core: Rename `cv::float16_t` to `cv::hfloat` #25217

fengyuentau · 2024-03-15T09:35:55Z

Merge with opencv/opencv_contrib#3697

Partially resolves #25210

Checklist:

Renam cv::float16_t to cv::fp16_t.
Add typedef fp16_t float16_t for backward compatibility.
Remove class fp16_t methods except contructor and operator float.

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

fengyuentau · 2024-03-15T09:36:27Z

modules/core/include/opencv2/core/cvdef.h

 #endif
 };

+typedef fp16_t float16_t;


I believe we need to drop this line when porting to 5.x.

vpisarev · 2024-03-15T15:46:31Z

@fengyuentau, it looks good. I have been thinking about those names today, and actually since those types are becoming the most important data types nowadays, maybe it would be nice to have shorter or at least more convenient to type names, e.g. 'hfloat' and 'bfloat' (same length as fp16_t and bf16_t, just easier to type). What do you think, @mshabunin, @opencv-alalek, @asmorkalov?

fengyuentau · 2024-03-15T15:55:32Z

e.g. 'hfloat' and 'bfloat'

Advantage is they look closer to float. I am okay to any option that does not conflict.

By the way, @vpisarev , the fromBits function is used in the following line

opencv/modules/core/src/persistence.cpp

Line 1812 in 625eeba

fval = (float) float16_t::fromBits(base64decoder.getUInt16());

I dont have a good idea how to workaround it if we are still going to remove fromBits from the data type class.

vpisarev · 2024-03-15T16:13:31Z

@fengyuentau, float16_t::fromBits should be replaced with external function hfloatFromBits(ushort x) or something like that.

fengyuentau · 2024-03-19T03:55:58Z

This PR is mostly done. Legacy CI showed there is a error check
https://pullrequest.opencv.org/buildbot/builders/precommit_linux64/builds/106635, but it should be fine since it's expected.

vpisarev · 2024-03-20T06:12:33Z

I'm going to merge this PR. @opencv-alalek, @asmorkalov, any objections?

opencv-alalek · 2024-04-02T11:59:50Z

modules/core/include/opencv2/core/cvdef.h

+#endif
+}
+
+typedef hfloat float16_t;


That should be available for EXTERNAL users only.
We should not use that in OpenCV code anywhere else.

Because these hits still lead to conflicts with future C++ compilers:

