cleanup

array.cu

@@ -70,7 +70,7 @@ int myCreateCUDAArray(uint8_t *tf_ptr){
 
     // check if the GPU did the same as the CPU
     bool workedCorrectly = true;
-    printf("CUDA kernel incrementing test:\n");
+    printf("CUDA kernel simple incrementing test:\n");
     for(int i=0;i<numberOfNumbers;i++)
     {
         if (numbers1[i] != numbers2[i])

main.cpp

@@ -32,27 +32,19 @@ using tensorflow::Status;
 using tensorflow::Tensor;
 
 
-Status loadGraph(unique_ptr<tensorflow::Session> *session){
+Status createGraphAndSession(unique_ptr<tensorflow::Session> *session){
 
     tensorflow::GraphDef graph_def;
 
     using namespace tensorflow;
-    using namespace tensorflow::ops;
 
     auto scope = Scope::NewRootScope();
 
     // TF likes power of 2
     tensorflow::TensorShape shape = tensorflow::TensorShape({256});
 
-    auto a  = Placeholder(scope.WithOpName("array_tensor_in"), DT_UINT8, Placeholder::Shape(shape));
-
-    //auto c0 = Const(scope,{256});
-    //auto c1 = Const(scope, (uint8_t)1, {});
-    //auto c2  = Fill(scope, c0, c1);
-
-    auto b  = Identity(scope.WithOpName("array_tensor_out"), a);
-    //auto b  = Identity(scope.WithOpName("array_tensor_out"), c2);
-    //auto b = Add(scope.WithOpName("array_tensor_out"),c2,a);
+    auto a  = ops::Placeholder(scope.WithOpName("array_tensor_in"), DT_UINT8, ops::Placeholder::Shape(shape));
+    auto b  = ops::Identity(scope.WithOpName("array_tensor_out"), a);
 
     TF_CHECK_OK(scope.ToGraphDef(&graph_def));
 
@@ -65,7 +57,7 @@ Status loadGraph(unique_ptr<tensorflow::Session> *session){
     session->reset(tensorflow::NewSession(session_options));
     Status session_create_status = (*session)->Create(graph_def);
     if (!session_create_status.ok()){
-        LOG(ERROR) << "loadGraph(): ERROR" << session_create_status;
+        LOG(ERROR) << "createGraphAndSession(): ERROR" << session_create_status;
     }
     return Status::OK();
 }
@@ -109,7 +101,7 @@ int main(int, char**) {
         LOG(INFO) << "Tensorflow built with CUDA, keep running" << endl;
     }
 
-    // check and init CUDA drivers and libs
+    // check and init CUDA drivers and libs?
     void *hHandleDriver = nullptr;
     CUresult cuda_res = cuInit(0, __CUDA_API_VERSION, hHandleDriver);
     if (cuda_res != CUDA_SUCCESS)
@@ -123,14 +115,17 @@ int main(int, char**) {
     }
     LOG(INFO) << "\033[1;32m" << "CUDA init: ok" << "\033[0m";
 
+
+    // construct graph and create TF session
     std::unique_ptr<tensorflow::Session> session;
+    createGraphAndSession(&session);
 
-    loadGraph(&session);
 
+    // do the opts and allocate tensor in GPU
+    // NOTE: must match with graph names
     const string inputLayer  = "array_tensor_in:0";
     const string outputLayer = "array_tensor_out:0";
 
-    // do the opts
     CallableOptions opts;
     Session::CallableHandle feed_gpu_fetch_cpu;
 
@@ -150,7 +145,6 @@ int main(int, char**) {
 
     // TF likes power of 2 and 256s
     tensorflow::TensorShape shape = tensorflow::TensorShape({256});
-    // allocate tensor on the GPU
     tensorflow::PlatformGpuId platform_gpu_id(0);
 
     tensorflow::GPUMemAllocator *sub_allocator =
@@ -165,32 +159,42 @@ int main(int, char**) {
 
     LOG(INFO) << "\033[1;37m" << "Is CUDA Tensor? " << (IsCUDATensor(inputTensor)?"\033[1;32myes":"\033[1;31mno") << "\033[0m";
 
+    // pointer to tensor data
     tensorflow::uint8 *p = inputTensor.flat<tensorflow::uint8>().data();
 
+    // run and check
+    vector<Tensor> outputs;
+
+    for(int i=0;i<2;i++){
+
+        cout << endl;
+
+        if (i==0){
+            LOG(INFO) << "\033[1;32m" << "RunCallable()... No feeding (zeroes)" << "\033[0m";
+        }
+        if (i==1){
+            LOG(INFO) << "\033[1;32m" << "RunCallable()... Feeding from CUDA kernel"<< "\033[0m";
             // CUDA kernel call
+            // NOTE: do not allocate memory for p inside the kernel
             myCreateCUDAArray(p);
-
-    vector<Tensor> outputs;
-    LOG(INFO) << "RunCallable()...";
+        }
 
         runStatus = session->RunCallable(feed_gpu_fetch_cpu, {inputTensor}, &outputs, nullptr);
-    if (!runStatus.ok())
-    {
+        if (!runStatus.ok()){
             LOG(ERROR) << "Running model failed: " << runStatus;
             return -1;
         }
-
         LOG(INFO) << "RunCallable() output:";
-
         LOG(INFO) << outputs[0].DebugString();
         auto tmap = outputs[0].tensor<uint8_t, 1>();
 
         cout << "\033[1;37m";
-    for (int d = 0; d < 256; d++) {
+        for (int d = 0; d < 256; d++){
             cout << (int) tmap(d);
             if (d!=255) cout << ", ";
         }
         cout << "\033[0m" << endl;
+    }
 
     session->ReleaseCallable(feed_gpu_fetch_cpu);