llvm/test/CodeGen/AMDGPU/propagate-flat-work-group-size.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --check-globals
   2 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -amdgpu-attributor %s | FileCheck %s
   3 ; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-attributor %s | FileCheck %s
   4
   5 ; Check propagation of amdgpu-flat-work-group-size attribute.
   6
   7 ; Called from a single kernel with 1,256
   8 define internal void @default_to_1_256() {
   9 ; CHECK-LABEL: define {{[^@]+}}@default_to_1_256
  10 ; CHECK-SAME: () #[[ATTR0:[0-9]+]] {
  11 ; CHECK-NEXT:    ret void
  12 ;
  13   ret void
  14 }
  15
  16 define amdgpu_kernel void @kernel_1_256() #0 {
  17 ; CHECK-LABEL: define {{[^@]+}}@kernel_1_256
  18 ; CHECK-SAME: () #[[ATTR0]] {
  19 ; CHECK-NEXT:    call void @default_to_1_256()
  20 ; CHECK-NEXT:    ret void
  21 ;
  22   call void @default_to_1_256()
  23   ret void
  24 }
  25
  26 ; Called from a single kernel with 64,128
  27 define internal void @default_to_64_128() {
  28 ; CHECK-LABEL: define {{[^@]+}}@default_to_64_128
  29 ; CHECK-SAME: () #[[ATTR1:[0-9]+]] {
  30 ; CHECK-NEXT:    ret void
  31 ;
  32   ret void
  33 }
  34
  35 define amdgpu_kernel void @kernel_64_128() #1 {
  36 ; CHECK-LABEL: define {{[^@]+}}@kernel_64_128
  37 ; CHECK-SAME: () #[[ATTR1]] {
  38 ; CHECK-NEXT:    call void @default_to_64_128()
  39 ; CHECK-NEXT:    call void @flat_group_64_64()
  40 ; CHECK-NEXT:    call void @default_to_64_256()
  41 ; CHECK-NEXT:    call void @flat_group_128_256()
  42 ; CHECK-NEXT:    ret void
  43 ;
  44   call void @default_to_64_128()
  45   call void @flat_group_64_64()
  46   call void @default_to_64_256()
  47   call void @flat_group_128_256()
  48   ret void
  49 }
  50
  51 ; Called from kernels with 128,512 and 512,512
  52 define internal void @default_to_128_512() {
  53 ; CHECK-LABEL: define {{[^@]+}}@default_to_128_512
  54 ; CHECK-SAME: () #[[ATTR2:[0-9]+]] {
  55 ; CHECK-NEXT:    ret void
  56 ;
  57   ret void
  58 }
  59
  60 ; This already has a strict bounds, but called from kernels with wider
  61 ; bounds, and should not be changed.
  62 define internal void @flat_group_64_64() #2 {
  63 ; CHECK-LABEL: define {{[^@]+}}@flat_group_64_64
  64 ; CHECK-SAME: () #[[ATTR3:[0-9]+]] {
  65 ; CHECK-NEXT:    ret void
  66 ;
  67   ret void
  68 }
  69
  70 ; 128,256 -> 128,128
  71 define internal void @flat_group_128_256() #3 {
  72 ; CHECK-LABEL: define {{[^@]+}}@flat_group_128_256
  73 ; CHECK-SAME: () #[[ATTR4:[0-9]+]] {
  74 ; CHECK-NEXT:    ret void
  75 ;
  76   ret void
  77 }
  78
  79 define internal void @flat_group_512_1024() #4 {
  80 ; CHECK-LABEL: define {{[^@]+}}@flat_group_512_1024
  81 ; CHECK-SAME: () #[[ATTR5:[0-9]+]] {
  82 ; CHECK-NEXT:    ret void
  83 ;
  84   ret void
  85 }
  86
  87 define amdgpu_kernel void @kernel_128_512() #5 {
  88 ; CHECK-LABEL: define {{[^@]+}}@kernel_128_512
  89 ; CHECK-SAME: () #[[ATTR2]] {
  90 ; CHECK-NEXT:    call void @default_to_128_512()
  91 ; CHECK-NEXT:    call void @flat_group_64_64()
  92 ; CHECK-NEXT:    ret void
  93 ;
  94   call void @default_to_128_512()
  95   call void @flat_group_64_64()
  96   ret void
  97 }
  98
  99 define amdgpu_kernel void @kernel_512_512() #6 {
 100 ; CHECK-LABEL: define {{[^@]+}}@kernel_512_512
 101 ; CHECK-SAME: () #[[ATTR6:[0-9]+]] {
 102 ; CHECK-NEXT:    call void @default_to_128_512()
 103 ; CHECK-NEXT:    call void @flat_group_512_1024()
 104 ; CHECK-NEXT:    ret void
 105 ;
 106   call void @default_to_128_512()
 107   call void @flat_group_512_1024()
 108   ret void
 109 }
 110
 111 ; Called from kernels with 128,256 and 64,128 => 64,256
 112 define internal void @default_to_64_256() {
 113 ; CHECK-LABEL: define {{[^@]+}}@default_to_64_256
 114 ; CHECK-SAME: () #[[ATTR7:[0-9]+]] {
 115 ; CHECK-NEXT:    ret void
 116 ;
 117   ret void
 118 }
 119
 120 ; The kernel's lower bound is higher than the callee's lower bound, so
 121 ; this should probably be illegal.
