llvm/test/CodeGen/SPIRV/transcoding/OpenCL/work_group_barrier.ll

   1 ; RUN: llc -O0 -mtriple=spirv32-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV
   2
   3 ;; This test checks that the backend is capable to correctly translate
   4 ;; sub_group_barrier built-in function [1] from cl_khr_subgroups extension into
   5 ;; corresponding SPIR-V instruction.
   6
   7 ;; __kernel void test_barrier_const_flags() {
   8 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE);
   9 ;;   work_group_barrier(CLK_GLOBAL_MEM_FENCE);
  10 ;;   work_group_barrier(CLK_IMAGE_MEM_FENCE);
  11 ;;
  12 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE | CLK_GLOBAL_MEM_FENCE);
  13 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE | CLK_IMAGE_MEM_FENCE);
  14 ;;   work_group_barrier(CLK_GLOBAL_MEM_FENCE | CLK_LOCAL_MEM_FENCE | CLK_IMAGE_MEM_FENCE);
  15 ;;
  16 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE, memory_scope_work_item);
  17 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE, memory_scope_work_group);
  18 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE, memory_scope_device);
  19 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE, memory_scope_all_svm_devices);
  20 ;;   work_group_barrier(CLK_LOCAL_MEM_FENCE, memory_scope_sub_group);
  21 ;;
  22   ;; barrier should also work (preserved for backward compatibility)
  23 ;;   barrier(CLK_GLOBAL_MEM_FENCE);
  24 ;; }
  25 ;;
  26 ;; __kernel void test_barrier_non_const_flags(cl_mem_fence_flags flags, memory_scope scope) {
  27   ;; FIXME: OpenCL spec doesn't require flags to be compile-time known
  28   ;; work_group_barrier(flags);
  29   ;; work_group_barrier(flags, scope);
  30 ;; }
  31
  32 ; CHECK-SPIRV: OpName %[[#TEST_CONST_FLAGS:]] "test_barrier_const_flags"
  33 ; CHECK-SPIRV: %[[#UINT:]] = OpTypeInt 32 0
  34
  35 ;; In SPIR-V, barrier is represented as OpControlBarrier [2] and OpenCL
  36 ;; cl_mem_fence_flags are represented as part of Memory Semantics [3], which
  37 ;; also includes memory order constraints. The backend applies some default
  38 ;; memory order for OpControlBarrier and therefore, constants below include a
  39 ;; bit more information than original source
  40
  41 ;; 0x10 SequentiallyConsistent + 0x100 WorkgroupMemory
  42 ; CHECK-SPIRV-DAG: %[[#LOCAL:]] = OpConstant %[[#UINT]] 272
  43 ;; 0x10 SequentiallyConsistent + 0x200 CrossWorkgroupMemory
  44 ; CHECK-SPIRV-DAG: %[[#GLOBAL:]] = OpConstant %[[#UINT]] 528
  45 ;; 0x10 SequentiallyConsistent + 0x800 ImageMemory
  46 ; CHECK-SPIRV-DAG: %[[#IMAGE:]] = OpConstant %[[#UINT]] 2064
  47 ;; 0x10 SequentiallyConsistent + 0x100 WorkgroupMemory + 0x200 CrossWorkgroupMemory
  48 ; CHECK-SPIRV-DAG: %[[#LOCAL_GLOBAL:]] = OpConstant %[[#UINT]] 784
  49 ;; 0x10 SequentiallyConsistent + 0x100 WorkgroupMemory + 0x800 ImageMemory
  50 ; CHECK-SPIRV-DAG: %[[#LOCAL_IMAGE:]] = OpConstant %[[#UINT]] 2320
  51 ;; 0x10 SequentiallyConsistent + 0x100 WorkgroupMemory + 0x200 CrossWorkgroupMemory + 0x800 ImageMemory
  52 ; CHECK-SPIRV-DAG: %[[#LOCAL_GLOBAL_IMAGE:]] = OpConstant %[[#UINT]] 2832
  53
  54 ;; Scopes [4]:
  55 ;; 2 Workgroup
  56 ; CHECK-SPIRV-DAG: %[[#SCOPE_WORK_GROUP:]] = OpConstant %[[#UINT]] 2
  57 ;; 4 Invocation
  58 ; CHECK-SPIRV-DAG: %[[#SCOPE_INVOCATION:]] = OpConstant %[[#UINT]] 4
  59 ;; 1 Device
  60 ; CHECK-SPIRV-DAG: %[[#SCOPE_DEVICE:]] = OpConstant %[[#UINT]] 1
  61 ;; 0 CrossDevice
  62 ; CHECK-SPIRV-DAG: %[[#SCOPE_CROSS_DEVICE:]] = OpConstant %[[#UINT]] 0
  63 ;; 3 Subgroup
  64 ; CHECK-SPIRV-DAG: %[[#SCOPE_SUBGROUP:]] = OpConstant %[[#UINT]] 3
  65
  66 ; CHECK-SPIRV: %[[#TEST_CONST_FLAGS]] = OpFunction %[[#]]
  67 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#LOCAL]]
  68 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#GLOBAL]]
  69 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#IMAGE]]
  70 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#LOCAL_GLOBAL]]
  71 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#LOCAL_IMAGE]]
  72 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#LOCAL_GLOBAL_IMAGE]]
  73 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_INVOCATION]] %[[#LOCAL]]
  74 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#LOCAL]]
  75 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_DEVICE]] %[[#LOCAL]]
  76 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_CROSS_DEVICE]] %[[#LOCAL]]
  77 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_SUBGROUP]] %[[#LOCAL]]
  78 ; CHECK-SPIRV: OpControlBarrier %[[#SCOPE_WORK_GROUP]] %[[#SCOPE_WORK_GROUP]] %[[#GLOBAL]]
  79
  80 define dso_local spir_kernel void @test_barrier_const_flags() local_unnamed_addr {
  81 entry:
  82   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 1)
  83   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 2)
  84   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 4)
  85   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 3)
  86   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 5)
  87   tail call spir_func void @_Z18work_group_barrierj(i32 noundef 7)
  88   tail call spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef 1, i32 noundef 0)
  89   tail call spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef 1, i32 noundef 1)
  90   tail call spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef 1, i32 noundef 2)
  91   tail call spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef 1, i32 noundef 3)
  92   tail call spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef 1, i32 noundef 4)
  93   tail call spir_func void @_Z7barrierj(i32 noundef 2)
  94   ret void
  95 }
  96
  97 declare spir_func void @_Z18work_group_barrierj(i32 noundef) local_unnamed_addr
  98
  99 declare spir_func void @_Z18work_group_barrierj12memory_scope(i32 noundef, i32 noundef) local_unnamed_addr
 100
 101 declare spir_func void @_Z7barrierj(i32 noundef) local_unnamed_addr
 102
 103 define dso_local spir_kernel void @test_barrier_non_const_flags(i32 noundef %flags, i32 noundef %scope) local_unnamed_addr {
 104 entry:
 105   ret void
 106 }
 107
 108 ;; References:
 109 ;; [1]: https://www.khronos.org/registry/OpenCL/sdk/2.0/docs/man/xhtml/work_group_barrier.html
 110 ;; [2]: https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#OpControlBarrier
 111 ;; [3]: https://www.khronos.org/registry/spir-v/specs/unified1/SPIRV.html#_a_id_memory_semantics__id_a_memory_semantics_lt_id_gt