Make filter compute shaders use 2D grid of no more than 64K blocks on a single axis - Githubissues

ngscopeclient / scopehal

Test and measurement hardware abstraction library and protocol decodes. This is the library only. Most users should use scopehal-apps.

BSD 3-Clause "New" or "Revised" License

216 stars 97 forks source link

Make filter compute shaders use 2D grid of no more than 64K blocks on a single axis #716

Open azonenberg opened 2 years ago

azonenberg commented 2 years ago

This should give us max compatibility with e.g. Intel cards.

On nvidia it's easy to just run up to 2G blocks on one axis, but Intel iGPUs have much smaller limits.

azonenberg commented 6 months ago

Shaders to refactor (this does NOT imply calling code has been refactored to match):

[ ] Convert16BitSamples: x=sample index
[ ] Convert8BitSamples: x=sample index
[ ] AddFilter: x = sample index
[ ] BlackmanHarrisWindow: x=sample index with block offset
[ ] ComplexBlackmanHarrisWindow: x=sample index with block offset
[ ] ComplexCosineSumWindow: x=sample index with block offset
[ ] ComplexRectangularWindow: x =sample index with block offset
[x] ComplexSpectrogramPostprocess: x=pixel index, y/z = FFT bin
[ ] ComplexToLogMagnitude: x=sample index
[ ] ComplexToMagnitude: x=sample index
[ ] CosineSumWindow: x=sample index with block offset
[ ] DeEmbedFilter: x=sample index
[ ] DeEmbedNormalization: x=sample index
[ ] DeEmbedOutOfPlace: x=sample index
[ ] FIRFIlter: x=sample index
[ ] RectangularWindow: x=sample index with block offset
[x] SpectrogramPostprocess: x=pixel index, y/z = FFT bin
[ ] SubtractFilter: x=sample index with P/N side input offsets
[ ] SubtractInPlace: x=sample index
[ ] SubtractOutOfPlace: x=sample index
[ ] UpsampleFilter: x=input sample index, y=subsample index
[ ] WaterfallFilter: x=width, y=height

Filters to refactor (this list is likely incomplete)

[ ] FFTFilter: always use offset in/out of zero
[ ] SpectrogramFilter: block based, but doesn't account for FFTs larger than thread block size