SYCL port

Jump to bottom Edit New page

maddyscientist edited this page Dec 15, 2021 · 2 revisions

Development is in the sycl branch https://github.com/lattice/quda/tree/feature/sycl

Changes from develop can be seen in the PR https://github.com/lattice/quda/pull/1168

Outstanding changes/issues

BlockReduce calls simplified to make it easier to implement in SYCL
reducer_t types added for reductions (reducer.h, transform_reduce.cuh)
multi blas Args may not fit in max_kernel_arg_size (using max_constant_size instead)
quda_target.h needs to be included from quda_internal.h
block size in dslash_coarse kernel must evenly divide threads
~~FAST_COMPILE_REDUCE version of block_orthogonalize.cu and restrictor.cu can't go larger than max_block_size~~