[TritonGEN] Use OCL builtins for subgroup block read/write

[whitneywhtsang]

Use intel_sub_group_block_[read|write] defined in https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_subgroups.html, https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_subgroups_char.html, https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_subgroups_short.html, https://registry.khronos.org/OpenCL/extensions/intel/cl_intel_subgroups_long.html, and https://github.com/KhronosGroup/OpenCL-Docs/blob/main/extensions/cl_intel_subgroup_local_block_io.asciidoc.

[victor-eds]

We can emulate wider vectors with several calls and avoid relying on GenISA intrinsics:

%res = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<32xi32>

is equivalent to:

%0 = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<8xi32>
%ptr1 = llvm.getelementptr inbounds %ptr[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%1 = triton_gen.simdblockread %ptr1 : (!llvm.ptr) -> vector<8xi32>
%ptr2 = llvm.getelementptr inbounds %ptr1[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%2 = triton_gen.simdblockread %ptr2 : (!llvm.ptr) -> vector<8xi32>
%ptr3 = llvm.getelementptr inbounds %ptr2[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%3 = triton_gen.simdblockread %ptr3 : (!llvm.ptr) -> vector<8xi32>
%res = // Vector concatenation %0 %1 %2 %3

[whitneywhtsang]

We can emulate wider vectors with several calls and avoid relying on GenISA intrinsics:

%res = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<32xi32>

is equivalent to:

%0 = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<8xi32>
%ptr1 = llvm.getelementptr inbounds %ptr[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%1 = triton_gen.simdblockread %ptr1 : (!llvm.ptr) -> vector<8xi32>
%ptr2 = llvm.getelementptr inbounds %ptr1[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%2 = triton_gen.simdblockread %ptr2 : (!llvm.ptr) -> vector<8xi32>
%ptr3 = llvm.getelementptr inbounds %ptr2[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%3 = triton_gen.simdblockread %ptr3 : (!llvm.ptr) -> vector<8xi32>
%res = // Vector concatenation %0 %1 %2 %3

Right, I have a local change for that, would like to do that in a separate PR.

[quintinwang5]

We can emulate wider vectors with several calls and avoid relying on GenISA intrinsics:

%res = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<32xi32>

is equivalent to:

%0 = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<8xi32>
%ptr1 = llvm.getelementptr inbounds %ptr[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%1 = triton_gen.simdblockread %ptr1 : (!llvm.ptr) -> vector<8xi32>
%ptr2 = llvm.getelementptr inbounds %ptr1[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%2 = triton_gen.simdblockread %ptr2 : (!llvm.ptr) -> vector<8xi32>
%ptr3 = llvm.getelementptr inbounds %ptr2[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%3 = triton_gen.simdblockread %ptr3 : (!llvm.ptr) -> vector<8xi32>
%res = // Vector concatenation %0 %1 %2 %3

Right, I have a local change for that, would like to do that in a separate PR.

How about its performance?

[whitneywhtsang]

We can emulate wider vectors with several calls and avoid relying on GenISA intrinsics:

%res = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<32xi32>

is equivalent to:

%0 = triton_gen.simdblockread %ptr : (!llvm.ptr) -> vector<8xi32>
%ptr1 = llvm.getelementptr inbounds %ptr[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%1 = triton_gen.simdblockread %ptr1 : (!llvm.ptr) -> vector<8xi32>
%ptr2 = llvm.getelementptr inbounds %ptr1[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%2 = triton_gen.simdblockread %ptr2 : (!llvm.ptr) -> vector<8xi32>
%ptr3 = llvm.getelementptr inbounds %ptr2[16] : (!llvm.ptr) -> !llvm.ptr, vector<8xi32>
%3 = triton_gen.simdblockread %ptr3 : (!llvm.ptr) -> vector<8xi32>
%res = // Vector concatenation %0 %1 %2 %3

Right, I have a local change for that, would like to do that in a separate PR.

How about its performance?

I would like to test its performance, but the SLM path is not working at the moment, so I cannot get a baseline.

[Dewei-Wang-sh]

anywhere to document what the current ocl built-ins we can use in triton?

[victor-eds]

Small NIT. LGTM.

[whitneywhtsang]

anywhere to document what the current ocl built-ins we can use in triton?

We can use any ocl built-ins documented in any OpenCL C extensions, if we want to use an OpenCL C builtin that is not available in any OpenCL C extensions, then we need to send a request to IGC team.