507 lines
19 KiB
Plaintext
507 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 8 .b8 texCoords[8];
|
|
.global .align 1 .b8 textureIDs[1];
|
|
.global .align 1 .b8 emissiveIDs[1];
|
|
.global .align 1 .b8 areaTexIDs[1];
|
|
.global .align 1 .b8 areaLights[1];
|
|
.global .align 4 .b8 payload[28];
|
|
.global .align 4 .b8 ray[36];
|
|
.global .align 4 .f32 t;
|
|
.global .align 4 .u32 bounceFlag;
|
|
.global .align 4 .f32 backFaceWeight;
|
|
.global .align 4 .f32 emissiveBoost;
|
|
.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_typeinfo9texCoordsE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7payloadE[8] = {82, 97, 121, 0, 28, 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 4 .b8 _ZN21rti_internal_typeinfo10bounceFlagE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14backFaceWeightE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo13emissiveBoostE[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_typename9texCoordsE[7] = {102, 108, 111, 97, 116, 50, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename7payloadE[15] = {82, 97, 121, 68, 97, 116, 97, 80, 114, 101, 118, 105, 101, 119, 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 .b8 _ZN21rti_internal_typename10bounceFlagE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14backFaceWeightE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename13emissiveBoostE[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_typeenum9texCoordsE = 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 4 .u32 _ZN21rti_internal_typeenum10bounceFlagE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14backFaceWeightE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum13emissiveBoostE = 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_semantic9texCoordsE[20] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 116, 101, 120, 67, 111, 111, 114, 100, 115, 0};
|
|
.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 _ZN21rti_internal_semantic10bounceFlagE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14backFaceWeightE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic13emissiveBoostE[1];
|
|
.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_annotation9texCoordsE[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];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10bounceFlagE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14backFaceWeightE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation13emissiveBoostE[1];
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.reg .pred %p<29>;
|
|
.reg .b16 %rs<4>;
|
|
.reg .f32 %f<227>;
|
|
.reg .b32 %r<39>;
|
|
.reg .b64 %rd<30>;
|
|
|
|
|
|
ld.global.f32 %f43, [t];
|
|
st.global.f32 [payload+12], %f43;
