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_merge_block_tbl[48]) = {
30 // Shift left and insert new last column in transposed 4x4 block.
31 1, 2, 3, 16, 5, 6, 7, 20, 9, 10, 11, 24, 13, 14, 15, 28,
32 // Shift left and insert two new columns in transposed 4x4 block.
33 2, 3, 16, 17, 6, 7, 20, 21, 10, 11, 24, 25, 14, 15, 28, 29,
34 // Shift left and insert three new columns in transposed 4x4 block.
35 3, 16, 17, 18, 7, 20, 21, 22, 11, 24, 25, 26, 15, 28, 29, 30
36 };
37
convolve4_4_h(const uint8x16_t samples,const int8x8_t filters,const uint8x16_t permute_tbl)38 static INLINE int16x4_t convolve4_4_h(const uint8x16_t samples,
39 const int8x8_t filters,
40 const uint8x16_t permute_tbl) {
41 // Permute samples ready for dot product.
42 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
43 uint8x16_t permuted_samples = vqtbl1q_u8(samples, permute_tbl);
44
45 int32x4_t sum =
46 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples, filters, 0);
47
48 // Further narrowing and packing is performed by the caller.
49 return vmovn_s32(sum);
50 }
51
convolve4_8_h(const uint8x16_t samples,const int8x8_t filters,const uint8x16x2_t permute_tbl)52 static INLINE uint8x8_t convolve4_8_h(const uint8x16_t samples,
53 const int8x8_t filters,
54 const uint8x16x2_t permute_tbl) {
55 // Permute samples ready for dot product.
56 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
57 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 }
58 uint8x16_t permuted_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]),
59 vqtbl1q_u8(samples, permute_tbl.val[1]) };
60
61 // First 4 output values.
62 int32x4_t sum0 =
63 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
64 // Second 4 output values.
65 int32x4_t sum1 =
66 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filters, 0);
67
68 // Narrow and re-pack.
69 int16x8_t sum = vcombine_s16(vmovn_s32(sum0), vmovn_s32(sum1));
70 // We halved the filter values so -1 from right shift.
71 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
72 }
73
convolve8_4_h(const uint8x16_t samples,const int8x8_t filters,const uint8x16x2_t permute_tbl)74 static INLINE int16x4_t convolve8_4_h(const uint8x16_t samples,
75 const int8x8_t filters,
76 const uint8x16x2_t permute_tbl) {
77 // Permute samples ready for dot product.
78 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
79 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 }
80 uint8x16_t permuted_samples[2] = { vqtbl1q_u8(samples, permute_tbl.val[0]),
81 vqtbl1q_u8(samples, permute_tbl.val[1]) };
82
83 int32x4_t sum =
84 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
85 sum = vusdotq_lane_s32(sum, permuted_samples[1], filters, 1);
86
87 // Further narrowing and packing is performed by the caller.
88 return vshrn_n_s32(sum, 1);
89 }
90
convolve8_8_h(const uint8x16_t samples,const int8x8_t filters,const uint8x16x3_t permute_tbl)91 static INLINE uint8x8_t convolve8_8_h(const uint8x16_t samples,
92 const int8x8_t filters,
93 const uint8x16x3_t permute_tbl) {
94 // Permute samples ready for dot product.
95 // { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 }
96 // { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 }
97 // { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 }
98 uint8x16_t permuted_samples[3] = { vqtbl1q_u8(samples, permute_tbl.val[0]),
99 vqtbl1q_u8(samples, permute_tbl.val[1]),
100 vqtbl1q_u8(samples, permute_tbl.val[2]) };
101
102 // First 4 output values.
103 int32x4_t sum0 =
104 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
105 sum0 = vusdotq_lane_s32(sum0, permuted_samples[1], filters, 1);
106 // Second 4 output values.
107 int32x4_t sum1 =
108 vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filters, 0);
109 sum1 = vusdotq_lane_s32(sum1, permuted_samples[2], filters, 1);
110
111 // Narrow and re-pack.
112 int16x8_t sum = vcombine_s16(vshrn_n_s32(sum0, 1), vshrn_n_s32(sum1, 1));
113 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
114 }
115
convolve_4tap_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,int w,int h,const int8x8_t filter)116 static INLINE void convolve_4tap_horiz_neon_i8mm(const uint8_t *src,
117 ptrdiff_t src_stride,
118 uint8_t *dst,
119 ptrdiff_t dst_stride, int w,
120 int h, const int8x8_t filter) {
121 if (w == 4) {
122 const uint8x16_t permute_tbl = vld1q_u8(dot_prod_permute_tbl);
123
124 do {
125 uint8x16_t s0, s1, s2, s3;
126 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
127
128 int16x4_t t0 = convolve4_4_h(s0, filter, permute_tbl);
129 int16x4_t t1 = convolve4_4_h(s1, filter, permute_tbl);
130 int16x4_t t2 = convolve4_4_h(s2, filter, permute_tbl);
131 int16x4_t t3 = convolve4_4_h(s3, filter, permute_tbl);
132 // We halved the filter values so -1 from right shift.
