This update includes various fixes and the following features:
preliminary support for block-pointers
block-pointers are lowered to memref buffers similarly to traditional triton pointer loads
support lowering for triton.get_num_programs by introducing extra arguments when launching triton kernels similarly to triton.get_program_id
improved lowering triton.reduce which previously only supports lowering float values
preliminary support for pointer arithmetic involving the modulo operator
the modulo operator use case is seen in the tutorial matmul example, where pointer offsets are being modded to prevent loading out-of-bound values; in such case, these values are wrapped around to the beginning of the buffer
This update includes various fixes and the following features:
triton.get_num_programs
by introducing extra arguments when launching triton kernels similarly totriton.get_program_id
triton.reduce
which previously only supports lowering float values