Skip to content

v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@Xeratec Xeratec released this 08 Jul 13:58
· 25 commits to devel since this release
v0.2.0
This tag was signed with the committer’s verified signature.
Xeratec Philip Wiese
SSH Key Fingerprint: xwt+VVmkWrqT6F7IhnZ4Df9+z5E+zKYaYy0c4fXIMIk
Verified
Learn about vigilant mode.
c7fd2a1
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release v0.2.0 (2025-07-08) #103

This release contains major architectural changes, new platform support, enhanced simulation workflows, floating-point kernel support, training infrastructure for CCT models, memory allocation strategies, and documentation improvements.

List of Pull Requests

  • Prepare v0.2.0 release #102
  • Add Luka as Code Owner #101
  • Fix CI, Docker Files, and Documentation Workflow #100
  • Chimera Platform Integration #96
  • Add Tutorial and Refactor README #97
  • Reduce Mean Float Template #92
  • Reshape Memory Freeing and Generic Float GEMM Fixes #91
  • Prepare for Release and Separate Dependencies #90
  • Fix input offsets calculation #89
  • Move PULP SDK to main branch/fork #88
  • Finite Lifetime for IO Tensors #51
  • Improved Memory Visualization and Multi-Layer Tiling Profiling #56
  • Fix Linting in CI and Reformat C Files #86
  • Fix Broken CMake Flow For pulp-sdk #87
  • Refactor Changelog For Release #85
  • ARM Docker Container and Minor Bug Fix #84
  • Added Kernel for Generic Float DW Conv2D #63
  • Autoselect Self-Hosted Runners if the Action is on Upstream #81
  • TEST_RECENT linking on MacOS #78
  • Add RV32IMF Picolibc support for Siracusa platform #66
  • Improve Documentation and VSCode Support #76
  • Debug Print Topology Pass and Code Transformation #75
  • Find all subdirectories of Deeploy when installing with pip install #70
  • Add milestone issue template #71
  • Bunch of fixes and changes #58
  • Add SoftHier platform #65
  • rv32imf_xpulpv2 ISA support for Siracusa platform #64
  • One LLVM To Compile Them All #60
  • One GVSoC to Simulate Them All #59
  • Add Support for CCT Last Layer Training with Embedding Dim 8-128 #55
  • Add CCT Classifier Training Support #53
  • L3 Bugs: DMA Struct Datatype and Maxpool Margin Error #45
  • DeepQuant Quantized Linear Support #54
  • Implemented Dequant Layer for Generic and Siracusa #52
  • Infinite Lifetime Buffers Considered in Tiling & Memory Allocation (+ Visualization) #44
  • Implemented Quant Layer for Generic and Siracusa #49
  • Increase maximal Mchan DMA transfer sizes from 64KiB to 128KiB #47
  • Add MiniMalloc and Decouple Memory Allocation and Tiling #40
  • Float CCT Bugs on L3 #37
  • Memory Allocation Strategies and Visualization #36
  • Add CODEOWNERS #42
  • Add Tiling Support to All CCT Kernels and Fix CCT Operators on Siracusa Platform for L2 #35
  • Add Fp gemm and Softmax for Snitch platform #31
  • Add Float Kernels for CCT #29
  • documentation deployment #34
  • main.c Float Cast Bugs #28
  • Add Float GEMM on PULP with Tiling #26
  • Add Float Support & Float GEMM for Generic #25
  • GVSOC support for the Snitch Cluster platform #23
  • Snitch Cluster Tiling Support #22
  • Snitch support integration #14
  • Update bibtex citation #20
  • the PR template location, bump min python to 3.10, change install command #17
  • Add pre-commit for python formatting #15
  • FP integration (v2) #12
  • shell for sequential tests of Generic, Cortex, and Mempool platforms #11
  • Add issue templates #10
  • Minor CI and Readme Improvements #8
  • Fix GHCR Link for Docker Build #7
  • neureka's ccache id #6
  • GitHub-based CI/CD Flow #4
  • Generic Softmax Kernel #2
  • Port GitLab CI #1

