Implement SPV_INTEL_hw_thread_queries extension

The extension adds BuiltIn decorations to query maximum number
of H/W threads per a sub-device (tile), a sub-device ID and a H/W
thread ID within the sub-device. These instructions combined can be
used to calculate a unique H/W thread on which the code is being launched.

Spec: #4747
Some info about Intel multi-tile cards:
https://github.com/intel/llvm/blob/sycl/sycl/doc/MultiTileCardWithLevelZero.md

Signed-off-by: Dmitry Sidorov <dmitry.sidorov@intel.com>

Original commit:
KhronosGroup/SPIRV-LLVM-Translator@ddb1c7a

Author: MrSidims

llvm-spirv/include/LLVMSPIRVExtensions.inc

@@ -46,3 +46,4 @@ EXT(SPV_INTEL_runtime_aligned)
 EXT(SPV_INTEL_arithmetic_fence)
 EXT(SPV_INTEL_bfloat16_conversion)
 EXT(SPV_INTEL_joint_matrix)
+EXT(SPV_INTEL_hw_thread_queries)

llvm-spirv/lib/SPIRV/SPIRVToOCL.cpp

@@ -92,6 +92,12 @@ void SPIRVToOCLBase::visitCallInst(CallInst &CI) {
                     << "BuiltinKind = " << BuiltinKind << '\n');
 
   if (BuiltinKind != SPIRVBuiltinVariableKind::BuiltInMax) {
+    if (static_cast<uint32_t>(BuiltinKind) >=
+            internal::BuiltInSubDeviceIDINTEL &&
+        static_cast<uint32_t>(BuiltinKind) <=
+            internal::BuiltInMaxHWThreadIDPerSubDeviceINTEL)
+      return;
+
     visitCallSPIRVBuiltin(&CI, BuiltinKind);
     return;
   }

llvm-spirv/lib/SPIRV/SPIRVWriter.cpp

@@ -1539,6 +1539,19 @@ LLVMToSPIRVBase::transValueWithoutDecoration(Value *V, SPIRVBasicBlock *BB,
     spv::BuiltIn Builtin = spv::BuiltInPosition;
     if (!GV->hasName() || !getSPIRVBuiltin(GV->getName().str(), Builtin))
       return BVar;
+    if (static_cast<uint32_t>(Builtin) >= internal::BuiltInSubDeviceIDINTEL &&
+        static_cast<uint32_t>(Builtin) <=
+            internal::BuiltInMaxHWThreadIDPerSubDeviceINTEL) {
+      if (!BM->isAllowedToUseExtension(
+              ExtensionID::SPV_INTEL_hw_thread_queries)) {
+        std::string ErrorStr = "Intel HW thread queries must be enabled by "
+                               "SPV_INTEL_hw_thread_queries extension.\n"
+                               "LLVM value that is being translated:\n";
+        getErrorLog().checkError(false, SPIRVEC_InvalidModule, V, ErrorStr);
+      }
+      BM->addExtension(ExtensionID::SPV_INTEL_hw_thread_queries);
+    }
+
     BVar->setBuiltin(Builtin);
     return BVar;
   }

llvm-spirv/lib/SPIRV/libSPIRV/SPIRVEnum.h

@@ -491,6 +491,12 @@ template <> inline void SPIRVMap<BuiltIn, SPIRVCapVec>::init() {
   ADD_VEC_INIT(BuiltInSubgroupLtMask, {CapabilityGroupNonUniformBallot});
   ADD_VEC_INIT(BuiltInVertexIndex, {CapabilityShader});
   ADD_VEC_INIT(BuiltInInstanceIndex, {CapabilityShader});
+  ADD_VEC_INIT(internal::BuiltInSubDeviceIDINTEL,
+               {internal::CapabilityHWThreadQueryINTEL});
+  ADD_VEC_INIT(internal::BuiltInHWThreadIDINTEL,
+               {internal::CapabilityHWThreadQueryINTEL});
+  ADD_VEC_INIT(internal::BuiltInMaxHWThreadIDPerSubDeviceINTEL,
+               {internal::CapabilityHWThreadQueryINTEL});
 }
 
