Skip to content

O
opencv
Commit 7688a182 authored by Alexander Alekhin's avatar Alexander Alekhin
Browse files
Options

Merge pull request #2840 from ilya-lavrenov:tapi_calchist

parents acebfcd2 33239fca
Hide whitespace changes
Inline Side-by-side
Showing with 147 additions and 62 deletions
modules/imgproc/src/histogram.cpp
View file @ 7688a182
...@@ -1477,42 +1477,44 @@ enum ...@@ -1477,42 +1477,44 @@ enum
static bool ocl_calcHist1(InputArray _src, OutputArray _hist, int ddepth = CV_32S) static bool ocl_calcHist1(InputArray _src, OutputArray _hist, int ddepth = CV_32S)
{ {
int compunits = ocl::Device::getDefault().maxComputeUnits(); const ocl::Device & dev = ocl::Device::getDefault();
size_t wgs = ocl::Device::getDefault().maxWorkGroupSize(); int compunits = dev.maxComputeUnits();
size_t wgs = dev.maxWorkGroupSize();
Size size = _src.size(); Size size = _src.size();
bool use16 = size.width % 16 == 0 && _src.offset() % 16 == 0 && _src.step() % 16 == 0; bool use16 = size.width % 16 == 0 && _src.offset() % 16 == 0 && _src.step() % 16 == 0;
int kercn = dev.isAMD() && use16 ? 16 : std::min(4, ocl::predictOptimalVectorWidth(_src));
ocl::Kernel k1("calculate_histogram", ocl::imgproc::histogram_oclsrc, ocl::Kernel k1("calculate_histogram", ocl::imgproc::histogram_oclsrc,
format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d -D cn=%d", format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d -D kercn=%d -D T=%s%s",
BINS, compunits, wgs, use16 ? 16 : 1)); BINS, compunits, wgs, kercn,
kercn == 4 ? "int" : ocl::typeToStr(CV_8UC(kercn)),
_src.isContinuous() ? " -D HAVE_SRC_CONT" : ""));
if (k1.empty()) if (k1.empty())
return false; return false;
_hist.create(BINS, 1, ddepth); _hist.create(BINS, 1, ddepth);
UMat src = _src.getUMat(), ghist(1, BINS * compunits, CV_32SC1), UMat src = _src.getUMat(), ghist(1, BINS * compunits, CV_32SC1),
hist = ddepth == CV_32S ? _hist.getUMat() : UMat(BINS, 1, CV_32SC1); hist = _hist.getUMat();
k1.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::PtrWriteOnly(ghist), (int)src.total()); k1.args(ocl::KernelArg::ReadOnly(src),
ocl::KernelArg::PtrWriteOnly(ghist), (int)src.total());
size_t globalsize = compunits * wgs; size_t globalsize = compunits * wgs;
if (!k1.run(1, &globalsize, &wgs, false)) if (!k1.run(1, &globalsize, &wgs, false))
return false; return false;
char cvt[40];
ocl::Kernel k2("merge_histogram", ocl::imgproc::histogram_oclsrc, ocl::Kernel k2("merge_histogram", ocl::imgproc::histogram_oclsrc,
format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d", BINS, compunits, (int)wgs)); format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d -D convertToHT=%s -D HT=%s",
BINS, compunits, (int)wgs, ocl::convertTypeStr(CV_32S, ddepth, 1, cvt),
ocl::typeToStr(ddepth)));
if (k2.empty()) if (k2.empty())
return false; return false;
k2.args(ocl::KernelArg::PtrReadOnly(ghist), ocl::KernelArg::PtrWriteOnly(hist)); k2.args(ocl::KernelArg::PtrReadOnly(ghist),
if (!k2.run(1, &wgs, &wgs, false)) ocl::KernelArg::WriteOnlyNoSize(hist));
return false;
if (hist.depth() != ddepth)
hist.convertTo(_hist, ddepth);
else
_hist.getUMatRef() = hist;
return true; return k2.run(1, &wgs, &wgs, false);
} }
static bool ocl_calcHist(InputArrayOfArrays images, OutputArray hist) static bool ocl_calcHist(InputArrayOfArrays images, OutputArray hist)
...@@ -3428,24 +3430,40 @@ namespace cv { ...@@ -3428,24 +3430,40 @@ namespace cv {
static bool ocl_equalizeHist(InputArray _src, OutputArray _dst) static bool ocl_equalizeHist(InputArray _src, OutputArray _dst)
{ {
size_t wgs = std::min<size_t>(ocl::Device::getDefault().maxWorkGroupSize(), BINS); const ocl::Device & dev = ocl::Device::getDefault();
