How to compile PTX code

Question

I need to modify the PTX code and compile it directly. The reason is that I want to have some specific instructions right after each other and it is difficult to write a cud

Answer 1

You can use NVTRC - it's easy!

_{Expanding on @ArtemB's answer:}

nVIDIA offers a real-time compilation (RTC) library. There's an example of how it's used as part of the CUDA samples; you can access it here.

The sample actually starts from CUDA code, but the intermediary step is creating a PTX code as a plain C string (`char *). From there, this is what you do, basically:

char* ptx;
size_t ptxSize;

// ... populate ptx and ptxSize somehow ...

CUcontext context;
CUdevice cuDevice;

// These next few lines simply initialize your work with the CUDA driver,
// they're not specific to PTX compilation
cuInit(0);
cuDeviceGet(&cuDevice, 0); // or some other device on your system
cuCtxCreate(&context, 0, cuDevice);

// The magic happens here:
CUmodule module;
cuModuleLoadDataEx(&module, ptx, 0, 0, 0));

// And here is how you use your compiled PTX
CUfunction kernel_addr;
cuModuleGetFunction(&kernel_addr, module, "my_kernel_name");
cuLaunchKernel(kernel_addr, 
   // launch parameters go here
   // kernel arguments go here
);

Notes:

• I've removed all error checks so as not to clutter the example code - but do check for errors in your code.
• You'll need to link your program with the NVRTC library - it's separate from the main CUDA and CUDA driver libs. On linux, it's called libnvrtc.so.