gpu_residency.hpp File Reference

Decide whether N frames of the analysis working set fit on the GPU. More...

#include <cstddef>
#include <string>

Go to the source code of this file.

Classes
struct	gpu::DeviceInfo
struct	gpu::Footprint
struct	gpu::Plan

Namespaces
namespace	gpu

Functions
DeviceInfo	gpu::probeDevice ()
Footprint	gpu::estimateFootprint (int nAtoms, int nFrames, int kMax=16, int maxSixRingsPerAtom=16)
int	gpu::maxResidentFrames (const DeviceInfo &dev, int nAtoms, int kMax=16, int maxSixRingsPerAtom=16, double safety=0.80)
Plan	gpu::planBatch (int nAtoms, int requestedFrames, int kMax=16, int maxSixRingsPerAtom=16, double safety=0.80)

Detailed Description

Decide whether N frames of the analysis working set fit on the GPU.

The offload unit is a batch of frames of the TUM ice-score path. A gpulite-style runtime probe reports free device memory. The footprint is positions, the mutual-graph CSR, the six-ring CSR, affiliation flags and per-atom HC/DDC labels. CHILL and (q_{lm}) are not in this working set.

Definition in file gpu_residency.hpp.