Solve error: ambiguous overload for ‘operator!=’ (operand types are ‘__half’ and ‘double’) by jiapei100 · Pull Request #23948 · opencv/opencv

jiapei100 · 2023-07-07T18:10:44Z

**Remove the error: error: ambiguous overload for ‘operator!=’ (operand types are ‘__half’ and ‘double’) **

Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

I agree to contribute to the project under Apache 2 License.
To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
The PR is proposed to the proper branch
There is a reference to the original bug report and related work
There is accuracy test, performance test and test data in opencv_extra repository, if applicable
Patch to opencv_extra has the same branch name.
The feature is well documented and sample code can be built with the project CMake

… types are ‘__half’ and ‘double’)

asmorkalov · 2023-07-11T07:54:43Z

@mshabunin could you take a look?

mshabunin · 2023-07-11T13:45:35Z

@jiapei100 , our CI CUDA builds have failed: https://github.com/opencv/opencv/actions/runs/5489317969/jobs/10008696594?pr=23948

jiapei100 · 2023-07-11T17:27:11Z

@mshabunin
Could be the reasons:

Ubuntu 20.04 vs. Ubuntu 22.04
Cuda 11.7 vs. Cuda 12.2

?????

BTW, are you building everything on your own OpenCV CI ?? Everything automatically??
That is cool...

…‘__half’ and ‘double’)

VadimLevin · 2023-07-12T09:38:19Z

modules/dnn/src/cuda4dnn/primitives/normalize_bbox.hpp

             * or we don't have to scale
             */
-            if (weight != (T)(1.0))
+            if (weight != static_cast(T)(1.0))


Wrong static_cast syntax, please refer to cppreference for more information about how it works.

static_cast<T>(1.0)

mshabunin · 2023-07-12T10:04:14Z

Perhaps original issue is no caused by missing type case, but by ambiguous constant type. Try to change to 1.f and 0.f.

cudawarped · 2023-07-13T07:18:38Z

Perhaps original issue is no caused by missing type case, but by ambiguous constant type. Try to change to 1.f and 0.f.

The static cast fixes the issue with CUDA 12.2 but due to the ambiguous constant type you mentioned causes an issue with earlier versions of CUDA.

The combination of both, i.e.

if (weight != static_cast<T>(1.0f))
if(nms_iou_threshold > static_cast<T>(0.0f))

or double

if (weight != static_cast<T>(1.0))
if(nms_iou_threshold > static_cast<T>(0.0))

works on Windows with Visual Studio 2022 for CUDA 12.1 (old math api) and CUDA 12.2.

Zacrain · 2023-07-14T00:57:56Z

Perhaps original issue is no caused by missing type case, but by ambiguous constant type. Try to change to 1.f and 0.f.

The static cast fixes the issue with CUDA 12.2 but due to the ambiguous constant type you mentioned causes an issue with earlier versions of CUDA.

The combination of both, i.e.
if (weight != static_cast<T>(1.0f))
if(nms_iou_threshold > static_cast<T>(0.0f))
or double
if (weight != static_cast<T>(1.0))
if(nms_iou_threshold > static_cast<T>(0.0))
works on Windows with Visual Studio 2022 for CUDA 12.1 (old math api) and CUDA 12.2.

If I am not mistaken, ISO C++ is clear about default primitive types of constant literals. 1 defaults to int and can be implicitly type-casted when assigning to, e.g,, double, 1.0 defaults to double, 1.0f to float and so on. So there should not be an "ambiguous constant type".
Therefore it shouldn't be necessary to specifically indicate the use of double or float when it is static_cast-ed to the target template type anyway. Maybe there are conversion functions missing for the "primitive" data types added by CUDA?
At this point I wonder whether this might be a problem of nvcc, although I couldn't find any info about that. Maybe that's intentional for some reason?

cudawarped · 2023-07-14T04:35:44Z

If I am not mistaken, ISO C++ is clear about default primitive types of constant literals. 1 defaults to int and can be implicitly type-casted when assigning to, e.g,, double, 1.0 defaults to double, 1.0f to float and so on. So there should not be an "ambiguous constant type".
Therefore it shouldn't be necessary to specifically indicate the use of double or float when it is static_cast-ed to the target template type anyway. Maybe there are conversion functions missing for the "primitive" data types added by CUDA?
At this point I wonder whether this might be a problem of nvcc, although I couldn't find any info about that. Maybe that's intentional for some reason?

