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
|
/*
* Copyright (c) 2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include <cstddef>
#include <cstdint>
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FP16_ARGS)
namespace arm_conv {
namespace depthwise {
void sve_fp16_nhwc_3x3_s2_output2x2_mla_depthfirst_direct_impl(
const unsigned int n_tile_rows,
const unsigned int n_tile_cols,
const __fp16 *inptr,
int64_t ld_input_row,
int64_t ld_input_col,
__fp16 *outptr,
int64_t ld_output_row,
int64_t ld_output_col,
const void *params,
unsigned int n_channels,
const __fp16 activation_min,
const __fp16 activation_max
)
{
struct Args
{
const uint64_t n_tile_rows, n_tile_cols;
const __fp16 *inptr;
const uint64_t ld_input_row;
const uint64_t ld_input_col;
__fp16 *outptr;
const uint64_t ld_output_row;
const uint64_t ld_output_col;
const void *params;
const __fp16 min, max;
uint64_t tile_i = 0, tile_j = 0;
Args(
const unsigned int n_tile_rows,
const unsigned int n_tile_cols,
const __fp16 *inptr,
int64_t ld_input_row,
int64_t ld_input_col,
__fp16 *outptr,
int64_t ld_output_row,
int64_t ld_output_col,
const void *params,
const float activation_min,
const float activation_max
) : n_tile_rows(n_tile_rows), n_tile_cols(n_tile_cols), inptr(inptr),
ld_input_row(ld_input_row), ld_input_col(ld_input_col), outptr(outptr),
ld_output_row(ld_output_row), ld_output_col(ld_output_col),
params(params), min(activation_min), max(activation_max)
{
}
};
Args params_struct(
n_tile_rows, n_tile_cols,
inptr, ld_input_row, ld_input_col,
outptr, ld_output_row, ld_output_col,
params, activation_min, activation_max
);
__asm__ __volatile__(
"ptrue p3.b\n"
"mov x7, #0x0\n"
"mov x8, #0x0\n"
"1:" // Tile loop
"str x7, [%x[params_struct], %[offsetof_args_tile_i]]\n"
"mov x23, #0x4\n"
"str x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
"mov x17, #0x2\n"
"ldr x16, [%x[params_struct], %[offsetof_args_params]]\n"
"mov x15, #0x0\n"
"ldr x22, [%x[params_struct], %[offsetof_args_ld_input_row]]\n"
"cnth x14\n"
"ldr x13, [%x[params_struct], %[offsetof_args_ld_input_col]]\n"
"sub x12, XZR, x14\n"
"ldr x21, [%x[params_struct], %[offsetof_args_inptr]]\n"
"mul x19, x7, x22\n" // offset = tile_i * ld_input_row
"ldr x20, [%x[params_struct], %[offsetof_args_ld_output_row]]\n"
"madd x19, x8, x13, x19\n" // offset += tile_j * ld_input_col
"ldr x11, [%x[params_struct], %[offsetof_args_ld_output_col]]\n"
"mul x19, x19, x23\n" // offset *= kernel_stride * output_size
"ldr x10, [%x[params_struct], %[offsetof_args_outptr]]\n"
"add x21, x21, x19, LSL #1\n" // inptr[0] += offset * sizeof(__fp16)
"ld1rh { z19.h }, p3/Z, [%x[params_struct], %[offsetof_args_min]]\n"
"add x9, x21, x22, LSL #1\n"
"ld1rh { z18.h }, p3/Z, [%x[params_struct], %[offsetof_args_max]]\n"
"add x28, x9, x22, LSL #1\n"
"ld1h { z17.h }, p3/Z, [x16]\n" // Load from weights and bias
"mov z31.d, z17.d\n"
"ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" // Load from weights and bias
"add x27, x28, x22, LSL #1\n"
"mov z30.d, z17.d\n"
"ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" // Load from weights and bias
"add x26, x27, x22, LSL #1\n"
"mov z29.d, z17.d\n"
"ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" // Load from weights and bias
"add x25, x13, x13\n"
"mov z28.d, z17.d\n"
"ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" // Load from weights and bias
"add x24, x25, x13\n"
"ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" // Load from weights and bias
"add x23, x24, x13\n"
"ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" // Load from weights and bias
"mul x19, x7, x20\n" // offset = tile_i * ld_output_row
"ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" // Load from weights and bias
"madd x19, x8, x11, x19\n" // offset += tile_j * ld_output_col
"whilelt p2.h, XZR, %x[n_channels]\n"
"ld1h { z9.h }, p2/Z, [x28, x25, LSL #1]\n" // Load input point (2, 2)
"ld1h { z10.h }, p2/Z, [x21]\n" // Load input point (0, 0)
"mul x19, x19, x17\n" // offset *= output_tile_size
"ld1h { z11.h }, p2/Z, [x21, x13, LSL #1]\n" // Load input point (0, 1)
"add x10, x10, x19, LSL #1\n" // outptrs[0] += offset * sizeof(__fp16)
"ld1h { z12.h }, p2/Z, [x21, x24, LSL #1]\n" // Load input point (0, 3)
"add x22, x10, x20, LSL #1\n"
"ld1h { z13.h }, p2/Z, [x21, x23, LSL #1]\n" // Load input point (0, 4)
"addvl x16, x16, #16\n"
"ld1h { z14.h }, p2/Z, [x9]\n" // Load input point (1, 0)
"cmp x14, %x[n_channels]\n"
"ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" // Load from weights and bias
"ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" // Load from weights and bias
"addvl x16, x16, #-6\n"
"ld1h { z15.h }, p2/Z, [x9, x13, LSL #1]\n" // Load input point (1, 1)
"ld1h { z16.h }, p2/Z, [x21, x25, LSL #1]\n" // Load input point (0, 2)
"bge 3f\n"
"2:" // Tile loop: Channel loop
"fmla z31.h, p3/M, z8.h, z9.h\n"
"ld1h { z17.h }, p3/Z, [x16]\n" // Load from weights and bias
"whilelt p1.h, x14, %x[n_channels]\n"
"fmla z30.h, p3/M, z6.h, z9.h\n"
"inch x12\n"
"fmla z29.h, p3/M, z2.h, z9.h\n"
"mov p0.b, p2.b\n"
"fmla z28.h, p3/M, z0.h, z9.h\n"
"inch x15\n"
"addvl x21, x21, #1\n"
"fmla z31.h, p3/M, z0.h, z10.h\n"
"ld1h { z10.h }, p1/Z, [x21]\n" // Load input point (0, 0)
"fmla z30.h, p3/M, z1.h, z12.h\n"
"ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (1, 4)
"inch x14\n"
"fmla z31.h, p3/M, z1.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (1, 3)
"fmla z30.h, p3/M, z2.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (1, 2)
"addvl x9, x9, #1\n"
"fmla z31.h, p3/M, z3.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x27]\n" // Load input point (3, 0)
"fmla z30.h, p3/M, z0.h, z16.h\n"
"fmla z29.h, p3/M, z3.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (3, 4)
"fmla z31.h, p3/M, z4.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x28]\n" // Load input point (2, 0)
"fmla z30.h, p3/M, z4.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" // Load input point (3, 1)
"fmla z29.h, p3/M, z0.h, z15.h\n"
"ld1h { z0.h }, p3/Z, [x16, #1, MUL VL]\n" // Load from weights and bias
"fmla z31.h, p3/M, z2.h, z16.h\n"
"ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" // Load input point (2, 1)
"fmla z30.h, p3/M, z5.h, z12.h\n"
"ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (2, 3)
"fmla z29.h, p3/M, z4.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (2, 4)
"addvl x28, x28, #1\n"
"fmla z31.h, p3/M, z5.h, z13.h\n"
"ld1h { z9.h }, p1/Z, [x28, x25, LSL #1]\n" // Load input point (2, 2)
"fmla z30.h, p3/M, z3.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (3, 3)
"fmla z29.h, p3/M, z1.h, z16.h\n"
"fmla z31.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x26]\n" // Load input point (4, 0)
"fmla z28.h, p3/M, z4.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" // Load input point (4, 1)
"fmla z30.h, p3/M, z7.h, z12.h\n"
"ld1h { z4.h }, p3/Z, [x16, #5, MUL VL]\n" // Load from weights and bias
"fmla z29.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" // Load input point (4, 2)
