1 /*
2 * Copyright (c) 2018, 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/aom_dsp_rtcd.h"
15
16 #include "aom_dsp/aom_filter.h"
17 #include "aom_dsp/x86/convolve_sse2.h"
18
av1_jnt_convolve_x_sse2(const uint8_t * src,int src_stride,uint8_t * dst0,int dst_stride0,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_q4,const int subpel_y_q4,ConvolveParams * conv_params)19 void av1_jnt_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst0,
20 int dst_stride0, int w, int h,
21 const InterpFilterParams *filter_params_x,
22 const InterpFilterParams *filter_params_y,
23 const int subpel_x_q4, const int subpel_y_q4,
24 ConvolveParams *conv_params) {
25 const int bd = 8;
26 CONV_BUF_TYPE *dst = conv_params->dst;
27 const int dst_stride = conv_params->dst_stride;
28 const int fo_horiz = filter_params_x->taps / 2 - 1;
29 const uint8_t *src_ptr = src - fo_horiz;
30 const int bits = FILTER_BITS - conv_params->round_1;
31 const __m128i left_shift = _mm_cvtsi32_si128(bits);
32 const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_0) >> 1);
33 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
34 const int w0 = conv_params->fwd_offset;
35 const int w1 = conv_params->bck_offset;
36 const __m128i wt0 = _mm_set1_epi16(w0);
37 const __m128i wt1 = _mm_set1_epi16(w1);
38 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
39 const int do_average = conv_params->do_average;
40 const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
41 const int offset_0 =
42 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
43 const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
44 const __m128i offset_const = _mm_set1_epi16(offset);
45 const int rounding_shift =
46 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
47 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
48 __m128i coeffs[4];
49
50 (void)filter_params_y;
51 (void)subpel_y_q4;
52
53 prepare_coeffs(filter_params_x, subpel_x_q4, coeffs);
54
55 if (w == 4) {
56 do {
57 const __m128i data = _mm_loadu_si128((__m128i *)src_ptr);
58 __m128i s[4];
59
60 s[0] = _mm_unpacklo_epi8(data, _mm_srli_si128(data, 1));
61 s[1] =
62 _mm_unpacklo_epi8(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3));
63 s[2] =
64 _mm_unpacklo_epi8(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5));
65 s[3] =
66 _mm_unpacklo_epi8(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7));
67 const __m128i res_lo = convolve_lo_x(s, coeffs);
68 const __m128i res_lo_round =
69 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
70 const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift);
71
72 const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_lo_shift);
73 const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
74
75 // Accumulate values into the destination buffer
76 if (do_average) {
77 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
78
79 const __m128i comp_avg_res =
80 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
81
82 const __m128i round_result = convolve_rounding(
83 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
84
85 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
86 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
87 } else {
88 _mm_store_si128((__m128i *)(&dst[0]), res_unsigned);
89 }
90 src_ptr += src_stride;
91 dst += dst_stride;
92 dst0 += dst_stride0;
93 } while (--h);
94 } else {
95 assert(!(w % 8));
96 int i = 0;
97 do {
98 int j = 0;
99 do {
100 const __m128i data =
101 _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
102 __m128i s[4];
103
104 // Filter even-index pixels
105 s[0] = data;
106 s[1] = _mm_srli_si128(data, 2);
107 s[2] = _mm_srli_si128(data, 4);
108 s[3] = _mm_srli_si128(data, 6);
109 const __m128i res_even = convolve_lo_x(s, coeffs);
110
111 // Filter odd-index pixels
112 s[0] = _mm_srli_si128(data, 1);
113 s[1] = _mm_srli_si128(data, 3);
114 s[2] = _mm_srli_si128(data, 5);
115 s[3] = _mm_srli_si128(data, 7);
116 const __m128i res_odd = convolve_lo_x(s, coeffs);
117
118 // Rearrange pixels back into the order 0 ... 7
119 const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
120 const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
121 const __m128i res_lo_round =
122 _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
123 const __m128i res_hi_round =
124 _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
125 const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift);
126 const __m128i res_hi_shift = _mm_sll_epi32(res_hi_round, left_shift);
127
128 const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
129 const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
130
131 // Accumulate values into the destination buffer
132 if (do_average) {
133 const __m128i data_ref_0 =
