【发布时间】：2017-04-27 22:37:04
【问题描述】：
我正在使用 opencv 3.1 cv::cuda 模板匹配，但 cv::cuda::minMaxLoc() 函数对我来说太慢了。我的匹配结果的最小尺寸为128x128，最大尺寸为512x512。平均而言，minMaxLoc() 对于 128x128 将需要 1.65 毫秒，而对于像 350x350 这样的操作则最多需要 25 毫秒，这太长了，因为这已经完成了数百次。
我知道我的匹配大小对于您通常在 GPU 中使用的大小可能太小了。但我想按照 Robert Crovella 在thrust::max_element slow in comparison cublasIsamax - More efficient implementation? 所做的测试，看看我是否可以获得更好的性能。
我的问题是，所有这些数据都是使用线性索引读取的，cv::cuda::PtrStepSzf不允许这样做（至少我没有找到方法）。我试图重塑我的匹配结果，但我不能这样做，因为数据不连续。我是否需要前往cudaMallocPitch 和cudaMemcpy2D如果是这种情况，我如何将cv::cuda::GPUMat 读取为cv::cuda::PtrStepSzf 对象？
__global__ void minLoc(const cv::cuda::PtrStepSzf data,
float* minVal,
float * minValLoc
)
{
int dsize = data.cols*data.rows
__shared__ volatile T vals[nTPB];
__shared__ volatile int idxs[nTPB];
__shared__ volatile int last_block;
int idx = threadIdx.x+blockDim.x*blockIdx.x;
last_block = 0;
T my_val = FLOAT_MIN;
int my_idx = -1;
// sweep from global memory
while (idx < dsize)
{
//data(idx) is an illegal call;The legal one is data(x,y)
// How do I do it?
if (data(idx) > my_val)
{
my_val = data(idx); my_idx = idx;
}
idx += blockDim.x*gridDim.x;
}
// ... rest of the kernel
}
void callMinLocKernel(cv::InputArray _input,
cv::Point minValLoc,
float minVal,
cv::cuda::Stream _stream)
{
const cv::cuda::GpuMat input = _input.getGpuMat();
dim3 cthreads(32, 32);
dim3 cblocks(
static_cast<int>(std::ceil(input1.size().width /
static_cast<double>(cthreads.x))),
static_cast<int>(std::ceil(input1.size().height /
static_cast<double>(cthreads.y))));
// code that creates and upload d_min, d_minLoc
float h_min = 9999;
int h_minLoc = -1;
float * d_min = 0;
int * d_minLoc = 0;
//gpuErrchk is defined on other place
gpuErrchk( cudaMalloc((void**)&d_min, sizeof(h_min)));
gpuErrchk( cudaMalloc((void**)&d_minLoc, sizeof(h_minLoc));
gpuErrchk( cudaMemcpy(d_min, &h_min, sizeof(h_min), cudaMemcpyHostToDevice) );
gpuErrchk( cudaMemcpy(d_minLoc, &h_minLoc, sizeof(h_minLoc), cudaMemcpyHostToDevice) );
cudaStream_t stream = cv::cuda::StreamAccessor::getStream(_stream);
minLoc<<<cblocks, cthreads, 0, stream>>>(input,d_min,d_minLoc);
gpuErrchk(cudaGetLastError());
//code to read the answer
gpuErrchk( cudaMemcpy(&h_min, d_min, sizeof(h_min), cudaMemcpyDeviceToHost) );
gpuErrchk( cudaMemcpy(&h_minLoc, d_minLoc, sizeof(h_minLoc), cudaMemcpyDeviceToHost) );
minValLoc = cv::point(h_minLoc/data.cols,h_minLoc%data.cols)
minVal = h_min;
}
int main()
{
//read Background and template
cv::Mat input = imread("cat.jpg",0);
cv::Mat templ = imread("catNose.jpg",0)
//convert to floats
cv::Mat float_input, float_templ;
input.convertTo(float_input,CV_32FC1);
input.convertTo(float_templ,CV_32FC1);
//upload Bckg and template to gpu
cv::cuda::GpuMat d_src,d_templ, d_match;
Size size = float_input.size();
d_src.upload(float_input);
d_templ.upload(float_templ);
double min_val, max_val;
Point min_loc, max_loc;
Ptr<cv::cuda::TemplateMatching> alg = cuda::createTemplateMatching(d_src.type(), CV_TM_SQDIFF);
alg->match(d_src, d_templ, d_match);
cv::cuda::Normalize(d_match,d_match);
//Too slow
//cv::cuda::minMaxLoc(d_match, &min_val, &max_val, &min_loc, &max_loc);
callMinLocKernel(d_match,min_val,min_loc);
return 0;
}
【问题讨论】：