From 40d0e0eda0ddb04099c58127c9d12ace78da8f6e Mon Sep 17 00:00:00 2001
From: yao <bitwangyaoyao@gmail.com>
Date: Sat, 13 Apr 2013 14:58:49 +0800
Subject: [PATCH] use host data when DEVICE_MEM_UHP is set (the risk of vary
align size is owned by users)
---
modules/ocl/include/opencv2/ocl.hpp | 6 ++-
.../ocl/include/opencv2/ocl/private/util.hpp | 3 +-
modules/ocl/src/initialization.cpp | 19 ++++---
modules/ocl/src/matrix_operations.cpp | 50 +++++--------------
4 files changed, 31 insertions(+), 47 deletions(-)
diff --git a/modules/ocl/include/opencv2/ocl.hpp b/modules/ocl/include/opencv2/ocl.hpp
index cd1429a63f..7cc55b8335 100644
--- a/modules/ocl/include/opencv2/ocl.hpp
+++ b/modules/ocl/include/opencv2/ocl.hpp
@@ -263,8 +263,10 @@ namespace cv
void create(Size size, int type);
//! allocates new oclMatrix with specified device memory type.
- void createEx(int rows, int cols, int type, DevMemRW rw_type, DevMemType mem_type);
- void createEx(Size size, int type, DevMemRW rw_type, DevMemType mem_type);
+ void createEx(int rows, int cols, int type,
+ DevMemRW rw_type, DevMemType mem_type, void* hptr = 0);
+ void createEx(Size size, int type, DevMemRW rw_type,
+ DevMemType mem_type, void* hptr = 0);
//! decreases reference counter;
// deallocate the data when reference counter reaches 0.
diff --git a/modules/ocl/include/opencv2/ocl/private/util.hpp b/modules/ocl/include/opencv2/ocl/private/util.hpp
index 9ac032a29a..786f2d6192 100644
--- a/modules/ocl/include/opencv2/ocl/private/util.hpp
+++ b/modules/ocl/include/opencv2/ocl/private/util.hpp
@@ -68,7 +68,8 @@ namespace cv
void CV_EXPORTS openCLMallocPitch(Context *clCxt, void **dev_ptr, size_t *pitch,
size_t widthInBytes, size_t height);
void CV_EXPORTS openCLMallocPitchEx(Context *clCxt, void **dev_ptr, size_t *pitch,
- size_t widthInBytes, size_t height, DevMemRW rw_type, DevMemType mem_type);
+ size_t widthInBytes, size_t height,
+ DevMemRW rw_type, DevMemType mem_type, void* hptr = 0);
void CV_EXPORTS openCLMemcpy2D(Context *clCxt, void *dst, size_t dpitch,
const void *src, size_t spitch,
size_t width, size_t height, openCLMemcpyKind kind, int channels = -1);
diff --git a/modules/ocl/src/initialization.cpp b/modules/ocl/src/initialization.cpp
index 37a69cece5..9394b7e9d0 100644
--- a/modules/ocl/src/initialization.cpp
+++ b/modules/ocl/src/initialization.cpp
@@ -163,7 +163,7 @@ namespace cv
{
releaseResources();
delete this;
- }
+ }
}
Impl* copy()
@@ -260,9 +260,8 @@ namespace cv
int setDevMemType(DevMemRW rw_type, DevMemType mem_type)
{
- if( (mem_type == DEVICE_MEM_PM && Context::getContext()->impl->unified_memory == 0) ||
- mem_type == DEVICE_MEM_UHP ||
- mem_type == DEVICE_MEM_CHP )
+ if( (mem_type == DEVICE_MEM_PM &&
+ Context::getContext()->impl->unified_memory == 0) )
return -1;
gDeviceMemRW = rw_type;
gDeviceMemType = mem_type;
@@ -432,11 +431,17 @@ namespace cv
}
void openCLMallocPitchEx(Context *clCxt, void **dev_ptr, size_t *pitch,
- size_t widthInBytes, size_t height, DevMemRW rw_type, DevMemType mem_type)
+ size_t widthInBytes, size_t height,
+ DevMemRW rw_type, DevMemType mem_type, void* hptr)
{
cl_int status;
- *dev_ptr = clCreateBuffer(clCxt->impl->oclcontext, gDevMemRWValueMap[rw_type]|gDevMemTypeValueMap[mem_type],
- widthInBytes * height, 0, &status);
+ if(hptr && (mem_type==DEVICE_MEM_UHP || mem_type==DEVICE_MEM_CHP))
+ *dev_ptr = clCreateBuffer(clCxt->impl->oclcontext,
+ gDevMemRWValueMap[rw_type]|gDevMemTypeValueMap[mem_type],
+ widthInBytes * height, hptr, &status);
+ else
+ *dev_ptr = clCreateBuffer(clCxt->impl->oclcontext, gDevMemRWValueMap[rw_type]|gDevMemTypeValueMap[mem_type],
+ widthInBytes * height, 0, &status);
openCLVerifyCall(status);
*pitch = widthInBytes;
}
diff --git a/modules/ocl/src/matrix_operations.cpp b/modules/ocl/src/matrix_operations.cpp
index 7cd596f333..4d697a2d50 100644
--- a/modules/ocl/src/matrix_operations.cpp
+++ b/modules/ocl/src/matrix_operations.cpp
@@ -177,15 +177,9 @@ void cv::ocl::oclMat::upload(const Mat &m)
Size wholeSize;
Point ofs;
m.locateROI(wholeSize, ofs);
- // int type = m.type();
- // if(m.oclchannels() == 3)
- //{
- // type = CV_MAKETYPE(m.depth(), 4);
- //}
- create(wholeSize, m.type());
-
if(m.channels() == 3)
{
+ create(wholeSize, m.type());
int pitch = wholeSize.width * 3 * m.elemSize1();
int tail_padding = m.elemSize1() * 3072;
int err;
@@ -195,35 +189,20 @@ void cv::ocl::oclMat::upload(const Mat &m)
openCLMemcpy2D(clCxt, temp, pitch, m.datastart, m.step, wholeSize.width * m.elemSize(), wholeSize.height, clMemcpyHostToDevice, 3);
convert_C3C4(temp, *this);
- //int* cputemp=new int[wholeSize.height*wholeSize.width * 3];
- //int* cpudata=new int[this->step*this->wholerows/sizeof(int)];
- //openCLSafeCall(clEnqueueReadBuffer(clCxt->impl->clCmdQueue, temp, CL_TRUE,
- // 0, wholeSize.height*wholeSize.width * 3* sizeof(int), cputemp, 0, NULL, NULL));
- //openCLSafeCall(clEnqueueReadBuffer(clCxt->impl->clCmdQueue, (cl_mem)data, CL_TRUE,
- // 0, this->step*this->wholerows, cpudata, 0, NULL, NULL));
- //for(int i=0;i<wholeSize.height;i++)
- //{
- // int *a = cputemp+i*wholeSize.width * 3,*b = cpudata + i*this->step/sizeof(int);
- // for(int j=0;j<wholeSize.width;j++)
- // {
- // if((a[3*j] != b[4*j])||(a[3*j+1] != b[4*j+1])||(a[3*j+2] != b[4*j+2]))
- // printf("rows=%d,cols=%d,cputtemp=%d,%d,%d;cpudata=%d,%d,%d\n",
- // i,j,a[3*j],a[3*j+1],a[3*j+2],b[4*j],b[4*j+1],b[4*j+2]);
- // }
- //}
- //delete []cputemp;
- //delete []cpudata;
openCLSafeCall(clReleaseMemObject(temp));
}
else
{
- openCLMemcpy2D(clCxt, data, step, m.datastart, m.step, wholeSize.width * elemSize(), wholeSize.height, clMemcpyHostToDevice);
+ // try to use host ptr
+ createEx(wholeSize, m.type(), gDeviceMemRW, gDeviceMemType, m.datastart);
+ if(gDeviceMemType!=DEVICE_MEM_UHP && gDeviceMemType!=DEVICE_MEM_CHP)
+ openCLMemcpy2D(clCxt, data, step, m.datastart, m.step,
+ wholeSize.width * elemSize(), wholeSize.height, clMemcpyHostToDevice);
}
rows = m.rows;
cols = m.cols;
offset = ofs.y * step + ofs.x * elemSize();
- //download_channels = m.channels();
}
void cv::ocl::oclMat::download(cv::Mat &m) const
@@ -908,9 +887,10 @@ oclMat cv::ocl::oclMat::reshape(int new_cn, int new_rows) const
}
-void cv::ocl::oclMat::createEx(Size size, int type, DevMemRW rw_type, DevMemType mem_type)
+void cv::ocl::oclMat::createEx(Size size, int type,
+ DevMemRW rw_type, DevMemType mem_type, void* hptr)
{
- createEx(size.height, size.width, type, rw_type, mem_type);
+ createEx(size.height, size.width, type, rw_type, mem_type, hptr);
}
void cv::ocl::oclMat::create(int _rows, int _cols, int _type)
@@ -918,16 +898,12 @@ void cv::ocl::oclMat::create(int _rows, int _cols, int _type)
createEx(_rows, _cols, _type, gDeviceMemRW, gDeviceMemType);
}
-void cv::ocl::oclMat::createEx(int _rows, int _cols, int _type, DevMemRW rw_type, DevMemType mem_type)
+void cv::ocl::oclMat::createEx(int _rows, int _cols, int _type,
+ DevMemRW rw_type, DevMemType mem_type, void* hptr)
{
clCxt = Context::getContext();
/* core logic */
_type &= Mat::TYPE_MASK;
- //download_channels = CV_MAT_CN(_type);
- //if(download_channels==3)
- //{
- // _type = CV_MAKE_TYPE((CV_MAT_DEPTH(_type)),4);
- //}
if( rows == _rows && cols == _cols && type() == _type && data )
return;
if( data )
@@ -943,8 +919,8 @@ void cv::ocl::oclMat::createEx(int _rows, int _cols, int _type, DevMemRW rw_type
size_t esz = elemSize();
void *dev_ptr;
- openCLMallocPitchEx(clCxt, &dev_ptr, &step, GPU_MATRIX_MALLOC_STEP(esz * cols), rows, rw_type, mem_type);
- //openCLMallocPitch(clCxt,&dev_ptr, &step, esz * cols, rows);
+ openCLMallocPitchEx(clCxt, &dev_ptr, &step, GPU_MATRIX_MALLOC_STEP(esz * cols),
+ rows, rw_type, mem_type, hptr);
if (esz * cols == step)
flags |= Mat::CONTINUOUS_FLAG;