Skip to content

O
opencv
Commit aa92be34 authored by marina.kolpakova's avatar marina.kolpakova
Browse files
Options

GK107 Policy

parent 580d8173
Show whitespace changes
Inline Side-by-side
Showing with 76 additions and 138 deletions
modules/gpu/src/cuda/isf-sc.cu
View file @ aa92be34
......@@ -161,12 +161,11 @@ namespace icf {
}
texture<float2, cudaTextureType2D, cudaReadModeElementType> troi;
#if defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 300
template<bool isUp>
__global__ void test_kernel_warp(const Level* levels, const Octave* octaves, const float* stages,
const Node* nodes, const float* leaves, Detection* objects, const uint ndetections, uint* ctr,
const int downscales)
{
template<typename Policy>
template<bool isUp>
__device void CascadeInvoker<Policy>::detect(Detection* objects, const uint ndetections, uint* ctr, const int downscales) const
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x;
......@@ -174,7 +173,8 @@ namespace icf {
__shared__ Level level;
// check POI
__shared__ volatile char roiCache[8];
__shared__ volatile char roiCache[Policy::STA_Y];
if (!threadIdx.y && !threadIdx.x)
((float2*)roiCache)[threadIdx.x] = tex2D(troi, blockIdx.y, x);
......@@ -187,24 +187,20 @@ namespace icf {
if(x >= level.workRect.x || y >= level.workRect.y) return;
Octave octave = octaves[level.octave];
int st = octave.index * octave.stages;
const int stEnd = st + 1024;
int st = level.octave * level.step;
const int stEnd = st + level.step;
float confidence = 0.f;
for(; st < stEnd; st += 32)
for(; st < stEnd; st += Policy::WARP)
{
const int nId = (st + threadIdx.x) * 3;
dprintf("\n\n%d: stage: %d %d\n",threadIdx.x, st, nId);
Node node = nodes[nId];
float threshold = rescale<isUp>(level, node);
int sum = get<isUp>(x, y + (node.threshold >> 28) * 120, node.rect);
int next = 1 + (int)(sum >= threshold);
dprintf("%d: go: %d (%d >= %f)\n\n" ,threadIdx.x, next, sum, threshold);
node = nodes[nId + next];
threshold = rescale<isUp>(level, node);
......@@ -213,21 +209,16 @@ namespace icf {
const int lShift = (next - 1) * 2 + (int)(sum >= threshold);
float impact = leaves[(st + threadIdx.x) * 4 + lShift];
dprintf("%d: decided: %d (%d >= %f) %d %f\n\n" ,threadIdx.x, next, sum, threshold, lShift, impact);
dprintf("%d: extracted stage: %f\n",threadIdx.x, stages[(st + threadIdx.x)]);
dprintf("%d: computed score: %f\n",threadIdx.x, impact);
#pragma unroll
// scan on shuffl functions
for (int i = 1; i < 32; i *= 2)
for (int i = 1; i < Policy::WARP; i *= 2)
{
const float n = __shfl_up(impact, i, 32);
const float n = __shfl_up(impact, i, Policy::WARP);
if (threadIdx.x >= i)
impact += n;
}
dprintf("%d: impact scaned %f\n" ,threadIdx.x, impact);
confidence += impact;
if(__any((confidence <= stages[(st + threadIdx.x)]))) st += 2048;
}
......@@ -235,86 +226,25 @@ namespace icf {
if(!threadIdx.x && st == stEnd && ((confidence - FLT_EPSILON) >= 0))
{
int idx = atomicInc(ctr, ndetections);
// store detection
objects[idx] = Detection(__float2int_rn(x * octave.shrinkage),
__float2int_rn(y * octave.shrinkage), level.objSize.x, level.objSize.y, confidence);
}
}
#else
template<bool isUp>
__global__ void test_kernel_warp(const Level* levels, const Octave* octaves, const float* stages,
const Node* nodes, const float* leaves, Detection* objects, const uint ndetections, uint* ctr,
const int downscales)
{
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int x = blockIdx.x * blockDim.x + threadIdx.x;
Level level = levels[blockIdx.z];
// if (blockIdx.z != 31) return;
if(x >= level.workRect.x || y >= level.workRect.y) return;
// int roi = tex2D(troi, x, y);
// printf("%d\n", roi);
// if (!roi) return;
Octave octave = octaves[level.octave];
int st = octave.index * octave.stages;
const int stEnd = st + 1000;//octave.stages;
float confidence = 0.f;
for(; st < stEnd; ++st)
{
dprintf("\n\nstage: %d\n", st);
const int nId = st * 3;
Node node = nodes[nId];