$ grep -Rn float16_t ./ ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:497: typedef __fp16 float16_t; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:498: const float16_t* a = (const float16_t*)_a; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:499: const float16_t* b = (const float16_t*)_b; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:500: float16_t* c = (float16_t*)_c; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:641: typedef __fp16 float16_t; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:643: const float16_t* a = (const float16_t*)_a; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:644: const float16_t* b = (const float16_t*)_b; ./modules/dnn/src/layers/cpu_kernels/conv_block.simd.hpp:646: const float16_t bias = (float16_t)_bias; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.cpp:88: esz = sizeof(float16_t); ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:438: typedef __fp16 float16_t; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:439: const float16_t* inwptr = (const float16_t*)_inwptr; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:440: const float16_t* wptr = (const float16_t*)_wptr; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:441: float16_t* outbuf = (float16_t*)_outbuf; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:594: typedef __fp16 float16_t; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:595: float16_t* outptr = (float16_t*)_outptr; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:760: typedef __fp16 float16_t; ./modules/dnn/src/layers/cpu_kernels/conv_winograd_f63.simd.hpp:761: const float16_t* inptr = (const float16_t*)_inptr; ./modules/dnn/src/layers/cpu_kernels/convolution.hpp:65: std::vector<float16_t> weightsBuf_FP16; ./modules/dnn/src/layers/cpu_kernels/convolution.hpp:66: std::vector<float16_t> weightsWinoBuf_FP16; ./modules/dnn/src/layers/cpu_kernels/convolution.hpp:67: float16_t* getWeightsFP16(); ./modules/dnn/src/layers/cpu_kernels/convolution.hpp:68: float16_t* getWeightsWinoFP16(); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:29:static inline void _cvt32f16f(const float* src, float16_t* dst, int len) ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:63: dst[j] = float16_t(src[j]); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:77:float16_t* FastConv::getWeightsFP16() ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:82:float16_t* FastConv::getWeightsWinoFP16() ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:274: float16_t* wptrWino_FP16 = nullptr; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:330: float16_t* wptr = wptrWino_FP16 + (g*Kg_nblocks + ki) * Cg *CONV_WINO_KBLOCK*CONV_WINO_AREA + ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:338: wptr[j] = (float16_t)kernelTm[i * CONV_WINO_ATOM_F16 + j]; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:374: float16_t* weightsPtr_FP16 = nullptr; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:403: float16_t* packed_wptr = weightsPtr_FP16 + DkHkWkCg * (startK + g * Kg_aligned_FP16); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:414: packed_wptr[k] = (float16_t)(*wptr); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:416: packed_wptr[k] = (float16_t)0.f; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:476: float16_t* inpbufC_FP16 = (float16_t *)inpbufC; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:477: if (esz == sizeof(float16_t)) ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:574: float16_t* inpbufC_FP16 = (float16_t *)inpbufC; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:575: if (esz == sizeof(float16_t)) ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:582: inpbufC_FP16[k*CONV_NR_FP16] = (float16_t)v0; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:583: inpbufC_FP16[k*CONV_NR_FP16+1] = (float16_t)v1; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:639: _cvt32f16f(inptr, (float16_t *)inpbuf, CONV_NR); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:653: _cvt32f16f(inptr, (float16_t *)inpbuf, slice_len); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:715: float16_t* inpbufC = (float16_t *)inpbuf + s0; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:719: inpbufC[w*CONV_NR] = (float16_t)inptrInC[imgofs]; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:776: float16_t* inpbufC = (float16_t *)inpbuf + s0; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:783: inpbufC[(h*Wk + w)*CONV_NR] = (float16_t)inptrInC[imgofs]; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:849: float16_t* inpbufC = (float16_t* )inpbuf + s0; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:858: inpbufC[((d*Hk + h)*Wk + w)*CONV_NR] = (float16_t)inptrInC[imgofs]; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:900: float16_t * inpbuf_ki_FP16 = (float16_t *)inpbuf + k * CONV_NR * Cg + i; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:1064: inpbuf_ki_FP16[0] = (float16_t)(*inptr_ki); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:1080: inpbuf_ki_FP16[0] = (float16_t)0.f; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:1268: esz = sizeof(float16_t); ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:1522: float16_t* cptr_f16 = (float16_t*)cbuf_task + stripe*CONV_NR; ./modules/dnn/src/layers/cpu_kernels/convolution.cpp:1558: const float16_t *cptr_fp16 = (const float16_t *)cbuf_task; ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1637: AutoBuffer<float16_t, 16> aligned_val; ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1640: float16_t* bufPtr = aligned_val.data(); ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1642: float16_t *fp16Ptr = (float16_t *)field.data(); ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1654: AutoBuffer<float16_t, 16> aligned_val; ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1655: if (!isAligned<sizeof(float16_t)>(val)) ./modules/dnn/src/onnx/onnx_graph_simplifier.cpp:1658: aligned_val.allocate(divUp(sz, sizeof(float16_t)));

float16_t used in convolution related code should be valid, which is exactly float16_t from arm_neon.h. I am not sure about the float16_t in onnx_graph_simplifier.cpp though.

This may conflict with compilers too: typedef __fp16 float16_t;

Need to use #if !defined(__OPENCV_BUILD) && !defined(OPENCV_HIDE_FLOAT16_T) compilation guard.

Rename remaining float16_t for future proof #25387 Resolves comment: #25217 (comment). `std::float16_t` and `std::bfloat16_t` are introduced since c++23: https://en.cppreference.com/w/cpp/types/floating-point. ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] 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 - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

* rename cv::float16_t to cv::fp16_t * add typedef fp16_t float16_t * remove zero(), bits() from fp16_t class * fp16_t -> hfloat * remove cv::float16_t::fromBits; add hfloatFromBits * undo changes in conv_winograd_f63.simd.hpp and conv_block.simd.hpp * undo some changes in dnn

…enaming Rename remaining float16_t for future proof opencv#25387 Resolves comment: opencv#25217 (comment). `std::float16_t` and `std::bfloat16_t` are introduced since c++23: https://en.cppreference.com/w/cpp/types/floating-point. ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] 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 - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

* rename cv::float16_t to cv::fp16_t * add typedef fp16_t float16_t * remove zero(), bits() from fp16_t class * fp16_t -> hfloat * remove cv::float16_t::fromBits; add hfloatFromBits * undo changes in conv_winograd_f63.simd.hpp and conv_block.simd.hpp * undo some changes in dnn

…enaming Rename remaining float16_t for future proof opencv#25387 Resolves comment: opencv#25217 (comment). `std::float16_t` and `std::bfloat16_t` are introduced since c++23: https://en.cppreference.com/w/cpp/types/floating-point. ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] 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 - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

* rename cv::float16_t to cv::fp16_t * add typedef fp16_t float16_t * remove zero(), bits() from fp16_t class * fp16_t -> hfloat * remove cv::float16_t::fromBits; add hfloatFromBits * undo changes in conv_winograd_f63.simd.hpp and conv_block.simd.hpp * undo some changes in dnn

