.visible .entry main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic( .param .u64 .ptr .global .align 16 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_0, .param .u64 .ptr .global .align 16 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_1, .param .u64 .ptr .global .align 16 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_2, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_3, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_4, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_5, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_6, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_7, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_8, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_9, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_10, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_11, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_12, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_13, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_14, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_15, .param .u32 main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_16 ) .maxntid 1024, 1, 1 // @main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic { .reg .pred %p<174>; .reg .b16 %rs<73>; .reg .b32 %r<648>; .reg .b64 %rd<150>; // %bb.0: ld.param.b32 %rd3, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_9]; ld.param.b32 %rd72, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_10]; shl.b64 %rd4, %rd72, 32; or.b64 %rd5, %rd4, %rd3; ld.param.b32 %rd76, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_13]; ld.param.b32 %rd77, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_14]; shl.b64 %rd78, %rd77, 32; or.b64 %rd80, %rd78, %rd76; mul.lo.s64 %rd8, %rd5, 28; mul.lo.s64 %rd9, %rd80, %rd8; mov.u32 %r30, %ctaid.x; cvt.u64.u32 %rd133, %r30; setp.le.u64 %p25, %rd9, %rd133; @%p25 bra $L__BB27_138; // %bb.1: // %.lr.ph23 ld.param.b64 %rd70, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_2]; ld.param.b64 %rd69, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_1]; ld.param.b64 %rd68, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_0]; ld.param.b32 %rd1, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_3]; ld.param.b32 %rd71, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_4]; shl.b64 %rd2, %rd71, 32; ld.param.b32 %rd73, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_12]; shl.b64 %rd74, %rd73, 32; ld.param.b32 %rd75, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_11]; or.b64 %rd6, %rd74, %rd75; ld.param.b32 %rd79, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_15]; ld.param.b32 %rd81, [main_graph_async_dispatch_33_softmax_Dx28xDxDxf32_generic_param_16]; shl.b64 %rd82, %rd81, 32; or.b64 %rd7, %rd82, %rd79; mov.u32 %r31, %nctaid.x; cvt.u64.u32 %rd10, %r31; mov.u32 %r32, %tid.x; max.u64 %rd83, %rd7, 1; add.s64 %rd84, %rd83, -1; shr.u64 %rd85, %rd84, 3; and.b64 %rd11, %rd85, 2251799813684224; shl.b32 %r1, %r32, 3; cvt.u64.u32 %rd12, %r1; shr.u32 %r33, %r32, 5; and.b32 %r2, %r32, 31; mul.wide.u32 %rd86, %r33, 4; mov.b64 %rd87, __dynamic_shared_memory__; add.s64 %rd13, %rd87, %rd86; mul.wide.u32 %rd88, %r2, 4; add.s64 %rd14, %rd87, %rd88; shl.b32 %r34, %r33, 8; shl.b32 %r35, %r2, 3; or.b32 %r3, %r35, %r34; add.s64 %rd15, %rd4, %rd3; mul.wide.u32 %rd89, %r32, 16; add.s64 %rd90, %rd2, %rd1; and.b64 %rd91, %rd90, 2305843009213693950; add.s64 %rd92, %rd89, %rd91; add.s64 %rd16, %rd69, %rd92; cvt.u64.u32 %rd93, %r3; add.s64 %rd17, %rd68, %rd93; mul.wide.u32 %rd94, %r3, 2; add.s64 %rd95, %rd94, %rd91; add.s64 %rd18, %rd69, %rd95; mul.wide.u32 %rd96, %r3, 4; add.s64 %rd19, %rd70, %rd96; shl.b64 %rd20, %rd7, 2; shl.b64 %rd21, %rd7, 1; setp.eq.b64 %p32, %rd7, 0; setp.ne.b32 %p58, %r2, 0; prmt.b32 %r396, %r397, %r398, 0x3340U; prmt.b32 %r143, %r144, %r145, 0x3340U; add.s32 %r45, %r1, 7; add.s32 %r46, %r1, 6; add.s32 %r47, %r1, 5; add.s32 %r48, %r1, 4; add.s32 %r49, %r1, 3; add.s32 %r50, %r1, 2; add.s32 %r51, %r1, 1; prmt.b32 %r55, %r56, %r57, 0x3340U; bra.uni $L__BB27_2; $L__BB27_137: // %._crit_edge21 // in Loop: Header=BB27_2 Depth=1 add.s64 %rd133, %rd133, %rd10; setp.lt.s64 %p173, %rd133, %rd9; @!%p173 bra $L__BB27_138; $L__BB27_2: // =>This Loop Header: Depth=1 // Child Loop BB27_13 Depth 2 // Child Loop BB27_50 Depth 2 // Child Loop BB27_88 Depth 2 or.b64 %rd97, %rd133, %rd8; and.b64 %rd98, %rd97, -4294967296; setp.ne.b64 %p26, %rd98, 0; @!%p26 bra $L__BB27_3; // %bb.4: // in Loop: Header=BB27_2 Depth=1 div.s64 %rd134, %rd133, %rd8; bra.uni $L__BB27_5; $L__BB27_3: // in Loop: Header=BB27_2 Depth=1 cvt.u32.u64 %r36, %rd8; cvt.u32.u64 %r37, %rd133; div.u32 %r38, %r37, %r36; cvt.u64.u32 %rd134, %r38; $L__BB27_5: // %.split // in Loop: Header=BB27_2 Depth=1 mul.lo.s64 %rd99, %rd134, %rd8; setp.ne.b64 %p27, %rd133, %rd99; sub.s64 %rd101, %rd133, %rd99; shr.s64 %rd102, %rd101, 63; and.b64 %rd103, %rd102, %rd8; add.s64 %rd23, %rd103, %rd101; or.b64 %rd104, %rd23, %rd5; and.b64 %rd105, %rd104, -4294967296; setp.ne.b64 %p30, %rd105, 0; @!%p30 bra $L__BB27_6; // %bb.7: // in Loop: Header=BB27_2 Depth=1 div.s64 %rd135, %rd23, %rd5; bra.uni $L__BB27_8; $L__BB27_6: // in Loop: Header=BB27_2 Depth=1 cvt.u32.u64 %r39, %rd5; cvt.u32.u64 %r40, %rd23; div.u32 %r41, %r40, %r39; cvt.u64.u32 %rd135, %r41; $L__BB27_8: // %.split.split // in Loop: Header=BB27_2 Depth=1 setp.lt.s64 %p28, %rd133, 0; and.pred %p29, %p28, %p27; selp.b64 %rd100, -1, 0, %p29; add.s64 %rd22, %rd134, %rd100; or.b64 %rd106, %rd133, %rd5; and.b64 %rd107, %rd106, -4294967296; setp.ne.b64 %p31, %rd107, 0; @!