Skip to content

O
opencv
Commit d00fa6b8 authored by Vladislav Vinogradov's avatar Vladislav Vinogradov
Browse files
Options

improved type dispatching in gpu arithm functions

parent 60ddaa56
Hide whitespace changes
Inline Side-by-side
Showing with 107 additions and 118 deletions
modules/gpu/src/element_operations.cpp
View file @ d00fa6b8
......@@ -89,92 +89,74 @@ void cv::gpu::addWeighted(const GpuMat&, double, const GpuMat&, double, double,
namespace
{
typedef NppStatus (*npp_arithm_8u_t)(const Npp8u* pSrc1, int nSrc1Step, const Npp8u* pSrc2, int nSrc2Step, Npp8u* pDst, int nDstStep, NppiSize oSizeROI, int nScaleFactor);
typedef NppStatus (*npp_arithm_16u_t)(const Npp16u* pSrc1, int nSrc1Step, const Npp16u* pSrc2, int nSrc2Step, Npp16u* pDst, int nDstStep, NppiSize oSizeROI, int nScaleFactor);
typedef NppStatus (*npp_arithm_16s_t)(const Npp16s* pSrc1, int nSrc1Step, const Npp16s* pSrc2, int nSrc2Step, Npp16s* pDst, int nDstStep, NppiSize oSizeROI, int nScaleFactor);
typedef NppStatus (*npp_arithm_32s_t)(const Npp32s* pSrc1, int nSrc1Step, const Npp32s* pSrc2, int nSrc2Step, Npp32s* pDst, int nDstStep, NppiSize oSizeROI, int nScaleFactor);
typedef NppStatus (*npp_arithm_32f_t)(const Npp32f* pSrc1, int nSrc1Step, const Npp32f* pSrc2, int nSrc2Step, Npp32f* pDst, int nDstStep, NppiSize oSizeROI);
template<int DEPTH> struct NppTypeTraits;
template<> struct NppTypeTraits<CV_8U> { typedef Npp8u npp_t; };
template<> struct NppTypeTraits<CV_8S> { typedef Npp8s npp_t; };
template<> struct NppTypeTraits<CV_16U> { typedef Npp16u npp_t; };
template<> struct NppTypeTraits<CV_16S> { typedef Npp16s npp_t; typedef Npp16sc npp_complex_type; };
template<> struct NppTypeTraits<CV_32S> { typedef Npp32s npp_t; typedef Npp32sc npp_complex_type; };
template<> struct NppTypeTraits<CV_32F> { typedef Npp32f npp_t; typedef Npp32fc npp_complex_type; };
template<> struct NppTypeTraits<CV_64F> { typedef Npp64f npp_t; typedef Npp64fc npp_complex_type; };
bool nppArithmCaller(const GpuMat& src1, const GpuMat& src2, GpuMat& dst,
npp_arithm_8u_t npp_func_8uc1, npp_arithm_8u_t npp_func_8uc4,
npp_arithm_16u_t npp_func_16uc1, npp_arithm_16u_t npp_func_16uc4,
npp_arithm_16s_t npp_func_16sc1, npp_arithm_16s_t npp_func_16sc4,
npp_arithm_32s_t npp_func_32sc1,
npp_arithm_32f_t npp_func_32fc1, npp_arithm_32f_t npp_func_32fc4,
cudaStream_t stream)
template <int DEPTH> struct NppArithmFunc
{
bool useNpp = (src1.depth() == CV_8U || src1.depth() == CV_16U || src1.depth() == CV_16S || src1.depth() == CV_32S || src1.depth() == CV_32F);
typedef typename NppTypeTraits<DEPTH>::npp_t npp_t;
typedef NppStatus (*func_t)(const npp_t* pSrc1, int nSrc1Step, const npp_t* pSrc2, int nSrc2Step, npp_t* pDst, int nDstStep, NppiSize oSizeROI, int nScaleFactor);
};
template <> struct NppArithmFunc<CV_32F>
{
typedef NppTypeTraits<CV_32F>::npp_t npp_t;
if (!useNpp)
return false;
typedef NppStatus (*func_t)(const Npp32f* pSrc1, int nSrc1Step, const Npp32f* pSrc2, int nSrc2Step, Npp32f* pDst, int nDstStep, NppiSize oSizeROI);
};
bool aligned = isAligned(src1.data, 16) && isAligned(src2.data, 16) && isAligned(dst.data, 16);
template <int DEPTH, typename NppArithmFunc<DEPTH>::func_t func> struct NppArithm
{
typedef typename NppArithmFunc<DEPTH>::npp_t npp_t;
NppiSize sz;
sz.width = src1.cols * src1.channels();
sz.height = src1.rows;
static void call(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream)
{
NppStreamHandler h(stream);
NppStreamHandler h(stream);
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
if (aligned && src1.depth() == CV_8U && (sz.width % 4) == 0)
{
sz.width /= 4;
nppSafeCall( func((const npp_t*)src1.data, static_cast<int>(src1.step), (const npp_t*)src2.data, static_cast<int>(src2.step),
(npp_t*)dst.data, static_cast<int>(dst.step), sz, 0) );
nppSafeCall( npp_func_8uc4(src1.ptr<Npp8u>(), static_cast<int>(src1.step), src2.ptr<Npp8u>(), static_cast<int>(src2.step),
dst.ptr<Npp8u>(), static_cast<int>(dst.step), sz, 0) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
else if (src1.depth() == CV_8U)
static void call(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream)
{
nppSafeCall( npp_func_8uc1(src1.ptr<Npp8u>(), static_cast<int>(src1.step), src2.ptr<Npp8u>(), static_cast<int>(src2.step),
dst.ptr<Npp8u>(), static_cast<int>(dst.step), sz, 0) );
call(src1, src2, dst, PtrStepb(), stream);
}
else if (aligned && src1.depth() == CV_16U && (sz.width % 4) == 0)
{
sz.width /= 4;
};
template <typename NppArithmFunc<CV_32F>::func_t func> struct NppArithm<CV_32F, func>
{
typedef typename NppArithmFunc<CV_32F>::npp_t npp_t;
nppSafeCall( npp_func_16uc4(src1.ptr<Npp16u>(), static_cast<int>(src1.step), src2.ptr<Npp16u>(), static_cast<int>(src2.step),
dst.ptr<Npp16u>(), static_cast<int>(dst.step), sz, 0) );
}
else if (src1.depth() == CV_16U)
static void call(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, const PtrStepb& mask, cudaStream_t stream)
{
nppSafeCall( npp_func_16uc1(src1.ptr<Npp16u>(), static_cast<int>(src1.step), src2.ptr<Npp16u>(), static_cast<int>(src2.step),
dst.ptr<Npp16u>(), static_cast<int>(dst.step), sz, 0) );
}
else if (aligned && src1.depth() == CV_16S && (sz.width % 4) == 0)
{
sz.width /= 4;
NppStreamHandler h(stream);
nppSafeCall( npp_func_16sc4(src1.ptr<Npp16s>(), static_cast<int>(src1.step), src2.ptr<Npp16s>(), static_cast<int>(src2.step),
dst.ptr<Npp16s>(), static_cast<int>(dst.step), sz, 0) );
}
else if (src1.depth() == CV_16S)
{
nppSafeCall( npp_func_16sc1(src1.ptr<Npp16s>(), static_cast<int>(src1.step), src2.ptr<Npp16s>(), static_cast<int>(src2.step),
dst.ptr<Npp16s>(), static_cast<int>(dst.step), sz, 0) );
}
else if (src1.depth() == CV_32S)
{
nppSafeCall( npp_func_32sc1(src1.ptr<Npp32s>(), static_cast<int>(src1.step), src2.ptr<Npp32s>(), static_cast<int>(src2.step),
dst.ptr<Npp32s>(), static_cast<int>(dst.step), sz, 0) );
}
else if (aligned && src1.depth() == CV_32F && (sz.width % 4) == 0)
{
sz.width /= 4;
NppiSize sz;
sz.width = src1.cols;
sz.height = src1.rows;
nppSafeCall( func((const npp_t*)src1.data, static_cast<int>(src1.step), (const npp_t*)src2.data, static_cast<int>(src2.step),
(npp_t*)dst.data, static_cast<int>(dst.step), sz) );
nppSafeCall( npp_func_32fc4(src1.ptr<Npp32f>(), static_cast<int>(src1.step), src2.ptr<Npp32f>(), static_cast<int>(src2.step),
dst.ptr<Npp32f>(), static_cast<int>(dst.step), sz) );
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
}
else // if (src1.depth() == CV_32F)
static void call(const DevMem2Db& src1, const DevMem2Db& src2, const DevMem2Db& dst, double scale, cudaStream_t stream)
{
nppSafeCall( npp_func_32fc1(src1.ptr<Npp32f>(), static_cast<int>(src1.step), src2.ptr<Npp32f>(), static_cast<int>(src2.step),
dst.ptr<Npp32f>(), static_cast<int>(dst.step), sz) );
call(src1, src2, dst, PtrStepb(), stream);
}
if (stream == 0)
cudaSafeCall( cudaDeviceSynchronize() );
return true;
}
};
}
