1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
|
//===-- SISchedule.td - SI Scheduling definitions -------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// MachineModel definitions for Southern Islands (SI)
//
//===----------------------------------------------------------------------===//
def : PredicateProlog<[{
const SIInstrInfo *TII =
static_cast<const SIInstrInfo*>(SchedModel->getInstrInfo());
(void)TII;
}]>;
def WriteBranch : SchedWrite;
def WriteExport : SchedWrite;
def WriteLDS : SchedWrite;
def WriteSALU : SchedWrite;
def WriteSMEM : SchedWrite;
def WriteVMEM : SchedWrite;
def WriteBarrier : SchedWrite;
def MIVGPRRead : SchedRead;
def MIMFMARead : SchedRead;
// Normal 16 or 32 bit VALU instructions
def Write32Bit : SchedWrite;
// Conversion to or from F32 (but not converting F64 to or from F32)
def WriteFloatCvt : SchedWrite;
// F16 or F32 transcendental instructions (these are quarter rate)
def WriteTrans32 : SchedWrite;
// Other quarter rate VALU instructions
def WriteQuarterRate32 : SchedWrite;
def WriteFloatFMA : SchedWrite;
// Slow quarter rate f64 instruction.
def WriteDouble : SchedWrite;
// half rate f64 instruction (same as v_add_f64)
def WriteDoubleAdd : SchedWrite;
// Conversion to or from f64 instruction
def WriteDoubleCvt : SchedWrite;
// F64 "transcendental" (actually only reciprocal and/or square root)
// instructions
def WriteTrans64 : SchedWrite;
// Half rate 64-bit instructions.
def Write64Bit : SchedWrite;
// Integer multiplications.
def WriteIntMul : SchedWrite;
// mAI multipass instructions.
def Write2PassMAI : SchedWrite;
def Write4PassMAI : SchedWrite;
def Write8PassMAI : SchedWrite;
def Write16PassMAI : SchedWrite;
def Write4PassDGEMM : SchedWrite;
def Write8PassDGEMM : SchedWrite;
// FIXME: Should there be a class for instructions which are VALU
// instructions and have VALU rates, but write to the SALU (i.e. VOPC
// instructions)
class SISchedMachineModel : SchedMachineModel {
let CompleteModel = 1;
// MicroOpBufferSize = 1 means that instructions will always be added
// the ready queue when they become available. This exposes them
// to the register pressure analysis.
let MicroOpBufferSize = 1;
let IssueWidth = 1;
let PostRAScheduler = 1;
// FIXME:Approximate 2 * branch cost. Try to hack around bad
// early-ifcvt heuristics. These need improvement to avoid the OOE
// heuristics.
int MispredictPenalty = 20;
}
def SIFullSpeedModel : SISchedMachineModel;
def SIQuarterSpeedModel : SISchedMachineModel;
def SIDPFullSpeedModel : SISchedMachineModel;
def SIDPGFX940FullSpeedModel : SISchedMachineModel;
def GFX10SpeedModel : SISchedMachineModel;
def GFX11SpeedModel : SISchedMachineModel;
// XXX: Are the resource counts correct?
def HWBranch : ProcResource<1> {
let BufferSize = 1;
}
def HWExport : ProcResource<1> {
let BufferSize = 1;
}
def HWLGKM : ProcResource<1> {
let BufferSize = 1;
}
def HWSALU : ProcResource<1> {
let BufferSize = 1;
}
def HWVMEM : ProcResource<1> {
let BufferSize = 1;
}
def HWVALU : ProcResource<1> {
let BufferSize = 1;
}
def HWTransVALU : ProcResource<1> { // Transcendental VALU
let BufferSize = 1;
}
def HWRC : ProcResource<1> { // Register destination cache
let BufferSize = 1;
}
def HWXDL : ProcResource<1> { // MFMA CU
let BufferSize = 0;
}
class HWWriteRes<SchedWrite write, list<ProcResourceKind> resources,
int latency> : WriteRes<write, resources> {
let Latency = latency;
}
class HWVALUWriteRes<SchedWrite write, int latency> :
HWWriteRes<write, [HWVALU], latency>;
def PredMIReadVGPR : SchedPredicate<[{TII->hasVGPRUses(*MI)}]>;
def MIReadVGPR : SchedReadVariant<[
SchedVar<PredMIReadVGPR, [MIVGPRRead]>,
SchedVar<NoSchedPred, [ReadDefault]>]>;
// The latency numbers are taken from AMD Accelerated Parallel Processing
// guide. They may not be accurate.
