Sohaib
71ae6e5605
fix: avgpool without counting padding ( #605 )
2023-02-26 07:13:00 -08:00
George Hotz
2b6d8e788d
pushing permutes is done
2023-02-25 21:37:46 -08:00
George Hotz
f4ee7d2cad
back to 196 kernels
2023-02-25 18:25:34 -08:00
George Hotz
187738f64f
test fix, wow fixing contig exposed so much
2023-02-25 18:01:14 -08:00
George Hotz
c3d19101cb
more debug info
2023-02-25 17:52:46 -08:00
George Hotz
6e98a172a0
fix broken contiguous
2023-02-25 17:41:49 -08:00
George Hotz
a44e8e4385
discard children on mop shuffle, 200 -> 196 kernels
2023-02-25 10:51:07 -08:00
George Hotz
1d01842232
remove fake test
2023-02-25 10:21:07 -08:00
George Hotz
8b96522e1d
instant identity removal
2023-02-25 09:46:04 -08:00
George Hotz
a8de233e12
only div, no reciprocal ( #601 )
...
* only div, no reciprocal
* remove reciprocal
* fix pad shuffling
2023-02-25 09:35:03 -08:00
Sohaib
d581a99d90
onnx: lrn ( #602 )
...
Co-authored-by: Sohaib Errabii <errabii.sohaib@gmail.com >
2023-02-25 09:24:53 -08:00
voidz
94bec40110
moved extras/jit.py -> tinygrad/jit.py ( #599 )
...
* moved extras/jit.py to tinygrad/jit.py
* fixed indent
* removed tinygrad.helpers.DEBUG from jit.py
2023-02-25 08:32:33 -08:00
Benedikt Mandelkow
7348e9a6c6
add restrict qualifier to inputs in c backend ( #593 )
...
* add restrict qualifier for clang backend convolution inputs/ outputs
see https://godbolt.org/z/Tb9jMxWfx for generated assembly
* enable more checks
* inline fmax to motivate the compiler to inline some more
* fix if else binding power
2023-02-25 08:32:21 -08:00
George Hotz
2c5e13a513
Reluless ( #600 )
...
* replace relu for maximum
* fix for other backend
* clean up RELU and GT0
* tests for maximum
* had to clean that up
* why reverse a maximum?
2023-02-25 01:21:16 -08:00
George Hotz
176ad29974
retain support for old onnx
2023-02-24 22:29:54 -08:00
George Hotz
f3386c7f09
improve symbolic, hlop conv output is simple now
2023-02-24 22:20:40 -08:00
George Hotz
9b5b17db54
shapetracker : always 0 out the strides if the shape is 1
2023-02-24 21:55:38 -08:00
George Hotz
f8f026e8bb
oversized expand for HLOP convs
2023-02-24 21:48:47 -08:00
George Hotz
2edfe64512
improve shapetracker tests
2023-02-24 21:07:53 -08:00
Marcello Fuschi
6d97d62ab3
Add PyCharm's .idea to .gitignore ( #597 )
2023-02-24 20:14:38 -08:00
George Hotz
da5643d024
rest of tests shouid be made to pass
2023-02-24 12:52:23 -08:00
George Hotz
85452fbaf3
onnx 58/109/208
2023-02-24 12:19:05 -08:00
George Hotz
e8a153e4e9
onnx : add a whole bunch of ops
2023-02-24 12:00:03 -08:00
George Hotz
f2486a7248
more onnx ops
2023-02-24 10:55:58 -08:00
George Hotz
4d0a3dd653
openpilot expand is bugged
2023-02-24 10:25:59 -08:00
George Hotz
2e56a4793e
rename log_softmax, support dim, fix onnx Softmax
2023-02-24 10:11:24 -08:00
George Hotz
5cdfeffe2c
fix shape test
2023-02-24 09:36:32 -08:00
George Hotz
3becefa218
fix onnx tests
2023-02-24 09:27:18 -08:00
George Hotz
e263c0c628
onnx : another model test is passing
2023-02-24 09:22:58 -08:00
George Hotz
d3feea302d
much cleaner way to write onnx ops
2023-02-24 08:46:28 -08:00
George Hotz
d3029c91c5
no rng for op test
2023-02-24 00:23:20 -08:00
George Hotz
f6d946853c
more bugfixes
2023-02-24 00:21:29 -08:00
George Hotz
b1b2d8f440
onnx : some op tests working
2023-02-23 23:58:13 -08:00
George Hotz
b287b1d529
fix yolov8 to get to ConvTranspose
2023-02-23 22:46:48 -08:00
George Hotz
2d59b25ead
onnx backend test : enable only the model tests
2023-02-23 22:36:26 -08:00
George Hotz
5b10dfcab8
onnx tests : 22/175/208
2023-02-23 22:00:16 -08:00
George Hotz
d8b6f241f1
external_test_onnx_backend
2023-02-23 21:55:07 -08:00
George Hotz
edaf878339
put rng with creation ops
2023-02-23 20:38:36 -08:00
George Hotz
8137e0b400
typo in __imatmul__ -> cleanup
2023-02-23 20:13:44 -08:00
George Hotz
4c54adeb18
adding tuples is fine
2023-02-23 19:42:48 -08:00
George Hotz
661812ffef
don't ignore type
2023-02-23 19:38:52 -08:00
George Hotz
10c6ccf7e0
better order in Tensor class
2023-02-23 19:33:37 -08:00
George Hotz
f2ca81c66d
reduce cleanup
2023-02-23 19:19:49 -08:00
George Hotz
8446f77fe9
image_conv2d_decorator
2023-02-23 18:49:23 -08:00
George Hotz
82caa2d5b7
MULACC -> FusedOp
2023-02-23 18:17:57 -08:00
George Hotz
758515dcc0
conv2d is an hlop ( #589 )
...
* conv2d is an hlop
* shorter conv
* KOPT=-1
* alt imp
* MULACC
* smarter mulacc
* pop conv
* 7x7 -> 5x5
* didn't fix, that's not going to work
* this is faster and matches old behavior
* oh, non lazy just won't work with mulacc
* mulacc in torch
* bool types were creeping in
* optimizer is actually better with hlop conv
* fix pushing permutes issue
* refactor einsum_mulacc
* fix up readme
* update readme
* _image_conv2d
* fix bias addition location
* pushing permutes gets back to 200 kernels
* conv cleanup
* disable hlop conv
* don't hide that in helpers
2023-02-23 17:52:31 -08:00
Sohaib
8835df7a5c
upgrade onnx to 1.13.0 ( #588 )
...
- remove protobuf from direct dependencies
- replace deprecated mapping.TENSOR_TYPE_TO_NP_TYPE
Co-authored-by: Sohaib Errabii <sohaib.errabii@ipops.io >
2023-02-23 13:59:23 -08:00
calledit
81f7c6800a
Added info on simdgroup availability ( #586 )
...
* Add info on simdgroup availability
* "osx" not "os x"
* Update metal_matmul.py
* Update metal_matmul.py
2023-02-23 13:59:02 -08:00
George Hotz
d22e19536b
onnx: support low quality Resize. stuck on ConvTranspose will have to wait for convless
2023-02-23 09:05:23 -08:00
George Hotz
ab3a2ae9a2
fix test_resnet in onnx now that maxpool works
2023-02-23 08:41:47 -08:00