%p31 bra $L__BB27_9; // %bb.10: // in Loop: Header=BB27_2 Depth=1 rem.s64 %rd136, %rd133, %rd5; bra.uni $L__BB27_11; $L__BB27_9: // in Loop: Header=BB27_2 Depth=1 cvt.u32.u64 %r42, %rd5; cvt.u32.u64 %r43, %rd133; rem.u32 %r44, %r43, %r42; cvt.u64.u32 %rd136, %r44; $L__BB27_11: // %.split.split.split // in Loop: Header=BB27_2 Depth=1 shr.s64 %rd108, %rd136, 63; and.b64 %rd109, %rd108, %rd5; add.s64 %rd24, %rd109, %rd136; mov.b32 %r633, 0fFFC00000; mad.lo.s64 %rd130, %rd15, %rd22, %rd24; shl.b64 %rd131, %rd135, 1; shl.b64 %rd132, %rd24, 1; @%p32 bra $L__BB27_46; // %bb.12: // %.lr.ph // in Loop: Header=BB27_2 Depth=1 mov.b64 %rd31, 7; mov.b64 %rd30, 6; mov.b64 %rd29, 5; mov.b64 %rd28, 4; mov.b64 %rd27, 3; mov.b64 %rd26, 2; mov.b64 %rd25, 1; cvt.u64.u32 %rd32, %r1; cvt.u64.u32 %rd33, %r51; cvt.u64.u32 %rd34, %r50; cvt.u64.u32 %rd35, %r49; cvt.u64.u32 %rd36, %r48; cvt.u64.u32 %rd37, %r47; cvt.u64.u32 %rd38, %r46; cvt.u64.u32 %rd39, %r45; mad.lo.s64 %rd40, %rd6, %rd130, %rd68; mad.lo.s64 %rd110, %rd22, 56, %rd131; mad.lo.s64 %rd111, %rd15, %rd110, %rd132; mad.lo.s64 %rd138, %rd7, %rd111, %rd16; mov.b32 %r633, 0fFFC00000; mov.b64 %rd140, -1024; mov.b64 %rd137, %rd7; mov.b64 %rd139, %rd12; bra.uni $L__BB27_13; $L__BB27_45: // %else82 // in Loop: Header=BB27_13 Depth=2 add.s64 %rd112, %rd30, %rd139; add.s64 %rd113, %rd31, %rd139; add.s64 %rd114, %rd28, %rd139; add.s64 %rd115, %rd29, %rd139; add.s64 %rd116, %rd25, %rd139; add.s64 %rd117, %rd27, %rd139; add.s64 %rd118, %rd26, %rd139; setp.lt.u64 %p41, %rd118, %rd7; setp.lt.u64 %p42, %rd117, %rd7; setp.lt.u64 %p43, %rd116, %rd7; setp.lt.u64 %p44, %rd139, %rd7; setp.lt.u64 %p45, %rd115, %rd7; setp.lt.u64 %p46, %rd114, %rd7; setp.lt.u64 %p47, %rd113, %rd7; setp.lt.u64 %p48, %rd112, %rd7; prmt.b32 %r65, %r631, 0, 0x7772U; and.b32 %r66, %r65, 1; setp.ne.b32 %p49, %r66, 0; prmt.b32 %r67, %r631, 0, 0x7773U; and.b32 %r68, %r67, 1; setp.ne.b32 %p50, %r68, 0; prmt.b32 %r69, %r631, 0, 0x7771U; and.b32 %r70, %r69, 1; setp.ne.b32 %p51, %r70, 0; and.b32 %r71, %r631, 1; setp.ne.b32 %p52, %r71, 0; prmt.b32 %r72, %r632, 0, 0x7771U; and.b32 %r73, %r72, 1; setp.ne.b32 %p53, %r73, 0; and.b32 %r74, %r632, 1; setp.ne.b32 %p54, %r74, 0; prmt.b32 %r75, %r632, 0, 0x7773U; and.b32 %r76, %r75, 1; setp.ne.b32 %p55, %r76, 0; prmt.b32 %r77, %r632, 0, 0x7772U; and.b32 %r78, %r77, 1; setp.ne.b32 %p56, %r78, 0; selp.f32 %r79, 0f00000000, 0fFF800000, %p56; selp.f32 %r80, 0f00000000, 0fFF800000, %p55; selp.f32 %r81, 0f00000000, 0fFF800000, %p54; selp.f32 %r82, 0f00000000, 0fFF800000, %p53; selp.f32 %r83, 0f00000000, 0fFF800000, %p52; selp.f32 %r84, 0f00000000, 0fFF800000, %p51; selp.f32 %r85, 0f00000000, 0fFF800000, %p50; selp.f32 %r86, 0f00000000, 0fFF800000, %p49; mov.b32 {%rs17, %rs18}, %r630; cvt.f32.f16 %r87, %rs17; cvt.f32.f16 %r88, %rs18; mov.b32 {%rs19, %rs20}, %r629; cvt.f32.f16 %r89, %rs19; cvt.f32.f16 %r90, %rs20; mov.b32 {%rs21, %rs22}, %r627; cvt.f32.f16 %r91, %rs21; cvt.f32.f16 %r92, %rs22; mov.b32 {%rs23, %rs24}, %r628; cvt.f32.f16 %r93, %rs24; cvt.f32.f16 %r94, %rs23; add.rn.f32 %r95, %r86, %r94; add.rn.f32 %r96, %r85, %r93; add.rn.f32 %r97, %r84, %r92; add.rn.f32 %r98, %r83, %r91; add.rn.f32 %r99, %r82, %r90; add.rn.f32 %r100, %r81, %r89; add.rn.f32 %r101, %r80, %r88; add.rn.f32 %r102, %r79, %r87; selp.f32 %r103, %r102, 0fFFC00000, %p48; selp.f32 %r104, %r101, 0fFFC00000, %p47; selp.f32 %r105, %r100, 0fFFC00000, %p46; selp.f32 %r106, %r99, 0fFFC00000, %p45; selp.f32 %r107, %r98, 0fFFC00000, %p44; selp.f32 %r108, %r97, 0fFFC00000, %p43; selp.f32 %r109, %r96, 0fFFC00000, %p42; selp.f32 %r110, %r95, 0fFFC00000, %p41; max.f32 %r111, %r110, %r109; max.f32 %r112, %r107, %r108; max.f32 %r113, %r112, %r111; max.f32 %r114, %r105, %r106; max.f32 %r115, %r103, %r104; max.f32 %r116, %r114, %r115; max.f32 %r117, %r113, %r116; max.f32 %r118, %r117, 0fFF800000; max.f32 %r633, %r118, %r633; add.s64 %rd140, %rd140, 1024; add.s64 %rd139, %rd139, 8192; add.s64 %rd138, %rd138, 16384; add.s64 %rd137, %rd137, -8192; setp.lt.u64 %p57, %rd140, %rd11; @!%p57 bra $L__BB27_46; $L__BB27_13: // Parent Loop BB27_2 Depth=1 // => This Inner Loop Header: Depth=2 min.s64 %rd41, %rd137, 8192; setp.gt.s64 %p1, %rd41, %rd32; // implicit-def: %r627 @!%p1 bra $L__BB27_15; // %bb.14: // %cond.load // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs1, [%rd138]; mov.b32 %r627, {%rs1, %rs2}; $L__BB27_15: // %else // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p2, %rd41, %rd33; @!%p2 bra $L__BB27_17; // %bb.16: // %cond.load39 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs3, [%rd138+2]; mov.b32 {%rs4, _}, %r627; mov.b32 %r627, {%rs4, %rs3}; $L__BB27_17: // %else40 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p3, %rd41, %rd34; // implicit-def: %r628 @!%p3 bra $L__BB27_19; // %bb.18: // %cond.load42 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs5, [%rd138+4]; mov.b32 {_, %rs6}, %r4; mov.b32 %r628, {%rs5, %rs6}; $L__BB27_19: // %else43 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p4, %rd41, %rd35; @!%p4 bra $L__BB27_21; // %bb.20: // %cond.load45 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs7, [%rd138+6]; mov.b32 {%rs8, _}, %r628; mov.b32 %r628, {%rs8, %rs7}; $L__BB27_21: // %else46 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p5, %rd41, %rd36; // implicit-def: %r629 @!%p5 bra $L__BB27_23; // %bb.22: // %cond.load48 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs9, [%rd138+8]; mov.b32 {_, %rs10}, %r5; mov.b32 %r629, {%rs9, %rs10}; $L__BB27_23: // %else49 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p6, %rd41, %rd37; @!%p6 bra $L__BB27_25; // %bb.24: // %cond.load51 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs11, [%rd138+10]; mov.b32 {%rs12, _}, %r629; mov.b32 %r629, {%rs12, %rs11}; $L__BB27_25: // %else52 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p7, %rd41, %rd38; // implicit-def: %r630 @!%p7 bra $L__BB27_27; // %bb.26: // %cond.load54 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs13, [%rd138+12]; mov.b32 {_, %rs14}, %r6; mov.b32 %r630, {%rs13, %rs14}; $L__BB27_27: // %else55 // in Loop: Header=BB27_13 Depth=2 setp.gt.s64 %p8, %rd41, %rd39; @!