Add sub operator

Author: Ziminli

src/ntops/kernels/sub.py

@@ -0,0 +1,17 @@
+import functools
+
+import ninetoothed
+from ninetoothed import Tensor
+
+from ntops.kernels.element_wise import arrangement
+
+
+def application(input, other, alpha, output):
+    output = input - alpha * other  # noqa: F841
+
+
+@functools.cache
+def make(ndim):
+    tensors = (Tensor(ndim), Tensor(ndim), Tensor(0), Tensor(ndim))
+
+    return ninetoothed.make(arrangement, application, tensors)

src/ntops/torch.py

@@ -32,6 +32,7 @@
 import ntops.kernels.sigmoid
 import ntops.kernels.sin
 import ntops.kernels.softmax
+import ntops.kernels.sub
 import ntops.kernels.tanh
 
 
@@ -384,6 +385,17 @@ def softmax(input, dim, dtype=None):
     return output
 
 
+def sub(input, other, *, alpha=1, out=None):
+    if out is None:
+        out = torch.empty_like(input)
+
+    kernel = ntops.kernels.sub.make(input.ndim)
+
+    kernel(input, other, alpha, out)
+
+    return out
+
+
 def tanh(input, *, out=None):
     if out is None:
         out = torch.empty_like(input)

tests/test_sub.py

@@ -0,0 +1,21 @@
+import pytest
+import torch
+
+import ntops.torch
+from tests.skippers import skip_if_cuda_not_available
+from tests.utils import gauss, generate_arguments
+
+
+@skip_if_cuda_not_available
+@pytest.mark.parametrize(*generate_arguments())
+def test_cuda(shape, dtype, atol, rtol):
+    device = "cuda"
+
+    input = torch.randn(shape, dtype=dtype, device=device)
+    other = torch.randn(shape, dtype=dtype, device=device)
+    alpha = gauss()
+
+    ninetoothed_output = ntops.torch.sub(input, other, alpha=alpha)
+    reference_output = torch.sub(input, other, alpha=alpha)
+
+    assert torch.allclose(ninetoothed_output, reference_output, atol=atol, rtol=rtol)