Merge pull request #58 from InfiniTensor/develop-matmul

Add `matmul` operator

Author: voltjia

src/ntops/torch/__init__.py

@@ -19,6 +19,7 @@
 from ntops.torch.layer_norm import layer_norm
 from ntops.torch.le import le
 from ntops.torch.lt import lt
+from ntops.torch.matmul import matmul
 from ntops.torch.mm import mm
 from ntops.torch.mul import mul
 from ntops.torch.ne import ne
@@ -58,6 +59,7 @@
     "layer_norm",
     "le",
     "lt",
+    "matmul",
     "mm",
     "mul",
     "ne",

src/ntops/torch/matmul.py

@@ -0,0 +1,18 @@
+import ntops
+
+
+def matmul(input, other, *, out=None):
+    assert input.ndim in (2, 3) and other.ndim in (2, 3), (
+        "Currently, only 2D and 3D tensors are supported."
+    )
+
+    if input.ndim == 2 and other.ndim == 2:
+        return ntops.torch.mm(input, other, out=out)
+
+    if input.ndim < 3:
+        input = input.unsqueeze(0)
+
+    if other.ndim < 3:
+        other = other.unsqueeze(0)
+
+    return ntops.torch.bmm(input, other, out=out)

tests/test_matmul.py

@@ -0,0 +1,22 @@
+import pytest
+import torch
+
+import ntops
+from tests.skippers import skip_if_cuda_not_available
+from tests.test_mm import generate_arguments
+
+
+@skip_if_cuda_not_available
+@pytest.mark.parametrize(*generate_arguments())
+@pytest.mark.parametrize("b", (None, 1, 2, 3))
+def test_matmul(b, m, n, k, dtype, device, rtol, atol):
+    input_shape = (b, m, k) if b is not None else (m, k)
+    other_shape = (b, k, n) if b is not None else (k, n)
+
+    input = torch.randn(input_shape, dtype=dtype, device=device)
+    other = torch.randn(other_shape, dtype=dtype, device=device)
+
+    ninetoothed_output = ntops.torch.matmul(input, other)
+    reference_output = torch.matmul(input, other)
+
+    assert torch.allclose(ninetoothed_output, reference_output, rtol=rtol, atol=atol)