[DAPHNE-#768] isNan: elementwise check for NaN elements in a matrix - Draft PR

StoeckOverflow commented 4 months ago

This Draft PR introduces the isNan function to the Daphne project, enabling element-wise NaN checks on matrices (see issue #768) . The function has been implemented and successfully tested in C++ unit tests.

However, there is an issue with invoking the isNan function from a Daphne script, which needs to be resolved.

Implemented Features

[X] Created an IsNanOp operation in DaphneIR as a subclass of EwUnaryOp (elementwise unary operation) (see src/ir/daphneir/DaphneOps.td).
[X] Created a DaphneDSL built-in function isNan and documented it (see src/parser/daphnedsl/DaphneDSLBuiltins.cpp and doc/DaphneDSL/Builtins.md).
[X] Created a DaphneLib function isNan for matrices and scalars and documented it (see src/api/python/daphne/operator/nodes/matrix.py, src/api/python/daphne/operator/nodes/scalar.py, and doc/DaphneLib/APIRef.md).
[X] Created a runtime kernel for isNan by extending the existing ewUnaryMat and ewUnarySca kernel with a new op code (see src/runtime/local/kernels/EwUnaryMat.h).
[X] Added unit test cases for the kernel and script-level test cases for DaphneDSL and DaphneLib.

Issues

The current implementation works in C++ unit tests but fails when invoked from a Daphne script with the following error:


bin/daphne test/api/cli/operations/isNan_1.daphne

[error]: Lowering pipeline error.RewriteToCallKernelOpPass failed with the following message [ no kernels registered for operation `ewIsNan` ]
    | Source file -> "/home/domin/daphne/test/api/cli/operations/isNan_1.daphne":1:6
    |
  1 | print(isNan(3));
    |       ^~~

PassManager failed module lowering, responsible IR written to module_fail.log.

This is the module_fail.log:

module {
  func.func @main() {
    %0 = "daphne.constant"() {value = 3 : si64} : () -> si64
    %1 = "daphne.constant"() {value = true} : () -> i1
    %2 = "daphne.constant"() {value = false} : () -> i1
    %3 = "daphne.constant"() {value = 0x7FF8000000000000 : f64} : () -> f64
    %4 = "daphne.constant"() {value = 94808314720080 : ui64} : () -> ui64
    %5 = "daphne.constant"() {value = 1 : index} : () -> index
    %6 = "daphne.constant"() {value = 6 : index} : () -> index
    %7 = "daphne.constant"() {value = 94808303431520 : ui64} : () -> ui64
    %8 = "daphne.constant"() {value = 94808303431392 : ui64} : () -> ui64
    %9 = "daphne.constant"() {value = 140725113146728 : ui64} : () -> ui64
    %c0_i32 = arith.constant 0 : i32
    %10 = "daphne.call_kernel"(%9, %8, %7, %c0_i32) {callee = "_createDaphneContext__DaphneContext__uint64_t__uint64_t__uint64_t"} : (ui64, ui64, ui64, i32) -> !daphne.DaphneContext
    %11 = "daphne.createDaphneContext"(%9, %8, %7) : (ui64, ui64, ui64) -> !daphne.DaphneContext
    %12 = "daphne.ewIsNan"(%0) : (si64) -> si64
    "daphne.print"(%12, %1, %2) : (si64, i1, i1) -> ()
    %13 = "daphne.ewIsNan"(%3) : (f64) -> f64
    "daphne.print"(%13, %1, %2) : (f64, i1, i1) -> ()
    %14 = "daphne.matrixConstant"(%4) : (ui64) -> !daphne.Matrix<6x1xf64>
    %15 = "daphne.reshape"(%14, %6, %5) : (!daphne.Matrix<6x1xf64>, index, index) -> !daphne.Matrix<6x1xf64>
    "daphne.decRef"(%14) : (!daphne.Matrix<6x1xf64>) -> ()
    %16 = "daphne.transpose"(%15) : (!daphne.Matrix<6x1xf64>) -> !daphne.Matrix<1x6xf64>
    "daphne.decRef"(%15) : (!daphne.Matrix<6x1xf64>) -> ()
    %17 = "daphne.ewIsNan"(%16) : (!daphne.Matrix<1x6xf64>) -> !daphne.Matrix<1x6xf64>
    "daphne.decRef"(%16) : (!daphne.Matrix<1x6xf64>) -> ()
    "daphne.print"(%17, %1, %2) : (!daphne.Matrix<1x6xf64>, i1, i1) -> ()
    "daphne.decRef"(%17) : (!daphne.Matrix<1x6xf64>) -> ()
    "daphne.destroyDaphneContext"() : () -> ()
    "daphne.return"() : () -> ()
  }
}

Next Steps and Feedback

I am currently investigating the kernel registration issue and ensuring proper integration. Any insights or suggestions on this matter would be greatly appreciated. Thank you for your time and assistance in reviewing this PR.

pdamme commented 4 months ago

Hi @StoeckOverflow, thanks for this pull request!

The error message [ no kernels registered for operation `ewIsNan` ] indicates that DAPHNE looks for a kernel that is registered for the operation ewIsNan (case-sensitive). However, the kernel catalog (materialized in lib/catalog.json) contains kernels for the operation ewIsnan. The simplest solution is to change the mnemonic of the DaphneIR operation to ewIsnan in src/ir/daphneir/DaphneOps.td.

For some background: The catalog file is generated automatically by src/runtime/local/kernels/genKernelInst.py from the information in src/runtime/local/kernels.json. In the latter file, we specify the op codes of the elementwise unary operations to pre-compile like they are spelled in the respective C++ enum, i.e., in upper case letters. When deriving the DaphneIR operation names, genKernelInst.py (line 153) takes the first letter in upper case and all remaining letter in lower case, which makes sense for virtually all op codes.

StoeckOverflow commented 4 months ago

Hi @pdamme, Thank you for your detailed explanation!

I have implemented the suggested change by modifying the mnemonic of the DaphneIR operation to ewIsnan in src/ir/daphneir/DaphneOps.td. I am pleased to report that this has resolved the issue. All tests have passed successfully, and the pull request is now ready for review.

daphne-eu / daphne