sgc-ruby-cuda 0.1.0

data/lib/cuda/runtime/error.rb

@@ -0,0 +1,197 @@
+#
+# Copyright (c) 2010-2011 Chung Shin Yee
+#
+#       shinyee@speedgocomputing.com
+#       http://www.speedgocomputing.com
+#       http://github.com/xman/sgc-ruby-cuda
+#       http://rubyforge.org/projects/rubycuda
+#
+# This file is part of SGC-Ruby-CUDA.
+#
+# SGC-Ruby-CUDA is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# SGC-Ruby-CUDA is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with SGC-Ruby-CUDA.  If not, see <http://www.gnu.org/licenses/>.
+#
+
+require 'cuda/runtime/ffi-cuda'
+
+
+module SGC
+module Cuda
+
+    module Error
+
+        class CudaStandardError < RuntimeError; end
+        class CudaMissingConfigurationError < CudaStandardError; end
+        class CudaMemoryAllocationError < CudaStandardError; end
+        class CudaInitializationError < CudaStandardError; end
+        class CudaLaunchFailureError < CudaStandardError; end
+        class CudaPriorLaunchFailureError < CudaStandardError; end # @deprecated
+        class CudaLaunchTimeoutError < CudaStandardError; end
+        class CudaLaunchOutOfResourcesError < CudaStandardError; end
+        class CudaInvalidDeviceFunctionError < CudaStandardError; end
+        class CudaInvalidConfigurationError < CudaStandardError; end
+        class CudaInvalidDeviceError < CudaStandardError; end
+        class CudaInvalidValueError < CudaStandardError; end
+        class CudaInvalidPitchValueError < CudaStandardError; end
+        class CudaInvalidSymbolError < CudaStandardError; end
+        class CudaMapBufferObjectFailedError < CudaStandardError; end
+        class CudaUnmapBufferObjectFailedError < CudaStandardError; end
+        class CudaInvalidHostPointerError < CudaStandardError; end
+        class CudaInvalidDevicePointerError < CudaStandardError; end
+        class CudaInvalidTextureError < CudaStandardError; end
+        class CudaInvalidTextureBindingError < CudaStandardError; end
+        class CudaInvalidChannelDescriptorError < CudaStandardError; end
+        class CudaInvalidMemcpyDirectionError < CudaStandardError; end
+        class CudaAddressOfConstantError < CudaStandardError; end # @deprecated
+        class CudaTextureFetchFailedError  < CudaStandardError; end  # @deprecated
+        class CudaTextureNotFoundError < CudaStandardError; end # @deprecated
+        class CudaSynchronizationError < CudaStandardError; end # @deprecated
+        class CudaInvalidFilterSettingError < CudaStandardError; end
+        class CudaInvalidNormSettingError < CudaStandardError; end
+        class CudaMixedDeviceExecutionError < CudaStandardError; end # @deprecated
+        class CudaCudartUnloadingError < CudaStandardError; end # @deprecated
+        class CudaUnknownError < CudaStandardError; end
+        class CudaNotYetImplementedError < CudaStandardError; end
+        class CudaMemoryValueTooLargeError < CudaStandardError; end # @deprecated
+        class CudaInvalidResourceHandleError < CudaStandardError; end
+        class CudaNotReadyError < CudaStandardError; end
+        class CudaInsufficientDriverError < CudaStandardError; end
+        class CudaSetOnActiveProcessError < CudaStandardError; end
+        class CudaInvalidSurfaceError < CudaStandardError; end
+        class CudaNoDeviceError < CudaStandardError; end
+        class CudaECCUncorrectableError < CudaStandardError; end
+        class CudaSharedObjectSymbolNotFoundError < CudaStandardError; end
+        class CudaSharedObjectInitFailedError < CudaStandardError; end
+        class CudaUnsupportedLimitError < CudaStandardError; end
+        class CudaDuplicateVariableNameError < CudaStandardError; end
+        class CudaDuplicateTextureNameError < CudaStandardError; end
+        class CudaDuplicateSurfaceNameError < CudaStandardError; end
+        class CudaDevicesUnavailableError < CudaStandardError; end
+        class CudaInvalidKernelImageError < CudaStandardError; end
+        class CudaNoKernelImageForDeviceError < CudaStandardError; end
+        class CudaIncompatibleDriverContextError < CudaStandardError; end
+        class CudaPeerAccessAlreadyEnabledError < CudaStandardError; end
+        class CudaPeerAccessNotEnabledError < CudaStandardError; end
+        class CudaDeviceAlreadyInUseError < CudaStandardError; end
+        class CudaProfilerDisabledError < CudaStandardError; end
+        class CudaProfilerNotInitializedError < CudaStandardError; end
+        class CudaProfilerAlreadyStartedError < CudaStandardError; end
+        class CudaProfilerAlreadyStoppedError < CudaStandardError; end
+        class CudaStartupFailureError < CudaStandardError; end
+        class CudaAPIFailureBaseError < CudaStandardError; end
+
+    end
+
+
+    # @param [Integer, CudaError] e A CUDA error value or label.
+    # @return [String] The error string of _e_.
