| Commit-Queue | +1 |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
📍 Job linux-r350-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/13fda419490000
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
📍 Job mac-m4-pro-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/142f9cc9490000
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
pinpoint doesn't seem super happy about this CL, in particular on x64...
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
pinpoint doesn't seem super happy about this CL, in particular on x64...
I haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Sam Parker-Haynespinpoint doesn't seem super happy about this CL, in particular on x64...
I haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
Not just sync-fs:
navier-stokes -3.8%
sync-fs -3.1%
gaussian-blur -2.8%
js-tokens -2.2%
threejs -1.6%
mandreel -1.4%
espree-wtb -1.2%
If you have access to the pinpoint links, then you can select "base: ..." where it says "Select a reference column" at the top, the sort the "exp" column, and then all lines with a red down arrow are regressions, and lines with a green up arrow are improvements (and the other lines you can kinda ignore: their changes are not statistically significant).
Worth noting that I'm running pinpoint with `--turbo-instruction-scheduling`, which is disabled by default. If the point of this CL was to improve builtins (since --turbo-instruction-scheduling is enabled by default in mksnapshot), then the pinpoints I've ran are useless to show whether this CL is effective or not.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Sam Parker-Haynespinpoint doesn't seem super happy about this CL, in particular on x64...
Darius MercadierI haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
Not just sync-fs:
navier-stokes -3.8%
sync-fs -3.1%
gaussian-blur -2.8%
js-tokens -2.2%
threejs -1.6%
mandreel -1.4%
espree-wtb -1.2%If you have access to the pinpoint links, then you can select "base: ..." where it says "Select a reference column" at the top, the sort the "exp" column, and then all lines with a red down arrow are regressions, and lines with a green up arrow are improvements (and the other lines you can kinda ignore: their changes are not statistically significant).
Worth noting that I'm running pinpoint with `--turbo-instruction-scheduling`, which is disabled by default. If the point of this CL was to improve builtins (since --turbo-instruction-scheduling is enabled by default in mksnapshot), then the pinpoints I've ran are useless to show whether this CL is effective or not.
Okay, thanks. So, it's two up and two down on the mac and mostly down on x64.
I'm not bothered about builtins really, I have a series of downstream patches that give me ~2% geomean improvement on a bunch of wasm codes. I have those features hidden behind an experimental flag to avoid this kind of trouble. I'm going to assume / hope that it's the change to `PopBestCandidate` that's causing the problem so I'll try hiding that behind a new flag.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Sam Parker-Haynespinpoint doesn't seem super happy about this CL, in particular on x64...
Darius MercadierI haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
Sam Parker-HaynesNot just sync-fs:
navier-stokes -3.8%
sync-fs -3.1%
gaussian-blur -2.8%
js-tokens -2.2%
threejs -1.6%
mandreel -1.4%
espree-wtb -1.2%If you have access to the pinpoint links, then you can select "base: ..." where it says "Select a reference column" at the top, the sort the "exp" column, and then all lines with a red down arrow are regressions, and lines with a green up arrow are improvements (and the other lines you can kinda ignore: their changes are not statistically significant).
Worth noting that I'm running pinpoint with `--turbo-instruction-scheduling`, which is disabled by default. If the point of this CL was to improve builtins (since --turbo-instruction-scheduling is enabled by default in mksnapshot), then the pinpoints I've ran are useless to show whether this CL is effective or not.
Okay, thanks. So, it's two up and two down on the mac and mostly down on x64.
I'm not bothered about builtins really, I have a series of downstream patches that give me ~2% geomean improvement on a bunch of wasm codes. I have those features hidden behind an experimental flag to avoid this kind of trouble. I'm going to assume / hope that it's the change to `PopBestCandidate` that's causing the problem so I'll try hiding that behind a new flag.
Would you mind kicking off the performance runs again please?
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
Sam Parker-Haynespinpoint doesn't seem super happy about this CL, in particular on x64...
