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/GpuCast.h"
25
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/ClComponentCast.h"
30
31 #include "src/common/utils/Log.h"
32
33 namespace arm_compute
34 {
35 namespace experimental
36 {
37 namespace dynamic_fusion
38 {
39 namespace
40 {
is_supported_op_helper(const GpuWorkloadContext & context,const ITensorInfo * src,const ITensorInfo * dst,const CastAttributes & attributes)41 Status is_supported_op_helper(const GpuWorkloadContext &context,
42 const ITensorInfo *src,
43 const ITensorInfo *dst,
44 const CastAttributes &attributes)
45 {
46 ARM_COMPUTE_RETURN_ERROR_ON(src == dst);
47 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
48
49 TensorInfo dst_info_to_validate;
50 const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate;
51
52 if(dst != nullptr)
53 {
54 dst_info_to_validate_ptr = dst;
55 }
56
57 auto_init_if_empty(dst_info_to_validate, src->clone()->set_data_type(attributes.data_type()));
58
59 // Check support level
60 // Data Type
61 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src,
62 1,
63 DataType::U8, DataType::S8, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8_PER_CHANNEL, DataType::S16,
64 DataType::U16, DataType::U32, DataType::S32, DataType::F16,
65 DataType::F32);
66 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(dst_info_to_validate_ptr,
67 1,
68 DataType::U8, DataType::S8, DataType::QASYMM8, DataType::S16,
69 DataType::U16, DataType::U32, DataType::S32, DataType::F16,
70 DataType::F32);
71
72 if(context.gpu_language() == GpuLanguage::OpenCL)
73 {
74 const auto cl_compile_ctx = context.cl_compile_context();
75 ARM_COMPUTE_RETURN_ERROR_ON(cl_compile_ctx == nullptr);
76 // Validate Cast Component
77 {
78 const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
79 auto settings = ClComponentCast::Settings();
80
81 ArgumentPack<ITensorInfo> arguments;
82 arguments.add_const_tensor(ACL_SRC_0, src);
83 arguments.add_const_tensor(ACL_DST_0, dst_info_to_validate_ptr);
84 ARM_COMPUTE_RETURN_ON_ERROR(ClComponentCast::validate(properties, arguments, attributes, settings));
85 }
86 }
87 else
88 {
89 ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
90 }
91
92 return Status{};
93 }
94 constexpr GpuOperatorType operator_type = GpuOperatorType::Simple;
95 } // namespace
96
is_supported_op(const GpuWorkloadContext & context,const ITensorInfo * src,const CastAttributes & attributes)97 Status GpuCast::is_supported_op(const GpuWorkloadContext &context,
98 const ITensorInfo *src,
99 const CastAttributes &attributes)
100 {
101 return is_supported_op_helper(context, src, nullptr, attributes);
102 }
103
validate_op(const GpuWorkloadSketch & sketch,const ITensorInfo * src,const CastAttributes & attributes)104 Status GpuCast::validate_op(const GpuWorkloadSketch &sketch,
105 const ITensorInfo *src,
106 const CastAttributes &attributes)
107 {
108 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
109 ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id());
110
111 // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object
112 TensorInfo dst_info_to_validate;
113
114 // Auto initialize dst tensor info
115 auto_init_if_empty(dst_info_to_validate, src->clone()->set_data_type(attributes.data_type()));
116
117 // Perform fusion test
118 // Pack tensor infos
119 ArgumentPack<ITensorInfo> tensors;
120 tensors.add_const_tensor(ACL_SRC_0, src);
121 tensors.add_const_tensor(ACL_DST_0, &dst_info_to_validate);
122 const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
123 ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
124 "Operator fusion test failed. This operator cannot be fused into the workload");
125
126 // Check if configuration is supported
127 return is_supported_op_helper(*sketch.gpu_context(), src, &dst_info_to_validate, attributes);
128 }
129
create_op(GpuWorkloadSketch & sketch,ITensorInfo * src,const CastAttributes & attributes)130 ITensorInfo *GpuCast::create_op(GpuWorkloadSketch &sketch,
131 ITensorInfo *src,
132 const CastAttributes &attributes)
133 {
134 ARM_COMPUTE_ERROR_ON_NULLPTR(src);
135 ARM_COMPUTE_LOG_PARAMS(src, attributes);
136 ARM_COMPUTE_ERROR_THROW_ON(GpuCast::validate_op(sketch, src, attributes));
137
138 ITensorInfo *dst = sketch.implementation().create_virtual_tensor();
139 ARM_COMPUTE_ERROR_ON_NULLPTR(dst);
140
141 // Auto initialize dst tensor info if empty
142 auto_init_if_empty(*dst, src->clone()->set_data_type(attributes.data_type()));
143
144 // Translate into components and add to component graph
145 GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph();
146 const auto *sketch_ctx = sketch.implementation().context();
147
148 if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
149 {
150 ARM_COMPUTE_ERROR_ON(sketch_ctx->cl_compile_context() == nullptr);
151
152 // Add Depthwise Conv2d Component
153 {
154 const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
155 auto settings = ClComponentCast::Settings();
156
157 ArgumentPack<ITensorInfo> arguments;
158 arguments.add_const_tensor(ACL_SRC_0, src);
159 arguments.add_const_tensor(ACL_DST_0, dst);
160 comp_graph.add_new_component<ClComponentCast>(properties, arguments, attributes, settings);
161 }
162 }
163 else
164 {
165 ARM_COMPUTE_ERROR("Unimplemented Gpu language");
166 }
167
168 // Set up fusion test by adding to the Operator Group
169 // Note this has to be performed after all the components have been successfully added to the component graph
170
171 // Pack tensor infos
172 ArgumentPack<ITensorInfo> tensors;
173 tensors.add_const_tensor(ACL_SRC_0, src);
174 tensors.add_const_tensor(ACL_DST_0, dst);
175
176 const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
177 sketch.implementation().operator_group().add_operator(op);
178
179 return dst;
180 }
181
182 } // namespace dynamic_fusion
183 } // namespace experimental
184 } // namespace arm_compute
185