634 lines · plain
1// RUN: mlir-translate --mlir-to-llvmir %s | FileCheck %s2 3// CHECK-LABEL: @nvvm_tcgen05_mma_cta_14llvm.func @nvvm_tcgen05_mma_cta_1(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1) {5 6 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 1, /* collector=discard */ i32 0)7 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d8 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)9 10 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 1, /* collector=discard */ i32 0)11 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d12 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)13 14 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 1, /* collector=discard */ i32 0)15 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d16 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)17 18 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 1, /* collector=discard */ i32 0)19 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d20 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)21 22 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 1, /* collector=lastuse */ i32 1)23 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d24 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)25 26 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 1, /* collector=lastuse */ i32 1)27 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d28 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)29 30 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 1, /* collector=lastuse */ i32 1)31 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d32 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)33 34 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 1, /* collector=lastuse */ i32 1)35 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d36 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)37 38 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 0, i32 1, i32 0)39 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d40 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)41 42 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 1, i32 1, i32 0)43 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d44 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)45 46 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 2, i32 1, i32 0)47 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d48 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)49 50 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 3, i32 1, i32 0)51 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d52 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)53 54 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 0, i32 1, i32 1)55 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d56 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)57 58 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 1, i32 1, i32 1)59 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d60 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)61 62 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 2, i32 1, i32 1)63 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d64 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)65 66 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 3, i32 1, i32 1)67 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d68 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)69 70 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 1, /* collector=fill */ i32 2)71 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d72 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)73 74 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 1, /* collector=fill */ i32 2)75 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d76 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)77 78 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 1, /* collector=fill */ i32 2)79 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d80 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)81 82 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 1, /* collector=fill */ i32 2)83 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d84 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)85 86 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 1, /* collector=use */ i32 3)87 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d88 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)89 90 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 1, /* collector=use */ i32 3)91 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d92 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)93 94 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 1, /* collector=use */ i32 3)95 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d96 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)97 98 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 1, /* collector=use */ i32 3)99 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d100 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)101 102 llvm.return103}104 105// CHECK-LABEL: @nvvm_tcgen05_mma_cta_2106llvm.func @nvvm_tcgen05_mma_cta_2(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1) {107 108 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 2, /* collector=discard */ i32 0)109 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d110 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)111 112 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 2, /* collector=discard */ i32 0)113 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d114 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)115 116 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 2, /* collector=discard */ i32 0)117 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d118 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)119 120 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 2, /* collector=discard */ i32 0)121 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d122 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)123 124 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 2, /* collector=lastuse */ i32 1)125 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d126 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)127 128 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 2, /* collector=lastuse */ i32 1)129 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d130 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)131 132 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 2, /* collector=lastuse */ i32 1)133 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d134 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)135 136 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 2, /* collector=lastuse */ i32 1)137 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d138 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)139 140 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 0, i32 2, i32 0)141 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d142 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)143 144 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 1, i32 2, i32 0)145 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d146 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)147 148 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 2, i32 2, i32 0)149 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d150 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)151 152 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 3, i32 2, i32 0)153 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d154 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)155 156 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 0, i32 2, i32 1)157 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d158 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)159 160 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 1, i32 2, i32 1)161 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d162 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)163 164 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 2, i32 2, i32 1)165 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d166 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)167 168 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i32 3, i32 2, i32 1)169 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d170 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)171 172 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 2, /* collector=fill */ i32 2)173 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d174 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)175 176 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 2, /* collector=fill */ i32 2)177 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d178 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)179 180 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 2, /* collector=fill */ i32 2)181 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d182 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)183 184 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 2, /* collector=fill */ i32 2)185 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d186 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)187 188 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f16 */ i32 0, /* cta_group= */ i32 2, /* collector=use */ i32 3)189 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d190 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)191 192 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=tf32 */ i32 1, /* cta_group= */ i32 2, /* collector=use */ i32 3)193 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d194 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)195 196 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=f8f6f4 */ i32 2, /* cta_group= */ i32 2, /* collector=use */ i32 3)197 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d198 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)199 200 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, /* kind=i8 */ i32 3, /* cta_group= */ i32 2, /* collector=use */ i32 3)201 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d202 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1)203 204 llvm.return205}206 207// CHECK-LABEL: @nvvm_tcgen05_mma_scale_d_imm_cta_1208llvm.func @nvvm_tcgen05_mma_scale_d_imm_cta_1(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1) {209 210 %scale_d_imm = llvm.mlir.constant(0:i64) : i64211 212 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 0)213 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm214 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)215 216 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 0)217 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm218 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)219 220 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 1)221 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm222 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)223 224 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 1)225 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm226 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)227 228 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 0)229 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm230 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)231 232 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 0)233 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm234 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)235 236 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 1)237 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm238 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)239 240 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 1)241 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm242 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)243 244 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 2)245 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm246 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)247 248 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 2)249 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm250 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)251 252 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 1, i32 3)253 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm254 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)255 256 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 1, i32 3)257 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm258 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)259 260 llvm.return261}262 263// CHECK-LABEL: @nvvm_tcgen05_mma_scale_d_imm_cta_2264llvm.func @nvvm_tcgen05_mma_scale_d_imm_cta_2(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1) {265 266 %scale_d_imm = llvm.mlir.constant(0:i64) : i64267 268 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 0)269 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm270 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)271 272 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 0)273 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm274 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)275 276 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 1)277 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm278 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)279 280 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 1)281 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm282 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)283 284 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 0)285 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm286 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)287 288 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 0)289 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm290 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)291 292 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 1)293 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm294 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)295 296 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 1)297 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm298 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)299 300 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 2)301 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm302 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)303 304 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 2)305 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm306 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)307 308 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 0, i32 2, i32 3)309 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm310 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)311 312 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, i32 1, i32 2, i32 3)313 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm314 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64)315 316 llvm.return317}318 319// CHECK-LABEL: @nvvm_tcgen05_mma_disable_output_lane_cta_1320llvm.func @nvvm_tcgen05_mma_disable_output_lane_cta_1(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %disableOutputLane : vector<4 x i32>) {321 322 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 0)323 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane324 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)325 326 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 0)327 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane328 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)329 330 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 0)331 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane332 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)333 334 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 0)335 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane336 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)337 338 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 1)339 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane340 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)341 342 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 1)343 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane344 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)345 346 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 1)347 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane348 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)349 350 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 1)351 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane352 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)353 354 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 0)355 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane356 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)357 358 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 0)359 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane360 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)361 362 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 0)363 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane364 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)365 366 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 0)367 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane368 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)369 370 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 1)371 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane372 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)373 374 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 1)375 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane376 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)377 378 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 1)379 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane380 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)381 382 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 1)383 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane384 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)385 386 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 2)387 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane388 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)389 390 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 2)391 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane392 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)393 394 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 2)395 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane396 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)397 398 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 2)399 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane400 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)401 402 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 0, i32 3)403 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane404 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)405 406 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 1, i32 3)407 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane408 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)409 410 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 2, i32 3)411 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane412 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)413 414 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <4 x i32> {{%[0-9]+}}, i32 3, i32 3)415 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane416 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<4 x i32>)417 418 llvm.return419}420 421// CHECK-LABEL: @nvvm_tcgen05_mma_disable_output_lane_cta_2422llvm.func @nvvm_tcgen05_mma_disable_output_lane_cta_2(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %disableOutputLane: vector<8 x i32>) {423 424 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 0)425 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane426 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)427 428 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 0)429 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane430 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)431 432 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 0)433 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane434 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)435 436 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 0)437 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane438 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)439 440 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 1)441 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane442 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)443 444 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 1)445 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane446 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)447 448 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 1)449 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane450 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)451 452 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 1)453 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane454 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)455 456 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 0)457 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane458 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)459 460 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 0)461 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane462 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)463 464 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 0)465 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane466 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)467 468 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 0)469 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane470 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)471 472 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 1)473 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane474 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)475 476 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 1)477 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane478 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)479 480 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 1)481 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane482 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)483 484 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 1)485 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane486 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)487 488 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 2)489 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane490 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)491 492 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 2)493 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane494 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)495 496 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 2)497 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane498 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)499 500 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 2)501 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane502 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)503 504 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 0, i32 3)505 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane506 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)507 508 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 1, i32 3)509 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane510 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)511 512 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 2, i32 3)513 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane514 {kind = #nvvm.tcgen05_mma_kind<f8f6f4>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)515 516 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, <8 x i32> {{%[0-9]+}}, i32 3, i32 3)517 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d mask = %disableOutputLane518 {kind = #nvvm.tcgen05_mma_kind<i8>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, vector<8 x i32>)519 520 llvm.return521}522 523// CHECK-LABEL: @nvvm_tcgen05_mma_scale_d_imm_disable_output_lane_cta_1524llvm.func @nvvm_tcgen05_mma_scale_d_imm_disable_output_lane_cta_1(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %disableOutputLane: vector<4 x i32>) {525 526 %scale_d_imm = llvm.mlir.constant(0:i64) : i64527 528 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 0)529 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane530 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)531 532 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 0)533 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane534 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)535 536 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 1)537 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane538 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)539 540 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 1)541 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane542 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)543 544 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 0)545 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane546 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)547 548 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 0)549 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane550 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)551 552 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 1)553 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane554 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)555 556 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 1)557 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane558 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)559 560 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 2)561 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane562 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)563 564 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 2)565 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane566 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)567 568 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 0, i32 3)569 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane570 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)571 572 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg1(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <4 x i32> {{%[0-9]+}}, i32 1, i32 3)573 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane574 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_1>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<4 x i32>)575 576 llvm.return577}578 579// CHECK-LABEL: @nvvm_tcgen05_mma_scale_d_imm_disable_output_lane_cta_2580llvm.func @nvvm_tcgen05_mma_scale_d_imm_disable_output_lane_cta_2(%d_tmem : !llvm.ptr<6>, %a_tmem: !llvm.ptr<6>, %adesc: i64, %b_desc: i64, %idesc: i32, %enable_input_d: i1, %disableOutputLane: vector<8 x i32>) {581 582 %scale_d_imm = llvm.mlir.constant(0:i64) : i64583 584 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 0)585 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane586 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)587 588 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 0)589 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane590 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)591 592 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 1)593 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane594 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)595 596 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 1)597 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane598 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)599 600 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 0)601 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane602 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)603 604 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 0)605 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane606 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)607 608 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 1)609 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane610 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)611 612 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2.ashift(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 1)613 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane614 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<lastuse>, aShift} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)615 616 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 2)617 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane618 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)619 620 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 2)621 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane622 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<fill>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)623 624 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 0, i32 3)625 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane626 {kind = #nvvm.tcgen05_mma_kind<f16>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)627 628 // CHECK: call void @llvm.nvvm.tcgen05.mma.tensor.scale_d.disable_output_lane.cg2(ptr addrspace(6) {{%[0-9]+}}, ptr addrspace(6) {{%[0-9]+}}, i64 {{%[0-9]+}}, i32 {{%[0-9]+}}, i1 {{%[0-9]+}}, i64 0, <8 x i32> {{%[0-9]+}}, i32 1, i32 3)629 nvvm.tcgen05.mma %d_tmem, %a_tmem, %b_desc, %idesc, %enable_input_d scale = %scale_d_imm mask = %disableOutputLane630 {kind = #nvvm.tcgen05_mma_kind<tf32>, ctaGroup = #nvvm.cta_group<cta_2>, collectorOp = #nvvm.tcgen05_mma_collectorop<use>} : (!llvm.ptr<6>, !llvm.ptr<6>, i64, i32, i1, i64, vector<8 x i32>)631 632 llvm.return633}634