Added

  • ChimeraDeployer, currently mainly a placeholder
  • Allocate templates for Chimera
  • ChimeraPlatform, using appropriate allocation templates and using the generic Parser + Binding for the Add node
  • Adder CI test for Chimera
  • Install flow for chimera-sdk via Makefile
  • DeeployChimeraMath library
  • Generic FP32 reduce mean bindings, parser, and template
  • New alias list parameter for buffer objects
  • New test, also included in the CI pipeline, for the reshape and skip connection situation
  • 'shape' parameter handling similar to the 'indices' parameter in the generic reshape template
  • Test the correcteness of the memory map generated by the tiler
  • Add attribute to VariableBuffer to distinguish I/Os
  • Add proper static memory allocation with finite lifetime for I/Os
  • The memory allocation visualization now displays the allocation for each level used
  • Tutorial section in the documentation
  • Guide on using the debug print topology pass and code transformation
  • VSCode configuration files for improved IDE support
  • Multi-branch GitHub Pages deployment support
  • Test for the DebugPrintTopologyPass.
  • Test for PrintInputGeneration, PrintOutputGeneration, MemoryAwarePrintInputGeneration, MemoryAwarePrintOutputGeneration
  • check for CMAKE variable and fallback to searching for cmake
  • tensor name mangling
  • identity operation removal
  • _unpack_const helper function to NodeParser to allow for node attributes that are direct Constant tensors or direct numpy values
  • load_file_to_local in dory_mem as a way to load values directly to a local memory (not ram). needed for copying values from flash to wmem needed for Neureka v2
  • Add the documentation.yml workflow to deploy doc pages.
  • Improved README with more detailed Getting Started section, a section listing related publications, and a list of supported platforms.
  • Schedule a CI run every 6 days at 2AM CET to refresh the cache (it expires after 7 days if unused).
  • Add the FloatImmediate AbstractType
  • Define fp64, fp32, fp16, and bf16
  • Add float binding for the Adder in the Generic platform
  • Add a FloatAdder test to the CI for Siracusa and Generic platforms
  • Extend testType.py with float tests
  • LIMITATION: Current LLVM compiler does not support bfp16 and fp16, these types are commented in the library header
  • cMake Flow for the Snitch Cluster
  • Added snitch_cluster to Makefile
  • New Snitch platform with testing application
  • Testrunner for tiled and untiled execution (testRunner_snitch.py, testRunner_tiled_snitch.py)
  • Minimal library with CycleCounter and utility function
  • Support for single-buffered tiling from L2.
  • Parsers, Templates, TypeCheckers, Layers, and TCF for the newly supported operators.
  • A code transformation pass to filter DMA cores or compute cores for an ExecutionBlock.
  • A code transformation pass to profile an ExecutionBlock.
  • Test for single kernels, both with and without tiling.
  • Adds the --debug flag to cargo install when installing Banshee to get the possibility of enabling the debug prints.
  • New tests for the snitch_cluster platform.
  • Add macros to main.c to disable printing and testing (convenient when running RTL simulations).
  • gvsoc in the Makefile and dockerfile
  • cmake flow for gvsoc
  • CI tests regarding Snitch run on GVSOC as well
  • Float Support for Constbuffer
  • Simple Float GEMM on Generic and PULP
  • FP GEMM to CI
  • FP GEMM Tiling on PULP
  • Add one new #define OUTPUTTYPE to testoutput.h
  • Float Template, binding and parser, test for Conv2D, LayerNorm, Div, Relu, Softmax, MaxPool, Matmul, Transpose, Gelu, Mul, Reshape, Gather, Squeeze, Padding
  • CCT model test to Generic Target
  • Math Lib link on Generic Target
  • New templates for GEMM and Softmax.
  • Added GEMM and Softmax to TargetLibraries, including case for GEMM with a transposed B matrix.
  • Added new CI tests for GEMM and Softmax.
  • Float Bindings, Tilers of CCT kernels for Pulp Target
  • Float Convolution, MaxPool Parser, Template, Kernel with HWC layout and padding integrated
  • Added tiling constraints for conv gather and layernorm and exisitng constraints for other kernels
  • profileuntiled arg
  • CCT onnx tests with img size of 16 and 32
  • CODEOWNERS file to control who is responsible for reviewing future PRs.
  • A visualization of the memory allocation solution generated by Deeploy at each level of memory. I use Plotpy to generate a static html file and save it to the DeeployState directory.
  • An initialization strategy for the variable in the tiling to randomize the variables related to the permutation matrix.
  • New interface to testRunner_tiled_siracusa to control the generation of the memory allocation visualization, the memory allocation strategy, and the search strategy.
  • Export a new docker container with plotpy as dependency.
  • Added multiple CCT settings for testing.
  • Added CCT L3 test to CI to ensure correctness for img size of 16 and 32.
