1; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s 2; RUN: llc -march=amdgcn -mcpu=fiji -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s 3 4declare half @llvm.maxnum.f16(half %a, half %b) 5declare <2 x half> @llvm.maxnum.v2f16(<2 x half> %a, <2 x half> %b) 6 7; GCN-LABEL: {{^}}maxnum_f16: 8; GCN: buffer_load_ushort v[[A_F16:[0-9]+]] 9; GCN: buffer_load_ushort v[[B_F16:[0-9]+]] 10; SI: v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]] 11; SI: v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]] 12; SI: v_max_f32_e32 v[[R_F32:[0-9]+]], v[[B_F32]], v[[A_F32]] 13; SI: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]] 14; VI: v_max_f16_e32 v[[R_F16:[0-9]+]], v[[B_F16]], v[[A_F16]] 15; GCN: buffer_store_short v[[R_F16]] 16; GCN: s_endpgm 17define amdgpu_kernel void @maxnum_f16( 18 half addrspace(1)* %r, 19 half addrspace(1)* %a, 20 half addrspace(1)* %b) { 21entry: 22 %a.val = load half, half addrspace(1)* %a 23 %b.val = load half, half addrspace(1)* %b 24 %r.val = call half @llvm.maxnum.f16(half %a.val, half %b.val) 25 store half %r.val, half addrspace(1)* %r 26 ret void 27} 28 29; GCN-LABEL: {{^}}maxnum_f16_imm_a: 30; GCN: buffer_load_ushort v[[B_F16:[0-9]+]] 31; SI: v_cvt_f32_f16_e32 v[[B_F32:[0-9]+]], v[[B_F16]] 32; SI: v_max_f32_e32 v[[R_F32:[0-9]+]], 0x40400000, v[[B_F32]] 33; SI: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]] 34; VI: v_max_f16_e32 v[[R_F16:[0-9]+]], 0x4200, v[[B_F16]] 35; GCN: buffer_store_short v[[R_F16]] 36; GCN: s_endpgm 37define amdgpu_kernel void @maxnum_f16_imm_a( 38 half addrspace(1)* %r, 39 half addrspace(1)* %b) { 40entry: 41 %b.val = load half, half addrspace(1)* %b 42 %r.val = call half @llvm.maxnum.f16(half 3.0, half %b.val) 43 store half %r.val, half addrspace(1)* %r 44 ret void 45} 46 47; GCN-LABEL: {{^}}maxnum_f16_imm_b: 48; GCN: buffer_load_ushort v[[A_F16:[0-9]+]] 49; SI: v_cvt_f32_f16_e32 v[[A_F32:[0-9]+]], v[[A_F16]] 50; SI: v_max_f32_e32 v[[R_F32:[0-9]+]], 4.0, v[[A_F32]] 51; SI: v_cvt_f16_f32_e32 v[[R_F16:[0-9]+]], v[[R_F32]] 52; VI: v_max_f16_e32 v[[R_F16:[0-9]+]], 4.0, v[[A_F16]] 53; GCN: buffer_store_short v[[R_F16]] 54; GCN: s_endpgm 55define amdgpu_kernel void @maxnum_f16_imm_b( 56 half addrspace(1)* %r, 57 half addrspace(1)* %a) { 58entry: 59 %a.val = load half, half addrspace(1)* %a 60 %r.val = call half @llvm.maxnum.f16(half %a.val, half 4.0) 61 store half %r.val, half addrspace(1)* %r 62 ret void 63} 64 65; GCN-LABEL: {{^}}maxnum_v2f16: 66; GCN: buffer_load_dword v[[A_V2_F16:[0-9]+]] 67; GCN: buffer_load_dword v[[B_V2_F16:[0-9]+]] 68 69; SI: v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]] 70; SI: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]] 71; SI: v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]] 72; SI: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]] 73 74; SI: v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]] 75; SI: v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]] 76; SI: v_max_f32_e32 v[[R_F32_0:[0-9]+]], v[[B_F32_0]], v[[A_F32_0]] 77; SI-DAG: v_max_f32_e32 v[[R_F32_1:[0-9]+]], v[[B_F32_1]], v[[A_F32_1]] 78; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]] 79; SI-DAG: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]] 80; SI: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]] 81; SI-NOT: and 82; SI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_HI]], v[[R_F16_0]] 83 84; VI-DAG: v_max_f16_e32 v[[R_F16_0:[0-9]+]], v[[B_V2_F16]], v[[A_V2_F16]] 