This is an archive of the discontinued LLVM Phabricator instance.

Differential D147145

[AMDGPU] Consistently use v2i16 for packed bf16 inputs. NFC.
ClosedPublic

Authored by foad on Mar 29 2023, 5:12 AM.

Details

Reviewers
rampitec
piotr
mariusz-sikora-at-amd
Group Reviewers
Restricted Project
Commits
rG70e2d631e178: [AMDGPU] Consistently use v2i16 for packed bf16 inputs. NFC.

Diff Detail

Event Timeline

foad created this revision.Mar 29 2023, 5:12 AM
Herald added a project: Restricted Project. · View Herald TranscriptMar 29 2023, 5:12 AM
Herald added subscribers: kosarev, StephenFan, kerbowa and 7 others. · View Herald Transcript
foad requested review of this revision.Mar 29 2023, 5:12 AM
Herald added a project: Restricted Project. · View Herald TranscriptMar 29 2023, 5:12 AM
Herald added subscribers: llvm-commits, wdng. · View Herald Transcript
foad added inline comments.Mar 29 2023, 5:16 AM
llvm/lib/Target/AMDGPU/FLATInstructions.td
735–736

FPAtomic is usually deduced from the type, so I override it here to ensure NFC.

However it is only used by the hazard detection code for "FPAtomicToDenormModeHazard" which is GFX10-only, and GFX10 does not have these instructions, so perhaps I should not bother to override FPAtomic here?

Harbormaster completed remote builds in B222486: Diff 509321.Mar 29 2023, 9:04 AM
rampitec accepted this revision.Mar 29 2023, 10:48 AM

LGTM

This revision is now accepted and ready to land.Mar 29 2023, 10:48 AM
This revision was landed with ongoing or failed builds.Mar 29 2023, 12:33 PM
Closed by commit rG70e2d631e178: [AMDGPU] Consistently use v2i16 for packed bf16 inputs. NFC. (authored by foad). · Explain Why
This revision was automatically updated to reflect the committed changes.
foad added a commit: rG70e2d631e178: [AMDGPU] Consistently use v2i16 for packed bf16 inputs. NFC..

Revision Contents

PathSize
llvm/
lib/
Target/
AMDGPU/
7 lines

Diff 509443

llvm/lib/Target/AMDGPU/FLATInstructions.td

Show First 20 Lines • Show All 726 Lines • ▼ Show 20 Lineslet SubtargetPredicate = isGFX90APlus in {
defm FLAT_ATOMIC_MAX_F64 : FLAT_Atomic_Pseudo<"flat_atomic_max_f64", VReg_64, f64>; defm FLAT_ATOMIC_MAX_F64 : FLAT_Atomic_Pseudo<"flat_atomic_max_f64", VReg_64, f64>;
defm GLOBAL_ATOMIC_ADD_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_add_f64", VReg_64, f64>; defm GLOBAL_ATOMIC_ADD_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_add_f64", VReg_64, f64>;
defm GLOBAL_ATOMIC_MIN_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_min_f64", VReg_64, f64>; defm GLOBAL_ATOMIC_MIN_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_min_f64", VReg_64, f64>;
defm GLOBAL_ATOMIC_MAX_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_max_f64", VReg_64, f64>; defm GLOBAL_ATOMIC_MAX_F64 : FLAT_Global_Atomic_Pseudo<"global_atomic_max_f64", VReg_64, f64>;
} // End SubtargetPredicate = isGFX90APlus } // End SubtargetPredicate = isGFX90APlus
let SubtargetPredicate = HasAtomicFlatPkAdd16Insts in { let SubtargetPredicate = HasAtomicFlatPkAdd16Insts in {
defm FLAT_ATOMIC_PK_ADD_F16 : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_f16", VGPR_32, v2f16>; defm FLAT_ATOMIC_PK_ADD_F16 : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_f16", VGPR_32, v2f16>;
defm FLAT_ATOMIC_PK_ADD_BF16 : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_bf16", VGPR_32, v2f16>; let FPAtomic = 1 in
defm FLAT_ATOMIC_PK_ADD_BF16 : FLAT_Atomic_Pseudo<"flat_atomic_pk_add_bf16", VGPR_32, v2i16>;
foadAuthorUnsubmitted
Done
ReplyInline Actions

FPAtomic is usually deduced from the type, so I override it here to ensure NFC.

However it is only used by the hazard detection code for "FPAtomicToDenormModeHazard" which is GFX10-only, and GFX10 does not have these instructions, so perhaps I should not bother to override FPAtomic here?

foad: FPAtomic is usually deduced from the type, so I override it here to ensure NFC. //However// it…
} // End SubtargetPredicate = HasAtomicFlatPkAdd16Insts } // End SubtargetPredicate = HasAtomicFlatPkAdd16Insts
let SubtargetPredicate = HasAtomicGlobalPkAddBF16Inst in let SubtargetPredicate = HasAtomicGlobalPkAddBF16Inst, FPAtomic = 1 in
defm GLOBAL_ATOMIC_PK_ADD_BF16 : FLAT_Global_Atomic_Pseudo<"global_atomic_pk_add_bf16", VGPR_32, v2f16>; defm GLOBAL_ATOMIC_PK_ADD_BF16 : FLAT_Global_Atomic_Pseudo<"global_atomic_pk_add_bf16", VGPR_32, v2i16>;
// GFX7-, GFX10-, GFX11-only flat instructions. // GFX7-, GFX10-, GFX11-only flat instructions.
let SubtargetPredicate = isGFX7GFX10GFX11 in { let SubtargetPredicate = isGFX7GFX10GFX11 in {
defm FLAT_ATOMIC_FCMPSWAP : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap", defm FLAT_ATOMIC_FCMPSWAP : FLAT_Atomic_Pseudo <"flat_atomic_fcmpswap",
VGPR_32, f32, v2f32, VReg_64>; VGPR_32, f32, v2f32, VReg_64>;
defm FLAT_ATOMIC_FMIN : FLAT_Atomic_Pseudo <"flat_atomic_fmin", defm FLAT_ATOMIC_FMIN : FLAT_Atomic_Pseudo <"flat_atomic_fmin",
▲ Show 20 Lines • Show All 1,662 Lines • Show Last 20 Lines