Bump llvm - transform.apply_registered_pass options params

[rolfmorel]

• [mlir][NFC] Simplify constant checks with isOneInteger and renamed isZeroInteger. llvm/llvm-project#139340

sed -i "s/isZeroIndex/isZeroInteger/g" **/*.h
sed -i "s/isZeroIndex/isZeroInteger/g" **/*.cpp

• [MLIR] Add bufferization state to getBufferType and resolveConflicts interface methods llvm/llvm-project#141466 & [MLIR] Add bufferization state class to OneShotBufferization pass llvm/llvm-project#141019

  • Add BufferizationState &state to bufferize and getBuffer

• [MLIR][Transform] apply_registered_pass op's options as a dict llvm/llvm-project#143159 & [MLIR][Transform] Allow ApplyRegisteredPassOp to take options as a param llvm/llvm-project#142683 & [MLIR][Transform] apply_registered_op fixes: arg order & python options auto-conversion llvm/llvm-project#143779

  • Updates to transform.apply_registered_pass and its Python-bindings

• [mlir][scf] Return replacements explicitly in SCFTilingResult. llvm/llvm-project#143217

  • tilingResult->mergeResult.replacements -> tilingResult->replacements

• [mlir][amx] Simplify intrinsic generation llvm/llvm-project#140559 & [mlir][amx] Restore conversion interface for AMX llvm/llvm-project#143871

  • Change CHECK lines which expected amx. ops to llvm.call_intrinsics & fix which enables conversion again.

[rolfmorel]

CI now fails with

error: failed to legalize operation 'cf.br' that was explicitly marked illegal
ERROR: Failed to lower IR to LLVM dialect

for the following for four tests:

Failed Tests (4):
  TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-gemm.mlir
  TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-mlp.mlir
  TPP_OPT :: BF16/Integration/avx512bf16/vector-contract-to-amx-gemm.mlir
  TPP_OPT :: BF16/Integration/avx512bf16/vector-contract-to-amx-mlp.mlir

However, on the compile server all the tests pass.

@shahidact , @arun-thmn , @adam-smnk , @rengolin - any clue what the issue might be?

[arun-thmn]

@rolfmorel
The below two tests should be removed as these tests targets amx and in wrong avx512bf16dp folder.
TPP_OPT :: BF16/Integration/avx512bf16/vector-contract-to-amx-gemm.mlir
TPP_OPT :: BF16/Integration/avx512bf16/vector-contract-to-amx-mlp.mlir

The other two tests skipped (not passed) on no amx (compile node) because it's target is AMX machine. On the login node spr it get invoked and on compile node no amx it get skipped.
TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-gemm.mlir
TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-mlp.mlir

[rolfmorel]

Okay. Removed the duplicated tests. The following two tests still fail on EMR (on the compile node and login node they are skipped) with the message error: failed to legalize operation 'cf.br' that was explicitly marked illegal -- ERROR: Failed to lower IR to LLVM dialect.

TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-gemm.mlir
TPP_OPT :: BF16/Integration/amx/vector-contract-to-amx-mlp.mlir

Looking at the IR the failure to convert/legalize the cf.br is related to !amx.tile<..> types not having been converted to !llvm.x86_amx:

    %145 = builtin.unrealized_conversion_cast %144 : !llvm.x86_amx to !amx.tile<16x16xf32>
    %146 = llvm.mul %101, %93 : i64
    %147 = llvm.add %146, %95 : i64
    %148 = llvm.getelementptr %115[%147] : (!llvm.ptr, i64) -> !llvm.ptr, f32
    %149 = llvm.call_intrinsic "llvm.x86.tileloadd64.internal"(%91, %90, %148, %89) : (i16, i16, !llvm.ptr, i64) -> !llvm.x86_amx
    %150 = builtin.unrealized_conversion_cast %149 : !llvm.x86_amx to !amx.tile<16x16xf32>
    %151 = llvm.mul %101, %93 : i64
    %152 = llvm.add %151, %101 : i64
    %153 = llvm.getelementptr %115[%152] : (!llvm.ptr, i64) -> !llvm.ptr, f32
    %154 = llvm.call_intrinsic "llvm.x86.tileloadd64.internal"(%91, %90, %153, %89) : (i16, i16, !llvm.ptr, i64) -> !llvm.x86_amx
    %155 = builtin.unrealized_conversion_cast %154 : !llvm.x86_amx to !amx.tile<16x16xf32>
    cf.br ^bb9(%96, %140, %145, %150, %155 : index, !amx.tile<16x16xf32>, !amx.tile<16x16xf32>, !amx.tile<16x16xf32>, !amx.tile<16x16xf32>)
  ^bb9(%156: index, %157: !amx.tile<16x16xf32>, %158: !amx.tile<16x16xf32>, %159: !amx.tile<16x16xf32>, %160: !amx.tile<16x16xf32>):  // 2 preds: ^bb8, ^bb10
    %161 = builtin.unrealized_conversion_cast %156 : index to i64
    %162 = llvm.icmp "slt" %161, %99 : i64
    llvm.cond_br %162, ^bb10, ^bb11
  ^bb10:  // pred: ^bb9
    %163 = builtin.unrealized_conversion_cast %160 : !amx.tile<16x16xf32> to !llvm.x86_amx
    %164 = builtin.unrealized_conversion_cast %159 : !amx.tile<16x16xf32> to !llvm.x86_amx
    %165 = builtin.unrealized_conversion_cast %158 : !amx.tile<16x16xf32> to !llvm.x86_amx
    %166 = builtin.unrealized_conversion_cast %157 : !amx.tile<16x16xf32> to !llvm.x86_amx

@adam-smnk, any ideas?

[adam-smnk]

It looks like type conversion for within non-amx ops is missing after I axed AMXDialectLLVMIRTranslationInterface.
I'll see if I can somehow add the type conversion without bringing the whole thing back.

[adam-smnk]

This should be fixed after llvm/llvm-project#143871

[rengolin]

This should be fixed after llvm/llvm-project#143871

Indeed. Best to re-bump now than change the tests and bump again later.

[rolfmorel]

Thanks for the quick fix upstream, @adam-smnk !