Feature wish list

[j-stephan]

In today's meeting we decided that we need a list of feature wishes that we can use for further development of alpaka. This issue will provide that overview and will be constantly open and pinned.

Please open new issues for any feature wish that you might have - they will subsequently be linked here.

Currently open wishes:

General API

• Improve wait API: add extended wait interface #32
• C++ concepts: Add support for concepts #286
• Assignment of ViewPlainPtr: ViewPlainPtr doesn't work as promised #655
• API simplification: simplification of the user interface #944
• Separate host-side and device-side Acc: Separate usage of Acc inside and outside of kernels #966
• Inform dynamic shared memory about work division: dynamic shared memory trait #993
• Vulkan as inspiration: [Meta] Consider studying Vulkan as API design input #1065
• Accelerator tags: Introduce accelerator tags #1246
• Tensor core abstractions: Add tensor core abstractions #1346

Backend-specific

Boost.Fiber

• parallelize fiber block execution: parallelize fiber block execution #22
• execute row by row: execute fibers row by row #34

CUDA/HIP

• improve automatic work division: improve automatic work division for CUDA devices #33
• Allow n-dimensional kernels with arbitrary n: CUDA full n-dimensional support #159
• hwloc for memory pinning: Use hwloc for CUDA independent memory pinning #634

General CPU

• Improve polling performance: polling: low performance with CPU accelerator and async streams #368

OpenMP

• Use OpenMP for both block- and thread-level parallelization: add an accelerator back-end that uses OpenMP 2 for block and thread parallelization #42

std::thread

• Parallelize over blocks: Make AccCpuThreads parallelize over blocks #1135

TBB

• Thread pinning: Adding thread pinning for the tbb back end #361

Build system / CI / Docs / Installation

• Precompiled headers: Precompiled Header Support #858
• Transition to HZDR CI: Renovating CI and usage of HZDR GitLab CI #967
• One CMake target per accelerator: CMake target for each accelerator #1019
• Improve Doxygen: [Discussion] Improving the Doxygen API #1059
• Improve examples: Improve readability of examples #1101
• Simplify CMake's accelerator names: Simplify accelerator names in CMake #1124

Device handling

• More device properties: extend device properties #21
• DeviceView / subdevices: add DeviceView concept #45

Host handling

• hwloc for querying host topology: use hwloc to query the topology of the host #36

Kernel handling

• Thread-safe kernel execution: make kernel execution thread safe #23

Kernel language

• alpaka equivalent of CUDA's cache hints (load data to the read-only cache): Add cache hints #1312
• No direct pointer access in kernels: abstract the memory access inside kernels #38
• Allocating static shared memory should not require an ID: Shared memory allocation without the use of an ID #81
• Fast math option for individual math functions: Math: User-Controlled Fast Math #297
• RNG should not be identical across different devices: random number generators not identical across accelerators #390
• Explicit vectorization inside kernels: Libraries for explicit vectorization that might be usable for the Alpaka element layer #652
• Counter-based RNG: Counter based random number generators? #710

Memory handling

• Copy only if different memory type: Add a method that copies memory only if it is in a different memory #28
• Unified memory: Add support for unified memory #296, Support for mapped/unified memory #612

[j-stephan]

Wish #1005 (concise atomic functions) was fulfilled by #1185.

[sbastrakov]

Secret Santa @bernhardmgruber