133 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS - 1);
134 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS - 1);
135
136 store_u8(dst + 0 * dst_stride, dst_stride, d01);
137 store_u8(dst + 2 * dst_stride, dst_stride, d23);
138
139 src += 4 * src_stride;
140 dst += 4 * dst_stride;
141 h -= 4;
142 } while (h != 0);
143 } else {
144 const uint8x16x2_t permute_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
145
146 do {
147 const uint8_t *s = src;
148 uint8_t *d = dst;
149 int width = w;
150
151 do {
152 uint8x16_t s0, s1, s2, s3;
153 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
154
155 uint8x8_t d0 = convolve4_8_h(s0, filter, permute_tbl);
156 uint8x8_t d1 = convolve4_8_h(s1, filter, permute_tbl);
157 uint8x8_t d2 = convolve4_8_h(s2, filter, permute_tbl);
158 uint8x8_t d3 = convolve4_8_h(s3, filter, permute_tbl);
159
160 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
161
162 s += 8;
163 d += 8;
164 width -= 8;
165 } while (width != 0);
166 src += 4 * src_stride;
167 dst += 4 * dst_stride;
168 h -= 4;
169 } while (h != 0);
170 }
171 }
172
convolve_8tap_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,int w,int h,const int8x8_t filter)173 static INLINE void convolve_8tap_horiz_neon_i8mm(const uint8_t *src,
174 ptrdiff_t src_stride,
175 uint8_t *dst,
176 ptrdiff_t dst_stride, int w,
177 int h, const int8x8_t filter) {
178 if (w == 4) {
179 const uint8x16x2_t permute_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
180
181 do {
182 uint8x16_t s0, s1, s2, s3;
183 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
184
185 int16x4_t t0 = convolve8_4_h(s0, filter, permute_tbl);
186 int16x4_t t1 = convolve8_4_h(s1, filter, permute_tbl);
187 int16x4_t t2 = convolve8_4_h(s2, filter, permute_tbl);
188 int16x4_t t3 = convolve8_4_h(s3, filter, permute_tbl);
189 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS - 1);
190 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS - 1);
191
192 store_u8(dst + 0 * dst_stride, dst_stride, d01);
193 store_u8(dst + 2 * dst_stride, dst_stride, d23);
194
195 src += 4 * src_stride;
196 dst += 4 * dst_stride;
197 h -= 4;
198 } while (h != 0);
199 } else {
200 const uint8x16x3_t permute_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
201
202 do {
203 const uint8_t *s = src;
204 uint8_t *d = dst;
205 int width = w;
206
207 do {
208 uint8x16_t s0, s1, s2, s3;
209 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
210
211 uint8x8_t d0 = convolve8_8_h(s0, filter, permute_tbl);
212 uint8x8_t d1 = convolve8_8_h(s1, filter, permute_tbl);
213 uint8x8_t d2 = convolve8_8_h(s2, filter, permute_tbl);
214 uint8x8_t d3 = convolve8_8_h(s3, filter, permute_tbl);
215
216 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
217
218 s += 8;
219 d += 8;
220 width -= 8;
221 } while (width != 0);
222 src += 4 * src_stride;
223 dst += 4 * dst_stride;
224 h -= 4;
225 } while (h != 0);
226 }
227 }
228
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)229 void vpx_convolve8_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
230 uint8_t *dst, ptrdiff_t dst_stride,
231 const InterpKernel *filter, int x0_q4,
232 int x_step_q4, int y0_q4, int y_step_q4,
233 int w, int h) {
234 assert((intptr_t)dst % 4 == 0);
235 assert(dst_stride % 4 == 0);
236 assert(x_step_q4 == 16);
237
238 (void)x_step_q4;
239 (void)y0_q4;
240 (void)y_step_q4;
241
242 if (vpx_get_filter_taps(filter[x0_q4]) <= 4) {
243 // Load 4-tap filter into first 4 elements of the vector.
244 // All 4-tap and bilinear filter values are even, so halve them to reduce
245 // intermediate precision requirements.
246 const int16x4_t x_filter = vld1_s16(filter[x0_q4] + 2);
247 const int8x8_t x_filter_4tap =
248 vshrn_n_s16(vcombine_s16(x_filter, vdup_n_s16(0)), 1);
249
250 convolve_4tap_horiz_neon_i8mm(src - 1, src_stride, dst, dst_stride, w, h,
251 x_filter_4tap);
252
253 } else {
254 const int8x8_t x_filter_8tap = vmovn_s16(vld1q_s16(filter[x0_q4]));
255
256 convolve_8tap_horiz_neon_i8mm(src - 3, src_stride, dst, dst_stride, w, h,
257 x_filter_8tap);
258 }
259 }
260
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)261 void vpx_convolve8_avg_horiz_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
262 uint8_t *dst, ptrdiff_t dst_stride,
263 const InterpKernel *filter, int x0_q4,
264 int x_step_q4, int y0_q4, int y_step_q4,
265 int w, int h) {
266 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[x0_q4]));
267
268 assert((intptr_t)dst % 4 == 0);
269 assert(dst_stride % 4 == 0);
270 assert(x_step_q4 == 16);
271
272 (void)x_step_q4;
273 (void)y0_q4;
274 (void)y_step_q4;
275
276 src -= 3;
277
278 if (w == 4) {
279 const uint8x16x2_t permute_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
280
281 do {
282 uint8x16_t s0, s1, s2, s3;
283 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
284
285 int16x4_t t0 = convolve8_4_h(s0, filters, permute_tbl);
