Add shared memory support for L0 path

[kurapov-peter]

We need to add shared memory support for Intel GPU using L0 backend.

1. Create an equivalent of the Nvidia GpuSharedMemoryTest.cpp for Intel GPU via L0
2. Activate SM on Backend / Execution Engine.
3. Check hardcoded parameters for each GPU architecture

[lmontigny]

Created branch: 116-add-shared-memory-support-for-l0-path

[lmontigny]

Moved to pakurapo/smem-2 branch, need to change the is_l0 flag to true in several place to avoid CPU fallback (and use GPU via L0). Otherwise reported address space casting issue

Clarify strange address space casting:
%dest_slot_adr_0 = addrspacecast i8 addrspace(3)* %2 to i32 addrspace(21875)*

[lmontigny]

Large address space issue related to the traits object?
solved using

    if (slot_bytes == sizeof(int32_t)) {
      // return traits.smemPointerType(llvm::Type::getInt32Ty(context));
      return llvm::Type::getInt32PtrTy(context, /*address_space=*/3)

generate:
%dest_slot_adr_0 = bitcast i8 addrspace(3)* %2 to i32 addrspace(3)*

Current issue:

/GpuSharedMemoryTestIntel.INFO
 2023-05-03T04:39:46.972076 I 1772468 0 0 L0Mgr.cpp:75 Discovered 1 driver(s) for L0 platform.
 2023-05-03T04:39:47.003210 I 1772468 0 0 L0Mgr.cpp:221 L0 module build log: error: IGC SPIRV Consumer does    not support the following opcode: 65534
 ^@

[lmontigny]

L0Mgr issue, it seems to be an invalid opcode as 65535 is not reserved by any vendor:
https://github.com/KhronosGroup/SPIRV-Headers/blob/main/include/spirv/spir-v.xml#L152

<ids type="opcode" start="6656" end="65535" comment="Opcode range reservable for future use by vendors"/>

Wrong data is read as an opcode earlier in the IGC SPIR-V process

[lmontigny]

Searching the invalid opcode, see spir-v disassembly

[lmontigny]

Invalid casting in the write_projection_int64 from RuntimeFunctions.cpp

/usr/bin/llvm-as-14: after.linking.before.insert_declaration.spirv.ll:1712:16: error: invalid cast opcode for cast from 'i8 addrspace(4)*' to 'i64*'
  %6 = bitcast i8 addrspace(4)* %0 to i64*, !ResultSetReductionCodegen.cpp !10

Need to create a custom method in genx.ll with the appropriate casting to addrspace(4)

[lmontigny]

Added write_projection_int64 in genx.ll with the casting to addrspacecast(4)
%6 = bitcast i8 addrspace(4)* %0 to i64 addrspace(4)*

write_projection_int32 is required in genx.ll as well.

Now same casting issue is appearing for agg_sum_double_skip_val, this function needs to be added in genx.ll as well. (with the casting addrspace(4)).

[lmontigny]

Fixed several function and added implementation in genx.ll

Need to fix this one:

/usr/bin/llvm-as-14: after.linking.spirv.ll:1978:13: error: '@reduce_from_smem_to_gmem' defined with type 'void (i64 addrspace(4)*, i64 addrspace(4)*, i32)*' but expected 'void (i64 addrspace(4)*, i64 addrspace(3)*, i32)*'
  call void @reduce_from_smem_to_gmem(i64 addrspace(4)* %output_buffer_ptr, i64 addrspace(3)* %smem_input_buffer_ptr, i32 14952)

[lmontigny]

Now on the shared_memory related issue:

$ /usr/bin/llvm-as-14 after.linking.spirv.ll
/usr/bin/llvm-as-14: after.linking.spirv.ll:1258:45: error: use of undefined value '@slm.buf.i64'
  %res = bitcast [1024 x i64] addrspace(3)* @slm.buf.i64 to i64 addrspace(3)*

Current genx.ll file:

; TODO(Petr): this should be dynamically sized depending on the arch, but at least 64KB
@slm.buf.i64 = internal local_unnamed_addr addrspace(3) global [1024 x i64] zeroinitializer, align 8

; https://registry.khronos.org/SPIR-V/specs/unified1/SPIRV.html#OpControlBarrier
declare void @__spirv_ControlBarrier(i32 %execution_scope, i32 %memory_scope, i32 %memory_semantics)

; L0 zeroes out the memory by default, so an empty implementation can work in simple cases
define i64 addrspace(3)* @init_shared_mem(i64 addrspace(4)* %agg_init_val, i32 noundef %groups_buffer_size) {
    %res = bitcast [1024 x i64] addrspace(3)* @slm.buf.i64 to i64 addrspace(3)*
    ret i64 addrspace(3)* null

[lmontigny]

Need to copy the global variables from genx.ll to the IR
Done following main...pakurapo/slm-copying-example

[lmontigny]

Initializer is not copied, need to add it separately

$ /usr/bin/llvm-as-14 after.linking.spirv.ll
/usr/bin/llvm-as-14: after.linking.spirv.ll:25:76: error: expected value token
@slm.buf.i64 = internal local_unnamed_addr addrspace(3) global [1024 x i64], align 8

Missing zeroinitializer at the end

[lmontigny]

Solved init issue using

 // Initialize shared memory buffer
  const auto slm_buffer = module_->getNamedGlobal("slm.buf.i64");
  CHECK(slm_buffer);
  llvm::ArrayType* ArrayTy_0 =
      llvm::ArrayType::get(llvm::IntegerType::get(module_->getContext(), 64), 1024);
  llvm::ConstantAggregateZero* const_array_2 =
      llvm::ConstantAggregateZero::get(ArrayTy_0);
  slm_buffer->setInitializer(const_array_2)

[lmontigny]

By running the test ./Tests/GpuSharedMemoryTestIntel:

hdk_log/./Tests/GpuSharedMemoryTestIntel =

 2023-05-24T05:52:15.323533 I 21622 0 0 L0Mgr.cpp:75 Discovered 1 driver(s) for L0 platform.
 2023-05-24T05:52:15.366145 I 21622 0 0 L0Mgr.cpp:221 L0 module build log: error: IGC SPIRV Consumer does not   support the following opcode: 65534

Need to rebase and/or clean the current branch with the Intel GPU tests