MulAdd implemented

src/MulAdd.cpp

@@ -0,0 +1,45 @@
+#include <iostream>
+
+#include "Layer.h"
+#include "kernels.h"
+
+namespace tkDNN {
+
+MulAdd::MulAdd(Network *net, dataDim_t input_dim, value_type mul, value_type add) : 
+    Layer(net, input_dim) {
+
+    this->mul = mul;
+    this->add = add;
+
+    int size = input_dim.tot();
+
+    // create a vector with all value setted to add 
+    value_type *add_vector_h = new value_type[size];
+    for(int i=0; i<size; i++)
+        add_vector_h[i] = add;
+
+    checkCuda( cudaMalloc(&add_vector, size*sizeof(value_type)));
+    checkCuda( cudaMemcpy(add_vector, add_vector_h, size*sizeof(value_type), cudaMemcpyHostToDevice));
+    delete [] add_vector_h;
+
+
+    checkCuda( cudaMalloc(&dstData, input_dim.tot()*sizeof(value_type)) );
+}
+
+MulAdd::~MulAdd() {
+
+    checkCuda( cudaFree(add_vector) );
+    checkCuda( cudaFree(dstData) );
+}
+
+value_type* MulAdd::infer(dataDim_t &dim, value_type* srcData) {
+
+    matrixMulAdd(net->cublasHandle, srcData, dstData, add_vector, input_dim.tot(), mul);
+    
+    //update data dimensions    
+    dim = output_dim;
+
+    return dstData;
+}
+
+}

src/utils.cpp

@@ -53,4 +53,14 @@ void matrixTranspose(cublasHandle_t handle, value_type* srcData, value_type* dst
     float const alpha(1.0);
     float const beta(0.0);
     checkERROR( cublasSgeam( handle, CUBLAS_OP_T, CUBLAS_OP_N, m, n, &alpha, A, n, &beta, A, m, clone, m ));
+}
+
+void matrixMulAdd(  cublasHandle_t handle, value_type* srcData, value_type* dstData, 
+                    value_type* add_vector, int dim, value_type mul) {
+
+    checkCuda( cudaMemcpy(dstData, add_vector, dim*sizeof(value_type), cudaMemcpyDeviceToDevice));
+        
+    value_type alpha = mul;
+    checkERROR( cublasSaxpy(handle, dim, &alpha, srcData, 1, dstData, 1));
+
 }