85; VI-DAG: v_max_f16_sdwa v[[R_F16_1:[0-9]+]], v[[A_V2_F16]], v[[B_V2_F16]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:WORD_1 86; VI-NOT: and 87; VI: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_1]], v[[R_F16_0]] 88 89; GCN: buffer_store_dword v[[R_V2_F16]] 90; GCN: s_endpgm 91define amdgpu_kernel void @maxnum_v2f16( 92 <2 x half> addrspace(1)* %r, 93 <2 x half> addrspace(1)* %a, 94 <2 x half> addrspace(1)* %b) { 95entry: 96 %a.val = load <2 x half>, <2 x half> addrspace(1)* %a 97 %b.val = load <2 x half>, <2 x half> addrspace(1)* %b 98 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> %b.val) 99 store <2 x half> %r.val, <2 x half> addrspace(1)* %r 100 ret void 101} 102 103; GCN-LABEL: {{^}}maxnum_v2f16_imm_a: 104; GCN-DAG: buffer_load_dword v[[B_V2_F16:[0-9]+]] 105; SI: v_cvt_f32_f16_e32 v[[B_F32_0:[0-9]+]], v[[B_V2_F16]] 106; SI: v_lshrrev_b32_e32 v[[B_F16_1:[0-9]+]], 16, v[[B_V2_F16]] 107; SI: v_cvt_f32_f16_e32 v[[B_F32_1:[0-9]+]], v[[B_F16_1]] 108; SI: v_max_f32_e32 v[[R_F32_0:[0-9]+]], 0x40400000, v[[B_F32_0]] 109; SI: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]] 110; SI: v_max_f32_e32 v[[R_F32_1:[0-9]+]], 4.0, v[[B_F32_1]] 111; SI: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]] 112; VI-DAG: v_mov_b32_e32 [[CONST4:v[0-9]+]], 0x4400 113; VI-DAG: v_max_f16_sdwa v[[R_F16_HI:[0-9]+]], v[[B_V2_F16]], [[CONST4]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD 114; VI-DAG: v_max_f16_e32 v[[R_F16_0:[0-9]+]], 0x4200, v[[B_V2_F16]] 115 116; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]] 117; GCN-NOT: and 118; GCN: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_HI]], v[[R_F16_0]] 119; GCN: buffer_store_dword v[[R_V2_F16]] 120; GCN: s_endpgm 121define amdgpu_kernel void @maxnum_v2f16_imm_a( 122 <2 x half> addrspace(1)* %r, 123 <2 x half> addrspace(1)* %b) { 124entry: 125 %b.val = load <2 x half>, <2 x half> addrspace(1)* %b 126 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> <half 3.0, half 4.0>, <2 x half> %b.val) 127 store <2 x half> %r.val, <2 x half> addrspace(1)* %r 128 ret void 129} 130 131; GCN-LABEL: {{^}}maxnum_v2f16_imm_b: 132; GCN-DAG: buffer_load_dword v[[A_V2_F16:[0-9]+]] 133; SI: v_cvt_f32_f16_e32 v[[A_F32_0:[0-9]+]], v[[A_V2_F16]] 134; SI: v_lshrrev_b32_e32 v[[A_F16_1:[0-9]+]], 16, v[[A_V2_F16]] 135; SI: v_cvt_f32_f16_e32 v[[A_F32_1:[0-9]+]], v[[A_F16_1]] 136; SI: v_max_f32_e32 v[[R_F32_0:[0-9]+]], 4.0, v[[A_F32_0]] 137; SI: v_cvt_f16_f32_e32 v[[R_F16_0:[0-9]+]], v[[R_F32_0]] 138; SI: v_max_f32_e32 v[[R_F32_1:[0-9]+]], 0x40400000, v[[A_F32_1]] 139; SI: v_cvt_f16_f32_e32 v[[R_F16_1:[0-9]+]], v[[R_F32_1]] 140; VI-DAG: v_mov_b32_e32 [[CONST3:v[0-9]+]], 0x4200 141; VI-DAG: v_max_f16_sdwa v[[R_F16_HI:[0-9]+]], v[[B_V2_F16]], [[CONST3]] dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD 142; VI-DAG: v_max_f16_e32 v[[R_F16_0:[0-9]+]], 4.0, v[[A_V2_F16]] 143 144; SI-DAG: v_lshlrev_b32_e32 v[[R_F16_HI:[0-9]+]], 16, v[[R_F16_1]] 145; GCN-NOT: and 146; GCN: v_or_b32_e32 v[[R_V2_F16:[0-9]+]], v[[R_F16_HI]], v[[R_F16_0]] 147; GCN: buffer_store_dword v[[R_V2_F16]] 148; GCN: s_endpgm 149define amdgpu_kernel void @maxnum_v2f16_imm_b( 150 <2 x half> addrspace(1)* %r, 151 <2 x half> addrspace(1)* %a) { 152entry: 153 %a.val = load <2 x half>, <2 x half> addrspace(1)* %a 154 %r.val = call <2 x half> @llvm.maxnum.v2f16(<2 x half> %a.val, <2 x half> <half 4.0, half 3.0>) 155 store <2 x half> %r.val, <2 x half> addrspace(1)* %r 156 ret void 157} 158