Scalar input can’t be float16_t

[Vernlium]

Hi all,
In Halide 14.0.0, I found that writing like below two all work well:

Input<uint8_t> Input3{"Input3"};
void generate() {
     Func5() = Input3;
}

and

Input<Buffer<float16_t>> Input3{"Input3"};

void generate() {
     Func5() = Input3(x,y)[0][0];
}

But writing like below failed:

Input<float16_t> Input3{"Input3"};
void generate() {
    Func5() = Input3;
}

And it report error is:

/include/Halide.h:25503:7: error: invalid use of 'Halide::Internal::GeneratorInputImplBase<Halide::float16_t, Halide::float16_t>' {aka 'void'}
 class GeneratorInput : public Internal::GeneratorInputImplBase<T> {
       ^~~~~~~~~~~~~~
./include/Halide.h:25508:11: error: 'Halide::GeneratorInput<Halide::float16_t>::Super' {aka 'void'} is not a class, struct, or union type
     using TBase = typename Super::TBase;
           ^~~~~
./include/Halide.h:25514:11: error: 'std::conditional<false, Halide::Internal::GeneratorInput_Buffer<Halide::float16_t>, void>::type' {aka 'void'} is not a class, struct, or union type
     using IntIfNonScalar =
           ^~~~~~~~~~~~~~
./testgen.cpp: In member function 'void HicvRandomTc0001::generate()':
./testgen.cpp:18:9: error: 'Func5' was not declared in this scope; did you mean 'Func'?
         Func5()=Input3();
         ^~~~~
         Func
./testgen.cpp:18:24: error: no match for call to '(Halide::Internal::GeneratorBase::Input<Halide::float16_t> {aka Halide::GeneratorInput<Halide::float16_t>}) ()'
         Func5()=Input3();
                        ^

Thank you in advance !

[abadams]

The issue with scalar float16 inputs is the ABI. We're not sure exactly how they should be passed on different platforms. Some ABIs put them in floating-point-specific registers, but in other cases Halide treats them as uint16s for passing them. For now you should just cast scalar float16 inputs to float32 on the C++ side and cast them back to float16 in the Halide code.

[Vernlium]

Thank you for your reply.

[steven-johnson]

We can/should improve the compile time error message to make this answer clearer. (Is there already an issue for this?)

…

On Fri, Jun 24, 2022, 5:19 PM Andrew Adams ***@***.***> wrote: The issue with scalar float16 inputs is the ABI. We're not sure exactly how they should be passed on different platforms. Some ABIs put them in floating-point-specific registers, but in other cases Halide treats them as uint16s for passing them. For now you should just cast scalar float16 inputs to float32 on the C++ side and cast them back to float16 in the Halide code. — Reply to this email directly, view it on GitHub <#6812 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AACQJ67SX7A3ASWGWQR3LKDVQXNY3ANCNFSM5ZWIESFQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>