1 //===-- SISchedule.td - SI Scheduling definitions -------------------------===//
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
7 //===----------------------------------------------------------------------===//
9 // MachineModel definitions for Southern Islands (SI)
11 //===----------------------------------------------------------------------===//
13 def : PredicateProlog<[{
14 const SIInstrInfo *TII =
15 static_cast<const SIInstrInfo*>(SchedModel->getInstrInfo());
19 def WriteBranch : SchedWrite;
20 def WriteExport : SchedWrite;
21 def WriteLDS : SchedWrite;
22 def WriteSALU : SchedWrite;
23 def WriteSMEM : SchedWrite;
24 def WriteVMEM : SchedWrite;
25 def WriteBarrier : SchedWrite;
27 def MIVGPRRead : SchedRead;
28 def MIMFMARead : SchedRead;
30 // Normal 16 or 32 bit VALU instructions
31 def Write32Bit : SchedWrite;
32 // Conversion to or from F32 (but not converting F64 to or from F32)
33 def WriteFloatCvt : SchedWrite;
34 // F16 or F32 transcendental instructions (these are quarter rate)
35 def WriteTrans32 : SchedWrite;
36 // Other quarter rate VALU instructions
37 def WriteQuarterRate32 : SchedWrite;
39 def WriteFloatFMA : SchedWrite;
41 // Slow quarter rate f64 instruction.
42 def WriteDouble : SchedWrite;
44 // half rate f64 instruction (same as v_add_f64)
45 def WriteDoubleAdd : SchedWrite;
47 // Conversion to or from f64 instruction
48 def WriteDoubleCvt : SchedWrite;
50 // F64 "transcendental" (actually only reciprocal and/or square root)
52 def WriteTrans64 : SchedWrite;
54 // Half rate 64-bit instructions.
55 def Write64Bit : SchedWrite;
57 // mAI multipass instructions.
58 def Write2PassMAI : SchedWrite;
59 def Write8PassMAI : SchedWrite;
60 def Write16PassMAI : SchedWrite;
62 // FIXME: Should there be a class for instructions which are VALU
63 // instructions and have VALU rates, but write to the SALU (i.e. VOPC
66 class SISchedMachineModel : SchedMachineModel {
67 let CompleteModel = 1;
68 // MicroOpBufferSize = 1 means that instructions will always be added
69 // the ready queue when they become available. This exposes them
70 // to the register pressure analysis.
71 let MicroOpBufferSize = 1;
73 let PostRAScheduler = 1;
75 // FIXME:Approximate 2 * branch cost. Try to hack around bad
76 // early-ifcvt heuristics. These need improvement to avoid the OOE
78 int MispredictPenalty = 20;
81 def SIFullSpeedModel : SISchedMachineModel;
82 def SIQuarterSpeedModel : SISchedMachineModel;
83 def GFX10SpeedModel : SISchedMachineModel;
85 // XXX: Are the resource counts correct?
86 def HWBranch : ProcResource<1> {
89 def HWExport : ProcResource<1> {
90 let BufferSize = 7; // Taken from S_WAITCNT
92 def HWLGKM : ProcResource<1> {
93 let BufferSize = 31; // Taken from S_WAITCNT
95 def HWSALU : ProcResource<1> {
98 def HWVMEM : ProcResource<1> {
99 let BufferSize = 15; // Taken from S_WAITCNT
101 def HWVALU : ProcResource<1> {
104 def HWRC : ProcResource<1> { // Register destination cache
108 class HWWriteRes<SchedWrite write, list<ProcResourceKind> resources,
109 int latency> : WriteRes<write, resources> {
110 let Latency = latency;
113 class HWVALUWriteRes<SchedWrite write, int latency> :
114 HWWriteRes<write, [HWVALU], latency>;
116 def PredMIReadVGPR : SchedPredicate<[{TII->hasVGPRUses(*MI)}]>;
118 def MIReadVGPR : SchedReadVariant<[
119 SchedVar<PredMIReadVGPR, [MIVGPRRead]>,
120 SchedVar<NoSchedPred, [ReadDefault]>]>;
122 // The latency numbers are taken from AMD Accelerated Parallel Processing
123 // guide. They may not be accurate.
125 // The latency values are 1 / (operations / cycle) / 4.
126 multiclass SICommonWriteRes {
128 def : HWWriteRes<WriteBranch, [HWBranch], 8>;
129 def : HWWriteRes<WriteExport, [HWExport], 4>;
130 def : HWWriteRes<WriteLDS, [HWLGKM], 5>; // Can be between 2 and 64
131 def : HWWriteRes<WriteSALU, [HWSALU], 1>;
132 def : HWWriteRes<WriteSMEM, [HWLGKM], 5>;
133 def : HWWriteRes<WriteVMEM, [HWVMEM], 80>;
134 def : HWWriteRes<WriteBarrier, [HWBranch], 500>; // XXX: Guessed ???
