1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
|
/*
* Copyright (c) 2014 The WebM project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#ifndef VPX_VPX_DSP_ARM_VPX_CONVOLVE8_NEON_H_
#define VPX_VPX_DSP_ARM_VPX_CONVOLVE8_NEON_H_
#include <arm_neon.h>
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/vpx_filter.h"
#if VPX_ARCH_AARCH64 && defined(__ARM_FEATURE_DOTPROD)
void vpx_convolve8_2d_horiz_neon_dotprod(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);
static INLINE int16x4_t convolve4_4_sdot_partial(const int8x16_t samples,
const int32x4_t correction,
const int8x8_t filters) {
/* Accumulate dot product into 'correction' to account for range clamp. */
int32x4_t sum = vdotq_lane_s32(correction, samples, filters, 0);
/* Further narrowing and packing is performed by the caller. */
return vmovn_s32(sum);
}
static INLINE int16x4_t convolve4_4_sdot(const uint8x16_t samples,
const int8x8_t filters,
const int32x4_t correction,
const uint8x16_t range_limit,
const uint8x16_t permute_tbl) {
/* Clamp sample range to [-128, 127] for 8-bit signed dot product. */
int8x16_t clamped_samples =
vreinterpretq_s8_u8(vsubq_u8(samples, range_limit));
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
int8x16_t permuted_samples = vqtbl1q_s8(clamped_samples, permute_tbl);
/* Accumulate dot product into 'correction' to account for range clamp. */
int32x4_t sum = vdotq_lane_s32(correction, permuted_samples, filters, 0);
/* Further narrowing and packing is performed by the caller. */
return vmovn_s32(sum);
}
static INLINE uint8x8_t convolve4_8_sdot_partial(const int8x16_t samples_lo,
const int8x16_t samples_hi,
const int32x4_t correction,
const int8x8_t filters) {
/* Sample range-clamping and permutation are performed by the caller. */
/* Accumulate dot product into 'correction' to account for range clamp. */
/* First 4 output values. */
int32x4_t sum0 = vdotq_lane_s32(correction, samples_lo, filters, 0);
/* Second 4 output values. */
int32x4_t sum1 = vdotq_lane_s32(correction, samples_hi, filters, 0);
/* Narrow and re-pack. */
int16x8_t sum = vcombine_s16(vmovn_s32(sum0), vmovn_s32(sum1));
/* We halved the filter values so -1 from right shift. */
return vqrshrun_n_s16(sum, FILTER_BITS - 1);
}
static INLINE uint8x8_t convolve4_8_sdot(const uint8x16_t samples,
const int8x8_t filters,
const int32x4_t correction,
const uint8x16_t range_limit,
const uint8x16x2_t permute_tbl) {
int8x16_t clamped_samples, permuted_samples[2];
/* Clamp sample range to [-128, 127] for 8-bit signed dot product. */
clamped_samples = vreinterpretq_s8_u8(vsubq_u8(samples, range_limit));
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_s8(clamped_samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_s8(clamped_samples, permute_tbl.val[1]);
/* Accumulate dot product into 'correction' to account for range clamp. */
/* First 4 output values. */
int32x4_t sum0 = vdotq_lane_s32(correction, permuted_samples[0], filters, 0);
/* Second 4 output values. */
int32x4_t sum1 = vdotq_lane_s32(correction, permuted_samples[1], filters, 0);
/* Narrow and re-pack. */
int16x8_t sum = vcombine_s16(vmovn_s32(sum0), vmovn_s32(sum1));
/* We halved the filter values so -1 from right shift. */
return vqrshrun_n_s16(sum, FILTER_BITS - 1);
}
static INLINE int16x4_t convolve8_4_sdot_partial(const int8x16_t samples_lo,
const int8x16_t samples_hi,
const int32x4_t correction,
const int8x8_t filters) {
/* Sample range-clamping and permutation are performed by the caller. */
int32x4_t sum;
/* Accumulate dot product into 'correction' to account for range clamp. */
