1 /*
2 * Copyright (c) 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/GpuReshape.h"
25 #include "arm_compute/core/Error.h"
26 #include "src/common/utils/Log.h"
27 #include "src/core/helpers/AutoConfiguration.h"
28 #include "src/dynamic_fusion/sketch/ArgumentPack.h"
29 #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
30 #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentReshape.h"
31
32 namespace arm_compute
33 {
34 namespace experimental
35 {
36 namespace dynamic_fusion
37 {
38 namespace
39 {
is_supported_op_helper(const GpuWorkloadContext & context,const ITensorInfo * src,const ITensorInfo * dst,const ReshapeAttributes & attributes)40 Status is_supported_op_helper(const GpuWorkloadContext &context,
41 const ITensorInfo *src,
42 const ITensorInfo *dst,
43 const ReshapeAttributes &attributes)
44 {
45 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
46
47 TensorInfo dst_info_to_validate;
48 const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate;
49
50 if(dst != nullptr)
51 {
52 dst_info_to_validate_ptr = dst;
53 }
54
55 auto_init_if_empty(dst_info_to_validate, src->clone()->set_tensor_shape(attributes.shape()));
56
57 // Check components
58 if(context.gpu_language() == GpuLanguage::OpenCL)
59 {
60 const auto cl_compile_ctx = context.cl_compile_context();
61 ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr);
62
63 // Validate GpuReshape Component
64 ArgumentPack<ITensorInfo> arguments;
65 arguments.add_const_tensor(ACL_SRC_0, src);
66 arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr);
67
68 ARM_COMPUTE_RETURN_ON_ERROR(ClComponentReshape::validate(arguments));
69 }
70 else
71 {
72 ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
73 }
74
75 return Status{};
76 }
77
78 GpuOperatorType operator_type = GpuOperatorType::Complex;
79 } // namespace
80
is_supported_op(const GpuWorkloadContext & context,const ITensorInfo * src,const Attributes & attributes)81 Status GpuReshape::is_supported_op(const GpuWorkloadContext &context,
82 const ITensorInfo *src,
83 const Attributes &attributes)
84 {
85 return is_supported_op_helper(context, src, nullptr, attributes);
86 }
87
validate_op(const GpuWorkloadSketch & sketch,const ITensorInfo * src,const Attributes & attributes)88 Status GpuReshape::validate_op(const GpuWorkloadSketch &sketch,
89 const ITensorInfo *src,
90 const Attributes &attributes)
91 {
92 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
93 ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id());
94
95 // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object
96 TensorInfo dst_info_to_validate;
97
98 // Auto initialize dst tensor info
99 auto_init_if_empty(dst_info_to_validate, src->clone()->set_tensor_shape(attributes.shape()));
100
101 // Perform fusion test
102 // Pack tensor infos
103 ArgumentPack<ITensorInfo> tensors;
104 tensors.add_const_tensor(ACL_SRC_0, src);
105 tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate);
106 const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
107 ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
108 "Operator fusion test failed. This operator cannot be fused into the workload");
109
110 // Check if configuration is supported
111 return is_supported_op_helper(*sketch.gpu_context(), src, &dst_info_to_validate, attributes);
112 }
113
create_op(GpuWorkloadSketch & sketch,ITensorInfo * src,const Attributes & attributes)114 ITensorInfo *GpuReshape::create_op(GpuWorkloadSketch &sketch,
115 ITensorInfo *src,
116 const Attributes &attributes)
117 {
118 ARM_COMPUTE_ERROR_ON_NULLPTR(src);
119 ARM_COMPUTE_LOG_PARAMS(src, attributes.shape());
120 ARM_COMPUTE_ERROR_THROW_ON(GpuReshape::validate_op(sketch, src, attributes));
121
122 ITensorInfo *dst = sketch.implementation().create_virtual_tensor();
123 ARM_COMPUTE_ERROR_ON_NULLPTR(dst);
124
125 auto_init_if_empty(*dst, src->clone()->set_tensor_shape(attributes.shape()));
126
127 // Translate into components and add to component graph
128 auto &comp_graph = sketch.implementation().component_graph();
129 const auto sketch_ctx = sketch.implementation().context();
130 if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
131 {
132 const auto cl_compile_ctx = sketch_ctx->cl_compile_context();
133 ARM_COMPUTE_ERROR_ON(cl_compile_ctx == nullptr);
134
135 // Add ElementwiseBinary Component
136 {
137 auto properties = IGpuKernelComponent::Properties();
138 properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
139
140 ArgumentPack<ITensorInfo> arguments;
141 arguments.add_const_tensor(ACL_SRC_0, src);
142 arguments.add_const_tensor(ACL_DST_0, dst);
143 comp_graph.add_new_component<ClComponentReshape>(properties, arguments);
144 }
145 }
146 else
147 {
148 ARM_COMPUTE_ERROR("Unimplemented Gpu language");
149 }
150 // Set up fusion test by adding to the Operator Group
151 // Note this has to be performed after all the components have been successfully added to the component graph
152
153 // Pack tensor infos
154 ArgumentPack<ITensorInfo> tensors;
155 tensors.add_const_tensor(ACL_SRC_0, src);
156 tensors.add_tensor(ACL_DST_0, dst);
157
158 const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
159 sketch.implementation().operator_group().add_operator(op);
160
161 return dst;
162 }
163 } // namespace dynamic_fusion
164 } // namespace experimental
165 } // namespace arm_compute
166