Lines Matching refs:DL

5 … llc -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-DL %s
6 …lc -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
7 …lc -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
100 ; GFX9-DL-LABEL: idot4_acc32:
101 ; GFX9-DL: ; %bb.0: ; %entry
102 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
103 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
104 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
105 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
106 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
107 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
108 ; GFX9-DL-NEXT: s_load_dword s0, s[2:3], 0x0
109 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
110 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
111 ; GFX9-DL-NEXT: v_dot4_i32_i8 v1, v1, v2, s0
112 ; GFX9-DL-NEXT: global_store_dword v0, v1, s[2:3]
113 ; GFX9-DL-NEXT: s_endpgm
115 ; GFX10-DL-LABEL: idot4_acc32:
116 ; GFX10-DL: ; %bb.0: ; %entry
117 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
118 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
119 ; GFX10-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
120 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
121 ; GFX10-DL-NEXT: s_clause 0x1
122 ; GFX10-DL-NEXT: global_load_dword v1, v0, s[4:5]
123 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[6:7]
124 ; GFX10-DL-NEXT: v_mov_b32_e32 v0, 0
125 ; GFX10-DL-NEXT: s_load_dword s2, s[0:1], 0x0
126 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
127 ; GFX10-DL-NEXT: v_dot4_i32_i8 v1, v1, v2, s2
128 ; GFX10-DL-NEXT: global_store_dword v0, v1, s[0:1]
129 ; GFX10-DL-NEXT: s_endpgm
292 ; GFX9-DL-LABEL: idot4_acc16:
293 ; GFX9-DL: ; %bb.0: ; %entry
294 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
295 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
296 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
297 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
298 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
299 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
300 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
301 ; GFX9-DL-NEXT: global_load_ushort v3, v0, s[2:3]
302 ; GFX9-DL-NEXT: s_waitcnt vmcnt(2)
303 ; GFX9-DL-NEXT: v_bfe_i32 v6, v1, 0, 8
304 ; GFX9-DL-NEXT: s_waitcnt vmcnt(1)
305 ; GFX9-DL-NEXT: v_bfe_i32 v7, v2, 0, 8
306 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v8, 8, v1
307 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v9, 8, v2
308 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v4, 16, v1
309 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v5, 16, v2
310 ; GFX9-DL-NEXT: v_bfe_i32 v8, v8, 0, 8
311 ; GFX9-DL-NEXT: v_bfe_i32 v9, v9, 0, 8
312 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
313 ; GFX9-DL-NEXT: v_mad_legacy_u16 v3, v6, v7, v3
314 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v1, 24, v1
315 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v2, 24, v2
316 ; GFX9-DL-NEXT: v_bfe_i32 v4, v4, 0, 8
317 ; GFX9-DL-NEXT: v_bfe_i32 v5, v5, 0, 8
318 ; GFX9-DL-NEXT: v_mad_legacy_u16 v3, v8, v9, v3
319 ; GFX9-DL-NEXT: v_bfe_i32 v1, v1, 0, 8
320 ; GFX9-DL-NEXT: v_bfe_i32 v2, v2, 0, 8
321 ; GFX9-DL-NEXT: v_mad_legacy_u16 v3, v4, v5, v3
322 ; GFX9-DL-NEXT: v_mad_legacy_u16 v1, v1, v2, v3
323 ; GFX9-DL-NEXT: global_store_short v0, v1, s[2:3]
324 ; GFX9-DL-NEXT: s_endpgm
326 ; GFX10-DL-LABEL: idot4_acc16:
327 ; GFX10-DL: ; %bb.0: ; %entry
328 ; GFX10-DL-NEXT: s_clause 0x1