sum = vdotq_lane_s32(correction, samples_lo, filters, 0);
sum = vdotq_lane_s32(sum, samples_hi, filters, 1);
/* Further narrowing and packing is performed by the caller. */
return vqmovn_s32(sum);
}
static INLINE int16x4_t convolve8_4_sdot(const uint8x16_t samples,
const int8x8_t filters,
const int32x4_t correction,
const uint8x16_t range_limit,
const uint8x16x2_t permute_tbl) {
int8x16_t clamped_samples, permuted_samples[2];
int32x4_t sum;
/* Clamp sample range to [-128, 127] for 8-bit signed dot product. */
clamped_samples = vreinterpretq_s8_u8(vsubq_u8(samples, range_limit));
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_s8(clamped_samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_s8(clamped_samples, permute_tbl.val[1]);
/* Accumulate dot product into 'correction' to account for range clamp. */
sum = vdotq_lane_s32(correction, permuted_samples[0], filters, 0);
sum = vdotq_lane_s32(sum, permuted_samples[1], filters, 1);
/* Further narrowing and packing is performed by the caller. */
return vqmovn_s32(sum);
}
static INLINE uint8x8_t convolve8_8_sdot_partial(const int8x16_t samples0_lo,
const int8x16_t samples0_hi,
const int8x16_t samples1_lo,
const int8x16_t samples1_hi,
const int32x4_t correction,
const int8x8_t filters) {
/* Sample range-clamping and permutation are performed by the caller. */
int32x4_t sum0, sum1;
int16x8_t sum;
/* Accumulate dot product into 'correction' to account for range clamp. */
/* First 4 output values. */
sum0 = vdotq_lane_s32(correction, samples0_lo, filters, 0);
sum0 = vdotq_lane_s32(sum0, samples0_hi, filters, 1);
/* Second 4 output values. */
sum1 = vdotq_lane_s32(correction, samples1_lo, filters, 0);
sum1 = vdotq_lane_s32(sum1, samples1_hi, filters, 1);
/* Narrow and re-pack. */
sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
static INLINE uint8x8_t convolve8_8_sdot(const uint8x16_t samples,
const int8x8_t filters,
const int32x4_t correction,
const uint8x16_t range_limit,
const uint8x16x3_t permute_tbl) {
int8x16_t clamped_samples, permuted_samples[3];
int32x4_t sum0, sum1;
int16x8_t sum;
/* Clamp sample range to [-128, 127] for 8-bit signed dot product. */
clamped_samples = vreinterpretq_s8_u8(vsubq_u8(samples, range_limit));
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_s8(clamped_samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_s8(clamped_samples, permute_tbl.val[1]);
/* { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } */
permuted_samples[2] = vqtbl1q_s8(clamped_samples, permute_tbl.val[2]);
/* Accumulate dot product into 'correction' to account for range clamp. */
/* First 4 output values. */
sum0 = vdotq_lane_s32(correction, permuted_samples[0], filters, 0);
sum0 = vdotq_lane_s32(sum0, permuted_samples[1], filters, 1);
/* Second 4 output values. */
sum1 = vdotq_lane_s32(correction, permuted_samples[1], filters, 0);
sum1 = vdotq_lane_s32(sum1, permuted_samples[2], filters, 1);
/* Narrow and re-pack. */
sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
#endif // VPX_ARCH_AARCH64 && defined(__ARM_FEATURE_DOTPROD)
#if VPX_ARCH_AARCH64 && defined(__ARM_FEATURE_MATMUL_INT8)
void 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);
static INLINE int16x4_t convolve4_4_usdot_partial(const uint8x16_t samples,
const int8x8_t filters) {
int32x4_t sum = vusdotq_lane_s32(vdupq_n_s32(0), samples, filters, 0);
/* Further narrowing and packing is performed by the caller. */
return vmovn_s32(sum);
}
static INLINE int16x4_t convolve4_4_usdot(const uint8x16_t samples,
const int8x8_t filters,
const uint8x16_t permute_tbl) {
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
