http://www.gdiamos.net/classes/translator/api/index.htmlOcelot: A Binary Translation Framework for PTX
0.3.996IntroductionThe emergence of GPGPU programming models such as CUDA and OpenCL provide avenues for general purpose application developers to leverage the computation capabilities of highly parallel SIMD GPU architectures. PTX defines a machine model and virtual ISA for these architectures that is used commercially by NVIDIA GPUs, but is also representative of other SIMD machine models. Ocelot seeks to develop a set of tools that enable the low level analysis of GPGPU applications as well a providing a JIT compiler for generic architectures. We have completed control and dataflow analysis modules for PTX as well as an emulator for the PTX 1.3-1.4 virtual machine compatible with CUDA 2.1-2.2 applications. We are currently working on a back end binary translator from PTX to various architectures.InstallationRequirementsThe following compiler tools are required (tested):
The following libraries are required to be installed:
The following external library headers are required to be on your search path:
We require a compiler that supports parts of the upcoming C++0x standard. Ocelot is rigorously tested Ubuntu 9.04 using GCC-4.3.2. We strive to be ANSI C++ compliant, but only the specified systems is actively tested. InstructionsFollow the standard autotools flow.
We assume that you have cuda installed in /usr/local/cuda. If not specify a different path with
Expected UsageOcelot as it stands includes a full featured parser for PTX 1.3-1.4, an emulator for PTX 1.3-1.4, and a complete implementation of the Cuda Runtime API 2.2 as defined in PATH_TO_CUDA/include/cuda_runtime_api.h . It also includes for generating the control flow graphs and dominator trees of PTX programs. There are several trace generator modules and a trace generator interface to the emulator that supports a callback after each emulated event.Compiling Ocelot will generate a libtool library libcudart.la that is a drop in replacement for NVIDIA's libcuda.so . Libtool allows us to convert the library to a static libcudart.a or a shared object libcudart.so . Note that Ocelot is built from several internal modules, so you will have to link against every module that you use rather than simply cudart. TestingOcelot has been validated against the entire CUDA 2.1 and 2.2 SDKs, including streams, events, asynchronous operations, multidimensional memory allocation, shared host/device memory, atomic/vote/reduction operations, and even OpenGL. It has also been validated against GPU-VISPL and several other internal applications.Under the tests directory, there are two separate projects cuda2.1 and cuda2.2 that build the CUDA SDKs and link them against Ocelot. Running make test in these directories will run a regression test for each SDK. CaveatsTo use Ocelot with any pre-compiled CUDA libraries (such as CUFFT or CUBLAS), the libraries must be compiled as shared objects and must be linked in the correct order. The order is PRECOMPILED_LIBRARIES OCELOT_LIBRARIES YOUR_PROGRAM . This is REQUIRED to ensure that global constructors are called in the correct order. It is an artifact of how CUDA is designed and impossible for us to change.The warp size in Ocelot is variable and can change depending on the version being used. Any code that relies on the warp size being 32 without checking the device properties will break. The implementation of texture instructions is not the same as NVIDIA's. We believe that our implementation is correct as far as the instructions are defined, however, the results will not be bit-accurate. The implementations of floating point rounding for single precision instructions is probably different from NVIDIA's. |
