#         "fallback_tensor_conversion_host.mlir": ["nomsan"],  # Can't instrument code in precompiled lib (cuDNN)
#         "kernel_fallback_op_handler.mlir": ["nomsan"],  # Can't instrument code in precompiled lib (cuDNN)
#         "mnist.mlir": ["nomsan"],  # Can't instrument code in precompiled lib (cuDNN)
#         "runtime_fallback_op_handler.mlir": ["nomsan"],  # Can't instrument code in precompiled lib (cuDNN)

# CuDNN: https://docs.nvidia.com/deeplearning/sdk/cudnn-install/index.html#ubuntu-network-installation
libcudnn9-dev-cuda-12=9.1.1.17-1
libcudnn9-cuda-12=9.1.1.17-1

# This can be removed once NVIDIA publishes a cuda-12.3.2 Docker image.
# For now it ensures that we install at least version 12.3.107 of PTXAS,
# since 12.3.103 has a bug.

       * Try adding the CUDA, CUPTI, and cuDNN installation directories to the %PATH% environment variable.
       * Refer [windows setup guide](https://www.tensorflow.org/install/gpu#windows_setup).
     * If you have above configuration and using _**Ubuntu/Linux**_ platform -
       * Try adding the CUDA, CUPTI, and cuDNN installation directories to the $LD_LIBRARY_PATH environment variable.

} {

// CHECK-LABEL: func @transposeConv2D_3x3_f16
func.func @transposeConv2D_3x3_f16(%input: tensor<1x28x28x64xf16>, %filter: tensor<3x3x64x64xf16>) -> tensor<1x26x26x64xf16> {
  // cuDNN prefers NCHW data format for spatial convolutions in f16 before
  // compute capability 7.0 (NVIDIA Tensor Cores).

  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"

  - type: input
    id: Compiler
    attributes:
      label: GCC/compiler version
      description: If compiling from source
  - type: input
    id: Cuda
    attributes:
      label: CUDA/cuDNN version
  - type: input
    id: Gpu
    attributes:
      label: GPU model and memory
      description: If compiling from source
  - type: textarea
    id: what-happened
    attributes:

-   type: input id: Compiler attributes: label: GCC/Compiler version
    description: if compiling from source placeholder: validations: required:
    false

-   type: input id: Cuda attributes: label: CUDA/cuDNN version description:
    placeholder: validations: required: false

-   type: input id: Gpu attributes: label: GPU model and memory description: if
    compiling from source placeholder: validations: required: false

} {

// CHECK-LABEL: func @transposeConv2D_3x3_f32
func.func @transposeConv2D_3x3_f32(%input: tensor<1x28x28x64xf32>, %filter: tensor<3x3x64x64xf32>) -> tensor<1x26x26x64xf32> {
  // cuDNN prefers NCHW data format for spatial convolutions.
  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NHWC",

  environ_cp['TF_CUDA_VERSION'] = tf_cuda_version


def set_tf_cudnn_version(environ_cp):
  """Set TF_CUDNN_VERSION."""
  ask_cudnn_version = (
      'Please specify the cuDNN version you want to use. '
      '[Leave empty to default to cuDNN %s]: ') % _DEFAULT_CUDNN_VERSION
  tf_cudnn_version = get_from_env_or_user_or_default(environ_cp,
                                                     'TF_CUDNN_VERSION',

    for more details.

## Deprecations

*   TensorFlow 1.2 may be the last time we build with cuDNN 5.1. Starting with
    TensorFlow 1.3, we will try to build all our prebuilt binaries with cuDNN
    6.0. While we will try to keep our source code compatible with cuDNN 5.1, it
    will be best effort.

## Breaking Changes to the API

  // b/111271662: MatrixSolve performance issues.
  // https://github.com/tensorflow/tensorflow/pull/31012:
  //    ResizeNearestNeighbor, ResizeBilinear, and ResizeBilinearGrad sometimes
  //    create convolutions too large for CuDNN to handle.
  return node.type_string() == "SelfAdjointEigV2" ||
         node.type_string() == "Svd" || node.type_string() == "Qr" ||
         node.type_string() == "MatrixInverse" ||