|
|
ld.global.f32 %f44, [normal];
|
|
ld.global.f32 %f45, [normal+4];
|
|
ld.global.f32 %f46, [normal+8];
|
|
st.global.f32 [payload+8], %f46;
|
|
st.global.f32 [payload+4], %f45;
|
|
st.global.f32 [payload], %f44;
|
|
ld.global.f32 %f47, [ray+12];
|
|
ld.global.f32 %f48, [ray+16];
|
|
mul.f32 %f49, %f45, %f48;
|
|
fma.rn.f32 %f50, %f44, %f47, %f49;
|
|
ld.global.f32 %f51, [ray+20];
|
|
fma.rn.f32 %f1, %f46, %f51, %f50;
|
|
setp.ge.f32 %p2, %f1, 0f00000000;
|
|
ld.global.f32 %f52, [backFaceWeight];
|
|
setp.eq.f32 %p3, %f52, 0f00000000;
|
|
and.pred %p4, %p2, %p3;
|
|
@%p4 bra BB0_23;
|
|
bra.uni BB0_1;
|
|
|
|
BB0_23:
|
|
mov.u32 %r38, 0;
|
|
st.global.u32 [payload+24], %r38;
|
|
st.global.u32 [payload+20], %r38;
|
|
st.global.u32 [payload+16], %r38;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_1:
|
|
ld.global.f32 %f2, [texCoords];
|
|
div.rn.f32 %f53, %f2, 0f41200000;
|
|
cvt.rzi.s32.f32 %r1, %f53;
|
|
setp.lt.s32 %p5, %r1, 0;
|
|
@%p5 bra BB0_6;
|
|
|
|
cvt.rzi.s32.f32 %r9, %f2;
|
|
cvt.rn.f32.s32 %f62, %r9;
|
|
sub.f32 %f58, %f2, %f62;
|
|
ld.global.f32 %f4, [texCoords+4];
|
|
cvt.s64.s32 %rd10, %r1;
|
|
mov.u64 %rd14, textureIDs;
|
|
cvta.global.u64 %rd3, %rd14;
|
|
mov.u32 %r7, 1;
|
|
mov.u32 %r8, 4;
|
|
mov.u64 %rd13, 0;
|
|
// inline asm
|
|
call (%rd2), _rt_buffer_get_64, (%rd3, %r7, %r8, %rd10, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd2];
|
|
mov.u32 %r6, 2;
|
|
mov.f32 %f61, 0f00000000;
|
|
// inline asm
|
|
call (%f54, %f55, %f56, %f57), _rt_texture_get_f_id, (%r5, %r6, %f58, %f4, %f61, %f61);
|
|
// inline asm
|
|
add.f32 %f63, %f54, 0fB4000000;
|
|
max.f32 %f64, %f63, %f61;
|
|
add.f32 %f65, %f55, 0fB4000000;
|
|
max.f32 %f66, %f65, %f61;
|
|
add.f32 %f67, %f56, 0fB4000000;
|
|
max.f32 %f68, %f67, %f61;
|
|
st.global.f32 [payload+16], %f64;
|
|
st.global.f32 [payload+20], %f66;
|
|
st.global.f32 [payload+24], %f68;
|
|
mov.u64 %rd15, emissiveIDs;
|
|
cvta.global.u64 %rd9, %rd15;
|
|
// inline asm
|
|
call (%rd8), _rt_buffer_get_64, (%rd9, %r7, %r8, %rd10, %rd13, %rd13, %rd13);
|
|
// inline asm
|
|
ld.u32 %r2, [%rd8];
|
|
setp.gt.s32 %p6, %r2, 0;
|
|
@%p6 bra BB0_4;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_4:
|
|
// inline asm
|
|
call (%r10, %r11, %r12, %r13), _rt_texture_get_u_id, (%r2, %r6, %f58, %f4, %f61, %f61);
|
|
// inline asm
|
|
cvt.u16.u32 %rs1, %r10;
|
|
cvt.u16.u32 %rs2, %r11;
|
|
cvt.u16.u32 %rs3, %r12;
|
|
// inline asm
|
|
{ cvt.f32.f16 %f73, %rs1;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f74, %rs2;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f75, %rs3;}
|
|
|
|
// inline asm
|
|
ld.global.f32 %f76, [emissiveBoost];
|
|
mul.f32 %f77, %f73, %f76;
|
|
mul.f32 %f78, %f74, %f76;
