1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI -check-prefix=SIVI %s 2; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI -check-prefix=SIVI %s 3; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -denormal-fp-math=preserve-sign -verify-machineinstrs -enable-unsafe-fp-math < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GFX9 %s 4 5; GCN-LABEL: {{^}}fptrunc_f32_to_f16: 6; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 7; GCN: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[A_F32]] 8; GCN: buffer_store_short v[[R_F16]] 9; GCN: s_endpgm 10define amdgpu_kernel void @fptrunc_f32_to_f16( 11 half addrspace(1)* %r, 12 float addrspace(1)* %a) { 13entry: 14 %a.val = load float, float addrspace(1)* %a 15 %r.val = fptrunc float %a.val to half 16 store half %r.val, half addrspace(1)* %r 17 ret void 18} 19 20; GCN-LABEL: {{^}}fptrunc_f64_to_f16: 21; GCN: buffer_load_dwordx2 v[[[A_F64_0:[0-9]+]]:[[A_F64_1:[0-9]+]]] 22; GCN: v_cvt_f32_f64_e32 v[[A_F32:[0-9]+]], v[[[A_F64_0]]:[[A_F64_1]]] 23; GCN: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[A_F32]] 24; GCN: buffer_store_short v[[R_F16]] 25; GCN: s_endpgm 26define amdgpu_kernel void @fptrunc_f64_to_f16( 27 half addrspace(1)* %r, 28 double addrspace(1)* %a) { 29entry: 30 %a.val = load double, double addrspace(1)* %a 31 %r.val = fptrunc double %a.val to half 32 store half %r.val, half addrspace(1)* %r 33 ret void 34} 35 36; GCN-LABEL: {{^}}fptrunc_v2f32_to_v2f16: 37; GCN: buffer_load_dwordx2 v[[[A_F32_0:[0-9]+]]:[[A_F32_1:[0-9]+]]] 38; GCN-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[A_F32_0]] 39; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[A_F32_1]] 40; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]] 41; SI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]] 42 43; VI-DAG: v_cvt_f16_f32_sdwa v[[R_F16_1:[0-9]+]], v[[A_F32_1]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD 44; VI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_1]] 45 46; GFX9-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[A_F32_1]] 47; GFX9: v_pack_b32_f16 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_1]] 48 49; GCN: buffer_store_dword v[[R_V2_F16]] 50; GCN: s_endpgm 51 52define amdgpu_kernel void @fptrunc_v2f32_to_v2f16( 53 <2 x half> addrspace(1)* %r, 54 <2 x float> addrspace(1)* %a) { 55entry: 56 %a.val = load <2 x float>, <2 x float> addrspace(1)* %a 57 %r.val = fptrunc <2 x float> %a.val to <2 x half> 58 store <2 x half> %r.val, <2 x half> addrspace(1)* %r 59 ret void 60} 61 62; GCN-LABEL: {{^}}fptrunc_v2f64_to_v2f16: 63; GCN: buffer_load_dwordx4 v[[[A_F64_0:[0-9]+]]:[[A_F64_3:[0-9]+]]] 64; GCN-DAG: v_cvt_f32_f64_e32 v[[A_F32_0:[0-9]+]], v[[[A_F64_0]]:{{[0-9]+}}] 65; GCN-DAG: v_cvt_f32_f64_e32 v[[A_F32_1:[0-9]+]], v[{{[0-9]+}}:[[A_F64_3]]] 66; VI: v_cvt_f16_f32_sdwa v[[R_F16_HI:[0-9]+]], v[[A_F32_1]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD 67; GCN-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[A_F32_0]] 68; 69; SI-DAG: v_cvt_f16_f32_e32 v[[CVTHI:[0-9]+]], v[[A_F32_1]] 70; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[CVTHI]] 71 72; SIVI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_0]], v[[R_F16_HI]] 73 74; GFX9-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[A_F32_1]] 75; GFX9: v_lshl_or_b32 v[[R_V2_F16:[0-9]+]], v[[R_F16_1]], 16, v[[R_F16_0]] 76 77; GCN: buffer_store_dword v[[R_V2_F16]] 78 79define amdgpu_kernel void @fptrunc_v2f64_to_v2f16( 80 <2 x half> addrspace(1)* %r, 81 <2 x double> addrspace(1)* %a) { 82entry: 83 %a.val = load <2 x double>, <2 x double> addrspace(1)* %a 84 %r.val = fptrunc <2 x double> %a.val to <2 x half> 85 store <2 x half> %r.val, <2 x half> addrspace(1)* %r 86 ret void 87} 88 89; GCN-LABEL: {{^}}fneg_fptrunc_f32_to_f16: 90; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 91; GCN: v_cvt_f16_f32_e64 v[[R_F16:[0-9]+]], -v[[A_F32]] 92; GCN: buffer_store_short v[[R_F16]] 93; GCN: s_endpgm 94define amdgpu_kernel void @fneg_fptrunc_f32_to_f16( 95 half addrspace(1)* %r, 96 float addrspace(1)* %a) { 97entry: 98 %a.val = load float, float addrspace(1)* %a 99 %a.fneg = fneg float %a.val 100 %r.val = fptrunc float %a.fneg to half 101 store half %r.val, half addrspace(1)* %r 102 ret void 103} 104 105; GCN-LABEL: {{^}}fabs_fptrunc_f32_to_f16: 106; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 107; GCN: v_cvt_f16_f32_e64 v[[R_F16:[0-9]+]], |v[[A_F32]]| 108; GCN: buffer_store_short v[[R_F16]] 109; GCN: s_endpgm 110define amdgpu_kernel void @fabs_fptrunc_f32_to_f16( 111 half addrspace(1)* %r, 112 float addrspace(1)* %a) { 113entry: 114 %a.val = load float, float addrspace(1)* %a 115 %a.fabs = call float @llvm.fabs.f32(float %a.val) 116 %r.val = fptrunc float %a.fabs to half 117 store half %r.val, half addrspace(1)* %r 118 ret void 119} 120 121; GCN-LABEL: {{^}}fneg_fabs_fptrunc_f32_to_f16: 122; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 123; GCN: v_cvt_f16_f32_e64 v[[R_F16:[0-9]+]], -|v[[A_F32]]| 124; GCN: buffer_store_short v[[R_F16]] 125; GCN: s_endpgm 126define amdgpu_kernel void @fneg_fabs_fptrunc_f32_to_f16( 127 half addrspace(1)* %r, 128 float addrspace(1)* %a) #0 { 129entry: 130 %a.val = load float, float addrspace(1)* %a 131 %a.fabs = call float @llvm.fabs.f32(float %a.val) 132 %a.fneg.fabs = fneg float %a.fabs 133 %r.val = fptrunc float %a.fneg.fabs to half 134 store half %r.val, half addrspace(1)* %r 135 ret void 136} 137 138; GCN-LABEL: {{^}}fptrunc_f32_to_f16_zext_i32: 139; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 140; GCN: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[A_F32]] 141; SIVI-NOT: v[[R_F16]] 142; GFX9-NOT: v_and_b32 143; GCN: buffer_store_dword v[[R_F16]] 144define amdgpu_kernel void @fptrunc_f32_to_f16_zext_i32( 145 i32 addrspace(1)* %r, 146 float addrspace(1)* %a) #0 { 147entry: 148 %a.val = load float, float addrspace(1)* %a 149 %r.val = fptrunc float %a.val to half 150 %r.i16 = bitcast half %r.val to i16 151 %zext = zext i16 %r.i16 to i32 152 store i32 %zext, i32 addrspace(1)* %r 153 ret void 154} 155 156; GCN-LABEL: {{^}}fptrunc_fabs_f32_to_f16_zext_i32: 157; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 158; GCN: v_cvt_f16_f32_e64 v[[R_F16:[0-9]+]], |v[[A_F32]]| 159; SIVI-NOT: v[[R_F16]] 160; GFX9-NOT: v_and_b32 161; GCN: buffer_store_dword v[[R_F16]] 162define amdgpu_kernel void @fptrunc_fabs_f32_to_f16_zext_i32( 163 i32 addrspace(1)* %r, 164 float addrspace(1)* %a) #0 { 165entry: 166 %a.val = load float, float addrspace(1)* %a 167 %a.fabs = call float @llvm.fabs.f32(float %a.val) 168 %r.val = fptrunc float %a.fabs to half 169 %r.i16 = bitcast half %r.val to i16 170 %zext = zext i16 %r.i16 to i32 171 store i32 %zext, i32 addrspace(1)* %r 172 ret void 173} 174 175; GCN-LABEL: {{^}}fptrunc_f32_to_f16_sext_i32: 176; GCN: buffer_load_dword v[[A_F32:[0-9]+]] 177; GCN: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[A_F32]] 178; GCN: v_bfe_i32 v[[R_F16_SEXT:[0-9]+]], v[[R_F16]], 0, 16 179; GCN: buffer_store_dword v[[R_F16_SEXT]] 180define amdgpu_kernel void @fptrunc_f32_to_f16_sext_i32( 181 i32 addrspace(1)* %r, 182 float addrspace(1)* %a) #0 { 183entry: 184 %a.val = load float, float addrspace(1)* %a 185 %r.val = fptrunc float %a.val to half 186 %r.i16 = bitcast half %r.val to i16 187 %zext = sext i16 %r.i16 to i32 188 store i32 %zext, i32 addrspace(1)* %r 189 ret void 190} 191 192declare float @llvm.fabs.f32(float) #1 193 194attributes #0 = { nounwind } 195attributes #1 = { nounwind readnone } 196