Compute Library
 21.11
GEMMLowp.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2021 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #include "arm_compute/core/Types.h"
31 #include "tests/NEON/Accessor.h"
32 #include "tests/NEON/Helper.h"
34 #include "tests/datasets/GEMMLowpFusedOffsetOutputDataset.h"
35 #include "tests/datasets/LargeGEMMLowpDataset.h"
36 #include "tests/datasets/ShapeDatasets.h"
37 #include "tests/datasets/SmallGEMMLowpDataset.h"
39 #include "tests/framework/Macros.h"
42 #include "tests/validation/fixtures/GEMMLowpFixture.h"
43 
44 namespace arm_compute
45 {
46 namespace test
47 {
48 namespace validation
49 {
50 TEST_SUITE(NEON)
51 TEST_SUITE(GEMMLowp)
52 TEST_SUITE(MatrixMultiplyCore)
53 using NEGEMMLowpMatrixMultiplyCoreFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
54 
55 DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::concat(datasets::SmallGEMMLowpDataset(), datasets::LargeGEMMLowpDataset()),
56  shape_a, shape_b, shape_c, a_offset, b_offset)
57 {
58  // Create tensors
59  Tensor a = create_tensor<Tensor>(shape_a, DataType::QASYMM8);
60  Tensor b = create_tensor<Tensor>(shape_b, DataType::QASYMM8);
61  Tensor c = create_tensor<Tensor>(shape_c, DataType::S32);
62 
63  a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset));
64  b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset));
65 
67  ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS);
68  ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS);
69 
70  // Create and configure function
71  NEGEMMLowpMatrixMultiplyCore gemmlowp_mm;
72  gemmlowp_mm.configure(&a, &b, nullptr, &c);
73 
74  // Validate padding is zero
75  validate(a.info()->padding(), PaddingSize());
76  validate(b.info()->padding(), PaddingSize());
77  validate(c.info()->padding(), PaddingSize());
78 }
79 
80 // *INDENT-OFF*
81 // clang-format off
83  framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Input not a multiple of 4
84  TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Mismatching data type
85  TensorInfo(TensorShape(20U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Invalid dimensions
86  TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Invalid dimensions
88  }),
94  })),
100  })),
101  framework::dataset::make("Expected", { true, false, false, false, true })),
102  a_info, b_info, output_info, expected)
103 {
104  // Lock tensors
105  Status status = NEGEMMLowpMatrixMultiplyCore::validate(&a_info.clone()->set_is_resizable(false),
106  &b_info.clone()->set_is_resizable(false),
107  nullptr,
108  &output_info.clone()->set_is_resizable(false));
110 }
111 // clang-format on
112 // *INDENT-ON*
113 
114 /** Test case for memory injection in @ref cpu::CpuGemmLowpMatrixMultiplyCore.
115  *
116  * Configure the operator once and inject memory at run-time in multiple executions.
