# Revision History - Archived

| Version | Date | Description |
| --- | --- | --- |
| 2.33.0 | March 2025 | <ul class="simple"><br><li><p>Tool: Added support for defer loading in case of quantization override. {127344}</p></li><br><li><p>Tool:Converter:ONNX: Added multi-graph deduplication to reduce context binary size and improve memory efficiency for LLMs.  {116193}</p></li><br><li><p>Tool:Converter:ONNX: Added support for moving tensors from constant nodes to initializers to improve memory efficiency for LLMs<br>{112211}</p></li><br><li><p>Tool:Converter:ONNX: Improved handling of external data loading for ONNX model initializers, including support for 0-D tensors.<br>{126087}</p></li><br><li><p>SNPE Core: Modified extract_record_from_path python function to only extract the QNN context binary part of a SNPE cache. {124728}</p></li><br><li><p>SNPE DSP/AIP: Fixed SNPE API Snpe_SNPEBuilder_SetInitCacheMode()/SNPEBuilder::setInitCacheMode() breakage for non HTP backends.</p></li><br><li><p>This is exercised via snpe-net-run option –enable_init_cache {129545}</p></li><br><li><p>SNPE DSP: Fixed a bug that prevented the correct execution of signed PDs (placeholder data) in the DSP runtime {109078}</p></li><br><li><p>SNPE DSP: Fixed an issue that prevented devices with DSP architecture v68 from entering sleep mode after running a model. {119271}</p></li><br><li><p>SNPE DSP:Fixed SNPE HTP stress test failures caused by VTCM timeout. {125724}</p></li><br><li><p>Core: Added performance profiling capabilities to the C and CPP sample applications {124508}</p></li><br><li><p>Core: Fixed bug in running signed PDs in DSP runtime {119032}</p></li><br><li><p>Core: Resolved inference time regressions observed in the HTP runtime when using qnn-net-run. {114832}</p></li><br><li><p>Op:HTP: Fixed accuracy failures with HTP FP16 custom MobileNet_v2 mixed precision models due to rounding issues<br>Observed in half-float to signed 16-bit conversion. {121932}</p></li><br><li><p>Tool: Ensured ChannelShuffle output transposed to NCHW when –preserve_io layout is enabled. {123922}</p></li><br><li><p>Tool:Converter: Added support for specifying batch size during model conversion when using the TensorFlow converter. {94854}</p></li><br><li><p>Tool:Converter: Corrected the handling of Bias tensor encodings in FC and MatMul Ops. {120720}</p></li><br><li><p>Tool:Converter: Corrected the layout override logic for the Select Op. {110665}</p></li><br><li><p>Tool:Converter: Fixed a bug in Op sequence matching for the GroupNorm Op {124757}</p></li><br><li><p>Tool:Converter: Fixed an issue in the Converter that ensures disable BN squash when conv node’s weight/bias overrides are present.<br>{124293}</p></li><br><li><p>Tool:Converter: Fixed an issue that conv2d bias gets incorrect scale when input[0] is overridden to float {116164}</p></li><br><li><p>Tool:Converter: Fixed an issue where folding transposes would cause graph output names to be lost. {125412}</p></li><br><li><p>Tool:Converter: Fixed argument parsing errors in QNN converter that caused failures for pytorch models on cloud<br>Observed when used with QNN-AIC tool. {126985}</p></li><br><li><p>Tool:Converter: Fixed the logic of “add_op_to_backend” in QnnCastTranslation {117228}</p></li><br><li><p>Tool:Converter: Resolved an issue where model input encoding was not correctly derived from Quantize/Dequantize Ops<br>after   quantization nodes were removed. {124268}</p></li><br><li><p>Tool:Converter:ONNX: Added support for three new einsum equations, expanding the range of supported ONNX models. {113767}</p></li><br><li><p>Tool:Converter:ONNX: Fixed a bug in axes format population for the Pool Op {118828}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue where GroupNorm Op was not getting correct gamma and beta tensor values,<br>which may have led to accuracy issues in optimized graphs. {119523}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue with the ElementwiseSelect Operator that resulted in incorrect input dimensions. {127541}</p></li><br><li><p>Tool:Converter:ONNX: Fixed graph name inconsistency in qairt lora converter workflow. {104805}</p></li><br><li><p>Tool:Converter:ONNX: Fixed propagation of user encodings in IdentityOp {126845}</p></li><br><li><p>Tool:Quantizer: Fixed an overflow issue in profiling data casting {123812}</p></li><br></ul> |
| 2.32.0 | February 2025 | <ul class="simple"><br><li><p>CPU: Added support for trilinear interpolation in the Resize 5D Op. {122218}</p></li><br><li><p>CPU: Resolved an accuracy issue with MobileBert models on certain chipsets when using the xnnpack matmul implementation. {121590}</p></li><br><li><p>GPU: Fixed an issue where multiple models failed on SM4250-IOT with a Graph Execution failure when using the GPU backend. {121569}</p></li><br><li><p>GPU: Resolved out of memory issues for specific models when running on the GPU backend. {108838}</p></li><br><li><p>HTA: Resolved an issue that caused models with reshape layers to fail on the AIP runtime. {124581}</p></li><br><li><p>HTP: Fixed a timeout issue in SNPE util. {120943}</p></li><br><li><p>Op:CPU: Added support for 5D tensors in Elementwise Comparison Ops. {125481}</p></li><br><li><p>SNPE Core: Resolved an issue that caused network resizing to fail in SNPE versions 2.27 and later. {123315}</p></li><br><li><p>SNPE HTA: Resolved a memory leak issue encountered while running SNPE stress tests on the AIP runtime. {123604}</p></li><br><li><p>SNPE: Fixed a bug where model version option during conversion is not being properly saved into generated DLC. {124903}</p></li><br><li><p>SNPE:CPU: Resolved a memory leak in the DepthwiseConv2d Op. {123629}</p></li><br><li><p>SNPE:GPU: Resolved an issue where SNPE-QNN coexistence stress tests were getting stuck with the GPU backend. {121582}</p></li><br><li><p>Tool:Converter: Added axis tracking for ExpandOp {118618}</p></li><br><li><p>Tool:Converter: Added new sequence for FC/matmul + squash batchnorm in case of optional bias in FC and matmul {119429}</p></li><br><li><p>Tool:Converter: Added support for constant input in BatchNorm Op. {117961}</p></li><br><li><p>Tool:Converter: Added support for constant scalar input in the PyIrConstant tensor. {117953}</p></li><br><li><p>Tool:Converter: Added support for Device NMS (QDetect) in the QNN converter for AIC. {68665}</p></li><br><li><p>Tool:Converter: Added support for partial IO for Preserve IO Datatype. The –preserve_io_datatype flag now accepts specific<br>inputs/outputs for datatype preservation. {117580}</p></li><br><li><p>Tool:Converter: Added support for QAIRT command-line arguments to specify desired input layout and output shape. {106635}</p></li><br><li><p>Tool:Converter: Added validation in match_base_layernorm for layernorm pattern matching {119491}</p></li><br><li><p>Tool:Converter: Enable Constant folding for OneHot Op {116983}</p></li><br><li><p>Tool:Converter: Enabled QnnIR backend for all toolchains having QNNHTP support. {111592}</p></li><br><li><p>Tool:Converter: Fixed an issue where bias was not correctly handled in the GEMM Op during model conversion. {120022}</p></li><br><li><p>Tool:Converter: Fixed an issue where per-channel quantization for Conv2D bias was not being performed when using the<br>–use_per_channel_quantization flag. {118497}</p></li><br><li><p>Tool:Converter: Fixed the broadcasting failure in the squash_eltwise_into_conv method in optimization {119746}</p></li><br><li><p>Tool:Converter: Removed dependency on the Python rich library. {124683}</p></li><br><li><p>Tool:Converter: Updated the Onnx Runtime Version from Onnx-1.17.1 to Onnx-1.18.0 {125506}</p></li><br><li><p>Tool:Converter:ONNX: Added LayoutInferer support for CustomOp, enabling layout transformations for custom operations based<br>on user-defined XML op definitions. {104942}</p></li><br><li><p>Tool:Converter:ONNX: Added support for the ONNX IsNaN Op. {115649}</p></li><br><li><p>Tool:Converter:ONNX: Addressed an issue where input preprocessing encodings were being ignored, leading to unquantized nodes<br>in the converted model. {124609}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue in constant folding for the Where Op. {122514}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue in static alpha conversion for the PReLU Op. {124487}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue in the QNN converter related to batchnorm operator. {115134}</p></li><br><li><p>Tool:Converter:ONNX: Fixed an issue where the converter would fail with an UnboundLocalError when processing subgraphs of the<br>BERT Large model with quantization overrides. {110453}</p></li><br><li><p>Tool:Converter:ONNX: Fixed constant folding logic in ReduceOp translation {125098}</p></li><br><li><p>Tool:Converter:ONNX: Resolved an issue where the converter was adding unnecessary int16 to int8 conversions for reshape ops in<br>float-fallback mode. {126844}</p></li><br><li><p>Tool:Converter:ONNX: Resolved issues in Slice and Concat Op translation. {123355}</p></li><br><li><p>Tool:Converter:ONNX: Support Onnx LSTM with pre-quantized weights and biases {114557}</p></li><br><li><p>Tool:Converter:Relay: Corrected an error in TFLite CustomOp conversion and Conv2D dequantization. {120231}</p></li><br><li><p>Tool:qnn-onnx-converter: Fixed an issue in the ONNX converter that caused errors when handling specific axis layouts in the<br>Buffer Op. {125932}</p></li><br></ul> |
| 2.31.0 | January 2025 | <ul class="simple"><br><li><p>Tool: Converter: Updated command-line argument names for QAIRT Converter and Quantizer to ensure consistency.</p></li><br><li><p>Tool: Added support to display average latency per inference in snpe-throughput-net-run.</p></li><br><li><p>Tool: Converter: Fixed an issue that caused inconsistent input tensor order when using the custom_io option due to the use of<br>std::set.</p></li><br><li><p>Tool: Converter: Fixed an issue that prevented RMSNorm from being folded when the epsilon tensor had more than one element.</p></li><br><li><p>Tool: Converter: Fixed an issue with the ‘axis’ value when merging a reshape_transpose_reshape pattern into a channelshuffle Op.</p></li><br><li><p>Tool: Converter: Added support for unsigned symmetric quantization of activations using the ‘–act_quantizer_schema<br>unsignedsymmetric’ option.</p></li><br><li><p>Tool: Converter: Fixed an issue where Batchnorm weights had an incorrect datatype when using asymmetric overridden encoding.</p></li><br><li><p>Tool: Converter: Fixed an issue where ReLU symmetric external encoding was not being applied correctly.</p></li><br><li><p>Tool: Converter: Corrected GRU Op optimization to properly apply encoding information.</p></li><br><li><p>Tool: Converter: Improved handling of empty inputs for GRU and LSTM Ops.</p></li><br><li><p>Tool: Converter: Updated ONNX framework version information in QNN documentation.</p></li><br><li><p>Tool: Converter: ONNX: Fixed an issue to handle static first input to Gemm Op.</p></li><br><li><p>Tool: Converter: TFLite: Added support for int64 bias in TFLite Conv2D Op.</p></li><br><li><p>Tool: Converter: TFLite: Added a pattern to dequantize constant expressions.</p></li><br><li><p>Tool: qnn-tflite-converter: Fixed an issue where the dequantize node had the wrong datatype in quantized TFLite models.</p></li><br></ul> |
