151 lines · plain
1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 62// REQUIRES: amdgpu-registered-target3// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1100 -target-feature +wavefrontsize64 -DWMMA_GFX1100_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX11004 5typedef float v4f __attribute__((ext_vector_type(4)));6typedef half v8h __attribute__((ext_vector_type(8)));7typedef half v16h __attribute__((ext_vector_type(16)));8typedef int v2i __attribute__((ext_vector_type(2)));9typedef int v4i __attribute__((ext_vector_type(4)));10typedef short v8s __attribute__((ext_vector_type(8)));11typedef short v16s __attribute__((ext_vector_type(16)));12 13#ifdef WMMA_GFX1100_TESTS14 15// Wave6416 17//18// amdgcn_wmma_f32_16x16x16_f1619//20 21// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f32_16x16x16_f16_w64(22// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {23// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]24// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16.v4f32.v16f16(<16 x half> [[A]], <16 x half> [[B]], <4 x float> [[C]])25// CHECK-GFX1100-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8:![0-9]+]]26// CHECK-GFX1100-NEXT: ret void27//28void test_amdgcn_wmma_f32_16x16x16_f16_w64(global v4f* out, v16h a, v16h b, v4f c)29{30 *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w64(a, b, c);31}32 33//34// amdgcn_wmma_f32_16x16x16_bf1635//36 37// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f32_16x16x16_bf16_w64(38// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <4 x float> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {39// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]40// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v4f32.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <4 x float> [[C]])41// CHECK-GFX1100-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]42// CHECK-GFX1100-NEXT: ret void43//44void test_amdgcn_wmma_f32_16x16x16_bf16_w64(global v4f* out, v16s a, v16s b, v4f c)45{46 *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w64(a, b, c);47}48 49//50// amdgcn_wmma_f16_16x16x16_f1651//52 53// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f16_16x16x16_f16_w64(54// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {55// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]56// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.v8f16.v16f16(<16 x half> [[A]], <16 x half> [[B]], <8 x half> [[C]], i1 true)57// CHECK-GFX1100-NEXT: store <8 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]58// CHECK-GFX1100-NEXT: ret void59//60void test_amdgcn_wmma_f16_16x16x16_f16_w64(global v8h* out, v16h a, v16h b, v8h c)61{62 *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w64(a, b, c, true);63}64 65//66// amdgcn_wmma_bf16_16x16x16_bf1667//68 69// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_bf16_16x16x16_bf16_w64(70// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <8 x i16> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {71// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]72// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v8i16.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <8 x i16> [[C]], i1 true)73// CHECK-GFX1100-NEXT: store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]74// CHECK-GFX1100-NEXT: ret void75//76void test_amdgcn_wmma_bf16_16x16x16_bf16_w64(global v8s* out, v16s a, v16s b, v8s c)77{78 *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w64(a, b, c, true);79}80 81//82// amdgcn_wmma_f16_16x16x16_f16_tied83//84 85// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f16_16x16x16_f16_tied_w64(86// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {87// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]88// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.tied.v8f16.v16f16(<16 x half> [[A]], <16 x half> [[B]], <8 x half> [[C]], i1 true)89// CHECK-GFX1100-NEXT: store <8 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]90// CHECK-GFX1100-NEXT: ret void91//92void test_amdgcn_wmma_f16_16x16x16_f16_tied_w64(global v8h* out, v16h a, v16h b, v8h c)93{94 *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w64(a, b, c, true);95}96 97//98// amdgcn_wmma_bf16_16x16x16_bf16_tied99//100 101// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64(102// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <8 x i16> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {103// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]104// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.tied.v8i16.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <8 x i16> [[C]], i1 true)105// CHECK-GFX1100-NEXT: store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]106// CHECK-GFX1100-NEXT: ret void107//108void test_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64(global v8s* out, v16s a, v16s b, v8s c)109{110 *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w64(a, b, c, true);111}112 113//114// amdgcn_wmma_i32_16x16x16_iu8115//116 117// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_i32_16x16x16_iu8_w64(118// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <4 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <4 x i32> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {119// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]120// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu8.v4i32.v4i32(i1 true, <4 x i32> [[A]], i1 true, <4 x i32> [[B]], <4 x i32> [[C]], i1 false)121// CHECK-GFX1100-NEXT: store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]122// CHECK-GFX1100-NEXT: ret void123//124void test_amdgcn_wmma_i32_16x16x16_iu8_w64(global v4i* out, v4i a, v4i b, v4i c)125{126 *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w64(true, a, true, b, c, false);127}128 129//130// amdgcn_wmma_i32_16x16x16_iu4131//132 133// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_i32_16x16x16_iu4_w64(134// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <4 x i32> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {135// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]136// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu4.v4i32.v2i32(i1 true, <2 x i32> [[A]], i1 true, <2 x i32> [[B]], <4 x i32> [[C]], i1 false)137// CHECK-GFX1100-NEXT: store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]138// CHECK-GFX1100-NEXT: ret void139//140void test_amdgcn_wmma_i32_16x16x16_iu4_w64(global v4i* out, v2i a, v2i b, v4i c)141{142 *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w64(true, a, true, b, c, false);143}144 145#endif146//.147// CHECK-GFX1100: [[META6:![0-9]+]] = !{!"omnipotent char", [[META7:![0-9]+]], i64 0}148// CHECK-GFX1100: [[META7]] = !{!"Simple C/C++ TBAA"}149// CHECK-GFX1100: [[CHAR_TBAA8]] = !{[[META6]], [[META6]], i64 0}150//.151