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
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
|
// RUN: mlir-opt %s -split-input-file -verify-diagnostics | FileCheck %s
// CHECK-LABEL: @nvvm_special_regs
func.func @nvvm_special_regs() -> i32 {
// CHECK: nvvm.read.ptx.sreg.tid.x : i32
%0 = nvvm.read.ptx.sreg.tid.x : i32
// CHECK: nvvm.read.ptx.sreg.tid.y : i32
%1 = nvvm.read.ptx.sreg.tid.y : i32
// CHECK: nvvm.read.ptx.sreg.tid.z : i32
%2 = nvvm.read.ptx.sreg.tid.z : i32
// CHECK: nvvm.read.ptx.sreg.ntid.x : i32
%3 = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.read.ptx.sreg.ntid.y : i32
%4 = nvvm.read.ptx.sreg.ntid.y : i32
// CHECK: nvvm.read.ptx.sreg.ntid.z : i32
%5 = nvvm.read.ptx.sreg.ntid.z : i32
// CHECK: nvvm.read.ptx.sreg.ctaid.x : i32
%6 = nvvm.read.ptx.sreg.ctaid.x : i32
// CHECK: nvvm.read.ptx.sreg.ctaid.y : i32
%7 = nvvm.read.ptx.sreg.ctaid.y : i32
// CHECK: nvvm.read.ptx.sreg.ctaid.z : i32
%8 = nvvm.read.ptx.sreg.ctaid.z : i32
// CHECK: nvvm.read.ptx.sreg.nctaid.x : i32
%9 = nvvm.read.ptx.sreg.nctaid.x : i32
// CHECK: nvvm.read.ptx.sreg.nctaid.y : i32
%10 = nvvm.read.ptx.sreg.nctaid.y : i32
// CHECK: nvvm.read.ptx.sreg.nctaid.z : i32
%11 = nvvm.read.ptx.sreg.nctaid.z : i32
llvm.return %0 : i32
}
// CHECK-LABEL: @nvvm_rcp
func.func @nvvm_rcp(%arg0: f32) -> f32 {
// CHECK: nvvm.rcp.approx.ftz.f %arg0 : f32
%0 = nvvm.rcp.approx.ftz.f %arg0 : f32
llvm.return %0 : f32
}
// CHECK-LABEL: @llvm_nvvm_barrier0
func.func @llvm_nvvm_barrier0() {
// CHECK: nvvm.barrier0
nvvm.barrier0
llvm.return
}
// CHECK-LABEL: @nvvm_shfl
func.func @nvvm_shfl(
%arg0 : i32, %arg1 : i32, %arg2 : i32,
%arg3 : i32, %arg4 : f32) -> i32 {
// CHECK: nvvm.shfl.sync bfly %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : i32 -> i32
%0 = nvvm.shfl.sync bfly %arg0, %arg3, %arg1, %arg2 : i32 -> i32
// CHECK: nvvm.shfl.sync bfly %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : f32 -> f32
%1 = nvvm.shfl.sync bfly %arg0, %arg4, %arg1, %arg2 : f32 -> f32
// CHECK: nvvm.shfl.sync up %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : f32 -> f32
%2 = nvvm.shfl.sync up %arg0, %arg4, %arg1, %arg2 : f32 -> f32
// CHECK: nvvm.shfl.sync down %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : f32 -> f32
%3 = nvvm.shfl.sync down %arg0, %arg4, %arg1, %arg2 : f32 -> f32
// CHECK: nvvm.shfl.sync idx %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} : f32 -> f32
%4 = nvvm.shfl.sync idx %arg0, %arg4, %arg1, %arg2 : f32 -> f32
llvm.return %0 : i32
}
// CHECK-LABEL: @nvvm_shfl_pred
func.func @nvvm_shfl_pred(
%arg0 : i32, %arg1 : i32, %arg2 : i32,
%arg3 : i32, %arg4 : f32) -> !llvm.struct<(i32, i1)> {
// CHECK: nvvm.shfl.sync bfly %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} {return_value_and_is_valid} : i32 -> !llvm.struct<(i32, i1)>
%0 = nvvm.shfl.sync bfly %arg0, %arg3, %arg1, %arg2 {return_value_and_is_valid} : i32 -> !llvm.struct<(i32, i1)>
// CHECK: nvvm.shfl.sync bfly %{{.*}}, %{{.*}}, %{{.*}}, %{{.*}} {return_value_and_is_valid} : f32 -> !llvm.struct<(f32, i1)>
%1 = nvvm.shfl.sync bfly %arg0, %arg4, %arg1, %arg2 {return_value_and_is_valid} : f32 -> !llvm.struct<(f32, i1)>
llvm.return %0 : !llvm.struct<(i32, i1)>
}
// CHECK-LABEL: @nvvm_vote(
func.func @nvvm_vote(%arg0 : i32, %arg1 : i1) -> i32 {
// CHECK: nvvm.vote.ballot.sync %{{.*}}, %{{.*}} : i32
%0 = nvvm.vote.ballot.sync %arg0, %arg1 : i32
