[Mesa-dev] R600/SI: Intrinsics for derivatives
Tom Stellard
tom at stellard.net
Sat Jun 8 17:08:08 PDT 2013
On Fri, Jun 07, 2013 at 05:48:05PM -0700, Tom Stellard wrote:
> On Fri, Jun 07, 2013 at 05:24:42PM +0200, Michel Dänzer wrote:
> >
> > The most important difference to the previous version of these is that
> > whole quad mode is now enabled and M0 initialized appropriately for the
> > LDS instructions, which now allows all of the relevant piglit tests to
> > pass.
> >
>
> Hi Michel,
>
> After I gave this series my r-b, I was reviewing your Mesa patches, and
> I suddenly had an idea for a better way to implement this. See my
> comments below:
>
> > From bb5adcd52cc5cadc308e85f635675199f5c02f35 Mon Sep 17 00:00:00 2001
> > From: =?UTF-8?q?Michel=20D=C3=A4nzer?= <michel.daenzer at amd.com>
> > Date: Thu, 21 Feb 2013 17:56:22 +0100
> > Subject: [PATCH 3/3] R600/SI: Support AMDGPU.ddx/y intrinsics
> > MIME-Version: 1.0
> > Content-Type: text/plain; charset=UTF-8
> > Content-Transfer-Encoding: 8bit
> >
> > Use LDS for calculating the deltas between neighbouring pixels.
> >
> > Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
> > ---
> > lib/Target/R600/SIISelLowering.cpp | 77 +++++++++++++++++++++++++++++++++++++-
> > lib/Target/R600/SIISelLowering.h | 6 +++
> > lib/Target/R600/SIInstructions.td | 42 ++++++++++++++++++++-
> > 3 files changed, 121 insertions(+), 4 deletions(-)
> >
> > diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
> > index ac6a4c3..7ea226a 100644
> > --- a/lib/Target/R600/SIISelLowering.cpp
> > +++ b/lib/Target/R600/SIISelLowering.cpp
> > @@ -249,7 +249,7 @@ SDValue SITargetLowering::LowerFormalArguments(
> >
> > MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> > MachineInstr * MI, MachineBasicBlock * BB) const {
> > -
> > + MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
> > MachineBasicBlock::iterator I = *MI;
> >
> > switch (MI->getOpcode()) {
> > @@ -257,7 +257,6 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> > return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
> > case AMDGPU::BRANCH: return BB;
> > case AMDGPU::SI_ADDR64_RSRC: {
> > - MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
> > unsigned SuperReg = MI->getOperand(0).getReg();
> > unsigned SubRegLo = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
> > unsigned SubRegHi = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
> > @@ -282,10 +281,84 @@ MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
> > MI->eraseFromParent();
> > break;
> > }
> > + case AMDGPU::SI_DD:
> > + LowerSI_DD(MI, *BB, I, MRI);
> > + break;
> > + case AMDGPU::SI_TID:
> > + LowerSI_TID(MI, *BB, I, MRI);
> > + break;
> > }
> > return BB;
> > }
> >
> > +void SITargetLowering::LowerSI_DD(MachineInstr *MI, MachineBasicBlock &BB,
> > + MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
> > + unsigned coord0 = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
> > + unsigned coord1 = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
> > + MachineOperand dst = MI->getOperand(0);
> > + MachineOperand coord = MI->getOperand(1);
> > + MachineOperand ldsaddr = MI->getOperand(2);
> > + MachineOperand ldsaddr0 = MI->getOperand(3);
