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
|
// Copyright (c) 2023, ARM Limited.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "generate_dot_product.h"
namespace
{
//---------------------------------------------------------------------------//
// MatMul //
//---------------------------------------------------------------------------//
void generateMatMulA(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
float* a = reinterpret_cast<float*>(data);
const uint32_t T = cfg.shape[0] * cfg.shape[1] * cfg.shape[2];
const uint32_t C = cfg.shape[2];
for (uint32_t t = 0; t < T; ++t)
{
a[t] = generator(t % C); // k = c
}
}
void generateMatMulB(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
float* b = reinterpret_cast<float*>(data);
const uint32_t T = cfg.shape[0] * cfg.shape[1] * cfg.shape[2];
const uint32_t C = cfg.shape[1];
const uint32_t W = cfg.shape[2];
for (uint32_t t = 0; t < T; ++t)
{
b[t] = generator((t / W) % C); // k = c
}
}
bool generateMatMul(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dataType != DType::DType_FP32)
{
WARNING("[Generator][DP][MatMul] Only supports FP32.");
return false;
}
if (cfg.shape.size() != 3)
{
WARNING("[Generator][DP][MatMul] Tensor shape expected 3 dimensions.");
return false;
}
if (cfg.inputPos > 1 || cfg.inputPos < 0)
{
WARNING("[Generator][DP][MatMul] Invalid input tensor slot position to operator.");
return false;
}
(cfg.inputPos == 0) ? generateMatMulA(cfg, generator, data, size) : generateMatMulB(cfg, generator, data, size);
return true;
}
//---------------------------------------------------------------------------//
// Conv2D //
//---------------------------------------------------------------------------//
bool generateConv2DInput(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dotProductInfo.kernel.size() != 2 || cfg.dotProductInfo.kernel[0] <= 0 || cfg.dotProductInfo.kernel[1] <= 0)
{
WARNING("[Generator][DP][Conv2D][Input] Missing or incorrect kernel size information.");
return false;
}
if (cfg.shape.size() != 4)
{
WARNING("[Generator][DP][Conv2D][Input] Tensor shape expected 4 dimensions.");
return false;
}
float* input = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t IH = cfg.shape[1];
const uint32_t IW = cfg.shape[2];
const uint32_t IC = cfg.shape[3];
const uint32_t KH = cfg.dotProductInfo.kernel[0];
const uint32_t KW = cfg.dotProductInfo.kernel[1];
for (int64_t t = 0; t < T; ++t)
{
uint32_t ic = t % IC;
uint32_t ix = (t / IC) % IW;
uint32_t iy = ((t / IC) / IW) % IH;
uint32_t k = ((iy % KH) * KW + (ix % KW)) * IC + ic;
input[t] = generator(k);
}
return true;
}
bool generateConv2DWeight(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.shape.size() != 4)
{
WARNING("[Generator][DP][Conv2D][Weight] Tensor shape expected 4 dimensions.");
return false;
}
float* weight = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t KH = cfg.shape[1];
const uint32_t KW = cfg.shape[2];
const uint32_t IC = cfg.shape[3];
for (int64_t t = 0; t < T; ++t)
{
uint32_t ic = t % IC;
uint32_t kx = (t / IC) % KW;
uint32_t ky = ((t / IC) / KW) % KH;
uint32_t k = (ky + KW * kx) * IC + ic;
weight[t] = generator(k);
}
return true;
}
bool generateConv2DBias(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.shape.size() != 1)
{
WARNING("[Generator][DP][Conv2D][Bias] Tensor shape expected 1 dimension.");
return false;
}
float* bias = reinterpret_cast<float*>(data);
const uint32_t T = cfg.shape[0];
for (uint32_t t = 0; t < T; ++t)
{
bias[t] = generator(2);
}
return true;
}
bool generateConv2D(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dataType != DType::DType_FP32)
{
WARNING("[Generator][DP][Conv2D] Only supports FP32.");
return false;
}
switch (cfg.inputPos)
{
case 0:
return generateConv2DInput(cfg, generator, data, size);
case 1:
return generateConv2DWeight(cfg, generator, data, size);
case 2:
return generateConv2DBias(cfg, generator, data, size);
default:
WARNING("[Generator][DP][Conv2D] Invalid input tensor slot position to operator.");
return false;
}
}
//---------------------------------------------------------------------------//
// Reduce Sum //
//---------------------------------------------------------------------------//
bool generateReduceSum(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dataType != DType::DType_FP32)
{
WARNING("[Generator][DP][ReduceSum] Only supports FP32.");
return false;
}
if (cfg.inputPos != 0)
{
WARNING("[Generator][DP][ReduceSum] Invalid input tensor slot position to operator.");
return false;
}
if (cfg.dotProductInfo.axis < 0 || static_cast<size_t>(cfg.dotProductInfo.axis) >= cfg.shape.size())
{
WARNING("[Generator][DP][ReduceSum] Invalid axis %d.", cfg.dotProductInfo.axis);
