From b84e25313e5dc7acbc03623e1e071e845047c111 Mon Sep 17 00:00:00 2001 From: Viet-Hoa Do Date: Tue, 13 Dec 2022 13:09:10 +0000 Subject: Add output operator for dynamic fusion * The output of the fused operator must be explicitly specified using GpuOutput operator. * Any temporary tensors used to connect the output of an operator to the input of another operator will be marked as no-alloc and won't be allocated as a tensor in the memory. Resolves: COMPMID-5771 Signed-off-by: Viet-Hoa Do Change-Id: I5ae8e800f8f737db23a055a92b01c4f1d78c3bb8 Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8794 Tested-by: Arm Jenkins Reviewed-by: SiCong Li Reviewed-by: Gian Marco Iodice Reviewed-by: Gunes Bayir Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../runtime/gpu/cl/ClWorkloadRuntime.cpp | 9 +- .../sketch/gpu/GpuKernelComponentGraph.cpp | 13 +- src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp | 27 ---- src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.h | 2 - .../sketch/gpu/operators/GpuOutput.cpp | 141 +++++++++++++++++++++ .../cl/ClTemplateElementwiseBinary.cpp | 5 +- .../gpu/template_writer/cl/ClTemplateWriter.cpp | 7 +- src/dynamic_fusion/sketch/utils/DependencyGraph.h | 27 ++++ 8 files changed, 193 insertions(+), 38 deletions(-) create mode 100644 src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp (limited to 'src') diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp index 36168d14f1..7e427fef72 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp +++ b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp @@ -132,10 +132,11 @@ Status create_aux_tensors(ClAuxTensors *aux_tensors, const GpuWorkloadSourceCode ARM_COMPUTE_ERROR_ON(tensor_info.id() != t_id); const auto aux_memory_info = workload_arg->memory_descriptor()->aux_memory_info; tensor_object = aux_tensors->add_aux_tensor(tensor_info, aux_memory_info); - } - if(tensor_object == nullptr) - { - return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Failed to construct an auxiliary tensor"); + + if(tensor_object == nullptr) + { + return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Failed to construct an auxiliary tensor"); + } } } return Status{}; diff --git a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp index 6e6422c957..1f90aab477 100644 --- a/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp +++ b/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.cpp @@ -42,15 +42,24 @@ namespace */ MemoryDescriptorMap assign_memory_descriptors(const std::map tensors, const DependencyGraph &graph) { + const auto all_tensors = graph.all_tensors(); + const auto src_tensors = graph.global_src_tensors(); + const auto dst_tensors = graph.global_dst_tensors(); + const auto interm_tensors = graph.intermediate_tensors(); + MemoryDescriptorMap mem_map{}; - for(auto t_id : graph.all_tensors()) + for(auto t_id : all_tensors) { const auto &tensor = tensors.at(t_id); // Only global src and dst tensors to the entire component graph are "User" tensors, which are user-specified memories - if(is_in(t_id, graph.global_src_tensors()) || is_in(t_id, graph.global_dst_tensors())) + if(is_in(t_id, src_tensors) || is_in(t_id, dst_tensors)) { mem_map[t_id] = MemoryDescriptor{ MemoryType::User }; } + else if(is_in(t_id, interm_tensors)) + { + mem_map[t_id] = MemoryDescriptor { MemoryType::NoAlloc }; + } else { AuxMemoryInfo aux_mem_info{ tensor->total_size() }; diff --git a/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp b/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp index c560e9a931..66760b3812 100644 --- a/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp +++ b/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.cpp @@ -40,7 +40,6 @@ namespace dynamic_fusion GpuLogicalKernel::GpuLogicalKernel(GpuComponentServices *services, const GpuKernelComponentGroup &components) : _services{ services }, _comp_group{ components }, _store_components{} { - add_load_store(); } GpuKernelSourceCode GpuLogicalKernel::write_kernel_code() @@ -57,32 +56,6 @@ GpuKernelSourceCode GpuLogicalKernel::write_kernel_code() return code; } - -void GpuLogicalKernel::add_load_store() -{ - const auto dst_tensors = _comp_group.get_dst_tensors(); - // Each dst tensor from the component group requires exactly one store