 template <> inline void SPIRVMap<MemorySemanticsMask, SPIRVCapVec>::init() {

llvm-spirv/lib/SPIRV/libSPIRV/SPIRVIsValidEnum.h

@@ -174,7 +174,7 @@ inline bool isValid(spv::FunctionParameterAttribute V) {
 }
 
 inline bool isValid(spv::BuiltIn V) {
-  switch (V) {
+  switch (static_cast<uint32_t>(V)) {
   case BuiltInPosition:
   case BuiltInPointSize:
   case BuiltInClipDistance:
@@ -273,6 +273,9 @@ inline bool isValid(spv::BuiltIn V) {
   case BuiltInSMCountNV:
   case BuiltInWarpIDNV:
   case BuiltInSMIDNV:
+  case internal::BuiltInSubDeviceIDINTEL:
+  case internal::BuiltInHWThreadIDINTEL:
+  case internal::BuiltInMaxHWThreadIDPerSubDeviceINTEL:
     return true;
   default:
     return false;

llvm-spirv/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h

@@ -306,6 +306,10 @@ template <> inline void SPIRVMap<BuiltIn, std::string>::init() {
   add(BuiltInWarpIDNV, "BuiltInWarpIDNV");
   add(BuiltInSMIDNV, "BuiltInSMIDNV");
   add(BuiltInMax, "BuiltInMax");
+  add(internal::BuiltInSubDeviceIDINTEL, "BuiltInSubDeviceIDINTEL");
+  add(internal::BuiltInHWThreadIDINTEL, "BuiltInHWThreadIDINTEL");
+  add(internal::BuiltInMaxHWThreadIDPerSubDeviceINTEL,
+      "BuiltInMaxHWThreadIDPerSubDeviceINTEL");
 }
 SPIRV_DEF_NAMEMAP(BuiltIn, SPIRVBuiltInNameMap)
 
@@ -582,6 +586,7 @@ template <> inline void SPIRVMap<Capability, std::string>::init() {
   add(internal::CapabilityFPArithmeticFenceINTEL, "FPArithmeticFenceINTEL");
   add(internal::CapabilityBfloat16ConversionINTEL, "Bfloat16ConversionINTEL");
   add(internal::CapabilityJointMatrixINTEL, "JointMatrixINTEL");
+  add(internal::CapabilityHWThreadQueryINTEL, "HWThreadQueryINTEL");
 }
 SPIRV_DEF_NAMEMAP(Capability, SPIRVCapabilityNameMap)
 

llvm-spirv/lib/SPIRV/libSPIRV/spirv_internal.hpp

@@ -76,6 +76,7 @@ enum InternalCapability {
   ICapTokenTypeINTEL = 6112,
   ICapBfloat16ConversionINTEL = 6115,
   ICapabilityJointMatrixINTEL = 6118,
+  ICapabilityHWThreadQueryINTEL = 6134,
   ICapFPArithmeticFenceINTEL = 6144
 };
 
@@ -98,12 +99,23 @@ constexpr LinkageType LinkageTypeInternal =
 
 enum InternalJointMatrixLayout { RowMajor, ColumnMajor, PackedA, PackedB };
 