> > + MachineOperand ldsdelta = MI->getOperand(4);
> > +
> > + // Write this thread's coordinate to LDS
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::DS_WRITE_B32))
> > + .addOperand(coord)
> > + .addImm(0) // LDS
> > + .addOperand(ldsaddr)
> > + .addOperand(coord)
> > + .addOperand(coord)
> > + .addImm(0)
> > + .addImm(0);
> > +
> > + // Read top right / bottom left thread's coordinate from LDS
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::DS_READ_B32), coord0)
> > + .addImm(0) // LDS
> > + .addOperand(ldsaddr0)
> > + .addOperand(ldsaddr0)
> > + .addOperand(ldsaddr0)
> > + .addOperand(ldsdelta)
> > + .addImm(0);
> > +
> > + // Read top left thread's coordinate from LDS
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::DS_READ_B32), coord1)
> > + .addImm(0) // LDS
> > + .addOperand(ldsaddr0)
> > + .addOperand(ldsaddr0)
> > + .addOperand(ldsaddr0)
> > + .addImm(0)
> > + .addImm(0);
> > +
> > + // Subtract top left coordinate from top right / bottom left
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_SUB_F32_e32))
> > + .addOperand(dst)
> > + .addReg(coord0)
> > + .addReg(coord1);
> > +
> > + MI->eraseFromParent();
> > +}
> > +
> > +void SITargetLowering::LowerSI_TID(MachineInstr *MI, MachineBasicBlock &BB,
> > + MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const {
> > + unsigned mbcnt_lo = MRI.createVirtualRegister(&AMDGPU::VReg_32RegClass);
> > + MachineOperand dst = MI->getOperand(0);
> > +
> > + // Get this thread's ID
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_MBCNT_LO_U32_B32_e64), mbcnt_lo)
> > + .addImm(0xffffffff)
> > + .addImm(0)
> > + .addImm(0)
> > + .addImm(0)
> > + .addImm(0)
> > + .addImm(0);
> > + BuildMI(BB, I, BB.findDebugLoc(I), TII->get(AMDGPU::V_MBCNT_HI_U32_B32_e32))
> > + .addOperand(dst)
> > + .addImm(0xffffffff)
> > + .addReg(mbcnt_lo);
> > +
> > + MI->eraseFromParent();
> > +}
> > +
> > EVT SITargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
> > return MVT::i1;
> > }
> > diff --git a/lib/Target/R600/SIISelLowering.h b/lib/Target/R600/SIISelLowering.h
> > index 9b263b9..1a6538d 100644
> > --- a/lib/Target/R600/SIISelLowering.h
> > +++ b/lib/Target/R600/SIISelLowering.h
> > @@ -26,6 +26,12 @@ class SITargetLowering : public AMDGPUTargetLowering {
> >
> > SDValue LowerParameter(SelectionDAG &DAG, EVT VT, SDLoc DL,
> > SDValue Chain, unsigned Offset) const;
> > +
> > + void LowerSI_DD(MachineInstr *MI, MachineBasicBlock &BB,
> > + MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
> > + void LowerSI_TID(MachineInstr *MI, MachineBasicBlock &BB,
> > + MachineBasicBlock::iterator I, MachineRegisterInfo & MRI) const;
> > +
> > SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const;
> > SDValue LowerSIGN_EXTEND(SDValue Op, SelectionDAG &DAG) const;
> > SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) const;
> > diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
> > index d8fbf3e..1126729 100644
> > --- a/lib/Target/R600/SIInstructions.td
> > +++ b/lib/Target/R600/SIInstructions.td
> > @@ -903,8 +903,8 @@ defm V_MAC_F32 : VOP2_32 <0x0000001f, "V_MAC_F32", []>;
> > defm V_MADMK_F32 : VOP2_32 <0x00000020, "V_MADMK_F32", []>;
> > defm V_MADAK_F32 : VOP2_32 <0x00000021, "V_MADAK_F32", []>;
