427 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80| FileCheck --check-prefixes=CHECK-PTX %s3; RUN: %if ptxas-sm_90 && ptxas-isa-8.0 %{ llc < %s -mtriple=nvptx64 -mcpu=sm_90 -mattr=+ptx80| %ptxas-verify -arch=sm_90 %}4 5target triple = "nvptx64-nvidia-cuda"6 7declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.tile.1d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i64 %ch, i1 %flag_ch);8declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.tile.2d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i64 %ch, i1 %flag_ch);9declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.tile.3d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 %flag_ch);10declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.tile.4d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 %flag_ch);11declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.tile.5d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 %flag_ch);12 13declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.im2col.3d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 %flag_ch);14declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.im2col.4d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 %flag_ch);15declare void @llvm.nvvm.cp.async.bulk.tensor.reduce.im2col.5d(ptr addrspace(3) %s, ptr %tm, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 %flag_ch);16 17; CHECK-LABEL: cp_async_bulk_tensor_reduce_tile_1d18define void @cp_async_bulk_tensor_reduce_tile_1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch) {19; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_tile_1d(20; CHECK-PTX: {21; CHECK-PTX-NEXT: .reg .b32 %r<2>;22; CHECK-PTX-NEXT: .reg .b64 %rd<4>;23; CHECK-PTX-EMPTY:24; CHECK-PTX-NEXT: // %bb.0:25; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_tile_1d_param_0];26; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_tile_1d_param_1];27; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_tile_1d_param_2];28; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_tile_1d_param_3];29; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.add.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;30; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.min.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;31; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.max.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;32; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.inc.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;33; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.dec.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;34; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.and.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;35; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.or.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;36; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.xor.tile.bulk_group.L2::cache_hint [%rd2, {%r1}], [%rd1], %rd3;37; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.add.tile.bulk_group [%rd2, {%r1}], [%rd1];38; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.min.tile.bulk_group [%rd2, {%r1}], [%rd1];39; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.max.tile.bulk_group [%rd2, {%r1}], [%rd1];40; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.inc.tile.bulk_group [%rd2, {%r1}], [%rd1];41; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.dec.tile.bulk_group [%rd2, {%r1}], [%rd1];42; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.and.tile.bulk_group [%rd2, {%r1}], [%rd1];43; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.or.tile.bulk_group [%rd2, {%r1}], [%rd1];44; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.1d.global.shared::cta.xor.tile.bulk_group [%rd2, {%r1}], [%rd1];45; CHECK-PTX-NEXT: ret;46 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)47 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)48 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)49 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)50 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)51 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)52 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)53 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 1)54 55 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)56 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)57 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)58 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)59 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)60 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)61 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)62 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.1d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i64 %ch, i1 0)63 ret void64}65 66; CHECK-LABEL: cp_async_bulk_tensor_reduce_tile_2d67define void @cp_async_bulk_tensor_reduce_tile_2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch) {68; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_tile_2d(69; CHECK-PTX: {70; CHECK-PTX-NEXT: .reg .b32 %r<3>;71; CHECK-PTX-NEXT: .reg .b64 %rd<4>;72; CHECK-PTX-EMPTY:73; CHECK-PTX-NEXT: // %bb.0:74; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_tile_2d_param_0];75; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_tile_2d_param_1];76; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_tile_2d_param_2];77; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_tile_2d_param_3];78; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_tile_2d_param_4];79; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.add.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;80; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.min.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;81; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.max.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;82; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.inc.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;83; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.dec.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;84; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.and.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;85; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.or.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;86; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.xor.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2}], [%rd1], %rd3;87; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.add.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];88; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.min.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];89; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.max.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];90; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.inc.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];91; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.dec.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];92; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.and.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];93; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.or.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];94; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.2d.global.shared::cta.xor.tile.bulk_group [%rd2, {%r1, %r2}], [%rd1];95; CHECK-PTX-NEXT: ret;96 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)97 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)98 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)99 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)100 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)101 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)102 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)103 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 1)104 105 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)106 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)107 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)108 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)109 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)110 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)111 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)112 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.2d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i64 %ch, i1 0)113 ret void114}115 116; CHECK-LABEL: cp_async_bulk_tensor_reduce_tile_3d117define void @cp_async_bulk_tensor_reduce_tile_3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch) {118; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_tile_3d(119; CHECK-PTX: {120; CHECK-PTX-NEXT: .reg .b32 %r<4>;121; CHECK-PTX-NEXT: .reg .b64 %rd<4>;122; CHECK-PTX-EMPTY:123; CHECK-PTX-NEXT: // %bb.0:124; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_tile_3d_param_0];125; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_tile_3d_param_1];126; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_tile_3d_param_2];127; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_tile_3d_param_3];128; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_tile_3d_param_4];129; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_tile_3d_param_5];130; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;131; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.min.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;132; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.max.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;133; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.inc.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;134; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.dec.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;135; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.and.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;136; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.or.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;137; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.xor.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;138; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];139; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.min.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];140; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.max.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];141; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.inc.