+enum InternalBuiltIn {
+  IBuiltInSubDeviceIDINTEL = 6135,
+  IBuiltInHWThreadIDINTEL = 6136,
+  IBuiltInMaxHWThreadIDPerSubDeviceINTEL = 6137
+};
+
 #define _SPIRV_OP(x, y) constexpr x x##y = static_cast<x>(I##x##y);
 _SPIRV_OP(Capability, JointMatrixINTEL)
 _SPIRV_OP(Op, TypeJointMatrixINTEL)
 _SPIRV_OP(Op, JointMatrixLoadINTEL)
 _SPIRV_OP(Op, JointMatrixStoreINTEL)
 _SPIRV_OP(Op, JointMatrixMadINTEL)
+
+_SPIRV_OP(Capability, HWThreadQueryINTEL)
+_SPIRV_OP(BuiltIn, SubDeviceIDINTEL)
+_SPIRV_OP(BuiltIn, HWThreadIDINTEL)
+_SPIRV_OP(BuiltIn, MaxHWThreadIDPerSubDeviceINTEL)
 #undef _SPIRV_OP
 
 constexpr Op OpForward = static_cast<Op>(IOpForward);

llvm-spirv/test/transcoding/SPV_INTEL_hw_thread_queries/intel_hw_thread_queries_function_call.ll

@@ -0,0 +1,41 @@
+; Test for SPV_INTEL_hw_thread_queries feature
+; SPIR-V friendly LLVM IR to SPIR-V translation and vice versa
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc --spirv-ext=+SPV_INTEL_hw_thread_queries -o %t.spv
+; RUN: llvm-spirv %t.spv -to-text -o %t.spt
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: llvm-spirv -r %t.spv -o %t.rev.bc
+; RUN: llvm-dis < %t.rev.bc | FileCheck %s --check-prefix=CHECK-LLVM
+; RUN: llvm-spirv -r --spirv-target-env=SPV-IR %t.spv -o %t.rev.bc
+; RUN: llvm-dis < %t.rev.bc | FileCheck %s --check-prefix=CHECK-LLVM
+
+; CHECK-SPIRV: Capability HWThreadQueryINTEL
+; CHECK-SPIRV: Extension "SPV_INTEL_hw_thread_queries"
+; CHECK-SPIRV: Decorate [[#Id1:]] BuiltIn 6135
+; CHECK-SPIRV: Decorate [[#Id2:]] BuiltIn 6136
+; CHECK-SPIRV: Decorate [[#Id3:]] BuiltIn 6137
+; CHECK-SPIRV: Variable [[#]] [[#Id1]]
+; CHECK-SPIRV: Variable [[#]] [[#Id2]]
+; CHECK-SPIRV: Variable [[#]] [[#Id3]]
+
+target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir-unknown-unknown"
+
+; Function Attrs: nounwind readnone
+define spir_kernel void @foo() {
+entry:
+  %0 = call spir_func i32 @_Z31__spirv_BuiltInSubDeviceIDINTELv() #1
+  %1 = call spir_func i32 @_Z30__spirv_BuiltInHWThreadIDINTELv() #1
+  %2 = call spir_func i32 @_Z45__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTELv() #1
+  ; CHECK-LLVM: call spir_func i32 @_Z31__spirv_BuiltInSubDeviceIDINTELv() #1
+  ; CHECK-LLVM: call spir_func i32 @_Z30__spirv_BuiltInHWThreadIDINTELv() #1
+  ; CHECK-LLVM: call spir_func i32 @_Z45__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTELv() #1
+  ret void
+}
+
+declare dso_local spir_func i32 @_Z31__spirv_BuiltInSubDeviceIDINTELv()
+
+declare dso_local spir_func i32 @_Z30__spirv_BuiltInHWThreadIDINTELv()
+
+declare dso_local spir_func i32 @_Z45__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTELv()

llvm-spirv/test/transcoding/SPV_INTEL_hw_thread_queries/intel_hw_thread_queries_load_from_global.ll

