9. Release Notes

9.1. TIDL_PSDK_8.6

  • New devices

    • Support AM62A - To compile model for AM62A, set platform argument to “AM62A”, see Compiling Models for details.

  • Bug fixes in C7x code generation

    • Fix DMAPass src on/off chip check

    • Fix scatter_nd code generation

9.2. TIDL_PSDK_8.5

  • Bug fixes in C7x code generation

    • Support non-contiguous access pattern in DMA

    • Support strided access pattern in SE

    • Support odd number of DMA blocks and iterations

    • Simplify resize2d index computation

    • Optimize maxpool2d with 1x1 pool size

  • TIDL offload

    • Rewrite broadcast 1D “add” as “bias_add” and offload to TIDL

    • Do not offload resize2d with asymmetric or non-power-of-2 scaling factor

    • Enabled TIDL batch processing

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

    • Offload sigmoid to TIDL

  • TVM extension

    • Support calling external function with DLTensor args

    • Demonstrate overwriting default TVM strategy for an operator

    • Demonstrate overwriting default TVM strategy for an operator for a special case

9.3. TIDL_PSDK_8.4

  • Bug fixes in C7x code generation

    • Fix streaming engine pass for loops containing multiple accesses to same tensor

    • Add nop() function to support Reshape op in TVM C runtime

    • Avoid overriding generic op strategy in “hls.py” (back-ported from upstream TVM)

    • Add C7x strategy for concatenate instead of using generic strategy

    • Do not vectorize inner loop with iterations less than vector length

    • Do not vectorize if loop body contains call

    • Fix vectorization factor to use largest data type in computation

    • Add 64-bit integer support in streaming engine config

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

    • Return tvm runtime create failure to OpenVX node

    • Apply tiling and dma schedule only to broadcast ops in injective.py

  • J721S2 support

  • Merge with upstream neo-ai tvm 1.11.2

  • Tensor debug support for TVM Arm runtime and TVM C7x runtime

9.4. TIDL_PSDK_8.2

  • Initial release of C7x code generation support

  • Merge with neo-ai-tvm 1.10.0