1 /*
2  * Copyright (c) 2022-2023 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 #ifndef TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE
25 #define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE
26 
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Types.h"
30 
31 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
32 #include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
33 #include "arm_compute/dynamic_fusion/sketch/attributes/Conv2dAttributes.h"
34 #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
35 #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h"
36 #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h"
37 
38 #include "tests/CL/CLAccessor.h"
39 #include "tests/framework/Fixture.h"
40 #include "tests/framework/Macros.h"
41 #include "tests/validation/Validation.h"
42 #include "tests/validation/reference/ConvolutionLayer.h"
43 #include "tests/validation/reference/Permute.h"
44 
45 using namespace arm_compute::experimental::dynamic_fusion;
46 
47 namespace arm_compute
48 {
49 namespace test
50 {
51 namespace validation
52 {
53 namespace
54 {
55 template <typename U>
fill(U && tensor,int i)56 void fill(U &&tensor, int i)
57 {
58     switch(tensor.data_type())
59     {
60         case DataType::F16:
61         {
62             arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f };
63             library->fill(tensor, distribution, i);
64             break;
65         }
66         case DataType::F32:
67         {
68             std::uniform_real_distribution<float> distribution(-1.0f, 1.0f);
69             library->fill(tensor, distribution, i);
70             break;
71         }
72         default:
73             library->fill_tensor_uniform(tensor, i);
74     }
75 }
76 
77 } // namespace
78 
79 /** General Conv2d fixture
80  *  Adapted from tests/validation/fixtures/ConvolutionLayerFixture.h
81  *  TODO: Parameterize to be fully backend agnostic: COMPMID-5760; remove Gpu from name
82  */
83 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
84 class DynamicFusionGpuConv2dValidationGenericFixture : public framework::Fixture
85 {
86 public:
87     using TBias = typename std::conditional < std::is_same<typename std::decay<T>::type, uint8_t>::value
88                   || std::is_same<typename std::decay<T>::type, int8_t>::value,
89                   int32_t, T >::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T
90 
91     template <typename...>
setup(TensorShape input_shape,TensorShape weights_shape,TensorShape bias_shape,TensorShape output_shape,const PadStrideInfo & info,const Size2D & dilation,DataType data_type,DataLayout data_layout,QuantizationInfo quantization_info,QuantizationInfo weight_quantization_info)92     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, const PadStrideInfo &info, const Size2D &dilation, DataType data_type,
93                DataLayout data_layout, QuantizationInfo quantization_info, QuantizationInfo weight_quantization_info)
94     {
95         ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout
96         const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, dilation);
97         _data_type                         = data_type;
98         _data_layout                       = data_layout;
99         _is_quantized                      = is_data_type_quantized_asymmetric(data_type);
100         _quantization_info                 = quantization_info;
101         _weight_quantization_info          = weight_quantization_info;
102         _bias_data_type                    = _is_quantized ? DataType::S32 : data_type;
103         _target                            = compute_target(input_shape, weights_shape, bias_shape, conv2d_attr);
104         _reference                         = compute_reference(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr);
105     }
106 
107 protected:
108     // Given input is in nchw format
compute_target(TensorShape input_shape,TensorShape weights_shape,const TensorShape & bias_shape,Conv2dAttributes conv2d_attr)109     TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, Conv2dAttributes conv2d_attr)
110     {
111         ARM_COMPUTE_ERROR_ON(_data_layout != DataLayout::NHWC);
112         permute(input_shape, PermutationVector(2U, 0U, 1U));
113         permute(weights_shape, PermutationVector(2U, 0U, 1U));
114         CLScheduler::get().default_reinit();
115 
116         // Create a new workload sketch
117         auto              cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
118         auto              gpu_ctx        = GpuWorkloadContext{ &cl_compile_ctx };
119         GpuWorkloadSketch sketch{ &gpu_ctx };
120 
121         // Create sketch tensors
122         TensorInfo input_info  = sketch.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout));
123         TensorInfo weight_info = sketch.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout));
124         TensorInfo bias_info   = sketch.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout));
125         TensorInfo dst_info    = sketch.create_tensor_info();
126 
127         ITensorInfo *ans_info = FunctionType::create_op(sketch, &input_info, &weight_info, &bias_info, conv2d_attr);
128         GpuOutput::create_op(sketch, ans_info, &dst_info);
129 
130         // Configure runtime
131         ClWorkloadRuntime runtime;
132         runtime.configure(sketch);
133         // (Important) Allocate auxiliary tensor memory if there are any
134         for(auto &data : runtime.get_auxiliary_tensors())
135         {
136             CLTensor     *tensor      = std::get<0>(data);
137             TensorInfo    info        = std::get<1>(data);
138             AuxMemoryInfo aux_mem_req = std::get<2>(data);
