numo-narray

numo-narray copied to clipboard

How to implement?

masa16

Lots of memcpy are issued on a.dot(b.transpose)

6

With numo-linalg (I did not investigate without numo-linalg), ```ruby a = Numo::SFloat.new(3, 4).seq(0) b = Numo::SFloat.new(3, 4).seq(0) a.dot(b.transpose) ``` issues 12 times of memcpy at https://github.com/ruby-numo/numo-narray/blob/7bba089c3114ff08d110e47ba8bc448aa775f0d4/ext/numo/narray/ndloop.c#L1127 This drastically slows down...

sonots

[Propose] Stop making a view of NArray at #aref (or Stop using idx)

1

In Numo, advanced indexing returns a view of a Numo NArray. ``` irb(main):007:0> a = Numo::Int32.new(3,4).seq => Numo::Int32#shape=[3,4] [[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]]...

sonots

[Propose] Return 0-dimensional array instead of Ruby numeric object

1

This is a feedback from Cumo implementation. https://speakerdeck.com/sonots/fast-numerical-computing-and-deep-learning-in-ruby-with-cumo?slide=42 Numo: ``` irb(main):015:0> a = Numo::Int32.new(3,4).seq => Numo::Int32#shape=[3,4] [[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]] irb(main):016:0> a.sum...

sonots

In the case of numpy, we can create a contiguous array from a broadcasted array as: ``` In [11]: a = np.arange(3).reshape(3, 1) In [12]: a.strides Out[12]: (8, 8) In...

sonots

Not sure whether this issue goes here or in `numo-linalg`. ```ruby require 'numo/narray' # => true a = Numo::DFloat.new(5).seq # => Numo::DFloat#shape=[5] # [0, 1, 2, 3, 4] b =...

giuse

Not sure whether this should go here or in `numo-linalg`. ```ruby require 'numo/narray' # => true a = Numo::DFloat.zeros(1,10) # => Numo::DFloat#shape=[1,10] # [[0, 0, 0, 0, 0, 0, 0,...

giuse

Bug in `Numo::DFloat#sort`

2

I include the exact values with which I found this bug to be reproducible, as it does not present itself with the usual `Numo::DFloat.new(20).rand`: ```ruby a = Numo::NArray[0.0, 0.07429146315444714, 0.0,...

giuse

Optimize marshaling in parallel computation by allowing constructing Numo arrays from pre-allocated data pointers, such as coming from views.

10

In modern machine learning, parallel computation is a necessity. In Ruby this currently implies using methods based on `#fork` rather than `#thread`, since the goal is distributing the computation on...

giuse

API and method return types unification by making Numo a tensor library [ex: Arithmetic overflow]

8

**[EDIT]:** What I initially thought being an arithmetic overflow problem revealed itself to be a precise design choice (plus missing features/bugs) which led to the following discussion. ------ Using specialized...

giuse