0

I use cudnn acceleration in my caffe program. I use cudnn 4 at the begin and it's working fine but when I updated the cudnn to version 5.0, the pow function doesn't work. The calling function is in the batch_norm layer as 

caffe_gpu_powx(variance_.count(), variance_.gpu_data(), Dtype(0.5), variance_.mutable_gpu_data());

And the data after calling doesn't change. The pow function is defined as below, as same as in the caffe github banch

template <typename Dtype>
\__global__ void powx_kernel(const int n, const Dtype* a,
    const Dtype alpha, Dtype* y)
 {
     CUDA_KERNEL_LOOP(index, n)
     {           
         y[index] = pow(a[index], alpha);  
     }  
}

template <>
void caffe_gpu_powx<float>(const int N, const float* a,
    const float alpha, float* y) {
    // NOLINT_NEXT_LINE(whitespace/operators)
    powx_kernel<float><<<CAFFE_GET_BLOCKS(N), CAFFE_CUDA_NUM_THREADS>>>(
      N, a, alpha, y);
}
Shai
  • 111,146
  • 38
  • 238
  • 371
excelgco
  • 11
  • 1

1 Answers1

1

I made a mistake that I have set the code generation to "compute_52,sm_52" at the begin with TITAN X, but now a lower GPU should be set to "compute_20,sm_20". And it's working fine now.

excelgco
  • 11
  • 1
  • That setting doesn't make sense. CUDNN requires a compute capability >=3.0 device to run. What device are you actually trying to run on? – talonmies Nov 10 '16 at 11:28