sgc-ruby-cuda 0.1.0

data/lib/cuda/runtime/function.rb

@@ -0,0 +1,161 @@
+#
+# 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/cuda'
+require 'cuda/runtime/error'
+require 'cuda/runtime/stream'
+require 'memory/pointer'
+require 'dl'
+
+
+module SGC
+module Cuda
+
+class CudaFunction
+
+    attr_reader :name
+
+
+    # Create an instance to function _name_.
+    def initialize(name)
+        @name = name
+    end
+
+
+    # @return [CudaFunctionAttributes] The attributes of this kernel function.
+    def attributes
+        a = CudaFunctionAttributes.new
+        status = API::cudaFuncGetAttributes(a.to_ptr, @name)
+        Pvt::handle_error(status, "Failed to query function attributes.")
+        a
+    end
+
+
+    # Set the preferred cache configuration to use for next launch on this kernel function.
+    # @param [CudaFunctionCache] conf The preferred cache configuration.
+    def cache_config=(conf)
+        status = API::cudaFuncSetCacheConfig(@name, conf)
+        Pvt::handle_error(status, "Failed to set function cache config: config = #{conf}.")
+    end
+
+
+    # Launch this kernel function with pre-configured settings.
+    # @return [Class] This class.
+    #
+    # @see .configure
+    def launch
+        status = API::cudaLaunch(@name)
+        Pvt::handle_error(status, "Failed to launch kernel function: name = #{@name}.")
+        self
+    end
+
+
+    # Configure the settings for the next kernel launch.
+    # @param [Dim3] grid_dim The 3D grid dimensions x, y, z to launch.
+    # @param [Dim3] block_dim The 3D block dimensions x, y, z to launch.
+    # @param [Integer] shared_mem_size Number of bytes of dynamic shared memory for each thread block.
+    # @param [Integer, CudaStream] stream The stream to launch this kernel function on.
+    #     Setting _stream_ to anything other than an instance of CudaStream will execute on the default stream 0.
+    # @return [Class] This class.
+    def self.configure(grid_dim, block_dim, shared_mem_size = 0, stream = 0)
+        s = Pvt::parse_stream(stream)
+        status = API::cudaConfigureCall(grid_dim, block_dim, shared_mem_size, s)
+        Pvt::handle_error(status, "Failed to configure kernel function launch settings.\n" +
+            "* #{grid_dim.x} x #{grid_dim.y} x #{grid_dim.z} grid\n" +
+            "* #{block_dim.x} x #{block_dim.y} x #{block_dim.z} blocks\n" +
+            "* shared memory size = #{shared_mem_size}")
+        self
+    end
+
+
+    # Set the argument list of subsequent kernel function launch.
+    # @param [Array] *args The list of arguments to pass to the kernel.
+    # @return [Class] This class.
+    def self.setup(*args)
+        offset = 0
+        args.each do |x|
+            case x
+                when Fixnum
+                    p = FFI::MemoryPointer.new(:int).write_int(x)
+                    size = 4
+                when Float
+                    p = FFI::MemoryPointer.new(:float).write_float(x)
+                    size = 4
+                when SGC::Memory::MemoryPointer
+                    p = x.ref
+                    size = FFI::MemoryPointer.size
+                else
+                    raise TypeError, "Invalid type of kernel parameters #{x}."
+            end
+            offset = align_up(offset, size)
+            status = API::cudaSetupArgument(p, size, offset)
+            Pvt::handle_error(status, "Failed to setup kernel argument for #{x}.")
+            offset += size
+        end
+        self
+    end
+
+
+    # Load a dynamic library with _name_ from dynamic library path.
+    # @param [String] name The name of the dynamic library to load.
+    #     For library libcudart.so, its name is cudart.
+    # @return [Class] This class.
+    def self.load_lib(name)
+        raise NotImplementedError
+    end
+
+
+    # Load a dynamic library from the given path.
+    # @param [String] path The path of the dynamic library to load.
+    # @return [Class] This class.
+    def self.load_lib_file(path)
+        @@libs << DL::dlopen(path)
+        self
+    end
+
+
+    # Unload all the loaded dynamic libraries.
+    # @return [Class] This class.
+    def self.unload_all_libs
+        @@libs.each do |h|
+            h.close
+        end
+        @@libs = []
+        self
+    end
+
+
+    @@libs = [] # @private
+
+private
+
+    def self.align_up(offset, alignment)
+        (offset + alignment - 1) & ~(alignment - 1)
+    end
+
+end
+
+end # module
+end # module

