_Float16 support

I’d like to start a discussion about how clang supports _Float16 for target architectures that don’t have direct support for 16-bit floating point arithmetic.

The current clang language extensions documentation says, “If half-precision instructions are unavailable, values will be promoted to single-precision, similar to the semantics of __fp16 except that the results will be stored in single-precision.” This is somewhat vague (to me) as to what is meant by promotion of values, and the part about results being stored in single-precision isn’t what actually happens.

Consider this example:

_Float16 x;

_Float16 f(_Float16 y, _Float16 z) {

x = y * z;

return x;

}

When compiling with “-march=core-avx2” that results (after some trivial cleanup) in this IR:

@x = global half 0xH0000, align 2

define half @f(half, half) {

%3 = fmul half %0, %1

store half %3, half* @x

ret half %3

}

That’s not too unreasonable I suppose, except for the fact that it hasn’t taken the lack of target support for half-precision arithmetic into account yet. That will happen in the selection DAG. The assembly code generated looks like this (with my annotations):

f: # @f

%bb.0:

vcvtps2ph xmm1, xmm1, 4 # Convert argument 1 from single to half

vcvtph2ps xmm1, xmm1 # Convert argument 1 back to single

vcvtps2ph xmm0, xmm0, 4 # Convert argument 0 from single to half

vcvtph2ps xmm0, xmm0 # Convert argument 0 back to single

vmulss xmm0, xmm0, xmm1 # xmm0 = xmm0*xmm1 (single precision)

vcvtps2ph xmm1, xmm0, 4 # Convert the single precision result to half

vmovd eax, xmm1 # Move the half precision result to eax

mov word ptr [rip + x], ax # Store the half precision result in the global, x

ret # Return the single precision result still in xmm0

.Lfunc_end0:

– End function

Something odd has happened here, and it may not be obvious what it is. This code begins by converting xmm0 and xmm1 from single to half and then back to single. The first conversion is happening because the back end decided that it needed to change the types of the parameters to single precision but the function body is expecting half precision values. However, since the target can’t perform the required computation with half precision values they must be converted back to single for the multiplication. The single precision result of the multiplication is converted to half precision to be stored in the global value, x, but the result is returned as single precision (via xmm0).

I’m not primarily worried about the extra conversions here. We can’t get rid of them because we can’t prove they aren’t rounding, but that’s a secondary issue. What I’m worried about is that we allowed/required the back end to improvise an ABI to satisfy the incoming IR, and the choice it made is questionable.

For a point of comparison, I looked at what gcc does. Currently, gcc only allows _Float16 in C, not C++, and if you try to use it with a target that doesn’t have native support for half-precision arithmetic, it tells you “’_Float16’ is not supported on this target.” That seems preferable to making up an ABI on the fly.

I haven’t looked at what happens with clang when compiling for other targets that don’t have native support for half-precision arithmetic, but I would imagine that similar problems exist.

Thoughts?

Thanks,

Andy

While looking at the codegen Andy showed, I notice that the initial SelectionDAG looks like this for x86-64.

t0: ch = EntryToken
t2: f32,ch = CopyFromReg t0, Register:f32 %0
t6: f16 = fp_round t2, TargetConstant:i64<1>
t4: f32,ch = CopyFromReg t0, Register:f32 %1
t7: f16 = fp_round t4, TargetConstant:i64<1>
t8: f16 = fmul t6, t7
t10: i64 = Constant<0>
t12: ch = store<(store 2 into @x)> t0, t8, GlobalAddress:i64<half* @x> 0, undef:i64
t13: f32 = fp_extend t8
t16: ch,glue = CopyToReg t12, Register:f32 $xmm0, t13
t17: ch = X86ISD::RET_FLAG t16, TargetConstant:i32<0>, Register:f32 $xmm0, t16:1

The FP_ROUNDs for the arguments each have the flag set that indicates that the fp_round doesn’t lose any information. This is the TargetConstant:i64<1> as the second operand.

