Provides global access to a CL context and command queue. More...

#include <CLScheduler.h>

Public Member Functions
	CLScheduler ()
	Constructor. More...

	CLScheduler (const CLScheduler &)=delete
	Prevent instances of this class from being copied (As this class contains pointers) More...

CLScheduler &	operator= (const CLScheduler &)=delete
	Prevent instances of this class from being copied (As this class contains pointers) More...

	~CLScheduler ()=default
	Default destructor. More...

void	default_init (ICLTuner cl_tuner=nullptr, CLGEMMHeuristicsHandle gemm_h=nullptr, CLBackendType cl_backend_type=CLBackendType::Native)
	Initialises the context and command queue used by the scheduler to default values and sets a default device and kernel path for the CLKernelLibrary. More...

void	default_init_with_context (cl::Device &device, cl::Context &ctx, ICLTuner cl_tuner=nullptr, CLGEMMHeuristicsHandle gemm_h=nullptr)
	Initialises the scheduler with context and device provided by the user. More...

void	default_reinit (ICLTuner cl_tuner=nullptr, CLGEMMHeuristicsHandle gemm_h=nullptr, CLBackendType cl_backend_type=CLBackendType::Native)
	Re-initializes the context and command queue used by the scheduler to default values and sets a default device and kernel path for the CLKernelLibrary. More...

void	enqueue (ICLKernel &kernel, bool flush=true)
	Schedule the execution of the passed kernel if possible. More...

void	enqueue_op (ICLKernel &kernel, ITensorPack &tensors, bool flush=true)
	Schedule the execution of the passed kernel if possible. More...

void	init (cl::Context context, cl::CommandQueue queue, const cl::Device &device, ICLTuner cl_tuner=nullptr, CLGEMMHeuristicsHandle gemm_h=nullptr, CLBackendType cl_backend_type=CLBackendType::Native)
	Initialises the context and command queue to be used by the scheduler. More...

cl::Context &	context ()
	Accessor for the associated CL context. More...

cl::CommandQueue &	queue ()
	Accessor for the associated CL command queue. More...

GPUTarget	target () const
	Get the target GPU. More...

CLGEMMHeuristicsHandle *	gemm_heuristics () const
	Accessor for the associated CLGEMMHeuristicsHandle. More...

void	set_context (cl::Context context)
	Accessor to set the CL context to be used by the scheduler. More...

void	set_queue (cl::CommandQueue queue)
	Accessor to set the CL command queue to be used by the scheduler. More...

void	set_target (GPUTarget target)
	Accessor to set target GPU to be used by the scheduler. More...

void	set_tuner (ICLTuner *tuner)
	Accessor to set the CL tuner to be used by the scheduler. More...

void	sync ()
	Blocks until all commands in the associated command queue have finished. More...

cl::Event	enqueue_sync_event ()
	Enqueues a marker into the associated command queue and return the event. More...

void	tune_kernel_static (ICLKernel &kernel)
	Tunes OpenCL kernel. More...

void	enable_job_chaining (int job_chaining_size)
	Enable job chaining. More...

bool	is_initialised () const

Static Public Member Functions
static CLScheduler &	get ()
	Access the scheduler singleton. More...

Detailed Description

Provides global access to a CL context and command queue.

Definition at line 43 of file CLScheduler.h.

Constructor & Destructor Documentation

◆ CLScheduler() [1/2]

CLScheduler ( )

Constructor.

Definition at line 97 of file CLScheduler.cpp.

     : _context(), _queue(), _target(GPUTarget::MIDGARD), _is_initialised(false), _cl_tuner(nullptr), _gemm_heuristics(nullptr), _backend_type(CLBackendType::Native), _job_chaining_enabled(false),
       _job_chaining_size(), _job_chaining_count(0)
 {
 }

◆ CLScheduler() [2/2]

CLScheduler ( const CLScheduler & )

delete

Prevent instances of this class from being copied (As this class contains pointers)

◆ ~CLScheduler()

~CLScheduler ( )

default

Default destructor.

