152 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 gfx1200 -target-feature +wavefrontsize32 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX12004 5typedef int v2i __attribute__((ext_vector_type(2)));6typedef int v4i __attribute__((ext_vector_type(4)));7typedef float v8f __attribute__((ext_vector_type(8)));8typedef half v8h __attribute__((ext_vector_type(8)));9typedef short v8s __attribute__((ext_vector_type(8)));10typedef int v8i __attribute__((ext_vector_type(8)));11typedef half v16h __attribute__((ext_vector_type(16)));12typedef short v16s __attribute__((ext_vector_type(16)));13 14// Wave3215 16// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_f16_w32(17// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <8 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {18// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]19// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.f16.v8f32.v8f16.v16f16.i32(<8 x half> [[A]], <16 x half> [[B]], <8 x float> [[C]], i32 [[INDEX]])20// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8:![0-9]+]]21// CHECK-GFX1200-NEXT: ret void22//23void test_amdgcn_swmmac_f32_16x16x32_f16_w32(global v8f* out, v8h a, v16h b, v8f c, int index)24{25 *out = __builtin_amdgcn_swmmac_f32_16x16x32_f16_w32(a, b, c, index);26}27 28// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf16_w32(29// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <8 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {30// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]31// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v8f32.v8i16.v16i16.i32(<8 x i16> [[A]], <16 x i16> [[B]], <8 x float> [[C]], i32 [[INDEX]])32// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]33// CHECK-GFX1200-NEXT: ret void34//35void test_amdgcn_swmmac_f32_16x16x32_bf16_w32(global v8f* out, v8s a, v16s b, v8f c, int index)36{37 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf16_w32(a, b, c, index);38}39 40// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f16_16x16x32_f16_w32(41// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <8 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <8 x half> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {42// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]43// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x half> @llvm.amdgcn.swmmac.f16.16x16x32.f16.v8f16.v8f16.v16f16.i32(<8 x half> [[A]], <16 x half> [[B]], <8 x half> [[C]], i32 [[INDEX]])44// CHECK-GFX1200-NEXT: store <8 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]45// CHECK-GFX1200-NEXT: ret void46//47void test_amdgcn_swmmac_f16_16x16x32_f16_w32(global v8h* out, v8h a, v16h b, v8h c, int index)48{49 *out = __builtin_amdgcn_swmmac_f16_16x16x32_f16_w32(a, b, c, index);50}51 52// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_bf16_16x16x32_bf16_w32(53// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 16)) [[OUT:%.*]], <8 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <8 x i16> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {54// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]55// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v8i16.v8i16.v16i16.i32(<8 x i16> [[A]], <16 x i16> [[B]], <8 x i16> [[C]], i32 [[INDEX]])56// CHECK-GFX1200-NEXT: store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 16, !tbaa [[CHAR_TBAA8]]57// CHECK-GFX1200-NEXT: ret void58//59void test_amdgcn_swmmac_bf16_16x16x32_bf16_w32(global v8s* out, v8s a, v16s b, v8s c, int index)60{61 *out = __builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w32(a, b, c, index);62}63 64// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x32_iu8_w32(65// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {66// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]67// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu8.v8i32.v2i32.v4i32.i32(i1 true, <2 x i32> [[A]], i1 true, <4 x i32> [[B]], <8 x i32> [[C]], i32 [[INDEX]], i1 true)68// CHECK-GFX1200-NEXT: store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]69// CHECK-GFX1200-NEXT: ret void70//71void test_amdgcn_swmmac_i32_16x16x32_iu8_w32(global v8i* out, v2i a, v4i b, v8i c, int index)72{73 *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu8_w32(true, a, true, b, c, index, true);74}75 76// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x32_iu4_w32(77// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], i32 noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <8 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {78// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]79// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu4.v8i32.i32.v2i32.i32(i1 true, i32 [[A]], i1 true, <2 x i32> [[B]], <8 x i32> [[C]], i32 [[INDEX]], i1 true)80// CHECK-GFX1200-NEXT: store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]81// CHECK-GFX1200-NEXT: ret void82//83void test_amdgcn_swmmac_i32_16x16x32_iu4_w32(global v8i* out, int a, v2i b, v8i c, int index)84{85 *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu4_w32(true, a, true, b, c, index, true);86}87 88// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_i32_16x16x64_iu4_w32(89// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x i32> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {90// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]91// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.swmmac.i32.16x16x64.iu4.v8i32.v2i32.v4i32.i32(i1 true, <2 x i32> [[A]], i1 true, <4 x i32> [[B]], <8 x i32> [[C]], i32 [[INDEX]], i1 true)92// CHECK-GFX1200-NEXT: store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]93// CHECK-GFX1200-NEXT: ret void94//95void test_amdgcn_swmmac_i32_16x16x64_iu4_w32(global v8i* out, v2i a, v4i b, v8i c, int index)96{97 *out = __builtin_amdgcn_swmmac_i32_16x16x64_iu4_w32(true, a, true, b, c, index, true);98}99 100// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w32(101// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {102// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]103// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.fp8.v8f32.v2i32.v4i32.i32(<2 x i32> [[A]], <4 x i32> [[B]], <8 x float> [[C]], i32 [[INDEX]])104// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]105// CHECK-GFX1200-NEXT: ret void106//107void test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w32(global v8f* out, v2i a, v4i b, v8f c, int index)108{109 *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w32(a, b, c, index);110}111 112// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w32(113// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {114// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]115// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.bf8.v8f32.v2i32.v4i32.i32(<2 x i32> [[A]], <4 x i32> [[B]], <8 x float> [[C]], i32 [[INDEX]])116// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]117// CHECK-GFX1200-NEXT: ret void118//119void test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w32(global v8f* out, v2i a, v4i b, v8f c, int index)120{121 *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w32(a, b, c, index);122}123 124// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w32(125// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {126// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]127// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.fp8.v8f32.v2i32.v4i32.i32(<2 x i32> [[A]], <4 x i32> [[B]], <8 x float> [[C]], i32 [[INDEX]])128// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]129// CHECK-GFX1200-NEXT: ret void130//131void test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w32(global v8f* out, v2i a, v4i b, v8f c, int index)132{133 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w32(a, b, c, index);134}135 136// CHECK-GFX1200-LABEL: define dso_local void @test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w32(137// CHECK-GFX1200-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]], i32 noundef [[INDEX:%.*]]) local_unnamed_addr #[[ATTR0]] {138// CHECK-GFX1200-NEXT: [[ENTRY:.*:]]139// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.bf8.v8f32.v2i32.v4i32.i32(<2 x i32> [[A]], <4 x i32> [[B]], <8 x float> [[C]], i32 [[INDEX]])140// CHECK-GFX1200-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]141// CHECK-GFX1200-NEXT: ret void142//143void test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w32(global v8f* out, v2i a, v4i b, v8f c, int index)144{145 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w32(a, b, c, index);146}147//.148// CHECK-GFX1200: [[META6:![0-9]+]] = !{!"omnipotent char", [[META7:![0-9]+]], i64 0}149// CHECK-GFX1200: [[META7]] = !{!"Simple C/C++ TBAA"}150// CHECK-GFX1200: [[CHAR_TBAA8]] = !{[[META6]], [[META6]], i64 0}151//.152