cl_khr_subgroups support

[sjw36]

We would like to enable the cl_khr_subgroups extension for the following functions:

These functions will map to Builtin constants:

• get_sub_group_size - SubgroupSize
• get_max_sub_group_size - SubgroupMaxSize
• get_num_sub_groups - NumSubgroups
• get_sub_group_id - SubgroupId
• get_sub_group_local_id - SubgroupLocalInvocationId

And these functions will map to spv::Op's accordingly:

• sub_group_all - OpGroupNonUniformAll
• sub_group_any - OpGroupNonUniformAny
• sub_group_broadcast - OpGroupNonUniformBroadcast
• sub_group_barrier - OpControlBarrier

And the reduction functions as follows:

• sub_group_<group_operation>_add
  • INT: OpGroupNonUniformIAdd
  • FLOAT: OpGroupNonUniformFAdd
• sub_group_<group_operation>_min
  • INT: OpGroupNonUniformSMin
  • UINT: OpGroupNonUniformUMin
  • FLOAT: OpGroupNonUniformFMin
• sub_group_<group_operation>_max
  • INT: OpGroupNonUniformSMax
  • UINT: OpGroupNonUniformUMax
  • FLOAT: OpGroupNonUniformFMax With <group_operation> correspondingly:
• reduce - GroupOperationReduce
• scan_exclusive - GroupOperationExclusiveScan
• scan_inclusive - GroupOperationInclusiveScan

Adding the cl_khr_subgroups extension support requires OpenCL 2.0 or greater (clang front-end already supports with -cl-std=CL2.0) and also requires SPIR-V 1.3 or greater support. A command line switch will be required to specify the -spirv-std for the output version.

Setting -spirv-std=1.4 or greater requires updating the EntryPointInterface with all GlobalVariable IDs.

All tests should be run with -spirv-std=1.5 to test conformance.

[sjw36]

The remainder of cl_khr_subgroups functions that should be implemented for full support:

• get_enqueued_num_subroups
• sub_group_reserve_read_pipe
• sub_group_reserve_write_pipe
• sub_group_commit_read_pipe
• sub_group_commit_write_pipe
• get_kernel_sub_group_count_for_ndrange
• get_kernel_max_sub_group_size_for_ndrange

[alan-baker]

#610 added sub_group_barrier support.

[dokipen3d]

Are there any plans to introduce support for the intel extension 'cl_intel_subgroups' which includes subgroup functions sub_group_shuffle function etc...?