Release v0.2.1

This release includes improvements to the tiling and DMA code generation, new networks and operators, improved CI workflows, migration to PyTest, and support for PyPi package releases.

Note: Since the release tag references the Docker container tagged with the release tag (ghcr.io/pulp-platform/deeploy:v0.2.1), the CI will initially fail. The Deeploy Docker image must be built after the release PR is merged and the CI restarted.

List of Pull Requests

• PyPi Package Deployment + Remove Banshee Dept #154
• PyTest Migration #144
• Update submodule pulp-nn-mixed #145
• Improve Profiling #138
• FP32 ReduceMean operator improvement #137
• Support for RMSNorm (Pow and Sqrt operators) #136
• Demo TinyViT compatibility with tiled Siracusa #124
• TinyViT on non-tiled Siracusa #117
• Support Fully Asynchronous DMAs #114
• Disallow shape inference #128
• Remove memory-aware node bindings #123
• Fix missing const's layout transformation and refactor NCHWtoNHWC passes #122
• Fix aliasing #125
• Support for 1D Autoencoder #98
• Refactor Logging for Improved Debugging #115
• Add reuse-tool as an SPDX license header linter #113
• Bug fixes, API Cleanup and Reduce Compiler Warning on PULP #112
• Fix PULP GEMM batch serialization #109
• Split CI Workflows by Platform and Task, Improve Formatting and Linting Reliability #108
• Refactor tiling code generation #105
• Change order of typeMatching entries #68
• Node Mangling to avoid duplication #93
• Prepare Post v0.2.0 Release #104
• Use Docker digests instead of arch-specific tags #106
• Fix Unsqueeze Op. when using ONNX opset 13 or higher (from attribute to input) #119
• Fix bias hoisting in generic GEMM with no bias #126

