Home

Welcome to the SPARTA wiki!

SPARTA architecture notes

Utility

Input file

Read from file the following formats

• sorted edge list (SNAP style)
• mtx supported
• VBR Output is a CSR. Features:
• from the input file to CSR;
• our CSR stores the number of elements of i-row and not the position of the first element for the i-row;
• shape = 0 e.g., no symmetric squared matrix
• boolean matrix = 0
  Need to be revised...

Reordering

It reads env configuration and uses OpenMP to convert into VBR according to a specific algorithm. Access to routine for computing statistics from CSR. Optimized versions that use accelerators goes into a separate file e.g., reordering.cu

reordering takes: CSR struct, algorithms (saad, SA-1 etc... ), pointer to VBR

return in VBR

• Variable block
• Fixed block

Examples

a main for each architecture supported (readfile - convert if any - multiply with proprietary libs that use tensor accelerators

• CUDA (cuBLAS - tensor cores)
• GraphCore (poplib - IPUs)
• IBM (? - ?)
• Intel (MKL - ?)
• ARM (? - ?)
• Specialized multiplications (CUDA etc... or others )

Benchmark

similar to examples

• Synthetic matrices
• Synthetic graphs
• Read files as implemented in examples

Applications

• TC
• DNN / GNN