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/Transforms/MemoryPromotion.h"
15 
16 #include "mlir/Dialect/Affine/LoopUtils.h"
17 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
18 #include "mlir/Dialect/GPU/IR/GPUDialect.h"
19 #include "mlir/Dialect/MemRef/IR/MemRef.h"
20 #include "mlir/Dialect/SCF/IR/SCF.h"
21 #include "mlir/IR/ImplicitLocOpBuilder.h"
22 #include "mlir/Pass/Pass.h"
23 
24 using namespace mlir;
25 using namespace mlir::gpu;
26 
27 /// Emits the (imperfect) loop nest performing the copy between "from" and "to"
28 /// values using the bounds derived from the "from" value. Emits at least
29 /// GPUDialect::getNumWorkgroupDimensions() loops, completing the nest with
30 /// single-iteration loops. Maps the innermost loops to thread dimensions, in
31 /// reverse order to enable access coalescing in the innermost loop.
32 static void insertCopyLoops(ImplicitLocOpBuilder &b, Value from, Value to) {
33   auto memRefType = from.getType().cast<MemRefType>();
34   auto rank = memRefType.getRank();
35 
36   SmallVector<Value, 4> lbs, ubs, steps;
37   Value zero = b.create<arith::ConstantIndexOp>(0);
38   Value one = b.create<arith::ConstantIndexOp>(1);
39 
40   // Make sure we have enough loops to use all thread dimensions, these trivial
41   // loops should be outermost and therefore inserted first.
42   if (rank < GPUDialect::getNumWorkgroupDimensions()) {
43     unsigned extraLoops = GPUDialect::getNumWorkgroupDimensions() - rank;
44     lbs.resize(extraLoops, zero);
45     ubs.resize(extraLoops, one);
46     steps.resize(extraLoops, one);
47   }
48 
49   // Add existing bounds.
50   lbs.append(rank, zero);
51   ubs.reserve(lbs.size());
52   steps.reserve(lbs.size());
53   for (auto idx = 0; idx < rank; ++idx) {
54     ubs.push_back(b.createOrFold<memref::DimOp>(
55         from, b.create<arith::ConstantIndexOp>(idx)));
56     steps.push_back(one);
57   }
58 
59   // Obtain thread identifiers and block sizes, necessary to map to them.
60   auto indexType = b.getIndexType();
61   SmallVector<Value, 3> threadIds, blockDims;
62   for (auto dim : {gpu::Dimension::x, gpu::Dimension::y, gpu::Dimension::z}) {
63     threadIds.push_back(b.create<gpu::ThreadIdOp>(indexType, dim));
64     blockDims.push_back(b.create<gpu::BlockDimOp>(indexType, dim));
65   }
66 
67   // Produce the loop nest with copies.
68   SmallVector<Value, 8> ivs(lbs.size());
69   mlir::scf::buildLoopNest(
70       b, b.getLoc(), lbs, ubs, steps,
71       [&](OpBuilder &b, Location loc, ValueRange loopIvs) {
72         ivs.assign(loopIvs.begin(), loopIvs.end());
73         auto activeIvs = llvm::makeArrayRef(ivs).take_back(rank);
74         Value loaded = b.create<memref::LoadOp>(loc, from, activeIvs);
75         b.create<memref::StoreOp>(loc, loaded, to, activeIvs);
76       });
77 
78   // Map the innermost loops to threads in reverse order.
79   for (const auto &en :
80        llvm::enumerate(llvm::reverse(llvm::makeArrayRef(ivs).take_back(
81            GPUDialect::getNumWorkgroupDimensions())))) {
82     Value v = en.value();
83     auto loop = cast<scf::ForOp>(v.getParentRegion()->getParentOp());
84     mapLoopToProcessorIds(loop, {threadIds[en.index()]},
85                           {blockDims[en.index()]});
86   }
87 }
88 
89 /// Emits the loop nests performing the copy to the designated location in the
90 /// beginning of the region, and from the designated location immediately before
91 /// the terminator of the first block of the region. The region is expected to
92 /// have one block. This boils down to the following structure
93 ///
94 ///   ^bb(...):
95 ///     <loop-bound-computation>
96 ///     for %arg0 = ... to ... step ... {
97 ///       ...
98 ///         for %argN = <thread-id-x> to ... step <block-dim-x> {
99 ///           %0 = load %from[%arg0, ..., %argN]
100 ///           store %0, %to[%arg0, ..., %argN]
101 ///         }
102 ///       ...
103 ///     }
104 ///     gpu.barrier
105 ///     <... original body ...>
106 ///     gpu.barrier
107 ///     for %arg0 = ... to ... step ... {
108 ///       ...
109 ///         for %argN = <thread-id-x> to ... step <block-dim-x> {
110 ///           %1 = load %to[%arg0, ..., %argN]
111 ///           store %1, %from[%arg0, ..., %argN]
112 ///         }
113 ///       ...
114 ///     }
115 ///
116 /// Inserts the barriers unconditionally since different threads may be copying
117 /// values and reading them. An analysis would be required to eliminate barriers
118 /// in case where value is only used by the thread that copies it. Both copies
119 /// are inserted unconditionally, an analysis would be required to only copy
120 /// live-in and live-out values when necessary. This copies the entire memref
121 /// pointed to by "from". In case a smaller block would be sufficient, the
122 /// caller can create a subview of the memref and promote it instead.
123 static void insertCopies(Region &region, Location loc, Value from, Value to) {
124   auto fromType = from.getType().cast<MemRefType>();
125   auto toType = to.getType().cast<MemRefType>();
126   (void)fromType;
127   (void)toType;
128   assert(fromType.getShape() == toType.getShape());
129   assert(fromType.getRank() != 0);
130   assert(llvm::hasSingleElement(region) &&
131          "unstructured control flow not supported");
132 
133   auto b = ImplicitLocOpBuilder::atBlockBegin(loc, &region.front());
134   insertCopyLoops(b, from, to);
135   b.create<gpu::BarrierOp>();
136 
137   b.setInsertionPoint(&region.front().back());
138   b.create<gpu::BarrierOp>();
139   insertCopyLoops(b, to, from);
140 }
141 
142 /// Promotes a function argument to workgroup memory in the given function. The
143 /// copies will be inserted in the beginning and in the end of the function.
144 void mlir::promoteToWorkgroupMemory(GPUFuncOp op, unsigned arg) {
145   Value value = op.getArgument(arg);
146   auto type = value.getType().dyn_cast<MemRefType>();
147   assert(type && type.hasStaticShape() && "can only promote memrefs");
148 
149   // Get the type of the buffer in the workgroup memory.
150   int workgroupMemoryAddressSpace = gpu::GPUDialect::getWorkgroupAddressSpace();
151   auto bufferType = MemRefType::get(type.getShape(), type.getElementType(), {},
152                                     workgroupMemoryAddressSpace);
153   Value attribution = op.addWorkgroupAttribution(bufferType, value.getLoc());
154 
155   // Replace the uses first since only the original uses are currently present.
156   // Then insert the copies.
157   value.replaceAllUsesWith(attribution);
158   insertCopies(op.getBody(), op.getLoc(), value, attribution);
159 }
160