 122 define amdgpu_kernel void @kernel_128_256() #3 {
 123 ; CHECK-LABEL: define {{[^@]+}}@kernel_128_256
 124 ; CHECK-SAME: () #[[ATTR4]] {
 125 ; CHECK-NEXT:    call void @default_to_64_256()
 126 ; CHECK-NEXT:    ret void
 127 ;
 128   call void @default_to_64_256()
 129   ret void
 130 }
 131
 132 ; 64,128 -> 64,128
 133 define internal void @merge_cycle_0() #1 {
 134 ; CHECK-LABEL: define {{[^@]+}}@merge_cycle_0
 135 ; CHECK-SAME: () #[[ATTR1]] {
 136 ; CHECK-NEXT:    call void @merge_cycle_1()
 137 ; CHECK-NEXT:    ret void
 138 ;
 139   call void @merge_cycle_1()
 140   ret void
 141 }
 142
 143 ; 128,256 -> 128,128
 144 define internal void @merge_cycle_1() #3 {
 145 ; CHECK-LABEL: define {{[^@]+}}@merge_cycle_1
 146 ; CHECK-SAME: () #[[ATTR4]] {
 147 ; CHECK-NEXT:    call void @merge_cycle_0()
 148 ; CHECK-NEXT:    ret void
 149 ;
 150   call void @merge_cycle_0()
 151   ret void
 152 }
 153
 154 define amdgpu_kernel void @kernel_64_256() #7 {
 155 ; CHECK-LABEL: define {{[^@]+}}@kernel_64_256
 156 ; CHECK-SAME: () #[[ATTR7]] {
 157 ; CHECK-NEXT:    call void @merge_cycle_0()
 158 ; CHECK-NEXT:    call void @default_captured_address()
 159 ; CHECK-NEXT:    call void @externally_visible_default()
 160 ; CHECK-NEXT:    [[F32:%.*]] = call float @bitcasted_function()
 161 ; CHECK-NEXT:    ret void
 162 ;
 163   call void @merge_cycle_0()
 164   call void @default_captured_address()
 165   call void @externally_visible_default()
 166   %f32 = call float @bitcasted_function()
 167   ret void
 168 }
 169
 170 define internal void @default_captured_address() {
 171 ; CHECK-LABEL: define {{[^@]+}}@default_captured_address
 172 ; CHECK-SAME: () #[[ATTR8:[0-9]+]] {
 173 ; CHECK-NEXT:    store volatile ptr @default_captured_address, ptr undef, align 8
 174 ; CHECK-NEXT:    ret void
 175 ;
 176   store volatile ptr @default_captured_address, ptr undef, align 8
 177   ret void
 178 }
 179
 180 define void @externally_visible_default() {
 181 ; CHECK-LABEL: define {{[^@]+}}@externally_visible_default
 182 ; CHECK-SAME: () #[[ATTR8]] {
 183 ; CHECK-NEXT:    ret void
 184 ;
 185   ret void
 186 }
 187
 188 ; 1,1024 -> 64,256
 189 define internal i32 @bitcasted_function() {
 190 ; CHECK-LABEL: define {{[^@]+}}@bitcasted_function
 191 ; CHECK-SAME: () #[[ATTR7]] {
 192 ; CHECK-NEXT:    ret i32 0
 193 ;
 194   ret i32 0
 195 }
 196
 197 attributes #0 = { "amdgpu-flat-work-group-size"="1,256" }
 198 attributes #1 = { "amdgpu-flat-work-group-size"="64,128" }
 199 attributes #2 = { "amdgpu-flat-work-group-size"="64,64" }
 200 attributes #3 = { "amdgpu-flat-work-group-size"="128,256" }
 201 attributes #4 = { "amdgpu-flat-work-group-size"="512,1024" }
 202 attributes #5 = { "amdgpu-flat-work-group-size"="128,512" }
 203 attributes #6 = { "amdgpu-flat-work-group-size"="512,512" }
 204 attributes #7 = { "amdgpu-flat-work-group-size"="64,256" }
 205 ;.
 206 ; CHECK: attributes #[[ATTR0]] = { "amdgpu-flat-work-group-size"="1,256" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 207 ; CHECK: attributes #[[ATTR1]] = { "amdgpu-flat-work-group-size"="64,128" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 208 ; CHECK: attributes #[[ATTR2]] = { "amdgpu-flat-work-group-size"="128,512" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 209 ; CHECK: attributes #[[ATTR3]] = { "amdgpu-flat-work-group-size"="64,64" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 210 ; CHECK: attributes #[[ATTR4]] = { "amdgpu-flat-work-group-size"="128,256" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 211 ; CHECK: attributes #[[ATTR5]] = { "amdgpu-flat-work-group-size"="512,1024" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 212 ; CHECK: attributes #[[ATTR6]] = { "amdgpu-flat-work-group-size"="512,512" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 213 ; CHECK: attributes #[[ATTR7]] = { "amdgpu-flat-work-group-size"="64,256" "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 214 ; CHECK: attributes #[[ATTR8]] = { "amdgpu-no-agpr" "amdgpu-no-completion-action" "amdgpu-no-default-queue" "amdgpu-no-dispatch-id" "amdgpu-no-dispatch-ptr" "amdgpu-no-flat-scratch-init" "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" }
 215 ;.