Error: Kernel argument type mismatch vol 2

[Patryk27]

Hi,

I've got following Rust and OpenCL code:

queue.kernel_builder("foo")
    .arg(Uint::new(123u32))
    .build()

__kernel void foo(uint bar) {
  //
}

... which ends up with this, quite unexpected and counter-intuitive, run-time error:

Kernel argument type mismatch. The argument named: 'bar' at index: [0] should be a 'uint' (ArgType { base_type: Uint, cardinality: One, is_ptr: false }).

Seems like the core of this problem lays in the ArgType::matches() method:

// kernel.rs:1754
let card_match = match self.cardinality {
    Cardinality::One => TypeId::of::<cl_uint>() == type_id,
    /* ... */
};

cl_uint is an alias for u32, which has different type-id than Uint - so even though this code is, in fact, correct and works just fine, it refuses to compile with the Intel's driver [1].

I was able to successfully fix this with TypeId::of::<cl_uint>() == type_id || TypeId::of::<Uint>() == type_id and I might be able to prepare a complete fix (for other types) too, if you let me know that I'm doing it the right way (TM) :-)

[1] The issue seems to occur only with the Intel's OpenCL driver - I've got an Nvidia card that runs the original code without any changes.

By the way: I'm aware of the #146 - I've created another one, because even though both problems end up with similar error message, the essence of both problems is different.

[c0gent]

Yeah your fix is exactly right and I'd be happy to merge a PR with those changes.

I'm curious to know why you're wrapping the u32 in a Uint (not that there's anything wrong with it).

[Patryk27]

Great, I'll try to prepare something later.

About the Uint: I've got a kernel where I'm passing together some Uint2s with Uints, so doing it like this:

queue.kernel_builder("foo")
    .arg(Uint::new(100))
    .arg(Uint2::new(200, 300))
    .arg(Uint2::new(400, 500))
    .build()

... seemed more logical to me than just .arg(100) (actually I've never even tried passing just the integer, until I've stepped on this bug and begin to experiment).

[Patryk27]

Hi,

I've come up with this: https://github.com/Patryk27/ocl/commit/03086863e95e43033ae67f1530801cb57032e1a3

It seems to work correctly + using the macro makes the entire method more legible IMO - could you please take a look?

Thanks!

[c0gent]

Yeah looks great. Go ahead and submit a pull request :)

[Patryk27]

D'oh, I forgot about the PR itself :-) #161