630 lines
19 KiB
Plaintext
630 lines
19 KiB
Plaintext
|
//
|
||
|
|
// Generated by NVIDIA NVVM Compiler
|
||
|
|
//
|
||
|
|
// Compiler Build ID: CL-23083092
|
||
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
||
|
|
// Based on LLVM 3.4svn
|
||
|
|
//
|
||
|
|
|
||
|
|
.version 6.1
|
||
|
|
.target sm_30
|
||
|
|
.address_size 64
|
||
|
|
|
||
|
|
// .globl _Z6oxMainv
|
||
|
|
.global .align 8 .b8 pixelID[8];
|
||
|
|
.global .align 8 .b8 resolution[8];
|
||
|
|
.global .align 4 .b8 normal[12];
|
||
|
|
.global .align 4 .b8 camPos[12];
|
||
|
|
.global .align 4 .b8 root[4];
|
||
|
|
.global .align 4 .u32 imageEnabled;
|
||
|
|
.global .texref lightmap;
|
||
|
|
.global .align 16 .b8 tileInfo[16];
|
||
|
|
.global .align 4 .u32 additive;
|
||
|
|
.global .align 4 .b8 payload[4];
|
||
|
|
.global .align 4 .b8 ray[36];
|
||
|
|
.global .align 4 .f32 t;
|
||
|
|
.global .align 1 .b8 rnd_seeds[1];
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7payloadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo3rayE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo1tE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
||
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7payloadE[8] = {82, 97, 121, 68, 97, 116, 97, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename3rayE[4] = {82, 97, 121, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename1tE[6] = {102, 108, 111, 97, 116, 0};
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7payloadE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum3rayE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum1tE = 4919;
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7payloadE[10] = {114, 116, 80, 97, 121, 108, 111, 97, 100, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic3rayE[13] = {114, 116, 67, 117, 114, 114, 101, 110, 116, 82, 97, 121, 0};
|
||
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic1tE[23] = {114, 116, 73, 110, 116, 101, 114, 115, 101, 99, 116, 105, 111, 110, 68, 105, 115, 116, 97, 110, 99, 101, 0};
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7payloadE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation3rayE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation1tE[1];
|
||
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
||
|
|
|
||
|
|
.visible .entry _Z6oxMainv(
|
||
|
|
|
||
|
|
)
|
||
|
|
{
|
||
|
|
.local .align 4 .b8 __local_depot0[32];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<31>;
|
||
|
|
.reg .f32 %f<192>;
|
||
|
|
.reg .b32 %r<212>;
|
||
|
|
.reg .b64 %rd<37>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd36, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd36;
|
||
|
|
ld.global.f32 %f62, [normal];
|
||
|
|
ld.global.f32 %f63, [normal+4];
|
||
|
|
