
98 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 5), zero);
99 const __m128i res_9 = _mm_madd_epi16(src_9, coeffs[4]);
100 const __m128i src_11 =
101 _mm_unpacklo_epi8(_mm_srli_si128(data_2, 7), zero);
102 const __m128i res_11 = _mm_madd_epi16(src_11, coeffs[5]);
103
104 const __m128i res_1357 = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
105 _mm_add_epi32(res_3, res_7));
106 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_9, res_11), res_1357);
107 res_odd =
108 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
109
110 // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
111 __m128i res = _mm_packs_epi32(res_even, res_odd);
112 _mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res);
113 }
114 }
115 }
116
117 /* Vertical filter */
118 {
119 prepare_coeffs_12tap(filter_params_y, subpel_y_qn, coeffs);
120
121 const __m128i sum_round =
122 _mm_set1_epi32((1 << offset_bits) + ((1 << conv_params->round_1) >> 1));
123 const __m128i sum_shift = _mm_cvtsi32_si128(conv_params->round_1);
124
125 const __m128i round_const = _mm_set1_epi32(
126 ((1 << bits) >> 1) - (1 << (offset_bits - conv_params->round_1)) -
127 ((1 << (offset_bits - conv_params->round_1)) >> 1));
128 const __m128i round_shift = _mm_cvtsi32_si128(bits);
129
130 for (i = 0; i < h; ++i) {
131 for (j = 0; j < w; j += 8) {
132 // Filter even-index pixels
133 const int16_t *data = &im_block[i * im_stride + j];
134 const __m128i src_0 =
135 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
136 *(__m128i *)(data + 1 * im_stride));
137 const __m128i src_2 =
138 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
139 *(__m128i *)(data + 3 * im_stride));
140 const __m128i src_4 =
141 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
142 *(__m128i *)(data + 5 * im_stride));
143 const __m128i src_6 =
144 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
145 *(__m128i *)(data + 7 * im_stride));
146 const __m128i src_8 =
147 _mm_unpacklo_epi16(*(__m128i *)(data + 8 * im_stride),
148 *(__m128i *)(data + 9 * im_stride));
149 const __m128i src_10 =
150 _mm_unpacklo_epi16(*(__m128i *)(data + 10 * im_stride),
151 *(__m128i *)(data + 11 * im_stride));
152
153 const __m128i res_0 = _mm_madd_epi16(src_0, coeffs[0]);
154 const __m128i res_2 = _mm_madd_epi16(src_2, coeffs[1]);
155 const __m128i res_4 = _mm_madd_epi16(src_4, coeffs[2]);
156 const __m128i res_6 = _mm_madd_epi16(src_6, coeffs[3]);
157 const __m128i res_8 = _mm_madd_epi16(src_8, coeffs[4]);
158 const __m128i res_10 = _mm_madd_epi16(src_10, coeffs[5]);
159
160 const __m128i res_0246 = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
161 _mm_add_epi32(res_4, res_6));
162 __m128i res_even =
163 _mm_add_epi32(_mm_add_epi32(res_8, res_10), res_0246);
164
165 // Filter odd-index pixels
166 const __m128i src_1 =
167 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
168 *(__m128i *)(data + 1 * im_stride));
169 const __m128i src_3 =
170 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
171 *(__m128i *)(data + 3 * im_stride));
172 const __m128i src_5 =
173 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
174 *(__m128i *)(data + 5 * im_stride));
175 const __m128i src_7 =
176 _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
177 *(__m128i *)(data + 7 * im_stride));
178 const __m128i src_9 =
179 _mm_unpackhi_epi16(*(__m128i *)(data + 8 * im_stride),
180 *(__m128i *)(data + 9 * im_stride));
181 const __m128i src_11 =
182 _mm_unpackhi_epi16(*(__m128i *)(data + 10 * im_stride),
183 *(__m128i *)(data + 11 * im_stride));
184
185 const __m128i res_1 = _mm_madd_epi16(src_1, coeffs[0]);
186 const __m128i res_3 = _mm_madd_epi16(src_3, coeffs[1]);
187 const __m128i res_5 = _mm_madd_epi16(src_5, coeffs[2]);
188 const __m128i res_7 = _mm_madd_epi16(src_7, coeffs[3]);
189 const __m128i res_9 = _mm_madd_epi16(src_9, coeffs[4]);
190 const __m128i res_11 = _mm_madd_epi16(src_11, coeffs[5]);
191
192 const __m128i res_1357 = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
193 _mm_add_epi32(res_3, res_7));
194 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_9, res_11), res_1357);
195
196 // Rearrange pixels back into the order 0 ... 7
197 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
198 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
199
200 __m128i res_lo_round =
