Add operators for +,-,*,/ for half with other scalar types
#969
Conversation
|
Can you comment more on the use case? What is the code you attempt to make work? |
|
Things like currently do not compile because the multiplication is ambiguous (the compiler could first create a |
dc42e2a to
4087bcc
Compare
|
The SYCL WG agreed that |
| return fp16::builtin_add(a._data, b._data), | ||
| return fp16::builtin_add(a._data, b._data)) | ||
| data = fp16::builtin_add(a._data, b._data), | ||
| data = fp16::builtin_add(a._data, b._data)); |
There was a problem hiding this comment.
Couldn't you simplify this along the lines of:
return detail::create_half(__hipsycl_backend_switch(
fp16::builtin_add(a._data, b._data),
....
));There was a problem hiding this comment.
Unfortunately not, because __hipsycl_backend_switch adds semicolons to its arguments, so this becomes something like
return detail::create_half(fp16::builtin_add(a._data, b._data););
which does not compile (this could of course be fixed by removing the semicolons in the definition of __hipsycl_backend_switch and adding them in the places where the macro is used). However, for the sscp backend, __hipsycl_backend_switch becomes an if statement, which again doesn't compile, and we can't just change that to a ternary operator because sometimes __hipsycl_backend_switch is used with return statements as its arguments and these are not allowed as expressions in the ternary operator.
There was a problem hiding this comment.
Ugh... even if SSCP might in principle work with ternary operator, it probably won't work with nvc++, where we need if target() branching.. So you are totally right :(
|
Can you resolve merge conflicts? |
This also required removing the `explicit` keyword for the constructors and adding three new constructors for `unsigned int`, `unsigned long` and `long`.
We can rely on implicit conversion here, because according to https://en.cppreference.com/w/cpp/language/static_cast, the static cast performs the same conversion as an implicit conversion (and implicit conversions from {integral types, double} -> float are valid according to https://en.cppreference.com/w/cpp/language/implicit_conversion)
Since we removed the half(fp16::half_storage) constructors, we must explicitly construct the half that is returned. Unfortunately, the macro `__hipsycl_backend_switch` contains an if statement for the host pass in the generic backend, so we cannot just write `auto data = __hipsycl_backend_switch(...)`.
2c17376 to
585e08c
Compare
This also required removing the
explicitkeyword for the constructors and adding three new constructors forunsigned int,unsigned longandlong.