Add pow operator

Author: voltjia

src/ntops/kernels/pow.py

@@ -0,0 +1,18 @@
+import functools
+
+import ninetoothed
+from ninetoothed import Tensor
+from ninetoothed.language import libdevice
+
+from ntops.kernels.element_wise import arrangement
+
+
+def application(input, exponent, output):
+    output = libdevice.pow(input, exponent)  # noqa: F841
+
+
+@functools.cache
+def make(ndim):
+    tensors = (Tensor(ndim), Tensor(ndim), Tensor(ndim))
+
+    return ninetoothed.make(arrangement, application, tensors)

src/ntops/torch.py

@@ -26,6 +26,7 @@
 import ntops.kernels.mul
 import ntops.kernels.ne
 import ntops.kernels.neg
+import ntops.kernels.pow
 import ntops.kernels.relu
 import ntops.kernels.rsqrt
 import ntops.kernels.sigmoid
@@ -314,6 +315,17 @@ def neg(input, *, out=None):
     return out
 
 
+def pow(input, exponent, *, out=None):
+    if out is None:
+        out = torch.empty_like(input)
+
+    kernel = ntops.kernels.pow.make(input.ndim)
+
+    kernel(input, exponent, out)
+
+    return out
+
+
 def relu(input, inplace=False):
     if inplace:
         output = input

tests/test_pow.py

@@ -0,0 +1,26 @@
+import pytest
+import torch
+
+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):
+    # TODO: Test for `float16` later.
+    if dtype is torch.float16:
+        return
+
+    device = "cuda"
+
+    input = torch.randn(shape, dtype=dtype, device=device)
+    exponent = torch.randn(shape, dtype=dtype, device=device)
+
+    ninetoothed_output = ntops.torch.pow(input, exponent)
+    reference_output = torch.pow(input, exponent)
+
+    assert torch.allclose(
+        ninetoothed_output, reference_output, atol=atol, rtol=rtol, equal_nan=True
+    )