sgc-ruby-cuda 0.1.0

data/lib/cuda/driver/function.rb

@@ -0,0 +1,293 @@
+#
+# Copyright (c) 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/driver/ffi-cu'
+require 'cuda/driver/error'
+
+
+module SGC
+module CU
+
+class CUFunction
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set the argument list of subsequent function call to _arg1_, _arg2_, *other_args.
+    # @param *args The list of arguments to pass to the kernel function.
+    def param=(*args)
+        offset = 0
+        args.flatten.each do |x|
+            case x
+                when Fixnum
+                    p = FFI::MemoryPointer.new(:int)
+                    p.write_int(x)
+                    size = 4
+                when Float
+                    p = FFI::MemoryPointer.new(:float)
+                    p.write_float(x)
+                    size = 4
+                when CUDevicePtr
+                    p = FFI::MemoryPointer.new(:CUDevicePtr)
+                    API::write_cudeviceptr(p, x.to_api.address)
+                    size = p.size
+                else
+                    raise TypeError, "Invalid type of argument #{x.to_s}."
+            end
+            offset = align_up(offset, size)
+            status = API::cuParamSetv(self.to_api, offset, p, size)
+            Pvt::handle_error(status, "Failed to set function parameters: offset = #{offset}, value = #{x}.")
+            offset += size
+        end
+
+        status = API::cuParamSetSize(self.to_api, offset)
+        Pvt::handle_error(status, "Failed to set function parameter size: size = #{offset}.")
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set a float parameter to the function's argument list at _offset_ with _value_.
+    # @param [Integer] offset Number of bytes to offset.
+    # @param [Float] value The floating-point value to set as the function parameter.
+    # @return [CUFunction] This function.
+    def param_setf(offset, value)
+        status = API::cuParamSetf(self.to_api, offset, value)
+        Pvt::handle_error(status, "Failed to set function float parameter: offset = #{offset}, value = #{value}.")
+        self
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set an integer parameter to the function's argument list at _offset_ with _value_.
+    # @param [Integer] offset Number of bytes to offset.
+    # @param [Integer] value The integer value to set as the function parameter.
+    # @return [CUFunction] This function.
+    def param_seti(offset, value)
+        status = API::cuParamSeti(self.to_api, offset, value)
+        Pvt::handle_error(status, "Failed to set function integer parameter: offset = #{offset}, value = #{value}")
+        self
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set an arbitrary data to the function's argument list at _offset_ with _ptr_ pointed _nbytes_ data.
+    # @param [Integer] offset Number of bytes to offset.
+    # @param [CUDevicePtr] ptr A device pointer pointing to an arbitrary data to be used as the function parameter.
+    # @param [Integer] nbytes The size of the arbitrary data in bytes.
+    # @return [CUFunction] This function.
+    def param_setv(offset, ptr, nbytes)
+        status = API::cuParamSetv(self.to_api, offset, ptr.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to set function arbitrary parameter: offset = #{offset}, size = #{nbytes}.")
+        self
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set the function parameter size to _nbytes_.
+    # @param [Integer] nbytes The parameter size in bytes.
+    # @return [CUFunction] This function.
+    def param_set_size(nbytes)
+        status = API::cuParamSetSize(self.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to set function parameter size: size = #{nbytes}.")
+        self
+    end
+
+
+    # @deprecated
+    #
+    def param_set_texref(texunit, texref)
+        raise NotImplementedError
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set the block dimensions to use for next launch.
+    # @overload block_shape=(xdim)
+    # @overload block_shape=(xdim, ydim)
+    # @overload block_shape=(xdim, ydim, zdim)
+    # @param [Integer] xdim The size of the x dimension.
+    # @param [Integer] ydim The size of the y dimension. Defaults to 1.
+    # @param [Integer] zdim The size of the z dimension. Defaults to 1.
+    def block_shape=(*args)
+        xdim, ydim, zdim = args.flatten
+        ydim = 1 if ydim.nil?
+        zdim = 1 if zdim.nil?
+        status = API::cuFuncSetBlockShape(self.to_api, xdim, ydim, zdim)
+        Pvt::handle_error(status, "Failed to set function block shape: (x,y,z) = (#{xdim},#{ydim},#{zdim}).")
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Set the dynamic shared-memory size to use for next launch.
+    # @param [Integer] nbytes Number of bytes.
+    def shared_size=(nbytes)
+        status = API::cuFuncSetSharedSize(self.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to set function shared memory size: #{nbytes}.")
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Launch this kernel function with 1x1x1 grid of blocks to execute on the current CUDA device.
+    # @return [CUFunction] This function.
+    def launch
+        status = API::cuLaunch(self.to_api)
+        Pvt::handle_error(status, "Failed to launch kernel function on 1x1x1 grid of blocks.")
+        self
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Launch this kernel function with grid dimensions (_xdim_, _ydim_) to execute on the current CUDA device.
+    # @overload launch_grid(xdim)
+    # @overload launch_grid(xdim, ydim)
+    # @param [Integer] xdim The x dimensional size of the grid to launch.
+    # @param [Integer] ydim The y dimensional size of the grid to launch. Defaults to 1.
+    # @return [CUFunction] This function.
+    def launch_grid(xdim, ydim = 1)
+        status = API::cuLaunchGrid(self.to_api, xdim, ydim)
+        Pvt::handle_error(status, "Failed to launch kernel function on #{xdim}x#{ydim} grid of blocks.")
+        self
+    end
+
+
+    # @deprecated Use {#launch_kernel}.
+    #
+    # Launch this kernel function with grid dimensions (_xdim_, _ydim_) on _stream_ asynchronously to execute
+    # on the current CUDA device. Setting _stream_ to anything other than an instance of CUStream
+    # will execute on the default stream 0.
+    # @overload launch_grid_async(xdim, stream)
+    # @overload launch_grid_async(xdim, ydim, stream)
+    # @param [Integer] xdim The x dimensional size 
+    def launch_grid_async(xdim, ydim = 1, stream)
+        s = Pvt::parse_stream(stream)
+        status = API::cuLaunchGridAsync(self.to_api, xdim, ydim, s)
+        Pvt::handle_error(status, "Failed to launch kernel function asynchronously on #{xdim}x#{ydim} grid of blocks.")
+        self
+    end
+
+
+    # @param [CUFunctionAttribute] attrib The attribute of the kernel function to query.
+    # @return [Integer] The particular attribute _attrib_ of this kernel function.
+    #
+    # @example Get function attribute.
+    #    func.attribute(:MAX_THREADS_PER_BLOCK)    #=> 512
+    #    func.attribute(:SHARED_SIZE_BYTES)        #=> 44
+    #    func.attribute(:NUM_REGS)                 #=> 3
+    def attribute(attrib)
+        p = FFI::MemoryPointer.new(:int)
+        status = API::cuFuncGetAttribute(p, attrib, self.to_api)
+        Pvt::handle_error(status, "Failed to query function attribute: attribute = #{attrib}.")
+        p.read_int
+    end
+
+
+    # Set the preferred cache configuration (CUFunctionCache) to use for next launch.
+    # @param [CUFunctionCache] conf The preferred cache configuration.
+    def cache_config=(conf)
+        status = API::cuFuncSetCacheConfig(self.to_api, conf)
+        Pvt::handle_error(status, "Failed to set function cache config: config = #{conf}.")
+    end
+
+    # Launch this kernel function with full configuration parameters and function parameters
+    # to execute on the current CUDA device.
+    # @param [Integer] grid_xdim The x dimensional size of the grid to launch.
+    # @param [Integer] grid_ydim The y dimensional size of the grid to launch.
+    # @param [Integer] grid_zdim The z dimensional size of the grid to launch.
+    # @param [Integer] block_xdim The x dimensional size of a block in the grid.
+    # @param [Integer] block_ydim The y dimensional size of a block in the grid.
+    # @param [Integer] block_zdim The z dimensional size of a block in the grid.
+    # @param [Integer] shared_mem_size Number of bytes of dynamic shared memory for each thread block.
