MCPcopy Create free account
hub / github.com/RightNow-AI/autokernel / kernel_fn

Function kernel_fn

kernels/softmax.py:57–83  ·  view source on GitHub ↗

Entry point called by bench.py. Must match reference.softmax_ref signature.

(x: torch.Tensor)

Source from the content-addressed store, hash-verified

55
56
57def kernel_fn(x: torch.Tensor) -> torch.Tensor:
58 """Entry point called by bench.py. Must match reference.softmax_ref signature."""
59 assert x.is_cuda
60
61 # Flatten to 2D for row-parallel processing
62 orig_shape = x.shape
63 if x.ndim == 1:
64 x = x.unsqueeze(0)
65 elif x.ndim > 2:
66 x = x.view(-1, x.shape[-1])
67
68 n_rows, n_cols = x.shape
69 output = torch.empty_like(x)
70
71 # Block size must be a power of 2 >= n_cols
72 BLOCK_SIZE = triton.next_power_of_2(n_cols)
73
74 grid = (n_rows,)
75 softmax_kernel[grid](
76 x, output,
77 n_cols,
78 x.stride(0),
79 output.stride(0),
80 BLOCK_SIZE=BLOCK_SIZE,
81 )
82
83 return output.view(orig_shape)

Callers

nothing calls this directly

Calls

no outgoing calls

Tested by

no test coverage detected