nvopencl_wrapper now works on machine with both AMD and NVIDIA GPUs. It will automatically compile the OpenCL kernels on the NVIDIA platform.

[git-p4: depot-paths = "//depot/gpgpu_sim_research/fermi/distribution/": change = 14558]

2 files changed, 20 insertions, 4 deletions

diff --git a/CHANGES b/CHANGES
index 8119f29..d46de2b 100644
--- a/CHANGES
+++ b/CHANGES
@@ -25,7 +25,10 @@ Version 3.1.1+edits (development branch) versus 3.1.1
   '=' and ';'.  This allow options with a long string of sub-options (e.g. DRAM
   timing) to have a more friendly format. 
 - Changed the DRAM timing option to a more reader-friendly format. 
-- Add support for the PTX v3 codes generated by new OpenCL drivers. 
+- Added support for the PTX version 3.0 codes generated by NVIDIA OpenCL
+  drivers newer than version 256.40. 
+- nvopencl_wrapper now works on machine with both AMD and NVIDIA GPUs.  It will
+  automatically compile the OpenCL Kernels via the NVIDIA platform. 
 - Bug Fixes:
     - Fixed a compile error that happens with newer gcc/g++ versions (4.7.1)
     - Fixed a bug with the association between cuobjdump output and cubin
diff --git a/libopencl/nvopencl_wrapper.cc b/libopencl/nvopencl_wrapper.cc
index 73e60bb..09d83c6 100644
--- a/libopencl/nvopencl_wrapper.cc
+++ b/libopencl/nvopencl_wrapper.cc
@@ -130,16 +130,29 @@ int main(int argc, const char **argv)
    platforms = (cl_platform_id*)malloc(num_platforms * sizeof(cl_platform_id));
    errcode = clGetPlatformIDs(num_platforms, platforms, NULL);
    if ( errcode != CL_SUCCESS ) myexit(3,"clGetPlatformIDs returned %d",errcode);
-   errcode = clGetPlatformInfo(platforms[0], CL_PLATFORM_NAME, 1024, &buffer, NULL);
+   unsigned use_platform = 0; 
+   if (num_platforms > 1) {
+      char platformName[1024]; 
+      printf("%s: Multiple OpenCL platforms found.  Searching for compatible platform...\n",PREAMBLE);
+      for (unsigned p = 0; p < num_platforms; p++) {
+         errcode = clGetPlatformInfo(platforms[p], CL_PLATFORM_NAME, 1024, &platformName, NULL);
+         if ( errcode != CL_SUCCESS ) myexit(3,"clGetPlatformInfo returned %d",errcode);
+         printf("%s:     OpenCL platform \'%s\'\n",PREAMBLE, platformName);
+         if (strstr(platformName, "NVIDIA") != NULL) {
+            use_platform = p; 
+         }
+      }
+   }
+   errcode = clGetPlatformInfo(platforms[use_platform], CL_PLATFORM_NAME, 1024, &buffer, NULL);
    if ( errcode != CL_SUCCESS ) myexit(3,"clGetPlatformInfo returned %d",errcode);
    printf("%s: Generating PTX using OpenCL platform \'%s\'\n",PREAMBLE,buffer);
 
-   errcode = clGetDeviceIDs(platforms[0], CL_DEVICE_TYPE_GPU, 0, NULL, &num_devices);
+   errcode = clGetDeviceIDs(platforms[use_platform], CL_DEVICE_TYPE_GPU, 0, NULL, &num_devices);
    if ( errcode != CL_SUCCESS ) myexit(4,"clGetDeviceIDs returned %d",errcode);
    printf("%s: found %u native OpenCL devices\n",PREAMBLE,num_devices);
 
    cl_device_id *devices = (cl_device_id *)malloc(num_devices * sizeof(cl_device_id) );
-   errcode = clGetDeviceIDs(platforms[0], CL_DEVICE_TYPE_GPU, num_devices, devices, NULL);
+   errcode = clGetDeviceIDs(platforms[use_platform], CL_DEVICE_TYPE_GPU, num_devices, devices, NULL);
    if ( errcode != CL_SUCCESS ) myexit(5,"clGetDeviceIDs returned %d",errcode);
    context = clCreateContext(0, num_devices, devices, NULL, NULL, &errcode);
    if ( errcode != CL_SUCCESS ) myexit(6,"clCreateContext returned %d",errcode);