ELF>@@HH|$8Ht$0T$,L$(LD$ DL$H$HD$H$HD$dH%(H$1HD$8H$HD$0H$HD$,H$HD$(H$HD$ H$HD$H$HD$H$HD$H$D$PD$TD$XD$\D$`D$dHL$HHT$@Ht$\H|$PtH$dH3%(uPHHt$PHt$hT$XT$pHL$\HL$tDD$dDD$|t$Ht$HL$H=HHt$t$HHH|$HHt$@T$ /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxsm_20_atomic_functions.hpp ~}~( z ~~~~}| { z~| }z}x|~~ t(~~ t  (}}{~~ t  ~~ ~z~yx(zx~~}~|z|wx y t t t t mros~t (  Jzz x  x  }i킀  ~}}}}y{x~w w  wu{bl w wk u s v n Uu t p~}p  tws u i  u p z{ ~|  wzzx~wu  qx zt{v{ l ptzdmzyxuxy xya zgpjzzy p z{|.version 6.2.target sm_30.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<27>;.reg .f64 %fd<15>;.reg .b64 %rd<49>;ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r4, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd17, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r5, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd18, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd1, %rd18;cvta.to.global.u64 %rd2, %rd17;cvta.to.global.u64 %rd47, %rd16;mov.u32 %r6, %ntid.x;mov.u32 %r7, %ctaid.x;mov.u32 %r8, %tid.x;mad.lo.s32 %r1, %r6, %r7, %r8;mov.u32 %r2, %ctaid.y;setp.ge.s32 %p1, %r1, %r3;@%p1 bra BB0_8;cvta.to.global.u64 %rd19, %rd14;mul.wide.s32 %rd20, %r2, 8;add.s64 %rd21, %rd19, %rd20;ld.global.u32 %r9, [%rd21];cvt.s64.s32 %rd4, %r9;mul.wide.s32 %rd22, %r9, 12;add.s64 %rd48, %rd47, %rd22;ld.global.u32 %r10, [%rd21+4];mul.wide.s32 %rd23, %r10, 12;add.s64 %rd7, %rd47, %rd23;add.s32 %r11, %r9, 2;mul.wide.s32 %rd24, %r11, 12;add.s64 %rd25, %rd47, %rd24;setp.gt.u64 %p2, %rd25, %rd7;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd8, %rd4, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd26, %rd47, %rd8;ld.global.u32 %r12, [%rd26+4];mad.lo.s32 %r13, %r12, %r5, %r1;mul.wide.s32 %rd27, %r13, 4;add.s64 %rd28, %rd2, %rd27;ld.global.u32 %r14, [%rd26+8];mad.lo.s32 %r15, %r14, %r3, %r1;mul.wide.s32 %rd29, %r15, 4;add.s64 %rd30, %rd1, %rd29;ld.global.u32 %r16, [%rd26+16];mad.lo.s32 %r17, %r16, %r5, %r1;mul.wide.s32 %rd31, %r17, 4;add.s64 %rd32, %rd2, %rd31;ld.global.u32 %r18, [%rd26+20];mad.lo.s32 %r19, %r18, %r3, %r1;mul.wide.s32 %rd33, %r19, 4;add.s64 %rd34, %rd1, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd26];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd28];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd26+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd47, %rd47, 24;add.s64 %rd35, %rd47, %rd8;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd7;@%p3 bra BB0_3;cvt.u32.u64 %r20, %rd4;mul.wide.s32 %rd37, %r20, 12;add.s64 %rd48, %rd47, %rd37;BB0_5:setp.eq.s64 %p4, %rd48, %rd7;@%p4 bra BB0_7;ld.global.u32 %r21, [%rd48+4];mad.lo.s32 %r22, %r21, %r5, %r1;mul.wide.s32 %rd38, %r22, 4;add.s64 %rd39, %rd2, %rd38;ld.global.u32 %r23, [%rd48+8];mad.lo.s32 %r24, %r23, %r3, %r1;mul.wide.s32 %rd40, %r24, 4;add.s64 %rd41, %rd1, %rd40;ld.global.f32 %f11, [%rd41];ld.global.f32 %f12, [%rd48];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd39];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:cvta.to.global.u64 %rd42, %rd15;mad.lo.s32 %r25, %r4, %r3, %r1;mul.wide.s32 %rd43, %r25, 4;add.s64 %rd44, %rd1, %rd43;ld.global.f32 %f16, [%rd44];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r26, %r2, %r3, %r1;mul.wide.s32 %rd45, %r26, 4;add.s64 %rd46, %rd42, %rd45;st.global.f32 [%rd46], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<57>;.reg .f64 %fd<15>;.reg .b64 %rd<66>;ld.param.u64 %rd13, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd14, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd15, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd16, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd17, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_18;cvta.to.global.u64 %rd20, %rd13;cvta.to.global.u64 %rd21, %rd16;cvta.to.global.u64 %rd64, %rd14;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r12, %r11, %r4, %r1;mul.wide.s32 %rd23, %r12, 4;add.s64 %rd24, %rd21, %rd23;mad.lo.s32 %r13, %r5, %r4, %r1;mul.wide.s32 %rd25, %r13, 4;add.s64 %rd26, %rd21, %rd25;ld.global.f32 %f1, [%rd26];ld.global.f32 %f7, [%rd24];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd27, %r11, 8;add.s64 %rd28, %rd20, %rd27;ld.global.u32 %r2, [%rd28];mul.wide.s32 %rd29, %r2, 12;add.s64 %rd65, %rd64, %rd29;ld.global.u32 %r14, [%rd28+4];mul.wide.s32 %rd30, %r14, 12;add.s64 %rd3, %rd64, %rd30;add.s32 %r15, %r2, 2;mul.wide.s32 %rd31, %r15, 12;add.s64 %rd32, %rd64, %rd31;setp.gt.u64 %p2, %rd32, %rd3;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd34, %rd64, %rd29;ld.global.u32 %r16, [%rd34+8];mad.lo.s32 %r21, %r16, %r4, %r1;cvta.to.global.u64 %rd35, %rd17;mul.wide.s32 %rd36, %r21, 4;add.s64 %rd37, %rd35, %rd36;ld.global.f32 %f8, [%rd37];ld.global.f32 %f9, [%rd34];mul.f32 %f10, %f9, %f8;ld.global.u32 %r22, [%rd34+4];mad.lo.s32 %r23, %r22, %r6, %r1;cvta.to.global.u64 %rd38, %rd15;mul.wide.s32 %rd39, %r23, 4;add.s64 %rd40, %rd38, %rd39;ld.global.f32 %f11, [%rd40];mul.f32 %f12, %f10, %f11;ld.global.u32 %r24, [%rd34+20];mad.lo.s32 %r25, %r24, %r4, %r1;mul.wide.s32 %rd41, %r25, 4;add.s64 %rd42, %rd35, %rd41;ld.global.f32 %f13, [%rd42];ld.global.f32 %f14, [%rd34+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd34+16];mad.lo.s32 %r26, %r3, %r6, %r1;mul.wide.s32 %rd43, %r26, 4;add.s64 %rd44, %rd38, %rd43;ld.global.f32 %f16, [%rd44];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r27, %r22, %r7, %r1;cvta.to.global.u64 %rd45, %rd19;mul.wide.s32 %rd46, %r27, 4;add.s64 %rd6, %rd45, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r28, %f19;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p4, %r29, %r30;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd6], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd6], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r35, %r3, %r7, %r1;mul.wide.s32 %rd48, %r35, 