Add raja perf benchmarks #2
Conversation
johnbowen42
force-pushed
the
add-raja-perf
branch
from
02a01e5 to
ffb1085
Compare
driver.py
Outdated
| @@ -211,12 +220,10 @@ def build_and_run(self, reps, profiler=None): | |||
| or self.exemode == "jitify" | |||
| ), "Expected aot or proteus or jitify for exemode" | |||
|
|
|||
| exe = f"{self.benchmark}-{self.exemode}.x" | |||
| self.clean() | |||
| #self.clean() | |||
There was a problem hiding this comment.
IIRC for hecbench we have to clean first because otherwise different compilation modes (aot, proteus) pick up pre-built, incorrect object files.
I think we should reinstate cleaning and add a clean key in the per benchmark or benchmark grouping if possible.
| # the build command is meant to be a full bash command to build the benchmark, eg | ||
| # `cmake -DCMAKE_BUILD_TYPE=Debug --build` or `make benchmark` | ||
| # If none is provided, it will default to `make` | ||
| self.build_command = 'make' if build_command == None else build_command |
There was a problem hiding this comment.
Remove the default. User should define explicitly. We should use the TOML dictionary hierarchically, so user can provide a build command that applies to a group of programs.
driver.py
Outdated
| config["inputs"], | ||
| args.compiler, | ||
| args.proteus_path, | ||
| env_configs, | ||
| env_configs |
There was a problem hiding this comment.
Use Black formatter or Ruff. It requires the comma here.
| for env in env_configs: | ||
| env["PROTEUS_INSTALL_PATH"] = proteus_install |
There was a problem hiding this comment.
build instructions for RAJA perf
There was a problem hiding this comment.
I assume programs are commented out because you are testing?
raja-perf-benchmarks.toml
Outdated
| #[ATOMIC.nvidia.aot] | ||
| #path = "benchmarks/RAJAPerf" | ||
| #exe = "./build/bin/raja-perf.exe" | ||
| #args = "--exclude-variants Base_Seq Lambda_Seq RAJA_Seq Base_CUDA RAJA_CUDA " |
There was a problem hiding this comment.
Two questions:
- Why do we do
--exclude-variantsinstead of doing--variantto only run the desired one? - Why do you exclude
RAJA_CUDA? That's the one we want to run without proteus enabled (same for HIP)
There was a problem hiding this comment.
- Legacy, I can refactor but it's a little tedious
- I am comparing
RAJA_CUDA(which always uses proteus) toLambda_CUDAin these benchmarks, as a hack to avoid modifying the benchmarks heavily to have an extraRAJA_CUDA_JITvariant.Lambda_CUDAandRAJA_CUDAare roughly on parity (by design) for these benchmarks. This can be fixed later and we can add aRAJA_CUDA_JITvariant to RAJA perf
| @@ -0,0 +1,1065 @@ | |||
| [build] | |||
| [build.nvidia] | |||
There was a problem hiding this comment.
We need build configurations with/without proteus
There was a problem hiding this comment.
see the above comment, I don't think we do. Ultimately we will add forall_jit to RAJA, and a RAJA_CUDA_JIT/RAJA_HIP_JIT variant to RAJAPerf
No description provided.