videoStreamInfer/src/infer/trt/classfier/classifier.cu

#include "infer/trt/classfier/classifier.hpp"
#include <mutex>
#include <vector>
#include <algorithm>

namespace cls
{


static __global__ void softmax(float *predict, int length, int *max_index) {
    extern __shared__ float shared_data[];
    float *shared_max_vals = shared_data;
    int *shared_max_indices = (int*)&shared_max_vals[blockDim.x];
    
    int tid = threadIdx.x;

    // 1. 找到最大值和最大值的下标，存储在共享内存中
    float max_val = -FLT_MAX;
    int max_idx = -1;
    for (int i = tid; i < length; i += blockDim.x) {
        if (predict[i] > max_val) {
            max_val = predict[i];
            max_idx = i;
        }
    }
    shared_max_vals[tid] = max_val;
    shared_max_indices[tid] = max_idx;
    __syncthreads();

    // 在所有线程间找到全局最大值和对应的下标
    if (tid == 0) {
        for (int i = 1; i < blockDim.x; i++) {
            if (shared_max_vals[i] > shared_max_vals[0]) {
                shared_max_vals[0] = shared_max_vals[i];
                shared_max_indices[0] = shared_max_indices[i];
            }
        }
        *max_index = shared_max_indices[0];
    }
    __syncthreads();

    max_val = shared_max_vals[0];

    // 2. 计算指数并求和
    float sum_exp = 0.0f;
    for (int i = tid; i < length; i += blockDim.x) {
        predict[i] = expf(predict[i] - max_val);
        sum_exp += predict[i];
    }
    shared_max_vals[tid] = sum_exp;
    __syncthreads();

    // 汇总所有线程的指数和
    if (tid == 0) {
        for (int i = 1; i < blockDim.x; i++) {
            shared_max_vals[0] += shared_max_vals[i];
        }
    }
    __syncthreads();
    float total_sum = shared_max_vals[0];

    // 3. 每个元素除以总和，得到 softmax 值
    for (int i = tid; i < length; i += blockDim.x) {
        predict[i] /= total_sum;
    }
}

static void classfier_softmax(float *predict, int length, int *max_index, cudaStream_t stream) {
    int block_size = 256;
    checkKernel(softmax<<<1, block_size, block_size * sizeof(float), stream>>>(predict, length, max_index));
}
    

bool ClassifierModelImpl::load(const std::string &engine_file, int gpu_id)
{
    trt_ = TensorRT::load(engine_file);
    device_id_ = gpu_id;
    if (trt_ == nullptr) return false;

    trt_->print();

    auto input_dim = trt_->static_dims(0);
    network_input_width_ = input_dim[3];
    network_input_height_ = input_dim[2];
    isdynamic_model_ = trt_->has_dynamic_dim();
    num_classes_ = trt_->static_dims(1)[1];
    float mean[3] = {0.485f, 0.456f, 0.406f};
    float std[3] = {0.229f, 0.224f, 0.225f};
    normalize_ = affine::Norm::mean_std(mean, std, 1 / 255.0f, affine::ChannelType::SwapRB);
    // normalize_ = affine::Norm::alpha_beta(1 / 255.0f, 0.0f, affine::ChannelType::SwapRB);
    return true;
}

void ClassifierModelImpl::preprocess(int ibatch, affine::CropResizeMatrix& matrix, int x, int y, int w, int h, void *stream)
{
    matrix.compute(
        std::make_tuple(w, h),
        std::make_tuple(network_input_width_, network_input_height_));
    size_t input_numel = network_input_width_ * network_input_height_ * 3;

    float *input_device = input_buffer_.gpu() + ibatch * input_numel;

    uint8_t *image_device = preprocess_buffer_.gpu();
    uint8_t *image_host   = preprocess_buffer_.cpu();

    float *affine_matrix_device = affine_matrix_.gpu();
    float *affine_matrix_host = affine_matrix_.cpu();
    
    cudaStream_t stream_ = (cudaStream_t)stream;
    memcpy(affine_matrix_host, matrix.d2i, sizeof(matrix.d2i));
    checkRuntime(cudaMemcpyAsync(affine_matrix_device, affine_matrix_host, sizeof(matrix.d2i),
                                cudaMemcpyHostToDevice, stream_));