return false;
}
float* input = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t axis = cfg.dotProductInfo.axis;
for (int64_t t = 0; t < T; ++t)
{
uint64_t k = t;
for (uint32_t d = cfg.shape.size() - 1; d > axis; --d)
{
k = k / cfg.shape[d];
}
k = k % cfg.shape[axis];
input[t] = generator(static_cast<int32_t>(k));
}
return true;
}
//---------------------------------------------------------------------------//
// Fully Connected //
//---------------------------------------------------------------------------//
bool generateFullyConnectedInput(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.shape.size() != 2)
{
WARNING("[Generator][DP][FullyConnected][Input] Tensor shape expected 2 dimensions.");
return false;
}
float* input = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t IC = cfg.shape[1];
for (int64_t t = 0; t < T; ++t)
{
uint32_t k = t % IC;
input[t] = generator(k);
}
return true;
}
bool generateFullyConnectedWeight(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.shape.size() != 2)
{
WARNING("[Generator][DP][FullyConnected][Weight] Tensor shape expected 2 dimensions.");
return false;
}
float* weight = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t IC = cfg.shape[1];
for (int64_t t = 0; t < T; ++t)
{
uint32_t k = t % IC;
weight[t] = generator(k);
}
return true;
}
bool generateFullyConnectedBias(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.shape.size() != 1)
{
WARNING("[Generator][DP][FullyConnected][Bias] Tensor shape expected 1 dimension.");
return false;
}
float* bias = reinterpret_cast<float*>(data);
const uint32_t T = cfg.shape[0];
for (uint32_t t = 0; t < T; ++t)
{
bias[t] = generator(2);
}
return true;
}
bool generateFullyConnected(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dataType != DType::DType_FP32)
{
WARNING("[Generator][DP][FullyConnected] Only supports FP32.");
return false;
}
switch (cfg.inputPos)
{
case 0:
return generateFullyConnectedInput(cfg, generator, data, size);
case 1:
return generateFullyConnectedWeight(cfg, generator, data, size);
case 2:
return generateFullyConnectedBias(cfg, generator, data, size);
default:
WARNING("[Generator][DP][FullyConnected] Invalid input tensor slot position to operator.");
return false;
}
}
//---------------------------------------------------------------------------//
// Avg Pool 2D //
//---------------------------------------------------------------------------//
bool generateAvgPool2D(const TosaReference::GenerateConfig& cfg,
TosaReference::IDotProductGenerator& generator,
void* data,
size_t size)
{
if (cfg.dataType != DType::DType_FP32)
{
WARNING("[Generator][DP][AvgPool2D] Only supports FP32.");
return false;
}
if (cfg.inputPos != 0)
{
WARNING("[Generator][DP][AvgPool2D] Invalid input tensor slot position to operator.");
return false;
}
if (cfg.dotProductInfo.kernel.size() != 2 || cfg.dotProductInfo.kernel[0] <= 0 || cfg.dotProductInfo.kernel[1] <= 0)
{
WARNING("[Generator][DP][AvgPool2D] Missing or incorrect kernel size information.");
return false;
}
if (cfg.shape.size() != 4)
{
WARNING("[Generator][DP][AvgPool2D] Tensor shape expected 4 dimensions.");
return false;
}
float* input = reinterpret_cast<float*>(data);
const int64_t T = TosaReference::numElementsFromShape(cfg.shape);
const uint32_t IH = cfg.shape[1];
const uint32_t IW = cfg.shape[2];
const uint32_t C = cfg.shape[3];
const uint32_t KY = cfg.dotProductInfo.kernel[0];
const uint32_t KX = cfg.dotProductInfo.kernel[1];
for (int64_t t = 0; t < T; ++t)
{
uint32_t c = t % C;
uint32_t ix = (t / C) % IW;
uint32_t iy = ((t / C) / IW) % IH;
uint32_t k = ((iy % KY) * KX + (ix % KX)) * C + c;
input[t] = generator(k);
}
return true;
}
} // namespace
namespace TosaReference
{
bool generateDotProduct(const GenerateConfig& cfg, void* data, size_t size)
{
auto generator = pickDotProductGenerator(cfg);
if (!generator)
{
WARNING("[Generator][DP] Requested generator could not be created!");
return false;
}
if (cfg.dotProductInfo.ks <= 0)
{
WARNING("[Generator][DP] Invalid test set kernel size %d.", cfg.dotProductInfo.ks);
return false;
}
// Select which generator to use
switch (cfg.opType)
{
case tosa::Op_MATMUL:
return generateMatMul(cfg, *generator, data, size);
case tosa::Op_CONV2D:
return generateConv2D(cfg, *generator, data, size);
case tosa::Op_REDUCE_SUM:
return generateReduceSum(cfg, *generator, data, size);
case tosa::Op_FULLY_CONNECTED:
return generateFullyConnected(cfg, *generator, data, size);
case tosa::Op_AVG_POOL2D:
return generateAvgPool2D(cfg, *generator, data, size);
default:
WARNING("[Generator][DP] Unsupported operator.");
return false;
}
return false;
}
} // namespace TosaReference
|
