Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Correctness issue in canonicalize-loops #431

Closed
wsmoses opened this issue Mar 4, 2025 · 3 comments
Closed

Correctness issue in canonicalize-loops #431

wsmoses opened this issue Mar 4, 2025 · 3 comments

Comments

@wsmoses
Copy link
Member

wsmoses commented Mar 4, 2025 

(base) wmoses@hydra:~/git/Enzyme-JaX$ ./bazel-bin/enzymexlamlir-opt "--pass-pipeline=builtin.module(func.func(canonicalize-loops),llvm-to-affine-access)" salt.mlir 
#set = affine_set<(d0, d1) : (d0 floordiv 16 + (d1 floordiv 12) * 16 + 1 >= 0, -(d0 floordiv 16) - (d1 floordiv 12) * 16 + 84 >= 0, d0 mod 16 + (d1 mod 12) * 16 >= 0, -(d0 mod 16) - (d1 mod 12) * 16 + 179 >= 0)>
#tbaa_root = #llvm.tbaa_root<id = "custom_tbaa">
#tbaa_type_desc = #llvm.tbaa_type_desc<id = "custom_tbaa_addrspace(1)", members = {<#tbaa_root, 0>}>
#tbaa_tag = #llvm.tbaa_tag<base_type = #tbaa_type_desc, access_type = #tbaa_type_desc, offset = 0>
module @"reactant_run!" attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @bar(%arg0: tensor<1x187x194xf64>, %arg1: tensor<34xf64>, %arg2: tensor<1x187x194xf64>) -> tensor<1x187x194xf64> {
    %0 = enzymexla.jit_call @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0" (%arg0, %arg1, %arg2) {output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 0, operand_tuple_indices = []>]} : (tensor<1x187x194xf64>, tensor<34xf64>, tensor<1x187x194xf64>) -> tensor<1x187x194xf64>
    return %0 : tensor<1x187x194xf64>
  }
  func.func private @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0"(%arg0: memref<1x187x194xf64, 1>, %arg1: memref<34xf64, 1>, %arg2: memref<1x187x194xf64, 1>) {
    %cst = arith.constant 0.000000e+00 : f64
    %0 = "enzymexla.memref2pointer"(%arg0) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    %1 = "enzymexla.memref2pointer"(%arg1) : (memref<34xf64, 1>) -> !llvm.ptr<1>
    %2 = "enzymexla.memref2pointer"(%arg2) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    affine.parallel (%arg3, %arg4) = (0, 0) to (72, 256) {
      affine.if #set(%arg4, %arg3) {
        %3 = "enzymexla.pointer2memref"(%1) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %4 = affine.load %3[26] {alignment = 16 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %5 = "enzymexla.pointer2memref"(%2) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %6 = affine.load %5[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 3104 + (%arg4 floordiv 16) * 178 - ((%arg3 + 1) floordiv 12) * 192 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %7 = arith.cmpf ole, %4, %6 {fastmathFlags = #llvm.fastmath<none>} : f64
        %8 = "enzymexla.pointer2memref"(%2) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %9 = affine.load %8[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 3104 + (%arg4 floordiv 16) * 178 - ((%arg3 + 1) floordiv 12) * 192 + 9900] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %10 = arith.cmpf ole, %4, %9 {fastmathFlags = #llvm.fastmath<none>} : f64
        %11 = arith.ori %7, %10 : i1
        %12 = "enzymexla.pointer2memref"(%0) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %13 = affine.load %12[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 3104 + (%arg4 floordiv 16) * 178 - ((%arg3 + 1) floordiv 12) * 192 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %14 = arith.select %11, %cst, %13 : f64
        %15 = "enzymexla.pointer2memref"(%0) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        affine.store %14, %15[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 3104 + (%arg4 floordiv 16) * 178 - ((%arg3 + 1) floordiv 12) * 192 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
      }
    }
    return
  }
}