4;add.s64 %rd7, %rd45, %rd48;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r36, %f22;shr.s32 %r37, %r36, 12;and.b32 %r38, %r36, 4095;setp.le.s32 %p6, %r37, %r38;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd7], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd7], %f5;BB1_11:add.s64 %rd64, %rd64, 24;add.s64 %rd65, %rd64, %rd29;add.s64 %rd51, %rd65, 24;setp.le.u64 %p7, %rd51, %rd3;@%p7 bra BB1_3;BB1_12:setp.eq.s64 %p8, %rd65, %rd3;@%p8 bra BB1_17;ld.global.u32 %r39, [%rd65+8];mad.lo.s32 %r44, %r39, %r4, %r1;cvta.to.global.u64 %rd53, %rd17;mul.wide.s32 %rd54, %r44, 4;add.s64 %rd55, %rd53, %rd54;ld.global.f32 %f24, [%rd55];ld.global.f32 %f25, [%rd65];mul.f32 %f26, %f25, %f24;ld.global.u32 %r45, [%rd65+4];mad.lo.s32 %r46, %r45, %r6, %r1;cvta.to.global.u64 %rd56, %rd15;mul.wide.s32 %rd57, %r46, 4;add.s64 %rd58, %rd56, %rd57;ld.global.f32 %f27, [%rd58];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r47, %r45, %r7, %r1;cvta.to.global.u64 %rd59, %rd19;mul.wide.s32 %rd60, %r47, 4;add.s64 %rd12, %rd59, %rd60;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_15;bra.uni BB1_14;BB1_15:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r48, %f30;shr.s32 %r49, %r48, 12;and.b32 %r50, %r48, 4095;setp.le.s32 %p10, %r49, %r50;@%p10 bra BB1_17;atom.global.add.f32 %f31, [%rd12], 0f38800000;bra.uni BB1_17;BB1_14:atom.global.add.f32 %f29, [%rd12], %f6;BB1_17:cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;cvta.to.global.u64 %rd61, %rd18;add.s64 %rd63, %rd61, %rd23;st.global.f32 [%rd63], %f32;BB1_18:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   #### @((  !  1 !  ? X    @LL H @! 8! 0! (!   !   ! !?80 @88 0! (!   !   ! !?80P, c ?3s[UU@>>?~32Br"]@(,,,0@ ,#1@@B‚"@ #P@!0`@ C! @@\!0 G""A0p@ C@@ A@ @A@%BB" A@PA@0! B A@PA@0!@@@2"]@(,,@ #@@,Gp2B"P@(] @@ 0@@ C@@ C1@@GB"\1 aC @(]0@(#Pc!@@GB2" 0  (\%%(C2B"(T! A \U1 \A  A0 _LHBB"] @(CATA1 ]0@(\A1  @(BB22"(A1! A1 h\a bP723B"fe@f0\Q@@ @@ C\V@@܂@ "BB"\R C@@\s@ \ d(C`@@\fG0BB"ܲp C`@@A\p 3‚"!@ C܁@@ \dXFhX\)X1XB"qP\&eXܑ Ύ$1@H(]`(BB"@@^F,X!#a0Xa?8DzBB"#qA,u!@ `hGBB"!qXCq@@q Ύ$\q @@A,B²"qX!#a0Xa?8#qGBB"A,a C\GH\A1! A1 Q` G"CHOLHCaT@(](OLHGB03"cQThA \Cq@@ G‚BB"р@ Cr@@\r Cb`@@bp AG"\ @ C@@\ Xa(X$XwB"1! Ύ$ H@@D,!XB²"!#a0Xa?8#qA,BpB"1=`QȀ? HB "  a "PB  b( 0a0H"B" @PS@@\! \"BBpB"Q9\29a@H(GBB"܁PH  hBPܑ`H(ܳ\PH 2R" A #0P@ 8P@ ( ײB2"@ܑpHb0P\PH ( PW2Rr"(Pa0PB  C ܡ0q$ a DzBB"q&1 c;]@a(Bd8G2B2"\H 0(A P P@(](b ҂BB" ܡ 1  8aCHBb", B\r r( 0 \ ,(2B"((A @  a, ܑ( ײB"`@Q(cAߎ$A w"dQ ((!P!PGBr"(](\a8@0BR8r"@00@1 ~0@^0@10@0\a;@0 0²"@0~019A AH! @`P (BB@B0"``<(]h``BB"C0 CA 1A ](!@ߎ!B2B2"a߀! b<P!b`\`2\2²²"$h#ܡ`@!a߰&`@炂‚"޲HC h#܁ h :‚"!bHC$h#܁ h :GBB"B 9!#1`#C!`a~0GB"(@~0@(A GrB"\bH 1 ܑ0\aH r(00H"Q(~0ޑ 0 0^0 0\$0Q` $‚B"0H H܀ 1` H`1H !@ǂB²"#܁P:B`19!`@#܁ @#܁GBBb"`Q:!@ܠ$0$1$1\R:#܁GBB"0``a80HBa8@G2B"\q`H\ 1qXS0hqx`\A|XwB"q8C0h@ h\qX1@QC\hGB ( (@@@2"]@(,,@ #@@,GpB2" 0c@@ @( #\#32"((  0 \A  A0 BB"_8HCA<0 0@(\0 @#|XG B2B" @(0@(((1\0C0 G222"!EHHC\MH(A EPB@C0B"A\b@@ BA@ C@@ܲ@ ܕ 'B"C\@@q@ 㜥 $(C\`@@\p Cu`@@GB B"X(E0]d(vp  (\(G`BPB2"T($a E]h(aCCH\GC"E"EHCMHܲ` XXe\XCOHrB2"$Xݿ8HqQ0Ca<1@H@DzBB2"0% \0 O8H]P@(cQ<`A GBB"A@@ C@@ A܃ 3B‚B"q@@ a@ C r@@`r Cb`@@ BS"bp \s9HXa(X"1 H@@@P@\ 0B0B2" PQ 0@@ 1C@@ P72B"  ((!`G"$hA`@`#AG" !~0\ 0 PQ8 0G²B"A~0\A Q@@8 GB"B8`aܳa0 aPr0GBB2"\8q1B\8$haHXG@B0B"\Q$hA ܡ88C\ahqQB"Xܡ\:AA`CAhG @@@@@. nh 0O;7)pMgpNupTO`p8P P0 Q0PQh;R0RtTu|Tx2V@@@W@ e @qPQ$$Xr#@chain-kernels.cuELF3\@rm##@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i__ocg_const$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt2@a L@="" ."p fs"D@ /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp ~}~( z ~~z (~ |({{(z {{| }(|zz|z~~ t  ~~ t((~~}(({~~~ t  4~~ ~z~yx(zx~~}~|z|wx y t t t t mros~t (  Jzz x  x  }i킀  ~}}}}qzz|w t qz xizo w wxv vnvwfnwp񀀀 탁}p~ { w wv o| x p񀀀 |} ~ wxzx~ws  qv zn {v{ f pqzamzyxuxy xy#] wcofww vo xyz.version 6.2.target sm_35.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   #### @((  !  1 !   X    @LL H @! 8! 0! (!   !   ! !80p @88 0! (!   !   ! !80PQ M ?3s[UU@>>?tLP@Č  `@d*\Q" *\Q6.`.`,\Q&.<j .`@,\Q .<*,+`"<.,+< 2+`((6+< ,DB(@86 @eB@@2:88B0*(028@2Xܬ<$@@[*& @e"<20" 䀠*2* ൬< 0 &(,t"<@!&$@e< "<&<*" ۴* 2@-0@0  *@!0@.0 %A* <&$ "  <䀤   $<0ܬ<,<> $h$倠$<<<<"<&<< @<,XX @  ""  :  << <:<<   %@$~@! @ ~A<<`=<? (۬<x=4h? >)۬<L? )<, >$<t @<\ @` @ @Q"& @2  ̠"&?H"0  ࠠ0  &$@<$H <|$< $ @ <0 ? $@>$X*@!0 ̘0 @($X>X  ⸠< @ ?!<@ Q<< @ ?!< @ <?t<<<<<<) ) ) \@4 )t.#t: -` > -<88 *Q ,Q" -`&-,* +` ĠԠ.+,(( 8@4? >4@@[ @ @㠀, @  <H<\8:8@8\0*Qt,@.` 䌀.8@ <<<8P<48X8 < 98\8:8@:8<@@<,8? *t,!*(>, @"@$ ? $$@! h*(*(@( (( PQ۬$(P8@ @<8<<<<<<@. np 8D|f 8)ptNgppOup P`pP Q0 Q0Rh;pS,S(Uu0Ux2V@@X@@ @f @@rRX~2@chain-kernels.cuELF3\`} x22@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i__ocg_const$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt2a L=" "."@ fs" `D /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp  {} ~0  (8 8y  ~{  ~|}(}|(|0 { {~| |||z |{|~~ t( 0~~0 t0 (~{(~~ t0 4 8y8s8ys(xxx xx xw  r~|~~|{|~}~}y t ~~~~~ (  Jz{ x  x  } i ~  z|  zoz ~ {|(|x(w 0w  t n  x{hmxvomvnnn x  nxn pm~zp ~0| p   z u v w v v  pz p ~큀 ~ r k({k(xxx w w  t n d~ zn kqz{gtfz`qri%xw x  wxw w sv w v타8|z.version 6.2.target sm_50.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   #### @((  !  1 !  (Hx   @LL H @! 8! 0! (!   !   ! !hP p @88 0! (!   !   ! !hPP!  ?3s[UU@>>? DLg'1?W7N"_7OW0['NB'OGcKW0[WcKgNgOG0[)8@'L7L @wNwO70[)8LLGK@K XL !I XL I @ |LW@'N'O70[GmKWL @gGNGO AW0[GNGO @G0[)8L"'H8)8LL L  'L7L!@@`)8L LG   g[ g[ ' g[g[ 0[ g[ g)[$ g/[g[g[ g[ g)['L @ \g/[70[ 7L G \'L"66  7LG\")6'L 6 g[/[W h[ G\7Lg\\\W\ Q@b@6)8W[ @g [(66 6w\\`\ 6G\ g\, w\   G  G     '4 @GNGO GO  w0[GNN @O )80[ @L NW0[ @O )8 L L w0[@)8 LgL@ )8 wL  gL  ? wL  'N '4 h\h\h\ h\ X\@ \'O0[ h\)86 LL'p\\\@_7@Rh8 Ag\)8g[@?7@'N@ 'Og0[ h\6)8LL\ @7@?Rh8g\)8g[@ 7@ '\8 \G\W f[ @b@66g[ @/[)6 6g\\'\G\W e[ @ G  ` GN GO  N0[O" )80[ L@ )8 L gL@  wL  'N'O0[ )8 Gh\h\ w\wh\6@ LLp\ @7@?Rh8\)8wg[@ 7@t< \gPx<qSp[x<p[\ gq[2_p['CHWY 7DH0@\@`L ?g\L 8 GbK\P WLg\  gP gbK@  qS WL p[H\ \ \ p[x<\\g q[?  