qazal
217c006103
buffer access on UOp [pr] ( #7665 )
...
* add .buffer access on uop
* rename to buf_uop
* start smaller
* ptr != buffer!!
2024-11-13 17:04:19 +08:00
qazal
5da149d23c
uop can have base [pr] ( #7666 )
2024-11-13 16:53:49 +08:00
qazal
ca99c67d78
refactors from the delete lazy diff [pr] ( #7664 )
...
* dedup parent shapetrackers [pr]
* arg -> dtype
* move to ops
* arg
2024-11-13 16:23:53 +08:00
chenyu
e6cfaaa496
metal benchmark JIT=2 -> JIT=1 ( #7661 )
2024-11-12 22:55:27 -05:00
chenyu
4c5f7ddf1f
flux set model path in args ( #7660 )
...
in addition to default downloading through fetch, add an arg to pass model path directly
2024-11-12 22:11:40 -05:00
chenyu
08706c2ea4
more readable rand [pr] ( #7659 )
...
no walrus inside walrus
2024-11-12 19:02:27 -05:00
chenyu
1884f021e3
add conv3x3 to speed_v_theoretical ( #7658 )
...
* add conv3x3 to speed_v_theoretical
* show test duration
2024-11-12 16:41:56 -05:00
ignaciosica
54c0abcb2b
cleaner code_for_op order [pr] ( #7653 )
...
* cleaner code_for_op order
* mantain unary-bin-tern order
* might as well reorder for cuda and amd
2024-11-12 15:13:56 -05:00
chenyu
962dafb467
use randn in speed_v_theoretical instead of rand ( #7656 )
...
* use randn in speed_v_theoretical instead of rand
this made green gemv 20% faster... but why?
* update threshold
2024-11-12 15:00:32 -05:00
chenyu
397a2e6eb6
no special case for int32 in truncate [pr] ( #7657 )
...
this masked an issue that idx is not data, and should never need truncate
2024-11-12 14:52:14 -05:00
chenyu
6159790ab8
add gemv to speed_v_theoretical ( #7654 )
...
* add gemv to speed_v_theoretical
getting ~300GB/s if we just count the memory of inputs and output
* better green numbers
* flip
2024-11-12 11:19:35 -05:00
qazal
e07d2d0966
skip TestBeamSearch.test_large_ast ( #7652 )
2024-11-12 20:52:22 +08:00
qazal
0f02573830
save lines in assign tracking [pr] ( #7651 )
2024-11-12 20:49:13 +08:00
qazal
fbad4900bf
move groups to uop [pr] ( #7640 )
...
* override group post chase [pr]
* key reduceop on ubuf
* fix type
2024-11-12 20:09:13 +08:00
George Hotz
4f1f823021
add tiny test for randomness + remove ulong buffers ( #7648 )
...
* add tiny test for randomness
* Tensor._device_seeds is a Tuple
* no tuple, just a 2 element tensor
* no more longs
* fix tests, and maybe ocelot works now
* NV still doesn't work. cleanup rules
* test + two more rules
2024-11-12 12:45:52 +08:00
chenyu
c06a5a9c72
Tensor.linspace raises for dtype.bool ( #7649 )
...
also fixed an assert when passing str dtype to randint
2024-11-11 23:05:14 -05:00
geohotstan
5eef59d732
add Tensor.linspace ( #7609 )
...
* add linspace
* shave off tests and forgot to add to docs crap
* WHOOPS
* better tests
2024-11-12 10:29:36 +08:00
chenyu
99f29e50b2
update speed_v_theoretical numbers ( #7647 )
...
better amd after set compute profile
2024-11-11 20:05:13 -05:00
chenyu
035e39f900
remove copied is_dtype_supported from onnx [pr] ( #7646 )
2024-11-11 19:20:32 -05:00
Ahmed Harmouche
9c63c3d8ab
These casts should only happen if these are supported ( #7644 )
2024-11-12 07:56:50 +08:00
chenyu
a88a15c7e8
setup perflevel in red CI ( #7645 )
...
runs v4.1 bert setup.
```
rocm-smi --setprofile compute
rocm-smi --setmclk 3
rocm-smi --setperflevel high
```
2024-11-11 18:44:55 -05:00
chenyu
773d5b60bf
beam benchmark tests ( #7638 )
...
* beam benchmark tests
* lower AMD number somehow
* less flaky
2024-11-11 18:11:18 -05:00
chenyu
bfab03288d
fix HALF=1 in test_speed_v_torch ( #7642 )
...
* fix HALF=1 in test_speed_v_torch
"operation cache defeats" adds 1 to all arg, which were centered around 0. adding 1 makes big matmul and matvec go inf.
fixed by subtract 1 after and bumpped tolerance for half input
* bigger tol for BIG=2, update CI too
* bigger tol
2024-11-11 14:29:37 -05:00
nimlgen
4d81b7952a
qcom match texture/sampler descriptors to OpenCL ( #7622 )
...
* qcom ioctl compare more regs
* bug fix
2024-11-11 21:56:51 +03:00
qazal
0b66a0d688
only lookup buf_uops in fuse.py [pr] ( #7641 )
2024-11-11 19:14:30 +02:00
qazal
08b9f055f2
don't need outputs in fuse.py [pr] ( #7639 )
2024-11-11 18:35:31 +02:00
George Hotz
b4cb6b89f9
hotfix: CI mac uses python 3.11
2024-11-11 23:42:35 +08:00
George Hotz
9648372ee6
hotfix: mac uses python 3.12
2024-11-11 23:23:48 +08:00
George Hotz
aaa8059aec
python 3.10 is minimum [pr] ( #7636 )
2024-11-11 23:05:50 +08:00
Kinvert
6a0ed46b1c
adding viz to env_vars docs ( #7630 )
2024-11-11 21:28:27 +08:00
George Hotz
d40673505f
new cloud is cloudy [pr] ( #7631 )
...
* new cloud is cloudy [pr]
* waste lines to add security
* safety, with speed and less lines
* timing and del
* lines
* cleanups
* restore CloudSession
* bump to 3.10
* quotes
* renderer security
2024-11-11 20:18:04 +08:00
qazal
766a680588
swizzle parents with graph rewrite ( #7625 )
...
* delete st_fixup
* refactor
* minimal diff
2024-11-11 16:50:38 +08:00
qazal
fec977b966
calling view on graph edges is fine [pr] ( #7632 )
2024-11-11 16:35:18 +08:00
George Hotz
bbc64bf305
x|(x&y) -> x ( #7629 )
...
* x|(x&y) -> x
* fix tests
2024-11-11 10:00:18 +08:00
uuuvn
94a484542b
Hook memoryview via class instead of a function ( #7627 )
2024-11-11 09:07:06 +08:00
qazal
a8da84cce0
recursive swizzle with just graph_rewrite [pr] ( #7626 )
2024-11-10 20:14:21 +02:00
qazal
7275cfb9d8
cleanup swizzle upats [pr] ( #7624 )
2024-11-10 17:05:27 +02:00
qazal
092a441748
test swizzle post permute ( #7623 )
...
* test swizzle post permute
* add st_fixup assert
2024-11-10 16:18:22 +02:00
George Hotz
745316493c
hotfix: add test_simple_conv2d_bias
2024-11-10 18:36:42 +08:00
George Hotz
44c1fd5661
add optional llvm opt [pr] ( #7619 )
2024-11-10 13:26:49 +08:00
George Hotz
0a411b4f68
replace llvm with new llvm ( #7616 )
...
* replace llvm with new llvm
* fix test_linearizer
* minor fixups
* fix alloca
* don't use alloca
* fix DEFINE_ACC
* lines
* comments and lines
* a little tighter
2024-11-10 11:28:52 +08:00
qazal
b61266eb97
late fusion spec for big graph [pr] ( #7613 )
2024-11-09 23:43:11 +08:00
qazal
9d6b03d691
early assert swizzle in kernel [pr] ( #7610 )
...
* early assert swizzle in kernel [pr]
* better
* note changes
* TestIndexing 2
2024-11-09 21:54:43 +08:00
chenyu
8ca422e21a
script to compare kernel opt with BEAM ( #7604 )
...
intersting that on m1 max hcopt wins BEAM 2 about 20% of the time
2024-11-08 17:40:28 -05:00
chenyu
573f145dcf
METAL raise RuntimeError with no compiler and bad src ( #7603 )
...
fixed BEAM if src is invalid on METAL. it currently only accept RuntimeError in `_time_program`
2024-11-08 17:09:12 -05:00
chenyu
74b4d1c1e1
rewrite idx again in real_strides after uop_given_valid ( #7600 )
...
uop_given_valid does not guarantee output to be flat. fixed one last real_strides test.
2024-11-08 14:30:32 -05:00
chenyu
c6189e38c1
simplify_valid in real_strides ( #7599 )
...
improved one more real_strides. after finishing the last one will think about always applying these in to_indexed_uops
2024-11-08 10:45:22 -05:00
George Hotz
d8691a4f03
lil touchups ( #7597 )
2024-11-08 22:31:43 +08:00
Ahmed Harmouche
e35226e698
Remove Ops.ALU ( #7595 )
2024-11-08 19:52:14 +08:00
Harald Schäfer
e7cbc29f48
openpilot benchmark: add cast from numpy to benchmark ( #7593 )
...
* openpilot benchmark: add cast from numpy to benchmark
* whitespace
* comment
2024-11-08 19:31:00 +08:00