|
|
mul.f32 %f79, %f75, %f76;
|
|
mul.f32 %f80, %f77, 0f437F0000;
|
|
cvt.rmi.f32.f32 %f81, %f80;
|
|
mul.f32 %f82, %f78, 0f437F0000;
|
|
cvt.rmi.f32.f32 %f83, %f82;
|
|
mul.f32 %f84, %f79, 0f437F0000;
|
|
cvt.rmi.f32.f32 %f85, %f84;
|
|
ld.global.f32 %f86, [payload+16];
|
|
add.f32 %f218, %f81, %f86;
|
|
ld.global.f32 %f87, [payload+20];
|
|
ld.global.f32 %f88, [payload+24];
|
|
st.global.f32 [payload+16], %f218;
|
|
add.f32 %f89, %f83, %f87;
|
|
st.global.f32 [payload+20], %f89;
|
|
add.f32 %f90, %f85, %f88;
|
|
st.global.f32 [payload+24], %f90;
|
|
bra.uni BB0_5;
|
|
|
|
BB0_6:
|
|
neg.s32 %r18, %r1;
|
|
cvt.s64.s32 %rd18, %r18;
|
|
mov.u64 %rd22, areaLights;
|
|
cvta.global.u64 %rd17, %rd22;
|
|
mov.u32 %r16, 1;
|
|
mov.u32 %r17, 16;
|
|
mov.u64 %rd21, 0;
|
|
// inline asm
|
|
call (%rd16), _rt_buffer_get_64, (%rd17, %r16, %r17, %rd18, %rd21, %rd21, %rd21);
|
|
// inline asm
|
|
ld.v4.f32 {%f93, %f94, %f95, %f222}, [%rd16];
|
|
mov.b32 %r19, %f93;
|
|
shr.s32 %r20, %r19, 16;
|
|
cvt.rn.f32.s32 %f97, %r20;
|
|
div.rn.f32 %f98, %f97, 0f437F0000;
|
|
mul.f32 %f219, %f98, %f94;
|
|
bfe.u32 %r21, %r19, 8, 8;
|
|
cvt.rn.f32.s32 %f99, %r21;
|
|
div.rn.f32 %f100, %f99, 0f437F0000;
|
|
mul.f32 %f220, %f94, %f100;
|
|
and.b32 %r22, %r19, 255;
|
|
cvt.rn.f32.s32 %f101, %r22;
|
|
div.rn.f32 %f102, %f101, 0f437F0000;
|
|
mul.f32 %f221, %f94, %f102;
|
|
setp.gtu.f32 %p8, %f222, 0f00000000;
|
|
@%p8 bra BB0_8;
|
|
|
|
neg.f32 %f222, %f222;
|
|
mov.u64 %rd29, areaTexIDs;
|
|
cvta.global.u64 %rd24, %rd29;
|
|
mov.u32 %r24, 4;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r16, %r24, %rd18, %rd21, %rd21, %rd21);
|
|
// inline asm
|
|
ld.u32 %r25, [%rd23];
|
|
ld.global.v2.f32 {%f111, %f112}, [texCoords];
|
|
neg.f32 %f114, %f111;
|
|
mul.lo.s32 %r27, %r1, -10;
|
|
cvt.rn.f32.s32 %f115, %r27;
|
|
sub.f32 %f107, %f114, %f115;
|
|
mov.u32 %r26, 2;
|
|
mov.f32 %f110, 0f00000000;
|
|
// inline asm
|
|
call (%f103, %f104, %f105, %f106), _rt_texture_get_f_id, (%r25, %r26, %f107, %f112, %f110, %f110);
|
|
// inline asm
|
|
mul.f32 %f219, %f219, %f103;
|
|
mul.f32 %f220, %f220, %f104;
|
|
mul.f32 %f221, %f221, %f105;
|
|
|
|
BB0_8:
|
|
ld.global.u32 %r28, [bounceFlag];
|
|
cvt.rn.f32.s32 %f24, %r28;
|
|
mov.f32 %f226, 0f3F800000;
|
|
setp.leu.f32 %p9, %f24, 0f00000000;
|
|
@%p9 bra BB0_22;
|
|
|
|
ld.global.f32 %f119, [t];
|
|
mul.f32 %f26, %f222, %f119;
|
|
mov.f32 %f120, 0f40000000;
|
|
cvt.rzi.f32.f32 %f121, %f120;
|
|
add.f32 %f122, %f121, %f121;
|
|
mov.f32 %f123, 0f40800000;
|
|
sub.f32 %f124, %f123, %f122;
|
|
abs.f32 %f27, %f124;
|
|
abs.f32 %f28, %f26;
|
|
setp.lt.f32 %p10, %f28, 0f00800000;
|
|
mul.f32 %f125, %f28, 0f4B800000;
|
|
