Improved profiling stability and reliability during model execution.
Improved profiling robustness and more accurate performance reporting.
Driver fixes improve clock measurement accuracy under multi‑threaded workloads.
Runtime robustness improvements reduce crashes during profiling and startup.
Expanded operator support (GEMM, Elemwise, Pack, Slice, SpaceToDepth/DepthToSpace).
Improved shape inference, layout handling, and graph optimization for better portability.
Runtime and driver enhancements improve execution stability and startup reliability.
Updated Hexagon simulator prebuilts improve simulation and validation coverage.
New custom‑op registration support improves extensibility.
Improved Elemwise Ops(Add/Sub/Mul): broadcast handling,shape inference,mixed‑dim tensors.
Reduce Op: More reliable axis handling when reducing tensors without keeping dimensions.
GEMM/FullyConnected Ops: Improved slicing,execution efficiency for dense compute workloads.
Norm Ops(LayerNorm/RMSNorm): Improved numeric stability and accuracy in quantized models.
Transpose,Reformat Ops: Bug fixes and broader layout support improve model portability.
More accurate output shape propagation reduces downstream mismatches.
Improved correctness of Quantized Op single‑channel and edge‑case quantization scenarios.
Runtime improvements that reduce failures when handling model input/output buffers.
Improved reliability during model initialization and execution start‑up.
Add support for setting Power modes (eNPU clocks, DSP voltage corners, etc).
Improved overall stability and accuracy across supported workloads.
Better reliability for convolution-based and quantized models.
Improved handling of pooling, padding, and shape transformations.
Increased robustness of elementwise ops (Add, Sub, Mul, broadcast patterns) execution.
Increased robustness of GEMM/MatMul execution.
Expanded support for a broader range of vision, audio, and transformer-style models.
Improved efficiency of matrix ops(GEMM/MatMul), elemwise ops, normalization flows.
Above improvements benefiting modern neural networks.
Enhanced stability for sequence-based workloads (GRU/RNNstyle models) and quantized models
Improved runtime performance, profiling, and mem-management for more predictable execution.
Improve LPAI profiling initialization using ADSP/ENPU clock.
Corrected errors in the LPAI Backend Op Definition Supplement.
Performance improvements for certain model patterns.
Better support for advanced normalization operations.
More accurate handling of tensor layouts and shapes.
Fixes to ensure correct outputs for select model types (e.g., GRU).
Improved memory‑usage reporting for easier model planning.
Improved runtime stability and error handling.
Enhanced support for performance-related data (e.g., clock info).
Better consistency with platform requirements.
General fixes that improve execution reliability.
Documentation enhancements for clarity and usability
Improved profiling by retrieving clock data during initialization
Updated internal system definitions (no user action required)
Corrected transpose operation handling before and after gather connections
Resolved Windows-specific conversion errors
Improved memory handling and DMA job optimization for faster execution.
Enhanced constant tensor prefetch and alignment for better runtime performance.
Improved tensor ROI handling and synchronization in DSP engine.
Improved accuracy in model computations.
Fixed memory handling and alignment issues.
Enhanced stability in layer operations and data processing.
Documentation: Improved LPAI documentation for easier integration.
Documentation: Added instructions for Windows on Snapdragon (WoS) execution.
Documentation: Updated guides for backend upgrade and profiling steps.
Improved system stability and performance
Broader support for model types and data formats
Enhanced compatibility with customer platforms
Better handling of intermediate data for smoother execution
Expanded support for key operations to improve model flexibility
Enhanced SoC logic improves robustness in handling new devices
Refined config file logic in LPAI Extension for multi-threaded stability
Add core selection config option
Support multi-type attributes
enable dspqueue for enpuV6 and enpuV5_1
Profiling improvements
PSRAM memory support
Support Async Execution
Add ubuf config option
Improvements on profiling
LPAI BE adapted to work with 5.9 enpu components
More profiling improvements
Reset Call handling improvements
Added OP support: RmsNorm, Buffer Op Padding Param
LPAI BE adapted to work with 5.8 enpu components
More profiling improvements
Added OP support: StatelessLSTM/GRU
Simplified Lpai BE library names
LPAI Documentation improvements
enpu compile,execute, simulator libraries update
LPAI BE dapted to work with 5.7 enpu components update
More profiling improvements
Added OP support:BatchToSpace,SpaceToBatch,DepthToSpace,SpaceToDepth,Quantize,Dequantize
Add device discovery mechanism across different LPAI Supported HW
Add support for LPAI graph early termination
Fix runtime stability issues
Improved Profiling information
Added OP support: Channel_Shuffle
Add Compiler + Sim for x86 Windows
Fix fastRPC bugs for hex79,hex81
Documentation
mem Register/deRegister
Add Support for Backend Custom properties
Add QNN LPAI island interface
Add model preparation and simulation support on x86 Windows platform
Align QNN LPAI config file defaults
Runtime check for compiled LPAI models older than v4.6, no longer supported
Prepare for LPAI Runtime Island Support
Add profiling support: graph execute time
Bug fix for unsupported float tensor
Bug fix for reset tensors as input
Op update: Add default support of antialias param for ResizeBilinear Op
Ops updates and bug-fixes: Split,LSTM,GRU,Batchnorm,Conv,StridedSlice,ReduceProd
Add profiling support: per layer execution time, layer fusion info, layer linking info
Add Op support: Buffer (Framer)
PerLayer Profiling support via configs
Version Checks for LPAI Execute
Start support for lightweight internal aDSP build
Memory allocation for internal aDSP Execute
Optimize memory planning
Fix memory leaks issue during model conversion
Add Op support: Power, Layernorm, Split
Fixed memory leak 32 properly if initialization is failed
Fixed stability issue related to multiple threads/processes
Add API profiler parameter
Added: Runtime eNPU5 Support, x86 generation, x86 Simulator and ARM execution
Added: LPAI BE Adaptor getMap() APIs
Added: Support to model visualizer, eNPU4 for linux_x86, windows,hexagonsim and adsp
Added: LPAI BE getSupportedOperations()
Fixes: TensorID duplicate errors during QNN change from tensor name to tensorID
Fixes: Sanitize quantizations with invalid parameters
Fixes: Fix resize op to account for annotations on sizes tensor
Added: Support to multiple loaded HW compiler versions
Fixes: Fix resize op to account for annotations on sizes tensor
Fixes: Deconv2D segfault data size calculated from model builder is incorrect.
Fixes: Fix segfault in case of min/max with constant tensor.
Fixes: Expand tensor shape when rank is different.
Fixes: Proper bias alignment for gemm with square output
Fixes: Improve layernorm accuracy for small hidden_size
Added: Support to different permutations of 1d bias matmul.
Added: New memory planning compiler pipeline
Fixes: per-channel quantization
See op support updates
Fixes: symmetric quantization and compression ratio calc
Added: support multiple nodes with same id, input name from onnx node for perlayer dump
Initial release