+    def get_error_string(e)
+        API::cudaGetErrorString(e)
+    end
+    module_function :get_error_string
+
+
+    # @return [Integer] The error value of the last CUDA error.
+    def get_last_error
+        API::cudaGetLastError
+    end
+    module_function :get_last_error
+
+
+    # Return the last CUDA error, but do not reset the error.
+    # @return [Integer] The error value of the last CUDA error.
+    def peek_at_last_error
+        API::cudaPeekAtLastError
+    end
+    module_function :peek_at_last_error
+
+    # @private
+    module Pvt
+
+        include Error
+
+
+        def self.handle_error(status, msg = nil)
+            status == CUDA_SUCCESS or raise @error_class_by_enum[API::CudaError[status]], API::cudaGetErrorString(status) + " : #{msg}"
+            nil
+        end
+
+
+        CUDA_SUCCESS = API::CudaError[:SUCCESS]
+        CUDA_ERROR_NOT_READY = API::CudaError[:ERROR_NOT_READY]
+
+        @error_class_by_enum = {
+            ERROR_MISSING_CONFIGURATION: CudaMissingConfigurationError,
+            ERROR_MEMORY_ALLOCATION: CudaMemoryAllocationError,
+            ERROR_INITIALIZATION_ERROR: CudaInitializationError,
+            ERROR_LAUNCH_FAILURE: CudaLaunchFailureError,
+            ERROR_PRIOR_LAUNCH_FAILURE: CudaPriorLaunchFailureError,
+            ERROR_LAUNCH_TIMEOUT: CudaLaunchTimeoutError,
+            ERROR_LAUNCH_OUT_OF_RESOURCES: CudaLaunchOutOfResourcesError,
+            ERROR_INVALID_DEVICE_FUNCTION: CudaInvalidDeviceFunctionError,
+            ERROR_INVALID_CONFIGURATION: CudaInvalidConfigurationError,
+            ERROR_INVALID_DEVICE: CudaInvalidDeviceError,
+            ERROR_INVALID_VALUE: CudaInvalidValueError,
+            ERROR_INVALID_PITCH_VALUE: CudaInvalidPitchValueError,
+            ERROR_INVALID_SYMBOL: CudaInvalidSymbolError,
+            ERROR_MAP_BUFFER_OBJECT_FAILED: CudaMapBufferObjectFailedError,
+            ERROR_UNMAP_BUFFER_OBJECT_FAILED: CudaUnmapBufferObjectFailedError,
+            ERROR_INVALID_HOST_POINTER: CudaInvalidHostPointerError,
+            ERROR_INVALID_DEVICE_POINTER: CudaInvalidDevicePointerError,
+            ERROR_INVALID_TEXTURE: CudaInvalidTextureError,
+            ERROR_INVALID_TEXTURE_BINDING: CudaInvalidTextureBindingError,
+            ERROR_INVALID_CHANNEL_DESCRIPTOR: CudaInvalidChannelDescriptorError,
+            ERROR_INVALID_MEMCPY_DIRECTION: CudaInvalidMemcpyDirectionError,
+            ERROR_ADDRESS_OF_CONSTANT: CudaAddressOfConstantError,
+            ERROR_TEXTURE_FETCH_FAILED: CudaTextureFetchFailedError,
+            ERROR_TEXTURE_NOT_BOUND: CudaTextureNotFoundError,
+            ERROR_SYNCHRONIZATION_ERROR: CudaSynchronizationError,
+            ERROR_INVALID_FILTER_SETTING: CudaInvalidFilterSettingError,
+            ERROR_INVALID_NORM_SETTING: CudaInvalidNormSettingError,
+            ERROR_MIXED_DEVICE_EXECUTION: CudaMixedDeviceExecutionError,
+            ERROR_CUDART_UNLOADING: CudaCudartUnloadingError,
+            ERROR_UNKNOWN: CudaUnknownError,
+            ERROR_NOT_YET_IMPLEMENTED: CudaNotYetImplementedError,
+            ERROR_MEMORY_VALUE_TOO_LARGE: CudaMemoryValueTooLargeError,
+            ERROR_INVALID_RESOURCE_HANDLE: CudaInvalidResourceHandleError,
+            ERROR_NOT_READY: CudaNotReadyError,
+            ERROR_INSUFFICIENT_DRIVER: CudaInsufficientDriverError,
+            ERROR_SET_ON_ACTIVE_PROCESS: CudaSetOnActiveProcessError,
+            ERROR_INVALID_SURFACE: CudaInvalidSurfaceError,
+            ERROR_NO_DEVICE: CudaNoDeviceError,
+            ERROR_ECC_UNCORRECTABLE: CudaECCUncorrectableError,
+            ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND: CudaSharedObjectSymbolNotFoundError,
+            ERROR_SHARED_OBJECT_INIT_FAILED: CudaSharedObjectInitFailedError,
+            ERROR_UNSUPPORTED_LIMIT: CudaUnsupportedLimitError,
+            ERROR_DUPLICATE_VARIABLE_NAME: CudaDuplicateVariableNameError,
+            ERROR_DUPLICATE_TEXTURE_NAME: CudaDuplicateTextureNameError,
+            ERROR_DUPLICATE_SURFACE_NAME: CudaDuplicateSurfaceNameError,
+            ERROR_DEVICES_UNAVAILABLE: CudaDevicesUnavailableError,
+            ERROR_INVALID_KERNEL_IMAGE: CudaInvalidKernelImageError,
+            ERROR_NO_KERNEL_IMAGE_FOR_DEVICE: CudaNoKernelImageForDeviceError,
+            ERROR_INCOMPATIBLE_DRIVER_CONTEXT: CudaIncompatibleDriverContextError,
+            ERROR_PEER_ACCESS_ALREADY_ENABLED: CudaPeerAccessAlreadyEnabledError,
+            ERROR_PEER_ACCESS_NOT_ENABLED: CudaPeerAccessNotEnabledError,
+            ERROR_DEVICE_ALREADY_IN_USE: CudaDeviceAlreadyInUseError,
+            ERROR_PROFILER_DISABLED: CudaProfilerDisabledError,
+            ERROR_PROFILER_NOT_INITIALIZED: CudaProfilerNotInitializedError,
+            ERROR_PROFILER_ALREADY_STARTED: CudaProfilerAlreadyStartedError,
+            ERROR_PROFILER_ALREADY_STOPPED: CudaProfilerAlreadyStoppedError,
+            ERROR_STARTUP_FAILURE: CudaStartupFailureError,
+            ERROR_API_FAILURE_BASE: CudaAPIFailureBaseError,
+        }
+
+    end
+
+end # module
+end # module

data/lib/cuda/runtime/event.rb

@@ -0,0 +1,117 @@
+#
+# Copyright (c) 2010-2011 Chung Shin Yee
+#
+#       shinyee@speedgocomputing.com
+#       http://www.speedgocomputing.com
+#       http://github.com/xman/sgc-ruby-cuda
+#       http://rubyforge.org/projects/rubycuda
+#
+# This file is part of SGC-Ruby-CUDA.
+#
+# SGC-Ruby-CUDA is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# SGC-Ruby-CUDA is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with SGC-Ruby-CUDA.  If not, see <http://www.gnu.org/licenses/>.
+#
+
+require 'cuda/runtime/ffi-cuda'
+require 'cuda/runtime/error'
+
+
+module SGC
+module Cuda
+
+class CudaEvent
+
+    # Create and return an event with _flags_.
+    # @overload create
+    # @overload create(flags)
+    # @return [CudaEvent] An event created with _flags_.
+    def self.create(*flags)
+        flags.empty? == false or flags = :DEFAULT
+        p = FFI::MemoryPointer.new(:CudaEvent)
+        f = CudaEventFlags.value(flags)
+        status = API::cudaEventCreateWithFlags(p, f)
+        Pvt::handle_error(status, "Failed to create event: flags = #{flags}")
+        new(p)
+    end
+
+
+    # Destroy this event.
+    def destroy
+        status = API::cudaEventDestroy(self.to_api)
+        Pvt::handle_error(status, "Failed to destroy event.")
+        API::write_cudaevent(@pevent, 0)
+        nil
+    end
+
+
+    # @return [Boolean] Return true if this event has been recorded. Otherwise, return false.
+    def query
+        status = API::cudaEventQuery(self.to_api)
+        if status == Pvt::CUDA_SUCCESS
+            return true
+        elsif status == Pvt::CUDA_ERROR_NOT_READY
+            return false
+        end
+        Pvt::handle_error(status, "Failed to query event.")
+        raise CudaStandardError, "Error handling fails to catch this error."
+    end
+
+
+    # Record this event asynchronously on _stream_.
+    # @param [Integer, CudaStream] stream The CUDA stream to record this event on.
+    #     Setting _stream_ on anything other than an instance of CudaStream will record on any stream.
+    # @return [CudaEvent] This event.
+    def record(stream = 0)
+        s = Pvt::parse_stream(stream)
+        status = API::cudaEventRecord(self.to_api, s)
+        Pvt::handle_error(status, "Failed to record event.")
+        self
+    end
+
+
+    # Block the calling CPU thread until this event has been recorded.
+    # @return [CudaEvent] This event.
+    def synchronize
+        status = API::cudaEventSynchronize(self.to_api)
+        Pvt::handle_error(status)
+        self
+    end
+
+
+    # Compute the elapsed time (ms) from _event_start_ to _event_end_.
+    # @param [CudaEvent] event_start The event corresponds to the start time.
+    # @param [CudaEvent] event_end The event corresponds to the end time.
+    # @return [Numeric] The elapsed time in ms.
+    def self.elapsed_time(event_start, event_end)
+        t = FFI::MemoryPointer.new(:float)
+        API::cudaEventElapsedTime(t, event_start.to_api, event_end.to_api)
+        t.read_float
+    end
+
+
+    # @private
+    def initialize(ptr)
+        @pevent = ptr
+    end
+    private_class_method :new
+
+
+    # @private
+    def to_api
+        API::read_cudaevent(@pevent)
+    end
+
+end
+
+end # module
+end # module

data/lib/cuda/runtime/ffi-cuda.rb

@@ -0,0 +1,588 @@
+#
+# Copyright (c) 2010-2011 Chung Shin Yee
+#
+#       shinyee@speedgocomputing.com
+#       http://www.speedgocomputing.com
+#       http://github.com/xman/sgc-ruby-cuda
+#       http://rubyforge.org/projects/rubycuda
+#
+# This file is part of SGC-Ruby-CUDA.
+#
+# SGC-Ruby-CUDA is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# SGC-Ruby-CUDA is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with SGC-Ruby-CUDA.  If not, see <http://www.gnu.org/licenses/>.
+#
+
+require 'ffi'
+require 'ffi/prettystruct'
+require 'helpers/interface/ienum'
+require 'helpers/flags'
+require 'helpers/klass'
+
+
+module SGC
+module Cuda
+module API
+
+    extend FFI::Library
+    ffi_lib "cudart"
+
+    class Enum
+        extend SGC::Helper::IEnum
+        extend SGC::Helper::FlagsValue
+
+        def self.inherited(subclass)
+            subclass.instance_eval %{
+                def symbols
+                    SGC::Cuda::API::#{SGC::Helper.classname(subclass)}.symbols
+                end
+
+                def [](*args)
+                    SGC::Cuda::API::#{SGC::Helper.classname(subclass)}[*args]
+                end
+            }
+        end
+    end
+
+    CudaError = enum(
+        :SUCCESS, 0,
+        :ERROR_MISSING_CONFIGURATION, 1,
+        :ERROR_MEMORY_ALLOCATION, 2,
+        :ERROR_INITIALIZATION_ERROR, 3,
+        :ERROR_LAUNCH_FAILURE, 4,
+        :ERROR_PRIOR_LAUNCH_FAILURE, 5, # Deprecated.
+        :ERROR_LAUNCH_TIMEOUT, 6,
+        :ERROR_LAUNCH_OUT_OF_RESOURCES, 7,
+        :ERROR_INVALID_DEVICE_FUNCTION, 8,
+        :ERROR_INVALID_CONFIGURATION, 9,
+        :ERROR_INVALID_DEVICE, 10,
+        :ERROR_INVALID_VALUE, 11,
+        :ERROR_INVALID_PITCH_VALUE, 12,
+        :ERROR_INVALID_SYMBOL, 13,
+        :ERROR_MAP_BUFFER_OBJECT_FAILED, 14,
+        :ERROR_UNMAP_BUFFER_OBJECT_FAILED, 15,
+        :ERROR_INVALID_HOST_POINTER, 16,
+        :ERROR_INVALID_DEVICE_POINTER, 17,
+        :ERROR_INVALID_TEXTURE, 18,
+        :ERROR_INVALID_TEXTURE_BINDING, 19,
+        :ERROR_INVALID_CHANNEL_DESCRIPTOR, 20,
+        :ERROR_INVALID_MEMCPY_DIRECTION, 21,
+        :ERROR_ADDRESS_OF_CONSTANT, 22, # Deprecated.
+        :ERROR_TEXTURE_FETCH_FAILED, 23, # Deprecated.
+        :ERROR_TEXTURE_NOT_BOUND, 24, # Deprecated.
+        :ERROR_SYNCHRONIZATION_ERROR, 25, # Deprecated.
+        :ERROR_INVALID_FILTER_SETTING, 26,
+        :ERROR_INVALID_NORM_SETTING, 27,
+        :ERROR_MIXED_DEVICE_EXECUTION, 28, # Deprecated.
+        :ERROR_CUDART_UNLOADING, 29, # Deprecated.
+        :ERROR_UNKNOWN, 30,
+        :ERROR_NOT_YET_IMPLEMENTED, 31,
+        :ERROR_MEMORY_VALUE_TOO_LARGE, 32, # Deprecated.
+        :ERROR_INVALID_RESOURCE_HANDLE, 33,
+        :ERROR_NOT_READY, 34,
+        :ERROR_INSUFFICIENT_DRIVER, 35,
+        :ERROR_SET_ON_ACTIVE_PROCESS, 36,
+        :ERROR_INVALID_SURFACE, 37,
+        :ERROR_NO_DEVICE, 38,
+        :ERROR_ECC_UNCORRECTABLE, 39,
+        :ERROR_SHARED_OBJECT_SYMBOL_NOT_FOUND, 40,
+        :ERROR_SHARED_OBJECT_INIT_FAILED, 41,
+        :ERROR_UNSUPPORTED_LIMIT, 42,
+        :ERROR_DUPLICATE_VARIABLE_NAME, 43,
+        :ERROR_DUPLICATE_TEXTURE_NAME, 44,
+        :ERROR_DUPLICATE_SURFACE_NAME, 45,
+        :ERROR_DEVICES_UNAVAILABLE, 46,
+        :ERROR_INVALID_KERNEL_IMAGE, 47,
+        :ERROR_NO_KERNEL_IMAGE_FOR_DEVICE, 48,
+        :ERROR_INCOMPATIBLE_DRIVER_CONTEXT, 49,
+        :ERROR_PEER_ACCESS_ALREADY_ENABLED, 50,
+        :ERROR_PEER_ACCESS_NOT_ENABLED, 51,
+        :ERROR_DEVICE_ALREADY_IN_USE, 54,
+        :ERROR_PROFILER_DISABLED, 55,
+        :ERROR_PROFILER_NOT_INITIALIZED, 56,
+        :ERROR_PROFILER_ALREADY_STARTED, 57,
+        :ERROR_PROFILER_ALREADY_STOPPED, 58,
+        :ERROR_STARTUP_FAILURE, 0x7F,
+        :ERROR_API_FAILURE_BASE, 10000,
+    )
+
+    CudaDeviceFlags = enum(
+        :SCHEDULE_AUTO, 0,
+        :SCHEDULE_SPIN, 1,
+        :SCHEDULE_YIELD, 2,
+        :SCHEDULE_BLOCKING_SYNC, 4,
+        :BLOCKING_SYNC, 4, # Deprecated. Use :SCHEDULE_BLOCKING_SYNC.
