diff --git a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu index fc5fa87283292783cf34bc91fba329223af86e83..c65fd8b0163b80459476b4ab3c61ce48580cc0ff 100644 --- a/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu +++ b/llvm/projects/hpvm-tensor-rt/tensor_runtime/include/tensor_utils.cu @@ -63,6 +63,20 @@ struct Tensor* getRuntimeTensor(struct Tensor* hpvm_tensor){ */ +/*void freeCudaPtr(void* cuda_ptr){ + + int i; + for(i = 0; i < tensors_ptr.size(); i++){ + if(tensors_ptr[i] == cuda_ptr) + break; + } + + tensors_ptr.erase(tensors_ptr.begin()+i); + cudaFree(cuda_ptr); +} +*/ + + void freeTensor(void* tensor_ptr){ Tensor* tensor = (Tensor*) tensor_ptr; @@ -536,9 +550,10 @@ void convertToFP32_offline(struct Tensor* tensor){ h2f((half*) tensor->gpu_half_data, tensor->num_elems, (float*) tensor->gpu_data); tensor->cur_type = float_type; - - cudaFree(tensor->gpu_half_data); - tensor->gpu_half_data = NULL; + + //freeCudaPtr(tensor->gpu_half_data); + //cudaFree(tensor->gpu_half_data); + //tensor->gpu_half_data = NULL; }