/ 
Peta-NN-0.2610090
/ Peta::NN::Backend::WebGPU

NAME

Peta::NN::Backend::WebGPU - Peta::NN on the graphics card

VERSION

version 0.2610090

SYNOPSIS

my $net = Peta::NN->new(input => 2, layers => [ [dense => 2] ], backend => 'gpu');

DESCRIPTION

Selected with backend => 'gpu'. Needs a pperl built with the webgpu feature and a usable adapter. For a small model on few samples the per-operation dispatch costs more than the arithmetic saves, and a network with backend => 'auto' comes here only when a few timed steps say it is faster ("train" in Peta::NN). It pays with wide layers, large batches and many samples: Peta::NN then keeps the samples of a training run on the card (see epoch_data below), and a step carries nothing there or back.

Single precision: results agree with the plain backend to about six digits. See Peta::NN::Backend for the operations.

METHODS

This class implements the backend interface described in "THE BACKEND INTERFACE" in Peta::NN::Backend and adds these methods to it.

adapter

What the graphics adapter the backend computes on says of itself, as a table.

epoch_data, epoch_order, batch_at, softmax_ce_at, epoch_loss

An epoch that stays on the card. epoch_data(\@token_rows, $per_row, \@classes, \@weights) puts all the samples of a training run on the card, once; epoch_order($data, \@order) the order they are taken in, once per epoch; batch_at($data, $start, $count) gives the token rows of a step as embed takes them, picked on the card; softmax_ce_at($logits, $data, $start) returns the gradient and leaves the losses on the card; epoch_loss($data) reads their sum when the epoch is over. Peta::NN trains this way when its backend has these methods.

define

define($name, \@bindings, $dimensions, $body, $functions): adds a compute shader under a name, for a module that brings its own (Peta::NN::Fused does). A binding is "r:name" (floats, read), "w:name" (floats, written), "u:name" (unsigned integers, read) or "x:name" (unsigned integers, written); the body is WGSL and sees the invocation as id and the parameter block as p; $functions, optional, is WGSL functions the body calls.

dispatch

dispatch($name, \@buffers, \@integers, $nx, $ny): runs a shader over a grid of invocations, with up to eight integers in its parameter block.

buffer

buffer($bytes): a storage buffer on the card holding the bytes, or, given a number, one of that many bytes.

AUTHOR

PetaMem s.r.o. <info@petamem.com>

COPYRIGHT

Copyright (c) 2026 PetaMem s.r.o.

LICENSE

This package is free software, dual-licensed under the Artistic License 2.0 and the BSD 2-Clause License. See the LICENSE file of the distribution.