ld.global.f32 %f64, [normal+8];
|
||
|
|
mov.u32 %r80, 7937;
|
||
|
|
mov.f32 %f191, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
call (%f58, %f59, %f60, %f61), _rt_transform_tuple, (%r80, %f62, %f63, %f64, %f191);
|
||
|
|
// inline asm
|
||
|
|
mul.f32 %f67, %f59, %f59;
|
||
|
|
fma.rn.f32 %f68, %f58, %f58, %f67;
|
||
|
|
fma.rn.f32 %f69, %f60, %f60, %f68;
|
||
|
|
sqrt.rn.f32 %f70, %f69;
|
||
|
|
rcp.rn.f32 %f71, %f70;
|
||
|
|
mul.f32 %f1, %f58, %f71;
|
||
|
|
mul.f32 %f2, %f59, %f71;
|
||
|
|
mul.f32 %f3, %f60, %f71;
|
||
|
|
ld.global.f32 %f72, [ray+12];
|
||
|
|
ld.global.f32 %f73, [t];
|
||
|
|
ld.global.f32 %f74, [ray+16];
|
||
|
|
ld.global.f32 %f75, [ray+20];
|
||
|
|
ld.global.f32 %f76, [ray];
|
||
|
|
fma.rn.f32 %f4, %f73, %f72, %f76;
|
||
|
|
ld.global.f32 %f77, [ray+4];
|
||
|
|
fma.rn.f32 %f5, %f73, %f74, %f77;
|
||
|
|
ld.global.f32 %f78, [ray+8];
|
||
|
|
fma.rn.f32 %f6, %f73, %f75, %f78;
|
||
|
|
abs.f32 %f79, %f1;
|
||
|
|
abs.f32 %f80, %f3;
|
||
|
|
setp.gt.f32 %p1, %f79, %f80;
|
||
|
|
neg.f32 %f81, %f2;
|
||
|
|
neg.f32 %f82, %f3;
|
||
|
|
selp.f32 %f83, %f81, 0f00000000, %p1;
|
||
|
|
selp.f32 %f84, %f1, %f82, %p1;
|
||
|
|
selp.f32 %f85, 0f00000000, %f2, %p1;
|
||
|
|
mul.f32 %f86, %f84, %f84;
|
||
|
|
fma.rn.f32 %f87, %f83, %f83, %f86;
|
||
|
|
fma.rn.f32 %f88, %f85, %f85, %f87;
|
||
|
|
sqrt.rn.f32 %f89, %f88;
|
||
|
|
rcp.rn.f32 %f90, %f89;
|
||
|
|
mul.f32 %f7, %f83, %f90;
|
||
|
|
mul.f32 %f8, %f84, %f90;
|
||
|
|
mul.f32 %f9, %f85, %f90;
|
||
|
|
mul.f32 %f91, %f3, %f8;
|
||
|
|
mul.f32 %f92, %f2, %f9;
|
||
|
|
sub.f32 %f10, %f91, %f92;
|
||
|
|
mul.f32 %f93, %f1, %f9;
|
||
|
|
mul.f32 %f94, %f3, %f7;
|
||
|
|
sub.f32 %f11, %f93, %f94;
|
||
|
|
mul.f32 %f95, %f2, %f7;
|
||
|
|
mul.f32 %f96, %f1, %f8;
|
||
|
|
sub.f32 %f12, %f95, %f96;
|
||
|
|
ld.global.v2.u32 {%r84, %r85}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd15, %r84;
|
||
|
|
cvt.u64.u32 %rd16, %r85;
|
||
|
|
mov.u64 %rd19, rnd_seeds;
|
||
|
|
cvta.global.u64 %rd14, %rd19;
|
||
|
|
mov.u32 %r81, 2;
|
||
|
|
mov.u32 %r82, 4;
|
||
|
|
mov.u64 %rd18, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd13), _rt_buffer_get_64, (%rd14, %r81, %r82, %rd15, %rd16, %rd18, %rd18);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r191, [%rd13];
|
||
|
|
mov.u32 %r83, 0;
|
||
|
|
st.global.u32 [payload], %r83;
|
||
|
|
mov.u32 %r188, %r83;
|
||
|
|
|
||
|
|
BB0_1:
|
||
|
|
cvt.rn.f32.s32 %f14, %r188;
|
||
|
|
mov.u32 %r190, %r83;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
mad.lo.s32 %r89, %r191, 1664525, 1013904223;
|
||
|
|
and.b32 %r90, %r89, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f97, %r90;
|
||
|
|
fma.rn.f32 %f98, %f97, 0f33800000, %f14;
|
||
|
|
mul.f32 %f99, %f98, 0f3DCCCCCD;
|
||
|
|