+        :MAP_HOST, 8,
+        :LMEM_RESIZE_TO_MAX, 16,
+    )
+
+    CudaEventFlags = enum(
+        :DEFAULT, 0,
+        :BLOCKING_SYNC, 1,
+        :DISABLE_TIMING, 2,
+    )
+
+    CudaHostAllocFlags = enum(
+        :DEFAULT, 0,
+        :PORTABLE, 1,
+        :MAPPED, 2,
+        :WRITE_COMBINED, 4,
+    )
+
+    CudaHostRegisterFlags = enum(
+        :DEFAULT, 0,
+        :PORTABLE, 1,
+        :MAPPED, 2,
+    )
+
+    CudaArrayFlags = enum(
+        :DEFAULT, 0x00,
+        :LAYERED, 0x01,
+        :SURFACE_LOAD_STORE, 0x02,
+    )
+
+    CudaMemoryType = enum(
+        :Host, 1,
+        :DEVICE, 2,
+    )
+
+    CudaMemcpyKind = enum(
+        :HOST_TO_HOST, 0,
+        :HOST_TO_DEVICE, 1,
+        :DEVICE_TO_HOST, 2,
+        :DEVICE_TO_DEVICE, 3,
+        :DEFAULT, 4,
+    )
+
+    CudaChannelFormatKind = enum(
+        :SIGNED, 0,
+        :UNSIGNED, 1,
+        :FLOAT, 2,
+        :None,3,
+    )
+
+    CudaFunctionCache = enum(
+        :PREFER_NONE, 0,
+        :PREFER_SHARED, 1,
+        :PREFER_L1, 2,
+    )
+
+    CudaLimit = enum(
+        :STACK_SIZE, 0x00,
+        :PRINTF_FIFO_SIZE, 0x01,
+        :MALLOC_HEAP_SIZE, 0x02,
+    )
+
+    CudaOutputMode = enum(
+        :KEY_VALUE_PAIR, 0x00,
+        :CSV, 0x01,
+    )
+
+    CudaComputeMode = enum(
+        :DEFAULT, 0,
+        :EXCLUSIVE, 1,
+        :PROHIBITED, 2,
+        :EXCLUSIVE_PROCESS, 3,
+    )
+
+    CudaSurfaceBoundaryMode = enum(
+        :ZERO, 0,
+        :CLAMP, 1,
+        :TRAP, 2,
+    )
+
+    CudaSurfaceFormatMode = enum(
+        :FORCED, 0,
+        :AUTO, 1,
+    )
+
+    CudaTextureAddressMode = enum(
+        :WRAP, 0,
+        :CLAMP, 1,
+        :MIRROR, 2,
+        :BORDER, 3,
+    )
+
+    CudaTextureFilterMode = enum(
+        :POINT, 0,
+        :LINEAR, 1,
+    )
+
+    CudaTextureReadMode = enum(
+        :ELEMENT_TYPE, 0,
+        :NORMALIZED_FLOAT, 1,
+    )
+
+    FFI::typedef :int, :enum
+    FFI::typedef :pointer, :CudaStream
+    FFI::typedef :pointer, :CudaEvent
+
+    def read_int(ptr); ptr.read_int; end
+    def read_long(ptr); ptr.read_long; end
+    def read_pointer(ptr); ptr.read_pointer; end
+
+    def write_int(ptr); ptr.write_int; end
+    def write_long(ptr); ptr.write_long; end
+    def write_pointer(ptr, value); ptr.write_pointer(value); end
+
+    alias read_size_t read_long
+    alias read_enum read_int
+    alias read_cudastream read_pointer
+    alias read_cudaevent read_pointer
+
+    alias write_size_t write_long
+    alias write_enum write_int
+    alias write_cudastream write_pointer
+    alias write_cudaevent write_pointer
+
+    module_function :read_size_t
+    module_function :read_enum
+    module_function :read_cudastream
+    module_function :read_cudaevent
+
+    module_function :write_size_t
+    module_function :write_enum
+    module_function :write_cudastream
+    module_function :write_cudaevent
+
+
+    class Dim3 < FFI::Struct
+        layout(
+            :array, [:uint, 3],
+        )
+
+        alias :init :initialize
+        alias :get :[]
+        alias :set :[]=
+        private :init, :get, :set
+
+        def initialize(x, y, z)
+            init
+            @array = get(:array)
+            @array[0], @array[1], @array[2] = x, y, z
+        end
+
+        def [](index); @array[index]; end
+        def []=(index, value); @array[index] = value; end
+
+        def x; @array[0]; end
+        def y; @array[1]; end
+        def z; @array[2]; end
+
+        def x=(value); @array[0] = value; end
+        def y=(value); @array[1] = value; end
+        def z=(value); @array[2] = value; end
+    end
+
+    class CudaDeviceProp < FFI::PrettyStruct
+        layout(
+            :name, [:char, 256],
+            :total_global_mem, :size_t,
+            :shared_mem_per_block, :size_t,
+            :regs_per_block, :int,
+            :warp_size, :int,
+            :mem_pitch, :size_t,
+            :max_threads_per_block, :int,
+            :max_threads_dim, [:int, 3],
+            :max_grid_size, [:int, 3],
+            :clock_rate, :int,
+            :total_const_mem, :size_t,
+            :major, :int,
+            :minor, :int,
+            :texture_alignment, :size_t,
+            :device_overlap, :int, # Deprecated. Use :async_engine_count.
