143 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx900 --denormal-fp-math=preserve-sign < %s | FileCheck -check-prefix=GFX9-FAST-DENORM %s3; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 --denormal-fp-math=preserve-sign < %s | FileCheck -check-prefix=GFX10-FAST-DENORM %s4 5; fold (fadd fast (fpext (fmul fast x, y)), z) -> (fma (fpext x), (fpext y), z)6; fold (fadd fast x, (fpext (fmul fast y, z))) -> (fma (fpext y), (fpext z), x)7 8define amdgpu_vs float @test_f16_f32_add_ext_mul(half inreg %x, half inreg %y, float inreg %z) {9; GFX9-FAST-DENORM-LABEL: test_f16_f32_add_ext_mul:10; GFX9-FAST-DENORM: ; %bb.0: ; %.entry11; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s112; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s213; GFX9-FAST-DENORM-NEXT: v_mad_mix_f32 v0, s0, v0, v1 op_sel_hi:[1,1,0]14; GFX9-FAST-DENORM-NEXT: ; return to shader part epilog15;16; GFX10-FAST-DENORM-LABEL: test_f16_f32_add_ext_mul:17; GFX10-FAST-DENORM: ; %bb.0: ; %.entry18; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s219; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v0, s0, s1, v0 op_sel_hi:[1,1,0]20; GFX10-FAST-DENORM-NEXT: ; return to shader part epilog21.entry:22 %a = fmul fast half %x, %y23 %b = fpext half %a to float24 %c = fadd fast float %b, %z25 ret float %c26}27 28define amdgpu_vs float @test_f16_f32_add_ext_mul_rhs(half inreg %x, half inreg %y, float inreg %z) {29; GFX9-FAST-DENORM-LABEL: test_f16_f32_add_ext_mul_rhs:30; GFX9-FAST-DENORM: ; %bb.0: ; %.entry31; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s132; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s233; GFX9-FAST-DENORM-NEXT: v_mad_mix_f32 v0, s0, v0, v1 op_sel_hi:[1,1,0]34; GFX9-FAST-DENORM-NEXT: ; return to shader part epilog35;36; GFX10-FAST-DENORM-LABEL: test_f16_f32_add_ext_mul_rhs:37; GFX10-FAST-DENORM: ; %bb.0: ; %.entry38; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s239; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v0, s0, s1, v0 op_sel_hi:[1,1,0]40; GFX10-FAST-DENORM-NEXT: ; return to shader part epilog41.entry:42 %a = fmul fast half %x, %y43 %b = fpext half %a to float44 %c = fadd fast float %z, %b45 ret float %c46}47 48define amdgpu_vs <5 x float> @test_5xf16_5xf32_add_ext_mul(<5 x half> inreg %x, <5 x half> inreg %y, <5 x float> inreg %z) {49; GFX9-FAST-DENORM-LABEL: test_5xf16_5xf32_add_ext_mul:50; GFX9-FAST-DENORM: ; %bb.0: ; %.entry51; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s352; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s453; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v2, s554; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v0, s0, v055; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v1, s1, v156; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v2, s2, v257; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v3, v058; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_sdwa v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_159; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v5, v160; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_sdwa v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_161; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v7, v262; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v0, s6, v363; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v1, s7, v464; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v2, s8, v565; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v3, s9, v666; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v4, s10, v767; GFX9-FAST-DENORM-NEXT: ; return to shader part epilog68;69; GFX10-FAST-DENORM-LABEL: test_5xf16_5xf32_add_ext_mul:70; GFX10-FAST-DENORM: ; %bb.0: ; %.entry71; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s672; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s773; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v2, s874; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v3, s975; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v4, s1076; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s11, s0, 1677; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s12, s1, 1678; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s13, s3, 1679; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s14, s4, 1680; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v0, s0, s3, v0 op_sel_hi:[1,1,0]81; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v1, s11, s13, v1 op_sel_hi:[1,1,0]82; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v2, s1, s4, v2 op_sel_hi:[1,1,0]83; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v3, s12, s14, v3 op_sel_hi:[1,1,0]84; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v4, s2, s5, v4 op_sel_hi:[1,1,0]85; GFX10-FAST-DENORM-NEXT: ; return to shader part epilog86.entry:87 %a = fmul fast <5 x half> %x, %y88 %b = fpext <5 x half> %a to <5 x float>89 %c = fadd fast <5 x float> %b, %z90 ret <5 x float> %c91}92 93define amdgpu_vs <6 x float> @test_6xf16_6xf32_add_ext_mul_rhs(<6 x half> inreg %x, <6 x half> inreg %y, <6 x float> inreg %z) {94; GFX9-FAST-DENORM-LABEL: test_6xf16_6xf32_add_ext_mul_rhs:95; GFX9-FAST-DENORM: ; %bb.0: ; %.entry96; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s397; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s498; GFX9-FAST-DENORM-NEXT: v_mov_b32_e32 v2, s599; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v0, s0, v0100; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v1, s1, v1101; GFX9-FAST-DENORM-NEXT: v_pk_mul_f16 v2, s2, v2102; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v3, v0103; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_sdwa v4, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1104; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v5, v1105; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_sdwa v6, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1106; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_e32 v7, v2107; GFX9-FAST-DENORM-NEXT: v_cvt_f32_f16_sdwa v8, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1108; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v0, s6, v3109; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v1, s7, v4110; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v2, s8, v5111; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v3, s9, v6112; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v4, s10, v7113; GFX9-FAST-DENORM-NEXT: v_add_f32_e32 v5, s11, v8114; GFX9-FAST-DENORM-NEXT: ; return to shader part epilog115;116; GFX10-FAST-DENORM-LABEL: test_6xf16_6xf32_add_ext_mul_rhs:117; GFX10-FAST-DENORM: ; %bb.0: ; %.entry118; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v0, s6119; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v1, s7120; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v2, s8121; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v3, s9122; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v4, s10123; GFX10-FAST-DENORM-NEXT: v_mov_b32_e32 v5, s11124; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s12, s0, 16125; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s13, s1, 16126; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s6, s2, 16127; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s14, s3, 16128; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s15, s4, 16129; GFX10-FAST-DENORM-NEXT: s_lshr_b32 s16, s5, 16130; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v0, s0, s3, v0 op_sel_hi:[1,1,0]131; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v1, s12, s14, v1 op_sel_hi:[1,1,0]132; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v2, s1, s4, v2 op_sel_hi:[1,1,0]133; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v3, s13, s15, v3 op_sel_hi:[1,1,0]134; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v4, s2, s5, v4 op_sel_hi:[1,1,0]135; GFX10-FAST-DENORM-NEXT: v_fma_mix_f32 v5, s6, s16, v5 op_sel_hi:[1,1,0]136; GFX10-FAST-DENORM-NEXT: ; return to shader part epilog137.entry:138 %a = fmul fast <6 x half> %x, %y139 %b = fpext <6 x half> %a to <6 x float>140 %c = fadd fast <6 x float> %z, %b141 ret <6 x float> %c142}143