(base) wmoses@hydra:~/git/Enzyme-JaX$ ./bazel-bin/enzymexlamlir-opt "--pass-pipeline=builtin.module(llvm-to-affine-access)" salt.mlir 
#set = affine_set<(d0, d1) : (d0 floordiv 16 + (d1 floordiv 12) * 16 + 1 >= 0, -(d0 floordiv 16) - (d1 floordiv 12) * 16 + 84 >= 0, d0 mod 16 + (d1 mod 12) * 16 >= 0, -(d0 mod 16) - (d1 mod 12) * 16 + 179 >= 0)>
#tbaa_root = #llvm.tbaa_root<id = "custom_tbaa">
#tbaa_type_desc = #llvm.tbaa_type_desc<id = "custom_tbaa_addrspace(1)", members = {<#tbaa_root, 0>}>
#tbaa_tag = #llvm.tbaa_tag<base_type = #tbaa_type_desc, access_type = #tbaa_type_desc, offset = 0>
module @"reactant_run!" attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @bar(%arg0: tensor<1x187x194xf64>, %arg1: tensor<34xf64>, %arg2: tensor<1x187x194xf64>) -> tensor<1x187x194xf64> {
    %0 = enzymexla.jit_call @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0" (%arg0, %arg1, %arg2) {output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 0, operand_tuple_indices = []>]} : (tensor<1x187x194xf64>, tensor<34xf64>, tensor<1x187x194xf64>) -> tensor<1x187x194xf64>
    return %0 : tensor<1x187x194xf64>
  }
  func.func private @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0"(%arg0: memref<1x187x194xf64, 1>, %arg1: memref<34xf64, 1>, %arg2: memref<1x187x194xf64, 1>) {
    %cst = arith.constant 0.000000e+00 : f64
    %0 = "enzymexla.memref2pointer"(%arg0) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    %1 = "enzymexla.memref2pointer"(%arg1) : (memref<34xf64, 1>) -> !llvm.ptr<1>
    %2 = "enzymexla.memref2pointer"(%arg2) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    affine.parallel (%arg3, %arg4) = (0, 0) to (72, 256) {
      affine.if #set(%arg4, %arg3) {
        %3 = "enzymexla.pointer2memref"(%1) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %4 = affine.load %3[26] {alignment = 16 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %5 = "enzymexla.pointer2memref"(%2) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %6 = affine.load %5[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 2912 + (%arg4 floordiv 16) * 178 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %7 = arith.cmpf ole, %4, %6 {fastmathFlags = #llvm.fastmath<none>} : f64
        %8 = "enzymexla.pointer2memref"(%2) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %9 = affine.load %8[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 2912 + (%arg4 floordiv 16) * 178 + 9900] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %10 = arith.cmpf ole, %4, %9 {fastmathFlags = #llvm.fastmath<none>} : f64
        %11 = arith.ori %7, %10 : i1
        %12 = "enzymexla.pointer2memref"(%0) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        %13 = affine.load %12[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 2912 + (%arg4 floordiv 16) * 178 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
        %14 = arith.select %11, %cst, %13 : f64
        %15 = "enzymexla.pointer2memref"(%0) : (!llvm.ptr<1>) -> memref<?xf64, 1>
        affine.store %14, %15[%arg3 * 16 + %arg4 + (%arg3 floordiv 12) * 2912 + (%arg4 floordiv 16) * 178 + 9901] {alignment = 8 : i64, ordering = 0 : i64, tbaa = [#tbaa_tag]} : memref<?xf64, 1>
      }
    }
    return
  }
}


