potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

Browse Source
This commit is contained in:
Alexander Alekhin
2019-02-19 19:36:46 +00:00
parent f1d9666a20 8cedc052ca
commit 8bde6aea4b
137 changed files with 9339 additions and 4401 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/CMakeLists.txt
+3
View File
@@ -5,6 +5,9 @@ ocv_add_dispatched_file(stat SSE4_2 AVX2)
ocv_add_dispatched_file(arithm SSE2 SSE4_1 AVX2 VSX3)
ocv_add_dispatched_file(convert SSE2 AVX2)
ocv_add_dispatched_file(convert_scale SSE2 AVX2)
ocv_add_dispatched_file(count_non_zero SSE2 AVX2)
ocv_add_dispatched_file(matmul SSE2 AVX2)
ocv_add_dispatched_file(sum SSE2 AVX2)
# dispatching for accuracy tests
ocv_add_dispatched_file_force_all(test_intrin128 TEST SSE2 SSE3 SSSE3 SSE4_1 SSE4_2 AVX FP16 AVX2)
modules/core/include/opencv2/core/cv_cpu_dispatch.h
+1
View File
@@ -15,6 +15,7 @@
#define CV_CPU_OPTIMIZATION_NAMESPACE cpu_baseline
#define CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN namespace cpu_baseline {
#define CV_CPU_OPTIMIZATION_NAMESPACE_END }
#define CV_CPU_BASELINE_MODE 1
#endif
modules/core/src/count_non_zero.cpp → modules/core/src/count_non_zero.dispatch.cpp
+10 -184
View File
@@ -7,190 +7,18 @@
#include "opencl_kernels_core.hpp"
#include "stat.hpp"
#include "count_non_zero.simd.hpp"
#include "count_non_zero.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
namespace cv {
template<typename T>
static int countNonZero_(const T* src, int len )
{
int i=0, nz = 0;
#if CV_ENABLE_UNROLLED
for(; i <= len - 4; i += 4 )
nz += (src[i] != 0) + (src[i+1] != 0) + (src[i+2] != 0) + (src[i+3] != 0);
#endif
for( ; i < len; i++ )
nz += src[i] != 0;
return nz;
}
static int countNonZero8u( const uchar* src, int len )
{
int i=0, nz = 0;
#if CV_SIMD
int len0 = len & -v_uint8::nlanes;
v_uint8 v_zero = vx_setzero_u8();
v_uint8 v_one = vx_setall_u8(1);
v_uint32 v_sum32 = vx_setzero_u32();
while (i < len0)
{
v_uint16 v_sum16 = vx_setzero_u16();
int j = i;
while (j < std::min(len0, i + 65280 * v_uint16::nlanes))
{
v_uint8 v_sum8 = vx_setzero_u8();
int k = j;
for (; k < std::min(len0, j + 255 * v_uint8::nlanes); k += v_uint8::nlanes)
v_sum8 += v_one & (vx_load(src + k) == v_zero);
v_uint16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_uint32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
for( ; i < len; i++ )
nz += src[i] != 0;
return nz;
}
static int countNonZero16u( const ushort* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_uint16 v_zero = vx_setzero_u16();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(v_reinterpret_as_s16(vx_load(src + k) == v_zero), v_reinterpret_as_s16(vx_load(src + k + v_uint16::nlanes) == v_zero));
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero32s( const int* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_int32 v_zero = vx_setzero_s32();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(
v_pack(vx_load(src + k ) == v_zero, vx_load(src + k + v_int32::nlanes) == v_zero),
v_pack(vx_load(src + k + 2*v_int32::nlanes) == v_zero, vx_load(src + k + 3*v_int32::nlanes) == v_zero)
);
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero32f( const float* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_float32 v_zero = vx_setzero_f32();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(
v_pack(v_reinterpret_as_s32(vx_load(src + k ) == v_zero), v_reinterpret_as_s32(vx_load(src + k + v_float32::nlanes) == v_zero)),
v_pack(v_reinterpret_as_s32(vx_load(src + k + 2*v_float32::nlanes) == v_zero), v_reinterpret_as_s32(vx_load(src + k + 3*v_float32::nlanes) == v_zero))
);
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero64f( const double* src, int len )
{
return countNonZero_(src, len);
}
typedef int (*CountNonZeroFunc)(const uchar*, int);
static CountNonZeroFunc getCountNonZeroTab(int depth)
{
static CountNonZeroFunc countNonZeroTab[] =
{
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero8u), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero8u),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero16u), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero16u),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero32s), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero32f),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero64f), 0
};
return countNonZeroTab[depth];
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getCountNonZeroTab, (depth),
CV_CPU_DISPATCH_MODES_ALL);
}
#ifdef HAVE_OPENCL
static bool ocl_countNonZero( InputArray _src, int & res )
{
@@ -288,9 +116,7 @@ static bool ipp_countNonZero( Mat &src, int &res )
}
#endif
} // cv::
int cv::countNonZero( InputArray _src )
int countNonZero(InputArray _src)
{
CV_INSTRUMENT_REGION();
@@ -324,10 +150,8 @@ int cv::countNonZero( InputArray _src )
return nz;
}
void cv::findNonZero( InputArray _src, OutputArray _idx )
void findNonZero(InputArray _src, OutputArray _idx)
{
CV_INSTRUMENT_REGION();
Mat src = _src.getMat();
CV_Assert( src.channels() == 1 && src.dims == 2 );
@@ -386,3 +210,5 @@ void cv::findNonZero( InputArray _src, OutputArray _idx )
if( !idxvec.empty() )
Mat(idxvec).copyTo(_idx);
}
} // namespace
modules/core/src/count_non_zero.simd.hpp
+201
View File
@@ -0,0 +1,201 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html
#include "precomp.hpp"
namespace cv {
typedef int (*CountNonZeroFunc)(const uchar*, int);
