[Mlir-commits] [mlir] 95d526f - [MLIR][Tosa] Fix argmax NaN propagate lowering (#133074)

[llvmlistbot at llvm.org]

Author: Thomas Preud'homme
Date: 2025-04-14T09:37:09+01:00
New Revision: 95d526f7f587cc7a3db785169967f8dad0ba3978

URL: https://github.com/llvm/llvm-project/commit/95d526f7f587cc7a3db785169967f8dad0ba3978
DIFF: https://github.com/llvm/llvm-project/commit/95d526f7f587cc7a3db785169967f8dad0ba3978.diff

LOG: [MLIR][Tosa] Fix argmax NaN propagate lowering (#133074)

In the propagate mode, NaN compare equal to each other so in case of
several NaNs the index of the first one needs to be returned. This
commit changes the index update condition to check that the current
index is not that of a NaN.

The commit also simplifies argmax NaN ignore lowering to only use OGT.
This prevent any update in case of NaN. The only case where the index of
a NaN is returned is when all values are NaN and this is covered by the
fact that the initial index value is 0 so no update will result in 0
being returned.

Added: 
    

Modified: 
    mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
    mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index e18fa849e9f30..9ca93ab28daed 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -2285,8 +2285,22 @@ class ArgMaxConverter : public OpRewritePattern<tosa::ArgMaxOp> {
 
           Value predicate;
           if (isa<FloatType>(inElementTy)) {
-            predicate = rewriter.create<arith::CmpFOp>(
-                nestedLoc, arith::CmpFPredicate::OGT, newValue, oldValue);
+            if (argmaxOp.getNanMode() == "IGNORE") {
+              // Only update index & max value for non NaN values. If all
+              // values are NaNs, the initial index will be return which is 0.
+              predicate = rewriter.create<arith::CmpFOp>(
+                  nestedLoc, arith::CmpFPredicate::OGT, newValue, oldValue);
+            } else {
+              // Update max value if either of the following is true:
+              // - new value is bigger
+              // - cur max is not NaN and new value is NaN
+              Value gt = rewriter.create<arith::CmpFOp>(
+                  nestedLoc, arith::CmpFPredicate::UGT, newValue, oldValue);
+              Value oldNonNaN = rewriter.create<arith::CmpFOp>(
+                  nestedLoc, arith::CmpFPredicate::ORD, oldValue, oldValue);
+              predicate = rewriter.create<arith::AndIOp>(
+                  nestedLoc, rewriter.getI1Type(), gt, oldNonNaN);
+            }
           } else if (isa<IntegerType>(inElementTy)) {
             predicate = rewriter.create<arith::CmpIOp>(
                 nestedLoc, arith::CmpIPredicate::sgt, newValue, oldValue);
@@ -2299,28 +2313,6 @@ class ArgMaxConverter : public OpRewritePattern<tosa::ArgMaxOp> {
               nestedLoc, predicate, newValue, oldValue);
           auto resultIndex = rewriter.create<arith::SelectOp>(
               nestedLoc, predicate, newIndex, oldIndex);
-
-          // Check if we need to materialize compare and select for the given
-          // NaN propagation mode.
-
-          // "PROPAGATE" matches the default NaN propagation mode of the arith
-          // dialect so no compare and select is required.
-          //
-          // In the case "IGNORE" we check if the current argument is NaN and
-          // select the old index and value otherwise take the updated index and
-          // value.
-          if (const auto nanMode = argmaxOp.getNanMode();
-              isa<FloatType>(inElementTy) && nanMode == "IGNORE") {
-            // Unordered comparison of NaN against itself will always return
-            // true.
-            Value isNaN = rewriter.create<arith::CmpFOp>(
-                argmaxOp.getLoc(), arith::CmpFPredicate::UNO, newValue,
-                newValue);
-            resultMax = rewriter.create<arith::SelectOp>(nestedLoc, isNaN,
-                                                         oldValue, resultMax);
-            resultIndex = rewriter.create<arith::SelectOp>(
-                nestedLoc, isNaN, oldIndex, resultIndex);
-          }
           nestedBuilder.create<linalg::YieldOp>(
               nestedLoc, ValueRange({resultIndex, resultMax}));
         });

diff  --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index e2ef6574fc661..bcf45e5271aaf 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -1525,7 +1525,9 @@ func.func @argmax(%arg0 : tensor<3x2xi32>, %arg1 : tensor<6xf32>) -> () {
   // CHECK: arith.constant -3.40282347E+38 : f32
   // CHECK: linalg.index
   // CHECK: arith.index_cast
-  // CHECK: arith.cmpf ogt
+  // CHECK: arith.cmpf ugt
+  // CHECK: arith.cmpf ord
+  // CHECK: andi
   // CHECK: select
   // CHECK: select
   // CHECK: linalg.yield
@@ -2230,12 +2232,12 @@ func.func @maximum_nan_ignore(%arg0: tensor<5x4xf32>, %arg1: tensor<5x4xf32>) ->
 // CHECK-LABEL: @argmax_nan_propagate
 func.func @argmax_nan_propagate(%arg0: tensor<5x4xf32>, %arg1: tensor<5x4xf32>) -> () {
   // CHECK: linalg.generic
-  // CHECK: arith.cmpf ogt
+  // CHECK: arith.cmpf ugt
+  // CHECK: arith.cmpf ord
+  // CHECK: andi
   // CHECK: arith.select
   // CHECK: arith.select
   // CHECK-NOT: arith.cmpf uno
-  // CHECK-NOT: arith.cmpf uno
-  // CHECK-NOT: arith.select
   // CHECK-NOT: arith.select
   // CHECK: linalg.yield
   %11 = tosa.argmax %arg0 {axis = 0 : i32, nan_mode = "PROPAGATE"} : (tensor<5x4xf32>)  -> tensor<4xi32>
@@ -2267,9 +2269,6 @@ func.func @argmax_nan_ignore(%arg0: tensor<5x4xf32>, %arg1: tensor<5x4xf32>) ->
   // CHECK: arith.cmpf ogt
   // CHECK: arith.select
   // CHECK: arith.select
-  // CHECK: arith.cmpf uno
-  // CHECK: arith.select
-  // CHECK: arith.select
   // CHECK: linalg.yield
   %12 = tosa.argmax %arg0 {axis = 0 : i32, nan_mode = "IGNORE"} : (tensor<5x4xf32>)  -> tensor<4xi32>
   return