mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2025-01-18 14:13:15 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Allow OpenCL acceleration in every OpenCV module

Browse Source
This commit is contained in:
Andrey Kamaev 2013-03-15 23:56:31 +04:00
parent d28df08eb0
commit 91ac9688a8
95 changed files with 1999 additions and 2066 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
CMakeLists.txt
View File

@ -782,7 +782,7 @@ if(HAVE_CUDA)
status(" Use fast math:" CUDA_FAST_MATH THEN YES ELSE NO)
endif()
if(HAVE_OPENCL AND BUILD_opencv_ocl)
if(HAVE_OPENCL)
status("")
status(" OpenCL")
if(OPENCL_INCLUDE_DIR)

15
cmake/OpenCVModule.cmake
View File

@ -432,10 +432,22 @@ macro(ocv_glob_module_sources)
file(GLOB lib_hdrs "include/opencv2/${name}/*.hpp" "include/opencv2/${name}/*.h")
file(GLOB lib_hdrs_detail "include/opencv2/${name}/detail/*.hpp" "include/opencv2/${name}/detail/*.h")
file(GLOB cl_kernels "src/opencl/*.cl")
source_group("Src" FILES ${lib_srcs} ${lib_int_hdrs})
source_group("Include" FILES ${lib_hdrs})
source_group("Include\\detail" FILES ${lib_hdrs_detail})
if(HAVE_OPENCL AND cl_kernels)
ocv_include_directories(${OPENCL_INCLUDE_DIRS})
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/kernels.cpp"
COMMAND ${CMAKE_COMMAND} -DCL_DIR="${CMAKE_CURRENT_SOURCE_DIR}/src/opencl" -DOUTPUT="${CMAKE_CURRENT_BINARY_DIR}/kernels.cpp" -P "${OpenCV_SOURCE_DIR}/cmake/cl2cpp.cmake"
DEPENDS ${cl_kernels} "${OpenCV_SOURCE_DIR}/cmake/cl2cpp.cmake")
source_group("Src\\OpenCL" FILES ${cl_kernels} "${CMAKE_CURRENT_BINARY_DIR}/kernels.cpp")
list(APPEND lib_srcs ${cl_kernels} "${CMAKE_CURRENT_BINARY_DIR}/kernels.cpp")
endif()
ocv_set_module_sources(${ARGN} HEADERS ${lib_hdrs} ${lib_hdrs_detail} SOURCES ${lib_srcs} ${lib_int_hdrs})
endmacro()
@ -449,6 +461,9 @@ macro(ocv_create_module)
if(NOT "${ARGN}" STREQUAL "SKIP_LINK")
target_link_libraries(${the_module} ${OPENCV_MODULE_${the_module}_DEPS} ${OPENCV_MODULE_${the_module}_DEPS_EXT} ${OPENCV_LINKER_LIBS} ${IPP_LIBS} ${ARGN})
if(HAVE_OPENCL AND OPENCL_LIBRARIES)
target_link_libraries(${the_module} ${OPENCL_LIBRARIES})
endif()
endif()
add_dependencies(opencv_modules ${the_module})

0
modules/ocl/cl2cpp.cmake → cmake/cl2cpp.cmake
View File

42
modules/ocl/CMakeLists.txt
View File

@ -3,45 +3,5 @@ if(NOT HAVE_OPENCL)
endif()
set(the_description "OpenCL-accelerated Computer Vision")
ocv_add_module(ocl opencv_core opencv_imgproc opencv_features2d opencv_objdetect opencv_video opencv_nonfree)
ocv_module_include_directories(${OPENCL_INCLUDE_DIRS})
file(GLOB CL_FILES "${CMAKE_CURRENT_SOURCE_DIR}/src/kernels/*.cl")
set(kernels_cpp "${CMAKE_CURRENT_BINARY_DIR}/kernels.cpp")
set(cl2cpp_script "${CMAKE_CURRENT_SOURCE_DIR}/cl2cpp.cmake")
add_custom_command(
OUTPUT ${kernels_cpp}
COMMAND ${CMAKE_COMMAND} -DCL_DIR="${CMAKE_CURRENT_SOURCE_DIR}/src/kernels" -DOUTPUT="${kernels_cpp}" -P ${cl2cpp_script}
DEPENDS ${CL_FILES} ${cl2cpp_script})
file(GLOB lib_hdrs "include/opencv2/${name}/*.hpp" "include/opencv2/${name}/*.h")
file(GLOB lib_srcs "src/*.cpp")
file(GLOB lib_int_hdrs "src/*.h*")
source_group("Include" FILES ${lib_hdrs})
source_group("Src\\Host" FILES ${lib_srcs} ${lib_int_hdrs} ${kernels_cpp})
ocv_define_module(ocl opencv_core opencv_imgproc opencv_features2d opencv_objdetect opencv_video opencv_nonfree)
ocv_warnings_disable(CMAKE_CXX_FLAGS -Wshadow)
ocv_set_module_sources(HEADERS ${lib_hdrs} SOURCES ${lib_int_hdrs} ${lib_srcs} ${kernels_cpp})
ocv_create_module(${OPENCL_LIBRARIES})
ocv_add_precompiled_headers(${the_module})
################################################################################################################
################################ OpenCL Module Tests ##################################################
################################################################################################################
file(GLOB test_srcs "test/*.cpp")
file(GLOB test_hdrs "test/*.hpp" "test/*.h")
ocv_add_accuracy_tests(FILES "Include" ${test_hdrs}
FILES "Src" ${test_srcs})
################################################################################################################
################################ OpenCL Module Performance ##################################################
################################################################################################################
file(GLOB perf_srcs "perf/*.cpp")
file(GLOB perf_hdrs "perf/*.hpp" "perf/*.h")
ocv_add_perf_tests(FILES "Include" ${perf_hdrs}
FILES "Src" ${perf_srcs})