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
CountNonZeroFunc getCountNonZeroTab(int depth);
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
template<typename T>
static int countNonZero_(const T* src, int len )
{
int i=0, nz = 0;
#if CV_ENABLE_UNROLLED
for(; i <= len - 4; i += 4 )
nz += (src[i] != 0) + (src[i+1] != 0) + (src[i+2] != 0) + (src[i+3] != 0);
#endif
for( ; i < len; i++ )
nz += src[i] != 0;
return nz;
}
static int countNonZero8u( const uchar* src, int len )
{
int i=0, nz = 0;
#if CV_SIMD
int len0 = len & -v_uint8::nlanes;
v_uint8 v_zero = vx_setzero_u8();
v_uint8 v_one = vx_setall_u8(1);
v_uint32 v_sum32 = vx_setzero_u32();
while (i < len0)
{
v_uint16 v_sum16 = vx_setzero_u16();
int j = i;
while (j < std::min(len0, i + 65280 * v_uint16::nlanes))
{
v_uint8 v_sum8 = vx_setzero_u8();
int k = j;
for (; k < std::min(len0, j + 255 * v_uint8::nlanes); k += v_uint8::nlanes)
v_sum8 += v_one & (vx_load(src + k) == v_zero);
v_uint16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_uint32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
for( ; i < len; i++ )
nz += src[i] != 0;
return nz;
}
static int countNonZero16u( const ushort* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_uint16 v_zero = vx_setzero_u16();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(v_reinterpret_as_s16(vx_load(src + k) == v_zero), v_reinterpret_as_s16(vx_load(src + k + v_uint16::nlanes) == v_zero));
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero32s( const int* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_int32 v_zero = vx_setzero_s32();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(
v_pack(vx_load(src + k ) == v_zero, vx_load(src + k + v_int32::nlanes) == v_zero),
v_pack(vx_load(src + k + 2*v_int32::nlanes) == v_zero, vx_load(src + k + 3*v_int32::nlanes) == v_zero)
);
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero32f( const float* src, int len )
{
int i = 0, nz = 0;
#if CV_SIMD
int len0 = len & -v_int8::nlanes;
v_float32 v_zero = vx_setzero_f32();
v_int8 v_one = vx_setall_s8(1);
v_int32 v_sum32 = vx_setzero_s32();
while (i < len0)
{
v_int16 v_sum16 = vx_setzero_s16();
int j = i;
while (j < std::min(len0, i + 32766 * v_int16::nlanes))
{
v_int8 v_sum8 = vx_setzero_s8();
int k = j;
for (; k < std::min(len0, j + 127 * v_int8::nlanes); k += v_int8::nlanes)
v_sum8 += v_one & v_pack(
v_pack(v_reinterpret_as_s32(vx_load(src + k ) == v_zero), v_reinterpret_as_s32(vx_load(src + k + v_float32::nlanes) == v_zero)),
v_pack(v_reinterpret_as_s32(vx_load(src + k + 2*v_float32::nlanes) == v_zero), v_reinterpret_as_s32(vx_load(src + k + 3*v_float32::nlanes) == v_zero))
);
v_int16 part1, part2;
v_expand(v_sum8, part1, part2);
v_sum16 += part1 + part2;
j = k;
}
v_int32 part1, part2;
v_expand(v_sum16, part1, part2);
v_sum32 += part1 + part2;
i = j;
}
nz = i - v_reduce_sum(v_sum32);
v_cleanup();
#endif
return nz + countNonZero_(src + i, len - i);
}
static int countNonZero64f( const double* src, int len )
{
return countNonZero_(src, len);
}
CountNonZeroFunc getCountNonZeroTab(int depth)
{
static CountNonZeroFunc countNonZeroTab[] =
{
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero8u), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero8u),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero16u), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero16u),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero32s), (CountNonZeroFunc)GET_OPTIMIZED(countNonZero32f),
(CountNonZeroFunc)GET_OPTIMIZED(countNonZero64f), 0
};
return countNonZeroTab[depth];
}
#endif
CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace
modules/core/src/matmul.dispatch.cpp
+1222
View File
File diff suppressed because it is too large Load Diff
modules/core/src/matmul.cpp → modules/core/src/matmul.simd.hpp
+176 -1146
View File
File diff suppressed because it is too large Load Diff
modules/core/src/sum.dispatch.cpp
+245
View File
@@ -0,0 +1,245 @@
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#include "stat.hpp"
#include "sum.simd.hpp"
#include "sum.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content
#undef HAVE_IPP
#undef CV_IPP_RUN_FAST
#define CV_IPP_RUN_FAST(f, ...)
#undef CV_IPP_RUN
#define CV_IPP_RUN(c, f, ...)
namespace cv
{
SumFunc getSumFunc(int depth)
{
CV_INSTRUMENT_REGION();
CV_CPU_DISPATCH(getSumFunc, (depth),
CV_CPU_DISPATCH_MODES_ALL);
}
#ifdef HAVE_OPENCL
bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask,
InputArray _src2, bool calc2, const Scalar & res2 )
{
CV_Assert(sum_op == OCL_OP_SUM || sum_op == OCL_OP_SUM_ABS || sum_op == OCL_OP_SUM_SQR);
const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0,
haveMask = _mask.kind() != _InputArray::NONE,
haveSrc2 = _src2.kind() != _InputArray::NONE;
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
kercn = cn == 1 && !haveMask ? ocl::predictOptimalVectorWidth(_src, _src2) : 1,
mcn = std::max(cn, kercn);
CV_Assert(!haveSrc2 || _src2.type() == type);
int convert_cn = haveSrc2 ? mcn : cn;
if ( (!doubleSupport && depth == CV_64F) || cn > 4 )
return false;
int ngroups = dev.maxComputeUnits(), dbsize = ngroups * (calc2 ? 2 : 1);
size_t wgs = dev.maxWorkGroupSize();
int ddepth = std::max(sum_op == OCL_OP_SUM_SQR ? CV_32F : CV_32S, depth),
dtype = CV_MAKE_TYPE(ddepth, cn);
CV_Assert(!haveMask || _mask.type() == CV_8UC1);
int wgs2_aligned = 1;
while (wgs2_aligned < (int)wgs)
