2328 lines
64 KiB
Plaintext
2328 lines
64 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 1 .b8 image[1];
|
||
|
|
.global .align 1 .b8 image_HDR[1];
|
||
|
|
.global .align 1 .b8 image_HDR2[1];
|
||
|
|
.global .align 1 .b8 image_Mask[1];
|
||
|
|
.global .align 1 .b8 image_RNM0[1];
|
||
|
|
.global .align 1 .b8 image_RNM1[1];
|
||
|
|
.global .align 1 .b8 image_RNM2[1];
|
||
|
|
.global .align 1 .b8 image_RNM3[1];
|
||
|
|
.global .align 1 .b8 uvpos[1];
|
||
|
|
.global .align 1 .b8 uvnormal[1];
|
||
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
||
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
||
|
|
.global .align 4 .f32 lightInvCutoff;
|
||
|
|
.global .align 4 .f32 lightPointSize;
|
||
|
|
.global .align 4 .b8 lightColor[12];
|
||
|
|
.global .align 1 .b8 rnd_seeds[1];
|
||
|
|
.global .align 1 .b8 rnd_seeds2[1];
|
||
|
|
.global .align 4 .u32 samples;
|
||
|
|
.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_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[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 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 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_typeenum19lightMeshBufferSizeE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation19lightMeshBufferSizeE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[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[36];
|
||
|
|
.reg .b64 %SP;
|
||
|
|
.reg .b64 %SPL;
|
||
|
|
.reg .pred %p<149>;
|
||
|
|
.reg .b16 %rs<153>;
|
||
|
|
.reg .f32 %f<988>;
|
||
|
|
.reg .b32 %r<413>;
|
||
|
|
.reg .b64 %rd<308>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd307, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd307;
|
||
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
||
|
|
abs.s32 %r2, %r1;
|
||
|
|
ld.global.u32 %r3, [samples];
|
||
|
|
shl.b32 %r4, %r3, 1;
|
||
|
|
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd22, %r103;
|
||
|
|
cvt.u64.u32 %rd23, %r104;
|
||
|
|
mov.u64 %rd26, uvnormal;
|
||
|
|
cvta.global.u64 %rd21, %rd26;
|
||
|
|
mov.u32 %r101, 2;
|
||
|
|
mov.u32 %r102, 4;
|
||
|
|
mov.u64 %rd25, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r101, %r102, %rd22, %rd23, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r5, [%rd20];
|
||
|
|
shr.u32 %r107, %r5, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r107;
|
||
|
|
and.b16 %rs7, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs8, %r5;
|
||
|
|
or.b16 %rs9, %rs8, %rs7;
|
||
|
|
setp.eq.s16 %p7, %rs9, 0;
|
||
|
|
mov.f32 %f916, 0f00000000;
|
||
|
|
mov.f32 %f917, %f916;
|
||
|
|
mov.f32 %f918, %f916;
|
||
|
|
@%p7 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs10, [%rd20+1];
|
||
|
|
and.b16 %rs12, %rs8, 255;
|
||
|
|
cvt.rn.f32.u16 %f217, %rs12;
|
||
|
|
div.rn.f32 %f218, %f217, 0f437F0000;
|
||
|
|
fma.rn.f32 %f219, %f218, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f220, %rs10;
|
||
|
|
div.rn.f32 %f221, %f220, 0f437F0000;
|
||
|
|
fma.rn.f32 %f222, %f221, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f223, %rs7;
|
||
|
|
div.rn.f32 %f224, %f223, 0f437F0000;
|
||
|
|
fma.rn.f32 %f225, %f224, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f226, %f222, %f222;
|
||
|
|
fma.rn.f32 %f227, %f219, %f219, %f226;
|
||
|
|
fma.rn.f32 %f228, %f225, %f225, %f227;
|
||
|
|
sqrt.rn.f32 %f229, %f228;
|
||
|
|
rcp.rn.f32 %f230, %f229;
|
||
|
|
mul.f32 %f916, %f219, %f230;
|
||
|
|
mul.f32 %f917, %f222, %f230;
|
||
|
|
mul.f32 %f918, %f225, %f230;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
||
|
|
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
|
||
|
|
add.s32 %r6, %r108, %r111;
|
||
|
|
add.s32 %r7, %r109, %r112;
|
||
|
|
setp.eq.f32 %p8, %f917, 0f00000000;
|
||
|
|
setp.eq.f32 %p9, %f916, 0f00000000;
|
||
|
|
and.pred %p10, %p9, %p8;
|
||
|
|
setp.eq.f32 %p11, %f918, 0f00000000;
|
||
|
|
and.pred %p12, %p10, %p11;
|
||
|
|
@%p12 bra BB0_140;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_140:
|
||
|
|
ld.global.u32 %r412, [imageEnabled];
|
||
|
|
and.b32 %r337, %r412, 1;
|
||
|
|
setp.eq.b32 %p140, %r337, 1;
|
||
|
|
@!%p140 bra BB0_142;
|
||
|
|
bra.uni BB0_141;
|
||
|
|
|
||
|
|
BB0_141:
|
||
|
|
cvt.u64.u32 %rd189, %r6;
|
||
|
|
cvt.u64.u32 %rd190, %r7;
|
||
|
|
mov.u64 %rd193, image;
|
||
|
|
cvta.global.u64 %rd188, %rd193;
|
||
|
|
// inline asm
|
||
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r101, %r102, %rd189, %rd190, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs85, 0;
|
||
|
|
st.v4.u8 [%rd187], {%rs85, %rs85, %rs85, %rs85};
|
||
|
|
ld.global.u32 %r412, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_142:
|
||
|
|
and.b32 %r340, %r412, 8;
|
||
|
|
setp.eq.s32 %p141, %r340, 0;
|
||
|
|
@%p141 bra BB0_144;
|
||
|
|
|
||
|
|
cvt.u64.u32 %rd197, %r7;
|
||
|
|
cvt.u64.u32 %rd196, %r6;
|
||
|
|
mov.u64 %rd200, image_Mask;
|
||
|
|
cvta.global.u64 %rd195, %rd200;
|
||
|
|
// inline asm
|
||
|
|
call (%rd194), _rt_buffer_get_64, (%rd195, %r101, %r101, %rd196, %rd197, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f873, 0f00000000;
|
||
|
|
cvt.rzi.u32.f32 %r343, %f873;
|
||
|
|
cvt.u16.u32 %rs86, %r343;
|
||
|
|
mov.u16 %rs87, 0;
|
||
|
|
st.v2.u8 [%rd194], {%rs86, %rs87};
|
||
|
|
ld.global.u32 %r412, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_144:
|
||
|
|
cvt.u64.u32 %rd18, %r6;
|
||
|
|
cvt.u64.u32 %rd19, %r7;
|
||
|
|
and.b32 %r344, %r412, 4;
|
||
|
|
setp.eq.s32 %p142, %r344, 0;
|
||
|
|
@%p142 bra BB0_148;
|
||
|
|
|
||
|
|
ld.global.u32 %r345, [additive];
|
||
|
|
setp.eq.s32 %p143, %r345, 0;
|
||
|
|
@%p143 bra BB0_147;
|
||
|
|
|
||
|
|
mov.u64 %rd213, image_HDR;
|
||
|
|
cvta.global.u64 %rd202, %rd213;
|
||
|
|
mov.u32 %r349, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd201), _rt_buffer_get_64, (%rd202, %r101, %r349, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs94, %rs95, %rs96, %rs97}, [%rd201];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f874, %rs94;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f875, %rs95;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f876, %rs96;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd207), _rt_buffer_get_64, (%rd202, %r101, %r349, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f877, %f874, 0f00000000;
|
||
|
|
add.f32 %f878, %f875, 0f00000000;
|
||
|
|
add.f32 %f879, %f876, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs93, %f879;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs92, %f878;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs91, %f877;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs98, 0;
|
||
|
|
st.v4.u16 [%rd207], {%rs91, %rs92, %rs93, %rs98};
|
||
|
|
bra.uni BB0_148;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd29, %r121;
|
||
|
|
cvt.u64.u32 %rd30, %r122;
|
||
|
|
mov.u64 %rd39, uvpos;
|
||
|
|
cvta.global.u64 %rd28, %rd39;
|
||
|
|
mov.u32 %r117, 12;
|
||
|
|
// inline asm
|
||
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r101, %r117, %rd29, %rd30, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f9, [%rd27+8];
|
||
|
|
ld.f32 %f8, [%rd27+4];
|
||
|
|
ld.f32 %f7, [%rd27];
|
||
|
|
mul.f32 %f237, %f7, 0f3456BF95;
|
||
|
|
mul.f32 %f238, %f8, 0f3456BF95;
|
||
|
|
mul.f32 %f239, %f9, 0f3456BF95;
|
||
|
|
abs.f32 %f240, %f916;
|
||
|
|
div.rn.f32 %f241, %f237, %f240;
|
||
|
|
abs.f32 %f242, %f917;
|
||
|
|
div.rn.f32 %f243, %f238, %f242;
|
||
|
|
abs.f32 %f244, %f918;
|
||
|
|
div.rn.f32 %f245, %f239, %f244;
|
||
|
|
abs.f32 %f246, %f241;
|
||
|
|
abs.f32 %f247, %f243;
|
||
|
|
abs.f32 %f248, %f245;
|
||
|
|
mov.f32 %f249, 0f38D1B717;
|
||
|
|
max.f32 %f250, %f246, %f249;
|
||
|
|
max.f32 %f251, %f247, %f249;
|
||
|
|
max.f32 %f252, %f248, %f249;
|
||
|
|
fma.rn.f32 %f10, %f916, %f250, %f7;
|
||
|
|
fma.rn.f32 %f11, %f917, %f251, %f8;
|
||
|
|
fma.rn.f32 %f12, %f918, %f252, %f9;
|
||
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd35, %r125;
|
||
|
|
cvt.u64.u32 %rd36, %r126;
|
||
|
|
mov.u64 %rd40, rnd_seeds;
|
||
|
|
cvta.global.u64 %rd34, %rd40;
|
||
|
|
// inline asm
|
||
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r101, %r102, %rd35, %rd36, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r388, [%rd33];
|
||
|
|
mov.pred %p13, 0;
|
||
|
|
mov.f32 %f17, 0f00000000;
|
||
|
|
mov.u32 %r384, 0;
|
||
|
|
setp.lt.s32 %p14, %r2, 1;
|
||
|
|
@%p14 bra BB0_4;
|
||
|
|
|
||
|
|
ld.global.f32 %f13, [lightPointSize];
|
||
|
|
mul.f32 %f14, %f10, 0f3456BF95;
|
||
|
|
mul.f32 %f15, %f11, 0f3456BF95;
|
||
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
||
|
|
mov.f32 %f17, 0f00000000;
|
||
|
|
mov.u32 %r384, 0;
|
||
|
|
abs.f32 %f381, %f15;
|
||
|
|
abs.f32 %f382, %f14;
|
||
|
|
max.f32 %f383, %f382, %f381;
|
||
|
|
abs.f32 %f384, %f16;
|
||
|
|
max.f32 %f385, %f383, %f384;
|
||
|
|
mov.f32 %f18, %f17;
|
||
|
|
mov.f32 %f19, %f17;
|
||
|
|
mov.f32 %f20, %f17;
|
||
|
|
mov.f32 %f21, %f17;
|
||
|
|
mov.f32 %f22, %f17;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
shl.b32 %r10, %r384, 1;
|
||
|
|
cvt.s64.s32 %rd43, %r10;
|
||
|
|
mov.u64 %rd53, lightMeshBuffer;
|
||
|
|
cvta.global.u64 %rd42, %rd53;
|
||
|
|
mov.u32 %r130, 1;
|
||
|
|
// inline asm
|
||
|
|
call (%rd41), _rt_buffer_get_64, (%rd42, %r130, %r117, %rd43, %rd25, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f259, [%rd41];
|
||
|
|
sub.f32 %f260, %f259, %f7;
|