329 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
330 ; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
331 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
332 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
333 ; GFX10-DL-NEXT: s_clause 0x1
334 ; GFX10-DL-NEXT: global_load_dword v1, v0, s[4:5]
335 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[6:7]
336 ; GFX10-DL-NEXT: v_mov_b32_e32 v0, 0
337 ; GFX10-DL-NEXT: global_load_ushort v3, v0, s[2:3]
338 ; GFX10-DL-NEXT: s_waitcnt vmcnt(2)
339 ; GFX10-DL-NEXT: v_bfe_i32 v4, v1, 0, 8
340 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v5, 8, v1
341 ; GFX10-DL-NEXT: s_waitcnt vmcnt(1)
342 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v6, 8, v2
343 ; GFX10-DL-NEXT: v_bfe_i32 v7, v2, 0, 8
344 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v8, 16, v1
345 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v9, 16, v2
346 ; GFX10-DL-NEXT: v_bfe_i32 v5, v5, 0, 8
347 ; GFX10-DL-NEXT: v_bfe_i32 v6, v6, 0, 8
348 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
349 ; GFX10-DL-NEXT: v_mad_u16 v3, v4, v7, v3
350 ; GFX10-DL-NEXT: v_bfe_i32 v4, v8, 0, 8
351 ; GFX10-DL-NEXT: v_bfe_i32 v7, v9, 0, 8
352 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v1, 24, v1
353 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v2, 24, v2
354 ; GFX10-DL-NEXT: v_mad_u16 v3, v5, v6, v3
355 ; GFX10-DL-NEXT: v_bfe_i32 v1, v1, 0, 8
356 ; GFX10-DL-NEXT: v_bfe_i32 v2, v2, 0, 8
357 ; GFX10-DL-NEXT: v_mad_u16 v3, v4, v7, v3
358 ; GFX10-DL-NEXT: v_mad_u16 v1, v1, v2, v3
359 ; GFX10-DL-NEXT: global_store_short v0, v1, s[2:3]
360 ; GFX10-DL-NEXT: s_endpgm
497 ; GFX9-DL-LABEL: idot4_acc8:
498 ; GFX9-DL: ; %bb.0: ; %entry
499 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
500 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
501 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
502 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
503 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
504 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
505 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
506 ; GFX9-DL-NEXT: global_load_ubyte v3, v0, s[2:3]
507 ; GFX9-DL-NEXT: s_waitcnt vmcnt(2)
508 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v4, 16, v1
509 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v6, 8, v1
510 ; GFX9-DL-NEXT: s_waitcnt vmcnt(1)
511 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v7, 8, v2
512 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v8, 24, v1
513 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
514 ; GFX9-DL-NEXT: v_mad_legacy_u16 v1, v1, v2, v3
515 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v5, 16, v2
516 ; GFX9-DL-NEXT: v_mad_legacy_u16 v1, v6, v7, v1
517 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v9, 24, v2
518 ; GFX9-DL-NEXT: v_mad_legacy_u16 v1, v4, v5, v1
519 ; GFX9-DL-NEXT: v_mad_legacy_u16 v1, v8, v9, v1
520 ; GFX9-DL-NEXT: global_store_byte v0, v1, s[2:3]
521 ; GFX9-DL-NEXT: s_endpgm
523 ; GFX10-DL-LABEL: idot4_acc8:
524 ; GFX10-DL: ; %bb.0: ; %entry
525 ; GFX10-DL-NEXT: s_clause 0x1
526 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
527 ; GFX10-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
528 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
529 ; GFX10-DL-NEXT: v_mov_b32_e32 v1, 0
530 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
531 ; GFX10-DL-NEXT: s_clause 0x1
532 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[4:5]
533 ; GFX10-DL-NEXT: global_load_dword v3, v0, s[6:7]
