GitLab

Retune radix sort, run length encoding, reduce by key, scan, select if, and
histogram for SM70 and SM80.

Thrust will be switching to `cub::DeviceScan` to replace its custom scan
implementation. This patch addresses some issues found by the Thrust
tests:

- Initialize unused `BlockLoad` items with values known to be in the input
  set. This fixes the `TestInclusiveScanWithIndirection` Thrust test by
  keeping the `plus_mod3` functor indices valid.
- Use `OffsetT` instead of `int` to hold indicies in `AgentScan`. This
  fixes the `Test*ScanWithBigIndexes` Thrust tests by not truncating
  the input problem size.
- Use `BLOCK_[STORE|LOAD]_WARP_TRANSPOSED_TIMESLICED` instead of
  `BLOCK_[STORE|LOAD]_WARP_TRANSPOSED` when the intermediate type is
  larger than 128 bytes. This keeps shared memory buffers from growing
  too large in the `TestScanWithLargeTypes` Thrust test.

Specifically, this ports the following PRs to CUB:

- NVIDIA/thrust#1297 Add install rule option for add_subdirectory users.
- NVIDIA/thrust#1298 Enforce semantic versioning in CMake version configs.
- NVIDIA/thrust#1300 Use FindPackageHandleStandardArgs in CMake config.

https://wg21.link/P0571

so update the README accordingly.

*nix folks.

going forward.

github.com/nvlabs/cub and github.com/thrust/cub

Speeds up clang sorts with 32-bit keys by a factor of 1.5-2x.  No change
to 64-bit keys, and no change with nvcc.

The only usage of sm_count was removed in ce261ed6

Related to thrust/thrust#1244.

This is a packaging usecase, when only install rules are needed.

See thrust/thrust#1211.

Setting CMP0104 to OLD wasn't enough to suppress the errors about
this property empty being uninitialized. Setting the prop to OFF
explicitly silences the errors.

Only bump to 3.18 when C++17 is enabled.

CMake doesn't recognize CUDA_STANDARD=17 until v3.18, and the
CUDA_STANDARD_REQUIRED property doesn't seem to work properly. See
CMake bug: https://gitlab.kitware.com/cmake/cmake/-/issues/20953

To allow the C++17 configs to actually use C++17, we need to bump our
minimum CMake version to 3.18.

- Deprecation warnings are disabled due to use of deprecated CUDA APIs
  such as cudaBindTexture. Filed NVlabs/cub#191 to track this.
- Bump CMake minimum version to 3.15.
- Remove workarounds for old cmake/compilers.
- Switch to CUB_ENABLE_DIALECT_CPPXX instead of CMAKE_CXX_STANDARD.
  - Like Thrust multiconfig, all can be enabled at once.
  - Uses Thrust's settings when Thrust multiconfig is enabled.
- Add CUB_ENABLE_* options to disable portions of the build.
  - HEADER_TESTING, TESTING, EXAMPLES
- Add sm_37, sm_80, remove sm_30, sm_32.
  - When building CUB in Thrust, just use thrust's flags.
- Sanitize variable names in AppendOptionIfAvailable.
  - GCC was failing to detect `-Werror`. Long story short, this fixes it.
- Add cub.compiler_interface target to store warning suppressions, etc.
- Find and use Thrust. We should move the Thrust code into Thrust.
- Check for THOROUGH, QUICK and QUICKER variants of tests, create new
  targets for them when found.
- Rem...

A future release of NVC++ will report an error for local static variables
that are defined in device functions and host/device functions, even if
the variable is defined within a host-only region of the function.  To
avoid errors, move the local static variable that stores the device count
from cub::DeviceCount() to a separate host-only function.

Fix #1217

- Add a derivative of Thrust's development model document.

Formerly used the assignment operator to copy to uninitialized
memory. But a non-trivial assignment operator requires the destination
object be in a valid state. So use placement new to construct the item
on the uninitialized bits.

Partial fix for thrust issue 1153
Also a similar fix in thrust.