wgs2_aligned <<= 1;
wgs2_aligned >>= 1;
static const char * const opMap[3] = { "OP_SUM", "OP_SUM_ABS", "OP_SUM_SQR" };
char cvt[2][40];
String opts = format("-D srcT=%s -D srcT1=%s -D dstT=%s -D dstTK=%s -D dstT1=%s -D ddepth=%d -D cn=%d"
" -D convertToDT=%s -D %s -D WGS=%d -D WGS2_ALIGNED=%d%s%s%s%s -D kercn=%d%s%s%s -D convertFromU=%s",
ocl::typeToStr(CV_MAKE_TYPE(depth, mcn)), ocl::typeToStr(depth),
ocl::typeToStr(dtype), ocl::typeToStr(CV_MAKE_TYPE(ddepth, mcn)),
ocl::typeToStr(ddepth), ddepth, cn,
ocl::convertTypeStr(depth, ddepth, mcn, cvt[0]),
opMap[sum_op], (int)wgs, wgs2_aligned,
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
haveMask ? " -D HAVE_MASK" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : "",
haveMask && _mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn,
haveSrc2 ? " -D HAVE_SRC2" : "", calc2 ? " -D OP_CALC2" : "",
haveSrc2 && _src2.isContinuous() ? " -D HAVE_SRC2_CONT" : "",
depth <= CV_32S && ddepth == CV_32S ? ocl::convertTypeStr(CV_8U, ddepth, convert_cn, cvt[1]) : "noconvert");
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
if (k.empty())
return false;
UMat src = _src.getUMat(), src2 = _src2.getUMat(),
db(1, dbsize, dtype), mask = _mask.getUMat();
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
dbarg = ocl::KernelArg::PtrWriteOnly(db),
maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
src2arg = ocl::KernelArg::ReadOnlyNoSize(src2);
if (haveMask)
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg);
}
else
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg);
}
size_t globalsize = ngroups * wgs;
if (k.run(1, &globalsize, &wgs, true))
{
typedef Scalar (*part_sum)(Mat m);
part_sum funcs[3] = { ocl_part_sum<int>, ocl_part_sum<float>, ocl_part_sum<double> },
func = funcs[ddepth - CV_32S];
Mat mres = db.getMat(ACCESS_READ);
if (calc2)
const_cast<Scalar &>(res2) = func(mres.colRange(ngroups, dbsize));
res = func(mres.colRange(0, ngroups));
return true;
}
return false;
}
#endif
#ifdef HAVE_IPP
static bool ipp_sum(Mat &src, Scalar &_res)
{
CV_INSTRUMENT_REGION_IPP();
#if IPP_VERSION_X100 >= 700
int cn = src.channels();
if (cn > 4)
return false;
size_t total_size = src.total();
int rows = src.size[0], cols = rows ? (int)(total_size/rows) : 0;
if( src.dims == 2 || (src.isContinuous() && cols > 0 && (size_t)rows*cols == total_size) )
{
IppiSize sz = { cols, rows };
int type = src.type();
typedef IppStatus (CV_STDCALL* ippiSumFuncHint)(const void*, int, IppiSize, double *, IppHintAlgorithm);
typedef IppStatus (CV_STDCALL* ippiSumFuncNoHint)(const void*, int, IppiSize, double *);
ippiSumFuncHint ippiSumHint =
type == CV_32FC1 ? (ippiSumFuncHint)ippiSum_32f_C1R :
type == CV_32FC3 ? (ippiSumFuncHint)ippiSum_32f_C3R :
type == CV_32FC4 ? (ippiSumFuncHint)ippiSum_32f_C4R :
0;
ippiSumFuncNoHint ippiSum =
type == CV_8UC1 ? (ippiSumFuncNoHint)ippiSum_8u_C1R :
type == CV_8UC3 ? (ippiSumFuncNoHint)ippiSum_8u_C3R :
type == CV_8UC4 ? (ippiSumFuncNoHint)ippiSum_8u_C4R :
type == CV_16UC1 ? (ippiSumFuncNoHint)ippiSum_16u_C1R :
type == CV_16UC3 ? (ippiSumFuncNoHint)ippiSum_16u_C3R :
type == CV_16UC4 ? (ippiSumFuncNoHint)ippiSum_16u_C4R :
type == CV_16SC1 ? (ippiSumFuncNoHint)ippiSum_16s_C1R :
type == CV_16SC3 ? (ippiSumFuncNoHint)ippiSum_16s_C3R :
type == CV_16SC4 ? (ippiSumFuncNoHint)ippiSum_16s_C4R :
0;
CV_Assert(!ippiSumHint || !ippiSum);
if( ippiSumHint || ippiSum )
{
Ipp64f res[4];
IppStatus ret = ippiSumHint ?
CV_INSTRUMENT_FUN_IPP(ippiSumHint, src.ptr(), (int)src.step[0], sz, res, ippAlgHintAccurate) :
CV_INSTRUMENT_FUN_IPP(ippiSum, src.ptr(), (int)src.step[0], sz, res);
if( ret >= 0 )
{
for( int i = 0; i < cn; i++ )
_res[i] = res[i];
return true;
}
}
}
#else
CV_UNUSED(src); CV_UNUSED(_res);
#endif
return false;
}
#endif
Scalar sum(InputArray _src)
{
CV_INSTRUMENT_REGION();
#if defined HAVE_OPENCL || defined HAVE_IPP
Scalar _res;
#endif
#ifdef HAVE_OPENCL
CV_OCL_RUN_(OCL_PERFORMANCE_CHECK(_src.isUMat()) && _src.dims() <= 2,
ocl_sum(_src, _res, OCL_OP_SUM),
_res)
#endif
Mat src = _src.getMat();
CV_IPP_RUN(IPP_VERSION_X100 >= 700, ipp_sum(src, _res), _res);
int k, cn = src.channels(), depth = src.depth();
SumFunc func = getSumFunc(depth);
CV_Assert( cn <= 4 && func != 0 );
const Mat* arrays[] = {&src, 0};
uchar* ptrs[1] = {};
NAryMatIterator it(arrays, ptrs);
Scalar s;
int total = (int)it.size, blockSize = total, intSumBlockSize = 0;
int j, count = 0;
AutoBuffer<int> _buf;
int* buf = (int*)&s[0];
size_t esz = 0;
bool blockSum = depth < CV_32S;
if( blockSum )
{
intSumBlockSize = depth <= CV_8S ? (1 << 23) : (1 << 15);
blockSize = std::min(blockSize, intSumBlockSize);
_buf.allocate(cn);
buf = _buf.data();
for( k = 0; k < cn; k++ )
buf[k] = 0;
esz = src.elemSize();
}
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
for( j = 0; j < total; j += blockSize )
{
int bsz = std::min(total - j, blockSize);
func( ptrs[0], 0, (uchar*)buf, bsz, cn );
count += bsz;
if( blockSum && (count + blockSize >= intSumBlockSize || (i+1 >= it.nplanes && j+bsz >= total)) )
{
for( k = 0; k < cn; k++ )
{
s[k] += buf[k];
buf[k] = 0;
}
count = 0;
}
ptrs[0] += bsz*esz;
}
}
return s;
}
} // namespace
modules/core/src/sum.cpp → modules/core/src/sum.simd.hpp
+14 -230
View File
@@ -4,17 +4,14 @@
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#include "stat.hpp"
#undef HAVE_IPP
#undef CV_IPP_RUN_FAST
#define CV_IPP_RUN_FAST(f, ...)