201 _mm_sra_epi32(_mm_add_epi32(res_lo, sum_round), sum_shift);
202 __m128i res_hi_round =
203 _mm_sra_epi32(_mm_add_epi32(res_hi, sum_round), sum_shift);
204
205 res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
206 round_shift);
207 res_hi_round = _mm_sra_epi32(_mm_add_epi32(res_hi_round, round_const),
208 round_shift);
209
210 const __m128i res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
211 const __m128i res = _mm_packus_epi16(res16, res16);
212
213 // Accumulate values into the destination buffer
214 __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
215
216 _mm_storel_epi64(p, res);
217 }
218 }
219 }
220 }
221
av1_convolve_2d_sr_sse2(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_qn,const int subpel_y_qn,ConvolveParams * conv_params)222 void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
223 int dst_stride, int w, int h,
224 const InterpFilterParams *filter_params_x,
225 const InterpFilterParams *filter_params_y,
226 const int subpel_x_qn, const int subpel_y_qn,
227 ConvolveParams *conv_params) {
228 if (filter_params_x->taps > 8) {
229 if (w < 8) {
230 av1_convolve_2d_sr_c(src, src_stride, dst, dst_stride, w, h,
231 filter_params_x, filter_params_y, subpel_x_qn,
232 subpel_y_qn, conv_params);
233 } else {
234 av1_convolve_2d_sr_12tap_sse2(src, src_stride, dst, dst_stride, w, h,
235 filter_params_x, filter_params_y,
236 subpel_x_qn, subpel_y_qn, conv_params);
237 }
238 } else {
239 const int bd = 8;
240
241 DECLARE_ALIGNED(16, int16_t,
242 im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
243 int im_h = h + filter_params_y->taps - 1;
244 int im_stride = MAX_SB_SIZE;
245 int i, j;
246 const int fo_vert = filter_params_y->taps / 2 - 1;
247 const int fo_horiz = filter_params_x->taps / 2 - 1;
248 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
249
250 const __m128i zero = _mm_setzero_si128();
251 const int bits =
252 FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
253 const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
254
255 assert(conv_params->round_0 > 0);
256
257 /* Horizontal filter */
258 {
259 const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
260 filter_params_x, subpel_x_qn & SUBPEL_MASK);
261 const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter);
262
263 // coeffs 0 1 0 1 2 3 2 3
264 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
265 // coeffs 4 5 4 5 6 7 6 7
266 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
267
268 // coeffs 0 1 0 1 0 1 0 1
269 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
270 // coeffs 2 3 2 3 2 3 2 3
271 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
272 // coeffs 4 5 4 5 4 5 4 5
273 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
274 // coeffs 6 7 6 7 6 7 6 7
275 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
276
277 const __m128i round_const = _mm_set1_epi32(
278 (1 << (bd + FILTER_BITS - 1)) + ((1 << conv_params->round_0) >> 1));
279 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
280
281 for (i = 0; i < im_h; ++i) {
282 for (j = 0; j < w; j += 8) {
283 const __m128i data =
284 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
285
286 // Filter even-index pixels
287 const __m128i src_0 = _mm_unpacklo_epi8(data, zero);
288 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
289 const __m128i src_2 =
290 _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero);
291 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
292 const __m128i src_4 =
293 _mm_unpacklo_epi8(_mm_srli_si128(data, 4), zero);
294 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
295 const __m128i src_6 =
296 _mm_unpacklo_epi8(_mm_srli_si128(data, 6), zero);
297 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
298
299 __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
300 _mm_add_epi32(res_2, res_6));
301 res_even =
302 _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
303
304 // Filter odd-index pixels
305 const __m128i src_1 =
306 _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero);
307 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
308 const __m128i src_3 =
309 _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero);
310 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
311 const __m128i src_5 =
312 _mm_unpacklo_epi8(_mm_srli_si128(data, 5), zero);
