portable arg{max,min}: optimize update check

We don't need a second isnan; see code comment. (This is a small optimization.)

ghstack-comment-id: 2686527528

ghstack-source-id: a34aa77
ghstack-comment-id: 2691808767
Pull Request resolved: pytorch/executorch#8863

Author: swolchok

kernels/portable/cpu/op_argmax.cpp

@@ -50,7 +50,10 @@ Tensor& argmax_out(
     for (const auto out_ix : c10::irange(out.numel())) {
       std::tuple<CTYPE, long> acc = reduce_over_dim<CTYPE>(
           [](CTYPE v, long ix, CTYPE acc_val, long acc_ix) {
-            if (!std::isnan(acc_val) && (std::isnan(v) || v > acc_val)) {
+            // the below condition as written is equivalent to
+            // !isnan(accval) && (isnan(v) || v > acc_val). See
+            // argument in op_argmin.cpp.
+            if (!std::isnan(acc_val) && !(v <= acc_val)) {
               acc_val = v;
               acc_ix = ix;
             }

kernels/portable/cpu/op_argmin.cpp

@@ -50,7 +50,17 @@ Tensor& argmin_out(
     for (const auto out_ix : c10::irange(out.numel())) {
       std::tuple<CTYPE, long> acc = reduce_over_dim<CTYPE>(
           [](CTYPE v, long ix, CTYPE acc_val, long acc_ix) {
-            if (!std::isnan(acc_val) && (std::isnan(v) || v < acc_val)) {
+            // the below condition as written is equivalent to !isnan(accval) &&
+            // (isnan(v) || v < acc_val). cases:
+            // - if neither acc_val nor v is NaN, !(v >= acc_val) is
+            //   trivially equivalent to v < acc_val.
+            // - if acc_val is NaN, the whole thing is trivially false.
+            // - if acc_val is not NaN and v is NaN, then v >= acc_val
+            // - is false because all comparisons involving NaN are
+            // - false, so the result is true. The result is trivially
+            // - true for the above condition that uses isnan(v) as
+            // - well.
+            if (!std::isnan(acc_val) && !(v >= acc_val)) {
               acc_val = v;
               acc_ix = ix;
             }

kernels/test/op_argmax_test.cpp

@@ -90,3 +90,16 @@ TEST_F(OpArgmaxTest, SanityCheckNullDim) {
   EXPECT_TENSOR_EQ(out, expected);
   // clang-format on
 }
+
+TEST_F(OpArgmaxTest, FirstNaNWins) {
+  TensorFactory<ScalarType::Float> tf_float;
+  Tensor in = tf_float.make({4}, {1, NAN, -4, NAN});
+
+  TensorFactory<ScalarType::Long> tf_long;
+  Tensor out = tf_long.zeros({});
+  Tensor expected = tf_long.make({}, {1});
+
+  Tensor ret = op_argmax_out(in, {}, false, out);
+  EXPECT_TENSOR_EQ(out, ret);
+  EXPECT_TENSOR_EQ(out, expected);
+}

kernels/test/op_argmin_test.cpp

@@ -90,3 +90,16 @@ TEST_F(OpArgminTest, SanityCheckNullDim) {
   EXPECT_TENSOR_EQ(out, expected);
   // clang-format on
 }
+
+TEST_F(OpArgminTest, FirstNaNWins) {
+  TensorFactory<ScalarType::Float> tf_float;
+  Tensor in = tf_float.make({4}, {1, NAN, -4, NAN});
+
+  TensorFactory<ScalarType::Long> tf_long;
+  Tensor out = tf_long.zeros({});
+  Tensor expected = tf_long.make({}, {1});
+
+  Tensor ret = op_argmin_out(in, {}, false, out);
+  EXPECT_TENSOR_EQ(out, ret);
+  EXPECT_TENSOR_EQ(out, expected);
+}