(base) wmoses@hydra:~/git/Enzyme-JaX$ cat salt.mlir 
#set = affine_set<(d0, d1) : (d0 floordiv 16 + (d1 floordiv 12) * 16 + 1 >= 0, -(d0 floordiv 16) - (d1 floordiv 12) * 16 + 84 >= 0, d0 mod 16 + (d1 mod 12) * 16 >= 0, -(d0 mod 16) - (d1 mod 12) * 16 + 179 >= 0)>
#tbaa_root = #llvm.tbaa_root<id = "custom_tbaa">
#tbaa_type_desc = #llvm.tbaa_type_desc<id = "custom_tbaa_addrspace(1)", members = {<#tbaa_root, 0>}>
#tbaa_tag = #llvm.tbaa_tag<base_type = #tbaa_type_desc, access_type = #tbaa_type_desc, offset = 0>
module @"reactant_run!" attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @bar(%arg0: tensor<1x187x194xf64>, %arg1: tensor<34xf64>, %arg2: tensor<1x187x194xf64>) -> tensor<1x187x194xf64> {
    %0 = enzymexla.jit_call @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0" (%arg0, %arg1, %arg2) {output_operand_aliases = [#stablehlo.output_operand_alias<output_tuple_indices = [], operand_index = 0, operand_tuple_indices = []>]} : (tensor<1x187x194xf64>, tensor<34xf64>, tensor<1x187x194xf64>) -> tensor<1x187x194xf64>
    return %0 : tensor<1x187x194xf64>
  }
  func.func private @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0"(%arg0: memref<1x187x194xf64, 1>, %arg1: memref<34xf64, 1>, %arg2: memref<1x187x194xf64, 1>) {
    %c51_i64 = arith.constant 51 : i64
    %c-1_i64 = arith.constant -1 : i64
    %c1 = arith.constant 1 : index
    %c12_i32 = arith.constant 12 : i32
    %c-12_i64 = arith.constant -12 : i64
    %c16_i16 = arith.constant 16 : i16
    %c0_i64 = arith.constant 0 : i64
    %c16_i64 = arith.constant 16 : i64
    %c194_i64 = arith.constant 194 : i64
    %c6_i64 = arith.constant 6 : i64
    %c5_i64 = arith.constant 5 : i64
    %cst = arith.constant 0.000000e+00 : f64
    %0 = "enzymexla.memref2pointer"(%arg0) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    %1 = "enzymexla.memref2pointer"(%arg1) : (memref<34xf64, 1>) -> !llvm.ptr<1>
    %2 = "enzymexla.memref2pointer"(%arg2) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
    affine.parallel (%arg3, %arg4) = (0, 0) to (72, 256) {
      %3 = arith.addi %arg3, %c1 : index
      %4 = arith.addi %arg4, %c1 : index
      %5 = arith.index_castui %3 : index to i64
      %6 = arith.addi %5, %c-1_i64 : i64
      %7 = arith.trunci %6 : i64 to i32
      %8 = arith.divui %7, %c12_i32 : i32
      %9 = arith.extui %8 : i32 to i64
      %10 = arith.muli %9, %c-12_i64 : i64
      %11 = arith.addi %10, %5 : i64
      %12 = arith.index_castui %4 : index to i64
      %13 = arith.addi %12, %c-1_i64 : i64
      %14 = arith.trunci %13 : i64 to i16
      %15 = arith.divui %14, %c16_i16 : i16
      %16 = arith.extui %15 : i16 to i64
      %17 = arith.subi %c0_i64, %16 : i64
      %18 = arith.addi %17, %11 : i64
      %19 = arith.addi %18, %c-1_i64 : i64
      %20 = arith.muli %19, %c16_i64 : i64
      %21 = arith.addi %12, %20 : i64
      %22 = arith.muli %9, %c16_i64 : i64
      %23 = arith.addi %22, %16 : i64
      affine.if #set(%arg4, %arg3) {
        %24 = llvm.getelementptr inbounds %1[26] : (!llvm.ptr<1>) -> !llvm.ptr<1>, f64
        %25 = llvm.load %24 {alignment = 16 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr<1> -> f64
        %26 = arith.addi %23, %c51_i64 : i64
        %27 = arith.muli %26, %c194_i64 : i64
        %28 = arith.addi %27, %21 : i64
        %29 = arith.addi %28, %c6_i64 : i64
        %30 = llvm.getelementptr inbounds %2[%29] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
        %31 = llvm.load %30 {alignment = 8 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr<1> -> f64
        %32 = arith.cmpf ole, %25, %31 {fastmathFlags = #llvm.fastmath<none>} : f64
        %33 = arith.addi %27, %21 : i64
        %34 = arith.addi %33, %c5_i64 : i64
        %35 = llvm.getelementptr inbounds %2[%34] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
        %36 = llvm.load %35 {alignment = 8 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr<1> -> f64
        %37 = arith.cmpf ole, %25, %36 {fastmathFlags = #llvm.fastmath<none>} : f64
        %38 = arith.ori %32, %37 : i1
        %39 = llvm.getelementptr inbounds %0[%29] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
        %40 = llvm.load %39 {alignment = 8 : i64, tbaa = [#tbaa_tag]} : !llvm.ptr<1> -> f64
        %41 = arith.select %38, %cst, %40 : f64
        llvm.store %41, %39 {alignment = 8 : i64, tbaa = [#tbaa_tag]} : f64, !llvm.ptr<1>
      }
    }
    return
  }
}
@wsmoses
Copy link
Member Author