534 ; GFX10-DL-NEXT: global_load_ubyte v4, v1, s[2:3]
535 ; GFX10-DL-NEXT: s_waitcnt vmcnt(2)
536 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v0, 8, v2
537 ; GFX10-DL-NEXT: s_waitcnt vmcnt(1)
538 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v5, 8, v3
539 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
540 ; GFX10-DL-NEXT: v_mad_u16 v4, v2, v3, v4
541 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v6, 16, v2
542 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v7, 16, v3
543 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v2, 24, v2
544 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v3, 24, v3
545 ; GFX10-DL-NEXT: v_mad_u16 v0, v0, v5, v4
546 ; GFX10-DL-NEXT: v_mad_u16 v0, v6, v7, v0
547 ; GFX10-DL-NEXT: v_mad_u16 v0, v2, v3, v0
548 ; GFX10-DL-NEXT: global_store_byte v1, v0, s[2:3]
549 ; GFX10-DL-NEXT: s_endpgm
681 ; GFX9-DL-LABEL: idot4_multiuse_mul1:
682 ; GFX9-DL: ; %bb.0: ; %entry
683 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
684 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
685 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
686 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
687 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
688 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
689 ; GFX9-DL-NEXT: s_load_dword s0, s[2:3], 0x0
690 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
691 ; GFX9-DL-NEXT: s_waitcnt vmcnt(1)
692 ; GFX9-DL-NEXT: v_bfe_i32 v3, v1, 0, 8
693 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
694 ; GFX9-DL-NEXT: v_bfe_i32 v4, v2, 0, 8
695 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
696 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v6, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
697 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
698 ; GFX9-DL-NEXT: v_mul_i32_i24_e32 v2, v3, v4
699 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
700 ; GFX9-DL-NEXT: v_mad_i32_i24 v3, v3, v4, s0
701 ; GFX9-DL-NEXT: v_add3_u32 v2, v5, v3, v2
702 ; GFX9-DL-NEXT: v_add3_u32 v1, v2, v6, v1
703 ; GFX9-DL-NEXT: global_store_dword v0, v1, s[2:3]
704 ; GFX9-DL-NEXT: s_endpgm
706 ; GFX10-DL-LABEL: idot4_multiuse_mul1:
707 ; GFX10-DL: ; %bb.0: ; %entry
708 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
709 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
710 ; GFX10-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
711 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
712 ; GFX10-DL-NEXT: s_clause 0x1
713 ; GFX10-DL-NEXT: global_load_dword v1, v0, s[4:5]
714 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[6:7]
715 ; GFX10-DL-NEXT: s_load_dword s2, s[0:1], 0x0
716 ; GFX10-DL-NEXT: s_waitcnt vmcnt(1)
717 ; GFX10-DL-NEXT: v_bfe_i32 v0, v1, 0, 8
718 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
719 ; GFX10-DL-NEXT: v_bfe_i32 v3, v2, 0, 8
720 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v4, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
721 ; GFX10-DL-NEXT: v_mul_i32_i24_e32 v5, v0, v3
722 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
723 ; GFX10-DL-NEXT: v_mad_i32_i24 v0, v0, v3, s2
724 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v3, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
725 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
726 ; GFX10-DL-NEXT: v_mov_b32_e32 v2, 0
727 ; GFX10-DL-NEXT: v_add3_u32 v0, v4, v0, v5
728 ; GFX10-DL-NEXT: v_add3_u32 v0, v0, v3, v1
729 ; GFX10-DL-NEXT: global_store_dword v2, v0, s[0:1]
