batched matmul pattern

add conv helpers
new validation tests for matmul

validation/operations/gen_tests.py

@@ -421,6 +421,53 @@ def matmul_batched_3d():
     save_model(model, "matmul/batched_3d", "matmul_batched_3d.onnx")
 
 
+def matmul_batched_3d_dynamic():
+    """Batched 3D MatMul with both operands provided at runtime."""
+    A = helper.make_tensor_value_info("A", TensorProto.FLOAT, [2, 2, 3])
+    B = helper.make_tensor_value_info("B", TensorProto.FLOAT, [2, 3, 4])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 2, 4])
+    node = helper.make_node("MatMul", ["A", "B"], ["Y"])
+    graph = helper.make_graph([node], "matmul_batched_3d_dynamic", [A, B], [Y])
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+    save_model(model, "matmul/batched_3d_dynamic", "matmul_batched_3d_dynamic.onnx")
+
+
+def matmul_batched_left_constant():
+    """Batched 3D MatMul with constant LHS and runtime RHS."""
+    rng = np.random.default_rng(70)
+    A = numpy_helper.from_array(rng.uniform(-1, 1, (2, 2, 3)).astype(np.float32), name="A")
+    B = helper.make_tensor_value_info("B", TensorProto.FLOAT, [2, 3, 4])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 2, 4])
+    node = helper.make_node("MatMul", ["A", "B"], ["Y"])
+    graph = helper.make_graph([node], "matmul_batched_left_constant", [B], [Y], initializer=[A])
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+    save_model(model, "matmul/batched_left_constant", "matmul_batched_left_constant.onnx")
+
+
+def matmul_batched_rhs_broadcast():
+    """Batched 3D MatMul with 2D constant RHS broadcast across batch."""
+    rng = np.random.default_rng(71)
+    A = helper.make_tensor_value_info("A", TensorProto.FLOAT, [2, 2, 3])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 2, 4])
+    B = numpy_helper.from_array(rng.uniform(-1, 1, (3, 4)).astype(np.float32), name="B")
+    node = helper.make_node("MatMul", ["A", "B"], ["Y"])
+    graph = helper.make_graph([node], "matmul_batched_rhs_broadcast", [A], [Y], initializer=[B])
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+    save_model(model, "matmul/batched_rhs_broadcast", "matmul_batched_rhs_broadcast.onnx")
+
+
+def matmul_batched_lhs_broadcast():
+    """Batched 3D MatMul with 2D runtime LHS broadcast across batched RHS."""
+    rng = np.random.default_rng(72)
+    A = helper.make_tensor_value_info("A", TensorProto.FLOAT, [2, 3])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 2, 4])
+    B = numpy_helper.from_array(rng.uniform(-1, 1, (2, 3, 4)).astype(np.float32), name="B")
+    node = helper.make_node("MatMul", ["A", "B"], ["Y"])
+    graph = helper.make_graph([node], "matmul_batched_lhs_broadcast", [A], [Y], initializer=[B])
+    model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 13)])
+    save_model(model, "matmul/batched_lhs_broadcast", "matmul_batched_lhs_broadcast.onnx")
+
+
 # ---------------------------------------------------------------------------
 # Pooling tests
 # ---------------------------------------------------------------------------
@@ -972,6 +1019,10 @@ if __name__ == "__main__":
     matmul_left_constant()
     matmul_dynamic()
     matmul_batched_3d()
+    matmul_batched_3d_dynamic()
+    matmul_batched_left_constant()
+    matmul_batched_rhs_broadcast()
+    matmul_batched_lhs_broadcast()
 
     print("\nGenerating Pooling tests:")
     maxpool_basic()