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
|
/*
* Copyright (c) 2022 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/CLKernelLibrary.h"
#include "arm_compute/core/TensorInfo.h"
#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
#include "arm_compute/dynamic_fusion/sketch/OperatorAttributes.h"
#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h"
#include "arm_compute/runtime/CL/CLScheduler.h"
#include "src/gpu/cl/operators/ClAdd.h"
#include "src/gpu/cl/operators/ClConv2d.h"
#include "tests/CL/CLAccessor.h"
#include "tests/framework/Asserts.h"
#include "tests/framework/Macros.h"
#include "tests/validation/CL/UNIT/dynamic_fusion/Utils.h"
#include "tests/validation/Validation.h"
#include "tests/validation/reference/ConvolutionLayer.h"
#include "tests/validation/reference/ElementwiseOperations.h"
#include "tests/validation/reference/Permute.h"
#ifdef ARM_COMPUTE_ASSERTS_ENABLED
#include "tests/SimpleTensorPrinter.h"
#endif /* ARM_COMPUTE_ASSERTS_ENABLED */
using namespace arm_compute::experimental::dynamic_fusion;
using namespace arm_compute::test::validation::utils;
namespace arm_compute
{
namespace test
{
namespace validation
{
TEST_SUITE(CL)
TEST_SUITE(INTEGRATION)
TEST_SUITE(DYNAMIC_FUSION)
TEST_CASE(Conv2d, framework::DatasetMode::ALL)
{
/* Computation:
* out = conv2d1x1(direct_conv)(input, weights, bias)
*/
CLScheduler::get().default_reinit();
const auto data_type = DataType::F32;
const auto data_layout = DataLayout::NHWC;
const auto t_input_shape = TensorShape(384, 12, 12);
const auto t_weight_shape = TensorShape(384, 1, 1, 16);
const auto t_dst_shape = TensorShape(16, 12, 12);
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
GpuWorkloadSketch sketch{ &gpu_ctx };
// Fuse conv2d
Conv2dAttributes conv2d_attr{};
auto input_info = sketch.create_tensor_info(t_input_shape, 1, data_type, data_layout);
auto weight_info = sketch.create_tensor_info(TensorInfo(t_weight_shape, 1, data_type, data_layout));
auto dst_info = sketch.create_tensor_info();
GpuConv2d::create_op(sketch, &input_info, &weight_info, nullptr, &dst_info, conv2d_attr);
// Configure runtime
ClWorkloadRuntime runtime;
runtime.configure(sketch);
// (Important) Allocate auxiliary tensor memory if there are any
// Instead of using ACL allocated memory, the user can choose to import memory into the tensors
for(auto &data : runtime.get_auxiliary_tensors())
{
CLTensor *tensor = data.first;
AuxMemoryInfo aux_mem_req = data.second;
tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
tensor->allocator()->allocate(); // Use ACL allocated memory
// auto buf = cl::Buffer();
// tensor->allocator()->import_memory(buf); // Or, import external memory
}
// Construct user tensors
CLTensor t_input{};
CLTensor t_weight{};
CLTensor t_dst{};
// Initialize user tensors
t_input.allocator()->init(input_info);
t_weight.allocator()->init(weight_info);
t_dst.allocator()->init(dst_info);
// Allocate and fill user tensors
// Instead of using ACL allocator, the user can choose to import memory into the tensors
t_input.allocator()->allocate();
t_weight.allocator()->allocate();
t_dst.allocator()->allocate();
fill<float>(CLAccessor(t_input), 0, library.get());
fill<float>(CLAccessor(t_weight), 1, library.get());
// Run runtime
runtime.run({ &t_input, &t_weight, &t_dst });
// Create reference