730 ; GFX10-DL-NEXT: s_endpgm
872 ; GFX9-DL-LABEL: idot4_acc32_vecMul:
873 ; GFX9-DL: ; %bb.0: ; %entry
874 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
875 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
876 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
877 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
878 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
879 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
880 ; GFX9-DL-NEXT: s_load_dword s0, s[2:3], 0x0
881 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
882 ; GFX9-DL-NEXT: s_waitcnt vmcnt(1)
883 ; GFX9-DL-NEXT: v_lshrrev_b16_e32 v3, 8, v1
884 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
885 ; GFX9-DL-NEXT: v_lshrrev_b16_e32 v4, 8, v2
886 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
887 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v3, sext(v3), sext(v4) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
888 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v6, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
889 ; GFX9-DL-NEXT: v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD sr…
890 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
891 ; GFX9-DL-NEXT: v_add3_u32 v2, v5, s0, v3
892 ; GFX9-DL-NEXT: v_add3_u32 v1, v2, v6, v1
893 ; GFX9-DL-NEXT: global_store_dword v0, v1, s[2:3]
894 ; GFX9-DL-NEXT: s_endpgm
896 ; GFX10-DL-LABEL: idot4_acc32_vecMul:
897 ; GFX10-DL: ; %bb.0: ; %entry
898 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
899 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
900 ; GFX10-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
901 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
902 ; GFX10-DL-NEXT: s_clause 0x1
903 ; GFX10-DL-NEXT: global_load_dword v1, v0, s[4:5]
904 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[6:7]
905 ; GFX10-DL-NEXT: s_load_dword s2, s[0:1], 0x0
906 ; GFX10-DL-NEXT: s_waitcnt vmcnt(1)
907 ; GFX10-DL-NEXT: v_lshrrev_b16 v0, 8, v1
908 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
909 ; GFX10-DL-NEXT: v_lshrrev_b16 v3, 8, v2
910 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v4, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
911 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v0, sext(v0), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD s…
912 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v3, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
913 ; GFX10-DL-NEXT: v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD s…
914 ; GFX10-DL-NEXT: v_mov_b32_e32 v2, 0
915 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
916 ; GFX10-DL-NEXT: v_add3_u32 v0, v4, s2, v0
917 ; GFX10-DL-NEXT: v_add3_u32 v0, v0, v3, v1
918 ; GFX10-DL-NEXT: global_store_dword v2, v0, s[0:1]
919 ; GFX10-DL-NEXT: s_endpgm
1071 ; GFX9-DL-LABEL: idot4_acc16_vecMul:
1072 ; GFX9-DL: ; %bb.0: ; %entry
1073 ; GFX9-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
1074 ; GFX9-DL-NEXT: s_load_dwordx2 s[2:3], s[0:1], 0x34
1075 ; GFX9-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
1076 ; GFX9-DL-NEXT: v_mov_b32_e32 v4, 0xffff
1077 ; GFX9-DL-NEXT: s_waitcnt lgkmcnt(0)
1078 ; GFX9-DL-NEXT: global_load_dword v1, v0, s[4:5]
1079 ; GFX9-DL-NEXT: global_load_dword v2, v0, s[6:7]
1080 ; GFX9-DL-NEXT: v_mov_b32_e32 v0, 0
1081 ; GFX9-DL-NEXT: global_load_ushort v3, v0, s[2:3]
1082 ; GFX9-DL-NEXT: s_waitcnt vmcnt(2)
1083 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v5, 16, v1
1084 ; GFX9-DL-NEXT: s_waitcnt vmcnt(1)
