| 1 | //===- EmulateAtomics.cpp - Emulate unsupported AMDGPU atomics ------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #include "mlir/Dialect/AMDGPU/Transforms/Passes.h" |
| 10 | |
| 11 | #include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h" |
| 12 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 13 | #include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h" |
| 14 | #include "mlir/Dialect/LLVMIR/ROCDLTargetInfo.h" |
| 15 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 16 | #include "mlir/IR/BuiltinAttributes.h" |
| 17 | #include "mlir/IR/TypeUtilities.h" |
| 18 | #include "mlir/Transforms/DialectConversion.h" |
| 19 | |
| 20 | namespace mlir::amdgpu { |
| 21 | #define GEN_PASS_DEF_AMDGPUEMULATEATOMICSPASS |
| 22 | #include "mlir/Dialect/AMDGPU/Transforms/Passes.h.inc" |
| 23 | } // namespace mlir::amdgpu |
| 24 | |
| 25 | using namespace mlir; |
| 26 | using namespace mlir::amdgpu; |
| 27 | |
| 28 | namespace { |
| 29 | struct AmdgpuEmulateAtomicsPass |
| 30 | : public amdgpu::impl::AmdgpuEmulateAtomicsPassBase< |
| 31 | AmdgpuEmulateAtomicsPass> { |
| 32 | using AmdgpuEmulateAtomicsPassBase< |
| 33 | AmdgpuEmulateAtomicsPass>::AmdgpuEmulateAtomicsPassBase; |
| 34 | void runOnOperation() override; |
| 35 | }; |
| 36 | |
| 37 | template <typename AtomicOp, typename ArithOp> |
| 38 | struct RawBufferAtomicByCasPattern : public OpConversionPattern<AtomicOp> { |
| 39 | using OpConversionPattern<AtomicOp>::OpConversionPattern; |
| 40 | using Adaptor = typename AtomicOp::Adaptor; |
| 41 | |
| 42 | LogicalResult |
| 43 | matchAndRewrite(AtomicOp atomicOp, Adaptor adaptor, |
| 44 | ConversionPatternRewriter &rewriter) const override; |
| 45 | }; |
| 46 | } // namespace |
| 47 | |
| 48 | namespace { |
| 49 | enum class DataArgAction : unsigned char { |
| 50 | Duplicate, |
| 51 | Drop, |
| 52 | }; |
| 53 | } // namespace |
| 54 | |
| 55 | // Fix up the fact that, when we're migrating from a general bugffer atomic |
| 56 | // to a load or to a CAS, the number of openrands, and thus the number of |
| 57 | // entries needed in operandSegmentSizes, needs to change. We use this method |
| 58 | // because we'd like to preserve unknown attributes on the atomic instead of |
| 59 | // discarding them. |
| 60 | static void patchOperandSegmentSizes(ArrayRef<NamedAttribute> attrs, |
| 61 | SmallVectorImpl<NamedAttribute> &newAttrs, |
| 62 | DataArgAction action) { |
| 63 | newAttrs.reserve(N: attrs.size()); |
| 64 | for (NamedAttribute attr : attrs) { |
| 65 | if (attr.getName().getValue() != "operandSegmentSizes" ) { |
| 66 | newAttrs.push_back(Elt: attr); |
| 67 | continue; |
| 68 | } |
| 69 | auto segmentAttr = cast<DenseI32ArrayAttr>(Val: attr.getValue()); |
| 70 | MLIRContext *context = segmentAttr.getContext(); |
| 71 | DenseI32ArrayAttr newSegments; |
| 72 | switch (action) { |
| 73 | case DataArgAction::Drop: |
| 74 | newSegments = DenseI32ArrayAttr::get( |
| 75 | context, content: segmentAttr.asArrayRef().drop_front()); |
| 76 | break; |
| 77 | case DataArgAction::Duplicate: { |
| 78 | SmallVector<int32_t> newVals; |
| 79 | ArrayRef<int32_t> oldVals = segmentAttr.asArrayRef(); |
| 80 | newVals.push_back(Elt: oldVals[0]); |
| 81 | newVals.append(in_start: oldVals.begin(), in_end: oldVals.end()); |
| 82 | newSegments = DenseI32ArrayAttr::get(context, content: newVals); |
| 83 | break; |
| 84 | } |
| 85 | } |
| 86 | newAttrs.push_back(Elt: NamedAttribute(attr.getName(), newSegments)); |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | template <typename OpTy> |
| 91 | static typename OpTy::Properties |
| 92 | getPropertiesFromAttrs(OpBuilder &builder, Location loc, |
| 93 | ArrayRef<NamedAttribute> attrs) { |
| 94 | typename OpTy::Properties properties{}; |
| 95 | OpTy::populateDefaultProperties( |
| 96 | OperationName(OpTy::getOperationName(), builder.getContext()), |
| 97 | properties); |
| 98 | LogicalResult result = |
| 99 | OpTy::setPropertiesFromAttr(properties, builder.getDictionaryAttr(value: attrs), |
| 100 | [&]() { return emitError(loc); }); |
| 101 | assert(succeeded(result) && "failed to convert operation properties" ); |
| 102 | (void)result; |
| 103 | return properties; |
| 104 | } |
| 105 | |
| 106 | // A helper function to flatten a vector value to a scalar containing its bits, |
| 107 | // returning the value itself if othetwise. |
| 108 | static Value flattenVecToBits(ConversionPatternRewriter &rewriter, Location loc, |
| 109 | Value val) { |
| 110 | auto vectorType = dyn_cast<VectorType>(Val: val.getType()); |
| 111 | if (!vectorType) |
| 112 | return val; |
| 113 | |
| 114 | int64_t bitwidth = |
| 115 | vectorType.getElementTypeBitWidth() * vectorType.getNumElements(); |
| 116 | Type allBitsType = rewriter.getIntegerType(width: bitwidth); |
| 117 | auto allBitsVecType = VectorType::get(shape: {1}, elementType: allBitsType); |
| 118 | Value bitcast = vector::BitCastOp::create(builder&: rewriter, location: loc, result: allBitsVecType, source: val); |
| 119 | Value scalar = vector::ExtractOp::create(builder&: rewriter, location: loc, source: bitcast, position: 0); |
| 120 | return scalar; |
| 121 | } |
| 122 | |
| 123 | template <typename AtomicOp, typename ArithOp> |
| 124 | LogicalResult RawBufferAtomicByCasPattern<AtomicOp, ArithOp>::matchAndRewrite( |
| 125 | AtomicOp atomicOp, Adaptor adaptor, |
| 126 | ConversionPatternRewriter &rewriter) const { |
| 127 | Location loc = atomicOp.getLoc(); |
| 128 | |
| 129 | NamedAttrList origProperties; |
| 130 | atomicOp->getName().walkInherentAttrs(atomicOp, |
| 131 | [&](StringRef name, Attribute &attr) { |
| 132 | origProperties.append(name, attr); |
| 133 | }); |
| 134 | ArrayRef<NamedAttribute> discardableAttrs = |
| 135 | atomicOp->getDiscardableAttrDictionary().getValue(); |
| 136 | ValueRange operands = adaptor.getOperands(); |
| 137 | Value data = operands.take_front()[0]; |
| 138 | ValueRange invariantArgs = operands.drop_front(); |
| 139 | Type dataType = data.getType(); |
| 140 | |
| 141 | SmallVector<NamedAttribute> loadAttrs; |
| 142 | patchOperandSegmentSizes(attrs: origProperties, newAttrs&: loadAttrs, action: DataArgAction::Drop); |
| 143 | auto loadProperties = |
| 144 | getPropertiesFromAttrs<RawBufferLoadOp>(builder&: rewriter, loc, attrs: loadAttrs); |
| 145 | Value initialLoad = |
| 146 | RawBufferLoadOp::create(builder&: rewriter, location: loc, resultTypes: TypeRange{dataType}, operands: invariantArgs, |
| 147 | properties: loadProperties, discardableAttributes: discardableAttrs); |
| 148 | Block *currentBlock = rewriter.getInsertionBlock(); |
| 149 | Block *afterAtomic = |
| 150 | rewriter.splitBlock(block: currentBlock, before: rewriter.getInsertionPoint()); |
| 151 | afterAtomic->addArgument(type: dataType, loc); |
| 152 | Block *loopBlock = rewriter.createBlock(insertBefore: afterAtomic, argTypes: {dataType}, locs: {loc}); |
| 153 | |
| 154 | rewriter.setInsertionPointToEnd(currentBlock); |
| 155 | cf::BranchOp::create(builder&: rewriter, location: loc, dest: loopBlock, destOperands: initialLoad); |
| 156 | |
| 157 | rewriter.setInsertionPointToEnd(loopBlock); |
| 158 | Value prevLoad = loopBlock->getArgument(i: 0); |
| 159 | Value operated = ArithOp::create(rewriter, loc, data, prevLoad); |
| 160 | dataType = operated.getType(); |
| 161 | |
| 162 | SmallVector<NamedAttribute> cmpswapAttrs; |
| 163 | patchOperandSegmentSizes(attrs: origProperties, newAttrs&: cmpswapAttrs, |
| 164 | action: DataArgAction::Duplicate); |
| 165 | SmallVector<Value> cmpswapArgs = {operated, prevLoad}; |
| 166 | cmpswapArgs.append(in_start: invariantArgs.begin(), in_end: invariantArgs.end()); |
| 167 | auto cmpswapProperties = getPropertiesFromAttrs<RawBufferAtomicCmpswapOp>( |
| 168 | builder&: rewriter, loc, attrs: cmpswapAttrs); |
| 169 | Value atomicRes = RawBufferAtomicCmpswapOp::create( |
| 170 | builder&: rewriter, location: loc, resultTypes: TypeRange{dataType}, operands: cmpswapArgs, properties: cmpswapProperties, |
| 171 | discardableAttributes: discardableAttrs); |
| 172 | |
| 173 | // We care about exact bitwise equality here, so do some bitcasts. |
| 174 | // These will fold away during lowering to the ROCDL dialect, where |
| 175 | // an int->float bitcast is introduced to account for the fact that cmpswap |
| 176 | // only takes integer arguments. |
| 177 | |
| 178 | Value prevLoadForCompare = flattenVecToBits(rewriter, loc, val: prevLoad); |
| 179 | Value atomicResForCompare = flattenVecToBits(rewriter, loc, val: atomicRes); |
| 180 | if (auto floatDataTy = dyn_cast<FloatType>(Val&: dataType)) { |
| 181 | Type equivInt = rewriter.getIntegerType(width: floatDataTy.getWidth()); |
| 182 | prevLoadForCompare = |
| 183 | arith::BitcastOp::create(builder&: rewriter, location: loc, out: equivInt, in: prevLoad); |
| 184 | atomicResForCompare = |
| 185 | arith::BitcastOp::create(builder&: rewriter, location: loc, out: equivInt, in: atomicRes); |
| 186 | } |
| 187 | Value canLeave = |
| 188 | arith::CmpIOp::create(builder&: rewriter, location: loc, predicate: arith::CmpIPredicate::eq, |
| 189 | lhs: atomicResForCompare, rhs: prevLoadForCompare); |
| 190 | cf::CondBranchOp::create(builder&: rewriter, location: loc, condition: canLeave, trueDest: afterAtomic, |
| 191 | trueOperands: ValueRange{prevLoad}, falseDest: loopBlock, falseOperands: atomicRes); |
| 192 | rewriter.replaceOp(atomicOp, ValueRange{afterAtomic->getArgument(i: 0)}); |
| 193 | return success(); |
| 194 | } |
| 195 | |
| 196 | /// Returns whether \p op can be lowered to a native buffer atomic fadd on |
| 197 | /// \p target. |
| 198 | static bool isFaddNativelySupported(const ROCDL::TargetInfo &target, |
| 199 | RawBufferAtomicFaddOp op) { |
| 200 | namespace AMDGPU = ::llvm::AMDGPU; |
| 201 | |
| 202 | // A target with only the no-return form can still perform the atomic, it |
| 203 | // just cannot report the old value, so it suffices when the result is dead. |
| 204 | bool hasFadd = target.has(feature: AMDGPU::FEAT_ATOMIC_FADD_RTN_INSTS) || |
| 205 | (target.has(feature: AMDGPU::FEAT_ATOMIC_FADD_NO_RTN_INSTS) && |
| 206 | op.getOldValue().use_empty()); |
| 207 | if (!hasFadd) |
| 208 | return false; |
| 209 | |
| 210 | // The packed 16-bit forms are separate instructions with their own features. |
| 211 | Type elemType = getElementTypeOrSelf(type: op.getValue().getType()); |
| 212 | if (isa<Float16Type>(Val: elemType)) |
| 213 | return target.has(feature: AMDGPU::FEAT_ATOMIC_BUFFER_GLOBAL_PK_ADD_F16_INSTS); |
| 214 | if (isa<BFloat16Type>(Val: elemType)) |
| 215 | return target.has(feature: AMDGPU::FEAT_ATOMIC_BUFFER_PK_ADD_BF16_INST); |
| 216 | return true; |
| 217 | } |
| 218 | |
| 219 | void mlir::amdgpu::populateAmdgpuEmulateAtomicsPatterns( |
| 220 | ConversionTarget &target, RewritePatternSet &patterns, |
| 221 | const ROCDL::TargetInfo &targetInfo, PatternBenefit benefit) { |
| 222 | namespace AMDGPU = ::llvm::AMDGPU; |
| 223 | |
| 224 | target.addDynamicallyLegalOp<RawBufferAtomicFaddOp>( |
| 225 | callback: [targetInfo](RawBufferAtomicFaddOp op) -> bool { |
| 226 | return isFaddNativelySupported(target: targetInfo, op); |
| 227 | }); |
| 228 | |
| 229 | // Floating-point min and max are only emulated on gfx9; later generations |
| 230 | // have them for every type this op accepts. f64 is the one gfx9 case that |
| 231 | // may be native. |
| 232 | if (targetInfo.isGeneration(major: 9)) { |
| 233 | target.addDynamicallyLegalOp<RawBufferAtomicFmaxOp>( |
| 234 | callback: [targetInfo](RawBufferAtomicFmaxOp op) -> bool { |
| 235 | return op.getValue().getType().isF64() && |
| 236 | targetInfo.has(feature: AMDGPU::FEAT_ATOMIC_FMIN_FMAX_GLOBAL_F64); |
| 237 | }); |
| 238 | } |
| 239 | patterns.add< |
| 240 | RawBufferAtomicByCasPattern<RawBufferAtomicFaddOp, arith::AddFOp>, |
| 241 | RawBufferAtomicByCasPattern<RawBufferAtomicFmaxOp, arith::MaximumFOp>, |
| 242 | RawBufferAtomicByCasPattern<RawBufferAtomicSmaxOp, arith::MaxSIOp>, |
| 243 | RawBufferAtomicByCasPattern<RawBufferAtomicUminOp, arith::MinUIOp>>( |
| 244 | arg: patterns.getContext(), args&: benefit); |
| 245 | } |
| 246 | |
| 247 | void AmdgpuEmulateAtomicsPass::runOnOperation() { |
| 248 | Operation *op = getOperation(); |
| 249 | FailureOr<ROCDL::TargetInfo> targetInfo = |
| 250 | ROCDL::TargetInfo::get(arch: ROCDL::resolveArchOption(arch, deprecatedAlias: chipset), |
| 251 | /*waveSize=*/0, emitError: [&] { return op->emitError(); }); |
| 252 | if (failed(Result: targetInfo)) |
| 253 | return signalPassFailure(); |
| 254 | |
| 255 | MLIRContext &ctx = getContext(); |
| 256 | ConversionTarget target(ctx); |
| 257 | RewritePatternSet patterns(&ctx); |
| 258 | target.markUnknownOpDynamicallyLegal( |
| 259 | fn: [](Operation *op) -> bool { return true; }); |
| 260 | |
| 261 | populateAmdgpuEmulateAtomicsPatterns(target, patterns, targetInfo: *targetInfo); |
| 262 | if (failed(Result: applyPartialConversion(op, target, patterns: std::move(patterns)))) |
| 263 | return signalPassFailure(); |
| 264 | } |
| 265 | |