icctips.txt

Intel Compiler Tips (v11.1.072)
February 3rd, 2011

General:
-Remember to deactivate -g for final version. Disables optimization flags.
-Try profile guided compilation, after debugging, using datasets of normal size.
-Try switching to -fp-model fast=2 (instead of default fast=1) for more aggressive fp optimizations.
-Use #pragma loop_count (n) and #pragma loop_count min(n), max(n), avg(n) to indicate approximate number of iterations for loops.
-Use #pragma distribute point and #pragma parallel.
-Try using -parallel to activate automatic parallelizations.
-Compile with -openmp-report 2 -opt-report 3 -vec-report 3 (and -par-report 3) and use reports to optimize with pragmas.

Specific:
-Use Intel Thread Profiler.

Test:
-Try -(a)xT and -(a)xP against -(a)xHOST.

Notes:
-Change -x(option) to -m(option) for non-Intel processors.

Cases:
-Copying with copy and long strides more efficient than with fors. Generally for long stride vs for loop.
-Use of COPY more efficient than memcpy.