Question

I made a macro to simplify CUDA kernel calls:

#define LAUNCH LAUNCH_ASYNC

#define LAUNCH_ASYNC(kernel_name, gridsize, blocksize, ...) \
    LOG("Async kernel launch: " #kernel_name);              \
    kernel_name <<< (gridsize), (blocksize) >>> (__VA_ARGS__);

#define LAUNCH_SYNC(kernel_name, gridsize, blocksize, ...)     \
    LOG("Sync kernel launch: " #kernel_name);                  \
    kernel_name <<< (gridsize), (blocksize) >>> (__VA_ARGS__); \
    cudaDeviceSynchronize();                                   \
    // error check, etc...

Usage:

LAUNCH(my_kernel, 32, 32, param1, param2)

LAUNCH(my_kernel<int>, 32, 32, param1, param2)

This works fine; with the first define I can enable synronous calls and error checking for debugging.

However it does not work with multiple template arguments like below:

LAUNCH(my_kernel<int,float>, 32, 32, param1, param3)

The error message I get in the line where I call the macro:

error : expected a ">"

Is it possible to make this macro work with multiple template arguments?

Answer 1

The problem is that the preprocessor knows nothing about angle bracket nesting, so it interprets the comma between them as macro argument separator.

If the kernel-launch syntax supports parentheses around the kernel name (I can’t check now, not on a CUDA machine), you could do this:

LAUNCH((my_kernel<int, float>), 32, 32, param1, param3)