blib/lib/Peta/NN/Backend.pm untested

filter: Statementstmt 0.0% 0 / 28 Branchbran 0.0% 0 / 6 Conditioncond 0.0% 0 / 4 MC/DCmcdc 0.0% 0 / 2 Subroutinesub 0.0% 0 / 6 Podpod 0.0% 0 / 5 totaltotal 0.0% 0 / 51 CC 4 SCAR 18.3
CC
4
Cov
0%
create
12.0
argmax_rows
6.0
names
2.0
CRAP 6.2
1 package Peta::NN::Backend;
2 # ABSTRACT: the engines Peta::NN can compute on
3  
4 # Where the arithmetic happens. A backend owns the tensors (a batch of rows:
5 # inputs, activations, weights, gradients) and implements a dozen operations
6 # on whole batches; the layers above only say which operation, never how.
7 #
8 # plain Perl arrays. The reference: every other backend is tested against it.
9 # pdl PDL ndarrays. The efficient CPU form.
10 # gpu WebGPU storage buffers and compute shaders, in 32-bit floats.
11 #
12 # What a backend implements (tensors are opaque to the caller):
13 #
14 # tensor(\@flat, $cols) rows of $cols numbers, row after row
15 # tokens(\@flat, $per_row) rows of token indices, for embed()
16 # flat($tensor) back to a flat Perl list
17 # zeros_like($tensor)
18 #
19 # affine($X, $W, $b) X W^T + b
20 # affine_grad($X, $W, $dY, $need_dx) (dW, db, dX), summed over the batch
21 # activate($kind, $X)
22 # activate_grad($kind, $Y, $dY) in terms of the OUTPUT $Y
23 # embed($E, $tokens) table rows, concatenated per sample
24 # embed_grad($E, $tokens, $dX) dE
25 #
26 # softmax_ce($logits, \@classes, \@weights) (summed loss, dlogits)
27 # mse($out, \@flat_targets, \@weights) (summed loss, dout)
28 # the weights, one per row, are optional
29 #
30 # decay($p, $factor) every value times $factor, in place
31 #
32 # sgd_update($p, $g, $velocity, $lr, $momentum, $scale) in place
33 # adam_update($p, $g, $m, $v, $rate, $b1, $b2, $eps, $scale) in place
34  
35 use v5.36;
36  
37 our $VERSION = '0.2610090';
38  
39 my %CLASS = (
40 plain => 'Peta::NN::Backend::Plain',
41 pdl => 'Peta::NN::Backend::PDL',
42 gpu => 'Peta::NN::Backend::WebGPU',
43 );
44  
45 # Where a network that is left the choice ('auto') starts: the first of
46 # these that is there. The GPU is not among them, since opening a device is
47 # not free; a network goes there from here when it is trained and a few
48 # timed steps say that pays (Peta::NN, train).
49 my @AUTO_ORDER = qw(pdl plain);
50  
51 sub names { return sort keys %CLASS }
52  
53 # A backend object, or undef when its engine is not in this perl or finds
54 # nothing to run on (no GPU adapter); $@ then says why.
55 sub try ($name) {
56 my $class = $CLASS{$name} // die "unknown backend '$name' (have: @{[ names() ]}, auto)\n";
57 return eval "require $class; $class->new";
58 }
59  
60 sub available { return grep { defined try($_) } names() }
61  
62 sub create ($name) {
63 if ($name eq 'auto') {
64 for my $candidate (@AUTO_ORDER) {
65 my $backend = try($candidate);
66 return $backend if $backend;
67 }
68 }
69 return try($name) // die "backend '$name' is not available in this perl: $@";
70 }
71  
72 # The index of the largest value in each row. Read through flat(), so every
73 # backend has it; a batch of outputs is small.
74 sub argmax_rows ($self, $tensor, $cols) {
75 my $flat = $self->flat($tensor);
76 my @best;
77 for (my $start = 0; $start < @$flat; $start += $cols) {
78 my $best = 0;
79 for my $i (1 .. $cols - 1) {
80 $best = $i if $flat->[ $start + $i ] > $flat->[ $start + $best ];
81 }
82 push @best, $best;
83 }
84 return \@best;
85 }
86  
87 1;
88  
89 __END__