Skip to content

[ISSUE-262, 85] Unfix the Compute Capabilities across CMakeLists, NVCC Compiler Options #921

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
stardriftfx wants to merge 5 commits into
base: SharedDevelopment
Choose a base branch
from
Open

[ISSUE-262, 85] Unfix the Compute Capabilities across CMakeLists, NVCC Compiler Options #921

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
7dc42dd
- updated nvcc compile flag to compile on local machine, if target no…
stardriftfx Mar 16, 2026
ed0faa7
Potential fix for pull request finding
stardriftfx Mar 17, 2026
0ff39e9
Apply suggestions from code review
stardriftfx Mar 17, 2026
9747fce
cuda enabled (not sure if this didnt get sent)
stardriftfx Mar 18, 2026
e9294d8
Merge branch '262-unfix-the-compute-capabilities-across-cmakelists' o…
stardriftfx Mar 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
26 changes: 17 additions & 9 deletions CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ cmake_minimum_required(VERSION 3.12)
#CONDITIONAL FLAG(to run simulation on CPU or GPU)
#
#For GPU:
#set(ENABLE_CUDA YES)
set(ENABLE_CUDA YES)
#
#For CPU:
#set(ENABLE_CUDA NO)
Expand All @@ -31,12 +31,13 @@ endif()
# to your desired architecture: \
#
#cmake -D ENABLE_CUDA=YES -D TARGET_ARCH=70 ..
#
#
#"YES" / GPU choice only available if CUDA library is installed and the GPU is CUDA capable.
#If no TARGET_ARCH is passed in then it will default to 37 which is the kepler architecture
#If TARGET_ARCH is not user-specified then it will default to native.
############################################################################################
if(NOT DEFINED TARGET_ARCH)
set(TARGET_ARCH 37)
# If target not specified by user, use local machine's GPU architecture
set(TARGET_ARCH "native")
endif()
Comment on lines 38 to 41
Copy link
Contributor Author

@stardriftfx stardriftfx Mar 17, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Tests passed successfully though.

Copy link
Contributor

@stiber stiber Mar 18, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This generates worrisome warnings that don't exist previously:

CMake Warning (dev) in CMakeLists.txt:
  Policy CMP0104 is not set: CMAKE_CUDA_ARCHITECTURES now detected for NVCC,
  empty CUDA_ARCHITECTURES not allowed.  Run "cmake --help-policy CMP0104"
  for policy details.  Use the cmake_policy command to set the policy and
  suppress this warning.

  CUDA_ARCHITECTURES is empty for target "Core".
This warning is for project developers.  Use -Wno-dev to suppress it.```


#CONDITIONAL FLAG to turn on the validation mode
Expand All @@ -50,14 +51,21 @@ if(ENABLE_CUDA)
message("\n----Generating Makefile for Graphitti GPU version----")
project(Graphitti LANGUAGES CXX CUDA C)
#Verify CUDA package is present
find_Package(CUDA REQUIRED)
find_package(CUDA REQUIRED)
#Set the USE_GPU preprocessor macro so that GPU code will be compiled.
add_compile_definitions(USE_GPU)
#Specify the CUDA architecture / gencode that will be targeted
### Set gencode and architecture variables to the correct values for your specific NVIDIA hardware
set(CMAKE_CUDA_ARCHITECTURES ${TARGET_ARCH})
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode=arch=compute_${TARGET_ARCH},code=sm_${TARGET_ARCH})

# Specify the CUDA architecture / gencode that will be targeted
# Set gencode and architecture variables to the correct values for your specific NVIDIA hardware
# Sets real and virtual architecture switch
# sm_${TARGET_ARCH} = Real architecture
# compute_${TARGET_ARCH} = Virtual architecture
# (embeds PTX code e.g. Parallel Thread Execution.
# If on a newer GPU, then the CUDA driver compiles a working binary)
set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS}
"-gencode=arch=compute_${TARGET_ARCH},code=sm_${TARGET_ARCH},compute_${TARGET_ARCH}")
message(STATUS "Using CUDA architecture: ${TARGET_ARCH}")


else()
message("\n----Generating Makefile for Graphitti CPU version----")
Expand Down
41 changes: 41 additions & 0 deletions docs/Developer/GPUArchLevels.md
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,41 @@
# GPU Architecture Levels
Copy link
Contributor

@stiber stiber Mar 18, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This doesn't appear to be linked in to the documentation.

Originally, Graphitti used CUDA Compute Capability (CC 3.5). This later changed to use a flexible architecture model, such that Graphitti can now be compiled for modern hardware (such as Volta or Lovelace). This allows the simulator to leverage modern GPU features while maintaining backwards compatibility for older cards by using conditional compilation.

# Supported Architectures

| Architecture | Compute Capability | Project Build Compatibility |
| :--- | :--- | :--- |
| **Kepler** | 3.5 / 3.7 | **Baseline**: Minimum version for backwards compatibility. |
| **Volta** | 7.0 | **Target**: Primary architecture for high-performance server runs. |
| **Ampere** | 8.0 / 8.6 | **Development**: Common for modern local development. |
| **Ada Lovelace**| 8.9 | **Current**: Latest generation available in the lab. |

## Compute Capability
* Binary Compatibility (cubin): Strictly backwards compatible. A binary for raiju (3.7) runs on ghidorah (8.9), but not vice versa.
- Backwards Compatibility: Any code compiled on an older architecture will work on newer ones.

- Forwards Compatibility (PTX): Any code compiled for a specific architecture will require that or a newer one to run.
- Parallel Thread Execution (PTX): PTX is included in lab builds.

# Specifying Target Architecture

By default, if not user-specified, `TARGET_ARCH` is set to `"native"`
which auto-detects and utilizes local hardware.

Legacy Support: -DTARGET_ARCH=35

Otachi Server: -DTARGET_ARCH=70

# Performance Notes
## Lab Servers Reference Table
Info retrieved from `nvidia-smi` (in terminal).

| Lab Server | GPU Model | Architecture | Compute Capability | Recommended `TARGET_ARCH` |
| :--- | :--- | :--- | :--- | :--- |
| **raiju** | Tesla K80 | **Kepler** | 3.7 | `37` |
| **otachi** | Tesla V100-PCIE-16GB | **Volta** | 7.0 | `70` |
| **ghidorah** | RTX 4500 Ada Generation | **Ada Lovelace** | 8.9 | `89` |

Note: Per project guidelines, conditional compilation adds structural complexity (code cruft). We only implement architecture-specific paths if they produce a measurable benefit.

- Example: Run a 5-10 minute simulation and then measure performance `nvidia-smi` and compare it to baseline results.
Loading