As far as I can tell, any caller of this would have an FP_EXTEND from f16 to f32 in their initial selection dag for calling this function. When the FP_EXTENDs are type legalized by DAGTypeLegalizer::PromoteFloatOp_FP_EXTEND, the FP_EXTEND will be removed completely with no replacement operations. I believe this means there is no guarantee that the f32 value passed in doesn’t contain precision beyond the range of f16. So the fp_round nodes saying no information is lost in the callee are not accurate.

While looking at the codegen Andy showed, I notice that the initial SelectionDAG looks like this for x86-64.

t0: ch = EntryToken
t2: f32,ch = CopyFromReg t0, Register:f32 %0
t6: f16 = fp_round t2, TargetConstant:i64<1>
t4: f32,ch = CopyFromReg t0, Register:f32 %1
t7: f16 = fp_round t4, TargetConstant:i64<1>
t8: f16 = fmul t6, t7
t10: i64 = Constant<0>
t12: ch = store<(store 2 into @x)> t0, t8, GlobalAddress:i64<half* @x> 0, undef:i64
t13: f32 = fp_extend t8
t16: ch,glue = CopyToReg t12, Register:f32 $xmm0, t13
t17: ch = X86ISD::RET_FLAG t16, TargetConstant:i32<0>, Register:f32 $xmm0, t16:1

The FP_ROUNDs for the arguments each have the flag set that indicates that the fp_round doesn’t lose any information. This is the TargetConstant:i64<1> as the second operand.

As far as I can tell, any caller of this would have an FP_EXTEND from f16 to f32 in their initial selection dag for calling this function. When the FP_EXTENDs are type legalized by DAGTypeLegalizer::PromoteFloatOp_FP_EXTEND, the FP_EXTEND will be removed completely with no replacement operations. I believe this means there is no guarantee that the f32 value passed in doesn’t contain precision beyond the range of f16. So the fp_round nodes saying no information is lost in the callee are not accurate.

That seems wrong to me from an ABI perspective; I would expect the burden to be on the caller to only pass a valid “half” value to a “half” parameter. But this leads back to Andy’s point: we’re inventing an ABI rule here.

The issue isn’t limited to calls either. If a half is a liveout of one basic block and used by another basic block. We’ll emit an fp_round with 1 for the second argument in the receiving basic block. But the producing basic block won’t have done anything to make it true.

Hey Andy,

I'd like to start a discussion about how clang supports _Float16 for target architectures that don't have direct support for 16-bit floating point arithmetic.

Thanks for bringing this up; we'd also like to get better support,
for sysv x86-64 specifically - AArch64 is mostly fine, and ARM is
usable with +fp16.

I'm not sure much of this discussion generalizes across platforms
though (beyond Craig's potential bug fix?). I guess the
"target-independent" question is: should we allow this kind of
"legalization" in the vreg assignment code at all? (I think that's
where it all comes from: RegsForValue, TLI::get*Register*)
It's convenient for experimental frontends: you can use weird types
(half, i3, ...) without worrying too much about it, and you usually
get something self-consistent out of the backend. But you eventually
need to worry about it and need to make the calling convention
explicit. But I guess that's a discussion for the other thread :wink:

The current clang language extensions documentation says, "If half-precision instructions are unavailable, values will be promoted to single-precision, similar to the semantics of __fp16 except that the results will be stored in single-precision." This is somewhat vague (to me) as to what is meant by promotion of values, and the part about results being stored in single-precision isn't what actually happens.

Consider this example:

_Float16 x;
_Float16 f(_Float16 y, _Float16 z) {
  x = y * z;
  return x;
}

When compiling with “-march=core-avx2” that results (after some trivial cleanup) in this IR:

@x = global half 0xH0000, align 2
define half @f(half, half) {
  %3 = fmul half %0, %1
  store half %3, half* @x
  ret half %3
}

