xref: /aosp_15_r20/external/ComputeLibrary/src/dynamic_fusion/sketch/gpu/operators/GpuClamp.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/GpuClamp.h"
25 
26 #include "arm_compute/core/experimental/Types.h"
27 
28 #include "src/core/helpers/AutoConfiguration.h"
29 #include "src/dynamic_fusion/sketch/ArgumentPack.h"
30 #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h"
31 #include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentActivation.h"
32 #include "src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateActivation.h"
33 
34 #include "src/common/utils/Log.h"
35 
36 namespace arm_compute
37 {
38 namespace experimental
39 {
40 namespace dynamic_fusion
41 {
42 namespace
43 {
is_supported_op_helper(const GpuWorkloadContext & context,const ITensorInfo * src,const ITensorInfo * dst,const ClampAttributes & attributes)44 Status is_supported_op_helper(const GpuWorkloadContext &context,
45                               const ITensorInfo        *src,
46                               const ITensorInfo        *dst,
47                               const ClampAttributes    &attributes)
48 {
49     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst);
50     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
51     ARM_COMPUTE_RETURN_ERROR_ON_MSG(attributes.max_val() < attributes.min_val(), "Maximum clamp value cannot be lower than minimum value");
52 
53     TensorInfo         dst_info_to_validate;
54     const ITensorInfo *dst_info_to_validate_ptr = &dst_info_to_validate;
55 
56     if(dst != nullptr)
57     {
58         dst_info_to_validate_ptr = dst;
59     }
60 
61     auto_init_if_empty(dst_info_to_validate, *src->clone());
62 
63     // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped
64     const ClComponentActivation::Attributes act_info
65     {
66         ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val()
67     };
68 
69     // Check components
70     if(context.gpu_language() == GpuLanguage::OpenCL)
71     {
72         // Validate Activation Component
73         const auto properties = IGpuKernelComponent::Properties().stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
74 
75         ArgumentPack<ITensorInfo> arguments;
76         arguments.add_const_tensor(ACL_SRC, src);
77         arguments.add_const_tensor(ACL_DST, dst_info_to_validate_ptr);
78         ARM_COMPUTE_RETURN_ON_ERROR(ClComponentActivation::validate(properties, arguments, act_info));
79     }
80     else
81     {
82         ARM_COMPUTE_RETURN_ERROR_MSG("Unimplemented Gpu language");
83     }
84     return Status{};
85 }
86 
87 constexpr GpuOperatorType operator_type = GpuOperatorType::Simple;
88 } // namespace
89 
is_supported_op(const GpuWorkloadContext & context,const ITensorInfo * src,const ClampAttributes & attributes)90 Status GpuClamp::is_supported_op(const GpuWorkloadContext &context,
91                                  const ITensorInfo        *src,
92                                  const ClampAttributes    &attributes)
93 {
94     return is_supported_op_helper(context, src, nullptr, attributes);
95 }
96 
validate_op(const GpuWorkloadSketch & sketch,const ITensorInfo * src,const ClampAttributes & attributes)97 Status GpuClamp::validate_op(const GpuWorkloadSketch &sketch,
98                              const ITensorInfo       *src,
99                              const ClampAttributes   &attributes)
100 {
101     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src);
102 
103     // Check if tensors have valid id, i.e. they are created from a sketch
104     ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id());
105 
106     // Refer to GpuConv2d::validate_op() for id-validness of this TensorInfo object
107     TensorInfo dst_info_to_validate;
108 
109     // Auto initialize dst tensor info
110     auto_init_if_empty(dst_info_to_validate, *src->clone());
111 
112     // Perform fusion test to check if the operator meets fusion constraints
113     ArgumentPack<ITensorInfo> tensors;
114     tensors.add_const_tensor(ACL_SRC, src);
115     tensors.add_const_tensor(ACL_DST, &dst_info_to_validate);
116     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
117     ARM_COMPUTE_RETURN_ERROR_ON_MSG(!sketch.implementation().operator_group().try_add_operator(op),
118                                     "Operator fusion test failed. This operator cannot be fused into the workload");
119 
120     // Check if configuration is supported
121     return is_supported_op_helper(*sketch.gpu_context(), src, &dst_info_to_validate, attributes);
122 }
123 
create_op(GpuWorkloadSketch & sketch,ITensorInfo * src,const ClampAttributes & attributes)124 ITensorInfo *GpuClamp::create_op(GpuWorkloadSketch     &sketch,
125                                  ITensorInfo           *src,
126                                  const ClampAttributes &attributes)
127 {
128     ARM_COMPUTE_ERROR_ON_NULLPTR(src);
129     ARM_COMPUTE_LOG_PARAMS(src, attributes);
130     ARM_COMPUTE_ERROR_THROW_ON(GpuClamp::validate_op(sketch, src, attributes));
131 
132     ITensorInfo *dst = sketch.implementation().create_virtual_tensor();
133     ARM_COMPUTE_ERROR_ON_NULLPTR(dst);
134 
135     // Auto initialize dst tensor
136     auto_init_if_empty(*dst, *src->clone());
137 
138     // Translate into components and add to component graph
139     GpuKernelComponentGraph &comp_graph = sketch.implementation().component_graph();
140 
141     // CLAMP operator is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped
142     const ClComponentActivation::Attributes act_info
143     {
144         ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val()
145     };
146 
147     const auto *const sketch_ctx = sketch.implementation().context();
148 
149     if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL)
150     {
151         // Add Activation Component
152         auto properties = IGpuKernelComponent::Properties();
153         properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run });
154 
155         ArgumentPack<ITensorInfo> arguments;
156         arguments.add_const_tensor(ACL_SRC, src);
157         arguments.add_const_tensor(ACL_DST, dst);
158         comp_graph.add_new_component<ClComponentActivation>(properties, arguments, act_info);
159     }
160     else
161     {
162         ARM_COMPUTE_ERROR("Unimplemented Gpu language");
163     }
164 
165     // Set up fusion test by adding to the Operator Group
166     // Note this has to be performed after all the components have been successfully added to the component graph
167 
168     // Pack tensor infos
169     ArgumentPack<ITensorInfo> tensors;
170     tensors.add_const_tensor(ACL_SRC, src);
171     tensors.add_const_tensor(ACL_DST, dst);
172 
173     const auto op = sketch.implementation().operator_group().new_operator(operator_type, tensors);
174     sketch.implementation().operator_group().add_operator(op);
175 
176     return dst;
177 }
178 
179 } // namespace dynamic_fusion
180 } // namespace experimental
181 } // namespace arm_compute
182