Hello,
I have a Jetson Xavier AGX with Jetpack 4.5 with TensorRT 7.1.3.0.
I have a convolution that wants to output nhwc followed by an activation that doesn’t care about dimension layout.
I have a custom plugin for my activation with the following configurePlugin:
bool supportsFormatCombination(int32_t pos, const nvinfer1::PluginTensorDesc* inOut, int32_t nbInputs, int32_t nbOutputs) const
{
const bool valid_dtype = (inOut[pos].type == nvinfer1::DataType::kHALF);
const bool valid_format = (inOut[pos].format == nvinfer1::TensorFormat::kHWC8) || (inOut[pos].format == nvinfer1::TensorFormat::kLINEAR);
return valid_dtype && valid_format;
}
I have a convolution prior to this operation that wants to use:
trt_volta_h884cudnn_128x128_ldg8_relu_exp_small_nhwc_tn_v1, if I allow kHWC8, then I get the following error during conversion:
[TensorRT] ERROR: ../rtSafe/cuda/cudaReformatRunner.cpp (251) - Assertion Error in combineDHtoH: 0 (nbSpatialDims == 3 && nbDims >= 5)
If I set my format to kLINEAR then everything works correctly, but I have a reformat operation before and after each activation. If I accept any input format, then I get the same error as above.
I have upgraded my jetpack 4.5 xavier to TensorRT 8 by extracting the debs from a Jetpack 4.6 install and apt installing them, this appears to work at first and resolves the previously mentioned error, but the model no longer gives usable output. From this post: How to create a plugin support input format::kHWC8? - #5 by pznucleus it would appear that upgrading to Tensorrt 7.2.2.3 should be sufficient, however I can’t find a build of Tensorrt 7.2.2.3 that works on Jetpack 4.5. My next attempt will be to try a Jetpack 4.6 ngc image to see if it is backwards compatible.