That’s not too unreasonable I suppose, except for the fact that it hasn’t taken the lack of target support for half-precision arithmetic into account yet. That will happen in the selection DAG. The assembly code generated looks like this (with my annotations):

f: # @f
# %bb.0:
       vcvtps2ph xmm1, xmm1, 4 # Convert argument 1 from single to half
        vcvtph2ps xmm1, xmm1 # Convert argument 1 back to single
        vcvtps2ph xmm0, xmm0, 4 # Convert argument 0 from single to half
        vcvtph2ps xmm0, xmm0 # Convert argument 0 back to single
        vmulss xmm0, xmm0, xmm1 # xmm0 = xmm0*xmm1 (single precision)
        vcvtps2ph xmm1, xmm0, 4 # Convert the single precision result to half
        vmovd eax, xmm1 # Move the half precision result to eax
        mov word ptr [rip + x], ax # Store the half precision result in the global, x
        ret # Return the single precision result still in xmm0
.Lfunc_end0:
                                        # -- End function

Something odd has happened here, and it may not be obvious what it is. This code begins by converting xmm0 and xmm1 from single to half and then back to single. The first conversion is happening because the back end decided that it needed to change the types of the parameters to single precision but the function body is expecting half precision values. However, since the target can’t perform the required computation with half precision values they must be converted back to single for the multiplication. The single precision result of the multiplication is converted to half precision to be stored in the global value, x, but the result is returned as single precision (via xmm0).

I’m not primarily worried about the extra conversions here. We can’t get rid of them because we can’t prove they aren’t rounding, but that’s a secondary issue. What I’m worried about is that we allowed/required the back end to improvise an ABI to satisfy the incoming IR, and the choice it made is questionable.

As Richard said, an ABI rule emerged from the implementation, and I
believe we should solidify it, so here's a simple strawman proposal:
pass scalars in the low 16 bits of SSE registers, don't change the
memory layout, and pack them in vectors of 16-bit elements. That
matches the only ISA extension so far (ph<>ps conversions), and fits
well with that (as opposed to i16 coercion) as well as vectors (as
opposed to f32 promotion). To my knowledge, there hasn't been any
alternative ABI proposal (but I haven't looked in 1 or 2 years). It's
interesting because we technically have no way of accessing scalars
(so we have the same problems as i8/i16 vector elements, but without
the saving grace of having matching GPRs - x86, or direct copies -
aarch64), and there are not even any scalar operations.

Any thoughts? We can suggest this to x86-psABI if folks think this is
a good idea. (I don't know about other ABIs or other architectures
though).

Concretely, this means no/little change in IRGen. As for the SDAG
implementation, this is an unusual situation. I've done some
experimentation a long time ago. We can make the types legal, even
though no operations are. It's relatively straightforward to promote
all operations (and we made sure that worked years ago for AArch64,
for the pre-v8.2 mode), but vectors are fun, because of build_vector
(where it helps to have the truncating behavior we have for integers,
but for fp), extract_vector_elt (where you need the matching extend),
and insert_vector_elt (which you have to lower using some movd and/or
pinsrw trickery, if you want to avoid the generic slow via-memory
fallback).
Alternatively, we can immediately, in call lowering/register
assignment logic (this covers the SDAG cross-BB vreg assignments Craig
mentions) promote to f32 "via" i16. I'm afraid I don't remember the
arguments one way or the other, I can dust off my old patches and put
them up on phabricator.

-Ahmed

It seems that there are several issues here:

1. Should the front end be concerned with whether or not the IR that it is emitting can be translated into a well-defined IR?
2. How should the selection DAG handle data types whose representation isn't defined by the ABI we're targeting?
3. What should the ABI do with half-precision floats?

Working backward...

The third question here is obviously target specific. I've talked to HJ Lu about this, and he's working on an update to the x86 psABI. I believe that his eventual proposal will follow the lines of what you (Ahmed) suggested below, but I'm not completely proficient at comprehending ABI definitions so there may be some subtlety that I am misunderstanding in what he told me. I also talked to Craig about would be involved in making the LLVM x86 backend handle 'half' values this way. That involves a good bit of work, but it can be done.

The second question above probably involves a mix of target-independent and target-specific code. Right now the selection DAG code is operating on the assumption that it needs to do *something* with any IR it is given. It tries to make a reasonable choice, and the choice is consistent and predictable but not necessarily what the user expects. It seems like we should at the very least be producing a diagnostic so the user knows what we did (or even just that we did something). Then there are the specific problems Craig has brought up with the way we're currently handling 'half' values. Would defining a legal f16 type take care of those problems?