llvm.return %0 : i32
}
// CHECK-LABEL: @llvm_nvvm_bar_warp_sync
func.func @llvm_nvvm_bar_warp_sync(%mask : i32) {
// CHECK: nvvm.bar.warp.sync %{{.*}}
nvvm.bar.warp.sync %mask : i32
llvm.return
}
// CHECK-LABEL: @nvvm_mma_m8n8k4_row_col_f32_f32
func.func @nvvm_mma_m8n8k4_row_col_f32_f32(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : f32, %c1 : f32, %c2 : f32, %c3 : f32, %c4 : f32, %c5 : f32, %c6 : f32, %c7 : f32) {
// CHECK: nvvm.mma.sync
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0, %b1] C[%c0, %c1, %c2, %c3, %c4, %c5, %c6, %c7]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 8, n = 8, k = 4>} : (vector<2xf16>, vector<2xf16>, f32) -> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32)>
llvm.return %0 : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32)>
}
// CHECK-LABEL: @nvvm_mma_m8n8k4_f16_f16
func.func @nvvm_mma_m8n8k4_f16_f16(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : vector<2xf16>, %c1 : vector<2xf16>, %c2 : vector<2xf16>, %c3 : vector<2xf16>) {
// CHECK: nvvm.mma.sync A[{{.*}}] B[{{.*}}] C[{{.*}}]
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0, %b1] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 8, n = 8, k = 4>} : (vector<2xf16>,vector<2xf16>,vector<2xf16>) -> !llvm.struct<(vector<2xf16>, vector<2xf16>, vector<2xf16>, vector<2xf16>)>
llvm.return %0 : !llvm.struct<(vector<2xf16>, vector<2xf16>, vector<2xf16>, vector<2xf16>)>
}
// CHECK-LABEL: @nvvm_mma_m8n8k16_s8_s8
func.func @nvvm_mma_m8n8k16_s8_s8(%a0 : i32, %b0 : i32,
%c0 : i32, %c1 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}] {intOverflowBehavior = #nvvm.mma_int_overflow<wrapped>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<s8>, shape = #nvvm.shape<m = 8, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32, i32)>
%0 = nvvm.mma.sync A[%a0] B[%b0] C[%c0, %c1]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<s8>,
intOverflowBehavior=#nvvm.mma_int_overflow<wrapped>,
shape = #nvvm.shape<m = 8, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32, i32)>
llvm.return %0 : !llvm.struct<(i32, i32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k8_f16_f16
func.func @nvvm_mma_m16n8k8_f16_f16(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%b0 : vector<2xf16>,
%c0 : vector<2xf16>, %c1 : vector<2xf16>) {
// CHECK: nvvm.mma.sync A[%{{.*}}, %{{.*}}] B[%{{.*}}] C[%{{.*}}, %{{.*}}] {{{.*}}} : (vector<2xf16>, vector<2xf16>, vector<2xf16>) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 16, n = 8, k = 8>} : (vector<2xf16>, vector<2xf16>, vector<2xf16>) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
llvm.return %0 : !llvm.struct<(vector<2xf16>, vector<2xf16>)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_f16_f16
func.func @nvvm_mma_m16n8k16_f16_f16(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%a2 : vector<2xf16>, %a3 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : vector<2xf16>, %c1 : vector<2xf16>) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}, {{.*}}, {{.*}}] B[{{.*}}, {{.*}}] C[{{.*}}, {{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, vector<2xf16>) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
%0 = nvvm.mma.sync A[%a0, %a1, %a2, %a3] B[%b0, %b1] C[%c0, %c1]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, vector<2xf16>) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
