Skip to content

[Bug] generated C++ Cube kernel hangs with version >= 0.18 #428

Open
#443
Open
[Bug] generated C++ Cube kernel hangs with version >= 0.18#428
#443
Labels
bugSomething isn't working
@MirkoDeVita98

Description

@MirkoDeVita98
MirkoDeVita98
opened on Apr 2, 2026

Component

EmitC / Codegen (lib/PTO/Transforms/PTOToEmitC.cpp)

Description

The generated C++ in versions >= 0.18 appears to preserve the same overall math,
tensor views, and tensor layout for this kernel, but it changes the pipeline event
synchronization (set_flag / wait_flag) and tail synchronization behavior.

The version generated by 0.17 works for this reproducer. In that working output,
the event sequence is initialized before entering the loops, and the kernel ends
with explicit waits that drain the outstanding event dependencies.

In the >= 0.18 output, the synchronization schedule is different. In the failing
variant, the first main loop begins with:

wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);

but there is no obvious pre-loop:

set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);

before that first wait on EVENT_ID3.

I compared the MLIR from the older working version and the newer failing version,
and for this kernel they appear identical in substance. The generated C++ differs
mainly in synchronization/event scheduling, which suggests the regression is
introduced downstream during EmitC/codegen or sync insertion rather than in the
high-level IR.

Reproducer archive:
bug_reproduce.zip

Reproduction

To fully reproduce see: https://github.com/huawei-csl/pto-dsl/tree/main/examples/aot/fast_inverse/block_inversion

module {
  func.func @tri_inv_block2x2_fp16(%arg0: !pto.ptr<f32>, %arg1: !pto.ptr<f16>, %arg2: !pto.ptr<f16>, %arg3: i32) {
    pto.section.cube {
      %c0 = arith.constant 0 : index
      %c1 = arith.constant 1 : index
      %c64 = arith.constant 64 : index
      %c32 = arith.constant 32 : index
      %0 = arith.index_cast %arg3 : i32 to index
      %1 = arith.subi %0, %c1 : index
      %2 = pto.get_block_idx
      %3 = arith.index_cast %2 : i64 to index
      %4 = pto.get_block_num
      %5 = arith.index_cast %4 : i64 to index
      %6 = arith.muli %5, %c64 : index
      %7 = arith.muli %3, %c64 : index
      %8 = arith.addi %7, %c32 : index
      %9 = pto.make_tensor_view %arg1, shape = [%6, %c64], strides = [%c64, %c1] : !pto.tensor_view<?x?xf16>
      %10 = pto.make_tensor_view %arg0, shape = [%6, %c64], strides = [%c64, %c1] : !pto.tensor_view<?x?xf32>
      %11 = pto.make_tensor_view %arg2, shape = [%c32, %c32], strides = [%c32, %c1] : !pto.tensor_view<?x?xf16>
      %12 = pto.partition_view %11, offsets = [%c0, %c0], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf16> -> !pto.partition_tensor_view<32x32xf16>
      %13 = pto.partition_view %9, offsets = [%7, %c0], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf16> -> !pto.partition_tensor_view<32x32xf16>
      %14 = pto.partition_view %9, offsets = [%8, %c0], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf16> -> !pto.partition_tensor_view<32x32xf16>
      %15 = pto.partition_view %9, offsets = [%8, %c32], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf16> -> !pto.partition_tensor_view<32x32xf16>
      %16 = pto.partition_view %10, offsets = [%7, %c0], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf32> -> !pto.partition_tensor_view<32x32xf32>
      %17 = pto.partition_view %10, offsets = [%8, %c0], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf32> -> !pto.partition_tensor_view<32x32xf32>
      %18 = pto.partition_view %10, offsets = [%8, %c32], sizes = [%c32, %c32] : !pto.tensor_view<?x?xf32> -> !pto.partition_tensor_view<32x32xf32>
      %19 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %20 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %21 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %22 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %23 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %24 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %25 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %26 = pto.alloc_tile : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>
      %27 = pto.alloc_tile : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>
      %28 = pto.alloc_tile : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>
      %29 = pto.alloc_tile : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>
      pto.tload ins(%12 : !pto.partition_tensor_view<32x32xf16>) outs(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%25 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%19 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%21 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tload ins(%13 : !pto.partition_tensor_view<32x32xf16>) outs(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      scf.for %arg4 = %c0 to %1 step %c1 {
        pto.tmov ins(%19 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
        pto.tmov ins(%25 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
        pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
        pto.tmov ins(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
        pto.tmatmul.acc ins(%29, %27, %28 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
        %30 = arith.addi %arg4, %c1 : index
        %31 = arith.cmpi slt, %30, %1 : index
        scf.if %31 {
          pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%19 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
          pto.tmov ins(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
          pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
          pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%20 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
        }
      }
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%19 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tstore ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%16 : !pto.partition_tensor_view<32x32xf32>)
      pto.tload ins(%15 : !pto.partition_tensor_view<32x32xf16>) outs(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      scf.for %arg4 = %c0 to %1 step %c1 {
        pto.tmov ins(%21 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
        pto.tmov ins(%25 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
        pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
        pto.tmov ins(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
        pto.tmatmul.acc ins(%29, %27, %28 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>, !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
        %30 = arith.addi %arg4, %c1 : index
        %31 = arith.cmpi slt, %30, %1 : index
        scf.if %31 {
          pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%21 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
          pto.tmov ins(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
          pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
          pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%22 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
        }
      }
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%21 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tstore ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%18 : !pto.partition_tensor_view<32x32xf32>)
      pto.tload ins(%14 : !pto.partition_tensor_view<32x32xf16>) outs(%23 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%21 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%23 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%26 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%26 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%19 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tmov ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%26 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%24 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%27 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>)
      pto.tmov ins(%26 : !pto.tile_buf<loc=mat, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=512, pad=0>) outs(%28 : !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>)
      pto.tmatmul ins(%27, %28 : !pto.tile_buf<loc=left, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=row_major, fractal=512, pad=0>, !pto.tile_buf<loc=right, dtype=f16, rows=32, cols=32, v_row=32, v_col=32, blayout=row_major, slayout=col_major, fractal=512, pad=0>) outs(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>)
      pto.tstore ins(%29 : !pto.tile_buf<loc=acc, dtype=f32, rows=32, cols=32, v_row=32, v_col=32, blayout=col_major, slayout=row_major, fractal=1024, pad=0>) outs(%17 : !pto.partition_tensor_view<32x32xf32>)
    }
    return
  }
}