selp.f32 %f126, 0fC3170000, 0fC2FE0000, %p10;
|
|
selp.f32 %f127, %f125, %f28, %p10;
|
|
mov.b32 %r29, %f127;
|
|
and.b32 %r30, %r29, 8388607;
|
|
or.b32 %r31, %r30, 1065353216;
|
|
mov.b32 %f128, %r31;
|
|
shr.u32 %r32, %r29, 23;
|
|
cvt.rn.f32.u32 %f129, %r32;
|
|
add.f32 %f130, %f126, %f129;
|
|
setp.gt.f32 %p11, %f128, 0f3FB504F3;
|
|
mul.f32 %f131, %f128, 0f3F000000;
|
|
add.f32 %f132, %f130, 0f3F800000;
|
|
selp.f32 %f133, %f131, %f128, %p11;
|
|
selp.f32 %f134, %f132, %f130, %p11;
|
|
add.f32 %f135, %f133, 0fBF800000;
|
|
add.f32 %f118, %f133, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f117,%f118;
|
|
// inline asm
|
|
add.f32 %f136, %f135, %f135;
|
|
mul.f32 %f137, %f117, %f136;
|
|
mul.f32 %f138, %f137, %f137;
|
|
mov.f32 %f139, 0f3C4CAF63;
|
|
mov.f32 %f140, 0f3B18F0FE;
|
|
fma.rn.f32 %f141, %f140, %f138, %f139;
|
|
mov.f32 %f142, 0f3DAAAABD;
|
|
fma.rn.f32 %f143, %f141, %f138, %f142;
|
|
mul.rn.f32 %f144, %f143, %f138;
|
|
mul.rn.f32 %f145, %f144, %f137;
|
|
sub.f32 %f146, %f135, %f137;
|
|
neg.f32 %f147, %f137;
|
|
add.f32 %f148, %f146, %f146;
|
|
fma.rn.f32 %f149, %f147, %f135, %f148;
|
|
mul.rn.f32 %f150, %f117, %f149;
|
|
add.f32 %f151, %f145, %f137;
|
|
sub.f32 %f152, %f137, %f151;
|
|
add.f32 %f153, %f145, %f152;
|
|
add.f32 %f154, %f150, %f153;
|
|
add.f32 %f155, %f151, %f154;
|
|
sub.f32 %f156, %f151, %f155;
|
|
add.f32 %f157, %f154, %f156;
|
|
mov.f32 %f158, 0f3F317200;
|
|
mul.rn.f32 %f159, %f134, %f158;
|
|
mov.f32 %f160, 0f35BFBE8E;
|
|
mul.rn.f32 %f161, %f134, %f160;
|
|
add.f32 %f162, %f159, %f155;
|
|
sub.f32 %f163, %f159, %f162;
|
|
add.f32 %f164, %f155, %f163;
|
|
add.f32 %f165, %f157, %f164;
|
|
add.f32 %f166, %f161, %f165;
|
|
add.f32 %f167, %f162, %f166;
|
|
sub.f32 %f168, %f162, %f167;
|
|
add.f32 %f169, %f166, %f168;
|
|
mul.rn.f32 %f170, %f123, %f167;
|
|
neg.f32 %f171, %f170;
|
|
fma.rn.f32 %f172, %f123, %f167, %f171;
|
|
fma.rn.f32 %f173, %f123, %f169, %f172;
|
|
mov.f32 %f174, 0f00000000;
|
|
fma.rn.f32 %f175, %f174, %f167, %f173;
|
|
add.rn.f32 %f176, %f170, %f175;
|
|
neg.f32 %f177, %f176;
|
|
add.rn.f32 %f178, %f170, %f177;
|
|
add.rn.f32 %f179, %f178, %f175;
|
|
mov.b32 %r33, %f176;
|
|
setp.eq.s32 %p12, %r33, 1118925336;
|
|
add.s32 %r34, %r33, -1;
|
|
mov.b32 %f180, %r34;
|
|
add.f32 %f181, %f179, 0f37000000;
|
|
selp.f32 %f182, %f180, %f176, %p12;
|
|
selp.f32 %f29, %f181, %f179, %p12;
|
|
mul.f32 %f183, %f182, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f184, %f183;
|
|
mov.f32 %f185, 0fBF317200;
|
|
fma.rn.f32 %f186, %f184, %f185, %f182;
|
|
mov.f32 %f187, 0fB5BFBE8E;
|
|
fma.rn.f32 %f188, %f184, %f187, %f186;
|
|
mul.f32 %f189, %f188, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f190, %f189;
|
|
