Post v13.5 (#470)

This commit is contained in:
Dheemanth
2026-11-04 12:48:54 -08:00
committed by GitHub
parent 5443602d89
commit 0ec046d5e9
53 changed files with 1750 additions and 183 deletions

View File

@@ -2,7 +2,7 @@
## Description
This sample demonstrates efficient all-pairs simulation of a gravitational n-body simulation in CUDA. This sample accompanies the GPU Gems 3 chapter "Fast N-Body Simulation with CUDA". With CUDA 5.5, performance on Tesla K20c has increased to over 1.8TFLOP/s single precision. Double Performance has also improved on all Kepler and Fermi GPU architectures as well. Starting in CUDA 4.0, the nBody sample has been updated to take advantage of new features to easily scale the n-body simulation across multiple GPUs in a single PC. Adding "-numbodies=<bodies>" to the command line will allow users to set # of bodies for simulation. Adding “-numdevices=<N>” to the command line option will cause the sample to use N devices (if available) for simulation. In this mode, the position and velocity data for all bodies are read from system memory using “zero copy” rather than from device memory. For a small number of devices (4 or fewer) and a large enough number of bodies, bandwidth is not a bottleneck so we can achieve strong scaling across these devices.
This sample demonstrates efficient all-pairs simulation of a gravitational n-body simulation in CUDA. This sample accompanies the GPU Gems 3 chapter "Fast N-Body Simulation with CUDA". With CUDA 5.5, performance on Tesla K20c has increased to over 1.8TFLOP/s single precision. Double Performance has also improved on all Kepler and Fermi GPU architectures as well. Starting in CUDA 4.0, the nBody sample has been updated to take advantage of new features to easily scale the n-body simulation across multiple GPUs in a single PC. Adding "-numbodies=<bodies>" to the command line will allow users to set # of bodies for simulation. Adding “-numdevices=<N>” to the command line option will cause the sample to use N devices (if available) for simulation. In this mode, GPU 0 holds the authoritative position and velocity buffers. Each other device keeps a local scratch copy of the read-only position array and refreshes it with one peer-to-peer copy per simulation step, so the all-pairs kernel reads every tile from local memory instead of re-fetching it over the peer link on every tile access; each device still writes its own share of results directly back to GPU 0 over the peer link. Each device issues that copy and its kernel on its own non-blocking stream so the GPUs overlap. This keeps per-step peer traffic proportional to the number of devices instead of the number of thread blocks, so multi-GPU runs scale on both NVLink and PCIe-only (no NVLink) systems.
## Key Concepts
@@ -23,7 +23,7 @@ x86_64, armv7l
## CUDA APIs involved
### [CUDA Runtime API](http://docs.nvidia.com/cuda/cuda-runtime-api/index.html)
cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceCount, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaStreamQuery, cudaEventDestroy, cudaEventElapsedTime, cudaDeviceCanAccessPeer, cudaEventCreate
cudaGraphicsUnmapResources, cudaSetDeviceFlags, cudaGraphicsResourceSetMapFlags, cudaGraphicsResourceGetMappedPointer, cudaGraphicsMapResources, cudaSetDevice, cudaEventSynchronize, cudaGetDeviceCount, cudaGetDeviceProperties, cudaDeviceSynchronize, cudaEventRecord, cudaGetDevice, cudaMemcpyToSymbol, cudaEventDestroy, cudaEventElapsedTime, cudaDeviceCanAccessPeer, cudaDeviceEnablePeerAccess, cudaMemcpyPeerAsync, cudaStreamCreateWithFlags, cudaStreamDestroy, cudaStreamSynchronize, cudaEventCreate
## Dependencies needed to build/run
[X11](../../../README.md#x11), [OpenGL](../../../README.md#opengl), [Freeglut](../../../README.md#freeglut), [GLEW](../../../README.md#glew)