+    # @param [Integer, CUStream] stream The stream to launch this kernel function on.
+    #     Setting _stream_ to anything other than an instance of CUStream will execute on the default stream 0.
+    # @param [Array<Fixnum, Float, CUDevicePtr>] params The list of parameters to pass in for the kernel function launch.
+    #     * A Fixnum is mapped to a C type int.
+    #     * A Float is mapped to a C type float.
+    # @return [CUFunction] This function.
+    #
+    # @todo Add support for other C data types for the kernel function parameters.
+    def launch_kernel(grid_xdim, grid_ydim, grid_zdim, block_xdim, block_ydim, block_zdim, shared_mem_size, stream, params)
+        p = parse_params(params)
+        s = Pvt::parse_stream(stream)
+        status = API::cuLaunchKernel(self.to_api, grid_xdim, grid_ydim, grid_zdim, block_xdim, block_ydim, block_zdim, shared_mem_size, s, p, nil)
+        Pvt::handle_error(status, "Failed to launch kernel function.\n" +
+            "* #{grid_xdim} x #{grid_ydim} x #{grid_zdim} grid\n" +
+            "* #{block_xdim} x #{block_ydim} x #{block_zdim} blocks\n" +
+            "* shared memory size = #{shared_mem_size}")
+        self
+    end
+
+
+    # @private
+    def initialize(ptr)
+        @pfunc = ptr
+    end
+    private_class_method :new
+
+
+    # @private
+    def to_api
+        API::read_cufunction(@pfunc)
+    end
+
+private
+
+    def parse_params(params)
+        params.is_a?(Array) or raise TypeError, "Expect _params_ an Array, but we get a #{params.class}."
+        params.size <= 0 and return nil
+
+        p = FFI::MemoryPointer.new(:pointer, params.size)
+        params.each_with_index do |x,i|
+            m = case x
+                when Fixnum
+                    FFI::MemoryPointer.new(:int).write_int(x)
+                when Float
+                    FFI::MemoryPointer.new(:float).write_float(x)
+                when CUDevicePtr
+                    ptr = FFI::MemoryPointer.new(:CUDevicePtr)
+                    API::write_cudeviceptr(ptr, x.to_api.address)
+                    ptr
+                else
+                    raise TypeError, "Invalid type of kernel parameter #{x.to_s}."
+            end
+            p[i].write_pointer(m)
+        end
+        p
+    end
+
+
+    def align_up(offset, alignment)
+        (offset + alignment - 1) & ~(alignment - 1)
+    end
+
+end
+
+end # module
+end # module

data/lib/cuda/driver/init.rb

@@ -0,0 +1,45 @@
+#
+# Copyright (c) 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/driver/ffi-cu'
+require 'cuda/driver/error'
+
+
+module SGC
+module CU
+
+class CUInit
+
+    # Initialize the CUDA driver API. This must be called before other CUDA driver functions.
+    # @param [Integer] flags Currently _flags_ must be set to zero.
+    def self.init(flags = 0)
+        status = API::cuInit(flags)
+        Pvt::handle_error(status, "Failed to initialize the CUDA driver API.")
+        nil
+    end
+
+end
+
+end # module
+end # module

data/lib/cuda/driver/memory.rb

@@ -0,0 +1,134 @@
+#
+# Copyright (c) 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/driver/ffi-cu'
+require 'cuda/driver/error'
+require 'cuda/driver/stream'
+
+
+module SGC
+module CU
+
+module CUMemory
+
+    # Copy _nbytes_ from the memory at _src_ptr_ to the memory at _dst_ptr_.
+    # The type of memory (host or device) is inferred from the pointer value.
+    def memcpy(dst_ptr, src_ptr, nbytes)
+        status = API::cuMemcpy(dst_ptr.to_api, src_ptr.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to copy memory: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy
+
+
+    # Copy _nbytes_ from the memory at _src_ptr_ to the memory at _dst_ptr_ on _stream_ asynchronously.
+    # The type of memory (host or device) is inferred from the pointer value.