865
modules/ocl/src/kernels/brute_force_match.cl
View File

@ -1,865 +0,0 @@
#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics:enable
#define MAX_FLOAT 1e7f
int bit1Count(float x)
{
int c = 0;
int ix = (int)x;
for (int i = 0 ; i < 32 ; i++)
{
c += ix & 0x1;
ix >>= 1;
}
return (float)c;
}
/* 2dim launch, global size: dim0 is (query rows + block_size - 1) / block_size * block_size, dim1 is block_size
local size: dim0 is block_size, dim1 is block_size.
*/
__kernel void BruteForceMatch_UnrollMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * max_desc_len;
int queryIdx = groupidx * block_size + lidy;
// load the query into local memory.
for (int i = 0 ; i < max_desc_len / block_size; i ++)
{
int loadx = lidx + i * block_size;
s_query[lidy * max_desc_len + loadx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
}
float myBestDistance = MAX_FLOAT;
int myBestTrainIdx = -1;
// loopUnrolledCached to find the best trainIdx and best distance.
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_train[lidx * block_size + lidy] = loadx < train_cols ? train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows && result < myBestDistance/* && mask(queryIdx, trainIdx)*/)
{
//bestImgIdx = imgIdx;
myBestDistance = result;
myBestTrainIdx = trainIdx;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)(sharebuffer);
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//find BestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance;
s_trainIdx[lidx] = myBestTrainIdx;
barrier(CLK_LOCAL_MEM_FENCE);
//reduce -- now all reduce implement in each threads.
for (int k = 0 ; k < block_size; k++)
{
if (myBestDistance > s_distance[k])
{
myBestDistance = s_distance[k];
myBestTrainIdx = s_trainIdx[k];
}
}
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = myBestTrainIdx;
bestDistance[queryIdx] = myBestDistance;
}
}
__kernel void BruteForceMatch_Match(
__global float *query,
__global float *train,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
float myBestDistance = MAX_FLOAT;
int myBestTrainIdx = -1;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
// loop
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
//Dist dist;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
//load query and train into local memory
s_query[lidy * block_size + lidx] = 0;
s_train[lidx * block_size + lidy] = 0;
if (loadx < query_cols)
{
s_query[lidy * block_size + lidx] = query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx];
s_train[lidx * block_size + lidy] = train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx];
}
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows && result < myBestDistance /*&& mask(queryIdx, trainIdx)*/)
{
//myBestImgidx = imgIdx;
myBestDistance = result;
myBestTrainIdx = trainIdx;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)sharebuffer;
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//findBestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance;
s_trainIdx[lidx] = myBestTrainIdx;
barrier(CLK_LOCAL_MEM_FENCE);
//reduce -- now all reduce implement in each threads.
for (int k = 0 ; k < block_size; k++)
{
if (myBestDistance > s_distance[k])
{
myBestDistance = s_distance[k];
myBestTrainIdx = s_trainIdx[k];
}
}
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = myBestTrainIdx;
bestDistance[queryIdx] = myBestDistance;
}
}
//radius_unrollmatch
__kernel void BruteForceMatch_RadiusUnrollMatch(
__global float *query,
__global float *train,
float maxDistance,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int bestTrainIdx_cols,
int step,
int ostep,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int groupidy = get_group_id(1);
const int queryIdx = groupidy * block_size + lidy;
const int trainIdx = groupidx * block_size + lidx;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; ++i)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_query[lidy * block_size + lidx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
s_train[lidx * block_size + lidy] = loadx < query_cols ? train[min(groupidx * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are three types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (queryIdx < query_rows && trainIdx < train_rows && result < maxDistance/* && mask(queryIdx, trainIdx)*/)
{
unsigned int ind = atom_inc(nMatches + queryIdx/*, (unsigned int) -1*/);
if (ind < bestTrainIdx_cols)
{
//bestImgIdx = imgIdx;
bestTrainIdx[queryIdx * (ostep / sizeof(int)) + ind] = trainIdx;
bestDistance[queryIdx * (ostep / sizeof(float)) + ind] = result;
}
}
}
//radius_match
__kernel void BruteForceMatch_RadiusMatch(
__global float *query,
__global float *train,
float maxDistance,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int bestTrainIdx_cols,
int step,
int ostep,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int groupidy = get_group_id(1);
const int queryIdx = groupidy * block_size + lidy;
const int trainIdx = groupidx * block_size + lidx;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; ++i)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_query[lidy * block_size + lidx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
s_train[lidx * block_size + lidy] = loadx < query_cols ? train[min(groupidx * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are three types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (queryIdx < query_rows && trainIdx < train_rows && result < maxDistance/* && mask(queryIdx, trainIdx)*/)
{
unsigned int ind = atom_inc(nMatches + queryIdx/*, (unsigned int) -1*/);
if (ind < bestTrainIdx_cols)
{
//bestImgIdx = imgIdx;
bestTrainIdx[queryIdx * (ostep / sizeof(int)) + ind] = trainIdx;
bestDistance[queryIdx * (ostep / sizeof(float)) + ind] = result;
}
}
}
__kernel void BruteForceMatch_knnUnrollMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
local float *s_query = sharebuffer;
local float *s_train = sharebuffer + block_size * max_desc_len;
// load the query into local memory.
for (int i = 0 ; i < max_desc_len / block_size; i ++)
{
int loadx = lidx + i * block_size;
s_query[lidy * max_desc_len + loadx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
}
float myBestDistance1 = MAX_FLOAT;
float myBestDistance2 = MAX_FLOAT;
int myBestTrainIdx1 = -1;
int myBestTrainIdx2 = -1;
//loopUnrolledCached
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
const int loadX = lidx + i * block_size;
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_train[lidx * block_size + lidy] = loadx < train_cols ? train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows)
{
if (result < myBestDistance1)
{
myBestDistance2 = myBestDistance1;
myBestTrainIdx2 = myBestTrainIdx1;
myBestDistance1 = result;
myBestTrainIdx1 = trainIdx;
}
else if (result < myBestDistance2)
{
myBestDistance2 = result;
myBestTrainIdx2 = trainIdx;
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
local float *s_distance = (local float *)sharebuffer;
local int *s_trainIdx = (local int *)(sharebuffer + block_size * block_size);
// find BestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance1;
s_trainIdx[lidx] = myBestTrainIdx1;
float bestDistance1 = MAX_FLOAT;
float bestDistance2 = MAX_FLOAT;
int bestTrainIdx1 = -1;
int bestTrainIdx2 = -1;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestDistance1 = val;
bestTrainIdx1 = s_trainIdx[i];
}
else if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
s_distance[lidx] = myBestDistance2;
s_trainIdx[lidx] = myBestTrainIdx2;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
myBestDistance1 = bestDistance1;
myBestDistance2 = bestDistance2;
myBestTrainIdx1 = bestTrainIdx1;
myBestTrainIdx2 = bestTrainIdx2;
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = (int2)(myBestTrainIdx1, myBestTrainIdx2);
bestDistance[queryIdx] = (float2)(myBestDistance1, myBestDistance2);
}
}
__kernel void BruteForceMatch_knnMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
local float *s_query = sharebuffer;
local float *s_train = sharebuffer + block_size * block_size;
float myBestDistance1 = MAX_FLOAT;
float myBestDistance2 = MAX_FLOAT;
int myBestTrainIdx1 = -1;
int myBestTrainIdx2 = -1;
//loop
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0.0f;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
//load query and train into local memory
s_query[lidy * block_size + lidx] = 0;
s_train[lidx * block_size + lidy] = 0;
if (loadx < query_cols)
{
s_query[lidy * block_size + lidx] = query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx];
s_train[lidx * block_size + lidy] = train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx];
}
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows /*&& mask(queryIdx, trainIdx)*/)
{
if (result < myBestDistance1)
{
myBestDistance2 = myBestDistance1;
myBestTrainIdx2 = myBestTrainIdx1;
myBestDistance1 = result;
myBestTrainIdx1 = trainIdx;
}
else if (result < myBestDistance2)
{
myBestDistance2 = result;
myBestTrainIdx2 = trainIdx;
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)sharebuffer;
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//findBestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance1;
s_trainIdx[lidx] = myBestTrainIdx1;
float bestDistance1 = MAX_FLOAT;
float bestDistance2 = MAX_FLOAT;
int bestTrainIdx1 = -1;
int bestTrainIdx2 = -1;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestDistance1 = val;
bestTrainIdx1 = s_trainIdx[i];
}
else if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
s_distance[lidx] = myBestDistance2;
s_trainIdx[lidx] = myBestTrainIdx2;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
myBestDistance1 = bestDistance1;
myBestDistance2 = bestDistance2;
myBestTrainIdx1 = bestTrainIdx1;
myBestTrainIdx2 = bestTrainIdx2;
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = (int2)(myBestTrainIdx1, myBestTrainIdx2);
bestDistance[queryIdx] = (float2)(myBestDistance1, myBestDistance2);
}
}
kernel void BruteForceMatch_calcDistanceUnrolled(
__global float *query,
__global float *train,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType)
{
/* Todo */
}
kernel void BruteForceMatch_calcDistance(
__global float *query,
__global float *train,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType)
{
/* Todo */
}
kernel void BruteForceMatch_findBestMatch(
__global float *allDist,
__global int *bestTrainIdx,
__global float *bestDistance,
int k,
int block_size
)
{
/* Todo */
}

