brintos

brintos / llvm-project-archived public Read only

0
0
Text · 24.9 KiB · 33665b6 Raw
325 lines · plain
1# RUN: llvm-mca -mtriple=amdgcn -mcpu=gfx950 --timeline --iterations=1 --timeline-max-cycles=0 < %s | FileCheck %s2 3# CHECK: Iterations:        14# CHECK: Instructions:      1335# CHECK: Total Cycles:      11016# CHECK: Total uOps:        1337 8v_mfma_f32_16x16x32_f16 a[0:3], a[0:3], a[0:3], a[0:3] blgp:19v_mfma_f32_16x16x32_f16 a[0:3], v[0:3], v[0:3], a[4:7]10v_mfma_f32_32x32x16_f16 v[0:15], v[0:3], v[0:3], v[0:15]11v_mfma_f32_32x32x16_f16 a[0:15], a[0:3], a[0:3], a[0:15] blgp:212v_mfma_f32_32x32x16_bf16 v[0:15], v[0:3], v[0:3], v[0:15]13v_mfma_f32_32x32x16_bf16 a[0:15], a[0:3], a[0:3], a[0:15] blgp:214v_mfma_i32_16x16x64_i8 a[0:3], a[0:3], a[0:3], a[0:3] blgp:115v_mfma_i32_16x16x64_i8 a[0:3], v[0:3], v[0:3], a[4:7]16v_mfma_i32_32x32x32_i8 v[0:15], v[0:3], v[0:3], v[0:15]17v_mfma_i32_32x32x32_i8 a[0:15], a[0:3], a[0:3], a[0:15] blgp:218v_mfma_f32_16x16x32_bf16 a[0:3], a[0:3], a[0:3], a[0:3] blgp:119v_mfma_f32_16x16x32_bf16 a[0:3], v[0:3], v[0:3], a[4:7]20 21v_mfma_ld_scale_b32 v0, v022 23;; FIXME: should have different cycle count depending on whether either matrix is f824;;  TODO: test vdc/adc25v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3]26v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3] blgp:127v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] blgp:228v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] blgp:329v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:7], v[0:3] blgp:430v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3] cbsz:131v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:232v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:333v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:7], v[4:11], v[0:3] cbsz:434v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:2 blgp:135v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] cbsz:1 blgp:236 37;; FIXME: should have different cycle count depending on whether either matrix is f838v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15]39v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] blgp:140v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:9], v[0:15] blgp:241v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:9], v[0:15] blgp:342v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:7], v[0:15] blgp:443v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] cbsz:144v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:245v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:346v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:7], v[4:11], v[0:15] cbsz:447v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:248v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] blgp:149 50;; FIXME: should have different cycle count depending on whether either matrix is f851v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3], v5, v552v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3], v5, v5 blgp:153v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:9], v[0:3], v5, v5 cbsz:2 blgp:254 55v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15], v5, v556v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15], v5, v5 cbsz:2 blgp:157v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:9], v[0:15], v5, v5 cbsz:2 blgp:258 59;;  TODO: These results are wrong60v_smfmac_f32_16x16x64_f16 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:161v_smfmac_f32_32x32x32_f16 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:162v_smfmac_f32_16x16x64_bf16 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:163v_smfmac_f32_32x32x32_bf16 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:164v_smfmac_i32_16x16x128_i8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:165v_smfmac_i32_32x32x64_i8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:166 