1 /*
2 * Copyright (c) 2023 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 <arm_neon.h>
12 #include <assert.h>
13
14 #include "./vpx_config.h"
15 #include "./vpx_dsp_rtcd.h"
16 #include "vpx/vpx_integer.h"
17 #include "vpx_dsp/arm/mem_neon.h"
18 #include "vpx_dsp/arm/transpose_neon.h"
19 #include "vpx_dsp/arm/vpx_convolve8_neon.h"
20 #include "vpx_dsp/vpx_filter.h"
21 #include "vpx_ports/mem.h"
22
23 DECLARE_ALIGNED(16, static const uint8_t, dot_prod_permute_tbl[48]) = {
24 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6,
25 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10,
26 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14
27 };
28
29 DECLARE_ALIGNED(16, static const uint8_t, dot_prod_tran_concat_tbl[32]) = {
30 0, 8, 16, 24, 1, 9, 17, 25, 2, 10, 18, 26, 3, 11, 19, 27,
31 4, 12, 20, 28, 5, 13, 21, 29, 6, 14, 22, 30, 7, 15, 23, 31
32 };
33
34 DECLARE_ALIGNED(16, static const uint8_t, dot_prod_merge_block_tbl[48]) = {
35 /* Shift left and insert new last column in transposed 4x4 block. */
36 1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28,
37 /* Shift left and insert two new columns in transposed 4x4 block. */
38 2, 3, 16, 17, 6, 7, 20, 21, 10, 11, 24, 25, 14, 15, 28, 29,
39 /* Shift left and insert three new columns in transposed 4x4 block. */
40 3, 16, 17, 18, 7, 20, 21, 22, 11, 24, 25, 26, 15, 28, 29, 30
41 };
42
vpx_convolve8_2d_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)43 void vpx_convolve8_2d_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
44 uint8_t *dst, ptrdiff_t dst_stride,
45 const InterpKernel *filter, int x0_q4,
46 int x_step_q4, int y0_q4, int y_step_q4,
47 int w, int h) {
48 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[x0_q4]));
49 uint8x16_t s0, s1, s2, s3;
50
51 assert((intptr_t)dst % 4 == 0);
52 assert(dst_stride % 4 == 0);
53 assert(x_step_q4 == 16);
54 assert(h % 4 == 3);
55
56 (void)x_step_q4;
57 (void)y0_q4;
58 (void)y_step_q4;
59
60 src -= 3;
61
62 if (w == 4) {
63 const uint8x16x2_t perm_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
64 int16x4_t d0, d1, d2, d3;
65 uint8x8_t d01, d23;
66
67 do {
68 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
69
70 d0 = convolve8_4_usdot(s0, filters, perm_tbl);
71 d1 = convolve8_4_usdot(s1, filters, perm_tbl);
72 d2 = convolve8_4_usdot(s2, filters, perm_tbl);
73 d3 = convolve8_4_usdot(s3, filters, perm_tbl);
74 d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
75 d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
76
77 store_u8(dst + 0 * dst_stride, dst_stride, d01);
78 store_u8(dst + 2 * dst_stride, dst_stride, d23);
79
80 src += 4 * src_stride;
81 dst += 4 * dst_stride;
82 h -= 4;
83 } while (h > 3);
84
85 /* Process final three rows (h % 4 == 3). See vpx_convolve_neon.c for
86 * further details on possible values of block height. */
87 load_u8_16x3(src, src_stride, &s0, &s1, &s2);
88
89 d0 = convolve8_4_usdot(s0, filters, perm_tbl);
90 d1 = convolve8_4_usdot(s1, filters, perm_tbl);
91 d2 = convolve8_4_usdot(s2, filters, perm_tbl);
92 d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
93 d23 = vqrshrun_n_s16(vcombine_s16(d2, vdup_n_s16(0)), FILTER_BITS);
94
95 store_u8(dst + 0 * dst_stride, dst_stride, d01);
96 store_u8_4x1(dst + 2 * dst_stride, d23);
97 } else {
98 const uint8x16x3_t perm_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
99 const uint8_t *s;
100 uint8_t *d;
101 int width;
102 uint8x8_t d0, d1, d2, d3;
103
104 do {
105 width = w;
106 s = src;
107 d = dst;
108 do {
109 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
110
111 d0 = convolve8_8_usdot(s0, filters, perm_tbl);
