What is maximum size of kernel code?

Drmike · February 18, 2010, 3:21pm

I implemented a multiple-precision (MP) float and integer library for CUDA. I use these MP functions within the kernel listed below. Thereby, the kernel computes wrong values for element k of vector c (it remains 0), if the thread block dimension is greater than 11x11. In device emulation mode it works. In addition, if the code is compiled for the device with the -g -G flags, then it works correctly too.

Nevertheless, element k of vector c is computed only once at Line 122 (last instruction); it depends on vector b2 and s (vectors b2 and s are computed correctly). It seems that the kernel is too large for thread block dimensions greater than 11x11. Afterward, I broke this kernel into two and then it worked.

As a consequence, it seems that that the maximum number of instructions per kernel depends on the thread block dimension. Is there an explanation for this behavior?

I have not yet started to convert my serial code to CUDA version, but I am also curious about the maximum kernel size possible. Most of my subroutines are very long polynomials and I’m wondering about how large a kernel can get. It looks like trial and error is the only way to find out for now??

avidday · February 18, 2010, 4:07pm

It is stated in Appendix A of the programming guide - 2 million PTX instructions per kernel.

Drmike · February 18, 2010, 4:44pm

Found it! I was searching for “instruction size” not “kernel size”. That’s huge, I should not have any problems! At least getting my kernels to fit - getting them to run correctly is another story…

Thanks avidday. Much appreciated.