139             tensor->allocator()->init(info, aux_mem_req.alignment);
140             tensor->allocator()->allocate(); // Use ACL allocated memory
141         }
142         // Construct user tensors
143         TensorType t_input{};
144         TensorType t_weight{};
145         TensorType t_bias{};
146         TensorType t_dst{};
147 
148         // Initialize user tensors
149         t_input.allocator()->init(input_info);
150         t_weight.allocator()->init(weight_info);
151         t_bias.allocator()->init(bias_info);
152         t_dst.allocator()->init(dst_info);
153 
154         // Allocate and fill user tensors
155         t_input.allocator()->allocate();
156         t_weight.allocator()->allocate();
157         t_bias.allocator()->allocate();
158         t_dst.allocator()->allocate();
159 
160         fill(AccessorType(t_input), 0);
161         fill(AccessorType(t_weight), 1);
162         fill(AccessorType(t_bias), 2);
163 
164         // Run runtime
165         runtime.run({ &t_input, &t_weight, &t_bias, &t_dst });
166         return t_dst;
167     }
168 
compute_reference(const TensorShape & input_shape,const TensorShape & weights_shape,const TensorShape & bias_shape,const TensorShape & output_shape,Conv2dAttributes conv2d_attr)169     SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape,
170                                       const TensorShape &output_shape, Conv2dAttributes conv2d_attr)
171     {
172         // Create reference
173         SimpleTensor<T>     src{ input_shape, _data_type, 1, _quantization_info };
174         SimpleTensor<T>     weight{ weights_shape, _data_type, 1, _weight_quantization_info };
175         SimpleTensor<TBias> bias{ bias_shape, _data_type, 1, _quantization_info };
176 
177         fill(src, 0);
178         fill(weight, 1);
179         fill(bias, 2);
180 
181         auto src_nchw          = src;
182         auto weights_nchw      = weight;
183         auto bias_nchw         = bias;
184         auto output_shape_nchw = output_shape;
185 
186         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,
187                                         DimensionRoundingType{});
188         auto dst_nchw = reference::convolution_layer(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, legacy_pad_stride, conv2d_attr.dilation());
189         return dst_nchw;
190     }
191 
192     TensorType       _target{};
193     SimpleTensor<T>  _reference{};
194     DataType         _data_type{};
195     DataType         _bias_data_type{};
196     DataLayout       _data_layout{};
197     QuantizationInfo _quantization_info{};
198     QuantizationInfo _weight_quantization_info{};
199     bool             _is_quantized = false;
200 };
201 
202 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
203 class DynamicFusionGpuConv2dValidationFixture : public DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
204 {
205 public:
206     template <typename...>
setup(TensorShape input_shape,TensorShape weights_shape,TensorShape output_shape,TensorShape bias_shape,const PadStrideInfo & info,const Size2D & dialation,DataType data_type,DataLayout data_layout,QuantizationInfo quantization_info)207     void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape output_shape, TensorShape bias_shape,
208                const PadStrideInfo &info, const Size2D &dialation, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info)
209     {
210         DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, output_shape, bias_shape, info, dialation,
211                                                                                                          data_type, data_layout, quantization_info, quantization_info);
212     }
213 };
214 
215 /** Specific Conv2d method: Direct Conv2d fixture
216  *  Adapted from tests/validation/fixtures/DirectConvolutionLayerFixture.h
217  *  TODO: Parameterize to be fully backend agnostic: COMPMID-5760
218  */
219 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
220 class DynamicFusionDirectConv2dValidationGenericFixture : public framework::Fixture
221 {
222 public:
223     using TBias = typename std::conditional < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T >::type;
224 
225     template <typename...>
setup(TensorShape input_shape,int stride_x,int stride_y,int pad_x,int pad_y,unsigned int kernel_size,unsigned int num_kernels,DataType data_type,QuantizationInfo quantization_info,DataLayout data_layout)226     void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels,
227                DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout)
228     {
229         ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout
230 
231         TensorShape         weights_shape(kernel_size, kernel_size, input_shape.z(), num_kernels);
232         const TensorShape   bias_shape(num_kernels);
233         const PadStrideInfo info(stride_x, stride_y, pad_x, pad_y, DimensionRoundingType::FLOOR);
234         const DataType      bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type;
235 
236         const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, { 1U, 1U } /* dilation */);
237 
238         TensorInfo input_info   = TensorInfo(input_shape, 1, data_type);
239         TensorInfo weights_info = TensorInfo(weights_shape, 1, data_type);
240 
241         const TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_info, weights_info, info);
242 
243         _target    = compute_target(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr, data_type, bias_data_type, quantization_info, data_layout);
244         _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info);
245     }
246 
247 protected:
compute_target(TensorShape input_shape,TensorShape weights_shape,const TensorShape & bias_shape,TensorShape output_shape,const Conv2dAttributes & conv2d_attr,DataType data_type,DataType bias_data_type,QuantizationInfo quantization_info,const DataLayout & data_layout)248     TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const Conv2dAttributes &conv2d_attr,
249                               DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info, const DataLayout &data_layout)
250     {
251         ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC);
252         ARM_COMPUTE_UNUSED(quantization_info);
253         // Dataset shapes are in NCHW layout
254         permute(input_shape, PermutationVector(2U, 0U, 1U));
255         permute(weights_shape, PermutationVector(2U, 0U, 1U));
256         permute(output_shape, PermutationVector(2U, 0U, 1U));
257 
258         auto              cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
259         auto              gpu_ctx        = GpuWorkloadContext{ &cl_compile_ctx };
260         GpuWorkloadSketch sketch{ &gpu_ctx };
261 
262         // Create sketch tensors
263         auto input_info  = sketch.create_tensor_info(TensorInfo(input_shape, 1, data_type, data_layout));
264         auto weight_info = sketch.create_tensor_info(TensorInfo(weights_shape, 1, data_type, data_layout));
265         auto bias_info   = sketch.create_tensor_info(TensorInfo(bias_shape, 1, bias_data_type, data_layout));
266         auto dst_info    = sketch.create_tensor_info();
267 
268         ITensorInfo *ans_info = FunctionType::create_op(sketch, &input_info, &weight_info, &bias_info, conv2d_attr);
269         GpuOutput::create_op(sketch, ans_info, &dst_info);
270 
271         // Configure runtime
272         ClWorkloadRuntime runtime;
273         runtime.configure(sketch);
274 
275         for(auto &data : runtime.get_auxiliary_tensors())
276         {
277             CLTensor     *tensor      = std::get<0>(data);
278             TensorInfo    info        = std::get<1>(data);
279             AuxMemoryInfo aux_mem_req = std::get<2>(data);
280             tensor->allocator()->init(info, aux_mem_req.alignment);
281             tensor->allocator()->allocate(); // Use ACL allocated memory
282         }
283         // Construct user tensors
284         TensorType t_input{};
285         TensorType t_weight{};
286         TensorType t_bias{};
287         TensorType t_dst{};
288 
289         // Initialize user tensors
290         t_input.allocator()->init(input_info);
291         t_weight.allocator()->init(weight_info);
292         t_bias.allocator()->init(bias_info);
293         t_dst.allocator()->init(dst_info);
294 
295         ARM_COMPUTE_ASSERT(t_input.info()->is_resizable());
296         ARM_COMPUTE_ASSERT(t_weight.info()->is_resizable());
297         ARM_COMPUTE_ASSERT(t_bias.info()->is_resizable());
298         ARM_COMPUTE_ASSERT(t_dst.info()->is_resizable());
299 
300         // Allocate and fill user tensors
301         t_input.allocator()->allocate();
302         t_weight.allocator()->allocate();
303         t_bias.allocator()->allocate();
304         t_dst.allocator()->allocate();
305 
306         ARM_COMPUTE_ASSERT(!t_input.info()->is_resizable());
307         ARM_COMPUTE_ASSERT(!t_weight.info()->is_resizable());
308         ARM_COMPUTE_ASSERT(!t_bias.info()->is_resizable());
309         ARM_COMPUTE_ASSERT(!t_dst.info()->is_resizable());
310 
311         fill(AccessorType(t_input), 0);
312         fill(AccessorType(t_weight), 1);
313         fill(AccessorType(t_bias), 2);
314 
315         // Run runtime
316         runtime.run({ &t_input, &t_weight, &t_bias, &t_dst });
317         return t_dst;
318     }
319 
compute_reference(const TensorShape & input_shape,const TensorShape & weights_shape,const TensorShape & bias_shape,const TensorShape & output_shape,const PadStrideInfo & info,DataType data_type,DataType bias_data_type,QuantizationInfo quantization_info)320     SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info,
321                                       DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info)
322     {
323         // Create reference
324         SimpleTensor<T>     src{ input_shape, data_type, 1, quantization_info };
325         SimpleTensor<T>     weights{ weights_shape, data_type, 1, quantization_info };
326         SimpleTensor<TBias> bias{ bias_shape, bias_data_type, 1, quantization_info };
327 
328         // Fill reference
329         fill(src, 0);
330         fill(weights, 1);
331         fill(bias, 2);
332 
333         SimpleTensor<T> dst = reference::convolution_layer<T>(src, weights, bias, output_shape, info);
334         return dst;
335     }
336     TensorType      _target{};
337     SimpleTensor<T> _reference{};
338 };
339 
340 template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
341 class DynamicFusionDirectConv2dValidationFixture : public DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
342 {
343 public:
344     template <typename...>
setup(TensorShape input_shape,int stride_x,int stride_y,int pad_x,int pad_y,unsigned int kernel_size,unsigned int num_kernels,DataType data_type,DataLayout data_layout)345     void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type,
346                DataLayout data_layout)
347     {
348         DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type,
349                                                                                                             QuantizationInfo(),
350                                                                                                             data_layout);
351     }
352 };
353 
354 } // namespace validation
355 } // namespace test
356 } // namespace arm_compute
357 #endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE */
358