mad.lo.s32 %r191, %r89, 1664525, 1013904223;
|
||
|
|
and.b32 %r91, %r191, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f100, %r91;
|
||
|
|
cvt.rn.f32.s32 %f101, %r190;
|
||
|
|
fma.rn.f32 %f102, %f100, 0f33800000, %f101;
|
||
|
|
mul.f32 %f103, %f102, 0f3DCCCCCD;
|
||
|
|
sqrt.rn.f32 %f16, %f99;
|
||
|
|
mul.f32 %f185, %f103, 0f40C90FDB;
|
||
|
|
abs.f32 %f18, %f185;
|
||
|
|
setp.neu.f32 %p2, %f18, 0f7F800000;
|
||
|
|
mov.f32 %f179, %f185;
|
||
|
|
@%p2 bra BB0_4;
|
||
|
|
|
||
|
|
mov.f32 %f104, 0f00000000;
|
||
|
|
mul.rn.f32 %f179, %f185, %f104;
|
||
|
|
|
||
|
|
BB0_4:
|
||
|
|
mul.f32 %f105, %f179, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r201, %f105;
|
||
|
|
cvt.rn.f32.s32 %f106, %r201;
|
||
|
|
neg.f32 %f107, %f106;
|
||
|
|
mov.f32 %f108, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f109, %f107, %f108, %f179;
|
||
|
|
mov.f32 %f110, 0f33A22168;
|
||
|
|
fma.rn.f32 %f111, %f107, %f110, %f109;
|
||
|
|
mov.f32 %f112, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f180, %f107, %f112, %f111;
|
||
|
|
abs.f32 %f113, %f179;
|
||
|
|
setp.leu.f32 %p3, %f113, 0f47CE4780;
|
||
|
|
@%p3 bra BB0_15;
|
||
|
|
|
||
|
|
mov.b32 %r8, %f179;
|
||
|
|
shr.u32 %r9, %r8, 23;
|
||
|
|
shl.b32 %r94, %r8, 8;
|
||
|
|
or.b32 %r10, %r94, -2147483648;
|
||
|
|
add.u64 %rd21, %SP, 4;
|
||
|
|
cvta.to.local.u64 %rd33, %rd21;
|
||
|
|
mov.u32 %r193, 0;
|
||
|
|
mov.u64 %rd32, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r192, -6;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r97, [%rd32];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r95, %r97, %r10, %r193;
|
||
|
|
madc.hi.u32 %r193, %r97, %r10, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd33], %r95;
|
||
|
|
add.s64 %rd33, %rd33, 4;
|
||
|
|
add.s64 %rd32, %rd32, 4;
|
||
|
|
add.s32 %r192, %r192, 1;
|
||
|
|
setp.ne.s32 %p4, %r192, 0;
|
||
|
|
@%p4 bra BB0_6;
|
||
|
|
|
||
|
|
and.b32 %r100, %r9, 255;
|
||
|
|
add.s32 %r101, %r100, -128;
|
||
|
|
shr.u32 %r102, %r101, 5;
|
||
|
|
and.b32 %r15, %r8, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd23, %rd21;
|
||
|
|
st.local.u32 [%rd23+24], %r193;
|
||
|
|
mov.u32 %r103, 6;
|
||
|
|
sub.s32 %r104, %r103, %r102;
|
||
|
|
mul.wide.s32 %rd24, %r104, 4;
|
||
|
|
add.s64 %rd6, %rd23, %rd24;
|
||
|
|
ld.local.u32 %r194, [%rd6];
|
||
|
|
ld.local.u32 %r195, [%rd6+-4];
|
||
|
|
and.b32 %r18, %r9, 31;
|
||
|
|
setp.eq.s32 %p5, %r18, 0;
|
||
|
|
@%p5 bra BB0_9;
|
||
|
|
|
||
|
|
mov.u32 %r105, 32;
|
||
|
|
sub.s32 %r106, %r105, %r18;
|
||
|
|
shr.u32 %r107, %r195, %r106;
|
||
|
|
shl.b32 %r108, %r194, %r18;
|
||
|
|
add.s32 %r194, %r107, %r108;
|
||
|
|
ld.local.u32 %r109, [%rd6+-8];
|
||
|
|
shr.u32 %r110, %r109, %r106;
|
||
|
|
shl.b32 %r111, %r195, %r18;
|
||
|
|
add.s32 %r195, %r110, %r111;
|
||
|
|
|
||
|
|
BB0_9:
|
||
|
|
shr.u32 %r112, %r195, 30;
|
||
|
|
shl.b32 %r113, %r194, 2;
|
||
|
|
add.s32 %r196, %r112, %r113;
|
||
|
|
shl.b32 %r24, %r195, 2;
|
||
|
|
shr.u32 %r114, %r196, 31;
|
||
|
|
shr.u32 %r115, %r194, 30;
|
||
|
|
add.s32 %r25, %r114, %r115;
|
||
|
|
setp.eq.s32 %p6, %r114, 0;
|
||
|
|
@%p6 bra BB0_10;
|
||
|
|
bra.uni BB0_11;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
mov.u32 %r197, %r15;
|
||
|
|
mov.u32 %r198, %r24;
|
||
|
|
bra.uni BB0_12;
|
||
|
|
|
||
|
|
BB0_11:
|
||
|
|
not.b32 %r116, %r196;
|
||
|
|
neg.s32 %r198, %r24;
|
||
|
|
setp.eq.s32 %p7, %r24, 0;
|
||
|
|
selp.u32 %r117, 1, 0, %p7;
|
||
|
|
add.s32 %r196, %r117, %r116;
|
||
|
|
xor.b32 %r197, %r15, -2147483648;
|
||
|
|
|
||
|
|
BB0_12:
|
||
|
|
clz.b32 %r200, %r196;
|
||
|
|
setp.eq.s32 %p8, %r200, 0;
|
||
|
|
shl.b32 %r118, %r196, %r200;
|
||
|
|
mov.u32 %r119, 32;
|
||
|
|
sub.s32 %r120, %r119, %r200;
|
||
|
|
shr.u32 %r121, %r198, %r120;
|
||
|
|
add.s32 %r122, %r121, %r118;
|
||
|
|
selp.b32 %r33, %r196, %r122, %p8;
|
||
|
|
mov.u32 %r123, -921707870;
|
||
|
|
mul.hi.u32 %r199, %r33, %r123;
|
||
|
|
setp.eq.s32 %p9, %r15, 0;
|
||
|
|
neg.s32 %r124, %r25;
|
||
|
|
selp.b32 %r201, %r25, %r124, %p9;
|
||
|
|
setp.lt.s32 %p10, %r199, 1;
|
||
|
|
@%p10 bra BB0_14;
|
||
|
|
|
||
|
|
mul.lo.s32 %r125, %r33, -921707870;
|
||
|
|
shr.u32 %r126, %r125, 31;
|
||
|
|
shl.b32 %r127, %r199, 1;
|
||
|
|
add.s32 %r199, %r126, %r127;
|
||
|
|
add.s32 %r200, %r200, 1;
|
||
|
|
|
||
|
|
BB0_14:
|
||
|
|
mov.u32 %r128, 126;
|
||
|
|
sub.s32 %r129, %r128, %r200;
|
||
|
|
shl.b32 %r130, %r129, 23;
|
||
|
|
add.s32 %r131, %r199, 1;
|
||
|
|
shr.u32 %r132, %r131, 7;
|
||
|
|
add.s32 %r133, %r132, 1;
|
||
|
|
shr.u32 %r134, %r133, 1;
|
||
|
|
add.s32 %r135, %r134, %r130;
|
||
|
|
or.b32 %r136, %r135, %r197;
|
||
|
|
mov.b32 %f180, %r136;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
mul.rn.f32 %f24, %f180, %f180;
|
||
|
|
add.s32 %r41, %r201, 1;
|
||
|
|
and.b32 %r42, %r41, 1;
|
||
|
|
setp.eq.s32 %p11, %r42, 0;
|
||
|
|
@%p11 bra BB0_17;
|
||
|
|
bra.uni BB0_16;
|
||
|
|
|
||
|
|
BB0_17:
|
||
|
|
mov.f32 %f116, 0f3C08839E;
|
||
|
|
mov.f32 %f117, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f181, %f117, %f24, %f116;
|
||
|
|
bra.uni BB0_18;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
mov.f32 %f114, 0fBAB6061A;
|
||
|
|
mov.f32 %f115, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f181, %f115, %f24, %f114;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
@%p11 bra BB0_20;
|
||
|
|
bra.uni BB0_19;
|
||
|
|
|
||
|
|
BB0_20:
|
||
|
|
mov.f32 %f121, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f122, %f181, %f24, %f121;
|
||
|
|
mov.f32 %f123, 0f00000000;
|
||
|
|
fma.rn.f32 %f182, %f122, %f24, %f123;
