Chiebot-Mirror
/
opencv
mirror of https://github.com/opencv/opencv.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity

Use BufferArea in more places

Browse Source
pull/16638/head
Maksim Shabunin 5 years ago
parent f9bd025722
commit bf96d8239d
6 changed files with 573 additions and 522 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Split View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 123
      modules/calib3d/src/rho.cpp
  2. 854
      modules/calib3d/src/stereosgbm.cpp
  3. 27
      modules/core/include/opencv2/core/utils/buffer_area.private.hpp
  4. 49
      modules/core/src/buffer_area.cpp
  5. 15
      modules/core/test/test_utils.cpp
  6. 27
      modules/features2d/src/fast.cpp

123
modules/calib3d/src/rho.cpp
Unescape View File

@ -55,7 +55,7 @@
#include <math.h>
#include <vector>
#include "rho.h"
#include "opencv2/core/utils/buffer_area.private.hpp"
@ -65,7 +65,6 @@ namespace cv{/* For C support, replace with extern "C" { */
/* Constants */
const int MEM_ALIGN = 32;
const size_t HSIZE = (3*3*sizeof(float));
const double MIN_DELTA_CHNG = 0.1;
// const double CHI_STAT = 2.706;
@ -312,16 +311,14 @@ struct RHO_HEST_REFC : RHO_HEST{
/* Levenberg-Marquardt Refinement */
struct{
float (* JtJ)[8]; /* JtJ matrix */
float (* tmp1)[8]; /* Temporary 1 */
float* JtJ; /* JtJ matrix */
float* tmp1; /* Temporary 1 */
float* Jte; /* Jte vector */
} lm;
/* Memory Management */
struct{
cv::Mat perObj;
cv::Mat perRun;
} mem;
utils::BufferArea runArea;
utils::BufferArea objArea;
/* Initialized? */
int initialized;
@ -659,16 +656,9 @@ inline int RHO_HEST_REFC::initialize(void){
fastSeed((uint64_t)~0);
initialized = 1;
int areAllAllocsSuccessful = !mem.perObj.empty();
if(!areAllAllocsSuccessful){
finalize();
}else{
initialized = 1;
}
return areAllAllocsSuccessful;
return true;
}
/**
@ -835,45 +825,14 @@ unsigned RHO_HEST_REFC::rhoHest(const float* src, /* Source points */
*/
inline void RHO_HEST_REFC::allocatePerObj(void){
/* We have known sizes */
size_t ctrl_smpl_sz = SMPL_SIZE*sizeof(*ctrl.smpl);
size_t curr_pkdPts_sz = SMPL_SIZE*2*2*sizeof(*curr.pkdPts);
size_t curr_H_sz = HSIZE;
size_t best_H_sz = HSIZE;
size_t lm_JtJ_sz = 8*8*sizeof(float);
size_t lm_tmp1_sz = 8*8*sizeof(float);
size_t lm_Jte_sz = 1*8*sizeof(float);
/* We compute offsets */
size_t total = 0;
#define MK_OFFSET(v) \
size_t v ## _of = total; \
total = alignSize(v ## _of + v ## _sz, MEM_ALIGN)
MK_OFFSET(ctrl_smpl);
MK_OFFSET(curr_pkdPts);
MK_OFFSET(curr_H);
MK_OFFSET(best_H);
MK_OFFSET(lm_JtJ);
MK_OFFSET(lm_tmp1);
MK_OFFSET(lm_Jte);
#undef MK_OFFSET
/* Allocate dynamic memory managed by cv::Mat */
mem.perObj.create(1, (int)(total + MEM_ALIGN), CV_8UC1);
/* Extract aligned pointer */
unsigned char* ptr = alignPtr(mem.perObj.data, MEM_ALIGN);
/* Assign pointers */
ctrl.smpl = (unsigned*) (ptr + ctrl_smpl_of);
curr.pkdPts = (float*) (ptr + curr_pkdPts_of);
curr.H = (float*) (ptr + curr_H_of);
best.H = (float*) (ptr + best_H_of);
lm.JtJ = (float(*)[8])(ptr + lm_JtJ_of);
lm.tmp1 = (float(*)[8])(ptr + lm_tmp1_of);
lm.Jte = (float*) (ptr + lm_Jte_of);
objArea.allocate(ctrl.smpl, SMPL_SIZE);
objArea.allocate(curr.pkdPts, SMPL_SIZE*2*2);
objArea.allocate(curr.H, HSIZE);
objArea.allocate(best.H, HSIZE);
objArea.allocate(lm.JtJ, 8*8);
objArea.allocate(lm.tmp1, 8*8);
objArea.allocate(lm.Jte, 1*8);
objArea.commit();
}
@ -885,30 +844,9 @@ inline void RHO_HEST_REFC::allocatePerObj(void){
*/
inline void RHO_HEST_REFC::allocatePerRun(void){
/* We have known sizes */
size_t best_inl_sz = arg.N;
size_t curr_inl_sz = arg.N;
/* We compute offsets */
size_t total = 0;
#define MK_OFFSET(v) \
size_t v ## _of = total; \
total = alignSize(v ## _of + v ## _sz, MEM_ALIGN)
MK_OFFSET(best_inl);
MK_OFFSET(curr_inl);
#undef MK_OFFSET
/* Allocate dynamic memory managed by cv::Mat */
mem.perRun.create(1, (int)(total + MEM_ALIGN), CV_8UC1);
/* Extract aligned pointer */
unsigned char* ptr = alignPtr(mem.perRun.data, MEM_ALIGN);
/* Assign pointers */
best.inl = (char*)(ptr + best_inl_of);
curr.inl = (char*)(ptr + curr_inl_of);
runArea.allocate(best.inl, arg.N);
runArea.allocate(curr.inl, arg.N);
runArea.commit();
}
@ -919,10 +857,7 @@ inline void RHO_HEST_REFC::allocatePerRun(void){
*/
inline void RHO_HEST_REFC::deallocatePerRun(void){
best.inl = NULL;
curr.inl = NULL;
mem.perRun.release();
runArea.release();
}
@ -933,15 +868,7 @@ inline void RHO_HEST_REFC::deallocatePerRun(void){
*/
inline void RHO_HEST_REFC::deallocatePerObj(void){
ctrl.smpl = NULL;
curr.pkdPts = NULL;
curr.H = NULL;
best.H = NULL;
lm.JtJ = NULL;
lm.tmp1 = NULL;
lm.Jte = NULL;
mem.perObj.release();
objArea.release();
}
@ -2144,7 +2071,7 @@ inline void RHO_HEST_REFC::refine(void){
*/
/* Find initial conditions */
sacCalcJacobianErrors(best.H, arg.src, arg.dst, best.inl, arg.N,
lm.JtJ, lm.Jte, &S);
(float(*)[8])lm.JtJ, lm.Jte, &S);
/*Levenberg-Marquardt Loop.*/
for(i=0;i<MAXLEVMARQITERS;i++){
@ -2169,11 +2096,11 @@ inline void RHO_HEST_REFC::refine(void){
* transpose) then multiply Jte in order to find dH.
*/
while(!sacChol8x8Damped(lm.JtJ, L, lm.tmp1)){
while(!sacChol8x8Damped((float(*)[8])lm.JtJ, L, (float(*)[8])lm.tmp1)){
L *= 2.0f;
}
sacTRInv8x8 (lm.tmp1, lm.tmp1);
sacTRISolve8x8(lm.tmp1, lm.Jte, dH);
sacTRInv8x8 ((float(*)[8])lm.tmp1, (float(*)[8])lm.tmp1);
sacTRISolve8x8((float(*)[8])lm.tmp1, lm.Jte, dH);
sacSub8x1 (newH, best.H, dH);
sacCalcJacobianErrors(newH, arg.src, arg.dst, best.inl, arg.N,
NULL, NULL, &newS);
@ -2204,7 +2131,7 @@ inline void RHO_HEST_REFC::refine(void){
S = newS;
memcpy(best.H, newH, sizeof(newH));
sacCalcJacobianErrors(best.H, arg.src, arg.dst, best.inl, arg.N,
lm.JtJ, lm.Jte, &S);
(float(*)[8])lm.JtJ, lm.Jte, &S);
}
}
}