data/lib/cuda/runtime/memory.rb

@@ -0,0 +1,110 @@
+#
+# 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'
+require 'memory/pointer'
+
+
+module SGC
+module Cuda
+
+class CudaDeviceMemory
+
+    # Allocate memory on the device.
+    # @param [Integer] nbytes The number of bytes of memory to allocate.
+    # @return [*SGC::Memory::MemoryPointer] A memory pointer to the allocated device memory. 
+    #
+    # @note The returned memory pointer is enabled to call _free_ method on itself.
+    def self.malloc(nbytes)
+        p = SGC::Memory::MemoryPointer.new
+        status = API::cudaMalloc(p.ref, nbytes)
+        Pvt::handle_error(status, "Failed to allocate memory on the device: nbytes = #{nbytes}")
+        p.instance_eval %{
+            def free
+                CudaDeviceMemory.free(self)
+            end
+        }
+        p
+    end
+
+
+    # Free the device memory at _devptr_.
+    # @param [*SGC::Memory::MemoryPointer] devptr The memory pointer pointing to the device memory to be freed.
+    def self.free(devptr)
+        status = API::cudaFree(devptr.ptr)
+        Pvt::handle_error(status, "Failed to free the device memory.")
+        devptr.ptr = 0
+        nil
+    end
+
+end
+
+
+module CudaMemory
+
+    # Copy _nbytes_ from the memory at _src_ptr_ to the memory at _dst_ptr_.
+    # @param [#ptr] dst_ptr Destination of the memory copy.
+    # @param [#ptr] src_ptr Source of the memory copy.
+    # @param [Integer] nbytes The number of bytes to copy.
+    # @param [Symbol] memcpy_kind The direction of the memory copy specified with one of the following:
+    #     * :HOST_TO_HOST
+    #     * :HOST_TO_DEVICE
+    #     * :DEVICE_TO_HOST
+    #     * :DEVICE_TO_DEVICE
+    def memcpy(dst_ptr, src_ptr, nbytes, memcpy_kind)
+        status = API::cudaMemcpy(dst_ptr.ptr, src_ptr.ptr, nbytes, memcpy_kind)
+        Pvt::handle_error(status, "Failed to copy memory.")
+    end
+    module_function :memcpy
+
+
+    # Copy _nbytes_ from the host memory at _src_ptr_ to the host memory at _dst_ptr_.
+    def memcpy_htoh(dst_ptr, src_ptr, nbytes)
+        memcpy(dst_ptr, src_ptr, nbytes, :HOST_TO_HOST)
+    end
+    module_function :memcpy_htoh
+
+    # Copy _nbytes_ from the host memory at _src_ptr_ to the device memory at _dst_ptr_.
+    def memcpy_htod(dst_ptr, src_ptr, nbytes)
+        memcpy(dst_ptr, src_ptr, nbytes, :HOST_TO_DEVICE)
+    end
+    module_function :memcpy_htod
+
+    # Copy _nbytes_ from the device memory at _src_ptr_ to the host memory at _dst_ptr_.
+    def memcpy_dtoh(dst_ptr, src_ptr, nbytes)
+        memcpy(dst_ptr, src_ptr, nbytes, :DEVICE_TO_HOST)
+    end
+    module_function :memcpy_dtoh
+
+    # Copy _nbytes_ from the device memory at _src_ptr_ to the device memory at _dst_ptr_.
+    def memcpy_dtod(dst_ptr, src_ptr, nbytes)
+        memcpy(dst_ptr, src_ptr, nbytes, :DEVICE_TO_DEVICE)
+    end
+    module_function :memcpy_dtod
+
+end
+
+end # module
+end # module

data/lib/cuda/runtime/rubycuda.rb

@@ -0,0 +1,34 @@
+#
+# 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/cuda'
+require 'cuda/runtime/error'
+require 'cuda/runtime/version'
+require 'cuda/runtime/device'
+require 'cuda/runtime/thread'
+require 'cuda/runtime/memory'
+require 'cuda/runtime/function'
+require 'cuda/runtime/stream'
+require 'cuda/runtime/event'