%p8 bra $L__BB27_29; // %bb.28: // %cond.load57 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b16 %rs15, [%rd138+14]; mov.b32 {%rs16, _}, %r630; mov.b32 %r630, {%rs16, %rs15}; $L__BB27_29: // %else58 // in Loop: Header=BB27_13 Depth=2 add.s64 %rd42, %rd40, %rd139; not.pred %p33, %p1; // implicit-def: %r631 @%p33 bra $L__BB27_31; // %bb.30: // %cond.load60 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r52, [%rd42]; prmt.b32 %r53, %r52, %r54, 0x3340U; prmt.b32 %r631, %r53, %r55, 0x5410U; $L__BB27_31: // %else61 // in Loop: Header=BB27_13 Depth=2 not.pred %p34, %p2; @%p34 bra $L__BB27_33; // %bb.32: // %cond.load63 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r58, [%rd42+1]; bfi.b32 %r631, %r58, %r631, 8, 8; $L__BB27_33: // %else64 // in Loop: Header=BB27_13 Depth=2 not.pred %p35, %p3; @%p35 bra $L__BB27_35; // %bb.34: // %cond.load66 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r59, [%rd42+2]; bfi.b32 %r631, %r59, %r631, 16, 8; $L__BB27_35: // %else67 // in Loop: Header=BB27_13 Depth=2 not.pred %p36, %p4; @%p36 bra $L__BB27_37; // %bb.36: // %cond.load69 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r60, [%rd42+3]; bfi.b32 %r631, %r60, %r631, 24, 8; $L__BB27_37: // %else70 // in Loop: Header=BB27_13 Depth=2 not.pred %p37, %p5; // implicit-def: %r632 @%p37 bra $L__BB27_39; // %bb.38: // %cond.load72 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r61, [%rd42+4]; bfi.b32 %r632, %r61, %r7, 0, 8; $L__BB27_39: // %else73 // in Loop: Header=BB27_13 Depth=2 not.pred %p38, %p6; @%p38 bra $L__BB27_41; // %bb.40: // %cond.load75 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r62, [%rd42+5]; bfi.b32 %r632, %r62, %r632, 8, 8; $L__BB27_41: // %else76 // in Loop: Header=BB27_13 Depth=2 not.pred %p39, %p7; @%p39 bra $L__BB27_43; // %bb.42: // %cond.load78 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r63, [%rd42+6]; bfi.b32 %r632, %r63, %r632, 16, 8; $L__BB27_43: // %else79 // in Loop: Header=BB27_13 Depth=2 not.pred %p40, %p8; @%p40 bra $L__BB27_45; // %bb.44: // %cond.load81 // in Loop: Header=BB27_13 Depth=2 ld.global.nc.b8 %r64, [%rd42+7]; bfi.b32 %r632, %r64, %r632, 24, 8; bra.uni $L__BB27_45; $L__BB27_46: // %._crit_edge // in Loop: Header=BB27_2 Depth=1 max.f32 %r119, %r633, 0fFF800000; shfl.sync.bfly.b32 %r120, %r119, 1, 31, -1; max.f32 %r121, %r119, %r120; shfl.sync.bfly.b32 %r122, %r121, 2, 31, -1; max.f32 %r123, %r121, %r122; shfl.sync.bfly.b32 %r124, %r123, 4, 31, -1; max.f32 %r125, %r123, %r124; shfl.sync.bfly.b32 %r126, %r125, 8, 31, -1; max.f32 %r127, %r125, %r126; shfl.sync.bfly.b32 %r128, %r127, 16, 31, -1; bar.sync 0; @%p58 bra $L__BB27_48; // %bb.47: // in Loop: Header=BB27_2 Depth=1 max.f32 %r8, %r127, %r128; st.shared.b32 [%rd13], %r8; $L__BB27_48: // in Loop: Header=BB27_2 Depth=1 bar.sync 0; ld.shared.b32 %r129, [%rd14]; max.f32 %r130, %r129, 0fFF800000; shfl.sync.bfly.b32 %r131, %r130, 1, 31, -1; max.f32 %r132, %r130, %r131; shfl.sync.bfly.b32 %r133, %r132, 2, 31, -1; max.f32 %r134, %r132, %r133; shfl.sync.bfly.b32 %r135, %r134, 4, 31, -1; max.f32 %r136, %r134, %r135; shfl.sync.bfly.b32 %r137, %r136, 8, 31, -1; max.f32 %r138, %r136, %r137; shfl.sync.bfly.b32 %r139, %r138, 16, 31, -1; max.f32 %r9, %r138, %r139; mov.b32 %r641, 0f00000000; add.s32 %r620, %r3, 7; add.s32 %r621, %r3, 6; add.s32 %r622, %r3, 5; add.s32 %r623, %r3, 4; add.s32 %r624, %r3, 3; add.s32 %r625, %r3, 2; add.s32 %r626, %r3, 1; @%p32 bra $L__BB27_84; // %bb.49: // %.lr.ph16 // in Loop: Header=BB27_2 Depth=1 mov.b64 %rd49, 7; mov.b64 %rd48, 6; mov.b64 %rd47, 5; mov.b64 %rd46, 4; mov.b64 %rd45, 3; mov.b64 %rd44, 2; mov.b64 %rd43, 1; cvt.u64.u32 %rd50, %r626; cvt.u64.u32 %rd51, %r625; cvt.u64.u32 %rd52, %r624; cvt.u64.u32 %rd53, %r623; cvt.u64.u32 %rd54, %r622; cvt.u64.u32 %rd55, %r621; cvt.u64.u32 %rd56, %r620; mad.lo.s64 %rd143, %rd6, %rd130, %rd17; mad.lo.s64 %rd119, %rd22, 56, %rd131; mad.lo.s64 %rd120, %rd15, %rd119, %rd132; mad.lo.s64 %rd142, %rd7, %rd120, %rd18; mov.b32 %r634, 0f00000000; mov.b64 %rd145, -1024; mov.b64 %rd141, %rd7; mov.b64 %rd144, %rd12; bra.uni $L__BB27_50; $L__BB27_82: // %else130 // in Loop: Header=BB27_50 Depth=2 add.s64 %rd121, %rd49, %rd144; add.s64 %rd122, %rd48, %rd144; add.s64 %rd123, %rd47, %rd144; add.s64 %rd124, %rd46, %rd144; add.s64 %rd125, %rd45, %rd144; add.s64 %rd126, %rd44, %rd144; add.s64 %rd127, %rd43, %rd144; setp.lt.u64 %p67, %rd144, %rd7; setp.lt.u64 %p68, %rd127, %rd7; setp.lt.u64 %p69, %rd126, %rd7; setp.lt.u64 %p70, %rd125, %rd7; setp.lt.u64 %p71, %rd124, %rd7; setp.lt.u64 %p72, %rd123, %rd7; setp.lt.u64 %p73, %rd122, %rd7; setp.lt.u64 %p74, %rd121, %rd7; and.b32 %r153, %r639, 1; setp.ne.b32 %p75, %r153, 0; prmt.b32 %r154, %r639, 0, 0x7771U; and.b32 %r155, %r154, 1; setp.ne.b32 %p76, %r155, 0; prmt.b32 %r156, %r639, 0, 0x7772U; and.b32 %r157, %r156, 1; setp.ne.b32 %p77, %r157, 0; prmt.b32 %r158, %r639, 0, 0x7773U; and.b32 %r159, %r158, 1; setp.ne.b32 %p78, %r159, 0; and.b32 %r160, %r640, 1; setp.ne.b32 %p79, %r160, 0; prmt.b32 %r161, %r640, 0, 0x7771U; and.b32 %r162, %r161, 1; setp.ne.b32 %p80, %r162, 0; prmt.b32 %r163, %r640, 0, 0x7772U; and.b32 %r164, %r163, 1; setp.ne.b32 %p81, %r164, 0; prmt.b32 %r165, %r640, 0, 0x7773U; and.b32 %r166, %r165, 1; setp.ne.b32 %p82, %r166, 0; selp.f32 %r167, 0f00000000, 0fFF800000, %p82; selp.f32 %r168, 0f00000000, 0fFF800000, %p81; selp.f32 %r169, 0f00000000, 0fFF800000, %p80; selp.f32 %r170, 0f00000000, 0fFF800000, %p79; selp.f32 %r171, 0f00000000, 0fFF800000, %p78; selp.f32 %r172, 0f00000000, 0fFF800000, %p77; selp.f32 %r173, 0f00000000, 0fFF800000, %p76; selp.f32 %r174, 0f00000000, 0fFF800000, %p75; mov.b32 {%rs41, %rs42}, %r638; cvt.f32.f16 %r175, %rs42; cvt.f32.f16 %r176, %rs41; mov.b32 {%rs43, %rs44}, %r637; cvt.f32.f16 %r177, %rs44; cvt.f32.f16 %r178, %rs43; mov.b32 {%rs45, %rs46}, %r636; cvt.f32.f16 %r179, %rs46; cvt.f32.f16 %r180, %rs45; mov.b32 {%rs47, %rs48}, %r635; cvt.f32.f16 %r181, %rs48; cvt.f32.f16 %r182, %rs47; add.rn.f32 %r183, %r174, %r182; add.rn.f32 %r184, %r173, %r181; add.rn.f32 %r185, %r172, %r180; add.rn.f32 %r186, %r171, %r179; add.rn.f32 %r187, %r170, %r178; add.rn.f32 %r188, %r169, %r177; add.rn.f32 %r189, %r168, %r176; add.rn.f32 %r190, %r167, %r175; sub.rn.f32 %r191, %r190, %r9; sub.rn.f32 %r192, %r189, %r9; sub.rn.f32 %r193, %r188, %r9; sub.rn.f32 %r194, %r187, %r9; sub.rn.f32 %r195, %r186, %r9; sub.rn.f32 %r196, %r185, %r9; sub.rn.f32 %r197, %r184, %r9; sub.rn.f32 %r198, %r183, %r9; setp.lt.f32 %p83, %r198, 0fC2AF999A; setp.lt.f32 %p84, %r197, 0fC2AF999A; setp.lt.f32 %p85, %r196, 0fC2AF999A; setp.lt.f32 %p86, %r195, 0fC2AF999A; setp.lt.f32 %p87, %r194, 0fC2AF999A; setp.lt.f32 %p88, %r193, 0fC2AF999A; setp.lt.f32 %p89, %r192, 0fC2AF999A; setp.lt.f32 %p90, %r191, 0fC2AF999A; selp.f32 %r199, 0fC2AF999A, %r191, %p90; selp.f32 %r200, 0fC2AF999A, %r192, %p89; selp.f32 %r201, 0fC2AF999A, %r193, %p88; selp.f32 %r202, 0fC2AF999A, %r194, %p87; selp.f32 %r203, 0fC2AF999A, %r195, %p86; selp.f32 %r204, 0fC2AF999A, %r196, %p85; selp.f32 %r205, 0fC2AF999A, %r197, %p84; selp.f32 %r206, 0fC2AF999A, %r198, %p83; setp.gt.f32 %p91, %r206, 0f42B1999A; setp.gt.f32 %p92, %r205, 0f42B1999A; setp.gt.f32 %p93, %r204, 0f42B1999A; setp.gt.f32 %p94, %r203, 0f42B1999A; setp.gt.f32 %p95, %r202, 0f42B1999A; setp.gt.f32 %p96, %r201, 0f42B1999A; setp.gt.f32 %p97, %r200, 0f42B1999A; setp.gt.f32 %p98, %r199, 0f42B1999A; selp.f32 %r207, 0f42B1999A, %r199, %p98; selp.f32 %r208, 0f42B1999A, %r200, %p97; selp.f32 %r209, 0f42B1999A, %r201, %p96; selp.f32 %r210, 0f42B1999A, %r202, %p95; selp.f32 %r211, 0f42B1999A, %r203, %p94; selp.f32 %r212, 0f42B1999A, %r204, %p93; selp.f32 %r213, 0f42B1999A, %r205, %p92; selp.f32 %r214, 0f42B1999A, %r206, %p91; fma.rn.f32 %r215, %r214, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r216, %r213, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r217, %r212, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r218, %r211, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r219, %r210, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r220, %r209, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r221, %r208, 0f3FB8AA3B, 0f3F000000; fma.rn.f32 %r222, %r207, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r223, %r215; cvt.rmi.f32.f32 %r224, %r216; cvt.rmi.f32.f32 %r225, %r217; cvt.rmi.f32.f32 %r226, %r218; cvt.rmi.f32.f32 %r227, %r219; cvt.rmi.f32.f32 %r228, %r220; cvt.rmi.f32.f32 %r229, %r221; cvt.rmi.f32.f32 %r230, %r222; setp.lt.f32 %p99, %r223, 0fC2FE0000; setp.lt.f32 %p100, %r224, 0fC2FE0000; setp.lt.f32 %p101, %r225, 0fC2FE0000; setp.lt.f32 %p102, %r226, 0fC2FE0000; setp.lt.f32 %p103, %r227, 0fC2FE0000; setp.lt.f32 %p104, %r228, 0fC2FE0000; setp.lt.f32 %p105, %r229, 0fC2FE0000; setp.lt.f32 %p106, %r230, 0fC2FE0000; selp.f32 %r231, 0fC2FE0000, %r230, %p106; selp.f32 %r232, 0fC2FE0000, %r229, %p105; selp.f32 %r233, 0fC2FE0000, %r228, %p104; selp.f32 %r234, 0fC2FE0000, %r227, %p103; selp.f32 %r235, 0fC2FE0000, %r226, %p102; selp.f32 %r236, 0fC2FE0000, %r225, %p101; selp.f32 %r237, 0fC2FE0000, %r224, %p100; selp.f32 %r238, 0fC2FE0000, %r223, %p99; setp.gt.f32 %p107, %r238, 0f42FE0000; setp.gt.f32 %p108, %r237, 0f42FE0000; setp.gt.f32 %p109, %r236, 0f42FE0000; setp.gt.f32 %p110, %r235, 0f42FE0000; setp.gt.f32 %p111, %r234, 0f42FE0000; setp.gt.f32 %p112, %r233, 0f42FE0000; setp.gt.f32 %p113, %r232, 0f42FE0000; setp.gt.f32 %p114, %r231, 0f42FE0000; selp.f32 %r239, 0f42FE0000, %r231, %p114; selp.f32 %r240, 0f42FE0000, %r232, %p113; selp.f32 %r241, 0f42FE0000, %r233, %p112; selp.f32 %r242, 0f42FE0000, %r234, %p111; selp.f32 %r243, 0f42FE0000, %r235, %p110; selp.f32 %r244, 0f42FE0000, %r236, %p109; selp.f32 %r245, 0f42FE0000, %r237, %p108; selp.f32 %r246, 0f42FE0000, %r238, %p107; fma.rn.f32 %r247, %r246, 0fBF318000, %r214; fma.rn.f32 %r248, %r245, 0fBF318000, %r213; fma.rn.f32 %r249, %r244, 0fBF318000, %r212; fma.rn.f32 %r250, %r243, 0fBF318000, %r211; fma.rn.f32 %r251, %r242, 0fBF318000, %r210; fma.rn.f32 %r252, %r241, 0fBF318000, %r209; fma.rn.f32 %r253, %r240, 0fBF318000, %r208; fma.rn.f32 %r254, %r239, 0fBF318000, %r207; fma.rn.f32 %r255, %r246, 0f395E8083, %r247; fma.rn.f32 %r256, %r245, 0f395E8083, %r248; fma.rn.f32 %r257, %r244, 0f395E8083, %r249; fma.rn.f32 %r258, %r243, 0f395E8083, %r250; fma.rn.f32 %r259, %r242, 0f395E8083, %r251; fma.rn.f32 %r260, %r241, 0f395E8083, %r252; fma.rn.f32 %r261, %r240, 0f395E8083, %r253; fma.rn.f32 %r262, %r239, 0f395E8083, %r254; fma.rn.f32 %r263, %r255, 0f39506967, 0f3AB743CE; fma.rn.f32 %r264, %r256, 0f39506967, 0f3AB743CE; fma.rn.f32 %r265, %r257, 0f39506967, 0f3AB743CE; fma.rn.f32 %r266, %r258, 0f39506967, 0f3AB743CE; fma.rn.f32 %r267, %r259, 0f39506967, 0f3AB743CE; fma.rn.f32 %r268, %r260, 0f39506967, 0f3AB743CE; fma.rn.f32 %r269, %r261, 0f39506967, 0f3AB743CE; fma.rn.f32 %r270, %r262, 0f39506967, 0f3AB743CE; fma.rn.f32 %r271, %r263, %r255, 0f3C088908; fma.rn.f32 %r272, %r264, %r256, 0f3C088908; fma.rn.f32 %r273, %r265, %r257, 0f3C088908; fma.rn.f32 %r274, %r266, %r258, 0f3C088908; fma.rn.f32 %r275, %r267, %r259, 0f3C088908; fma.rn.f32 %r276, %r268, %r260, 0f3C088908; fma.rn.f32 %r277, %r269, %r261, 0f3C088908; fma.rn.f32 %r278, %r270, %r262, 0f3C088908; fma.rn.f32 %r279, %r271, %r255, 0f3D2AA9C1; fma.rn.f32 %r280, %r272, %r256, 0f3D2AA9C1; fma.rn.f32 %r281, %r273, %r257, 0f3D2AA9C1; fma.rn.f32 %r282, %r274, %r258, 0f3D2AA9C1; fma.rn.f32 %r283, %r275, %r259, 0f3D2AA9C1; fma.rn.f32 %r284, %r276, %r260, 0f3D2AA9C1; fma.rn.f32 %r285, %r277, %r261, 0f3D2AA9C1; fma.rn.f32 %r286, %r278, %r262, 0f3D2AA9C1; fma.rn.f32 %r287, %r279, %r255, 0f3E2AAAAA; fma.rn.f32 %r288, %r280, %r256, 0f3E2AAAAA; fma.rn.f32 %r289, %r281, %r257, 0f3E2AAAAA; fma.rn.f32 %r290, %r282, %r258, 0f3E2AAAAA; fma.rn.f32 %r291, %r283, %r259, 0f3E2AAAAA; fma.rn.f32 %r292, %r284, %r260, 0f3E2AAAAA; fma.rn.f32 %r293, %r285, %r261, 0f3E2AAAAA; fma.rn.f32 %r294, %r286, %r262, 0f3E2AAAAA; fma.rn.f32 %r295, %r287, %r255, 0f3F000000; fma.rn.f32 %r296, %r288, %r256, 0f3F000000; fma.rn.f32 %r297, %r289, %r257, 0f3F000000; fma.rn.f32 %r298, %r290, %r258, 0f3F000000; fma.rn.f32 %r299, %r291, %r259, 0f3F000000; fma.rn.f32 %r300, %r292, %r260, 0f3F000000; fma.rn.f32 %r301, %r293, %r261, 0f3F000000; fma.rn.f32 %r302, %r294, %r262, 0f3F000000; mul.rn.f32 %r303, %r262, %r262; mul.rn.f32 %r304, %r261, %r261; mul.rn.f32 %r305, %r260, %r260; mul.rn.f32 %r306, %r259, %r259; mul.rn.f32 %r307, %r258, %r258; mul.rn.f32 %r308, %r257, %r257; mul.rn.f32 %r309, %r256, %r256; mul.rn.f32 %r310, %r255, %r255; fma.rn.f32 %r311, %r295, %r310, %r255; fma.rn.f32 %r312, %r296, %r309, %r256; fma.rn.f32 %r313, %r297, %r308, %r257; fma.rn.f32 %r314, %r298, %r307, %r258; fma.rn.f32 %r315, %r299, %r306, %r259; fma.rn.f32 %r316, %r300, %r305, %r260; fma.rn.f32 %r317, %r301, %r304, %r261; fma.rn.f32 %r318, %r302, %r303, %r262; add.rn.f32 %r319, %r318, 0f3F800000; add.rn.f32 %r320, %r317, 0f3F800000; add.rn.f32 %r321, %r316, 0f3F800000; add.rn.f32 %r322, %r315, 0f3F800000; add.rn.f32 %r323, %r314, 0f3F800000; add.rn.f32 %r324, %r313, 0f3F800000; add.rn.f32 %r325, %r312, 0f3F800000; add.rn.f32 %r326, %r311, 0f3F800000; cvt.rzi.s32.f32 %r327, %r239; cvt.rzi.s32.f32 %r328, %r240; cvt.rzi.s32.f32 %r329, %r241; cvt.rzi.s32.f32 %r330, %r242; cvt.rzi.s32.f32 %r331, %r243; cvt.rzi.s32.f32 %r332, %r244; cvt.rzi.s32.f32 %r333, %r245; cvt.rzi.s32.f32 %r334, %r246; shl.b32 %r335, %r334, 23; shl.b32 %r336, %r333, 23; shl.b32 %r337, %r332, 23; shl.b32 %r338, %r331, 23; shl.b32 %r339, %r330, 23; shl.b32 %r340, %r329, 23; shl.b32 %r341, %r328, 23; shl.b32 %r342, %r327, 23; add.s32 %r343, %r342, 1065353216; add.s32 %r344, %r341, 1065353216; add.s32 %r345, %r340, 1065353216; add.s32 %r346, %r339, 1065353216; add.s32 %r347, %r338, 1065353216; add.s32 %r348, %r337, 1065353216; add.s32 %r349, %r336, 1065353216; add.s32 %r350, %r335, 1065353216; mul.rn.f32 %r351, %r326, %r350; mul.rn.f32 %r352, %r325, %r349; mul.rn.f32 %r353, %r324, %r348; mul.rn.f32 %r354, %r323, %r347; mul.rn.f32 %r355, %r322, %r346; mul.rn.f32 %r356, %r321, %r345; mul.rn.f32 %r357, %r320, %r344; mul.rn.f32 %r358, %r319, %r343; selp.f32 %r359, %r358, 0f00000000, %p74; selp.f32 %r360, %r357, 0f00000000, %p73; selp.f32 %r361, %r356, 0f00000000, %p72; selp.f32 %r362, %r355, 0f00000000, %p71; selp.f32 %r363, %r354, 0f00000000, %p70; selp.f32 %r364, %r353, 0f00000000, %p69; selp.f32 %r365, %r352, 0f00000000, %p68; selp.f32 %r366, %r351, 0f00000000, %p67; add.rn.f32 %r367, %r366, 0f00000000; add.rn.f32 %r368, %r367, %r365; add.rn.f32 %r369, %r368, %r364; add.rn.f32 %r370, %r369, %r363; add.rn.f32 %r371, %r370, %r362; add.rn.f32 %r372, %r371, %r361; add.rn.f32 %r373, %r372, %r360; add.rn.f32 %r374, %r373, %r359; add.rn.f32 %r634, %r634, %r374; add.s64 %rd145, %rd145, 1024; add.s64 %rd144, %rd144, 8192; add.s64 %rd143, %rd143, 8192; add.s64 %rd142, %rd142, 16384; add.s64 %rd141, %rd141, -8192; setp.lt.u64 %p115, %rd145, %rd11; @!%p115 bra $L__BB27_83; $L__BB27_50: // Parent Loop BB27_2 Depth=1 // => This Inner Loop Header: Depth=2 min.s64 %rd57, %rd141, 8192; setp.gt.s64 %p9, %rd57, %rd93; // implicit-def: %r635 @!%p9 bra $L__BB27_52; // %bb.51: // %cond.load84 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs25, [%rd142]; mov.b32 %r635, {%rs25, %rs26}; $L__BB27_52: // %else85 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p10, %rd57, %rd50; @!%p10 bra $L__BB27_54; // %bb.53: // %cond.load87 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs27, [%rd142+2]; mov.b32 {%rs28, _}, %r635; mov.b32 %r635, {%rs28, %rs27}; $L__BB27_54: // %else88 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p11, %rd57, %rd51; // implicit-def: %r636 @!%p11 bra $L__BB27_56; // %bb.55: // %cond.load90 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs29, [%rd142+4]; mov.b32 {_, %rs30}, %r10; mov.b32 %r636, {%rs29, %rs30}; $L__BB27_56: // %else91 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p12, %rd57, %rd52; @!%p12 bra $L__BB27_58; // %bb.57: // %cond.load93 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs31, [%rd142+6]; mov.b32 {%rs32, _}, %r636; mov.b32 %r636, {%rs32, %rs31}; $L__BB27_58: // %else94 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p13, %rd57, %rd53; // implicit-def: %r637 @!%p13 bra $L__BB27_60; // %bb.59: // %cond.load96 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs33, [%rd142+8]; mov.b32 {_, %rs34}, %r11; mov.b32 %r637, {%rs33, %rs34}; $L__BB27_60: // %else97 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p14, %rd57, %rd54; @!%p14 bra $L__BB27_62; // %bb.61: // %cond.load99 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs35, [%rd142+10]; mov.b32 {%rs36, _}, %r637; mov.b32 %r637, {%rs36, %rs35}; $L__BB27_62: // %else100 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p15, %rd57, %rd55; // implicit-def: %r638 @!%p15 bra $L__BB27_64; // %bb.63: // %cond.load102 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs37, [%rd142+12]; mov.b32 {_, %rs38}, %r12; mov.b32 %r638, {%rs37, %rs38}; $L__BB27_64: // %else103 // in Loop: Header=BB27_50 Depth=2 setp.gt.s64 %p16, %rd57, %rd56; @!%p16 bra $L__BB27_66; // %bb.65: // %cond.load105 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b16 %rs39, [%rd142+14]; mov.b32 {%rs40, _}, %r638; mov.b32 %r638, {%rs40, %rs39}; $L__BB27_66: // %else106 // in Loop: Header=BB27_50 Depth=2 not.pred %p59, %p9; // implicit-def: %r639 @%p59 bra $L__BB27_68; // %bb.67: // %cond.load108 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r140, [%rd143]; prmt.b32 %r141, %r140, %r142, 0x3340U; prmt.b32 %r639, %r141, %r143, 0x5410U; $L__BB27_68: // %else109 // in Loop: Header=BB27_50 Depth=2 not.pred %p60, %p10; @%p60 bra $L__BB27_70; // %bb.69: // %cond.load111 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r146, [%rd143+1]; bfi.b32 %r639, %r146, %r639, 8, 8; $L__BB27_70: // %else112 // in Loop: Header=BB27_50 Depth=2 not.pred %p61, %p11; @%p61 bra $L__BB27_72; // %bb.71: // %cond.load114 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r147, [%rd143+2]; bfi.b32 %r639, %r147, %r639, 16, 8; $L__BB27_72: // %else115 // in Loop: Header=BB27_50 Depth=2 not.pred %p62, %p12; @%p62 bra $L__BB27_74; // %bb.73: // %cond.load117 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r148, [%rd143+3]; bfi.b32 %r639, %r148, %r639, 24, 8; $L__BB27_74: // %else118 // in Loop: Header=BB27_50 Depth=2 not.pred %p63, %p13; // implicit-def: %r640 @%p63 bra $L__BB27_76; // %bb.75: // %cond.load120 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r149, [%rd143+4]; bfi.b32 %r640, %r149, %r13, 0, 8; $L__BB27_76: // %else121 // in Loop: Header=BB27_50 Depth=2 not.pred %p64, %p14; @%p64 bra $L__BB27_78; // %bb.77: // %cond.load123 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r150, [%rd143+5]; bfi.b32 %r640, %r150, %r640, 8, 8; $L__BB27_78: // %else124 // in Loop: Header=BB27_50 Depth=2 not.pred %p65, %p15; @%p65 bra $L__BB27_80; // %bb.79: // %cond.load126 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r151, [%rd143+6]; bfi.b32 %r640, %r151, %r640, 16, 8; $L__BB27_80: // %else127 // in Loop: Header=BB27_50 