count-send-latency
count-send-latency
Select how to model SEND instruction
latency.
Syntax
--count-send-latency
=
<string>
Arguments
<string>
is one of the
following:
Argument
| Description
|
---|---|
all
| Assume each SEND instruction has an uncovered latency.
|
first
| Assume only the first SEND instruction in a thread has an
uncovered latency.
|
off
| Do not model SEND instruction latency.
|
Default
off
Usage
Use
first
for CPU-to-GPU offload
modeling.
Use
all
for GPU-to-GPU offload
modeling.
Example
Model performance of your application on a target
device assuming only the first send instruction has an uncovered latency.
advisor --collect=projection --count-send-latency=first --project-dir=./advi_results