-
Notifications
You must be signed in to change notification settings - Fork 17.9k
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.
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
cmd/compile: add ilp pass to help balance commutative expressions aiding in instruction level parallelism #60278
base: master
Are you sure you want to change the base?
Conversation
This PR (HEAD: e764f20) has been imported to Gerrit for code review. Please visit https://go-review.googlesource.com/c/go/+/496095 to see it. Tip: You can toggle comments from me using the |
Message from Wayne Zuo: Patch Set 1: (2 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Ryan Berger: Patch Set 1: (2 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Wayne Zuo: Patch Set 1: (1 comment) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Ryan Berger: Patch Set 1: (1 comment) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Keith Randall: Patch Set 1: (10 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from David Chase: Patch Set 1: (2 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Ryan Berger: Patch Set 1: (12 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Ryan Berger: Patch Set 1: (2 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
514a356
to
cbc1baa
Compare
This PR (HEAD: cbc1baa) has been imported to Gerrit for code review. Please visit Gerrit at https://go-review.googlesource.com/c/go/+/496095. Important tips:
|
Message from tricium-prod@appspot.gserviceaccount.com: Patch Set 2: (1 comment) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
Message from Ryan Berger: Patch Set 2: (12 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
This PR (HEAD: a79b267) has been imported to Gerrit for code review. Please visit Gerrit at https://go-review.googlesource.com/c/go/+/496095. Important tips:
|
Message from Keith Randall: Patch Set 3: (4 comments) Please don’t reply on this GitHub thread. Visit golang.org/cl/496095. |
…ions Currently the compiler groups expressions with commutative operations such as a + b + c + d as so: (a + (b + (c + d))) which is suboptimal for CPU instruction pipelining. This pass balances commutative expressions as shown above to (a + b) + (c + d) to optimally pipeline them. It also attempts to reassociate constants to as far right of the commutative expression as possible for better constant folding opportunities. Below is a benchmark from crypto/md5 on an MacBook Pro M2: trunk reassociate Hash1K-8 433.7Mi ± 0% 499.4Mi ± 4% +15.17% (p=0.000 n=10) Hash8K-8 454.3Mi ± 1% 524.9Mi ± 1% +15.53% (p=0.000 n=10) .... geomean 284.4Mi 327.5Mi +15.15% Other CPU architectures tried showed very little change (+/-1%) on this particular benchmark but tight mathematical code stands to gain greatly from this optimization Fixes golang#49331
This PR (HEAD: cdffffd) has been imported to Gerrit for code review. Please visit Gerrit at https://go-review.googlesource.com/c/go/+/496095. Important tips:
|
Currently the compiler groups expressions with commutative operations such as a + b + c + d as so: (a + (b + (c + d))) which is suboptimal for CPU instruction pipelining.
This pass balances commutative expressions as shown above to (a + b) + (c + d) to optimally pipeline them.
Below is a benchmark from crypto/md5 on an MacBook Pro M2:
Hash1K-8 433.7Mi ± 0% 499.4Mi ± 4% +15.17% (p=0.000 n=10)
Hash8K-8 454.3Mi ± 1% 524.9Mi ± 1% +15.53% (p=0.000 n=10)
....
geomean 284.4Mi 327.5Mi +15.15%
Other CPU architectures tried showed very little change (+/-1%) on this particular benchmark but tight mathematical code stands to gain greatly from this optimization
Fixes #49331