generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

added support of all data types into gpu::minMax

Browse Source
This commit is contained in:
Alexey Spizhevoy
2010-11-24 08:55:52 +00:00
parent 3c0cc087d6
commit 282e01cb4a
5 changed files with 230 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
modules/gpu/src/arithm.cpp
View File

@@ -489,55 +489,45 @@ Scalar cv::gpu::sum(const GpuMat& src)
////////////////////////////////////////////////////////////////////////
// minMax
namespace
{
void minMax_c1(const GpuMat& src, double* minVal, double* maxVal)
{
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
Npp8u min_res, max_res;
nppSafeCall( nppiMinMax_8u_C1R(src.ptr<Npp8u>(), src.step, sz, &min_res, &max_res) );
if (minVal)
*minVal = min_res;
if (maxVal)
*maxVal = max_res;
}
void minMax_c4(const GpuMat& src, double* minVal, double* maxVal)
{
NppiSize sz;
sz.width = src.cols;
sz.height = src.rows;
Npp8u* cuMem;
cuMem = nppsMalloc_8u(8);
nppSafeCall( nppiMinMax_8u_C4R(src.ptr<Npp8u>(), src.step, sz, cuMem, cuMem + 4) );
if (minVal)
cudaMemcpy(minVal, cuMem, 4 * sizeof(Npp8u), cudaMemcpyDeviceToHost);
if (maxVal)
cudaMemcpy(maxVal, cuMem + 4, 4 * sizeof(Npp8u), cudaMemcpyDeviceToHost);
nppsFree(cuMem);
}
}
namespace cv { namespace gpu { namespace mathfunc {
template <typename T>
void min_max_caller(const DevMem2D src, double* minval, double* maxval);
}}}
void cv::gpu::minMax(const GpuMat& src, double* minVal, double* maxVal)
{
typedef void (*minMax_t)(const GpuMat& src, double* minVal, double* maxVal);
static const minMax_t minMax_callers[] = {0, minMax_c1, 0, 0, minMax_c4};
CV_Assert(src.channels() == 1);
CV_Assert(!"disabled until fix npp bug");
CV_Assert(src.type() == CV_8UC1 || src.type() == CV_8UC4);
minMax_callers[src.channels()](src, minVal, maxVal);
double maxVal_;
if (!maxVal)
maxVal = &maxVal_;
switch (src.type())
{
case CV_8U:
mathfunc::min_max_caller<unsigned char>(src, minVal, maxVal);
break;
case CV_8S:
mathfunc::min_max_caller<signed char>(src, minVal, maxVal);
break;
case CV_16U:
mathfunc::min_max_caller<unsigned short>(src, minVal, maxVal);
break;
case CV_16S:
mathfunc::min_max_caller<signed short>(src, minVal, maxVal);
break;
case CV_32S:
mathfunc::min_max_caller<int>(src, minVal, maxVal);
break;
case CV_32F:
mathfunc::min_max_caller<float>(src, minVal, maxVal);
break;
case CV_64F:
mathfunc::min_max_caller<double>(src, minVal, maxVal);
break;
default:
CV_Error(CV_StsBadArg, "Unsupported type");
}
}
////////////////////////////////////////////////////////////////////////