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];142; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.dec.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];143; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.and.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];144; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.or.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];145; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.xor.tile.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];146; CHECK-PTX-NEXT: ret;147 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)148 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)149 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)150 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)151 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)152 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)153 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)154 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)155 156 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)157 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)158 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)159 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)160 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)161 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)162 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)163 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)164 ret void165}166 167; CHECK-LABEL: cp_async_bulk_tensor_reduce_tile_4d168define void @cp_async_bulk_tensor_reduce_tile_4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch) {169; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_tile_4d(170; CHECK-PTX: {171; CHECK-PTX-NEXT: .reg .b32 %r<5>;172; CHECK-PTX-NEXT: .reg .b64 %rd<4>;173; CHECK-PTX-EMPTY:174; CHECK-PTX-NEXT: // %bb.0:175; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_tile_4d_param_0];176; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_tile_4d_param_1];177; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_tile_4d_param_2];178; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_tile_4d_param_3];179; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_tile_4d_param_4];180; CHECK-PTX-NEXT: ld.param.b32 %r4, [cp_async_bulk_tensor_reduce_tile_4d_param_5];181; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_tile_4d_param_6];182; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;183; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.min.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;184; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.max.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;185; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.inc.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;186; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.dec.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;187; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.and.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;188; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.or.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;189; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.xor.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;190; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];191; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.min.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];192; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.max.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];193; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.inc.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];194; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.dec.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];195; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.and.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];196; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.or.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];197; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.xor.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];198; CHECK-PTX-NEXT: ret;199 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)200 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)201 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)202 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)203 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)204 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)205 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)206 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)207 208 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)209 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)210 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)211 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)212 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)213 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)214 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)215 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)216 ret void217}218 219; CHECK-LABEL: cp_async_bulk_tensor_reduce_tile_5d220define void @cp_async_bulk_tensor_reduce_tile_5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch) {221; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_tile_5d(222; CHECK-PTX: {223; CHECK-PTX-NEXT: .reg .b32 %r<6>;224; CHECK-PTX-NEXT: .reg .b64 %rd<4>;225; CHECK-PTX-EMPTY:226; CHECK-PTX-NEXT: // %bb.0:227; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_tile_5d_param_0];228; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_tile_5d_param_1];229; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_tile_5d_param_2];230; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_tile_5d_param_3];231; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_tile_5d_param_4];232; CHECK-PTX-NEXT: ld.param.b32 %r4, [cp_async_bulk_tensor_reduce_tile_5d_param_5];233; CHECK-PTX-NEXT: ld.param.b32 %r5, [cp_async_bulk_tensor_reduce_tile_5d_param_6];234; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_tile_5d_param_7];235; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.add.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;236; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.min.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;237; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.max.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;238; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.inc.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;239; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.dec.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;240; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.and.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;241; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.or.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;242; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.xor.tile.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;243; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.add.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];244; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.min.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];245; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.max.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];246; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.inc.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];247; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.dec.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];248; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.and.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];249; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.or.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];250; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.xor.tile.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];251; CHECK-PTX-NEXT: ret;252 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)253 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)254 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)255 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)256 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)257 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)258 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)259 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)260 261 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)262 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)263 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)264 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)265 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)266 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)267 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)268 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.tile.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)269 ret void270}271 272; CHECK-LABEL: cp_async_bulk_tensor_reduce_im2col_3d273define void @cp_async_bulk_tensor_reduce_im2col_3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch) {274; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_im2col_3d(275; CHECK-PTX: {276; CHECK-PTX-NEXT: .reg .b32 %r<4>;277; CHECK-PTX-NEXT: .reg .b64 %rd<4>;278; CHECK-PTX-EMPTY:279; CHECK-PTX-NEXT: // %bb.0:280; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_im2col_3d_param_0];281; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_im2col_3d_param_1];282; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_im2col_3d_param_2];283; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_im2col_3d_param_3];284; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_im2col_3d_param_4];285; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_im2col_3d_param_5];286; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;287; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.min.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;288; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.max.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;289; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.inc.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;290; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.dec.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;291; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.and.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;292; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.or.