Incorrect Cost Calculations for Shuffle Ports on Icelake-client / Icelake-server

[Omnisip]

Bugzilla Link	48110
Version	trunk
OS	All
Blocks	#31672
CC	@adibiagio,@topperc,@RKSimon,@MattPD,@phoebewang,@rotateright

Extended Description

Hi there,

It looks like llvm-mca is treating icelake-client and icelake-server as having only one shuffle port. This causes incorrect cost calculations for any operation that would use said shuffle ports. A block diagram showing the architecture can be found here: https://en.wikichip.org/wiki/intel/microarchitectures/sunny_cove (sunny cove is the core used by both icelake-server and icelake-client). Similarly, you can find the uops usage for any given shuffle here: https://uops.info/table.html?search=shuf&cb_lat=on&cb_tp=on&cb_uops=on&cb_ports=on&cb_SKL=on&cb_ICL=on&cb_measurements=on&cb_base=on&cb_avx=on

If the code generating the costs for LLVM-MCA is used by anything else in the toolchain, it's likely this will yield a performance benefit for other users targeting icelake-server and icelake-client.

[topperc]

llvm-mca and the reset of llvm is using the skylake-server scheduling information for icelake-client and server.

[RKSimon]

This isn't an llvm-mca issue, it just shows the problem - as Craig said icelake doesn't have its own scheduler models.

Ideally we need someone who has the CPUs that can fork the skylake models, adjust them based off llvm-exegesis reports and the intel hw docs, and then confirm that it actually improves performance.

[RKSimon]

https://reviews.llvm.org/D108914

I've put up a candidate patch that creates a Icelake model as a direct copy of the existing SkylakeServer model - this would then allow iterative improvements to address microarch differences such as this.

[RKSimon]

https://reviews.llvm.org/D108914

I've put up a candidate patch that creates a Icelake model as a direct copy
of the existing SkylakeServer model - this would then allow iterative
improvements to address microarch differences such as this.

This landed at 9e2d14c

[RKSimon]

https://reviews.llvm.org/D115547 fixes some of the integer shuffles

[RKSimon]

https://reviews.llvm.org/D115752 fixes some of the fp shuffles

[RKSimon]

Fixed by 74d1fc7 and 67cce1c