[libspirv] reuse clc function in __spirv_ocl_normalize/fast_normalize

[wenju-he]

Before this PR, libspirv-nvptx64--nvidiacl uses __nv_rsqrtf/isinf/copysign
in these two functions, and libspirv-amdgcn--amdhsa uses __ocml_rsqrt/copysign.
We can upstream use of target built-in to CLC library:
Use of __nv_rsqrtf/isinf is to be upstreamed by llvm/llvm-project#150174.
Then the changes to libspirv-nvptx64--nvidiacl should be minimized.

Generic implementation of __clc_rsqrt is better than __ocml_rsqrt according to
llvm/llvm-project#152436.
Use of these target built-in scalarize llvm vector intrinsic. So it is
likely better to use generic implementation.
In addition, use of generic implementation aligns with OpenCL library.

Delete libspirv/generic/math/floatn.inc.
Delete unused clc/relational/floatn.inc and libspirv/generic/math/minmag.inc.

[frasercrmck]

Am I right in thinking the NVPTX/AMDGPU-specific builtins are not coming from __spirv_ocl_(fast)?normalize but from builtins those are calling? It's a bit of a shame that we need to duplicate the "wrong" function to work around this, to avoid impacting the OpenCL implementations.

We obviously don't want to recompile the entire CLC library for both OpenCL and SPIR-V NVPTX builtins libraries, when only a couple of functions may change. I wonder if we could somehow swap in/out just those builtins with differing implementations but keep the 99% identical module somehow. It's just an idea at this stage so I'm not sure how it'd work.

Even so, that's future work and needn't hold up this patch.

[wenju-he]

Am I right in thinking the NVPTX/AMDGPU-specific builtins are not coming from __spirv_ocl_(fast)?normalize but from builtins those are calling?

Yes

It's a bit of a shame that we need to duplicate the "wrong" function to work around this, to avoid impacting the OpenCL implementations.

I have

• created PR [libclc] Implement __clc_rsqrt with __ocml_rsqrt_* functions llvm/llvm-project#152436 to upstream use of _ocml_rsqrt. The PR is closed because use of _ocml_rsqrt should be deleted.
• updated PR [libclc] Implement clc_log/sinpi/sqrt with __nv_* functions llvm/llvm-project#150174 to __clc_rsqrt/isinf with _nv* functions.

Another LLVM IR change is copysign. NVPTX/AMDGPU-specific only provides scalar copysign. So it is probably not worthy to use target-specific copysign to scalarize vector llvm.copysign intrinsic.

We obviously don't want to recompile the entire CLC library for both OpenCL and SPIR-V NVPTX builtins libraries, when only a couple of functions may change. I wonder if we could somehow swap in/out just those builtins with differing implementations but keep the 99% identical module somehow. It's just an idea at this stage so I'm not sure how it'd work.

I've deleted customization files for NVPTX/AMDGPU.

[wenju-he]

@intel/llvm-gatekeepers please merge, thanks