"fmla z31.h, p3/M, z7.h, z16.h\n"
"ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (3, 2)
"addvl x27, x27, #1\n"
"fmla z28.h, p3/M, z1.h, z12.h\n"
"ld1h { z12.h }, p1/Z, [x21, x24, LSL #1]\n" // Load input point (0, 3)
"fmla z30.h, p3/M, z8.h, z11.h\n"
"ld1h { z1.h }, p3/Z, [x16, #2, MUL VL]\n" // Load from weights and bias
"fmla z29.h, p3/M, z7.h, z13.h\n"
"ld1h { z13.h }, p1/Z, [x21, x23, LSL #1]\n" // Load input point (0, 4)
"fmax z31.h, p3/M, z31.h, z19.h\n"
"fmla z28.h, p3/M, z5.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (4, 3)
"fmax z30.h, p3/M, z30.h, z19.h\n"
"fmla z29.h, p3/M, z5.h, z16.h\n"
"ld1h { z5.h }, p3/Z, [x16, #6, MUL VL]\n" // Load from weights and bias
"fmin z31.h, p3/M, z31.h, z18.h\n"
"st1h { z31.h }, p0, [x10]\n" // Store output point (0, 0)
"mov z31.d, z17.d\n"
"fmla z28.h, p3/M, z2.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (4, 4)
"whilelt p2.h, x15, %x[n_channels]\n"
"fmla z29.h, p3/M, z8.h, z15.h\n"
"ld1h { z2.h }, p3/Z, [x16, #3, MUL VL]\n" // Load from weights and bias
"addvl x26, x26, #1\n"
"fmin z30.h, p3/M, z30.h, z18.h\n"
"st1h { z30.h }, p0, [x10, x11, LSL #1]\n" // Store output point (0, 1)
"mov z30.d, z17.d\n"
"addvl x10, x10, #1\n"
"fmla z28.h, p3/M, z3.h, z16.h\n"
"ld1h { z16.h }, p1/Z, [x21, x25, LSL #1]\n" // Load input point (0, 2)
"cmp x14, %x[n_channels]\n"
"fmax z29.h, p3/M, z29.h, z19.h\n"
"ld1h { z3.h }, p3/Z, [x16, #4, MUL VL]\n" // Load from weights and bias
"fmla z28.h, p3/M, z7.h, z14.h\n"
"ld1h { z14.h }, p1/Z, [x9]\n" // Load input point (1, 0)
"fmin z29.h, p3/M, z29.h, z18.h\n"
"st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0)
"mov z29.d, z17.d\n"
"fmla z28.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p1/Z, [x9, x13, LSL #1]\n" // Load input point (1, 1)
"ld1h { z6.h }, p3/Z, [x16, #7, MUL VL]\n" // Load from weights and bias
"fmla z28.h, p3/M, z8.h, z11.h\n"
"ld1h { z11.h }, p1/Z, [x21, x13, LSL #1]\n" // Load input point (0, 1)
"addvl x16, x16, #16\n"
"fmax z28.h, p3/M, z28.h, z19.h\n"
"ld1h { z7.h }, p3/Z, [x16, #-8, MUL VL]\n" // Load from weights and bias
"ld1h { z8.h }, p3/Z, [x16, #-7, MUL VL]\n" // Load from weights and bias
"fmin z28.h, p3/M, z28.h, z18.h\n"
"st1h { z28.h }, p0, [x22, x11, LSL #1]\n" // Store output point (1, 1)
"mov z28.d, z17.d\n"
"addvl x22, x22, #1\n"
"addvl x16, x16, #-6\n"
"blt 2b\n"
"3:" // Tile loop: Channel tail
"fmla z31.h, p3/M, z8.h, z9.h\n"
"ldr x7, [%x[params_struct], %[offsetof_args_tile_i]]\n"
"mov p0.b, p2.b\n"
"fmla z30.h, p3/M, z6.h, z9.h\n"
"ldr x8, [%x[params_struct], %[offsetof_args_tile_j]]\n"
"add x21, x7, #0x1\n"
"fmla z29.h, p3/M, z2.h, z9.h\n"
"ldr x20, [%x[params_struct], %[offsetof_args_n_tile_rows]]\n"
"fmla z28.h, p3/M, z0.h, z9.h\n"
"ldr x19, [%x[params_struct], %[offsetof_args_n_tile_cols]]\n"
"add x8, x8, #0x1\n"
"fmla z31.h, p3/M, z0.h, z10.h\n"
"cmp x8, x19\n"
"fmla z30.h, p3/M, z1.h, z12.h\n"
"ld1h { z12.h }, p2/Z, [x9, x23, LSL #1]\n" // Load input point (1, 4)
"fmla z31.h, p3/M, z1.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x9, x24, LSL #1]\n" // Load input point (1, 3)
"csel x8, x8, XZR, LT\n"
"fmla z30.h, p3/M, z2.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x9, x25, LSL #1]\n" // Load input point (1, 2)
"csel x7, x7, x21, LT\n"
"fmla z31.h, p3/M, z3.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x27]\n" // Load input point (3, 0)
"cmp x7, x20\n"
"fmla z30.h, p3/M, z0.h, z16.h\n"
"fmla z29.h, p3/M, z3.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x27, x23, LSL #1]\n" // Load input point (3, 4)
