AMDGPU: Add AMDGPUSubtarget::getEUsPerCU()

[nhaehnle]

Summary

We will use this for more accurate occupancy computations. Note that
IsaInfo takes WGP mode vs. CU mode into account on gfx10+.

[foad]

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

213

Since "CU" is misleading, can you think of a better name for "unit onto which workgroups are mapped"?

[nhaehnle]

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

213

I was wondering this as well and am open to suggestions.

Work Group Processor obviously doesn't work.

Work Group Compute Unit (WGCU)? Work Group Unit (WGU)? Work Group Execution Unit (WGEU)? (And make the EU a Wave Execution Unit?)

[foad]

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

213

It's a shame there isn't an existing term. Your ideas sound OK to me, but none of them is so obviously great that we should immediately adopt it (IMO).

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

213

Should include the expanded forms of these acronyms in the comments. Also /// for doxygen I think?

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

protected:
   bool HasVOP3PInsts = false;
   bool HasMulI24 = true;
   bool HasMulU24 = true;
   bool HasSMulHi = false;
   bool HasInv2PiInlineImm = false;
   bool HasFminFmaxLegacy = true;
   bool EnablePromoteAlloca = false;
   bool HasTrigReducedRange = false;
+  unsigned EUsPerCU = 4;
   unsigned MaxWavesPerEU = 10;
   unsigned LocalMemorySize = 0;
   char WavefrontSizeLog2 = 0;
 
 public:
   AMDGPUSubtarget(const Triple &TT);
 
   static const AMDGPUSubtarget &get(const MachineFunction &MF);
public:
   unsigned getWavefrontSizeLog2() const {
     return WavefrontSizeLog2;
   }
 
   unsigned getLocalMemorySize() const {
     return LocalMemorySize;
   }
 
+  /// Number of SIMDs/EUs (execution units) per "CU" ("compute unit"), where the
+  /// "CU" is the unit onto which workgroups are mapped. This takes WGP mode vs.
+  /// CU mode into account.
+  unsigned getEUsPerCU() const { return EUsPerCU; }
+
   Align getAlignmentForImplicitArgPtr() const {
     return isAmdHsaOS() ? Align(8) : Align(4);
   }
 
   /// Returns the offset in bytes from the start of the input buffer
   ///        of the first explicit kernel argument.
   unsigned getExplicitKernelArgOffset(const Function &F) const {
     switch (TargetTriple.getOS()) {

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     TargetTriple(TT),
     TargetID(*this),
     InstrItins(getInstrItineraryForCPU(GPU)),
     InstrInfo(initializeSubtargetDependencies(TT, GPU, FS)),
     TLInfo(TM, *this),
     FrameLowering(TargetFrameLowering::StackGrowsUp, getStackAlignment(), 0) {
   // clang-format on
   MaxWavesPerEU = AMDGPU::IsaInfo::getMaxWavesPerEU(this);
+  EUsPerCU = AMDGPU::IsaInfo::getEUsPerCU(this);
   CallLoweringInfo.reset(new AMDGPUCallLowering(*getTargetLowering()));
   InlineAsmLoweringInfo.reset(new InlineAsmLowering(getTargetLowering()));
   Legalizer.reset(new AMDGPULegalizerInfo(*this, TM));
   RegBankInfo.reset(new AMDGPURegisterBankInfo(*this));
   InstSelector.reset(new AMDGPUInstructionSelector(
   *this, *static_cast<AMDGPURegisterBankInfo *>(RegBankInfo.get()), TM));
 }