Skip to content

Latest commit

 

History

History
294 lines (186 loc) · 14.1 KB

api.md

File metadata and controls

294 lines (186 loc) · 14.1 KB
Raw

CLCudaAPI: API reference

This file describes the high-level API for both the CUDA and OpenCL back-end of the CLCudaAPI headers. On top of the described API, each class has a constructor which takes the regular OpenCL or CUDA data-type and transforms it into a CLCudaAPI class. Furthermore, each class also implements a () operator which returns the regular OpenCL or CUDA data-type.

CLCudaAPI::Event

Constructor(s):

  • Event(): Creates a new event, to be used for example when timing kernels.

Public method(s):

  • void WaitForCompletion() const: Waits for completion of an event (OpenCL) or does nothing (CUDA).

  • float GetElapsedTime() const: Retrieves the elapsed time in milliseconds of the last recorded event (e.g. a device kernel). This method first makes sure that the last event is finished before computing the elapsed time.

CLCudaAPI::Platform

Constructor(s):

  • Platform(const size_t platform_id): When using the OpenCL back-end, this initializes a new OpenCL platform (e.g. AMD SDK, Intel SDK, NVIDIA SDK) specified by the integer platform_id. When using the CUDA back-end, this initializes the CUDA driver API. The platform_id argument is ignored: there is only one platform.

Public method(s):

  • std::string Name() const: Retrieves the name of the platform.

  • std::string Vendor() const: Retrieves the name of the vendor of the platform.

  • std::string Version() const: Retrieves which version of an OpenCL platform is used (OpenCL back-end) or which CUDA driver is used (CUDA back-end).

  • size_t NumDevices() const: Retrieves the number of devices on this platform.

Non-member function(s):

  • std::vector<Platform> GetAllPlatforms(): Retrieves a vector containing all available platforms.

CLCudaAPI::Device

Constructor(s):

  • Device(const Platform &platform, const size_t device_id): Initializes a new OpenCL or CUDA device on the specified platform. The device_id defines which device should be selected.

Public method(s):

  • RawPlatformID PlatformID() const: Retrieves the raw cl_platform_id ID of the platform used (OpenCL back-end) or a 0 size_t in case of the CUDA back-end.

  • std::string Version() const: Retrieves which version of the OpenCL standard is supported (OpenCL back-end) or which CUDA driver is used (CUDA back-end).

  • size_t VersionNumber() const: The same as the Version() method, but without text, just the numeric value.

  • std::string Vendor() const: Retrieves the name of the vendor of the device.

  • std::string Name() const: Retrieves the name of the device.

  • std::string Type() const: Retrieves the type of the devices. Possible return values are 'CPU', 'GPU', 'accelerator', or 'default'.

  • size_t MaxWorkGroupSize() const: Retrieves the maximum total number of threads in an OpenCL work-group or CUDA thread-block.

  • size_t MaxWorkItemDimensions() const: Retrieves the maximum number of dimensions (e.g. 2D or 3D) in an OpenCL work-group or CUDA thread-block.

  • unsigned long LocalMemSize() const: Retrieves the maximum amount of on-chip scratchpad memory ('local memory') available to a single OpenCL work-group or CUDA thread-block.

  • std::string Capabilities() const: In case of the OpenCL back-end, this returns a list of the OpenCL extensions supported. For CUDA, this returns the device capability (e.g. SM 3.5).

  • bool HasExtension(const std::string &extension) const: In case of the OpenCL back-end, queries whether a certain extension is present (as reported by Capabilities()). For CUDA, this always returns false.

  • bool SupportsFP64() const: Returns whether or not double-precision floating-point 64-bit is supported by the device.

  • bool SupportsFP16() const: Returns whether or not half-precision floating-point 16-bit is supported by the device.

  • size_t CoreClock() const: Retrieves the device's core clock frequency in MHz.

  • size_t ComputeUnits() const: Retrieves the number of compute units (OpenCL terminology) or multi-processors (CUDA terminology) in the device.

  • unsigned long MemorySize() const: Retrieves the total global memory size.

  • unsigned long MaxAllocSize() const: Retrieves the maximum amount of allocatable global memory per allocation.

  • size_t MemoryClock() const: Retrieves the device's memory clock frequency in MHz (CUDA back-end) or 0 (OpenCL back-end).

