Add relu operator

Author: stever178

src/ntops/kernels/relu.py

@@ -0,0 +1,15 @@
+import functools
+
+import ninetoothed
+from ninetoothed import Tensor
+
+from ntops.kernels.element_wise import arrangement
+
+
+def application(input, output):
+    output = max(0.0, input)  # noqa: F841
+
+
+@functools.cache
+def make(ndim):
+    return ninetoothed.make(arrangement, application,(Tensor(ndim), Tensor(ndim)))

src/ntops/torch.py

@@ -9,6 +9,7 @@
 import ntops.kernels.gelu
 import ntops.kernels.mm
 import ntops.kernels.mul
+import ntops.kernels.relu
 import ntops.kernels.rsqrt
 
 
@@ -119,6 +120,17 @@ def mul(input, other, *, out=None):
     return out
 
 
+def relu(input, *, out=None):
+    if out is None:
+        out = torch.empty_like(input)
+
+    kernel = ntops.kernels.relu.make(input.ndim)
+
+    kernel(input, out)
+
+    return out
+
+
 def rsqrt(input, *, out=None):
     if out is None:
         out = torch.empty_like(input)

tests/test_relu.py

@@ -0,0 +1,20 @@
+import pytest
+import torch
+import torch.nn.functional as F
+
+import ntops.torch
+from tests.skippers import skip_if_cuda_not_available
+from tests.utils import 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)
+
+    ninetoothed_output = ntops.torch.relu(input)
+    reference_output = F.relu(input)
+
+    assert torch.allclose(ninetoothed_output, reference_output, atol=atol, rtol=rtol)