[Microbenchmarks] Add benchmark for conditional scalar assignment autovec #295
Conversation
…ovec Benchmarks with vs. without autovec for a loop containing conditional scalar assignment (plus a little extra arithmetic as a 'work payload').
|
Microbenchmark for FindLast/CSA autovec, as requested on llvm/llvm-project#158088 With just the conditional assignment in the loop, there was no noticeable performance difference. However, when I added a small arithmetic payload I saw a noticeable difference, especially for uint8t. |
MacDue
left a comment
MacDue
left a comment
There was a problem hiding this comment.
Generally seems reasonable to me (bar a few nits), but I've not added a benchmark before, so wait and see if there's any more comments.
| // Pick out-of-range default value. | ||
| T Result = 101; |
There was a problem hiding this comment.
Took a moment to see that "out-of-range" here was referring to the range of the input A values. Could you clarify that?
| @@ -0,0 +1,118 @@ | |||
| #include <iostream> | |||
There was a problem hiding this comment.
Was going to comment about the license header, but it seems that's not done here (looking at other files).
There was a problem hiding this comment.
Yeah, I wondered about that too.
| run_csa_autovec(&A[0], &B[0], &C[0], Threshold); | ||
| benchmark::DoNotOptimize(A); | ||
| benchmark::DoNotOptimize(B); | ||
| benchmark::DoNotOptimize(C); | ||
| benchmark::ClobberMemory(); |
There was a problem hiding this comment.
Not sure if it makes a difference, but other benchmarks seem to do these first:
| run_csa_autovec(&A[0], &B[0], &C[0], Threshold); | |
| benchmark::DoNotOptimize(A); | |
| benchmark::DoNotOptimize(B); | |
| benchmark::DoNotOptimize(C); | |
| benchmark::ClobberMemory(); | |
| benchmark::DoNotOptimize(A); | |
| benchmark::DoNotOptimize(B); | |
| benchmark::DoNotOptimize(C); | |
| benchmark::ClobberMemory(); | |
| run_csa_autovec(&A[0], &B[0], &C[0], Threshold); |
There was a problem hiding this comment.
MathFunctions.cpp and RuntimeChecks.cpp seem to run the test function first as well, so there doesn't seem to be agreement on this.
| // for 'A' in init_data below. | ||
| T Result = 101; | ||
| for (unsigned i = 0; i < ITERATIONS; i++) { | ||
| // Do some work to make the difference noticeable |
There was a problem hiding this comment.
could you add a few more variations, like the minimal case with just a CAS and multiple independent CAS?
| } | ||
| } | ||
|
|
||
| // Add add auto-vectorized and disabled vectorization benchmarks for math |
There was a problem hiding this comment.
The comment needs updating, currently passes only ty and Threshold, but it might be helpful to also pass a function if it helps to reduce the duplication for additional patterns
3 participants
Benchmarks with vs. without autovec for a loop containing conditional
scalar assignment (plus a little extra arithmetic as a 'work payload').