Add blocked convolution - #768

Open

FedyuninV wants to merge 79 commits into

neoml-lib:masterfrom

FedyuninV:BlockedConvExperiments

Commits on Aug 26, 2022

Fix current progress
Valeriy Fedyunin
committed
89574a9
Fix a couple more bugs
Valeriy Fedyunin
committed
f1e664b
Organize tests in a better way
Valeriy Fedyunin
committed
75bccb1
Fix Trivial/PaddingHeight test
Valeriy Fedyunin
committed
05ac4d9
HELLL YEAHHHH It passes all the given test!!!
Valeriy Fedyunin
committed
c52427d
Add tests from yolox
Valeriy Fedyunin
committed
5f68ebe

Commits on Aug 29, 2022

Add simple performance check
Valeriy Fedyunin
committed
93550c2
Use COMPUTE_BLOCK macro
Valeriy Fedyunin
committed
affcfe1
Move ClearBlock to macro
Valeriy Fedyunin
committed
96b189c
Move ProcessFilterCountN to macro
Valeriy Fedyunin
committed
497175b
Move PROCESS_OUTPUT_COUNT_N to macro
Valeriy Fedyunin
committed
40601bd

Commits on Aug 31, 2022

Optimize PackData and UnpackData functions
Valeriy Fedyunin
committed
4bc7630
Fix convDesc leak
Valeriy Fedyunin
committed
5d43103
Refactor measurements
Valeriy Fedyunin
committed
6b019b2

Commits on Sep 1, 2022

Add temp code
Valeriy Fedyunin
committed
ecf9efc

Commits on Sep 24, 2022

DEBUG (expand ISimdMathEngine interface)
Valeriy Fedyunin
committed
21eab6e
DEBUG add AvxTestDesktop
Valeriy Fedyunin
committed
f76fbf0

Commits on Sep 25, 2022

Test commit
Valeriy Fedyunin
committed
8d3cda2

Commits on Sep 27, 2022

Add blocked convolution with macro to NeoMathEngineAvx
Valeriy Fedyunin
committed
3c4b03d
Fix compilation warning in NeoMathEngine test
Valeriy Fedyunin
committed
6ae35b5
Add blocked convolution test for NeoMathEngineAvx
Valeriy Fedyunin
committed
aa50778

Commits on Oct 9, 2022

EXPERIMENTAL: start moving to JIT
Valeriy Fedyunin
committed
6af3f0d

Commits on Oct 17, 2022

Use single generator for all of the broadcast values
Valeriy Fedyunin
committed
9aac8e6
Remove unused code
Valeriy Fedyunin
committed
d20c990

Commits on Oct 18, 2022

Pass arguments to JIT via struct
Valeriy Fedyunin
committed
69e8d29

Commits on Oct 19, 2022

Move kernelHeight and kernelWidth loops to JIT
Valeriy Fedyunin
committed
0082a87
Move CLEAR_BLOCK macro to JIT
Valeriy Fedyunin
committed
2299071
Move postProcessing to JIT
Valeriy Fedyunin
committed
76a7519

Commits on Oct 21, 2022

Use loops during JIT
Valeriy Fedyunin
committed
b730e7d
Even less lines
Valeriy Fedyunin
committed
f56f27c

Commits on Oct 22, 2022

Manually assign registers
Valeriy Fedyunin
committed
f303189

Commits on Oct 23, 2022

Use RSP instead of Param1
Valeriy Fedyunin
committed
0983e8f
Shorten the macro (a little bit)
Valeriy Fedyunin
committed
e973333
Wrap generator into C++ class (will be useful in future)
Valeriy Fedyunin
committed
c7cdfae
Move filterCount and outputCount to generation parameters
Valeriy Fedyunin
committed
610d32a
Move PROCESS_FILTER_COUNT_N to JIT
Valeriy Fedyunin
committed
e920418
Move all we need to JIT!!!
Valeriy Fedyunin
committed
d92b687