+            :multi_processor_count, :int,
+            :kernel_exec_timeout_enabled, :int,
+            :integrated, :int,
+            :can_map_host_memory, :int,
+            :compute_mode, :int,
+            :max_texture1d, :int,
+            :max_texture2d, [:int, 2],
+            :max_texture3d, [:int, 3],
+            :max_texture1d_layered, [:int, 2],
+            :max_texture2d_layered, [:int, 3],
+            :surface_alignment, :size_t,
+            :concurrent_kernels, :int,
+            :ecc_enabled, :int,
+            :pci_bus_id, :int,
+            :pci_device_id, :int,
+            :tcc_driver, :int,
+            :async_engine_count, :int,
+            :unified_addressing, :int,
+            :memory_clock_rate, :int,
+            :memory_bus_width, :int,
+            :l2_cache_size, :int,
+            :max_threads_per_multi_processor, :int,
+        )
+    end
+
+    class CudaFunctionAttributes < FFI::PrettyStruct
+        layout(
+            :shared_size_bytes, :size_t,
+            :const_size_bytes, :size_t,
+            :local_size_bytes, :size_t,
+            :max_threads_per_block, :int,
+            :num_regs, :int,
+            :ptx_version, :int,
+            :binary_version, :int,
+        )
+    end
+
+    class CudaPointerAttributes < FFI::PrettyStruct
+        layout(
+            :memory_type, CudaMemoryType,
+            :device, :int,
+            :device_pointer, :pointer,
+            :host_pointer, :pointer,
+        )
+    end
+
+    class CudaChannelFormatDesc < FFI::PrettyStruct
+        layout(
+            :x, :int,
+            :y, :int,
+            :z, :int,
+            :w, :int,
+            :f, CudaChannelFormatKind,
+        )
+    end
+
+    class CudaPitchedPtr < FFI::PrettyStruct
+        layout(
+            :ptr, :pointer,
+            :pitch, :size_t,
+            :xsize, :size_t,
+            :ysize, :size_t,
+        )
+    end
+
+    class CudaPos < FFI::PrettyStruct
+        layout(
+            :x, :size_t,
+            :y, :size_t,
+            :z, :size_t,
+        )
+    end
+
+    class CudaExtent < FFI::PrettyStruct
+        layout(
+            :width, :size_t,
+            :height, :size_t,
+            :depth, :size_t,
+        )
+    end
+
+    class CudaMemcpy3DParms < FFI::PrettyStruct
+        layout(
+            :src_array, :pointer,
+            :src_pos, CudaPos,
+            :src_ptr, CudaPitchedPtr,
+            :dst_array, :pointer,
+            :dst_pos, CudaPos,
+            :dst_ptr, CudaPitchedPtr,
+            :extent, CudaExtent,
+            :kind, CudaMemcpyKind,
+        )
+    end
+
+    class CudaMemcpy3DPeerParms < FFI::PrettyStruct
+        layout(
+            :src_array, :pointer,
+            :src_pos, CudaPos,
+            :src_ptr, CudaPitchedPtr,
+            :src_device, :int,
+            :dst_array, :pointer,
+            :dst_pos, CudaPos,
+            :dst_ptr, CudaPitchedPtr,
+            :dst_device, :int,
+            :extent, CudaExtent,
+        )
+    end
+
+    class TextureReference < FFI::PrettyStruct
+        layout(
+            :normalized, :int,
+            :filter_mode, CudaTextureFilterMode,
+            :address_mode, [CudaTextureAddressMode, 3],
+            :channel_desc, CudaChannelFormatDesc,
+            :srgb, :int,
+            :__cuda_reserved, [:int, 15],
+        )
+    end
+
+    class SurfaceReference < FFI::PrettyStruct
+        layout(
+            :channel_desc, CudaChannelFormatDesc,
+        )
+    end
+
+    # CUDA Version Management.
+    attach_function :cudaDriverGetVersion, [:pointer], :enum
+    attach_function :cudaRuntimeGetVersion, [:pointer], :enum
+
+    # CUDA Error Handling.
