Skip to content

Commit 23b4e1e

Browse files
Update IDEA.md
1 parent fa3ad31 commit 23b4e1e

1 file changed

Lines changed: 4 additions & 0 deletions

File tree

IDEA.md

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -45,6 +45,7 @@ wrapper, tests, examples, docs, and a validated Makefile workflow.
4545
- direct ABS grab
4646

4747
```
48+
4849
With a real-only Bruun interior, fused depth-first traversal, and leaf codelets that currently materialize re/im complex bins at the bottom. The leaf packers are where X[k].re and X[k].im are finally produced, and the recursive transform bottoms out in fused depth-3 codelets.
4950
5051
The magnitude-only version should not do this:
@@ -136,10 +137,13 @@ standard power: much cheaper than standard complex
136137
FFTW+power: slower than FFTW alone
137138
138139
the route is real. The magnitude-only case strengthens Bruun’s advantage because you can throw phase away before it ever hits memory.
140+
139141
```
140142

141143
- High-Speed DCT
144+
142145
```
146+
143147
It is a Chebyshev-Bruun kernel.
144148
145149
DCTs are already the reciprocal-even half of a Fourier transform. Ordinary FFT-based DCT code pays for a cyclic complex object, then projects down to cosines. BFFT is already closer to the projection because its real quadratic leaves are conjugate-pair planes. A DCT kernel can go one step further: never build the conjugate-pair plane as a complex output. Work directly in the fixed reciprocal basis where a leaf is just evaluation at x = cos theta.

0 commit comments

Comments
 (0)