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

moved GpuMat and DevMem2D to core module, some code refactoring

Browse Source
This commit is contained in:
Vladislav Vinogradov
2011-11-09 13:13:52 +00:00
parent 8a148e39f0
commit fcfa72081e
95 changed files with 18889 additions and 18485 deletions
Download Patch File Download Diff File Expand all files Collapse all files

300
modules/gpu/src/cuda/mathfunc.cu
View File

@@ -42,178 +42,174 @@
#include "internal_shared.hpp"
using namespace cv::gpu;
BEGIN_OPENCV_DEVICE_NAMESPACE
#ifndef CV_PI
#define CV_PI 3.1415926535897932384626433832795f
#endif
namespace mathfunc {
//////////////////////////////////////////////////////////////////////////////////////
// Cart <-> Polar
namespace cv { namespace gpu { namespace mathfunc
struct Nothing
{
struct Nothing
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
static __device__ __forceinline__ void calc(int, int, float, float, float*, size_t, float)
{
}
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = sqrtf(x_data * x_data + y_data * y_data);
}
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
};
struct Magnitude
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
dst[y * dst_step + x] = ::sqrtf(x_data * x_data + y_data * y_data);
}
template <typename Mag, typename Angle>
void cartToPolar_caller(const DevMem2Df& x, const DevMem2Df& y, const DevMem2Df& mag, const DevMem2Df& angle, bool angleInDegrees, cudaStream_t stream)
};
struct MagnitudeSqr
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
dst[y * dst_step + x] = x_data * x_data + y_data * y_data;
}
void cartToPolar_gpu(const DevMem2Df& x, const DevMem2Df& y, const DevMem2Df& mag, bool magSqr, const DevMem2Df& angle, bool angleInDegrees, cudaStream_t stream)
};
struct Atan2
{
static __device__ __forceinline__ void calc(int x, int y, float x_data, float y_data, float* dst, size_t dst_step, float scale)
{
float angle = ::atan2f(y_data, x_data);
angle += (angle < 0) * 2.0 * CV_PI;
dst[y * dst_step + x] = scale * angle;
}
};
template <typename Mag, typename Angle>
__global__ void cartToPolar(const float* xptr, size_t x_step, const float* yptr, size_t y_step,
float* mag, size_t mag_step, float* angle, size_t angle_step, float scale, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float x_data = xptr[y * x_step + x];
float y_data = yptr[y * y_step + x];
Mag::calc(x, y, x_data, y_data, mag, mag_step, scale);
Angle::calc(x, y, x_data, y_data, angle, angle_step, scale);
}
}
struct NonEmptyMag
{
static __device__ __forceinline__ float get(const float* mag, size_t mag_step, int x, int y)
{
return mag[y * mag_step + x];
}
};
struct EmptyMag
{
static __device__ __forceinline__ float get(const float*, size_t, int, int)
{
return 1.0f;
}
};
template <typename Mag>
__global__ void polarToCart(const float* mag, size_t mag_step, const float* angle, size_t angle_step, float scale,
float* xptr, size_t x_step, float* yptr, size_t y_step, int width, int height)
{
const int x = blockDim.x * blockIdx.x + threadIdx.x;
const int y = blockDim.y * blockIdx.y + threadIdx.y;
if (x < width && y < height)
{
float mag_data = Mag::get(mag, mag_step, x, y);
float angle_data = angle[y * angle_step + x];
float sin_a, cos_a;
::sincosf(scale * angle_data, &sin_a, &cos_a);
xptr[y * x_step + x] = mag_data * cos_a;
yptr[y * y_step + x] = mag_data * sin_a;
}
}
template <typename Mag, typename Angle>
void cartToPolar_caller(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(x.cols, threads.x);
grid.y = divUp(x.rows, threads.y);
const float scale = angleInDegrees ? (float)(180.0f / CV_PI) : 1.f;
cartToPolar<Mag, Angle><<<grid, threads, 0, stream>>>(
x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(),
mag.data, mag.step/mag.elemSize(), angle.data, angle.step/angle.elemSize(), scale, x.cols, x.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void cartToPolar_gpu(DevMem2Df x, DevMem2Df y, DevMem2Df mag, bool magSqr, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df x, DevMem2Df y, DevMem2Df mag, DevMem2Df angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
typedef void (*caller_t)(const DevMem2Df& x, const DevMem2Df& y, const DevMem2Df& mag, const DevMem2Df& angle, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2][2][2] =
{
{
{
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
cartToPolar_caller<Magnitude, Atan2>,
cartToPolar_caller<Magnitude, Nothing>
},
{
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
cartToPolar_caller<MagnitudeSqr, Atan2>,
cartToPolar_caller<MagnitudeSqr, Nothing>,
}
};
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag>
void polarToCart_caller(const DevMem2Df& mag, const DevMem2Df& angle, const DevMem2Df& x, const DevMem2Df& y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void polarToCart_gpu(const DevMem2Df& mag, const DevMem2Df& angle, const DevMem2Df& x, const DevMem2Df& y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(const DevMem2Df& mag, const DevMem2Df& angle, const DevMem2Df& x, const DevMem2Df& y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
},
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>
},
{
cartToPolar_caller<Nothing, Atan2>,
cartToPolar_caller<Nothing, Nothing>,
}
}
};
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
}}}
callers[mag.data == 0][magSqr][angle.data == 0](x, y, mag, angle, angleInDegrees, stream);
}
template <typename Mag>
void polarToCart_caller(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
dim3 threads(32, 8, 1);
dim3 grid(1, 1, 1);
grid.x = divUp(mag.cols, threads.x);
grid.y = divUp(mag.rows, threads.y);
const float scale = angleInDegrees ? (float)(CV_PI / 180.0f) : 1.0f;
polarToCart<Mag><<<grid, threads, 0, stream>>>(mag.data, mag.step/mag.elemSize(),
angle.data, angle.step/angle.elemSize(), scale, x.data, x.step/x.elemSize(), y.data, y.step/y.elemSize(), mag.cols, mag.rows);
cudaSafeCall( cudaGetLastError() );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
void polarToCart_gpu(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream)
{
typedef void (*caller_t)(DevMem2Df mag, DevMem2Df angle, DevMem2Df x, DevMem2Df y, bool angleInDegrees, cudaStream_t stream);
static const caller_t callers[2] =
{
polarToCart_caller<NonEmptyMag>,
polarToCart_caller<EmptyMag>
};
callers[mag.data == 0](mag, angle, x, y, angleInDegrees, stream);
}
} // namespace mathfunc
END_OPENCV_DEVICE_NAMESPACE