286 int16x4_t t1 = convolve8_4_h(s1, filters, permute_tbl);
287 int16x4_t t2 = convolve8_4_h(s2, filters, permute_tbl);
288 int16x4_t t3 = convolve8_4_h(s3, filters, permute_tbl);
289 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS - 1);
290 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(t2, t3), FILTER_BITS - 1);
291
292 uint8x8_t dd01 = load_u8(dst + 0 * dst_stride, dst_stride);
293 uint8x8_t dd23 = load_u8(dst + 2 * dst_stride, dst_stride);
294
295 d01 = vrhadd_u8(d01, dd01);
296 d23 = vrhadd_u8(d23, dd23);
297
298 store_u8(dst + 0 * dst_stride, dst_stride, d01);
299 store_u8(dst + 2 * dst_stride, dst_stride, d23);
300
301 src += 4 * src_stride;
302 dst += 4 * dst_stride;
303 h -= 4;
304 } while (h != 0);
305 } else {
306 const uint8x16x3_t permute_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
307
308 do {
309 const uint8_t *s = src;
310 uint8_t *d = dst;
311 int width = w;
312
313 do {
314 uint8x16_t s0, s1, s2, s3;
315 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
316
317 uint8x8_t d0 = convolve8_8_h(s0, filters, permute_tbl);
318 uint8x8_t d1 = convolve8_8_h(s1, filters, permute_tbl);
319 uint8x8_t d2 = convolve8_8_h(s2, filters, permute_tbl);
320 uint8x8_t d3 = convolve8_8_h(s3, filters, permute_tbl);
321
322 uint8x8_t dd0, dd1, dd2, dd3;
323 load_u8_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3);
324
325 d0 = vrhadd_u8(d0, dd0);
326 d1 = vrhadd_u8(d1, dd1);
327 d2 = vrhadd_u8(d2, dd2);
328 d3 = vrhadd_u8(d3, dd3);
329
330 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
331
332 s += 8;
333 d += 8;
334 width -= 8;
335 } while (width != 0);
336 src += 4 * src_stride;
337 dst += 4 * dst_stride;
338 h -= 4;
339 } while (h != 0);
340 }
341 }
342
transpose_concat_4x4(uint8x8_t a0,uint8x8_t a1,uint8x8_t a2,uint8x8_t a3,uint8x16_t * b)343 static INLINE void transpose_concat_4x4(uint8x8_t a0, uint8x8_t a1,
344 uint8x8_t a2, uint8x8_t a3,
345 uint8x16_t *b) {
346 // Transpose 8-bit elements and concatenate result rows as follows:
347 // a0: 00, 01, 02, 03, XX, XX, XX, XX
348 // a1: 10, 11, 12, 13, XX, XX, XX, XX
349 // a2: 20, 21, 22, 23, XX, XX, XX, XX
350 // a3: 30, 31, 32, 33, XX, XX, XX, XX
351 //
352 // b: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
353
354 uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0));
355 uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0));
356 uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0));
357 uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0));
358
359 uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0];
360 uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0];
361
362 uint16x8_t a0123 =
363 vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23)).val[0];
364
365 *b = vreinterpretq_u8_u16(a0123);
366 }
367
transpose_concat_8x4(uint8x8_t a0,uint8x8_t a1,uint8x8_t a2,uint8x8_t a3,uint8x16_t * b0,uint8x16_t * b1)368 static INLINE void transpose_concat_8x4(uint8x8_t a0, uint8x8_t a1,
369 uint8x8_t a2, uint8x8_t a3,
370 uint8x16_t *b0, uint8x16_t *b1) {
371 // Transpose 8-bit elements and concatenate result rows as follows:
372 // a0: 00, 01, 02, 03, 04, 05, 06, 07
373 // a1: 10, 11, 12, 13, 14, 15, 16, 17
374 // a2: 20, 21, 22, 23, 24, 25, 26, 27
375 // a3: 30, 31, 32, 33, 34, 35, 36, 37
376 //
377 // b0: 00, 10, 20, 30, 01, 11, 21, 31, 02, 12, 22, 32, 03, 13, 23, 33
378 // b1: 04, 14, 24, 34, 05, 15, 25, 35, 06, 16, 26, 36, 07, 17, 27, 37
379
380 uint8x16_t a0q = vcombine_u8(a0, vdup_n_u8(0));
381 uint8x16_t a1q = vcombine_u8(a1, vdup_n_u8(0));
382 uint8x16_t a2q = vcombine_u8(a2, vdup_n_u8(0));
383 uint8x16_t a3q = vcombine_u8(a3, vdup_n_u8(0));
384
385 uint8x16_t a01 = vzipq_u8(a0q, a1q).val[0];
386 uint8x16_t a23 = vzipq_u8(a2q, a3q).val[0];
387
388 uint16x8x2_t a0123 =
389 vzipq_u16(vreinterpretq_u16_u8(a01), vreinterpretq_u16_u8(a23));
390
391 *b0 = vreinterpretq_u8_u16(a0123.val[0]);
392 *b1 = vreinterpretq_u8_u16(a0123.val[1]);
393 }
394
convolve8_4_v(const uint8x16_t samples_lo,const uint8x16_t samples_hi,const int8x8_t filters)395 static INLINE int16x4_t convolve8_4_v(const uint8x16_t samples_lo,
396 const uint8x16_t samples_hi,
397 const int8x8_t filters) {
398 // Sample permutation is performed by the caller.
399 int32x4_t sum = vusdotq_lane_s32(vdupq_n_s32(0), samples_lo, filters, 0);
400 sum = vusdotq_lane_s32(sum, samples_hi, filters, 1);
401
402 // Further narrowing and packing is performed by the caller.
403 return vshrn_n_s32(sum, 1);
404 }
405
convolve8_8_v(const uint8x16_t samples0_lo,const uint8x16_t samples0_hi,const uint8x16_t samples1_lo,const uint8x16_t samples1_hi,const int8x8_t filters)406 static INLINE uint8x8_t convolve8_8_v(const uint8x16_t samples0_lo,
407 const uint8x16_t samples0_hi,
408 const uint8x16_t samples1_lo,
409 const uint8x16_t samples1_hi,
410 const int8x8_t filters) {
411 // Sample permutation is performed by the caller.
412
413 // First 4 output values.
414 int32x4_t sum0 = vusdotq_lane_s32(vdupq_n_s32(0), samples0_lo, filters, 0);
415 sum0 = vusdotq_lane_s32(sum0, samples0_hi, filters, 1);
416 // Second 4 output values.
417 int32x4_t sum1 = vusdotq_lane_s32(vdupq_n_s32(0), samples1_lo, filters, 0);
418 sum1 = vusdotq_lane_s32(sum1, samples1_hi, filters, 1);
419
420 // Narrow and re-pack.
421 int16x8_t sum = vcombine_s16(vshrn_n_s32(sum0, 1), vshrn_n_s32(sum1, 1));
422 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
423 }
424
convolve_8tap_vert_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,int w,int h,const int8x8_t filter)425 static INLINE void convolve_8tap_vert_neon_i8mm(const uint8_t *src,
426 ptrdiff_t src_stride,
427 uint8_t *dst,
428 ptrdiff_t dst_stride, int w,
429 int h, const int8x8_t filter) {
430 const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl);
431 if (w == 4) {
432 uint8x8_t s0, s1, s2, s3, s4, s5, s6;
433 load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
434 src += 7 * src_stride;
435
436 // This operation combines a conventional transpose and the sample permute
437 // (see horizontal case) required before computing the dot product.
438 uint8x16_t s0123, s1234, s2345, s3456;
439 transpose_concat_4x4(s0, s1, s2, s3, &s0123);
440 transpose_concat_4x4(s1, s2, s3, s4, &s1234);
441 transpose_concat_4x4(s2, s3, s4, s5, &s2345);
442 transpose_concat_4x4(s3, s4, s5, s6, &s3456);
443
444 do {
445 uint8x8_t s7, s8, s9, s10;
446 load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10);
447
448 uint8x16_t s78910;
449 transpose_concat_4x4(s7, s8, s9, s10, &s78910);
450
451 // Merge new data into block from previous iteration.
452 uint8x16x2_t samples_LUT = { { s3456, s78910 } };
453 uint8x16_t s4567 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
454 uint8x16_t s5678 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
455 uint8x16_t s6789 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
456
457 int16x4_t d0 = convolve8_4_v(s0123, s4567, filter);
458 int16x4_t d1 = convolve8_4_v(s1234, s5678, filter);
459 int16x4_t d2 = convolve8_4_v(s2345, s6789, filter);
460 int16x4_t d3 = convolve8_4_v(s3456, s78910, filter);
461 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
462 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
463
464 store_u8(dst + 0 * dst_stride, dst_stride, d01);
465 store_u8(dst + 2 * dst_stride, dst_stride, d23);
466
467 // Prepare block for next iteration - re-using as much as possible.
468 // Shuffle everything up four rows.
469 s0123 = s4567;
470 s1234 = s5678;
471 s2345 = s6789;
472 s3456 = s78910;
473
474 src += 4 * src_stride;
475 dst += 4 * dst_stride;
476 h -= 4;
477 } while (h != 0);
478 } else {
479 do {
480 const uint8_t *s = src;
481 uint8_t *d = dst;
482 int height = h;
483
484 uint8x8_t s0, s1, s2, s3, s4, s5, s6;
485 load_u8_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
486 s += 7 * src_stride;
487
488 // This operation combines a conventional transpose and the sample permute
489 // (see horizontal case) required before computing the dot product.
490 uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
491 s3456_lo, s3456_hi;
492 transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi);
493 transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi);
494 transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi);
495 transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi);
496
497 do {
498 uint8x8_t s7, s8, s9, s10;
499 load_u8_8x4(s, src_stride, &s7, &s8, &s9, &s10);
500
501 uint8x16_t s78910_lo, s78910_hi;
502 transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi);
503
504 // Merge new data into block from previous iteration.
505 uint8x16x2_t samples_LUT = { { s3456_lo, s78910_lo } };
506 uint8x16_t s4567_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
507 uint8x16_t s5678_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
508 uint8x16_t s6789_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
509
510 samples_LUT.val[0] = s3456_hi;
511 samples_LUT.val[1] = s78910_hi;
512 uint8x16_t s4567_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
513 uint8x16_t s5678_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
514 uint8x16_t s6789_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
515
516 uint8x8_t d0 =
517 convolve8_8_v(s0123_lo, s4567_lo, s0123_hi, s4567_hi, filter);
518 uint8x8_t d1 =
519 convolve8_8_v(s1234_lo, s5678_lo, s1234_hi, s5678_hi, filter);
520 uint8x8_t d2 =
521 convolve8_8_v(s2345_lo, s6789_lo, s2345_hi, s6789_hi, filter);
522 uint8x8_t d3 =
523 convolve8_8_v(s3456_lo, s78910_lo, s3456_hi, s78910_hi, filter);
524
525 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
526
527 // Prepare block for next iteration - re-using as much as possible.
528 // Shuffle everything up four rows.