llvm.return %0 : !llvm.struct<(vector<2xf16>, vector<2xf16>)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_f32_f16
func.func @nvvm_mma_m16n8k16_f32_f16(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%a2 : vector<2xf16>, %a3 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : vector<2xf16>, %c1 : vector<2xf16>) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}, {{.*}}, {{.*}}] B[{{.*}}, {{.*}}] C[{{.*}}, {{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, vector<2xf16>) -> !llvm.struct<(f32, f32, f32, f32)>
%0 = nvvm.mma.sync A[%a0, %a1, %a2, %a3] B[%b0, %b1] C[%c0, %c1]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>,vector<2xf16>,vector<2xf16>) -> !llvm.struct<(f32, f32, f32, f32)>
llvm.return %0 : !llvm.struct<(f32, f32, f32, f32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_f16_f32
func.func @nvvm_mma_m16n8k16_f16_f32(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%a2 : vector<2xf16>, %a3 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : f32, %c1 : f32, %c2 : f32, %c3 : f32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}, {{.*}}, {{.*}}] B[{{.*}}, {{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, f32) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
%0 = nvvm.mma.sync A[%a0, %a1, %a2, %a3] B[%b0, %b1] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, f32) -> !llvm.struct<(vector<2xf16>, vector<2xf16>)>
llvm.return %0 : !llvm.struct<(vector<2xf16>, vector<2xf16>)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_f32_f32
func.func @nvvm_mma_m16n8k16_f32_f32(%a0 : vector<2xf16>, %a1 : vector<2xf16>,
%a2 : vector<2xf16>, %a3 : vector<2xf16>,
%b0 : vector<2xf16>, %b1 : vector<2xf16>,
%c0 : f32, %c1 : f32, %c2 : f32, %c3 : f32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}, {{.*}}, {{.*}}] B[{{.*}}, {{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, f32) -> !llvm.struct<(f32, f32, f32, f32)>
%0 = nvvm.mma.sync A[%a0, %a1, %a2, %a3] B[%b0, %b1] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (vector<2xf16>, vector<2xf16>, f32) -> !llvm.struct<(f32, f32, f32, f32)>
llvm.return %0 : !llvm.struct<(f32, f32, f32, f32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k4_tf32_f32
func.func @nvvm_mma_m16n8k4_tf32_f32(%a0 : i32, %a1 : i32,
%b0 : i32,
%c0 : f32, %c1 : f32, %c2 : f32, %c3 : f32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<tf32>, multiplicandBPtxType = #nvvm.mma_type<tf32>, shape = #nvvm.shape<m = 16, n = 8, k = 4>} : (i32, i32, f32) -> !llvm.struct<(f32, f32, f32, f32)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<tf32>, multiplicandBPtxType = #nvvm.mma_type<tf32>,
shape = #nvvm.shape<m = 16, n = 8, k = 4>} : (i32, i32, f32) -> !llvm.struct<(f32, f32, f32, f32)>
llvm.return %0 : !llvm.struct<(f32, f32, f32, f32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_s8_s8
func.func @nvvm_mma_m16n8k16_s8_s8(%a0 : i32, %a1 : i32, %b0 : i32,
%c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {intOverflowBehavior = #nvvm.mma_int_overflow<wrapped>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<s8>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32, i32, i32, i32)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<s8>,
intOverflowBehavior=#nvvm.mma_int_overflow<wrapped>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k16_s8_u8
func.func @nvvm_mma_m16n8k16_s8_u8(%a0 : i32, %a1 : i32,
