|
| 1 | +From c1cefd2c2b233132904b8f31ae7ca68965ecc431 Mon Sep 17 00:00:00 2001 |
| 2 | +From: haonanya < [email protected]> |
| 3 | +Date: Mon, 28 Feb 2022 22:27:54 +0800 |
| 4 | +Subject: [PATCH] Add support for split barriers extension |
| 5 | + SPV_INTEL_split_barrier |
| 6 | + |
| 7 | +Signed-off-by: haonanya < [email protected]> |
| 8 | +--- |
| 9 | + include/LLVMSPIRVExtensions.inc | 1 + |
| 10 | + lib/SPIRV/OCLToSPIRV.cpp | 38 +++++++++++++++++++++++++++ |
| 11 | + lib/SPIRV/OCLUtil.cpp | 7 ++++- |
| 12 | + lib/SPIRV/OCLUtil.h | 1 + |
| 13 | + lib/SPIRV/SPIRVReader.cpp | 3 ++- |
| 14 | + lib/SPIRV/SPIRVToOCL.cpp | 4 +++ |
| 15 | + lib/SPIRV/SPIRVToOCL.h | 13 +++++++++ |
| 16 | + lib/SPIRV/SPIRVToOCL12.cpp | 13 +++++++++ |
| 17 | + lib/SPIRV/SPIRVToOCL20.cpp | 22 ++++++++++++++++ |
| 18 | + lib/SPIRV/libSPIRV/SPIRVInstruction.h | 19 +++++++++++++- |
| 19 | + lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h | 1 + |
| 20 | + lib/SPIRV/libSPIRV/SPIRVOpCode.h | 5 ++++ |
| 21 | + lib/SPIRV/libSPIRV/SPIRVOpCodeEnum.h | 2 ++ |
| 22 | + 13 files changed, 126 insertions(+), 3 deletions(-) |
| 23 | + |
| 24 | +diff --git a/include/LLVMSPIRVExtensions.inc b/include/LLVMSPIRVExtensions.inc |
| 25 | +index 3faf76e2..47b5c268 100644 |
| 26 | +--- a/include/LLVMSPIRVExtensions.inc |
| 27 | ++++ b/include/LLVMSPIRVExtensions.inc |
| 28 | +@@ -36,3 +36,4 @@ EXT(SPV_INTEL_variable_length_array) |
| 29 | + EXT(SPV_INTEL_fp_fast_math_mode) |
| 30 | + EXT(SPV_INTEL_optnone) |
| 31 | + EXT(SPV_INTEL_memory_access_aliasing) |
| 32 | ++EXT(SPV_INTEL_split_barrier) |
| 33 | +diff --git a/lib/SPIRV/OCLToSPIRV.cpp b/lib/SPIRV/OCLToSPIRV.cpp |
| 34 | +index 1a476fbf..71484429 100644 |
| 35 | +--- a/lib/SPIRV/OCLToSPIRV.cpp |
| 36 | ++++ b/lib/SPIRV/OCLToSPIRV.cpp |
| 37 | +@@ -260,6 +260,9 @@ public: |
| 38 | + void visitSubgroupAVCBuiltinCallWithSampler(CallInst *CI, |
| 39 | + StringRef DemangledName); |
| 40 | + |
| 41 | ++ /// For cl_intel_split_work_group_barrier built-ins: |
| 42 | ++ void visitCallSplitBarrierINTEL(CallInst *CI, StringRef DemangledName); |
| 43 | ++ |
| 44 | + void visitCallLdexp(CallInst *CI, StringRef MangledName, |
| 45 | + StringRef DemangledName); |
| 46 | + |
| 47 | +@@ -560,6 +563,10 @@ void OCLToSPIRVBase::visitCallInst(CallInst &CI) { |
| 48 | + visitSubgroupImageMediaBlockINTEL(&CI, DemangledName); |
| 49 | + return; |
| 50 | + } |
| 51 | ++ if (DemangledName.find(kOCLBuiltinName::SplitBarrierINTELPrefix) == 0) { |
| 52 | ++ visitCallSplitBarrierINTEL(&CI, DemangledName); |
| 53 | ++ return; |
| 54 | ++ } |