…enaming Rename remaining float16_t for future proof opencv#25387 Resolves comment: opencv#25217 (comment). `std::float16_t` and `std::bfloat16_t` are introduced since c++23: https://en.cppreference.com/w/cpp/types/floating-point. ### Pull Request Readiness Checklist See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request - [x] I agree to contribute to the project under Apache 2 License. - [x] 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 - [x] The PR is proposed to the proper branch - [x] There is a reference to the original bug report and related work - [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [x] The feature is well documented and sample code can be built with the project CMake

fengyuentau added 2 commits March 15, 2024 17:22

rename cv::float16_t to cv::fp16_t

26adc67

add typedef fp16_t float16_t

ec4fbc1

fengyuentau added category: core port to 5.x is needed Label for maintainers. Authors of PR can ignore this labels Mar 15, 2024

fengyuentau added this to the 4.10.0 milestone Mar 15, 2024

fengyuentau requested review from asmorkalov, opencv-alalek and vpisarev March 15, 2024 09:35

fengyuentau commented Mar 15, 2024

View reviewed changes

fengyuentau mentioned this pull request Mar 15, 2024

rename cv::float16_t to cv::hfloat opencv/opencv_contrib#3697

Merged

6 tasks

remove zero(), bits() from fp16_t class

b22bd59

fp16_t -> hfloat

8bc4963

fengyuentau changed the title ~~core: Rename cv::float16_t to cv::fp16_t~~ core: Rename cv::float16_t to cv::hfloat Mar 18, 2024

fengyuentau added 2 commits March 18, 2024 15:09

remove cv::float16_t::fromBits; add hfloatFromBits

5346de8

undo changes in conv_winograd_f63.simd.hpp and conv_block.simd.hpp

3ca65b8

undo some changes in dnn

4b3bbe1

fengyuentau mentioned this pull request Mar 19, 2024

5.x core: rename cv::bfloat16_t to cv::bfloat #25232

Merged

6 tasks

vpisarev approved these changes Mar 20, 2024

View reviewed changes

vpisarev mentioned this pull request Mar 20, 2024

(5.x) Merge 4.x #25168

Closed

vpisarev self-requested a review March 20, 2024 12:04

vpisarev approved these changes Mar 20, 2024

View reviewed changes

vpisarev merged commit 3afe8dd into opencv:4.x Mar 21, 2024

fengyuentau deleted the float16_datatype_renaming branch March 26, 2024 07:52

asmorkalov mentioned this pull request Apr 1, 2024

5.x merge 4.x #25305

Merged

opencv-alalek reviewed Apr 2, 2024

View reviewed changes

fengyuentau mentioned this pull request Apr 10, 2024

Rename remaining float16_t for future proof #25387

Merged

6 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

core: Rename `cv::float16_t` to `cv::hfloat` #25217

core: Rename `cv::float16_t` to `cv::hfloat` #25217

Uh oh!

fengyuentau commented Mar 15, 2024 •

edited

Loading

Uh oh!

fengyuentau Mar 15, 2024

Uh oh!

vpisarev commented Mar 15, 2024

Uh oh!

fengyuentau commented Mar 15, 2024

Uh oh!

vpisarev commented Mar 15, 2024

Uh oh!

fengyuentau commented Mar 19, 2024

Uh oh!

vpisarev commented Mar 20, 2024

Uh oh!

opencv-alalek Apr 2, 2024

Uh oh!

fengyuentau Apr 2, 2024

Uh oh!

opencv-alalek Apr 2, 2024

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

core: Rename cv::float16_t to cv::hfloat #25217

core: Rename cv::float16_t to cv::hfloat #25217

Uh oh!

Conversation

fengyuentau commented Mar 15, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Pull Request Readiness Checklist

Uh oh!

fengyuentau Mar 15, 2024

Choose a reason for hiding this comment

Uh oh!

vpisarev commented Mar 15, 2024

Uh oh!

fengyuentau commented Mar 15, 2024

Uh oh!

vpisarev commented Mar 15, 2024

Uh oh!

fengyuentau commented Mar 19, 2024

Uh oh!

vpisarev commented Mar 20, 2024

Uh oh!

opencv-alalek Apr 2, 2024

Choose a reason for hiding this comment

Uh oh!

fengyuentau Apr 2, 2024

Choose a reason for hiding this comment

Uh oh!

opencv-alalek Apr 2, 2024

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

core: Rename `cv::float16_t` to `cv::hfloat` #25217

core: Rename `cv::float16_t` to `cv::hfloat` #25217

fengyuentau commented Mar 15, 2024 •

edited

Loading