%b0 : i32,
%c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {intOverflowBehavior = #nvvm.mma_int_overflow<satfinite>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<u8>, shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32, i32, i32, i32)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<s8>, multiplicandBPtxType = #nvvm.mma_type<u8>,
intOverflowBehavior=#nvvm.mma_int_overflow<satfinite>,
shape = #nvvm.shape<m = 16, n = 8, k = 16>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k256_b1_b1
func.func @nvvm_mma_m16n8k256_b1_b1(%a0 : i32, %a1 : i32, %a2 : i32, %a3 : i32,
%b0 : i32, %b1 : i32,
%c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}, {{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {b1Op = #nvvm.mma_b1op<xor_popc>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>, shape = #nvvm.shape<m = 16, n = 8, k = 256>} : (i32, i32, i32) -> !llvm.struct<(i32, i32, i32, i32)>
%0 = nvvm.mma.sync A[%a0, %a1, %a2, %a3] B[%b0, %b1] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>,
b1Op = #nvvm.mma_b1op<xor_popc>, shape = #nvvm.shape<m = 16, n = 8, k = 256>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k128_b1_b1
func.func @nvvm_mma_m16n8k128_b1_b1(%a0 : i32, %a1 : i32,
%b0 : i32,
%c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {b1Op = #nvvm.mma_b1op<xor_popc>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>, shape = #nvvm.shape<m = 16, n = 8, k = 128>} : (i32, i32, i32) -> !llvm.struct<(i32, i32, i32, i32)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>,
b1Op = #nvvm.mma_b1op<xor_popc>,
shape = #nvvm.shape<m = 16, n = 8, k = 128>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
}
// CHECK-LABEL: @nvvm_mma_m8n8k128_b1_b1
func.func @nvvm_mma_m8n8k128_b1_b1(%a0 : i32,
%b0 : i32,
%c0 : i32, %c1 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}] {b1Op = #nvvm.mma_b1op<xor_popc>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>, shape = #nvvm.shape<m = 8, n = 8, k = 128>} : (i32, i32, i32) -> !llvm.struct<(i32, i32)>
%0 = nvvm.mma.sync A[%a0] B[%b0] C[%c0, %c1]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<b1>, multiplicandBPtxType = #nvvm.mma_type<b1>,
b1Op = #nvvm.mma_b1op<xor_popc>, shape = #nvvm.shape<m = 8, n = 8, k = 128>} : (i32, i32, i32) -> !llvm.struct<(i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32)>
}
// CHECK-LABEL: @nvvm_mma_m16n8k32_s4_s4
func.func @nvvm_mma_m16n8k32_s4_s4(%a0 : i32, %a1 : i32,
%b0 : i32,
%c0 : i32, %c1 : i32, %c2 : i32, %c3 : i32) {
// CHECK: nvvm.mma.sync A[{{.*}}, {{.*}}] B[{{.*}}] C[{{.*}}, {{.*}}, {{.*}}, {{.*}}] {intOverflowBehavior = #nvvm.mma_int_overflow<wrapped>, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>, multiplicandAPtxType = #nvvm.mma_type<s4>, multiplicandBPtxType = #nvvm.mma_type<s4>, shape = #nvvm.shape<m = 16, n = 8, k = 32>} : (i32, i32, i32) -> !llvm.struct<(i32, i32, i32, i32)>
%0 = nvvm.mma.sync A[%a0, %a1] B[%b0] C[%c0, %c1, %c2, %c3]
{layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<col>,
multiplicandAPtxType = #nvvm.mma_type<s4>, multiplicandBPtxType = #nvvm.mma_type<s4>,
intOverflowBehavior=#nvvm.mma_int_overflow<wrapped>,
shape = #nvvm.shape<m = 16, n = 8, k = 32>} : (i32, i32, i32) -> !llvm.struct<(i32,i32,i32,i32)>
llvm.return %0 : !llvm.struct<(i32,i32,i32,i32)>
}
// CHECK-LABEL: @nvvm_wmma_load_tf32
