triton.experimental.gluon.language.bank_conflicts

triton.experimental.gluon.language.bank_conflicts(distr_ty, shared_ty, _semantic=None) → int

Count the bank conflicts per wavefront of each instruction generated when reading/writing the distributed tensor from/to the shared memory descriptor using ld.shared/st.shared instructions.

We define a bank conflict of N to be the excess number of memory accesses that each wavefront needs to access the shared memory descriptor. When one uses no ld/st vectorization, this is equal to t he number of excess memory accesses per instruction.

Parameters:

• distr_ty (distributed_type) – The distributed tensor.

• shared_ty (shared_memory_descriptor_type) – The shared memory descriptor.

Returns:

The number of bank conflicts.

Return type:

int