112 d1 = convolve8_8_usdot(s1, filters, perm_tbl);
113 d2 = convolve8_8_usdot(s2, filters, perm_tbl);
114 d3 = convolve8_8_usdot(s3, filters, perm_tbl);
115
116 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
117
118 s += 8;
119 d += 8;
120 width -= 8;
121 } while (width > 0);
122 src += 4 * src_stride;
123 dst += 4 * dst_stride;
124 h -= 4;
125 } while (h > 3);
126
127 /* Process final three rows (h % 4 == 3). See vpx_convolve_neon.c for
128 * further details on possible values of block height. */
129 width = w;
130 s = src;
131 d = dst;
132 do {
133 load_u8_16x3(s, src_stride, &s0, &s1, &s2);
134
135 d0 = convolve8_8_usdot(s0, filters, perm_tbl);
136 d1 = convolve8_8_usdot(s1, filters, perm_tbl);
137 d2 = convolve8_8_usdot(s2, filters, perm_tbl);
138
139 store_u8_8x3(d, dst_stride, d0, d1, d2);
140
141 s += 8;
142 d += 8;
143 width -= 8;
144 } while (width > 0);
145 }
146 }
147
vpx_convolve8_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)148 void vpx_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
149 uint8_t *dst, ptrdiff_t dst_stride,
150 const InterpKernel *filter, int x0_q4,
151 int x_step_q4, int y0_q4, int y_step_q4,
152 int w, int h) {
153 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[x0_q4]));
154 uint8x16_t s0, s1, s2, s3;
155
156 assert((intptr_t)dst % 4 == 0);
157 assert(dst_stride % 4 == 0);
158 assert(x_step_q4 == 16);
159
160 (void)x_step_q4;
161 (void)y0_q4;
162 (void)y_step_q4;
163
164 src -= 3;
165
166 if (w == 4) {
167 const uint8x16x2_t perm_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
168 do {
169 int16x4_t t0, t1, t2, t3;
170 uint8x8_t d01, d23;
171
172 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
173
174 t0 = convolve8_4_usdot(s0, filters, perm_tbl);
175 t1 = convolve8_4_usdot(s1, filters, perm_tbl);
176 t2 = convolve8_4_usdot(s2, filters, perm_tbl);
177 t3 = convolve8_4_usdot(s3, filters, perm_tbl);
178 d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS);
179 d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS);
180
181 store_u8(dst + 0 * dst_stride, dst_stride, d01);
182 store_u8(dst + 2 * dst_stride, dst_stride, d23);
183
184 src += 4 * src_stride;
185 dst += 4 * dst_stride;
186 h -= 4;
187 } while (h != 0);
188 } else {
189 const uint8x16x3_t perm_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
190 const uint8_t *s;
191 uint8_t *d;
192 int width;
193 uint8x8_t d0, d1, d2, d3;
194
195 do {
196 width = w;
197 s = src;
198 d = dst;
199 do {
200 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
201
202 d0 = convolve8_8_usdot(s0, filters, perm_tbl);
203 d1 = convolve8_8_usdot(s1, filters, perm_tbl);
204 d2 = convolve8_8_usdot(s2, filters, perm_tbl);
205 d3 = convolve8_8_usdot(s3, filters, perm_tbl);
206
207 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
208
209 s += 8;
210 d += 8;
211 width -= 8;
212 } while (width != 0);
213 src += 4 * src_stride;
214 dst += 4 * dst_stride;
215 h -= 4;
216 } while (h != 0);
217 }
218 }
219
vpx_convolve8_avg_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)220 void vpx_convolve8_avg_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
221 uint8_t *dst, ptrdiff_t dst_stride,
222 const InterpKernel *filter, int x0_q4,
223 int x_step_q4, int y0_q4, int y_step_q4,
224 int w, int h) {
225 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[x0_q4]));
226 uint8x16_t s0, s1, s2, s3;
227
228 assert((intptr_t)dst % 4 == 0);
229 assert(dst_stride % 4 == 0);
230 assert(x_step_q4 == 16);
231
232 (void)x_step_q4;
233 (void)y0_q4;
234 (void)y_step_q4;
235
236 src -= 3;
237
238 if (w == 4) {
239 const uint8x16x2_t perm_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
240 do {
