Simplify use of OpenCL extensions

[VarLad]

could not load symbol "clHostMemAllocINTEL":

~/.julia/artifacts/c99d19d2f6dd75af3e997c4ecc4ac85e410cb7a7/lib/libOpenCL.so: undefined symbol: clHostMemAllocINTEL

To reproduce:

using OpenCL
error = Ref{Int32}()
al = cl.clHostMemAllocINTEL(cl.context(), C_NULL, 10000, 1, error)

Could this be an issue in using the function itself (wrong parameters or something)? 🤔

[maleadt]

I think this is intended. The ICD we use only provides entry-points for functions from the spec, and for extensions one needs to use clGetExtensionFunctionAddressForPlatform to fetch the entrypoint address:

julia> using OpenCL, pocl_jll

julia> ptr = cl.clGetExtensionFunctionAddressForPlatform(cl.platform(), "clHostMemAllocINTEL")
Ptr{Nothing} @0x000000013131f000

julia> error = Ref{Int32}()
Base.RefValue{Int32}(219365504)

julia> ccall(ptr, Ptr{Cvoid}, (cl.cl_context, Ptr{cl.cl_mem_properties_intel}, Csize_t, cl.cl_uint, Ptr{cl.cl_int}), cl.context(), C_NULL, 1000, 1, error)
Ptr{Nothing} @0x0000000121016e00

julia> error[]
0

It's a bit annoying that we can't re-use the existing Clang.jl-generated headers for this, although for the handful of USM functions you need it's probably fine duplicating those ccalls.

[VarLad]

@maleadt Would you have an example of how to pass the properties flags properly? I'm not sure how this works...
It seems to work differently from the svm flags in the current implementation
Unable to figure out how to use these from the docs:

[maleadt]

AFAIU it behaves like context properties, so you pass a list of properties starting with CL_MEM_ALLOC_FLAGS_INTEL, then a flags bitmask, and terminate with NULL.

julia> using OpenCL, pocl_jll

julia> fun = cl.clGetExtensionFunctionAddressForPlatform(cl.platform(), "clHostMemAllocINTEL")
Ptr{Nothing} @0x0000000135a78178

julia> error = Ref{Int32}();

julia> flags = cl.CL_MEM_ALLOC_WRITE_COMBINED_INTEL;

julia> properties = cl.cl_mem_properties_intel[cl.CL_MEM_ALLOC_FLAGS_INTEL, flags, 0];

julia> buf = ccall(fun, Ptr{Cvoid},
                   (cl.cl_context, Ptr{cl.cl_mem_properties_intel}, Csize_t, cl.cl_uint, Ptr{cl.cl_int}),
                   cl.context(), properties, 1000, 1, error)
Ptr{Nothing} @0x00000001322cee00

julia> @assert error[] == cl.CL_SUCCESS

You can also look at how PoCL implements this:

• properties parsing: https://github.com/pocl/pocl/blob/1990f7556242c0d55750b5bde7262f7ddaebe213/lib/CL/clMemAllocINTEL.c#L40-L58
• contained flag handling: https://github.com/pocl/pocl/blob/1990f7556242c0d55750b5bde7262f7ddaebe213/lib/CL/clMemAllocINTEL.c#L99-L118

[VarLad]

@maleadt I've been trying to use clGetMemAllocInfoINTEL, but so far, all my efforts to try and get any info out of that function are not working as I get CL_INVALID_VALUE (-30)

Here's an MWE for that:

    using OpenCL, pocl_jll

    ocl_extension(s) = cl.clGetExtensionFunctionAddressForPlatform(cl.platform(), s)

    function clDeviceMemAllocINTEL(context, device, properties, size, alignment, errcode_ret)
	ocl_intel = ocl_extension("clDeviceMemAllocINTEL")
	
        @ccall $ocl_intel(context::cl.cl_context, device::cl.cl_device_id, properties::Ptr{cl.cl_mem_properties_intel}, size::Csize_t, alignment::cl.cl_uint, errcode_ret::Ptr{cl.cl_int})::Ptr{Cvoid}
    end

    function clGetMemAllocInfoINTEL(context, ptr, param_name, param_value_size, param_value, param_value_size_ret)
	ocl_intel = ocl_extension("clGetMemAllocInfoINTEL")
	
        @ccall $ocl_intel(context::cl.cl_context, ptr::Ptr{Cvoid}, param_name::cl.cl_mem_info_intel, param_value_size::Csize_t, param_value::Ptr{Cvoid}, param_value_size_ret::Ptr{Csize_t})::cl.cl_int
    end     
    begin    
        ctx = cl.context()
        dev = cl.device()
        bytesize = 1000
        alignment = 1
        error_code = Ref{Int32}()
        flags = cl.CL_MEM_ALLOC_WRITE_COMBINED_INTEL

	ptr = clDeviceMemAllocINTEL(ctx, dev,cl.cl_mem_properties_intel[cl.CL_MEM_ALLOC_FLAGS_INTEL, flags, 0], bytesize, alignment, error_code)
	
	@assert error_code[] == cl.CL_SUCCESS

	result = Ref{Cvoid}()

	success = clGetMemAllocInfoINTEL(ctx, ptr, cl.CL_MEM_ALLOC_BASE_PTR_INTEL, 
        sizeof(cl.cl_mem_info_intel), result, C_NULL)

        @assert success == cl.CL_SUCCESS
    end

The last assertion fails...
I've also tried defining result as:

result = Ref{cl.cl_mem_info_intel}()

but that gives me the same error...

Finally, it works if I instead define result as:

result = Ptr{Cvoid}()

But I don't think setting a NULL pointer is a valid solution...

[maleadt]

Run with POCL_DEBUG=all:

[2024-11-26 14:32:46.175664000] PoCL: in fn clGetMemAllocInfoINTEL at line 53:
 |     ERROR | CL_INVALID_VALUE param_value_size (4) smaller than actual size (8)

[maleadt]

Will be done as part of #264