How is OpenCL floatn different from a custom struct - struct

Question

I'm having problems passing a structure of vectors to an OpenCL kernel.
Long story short, I have found that I CAN pass a C structure:
typedef struct {
cl_float4 vec;
} my_type;
. . . to a kernel, if I declare it like:
typedef struct { float s[4]; } my_float4;
typedef struct { my_float4 vec; } my_type; //use custom float4
__kernel void function(const my_type test) {}
. . . but NOT, if I declare it like:
typedef struct { float4 vec; } my_type; //use built-in float4
__kernel void function(const my_type test) {}
So my question is, what makes a float4 different and special? In the second example, I get CL_INVALID_ARG_SIZE on clSetKernelArg. In the host code, I'm using cl_float4--isn't that supposed to be paired with float4 in the kernel?
EDIT: by request, the clSetKernelArg code (error handling cropped):
template <typename type_data> void set_argument(int arg_index, const type_data* data) {
cl_int err = clSetKernelArg(kernel,arg_index,sizeof(type_data),data);
if (err!=CL_SUCCESS) throw "ERROR"; /*simplified*/
}
I call the function with a struct declared exactly as above:
my_type test;
my_kernel->set_argument(0,&test);

Answer 1

This sounds like a bug with your implementation - you should be able to pass that user-defined type to a kernel without issue. What platform / SDK are you using? I was able to successfully declare this type in host code:
typedef struct _mytype {
cl_float4 val;
} mytype;
... and then later use it:
mytype t;
status = clSetKernelArg(kernel, argc++, sizeof(mytype), &t);
check_status("clSetKernelArg:mytype t", status);
Where check_status does what you'd expect: checks for a non-CL_SUCCESS return value.
The kernel looks like this:
typedef struct _mytype {
float4 value;
} mytype;
kernel void do_it(/* ... */, const mytype test) { /* ... */ }
This compiles and executes, giving correct results on the CPU device.