  • size_t MemoryBusWidth() const: Retrieves the device's memory bus-width in bits (CUDA back-end) or 0 (OpenCL back-end).

  • bool IsLocalMemoryValid(const size_t local_mem_usage) const: Given a requested amount of local on-chip scratchpad memory, this method returns whether or not this is a valid configuration for this particular device.

  • bool IsThreadConfigValid(const std::vector<size_t> &local) const: Given a requested OpenCL work-group or CUDA thread-block configuration local, this method returns whether or not this is a valid configuration for this particular device.

  • bool IsCPU() const: Determines whether this device is of the CPU type.

  • bool IsGPU() const: Determines whether this device is of the GPU type.

  • bool IsAMD() const: Determines whether this device is of the AMD brand.

  • bool IsNVIDIA() const: Determines whether this device is of the NVIDIA brand.

  • bool IsIntel() const: Determines whether this device is of the Intel brand.

  • bool IsARM() const: Determines whether this device is of the ARM brand.

  • std::string AMDBoardName() const: Returns the value of CL_DEVICE_BOARD_NAME_AMD if present. For the CUDA back-end, this always returns an empty string.

  • std::string NVIDIAComputeCapability() const: Returns the compute capability of an NVIDIA GPU, e.g. SM3.5. For the CUDA back-end, this returns the same as the Capabilities() method.

CLCudaAPI::Context

Constructor(s):

  • Context(const Device &device): Initializes a new context on a given device. On top of this context, CLCudaAPI can create new programs, queues and buffers.

CLCudaAPI::Program

Constructor(s):

  • Program(const Context &context, std::string source): Creates a new OpenCL or CUDA program on a given context. A program is a collection of one or more device kernels which form a single compilation unit together. The device-code is passed as a string. Such a string can for example be generated, hard-coded, or read from file at run-time. If passed as an r-value (e.g. using std::move), the device-code string is moved instead of copied into the class' member variable.

  • Program(const Device &device, const Context &context, const std::string& binary): As above, but now the program is constructed based on an already compiled IR or binary of the device kernels. This requires a context corresponding to the binary. This constructor for OpenCL is based on the clCreateProgramWithBinary function.

Public method(s):

  • void Build(const Device &device, std::vector<std::string> &options): This method invokes the OpenCL or CUDA compiler to build the program at run-time for a specific target device. Depending on the back-end, specific options can be passed to the compiler in the form of the options vector. Compilation errors generated by the run-time compiler result in an std::runtime_error exception, which can be caught.

  • std::string GetBuildInfo(const Device &device) const: Retrieves all compiler warnings and errors generated by the build process.

  • std::string GetIR() const: Retrieves the intermediate representation (IR) of the compiled program. When using the CUDA back-end, this returns the PTX-code. For the OpenCL back-end, this returns either an IR (e.g. PTX) or a binary. This is different per OpenCL implementation.

CLCudaAPI::Queue

Constructor(s):

  • Queue(const Context &context, const Device &device): Creates a new queue to enqueue kernel launches and device memory operations. This is analogous to an OpenCL command queue or a CUDA stream.

Public method(s):

  • void Finish(Event &event) const and void Finish() const: Completes all tasks in the queue. In the case of the CUDA back-end, the first form additionally synchronizes on the specified event.

  • Context GetContext() const: Retrieves the CUDA/OpenCL context associated with this queue.

  • Device GetDevice() const: Retrieves the CUDA/OpenCL device associated with this queue.

template <typename T> CLCudaAPI::BufferHost

Constructor(s):

  • BufferHost(const Context &, const size_t size): Initializes a new linear 1D memory buffer on the host of type T. This buffer is allocated with a fixed number of elements given by size. Note that the buffer's elements are not initialized. In the case of the CUDA back-end, this host buffer is implemented as page-locked memory. The OpenCL back-end uses a regular std::vector container.

Public method(s):

  • size_t GetSize() const: Retrieves the allocated size in bytes.

  • Several std::vector methods: Adds some compatibility with std::vector by implementing the size, begin, end, operator[], and data methods.

template <typename T> CLCudaAPI::Buffer

Constants(s):

  • enum class BufferAccess { kReadOnly, kWriteOnly, kReadWrite, kNotOwned } Defines the different access types for the buffers. Writing to a read-only buffer will throw an error, as will reading from a write-only buffer. A buffer which is of type kNotOwned will not be automatically freed afterwards.