wsmoses commented Mar 4, 2025

ToRem of this seems to be wrong:

func.func private @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0"(%arg0: memref<1x187x194xf64, 1>, %arg1: memref<34xf64, 1>, %arg2: memref<1x187x194xf64, 1>) {
  %c51_i64 = arith.constant 51 : i64
  %c-1_i64 = arith.constant -1 : i64
  %c1 = arith.constant 1 : index
  %c12_i32 = arith.constant 12 : i32
  %c-12_i64 = arith.constant -12 : i64
  %c16_i16 = arith.constant 16 : i16
  %c0_i64 = arith.constant 0 : i64
  %c16_i64 = arith.constant 16 : i64
  %c194_i64 = arith.constant 194 : i64
  %c6_i64 = arith.constant 6 : i64
  %c5_i64 = arith.constant 5 : i64
  %cst = arith.constant 0.000000e+00 : f64
  %0 = "enzymexla.memref2pointer"(%arg0) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
  %1 = "enzymexla.memref2pointer"(%arg1) : (memref<34xf64, 1>) -> !llvm.ptr<1>
  %2 = "enzymexla.memref2pointer"(%arg2) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
  affine.parallel (%arg3, %arg4) = (0, 0) to (72, 256) {
    %3 = arith.addi %arg3, %c1 : index
    %4 = arith.addi %arg4, %c1 : index
    %5 = arith.index_castui %3 : index to i64
    %c0 = arith.constant 0 : index
    %6 = arith.addi %arg3, %c0 : index
    %7 = arith.index_castui %6 : index to i64
    %8 = arith.index_castui %6 : index to i32
    %c12 = arith.constant 12 : index
    %9 = arith.divui %6, %c12 : index
    %10 = arith.index_castui %9 : index to i32
    %11 = arith.index_castui %9 : index to i64
    %c-12 = arith.constant -12 : index
    %12 = arith.muli %9, %c-12 : index
    %13 = arith.index_cast %12 : index to i64
    %14 = arith.addi %12, %3 : index
    %15 = arith.index_cast %14 : index to i64
    %16 = arith.index_castui %4 : index to i64
    %c16 = arith.constant 16 : index
    %17 = arith.divui %arg4, %c16 : index
    %18 = arith.index_castui %17 : index to i64
    %19 = arith.subi %c0_i64, %18 : i64
    %20 = arith.addi %19, %15 : i64
    %21 = arith.addi %20, %c-1_i64 : i64
    %22 = arith.muli %21, %c16_i64 : i64
    %23 = arith.addi %16, %22 : i64
    %24 = arith.muli %11, %c16_i64 : i64
    %25 = arith.addi %24, %18 : i64
    affine.if affine_set<(d0, d1) : (d0 floordiv 16 + (d1 floordiv 12) * 16 + 1 >= 0, -(d0 floordiv 16) - (d1 floordiv 12) * 16 + 84 >= 0, d0 mod 16 + (d1 mod 12) * 16 >= 0, -(d0 mod 16) - (d1 mod 12) * 16 + 179 >= 0)>(%arg4, %arg3) {
      %26 = llvm.getelementptr inbounds %1[26] : (!llvm.ptr<1>) -> !llvm.ptr<1>, f64
      %27 = llvm.load %26 {alignment = 16 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %28 = arith.addi %25, %c51_i64 : i64
      %29 = arith.muli %28, %c194_i64 : i64
      %30 = arith.addi %29, %23 : i64
      %31 = arith.addi %30, %c6_i64 : i64
      %32 = llvm.getelementptr inbounds %2[%31] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %33 = llvm.load %32 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %34 = arith.cmpf ole, %27, %33 {fastmathFlags = #llvm.fastmath<none>} : f64
      %35 = arith.addi %29, %23 : i64
      %36 = arith.addi %35, %c5_i64 : i64
      %37 = llvm.getelementptr inbounds %2[%36] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %38 = llvm.load %37 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %39 = arith.cmpf ole, %27, %38 {fastmathFlags = #llvm.fastmath<none>} : f64
      %40 = arith.ori %34, %39 : i1
      %41 = llvm.getelementptr inbounds %0[%31] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %42 = llvm.load %41 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %43 = arith.select %40, %cst, %42 : f64
      llvm.store %43, %41 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>
]} : f64, !llvm.ptr<1>
    }
  }
  return
}

