This document contains the FAQ asked by our customers about Roofline feature of Intel® Advisor.
How to use Roofline analysis?
How do you count FLOPS?
How to do roofline analysis for multithreaded application?
How high KNL bandwidth memory is treated in Roofline?
Does ITT* API works for roofline feature?
How to store Roofline analysis data to file?
Does the command line interface to specify only a particular loops works for roofline analysis?
How to use Roofline analysis?
It is described in the article https://software.intel.com/en-us/articles/getting-started-with-intel-advisor-roofline-feature
How do you count FLOPS?
We use classic approach for computing FLOPS.
We count fp-operations +,-,*,/ as 1 and FMA as 2 operations.
If a SIMD fp-operation is used number fp-operation is multiplied by number of vector elements used in the operation.
Single and double precision FLOPS are not counted into separate metrics, the single cumulative metrics is computed.
We count actual instructions got executed.
So, if math function is an instruction then it’s counted as 1 (exp, rsp, sqrt). If it is a math library call (sin, cos, …. ) then we count actual number of compute instructions in Chebyshev/Taylor decomposition in the routine implementation. But the FLOPS data will be assigned to the function call itself and not to the loop called the function. We call it selftime-based FLOPS counting. More information on selftime-based FLOPS counting can be found in the article https://software.intel.com/en-us/articles/selftime-based-flops-computing-vectorization-advisor.
How to do roofline analysis for multithreaded application?
...
How high KNL bandwidth memory is treated in Roofline?
...
How to store Roofline analysis data to file?
This is not possible so far. Contact vector_advisor@intel.com if you feel this is a critical feature for you.
Does ITT* API works for roofline feature?
It should work in general. Could be problems for Trip Counts support.But it works for Survey, therefore FLOPS will not be computed for excluded loops.
Does the command line interface to specify only a particular loops works for roofline analysis?
No. It is not implemented. Contact vector_advisor@intel.com if you feel this is a critical feature for you.