better rt inference

include/NetworkRT.h

@@ -32,6 +32,7 @@ public:
         Do inferece
     */
     dnnType* infer(dataDim_t &dim, dnnType* data);
+    void enqueue();    
 
     nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Layer *l);
     nvinfer1::ILayer* convert_layer(nvinfer1::ITensor *input, Conv2d *l);
@@ -62,4 +63,4 @@ template<typename T> T readBUF(const char*& buffer)
 }
 
 }
-#endif //NETWORKRT_H
+#endif //NETWORKRT_H

src/NetworkRT.cpp

@@ -139,6 +139,10 @@ dnnType* NetworkRT::infer(dataDim_t &dim, dnnType* data) {
     return output;
 }
 
+void NetworkRT::enqueue() {
+    contextRT->enqueue(1, buffersRT, stream, nullptr);
+}
+
 ILayer* NetworkRT::convert_layer(ITensor *input, Layer *l) {
 
     layerType_t type = l->getLayerType();

tests/test_rtinference/rtinference.cpp

@@ -1,24 +1,33 @@
 #include<iostream>
 #include "tkdnn.h"
+#include <stdlib.h>     /* srand, rand */
 
 int main(int argc, char *argv[]) {
 
     if(argc < 2 || !fileExist(argv[1]))
         FatalError("unable to read serialRT file");
 
+    //always same test 
+    srand (0); 
+
     //convert network to tensorRT
     tkDNN::NetworkRT netRT(NULL, argv[1]);
 
-    tkDNN::dataDim_t dim = netRT.input_dim;
+    dnnType *input = new float[netRT.input_dim.tot()];
-    dnnType *data;
+    dnnType *output = new float[netRT.input_dim.tot()];
-    checkCuda(cudaMalloc(&data, dim.tot()*sizeof(dnnType)));
 
-    printCenteredTitle(" TENSORRT inference ", '=', 30); {
+    printCenteredTitle(" TENSORRT inference ", '=', 30); 
-        dim.print();
+    for(int i=0; i<100; i++) {
+        for(int j=0; j<netRT.input_dim.tot(); j++)
+            input[j] = ((float) rand() / (RAND_MAX));
         TIMER_START
-        data = netRT.infer(dim, data);
+        checkCuda(  cudaMemcpyAsync(netRT.buffersRT[netRT.buf_input_idx], input, 
+                    netRT.input_dim.tot()*sizeof(float), cudaMemcpyHostToDevice, netRT.stream));
+        netRT.enqueue();
+        checkCuda(  cudaMemcpyAsync(output, netRT.buffersRT[netRT.buf_output_idx], 
+                                    netRT.output_dim.tot()*sizeof(float), cudaMemcpyDeviceToHost, netRT.stream));
+        cudaStreamSynchronize(netRT.stream);
         TIMER_STOP
-        dim.print();
     }
     
     return 0;