1085 ; GFX9-DL-NEXT: v_lshrrev_b32_e32 v6, 16, v2
1086 ; GFX9-DL-NEXT: v_ashrrev_i16_e32 v7, 8, v1
1087 ; GFX9-DL-NEXT: v_ashrrev_i16_e32 v8, 8, v2
1088 ; GFX9-DL-NEXT: v_and_b32_sdwa v2, v4, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWO…
1089 ; GFX9-DL-NEXT: v_and_b32_sdwa v1, v4, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWO…
1090 ; GFX9-DL-NEXT: v_lshl_or_b32 v2, v8, 16, v2
1091 ; GFX9-DL-NEXT: v_lshl_or_b32 v1, v7, 16, v1
1092 ; GFX9-DL-NEXT: v_ashrrev_i16_e32 v9, 8, v5
1093 ; GFX9-DL-NEXT: v_ashrrev_i16_e32 v10, 8, v6
1094 ; GFX9-DL-NEXT: v_and_b32_sdwa v6, v4, sext(v6) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWO…
1095 ; GFX9-DL-NEXT: v_and_b32_sdwa v4, v4, sext(v5) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWO…
1096 ; GFX9-DL-NEXT: v_pk_mul_lo_u16 v1, v1, v2
1097 ; GFX9-DL-NEXT: v_lshl_or_b32 v5, v10, 16, v6
1098 ; GFX9-DL-NEXT: v_lshl_or_b32 v4, v9, 16, v4
1099 ; GFX9-DL-NEXT: s_waitcnt vmcnt(0)
1100 ; GFX9-DL-NEXT: v_add_u16_e32 v3, v1, v3
1101 ; GFX9-DL-NEXT: v_pk_mul_lo_u16 v2, v4, v5
1102 ; GFX9-DL-NEXT: v_add_u16_sdwa v1, v3, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src…
1103 ; GFX9-DL-NEXT: v_add_u16_e32 v1, v1, v2
1104 ; GFX9-DL-NEXT: v_add_u16_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src…
1105 ; GFX9-DL-NEXT: global_store_short v0, v1, s[2:3]
1106 ; GFX9-DL-NEXT: s_endpgm
1108 ; GFX10-DL-LABEL: idot4_acc16_vecMul:
1109 ; GFX10-DL: ; %bb.0: ; %entry
1110 ; GFX10-DL-NEXT: s_load_dwordx4 s[4:7], s[0:1], 0x24
1111 ; GFX10-DL-NEXT: v_lshlrev_b32_e32 v0, 2, v0
1112 ; GFX10-DL-NEXT: s_load_dwordx2 s[0:1], s[0:1], 0x34
1113 ; GFX10-DL-NEXT: v_mov_b32_e32 v4, 0xffff
1114 ; GFX10-DL-NEXT: s_waitcnt lgkmcnt(0)
1115 ; GFX10-DL-NEXT: s_clause 0x1
1116 ; GFX10-DL-NEXT: global_load_dword v1, v0, s[4:5]
1117 ; GFX10-DL-NEXT: global_load_dword v2, v0, s[6:7]
1118 ; GFX10-DL-NEXT: v_mov_b32_e32 v0, 0
1119 ; GFX10-DL-NEXT: global_load_ushort v3, v0, s[0:1]
1120 ; GFX10-DL-NEXT: s_waitcnt vmcnt(2)
1121 ; GFX10-DL-NEXT: v_ashrrev_i16 v5, 8, v1
1122 ; GFX10-DL-NEXT: s_waitcnt vmcnt(1)
1123 ; GFX10-DL-NEXT: v_ashrrev_i16 v6, 8, v2
1124 ; GFX10-DL-NEXT: v_and_b32_sdwa v7, v4, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DW…
1125 ; GFX10-DL-NEXT: v_and_b32_sdwa v8, v4, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DW…
1126 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v1, 16, v1
1127 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v2, 16, v2
1128 ; GFX10-DL-NEXT: v_lshl_or_b32 v6, v6, 16, v7
1129 ; GFX10-DL-NEXT: v_lshl_or_b32 v5, v5, 16, v8
1130 ; GFX10-DL-NEXT: v_ashrrev_i16 v7, 8, v1
1131 ; GFX10-DL-NEXT: v_ashrrev_i16 v8, 8, v2
1132 ; GFX10-DL-NEXT: v_and_b32_sdwa v2, v4, sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DW…
1133 ; GFX10-DL-NEXT: v_and_b32_sdwa v1, v4, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DW…
1134 ; GFX10-DL-NEXT: v_pk_mul_lo_u16 v4, v5, v6
1135 ; GFX10-DL-NEXT: v_lshl_or_b32 v2, v8, 16, v2
1136 ; GFX10-DL-NEXT: v_lshl_or_b32 v1, v7, 16, v1
1137 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v5, 16, v4
1138 ; GFX10-DL-NEXT: s_waitcnt vmcnt(0)
1139 ; GFX10-DL-NEXT: v_add_nc_u16 v3, v4, v3
1140 ; GFX10-DL-NEXT: v_pk_mul_lo_u16 v1, v1, v2
1141 ; GFX10-DL-NEXT: v_add_nc_u16 v2, v3, v5
1142 ; GFX10-DL-NEXT: v_lshrrev_b32_e32 v3, 16, v1
1143 ; GFX10-DL-NEXT: v_add_nc_u16 v1, v2, v1
1144 ; GFX10-DL-NEXT: v_add_nc_u16 v1, v1, v3
1145 ; GFX10-DL-NEXT: global_store_short v0, v1, s[0:1]
1146 ; GFX10-DL-NEXT: s_endpgm