commaai
/
tinygrad
mirror of https://github.com/commaai/tinygrad.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
6,388 Commits 10 Branches 0 Tags 52 MiB
Commit Graph

25 Commits

Author SHA1 Message Date
wozeparrot c100f3d406
default threefry (#6116) 2024-09-25 17:45:13 +08:00
chenyu a37e92081a
fix unrolled arange folding (#6606) 
* fix unrolled arange folding

also added flop test to test_arange to make sure it's 0 flop

* skip PTX
 2024-09-19 09:03:01 -04:00
chenyu 15c4d4f406
fold unrolled arange div pattern (#6465) 2024-09-10 22:35:52 -04:00
Roelof van Dijk ad4b3b457f
bump limit for test_llama_embedding_opt (#6332) 2024-08-31 10:03:43 -04:00
George Hotz 16f420f7a7
split full_graph_rewrite and linearize_uop [run_process_replay] (#6215) 
* split full_graph_rewrite and linearize_uop

* fix tests

* graph rewrite in test uops

* add types
 2024-08-20 20:12:33 -07:00
George Hotz a5d79688db
fix indexing out of bounds (#6208) 
* fix indeing out of bounds

* 5 ops per access is fine
 2024-08-20 11:34:56 -07:00
chenyu 4451bcaf95
update test_arange test_llama_embedding_opt (#6207) 
non CI uses larger embedding, still same orders of magnitude
 2024-08-20 13:58:43 -04:00
George Hotz cf7d3c1eb8
fix tests locally on metal (#6025) 
* remove contiguous child, it was breaking tests locally

* hmm, it's still needed

* include NOOPT in method cache key
 2024-08-10 12:36:22 -07:00
qazal 45b1761175
smaller test_llama_embedding + assert correctness (#5986) 
* smaller test_llama_embedding in CI

* test correctness
 2024-08-08 22:11:29 +03:00
George Hotz 6d1fdcfce2
don't reduce the same thing in a vector (#5950) 
* don't reduce the same thing over and over

* cleaner way to write it that doesn't loop
 2024-08-06 16:59:15 -07:00
George Hotz 3e1336957d
test arange with all opts (#5923) 
* test arange with all opts

* Update test_arange.py

* Update test_arange.py

* Update test_arange.py

* Update test_arange.py

* Update test_arange.py
 2024-08-05 18:38:25 -07:00
George Hotz 5d17f54e3c
fast mnist indexing (#5921) 
* fast mnist indexing

* more tests

* remove those tests, new indexing rule
 2024-08-05 13:55:15 -07:00
George Hotz e81c18f494
make the arange test check correctness [run_process_replay] (#5920) 2024-08-05 13:41:06 -07:00
George Hotz 42f599870c
unroll arange is broken (#5918) 
* unroll arange is broken

* fix unrolled arange

* one more test
 2024-08-05 12:15:07 -07:00
qazal 65fa86901a
indexing fusion 2 (#5888) 
* arange fusion

* kernels that fuse

* tests
 2024-08-03 13:13:39 +03:00
qazal 4e070a2c89
start work on indexing fusion (#5590) 
* start base

* the views add up

base reduceop st:
ShapeTracker(views=(View(shape=(60000, 1), strides=(1, 0), offset=0, mask=None, contiguous=True),))

top st:

ShapeTracker(views=(View(shape=(512, 6000, 1, 28, 28, 10), strides=(0, 1, 0, 0, 0, 6000), offset=0, mask=None, contiguous=False), View(shape=(512, 6000, 1, 28, 28, 10), strides=(47040000, 784, 0, 28, 1, 4704000), offset=0, mask=None, contiguous=False)))

merged buf.st+st:
ShapeTracker(views=(View(shape=(512, 6000, 1, 28, 28, 10), strides=(0, 1, 0, 0, 0, 6000), offset=0, mask=None, contiguous=False), View(shape=(512, 6000, 1, 28, 28, 10), strides=(47040000, 784, 0, 28, 1, 4704000), offset=0, mask=None, contiguous=False)))

