carbs

Here I have organized my thoughts as I was performing an experiment on Gnocchi - the database

• What and Why
  Its an experiment to explore General-Purpose-GPU compute space where I try to shift computation from CPU to GPU for the OpenSource project Gnocchi

• How
  This part goes into implementation details. It explains the iterations based on profiling results.

• Now What
  It talks about Gpu-fication of other parts of the library. It also has my learning and conclusions from the experiment.

Repo Readme:

This repo has experiments and analysis for implementing various kernels for the carbonara library from the Gnocchi project.

The repo is organized into various branches.

• v12 This branch has the version v1 and v2
• V3 (final) this branch is for version v3 and also the final implementation
• v3.1 This was intermediate improvement over v3 but the speed-up was not great
• v4 This implements the kernel using stream, a more advanced approach
• threaded Another approach where work is launched from different threads, but overhead is too high

Originally published here:

Post 1, Post 2, Post 3