| 2.30.0 | December 2024 | <ul class="simple"><br><li><p>Tool: Converter: Added framework tracing validation to check if all ops and tensors from framework model are traced.</p></li><br><li><p>Tool: Converter: Converter now correctly preserves input tensor datatypes based on the ‘preserve_io datatype’ command-line option.</p></li><br><li><p>Tool: Converter: TensorFlow: Resolved an issue in GenericBatchNorm node fusion.</p></li><br><li><p>Tool: Converter: Resolved a converter issue that caused a failure when using FP16 mix precision due to static shared tensors<br>having inconsistent types.</p></li><br><li><p>Tool: Converter: qnn-onnx-converter: Resolved an issue in framework tracing that caused some tensors to be missed.</p></li><br><li><p>CPU:  Corrected the output index for the MultiClass NMS Op.</p></li><br><li><p>SNPE:DSP: Resolved an issue where running an FP16 model with ‘linting’ or ‘detailed’ profiling levels resulted<br>in a ‘QNN_GRAPH_ERROR_SET_PROFILE’ error.</p></li><br></ul> |
| 2.29.0 | November 2024 | <ul class="simple"><br><li><p>Added 16KB alignment support for Android libraries in QAIRT SDK to enhance memory management.</p></li><br><li><p>DSP: Fixed TF/TFLite model preparation failures on QCS6490.</p></li><br><li><p>HTP: Resolved a race condition during thread group creation, preventing thread exhaustion under heavy system load.</p></li><br><li><p>Tool: Converter: Fixed bug in quantization of Elementwise Binary Ops when the output is non-quantizable and one of the inputs has<br>quantized data type while other input is float32.</p></li><br><li><p>Tool: Converter: Fixed the ONNX converter’s incorrect quantization setting for the third input of the ScatterElements Op</p></li><br><li><p>Tool: Converter: Updated the axis tracking logic for the RoiAlign Op.</p></li><br><li><p>Tool: Converter: Fixed an issue in the Converter that ensures correct assignment of the graph.preserve_io_datatype_passed and<br>graph.preserve_io_datatype parameters.</p></li><br><li><p>Tool: Converter: Mapping int64 inputs to int32 inputs without inserting extra cast.</p></li><br><li><p>Tool: Converter:  Fixed the bug in ElementwiseProduct Optimization.</p></li><br><li><p>Tool: qnn-accuracy-debugger: Fixed bug in qnn-accuracy-debugger when sanitize tensor name following converter’s node naming<br>conventions.</p></li><br></ul> |
| 2.28.0 | October 2024 | <ul class="simple"><br><li><p>SNPE: Fixed high DSP core clock retention when using multiple models with different performance profiles.</p></li><br><li><p>SNPE: Fixed initialization with buffer via DLC API &amp;  builder API behavior on a DLC with valid cache; also addresses<br>init time regression.</p></li><br><li><p>SNPE Core: Added the Snpe_SNPE_SetExecutionPriorityHint() / SNPE::setExecutionPriorityHint() API to allow changing the priority of<br>inferences after network initialization. The snpe-throughput-net-run –priority_hint option now accepts a list of priorities.</p></li><br><li><p>SNPE Core: Added a warning in the logs to indicate missing output tensors when output is not explicitly set via the builder API.</p></li><br><li><p>API: Fixed SNPE execution failure when loading models from a buffer using Snpe_DlContainer_OpenBuffer().</p></li><br><li><p>Tool: Quantizer: Added support for Int32 Quantization Override.</p></li><br><li><p>Tool: Converter: Fixed a shape mismatch error for the Concat Op that occurred under specific conditions involving continuous<br>Concat Ops and Nontrivial layouts.</p></li><br><li><p>Tool: Converter: Fixed RMSNorm fusion for models where the topological order of nodes differs from their sequential order.</p></li><br><li><p>Tool: Converter: Relay: Added support for Quantized BatchMatmul Op in TFLite Converter.</p></li><br><li><p>Tool: Converter: Onnx: Fixed model conversion failures for models with Concat and GridSample operations with varying input layouts.</p></li><br><li><p>Tool: Converter: Fixed incorrect layout setting for Transpose Op’s output tensors. Note: The output tensor’s layout should not<br>be set to Nontrivial in the axis tracking optimization function if the function does not change anything.</p></li><br><li><p>Tool: Converter: Fixed an encoding override issue.</p></li><br><li><p>Tool: Converter: qnn-onnx-converter: Fixed converter failure for Gather Op.</p></li><br><li><p>HTP: Fixed power config ID leak when using SNPE and QNN together that was causing stability issues.</p></li><br><li><p>HTP: Added generic event WER (Windows Error Reporting) for critical SDK errors. WER generation is automatic and always enabled, but<br>submission is controlled by Windows OS privacy settings. Visit “Windows Error Reporting” page in SDK documentation for more details.</p></li><br><li><p>DSP: Fixed crashes that occurred intermittently while running stress tests with interceptors enabled in different performance<br>profiles.</p></li><br></ul> |
| 2.27.0 | September 2024 | <ul class="simple"><br><li><p>SDK: Doc: Updated SDK documentation to include information about using SSH/SCP commands for OE Linux based targets.</p></li><br><li><p>SDK: License: Separated the license restrictions section into two parts: one for general restrictions and another<br>for a prohibited items list.</p></li><br><li><p>SDK: Added support for dynamic tensor shapes in the DLC format.</p></li><br><li><p>Tool: onnx-simplifier: Added the following HTP-specific post-quantization adaptations:<br>* Output transposed keycache: Avoids repetitive transpose of key state tensors.<br>* Output new key value only: Reduces memory traffic.</p></li><br><li><p>Tool: Converter: Added support for 6D ReshapeOp, ElementwiseUnaryOp, ElementwiseOp, ReduceOp, and GatherOp.</p></li><br><li><p>Tool: Converter: Added converter support for the QLinearConv Op.</p></li><br><li><p>Tool: Converter: Added converter support for the QLinearMatMul Op.</p></li><br><li><p>Tool: Converter:  Fixed an issue where the conv+bn fusion was not being disabled when the conv node was the graph output.</p></li><br><li><p>Tool: Converter: Fixed an issue where L2Norm had the wrong axis after sequence matching.</p></li><br><li><p>Tool: Converter: Fixed an issue where the Topk Op’s K value was invalid.</p></li><br><li><p>Tool: Converter: Fixed a bug where quantization overrides for LSTM/GRU Ops were not propagated correctly during Op expansion.</p></li><br><li><p>Tool: Converter: ONNX:  Added translation for if Op.</p></li><br><li><p>Tool: Converter: ONNX: Added support for pattern matching matmul with bias in MHA to SHA conversion.</p></li><br><li><p>Tool: Converter: Fixed GELU fusion for models where the topological order of nodes differs from their sequential order.</p></li><br><li><p>Tool: Converter: Fixed a GroupNorm optimization issue when  the gamma/beta shape was modified.</p></li><br><li><p>Tool: Converter: Fixed bug in GroupNorm fusion for the VAE Encoder model.</p></li><br><li><p>Tool: API: Added optional arguments to the simplify API.</p></li><br><li><p>Tool: Quantizer: Fixed a bug in per-channel bias with float_fallback.</p></li><br><li><p>HTP: Added online prepare support for OE Linux targets based on gcc9.3 toolchain.</p></li><br><li><p>HTP: Fixed Inference time regressions on vgg16 and other models.</p></li><br><li><p>HTP: Fixed a performance issue in Depthwise Convolution when it is the first layer of the model and quantized to int8.</p></li><br><li><p>Core: Added –platform_options=”deviceId:1” support for multi NSP devices in HTP runtime for snpe-net-run and<br>snpe-throughput-net-run.</p></li><br></ul> |
| 2.26.0 | August 2024 | <ul class="simple"><br><li><p>SNPE Core: Optimized memory footprint for networks when input/output user buffers match the tensor data type.<br>For mismatched data types, the first inference will be slower due to memory allocation for conversion.</p></li><br><li><p>Tool: Converter: Onnx: Added Conversion support for “largest” attribute in TopK Op.</p></li><br><li><p>DSP: Upgraded  Hexnnv2 to DSPCore1.53.0.</p></li><br><li><p>Tool: Converter: Added support for antialias attribute of ONNX Resize operator with linear interpolation mode.<br>This is only supported with 4D inputs currently.</p></li><br><li><p>Tool: Converter: ONNX: Added functionality to output only the last logit.</p></li><br><li><p>Tool: Converter: ONNX: Enhanced model conversion efficiency by eliminating superfluous Transpose nodes around Elementwise Op.</p></li><br><li><p>Tool: Converter: ONNX:  Enabled support for additional Einsum equations.</p></li><br><li><p>Op: CPU: Added 6d elementwise Ops.</p></li><br><li><p>SNPE Core: Fixed a bug in user buffer data type conversion from ufxp_8 to uint_8/16/64 and int_8/16/64.</p></li><br><li><p>Op: CPU: Added int8 support for RMS Op.</p></li><br><li><p>Core: Updated zlib version to 1.3.1 to fix CVE-2023-6992 and CVE-2022-37434.</p></li><br><li><p>Core: Added support to read output tensor names from the input list in snpe-parallel-run.</p></li><br><li><p>Tool: Converter: Mapped the cast op to constant op in case of static input to the cast op.</p></li><br><li><p>Tool: Quantizer: Fixed an accuracy issue caused by squashing Relu when using a symmetric quantization profile.</p></li><br><li><p>Core: Improved memory-mapped user buffer registration API to handle duplicate address/offset gracefully,<br>particularly in recurrent networks.</p></li><br><li><p>Core: SNPE: Corrected CSV data display issue in snpe_bench.py script, ensuring accurate result depiction since version 2.22.0.</p></li><br><li><p>Tool: Converter: Fixed a bug in applying quantization overrides when a RMSNorm pattern is folded into RMSNorm QNN Operator.</p></li><br><li><p>Tool: Converter: Added support for string datatype in customOp.</p></li><br><li><p>Tool: Converter: ONNX: Added folding support for new RMSNorm patterns.</p></li><br><li><p>Tool: Converter: Qnnx: Fixed conversion failure due to axis tracking for specific models with qairt-converter.</p></li><br><li><p>Tool: Converter: TFLite: Fixed multiple Converter and Quantizer issues for the FullyConnected Op in QNN TFLite Converter.</p></li><br><li><p>Tool: Quantizer: Qairt: Fixed a bug in applying quantization overrides for static input tensors of data invariant operators.</p></li><br><li><p>Tool: Quantizer: Fixed a bug in converting tensor data from FP32 -&gt; FP16 for FP16 overrides.</p></li><br></ul> |
| 2.25.0 | July 2024 | <ul class="simple"><br><li><p>Tool: Converter: Relay: Added new Op Support for BatchToSpace and SpaceToBatch Ops to the TFLite Converter.</p></li><br><li><p>Tool: Converter: Optimized the implementation of expand LSTM Op structure in the converter.</p></li><br><li><p>Tool: Quantizer: Fixed accuracy drop at output of Cast Op (INT32 -&gt; uFxp8) by inserting Quantize (FP32 -&gt; uFxp8) Op after Cast<br>(INT -&gt; FP32).</p></li><br><li><p>Tool: Converter: Fixed issue where some models with LSTM and NTF format input failed to convert.</p></li><br><li><p>Tool: Converter: Fixed broadcasting error for constant input to Quantize/Dequantize Linear ONNX Ops, ensuring correct input<br>handling.</p></li><br><li><p>Tool: Converter: ONNX: Mapped RMSNorm pattern in ONNX networks to a QNN RMSNorm Op.</p></li><br><li><p>Tool: Converter: Fixed an issue in the calculation of padding for deconv.</p></li><br><li><p>Core: Fixed performance regressions in memory mapped use case with snpe-throughput-net-run.</p></li><br><li><p>Core: Fixed CSV format data display issue in snpe_bench.py, which was broken since version 2.22.0.</p></li><br><li><p>Core: Fixed snpe-net-run to not automatically round float32 userbuffers before conversion to int32/uint32 tensor data types.</p></li><br><li><p>Op: HTP: Improved memory footprint and performance for Conv3D Op.</p></li><br><li><p>Tool: Added Python APIs for converter, optimizer, and quantizer.</p></li><br><li><p>Tool: Converter: Added op support for RMS Normalization.</p></li><br><li><p>Tool: QAIRT: Added a feature to preserve the input/output layout and datatype as in the source framework model.<br>Added a new commandline argument to invoke the feature.</p></li><br><li><p>Tool: snpe-dlc-info: Enabled dumping of framework trace info from DLC file.<br>Added an argument to enable or disable this functionality.</p></li><br><li><p>Tool: Converter: Added fix to remove identity patterns emerging from a sequence of Reshape and Transpose ops.</p></li><br><li><p>CPU: Fixed memory leak for XNNPACK operator.</p></li><br><li><p>Tool: Converter: Added fix to remove additional transpose operations around RMSNorm operation.</p></li><br><li><p>Tool: Converter:  Fixed axes order mismatch in LayerNorm Op and conversion issue in TransposeConv1d Op.</p></li><br><li><p>HTA: Improved Validator for elementwise Op.</p></li><br></ul> |