* p1

* some cleanups

* more cleanups

* one kernel

* more

* late fuse arange

* less lines

* more work

* fix st strides 1

* update test_schedule, start argmax

* test_tiny_argmax

* add FUSE_ARANGE

* more cleanup

* add utils

* reduce merging

* fix axis and fold if needed

* more fusion

* need to figure this out

* now fixing all of these

* todos+save a line

* ready for p1
 2024-07-25 13:23:38 +03:00
George Hotz dc21e63bd2
test: put conv in one reduce (#4441) 
* test: put conv in one reduce

* put reduce at the end

* more expand

* generic, and that expand was breaking things

* ratio

* don't undo the expand

* arg 1

* strides

* warning, for resnet

* warning removed

* disable cast

* handle cast

* op

* err, that's right

* fixup

* fix that

* a test to play with

* add double_reduces

* working up to final reshape

* fold the last reshape

* moved to schedule

* fix axis

* ci, need to bring arange back

* FUSE_CONV_BW maybe

* valid in 3.9

* test_expand_reduce_is_folded_on_different_axes

* add FUSE_CONV_BW=1

* test_fold_batchnorm_backward

* test_sgd_4convs_fuse

---------

Co-authored-by: qazal <qazal.software@gmail.com>
 2024-07-22 12:16:13 +03:00
George Hotz d1a7279605
indexing fold with casted bool (#5551) 
* cast bool is where

* universal transform is wrong
 2024-07-18 10:02:29 -07:00
qazal 0ad1672d5f
fuse indexing (LazyOp creation) (#5506) 
* bring FUSE_AS_ONE_KERNEL back

* operands need reshape?

* fused but arange didnt fold

* something deeply wrong

* yay, fused

* derive broadcasts

* s/input/reduce_input

* _fixup_ones proved a point

* this is what it takes

* down to 3 required reshapes:

1. output_shape
2. the second reduce merge dims
3. remove dims for above reshape

* start real reshapes

* resolve shape in the edges pre lazyop

* outputs are the same shape

* rewrite1: just the reduce

* more correct

* fuse_as_one_kernel

* closer

* this passes

* dont rerun info

* dont need these

* not needed
 2024-07-18 14:09:17 +03:00
qazal e22b377839
generalize FUSE_AS_ONE_KERNEL in the scheduler (#5397) 
* test: use const

* hotfix: base

* asserts

* dont push through reshape

* cleanup

* dont need the cache

* test_reduceop_reshape_dont_push and test_index_fused are next
 2024-07-12 10:23:16 +03:00
George Hotz 3a2b5a75d2
improve single kernel indexing (#5398) 
* improve single kernel indexing

* metadata in graph (#5399)

* indexing is O(1)

* add failing test

* ugh, that all needs to be replaced with symbolic

* broken on ptx, it's fine

---------

Co-authored-by: wozeparrot <wozeparrot@gmail.com>
 2024-07-11 19:00:57 -07:00
George Hotz c2da4454cd
indexing getting better (#5389) 
* indexing getting better [run_process_replay] [no_assert]

* fix test

* test_arange_2_reduce is a simpler test

* put that print back, NOOPT

* don't merge reduces (they could be different reduces)

* FUSE_AS_ONE_KERNEL

* fix tests

* fix test_var_multireduce

* w/e put that there

* fails on others too

* fix test, revert UNMUL change

* in case order matters

* one kernel indexing works

* one kernel indexing works (test other)
 2024-07-11 16:41:51 -07:00
chenyu f1bf916b8a
apply NOOPT in test_arange complexity (#4774) 
with hcopt, arange(2560) uses less ops than arange(256)
 2024-05-29 23:12:35 -04:00
George-the-1st 0627e26140
Added missing unittest execution code (#4400) 
same code as on every other test file, just missing from this one for some reason.
 2024-05-02 22:34:30 -04:00
chenyu a6ed2ae3c6
use old cumsum optimization for arange (#3813) 
revert to old cumsum opt while phi simplification is disabled.

added a flops complexity test for this
 2024-03-18 20:01:03 -04:00