Batching models

GLASS is single-block by design: you launch one block per independent problem, and the block’s threads cooperate over data already in shared/global memory. But many callers have many small problems to run at once, and GLASS offers several distinct ways to pack them. They differ in how problems map onto threads/warps/blocks and in what stride/offset metadata they take — pick by how your problems are laid out in memory and how many threads you can give each one.

Table 3 The batching models

Model	Functions	When to use
One block per problem (the base model)	every plain op (gemm / gemv / syrk / posv / …)	The default. One <<<num_problems, threads>>> launch; each block owns one problem. Use when each problem is big enough to keep a block busy.
Block-per-problem, strided sub-blocks	gemm_strided / gemv_strided (and axpy_strided / copy_strided for L1)	One block still owns one problem, but the operands are sub-blocks of a larger column-major matrix addressed with explicit leading dimensions (row strides). Use to operate on a tile in place without copying it out.
Flattened-batch, block-stride indexed	gemm_batched_indexed	A flat batch of GEMMs whose operands live at arbitrary offsets given by an index array; blocks stride over the batch. Use for gather-style batches where problem i’s pointers are not a fixed stride apart.
K-way fused, interleaved in one block	inv / potrf (K-way overloads; inv2 / inv3 wrappers)	A single block factors/inverts K independent matrices at once by interleaving their sweeps over one shared row loop. Use when K is small and each matrix is too small to fill a block alone — fills the block by fusing across problems instead of across rows.
TC-group SIMT-1D batched	the batched-1D GEMM APIs (see Batched-1D GEMM APIs)	Designed for kernels with a single 1D thread block that must run a batch of tiny GEMMs; threads are partitioned into per-problem groups. Use inside an existing 1D-block kernel that can’t relaunch.
Warp-per-problem	glass::warp:: ops (dot / gemv / trsv / potrf / posv / …)	One 32-lane warp owns one problem; pack many warps into a block to run many problems concurrently (threadIdx.y selects the warp). Use for the smallest problems (N ≈ 7–32) where a whole block per problem would waste lanes. See Contraction-parallel ops (the *_reduced family).
Segmented	gemv_segmented	One matvec whose output rows are partitioned into contiguous segments with per-segment row counts. Use for ragged/block-structured matvecs.

Naming convention

All batching variants follow operation_qualifier — the operation first, the batching qualifier as a suffix (gemm_strided, gemv_segmented, gemm_batched_indexed), and the file name matches the function name. The qualifier names how the problems are packed, never the scope (scope is the namespace — glass:: vs glass::warp::; see Namespaces, suffixes, and flags).