[SYCL] Optimize NDRDescT by removing sycl::range, sycl::id and padding

[DBDuncan]

sycl::range and sycl::id perform validity checks every time setting them. Use std::array instead as dimensions should already be valid. In addition, remove explicitly padding dimensions smaller than 3 and get number of dimensions from template argument instead of function argument.

[aelovikov-intel]

Can we remove the throw from these SYCL classes instead?

[aelovikov-intel]

I know it wasn't you who messed up the sorting here, but please either remove unnecessary changes or clean it up with a preceding PR to just restore the sorting.

[Pennycook]

This looks really weird to me. I know you didn't introduce this SetNumWorkGroups thing, but it's odd.

From a quick glance, it looks like:

• We always store the range passed to the constructor, but potentially in different places.
• NumWorkGroups is only used by hierarchical parallelism (parallel_for_work_group, specifically).

Could we flip the logic here, so that the constructor always unconditionally stores into GlobalSize, and the parallel_for_work_group code knows to read GlobalSize instead of NumWorkGroups?

[DBDuncan]

I have been looking into this. There is some confusing stuff going on in the unmodified version of handler.cpp:

llvm/sycl/source/handler.cpp

Lines 1037 to 1079 in 22c8d2f

	case kernel_param_kind_t::kind_stream: {
	// Stream contains several accessors inside.
	stream *S = static_cast<stream *>(Ptr);
	detail::AccessorBaseHost *GBufBase =
	static_cast<detail::AccessorBaseHost *>(&S->GlobalBuf);
	detail::AccessorImplPtr GBufImpl = detail::getSyclObjImpl(*GBufBase);
	detail::Requirement *GBufReq = GBufImpl.get();
	addArgsForGlobalAccessor(
	GBufReq, Index, IndexShift, Size, IsKernelCreatedFromSource,
	impl->MNDRDesc.GlobalSize.size(), impl->MArgs, IsESIMD);
	++IndexShift;
	detail::AccessorBaseHost *GOffsetBase =
	static_cast<detail::AccessorBaseHost *>(&S->GlobalOffset);
	detail::AccessorImplPtr GOfssetImpl = detail::getSyclObjImpl(*GOffsetBase);
	detail::Requirement *GOffsetReq = GOfssetImpl.get();
	addArgsForGlobalAccessor(
	GOffsetReq, Index, IndexShift, Size, IsKernelCreatedFromSource,
	impl->MNDRDesc.GlobalSize.size(), impl->MArgs, IsESIMD);
	++IndexShift;
	detail::AccessorBaseHost *GFlushBase =
	static_cast<detail::AccessorBaseHost *>(&S->GlobalFlushBuf);
	detail::AccessorImplPtr GFlushImpl = detail::getSyclObjImpl(*GFlushBase);
	detail::Requirement *GFlushReq = GFlushImpl.get();
	size_t GlobalSize = impl->MNDRDesc.GlobalSize.size();
	// If work group size wasn't set explicitly then it must be recieved
	// from kernel attribute or set to default values.
	// For now we can't get this attribute here.
	// So we just suppose that WG size is always default for stream.
	// TODO adjust MNDRDesc when device image contains kernel's attribute
	if (GlobalSize == 0) {
	// Suppose that work group size is 1 for every dimension
	GlobalSize = impl->MNDRDesc.NumWorkGroups.size();
	}
	addArgsForGlobalAccessor(GFlushReq, Index, IndexShift, Size,
	IsKernelCreatedFromSource, GlobalSize, impl->MArgs,
	IsESIMD);
	++IndexShift;
	addArg(kernel_param_kind_t::kind_std_layout, &S->FlushBufferSize,
	sizeof(S->FlushBufferSize), Index + IndexShift);
	break;

It looks like to me that it is expected that it can be the case the GlobalSize is zero. This means that addArgsForGlobalAccessor are called with the size argument set to zero and then later on GlobalSize is checked if it is zero and it it is it is set to the size of NumWorkGroups.

I am not quite sure how this is working. I would have expected there to be issues passing size of zero to addArgsForGlobalAccessor.

[DBDuncan]

The only other place that NumWorkGroups is used is in adjectNDRangePerKernel in sycl/source/detail/scheduler/commands.cpp

[DBDuncan]

I think it is because AccImpl->PerWI just happens to be false so GlobalSize is not used. Not sure what that variable is meant to signal.

[Pennycook]

Can you try using a single variable, and see if anything breaks?

If the other parts of the code are checking for zero GlobalSize and then reading NumWorkGroups instead, it seems like you could just remove the check and read GlobalSize unconditionally,

[DBDuncan]

There are a number of tests that depend on extra LocalSize dimensions higher than Dims_ being set to zero or one depending on whether LocalSizes[I] is zero or not respectively. RequiredWGSize.NoRequiredSize and RequiredWGSize.HasRequiredSize always fail if extra LocalSize dimensions are always set to 1 and various tests such as work_group_size_prop.cpp and six others fail if extra LocalSize dimensions are always set to zero. This preserves the old behaviour.

[DBDuncan]

It seems strange to me that this was introduced in the first place. It really should not matter what the value of dimensions higher than Dims_ are and should just be ignored. But now a number of tests depend on this behaviour.

[DBDuncan]

Working on improving performance with this PR has lead me to hopefully make it more explicit in what values the members of NDRDescT are set to when constructed. There are a lot of tests that rely on 1 or 0 to be set for dimensions that are not used and other behaviour I have had to preserve to get the CI to pass.

Not looking to fix the root cause of that here but to at least hopefully make it more obvious what is going on.