creating branch for adding support for CUDA 3.x and Fermi

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 6829]

1 files changed, 43 insertions, 0 deletions

diff --git a/benchmarks/OpenCL/README b/benchmarks/OpenCL/README
new file mode 100644
index 0000000..1fa6cee
--- /dev/null
+++ b/benchmarks/OpenCL/README
@@ -0,0 +1,43 @@
+To run OpenCL applications on GPGPU-Sim, do the following:
+
+1. Build the OpenCL application as you would to run it on an NVIDIA GPU.
+2. Set GPGPUSIM_ROOT to the top level of your GPGPU-Sim installation.
+3. Set NVOPENCL_LIBDIR to the location of libOpenCL.so distributed by NVIDIA
+   (usually /usr/lib or /usr/lib64)
+4. Run make in the top level GPGPU-Sim directory (after step 3)
+5. Ensure GPGPUSIM_ROOT/lib is before /usr/lib and /usr/lib64 in your 
+   LD_LIBRARY_PATH
+6. Copy appropriate gpgpusim.config and interconnect configuration files 
+   to the directory you will run the OpenCL application in.
+7. Run the OpenCL application as built in step 1.
+
+The following NVIDIA Compute SDK 2.3a OpenCL applications have been tested 
+(using reduced input sets):
+
+oclBandwidthTest
+oclBlackScholes
+oclReduction
+oclDCT8x8
+oclDotProduct
+oclHistogram
+oclMersenneTwister
+oclMatrixMul
+oclMatVecMul 
+oclSortingNetworks
+oclTranspose
+oclVectorAdd
+oclConvolutionSeparable
+oclQuasirandomGenerator
+oclParticles (w/ -qatest)
+
+The LLVM based PTX compiler in the current OpenCL enabled NVIDIA driver
+(190.29) will occassionally move code in such a way that a barrier becomes
+control dependent upon a branch that can diverge, leading to a deadlock. This
+has been observed for oclScan and oclRadixSort.  
+
+The following are known to not work currently:
+
+oclRadixSort		: barrier control dependence deadlock 
+oclScan			: barrier control dependence deadlock
+
+The other benchmarks have not been tested, or not tested to completion.