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 "src/dynamic_fusion/sketch/gpu/operators/internal/GpuElementwiseBinaryCommon.h"
25 #include "src/common/utils/Log.h"
26 #include "src/core/helpers/AutoConfiguration.h"
27 #include "src/dynamic_fusion/sketch/ArgumentPack.h"
28 #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
29 #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentElementwiseBinary.h"
30 
31 namespace arm_compute
32 {
33 namespace experimental
34 {
35 namespace dynamic_fusion
36 {
37 namespace
38 {
calculate_and_init_dst_if_empty(ITensorInfo * dst,const ITensorInfo * lhs,const ITensorInfo * rhs)39 void calculate_and_init_dst_if_empty(ITensorInfo *dst, const ITensorInfo *lhs, const ITensorInfo *rhs)
40 {
41     if(dst->total_size() == 0U)
42     {
43         const std::pair<TensorShape, ValidRegion> broadcast_pair = ITensorInfo::broadcast_shape_and_valid_region(*lhs, *rhs);
44         auto_init_if_empty(*dst, lhs->clone()->set_tensor_shape(broadcast_pair.first));
45     }
46 }
47 
is_supported_op_helper(const GpuWorkloadContext & context,const ITensorInfo * lhs,const ITensorInfo * rhs,const ITensorInfo * dst,const ElementwiseBinaryCommonAttributes & attributes)48 Status is_supported_op_helper(const GpuWorkloadContext                &context,
49                               const ITensorInfo                       *lhs,
50                               const ITensorInfo                       *rhs,
51                               const ITensorInfo                       *dst,
52                               const ElementwiseBinaryCommonAttributes &attributes)
53 {
54     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs);
55 
56     TensorInfo         dst_info_to_validate;
57     const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate;
58 
59     if(dst != nullptr)
60     {
61         dst_info_to_validate_ptr = dst;
62     }
63 
64     calculate_and_init_dst_if_empty(&dst_info_to_validate, lhs, rhs);
65 
66     // Check components
67     if(context.gpu_language() == GpuLanguage::OpenCL)
68     {
69         const auto cl_compile_ctx = context.cl_compile_context();
70         ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr);
71 
72         // Validate ElementwiseBinary Component
73         {
74             ArgumentPack<ITensorInfo> arguments;
75             arguments.add_const_tensor(ACL_SRC_0, lhs);
76             arguments.add_const_tensor(ACL_SRC_1, rhs);
77             arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr);
78 
79             ARM_COMPUTE_RETURN_ON_ERROR(ClComponentElementwiseBinary::validate(arguments, attributes));
80         }
81     }
82     else
83     {
84         ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
85     }
86 
87     return Status{};
88 }
89 
90 GpuOperatorType operator_type = GpuOperatorType::Simple;
91 } // namespace
92 
operation(const ElementwiseBinaryCommonAttributes::ElementwiseOp & operation)93 ElementwiseBinaryCommonAttributes &ElementwiseBinaryCommonAttributes::operation(const ElementwiseBinaryCommonAttributes::ElementwiseOp &operation)
94 {
95     _operation = operation;
96     return *this;
97 }
98 
operation() const99 ElementwiseBinaryCommonAttributes::ElementwiseOp ElementwiseBinaryCommonAttributes::operation() const
100 {
101     return _operation;
102 }
103 
is_supported_op(const GpuWorkloadContext & context,const ITensorInfo * lhs,const ITensorInfo * rhs,const ElementwiseBinaryCommonAttributes & attributes)104 Status GpuElementwiseBinaryCommon::is_supported_op(const GpuWorkloadContext                &context,
105                                                    const ITensorInfo                       *lhs,
106                                                    const ITensorInfo                       *rhs,
107                                                    const ElementwiseBinaryCommonAttributes &attributes)
108 {
109     return is_supported_op_helper(context, lhs, rhs, nullptr, attributes);
110 }
111 
validate_op(const GpuWorkloadSketch & sketch,const ITensorInfo * lhs,const ITensorInfo * rhs,const ElementwiseBinaryCommonAttributes & attributes)112 Status GpuElementwiseBinaryCommon::validate_op(const GpuWorkloadSketch                 &sketch,
