- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
For some reason points on the roofline model corresponding to my loops are out of bounds of the roofline model
How it is possible? How to interpret this result? Microarchitecture analysis of vtune amplifier shows a lot of space for optimization.
Link Copied
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hi,
Could You please provide project directory?
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
Hello!
What is a total duration of your test case? Does it run (without advisor) for 1 or 2 seconds, for several minutes or for several hours?
The vertical coordinate of dots is FLOP/Seconds. If seconds --> 0 , then FLOP/Seconds --> infinite. The only case when seconds --> 0 is when the time spent in those kernels is smaller or same as "sampling interval" (10-50 milliseconds). In cases like that hotspots statistical sampling will not properly capture the time spent in the kernels and therefore FLOP/S will be "exaggerated".
You have to run your code for at least 10 or 20 seconds to make the case statistically representative and have correct FLOP/S data.
- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page