Back to PBT run

test.mlir

// RUN: loom-raise-opt --loom-lower-graph-memory --mlir-print-op-generic %s | FileCheck %s

dataflow.graph private @pair_0(%start: none, %lb: i64, %ub: i64, %step: i64,
    %a: memref<?xi32>, %b: memref<?xi32>) -> ()
    attributes {input_segments = array<i32: 3, 0, 2>,
                result_segments = array<i32: 0, 0, 0>} {
  scf.for %i = %lb to %ub step %step : i64 {
    %scaled = arith.muli %i, %step : i64
    %idx = arith.index_cast %scaled : i64 to index
  }
  dataflow.graph.return %start : none
}

dataflow.graph private @acc_1(%start: none, %lb: i64, %ub: i64, %step: i64, %init: i32,
    %a: memref<?xi32>) -> (i32)
    attributes {input_segments = array<i32: 4, 0, 1>,
                result_segments = array<i32: 1, 0, 0>} {
  %total = scf.for %i = %lb to %ub step %step
      iter_args(%state = %init) -> (i32) : i64 {
    %shifted = arith.addi %i, %lb : i64
    %idx = arith.index_cast %shifted : i64 to index
    %aloaded = memref.load %a[%idx] : memref<?xi32>
    %asum = arith.addi %state, %aloaded : i32
    memref.store %asum, %a[%idx] : memref<?xi32>
    scf.yield %asum : i32
  }
  dataflow.graph.return %start, %total : none, i32
}

dataflow.graph private @pair_2(%start: none, %lb: i64, %ub: i64, %step: i64,
    %a: memref<?xi32>, %b: memref<?xi32>) -> ()
    attributes {input_segments = array<i32: 3, 0, 2>,
                result_segments = array<i32: 0, 0, 0>} {
  dataflow.graph.return %start : none
}