// The latency values are 1 / (operations / cycle) / 4.
multiclass SICommonWriteRes {
let RetireOOO = 1 in { // llvm-mca specific flag
def : HWWriteRes<WriteBranch, [HWBranch], 8>;
def : HWWriteRes<WriteExport, [HWExport], 4>;
def : HWWriteRes<WriteLDS, [HWLGKM], 5>; // Can be between 2 and 64
def : HWWriteRes<WriteSALU, [HWSALU], 1>;
def : HWWriteRes<WriteSMEM, [HWLGKM], 5>;
def : HWWriteRes<WriteVMEM, [HWVMEM], 80>;
def : HWWriteRes<WriteBarrier, [HWBranch], 500>; // XXX: Guessed ???
def : HWVALUWriteRes<Write32Bit, 1>;
def : HWVALUWriteRes<WriteFloatCvt, 4>;
def : HWVALUWriteRes<WriteTrans32, 4>;
def : HWVALUWriteRes<WriteQuarterRate32, 4>;
def : HWVALUWriteRes<Write4PassDGEMM, 4>;
def : HWVALUWriteRes<Write8PassDGEMM, 16>;
let ResourceCycles = [2] in
def : HWWriteRes<Write2PassMAI, [HWXDL], 2>;
let ResourceCycles = [4] in
def : HWWriteRes<Write4PassMAI, [HWXDL], 4>;
let ResourceCycles = [8] in
def : HWWriteRes<Write8PassMAI, [HWXDL], 8>;
let ResourceCycles = [16] in
def : HWWriteRes<Write16PassMAI, [HWXDL], 16>;
} // End RetireOOO = 1
def : ReadAdvance<MIVGPRRead, -2>;
// Technically mfma reads can be from 0 to 4 cycles but that does not make
// sense to model because its register setup is huge. In particular if we
// properly model read advance as -2 for a vgpr read it will result in a
// bad scheduling of acc writes before that mfma. To avoid it we would
// need to consume 2 or 4 more vgprs to be initialized before the acc
// write sequence. Just assume worst case here.
def : ReadAdvance<MIMFMARead, -4>;
}
def PredIsVGPR32Copy : SchedPredicate<[{TII->isVGPRCopy(*MI) && TII->getOpSize(*MI, 0) <= 32}]>;
def PredIsVGPR64Copy : SchedPredicate<[{TII->isVGPRCopy(*MI) && TII->getOpSize(*MI, 0) > 32}]>;
def WriteCopy : SchedWriteVariant<[
SchedVar<PredIsVGPR32Copy, [Write32Bit]>,
SchedVar<PredIsVGPR64Copy, [Write64Bit]>,
SchedVar<NoSchedPred, [WriteSALU]>]>;
let SchedModel = SIFullSpeedModel in {
defm : SICommonWriteRes;
let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<Write64Bit, 2>;
def : HWVALUWriteRes<WriteIntMul, 4>;
def : HWVALUWriteRes<WriteFloatFMA, 1>;
def : HWVALUWriteRes<WriteDouble, 4>;
def : HWVALUWriteRes<WriteDoubleAdd, 2>;
def : HWVALUWriteRes<WriteDoubleCvt, 4>;
def : HWVALUWriteRes<WriteTrans64, 4>;
} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
} // End SchedModel = SIFullSpeedModel
let SchedModel = SIQuarterSpeedModel in {
defm : SICommonWriteRes;
let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<Write64Bit, 2>;
def : HWVALUWriteRes<WriteIntMul, 4>;
def : HWVALUWriteRes<WriteFloatFMA, 16>;
def : HWVALUWriteRes<WriteDouble, 16>;
def : HWVALUWriteRes<WriteDoubleAdd, 8>;
def : HWVALUWriteRes<WriteDoubleCvt, 4>;
def : HWVALUWriteRes<WriteTrans64, 16>;
} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
def : InstRW<[Write64Bit, MIReadVGPR], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
def : InstRW<[Write2PassMAI, MIMFMARead], (instregex "^V_MFMA_..._4X4X")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_..._16X16X")>;
def : InstRW<[Write16PassMAI, MIMFMARead], (instregex "^V_MFMA_..._32X32X")>;
} // End SchedModel = SIQuarterSpeedModel
let SchedModel = SIDPFullSpeedModel in {
defm : SICommonWriteRes;
let RetireOOO = 1 in { // llvm-mca specific flag
def : HWVALUWriteRes<WriteFloatFMA, 1>;
def : HWVALUWriteRes<WriteDouble, 1>;
def : HWVALUWriteRes<WriteDoubleAdd, 1>;
def : HWVALUWriteRes<WriteDoubleCvt, 1>;
def : HWVALUWriteRes<WriteTrans64, 4>;
def : HWVALUWriteRes<WriteIntMul, 1>;
def : HWVALUWriteRes<Write64Bit, 1>;
} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
def : InstRW<[Write64Bit], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
def : InstRW<[Write2PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_4X4X")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_16X16X")>;
def : InstRW<[Write16PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_32X32X")>;
def : InstRW<[Write4PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_4X4X")>;
def : InstRW<[Write8PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_16X16X")>;
} // End SchedModel = SIDPFullSpeedModel
let SchedModel = SIDPGFX940FullSpeedModel in {
defm : SICommonWriteRes;
def : HWVALUWriteRes<WriteFloatFMA, 1>;
def : HWVALUWriteRes<WriteDouble, 1>;
def : HWVALUWriteRes<WriteDoubleAdd, 1>;
def : HWVALUWriteRes<WriteDoubleCvt, 1>;
def : HWVALUWriteRes<WriteTrans64, 4>;
def : HWVALUWriteRes<WriteIntMul, 1>;
def : HWVALUWriteRes<Write64Bit, 1>;
def : InstRW<[WriteCopy], (instrs COPY)>;
def : InstRW<[Write64Bit], (instregex "^V_ACCVGPR_WRITE_B32_e64$")>;
def : InstRW<[Write2PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_4X4X")>;
def : InstRW<[Write4PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_16X16X8X")>;
def : InstRW<[Write4PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_16X16X16")>;
def : InstRW<[Write4PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_16X16X32")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_16X16X[14][FBI]")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_32X32X4XF")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_32X32X8")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_32X32X16")>;
def : InstRW<[Write16PassMAI, MIMFMARead], (instregex "^V_MFMA_.32_32X32X[124][FBI]")>;
def : InstRW<[Write4PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_4X4X")>;
def : InstRW<[Write8PassDGEMM, MIMFMARead], (instregex "^V_MFMA_.64_16X16X")>;
def : InstRW<[Write4PassMAI, MIMFMARead], (instregex "^V_SMFMAC_.32_16X16X")>;
def : InstRW<[Write8PassMAI, MIMFMARead], (instregex "^V_SMFMAC_.32_32X32X")>;
} // End SchedModel = SIDPGFX940FullSpeedModel
let SchedModel = GFX10SpeedModel in {
// The latency values are 1 / (operations / cycle).