|
||
|
|
bra.uni BB0_21;
|
||
|
|
|
||
|
|
BB0_19:
|
||
|
|
mov.f32 %f118, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f119, %f181, %f24, %f118;
|
||
|
|
mov.f32 %f120, 0fBF000000;
|
||
|
|
fma.rn.f32 %f182, %f119, %f24, %f120;
|
||
|
|
|
||
|
|
BB0_21:
|
||
|
|
fma.rn.f32 %f183, %f182, %f180, %f180;
|
||
|
|
@%p11 bra BB0_23;
|
||
|
|
|
||
|
|
mov.f32 %f124, 0f3F800000;
|
||
|
|
fma.rn.f32 %f183, %f182, %f24, %f124;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
and.b32 %r137, %r41, 2;
|
||
|
|
setp.eq.s32 %p14, %r137, 0;
|
||
|
|
@%p14 bra BB0_25;
|
||
|
|
|
||
|
|
mov.f32 %f125, 0f00000000;
|
||
|
|
mov.f32 %f126, 0fBF800000;
|
||
|
|
fma.rn.f32 %f183, %f183, %f126, %f125;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
@%p2 bra BB0_27;
|
||
|
|
|
||
|
|
mov.f32 %f127, 0f00000000;
|
||
|
|
mul.rn.f32 %f185, %f185, %f127;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
mul.f32 %f128, %f185, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r211, %f128;
|
||
|
|
cvt.rn.f32.s32 %f129, %r211;
|
||
|
|
neg.f32 %f130, %f129;
|
||
|
|
fma.rn.f32 %f132, %f130, %f108, %f185;
|
||
|
|
fma.rn.f32 %f134, %f130, %f110, %f132;
|
||
|
|
fma.rn.f32 %f186, %f130, %f112, %f134;
|
||
|
|
abs.f32 %f136, %f185;
|
||
|
|
setp.leu.f32 %p16, %f136, 0f47CE4780;
|
||
|
|
@%p16 bra BB0_38;
|
||
|
|
|
||
|
|
mov.b32 %r44, %f185;
|
||
|
|
shr.u32 %r45, %r44, 23;
|
||
|
|
shl.b32 %r140, %r44, 8;
|
||
|
|
or.b32 %r46, %r140, -2147483648;
|
||
|
|
add.u64 %rd26, %SP, 4;
|
||
|
|
cvta.to.local.u64 %rd35, %rd26;
|
||
|
|
mov.u32 %r203, 0;
|
||
|
|
mov.u64 %rd34, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r202, -6;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r143, [%rd34];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r141, %r143, %r46, %r203;
|
||
|
|
madc.hi.u32 %r203, %r143, %r46, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd35], %r141;
|
||
|
|
add.s64 %rd35, %rd35, 4;
|
||
|
|
add.s64 %rd34, %rd34, 4;
|
||
|
|
add.s32 %r202, %r202, 1;
|
||
|
|
setp.ne.s32 %p17, %r202, 0;
|
||
|
|
@%p17 bra BB0_29;
|
||
|
|
|
||
|
|
and.b32 %r146, %r45, 255;
|
||
|
|
add.s32 %r147, %r146, -128;
|
||
|
|
shr.u32 %r148, %r147, 5;
|
||
|
|
and.b32 %r51, %r44, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd28, %rd26;
|
||
|
|
st.local.u32 [%rd28+24], %r203;
|
||
|
|
mov.u32 %r149, 6;
|
||
|
|
sub.s32 %r150, %r149, %r148;
|
||
|
|
mul.wide.s32 %rd29, %r150, 4;
|
||
|
|
add.s64 %rd12, %rd28, %rd29;
|
||
|
|
ld.local.u32 %r204, [%rd12];
|
||
|
|
ld.local.u32 %r205, [%rd12+-4];
|
||
|
|
and.b32 %r54, %r45, 31;
|
||
|
|
setp.eq.s32 %p18, %r54, 0;
|
||
|
|
@%p18 bra BB0_32;
|
||
|
|
|
||
|
|
mov.u32 %r151, 32;
|
||
|
|
sub.s32 %r152, %r151, %r54;
|
||
|
|
shr.u32 %r153, %r205, %r152;
|
||
|
|
