61 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -mcpu=sm_90 -mattr=+ptx90 | FileCheck %s3 4target triple = "nvptx64-nvidia-cuda"5 6; Test "blocksareclusters" attribute with full "reqntid" and "cluster_dim"7; attributes.8define ptx_kernel void @kernel1(ptr %input, ptr %output) #0 #1 #2 {9; CHECK-LABEL: kernel1(10; CHECK: .reqntid 1024, 1, 111; CHECK-NEXT: .reqnctapercluster 2, 2, 212; CHECK-NEXT: .blocksareclusters13; CHECK-NEXT: {14; CHECK-EMPTY:15; CHECK-EMPTY:16; CHECK-NEXT: // %bb.0:17; CHECK-NEXT: ret;18 ret void19}20 21; Test "blocksareclusters" attribute with single dimension "reqntid" and22; "cluster_dim" attributes.23define ptx_kernel void @kernel2(ptr %input, ptr %output) #0 #3 #4 {24; CHECK-LABEL: kernel2(25; CHECK: .reqntid 102426; CHECK-NEXT: .reqnctapercluster 227; CHECK-NEXT: .blocksareclusters // @kernel228; CHECK-NEXT: {29; CHECK-EMPTY:30; CHECK-EMPTY:31; CHECK-NEXT: // %bb.0:32; CHECK-NEXT: ret;33 ret void34}35 36; Test "blocksareclusters" attribute with two dimensions(not z dimension)37; "reqntid" and "cluster_dim" attributes.38define ptx_kernel void @kernel3(ptr %input, ptr %output) #0 #5 #6 {39; CHECK-LABEL: kernel3(40; CHECK: .reqntid 512, 241; CHECK-NEXT: .reqnctapercluster 2, 242; CHECK-NEXT: .blocksareclusters // @kernel343; CHECK-NEXT: {44; CHECK-EMPTY:45; CHECK-EMPTY:46; CHECK-NEXT: // %bb.0:47; CHECK-NEXT: ret;48 ret void49}50 51attributes #0 = { "nvvm.blocksareclusters" }52 53attributes #1 = { "nvvm.reqntid"="1024,1,1" }54attributes #2 = { "nvvm.cluster_dim"="2,2,2" }55 56attributes #3 = { "nvvm.reqntid"="1024" }57attributes #4 = { "nvvm.cluster_dim"="2" }58 59attributes #5 = { "nvvm.reqntid"="512,2" }60attributes #6 = { "nvvm.cluster_dim"="2,2" }61