Skip to content

Commit 11d4893

Browse files
charlesmartin14charlesmartin14
committed
added Mac M1/M2 swtich to numpy
1 parent 44ce483 commit 11d4893

9 files changed

+343
-18
lines changed

‎checks/check_mac_M12.py

+24
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,24 @@
1+
2+
import platform
3+
import warnings
4+
5+
use_linalg = False
6+
mac_arch = platform.machine()
7+
if mac_arch == 'arm64':
8+
9+
with warnings.catch_warnings():
10+
Warningskill.filterwarnings("ignore",category=DeprecationWarning)
11+
import numpy.distutils.system_info as sysinfo
12+
info = sysinfo.get_info('accelerate')
13+
if info is not None and len(info)>0:
14+
for x in info['extra_link_args']:
15+
if 'Accelerate' in x:
16+
use_linalg = True
17+
18+
print(f"Accelerate found? , use linalg={use_linalg}")
19+
20+
21+
22+
23+
24+

‎checks/check_scipy_svd.py

+27
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,27 @@
1+
import numpy as np
2+
import scipy.linalg as la
3+
import time
4+
5+
# Generate random matrix with 1000 rows and 500 columns
6+
A = np.random.rand(1000, 500)
7+
8+
# Test numpy.linalg.svd
9+
start_time = time.time()
10+
U, S, VT = np.linalg.svd(A, full_matrices=False)
11+
end_time = time.time()
12+
np_time = end_time - start_time
13+
print(f"np.linalg.svd took {np_time:.5f} seconds")
14+
15+
# Test scipy.linalg.svd
16+
start_time = time.time()
17+
U, S, VT = la.svd(A, full_matrices=False)
18+
end_time = time.time()
19+
scipy_time = end_time - start_time
20+
print(f"scipy.linalg.svd took {scipy_time:.5f} seconds")
21+
22+
# Compare the execution times
23+
if np_time < scipy_time:
24+
print("numpy.linalg.svd is faster")
25+
else:
26+
print("scipy.linalg.svd is faster")
27+

‎checks/check_svd_perf.py

+17
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
# https://developer.apple.com/forums/thread/695963
2+
3+
import time
4+
import numpy as np
5+
np.random.seed(42)
6+
a = np.random.uniform(size=(300, 300))
7+
runtimes = 10
8+
9+
timecosts = []
10+
for _ in range(runtimes):
11+
s_time = time.time()
12+
for i in range(10):
13+
a += 1
14+
np.linalg.svd(a)
15+
timecosts.append(time.time() - s_time)
16+
17+
print(f'mean of {runtimes} runs: {np.mean(timecosts):.5f}s')

‎setup.mac

+146
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,146 @@
1+
# Macbook Pro M2
2+
#
3+
4+
5+
# Make sure conda has the channels
6+
# channels:
7+
# - apple
8+
# - conda-forge
9+
#
10+
conda config --add channels apple
11+
conda config --add channels conda-forge
12+
#
13+
# check the channels
14+
# (I think apple needs to be first..not sure yet)
15+
#
16+
conda config --get
17+
18+
#
19+
# install a python ARM version
20+
#
21+
CONDA_SUBDIR=osx-arm64 conda create -n ww0.7 python=3.10
22+
23+
#
24+
# check the platform: 'macOS-13.2.1-arm64-arm-64bit'
25+
#
26+
pip3 install ipython
27+
ipython
28+
import platform
29+
platform.platform()
30+
31+
#
32+
# Veclib Perform floating-point arithmetic, transcendental, and trigonometric functions on 128-bit vectors.
33+
# I guess I need this ? not sure
34+
#
35+
brew install veclibfort
36+
37+
38+
#
39+
# install numpy from the apple channel
40+
# this uses libblas, etc
41+
#
42+
# https://developer.apple.com/forums/thread/695963
43+
#
44+
conda install cython pybind11
45+
pip3 install --no-binary :all: --no-use-pep517 numpy
46+
47+
48+
#
49+
# If NumPy is using the Accelerate framework, you should see a line in the output that says accelerate_info: {'libraries': ['System', 'Accelerate'], 'library_dirs': ['/usr/lib']}.
50+
#
51+
import numpy as np
52+
np.__config__.show()
53+
#
54+
# you should see
55+
#
56+
# lapack_opt_info:
57+
# extra_compile_args = ['-I/System/Library/Frameworks/vecLib.framework/Headers']
58+
# extra_link_args = ['-Wl,-framework', '-Wl,Accelerate']
59+
# define_macros = [('NO_ATLAS_INFO', 3), ('HAVE_CBLAS', None)]
60+
# Supported SIMD extensions in this NumPy install:
61+
# baseline = NEON,NEON_FP16,NEON_VFPV4,ASIMD
62+
# found = ASIMDHP,ASIMDDP
63+
# not found = ASIMDFHM
64+
#
65+
66+
#
67+
# test the performance of the SVD
68+
# average speed should be
69+
#
70+
# mean of 10 runs: 0.08435s
71+
#
72+
python checks/check_svd_perf.py
73+
74+
# install the tensorflow dependencies
75+
# note: numpy must match
76+
#
77+
# this will not work
78+
#conda install -c apple tensorflow-deps
79+
#pip3 install tensorflow-macos
80+
#pip3 install tensorflow-metal
81+
82+
# instead,use
83+
SYSTEM_VERSION_COMPAT=0 pip3 install tensorflow-macos tensorflow-metal
84+
85+
#
86+
# test that tensorflow can be imported and the GPU and CPU are present
87+
#
88+
# [PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'),
89+
# PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]
90+
#
91+
ipython
92+
import tensorflow as tf
93+
tf.__version__ (currently '2.11.0')
94+
tf.config.list_physical_devices()
95+
96+
#
97+
# install pytorch
98+
# https://towardsdatascience.com/installing-pytorch-on-apple-m1-chip-with-gpu-acceleration-3351dc44d67c
99+
#
100+
pip3 install torch torchvision torchaudio
101+
102+
# check that the MPS beckend is available and built
103+
ipython
104+
105+
import torch
106+
torch.backends.mps.is_available()
107+
torch.backends.mps.is_built()
108+
109+
110+
#
111+
# install scipy (can not be accelerated)
112+
#
113+
# scipy needs OPENBLAS, it can not use the Accelerate library
114+
# Note: we can not both install pythran and tensorflow, so we can not build scipy from scratch
115+
116+
brew install openblas gfortran
117+
OPENBLAS="$(brew --prefix openblas)" pip3 install scipy
118+
119+
#
120+
# test the performance of the scipy SVD
121+
#
122+
python checks/check_scipy_svd.py
123+
#
124+
# currently I find that np.linalg.svd is faster:
125+
#
126+
# python checks/checkscipy_svd.py
127+
# np.linalg.svd took 0.05757 seconds
128+
# scipy.linalg.svd took 0.18015 seconds
129+
# numpy.linalg.svd is faster
130+
#
131+
132+
# confirm the linalg is found so the code can switch on this
133+
python checks/check_mac_M12.py
134+
135+
#
136+
# install remaining packages
137+
#
138+
export OPENBLAS=$(/opt/homebrew/bin/brew --prefix openblas)
139+
export CFLAGS="-falign-functions=8 ${CFLAGS}"
140+
pip3 install scikit-learn pandas powerlaw
141+
142+
# for testing and notebooks
143+
pip3 install matplotlib matplotlib-inline tqdm transformers
144+
#
145+
146+

‎tests/test.py

+49-7
Original file line numberDiff line numberDiff line change
@@ -2970,7 +2970,7 @@ def test_conv2d_fft(self):
29702970
details = self.watcher.analyze(layers=[self.first_layer], conv2d_fft=True)
29712971
actual = details.alpha.to_numpy()[0]
29722972
expected = 2.144
2973-
self.assertAlmostEqual(actual,expected, places=4)
2973+
self.assertAlmostEqual(actual,expected, places=3)
29742974

29752975

29762976
class Test_VGG11_StateDict_Alpha_w_WWFit(Test_VGG11_Alpha_w_WWFit):
@@ -3174,6 +3174,30 @@ def setUp(self):
31743174
"""
31753175
print("\n-------------------------------------\nIn Test_RMT_Util:", self._testMethodName)
31763176

3177+
3178+
def test_has_mac_accelerate(self):
3179+
"""Only useful if on a mac M1/M2
3180+
3181+
Note: this test uses a deprecated method that needs to eventually be replaced
3182+
"""
3183+
3184+
expected_has_accelerate = False
3185+
3186+
import platform
3187+
import numpy.distutils.system_info as sysinfo
3188+
3189+
mac_arch = platform.machine()
3190+
if mac_arch == 'arm64':
3191+
info = sysinfo.get_info('accelerate')
3192+
if info is not None and len(info)>0:
3193+
for x in info['extra_link_args']:
3194+
if 'Accelerate' in x:
3195+
expected_has_accelerate = True
3196+
3197+
actual_has_accelerate = RMT_Util.has_mac_accelerate()
3198+
self.assertEqual(expected_has_accelerate, actual_has_accelerate)
3199+
return
3200+
31773201

31783202
def test_vector_entropy(self):
31793203
u = np.array([1,1,1,1])
@@ -3363,6 +3387,8 @@ def get_weights_and_biases_from_Onnx(self):
33633387
class Test_PyTorchSVD(Test_Base):
33643388
"""
33653389
Tests for discrepancies between the scipy and torch implementations of SVD.
3390+
3391+
This has to be modified to support MPS also
33663392
"""
33673393

33683394
def setUp(self):
@@ -3378,7 +3404,7 @@ def setUp(self):
33783404

33793405
def test_torch_svd(self):
33803406
if RMT_Util._svd_full_fast is RMT_Util._svd_full_accurate:
3381-
print("Warning: Unable to test PyTorch SVD method because torch / CUDA not available")
3407+
print("Warning: Unable to test PyTorch SVD method because torch / CUDA or MPS not available")
33823408
return
33833409

33843410
model = models.vgg11(weights='VGG11_Weights.IMAGENET1K_V1')
@@ -3411,21 +3437,38 @@ def test_bad_methods(self):
34113437
with self.assertRaises(AssertionError, msg="svd_vals accepted a bad method"):
34123438
RMT_Util.svd_vals(W, "BAD_METHOD")
34133439

3440+
34143441
def torch_available(self):
3442+
available = False
34153443
try:
34163444
from torch.cuda import is_available
3417-
return is_available()
3445+
print(f"torch cuda available ? {is_available()}")
3446+
if is_available():
3447+
return True
34183448
except ImportError:
3419-
return False
3449+
available = False
3450+
#
3451+
# Torch is available, but CUDA is NOT
3452+
# MPS (MAc M1/2) does not yet support SVD or EIG
3453+
# try:
3454+
# from torch.backends.mps import is_built
3455+
# print(f"torch mps built ? {is_built()}")
3456+
# if is_built():
3457+
# return True
3458+
# except ImportError:
3459+
# pass
3460+
3461+
return available
3462+
34203463

34213464
def test_torch_availability(self):
34223465
if self.torch_available():
3423-
print("torch is available and cuda is available")
3466+
print("torch is available and cuda or mps is available")
34243467
self.assertFalse(RMT_Util._eig_full_accurate is RMT_Util._eig_full_fast)
34253468
self.assertFalse(RMT_Util._svd_full_accurate is RMT_Util._svd_full_fast)
34263469
self.assertFalse(RMT_Util._svd_vals_accurate is RMT_Util._svd_vals_fast)
34273470
else:
3428-
print("torch is not available or cuda is not available")
3471+
print("torch is not available or cuda or mps is not available")
34293472
self.assertTrue(RMT_Util._eig_full_accurate is RMT_Util._eig_full_fast)
34303473
self.assertTrue(RMT_Util._svd_full_accurate is RMT_Util._svd_full_fast)
34313474
self.assertTrue(RMT_Util._svd_vals_accurate is RMT_Util._svd_vals_fast)
@@ -3668,6 +3711,5 @@ def test_column_names_analyze_mp_fit(self):
36683711

36693712

36703713

3671-
36723714
if __name__ == '__main__':
36733715
unittest.main()

‎tests/test_tf.py

+11
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,11 @@
1+
import tensorflow as tf
2+
3+
tf.compat.v1.disable_eager_execution()
4+
5+
a = tf.constant(2)
6+
b = tf.constant(3)
7+
c = tf.add(a, b)
8+
9+
with tf.compat.v1.Session() as sess:
10+
print(sess.run(c))
11+

‎tests/tt.py

+17
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,17 @@
1+
# https://developer.apple.com/forums/thread/695963
2+
3+
import time
4+
import numpy as np
5+
np.random.seed(42)
6+
a = np.random.uniform(size=(300, 300))
7+
runtimes = 10
8+
9+
timecosts = []
10+
for _ in range(runtimes):
11+
s_time = time.time()
12+
for i in range(10):
13+
a += 1
14+
np.linalg.svd(a)
15+
timecosts.append(time.time() - s_time)
16+
17+
print(f'mean of {runtimes} runs: {np.mean(timecosts):.5f}s')

0 commit comments

Comments
 (0)