| 2.24.0 | June 2024 | <ul class="simple"><br><li><p>Added onnx-simplifier and onnx-runtime versions to sdk.yaml.</p></li><br><li><p>Tools: Quantizer: Updated SDK documentation for failing conditions of QAT encodings.</p></li><br><li><p>Core : Added more logs for custom and preset perf profiles.</p></li><br><li><p>Tool: Added infrastructure to support –backend and –soc_model options in qairt-converter &amp; qairt-quantizer tools,<br>enabling the generation of graphs suitable for the chosen backend and SOC</p></li><br><li><p>Core: Add HMX voting support to SNPE.</p></li><br><li><p>Tools: Converters: Fixed an issue in Converter to allow for the Graph input datatype to be correctly updated to FP16 from FP32.</p></li><br><li><p>Tool: SNPE Quantizer: Added fix for Segmentation Fault issue when using algorithms cle flag.</p></li><br><li><p>SDK: Add version information to libraries and executable files.</p></li><br><li><p>Tool:Converters: Fixed performance regressions due to redundant transpose ops introduced during graph optimizations.</p></li><br><li><p>Core: Fixed memory leak in CPU and GPU runtimes while creating and deleting network in a loop.</p></li><br><li><p>Tool: Converter: Fixed Op Config Validation issue when translating grouped ConvTranspose / Deconvolution Op.</p></li><br></ul> |
| 2.23.0 | May 2024 | <ul class="simple"><br><li><p>Tool:Converter: Update clear help message for argument “–enable_framework_trace”.</p></li><br><li><p>Tool:Converter:  Disable framework trace for other converters than onnx converter.</p></li><br><li><p>Core: Added support for traceinfo in dlc.</p></li><br><li><p>Tool:Converter: Added implementation for framework op tracking for graph quantization optimization stage.</p></li><br><li><p>Core: Added perf API support to snpe-throughput-net-run.</p></li><br><li><p>Op:HTP: support TCM for ConvLayer.opt.grpconv_weights.</p></li><br><li><p>Tool: Converters: Relay: Fixed a tflite conversion failure by adding dequantize reduce pattern pass.</p></li><br><li><p>Core: Fixed memory leaks in Android APK with DSP runtime in every classification (inference) request.</p></li><br><li><p>Tools: Fixed a bug in snpe-dlc-graph-prepare where passing –set_output_layers would add the last op of the network even if<br>not passed explicitly.</p></li><br></ul> |
| 2.22.0 | April 2024 | <ul class="simple"><br><li><p>Tools: Converter: Adding reduction attribute as none in case of attribute is not available in original graph.</p></li><br><li><p>Tools: Converter: Div op support is added in Tensorflow converter.</p></li><br><li><p>Tool:snpe-accuracy-debugger: Enable windows x86 Native support for integrated Quant Checker for SNPE SDK.</p></li><br><li><p>Tools: Converters: Cleanup of old references for Binary Coarse op.</p></li><br><li><p>Core: Support 5D tensors for transpose fp16 op in HTP.</p></li><br><li><p>DSP: Introducing low level performance APIs for DSP thus enabling custom performance profile settings for init, inference,<br>de-init and from inference to inference. Also enabling the ability to overwrite partially a preset profile.</p></li><br><li><p>Tool:quantizer: Fixed issue observed when bias of conv op need to be per-channel quantized in mix-precision mode.</p></li><br><li><p>Tools: Converter: Fixed the small bug in onnx softmax translation.</p></li><br><li><p>Core: snpe-dlc-graph-prepare - Fixed a bug with multiple SoC prepare for certain networks.</p></li><br><li><p>API: Added new APIs to quantize and dequantize buffers - Snpe_Util_Convert_Float32ToTfN / IUserBufferFactory::Float32ToTfN,<br>Snpe_Util_Convert_TfNToFloat32 / TfNToFloat32. Please read the API docs for details on usage.</p></li><br><li><p>SDK: Updated Android NDK version to android-ndk-r26c for compiling SNPE/QNN SDK for Android based targets.</p></li><br><li><p>Tool: Converter: Onnx: Fix axis tracking issue for TransposeConv2d.</p></li><br><li><p>Core: Fixed multiple timestamps in logs in android logcat.</p></li><br><li><p>Tools: Pytorch &amp; TFlite Converter: Fix incorrect rounding behavior.</p></li><br><li><p>Tools: Converters: PyTorch: Added support for TransposeConv3d.</p></li><br><li><p>HTP: Fixed accuracy issue for the pattern: Batchnorm -&gt; Relu -&gt; Concat.</p></li><br></ul> |
| 2.21.0 | March 2024 | <ul class="simple"><br><li><p>SNPE library migrated to use static libc++ for Android platform.</p></li><br><li><p>Tool: Quantizer: Added a new standalone qairt-quantizer tool equivalent to snpe-dlc-quant. This new tool takes a float DLC and<br>produce a Quantized or Mixed Precision DLC.</p></li><br><li><p>Tool:Converters: Added support for sparse tensors.</p></li><br><li><p>Core: Renamed DSP_v68 folder at &lt;SNPE_SDK&gt;/examples/SNPE/NativeCpp/UdoExample/&lt;OpName&gt;/src to HTP.</p></li><br><li><p>Tools: Converter: enable 16bits QuantizeLinear/DequantizerLinear in Onnx converter.</p></li><br><li><p>Tools: Converters: ONNX: Added support for ThresholdedRelu ONNX op.</p></li><br><li><p>Tools: Converter: Downcast to_type for Cast op from int64 to int32.</p></li><br><li><p>Tool: Converter: Added PyTorch ChannelShuffle support.</p></li><br><li><p>Tool: Converter: Added TFLite LOCAL_RESPONSE_NORMALIZATION support.</p></li><br><li><p>Tool: Converter: Onnx: Support FP16 model conversion.</p></li><br><li><p>Tools: Converter: TFLite: Fixed an encoding mismatch issue at input and output layers when converting a pre-quantized tflite network.</p></li><br><li><p>Tool:snpe-dlc-quantize: Fix the bug causing snpe-dlc-quantize to fail with multi-dot dlc filenames.</p></li><br><li><p>Core: Fixed Snpe_Util_SetSNPEStorageLocation/ SNPEFactory::setSNPEStorageLocation to not create duplicate kernel<br>repo file (GpuKernelRepo.pb) for GPU and delegate it to the backend (gpukernelcache.qti.aisw)</p></li><br><li><p>Core: Fixed resource release in burst and sustained_high_performance mode.</p></li><br><li><p>Tool: Converter: Onnx: Bug fixed to support different layout of axis in TopK.</p></li><br><li><p>Tool: Pytorch Converter: Fixed scalar indices issue for gather op.</p></li><br><li><p>Addressed SSR (SubSystem Restart) occurring during SNPE_Execute().</p></li><br><li><p>Tools: Converter: Added qairt-converter tool. This converter tool takes a Pytorch/Onnx/Tensorflow/TFLite network and<br>converts it to a DLC file representing the QNN graph format that can enable inference on Qualcomm AI IP/HW.<br>Please refer Documentation or AppNote for more details.</p></li><br></ul> |
| 2.20.0 | February 2024 | <ul class="simple"><br><li><p>Tools: Converters: Added Converter support for MaskedSoftmax Operator.</p></li><br><li><p>Core: Added new option –userbuffer_memorymapped_shared to snpe-parallel-run to enable sharing memory mapped user buffers.<br>allowing different tensors to register using the same address/file descriptor and a unique byte offset.</p></li><br><li><p>Tools:qnn-pytorch-converter: Enabled preserve_io feature.</p></li><br><li><p>Tools: Converters: support hardsigmoid in onnx converter.</p></li><br><li><p>Tools: snpe-dlc-info updated to display output tensors and unconsumed internal tensors in separate tables.</p></li><br><li><p>Tools: Added support for snpe-diagview tool for OE Linux GCC11.2 toolchain based targets.</p></li><br><li><p>Core : Added CPU INT8 support to Softmax UDO example.</p></li><br><li><p>Tools: qnn-pytorch-converter: add support of aten::upsample_linear1d for pytorch converter.</p></li><br><li><p>Tools: Converter: PyTorch: added PixelShuffle support for pytorch converter.</p></li><br><li><p>Core: Fixed snpe-diagview –chrometrace stats computations related to averaging when multiple input sets are involved.</p></li><br><li><p>Core: Updated the sample app MemoryMappedUserBuffer to reflect proper usage of memory mapped userbuffer.<br>registration/de-registration, once in SNPE’s lifecycle and not once per execute.</p></li><br></ul> |
| 2.19.0 | January 2024 | <ul class="simple"><br><li><p>Core: Added support for sharing regular userbuffers (non memory-mapped) via single buffer and offsets.</p></li><br><li><p>SDK: Added supported SOC table to SDK documentation.</p></li><br><li><p>Core: Added new option –validate_cache to snpe-net-run and snpe-throughput-net-run to validate<br>HTP cache before network initialization.</p></li><br><li><p>Tools: Converters: Added the support for assigning input dtype in PyTorch converter.</p></li><br><li><p>Tool: Converter: Pytorch: Fixed the issue of duplicate buffer names for two identical transpose ops.</p></li><br><li><p>Tools: Fixed bug to correctly convert shared static tensor to FP16.</p></li><br><li><p>Tools: Converter: Supported optional initial_h and initial_c in Onnx bidirectional LSTM.</p></li><br><li><p>Tools: Converter: TFlite: Fixed data type mismatch issue for TFLite pre-quantized model.</p></li><br></ul> |
| 2.18.0 | December 2023 | <ul class="simple"><br><li><p>Tools: Converters: Updated squashing logic to avoid removing model outputs.</p></li><br><li><p>Core: Added new SNPE builder API s Snpe_SNPEBuilder_SetCacheCompatibilityMode() / SNPEBuilder::setCacheCompatibilityMode()<br>to set HTP cache compatibility mode for cache selection.</p></li><br><li><p>Tools: Converter: PyTorch: Raise an error for custom op support in SNPE product.</p></li><br><li><p>Core : Added Relu UDO example to SNPE sdk.</p></li><br><li><p>Tools: snpe-dlc-graph-prepare - Added a new option –num_hvx_threads for reserving no of HVX threads for a graph running on HTP.</p></li><br><li><p>Tools: Converters: support group_norm in pytorch converter.</p></li><br><li><p>Core: ArgMax example added to UDO examples in SNPE SDK.</p></li><br><li><p>Core: Updated CAPI Sample App for Memory Mapped User Buffer to demonstrate usage of dmabuf (libdmabufheap.so).</p></li><br><li><p>Tools: Converters: Add Gather and Take support in pytorch converter.</p></li><br><li><p>Tools: Converters: PyTorch: Added support for AvgPool1d/MaxPool1d/AvgPool3d/MaxPool3d/AdaptiveAvgPool1d/<br>AdaptiveMaxPool1d/GlobalAvgPool1d/GlobalMaxPool1d.</p></li><br><li><p>Tools: Converter: Onnx: Enforce h/c input buffers of LSTM to be NONTRIVIAL.</p></li><br><li><p>Tools: Converter: Support BROADCAST_TO in tvm tflite frontend.</p></li><br></ul> |
| 2.17.0 | November 2023 | <ul class="simple"><br><li><p>Core: Added –userbuffer_memorymapped_shared option in snpe-net-run and snpe-throughput-net-run to exercise shared memory<br>mapped buffers.</p></li><br><li><p>Added product and OS info to SDK.yaml file.</p></li><br><li><p>Core: Added new APIs Snpe_UserMemoryMap_AddFdOffset / UserMemoryMap::add() and<br>Snpe_Util_CreateUserBufferShared() / IUserBufferFactory::createUserBufferShared() to enable sharing memory mapped user buffers<br>allowing different tensors to register using the same address/file descriptor and an unique byte offset.</p></li><br><li><p>Tools: Converters: Updated algorithm to fix Tensor Layout from Constant operator when it is located ahead of Concat Operator.</p></li><br><li><p>HTP: updated backend extensions config - changed graph object to graph array to allow different graphs have different set of<br>properties.</p></li><br><li><p>Tools: Converter: Fixed param name parsing issue in pytorch converter.</p></li><br><li><p>Tools: Pytorch converter: Added support for OneHot op.</p></li><br><li><p>SNPE Core: Added support for targetSdkVersion 32 in SNPE apk.</p></li><br><li><p>Core: Update MemoryMappedUserBuffer sample app to demonstrate shared buffer usage. Added support for named input parsing<br>in all the sample apps.</p></li><br><li><p>Core: Enhanced error messages for API failures in userlogs as well as Snpe_ErrorCode_GetLastErrorString() /<br>DlSystem::getLastErrorString()</p></li><br><li><p>Tools:ONNX Converter: Fixed WhereOp axis format issue.</p></li><br><li><p>Tools: Converter: fix tensorflow strided_slice conversion for out of range start/end.</p></li><br><li><p>DSP Runtime: Added DSP reset / Subsystem Reset (SSR) error handling for logging API: Snpe_Util_SetLogLevel() /<br>SNPEFactory::setLogLevel().</p></li><br><li><p>Core: Fixed registration of memorymapped user buffer with multiple addresses against the same tensor name.</p></li><br></ul> |
| 2.16.0 | October 2023 | <ul class="simple"><br><li><p>SDK: Add support for Mobile SoC: SM8650</p></li><br><li><p>Core: Added new APIs Snpe_UserMemoryMap_AddFdOffset / UserMemoryMap::add() to enable sharing memory mapped user buffers allowing<br>different tensors to register using the same address/file descriptor and an unique byte offset.</p></li><br><li><p>Core: Added pre-emption count captured as yield count in logs in SNPE.</p></li><br><li><p>Core: Fixed redundant rpc memory allocations during graph initialization for memory-mapped (zero copy) buffer usage by providing<br>hints during graph prepare. snpe-dlc-graph-prepare option –memorymapped_buffer_hint introduced.</p></li><br><li><p>SDK: Added sdk.yaml in SDK to capture build and other version info.</p></li><br><li><p>Tool: Common: Quantizer: Added a Quantizer pass to make static inputs of Elementwise Op float if the output is overridden to float.</p></li><br><li><p>Core: Performance improvements done in network initialization time - up to 1.5x speed up for certain networks with HTP offline cache.</p></li><br><li><p>Tools: Converters: Added broadcast support for layernorm op weights and bias.</p></li><br><li><p>Tools: Converters: Added rectangular SpaceToDepth op support to handle SpaceToDepth pattern in Pytorch model.</p></li><br><li><p>DSP Runtime: Added HTP DLBC(Deep Learning Bandwidth Compression) option for graph preparation.</p></li><br><li><p>Tools: Converters: Added batch_norm ND support in tflite/pytorch converter.</p></li><br><li><p>Tools: PyTorch Converter: Add support for custom op in QNN product.</p></li><br><li><p>Tools: enable LSTM operator in SNPE.</p></li><br><li><p>Tools: Converters: Onnx: Fixed conversion failure for gather op with scalar indices.</p></li><br><li><p>Tools: Quantizer: Fixed an issue by not converting Cast to Convert if next op is float.</p></li><br><li><p>Core: Fixed Snpe_SNPEBuilder_SetInputDimensions() / SNPEBuilder::setInputDimensions() /snpe-net-run option –input_name xxx<br>–input_dimensions yyy to evaluate new dims when a compatible dsp cache record is present. If new dims are accepted offline cache<br>will be rejected in favor of online preparation.</p></li><br><li><p>Tools: PyTorch Converter: Fixed parameter quantization override.</p></li><br><li><p>Tools: Converters: Fixed a conversion failure when folding Concat Ops.</p></li><br><li><p>Tools: Converters: Pytorch: Fixed an issue with applying overrides.</p></li><br><li><p>Tools: Converters: Onnx: Fixed a conversion failure when Onnx inferShape API returns an empty graph.</p></li><br><li><p>Tools: Converters: Onnx: Fixed a quantization failure for networks having Float16 activations.</p></li><br><li><p>Core: Implemented coexistence of DSP/HTP cache records prepared with different input dimensions. Added option to specify input<br>dimensions in snpe-dlc-graph-prepare. Cache selection logic updated to match dimensions passed during graph initialization.</p></li><br><li><p>Tools: Converters: Onnx: Add support for Gather Op with negative indices.</p></li><br><li><p>Tools: Converters: Updated the validation to see if the weights of FC and BN are eligible for optimization of BN into FC.</p></li><br><li><p>Core: Logging from backends is made conditional based on SNPE logging API invocation.</p></li><br><li><p>Docs: Updated inceptionv3 documentation to include LU / LE toolchains.</p></li><br><li><p>SNPE HTA: Added support of Pooling 16bit for large dimensions.</p></li><br><li><p>HTP: fixed graph prepare issue due to edge mod pad.</p></li><br></ul> |
| 2.15.0 | September 2023 | <ul class="simple"><br><li><p>Tools: snpe-dlc-info tensor columns rearranged and HTP cache info section updated to display UDO information,<br>Optimization level etc.</p></li><br><li><p>Core: snpe-dlc-graph-prepare updated to overwrite existing cache with similar signature but from older cache version by default</p></li><br><li><p>SNPE GPU: Extreme power saver performance profile fixed to map to the lowest profile on the SoC instead of highest.</p></li><br><li><p>SNPE AIP: Extreme power saver performance profile now maps to lowest profile available on the SoC.</p></li><br><li><p>SDK: Fixed broken links in PSNPE C API documentation.</p></li><br><li><p>SNPE DSP: Extreme power saver performance profile for DSP v66 devices now maps to lowest performance profile available on the SoC.</p></li><br><li><p>Core: Memory Mapped Userbuffer Sample App - added error handling for incompatible data types.</p></li><br><li><p>Core: Fixed –debug not emitting intermediate tensors for offline cache based execution.</p></li><br></ul> |
| 2.14.0 | August 2023 | <ul class="simple"><br><li><p>Core: HTP offline cache records prepared for DSP architecture v68, v69 (sm8350/sm8450) will be rejected on SoCs<br>with DSP architecture v73 and above (sm8550).</p></li><br><li><p>Tools: Converter: Onnx: Added default attribute perm for Transpose Op.</p></li><br><li><p>Tools: Converter: Tensor with no consumers and not an actual graph output will be set to NATIVE for QNN Onnx Converter.</p></li><br><li><p>Tools: Converter: Allow only output tensors in the source model to be marked as QNN_TENSOR_TYPE_APP_READ. All other tensors with<br>zero consumers will change from being APP_READ to NATIVE.</p></li><br><li><p>Tools: Converters: Onnx: Added negative max_output_boxes_per_class parameter support for NonMaxSuppression.</p></li><br><li><p>Tools: Converter: update tvm version to support pytorch 1.13 version.</p></li><br><li><p>SDK: Update documentation contents for standalone SDK.</p></li><br><li><p>Tools: Converters: Added a Graph pass that matches Space2Depth Op (CRD &amp; DCR) from Reshape - Transpose - Reshape pattern.</p></li><br><li><p>Tools: quantizer: Avoid act’s bw changing according the weight/bias’s bw.</p></li><br><li><p>Op:HTP: added uint8 support for maxpool w77s44p00.</p></li><br><li><p>Core: SNPE de-initialization moved to a separate thread for all profiling levels to better affine to faster CPU core(s)<br>thus improving de-init time for most graphs.</p></li><br><li><p>Tools: Converters: Resolved OpValidation error related to LayerNorm Op caused due to the unsqueezed Gamma/Beta<br>tensor being &gt; 1D rank.</p></li><br></ul> |
| 2.13.0 | July 2023 | <ul class="simple"><br><li><p>SDK: Updated Revision history formatting in SNPE docs.</p></li><br><li><p>Tools: Converter: GRU weights are shared across time unrolling step.</p></li><br><li><p>Documents: Update latest PyTorch Op support.</p></li><br><li><p>Tools: Converter: Allow only output tensors in the source model to be marked as QNN_TENSOR_TYPE_APP_READ.<br>All other tensors with zero consumers will change from being APP_READ to NATIVE.</p></li><br><li><p>SNPE DSP: added support for uint8 window7x7 stride3x3 maxpool ops on HTP.</p></li><br><li><p>Tools: snpe-dlc-graph-prepare - introducing new option –optimization_level. Higher optimization levels incur longer<br>prepare time but yields more optimal graph and hence faster execution time for most graphs.</p></li><br><li><p>Tools: Converter: Changed the logic for converting 1dOp into 2DOp by expanding along H dimension instead of W dimension.</p></li><br><li><p>Op:DSP: added support for logSoftmax.</p></li><br><li><p>Tools: Converter: Changed the translation of FloorDiv operator to ElementWiseDivide if the datatype of input is Int32.</p></li><br><li><p>Tools: Introducing –userbuffer_memory_mapped option in snpe-net-run, snpe-throughput-net-run and snpe-parallel-run for general<br>memory mapped userbuffer use cases(like ionbuffers in Android).</p></li><br><li><p>Tool: TF Converter: added support for conv2d_transpose layer with asymmetric strides.</p></li><br><li><p>Tools: TF Converter: Support optimized Gelu pattern that contains Mul instead of Realdiv.</p></li><br><li><p>API: Generic APIs added for memory-mapped userbuffers in lieu of existing ion buffer registration/de-registration APIs.</p></li><br><li><p>Core: Added CAPI based Sample Apps for userbuffer and memory-mapped buffers (like ion buffer).</p></li><br><li><p>HTP: Fixed bug in ReduceMean optimization during prepare.</p></li><br><li><p>Tools: snpe-diagview - fixed “Snpe Accelerator Time” and “Accelerator Time” data being larger than “Total Inference Time”.</p></li><br><li><p>SDK : libCalculator_Skel.so added to lib/hexagon-v68/unsigned and lib/hexagon-v69/unsigned folders of SNPE SDK.</p></li><br></ul> |
| 2.12.0 | June 2023 | <ul class="simple"><br><li><p>Tools: Converters: Added a new transformation to change MatMul into FullyConnected even without Bias.</p></li><br><li><p>Tools: Converter: TFlite: Added a fix to account for the difference in the offset sign and usage when quantizing tensors.</p></li><br><li><p>SNPE DSP: Introduced new extreme power saver performance profile to enable ultra low power inferencing usecases on HTP runtime.</p></li><br><li><p>SNPE Core: All binaries are now built with libc++, not libstdc++ for X86/Linux.</p></li><br><li><p>SNPE Core: All binaries are now built with clang9 instead of clang7.</p></li><br><li><p>Tools: Converters: Modified the output names generated by Pytorch Converter and TFlite Converters.<br>Also changed the axis tracking behavior to match the TF &amp; Onnx Converters.<br>This may change the name and layout for the output layer of the model.</p></li><br><li><p>API: Fixed Snpe_SNPEBuilder_SetTimeOut/SNPEBuilder::setTimeOut() to failure when Snpe_SNPE_ExecuteUserBuffers()/SNPE::execute()<br>fails to return within the timeout duration.</p></li><br><li><p>Tools: Quantizer: Fixed an issue that prevented weights &amp; bias inputs of Batchnorm from being set as FP16.</p></li><br><li><p>Tools: Quantizer: Fixed an error related to locking the WeakPtr associated with the Bias tensor to Convolution Op.</p></li><br></ul> |
| 2.11.0 | May 2023 | <ul class="simple"><br><li><p>SDK: dependencies.sh changed to check-linux-dependency.sh and check_python_depends.sh changed to check-python-dependency.sh.<br>Also envcheck.sh added.</p></li><br><li><p>Core: SNPE C++ APIs are deprecated. To maintain backwards compatibility C++ header-only wrapper APIs are included in the SDK<br>that invokes SNPE C APIs internally.</p></li><br><li><p>Op: ONNX converter: added support for Mod.</p></li><br><li><p>CPU: INT8 support enabled for LA targets.</p></li><br><li><p>Core: Caffe source framework models are no longer supported in SNPE.</p></li><br><li><p>Core: arm-32 platform is no longer supported in SNPE.</p></li><br><li><p>SDK directory structure is updated.</p></li><br><li><p>Documentation refreshed with a new look and feel and contents are enhanced.</p></li><br><li><p>Core: New C API s added to match the deprecated C++ API capabilities.</p></li><br><li><p>Core: Init time improvements for most models with HTP offline cache record.<br>(Note that the offline cache needs to be regenerated to take advantage of this improvement)</p></li><br><li><p>Core :Native Cpp example for Platform Validator with C APIs is now functional.</p></li><br><li><p>Tools: snpe-net-run now allows –debug when input list has output op names (# )or output tensor names (% ) specified<br>in the first line.</p></li><br></ul> |
| 2.10.0 | April 2023 | <ul class="simple"><br><li><p>GPU Runtime: Support Pack operation with 1 input.</p></li><br><li><p>Core: Updated C API documentation for ITensor/Userbuffer creation indicating data size.</p></li><br><li><p>Core: setLogLevel() API hooked up to the runtimes for updating logging level after creating logger handle.</p></li><br><li><p>Tools: snpe-throughput-net-run now supports –userbuffer_auto option (similar to snpe-net-run) for automatic IO tensor<br>data type detection.</p></li><br><li><p>Tools: Converters: Added a new optimization sequence to squash BatchNorm into FullyConnected.</p></li><br><li><p>HTP: Fixed issue with ElementwiseSin.</p></li><br><li><p>Tools: Fix the converter issue for GRU op.</p></li><br><li><p>SNPE AIP: Fixed perf profile setting for multithread scenario.</p></li><br></ul> |
| 2.9.0 | March 2023 | <ul class="simple"><br><li><p>Core: Added new C API Snpe_SNPE_GetInputDimensionsOfFirstTensor() to facilitate retrieving Input dimension without Input tensor name.</p></li><br><li><p>Tools: ONNX converter: Added support for NonMaxSuppression op.</p></li><br><li><p>Tools: snpe-dlc-graph-prepare fix benign error message during offline prepare for v68 based SoC s (–htp_socs sm8350, sm7350 etc)</p></li><br></ul> |
| 2.8.0 | February 2023 | <ul class="simple"><br><li><p>Tools: Converters: Onnx: Added support for Sign.</p></li><br><li><p>HTP: solve vtcm overflow issue happened when change data layout: from uint8 flat to uint8 crouton in tcm.</p></li><br><li><p>Tool:ONNX Converter: Fixed TransposeOp input axis format NT issue.</p></li><br></ul> |
| 2.7.0 | January 2023 | <ul class="simple"><br><li><p>Tools: Converters: Fixed a bug in the optimization that merges Matmul + Reshape + Add to FC Op that would incorrectly insert the</p></li><br><li><p>FC Op before the Constant Bias Op.</p></li><br></ul> |
| 2.6.0 | December 2022 | <ul class="simple"><br><li><p>Tools: onnx converter: support conv’s input data is Initializer.</p></li><br><li><p>DSP: Improve execute time of dynamic depthwise convolution with uint8 weights.</p></li><br><li><p>Core: Added error handling based on buffer data size in execute().</p></li><br></ul> |
| 2.5.0 | December 2022 | <ul class="simple"><br><li><p>Tools: Added new options for snpe-net-run and snpe-parallel-run –use_native_input_files and –use_native_output_files to support<br>inputs in their native format as opposed to default float32 format.</p></li><br><li><p>Tools: Added new flag –userbuffer_auto in snpe-parallel-run to automatically detect and use the right buffer type based on tensor<br>data type in the model.</p></li><br><li><p>Documentation: SNPE1 to SNPE2 migration guide is added.</p></li><br><li><p>Tools: snpe-throughput-net-run - capturing the status of lost thread in the result summary.</p></li><br><li><p>Tools: snpe-dlc-quant: Fixed abnormal DLC size increase when axis quantization is used.</p></li><br><li><p>Tools: Tensorflow Converter: Fixed issues with per-channel quantization of weights: set is_symmetric = true by default, added param<br>“axis” and “is_symmetric” into weight encodings info.</p></li><br><li><p>HTP: solve vtcm overflow for transposeconv2d layer whose groups &gt; 1, in depth= out depth, padding =0 and groups != in depth.</p></li><br></ul> |
| 2.4.1 | October 2022 | <ul class="simple"><br><li><p>Tools: New tools - snpe-architecture-checker &amp; snpe-quantization-checker are added.</p></li><br><li><p>snpe-net-run: Added new flag –userbuffer_auto to automatically detect and use the right buffer type based on tensor data type in the<br>model</p></li><br><li><p>SNPE Core: Enabled logging in Op validation.</p></li><br><li><p>SDK: Added missing documentation files for snpe-quantization-checker.</p></li><br><li><p>GPU Runtime: Improved network initialization time in subsequent runs on GPU when using setInitCacheMode.</p></li><br><li><p>Tools: ONNX Converter: fixed issue related to missing Cast operation.</p></li><br><li><p>Tools: Missing files for snpe-quantization-checker have been added to the SDK.</p></li><br><li><p>Tools: Fixed functional failure for snpe-architecture-checker.</p></li><br><li><p>Tools: Quantizer: Improve Error handling to remove ‘uncaught exception’ errors.</p></li><br><li><p>Tools: Fixed bug in snpe-dlc-quantize with option –axis_quant and –enable_htp when multiple socs are passed using –htp_socs.</p></li><br><li><p>GPU Runtime: Fixed validation errors for Concat op with large dimensions.</p></li><br><li><p>GPU Runtime: Improved accuracy in models having Concat op with large dimensions.</p></li><br><li><p>DSP Runtime: Bug fix in running HTP FP16 networks on non fp16 supported SoCs (like sm8350, sm7350)</p></li><br><li><p>GPU Runtime: Fixed verifier issue in Softmax2UdoPackage.</p></li><br><li><p>GPU Runtime: Improved network initialization time in subsequent runs on GPU when using netrun –storage_dir option.</p></li><br></ul> |
| 2.3.1 | September 2022 | <ul class="simple"><br><li><p>Tools: Converters: Onnx: Added 5D tensor support for PoolMax3d.</p></li><br><li><p>Tools: GoogleNAS: Added support for utilizing the GoogleNAS service with SNPE hardware in the loop (HIL).</p></li><br><li><p>Tools: Quantizer: Added fix to use default activation bitwidth for static tensors instead of default parameter, except for static<br>tensor that are known to be parameters like convolution weights and bias</p></li><br><li><p>SNPE Core: Fix online dequantization of int4 axis quant dlc when ran on CPU/GPU.</p></li><br><li><p>SNPE Core: Fixed stability with concurrency use cases.</p></li><br><li><p>GPU Runtime: Fixed accuracy issues related to tensor memory optimization.</p></li><br><li><p>Tools: Quantizer: Fixed issue observed with int4 weight override support.</p></li><br></ul> |
| 2.2.1 | August 2022 | <ul class="simple"><br><li><p>Core: Added userlogs ( –userlogs=warn) for Op validation failures for both offline and online prepare thereby making it easier to<br>track fallback.</p></li><br><li><p>Core: HTP Offline Cache Blob backward compatibility - Snpe Version check relaxed from SNPE-2.2.1 onwards.</p></li><br><li><p>Tool: Converters: Added DepthToSpace DCR/CRD pattern that matched reshape, transpose, reshape nodes.</p></li><br><li><p>Core: Fixed dlc-info to display per axis encoding information for axis_quant dlcs.</p></li><br><li><p>Tools: Quantizer: Added support for CLE quantization algorithm.</p></li><br><li><p>Core: snpe-dlc-graph-prepare bug fixes-bound –vtcm_override to the maximum VTCM for each SOC chipset requested instead of<br>a hardcoded<br>8MB. Limit to 1 cache record per SoC in the dlc</p></li><br><li><p>Core: Fix runtime de-quantization of weights and biases for axis quantized dlcs when executing in floating point backends (CPU/GPU).</p></li><br><li><p>Tool: Onnx Converter: Added axis tracking edge case fixes for Concat and MatMul operations.</p></li><br><li><p>Core: Added protection for loading malicious dlc file.</p></li><br><li><p>Converter: change the output dims as the node output axis format order.</p></li><br><li><p>Core: SNPE::Execute() API updated to validate input/output buffer map size before proceeding.|</p></li><br><li><p>Core: snape-dlc-quantize - fixed error in handling % in input list.</p></li><br><li><p>Tools: snpe-dlc-quantize miscellaneous bug fixes with –output_dlc option.</p></li><br><li><p>Tools: Converter: Resolved bug that caused failure to override weight encodings for Conv Ops.</p></li><br><li><p>Tools: Quantizer: Fixed issues related to axis quantization when the model contains TransposeConv2D.</p></li><br><li><p>Tools: Converter: Fixed bug in elementwise min and max sequence optimization.</p></li><br></ul> |
| 2.1.1 | July 2022 | <ul class="simple"><br><li><p>Core: Re-Enable LSTM support for CPU, GPU (HTP will follow).</p></li><br><li><p>DSP Runtime: Implemented rules for coexistence and selection of multiple cache records for HTP based on VTCM size, DSP Architecture,<br>and SoC</p></li><br><li><p>Tools: Converter: Added optimization to fold scalar min + max to ReluMinMax.</p></li><br><li><p>Tools: Quantizer: Re-enabled support for overriding activation quantization (overriding weight quantization will follow).</p></li><br><li><p>Tools: Quantizer: Fixed missing skip_quantization command line argument in the new snpe-dlc-quantize shell script.</p></li><br><li><p>Tools: Quantizer: Fixed axis quantization failure.</p></li><br><li><p>Tools: Quantizer: Fixed issues with quantizing inputs to the gather op.</p></li><br><li><p>Tools: Converter &amp; Quantizer: Update converter and quantizer to persist the command used in the DLC that can be displayed in<br>snpe-dlc-info.</p></li><br><li><p>Tools: DLC Viewer: Fixed to support the new DLC Format.</p></li><br><li><p>C API: Added new Snpe_DlContainer_OpenBuffer() to support loading a model from a buffer.</p></li><br><li><p>Docs: Fixed C API documentation related to creating a User Buffer.</p></li><br><li><p>Core: Change default option for SNPEFactory::isRuntimeAvailable() to UNSIGNEDPD_CHECK from NORMAL_CHECK. Note that this also affects<br>the C API.</p></li><br><li><p>Core: Re-enable NV21 input processing support.</p></li><br></ul> |
| 2.0.1 | June 2022 | <ul class="simple"><br><li><p>Added support for SM8550.</p></li><br><li><p>Added new C API. This API is in addition to the C++ API. Note that the APIs cannot be mixed, all code should use one or the other.</p></li><br><li><p>Updated the DLC internal format to use ‘ops’ rather than ‘layers’ to more closely align the graph definition with QNN.</p></li><br></ul> |
| 1.64.0 | June 2022 | Tool: Onnx Converter : Reenabled converter command line input dtype to take precedence over model specified.<br>GPU: Improved accuracy in deepsort model, Resolved issues with Conv + elu op fusion.<br>Tools: Quantizer: Fixed issue observed with applying 8-bit overrides using 16-bit default activation quantization encodings.<br>SNPE Core: Fixed failure to select HTP offline cache for certain multi-subnet network topologies. |
| 1.63.0 | May 2022 | SNPE Core: Support PRELU bias broadcasting in SNPE.<br>SNPE Core : snpe-diagview tool updated to display actual units (like cycles) instead of usec by default.<br>SNPE Core: Open GL buffers supported for GPU backend.<br>SNPE Core : Fixed Zip utility’s std::istream index to internal extensible array to be const for every container(DLC) load. |
| 1.62.0 | April 2022 | DSP Runtime: Perf improvement for FP16 models on HTP.<br>Added GatherV2 support for SNPE-QNN-DSP.<br>Tools: Converters: Added 5D tensor annotations NCDHW and NDHWC support.<br>Tools: Converters: TF: Fixed issue with translating explicit padding from Conv Op.<br>Tools: Converters: Onnx: Fixed Onnx Concat axis.<br>Tools: ONNX Converters: Fixed implementation details for Conv1D and Pool1D Ops.<br>Tools: Converters: Onnx: Added optimization folding continuous reshapes. |
| 1.61.0 | March 2022 | Tools: Converters: Onnx: Enabled support to handle custom op inputs correctly when the default values are provided.<br>Tools: ONNX Converter: Added support to resolve static ONNX Cast operation as Constant.<br>CPU Runtime: Supported CRD mode for depthtospace(pixelshuffle).<br>Improved performance of loading DLC from a memory buffer.<br>Fixed scale calculation for ONNX Resize Operator for align\_corner mode.<br>Also overrides Resize input axis format as per source axis order. |
| 1.60.0 | February 2022 | Tools: Converter: Added ONNX Gemm transA and transB support.<br>Native sample code is updated to take static quantization parameters for quantized input buffers.<br>libSNPE.so, libcalculator.so, libplatformValidatorShared.so, libnpe\_dsp\_domains\_v2.so - libraries generated with gcc8.2 and<br>gcc9.3 toolchain - are now compiled with additional read-only relocation compiler flags.<br>HTP: Fixed issue with Cast op usage in certain configurations.<br>ONNX Converter: Improvements to handle different input axis layouts. |
| 1.59.0 | January 2022 | DSP Runtime : Added support for edge padding from SNPE side.<br>Tools: ONNX Converter: Limited support for Expand operator when it can be interpreted as a noop from operator attributes.<br>Tools: ONNX Converter: Added support for ScatterND.<br>Tool: Quantizer: Fixed duplicate Convert layer Id issue observed in generated DLC when multiple Convert layers feed into a single layer.<br>Tool: ONNX Converter: Fixed handling of models with inputs of unknown shape.<br>Tools: ONNX Converter: Resolves issue where Shape operator translation could fail if the input was part of the initializer list. |
| 1.58.0 | December 2021 | Tools: Converter: Enabled broadcasting of weights and bias for BatchNorm layer to match channel dimensions. |
| 1.57.0 | November 2021 | Tool: Onnx Converter: Added support in dry-run mode to handle reporting ops that are not in onnx schema domain.<br>Tool: Converter: - Updated inaccurate macs/params calculations for Ops per re-analysis.<br>CPU Runtime: Set the max detections to keep top K for Caffe SSD network.<br>Tool Converter: Removed obsolete ssd\_permute\_param parameter in caffe converter permute translation.<br>SNPE DSP: Fix axis quantization not adding all the fixedPointParam of output to bufferDeltas.<br>Tool Converter: Fixed coefficient input broadcasting issue for ONNX Prelu operation.<br>Tool Converter: Fixed axis tracking bug for permute when input is btf format. |
| 1.56.2 | October 2021 | DSP Runtime: Caffe SSD models can now run fully on HTP, but show some performance issues.<br>Tool: Converter: Added new layernorm sequence for pattern matching and added a constraint to enforce MatMul layer’s constant<br>second input to 8-bit tensor in quantized model. |
| 1.55.0 | September 2021 | Added support for the OneHot operation across the SNPE converters with runtime support available on SNPE CPU.<br>Tool: ONNX Converter: Added support for LSTM & CRNN.<br>DSP Runtime: Added support for LSTM.<br>Tools: Converters: Added support for Caffe Power scale/shift parameters.<br>SNPE DSP: Fixed the issue of invalid cache record added to DLC while doing offline prepare for HTP.<br>Tools: Converters: Fixed Softmax and Reduction Ops to have default case for output\_buf axis format. |
| 1.54.0 | August 2021 | Tool: TF Converter: Added support for detecting eltwise pattern for batchnorm layer with fakequant inputs.<br>Tools: Converters: Adds support for Caffe Reduction layer Sum and Mean Ops.<br>Tool: Quantizer: Added support to make Convert Operator upscale and downscale quantization parameters loss free.<br>ONNX Converter: Add support for LSTM & CRNN in converters.<br>DSP Runtime: Add support for LSTM.<br>Tool: Converters: Added batch dimension to anchor input data conversion from tensorflow corner style to center style for<br>DetectionOutput operation optimization.<br>Tool: ONNX Converter: Added support to pre-apply ONNX batchnorm scale and bias quantization encodings before getting consumed by<br>Converter to compute weights and bias.<br>Add support for reverse engineering SAME padding mode from the explicit pad values. |
| 1.53.2 | July 2021 | Tool: Quantizer: Added support for fake quant operators in snpe-dlc-quantize.<br>Tools: TF Converter: Support for logical\_and, equal, greater, greater\_equal, less, less\_equal, not\_equal, logical\_or, select.<br>Tool: TensorFlow Converter: Added support for Identity nodes that act as graph output nodes.<br>Tool:ONNX converter: Fixed incorrect default bias shape for ConvTranspose translation. |
| 1.52.0 | June 2021 | Tools: Converters: Removes pre-broadcasting of constant tensors resulting in smaller file sizes in converter output.<br>Tool: Converter: Added Converter support for Nd Reshape layer.<br>Tool: Converter: Added CastOp support for TF.<br>Tool: Converter: Added support for static subgraph resolutions at conversion time.<br>Tool: Converter: Added support for tensor dtype for TF fill op translation.<br>SNPE DSP: Fixed variance accuracy loss in InstanceNormalization on HTP.<br>SNPE GPU : Added optimized kernel for ReduceMean Operation.<br>Tool: Converter: Fixed bug in TF fullyconnected translation where input was intermittently out-of-order.<br>SNPE DSP: Fixed the issue of freeing the uninitialized pointer that is leading to random crash.<br>SNPE DSP: Optimized specific unpack-&gt;elementwise sequences for certain models on HTP. |
| 1.51.0 | May 2021 | Tool:Converter: Added supported for Onnx WhereOp.<br>Added support for edge padding type for pad operation in GPU runtime.<br>SNPE DSP: Enabled support for ElementWiseUnary abs layer on HTP.<br>GPU Runtime: Added support for asymmetric reflect padding for pad operation.<br>UDO: Allow users to specify a different datatype for each core in single config file.<br>UDO: HTML documentation & sample app is updated to provide example for loading UDO package.<br>DSP Runtime: Fixed the context leak on HTP targets during repeated init/deinit scenarios.<br>SNPE: Init stage is optimized to be done faster.<br>SNPE DSP: Optimized maxpool with stride 2x1 on HTP.<br>SNPE DSP: Optimized the big sized concat ops to fit into memory.<br>SNPE DSP: Optimized the init on HTP.<br>SNPE DSP: Graph prepare is optimized for HTP targets to be able to run bigger graphs.<br>SNPE DSP: Fixed the issue with CDSP not going to sleep when the model is de-initialized. |
| 1.50.0 | April 2021 | Tool: Quantizer: Added SNPE Quantizer support for is\_symmetric field used in updated AIMET specification.<br>DSP Runtime: Improved instance norm op accuracy when input size is big.<br>DSP Runtime: Enabled edge padding support for v65/v66 targets.<br>Tool: Tensorflow Converter: Resolved Xiaomi issue where TF Mul was not being translated correctly. |
| 1.49.0 | March 2021 | ONNX Converter: Added support for ONNX 1.6 (Opset 11)<br>TF Converter: Added support for TF2.3 models.<br>TFLite Converter: Add initial TFLite converter.<br>ONNX Converter: Add support for YOLOv2, YOLOv3, tiny-YOLOv3, and YOLOv5.<br>DSP Runtime: Optimize conversion performance to/from 16-bit quantized values on HTP.<br>Converters: Improve detection and removal of unconnected nodes.<br>ONNX Converter: Add support for DETR model.<br>AIP Runtime: Optimized the input and output data format conversion times for specific depth configurations for models having 16bit<br>activations.<br>DSP Runtime: Enabled support for Matmul on HTP.<br>snpe-throughput-net-run: Fix input\_list processing when using multiple batches.<br>TF Converter: Fixed inconsistent network topology flow differing between runs for larger models with forking nodes.<br>DLC Quantizer: Fixed a race condition that might result in integer overflow.<br>Android Sample App: Fix to work correctly when multiple models are packaged, with only some requiring UDO.<br>snpe-diagview: Fixed crash bug when using circumstances AIP runtime networks with UB\_FLOAT and UB\_TF8 buffer modes with init caching.<br>DSP Runtime: Additional model support with offline prepare. |
| 1.48.0 | February 2021 | SDK: Migrated to use Ubuntu 18.04 as the host platform.<br>SDK: Updated dependencies.sh and check\_python\_dependencies.sh for the transition to Ubuntu 18.04 - Python 3.6, and libc++9.<br>SDK: Removed the system variants for the DSP stub libraries.<br>Tool: Switched diagnostic logging (SNPEDiag.log) to a new file format.<br>Added static buffer mapping support for frequently used buffers in DSP runtime.<br>Improved instance norm op accuracy when input size is big in DSP runtime.<br>Added support for Unsigned PD with the AIP runtime.<br>Tools: Converters: Fixed a bug which might prevent applying quantization overrides to a model.<br>Fixed NMS op code in HTP core.<br>Input node followed by concat node is optimized in HTP.<br>SNPE DSP: Fixed Unpack Layer indexing error on HTP.<br>DSP Core: Fixed overflow issue in instance norm op when variance is too small. |
| 1.47.0 | January 2021 | Added support for TF 1.15 for NonMaxSuppressionV3 translation in TF converter.<br>Added support for Normalize layer translation in Caffe converter. |
| 1.46.0 | December 2020 | Improved CDSP power voting by using client specific context id.<br>SNPE DSP: Improved argmax op performance by optimizing l2 cache prefetch and replacing int to float cast op.<br>Improved the input/output data conversion times on AIP runtime for specific depth configurations.<br>Enabled the support for random inputs for networks having more than one input layer in SDK benchmarking scriptsTool:<br>qnn-tensorflow-converter: Removed ?allow\_unconsumed\_nodes option from TF converter as it is now the default.<br>Enabled elementwise sub and div on HTP. |
| 1.45.0 | November 2020 | Beginning with SNPE 1.45.0, users must install libc++1-8 using apt-get or other package manager in order to perform offline cache<br>generation for HTP.<br>Optimized shallow convolution (depth &lt;= 4) in inputsupernode for v66 DSP.<br>Remapped previous converter translation of Caffe Tile layer as ConcatOp to TileOp in Caffe Converter.<br>Fixed small accuracy regression on VGG and Flownet models in DSP runtime.<br>Named the HTA threads created on CDSP appropriately.<br>Improved logging when libcdsprpc cannot be found.<br>Fixed the issue with AIP runtime being unavailable on the Android R platforms.<br>Fixed the issues with HTA metadata generation of conv2d op. |
| 1.44.0 | October 2020 | Optimized concat performance when input size is very big in DSP runtime.<br>Optimized slice performance when split 3 channel RGB input in DSP runtime.<br>Removed support for proposal layer in DSP runtime.<br>Added SNPE converter support for Softmax axis parameter.<br>Added support for consuming AIMET/custom quantization encodings to override quantizer generated encodings.<br>Fixed an issue on graphs where the final node in a graph was an elementwise operation with more than two inputs.<br>Fixed bug where output\_shapes were calculated as float values for DepthToSpace and SpaceToDepth Ops.<br>Changed ONNX converter to not allow negative or placeholder dimensions.<br>Fixed potential issues with some models where QAT nodes may not get propagated properly to the final converted model. |
| 1.43.0 | September 2020 | Improved the input/output conversion times for models having depth as 4 on AIP runtime.<br>Enabled initial support for constant layers along with elementwise Op on HTA.<br>Added support for opaque float concat operation in SNPE DSP concat layer.<br>Added support for Caffe’s “Clip” layer in the caffe converter.<br>Added int16 example to snpe-sample app.<br>Fixed the crash while running multi-threading applications with user buffer mode on AIP runtime.<br>Fixed bug in ONNX converter that used a hard-coded name for the sequence length input of the LSTM operator.<br>Fixed bug in ONNX converter for Unsqueeze layer, which got a key-error with static inputs.<br>Fixed the bug in l2\_fetch usage during output conversion which improved the performance significantly for some models running on<br>AIP runtime.<br>Fixed the issue with generation of HTA enabled dlc for denoise model.<br>Fixed the segmentation fault issue during dlc generation with specific inputs on HTA.<br>Fixed issue with PlatformValidator.hpp reference to non-existent #include. |
| 1.42.2 | September 2020 | Fixed the bug in l2\_fetch usage during output conversion which improved the performance significantly for some models running on<br>AIP runtime. |
| 1.42.0 | August 2020 | Removed V60 DSP libs from SNPE SDK.<br>Enabled the AIP runtime support for generating the intermediate outputs from HTA with online compiler.<br>Enabled multithread for re-quantize process in DSP runtime.<br>Added optional parameter to set the hysteris period for sustained high and burst profiles in DSP runtime.<br>Added support for opaque float concat operation in SNPE DSP concat layer.<br>Fixed bug in UserBufferTF8 where retrieving the encoding would always return null.<br>Fixed box decoder performance issue on mobilenet v2 ssd model for DSP runtime.<br>Fixed tanh performance issue by replacing QuantizedTanh\_8\_ref with QuantizedTanh\_8 op in DSP runtime. |
| 1.41.0 | July 2020 | Added MatMul support on the CPU runtime.<br>Added support for new version of 7250 with integrated PMIC module.<br>User Defined Operations(UDO) with weight parameters have been added to demonstrate both quantization and network execution on CPU<br>and DSP runtime cores<br>respectively.<br>Optimized tile Op in DSP runtime, that used 2d memcpy for w-d plane tiling and HVX for tiling along depth.<br>Fixed stack overflow issue in concat layer in DSP runtime.<br>Fixed issue with input for multibatch in DSP runtime.<br>Fixed issue in TF converter that prevented FusedBatchNorm operations from being merged into previous Convolution layer.<br>Fixed DSP crash issue due to stack overflow Concat layer preparation. |
| 1.40.0 | June 2020 | Added DSP Graph Caching support for AIP models with HVX subnets.<br>Upgraded DSP to use Hexagon SDK 3.5.2 toolchain.<br>Added support for 16 bit UDO layers in DSP.<br>Added support for large average pooling, reduce\_mean layer and improved elemetnwise\_mul support for larger tensor size.<br>Fixed the issue with buffer ordering during the execution of batched models on AIP runtime.<br>Fixed issue with SsdDetectionOut when number of classes is only 1.<br>Fixed accuracy issue with Correlation 1D op.<br>Fixed improper processing when 16bit input quantization is used in certain cases.<br>Fixed scaling logic in convert\_16 op. |
| 1.39.1 | May 2020 | Fixed the performance regression of Mobilenet SSD model on AIP runtime. |
| 1.39.0 | May 2020 | The SNPE license (LICENSE.pdf) has been updated, please review it for more details. Additionally the REDIST.txt has been removed, as<br>redistribution is covered in the license.<br>Added graph caching support which improves init times for DSP & AIP networks. (DSP subnet with in AIP is not supported)<br>Optimized Prelu to reduce saturation loss during re-quantization at prelu by using cubic approximation in AIP runtime.<br>Fixed the input conversions to allocate the required buffers during initialization itself, to improve the inference time for<br>AIP runtime.<br>Fixed potential bug with freeing threads in DSP runtime.<br>Added additional logging messages for debugging in DSP runtime.<br>Added support for the AIP runtime in the SNPE sample “snpe-sample”.<br>Added support for BBox transform layer in Caffe2 converter.<br>Added new opset support in the ONNX converter: ArgMax, ArgMin, Concat, PRelu, ReduceMean, ReduceMax, ReduceMin, ReduceSum,<br>Squeeze, Unsqueeze, MatMul, Flatten, Max, Split, Clip.<br>Added support for the fixed-point version of the MobileNetV3 model with H-Swish neuron in TF converter.<br>Improved support of resizing in Crop layer for TF and Caffe converter by introducing new ?counts? parameter.<br>Fixed issue of incorrect UDO tensor datatype in quantizer.<br>Fixed issue with setting the performance profile mode for HTA from AIP runtime in multi-threading use cases that could cause<br>performance to drop.<br>Fixed issue with snpe\_bench.py memory profiling. |
| 1.38.0 | April 2020 | Enabled FC/MatMul to use VTCM if available in DSP.<br>Optimized 16-bit MeanVarianceNormalize in DSP runtime.<br>Added support batchwise scalar divide operation in DSP runtime.<br>Optimized Hard-swish operator for mobilenetV3.<br>Added support for EltwiseMin layer for ONNX converter and CPU runtime.<br>Added support for Onnx BatchNorm layer (OpVer 9, 12) in Onnx Converters.<br>Caffe preprocessing subtract\_mean layer is added. If specified, converter will enable preprocessing specified by a data layer<br>transform\_param subtract\_mean.<br>ONNX softmax converter support only existed for rank &lt;= 2. Support for tensors rank &lt;= 4 was added.<br>Enabled the end-user / developer to request the use of an unsigned process domain to avoid the requirement of signed libraries for<br>SNPE execution on 8250 and newer devices.<br>Removed autoquantization for classes output in MultiClassNMS layer and added support for float addition in ElementwiseOp layer to<br>handle this case.<br>Fixed the issue with enabling stats for AIP runtime on models where number of layers in HTA subnet is more than SNPE layers.<br>Fixed the output conversions to allocate the required buffers during initialization itself in AIP runtime, to improve the<br>inference time.<br>Enabled honoring of padding information from the HTA driver which is pre-computed by AIP runtime earlier, to unblock execution of<br>more models.<br>Fixed the issue with output buffer id while converting depth2space to deconv on HTA.<br>Fixed a bug during graph transformation while folding the batchnorm on HTA.<br>Increased DCVS relaxed sleep latency duration, this will let power system know that CDSP can goto deeper sleep state. If there is no<br>active request for inferencing, it is better for system to go in deeper sleep state. |
| 1.37.0 | March 2020 | Enabled the online compiler support for HTA 1.x family of devices.<br>AIP performance profiles behavior is aligned similar to DSP runtime for reduced power consumption in case of inference inactivity.<br>ONNX Converter: Added support for Onnx Pad layer (OpVer 11).<br>Added support for the h-swish layer used by MobileNet V3.<br>Removed support for the Generate Proposals, ROI Align, and ROI Proposal layers.<br>Added improved support for the reporting of Exceptions in the Java API.<br>Updated the DSP UDO header file to be compatible with SNPE 1.37.0.<br>The DSP UDO support is updated to be compatible with Hexagon SDK 3.5.1.<br>The network creation action was moved onto another thread to avoid impacting the affinity for the main thread of the calling program.<br>Snpe-dlc-info: Fixed issue in MACs calculation error for deconvolution layer.<br>Avoid crash on SDM845 and other v65 targets when unable to retrieve VTCM memory.<br>Fixed an issue in the TensorFlow converter where the weights in the Fully Connected layer were incorrectly transposed.<br>Fixed the support for using DSP UDO with the AIP runtime. Previously, the UDO packages would not be properly loaded in the AIP runtime.<br>Fixed DiagLog data for a UDO on GPU, where it did not report proper values for start and stop.<br>Enable support for keras batchnorm with empty mean and variance to a default values.<br>Fixed a memory leak when using IsRuntimeAvailable() with the VOLATILE\_CHECK for the DSP runtime. |
| 1.36.0 | February 2020 | Added Java API extension to register UDO package with SNPE.<br>snpe-dlc-info now prints the command-line that was used to quantize the DLC if applicable.<br>Added support to handle UDO layers with multiple TF8 outputs with different quantization parameters.<br>Added support for an additional profiling level (moderate) for SNPE benchmarking script and associated snpe-net-run executable<br>for tracking initialization time metrics.<br>Upgraded DSP to use Hexagon SDK 3.5.1 toolchain.<br>Extend Platform Validator to detect HTA API version.<br>Add VOLATILE\_CHECK Mode for SNPE DSP Runtime Checking to query runtime availability in each call instead of giving cached result.<br>Performance modes like LOW\_POWER\_SAVER, HIGH\_POWER\_SAVER, LOW\_BALANCED added for CPU runtime.<br>Fixed bug with propagation of model version during conversion.<br>Fixed the issue with selecting the correct output shape during graph transformation while inserting1x1 conv2d for different<br>input format.<br>Fixed the issue with allocation of layer descriptor while loading the network on HTA. |
| 1.35.0 | January 2020 | Introduce the User-Defined Operations (UDO) feature.<br>Added support for SDM720G/SM7125.<br>Added support to snpe-throughput-net-run for UserBuffer input tensors (both INT8 and INT16).<br>Input batching support is added for networks that can run completely on AIP runtime.<br>Add support for the tf.stack and tf.unstack ops to the DSP and CPU runtimes.<br>Add support for the tf.stack, tf.unstack, tf.floor, tf.minimum to the TF converter.<br>Fixed some small memory leaks that are seen when repeatedly calling dlopen()/dlclose() on libSNPE.so.<br>Updated the Deconvolution operation on DSP with a new kernel that improves performance on various kernel sizes and strides.<br>Fix ssd\_detection CDSP crash on DSP runtime.<br>Updated the HTA to partition the input layer, if it has a connection to a layer that is not included in the same partition.<br>Improved the tiling configuration support for depth wise convolution layer. |
| 1.34.0 | January 2020 | Initial support for ops with 16-bit activations using HTA in both snpe-dlc-quantize and in the SNPE AIP runtime.<br>New option for snpe-net-run to automatically turn unconsumed tensors of the network (tensors that are not inputs to a layer)<br>into network outputs.<br>Fixed inconsistent results on SM8250 in certain cases for depthwise convolutions.<br>Add support for the depth2space operation on the GPU.<br>Using optimized Softmax implementation in AIP networks when input activation has more than 5000 elements.<br>Truncate detection output on DSP to return valid data only.<br>Ensure weights are properly flushed to DDR for use during inference in the DSP runtime.<br>Fix support for NV21 encoding in the DSP runtime. |
| 1.33.2 | November 2019 | Address accuracy issues for Deconvolution in the AIP runtime.<br>Changed behavior of Crop layer resize, so it retains the number of copied elements on each dimension.<br>Make quantizer ?override\_params work for AIP.<br>Reordered PerformanceProfile\_t to be ABI compatible with 1.32.0.<br>Using optimized Softmax implementation in AIP networks when input activation has more than 5000 elements. |
| 1.33.1 | November 2019 | Fixed a build issue that incorrectly removed Symphony. |
| 1.33.0 | November 2019 | New performance modes have been added:<br>LOW\_POWER\_SAVER: Run in lower clock than POWER\_SAVER, at the expense of performance.<br>HIGH\_POWER\_SAVER: Run in higher clock and provides better performance than POWER\_SAVER.<br>LOW\_BALANCED: Run in lower balanced mode, provides lower performance than BALANCED.<br>snpe-dlc-info adds a summary of the layer types in use in the model.<br>Updated to use new BLAS functionality that leverages OpenMP. This adds a new dependency on the OpenMP shared library for<br>Linux platforms.<br>Added 32-bit bias support.<br>Support init caching for SSD output layer on DSP.<br>Fix memory leak causing increasing init time on DSP.<br>Add converter support for dilated convolution when used with fakequant nodes.<br>Multiple bugs fixed in snpe-onnx-to-dlc that were causing errors for models having torch.Mul op.<br>Extends TF converter support to NMSv1 Op in addition to existing support for v2 and v3 NMS Ops.<br>Tensorflow conversion bug fixed in infer\_shape for StridedSlice Op. output\_shape should not be a list of shapes but the shape<br>of the one output.<br>Fix bug with propagation of model version during conversion.<br>If burst mode is set, set thread affinity to Big Cores during init and de-init, and restore to the previous setting after the<br>actions are complete.<br>Fix segfault when using user buffers with a resizable dimension. |
