[mlir][tosa] Add tosa.reduce_any and tosa.reduce_all linalg lowering

[rsuderman]

Summary

Added lowerings for Tosa's reduce boolean operations. This includes a fix to
maintain the output rank of reduce operations.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

if (isa<tosa::ReduceMaxOp>(op) && elementTy.isa<FloatType>())
     return rewriter.getFloatAttr(
         elementTy, APFloat::getLargest(
                        elementTy.cast<FloatType>().getFloatSemantics(), true));
 
   if (isa<tosa::ReduceMaxOp>(op) && elementTy.isa<IntegerType>())
     return rewriter.getIntegerAttr(
         elementTy, APInt::getSignedMinValue(elementTy.getIntOrFloatBitWidth()));
 
+  if (isa<tosa::ReduceAllOp>(op) && elementTy.isInteger(1))
+    return rewriter.getIntegerAttr(elementTy, APInt::getAllOnesValue(1));
+
+  if (isa<tosa::ReduceAnyOp>(op) && elementTy.isInteger(1))
+    return rewriter.getIntegerAttr(elementTy, APInt::getNullValue(1));
+
   if (isa<tosa::ArgMaxOp>(op) && elementTy.isa<FloatType>())
     return rewriter.getFloatAttr(
         elementTy, APFloat::getLargest(
                        elementTy.cast<FloatType>().getFloatSemantics(), true));
 
   if (isa<tosa::ArgMaxOp>(op) && elementTy.isa<IntegerType>())
     return rewriter.getIntegerAttr(
         elementTy, APInt::getSignedMinValue(elementTy.getIntOrFloatBitWidth()));
static Value createLinalgBodyCalculationForReduceOp(Operation *op,
   }
 
   if (isa<tosa::ReduceMaxOp>(op) && elementTy.isa<IntegerType>()) {
     auto predicate = rewriter.create<mlir::CmpIOp>(loc, CmpIPredicate::sgt,
                                                    args[0], args[1]);
     return rewriter.create<mlir::SelectOp>(loc, predicate, args[0], args[1]);
   }
 
+  if (isa<tosa::ReduceAllOp>(op) && elementTy.isInteger(1))
+    return rewriter.create<mlir::AndOp>(loc, args);
+
+  if (isa<tosa::ReduceAnyOp>(op) && elementTy.isInteger(1))
+    return rewriter.create<mlir::OrOp>(loc, args);
+
   return {};
 }
 
 // Performs the match and rewrite for reduction operations. This includes
 // declaring a correctly sized initial value, and the linalg.generic operation
 // that reduces across the specified axis.
 static LogicalResult reduceMatchAndRewriteHelper(Operation *op, uint64_t axis,
                                                  PatternRewriter &rewriter) {
static LogicalResult reduceMatchAndRewriteHelper(Operation *op, uint64_t axis,
   SmallVector<StringRef, 4> iteratorTypes;
   for (unsigned int i = 0, rank = inputTy.getRank(); i != rank; ++i) {
     srcExprs.push_back(mlir::getAffineDimExpr(i, rewriter.getContext()));
 
     iteratorTypes.push_back(axis == i ? getReductionIteratorTypeName()
                                       : getParallelIteratorTypeName());
     if (axis != i)
       dstExprs.push_back(mlir::getAffineDimExpr(i, rewriter.getContext()));
+    else
+      dstExprs.push_back(rewriter.getAffineConstantExpr(0));
   }
 
   bool didEncounterError = false;
   auto maps = AffineMap::inferFromExprList({srcExprs, dstExprs});
   auto linalgOp = rewriter.create<linalg::GenericOp>(
       loc, resultTy, input, filledTensor, maps, iteratorTypes,
       [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange blockArgs) {
         auto result = createLinalgBodyCalculationForReduceOp(
patterns->add<
       PointwiseConverter<tosa::LogicalLeftShiftOp>,
       PointwiseConverter<tosa::LogicalRightShiftOp>,
       PointwiseConverter<tosa::SelectOp>, PointwiseConverter<tosa::GreaterOp>,
       PointwiseConverter<tosa::GreaterEqualOp>,
       PointwiseConverter<tosa::MaximumOp>, PointwiseConverter<tosa::MinimumOp>,
       PointwiseConverter<tosa::CeilOp>, PointwiseConverter<tosa::FloorOp>,
       PointwiseConverter<tosa::ClampOp>, PointwiseConverter<tosa::ReluNOp>,
       PointwiseConverter<tosa::SigmoidOp>, IdentityNConverter<tosa::IdentityOp>,
-      IdentityNConverter<tosa::IdentityNOp>, ReduceConverter<tosa::ReduceMinOp>,
+      IdentityNConverter<tosa::IdentityNOp>, ReduceConverter<tosa::ReduceAllOp>,
+      ReduceConverter<tosa::ReduceAnyOp>, ReduceConverter<tosa::ReduceMinOp>,
       ReduceConverter<tosa::ReduceMaxOp>, ReduceConverter<tosa::ReduceSumOp>,
       ReduceConverter<tosa::ReduceProdOp>, ArgMaxConverter, ConcatConverter,
       PadConverter, ReshapeConverter, RescaleConverter, ReverseConverter,
       TileConverter, TransposeConverter, MatMulConverter,
       FullyConnectedConverter>(patterns->getContext());
 }

