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
|
/*
* Copyright (c) 2019-2020 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 "arm_compute/core/CL/kernels/CLGEMMMatrixMultiplyNativeKernel.h"
#include "arm_compute/core/KernelDescriptors.h"
#include "arm_compute/core/Types.h"
#include "arm_compute/core/utils/misc/ShapeCalculator.h"
#include "arm_compute/runtime/CL/CLTensor.h"
#include "arm_compute/runtime/CL/CLTensorAllocator.h"
#include "tests/CL/CLAccessor.h"
#include "tests/CL/Helper.h"
#include "tests/PaddingCalculator.h"
#include "tests/datasets/ShapeDatasets.h"
#include "tests/framework/Asserts.h"
#include "tests/framework/Macros.h"
#include "tests/framework/datasets/Datasets.h"
#include "tests/validation/Validation.h"
#include "tests/validation/fixtures/GEMMFixture.h"
namespace arm_compute
{
namespace test
{
namespace validation
{
using namespace arm_compute::misc::shape_calculator;
// Create function for CLGEMMMatrixMultiplyNativeKernel
using CLGEMMMatrixMultiplyNative = CLSynthetizeFunction<CLGEMMMatrixMultiplyNativeKernel>;
// Fixture for CLGEMMMatrixMultiplyNative
template <typename T>
using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyNativeValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
// Fixture for CLGEMMMatrixMultiplyNative3D
template <typename T>
using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiplyNative3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>;
namespace
{
// *INDENT-OFF*
// clang-format off
RelativeTolerance<float> rel_tolerance_f32(0.001f);
constexpr float abs_tolerance_f32(0.0001f);
/** Alpha values to test - Precommit */
const auto a_values = framework::dataset::make("alpha", {1.0f, -0.75f} );
/** Beta values to test - Precommit */
const auto beta_values = framework::dataset::make("beta", {-0.75f, 0.0f} );
/** M values to test */
const auto m_values = framework::dataset::make("M", 37);
/** M_W values to test */
const auto m_w_values = framework::dataset::make("M_W", 5);
/** M_H values to test */
const auto m_h_values = framework::dataset::make("M_H", 7);
/** N values to test */
const auto n_values = framework::dataset::make("N", 51);
/** K values to test */
const auto k_values = framework::dataset::make("K", 23);
/** Batch size values to test */
const auto b_values = framework::dataset::make("batch_size", 1, 3);
/** Activation values to test */
const auto act_values = framework::dataset::make("Activation",
{
ActivationLayerInfo(),
ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f),
});
/** M0 values to test - Precommit */
const auto m0_values_precommit = framework::dataset::make("M0", { 4, 6 });
/** N0 values to test - Precommit */
const auto n0_values_precommit = framework::dataset::make("N0", { 4 });
/** K0 values to test - Precommit */
const auto k0_values_precommit = framework::dataset::make("K0", { 4 });
/** H0 values to test - Precommit */
const auto h0_values_precommit = framework::dataset::make("H0", 1, 3);
/** M0 values to test - Nightly */
const auto m0_values_nightly = framework::dataset::make("M0", 1, 8);
/** N0 values to test - Nightly */
const auto n0_values_nightly = framework::dataset::make("N0", { 2, 3, 4, 8 });
/** K0 values to test - Nightly */
const auto k0_values_nightly = framework::dataset::make("K0", { 2, 3, 4, 8 });
/** Broadcast bias from vector to matrix */
const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { false, true } );
/** Boundary handling cases for testing partial/non-partial (full) block dimensions, resulting from different combinations
* of M, M0, N and N0 values.
* M0 and N0 are kept constant, while the different test cases need to vary M and N.
*
* Eg. M = 64 and N = 33 result in a block dimension that has no partial blocks (all full blocks) in Y dimension and
* parital blocks in X dimension.
*/
const auto boundary_handling_cases = combine(combine(combine(combine(combine(combine(combine(combine(combine(
// Large k to force potential out-of-bound reads on input0
framework::dataset::make("K", 315),
// Batch size == 1 to force potential out-of-bound reads on input0
framework::dataset::make("batch_size", 1)),
framework::dataset::make("M0", 4)),
framework::dataset::make("N0", 4)),
framework::dataset::make("K0", 4)),
// Only need to test F32 as F16 shares identical boundary handling logics
framework::dataset::make("DataType", DataType::F32)),
framework::dataset::make("alpha", -0.75f )),
framework::dataset::make("beta", -0.35f )),
broadcast_bias_values),
framework::dataset::make("Activation", ActivationLayerInfo()));
/** Configuration test */
void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, bool broadcast_bias, DataType data_type, const ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
const unsigned int K = k_value;
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0_value;
lhs_info.k0 = k0_value;
GEMMRHSMatrixInfo rhs_info;
rhs_info.n0 = n0_value;
rhs_info.k0 = k0_value;
GEMMKernelInfo kernel_info;
kernel_info.m = M;
kernel_info.n = N;
kernel_info.k = K;
kernel_info.broadcast_bias = broadcast_bias;
kernel_info.activation_info = act_info;
const TensorShape lhs_shape(K, M, b_value);
const TensorShape rhs_shape(N, K, b_value);
const TensorShape bias_shape(N,
broadcast_bias? 1 : M,
broadcast_bias? 1 : b_value);
const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
TensorInfo(rhs_shape, 1, data_type),
kernel_info);
// Create tensors
CLTensor lhs = create_tensor<CLTensor>(lhs_shape, data_type);
CLTensor rhs = create_tensor<CLTensor>(rhs_shape, data_type);
CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type);
CLTensor dst = create_tensor<CLTensor>(dst_shape, data_type);
ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
// Create and configure function
CLGEMMMatrixMultiplyNative gemm;
