barry-jin opened a new pull request #20087:
URL: https://github.com/apache/incubator-mxnet/pull/20087
## Description ##
Adopt packedfunc based ffi on some frequently used numpy_extension
operators. Some benchmarks are as follows:
npx.softmax and npx.log_softmax
```python
setup = """
from mxnet import np, npx
npx.set_np()
a = np.ones((2, 2))
"""
stmt = """
npx.softmax(a)
"""
legacy
7.195554300000006e-05
New FFI
5.542412399999996e-05
setup = """
from mxnet import np, npx
npx.set_np()
a = np.array([1, 2, .1])
"""
stmt = """
npx.log_softmax(a)
"""
legacy
7.109741400000047e-05
New FFI
5.514604400000067e-05
```
npx.activation
```python
setup = """
from mxnet import np, npx, img
from mxnet.ndarray.numpy import _internal as _npi
from mxnet.ndarray.numpy import _api_internal
npx.set_np()
x = np.arange(8).reshape((2, 2, 2))
"""
stmt = """
o = npx.activation(x, act_type='sigmoid')
"""
timer = timeit.Timer(setup=setup,
stmt=stmt)
num_repeat = 1000
print(min(timer.repeat(repeat=10, number=num_repeat)) / num_repeat)
legacy
0.00010351458600000019
New FFI
4.998078599999989e-05
```
npx.batch_norm
```python
setup = """
from mxnet import np, npx, img
from mxnet.ndarray.numpy import _internal as _npi
from mxnet.ndarray.numpy import _api_internal
npx.set_np()
a = np.zeros((2,2))
b = np.ones((2,2))
c = np.ones((2,2))
d = np.ones((2,2))
e = np.ones((2,2))
"""
stmt = """
o = npx.batch_norm(a, b, c, d, e,
eps=1e-3,
momentum=0.9,
fix_gamma=True,
use_global_stats=False,
output_mean_var=False,
axis=1,
cudnn_off=False,
min_calib_range=None,
max_calib_range=None)
"""
timer = timeit.Timer(setup=setup,
stmt=stmt)
num_repeat = 1000
print(min(timer.repeat(repeat=10, number=num_repeat)) / num_repeat)
legacy
0.00017782039200000012
New FFI
8.22070250000002e-05
```
npx.fully_connected
```python
setup = """
from mxnet import np, npx, img
from mxnet.ndarray.numpy import _internal as _npi
from mxnet.ndarray.numpy import _api_internal
npx.set_np()
data = np.arange(8).reshape((4,2))
weight = np.arange(4).reshape((2,2))
bias = np.arange(2).reshape((2,))
num_hidden = bias.shape[0]
"""
stmt = """
o = npx.fully_connected(data, weight, bias, num_hidden=num_hidden,
no_bias=False)
"""
timer = timeit.Timer(setup=setup,
stmt=stmt)
num_repeat = 1000
print(min(timer.repeat(repeat=10, number=num_repeat)) / num_repeat)
legacy
9.581770600000006e-05
New FFI
6.12347289999997e-05
```
## Checklist ##
### Essentials ###
- [x] PR's title starts with a category (e.g. [BUGFIX], [MODEL], [TUTORIAL],
[FEATURE], [DOC], etc)
- [x] Changes are complete (i.e. I finished coding on this PR)
- [x] All changes have test coverage
- [x] Code is well-documented
### Changes ###
- [ ] Feature1, tests, (and when applicable, API doc)
- [ ] Feature2, tests, (and when applicable, API doc)
## Comments ##
- If this change is a backward incompatible change, why must this change be
made.
- Interesting edge cases to note here
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
