INTEL Workshop on "in-core" performance

Code examples

SMatrix

The other examples (see Attachments) are compiled like this: g++ -I standaloneMinimal/root/ fse_Assign_Sim_M_C_add_Sim_K_V.cxx

The data size can be changed in some of the examples with -DDIM_L=, ex: g++ -I standaloneMinimal/root/ fse_Assign_Sim_M_C_add_Sim_K_V.cxx -DDIM_L=2

-- RuneJensen - 2009-09-07

Geant4

• G4AffineTransform.zip: G4AffineTransform. Several code versions included, also SSE intrinsic.

This code shows usage of SSE intrinsics (double precision), but little runtime improvement.

-- RuneJensen - 2009-09-08

CMS geometrical transformations

• match.tar.bz2: This is an example of matrix-vector multiplication in a real world context. Part of the "project" routine of CMSSW has been rewritten using intel SIMD instrinsics. The performance of the original code has been compared to the SIMD code (read perf_res.txt).

-- VincenzoInnocente - 2009-09-06

• match_new.tar.bz2: New version of the matrix-multiplication code! SIMD code is unchanged but benchmarking is done more carefully using Rune's trick. Stalls are thus removed and finally SIMD code shows its benefits in performance (perf_res.txt).

• example.tar.bz2: Yet another version of the match example, including the missing part from the original CMSSW code.