uint8x16_t permuted_samples = vqtbl1q_u8(samples, permute_tbl);
int32x4_t sum =
vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples, filters, 0);
/* Further narrowing and packing is performed by the caller. */
return vmovn_s32(sum);
}
static INLINE uint8x8_t convolve4_8_usdot_partial(const uint8x16_t samples_lo,
const uint8x16_t samples_hi,
const int8x8_t filters) {
/* Sample permutation is performed by the caller. */
/* First 4 output values. */
int32x4_t sum0 = vusdotq_lane_s32(vdupq_n_s32(0), samples_lo, filters, 0);
/* Second 4 output values. */
int32x4_t sum1 = vusdotq_lane_s32(vdupq_n_s32(0), samples_hi, filters, 0);
/* Narrow and re-pack. */
int16x8_t sum = vcombine_s16(vmovn_s32(sum0), vmovn_s32(sum1));
/* We halved the filter values so -1 from right shift. */
return vqrshrun_n_s16(sum, FILTER_BITS - 1);
}
static INLINE uint8x8_t convolve4_8_usdot(const uint8x16_t samples,
const int8x8_t filters,
const uint8x16x2_t permute_tbl) {
uint8x16_t permuted_samples[2];
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_u8(samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_u8(samples, permute_tbl.val[1]);
/* First 4 output values. */
int32x4_t sum0 =
vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
/* Second 4 output values. */
int32x4_t sum1 =
vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filters, 0);
/* Narrow and re-pack. */
int16x8_t sum = vcombine_s16(vmovn_s32(sum0), vmovn_s32(sum1));
/* We halved the filter values so -1 from right shift. */
return vqrshrun_n_s16(sum, FILTER_BITS - 1);
}
static INLINE int16x4_t convolve8_4_usdot_partial(const uint8x16_t samples_lo,
const uint8x16_t samples_hi,
const int8x8_t filters) {
/* Sample permutation is performed by the caller. */
int32x4_t sum;
sum = vusdotq_lane_s32(vdupq_n_s32(0), samples_lo, filters, 0);
sum = vusdotq_lane_s32(sum, samples_hi, filters, 1);
/* Further narrowing and packing is performed by the caller. */
return vqmovn_s32(sum);
}
static INLINE int16x4_t convolve8_4_usdot(const uint8x16_t samples,
const int8x8_t filters,
const uint8x16x2_t permute_tbl) {
uint8x16_t permuted_samples[2];
int32x4_t sum;
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_u8(samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_u8(samples, permute_tbl.val[1]);
sum = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
sum = vusdotq_lane_s32(sum, permuted_samples[1], filters, 1);
/* Further narrowing and packing is performed by the caller. */
return vqmovn_s32(sum);
}
static INLINE uint8x8_t convolve8_8_usdot_partial(const uint8x16_t samples0_lo,
const uint8x16_t samples0_hi,
const uint8x16_t samples1_lo,
const uint8x16_t samples1_hi,
const int8x8_t filters) {
/* Sample permutation is performed by the caller. */
int32x4_t sum0, sum1;
int16x8_t sum;
/* First 4 output values. */
sum0 = vusdotq_lane_s32(vdupq_n_s32(0), samples0_lo, filters, 0);
sum0 = vusdotq_lane_s32(sum0, samples0_hi, filters, 1);
/* Second 4 output values. */
sum1 = vusdotq_lane_s32(vdupq_n_s32(0), samples1_lo, filters, 0);
sum1 = vusdotq_lane_s32(sum1, samples1_hi, filters, 1);
/* Narrow and re-pack. */
sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
static INLINE uint8x8_t convolve8_8_usdot(const uint8x16_t samples,
const int8x8_t filters,
const uint8x16x3_t permute_tbl) {
uint8x16_t permuted_samples[3];
int32x4_t sum0, sum1;
int16x8_t sum;
/* Permute samples ready for dot product. */
/* { 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6 } */
permuted_samples[0] = vqtbl1q_u8(samples, permute_tbl.val[0]);
/* { 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10 } */