component - for(const auto &dst_tensor : dst_tensors) - { - ArgumentPack tensors; - // Pass same destination tensor to both source and destination of the store component - // In other words, the addition of a store component does not create a new dst tensor - // This way we avoid the issue of the dst tensor of the component group differs from that of a logical kernel - // This may seem to violate the acyclic-ness of the component graph. But it is fine because at the point of - // the construction of the logical kernel, we do not need a graph representation of components anymore - // (the graph has been serialized) - tensors.add_const_tensor(ACL_SRC_0, dst_tensor); - tensors.add_const_tensor(ACL_DST_0, dst_tensor); - - auto store = _services->component_factory().create( - _comp_group.get_root_component()->properties(), // Store component share the same properties as that of the root component - tensors); - _store_components.push_back(std::move(store)); - auto success = _comp_group.add_component(_store_components.back().get()); - ARM_COMPUTE_UNUSED(success); - ARM_COMPUTE_ERROR_ON(!success); // It's guaranteed that any load store insertion should be successful - } -} } // namespace dynamic_fusion } // namespace experimental } // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.h b/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.h index 4ce4443f60..2654224329 100644 --- a/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.h +++ b/src/dynamic_fusion/sketch/gpu/GpuLogicalKernel.h @@ -65,8 +65,6 @@ public: GpuKernelSourceCode write_kernel_code(); private: - void add_load_store(); - GpuComponentServices *_services; GpuKernelComponentGroup _comp_group{}; std::vector> _store_components{}; diff --git a/src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp b/src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp new file mode 100644 index 0000000000..017536df6c --- /dev/null +++ b/src/dynamic_fusion/sketch/gpu/operators/GpuOutput.cpp @@ -0,0 +1,141 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "src/core/helpers/AutoConfiguration.h" +#include "src/common/utils/Log.h" + +#include "src/dynamic_fusion/sketch/ArgumentPack.h" +#include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" +#include "src/dynamic_fusion/sketch/gpu/components/cl/ClComponentStore.h" + +namespace arm_compute +{ +namespace experimental +{ +namespace dynamic_fusion +{ +namespace +{ +constexpr GpuOperatorType operator_type = GpuOperatorType::Simple; +} + +Status GpuOutput::is_supported_op(const GpuWorkloadContext &context, + const ITensorInfo *src, + const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + + // Initialize the destination tensor info. + TensorInfo dst_to_validate = *dst; + auto_init_if_empty(dst_to_validate, *src); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, &dst_to_validate); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, &dst_to_validate); + + ARM_COMPUTE_UNUSED(context); + return Status{}; +} + +Status GpuOutput::validate_op(const GpuWorkloadSketch &sketch, + const ITensorInfo *src, + const ITensorInfo *dst) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, dst); + ARM_COMPUTE_RETURN_ERROR_ON(!src->has_valid_id()); + ARM_COMPUTE_RETURN_ERROR_ON(!dst->has_valid_id()); + + // Initialize the destination tensor info. + TensorInfo dst_to_validate = *dst; + auto_init_if_empty(dst_to_validate, *src); + + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, &dst_to_validate); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(src, &dst_to_validate); + + // Perform fusion test. + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC_0, src); + tensors.add_const_tensor(ACL_DST_0, &dst_to_validate); + + const auto group = sketch.implementation().operator_group(); + const auto op = group.new_operator(operator_type, tensors); + const auto success = group.try_add_operator(op); + + ARM_COMPUTE_RETURN_ERROR_ON_MSG(!success, "This operator cannot be fused into the workload."); + ARM_COMPUTE_UNUSED(success); + + const auto status = is_supported_op(*sketch.gpu_context(), src, dst); + return status; +} + +void GpuOutput::create_op(GpuWorkloadSketch &sketch, + ITensorInfo *src, + ITensorInfo *dst) +{ + ARM_COMPUTE_LOG_PARAMS(src, dst); + ARM_COMPUTE_ERROR_THROW_ON(GpuOutput::validate_op(sketch, src, dst)); + + // Auto initialize dst tensor info if empty + auto_init_if_empty(*dst, *src); + + // Translate into components and add to component graph + auto &comp_graph = sketch.implementation().component_graph(); + const auto sketch_ctx = sketch.implementation().context(); + + if(sketch_ctx->gpu_language() == GpuLanguage::OpenCL) + { + ARM_COMPUTE_ERROR_ON(sketch_ctx->cl_compile_context() == nullptr); + + // Add store component + { + IGpuKernelComponent::Properties properties; + properties.stage(UnitWorkloadStage{ UnitWorkloadStage::Stage::Run }); + + ArgumentPack arguments; + arguments.add_const_tensor(ACL_SRC_0, src); + arguments.add_const_tensor(ACL_DST_0, dst); + comp_graph.add_new_component(properties, arguments); + } + } + else + { + ARM_COMPUTE_ERROR("Unimplemented Gpu language"); + } + + // Set up fusion test by adding to the Operator Group + // Note this has to be performed after all the components have been successfully added to the component graph + + // Pack tensor infos + ArgumentPack tensors; + tensors.add_const_tensor(ACL_SRC_0, src); + tensors.add_const_tensor(ACL_DST_0, dst); + + const Operator op = sketch.implementation().operator_group().new_operator(operator_type, tensors); + sketch.implementation().operator_group().add_operator(op); +} + +} // namespace dynamic_fusion +} // namespace experimental +} // namespace arm_compute diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp index 996bf15d01..6c1e0fb1de 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateElementwiseBinary.cpp @@ -110,8 +110,8 @@ R"_( R"_( LOOP_UNROLLING(int, i, 0, 1, M0, { - g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{dst}}_w * {{dst}}_h) - 1); - g_dst_indirect_y[i].v += g_ind_2 * (int)({{dst}}_w * {{dst}}_h); + g_dst_indirect_y[i].v = (uint)min(g_ind_1 + i, (int)({{out}}_w * {{out}}_h) - 1); + g_dst_indirect_y[i].v += g_ind_2 * (int)({{out}}_w * {{out}}_h); }) } //------------------ END KERNEL {{meta_kernel_id}} ELTWISE_OP --------------------- @@ -194,6 +194,7 @@ TagLUT ClTemplateElementwiseBinary::get_tag_lut(const GpuKernelVariableTable &vt lut["lhs"] = vtable.get_variable(_lhs); lut["rhs"] = vtable.get_variable(_rhs); lut["dst"] = vtable.get_variable(_dst); + lut["out"] = vtable.get_variable(comp_group.get_dst_tensors().front()); } else { diff --git a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp index cb643a741d..0afd0e7581 100644 --- a/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp +++ b/src/dynamic_fusion/sketch/gpu/template_writer/cl/ClTemplateWriter.cpp @@ -179,7 +179,11 @@ std::string ClTemplateWriter::write_code() code += macros; } - code += write_kernel_signature(_vtable.get_variable_list(_components.get_argument_tensors())); + auto arguments = _components.get_argument_tensors(); + std::sort(arguments.begin(), arguments.end(), [](const ITensorInfo *l, const ITensorInfo *r) { + return l->id() < r->id(); + }); + code += write_kernel_signature(_vtable.get_variable_list(arguments)); code += "\n{\n\n"; @@ -190,6 +194,7 @@ std::string ClTemplateWriter::write_code() for(const auto &component_code : component_codes) { code += component_code; + code += "\n"; } code += "}\n"; diff --git a/src/dynamic_fusion/sketch/utils/DependencyGraph.h b/src/dynamic_fusion/sketch/utils/DependencyGraph.h index 03678defae..633c5e4263 100644 --- a/src/dynamic_fusion/sketch/utils/DependencyGraph.h +++ b/src/dynamic_fusion/sketch/utils/DependencyGraph.h @@ -417,6 +417,33 @@ public: } return tensors; } + /** Get intermediate tensors of the whole graph. + * + * @return std::vector + */ + std::vector intermediate_tensors() const + { + std::vector tensors; + + // If a tensor is used to connect the input of an operator and the output of another operator, + // it is not allocated in the memory. The tensor exists as a temporary variable only. + for(auto src_tensor : _adj_src_ops) + { + if(!src_tensor.second.empty()) + { + const auto dst_tensor = _adj_dst_ops.find(src_tensor.first); + if(dst_tensor != _adj_dst_ops.end()) + { + if(!dst_tensor->second.empty()) + { + tensors.push_back(src_tensor.first); + } + } + } + } + + return tensors; + } /** Get all root ops. Root ops can also be referred to as "src ops" of the whole graph * * @return std::vector -- cgit v1.2.1