+    def memcpy_async(dst_ptr, src_ptr, nbytes, stream)
+        s = Pvt::parse_stream(stream)
+        status = API::cuMemcpyAsync(dst_ptr.to_api, src_ptr.to_api, nbytes, s)
+        Pvt::handle_error(status, "Failed to copy memory asynchronously: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy_async
+
+
+    # Copy _nbytes_ from the host memory at _src_mem_ to the device memory at _dst_devptr_.
+    def memcpy_htod(dst_devptr, src_mem, nbytes)
+        status = API::cuMemcpyHtoD(dst_devptr.to_api, src_mem.ptr, nbytes)
+        Pvt::handle_error(status, "Failed to copy memory from host to device: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy_htod
+
+
+    # Copy _nbytes_ from the host memory at _src_mem_ to the device memory at _dst_devptr_ on _stream_ asynchronously.
+    #
+    # @note The _src_mem_ should be *page-locked* memory.
+    # @note Not implemented yet.
+    def memcpy_htod_async(dst_devptr, src_mem, nbytes, stream)
+        s = Pvt::parse_stream(stream)
+        status = API::cuMemcpyHtoDAsync(dst_devptr.to_api, src_mem.ptr, nbytes, s)
+        Pvt::handle_error(status, "Failed to copy memory from host to device asynchronously: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy_htod_async
+
+
+    # Copy _nbytes_ from the device memory at _src_devptr_ to the host memory at _dst_mem_.
+    def memcpy_dtoh(dst_mem, src_devptr, nbytes)
+        status = API::cuMemcpyDtoH(dst_mem.ptr, src_devptr.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to copy memory from device to host: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy_dtoh
+
+
+    # Copy _nbytes_ from the device memory at _src_devptr_ to the host memory at _dst_mem_ on _stream_ asynchronously.
+    #
+    # @note The _dst_mem_ should be *page-locked* memory.
+    # @note Not implemented yet.
+    def memcpy_dtoh_async(dst_mem, src_devptr, nbytes, stream)
+        s = Pvt::parse_stream(stream)
+        status = API::cuMemcpyDtoHAsync(dst_mem.ptr, src_devptr.to_api, nbytes, s)
+        Pvt::handle_error(status, "Failed to copy memory from device to host asynchronously: size = #{nbytes}")
+        nil
+    end
+    module_function :memcpy_dtoh_async
+
+
+    # Copy _nbytes_ from the device memory at _src_devptr_ to the device memory at _dst_devptr_ asynchronously.
+    def memcpy_dtod(dst_devptr, src_devptr, nbytes)
+        status = API::cuMemcpyDtoD(dst_devptr.to_api, src_devptr.to_api, nbytes)
+        Pvt::handle_error(status, "Failed to copy memory from device to device asynchronously: size = #{nbytes}.")
+        nil
+    end
+    module_function :memcpy_dtod
+
+
+    # Copy _nbytes_ from the device memory at _src_devptr_ to the device memory at _dst_devptr_ on _stream_ asynchronously.
+    #
+    # @note Not implemented yet.
+    def memcpy_dtod_async(dst_devptr, src_devptr, nbytes, stream)
+        s = Pvt::parse_stream(stream)
+        status = API::cuMemcpyDtoDAsync(dst_devptr.to_api, src_devptr.to_api, nbytes, s)
+        Pvt::handle_error(status, "Failed to copy memory from device to device asynchronously: size = #{nbytes}.")
+        nil
+    end
+    module_function :memcpy_dtod_async
+
+
+    # @return [Hash{ :free, :total }] A hash with the amount of free and total device memory in bytes.
+    def mem_info
+        pfree = FFI::MemoryPointer.new(:size_t)
+        ptotal = FFI::MemoryPointer.new(:size_t)
+        status = API::cuMemGetInfo(pfree, ptotal)
+        Pvt::handle_error(status, "Failed to get memory information.")
+        { free: API::read_size_t(pfree), total: API::read_size_t(ptotal) }
+    end
+    module_function :mem_info
+
+end
+
+end # module
+end # module