p[ [@  whK\ \@"\ WL \\B<\ r[ gr[ [\6\@6  \    L\1D?\\\ X w\ Gr[Gr[ [ \"d6 w\@ / \ \ w\g\(\ \gr[6?r['[g\t \@ [t@\ gP \[\  6i\ y\@ \\g\@G\@ \@?0Y\GPQ(WY7YWYYWYGY  g\\@`7\ w8  w8 h6   h6\002@\:@Lw[ 6@000@\4@60AL!P0@60AL!P +@ c[ c[ \22$\ GP 0Y\ QW \ Y@Y Y Y 8 YG\ WYw8 g\ b6@ i6@ c6@ c7@@ WY k[    WY WY: \wH\ ?W[k[K[!PG(\(88<GG\G\7G\@'@7 \@ H' H  gX\ @PPPPP |LW@'N'O70[GmK  A g )8 L@LG\ G    'L'"@66["@ /[ 6 6 @)66 ["@ /[ )67L '\\ 6   \ 6 g\'\B@[/[ )6  g\\6  h[ 6 w \ '\\\g\6@ @6)8"'[W [(6 6'\6 'L7L\ 7\\ 6@7\'\   G  G ` '4GN"@ GO GN GO 0[ W 0[ N"@ O)8L  0[ )8N#OL   LW 0[ )8 L gL  ")8 wLgL\W\ wL \ W\ \ f[`\74 gh\@4wh\h\`G Y g\'p\ @b@66g[ @/[)6 6g\\\\ e[@@ G  + GN GON@0[ O )80[L )8  L gL\@ wL  |Gh\h\ \p\WL @GNGOW0[)8LL  hK@@ `tW\GPQ0Y\WY@D GN g\  GO w0[G\ )8 LL '\ H8 (8j[@DH8k[HP@2GPQ0Y\gY2@7h6@@t  #GP8 Q 0Y\ Y Y 7  X gH\   \ G\ g?[ @ g(\'0B80@8 APG d[ PW(\H8'@?GP@ @PPPPP@. np 8 8)pOgpPuphQ`pLR S0 0S0`Sh;T,TV uVx2 X Y m   `}`S$$X~4@chain-kernels.cuELF3\`} x44@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i__ocg_const$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt2a L=" "."@ fs" `D /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp  {} ~0  (8 8y  ~{  ~|}(}|(|0 { {~| |||z |{|~~ t( 0~~0 t0 (~{(~~ t0 4 8y8s8ys(xxx xx xw  r~|~~|{|~}~}y t ~~~~~ (  Jz{ x  x  } i ~  z|  zoz ~ {|(|x(w 0w  t n  x{hmxvomvnnn x  nxn pm~zp ~0| p   z u v w v v  pz p ~큀 ~ r k({k(xxx w w  t n d~ zn kqz{gtfz`qri%xw x  wxw w sv w v타8|z.version 6.2.target sm_52.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   #### @((  !  1 !  (Hx   @LL H @! 8! 0! (!   !   ! !hP p @88 0! (!   !   ! !hPP!  ?3s[UU@>>? DLg'1?W7N"_7OW0['NB'OGcKW0[WcKgNgOG0[)8@'L7L @wNwO70[)8LLGK@K XL !I XL I @ |LW@'N'O70[GmKWL @gGNGO AW0[GNGO @G0[)8L"'H8)8LL L  'L7L!@@`)8L LG   g[ g[ ' g[g[ 0[ g[ g)[$ g/[g[g[ g[ g)['L @ \g/[70[ 7L G \'L"66  7LG\")6'L 6 g[/[W h[ G\7Lg\\\W\ Q@b@6)8W[ @g [(66 6w\\`\ 6G\ g\, w\   G  G     '4 @GNGO GO  w0[GNN @O )80[ @L NW0[ @O )8 L L w0[@)8 LgL@ )8 wL  gL  ? wL  'N '4 h\h\h\ h\ X\@ \'O0[ h\)86 LL'p\\\@_7@Rh8 Ag\)8g[@?7@'N@ 'Og0[ h\6)8LL\ @7@?Rh8g\)8g[@ 7@ '\8 \G\W f[ @b@66g[ @/[)6 6g\\'\G\W e[ @ G  ` GN GO  N0[O" )80[ L@ )8 L gL@  wL  'N'O0[ )8 Gh\h\ w\wh\6@ LLp\ @7@?Rh8\)8wg[@ 7@t< \gPx<qSp[x<p[\ gq[2_p['CHWY 7DH0@\@`L ?g\L 8 GbK\P WLg\  gP gbK@  qS WL p[H\ \ \ p[x<\\g q[?  p[ [@  whK\ \@"\ WL \\B<\ r[ gr[ [\6\@6  \    L\1D?\\\ X w\ Gr[Gr[ [ \"d6 w\@ / \ \ w\g\(\ \gr[6?r['[g\t \@ [t@\ gP \[\  6i\ y\@ \\g\@G\@ \@?0Y\GPQ(WY7YWYYWYGY  g\\@`7\ w8  w8 h6   h6\002@\:@Lw[ 6@000@\4@60AL!P0@60AL!P +@ c[ c[ \22$\ GP 0Y\ QW \ Y@Y Y Y 8 YG\ WYw8 g\ b6@ i6@ c6@ c7@@ WY k[    WY WY: \wH\ ?W[k[K[!PG(\(88<GG\G\7G\@'@7 \@ H' H  gX\ @PPPPP |LW@'N'O70[GmK  A g )8 L@LG\ G    'L'"@66["@ /[ 6 6 @)66 ["@ /[ )67L '\\ 6   \ 6 g\'\B@[/[ )6  g\\6  h[ 6 w \ '\\\g\6@ @6)8"'[W [(6 6'\6 'L7L\ 7\\ 6@7\'\   G  G ` '4GN"@ GO GN GO 0[ W 0[ N"@ O)8L  0[ )8N#OL   LW 0[ )8 L gL  ")8 wLgL\W\ wL \ W\ \ f[`\74 gh\@4wh\h\`G Y g\'p\ @b@66g[ @/[)6 6g\\\\ e[@@ G  + GN GON@0[ O )80[L )8  L gL\@ wL  |Gh\h\ \p\WL @GNGOW0[)8LL  hK@@ `tW\GPQ0Y\WY@D GN g\  GO w0[G\ )8 LL '\ H8 (8j[@DH8k[HP@2GPQ0Y\gY2@7h6@@t  #GP8 Q 0Y\ Y Y 7  X gH\   \ G\ g?[ @ g(\'0B80@8 APG d[ PW(\H8'@?GP@ @PPPPP@. np 8 8)pOgpPuphQ`pLR S0 0S0`Sh;T,TV uVx2 X Y m   `}`S$$X~<@chain-kernels.cuELF3\~x<<@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i__ocg_const$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt2a L=" "."Xh fs"8HD /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp  {} ~0   8 y  ~  }0}|(|8{ |{ ~| |}||}z x}~~ t0 (~~( t( (~{~~~ t( 4~ 8 y8s0ys0xxx xx xw  r~~~|~|퀀~ t t t t t ~~~~ (  Jz{ x  x  } (i( ~  z|  z{oz ~ |0|x(w 8w  t n  x{hmxvomvnnn x  nxn pm#ezwzp ~( p   {z u v w v v  o| xp } ~ r k~(k0xxx w w  t n d~ zn kqzgtfz`qri%x  m  x|톃 wxw w sv w v}8}}z .version 6.2.target sm_60.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   ####* @((  !  1 !  (H *  @LL H @! 8! 0! (!   !   ! !pX p* @88 0! (!   !   ! !pX0P#  ?3s[UU@>>? DLg'1?W7N"_7OW0['NB'OGcKW0[WcKgNgOG0[)8@'L7L wNwO70[)8LLGK@K XL !I P@PPXL I PPP@ |LW@'N'O70[@GmKPWLgGN?GOW0[GN"GOG0[)8 @L'H8)8LL  L 'L!7L@`)8@LLG  g[& g[' g["g[0[ g[& g)[ g/[g[&g[g[ g)[ 'L \g/[ 70[7L G \ @'L66@ 7L @G\)6'L" 6g[/[ W h[G\7L g\\\W\Q@ @6)8"W[g [(666w\ \\ 6G\ g\,  w\  G   G     '4GNGO  GOw0[GNaNO )8 0[L N@W0[O )8 LL  w0[)8 L gL )8@ wL  gL@  wL  `?'O'4 h\h\'N h\ h\0[ X\ @ h\)8L  6 G\L2@'p\\g\ @ 7@? Rh8g\)8g[@ 7@"'N 'Og0[h\6)8LL\@7@t@Rh8g\)8g[@@7@'\ 8 \G\W f[@ @66"g[/[)6  6g\\'\G\W e[@  G  GN GO N0[ O )80[ L )8 L@ gLw\   wL  'N 'O0[)8LLWh\ Th\h\ \@?6Gp\g\Cw\@7@t@Rh8\)8wg[@@7@ \ x<gPqSx<p[p[\@? gq[p['CHWY 7DH0@\@`Lg\L  6GbK@\ WL g\ gP gbK  qS WL2@ p[ \ \x< p[\\<<g q[  p[ [ @ whK\;H$ \ WL \6@\(\\ r[>? gr[ [\6\@ 6 \    L \\\D\0\ w\ Gr[?Gr[ [ \  d6w\  /  \ \ w\ \\\2<\gr[r[G[\ \@ [@\<<gP[ 6i\ y\ \\g\@?G\@ \@0Y\AGPQWY@7YWYYWYGY g\\@`7\ w8  w8 h6   h6\002@\:@Lw[7@000@\ 4@60AL!P0@60AL!P+@ c[ c[\ 22@P$\GP? 0Y\ QW \ YY  Y Y8 YG\ WYw8 g\ b6@ i6@ c6@ c7@ WY  k[    WY WY@: \wH\W[k[K[!PG(\(88<GG\G\7G\@'@7 \@ H' H  gX\ @PPP |LW@'N'O70[@GmKP?  g )8  LLG\  G ?  'L!@'66@@[ /[ 6" 6)66 @ [ /[ )6 7L'\\  6   \6 g\ @'\[/[" )6 g\\ 6 h[ 6  w \'\\\g\7@ 6"@)8'[W [(66'\ 6'L7L \7\\ 67\'\  G   G  '4 @(GN GO GNB GO0[ W 0[A@ N O)8 L 0[ )8 @NOL LW 0[@ )8 L gL@  )8 wL@gL\W\ wL  \ 74` gh\4   W\wh\h\  \G Y f[@ g\\0'p\@ @66"g[/[)6  6g\\\\ e[ @ G   GN GO N0[ O  )80[L  )8 L gL`\ wL   Gh\h\? \p\WL  GNGO @W0[)8LL hK@t@ `GPQ0Y\gY @ GN GO`  W\ g0[G\? )8 LL2`'\\G\ H8 (8j[@H8 k[HP@@t2GPQ0Y\gY2@7h6@ ? #GP Q  0Y\ Y Y @ 7  X  gH\  \ G\ g?