854
modules/calib3d/src/stereosgbm.cpp
View File

File diff suppressed because it is too large Load Diff

27
modules/core/include/opencv2/core/utils/buffer_area.private.hpp
Unescape View File

@ -74,6 +74,25 @@ public:
allocate_((void**)(&ptr), static_cast<ushort>(sizeof(T)), count, alignment);
}
/** @brief Fill one of buffers with zeroes
@param ptr pointer to memory block previously added using BufferArea::allocate
BufferArea::commit must be called before using this method
*/
template <typename T>
void zeroFill(T*&ptr)
{
CV_Assert(ptr);
zeroFill_((void**)&ptr);
}
/** @brief Fill all buffers with zeroes
BufferArea::commit must be called before using this method
*/
void zeroFill();
/** @brief Allocate memory and initialize all bound pointers
Each pointer bound to the area with the BufferArea::allocate will be initialized and will be set
@ -83,10 +102,18 @@ public:
*/
void commit();
/** @brief Release all memory and unbind all pointers
All memory will be freed and all pointers will be reset to NULL and untied from the area allowing
to call `allocate` and `commit` again.
*/
void release();
private:
BufferArea(const BufferArea &); // = delete
BufferArea &operator=(const BufferArea &); // = delete
void allocate_(void **ptr, ushort type_size, size_t count, ushort alignment);
void zeroFill_(void **ptr);
private:
class Block;

49
modules/core/src/buffer_area.cpp
Unescape View File

@ -66,6 +66,16 @@ public:
*ptr = buf;
return static_cast<void*>(static_cast<uchar*>(*ptr) + type_size * count);
}
bool operator==(void **other) const
{
CV_Assert(ptr && other);
return *ptr == *other;
}
void zeroFill() const
{
CV_Assert(ptr && *ptr);
memset(static_cast<uchar*>(*ptr), 0, count * type_size);
}
private:
void **ptr;
void * raw_mem;
@ -85,10 +95,7 @@ BufferArea::BufferArea(bool safe_) :
BufferArea::~BufferArea()
{
for(std::vector<Block>::const_iterator i = blocks.begin(); i != blocks.end(); ++i)
i->cleanup();
if (oneBuf)
fastFree(oneBuf);
release();
}
void BufferArea::allocate_(void **ptr, ushort type_size, size_t count, ushort alignment)
@ -100,6 +107,26 @@ void BufferArea::allocate_(void **ptr, ushort type_size, size_t count, ushort al
totalSize += blocks.back().getByteCount();
}
void BufferArea::zeroFill_(void **ptr)
{
for(std::vector<Block>::const_iterator i = blocks.begin(); i != blocks.end(); ++i)
{
if (*i == ptr)
{
i->zeroFill();
break;
}
}
}
void BufferArea::zeroFill()
{
for(std::vector<Block>::const_iterator i = blocks.begin(); i != blocks.end(); ++i)
{
i->zeroFill();
}
}
void BufferArea::commit()
{
if (!safe)
@ -116,6 +143,20 @@ void BufferArea::commit()
}
}
void BufferArea::release()
{
for(std::vector<Block>::const_iterator i = blocks.begin(); i != blocks.end(); ++i)
{
i->cleanup();
}
blocks.clear();
if (oneBuf)
{
fastFree(oneBuf);
oneBuf = 0;
}
}
//==================================================================================================
}} // cv::utils::