Added

• The publish.yml action to build a branch and push it to PyPi. The action is automatically triggered when a tag with the "v*" format is emitted.
• I created a release of Banshee so we don't need to rebuild it over and over. The Makefile now pulls that release depending on the platform.
• I bumped the onnx-graphsurgeon version such that we don't need to use NVIDIA's PyPi index anymore.
• _export_graph assigns their export type to the tensors before export.
• pytest and pytest-xdist as dependencies of Deeploy.
• A pytest.ini for the global configuration of PyTest for the project.
• conftest.py to define CLI args for PyTest for the whole project, it also defines a set of global fixtures and markers.
• pytestRunner.py contains helper functions and fixtures for the whole project.
• test_platforms.py lists the E2E tests and sorts them into marked categories (per platform and per kernel/model).
• Each platform has a test config file where a list or a dict describes the tests.
• Support for unknown number of data dimensions in the tiler
• Parallelization support for the FP32 ReduceMean operator on PULPOpen
• Extensive testing for the ReduceMean operator
• Pass to remove ReduceMean operators that don't change data content, but only its shape
• Support for RMSNorm operation via operator decomposition.
• Added Pow (Power) and Sqrt (Square Root) operation support (Parsers, Layers, Bindings, Templates, and FP32 Kernels) for the Generic platform.
• Support for input tiling for PULP FP regular and DW conv 2D.
• CI tests for tiled Siracusa FP regular and DW conv 2D, with and without bias, for skip connections, and for the demo version of TinyViT.
• Documentation for PULP FP regular and DW conv 2D and MatMul tile constraints.
• PULP ReduceMean and Slice tile constraints.
• PULP 2D FP DW conv Im2Col template and kernel, with bias support.
• Bias support for PULP 2D FP regular conv Im2Col in template & kernel.
• PULP FP DW conv 2D parser.
• FP conv 2D (simple & DW), reshape & skip connection, and TinyViT demo tests to the non-tiled Siracusa CI pipeline.
• FP bindings and mappings for PULP slice, DW conv 2D, and reduce mean operations.
• FP PULP DW conv lowering optimization pass similar to the existent one for integer version.
• RemoveEmptyConvBiasPass to the PULP optimizer.
• Add manual type inference feature (CLI: --input-type-map/--input-offset-map) to resolve ambiguities when test inputs are not representative enough
• Added a testTypeInferenceDifferentTypes test case to validate type inference for different input types
• Added _mangleNodeNames function to avoid duplicate node mappings
• Output Docker image digests per platform (amd64, arm64) after build, which is used to construct the multi-arch Docker manifest. This preventes registry clutter caused by unnecessary per-architecture Docker tags.
• AsyncDma abstraction of DMA's
• test runner per DMA and a script that tests all the DMA's
• generic Single/DoubleBufferingTilingCodeGeneration classes
• TilingVariableReplacementUpdate class that updates the variable replacement refs
• TilingHoistingMixIn class that encapsulates all the hoisting helper functions of tiling
• sorting of input memory allocations to allow references that live in the same memory level as the memory they are referencing
• a function that tests the tiling solution for correctness which currently only tests buffer allocation for byte alignment
• IntrospectiveCodeTransformation: _indexPointer(), indexVars(), dereferenceVars(). The *Vars functions index/dereference a list of variables (useful for tiling)
• NetworkContext: unravelReference() that unravels a _ReferenceBuffer until the base buffer
• NetworkContext: is_object() - helper function that determines whether the string represents a name of a local or global object
• NetworkContext: is_buffer() - helper function that determines whether the string represents a name of a buffer
• missing checks for environment variables
• _permuteHyperRectangle helper function
• Added CI badges to the README
• Added YAML linting to CI
• Added missing license headers and C header include guards
• Extended the pre-commit hooks to remove trailing whitespace, check licenses, format and lint files
• Reshape operator support for PULP (ReshapeTemplate in bindings)
• Missing class attributes in Closure.py
• reuse_skip_wrapper.py to manually skip files
• Centralized logging with DEFAULT_LOGGER, replacing print statements
• Debug logs for type checking/parsing; __repr__ for core classes
• Buffer utilities: checkNumLevels validation and sizeInBytes method
• Per–memory-level usage tracking and worst-case reporting in NetworkContext
• Memory/I/O summaries and input/output logging in deployers
• RequantHelpers.py for Neureka's TileConstraints
• Added assertion that all the graph tensors after lowering have a shape annotated
• Added testFloatGEMMnobias
• Profiling support and optional comments in generated DMA code for better traceability
• Added new waiting-strategy logic with fine-grained PerTensorWaitingStrategy
• PULPClusterEngine now accepts a n_cores parameter to set the number of cores used
• annotateNCores method to PULPDeployer that adds an n_cores key to all PULPClusterEngine templates' operatorRepresentations
• Calculate non-kernel overhead and show total time spent during profiling