+    attach_function :cudaGetErrorString, [CudaError], :string
+    attach_function :cudaGetLastError, [], :enum
+    attach_function :cudaPeekAtLastError, [], :enum
+
+    # CUDA Device Management.
+    attach_function :cudaChooseDevice, [:pointer, :pointer], :enum
+    attach_function :cudaDeviceGetCacheConfig, [:pointer], :enum
+    attach_function :cudaDeviceGetLimit, [:pointer, CudaLimit], :enum
+    attach_function :cudaDeviceReset, [], :enum
+    attach_function :cudaDeviceSetCacheConfig, [CudaFunctionCache], :enum
+    attach_function :cudaDeviceSetLimit, [CudaLimit, :size_t], :enum
+    attach_function :cudaDeviceSynchronize, [], :enum
+    attach_function :cudaGetDevice, [:pointer], :enum
+    attach_function :cudaGetDeviceCount, [:pointer], :enum
+    attach_function :cudaGetDeviceProperties, [:pointer, :int], :enum
+    attach_function :cudaSetDevice, [:int], :enum
+    attach_function :cudaSetDeviceFlags, [:uint], :enum
+    attach_function :cudaSetValidDevices, [:pointer, :int], :enum
+
+    # CUDA Thread Management.
+    # Deprecated.
+    attach_function :cudaThreadExit, [], :enum
+    attach_function :cudaThreadGetCacheConfig, [:pointer], :enum
+    attach_function :cudaThreadGetLimit, [:pointer, CudaLimit], :enum
+    attach_function :cudaThreadSetCacheConfig, [CudaFunctionCache], :enum
+    attach_function :cudaThreadSetLimit, [CudaLimit, :size_t], :enum
+    attach_function :cudaThreadSynchronize, [], :enum
+
+    # CUDA Memory Management.
+    attach_function :cudaFree, [:pointer], :enum
+    attach_function :cudaFreeArray, [:pointer], :enum
+    attach_function :cudaFreeHost, [:pointer], :enum
+    attach_function :cudaGetSymbolAddress, [:pointer, :string], :enum
+    attach_function :cudaGetSymbolSize, [:pointer, :string], :enum
+    attach_function :cudaHostAlloc, [:pointer, :size_t, :uint], :enum
+    attach_function :cudaHostGetDevicePointer, [:pointer, :pointer, :uint], :enum
+    attach_function :cudaHostGetFlags, [:pointer, :pointer], :enum
+    attach_function :cudaHostRegister, [:pointer, :size_t, :uint], :enum
+    attach_function :cudaHostUnregister, [:pointer], :enum
+    attach_function :cudaMalloc, [:pointer, :size_t], :enum
+    attach_function :cudaMalloc3D, [:pointer, CudaExtent.by_value], :enum
+    attach_function :cudaMalloc3DArray, [:pointer, :pointer, CudaExtent.by_value, :uint], :enum
+    attach_function :cudaMallocArray, [:pointer, :pointer, :size_t, :size_t, :uint], :enum
+    attach_function :cudaMallocHost, [:pointer, :size_t], :enum
+    attach_function :cudaMallocPitch, [:pointer, :pointer, :size_t, :size_t], :enum
+    attach_function :cudaMemcpy, [:pointer, :pointer, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpy2D, [:pointer, :size_t, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpy2DArrayToArray, [:pointer, :size_t, :size_t, :pointer, :size_t, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpy2DAsync, [:pointer, :size_t, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpy2DFromArray, [:pointer, :size_t, :pointer, :size_t, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpy2DFromArrayAsync, [:pointer, :size_t, :pointer, :size_t, :size_t, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpy2DToArray, [:pointer, :size_t, :size_t, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpy2DToArrayAsync, [:pointer, :size_t, :size_t, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpy3D, [:pointer], :enum
+    attach_function :cudaMemcpy3DAsync, [:pointer, :CudaStream], :enum
+    attach_function :cudaMemcpy3DPeer, [:pointer], :enum
+    attach_function :cudaMemcpy3DPeerAsync, [:pointer, :CudaStream], :enum
+    attach_function :cudaMemcpyArrayToArray, [:pointer, :size_t, :size_t, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpyAsync, [:pointer, :pointer, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpyFromArray, [:pointer, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpyFromArrayAsync, [:pointer, :pointer, :size_t, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpyFromSymbol, [:pointer, :string, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpyFromSymbolAsync, [:pointer, :string, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpyPeer, [:pointer, :int, :pointer, :int, :size_t], :enum
+    attach_function :cudaMemcpyPeerAsync, [:pointer, :int, :pointer, :int, :size_t, :CudaStream], :enum