SimpleTensor<float> ref_t_input{ t_input_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC };
SimpleTensor<float> ref_t_weight{ t_weight_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC };
SimpleTensor<float> ref_t_bias_placeholder{ t_dst_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC };
// Fill reference
fill<float>(ref_t_input, 0, library.get());
fill<float>(ref_t_weight, 1, library.get());
auto ref_t_input_nchw = reference::permute(ref_t_input, PermutationVector(1U, 2U, 0U));
auto ref_t_weight_nchw = reference::permute(ref_t_weight, PermutationVector(1U, 2U, 0U));
auto ref_t_bias_placeholder_nchw = reference::permute(ref_t_bias_placeholder, PermutationVector(1U, 2U, 0U));
auto t_dst_shape_nchw = t_dst_shape;
permute(t_dst_shape_nchw, PermutationVector(1U, 2U, 0U));
PadStrideInfo legacy_pad_stride(conv2d_attr.stride().x(), conv2d_attr.stride().y(), conv2d_attr.pad().left, conv2d_attr.pad().right, conv2d_attr.pad().top, conv2d_attr.pad().bottom,
DimensionRoundingType{});
auto ref_t_dst_nchw = reference::convolution_layer(ref_t_input_nchw, ref_t_weight_nchw, ref_t_bias_placeholder_nchw, t_dst_shape_nchw, legacy_pad_stride, conv2d_attr.dilation());
const auto ref_t_dst = reference::permute(ref_t_dst_nchw, PermutationVector(2U, 0U, 1U));
RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */
validate(CLAccessor(t_dst), ref_t_dst_nchw, tolerance_f32);
}
TEST_SUITE(Invalid_Fusion_Should_Fail)
TEST_CASE(Multiple_Complex_Ops_0, framework::DatasetMode::ALL)
{
/* Computation:
* out = conv2d(conv2d(l0_input, l0_weight), l1_weight)
*/
CLScheduler::get().default_reinit();
const auto data_type = DataType::F32;
const auto data_layout = DataLayout::NHWC;
const auto t_input_shape = TensorShape(384, 12, 12);
const auto t_weight_shape = TensorShape(384, 1, 1, 16);
const auto t_dst_shape = TensorShape(16, 12, 12);
auto t_input_info = TensorInfo(t_input_shape, 1, data_type, data_layout);
auto t_weight_info = TensorInfo(t_weight_shape, 1, data_type, data_layout);
auto t_dst_info = TensorInfo();
Conv2dAttributes conv2d_attr{};
// Create a new workload sketch
auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx };
GpuWorkloadSketch sketch{ &gpu_ctx };
// Create tensor infos
auto input_info = sketch.create_tensor_info(t_input_shape, 1, data_type, data_layout);
auto weight_info = sketch.create_tensor_info(TensorInfo(t_weight_shape, 1, data_type, data_layout));
auto dst_info = sketch.create_tensor_info();
// Fuse conv2d into the workload
{
// Validate operator
const auto success = GpuConv2d::validate_op(sketch, &input_info, &weight_info, nullptr, &dst_info, conv2d_attr);
ARM_COMPUTE_EXPECT(bool(success), framework::LogLevel::ERRORS);
GpuConv2d::create_op(sketch, &input_info, &weight_info, nullptr, &dst_info, conv2d_attr);
}
// Create tensor infos
auto weight_info_2 = sketch.create_tensor_info(t_weight_info);
auto dst_info_2 = sketch.create_tensor_info();
// Fuse conv2d into the workload
{
// Validate operator, should fail
const auto success = GpuConv2d::validate_op(sketch, &dst_info, &weight_info_2, nullptr, &dst_info_2, conv2d_attr);
ARM_COMPUTE_EXPECT(!bool(success), framework::LogLevel::ERRORS);
}
}
TEST_SUITE_END() // Invalid_Fusion_Should_Fail
TEST_SUITE_END() // DYNAMIC_FUSION
TEST_SUITE_END() // INTEGRATION
TEST_SUITE_END() // CL
} // namespace validation
} // namespace test
} // namespace arm_compute
|