Constructor(s):

  • Buffer(const Context &context, const BufferAccess access, const size_t size): Initializes a new linear 1D memory buffer on the device of type T. This buffer is allocated with a fixed number of elements given by size. Note that the buffer's elements are not initialized. The buffer can be read-only, write-only, read-write, or not-owned as specified by the access argument.

  • Buffer(const Context &context, const size_t size): As above, but now defaults to read-write access.

  • template <typename Iterator> Buffer(const Context &context, const Queue &queue, Iterator start, Iterator end): Creates a new buffer based on data in a linear C++ container (such as std::vector). The size is determined by the difference between the end and start iterators. This method both creates a new buffer and writes data to it. It synchronises the queue before returning.

Public method(s):

  • void ReadAsync(const Queue &queue, const size_t size, T* host) const and void ReadAsync(const Queue &queue, const size_t size, std::vector<T> &host) and void ReadAsync(const Queue &queue, const size_t size, BufferHost<T> &host): Copies size elements from the current device buffer to the target host buffer. The host buffer has to be pre-allocated with a size of at least size elements. This method is a-synchronous: it can return before the copy operation is completed.

  • void Read(const Queue &queue, const size_t size, T* host) const and void Read(const Queue &queue, const size_t size, std::vector<T> &host) and void Read(const Queue &queue, const size_t size, BufferHost<T> &host): As above, but now completes the operation before returning.

  • void WriteAsync(const Queue &queue, const size_t size, const T* host) and void WriteAsync(const Queue &queue, const size_t size, const std::vector<T> &host) and void WriteAsync(const Queue &queue, const size_t size, const BufferHost<T> &host): Copies size elements from a host buffer to the current device buffer. The device buffer has to be pre-allocated with a size of at least size elements. This method is a-synchronous: it can return before the copy operation is completed.

  • void Write(const Queue &queue, const size_t size, const T* host) and void Write(const Queue &queue, const size_t size, const std::vector<T> &host) and void Write(const Queue &queue, const size_t size, const BufferHost<T> &host): As above, but now completes the operation before returning.

  • void CopyToAsync(const Queue &queue, const size_t size, const Buffer<T> &destination) const: Copies size elements from the current device buffer to another device buffer given by destination. The destination buffer has to be pre-allocated with a size of at least size elements. This method is a-synchronous: it can return before the copy operation is completed.

  • void CopyTo(const Queue &queue, const size_t size, const Buffer<T> &destination) const: As above, but now completes the operation before returning.

  • size_t GetSize() const: Retrieves the allocated size in bytes.

CLCudaAPI::Kernel

Constructor(s):

  • Kernel(const Program &program, const std::string &name): Retrieves a new kernel from a compiled program. The kernel name is given as the string name.

Public method(s):

  • template <typename T> void SetArgument(const size_t index, const T &value): Method to set a kernel argument (l-value or r-value). The argument index specifies the position in the list of kernel arguments. The argument value can also be a CLCudaAPI::Buffer.

  • template <typename... Args> void SetArguments(Args&... args): As above, but now sets all arguments in one go, starting at index 0. This overwrites any previous arguments (if any). The parameter pack args takes any number of arguments of different types, including CLCudaAPI::Buffer.

  • unsigned long LocalMemUsage(const Device &device) const: Retrieves the amount of on-chip scratchpad memory (local memory in OpenCL, shared memory in CUDA) required by this specific kernel.

  • std::string GetFunctionName() const : Retrieves the name of the kernel (OpenCL only).

  • Launch(const Queue &queue, const std::vector<size_t> &global, const std::vector<size_t> &local, Event &event): Launches a kernel onto the specified queue. This kernel launch is a-synchronous: this method can return before the device kernel is completed. The total number of threads launched is equal to the global vector; the number of threads per OpenCL work-group or CUDA thread-block is given by the local vector. The elapsed time is recorded into the event argument.

  • Launch(const Queue &queue, const std::vector<size_t> &global, const std::vector<size_t> &local, Event &event, std::vector<Event>& waitForEvents): As above, but now this kernel is only launched after the other specified events have finished (OpenCL only). If local is empty, the kernel-size is determined automatically (OpenCL only).