I think the use of the word ambiguous has itself become ambiguous!

I may be wrong but my interpretation of the build error

Ubuntu2004-x64-CUDA / BuildAndTest output

@jiapei100 , our CI CUDA builds have failed: https://github.com/opencv/opencv/actions/runs/5489317969/jobs/10008696594?pr=23948

/home/ci/opencv/modules/dnn/src/layers/../cuda4dnn/primitives/region.hpp:124:37: error: call of overloaded '__half(int)' is ambiguous
124 | if (nms_iou_threshold > (T)(0)) {
| ^~~~~~
In file included from /usr/local/cuda/include/cuda_fp16.h:3790,
from /usr/local/cuda/include/cublas_api.h:75,
from /usr/local/cuda/include/cublas_v2.h:65,
from /home/ci/opencv/modules/dnn/src/layers/../cuda4dnn/csl/cublas.hpp:14,
from /home/ci/opencv/modules/dnn/src/layers/../op_cuda.hpp:11,
from /home/ci/opencv/modules/dnn/src/layers/region_layer.cpp:44:
/usr/local/cuda/include/cuda_fp16.hpp:202:25: note: candidate: '__half::__half(double)'
202 | CUDA_HOSTDEVICE __half(const double f) { __x = __double2half(f).__x; }
| ^~~~~~
/usr/local/cuda/include/cuda_fp16.hpp:201:25: note: candidate: '__half::__half(float)'
201 | CUDA_HOSTDEVICE __half(const float f) { __x = __float2half(f).__x; }
| ^~~~~~
/usr/local/cuda/include/cuda_fp16.hpp:179:26: note: candidate: 'constexpr __half::__half(const __half&)'
179 | struct CUDA_ALIGN(2) __half {
| ^~~~~~

is that there is no conversion from int to half in CUDA 12.1 so it ambiguous as to whether it should cast the constant 0 to a float or a double as theses are the only two overloads available.

cudawarped · 2023-07-17T09:26:45Z

@mshabunin
Could be the reasons:
Ubuntu 20.04 vs. Ubuntu 22.04
Cuda 11.7 vs. Cuda 12.2

@jiapei100 The CI uses an older version of CUDA, can you apply the changes suggested by @mshabunin in addition to your changes, e.g.

if (weight != static_cast<T>(1.0f))
if (nms_iou_threshold > static_cast<T>(0.0f))

to this PR to see if it compiles successfully so we can update OpenCV to be compatilble with CUDA 12.2 please.

ibrewster · 2023-07-26T18:45:10Z

For what it's worth, when I tried manually applying the suggested changes to my copy of the source on my Ubuntu 22.04 box with Cuda 12.2, it did resolve the errors being generated by those lines, however I ran into other issues later in compilation with different files (looks like the python bindings). Different issue, so not including the details unless requested...

cudawarped · 2023-07-27T04:28:09Z

@jiapei100 Kindly asking you to fix this PR so that OpenCV can support the latest version of CUDA.

cudawarped · 2023-08-01T10:10:25Z

@VadimLevin, @mshabunin is it possilble for the maintainers to apply the fixes discussed here (also shown here on my forked branch) to see if this fixes the issue and passes all the relavent CI tests.

CUDA 12.2 was released over a month ago with update 1 released yesterday and @jiapei100 seems to have gone awol. It seems a shame to let this fix go stale when 🤞 it should be such a simple fix.

Remove the error: error: ambiguous overload for ‘operator!=’ (operand…

d20bb0c

… types are ‘__half’ and ‘double’)

jiapei100 mentioned this pull request Jul 7, 2023

OpenCV does not support protobuf 3.x #23791

Closed

4 tasks

asmorkalov added the category: build/install label Jul 10, 2023

asmorkalov changed the title ~~Remove the error: error: ambiguous overload for ‘operator!=’ (operand…~~ Solve error: ambiguous overload for ‘operator!=’ (operand types are ‘__half’ and ‘double’) Jul 11, 2023

asmorkalov requested a review from mshabunin July 11, 2023 07:54

asmorkalov added this to the 4.9.0 milestone Jul 11, 2023

Remove error: ambiguous overload for ‘operator!=’ (operand types are …

94f7bf9

…‘__half’ and ‘double’)

VadimLevin reviewed Jul 12, 2023

View reviewed changes

opencv-alalek added the category: gpu/cuda (contrib) OpenCV 4.0+: moved to opencv_contrib label Jul 13, 2023

opencv-alalek mentioned this pull request Jul 13, 2023

CUDA 12.2 fp16 dnn compilation error #23893

Closed

4 tasks

asmorkalov mentioned this pull request Jul 13, 2023

Opencv Build form source #23984

Closed

4 tasks

cudawarped mentioned this pull request Aug 4, 2023

cuda: fix for compatibility with CUDA Toolkit >= 12.2.0 #24104

Merged

6 tasks

asmorkalov closed this Aug 4, 2023

opencv-alalek removed this from the 4.9.0 milestone Aug 4, 2023

cudawarped mentioned this pull request Jan 10, 2024

Cannot build OpenCV 4.9.0 on Ubuntu 22 with Cuda 12.2.1 and CuDNN 8.9.7 #24839

Closed

4 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Solve error: ambiguous overload for ‘operator!=’ (operand types are ‘__half’ and ‘double’)#23948

Solve error: ambiguous overload for ‘operator!=’ (operand types are ‘__half’ and ‘double’)#23948
jiapei100 wants to merge 2 commits intoopencv:4.xfrom
LongerVision:4.x

jiapei100 commented Jul 7, 2023 •

edited by asmorkalov

Loading

Uh oh!

asmorkalov commented Jul 11, 2023

Uh oh!

mshabunin commented Jul 11, 2023

Uh oh!

jiapei100 commented Jul 11, 2023

Uh oh!

VadimLevin Jul 12, 2023

Uh oh!

mshabunin commented Jul 12, 2023 •

edited

Loading

Uh oh!

cudawarped commented Jul 13, 2023

Uh oh!

Zacrain commented Jul 14, 2023 •

edited

Loading

Uh oh!

cudawarped commented Jul 14, 2023

Uh oh!

cudawarped commented Jul 17, 2023

Uh oh!

ibrewster commented Jul 26, 2023

Uh oh!

cudawarped commented Jul 27, 2023

Uh oh!

cudawarped commented Aug 1, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

Uh oh!

Conversation

jiapei100 commented Jul 7, 2023 • edited by asmorkalov Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Readiness Checklist

Uh oh!

asmorkalov commented Jul 11, 2023

Uh oh!

mshabunin commented Jul 11, 2023

Uh oh!

jiapei100 commented Jul 11, 2023

Uh oh!

VadimLevin Jul 12, 2023

Choose a reason for hiding this comment

Uh oh!

mshabunin commented Jul 12, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cudawarped commented Jul 13, 2023

Uh oh!

Zacrain commented Jul 14, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cudawarped commented Jul 14, 2023

Uh oh!

cudawarped commented Jul 17, 2023

Uh oh!

ibrewster commented Jul 26, 2023

Uh oh!

cudawarped commented Jul 27, 2023

Uh oh!

cudawarped commented Aug 1, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

jiapei100 commented Jul 7, 2023 •

edited by asmorkalov

Loading

mshabunin commented Jul 12, 2023 •

edited

Loading

Zacrain commented Jul 14, 2023 •

edited

Loading