134 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
135
136 const __m128i comp_avg_res =
137 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
138
139 const __m128i round_result = convolve_rounding(
140 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
141
142 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
143 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
144 } else {
145 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
146 }
147 j += 8;
148 } while (j < w);
149 } while (++i < h);
150 }
151 }
152
av1_jnt_convolve_y_sse2(const uint8_t * src,int src_stride,uint8_t * dst0,int dst_stride0,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_q4,const int subpel_y_q4,ConvolveParams * conv_params)153 void av1_jnt_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst0,
154 int dst_stride0, int w, int h,
155 const InterpFilterParams *filter_params_x,
156 const InterpFilterParams *filter_params_y,
157 const int subpel_x_q4, const int subpel_y_q4,
158 ConvolveParams *conv_params) {
159 const int bd = 8;
160 CONV_BUF_TYPE *dst = conv_params->dst;
161 const int dst_stride = conv_params->dst_stride;
162 const int fo_vert = filter_params_y->taps / 2 - 1;
163 const uint8_t *src_ptr = src - fo_vert * src_stride;
164 const int bits = FILTER_BITS - conv_params->round_0;
165 const __m128i left_shift = _mm_cvtsi32_si128(bits);
166 const __m128i wt0 = _mm_set1_epi16(conv_params->fwd_offset);
167 const __m128i wt1 = _mm_set1_epi16(conv_params->bck_offset);
168 const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
169 const int do_average = conv_params->do_average;
170 const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
171 const int offset_0 =
172 bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
173 const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
174 const __m128i offset_const = _mm_set1_epi16(offset);
175 const int rounding_shift =
176 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
177 const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
178 const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_1) >> 1);
179 const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
180 __m128i coeffs[4];
181
182 (void)filter_params_x;
183 (void)subpel_x_q4;
184
185 prepare_coeffs(filter_params_y, subpel_y_q4, coeffs);
186
187 if (w == 4) {
188 __m128i s[8], src6, res, res_shift;
189 src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 6 * src_stride));
190 s[0] = _mm_unpacklo_epi8(
191 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 0 * src_stride)),
192 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)));
193 s[1] = _mm_unpacklo_epi8(
194 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)),
195 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)));
196 s[2] = _mm_unpacklo_epi8(
197 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)),
198 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)));
199 s[3] = _mm_unpacklo_epi8(
200 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)),
201 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)));
202 s[4] = _mm_unpacklo_epi8(
203 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)),
204 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)));
205 s[5] = _mm_unpacklo_epi8(
206 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)), src6);
207
208 do {
209 s[6] = _mm_unpacklo_epi8(
210 src6, _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)));
211 src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 8 * src_stride));
212 s[7] = _mm_unpacklo_epi8(
213 _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)), src6);
214
215 res = convolve_lo_y(s + 0, coeffs);
216 res_shift = _mm_sll_epi32(res, left_shift);
217 res_shift =
218 _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift);
219
220 __m128i res_16b = _mm_packs_epi32(res_shift, res_shift);
221 __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
222
223 // Accumulate values into the destination buffer
224 if (do_average) {
225 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
226
227 const __m128i comp_avg_res =
228 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
229
230 const __m128i round_result = convolve_rounding(
231 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
232
233 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
234 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
235
236 } else {
237 _mm_store_si128((__m128i *)dst, res_unsigned);
238 }
239
240 src_ptr += src_stride;
241 dst += dst_stride;
242 dst0 += dst_stride0;
243
244 res = convolve_lo_y(s + 1, coeffs);
245 res_shift = _mm_sll_epi32(res, left_shift);
246 res_shift =
247 _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift);
248
249 res_16b = _mm_packs_epi32(res_shift, res_shift);
250 res_unsigned = _mm_add_epi16(res_16b, offset_const);
251
252 // Accumulate values into the destination buffer
253 if (do_average) {
254 const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
255
256 const __m128i comp_avg_res =
257 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
258
259 const __m128i round_result = convolve_rounding(
260 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
261
262 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
263 *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
264
265 } else {
266 _mm_store_si128((__m128i *)dst, res_unsigned);
267 }
268
269 src_ptr += src_stride;
270 dst += dst_stride;
271 dst0 += dst_stride0;
272
273 s[0] = s[2];
274 s[1] = s[3];
275 s[2] = s[4];
276 s[3] = s[5];
277 s[4] = s[6];
278 s[5] = s[7];
279 h -= 2;
280 } while (h);
281 } else {
282 assert(!(w % 8));
283 int j = 0;
284 do {
285 __m128i s[8], src6, res_lo, res_hi, res_lo_shift, res_hi_shift;
286 const uint8_t *data = &src_ptr[j];
287
288 src6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride));
289 s[0] = _mm_unpacklo_epi8(
290 _mm_loadl_epi64((__m128i *)(data + 0 * src_stride)),
291 _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)));
292 s[1] = _mm_unpacklo_epi8(
293 _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)),
294 _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)));
295 s[2] = _mm_unpacklo_epi8(
296 _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)),
297 _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)));
298 s[3] = _mm_unpacklo_epi8(
299 _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)),
300 _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)));
301 s[4] = _mm_unpacklo_epi8(
302 _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)),
303 _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)));
304 s[5] = _mm_unpacklo_epi8(
305 _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)), src6);
306
307 int i = 0;
308 do {
309 data = &src_ptr[i * src_stride + j];
310 s[6] = _mm_unpacklo_epi8(
311 src6, _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)));
312 src6 = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride));
313 s[7] = _mm_unpacklo_epi8(
314 _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)), src6);
315
316 res_lo = convolve_lo_y(s, coeffs); // Filter low index pixels
317 res_hi = convolve_hi_y(s, coeffs); // Filter high index pixels
318 res_lo_shift = _mm_sll_epi32(res_lo, left_shift);
319 res_hi_shift = _mm_sll_epi32(res_hi, left_shift);
320 res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const),
321 round_shift);
322 res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const),
323 round_shift);
324
325 __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
326 __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
327
328 // Accumulate values into the destination buffer
329 if (do_average) {
330 const __m128i data_ref_0 =
331 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
332
333 const __m128i comp_avg_res =
334 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
335
336 const __m128i round_result = convolve_rounding(
337 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
338
339 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
340 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
341 } else {
342 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
343 }
344 i++;
345
346 res_lo = convolve_lo_y(s + 1, coeffs); // Filter low index pixels
347 res_hi = convolve_hi_y(s + 1, coeffs); // Filter high index pixels
348 res_lo_shift = _mm_sll_epi32(res_lo, left_shift);
349 res_hi_shift = _mm_sll_epi32(res_hi, left_shift);
350 res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const),
351 round_shift);
352 res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const),
353 round_shift);
354 res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
355 res_unsigned = _mm_add_epi16(res_16b, offset_const);
356
357 // Accumulate values into the destination buffer
358 if (do_average) {
359 __m128i data_ref_0 =
360 _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
361
362 const __m128i comp_avg_res =
363 comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
364
365 const __m128i round_result = convolve_rounding(
366 &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
367
368 const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
369 _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
370 } else {
371 _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
372 }
373 i++;
374
375 s[0] = s[2];
376 s[1] = s[3];
377 s[2] = s[4];
378 s[3] = s[5];
379 s[4] = s[6];
380 s[5] = s[7];
381 } while (i < h);
382 j += 8;
383 } while (j < w);
384 }
385 }
386