diff options
| -rw-r--r-- | libcuda/cuda_runtime_api.cc | 24 |
1 files changed, 20 insertions, 4 deletions
diff --git a/libcuda/cuda_runtime_api.cc b/libcuda/cuda_runtime_api.cc index cbe8a11..d8e7832 100644 --- a/libcuda/cuda_runtime_api.cc +++ b/libcuda/cuda_runtime_api.cc @@ -336,10 +336,23 @@ class _cuda_device_id *GPGPUSim_Init() prop->minor = 2; prop->totalGlobalMem = 0x80000000 /* 2 GB */; prop->memPitch = 0; - prop->maxThreadsPerBlock = 512; - prop->maxThreadsDim[0] = 512; - prop->maxThreadsDim[1] = 512; - prop->maxThreadsDim[2] = 512; + + if(prop->major >= 2) { + prop->maxThreadsPerBlock = 1024; + prop->maxThreadsDim[0] = 1024; + prop->maxThreadsDim[1] = 1024; + } + else + { + prop->maxThreadsPerBlock = 512; + prop->maxThreadsDim[0] = 512; + prop->maxThreadsDim[1] = 512; + } + + prop->maxThreadsDim[2] = 64; + prop->maxGridSize[0] = 0x40000000; + prop->maxGridSize[1] = 0x40000000; + prop->maxGridSize[2] = 0x40000000; prop->maxGridSize[0] = 0x40000000; prop->maxGridSize[1] = 0x40000000; prop->maxGridSize[2] = 0x40000000; @@ -352,6 +365,9 @@ class _cuda_device_id *GPGPUSim_Init() #if (CUDART_VERSION >= 2010) prop->multiProcessorCount = the_gpu->get_config().num_shader(); #endif +#if (CUDART_VERSION >= 4000) + prop->maxThreadsPerMultiProcessor = the_gpu->threads_per_core(); +#endif the_gpu->set_prop(prop); the_device = new _cuda_device_id(the_gpu); } |