add.f32 %f191, %f184, 0f00000000;
|
|
ex2.approx.f32 %f192, %f191;
|
|
mul.f32 %f193, %f190, %f192;
|
|
setp.lt.f32 %p13, %f182, 0fC2D20000;
|
|
selp.f32 %f194, 0f00000000, %f193, %p13;
|
|
setp.gt.f32 %p14, %f182, 0f42D20000;
|
|
selp.f32 %f223, 0f7F800000, %f194, %p14;
|
|
setp.eq.f32 %p15, %f223, 0f7F800000;
|
|
@%p15 bra BB0_11;
|
|
|
|
fma.rn.f32 %f223, %f223, %f29, %f223;
|
|
|
|
BB0_11:
|
|
setp.lt.f32 %p16, %f26, 0f00000000;
|
|
setp.eq.f32 %p17, %f27, 0f3F800000;
|
|
and.pred %p1, %p16, %p17;
|
|
mov.b32 %r35, %f223;
|
|
xor.b32 %r36, %r35, -2147483648;
|
|
mov.b32 %f195, %r36;
|
|
selp.f32 %f225, %f195, %f223, %p1;
|
|
setp.eq.f32 %p18, %f26, 0f00000000;
|
|
@%p18 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f198, %f26, %f26;
|
|
selp.f32 %f225, %f198, 0f00000000, %p17;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_3:
|
|
ld.global.f32 %f218, [payload+16];
|
|
|
|
BB0_5:
|
|
selp.f32 %f91, 0fBF800000, 0f3F800000, %p2;
|
|
mul.f32 %f92, %f91, %f218;
|
|
st.global.f32 [payload+16], %f92;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p19, %f26, 0f00000000;
|
|
@%p19 bra BB0_15;
|
|
|
|
cvt.rzi.f32.f32 %f197, %f123;
|
|
setp.neu.f32 %p20, %f197, 0f40800000;
|
|
selp.f32 %f225, 0f7FFFFFFF, %f225, %p20;
|
|
|
|
BB0_15:
|
|
add.f32 %f199, %f28, 0f40800000;
|
|
mov.b32 %r37, %f199;
|
|
setp.lt.s32 %p22, %r37, 2139095040;
|
|
@%p22 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p23, %f28, 0f7F800000;
|
|
@%p23 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f225, %f26, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p24, %f28, 0f7F800000;
|
|
@%p24 bra BB0_20;
|
|
|
|
selp.f32 %f225, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.f32 %f200, 0f3F800000;
|
|
sub.f32 %f201, %f200, %f225;
|
|
setp.eq.f32 %p25, %f26, 0f3F800000;
|
|
selp.f32 %f202, 0f00000000, %f201, %p25;
|
|
cvt.sat.f32.f32 %f226, %f202;
|
|
setp.leu.f32 %p26, %f24, 0f3F800000;
|
|
@%p26 bra BB0_22;
|
|
|
|
setp.lt.f32 %p27, %f95, 0f00000000;
|
|
neg.f32 %f203, %f95;
|
|
selp.f32 %f204, %f203, %f95, %p27;
|
|
mul.f32 %f226, %f204, %f226;
|
|
|
|
BB0_22:
|
|
selp.f32 %f205, 0f00000000, 0f3F800000, %p2;
|
|
mul.f32 %f206, %f205, %f219;
|
|
mul.f32 %f207, %f206, %f226;
|
|
mul.f32 %f208, %f205, %f220;
|
|
mul.f32 %f209, %f208, %f226;
|
|
mul.f32 %f210, %f205, %f221;
|
|
mul.f32 %f211, %f210, %f226;
|
|
mul.f32 %f212, %f207, 0f437F0000;
|
|
mul.f32 %f213, %f209, 0f437F0000;
|
|
mul.f32 %f214, %f211, 0f437F0000;
|
|
cvt.rmi.f32.f32 %f215, %f212;
|
|
cvt.rmi.f32.f32 %f216, %f213;
|
|
cvt.rmi.f32.f32 %f217, %f214;
|
|
st.global.f32 [payload+16], %f215;
|
|
st.global.f32 [payload+20], %f216;
|
|
st.global.f32 [payload+24], %f217;
|
|
|
|
BB0_24:
|
|
ret;
|
|
}
|
|
|
|
|