136 def : HWVALUWriteRes<Write32Bit, 1>;
137 def : HWVALUWriteRes<Write64Bit, 2>;
138 def : HWVALUWriteRes<WriteFloatCvt, 4>;
139 def : HWVALUWriteRes<WriteTrans32, 4>;
140 def : HWVALUWriteRes<WriteQuarterRate32, 4>;
141 def : HWVALUWriteRes<Write2PassMAI, 2>;
142 def : HWVALUWriteRes<Write8PassMAI, 8>;
143 def : HWVALUWriteRes<Write16PassMAI, 16>;
145 def : ReadAdvance<MIVGPRRead, -2>;
146 def : InstRW<[Write64Bit, MIReadVGPR], (instregex "^V_ACCVGPR_WRITE_B32$")>;
148 // Technically mfma reads can be from 0 to 4 cycles but that does not make
149 // sense to model because its register setup is huge. In particular if we
150 // properly model read advance as -2 for a vgpr read it will result in a
151 // bad scheduling of acc writes before that mfma. To avoid it we would
152 // need to consume 2 or 4 more vgprs to be initialized before the acc
153 // write sequence. Just assume worst case here.
154 def : ReadAdvance<MIMFMARead, -4>;
156 def : InstRW<[Write2PassMAI, MIMFMARead], (instregex "^V_MFMA_..._4X4X")>;
157 def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_..._16X16X")>;
158 def : InstRW<[Write16PassMAI, MIMFMARead], (instregex "^V_MFMA_..._32X32X")>;
161 def PredIsVGPR32Copy : SchedPredicate<[{TII->isVGPRCopy(*MI) && TII->getOpSize(*MI, 0) <= 32}]>;
162 def PredIsVGPR64Copy : SchedPredicate<[{TII->isVGPRCopy(*MI) && TII->getOpSize(*MI, 0) > 32}]>;
163 def WriteCopy : SchedWriteVariant<[
164 SchedVar<PredIsVGPR32Copy, [Write32Bit]>,
165 SchedVar<PredIsVGPR64Copy, [Write64Bit]>,
166 SchedVar<NoSchedPred, [WriteSALU]>]>;
168 let SchedModel = SIFullSpeedModel in {
170 defm : SICommonWriteRes;
172 def : HWVALUWriteRes<WriteFloatFMA, 1>;
173 def : HWVALUWriteRes<WriteDouble, 4>;
174 def : HWVALUWriteRes<WriteDoubleAdd, 2>;
175 def : HWVALUWriteRes<WriteDoubleCvt, 4>;
176 def : HWVALUWriteRes<WriteTrans64, 4>;
178 def : InstRW<[WriteCopy], (instrs COPY)>;
180 } // End SchedModel = SIFullSpeedModel
182 let SchedModel = SIQuarterSpeedModel in {
184 defm : SICommonWriteRes;
186 def : HWVALUWriteRes<WriteFloatFMA, 16>;
187 def : HWVALUWriteRes<WriteDouble, 16>;
188 def : HWVALUWriteRes<WriteDoubleAdd, 8>;
189 def : HWVALUWriteRes<WriteDoubleCvt, 4>;
190 def : HWVALUWriteRes<WriteTrans64, 16>;
192 def : InstRW<[WriteCopy], (instrs COPY)>;
194 } // End SchedModel = SIQuarterSpeedModel
196 let SchedModel = GFX10SpeedModel in {
198 // The latency values are 1 / (operations / cycle).
199 // Add 1 stall cycle for VGPR read.
200 def : HWWriteRes<Write32Bit, [HWVALU, HWRC], 5>;
201 def : HWWriteRes<WriteFloatCvt, [HWVALU, HWRC], 5>;
202 def : HWWriteRes<Write64Bit, [HWVALU, HWRC], 6>;
203 def : HWWriteRes<WriteTrans32, [HWVALU, HWRC], 10>;
204 def : HWWriteRes<WriteQuarterRate32, [HWVALU, HWRC], 8>;
205 def : HWWriteRes<WriteFloatFMA, [HWVALU, HWRC], 5>;
206 def : HWWriteRes<WriteDouble, [HWVALU, HWRC], 22>;
207 def : HWWriteRes<WriteDoubleAdd, [HWVALU, HWRC], 22>;
208 def : HWWriteRes<WriteDoubleCvt, [HWVALU, HWRC], 22>;
209 def : HWWriteRes<WriteTrans64, [HWVALU, HWRC], 24>;
211 def : HWWriteRes<WriteBranch, [HWBranch], 32>;
212 def : HWWriteRes<WriteExport, [HWExport, HWRC], 16>;
213 def : HWWriteRes<WriteLDS, [HWLGKM, HWRC], 20>;
214 def : HWWriteRes<WriteSALU, [HWSALU, HWRC], 2>;
215 def : HWWriteRes<WriteSMEM, [HWLGKM, HWRC], 20>;
216 def : HWWriteRes<WriteVMEM, [HWVMEM, HWRC], 320>;
217 def : HWWriteRes<WriteBarrier, [HWBranch], 2000>;
219 def : InstRW<[WriteCopy], (instrs COPY)>;
221 } // End SchedModel = GFX10SpeedModel