Changed

• Rename package name from PULP-Deeploy to deeploy-pulp.
• Each CI workflow has been simplified to call the pytest suite with certain markers.
• Structure of Tests subdir for improved ordering
• Structure of .gitignore file for improved ordering
• Decreased L1 maximal memory limit for CI pipeline tests where compatible thanks to the implementation of Conv2D input tiling support.
• Reduced size of reshape & skip connection test, for non-tiled Siracusa memory compatibility.
• Replaced platform-specific tags (*-amd64, *-arm64) with direct digest references in Noelware/docker-manifest-action.
• mchan HAL is now reduced to bare-bones
• refactor of the IntrospectiveCodeTransformation to work on the Mako template
• refactor of memory allocation code transformation passes
• _ReferenceBuffer accepts an optional offset argument to offset the reference
• NetworkContext: hoistReference - accepts the actual buffer as reference instead of name, accepts shape, offset, and override_type arguments, and returns the actual buffer, not its name
• _mangleNodeRep -> _mangleOpRepr - the canonical name we use is OperatorRepresentation. NodeRep and ParseDict are old iterations of the name.
• rename of permutation functions to follow this convention: permute is an action that permutes something, permutation is a function that generates a permutation
• _permuteList to just _permute
• removed manual buffer name mangling since we do it in the ExecutionBlock generate() function, simplifies templates
• we now check that buffer shapes/hyperrectangles/tiling ranks match which required changing a few serializeTilingSolution functions to preserve the same shape rank
• big refactor of the code generation part of the TilingExtension and needed changes to PULPOpen and Snitch due to it
• PULPClusterTilingSB and PULPClusterTilingDB now allow for transfers of any rank (dimensionality)
• PULP's final output diff is now calculated as absolute error, instead of just subtraction
• common code generation code between testMVP/generateNetwork/... was extracted into a single generateTestNetwork function
• in some functions, instead of passing the name of a buffer, the actual buffer is just passed
• tile function allows overriding the optimizer with external tilingSolution and memoryMap
• refactor of the permutation functions for clarity
• Split CI into multiple workflow files: one per platform, one for lint & license, one for general Deeploy tests, one for infrastructure, and two for Docker flows, improving maintainability and status reporting
• Extended CI to check license in cMake and YAML files
• Removed all trailing whitespace
• Removed unnecessary includes from the PULP platform header list, such as DeeployBasicMath.h, for cleaner code generation
• Changed types and added correct casts to fix many compiler warnings in the PULP target library
• Use reuse-tool in pre-commit, CI, and Makefile for SPDX license header linting
• Deployer workflow now uses prepare(...) instead of generateFunction(...).
• Removed fromVariableBuffer
• Refactored hoistConstant
• Refactored TransientBuffer's __init__
• Refactor of the NCHWtoNHWC passes
• Removed NodeMemoryLevelChecker, MemoryAwareNodeBinding
• Removed _parseNode from MemoryNetworkDeployer since we don't need the annotations before typeChecking anymore
• Removed Wmem variants of bindings and tile constraints from Neureka
• Disabled ICCT_ITA_8 MemPool test because it was using a lowering that created shapeless tensors
• Added missing shape annotation to the testTypeInferenceDifferentTypes
• Refactored DMA code generation (SnitchDma, Mchan) to correctly overlap transfers and compute in double-buffering mode
• changed _mapNode to _selectEngine which reduces the responsibility of that function to, as the name states, just engine selection
• Print kernel profiling information for all memory levels

Fixed

• Update install.md to remove rust mention and fix test command.
• Update README.md to remove reference to NVIDIA's PyPi index.
• nvidia-pyindex was broken as it now tries to build the wheel to respect the new policy on packages using pyproject. Instead of installing this package, we just add the https://pypi.ngc.nvidia.com channel to the pip config file.
• Pin versions of broken dependencies of Banshee.
• Fixed ReduceMean parallelization and tiling issues described in Issue #134.
• Fixed PULP FP32 regular and DW Conv2D, and MatMul tile constraints.
• Fixed type casting for tiling code generation.
• Fixed bug in buffer name identification in code generation for tests with L3 default memory level.
• PULP GELU kernel to use tanh approximation.
• Fixed bug for non-batched elements in the PULPOpen FP GEMM and matmul templates.
• Added underscore to the beginning of closure names to avoid naming issues when they start with unsupported first characters (like numbers).
• Data types in the PULPOpen FP add and mul templates.
• Prevent node duplication for graphs generated via GraphSurgeon
• Resolved issue with missing id in the Build Cache for Docker step, used in the Inject build-cache step.
• Fix license CI check and prevent potential issues with jq installation
• PULP Gemm batch variable serialization
• Fixed multiple typos in variable and method names, such as changing includeGobalReferences to includeGlobalReferences and dicardedMappers to discardedMappers
• Corrected method usage in importDeeployState to call NetworkContext.importNetworkContext instead of the incorrect method name
• Correctly return signProp from setupDeployer instead of hardcoding the value to False in testMVP.py
• Fixed Unsqueeze Op. when using ONNX opset 13 or higher (from attribute to input)
• Fixed aliasing
• Missing layout transformation of the const's (bias, mul, add, shift in Conv/RequantizedConv)
• Keep mul/add rank of requantized Neureka tile constraints
• Fix bias hoisting in generic GEMM with no bias
• DMA synchronization bug causing reduced DB performance on memory-bound kernels.

Removed

• Delete outdated and unused .gitlab-ci.yml file
• dory_dma.c and dory_dma.h