8. Release Notes

8.1. Software Manifest

The Software Manifest lists all open-source components included in this release, along with their versions and license information.

8.1.1. TIDL_PSDK_11.2

  • New features

    • Updated TI TVM fork to upstream TVM version 0.18.0 (from 0.12.0)

    • Replaced neo-ai-dlr with TVM runtime for inference across all interfaces and example applications

    • Added TVMC command-line interface for simplified model compilation on C7™ NPU

    • Added TIDL offload support for object detection post-processing (ODPostProc) via TIDL Meta Architecture

    • Added TIDL offload support for Transformer layer patterns (Layer Normalization, GeLU)

  • Bug fixes

    • Revamped TIDL offload compilation mechanism for improved reliability

8.1.2. TIDL_PSDK_8.6

  • New devices

    • Added AM62A support. To compile a model for AM62A, set the platform argument to "AM62A"; see Compiling Models for details.

  • Bug fixes in C7™ NPU code generation

    • Fixed DMAPass on/off-chip source check

    • Fixed scatter_nd code generation

8.1.3. TIDL_PSDK_8.5

  • Bug fixes in C7™ NPU code generation

    • Added support for non-contiguous access patterns in DMA

    • Added support for strided access patterns in the Streaming Engine

    • Added support for odd numbers of DMA blocks and iterations

    • Simplified resize2d index computation

    • Optimized maxpool2d with 1x1 pool size

  • TIDL offload improvements

    • Rewrote broadcast 1D add as bias_add and offloaded to TIDL

    • Disabled TIDL offload for resize2d with asymmetric or non-power-of-2 scaling factors

    • Enabled TIDL batch processing

    • Rewrote conv2d with 1x2 strides as conv2d with 1x1 strides followed by maxpool2d with 1x2 strides

    • Added TIDL offload support for sigmoid

  • TVM extensions

    • Added support for calling external functions with DLTensor arguments

    • Added example demonstrating how to override the default TVM strategy for an operator

    • Added example demonstrating how to override the default TVM strategy for a specific case

8.1.4. TIDL_PSDK_8.4

  • Bug fixes in C7™ NPU code generation

    • Fixed Streaming Engine pass for loops with multiple accesses to the same tensor

    • Added nop() function to support the Reshape operator in TVM C runtime

    • Avoided overriding generic operator strategy in hls.py (back-ported from upstream TVM)

    • Added C7™ NPU strategy for concatenate instead of using the generic strategy

    • Disabled vectorization of inner loops with fewer iterations than the vector length

    • Disabled vectorization when the loop body contains a function call

    • Fixed vectorization factor to use the largest data type in the computation

    • Added 64-bit integer support in Streaming Engine configuration

    • Fixed TVM C runtime to support more than 255 functions/layers

    • Returned TVM runtime creation failures to the OpenVX node

    • Applied tiling and DMA schedule only to broadcast ops in injective.py

  • Added J721S2 device support

  • Merged with upstream neo-ai-tvm 1.11.2

  • Added tensor debug support for TVM Arm runtime and TVM C7™ NPU runtime

8.1.5. TIDL_PSDK_8.2

  • Initial release of C7™ NPU code generation support

  • Merged with neo-ai-tvm 1.10.0