[CPU] Dump human-readable asm code in TRITON_CACHE_DIR

[minjang]

I think it'd be much better to dump generated assembly code in the cache directory.

.asm file is now dropped at TRITON_CACHE_DIR.

> % ll /home/minjang/temp/triton-cache/888bdd8e7af20d785cf333479fbae3dd400aca0d11d7ceec5739759a8b2da4b2/
total 276
-rw-r--r-- 1 minjang users 169178 Jun 10 15:27 add_kernel.asm
-rw-r--r-- 1 minjang users  10724 Jun 10 15:27 add_kernel.bc
-rw-r--r-- 1 minjang users    394 Jun 10 15:27 add_kernel.json
-rw-r--r-- 1 minjang users  43591 Jun 10 15:27 add_kernel.llir
-rw-r--r-- 1 minjang users  37239 Jun 10 15:27 add_kernel.ttcir
-rw-r--r-- 1 minjang users   3418 Jun 10 15:27 add_kernel.ttir
-rw-r--r-- 1 minjang users    823 Jun 10 15:27 __grp__add_kernel.json

[ienkovich]

Why do you want to generate ASM in the common compiler.py instead of the backend make_bc method? If it's used for debugging purposes only, then there is no need to put it to metadata_group.

[minjang]

Why do you want to generate ASM in the common compiler.py instead of the backend make_bc method? If it's used for debugging purposes only, then there is no need to put it to metadata_group.

Oh yeah, my bad. My intention was to create an asm file in the cache directory. It turned out I can simply do it under make_bc. Fixed it. It's still behind TRITON_CPU_ASM_DUMP as it turned out translate_to_host_asm may take a long time. Now the change is only under make_bc.

[ienkovich]

Yes, translate_to_host_asm actually runs the whole LLVM compilation pipeline, so it's an additional compilation of the kernel. I think we had a task to change the compilation pipeline, so we can eventually transition to a better flow: LLVM MLIR -> LLVM IR -> ASM -> shared object.

Your patch looks good though! Thanks!