Re:AMA: Explaining my 750 line compiler+runtime designed to ...
by pavanky on Sunday March 05, 2017 @11:46PM
(#13800759)
Not the author but as someone who worked on the arrayfire library I think I have some input to this question.
Currently a lot of math functions are already compiled using a fairly simple just in time compiler inside arrayfire. This results in complex mathematical operations being fused into single kernels.
That said, the number of functions that are supported this way are fairly limited and I am going to try and add more of them.
Anyway here are a couple of links that may be helpful in trying to understand the overheads involved. The code shown [2] compiles to a single CUDA / OpenCL kernel.
Currently a lot of math functions are already compiled using a fairly simple just in time compiler inside arrayfire. This results in complex mathematical operations being fused into single kernels.
That said, the number of functions that are supported this way are fairly limited and I am going to try and add more of them.
Anyway here are a couple of links that may be helpful in trying to understand the overheads involved. The code shown [2] compiles to a single CUDA / OpenCL kernel.
[1] https://arrayfire.com/benchmarking-parallel-vector-libraries...
[2] https://arrayfire.com/performance-improvements-to-jit-in-arr...