Commits on Oct 24, 2022

Merge branch 'master' into BlockedConvExperiments
Valeriy Fedyunin
committed
f7b74e0
Disable NeoAvxTestDesktop for 32-bit and for FineObjects
Valeriy Fedyunin
committed
7497091
Disable BlockedConv tests for FineObjects
Valeriy Fedyunin
committed
0456104
Use blocked conv when possible
Valeriy Fedyunin
committed
74e15b2

Commits on Oct 25, 2022

DEBUG: temporary disable blocked conv
Valeriy Fedyunin
committed
874ecfc
Extend test set
Valeriy Fedyunin
committed
f5be1d8
Enable blocked convolution when heuristics are OK
Valeriy Fedyunin
committed
f97f47f

Commits on Oct 26, 2022

Add more statistics to BlockedConv Real tests
Valeriy Fedyunin
committed
fc7df2e
Overflow protection
Valeriy Fedyunin
committed
93a5ac6
Reduce the number of ZEROUPPER calls
Valeriy Fedyunin
committed
98dfc10
Better tuning
Valeriy Fedyunin
committed
b5b0c3a
Reduce memory consumption
Valeriy Fedyunin
committed
1726218
Add even more tests to NeoAvxTestDesktop
Valeriy Fedyunin
committed
85c40b5

Commits on Oct 28, 2022

Start cleaning up this mess...
Valeriy Fedyunin
committed
067e1c8
Use calcOutputPad for width
Valeriy Fedyunin
committed
011332d
Don't depend on the order of fields in CBlockedConvGen::CParams
Valeriy Fedyunin
committed
64bcd1f
More parameter renamings and optimizations
Valeriy Fedyunin
committed
1d3d647
Fix comment
Valeriy Fedyunin
committed
247521c
Use more clear indexes for Ymm accumulators
Valeriy Fedyunin
committed
a8d9b79
Clarify bias Ymms
Valeriy Fedyunin
committed
9621ba2

Commits on Oct 29, 2022

Clarify a bit more
Valeriy Fedyunin
committed
9fa5a15
Improve naming
Valeriy Fedyunin
committed
572e277
Optimize passing arguments runConv
Valeriy Fedyunin
committed
3963962
Add multithreading
Valeriy Fedyunin
committed
cf1917d
Remove fixed TODO comment
Valeriy Fedyunin
committed
2385013

Commits on Oct 31, 2022

Turn on all the heuristics
Valeriy Fedyunin
committed
008a9fa
Use NeoML terms
Valeriy Fedyunin
committed
f75dffb
Reorder fields
Valeriy Fedyunin
committed
7090124
More renamings
Valeriy Fedyunin
committed
6c5700f
Add more comments
Valeriy Fedyunin
committed
63fe462

Commits on Nov 14, 2022

Fix some naming and add some comments
Valeriy Fedyunin
committed
9a9272e
Minor optimization and even more comments
Valeriy Fedyunin
committed
d0b54ca
Fix test skip condition
Valeriy Fedyunin
committed
857d06c
Fix non-MSVC compilation
Valeriy Fedyunin
committed
ede4132
Reduce diff size
Valeriy Fedyunin
committed
25fc832

Commits on Nov 15, 2022

Refactor tests
Valeriy Fedyunin
committed
af96c8b
Add more heuristics for blocked convolution
Valeriy Fedyunin
committed
591a047
Grammar fix
Valeriy Fedyunin
committed
75aa7ac
Disable AvxTestDesktop everywhere but Windows
Valeriy Fedyunin
committed
ab0638c

Commits on Dec 10, 2022

Merge branch 'master' into BlockedConvExperiments
Valeriy Fedyunin
committed
f48de7c
Fix compilation
Valeriy Fedyunin
committed
69dc1f1

Commits on Dec 19, 2022

Disable blocked conv for AVX512
Valeriy Fedyunin
committed
1db000d