llvm/test/CodeGen/AMDGPU/propagate-waves-per-eu.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes --check-globals --version 2
   2 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -amdgpu-attributor %s | FileCheck %s
   3
   4 ; Check propagation of amdgpu-flat-work-group-size attribute.
   5
   6 ; Called from a single kernel with 1,8
   7 define internal void @default_to_1_8_a() {
   8 ; CHECK-LABEL: define internal void @default_to_1_8_a
   9 ; CHECK-SAME: () #[[ATTR0:[0-9]+]] {
  10 ; CHECK-NEXT:    ret void
  11 ;
  12   ret void
  13 }
  14
  15 define amdgpu_kernel void @kernel_1_8() #0 {
  16 ; CHECK-LABEL: define amdgpu_kernel void @kernel_1_8
  17 ; CHECK-SAME: () #[[ATTR0]] {
  18 ; CHECK-NEXT:    call void @default_to_1_8_a()
  19 ; CHECK-NEXT:    ret void
  20 ;
  21   call void @default_to_1_8_a()
  22   ret void
  23 }
  24
  25 ; Called from a single kernel with 1,2
  26 define internal void @default_to_1_2() {
  27 ; CHECK-LABEL: define internal void @default_to_1_2
  28 ; CHECK-SAME: () #[[ATTR1:[0-9]+]] {
  29 ; CHECK-NEXT:    ret void
  30 ;
  31   ret void
  32 }
  33
  34 define amdgpu_kernel void @kernel_1_2() #1 {
  35 ; CHECK-LABEL: define amdgpu_kernel void @kernel_1_2
  36 ; CHECK-SAME: () #[[ATTR1]] {
  37 ; CHECK-NEXT:    call void @default_to_1_2()
  38 ; CHECK-NEXT:    call void @flat_group_1_1()
  39 ; CHECK-NEXT:    call void @default_to_1_8_b()
  40 ; CHECK-NEXT:    call void @flat_group_2_8()
  41 ; CHECK-NEXT:    ret void
  42 ;
  43   call void @default_to_1_2()
  44   call void @flat_group_1_1()
  45   call void @default_to_1_8_b()
  46   call void @flat_group_2_8()
  47   ret void
  48 }
  49
  50 ; Called from a single kernel with 1,4
  51 define internal void @default_to_1_4() {
  52 ; CHECK-LABEL: define internal void @default_to_1_4
  53 ; CHECK-SAME: () #[[ATTR2:[0-9]+]] {
  54 ; CHECK-NEXT:    ret void
  55 ;
  56   ret void
  57 }
  58
  59 define amdgpu_kernel void @kernel_1_4() #2 {
  60 ; CHECK-LABEL: define amdgpu_kernel void @kernel_1_4
  61 ; CHECK-SAME: () #[[ATTR2]] {
  62 ; CHECK-NEXT:    call void @default_to_1_4()
  63 ; CHECK-NEXT:    ret void
  64 ;
  65   call void @default_to_1_4()
  66   ret void
  67 }
  68
  69 ; Called from kernels with 2,9 and 9,9
  70 define internal void @default_to_2_9() {
  71 ; CHECK-LABEL: define internal void @default_to_2_9
  72 ; CHECK-SAME: () #[[ATTR3:[0-9]+]] {
  73 ; CHECK-NEXT:    ret void
  74 ;
  75   ret void
  76 }
  77
  78 ; This already has strict bounds, but called from kernels with wider
  79 ; bounds, and should not be changed.
  80 define internal void @flat_group_1_1() #3 {
  81 ; CHECK-LABEL: define internal void @flat_group_1_1
  82 ; CHECK-SAME: () #[[ATTR4:[0-9]+]] {
  83 ; CHECK-NEXT:    ret void
  84 ;
  85   ret void
  86 }
  87
  88 ; 2,8 -> 2,2
  89 define internal void @flat_group_2_8() #4 {
  90 ; CHECK-LABEL: define internal void @flat_group_2_8
  91 ; CHECK-SAME: () #[[ATTR5:[0-9]+]] {
  92 ; CHECK-NEXT:    ret void
  93 ;
  94   ret void
  95 }
  96
  97 ; 9,10 -> 9,9
  98 define internal void @flat_group_9_10() #5 {
  99 ; CHECK-LABEL: define internal void @flat_group_9_10
 100 ; CHECK-SAME: () #[[ATTR6:[0-9]+]] {
 101 ; CHECK-NEXT:    ret void
 102 ;
 103   ret void
 104 }
 105
 106 define amdgpu_kernel void @kernel_2_9() #6 {
 107 ; CHECK-LABEL: define amdgpu_kernel void @kernel_2_9
 108 ; CHECK-SAME: () #[[ATTR3]] {
 109 ; CHECK-NEXT:    call void @default_to_2_9()
 110 ; CHECK-NEXT:    call void @flat_group_1_1()
 111 ; CHECK-NEXT:    ret void
 112 ;
 113   call void @default_to_2_9()
 114   call void @flat_group_1_1()
 115   ret void
 116 }
 117
 118 define amdgpu_kernel void @kernel_9_9() #7 {
 119 ; CHECK-LABEL: define amdgpu_kernel void @kernel_9_9
 120 ; CHECK-SAME: () #[[ATTR6]] {
 121 ; CHECK-NEXT:    call void @default_to_2_9()
 122 ; CHECK-NEXT:    call void @flat_group_9_10()
 123 ; CHECK-NEXT:    ret void
 124 ;
 125   call void @default_to_2_9()
 126   call void @flat_group_9_10()
 127   ret void
 128 }
 129
 130 ; Called from kernels with 2,8 and 1,2 => 1,8
 131 define internal void @default_to_1_8_b() {
 132 ; CHECK-LABEL: define internal void @default_to_1_8_b
 133 ; CHECK-SAME: () #[[ATTR0]] {
 134 ; CHECK-NEXT:    ret void
 135 ;
 136   ret void
 137 }
 138
 139 ; The kernel's lower bound is higher than the callee's lower bound, so
 140 ; this should probably be illegal.
 141 define amdgpu_kernel void @kernel_2_8() #4 {
 142 ; CHECK-LABEL: define amdgpu_kernel void @kernel_2_8
 143 ; CHECK-SAME: () #[[ATTR7:[0-9]+]] {
 144 ; CHECK-NEXT:    call void @default_to_1_8_a()
 145 ; CHECK-NEXT:    call void @default_to_1_8_b()
 146 ; CHECK-NEXT:    ret void
 147 ;
 148   call void @default_to_1_8_a()
 149   call void @default_to_1_8_b()
 150   ret void
 151 }
 152
 153 ; 1,2 -> 2,2
 154 define internal void @merge_cycle_0() #1 {
 155 ; CHECK-LABEL: define internal void @merge_cycle_0
 156 ; CHECK-SAME: () #[[ATTR5]] {
 157 ; CHECK-NEXT:    call void @merge_cycle_1()
 158 ; CHECK-NEXT:    ret void
 159 ;
 160   call void @merge_cycle_1()
 161   ret void
 162 }
 163
 164 ; Called from 1,2 + 3,8
 165 ; 2,8 -> 2,8
 166 define internal void @merge_cycle_1() #4 {
 167 ; CHECK-LABEL: define internal void @merge_cycle_1
 168 ; CHECK-SAME: () #[[ATTR7]] {
 169 ; CHECK-NEXT:    call void @merge_cycle_0()
 170 ; CHECK-NEXT:    ret void
 171 ;
 172   call void @merge_cycle_0()
 173   ret void
 174 }
 175
 176 define amdgpu_kernel void @kernel_3_8() #8 {
 177 ; CHECK-LABEL: define amdgpu_kernel void @kernel_3_8
 178 ; CHECK-SAME: () #[[ATTR8:[0-9]+]] {
 179 ; CHECK-NEXT:    call void @merge_cycle_0()
 180 ; CHECK-NEXT:    call void @default_captured_address()
 181 ; CHECK-NEXT:    call void @externally_visible_default()
 182 ; CHECK-NEXT:    [[F32:%.*]] = call float @bitcasted_function()
 183 ; CHECK-NEXT:    ret void
 184 ;
 185   call void @merge_cycle_0()
 186   call void @default_captured_address()
 187   call void @externally_visible_default()
 188   %f32 = call float @bitcasted_function()
 189   ret void
 190 }
 191
 192 define internal void @default_captured_address() {
 193 ; CHECK-LABEL: define internal void @default_captured_address
 194 ; CHECK-SAME: () #[[ATTR9:[0-9]+]] {
 195 ; CHECK-NEXT:    store volatile ptr @default_captured_address, ptr undef, align 8
 196 ; CHECK-NEXT:    ret void
 197 ;
 198   store volatile ptr @default_captured_address, ptr undef, align 8
 199   ret void
 200 }
 201
 202 define void @externally_visible_default() {
 203 ; CHECK-LABEL: define void @externally_visible_default
 204 ; CHECK-SAME: () #[[ATTR9]] {
 205 ; CHECK-NEXT:    ret void
 206 ;
 207   ret void
 208 }
 209
 210 ; 1,10 -> 3,8
 211 define internal i32 @bitcasted_function() {
 212 ; CHECK-LABEL: define internal i32 @bitcasted_function
 213 ; CHECK-SAME: () #[[ATTR8]] {
 214 ; CHECK-NEXT:    ret i32 0
 215 ;
 216   ret i32 0
 217 }
 218
 219 define internal void @called_from_invalid_bounds_0() {
 220 ; CHECK-LABEL: define internal void @called_from_invalid_bounds_0
 221 ; CHECK-SAME: () #[[ATTR10:[0-9]+]] {
 222 ; CHECK-NEXT:    ret void
 223 ;
 224   ret void
 225 }
 226
 227 define internal void @called_from_invalid_bounds_1() {
 228 ; CHECK-LABEL: define internal void @called_from_invalid_bounds_1
 229 ; CHECK-SAME: () #[[ATTR10]] {
 230 ; CHECK-NEXT:    ret void
 231 ;
 232   ret void
 233 }
 234
 235 ; Invalid range for amdgpu-waves-per-eu
 236 define amdgpu_kernel void @kernel_invalid_bounds_0_8() #9 {
 237 ; CHECK-LABEL: define amdgpu_kernel void @kernel_invalid_bounds_0_8
 238 ; CHECK-SAME: () #[[ATTR11:[0-9]+]] {
 239 ; CHECK-NEXT:    call void @called_from_invalid_bounds_0()
 240 ; CHECK-NEXT:    ret void
 241 ;
 242   call void @called_from_invalid_bounds_0()
 243   ret void
 244 }
 245
 246 ; Invalid range for amdgpu-waves-per-eu
 247 define amdgpu_kernel void @kernel_invalid_bounds_1_123() #10 {
 248 ; CHECK-LABEL: define amdgpu_kernel void @kernel_invalid_bounds_1_123
 249 ; CHECK-SAME: () #[[ATTR12:[0-9]+]] {
 250 ; CHECK-NEXT:    call void @called_from_invalid_bounds_1()
 251 ; CHECK-NEXT:    ret void
 252 ;
 253   call void @called_from_invalid_bounds_1()
 254   ret void
 255 }
 256
 257 ; XXX - Why is the maximum not 6?
 258 ; The 512 maximum workgroup size implies a minimum occupancy of 2. The
 259 ; implied minimum waves-per-eu should not be 3
 260 ; -> 2,10
 261 define void @larger_group_size_implies_lower_minimum() #11 {
 262 ; CHECK-LABEL: define void @larger_group_size_implies_lower_minimum
 263 ; CHECK-SAME: () #[[ATTR13:[0-9]+]] {
 264 ; CHECK-NEXT:    ret void
 265 ;
 266   ret void
 267 }
 268
 269 define amdgpu_kernel void @kernel_3_6() #12 {
 270 ; CHECK-LABEL: define amdgpu_kernel void @kernel_3_6
 271 ; CHECK-SAME: () #[[ATTR14:[0-9]+]] {
 272 ; CHECK-NEXT:    call void @larger_group_size_implies_lower_minimum()
 273 ; CHECK-NEXT:    ret void
 274 ;