  • Added NaN check for deeploytest diff to improve result validation.
  • Installation and compilation flow for MiniMalloc through Makefile.
  • Adapt the docker to install MiniMalloc and declare necessary symbols.
  • Add the constraintTileBuffersWithOverlappingLifetime method to the memory scheduler to add the necessary memory constraint when we decouple memory allocation and tiling.
  • Add the minimalloc method to the Tiler class. MiniMalloc comes as a precompiled cpp library using CSV for I/O. Hence, this method converts Deeploy's memory map to MiniMalloc's CSV representation, calls a subprocess to run MiniMalloc, reads the output CSV, and translates it back to Deeploy's memory map.
  • Add MiniMalloc to the memory allocation strategies and add a new argument to the test runner to control the L2 size.
  • New Quant operation to handle quantization pattern in ONNX models
  • Implementation for both Generic and Siracusa targets in the Deeploy framework
  • Custom QuantPatternPass class to replace matched patterns with a single Quant operator
  • Parser implementation in Parsers.py to extract quantization parameters
  • C template implementation in QuantTemplate.py for efficient quantization
  • Type checker implementation in TypeCheckers.py to handle bit-width and signedness
  • New Dequant operation to handle dequantization pattern in ONNX models
  • Implementation for both Generic and Siracusa targets in the Deeploy framework
  • Custom DequantPatternPass class to replace matched patterns with a single Dequant operator
  • Parser implementation in Parsers.py to extract dequantization parameters
  • C template implementation in DequantTemplate.py for efficient dequantization
  • Type checker implementation in TypeCheckers.py to handle bit-width and signedness
  • New Test Cases: Added and passed tests for 16×16 64 and 16×16 128 configurations to validate correctness.
  • New _sanitizeGraphNames function to sanitize the names of the nodes and tensors of the graph
  • Implementation for both Generic and Siracusa targets in the Deeploy framework
  • Modified the binding of dequant in Bindings.py to handle int32 after GEMM operation
  • New test cases: testTrainCCT/CCT_GEMM_Weight_Bias_1_16_16_8, testFloatReduceSum, testFloatSoftmaxGrad, testFloatSoftmaxCrossEntropy, testFloatSoftmaxCrossEntropyGrad
  • New kernels: SoftmaxCrossEntropy, SoftmaxCrossEntropyGrad, SoftmaxGrad, ReduceSum
  • Refinements in operator parsers and computeShape logic for: Softmax, Mul, Reducesum
  • New _sanitizeGraphNames function to sanitize the names of the nodes and tensors of the graph
  • Implementation for both Generic and Siracusa targets in the Deeploy framework
  • Modified the binding of dequant in Bindings.py to handle int32 after GEMM operation
  • Support for SoftmaxCrossEntropyLoss and SoftmaxCrossEntropyLossGrad with tiling.
  • Implementation of SGD updates for CCT training.
  • Test for one iteration of CCT last-layer training with dimensions from 8 to 128.
  • All Banshee dependencies now have a frozen version. This improves maintainability as some packages get yanked for the old versions of Rust.
  • Increase the L2 buffer size for loading files from Flash to RAM. This speeds up the simulation setup time.
  • Align the GVSoC simulation command and build command for the new version.
  • Bump new version of GVSoC and PULP-SDK
  • Build flow and its Docker integration for LLVM 15 tagged `15.0.0-snitch-0.1.0'
  • Picolibc build flow for v32im, v32ima, rv32imc and rv32imafd. Previously, it was only for rv32imc.
  • LLVM Compiler RT for rv32im, rv32ima, and rv32imafd.
  • Appropriate linking of picolibc and compiler RT.
  • Build and install a flow for XTensor, XTL, and XSIMD. These libraries are used in some GVSoC models, and they used to live in the PULP SDK, as a header-only library. Keeping only the library headers in the PULP SDK makes it hard to bump new versions.
  • Adds RV32IMF Picolib to the toolchain
  • Generic float DW Conv2D kernel and bindings.
  • Bias handling and computation for regular and DW Conv2D.
  • Empty bias handling for generic regular and DW Conv2D.
  • Tests for Conv2D regular and DW, with and without bias (and included them in the CI pipeline).
  • BuildDockerToolchain.yml to build Toolchain Docker container
  • BuildDockerDeeploy.yml to build Deeploy Docker container
  • Add support for linux/arm64 containers
  • Added caching to speed up container builds
  • Makefile to simplify local container build
  • Add helper script to generate a baseline changelog.
  • SoftHier Deeploy Targets, including Deployer, Platform, and Templates
  • SoftHier cmake compilation flow
  • SoftHier CI task
  • Parallel implementations of the following operators on Siracusa: Matmul, Softmax, Gelu, Conv, Layernorm, Maxpool, Add, Mul,and Relu
  • Gelu with Sigmoid implementation
  • ComputeOp support for multiple float kernels: Maxpool, Relu, and Mul
  • dev-requirements.txt tracking the dependencies of the build system, linting, documentation, and QOL.

