[code-reflection] Integrated: [hat] Initial support for 16-bits float types (half)

[Juan Fumero]

On Tue, 21 Oct 2025 08:40:11 GMT, Juan Fumero <jfumero at openjdk.org> wrote:

> This PR extends HAT with support for OpenCL and CUDA `half` types. 
> 
> From the Java side:
> 
> 
> F16Array.F16 ha = a.array(kernelContext.gix);
> F16Array.F16 hb = b.array(kernelContext.gix);
> hb.value(ha.value());
> 
> 
> We can also perform some common operations with half types:
> 
> 
> F16Array.F16 ha = a.array(kernelContext.gix);
> F16Array.F16 hb = b.array(kernelContext.gix);
> 
> F16Array.F16 r1 = F16.mul(ha, hb);
> F16Array.F16 r2 = F16.div(ha, hb);
> F16Array.F16 r3 = F16.sub(ha, hb);
> F16Array.F16 r4 = F16.add(r1, r2);
> F16Array.F16 r5 = F16.mul(r4, r3);
> F16Array.F16 hC = c.array(kernelContext.gix);
> 
> 
> To run with FP16, HAT needs to run GPUs with HALF support for both OpenCL and CUDA backends. For example, for OpenCL, it needs to run on devices that support the `cl_khr_fp16` macro. 
> 
> In near future, we will extend with more operations. 
> 
> How to test (OpenCL):
> 
> 
> HAT=SHOW_CODE,SHOW_COMPILATION_PHASES java @hat/otest ffi-opencl hat.test.TestFP16Type
> 
> 
> CUDA: 
> 
> 
> HAT=SHOW_CODE,SHOW_COMPILATION_PHASES java @hat/otest ffi-cuda hat.test.TestFP16Type

This pull request has now been integrated.

Changeset: 5128c70b
Author:    Juan Fumero <jfumero at openjdk.org>
URL:       https://git.openjdk.org/babylon/commit/5128c70b87c209d781307cf583826a2f5b8a1a8d
Stats:     986 lines in 27 files changed: 929 ins; 44 del; 13 mod

[hat] Initial support for 16-bits float types (half)

-------------

PR: https://git.openjdk.org/babylon/pull/624