func.func @nvvm_wmma_load_tf32(%arg0: !llvm.ptr, %arg1 : i32) -> !llvm.struct<(i32, i32, i32, i32)> {
// CHECK: nvvm.wmma.load {{.*}} {eltype = #nvvm.mma_type<tf32>, frag = #nvvm.mma_frag<a>, k = 8 : i32, layout = #nvvm.mma_layout<row>, m = 16 : i32, n = 16 : i32}
%0 = nvvm.wmma.load %arg0, %arg1
{eltype = #nvvm.mma_type<tf32>, frag = #nvvm.mma_frag<a>, k = 8 : i32, layout = #nvvm.mma_layout<row>, m = 16 : i32, n = 16 : i32}
: (!llvm.ptr) -> !llvm.struct<(i32, i32, i32, i32)>
llvm.return %0 : !llvm.struct<(i32, i32, i32, i32)>
}
// CHECK-LABEL: @nvvm_wmma_mma
func.func @nvvm_wmma_mma(%0 : i32, %1 : i32, %2 : i32, %3 : i32, %4 : i32, %5 : i32,
%6 : i32, %7 : i32, %8 : f32, %9 : f32, %10 : f32,
%11 : f32, %12 : f32, %13 : f32, %14 : f32, %15 : f32)
-> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32)> {
// CHECK: nvvm.wmma.mma {{.*}} {eltypeA = #nvvm.mma_type<tf32>, eltypeB = #nvvm.mma_type<f32>, k = 8 : i32, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<row>, m = 16 : i32, n = 16 : i32}
%r = nvvm.wmma.mma %0, %1, %2, %3, %4, %5, %6, %7, %8, %9, %10, %11, %12, %13, %14, %15
{eltypeA = #nvvm.mma_type<tf32>, eltypeB = #nvvm.mma_type<f32>, k = 8 : i32, layoutA = #nvvm.mma_layout<row>, layoutB = #nvvm.mma_layout<row>, m = 16 : i32, n = 16 : i32}
: (i32, i32, i32, i32, i32, i32, i32, i32, f32, f32, f32, f32, f32, f32, f32, f32)
-> !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32)>
llvm.return %r : !llvm.struct<(f32, f32, f32, f32, f32, f32, f32, f32)>
}
// CHECK-LABEL: @cp_async
llvm.func @cp_async(%arg0: !llvm.ptr<3>, %arg1: !llvm.ptr<1>) {
// CHECK: nvvm.cp.async.shared.global %{{.*}}, %{{.*}}, 16, cache = ca
nvvm.cp.async.shared.global %arg0, %arg1, 16, cache = ca : !llvm.ptr<3>, !llvm.ptr<1>
// CHECK: nvvm.cp.async.shared.global %{{.*}}, %{{.*}}, 16, cache = cg
nvvm.cp.async.shared.global %arg0, %arg1, 16, cache = cg : !llvm.ptr<3>, !llvm.ptr<1>
// CHECK: nvvm.cp.async.commit.group
nvvm.cp.async.commit.group
// CHECK: nvvm.cp.async.wait.group 0
nvvm.cp.async.wait.group 0
llvm.return
}
// CHECK-LABEL: llvm.func @ld_matrix
llvm.func @ld_matrix(%arg0: !llvm.ptr<3>) {
// CHECK: nvvm.ldmatrix %{{.*}} {layout = #nvvm.mma_layout<row>, num = 1 : i32} : (!llvm.ptr<3>) -> i32
%l1 = nvvm.ldmatrix %arg0 {num = 1 : i32, layout = #nvvm.mma_layout<row>} : (!llvm.ptr<3>) -> i32
// CHECK: nvvm.ldmatrix %{{.*}} {layout = #nvvm.mma_layout<row>, num = 2 : i32} : (!llvm.ptr<3>) -> !llvm.struct<(i32, i32)>
%l2 = nvvm.ldmatrix %arg0 {num = 2 : i32, layout = #nvvm.mma_layout<row>} : (!llvm.ptr<3>) -> !llvm.struct<(i32, i32)>
// CHECK: nvvm.ldmatrix %{{.*}} {layout = #nvvm.mma_layout<row>, num = 4 : i32} : (!llvm.ptr<3>) -> !llvm.struct<(i32, i32, i32, i32)>
%l4 = nvvm.ldmatrix %arg0 {num = 4 : i32, layout = #nvvm.mma_layout<row>} : (!llvm.ptr<3>) -> !llvm.struct<(i32, i32, i32, i32)>
llvm.return
}
// CHECK-LABEL: llvm.func @redux_sync
llvm.func @redux_sync(%value : i32, %offset : i32) -> i32 {
// CHECK: nvvm.redux.sync add %{{.*}}
%r1 = nvvm.redux.sync add %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync max %{{.*}}
%r2 = nvvm.redux.sync max %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync min %{{.*}}
%r3 = nvvm.redux.sync min %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync umax %{{.*}}
%r5 = nvvm.redux.sync umax %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync umin %{{.*}}
