theano error cannot convert 'cudnnConvolutionFwdAlgo_t*' to 'cudnnConvolutionFwdAlgoPerf_t

[behnam m]

My installed packages are:
CUDA 11.7, cudnn 8.5, theano 1.0.5

When I run the convolution network in the test file of Michael Nielsen's tutorial on Deep Learning:

    import network3
    from network3 import Network, ConvPoolLayer, FullyConnectedLayer, SoftmaxLayer
    training_data, validation_data, test_data = network3.load_data_shared()
    mini_batch_size = 10
    net = Network([
            ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),filter_shape=(20, 1, 5, 5), poolsize=(2, 2)),FullyConnectedLayer(n_in=20*12*12, n_out=100),SoftmaxLayer(n_in=100, n_out=10)], mini_batch_size)
    net.SGD(training_data, 60, mini_batch_size, 0.1,

                validation_data, test_data) 

   Problem occurred during compilation with the command line below:

dnn_fwd.c:326:60: error: invalid conversion from 'size_t {aka long unsigned int}' to 'int*' [-fpermissive]
    dnn_fwd.c:326:60: error: cannot convert 'cudnnConvolutionFwdAlgo_t*' to 'cudnnConvolutionFwdAlgoPerf_t* {aka cudnnConvolutionFwdAlgoPerfStruct*}' for argument '8' to 'cudnnStatus_t cudnnGetConvolutionForwardAlgorithm_v7(cudnnHandle_t, cudnnTensorDescriptor_t, cudnnFilterDescriptor_t, cudnnConvolutionDescriptor_t, cudnnTensorDescriptor_t, int, int*, cudnnConvolutionFwdAlgoPerf_t*)'

I understand that incompatibility of packages should be the reason, however I was wondering if there is any way to resolve this by modify the module or the theano's config file without downgrading the packages?

[Max Home]

I think I solved this issue. It popped up when I updated to CUDA11.      I had to patch up

three functions  dnn_fwd.c, dnn_gi.c, dnn_gw.c. I added the config option     gcc.cxxflags=-DCUDAVERSION=11, then added checks in the

code.  It's can use cleaning up.  If someone wants, they can take this fix and check it in to the repo.

cheers, Paul

In dnn_fwd.c, I put in this check:

        #if CUDAVERSION  > 10
           // PMB
           #define CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT 2
           int ret_alg_cnt;
           cudnnConvolutionFwdAlgoPerf_t ta;

           err = cudnnGetConvolutionForwardAlgorithm_v7(
             params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(kerns),
             desc, APPLY_SPECIFIC(output),
             CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, &ret_alg_cnt, &ta);
             algo=ta.algo;
        #else
           err = cudnnGetConvolutionForwardAlgorithm(
             params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(kerns),
             desc, APPLY_SPECIFIC(output),
             CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT, maxfree, &algo);
        #endif

In dnn_gi.c, I put in this check:

#if CUDAVERSION > 10

           int ret_alg_cnt;
           cudnnConvolutionBwdDataAlgoPerf_t at;
           #define CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT 2
           err = cudnnGetConvolutionBackwardDataAlgorithm_v7(
             params->handle, APPLY_SPECIFIC(kerns), APPLY_SPECIFIC(output),
             desc, APPLY_SPECIFIC(input),
             CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,&ret_alg_cnt, &at);
             algo=at.algo;
        #else
        err = cudnnGetConvolutionBackwardDataAlgorithm(
          params->handle, APPLY_SPECIFIC(kerns), APPLY_SPECIFIC(output),
          desc, APPLY_SPECIFIC(input),
          CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT, maxfree, &algo);

        #endif

In dnn_gw.c, I put in this check:

       #if CUDAVERSION > 10
            // PMB
            #define CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT 2
            cudnnConvolutionBwdFilterAlgoPerf_t ta;
            int alg_cnt;
            err = cudnnGetConvolutionBackwardFilterAlgorithm_v7(
              params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(output),
              desc, APPLY_SPECIFIC(kerns),
              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, &alg_cnt, &ta);
              algo=ta.algo;
        #else
            err = cudnnGetConvolutionBackwardFilterAlgorithm(
              params->handle, APPLY_SPECIFIC(input), APPLY_SPECIFIC(output),
              desc, APPLY_SPECIFIC(kerns),
              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT, maxfree, &algo);

        #endif