| 55 | + // Handle 'cl_intel_device_side_avc_motion_estimation' extension built-ins |
| 56 | + if (DemangledName.find(kOCLSubgroupsAVCIntel::Prefix) == 0 || |
| 57 | + // Workaround for a bug in the extension specification |
| 58 | +@@ -1890,6 +1897,37 @@ void OCLToSPIRVBase::visitSubgroupAVCBuiltinCallWithSampler( |
| 59 | + &Attrs); |
| 60 | + } |
| 61 | + |
| 62 | ++void OCLToSPIRVBase::visitCallSplitBarrierINTEL(CallInst *CI, |
| 63 | ++ StringRef DemangledName) { |
| 64 | ++ auto Lit = getBarrierLiterals(CI); |
| 65 | ++ AttributeList Attrs = CI->getCalledFunction()->getAttributes(); |
| 66 | ++ Op OpCode = |
| 67 | ++ StringSwitch<Op>(DemangledName) |
| 68 | ++ .Case("intel_work_group_barrier_arrive", OpControlBarrierArriveINTEL) |
| 69 | ++ .Case("intel_work_group_barrier_wait", OpControlBarrierWaitINTEL) |
| 70 | ++ .Default(OpNop); |
| 71 | ++ |
| 72 | ++ mutateCallInstSPIRV( |
| 73 | ++ M, CI, |
| 74 | ++ [=](CallInst *, std::vector<Value *> &Args) { |
| 75 | ++ Args.resize(3); |
| 76 | ++ // Execution scope |
| 77 | ++ Args[0] = addInt32(map<Scope>(std::get<2>(Lit))); |
| 78 | ++ // Memory scope |
| 79 | ++ Args[1] = addInt32(map<Scope>(std::get<1>(Lit))); |
| 80 | ++ // Memory semantics |
| 81 | ++ // OpControlBarrierArriveINTEL -> Release, |
| 82 | ++ // OpControlBarrierWaitINTEL -> Acquire |
| 83 | ++ unsigned MemFenceFlag = std::get<0>(Lit); |
| 84 | ++ OCLMemOrderKind MemOrder = OpCode == OpControlBarrierArriveINTEL |
| 85 | ++ ? OCLMO_release |
| 86 | ++ : OCLMO_acquire; |
| 87 | ++ Args[2] = addInt32(mapOCLMemSemanticToSPIRV(MemFenceFlag, MemOrder)); |
| 88 | ++ return getSPIRVFuncName(OpCode); |
| 89 | ++ }, |
| 90 | ++ &Attrs); |
| 91 | ++} |
| 92 | ++ |
| 93 | + void OCLToSPIRVBase::visitCallLdexp(CallInst *CI, StringRef MangledName, |
| 94 | + StringRef DemangledName) { |
| 95 | + auto Args = getArguments(CI); |
| 96 | +diff --git a/lib/SPIRV/OCLUtil.cpp b/lib/SPIRV/OCLUtil.cpp |
| 97 | +index 35277022..01083173 100644 |
| 98 | +--- a/lib/SPIRV/OCLUtil.cpp |
| 99 | ++++ b/lib/SPIRV/OCLUtil.cpp |
| 100 | +@@ -421,6 +421,9 @@ template <> void SPIRVMap<std::string, Op, SPIRVInstruction>::init() { |
| 101 | + _SPIRV_OP(bitfield_extract_signed, BitFieldSExtract) |
| 102 | + _SPIRV_OP(bitfield_extract_unsigned, BitFieldUExtract) |
| 103 | + _SPIRV_OP(bit_reverse, BitReverse) |
| 104 | ++ // cl_khr_split_work_group_barrier |
| 105 | ++ _SPIRV_OP(intel_work_group_barrier_arrive, ControlBarrierArriveINTEL) |
| 106 | ++ _SPIRV_OP(intel_work_group_barrier_wait, ControlBarrierWaitINTEL) |
| 107 | + #undef _SPIRV_OP |
| 108 | + } |
| 109 | + |
| 110 | +@@ -1026,7 +1029,9 @@ public: |
