150 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 -DWMMA_GFX1100_TESTS -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX11004 5typedef float v8f __attribute__((ext_vector_type(8)));6typedef half v16h __attribute__((ext_vector_type(16)));7typedef int v2i __attribute__((ext_vector_type(2)));8typedef int v4i __attribute__((ext_vector_type(4)));9typedef int v8i __attribute__((ext_vector_type(8)));10typedef short v16s __attribute__((ext_vector_type(16)));11 12#ifdef WMMA_GFX1100_TESTS13 14// Wave3215 16//17// amdgcn_wmma_f32_16x16x16_f1618//19 20// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f32_16x16x16_f16_w32(21// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0:[0-9]+]] {22// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]23// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16.v8f32.v16f16(<16 x half> [[A]], <16 x half> [[B]], <8 x float> [[C]])24// CHECK-GFX1100-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8:![0-9]+]]25// CHECK-GFX1100-NEXT: ret void26//27void test_amdgcn_wmma_f32_16x16x16_f16_w32(global v8f* out, v16h a, v16h b, v8f c)28{29 *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32(a, b, c);30}31 32//33// amdgcn_wmma_f32_16x16x16_bf1634//35 36// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f32_16x16x16_bf16_w32(37// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <8 x float> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {38// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]39// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v8f32.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <8 x float> [[C]])40// CHECK-GFX1100-NEXT: store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]41// CHECK-GFX1100-NEXT: ret void42//43void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v16s a, v16s b, v8f c)44{45 *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32(a, b, c);46}47 48//49// amdgcn_wmma_f16_16x16x16_f1650//51 52// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f16_16x16x16_f16_w32(53// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <16 x half> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {54// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]55// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <16 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.v16f16.v16f16(<16 x half> [[A]], <16 x half> [[B]], <16 x half> [[C]], i1 true)56// CHECK-GFX1100-NEXT: store <16 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]57// CHECK-GFX1100-NEXT: ret void58//59void test_amdgcn_wmma_f16_16x16x16_f16_w32(global v16h* out, v16h a, v16h b, v16h c)60{61 *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32(a, b, c, true);62}63 64//65// amdgcn_wmma_bf16_16x16x16_bf1666//67 68// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_bf16_16x16x16_bf16_w32(69// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <16 x i16> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {70// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]71// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <16 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v16i16.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <16 x i16> [[C]], i1 true)72// CHECK-GFX1100-NEXT: store <16 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]73// CHECK-GFX1100-NEXT: ret void74//75void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v16s* out, v16s a, v16s b, v16s c)76{77 *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32(a, b, c, true);78}79 80//81// amdgcn_wmma_f16_16x16x16_f16_tied82//83 84// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_f16_16x16x16_f16_tied_w32(85// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x half> noundef [[A:%.*]], <16 x half> noundef [[B:%.*]], <16 x half> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {86// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]87// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <16 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.tied.v16f16.v16f16(<16 x half> [[A]], <16 x half> [[B]], <16 x half> [[C]], i1 true)88// CHECK-GFX1100-NEXT: store <16 x half> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]89// CHECK-GFX1100-NEXT: ret void90//91void test_amdgcn_wmma_f16_16x16x16_f16_tied_w32(global v16h* out, v16h a, v16h b, v16h c)92{93 *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_tied_w32(a, b, c, true);94}95 96//97// amdgcn_wmma_bf16_16x16x16_bf16_tied98//99 100// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32(101// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <16 x i16> noundef [[A:%.*]], <16 x i16> noundef [[B:%.*]], <16 x i16> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {102// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]103// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <16 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.tied.v16i16.v16i16(<16 x i16> [[A]], <16 x i16> [[B]], <16 x i16> [[C]], i1 true)104// CHECK-GFX1100-NEXT: store <16 x i16> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]105// CHECK-GFX1100-NEXT: ret void106//107void test_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32(global v16s* out, v16s a, v16s b, v16s c)108{109 *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_tied_w32(a, b, c, true);110}111 112//113// amdgcn_wmma_i32_16x16x16_iu8114//115 116// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_i32_16x16x16_iu8_w32(117// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <4 x i32> noundef [[A:%.*]], <4 x i32> noundef [[B:%.*]], <8 x i32> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {118// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]119// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu8.v8i32.v4i32(i1 true, <4 x i32> [[A]], i1 true, <4 x i32> [[B]], <8 x i32> [[C]], i1 false)120// CHECK-GFX1100-NEXT: store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]121// CHECK-GFX1100-NEXT: ret void122//123void test_amdgcn_wmma_i32_16x16x16_iu8_w32(global v8i* out, v4i a, v4i b, v8i c)124{125 *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a, true, b, c, false);126}127 128//129// amdgcn_wmma_i32_16x16x16_iu4130//131 132// CHECK-GFX1100-LABEL: define dso_local void @test_amdgcn_wmma_i32_16x16x16_iu4_w32(133// CHECK-GFX1100-SAME: ptr addrspace(1) noundef writeonly captures(none) initializes((0, 32)) [[OUT:%.*]], <2 x i32> noundef [[A:%.*]], <2 x i32> noundef [[B:%.*]], <8 x i32> noundef [[C:%.*]]) local_unnamed_addr #[[ATTR0]] {134// CHECK-GFX1100-NEXT: [[ENTRY:.*:]]135// CHECK-GFX1100-NEXT: [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu4.v8i32.v2i32(i1 true, <2 x i32> [[A]], i1 true, <2 x i32> [[B]], <8 x i32> [[C]], i1 false)136// CHECK-GFX1100-NEXT: store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT]], align 32, !tbaa [[CHAR_TBAA8]]137// CHECK-GFX1100-NEXT: ret void138//139void test_amdgcn_wmma_i32_16x16x16_iu4_w32(global v8i* out, v2i a, v2i b, v8i c)140{141 *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w32(true, a, true, b, c, false);142}143 144#endif145//.146// CHECK-GFX1100: [[META6:![0-9]+]] = !{!"omnipotent char", [[META7:![0-9]+]], i64 0}147// CHECK-GFX1100: [[META7]] = !{!"Simple C/C++ TBAA"}148// CHECK-GFX1100: [[CHAR_TBAA8]] = !{[[META6]], [[META6]], i64 0}149//.150