[V1.0] Development Work

[luigifcruz]

The scope of this pull request is to remove unnecessary complexity from BLADE. This aims to simplify maintainability and facilitate integration with third-party frameworks. The end result should retain the performance while reducing the boilerplate code necessary to create pipelines and schedule work. The Python interface will also receive an overhaul to match C++ API in both performance and features. Higher-level applications such as the CLI will also be rewritten using the Python bindings instead of the C++ API.

It isn't in the scope of this PR to modify the code of existing modules. This will be addressed in future versions of BLADE. Features such as support for multiple ArrayTensor data layouts like ATFP and AFTP, the addition of half-precision computation support, and in-place array operations.

• Removal of CUDA Graphs. [Unnecessary Complexity]
• Removal of CLI. [To-be Rewritten in Python]
• Remove support for nested Pipelines. [In Favor Of Variable Rate Pipelines]
• Replace pre-made Pipelines with Bundles. [Better Single Pipeline Module Support]
• Refactor Runner with buffer management.
• Modernization of the Python API. Possibly with nanobind. [Too Low-level Currently]
• Add variable rate pipeline modules. [Gather, Copy, Permutation]
• Add an open-source license.
• Write updated documentation. [Readme]

Closes #65 and #61.

[luigifcruz]

It will be necessary to refactor the Blade::Runner class to add support for handling asynchronous memory copies from host to device locally instead of deferring it to the pipeline execution. This will result in less VRAM usage with no expected impact on performance.

[luigifcruz]

To ensure a level of stability for BLADE users, it is essential to keep track of performance regressions. In previous versions, benchmark results were manually saved to a file inside the repository tree. This PR adds the ability to automatically run benchmarks and tests using a standard server. This makes it possible to graph the benchmark results of components from the past and see how they have changed over time.

Automated benchmarks and tests will be run before any pull request is merged into the main branch to ensure quality. Approved work-in-progress pull requests will also have the ability to run these automated tasks to keep track of regressions as they occur.

For consistent results, a self-hosted machine dedicated to this task will serve as the standard server. Its hardware configuration will be representative of the hardware currently used in production. By using the results produced by this server, it will be possible to extrapolate the performance for other hardware configurations. This capability will help better understand how to deploy the pipeline in a heterogeneous server hardware configuration environment.

[luigifcruz]

All tests are passing with no actionable TODO for this PR. Merging!