GPU assembler generation diff vs compilers

[morousg]

I would like to see how different compilers generate different SASS code, and check:

1. Resulting performance
2. Differences in decisions made about function calls with tuples. (local memory accesses)
3. Register usage (reusing more or less registers)
4. Total amount of instructions
5. Amount of FMMA and other instructions

[morousg]

@albertandaluz it would be great to create a GPU-enabled Docker container that has Compiler Explorer installed, with both CUDA and Clang-CUDA.

Add the FKL (LTS) as header based

Let's limit it to x86 and Desktop GPU's on Linux (could be WSL2)