[DAE][SYCL] Emit MD instead of updating integration header

[DenisBakhvalov]

No description provided.

[bader]

@kbobrovs, do you know why kmeans test fails here?
http://ci.llvm.intel.com:8010/#/builders/37/builds/2446/steps/15/logs/FAIL__SYCL__kmeans_cpp

[kbobrovs]

@kbobrovs, do you know why kmeans test fails here?
http://ci.llvm.intel.com:8010/#/builders/37/builds/2446/steps/15/logs/FAIL__SYCL__kmeans_cpp

@bader - no, I don't. Is this reproduced with this PR only? @NikitaRudenkoIntel - do you have any ideas?

#6 0x00007fdf0665842a __assert_fail_base /build/glibc-2ORdQG/glibc-2.27/assert/assert.c:89:0
 #7 0x00007fdf066584a2 (/lib/x86_64-linux-gnu/libc.so.6+0x304a2)
 #8 0x000055e8034a32e1 llvm::genx::GenXSPIRVWriterAdaptor::runOnFunction(llvm::Function&) (/localdisk2/sycl_ci/buildbot/worker/Lit_With_Cuda/llvm.obj/bin/clang-12+0x2dbd2e1)
 #9 0x000055e8034a3745 llvm::genx::GenXSPIRVWriterAdaptor::runOnModule(llvm::Module&) (/localdisk2/sycl_ci/buildbot/worker/Lit_With_Cuda/llvm.obj/bin/clang-12+0x2dbd745)

[NikitaRudenkoIntel]

Will take a look

[bader]

@DenisBakhvalov, could you apply suggestion from this comment: #2226 (comment)?

I think this patch might conflict with #2207. With this patch we apply DAE unconditionally, but there are missing parts of the implementation to make it work (e.g. integration with the runtime).
Can we temporally exclude this new pass from standard optimization pipeline until the rest of the framework is ready?

[NikitaRudenkoIntel]

For some reason this patch breaks genx.kernels metadata. The first argument of KernelMD should be a reference to kernel. This is how it looks like without this patch:

!genx.kernels = !{!4, !7}
!4 = !{void (%struct._ZTS5Point.Point addrspace(1)*, %struct._ZTS8Centroid.Centroid addrspace(1)*, %struct._ZTS5Accum.Accum addrspace(1)*, i1, i8 addrspace(1)*)* @"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE_clES3_E6kMeans", !"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE_clES3_E6kMeans", !5, i32 0, i32 0, !5, !6}
!7 = !{void (%struct._ZTS8Centroid.Centroid addrspace(1)*, %struct._ZTS5Accum.Accum addrspace(1)*, i8 addrspace(1)*)* @"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE0_clES3_E16kCompCentroidPos", !"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE0_clES3_E16kCompCentroidPos", !8, i32 0, i32 0, !8, !9}

And this is with the patch:

!genx.kernels = !{!4, !7}
!4 = distinct !{null, !"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE_clES3_E6kMeans", !5, i32 0, i32 0, !5, !6}
!7 = distinct !{null, !"_ZTSZZZ4mainENK3$_0clEbENKUlRN2cl4sycl7handlerEE0_clES3_E16kCompCentroidPos", !8, i32 0, i32 0, !8, !9}

The references to kernels are missing, which eventually causes the failure.

[bader]

Can we temporally exclude this new pass from standard optimization pipeline until the rest of the framework is ready?

Another option is to add a command line option to control this optimization with "off" value by default.

[DenisBakhvalov]

@NikitaRudenkoIntel, thanks for the analysis. It turns out that we cannot run DAE in ESIMD context since the pointers to SPIR kernel functions are saved in !genx.kernels metadata. DAE pass creates new function prototypes that break metadata.

@bader, I removed the pass from the standard pipeline and put it in the place that is specific for SYCL passes and guarded it under a new temporary option: -fenable-sycl-dae. This option is only used until other parts of the solution are not in place.

[bader]

@NikitaRudenkoIntel, thanks for the analysis. It turns out that we cannot run DAE in ESIMD context since the pointers to SPIR kernel functions are saved in !genx.kernels metadata. DAE pass creates new function prototypes that break metadata.

Do I understand it correctly, that fixed version doesn't apply DAE in ESIMD mode?

@bader, I removed the pass from the standard pipeline and put it in the place that is specific for SYCL passes and guarded it under a new temporary option: -fenable-sycl-dae. This option is only used until other parts of the solution are not in place.

Thanks. I'll open an issue to enable this optimization for non-SPIR kernels when full implementation is in place.

[DenisBakhvalov]

@NikitaRudenkoIntel, thanks for the analysis. It turns out that we cannot run DAE in ESIMD context since the pointers to SPIR kernel functions are saved in !genx.kernels metadata. DAE pass creates new function prototypes that break metadata.

Do I understand it correctly, that fixed version doesn't apply DAE in ESIMD mode?

Yes.

[elizabethandrews]

FE changes LGTM. I did not check llvm/ changes