#undef CV_IPP_RUN
#define CV_IPP_RUN(c, f, ...)
namespace cv {
CV_CPU_OPTIMIZATION_NAMESPACE_BEGIN
namespace cv
{
SumFunc getSumFunc(int depth);
#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
template <typename T, typename ST>
struct Sum_SIMD
@@ -415,25 +412,25 @@ static int sum_(const T* src0, const uchar* mask, ST* dst, int len, int cn )
static int sum8u( const uchar* src, const uchar* mask, int* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum8s( const schar* src, const uchar* mask, int* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum16u( const ushort* src, const uchar* mask, int* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum16s( const short* src, const uchar* mask, int* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum32s( const int* src, const uchar* mask, double* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum32f( const float* src, const uchar* mask, double* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
static int sum64f( const double* src, const uchar* mask, double* dst, int len, int cn )
{ return sum_(src, mask, dst, len, cn); }
{ CV_INSTRUMENT_REGION(); return sum_(src, mask, dst, len, cn); }
SumFunc getSumFunc(int depth)
{
@@ -449,220 +446,7 @@ SumFunc getSumFunc(int depth)
return sumTab[depth];
}
#ifdef HAVE_OPENCL
bool ocl_sum( InputArray _src, Scalar & res, int sum_op, InputArray _mask,
InputArray _src2, bool calc2, const Scalar & res2 )
{
CV_Assert(sum_op == OCL_OP_SUM || sum_op == OCL_OP_SUM_ABS || sum_op == OCL_OP_SUM_SQR);
const ocl::Device & dev = ocl::Device::getDefault();
bool doubleSupport = dev.doubleFPConfig() > 0,
haveMask = _mask.kind() != _InputArray::NONE,
haveSrc2 = _src2.kind() != _InputArray::NONE;
int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type),
kercn = cn == 1 && !haveMask ? ocl::predictOptimalVectorWidth(_src, _src2) : 1,
mcn = std::max(cn, kercn);
CV_Assert(!haveSrc2 || _src2.type() == type);
int convert_cn = haveSrc2 ? mcn : cn;
if ( (!doubleSupport && depth == CV_64F) || cn > 4 )
return false;
int ngroups = dev.maxComputeUnits(), dbsize = ngroups * (calc2 ? 2 : 1);
size_t wgs = dev.maxWorkGroupSize();
int ddepth = std::max(sum_op == OCL_OP_SUM_SQR ? CV_32F : CV_32S, depth),
dtype = CV_MAKE_TYPE(ddepth, cn);
CV_Assert(!haveMask || _mask.type() == CV_8UC1);
int wgs2_aligned = 1;
while (wgs2_aligned < (int)wgs)
wgs2_aligned <<= 1;
wgs2_aligned >>= 1;
static const char * const opMap[3] = { "OP_SUM", "OP_SUM_ABS", "OP_SUM_SQR" };
char cvt[2][40];
String opts = format("-D srcT=%s -D srcT1=%s -D dstT=%s -D dstTK=%s -D dstT1=%s -D ddepth=%d -D cn=%d"
" -D convertToDT=%s -D %s -D WGS=%d -D WGS2_ALIGNED=%d%s%s%s%s -D kercn=%d%s%s%s -D convertFromU=%s",
ocl::typeToStr(CV_MAKE_TYPE(depth, mcn)), ocl::typeToStr(depth),
ocl::typeToStr(dtype), ocl::typeToStr(CV_MAKE_TYPE(ddepth, mcn)),
ocl::typeToStr(ddepth), ddepth, cn,
ocl::convertTypeStr(depth, ddepth, mcn, cvt[0]),
opMap[sum_op], (int)wgs, wgs2_aligned,
doubleSupport ? " -D DOUBLE_SUPPORT" : "",
haveMask ? " -D HAVE_MASK" : "",
_src.isContinuous() ? " -D HAVE_SRC_CONT" : "",
haveMask && _mask.isContinuous() ? " -D HAVE_MASK_CONT" : "", kercn,
haveSrc2 ? " -D HAVE_SRC2" : "", calc2 ? " -D OP_CALC2" : "",
haveSrc2 && _src2.isContinuous() ? " -D HAVE_SRC2_CONT" : "",
depth <= CV_32S && ddepth == CV_32S ? ocl::convertTypeStr(CV_8U, ddepth, convert_cn, cvt[1]) : "noconvert");
ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
if (k.empty())
return false;
UMat src = _src.getUMat(), src2 = _src2.getUMat(),
db(1, dbsize, dtype), mask = _mask.getUMat();
ocl::KernelArg srcarg = ocl::KernelArg::ReadOnlyNoSize(src),
dbarg = ocl::KernelArg::PtrWriteOnly(db),
maskarg = ocl::KernelArg::ReadOnlyNoSize(mask),
src2arg = ocl::KernelArg::ReadOnlyNoSize(src2);
if (haveMask)
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, maskarg);
}
else
{
if (haveSrc2)
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg, src2arg);
else
k.args(srcarg, src.cols, (int)src.total(), ngroups, dbarg);
}
size_t globalsize = ngroups * wgs;
if (k.run(1, &globalsize, &wgs, true))
{
typedef Scalar (*part_sum)(Mat m);
part_sum funcs[3] = { ocl_part_sum<int>, ocl_part_sum<float>, ocl_part_sum<double> },
func = funcs[ddepth - CV_32S];
Mat mres = db.getMat(ACCESS_READ);
if (calc2)
const_cast<Scalar &>(res2) = func(mres.colRange(ngroups, dbsize));
res = func(mres.colRange(0, ngroups));
return true;
}
return false;
}
#endif
#ifdef HAVE_IPP
static bool ipp_sum(Mat &src, Scalar &_res)
{
CV_INSTRUMENT_REGION_IPP();
#if IPP_VERSION_X100 >= 700
int cn = src.channels();
if (cn > 4)
return false;
size_t total_size = src.total();
