[mlir][tosa] Add support for EXT-DOUBLEROUND and EXT-INEXACTROUND

This commit adds a concept of EXT-DOUBLEROUND and EXT-INEXACTROUND
to the dialect. It also converts the "double_round" attribute on rescale
to a string type "rounding_mode" attribute with the following options:
"DOUBLE_ROUND", "SINGLE_ROUND", "INEXACT_ROUND".

The validation pass has been updated to ensure "DOUBLE_ROUND" and
"INEXACT_ROUND" are only valid when their extensions are available.

Finally, lowerings to arith and linalg have been updated such that
a lowering for "INEXACT_ROUND" is not currently supported.

Co-authored-by: TatWai Chong <[email protected]>

Author: lhutton1

mlir/include/mlir/Dialect/Tosa/IR/TosaOpBase.td

@@ -226,6 +226,8 @@ class Tosa_I32EnumAttr<string name, string description, string mnemonic,
 // FFT          : Fast Fourier Transform operations.
 // VARIABLE     : Stateful variable operations.
 // CONTROLFLOW  : Control Flow operations.
+// DOUBLEROUND  : Adds double rounding support to the RESCALE operator.
+// INEXACTROUND : Adds inexact rounding support to the RESCALE operator.
 //===----------------------------------------------------------------------===//
 
 def Tosa_NONE : I32EnumAttrCase<"none", 0>;
@@ -241,11 +243,14 @@ def Tosa_EXT_FP8E5M2      : I32EnumAttrCase<"fp8e5m2", 5>;
 def Tosa_EXT_FFT          : I32EnumAttrCase<"fft", 6>;
 def Tosa_EXT_VARIABLE     : I32EnumAttrCase<"variable", 7>;
 def Tosa_EXT_CONTROLFLOW  : I32EnumAttrCase<"controlflow", 8>;
+def Tosa_EXT_DOUBLEROUND  : I32EnumAttrCase<"doubleround", 9>;
+def Tosa_EXT_INEXACTROUND : I32EnumAttrCase<"inexactround", 10>;
 
 def Tosa_ExtensionAttr
     : Tosa_I32EnumAttr<"Extension", "supported TOSA extensions", "ext", [
       Tosa_EXT_INT16, Tosa_EXT_INT4, Tosa_EXT_BF16, Tosa_EXT_FP8E4M3,
-      Tosa_EXT_FP8E5M2, Tosa_EXT_FFT, Tosa_EXT_VARIABLE, Tosa_EXT_CONTROLFLOW, Tosa_EXT_NONE
+      Tosa_EXT_FP8E5M2, Tosa_EXT_FFT, Tosa_EXT_VARIABLE, Tosa_EXT_CONTROLFLOW,
+      Tosa_EXT_DOUBLEROUND, Tosa_EXT_INEXACTROUND, Tosa_EXT_NONE
     ]>;
 
 def Tosa_ExtensionArrayAttr

mlir/include/mlir/Dialect/Tosa/IR/TosaOps.td

@@ -2310,7 +2310,7 @@ def Tosa_RescaleOp : Tosa_InferShapedTypeOp<"rescale"> {
     I32Attr:$input_zp,
     I32Attr:$output_zp,
     BoolAttr:$scale32,
-    BoolAttr:$double_round,
+    Tosa_RoundingTypeAttr:$rounding_mode,
     BoolAttr:$per_channel,
     BoolAttr: $input_unsigned,
     BoolAttr: $output_unsigned

mlir/include/mlir/Dialect/Tosa/IR/TosaProfileCompliance.h

@@ -136,6 +136,8 @@ class TosaProfileCompliance {
     switch (ext) {
     case Extension::int16:
     case Extension::int4:
+    case Extension::doubleround:
+    case Extension::inexactround:
       return {Profile::pro_int};
     case Extension::bf16:
     case Extension::fp8e4m3:

mlir/include/mlir/Dialect/Tosa/IR/TosaTypesBase.td

@@ -247,6 +247,13 @@ def Tosa_NanPropagationAttr : StringBasedAttr<
           "::llvm::cast<StringAttr>($_self).getValue() == \"IGNORE\"">,
     "Supported NaN propagation strategies">;
 
+// Rounding mode for tosa.rescale
+def Tosa_RoundingTypeAttr : StringBasedAttr<
+    CPred<"::llvm::cast<StringAttr>($_self).getValue() == \"SINGLE_ROUND\"  || " #
+          "::llvm::cast<StringAttr>($_self).getValue() == \"INEXACT_ROUND\" || " #
+          "::llvm::cast<StringAttr>($_self).getValue() == \"DOUBLE_ROUND\"">,
+    "Supported rounding modes">;
+
 def Tosa_TensorTypeAttr : TypeAttrBase<"TensorType", "Tensor type attribute">;
 
 // Tensor to buffer types.

mlir/include/mlir/Dialect/Tosa/IR/TosaUtilOps.td

@@ -44,7 +44,7 @@ def Tosa_ApplyScaleOp :
     Tosa_IntLike:$value,
     Tosa_IntLike:$multiplier,
     Tosa_Int8Like:$shift,
-    BoolAttr:$double_round
+    Tosa_RoundingTypeAttr:$rounding_mode
   );
 
   let results = (outs

mlir/lib/Conversion/TosaToArith/TosaToArith.cpp

@@ -65,6 +65,11 @@ class ApplyScaleGenericOpConverter
 
   LogicalResult matchAndRewrite(tosa::ApplyScaleOp op,
                                 PatternRewriter &rewriter) const final {
+    StringRef roundingMode = op.getRoundingMode();
+    if (roundingMode != "DOUBLE_ROUND" && roundingMode != "SINGLE_ROUND") {
+      return failure();
+    }
+
     Location loc = op.getLoc();
     Value value = op.getValue();
     Value multiplier32 = op.getMultiplier();
@@ -96,7 +101,7 @@ class ApplyScaleGenericOpConverter
     multiply64 = rewriter.create<arith::AddIOp>(loc, multiply64, round);
 
     // Apply double rounding if necessary.
-    if (op.getDoubleRound()) {
+    if (op.getRoundingMode() == "DOUBLE_ROUND") {
       int64_t roundInt = 1 << 30;
       Value roundUp = getConstantValue(loc, i64Ty, roundInt, rewriter);
       Value roundDown = getConstantValue(loc, i64Ty, -roundInt, rewriter);
@@ -125,6 +130,11 @@ class ApplyScale32BitOpConverter : public OpRewritePattern<tosa::ApplyScaleOp> {
 
   LogicalResult matchAndRewrite(tosa::ApplyScaleOp op,
                                 PatternRewriter &rewriter) const final {
+    StringRef roundingMode = op.getRoundingMode();
+    if (roundingMode != "DOUBLE_ROUND" && roundingMode != "SINGLE_ROUND") {
+      return failure();
+    }
+
     Location loc = op.getLoc();
 
     Type resultTy = op.getType();
@@ -170,7 +180,7 @@ class ApplyScale32BitOpConverter : public OpRewritePattern<tosa::ApplyScaleOp> {
         rewriter.create<arith::SelectOp>(loc, shiftOver32, shiftHighR, zero32);
 
     // Conditionally perform our double round.
-    if (op.getDoubleRound()) {
+    if (op.getRoundingMode() == "DOUBLE_ROUND") {
       Value negOne32 = getConstantValue(loc, i32Ty, -1, rewriter);
       Value valuePositive = rewriter.create<arith::CmpIOp>(
           loc, arith::CmpIPredicate::sge, value32, zero32);

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

@@ -170,7 +170,7 @@ static Value createLinalgBodyCalculationForElementwiseOp(
 
         auto result = rewriter.create<tosa::ApplyScaleOp>(
             loc, rewriter.getI32Type(), a, b, shiftConst,
-            rewriter.getBoolAttr(false));
+            rewriter.getStringAttr("SINGLE_ROUND"));
 
         if (elementTy.isInteger(32))
           return result;
@@ -1374,7 +1374,11 @@ class RescaleConverter : public OpRewritePattern<tosa::RescaleOp> {
     unsigned rank = inputTy.getRank();
 
     // This is an illegal configuration. terminate and log an error
-    if (op.getDoubleRound() && !op.getScale32())
+    if (op.getRoundingMode() == "INEXACT_ROUND")
+      return rewriter.notifyMatchFailure(
+          op, "tosa.rescale with rounding mode = 'INEXACT_ROUND' is not "
+              "currently supported");
+    if (op.getRoundingMode() == "DOUBLE_ROUND" && !op.getScale32())
       return rewriter.notifyMatchFailure(
           op, "tosa.rescale requires scale32 for double_round to be true");
 
@@ -1418,9 +1422,13 @@ class RescaleConverter : public OpRewritePattern<tosa::RescaleOp> {
 
     // Double round only occurs if shift is greater than 31, check that this
     // is ever true.
+
     bool doubleRound =
-        op.getDoubleRound() &&
+        op.getRoundingMode() == "DOUBLE_ROUND" &&
         llvm::any_of(shiftValues, [](int32_t v) { return v > 31; });
+    StringAttr roundingMode = doubleRound
+                                  ? rewriter.getStringAttr("DOUBLE_ROUND")
+                                  : rewriter.getStringAttr("SINGLE_ROUND");
 
     SmallVector<AffineMap> indexingMaps = {
         rewriter.getMultiDimIdentityMap(rank)};
@@ -1516,7 +1524,7 @@ class RescaleConverter : public OpRewritePattern<tosa::RescaleOp> {
 
           value = nestedBuilder.create<tosa::ApplyScaleOp>(
               loc, nestedBuilder.getI32Type(), value, multiplier, shift,
-              nestedBuilder.getBoolAttr(doubleRound));
+              roundingMode);
 
           // Move to the new zero-point.
           value =

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp

@@ -1003,9 +1003,9 @@ class AvgPool2dConverter : public OpRewritePattern<tosa::AvgPool2dOp> {
 
             auto scaled =
                 rewriter
-                    .create<tosa::ApplyScaleOp>(loc, rewriter.getI32Type(),
-                                                poolVal, multiplier, shift,
-                                                rewriter.getBoolAttr(false))
+                    .create<tosa::ApplyScaleOp>(
+                        loc, rewriter.getI32Type(), poolVal, multiplier, shift,
+                        rewriter.getStringAttr("SINGLE_ROUND"))
                     .getResult();
 
             // If we have quantization information we need to apply output

mlir/lib/Dialect/Tosa/Transforms/TosaValidation.cpp

@@ -104,6 +104,7 @@ struct TosaValidation : public tosa::impl::TosaValidationBase<TosaValidation> {
   }
 
   LogicalResult applyLevelCheck(Operation *op);
+  LogicalResult applyAttributeCheck(Operation *op);
 
   // check variable read/write data types against variable declarations
   LogicalResult applyVariableCheck(Operation *op);
@@ -386,6 +387,25 @@ struct TosaValidation : public tosa::impl::TosaValidationBase<TosaValidation> {
     return true;
   }
 
+  bool attributeCheckRescale(Operation *op) {
+    if (auto rescale = dyn_cast<tosa::RescaleOp>(op)) {
+      if (rescale.getRoundingMode() == "DOUBLE_ROUND" &&
+          !targetEnv.allows(Extension::doubleround)) {
+        op->emitOpError()
+            << "failed attribute check: rounding_mode = DOUBLE_ROUND "
+            << "requires extension [doubleround]";
+        return false;
+      } else if (rescale.getRoundingMode() == "INEXACT_ROUND" &&
+                 !targetEnv.allows(Extension::inexactround)) {
+        op->emitOpError()
+            << "failed attribute check: rounding_mode = INEXACT_ROUND "
+            << "requires extension [inexactround]";
+        return false;
+      }
+    }
+    return true;
+  }
+
   // configure profile and level values from pass options profileName and
   // levelName
   void configLevelAndProfile() {
@@ -415,7 +435,8 @@ struct TosaValidation : public tosa::impl::TosaValidationBase<TosaValidation> {
         } else {
           llvm::errs() << "unknown TOSA extension name passed in: " << ext
                        << ", supported extension are int16, int4, bf16, "
-                       << "fp8e4m3, fp8e5m2, fft, variable and controlflow\n";
+                       << "fp8e4m3, fp8e5m2, fft, variable, controlflow, "
+                       << "doubleround and inexactround\n";
           return signalPassFailure();
         }
       }
@@ -642,6 +663,12 @@ LogicalResult TosaValidation::applyLevelCheck(Operation *op) {
   return success();
 }
 
+LogicalResult TosaValidation::applyAttributeCheck(Operation *op) {
+  if (!attributeCheckRescale(op))
+    return failure();
+  return success();
+}
+
 inline bool CompatibleTypes(const mlir::Type &type,
                             const mlir::Type &declaredType) {
   // for now, simply use type equality comparison
@@ -936,6 +963,10 @@ void TosaValidation::runOnOperation() {
     if (failed(applyLevelCheck(op)))
       signalPassFailure();
 
+    // check additional attribute restrictions
+    if (failed(applyAttributeCheck(op)))
+      signalPassFailure();
+
     // do variable type checks
     if (failed(applyVariableCheck(op)))
       signalPassFailure();

mlir/test/Conversion/TosaToArith/tosa-to-arith-invalid.mlir

@@ -0,0 +1,8 @@
+// RUN: mlir-opt --split-input-file --tosa-to-arith="include-apply-rescale=true use-32-bit=true" %s -verify-diagnostics
+
+// CHECK-LABEL: @apply_scale_unsupported_inexact_round
+func.func @apply_scale_unsupported_inexact_round(%arg0 : i64, %arg1 : i32, %arg2 : i8) -> (i32) {
+  // expected-error@+1 {{failed to legalize operation 'tosa.apply_scale'}}
+  %res = tosa.apply_scale %arg0, %arg1, %arg2 {rounding_mode = "INEXACT_ROUND"} : (i64, i32, i8) -> i32
+  return %res : i32
+}