Skip to content

Fix FFT failure for array sizes 2^21 and 2^22 #2997

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Anri-Lombard wants to merge 1 commit into
base: main
Choose a base branch
from
+10 −1
Open

Fix FFT failure for array sizes 2^21 and 2^22 #2997

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 2 additions & 1 deletion mlx/backend/metal/fft.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@ namespace mlx::core {
using MTLFC = std::tuple<const void*, MTL::DataType, NS::UInteger>;

#define MAX_STOCKHAM_FFT_SIZE 4096
#define MAX_SAFE_FFT_SIZE 1024
#define MAX_RADER_FFT_SIZE 2048
#define MAX_BLUESTEIN_FFT_SIZE 2048
// Threadgroup memory batching improves throughput for small n
Expand Down Expand Up @@ -121,7 +122,7 @@ FFTPlan plan_fft(int n) {
int remaining_n = n;

// Four Step FFT when N is too large for shared mem.
if (n > MAX_STOCKHAM_FFT_SIZE && is_power_of_2(n)) {
if (n > MAX_SAFE_FFT_SIZE && is_power_of_2(n)) {
// For power's of two we have a fast, no transpose four step implementation.
plan.four_step = true;
// Rough heuristic for choosing faster powers of two when we can
Expand Down
8 changes: 8 additions & 0 deletions python/tests/test_fft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -318,6 +318,14 @@ def g(x):
dgdx = torch.func.grad(g)(torch.tensor(x))
self.assertLess((dfdx - dgdx).abs().max() / dgdx.abs().mean(), 1e-4)

def test_fft_large_powers_of_two(self):
for power in [21, 22, 23]:
size = 2**power
x = mx.ones(size, dtype=mx.complex64)
result = mx.fft.fft(x, stream=mx.gpu)
mx.eval(result)
self.assertEqual(result.shape[0], size)


if __name__ == "__main__":
mlx_tests.MLXTestRunner()