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

used new device layer for cv::gpu::reduce

Browse Source
This commit is contained in:
Vladislav Vinogradov
2013-08-27 12:28:01 +04:00
parent 31a7814395
commit 1ef211b889
6 changed files with 316 additions and 468 deletions
Download Patch File Download Diff File Expand all files Collapse all files

44
modules/cudev/include/opencv2/cudev/grid/detail/reduce_to_column.hpp
View File

@@ -54,12 +54,52 @@ namespace cv { namespace cudev {
namespace grid_reduce_to_vec_detail
{
template <int BLOCK_SIZE, typename work_type, typename work_elem_type, class Reductor, int cn> struct Reduce;
template <int BLOCK_SIZE, typename work_type, typename work_elem_type, class Reductor> struct Reduce<BLOCK_SIZE, work_type, work_elem_type, Reductor, 1>
{
__device__ __forceinline__ static void call(work_elem_type smem[1][BLOCK_SIZE], work_type& myVal)
{
typename Reductor::template rebind<work_elem_type>::other op;
blockReduce<BLOCK_SIZE>(smem[0], myVal, threadIdx.x, op);
}
};
template <int BLOCK_SIZE, typename work_type, typename work_elem_type, class Reductor> struct Reduce<BLOCK_SIZE, work_type, work_elem_type, Reductor, 2>
{
__device__ __forceinline__ static void call(work_elem_type smem[2][BLOCK_SIZE], work_type& myVal)
{
typename Reductor::template rebind<work_elem_type>::other op;
blockReduce<BLOCK_SIZE>(smem_tuple(smem[0], smem[1]), tie(myVal.x, myVal.y), threadIdx.x, make_tuple(op, op));
}
};
template <int BLOCK_SIZE, typename work_type, typename work_elem_type, class Reductor> struct Reduce<BLOCK_SIZE, work_type, work_elem_type, Reductor, 3>
{
__device__ __forceinline__ static void call(work_elem_type smem[3][BLOCK_SIZE], work_type& myVal)
{
typename Reductor::template rebind<work_elem_type>::other op;
blockReduce<BLOCK_SIZE>(smem_tuple(smem[0], smem[1], smem[2]), tie(myVal.x, myVal.y, myVal.z), threadIdx.x, make_tuple(op, op, op));
}
};
template <int BLOCK_SIZE, typename work_type, typename work_elem_type, class Reductor> struct Reduce<BLOCK_SIZE, work_type, work_elem_type, Reductor, 4>
{
__device__ __forceinline__ static void call(work_elem_type smem[4][BLOCK_SIZE], work_type& myVal)
{
typename Reductor::template rebind<work_elem_type>::other op;
blockReduce<BLOCK_SIZE>(smem_tuple(smem[0], smem[1], smem[2], smem[3]), tie(myVal.x, myVal.y, myVal.z, myVal.w), threadIdx.x, make_tuple(op, op, op, op));
}
};
template <class Reductor, int BLOCK_SIZE, class SrcPtr, typename ResType, class MaskPtr>
__global__ void reduceToColumn(const SrcPtr src, ResType* dst, const MaskPtr mask, const int cols)
{
typedef typename Reductor::work_type work_type;
typedef typename VecTraits<work_type>::elem_type work_elem_type;
const int cn = VecTraits<work_type>::cn;
__shared__ work_type smem[BLOCK_SIZE];
__shared__ work_elem_type smem[cn][BLOCK_SIZE];
const int y = blockIdx.x;
@@ -75,7 +115,7 @@ namespace grid_reduce_to_vec_detail
}
}
blockReduce<BLOCK_SIZE>(smem, myVal, threadIdx.x, op);
Reduce<BLOCK_SIZE, work_type, work_elem_type, Reductor, cn>::call(smem, myVal);
if (threadIdx.x == 0)
dst[y] = saturate_cast<ResType>(Reductor::result(myVal, cols));

29
modules/cudev/include/opencv2/cudev/grid/reduce_to_vec.hpp
View File

@@ -49,6 +49,7 @@
#include "../common.hpp"
#include "../util/vec_traits.hpp"
#include "../util/limits.hpp"
#include "../util/saturate_cast.hpp"
#include "../ptr2d/traits.hpp"
#include "../ptr2d/gpumat.hpp"
#include "../ptr2d/mask.hpp"
@@ -62,6 +63,11 @@ template <typename T> struct Sum : plus<T>
{
typedef T work_type;
template <typename U> struct rebind
{
typedef Sum<U> other;
};
__device__ __forceinline__ static T initialValue()
{
return VecTraits<T>::all(0);
@@ -77,14 +83,19 @@ template <typename T> struct Avg : plus<T>
{
typedef T work_type;
template <typename U> struct rebind
{
typedef Avg<U> other;
};
__device__ __forceinline__ static T initialValue()
{
return VecTraits<T>::all(0);
}
__device__ __forceinline__ static T result(T r, int sz)
__device__ __forceinline__ static T result(T r, float sz)
{
return r / sz;
return saturate_cast<T>(r / sz);
}
};
@@ -92,6 +103,11 @@ template <typename T> struct Min : minimum<T>
{
typedef T work_type;
template <typename U> struct rebind
{
typedef Min<U> other;
};
__device__ __forceinline__ static T initialValue()
{
return VecTraits<T>::all(numeric_limits<typename VecTraits<T>::elem_type>::max());
@@ -107,6 +123,11 @@ template <typename T> struct Max : maximum<T>
{
typedef T work_type;
template <typename U> struct rebind
{
typedef Max<U> other;
};
__device__ __forceinline__ static T initialValue()
{
return VecTraits<T>::all(-numeric_limits<typename VecTraits<T>::elem_type>::max());
@@ -158,7 +179,7 @@ __host__ void gridReduceToColumn_(const SrcPtr& src, GpuMat_<ResType>& dst, cons
CV_Assert( getRows(mask) == rows && getCols(mask) == cols );
createContinuous(rows, 1, DataType<ResType>::type, dst);
dst.create(1, rows);
grid_reduce_to_vec_detail::reduceToColumn<Reductor, Policy>(shrinkPtr(src),
dst[0],
@@ -173,7 +194,7 @@ __host__ void gridReduceToColumn_(const SrcPtr& src, GpuMat_<ResType>& dst, Stre
const int rows = getRows(src);
const int cols = getCols(src);
createContinuous(rows, 1, DataType<ResType>::type, dst);
dst.create(1, rows);
grid_reduce_to_vec_detail::reduceToColumn<Reductor, Policy>(shrinkPtr(src),
dst[0],