xref: /aosp_15_r20/external/ComputeLibrary/src/dynamic_fusion/sketch/gpu/operators/GpuConv2d.cpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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 #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h"
25 
26 #include "arm_compute/core/KernelDescriptors.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
29 #include "arm_compute/runtime/CL/CLScheduler.h"
30 
31 #include "src/common/utils/Log.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/dynamic_fusion/sketch/ArgumentPack.h"
34 #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
35 #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentDirectConv2d.h"
36 #include "src/gpu/cl/kernels/gemm/ClGemmHelpers.h"
37 #include "src/runtime/heuristics/direct_conv/ClDirectConvKernelConfig.h"
38 #include "src/runtime/heuristics/direct_conv/IClDirectConvKernelConfig.h"
39 
40 namespace arm_compute
41 {
42 namespace experimental
43 {
44 namespace dynamic_fusion
45 {
46 namespace
47 {
export_to_cl_image_support(const ITensorInfo * tensor,GPUTarget gpu_target,const cl::Device & device,DataLayout data_layout)48 bool export_to_cl_image_support(const ITensorInfo *tensor, GPUTarget gpu_target, const cl::Device &device, DataLayout data_layout)
49 {
50     if(tensor->tensor_shape()[0] % 4 || (data_layout != DataLayout::NHWC))
51     {
52         return false;
53     }
54 
55     // If not floating point
56     if(!is_data_type_float(tensor->data_type()))
57     {
58         return false;
59     }
60 
61     if(gpu_target == GPUTarget::G71 || get_arch_from_target(gpu_target) == GPUTarget::MIDGARD)
62     {
63         return false;
64     }
65 
66     // Check if the cl_khr_image2d_from_buffer extension is supported on the target platform
67     if(!image2d_from_buffer_supported(device))
68     {
69         return false;
70     }
71 
72     // Check cl image pitch alignment
73     if(get_cl_image_pitch_alignment(device) == 0)
74     {
75         return false;
76     }
77 
78     const size_t image_w     = tensor->tensor_shape()[0] / 4;
79     const size_t image_h     = tensor->tensor_shape()[1] * tensor->tensor_shape()[2] * tensor->tensor_shape()[3];
80     const size_t max_image_w = device.getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>();
81     const size_t max_image_h = device.getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>();
82 
83     if(image_w > max_image_w || image_h > max_image_h)
84     {
85         return false;
86     }
87 
88     return true;
89 }
90 
config_direct_convolution_nhwc(const ITensorInfo * src,const ITensorInfo * weights,const PadStrideInfo & conv_info)91 DirectConvComputeKernelInfo config_direct_convolution_nhwc(const ITensorInfo *src, const ITensorInfo *weights, const PadStrideInfo &conv_info)
92 {
93     // Get GPU target
94     GPUTarget gpu_target = CLScheduler::get().target();
95 
96     std::unique_ptr<arm_compute::cl_direct_conv::IClDirectConvKernelConfig> t = arm_compute::cl_direct_conv::ClDirectConvKernelConfigurationFactory::create(gpu_target);
97 
98     return t->configure(src, weights, conv_info);
99 }
100 
calculate_and_init_dst_if_empty(ITensorInfo * dst,const ITensorInfo * src,const ITensorInfo * wei,const Conv2dAttributes & attributes)101 void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *src, const ITensorInfo *wei, const Conv2dAttributes &attributes)
102 {
103     if(dst->total_size() == 0U)
104     {
105         const auto shape = misc::shape_calculator::compute_deep_convolution_shape(src->tensor_shape(), src->data_layout(), wei->tensor_shape(),
106                                                                                   PadStrideInfo(attributes.stride().x(), attributes.stride().y(), attributes.pad().left,
107                                                                                                 attributes.pad().right,
108                                                                                                 attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR)); // use the default DimensionRoundingType
109 
110         auto_init_if_empty(*dst, src->clone()->set_tensor_shape(shape));
111     }
112 }
113 
114 /* A helper method to reduce the duplication in dst tensor initialization
115 *  when calling validate()
116 */
is_supported_op_helper(const GpuWorkloadContext & context,const ITensorInfo * src,const ITensorInfo * wei,const ITensorInfo * bia,const ITensorInfo * dst,const Conv2dAttributes & attributes)117 Status is_supported_op_helper(const GpuWorkloadContext &context,
118                               const ITensorInfo        *src,
119                               const ITensorInfo        *wei,
120                               const ITensorInfo        *bia,
121                               const ITensorInfo        *dst,
122                               const Conv2dAttributes   &attributes)
123 {
124     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei);
125 
126     TensorInfo         dst_info_to_validate;
127     const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate;
128 
129     const DataLayout data_layout = src->data_layout();
130     if(dst != nullptr)
131     {
132         dst_info_to_validate_ptr = dst;
133     }
134 
135     calculate_and_init_dst_if_empty(&dst_info_to_validate, src, wei, attributes);
136 
137     // Check support level
138     // Data type
139     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
140     // Data layout
141     ARM_COMPUTE_RETURN_ERROR_ON_DATA_LAYOUT_NOT_IN(src, DataLayout::NHWC);
142 
143     // Check components
144     const auto gpu_target = context.gpu_target();
145     if(context.gpu_language() == GpuLanguage::OpenCL)
146     {
147         const auto cl_compile_ctx = context.cl_compile_context();
148         ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr);
149         // Validate Direct Conv2d Component
150         {
151             const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
152             auto       settings   = ClComponentDirectConv2d::Settings();
153 
154             settings.export_to_cl_image(
155                 export_to_cl_image_support(src, gpu_target, cl_compile_ctx->get_device(), data_layout));
156 
157             settings.fast_relaxed_math(
158                 (gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST)
159                 && (dst_info_to_validate_ptr->data_type() == DataType::F32 || dst_info_to_validate_ptr->data_type() == DataType::F16));
160 
161             ArgumentPack<ITensorInfo> arguments;
162             arguments.add_const_tensor(ACL_SRC_0, src);
163             arguments.add_const_tensor(ACL_SRC_1, wei);
164             arguments.add_const_tensor(ACL_SRC_2, bia);
165             arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr);
166             ARM_COMPUTE_RETURN_ON_ERROR(ClComponentDirectConv2d::validate(properties, arguments, attributes, settings));
167         }
168     }
169     else
170     {
171         ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
172     }
173     return Status{};
174 }
175 
176 constexpr GpuOperatorType operator_type = GpuOperatorType::Complex;
177 } // namespace
178 
is_supported_op(const GpuWorkloadContext & context,const ITensorInfo * src,const ITensorInfo * wei,const ITensorInfo * bia,const Conv2dAttributes & attributes)179 Status GpuConv2d::is_supported_op(const GpuWorkloadContext &context,
180                                   const ITensorInfo        *src,
181                                   const ITensorInfo        *wei,
182                                   const ITensorInfo        *bia,
183                                   const Conv2dAttributes   &attributes)
184 {
185     return is_supported_op_helper(context, src, wei, bia, nullptr, attributes);
186 }
187 
validate_op(const GpuWorkloadSketch & sketch,const ITensorInfo * src,const ITensorInfo * wei,const ITensorInfo * bia,const Conv2dAttributes & attributes)188 Status GpuConv2d::validate_op(const GpuWorkloadSketch &sketch,
189                               const ITensorInfo       *src,
190                               const ITensorInfo       *wei,
191                               const ITensorInfo       *bia,
192                               const Conv2dAttributes &attributes)
193 {
194     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, wei);
195 
196     // Check if tensors have valid id. I.e. they are created from a sketch
197     ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id() || !wei->has_valid_id());
198     if(bia != nullptr)
199     {
200         ARM_COMPUTE_RETURN_ERROR_ON(!bia->has_valid_id());
201     }
202 
203     // This tensor info will have invalid id but because all the existing tensors in the
204     // sketch have valid ids and the DependencyGraph implementation has no notion of validness
205     // regarding tensor ids, it'll be just another tensor id and will validate
206     // Additionally, a new dst id is added every time in create_op, thus there's no need to validate it
207     TensorInfo dst_info_to_validate;
208 
209     // Auto initialize dst tensor info
210     calculate_and_init_dst_if_empty(&dst_info_to_validate, src, wei, attributes);
211 
212     // Perform fusion test
213     // Check if operator meets fusion constraints
214     ArgumentPack<ITensorInfo> tensors;
215     tensors.add_const_tensor(ACL_SRC_0, src);
