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

added support of 4-channels images to StereoConstantSpaceBP.

Browse Source 
refactored transpose_gpu, made it non template function.
This commit is contained in:
Vladislav Vinogradov
2010-12-08 07:23:59 +00:00
parent c18aa438ec
commit 905e5f1739
5 changed files with 32 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
modules/gpu/src/cuda/constantspacebp.cu
View File

@@ -99,8 +99,15 @@ namespace cv { namespace gpu { namespace csbp
/////////////////////// init data cost ////////////////////////
///////////////////////////////////////////////////////////////
template <int channels>
struct DataCostPerPixel
template <int channels> struct DataCostPerPixel;
template <> struct DataCostPerPixel<1>
{
static __device__ float compute(const uchar* left, const uchar* right)
{
return fmin(cdata_weight * abs((int)*left - *right), cdata_weight * cmax_data_term);
}
};
template <> struct DataCostPerPixel<3>
{
static __device__ float compute(const uchar* left, const uchar* right)
{
@@ -111,13 +118,18 @@ namespace cv { namespace gpu { namespace csbp
return fmin(cdata_weight * (tr + tg + tb), cdata_weight * cmax_data_term);
}
};
template <>
struct DataCostPerPixel<1>
template <> struct DataCostPerPixel<4>
{
static __device__ float compute(const uchar* left, const uchar* right)
{
return fmin(cdata_weight * abs((int)*left - *right), cdata_weight * cmax_data_term);
uchar4 l = *((const uchar4*)left);
uchar4 r = *((const uchar4*)right);
float tb = 0.114f * abs((int)l.x - r.x);
float tg = 0.587f * abs((int)l.y - r.y);
float tr = 0.299f * abs((int)l.z - r.z);
return fmin(cdata_weight * (tr + tg + tb), cdata_weight * cmax_data_term);
}
};
@@ -327,6 +339,7 @@ namespace cv { namespace gpu { namespace csbp
{
case 1: init_data_cost<T, 1><<<grid, threads, 0, stream>>>(h, w, level); break;
case 3: init_data_cost<T, 3><<<grid, threads, 0, stream>>>(h, w, level); break;
case 4: init_data_cost<T, 4><<<grid, threads, 0, stream>>>(h, w, level); break;
default: cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
}
@@ -345,6 +358,7 @@ namespace cv { namespace gpu { namespace csbp
{
case 1: init_data_cost_reduce<T, winsz, 1><<<grid, threads, smem_size, stream>>>(level, rows, cols, h); break;
case 3: init_data_cost_reduce<T, winsz, 3><<<grid, threads, smem_size, stream>>>(level, rows, cols, h); break;
case 4: init_data_cost_reduce<T, winsz, 4><<<grid, threads, smem_size, stream>>>(level, rows, cols, h); break;
default: cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
}
@@ -517,6 +531,7 @@ namespace cv { namespace gpu { namespace csbp
{
case 1: compute_data_cost<T, 1><<<grid, threads, 0, stream>>>(disp_selected_pyr, data_cost, h, w, level, nr_plane); break;
case 3: compute_data_cost<T, 3><<<grid, threads, 0, stream>>>(disp_selected_pyr, data_cost, h, w, level, nr_plane); break;
case 4: compute_data_cost<T, 4><<<grid, threads, 0, stream>>>(disp_selected_pyr, data_cost, h, w, level, nr_plane); break;
default: cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
}
@@ -536,6 +551,7 @@ namespace cv { namespace gpu { namespace csbp
{
case 1: compute_data_cost_reduce<T, winsz, 1><<<grid, threads, smem_size, stream>>>(disp_selected_pyr, data_cost, level, rows, cols, h, nr_plane); break;
case 3: compute_data_cost_reduce<T, winsz, 3><<<grid, threads, smem_size, stream>>>(disp_selected_pyr, data_cost, level, rows, cols, h, nr_plane); break;
case 4: compute_data_cost_reduce<T, winsz, 4><<<grid, threads, smem_size, stream>>>(disp_selected_pyr, data_cost, level, rows, cols, h, nr_plane); break;
default: cv::gpu::error("Unsupported channels count", __FILE__, __LINE__);
}
}

17
modules/gpu/src/cuda/mathfunc.cu
View File

@@ -1254,10 +1254,9 @@ namespace cv { namespace gpu { namespace mathfunc
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// transpose
template <typename T>
__global__ void transpose(const DevMem2D_<T> src, PtrStep_<T> dst)
__global__ void transpose(const DevMem2Di src, PtrStepi dst)
{
__shared__ T s_mem[16 * 17];
__shared__ int s_mem[16 * 17];
int x = blockIdx.x * blockDim.x + threadIdx.x;
int y = blockIdx.y * blockDim.y + threadIdx.y;
@@ -1280,22 +1279,14 @@ namespace cv { namespace gpu { namespace mathfunc
}
}
template <typename T>
void transpose_gpu(const DevMem2D& src, const DevMem2D& dst)
void transpose_gpu(const DevMem2Di& src, const DevMem2Di& dst)
{
dim3 threads(16, 16, 1);
dim3 grid(divUp(src.cols, 16), divUp(src.rows, 16), 1);
transpose<T><<<grid, threads>>>((DevMem2D_<T>)src, (DevMem2D_<T>)dst);
transpose<<<grid, threads>>>(src, dst);
cudaSafeCall( cudaThreadSynchronize() );
}
template void transpose_gpu<uchar4 >(const DevMem2D& src, const DevMem2D& dst);
template void transpose_gpu<char4 >(const DevMem2D& src, const DevMem2D& dst);
template void transpose_gpu<ushort2>(const DevMem2D& src, const DevMem2D& dst);
template void transpose_gpu<short2 >(const DevMem2D& src, const DevMem2D& dst);
template void transpose_gpu<int >(const DevMem2D& src, const DevMem2D& dst);
template void transpose_gpu<float >(const DevMem2D& src, const DevMem2D& dst);
//////////////////////////////////////////////////////////////////////////////////////////////////////////
// min/max