Depth=2 not.pred %p66, %p16; @%p66 bra $L__BB27_82; // %bb.81: // %cond.load129 // in Loop: Header=BB27_50 Depth=2 ld.global.nc.b8 %r152, [%rd143+7]; bfi.b32 %r640, %r152, %r640, 24, 8; bra.uni $L__BB27_82; $L__BB27_83: // %._crit_edge17.loopexit // in Loop: Header=BB27_2 Depth=1 add.rn.f32 %r641, %r634, 0f00000000; $L__BB27_84: // %._crit_edge17 // in Loop: Header=BB27_2 Depth=1 shfl.sync.bfly.b32 %r375, %r641, 1, 31, -1; add.rn.f32 %r376, %r641, %r375; shfl.sync.bfly.b32 %r377, %r376, 2, 31, -1; add.rn.f32 %r378, %r376, %r377; shfl.sync.bfly.b32 %r379, %r378, 4, 31, -1; add.rn.f32 %r380, %r378, %r379; shfl.sync.bfly.b32 %r381, %r380, 8, 31, -1; add.rn.f32 %r382, %r380, %r381; shfl.sync.bfly.b32 %r383, %r382, 16, 31, -1; bar.sync 0; @%p58 bra $L__BB27_86; // %bb.85: // in Loop: Header=BB27_2 Depth=1 add.rn.f32 %r14, %r382, %r383; st.shared.b32 [%rd13+136], %r14; $L__BB27_86: // in Loop: Header=BB27_2 Depth=1 bar.sync 0; ld.shared.b32 %r384, [%rd14+136]; add.rn.f32 %r385, %r384, 0f00000000; shfl.sync.bfly.b32 %r386, %r385, 1, 31, -1; add.rn.f32 %r387, %r386, %r385; shfl.sync.bfly.b32 %r388, %r387, 2, 31, -1; add.rn.f32 %r389, %r388, %r387; shfl.sync.bfly.b32 %r390, %r389, 4, 31, -1; add.rn.f32 %r391, %r390, %r389; shfl.sync.bfly.b32 %r392, %r391, 8, 31, -1; add.rn.f32 %r15, %r392, %r391; shfl.sync.bfly.b32 %r16, %r15, 16, 31, -1; @%p32 bra $L__BB27_137; // %bb.87: // %.lr.ph20 // in Loop: Header=BB27_2 Depth=1 add.rn.f32 %r17, %r16, %r15; cvt.u64.u32 %rd58, %r626; cvt.u64.u32 %rd59, %r625; cvt.u64.u32 %rd60, %r624; cvt.u64.u32 %rd61, %r623; cvt.u64.u32 %rd62, %r622; cvt.u64.u32 %rd63, %r621; cvt.u64.u32 %rd64, %r620; mad.lo.s64 %rd65, %rd6, %rd130, %rd17; mad.lo.s64 %rd128, %rd22, 28, %rd135; mad.lo.s64 %rd129, %rd15, %rd128, %rd24; mad.lo.s64 %rd148, %rd20, %rd129, %rd19; mad.lo.s64 %rd147, %rd21, %rd129, %rd18; mov.b64 %rd149, 0; mov.b64 %rd146, %rd7; bra.uni $L__BB27_88; $L__BB27_136: // %else194 // in Loop: Header=BB27_88 Depth=2 add.s64 %rd149, %rd149, 8192; add.s64 %rd148, %rd148, 32768; add.s64 %rd147, %rd147, 16384; add.s64 %rd146, %rd146, -8192; setp.lt.u64 %p172, %rd149, %rd7; @!%p172 bra $L__BB27_137; $L__BB27_88: // Parent Loop BB27_2 Depth=1 // => This Inner Loop Header: Depth=2 min.s64 %rd66, %rd146, 8192; setp.gt.u64 %p17, %rd66, %rd93; // implicit-def: %r642 @!%p17 bra $L__BB27_90; // %bb.89: // %cond.load132 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs49, [%rd147]; mov.b32 %r642, {%rs49, %rs50}; $L__BB27_90: // %else133 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p18, %rd66, %rd58; @!%p18 bra $L__BB27_92; // %bb.91: // %cond.load135 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs51, [%rd147+2]; mov.b32 {%rs52, _}, %r642; mov.b32 %r642, {%rs52, %rs51}; $L__BB27_92: // %else136 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p19, %rd66, %rd59; // implicit-def: %r643 @!%p19 bra $L__BB27_94; // %bb.93: // %cond.load138 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs53, [%rd147+4]; mov.b32 {_, %rs54}, %r18; mov.b32 %r643, {%rs53, %rs54}; $L__BB27_94: // %else139 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p20, %rd66, %rd60; @!%p20 bra $L__BB27_96; // %bb.95: // %cond.load141 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs55, [%rd147+6]; mov.b32 {%rs56, _}, %r643; mov.b32 %r643, {%rs56, %rs55}; $L__BB27_96: // %else142 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p21, %rd66, %rd61; // implicit-def: %r644 @!%p21 bra $L__BB27_98; // %bb.97: // %cond.load144 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs57, [%rd147+8]; mov.b32 {_, %rs58}, %r19; mov.b32 %r644, {%rs57, %rs58}; $L__BB27_98: // %else145 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p22, %rd66, %rd62; @!%p22 bra $L__BB27_100; // %bb.99: // %cond.load147 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs59, [%rd147+10]; mov.b32 {%rs60, _}, %r644; mov.b32 %r644, {%rs60, %rs59}; $L__BB27_100: // %else148 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p23, %rd66, %rd63; // implicit-def: %r645 @!%p23 bra $L__BB27_102; // %bb.101: // %cond.load150 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs61, [%rd147+12]; mov.b32 {_, %rs62}, %r20; mov.b32 %r645, {%rs61, %rs62}; $L__BB27_102: // %else151 // in Loop: Header=BB27_88 Depth=2 setp.gt.u64 %p24, %rd66, %rd64; @!%p24 bra $L__BB27_104; // %bb.103: // %cond.load153 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b16 %rs63, [%rd147+14]; mov.b32 {%rs64, _}, %r645; mov.b32 %r645, {%rs64, %rs63}; $L__BB27_104: // %else154 // in Loop: Header=BB27_88 Depth=2 add.s64 %rd67, %rd65, %rd149; not.pred %p116, %p17; // implicit-def: %r646 @%p116 bra $L__BB27_106; // %bb.105: // %cond.load156 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r393, [%rd67]; prmt.b32 %r394, %r393, %r395, 0x3340U; prmt.b32 %r646, %r394, %r396, 0x5410U; $L__BB27_106: // %else157 // in Loop: Header=BB27_88 Depth=2 not.pred %p117, %p18; @%p117 bra $L__BB27_108; // %bb.107: // %cond.load159 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r399, [%rd67+1]; bfi.b32 %r646, %r399, %r646, 8, 8; $L__BB27_108: // %else160 // in Loop: Header=BB27_88 Depth=2 not.pred %p118, %p19; @%p118 bra $L__BB27_110; // %bb.109: // %cond.load162 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r400, [%rd67+2]; bfi.b32 %r646, %r400, %r646, 16, 8; $L__BB27_110: // %else163 // in Loop: Header=BB27_88 Depth=2 not.pred %p119, %p20; @%p119 bra $L__BB27_112; // %bb.111: // %cond.load165 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r401, [%rd67+3]; bfi.b32 %r646, %r401, %r646, 24, 8; $L__BB27_112: // %else166 // in Loop: Header=BB27_88 Depth=2 not.pred %p120, %p21; // implicit-def: %r647 @%p120 bra $L__BB27_114; // %bb.113: // %cond.load168 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r402, [%rd67+4]; bfi.b32 %r647, %r402, %r21, 0, 8; $L__BB27_114: // %else169 // in Loop: Header=BB27_88 Depth=2 not.pred %p121, %p22; @%p121 bra $L__BB27_116; // %bb.115: // %cond.load171 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r403, [%rd67+5]; bfi.b32 %r647, %r403, %r647, 8, 8; $L__BB27_116: // %else172 // in Loop: Header=BB27_88 Depth=2 not.pred %p122, %p23; @%p122 