User Guide

Contents

atomic-access-pattern

Select atomic access pattern to model for a target device.

Syntax

--atomic-access-pattern
=
<string>
Arguments
Argument
Description
sequential
Increase atomic throughput by 16.
partial_sums_16
All threads block each other, SIMD lanes are reduced
separately
.
same
All threads block each other,
all
SIMD lanes need to be reduced.
Default
partial_sums_16
Actions Modified
collect
=projection
collect
=offload
Usage
  • Set to
    sequential
    for a reduction with 16 partial sums.
  • Set to
    partial_sums_16
    or
    same
    for a naive reduction.
Example
Run Performance Modeling and model sequential atomic access pattern.
$ advisor --collect=projection --project-dir=./advi_results --atomic-access-pattern=sequential

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.