Use the MLIR snippet in the reproducer and run:

ptoas --enable-insert-sync inverse.pto -o inverse.cpp

Additional note:
The MLIR for the working and failing versions appears unchanged for this kernel.
The generated C++ differs mainly in event scheduling and tail synchronization.

Expected behavior

The generated C++ should preserve a valid event schedule for the kernel, matching
the behavior of the previously working version.

In particular:

  • every wait_flag(...) should have a matching earlier set_flag(...)
  • events used inside loops should be properly initialized before the first iteration
  • end-of-kernel synchronization should ensure all pending work is finished
  • enabling --enable-insert-sync should not introduce hangs, invalid waits, or
    nondeterministic behavior

Actual behavior / error logs

The generated C++ has a different synchronization schedule from the previously
working output.

In the failing generated version, the first main loop contains:

wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);

but there is no obvious earlier pre-loop:

set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);

before the first iteration.

The old working version did not have this issue and used a fully initialized and
explicitly drained event schedule.

This appears to cause invalid synchronization behavior in the generated kernel,
potentially resulting in:

  • hang / deadlock
  • dependence on stale event state
  • nondeterministic execution

Tail behavior also changed.

Old working version ended with:

pipe_barrier(PIPE_ALL);
wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
wait_flag(PIPE_FIX, PIPE_M, EVENT_ID2);
wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
wait_flag(PIPE_FIX, PIPE_M, EVENT_ID6);

The failing generated version instead ends with:

ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);

which lowers to a barrier-only path and no longer explicitly drains those event
dependencies in the same way.

Git commit

b46694d

Host platform

Linux (aarch64)

Target Ascend arch (if relevant)

a3

PTOAS build level (if relevant)

level2

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Type

    No type

    Projects

    pto project

    Status

    In Progress

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions