[Mesa-dev] Adding a SPIR-V target to libclc

Fri Feb 23 00:51:29 UTC 2018

Jan Vesely <jan.vesely at rutgers.edu> writes:

> On Thu, 2018-02-15 at 22:50 -0800, Francisco Jerez wrote:
>> Jan Vesely <jan.vesely at rutgers.edu> writes:
>> 
>> > On Thu, 2018-02-15 at 20:36 -0800, Francisco Jerez wrote:
>> > > Jan Vesely <jan.vesely at rutgers.edu> writes:
>> > > 
>> > > > On Thu, 2018-02-08 at 15:56 -0800, Francisco Jerez wrote:
>> > > > > Jan Vesely <jan.vesely at rutgers.edu> writes:
>> > > > > 
>> > > > > > On Thu, 2018-02-08 at 23:16 +0100, Pierre Moreau wrote:
>> > > > > > > (Moving the conversation to its own thread.)
>> > > > > > > 
>> > > > > > > > target agnostic libclc is rather difficult to do. CLC includes 3 levels
>> > > > > > > > of precision on float (fp32) operands; full, half, native. The
>> > > > > > > > implementation of each depends on capabilites of specific device (e.g.
>> > > > > > > > vega(VI+?) can do 1 ULP log2/exp2 in hw, other targets need sw
>> > > > > > > > implementation to meet CLC requirement of 3ulp). Any conversion backend
>> > > > > > > > would thus need to implement sw versions of math builtins for targets
>> > > > > > > > that can't perform the op in HW.
>> > > > > > > 
>> > > > > > > My initial thought for the target agnostic libclc, was to just provide some
>> > > > > > > (fake?) implementations of OpenCL built-in functions to make clang happy and
>> > > > > > > let me compile kernels using “get_global_id()”, as well as include headers
>> > > > > > > defining OpenCL specific types like “float4” or others. If there is another
>> > > > > > > (better?) way to achieve this, I am all ears. (There is probably one, as I had
>> > > > > > > no issues when using the Khronos LLVM/clang fork rather than Tomeu’s
>> > > > > > > out-of-tree module, the former having also some bits and pieces in clang.)
>> > > > > > 
>> > > > > > I don't think you need libclc for this. workitem IDs are
>> > > > > > platform/device specific, and iiuc SPIR-V builtins should handle it in
>> > > > > > an abstract way [0]. any conversion consuming SPIR-V needs to replace
>> > > > > > those with device/platform specific way of obtaining the information.
>> > > > > > you can also use clang's clc header to provide data types [1].
>> > > > > > 
>> > > > > > 
>> > > > > > [0] https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#B
>> > > > > > uiltIn
>> > > > > > [1] https://github.com/llvm-mirror/clang/blob/master/lib/Headers/opencl
>> > > > > > -c.h
>> > > > > > 
>> > > > > > > 
>> > > > > > > > Extending the current libclc to provide target specific SPIR-V binaries
>> > > > > > > > in addition to/in place of LLVM IR is rather straightforward. Adding
>> > > > > > > > additional targets it's more work since it relies on clang to support
>> > > > > > > > those targets.
>> > > > > > > 
>> > > > > > > I’m curious how those target specific SPIR-V binaries would look like. I can
>> > > > > > > imagine how some functions like “OpSign” could be implemented using other
>> > > > > > > SPIR-V functions, but how would you handle something like “get_local_id()”? If
>> > > > > > > you define it as the built-in “LocalInvocationId” and don’t supply an
>> > > > > > > implementation of it, then you lose the target specificness. On the other hand,
>> > > > > > > if you want to keep it device-specific, how would you express that in SPIR-V?
>> > > > > > 
>> > > > > > getting IDs is not a problem. SPIR-V should provide builtins for that.
>> > > > > > 
>> > > > > > The problem I had in mind is when SPIR-V binary calls e.g. exp2(). You
>> > > > > > can either assume that the op needs CLC precision (3 ulp), or device
>> > > > > > native precision.
>> > > > > 
>> > > > > That's up to the SPIR-V extended instruction set specification to define
>> > > > > what precision the exp2 built-in is supposed to have.  
>> > > > > 
>> > > > > > SPIR-V binary can also call exp2(fp64), which does not have an
>> > > > > > equivalent GPU instruction.
>> > > > > 
>> > > > > Then it should probably be lowered by the SPIR-V front-end, right?
>> > > > 
>> > > > I'm not sure what you mean by "spir-v frontend". If it's the tool that
>> > > > generates SPIR-V, then no, not really. 
>> > > 
>> > > No, I meant the SPIR-V front-end of the driver (or whatever translation
>> > > pass in control of the driver is translating machine-agnostic SPIR-V
>> > > into some other more hardware-specific representation of the program).
>> > 
>> > OK. my question still stands. How does generic SPIR-V based libclc
>> > help the process?
>> > 
>> 
>> That I can think of now, it would remove the need for maintaining any
>> target-specific knowledge in libclc, for plumbing target-specific
>> information in order to select the right libclc flavour at link time,
>
> it would only move the specific decisions to SPIR-V lowering time.
> I understand the advantage of cross language usefulness, but I'm not
> sure how practical it is.  Taking the below example of exp2(fp64). CLC
> requires precision <= 2ulp, other languages might have different
> requirements. Thus to achieve good performance, you'd need to lower to
> different routine for each precision requirement.
>