| 111 | + } else if (NameRef.contains("barrier")) { |
| 112 | + addUnsignedArg(0); |
| 113 | + if (NameRef.equals("work_group_barrier") || |
| 114 | +- NameRef.equals("sub_group_barrier")) |
| 115 | ++ NameRef.equals("sub_group_barrier") || |
| 116 | ++ NameRef.equals("intel_work_group_barrier_arrive") || |
| 117 | ++ NameRef.equals("intel_work_group_barrier_wait")) |
| 118 | + setEnumArg(1, SPIR::PRIMITIVE_MEMORY_SCOPE); |
| 119 | + } else if (NameRef.startswith("atomic_work_item_fence")) { |
| 120 | + addUnsignedArg(0); |
| 121 | +diff --git a/lib/SPIRV/OCLUtil.h b/lib/SPIRV/OCLUtil.h |
| 122 | +index 1d9e4f93..0f5c53ef 100644 |
| 123 | +--- a/lib/SPIRV/OCLUtil.h |
| 124 | ++++ b/lib/SPIRV/OCLUtil.h |
| 125 | +@@ -304,6 +304,7 @@ const static char SubgroupBlockWriteINTELPrefix[] = |
| 126 | + "intel_sub_group_block_write"; |
| 127 | + const static char SubgroupImageMediaBlockINTELPrefix[] = |
| 128 | + "intel_sub_group_media_block"; |
| 129 | ++const static char SplitBarrierINTELPrefix[] = "intel_work_group_barrier_"; |
| 130 | + const static char LDEXP[] = "ldexp"; |
| 131 | + } // namespace kOCLBuiltinName |
| 132 | + |
| 133 | +diff --git a/lib/SPIRV/SPIRVReader.cpp b/lib/SPIRV/SPIRVReader.cpp |
| 134 | +index 6ef88cb8..51a3e563 100644 |
| 135 | +--- a/lib/SPIRV/SPIRVReader.cpp |
| 136 | ++++ b/lib/SPIRV/SPIRVReader.cpp |
| 137 | +@@ -2877,7 +2877,8 @@ Instruction *SPIRVToLLVM::transBuiltinFromInst(const std::string &FuncName, |
| 138 | + if (isFuncNoUnwind()) |
| 139 | + Func->addFnAttr(Attribute::NoUnwind); |
| 140 | + auto OC = BI->getOpCode(); |
| 141 | +- if (isGroupOpCode(OC) || isIntelSubgroupOpCode(OC)) |
| 142 | ++ if (isGroupOpCode(OC) || isIntelSubgroupOpCode(OC) || |
| 143 | ++ isSplitBarrierINTELOpCode(OC)) |
| 144 | + Func->addFnAttr(Attribute::Convergent); |
| 145 | + } |
| 146 | + auto Call = |
| 147 | +diff --git a/lib/SPIRV/SPIRVToOCL.cpp b/lib/SPIRV/SPIRVToOCL.cpp |
| 148 | +index 2e6b73f5..dafbfe0b 100644 |
| 149 | +--- a/lib/SPIRV/SPIRVToOCL.cpp |
| 150 | ++++ b/lib/SPIRV/SPIRVToOCL.cpp |
| 151 | +@@ -107,6 +107,10 @@ void SPIRVToOCLBase::visitCallInst(CallInst &CI) { |
| 152 | + if (OC == OpControlBarrier) { |
| 153 | + visitCallSPIRVControlBarrier(&CI); |
| 154 | + } |
| 155 | ++ if (isSplitBarrierINTELOpCode(OC)) { |
| 156 | ++ visitCallSPIRVSplitBarrierINTEL(&CI, OC); |
| 157 | ++ return; |
| 158 | ++ } |
| 159 | + if (isAtomicOpCode(OC)) { |
| 160 | + visitCallSPIRVAtomicBuiltin(&CI, OC); |
| 161 | + return; |
| 162 | +diff --git a/lib/SPIRV/SPIRVToOCL.h b/lib/SPIRV/SPIRVToOCL.h |
| 163 | +index 380459fa..af64f645 100644 |
| 164 | +--- a/lib/SPIRV/SPIRVToOCL.h |
| 165 | ++++ b/lib/SPIRV/SPIRVToOCL.h |
| 166 | +@@ -215,6 +215,11 @@ public: |
| 167 | + /// - OCL1.2: barrier |
| 168 | + virtual void visitCallSPIRVControlBarrier(CallInst *CI) = 0; |
| 169 | + |
| 170 | ++ /// Transform split __spirv_ControlBarrier barrier to: |
| 171 | ++ /// - OCL2.0: overload with a memory_scope argument |
| 172 | ++ /// - OCL1.2: overload with no memory_scope argument |
| 173 | ++ virtual void visitCallSPIRVSplitBarrierINTEL(CallInst *CI, Op OC) = 0; |
| 174 | ++ |
| 175 | + /// Transform __spirv_EnqueueKernel to __enqueue_kernel |
| 176 | + virtual void visitCallSPIRVEnqueueKernel(CallInst *CI, Op OC) = 0; |
| 177 | + |
| 178 | +@@ -283,6 +288,10 @@ public: |
| 179 | + /// barrier(flag(sema)) |
| 180 | + void visitCallSPIRVControlBarrier(CallInst *CI) override; |
| 181 | + |
| 182 | ++ /// Transform split __spirv_ControlBarrier barrier to overloads without a |
| 183 | ++ /// memory_scope argument. |
| 184 | ++ void visitCallSPIRVSplitBarrierINTEL(CallInst *CI, Op OC) override; |
| 185 | ++ |
| 186 | + /// Transform __spirv_OpAtomic functions. It firstly conduct generic |
| 187 | + /// mutations for all builtins and then mutate some of them seperately |
| 188 | + Instruction *visitCallSPIRVAtomicBuiltin(CallInst *CI, Op OC) override; |
| 189 | +@@ -372,6 +381,10 @@ public: |
| 190 | + /// sub_group_barrier(flag(sema), map(memScope)) |
| 191 | + void visitCallSPIRVControlBarrier(CallInst *CI) override; |
| 192 | + |
| 193 | ++ /// Transform split __spirv_ControlBarrier barrier to overloads with a |
| 194 | ++ /// memory_scope argument. |
| 195 | ++ void visitCallSPIRVSplitBarrierINTEL(CallInst *CI, Op OC) override; |
| 196 | ++ |
| 197 | + /// Transform __spirv_Atomic* to atomic_*. |
| 198 | + /// __spirv_Atomic*(atomic_op, scope, sema, ops, ...) => |
| 199 | + /// atomic_*(generic atomic_op, ops, ..., order(sema), map(scope)) |
| 200 | +diff --git a/lib/SPIRV/SPIRVToOCL12.cpp b/lib/SPIRV/SPIRVToOCL12.cpp |
| 201 | +index f34564b0..8f510d56 100644 |
| 202 | +--- a/lib/SPIRV/SPIRVToOCL12.cpp |
| 203 | ++++ b/lib/SPIRV/SPIRVToOCL12.cpp |
| 204 | +@@ -102,6 +102,19 @@ void SPIRVToOCL12Base::visitCallSPIRVControlBarrier(CallInst *CI) { |
| 205 | + &Attrs); |
| 206 | + } |
| 207 | + |
| 208 | ++void SPIRVToOCL12Base::visitCallSPIRVSplitBarrierINTEL(CallInst *CI, Op OC) { |
| 209 | ++ AttributeList Attrs = CI->getCalledFunction()->getAttributes(); |
| 210 | ++ mutateCallInstOCL( |
| 211 | ++ M, CI, |
| 212 | ++ [=](CallInst *, std::vector<Value *> &Args) { |
| 213 | ++ Value *MemFenceFlags = |
| 214 | ++ SPIRV::transSPIRVMemorySemanticsIntoOCLMemFenceFlags(Args[2], CI); |
| 215 | ++ Args.assign(1, MemFenceFlags); |
| 216 | ++ return OCLSPIRVBuiltinMap::rmap(OC); |
| 217 | ++ }, |
| 218 | ++ &Attrs); |
| 219 | ++} |
| 220 | ++ |
| 221 | + Instruction *SPIRVToOCL12Base::visitCallSPIRVAtomicIncDec(CallInst *CI, Op OC) { |
| 222 | + AttributeList Attrs = CI->getCalledFunction()->getAttributes(); |
| 223 | + return mutateCallInstOCL( |
| 224 | +diff --git a/lib/SPIRV/SPIRVToOCL20.cpp b/lib/SPIRV/SPIRVToOCL20.cpp |
| 225 | +index c9bb2829..3186c0ea 100644 |
| 226 | +--- a/lib/SPIRV/SPIRVToOCL20.cpp |
| 227 | ++++ b/lib/SPIRV/SPIRVToOCL20.cpp |
| 228 | +@@ -121,6 +121,28 @@ void SPIRVToOCL20Base::visitCallSPIRVControlBarrier(CallInst *CI) { |
| 229 | + &Attrs); |
| 230 | + } |
| 231 | + |
| 232 | ++void SPIRVToOCL20Base::visitCallSPIRVSplitBarrierINTEL(CallInst *CI, Op OC) { |
| 233 | ++ AttributeList Attrs = CI->getCalledFunction()->getAttributes(); |
| 234 | ++ mutateCallInstOCL( |
| 235 | ++ M, CI, |
| 236 | ++ [=](CallInst *, std::vector<Value *> &Args) { |
| 237 | ++ auto GetArg = [=](unsigned I) { |
| 238 | ++ return cast<ConstantInt>(Args[I])->getZExtValue(); |
| 239 | ++ }; |
| 240 | ++ Value *MemScope = |
| 241 | ++ getInt32(M, rmap<OCLScopeKind>(static_cast<Scope>(GetArg(1)))); |
| 242 | ++ Value *MemFenceFlags = |
| 243 | ++ SPIRV::transSPIRVMemorySemanticsIntoOCLMemFenceFlags(Args[2], CI); |
| 244 | ++ |
| 245 | ++ Args.resize(2); |
| 246 | ++ Args[0] = MemFenceFlags; |
| 247 | ++ Args[1] = MemScope; |
| 248 | ++ |
| 249 | ++ return OCLSPIRVBuiltinMap::rmap(OC); |
| 250 | ++ }, |
| 251 | ++ &Attrs); |
| 252 | ++} |
| 253 | ++ |
| 254 | + std::string SPIRVToOCL20Base::mapFPAtomicName(Op OC) { |
| 255 | + assert(isFPAtomicOpCode(OC) && "Not intended to handle other opcodes than " |
| 256 | + "AtomicF{Add/Min/Max}EXT!"); |
| 257 | +diff --git a/lib/SPIRV/libSPIRV/SPIRVInstruction.h b/lib/SPIRV/libSPIRV/SPIRVInstruction.h |
| 258 | +index 2b4830c4..2e9bb5f6 100644 |
| 259 | +--- a/lib/SPIRV/libSPIRV/SPIRVInstruction.h |
| 260 | ++++ b/lib/SPIRV/libSPIRV/SPIRVInstruction.h |
| 261 | +@@ -3177,6 +3177,23 @@ _SPIRV_OP(VariableLengthArray, true, 4) |
| 262 | + _SPIRV_OP(SaveMemory, true, 3) |
| 263 | + _SPIRV_OP(RestoreMemory, false, 2) |
| 264 | + #undef _SPIRV_OP |
| 265 | +-} // namespace SPIRV |
| 266 | + |
| 267 | ++class SPIRVSplitBarrierINTELBase : public SPIRVInstTemplateBase { |
| 268 | ++protected: |
| 269 | ++ SPIRVCapVec getRequiredCapability() const override { |
| 270 | ++ return getVec(CapabilitySplitBarrierINTEL); |
| 271 | ++ } |
| 272 | ++ |
| 273 | ++ llvm::Optional<ExtensionID> getRequiredExtension() const override { |
| 274 | ++ return ExtensionID::SPV_INTEL_split_barrier; |
| 275 | ++ } |
| 276 | ++}; |
| 277 | ++ |
| 278 | ++#define _SPIRV_OP(x, ...) \ |
| 279 | ++ typedef SPIRVInstTemplate<SPIRVSplitBarrierINTELBase, Op##x, __VA_ARGS__> \ |
| 280 | ++ SPIRV##x; |
| 281 | ++_SPIRV_OP(ControlBarrierArriveINTEL, false, 4) |
| 282 | ++_SPIRV_OP(ControlBarrierWaitINTEL, false, 4) |
| 283 | ++#undef _SPIRV_OP |
| 284 | ++} // namespace SPIRV |
| 285 | + #endif // SPIRV_LIBSPIRV_SPIRVINSTRUCTION_H |
| 286 | +diff --git a/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h b/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h |
| 287 | +index 9d6f4070..b7a5f8b7 100644 |
| 288 | +--- a/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h |
| 289 | ++++ b/lib/SPIRV/libSPIRV/SPIRVNameMapEnum.h |
| 290 | +@@ -565,6 +565,7 @@ template <> inline void SPIRVMap<Capability, std::string>::init() { |
| 291 | + add(internal::CapabilityOptNoneINTEL, "OptNoneINTEL"); |
| 292 | + add(internal::CapabilityMemoryAccessAliasingINTEL, |
| 293 | + "MemoryAccessAliasingINTEL"); |
| 294 | ++ add(CapabilitySplitBarrierINTEL, "SplitBarrierINTEL"); |
| 295 | + |
| 296 | + add(CapabilityMax, "Max"); |
| 297 | + } |
| 298 | +diff --git a/lib/SPIRV/libSPIRV/SPIRVOpCode.h b/lib/SPIRV/libSPIRV/SPIRVOpCode.h |
| 299 | +index f400f87e..18b446be 100644 |
| 300 | +--- a/lib/SPIRV/libSPIRV/SPIRVOpCode.h |
| 301 | ++++ b/lib/SPIRV/libSPIRV/SPIRVOpCode.h |
| 302 | +@@ -243,6 +243,11 @@ inline bool isEventOpCode(Op OpCode) { |
| 303 | + return OpRetainEvent <= OpCode && OpCode <= OpCaptureEventProfilingInfo; |
| 304 | + } |
| 305 | + |
| 306 | ++inline bool isSplitBarrierINTELOpCode(Op OpCode) { |
| 307 | ++ return OpCode == OpControlBarrierArriveINTEL || |
| 308 | ++ OpCode == OpControlBarrierWaitINTEL; |
| 309 | ++} |
| 310 | ++ |
| 311 | + } // namespace SPIRV |
| 312 | + |
| 313 | + #endif // SPIRV_LIBSPIRV_SPIRVOPCODE_H |
| 314 | +diff --git a/lib/SPIRV/libSPIRV/SPIRVOpCodeEnum.h b/lib/SPIRV/libSPIRV/SPIRVOpCodeEnum.h |
| 315 | +index 3e841009..aa282e3e 100644 |
| 316 | +--- a/lib/SPIRV/libSPIRV/SPIRVOpCodeEnum.h |
| 317 | ++++ b/lib/SPIRV/libSPIRV/SPIRVOpCodeEnum.h |
| 318 | +@@ -540,3 +540,5 @@ _SPIRV_OP(WritePipeBlockingINTEL, 5947) |
| 319 | + _SPIRV_OP(FPGARegINTEL, 5949) |
| 320 | + _SPIRV_OP(AtomicFAddEXT, 6035) |
| 321 | + _SPIRV_OP(TypeBufferSurfaceINTEL, 6086) |
| 322 | ++_SPIRV_OP(ControlBarrierArriveINTEL, 6142) |
| 323 | ++_SPIRV_OP(ControlBarrierWaitINTEL, 6143) |
| 324 | +-- |
| 325 | +2.17.1 |
| 326 | + |
0 commit comments