iree: Compilation error in linalg-tensors matmul test
Manifested in one build run on Windows only - may be an indicator of non-determinism.
2021-01-07T00:04:00.7571560Z [3108/3308] Generating check_llvm-ir_llvm_matmul.mlir_module.module
2021-01-07T00:04:00.7572709Z FAILED: iree/test/e2e/linalg_tensor_ops/check_llvm-ir_llvm_matmul.mlir_module.module
2021-01-07T00:04:00.7576590Z cmd.exe /C "cd /D D:\a\iree\iree\iree-build\iree\test\e2e\linalg_tensor_ops && D:\a\iree\iree\iree-build\iree\tools\iree-translate.exe -iree-mlir-to-vm-bytecode-module --iree-hal-target-backends=dylib-llvm-aot -iree-flow-dispatch-linalg-on-tensors -iree-flow-dispatch-linalg-on-tensors-tile-sizes=1,1 -iree-codegen-llvm-experimental-linalg-on-tensors D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir -o check_llvm-ir_llvm_matmul.mlir_module.module"
2021-01-07T00:04:00.7581181Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7582575Z %D = linalg.matmul ins(%A, %B: tensor<2x3xf32>, tensor<3x4xf32>)
2021-01-07T00:04:00.7583142Z ^
2021-01-07T00:04:00.7584389Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: note: see current operation: %2 = flow.dispatch.input.load %arg0, offsets = [%arg4, %c0], sizes = [%1, %c3], strides = [%c1, %c1] : !flow.dispatch.input<2x3xf32> -> tensor<?x3xf32>
2021-01-07T00:04:00.7586492Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7587868Z %D = linalg.matmul ins(%A, %B: tensor<2x3xf32>, tensor<3x4xf32>)
2021-01-07T00:04:00.7588433Z ^
2021-01-07T00:04:00.7589721Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: note: see current operation: %8 = flow.dispatch.input.load %arg1, offsets = [%c0, %arg5], sizes = [%c3, %7], strides = [%c1, %c1] : !flow.dispatch.input<3x4xf32> -> tensor<3x?xf32>
2021-01-07T00:04:00.7591789Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7593696Z %D = linalg.matmul ins(%A, %B: tensor<2x3xf32>, tensor<3x4xf32>)
2021-01-07T00:04:00.7594273Z ^
2021-01-07T00:04:00.7595533Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: note: see current operation: %13 = flow.dispatch.input.load %arg2, offsets = [%arg4, %arg5], sizes = [%1, %7], strides = [%c1, %c1] : !flow.dispatch.input<2x4xf32> -> tensor<?x?xf32>
2021-01-07T00:04:00.7597580Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: remark: unable to materialize shape calculation (unsupported op 'linalg.matmul'?): falling back to runtime resolution
2021-01-07T00:04:00.7598817Z %D = linalg.matmul ins(%A, %B: tensor<2x3xf32>, tensor<3x4xf32>)
2021-01-07T00:04:00.7599372Z ^
2021-01-07T00:04:00.7600553Z D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir:8:8: note: see current operation: %19 = linalg.matmul {__internal_linalg_transform__ = "workgroup"} ins(%6, %12 : tensor<?x3xf32>, tensor<3x?xf32>) outs(%18 : tensor<?x?xf32>) -> tensor<?x?xf32>
2021-01-07T00:04:00.7602064Z PLEASE submit a bug report to https://bugs.llvm.org/ and include the crash backtrace.
2021-01-07T00:04:00.7602867Z Stack dump:
2021-01-07T00:04:00.7606318Z 0. Program arguments: D:\\a\\iree\\iree\\iree-build\\iree\\tools\\iree-translate.exe -iree-mlir-to-vm-bytecode-module --iree-hal-target-backends=dylib-llvm-aot -iree-flow-dispatch-linalg-on-tensors -iree-flow-dispatch-linalg-on-tensors-tile-sizes=1,1 -iree-codegen-llvm-experimental-linalg-on-tensors D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir -o check_llvm-ir_llvm_matmul.mlir_module.module
2021-01-07T00:04:00.7609873Z 1. 2. In-Flight Diagnostics:
2021-01-07T00:04:00.7611020Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): error: expected dispatch function to have a single call operation
2021-01-07T00:04:00.7612411Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): error: expected dispatch function to have a single call operation
2021-01-07T00:04:00.7614153Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7616289Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7618322Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): remark: unable to materialize shape calculation (unsupported op 'flow.dispatch.input.load'?): falling back to runtime resolution
2021-01-07T00:04:00.7620279Z loc("D:/a/iree/iree/main_checkout/iree/test/e2e/linalg_tensor_ops/matmul.mlir":8:8): remark: unable to materialize shape calculation (unsupported op 'linalg.matmul'?): falling back to runtime resolution
About this issue
- Original URL
- State: closed
- Created 3 years ago
- Comments: 23 (16 by maintainers)
Commits related to this issue
- Fix #4409 — committed to nicolasvasilache/iree by deleted user 3 years ago
- Fix #4409 — committed to nicolasvasilache/iree by deleted user 3 years ago
- Bump benefit of pattern to convert from func -> llvm.func in ConvertToLLVM2 Bumping the pattern makes the pattern in IREE trigger before a similar pattern in MLIR core that is added to the pattern li... — committed to MaheshRavishankar/iree by deleted user 3 years ago
- Bump benefit of pattern to convert from func -> llvm.func in ConvertToLLVM2 (#4572) Bumping the pattern makes the pattern in IREE trigger before a similar pattern in MLIR core that is added to the p... — committed to iree-org/iree by MaheshRavishankar 3 years ago
Ahh good catch - that would explain the differences between compilers. I did notice there’s an
LLVM_REVERSE_ITERATION:BOOL
flag that causes all unordered containers to be iterated in reverse order and that may be useful to have on a CI to catch issues like this vs. how we are doing it now by just compiling with different compilers that iterate differently 😛