| 1.32.0 | Oct 2019 | Add Caffe MVN Layer support in the Caffe Converter, CPU Runtime, and DSP Runtime<br>snpe-dlc-quantize: Enable the use of quantization parameters calculated during training when using dlc quantizer. To override the<br>SNPE generated<br>quantization parameters pass ?override\_params to [|](https://docs.qualcomm.com/doc/80-63442-10/topic/revision_history_archived.html#id1)snpe-dlc-quantize.<br>Removed deprecated command line arguments from converters. All three converters now require passing -i/?input\_network for model<br>input paths.<br>snpe-dlc-diff: Added command-line option [?diff\_by\_id/-i] to snpe-dlc-diff. This option allows users to compare 2 models in<br>order (sorted by id)<br>Added support for L2Norm layer to TensorFlow converter<br>Optimized the DSP performance for the ‘Space To Depth’ layer<br>Add support in the Java API for setInitCacheEnabled(), and setStorageDirectory() to enable DLC caching support.<br>Allow graceful recovery after a fastrpc error - Recreate the userPD after the cDSP crashes so that the user can continue on the<br>SNPE process with subsequent<br>instances, instead of having to close the SNPE process. Note: [|](https://docs.qualcomm.com/doc/80-63442-10/topic/revision_history_archived.html#id3)all the instance associated to the previous userPD will be lost.<br>snpe-dlc-viewer: Associate each layer type to a fixed color for consistency when using snpe-dlc-viewer<br>Split the SNPE isRuntimeAvailable method into two separate functions to improve backward compatibility with existing client<br>binaries that were built against the older signature.<br>TF Converter: Fix Elementwise Broadcast support<br>ONNX Converter: Fixed bug where output dimension was incorrect when keep\_dims parameter was set to False for Argmax,<br>ReduceSum and ReduceMax.<br>ONNX Converter: Fixed bug where pad attribute was not properly parsed for Deconv Op.<br>Caffe Converter: Fixed bug when converting SSD-based models when using Python 3.<br>TF Converter: Fixed bug where converter was removing const Op input to reshape op when passed through identity op(s).<br>i.e const-&gt; identity -&gt; reshape.<br>Fixed bug where getOutputSize() would give the wrong result on output tensors in UserBuffer mode |
| 1.31.0 | September 2019 | New patterns were added to enable running the CLE algorithm on more op patterns and model architectures.<br>Added Tensorflow converter support for Caffe-style SSD networks.<br>Added support for HeatmapMaxKeypoint layer in the CPU runtime.<br>Added support for ROI Align layer in CPU runtime.<br>Added initial L2Norm layer support in CPU runtime. No support for axis parameter yet: normalization is performed along the<br>inner-most dimension of the input tensor.<br>Support for single-input Concatenation layers was added to CPU, GPU and DSP.<br>Changed determination of number of batch dimensions in the Fully Connected layer so rank greater than 1 is always assumed to mean<br>that there is 1 batch dimension.<br>Removed constraint on the LSTM layer in the GPU runtime that prevented batch mode operation.<br>Added support for Leaky-RELU in the TensorFlow converter. Both the actual Leaky-Relu op and the elementwise op representation<br>are supported and map to SNPE’s Prelu op.<br>Added Argmax support to the Caffe converter, and optimized performance on the DSP runtime.<br>Added new column to snpe-dlc-info that displays the supported runtimes for each layer.<br>Fixed an edge case where in certain conditions OpenCL would return CL\_INVALID\_WORK\_GROUP\_SIZE.<br>Made isRuntimeAvailable Java API thread-safe.<br>Replace unstable image from sample Android classifier application data set with an image that is more consistent. |
| 1.30.0 | August 2019 | Documentation has been added to reflect the new common converter command line options for input processing; Converters now propagate<br>required batchnorm information for performing quantization optimizations; Support for the new bias correction quantization optimization<br>which adjusts biases by analyzing float vs quantized activation errors and adjusting the model to compensate; ONNX converter now<br>filters single input Concats as a no ops as SNPE didn’t support them; Converter input processing now uniformly handles different input<br>types and encodings; ONNX converter now supports the ConvTranspose ‘output\_padding’ attribute by adding an additional pad layer<br>after the ConvTranspose op; Integrates the latest flatbuffer 1.11 library which brings speed improvements and options for model size<br>reduction; GPU size limitations with the ArgMax op (when setting the keepDims op attribute to false) can be worked around by enabling<br>CPU fallback; Fixed DSP error with MobileNet SSD on QCS403 and QCS405; Fixed the issue with partitioning of deconv layer in HTA; |
| 1.29.0 | July 2019 | Added support for dlc reorder tool;Optimization of HTA d32 conversions;Added tf space\_to\_depth op for SNPE CPU and DSP runtime;<br>Benchmarking scripts enhanced for showing further break down of execution time, across various components;Added support for additional<br>ONNX binary element-wise ops;Optimized deconv layer for improving performance;Fixed an issue related to runtime error in DSP runtime;<br>Performance Optimization of SNPE GPU Runtime for Shufflenet V2 by using profiling level config |
| 1.28.0 | June 2019 | Added an optional argument to isRuntimeAvailable for the DSP runtime so that it doesn’t activate the DSP; Allow UB\_T8 and UB\_FLOAT<br>output for snpe-net-run; Added a new command line option for snpe-dlc-diff to check layer names; Updated the –dlc argument to<br>–output\_path for snpe-caffe-to-dlc to align with the ONNX converter; Added –dry\_run argument to snpe-onnx-to-dlc to allow evaluation<br>for successful conversion on an ONNX model; Added support for the gather op in the DSP runtime; Added support to convert the TF<br>MobileNet-V1-FPN-SSD model; Fixed a memory leak in the DSP runtime that is seen when repeatedly loading and unloading a network;<br>Addressed issues on V66 DSPs related to acquiring VTCM memory; Fixed an issue related to multiple inputs for the Caffe converter; Fixed<br>an issue in the TF converter related to element-wise sun and the atrous parameter; Fixed an issue in the TF converter related to<br>tf.crop\_and\_resize when there are only 2 inputs.; Fixed additional cases of uncaught exceptions with the aarch64-android-clang6.0<br>platform; |
| 1.27.0 | May 2019 | Added new APIs support for setting output tensor names to snpeBuilder and to fetch output tensor names for a given output layer name;<br>Improved the peak memory usage with DLC v3 format; Fixed few issues with performance and runtime failures on DSP runtime; Fixed few<br>issues and improved error handling for platform validator; Fixed the issues with Pooling and Instance norm layers of Tensorflow<br>converter; Removed [\*](https://docs.qualcomm.com/doc/80-63442-10/topic/revision_history_archived.html#id5)-android-gcc4.9 platform support. This compiler has been retired for the Android NDK, so all support is<br>transitioning to using Clang for Android; Removed arm-linux-gcc4.8hf platform. The development platform has been retired; |
| 1.26.0 | Apr 2019 | Added support for the ONNX Gather Op in the ONNX Converter and CPU runtime; Optimized DeConvolution Layer for the DSP runtime; Support<br>for tf.nn.moments in the TF converter, CPU and DSP runtimes; Added TF Reflect Pad support for the DSP runtime; Add symmetric quantizer<br>option in snpe-dlc-quantize; Add support for batch &gt; 1 when using the Scale Layer on the DSP runtime; Updated Platform Validator python<br>script to be OS-independent; Added additional optimizations for HTA input conversion; |
| 1.25.0 | Mar 2019 | Updated DLC format to improve load time performance and memory consumption. Old DLCs will continue to work as is, but new DLCs<br>generated from 1.25 will use the new format; Added support for optimized; MultiClassNms and ArgMax ops on DSP runtime; Added option to<br>request larger memory allocations on the DSP for improved init time, at the expense of more memory use; Improved concurrency for<br>multiple; SNPE objects running simultaneously on DSP; Improvements when using priority control on DSP; Added support for channel<br>shuffle and ArgMax in the ONNX converter; Support multiple subnets within the AIP runtime; |
| 1.24.0 | Feb 2019 | Adding setProfilingLevel API support for AIP and CPU runtimes; Various stability issues on aip runtimes are addressed;Added support for<br>Snapdragon 712;Support multi inputs and multiple outputs on each SNPE AIP?s subnet |
| 1.23.0 | Jan 2019 | Upgrade to Android NDK r17c to build SNPE; Improving initialization and de-initialization times; Various DSP timing fixes; Addressed<br>some DSP concurrency edge cases that could impact output values; TF converter support for non max suppression, crop and resize Ops |
| 1.22.0 | Nov 2018 | Support for several new ops on DSP runtime; Upgrade to Android NDK r16b to build SNPE; setProfilingLevel API support in DSP runtime;<br>Added new tool snpe-throughput-net-run |
| 1.21.0 | Oct 2018 | Tensorflow converter and CPU runtime support for various ops; DSP runtime support for Eltwise Realdiv and Square ops; GPU support for<br>resize\_align\_corners layer |
| 1.20.0 | Sep 2018 | Support for QCS605 LE platform; NDK version upgrade to r14b; Tensorflow converter support for elementwise sqrt and softmax with<br>dimension &gt; 2; Platform validation command line tool |
| 1.19.0 | Aug 2018 | ELU op support for Tensorflow/Onnx Converters and CPU/GPU runtimes; BoxWithNMSLimit and BBoxTransform ops support in caffe2 converter;<br>Support for Caffe Power Layer in GPU |
| 1.18.0 | Jul 2018 | Support for pad and elementwise subtraction on GPU; ONNX converter support for shape and pad ops; Tensorflow converter support for<br>additional ops |
| 1.17.0 | Jun 2018 | Support for Scale Layer in Caffe converter and DSP runtime, DSP support for batch&gt;1 and ChannelShuffle, Updated SDK examples for<br>Inception v3 2016 model |
| 1.16.2 | May 2018 | Remove linkage to libstdc++.so in DSP loader libraries |
| 1.16.1 | May 2018 | Remove linkage to libstdc++.so, DSP runtime fixes, fix for 1D BatchNorm |
| 1.16.0 | May 2018 | Batch&gt;1 support (except DSP runtime); layer optimizations for DSP runtime; Caffe2 ChannelShuffle support (except DSP runtime) |
| 1.15.2 | Mar 2018 | Fix for GPU runtime memory leak and reshape to/from 1D |
| 1.15.1 | Apr 2018 | Fix for converter for instance normalization followed by scale |
| 1.15.0 | Apr 2018 | Support for instance normalization for Caffe and Caffe2, MobilenetSSD (Caffe) |
| 1.14.1 | Mar 2018 | Minor fixes |
| 1.14.0 | Mar 2018 | ONNX converter (alpha), multiple enhancements and fixes |
| 1.13.0 | Feb 2018 | GPU and DSP v65 performance improvements. GPU floating point 16 support. |
| 1.12.0 | Jan 2018 | Support for Android LLVM/libc++, MobilenetSSD (TensorFlow) |
| 1.10.1 | Dec 2017 | Fix a bug in the DSP runtime when using mixed userbuffer input types |
| 1.10.0 | Dec 2017 | Support for Mobilenet on DSP, enhanced DSP runtime, Snapdragon Flight Board, updates for UserBuffers |
| 1.8.0 | Nov 2017 | Mobilenet support on CPU, GPU, Support for Snapdragon 636 and Android 64 bit |
| 1.6.0 | Oct 2017 | Support for Snapdragon 450, minor updates and fixes |
| 1.4.0 | Aug 2017 | Support for Snapdragon 630, FasterRCNN and ADSP on AGL |
| 1.2.2 | July 2017 | QDN release |
| 1.2.0 | June 2017 | Beta Caffe2 Converter |
| 1.0.2 | May 2017 | Support for 820AGL platform, Snapdragon 660, and Compute DSP on Android |
| 1.0.1 | Apr 2017 | Documentation update only |
| 1.0 | Apr 2017 |  |

Last Published: Aug 06, 2026