  * Pattern {anonymous}::{anonymous}::ToRem : 'arith.addi -> ()' {
Trying to match "{anonymous}::{anonymous}::ToRem"
    ** Insert  : 'arith.remui'(0x5661ebcabd50)
    ** Replace : 'arith.addi'(0x5661ebcabc40)
    ** Modified: 'arith.index_cast'(0x5661ebcac330)
    ** Erase   : 'arith.addi'(0x5661ebcabc40)
"{anonymous}::{anonymous}::ToRem" result 1
  } -> success : pattern applied successfully
// *** IR Dump After Pattern Application ***
func.func private @"##call__Z28gpu__mask_immersed_field_xy_16CompilerMetadataI10StaticSizeI9_180__85_E12DynamicCheckvv7NDRangeILi2ES0_I7_12__6_ES0_I8_16__16_EvvEE5FieldI4Face6Centervvvv11OffsetArrayI7Float64Li3E13CuTracedArrayISC_Li3ELi1E13_194__187__1_EESC_vvvE5TupleIS9_SA_vE20ImmersedBoundaryGridISC_8Periodic14RightConnected7Bounded28OrthogonalSphericalShellGridISC_SK_SL_SM_28StaticVerticalDiscretizationISB_ISC_Li1ESD_ISC_Li1ELi1E5_35__EESB_ISC_Li1ESD_ISC_Li1ELi1E5_34__EESQ_SS_ESB_ISC_Li2ESD_ISC_Li2ELi1E10_194__187_EE8TripolarI5Int64SX_SX_EvE16GridFittedBottomIS8_ISA_SA_vvvvSF_SC_vvvE23CenterImmersedConditionEvvvESC_SX_#802$par0"(%arg0: memref<1x187x194xf64, 1>, %arg1: memref<34xf64, 1>, %arg2: memref<1x187x194xf64, 1>) {
  %c51_i64 = arith.constant 51 : i64
  %c-1_i64 = arith.constant -1 : i64
  %c1 = arith.constant 1 : index
  %c12_i32 = arith.constant 12 : i32
  %c-12_i64 = arith.constant -12 : i64
  %c16_i16 = arith.constant 16 : i16
  %c0_i64 = arith.constant 0 : i64
  %c16_i64 = arith.constant 16 : i64
  %c194_i64 = arith.constant 194 : i64
  %c6_i64 = arith.constant 6 : i64
  %c5_i64 = arith.constant 5 : i64
  %cst = arith.constant 0.000000e+00 : f64
  %0 = "enzymexla.memref2pointer"(%arg0) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
  %1 = "enzymexla.memref2pointer"(%arg1) : (memref<34xf64, 1>) -> !llvm.ptr<1>
  %2 = "enzymexla.memref2pointer"(%arg2) : (memref<1x187x194xf64, 1>) -> !llvm.ptr<1>
  affine.parallel (%arg3, %arg4) = (0, 0) to (72, 256) {
    %3 = arith.addi %arg3, %c1 : index
    %4 = arith.addi %arg4, %c1 : index
    %5 = arith.index_castui %3 : index to i64
    %c0 = arith.constant 0 : index
    %6 = arith.addi %arg3, %c0 : index
    %7 = arith.index_castui %6 : index to i64
    %8 = arith.index_castui %6 : index to i32
    %c12 = arith.constant 12 : index
    %9 = arith.divui %6, %c12 : index
    %10 = arith.index_castui %9 : index to i32
    %11 = arith.index_castui %9 : index to i64
    %c-12 = arith.constant -12 : index
    %12 = arith.muli %9, %c-12 : index
    %13 = arith.index_cast %12 : index to i64
