Fix dot general flop counting

jax/_src/pallas/BUILD

@@ -56,3 +56,12 @@ py_library(
         "//jax/_src/lib",
     ] + py_deps("numpy"),
 )
+
+py_test(
+    name = "cost_estimate_test",
+    srcs = ["test.py"],
+    deps = [
+        ":pallas",
+        "//jax",
+    ] + py_deps("absl/testing"),
+)

jax/_src/pallas/cost_estimate.py

@@ -204,11 +204,14 @@ def dot_general_cost_rule(ctx: Context,
   assert len(lhs_contracting_dims) == len(rhs_contracting_dims)
   assert len(lhs_batch_dims) == len(rhs_batch_dims)
   flops = 1
-  # Flops along a contracting dim is 2*dim (addition and multiplication)
+
+  contracting_size=1
   for i in range(len(lhs_contracting_dims)):
-    lhs_dim, rhs_dim = lhs_contracting_dims[i], rhs_contracting_dims[i]
+    lhs_dim, rhs_dim=lhs_contracting_dims[i], rhs_contracting_dims[i]
     assert x_shape[lhs_dim] == y_shape[rhs_dim]
-    flops *= 2 * x_shape[lhs_dim]
+    contracting_size *= x_shape[lhs_dim]
+
+  flops *= 2 * contracting_size
   # Now we handle all other dimensions.
   for i, lhs_dim in enumerate(x_shape):
     if i in lhs_contracting_dims:

jax/_src/pallas/cost_estimate_test.py

@@ -0,0 +1,63 @@
+# Copyright 2024 The JAX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for cost_estimate.py."""
+
+from absl.testing import absltest
+import jax
+import jax.numpy as jnp
+import jax.experimental.pallas as pl
+
+
+class CostEstimateTest(absltest.TestCase):
+    """Tests for Pallas cost estimation."""
+
+    def test_dot_general_single_contracting_dim(self):
+        """Test FLOP counting with single contracting dimension."""
+
+        def matmul(x, y):
+            return jnp.einsum("mk,kn->mn", x, y)
+
+        x = jax.random.normal(jax.random.key(0), (64, 128))
+        y = jax.random.normal(jax.random.key(1), (128, 256))
+
+        xla_flops = jax.jit(matmul).lower(x, y).compile().cost_analysis()["flops"]
+        pl_flops = pl.estimate_cost(matmul, x, y).flops
+
+        self.assertEqual(xla_flops, pl_flops)
+
+    def test_dot_general_multiple_contracting_dims(self):
+        """Test FLOP counting with multiple contracting dimensions.
+
+        This is a regression test for https://github.com/jax-ml/jax/issues/33388
+        where FLOPs were incorrectly doubled for each contracting dimension.
+        """
+
+        def test(x, y):
+            return jnp.einsum("...mk,...kn->mn", x, y)
+
+        x = jax.random.normal(jax.random.key(0), (2, 64, 128))
+        y = jax.random.normal(jax.random.key(1), (2, 128, 256))
+
+        xla_flops = jax.jit(test).lower(x, y).compile().cost_analysis()["flops"]
+        pl_flops = pl.estimate_cost(test, x, y).flops
+
+        # Expected: 64 * 256 * (2 * 128) * 2 = 8,388,608
+        expected = 64 * 256 * 2 * 128 * 2
+
+        self.assertEqual(xla_flops, expected)
+        self.assertEqual(pl_flops, expected)
+
+
+if __name__ == "__main__":
+    absltest.main()

jax/_src/pallas/test.py

@@ -0,0 +1,63 @@
+# Copyright 2024 The JAX Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Tests for cost_estimate.py."""
+
+from absl.testing import absltest
+import jax
+import jax.numpy as jnp
+import jax.experimental.pallas as pl
+
+
+class CostEstimateTest(absltest.TestCase):
+    """Tests for Pallas cost estimation."""
+
+    def test_dot_general_single_contracting_dim(self):
+        """Test FLOP counting with single contracting dimension."""
+
+        def matmul(x, y):
+            return jnp.einsum("mk,kn->mn", x, y)
+
+        x = jax.random.normal(jax.random.key(0), (64, 128))
+        y = jax.random.normal(jax.random.key(1), (128, 256))
+
+        xla_flops = jax.jit(matmul).lower(x, y).compile().cost_analysis()["flops"]
+        pl_flops = pl.estimate_cost(matmul, x, y).flops
+
+        self.assertEqual(xla_flops, pl_flops)
+
+    def test_dot_general_multiple_contracting_dims(self):
+        """Test FLOP counting with multiple contracting dimensions.
+
+        This is a regression test for https://github.com/jax-ml/jax/issues/33388
+        where FLOPs were incorrectly doubled for each contracting dimension.
+        """
+
+        def test(x, y):
+            return jnp.einsum("...mk,...kn->mn", x, y)
+
+        x = jax.random.normal(jax.random.key(0), (2, 64, 128))
+        y = jax.random.normal(jax.random.key(1), (2, 128, 256))
+
+        xla_flops = jax.jit(test).lower(x, y).compile().cost_analysis()["flops"]
+        pl_flops = pl.estimate_cost(test, x, y).flops
+
+        # Expected: 64 * 256 * (2 * 128) * 2 = 8,388,608
+        expected = 64 * 256 * 2 * 128 * 2
+
+        self.assertEqual(xla_flops, expected)
+        self.assertEqual(pl_flops, expected)
+
+
+if __name__ == "__main__":
+    absltest.main()