67v_smfmac_f32_16x16x128_bf8_bf8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:168v_smfmac_f32_16x16x128_bf8_fp8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:169v_smfmac_f32_16x16x128_fp8_bf8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:170v_smfmac_f32_16x16x128_fp8_fp8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:171 72v_smfmac_f32_32x32x64_bf8_bf8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:173v_smfmac_f32_32x32x64_bf8_fp8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:174v_smfmac_f32_32x32x64_fp8_bf8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:175v_smfmac_f32_32x32x64_fp8_fp8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:176 77v_mfma_f32_16x16x4_f32 a[0:3], v0, v1, a[2:5]78v_mfma_f32_16x16x4_f32 v[0:3], v0, v1, v[2:5]79 80v_mfma_f32_32x32x2_f32 a[0:15], v0, v1, a[18:33]81v_mfma_f32_32x32x2_f32 v[0:15], v0, v1, v[18:33]82 83v_mfma_f64_4x4x4_4b_f64 a[0:1], v[0:1], a[2:3], a[2:3]84v_mfma_f64_4x4x4_4b_f64 v[0:1], v[0:1], v[2:3], v[2:3]85 86v_mfma_f64_16x16x4_f64 a[0:7], v[0:1], v[2:3], a[0:7]87v_mfma_f64_16x16x4_f64 v[0:7], v[0:1], v[2:3], v[0:7]88 89v_mfma_f32_16x16x16_f16 v[0:3], v[4:5], v[6:7], v[0:3]90v_mfma_f32_16x16x16_f16 a[0:3], v[4:5], v[6:7], a[0:3]91 92v_mfma_f32_32x32x8_f16 v[0:15], v[4:5], v[6:7], v[0:15]93v_mfma_f32_32x32x8_f16 a[0:15], v[4:5], v[6:7], a[0:15]94 95v_mfma_f32_16x16x16_bf16 v[0:3], v[4:5], v[6:7], v[0:3]96v_mfma_f32_16x16x16_bf16 a[0:3], v[4:5], v[6:7], a[0:3]97 98v_mfma_f32_32x32x8_bf16 v[0:15], v[4:5], v[6:7], v[0:15]99v_mfma_f32_32x32x8_bf16 a[0:15], v[4:5], v[6:7], a[0:15]100 101v_mfma_i32_16x16x32_i8 v[0:3], v[4:5], v[6:7], v[0:3]102v_mfma_i32_16x16x32_i8 a[0:3], v[4:5], v[6:7], a[0:3]103 104v_mfma_i32_32x32x16_i8 v[0:15], v[2:3], v[4:5], v[0:15]105v_mfma_i32_32x32x16_i8 a[0:15], v[2:3], v[4:5], a[0:15]106 107v_mfma_f32_4x4x4_16b_f16 v[0:3], v[0:1], v[2:3], v[2:5]108v_mfma_f32_4x4x4_16b_f16 a[0:3], v[0:1], v[2:3], a[2:5]109 110v_mfma_f32_16x16x4_4b_f16 v[0:15], v[2:3], v[4:5], v[18:33]111v_mfma_f32_16x16x4_4b_f16 a[0:15], v[2:3], v[4:5], a[18:33]112 113v_mfma_f32_32x32x4_2b_f16 v[0:31], v[0:1], v[2:3], v[34:65]114v_mfma_f32_32x32x4_2b_f16 a[0:31], v[0:1], v[2:3], a[34:65]115 116v_mfma_f32_4x4x4_16b_bf16 v[0:3], v[0:1], v[2:3], v[2:5]117v_mfma_f32_4x4x4_16b_bf16 a[0:3], v[0:1], v[2:3], a[2:5]118 119v_mfma_f32_16x16x4_4b_bf16 v[0:15], v[2:3], v[4:5], v[18:33]120v_mfma_f32_16x16x4_4b_bf16 a[0:15], v[2:3], v[4:5], a[18:33]121 122v_mfma_f32_32x32x4_2b_bf16 v[0:31], v[0:1], v[2:3], v[34:65]123v_mfma_f32_32x32x4_2b_bf16 a[0:31], v[0:1], v[2:3], a[34:65]124 125v_mfma_f32_4x4x1_16b_f32 v[0:3], v0, v1, v[2:5]126v_mfma_f32_4x4x1_16b_f32 a[0:3], v0, v1, a[2:5]127 128v_mfma_f32_16x16x1_4b_f32 v[0:15], v0, v1, v[18:33]129v_mfma_f32_16x16x1_4b_f32 a[0:15], v0, v1, a[18:33]130 131v_mfma_f32_16x16x4_f32 v[0:3], v0, v1, v[2:5]132v_mfma_f32_16x16x4_f32 a[0:3], v0, v1, a[2:5]133 134v_mfma_f32_32x32x1_2b_f32 v[0:31], v0, v1, v[34:65] blgp:7135v_mfma_f32_32x32x1_2b_f32 a[0:31], v0, v1, a[34:65] blgp:7136 137v_mfma_f32_32x32x2_f32 v[0:15], v0, v1, v[18:33]138v_mfma_f32_32x32x2_f32 a[0:15], v0, v1, a[18:33]139 140v_mfma_i32_4x4x4_16b_i8 v[0:3], v0, v1, v[2:5]141v_mfma_i32_4x4x4_16b_i8 a[0:3], v0, v1, a[2:5]142 143v_mfma_i32_16x16x4_4b_i8 v[0:15], v0, v1, v[18:33]144v_mfma_i32_16x16x4_4b_i8 a[0:15], v0, v1, a[18:33]145 146v_mfma_i32_32x32x4_2b_i8 v[0:31], v0, v1, v[34:65]147v_mfma_i32_32x32x4_2b_i8 a[0:31], v0, v1, a[34:65]148 149v_smfmac_f32_16x16x32_f16 v[10:13], a[2:3], v[4:7], v0 cbsz:3 abid:1150v_smfmac_f32_16x16x32_f16 a[10:13], v[2:3], a[4:7], v1151 152v_smfmac_f32_32x32x16_f16 v[10:25], a[2:3], v[4:7], v2 cbsz:3 abid:1153v_smfmac_f32_32x32x16_f16 a[10:25], v[2:3], a[4:7], v3154 155v_smfmac_f32_16x16x32_bf16 v[10:13], a[2:3], v[4:7], v4 