471 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=instcombine < %s | FileCheck %s3 4; ------------------------------------------------------------------------------------5; Incorrect signature for format cases (IR vector too large) wmma.f32.16x16x128.f8f6f46; ------------------------------------------------------------------------------------7 8define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {9; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(10; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {11; CHECK-NEXT: [[BB:.*:]]12; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>13; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 2, <12 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]])14; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3215; CHECK-NEXT: ret void16;17bb:18 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 2, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C)19 store <8 x float> %res, ptr addrspace(1) %out20 ret void21}22 23define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {24; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(25; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {26; CHECK-NEXT: [[BB:.*:]]27; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>28; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 2, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]])29; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3230; CHECK-NEXT: ret void31;32bb:33 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C)34 store <8 x float> %res, ptr addrspace(1) %out35 ret void36}37 38define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {39; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(40; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {41; CHECK-NEXT: [[BB:.*:]]42; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>43; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 3, <12 x i32> [[TMP0]], i32 1, <16 x i32> [[B]], i16 0, <8 x float> [[C]])44; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3245; CHECK-NEXT: ret void46;47bb:48 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 1, <16 x i32> %B, i16 0, <8 x float> %C)49 store <8 x float> %res, ptr addrspace(1) %out50 ret void51}52 53define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {54; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(55; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {56; CHECK-NEXT: [[BB:.*:]]57; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>58; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 3, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]])59; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3260; CHECK-NEXT: ret void61;62bb:63 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 3, <16 x i32> %B, i16 0, <8 x float> %C)64 store <8 x float> %res, ptr addrspace(1) %out65 ret void66}67 68define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {69; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(70; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {71; CHECK-NEXT: [[BB:.*:]]72; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>73; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]])74; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3275; CHECK-NEXT: ret void76;77bb:78 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C)79 store <8 x float> %res, ptr addrspace(1) %out80 ret void81}82 83define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {84; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(85; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {86; CHECK-NEXT: [[BB:.*:]]87; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>88; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]])89; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 3290; CHECK-NEXT: ret void91;92bb:93 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <16 x i32> %B, i16 0, <8 x float> %C)94 store <8 x float> %res, ptr addrspace(1) %out95 ret void96}97 98define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {99; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(100; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {101; CHECK-NEXT: [[BB:.*:]]102; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>103; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]])104; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32105; CHECK-NEXT: ret void106;107bb:108 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C)109 store <8 x float> %res, ptr addrspace(1) %out110 ret void111}112 113define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {114; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(115; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {116; CHECK-NEXT: [[BB:.*:]]117; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>118; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]])119; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32120; CHECK-NEXT: ret void121;122bb:123 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C)124 store <8 x float> %res, ptr addrspace(1) %out125 ret void126}127 128define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {129; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(130; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {131; CHECK-NEXT: [[BB:.*:]]132; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>133; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>134; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v8i32.v12i32(i32 4, <8 x i32> [[TMP0]], i32 2, <12 x i32> [[TMP1]], i16 0, <8 x float> [[C]])135; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32136; CHECK-NEXT: ret void137;138bb:139 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C)140 store <8 x float> %res, ptr addrspace(1) %out141 ret void142}143 144define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {145; CHECK-LABEL: define amdgpu_ps void @test_wmma_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(146; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {147; CHECK-NEXT: [[BB:.*:]]148; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>149; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>150; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v12i32.v8i32(i32 3, <12 x i32> [[TMP0]], i32 4, <8 x i32> [[TMP1]], i16 0, <8 x float> [[C]])151; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32152; CHECK-NEXT: ret void153;154bb:155 %res = call <8 x float> @llvm.amdgcn.wmma.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C)156 store <8 x float> %res, ptr addrspace(1) %out157 ret void158}159 160; ------------------------------------------------------------------------------------------161; Incorrect signature for format cases (IR vector too large) wmma.scale.f32.16x16x128.f8f6f4162; ------------------------------------------------------------------------------------------163 164define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {165; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(166; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {167; CHECK-NEXT: [[BB:.*:]]168; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>169; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 2, <12 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)170; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32171; CHECK-NEXT: ret void172;173bb:174 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 2, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)175 store <8 x float> %res, ptr addrspace(1) %out176 ret void177}178 179define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {180; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(181; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {182; CHECK-NEXT: [[BB:.*:]]183; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>184; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 2, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)185; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32186; CHECK-NEXT: ret void187;188bb:189 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)190 store <8 x float> %res, ptr addrspace(1) %out191 ret void192}193 194define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {195; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(196; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {197; CHECK-NEXT: [[BB:.*:]]198; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>199; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 3, <12 x i32> [[TMP0]], i32 1, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)200; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32201; CHECK-NEXT: ret void202;203bb:204 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 1, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)205 store <8 x float> %res, ptr addrspace(1) %out206 ret void207}208 209define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {210; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(211; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {212; CHECK-NEXT: [[BB:.*:]]213; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>214; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 3, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)215; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32216; CHECK-NEXT: ret void217;218bb:219 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 3, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)220 store <8 x float> %res, ptr addrspace(1) %out221 ret void222}223 224define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {225; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(226; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {227; CHECK-NEXT: [[BB:.*:]]228; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>229; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)230; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32231; CHECK-NEXT: ret void232;233bb:234 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)235 store <8 x float> %res, ptr addrspace(1) %out236 ret void237}238 239define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {240; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(241; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {242; CHECK-NEXT: [[BB:.*:]]243; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>244; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)245; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32246; CHECK-NEXT: ret void247;248bb:249 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)250 store <8 x float> %res, ptr addrspace(1) %out251 ret void252}253 254define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {255; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(256; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {257; CHECK-NEXT: [[BB:.*:]]258; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>259; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)260; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32261; CHECK-NEXT: ret void262;263bb:264 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)265 store <8 x float> %res, ptr addrspace(1) %out266 ret void267}268 269define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {270; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(271; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {272; CHECK-NEXT: [[BB:.*:]]273; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>274; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)275; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32276; CHECK-NEXT: ret void277;278bb:279 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)280 store <8 x float> %res, ptr addrspace(1) %out281 ret void282}283 284define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {285; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(286; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {287; CHECK-NEXT: [[BB:.*:]]288; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>289; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>290; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v8i32.v12i32(i32 4, <8 x i32> [[TMP0]], i32 2, <12 x i32> [[TMP1]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)291; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32292; CHECK-NEXT: ret void293;294bb:295 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)296 store <8 x float> %res, ptr addrspace(1) %out297 ret void298}299 300define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {301; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(302; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {303; CHECK-NEXT: [[BB:.*:]]304; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>305; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>306; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v12i32.v8i32(i32 3, <12 x i32> [[TMP0]], i32 4, <8 x i32> [[TMP1]], i16 0, <8 x float> [[C]], i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)307; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32308; CHECK-NEXT: ret void309;310bb:311 %res = call <8 x float> @llvm.amdgcn.wmma.scale.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i32 0, i32 0, i32 0, i32 0, i1 false, i1 false)312 store <8 x float> %res, ptr addrspace(1) %out313 ret void314}315 316; --------------------------------------------------------------------------------------------317; Incorrect signature for format cases (IR vector too large) wmma.scale16.f32.16x16x128.f8f6f4318; --------------------------------------------------------------------------------------------319 320define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {321; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp6___v16i32_fp8(322; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {323; CHECK-NEXT: [[BB:.*:]]324; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>325; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 2, <12 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)326; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32327; CHECK-NEXT: ret void328;329bb:330 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 2, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)331 store <8 x float> %res, ptr addrspace(1) %out332 ret void333}334 335define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {336; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_fp6(337; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {338; CHECK-NEXT: [[BB:.*:]]339; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>340; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 2, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)341; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32342; CHECK-NEXT: ret void343;344bb:345 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)346 store <8 x float> %res, ptr addrspace(1) %out347 ret void348}349 350define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {351; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf6___v16i32_bf8(352; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {353; CHECK-NEXT: [[BB:.*:]]354; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>355; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v12i32.v16i32(i32 3, <12 x i32> [[TMP0]], i32 1, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)356; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32357; CHECK-NEXT: ret void358;359bb:360 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 1, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)361 store <8 x float> %res, ptr addrspace(1) %out362 ret void363}364 365define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {366; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf8___v16i32_bf6(367; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {368; CHECK-NEXT: [[BB:.*:]]369; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>370; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v12i32(i32 1, <16 x i32> [[A]], i32 3, <12 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)371; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32372; CHECK-NEXT: ret void373;374bb:375 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 1, <16 x i32> %A, i32 3, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)376 store <8 x float> %res, ptr addrspace(1) %out377 ret void378}379 380define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {381; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp4___v16i32_fp8(382; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {383; CHECK-NEXT: [[BB:.*:]]384; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>385; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)386; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32387; CHECK-NEXT: ret void388;389bb:390 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <16 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)391 store <8 x float> %res, ptr addrspace(1) %out392 ret void393}394 395define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(<16 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {396; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp8___v16i32_fp4(397; CHECK-SAME: <16 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {398; CHECK-NEXT: [[BB:.*:]]399; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>400; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)401; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32402; CHECK-NEXT: ret void403;404bb:405 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)406 store <8 x float> %res, ptr addrspace(1) %out407 ret void408}409 410define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {411; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp8(412; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {413; CHECK-NEXT: [[BB:.*:]]414; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>415; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v16i32(i32 4, <8 x i32> [[TMP0]], i32 0, <16 x i32> [[B]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)416; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32417; CHECK-NEXT: ret void418;419bb:420 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 0, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)421 store <8 x float> %res, ptr addrspace(1) %out422 ret void423}424 425define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {426; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_fp8___v12i32_fp4(427; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {428; CHECK-NEXT: [[BB:.*:]]429; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>430; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v8i32(i32 0, <16 x i32> [[A]], i32 4, <8 x i32> [[TMP0]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)431; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32432; CHECK-NEXT: ret void433;434bb:435 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 0, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)436 store <8 x float> %res, ptr addrspace(1) %out437 ret void438}439 440define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(<12 x i32> %A, <16 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {441; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v12i32_fp4___v16i32_fp6(442; CHECK-SAME: <12 x i32> [[A:%.*]], <16 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {443; CHECK-NEXT: [[BB:.*:]]444; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <12 x i32> [[A]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>445; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <16 x i32> [[B]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>446; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v8i32.v12i32(i32 4, <8 x i32> [[TMP0]], i32 2, <12 x i32> [[TMP1]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)447; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32448; CHECK-NEXT: ret void449;450bb:451 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 4, <12 x i32> %A, i32 2, <16 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)452 store <8 x float> %res, ptr addrspace(1) %out453 ret void454}455 456define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(<16 x i32> %A, <12 x i32> %B, <8 x float> %C, ptr addrspace(1) %out) {457; CHECK-LABEL: define amdgpu_ps void @test_wmma_scale16_f32_16x16x128_f8f6f4___v16i32_bf6___v12i32_fp4(458; CHECK-SAME: <16 x i32> [[A:%.*]], <12 x i32> [[B:%.*]], <8 x float> [[C:%.*]], ptr addrspace(1) [[OUT:%.*]]) {459; CHECK-NEXT: [[BB:.*:]]460; CHECK-NEXT: [[TMP0:%.*]] = shufflevector <16 x i32> [[A]], <16 x i32> poison, <12 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11>461; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <12 x i32> [[B]], <12 x i32> poison, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>462; CHECK-NEXT: [[RES:%.*]] = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v12i32.v8i32(i32 3, <12 x i32> [[TMP0]], i32 4, <8 x i32> [[TMP1]], i16 0, <8 x float> [[C]], i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)463; CHECK-NEXT: store <8 x float> [[RES]], ptr addrspace(1) [[OUT]], align 32464; CHECK-NEXT: ret void465;466bb:467 %res = call <8 x float> @llvm.amdgcn.wmma.scale16.f32.16x16x128.f8f6f4.v8f32.v16i32.v16i32(i32 3, <16 x i32> %A, i32 4, <12 x i32> %B, i16 0, <8 x float> %C, i32 0, i32 0, i64 0, i32 0, i32 0, i64 0, i1 false, i1 false)468 store <8 x float> %res, ptr addrspace(1) %out469 ret void470}471