permuted_samples[1] = vqtbl1q_u8(samples, permute_tbl.val[1]);
/* { 8, 9, 10, 11, 9, 10, 11, 12, 10, 11, 12, 13, 11, 12, 13, 14 } */
permuted_samples[2] = vqtbl1q_u8(samples, permute_tbl.val[2]);
/* First 4 output values. */
sum0 = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[0], filters, 0);
sum0 = vusdotq_lane_s32(sum0, permuted_samples[1], filters, 1);
/* Second 4 output values. */
sum1 = vusdotq_lane_s32(vdupq_n_s32(0), permuted_samples[1], filters, 0);
sum1 = vusdotq_lane_s32(sum1, permuted_samples[2], filters, 1);
/* Narrow and re-pack. */
sum = vcombine_s16(vqmovn_s32(sum0), vqmovn_s32(sum1));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
#endif // VPX_ARCH_AARCH64 && defined(__ARM_FEATURE_MATMUL_INT8)
static INLINE int16x4_t convolve4_4(const int16x4_t s0, const int16x4_t s1,
const int16x4_t s2, const int16x4_t s3,
const int16x4_t filters) {
int16x4_t sum = vmul_lane_s16(s0, filters, 0);
sum = vmla_lane_s16(sum, s1, filters, 1);
sum = vmla_lane_s16(sum, s2, filters, 2);
sum = vmla_lane_s16(sum, s3, filters, 3);
return sum;
}
static INLINE uint8x8_t convolve4_8(const int16x8_t s0, const int16x8_t s1,
const int16x8_t s2, const int16x8_t s3,
const int16x4_t filters) {
int16x8_t sum = vmulq_lane_s16(s0, filters, 0);
sum = vmlaq_lane_s16(sum, s1, filters, 1);
sum = vmlaq_lane_s16(sum, s2, filters, 2);
sum = vmlaq_lane_s16(sum, s3, filters, 3);
/* We halved the filter values so -1 from right shift. */
return vqrshrun_n_s16(sum, FILTER_BITS - 1);
}
static INLINE int16x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
const int16x4_t s2, const int16x4_t s3,
const int16x4_t s4, const int16x4_t s5,
const int16x4_t s6, const int16x4_t s7,
const int16x8_t filters) {
const int16x4_t filters_lo = vget_low_s16(filters);
const int16x4_t filters_hi = vget_high_s16(filters);
int16x4_t sum;
sum = vmul_lane_s16(s0, filters_lo, 0);
sum = vmla_lane_s16(sum, s1, filters_lo, 1);
sum = vmla_lane_s16(sum, s2, filters_lo, 2);
sum = vmla_lane_s16(sum, s5, filters_hi, 1);
sum = vmla_lane_s16(sum, s6, filters_hi, 2);
sum = vmla_lane_s16(sum, s7, filters_hi, 3);
sum = vqadd_s16(sum, vmul_lane_s16(s3, filters_lo, 3));
sum = vqadd_s16(sum, vmul_lane_s16(s4, filters_hi, 0));
return sum;
}
static INLINE uint8x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
const int16x8_t s2, const int16x8_t s3,
const int16x8_t s4, const int16x8_t s5,
const int16x8_t s6, const int16x8_t s7,
const int16x8_t filters) {
const int16x4_t filters_lo = vget_low_s16(filters);
const int16x4_t filters_hi = vget_high_s16(filters);
int16x8_t sum;
sum = vmulq_lane_s16(s0, filters_lo, 0);
sum = vmlaq_lane_s16(sum, s1, filters_lo, 1);
sum = vmlaq_lane_s16(sum, s2, filters_lo, 2);
sum = vmlaq_lane_s16(sum, s5, filters_hi, 1);
sum = vmlaq_lane_s16(sum, s6, filters_hi, 2);
sum = vmlaq_lane_s16(sum, s7, filters_hi, 3);
sum = vqaddq_s16(sum, vmulq_lane_s16(s3, filters_lo, 3));
sum = vqaddq_s16(sum, vmulq_lane_s16(s4, filters_hi, 0));
return vqrshrun_n_s16(sum, FILTER_BITS);
}
static INLINE uint8x8_t scale_filter_8(const uint8x8_t *const s,
const int16x8_t filters) {
int16x8_t ss[8];
ss[0] = vreinterpretq_s16_u16(vmovl_u8(s[0]));
ss[1] = vreinterpretq_s16_u16(vmovl_u8(s[1]));
ss[2] = vreinterpretq_s16_u16(vmovl_u8(s[2]));
ss[3] = vreinterpretq_s16_u16(vmovl_u8(s[3]));
ss[4] = vreinterpretq_s16_u16(vmovl_u8(s[4]));
ss[5] = vreinterpretq_s16_u16(vmovl_u8(s[5]));
ss[6] = vreinterpretq_s16_u16(vmovl_u8(s[6]));
ss[7] = vreinterpretq_s16_u16(vmovl_u8(s[7]));
return convolve8_8(ss[0], ss[1], ss[2], ss[3], ss[4], ss[5], ss[6], ss[7],
filters);
}
#endif // VPX_VPX_DSP_ARM_VPX_CONVOLVE8_NEON_H_
|