cbsz:3 abid:1156v_smfmac_f32_16x16x32_bf16 a[10:13], v[2:3], a[4:7], v5157 158v_smfmac_i32_16x16x64_i8 v[10:13], a[2:3], v[4:7], v8 cbsz:3 abid:1159v_smfmac_i32_16x16x64_i8 a[10:13], v[2:3], a[4:7], v9160 161v_smfmac_i32_32x32x32_i8 v[10:25], a[2:3], v[4:7], v10 cbsz:3 abid:1162v_smfmac_i32_32x32x32_i8 a[10:25], v[2:3], a[4:7], v11163 164v_mfma_f32_16x16x32_bf8_bf8 v[0:3], v[2:3], v[4:5], v[0:3]165v_mfma_f32_16x16x32_bf8_bf8 a[0:3], v[2:3], v[4:5], a[0:3]166 167v_mfma_f32_16x16x32_bf8_fp8 v[0:3], v[2:3], v[4:5], v[0:3]168v_mfma_f32_16x16x32_bf8_fp8 a[0:3], v[2:3], v[4:5], a[0:3]169 170v_mfma_f32_16x16x32_fp8_bf8 v[0:3], v[2:3], v[4:5], v[0:3]171v_mfma_f32_16x16x32_fp8_bf8 a[0:3], v[2:3], v[4:5], a[0:3]172 173v_mfma_f32_16x16x32_fp8_fp8 v[0:3], v[2:3], v[4:5], v[0:3]174v_mfma_f32_16x16x32_fp8_fp8 a[0:3], v[2:3], v[4:5], a[0:3]175 176v_mfma_f32_32x32x16_bf8_bf8 v[0:15], v[2:3], v[4:5], v[0:15]177v_mfma_f32_32x32x16_fp8_bf8 v[0:15], v[2:3], v[4:5], v[0:15]178v_mfma_f32_32x32x16_bf8_fp8 v[0:15], v[2:3], v[4:5], v[0:15]179v_mfma_f32_32x32x16_fp8_fp8 v[0:15], v[2:3], v[4:5], v[0:15]180 181v_smfmac_f32_16x16x64_bf8_bf8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1182v_smfmac_f32_16x16x64_bf8_fp8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1183v_smfmac_f32_16x16x64_fp8_bf8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1184v_smfmac_f32_16x16x64_fp8_fp8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1185 186v_smfmac_f32_32x32x32_bf8_bf8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1187v_smfmac_f32_32x32x32_bf8_fp8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1188v_smfmac_f32_32x32x32_fp8_bf8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1189v_smfmac_f32_32x32x32_fp8_fp8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1190 191# CHECK:     [0]    [1]    [2]    [3]    [4]    [5]    [6]    Instructions:192# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x32_f16 a[0:3], a[0:3], a[0:3], a[0:3] blgp:1193# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x32_f16 a[0:3], v[0:3], v[0:3], a[4:7]194# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x16_f16 v[0:15], v[0:3], v[0:3], v[0:15]195# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x16_f16 a[0:15], a[0:3], a[0:3], a[0:15] blgp:2196# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x16_bf16 v[0:15], v[0:3], v[0:3], v[0:15]197# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x16_bf16 a[0:15], a[0:3], a[0:3], a[0:15] blgp:2198# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_i32_16x16x64_i8 a[0:3], a[0:3], a[0:3], a[0:3] blgp:1199# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_i32_16x16x64_i8 a[0:3], v[0:3], v[0:3], a[4:7]200# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_i32_32x32x32_i8 v[0:15], v[0:3], v[0:3], v[0:15]201# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_i32_32x32x32_i8 a[0:15], a[0:3], a[0:3], a[0:15] blgp:2202# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x32_bf16 a[0:3], a[0:3], a[0:3], a[0:3] blgp:1203# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x32_bf16 a[0:3], v[0:3], v[0:3], a[4:7]204# CHECK-NEXT: -      -      -      -     1.00    -      -     v_mfma_ld_scale_b32 v0, v0205# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3]206# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3] blgp:1207# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] blgp:2208# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] blgp:3209# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:7], v[0:3] blgp:4210# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3] cbsz:1211# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:2212# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:3213# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:7], v[4:11], v[0:3] cbsz:4214# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:11], v[0:3] cbsz:2 