313 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
314 const __m128i src_7 =
315 _mm_unpacklo_epi8(_mm_srli_si128(data, 7), zero);
316 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
317
318 __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
319 _mm_add_epi32(res_3, res_7));
320 res_odd =
321 _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
322
323 // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
324 __m128i res = _mm_packs_epi32(res_even, res_odd);
325 _mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res);
326 }
327 }
328 }
329
330 /* Vertical filter */
331 {
332 const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
333 filter_params_y, subpel_y_qn & SUBPEL_MASK);
334 const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
335
336 // coeffs 0 1 0 1 2 3 2 3
337 const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
338 // coeffs 4 5 4 5 6 7 6 7
339 const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
340
341 // coeffs 0 1 0 1 0 1 0 1
342 const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
343 // coeffs 2 3 2 3 2 3 2 3
344 const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
345 // coeffs 4 5 4 5 4 5 4 5
346 const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
347 // coeffs 6 7 6 7 6 7 6 7
348 const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
349
350 const __m128i sum_round = _mm_set1_epi32(
351 (1 << offset_bits) + ((1 << conv_params->round_1) >> 1));
352 const __m128i sum_shift = _mm_cvtsi32_si128(conv_params->round_1);
353
354 const __m128i round_const = _mm_set1_epi32(
355 ((1 << bits) >> 1) - (1 << (offset_bits - conv_params->round_1)) -
356 ((1 << (offset_bits - conv_params->round_1)) >> 1));
357 const __m128i round_shift = _mm_cvtsi32_si128(bits);
358
359 for (i = 0; i < h; ++i) {
360 for (j = 0; j < w; j += 8) {
361 // Filter even-index pixels
362 const int16_t *data = &im_block[i * im_stride + j];
363 const __m128i src_0 =
364 _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
365 *(__m128i *)(data + 1 * im_stride));
366 const __m128i src_2 =
367 _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
368 *(__m128i *)(data + 3 * im_stride));
369 const __m128i src_4 =
370 _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
371 *(__m128i *)(data + 5 * im_stride));
372 const __m128i src_6 =
373 _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
374 *(__m128i *)(data + 7 * im_stride));
375
376 const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
377 const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
378 const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
379 const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
380
381 const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
382 _mm_add_epi32(res_4, res_6));
383
384 // Filter odd-index pixels
385 const __m128i src_1 =
386 _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
387 *(__m128i *)(data + 1 * im_stride));
388 const __m128i src_3 =
389 _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
390 *(__m128i *)(data + 3 * im_stride));
391 const __m128i src_5 =
392 _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
393 *(__m128i *)(data + 5 * im_stride));
394 const __m128i src_7 =
395 _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
396 *(__m128i *)(data + 7 * im_stride));
397
398 const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
399 const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
400 const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
401 const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
402
403 const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
404 _mm_add_epi32(res_5, res_7));
405
406 // Rearrange pixels back into the order 0 ... 7
407 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
408 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
409
410 __m128i res_lo_round =
411 _mm_sra_epi32(_mm_add_epi32(res_lo, sum_round), sum_shift);
412 __m128i res_hi_round =
413 _mm_sra_epi32(_mm_add_epi32(res_hi, sum_round), sum_shift);
414
415 res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
416 round_shift);
417 res_hi_round = _mm_sra_epi32(_mm_add_epi32(res_hi_round, round_const),
418 round_shift);
419
420 const __m128i res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