529 s0123_lo = s4567_lo;
530 s0123_hi = s4567_hi;
531 s1234_lo = s5678_lo;
532 s1234_hi = s5678_hi;
533 s2345_lo = s6789_lo;
534 s2345_hi = s6789_hi;
535 s3456_lo = s78910_lo;
536 s3456_hi = s78910_hi;
537
538 s += 4 * src_stride;
539 d += 4 * dst_stride;
540 height -= 4;
541 } while (height != 0);
542 src += 8;
543 dst += 8;
544 w -= 8;
545 } while (w != 0);
546 }
547 }
548
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)549 void vpx_convolve8_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
550 uint8_t *dst, ptrdiff_t dst_stride,
551 const InterpKernel *filter, int x0_q4,
552 int x_step_q4, int y0_q4, int y_step_q4,
553 int w, int h) {
554 assert((intptr_t)dst % 4 == 0);
555 assert(dst_stride % 4 == 0);
556 assert(y_step_q4 == 16);
557
558 (void)x0_q4;
559 (void)x_step_q4;
560 (void)y_step_q4;
561
562 if (vpx_get_filter_taps(filter[y0_q4]) <= 4) {
563 const int16x8_t y_filter = vld1q_s16(filter[y0_q4]);
564
565 convolve_4tap_vert_neon(src - src_stride, src_stride, dst, dst_stride, w, h,
566 y_filter);
567 } else {
568 const int8x8_t y_filter = vmovn_s16(vld1q_s16(filter[y0_q4]));
569
570 convolve_8tap_vert_neon_i8mm(src - 3 * src_stride, src_stride, dst,
571 dst_stride, w, h, y_filter);
572 }
573 }
574
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)575 void vpx_convolve8_avg_vert_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
576 uint8_t *dst, ptrdiff_t dst_stride,
577 const InterpKernel *filter, int x0_q4,
578 int x_step_q4, int y0_q4, int y_step_q4,
579 int w, int h) {
580 const int8x8_t filters = vmovn_s16(vld1q_s16(filter[y0_q4]));
581 const uint8x16x3_t merge_block_tbl = vld1q_u8_x3(dot_prod_merge_block_tbl);
582
583 assert((intptr_t)dst % 4 == 0);
584 assert(dst_stride % 4 == 0);
585 assert(y_step_q4 == 16);
586
587 (void)x0_q4;
588 (void)x_step_q4;
589 (void)y_step_q4;
590
591 src -= 3 * src_stride;
592
593 if (w == 4) {
594 uint8x8_t s0, s1, s2, s3, s4, s5, s6;
595 load_u8_8x7(src, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
596 src += 7 * src_stride;
597
598 // This operation combines a conventional transpose and the sample permute
599 // (see horizontal case) required before computing the dot product.
600 uint8x16_t s0123, s1234, s2345, s3456;
601 transpose_concat_4x4(s0, s1, s2, s3, &s0123);
602 transpose_concat_4x4(s1, s2, s3, s4, &s1234);
603 transpose_concat_4x4(s2, s3, s4, s5, &s2345);
604 transpose_concat_4x4(s3, s4, s5, s6, &s3456);
605
606 do {
607 uint8x8_t s7, s8, s9, s10;
608 load_u8_8x4(src, src_stride, &s7, &s8, &s9, &s10);
609
610 uint8x16_t s78910;
611 transpose_concat_4x4(s7, s8, s9, s10, &s78910);
612
613 // Merge new data into block from previous iteration.
614 uint8x16x2_t samples_LUT = { { s3456, s78910 } };
615 uint8x16_t s4567 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
616 uint8x16_t s5678 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
617 uint8x16_t s6789 = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
618
619 int16x4_t d0 = convolve8_4_v(s0123, s4567, filters);
620 int16x4_t d1 = convolve8_4_v(s1234, s5678, filters);
621 int16x4_t d2 = convolve8_4_v(s2345, s6789, filters);
622 int16x4_t d3 = convolve8_4_v(s3456, s78910, filters);
623 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
624 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
625
626 uint8x8_t dd01 = load_u8(dst + 0 * dst_stride, dst_stride);
627 uint8x8_t dd23 = load_u8(dst + 2 * dst_stride, dst_stride);
628
629 d01 = vrhadd_u8(d01, dd01);
630 d23 = vrhadd_u8(d23, dd23);
631
632 store_u8(dst + 0 * dst_stride, dst_stride, d01);
633 store_u8(dst + 2 * dst_stride, dst_stride, d23);
634
635 // Prepare block for next iteration - re-using as much as possible.
636 // Shuffle everything up four rows.
637 s0123 = s4567;
638 s1234 = s5678;
639 s2345 = s6789;
640 s3456 = s78910;
641
642 src += 4 * src_stride;
643 dst += 4 * dst_stride;
644 h -= 4;
645 } while (h != 0);
646 } else {
647 do {
648 const uint8_t *s = src;
649 uint8_t *d = dst;
650 int height = h;
651
652 uint8x8_t s0, s1, s2, s3, s4, s5, s6;
653 load_u8_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
654 s += 7 * src_stride;
655
656 // This operation combines a conventional transpose and the sample permute
657 // (see horizontal case) required before computing the dot product.
658 uint8x16_t s0123_lo, s0123_hi, s1234_lo, s1234_hi, s2345_lo, s2345_hi,
659 s3456_lo, s3456_hi;
660 transpose_concat_8x4(s0, s1, s2, s3, &s0123_lo, &s0123_hi);
661 transpose_concat_8x4(s1, s2, s3, s4, &s1234_lo, &s1234_hi);
662 transpose_concat_8x4(s2, s3, s4, s5, &s2345_lo, &s2345_hi);
663 transpose_concat_8x4(s3, s4, s5, s6, &s3456_lo, &s3456_hi);
664
665 do {
666 uint8x8_t s7, s8, s9, s10;
667 load_u8_8x4(s, src_stride, &s7, &s8, &s9, &s10);
668
669 uint8x16_t s78910_lo, s78910_hi;
670 transpose_concat_8x4(s7, s8, s9, s10, &s78910_lo, &s78910_hi);
671
672 // Merge new data into block from previous iteration.
673 uint8x16x2_t samples_LUT = { { s3456_lo, s78910_lo } };
674 uint8x16_t s4567_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
675 uint8x16_t s5678_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
676 uint8x16_t s6789_lo = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
677
678 samples_LUT.val[0] = s3456_hi;
679 samples_LUT.val[1] = s78910_hi;
680 uint8x16_t s4567_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[0]);
681 uint8x16_t s5678_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[1]);
682 uint8x16_t s6789_hi = vqtbl2q_u8(samples_LUT, merge_block_tbl.val[2]);
683
684 uint8x8_t d0 =
685 convolve8_8_v(s0123_lo, s4567_lo, s0123_hi, s4567_hi, filters);
686 uint8x8_t d1 =
687 convolve8_8_v(s1234_lo, s5678_lo, s1234_hi, s5678_hi, filters);
688 uint8x8_t d2 =
689 convolve8_8_v(s2345_lo, s6789_lo, s2345_hi, s6789_hi, filters);
690 uint8x8_t d3 =
691 convolve8_8_v(s3456_lo, s78910_lo, s3456_hi, s78910_hi, filters);
692
693 uint8x8_t dd0, dd1, dd2, dd3;
694 load_u8_8x4(d, dst_stride, &dd0, &dd1, &dd2, &dd3);
695
696 d0 = vrhadd_u8(d0, dd0);
697 d1 = vrhadd_u8(d1, dd1);
698 d2 = vrhadd_u8(d2, dd2);
699 d3 = vrhadd_u8(d3, dd3);
700
701 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
702
703 /* Prepare block for next iteration - re-using as much as possible. */
704 /* Shuffle everything up four rows. */
705 s0123_lo = s4567_lo;
706 s0123_hi = s4567_hi;
707 s1234_lo = s5678_lo;
708 s1234_hi = s5678_hi;
709 s2345_lo = s6789_lo;
710 s2345_hi = s6789_hi;
711 s3456_lo = s78910_lo;
712 s3456_hi = s78910_hi;
713
714 s += 4 * src_stride;
715 d += 4 * dst_stride;
716 height -= 4;
717 } while (height != 0);
718 src += 8;
719 dst += 8;
720 w -= 8;
721 } while (w != 0);
722 }
723 }
724
convolve_4tap_2d_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,int w,int h,const int8x8_t x_filter,const uint8x8_t y_filter)725 static INLINE void convolve_4tap_2d_neon_i8mm(const uint8_t *src,
726 ptrdiff_t src_stride,
727 uint8_t *dst,
728 ptrdiff_t dst_stride, int w,
729 int h, const int8x8_t x_filter,
730 const uint8x8_t y_filter) {
731 // Neon does not have lane-referencing multiply or multiply-accumulate
732 // instructions that operate on vectors of 8-bit elements. This means we have
733 // to duplicate filter taps into a whole vector and use standard multiply /
734 // multiply-accumulate instructions.
735 const uint8x8_t y_filter_taps[4] = { vdup_lane_u8(y_filter, 2),
736 vdup_lane_u8(y_filter, 3),
737 vdup_lane_u8(y_filter, 4),
738 vdup_lane_u8(y_filter, 5) };
739
740 if (w == 4) {
741 const uint8x16_t permute_tbl = vld1q_u8(dot_prod_permute_tbl);
742
743 uint8x16_t h_s0, h_s1, h_s2;
744 load_u8_16x3(src, src_stride, &h_s0, &h_s1, &h_s2);
745
746 int16x4_t t0 = convolve4_4_h(h_s0, x_filter, permute_tbl);
747 int16x4_t t1 = convolve4_4_h(h_s1, x_filter, permute_tbl);
748 int16x4_t t2 = convolve4_4_h(h_s2, x_filter, permute_tbl);
749 // We halved the filter values so -1 from right shift.
750 uint8x8_t v_s01 = vqrshrun_n_s16(vcombine_s16(t0, t1), FILTER_BITS - 1);
751 uint8x8_t v_s12 = vqrshrun_n_s16(vcombine_s16(t1, t2), FILTER_BITS - 1);
752
753 src += 3 * src_stride;
754
755 do {
756 uint8x16_t h_s3, h_s4, h_s5, h_s6;
757 load_u8_16x4(src, src_stride, &h_s3, &h_s4, &h_s5, &h_s6);
758
759 int16x4_t t3 = convolve4_4_h(h_s3, x_filter, permute_tbl);
760 int16x4_t t4 = convolve4_4_h(h_s4, x_filter, permute_tbl);
761 int16x4_t t5 = convolve4_4_h(h_s5, x_filter, permute_tbl);
762 int16x4_t t6 = convolve4_4_h(h_s6, x_filter, permute_tbl);
763 // We halved the filter values so -1 from right shift.
764 uint8x8_t v_s34 = vqrshrun_n_s16(vcombine_s16(t3, t4), FILTER_BITS - 1);
765 uint8x8_t v_s56 = vqrshrun_n_s16(vcombine_s16(t5, t6), FILTER_BITS - 1);
766 uint8x8_t v_s23 = vext_u8(v_s12, v_s34, 4);
767 uint8x8_t v_s45 = vext_u8(v_s34, v_s56, 4);
768
769 uint8x8_t d01 = convolve4_8(v_s01, v_s12, v_s23, v_s34, y_filter_taps);
770 uint8x8_t d23 = convolve4_8(v_s23, v_s34, v_s45, v_s56, y_filter_taps);
771
772 store_unaligned_u8(dst + 0 * dst_stride, dst_stride, d01);
773 store_unaligned_u8(dst + 2 * dst_stride, dst_stride, d23);
774
775 v_s01 = v_s45;
776 v_s12 = v_s56;
777 src += 4 * src_stride;
778 dst += 4 * dst_stride;
779 h -= 4;
780 } while (h != 0);
781 } else {
782 const uint8x16x2_t permute_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
783
784 do {
785 const uint8_t *s = src;
786 uint8_t *d = dst;
787 int height = h;
788
789 uint8x16_t h_s0, h_s1, h_s2;
790 load_u8_16x3(s, src_stride, &h_s0, &h_s1, &h_s2);
791
792 uint8x8_t v_s0 = convolve4_8_h(h_s0, x_filter, permute_tbl);
793 uint8x8_t v_s1 = convolve4_8_h(h_s1, x_filter, permute_tbl);
794 uint8x8_t v_s2 = convolve4_8_h(h_s2, x_filter, permute_tbl);
795
796 s += 3 * src_stride;
797
798 do {
799 uint8x16_t h_s3, h_s4, h_s5, h_s6;
800 load_u8_16x4(s, src_stride, &h_s3, &h_s4, &h_s5, &h_s6);
801
802 uint8x8_t v_s3 = convolve4_8_h(h_s3, x_filter, permute_tbl);
803 uint8x8_t v_s4 = convolve4_8_h(h_s4, x_filter, permute_tbl);
804 uint8x8_t v_s5 = convolve4_8_h(h_s5, x_filter, permute_tbl);
805 uint8x8_t v_s6 = convolve4_8_h(h_s6, x_filter, permute_tbl);
806
807 uint8x8_t d0 = convolve4_8(v_s0, v_s1, v_s2, v_s3, y_filter_taps);
808 uint8x8_t d1 = convolve4_8(v_s1, v_s2, v_s3, v_s4, y_filter_taps);
809 uint8x8_t d2 = convolve4_8(v_s2, v_s3, v_s4, v_s5, y_filter_taps);
810 uint8x8_t d3 = convolve4_8(v_s3, v_s4, v_s5, v_s6, y_filter_taps);
811
812 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
813
814 v_s0 = v_s4;
815 v_s1 = v_s5;
816 v_s2 = v_s6;
817 s += 4 * src_stride;
818 d += 4 * dst_stride;
819 height -= 4;
820 } while (height != 0);
821 src += 8;
822 dst += 8;
823 w -= 8;
824 } while (w != 0);
825 }
826 }
827
convolve_8tap_2d_horiz_neon_i8mm(const uint8_t * src,ptrdiff_t src_stride,uint8_t * dst,ptrdiff_t dst_stride,int w,int h,const int8x8_t filter)828 static INLINE void convolve_8tap_2d_horiz_neon_i8mm(
829 const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
830 ptrdiff_t dst_stride, int w, int h, const int8x8_t filter) {
831 if (w == 4) {
832 const uint8x16x2_t permute_tbl = vld1q_u8_x2(dot_prod_permute_tbl);
833
834 do {
835 uint8x16_t s0, s1, s2, s3;
836 load_u8_16x4(src, src_stride, &s0, &s1, &s2, &s3);
837
838 int16x4_t d0 = convolve8_4_h(s0, filter, permute_tbl);
839 int16x4_t d1 = convolve8_4_h(s1, filter, permute_tbl);
840 int16x4_t d2 = convolve8_4_h(s2, filter, permute_tbl);
841 int16x4_t d3 = convolve8_4_h(s3, filter, permute_tbl);
842 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
843 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
844
845 store_u8(dst + 0 * dst_stride, dst_stride, d01);
846 store_u8(dst + 2 * dst_stride, dst_stride, d23);
847
848 src += 4 * src_stride;
849 dst += 4 * dst_stride;
850 h -= 4;
851 } while (h > 3);
852
853 // Process final three rows (h % 4 == 3). See vpx_convolve_neon_i8mm()
854 // below for further details on possible values of block height.
855 uint8x16_t s0, s1, s2;
856 load_u8_16x3(src, src_stride, &s0, &s1, &s2);
857
858 int16x4_t d0 = convolve8_4_h(s0, filter, permute_tbl);
859 int16x4_t d1 = convolve8_4_h(s1, filter, permute_tbl);
860 int16x4_t d2 = convolve8_4_h(s2, filter, permute_tbl);
861 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
862 uint8x8_t d23 =
863 vqrshrun_n_s16(vcombine_s16(d2, vdup_n_s16(0)), FILTER_BITS - 1);
864
865 store_u8(dst + 0 * dst_stride, dst_stride, d01);
866 store_u8_4x1(dst + 2 * dst_stride, d23);
867 } else {
868 const uint8x16x3_t permute_tbl = vld1q_u8_x3(dot_prod_permute_tbl);
869
870 do {
871 const uint8_t *s = src;
872 uint8_t *d = dst;
873 int width = w;
874
875 do {
876 uint8x16_t s0, s1, s2, s3;
877 load_u8_16x4(s, src_stride, &s0, &s1, &s2, &s3);
878
879 uint8x8_t d0 = convolve8_8_h(s0, filter, permute_tbl);
880 uint8x8_t d1 = convolve8_8_h(s1, filter, permute_tbl);
881 uint8x8_t d2 = convolve8_8_h(s2, filter, permute_tbl);
882 uint8x8_t d3 = convolve8_8_h(s3, filter, permute_tbl);
883
884 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
885
886 s += 8;
887 d += 8;
888 width -= 8;
889 } while (width > 0);
890 src += 4 * src_stride;
891 dst += 4 * dst_stride;
892 h -= 4;
893 } while (h > 3);
894
895 // Process final three rows (h % 4 == 3). See vpx_convolve_neon_i8mm()
896 // below for further details on possible values of block height.
