Use cuTENSOR in cupy.prod, cupy.max, cupy.min, cupy.ptp and cupy.mean
#3765
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Nice! I was wondering when this PR would show up! 😁 Have you done any benchmarks? |
kmaehashi
added
the
to-be-backported
kmaehashi
added
cat:performance
|
pfnCI, test this please. |
|
Jenkins, test this please |
|
Jenkins CI test (for commit be9bd9c, target branch master) failed with status FAILURE. |
kmaehashi
requested changes
Sep 28, 2020
|
Jenkins, test this please. |
|
Jenkins CI test (for commit 5285907, target branch master) failed with status FAILURE. |
|
@asi1024 Could you check test failures? |
Co-authored-by: Kenichi Maehashi <webmaster@kenichimaehashi.com>
kmaehashi
added
the
blocking
asi1024
force-pushed
the
use-cutensor
branch
from
5285907
to
2ccad4d
Compare
asi1024
changed the title
cupy.prod, cupy.max, cupy.min, cupy.ptp and cupy.meancupy.prod, cupy.max, cupy.min, cupy.ptp and cupy.mean
|
Rebased. |
|
pfnCI, test this please. |
asi1024
changed the title
cupy.prod, cupy.max, cupy.min, cupy.ptp and cupy.meancupy.prod, cupy.max, cupy.min, cupy.ptp and cupy.mean
|
pfnCI, test this please. |
|
Jenkins CI test (for commit a45dae1, target branch master) failed with status FAILURE. |
|
pfnCI, test this please. |
|
Jenkins CI test (for commit a45dae1, target branch master) failed with status FAILURE. |
kmaehashi
reviewed
Sep 29, 2020
cupy/core/_routines_statistics.pyx
Outdated
| @@ -7,6 +7,7 @@ import cupy | |||
| from cupy.core import _reduction | |||
| from cupy.core._reduction import create_reduction_func | |||
| from cupy.core._reduction import ReductionKernel | |||
| from cupy import cutensor | |||
There was a problem hiding this comment.
This line needs to be removed.
|
pfnCI, test this please. |
|
Jenkins CI test (for commit 28ba922, target branch master) failed with status FAILURE. |
|
Jenkins, test this please. |
|
Jenkins CI test (for commit 28ba922, target branch master) failed with status FAILURE. |
|
pfnCI, test this please. |
|
Jenkins CI test (for commit 28ba922, target branch master) succeeded! |
|
LGTM! |
kmaehashi
added a commit
to kmaehashi/cupy
that referenced
this pull request
Sep 30, 2020
Use cuTENSOR in `cupy.prod`, `cupy.max`, `cupy.min`, `cupy.ptp` and `cupy.mean`
|
Looks like each backend has its own strength and there's no clear dominator. import cupy as cp
import numpy as np
from cupyx.time import repeat
##cp.show_config()
#CUB_device_targets = ['sum', 'prod', 'min', 'max', 'argmax', 'argmin', 'cumsum', 'cumprod', 'mean']
#full_targets = CUB_device_targets + \
# ['amin', 'amax', 'nanmin', 'nanmax', 'nanargmin', 'nanargmax',
# 'nanmean',
# 'var', 'nanvar', 'nansum', 'nanprod',
# 'all', 'any', 'count_nonzero']
CUB_device_targets = ['sum', 'prod', 'min', 'max', 'ptp', 'mean']
full_targets = CUB_device_targets
dtypes = [cp.float32, cp.float64, cp.complex64, cp.complex128]
shape = (512, 512, 512)
axes = ((2,), (1, 2), (0, 1, 2))
for dtype in dtypes:
a = cp.random.random(shape)
if dtype in (cp.complex64, cp.complex128):
a = a + 1j*cp.random.random(shape)
a = a.astype(dtype)
for target in full_targets:
for axis in axes:
if target in ('argmax', 'argmin', 'nanargmax', 'nanargmin', 'cumsum', 'cumprod'):
if len(axis) != a.ndim:
continue
else:
axis = None # NumPy limitation
if dtype in (cp.complex64, cp.complex128) and target in ('nanmin', 'nanmax', 'var', 'nanvar'):
continue
print(f"testing {target} with dtype={dtype} and axes={axis}...")
func_cp = getattr(cp, target)
func_np = getattr(np, target)
cp.core.set_routine_accelerators([])
cp.core.set_reduction_accelerators([])
print(repeat(func_cp, (a, axis), n_repeat=20, name=f'no acce, {target}'))
if not cp.allclose(func_cp(a, axis), func_np(cp.asnumpy(a), axis)):
print(f"WARNING: CuPy's kernel might have a problem with {target} and {dtype}")
if target in CUB_device_targets:
cp.core.set_routine_accelerators(['cub'])
cp.core.set_reduction_accelerators([])
print(repeat(func_cp, (a, axis), n_repeat=20, name=f'CUB device, {target}'))
if not cp.allclose(func_cp(a, axis), func_np(cp.asnumpy(a), axis)):
print(f"WARNING: CUB device might have a problem with {target} and {dtype}")
cp.core.set_routine_accelerators([])
cp.core.set_reduction_accelerators(['cub'])
print(repeat(func_cp, (a, axis), n_repeat=20, name=f'CUB block, {target}'))
if not cp.allclose(func_cp(a, axis), func_np(cp.asnumpy(a), axis)):
print(f"WARNING: CUB block might have a problem with {target} and {dtype}")
cp.core.set_routine_accelerators(['cutensor'])
cp.core.set_reduction_accelerators([])
print(repeat(func_cp, (a, axis), n_repeat=20, name=f'cuTENSOR, {target}'))
if not cp.allclose(func_cp(a, axis), func_np(cp.asnumpy(a), axis)):
print(f"WARNING: cuTENSOR might have a problem with {target} and {dtype}")Output (nn the master branch + GTX 2080 Ti + CUDA 10.2): |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Blocked by #3700 and #3732.