Module throughput 

Throughput utilities.

Modules

compute_cmma

compute_direct

launch_overhead

memory_direct

memory_probe

memory_read

memory_write

Structs

LaunchConfig

Hardware execution parameters for launching a compute kernel.

Functions

compute_bound

The compute half of a roofline: work’s operations against the peak of compute_key, close enough to it past threshold.

device_throughput

Measure peak throughput on device for each of the given keys.

measure_launch_overhead

What one launch costs on top of its kernel, which a time limit allows for.

measure_memory_curve

Measure the memory ceiling across a range of working sets, from a few kilobytes up to as much as the device will allocate.

measure_peak_throughput

Computes the peak throughput for a given runtime and key.

memory_bound

The memory half of a roofline: work’s bytes against the ceiling in the direction access names, at the work’s own footprint.

roofline_bounds

Both halves of the roofline for a Work amount, with the memory ceiling a copy’s.