cuDNN API的使用与测试-以二维卷积+Relu激活函数为例

一、cudnn介绍

NVIDIA CUDA深度神经网络库（cuDNN）是适用于英伟达GPU的用于深度神经网络的cuda函数库。cuDNN为神经网络中常见的算子提供了高度优化的实现，例如卷积，池化，归一化层和激活层的前向传播和反向传播的实现。 全球的深度学习研究人员和框架开发人员都依赖cuDNN来实现高性能GPU加速。它使他们可以专注于训练神经网络和开发软件应用程序，而不必花时间在底层GPU性能调整上。cuDNN广泛应用于深度学习框架上，包括Caffe2，Chainer，Keras，MATLAB，MxNet，PyTorch和TensorFlow。

原版的cuda安装包是不包含cudnn的，需要用户额外下载和安装。cudnn的安装过程不再文本内容的范围以内，还请读者仔细搜索下载安装。本文的例子是基于cudnn8.0版本，demo地址为： https://github.com/thb1314/cudnn_conv_relu 更多学习资料请参考: https://docs.nvidia.com/deeplearning/cudnn/

二、cudnn API的使用步骤

使用 cuDNN 的应用程序必须通过调用 cudnnCreate() 来初始化库上下文的句柄。这个句柄被显式地传递给对 GPU 数据进行操作的每个库函数。一旦应用程序完成使用 cuDNN，它可以使用 cudnnDestroy() 释放与库句柄关联的资源。这种方法允许用户在使用多个host主机线程（比如cpu线程）、GPU 或者 CUDA stream时显式控制库的功能。 cudnn API使用步骤一般如下：

1. 创建cuDNN句柄

cudnnStatus_t cudnnCreate(cudnnHandle_t *handle)

1. 以Host方式调用在Device上运行的函数

比如卷积运算： cudnnConvolutionForward等

1. 释放cuDNN句柄

cudnnStatus_t cudnnDestroy(cudnnHandle_t handle) 其他函数 比如在创建cudnn的时候可以设置或者互殴cudnn所用的stream

• cudnnStatus_t cudnnSetStream( cudnnHandle_t handle, cudaStream_t streamId)
• cudnnStatus_t cudnnGetStream( cudnnHandle_t handle, cudaStream_t *streamId)

三、二维卷积的实现案例

本小节介绍如何使用cuDNN的卷积相关api来实现二维卷积运算，本文还是会使用在cublas章节采用的一个张量管理的封装类。具体步骤如下：

1. 创建卷积所需要的输入Tensor，卷积权重和卷积偏执项。
2. 创建cudnn handle并且设置handle的stream
3. 设置卷积输入相关Tensor的形状描述符（使用TensorDescriptor系列API）和卷积配置相关的描述符
4. 根据当前配合搜索卷积计算方法并申请相关workspace
5. 调用cudnnConvolutionBiasActivationForward完成卷积计算
6. 释放workspace显存占用，销毁cudnn 句柄

核心代码如下:

