batchlas::ops::spmm Namespace Reference

Functions

candidates()

template<class T>

auto batchlas::ops::spmm::candidates ( )

constexpr

Every compiled choice, in tie-break order (§6.3).

Definition at line 20 of file choice.hh.

Types

SpmmChoice

using batchlas::ops::spmm::SpmmChoice = std::variant<Direct, Vendor>

fieldless: body (by transA), column block, pair load derived

Definition at line 17 of file choice.hh.

Variables

last_resort

std::array<std::string_view, 2> batchlas::ops::spmm::last_resort {"vendor", "direct"}

inlineconstexpr

vendor: every format

Definition at line 22 of file choice.hh.

spec

select::OpSpec batchlas::ops::spmm::spec {Op::spmm, select::Lib::sparse, {last_resort}}

inlineconstexpr

op, library, rules

Definition at line 23 of file choice.hh.

key_names

std::array<std::string_view, 5> batchlas::ops::spmm::key_names

inlineconstexpr

Initial value:

{
    "transA:exact", "transB:exact", "m:log", "nrhs:log", "batch:log"}

ConjTrans folds to T; m = A.rows() as stored; no nnz key (device memory). Work ~ m nrhs batch.

Definition at line 26 of file choice.hh.

grid_m

std::array<int, 5> batchlas::ops::spmm::grid_m {1, 16, 256, 4096, 65536}

inlineconstexpr

The transcriber (tuned/README.md) spells the same grid; the transcribed predicates read no extent.

Definition at line 30 of file choice.hh.

grid_nrhs

std::array<int, 5> batchlas::ops::spmm::grid_nrhs {1, 2, 4, 16, 64}

inlineconstexpr

Definition at line 31 of file choice.hh.

grid_batch

std::array<int, 5> batchlas::ops::spmm::grid_batch {1, 8, 128, 1024, 16384}

inlineconstexpr

Definition at line 32 of file choice.hh.

Classes
struct	Direct
	sycl_spmm::spmm_native_csr: CSR only, no GPU gate (native_cpu too) More...
struct	Vendor
	backend::spmm_vendor (cuSPARSE, rocSPARSE, netlib) less known-defect shapes More...