vulkan : support conv_2d_dw with f16 weights

[Acly]

Small change to support OP_CONV_2D_DW with F16 weights like most other ops.

[0cc4m]

Thank you! Looks good, but to make sure the results are correct, can you add f16 tests to test-backend-ops?

Edit: If I manually set the test's kernel tensor to f16 it fails all tests, but not sure if I overlooked a step.

[Acly]

CPU doesn't actually support F16 for this OP, which makes it difficult to test with backend-ops
(I think it doesn't check the type anywhere so you just get wrong results because F16 data is interpreted as F32)

There is a separate test-conv2d-dw test in ggml repo I wanted to use, it checks against a reference implementation. But those ggml tests don't seem to exist in llama.cpp repo, and PRs in ggml directly are discouraged...

I'm open for ideas, maybe there's an easy way to check F16 results against the F32 version that I missed. I'm actively using those OPs here but there's always a chance I mess up during upstreaming so a test would be nice.

[0cc4m]

Fair enough, I didn't see that. The missing type check seems like an oversight. Maybe we can add support for f16 on CPU later and then add the test. Considering you didn't change the shader and the fp32 version works fine, the code should be fine.

vision.cpp looks like a neat project! Let me know if there's more things that can be done with the backend to enable or optimize for those models. Most of our issue and performance reports come from LLMs, some from Stable Diffusion, cool to see the backend being used for more kinds of models.