wildjames / predictive_coding_rs / 26040019395

90%

Build Type

push

github

Committed by web-flow

Commit Message

Implement the remaining GPU kernels (#18)

* Move top layer error calc off the CPU and into the GPU

* Move the model error calculation to the GPU. Error per layer is calculated there, so we only need to move N values off the GPU for a model with N layers

* Move the value change coalescing onto the GPU

* Add training handlers for GPU. Move some more processing out to the GPU properly. Minor refactoring to add profiling to the benchmarking code.

* Stop committing benchmark data to source control

* remove gpu playground program. Update the devcontainer to be able to run the GPU pipeline

* Update configs and docs

* Remove debugging block as it may cause compile errors

* Compute error sum on the GPU to minimise data transfer there

* Add weight update clipping to help prevent divergence

* todo note

* use minimum buffer count

* Move duplicated code out to a macro

* Reduce duplication by moving code to shared functions

* report weight clip

* define the top group only once, and reuse it

* add fixme

Coverage Stats

268 of 288 new or added lines in 12 files covered. (93.06%)

6 existing lines in 2 files now uncovered.

2080 of 2320 relevant lines covered (89.66%)

14.25 hits per line