1 ; RUN: llc -march=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck %s --check-prefix=GFX906 2 3 declare float @llvm.amdgcn.fdot2(<2 x half> %a, <2 x half> %b, float %c, i1 %clamp) 4 5 ; GFX906-LABEL: {{^}}test_llvm_amdgcn_fdot2_clamp 6 ; GFX906: v_dot2_f32_f16 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}} clamp{{$}} 7 define amdgpu_kernel void @test_llvm_amdgcn_fdot2_clamp( 8 float addrspace(1)* %r, 9 <2 x half> addrspace(1)* %a, 10 <2 x half> addrspace(1)* %b, 11 float addrspace(1)* %c) { 12 entry: 13 %a.val = load <2 x half>, <2 x half> addrspace(1)* %a 14 %b.val = load <2 x half>, <2 x half> addrspace(1)* %b 15 %c.val = load float, float addrspace(1)* %c 16 %r.val = call float @llvm.amdgcn.fdot2(<2 x half> %a.val, <2 x half> %b.val, float %c.val, i1 1) 17 store float %r.val, float addrspace(1)* %r 18 ret void 19 } 20 21 ; GFX906-LABEL: {{^}}test_llvm_amdgcn_fdot2_no_clamp 22 ; GFX906: v_dot2_f32_f16 v{{[0-9]+}}, s{{[0-9]+}}, v{{[0-9]+}}, v{{[0-9]+}}{{$}} 23 define amdgpu_kernel void @test_llvm_amdgcn_fdot2_no_clamp( 24 float addrspace(1)* %r, 25 <2 x half> addrspace(1)* %a, 26 <2 x half> addrspace(1)* %b, 27 float addrspace(1)* %c) { 28 entry: 29 %a.val = load <2 x half>, <2 x half> addrspace(1)* %a 30 %b.val = load <2 x half>, <2 x half> addrspace(1)* %b 31 %c.val = load float, float addrspace(1)* %c 32 %r.val = call float @llvm.amdgcn.fdot2(<2 x half> %a.val, <2 x half> %b.val, float %c.val, i1 0) 33 store float %r.val, float addrspace(1)* %r 34 ret void 35 } 36