gemm.configure(&lhs, &rhs, &bias, &dst, 1.0f, 1.0f, lhs_info, rhs_info, kernel_info);
}
/** Zero padding test */
bool validate_zero_padding(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, bool broadcast_bias, DataType data_type, const ActivationLayerInfo &act_info)
{
const unsigned int M = m_value;
const unsigned int N = n_value;
const unsigned int K = k_value;
GEMMLHSMatrixInfo lhs_info;
lhs_info.m0 = m0_value;
lhs_info.k0 = k0_value;
GEMMRHSMatrixInfo rhs_info;
rhs_info.n0 = n0_value;
rhs_info.k0 = k0_value;
GEMMKernelInfo kernel_info;
kernel_info.m = M;
kernel_info.n = N;
kernel_info.k = K;
kernel_info.broadcast_bias = broadcast_bias;
kernel_info.activation_info = act_info;
const TensorShape lhs_shape(K, M, b_value);
const TensorShape rhs_shape(N, K, b_value);
const TensorShape bias_shape(N,
broadcast_bias? 1 : M,
broadcast_bias? 1 : b_value);
const TensorShape dst_shape = compute_mm_shape(TensorInfo(lhs_shape, 1, data_type),
TensorInfo(rhs_shape, 1, data_type),
kernel_info);
// Create tensors
CLTensor lhs = create_tensor<CLTensor>(lhs_shape, data_type);
CLTensor rhs = create_tensor<CLTensor>(rhs_shape, data_type);
CLTensor bias = create_tensor<CLTensor>(bias_shape, data_type);
CLTensor dst = create_tensor<CLTensor>(dst_shape, data_type);
ARM_COMPUTE_EXPECT(lhs.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(rhs.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS);
ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS);
// Create and configure function
CLGEMMMatrixMultiplyNative gemm;
gemm.configure(&lhs, &rhs, &bias, &dst, 1.0f, 1.0f, lhs_info, rhs_info, kernel_info);
// Padding can be added along rhs and bias's X dimension
return dst.info()->padding().empty() && lhs.info()->padding().empty() && bias.info()->padding().bottom == 0 && bias.info()->padding().top == 0;
}
} // namespace
TEST_SUITE(CL)
TEST_SUITE(GEMMMatrixMultiplyNative)
TEST_SUITE(Float)
TEST_SUITE(FP32)
DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
framework::dataset::make("batch_size", 1)),
m0_values_precommit),
n0_values_precommit),
k0_values_precommit),
broadcast_bias_values),
act_values),
m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, act_value)
{
validate_configuration(m_value, n_value, k_value, b_value, m0_value, n0_value, k0_value, broadcast_bias, DataType::F32, act_value);
}
/** Validate zero padding tests
*
* A series of validation tests to check that no padding is added as part of configuration for 4 different scenarios.
*
* Checks performed in order:
* - No partial blocks in both x and y dimensions
* - Partial blocks in x dimension
* - Partial blocks in y dimension
* - Partial blocks in both x and y dimensions
* - No blocks in both x and y dimensions, scalar store (N0==1)
* - Special case: partial_n0 == 5 (vstore1 should be invoked instead of vstore_partial_1)
*/
DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip(zip(zip(
framework::dataset::make("M", { 24, 64, 101, 1, 50, 256, }),
framework::dataset::make("N", { 48, 29, 16, 122, 20, 21, })),
framework::dataset::make("M0", { 4, 8, 7, 2, 1, 8, })),
framework::dataset::make("N0", { 4, 4, 16, 3, 1, 8, })),
m_value, n_value, m0_value, n0_value)
{
bool status = validate_zero_padding(m_value, n_value, 23, 1, m0_value, n0_value, 4, false, DataType::F32, ActivationLayerInfo());
ARM_COMPUTE_EXPECT(status, framework::LogLevel::ERRORS);
}
FIXTURE_DATA_TEST_CASE(RunSmallBoundaryHandlingPartialInXPartialInY, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
combine(combine(
framework::dataset::make("M", 3),
framework::dataset::make("N", 1)),
boundary_handling_cases))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmallBoundaryHandlingPartialInXFullInY, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
combine(combine(
framework::dataset::make("M", 64),
framework::dataset::make("N", 51)),
boundary_handling_cases))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmallBoundaryHandlingFullInXFullInY, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
combine(combine(
framework::dataset::make("M", 64),
framework::dataset::make("N", 32)),
boundary_handling_cases))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmallBoundaryHandlingFullInXPartialInY, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
combine(combine(
framework::dataset::make("M", 37),
framework::dataset::make("N", 32)),
boundary_handling_cases))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL,
combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
b_values),
m0_values_precommit),
n0_values_precommit),
k0_values_precommit),
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
broadcast_bias_values),
act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::DISABLED,
combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_values,
n_values),
k_values),
b_values),
m0_values_nightly),
n0_values_nightly),
k0_values_nightly),
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
broadcast_bias_values),
act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL,
combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
k_values),
b_values),
m0_values_precommit),
n0_values_precommit),
k0_values_precommit),
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::DISABLED,
combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(
m_w_values,
m_h_values),
n_values),
k_values),
b_values),
m0_values_nightly),
n0_values_nightly),
k0_values_nightly),
framework::dataset::make("DataType", DataType::F32)),
a_values),
beta_values),
act_values))
{
// Validate output
validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32);
}
TEST_SUITE_END() // FP32
TEST_SUITE_END() // Float
TEST_SUITE_END() // GEMMMatrixMulipltyNative
TEST_SUITE_END() // CL
} // namespace validation
} // namespace test
} // namespace arm_compute
|
