Module cublas 

Sys-level safe wrappers for the cuBLAS entry points cudarc 0.19 doesn’t expose through its safe layer.

Wrapped today (Phase 1 cuBLAS slice):

• cublasGemmEx, cublasGemmStridedBatchedEx
• cublasAxpyEx, cublasScalEx, cublasNrm2Ex, cublasDotEx
• cublasIamaxEx, cublasIaminEx, cublasAsumEx
• cublasCopyEx, cublasSwapEx, cublasRotEx
• cublasGemv_v2/cublasDgemv_v2, cublasSger_v2/cublasDger_v2
• cublasSgeam/cublasDgeam
• cublasSsyrk_v2/cublasDsyrk_v2
• cublasStrsm_v2/cublasDtrsm_v2

All callers must hold the cuBLAS handle’s stream current on the same OS thread. The atomr-accel-cuda actor pipeline guarantees that via GpuDispatcher.

Functions

asum_ex^⚠

axpy_ex^⚠

copy_ex^⚠

dgeam^⚠

dgemv^⚠

dger^⚠

dot_ex^⚠

dsyrk^⚠

dtrsm^⚠

gemm_ex^⚠

cublasGemmEx — typed-erased gemm with a separate compute type.

gemm_strided_batched_ex^⚠

cublasGemmStridedBatchedEx — typed-erased strided-batched gemm.

iamax_ex^⚠

iamin_ex^⚠

nrm2_ex^⚠

rot_ex^⚠

scal_ex^⚠

sgeam^⚠

cublasSgeam / cublasDgeam — matrix add/scale: C = α·op(A) + β·op(B).

sgemv^⚠

sger^⚠

ssyrk^⚠

strsm^⚠

swap_ex^⚠