[ g(\'0B80@8AP`G d[ PW(\H8D'@GP@ @PPP@. np 8/  8)p0Pgp,QupQ`pR pS0 S0Sh;8U,dUV uVx2X Z @n   ~S$$X~=@chain-kernels.cuELF3\~x==@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant2._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i__ocg_const$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant2._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt2a L=" "."Xh fs"8HD /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp  {} ~0   8 y  ~  }0}|(|8{ |{ ~| |}||}z x}~~ t0 (~~( t( (~{~~~ t( 4~ 8 y8s0ys0xxx xx xw  r~~~|~|퀀~ t t t t t ~~~~ (  Jz{ x  x  } (i( ~  z|  z{oz ~ |0|x(w 8w  t n  x{hmxvomvnnn x  nxn pm#ezwzp ~( p   {z u v w v v  o| xp } ~ r k~(k0xxx w w  t n d~ zn kqzgtfz`qri%x  m  x|톃 wxw w sv w v}8}}z .version 6.2.target sm_61.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}##   #   ####* @((  !  1 !  (H *  @LL H @! 8! 0! (!   !   ! !pX p* @88 0! (!   !   ! !pX0P#  ?3s[UU@>>? DLg'1?W7N"_7OW0['NB'OGcKW0[WcKgNgOG0[)8@'L7L wNwO70[)8LLGK@K XL !I P@PPXL I PPP@ |LW@'N'O70[@GmKPWLgGN?GOW0[GN"GOG0[)8 @L'H8)8LL  L 'L!7L@`)8@LLG  g[& g[' g["g[0[ g[& g)[ g/[g[&g[g[ g)[ 'L \g/[ 70[7L G \ @'L66@ 7L @G\)6'L" 6g[/[ W h[G\7L g\\\W\Q@ @6)8"W[g [(666w\ \\ 6G\ g\,  w\  G   G     '4GNGO  GOw0[GNaNO )8 0[L N@W0[O )8 LL  w0[)8 L gL )8@ wL  gL@  wL  `?'O'4 h\h\'N h\ h\0[ X\ @ h\)8L  6 G\L2@'p\\g\ @ 7@? Rh8g\)8g[@ 7@"'N 'Og0[h\6)8LL\@7@t@Rh8g\)8g[@@7@'\ 8 \G\W f[@ @66"g[/[)6  6g\\'\G\W e[@  G  GN GO N0[ O )80[ L )8 L@ gLw\   wL  'N 'O0[)8LLWh\ Th\h\ \@?6Gp\g\Cw\@7@t@Rh8\)8wg[@@7@ \ x<gPqSx<p[p[\@? gq[p['CHWY 7DH0@\@`Lg\L  6GbK@\ WL g\ gP gbK  qS WL2@ p[ \ \x< p[\\<<g q[  p[ [ @ whK\;H$ \ WL \6@\(\\ r[>? gr[ [\6\@ 6 \    L \\\D\0\ w\ Gr[?Gr[ [ \  d6w\  /  \ \ w\ \\\2<\gr[r[G[\ \@ [@\<<gP[ 6i\ y\ \\g\@?G\@ \@0Y\AGPQWY@7YWYYWYGY g\\@`7\ w8  w8 h6   h6\002@\:@Lw[7@000@\ 4@60AL!P0@60AL!P+@ c[ c[\ 22@P$\GP? 0Y\ QW \ YY  Y Y8 YG\ WYw8 g\ b6@ i6@ c6@ c7@ WY  k[    WY WY@: \wH\W[k[K[!PG(\(88<GG\G\7G\@'@7 \@ H' H  gX\ @PPP |LW@'N'O70[@GmKP?  g )8  LLG\  G ?  'L!@'66@@[ /[ 6" 6)66 @ [ /[ )6 7L'\\  6   \6 g\ @'\[/[" )6 g\\ 6 h[ 6  w \'\\\g\7@ 6"@)8'[W [(66'\ 6'L7L \7\\ 67\'\  G   G  '4 @(GN GO GNB GO0[ W 0[A@ N O)8 L 0[ )8 @NOL LW 0[@ )8 L gL@  )8 wL@gL\W\ wL  \ 74` gh\4   W\wh\h\  \G Y f[@ g\\0'p\@ @66"g[/[)6  6g\\\\ e[ @ G   GN GO N0[ O  )80[L  )8 L gL`\ wL   Gh\h\? \p\WL  GNGO @W0[)8LL hK@t@ `GPQ0Y\gY @ GN GO`  W\ g0[G\? )8 LL2`'\\G\ H8 (8j[@H8 k[HP@@t2GPQ0Y\gY2@7h6@ ? #GP Q  0Y\ Y Y @ 7  X  gH\  \ G\ g?[ g(\'0B80@8AP`G d[ PW(\H8D'@GP@ @PPP@. np 8/  8)p0Pgp,QupQ`pR pS0 S0Sh;8U,dUV uVx2X Z @n   ~S$$XF@chain-kernels.cuELF3\@~FF@8@.shstrtab.strtab.symtab.symtab_shndx.nv.info.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.debug_frame.rel.debug_frame.shstrtab.strtab.symtab.symtab_shndx.nv.info_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.text._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.info._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.shared._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf.nv.constant0._Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.text._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.info._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i.nv.shared._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm20_div_f64_slowpath_v2$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32$_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i$__cuda_sm3x_div_rn_noftz_f32_slowpath.nv.constant0._Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.text._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.info._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.nv.shared._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf$__cuda_sm20_rcp_rn_f32_slowpath.nv.constant0._Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf.debug_line.rel.debug_line.nv_debug_line_sass.rel.nv_debug_line_sass.nv_debug_ptx_txt.debug_frame.rel.debug_frame2aL" D"p"; "r+C /local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chain/usr/local/cuda/includechain-kernels.cuxcuda_device_runtime_api.hrsm_20_atomic_functions.hpp    0~     0  v  00 |0| z } x0~~   t ~ ~0   t ~~ ~   t 40  0 0 x0s  w z y 0t t t  t 0t  s } t E  J z  zx  i  ~ m 0 w  q xfho   ocx  [u x p   p     0 y w v y z p    ~  w  g  mq nzqxna!vx v nzz   wfr o z.version 6.2.target sm_70.address_size 64.entry _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf(.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4,.param .u32 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6,.param .u64 _Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7){.reg .pred %p<5>;.reg .f32 %f<19>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<52>;ld.param.u64 %rd12, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_0];ld.param.u64 %rd16, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_1];ld.param.u32 %r1, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_2];ld.param.u32 %r2, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_3];ld.param.u64 %rd13, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_4];ld.param.u32 %r3, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_5];ld.param.u64 %rd14, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_6];ld.param.u64 %rd15, [_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_param_7];cvta.to.global.u64 %rd50, %rd16;mov.u32 %r4, %ntid.x;mov.u32 %r5, %ctaid.x;mov.u32 %r6, %tid.x;mad.lo.s32 %r7, %r4, %r5, %r6;setp.ge.s32 %p1, %r7, %r1;@%p1 bra BB0_8;mov.u32 %r8, %ctaid.y;cvta.to.global.u64 %rd17, %rd12;mul.wide.s32 %rd18, %r8, 8;add.s64 %rd19, %rd17, %rd18;ld.global.u32 %r9, [%rd19];cvt.s64.s32 %rd2, %r9;mul.wide.s32 %rd20, %r9, 12;add.s64 %rd51, %rd50, %rd20;ld.global.u32 %r10, [%rd19+4];mul.wide.s32 %rd21, %r10, 12;add.s64 %rd5, %rd50, %rd21;add.s32 %r11, %r9, 2;mul.wide.s32 %rd22, %r11, 12;add.s64 %rd23, %rd50, %rd22;setp.gt.u64 %p2, %rd23, %rd5;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB0_5;mul.lo.s64 %rd6, %rd2, 12;mov.f64 %fd13, 0d0000000000000000;BB0_3:add.s64 %rd24, %rd50, %rd6;ld.global.u32 %r12, [%rd24+4];mad.lo.s32 %r17, %r12, %r3, %r7;cvta.to.global.u64 %rd25, %rd13;mul.wide.s32 %rd26, %r17, 4;add.s64 %rd27, %rd25, %rd26;ld.global.u32 %r18, [%rd24+8];mad.lo.s32 %r19, %r18, %r1, %r7;cvta.to.global.u64 %rd28, %rd14;mul.wide.s32 %rd29, %r19, 4;add.s64 %rd30, %rd28, %rd29;ld.global.u32 %r20, [%rd24+16];mad.lo.s32 %r21, %r20, %r3, %r7;mul.wide.s32 %rd31, %r21, 4;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r22, [%rd24+20];mad.lo.s32 %r23, %r22, %r1, %r7;mul.wide.s32 %rd33, %r23, 4;add.s64 %rd34, %rd28, %rd33;ld.global.f32 %f1, [%rd30];ld.global.f32 %f2, [%rd24];mul.f32 %f3, %f2, %f1;ld.global.f32 %f4, [%rd27];ld.global.f32 %f5, [%rd34];ld.global.f32 %f6, [%rd24+12];mul.f32 %f7, %f6, %f5;ld.global.f32 %f8, [%rd32];mul.f32 %f9, %f8, %f7;fma.rn.f32 %f10, %f4, %f3, %f9;cvt.f64.f32 %fd8, %f10;add.f64 %fd13, %fd13, %fd8;add.s64 %rd50, %rd50, 24;add.s64 %rd35, %rd50, %rd6;add.s64 %rd36, %rd35, 24;setp.le.u64 %p3, %rd36, %rd5;@%p3 bra BB0_3;cvt.u32.u64 %r24, %rd2;mul.wide.s32 %rd37, %r24, 12;add.s64 %rd51, %rd50, %rd37;BB0_5:setp.eq.s64 %p4, %rd51, %rd5;@%p4 bra BB0_7;ld.global.u32 %r25, [%rd51+4];mad.lo.s32 %r30, %r25, %r3, %r7;cvta.to.global.u64 %rd38, %rd13;mul.wide.s32 %rd39, %r30, 4;add.s64 %rd40, %rd38, %rd39;ld.global.u32 %r31, [%rd51+8];mad.lo.s32 %r32, %r31, %r1, %r7;cvta.to.global.u64 %rd41, %rd14;mul.wide.s32 %rd42, %r32, 4;add.s64 %rd43, %rd41, %rd42;ld.global.f32 %f11, [%rd43];ld.global.f32 %f12, [%rd51];mul.f32 %f13, %f12, %f11;ld.global.f32 %f14, [%rd40];mul.f32 %f15, %f14, %f13;cvt.f64.f32 %fd9, %f15;add.f64 %fd13, %fd13, %fd9;BB0_7:mad.lo.s32 %r37, %r2, %r1, %r7;cvta.to.global.u64 %rd44, %rd14;mul.wide.s32 %rd45, %r37, 4;add.s64 %rd46, %rd44, %rd45;ld.global.f32 %f16, [%rd46];rcp.rn.f32 %f17, %f16;cvt.f64.f32 %fd10, %f17;mul.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f18, %fd11;mad.lo.s32 %r39, %r8, %r1, %r7;cvta.to.global.u64 %rd47, %rd15;mul.wide.s32 %rd48, %r39, 4;add.s64 %rd49, %rd47, %rd48;st.global.f32 [%rd49], %f18;BB0_8:ret;}.entry _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i(.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8,.param .u64 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9,.param .u32 _Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10){.reg .pred %p<11>;.reg .f32 %f<33>;.reg .b32 %r<40>;.reg .f64 %fd<15>;.reg .b64 %rd<61>;ld.param.u64 %rd18, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_0];ld.param.u64 %rd21, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_1];ld.param.u32 %r4, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_2];ld.param.u32 %r5, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_3];ld.param.u64 %rd22, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_4];ld.param.u32 %r6, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_5];ld.param.u64 %rd19, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_6];ld.param.u64 %rd23, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_7];ld.param.u64 %rd20, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_8];ld.param.u64 %rd24, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_9];ld.param.u32 %r7, [_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_param_10];cvta.to.global.u64 %rd1, %rd24;cvta.to.global.u64 %rd2, %rd22;cvta.to.global.u64 %rd3, %rd23;cvta.to.global.u64 %rd59, %rd21;mov.u32 %r8, %ntid.x;mov.u32 %r9, %ctaid.x;mov.u32 %r10, %tid.x;mad.lo.s32 %r1, %r8, %r9, %r10;setp.ge.s32 %p1, %r1, %r4;@%p1 bra BB1_19;cvta.to.global.u64 %rd25, %rd18;cvta.to.global.u64 %rd26, %rd19;mov.u32 %r11, %ctaid.y;mad.lo.s32 %r2, %r11, %r4, %r1;mul.wide.s32 %rd27, %r2, 4;add.s64 %rd28, %rd26, %rd27;mad.lo.s32 %r12, %r5, %r4, %r1;mul.wide.s32 %rd29, %r12, 4;add.s64 %rd30, %rd26, %rd29;ld.global.f32 %f1, [%rd30];ld.global.f32 %f7, [%rd28];div.rn.f32 %f2, %f7, %f1;mul.wide.s32 %rd31, %r11, 8;add.s64 %rd32, %rd25, %rd31;ld.global.u32 %r13, [%rd32];cvt.s64.s32 %rd5, %r13;mul.wide.s32 %rd33, %r13, 12;add.s64 %rd60, %rd59, %rd33;ld.global.u32 %r14, [%rd32+4];mul.wide.s32 %rd34, %r14, 12;add.s64 %rd8, %rd59, %rd34;add.s32 %r15, %r13, 2;mul.wide.s32 %rd35, %r15, 12;add.s64 %rd36, %rd59, %rd35;setp.gt.u64 %p2, %rd36, %rd8;mov.f64 %fd13, 0d0000000000000000;@%p2 bra BB1_13;mul.lo.s64 %rd9, %rd5, 12;mov.f64 %fd13, 0d0000000000000000;BB1_3:add.s64 %rd37, %rd59, %rd9;ld.global.u32 %r16, [%rd37+8];mad.lo.s32 %r17, %r16, %r4, %r1;mul.wide.s32 %rd38, %r17, 4;add.s64 %rd39, %rd3, %rd38;ld.global.f32 %f8, [%rd39];ld.global.f32 %f9, [%rd37];mul.f32 %f10, %f9, %f8;ld.global.u32 %r18, [%rd37+4];mad.lo.s32 %r19, %r18, %r6, %r1;mul.wide.s32 %rd40, %r19, 4;add.s64 %rd41, %rd2, %rd40;ld.global.f32 %f11, [%rd41];mul.f32 %f12, %f10, %f11;ld.global.u32 %r20, [%rd37+20];mad.lo.s32 %r21, %r20, %r4, %r1;mul.wide.s32 %rd42, %r21, 4;add.s64 %rd43, %rd3, %rd42;ld.global.f32 %f13, [%rd43];ld.global.f32 %f14, [%rd37+12];mul.f32 %f15, %f14, %f13;ld.global.u32 %r3, [%rd37+16];mad.lo.s32 %r22, %r3, %r6, %r1;mul.wide.s32 %rd44, %r22, 4;add.s64 %rd45, %rd2, %rd44;ld.global.f32 %f16, [%rd45];mul.f32 %f3, %f15, %f16;add.f32 %f17, %f12, %f3;cvt.f64.f32 %fd8, %f17;add.f64 %fd13, %fd13, %fd8;mul.f32 %f4, %f2, %f12;mad.lo.s32 %r23, %r18, %r7, %r1;mul.wide.s32 %rd46, %r23, 4;add.s64 %rd11, %rd1, %rd46;setp.ltu.f32 %p3, %f4, 0f38800000;@%p3 bra BB1_5;bra.uni BB1_4;BB1_5:mul.f32 %f19, %f4, 0f52800000;cvt.rzi.s32.f32 %r24, %f19;shr.s32 %r25, %r24, 12;and.b32 %r26, %r24, 4095;setp.le.s32 %p4, %r25, %r26;@%p4 bra BB1_7;atom.global.add.f32 %f20, [%rd11], 0f38800000;bra.uni BB1_7;BB1_4:atom.global.add.f32 %f18, [%rd11], %f4;BB1_7:mul.f32 %f5, %f2, %f3;mad.lo.s32 %r27, %r3, %r7, %r1;mul.wide.s32 %rd47, %r27, 4;add.s64 %rd12, %rd1, %rd47;setp.ltu.f32 %p5, %f5, 