> > //defm V_BCNT_U32_B32 : VOP2_32 <0x00000022, "V_BCNT_U32_B32", []>;
> > -//defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
> > -//defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
> > +defm V_MBCNT_LO_U32_B32 : VOP2_32 <0x00000023, "V_MBCNT_LO_U32_B32", []>;
> > +defm V_MBCNT_HI_U32_B32 : VOP2_32 <0x00000024, "V_MBCNT_HI_U32_B32", []>;
> >
> > let isCommutable = 1, Defs = [VCC] in { // Carry-out goes to VCC
> > defm V_ADD_I32 : VOP2b_32 <0x00000025, "V_ADD_I32",
> > @@ -1110,6 +1110,24 @@ def LOAD_CONST : AMDGPUShaderInst <
> > [(set GPRF32:$dst, (int_AMDGPU_load_const imm:$src))]
> > >;
> >
> > +let usesCustomInserter = 1 in {
> > +
> > +def SI_DD : InstSI <
> > + (outs VReg_32:$dst),
> > + (ins VReg_32:$src, VReg_32:$lds_addr, VReg_32:$lds_addr0, i8imm:$ldsdelta),
> > + "SI_DD $src, $lds_addr, $lds_addr0, $ldsdelta",
> > + []
> > +>;
> > +
> > +def SI_TID : InstSI <
> > + (outs VReg_32:$dst),
> > + (ins),
> > + "SI_TID",
> > + []
> > +>;
> > +
> > +} // end usesCustomInserter
> > +
> > // SI Psuedo instructions. These are used by the CFG structurizer pass
> > // and should be lowered to ISA instructions prior to codegen.
> >
> > @@ -1544,6 +1562,26 @@ def : Pat <
> > sub3)
> > >;
> >
> > +class DDXY <Intrinsic name, bits<4> ldsdelta> : Pat <
> > + (name v4f32:$src, imm, imm, imm),
> > + (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)),
> > + (SI_DD (EXTRACT_SUBREG $src, sub0), (V_LSHLREV_B32_e32 2, (SI_TID)),
> > + (V_AND_B32_e32 0xfffffff0, (V_LSHLREV_B32_e32 2, (SI_TID))),
> > + ldsdelta), sub0),
> > + (SI_DD (EXTRACT_SUBREG $src, sub1), (V_LSHLREV_B32_e32 2, (SI_TID)),
> > + (V_AND_B32_e32 0xfffffff0, (V_LSHLREV_B32_e32 2, (SI_TID))),
> > + ldsdelta), sub1),
> > + (SI_DD (EXTRACT_SUBREG $src, sub2), (V_LSHLREV_B32_e32 2, (SI_TID)),
> > + (V_AND_B32_e32 0xfffffff0, (V_LSHLREV_B32_e32 2, (SI_TID))),
> > + ldsdelta), sub2),
> > + (SI_DD (EXTRACT_SUBREG $src, sub3), (V_LSHLREV_B32_e32 2, (SI_TID)),
> > + (V_AND_B32_e32 0xfffffff0, (V_LSHLREV_B32_e32 2, (SI_TID))),
> > + ldsdelta), sub3)
> > +>;
>
> Based on this pattern, I don't think you need to use a ddx/ddy intrinsic
> here. All of the instructions you are lowering DDX/DDY to have an
> equivalent LLVM IR instruction or LLVM intrinsic.
>
> For the DS_READ and DS_WRITE instructions all you need to do is emit
> load/stores to the local address space and then add patterns for those
> int the backend. As an added bonus this will add support for OpenCL
> local address spaces. I think the rest of the instructions are pretty straight
> forward (unless I've overlooked something). Let me know if you have any
> questions.
I did overlook something. You will need to add an intrinsic for thread
id in order to implement ddx/ddy completely in LLVM IR, but I still
think it is the best way.
-Tom
> > +
> > +def : DDXY<int_AMDGPU_ddx, 4>;
> > +def : DDXY<int_AMDGPU_ddy, 8>;
> > +
> > def : Pat <
> > (i32 (sext i1:$src0)),
> > (V_CNDMASK_B32_e64 (i32 0), (i32 -1), $src0)
> > --
> > 1.8.3
> >
>
> > _______________________________________________
> > llvm-commits mailing list
> > llvm-commits at cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
More information about the mesa-dev
mailing list