int compunits = dev.maxComputeUnits();
size_t wgs = dev.maxWorkGroupSize();
Size size = _src.size();
bool use16 = size.width % 16 == 0 && _src.offset() % 16 == 0 && _src.step() % 16 == 0;
int kercn = dev.isAMD() && use16 ? 16 : std::min(4, ocl::predictOptimalVectorWidth(_src));
// calculation of histogram ocl::Kernel k1("calculate_histogram", ocl::imgproc::histogram_oclsrc,
UMat hist; format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d -D kercn=%d -D T=%s%s",
if (!ocl_calcHist1(_src, hist)) BINS, compunits, wgs, kercn,
kercn == 4 ? "int" : ocl::typeToStr(CV_8UC(kercn)),
_src.isContinuous() ? " -D HAVE_SRC_CONT" : ""));
if (k1.empty())
return false; return false;
UMat lut(1, 256, CV_8UC1); UMat src = _src.getUMat(), ghist(1, BINS * compunits, CV_32SC1);
ocl::Kernel k("calcLUT", ocl::imgproc::histogram_oclsrc,
format("-D BINS=%d -D HISTS_COUNT=1 -D WGS=%d", BINS, (int)wgs)); k1.args(ocl::KernelArg::ReadOnly(src),
if (k.empty()) ocl::KernelArg::PtrWriteOnly(ghist), (int)src.total());
size_t globalsize = compunits * wgs;
if (!k1.run(1, &globalsize, &wgs, false))
return false; return false;
k.args(ocl::KernelArg::PtrWriteOnly(lut), wgs = std::min<size_t>(ocl::Device::getDefault().maxWorkGroupSize(), BINS);
ocl::KernelArg::PtrReadOnly(hist), (int)_src.total()); UMat lut(1, 256, CV_8UC1);
ocl::Kernel k2("calcLUT", ocl::imgproc::histogram_oclsrc,
format("-D BINS=%d -D HISTS_COUNT=%d -D WGS=%d",
BINS, compunits, (int)wgs));
k2.args(ocl::KernelArg::PtrWriteOnly(lut),
ocl::KernelArg::PtrReadOnly(ghist), (int)_src.total());
// calculation of LUT // calculation of LUT
if (!k.run(1, &wgs, &wgs, false)) if (!k2.run(1, &wgs, &wgs, false))
return false; return false;
// execute LUT transparently // execute LUT transparently
......
modules/imgproc/src/opencl/histogram.cl
View file @ 7688a182
...@@ -37,86 +37,153 @@ ...@@ -37,86 +37,153 @@
// //
// //
#ifndef cn #ifndef kercn
#define cn 1 #define kercn 1
#endif #endif
#if cn == 16 #ifndef T
#define T uchar16
#else
#define T uchar #define T uchar
#endif #endif
#define noconvert
__kernel void calculate_histogram(__global const uchar * src, int src_step, int src_offset, int src_rows, int src_cols, __kernel void calculate_histogram(__global const uchar * src, int src_step, int src_offset, int src_rows, int src_cols,
__global uchar * hist, int total) __global uchar * histptr, int total)
{ {
int lid = get_local_id(0); int lid = get_local_id(0);
int id = get_global_id(0) * cn; int id = get_global_id(0) * kercn;
int gid = get_group_id(0); int gid = get_group_id(0);
__local int localhist[BINS]; __local int localhist[BINS];
#pragma unroll
for (int i = lid; i < BINS; i += WGS) for (int i = lid; i < BINS; i += WGS)
localhist[i] = 0; localhist[i] = 0;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
for (int grain = HISTS_COUNT * WGS * cn; id < total; id += grain) int src_index;
for (int grain = HISTS_COUNT * WGS * kercn; id < total; id += grain)
{ {
int src_index = mad24(id / src_cols, src_step, src_offset + id % src_cols); #ifdef HAVE_SRC_CONT
#if cn == 1 src_index = id;
atomic_inc(localhist + convert_int(src[src_index]));
#else #else
src_index = mad24(id / src_cols, src_step, src_offset + id % src_cols);
#endif
#if kercn == 1
atomic_inc(localhist + convert_int(src[src_index]));
#elif kercn == 4
int value = *(__global const int *)(src + src_index);
atomic_inc(localhist + (value & 0xff));
atomic_inc(localhist + ((value >> 8) & 0xff));
atomic_inc(localhist + ((value >> 16) & 0xff));
atomic_inc(localhist + ((value >> 24) & 0xff));
#elif kercn >= 2
T value = *(__global const T *)(src + src_index); T value = *(__global const T *)(src + src_index);
