56 lines · plain
1; RUN: llc -mtriple=amdgcn < %s | FileCheck -check-prefix=GCN %s2 3; GCN-LABEL: {{^}}global_truncstore_f64_to_f16:4; GCN: s_endpgm5define amdgpu_kernel void @global_truncstore_f64_to_f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {6 %val = load double, ptr addrspace(1) %in7 %cvt = fptrunc double %val to half8 store half %cvt, ptr addrspace(1) %out9 ret void10}11 12; GCN-LABEL: {{^}}global_truncstore_v2f64_to_v2f16:13; GCN: s_endpgm14define amdgpu_kernel void @global_truncstore_v2f64_to_v2f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {15 %val = load <2 x double>, ptr addrspace(1) %in16 %cvt = fptrunc <2 x double> %val to <2 x half>17 store <2 x half> %cvt, ptr addrspace(1) %out18 ret void19}20 21; GCN-LABEL: {{^}}global_truncstore_v3f64_to_v3f16:22; GCN: s_endpgm23define amdgpu_kernel void @global_truncstore_v3f64_to_v3f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {24 %val = load <3 x double>, ptr addrspace(1) %in25 %cvt = fptrunc <3 x double> %val to <3 x half>26 store <3 x half> %cvt, ptr addrspace(1) %out27 ret void28}29 30; GCN-LABEL: {{^}}global_truncstore_v4f64_to_v4f16:31; GCN: s_endpgm32define amdgpu_kernel void @global_truncstore_v4f64_to_v4f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {33 %val = load <4 x double>, ptr addrspace(1) %in34 %cvt = fptrunc <4 x double> %val to <4 x half>35 store <4 x half> %cvt, ptr addrspace(1) %out36 ret void37}38 39; GCN-LABEL: {{^}}global_truncstore_v8f64_to_v8f16:40; GCN: s_endpgm41define amdgpu_kernel void @global_truncstore_v8f64_to_v8f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {42 %val = load <8 x double>, ptr addrspace(1) %in43 %cvt = fptrunc <8 x double> %val to <8 x half>44 store <8 x half> %cvt, ptr addrspace(1) %out45 ret void46}47 48; GCN-LABEL: {{^}}global_truncstore_v16f64_to_v16f16:49; GCN: s_endpgm50define amdgpu_kernel void @global_truncstore_v16f64_to_v16f16(ptr addrspace(1) %out, ptr addrspace(1) %in) #0 {51 %val = load <16 x double>, ptr addrspace(1) %in52 %cvt = fptrunc <16 x double> %val to <16 x half>53 store <16 x half> %cvt, ptr addrspace(1) %out54 ret void55}56