int rows = src.size[0], cols = rows ? (int)(total_size/rows) : 0;
if( src.dims == 2 || (src.isContinuous() && cols > 0 && (size_t)rows*cols == total_size) )
{
IppiSize sz = { cols, rows };
int type = src.type();
typedef IppStatus (CV_STDCALL* ippiSumFuncHint)(const void*, int, IppiSize, double *, IppHintAlgorithm);
typedef IppStatus (CV_STDCALL* ippiSumFuncNoHint)(const void*, int, IppiSize, double *);
ippiSumFuncHint ippiSumHint =
type == CV_32FC1 ? (ippiSumFuncHint)ippiSum_32f_C1R :
type == CV_32FC3 ? (ippiSumFuncHint)ippiSum_32f_C3R :
type == CV_32FC4 ? (ippiSumFuncHint)ippiSum_32f_C4R :
0;
ippiSumFuncNoHint ippiSum =
type == CV_8UC1 ? (ippiSumFuncNoHint)ippiSum_8u_C1R :
type == CV_8UC3 ? (ippiSumFuncNoHint)ippiSum_8u_C3R :
type == CV_8UC4 ? (ippiSumFuncNoHint)ippiSum_8u_C4R :
type == CV_16UC1 ? (ippiSumFuncNoHint)ippiSum_16u_C1R :
type == CV_16UC3 ? (ippiSumFuncNoHint)ippiSum_16u_C3R :
type == CV_16UC4 ? (ippiSumFuncNoHint)ippiSum_16u_C4R :
type == CV_16SC1 ? (ippiSumFuncNoHint)ippiSum_16s_C1R :
type == CV_16SC3 ? (ippiSumFuncNoHint)ippiSum_16s_C3R :
type == CV_16SC4 ? (ippiSumFuncNoHint)ippiSum_16s_C4R :
0;
CV_Assert(!ippiSumHint || !ippiSum);
if( ippiSumHint || ippiSum )
{
Ipp64f res[4];
IppStatus ret = ippiSumHint ?
CV_INSTRUMENT_FUN_IPP(ippiSumHint, src.ptr(), (int)src.step[0], sz, res, ippAlgHintAccurate) :
CV_INSTRUMENT_FUN_IPP(ippiSum, src.ptr(), (int)src.step[0], sz, res);
if( ret >= 0 )
{
for( int i = 0; i < cn; i++ )
_res[i] = res[i];
return true;
}
}
}
#else
CV_UNUSED(src); CV_UNUSED(_res);
#endif
return false;
}
#endif
} // cv::
cv::Scalar cv::sum( InputArray _src )
{
CV_INSTRUMENT_REGION();
#if defined HAVE_OPENCL || defined HAVE_IPP
Scalar _res;
#endif
#ifdef HAVE_OPENCL
CV_OCL_RUN_(OCL_PERFORMANCE_CHECK(_src.isUMat()) && _src.dims() <= 2,
ocl_sum(_src, _res, OCL_OP_SUM),
_res)
#endif
Mat src = _src.getMat();
CV_IPP_RUN(IPP_VERSION_X100 >= 700, ipp_sum(src, _res), _res);
int k, cn = src.channels(), depth = src.depth();
SumFunc func = getSumFunc(depth);
CV_Assert( cn <= 4 && func != 0 );
const Mat* arrays[] = {&src, 0};
uchar* ptrs[1] = {};
NAryMatIterator it(arrays, ptrs);
Scalar s;
int total = (int)it.size, blockSize = total, intSumBlockSize = 0;
int j, count = 0;
AutoBuffer<int> _buf;
int* buf = (int*)&s[0];
size_t esz = 0;
bool blockSum = depth < CV_32S;
if( blockSum )
{
intSumBlockSize = depth <= CV_8S ? (1 << 23) : (1 << 15);
blockSize = std::min(blockSize, intSumBlockSize);
_buf.allocate(cn);
buf = _buf.data();
for( k = 0; k < cn; k++ )
buf[k] = 0;
esz = src.elemSize();
}
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
for( j = 0; j < total; j += blockSize )
{
int bsz = std::min(total - j, blockSize);
func( ptrs[0], 0, (uchar*)buf, bsz, cn );
count += bsz;
if( blockSum && (count + blockSize >= intSumBlockSize || (i+1 >= it.nplanes && j+bsz >= total)) )
{
for( k = 0; k < cn; k++ )
{
s[k] += buf[k];
buf[k] = 0;
}
count = 0;
}
ptrs[0] += bsz*esz;
}
}
return s;
}
CV_CPU_OPTIMIZATION_NAMESPACE_END
} // namespace
modules/dnn/perf/perf_net.cpp
+1 -2
View File
@@ -124,8 +124,7 @@ PERF_TEST_P_(DNNTestNetwork, SSD)
PERF_TEST_P_(DNNTestNetwork, OpenFace)
{
if (backend == DNN_BACKEND_HALIDE ||
(backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
(backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
(backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
throw SkipTestException("");
processNet("dnn/openface_nn4.small2.v1.t7", "", "",
Mat(cv::Size(96, 96), CV_32FC3));
modules/dnn/src/dnn.cpp
+5 -15
View File
@@ -736,9 +736,9 @@ struct DataLayer : public Layer
biases->set(biasesVec);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ScaleShiftLayer ieLayer(name);
ieLayer.setWeights(weights);
ieLayer.setBiases(biases);
InferenceEngine::Builder::Layer ieLayer = InferenceEngine::Builder::ScaleShiftLayer(name);
addConstantData("weights", weights, ieLayer);
addConstantData("biases", biases, ieLayer);
#else
InferenceEngine::LayerParams lp;
lp.name = name;
@@ -1701,25 +1701,15 @@ struct Net::Impl
preferableTarget == DNN_TARGET_FPGA) && !fused)
{
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
bool hasWeights = false;
for (const std::string& name : {"weights", "biases"})
{
auto it = ieNode->layer.getParameters().find(name);
if (it != ieNode->layer.getParameters().end())
{
InferenceEngine::Blob::CPtr bp = it->second.as<InferenceEngine::Blob::CPtr>();
it->second = (InferenceEngine::Blob::CPtr)convertFp16(std::const_pointer_cast<InferenceEngine::Blob>(bp));
hasWeights = true;
InferenceEngine::Blob::Ptr bp = it->second.as<InferenceEngine::Blob::Ptr>();
