Merge pull request #36 from InfiniTensor/develop-silu

Add `silu` operator

Author: voltjia

src/ntops/kernels/silu.py

@@ -0,0 +1,18 @@
+import functools
+
+import ninetoothed
+import ninetoothed.language as ntl
+from ninetoothed import Tensor
+
+from ntops.kernels.element_wise import arrangement
+
+
+def application(input, output):
+    output = input / (1 + ntl.exp(-ntl.cast(input, ntl.float32)))  # noqa: F841
+
+
+@functools.cache
+def make(ndim):
+    tensors = (Tensor(ndim), Tensor(ndim))
+
+    return ninetoothed.make(arrangement, application, tensors)

src/ntops/torch.py

@@ -30,6 +30,7 @@
 import ntops.kernels.relu
 import ntops.kernels.rsqrt
 import ntops.kernels.sigmoid
+import ntops.kernels.silu
 import ntops.kernels.sin
 import ntops.kernels.softmax
 import ntops.kernels.sub
@@ -362,6 +363,19 @@ def sigmoid(input, *, out=None):
     return out
 
 
+def silu(input, inplace=False):
+    if inplace:
+        output = input
+    else:
+        output = torch.empty_like(input)
+
+    kernel = ntops.kernels.silu.make(input.ndim)
+
+    kernel(input, output)
+
+    return output
+
+
 def sin(input, *, out=None):
     if out is None:
         out = torch.empty_like(input)

tests/test_silu.py

@@ -0,0 +1,21 @@
+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)
+
+    # TODO: Add `inplace` tests later.
+    ninetoothed_output = ntops.torch.silu(input)
+    reference_output = F.silu(input)
+
+    assert torch.allclose(ninetoothed_output, reference_output, atol=atol, rtol=rtol)