kAtomics Activity
Geometric kernel for CAD programs using an atomic approach
Status: Alpha
Brought to you by:
jrheinlaender
Activity for kAtomics
-
Jan Rheinlaender committed [38c492]
Create repository and add Code::Blocks GUI files
-
Add generic logging support with log4cpp
-
Implement Chunk classes and the createPrism method
-
Convert AtomicChunk to a struct for compatibility with OpenCL code
-
Fix bug in index creation for listchunks
-
Export geometry in simplified OpenInventor format for debugging purposes
-
Break up the Chunk class hierarchy into three specific classes (Geometry, ListChunk, AtomicChunk)
-
Update logger class
-
Initialize OpenCL and prepare contexts, queues and programs
-
Move utility methods to separate file
-
Rework ListChunk class to use OpenCL
-
Template the Point definition in utility.h and use it everywhere
-
Improve calculation of required allocation size for atomic chunks
-
Create new class ClDeviceQueue to collect all relevant OpenCl data in one structure
-
Cosmetic improvements
-
Two optimizations
-
Store chunk list in a plain vector instead of a hash map.
-
Code cosmetics and improved debug/timing information
-
Unify code for all 3 calculation methods
-
Use vector data types to simplify code
-
Try to unify the code to compile with different CPUs and GPUs without conditional compilation
-
Implement memory layout variant 1 without OpenCL (single-threaded)
-
Improve error handling in ClDriver
-
Miscellaneous fixes
-
Prism creation: Implement draft parallel algorithm
-
Prism creation: Improve draft parallel algorithm (memory usage)
-
Handle OpenCL intialization once for class ListChunk
-
Prism display: Precalculate all unique subchunks and re-use the nodes
-
Prism creation: Improve draft parallel algorithm (flat list for uniqueing step)
-
Prism creation: Improve draft parallel algorithm with memory access coalescence
-
Change namespace for utilities to Util::
-
Implement first part of prism creation in an OpenCL kernel
-
Silence C++ parser warnings
-
Use C interface for accessing OpenCL because C++ wrapper gives trouble with nVidia CUDA devices
-
Move OpenCL datatype definitions to a separate header
-
Detect CMake debug build and evaluate the define in the source code
-
Check maximum allowed buffer size before allocating it
-
Move OpenCL kernels to separate files
-
Implement ClBuffer class to wrap OpenCL buffer functionality for SVM und non-SVM devices
-
Implement kernel for prism creation
-
Improve enqueuing of kernels
-
Some code refactoring
-
Implement kernels for uniqueing and collecting of subchunks
-
Fix minor issues
-
Code cosmetics
-
The collection step after eliminating duplicate subchunks can be optimized away
-
Implement logic to pad and sanitize workitem and workgroup sizes
-
Rework kernels to handle splitting of the workitems into workgroups either by user request or by necessity (device cannot handle arbitrary workgroup sizes)
-
Implement benchmarking for createPrism()
-
Fix classes that manage resources to properly handle copying and moving
-
Keep track of whether a buffer is mapped / unmapped
-
Support two GPUs in ClDriver
-
Put all timing code into the BENCHMARK conditional
-
Listchunk: Ensure that buffers are correctly mapped / unmapped
-
Implement public header atomics.h that includes all relevant headers
-
Allow turning off OpenCL features (mainly usefull for debugging and benchmarking)
-
Implement some utility methods
-
Cleanup code and commit minor forgotten items
-
Improve timing and benchmarking in class geometry
-
Update CMakeLists.txt for compiling the benchmark
-
Default-initialize atomic chunks
-
Improve mapping and unmapping of buffers
-
Implement inversion of listchunks
-
Implement merging of listchunks with a boolean operation
-
Define type idx6_t for geometry
-
Implement merging of subchunks in geometry
-
Some bits and pieces
-
Fix two bugs that crashed the kernels
-
Cleanup
-
Fix error in kernels
-
Fix more bugs in kernels
-
Improve documentation / debugging
-
Improve listchunk merge and fix some bugs
-
Fix bugs and minor cleanup
-
Allow setting the dimensions of the geometry in the constructor
-
Implement subchunk merging in class Geometry
-
ClBuffer: Use std::allocator instead of malloc
-
Allow turning off CL for testing/debugging purposes
-
Geometry: Remove member currentChunk and use reserve() / resize() more efficiently instead
-
Geometry: Fix merge algorithm to make a copy of existing chunks before modifying them
-
Numerous minor fixes
-
Implement copy constructor and assignment operator for listchunks
-
Implement a method to copy a CL buffer
-
ClDriver: Implement method to get buffer size and use it to improve the code
-
Improve atomic chunk inversion/merging by a linear iteration
-
Improve mapping and resizing of CL buffers when creatinga prism in listchunk
-
Improve method for creating Open Inventor geometry, make it more compatible with FreeCAD
-
Implement simple method to return Open Inventor geometry of an atomic chunk
-
Improve method for merging listchunks
-
Improve method for creating Open Inventor geometry for list chunks
-
Minor fixes and forgotten items
-
Allow setting the granularity of Open Inventor rendering of geometry
-
Preserve some experimental and obsolete code
-
Implement simple method for getting the bounding box of geometry (experimental code)
-
Update test code
1