897 const uint8_t *s = src;
898 uint8_t *d = dst;
899 int width = w;
900
901 do {
902 uint8x16_t s0, s1, s2;
903 load_u8_16x3(s, src_stride, &s0, &s1, &s2);
904
905 uint8x8_t d0 = convolve8_8_h(s0, filter, permute_tbl);
906 uint8x8_t d1 = convolve8_8_h(s1, filter, permute_tbl);
907 uint8x8_t d2 = convolve8_8_h(s2, filter, permute_tbl);
908
909 store_u8_8x3(d, dst_stride, d0, d1, d2);
910
911 s += 8;
912 d += 8;
913 width -= 8;
914 } while (width > 0);
915 }
916 }
917
vpx_convolve8_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)918 void vpx_convolve8_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
919 uint8_t *dst, ptrdiff_t dst_stride,
920 const InterpKernel *filter, int x0_q4,
921 int x_step_q4, int y0_q4, int y_step_q4, int w,
922 int h) {
923 assert(x_step_q4 == 16);
924 assert(y_step_q4 == 16);
925
926 (void)x_step_q4;
927 (void)y_step_q4;
928
929 const int x_filter_taps = vpx_get_filter_taps(filter[x0_q4]) <= 4 ? 4 : 8;
930 const int y_filter_taps = vpx_get_filter_taps(filter[y0_q4]) <= 4 ? 4 : 8;
931 // Account for needing filter_taps / 2 - 1 lines prior and filter_taps / 2
932 // lines post both horizontally and vertically.
933 const ptrdiff_t horiz_offset = x_filter_taps / 2 - 1;
934 const ptrdiff_t vert_offset = (y_filter_taps / 2 - 1) * src_stride;
935
936 if (x_filter_taps == 4 && y_filter_taps == 4) {
937 const int16x4_t x_filter = vld1_s16(filter[x0_q4] + 2);
938 const int16x8_t y_filter = vld1q_s16(filter[y0_q4]);
939
940 // 4-tap and bilinear filter values are even, so halve them to reduce
941 // intermediate precision requirements.
942 const int8x8_t x_filter_4tap =
943 vshrn_n_s16(vcombine_s16(x_filter, vdup_n_s16(0)), 1);
944 const uint8x8_t y_filter_4tap =
945 vshrn_n_u16(vreinterpretq_u16_s16(vabsq_s16(y_filter)), 1);
946
947 convolve_4tap_2d_neon_i8mm(src - horiz_offset - vert_offset, src_stride,
948 dst, dst_stride, w, h, x_filter_4tap,
949 y_filter_4tap);
950 return;
951 }
952
953 // Given our constraints: w <= 64, h <= 64, taps <= 8 we can reduce the
954 // maximum buffer size to 64 * (64 + 7).
955 DECLARE_ALIGNED(32, uint8_t, im_block[64 * 71]);
956 const int im_stride = 64;
957 const int im_height = h + SUBPEL_TAPS - 1;
958
959 const int8x8_t x_filter_8tap = vmovn_s16(vld1q_s16(filter[x0_q4]));
960 const int8x8_t y_filter_8tap = vmovn_s16(vld1q_s16(filter[y0_q4]));
961
962 convolve_8tap_2d_horiz_neon_i8mm(src - horiz_offset - vert_offset, src_stride,
963 im_block, im_stride, w, im_height,
964 x_filter_8tap);
965
966 convolve_8tap_vert_neon_i8mm(im_block, im_stride, dst, dst_stride, w, h,
967 y_filter_8tap);
968 }
969
vpx_convolve8_avg_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)970 void vpx_convolve8_avg_neon_i8mm(const uint8_t *src, ptrdiff_t src_stride,
971 uint8_t *dst, ptrdiff_t dst_stride,
972 const InterpKernel *filter, int x0_q4,
973 int x_step_q4, int y0_q4, int y_step_q4, int w,
974 int h) {
975 DECLARE_ALIGNED(32, uint8_t, im_block[64 * 71]);
976 const int im_stride = 64;
977
978 // Averaging convolution always uses an 8-tap filter.
979 // Account for the vertical phase needing 3 lines prior and 4 lines post.
980 const int im_height = h + SUBPEL_TAPS - 1;
981 const ptrdiff_t offset = SUBPEL_TAPS / 2 - 1;
982
983 assert(y_step_q4 == 16);
984 assert(x_step_q4 == 16);
985
986 const int8x8_t x_filter_8tap = vmovn_s16(vld1q_s16(filter[x0_q4]));
987
988 convolve_8tap_2d_horiz_neon_i8mm(src - offset - offset * src_stride,
989 src_stride, im_block, im_stride, w,
990 im_height, x_filter_8tap);
991
992 vpx_convolve8_avg_vert_neon_i8mm(im_block + offset * im_stride, im_stride,
993 dst, dst_stride, filter, x0_q4, x_step_q4,
994 y0_q4, y_step_q4, w, h);
995 }
996