Fix dialect registration in neura interpreter

[TimJZ]

Summary

This PR fixes a dialect registration issue in neura-interpreter.

Previously, neura-interpreter could not parse Neura MLIR files generated from the normal LLVM-import/e2e flow because the imported module may contain dlti.dl_spec metadata. This caused the interpreter to fail before it could execute the Neura dataflow IR.

The failure looked like:

'dlti' attribute created with unregistered dialect
[neura-interpreter] Failed to parse MLIR input file

After this fix, neura-interpreter can parse the mapped ReLU IR generated by the existing mlir-neura-opt pipeline and proceed into dataflow execution.

Testing

Tested manually on the ReLU e2e lowering flow.

1. Rebuild the updated tools

cd ~/dev/neura
ninja -C build neura-interpreter mlir-neura-opt

2. Generate LLVM IR for the ReLU kernel

mkdir -p /tmp/neura-step2-relu
cd /tmp/neura-step2-relu

clang++ -S -emit-llvm -O3 \
  -fno-vectorize -fno-unroll-loops \
  -std=c++17 \
  -o relu_kernel_full.ll \
  ~/dev/neura/test/benchmark/CGRA-Bench/kernels/relu/relu_int.cpp

3. Extract the kernel function and import it into MLIR

llvm-extract --rfunc=".*kernel.*" \
  relu_kernel_full.ll \
  -o relu_kernel_only.ll

mlir-translate --import-llvm \
  relu_kernel_only.ll \
  -o relu_kernel_imported.mlir

mlir-translate may print warnings such as:

warning: unhandled data layout token: m:e
warning: unhandled data layout token: n8:16:32:64

These warnings are not fatal; relu_kernel_imported.mlir is still generated.

4. Lower and map the imported MLIR to Neura IR

mlir-neura-opt relu_kernel_imported.mlir \
  --assign-accelerator \
  --lower-llvm-to-neura \
  --promote-input-arg-to-const \
  --fold-constant \
  --canonicalize-return \
  --canonicalize-live-in \
  --leverage-predicated-value \
  --transform-ctrl-to-data-flow \
  --fold-constant \
  --insert-data-mov \
  --map-to-accelerator="mapping-strategy=heuristic" \
  --architecture-spec=$HOME/dev/neura/test/arch_spec/architecture.yaml \
  -o relu_neura_mapped.mlir

Confirmed that the mapped IR was generated:

ls -lh relu_neura_mapped.mlir

Observed:

-rw-r--r-- 1 gnahzmit gnahzmit 15K May 19 18:18 relu_neura_mapped.mlir

5. Verify the interpreter behavior

Before this fix, running the interpreter on the mapped ReLU IR failed during parsing:

neura-interpreter relu_neura_mapped.mlir

Previous error:

'dlti' attribute created with unregistered dialect
[neura-interpreter] Failed to parse MLIR input file

After this fix, the interpreter successfully parses the mapped ReLU IR and begins dataflow execution:

neura-interpreter relu_neura_mapped.mlir --verbose --dataflow

Observed output includes:

[neura-interpreter]  Initial pending operation: %5 = neura.reserve ...
[neura-interpreter]  Initial pending operation: %1 = "neura.grant_once"() ...
[neura-interpreter]  DFG Iteration 0 - Beginning
[neura-interpreter]  Executing operation: %5 = neura.reserve ...

This confirms that the dialect registration issue is fixed.

The interpreter now proceeds past MLIR parsing and exposes the next execution-level issue:

[neura-interpreter]  Executing operation: neura.yield ...
[neura-interpreter]  Unhandled op: neura.yield ...
[neura-interpreter]  Execution failed

That neura.yield handling is a separate follow-up issue and is not addressed by this PR.

[tancheng]

Can you add

// WIP: https://github.com/coredac/neura/issues/5
// RN: neura-interpreter %t-neura.mlir >> %t-dumped_output.txt

into a e2e test you wanna enable interpreter?

[TimJZ]

Added // WIP and // RN markers to test/e2e/relu/relu_kernel.mlir pointing to issue #5.