1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py 2// REQUIRES: amdgpu-registered-target 3// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1200 -target-feature +wavefrontsize64 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX1200 4 5typedef int v2i __attribute__((ext_vector_type(2))); 6typedef int v4i __attribute__((ext_vector_type(4))); 7typedef float v4f __attribute__((ext_vector_type(4))); 8typedef half v4h __attribute__((ext_vector_type(4))); 9typedef short v4s __attribute__((ext_vector_type(4))); 10typedef half v8h __attribute__((ext_vector_type(8))); 11typedef short v8s __attribute__((ext_vector_type(8))); 12 13// Wave64 14 15// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_f16_w64( 16// CHECK-GFX1200-NEXT: entry: 17// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.f16.v4f32.v4f16.v8f16.i16(<4 x half> [[A:%.*]], <8 x half> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 18// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4:![0-9]+]] 19// CHECK-GFX1200-NEXT: ret void 20// 21void test_amdgcn_swmmac_f32_16x16x32_f16_w64(global v4f* out, v4h a, v8h b, v4f c, short index) 22{ 23 *out = __builtin_amdgcn_swmmac_f32_16x16x32_f16_w64(a, b, c, index); 24} 25 26// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_bf16_w64( 27// CHECK-GFX1200-NEXT: entry: 28// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf16.v4f32.v4i16.v8i16.i16(<4 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 29// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 30// CHECK-GFX1200-NEXT: ret void 31// 32void test_amdgcn_swmmac_f32_16x16x32_bf16_w64(global v4f* out, v4s a, v8s b, v4f c, short index) 33{ 34 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf16_w64(a, b, c, index); 35} 36 37// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f16_16x16x32_f16_w64( 38// CHECK-GFX1200-NEXT: entry: 39// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x half> @llvm.amdgcn.swmmac.f16.16x16x32.f16.v4f16.v4f16.v8f16.i16(<4 x half> [[A:%.*]], <8 x half> [[B:%.*]], <4 x half> [[C:%.*]], i16 [[INDEX:%.*]]) 40// CHECK-GFX1200-NEXT: store <4 x half> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]] 41// CHECK-GFX1200-NEXT: ret void 42// 43void test_amdgcn_swmmac_f16_16x16x32_f16_w64(global v4h* out, v4h a, v8h b, v4h c, short index) 44{ 45 *out = __builtin_amdgcn_swmmac_f16_16x16x32_f16_w64(a, b, c, index); 46} 47 48// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_bf16_16x16x32_bf16_w64( 49// CHECK-GFX1200-NEXT: entry: 50// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i16> @llvm.amdgcn.swmmac.bf16.16x16x32.bf16.v4i16.v4i16.v8i16.i16(<4 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <4 x i16> [[C:%.*]], i16 [[INDEX:%.*]]) 51// CHECK-GFX1200-NEXT: store <4 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 8, !tbaa [[TBAA4]] 52// CHECK-GFX1200-NEXT: ret void 53// 54void test_amdgcn_swmmac_bf16_16x16x32_bf16_w64(global v4s* out, v4s a, v8s b, v4s c, short index) 55{ 56 *out = __builtin_amdgcn_swmmac_bf16_16x16x32_bf16_w64(a, b, c, index); 57} 58 59// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_i32_16x16x32_iu8_w64( 60// CHECK-GFX1200-NEXT: entry: 61// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu8.v4i32.i32.v2i32.i16(i1 true, i32 [[A:%.*]], i1 true, <2 x i32> [[B:%.*]], <4 x i32> [[C:%.*]], i16 [[INDEX:%.*]], i1 true) 62// CHECK-GFX1200-NEXT: store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 63// CHECK-GFX1200-NEXT: ret void 64// 65void test_amdgcn_swmmac_i32_16x16x32_iu8_w64(global v4i* out, int a, v2i b, v4i c, short index) 66{ 67 *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu8_w64(true, a, true, b, c, index, true); 68} 69 70// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_i32_16x16x32_iu4_w64( 71// CHECK-GFX1200-NEXT: entry: 72// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x32.iu4.v4i32.i32.i32.i16(i1 true, i32 [[A:%.*]], i1 true, i32 [[B:%.*]], <4 x i32> [[C:%.*]], i16 [[INDEX:%.*]], i1 true) 73// CHECK-GFX1200-NEXT: store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 74// CHECK-GFX1200-NEXT: ret void 75// 76void test_amdgcn_swmmac_i32_16x16x32_iu4_w64(global v4i* out, int a, int b, v4i c, short index) 77{ 78 *out = __builtin_amdgcn_swmmac_i32_16x16x32_iu4_w64(true, a, true, b, c, index, true); 79} 80 81// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_i32_16x16x64_iu4_w64( 82// CHECK-GFX1200-NEXT: entry: 83// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x i32> @llvm.amdgcn.swmmac.i32.16x16x64.iu4.v4i32.i32.v2i32.i16(i1 true, i32 [[A:%.*]], i1 true, <2 x i32> [[B:%.*]], <4 x i32> [[C:%.*]], i16 [[INDEX:%.*]], i1 true) 84// CHECK-GFX1200-NEXT: store <4 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 85// CHECK-GFX1200-NEXT: ret void 86// 87void test_amdgcn_swmmac_i32_16x16x64_iu4_w64(global v4i* out, int a, v2i b, v4i c, short index) 88{ 89 *out = __builtin_amdgcn_swmmac_i32_16x16x64_iu4_w64(true, a, true, b, c, index, true); 90} 91 92// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64( 93// CHECK-GFX1200-NEXT: entry: 94// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.fp8.v4f32.i32.v2i32.i16(i32 [[A:%.*]], <2 x i32> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 95// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 96// CHECK-GFX1200-NEXT: ret void 97// 98void test_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64(global v4f* out, int a, v2i b, v4f c, short index) 99{ 100 *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_fp8_w64(a, b, c, index); 101} 102 103// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64( 104// CHECK-GFX1200-NEXT: entry: 105// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.fp8.bf8.v4f32.i32.v2i32.i16(i32 [[A:%.*]], <2 x i32> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 106// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 107// CHECK-GFX1200-NEXT: ret void 108// 109void test_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64(global v4f* out, int a, v2i b, v4f c, short index) 110{ 111 *out = __builtin_amdgcn_swmmac_f32_16x16x32_fp8_bf8_w64(a, b, c, index); 112} 113 114// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64( 115// CHECK-GFX1200-NEXT: entry: 116// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.fp8.v4f32.i32.v2i32.i16(i32 [[A:%.*]], <2 x i32> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 117// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 118// CHECK-GFX1200-NEXT: ret void 119// 120void test_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64(global v4f* out, int a, v2i b, v4f c, short index) 121{ 122 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_fp8_w64(a, b, c, index); 123} 124 125// CHECK-GFX1200-LABEL: @test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64( 126// CHECK-GFX1200-NEXT: entry: 127// CHECK-GFX1200-NEXT: [[TMP0:%.*]] = tail call <4 x float> @llvm.amdgcn.swmmac.f32.16x16x32.bf8.bf8.v4f32.i32.v2i32.i16(i32 [[A:%.*]], <2 x i32> [[B:%.*]], <4 x float> [[C:%.*]], i16 [[INDEX:%.*]]) 128// CHECK-GFX1200-NEXT: store <4 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]] 129// CHECK-GFX1200-NEXT: ret void 130// 131void test_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64(global v4f* out, int a, v2i b, v4f c, short index) 132{ 133 *out = __builtin_amdgcn_swmmac_f32_16x16x32_bf8_bf8_w64(a, b, c, index); 134} 135