117  *
118  * Checks performed in order:
119  * - Both runs compute the same output
120  */
121 TEST_CASE(MemoryInjection, framework::DatasetMode::ALL)
122 {
123  auto gemm = std::make_unique<cpu::CpuGemmLowpMatrixMultiplyCore>();
124  auto a_info = TensorInfo(TensorShape(32U, 72U), 1, DataType::QASYMM8);
125  auto b_info = TensorInfo(TensorShape(17U, 32U), 1, DataType::QASYMM8);
126  auto dst_info = TensorInfo(TensorShape(17U, 72U), 1, DataType::S32);
127  a_info.set_quantization_info(QuantizationInfo(1.0f / 255, -9));
128  b_info.set_quantization_info(QuantizationInfo(1.0f / 255, 1));
129  const auto gemm_info = GEMMInfo{};
130  gemm->configure(&a_info, &b_info, nullptr, &dst_info, gemm_info);
131 
132  // telhs are newly created every call of this lambda function
133  auto a = create_tensor<Tensor>(a_info);
134  auto b = create_tensor<Tensor>(b_info);
135  auto dst = create_tensor<Tensor>(dst_info);
136  a.allocator()->allocate();
137  b.allocator()->allocate();
138  dst.allocator()->allocate();
139 
140  ITensorPack run_pack =
141  {
142  { TensorType::ACL_SRC_0, &a },
143  { TensorType::ACL_SRC_1, &b },
144  { TensorType::ACL_DST, &dst }
145  };
146  ITensorPack prep_pack =
147  {
148  { TensorType::ACL_SRC_1, &b },
149  };
150 
151  auto mg = MemoryGroup{};
152  auto ws = manage_workspace<Tensor>(gemm->workspace(), mg, run_pack, prep_pack);
153 
154  auto run_conv = [&]() -> Tensor
155  {
156  auto dst = create_tensor<Tensor>(dst_info);
157  dst.allocator()->allocate();
158  run_pack.add_tensor(TensorType::ACL_DST, &dst);
159 
160  library->fill_tensor_value(Accessor(a), static_cast<uint8_t>(1));
161  library->fill_tensor_value(Accessor(b), static_cast<uint8_t>(2));
162  // This operator is configured once and captured by this lambda.
163  gemm->prepare(prep_pack);
164  gemm->run(run_pack);
165  return dst;
166  };
167  auto result_0 = run_conv();
168  auto result_1 = run_conv();
169  for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i)
170  {
171  ARM_COMPUTE_EXPECT(((uint8_t *)result_0.buffer())[i] == ((uint8_t *)result_1.buffer())[i], framework::LogLevel::ERRORS);
172  }
173 }
174 
175 /** Test case for memory injection in @ref NEGEMMLowpMatrixMultiplyCore.
176  *
177  * Make sure @ref NEGEMMLowpMatrixMultiplyCore still works through injecting the memory at configure time using the old API.
178  *
179  * Checks performed in order:
180  * - Both runs compute the same output
181  */
182 TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL)
183 {
184  auto gemm = std::make_unique<NEGEMMLowpMatrixMultiplyCore>();
185  auto a_info = TensorInfo(TensorShape(32U, 72U), 1, DataType::QASYMM8);
186  auto b_info = TensorInfo(TensorShape(17U, 32U), 1, DataType::QASYMM8);
187  auto dst_info = TensorInfo(TensorShape(17U, 72U), 1, DataType::S32);
188  a_info.set_quantization_info(QuantizationInfo(1.0f / 255, -9));
189  b_info.set_quantization_info(QuantizationInfo(1.0f / 255, 1));
190  const auto gemm_info = GEMMInfo{};
191  auto run_conv = [&]()
192  {
193  auto a = create_tensor<Tensor>(a_info);
194  auto b = create_tensor<Tensor>(b_info);
195  auto dst = create_tensor<Tensor>(dst_info);
196  gemm->configure(&a, &b, nullptr, &dst, gemm_info);
197  a.allocator()->allocate();
198  b.allocator()->allocate();
199  dst.allocator()->allocate();
200  library->fill_tensor_value(Accessor(a), static_cast<uint8_t>(1));
201  library->fill_tensor_value(Accessor(b), static_cast<uint8_t>(2));
202  gemm->run();
203  return dst;
204  };
205  auto result_0 = run_conv();
206  auto result_1 = run_conv();
207  for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i)
208  {
209  ARM_COMPUTE_EXPECT(((uint8_t *)result_0.buffer())[i] == ((uint8_t *)result_1.buffer())[i], framework::LogLevel::ERRORS);
210  }
211 }
212 
214 {
215  // Validate output
216  validate(Accessor(_target), _reference);
217 }
218 
220 {
221  // Validate output
222  validate(Accessor(_target), _reference);
223 }
224 
225 using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>;
226 TEST_SUITE(FusedOffsetOutput)
227 FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(),
228  framework::dataset::make("DataType", { DataType::QASYMM8 })))
229 {
230  // Validate output
231  validate(Accessor(_target), _reference);
232 }
233 
236 {
237  // Validate output
238  validate(Accessor(_target), _reference);
239 }
240 TEST_SUITE_END() // FusedOffsetOutput
241 TEST_SUITE_END() // MatrixMultiplyCore
242 TEST_SUITE_END() // GEMMLowp
243 TEST_SUITE_END() // NEON
244 } // namespace validation
245 } // namespace test
246 } // namespace arm_compute
Shape of a tensor.