dprintf("Node: [%d %d %d %d] %d %d\n", node.rect.x, node.rect.y, node.rect.z, node.rect.w,
node.threshold >> 28, node.threshold & 0x0FFFFFFFU);
float threshold = rescale<isUp>(level, node);
int sum = get<isUp>(x, y + (node.threshold >> 28) * 121, node.rect);
dprintf("Node: [%d %d %d %d] %f\n", node.rect.x, node.rect.y, node.rect.z,
node.rect.w, threshold);
int next = 1 + (int)(sum >= threshold);
dprintf("go: %d (%d >= %f)\n\n" ,next, sum, threshold);
node = nodes[nId + next];
threshold = rescale<isUp>(level, node);
sum = get<isUp>(x, y + (node.threshold >> 28) * 121, node.rect);
const int lShift = (next - 1) * 2 + (int)(sum >= threshold);
float impact = leaves[st * 4 + lShift];
confidence += impact;
if (confidence <= stages[st]) st = stEnd + 10;
dprintf("decided: %d (%d >= %f) %d %f\n\n" ,next, sum, threshold, lShift, impact);
dprintf("extracted stage: %f\n", stages[st]);
dprintf("computed score: %f\n\n", confidence);
objects[idx] = Detection(__float2int_rn(x * Policy::SHRINKAGE),
__float2int_rn(y * Policy::SHRINKAGE), level.objSize.x, level.objSize.y, confidence);
}
}
if(st == stEnd)
{
int idx = atomicInc(ctr, ndetections);
// store detection
objects[idx] = Detection(__float2int_rn(x * octave.shrinkage),
__float2int_rn(y * octave.shrinkage), level.objSize.x, level.objSize.y, confidence);
}
}
#endif
template<typename Policy, bool isUp>
__global__ void soft_cascade(const CascadeInvoker<Policy> invoker, Detection* objects, const uint n, uint* ctr, const int downs)
{
invoker.template detect<isUp>(objects, n, ctr, downs);
}
template<>
void CascadeInvoker<CascadePolicy>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
template<typename Policy>
void CascadeInvoker<Policy>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
PtrStepSz<uchar4> objects, PtrStepSzi counter, const int downscales, const int scale, const cudaStream_t& stream) const
{
{
int fw = 160;
int fh = 120;
dim3 block(32, 8);
dim3 grid(fw, fh / 8, (scale == -1) ? downscales : 1);
dim3 grid(fw, fh / Policy::STA_Y, (scale == -1) ? downscales : 1);
uint* ctr = (uint*)(counter.ptr(0));
Detection* det = (Detection*)objects.ptr();
......@@ -326,20 +256,22 @@ namespace icf {
cudaChannelFormatDesc desc_roi = cudaCreateChannelDesc<float2>();
cudaSafeCall( cudaBindTexture2D(0, troi, roi.data, desc_roi, roi.cols / 8, roi.rows, roi.step));
const CascadeInvoker<Policy> inv = *this;
if (scale == -1)
{
test_kernel_warp<false><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, 0);
soft_cascade<Policy, false><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, 0);
cudaSafeCall( cudaGetLastError());
grid = dim3(fw, fh / 8, 47 - downscales);
test_kernel_warp<true><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, downscales);
grid = dim3(fw, fh / Policy::STA_Y, scales - downscales);
soft_cascade<Policy, true><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, downscales);
}
else
{
if (scale >= downscales)
test_kernel_warp<true><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
soft_cascade<Policy, true><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, scale);
else
test_kernel_warp<false><<<grid, block, 0, stream>>>(levels, octaves, stages, nodes, leaves, det, max_det, ctr, scale);
soft_cascade<Policy, false><<<grid, Policy::block(), 0, stream>>>(inv, det, max_det, ctr, scale);
}
if (!stream)
......@@ -347,6 +279,10 @@ namespace icf {
cudaSafeCall( cudaGetLastError());
cudaSafeCall( cudaDeviceSynchronize());
}
}
}
template void CascadeInvoker<GK107PolicyX4>::operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv,
PtrStepSz<uchar4> objects, PtrStepSzi counter, const int downscales, const int scale, const cudaStream_t& stream) const;
}
}}}
\ No newline at end of file
modules/gpu/src/icf.hpp
View file @ aa92be34
......@@ -72,9 +72,9 @@ struct __align__(16) Octave
struct __align__(8) Level //is actually 24 bytes
{
int octave;
int step;