%r6 = nvvm.redux.sync umin %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync and %{{.*}}
%r7 = nvvm.redux.sync and %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync or %{{.*}}
%r8 = nvvm.redux.sync or %value, %offset : i32 -> i32
// CHECK: nvvm.redux.sync xor %{{.*}}
%r9 = nvvm.redux.sync xor %value, %offset : i32 -> i32
llvm.return %r1 : i32
}
// -----
// expected-error@below {{attribute attached to unexpected op}}
func.func private @expected_llvm_func() attributes { nvvm.kernel }
// -----
llvm.func private @mbarrier_init_generic(%barrier: !llvm.ptr) {
%count = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.mbarrier.init %{{.*}}, %{{.*}} : !llvm.ptr, i32
nvvm.mbarrier.init %barrier, %count : !llvm.ptr, i32
llvm.return
}
llvm.func private @mbarrier_init_shared(%barrier: !llvm.ptr<3>) {
%count = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.mbarrier.init.shared %{{.*}}, %{{.*}} : !llvm.ptr<3>, i32
nvvm.mbarrier.init.shared %barrier, %count : !llvm.ptr<3>, i32
llvm.return
}
llvm.func private @mbarrier_inval_generic(%barrier: !llvm.ptr) {
// CHECK: nvvm.mbarrier.inval %{{.*}} : !llvm.ptr
nvvm.mbarrier.inval %barrier : !llvm.ptr
llvm.return
}
llvm.func private @mbarrier_inval_shared(%barrier: !llvm.ptr<3>) {
// CHECK: nvvm.mbarrier.inval.shared %{{.*}} : !llvm.ptr<3>
nvvm.mbarrier.inval.shared %barrier : !llvm.ptr<3>
llvm.return
}
llvm.func private @mbarrier_arrive(%barrier: !llvm.ptr) {
// CHECK: nvvm.mbarrier.arrive %{{.*}} : !llvm.ptr
%0 = nvvm.mbarrier.arrive %barrier : !llvm.ptr -> i64
llvm.return
}
llvm.func private @mbarrier_arrive_shared(%barrier: !llvm.ptr<3>) {
// CHECK: nvvm.mbarrier.arrive.shared %{{.*}} : !llvm.ptr<3>
%0 = nvvm.mbarrier.arrive.shared %barrier : !llvm.ptr<3> -> i64
llvm.return
}
llvm.func private @mbarrier_arrive_nocomplete(%barrier: !llvm.ptr) {
%count = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.mbarrier.arrive.nocomplete %{{.*}} : !llvm.ptr
%0 = nvvm.mbarrier.arrive.nocomplete %barrier, %count : !llvm.ptr, i32 -> i64
llvm.return
}
llvm.func private @mbarrier_arrive_nocomplete_shared(%barrier: !llvm.ptr<3>) {
%count = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.mbarrier.arrive.nocomplete.shared %{{.*}} : !llvm.ptr<3>
%0 = nvvm.mbarrier.arrive.nocomplete.shared %barrier, %count : !llvm.ptr<3>, i32 -> i64
llvm.return
}
llvm.func private @mbarrier_test_wait(%barrier: !llvm.ptr, %token : i64) -> i1 {
// CHECK: nvvm.mbarrier.test.wait %{{.*}}
%isComplete = nvvm.mbarrier.test.wait %barrier, %token : !llvm.ptr, i64 -> i1
llvm.return %isComplete : i1
}
llvm.func private @mbarrier_test_wait_shared(%barrier: !llvm.ptr<3>, %token : i64) {
%count = nvvm.read.ptx.sreg.ntid.x : i32
// CHECK: nvvm.mbarrier.test.wait.shared %{{.*}}
%isComplete = nvvm.mbarrier.test.wait.shared %barrier, %token : !llvm.ptr<3>, i64 -> i1
llvm.return
}
// CHECK-LABEL : @wgmma_fence_aligned
func.func @wgmma_fence_aligned() {
// CHECK : nvvm.wgmma.fence.aligned
nvvm.wgmma.fence.aligned
return
}
// CHECK-LABEL : @wgmma_commit_group_sync_aligned
func.func @wgmma_commit_group_sync_aligned() {
// CHECK : nvvm.wgmma.commit.group.sync.aligned
nvvm.wgmma.commit.group.sync.aligned
return
}
// CHECK-LABEL : @wgmma_commit_group_sync_aligned
func.func @wgmma_wait_group_sync_aligned() {
// CHECK : nvvm.wgmma.wait.group.sync.aligned
nvvm.wgmma.wait.group.sync.aligned 0
return
}
|