 275   call void @larger_group_size_implies_lower_minimum()
 276   ret void
 277 }
 278
 279 ; 3,6 -> 6,9
 280 define internal void @refine_upper_func_3_6() #13 {
 281 ; CHECK-LABEL: define internal void @refine_upper_func_3_6
 282 ; CHECK-SAME: () #[[ATTR15:[0-9]+]] {
 283 ; CHECK-NEXT:    ret void
 284 ;
 285   ret void
 286 }
 287
 288 ; 4,8 -> 6,8
 289 define internal void @refine_lower_func_4_8() #14 {
 290 ; CHECK-LABEL: define internal void @refine_lower_func_4_8
 291 ; CHECK-SAME: () #[[ATTR16:[0-9]+]] {
 292 ; CHECK-NEXT:    call void @refine_upper_func_3_6()
 293 ; CHECK-NEXT:    ret void
 294 ;
 295   call void @refine_upper_func_3_6()
 296   ret void
 297 }
 298
 299 define amdgpu_kernel void @kernel_foo_6_8() #15 {
 300 ; CHECK-LABEL: define amdgpu_kernel void @kernel_foo_6_8
 301 ; CHECK-SAME: () #[[ATTR16]] {
 302 ; CHECK-NEXT:    call void @refine_upper_func_3_6()
 303 ; CHECK-NEXT:    call void @refine_lower_func_4_8()
 304 ; CHECK-NEXT:    call void @func_9_10_a()
 305 ; CHECK-NEXT:    ret void
 306 ;
 307   call void @refine_upper_func_3_6()
 308   call void @refine_lower_func_4_8()
 309   call void @func_9_10_a()
 310   ret void
 311 }
 312
 313 ; 5,5 -> 5,5
 314 define internal void @func_5_5() #16 {
 315 ; CHECK-LABEL: define internal void @func_5_5
 316 ; CHECK-SAME: () #[[ATTR17:[0-9]+]] {
 317 ; CHECK-NEXT:    ret void
 318 ;
 319   ret void
 320 }
 321
 322 ; 5,8 -> 8,8
 323 define internal void @func_5_8() #17 {
 324 ; CHECK-LABEL: define internal void @func_5_8
 325 ; CHECK-SAME: () #[[ATTR18:[0-9]+]] {
 326 ; CHECK-NEXT:    ret void
 327 ;
 328   ret void
 329 }
 330
 331 ; 9,10 -> 9,10
 332 define internal void @func_9_10_a() #18 {
 333 ; CHECK-LABEL: define internal void @func_9_10_a
 334 ; CHECK-SAME: () #[[ATTR19:[0-9]+]] {
 335 ; CHECK-NEXT:    ret void
 336 ;
 337   ret void
 338 }
 339
 340 ; 9,10 -> 9,9
 341 define internal void @func_9_10_b() #18 {
 342 ; CHECK-LABEL: define internal void @func_9_10_b
 343 ; CHECK-SAME: () #[[ATTR20:[0-9]+]] {
 344 ; CHECK-NEXT:    ret void
 345 ;
 346   ret void
 347 }
 348
 349 define amdgpu_kernel void @kernel_bar_8_9() #19 {
 350 ; CHECK-LABEL: define amdgpu_kernel void @kernel_bar_8_9
 351 ; CHECK-SAME: () #[[ATTR21:[0-9]+]] {
 352 ; CHECK-NEXT:    call void @refine_upper_func_3_6()
 353 ; CHECK-NEXT:    call void @func_5_5()
 354 ; CHECK-NEXT:    call void @func_9_10_b()
 355 ; CHECK-NEXT:    call void @func_5_8()
 356 ; CHECK-NEXT:    call void @externally_visible()
 357 ; CHECK-NEXT:    ret void
 358 ;
 359   call void @refine_upper_func_3_6()
 360   call void @func_5_5()
 361   call void @func_9_10_b()
 362   call void @func_5_8()
 363   call void @externally_visible()
 364   ret void
 365 }
 366
 367 ; This is an optimization hint based on users, so it's not strictly
 368 ; required that all callers be visible.
 369 define void @externally_visible() {
 370 ; CHECK-LABEL: define void @externally_visible
 371 ; CHECK-SAME: () #[[ATTR9]] {
 372 ; CHECK-NEXT:    ret void
 373 ;
 374   ret void
 375 }
 376
 377
 378 ; Use a 1 wave workgroup so there is no interaction by the workgroup
 379 ; size on the implied waves per EU.
 380
 381 attributes #0 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="1,8" }
 382 attributes #1 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="1,2" }
 383 attributes #2 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="1,4" }
 384 attributes #3 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="1,1" }
 385 attributes #4 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="2,8" }
 386 attributes #5 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="9,10" }
 387 attributes #6 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="2,9" }
 388 attributes #7 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="9,9" }
 389 attributes #8 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="3,8" }
 390 attributes #9 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="0,8" }
 391 attributes #10 = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-waves-per-eu"="1,123" }
 392 attributes #11 = { "amdgpu-flat-work-group-size"="1,512" }
 393 attributes #12 = { "amdgpu-flat-work-group-size"="1,512" "amdgpu-waves-per-eu"="3,6" }
 394 attributes #13 = { "amdgpu-waves-per-eu"="3,6" }
 395 attributes #14 = { "amdgpu-waves-per-eu"="4,8" }
 396 attributes #15 = { "amdgpu-waves-per-eu"="6,8" }
 397 attributes #16 = { "amdgpu-waves-per-eu"="5,5" }
 398 attributes #17 = { "amdgpu-waves-per-eu"="5,8" }
 399 attributes #18 = { "amdgpu-waves-per-eu"="9,10" }
 400 attributes #19 = { "amdgpu-waves-per-eu"="8,9" }
 401 ;.
 402 ; CHECK: attributes #[[ATTR0]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="1,8" "uniform-work-group-size"="false" }
 403 ; CHECK: attributes #[[ATTR1]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="1,2" "uniform-work-group-size"="false" }
 404 ; CHECK: attributes #[[ATTR2]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="1,4" "uniform-work-group-size"="false" }
 405 ; CHECK: attributes #[[ATTR3]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="2,9" "uniform-work-group-size"="false" }
 406 ; CHECK: attributes #[[ATTR4]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="1,1" "uniform-work-group-size"="false" }
 407 ; CHECK: attributes #[[ATTR5]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="2,2" "uniform-work-group-size"="false" }
 408 ; CHECK: attributes #[[ATTR6]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="9,9" "uniform-work-group-size"="false" }
 409 ; CHECK: attributes #[[ATTR7]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="2,8" "uniform-work-group-size"="false" }
 410 ; CHECK: attributes #[[ATTR8]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="3,8" "uniform-work-group-size"="false" }
 411 ; CHECK: attributes #[[ATTR9]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" }
 412 ; CHECK: attributes #[[ATTR10]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "uniform-work-group-size"="false" }
 413 ; CHECK: attributes #[[ATTR11]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="0,8" "uniform-work-group-size"="false" }
 414 ; CHECK: attributes #[[ATTR12]] = { "amdgpu-flat-work-group-size"="1,64" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="1,123" "uniform-work-group-size"="false" }
 415 ; CHECK: attributes #[[ATTR13]] = { "amdgpu-flat-work-group-size"="1,512" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="2,10" "uniform-work-group-size"="false" }
 416 ; CHECK: attributes #[[ATTR14]] = { "amdgpu-flat-work-group-size"="1,512" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="3,6" "uniform-work-group-size"="false" }
 417 ; CHECK: attributes #[[ATTR15]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="6,9" "uniform-work-group-size"="false" }
 418 ; CHECK: attributes #[[ATTR16]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="6,8" "uniform-work-group-size"="false" }
 419 ; CHECK: attributes #[[ATTR17]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="5,5" "uniform-work-group-size"="false" }
 420 ; CHECK: attributes #[[ATTR18]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="8,8" "uniform-work-group-size"="false" }
 421 ; CHECK: attributes #[[ATTR19]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="9,10" "uniform-work-group-size"="false" }
 422 ; CHECK: attributes #[[ATTR20]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="9,9" "uniform-work-group-size"="false" }
 423 ; CHECK: attributes #[[ATTR21]] = { "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-heap-ptr" "amdgpu-no-hostcall-ptr" "amdgpu-no-implicitarg-ptr" "amdgpu-no-lds-kernel-id" "amdgpu-no-multigrid-sync-arg" "amdgpu-no-queue-ptr" "amdgpu-no-workgroup-id-x" "amdgpu-no-workgroup-id-y" "amdgpu-no-workgroup-id-z" "amdgpu-no-workitem-id-x" "amdgpu-no-workitem-id-y" "amdgpu-no-workitem-id-z" "amdgpu-waves-per-eu"="8,9" "uniform-work-group-size"="false" }
 424 ;.