[AIE2P] Combine VST.PUSH.CONV #351
Conversation
abhinay-anubola
requested review from
abnikant,
andcarminati,
F-Stuckmann,
gbossu,
katerynamuts,
khallouh,
konstantinschwarz,
martien-de-jong,
niwinanto,
SagarMaheshwari99 and
stephenneuendorffer
as code owners
abhinay-anubola
force-pushed
the
sanubola.combine.fifo.store.conv
branch
3 times, most recently
from
c96965c to
9effa8e
Compare
| for (MachineInstr &MI : MBB) { | ||
| if (AIEII->isFifoStoreConvOpcode(MI.getOpcode())) { | ||
| const Register DstReg = MI.getOperand(0).getReg(); | ||
| Impl(DstReg); |
There was a problem hiding this comment.
Do we also have to check, that there is only one user of DstReg, i.e. nobody is using the intermediate results of the conversion except for VST_PUSH?
There was a problem hiding this comment.
Yes we have to check, Iam checking that in AIE2PInstructionSelector.cpp
if (!canCombineCONV(StoreI, *ConvOp) ||
StoreI.getParent() != ConvOp->getParent() || !MRI.hasOneUse(ConvResult))
return false;
| %2:ptrregbank(p0) = IMPLICIT_DEF | ||
| %3:fiforegbank(<32 x s32>) = IMPLICIT_DEF | ||
| %4:gprregbank(s32) = IMPLICIT_DEF | ||
| %5:vregbank(<64 x s8>), %6:gprregbank(<8 x s8>) = G_INTRINSIC_W_SIDE_EFFECTS intrinsic(@llvm.aie2p.v64accfloat.to.v64bfp16ebs8), %1(<64 x s32>) |
There was a problem hiding this comment.
This seems to be testing the same intrinsic as the test below. I guess you meant to test aie2p.v64accfloat.to.v64bfp16ebs16 instead? Maybe adjust the test name to reflect the change as well.
There was a problem hiding this comment.
We have different stores in both tests aie2p.fifo.st.push.576.bfp16 and aie2p.fifo.st.push.544.bfp16 with same conv
| %8:mstfifo, %7:mpfs, %9:mr26_fifo_st = VST_PUSH_544_CONV_bfp16ebs16_fp32 %3, %1, %2, %4, implicit-def $srf2bflags, implicit-def $srfifo_of, implicit $crf2bmask, implicit $crrnd | ||
| %10:em = IMPLICIT_DEF | ||
| %11:mstfifo, %12:mpfs, %13:mr26_fifo_st = VST_FLUSH_512_fifo_1d_flush %8, %7, %9, %10, implicit-def $srfifo_of | ||
| %14:mstfifo, %15:mpfs, %16:mr26_fifo_st = VST_FLUSH_512_normal_flush %11, %12, %13, implicit-def $srfifo_of |
There was a problem hiding this comment.
Do we need the VST_FLUSH_512_normal_flush in the 1d/2d/3d tests? I feel this is only makes the test more complex
There was a problem hiding this comment.
There is case in Conv_bfp where VST_FLUSH_512_normal_flush is followed by VST_FLUSH_512_2D
llvm/test/CodeGen/AIE/aie2p/GlobalIsel/post-select-optimize-vst.flush.conv.mir
Show resolved
Hide resolved
abhinay-anubola
force-pushed
the
sanubola.combine.fifo.store.conv
branch
from
9effa8e to
19288c1
Compare
Add VST.PUSH.CONV combine in AIE2PInstructionSelector
Modify VST.FLUSH to VST.FLUSH.CONV in AIEPostSelectOptimize
Corresponding MIR tests were added