GEMM methods leaks events if they need more kernels than command queues passed. For example, if clblasGemm is called with a single command queue and it needs all 4 kernels (tile, row, col, corner), will overwrite the event of each kernel enqueue with the event for the next.
A similar issue is probably present in other kernels (e.g. TRSV kernels, or anything else that uses makeSolutionSeq).
kknox
commented
7 years ago
GEMM methods leaks events if they need more kernels than command queues passed. For example, if clblasGemm is called with a single command queue and it needs all 4 kernels (tile, row, col, corner), will overwrite the event of each kernel enqueue with the event for the next.
A similar issue is probably present in other kernels (e.g. TRSV kernels, or anything else that uses makeSolutionSeq).
Probably related to the already merged PR #300