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;293; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.xor.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3}], [%rd1], %rd3;294; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.add.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];295; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.min.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];296; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.max.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];297; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.inc.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];298; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.dec.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];299; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.and.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];300; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.or.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];301; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.3d.global.shared::cta.xor.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3}], [%rd1];302; CHECK-PTX-NEXT: ret;303 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)304 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)305 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)306 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)307 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)308 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)309 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)310 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 1)311 312 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)313 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)314 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)315 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)316 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)317 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)318 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)319 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.3d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i64 %ch, i1 0)320 ret void321}322 323; CHECK-LABEL: cp_async_bulk_tensor_reduce_im2col_4d324define void @cp_async_bulk_tensor_reduce_im2col_4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch) {325; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_im2col_4d(326; CHECK-PTX: {327; CHECK-PTX-NEXT: .reg .b32 %r<5>;328; CHECK-PTX-NEXT: .reg .b64 %rd<4>;329; CHECK-PTX-EMPTY:330; CHECK-PTX-NEXT: // %bb.0:331; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_im2col_4d_param_0];332; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_im2col_4d_param_1];333; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_im2col_4d_param_2];334; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_im2col_4d_param_3];335; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_im2col_4d_param_4];336; CHECK-PTX-NEXT: ld.param.b32 %r4, [cp_async_bulk_tensor_reduce_im2col_4d_param_5];337; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_im2col_4d_param_6];338; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;339; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.min.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;340; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.max.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;341; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.inc.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;342; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.dec.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;343; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.and.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;344; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.or.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;345; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.xor.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4}], [%rd1], %rd3;346; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.add.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];347; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.min.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];348; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.max.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];349; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.inc.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];350; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.dec.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];351; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.and.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];352; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.or.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];353; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.4d.global.shared::cta.xor.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4}], [%rd1];354; CHECK-PTX-NEXT: ret;355 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)356 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)357 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)358 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)359 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)360 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)361 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)362 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 1)363 364 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)365 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)366 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)367 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)368 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)369 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)370 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)371 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.4d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i64 %ch, i1 0)372 ret void373}374 375; CHECK-LABEL: cp_async_bulk_tensor_reduce_im2col_5d376define void @cp_async_bulk_tensor_reduce_im2col_5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch) {377; CHECK-PTX-LABEL: cp_async_bulk_tensor_reduce_im2col_5d(378; CHECK-PTX: {379; CHECK-PTX-NEXT: .reg .b32 %r<6>;380; CHECK-PTX-NEXT: .reg .b64 %rd<4>;381; CHECK-PTX-EMPTY:382; CHECK-PTX-NEXT: // %bb.0:383; CHECK-PTX-NEXT: ld.param.b64 %rd1, [cp_async_bulk_tensor_reduce_im2col_5d_param_0];384; CHECK-PTX-NEXT: ld.param.b64 %rd2, [cp_async_bulk_tensor_reduce_im2col_5d_param_1];385; CHECK-PTX-NEXT: ld.param.b32 %r1, [cp_async_bulk_tensor_reduce_im2col_5d_param_2];386; CHECK-PTX-NEXT: ld.param.b32 %r2, [cp_async_bulk_tensor_reduce_im2col_5d_param_3];387; CHECK-PTX-NEXT: ld.param.b32 %r3, [cp_async_bulk_tensor_reduce_im2col_5d_param_4];388; CHECK-PTX-NEXT: ld.param.b32 %r4, [cp_async_bulk_tensor_reduce_im2col_5d_param_5];389; CHECK-PTX-NEXT: ld.param.b32 %r5, [cp_async_bulk_tensor_reduce_im2col_5d_param_6];390; CHECK-PTX-NEXT: ld.param.b64 %rd3, [cp_async_bulk_tensor_reduce_im2col_5d_param_7];391; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.add.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;392; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.min.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;393; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.max.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;394; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.inc.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;395; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.dec.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;396; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.and.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;397; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.or.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;398; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.xor.im2col_no_offs.bulk_group.L2::cache_hint [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1], %rd3;399; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.add.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];400; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.min.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];401; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.max.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];402; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.inc.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];403; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.dec.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];404; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.and.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];405; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.or.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];406; CHECK-PTX-NEXT: cp.reduce.async.bulk.tensor.5d.global.shared::cta.xor.im2col_no_offs.bulk_group [%rd2, {%r1, %r2, %r3, %r4, %r5}], [%rd1];407; CHECK-PTX-NEXT: ret;408 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)409 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)410 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)411 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)412 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)413 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)414 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)415 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 1)416 417 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.add.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)418 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.min.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)419 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.max.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)420 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.inc.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)421 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.dec.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)422 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.and.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)423 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.or.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)424 tail call void @llvm.nvvm.cp.async.bulk.tensor.reduce.xor.im2col.5d(ptr addrspace(3) %src, ptr %tmap, i32 %d0, i32 %d1, i32 %d2, i32 %d3, i32 %d4, i64 %ch, i1 0)425 ret void426}427