Memory leak in strided, valid-padded convolution with dynamic input shape #428
Description
Description
GPU memory footprint of program keeps growing when running a strided, valid-padded convolution with dynamic input shape.
Environment
TensorRT Version: 7.0.0.11
GPU Type: Titan X
Nvidia Driver Version: 436.30
CUDA Version: 10.0
CUDNN Version: 7.6.3.30
Operating System + Version: Windows 10 1903
Python Version (if applicable): 3.7.6
TensorFlow Version (if applicable): 1.14.0
PyTorch Version (if applicable): N/A
Baremetal or Container (if container which image + tag): N/A
Relevant Files
Steps To Reproduce
- Create a TF model
import tensorflow as tf
input = tf.placeholder(tf.float32, [1, None, None, 1], name='x')
layer = tf.nn.conv2d(input, tf.zeros((2, 2, 1, 1)), (1, 2, 2, 1), 'VALID')
output = tf.reduce_sum(layer, axis=[1, 2, 3], name='y')
with tf.Session() as sess:
g = tf.get_default_graph().as_graph_def()
tf.train.write_graph(g, ".", "leak.pb", as_text=False)- Convert to ONNX
python -m tf2onnx.convert --input leak.pb --output leak.onnx --inputs x:0 --outputs y:0 --opset 11
- Build engine
trtexec --onnx=leak.onnx --saveEngine=leak.trt --explicitBatch --minShapes='x:0':1x2x2x1 --optShapes='x:0':1x128x128x1 --maxShapes='x:0':1x256x256x1
- Run inference
# include <array>
# include <fstream>
# include <random>
# include <cuda_runtime_api.h>
# include <NvInfer.h>
# include <samples/common/buffers.h>
class MyLogger : public nvinfer1::ILogger
{
void log(Severity severity, char const * msg) override
{
std::cout << msg << std::endl;
}
} logger;
int main()
{
std::ifstream input("leak.trt", std::ios::binary);
std::vector<unsigned char> model(std::istreambuf_iterator<char>(input), {});
auto runtime = nvinfer1::createInferRuntime(logger);
auto engine = runtime->deserializeCudaEngine(model.data(), model.size());
auto context = engine->createExecutionContext();
auto x_idx = engine->getBindingIndex("x:0");
auto y_idx = engine->getBindingIndex("y:0");
samplesCommon::ManagedBuffer x{ nvinfer1::DataType::kFLOAT, nvinfer1::DataType::kFLOAT };
samplesCommon::HostBuffer y_host(1, nvinfer1::DataType::kFLOAT);
samplesCommon::DeviceBuffer y_device(1, nvinfer1::DataType::kFLOAT);
std:random_device rd;
std::mt19937 gen(rd());
std::uniform_int_distribution<> dis(1, 30);
std::array<void *, 2> buffer;
while (true)
{
auto x_shape = nvinfer1::Dims4(1, dis(gen) * 2, dis(gen) * 2, 1);
x.hostBuffer.resize(x_shape);
std::fill_n(static_cast<float*>(x.hostBuffer.data()), samplesCommon::volume(x_shape), 0.f);
x.deviceBuffer.resize(x_shape);
cudaMemcpy(x.deviceBuffer.data(), x.hostBuffer.data(), x.hostBuffer.nbBytes(), cudaMemcpyHostToDevice);
context->setBindingDimensions(x_idx, x_shape);
buffer[x_idx] = x.deviceBuffer.data();
buffer[y_idx] = y_device.data();
context->executeV2(buffer.data());
}
return 0;
}This results in a slow but steady increase of GPU memory.
Note that I don't have a leak no more
- if the convolution padding is changed to
'SAME', - or if the stride is changed to
(1, 1, 1, 1).
For the reasons above, it is possible that this issue is different from #351 , which has been fixed, since #351 is demonstrated on a non-strided, padded convolution.