// 123
int test_conv_relu() {
    ::srand(::time(0));
    std::cout << "CUDNN_VERSION:" << CUDNN_VERSION << std::endl;
    // 设定输入输出tensor的维度参数
    constexpr int batch_size = 4;
    constexpr int channel_in = 3;
    constexpr int height_in = 112;
    constexpr int width_in = 112;
    constexpr int channel_out = 15;
    constexpr int height_out = 112;
    constexpr int width_out = 112;
    constexpr int kernel_h = 1;
    constexpr int kernel_w = 1;
    // 构造相关Tensor
    // input
    TRT::Tensor q_tensor(std::vector<int>{batch_size, channel_in, height_in, width_in});
    // kernel input
    TRT::Tensor kernel_tensor(std::vector<int>{channel_out, channel_in, kernel_h, kernel_w});
    // bias
    TRT::Tensor bias_tensor(std::vector<int>{channel_out});
    TRT::Tensor z_tensor(std::vector<int>{batch_size, channel_out, height_out, width_out});
    // output
    TRT::Tensor out_tensor(std::vector<int>{batch_size, channel_out, height_out, width_out});
    auto qptr_cpu = q_tensor.cpu<float>();
    for(int i = 0; i < q_tensor.numel(); ++i)
    {
        qptr_cpu[i] = float(rand() % 100000) / 100000;
    }
    q_tensor.save_to_file("q_tensor.npz");
    auto biasptr_cpu = bias_tensor.cpu<float>();
    for(int i = 0; i < bias_tensor.numel(); ++i)
    {
        biasptr_cpu[i] = float(rand() % 100000) / 100000;
    }
    bias_tensor.save_to_file("bias_tensor.npz");
    auto kernelptr_cpu = kernel_tensor.cpu<float>();
    for(int i = 0; i < kernel_tensor.numel(); ++i)
    {
        kernelptr_cpu[i] = float(rand() % 100000) / 100000;
    }
    kernel_tensor.save_to_file("kernel_tensor.npz");
    auto qptr_gpu = q_tensor.to_gpu(true).gpu<float>();
    auto bias_gpu = bias_tensor.to_gpu(true).gpu<float>();
    auto kernel_gpu = kernel_tensor.to_gpu(true).gpu<float>();
    auto outptr_gpu = out_tensor.to_gpu().gpu<float>();
    cudaStream_t stream = out_tensor.get_stream();
    // 创建cudnn句柄并设置handle的stream
    cudnnHandle_t cudnn;
    checkCUDNN(cudnnCreate(&cudnn));
    checkCUDNN(cudnnSetStream(cudnn, stream));
    // y = act ( alpha1 * conv(x) + alpha2 * z + bias )
    const float alpha1 = 1;
    const float alpha2 = 0;
    // 设置输入Tensor描述符
    cudnnTensorDescriptor_t input_descriptor;
    checkCUDNN(cudnnCreateTensorDescriptor(&input_descriptor));
    checkCUDNN(cudnnSetTensor4dDescriptor(input_descriptor,
                                          /*format=*/CUDNN_TENSOR_NCHW,
                                          /*dataType=*/CUDNN_DATA_FLOAT,
                                          /*batch_size=*/batch_size,
                                          /*channels=*/channel_in,
                                          /*image_height=*/height_in,
                                          /*image_width=*/width_in));
    // 设置输出Tensor描述符
    cudnnTensorDescriptor_t output_descriptor;
    checkCUDNN(cudnnCreateTensorDescriptor(&output_descriptor));
    checkCUDNN(cudnnSetTensor4dDescriptor(output_descriptor,
                                      /*format=*/CUDNN_TENSOR_NCHW,
                                      /*dataType=*/CUDNN_DATA_FLOAT,
                                      /*batch_size=*/batch_size,
                                      /*channels=*/channel_out,
                                      /*image_height=*/height_out,
                                      /*image_width=*/width_out));
    // 设置bias描述符
    cudnnTensorDescriptor_t bias_descriptor;
    checkCUDNN(cudnnCreateTensorDescriptor(&bias_descriptor));
    checkCUDNN(cudnnSetTensor4dDescriptor(bias_descriptor,
                                      /*format=*/CUDNN_TENSOR_NCHW,
                                      /*dataType=*/CUDNN_DATA_FLOAT,
                                      /*batch_size=*/1,
                                      /*channels=*/channel_out,
                                      /*image_height=*/1,
                                      /*image_width=*/1));
    // 设置z描述符
    // // y = act ( alpha1 * conv(x) + alpha2 * z + bias ) 这里用不到
    cudnnTensorDescriptor_t z_descriptor;
    checkCUDNN(cudnnCreateTensorDescriptor(&z_descriptor));
    checkCUDNN(cudnnSetTensor4dDescriptor(z_descriptor,
                                      /*format=*/CUDNN_TENSOR_NCHW,
                                      /*dataType=*/CUDNN_DATA_FLOAT,
                                      /*batch_size=*/batch_size,
                                      /*channels=*/channel_out,
                                      /*image_height=*/height_out,
                                      /*image_width=*/width_out));
    // 设置conv weight的描述
    cudnnFilterDescriptor_t kernel_descriptor;
    checkCUDNN(cudnnCreateFilterDescriptor(&kernel_descriptor));
    checkCUDNN(cudnnSetFilter4dDescriptor(kernel_descriptor,
                                          /*dataType=*/CUDNN_DATA_FLOAT,
                                          /*format=*/CUDNN_TENSOR_NCHW,
                                          /*out_channels=*/channel_out,
                                          /*in_channels=*/channel_in,
                                          /*kernel_height=*/kernel_h,
                                          /*kernel_width=*/kernel_w));
    // 设置卷积相关参数
    cudnnConvolutionDescriptor_t convolution_descriptor;
    checkCUDNN(cudnnCreateConvolutionDescriptor(&convolution_descriptor));
    checkCUDNN(cudnnSetConvolution2dDescriptor(convolution_descriptor,
                                              /*pad_height=*/0,
                                              /*pad_width=*/0,
                                              /*vertical_stride=*/1,
                                              /*horizontal_stride=*/1,
                                              /*dilation_height=*/1,
                                              /*dilation_width=*/1,
                                              /*mode=*/CUDNN_CROSS_CORRELATION,
                                              /*computeType=*/CUDNN_DATA_FLOAT));
    // 设置激活层相关参数
    cudnnActivationDescriptor_t activation_descriptor;
    checkCUDNN(cudnnCreateActivationDescriptor(&activation_descriptor));
    checkCUDNN(cudnnSetActivationDescriptor(activation_descriptor,
                                            /*mode=*/CUDNN_ACTIVATION_RELU,
                                            /*reluNanOpt=*/CUDNN_PROPAGATE_NAN,
                                            /*relu_coef=*/0));
    // 获取卷积计算算法相关参数和workspace
    int cnt = 0;
    cudnnGetConvolutionForwardAlgorithmMaxCount(cudnn, &cnt);
    std::cout << "cnt: " << cnt << std::endl;
    cudnnConvolutionFwdAlgoPerf_t convolution_algorithm;
    int ret_cnt = 0;
    checkCUDNN(cudnnGetConvolutionForwardAlgorithm_v7(cudnn,
                                            input_descriptor,
                                            kernel_descriptor,
                                            convolution_descriptor,
                                            output_descriptor,
                                            1,
                                            &ret_cnt,
                                            &convolution_algorithm));
    size_t workspace_bytes = 0;
    checkCUDNN(cudnnGetConvolutionForwardWorkspaceSize(cudnn,
                                                      input_descriptor,
                                                      kernel_descriptor,
                                                      convolution_descriptor,
                                                      output_descriptor,
                                                      convolution_algorithm.algo,
                                                      &workspace_bytes));
    void* d_workspace{nullptr};
    cudaMalloc(&d_workspace, workspace_bytes);
    // 执行卷积运算
    checkCUDNN(cudnnConvolutionBiasActivationForward(
        cudnn, &alpha1, input_descriptor, qptr_gpu, kernel_descriptor, kernel_gpu,
        convolution_descriptor, convolution_algorithm.algo, d_workspace, workspace_bytes,
        &alpha2, z_descriptor, outptr_gpu,
        bias_descriptor, bias_gpu, activation_descriptor, output_descriptor, outptr_gpu));
    out_tensor.to_cpu(true);
    out_tensor.save_to_file("out_tensor.npz");
    // 销毁描述符和句柄
    cudnnDestroyTensorDescriptor(input_descriptor);
    cudnnDestroyTensorDescriptor(z_descriptor);
    cudnnDestroyTensorDescriptor(output_descriptor);
    cudnnDestroyTensorDescriptor(bias_descriptor);
    cudnnDestroyFilterDescriptor(kernel_descriptor);
    cudnnDestroyConvolutionDescriptor(convolution_descriptor);
    cudnnDestroy(cudnn);
    cudaFree(d_workspace);
    return 0;
}

四、验证环节

验证环节我们可以采用pytorch api来验证cudnn的实现是否能与cpu上torch的计算结果对齐。主要还是用到npy文件来保存，相关py代码如下：

import numpy as np
import torch.nn.functional as F
import torch
def _load_tensor(file):
    with open(file, "rb") as f:
        binary_data = f.read()
    magic_number, ndims, dtype = np.frombuffer(binary_data, np.uint32, count=3, offset=0)
    assert magic_number == 0xFCCFE2E2, f"{file} not a tensor file."
    dims = np.frombuffer(binary_data, np.uint32, count=ndims, offset=3 * 4)
    if dtype == 0:
        np_dtype = np.float32
    elif dtype == 1:
        np_dtype = np.float16
    else:
        assert False, f"Unsupport dtype = {dtype}, can not convert to numpy dtype"
    return np.frombuffer(binary_data, np_dtype, offset=(ndims + 3) * 4).reshape(*dims)
def load_tensor(file):
    if file.endswith("npz"):
        return np.load(file)['data']
    elif file.endswith("npy"):
        return np.load(file)
    else:
        return _load_tensor(file)
def test():
    input_tensor = load_tensor('q_tensor.npz')
    weight_tensor = load_tensor('kernel_tensor.npz')
    bias_tensor = load_tensor('bias_tensor.npz')
    out_tensor = load_tensor('out_tensor.npz')
    input_tensor = torch.as_tensor(input_tensor).float()
    weight_tensor = torch.as_tensor(weight_tensor).float()
    bias_tensor = torch.as_tensor(bias_tensor).float()
    out_tensor = torch.as_tensor(out_tensor).float()
    print(input_tensor.shape)
    print(weight_tensor.shape)
    print(bias_tensor.shape)
    print(out_tensor.shape)
    # 关键方法 F.conv2d
    out_tensor_torch = F.conv2d(input_tensor, weight_tensor, bias_tensor, stride=1)
    out_tensor_torch = F.relu(out_tensor_torch)
    print(torch.abs(out_tensor_torch - out_tensor).max())

if __name__ == "__main__":
    test()

五、总结

本文以二维卷积运算为例简单介绍了cudnn API的使用与验证。cuDNN为cuda开发者提供便利，一些专业的cuda开发者在写cuda算子的时候常常做的一件事就是先看看cudnn上有没有实现，并且验证自己的实现与cudnn的实现相比哪个更优。在日常的cuda开发中，对cudnn的熟悉是必不可少的环节，官网的例子较为丰富，遇到哪个api不会用的时候直接在github上搜索就可以找到用例。

---

最后更新: March 21, 2024
创建日期: March 21, 2024