Member Function Documentation

◆ context()

cl::Context & context ( )

Accessor for the associated CL context.

Returns: A CL context.

Definition at line 32 of file CLScheduler.cpp.

References ARM_COMPUTE_ERROR_ON, CLKernelLibrary::context(), and CLKernelLibrary::get().

Referenced by CLTensorAllocator::import_memory(), arm_compute::restore_program_cache_from_file(), and Framework::run().

 {
     ARM_COMPUTE_ERROR_ON(!_is_initialised);
     _context = CLKernelLibrary::get().context();
     return _context;
 }

◆ default_init()

void default_init	(	ICLTuner *	cl_tuner = `nullptr`,
		CLGEMMHeuristicsHandle *	gemm_h = `nullptr`,
		CLBackendType	cl_backend_type = `CLBackendType::Native`
	)

Initialises the context and command queue used by the scheduler to default values and sets a default device and kernel path for the CLKernelLibrary.

Parameters

[in]	cl_tuner	(Optional) Pointer to ICLTuner (default=nullptr)
[in]	gemm_h	(Optional) Pointer to CLGEMMHeuristicsHandle (default = nullptr)
[in]	cl_backend_type	(Optional) Type of backend to use (default = CLBackendType::Native)

Definition at line 122 of file CLScheduler.cpp.

References ARM_COMPUTE_ERROR_ON_MSG, arm_compute::create_opencl_context_and_device(), CLKernelLibrary::get(), CLKernelLibrary::init(), CLScheduler::init(), and CLScheduler::queue().

Referenced by CLScheduler::default_reinit(), CLDeviceBackend::initialize_backend(), and arm_compute::restore_program_cache_from_file().

 {
     if(!_is_initialised)
     {
         cl::Context ctx;
         cl::Device  dev;
         cl_int      err;
         std::tie(ctx, dev, err) = create_opencl_context_and_device(cl_backend_type);
         ARM_COMPUTE_ERROR_ON_MSG(err != CL_SUCCESS, "Failed to create OpenCL context");
         cl::CommandQueue queue = cl::CommandQueue(ctx, dev);
         CLKernelLibrary::get().init("./cl_kernels/", ctx, dev);
         init(ctx, queue, dev, cl_tuner, gemm_h);
     }
 
     // Set CL tuner and GEMM heuristics
     _cl_tuner        = cl_tuner;
     _gemm_heuristics = gemm_h;
 }

◆ default_init_with_context()

void default_init_with_context	(	cl::Device &	device,
		cl::Context &	ctx,
		ICLTuner *	cl_tuner = `nullptr`,
		CLGEMMHeuristicsHandle *	gemm_h = `nullptr`
	)

Initialises the scheduler with context and device provided by the user.

Parameters

[in]	device	OpenCL device to be used
[in]	ctx	OpenCL ctx to be used
[in]	cl_tuner	(Optional) Pointer to ICLTuner (default=nullptr)
[in]	gemm_h	(Optional) Pointer to CLGEMMHeuristicsHandle (default = nullptr)

Definition at line 110 of file CLScheduler.cpp.

References CLKernelLibrary::get(), CLKernelLibrary::init(), CLScheduler::init(), and CLScheduler::queue().

Referenced by main(), and arm_compute::utils::run_example().

 {
     if(!_is_initialised)
     {
         const std::string cl_kernels_folder("./cl_kernels/");
         cl::CommandQueue  queue = cl::CommandQueue(ctx, device);
         CLKernelLibrary::get().init(cl_kernels_folder, ctx, device);
         init(ctx, queue, device, cl_tuner, gemm_h);
         _cl_tuner = cl_tuner;
     }
 }

◆ default_reinit()

void default_reinit	(	ICLTuner *	cl_tuner = `nullptr`,
		CLGEMMHeuristicsHandle *	gemm_h = `nullptr`,
		CLBackendType	cl_backend_type = `CLBackendType::Native`
	)

Re-initializes the context and command queue used by the scheduler to default values and sets a default device and kernel path for the CLKernelLibrary.

