Bump DPC++ and hipSYCL versions used for CI

[psalz]

This bumps the version of DPC++ to 6699a5d (December 8) and hipSYCL to b836149 (December 9). I've also updated the CI filter for DPC++, which is now able to compile 12 additional test categories, failing for 2 new ones, for a total of 5 categories that do not compile yet.

I've also adjusted the CI workflow to encode an implementation's version into the Docker image tag (i.e. there is now a dpcpp-6699a5d tag, as opposed to just dpcpp) so that submitting this PR won't affect current CI runs (until it is merged).

Disable DPC++ range rounding feature to speed up compilation.

[psalz]

Shoot, of course we now need to adjust the naming scheme for the ComputeCpp image as well. @ProGTX who should I talk to regarding this going forward? You, Scott Straughan, Rod Burns..? Also, can we bump the image to 2.8?

[bader]

Compilation with dpcpp takes 148 minutes (~2.5 hours). Is this acceptable for pre-commit?

[psalz]

Compilation with dpcpp takes 148 minutes (~2.5 hours). Is this acceptable for pre-commit?

No, not really. These are the top 5 culprits:

1971.5 accessor_api_buffer_core.cpp.o (../tests/accessor/accessor_api_buffer_core.cpp)
1709.1 handler_copy.cpp.o (../tests/handler/handler_copy.cpp)
532.1 math_builtin_integer.cpp.o (tests/math_builtin_api/math_builtin_integer.cpp)
469.2 nd_item_async_work_group_copy.cpp.o (../tests/nd_item/nd_item_async_work_group_copy.cpp)
448.3 group_async_work_group_copy.cpp.o (../tests/group/group_async_work_group_copy.cpp)

I will take a look at the top two, I remember discussing long build times in the context of the handler_copy tests a while back. We'll probably have to seriously cull the set of tested combinations in non-full-conformance mode.

Edit: Still, we need to pick up discussions about improving the CI infrastructure early next year.

[alexbatashev]

Compilation with dpcpp takes 148 minutes (~2.5 hours). Is this acceptable for pre-commit?

No, not really. These are the top 5 culprits:

1971.5 accessor_api_buffer_core.cpp.o (../tests/accessor/accessor_api_buffer_core.cpp)
1709.1 handler_copy.cpp.o (../tests/handler/handler_copy.cpp)
532.1 math_builtin_integer.cpp.o (tests/math_builtin_api/math_builtin_integer.cpp)
469.2 nd_item_async_work_group_copy.cpp.o (../tests/nd_item/nd_item_async_work_group_copy.cpp)
448.3 group_async_work_group_copy.cpp.o (../tests/group/group_async_work_group_copy.cpp)

This is mostly spent in the kernel function instantiation. One hack, that you can employ here is passing -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ which halves compile time for at least the first test out of these five. I guess, the other tests will see a similar speedup.

[bader]

Compilation with dpcpp takes 148 minutes (~2.5 hours). Is this acceptable for pre-commit?

No, not really. These are the top 5 culprits:

1971.5 accessor_api_buffer_core.cpp.o (../tests/accessor/accessor_api_buffer_core.cpp)
1709.1 handler_copy.cpp.o (../tests/handler/handler_copy.cpp)
532.1 math_builtin_integer.cpp.o (tests/math_builtin_api/math_builtin_integer.cpp)
469.2 nd_item_async_work_group_copy.cpp.o (../tests/nd_item/nd_item_async_work_group_copy.cpp)
448.3 group_async_work_group_copy.cpp.o (../tests/group/group_async_work_group_copy.cpp)

This is mostly spent in the kernel function instantiation. One hack, that you can employ here is passing -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ which halves compile time for at least the first test out of these five. I guess, the other tests will see a similar speedup.

Applied in d752a19.
I think DPC++ should improve default compilation time.

[alexbatashev]

@psalz we now build docker images right in the intel/llvm repository, you can re-use those if you want. See https://github.com/intel/llvm/pkgs/container/llvm%2Fsycl_ubuntu2004_nightly

latest tag comes with all Intel drivers pre-installed, and no-drivers tag bundles compiler toolchain only. Also, there're tags for each commit hash, that was used to build the image.

[bader]

This is mostly spent in the kernel function instantiation. One hack, that you can employ here is passing -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ which halves compile time for at least the first test out of these five. I guess, the other tests will see a similar speedup.

It looks like the right way to disable this feature is to use -fsycl-disable-range-rounding.

[bader]

@alexbatashev, -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ saved ~10 minutes overall.