Because of SPIR-V extended instruction sets, you do know at SPIR-V
translation time what the required precision bound is, so the
translation logic can still provide a more optimal implementation for
certain APIs -- Or not, in the most common case where the performance
difference isn't clear enough to justify separate lowering for each
flavour of SPIR-V, doing things at SPIR-V translation time allows you
not to bother to write a separate implementation for separate APIs until
you have enough evidence that it's helpful.

>> and it would allow solving common problems in a place where there is a
>> chance that the solution could be shared among different drivers and
>> APIs (e.g. the exp2(fp64) lowering example you mentioned earlier is not
>> exclusively useful to CL).
>
> the exp2 example is not something that could be addressed in generic
> SPIR-V libclc, since the decision is hw specific.

But it can be addressed at SPIR-V translation time with minimal
hardware-specific knowledge, and with largely target-independent logic,
like the lowering of other fp64 intrinsics is done these days.

> Sure we can provide implementation of all CLC builtins using only the
> core SPIR-V operations,

That's fine for the built-ins that admit a compliant and reasonably
efficient implementation in terms of core SPIR-V exclusively, but I
wasn't arguing for lowering all of them in terms of core SPIR-V.

> but if a SPIR-V input uses clc extended instructions the same
> functionality would have to be implemented in SPIR-V lowering anyway,
> so it's just simpler to implement libclc as single op wrappers over
> CLC extended ops.

No objection against implementing libclc as a pile of one-liners written
in terms of CLC extended SPIR-V ops -- That's fully target-independent.

> Am I missing anything?
>
> Jan
>
>> 
>> > Jan
>> > 
>> > > 
>> > > > My understanding is that those are run prior to application
>> > > > distribution, and therefore have no information about the target HW.
>> > > > 
>> > > > So if a program imports "CLC.std.11" extended instruction set to get
>> > > > access CLC builtin functions. What would a generic SPIR-V libclc
>> > > > provide?
>> > > > 
>> > > > > 
>> > > > > > It's easier to translate these to libclc function calls (combined with
>> > > > > > the right library implementation of the exp2 builtin), than try to
>> > > > > > generate exp2 algorithm when converting to NIR (or anything else
>> > > > > > really).
>> > > > > > 
>> > > > > 
>> > > > > But the SPIR-V front-end will need to lower that in terms of
>> > > > > instructions supported by the back-end anyway in order to be able to
>> > > > > handle general SPIR-V shaders as input, right?  So why re-implement the
>> > > > > lowering for those operations in libclc in a way that's only going to be
>> > > > > useful for the OpenCL C language but not for other APIs?
>> > > > > 
>> > > > > > The current libclc mostly assumes that LLVM ops are done in device
>> > > > > > native precision, and provides sw implementation of operations that
>> > > > > > don't have conformant device instruction.
>> > > > > 
>> > > > > But I don't think there is any disadvantage from having a libclc
>> > > > > implementation that doesn't make any precision assumptions beyond what
>> > > > > is stated in the SPIR-V spec.  In fact that would have the IMO more
>> > > > > desirable advantage that you could re-use one and the same libclc
>> > > > > implementation for *all* back-ends that want SPIR-V as input.
>> > > > 
>> > > > Sure, a compiler-rt library would be more useful (usable by multiple
>> > > > languages). However, unlike target specific libclc, it's not available
>> > > > atm.
>> > > > 
>> > > > Jan
>> > > > 
>> > > > > 
>> > > > > > This obviates the need for compiler-rt library.  And alternative
>> > > > > > approach is to assume that the ops provide full precision and use
>> > > > > > target intrinsics for native precision. it's still target specific if
>> > > > > > a library call uses the former or the latter.
>> > > > > > 
>> > > > > > regards,
>> > > > > > Jan
>> > > > > > 
>> > > > > > > 
>> > > > > > > Regards,
>> > > > > > > Pierre
>> > > > > 
>> > > > > _______________________________________________
>> > > > > mesa-dev mailing list
>> > > > > mesa-dev at lists.freedesktop.org
>> > > > > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
>> > > > 
>> > > > -- 
>> > > > Jan Vesely <jan.vesely at rutgers.edu>
>> > 
>> > -- 
>> > Jan Vesely <jan.vesely at rutgers.edu>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 227 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20180222/c1634b04/attachment-0001.sig>