+    attach_function :cudaMemcpyToArray, [:pointer, :size_t, :size_t, :pointer, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpyToArrayAsync, [:pointer, :size_t, :size_t, :pointer, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemcpyToSymbol, [:string, :pointer, :size_t, :size_t, CudaMemcpyKind], :enum
+    attach_function :cudaMemcpyToSymbolAsync, [:string, :pointer, :size_t, :size_t, CudaMemcpyKind, :CudaStream], :enum
+    attach_function :cudaMemGetInfo, [:pointer, :pointer], :enum
+    attach_function :cudaMemset, [:pointer, :int, :size_t], :enum
+    attach_function :cudaMemset2D, [:pointer, :size_t, :int, :size_t, :size_t], :enum
+    attach_function :cudaMemset2DAsync, [:pointer, :size_t, :int, :size_t, :size_t, :CudaStream], :enum
+    attach_function :cudaMemset3D, [CudaPitchedPtr.by_value, :int, CudaExtent.by_value], :enum
+    attach_function :cudaMemset3DAsync, [CudaPitchedPtr.by_value, :int, CudaExtent.by_value, :CudaStream], :enum
+    attach_function :cudaMemsetAsync, [:pointer, :int, :size_t, :CudaStream], :enum
+    # attach_function :make_cudaExtent, [:size_t, :size_t, :size_t], CudaExtent
+    # attach_function :make_cudaPitchedPtr, [:pointer, :size_t, :size_t, :size_t], CudaPitchedPtr
+    # attach_function :make_cudaPos, [:size_t, :size_t, :size_t], CudaPos
+
+    def make_cudaExtent(w, h, d)
+        e = CudaExtent.new
+        e[:width], e[:height], e[:depth] = w, h, d
+        e
+    end
+    module_function :make_cudaExtent
+
+    def make_cudaPitchedPtr(d, p, xsz, ysz)
+        s = CudaPitchedPtr.new
+        s[:ptr] = d
+        s[:pitch] = p
+        s[:xsize] = xsz
+        s[:ysize] = ysz
+        s
+    end
+    module_function :make_cudaPitchedPtr
+
+    def make_cudaPos(x, y, z)
+        p = CudaPos.new
+        p[:x] = x
+        p[:y] = y
+        p[:z] = z
+        p
+    end
+    module_function :make_cudaPos
+
+    # CUDA Unified Addressing.
+    attach_function :cudaPointerGetAttributes, [:pointer, :pointer], :enum
+
+    # CUDA Peer Device Memory Access.
+    attach_function :cudaDeviceCanAccessPeer, [:pointer, :int, :int], :enum
+    attach_function :cudaDeviceDisablePeerAccess, [:int], :enum
+    attach_function :cudaDeviceEnablePeerAccess, [:int, :uint], :enum
+
+    # CUDA Execution Control.
+    attach_function :cudaConfigureCall, [Dim3.by_value, Dim3.by_value, :size_t, :CudaStream], :enum
+    attach_function :cudaFuncGetAttributes, [:pointer, :string], :enum
+    attach_function :cudaFuncSetCacheConfig, [:string, CudaFunctionCache], :enum
+    attach_function :cudaLaunch, [:string], :enum
+    attach_function :cudaSetDoubleForDevice, [:pointer], :enum
+    attach_function :cudaSetDoubleForHost, [:pointer], :enum
+    attach_function :cudaSetupArgument, [:pointer, :size_t, :size_t], :enum
+
+    # CUDA Stream Management.
+    attach_function :cudaStreamCreate, [:pointer], :enum
+    attach_function :cudaStreamDestroy, [:CudaStream], :enum
+    attach_function :cudaStreamQuery, [:CudaStream], :enum
+    attach_function :cudaStreamSynchronize, [:CudaStream], :enum
+    attach_function :cudaStreamWaitEvent, [:CudaStream, :CudaEvent, :uint], :enum
+
+    # CUDA Event Management.
+    attach_function :cudaEventCreate, [:pointer], :enum
+    attach_function :cudaEventCreateWithFlags, [:pointer, :uint], :enum
+    attach_function :cudaEventDestroy, [:CudaEvent], :enum
+    attach_function :cudaEventElapsedTime, [:pointer, :CudaEvent, :CudaEvent], :enum
+    attach_function :cudaEventQuery, [:CudaEvent], :enum
+    attach_function :cudaEventRecord, [:CudaEvent, :CudaStream], :enum
+    attach_function :cudaEventSynchronize, [:CudaEvent], :enum
+
+    # CUDA Texture Reference Management.
+    attach_function :cudaBindTexture, [:pointer, :pointer, :pointer, :pointer, :size_t], :enum
+    attach_function :cudaBindTexture2D, [:pointer, :pointer, :pointer, :pointer, :size_t, :size_t, :size_t], :enum
+    attach_function :cudaBindTextureToArray, [:pointer, :pointer, :pointer], :enum
+    attach_function :cudaCreateChannelDesc, [:int, :int, :int, :int, CudaChannelFormatKind], CudaChannelFormatDesc.by_value
+    attach_function :cudaGetChannelDesc, [:pointer, :pointer], :enum
+    attach_function :cudaGetTextureAlignmentOffset, [:pointer, :pointer], :enum
+    attach_function :cudaGetTextureReference, [:pointer, :string], :enum
+    attach_function :cudaUnbindTexture, [:pointer], :enum
+
+    # CUDA Surface Reference Management.
+    attach_function :cudaBindSurfaceToArray, [:pointer, :pointer, :pointer], :enum
+    attach_function :cudaGetSurfaceReference, [:pointer, :string], :enum
+
+end # module
+end # module
+end # module