The first question exposes my lack of understanding of the proper role of the front end. It isn't clear to me what responsibility the front end has for enforcing conformance to the ABI. As a user of the compiler, I would like the compiler to tell me when code I've written can't be represented using the ABI I am targeting. Whether the front end should detect this or the backend, I don't know. I suppose it's also an open question how strictly this should be enforced. Is it a warning that can be elevated to an error at the users' discretion? Is it something that should be blocked by default but enabled by a user-specified option? Should it always be rejected?

-Andy

Right. IR and SelectionDAG representational choices aside, it seems to me
that, like GCC, Clang should not be permitting _Float16 on any target that
doesn't specify an ABI for it, because otherwise we're just creating
future compatibility problems for that target. I'm surprised and disappointed
that it wasn't implemented this way.

Unlike GCC, of course, we would implement it in all language modes on the
target, since there's zero reason to make it C-specific.

As for those internal representational choices: I'll leave SelectionDAG
up to the backend engineers, but I think that in IR, a half argument should
clearly correspond to a direct representation whenever hardware support for
half exists. If the ABI calls for the type to be promoted and passed as a
float, that should be done in the frontend, just as is done for small integer
types. It would then make sense to have an attribute (fpext?) for optimization
purposes that says that a parameter is guaranteed to be a promotion of a
smaller type; Clang could use this whenever it's allowed by the psABI.

John.

Hello,

I added _Float16 support to Clang and codegen support in the AArch64 and ARM backends, but have not looked into x86. Ahmed is right: AArch64 is fine, only a few ACLE intrinsics are missing. ARM has rough edges: scalar codegen should be mostly fine, vector codegen needs some more work.

Implementation for AArch64 was mostly straightforward (it only has hard float ABI, and has half register/type support), but for ARM it was a huge pain to plumb f16 support because of different ABIs (hard/soft), different architecture extensions of FP and FP16 support, and the existence of another half-precision type with different semantics. Sounds like you’re doing a similar exercise, and yes, argument passing was one of the trickiest parts.

IR and SelectionDAG representational choices aside, it seems to me that,

like GCC, Clang should not be permitting _Float16 on any target that doesn’t

specify an ABI for it, because otherwise we’re just creating future compatibility

problems for that target. I’m surprised and disappointed that it wasn’t implemented

this way.

Apologies, I missed that.

Sjoerd.

IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

Hello,

I added _Float16 support to Clang and codegen support in the AArch64 and ARM backends, but have not looked into x86. Ahmed is right: AArch64 is fine, only a few ACLE intrinsics are missing. ARM has rough edges: scalar codegen should be mostly fine, vector codegen needs some more work.

Implementation for AArch64 was mostly straightforward (it only has hard float ABI, and has half register/type support), but for ARM it was a huge pain to plumb f16 support because of different ABIs (hard/soft), different architecture extensions of FP and FP16 support, and the existence of another half-precision type with different semantics. Sounds like you're doing a similar exercise, and yes, argument passing was one of the trickiest parts.

IR and SelectionDAG representational choices aside, it seems to me that,

like GCC, Clang should not be permitting _Float16 on any target that doesn't

specify an ABI for it, because otherwise we're just creating future compatibility

problems for that target. I'm surprised and disappointed that it wasn't implemented

this way.

Apologies, I missed that.

It's alright, oversights happen (in both patch-writing and review). Can we get a volunteer to do the work to restrict this now? I'm a little crushed.

John.

Since Andy is my architect, I'm probably responsible (aka, being Volun-told :)) for this. Do we have a good idea which targets should currently have support? Is it just AArch64 and ARM?

-Erich

Woops, dropped llvm-dev :confused:

Disable up for review here: https://reviews.llvm.org/D57188

I can start looking Friday/Monday. But if it’s more urgent, perhaps someone else might want to have a look.

Sjoerd. IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

Thanks! Reviewing.

John.