atomic_inc(localhist + convert_int(value.s0)); atomic_inc(localhist + value.s0);
atomic_inc(localhist + convert_int(value.s1)); atomic_inc(localhist + value.s1);
atomic_inc(localhist + convert_int(value.s2)); #if kercn >= 4
atomic_inc(localhist + convert_int(value.s3)); atomic_inc(localhist + value.s2);
atomic_inc(localhist + convert_int(value.s4)); atomic_inc(localhist + value.s3);
atomic_inc(localhist + convert_int(value.s5)); #if kercn >= 8
atomic_inc(localhist + convert_int(value.s6)); atomic_inc(localhist + value.s4);
atomic_inc(localhist + convert_int(value.s7)); atomic_inc(localhist + value.s5);
atomic_inc(localhist + convert_int(value.s8)); atomic_inc(localhist + value.s6);
atomic_inc(localhist + convert_int(value.s9)); atomic_inc(localhist + value.s7);
atomic_inc(localhist + convert_int(value.sA)); #if kercn == 16
atomic_inc(localhist + convert_int(value.sB)); atomic_inc(localhist + value.s8);
atomic_inc(localhist + convert_int(value.sC)); atomic_inc(localhist + value.s9);
atomic_inc(localhist + convert_int(value.sD)); atomic_inc(localhist + value.sA);
atomic_inc(localhist + convert_int(value.sE)); atomic_inc(localhist + value.sB);
atomic_inc(localhist + convert_int(value.sF)); atomic_inc(localhist + value.sC);
atomic_inc(localhist + value.sD);
atomic_inc(localhist + value.sE);
atomic_inc(localhist + value.sF);
#endif
#endif
#endif
#endif #endif
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
__global int * hist = (__global int *)(histptr + gid * BINS * (int)sizeof(int));
#pragma unroll
for (int i = lid; i < BINS; i += WGS) for (int i = lid; i < BINS; i += WGS)
*(__global int *)(hist + mad24(gid, BINS * (int)sizeof(int), i * (int)sizeof(int))) = localhist[i]; hist[i] = localhist[i];
} }
__kernel void merge_histogram(__global const int * ghist, __global int * hist) #ifndef HT
#define HT int
#endif
#ifndef convertToHT
#define convertToHT noconvert
#endif
__kernel void merge_histogram(__global const int * ghist, __global uchar * histptr, int hist_step, int hist_offset)
{ {
int lid = get_local_id(0); int lid = get_local_id(0);
__global HT * hist = (__global HT *)(histptr + hist_offset);
#if WGS >= BINS
HT res = (HT)(0);
#else
#pragma unroll #pragma unroll
for (int i = lid; i < BINS; i += WGS) for (int i = lid; i < BINS; i += WGS)
hist[i] = 0; hist[i] = (HT)(0);
barrier(CLK_LOCAL_MEM_FENCE); #endif
#pragma unroll #pragma unroll
for (int i = 0; i < HISTS_COUNT; ++i) for (int i = 0; i < HISTS_COUNT; ++i)
{ {
#pragma unroll #pragma unroll
for (int j = lid; j < BINS; j += WGS) for (int j = lid; j < BINS; j += WGS)
hist[j] += ghist[mad24(i, BINS, j)]; #if WGS >= BINS
barrier(CLK_LOCAL_MEM_FENCE); res += convertToHT(ghist[j]);
#else
hist[j] += convertToHT(ghist[j]);
#endif
ghist += BINS;
} }
#if WGS >= BINS
if (lid < BINS)
*(__global HT *)(histptr + mad24(lid, hist_step, hist_offset)) = res;
#endif
} }
__kernel void calcLUT(__global uchar * dst, __constant int * hist, int total) __kernel void calcLUT(__global uchar * dst, __global const int * ghist, int total)
{ {
int lid = get_local_id(0); int lid = get_local_id(0);
__local int sumhist[BINS]; __local int sumhist[BINS];
__local float scale; __local float scale;
sumhist[lid] = hist[lid]; #if WGS >= BINS
int res = 0;
#else
#pragma unroll
for (int i = lid; i < BINS; i += WGS)
sumhist[i] = 0;
#endif
#pragma unroll
for (int i = 0; i < HISTS_COUNT; ++i)
{
#pragma unroll
for (int j = lid; j < BINS; j += WGS)
#if WGS >= BINS
res += ghist[j];
#else
sumhist[j] += ghist[j];
#endif
ghist += BINS;
}
#if WGS >= BINS
if (lid < BINS)
sumhist[lid] = res;
#endif
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if (lid == 0) if (lid == 0)
......
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