diff options
Diffstat (limited to 'src/dynamic_fusion/runtime')
5 files changed, 100 insertions, 65 deletions
diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp index 15a5632d0b..9ca20fa152 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp +++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.cpp @@ -22,14 +22,15 @@ * SOFTWARE. */ #include "ClKernelRuntime.h" + #include "arm_compute/core/CL/ICLTensor.h" + #include "src/core/CL/CLUtils.h" #ifdef ACL_INTERNAL_TEST_CKW_IN_DF #include "src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h" #endif // ACL_INTERNAL_TEST_CKW_IN_DF #include "src/dynamic_fusion/sketch/gpu/GpuKernelSourceCode.h" #include "src/gpu/cl/ClKernelLibrary.h" - #include "support/Cast.h" namespace arm_compute { @@ -43,13 +44,12 @@ void ClKernelRuntime::configure(const ClCompileContext &compile_ctx, const GpuKe { // Create kernel from kernel source string opencl::ClKernelLibrary &klib = opencl::ClKernelLibrary::get(); - _kernel = static_cast<cl::Kernel>(compile_ctx.create_kernel(code.name(), - code.name(), // program name has to be provided to differentiate between different unfusable components' kernels. - // Each program contains exactly one kernel - code.code(), - klib.kernel_path() /* Kernel path: Used in cases of embedded kernels */, - code.build_options().options(), - false /* Is source binary */)); + _kernel = static_cast<cl::Kernel>(compile_ctx.create_kernel( + code.name(), + code.name(), // program name has to be provided to differentiate between different unfusable components' kernels. + // Each program contains exactly one kernel + code.code(), klib.kernel_path() /* Kernel path: Used in cases of embedded kernels */, + code.build_options().options(), false /* Is source binary */)); // Configure execution window IClKernel::configure_internal(code.window()); @@ -63,11 +63,15 @@ void ClKernelRuntime::configure(const ClCompileContext &compile_ctx, const GpuKe #ifndef ACL_INTERNAL_TEST_CKW_IN_DF -inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images) +inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, + const GpuKernelArgumentInfo &arg, + const ICLTensor *tensor, + const Window &arg_slice, + std::vector<cl::Image2D> &cl_images) { ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); - switch(arg.type) + switch (arg.type) { case GpuKernelArgumentInfo::Type::Scalar: { @@ -95,9 +99,13 @@ inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKer } case GpuKernelArgumentInfo::Type::Image_Export_To_ClImage2D: { - const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3)); + const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * + tensor->info()->dimension(2) * + tensor->info()->dimension(3)); const size_t image_row_pitch = tensor->info()->strides_in_bytes()[1]; - cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly); + cl::Image2D tensor_image2d = + create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, + tensor->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly); cl_images.push_back(tensor_image2d); _kernel.setArg(idx++, tensor_image2d); break; @@ -111,9 +119,13 @@ inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKer } case GpuKernelArgumentInfo::Type::Image_3D_Export_To_ClImage2D: { - const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * tensor->info()->dimension(2) * tensor->info()->dimension(3)); + const TensorShape shape2d(tensor->info()->dimension(0) / 4, tensor->info()->dimension(1) * + tensor->info()->dimension(2) * + tensor->info()->dimension(3)); const size_t image_row_pitch = tensor->info()->strides_in_bytes()[1]; - cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, tensor->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly); + cl::Image2D tensor_image2d = + create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), shape2d, + tensor->info()->data_type(), image_row_pitch, CLImage2DType::ReadOnly); cl_images.push_back(tensor_image2d); _kernel.setArg(idx++, tensor_image2d); _kernel.setArg<cl_uint>(idx++, static_cast<unsigned int>(tensor->info()->strides_in_bytes()[2])); @@ -142,8 +154,9 @@ inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKer const size_t image_h = tensor->info()->tensor_shape().total_size_upper(1); const size_t image_stride_y = tensor->info()->strides_in_bytes()[1]; - cl::Image2D tensor_image2d = create_image2d_from_buffer(CLKernelLibrary::get().context(), tensor->cl_buffer(), - TensorShape(image_w, image_h), tensor->info()->data_type(), image_stride_y, CLImage2DType::ReadOnly); + cl::Image2D tensor_image2d = create_image2d_from_buffer( + CLKernelLibrary::get().context(), tensor->cl_buffer(), TensorShape(image_w, image_h), + tensor->info()->data_type(), image_stride_y, CLImage2DType::ReadOnly); cl_images.push_back(tensor_image2d); _kernel.setArg(idx++, tensor_image2d); @@ -170,13 +183,16 @@ inline void ClKernelRuntime::add_tensor_argument(unsigned int &idx, const GpuKer } #else // ACL_INTERNAL_TEST_CKW_IN_DF -inline void ClKernelRuntime::add_kernel_argument(unsigned int &idx, const GpuKernelArgumentBinding &arg, const ICLTensor *tensor, std::vector<cl::Image2D> &cl_images) +inline void ClKernelRuntime::add_kernel_argument(unsigned int &idx, + const GpuKernelArgumentBinding &arg, + const ICLTensor *tensor, + std::vector<cl::Image2D> &cl_images) { - switch(arg.type()) + switch (arg.type()) { case GpuKernelArgumentBinding::Type::TensorStorage: { - switch(arg.tensor_storage_type()) + switch (arg.tensor_storage_type()) { case TensorStorageType::ClBufferUint8Ptr: { @@ -238,7 +254,7 @@ void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::Com // CLImages created from tensor arguments. Need to be retained until enqueue std::vector<cl::Image2D> cl_images; #ifndef ACL_INTERNAL_TEST_CKW_IN_DF - for(auto id_arg : _arguments) + for (auto id_arg : _arguments) { const auto arg = id_arg.second; auto tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(id_arg.first)); @@ -248,7 +264,7 @@ void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::Com } #else // ACL_INTERNAL_TEST_CKW_IN_DF - for(const auto &arg : _arguments) + for (const auto &arg : _arguments) { auto tensor = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(arg.id())); ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); @@ -259,8 +275,7 @@ void ClKernelRuntime::run_op(ITensorPack &tensors, const Window &window, cl::Com // Dispatch kernel enqueue(queue, *this, slice, lws_hint(), use_dummy_work_items); - } - while(skip_sliding_window && window.slide_window_slice_3D(slice)); + } while (skip_sliding_window && window.slide_window_slice_3D(slice)); } } // namespace dynamic_fusion diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h index 92e73503ce..e78567eb9d 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h +++ b/src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h @@ -68,7 +68,11 @@ private: * @param[in] arg_slice Window the kernel will be run on * @param[out] cl_images Extra cl images created from the tensor (will need to be retained until the kernel is enqueued) */ - inline void add_tensor_argument(unsigned int &idx, const GpuKernelArgumentInfo &arg, const ICLTensor *tensor, const Window &arg_slice, std::vector<cl::Image2D> &cl_images); + inline void add_tensor_argument(unsigned int &idx, + const GpuKernelArgumentInfo &arg, + const ICLTensor *tensor, + const Window &arg_slice, + std::vector<cl::Image2D> &cl_images); #else // ACL_INTERNAL_TEST_CKW_IN_DF /** Set a kernel argument as part of a tensor * @@ -77,7 +81,10 @@ private: * @param[in] tensor Tensor of which the kernel argument @p arg is a part of * @param[out] cl_images Extra cl images created from the tensor (will need to be retained until the kernel is enqueued) */ - inline void add_kernel_argument(unsigned int &idx, const GpuKernelArgumentBinding &arg, const ICLTensor *tensor, std::vector<cl::Image2D> &cl_images); + inline void add_kernel_argument(unsigned int &idx, + const GpuKernelArgumentBinding &arg, + const ICLTensor *tensor, + std::vector<cl::Image2D> &cl_images); #endif // ACL_INTERNAL_TEST_CKW_IN_DF private: diff --git a/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp index cd21b10180..ba39ff4c9d 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp +++ b/src/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.cpp @@ -25,6 +25,7 @@ #include "arm_compute/core/experimental/Types.h" #include "arm_compute/runtime/CL/CLTensor.h" + #include "src/dynamic_fusion/runtime/gpu/cl/ClKernelRuntime.h" #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSketchImpl.h" #include "src/dynamic_fusion/sketch/gpu/GpuWorkloadSourceCode.h" @@ -55,14 +56,14 @@ public: { DataView() = default; DataView(CLTensor *tensor, const TensorInfo &tensor_info, const AuxMemoryInfo &memory_info) - : tensor{ tensor }, tensor_info{ tensor_info }, memory_info{ memory_info } + : tensor{tensor}, tensor_info{tensor_info}, memory_info{memory_info} { } - ~DataView() = default; - DataView(const DataView &other) = default; + ~DataView() = default; + DataView(const DataView &other) = default; DataView &operator=(const DataView &other) = default; DataView(DataView &&other) = default; - DataView &operator=(DataView &&other) = default; + DataView &operator=(DataView &&other) = default; CLTensor *tensor{}; /**< Pointer to the auxiliary tensor */ TensorInfo tensor_info{}; /**< Associated tensor info */ AuxMemoryInfo memory_info{}; /**< Memory requirement */ @@ -92,7 +93,7 @@ private: { const auto t_id = tensor_info.id(); auto find_tensor_pair = _owned_tensors.find(t_id); - if(find_tensor_pair != _owned_tensors.end()) + if (find_tensor_pair != _owned_tensors.end()) { return find_tensor_pair->second.get(); } @@ -107,7 +108,7 @@ private: } std::map<ITensorInfo::Id, std::unique_ptr<CLTensor>> _owned_tensors{}; - std::vector<DataView> _tensors{}; + std::vector<DataView> _tensors{}; }; /** Construct auxiliary tensors required by @ref GpuWorkloadSourceCode * @@ -120,12 +121,12 @@ private: */ Status create_aux_tensors(ClAuxTensors *aux_tensors, const GpuWorkloadSourceCode &code) { - for(auto t_id : code.tensors()) + for (auto t_id : code.tensors()) { // Get tensor object const auto workload_arg = code.query_tensor(t_id); ICLTensor *tensor_object = nullptr; - if(workload_arg->memory_descriptor()->memory_type == MemoryType::Auxiliary) + if (workload_arg->memory_descriptor()->memory_type == MemoryType::Auxiliary) { // Create aux tensor CLTensor object const TensorInfo tensor_info = *workload_arg->tensor_info(); @@ -133,7 +134,7 @@ Status create_aux_tensors(ClAuxTensors *aux_tensors, const GpuWorkloadSourceCode 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) + if (tensor_object == nullptr) { return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Failed to construct an auxiliary tensor"); } @@ -156,7 +157,7 @@ public: ITensorPack *find_tensor_pack(UnitWorkloadId uwk_id) { auto tensor_pack = _tensor_packs.find(uwk_id); - if(tensor_pack != _tensor_packs.end()) + if (tensor_pack != _tensor_packs.end()) { return &(tensor_pack->second); } @@ -173,7 +174,10 @@ public: return _tensor_packs.at(uwk_id); } - friend Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &code, const std::vector<CLTensor *> &user_tensors, const ClAuxTensors &aux_tensors); + friend Status create_tensor_lut(ClTensorLUT *tensor_lut, + const GpuWorkloadSourceCode &code, + const std::vector<CLTensor *> &user_tensors, + const ClAuxTensors &aux_tensors); private: /** Add a tensor pack and associate it with @ref UnitWorkloadId @p uwk_id @@ -197,19 +201,22 @@ private: * * @return Status */ -Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &code, const std::vector<CLTensor *> &user_tensors, const ClAuxTensors &aux_tensors) +Status create_tensor_lut(ClTensorLUT *tensor_lut, + const GpuWorkloadSourceCode &code, + const std::vector<CLTensor *> &user_tensors, + const ClAuxTensors &aux_tensors) { // Combine user tensors and aux tensors std::map<ITensorInfo::Id, CLTensor *> tensor_map; - for(auto tensor : user_tensors) + for (auto tensor : user_tensors) { const auto t_id = tensor->info()->id(); - if(tensor_map.find(t_id) != tensor_map.end()) + if (tensor_map.find(t_id) != tensor_map.end()) { // In case of elementwise in-place: give another Id to the In/Out tensor when passed again std::vector<ITensorInfo::Id> ids; - for(auto &t : tensor_map) + for (auto &t : tensor_map) { ids.push_back(t.first); } @@ -221,11 +228,11 @@ Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &c tensor_map[t_id] = tensor; } } - for(const auto &data : aux_tensors.get_tensors()) + for (const auto &data : aux_tensors.get_tensors()) { const auto t_id = data.tensor_info.id(); const auto tensor = data.tensor; - if(tensor_map.find(t_id) != tensor_map.end()) + if (tensor_map.find(t_id) != tensor_map.end()) { return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Clashing tensor ids"); } @@ -233,25 +240,25 @@ Status create_tensor_lut(ClTensorLUT *tensor_lut, const GpuWorkloadSourceCode &c } // Add tensor objects into corresponding tensor packs - for(auto id_tensor : tensor_map) + for (auto id_tensor : tensor_map) { const auto t_id = id_tensor.first; const auto tensor_object = id_tensor.second; - if(tensor_object == nullptr) + if (tensor_object == nullptr) { return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Trying to add a nullptr into the tensor packs"); } - if(tensor_object->allocator()->info().total_size() == 0U) + if (tensor_object->allocator()->info().total_size() == 0U) { return ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "No allocated memory found in tensor"); } - for(auto uwk_id : code.get_unit_workloads_from_tensor(t_id)) + for (auto uwk_id : code.get_unit_workloads_from_tensor(t_id)) { ITensorPack *tensor_pack = tensor_lut->find_tensor_pack(uwk_id); - if(tensor_pack == nullptr) + if (tensor_pack == nullptr) { - tensor_lut->add_tensor_pack(uwk_id, ITensorPack{ { t_id, tensor_object } }); + tensor_lut->add_tensor_pack(uwk_id, ITensorPack{{t_id, tensor_object}}); } else { @@ -269,15 +276,14 @@ struct ClWorkloadRuntime::Implementation { std::map<UnitWorkloadId, std::unique_ptr<ClKernelRuntime>> _kernels{}; std::map<UnitWorkloadId, std::unique_ptr<ClKernelRuntime>> _kernels_prep{}; - bool _is_configured{ false }; - bool _is_prepared{ false }; - ClTensorLUT _tensor_lut{}; - ClAuxTensors _aux_tensors{}; - GpuWorkloadSourceCode _source_code{}; + bool _is_configured{false}; + bool _is_prepared{false}; + ClTensorLUT _tensor_lut{}; + ClAuxTensors _aux_tensors{}; + GpuWorkloadSourceCode _source_code{}; }; -ClWorkloadRuntime::ClWorkloadRuntime() - : _impl{ std::make_unique<Implementation>() } +ClWorkloadRuntime::ClWorkloadRuntime() : _impl{std::make_unique<Implementation>()} { } @@ -286,18 +292,19 @@ ClWorkloadRuntime::~ClWorkloadRuntime() = default; Status ClWorkloadRuntime::configure(const GpuWorkloadSketch &sketch) { ARM_COMPUTE_RETURN_ERROR_ON_MSG(_impl->_is_configured, "ClWorkloadRuntime cannot be re-configured"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(sketch.gpu_context()->gpu_language() != GpuLanguage::OpenCL, "ClWorkloadRuntime cannot be configured with non-OpenCL workload