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

func @test_transpose(%arg0: tensor<1x2x3xi32>) -> () {
   // CHECK: }
   %1 = "tosa.transpose"(%arg0, %0) : (tensor<1x2x3xi32>, tensor<3xi32>) -> (tensor<2x3x1xi32>)
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d1)>
-// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (0, d1)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0, 0)>
 
 // CHECK-LABEL: @reduce_float
 // CHECK-SAME: [[ARG0:%.+]]: tensor<5x4xf32>
 func @reduce_float(%arg0: tensor<5x4xf32>) -> () {
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [4]
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 4]
   // CHECK: [[CST0:%.+]] = constant 0.0
   // CHECK: [[FILL:%.+]] = linalg.fill([[INIT]], [[CST0]])
-  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["reduction", "parallel"]} ins([[ARG0]] : tensor<5x4xf32>) outs([[FILL]] : tensor<4xf32>)
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["reduction", "parallel"]} ins([[ARG0]] : tensor<5x4xf32>) outs([[FILL]] : tensor<1x4xf32>)
   // CHECK: ^bb0(%arg1: f32, %arg2: f32)
   // CHECK:   [[RES:%.+]] = addf %arg1, %arg2 : f32
   // CHECK:   linalg.yield [[RES]] : f32
-  %0 = "tosa.reduce_sum"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<4xf32>
+  %0 = "tosa.reduce_sum"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<1x4xf32>
 
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5]
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5, 1]
   // CHECK: [[CST0:%.+]] = constant 0.0
   // CHECK: [[FILL:%.+]] = linalg.fill([[INIT]], [[CST0]])
-  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP2]]], iterator_types = ["parallel", "reduction"]} ins([[ARG0]] : tensor<5x4xf32>) outs([[FILL]] : tensor<5xf32>)
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP2]]], iterator_types = ["parallel", "reduction"]} ins([[ARG0]] : tensor<5x4xf32>) outs([[FILL]] : tensor<5x1xf32>)
   // CHECK: ^bb0(%arg1: f32, %arg2: f32)
   // CHECK:   [[RES:%.+]] = addf %arg1, %arg2 : f32
   // CHECK:   linalg.yield [[RES]] : f32
-  %1 = "tosa.reduce_sum"(%arg0) {axis = 1 : i64} : (tensor<5x4xf32>) -> tensor<5xf32>
+  %1 = "tosa.reduce_sum"(%arg0) {axis = 1 : i64} : (tensor<5x4xf32>) -> tensor<5x1xf32>
 
   // CHECK: constant 1.0
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: mulf
-  %2 = "tosa.reduce_prod"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<4xf32>
+  %2 = "tosa.reduce_prod"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<1x4xf32>
 
   // CHECK: constant 3.40282347E+38 : f32
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: cmpf olt
   // CHECK: select
-  %3 = "tosa.reduce_min"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<4xf32>
+  %3 = "tosa.reduce_min"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<1x4xf32>
 
   // CHECK: constant -3.40282347E+38 : f32
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: cmpf ogt
   // CHECK: select
-  %4 = "tosa.reduce_max"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<4xf32>
+  %4 = "tosa.reduce_max"(%arg0) {axis = 0 : i64} : (tensor<5x4xf32>) -> tensor<1x4xf32>
   return
 }
 
 // -----
 
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d1)>
-// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (0, d1)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0, 0)>
 
