2a7ca101d7
Sparse-checkout from NVIDIA/cccl main branch to complete cccl_upstream: Added: - python/cuda_cccl/ (226 files) — Python bindings for device-level algorithms Critical for muh toolchain: cuda.compute.reduce_into, scan, radix_sort, etc. Includes 204 .py files with full test coverage for all 27 algorithms - ci/ (163 files) — Build/test infrastructure build_cub.sh, test_cub.sh, build_and_test_targets.sh, matrix.yaml Directly maps to our [INFRA-CI] and [INFRA-BUILD] items - .agent/skills/ (7 files) — NVIDIA's own agent skills for CCCL cccl-style/SKILL.md, cccl-test/SKILL.md, sass-diff/SKILL.md - docs/ (491 files) — Official CCCL documentation CI references, CMake guides, Python compute docs, libcudacxx PTX docs - test/ (12 files) — Top-level integration tests (cuda_smoke, stdpar) - Root configs: .clang-format, .clang-tidy, CONTRIBUTING.md, pyproject.toml - CLAUDE.md symlink → AGENTS.md (NVIDIA's standard) cccl_upstream now mirrors full NVIDIA/cccl structure: Before: 42M (cub + thrust + libcudacxx + cudax + c + examples + benchmarks) After: 53M (+python +ci +docs +.agent +test +configs) This completes the CCCL base needed for: - [muh-bench] items: ci/util/build_and_test_targets.sh for targeted builds - [CCCL-verify] items: python/cuda_cccl/tests/ as reference implementations - [CCCL-test] items: ci/test_cub.sh, ci/test_thrust.sh - Agent workflow: .agent/skills/ for consistent style and test patterns
107 lines
3.1 KiB
ReStructuredText
107 lines
3.1 KiB
ReStructuredText
.. _cccl-runtime-device:
|
|
|
|
Devices
|
|
========
|
|
|
|
:cpp:class:`cuda::device_ref`
|
|
-------------------------------
|
|
.. _cccl-runtime-device-device-ref:
|
|
|
|
:cpp:class:`cuda::device_ref` is a lightweight, non-owning handle to a CUDA device ordinal. It allows to query
|
|
information about a device and serves as an argument to other runtime APIs which are tied to a specific device.
|
|
It offers:
|
|
|
|
- ``get()``: native device ordinal
|
|
- ``name()``: device name
|
|
- ``init()``: initialize the device context
|
|
- ``peers()``: list peers for which peer access can be enabled
|
|
- ``has_peer_access_to(cuda::device_ref)``: query if peer access can be enabled to the given device
|
|
- ``attribute(attr)`` / ``attribute<::cudaDeviceAttr>()``: attribute queries
|
|
|
|
Availability: CCCL 3.1.0 / CUDA 13.1
|
|
|
|
:cpp:var:`cuda::devices`
|
|
----------------------------
|
|
.. _cccl-runtime-device-devices:
|
|
|
|
:cpp:var:`cuda::devices` is a random-access view of all available CUDA devices in the form of
|
|
:cpp:class:`cuda::device_ref` objects. It
|
|
provides indexing, size, and iteration for use
|
|
in range-based loops.
|
|
|
|
Availability: CCCL 3.1.0 / CUDA 13.1
|
|
|
|
Example:
|
|
|
|
.. code:: cpp
|
|
|
|
#include <cuda/devices>
|
|
#include <iostream>
|
|
|
|
void print_devices() {
|
|
for (auto& dev : cuda::devices) {
|
|
std::cout << "Device " << dev.get() << ": " << dev.name() << '\n';
|
|
}
|
|
}
|
|
|
|
Device attributes
|
|
-----------------
|
|
.. _cccl-runtime-device-attributes:
|
|
|
|
``cuda::device_attributes`` provides strongly-typed attribute query objects usable with
|
|
:cpp:func:`cuda::device_ref::attribute`. Selected examples:
|
|
|
|
- ``compute_capability``
|
|
- ``multiprocessor_count``
|
|
- ``concurrent_managed_access``
|
|
- ``clock_rate``
|
|
- ``numa_id``
|
|
|
|
Availability: CCCL 3.1.0 / CUDA 13.1
|
|
|
|
Example:
|
|
|
|
.. code:: cpp
|
|
|
|
#include <cuda/devices>
|
|
|
|
int get_max_blocks_on_device(cuda::device_ref dev) {
|
|
return cuda::device_attributes::multiprocessor_count(dev) * cuda::device_attributes::blocks_per_multiprocessor(dev);
|
|
}
|
|
|
|
:cpp:any:`cuda::arch_traits`
|
|
--------------------------------
|
|
.. _cccl-runtime-device-arch-traits:
|
|
|
|
Per-architecture trait accessors providing limits and capabilities common to all devices of an architecture.
|
|
Compared to ``cuda::device_attributes``, :cpp:any:`cuda::arch_traits` provide a compile-time accessible
|
|
structure that describes common characteristics of all devices of an architecture, while attributes are run-time
|
|
queries of a single characteristic of a specific device.
|
|
|
|
- :cpp:any:`cuda::arch_traits` and :cpp:any:`cuda::arch_traits_for` (compile-time and run-time forms).
|
|
- Returns a :cpp:struct:`cuda::arch_traits_t` with fields like
|
|
``max_threads_per_block``, ``max_shared_memory_per_block``, ``cluster_supported`` and other capability flags.
|
|
- Traits for the current architecture can be accessed with :cpp:func:`cuda::device::current_arch_traits`
|
|
|
|
Availability: CCCL 3.1.0 / CUDA 13.1
|
|
|
|
Example:
|
|
|
|
.. code:: cpp
|
|
|
|
#include <cuda/devices>
|
|
|
|
template <cuda::arch_id Arch>
|
|
__device__ void fn() {
|
|
auto traits = cuda::arch_traits<Arch>();
|
|
if constexpr (traits.cluster_supported) {
|
|
// cluster specific code
|
|
} else {
|
|
// non-cluster code
|
|
}
|
|
}
|
|
|
|
__global__ void kernel() {
|
|
fn<cuda::arch_id::sm_90>();
|
|
}
|