blgp:1215# CHECK-NEXT: -      -      -      -      -      -     4.00  v_mfma_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:9], v[0:3] cbsz:1 blgp:2216# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15]217# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] blgp:1218# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:9], v[0:15] blgp:2219# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:9], v[0:15] blgp:3220# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:7], v[0:15] blgp:4221# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] cbsz:1222# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:2223# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:3224# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:7], v[4:11], v[0:15] cbsz:4225# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15] cbsz:2226# CHECK-NEXT: -      -      -      -      -      -     8.00  v_mfma_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15] blgp:1227# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3], v5, v5 op_sel_hi:[0,0,0]228# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:11], v[4:11], v[0:3], v5, v5 op_sel_hi:[0,0,0] blgp:1229# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_scale_f32_16x16x128_f8f6f4 v[0:3], v[4:9], v[4:9], v[0:3], v5, v5 op_sel_hi:[0,0,0] cbsz:2 blgp:2230# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:11], v[4:11], v[0:15], v5, v5 op_sel_hi:[0,0,0]231# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:11], v[0:15], v5, v5 op_sel_hi:[0,0,0] cbsz:2 blgp:1232# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_scale_f32_32x32x64_f8f6f4 v[0:15], v[4:9], v[4:9], v[0:15], v5, v5 op_sel_hi:[0,0,0] cbsz:2 blgp:2233# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_f16 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1234# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_f16 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1235# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_bf16 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1236# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_bf16 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1237# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_i32_16x16x128_i8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1238# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_i32_32x32x64_i8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1239# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x128_bf8_bf8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1240# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x128_bf8_fp8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1241# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x128_fp8_bf8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1242# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x128_fp8_fp8 v[10:13], a[2:5], v[4:11], v3 cbsz:3 abid:1243# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x64_bf8_bf8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1244# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x64_bf8_fp8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1245# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x64_fp8_bf8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1246# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x64_fp8_fp8 v[10:25], a[2:5], v[4:11], v3 cbsz:3 abid:1247# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_f32 a[0:3], v0, v1, a[2:5]248# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_f32 v[0:3], v0, v1, v[2:5]249# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x2_f32 a[0:15], v0, v1, a[18:33]250# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x2_f32 v[0:15], v0, v1, v[18:33]251# CHECK-NEXT: -      -      -      -     4.00    -      -     v_mfma_f64_4x4x4_4b_f64 a[0:1], v[0:1], a[2:3], a[2:3]252# CHECK-NEXT: -      -      -      -     4.00    -      -     v_mfma_f64_4x4x4_4b_f64 v[0:1], v[0:1], v[2:3], v[2:3]253# CHECK-NEXT: -      -      -      -     16.00   -      -     v_mfma_f64_16x16x4_f64 a[0:7], v[0:1], v[2:3], a[0:7]254# CHECK-NEXT: -      -      -      -     16.00   -      -     v_mfma_f64_16x16x4_f64 v[0:7], v[0:1], v[2:3], v[0:7]255# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x16_f16 v[0:3], v[4:5], v[6:7], v[0:3]256# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x16_f16 a[0:3], v[4:5], v[6:7], a[0:3]257# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x8_f16 v[0:15], v[4:5], v[6:7], v[0:15]258# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x8_f16 a[0:15], v[4:5], v[6:7], a[0:15]259# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x16_bf16 v[0:3], v[4:5], v[6:7], v[0:3]260# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x16_bf16 a[0:3], v[4:5], v[6:7], a[0:3]261# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x8_bf16 v[0:15], v[4:5], v[6:7], v[0:15]262# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x8_bf16 a[0:15], v[4:5], v[6:7], a[0:15]263# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_i32_16x16x32_i8 v[0:3], v[4:5], v[6:7], v[0:3]264# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_i32_16x16x32_i8 a[0:3], v[4:5], v[6:7], a[0:3]265# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_i32_32x32x16_i8 v[0:15], v[2:3], v[4:5], v[0:15]266# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_i32_32x32x16_i8 a[0:15], v[2:3], v[4:5], a[0:15]267# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x4_16b_f16 v[0:3], v[0:1], v[2:3], v[2:5]268# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x4_16b_f16 a[0:3], v[0:1], v[2:3], a[2:5]269# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_4b_f16 v[0:15], v[2:3], v[4:5], v[18:33]270# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_4b_f16 a[0:15], v[2:3], v[4:5], a[18:33]271# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x4_2b_f16 v[0:31], v[0:1], v[2:3], v[34:65]272# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x4_2b_f16 a[0:31], v[0:1], v[2:3], a[34:65]273# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x4_16b_bf16 v[0:3], v[0:1], v[2:3], v[2:5]274# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x4_16b_bf16 a[0:3], v[0:1], v[2:3], a[2:5]275# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_4b_bf16 v[0:15], v[2:3], v[4:5], v[18:33]276# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_4b_bf16 a[0:15], v[2:3], v[4:5], a[18:33]277# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x4_2b_bf16 v[0:31], v[0:1], v[2:3], v[34:65]278# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x4_2b_bf16 a[0:31], v[0:1], v[2:3], a[34:65]279# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x1_16b_f32 v[0:3], v0, v1, v[2:5]280# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_f32_4x4x1_16b_f32 a[0:3], v0, v1, a[2:5]281# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x1_4b_f32 v[0:15], v0, v1, v[18:33]282# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x1_4b_f32 a[0:15], v0, v1, a[18:33]283# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_f32 v[0:3], v0, v1, v[2:5]284# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_16x16x4_f32 a[0:3], v0, v1, a[2:5]285# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x1_2b_f32 v[0:31], v0, v1, v[34:65] blgp:7286# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x1_2b_f32 a[0:31], v0, v1, a[34:65] blgp:7287# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x2_f32 v[0:15], v0, v1, v[18:33]288# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_f32_32x32x2_f32 a[0:15], v0, v1, a[18:33]289# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_i32_4x4x4_16b_i8 