15
modules/core/test/test_utils.cpp
Unescape View File

@ -337,6 +337,21 @@ TEST_P(BufferArea, basic)
ASSERT_TRUE(dbl_ptr != NULL);
EXPECT_EQ((size_t)0, (size_t)int_ptr % sizeof(int));
EXPECT_EQ((size_t)0, (size_t)dbl_ptr % sizeof(double));
for (size_t i = 0; i < SZ; ++i)
{
int_ptr[i] = (int)i + 1;
uchar_ptr[i] = (uchar)i + 1;
dbl_ptr[i] = (double)i + 1;
}
area.zeroFill(int_ptr);
area.zeroFill(uchar_ptr);
area.zeroFill(dbl_ptr);
for (size_t i = 0; i < SZ; ++i)
{
EXPECT_EQ((int)0, int_ptr[i]);
EXPECT_EQ((uchar)0, uchar_ptr[i]);
EXPECT_EQ((double)0, dbl_ptr[i]);
}
}
EXPECT_TRUE(int_ptr == NULL);
EXPECT_TRUE(uchar_ptr == NULL);

27
modules/features2d/src/fast.cpp
Unescape View File

@ -47,6 +47,7 @@ The references are:
#include "opencl_kernels_features2d.hpp"
#include "hal_replacement.hpp"
#include "opencv2/core/hal/intrin.hpp"
#include "opencv2/core/utils/buffer_area.private.hpp"
#include "opencv2/core/openvx/ovx_defs.hpp"
@ -80,20 +81,26 @@ void FAST_t(InputArray _img, std::vector<KeyPoint>& keypoints, int threshold, bo
for( i = -255; i <= 255; i++ )
threshold_tab[i+255] = (uchar)(i < -threshold ? 1 : i > threshold ? 2 : 0);
AutoBuffer<uchar> _buf((img.cols+16)*3*(sizeof(int) + sizeof(uchar)) + 128);
uchar* buf[3];
buf[0] = _buf.data(); buf[1] = buf[0] + img.cols; buf[2] = buf[1] + img.cols;
int* cpbuf[3];
cpbuf[0] = (int*)alignPtr(buf[2] + img.cols, sizeof(int)) + 1;
cpbuf[1] = cpbuf[0] + img.cols + 1;
cpbuf[2] = cpbuf[1] + img.cols + 1;
memset(buf[0], 0, img.cols*3);
uchar* buf[3] = { 0 };
int* cpbuf[3] = { 0 };
utils::BufferArea area;
for (unsigned idx = 0; idx < 3; ++idx)
{
area.allocate(buf[idx], img.cols);
area.allocate(cpbuf[idx], img.cols + 1);
}
area.commit();
for (unsigned idx = 0; idx < 3; ++idx)
{
memset(buf[idx], 0, img.cols);
}
for(i = 3; i < img.rows-2; i++)
{
const uchar* ptr = img.ptr<uchar>(i) + 3;
uchar* curr = buf[(i - 3)%3];
int* cornerpos = cpbuf[(i - 3)%3];
int* cornerpos = cpbuf[(i - 3)%3] + 1; // cornerpos[-1] is used to store a value
memset(curr, 0, img.cols);
int ncorners = 0;
@ -266,7 +273,7 @@ void FAST_t(InputArray _img, std::vector<KeyPoint>& keypoints, int threshold, bo
const uchar* prev = buf[(i - 4 + 3)%3];
const uchar* pprev = buf[(i - 5 + 3)%3];
cornerpos = cpbuf[(i - 4 + 3)%3];
cornerpos = cpbuf[(i - 4 + 3)%3] + 1; // cornerpos[-1] is used to store a value
ncorners = cornerpos[-1];
for( k = 0; k < ncorners; k++ )

Write Preview
Loading…
Cancel
Save