Using High-Density Mesh (HDM) solution selection with hipBLASLt#
hipBLASLt can select GEMM kernels from a High-Density Mesh (HDM) library, which is a dense, precomputed lookup table that maps problem sizes (M, N, batch count, K) to the best-performing kernel found by exhaustive benchmarking.
Unlike prediction-based selections, which use a model to estimate the best kernel, HDM selection returns the actual benchmark winner for a close problem size in the table.
Note
HDM kernel selection is also faster than prediction-based selection and generally results in superior performance for the GEMM types it covers.
Enabling HDM selection#
Set the TENSILE_USE_MESHBASED environment variable to 1 before launching
your application:
export TENSILE_USE_MESHBASED=1
When enabled, hipBLASLt checks whether an HDM library exists for the current GEMM type. If one exists, HDM is used for kernel selection. If not, selection falls through to the next available library (for example, Origami prediction or the free-size fallback), so enabling the variable is always safe.
To disable HDM selection, unset the variable or set it to 0:
export TENSILE_USE_MESHBASED=0
Verifying HDM selection is active#
To confirm that HDM selection is being used, set the TENSILE_DB debug
variable to 0x8006. This prints the selected kernel name and its matching
tag for each GEMM call:
export TENSILE_USE_MESHBASED=1
export TENSILE_DB=0x8006
When HDM selection is active, the output includes [MatchingTag: ContractionSolution::MatchingTag::MeshBased]:
Running kernel: <kernel_name> [MatchingTag: ContractionSolution::MatchingTag::MeshBased]