Parameters

[in]	cl_tuner	(Optional) Pointer to ICLTuner (default=nullptr)
[in]	gemm_h	(Optional) Pointer to CLGEMMHeuristicsHandle (default = nullptr)
[in]	cl_backend_type	(Optional) Type of backend to use (default = CLBackendType::Native)

Definition at line 141 of file CLScheduler.cpp.

References CLScheduler::default_init().

Referenced by arm_compute::test::validation::TEST_CASE().

 {
     _is_initialised = false;
 
     default_init(cl_tuner, gemm_h, cl_backend_type);
 }

◆ enable_job_chaining()

void enable_job_chaining ( int job_chaining_size )

Enable job chaining.

The command queue will only be flushed when job_chaining_size kernels have been enqueued.

Parameters

[in] job_chaining_size Kernels to enqueue before flushing

Definition at line 216 of file CLScheduler.cpp.

 {
     _job_chaining_enabled = true;
     _job_chaining_size    = job_chaining_size;
 }

◆ enqueue()

void enqueue	(	ICLKernel &	kernel,
		bool	flush = `true`
	)

Schedule the execution of the passed kernel if possible.

Parameters

[in]	kernel	Kernel to execute.
[in]	flush	(Optional) Specifies if the command queue will be flushed after running the kernel. This will be ignored if job chaining is enabled.

Definition at line 205 of file CLScheduler.cpp.

References arm_compute::test::validation::pack.

Referenced by CLSpaceToDepthLayer::run(), CLDeconvolutionLayerUpsample::run(), CLFFT1D::run(), CLStackLayer::run(), CLL2NormalizeLayer::run(), CLNormalizationLayer::run(), CLArgMinMaxLayer::run(), CLPadLayer::run(), CLReductionOperation::run(), CLDepthwiseConvolutionLayer::run(), CLMaxUnpoolingLayer::run(), CLFuseBatchNormalization::run(), CLBatchNormalizationLayer::run(), CLBatchToSpaceLayer::run(), CLSpaceToBatchLayer::run(), CLGEMMDeconvolutionLayer::run(), CLGenerateProposalsLayer::run(), CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::run(), CLQLSTMLayer::run(), ClSynthetizeOperatorWithBorder< K >::run(), and arm_compute::schedule_kernel_on_ctx().

 {
     ITensorPack pack;
     enqueue_common(kernel, pack, flush);
 }

◆ enqueue_op()

void enqueue_op	(	ICLKernel &	kernel,
		ITensorPack &	tensors,
		bool	flush = `true`
	)

Schedule the execution of the passed kernel if possible.

Parameters

[in]	kernel	Kernel to execute.
[in]	tensors	Vector containing the tensors to operate on.
[in]	flush	(Optional) Specifies if the command queue will be flushed after running the kernel. This will be ignored if job chaining is enabled.

Definition at line 211 of file CLScheduler.cpp.

Referenced by ClWorkloadRuntime::configure(), ClGemm::prepare(), ClGemmLowpMatrixMultiplyCore::prepare(), ClWinogradConv2d::prepare(), ClGemmConv2d::prepare(), CLQLSTMLayer::prepare(), ClDequantize::run(), ClQuantize::run(), ICLOperator::run(), ClScale::run(), ClWorkloadRuntime::run(), ClSoftmax::run(), ClConcatenate::run(), ClDirectConv2d::run(), ClDirectConv3d::run(), ClMatMul::run(), ClTransposedConvolution::run(), ClGemmLowpOutputStage::run(), ClIndirectConv2d::run(), ClGemm::run(), ClGemmLowpMatrixMultiplyCore::run(), ClWinogradConv2d::run(), CLSynthetizeOperatorInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::run(), ClGemmConv2d::run(), CLLSTMLayer::run(), and ClSynthetizeOperatorWithBorder< K >::run().

 {
     enqueue_common(kernel, tensors, flush);
 }

◆ enqueue_sync_event()

cl::Event enqueue_sync_event ( )

Enqueues a marker into the associated command queue and return the event.

Returns: An event that can be waited on to block the executing thread.

Definition at line 75 of file CLScheduler.cpp.

 {
     cl::Event event;
     _queue.enqueueMarker(&event);
     return event;
 }

◆ gemm_heuristics()

CLGEMMHeuristicsHandle * gemm_heuristics ( ) const

Accessor for the associated CLGEMMHeuristicsHandle.

Returns: Pointer to CLGEMMHeuristicsHandle

Definition at line 50 of file CLScheduler.cpp.

Referenced by arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_native(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_kernel(), and CLDeviceBackend::setup_backend_context().

 {
     return _gemm_heuristics;
 }

◆ get()

CLScheduler & get ( )

static

Access the scheduler singleton.

This method has been deprecated and will be removed in future releases

Returns: The scheduler

Definition at line 103 of file CLScheduler.cpp.

References arm_compute::opencl_is_available().

Referenced by CLTuner::add_tuning_params(), CLBufferAllocator::allocate(), ClQueue::cl_queue(), CLBufferMemoryRegion::CLBufferMemoryRegion(), ClScale::configure(), ClPool2d::configure(), ClPool3d::configure(), ClWorkloadRuntime::configure(), ClSoftmax::configure(), ClDirectConv2d::configure(), CLPriorBoxLayer::configure(), ClTransposedConvolution::configure(), ClMatMul::configure(), CLRange::configure(), ClIndirectConv2d::configure(), CLFFT1D::configure(), CLDepthwiseConvolutionLayer::configure(), ClGemm::configure(), ClGemmLowpMatrixMultiplyCore::configure(), CLSynthetizeOperatorInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::configure(), ClWinogradConv2d::configure(), CLCropResize::configure(), ClGemmConv2d::configure(), ClConv2d::configure(), CLConvolutionLayer::configure(), CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::configure(), arm_compute::test::validation::DATA_TEST_CASE(), ClQueue::finish(), CLTensorAllocator::import_memory(), CLDeviceBackend::initialize_backend(), main(), CLArray< cl_int >::map(), CLSubTensor::map(), CLTensor::map(), OpenCLClock< output_timestamps >::OpenCLClock(), ClGemm::prepare(), ClGemmLowpMatrixMultiplyCore::prepare(), ClWinogradConv2d::prepare(), ClGemmConv2d::prepare(), CLFFTConvolutionLayer::prepare(), CLQLSTMLayer::prepare(), arm_compute::restore_program_cache_from_file(), ClDequantize::run(), ClQuantize::run(), CLSplit::run(), ICLOperator::run(), ClScale::run(), ClWorkloadRuntime::run(), ClSoftmax::run(), ClConcatenate::run(), ClDirectConv2d::run(), ClDirectConv3d::run(), ClMatMul::run(), ClTransposedConvolution::run(), ClGemmLowpOutputStage::run(), CLSpaceToDepthLayer::run(), ClIndirectConv2d::run(), CLDeconvolutionLayerUpsample::run(), CLFFT1D::run(), ClGemm::run(), CLStackLayer::run(), ClGemmLowpMatrixMultiplyCore::run(), CLL2NormalizeLayer::run(), CLNormalizationLayer::run(), CLArgMinMaxLayer::run(), CLPadLayer::run(), CLReductionOperation::run(), CLSynthetizeOperatorInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::run(), ClWinogradConv2d::run(), CLDepthwiseConvolutionLayer::run(), CLMaxUnpoolingLayer::run(), ClGemmConv2d::run(), CLFuseBatchNormalization::run(), CLCropResize::run(), CLBatchNormalizationLayer::run(), CLBatchToSpaceLayer::run(), CLSpaceToBatchLayer::run(), CLGEMMDeconvolutionLayer::run(), CLGenerateProposalsLayer::run(), CLSynthetizeFunctionInitOutputWithZeroAndWithZeroConstantBorder< K, bordersize >::run(), CLLSTMLayer::run(), CLQLSTMLayer::run(), ClSynthetizeOperatorWithBorder< K >::run(), Framework::run(), arm_compute::utils::run_example(), arm_compute::save_program_cache_to_file(), arm_compute::schedule_kernel_on_ctx(), ClQueue::scheduler(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_native(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(), arm_compute::cl_gemm::auto_heuristics::select_mlgo_gemm_kernel(), ClContext::set_cl_ctx(), ClQueue::set_cl_queue(), CLTensorAllocator::set_global_allocator(), CLDeviceBackend::setup_backend_context(), CLDeviceBackend::sync(), arm_compute::test::sync_if_necessary(), arm_compute::test::validation::TEST_CASE(), OpenCLClock< output_timestamps >::test_measurements(), CLArray< cl_int >::unmap(), CLSubTensor::unmap(), CLTensor::unmap(), CLBufferMemoryRegion::unmap(), ClMatMul::validate(), ClGemm::validate(), ClGemmLowpMatrixMultiplyCore::validate(), CLDepthwiseConvolutionLayer::validate(), ClConv2d::validate(), CLGenerateProposalsLayer::validate(), and CLConvolutionLayer::validate().