// CHECK: "builtin.module"() ({
// CHECK-NEXT:   "dataflow.graph"() <{function_type = (i64, i64, i64, memref<?xi32>, memref<?xi32>) -> (), input_segments = array<i32: 3, 0, 2>, result_segments = array<i32: 0, 0, 0>, sym_name = "pair_0", sym_visibility = "private"}> ({
// CHECK-NEXT:   ^bb0(%arg12: none, %arg13: i64, %arg14: i64, %arg15: i64, %arg16: memref<?xi32>, %arg17: memref<?xi32>):
// CHECK-NEXT:     %26:2 = "dataflow.stream"(%arg13, %arg14, %arg15) <{predicate = 2 : i64, step_kind = 0 : i32}> : (i64, i64, i64) -> (i64, i1)
// CHECK-NEXT:     %27 = "dataflow.carry"(%26#1, %arg12, %28#1) : (i1, none, none) -> none
// CHECK-NEXT:     %28:2 = "dataflow.demux"(%26#1, %27) : (i1, none) -> (none, none)
// CHECK-NEXT:     %29 = "dataflow.invariant"(%26#1, %arg15) : (i1, i64) -> i64
// CHECK-NEXT:     %30:2 = "dataflow.gate"(%26#1, %29) : (i1, i64) -> (i1, i64)
// CHECK-NEXT:     %31:2 = "dataflow.demux"(%30#0, %30#1) : (i1, i64) -> (i64, i64)
// CHECK-NEXT:     %32 = "arith.cmpi"(%arg13, %arg14) <{predicate = 2 : i64}> : (i64, i64) -> i1
// CHECK-NEXT:     %33:2 = "dataflow.demux"(%32, %28#0) : (i1, none) -> (none, none)
// CHECK-NEXT:     %34:2 = "dataflow.sync"(%33#1, %31#0) : (none, i64) -> (none, i64)
// CHECK-NEXT:     %35 = "dataflow.mux"(%32, %33#0, %34#0) : (i1, none, none) -> none
// CHECK-NEXT:     "dataflow.graph.return"(%35) <{operandSegmentSizes = array<i32: 0, 0, 0, 1>}> : (none) -> ()
// CHECK-NEXT:   }) : () -> ()
// CHECK-NEXT:   "dataflow.graph"() <{function_type = (i64, i64, i64, i32, memref<?xi32>) -> i32, input_segments = array<i32: 4, 0, 1>, result_segments = array<i32: 1, 0, 0>, sym_name = "acc_1", sym_visibility = "private"}> ({
// CHECK-NEXT:   ^bb0(%arg6: none, %arg7: i64, %arg8: i64, %arg9: i64, %arg10: i32, %arg11: memref<?xi32>):
// CHECK-NEXT:     %0:2 = "dataflow.stream"(%arg7, %arg8, %arg9) <{predicate = 2 : i64, step_kind = 0 : i32}> : (i64, i64, i64) -> (i64, i1)
// CHECK-NEXT:     %1 = "dataflow.carry"(%0#1, %arg6, %2#1) : (i1, none, none) -> none
// CHECK-NEXT:     %2:2 = "dataflow.demux"(%0#1, %1) : (i1, none) -> (none, none)
// CHECK-NEXT:     %3 = "dataflow.carry"(%0#1, %arg10, %18) : (i1, i32, i32) -> i32
// CHECK-NEXT:     %4:2 = "dataflow.demux"(%0#1, %3) : (i1, i32) -> (i32, i32)
// CHECK-NEXT:     %5 = "dataflow.invariant"(%0#1, %arg7) : (i1, i64) -> i64
// CHECK-NEXT:     %6:2 = "dataflow.gate"(%0#1, %5) : (i1, i64) -> (i1, i64)
// CHECK-NEXT:     %7:2 = "dataflow.demux"(%6#0, %6#1) : (i1, i64) -> (i64, i64)
// CHECK-NEXT:     %8 = "dataflow.carry"(%0#1, %arg6, %19) : (i1, none, none) -> none
// CHECK-NEXT:     %9 = "dataflow.carry"(%0#1, %arg6, %19) : (i1, none, none) -> none
// CHECK-NEXT:     %10:2 = "dataflow.demux"(%0#1, %8) : (i1, none) -> (none, none)
// CHECK-NEXT:     %11:2 = "dataflow.demux"(%0#1, %9) : (i1, none) -> (none, none)
// CHECK-NEXT:     %12 = "arith.index_cast"(%0#0) : (i64) -> index
// CHECK-NEXT:     %13 = "arith.index_cast"(%6#1) : (i64) -> index
// CHECK-NEXT:     %14 = "arith.addi"(%12, %13) <{overflowFlags = #arith.overflow<none>}> : (index, index) -> index
// CHECK-NEXT:     %15:2 = "dataflow.sync"(%2#1, %10#1) : (none, none) -> (none, none)
// CHECK-NEXT:     %16:2 = "dataflow.load"(%arg11, %14, %15#0) : (memref<?xi32>, index, none) -> (i32, none)
// CHECK-NEXT:     %17:2 = "dataflow.sync"(%11#1, %16#1) : (none, none) -> (none, none)
// CHECK-NEXT:     %18 = "arith.addi"(%4#1, %16#0) <{overflowFlags = #arith.overflow<none>}> : (i32, i32) -> i32
// CHECK-NEXT:     %19 = "dataflow.store"(%arg11, %14, %18, %17#0) : (memref<?xi32>, index, i32, none) -> none
// CHECK-NEXT:     %20 = "arith.cmpi"(%arg7, %arg8) <{predicate = 2 : i64}> : (i64, i64) -> i1
// CHECK-NEXT:     %21:2 = "dataflow.demux"(%20, %2#0) : (i1, none) -> (none, none)
// CHECK-NEXT:     %22:2 = "dataflow.sync"(%21#1, %7#0) : (none, i64) -> (none, i64)
// CHECK-NEXT:     %23 = "dataflow.mux"(%20, %21#0, %22#0) : (i1, none, none) -> none
// CHECK-NEXT:     %24:2 = "dataflow.sync"(%23, %11#0) : (none, none) -> (none, none)
// CHECK-NEXT:     %25:2 = "dataflow.sync"(%24#0, %4#0) : (none, i32) -> (none, i32)
// CHECK-NEXT:     "dataflow.graph.return"(%25#1, %25#0) <{operandSegmentSizes = array<i32: 1, 0, 0, 1>}> : (i32, none) -> ()
// CHECK-NEXT:   }) : () -> ()
// CHECK-NEXT:   "dataflow.graph"() <{function_type = (i64, i64, i64, memref<?xi32>, memref<?xi32>) -> (), input_segments = array<i32: 3, 0, 2>, result_segments = array<i32: 0, 0, 0>, sym_name = "pair_2", sym_visibility = "private"}> ({
// CHECK-NEXT:   ^bb0(%arg0: none, %arg1: i64, %arg2: i64, %arg3: i64, %arg4: memref<?xi32>, %arg5: memref<?xi32>):
// CHECK-NEXT:     "dataflow.graph.return"(%arg0) <{operandSegmentSizes = array<i32: 0, 0, 0, 1>}> : (none) -> ()
// CHECK-NEXT:   }) : () -> ()
// CHECK-NEXT: }) : () -> ()
// CHECK-EMPTY: