37 lines · plain
1; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx83 | FileCheck --check-prefixes=CHECK %s2; RUN: %if ptxas-sm_90 && ptxas-isa-8.3 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx83 | %ptxas-verify -arch=sm_90 %}3 4; CHECK-LABEL: test_fence_proxy_tensormap_generic_release5define void @test_fence_proxy_tensormap_generic_release() {6 ; CHECK: fence.proxy.tensormap::generic.release.cta;7 call void @llvm.nvvm.fence.proxy.tensormap_generic.release.cta();8 9 ; CHECK: fence.proxy.tensormap::generic.release.cluster;10 call void @llvm.nvvm.fence.proxy.tensormap_generic.release.cluster();11 12 ; CHECK: fence.proxy.tensormap::generic.release.gpu;13 call void @llvm.nvvm.fence.proxy.tensormap_generic.release.gpu();14 15 ; CHECK: fence.proxy.tensormap::generic.release.sys;16 call void @llvm.nvvm.fence.proxy.tensormap_generic.release.sys();17 18 ret void19}20 21; CHECK-LABEL: test_fence_proxy_tensormap_generic_acquire22define void @test_fence_proxy_tensormap_generic_acquire(ptr addrspace(0) %addr) {23 ; CHECK: fence.proxy.tensormap::generic.acquire.cta [%rd{{[0-9]+}}], 128;24 call void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.cta(ptr addrspace(0) %addr, i32 128);25 26 ; CHECK: fence.proxy.tensormap::generic.acquire.cluster [%rd{{[0-9]+}}], 128;27 call void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.cluster(ptr addrspace(0) %addr, i32 128);28 29 ; CHECK: fence.proxy.tensormap::generic.acquire.gpu [%rd{{[0-9]+}}], 128;30 call void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.gpu(ptr addrspace(0) %addr, i32 128);31 32 ; CHECK: fence.proxy.tensormap::generic.acquire.sys [%rd{{[0-9]+}}], 128;33 call void @llvm.nvvm.fence.proxy.tensormap_generic.acquire.sys(ptr addrspace(0) %addr, i32 128);34 35 ret void36}37