 // CHECK-LABEL: @reduce_int
 // CHECK-SAME: [[ARG0:%.+]]: tensor<5x4xi32>
 func @reduce_int(%arg0: tensor<5x4xi32>) -> () {
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [4]
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 4]
   // CHECK: [[CST0:%.+]] = constant 0
   // CHECK: [[FILL:%.+]] = linalg.fill([[INIT]], [[CST0]])
-  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["reduction", "parallel"]} ins([[ARG0]] : tensor<5x4xi32>) outs([[FILL]] : tensor<4xi32>)
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["reduction", "parallel"]} ins([[ARG0]] : tensor<5x4xi32>) outs([[FILL]] : tensor<1x4xi32>)
   // CHECK: ^bb0(%arg1: i32, %arg2: i32)
   // CHECK:   [[RES:%.+]] = addi %arg1, %arg2 : i32
   // CHECK:   linalg.yield [[RES]] : i32
-  %0 = "tosa.reduce_sum"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<4xi32>
+  %0 = "tosa.reduce_sum"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<1x4xi32>
 
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5]
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [5, 1]
   // CHECK: [[CST0:%.+]] = constant 0
   // CHECK: [[FILL:%.+]] = linalg.fill([[INIT]], [[CST0]])
-  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP2]]], iterator_types = ["parallel", "reduction"]} ins([[ARG0]] : tensor<5x4xi32>) outs([[FILL]] : tensor<5xi32>)
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP2]]], iterator_types = ["parallel", "reduction"]} ins([[ARG0]] : tensor<5x4xi32>) outs([[FILL]] : tensor<5x1xi32>)
   // CHECK: ^bb0(%arg1: i32, %arg2: i32)
   // CHECK:   [[RES:%.+]] = addi %arg1, %arg2 : i32
   // CHECK:   linalg.yield [[RES]] : i32
-  %1 = "tosa.reduce_sum"(%arg0) {axis = 1 : i64} : (tensor<5x4xi32>) -> tensor<5xi32>
+  %1 = "tosa.reduce_sum"(%arg0) {axis = 1 : i64} : (tensor<5x4xi32>) -> tensor<5x1xi32>
 
   // CHECK: constant 1
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: muli
-  %2 = "tosa.reduce_prod"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<4xi32>
+  %2 = "tosa.reduce_prod"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<1x4xi32>
 
   // CHECK: constant 2147483647 : i32
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: cmpi slt
   // CHECK: select
-  %3 = "tosa.reduce_min"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<4xi32>
+  %3 = "tosa.reduce_min"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<1x4xi32>
 
   // CHECK: constant -2147483648 : i32
   // CHECK: linalg.fill
   // CHECK: linalg.generic
   // CHECK: cmpi sgt
   // CHECK: select
-  %4 = "tosa.reduce_max"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<4xi32>
+  %4 = "tosa.reduce_max"(%arg0) {axis = 0 : i64} : (tensor<5x4xi32>) -> tensor<1x4xi32>
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (0, d1)>
+
+// CHECK-LABEL: @reduce_bool
+// CHECK-SAME: [[ARG0:%.+]]: tensor<5x4xi1>
+func @reduce_bool(%arg0: tensor<5x4xi1>) -> () {
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 4]
+  // CHECK: [[CST0:%.+]] = constant true
+  // CHECK: [[FILL:%.+]] = linalg.fill([[INIT]], [[CST0]])
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["reduction", "parallel"]} ins([[ARG0]] : tensor<5x4xi1>) outs([[FILL]] : tensor<1x4xi1>)
+  // CHECK: ^bb0(%arg1: i1, %arg2: i1)
+  // CHECK:   [[RES:%.+]] = and %arg1, %arg2 : i1
+  // CHECK:   linalg.yield [[RES]] : i1
+  %0 = "tosa.reduce_all"(%arg0) {axis = 0 : i64} : (tensor<5x4xi1>) -> tensor<1x4xi1>
+
+  // CHECK: constant false
+  // CHECK: linalg.fill
+  // CHECK: linalg.generic
+  // CHECK: or
+  %1 = "tosa.reduce_any"(%arg0) {axis = 0 : i64} : (tensor<5x4xi1>) -> tensor<1x4xi1>
+
   return
 }
 
 // -----
 
 // CHECK-LABEL: @concat
 func @concat(%arg0: tensor<5x1xf32>, %arg1: tensor<6x1xf32>) -> () {
   // CHECK: [[AXIS:%.+]] = constant 0