Bug/789 pow binary op performance

[coquelin77]

Description

Optimizations for pow

Issue/s resolved: #789 although resolved is a strong word...more work required

Changes proposed:

• avoid calling _binary_op unless absolutely necessary and use a simpler switch instead

Type of change

Optimization

Due Diligence

• [x] All split configurations tested
• [x] Multiple dtypes tested in relevant functions
  • this may have an effect on the dtype results. however the tests ran clean with multiple different operations in var which requires this
• [x] Documentation updated (if needed)
• [ ] Updated changelog.md under the title "Pending Additions"

Does this change modify the behaviour of other functions? If so, which?

YES! comm.chunk now returns 4 parameters instead of 3!

[codecov[bot]]

Codecov Report

Merging #793 (b9f833d) into main (05a2acd) will decrease coverage by 0.18%. The diff coverage is 85.89%.

@@            Coverage Diff             @@
##             main     #793      +/-   ##
==========================================
- Coverage   91.12%   90.93%   -0.19%     
==========================================
  Files          65       65              
  Lines        9976    10143     +167     
==========================================
+ Hits         9091     9224     +133     
- Misses        885      919      +34     

Flag	Coverage Δ
gpu	?
unit	90.93% <85.89%> (-0.17%)	:arrow_down:

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
heat/core/_operations.py	96.04% <ø> (ø)
heat/core/stride_tricks.py	81.53% <50.00%> (ø)
heat/core/arithmetics.py	90.86% <81.35%> (-7.74%)	:arrow_down:
heat/core/communication.py	89.96% <100.00%> (+0.01%)	:arrow_up:
heat/core/dndarray.py	96.66% <100.00%> (+<0.01%)	:arrow_up:
heat/core/factories.py	99.23% <100.00%> (-0.01%)	:arrow_down:
heat/core/indexing.py	100.00% <100.00%> (ø)
heat/core/io.py	89.39% <100.00%> (-0.05%)	:arrow_down:
heat/core/linalg/basics.py	94.22% <100.00%> (ø)
heat/core/manipulations.py	98.63% <100.00%> (ø)
... and 6 more

[coquelin77]

rerun tests

[coquelin77]

rerun tests

[coquelin77]

rerun tests

[ghost]

👇 Click on the image for a new way to code review

Legend

[coquelin77]

New numbers:

local tests: 4 procs - a = ht.random.random((10000, 10000), dtype=ht.float32, split=split) numbers are avg of 10 runs

• pow (new / old)
  • a ** a: 0.2157 / 0.2207
  • 2.5 ** a: 0.2099 / 0.2051
  • a ** 2.5: 0.1986 / 0.2002
  • a ** 2: 0.0709 / 0.2390
• add (new / old)
  • a + a: 0.0531 / 0.0697
  • 2.5 + a: 0.0571 / 0.0603
  • a + 2.5: 0.0597 / 0.0615
  • a + 2: 0.0578 / 0.0592
• sub (new / old)
  • a - a: 0.0590 / 0.0689*
  • 2.5 - a: 0.0584 / 0.0570
  • a - 2.5: 0.0571 / 0.0583
  • a - 2: 0.0587 / 0.0593
• div (new / old)
  • a / a: 0.0582 / 0.0580
  • 2.5 a: 0.0582 / 0.0574
  • a / 2.5: 0.0615 / 0.0596
  • a / 2: 0.0540 / 0.0601
• floordiv (new / old)
  • a // a: 0.1291 / 0.1374
  • 2.5 // a: 0.1342 / 0.1371
  • a // 2.5: 0.1339 / 0.1406
  • a // 2: 0.1886 / 0.1922
• mul (new / old)
  • a * a: 0.0557 / 0.0579
  • 2.5 * a: 0.0580 / 0.0572
  • a * 2.5: 0.0600 / 0.0595
  • a * 2: 0.0600 / 0.0581

[coquelin77]

@mtar can you have a look at why this is failing? i dont know why gpu isnt recognized. maybe its a but somewhere that im missing

[ClaudiaComito]

Some of this PR has been superseded by the latest implementation of _operations.__binary_op with distribution sanitation #902 , but a lot of it is still relevant. I will make the necessary changes this week, @coquelin77 please scream if you prefer a review and to introduce the changes yourself.

[ClaudiaComito]

Closing this as too stale to update, changes implemented in #1141