241 int16x4_t t0, t1, t2, t3;
242 uint8x8_t d01, d23, dd01, dd23;
243
244 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
245
246 t0 = convolve8_4_usdot(s0, filters, perm_tbl);
247 t1 = convolve8_4_usdot(s1, filters, perm_tbl);
248 t2 = convolve8_4_usdot(s2, filters, perm_tbl);
249 t3 = convolve8_4_usdot(s3, filters, perm_tbl);
250 d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS);
251 d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS);
252
253 dd01 = load_u8(dst + 0 * dst_stride, dst_stride);
254 dd23 = load_u8(dst + 2 * dst_stride, dst_stride);
255
256 d01 = vrhadd_u8(d01, dd01);
257 d23 = vrhadd_u8(d23, dd23);
258
259 store_u8(dst + 0 * dst_stride, dst_stride, d01);
260 store_u8(dst + 2 * dst_stride, dst_stride, d23);
261
262 src += 4 * src_stride;
263 dst += 4 * dst_stride;
264 h -= 4;
265 } while (h != 0);
266 } else {
267 const uint8x16x3_t perm_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
268 const uint8_t *s;
269 uint8_t *d;
270 int width;
271 uint8x8_t d0, d1, d2, d3, dd0, dd1, dd2, dd3;
272
273 do {
274 width = w;
275 s = src;
276 d = dst;
277 do {
278 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
279
280 d0 = convolve8_8_usdot(s0, filters, perm_tbl);
281 d1 = convolve8_8_usdot(s1, filters, perm_tbl);
282 d2 = convolve8_8_usdot(s2, filters, perm_tbl);
283 d3 = convolve8_8_usdot(s3, filters, perm_tbl);
284
285 load_u8_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3);
286
287 d0 = vrhadd_u8(d0, dd0);
288 d1 = vrhadd_u8(d1, dd1);
289 d2 = vrhadd_u8(d2, dd2);
290 d3 = vrhadd_u8(d3, dd3);
291
292 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
293
294 s += 8;
295 d += 8;
296 width -= 8;
297 } while (width != 0);
298 src += 4 * src_stride;
299 dst += 4 * dst_stride;
300 h -= 4;
301 } while (h != 0);
302 }
303 }
304
transpose_concat_4x4(uint8x8_t a0,uint8x8_t a1,uint8x8_t a2,uint8x8_t a3,uint8x16_t * b,const uint8x16_t permute_tbl)305 static INLINE void transpose_concat_4x4(uint8x8_t a0, uint8x8_t a1,
306 uint8x8_t a2, uint8x8_t a3,
307 uint8x16_t *b,
308 const uint8x16_t permute_tbl) {
309 /* Transpose 8-bit elements and concatenate result rows as follows:
310 * a0: 00, 01, 02, 03, XX, XX, XX, XX
311 * a1: 10, 11, 12, 13, XX, XX, XX, XX
312 * a2: 20, 21, 22, 23, XX, XX, XX, XX
313 * a3: 30, 31, 32, 33, XX, XX, XX, XX
314 *
315 * b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
316 *
317 * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
318 * as an argument is preferable to loading it directly from memory as this
319 * inline helper is called many times from the same parent function.
320 */
321
322 uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } };
323 *b = vqtbl2q_u8(samples, permute_tbl);
324 }
325
transpose_concat_8x4(uint8x8_t a0,uint8x8_t a1,uint8x8_t a2,uint8x8_t a3,uint8x16_t * b0,uint8x16_t * b1,const uint8x16x2_t permute_tbl)326 static INLINE void transpose_concat_8x4(uint8x8_t a0, uint8x8_t a1,
327 uint8x8_t a2, uint8x8_t a3,
328 uint8x16_t *b0, uint8x16_t *b1,
329 const uint8x16x2_t permute_tbl) {
330 /* Transpose 8-bit elements and concatenate result rows as follows:
331 * a0: 00, 01, 02, 03, 04, 05, 06, 07
332 * a1: 10, 11, 12, 13, 14, 15, 16, 17
333 * a2: 20, 21, 22, 23, 24, 25, 26, 27
334 * a3: 30, 31, 32, 33, 34, 35, 36, 37
335 *
336 * b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
337 * b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37
338 *
339 * The 'permute_tbl' is always 'dot_prod_tran_concat_tbl' above. Passing it
340 * as an argument is preferable to loading it directly from memory as this
341 * inline helper is called many times from the same parent function.
342 */
343
344 uint8x16x2_t samples = { { vcombine_u8(a0, a1), vcombine_u8(a2, a3) } };
345 *b0 = vqtbl2q_u8(samples, permute_tbl.val[0]);
346 *b1 = vqtbl2q_u8(samples, permute_tbl.val[1]);
347 }
348
vpx_convolve8_vert_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)349 void vpx_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
350 uint8_t *dst, ptrdiff_t dst_stride,
351 const InterpKernel *filter, int x0_q4,
352 int x_step_q4, int y0_q4, int y_step_q4,
353 int w, int h) {
354 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[y0_q4]));
355 const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl);
356 uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
357 uint8x16x2_t samples_LUT;
358
359 assert((intptr_t)dst % 4 == 0);
360 assert(dst_stride % 4 == 0);
361 assert(y_step_q4 == 16);
362
363 (void)x0_q4;
364 (void)x_step_q4;
365 (void)y_step_q4;
366
367 src -= 3 * src_stride;
368
369 if (w == 4) {
370 const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl);
371 uint8x16_t s0123, s1234, s2345, s3456, s4567, s5678, s6789, s78910;
372 int16x4_t d0, d1, d2, d3;
373 uint8x8_t d01, d23;
374
375 load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
376 src += 7 * src_stride;
377
378 s7 = vdup_n_u8(0);
379 s8 = vdup_n_u8(0);
380 s9 = vdup_n_u8(0);
381
382 /* This operation combines a conventional transpose and the sample permute
383 * (see horizontal case) required before computing the dot product.
384 */
385 transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl);
386 transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl);
387 transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl);
388 transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl);
389 transpose_concat_4x4(s4, s5, s6, s7, &s4567, tran_concat_tbl);
390 transpose_concat_4x4(s5, s6, s7, s8, &s5678, tran_concat_tbl);
391 transpose_concat_4x4(s6, s7, s8, s9, &s6789, tran_concat_tbl);
392
393 do {
394 load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10);
395
396 transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl);
397
398 /* Merge new data into block from previous iteration. */
399 samples_LUT.val[0] = s3456;
400 samples_LUT.val[1] = s78910;
401 s4567 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
402 s5678 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
403 s6789 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
404
405 d0 = convolve8_4_usdot_partial(s0123, s4567, filters);
406 d1 = convolve8_4_usdot_partial(s1234, s5678, filters);
407 d2 = convolve8_4_usdot_partial(s2345, s6789, filters);
408 d3 = convolve8_4_usdot_partial(s3456, s78910, filters);
409 d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
410 d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
411
412 store_u8(dst + 0 * dst_stride, dst_stride, d01);
413 store_u8(dst + 2 * dst_stride, dst_stride, d23);
414
415 /* Prepare block for next iteration - re-using as much as possible. */
416 /* Shuffle everything up four rows. */
417 s0123 = s4567;
418 s1234 = s5678;
419 s2345 = s6789;
420 s3456 = s78910;
421
422 src += 4 * src_stride;
423 dst += 4 * dst_stride;
424 h -= 4;
425 } while (h != 0);
426 } else {
427 const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl);
428 uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
429 s3456_lo, s3456_hi, s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo,
430 s6789_hi, s78910_lo, s78910_hi;
431 uint8x8_t d0, d1, d2, d3;
432 const uint8_t *s;
433 uint8_t *d;
434 int height;
435
436 do {
437 height = h;
438 s = src;
439 d = dst;
440
441 load_u8_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
442 s += 7 * src_stride;
443
444 s7 = vdup_n_u8(0);
445 s8 = vdup_n_u8(0);
446 s9 = vdup_n_u8(0);
447
448 /* This operation combines a conventional transpose and the sample permute
449 * (see horizontal case) required before computing the dot product.