float relScale;
float shrScale; // used for marking detection
float scaling[2]; // calculated according to Dollal paper
// for 640x480 we can not get overflow
......@@ -115,31 +115,41 @@ struct __align__(16) Detection
: x(_x), y(_y), w(_w), h(_h), confidence(c), kind(0) {};
};
struct CascadePolicy
struct GK107PolicyX4
{
enum {STA_X = 32, STA_Y = 8};
enum {WARP = 32, STA_X = WARP, STA_Y = 8, SHRINKAGE = 4};
static const dim3 block()
{
return dim3(GK107PolicyX4::STA_X, GK107PolicyX4::STA_Y);
}
};
template<typename Policy>
struct CascadeInvoker
{
CascadeInvoker(): levels(0), octaves(0), stages(0), nodes(0), leaves(0) {}
CascadeInvoker(): levels(0), stages(0), nodes(0), leaves(0), scales(0) {}
CascadeInvoker(const PtrStepSzb& _levels, const PtrStepSzb& _octaves, const PtrStepSzf& _stages,
const PtrStepSzb& _nodes, const PtrStepSzf& _leaves)
: levels((const Level*)_levels.ptr()), octaves((const Octave*)_octaves.ptr()), stages((const float*)_stages.ptr()),
nodes((const Node*)_nodes.ptr()), leaves((const float*)_leaves.ptr())
: levels((const Level*)_levels.ptr()),
stages((const float*)_stages.ptr()),
nodes((const Node*)_nodes.ptr()), leaves((const float*)_leaves.ptr()),
scales(_levels.cols / sizeof(Level))
{}
const Level* levels;
const Octave* octaves;
const float* stages;
const Node* nodes;
const float* leaves;
int scales;
void operator()(const PtrStepSzb& roi, const PtrStepSzi& hogluv, PtrStepSz<uchar4> objects,
PtrStepSzi counter, const int downscales, const int csale = -1, const cudaStream_t& stream = 0) const;
template<bool isUp>
__device void detect(Detection* objects, const uint ndetections, uint* ctr, const int downscales) const;
};
}
......
modules/gpu/src/softcascade.cpp
View file @ aa92be34
......@@ -63,7 +63,7 @@ void cv::gpu::SCascade::read(const FileNode& fn) { Algorithm::read(fn); }
#include <icf.hpp>
cv::gpu::device::icf::Level::Level(int idx, const Octave& oct, const float scale, const int w, const int h)
: octave(idx), relScale(scale / oct.scale), shrScale (relScale / (float)oct.shrinkage)
: octave(idx), step(oct.stages), relScale(scale / oct.scale)
{
workRect.x = round(w / (float)oct.shrinkage);
workRect.y = round(h / (float)oct.shrinkage);
......@@ -100,7 +100,7 @@ namespace imgproc {
struct cv::gpu::SCascade::Fields
{
static Fields* parseCascade(const FileNode &root, const float mins, const float maxs)
static Fields* parseCascade(const FileNode &root, const float mins, const float maxs, const int totals)
{
static const char *const SC_STAGE_TYPE = "stageType";
static const char *const SC_BOOST = "BOOST";
......@@ -120,10 +120,7 @@ struct cv::gpu::SCascade::Fields
static const char *const SC_ORIG_H = "height";
int origWidth = (int)root[SC_ORIG_W];
CV_Assert(origWidth == ORIG_OBJECT_WIDTH);
int origHeight = (int)root[SC_ORIG_H];
CV_Assert(origHeight == ORIG_OBJECT_HEIGHT);
static const char *const SC_OCTAVES = "octaves";
static const char *const SC_STAGES = "stages";
......@@ -142,7 +139,6 @@ struct cv::gpu::SCascade::Fields
static const char * const SC_F_CHANNEL = "channel";
static const char * const SC_F_RECT = "rect";
FileNode fn = root[SC_OCTAVES];
if (fn.empty()) return false;
......@@ -167,8 +163,8 @@ struct cv::gpu::SCascade::Fields
ushort nstages = saturate_cast<ushort>((int)fns[SC_OCT_STAGES]);
ushort2 size;
size.x = cvRound(ORIG_OBJECT_WIDTH * scale);
size.y = cvRound(ORIG_OBJECT_HEIGHT * scale);
size.x = cvRound(origWidth * scale);
size.y = cvRound(origHeight * scale);
shrinkage = saturate_cast<ushort>((int)fns[SC_OCT_SHRINKAGE]);
Octave octave(octIndex, nstages, shrinkage, size, scale);
......@@ -245,11 +241,11 @@ struct cv::gpu::SCascade::Fields
CV_Assert(!hleaves.empty());
std::vector<Level> vlevels;
float logFactor = (::log(maxs) - ::log(mins)) / (TOTAL_SCALES -1);
float logFactor = (::log(maxs) - ::log(mins)) / (totals -1);
float scale = mins;
int downscales = 0;
for (int sc = 0; sc < TOTAL_SCALES; ++sc)
