This is an archive of the discontinued LLVM Phabricator instance.

Differential D49934

AMDGPU: Allow fp32-denormals feaure for r600 targets
ClosedPublic

Authored by jvesely on Jul 27 2018, 1:38 PM.

Details

Reviewers
arsenm
tstellar
Commits
rG7f4ef00eb20e: Merging r338569: --------------------------------------------------------------…
rG5ba1b4bdab18: AMDGPU: Allow fp32-denormals feature for r600 targets
rL339103: Merging r338569:
rL338569: AMDGPU: Allow fp32-denormals feature for r600 targets
Summary

This was accidentally removed in r335942.

Diff Detail

Repository
rL LLVM

Event Timeline

jvesely created this revision.Jul 27 2018, 1:38 PM
Herald added subscribers: jfb, t-tye, tpr and 5 others. · View Herald TranscriptJul 27 2018, 1:38 PM
jvesely mentioned this in D49650: Targets/AMDGPU: Don't set fp32-denormals feature for r600.Jul 27 2018, 2:04 PM
jvesely mentioned this in D49649: AMDGPU/R600: Don't set fp32-denormals feature for r600.
arsenm accepted this revision.Jul 30 2018, 5:22 AM

LGTM

This revision is now accepted and ready to land.Jul 30 2018, 5:22 AM
arsenm added inline comments.Jul 30 2018, 5:23 AM
test/CodeGen/AMDGPU/target-features.ll
1–3

This isn't really a .ll or codegen test. I'm not sure where this should really go.

It also won't stay updated once more features are added, so maybe the test should just be dropped

Closed by commit rL338569: AMDGPU: Allow fp32-denormals feature for r600 targets (authored by jvesely). · Explain WhyAug 1 2018, 8:05 AM
This revision was automatically updated to reflect the committed changes.
jvesely added a comment.Aug 1 2018, 8:06 AM

Merged without the test. thanks

Revision Contents

PathSize
lib/
Target/
AMDGPU/
9 lines
9 lines
test/
CodeGen/
AMDGPU/
118 lines

Diff 157756

lib/Target/AMDGPU/AMDGPU.td

Context not available.
// Subtarget Features (options and debugging) // Subtarget Features (options and debugging)
//===------------------------------------------------------------===// //===------------------------------------------------------------===//
// Some instructions do not support denormals despite this flag. Using
// fp32 denormals also causes instructions to run at the double
// precision rate for the device.
def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
"FP32Denormals",
"true",
"Enable single precision denormal handling"
>;
// Denormal handling for fp64 and fp16 is controlled by the same // Denormal handling for fp64 and fp16 is controlled by the same
// config register when fp16 supported. // config register when fp16 supported.
// TODO: Do we need a separate f16 setting when not legal? // TODO: Do we need a separate f16 setting when not legal?
Context not available.

lib/Target/AMDGPU/AMDGPUFeatures.td

Context not available.
"Enable single precision FMA (not as fast as mul+add, but fused)" "Enable single precision FMA (not as fast as mul+add, but fused)"
>; >;
// Some instructions do not support denormals despite this flag. Using
// fp32 denormals also causes instructions to run at the double
// precision rate for the device.
def FeatureFP32Denormals : SubtargetFeature<"fp32-denormals",
"FP32Denormals",
"true",
"Enable single precision denormal handling"
>;
class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature< class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature<
"localmemorysize"#Value, "localmemorysize"#Value,
"LocalMemorySize", "LocalMemorySize",
Context not available.

test/CodeGen/AMDGPU/target-features.ll

  • This file was added.
; RUN: llc -march=amdgcn -mattr=help 2>&1 | FileCheck %s --check-prefix=GCN
; RUN: llc -march=r600 -mattr=help 2>&1 | FileCheck %s --check-prefix=R600
arsenmUnsubmitted
Not Done
ReplyInline Actions

This isn't really a .ll or codegen test. I'm not sure where this should really go.

It also won't stay updated once more features are added, so maybe the test should just be dropped

arsenm: This isn't really a .ll or codegen test. I'm not sure where this should really go. It also…
GCN: 16-bit-insts
GCN: DumpCode
GCN: GFX9
GCN: SEA_ISLANDS
GCN: SOUTHERN_ISLANDS
GCN: VOLCANIC_ISLANDS
GCN: add-no-carry-insts
GCN: amdgpu-debugger-emit-prologue
GCN: amdgpu-debugger-insert-nops
GCN: aperture-regs
GCN: auto-waitcnt-before-barrier
GCN: ci-insts
GCN: code-object-v3
GCN: d16-preserves-unused-bits
GCN: dl-insts
GCN: dpp
GCN: dumpcode
GCN: dx10-clamp
GCN: enable-ds128
GCN: fast-fmaf
GCN: flat-address-space
GCN: flat-for-global
GCN: flat-global-insts
GCN: flat-inst-offsets
GCN: flat-scratch-insts
GCN: fma-mix-insts
GCN: fmaf
GCN: fp-exceptions
GCN: fp16-denormals
GCN: fp32-denormals
GCN: fp64
GCN: fp64-denormals
GCN: fp64-fp16-denormals
GCN: gcn
GCN: gcn3-encoding
GCN: gfx9-insts
GCN: half-rate-64-ops
GCN: huge-private-buffer
GCN: int-clamp-insts
GCN: inv-2pi-inline-imm
GCN: isaver6.0.0
GCN: isaver6.0.1
GCN: isaver7.0.0
GCN: isaver7.0.1
GCN: isaver7.0.2
GCN: isaver7.0.3
GCN: isaver7.0.4
GCN: isaver8.0.1
GCN: isaver8.0.2
GCN: isaver8.0.3
GCN: isaver8.1.0
GCN: isaver9.0.0
GCN: isaver9.0.2
GCN: isaver9.0.4
GCN: isaver9.0.6
GCN: ldsbankcount16
GCN: ldsbankcount32
GCN: load-store-opt
GCN: localmemorysize0
GCN: localmemorysize32768
GCN: localmemorysize65536
GCN: mad-mix-insts
GCN: max-private-element-size-16
GCN: max-private-element-size-4
GCN: max-private-element-size-8
GCN: mimg-r128
GCN: movrel
GCN: promote-alloca
GCN: s-memrealtime
GCN: scalar-atomics
GCN: scalar-stores
GCN: sdwa
GCN: sdwa-mav
GCN: sdwa-omod
GCN: sdwa-out-mods-vopc
GCN: sdwa-scalar
GCN: sdwa-sdst
GCN: sgpr-init-bug
GCN: si-scheduler
GCN: trap-handler
GCN: unaligned-buffer-access
GCN: unaligned-scratch-access
GCN: unpacked-d16-vmem
GCN: unsafe-ds-offset-folding
GCN: vgpr-index-mode
GCN: vgpr-spilling
GCN: vop3p
GCN: wavefrontsize16
GCN: wavefrontsize32
GCN: wavefrontsize64
GCN: xnack
R600: EVERGREEN
R600: HasVertexCache
R600: NORTHERN_ISLANDS
R600: R600
R600: R600ALUInst
R600: R700
R600: caymanISA
R600: cfalubug
R600: dx10-clamp
R600: fetch16
R600: fetch8
R600: fmaf
R600: fp32-denormals
R600: fp64
R600: localmemorysize0
R600: localmemorysize32768
R600: localmemorysize65536
R600: promote-alloca
R600: wavefrontsize16
R600: wavefrontsize32
R600: wavefrontsize64