// Add 1 stall cycle for VGPR read.
let RetireOOO = 1 in { // llvm-mca specific flag
def : HWWriteRes<Write32Bit, [HWVALU, HWRC], 5>;
def : HWWriteRes<WriteFloatCvt, [HWVALU, HWRC], 5>;
def : HWWriteRes<Write64Bit, [HWVALU, HWRC], 6>;
def : HWWriteRes<WriteTrans32, [HWTransVALU, HWRC], 10>;
def : HWWriteRes<WriteQuarterRate32, [HWVALU, HWRC], 8>;
def : HWWriteRes<WriteFloatFMA, [HWVALU, HWRC], 5>;
def : HWWriteRes<WriteDouble, [HWVALU, HWRC], 22>;
def : HWWriteRes<WriteDoubleAdd, [HWVALU, HWRC], 22>;
def : HWWriteRes<WriteDoubleCvt, [HWVALU, HWRC], 22>;
def : HWWriteRes<WriteIntMul, [HWVALU, HWRC], 8>;
def : HWWriteRes<WriteTrans64, [HWVALU, HWTransVALU, HWRC], 24>;
def : HWWriteRes<WriteBranch, [HWBranch], 32>;
def : HWWriteRes<WriteExport, [HWExport, HWRC], 16>;
def : HWWriteRes<WriteLDS, [HWLGKM, HWRC], 20>;
def : HWWriteRes<WriteSALU, [HWSALU, HWRC], 2>;
def : HWWriteRes<WriteSMEM, [HWLGKM, HWRC], 20>;
def : HWWriteRes<WriteVMEM, [HWVMEM, HWRC], 320>;
def : HWWriteRes<WriteBarrier, [HWBranch], 2000>;
} // End RetireOOO = 1
def : InstRW<[WriteCopy], (instrs COPY)>;
} // End SchedModel = GFX10SpeedModel
let SchedModel = GFX11SpeedModel in {
def : HWWriteRes<Write32Bit, [HWVALU, HWRC], 5>;
def : HWWriteRes<WriteFloatCvt, [HWVALU, HWRC], 5>;
def : HWWriteRes<Write64Bit, [HWVALU, HWRC], 6>;
def : HWWriteRes<WriteTrans32, [HWVALU, HWRC], 10>;
def : HWWriteRes<WriteQuarterRate32, [HWVALU, HWRC], 8>;
def : HWWriteRes<WriteFloatFMA, [HWVALU, HWRC], 5>;
def : HWWriteRes<WriteDouble, [HWVALU, HWRC], 38>;
def : HWWriteRes<WriteDoubleAdd, [HWVALU, HWRC], 38>;
def : HWWriteRes<WriteDoubleCvt, [HWVALU, HWRC], 38>;
def : HWWriteRes<WriteIntMul, [HWVALU, HWRC], 8>;
def : HWWriteRes<WriteTrans64, [HWVALU, HWRC], 40>;
def : HWWriteRes<WriteBranch, [HWBranch], 32>;
def : HWWriteRes<WriteExport, [HWExport, HWRC], 16>;
def : HWWriteRes<WriteLDS, [HWLGKM, HWRC], 20>;
def : HWWriteRes<WriteSALU, [HWSALU, HWRC], 2>;
def : HWWriteRes<WriteSMEM, [HWLGKM, HWRC], 20>;
def : HWWriteRes<WriteVMEM, [HWVMEM, HWRC], 320>;
def : HWWriteRes<WriteBarrier, [HWBranch], 2000>;
def : InstRW<[WriteCopy], (instrs COPY)>;
} // End SchedModel = GFX11SpeedModel
|
