NASP/cuda-samples
84
0
Fork 0
mirror of https://github.com/NVIDIA/cuda-samples.git synced 2026-10-13 00:08:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

CUDA 13.4 samples update - v13.4-public

Browse Source 
Release 13.4 of the CUDA Samples supported by CUDA Toolkit 13.4.
See Changelog for more information.
This commit is contained in:
Dheemanth
2026-09-09 15:07:08 -07:00
committed by GitHub
parent b7c5481c55
commit 5443602d89
871 changed files with 6843 additions and 321849 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
python/2_CoreConcepts/launchConfigTuning/README.md
View File

@@ -94,10 +94,10 @@ elapsed_ms = (end_event - start_event) / n_iterations
The benchmark loops in this sample read kernel results back from
`ManagedMemoryResource` allocations between launches, which requires the
device property `concurrent_managed_access=True`. This is only supported on
Linux with HMM (Pascal and newer). On Windows (WDDM/MCDM/TCC) the property
is `False`, so the sample exits early with a waive message and exit code
`2`.
device property `concurrent_managed_access=True`. Devices that report `False`
— Windows (WDDM/MCDM/TCC), and older integrated GPUs such as Jetson Orin —
keep managed allocations GPU-exclusive, so the sample reports the missing
capability and exits with code `2`.
## Installation

15
python/2_CoreConcepts/launchConfigTuning/launchConfigTuning.py
View File

@@ -341,14 +341,6 @@ def main():
3. Benchmarking different thread block configurations
4. Finding optimal threads-per-block for various kernel types
"""
if sys.platform == "win32":
print(
"This sample relies on ManagedMemoryResource with concurrent host "
"access, which is not supported on Windows "
"(concurrent_managed_access=False). Waiving this sample."
)
sys.exit(2)
print("=" * 60)
print("Launch Configuration Tuning (cuda.core)")
print("Finding the Best Block Size for Your Kernel")
@@ -358,6 +350,13 @@ def main():
device = Device(0)
device.set_current()
# This sample builds host NumPy views over managed memory. Devices without
# concurrent managed access keep managed allocations GPU-exclusive while
# the GPU is active, so those host views fault.
if not device.properties.concurrent_managed_access:
print("Concurrent managed memory access is not supported on this platform.")
sys.exit(2)
# Print GPU information
print_gpu_info(device)