Change the determination of parameters of macro-kernel
authorRoman Gareev <gareevroman@gmail.com>
Wed, 21 Dec 2016 12:51:12 +0000 (12:51 +0000)
committerRoman Gareev <gareevroman@gmail.com>
Wed, 21 Dec 2016 12:51:12 +0000 (12:51 +0000)
commitbe5299af0b88b74bf986d223883c0992bd0d993f
tree4636ffa178e20ed0b63a26ebb0f5a437c91dbd78
parent85e12d285188de3f02c82b0bd1da8abfb5ad12f3
Change the determination of parameters of macro-kernel

Typically processor architectures do not include an L3 cache, which means that
Nc, the parameter of the micro-kernel, is, for all practical purposes,
redundant ([1]). However, its small values can cause the redundant packing of
the same elements of the matrix A, the first operand of the matrix
multiplication. At the same time, big values of the parameter Nc can cause
segmentation faults in case the available stack is exceeded.

This patch adds an option to specify the parameter Nc as a multiple of
the parameter of the micro-kernel Nr.

In case of Intel Core i7-3820 SandyBridge and the following options,

clang -O3 gemm.c -I utilities/ utilities/polybench.c -DPOLYBENCH_TIME
-march=native -mllvm -polly -mllvm -polly-pattern-matching-based-opts=true
-DPOLYBENCH_USE_SCALAR_LB -mllvm -polly-target-cache-level-associativity=8,8
-mllvm -polly-target-cache-level-sizes=32768,262144 -mllvm
-polly-target-latency-vector-fma=8

it helps to improve the performance from 11.303 GFlops/sec (39,247% of
theoretical peak) to 17.896 GFlops/sec (62,14% of theoretical peak).

Refs.:

[1] - http://www.cs.utexas.edu/users/flame/pubs/TOMS-BLIS-Analytical.pdf

Reviewed-by: Tobias Grosser <tobias@grosser.es>
Differential Revision: https://reviews.llvm.org/D28019

llvm-svn: 290256
polly/lib/Transform/ScheduleOptimizer.cpp
polly/test/ScheduleOptimizer/mat_mul_pattern_data_layout.ll
polly/test/ScheduleOptimizer/mat_mul_pattern_data_layout_2.ll
polly/test/ScheduleOptimizer/pattern-matching-based-opts_3.ll