better_cpu_comparison.md

Below are the timing values in seconds for the kernel matvecs, invovlving different strategies in which the matvec is performed.

The time is in seconds in each table.

POCL and AMD are the 2 OpenCL implementations on which the timings are done.

Single Core

Kernel	POCL	AMD	MatFree	PyOP2(SpMV)
Mass	0.122	0.136	0.031	0.011
Laplace	0.124	0.125	0.035	0.011
Hyperelasticity	0.268	0.264	0.105	0.027

Cost of atomics on single core CPU

POCL:

Kernel	With	Without
Mass	0.122	0.064
Laplace	0.124	0.041
Hyperelasticity	0.268	0.091

AMD:

Kernel	With	Without
Mass	0.136	0.073
Laplace	0.125	0.057
Hyperelasticity	0.264	0.164

Cost of atmoics on Multicore CPU with Vectorization

Added MatFree just to compare what number are we chasing.

Kernel	With	Without	MatFree
Mass	0.029	0.009	0.002
Laplace	0.020	0.010	0.002
Hyperelasticity	0.041	0.019	0.009

New Lower bounds on bandwidth(using Footprint Measurement)

• Mass: 52GBps(was 1125 GBps)
• Laplace: 53GBps(was 270 GBps)
• Hyperelasticity: 18GBps(was 206 GBps)