You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
It will probably be generally useful to include a page (or pages) in the lib_xs3_math documentation regarding issues that may be encountered where the BFP implementation of an algorithm has intermediate results with less precision than is desirable or would be ideal.
These might include:
Problems that may be encountered when using vectors with a large dynamic range (e.g. most frequency-domain vectors)
Problems fundamental to BFP such as multiplying two vectors where each vector's headroom is determined by an element at a different index (which is opaque to the BFP functions) and so the multiply makes incorrect (but necessary!) assumptions about the worst case result (to avoid saturation)
Problems specific to the XS3 architecture. For example, the fact that 32-bit multiplies (by the VPU) always include an implicit 30-bit right-shift.
(Note that #89 was a case involving all three of these.)
The documentation page(s) should include simplified examples that demonstrate the problem(s) and should suggest possible remedies (such as splitting up BFP vectors or element-wise scaling of vector components)
The text was updated successfully, but these errors were encountered:
This comes out of issue #89.
It will probably be generally useful to include a page (or pages) in the lib_xs3_math documentation regarding issues that may be encountered where the BFP implementation of an algorithm has intermediate results with less precision than is desirable or would be ideal.
These might include:
(Note that #89 was a case involving all three of these.)
The documentation page(s) should include simplified examples that demonstrate the problem(s) and should suggest possible remedies (such as splitting up BFP vectors or element-wise scaling of vector components)
The text was updated successfully, but these errors were encountered: