fp16 implementation, TODO deallocate in LayerWgs

include/Layer.h

@@ -80,9 +80,19 @@ public:
 
     //batchnorm
     bool batchnorm;
+    dnnType *power_h;
     dnnType *scales_h,   *scales_d;
     dnnType *mean_h,     *mean_d;
     dnnType *variance_h, *variance_d;
+
+    //fp16
+    __half *data16_h, *bias16_h;
+    __half *data16_d, *bias16_d;
+
+    __half *power16_h,    *power16_d;
+    __half *scales16_h,   *scales16_d;
+    __half *mean16_h,     *mean16_d;
+    __half *variance16_h, *variance16_d;
 };
 
 

include/Network.h

@@ -58,7 +58,9 @@ public:
 
     dataDim_t input_dim;
     dataDim_t getOutputDim();
+
+    bool fp16;
 };
 
 }
-#endif //NETWORK_H
+#endif //NETWORK_H

include/kernels.h

@@ -12,4 +12,6 @@ void reorgForward(  dnnType* srcData, dnnType* dstData,
 void softmaxForward(float *input, int n, int batch, int batch_offset, 
                     int groups, int group_offset, int stride, float temp, float *output, cudaStream_t stream = cudaStream_t(0));
 
+
+void float2half(float* srcData, __half* dstData, int size, const cudaStream_t stream = cudaStream_t(0));
 #endif //KERNELS_H