Darius MercadierI haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
Sam Parker-HaynesNot just sync-fs:
navier-stokes -3.8%
sync-fs -3.1%
gaussian-blur -2.8%
js-tokens -2.2%
threejs -1.6%
mandreel -1.4%
espree-wtb -1.2%If you have access to the pinpoint links, then you can select "base: ..." where it says "Select a reference column" at the top, the sort the "exp" column, and then all lines with a red down arrow are regressions, and lines with a green up arrow are improvements (and the other lines you can kinda ignore: their changes are not statistically significant).
Worth noting that I'm running pinpoint with `--turbo-instruction-scheduling`, which is disabled by default. If the point of this CL was to improve builtins (since --turbo-instruction-scheduling is enabled by default in mksnapshot), then the pinpoints I've ran are useless to show whether this CL is effective or not.
Sam Parker-HaynesOkay, thanks. So, it's two up and two down on the mac and mostly down on x64.
I'm not bothered about builtins really, I have a series of downstream patches that give me ~2% geomean improvement on a bunch of wasm codes. I have those features hidden behind an experimental flag to avoid this kind of trouble. I'm going to assume / hope that it's the change to `PopBestCandidate` that's causing the problem so I'll try hiding that behind a new flag.
Would you mind kicking off the performance runs again please?
Done (https://pinpoint-dot-chromeperf.appspot.com/job/10d53823490000 and https://pinpoint-dot-chromeperf.appspot.com/job/12473935490000; which should notify on this CL once they finish)
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
📍 Job linux-r350-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/12473935490000
📍 Job mac-m4-mini-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/10d53823490000
Sam Parker-Haynespinpoint doesn't seem super happy about this CL, in particular on x64...
Darius MercadierI haven't really looked a pinpoint before... is it just sync-fs for x64 that is the problem?
Sam Parker-HaynesNot just sync-fs:
navier-stokes -3.8%
sync-fs -3.1%
gaussian-blur -2.8%
js-tokens -2.2%
threejs -1.6%
mandreel -1.4%
espree-wtb -1.2%If you have access to the pinpoint links, then you can select "base: ..." where it says "Select a reference column" at the top, the sort the "exp" column, and then all lines with a red down arrow are regressions, and lines with a green up arrow are improvements (and the other lines you can kinda ignore: their changes are not statistically significant).
Worth noting that I'm running pinpoint with `--turbo-instruction-scheduling`, which is disabled by default. If the point of this CL was to improve builtins (since --turbo-instruction-scheduling is enabled by default in mksnapshot), then the pinpoints I've ran are useless to show whether this CL is effective or not.
Sam Parker-HaynesOkay, thanks. So, it's two up and two down on the mac and mostly down on x64.
I'm not bothered about builtins really, I have a series of downstream patches that give me ~2% geomean improvement on a bunch of wasm codes. I have those features hidden behind an experimental flag to avoid this kind of trouble. I'm going to assume / hope that it's the change to `PopBestCandidate` that's causing the problem so I'll try hiding that behind a new flag.
Darius MercadierWould you mind kicking off the performance runs again please?
Done (https://pinpoint-dot-chromeperf.appspot.com/job/10d53823490000 and https://pinpoint-dot-chromeperf.appspot.com/job/12473935490000; which should notify on this CL once they finish)
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
😿 Job mac-m4-mini-perf/jetstream-main.crossbench failed.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/14a92dfd490000
📍 Job linux-r350-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/135eb1fd490000
So, it looks like the issues are resolved on x64, but any idea why the mac build failed?
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
I restarted the mac pinpoint, maybe it was just a fluke.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
📍 Job mac-m4-mini-perf/jetstream-main.crossbench complete.
See results at: https://pinpoint-dot-chromeperf.appspot.com/job/17b4358b490000
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Code-Review | +1 |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
[compiler] Instruction scheduler tweaks
- Don't add function parameter nops to the graph, just schedule them.
- Introduce data_successors, and use them to compute total latencies.
- Introduce a tie-breaking heuristic to PopBestCandidate, controlled via
the new flag: --experimental-turbo-instruction-scheduling.
- Introduce tracing.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |