572 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc < %s -o - -mcpu=sm_100a -march=nvptx64 -mattr=+ptx86 | FileCheck %s3; RUN: llc < %s -o - -mcpu=sm_101a -march=nvptx64 -mattr=+ptx86 | FileCheck %s4; RUN: llc < %s -o - -mcpu=sm_110a -march=nvptx64 -mattr=+ptx90 | FileCheck %s5; RUN: %if ptxas-sm_100a && ptxas-isa-8.6 %{ llc < %s -march=nvptx64 -mattr=+ptx86 -mcpu=sm_100a | %ptxas-verify -arch=sm_100a %}6; RUN: %if ptxas-sm_101a && ptxas-isa-8.6 %{ llc < %s -march=nvptx64 -mattr=+ptx86 -mcpu=sm_101a | %ptxas-verify -arch=sm_101a %}7; RUN: %if ptxas-sm_110a && ptxas-isa-9.0 %{ llc < %s -march=nvptx64 -mattr=+ptx90 -mcpu=sm_110a | %ptxas-verify -arch=sm_110a %}8 9define void @tcgen05_mma_ws_fp16(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d) {10; CHECK-LABEL: tcgen05_mma_ws_fp16(11; CHECK: {12; CHECK-NEXT: .reg .pred %p<2>;13; CHECK-NEXT: .reg .b16 %rs<3>;14; CHECK-NEXT: .reg .b32 %r<4>;15; CHECK-NEXT: .reg .b64 %rd<3>;16; CHECK-EMPTY:17; CHECK-NEXT: // %bb.0:18; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_fp16_param_5];19; CHECK-NEXT: and.b16 %rs2, %rs1, 1;20; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;21; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_fp16_param_0];22; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_fp16_param_2];23; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_fp16_param_3];24; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_fp16_param_4];25; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;26; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_fp16_param_1];27; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;28; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;29; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;30; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;31; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;32; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;33; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;34; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], [%r3], %rd2, %r2, %p1;35; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;36; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;37; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;38; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;39; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;40; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;41; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;42; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;43; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;44; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;45; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;46; CHECK-NEXT: ret;47 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)48 49 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)50 51 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)52 53 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)54 55 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)56 57 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 1)58 59 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 0)60 61 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 1)62 63 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 1)64 65 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 2)66 67 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 2)68 69 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 2)70 71 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 2)72 73 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 2)74 75 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)76 77 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)78 79 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)80 81 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)82 83 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)84 85 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 0, i32 0, i32 3)86 ret void87}88 89define void @tcgen05_mma_ws_fp16_zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask) {90; CHECK-LABEL: tcgen05_mma_ws_fp16_zero_col_mask(91; CHECK: {92; CHECK-NEXT: .reg .pred %p<2>;93; CHECK-NEXT: .reg .b16 %rs<3>;94; CHECK-NEXT: .reg .b32 %r<4>;95; CHECK-NEXT: .reg .b64 %rd<4>;96; CHECK-EMPTY:97; CHECK-NEXT: // %bb.0:98; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_fp16_zero_col_mask_param_5];99; CHECK-NEXT: and.b16 %rs2, %rs1, 1;100; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;101; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_fp16_zero_col_mask_param_0];102; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_fp16_zero_col_mask_param_2];103; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_fp16_zero_col_mask_param_3];104; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_fp16_zero_col_mask_param_4];105; CHECK-NEXT: ld.param.b64 %rd3, [tcgen05_mma_ws_fp16_zero_col_mask_param_6];106; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1, %rd3;107; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_fp16_zero_col_mask_param_1];108; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1, %rd3;109; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1, %rd3;110; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1, %rd3;111; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1, %rd3;112; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1, %rd3;113; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1, %rd3;114; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1, %rd3;115; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::lastuse [%r1], [%r3], %rd2, %r2, %p1, %rd3;116; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1, %rd3;117; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1, %rd3;118; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1, %rd3;119; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1, %rd3;120; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1, %rd3;121; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1, %rd3;122; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1, %rd3;123; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1, %rd3;124; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1, %rd3;125; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1, %rd3;126; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f16.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1, %rd3;127; CHECK-NEXT: ret;128 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)129 130 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)131 132 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)133 134 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)135 136 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)137 138 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 1)139 140 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 0)141 142 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 1)143 144 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 1)145 146 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 2)147 148 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 2)149 150 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 2)151 152 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 2)153 154 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 2)155 156 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)157 158 call void @llvm.nvvm.tcgen05.mma.ws.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)159 160 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)161 162 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)163 164 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)165 166 call void @llvm.nvvm.tcgen05.mma.ws.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, i32 0, i32 0, i32 3)167 ret void168}169 170define void @tcgen05_mma_ws_sp_fp16(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta) {171; CHECK-LABEL: tcgen05_mma_ws_sp_fp16(172; CHECK: {173; CHECK-NEXT: .reg .pred %p<2>;174; CHECK-NEXT: .reg .b16 %rs<3>;175; CHECK-NEXT: .reg .b32 %r<5>;176; CHECK-NEXT: .reg .b64 %rd<3>;177; CHECK-EMPTY:178; CHECK-NEXT: // %bb.0:179; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_sp_fp16_param_5];180; CHECK-NEXT: and.b16 %rs2, %rs1, 1;181; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;182; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_sp_fp16_param_0];183; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_sp_fp16_param_2];184; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_sp_fp16_param_3];185; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_sp_fp16_param_4];186; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_sp_fp16_param_6];187; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, [%r3], %r2, %p1;188; CHECK-NEXT: ld.param.b32 %r4, [tcgen05_mma_ws_sp_fp16_param_1];189; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, [%r3], %r2, %p1;190; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1;191; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1;192; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1;193; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, [%r3], %r2, %p1;194; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1;195; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, [%r3], %r2, %p1;196; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], [%r4], %rd2, [%r3], %r2, %p1;197; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1;198; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, [%r3], %r2, %p1;199; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, [%r3], %r2, %p1;200; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1;201; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1;202; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, [%r3], %r2, %p1;203; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, [%r3], %r2, %p1;204; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1;205; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1;206; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1;207; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1;208; CHECK-NEXT: ret;209 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)210 211 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)212 213 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)214 215 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)216 217 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)218 219 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 1)220 221 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 0)222 223 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 1)224 225 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 1)226 227 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 2)228 229 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 2)230 231 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 2)232 233 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 2)234 235 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 2)236 237 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)238 239 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)240 241 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)242 243 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)244 245 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)246 247 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i32 0, i32 0, i32 3)248 ret void249}250 251define void @tcgen05_mma_ws_sp_fp16_zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i64 %zero_col_mask, ptr addrspace(6) %spmeta) {252; CHECK-LABEL: tcgen05_mma_ws_sp_fp16_zero_col_mask(253; CHECK: {254; CHECK-NEXT: .reg .pred %p<2>;255; CHECK-NEXT: .reg .b16 %rs<3>;256; CHECK-NEXT: .reg .b32 %r<5>;257; CHECK-NEXT: .reg .b64 %rd<4>;258; CHECK-EMPTY:259; CHECK-NEXT: // %bb.0:260; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_5];261; CHECK-NEXT: and.b16 %rs2, %rs1, 1;262; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;263; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_0];264; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_2];265; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_3];266; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_4];267; CHECK-NEXT: ld.param.b64 %rd3, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_6];268; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_7];269; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;270; CHECK-NEXT: ld.param.b32 %r4, [tcgen05_mma_ws_sp_fp16_zero_col_mask_param_1];271; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;272; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;273; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;274; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;275; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;276; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::discard [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;277; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;278; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::lastuse [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;279; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;280; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;281; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;282; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;283; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::fill [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;284; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;285; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], %rd1, %rd2, [%r3], %r2, %p1, %rd3;286; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;287; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;288; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;289; CHECK-NEXT: tcgen05.mma.ws.sp.cta_group::1.kind::f16.collector::b0::use [%r1], [%r4], %rd2, [%r3], %r2, %p1, %rd3;290; CHECK-NEXT: ret;291 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)292 293 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)294 295 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)296 297 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)298 299 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)300 301 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 1)302 303 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 0)304 305 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 1)306 307 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 1)308 309 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 2)310 311 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 2)312 313 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 2)314 315 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 2)316 317 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 2)318 319 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)320 321 call void @llvm.nvvm.tcgen05.mma.ws.sp.shared.zero_col_mask(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)322 323 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)324 325 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)326 327 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)328 329 call void @llvm.nvvm.tcgen05.mma.ws.sp.tensor.zero_col_mask(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, ptr addrspace(6) %spmeta, i64 %zero_col_mask, i32 0, i32 0, i32 3)330 ret void331}332 333define void @tcgen05_mma_ws_tf32(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d) {334; CHECK-LABEL: tcgen05_mma_ws_tf32(335; CHECK: {336; CHECK-NEXT: .reg .pred %p<2>;337; CHECK-NEXT: .reg .b16 %rs<3>;338; CHECK-NEXT: .reg .b32 %r<4>;339; CHECK-NEXT: .reg .b64 %rd<3>;340; CHECK-EMPTY:341; CHECK-NEXT: // %bb.0:342; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_tf32_param_5];343; CHECK-NEXT: and.b16 %rs2, %rs1, 1;344; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;345; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_tf32_param_0];346; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_tf32_param_2];347; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_tf32_param_3];348; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_tf32_param_4];349; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;350; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_tf32_param_1];351; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;352; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;353; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;354; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;355; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;356; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;357; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;358; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::lastuse [%r1], [%r3], %rd2, %r2, %p1;359; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;360; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;361; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;362; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;363; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;364; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;365; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;366; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;367; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;368; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;369; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::tf32.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;370; CHECK-NEXT: ret;371 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)372 373 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)374 375 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)376 377 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)378 379 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)380 381 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 1)382 383 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 0)384 385 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 1)386 387 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 1)388 389 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 2)390 391 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 2)392 393 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 2)394 395 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 2)396 397 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 2)398 399 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)400 401 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)402 403 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)404 405 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)406 407 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)408 409 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 1, i32 0, i32 3)410 ret void411}412 413define void @tcgen05_mma_ws_f8f6f4(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d) {414; CHECK-LABEL: tcgen05_mma_ws_f8f6f4(415; CHECK: {416; CHECK-NEXT: .reg .pred %p<2>;417; CHECK-NEXT: .reg .b16 %rs<3>;418; CHECK-NEXT: .reg .b32 %r<4>;419; CHECK-NEXT: .reg .b64 %rd<3>;420; CHECK-EMPTY:421; CHECK-NEXT: // %bb.0:422; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_f8f6f4_param_5];423; CHECK-NEXT: and.b16 %rs2, %rs1, 1;424; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;425; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_f8f6f4_param_0];426; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_f8f6f4_param_2];427; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_f8f6f4_param_3];428; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_f8f6f4_param_4];429; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;430; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_f8f6f4_param_1];431; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;432; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;433; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;434; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;435; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;436; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;437; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;438; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::lastuse [%r1], [%r3], %rd2, %r2, %p1;439; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;440; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;441; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;442; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;443; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;444; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;445; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;446; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;447; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;448; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;449; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::f8f6f4.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;450; CHECK-NEXT: ret;451 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)452 453 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)454 455 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)456 457 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)458 459 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)460 461 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 1)462 463 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 0)464 465 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 1)466 467 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 1)468 469 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 2)470 471 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 2)472 473 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 2)474 475 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 2)476 477 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 2)478 479 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)480 481 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)482 483 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)484 485 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)486 487 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)488 489 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 2, i32 0, i32 3)490 ret void491}492 493define void @tcgen05_mma_ws_i8(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d) {494; CHECK-LABEL: tcgen05_mma_ws_i8(495; CHECK: {496; CHECK-NEXT: .reg .pred %p<2>;497; CHECK-NEXT: .reg .b16 %rs<3>;498; CHECK-NEXT: .reg .b32 %r<4>;499; CHECK-NEXT: .reg .b64 %rd<3>;500; CHECK-EMPTY:501; CHECK-NEXT: // %bb.0:502; CHECK-NEXT: ld.param.b8 %rs1, [tcgen05_mma_ws_i8_param_5];503; CHECK-NEXT: and.b16 %rs2, %rs1, 1;504; CHECK-NEXT: setp.ne.b16 %p1, %rs2, 0;505; CHECK-NEXT: ld.param.b32 %r1, [tcgen05_mma_ws_i8_param_0];506; CHECK-NEXT: ld.param.b64 %rd1, [tcgen05_mma_ws_i8_param_2];507; CHECK-NEXT: ld.param.b64 %rd2, [tcgen05_mma_ws_i8_param_3];508; CHECK-NEXT: ld.param.b32 %r2, [tcgen05_mma_ws_i8_param_4];509; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;510; CHECK-NEXT: ld.param.b32 %r3, [tcgen05_mma_ws_i8_param_1];511; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], %rd1, %rd2, %r2, %p1;512; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;513; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;514; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;515; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;516; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::discard [%r1], [%r3], %rd2, %r2, %p1;517; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::lastuse [%r1], %rd1, %rd2, %r2, %p1;518; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::lastuse [%r1], [%r3], %rd2, %r2, %p1;519; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;520; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;521; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::fill [%r1], %rd1, %rd2, %r2, %p1;522; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;523; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::fill [%r1], [%r3], %rd2, %r2, %p1;524; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;525; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], %rd1, %rd2, %r2, %p1;526; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;527; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;528; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;529; CHECK-NEXT: tcgen05.mma.ws.cta_group::1.kind::i8.collector::b0::use [%r1], [%r3], %rd2, %r2, %p1;530; CHECK-NEXT: ret;531 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)532 533 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)534 535 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)536 537 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)538 539 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)540 541 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 1)542 543 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 0)544 545 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 1)546 547 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 1)548 549 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 2)550 551 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 2)552 553 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 2)554 555 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 2)556 557 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 2)558 559 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)560 561 call void @llvm.nvvm.tcgen05.mma.ws.shared(ptr addrspace(6) %dtmem, i64 %ashared, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)562 563 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)564 565 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)566 567 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)568 569 call void @llvm.nvvm.tcgen05.mma.ws.tensor(ptr addrspace(6) %dtmem, ptr addrspace(6) %atensor, i64 %b, i32 %idesc, i1 %enable_inp_d, i32 3, i32 0, i32 3)570 ret void571}572