for (int sc = 0; sc < totals; ++sc)
{
int width = ::std::max(0.0f, FRAME_WIDTH - (origWidth * scale));
int height = ::std::max(0.0f, FRAME_HEIGHT - (origHeight * scale));
......@@ -302,7 +298,7 @@ struct cv::gpu::SCascade::Fields
leaves.upload(hleaves);
levels.upload(hlevels);
invoker = device::icf::CascadeInvoker<device::icf::CascadePolicy>(levels, octaves, stages, nodes, leaves);
invoker = device::icf::CascadeInvoker<device::icf::GK107PolicyX4>(levels, octaves, stages, nodes, leaves);
}
......@@ -456,16 +452,13 @@ public:
GpuMat sobelBuf;
device::icf::CascadeInvoker<device::icf::CascadePolicy> invoker;
device::icf::CascadeInvoker<device::icf::GK107PolicyX4> invoker;
enum { BOOST = 0 };
enum
{
FRAME_WIDTH = 640,
FRAME_HEIGHT = 480,
TOTAL_SCALES = 55,
ORIG_OBJECT_WIDTH = 64,
ORIG_OBJECT_HEIGHT = 128,
HOG_BINS = 6,
LUV_BINS = 3,
HOG_LUV_BINS = 10
......@@ -480,21 +473,19 @@ cv::gpu::SCascade::~SCascade() { delete fields; }
bool cv::gpu::SCascade::load(const FileNode& fn)
{
if (fields) delete fields;
fields = Fields::parseCascade(fn, minScale, maxScale);
fields = Fields::parseCascade(fn, minScale, maxScale, scales);
return fields != 0;
}
void cv::gpu::SCascade::detect(InputArray image, InputArray _rois, OutputArray _objects, Stream& s) const
{
CV_Assert(fields);
const GpuMat colored = image.getGpuMat();
// only color images are supperted
CV_Assert(colored.type() == CV_8UC3 || colored.type() == CV_32SC1);
GpuMat rois = _rois.getGpuMat(), objects = _objects.getGpuMat();
// we guess user knows about shrincage
// CV_Assert((rois.size().width == getRoiSize().height) && (rois.type() == CV_8UC1));
Fields& flds = *fields;
if (colored.type() == CV_8UC3)
......@@ -518,15 +509,13 @@ void cv::gpu::SCascade::detect(InputArray image, InputArray _rois, OutputArray _
void cv::gpu::SCascade::detect(InputArray image, InputArray _rois, OutputArray _objects, const int level, Stream& s) const
{
CV_Assert(fields);
const GpuMat colored = image.getGpuMat();
// only color images are supperted
CV_Assert(colored.type() == CV_8UC3 || colored.type() == CV_32SC1);
// we guess user knows about shrincage
// CV_Assert((rois.size().width == getRoiSize().height) && (rois.type() == CV_8UC1));
Fields& flds = *fields;
if (colored.type() == CV_8UC3)
{
// only this window size allowed
......@@ -549,6 +538,8 @@ void cv::gpu::SCascade::detect(InputArray image, InputArray _rois, OutputArray _
void cv::gpu::SCascade::genRoi(InputArray _roi, OutputArray _mask, Stream& stream) const
{
CV_Assert(fields);
const GpuMat roi = _roi.getGpuMat();
_mask.create( roi.cols / 4, roi.rows / 4, roi.type() );
GpuMat mask = _mask.getGpuMat();
......
modules/gpu/test/test_softcascade.cpp
View file @ aa92be34
......@@ -172,7 +172,7 @@ GPU_TEST_P(SCascadeTestRoi, detect,
sub.setTo(1);
cv::rectangle(result, r, cv::Scalar(0, 0, 255, 255), 1);
}
objectBoxes.setTo(0);
cascade.genRoi(rois, trois);
cascade.detect(colored, trois, objectBoxes);
......@@ -222,7 +222,7 @@ GPU_TEST_P(SCascadeTestLevel, detect,
cv::gpu::GpuMat trois;
cascade.genRoi(rois, trois);
objectBoxes.setTo(0);
int level = GET_PARAM(3);
cascade.detect(colored, trois, objectBoxes, level);
......@@ -281,7 +281,7 @@ GPU_TEST_P(SCascadeTestAll, detect,
cv::gpu::GpuMat trois;
cascade.genRoi(rois, trois);
objectBoxes.setTo(0);
cascade.detect(colored, trois, objectBoxes);
typedef cv::gpu::SCascade::Detection Detection;
......@@ -321,7 +321,7 @@ GPU_TEST_P(SCascadeTestAll, detectOnIntegral,
cv::gpu::GpuMat trois;
cascade.genRoi(rois, trois);
objectBoxes.setTo(0);
cascade.detect(hogluv, trois, objectBoxes);
typedef cv::gpu::SCascade::Detection Detection;
......@@ -357,7 +357,7 @@ GPU_TEST_P(SCascadeTestAll, detectStream,
cv::gpu::GpuMat trois;
cascade.genRoi(rois, trois, s);
objectBoxes.setTo(0);
cascade.detect(colored, trois, objectBoxes, s);
cudaDeviceSynchronize();
......
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