Fixed RNN (LSTM, RNN, CRNN, GRU) for CUDNN_HALF=1

src/crnn_layer.c

@@ -215,6 +215,16 @@ void forward_crnn_layer_gpu(layer l, network_state state)
     layer input_layer = *(l.input_layer);
     layer self_layer = *(l.self_layer);
     layer output_layer = *(l.output_layer);
+    /*
+#ifdef CUDNN_HALF
+// slow and bad
+    s.index = state.index;
+    s.net = state.net;
+    cuda_convert_f32_to_f16(input_layer.weights_gpu, input_layer.c*input_layer.n*input_layer.size*input_layer.size, input_layer.weights_gpu16);
+    cuda_convert_f32_to_f16(self_layer.weights_gpu, self_layer.c*self_layer.n*self_layer.size*self_layer.size, self_layer.weights_gpu16);
+    cuda_convert_f32_to_f16(output_layer.weights_gpu, output_layer.c*output_layer.n*output_layer.size*output_layer.size, output_layer.weights_gpu16);
+#endif  //CUDNN_HALF
+    */
 
     fill_ongpu(l.outputs * l.batch * l.steps, 0, output_layer.delta_gpu, 1);
     fill_ongpu(l.hidden * l.batch * l.steps, 0, self_layer.delta_gpu, 1);

src/network_kernels.cu

@@ -1,6 +1,7 @@
-#include "cuda_runtime.h"
+//#include "cuda_runtime.h"
-#include "curand.h"
+//#include "curand.h"
-#include "cublas_v2.h"
+//#include "cublas_v2.h"
+#include "cuda.h"
 
 extern "C" {
 #include <stdio.h>
@@ -61,7 +62,7 @@ void forward_network_gpu(network net, network_state state)
         //printf("\n layer %d - type: %d - \n", i, l.type);
         //start_timer();
         l.forward_gpu(l, state);
-        //cudaDeviceSynchronize();
+        //CHECK_CUDA(cudaDeviceSynchronize());
         //stop_timer_and_show();
 
         if(net.wait_stream)
@@ -152,7 +153,7 @@ void forward_backward_network_gpu(network net, float *x, float *y)
     int i;
     for (i = 0; i < net.n; ++i) {
         layer l = net.layers[i];
-        if (l.weights_gpu) {
+        if (l.weights_gpu && l.weights_gpu16) {
             assert((l.c*l.n*l.size*l.size) > 0);
             cuda_convert_f32_to_f16(l.weights_gpu, l.c*l.n*l.size*l.size, l.weights_gpu16);
         }