////////////////////////////////////////////////////////////////////////
......@@ -206,6 +188,18 @@ void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const Gpu
{0/*add_gpu<double, unsigned char>*/, 0/*add_gpu<double, signed char>*/, 0/*add_gpu<double, unsigned short>*/, 0/*add_gpu<double, short>*/, 0/*add_gpu<double, int>*/, 0/*add_gpu<double, float>*/, add_gpu<double, double>}
};
static const func_t npp_funcs[7] =
{
NppArithm<CV_8U, nppiAdd_8u_C1RSfs>::call,
0,
NppArithm<CV_16U, nppiAdd_16u_C1RSfs>::call,
NppArithm<CV_16S, nppiAdd_16s_C1RSfs>::call,
NppArithm<CV_32S, nppiAdd_32s_C1RSfs>::call,
NppArithm<CV_32F, nppiAdd_32f_C1R>::call,
add_gpu<double, double>
};
CV_Assert(src1.type() != CV_8S);
CV_Assert(src1.type() == src2.type() && src1.size() == src2.size());
CV_Assert(mask.empty() || (src1.channels() == 1 && mask.size() == src1.size() && mask.type() == CV_8U));
......@@ -218,16 +212,8 @@ void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const Gpu
if (mask.empty() && dst.type() == src1.type())
{
if (nppArithmCaller(src1, src2, dst,
nppiAdd_8u_C1RSfs, nppiAdd_8u_C4RSfs,
nppiAdd_16u_C1RSfs, nppiAdd_16u_C4RSfs,
nppiAdd_16s_C1RSfs, nppiAdd_16s_C4RSfs,
nppiAdd_32s_C1RSfs,
nppiAdd_32f_C1R, nppiAdd_32f_C4R,
stream))
{
return;
}
npp_funcs[src1.depth()](src1.reshape(1), src2.reshape(1), dst.reshape(1), PtrStepb(), stream);
return;
}
const func_t func = funcs[src1.depth()][dst.depth()];
......@@ -238,15 +224,6 @@ void cv::gpu::add(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, const Gpu
namespace
{
template<int type> struct NppTypeTraits;
template<> struct NppTypeTraits<CV_8U> { typedef Npp8u npp_t; };
template<> struct NppTypeTraits<CV_8S> { typedef Npp8s npp_t; };
template<> struct NppTypeTraits<CV_16U> { typedef Npp16u npp_t; };
template<> struct NppTypeTraits<CV_16S> { typedef Npp16s npp_t; typedef Npp16sc npp_complex_type; };
template<> struct NppTypeTraits<CV_32S> { typedef Npp32s npp_t; typedef Npp32sc npp_complex_type; };
template<> struct NppTypeTraits<CV_32F> { typedef Npp32f npp_t; typedef Npp32fc npp_complex_type; };
template<> struct NppTypeTraits<CV_64F> { typedef Npp64f npp_t; typedef Npp64fc npp_complex_type; };
template<int DEPTH, int cn> struct NppArithmScalarFunc
{
typedef typename NppTypeTraits<DEPTH>::npp_t npp_t;
......@@ -485,6 +462,18 @@ void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, cons
{0/*subtract_gpu<double, unsigned char>*/, 0/*subtract_gpu<double, signed char>*/, 0/*subtract_gpu<double, unsigned short>*/, 0/*subtract_gpu<double, short>*/, 0/*subtract_gpu<double, int>*/, 0/*subtract_gpu<double, float>*/, subtract_gpu<double, double>}
};
static const func_t npp_funcs[7] =
{
NppArithm<CV_8U, nppiSub_8u_C1RSfs>::call,
0,
NppArithm<CV_16U, nppiSub_16u_C1RSfs>::call,
NppArithm<CV_16S, nppiSub_16s_C1RSfs>::call,
NppArithm<CV_32S, nppiSub_32s_C1RSfs>::call,
NppArithm<CV_32F, nppiSub_32f_C1R>::call,
subtract_gpu<double, double>
};
CV_Assert(src1.type() != CV_8S);
CV_Assert(src1.type() == src2.type() && src1.size() == src2.size());
CV_Assert(mask.empty() || (src1.channels() == 1 && mask.size() == src1.size() && mask.type() == CV_8U));
......@@ -497,16 +486,8 @@ void cv::gpu::subtract(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, cons
if (mask.empty() && dst.type() == src1.type())
{
if (nppArithmCaller(src2, src1, dst,
nppiSub_8u_C1RSfs, nppiSub_8u_C4RSfs,
nppiSub_16u_C1RSfs, nppiSub_16u_C4RSfs,
nppiSub_16s_C1RSfs, nppiSub_16s_C4RSfs,