[alexbatashev]

@bader handler is not the only template-heavy thing we have. Here's before disabling range rounding:

1971.5 accessor_api_buffer_core.cpp.o (../tests/accessor/accessor_api_buffer_core.cpp)
1709.1 handler_copy.cpp.o (../tests/handler/handler_copy.cpp)
532.1 math_builtin_integer.cpp.o (tests/math_builtin_api/math_builtin_integer.cpp)
469.2 nd_item_async_work_group_copy.cpp.o (../tests/nd_item/nd_item_async_work_group_copy.cpp)
448.3 group_async_work_group_copy.cpp.o (../tests/group/group_async_work_group_copy.cpp)
371.3 vector_swizzles_sycl__half.cpp.o (tests/vector_swizzles/vector_swizzles_sycl__half.cpp)
367.7 vector_swizzles_sycl__byte.cpp.o (tests/vector_swizzles/vector_swizzles_sycl__byte.cpp)
367.3 vector_swizzles_int8_t.cpp.o (tests/vector_swizzles/vector_swizzles_int8_t.cpp)
364.0 vector_swizzles_int32_t.cpp.o (tests/vector_swizzles/vector_swizzles_int32_t.cpp)
361.1 vector_swizzles_char.cpp.o (tests/vector_swizzles/vector_swizzles_char.cpp)
360.2 vector_swizzles_double.cpp.o (tests/vector_swizzles/vector_swizzles_double.cpp)
360.1 vector_swizzles_int.cpp.o (tests/vector_swizzles/vector_swizzles_int.cpp)
344.6 vector_swizzles_float.cpp.o (tests/vector_swizzles/vector_swizzles_float.cpp)
311.3 accessor_api_buffer_fp16.cpp.o (../tests/accessor/accessor_api_buffer_fp16.cpp)
310.7 accessor_api_buffer_fp64.cpp.o (../tests/accessor/accessor_api_buffer_fp64.cpp)
291.1 accessor_constructors_buffer.cpp.o (../tests/accessor/accessor_constructors_buffer.cpp)

And here's the after:

1456.0 accessor_api_buffer_core.cpp.o (../tests/accessor/accessor_api_buffer_core.cpp)
1445.6 handler_copy.cpp.o (../tests/handler/handler_copy.cpp)
507.2 math_builtin_integer.cpp.o (tests/math_builtin_api/math_builtin_integer.cpp)
468.7 nd_item_async_work_group_copy.cpp.o (../tests/nd_item/nd_item_async_work_group_copy.cpp)
445.2 group_async_work_group_copy.cpp.o (../tests/group/group_async_work_group_copy.cpp)
362.5 vector_swizzles_int32_t.cpp.o (tests/vector_swizzles/vector_swizzles_int32_t.cpp)
362.0 vector_swizzles_sycl__half.cpp.o (tests/vector_swizzles/vector_swizzles_sycl__half.cpp)
359.3 vector_swizzles_double.cpp.o (tests/vector_swizzles/vector_swizzles_double.cpp)
358.5 vector_swizzles_sycl__byte.cpp.o (tests/vector_swizzles/vector_swizzles_sycl__byte.cpp)
355.3 vector_swizzles_int.cpp.o (tests/vector_swizzles/vector_swizzles_int.cpp)
354.9 vector_swizzles_int8_t.cpp.o (tests/vector_swizzles/vector_swizzles_int8_t.cpp)
353.6 vector_swizzles_float.cpp.o (tests/vector_swizzles/vector_swizzles_float.cpp)
342.1 vector_swizzles_char.cpp.o (tests/vector_swizzles/vector_swizzles_char.cpp)
280.5 accessor_constructors_buffer.cpp.o (../tests/accessor/accessor_constructors_buffer.cpp)
276.9 accessor_api_buffer_fp16.cpp.o (../tests/accessor/accessor_api_buffer_fp16.cpp)
272.3 accessor_api_buffer_fp64.cpp.o (../tests/accessor/accessor_api_buffer_fp64.cpp)

intel/llvm#5127 should also give a small improvement for the first two tests. The trick with disabling this-free-function queries on host device I showed you earlier in the morning also gives nice speedup to those two tests (I'm yet to test all three tricks applied to see the final difference). The next big thing is sycl::vec class, which probably also has some template abuse. There's clearly a lot of room for improvement in terms of compile times.

Edit: applying all three hacks drives compile times for accessor_api_buffer_core.cpp from ~9 minutes to ~130 seconds. Considering, that GitHub hosted runners compile this file for over 30 minutes, 4.5 time sounds like a big potential improvement here.

[psalz]

Thanks guys for looking into this.

applying all three hacks drives compile times for accessor_api_buffer_core.cpp from ~9 minutes to ~130 seconds. Considering, that GitHub hosted runners compile this file for over 30 minutes, 4.5 time sounds like a big potential improvement here.

So we already have the -D__SYCL_DISABLE_PARALLEL_FOR_RANGE_ROUNDING__ setting, and I can bump the DPC++ version to include intel/llvm#5127. What would be the third hack then? Is this something we can apply here?

@psalz we now build docker images right in the intel/llvm repository, you can re-use those if you want. See https://github.com/intel/llvm/pkgs/container/llvm%2Fsycl_ubuntu2004_nightly

Good to know, thanks. Since we don't really need any special setup within the containers we can probably switch to using this. I'll make sure to look into it when I have some time.

[alexbatashev]

What would be the third hack then? Is this something we can apply here?

This requires some work on the clang frontend side. I'll ping you once I have more to share.

[psalz]

Closed in favor of #241, let's move discussions on compilation time to a dedicated PR.