Notable User Facing Changes
---------------------------
- support for LLVM 13
- CMake: Inter-Procedural Optimization is enabled on code of runtime library
(libpocl.so is compiled with -flto on systems that support it).
- LTTng tracing improved - more command types are traced, and also
some synchronous API calls (like clCreateBuffer) are traced.
- poclcc, tests and examples can be disabled with CMake options
- Valgrind support improved by making Valgrind aware of pocl's
reference counting of cl_* objects
- kernels which are called by kernels are now force-inlined
- Support for NetBSD.
- Support for Unix systems without libdl.
- PoCL can now (optionally) respond to SIGUSR2 by printing
some live debug information.
- improved SPIR support for CUDA devices
Notable Bug Fixes
-----------------
- Fixed a potential crash on Unix systems without sysfs mounted.
- Fixed compilation errors when building on macOS.
- Fixed POCL_FAST_INIT macro; POCL_INIT_LOCK must be invoked with only one argument.
- Fix bin/poclcc to not depend on OpenCL 2.0 symbols
- Fixed miscompilation in kernel loops with multiple conditionals with barriers in them.
Other
-----
- Add cmake options PARALLEL_COMPILE_JOBS, PARALLEL_LINK_JOBS to
use ninja's seperate compile and link job pools.
- Improve memory architecture, buffer migration and allocation.
Buffers are now allocated on a device when first used
(previously each buffer was allocated on every device in context).
- the single global LLVMContext was replaced with
multiple LLVMContexts, one per OpenCL cl_context.
OpenCL code can now be compiled in parallel
when using separate cl_contexts. This feature
is disabled by default since it significantly slowed
down PyOpenCL. This should be resolved by separating
LLVM compilation in their own threads in the future.
- a new OpenCL extension was added to PoCL: cl_pocl_content_size.
The extension allows the user to give optimization hint to PoCL,
which will be used internally by PoCL to optimize buffer transfers
between multiple devices.
38dc6152c6