Fix arch guards in a few examples #2567
Conversation
Some examples build their own kernel layer (FMHA, DistGEMM), and they must guard them according to the supported features. Issue: NVIDIA#2559
The cmake config already guards targets by checking if we're building for SM100A (although unsure whether it's an exact match or not). For safety, it's best to just have the arch-specific MMA guards in place.
| @@ -507,6 +507,9 @@ struct Sm100FmhaMlaKernelTmaWarpspecialized { | |||
|
|
|||
|
|
|||
| CUTLASS_DEVICE void operator()(Params const& params, char* smem_raw) { | |||
| #if ! defined(__CUDA_ARCH_FEAT_SM100_ALL) | |||
There was a problem hiding this comment.
@alihassanijr , this only covers 100a but not 100f. You could take a look at launch control header file for the 100f macro.
There was a problem hiding this comment.
Wouldn't using the family macros break builds with CTK 12.8 and earlier?
There was a problem hiding this comment.
Both CUTLASS_ARCH_MMA_SM100F_SUPPORTED and CUTLASS_ARCH_MMA_SM100F_ENABLED are conditioned on CTK >= 12.9, so Sm100 users with CTK 12.8 will wind up with empty kernels.
There was a problem hiding this comment.
yes. so if < 12.9, 100a; else 100a || 100f
There was a problem hiding this comment.
or is CUDA_ARCH_FAMILY(1000) just a false in 12.8?
There was a problem hiding this comment.
You're right; defined(CUTLASS_ARCH_MMA_SM100A_ENABLED) || defined(CUTLASS_ARCH_MMA_SM100F_ENABLED) should work -- they're already conditioned on the correct CTK compiler version.
Some examples build their own kernel layer (FMHA, DistGEMM), and they must guard them according to the supported features.
Issue: #2559
Will close: #2559 #2558
CC @hwu36