216     tensors.add_const_tensor(ACL_SRC_1, wei);
217     tensors.add_const_tensor(ACL_SRC_2, bia);
218     tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate);
219     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
220     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
221                                     "Operator fusion test failed. This operator cannot be fused into the workload");
222 
223     // Check if configuration is supported
224     return is_supported_op_helper(*sketch.gpu_context(), src, wei, bia, &dst_info_to_validate, attributes);
225 }
226 
create_op(GpuWorkloadSketch & sketch,ITensorInfo * src,ITensorInfo * wei,ITensorInfo * bia,const Conv2dAttributes & attributes)227 ITensorInfo *GpuConv2d::create_op(GpuWorkloadSketch      &sketch,
228                                   ITensorInfo            *src,
229                                   ITensorInfo            *wei,
230                                   ITensorInfo            *bia,
231                                   const Conv2dAttributes &attributes)
232 {
233     ARM_COMPUTE_LOG_PARAMS(src, wei, bia, attributes);
234     PadStrideInfo conv_info(attributes.stride().x(), attributes.stride().y(), attributes.pad().left,
235                             attributes.pad().right,
236                             attributes.pad().top, attributes.pad().bottom, DimensionRoundingType::FLOOR);
237     // Initialize the direct convolution descriptor
238     const DirectConvComputeKernelInfo desc = config_direct_convolution_nhwc(src, wei, conv_info);
239 
240     ITensorInfo *dst = sketch.implementation().create_virtual_tensor();
241 
242     // Assert validation
243     ARM_COMPUTE_ERROR_THROW_ON(GpuConv2d::validate_op(sketch, src, wei, bia, attributes));
244     ARM_COMPUTE_ERROR_ON_NULLPTR(src, wei, dst);
245 
246     // Auto initialize dst tensor
247     calculate_and_init_dst_if_empty(dst, src, wei, attributes);
248 
249     // Translate into components and add to component graph
250     auto &comp_graph = sketch.implementation().component_graph();
251 
252     const auto sketch_ctx = sketch.implementation().context();
253 
254     const auto data_layout = src->data_layout();
255     const auto gpu_target  = sketch_ctx->gpu_target();
256 
257     if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
258     {
259         const auto cl_compile_ctx = sketch_ctx->cl_compile_context();
260         ARM_COMPUTE_ERROR_ON(cl_compile_ctx == nullptr);
261 
262         // Add Direct Conv2d Component
263         {
264             auto properties = IGpuKernelComponent::Properties();
265             properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
266 
267             auto settings = ClComponentDirectConv2d::Settings();
268 
269             settings.export_to_cl_image(
270                 export_to_cl_image_support(src, gpu_target, cl_compile_ctx->get_device(), data_layout));
271 
272             settings.fast_relaxed_math(
273                 (gpu_target != GPUTarget::G71 && (gpu_target & GPUTarget::GPU_ARCH_MASK) == GPUTarget::BIFROST)
274                 && (dst->data_type() == DataType::F32 || dst->data_type() == DataType::F16));
275 
276             if(settings.export_to_cl_image())
277             {
278                 arm_compute::opencl::kernels::gemm::update_padding_for_cl_image(wei);
279             }
280 
281             settings.direct_conv_descriptor(desc);
282 
283             ArgumentPack<ITensorInfo> arguments;
284             arguments.add_const_tensor(ACL_SRC_0, src);
285             arguments.add_const_tensor(ACL_SRC_1, wei);
286             arguments.add_const_tensor(ACL_SRC_2, bia);
287             arguments.add_const_tensor(ACL_DST_0, dst);
288             comp_graph.add_new_component<ClComponentDirectConv2d>(properties, arguments, attributes, settings);
289         }
290     }
291     else
292     {
293         ARM_COMPUTE_ERROR("Unimplemented Gpu language");
294     }
295 
296     // Set up fusion test by adding to the Operator Group
297     // Note this has to be performed after all the components have been successfully added to the component graph
298 
299     // Pack tensor infos
300     ArgumentPack<ITensorInfo> tensors;
301     tensors.add_const_tensor(ACL_SRC_0, src);
302     tensors.add_const_tensor(ACL_SRC_1, wei);
303     tensors.add_const_tensor(ACL_SRC_2, bia);
304     tensors.add_const_tensor(ACL_DST_0, dst);
305 
306     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
307     sketch.implementation().operator_group().add_operator(op);
308 
309     return dst;
310 }
311 
312 } // namespace dynamic_fusion
313 } // namespace experimental
314 } // namespace arm_compute
315