list of missing features compared to mainline alpaka

[psychocoderHPC]

general

• random numbers Random number generation. #107 Uniform random numbers (refactored) #365 Normal distributed random numbers #386
• INTEL TBB support TBB backend implementation #321
• SYCL support Support one api intel gpu/cpu #58
• SYCL global device nemory refactor device global memory #340 support for copy from and to global device memory. #345
• host only compile (allow alloc,memset,...) without the backend compiler e.g. nvcc
• standalone header
  • In principle, this feature should not be required because it is simply possible to include alpaka/alpaka.hpp because backends are enabled in C++ and not via CMake options or preprocessor macros.
  • any example can be copied into godbolt: https://alpaka3.readthedocs.io/en/latest/basic/example.html#play-online-with-compiler-explorer
• cuda or hip only mode (question is, should we still support it, if yes why?)
• GitLab CI
• version numbers, including CMake add version number to alpaka3 #376

onHost

• events add event support #222
• wait(device) Add alpaka::wait(device) synchronization helper and tests #221
• non owning Views of data
• SubViews of data/Views
• enqueue host task on all backends host task enqueue for all backends #218
• OpenMP scheduling selection
• is it possible to write an queue which behaves like https://github.com/alpaka-group/alpaka/blob/25bb7b2e1fa29005e753883257d82e84e0610f09/test/unit/queue/src/CollectiveQueue.cpp#L20
• allocAsync() add onHost::allocAsync(), onHost::allocLikeAsync() and addDestructorAction() to ManagedView #192
• mapped memory allocation add onHost::allocMapped() #204 and managed memory add onHost::allocManaged() #203
• blocking queues Blocking queue implementation #229

onAcc

• warp operations Add Warp ops #339 warp operations #353
• threadfence Add memoryFence, the "memory ordering API", with support for different scopes #258
• atomic definition per hierarchy Allow execution of atomic functions as non-atomic within specialized … #360
  • we can currently not transform CpuOmpBlocks atomics between threads to a non atomic operation
• intrinsic popc Add intrinsic support under onAcc namespace #369

[mehmetyusufoglu]

random numbers: philox integer uniform generator and unifrom real transporter: #107
Box-Müller transformer from uniform to normal real distribution : #111

[mehmetyusufoglu]

Question: threadFence is used in alpaka for memory-visibility ordering for a thread's operations, it is not exactly a "fence" but strangely Cuda and C++ prefered name threadFence. Should i use memoryFence or continue using threadFence at alpaka3?

CUDA/HIP: __threadfence(), __threadfence_block(), __threadfence_system() (actually device-memory visibility order, not a barrier).
SYCL: sycl::atomic_fence(...)
C++: std::atomic_thread_fence(...)

[mehmetyusufoglu]

threadFence is used in alpaka for memory-visibility ordering for a thread's operations, it is not exactly a "fence" but strangely Cuda and C++ prefered name threadFence. Should i use memoryFence or continue using threadFence at alpaka3?

CUDA/HIP: __threadfence(), __threadfence_block(), __threadfence_system() (actually device-memory visibility order, not a barrier). SYCL: sycl::atomic_fence(...) C++: std::atomic_thread_fence(...)

Ok i will use threadFence but add a comment saying: "non‑blocking memory visibility fence."

[psychocoderHPC]

Question: threadFence is used in alpaka for memory-visibility ordering for a thread's operations, it is not exactly a "fence" but strangely Cuda and C++ prefered name threadFence. Should i use memoryFence or continue using threadFence at alpaka3?

CUDA/HIP: __threadfence(), __threadfence_block(), __threadfence_system() (actually device-memory visibility order, not a barrier). SYCL: sycl::atomic_fence(...) C++: std::atomic_thread_fence(...)

I would say keep the naming for now

[mehmetyusufoglu]

Intel TBB support by Rene and Me;
#321