Graphene CLI + quantization research (#1320)

* Implement skeleton for graphene-cli

* Configure gpu surface on non wasm32 targets

* Create window with full hd size

* Create window using the graphen-cli

* Use window size for surface creation

* Reuse surface configuration

* Reduce window size for native applications to 800x600

* Add compute pipeline test

* Poll wgpu execution externally

* Remove cache node after texture upload

* Add profiling instructions

* Add more debug markers

* Evaluate extract node before flattening the network

* Reenable hue saturation node for compilation

* Make hue saturation node work on the gpu + make f32 default for user inputs

* Add version of test files without caching

* Only dispatch each workgroup not pixel

* ICE

* Add quantization to gpu code

* Fix quantization

* Load images at graph runtime

* Fix quantization calculation

* Feature gate quantization

* Use git version of autoquant

* Add license to `graphene-cli`

* Fix graphene-cli test case

* Ignore tests on non unix platforms

* Fix flattening test
This commit is contained in:
Dennis Kobert
2023-07-04 17:04:09 +02:00
committed by Keavon Chambers
parent 61c5dd1f88
commit 3c2d371173
57 changed files with 10169 additions and 845 deletions
@@ -9,13 +9,12 @@ pub struct Compiler {}
impl Compiler {
pub fn compile(&self, mut network: NodeNetwork, resolve_inputs: bool) -> impl Iterator<Item = ProtoNetwork> {
let node_ids = network.nodes.keys().copied().collect::<Vec<_>>();
println!("flattening");
let node_ids = network.nodes.keys().copied().collect::<Vec<_>>();
for id in node_ids {
network.flatten(id);
}
network.remove_redundant_id_nodes();
network.resolve_extract_nodes();
network.remove_dead_nodes();
let proto_networks = network.into_proto_networks();
proto_networks.map(move |mut proto_network| {