0
modules/ocl/src/kernels/arithm_2_mat.cl → modules/ocl/src/opencl/arithm_2_mat.cl
View File

0
modules/ocl/src/kernels/arithm_LUT.cl → modules/ocl/src/opencl/arithm_LUT.cl
View File

0
modules/ocl/src/kernels/arithm_absdiff.cl → modules/ocl/src/opencl/arithm_absdiff.cl
View File

0
modules/ocl/src/kernels/arithm_add.cl → modules/ocl/src/opencl/arithm_add.cl
View File

0
modules/ocl/src/kernels/arithm_addWeighted.cl → modules/ocl/src/opencl/arithm_addWeighted.cl
View File

0
modules/ocl/src/kernels/arithm_add_scalar.cl → modules/ocl/src/opencl/arithm_add_scalar.cl
View File

0
modules/ocl/src/kernels/arithm_add_scalar_mask.cl → modules/ocl/src/opencl/arithm_add_scalar_mask.cl
View File

1
modules/ocl/src/kernels/arithm_bitwise_and.cl → modules/ocl/src/opencl/arithm_bitwise_and.cl
View File

@ -320,4 +320,3 @@ __kernel void arithm_bitwise_and_D6 (__global char *src1, int src1_step, int src
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_and_mask.cl → modules/ocl/src/opencl/arithm_bitwise_and_mask.cl
View File

@ -1135,4 +1135,3 @@ __kernel void arithm_bitwise_and_with_mask_C4_D6 (__global char *src1, int src1_
}
}
#endif

0
modules/ocl/src/kernels/arithm_bitwise_and_scalar.cl → modules/ocl/src/opencl/arithm_bitwise_and_scalar.cl
View File

1
modules/ocl/src/kernels/arithm_bitwise_and_scalar_mask.cl → modules/ocl/src/opencl/arithm_bitwise_and_scalar_mask.cl
View File

@ -1055,4 +1055,3 @@ __kernel void arithm_s_bitwise_and_with_mask_C4_D6 (__global short *src1, int sr
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_not.cl → modules/ocl/src/opencl/arithm_bitwise_not.cl
View File

@ -255,4 +255,3 @@ __kernel void arithm_bitwise_not_D6 (__global char *src, int src_step, int src_o
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_or.cl → modules/ocl/src/opencl/arithm_bitwise_or.cl
View File

@ -276,4 +276,3 @@ __kernel void arithm_bitwise_or_D6 (__global char *src1, int src1_step, int src1
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_or_mask.cl → modules/ocl/src/opencl/arithm_bitwise_or_mask.cl
View File

@ -1135,4 +1135,3 @@ __kernel void arithm_bitwise_or_with_mask_C4_D6 (__global char *src1, int src1_s
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_or_scalar.cl → modules/ocl/src/opencl/arithm_bitwise_or_scalar.cl
View File

@ -911,4 +911,3 @@ __kernel void arithm_s_bitwise_or_C4_D6 (__global short *src1, int src1_step, in
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_or_scalar_mask.cl → modules/ocl/src/opencl/arithm_bitwise_or_scalar_mask.cl
View File

@ -1078,4 +1078,3 @@ __kernel void arithm_s_bitwise_or_with_mask_C4_D6 (__global short *src1, int src
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_xor.cl → modules/ocl/src/opencl/arithm_bitwise_xor.cl
View File

@ -324,4 +324,3 @@ __kernel void arithm_bitwise_xor_D6 (__global char *src1, int src1_step, int src
}
}
#endif

1
modules/ocl/src/kernels/arithm_bitwise_xor_mask.cl → modules/ocl/src/opencl/arithm_bitwise_xor_mask.cl
View File

@ -1135,4 +1135,3 @@ __kernel void arithm_bitwise_xor_with_mask_C4_D6 (__global char *src1, int src1_
}
}
#endif

0
modules/ocl/src/kernels/arithm_bitwise_xor_scalar.cl → modules/ocl/src/opencl/arithm_bitwise_xor_scalar.cl
View File

1
modules/ocl/src/kernels/arithm_bitwise_xor_scalar_mask.cl → modules/ocl/src/opencl/arithm_bitwise_xor_scalar_mask.cl
View File

@ -1055,4 +1055,3 @@ __kernel void arithm_s_bitwise_xor_with_mask_C4_D6 (__global short *src1, int sr
}
}
#endif

0
modules/ocl/src/kernels/arithm_cartToPolar.cl → modules/ocl/src/opencl/arithm_cartToPolar.cl
View File

1
modules/ocl/src/kernels/arithm_compare_eq.cl → modules/ocl/src/opencl/arithm_compare_eq.cl
View File

@ -954,4 +954,3 @@ __kernel void arithm_compare_ge_D6 (__global double *src1, int src1_step, int sr
}
}
#endif

2
modules/ocl/src/kernels/arithm_compare_ne.cl → modules/ocl/src/opencl/arithm_compare_ne.cl
View File

@ -952,5 +952,3 @@ __kernel void arithm_compare_le_D6 (__global double *src1, int src1_step, int sr
}
}
#endif

2
modules/ocl/src/kernels/arithm_div.cl → modules/ocl/src/opencl/arithm_div.cl
View File

@ -455,5 +455,3 @@ __kernel void arithm_s_div_D6 (__global double *src, int src_step, int src_offse
}
}
#endif

0
modules/ocl/src/kernels/arithm_exp.cl → modules/ocl/src/opencl/arithm_exp.cl
View File

0
modules/ocl/src/kernels/arithm_flip.cl → modules/ocl/src/opencl/arithm_flip.cl
View File

0
modules/ocl/src/kernels/arithm_flip_rc.cl → modules/ocl/src/opencl/arithm_flip_rc.cl
View File

0
modules/ocl/src/kernels/arithm_log.cl → modules/ocl/src/opencl/arithm_log.cl
View File

0
modules/ocl/src/kernels/arithm_magnitude.cl → modules/ocl/src/opencl/arithm_magnitude.cl
View File

0
modules/ocl/src/kernels/arithm_magnitudeSqr.cl → modules/ocl/src/opencl/arithm_magnitudeSqr.cl
View File

0
modules/ocl/src/kernels/arithm_minMax.cl → modules/ocl/src/opencl/arithm_minMax.cl
View File

0
modules/ocl/src/kernels/arithm_minMaxLoc.cl → modules/ocl/src/opencl/arithm_minMaxLoc.cl
View File

1
modules/ocl/src/kernels/arithm_minMaxLoc_mask.cl → modules/ocl/src/opencl/arithm_minMaxLoc_mask.cl
View File

@ -240,4 +240,3 @@ __kernel void arithm_op_minMaxLoc_mask (int cols,int invalid_cols,int offset,int
dst[gid + 3 * groupnum] = CONVERT_RES_TYPE(lm_maxloc[0]);
}
}