sketch"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(sketch.gpu_context()->gpu_language() != GpuLanguage::OpenCL, + "ClWorkloadRuntime cannot be configured with non-OpenCL workload sketch"); // Generate source code _impl->_source_code = sketch.implementation().generate_source_code(); // Configure unit workload from source code - for(auto uwk_id : _impl->_source_code.unit_workloads()) + for (auto uwk_id : _impl->_source_code.unit_workloads()) { const auto work = _impl->_source_code.query_unit_workload(uwk_id); const auto stage = work.stage().stage; auto k = std::make_unique<ClKernelRuntime>(); k->configure(*sketch.gpu_context()->cl_compile_context(), work.code()); - switch(stage) + switch (stage) { case UnitWorkloadStage::Stage::Run: { @@ -323,9 +330,9 @@ Status ClWorkloadRuntime::configure(const GpuWorkloadSketch &sketch) void ClWorkloadRuntime::prepare() { - if(!_impl->_is_prepared) + if (!_impl->_is_prepared) { - for(auto &id_kernel_pair : _impl->_kernels_prep) + for (auto &id_kernel_pair : _impl->_kernels_prep) { const bool flush_queue = false; const auto uwk_id = id_kernel_pair.first; @@ -344,7 +351,7 @@ Status ClWorkloadRuntime::run(const std::vector<CLTensor *> &tensors) const auto st = create_tensor_lut(&_impl->_tensor_lut, _impl->_source_code, tensors, _impl->_aux_tensors); ARM_COMPUTE_RETURN_ON_ERROR(st); prepare(); - for(auto &id_kernel_pair : _impl->_kernels) + for (auto &id_kernel_pair : _impl->_kernels) { // Flush the command queue on the last kernel const bool flush_queue = false; @@ -358,7 +365,7 @@ Status ClWorkloadRuntime::run(const std::vector<CLTensor *> &tensors) std::vector<std::tuple<CLTensor *, TensorInfo, AuxMemoryInfo>> ClWorkloadRuntime::get_auxiliary_tensors() { std::vector<std::tuple<CLTensor *, TensorInfo, AuxMemoryInfo>> aux_tensors; - for(const auto &data : _impl->_aux_tensors.get_tensors()) + for (const auto &data : _impl->_aux_tensors.get_tensors()) { aux_tensors.emplace_back(data.tensor, data.tensor_info, data.memory_info); } diff --git a/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.cpp b/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.cpp index 84fb279237..7044b0ea66 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.cpp +++ b/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.cpp @@ -30,14 +30,17 @@ namespace experimental { namespace dynamic_fusion { -void cl_add_tensor_component_argument(cl::Kernel &kernel, unsigned int &idx, const ICLTensor *tensor, TensorComponentType component) +void cl_add_tensor_component_argument(cl::Kernel &kernel, + unsigned int &idx, + const ICLTensor *tensor, + TensorComponentType component) { ARM_COMPUTE_ERROR_ON(tensor == nullptr); const auto *info = tensor->info(); const auto &strides = info->strides_in_bytes(); - switch(component) + switch (component) { case TensorComponentType::OffsetFirstElement: kernel.setArg<cl_uint>(idx++, info->offset_first_element_in_bytes()); diff --git a/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h b/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h index 4cbb157a48..306d547acb 100644 --- a/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h +++ b/src/dynamic_fusion/runtime/gpu/cl/ckw_driver/GpuCkwKernelArgumentsHelpers.h @@ -42,7 +42,10 @@ namespace dynamic_fusion * @param[in] tensor Tensor from which to access the tensor component. * @param[in] component Tensor component to select such as tensor dimensions, strides, etc. */ -void cl_add_tensor_component_argument(cl::Kernel &kernel, unsigned int &idx, const ICLTensor *tensor, TensorComponentType component); +void cl_add_tensor_component_argument(cl::Kernel &kernel, + unsigned int &idx, + const ICLTensor *tensor, + TensorComponentType component); /** Add an OpenCL buffer object to the kernel's arguments at the specified index @p idx. * |