Removing cl_khr_fp16

[LWisteria]

As of #224 and #261, the current solution for "half" needs cl_khr_fp16.

However, it must not need it theoretically and could be removed.

[LWisteria]

It was inserted on ac3adc9 to solve #131 .

@vorj @nsakabe-fixstars Can you remember and explain the reason?
I remember the reason is just to avoid the problem that sqrt of int8 and int16 is stored as the half format.
It is not enough to prove its necessity because only storing half format uses no half operation which needs to set cl_khr_fp16.

[vorj]

int8 and int16

No, problematic types are int8 , uint8 , and bool . Some functions have only floating type(s) for input/output , then the types are converted to half .
int16 is converted to float , so it has no problem.
The problematic functions list has been shown as test cases that are failed when we disable cl_khr_fp16 , by @nsakabe-fixstars.

It is not enough to prove its necessity because only storing half format uses no half operation which needs to set cl_khr_fp16.

I really think so, but NVIDIA doesn't .

[LWisteria]

@vorj Thanks.

Your previous comment

In OpenCL C on NVIDIA GPU, enabling cl_khr_fp16 is required to copy a half value.

must be checked on the current titanv (CUDA 10).

Furthermore, to solve the problem with only the copy operation, for example, we can treat fp16 as int16. I remember there were some reasons to disturb such solutions but I don't remember what they actually were.

Anyway, I don't think that investigation for the solution to avoid it is perfect. The priority to pass the tests was higher than enough investigation at that time.
Now it's time to try it...!

[vorj]

Your previous comment

In OpenCL C on NVIDIA GPU, enabling cl_khr_fp16 is required to copy a half value.

must be checked on the current titanv (CUDA 10).

Please see the secondary machine error log for #260 .
I've shown the extracts of the errors below:

warning: cannot enable cl_khr_fp_16 extension on this platform - ignoring.
error: declaring function return value of type 'half' is not allowed; did you forget * ?
error: loading directly from pointer to type 'const __attribute__((address_space(16776963))) half' is not allowed

It seems that NVIDIA doesn't allowed to copy half value yet (cl_khr_fp16 enable is just ignored). Originally we use cl_khr_fp16 only to enable copying and we've never added half arithmetic operations, so it's obvious. Of course, we can pass the test cases without enabling cl_khr_fp16 on the primary machine.

[vorj]

#269 fixed this issue, so I close it.