Bench/bowen/desul raja atomics #1624
Conversation
johnbowen42
force-pushed
the
bench/bowen/desul-raja-atomics
branch
from
dad5812
to
7d0da79
Compare
johnbowen42
force-pushed
the
bench/bowen/desul-raja-atomics
branch
from
7e171df
to
c4a9582
Compare
|
|
||
| template<int BLOCK_SZ> | ||
| struct ExecPolicyGPU { | ||
| using policy = RAJA::cuda_exec<BLOCK_SZ>; |
There was a problem hiding this comment.
This is a synchronous policy.
| void TimeAtomicOp(int num_iterations = 2, int array_size = 100) { | ||
| RAJA::Timer timer; | ||
|
|
||
| for (int i = 0; i < num_iterations; ++i) { |
There was a problem hiding this comment.
It might be good to not time the first run of the kernel. Then time multiple iterations of the loop, running asynchronously for gpus, instead of timing each loop individually.
|
|
||
| template<int BLOCK_SZ> | ||
| struct ExecPolicyGPU { | ||
| using policy = RAJA::cuda_exec<BLOCK_SZ>; |
There was a problem hiding this comment.
| using policy = RAJA::cuda_exec<BLOCK_SZ>; | |
| using policy = RAJA::cuda_exec<BLOCK_SZ, true /*asynchronous*/>; |
By default, RAJA exec policies are synchronous. To make them asynchronous, you add a template parameter with a bool true value.
| #include "desul/atomics.hpp" | ||
| #include "RAJA/util/Timer.hpp" | ||
|
|
||
| #define N 1000000000 |
There was a problem hiding this comment.
I think N should be a command line arg to select at run time.
| // GPU benchmarks | ||
| std::cout << "Executing CUDA benchmarks" << std::endl; | ||
| std::cout << INDENT << "Executing atomic add benchmarks" << std::endl; | ||
| TimeAtomicOp<ExecPolicyGPU<64>::policy, int, AtomicAdd<int, typename GPUAtomic::policy>, true>(4); |
There was a problem hiding this comment.
It may be cleaner and easier to work with to set a constexpr thread block size variable to a default value, such as 256, at the top of the file. I think we agreed in a recent group meeting that thread block size doesn't have a significant performance impact for the simple kernels in this file. @MrBurmark what do you think?
There was a problem hiding this comment.
That sounds good to me, then we could vary it fairly easily if we wanted to.
There was a problem hiding this comment.
And we could follow the pattern in RAJA Perf if we wanted to try block size sweeps.
There was a problem hiding this comment.
I added a for_each_type function into RAJA for just these kind of use cases https://github.com/LLNL/RAJA/blob/develop/include/RAJA/util/for_each.hpp#L88
Comparison of RAJA's implementation of atomic operations with desul