    %14 = arith.remui %3, %c12 : index
    %15 = arith.index_cast %14 : index to i64
    %16 = arith.index_castui %4 : index to i64
    %c16 = arith.constant 16 : index
    %17 = arith.divui %arg4, %c16 : index
    %18 = arith.index_castui %17 : index to i64
    %19 = arith.subi %c0_i64, %18 : i64
    %20 = arith.addi %19, %15 : i64
    %21 = arith.addi %20, %c-1_i64 : i64
    %22 = arith.muli %21, %c16_i64 : i64
    %23 = arith.addi %16, %22 : i64
    %24 = arith.muli %11, %c16_i64 : i64
    %25 = arith.addi %24, %18 : i64
    affine.if affine_set<(d0, d1) : (d0 floordiv 16 + (d1 floordiv 12) * 16 + 1 >= 0, -(d0 floordiv 16) - (d1 floordiv 12) * 16 + 84 >= 0, d0 mod 16 + (d1 mod 12) * 16 >= 0, -(d0 mod 16) - (d1 mod 12) * 16 + 179 >= 0)>(%arg4, %arg3) {
      %26 = llvm.getelementptr inbounds %1[26] : (!llvm.ptr<1>) -> !llvm.ptr<1>, f64
      %27 = llvm.load %26 {alignment = 16 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %28 = arith.addi %25, %c51_i64 : i64
      %29 = arith.muli %28, %c194_i64 : i64
      %30 = arith.addi %29, %23 : i64
      %31 = arith.addi %30, %c6_i64 : i64
      %32 = llvm.getelementptr inbounds %2[%31] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %33 = llvm.load %32 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %34 = arith.cmpf ole, %27, %33 {fastmathFlags = #llvm.fastmath<none>} : f64
      %35 = arith.addi %29, %23 : i64
      %36 = arith.addi %35, %c5_i64 : i64
      %37 = llvm.getelementptr inbounds %2[%36] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %38 = llvm.load %37 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %39 = arith.cmpf ole, %27, %38 {fastmathFlags = #llvm.fastmath<none>} : f64
      %40 = arith.ori %34, %39 : i1
      %41 = llvm.getelementptr inbounds %0[%31] : (!llvm.ptr<1>, i64) -> !llvm.ptr<1>, f64
      %42 = llvm.load %41 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : !llvm.ptr<1> -> f64
      %43 = arith.select %40, %cst, %42 : f64
      llvm.store %43, %41 {alignment = 8 : i64, tbaa = [#llvm.tbaa_tag<base_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, access_type = <id = "custom_tbaa_addrspace(1)", members = {<#llvm.tbaa_root<id = "custom_tbaa">, 0>}>, offset = 0>]} : f64, !llvm.ptr<1>
    }
  }
  return
}

@wsmoses
Copy link
Member Author

wsmoses commented Mar 4, 2025

found bug

@wsmoses
Copy link
Member Author

wsmoses commented Mar 4, 2025

x/ref #432

@wsmoses wsmoses closed this as completed Mar 4, 2025
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
1 participant
@wsmoses