 {
     std::call_once(_initialize_symbols, opencl_is_available);
     static CLScheduler scheduler;
     return scheduler;
 }

◆ init()

void init	(	cl::Context	context,
		cl::CommandQueue	queue,
		const cl::Device &	device,
		ICLTuner *	cl_tuner = `nullptr`,
		CLGEMMHeuristicsHandle *	gemm_h = `nullptr`,
		CLBackendType	cl_backend_type = `CLBackendType::Native`
	)

Initialises the context and command queue to be used by the scheduler.

Parameters

[in]	context	A CL context.
[in]	queue	A CL command queue.
[in]	device	A CL device.
[in]	cl_tuner	(Optional) Pointer to OpenCL tuner (default=nullptr) Note: It is caller's responsibility to release the allocated memory for CLTuner
[in]	gemm_h	(Optional) Pointer to CLGEMMHeuristicsHandle (default = nullptr)
[in]	cl_backend_type	(Optional) Type of backend to use (default = CLBackendType::Native)

Definition at line 154 of file CLScheduler.cpp.

References ARM_COMPUTE_ERROR_ON_MSG, ITensorPack::empty(), arm_compute::get_target_from_device(), ICLKernel::run(), ICLKernel::run_op(), CLScheduler::set_context(), ICLTuner::tune_kernel_dynamic(), and IKernel::window().

Referenced by CLScheduler::default_init(), and CLScheduler::default_init_with_context().

 {
     set_context(std::move(context));
     _queue           = std::move(queue);
     _target          = get_target_from_device(device);
     _is_initialised  = true;
     _cl_tuner        = cl_tuner;
     _gemm_heuristics = gemm_h;
     _backend_type    = cl_backend_type;
 }

◆ is_initialised()

bool is_initialised ( ) const

Definition at line 90 of file CLScheduler.cpp.

Referenced by arm_compute::restore_program_cache_from_file().

 {
     return _is_initialised;
 }

◆ operator=()

CLScheduler& operator= ( const CLScheduler & )

delete

Prevent instances of this class from being copied (As this class contains pointers)

◆ queue()

cl::CommandQueue & queue ( )

Accessor for the associated CL command queue.

Returns: A CL command queue.

Definition at line 39 of file CLScheduler.cpp.

References ARM_COMPUTE_ERROR_ON.

Referenced by CLTuner::add_tuning_params(), ClQueue::cl_queue(), arm_compute::test::validation::DATA_TEST_CASE(), CLScheduler::default_init(), CLScheduler::default_init_with_context(), ClQueue::finish(), CLTensor::map(), OpenCLClock< output_timestamps >::OpenCLClock(), ClGemmLowpMatrixMultiplyCore::prepare(), ClWinogradConv2d::prepare(), CLFFTConvolutionLayer::prepare(), CLQLSTMLayer::prepare(), CLSplit::run(), Framework::run(), CLTensorAllocator::set_global_allocator(), arm_compute::test::validation::TEST_CASE(), OpenCLClock< output_timestamps >::test_measurements(), and CLTensor::unmap().