@@ -0,0 +1,39 @@
+; Test for SPV_INTEL_hw_thread_queries feature
+; SPIR-V friendly LLVM IR to SPIR-V translation and vice versa
+; RUN: llvm-as %s -o %t.bc
+; RUN: llvm-spirv %t.bc --spirv-ext=+SPV_INTEL_hw_thread_queries -o %t.spv
+; RUN: llvm-spirv %t.spv -to-text -o %t.spt
+; RUN: FileCheck < %t.spt %s --check-prefix=CHECK-SPIRV
+
+; RUN: llvm-spirv -r %t.spv -o %t.rev.bc
+; RUN: llvm-dis < %t.rev.bc | FileCheck %s --check-prefix=CHECK-LLVM
+; RUN: llvm-spirv -r --spirv-target-env=SPV-IR %t.spv -o %t.rev.bc
+; RUN: llvm-dis < %t.rev.bc | FileCheck %s --check-prefix=CHECK-LLVM
+
+; CHECK-SPIRV: Capability HWThreadQueryINTEL
+; CHECK-SPIRV: Extension "SPV_INTEL_hw_thread_queries"
+; CHECK-SPIRV: Decorate [[#Id1:]] BuiltIn 6135
+; CHECK-SPIRV: Decorate [[#Id2:]] BuiltIn 6136
+; CHECK-SPIRV: Decorate [[#Id3:]] BuiltIn 6137
+; CHECK-SPIRV: Variable [[#]] [[#Id1]]
+; CHECK-SPIRV: Variable [[#]] [[#Id2]]
+; CHECK-SPIRV: Variable [[#]] [[#Id3]]
+
+target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir-unknown-unknown"
+
+@__spirv_BuiltInSubDeviceIDINTEL = external addrspace(1) global i32
+@__spirv_BuiltInHWThreadIDINTEL = external addrspace(1) global i32
+@__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTEL = external addrspace(1) global i32
+
+; Function Attrs: nounwind readnone
+define spir_kernel void @foo() {
+entry:
+  %0 = load i32, i32 addrspace(4)* addrspacecast (i32 addrspace(1)* @__spirv_BuiltInSubDeviceIDINTEL to i32 addrspace(4)*), align 4
+  %1 = load i32, i32 addrspace(4)* addrspacecast (i32 addrspace(1)* @__spirv_BuiltInHWThreadIDINTEL to i32 addrspace(4)*), align 4
+  %2 = load i32, i32 addrspace(4)* addrspacecast (i32 addrspace(1)* @__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTEL to i32 addrspace(4)*), align 4
+  ; CHECK-LLVM: call spir_func i32 @_Z31__spirv_BuiltInSubDeviceIDINTELv() #1
+  ; CHECK-LLVM: call spir_func i32 @_Z30__spirv_BuiltInHWThreadIDINTELv() #1
+  ; CHECK-LLVM: call spir_func i32 @_Z45__spirv_BuiltInMaxHWThreadIDPerSubDeviceINTELv() #1
+  ret void
+}

llvm-spirv/test/transcoding/SPV_INTEL_hw_thread_queries/negative_intel_hw_thread_queries.ll

@@ -0,0 +1,21 @@
+; Negative test for SPV_INTEL_hw_thread_queries feature
+; Check for errors in case if the extension is not enabled, but the appropriate
+; SPV-IR patter is found
+; RUN: llvm-as %s -o %t.bc
+; RUN: not llvm-spirv %t.bc -o %t.spv 2>&1 | FileCheck %s
+
+; CHECK: InvalidModule: Invalid SPIR-V module: Intel HW thread queries must be enabled by SPV_INTEL_hw_thread_queries extension.
+; CHECK: LLVM value that is being translated:
+; CHECK: @__spirv_BuiltInSubDeviceIDINTEL = external addrspace(1) global i32
+
+target datalayout = "e-p:32:32-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024"
+target triple = "spir-unknown-unknown"
+
+@__spirv_BuiltInSubDeviceIDINTEL = external addrspace(1) global i32
+
+; Function Attrs: nounwind readnone
+define spir_kernel void @foo() {
+entry:
+  %0 = load i32, i32 addrspace(4)* addrspacecast (i32 addrspace(1)* @__spirv_BuiltInSubDeviceIDINTEL to i32 addrspace(4)*), align 4
+  ret void
+}