113                                                const ITensorInfo                       *lhs,
114                                                const ITensorInfo                       *rhs,
115                                                const ElementwiseBinaryCommonAttributes &attributes)
116 {
117     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(lhs, rhs);
118     ARM_COMPUTE_RETURN_ERROR_ON(!lhs->has_valid_id() || !rhs->has_valid_id());
119 
120     // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object
121     TensorInfo dst_info_to_validate;
122 
123     // Auto initialize dst tensor info
124     calculate_and_init_dst_if_empty(&dst_info_to_validate, lhs, rhs);
125 
126     // Perform fusion test
127     // Pack tensor infos
128     ArgumentPack<ITensorInfo> tensors;
129     tensors.add_const_tensor(ACL_SRC_0, lhs);
130     tensors.add_const_tensor(ACL_SRC_1, rhs);
131     tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate);
132     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
133     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
134                                     "Operator fusion test failed. This operator cannot be fused into the workload");
135 
136     // Check if configuration is supported
137     return is_supported_op_helper(*sketch.gpu_context(), lhs, rhs, &dst_info_to_validate, attributes);
138 }
139 
create_op(GpuWorkloadSketch & sketch,ITensorInfo * lhs,ITensorInfo * rhs,const ElementwiseBinaryCommonAttributes & attributes)140 ITensorInfo *GpuElementwiseBinaryCommon::create_op(GpuWorkloadSketch                       &sketch,
141                                                    ITensorInfo                             *lhs,
142                                                    ITensorInfo                             *rhs,
143                                                    const ElementwiseBinaryCommonAttributes &attributes)
144 {
145     ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, rhs);
146     ARM_COMPUTE_LOG_PARAMS(lhs, rhs);
147     ARM_COMPUTE_ERROR_THROW_ON(GpuElementwiseBinaryCommon::validate_op(sketch, lhs, rhs, attributes));
148 
149     ITensorInfo *dst = sketch.implementation().create_virtual_tensor();
150     ARM_COMPUTE_ERROR_ON_NULLPTR(dst);
151 
152     // Auto initialize dst tensor
153     calculate_and_init_dst_if_empty(dst, lhs, rhs);
154 
155     // Translate into components and add to component graph
156     auto &comp_graph = sketch.implementation().component_graph();
157 
158     const auto sketch_ctx = sketch.implementation().context();
159 
160     if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
161     {
162         ARM_COMPUTE_ERROR_ON_NULLPTR(sketch_ctx->cl_compile_context());
163 
164         // Add ElementwiseBinary Component
165         {
166             auto properties = IGpuKernelComponent::Properties();
167             properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
168 
169             ArgumentPack<ITensorInfo> arguments;
170             arguments.add_const_tensor(ACL_SRC_0, lhs);
171             arguments.add_const_tensor(ACL_SRC_1, rhs);
172             arguments.add_const_tensor(ACL_DST_0, dst);
173             comp_graph.add_new_component<ClComponentElementwiseBinary>(properties, arguments, attributes);
174         }
175     }
176     else
177     {
178         ARM_COMPUTE_ERROR("Unimplemented Gpu language");
179     }
180 
181     // Set up fusion test by adding to the Operator Group
182     // Note this has to be performed after all the components have been successfully added to the component graph
183 
184     // Pack tensor infos
185     ArgumentPack<ITensorInfo> tensors;
186     tensors.add_const_tensor(ACL_SRC_0, lhs);
187     tensors.add_const_tensor(ACL_SRC_1, rhs);
188     tensors.add_tensor(ACL_DST_0, dst);
189     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
190     sketch.implementation().operator_group().add_operator(op);
191 
192     return dst;
193 }
194 
195 } // namespace dynamic_fusion
196 } // namespace experimental
197 } // namespace arm_compute
198