"fmla z31.h, p3/M, z4.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x28]\n" // Load input point (2, 0)
"fmla z30.h, p3/M, z4.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x27, x13, LSL #1]\n" // Load input point (3, 1)
"fmla z29.h, p3/M, z0.h, z15.h\n"
"fmla z31.h, p3/M, z2.h, z16.h\n"
"ld1h { z16.h }, p2/Z, [x28, x13, LSL #1]\n" // Load input point (2, 1)
"fmla z30.h, p3/M, z5.h, z12.h\n"
"ld1h { z12.h }, p2/Z, [x28, x24, LSL #1]\n" // Load input point (2, 3)
"fmla z29.h, p3/M, z4.h, z11.h\n"
"ld1h { z11.h }, p2/Z, [x28, x23, LSL #1]\n" // Load input point (2, 4)
"fmla z31.h, p3/M, z5.h, z13.h\n"
"fmla z30.h, p3/M, z3.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x27, x24, LSL #1]\n" // Load input point (3, 3)
"fmla z29.h, p3/M, z1.h, z16.h\n"
"fmla z31.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x26]\n" // Load input point (4, 0)
"fmla z28.h, p3/M, z4.h, z13.h\n"
"ld1h { z13.h }, p2/Z, [x26, x13, LSL #1]\n" // Load input point (4, 1)
"fmla z30.h, p3/M, z7.h, z12.h\n"
"fmla z29.h, p3/M, z6.h, z15.h\n"
"ld1h { z15.h }, p2/Z, [x26, x25, LSL #1]\n" // Load input point (4, 2)
"fmla z31.h, p3/M, z7.h, z16.h\n"
"ld1h { z16.h }, p2/Z, [x27, x25, LSL #1]\n" // Load input point (3, 2)
"fmla z28.h, p3/M, z1.h, z12.h\n"
"fmla z30.h, p3/M, z8.h, z11.h\n"
"fmla z29.h, p3/M, z7.h, z13.h\n"
"fmax z31.h, p3/M, z31.h, z19.h\n"
"fmla z28.h, p3/M, z5.h, z14.h\n"
"ld1h { z14.h }, p2/Z, [x26, x24, LSL #1]\n" // Load input point (4, 3)
"fmax z30.h, p3/M, z30.h, z19.h\n"
"fmla z29.h, p3/M, z5.h, z16.h\n"
"fmin z31.h, p3/M, z31.h, z18.h\n"
"st1h { z31.h }, p0, [x10]\n" // Store output point (0, 0)
"fmla z28.h, p3/M, z2.h, z11.h\n"
"fmla z29.h, p3/M, z8.h, z15.h\n"
"ld1h { z11.h }, p2/Z, [x26, x23, LSL #1]\n" // Load input point (4, 4)
"fmin z30.h, p3/M, z30.h, z18.h\n"
"st1h { z30.h }, p0, [x10, x11, LSL #1]\n" // Store output point (0, 1)
"fmla z28.h, p3/M, z3.h, z16.h\n"
"fmax z29.h, p3/M, z29.h, z19.h\n"
"fmla z28.h, p3/M, z7.h, z14.h\n"
"fmin z29.h, p3/M, z29.h, z18.h\n"
"st1h { z29.h }, p0, [x22]\n" // Store output point (1, 0)
"fmla z28.h, p3/M, z6.h, z15.h\n"
"fmla z28.h, p3/M, z8.h, z11.h\n"
"fmax z28.h, p3/M, z28.h, z19.h\n"
"fmin z28.h, p3/M, z28.h, z18.h\n"
"st1h { z28.h }, p0, [x22, x11, LSL #1]\n" // Store output point (1, 1)
"blt 1b\n"
:
: [n_channels] "r" ((unsigned long) n_channels), [offsetof_args_inptr] "I" (offsetof(Args, inptr)), [offsetof_args_ld_input_col] "I" (offsetof(Args, ld_input_col)), [offsetof_args_ld_input_row] "I" (offsetof(Args, ld_input_row)), [offsetof_args_ld_output_col] "I" (offsetof(Args, ld_output_col)), [offsetof_args_ld_output_row] "I" (offsetof(Args, ld_output_row)), [offsetof_args_max] "I" (offsetof(Args, max)), [offsetof_args_min] "I" (offsetof(Args, min)), [offsetof_args_n_tile_cols] "I" (offsetof(Args, n_tile_cols)), [offsetof_args_n_tile_rows] "I" (offsetof(Args, n_tile_rows)), [offsetof_args_outptr] "I" (offsetof(Args, outptr)), [offsetof_args_params] "I" (offsetof(Args, params)), [offsetof_args_tile_i] "I" (offsetof(Args, tile_i)), [offsetof_args_tile_j] "I" (offsetof(Args, tile_j)), [params_struct] "r" (¶ms_struct)
: "cc", "memory", "p0", "p1", "p2", "p3", "x7", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x17", "x19", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z28", "z29", "z30", "z31"
);
}
} // namespace depthwise
} // namespace arm_conv
#endif // defined(__ARM_FEATURE_SVE) && defined(__ARM_FP16_ARGS)
|