shl.b32 %r154, %r204, %r54;
|
||
|
|
add.s32 %r204, %r153, %r154;
|
||
|
|
ld.local.u32 %r155, [%rd12+-8];
|
||
|
|
shr.u32 %r156, %r155, %r152;
|
||
|
|
shl.b32 %r157, %r205, %r54;
|
||
|
|
add.s32 %r205, %r156, %r157;
|
||
|
|
|
||
|
|
BB0_32:
|
||
|
|
shr.u32 %r158, %r205, 30;
|
||
|
|
shl.b32 %r159, %r204, 2;
|
||
|
|
add.s32 %r206, %r158, %r159;
|
||
|
|
shl.b32 %r60, %r205, 2;
|
||
|
|
shr.u32 %r160, %r206, 31;
|
||
|
|
shr.u32 %r161, %r204, 30;
|
||
|
|
add.s32 %r61, %r160, %r161;
|
||
|
|
setp.eq.s32 %p19, %r160, 0;
|
||
|
|
@%p19 bra BB0_33;
|
||
|
|
bra.uni BB0_34;
|
||
|
|
|
||
|
|
BB0_33:
|
||
|
|
mov.u32 %r207, %r51;
|
||
|
|
mov.u32 %r208, %r60;
|
||
|
|
bra.uni BB0_35;
|
||
|
|
|
||
|
|
BB0_34:
|
||
|
|
not.b32 %r162, %r206;
|
||
|
|
neg.s32 %r208, %r60;
|
||
|
|
setp.eq.s32 %p20, %r60, 0;
|
||
|
|
selp.u32 %r163, 1, 0, %p20;
|
||
|
|
add.s32 %r206, %r163, %r162;
|
||
|
|
xor.b32 %r207, %r51, -2147483648;
|
||
|
|
|
||
|
|
BB0_35:
|
||
|
|
clz.b32 %r210, %r206;
|
||
|
|
setp.eq.s32 %p21, %r210, 0;
|
||
|
|
shl.b32 %r164, %r206, %r210;
|
||
|
|
mov.u32 %r165, 32;
|
||
|
|
sub.s32 %r166, %r165, %r210;
|
||
|
|
shr.u32 %r167, %r208, %r166;
|
||
|
|
add.s32 %r168, %r167, %r164;
|
||
|
|
selp.b32 %r69, %r206, %r168, %p21;
|
||
|
|
mov.u32 %r169, -921707870;
|
||
|
|
mul.hi.u32 %r209, %r69, %r169;
|
||
|
|
setp.eq.s32 %p22, %r51, 0;
|
||
|
|
neg.s32 %r170, %r61;
|
||
|
|
selp.b32 %r211, %r61, %r170, %p22;
|
||
|
|
setp.lt.s32 %p23, %r209, 1;
|
||
|
|
@%p23 bra BB0_37;
|
||
|
|
|
||
|
|
mul.lo.s32 %r171, %r69, -921707870;
|
||
|
|
shr.u32 %r172, %r171, 31;
|
||
|
|
shl.b32 %r173, %r209, 1;
|
||
|
|
add.s32 %r209, %r172, %r173;
|
||
|
|
add.s32 %r210, %r210, 1;
|
||
|
|
|
||
|
|
BB0_37:
|
||
|
|
mov.u32 %r174, 126;
|
||
|
|
sub.s32 %r175, %r174, %r210;
|
||
|
|
shl.b32 %r176, %r175, 23;
|
||
|
|
add.s32 %r177, %r209, 1;
|
||
|
|
shr.u32 %r178, %r177, 7;
|
||
|
|
add.s32 %r179, %r178, 1;
|
||
|
|
shr.u32 %r180, %r179, 1;
|
||
|
|
add.s32 %r181, %r180, %r176;
|
||
|
|
or.b32 %r182, %r181, %r207;
|
||
|
|
mov.b32 %f186, %r182;
|
||
|
|
|
||
|
|
BB0_38:
|
||
|
|
mul.rn.f32 %f41, %f186, %f186;
|
||
|
|
and.b32 %r77, %r211, 1;
|
||
|
|
setp.eq.s32 %p24, %r77, 0;
|
||
|
|
@%p24 bra BB0_40;
|
||
|
|
bra.uni BB0_39;
|
||
|
|
|
||
|
|
BB0_40:
|
||
|
|
mov.f32 %f139, 0f3C08839E;
|
||
|
|
mov.f32 %f140, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f187, %f140, %f41, %f139;
|
||
|
|
bra.uni BB0_41;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
mov.f32 %f137, 0fBAB6061A;
|
||
|
|
mov.f32 %f138, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f187, %f138, %f41, %f137;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
@%p24 bra BB0_43;
|
||
|
|
bra.uni BB0_42;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