data/lib/cuda/runtime/stream.rb

@@ -0,0 +1,126 @@
+#
+# 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 CudaStream
+
+    # Create and return a CUDA stream.
+    # @return [CudaStream] A CUDA stream.
+    def self.create
+        p = FFI::MemoryPointer.new(:CudaStream)
+        status = API::cudaStreamCreate(p)
+        Pvt::handle_error(status, "Failed to create CUDA stream.")
+        new(p)
+    end
+
+
+    # Destroy this CUDA stream.
+    def destroy
+        status = API::cudaStreamDestroy(self.to_api)
+        Pvt::handle_error(status, "Failed to destroy this CUDA stream.")
+        API::write_cudastream(@pstream, 0)
+        nil
+    end
+
+
+    # @return [Boolean] Return true if all operations in this CUDA stream have completed. Otherwise, return false.
+    def query
+        status = API::cudaStreamQuery(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 stream.")
+        raise CudaStandardError, "Error handling fails to catch this error."
+    end
+
+
+    # Block the calling CPU thread until all operations in this CUDA stream complete.
+    # @return [CudaStream] This CUDA stream.
+    def synchronize
+        status = API::cudaStreamSynchronize(self.to_api)
+        Pvt::handle_error(status)
+        self
+    end
+
+
+    # Let all future operations submitted to this CUDA stream wait until _event_ complete before beginning execution.
+    # @overload wait_event(event)
+    # @overload wait_event(event, flags)
+    # @param [CudaEvent] event The event to wait for.
+    # @param [Integer] flags Currently _flags_ must be set to zero.
+    # @return [CudaStream] This CUDA stream.
+    def wait_event(event, flags = 0)
+        status = API::cudaStreamWaitEvent(self.to_api, event.to_api, flags)
+        Pvt::handle_error(status, "Failed to make this CUDA stream's future operations to wait event: flags = #{flags}.")
+        self
+    end
+
+
+    # Let all future operations submitted to any CUDA stream wait until _event_ complete before beginning execution.
+    # @overload wait_event(event)                      
+    # @overload wait_event(event, flags)               
+    # @param (see CudaStream#wait_event) 
+    def self.wait_event(event, flags = 0)
+        status = API::cudaStreamWaitEvent(nil, event.to_api, flags)
+        Pvt::handle_error(status, "Failed to make any CUDA stream's future operations to wait event: flags = #{flags}.")
+        nil
+    end
+
+
+    # @private
+    def initialize(ptr)
+        @pstream = ptr
+    end
+    private_class_method :new
+
+
+    # @private
+    def to_api
+        API::read_cudastream(@pstream)
+    end
+
+end
+
+# @private
+module Pvt
+
+    def self.parse_stream(stream)
+        if stream.kind_of?(CudaStream)
+            return stream.to_api
+        end
+        nil
+    end
+
+end
+
+end # module
+end # module

data/lib/cuda/runtime/thread.rb

@@ -0,0 +1,81 @@
+#
+# 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/cuda'
+require 'cuda/runtime/error'
+
+
+module SGC
+module Cuda
+
+# @deprecated
+class CudaThread
+
+    def self.exit
+        status = API::cudaThreadExit
+        Pvt::handle_error(status)
+        self
+    end
+
+
+    def self.cache_config
+        p = FFI::MemoryPointer.new(:int)
+        status = API::cudaThreadGetCacheConfig(p)
+        Pvt::handle_error(status)
+        CudaFunctionCache[p.read_int]
+    end
+
+
+    def self.cache_config=(config)
+        status = API::cudaThreadSetCacheConfig(config)
+        Pvt::handle_error(status)
+    end
+
+
+    def self.limit(limit)
+        p = FFI::MemoryPointer.new(:size_t)
+        status = API::cudaThreadGetLimit(p, limit)
+        Pvt::handle_error(status)
+        p.read_long
+    end
+
+
+    def self.limit=(*limit_value_pair)
+        limit, value = limit_value_pair.flatten
+        status = API::cudaThreadSetLimit(limit, value)
+        Pvt::handle_error(status)
+    end
+
+
+    def self.synchronize
+        status = API::cudaThreadSynchronize
+        Pvt::handle_error(status)
+        self
+    end
+
+end
+
+end # module
+end # module