nppiSub_32s_C1RSfs,
nppiSub_32f_C1R, nppiSub_32f_C4R,
stream))
{
return;
}
npp_funcs[src1.depth()](src2.reshape(1), src1.reshape(1), dst.reshape(1), PtrStepb(), stream);
return;
}
const func_t func = funcs[src1.depth()][dst.depth()];
......@@ -604,6 +585,17 @@ void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, doub
{0/*multiply_gpu<double, unsigned char>*/, 0/*multiply_gpu<double, signed char>*/, 0/*multiply_gpu<double, unsigned short>*/, 0/*multiply_gpu<double, short>*/, 0/*multiply_gpu<double, int>*/, 0/*multiply_gpu<double, float>*/, multiply_gpu<double, double>}
};
static const func_t npp_funcs[7] =
{
NppArithm<CV_8U, nppiMul_8u_C1RSfs>::call,
0,
NppArithm<CV_16U, nppiMul_16u_C1RSfs>::call,
NppArithm<CV_16S, nppiMul_16s_C1RSfs>::call,
NppArithm<CV_32S, nppiMul_32s_C1RSfs>::call,
NppArithm<CV_32F, nppiMul_32f_C1R>::call,
multiply_gpu<double, double>
};
cudaStream_t stream = StreamAccessor::getStream(s);
if (src1.type() == CV_8UC4 && src2.type() == CV_32FC1)
......@@ -624,6 +616,7 @@ void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, doub
}
else
{
CV_Assert(src1.type() != CV_8S);
CV_Assert(src1.type() == src2.type() && src1.size() == src2.size());
if (dtype < 0)
......@@ -633,16 +626,8 @@ void cv::gpu::multiply(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, doub
if (scale == 1 && dst.type() == src1.type())
{
if (nppArithmCaller(src1, src2, dst,
nppiMul_8u_C1RSfs, nppiMul_8u_C4RSfs,
nppiMul_16u_C1RSfs, nppiMul_16u_C4RSfs,
nppiMul_16s_C1RSfs, nppiMul_16s_C4RSfs,
nppiMul_32s_C1RSfs,
nppiMul_32f_C1R, nppiMul_32f_C4R,
stream))
{
return;
}
npp_funcs[src1.depth()](src1.reshape(1), src2.reshape(1), dst.reshape(1), 1, stream);
return;
}
const func_t func = funcs[src1.depth()][dst.depth()];
......@@ -749,6 +734,17 @@ void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double
{0/*divide_gpu<double, unsigned char>*/, 0/*divide_gpu<double, signed char>*/, 0/*divide_gpu<double, unsigned short>*/, 0/*divide_gpu<double, short>*/, 0/*divide_gpu<double, int>*/, 0/*divide_gpu<double, float>*/, divide_gpu<double, double>}
};
static const func_t npp_funcs[7] =
{
NppArithm<CV_8U, nppiDiv_8u_C1RSfs>::call,
0,
NppArithm<CV_16U, nppiDiv_16u_C1RSfs>::call,
NppArithm<CV_16S, nppiDiv_16s_C1RSfs>::call,
NppArithm<CV_32S, nppiDiv_32s_C1RSfs>::call,
NppArithm<CV_32F, nppiDiv_32f_C1R>::call,
divide_gpu<double, double>
};
cudaStream_t stream = StreamAccessor::getStream(s);
if (src1.type() == CV_8UC4 && src2.type() == CV_32FC1)
......@@ -768,7 +764,8 @@ void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double
multiply_gpu(static_cast<DevMem2D_<short4> >(src1), static_cast<DevMem2Df>(src2), static_cast<DevMem2D_<short4> >(dst), stream);
}
else
{
{
CV_Assert(src1.type() != CV_8S);
CV_Assert(src1.type() == src2.type() && src1.size() == src2.size());
if (dtype < 0)
......@@ -778,16 +775,8 @@ void cv::gpu::divide(const GpuMat& src1, const GpuMat& src2, GpuMat& dst, double
if (scale == 1 && dst.type() == src1.type())
{
if (nppArithmCaller(src2, src1, dst,
nppiDiv_8u_C1RSfs, nppiDiv_8u_C4RSfs,
nppiDiv_16u_C1RSfs, nppiDiv_16u_C4RSfs,
nppiDiv_16s_C1RSfs, nppiDiv_16s_C4RSfs,
nppiDiv_32s_C1RSfs,
nppiDiv_32f_C1R, nppiDiv_32f_C4R,
stream))
{
return;
}
npp_funcs[src1.depth()](src2.reshape(1), src1.reshape(1), dst.reshape(1), 1, stream);
return;
}
const func_t func = funcs[src1.depth()][dst.depth()];
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment