Check cl_khr_fp16 support #225

vorj · 2019-04-16T07:02:54Z

Task of #224 .
ClPy in this PR checks that the device supports #pragma OPENCL EXTENSION cl_khr_fp16: enable or not before first running kernel, and enable or disable cl_khr_fp16 adequately.

Additionally, we also have the problem that test_helper.py blames DeprecationWarning with pytest>=5 . I also fixed this issue.

Remaining tasks:

skip tests which have half output when the device doesn't support cl_khr_fp16

jenkins-maekawa

Test (commit 8f24069) failed in vega.

vorj · 2019-04-16T07:04:50Z

@nsakabe-fixstars Please check this branch. Does ClPy of this branch work on your environment?

jenkins-maekawa

Test (commit 8f24069) failed in titanv.

jenkins-maekawa

Test (commit acc786b) failed in vega.

jenkins-maekawa

Test (commit acc786b) failed in titanv.

jenkins-maekawa

Test (commit 7b0637e) failed in vega.

jenkins-maekawa

Test (commit 3c0be6a) failed in vega.

jenkins-maekawa

Test (commit 7b0637e) passed in titanv.

jenkins-maekawa

Test (commit 7b0637e) passed in titanv.

ghost · 2019-04-16T08:49:10Z

@vorj I confirmed it working on NVIDIA GPU w/o fp16 feature. Thank you.

jenkins-maekawa

Test (commit 695f653) passed in vega.

jenkins-maekawa

Test (commit 3c0be6a) passed in titanv.

vorj · 2019-04-19T16:42:21Z

@nsakabe-fixstars It's time to pass the baton. This PR has a remaining task, so please push this forward. You have!

ghost · 2019-05-23T06:44:54Z

skip tests which have half output

I ran the tests with 1080Ti(w/o fp16). Needed to skip:

Need to skip cases:

test_fusion.py::TestFusionTrigonometric::test_arctan FAILED              [ 15%]
test_fusion.py::TestFusionTrigonometric::test_arctan2 FAILED             [ 16%]
test_fusion.py::TestFusionTrigonometric::test_cos FAILED                 [ 16%]
test_fusion.py::TestFusionTrigonometric::test_deg2rad FAILED             [ 16%]
test_fusion.py::TestFusionTrigonometric::test_hypot FAILED               [ 16%]
test_fusion.py::TestFusionTrigonometric::test_rad2deg FAILED             [ 16%]
test_fusion.py::TestFusionTrigonometric::test_sin FAILED                 [ 16%]
test_fusion.py::TestFusionTrigonometric::test_tan FAILED                 [ 16%]
test_fusion.py::TestFusionHyperbolic::test_arcsinh FAILED                [ 16%]
test_fusion.py::TestFusionHyperbolic::test_cosh FAILED                   [ 17%]
test_fusion.py::TestFusionHyperbolic::test_sinh FAILED                   [ 17%]
test_fusion.py::TestFusionHyperbolic::test_tanh FAILED                   [ 17%]
test_fusion.py::TestFusionRounding::test_ceil FAILED                     [ 17%]
test_fusion.py::TestFusionRounding::test_fix FAILED                      [ 17%]
test_fusion.py::TestFusionRounding::test_floor FAILED                    [ 17%]
test_fusion.py::TestFusionRounding::test_rint FAILED                     [ 17%]
test_fusion.py::TestFusionRounding::test_rint_negative FAILED            [ 17%]
test_fusion.py::TestFusionRounding::test_trunc FAILED                    [ 17%]
test_fusion.py::TestFusionExplog::test_exp FAILED                        [ 17%]
test_fusion.py::TestFusionExplog::test_exp2 FAILED                       [ 18%]
test_fusion.py::TestFusionExplog::test_expm1 FAILED                      [ 18%]
test_fusion.py::TestFusionExplog::test_log FAILED                        [ 18%]
test_fusion.py::TestFusionExplog::test_log10 FAILED                      [ 18%]
test_fusion.py::TestFusionExplog::test_log1p FAILED                      [ 18%]
test_fusion.py::TestFusionExplog::test_log2 FAILED                       [ 18%]
test_fusion.py::TestFusionExplog::test_logaddexp FAILED                  [ 18%]
test_fusion.py::TestFusionExplog::test_logaddexp2 FAILED                 [ 18%]
test_fusion.py::TestFusionFloating::test_copysign FAILED                 [ 18%]
test_fusion.py::TestFusionFloating::test_nextafter FAILED                [ 19%]
test_fusion.py::TestFusionMisc::test_sqrt FAILED                         [ 23%]
test_ndarray_complex_ops.py::TestAngle::test_angle FAILED                [ 64%]
test_arithmetic.py::TestArithmetic::test_angle FAILED                    [  1%]
test_explog.py::TestExplog::test_exp FAILED                              [ 12%]
test_explog.py::TestExplog::test_exp2 FAILED                             [ 12%]
test_explog.py::TestExplog::test_expm1 FAILED                            [ 13%]
test_explog.py::TestExplog::test_log FAILED                              [ 13%]
test_explog.py::TestExplog::test_log10 FAILED                            [ 14%]
test_explog.py::TestExplog::test_log1p FAILED                            [ 14%]
test_explog.py::TestExplog::test_log2 FAILED                             [ 15%]
test_explog.py::TestExplog::test_logaddexp FAILED                        [ 15%]
test_explog.py::TestExplog::test_logaddexp2 FAILED                       [ 16%]
test_floating.py::TestFloating::test_copysign FAILED                     [ 16%]
test_floating.py::TestFloating::test_nextafter FAILED                    [ 18%]
test_hyperbolic.py::TestHyperbolic::test_arcsinh FAILED                  [ 20%]
test_hyperbolic.py::TestHyperbolic::test_cosh FAILED                     [ 21%]
test_hyperbolic.py::TestHyperbolic::test_sinh FAILED                     [ 21%]
test_hyperbolic.py::TestHyperbolic::test_tanh FAILED                     [ 22%]
test_misc.py::TestMisc::test_sqrt FAILED                                 [ 56%]
test_rounding.py::TestRounding::test_ceil FAILED                         [ 57%]
test_rounding.py::TestRounding::test_fix FAILED                          [ 57%]
test_rounding.py::TestRounding::test_floor FAILED                        [ 58%]
test_rounding.py::TestRounding::test_rint FAILED                         [ 59%]
test_rounding.py::TestRounding::test_rint_negative FAILED                [ 59%]
test_rounding.py::TestRounding::test_trunc FAILED                        [ 60%]
test_trigonometric.py::TestTrigonometric::test_arctan FAILED             [ 89%]
test_trigonometric.py::TestTrigonometric::test_arctan2 FAILED            [ 90%]
test_trigonometric.py::TestTrigonometric::test_cos FAILED                [ 90%]
test_trigonometric.py::TestTrigonometric::test_deg2rad FAILED            [ 91%]
test_trigonometric.py::TestTrigonometric::test_hypot FAILED              [ 91%]
test_trigonometric.py::TestTrigonometric::test_rad2deg FAILED            [ 92%]
test_trigonometric.py::TestTrigonometric::test_sin FAILED                [ 92%]
test_trigonometric.py::TestTrigonometric::test_tan FAILED                [ 93%]

vorj · 2020-02-04T06:03:49Z

@LWisteria The CI has passed. Please review this PR.

vorj · 2020-02-04T06:07:18Z

Of course this PR is too ad-hoc, so we should use this as temporary solution and will revert this in the future if this PR will be merged.

LWisteria

Skipping CI if no fp16 support is a bad idea. I won't approve of this PR.

LWisteria · 2020-02-04T06:12:14Z

@vorj Removing CI tests needing fp16 support looks better for me.

vorj · 2020-02-04T06:16:20Z

@LWisteria You mean that we should remove all of codes related to half (temporarily), right?