it->second = convertFp16(std::const_pointer_cast<InferenceEngine::Blob>(bp));
}
}
if (!hasWeights)
{
InferenceEngine::Blob::Ptr blob = InferenceEngine::make_shared_blob<int16_t>(
InferenceEngine::Precision::FP16,
InferenceEngine::Layout::C, {1});
blob->allocate();
ieNode->layer.getParameters()["weights"] = (InferenceEngine::Blob::CPtr)blob;
}
#else
auto& blobs = ieNode->layer.getConstantData();
if (blobs.empty())
modules/dnn/src/layers/batch_norm_layer.cpp
+3 -4
View File
@@ -350,11 +350,10 @@ public:
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ScaleShiftLayer ieLayer(name);
InferenceEngine::Builder::Layer ieLayer = InferenceEngine::Builder::ScaleShiftLayer(name);
const size_t numChannels = weights_.total();
ieLayer.setWeights(wrapToInfEngineBlob(weights_, {numChannels}, InferenceEngine::Layout::C));
ieLayer.setBiases(wrapToInfEngineBlob(bias_, {numChannels}, InferenceEngine::Layout::C));
addConstantData("weights", wrapToInfEngineBlob(weights_, {numChannels}, InferenceEngine::Layout::C), ieLayer);
addConstantData("biases", wrapToInfEngineBlob(bias_, {numChannels}, InferenceEngine::Layout::C), ieLayer);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
InferenceEngine::LayerParams lp;
modules/dnn/src/layers/blank_layer.cpp
+3 -1
View File
@@ -125,7 +125,9 @@ public:
ieLayer.getParameters()["axis"] = input->dims.size() - 1;
ieLayer.getParameters()["out_sizes"] = input->dims[0];
}
ieLayer.setInputPorts(std::vector<InferenceEngine::Port>(1));
std::vector<size_t> shape(input->dims);
std::reverse(shape.begin(), shape.end());
ieLayer.setInputPorts({InferenceEngine::Port(shape)});
ieLayer.setOutputPorts(std::vector<InferenceEngine::Port>(1));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
#else
modules/dnn/src/layers/convolution_layer.cpp
+8 -9
View File
@@ -563,11 +563,11 @@ public:
ieLayer.setGroup((size_t)group);
ieLayer.setOutDepth((size_t)outCn);
ieLayer.setWeights(ieWeights);
if (ieBiases)
ieLayer.setBiases(ieBiases);
InferenceEngine::Builder::Layer l = ieLayer;
addConstantData("weights", ieWeights, l);
if (ieBiases)
addConstantData("biases", ieBiases, l);
if (!padMode.empty())
l.getParameters()["auto_pad"] = padMode == "VALID" ? std::string("valid") : std::string("same_upper");
@@ -1795,12 +1795,11 @@ public:
ieLayer.setGroup((size_t)group);
ieLayer.setOutDepth((size_t)numOutput);
ieLayer.setWeights(wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW));
InferenceEngine::Builder::Layer l = ieLayer;
addConstantData("weights", wrapToInfEngineBlob(blobs[0], InferenceEngine::Layout::OIHW), l);
if (hasBias())
{
ieLayer.setBiases(wrapToInfEngineBlob(blobs[1], {(size_t)numOutput}, InferenceEngine::Layout::C));
}
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
addConstantData("biases", wrapToInfEngineBlob(blobs[1], {(size_t)numOutput}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
const int outGroupCn = blobs[0].size[1]; // Weights are in IOHW layout
const int group = numOutput / outGroupCn;
modules/dnn/src/layers/elementwise_layers.cpp
+3 -3
View File
@@ -1210,10 +1210,10 @@ struct ChannelsPReLUFunctor
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::Layer initInfEngineBuilderAPI()
{
InferenceEngine::Builder::PReLULayer ieLayer("");
InferenceEngine::Builder::Layer l = InferenceEngine::Builder::PReLULayer("");
const size_t numChannels = scale.total();
ieLayer.setWeights(wrapToInfEngineBlob(scale, {numChannels}, InferenceEngine::Layout::C));
return ieLayer;
addConstantData("weights", wrapToInfEngineBlob(scale, {numChannels}, InferenceEngine::Layout::C), l);
return l;
}
#else
InferenceEngine::CNNLayerPtr initInfEngine(InferenceEngine::LayerParams& lp)
modules/dnn/src/layers/fully_connected_layer.cpp
+4 -3
View File
@@ -448,11 +448,12 @@ public:
const int outNum = blobs[0].size[0];
ieLayer.setOutputNum(outNum);
ieLayer.setWeights(wrapToInfEngineBlob(blobs[0], {(size_t)blobs[0].size[0], (size_t)blobs[0].size[1], 1, 1}, InferenceEngine::Layout::OIHW));
InferenceEngine::Builder::Layer l = ieLayer;
addConstantData("weights", wrapToInfEngineBlob(blobs[0], {(size_t)blobs[0].size[0], (size_t)blobs[0].size[1], 1, 1}, InferenceEngine::Layout::OIHW), l);
if (blobs.size() > 1)
ieLayer.setBiases(wrapToInfEngineBlob(blobs[1], {(size_t)outNum}, InferenceEngine::Layout::C));
addConstantData("biases", wrapToInfEngineBlob(blobs[1], {(size_t)outNum}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
modules/dnn/src/layers/lrn_layer.cpp
+1 -1
View File
@@ -93,7 +93,7 @@ public:
{
return backendId == DNN_BACKEND_OPENCV ||
backendId == DNN_BACKEND_HALIDE ||
