[Relay][ONNX] Batch_matmul to dense optimization (#8440)

* [ONNX]Add batch_matmul to dense optimization

* Add extra check to avoid unnecessary reshape

Co-authored-by: Ubuntu <ubuntu@ip-172-31-14-16.us-west-2.compute.internal>

python/tvm/relay/frontend/onnx.py

@@ -678,25 +678,34 @@ def _impl_v1(cls, inputs, attr, params):
         # When performing a batch matmul, we need to properly handle N-dim shapes.
         if a_rank > 2 or b_rank > 2:
 
-            def flatten_to_3d(x, x_shape):
+            def flatten_to_nd(x, x_shape, nd=3):
                 ndims = infer_shape(x_shape)[0]
+                if ndims == nd:
+                    return x
                 newshape = _op.concatenate(
                     [
                         _expr.const([-1], dtype=infer_type(x_shape).checked_type.dtype),
-                        _op.strided_slice(x_shape, [ndims - 2], [ndims]),
+                        _op.strided_slice(x_shape, [ndims - nd + 1], [ndims]),
                     ],
                     0,
                 )
                 out = _op.reshape(x, fold_constant(newshape))
                 return out
 
-            # Convert a and b into 3 dimensional tensors.
-            a = flatten_to_3d(inputs[0], a_shape)
-            b = flatten_to_3d(inputs[1], b_shape)
-            # Transpose matrix dimensions of b.
-            b = _op.transpose(b, [0, 2, 1])
-            # Perform a batch matmul.
-            output = _op.nn.batch_matmul(a, b)
+            b_type = infer_type(inputs[1])
+            # Convert to dense if the second matrix is 2d and non-dynamic
+            if b_rank == 2 and not _ty.is_dynamic(b_type.checked_type):
+                a = flatten_to_nd(inputs[0], a_shape, 2)
+                b = _op.transpose(inputs[1])
+                output = _op.nn.dense(a, b)
+            else:
+                # Convert a and b into 3 dimensional tensors.
+                a = flatten_to_nd(inputs[0], a_shape, 3)
+                b = flatten_to_nd(inputs[1], b_shape, 3)
+                # Transpose matrix dimensions of b.
+                b = _op.transpose(b, [0, 2, 1])
+                # Perform a batch matmul.
+                output = _op.nn.batch_matmul(a, b)
             # Determine the output batch dimension.
             if a_rank > b_rank:
                 out_batch = _op.strided_slice(a_shape, [0], [a_rank - 2])

tests/python/frontend/onnx/test_forward.py

@@ -1173,8 +1173,7 @@ def verify_batch_matmul(a_shape, b_shape, out_shape, target, dev):
     verify_with_ort_with_inputs(model, [a_array, b_array], use_vm=True, targets=[target])
 
 
-# TODO(mbrookhart, electriclilies): Add CUDA as a target once batch matmul is fixed
-@tvm.testing.parametrize_targets("llvm")
+@tvm.testing.uses_gpu
 def test_batch_matmul(target, dev):
     verify_batch_matmul((2, 3, 4, 3), (2, 3, 3, 4), (2, 3, 4, 4), target, dev)
     verify_batch_matmul((2, 4, 3), (3, 4), (2, 4, 4), target, dev)
@@ -1183,6 +1182,8 @@ def test_batch_matmul(target, dev):
     verify_batch_matmul((4, 3), (2, 3, 4), (2, 4, 4), target, dev)
     verify_batch_matmul((2, 4, 3), (1, 3, 4), (2, 4, 4), target, dev)
     verify_batch_matmul((1, 4, 3), (2, 3, 4), (2, 4, 4), target, dev)
+    verify_batch_matmul((4, 32, 16), (16, 32), (4, 32, 32), target, dev)
+    verify_batch_matmul((4, 32, 16, 32), (32, 16), (4, 32, 16, 16), target, dev)
 
 
 def verify_simple_dynamic_model(a_shape, b_shape, target, dev):
@@ -1221,7 +1222,6 @@ def verify_model(ex, a_shape, b_shape):
     b_anys = [relay.Any()] * len(b_shape)
 
     mod, params = relay.frontend.from_onnx(model, {"a": a_anys, "b": b_anys})
-
     ex = relay.create_executor("vm", mod=mod, device=dev, target=target)
     verify_model(ex, a_shape, b_shape)
     verify_model(ex, [a * 2 for a in a_shape], [b * 2 for b in b_shape])
@@ -4955,3 +4955,4 @@ def test_qlinearadd():
     test_reverse_sequence()
     test_eyelike()
     test_qlinearconv()
+    test_batch_matmul()