LWisteria · 2020-02-04T07:15:52Z

@vorj Actually, you don't need to do that. It's just ideally better than this PR, that's why I won't approve.

vorj · 2020-02-04T07:37:52Z

@LWisteria I didn't talk about my task, I just want to understand what you told more correctly.

BTW, for my information, I want to know the usecase of skipping test, because it seems that my opinion about that is wrong.
Do you know the location for some good description about that?

LWisteria · 2020-02-04T10:37:05Z

@vorj As you know, the primary reason for the problem is ClPy's insufficiency of fp16 support.
Therefore, skipping some tests and/or disabling cl_khr_fp16 if the device doesn't support it will solve nothing, but just only pass CI.

Only passing CI must not be your goal.

vorj · 2020-02-04T10:59:33Z

@LWisteria

Therefore, skipping some tests and/or disabling cl_khr_fp16 if the device doesn't support it will solve nothing, but just only pass CI.

This sentence is true, then what the difference between removing the test cases and skipping them? IMHO, removing tests will solve nothing as well as this PR.

Only passing CI must not be your goal.

I haven't said that, I told that we should revert this in the future if we will merge this PR. I have never thought that this PR solves the core issue, but I think that broken CI should be fixed in short order.
As you said at #261 , the CI issue is caused by the environment broken of our secondary machine, but it means that users who install CUDA 10.2 can't use ClPy at thier own environment for now. I think that this should be fixed.

LWisteria · 2020-02-04T11:17:29Z

the difference between removing the test cases and skipping them

I meant that removing the test from CI, not from unit-tests itself, which guarantees functionalities of current and/or future ClPy.
Modifying the tests means that we decided to change our goal, to be compatible with CuPy as much as possible.

We don't have enough discussions to do that. And I won't agree with it now.

I think that this should be fixed.

I agree. However, we must not treat symptoms, remove the bottom causes.

We don't have to hurry so much. Now you waste your costs for the symptomatic treatment.

vorj · 2020-02-04T11:34:39Z

@LWisteria

I meant that removing the test from CI, not from unit-tests itself, which guarantees functionalities of current and/or future ClPy.
Modifying the tests means that we decided to change our goal, to be compatible with CuPy as much as possible.

Ah, OK, I understand. Thank you for your explanation.
BTW, I think that it will be painful task that removing the test cases from CI, because we will need to list so many valid test cases like test_fusion.py::TestFusionFloating::test_signbit , and so on.
(Of course, I don't think that we should push forward this PR just because the substitute task will be painful.)

We don't have to hurry so much.

Hmm, well, any user doesn't make the ticket like #239 recently, so I kind of agree.

vorj · 2020-02-04T11:37:34Z

Now you waste your costs for the symptomatic treatment.

No, I had already wasted (it's just joke).

LWisteria · 2020-02-22T00:33:13Z

After #269, this PR is obsolete. I'm closing but not deleting this branch for the further development.

vorj assigned ghost Apr 16, 2019

jenkins-maekawa suggested changes Apr 16, 2019

View reviewed changes

vorj force-pushed the 224-check-cl_khr_fp16-support branch from b8cada7 to 8e5154b Compare April 16, 2019 07:40

jenkins-maekawa suggested changes Apr 16, 2019

View reviewed changes

vorj force-pushed the 224-check-cl_khr_fp16-support branch from 8e5154b to bbc8af3 Compare April 16, 2019 08:04

jenkins-maekawa suggested changes Apr 16, 2019

View reviewed changes

vorj force-pushed the 224-check-cl_khr_fp16-support branch from bbc8af3 to ddb430d Compare April 16, 2019 08:12

jenkins-maekawa suggested changes Apr 16, 2019

View reviewed changes

jenkins-maekawa approved these changes Apr 16, 2019

View reviewed changes

ghost assigned vorj and unassigned ghost Apr 16, 2019

vorj added 3 commits April 18, 2019 15:59

implement clpy.backend.opencl.env.supports.cl_khr_fp16

84bed8b

define __CLPY_ENABLE_CL_KHR_FP16 when the device supports cl_khr_fp16

5caa98b

disable cl_khr_fp16 when __CLPY_ENABLE_CL_KHR_FP16 is not defined

85fc8bb

vorj force-pushed the 224-check-cl_khr_fp16-support branch from ddb430d to 85fc8bb Compare April 18, 2019 08:15

jenkins-maekawa approved these changes Apr 18, 2019

View reviewed changes

vorj assigned ghost and unassigned vorj Apr 19, 2019

ghost removed their assignment Jun 27, 2019

LWisteria mentioned this pull request Jul 18, 2019

OpenCLがインストールできていなかった / Chainerのバージョン違い / エイリアス指定忘れ / Intel GPUなどfp16サポートしていないデバイス #239

Closed

vorj added 2 commits February 4, 2020 00:44

add no_8bit_integer option to testing.helper.for_dtypes

1af1dea

add skip_when_disabled_cl_khr_fp16

69c3a83

vorj force-pushed the 224-check-cl_khr_fp16-support branch from 2c3075b to bf0304f Compare February 4, 2020 04:18

vorj and others added 3 commits February 4, 2020 14:05

separate 8bit integer test cases

71522b0

add @testing.skip_when_disabled_cl_khr_fp16 to 8bit integer test cases

390cbb0

fix warning error of importing imp

43e8d6f

vorj force-pushed the 224-check-cl_khr_fp16-support branch from bf0304f to 43e8d6f Compare February 4, 2020 05:22

vorj marked this pull request as ready for review February 4, 2020 05:33

Merge branch 'clpy' into 224-check-cl_khr_fp16-support

0cc0349

vorj assigned LWisteria Feb 4, 2020

vorj requested a review from LWisteria February 4, 2020 06:04

LWisteria requested changes Feb 4, 2020

View reviewed changes

LWisteria assigned vorj and unassigned LWisteria Feb 4, 2020

vorj mentioned this pull request Feb 4, 2020

Removing cl_khr_fp16 #264

Closed

LWisteria closed this Feb 22, 2020

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Check cl_khr_fp16 support #225

Check cl_khr_fp16 support #225

vorj commented Apr 16, 2019 •

edited

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

vorj commented Apr 16, 2019

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment

jenkins-maekawa left a comment

ghost commented Apr 16, 2019

jenkins-maekawa left a comment

jenkins-maekawa left a comment

vorj commented Apr 19, 2019 •

edited

Loading

ghost commented May 23, 2019 •

edited by ghost

Loading

vorj commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria left a comment

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020 •

edited

Loading

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020 •

edited

Loading

vorj commented Feb 4, 2020 •

edited

Loading

LWisteria commented Feb 22, 2020

Check cl_khr_fp16 support #225

Check cl_khr_fp16 support #225

Conversation

vorj commented Apr 16, 2019 • edited Loading

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

vorj commented Apr 16, 2019

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

jenkins-maekawa left a comment • edited by vorj Loading

Choose a reason for hiding this comment

jenkins-maekawa left a comment

Choose a reason for hiding this comment

jenkins-maekawa left a comment

Choose a reason for hiding this comment

ghost commented Apr 16, 2019

jenkins-maekawa left a comment

Choose a reason for hiding this comment

jenkins-maekawa left a comment

Choose a reason for hiding this comment

vorj commented Apr 19, 2019 • edited Loading

ghost commented May 23, 2019 • edited by ghost Loading

vorj commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria left a comment

Choose a reason for hiding this comment

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020 • edited Loading

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020

LWisteria commented Feb 4, 2020

vorj commented Feb 4, 2020 • edited Loading

vorj commented Feb 4, 2020 • edited Loading

LWisteria commented Feb 22, 2020

vorj commented Apr 16, 2019 •

edited

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

jenkins-maekawa left a comment •

edited by vorj

Loading

vorj commented Apr 19, 2019 •

edited

Loading

ghost commented May 23, 2019 •

edited by ghost

Loading

vorj commented Feb 4, 2020 •

edited

Loading

vorj commented Feb 4, 2020 •

edited

Loading

vorj commented Feb 4, 2020 •

edited

Loading