systemml git commit: [SYSTEMML-1903] Fix codegen row candidate exploration w/ row indexing

2017-10-16 Thread mboehm7
Repository: systemml Updated Branches: refs/heads/master 5adb330de -> 3a7f38e58 [SYSTEMML-1903] Fix codegen row candidate exploration w/ row indexing For Autoencoder w/ batch=512, the codegen fuse-all heuristic was failing on cplan construction due to unknown row type. The root cause was an in

systemml git commit: [SYSTEMML-540] Reduce the number of unknowns in ConvolutionOp

2017-10-16 Thread niketanpansare
Repository: systemml Updated Branches: refs/heads/master 2ca2d8aa7 -> 5adb330de [SYSTEMML-540] Reduce the number of unknowns in ConvolutionOp - This commit reduces the unknowns during dynamic recompilation by inferring the input's height/width of ConvolutionOp based on its parent's output's he

[1/3] systemml git commit: [SYSTEMML-1963] Fix missing codegen dense-sparse vector primitives

2017-10-16 Thread mboehm7
Repository: systemml Updated Branches: refs/heads/master 06b4b9d5f -> 2ca2d8aa7 [SYSTEMML-1963] Fix missing codegen dense-sparse vector primitives This patch fixes issues that showed up with the fuse-all heuristic on different scenarios of ALS-CG. In particular, this adds missing vector primit

[2/3] systemml git commit: [SYSTEMML-1836] Fix unnecessary GC overhead in codegen row operations

2017-10-16 Thread mboehm7
[SYSTEMML-1836] Fix unnecessary GC overhead in codegen row operations This patch fixes issues of huge garbage collection overhead in special cases of codegen row operations with relatively large intermediate rows, matrix side inputs, and especially for multi-threaded operations. The major problem

[3/3] systemml git commit: [SYSTEMML-1964] Extended codegen outer template and rework close types

2017-10-16 Thread mboehm7
[SYSTEMML-1964] Extended codegen outer template and rework close types This patch makes a major change to the codegen outer template OFMC conditions in order to increase its applicability, which is crucial for sparsity exploitation in algorithms such as ALS-CG. In order to guarantee correctness, t