Add create gpu buffer helper function

src/ngraph/runtime/gpu/gpu_external_function.cpp

@@ -553,7 +553,8 @@ void runtime::gpu::GPU_ExternalFunction::compile()
             size_t temp_pool_size = current_function->get_temporary_pool_size();
             writer << "// Allocate the memory pool\n";
             // TODO memory pool malloc.
-            writer << "void** pool_base_ptr;\n";
+            writer << "void** pool_base_ptr = runtime::gpu::create_gpu_buffer(" << temp_pool_size
+                   << ");\n";
 
             // Add temporaries to the variable name map
             for (shared_ptr<Node> node : current_function->get_ordered_ops())

src/ngraph/runtime/gpu/gpu_util.cpp

@@ -45,6 +45,13 @@ void runtime::gpu::check_cuda_errors(CUresult err)
     assert(err == CUDA_SUCCESS);
 }
 
+void** runtime::gpu::create_gpu_buffer(size_t buffer_size)
+{
+    void** allocated_buffer_pool;
+    cudaMalloc(&allocated_buffer_pool, buffer_size);
+    return allocated_buffer_pool;
+}
+
 void runtime::gpu::cuda_memcpyDtD(void* d, void* s, size_t element_count, size_t element_size)
 {
     size_t size_in_bytes = element_size * element_count;

src/ngraph/runtime/gpu/gpu_util.hpp

@@ -22,6 +22,7 @@ namespace ngraph
         {
             void print_gpu_f32_tensor(void* p, size_t element_count, size_t element_size);
             void check_cuda_errors(CUresult err);
+            void** create_gpu_buffer(size_t buffer_size);
             void cuda_memcpyDtD(void* d, void* s, size_t element_count, size_t element_size);
             void cuda_memcpyHtD(void* d, void* s, size_t buffer_size);
         }