When used msvc clang-cl, kernel a64_fp16_4x4_3x3 isn't built

Question

When used msvc clang-cl, kernel a64_fp16_4x4_3x3 isn't built

allnes opened this issue 4 months ago · 6 comments

Hi,
I used Windows system on Snapdragon ARM CPU and noticed that kernel can't build:
arm_conv::winograd::weight_transform::a64_fp16_4x4_3x3
and other fp16 kernels in:
src/core/NEON/kernels/convolution/winograd/input_transforms
src/core/NEON/kernels/convolution/winograd/output_transforms
src/core/NEON/kernels/convolution/winograd/weight_transforms
This problem create failure of our build in OpenVINO on this system.
Could you tell are there reasons for that?

Answer 1 · 2024-08-12T13:08:58.000Z

Hi @allnes

Presently we support only armv8a targets for Windows, could you please let us know which options are you using to build ACL when you see this error?

Answer 2 · 2024-08-12T16:01:35.000Z

Hi this command was generated:
scons neon=1 opencl=0 openmp=0 cppthreads=1 examples=0 Werror=0 gemm_tuner=0 reference_openmp=0 validation_tests=0 benchmark_tests=0 data_layout_support=all arch=arm64-v8.2-a estate=64 os=windows build=native

Answer 3 · 2024-08-12T16:03:55.000Z

armv8a arch disable fp16 kernels, do you have plan to support arm64-v8.2-a instruction set?

Answer 4 · 2024-08-13T08:42:01.000Z

HI @allnes

Yes, we will look into adding support for arm64-v8.2-a

Answer 5 · 2024-08-14T11:44:06.000Z

Hi @allnes

The following patch fixes the linker errors.

Hope this helps.

Answer 6 · 2024-08-14T15:33:57.000Z

Hi @morgolock

Thank you, patch is worked, we'll wait for the release