 {
     ARM_COMPUTE_ERROR_ON(!_is_initialised);
     return _queue;
 }

◆ set_context()

void set_context ( cl::Context context )

Accessor to set the CL context to be used by the scheduler.

Parameters

[in] context A CL context.

Definition at line 148 of file CLScheduler.cpp.

References CLKernelLibrary::get(), and CLKernelLibrary::set_context().

Referenced by CLScheduler::init(), Framework::run(), and ClContext::set_cl_ctx().

 {
     _context = std::move(context);
     CLKernelLibrary::get().set_context(_context);
 }

◆ set_queue()

void set_queue ( cl::CommandQueue queue )

Accessor to set the CL command queue to be used by the scheduler.

Parameters

[in] queue A CL command queue.

Definition at line 55 of file CLScheduler.cpp.

Referenced by OpenCLClock< output_timestamps >::OpenCLClock(), Framework::run(), and ClQueue::set_cl_queue().

 {
     _queue = std::move(queue);
 }

◆ set_target()

void set_target ( GPUTarget target )

Accessor to set target GPU to be used by the scheduler.

Parameters

[in] target The target GPU.

Definition at line 60 of file CLScheduler.cpp.

References CLScheduler::target().

 {
     _target = target;
 }

◆ set_tuner()

void set_tuner ( ICLTuner * tuner )

Accessor to set the CL tuner to be used by the scheduler.

Parameters

[in] tuner A CL tuner

Definition at line 65 of file CLScheduler.cpp.

 {
     _cl_tuner = tuner;
 }

◆ sync()

void sync ( )

Blocks until all commands in the associated command queue have finished.

Definition at line 70 of file CLScheduler.cpp.

Referenced by CLCropResize::configure(), main(), CLCropResize::run(), CLDeviceBackend::sync(), arm_compute::test::sync_if_necessary(), and arm_compute::test::validation::TEST_CASE().

 {
     _queue.finish();
 }

◆ target()

GPUTarget target ( ) const

Get the target GPU.

Returns: The target GPU.

Definition at line 45 of file CLScheduler.cpp.

Referenced by ClMatMul::configure(), CLDepthwiseConvolutionLayer::configure(), ClGemm::configure(), ClGemmLowpMatrixMultiplyCore::configure(), CLConvolutionLayer::configure(), CLScheduler::set_target(), ClMatMul::validate(), ClGemm::validate(), ClGemmLowpMatrixMultiplyCore::validate(), CLDepthwiseConvolutionLayer::validate(), ClConv2d::validate(), and CLConvolutionLayer::validate().

 {
     return _target;
 }

◆ tune_kernel_static()

void tune_kernel_static ( ICLKernel & kernel )

Tunes OpenCL kernel.

Parameters

[in] kernel Kernel to tune

Definition at line 82 of file CLScheduler.cpp.

References ICLTuner::tune_kernel_static().

Referenced by ClScale::configure(), ClPool2d::configure(), ClPool3d::configure(), ClDirectConv2d::configure(), CLRange::configure(), and ClGemmConv2d::configure().

 {
     if(_cl_tuner != nullptr)
     {
         _cl_tuner->tune_kernel_static(kernel);
     }
 }

The documentation for this class was generated from the following files:

arm_compute/runtime/CL/CLScheduler.h
src/runtime/CL/CLScheduler.cpp

Public Member Functions

Static Public Member Functions

Detailed Description

Constructor & Destructor Documentation

◆ CLScheduler() [1/2]

◆ CLScheduler() [2/2]

◆ ~CLScheduler()

Member Function Documentation

◆ context()

◆ default_init()

◆ default_init_with_context()

◆ default_reinit()

◆ enable_job_chaining()

◆ enqueue()

◆ enqueue_op()

◆ enqueue_sync_event()

◆ gemm_heuristics()

◆ get()

◆ init()

◆ is_initialised()

◆ operator=()

◆ queue()

◆ set_context()

◆ set_queue()

◆ set_target()

◆ set_tuner()

◆ sync()

◆ target()

◆ tune_kernel_static()