daisytuner / docc / 26871523909
61%

DEFAULT BRANCH: main
Ran
Jobs 0
Files 0
Run time –
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
26871523909

push

github

web-flow 
Tensor node backport & quantization support (#725)

 * Updated Conv & Pooling Tensor Nodes to supporting quantization, sharing their base-impl and using new ptr format
 * Updating all unary, binary and cmath eltwise tensor nodes to all inherit from the more flexible ElementWiseDataflowTensorNode.
 + ElementWiseDataflowTensorNode: Can handle arbitrary amounts of inputs, with x tensor inputs, y further, non-tensor inputs, with support for optional parameters. Handles all the inputs and outputs, leaving only the core data flow.
 ! found a bunch of bugs where existing eltwise nodes had incomplete and invalid implementations, prevent them from running faulty
 * Updated all reduce nodes to new ptr structure (they could also do with a base-class, but not now)
 + Added Pointer metadata to the updated nodes
 * Update PyStructuredSdfgBuilder to use the new ptr-edges, fixed new TensorNodes use builder.clear_code_node to ensure removal of node without double-free etc.
 ~ Fixed: use utility function to correctly create transient temporary containers to use in expand. Was missing container creation
 ~ Fixed: use after free in eltwise_test
 * also extracted some reusable code from ConvNode.expand to reuse in CudaConvExpander.
 * CudaConvExpander will now decline to expand cases instead of falling back to essentially copy of ConvNode.expand()
 * Cherry-picked updated docker container names
 * Brought SpatialTensorNodes (conv & pooling) in line with Matmul's handling of quantization (infer uniform type from edges by default)
Source Files on build 26871523909
Detailed source file information is not available for this build.