v[0:3], v0, v1, v[2:5]290# CHECK-NEXT: -      -      -      -      -      -     2.00   v_mfma_i32_4x4x4_16b_i8 a[0:3], v0, v1, a[2:5]291# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_i32_16x16x4_4b_i8 v[0:15], v0, v1, v[18:33]292# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_i32_16x16x4_4b_i8 a[0:15], v0, v1, a[18:33]293# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_i32_32x32x4_2b_i8 v[0:31], v0, v1, v[34:65]294# CHECK-NEXT: -      -      -      -      -      -     16.00  v_mfma_i32_32x32x4_2b_i8 a[0:31], v0, v1, a[34:65]295# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x32_f16 v[10:13], a[2:3], v[4:7], v0 cbsz:3 abid:1296# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x32_f16 a[10:13], v[2:3], a[4:7], v1297# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x16_f16 v[10:25], a[2:3], v[4:7], v2 cbsz:3 abid:1298# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x16_f16 a[10:25], v[2:3], a[4:7], v3299# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x32_bf16 v[10:13], a[2:3], v[4:7], v4 cbsz:3 abid:1300# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x32_bf16 a[10:13], v[2:3], a[4:7], v5301# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_i32_16x16x64_i8 v[10:13], a[2:3], v[4:7], v8 cbsz:3 abid:1302# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_i32_16x16x64_i8 a[10:13], v[2:3], a[4:7], v9303# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_i32_32x32x32_i8 v[10:25], a[2:3], v[4:7], v10 cbsz:3 abid:1304# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_i32_32x32x32_i8 a[10:25], v[2:3], a[4:7], v11305# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_bf8_bf8 v[0:3], v[2:3], v[4:5], v[0:3]306# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_bf8_bf8 a[0:3], v[2:3], v[4:5], a[0:3]307# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_bf8_fp8 v[0:3], v[2:3], v[4:5], v[0:3]308# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_bf8_fp8 a[0:3], v[2:3], v[4:5], a[0:3]309# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_fp8_bf8 v[0:3], v[2:3], v[4:5], v[0:3]310# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_fp8_bf8 a[0:3], v[2:3], v[4:5], a[0:3]311# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_fp8_fp8 v[0:3], v[2:3], v[4:5], v[0:3]312# CHECK-NEXT: -      -      -      -      -      -     4.00   v_mfma_f32_16x16x32_fp8_fp8 a[0:3], v[2:3], v[4:5], a[0:3]313# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x16_bf8_bf8 v[0:15], v[2:3], v[4:5], v[0:15]314# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x16_fp8_bf8 v[0:15], v[2:3], v[4:5], v[0:15]315# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x16_bf8_fp8 v[0:15], v[2:3], v[4:5], v[0:15]316# CHECK-NEXT: -      -      -      -      -      -     8.00   v_mfma_f32_32x32x16_fp8_fp8 v[0:15], v[2:3], v[4:5], v[0:15]317# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_bf8_bf8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1318# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_bf8_fp8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1319# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_fp8_bf8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1320# CHECK-NEXT: -      -      -      -      -      -     4.00   v_smfmac_f32_16x16x64_fp8_fp8 v[0:3], a[2:3], v[4:7], v1 cbsz:3 abid:1321# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_bf8_bf8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1322# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_bf8_fp8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1323# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_fp8_bf8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1324# CHECK-NEXT: -      -      -      -      -      -     8.00   v_smfmac_f32_32x32x32_fp8_fp8 v[0:15], v[2:3], v[4:7], v1 cbsz:3 abid:1325