421 const __m128i res = _mm_packus_epi16(res16, res16);
422
423 // Accumulate values into the destination buffer
424 __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
425
426 if (w == 2) {
427 *(uint16_t *)p = (uint16_t)_mm_cvtsi128_si32(res);
428 } else if (w == 4) {
429 *(int *)p = _mm_cvtsi128_si32(res);
430 } else {
431 _mm_storel_epi64(p, res);
432 }
433 }
434 }
435 }
436 }
437 }
438
av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t * src,int src_stride,uint8_t * dst0,int dst_stride0,int w,int h,ConvolveParams * conv_params)439 void av1_dist_wtd_convolve_2d_copy_sse2(const uint8_t *src, int src_stride,
440 uint8_t *dst0, int dst_stride0, int w,
441 int h, ConvolveParams *conv_params) {
442 const int bd = 8;
443 CONV_BUF_TYPE *dst = conv_params->dst;
444 int dst_stride = conv_params->dst_stride;
445
446 const int bits =
447 FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
448 const int do_average = conv_params->do_average;
449 const int use_dist_wtd_comp_avg = conv_params->use_dist_wtd_comp_avg;
450 const __m128i zero = _mm_setzero_si128();
451 const __m128i left_shift = _mm_cvtsi32_si128(bits);
452 int i, j;
453
454 const int w0 = conv_params->fwd_offset;
455 const int w1 = conv_params->bck_offset;
456 const __m128i wt0 = _mm_set1_epi16(w0);
457 const __m128i wt1 = _mm_set1_epi16(w1);
458 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
459
460 const int offset_0 =
461 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
462 const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
463 const __m128i offset_const = _mm_set1_epi16(offset);
464 const int rounding_shift =
465 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
466 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
467
468 assert((w % 4) == 0);
469
470 if (!(w % 16)) {
471 for (i = 0; i < h; ++i) {
472 for (j = 0; j < w; j += 16) {
473 const __m128i d8 = _mm_loadu_si128((__m128i *)&src[j]);
474
475 const __m128i d16_lo = _mm_unpacklo_epi8(d8, zero);
476 const __m128i d16_hi = _mm_unpackhi_epi8(d8, zero);
477
478 const __m128i res_lo = _mm_sll_epi16(d16_lo, left_shift);
479 const __m128i res_unsigned_lo = _mm_add_epi16(res_lo, offset_const);
480
481 const __m128i res_hi = _mm_sll_epi16(d16_hi, left_shift);
482 const __m128i res_unsigned_hi = _mm_add_epi16(res_hi, offset_const);
483
484 if (do_average) {
485 const __m128i data_ref_0_lo = _mm_loadu_si128((__m128i *)(&dst[j]));
486 const __m128i data_ref_0_hi =
487 _mm_loadu_si128((__m128i *)(&dst[j + 8]));
488
489 const __m128i comp_avg_res_lo = comp_avg(
490 &data_ref_0_lo, &res_unsigned_lo, &wt, use_dist_wtd_comp_avg);
491
492 const __m128i round_result_lo = convolve_rounding(
493 &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
494
495 const __m128i comp_avg_res_hi = comp_avg(
496 &data_ref_0_hi, &res_unsigned_hi, &wt, use_dist_wtd_comp_avg);
497
498 const __m128i round_result_hi = convolve_rounding(
499 &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
500
501 const __m128i res_8 =
502 _mm_packus_epi16(round_result_lo, round_result_hi);
503
504 _mm_store_si128((__m128i *)(&dst0[j]), res_8);
505 } else {
506 _mm_store_si128((__m128i *)(&dst[j]), res_unsigned_lo);
507 _mm_store_si128((__m128i *)(&dst[j + 8]), res_unsigned_hi);
508 }
509 }
510 src += src_stride;
511 dst += dst_stride;
512 dst0 += dst_stride0;
513 }
514 } else {
515 for (i = 0; i < h; ++i) {
516 for (j = 0; j < w; j += 8) {
517 const __m128i d8 = _mm_loadl_epi64((__m128i *)&src[j]);
518 const __m128i d16_0 = _mm_unpacklo_epi8(d8, zero);
519
520 const __m128i res = _mm_sll_epi16(d16_0, left_shift);
521 const __m128i res_unsigned = _mm_add_epi16(res, offset_const);
522
523 if (do_average) {
524 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)(&dst[j]));
525
526 const __m128i comp_avg_res =
527 comp_avg(&data_ref_0, &res_unsigned, &wt, use_dist_wtd_comp_avg);
528
529 const __m128i round_result = convolve_rounding(
530 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
531
532 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
533
534 if (w > 4)
535 _mm_storel_epi64((__m128i *)(&dst0[j]), res_8);
536 else
537 *(int *)(&dst0[j]) = _mm_cvtsi128_si32(res_8);
538 } else {
539 _mm_store_si128((__m128i *)(&dst[j]), res_unsigned);
540 }
541 }
542 src += src_stride;
543 dst += dst_stride;
544 dst0 += dst_stride0;
545 }
546 }
547 }
548