128
modules/gpu/src/cuda/mathfunc.cu
View File

@@ -393,4 +393,132 @@ namespace cv { namespace gpu { namespace mathfunc
{
bitwise_bin_op<BIN_OP_XOR>(rows, cols, src1, src2, dst, elem_size, Mask8U(mask), stream);
}
//////////////////////////////////////////////////////////////////////////////
// Min max
enum { MIN, MAX };
template <typename T, int op> struct Cmp {};
template <typename T>
struct Cmp<T, MIN>
{
static __device__ void call(unsigned int tid, unsigned int offset, volatile T* optval)
{
T val = optval[tid + offset];
if (val < optval[tid]) optval[tid] = val;
//optval[tid] = min(optval[tid], optval[tid + offset]);
}
};
template <typename T>
struct Cmp<T, MAX>
{
static __device__ void call(unsigned int tid, unsigned int offset, volatile T* optval)
{
T val = optval[tid + offset];
if (val > optval[tid]) optval[tid] = val;
//optval[tid] = max(optval[tid], optval[tid + offset]);
}
};
template <int nthreads, typename Cmp, typename T>
__global__ void opt_kernel(int cols, int rows, const PtrStep src, PtrStep optval)
{
__shared__ T soptval[nthreads];
unsigned int x0 = blockIdx.x * blockDim.x;
unsigned int y0 = blockIdx.y * blockDim.y;
unsigned int tid = threadIdx.y * blockDim.x + threadIdx.x;
if (x0 + threadIdx.x < cols && y0 + threadIdx.y < rows)
soptval[tid] = ((const T*)src.ptr(y0 + threadIdx.y))[x0 + threadIdx.x];
else
soptval[tid] = ((const T*)src.ptr(y0))[x0];
__syncthreads();
if (nthreads >= 512) if (tid < 256) { Cmp::call(tid, 256, soptval); __syncthreads(); }
if (nthreads >= 256) if (tid < 128) { Cmp::call(tid, 128, soptval); __syncthreads(); }
if (nthreads >= 128) if (tid < 64) { Cmp::call(tid, 64, soptval); __syncthreads(); }
if (tid < 32)
{
if (nthreads >= 64) Cmp::call(tid, 32, soptval);
if (nthreads >= 32) Cmp::call(tid, 16, soptval);
if (nthreads >= 16) Cmp::call(tid, 8, soptval);
if (nthreads >= 8) Cmp::call(tid, 4, soptval);
if (nthreads >= 4) Cmp::call(tid, 2, soptval);
if (nthreads >= 2) Cmp::call(tid, 1, soptval);
}
if (tid == 0) ((T*)optval.ptr(blockIdx.y))[blockIdx.x] = soptval[0];
}
template <typename T>
void min_max_caller(const DevMem2D src, double* minval, double* maxval)
{
dim3 threads(32, 8);
// Allocate memory for aux. buffers
DevMem2D minval_buf[2]; DevMem2D maxval_buf[2];
minval_buf[0].cols = divUp(src.cols, threads.x);
minval_buf[0].rows = divUp(src.rows, threads.y);
minval_buf[1].cols = divUp(minval_buf[0].cols, threads.x);
minval_buf[1].rows = divUp(minval_buf[0].rows, threads.y);
maxval_buf[0].cols = divUp(src.cols, threads.x);
maxval_buf[0].rows = divUp(src.rows, threads.y);
maxval_buf[1].cols = divUp(maxval_buf[0].cols, threads.x);
maxval_buf[1].rows = divUp(maxval_buf[0].rows, threads.y);
cudaSafeCall(cudaMallocPitch(&minval_buf[0].data, &minval_buf[0].step, minval_buf[0].cols * sizeof(T), minval_buf[0].rows));
cudaSafeCall(cudaMallocPitch(&minval_buf[1].data, &minval_buf[1].step, minval_buf[1].cols * sizeof(T), minval_buf[1].rows));
cudaSafeCall(cudaMallocPitch(&maxval_buf[0].data, &maxval_buf[0].step, maxval_buf[0].cols * sizeof(T), maxval_buf[0].rows));
cudaSafeCall(cudaMallocPitch(&maxval_buf[1].data, &maxval_buf[1].step, maxval_buf[1].cols * sizeof(T), maxval_buf[1].rows));
int curbuf = 0;
dim3 cursize(src.cols, src.rows);
dim3 grid(divUp(cursize.x, threads.x), divUp(cursize.y, threads.y));
opt_kernel<256, Cmp<T, MIN>, T><<<grid, threads>>>(cursize.x, cursize.y, src, minval_buf[curbuf]);
opt_kernel<256, Cmp<T, MAX>, T><<<grid, threads>>>(cursize.x, cursize.y, src, maxval_buf[curbuf]);
cursize = grid;
while (cursize.x > 1 || cursize.y > 1)
{
grid.x = divUp(cursize.x, threads.x);
grid.y = divUp(cursize.y, threads.y);
opt_kernel<256, Cmp<T, MIN>, T><<<grid, threads>>>(cursize.x, cursize.y, minval_buf[curbuf], minval_buf[1 - curbuf]);
opt_kernel<256, Cmp<T, MAX>, T><<<grid, threads>>>(cursize.x, cursize.y, maxval_buf[curbuf], maxval_buf[1 - curbuf]);
curbuf = 1 - curbuf;
cursize = grid;
}
cudaSafeCall(cudaThreadSynchronize());
// Copy results from device to host
T minval_, maxval_;
cudaSafeCall(cudaMemcpy(&minval_, minval_buf[curbuf].ptr(0), sizeof(T), cudaMemcpyDeviceToHost));
cudaSafeCall(cudaMemcpy(&maxval_, maxval_buf[curbuf].ptr(0), sizeof(T), cudaMemcpyDeviceToHost));
*minval = minval_;
*maxval = maxval_;
// Release aux. buffers
cudaSafeCall(cudaFree(minval_buf[0].data));
cudaSafeCall(cudaFree(minval_buf[1].data));
cudaSafeCall(cudaFree(maxval_buf[0].data));
cudaSafeCall(cudaFree(maxval_buf[1].data));
}
template void min_max_caller<unsigned char>(const DevMem2D, double*, double*);
template void min_max_caller<signed char>(const DevMem2D, double*, double*);
template void min_max_caller<unsigned short>(const DevMem2D, double*, double*);
template void min_max_caller<signed short>(const DevMem2D, double*, double*);
template void min_max_caller<int>(const DevMem2D, double*, double*);
template void min_max_caller<float>(const DevMem2D, double*, double*);
template void min_max_caller<double>(const DevMem2D, double*, double*);
}}}