Changed

  • Bump the CMake version to 3.24 as required for the chimera-sdk
  • Bump GVSoC's version and add chimera simulation target
  • Rename the generic source util to utils to avoid name collision with chimera-sdk
  • Moved PULP SDK from Victor-Jung/pulp-sdk branch deeploy to pulp-platform/pulp-sdk branch main.
  • Memory arena buffers are now declared at the beginning of the InitNetwork function
  • Tiling profiling is now an ON/OFF version where you get the I/O DMA time for each DMA call
  • The profiling strings are const static, such that they are stored in .rodata
  • Adapt the select docker image stage to also select a runner depending on github.repository
  • Adapt the jobs and reusable workflows to use the selected runner.
  • Updated README.md description to use a persistent development container
  • Symlinking of the latest build and source files into TEST_RECENT
  • Disabled CMAKE_VERBOSE_MAKEFILE by default for cleaner builds.
  • Refactored IntrospectiveCodeTransformationMixIn to allow extracting dynamic references to global variables
  • duplicateConstants now also duplicates constant nodes
  • check float output define in DeeployTest Generic platform
  • kernel_shape now inferred from weight shape if not present as per ONNX spec
  • USE_NEUREKA moved into TargetLibraries where it's closer to pulp-nnx
  • hex dumping logic for pulp platforms in prep for neureka v2 where I need to save weights to flash and move them during runtime to wmem
  • add_gvsoc_emulation macro now requires an additional target argument and abstracted adding flags to gvsoc through the GVSOC_EXTRA_FLAGS variable
  • Updated README.md with direct link to the documentation page.
  • Update the Banshee's commit to include a recent PR.
  • Add the possibility of changing the simulator when using the snitch-tiled test runner.
  • Add the RTL library to the snitch_cluster build process in the Makefile, required for GVSOC simulation
  • float infinity macro #define inf
  • Signprop depend on float check and platform
  • Adapted snitch Bindings and Platform files.
  • Removed unused TilerAwareDeployer class.
  • Regenerated CCT ONNX files without "output" & "input" in their names to avoid triggering the dumphex parser bug.
  • Regenerated CCT ONNX file with 3 branches for attention, transforming the attention computation graph into three branches.
  • Changed code generation for Hex output to properly handle float values.
  • Enhanced layernorm operator to support three outputs (layernormout, mean, std) for compatibility with training-related layernormgrad in the future.
  • Modified the outputs of LayerNorm and SoftmaxCrossEntropyLoss nodes to a single output for better tiling compatibility.
  • Added SGD parameter updates to the CCT training graph.
  • Officially depreciate Banshee as a simulator for Snitch Cluster in the CI. Maintaining this is a burden and unnecessary, as GVSoC is now the standard simulator. Additionally, newer versions of the Snitch runtime don't support Banshee anymore.
  • Bump XTensor's version to 0.25.0 to fix a bug with Intel's SSE.
  • Update snitch cluster patch to link to picolibc and add explicit target.
  • Update README to include Snitch in the Getting Started and the D&T Journal.
  • The ISA for the Siracusa platform has been updated from rv32imc_zfinx_xpulpv2 to rv32imf_xpulpv2.
  • All floating-point comparison tasks in deeploytest.c are now offloaded to Cluster 0 for execution.
  • Split the original build flow into two container
  • Refactor changelog for better readability
  • Reformatted all C files
  • Prepare pyproject.toml for a proper pip package release.
  • Packages listed in dev-requirements.txt are installed in the final stage of the Deeploy container.

