SWAN Version 0.1 12 Mar 2010

THIS TOOL IS UNSUPPORTED AND NO LONGER USED BY US. DO NOT ASSUME THAT IT WORKS WITH CURRENT CUDA VERSIONS. JUST TAKE THE CODE AS AN INDICATION OF HOW IT WORKED.

INTRODUCTION

Swan is a simple tool for aiding a port from CUDA to OpenCL. It has the following benefits:

• Reduce reliance on nvcc for CUDA code by replacing the CUDA runtime extensions
• Automatic code generation of entry point functions for kernels, to replace <<<>>> syntax
• Source translation of CUDA kernels to OpenCL.
• Common API for both CUDA and OpenCL targets.

BUILDING

• In the top level directory (tld), edit config.mk and set the paths appropriately OpenCL and CUDA libraries may be selectively enabled with "OPENCL=yes|no" and "CUDA=yes|no"

• Set LD_LIBRARY_PATH to point to libOpenCL and other libs.

• Set PATH to point to (tld)/bin

• To build library and examples: $ cd (tld) $ make

  If all went well, you will have lib/libswan_(cuda|ocl).(so|a) and CUDA and/or OpenCL versions of each example (.nv and .ocl suffices)

• Build just the examples (having already built the library): $ cd (tld) $ make examples

EXAMPLES

Each example will build in OpenCL and CUDA variants, with suffices .ocl and .nv respectively.

• examples/vecadd A simple vector addition

• examples/vecadd_global A simple vector addition showing use of device/constant global data

• examples/shmem Shows use of dynamic shared memory

• examples/interp 1D Texturing, with arrays and linear memory (CUDA only)

PORTING EXISTING CUDA CODE

How to use in your own code:

• Put the kernels in a separate file, e.g. kernel.cu

• Include the compiled code in the host source: #include "kernel.kh"

• Replace CUDA-style kernel launches with calls to the entry point functions. The formal definitions of these functions are sytematically derived from the kernel source. For example, a kernel:

  global void func1( float4 *in, int *out, int N );

  will have the following entry points:

  void k_func1( block_config_t grid, block_config_t block, int shmem, float4 *in, int *out, int N );

  and

  void k_func1_async( block_config_t grid, block_config_t block, int shmem, float4 *in, int *out, int N );

  grid, block and shmem have the same definitions as in the CUDA launch syntax <<< grid, block, shmem >>>. block_config_t has 3 elements .x,.y,.z, like a dim3.

• Replace cuda API calls with Swan equivalents (See swan_api.h)

GENERATING CUDA / OPENCL EXECUTABLES

• For cross-compiling CUDA kernels to OpenCL: swan --opencl kernel.kh kernel.cu

• For compiling CUDA kernels to CUDA: swan --cuda kernel.kh kernel.cu

• For compiling OpenCL kernels to OpenCL: swan --opencl-direct kernel.kh kernel.cu

  This will produce a source code header file that contains a binary object containing the compiled kernel source and a set of functions for calling the kernels.

• Compile

• Link against the appropriate swan library (libswan_ocl or libswan_cuda).