bra $L__BB27_118; // %bb.117: // %cond.load174 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r404, [%rd67+6]; bfi.b32 %r647, %r404, %r647, 16, 8; $L__BB27_118: // %else175 // in Loop: Header=BB27_88 Depth=2 not.pred %p123, %p24; @%p123 bra $L__BB27_120; // %bb.119: // %cond.load177 // in Loop: Header=BB27_88 Depth=2 ld.global.nc.b8 %r405, [%rd67+7]; bfi.b32 %r647, %r405, %r647, 24, 8; $L__BB27_120: // %else178 // in Loop: Header=BB27_88 Depth=2 mov.b32 {%rs71, %rs72}, %r642; @%p116 bra $L__BB27_122; // %bb.121: // %cond.store // in Loop: Header=BB27_88 Depth=2 and.b32 %r406, %r646, 1; setp.ne.b32 %p124, %r406, 0; selp.f32 %r427, 0f00000000, 0fFF800000, %p124; cvt.f32.f16 %r435, %rs71; add.rn.f32 %r436, %r427, %r435; sub.rn.f32 %r451, %r436, %r9; setp.lt.f32 %p132, %r451, 0fC2AF999A; selp.f32 %r459, 0fC2AF999A, %r451, %p132; setp.gt.f32 %p140, %r459, 0f42B1999A; selp.f32 %r467, 0f42B1999A, %r459, %p140; fma.rn.f32 %r468, %r467, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r476, %r468; setp.lt.f32 %p148, %r476, 0fC2FE0000; selp.f32 %r491, 0fC2FE0000, %r476, %p148; setp.gt.f32 %p156, %r491, 0f42FE0000; selp.f32 %r499, 0f42FE0000, %r491, %p156; fma.rn.f32 %r500, %r499, 0fBF318000, %r467; fma.rn.f32 %r508, %r499, 0f395E8083, %r500; fma.rn.f32 %r516, %r508, 0f39506967, 0f3AB743CE; fma.rn.f32 %r524, %r516, %r508, 0f3C088908; fma.rn.f32 %r532, %r524, %r508, 0f3D2AA9C1; fma.rn.f32 %r540, %r532, %r508, 0f3E2AAAAA; fma.rn.f32 %r548, %r540, %r508, 0f3F000000; mul.rn.f32 %r563, %r508, %r508; fma.rn.f32 %r564, %r548, %r563, %r508; add.rn.f32 %r572, %r564, 0f3F800000; cvt.rzi.s32.f32 %r580, %r499; shl.b32 %r595, %r580, 23; add.s32 %r596, %r595, 1065353216; mul.rn.f32 %r611, %r572, %r596; div.rn.f32 %r612, %r611, %r17; setp.num.f32 %p171, %r612, %r612; selp.f32 %r22, %r612, 0f00000000, %p171; st.global.b32 [%rd148], %r22; $L__BB27_122: // %else180 // in Loop: Header=BB27_88 Depth=2 @%p117 bra $L__BB27_124; // %bb.123: // %cond.store181 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r407, %r646, 0, 0x7771U; and.b32 %r408, %r407, 1; setp.ne.b32 %p125, %r408, 0; selp.f32 %r426, 0f00000000, 0fFF800000, %p125; cvt.f32.f16 %r434, %rs72; add.rn.f32 %r437, %r426, %r434; sub.rn.f32 %r450, %r437, %r9; setp.lt.f32 %p133, %r450, 0fC2AF999A; selp.f32 %r458, 0fC2AF999A, %r450, %p133; setp.gt.f32 %p141, %r458, 0f42B1999A; selp.f32 %r466, 0f42B1999A, %r458, %p141; fma.rn.f32 %r469, %r466, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r477, %r469; setp.lt.f32 %p149, %r477, 0fC2FE0000; selp.f32 %r490, 0fC2FE0000, %r477, %p149; setp.gt.f32 %p157, %r490, 0f42FE0000; selp.f32 %r498, 0f42FE0000, %r490, %p157; fma.rn.f32 %r501, %r498, 0fBF318000, %r466; fma.rn.f32 %r509, %r498, 0f395E8083, %r501; fma.rn.f32 %r517, %r509, 0f39506967, 0f3AB743CE; fma.rn.f32 %r525, %r517, %r509, 0f3C088908; fma.rn.f32 %r533, %r525, %r509, 0f3D2AA9C1; fma.rn.f32 %r541, %r533, %r509, 0f3E2AAAAA; fma.rn.f32 %r549, %r541, %r509, 0f3F000000; mul.rn.f32 %r562, %r509, %r509; cvt.rzi.s32.f32 %r584, %r498; fma.rn.f32 %r565, %r549, %r562, %r509; shl.b32 %r591, %r584, 23; add.rn.f32 %r576, %r565, 0f3F800000; add.s32 %r600, %r591, 1065353216; mul.rn.f32 %r607, %r576, %r600; div.rn.f32 %r616, %r607, %r17; setp.num.f32 %p167, %r616, %r616; selp.f32 %r23, %r616, 0f00000000, %p167; st.global.b32 [%rd148+4], %r23; $L__BB27_124: // %else182 // in Loop: Header=BB27_88 Depth=2 mov.b32 {%rs69, %rs70}, %r643; @%p118 bra $L__BB27_126; // %bb.125: // %cond.store183 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r409, %r646, 0, 0x7772U; and.b32 %r410, %r409, 1; setp.ne.b32 %p126, %r410, 0; selp.f32 %r425, 0f00000000, 0fFF800000, %p126; cvt.f32.f16 %r433, %rs69; add.rn.f32 %r438, %r425, %r433; sub.rn.f32 %r449, %r438, %r9; setp.lt.f32 %p134, %r449, 0fC2AF999A; selp.f32 %r457, 0fC2AF999A, %r449, %p134; setp.gt.f32 %p142, %r457, 0f42B1999A; selp.f32 %r465, 0f42B1999A, %r457, %p142; fma.rn.f32 %r470, %r465, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r478, %r470; setp.lt.f32 %p150, %r478, 0fC2FE0000; selp.f32 %r489, 0fC2FE0000, %r478, %p150; setp.gt.f32 %p158, %r489, 0f42FE0000; selp.f32 %r497, 0f42FE0000, %r489, %p158; fma.rn.f32 %r502, %r497, 0fBF318000, %r465; fma.rn.f32 %r510, %r497, 0f395E8083, %r502; fma.rn.f32 %r518, %r510, 0f39506967, 0f3AB743CE; fma.rn.f32 %r526, %r518, %r510, 0f3C088908; fma.rn.f32 %r534, %r526, %r510, 0f3D2AA9C1; fma.rn.f32 %r542, %r534, %r510, 0f3E2AAAAA; fma.rn.f32 %r550, %r542, %r510, 0f3F000000; mul.rn.f32 %r561, %r510, %r510; cvt.rzi.s32.f32 %r583, %r497; fma.rn.f32 %r566, %r550, %r561, %r510; shl.b32 %r592, %r583, 23; add.rn.f32 %r575, %r566, 0f3F800000; add.s32 %r599, %r592, 1065353216; mul.rn.f32 %r608, %r575, %r599; div.rn.f32 %r615, %r608, %r17; setp.num.f32 %p168, %r615, %r615; selp.f32 %r24, %r615, 0f00000000, %p168; st.global.b32 [%rd148+8], %r24; $L__BB27_126: // %else184 // in Loop: Header=BB27_88 Depth=2 @%p119 bra $L__BB27_128; // %bb.127: // %cond.store185 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r411, %r646, 0, 0x7773U; and.b32 %r412, %r411, 1; setp.ne.b32 %p127, %r412, 0; selp.f32 %r424, 0f00000000, 0fFF800000, %p127; cvt.f32.f16 %r432, %rs70; add.rn.f32 %r439, %r424, %r432; sub.rn.f32 %r448, %r439, %r9; setp.lt.f32 %p135, %r448, 0fC2AF999A; selp.f32 %r456, 0fC2AF999A, %r448, %p135; setp.gt.f32 %p143, %r456, 0f42B1999A; selp.f32 %r464, 0f42B1999A, %r456, %p143; fma.rn.f32 %r471, %r464, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r479, %r471; setp.lt.f32 %p151, %r479, 0fC2FE0000; selp.f32 %r488, 0fC2FE0000, %r479, %p151; setp.gt.f32 %p159, %r488, 0f42FE0000; selp.f32 %r496, 0f42FE0000, %r488, %p159; fma.rn.f32 %r503, %r496, 0fBF318000, %r464; fma.rn.f32 %r511, %r496, 0f395E8083, %r503; fma.rn.f32 %r519, %r511, 0f39506967, 0f3AB743CE; fma.rn.f32 %r527, %r519, %r511, 0f3C088908; fma.rn.f32 %r535, %r527, %r511, 0f3D2AA9C1; fma.rn.f32 %r543, %r535, %r511, 0f3E2AAAAA; fma.rn.f32 %r551, %r543, %r511, 0f3F000000; mul.rn.f32 %r560, %r511, %r511; cvt.rzi.s32.f32 %r582, %r496; fma.rn.f32 %r567, %r551, %r560, %r511; shl.b32 %r593, %r582, 23; add.rn.f32 %r574, %r567, 0f3F800000; add.s32 %r598, %r593, 1065353216; mul.rn.f32 %r609, %r574, %r598; div.rn.f32 %r614, %r609, %r17; setp.num.f32 %p169, %r614, %r614; selp.f32 %r25, %r614, 0f00000000, %p169; st.global.b32 [%rd148+12], %r25; $L__BB27_128: // %else186 // in Loop: Header=BB27_88 Depth=2 mov.b32 {%rs67, %rs68}, %r644; @%p120 bra $L__BB27_130; // %bb.129: // %cond.store187 // in Loop: Header=BB27_88 Depth=2 and.b32 %r413, %r647, 1; setp.ne.b32 %p128, %r413, 0; selp.f32 %r423, 0f00000000, 0fFF800000, %p128; cvt.f32.f16 %r431, %rs67; add.rn.f32 %r440, %r423, %r431; sub.rn.f32 %r447, %r440, %r9; setp.lt.f32 %p136, %r447, 0fC2AF999A; selp.f32 %r455, 0fC2AF999A, %r447, %p136; setp.gt.f32 %p144, %r455, 0f42B1999A; selp.f32 %r463, 0f42B1999A, %r455, %p144; fma.rn.f32 %r472, %r463, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r480, %r472; setp.lt.f32 %p152, %r480, 0fC2FE0000; selp.f32 %r487, 0fC2FE0000, %r480, %p152; setp.gt.f32 %p160, %r487, 0f42FE0000; selp.f32 %r495, 0f42FE0000, %r487, %p160; fma.rn.f32 %r504, %r495, 0fBF318000, %r463; fma.rn.f32 %r512, %r495, 0f395E8083, %r504; fma.rn.f32 %r520, %r512, 0f39506967, 0f3AB743CE; fma.rn.f32 %r528, %r520, %r512, 0f3C088908; fma.rn.f32 %r536, %r528, %r512, 0f3D2AA9C1; fma.rn.f32 %r544, %r536, %r512, 0f3E2AAAAA; fma.rn.f32 %r552, %r544, %r512, 0f3F000000; mul.rn.f32 %r559, %r512, %r512; cvt.rzi.s32.f32 %r581, %r495; fma.rn.f32 %r568, %r552, %r559, %r512; shl.b32 %r594, %r581, 23; add.rn.f32 %r573, %r568, 0f3F800000; add.s32 %r597, %r594, 1065353216; mul.rn.f32 %r610, %r573, %r597; div.rn.f32 %r613, %r610, %r17; setp.num.f32 %p170, %r613, %r613; selp.f32 %r26, %r613, 0f00000000, %p170; st.global.b32 [%rd148+16], %r26; $L__BB27_130: // %else188 // in Loop: Header=BB27_88 Depth=2 @%p121 bra $L__BB27_132; // %bb.131: // %cond.store189 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r414, %r647, 0, 0x7771U; and.b32 %r415, %r414, 1; setp.ne.b32 %p129, %r415, 0; selp.f32 %r422, 0f00000000, 0fFF800000, %p129; cvt.f32.f16 %r430, %rs68; add.rn.f32 %r441, %r422, %r430; sub.rn.f32 %r446, %r441, %r9; setp.lt.f32 %p137, %r446, 0fC2AF999A; selp.f32 %r454, 0fC2AF999A, %r446, %p137; setp.gt.f32 %p145, %r454, 0f42B1999A; selp.f32 %r462, 0f42B1999A, %r454, %p145; fma.rn.f32 %r473, %r462, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r481, %r473; setp.lt.f32 %p153, %r481, 0fC2FE0000; selp.f32 %r486, 0fC2FE0000, %r481, %p153; setp.gt.f32 %p161, %r486, 0f42FE0000; selp.f32 %r494, 0f42FE0000, %r486, %p161; fma.rn.f32 %r505, %r494, 0fBF318000, %r462; fma.rn.f32 %r513, %r494, 0f395E8083, %r505; fma.rn.f32 %r521, %r513, 0f39506967, 0f3AB743CE; fma.rn.f32 %r529, %r521, %r513, 0f3C088908; fma.rn.f32 %r537, %r529, %r513, 0f3D2AA9C1; fma.rn.f32 %r545, %r537, %r513, 0f3E2AAAAA; fma.rn.f32 %r553, %r545, %r513, 0f3F000000; mul.rn.f32 %r558, %r513, %r513; cvt.rzi.s32.f32 %r586, %r494; fma.rn.f32 %r569, %r553, %r558, %r513; shl.b32 %r589, %r586, 23; add.rn.f32 %r578, %r569, 0f3F800000; add.s32 %r602, %r589, 1065353216; mul.rn.f32 %r605, %r578, %r602; div.rn.f32 %r618, %r605, %r17; setp.num.f32 %p165, %r618, %r618; selp.f32 %r27, %r618, 0f00000000, %p165; st.global.b32 [%rd148+20], %r27; $L__BB27_132: // %else190 // in Loop: Header=BB27_88 Depth=2 mov.b32 {%rs65, %rs66}, %r645; @%p122 bra $L__BB27_134; // %bb.133: // %cond.store191 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r416, %r647, 0, 0x7772U; and.b32 %r417, %r416, 1; setp.ne.b32 %p130, %r417, 0; selp.f32 %r421, 0f00000000, 0fFF800000, %p130; cvt.f32.f16 %r429, %rs65; add.rn.f32 %r442, %r421, %r429; sub.rn.f32 %r445, %r442, %r9; setp.lt.f32 %p138, %r445, 0fC2AF999A; selp.f32 %r453, 0fC2AF999A, %r445, %p138; setp.gt.f32 %p146, %r453, 0f42B1999A; selp.f32 %r461, 0f42B1999A, %r453, %p146; fma.rn.f32 %r474, %r461, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r482, %r474; setp.lt.f32 %p154, %r482, 0fC2FE0000; selp.f32 %r485, 0fC2FE0000, %r482, %p154; setp.gt.f32 %p162, %r485, 0f42FE0000; selp.f32 %r493, 0f42FE0000, %r485, %p162; fma.rn.f32 %r506, %r493, 0fBF318000, %r461; fma.rn.f32 %r514, %r493, 0f395E8083, %r506; fma.rn.f32 %r522, %r514, 0f39506967, 0f3AB743CE; fma.rn.f32 %r530, %r522, %r514, 0f3C088908; fma.rn.f32 %r538, %r530, %r514, 0f3D2AA9C1; fma.rn.f32 %r546, %r538, %r514, 0f3E2AAAAA; fma.rn.f32 %r554, %r546, %r514, 0f3F000000; mul.rn.f32 %r557, %r514, %r514; cvt.rzi.s32.f32 %r585, %r493; fma.rn.f32 %r570, %r554, %r557, %r514; shl.b32 %r590, %r585, 23; add.rn.f32 %r577, %r570, 0f3F800000; add.s32 %r601, %r590, 1065353216; mul.rn.f32 %r606, %r577, %r601; div.rn.f32 %r617, %r606, %r17; setp.num.f32 %p166, %r617, %r617; selp.f32 %r28, %r617, 0f00000000, %p166; st.global.b32 [%rd148+24], %r28; $L__BB27_134: // %else192 // in Loop: Header=BB27_88 Depth=2 @%p123 bra $L__BB27_136; // %bb.135: // %cond.store193 // in Loop: Header=BB27_88 Depth=2 prmt.b32 %r418, %r647, 0, 0x7773U; and.b32 %r419, %r418, 1; setp.ne.b32 %p131, %r419, 0; selp.f32 %r420, 0f00000000, 0fFF800000, %p131; cvt.f32.f16 %r428, %rs66; add.rn.f32 %r443, %r420, %r428; sub.rn.f32 %r444, %r443, %r9; setp.lt.f32 %p139, %r444, 0fC2AF999A; selp.f32 %r452, 0fC2AF999A, %r444, %p139; setp.gt.f32 %p147, %r452, 0f42B1999A; selp.f32 %r460, 0f42B1999A, %r452, %p147; fma.rn.f32 %r475, %r460, 0f3FB8AA3B, 0f3F000000; cvt.rmi.f32.f32 %r483, %r475; setp.lt.f32 %p155, %r483, 0fC2FE0000; selp.f32 %r484, 0fC2FE0000, %r483, %p155; setp.gt.f32 %p163, %r484, 0f42FE0000; selp.f32 %r492, 0f42FE0000, %r484, %p163; fma.rn.f32 %r507, %r492, 0fBF318000, %r460; fma.rn.f32 %r515, %r492, 0f395E8083, %r507; fma.rn.f32 %r523, %r515, 0f39506967, 0f3AB743CE; fma.rn.f32 %r531, %r523, %r515, 0f3C088908; fma.rn.f32 %r539, %r531, %r515, 0f3D2AA9C1; fma.rn.f32 %r547, %r539, %r515, 0f3E2AAAAA; fma.rn.f32 %r555, %r547, %r515, 0f3F000000; mul.rn.f32 %r556, %r515, %r515; cvt.rzi.s32.f32 %r587, %r492; fma.rn.f32 %r571, %r555, %r556, %r515; shl.b32 %r588, %r587, 23; add.rn.f32 %r579, %r571, 0f3F800000; add.s32 %r603, %r588, 1065353216; mul.rn.f32 %r604, %r579, %r603; div.rn.f32 %r619, %r604, %r17; setp.num.f32 %p164, %r619, %r619; selp.f32 %r29, %r619, 0f00000000, %p164; st.global.b32 [%rd148+28], %r29; bra.uni $L__BB27_136; $L__BB27_138: // %._crit_edge24 ret; // -- End function }