Brice Videau

> And a number of the examples shipped with `pocl` run with/against Apple's implementation. Those that do actually seem to run a lot faster. Those that don't often fail with...

> Remember that CL_INVALID_PLATFORM error in the logs above? That came from `clGetPlatformInfo` because your code retrieves the pointer to that function dynamically and `ocd_icd_wrapper` is not a traditional wrapper...

> Where exactly? For `clGetPlatformInfo` it first tries to load it dynamically from the library: https://github.com/OCL-dev/ocl-icd/blob/fdde6677b21329432db8b481e2637cd10f7d3cb2/ocl_icd_loader.c#L457 That call will *not* fail with OpenCL.framework, and will give you a pointer to...

Indeed, this is my understanding as well. It just took me way too long to make the connection. > Avoiding this is possible, but you'd have to overload the function....

To help with the review, I tried to create a minimal diff between the current tip of master and the headers generated her. I did this using `clang-format` and `indent`....

I would advise 1 change: I would create 2 macros guarding prototype declarations: * `CL_NO_EXPORTED_PROTOTYPES` to guard entry points that are exported by the loader, including extensions which are exported...

For `cl_ext_image_from_buffer` the `#ifdef CL_VERSION_3_0` guard has disappeared. Which is correct?

~Guards `#if defined(_WIN32)` have been added around `cl_dx9_surface_info_khr` definition, it looks the right thing to do, just checking this is intended.~

`clCreateProgramWithILKHR_fn` `CL_API_SUFFIX__VERSION_1_2` got removed, and was previously inconsistent with `clCreateProgramWithILKHR`.

`#ifdef CL_VERSION_1_2` of `cl_arm_get_core_id` got removed.