0f38800000;@%p5 bra BB1_9;bra.uni BB1_8;BB1_9:mul.f32 %f22, %f5, 0f52800000;cvt.rzi.s32.f32 %r28, %f22;shr.s32 %r29, %r28, 12;and.b32 %r30, %r28, 4095;setp.le.s32 %p6, %r29, %r30;@%p6 bra BB1_11;atom.global.add.f32 %f23, [%rd12], 0f38800000;bra.uni BB1_11;BB1_8:atom.global.add.f32 %f21, [%rd12], %f5;BB1_11:add.s64 %rd59, %rd59, 24;add.s64 %rd48, %rd59, %rd9;add.s64 %rd49, %rd48, 24;setp.le.u64 %p7, %rd49, %rd8;@%p7 bra BB1_3;cvt.u32.u64 %r31, %rd5;mul.wide.s32 %rd50, %r31, 12;add.s64 %rd60, %rd59, %rd50;BB1_13:setp.eq.s64 %p8, %rd60, %rd8;@%p8 bra BB1_18;ld.global.u32 %r32, [%rd60+8];mad.lo.s32 %r33, %r32, %r4, %r1;mul.wide.s32 %rd51, %r33, 4;add.s64 %rd52, %rd3, %rd51;ld.global.f32 %f24, [%rd52];ld.global.f32 %f25, [%rd60];mul.f32 %f26, %f25, %f24;ld.global.u32 %r34, [%rd60+4];mad.lo.s32 %r35, %r34, %r6, %r1;mul.wide.s32 %rd53, %r35, 4;add.s64 %rd54, %rd2, %rd53;ld.global.f32 %f27, [%rd54];mul.f32 %f28, %f26, %f27;cvt.f64.f32 %fd9, %f28;add.f64 %fd13, %fd13, %fd9;mul.f32 %f6, %f2, %f28;mad.lo.s32 %r36, %r34, %r7, %r1;mul.wide.s32 %rd55, %r36, 4;add.s64 %rd17, %rd1, %rd55;setp.ltu.f32 %p9, %f6, 0f38800000;@%p9 bra BB1_16;bra.uni BB1_15;BB1_16:mul.f32 %f30, %f6, 0f52800000;cvt.rzi.s32.f32 %r37, %f30;shr.s32 %r38, %r37, 12;and.b32 %r39, %r37, 4095;setp.le.s32 %p10, %r38, %r39;@%p10 bra BB1_18;atom.global.add.f32 %f31, [%rd17], 0f38800000;bra.uni BB1_18;BB1_15:atom.global.add.f32 %f29, [%rd17], %f6;BB1_18:cvta.to.global.u64 %rd56, %rd20;cvt.f64.f32 %fd10, %f1;div.rn.f64 %fd11, %fd13, %fd10;cvt.rn.f32.f64 %f32, %fd11;add.s64 %rd58, %rd56, %rd27;st.global.f32 [%rd58], %f32;BB1_19:ret;}.entry _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf(.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0,.param .f32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2,.param .align 4 .b8 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3[12],.param .u32 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4,.param .u64 _Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5){.reg .pred %p<6>;.reg .f32 %f<13>;.reg .b32 %r<15>;.reg .b64 %rd<9>;ld.param.f32 %f2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_0];ld.param.f32 %f3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_1];ld.param.u64 %rd2, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_2];ld.param.u32 %r5, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+8];ld.param.u32 %r3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3];ld.param.u32 %r4, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_3+4];ld.param.u32 %r6, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_4];ld.param.u64 %rd3, [_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_param_5];mov.u32 %r7, %ntid.x;mov.u32 %r8, %ctaid.x;mov.u32 %r9, %tid.x;mad.lo.s32 %r1, %r7, %r8, %r9;mov.u32 %r10, %ntid.y;mov.u32 %r11, %ctaid.y;mov.u32 %r12, %tid.y;mad.lo.s32 %r2, %r10, %r11, %r12;setp.lt.s32 %p1, %r1, %r4;setp.lt.s32 %p2, %r2, %r3;and.pred %p3, %p1, %p2;@!%p3 bra BB2_5;bra.uni BB2_1;BB2_1:cvta.to.global.u64 %rd4, %rd3;mad.lo.s32 %r13, %r2, %r5, %r1;mad.lo.s32 %r14, %r2, %r6, %r1;cvta.to.global.u64 %rd5, %rd2;mul.wide.s32 %rd6, %r13, 4;add.s64 %rd7, %rd5, %rd6;neg.f32 %f4, %f2;ld.global.f32 %f1, [%rd7];setp.lt.f32 %p4, %f1, %f4;mul.wide.s32 %rd8, %r14, 4;add.s64 %rd1, %rd4, %rd8;@%p4 bra BB2_4;bra.uni BB2_2;BB2_4:add.f32 %f9, %f1, %f2;mul.f32 %f10, %f9, %f3;ld.global.f32 %f11, [%rd1];sub.f32 %f12, %f11, %f10;st.global.f32 [%rd1], %f12;bra.uni BB2_5;BB2_2:setp.leu.f32 %p5, %f1, %f2;@%p5 bra BB2_5;sub.f32 %f5, %f1, %f2;mul.f32 %f6, %f5, %f3;ld.global.f32 %f7, [%rd1];sub.f32 %f8, %f7, %f6;st.global.f32 [%rd1], %f8;BB2_5:ret;}(| (((0 (t(| (((0  (8|( (((((  p 8|( (((((  H`8|( (((((  (| (((0 (8|( (((((  p#   #   ##### `((  !  1 !  @  `LL H @! 8! 0! (!   !   ! !`   `88 0! (!   !   ! !`i x' H8`  z y&"y""y%by!b$z$z/ z\p z]ppMx$z^%vZs$z_%v` zXOG p zXM s!vX#zYOsMys!vX#zYOsMyGy$v y%"y!$$z z\pbM y&"$v]$t $z\$z\%vb %vb s(s"EyP$vZz[xPDy0Ay$t%vXss$t xO%vZ %vZ  rp@ rA%vZ xG xx$x %x $r$xss s(sb$tshs b$z\O$z \%vd$z `%v d$z`%v ^%v^ss s (sb$zjEy`%vh r@O r @ r @ r@!r r@s " x8)rG yGGy xR@s $x x  r p2G x8yGAy r@ x8$zjEy%vhG yGGy xR@s $x x  r p2G $t8yGAyx$r$r$r$x %x $rxr rp0 r1G 0%x  rp  r#EyG ss$t s "$z\O$z`%vd %v^ s(s r  @ r @O r@s  & x8$z j%v h )r G y GGy xR@s $x x rp2G $t8y GAys "$ts *+t?+r+r(r+r +r " x3s[#r  xUUBEy@G xP Dy`Ays0"$t%vfsMyx @x_ x@px $r$r $r$r (r s "xx x>px+t ?+r d$r $r $r$r+r .(r (r+r +r  *r "EyG  x>p@rx(r J(r (rH.(r +r  H.+r P*r @B.r  xrG$t PX x $rx ?xx(r"$r(r +r dr+r  P*r @B.rxxx$r  xp x rr  r $rr$r +r H+r  P*r @B.rr Gy*r G s"r*rB*$$(r "Gy(r Ay$tr$rPys$G s"!r#t?#r#r#r#r#r#r$tPyEyP$rr xDy@Ay$t$rPyx xxxx x  x p@ x pDpEy$G x x xp5pB G x ȀBG x" x " x"BGpxxrTBGx xr4pB G  r p r p$#_# _@Ayx $x  s "!r x$x#t ? #r #r  #r #r #r x #r $x  xxr x xpEy$rG P xpBEyB G  xpBG xpxB G Ay#r x  x#r x#r @ rpR$r   rpRr   rr  r pRrxp5pxxxr$xrGy@xxGyr Ay$tPyx H$txPy@x H$tPyx$tPy$t!r PyGyyyz y%"y!$$z z\pbM y&"$t%vXssx xO%vZ %vZ  rp@ rA%v Z xG $vZz[xx $x %x  $r$x s s s (s bxs b$z \O$z \%v b$z`$z`%vb%v^s %v^s ss(sbx$rr  r @Or  r@ȏ%x   r@#rx s  "$x   r p0 r 1)r V0G %x  ? r p  r #  z]$t$z\%vbs"$\O%b$`ȏ%^ xx xp@Ey @O @̏  $) "G0xDyGy@s$#t!r#rAys "$z\(rdxs0".%vdsMyxx rpPEyEyB G $x rpR"BGAy#x_s$#t!r#r#x_Gyx xp@G x x ?s "x #t  !r #r @#r x r  r x x r r r  r  x x xpUxp0px r p r $xGys$AyrxPyGyyyy@ I  G*IVP 8h]L)pPOgpLPupPpQ `R0> R0u Rp0S;TdV2XZ s  @0S**_Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf_Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_PfCbF'@@@\b v int  x} y} z} y x}y}z}I^}}}r} }      !"#$%&'()*+,-./01236789:;<=>?@ABCDEFGHIJKLMNOPQR' ! &G $ % & ' ( )} +d ,v =d >v@\rv b  d b  d std!   } !_Tp" 8*" 9# f# # # # !# 6# L# b# # # # # # ## C# d# v# # # # # # # -# # # G# b# # x# # # 0# 1# 2# 3# 5# 6 # 7# 8# :# ;# <# =# ?l# @V# Bz# C# D# E# G*# H5# I@# JK# L# M# N# O# Qw# Ra$%&# # G# b# x# # # 'div  - (;f)< rem= >A(C)Dd remEd Fq(M)N remO P    0 % *###)+,absE@#&*-5\#'G-7x#(c-9#)-;#*.cos>#+.sin@#,.tanB #--G% #. -IA #/, -K] #0H .exp_y #1d -b  #2 -e #3 .logh #4 -k #5 -n  #6 .powA #7' -] #8H -y #9d - #: - #; - #< -U #N -W #O -Y& #P -B #Q- -c #RI .erf #Sj - #T - #U -w #V /F #W ,fmaO #X /I> #Y# /L` #ZE - #[g / #\ - #] /< #^ /B #_ -z #` -, #a -}H #b3 /:de #cO /@d #dl /& #e / #f / #g /#h /3, #i /I#j3/*f#kP/"d#lm/#m-#n/.#o/P01U!-6/iL/ndb-b##kk,divQf/w\/Sdd/k-qkk/Ck2;dbkk3lv42}-u\-d-v/ -k'\'k-SG\/Wb/ux---{--~#'#+#.d#6*#7!#86#9L#:b#<#>#@#C#D#E##GC#Hv#J#K#L#M#N#P#Q-"c#*"d# "e#d "f# "g#H "i#' "k# "l# "m# "n# "o#d "p# "q# "r#c"s#"t#, "u#G"v#"w# "x#"y#"z#"{#H GHIJK   % ; + , - /d 6 7 8} :v D Fd Gd Hd Q Sv Tv Uv ad dv oF pQ  .)/)0)1 2 N)O)P Q %- (?)))+),TvD5T6\ (`)a)b)c)db f~7vl78> 96--!T:val63;<=r>?@t'AUsAT AQ AR AX AY0@t'*AUsAT AQ AR AX AY0Bt'AUsAT AQ AR AX AY0@'AU BAU C\DE\FGH'FI'I'I'FI'I'I'L < dJJJFJdJJLKjLMVN__fzKKKk~K~O&%w?S%?H%?=%?4%O&%?S%?H%?=%?4%O$+?%? %?%P$P$?$B'AU AYB'AUATAQ~AR~Q';bzvLDEE EEEFEEFEFELELE?<oJJ JJJFJJFJFJLJLJKoLMaN__fzK~K~Kk~K~O&%?S%?H%?=%?4%O&%?S%?H%?=%?4%O$6?%? %?%P$P$~?$B'AU AYB'AU~AT~AQ~AR~Q'bv DhEhEh EhEhEhFEhEhFEhL< m J J J J J FJ J FJ LK  L M N__f zK ~K ~K k~K ~O&% '?S%?H%?=%?4%O&% g?S%?H%?=%?4%O$ ?%? %?%P$P$?$B'AU AYB'AU~AT~AQ~AR~Q'b$ vR P Q'Ds EbS8!TGr8TBl8U8U9U9FTdim:U:V;LW<!?9?-P!L?? ?BRAa-Ab-AUsATLAQvAR|@'!AX0AY0Q'S{k#TGrPTBl@UU VVVFVVF VF(VL0VL8VWOR#PP8P0P(P PPPPPVPSBAUsATvAQARAXAYB'AX0AY0SU$TGrPTBl@UU VVVFVV F V L(W) $P(P PPPPPVPSBAUsATvAQARAXAYB'AX0AY0X)&%I)#I*I*I+\I+kI,Y54%_%Z[vx}[vy}[vz}\&%p%%]4%]=%]H%]S%D%:in\^6&??????PPBAUUATTAQQARRAXXAYY^#&??????PPPPP BAUUATTAQQARRAXXAYY^ t'?? ?P!`?-?9BRAa-Ab-AUUATwAQRARX__`> _Ma_` % $ > 4: ; I &I5I4: ; I I : ; I $ >   : ;  : ; I8 : ;I : ; : ;I8 .?: ;n<dI4I.?: ;nI<d> I: ; ( (<;I!I/ 4: ; I?<4: ;I?<9: ; 9: ; :: ;  > I: ; !/I": ;#: ; $9: ; %9: ;&:: ;'.?: ; nI<( : ; n) : ; I8 *I+&,.?: ;I<-.?: ; I<..?: ; I</.?: ;I<01.?: ;nI<2.?: ;<3.?: ;<4.?: ;I<54: ; I64: ;I7!I/84: ;I 9.: ; I :: ; I; I<.: ; @B= >1X Y ?1@1ABB1CD.: ; E: ; IF G5H.: ; I: ;IJ: ; IK4: ; IL4: ; IM UN4: ; IO1X Y P1Q1R.: ;@BS.?: ;@BT: ;IU: ;IV: ;IW1X YX.: ;I Y.G dZI4[: ;I\.1nd]1^.1@B_.?<n: ; `.?<n: ;a.?<nGsPsS`a`b`U@` T HT`QP`Rp012345611\t~\t~\iitYt{6{ UTQRXYw wp012!3!.4.;5;H6HU7Ub8bo9o:=;1~1~1~1~&&1Y188 UTQRXYP]0]j1jw2w34567m81~1~Fa~Fa~FVVaYah h a0b0XSX0SYVY0V +\ +V+S++ $X$2S24X4S Y3V34Y4VXSX SYVY Vm}U}Um}T}Tm}Q}Qm}R}Rm}X}Xm}Y}Y=ZUZ`U=ZTZ`T=ZQZ`Q=ZRZ`R=ZXZ`X=ZYZ`Ya-b-UURPRXX,Fb2\uz /tmp/usr/local/cuda/include/usr/local/cuda/include/crt/usr/lib/gcc/x86_64-linux-gnu/7/include/usr/include/x86_64-linux-gnu/bits/usr/include/usr/include/c++/7/bits/usr/include/c++/7/usr/include/x86_64-linux-gnu/c++/7/bits../cudamatrix../chaintmpxft_00007e54_00000000-5_chain-kernels.compute_70.cudafe1.stub.cvector_types.hcuda_runtime_api.hchain-kernels.cuhost_runtime.hstddef.hdriver_types.htypes.htime.hcpp_type_traits.hcmathcstdlibcstdintc++config.h math.hstdlib.hstdint-intn.hmath.hmathcalls.hstdlib-float.hstdlib-bsearch.hstdlib.hstdlib.hmath_functions.hdevice_launch_parameters.hstdint-uintn.hstdint.hcu-matrixdim.h chain-datastruct.h fatBinaryCtl.htmpxft_00007e54_00000000-2_chain-kernels.fatbin.cdevice_functions.h  ={|0:^E"XJ~XW|0:gE"XJ~XN|0:^EX TXLUj~t) .ltO.4hk+.}X WGNU C++11 7.4.0 -msse -msse2 -m64 -mtune=generic -march=x86-64 -g -O1 -std=c++11 -std=c++11 -fPIC -fstack-protector-strongint8_tthis_betasize_tcudaErrorCooperativeLaunchTooLarge_penalize_out_of_range__valuecudaErrorInvalidDeviceFunctioncudaErrorInvalidSurfaceuint_fast64_tint32_t__par0__par1__par2__par3__par4__par5__par6__par7cols__par9cudaErrorMissingConfigurationsqrt__device_stub__Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_PflldivcudaErrorProfilerAlreadyStartedcudaErrorInvalidMemcpyDirection__gridDimfminstrtoldcudaErrorInvalidConfigurationstrtolldiv_t__cuda_0cudaErrorTextureNotBoundat_quick_exit__cuda_9cudaErrorDuplicateVariableNameint_fast8_t__args_arrcudaErrorDeviceAlreadyInUse__compar_fn_tcudaErrorMisalignedAddresscudaErrorInvalidSymbolatoll____nv_dummy_param_ref__is_integeruint_fast32_tprev_alpha__uint8_tout_deriv__sti____cudaRegisterAllnexttowardatofatoigridDimatolcudaErrorInsufficientDrivercudaErrorSharedObjectSymbolNotFound__device_builtin_variable_threadIdxlong int7lldiv_twctomb__cudaRegisterFunctionrand__reftransition_probdatatanhcudaErrorInvalidValuecallback_fpcudaErrorSyncDepthExceededuint_least64_tsigned charuint8_tnearbyint__timezonenextaftercudaErrorLaunchOutOfResourcessystemnum_hmm_statescudaLaunchKernelscaleunsigned charstridecudaErrorLaunchFileScopedSurfcopysignfdimcudaErrorHostMemoryAlreadyRegistered__args_idxfmodcudaErrorLaunchFileScopedTexlogbcudaErrorIllegalAddress__tznamecudaErrorMemoryValueTooLargelrintcoshcharcudaErrorInvalidPitchValuecudaErrorCudartUnloadingldiv_ZN9__gnu_cxx3divExxcudaErrorInvalidTextureuint_least32_t__uint16_tcudaErrorDevicesUnavailablelog10cudaErrorPeerAccessUnsupportedcudaErrorUnsupportedLimitsinh5div_ttimezonerowscudaErrorPeerAccessNotEnableddecltype(nullptr)fatbinDataint_least16_tmblenatan2cudaErrorMapBufferObjectFailedlog1pstream__nv_inited_managed_rt__cudaAddressOfdim3limitatanh_cuda_chain_hmm_forwardcudaErrorNoDeviceInt32Pair__cudaPushCallConfigurationstrtoulcudaErrorJitCompilerNotFound_ZN4dim3C2EjjjtgammacudaErrorInvalidFilterSettingcudaErrorInitializationErrorcudaStream_tcudaErrorInvalidPtxcudaErrorLaunchMaxDepthExceededcudaErrorProfilerAlreadyStoppedcudaErrorSetOnActiveProcess__nv_save_fatbinhandle_for_managed_rtfloorlog2double_t__blockDimcudaErrorInvalidPccuda_chain_hmm_backwardBaseFloatfabs__cuda_1__cuda_2__cuda_3__cuda_4__cuda_5__cuda_6__cuda_7__cuda_8__uint64_tgetenvuint_least16_tcudaErrorECCUncorrectablelong unsigned intlldiv_tcudaErrorInvalidDevicePointerparamint_least32_tcudaErrorLaunchFailure6ldiv_tcudaErrorUnknownint_least64_tthisprobserfcint32_cudauintptr_tasinhnext_betacudaError_tcudaLaunchKernel_ptszcudaErrorargscudaErrorInvalidChannelDescriptor__device_builtin_variable_warpSize__fatDeviceTextcudaErrorNotSupported__nv_fatbinhandle_for_managed_rt__device_builtin_variable_blockDimcudaErrorIllegalInstruction_ZN4dim3C4Ejjjout_stridein_data__uint32_t__stack_chk_failcudaErrorNotYetImplementedlong long intforward_transitionscudaErrorInvalidAddressSpace__fatBinC_Wrapper_tsecondcuda_chain_hmm_forwarddouble_ZN4dim3C4E5uint3__cudaRegisterFatBinarysharedMemuintmax_t/local_disk/orion/ontrac/yannick/kaldi_20190717/kaldi/src/chaincuda_penalize_out_of_rangecudaErrorNoKernelImageForDevicecudaErrorInvalidGraphicsContextround__cudaFatCubinHandlefloatcudaErrorNotReadynum_sequences__intmax_tcudaErrorMemoryAllocationunsigned int_cuda_chain_hmm_backwardllroundintptr_t__NV_MODULE_IDtznamecudaErrorApiFailureBasewcstombs__device_stub__Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_iremquocudaErrorProfilerDisabledcudaErrorInvalidResourceHandlecudaErrorAssertint_fast16_tfirstcbrtbackward_transitionspdf_id__gnu_cxxscalblnuint_least8_tcudaSuccessint_least8_tstrtoullqsortcudaErrorStartupFailureintmax_tlong doublefloat_tprob_stridetransitionsilogb19__fatBinC_Wrapper_tcudaErrorProfilerNotInitializedint_fast32_tldiv_t__cudaUnregisterFatBinarycudaErrorMixedDeviceExecutionatanbsearch__int8_tcudaErrorTooManyPeerslong long unsigned intsigngamldexpcudaErrorDuplicateSurfaceNameuint16_tfrexpcudaErrorAddressOfConstantcudaErrorIncompatibleDriverContextacosexpm1__module_id_strint_fast64_tCUstream_st__cudaPopCallConfigurationmbtowc__cudaUnregisterBinaryUtilcudaErrorPeerAccessAlreadyEnabledasin__cxx11MatrixDim_truncquotversionacoshcudaErrorNvlinkUncorrectable__int16_tcudaErrorLaunchTimeoutthis_alphaDenominatorGraphTransitionfmaxrintlgammauint3strtodstrtofcudaErrorHostMemoryNotRegisteredblockDimuint_fast8_tstrtollroundshort int__device_builtin_variable_blockIdxscalbncudaErrorSynchronizationErroruint64_tmodffuncint16_tatexit__stream_ZN4dim3cv5uint3Ev__sharedMem__uintmax_thypot__cuda_10cudaErrorNotPermittedsrandcudaErrorInvalidKernelImage__daylightlog_prob_deriv_stride__device_stub__Z22_penalize_out_of_rangeffPKf10MatrixDim_iPf__int64_t__nv_cudaEntityRegisterCallbackcudaErrorTextureFetchFailedmbstowcscudaErrorUnmapBufferObjectFailed__par10getdate_errfilename_or_fatbinsuint32_tcudaErrorInvalidNormSettingcudaErrorInvalidDevice__par8uint_fast16_tMatrixDim__device_builtin_variable_gridDimceil/tmp/tmpxft_00007e54_00000000-5_chain-kernels.compute_70.cudafe1.cppshort unsigned intmagiccudaErrorDuplicateTextureNamecudaErrorOperatingSystem__int32_tint64_tcudaErrorHardwareStackErrorcudaErrorInvalidHostPointercudaErrorInvalidTextureBindingdaylightcudaErrorLaunchPendingCountExceededexp2operator uint3cudaErrorSharedObjectInitFailed__T0cudaErrorPriorLaunchFailureremainderllrintlog_prob_derivquick_exitwchar_thmm_stateGCC: (Ubuntu 7.4.0-1ubuntu1~18.04.1) 7.4.0zRx (mG A jDXHDDD I(hG\ A sDX$#DE D(D0D8D@I$ G A jDU D [DP0UAAD@a AAA DHDPa@<H{AAD@a AAA DHHPGXG`GhGpg@0BAA Dpg  AABA DAXB B(B0iBB B(B0iBB B(B0iMALmm7=#:`  0    050:?UpU {/Ftmpxft_00007e54_00000000-5_chain-kernels.compute_70.cudafe1.cppfatbinData_ZL94__device_stub__Z23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_PfPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_ZL23_cuda_chain_hmm_forwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_Pf_ZL102__device_stub__Z24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_iPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_ZL24_cuda_chain_hmm_backwardPK9Int32PairPK26DenominatorGraphTransitioniiPKfiS6_S6_PfS7_i_ZL60__device_stub__Z22_penalize_out_of_rangeffPKf10MatrixDim_iPfffPKfR10MatrixDim_iPf_ZL22_penalize_out_of_rangeffPKf10MatrixDim_iPf_ZL26__cudaUnregisterBinaryUtilv_ZL20__cudaFatCubinHandle_ZL24__sti____cudaRegisterAllv_ZL15__fatDeviceText.LC0.LC1.LC2_GLOBAL_OFFSET_TABLE___cudaPopCallConfigurationcudaLaunchKernel_ptsz__stack_chk_fail__cudaUnregisterFatBinarycuda_chain_hmm_forward__cudaPushCallConfigurationcuda_chain_hmm_backwardcuda_penalize_out_of_range__cudaRegisterFatBinary__cudaRegisterFunctionatexit"^#i$".#9$"q#|$%'%'',$49*C^p+++,0   5  ) 0  5 C :R  g :l y >    Y  z t R   t 6 =  _ f  ~ e J     f       &   E   ? %   j& , X 2 8 > . D J P EV z\ b h n at Zz   7           $   t       c    4        " ( . 4 : @ F L nR X <^ d  j p  v  | t         Q   V    c  R     &  1 <  G  R is  ~   O 3 &   Y  b   x  + m 6 E N g ku  ~  @ C  P  { 2   H ;d    5- I     (I e 0  b     . AJ   ` @ $ cF 8h u  m  /  h 4 P /m       4 hQ  n l  }  :  " 7 M c    s } a $ D e w j  z  I ,  n. LH c <y c   =D O Z  e p P {      o |    t       . 3  V+ 6 A nL  W 7b  m x      / Z g  # ;   #   H u' [3 Z Fd @i  v        d s   : @0WGl {G t= 00+BmO\s= %  o  q     "  .  :  S Y`p "z 6 ) Y 0 | 7  >  E 9   p ` .  v  ]) A<W  `  i  r 0| V V V z\    9 u,W  q                  { w0 ": e? )I N 0X ] 7g l >v I{ E l L  S    _      W 0 . v % ]4 AGb Wk Wt W} {       5 2%=7b=  q                "   )   0* "/ 79 [> >H M EW \ Lf k Su z   "   . v  ] A      "  ,G 0 P 0 Y 0 b T lF z       bm m% , C Q B] v h ot H {   ? S w     ! .! m 7!  H!  Q!  Z!  _!)!!0!  ! ! ` "  " "  "  -" 3<"  K" Z" =i"  x" " " "4#S#)l#  s## F#  # # =#  # 3#  # $ |$ =$$z$$$ B$ -$ $ % ) %  % 5%  d% n% < %m% % % % G% % %~;&=V& _& +h& dq& z& & &[&' H' ' +' 4' C9'u' Sy' S'  '  '  '  ' o' o' & ' & ' G' G' -' -L=& m  Lml=`L 0.symtab.strtab.shstrtab.rela.text.data.bss.nv_fatbin.rodata.str1.8.rela.init_array.rela.nvFatBinSegment.rela.debug_info.debug_abbrev.rela.debug_loc.rela.debug_aranges.debug_ranges.rela.debug_line.debug_str.comment.note.GNU-stack.rela.eh_frame @@X&(,(1(e<28lP mK@a(m\@ w@m'r@h7 |@p H0@ 0Q@ 0ڱ0,@8! 0M