(backendId == DNN_BACKEND_INFERENCE_ENGINE && (preferableTarget != DNN_TARGET_MYRIAD || type == CHANNEL_NRM)) ||
backendId == DNN_BACKEND_INFERENCE_ENGINE ||
(backendId == DNN_BACKEND_VKCOM && haveVulkan() && (size % 2 == 1) && (type == CHANNEL_NRM));
}
modules/dnn/src/layers/normalize_bbox_layer.cpp
+4 -8
View File
@@ -67,14 +67,10 @@ public:
{
if (pnorm != 2)
return false;
if (!blobs.empty())
return true;
if (preferableTarget == DNN_TARGET_MYRIAD)
return !acrossSpatial;
return startAxis == 1 && (!acrossSpatial || endAxis > 1);
return preferableTarget == DNN_TARGET_MYRIAD ? !acrossSpatial : startAxis == 1;
}
else
return backendId == DNN_BACKEND_OPENCV;
return backendId == DNN_BACKEND_OPENCV;
}
bool getMemoryShapes(const std::vector<MatShape> &inputs,
@@ -295,7 +291,7 @@ public:
l.getParameters()["channel_shared"] = blobs[0].total() == 1;
}
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
l.getParameters()["weights"] = (InferenceEngine::Blob::CPtr)weights;
l.getParameters()["weights"] = weights;
#else
l.addConstantData("weights", weights);
#endif
modules/dnn/src/layers/prior_box_layer.cpp
+3 -3
View File
@@ -539,12 +539,12 @@ public:
if (_stepX == _stepY)
{
l.getParameters()["step"] = _stepX;
l.getParameters()["step_h"] = 0.0;
l.getParameters()["step_w"] = 0.0;
l.getParameters()["step_h"] = 0.0f;
l.getParameters()["step_w"] = 0.0f;
}
else
{
l.getParameters()["step"] = 0.0;
l.getParameters()["step"] = 0.0f;
l.getParameters()["step_h"] = _stepY;
l.getParameters()["step_w"] = _stepX;
}
modules/dnn/src/layers/resize_layer.cpp
+2 -3
View File
@@ -54,12 +54,11 @@ public:
#ifdef HAVE_INF_ENGINE
if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
{
return (interpolation == "nearest" && preferableTarget != DNN_TARGET_MYRIAD) ||
return (interpolation == "nearest" && scaleWidth == scaleHeight) ||
(interpolation == "bilinear" && INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R4));
}
else
#endif
return backendId == DNN_BACKEND_OPENCV;
return backendId == DNN_BACKEND_OPENCV;
}
virtual void finalize(InputArrayOfArrays inputs_arr, OutputArrayOfArrays outputs_arr) CV_OVERRIDE
modules/dnn/src/layers/scale_layer.cpp
+5 -5
View File
@@ -198,13 +198,13 @@ public:
{
#ifdef HAVE_INF_ENGINE
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InferenceEngine::Builder::ScaleShiftLayer ieLayer(name);
InferenceEngine::Builder::Layer l = InferenceEngine::Builder::ScaleShiftLayer(name);
CV_Assert(!blobs.empty());
const size_t numChannels = blobs[0].total();
if (hasWeights)
{
ieLayer.setWeights(wrapToInfEngineBlob(blobs[0], {numChannels}, InferenceEngine::Layout::C));
addConstantData("weights", wrapToInfEngineBlob(blobs[0], {numChannels}, InferenceEngine::Layout::C), l);
}
else
{
@@ -214,11 +214,11 @@ public:
std::vector<float> ones(numChannels, 1);
weights->set(ones);
ieLayer.setWeights(weights);
addConstantData("weights", weights, l);
}
if (hasBias)
ieLayer.setBiases(wrapToInfEngineBlob(blobs.back(), {numChannels}, InferenceEngine::Layout::C));
return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
addConstantData("biases", wrapToInfEngineBlob(blobs.back(), {numChannels}, InferenceEngine::Layout::C), l);
return Ptr<BackendNode>(new InfEngineBackendNode(l));
#else
InferenceEngine::LayerParams lp;
lp.name = name;
modules/dnn/src/op_inf_engine.cpp
+72 -3
View File
@@ -18,6 +18,11 @@ namespace cv { namespace dnn {
#ifdef HAVE_INF_ENGINE
// For networks with input layer which has an empty name, IE generates a name id[some_number].
// OpenCV lets users use an empty input name and to prevent unexpected naming,
// we can use some predefined name.
static std::string kDefaultInpLayerName = "empty_inp_layer_name";
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
InfEngineBackendNode::InfEngineBackendNode(const InferenceEngine::Builder::Layer& _layer)
: BackendNode(DNN_BACKEND_INFERENCE_ENGINE), layer(_layer) {}
@@ -90,7 +95,7 @@ void InfEngineBackendNet::connect(const std::vector<Ptr<BackendWrapper> >& input
it = layers.find(inpName);
if (it == layers.end())
{
InferenceEngine::Builder::InputLayer inpLayer(inpName);
InferenceEngine::Builder::InputLayer inpLayer(!inpName.empty() ? inpName : kDefaultInpLayerName);
std::vector<size_t> shape(inp->blob->dims());
std::reverse(shape.begin(), shape.end());
@@ -119,6 +124,14 @@ void InfEngineBackendNet::init(int targetId)
for (int id : unconnectedLayersIds)
{
InferenceEngine::Builder::OutputLayer outLayer("myconv1");
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
// Inference Engine determines network precision by ports.
InferenceEngine::Precision p = (targetId == DNN_TARGET_MYRIAD ||
targetId == DNN_TARGET_OPENCL_FP16) ?