450 */
451 transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi,
452 tran_concat_tbl);
453 transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi,
454 tran_concat_tbl);
455 transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi,
456 tran_concat_tbl);
457 transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi,
458 tran_concat_tbl);
459 transpose_concat_8x4(s4, s5, s6, s7, &s4567_lo, &s4567_hi,
460 tran_concat_tbl);
461 transpose_concat_8x4(s5, s6, s7, s8, &s5678_lo, &s5678_hi,
462 tran_concat_tbl);
463 transpose_concat_8x4(s6, s7, s8, s9, &s6789_lo, &s6789_hi,
464 tran_concat_tbl);
465
466 do {
467 load_u8_8x4(s, src_stride, &s7, &s8, &s9, &s10);
468
469 transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi,
470 tran_concat_tbl);
471
472 /* Merge new data into block from previous iteration. */
473 samples_LUT.val[0] = s3456_lo;
474 samples_LUT.val[1] = s78910_lo;
475 s4567_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
476 s5678_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
477 s6789_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
478
479 samples_LUT.val[0] = s3456_hi;
480 samples_LUT.val[1] = s78910_hi;
481 s4567_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
482 s5678_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
483 s6789_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
484
485 d0 = convolve8_8_usdot_partial(s0123_lo, s4567_lo, s0123_hi, s4567_hi,
486 filters);
487 d1 = convolve8_8_usdot_partial(s1234_lo, s5678_lo, s1234_hi, s5678_hi,
488 filters);
489 d2 = convolve8_8_usdot_partial(s2345_lo, s6789_lo, s2345_hi, s6789_hi,
490 filters);
491 d3 = convolve8_8_usdot_partial(s3456_lo, s78910_lo, s3456_hi, s78910_hi,
492 filters);
493
494 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
495
496 /* Prepare block for next iteration - re-using as much as possible. */
497 /* Shuffle everything up four rows. */
498 s0123_lo = s4567_lo;
499 s0123_hi = s4567_hi;
500 s1234_lo = s5678_lo;
501 s1234_hi = s5678_hi;
502 s2345_lo = s6789_lo;
503 s2345_hi = s6789_hi;
504 s3456_lo = s78910_lo;
505 s3456_hi = s78910_hi;
506
507 s += 4 * src_stride;
508 d += 4 * dst_stride;
509 height -= 4;
510 } while (height != 0);
511 src += 8;
512 dst += 8;
513 w -= 8;
514 } while (w != 0);
515 }
516 }
517
vpx_convolve8_avg_vert_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,const InterpKernel * filter,int x0_q4,int x_step_q4,int y0_q4,int y_step_q4,int w,int h)518 void vpx_convolve8_avg_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
519 uint8_t *dst, ptrdiff_t dst_stride,
520 const InterpKernel *filter, int x0_q4,
521 int x_step_q4, int y0_q4, int y_step_q4,
522 int w, int h) {
523 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[y0_q4]));
524 const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl);
525 uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
526 uint8x16x2_t samples_LUT;
527
528 assert((intptr_t)dst % 4 == 0);
529 assert(dst_stride % 4 == 0);
530 assert(y_step_q4 == 16);
531
532 (void)x0_q4;
533 (void)x_step_q4;
534 (void)y_step_q4;
535
536 src -= 3 * src_stride;
537
538 if (w == 4) {
539 const uint8x16_t tran_concat_tbl = vld1q_u8(dot_prod_tran_concat_tbl);
540 uint8x16_t s0123, s1234, s2345, s3456, s4567, s5678, s6789, s78910;
541 int16x4_t d0, d1, d2, d3;
542 uint8x8_t d01, d23, dd01, dd23;
543
544 load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
545 src += 7 * src_stride;
546
547 s7 = vdup_n_u8(0);
548 s8 = vdup_n_u8(0);
549 s9 = vdup_n_u8(0);
550
551 /* This operation combines a conventional transpose and the sample permute
552 * (see horizontal case) required before computing the dot product.
553 */
554 transpose_concat_4x4(s0, s1, s2, s3, &s0123, tran_concat_tbl);
555 transpose_concat_4x4(s1, s2, s3, s4, &s1234, tran_concat_tbl);
556 transpose_concat_4x4(s2, s3, s4, s5, &s2345, tran_concat_tbl);
557 transpose_concat_4x4(s3, s4, s5, s6, &s3456, tran_concat_tbl);
558 transpose_concat_4x4(s4, s5, s6, s7, &s4567, tran_concat_tbl);
559 transpose_concat_4x4(s5, s6, s7, s8, &s5678, tran_concat_tbl);
560 transpose_concat_4x4(s6, s7, s8, s9, &s6789, tran_concat_tbl);
561
562 do {
563 load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10);
564
565 transpose_concat_4x4(s7, s8, s9, s10, &s78910, tran_concat_tbl);
566
567 /* Merge new data into block from previous iteration. */
568 samples_LUT.val[0] = s3456;
569 samples_LUT.val[1] = s78910;
570 s4567 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
571 s5678 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
572 s6789 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
573
574 d0 = convolve8_4_usdot_partial(s0123, s4567, filters);
575 d1 = convolve8_4_usdot_partial(s1234, s5678, filters);
576 d2 = convolve8_4_usdot_partial(s2345, s6789, filters);
577 d3 = convolve8_4_usdot_partial(s3456, s78910, filters);
578 d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
579 d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
580
581 dd01 = load_u8(dst + 0 * dst_stride, dst_stride);
582 dd23 = load_u8(dst + 2 * dst_stride, dst_stride);
583
584 d01 = vrhadd_u8(d01, dd01);
585 d23 = vrhadd_u8(d23, dd23);
586
587 store_u8(dst + 0 * dst_stride, dst_stride, d01);
588 store_u8(dst + 2 * dst_stride, dst_stride, d23);
589
590 /* Prepare block for next iteration - re-using as much as possible. */
591 /* Shuffle everything up four rows. */
592 s0123 = s4567;
593 s1234 = s5678;
594 s2345 = s6789;
595 s3456 = s78910;
596
597 src += 4 * src_stride;
598 dst += 4 * dst_stride;
599 h -= 4;
600 } while (h != 0);
601 } else {
602 const uint8x16x2_t tran_concat_tbl = vld1q_u8_x2(dot_prod_tran_concat_tbl);
603 uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
604 s3456_lo, s3456_hi, s4567_lo, s4567_hi, s5678_lo, s5678_hi, s6789_lo,
605 s6789_hi, s78910_lo, s78910_hi;
606 uint8x8_t d0, d1, d2, d3, dd0, dd1, dd2, dd3;
607 const uint8_t *s;
608 uint8_t *d;
609 int height;
610
611 do {
612 height = h;
613 s = src;
614 d = dst;
615
616 load_u8_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
617 s += 7 * src_stride;
618
619 s7 = vdup_n_u8(0);
620 s8 = vdup_n_u8(0);
621 s9 = vdup_n_u8(0);
622
623 /* This operation combines a conventional transpose and the sample permute
624 * (see horizontal case) required before computing the dot product.
625 */
626 transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi,
627 tran_concat_tbl);
628 transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi,
629 tran_concat_tbl);
630 transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi,
631 tran_concat_tbl);
632 transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi,
633 tran_concat_tbl);
634 transpose_concat_8x4(s4, s5, s6, s7, &s4567_lo, &s4567_hi,
635 tran_concat_tbl);
636 transpose_concat_8x4(s5, s6, s7, s8, &s5678_lo, &s5678_hi,
637 tran_concat_tbl);
638 transpose_concat_8x4(s6, s7, s8, s9, &s6789_lo, &s6789_hi,
639 tran_concat_tbl);
640
641 do {
642 load_u8_8x4(s, src_stride, &s7, &s8, &s9, &s10);
643
644 transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi,
645 tran_concat_tbl);
646
647 /* Merge new data into block from previous iteration. */
648 samples_LUT.val[0] = s3456_lo;
649 samples_LUT.val[1] = s78910_lo;
650 s4567_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
651 s5678_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
652 s6789_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
653
654 samples_LUT.val[0] = s3456_hi;
655 samples_LUT.val[1] = s78910_hi;
656 s4567_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
657 s5678_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
658 s6789_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
659
660 d0 = convolve8_8_usdot_partial(s0123_lo, s4567_lo, s0123_hi, s4567_hi,
661 filters);
662 d1 = convolve8_8_usdot_partial(s1234_lo, s5678_lo, s1234_hi, s5678_hi,
663 filters);
664 d2 = convolve8_8_usdot_partial(s2345_lo, s6789_lo, s2345_hi, s6789_hi,
665 filters);
666 d3 = convolve8_8_usdot_partial(s3456_lo, s78910_lo, s3456_hi, s78910_hi,
667 filters);
668
669 load_u8_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3);
670
671 d0 = vrhadd_u8(d0, dd0);
672 d1 = vrhadd_u8(d1, dd1);
673 d2 = vrhadd_u8(d2, dd2);
674 d3 = vrhadd_u8(d3, dd3);
675
676 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
677
678 /* Prepare block for next iteration - re-using as much as possible. */
679 /* Shuffle everything up four rows. */
680 s0123_lo = s4567_lo;
681 s0123_hi = s4567_hi;
682 s1234_lo = s5678_lo;
683 s1234_hi = s5678_hi;
684 s2345_lo = s6789_lo;
685 s2345_hi = s6789_hi;
686 s3456_lo = s78910_lo;
687 s3456_hi = s78910_hi;
688
689 s += 4 * src_stride;
690 d += 4 * dst_stride;
691 height -= 4;
692 } while (height != 0);
693 src += 8;
694 dst += 8;
695 w -= 8;
696 } while (w != 0);
697 }
698 }
699