Alejandro Saucedo

Currently if empty array of tensors is passed to the OpAlgoBase, it would cause an exception as dispatch is expected to be inferred to 1, so need to handle the...

Currently the way that memory is mapped into host memory is always flushed. This in theory is only necessary if the host bit is not coherent, so should ensure this...

Currently the kp::Algorithm doesn't allow for the internal components to be provisioned by default, and should be explored how these can be exposed so components like pipelines can be provided...

Currently there are commands such as the tensor's recordCopyFrom command that can be potentially executed with resources that may hvae different devices. This would encompass adding functionality to check that...

Currently tensors can be initialised and this can be checked with the isInit function. We should explore how to also guard from the tensors beign out-of-sync when operations that modify...

Currently most of the Kompute components should not be passed using assignment or copy constructor operators. Also potentially deleting the base constructor on most classes that should not be created...

Explore tools like https://github.com/gperftools/gperftools to add as unit testing for memory profiling

Currently the recordBufferMemoryBarrier from the Tensor class requires raw vulkan flags to set up the barrier. There are higher level patterns and information that the tensor itself holds that could...

Currently Tensors only use buffers to store data, we'll have to look into adding capabilities to support imageBuffer, but will also be important to identify the key usecases that this...

Currently only 1 dimensional vectors are supported, looking to add extra support for multi-dimensional vectors