Definition: TensorShape.h:39
SimpleTensor< float > b
Definition: DFT.cpp:157
ARM_COMPUTE_EXPECT(has_error==expected, framework::LogLevel::ERRORS)
std::enable_if< is_container< T >::value, ContainerDataset< T > >::type make(std::string name, T &&values)
Helper function to create a ContainerDataset.
Status class.
Definition: Error.h:52
Copyright (c) 2017-2021 Arm Limited.
TensorAllocator * allocator()
Return a pointer to the tensor&#39;s allocator.
Definition: Tensor.cpp:48
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor&#39;s metadata.
Definition: Tensor.cpp:33
1 channel, 1 S32 per channel
virtual bool is_resizable() const =0
Flag indicating whether the size of the tensor can be changed.
Quantization information.
DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16), }), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(30U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16, QuantizationInfo(1.f/32768.f, 0)), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16, QuantizationInfo(1.f/32768.f, 0)), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QSYMM16, QuantizationInfo(1.f/32768.f, 0)), })), framework::dataset::make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQRT), })), framework::dataset::make("Expected", { false, true, true, true, false, false, true, true, false })), input_info, output_info, act_info, expected)
Accessor implementation for Tensor objects.
Definition: Accessor.h:35
DatasetMode
Possible dataset modes.
Definition: DatasetModes.h:40
GEMMLowpMatrixMultiplyCoreValidationFixture< Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore > NEGEMMLowpMatrixMultiplyCoreFixture
Definition: GEMMLowp.cpp:53
std::unique_ptr< AssetsLibrary > library
Definition: main.cpp:76
TEST_SUITE_END() FIXTURE_DATA_TEST_CASE(RunSmall
[CLActivationLayer Test snippet]
quantized, asymmetric fixed-point 8-bit number unsigned
void allocate() override
Allocate size specified by TensorInfo of CPU memory.
Basic implementation of the tensor interface.
Definition: Tensor.h:37
TEST_SUITE(U8_to_S8) FIXTURE_DATA_TEST_CASE(RunSmall
validate(CLAccessor(output_state), expected_output)
virtual PaddingSize padding() const =0
Padding of tensor.
virtual ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info)=0
Set the quantization settings (scale and offset) of the tensor.
UniqueGemmCommon< Top, Tret > gemm(const GemmArgs &args, const OutputStage &os)
BorderSize PaddingSize
Container for 2D padding size.
Definition: Types.h:384
GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture< Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore > NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture
Definition: GEMMLowp.cpp:225
FIXTURE_DATA_TEST_CASE(RunSmall, CLAbsLayerFixture< half >, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)))
Definition: AbsLayer.cpp:50
Tensor packing service.
Definition: ITensorPack.h:39
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:43
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of NEGEMMLowpMatrixMultiply...
GEMM information class.
Definition: Types.h:1974
JoinDataset< T, U > concat(T &&dataset1, U &&dataset2)
Helper function to create a JoinDataset.
Definition: JoinDataset.h:160
TEST_CASE(FusedActivation, framework::DatasetMode::ALL)
Validate fused activation expecting the following behaviours:
zip(zip(framework::dataset::make("Weights", { TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U, 1U), 1, DataType::F32), }), framework::dataset::make("MVBGInfo",{ TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F16), TensorInfo(TensorShape(5U), 1, DataType::F32), })), framework::dataset::make("Expected", { true, false, false}))
DataType
Available data types.
Definition: Types.h:79
Function to run Gemm on quantized types.
combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F32)))
Definition: AbsLayer.cpp:65