1
modules/ocl/src/kernels/arithm_minMax_mask.cl → modules/ocl/src/opencl/arithm_minMax_mask.cl
View File

@ -194,4 +194,3 @@ __kernel void arithm_op_minMax_mask (int cols,int invalid_cols,int offset,int el
dst[gid + groupnum] = localmem_max[0];
}
}

0
modules/ocl/src/kernels/arithm_mul.cl → modules/ocl/src/opencl/arithm_mul.cl
View File

0
modules/ocl/src/kernels/arithm_nonzero.cl → modules/ocl/src/opencl/arithm_nonzero.cl
View File

0
modules/ocl/src/kernels/arithm_phase.cl → modules/ocl/src/opencl/arithm_phase.cl
View File

0
modules/ocl/src/kernels/arithm_polarToCart.cl → modules/ocl/src/opencl/arithm_polarToCart.cl
View File

0
modules/ocl/src/kernels/arithm_pow.cl → modules/ocl/src/opencl/arithm_pow.cl
View File

0
modules/ocl/src/kernels/arithm_sub.cl → modules/ocl/src/opencl/arithm_sub.cl
View File

0
modules/ocl/src/kernels/arithm_sub_scalar.cl → modules/ocl/src/opencl/arithm_sub_scalar.cl
View File

0
modules/ocl/src/kernels/arithm_sub_scalar_mask.cl → modules/ocl/src/opencl/arithm_sub_scalar_mask.cl
View File

1
modules/ocl/src/kernels/arithm_sum.cl → modules/ocl/src/opencl/arithm_sum.cl
View File

@ -203,4 +203,3 @@ __kernel void arithm_op_sum (int cols,int invalid_cols,int offset,int elemnum,in
dst[gid] = localmem_sum[0];
}
}

1
modules/ocl/src/kernels/arithm_sum_3.cl → modules/ocl/src/opencl/arithm_sum_3.cl
View File

@ -245,4 +245,3 @@ __kernel void arithm_op_sum_3 (int cols,int invalid_cols,int offset,int elemnum,
dst[gid*3+2] = localmem_sum3[0];
}
}

0
modules/ocl/src/kernels/arithm_transpose.cl → modules/ocl/src/opencl/arithm_transpose.cl
View File

1
modules/ocl/src/kernels/blend_linear.cl → modules/ocl/src/opencl/blend_linear.cl
View File

@ -138,4 +138,3 @@ __kernel void BlendLinear_C4_D5(
dst[pos] = (img1[pos] * w1 + img2[pos] * w2) / (w1 + w2 + 1e-5f);
}
}

865
modules/ocl/src/opencl/brute_force_match.cl Normal file
View File

@ -0,0 +1,865 @@
#pragma OPENCL EXTENSION cl_khr_global_int32_base_atomics:enable
#define MAX_FLOAT 1e7f
int bit1Count(float x)
{
int c = 0;
int ix = (int)x;
for (int i = 0 ; i < 32 ; i++)
{
c += ix & 0x1;
ix >>= 1;
}
return (float)c;
}
/* 2dim launch, global size: dim0 is (query rows + block_size - 1) / block_size * block_size, dim1 is block_size
local size: dim0 is block_size, dim1 is block_size.
*/
__kernel void BruteForceMatch_UnrollMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * max_desc_len;
int queryIdx = groupidx * block_size + lidy;
// load the query into local memory.
for (int i = 0 ; i < max_desc_len / block_size; i ++)
{
int loadx = lidx + i * block_size;
s_query[lidy * max_desc_len + loadx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
}
float myBestDistance = MAX_FLOAT;
int myBestTrainIdx = -1;
// loopUnrolledCached to find the best trainIdx and best distance.
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_train[lidx * block_size + lidy] = loadx < train_cols ? train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows && result < myBestDistance/* && mask(queryIdx, trainIdx)*/)
{
//bestImgIdx = imgIdx;
myBestDistance = result;
myBestTrainIdx = trainIdx;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)(sharebuffer);
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//find BestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance;
s_trainIdx[lidx] = myBestTrainIdx;
barrier(CLK_LOCAL_MEM_FENCE);
//reduce -- now all reduce implement in each threads.
for (int k = 0 ; k < block_size; k++)
{
if (myBestDistance > s_distance[k])
{
myBestDistance = s_distance[k];
myBestTrainIdx = s_trainIdx[k];
}
}
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = myBestTrainIdx;
bestDistance[queryIdx] = myBestDistance;
}
}
__kernel void BruteForceMatch_Match(
__global float *query,
__global float *train,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
float myBestDistance = MAX_FLOAT;
int myBestTrainIdx = -1;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
// loop
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
//Dist dist;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
//load query and train into local memory
s_query[lidy * block_size + lidx] = 0;
s_train[lidx * block_size + lidy] = 0;
if (loadx < query_cols)
{
s_query[lidy * block_size + lidx] = query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx];
s_train[lidx * block_size + lidy] = train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx];
}
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows && result < myBestDistance /*&& mask(queryIdx, trainIdx)*/)
{
//myBestImgidx = imgIdx;
myBestDistance = result;
myBestTrainIdx = trainIdx;
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)sharebuffer;
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//findBestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance;
s_trainIdx[lidx] = myBestTrainIdx;
barrier(CLK_LOCAL_MEM_FENCE);
//reduce -- now all reduce implement in each threads.
for (int k = 0 ; k < block_size; k++)
{
if (myBestDistance > s_distance[k])
{
myBestDistance = s_distance[k];
myBestTrainIdx = s_trainIdx[k];
}
}
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = myBestTrainIdx;
bestDistance[queryIdx] = myBestDistance;
}
}
//radius_unrollmatch
__kernel void BruteForceMatch_RadiusUnrollMatch(
__global float *query,
__global float *train,
float maxDistance,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int bestTrainIdx_cols,
int step,
int ostep,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int groupidy = get_group_id(1);
const int queryIdx = groupidy * block_size + lidy;
const int trainIdx = groupidx * block_size + lidx;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; ++i)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_query[lidy * block_size + lidx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
s_train[lidx * block_size + lidy] = loadx < query_cols ? train[min(groupidx * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are three types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (queryIdx < query_rows && trainIdx < train_rows && result < maxDistance/* && mask(queryIdx, trainIdx)*/)
{
unsigned int ind = atom_inc(nMatches + queryIdx/*, (unsigned int) -1*/);
if (ind < bestTrainIdx_cols)
{
//bestImgIdx = imgIdx;
bestTrainIdx[queryIdx * (ostep / sizeof(int)) + ind] = trainIdx;
bestDistance[queryIdx * (ostep / sizeof(float)) + ind] = result;
}
}
}
//radius_match
__kernel void BruteForceMatch_RadiusMatch(
__global float *query,
__global float *train,
float maxDistance,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int bestTrainIdx_cols,
int step,
int ostep,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int groupidy = get_group_id(1);
const int queryIdx = groupidy * block_size + lidy;
const int trainIdx = groupidx * block_size + lidx;
__local float *s_query = sharebuffer;
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; ++i)
{
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_query[lidy * block_size + lidx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
s_train[lidx * block_size + lidy] = loadx < query_cols ? train[min(groupidx * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are three types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if (queryIdx < query_rows && trainIdx < train_rows && result < maxDistance/* && mask(queryIdx, trainIdx)*/)
{
unsigned int ind = atom_inc(nMatches + queryIdx/*, (unsigned int) -1*/);
if (ind < bestTrainIdx_cols)
{
//bestImgIdx = imgIdx;
bestTrainIdx[queryIdx * (ostep / sizeof(int)) + ind] = trainIdx;
bestDistance[queryIdx * (ostep / sizeof(float)) + ind] = result;
}
}
}
__kernel void BruteForceMatch_knnUnrollMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
local float *s_query = sharebuffer;
local float *s_train = sharebuffer + block_size * max_desc_len;
// load the query into local memory.
for (int i = 0 ; i < max_desc_len / block_size; i ++)
{
int loadx = lidx + i * block_size;
s_query[lidy * max_desc_len + loadx] = loadx < query_cols ? query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx] : 0;
}
float myBestDistance1 = MAX_FLOAT;
float myBestDistance2 = MAX_FLOAT;
int myBestTrainIdx1 = -1;
int myBestTrainIdx2 = -1;
//loopUnrolledCached
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
const int loadX = lidx + i * block_size;
//load a block_size * block_size block into local train.
const int loadx = lidx + i * block_size;
s_train[lidx * block_size + lidy] = loadx < train_cols ? train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx] : 0;
//synchronize to make sure each elem for reduceIteration in share memory is written already.
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows)
{
if (result < myBestDistance1)
{
myBestDistance2 = myBestDistance1;
myBestTrainIdx2 = myBestTrainIdx1;
myBestDistance1 = result;
myBestTrainIdx1 = trainIdx;
}
else if (result < myBestDistance2)
{
myBestDistance2 = result;
myBestTrainIdx2 = trainIdx;
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
local float *s_distance = (local float *)sharebuffer;
local int *s_trainIdx = (local int *)(sharebuffer + block_size * block_size);
// find BestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance1;
s_trainIdx[lidx] = myBestTrainIdx1;
float bestDistance1 = MAX_FLOAT;
float bestDistance2 = MAX_FLOAT;
int bestTrainIdx1 = -1;
int bestTrainIdx2 = -1;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestDistance1 = val;
bestTrainIdx1 = s_trainIdx[i];
}
else if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
s_distance[lidx] = myBestDistance2;
s_trainIdx[lidx] = myBestTrainIdx2;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
myBestDistance1 = bestDistance1;
myBestDistance2 = bestDistance2;
myBestTrainIdx1 = bestTrainIdx1;
myBestTrainIdx2 = bestTrainIdx2;
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = (int2)(myBestTrainIdx1, myBestTrainIdx2);
bestDistance[queryIdx] = (float2)(myBestDistance1, myBestDistance2);
}
}
__kernel void BruteForceMatch_knnMatch(
__global float *query,
__global float *train,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType
)
{
const int lidx = get_local_id(0);
const int lidy = get_local_id(1);
const int groupidx = get_group_id(0);
const int queryIdx = groupidx * block_size + lidy;
local float *s_query = sharebuffer;
local float *s_train = sharebuffer + block_size * block_size;
float myBestDistance1 = MAX_FLOAT;
float myBestDistance2 = MAX_FLOAT;
int myBestTrainIdx1 = -1;
int myBestTrainIdx2 = -1;
//loop
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0.0f;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
//load query and train into local memory
s_query[lidy * block_size + lidx] = 0;
s_train[lidx * block_size + lidy] = 0;
if (loadx < query_cols)
{
s_query[lidy * block_size + lidx] = query[min(queryIdx, query_rows - 1) * (step / sizeof(float)) + loadx];
s_train[lidx * block_size + lidy] = train[min(t * block_size + lidy, train_rows - 1) * (step / sizeof(float)) + loadx];
}
barrier(CLK_LOCAL_MEM_FENCE);
/* there are threee types in the reducer. the first is L1Dist, which to sum the abs(v1, v2), the second is L2Dist, which to
sum the (v1 - v2) * (v1 - v2), the third is humming, which to popc(v1 ^ v2), popc is to count the bits are set to 1*/
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
const int trainIdx = t * block_size + lidx;
if (queryIdx < query_rows && trainIdx < train_rows /*&& mask(queryIdx, trainIdx)*/)
{
if (result < myBestDistance1)
{
myBestDistance2 = myBestDistance1;
myBestTrainIdx2 = myBestTrainIdx1;
myBestDistance1 = result;
myBestTrainIdx1 = trainIdx;
}
else if (result < myBestDistance2)
{
myBestDistance2 = result;
myBestTrainIdx2 = trainIdx;
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
__local float *s_distance = (__local float *)sharebuffer;
__local int *s_trainIdx = (__local int *)(sharebuffer + block_size * block_size);
//findBestMatch
s_distance += lidy * block_size;
s_trainIdx += lidy * block_size;
s_distance[lidx] = myBestDistance1;
s_trainIdx[lidx] = myBestTrainIdx1;
float bestDistance1 = MAX_FLOAT;
float bestDistance2 = MAX_FLOAT;
int bestTrainIdx1 = -1;
int bestTrainIdx2 = -1;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
bestTrainIdx2 = bestTrainIdx1;
bestDistance1 = val;
bestTrainIdx1 = s_trainIdx[i];
}
else if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
barrier(CLK_LOCAL_MEM_FENCE);
s_distance[lidx] = myBestDistance2;
s_trainIdx[lidx] = myBestTrainIdx2;
barrier(CLK_LOCAL_MEM_FENCE);
if (lidx == 0)
{
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance2)
{
bestDistance2 = val;
bestTrainIdx2 = s_trainIdx[i];
}
}
}
myBestDistance1 = bestDistance1;
myBestDistance2 = bestDistance2;
myBestTrainIdx1 = bestTrainIdx1;
myBestTrainIdx2 = bestTrainIdx2;
if (queryIdx < query_rows && lidx == 0)
{
bestTrainIdx[queryIdx] = (int2)(myBestTrainIdx1, myBestTrainIdx2);
bestDistance[queryIdx] = (float2)(myBestDistance1, myBestDistance2);
}
}
kernel void BruteForceMatch_calcDistanceUnrolled(
__global float *query,
__global float *train,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,
int max_desc_len,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType)
{
/* Todo */
}
kernel void BruteForceMatch_calcDistance(
__global float *query,
__global float *train,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,
int query_rows,
int query_cols,
int train_rows,
int train_cols,
int step,
int distType)
{
/* Todo */
}
kernel void BruteForceMatch_findBestMatch(
__global float *allDist,
__global int *bestTrainIdx,
__global float *bestDistance,
int k,
int block_size
)
{
/* Todo */
}

