PV1 ⊧ GPUDrano, sometimes stylized as GPU Dranodetects uncoalesced memory accesses in CUDA programs

Static analysis tool to detect uncoalesced global memory accesses in CUDA programs

Application domain/field

• Graphics Processing Units (GPUs)
• Static analysis
• Uncoalesced memory accesses
• Performance bugs
• Dataflow analysis
• Abstract interpretation

Type of tool

Bug detector?

Expected input

LLVM IR

Format:

.ll file(s)

Expected output

GPUDrano will report all accesses that might be uncoalesced for each GPU kernel.

Internals

Load and store instructions that reference the GPU's global memory, should obey a certain structure to ensure optimal performance. If not, it might result in underutilization and performance problems. GPUDrano tries to find uncoalesced accesses. Uncoalesced memory accesses occur when the the memory accesses are not contained within a single cache block.