mov.f32 %f144, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f145, %f187, %f41, %f144;
|
||
|
|
mov.f32 %f146, 0f00000000;
|
||
|
|
fma.rn.f32 %f188, %f145, %f41, %f146;
|
||
|
|
bra.uni BB0_44;
|
||
|
|
|
||
|
|
BB0_42:
|
||
|
|
mov.f32 %f141, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f142, %f187, %f41, %f141;
|
||
|
|
mov.f32 %f143, 0fBF000000;
|
||
|
|
fma.rn.f32 %f188, %f142, %f41, %f143;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
fma.rn.f32 %f189, %f188, %f186, %f186;
|
||
|
|
@%p24 bra BB0_46;
|
||
|
|
|
||
|
|
mov.f32 %f147, 0f3F800000;
|
||
|
|
fma.rn.f32 %f189, %f188, %f41, %f147;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
and.b32 %r183, %r211, 2;
|
||
|
|
setp.eq.s32 %p27, %r183, 0;
|
||
|
|
@%p27 bra BB0_48;
|
||
|
|
|
||
|
|
mov.f32 %f148, 0f00000000;
|
||
|
|
mov.f32 %f149, 0fBF800000;
|
||
|
|
fma.rn.f32 %f189, %f189, %f149, %f148;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
mul.f32 %f150, %f16, %f183;
|
||
|
|
mul.f32 %f151, %f150, %f150;
|
||
|
|
mov.f32 %f152, 0f3F800000;
|
||
|
|
sub.f32 %f153, %f152, %f151;
|
||
|
|
mul.f32 %f154, %f16, %f189;
|
||
|
|
mul.f32 %f155, %f154, %f154;
|
||
|
|
sub.f32 %f156, %f153, %f155;
|
||
|
|
mov.f32 %f157, 0f00000000;
|
||
|
|
max.f32 %f158, %f157, %f156;
|
||
|
|
sqrt.rn.f32 %f159, %f158;
|
||
|
|
mul.f32 %f160, %f7, %f154;
|
||
|
|
mul.f32 %f161, %f8, %f154;
|
||
|
|
mul.f32 %f162, %f9, %f154;
|
||
|
|
fma.rn.f32 %f163, %f10, %f150, %f160;
|
||
|
|
fma.rn.f32 %f164, %f11, %f150, %f161;
|
||
|
|
fma.rn.f32 %f165, %f12, %f150, %f162;
|
||
|
|
fma.rn.f32 %f53, %f1, %f159, %f163;
|
||
|
|
fma.rn.f32 %f54, %f2, %f159, %f164;
|
||
|
|
fma.rn.f32 %f55, %f3, %f159, %f165;
|
||
|
|
setp.leu.f32 %p28, %f54, 0f00000000;
|
||
|
|
@%p28 bra BB0_50;
|
||
|
|
|
||
|
|
add.u64 %rd30, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd31, %rd30;
|
||
|
|
mov.u32 %r187, 1065353216;
|
||
|
|
st.local.u32 [%rd31], %r187;
|
||
|
|
ld.global.u32 %r184, [root];
|
||
|
|
mov.u32 %r185, 1;
|
||
|
|
mov.f32 %f172, 0f3A83126F;
|
||
|
|
mov.f32 %f173, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r184, %f4, %f5, %f6, %f53, %f54, %f55, %r185, %f172, %f173, %rd30, %r82);
|
||
|
|
// inline asm
|
||
|
|
ld.global.f32 %f174, [payload];
|
||
|
|
ld.local.f32 %f175, [%rd31];
|
||
|
|
add.f32 %f191, %f175, %f174;
|
||
|
|
st.global.f32 [payload], %f191;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
add.s32 %r190, %r190, 1;
|
||
|
|
setp.lt.s32 %p29, %r190, 10;
|
||
|
|
@%p29 bra BB0_2;
|
||
|
|
|
||
|
|
add.s32 %r188, %r188, 1;
|
||
|
|
setp.lt.s32 %p30, %r188, 10;
|
||
|
|
@%p30 bra BB0_1;
|
||
|
|
|
||
|
|
div.rn.f32 %f176, %f191, 0f42C80000;
|
||
|
|
st.global.f32 [payload], %f176;
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|