Francesco Gatti
53 lines
1.8 KiB
C++
53 lines
1.8 KiB
C++
#include <iostream>
|
|
|
|
#include "Layer.h"
|
|
#include "kernels.h"
|
|
|
|
namespace tkDNN {
|
|
|
|
Activation::Activation(Network *net, dataDim_t input_dim, tkdnnActivationMode_t act_mode) :
|
|
Layer(net, input_dim) {
|
|
|
|
this->act_mode = act_mode;
|
|
checkCuda( cudaMalloc(&dstData, input_dim.tot()*sizeof(value_type)) );
|
|
|
|
checkCUDNN( cudnnSetTensor4dDescriptor(srcTensorDesc,
|
|
net->tensorFormat,
|
|
net->dataType,
|
|
input_dim.n*input_dim.l,
|
|
input_dim.c,
|
|
input_dim.h, input_dim.w) );
|
|
checkCUDNN( cudnnSetTensor4dDescriptor(dstTensorDesc,
|
|
net->tensorFormat,
|
|
net->dataType,
|
|
input_dim.n*input_dim.l,
|
|
input_dim.c,
|
|
input_dim.h, input_dim.w) );
|
|
}
|
|
|
|
Activation::~Activation() {
|
|
|
|
checkCuda( cudaFree(dstData) );
|
|
}
|
|
|
|
value_type* Activation::infer(dataDim_t &dim, value_type* srcData) {
|
|
|
|
if(act_mode == ACTIVATION_ELU) {
|
|
activationELUForward(srcData, dstData, dim.tot());
|
|
|
|
} else {
|
|
value_type alpha = value_type(1);
|
|
value_type beta = value_type(0);
|
|
checkCUDNN( cudnnActivationForward(net->cudnnHandle,
|
|
cudnnActivationMode_t(act_mode),
|
|
&alpha,
|
|
srcTensorDesc,
|
|
srcData,
|
|
&beta,
|
|
dstTensorDesc,
|
|
dstData) );
|
|
}
|
|
return dstData;
|
|
}
|
|
|
|
} |