1
modules/ocl/src/kernels/build_warps.cl → modules/ocl/src/opencl/build_warps.cl
View File

@ -234,4 +234,3 @@ __kernel
map_y[y * step_y + x] = ycoo;
}
}

0
modules/ocl/src/kernels/convertC3C4.cl → modules/ocl/src/opencl/convertC3C4.cl
View File

0
modules/ocl/src/kernels/cvt_color.cl → modules/ocl/src/opencl/cvt_color.cl
View File

0
modules/ocl/src/kernels/filter_sep_col.cl → modules/ocl/src/opencl/filter_sep_col.cl
View File

2
modules/ocl/src/kernels/filter_sep_row.cl → modules/ocl/src/opencl/filter_sep_row.cl
View File

@ -466,5 +466,3 @@ __kernel __attribute__((reqd_work_group_size(LSIZE0,LSIZE1,1))) void row_filter_
dst[start_addr] = sum;
}
}

0
modules/ocl/src/kernels/filtering_boxFilter.cl → modules/ocl/src/opencl/filtering_boxFilter.cl
View File

0
modules/ocl/src/kernels/filtering_laplacian.cl → modules/ocl/src/opencl/filtering_laplacian.cl
View File

0
modules/ocl/src/kernels/filtering_morph.cl → modules/ocl/src/opencl/filtering_morph.cl
View File

4
modules/ocl/src/kernels/haarobjectdetect.cl → modules/ocl/src/opencl/haarobjectdetect.cl
View File

@ -559,7 +559,3 @@ if(result)
}
}
*/

1
modules/ocl/src/kernels/haarobjectdetect_scaled2.cl → modules/ocl/src/opencl/haarobjectdetect_scaled2.cl
View File

@ -283,4 +283,3 @@ __kernel void gpuscaleclassifier(global GpuHidHaarTreeNode *orinode, global GpuH
newnode[counter].alpha[0] = t1.alpha[0];
newnode[counter].alpha[1] = t1.alpha[1];
}

0
modules/ocl/src/kernels/imgproc_bilateral.cl → modules/ocl/src/opencl/imgproc_bilateral.cl
View File

0
modules/ocl/src/kernels/imgproc_calcHarris.cl → modules/ocl/src/opencl/imgproc_calcHarris.cl
View File

0
modules/ocl/src/kernels/imgproc_calcMinEigenVal.cl → modules/ocl/src/opencl/imgproc_calcMinEigenVal.cl
View File

0
modules/ocl/src/kernels/imgproc_canny.cl → modules/ocl/src/opencl/imgproc_canny.cl
View File

0
modules/ocl/src/kernels/imgproc_columnsum.cl → modules/ocl/src/opencl/imgproc_columnsum.cl
View File

2
modules/ocl/src/kernels/imgproc_convolve.cl → modules/ocl/src/opencl/imgproc_convolve.cl
View File

@ -107,5 +107,3 @@ __kernel void convolve_D5 (__global float *src, __global float *temp1, __global
dst[gy*(dst_step >> 2)+gx] = res;
}
}

0
modules/ocl/src/kernels/imgproc_copymakeboder.cl → modules/ocl/src/opencl/imgproc_copymakeboder.cl
View File

1
modules/ocl/src/kernels/imgproc_histogram.cl → modules/ocl/src/opencl/imgproc_histogram.cl
View File

@ -267,4 +267,3 @@ __kernel __attribute__((reqd_work_group_size(256,1,1)))void equalizeHist(
}
}
*/

0
modules/ocl/src/kernels/imgproc_integral.cl → modules/ocl/src/opencl/imgproc_integral.cl
View File

0
modules/ocl/src/kernels/imgproc_integral_sum.cl → modules/ocl/src/opencl/imgproc_integral_sum.cl
View File

1
modules/ocl/src/kernels/imgproc_median.cl → modules/ocl/src/opencl/imgproc_median.cl
View File

@ -484,4 +484,3 @@ __kernel void medianFilter5_C1_D5(__global float * src, __global float * dst, i
dst[dstOffset + get_global_id(1)*dstStep + get_global_id(0)]=p12;
}
#undef op(a,b)

3
modules/ocl/src/kernels/imgproc_remap.cl → modules/ocl/src/opencl/imgproc_remap.cl
View File

@ -980,6 +980,3 @@ __kernel void remapLNF1Constant_C4_D5(__global float * dst, __global float const
}
}

1
modules/ocl/src/kernels/imgproc_resize.cl → modules/ocl/src/opencl/imgproc_resize.cl
View File

@ -411,4 +411,3 @@ __kernel void resizeNN_C4_D5(__global float4 * dst, __global float4 * src,
dst[dpos] = src[spos];
}

1
modules/ocl/src/kernels/imgproc_threshold.cl → modules/ocl/src/opencl/imgproc_threshold.cl
View File

@ -150,4 +150,3 @@ __kernel void threshold_C1_D5(__global const float * restrict src, __global floa
}
}
}

0
modules/ocl/src/kernels/imgproc_warpAffine.cl → modules/ocl/src/opencl/imgproc_warpAffine.cl
View File

1
modules/ocl/src/kernels/imgproc_warpPerspective.cl → modules/ocl/src/opencl/imgproc_warpPerspective.cl
View File

@ -682,4 +682,3 @@ __kernel void warpPerspectiveCubic_C4_D5(__global float4 * src, __global float4
}
}
}

0
modules/ocl/src/kernels/interpolate_frames.cl → modules/ocl/src/opencl/interpolate_frames.cl
View File

1
modules/ocl/src/kernels/match_template.cl → modules/ocl/src/opencl/match_template.cl
View File

@ -821,4 +821,3 @@ void matchTemplate_Prepared_CCOFF_NORMED_C4_D0
res[res_idx] = normAcc(num, denum);
}
}

1
modules/ocl/src/kernels/meanShift.cl → modules/ocl/src/opencl/meanShift.cl
View File

@ -240,4 +240,3 @@ __kernel void meanshiftproc_kernel( __global uchar4* in, __global uchar4* outr,
// outsp[basesp] =(short2)((short)x0,(short)y0);
}
}

0
modules/ocl/src/kernels/merge_mat.cl → modules/ocl/src/opencl/merge_mat.cl
View File

0
modules/ocl/src/kernels/moments.cl → modules/ocl/src/opencl/moments.cl
View File

0
modules/ocl/src/kernels/nonfree_surf.cl → modules/ocl/src/opencl/nonfree_surf.cl
View File

0
modules/ocl/src/kernels/objdetect_hog.cl → modules/ocl/src/opencl/objdetect_hog.cl
View File

0
modules/ocl/src/kernels/operator_convertTo.cl → modules/ocl/src/opencl/operator_convertTo.cl
View File

0
modules/ocl/src/kernels/operator_copyToM.cl → modules/ocl/src/opencl/operator_copyToM.cl
View File

0
modules/ocl/src/kernels/operator_setTo.cl → modules/ocl/src/opencl/operator_setTo.cl
View File

1
modules/ocl/src/kernels/operator_setToM.cl → modules/ocl/src/opencl/operator_setToM.cl
View File

@ -57,4 +57,3 @@ __kernel void set_to_with_mask(
}
}

0
modules/ocl/src/kernels/pyr_down.cl → modules/ocl/src/opencl/pyr_down.cl
View File

0
modules/ocl/src/kernels/pyr_up.cl → modules/ocl/src/opencl/pyr_up.cl
View File

0
modules/ocl/src/kernels/pyrlk.cl → modules/ocl/src/opencl/pyrlk.cl
View File

0
modules/ocl/src/kernels/pyrlk_no_image.cl → modules/ocl/src/opencl/pyrlk_no_image.cl
View File

0
modules/ocl/src/kernels/split_mat.cl → modules/ocl/src/opencl/split_mat.cl
View File

0
modules/ocl/src/kernels/stereobm.cl → modules/ocl/src/opencl/stereobm.cl
View File