global Variable for sizing

[SReiter149]

Hi I'm new to Fxpmath,

I was digging around in the code and I noticed that the default sizing is optimal, I was wondering if there was an easy way to easily set all variables to output to 'same' sizing instead? I have too many variables to manually do that for everyone so I was hoping that you guys have a global variable that will let me set that as the default instead.

[SReiter149]

here is an example piece of code as to why I am asking this question:

x = Fxp(np.random.random((1,400)),True,32,10) y = Fxp(np.random.random((400,128)),True,32,10) z = np.dot(x,y) print(z.info(3))

showing the info shows that the word bits are 73 and I understand that you can't have more than 64 bits because of NumPy but I was hoping there was a way to keep it below 64, preferably keeping it to 32 but I just want a result that isn't all 0s.

[francof2a]

Hello @Hamuel101,
You're right about that default sizing is optimal, and there is a way to define a template for new Fxp objects:

from fxpmath import Fxp

Fxp.template = Fxp(None, True, 32, 10, op_sizing='same') # enable template

x = Fxp(np.random.random((1,400))) 
y = Fxp(np.random.random((400,128)))

Fxp.template = None # disable template

But checking with your example, I found that config parameters are set with default values, so op_sizing is set with 'optimal' again (I'm going to open an issue with this).

There is another concern is about your example. The function np.dot doesn't know how to resolve sizing from op_sizing property of x or y (it could be another issue), you can use some alternatives by now:

z = x.dot(y) # op_sizing of x is used

or

import fxpmath as fxp

z = fxp.dot(x, y, sizing='same')

Any suggestion is very welcome!

[SReiter149]

thanks! That worked

[SReiter149]

Apologies for another question. I was wondering if there is a way to have optimal sizing but up to a bit cap, so for example it can have a max n_word of 32 but have any amount of n_bits?

[francof2a]

Not problem about questions!
By now there is not a limit for output sizing. It could be added in next revision.

The only limit implemented by now is n_word_max in config, but it's only used when a new Fxp is created without a specified size.
I don't know details about your needs, but a workaround could be resize the Fxp obtained from an operation under a condition:

# ...

N_WORD_MAX = 16

# ...

z.resize(n_word=min(z.n_word, N_WORD_MAX)) # where z is the Fxp object to limit its size

I think other way is using on_value_change callback previously defined in template. Inside that callback you can execute the code above to limit the size. (I have not tested this, but if you are interested, I could do it).

[SReiter149]

ok thanks! I will look into this and see what I find.

[SReiter149]

all of your suggestions have been very helpful and I am now running into an issue with division. I have been doing as you suggested with fxp.dot and other functions like that. I am coming across a problem where I can't find a fxp.div or fxp.divide or anything like that. So I tried doing fxp.mul(x,1/y)but 1/y goes outside of the range I am allowing it so overflows.

Fxp.template = Fxp(None, True,8 ,7, op_sizing='same') x = Fxp(0.635432456768467845673456) y = Fxp(0.13451234523651234) z = fxp.mul(x,1/y) print(z)

it prints: 0.62786865234375

I can only give it 8 bits for other reasons so can not give integer bits for this purpose. Any suggestions?

On a side note can I specify rounding in the template like, this doesn't seem to work either: Fxp.template = Fxp(None, True,8 ,7, op_sizing='same',rounding = 'around')

[francof2a]

The functions for division are floordiv (// python operator) and truediv (/ python operator), these functions are from fxpmath.functions. Numpy function that are implemented by those functions are floor_divide and true_divide, correspondingly.

Following examples must be working:

z = x / y
z = np.true_divide(x, y)
z = fxp.truediv(x, y)

The np.divide implements np.true_divide. It's true that divide is not mapped (dispatched) by now, but it will be added now you realized it!

About template issue, could you tell me which fxpmath version are you using?

import fxpmath as fxp

fxp.__version__

[SReiter149]

yeah sure, I am on version 0.4.0.

[francof2a]

Last version 0.4.2 fixes some template issues. So, update maybe works.
keep me posted, please.

[SReiter149]

Yeah, updating it helped. Thanks!

[SReiter149]

I have another question. again apologies for your time.

Fxp.template = Fxp(None, True, 8,7, op_sizing='same',rounding = 'around') x = np.random.random((4,3,5,7)) y = Fxp(x) z = np.transpose(y, (3,2,0,1)) print(z.shape) Fxp.template = None

Error: TypeError: out must be a Fxp object!

Is there a reason you guys raise an error instead of just turning the output into an Fxp object? I feel like you guys could just say

if not isinstance(out, Fxp): out = Fxp(out))

or else is there a way to transpose and Fxp object?

[francof2a]

No problem at all about questions!

By now, the transpose method takes axes parameters only as kwarg, so:

z = np.transpose(y, axes=(3,2,0,1))
z = y.transpose(axes=(3,2,0,1))

should work.

If no axes is specified, following must work also:

z = y.T

PS: There is no problem about open a new discussion if the new question is not about original topic.