1 //===- MemoryPromotion.cpp - Utilities for moving data across GPU memories ===//
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 // This file implements utilities that allow one to create IR moving the data
10 // across different levels of the GPU memory hierarchy.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "mlir/Dialect/GPU/MemoryPromotion.h"
15 #include "mlir/Dialect/Affine/LoopUtils.h"
16 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
17 #include "mlir/Dialect/GPU/GPUDialect.h"
18 #include "mlir/Dialect/MemRef/IR/MemRef.h"
19 #include "mlir/Dialect/SCF/SCF.h"
20 #include "mlir/IR/ImplicitLocOpBuilder.h"
21 #include "mlir/Pass/Pass.h"
22 
23 using namespace mlir;
24 using namespace mlir::gpu;
25 
26 /// Emits the (imperfect) loop nest performing the copy between "from" and "to"
27 /// values using the bounds derived from the "from" value. Emits at least
28 /// GPUDialect::getNumWorkgroupDimensions() loops, completing the nest with
29 /// single-iteration loops. Maps the innermost loops to thread dimensions, in
30 /// reverse order to enable access coalescing in the innermost loop.
31 static void insertCopyLoops(ImplicitLocOpBuilder &b, Value from, Value to) {
32   auto memRefType = from.getType().cast<MemRefType>();
33   auto rank = memRefType.getRank();
34 
35   SmallVector<Value, 4> lbs, ubs, steps;
36   Value zero = b.create<arith::ConstantIndexOp>(0);
37   Value one = b.create<arith::ConstantIndexOp>(1);
38 
39   // Make sure we have enough loops to use all thread dimensions, these trivial
40   // loops should be outermost and therefore inserted first.
41   if (rank < GPUDialect::getNumWorkgroupDimensions()) {
42     unsigned extraLoops = GPUDialect::getNumWorkgroupDimensions() - rank;
43     lbs.resize(extraLoops, zero);
44     ubs.resize(extraLoops, one);
45     steps.resize(extraLoops, one);
46   }
47 
48   // Add existing bounds.
49   lbs.append(rank, zero);
50   ubs.reserve(lbs.size());
51   steps.reserve(lbs.size());
52   for (auto idx = 0; idx < rank; ++idx) {
53     ubs.push_back(b.createOrFold<memref::DimOp>(
54         from, b.create<arith::ConstantIndexOp>(idx)));
55     steps.push_back(one);
56   }
57 
58   // Obtain thread identifiers and block sizes, necessary to map to them.
59   auto indexType = b.getIndexType();
60   SmallVector<Value, 3> threadIds, blockDims;
61   for (auto dim : {gpu::Dimension::x, gpu::Dimension::y, gpu::Dimension::z}) {
62     threadIds.push_back(b.create<gpu::ThreadIdOp>(indexType, dim));
63     blockDims.push_back(b.create<gpu::BlockDimOp>(indexType, dim));
64   }
65 
66   // Produce the loop nest with copies.
67   SmallVector<Value, 8> ivs(lbs.size());
68   mlir::scf::buildLoopNest(
69       b, b.getLoc(), lbs, ubs, steps,
70       [&](OpBuilder &b, Location loc, ValueRange loopIvs) {
71         ivs.assign(loopIvs.begin(), loopIvs.end());
72         auto activeIvs = llvm::makeArrayRef(ivs).take_back(rank);
73         Value loaded = b.create<memref::LoadOp>(loc, from, activeIvs);
74         b.create<memref::StoreOp>(loc, loaded, to, activeIvs);
75       });
76 
77   // Map the innermost loops to threads in reverse order.
78   for (const auto &en :
79        llvm::enumerate(llvm::reverse(llvm::makeArrayRef(ivs).take_back(
80            GPUDialect::getNumWorkgroupDimensions())))) {
81     Value v = en.value();
82     auto loop = cast<scf::ForOp>(v.getParentRegion()->getParentOp());
83     mapLoopToProcessorIds(loop, {threadIds[en.index()]},
84                           {blockDims[en.index()]});
85   }
86 }
87 
88 /// Emits the loop nests performing the copy to the designated location in the
89 /// beginning of the region, and from the designated location immediately before
90 /// the terminator of the first block of the region. The region is expected to
91 /// have one block. This boils down to the following structure
92 ///
93 ///   ^bb(...):
94 ///     <loop-bound-computation>
95 ///     for %arg0 = ... to ... step ... {
96 ///       ...
97 ///         for %argN = <thread-id-x> to ... step <block-dim-x> {
98 ///           %0 = load %from[%arg0, ..., %argN]
99 ///           store %0, %to[%arg0, ..., %argN]
100 ///         }
101 ///       ...
102 ///     }
103 ///     gpu.barrier
104 ///     <... original body ...>
105 ///     gpu.barrier
106 ///     for %arg0 = ... to ... step ... {
107 ///       ...
108 ///         for %argN = <thread-id-x> to ... step <block-dim-x> {
109 ///           %1 = load %to[%arg0, ..., %argN]
110 ///           store %1, %from[%arg0, ..., %argN]
111 ///         }
112 ///       ...
113 ///     }
114 ///
115 /// Inserts the barriers unconditionally since different threads may be copying
116 /// values and reading them. An analysis would be required to eliminate barriers
117 /// in case where value is only used by the thread that copies it. Both copies
118 /// are inserted unconditionally, an analysis would be required to only copy
119 /// live-in and live-out values when necessary. This copies the entire memref
120 /// pointed to by "from". In case a smaller block would be sufficient, the
121 /// caller can create a subview of the memref and promote it instead.
122 static void insertCopies(Region &region, Location loc, Value from, Value to) {
123   auto fromType = from.getType().cast<MemRefType>();
124   auto toType = to.getType().cast<MemRefType>();
125   (void)fromType;
126   (void)toType;
127   assert(fromType.getShape() == toType.getShape());
128   assert(fromType.getRank() != 0);
129   assert(llvm::hasSingleElement(region) &&
130          "unstructured control flow not supported");
131 
132   auto b = ImplicitLocOpBuilder::atBlockBegin(loc, &region.front());
133   insertCopyLoops(b, from, to);
134   b.create<gpu::BarrierOp>();
135 
136   b.setInsertionPoint(&region.front().back());
137   b.create<gpu::BarrierOp>();
138   insertCopyLoops(b, to, from);
139 }
140 
141 /// Promotes a function argument to workgroup memory in the given function. The
142 /// copies will be inserted in the beginning and in the end of the function.
143 void mlir::promoteToWorkgroupMemory(GPUFuncOp op, unsigned arg) {
144   Value value = op.getArgument(arg);
145   auto type = value.getType().dyn_cast<MemRefType>();
146   assert(type && type.hasStaticShape() && "can only promote memrefs");
147 
148   // Get the type of the buffer in the workgroup memory.
149   int workgroupMemoryAddressSpace = gpu::GPUDialect::getWorkgroupAddressSpace();
150   auto bufferType = MemRefType::get(type.getShape(), type.getElementType(), {},
151                                     workgroupMemoryAddressSpace);
152   Value attribution = op.addWorkgroupAttribution(bufferType, value.getLoc());
153 
154   // Replace the uses first since only the original uses are currently present.
155   // Then insert the copies.
156   value.replaceAllUsesWith(attribution);
157   insertCopies(op.getBody(), op.getLoc(), value, attribution);
158 }
159