Fixed

  • DW Conv2D kernel header to avoid warnings
  • FP casting in GELU kernel to avoid warnings
  • Buffer deallocation to only happen when all its aliases are not live anymore (the data stored there is not needed anymore, not even by other nodes)
  • GEMM Generic float template to iterate through terms only when they actually contain multiple matrices
  • Fix the PULP Deployer where outputs were unecessary loaded in L3
  • Fix the lifetime computation of aliased buffers
  • Removed unsupported -MMD compiler flag in LLVM-based toolchains.
  • Fix DebugPrint topology pass
  • Fix PrintInput code transformations to work with global variables
  • RequantShift when log2d is 0
  • missing math.h headers
  • clang on mac doesn't support -Wl,--gc-sections flag, moved it into each target and for host it's checking now for host system
  • --ffast-math caused numerical errors on generic so moved into each target and removed from that one since I'm imagining it as the debug target
  • Gather kernel on generic target
  • Update the link of the Docker container used to run the CI with the Docker published by this repo instead of my fork.
  • Add a retry on timeout step for large network tests. This is a temporary fix to address the sporadic freeze happening at the compilation stage, see this issue.
  • Float bug on Testslice, CMSIS TestUtil, DivInterger
  • AbstractDatayType Float Bugs
  • Change main.c to use OUTPUTTYPE instead of float
  • MaxPool Padding Extract Pass for float and interger
  • Testinput, testoutput, weight type casted from double to float warning
  • Relaxed the error threshold between expected and actual values in deeploytest.
  • CycleMeasure Pass for Siracusa Untiled Profilling
  • GEMM Tiling Constraints transA and `transB' not supported
  • MatMul layer Multi-Dimensional Input Issue
  • Add Layer for Broadcasted Bias
  • Resolved an issue where concatenation of float32 with f caused inf errors during code generation
  • Fixed a bug in the MemoryScheduler where the CP problem was solved more time that it was needed.
  • Updated printinput nodetemplate for float handling.
  • Fix testMVP.py to get a proper should fail test.
  • Maxpool Tile Calculation Error: The last dimension padding was incorrectly calculated due to L3 wraptiling solution. This has been fixed by updating serializeTilingSolution of Maxpool to avoid incorrect padding of Maxpool and prevent potential DMA 3D transfer issues of Maxpool.
  • DMA 1D Copy Assertion Issue: Updated the DMA length datatype from uint16 to uint32 to avoid assertion failures when dealing with large block transfers.
  • Deeploy subdirectories installed when installing Deeploy with pip install
  • Fix linking TEST_RECENT on MacOS
  • Fixed broken VSCode launch configuration
  • Fixed broken pulp-sdk hash
  • Fix issue with building banshee on `linux/arm
  • Removed i3c related files from the pulp-sdk CMake flow
  • Fixed C-code linting stage in CI
  • Input offset height and width calculation for tiled PULPOpen convolution kernels

Removed

  • Removed commented code from generic allocation templates
  • Remove the link to the precompiled LLVM 12 in the testRunner for Snitch and in the CI.
  • Remove the sourcing of the cursed PULP SDK script.
  • Commented IPython breakpoints.
Assets 2
Loading