    affine::warp_affine_bilinear_and_normalize_plane(image_device, image.width * 3, image.width,
        image.height, input_device, network_input_width_,
        network_input_height_, affine_matrix_device, 114,
        normalize_, stream_);
    checkRuntime(cudaStreamSynchronize(stream_));
}

virtual Result ClassifierModelImpl::forward(const tensor::Image &image, void *stream)
{
    return;
}
virtual Result ClassifierModelImpl::forward(const tensor::Image &image, int slice_width, int slice_height, float overlap_width_ratio, float overlap_height_ratio, void *stream)
{
    return;
}
virtual Result ClassifierModelImpl::forward(const tensor::Image &image, data::BoxArray& boxes, void *stream)
{
    std::lock_guard<std::mutex> lock(mutex_);
    std::vector<data::Box*> classfier_boxes_ptr;
    for (auto& box : boxes)
    {
        if (std::find(box.label, class_names_.begin(), class_names_.end()) != class_names_.end())
        {
            classfier_boxes_ptr.push_back(&box);
        }
    }
    int num_image = classfier_boxes_ptr.size();
    if (num_image == 0){ return; }

    auto input_dims = trt_->static_dims(0);
    int infer_batch_size = input_dims[0];
    if (infer_batch_size != num_image) 
    {
        if (isdynamic_model_) 
        {
            infer_batch_size = num_image;
            input_dims[0] = num_image;
            if (!trt_->set_run_dims(0, input_dims)) 
            {
                printf("Fail to set run dims\n");
                return;
            }
        } 
        else 
        {
            if (infer_batch_size < num_image) 
            {
                printf(
                    "When using static shape model, number of images[%d] must be "
                    "less than or equal to the maximum batch[%d].",
                    num_image, infer_batch_size);
                return;
            }
        }
    }
    adjust_memory(num_image, image.width, image.height);
    
    uint8_t *image_device = preprocess_buffer_.gpu();
    uint8_t *image_host   = preprocess_buffer_.cpu();
    size_t size_image     = image.width * image.height * 3;
    cudaStream_t stream_  = (cudaStream_t)stream;
    memcpy(image_host, image.bgrptr, size_image);
    checkRuntime(
        cudaMemcpyAsync(image_device, image_host, size_image, cudaMemcpyHostToDevice, stream_));
    affine::CropResizeMatrix crmatrix;
    for(int i = 0; i < num_image; i++)
    {
        data::Box* box_ptr = classfier_boxes_ptr[i];
        int x = (int)box_ptr->left;
        int y = (int)box_ptr->top;
        int w = (int)box_ptr->right - x;
        int h = (int)box_ptr->bottom - y;
        preprocess(i, crmatrix, x, y, w, h, stream);
    }

    #ifdef TRT10
    if (!trt_->forward(std::unordered_map<std::string, const void *>{
            { "input", input_buffer_.gpu() }, 
            { "output", output_buffer_.gpu() }
        }, stream_))
    {
        printf("Failed to tensorRT forward.\n");
        return cv::Mat();
    }
    #else
    std::vector<void *> bindings{input_buffer_.gpu(), output_buffer_.gpu()};
    if (!trt_->forward(bindings, stream)) 
    {
        printf("Failed to tensorRT forward.");
        return cv::Mat();
    }
    #endif

    for (int ib = 0; ib < num_image; ++ib) 
    {
        float *output_buffer_device = output_buffer_.gpu() + ib * num_classes_;
        int *classes_indices_device  = classes_indices_.gpu() + ib;
        classfier_softmax(output_buffer_device, num_classes_, classes_indices_device, stream_);
    }

    checkRuntime(cudaMemcpyAsync(output_buffer_.cpu(), output_buffer_.gpu(),
                                 output_buffer_.gpu_bytes(), cudaMemcpyDeviceToHost, stream_));
    checkRuntime(cudaMemcpyAsync(classes_indices_.cpu(), classes_indices_.gpu(),
                                 classes_indices_.gpu_bytes(), cudaMemcpyDeviceToHost, stream_));
    checkRuntime(cudaStreamSynchronize(stream_));

    for (int ib = 0; ib < num_image; ++ib) 
    {
        int *max_index = classes_indices_.cpu() + ib;
        int index = *max_index;
        classfier_boxes_ptr[ib]->label = class_names_[index];
    }
}

}