InferenceEngine::Precision::FP16 :
InferenceEngine::Precision::FP32;
outLayer.setPort(InferenceEngine::Port({}, p));
#endif
netBuilder.addLayer({InferenceEngine::PortInfo(id)}, outLayer);
}
cnn = InferenceEngine::CNNNetwork(InferenceEngine::Builder::convertToICNNNetwork(netBuilder.build()));
@@ -167,12 +180,56 @@ void InfEngineBackendNet::init(int targetId)
initPlugin(cnn);
}
void InfEngineBackendNet::addLayer(const InferenceEngine::Builder::Layer& layer)
void InfEngineBackendNet::addLayer(InferenceEngine::Builder::Layer& layer)
{
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
// Add weights to network and connect them after input blobs.
std::map<std::string, InferenceEngine::Parameter>& params = layer.getParameters();
std::vector<int> blobsIds;
std::vector<int> portIds;
for (const std::string& name : {"weights", "biases"})
{
bool asInput = false;
int portId = 0;
for (int i = 0; i < layer.getInputPorts().size(); ++i)
{
const auto& port = layer.getInputPorts()[i];
auto it = port.getParameters().find("type");
if (it != port.getParameters().end() && it->second == name)
{
portId = i;
asInput = true;
break;
}
}
if (!asInput)
continue;
auto it = params.find(name);
if (it != params.end())
{
InferenceEngine::Blob::Ptr blob = it->second.as<InferenceEngine::Blob::Ptr>();
params.erase(it);
int blobId = netBuilder.addLayer(InferenceEngine::Builder::ConstLayer(name).setData(blob));
blobsIds.push_back(blobId);
portIds.push_back(portId);
}
}
#endif
int id = netBuilder.addLayer(layer);
const std::string& layerName = layer.getName();
CV_Assert(layers.insert({layerName, id}).second);
unconnectedLayersIds.insert(id);
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
// By default, all the weights are connected to last ports ids.
for (int i = 0; i < blobsIds.size(); ++i)
{
netBuilder.connect((size_t)blobsIds[i], {(size_t)id, portIds[i]});
}
#endif
}
void InfEngineBackendNet::addOutput(const std::string& name)
@@ -705,7 +762,7 @@ void InfEngineBackendNet::addBlobs(const std::vector<Ptr<BackendWrapper> >& ptrs
{
std::string name = wrapper->dataPtr->name;
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
name = name.empty() ? "id1" : name; // TODO: drop the magic input name.
name = name.empty() ? kDefaultInpLayerName : name;
#endif
allBlobs.insert({name, wrapper->blob});
}
@@ -776,6 +833,18 @@ InferenceEngine::Blob::Ptr convertFp16(const InferenceEngine::Blob::Ptr& blob)
return halfs;
}
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
void addConstantData(const std::string& name, InferenceEngine::Blob::Ptr data,
InferenceEngine::Builder::Layer& l)
{
#if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2018R5)
l.getParameters()[name] = data;
#else
l.addConstantData(name, data);
#endif
}
#endif
#endif // HAVE_INF_ENGINE
bool haveInfEngine()
modules/dnn/src/op_inf_engine.hpp
+5 -1
View File
@@ -162,7 +162,7 @@ public:
InfEngineBackendNet(InferenceEngine::CNNNetwork& net);
void addLayer(const InferenceEngine::Builder::Layer& layer);
void addLayer(InferenceEngine::Builder::Layer& layer);
void addOutput(const std::string& name);
@@ -255,6 +255,10 @@ Mat infEngineBlobToMat(const InferenceEngine::Blob::Ptr& blob);
// Allocates memory for a new blob.
InferenceEngine::Blob::Ptr convertFp16(const InferenceEngine::Blob::Ptr& blob);
#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2018R5)
void addConstantData(const std::string& name, InferenceEngine::Blob::Ptr data, InferenceEngine::Builder::Layer& l);
#endif
// This is a fake class to run networks from Model Optimizer. Objects of that
// class simulate responses of layers are imported by OpenCV and supported by
// Inference Engine. The main difference is that they do not perform forward pass.
modules/dnn/test/test_halide_layers.cpp
+2 -1
View File
@@ -695,7 +695,8 @@ TEST_P(Eltwise, Accuracy)
Target targetId = get<1>(get<4>(GetParam()));
#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE > 2018050000
if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_OPENCL)
if (backendId == DNN_BACKEND_INFERENCE_ENGINE &&
(targetId == DNN_TARGET_OPENCL || targetId == DNN_TARGET_OPENCL_FP16))
throw SkipTestException("");
#endif
modules/dnn/test/test_tf_importer.cpp
-2
View File
@@ -558,8 +558,6 @@ TEST_P(Test_TensorFlow_layers, split)
TEST_P(Test_TensorFlow_layers, resize_nearest_neighbor)
{
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_MYRIAD)
throw SkipTestException("");
runTensorFlowNet("resize_nearest_neighbor");
runTensorFlowNet("keras_upsampling2d");
}
modules/dnn/test/test_torch_importer.cpp
-2
View File
@@ -277,8 +277,6 @@ TEST_P(Test_Torch_nets, OpenFace_accuracy)
throw SkipTestException("");
#endif
checkBackend();
if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
throw SkipTestException("");
const string model = findDataFile("dnn/openface_nn4.small2.v1.t7", false);
Net net = readNetFromTorch(model);
modules/flann/include/opencv2/flann/dist.h
+2 -2
View File
@@ -47,7 +47,7 @@ typedef unsigned __int64 uint64_t;
# include <Intrin.h>
#endif
#ifdef __ARM_NEON__
#if defined(__ARM_NEON__) && !defined(__CUDACC__)
# include "arm_neon.h"
#endif
@@ -425,7 +425,7 @@ struct Hamming
ResultType operator()(Iterator1 a, Iterator2 b, size_t size, ResultType /*worst_dist*/ = -1) const
{
ResultType result = 0;
#ifdef __ARM_NEON__
#if defined(__ARM_NEON__) && !defined(__CUDACC__)
{
uint32x4_t bits = vmovq_n_u32(0);
for (size_t i = 0; i < size; i += 16) {
modules/photo/include/opencv2/photo.hpp
+2 -2
View File
@@ -328,8 +328,8 @@ class CV_EXPORTS_W Tonemap : public Algorithm
public:
/** @brief Tonemaps image
@param src source image - 32-bit 3-channel Mat
@param dst destination image - 32-bit 3-channel Mat with values in [0, 1] range
@param src source image - CV_32FC3 Mat (float 32 bits 3 channels)
@param dst destination image - CV_32FC3 Mat with values in [0, 1] range
*/
CV_WRAP virtual void process(InputArray src, OutputArray dst) = 0;
modules/photo/src/tonemap.cpp
+1
View File
@@ -66,6 +66,7 @@ public:
Mat src = _src.getMat();
CV_Assert(!src.empty());
CV_Assert(_src.dims() == 2 && _src.type() == CV_32FC3);
_dst.create(src.size(), CV_32FC3);
Mat dst = _dst.getMat();