||
|
|
ld.f32 %f261, [%rd41+4];
|
||
|
|
sub.f32 %f262, %f261, %f8;
|
||
|
|
ld.f32 %f263, [%rd41+8];
|
||
|
|
sub.f32 %f264, %f263, %f9;
|
||
|
|
mul.f32 %f265, %f262, %f262;
|
||
|
|
fma.rn.f32 %f266, %f260, %f260, %f265;
|
||
|
|
fma.rn.f32 %f267, %f264, %f264, %f266;
|
||
|
|
sqrt.rn.f32 %f23, %f267;
|
||
|
|
rcp.rn.f32 %f268, %f23;
|
||
|
|
mul.f32 %f24, %f260, %f268;
|
||
|
|
mul.f32 %f25, %f262, %f268;
|
||
|
|
mul.f32 %f26, %f264, %f268;
|
||
|
|
mul.f32 %f269, %f23, %f23;
|
||
|
|
div.rn.f32 %f27, %f13, %f269;
|
||
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd54, %r135;
|
||
|
|
cvt.u64.u32 %rd55, %r134;
|
||
|
|
and.b64 %rd49, %rd55, 255;
|
||
|
|
and.b64 %rd50, %rd54, 255;
|
||
|
|
mov.u64 %rd56, rnd_seeds2;
|
||
|
|
cvta.global.u64 %rd48, %rd56;
|
||
|
|
mov.u32 %r133, 3;
|
||
|
|
// inline asm
|
||
|
|
call (%rd47), _rt_buffer_get_64, (%rd48, %r101, %r133, %rd49, %rd50, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
setp.eq.s32 %p15, %r3, 0;
|
||
|
|
@%p15 bra BB0_8;
|
||
|
|
|
||
|
|
ld.u8 %rs14, [%rd47];
|
||
|
|
cvt.rn.f32.u16 %f270, %rs14;
|
||
|
|
div.rn.f32 %f271, %f270, 0fC37F0000;
|
||
|
|
fma.rn.f32 %f272, %f271, 0f3F333333, 0f3F800000;
|
||
|
|
setp.gt.f32 %p17, %f27, %f272;
|
||
|
|
mov.pred %p148, -1;
|
||
|
|
@%p17 bra BB0_23;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
ld.global.f32 %f275, [lightInvCutoff];
|
||
|
|
mul.f32 %f28, %f23, %f275;
|
||
|
|
mov.f32 %f279, 0f40800000;
|
||
|
|
abs.f32 %f30, %f28;
|
||
|
|
setp.lt.f32 %p18, %f30, 0f00800000;
|
||
|
|
mul.f32 %f281, %f30, 0f4B800000;
|
||
|
|
selp.f32 %f282, 0fC3170000, 0fC2FE0000, %p18;
|
||
|
|
selp.f32 %f283, %f281, %f30, %p18;
|
||
|
|
mov.b32 %r138, %f283;
|
||
|
|
and.b32 %r139, %r138, 8388607;
|
||
|
|
or.b32 %r140, %r139, 1065353216;
|
||
|
|
mov.b32 %f284, %r140;
|
||
|
|
shr.u32 %r141, %r138, 23;
|
||
|
|
cvt.rn.f32.u32 %f285, %r141;
|
||
|
|
add.f32 %f286, %f282, %f285;
|
||
|
|
setp.gt.f32 %p19, %f284, 0f3FB504F3;
|
||
|
|
mul.f32 %f287, %f284, 0f3F000000;
|
||
|
|
add.f32 %f288, %f286, 0f3F800000;
|
||
|
|
selp.f32 %f289, %f287, %f284, %p19;
|
||
|
|
selp.f32 %f290, %f288, %f286, %p19;
|
||
|
|
add.f32 %f291, %f289, 0fBF800000;
|
||
|
|
add.f32 %f274, %f289, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f273,%f274;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f292, %f291, %f291;
|
||
|
|
mul.f32 %f293, %f273, %f292;
|
||
|
|
mul.f32 %f294, %f293, %f293;
|
||
|
|
mov.f32 %f295, 0f3C4CAF63;
|
||
|
|
mov.f32 %f296, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f297, %f296, %f294, %f295;
|
||
|
|
mov.f32 %f298, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f299, %f297, %f294, %f298;
|
||
|
|
mul.rn.f32 %f300, %f299, %f294;
|
||
|
|
mul.rn.f32 %f301, %f300, %f293;
|
||
|
|
sub.f32 %f302, %f291, %f293;
|
||
|
|
neg.f32 %f303, %f293;
|
||
|
|
add.f32 %f304, %f302, %f302;
|
||
|
|
fma.rn.f32 %f305, %f303, %f291, %f304;
|
||
|
|
mul.rn.f32 %f306, %f273, %f305;
|
||
|
|
add.f32 %f307, %f301, %f293;
|
||
|
|
sub.f32 %f308, %f293, %f307;
|
||
|
|
add.f32 %f309, %f301, %f308;
|
||
|
|
add.f32 %f310, %f306, %f309;
|
||
|
|
add.f32 %f311, %f307, %f310;
|
||
|
|
sub.f32 %f312, %f307, %f311;
|
||
|
|
add.f32 %f313, %f310, %f312;
|
||
|
|
mov.f32 %f314, 0f3F317200;
|
||
|
|
mul.rn.f32 %f315, %f290, %f314;
|
||
|
|
mov.f32 %f316, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f317, %f290, %f316;
|
||
|
|
add.f32 %f318, %f315, %f311;
|
||
|
|
sub.f32 %f319, %f315, %f318;
|
||
|
|
add.f32 %f320, %f311, %f319;
|
||
|
|
add.f32 %f321, %f313, %f320;
|
||
|
|
add.f32 %f322, %f317, %f321;
|
||
|
|
add.f32 %f323, %f318, %f322;
|
||
|
|
sub.f32 %f324, %f318, %f323;
|
||
|
|
add.f32 %f325, %f322, %f324;
|
||
|
|
mul.rn.f32 %f326, %f279, %f323;
|
||
|
|
neg.f32 %f327, %f326;
|
||
|
|
fma.rn.f32 %f328, %f279, %f323, %f327;
|
||
|
|
fma.rn.f32 %f329, %f279, %f325, %f328;
|
||
|
|
mov.f32 %f330, 0f00000000;
|
||
|
|
fma.rn.f32 %f331, %f330, %f323, %f329;
|
||
|
|
add.rn.f32 %f332, %f326, %f331;
|
||
|
|
neg.f32 %f333, %f332;
|
||
|
|
add.rn.f32 %f334, %f326, %f333;
|
||
|
|
add.rn.f32 %f335, %f334, %f331;
|
||
|
|
mov.b32 %r142, %f332;
|
||
|
|
setp.eq.s32 %p20, %r142, 1118925336;
|
||
|
|
add.s32 %r143, %r142, -1;
|
||
|
|
mov.b32 %f336, %r143;
|
||
|
|
add.f32 %f337, %f335, 0f37000000;
|
||
|
|
selp.f32 %f338, %f336, %f332, %p20;
|
||
|
|
selp.f32 %f31, %f337, %f335, %p20;
|
||
|
|
mul.f32 %f339, %f338, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f340, %f339;
|
||
|
|
mov.f32 %f341, 0fBF317200;
|
||
|
|
fma.rn.f32 %f342, %f340, %f341, %f338;
|
||
|
|
mov.f32 %f343, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f344, %f340, %f343, %f342;
|
||
|
|
mul.f32 %f345, %f344, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f346, %f345;
|
||
|
|
add.f32 %f347, %f340, 0f00000000;
|
||
|
|
ex2.approx.f32 %f348, %f347;
|
||
|
|
mul.f32 %f349, %f346, %f348;
|
||
|
|
setp.lt.f32 %p21, %f338, 0fC2D20000;
|
||
|
|
selp.f32 %f350, 0f00000000, %f349, %p21;
|
||
|
|
setp.gt.f32 %p22, %f338, 0f42D20000;
|
||
|
|
selp.f32 %f925, 0f7F800000, %f350, %p22;
|
||
|
|
setp.eq.f32 %p23, %f925, 0f7F800000;
|
||
|
|
@%p23 bra BB0_10;
|
||
|
|
|
||
|
|
fma.rn.f32 %f925, %f925, %f31, %f925;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
mov.f32 %f914, 0f40000000;
|
||
|
|
cvt.rzi.f32.f32 %f913, %f914;
|
||
|
|
add.f32 %f912, %f913, %f913;
|
||
|
|
mov.f32 %f911, 0f40800000;
|
||
|
|
sub.f32 %f910, %f911, %f912;
|
||
|
|
abs.f32 %f909, %f910;
|
||
|
|
setp.lt.f32 %p24, %f28, 0f00000000;
|
||
|
|
setp.eq.f32 %p25, %f909, 0f3F800000;
|
||
|
|
and.pred %p1, %p24, %p25;
|
||
|
|
mov.b32 %r144, %f925;
|
||
|
|
xor.b32 %r145, %r144, -2147483648;
|
||
|
|
mov.b32 %f351, %r145;
|
||
|
|
selp.f32 %f927, %f351, %f925, %p1;
|
||
|
|
setp.eq.f32 %p26, %f28, 0f00000000;
|
||
|
|
@%p26 bra BB0_13;
|
||
|
|
bra.uni BB0_11;
|
||
|
|
|
||
|
|
BB0_13:
|
||
|
|
add.f32 %f354, %f28, %f28;
|
||
|
|
selp.f32 %f927, %f354, 0f00000000, %p25;
|
||
|
|
bra.uni BB0_14;
|
||
|
|
|
||
|
|
BB0_11:
|
||
|
|
setp.geu.f32 %p27, %f28, 0f00000000;
|
||
|
|
@%p27 bra BB0_14;
|
||
|
|
|
||
|
|
mov.f32 %f915, 0f40800000;
|
||
|
|
cvt.rzi.f32.f32 %f353, %f915;
|
||
|
|
setp.neu.f32 %p28, %f353, 0f40800000;
|
||
|
|
selp.f32 %f927, 0f7FFFFFFF, %f927, %p28;
|
||
|
|
|
||
|
|
BB0_14:
|
||
|
|
add.f32 %f355, %f30, 0f40800000;
|
||
|
|
mov.b32 %r146, %f355;
|
||
|
|
setp.lt.s32 %p30, %r146, 2139095040;
|
||
|
|
@%p30 bra BB0_19;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p31, %f30, 0f7F800000;
|
||
|
|
@%p31 bra BB0_18;
|
||
|
|
bra.uni BB0_16;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
add.f32 %f927, %f28, 0f40800000;
|
||
|
|
bra.uni BB0_19;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
setp.neu.f32 %p32, %f30, 0f7F800000;
|
||
|
|
@%p32 bra BB0_19;
|
||
|
|
|
||
|
|
selp.f32 %f927, 0fFF800000, 0f7F800000, %p1;
|
||
|
|
|
||
|
|
BB0_19:
|
||
|
|
mov.u32 %r381, 1;
|
||
|
|
mov.u64 %rd302, lightMeshBuffer;
|
||
|
|
cvta.global.u64 %rd301, %rd302;
|
||
|
|
shl.b32 %r380, %r384, 1;
|
||
|
|
mov.f32 %f356, 0f3F800000;
|
||
|
|
sub.f32 %f357, %f356, %f927;
|
||
|
|
setp.eq.f32 %p33, %f28, 0f3F800000;
|
||
|
|
selp.f32 %f358, 0f00000000, %f357, %p33;
|
||
|
|
cvt.sat.f32.f32 %f359, %f358;
|
||
|
|
mul.f32 %f360, %f27, %f359;
|
||
|
|
add.s32 %r149, %r380, 1;
|
||
|
|
cvt.s64.s32 %rd59, %r149;
|
||
|
|
// inline asm
|
||
|
|
call (%rd57), _rt_buffer_get_64, (%rd301, %r381, %r117, %rd59, %rd25, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f361, [%rd57];
|
||
|
|
mul.f32 %f362, %f24, %f361;
|
||
|
|
ld.f32 %f363, [%rd57+4];
|
||
|
|
mul.f32 %f364, %f25, %f363;
|
||
|
|
neg.f32 %f365, %f364;
|
||
|
|
sub.f32 %f366, %f365, %f362;
|
||
|
|
ld.f32 %f367, [%rd57+8];
|
||
|
|
mul.f32 %f368, %f26, %f367;
|
||
|
|
sub.f32 %f369, %f366, %f368;
|
||
|
|
cvt.sat.f32.f32 %f370, %f369;
|
||
|
|
mul.f32 %f42, %f360, %f370;
|
||
|
|
mul.f32 %f371, %f917, %f25;
|
||
|
|
fma.rn.f32 %f372, %f916, %f24, %f371;
|
||
|
|
fma.rn.f32 %f43, %f918, %f26, %f372;
|
||
|
|
setp.leu.f32 %p34, %f42, 0f3727C5AC;
|
||
|
|
@%p34 bra BB0_21;
|
||
|
|
|
||
|
|
mov.u32 %r382, 1;
|
||
|
|
add.u64 %rd64, %SP, 4;
|
||
|
|
cvta.to.local.u64 %rd65, %rd64;
|
||
|
|
max.f32 %f379, %f385, %f249;
|
||
|
|
sub.f32 %f380, %f23, %f379;
|
||
|
|
mov.u32 %r153, 1065353216;
|
||
|
|
st.local.u32 [%rd65], %r153;
|
||
|
|
ld.global.u32 %r150, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r150, %f10, %f11, %f12, %f24, %f25, %f26, %r382, %f379, %f380, %rd64, %r102);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f387, [%rd65];
|
||
|
|
mul.f32 %f388, %f42, 0f3EA2F983;
|
||
|
|
mul.f32 %f389, %f388, %f387;
|
||
|
|
cvt.sat.f32.f32 %f390, %f43;
|
||
|
|
fma.rn.f32 %f22, %f390, %f389, %f22;
|
||
|
|
fma.rn.f32 %f20, %f24, %f389, %f20;
|
||
|
|
fma.rn.f32 %f19, %f25, %f389, %f19;
|
||
|
|
fma.rn.f32 %f18, %f26, %f389, %f18;
|
||
|
|
add.f32 %f21, %f21, %f389;
|
||
|
|
add.f32 %f17, %f17, %f387;
|
||
|
|
|
||
|
|
BB0_21:
|
||
|
|
add.s32 %r384, %r384, 1;
|
||
|
|
setp.lt.s32 %p36, %r384, %r2;
|
||
|
|
@%p36 bra BB0_6;
|
||
|
|
bra.uni BB0_22;
|
||
|
|
|
||
|
|
BB0_4:
|
||
|
|
mov.f32 %f18, %f17;
|
||
|
|
mov.f32 %f19, %f17;
|
||
|
|
mov.f32 %f20, %f17;
|
||
|
|
mov.f32 %f21, %f17;
|
||
|
|
mov.f32 %f22, %f17;
|
||
|
|
|
||
|
|
BB0_22:
|
||
|
|
mov.pred %p148, %p13;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
cvt.rn.f32.s32 %f391, %r384;
|
||
|
|
mov.f32 %f392, 0f3F800000;
|
||
|
|
max.f32 %f393, %f391, %f392;
|
||
|
|
div.rn.f32 %f970, %f22, %f393;
|
||
|
|
div.rn.f32 %f975, %f17, %f393;
|
||
|
|
div.rn.f32 %f971, %f21, %f393;
|
||
|
|
div.rn.f32 %f972, %f20, %f393;
|
||
|
|
div.rn.f32 %f973, %f19, %f393;
|
||
|
|
div.rn.f32 %f974, %f18, %f393;
|
||
|
|
@!%p148 bra BB0_76;
|
||
|
|
bra.uni BB0_24;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
mov.f32 %f946, 0f00000000;
|
||
|
|
setp.lt.s32 %p37, %r3, 1;
|
||
|
|
mov.f32 %f947, %f946;
|
||
|
|
mov.f32 %f948, %f946;
|
||
|
|
mov.f32 %f949, %f946;
|
||
|
|
mov.f32 %f950, %f946;
|
||
|
|
mov.f32 %f951, %f946;
|
||
|
|
@%p37 bra BB0_75;
|
||
|
|
|
||
|
|
cvt.rn.f32.s32 %f406, %r4;
|
||
|
|
rcp.rn.f32 %f68, %f406;
|
||
|
|
add.u64 %rd66, %SP, 8;
|
||
|
|
cvta.to.local.u64 %rd3, %rd66;
|
||
|
|
mul.f32 %f69, %f10, 0f3456BF95;
|
||
|
|
mul.f32 %f70, %f11, 0f3456BF95;
|
||
|
|
mul.f32 %f71, %f12, 0f3456BF95;
|
||
|
|
add.u64 %rd67, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd4, %rd67;
|
||
|
|
mov.f32 %f946, 0f00000000;
|
||
|
|
mov.u32 %r154, 0;
|
||
|
|
abs.f32 %f407, %f70;
|
||
|
|
abs.f32 %f408, %f69;
|
||
|
|
max.f32 %f409, %f408, %f407;
|
||
|
|
abs.f32 %f410, %f71;
|
||
|
|
max.f32 %f411, %f409, %f410;
|
||
|
|
mov.u32 %r385, %r154;
|
||
|
|
mov.f32 %f947, %f946;
|
||
|
|
mov.f32 %f948, %f946;
|
||
|
|
mov.f32 %f949, %f946;
|
||
|
|
mov.f32 %f950, %f946;
|
||
|
|
mov.f32 %f951, %f946;
|
||
|
|
|
||
|
|
BB0_26:
|
||
|
|
cvt.rn.f32.s32 %f78, %r385;
|
||
|
|
max.f32 %f79, %f411, %f249;
|
||
|
|
mov.u32 %r387, %r154;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
mad.lo.s32 %r156, %r388, 1664525, 1013904223;
|
||
|
|
and.b32 %r157, %r156, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f413, %r157;
|
||
|
|
fma.rn.f32 %f414, %f413, 0f33800000, %f78;
|
||
|
|
mul.f32 %f415, %f68, %f414;
|
||
|
|
mad.lo.s32 %r388, %r156, 1664525, 1013904223;
|
||
|
|
and.b32 %r158, %r388, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f416, %r158;
|
||
|
|
cvt.rn.f32.s32 %f417, %r387;
|
||
|
|
fma.rn.f32 %f418, %f416, 0f33800000, %f417;
|
||
|
|
mul.f32 %f419, %f68, %f418;
|
||
|
|
fma.rn.f32 %f86, %f415, 0fC0000000, 0f3F800000;
|
||
|
|
mul.f32 %f420, %f86, %f86;
|
||
|
|
sub.f32 %f422, %f392, %f420;
|
||
|
|
mov.f32 %f423, 0f00000000;
|
||
|
|
max.f32 %f424, %f423, %f422;
|
||
|
|
sqrt.rn.f32 %f87, %f424;
|
||
|
|
mul.f32 %f958, %f419, 0f40C90FDB;
|
||
|
|
abs.f32 %f89, %f958;
|
||
|
|
setp.neu.f32 %p38, %f89, 0f7F800000;
|
||
|
|
mov.f32 %f952, %f958;
|
||
|
|
@%p38 bra BB0_29;
|
||
|
|
|
||
|
|
mul.rn.f32 %f952, %f958, %f423;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
mul.f32 %f426, %f952, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r398, %f426;
|
||
|
|
cvt.rn.f32.s32 %f427, %r398;
|
||
|
|
neg.f32 %f428, %f427;
|
||
|
|
mov.f32 %f429, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f430, %f428, %f429, %f952;
|
||
|
|
mov.f32 %f431, 0f33A22168;
|
||
|
|
fma.rn.f32 %f432, %f428, %f431, %f430;
|
||
|
|
mov.f32 %f433, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f953, %f428, %f433, %f432;
|
||
|
|
abs.f32 %f434, %f952;
|
||
|
|
setp.leu.f32 %p39, %f434, 0f47CE4780;
|
||
|
|
@%p39 bra BB0_40;
|
||
|
|
|
||
|
|
mov.b32 %r19, %f952;
|
||
|
|
shr.u32 %r20, %r19, 23;
|
||
|
|
shl.b32 %r161, %r19, 8;
|
||
|
|
or.b32 %r21, %r161, -2147483648;
|
||
|
|
mov.u32 %r390, 0;
|
||
|
|
mov.u64 %rd303, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r389, -6;
|
||
|
|
mov.u64 %rd304, %rd3;
|
||
|
|
|
||
|
|
BB0_31:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r164, [%rd303];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r162, %r164, %r21, %r390;
|
||
|
|
madc.hi.u32 %r390, %r164, %r21, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd304], %r162;
|
||
|
|
add.s64 %rd304, %rd304, 4;
|
||
|
|
add.s64 %rd303, %rd303, 4;
|
||
|
|
add.s32 %r389, %r389, 1;
|
||
|
|
setp.ne.s32 %p40, %r389, 0;
|
||
|
|
@%p40 bra BB0_31;
|
||
|
|
|
||
|
|
and.b32 %r167, %r20, 255;
|
||
|
|
add.s32 %r168, %r167, -128;
|
||
|
|
shr.u32 %r169, %r168, 5;
|
||
|
|
and.b32 %r26, %r19, -2147483648;
|
||
|
|
st.local.u32 [%rd3+24], %r390;
|
||
|
|
mov.u32 %r170, 6;
|
||
|
|
sub.s32 %r171, %r170, %r169;
|
||
|
|
mul.wide.s32 %rd69, %r171, 4;
|
||
|
|
add.s64 %rd9, %rd3, %rd69;
|
||
|
|
ld.local.u32 %r391, [%rd9];
|
||
|
|
ld.local.u32 %r392, [%rd9+-4];
|
||
|
|
and.b32 %r29, %r20, 31;
|
||
|
|
setp.eq.s32 %p41, %r29, 0;
|
||
|
|
@%p41 bra BB0_34;
|
||
|
|
|
||
|
|
mov.u32 %r172, 32;
|
||
|
|
sub.s32 %r173, %r172, %r29;
|
||
|
|
shr.u32 %r174, %r392, %r173;
|
||
|
|
shl.b32 %r175, %r391, %r29;
|
||
|
|
add.s32 %r391, %r174, %r175;
|
||
|
|
ld.local.u32 %r176, [%rd9+-8];
|
||
|
|
shr.u32 %r177, %r176, %r173;
|
||
|
|
shl.b32 %r178, %r392, %r29;
|
||
|
|
add.s32 %r392, %r177, %r178;
|
||
|
|
|
||
|
|
BB0_34:
|
||
|
|
shr.u32 %r179, %r392, 30;
|
||
|
|
shl.b32 %r180, %r391, 2;
|
||
|
|
add.s32 %r393, %r179, %r180;
|
||
|
|
shl.b32 %r35, %r392, 2;
|
||
|
|
shr.u32 %r181, %r393, 31;
|
||
|
|
shr.u32 %r182, %r391, 30;
|
||
|
|
add.s32 %r36, %r181, %r182;
|
||
|
|
setp.eq.s32 %p42, %r181, 0;
|
||
|
|
@%p42 bra BB0_35;
|
||
|
|
bra.uni BB0_36;
|
||
|
|
|
||
|
|
BB0_35:
|
||
|
|
mov.u32 %r394, %r26;
|
||
|
|
mov.u32 %r395, %r35;
|
||
|
|
bra.uni BB0_37;
|
||
|
|
|
||
|
|
BB0_36:
|
||
|
|
not.b32 %r183, %r393;
|
||
|
|
neg.s32 %r395, %r35;
|
||
|
|
setp.eq.s32 %p43, %r35, 0;
|
||
|
|
selp.u32 %r184, 1, 0, %p43;
|
||
|
|
add.s32 %r393, %r184, %r183;
|
||
|
|
xor.b32 %r394, %r26, -2147483648;
|
||
|
|
|
||
|
|
BB0_37:
|
||
|
|
clz.b32 %r397, %r393;
|
||
|
|
setp.eq.s32 %p44, %r397, 0;
|
||
|
|
shl.b32 %r185, %r393, %r397;
|
||
|
|
mov.u32 %r186, 32;
|
||
|
|
sub.s32 %r187, %r186, %r397;
|
||
|
|
shr.u32 %r188, %r395, %r187;
|
||
|
|
add.s32 %r189, %r188, %r185;
|
||
|
|
selp.b32 %r44, %r393, %r189, %p44;
|
||
|
|
mov.u32 %r190, -921707870;
|
||
|
|
mul.hi.u32 %r396, %r44, %r190;
|
||
|
|
setp.eq.s32 %p45, %r26, 0;
|
||
|
|
neg.s32 %r191, %r36;
|
||
|
|
selp.b32 %r398, %r36, %r191, %p45;
|
||
|
|
setp.lt.s32 %p46, %r396, 1;
|
||
|
|
@%p46 bra BB0_39;
|
||
|
|
|
||
|
|
mul.lo.s32 %r192, %r44, -921707870;
|
||
|
|
shr.u32 %r193, %r192, 31;
|
||
|
|
shl.b32 %r194, %r396, 1;
|
||
|
|
add.s32 %r396, %r193, %r194;
|
||
|
|
add.s32 %r397, %r397, 1;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
mov.u32 %r195, 126;
|
||
|
|
sub.s32 %r196, %r195, %r397;
|
||
|
|
shl.b32 %r197, %r196, 23;
|
||
|
|
add.s32 %r198, %r396, 1;
|
||
|
|
shr.u32 %r199, %r198, 7;
|
||
|
|
add.s32 %r200, %r199, 1;
|
||
|
|
shr.u32 %r201, %r200, 1;
|
||
|
|
add.s32 %r202, %r201, %r197;
|
||
|
|
or.b32 %r203, %r202, %r394;
|
||
|
|
mov.b32 %f953, %r203;
|
||
|
|
|
||
|
|
BB0_40:
|
||
|
|
mul.rn.f32 %f95, %f953, %f953;
|
||
|
|
add.s32 %r52, %r398, 1;
|
||
|
|
and.b32 %r53, %r52, 1;
|
||
|
|
setp.eq.s32 %p47, %r53, 0;
|
||
|
|
@%p47 bra BB0_42;
|
||
|
|
bra.uni BB0_41;
|
||
|
|
|
||
|
|
BB0_42:
|
||
|
|
mov.f32 %f437, 0f3C08839E;
|
||
|
|
mov.f32 %f438, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f954, %f438, %f95, %f437;
|
||
|
|
bra.uni BB0_43;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
mov.f32 %f435, 0fBAB6061A;
|
||
|
|
mov.f32 %f436, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f954, %f436, %f95, %f435;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
@%p47 bra BB0_45;
|
||
|
|
bra.uni BB0_44;
|
||
|
|
|
||
|
|
BB0_45:
|
||
|
|
mov.f32 %f442, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f443, %f954, %f95, %f442;
|
||
|
|
fma.rn.f32 %f955, %f443, %f95, %f423;
|
||
|
|
bra.uni BB0_46;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
mov.f32 %f439, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f440, %f954, %f95, %f439;
|
||
|
|
mov.f32 %f441, 0fBF000000;
|
||
|
|
fma.rn.f32 %f955, %f440, %f95, %f441;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
fma.rn.f32 %f956, %f955, %f953, %f953;
|
||
|
|
@%p47 bra BB0_48;
|
||
|
|
|
||
|
|
fma.rn.f32 %f956, %f955, %f95, %f392;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
and.b32 %r204, %r52, 2;
|
||
|
|
setp.eq.s32 %p50, %r204, 0;
|
||
|
|
@%p50 bra BB0_50;
|
||
|
|
|
||
|
|
mov.f32 %f447, 0fBF800000;
|
||
|
|
fma.rn.f32 %f956, %f956, %f447, %f423;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
@%p38 bra BB0_52;
|
||
|
|
|
||
|
|
mul.rn.f32 %f958, %f958, %f423;
|
||
|
|
|
||
|
|
BB0_52:
|
||
|
|
mul.f32 %f449, %f958, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r408, %f449;
|
||
|
|
cvt.rn.f32.s32 %f450, %r408;
|
||
|
|
neg.f32 %f451, %f450;
|
||
|
|
fma.rn.f32 %f453, %f451, %f429, %f958;
|
||
|
|
fma.rn.f32 %f455, %f451, %f431, %f453;
|
||
|
|
fma.rn.f32 %f959, %f451, %f433, %f455;
|
||
|
|
abs.f32 %f457, %f958;
|
||
|
|
setp.leu.f32 %p52, %f457, 0f47CE4780;
|
||
|
|
@%p52 bra BB0_63;
|
||
|
|
|
||
|
|
mov.b32 %r55, %f958;
|
||
|
|
shr.u32 %r56, %r55, 23;
|
||
|
|
shl.b32 %r207, %r55, 8;
|
||
|
|
or.b32 %r57, %r207, -2147483648;
|
||
|
|
mov.u32 %r400, 0;
|
||
|
|
mov.u64 %rd305, __cudart_i2opi_f;
|
||
|
|
mov.u32 %r399, -6;
|
||
|
|
mov.u64 %rd306, %rd3;
|
||
|
|
|
||
|
|
BB0_54:
|
||
|
|
.pragma "nounroll";
|
||
|
|
ld.const.u32 %r210, [%rd305];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r208, %r210, %r57, %r400;
|
||
|
|
madc.hi.u32 %r400, %r210, %r57, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd306], %r208;
|
||
|
|
add.s64 %rd306, %rd306, 4;
|
||
|
|
add.s64 %rd305, %rd305, 4;
|
||
|
|
add.s32 %r399, %r399, 1;
|
||
|
|
setp.ne.s32 %p53, %r399, 0;
|
||
|
|
@%p53 bra BB0_54;
|
||
|
|
|
||
|
|
and.b32 %r213, %r56, 255;
|
||
|
|
add.s32 %r214, %r213, -128;
|
||
|
|
shr.u32 %r215, %r214, 5;
|
||
|
|
and.b32 %r62, %r55, -2147483648;
|
||
|
|
st.local.u32 [%rd3+24], %r400;
|
||
|
|
mov.u32 %r216, 6;
|
||
|
|
sub.s32 %r217, %r216, %r215;
|
||
|
|
mul.wide.s32 %rd71, %r217, 4;
|
||
|
|
add.s64 %rd14, %rd3, %rd71;
|
||
|
|
ld.local.u32 %r401, [%rd14];
|
||
|
|
ld.local.u32 %r402, [%rd14+-4];
|
||
|
|
and.b32 %r65, %r56, 31;
|
||
|
|
setp.eq.s32 %p54, %r65, 0;
|
||
|
|
@%p54 bra BB0_57;
|
||
|
|
|
||
|
|
mov.u32 %r218, 32;
|
||
|
|
sub.s32 %r219, %r218, %r65;
|
||
|
|
shr.u32 %r220, %r402, %r219;
|
||
|
|
shl.b32 %r221, %r401, %r65;
|
||
|
|
add.s32 %r401, %r220, %r221;
|
||
|
|
ld.local.u32 %r222, [%rd14+-8];
|
||
|
|
shr.u32 %r223, %r222, %r219;
|
||
|
|
shl.b32 %r224, %r402, %r65;
|
||
|
|
add.s32 %r402, %r223, %r224;
|
||
|
|
|
||
|
|
BB0_57:
|
||
|
|
shr.u32 %r225, %r402, 30;
|
||
|
|
shl.b32 %r226, %r401, 2;
|
||
|
|
add.s32 %r403, %r225, %r226;
|
||
|
|
shl.b32 %r71, %r402, 2;
|
||
|
|
shr.u32 %r227, %r403, 31;
|
||
|
|
shr.u32 %r228, %r401, 30;
|
||
|
|
add.s32 %r72, %r227, %r228;
|
||
|
|
setp.eq.s32 %p55, %r227, 0;
|
||
|
|
@%p55 bra BB0_58;
|
||
|
|
bra.uni BB0_59;
|
||
|
|
|
||
|
|
BB0_58:
|
||
|
|
mov.u32 %r404, %r62;
|
||
|
|
mov.u32 %r405, %r71;
|
||
|
|
bra.uni BB0_60;
|
||
|
|
|
||
|
|
BB0_59:
|
||
|
|
not.b32 %r229, %r403;
|
||
|
|
neg.s32 %r405, %r71;
|
||
|
|
setp.eq.s32 %p56, %r71, 0;
|
||
|
|
selp.u32 %r230, 1, 0, %p56;
|
||
|
|
add.s32 %r403, %r230, %r229;
|
||
|
|
xor.b32 %r404, %r62, -2147483648;
|
||
|
|
|
||
|
|
BB0_60:
|
||
|
|
clz.b32 %r407, %r403;
|
||
|
|
setp.eq.s32 %p57, %r407, 0;
|
||
|
|
shl.b32 %r231, %r403, %r407;
|
||
|
|
mov.u32 %r232, 32;
|
||
|
|
sub.s32 %r233, %r232, %r407;
|
||
|
|
shr.u32 %r234, %r405, %r233;
|
||
|
|
add.s32 %r235, %r234, %r231;
|
||
|
|
selp.b32 %r80, %r403, %r235, %p57;
|
||
|
|
mov.u32 %r236, -921707870;
|
||
|
|
mul.hi.u32 %r406, %r80, %r236;
|
||
|
|
setp.eq.s32 %p58, %r62, 0;
|
||
|
|
neg.s32 %r237, %r72;
|
||
|
|
selp.b32 %r408, %r72, %r237, %p58;
|
||
|
|
setp.lt.s32 %p59, %r406, 1;
|
||
|
|
@%p59 bra BB0_62;
|
||
|
|
|
||
|
|
mul.lo.s32 %r238, %r80, -921707870;
|
||
|
|
shr.u32 %r239, %r238, 31;
|
||
|
|
shl.b32 %r240, %r406, 1;
|
||
|
|
add.s32 %r406, %r239, %r240;
|
||
|
|
add.s32 %r407, %r407, 1;
|
||
|
|
|
||
|
|
BB0_62:
|
||
|
|
mov.u32 %r241, 126;
|
||
|
|
sub.s32 %r242, %r241, %r407;
|
||
|
|
shl.b32 %r243, %r242, 23;
|
||
|
|
add.s32 %r244, %r406, 1;
|
||
|
|
shr.u32 %r245, %r244, 7;
|
||
|
|
add.s32 %r246, %r245, 1;
|
||
|
|
shr.u32 %r247, %r246, 1;
|
||
|
|
add.s32 %r248, %r247, %r243;
|
||
|
|
or.b32 %r249, %r248, %r404;
|
||
|
|
mov.b32 %f959, %r249;
|
||
|
|
|
||
|
|
BB0_63:
|
||
|
|
mul.rn.f32 %f112, %f959, %f959;
|
||
|
|
and.b32 %r88, %r408, 1;
|
||
|
|
setp.eq.s32 %p60, %r88, 0;
|
||
|
|
@%p60 bra BB0_65;
|
||
|
|
bra.uni BB0_64;
|
||
|
|
|
||
|
|
BB0_65:
|
||
|
|
mov.f32 %f460, 0f3C08839E;
|
||
|
|
mov.f32 %f461, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f960, %f461, %f112, %f460;
|
||
|
|
bra.uni BB0_66;
|
||
|
|
|
||
|
|
BB0_64:
|
||
|
|
mov.f32 %f458, 0fBAB6061A;
|
||
|
|
mov.f32 %f459, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f960, %f459, %f112, %f458;
|
||
|
|
|
||
|
|
BB0_66:
|
||
|
|
@%p60 bra BB0_68;
|
||
|
|
bra.uni BB0_67;
|
||
|
|
|
||
|
|
BB0_68:
|
||
|
|
mov.f32 %f465, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f466, %f960, %f112, %f465;
|
||
|
|
fma.rn.f32 %f961, %f466, %f112, %f423;
|
||
|
|
bra.uni BB0_69;
|
||
|
|
|
||
|
|
BB0_67:
|
||
|
|
mov.f32 %f462, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f463, %f960, %f112, %f462;
|
||
|
|
mov.f32 %f464, 0fBF000000;
|
||
|
|
fma.rn.f32 %f961, %f463, %f112, %f464;
|
||
|
|
|
||
|
|
BB0_69:
|
||
|
|
fma.rn.f32 %f962, %f961, %f959, %f959;
|
||
|
|
@%p60 bra BB0_71;
|
||
|
|
|
||
|
|
fma.rn.f32 %f962, %f961, %f112, %f392;
|
||
|
|
|
||
|
|
BB0_71:
|
||
|
|
and.b32 %r250, %r408, 2;
|
||
|
|
setp.eq.s32 %p63, %r250, 0;
|
||
|
|
@%p63 bra BB0_73;
|
||
|
|
|
||
|
|
mov.f32 %f470, 0fBF800000;
|
||
|
|
fma.rn.f32 %f962, %f962, %f470, %f423;
|
||
|
|
|
||
|
|
BB0_73:
|
||
|
|
mul.f32 %f474, %f87, %f956;
|
||
|
|
mov.u32 %r252, 0;
|
||
|
|
st.local.u32 [%rd4], %r252;
|
||
|
|
ld.global.u32 %r251, [root];
|
||
|
|
mul.f32 %f475, %f87, %f962;
|
||
|
|
mov.f32 %f478, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f474, %f475, %f86, %r252, %f79, %f478, %rd67, %r102);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f479, [%rd4];
|
||
|
|
setp.lt.f32 %p64, %f479, 0f00000000;
|
||
|
|
selp.f32 %f480, 0f00000000, %f479, %p64;
|
||
|
|
selp.f32 %f481, 0f00000000, 0f3F800000, %p64;
|
||
|
|
add.f32 %f946, %f946, %f481;
|
||
|
|
fma.rn.f32 %f949, %f474, %f480, %f949;
|
||
|
|
fma.rn.f32 %f948, %f475, %f480, %f948;
|
||
|
|
fma.rn.f32 %f947, %f86, %f480, %f947;
|
||
|
|
add.f32 %f950, %f950, %f480;
|
||
|
|
mul.f32 %f482, %f917, %f475;
|
||
|
|
fma.rn.f32 %f483, %f916, %f474, %f482;
|
||
|
|
fma.rn.f32 %f484, %f918, %f86, %f483;
|
||
|
|
cvt.sat.f32.f32 %f485, %f484;
|
||
|
|
fma.rn.f32 %f951, %f480, %f485, %f951;
|
||
|
|
add.s32 %r387, %r387, 1;
|
||
|
|
setp.lt.s32 %p65, %r387, %r4;
|
||
|
|
@%p65 bra BB0_27;
|
||
|
|
|
||
|
|
add.s32 %r385, %r385, 1;
|
||
|
|
setp.lt.s32 %p66, %r385, %r4;
|
||
|
|
@%p66 bra BB0_26;
|
||
|
|
|
||
|
|
BB0_75:
|
||
|
|
mul.lo.s32 %r254, %r4, %r4;
|
||
|
|
cvt.rn.f32.s32 %f486, %r254;
|
||
|
|
div.rn.f32 %f487, %f951, %f486;
|
||
|
|
div.rn.f32 %f975, %f946, %f486;
|
||
|
|
div.rn.f32 %f971, %f950, %f486;
|
||
|
|
div.rn.f32 %f972, %f949, %f486;
|
||
|
|
div.rn.f32 %f973, %f948, %f486;
|
||
|
|
div.rn.f32 %f974, %f947, %f486;
|
||
|
|
add.f32 %f970, %f487, %f487;
|
||
|
|
|
||
|
|
BB0_76:
|
||
|
|
setp.gt.s32 %p67, %r1, -1;
|
||
|
|
selp.f32 %f148, %f975, %f970, %p67;
|
||
|
|
ld.global.u32 %r410, [imageEnabled];
|
||
|
|
and.b32 %r255, %r410, 8;
|
||
|
|
setp.eq.s32 %p68, %r255, 0;
|
||
|
|
@%p68 bra BB0_89;
|
||
|
|
|
||
|
|
cvt.u64.u32 %rd75, %r6;
|
||
|
|
cvt.u64.u32 %rd76, %r7;
|
||
|
|
mov.u64 %rd79, image_Mask;
|
||
|
|
cvta.global.u64 %rd74, %rd79;
|
||
|
|
// inline asm
|
||
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r101, %r101, %rd75, %rd76, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f490, 0f3E68BA2E;
|
||
|
|
cvt.rzi.f32.f32 %f491, %f490;
|
||
|
|
fma.rn.f32 %f492, %f491, 0fC0000000, 0f3EE8BA2E;
|
||
|
|
abs.f32 %f149, %f492;
|
||
|
|
abs.f32 %f150, %f148;
|
||
|
|
setp.lt.f32 %p69, %f150, 0f00800000;
|
||
|
|
mul.f32 %f493, %f150, 0f4B800000;
|
||
|
|
selp.f32 %f494, 0fC3170000, 0fC2FE0000, %p69;
|
||
|
|
selp.f32 %f495, %f493, %f150, %p69;
|
||
|
|
mov.b32 %r258, %f495;
|
||
|
|
and.b32 %r259, %r258, 8388607;
|
||
|
|
or.b32 %r260, %r259, 1065353216;
|
||
|
|
mov.b32 %f496, %r260;
|
||
|
|
shr.u32 %r261, %r258, 23;
|
||
|
|
cvt.rn.f32.u32 %f497, %r261;
|
||
|
|
add.f32 %f498, %f494, %f497;
|
||
|
|
setp.gt.f32 %p70, %f496, 0f3FB504F3;
|
||
|
|
mul.f32 %f499, %f496, 0f3F000000;
|
||
|
|
add.f32 %f500, %f498, 0f3F800000;
|
||
|
|
selp.f32 %f501, %f499, %f496, %p70;
|
||
|
|
selp.f32 %f502, %f500, %f498, %p70;
|
||
|
|
add.f32 %f503, %f501, 0fBF800000;
|
||
|
|
add.f32 %f489, %f501, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f488,%f489;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f504, %f503, %f503;
|
||
|
|
mul.f32 %f505, %f488, %f504;
|
||
|
|
mul.f32 %f506, %f505, %f505;
|
||
|
|
mov.f32 %f507, 0f3C4CAF63;
|
||
|
|
mov.f32 %f508, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f509, %f508, %f506, %f507;
|
||
|
|
mov.f32 %f510, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f511, %f509, %f506, %f510;
|
||
|
|
mul.rn.f32 %f512, %f511, %f506;
|
||
|
|
mul.rn.f32 %f513, %f512, %f505;
|
||
|
|
sub.f32 %f514, %f503, %f505;
|
||
|
|
neg.f32 %f515, %f505;
|
||
|
|
add.f32 %f516, %f514, %f514;
|
||
|
|
fma.rn.f32 %f517, %f515, %f503, %f516;
|
||
|
|
mul.rn.f32 %f518, %f488, %f517;
|
||
|
|
add.f32 %f519, %f513, %f505;
|
||
|
|
sub.f32 %f520, %f505, %f519;
|
||
|
|
add.f32 %f521, %f513, %f520;
|
||
|
|
add.f32 %f522, %f518, %f521;
|
||
|
|
add.f32 %f523, %f519, %f522;
|
||
|
|
sub.f32 %f524, %f519, %f523;
|
||
|
|
add.f32 %f525, %f522, %f524;
|
||
|
|
mov.f32 %f526, 0f3F317200;
|
||
|
|
mul.rn.f32 %f527, %f502, %f526;
|
||
|
|
mov.f32 %f528, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f529, %f502, %f528;
|
||
|
|
add.f32 %f530, %f527, %f523;
|
||
|
|
sub.f32 %f531, %f527, %f530;
|
||
|
|
add.f32 %f532, %f523, %f531;
|
||
|
|
add.f32 %f533, %f525, %f532;
|
||
|
|
add.f32 %f534, %f529, %f533;
|
||
|
|
add.f32 %f535, %f530, %f534;
|
||
|
|
sub.f32 %f536, %f530, %f535;
|
||
|
|
add.f32 %f537, %f534, %f536;
|
||
|
|
mov.f32 %f538, 0f3EE8BA2E;
|
||
|
|
mul.rn.f32 %f539, %f538, %f535;
|
||
|
|
neg.f32 %f540, %f539;
|
||
|
|
fma.rn.f32 %f541, %f538, %f535, %f540;
|
||
|
|
fma.rn.f32 %f542, %f538, %f537, %f541;
|
||
|
|
mov.f32 %f543, 0f00000000;
|
||
|
|
fma.rn.f32 %f544, %f543, %f535, %f542;
|
||
|
|
add.rn.f32 %f545, %f539, %f544;
|
||
|
|
neg.f32 %f546, %f545;
|
||
|
|
add.rn.f32 %f547, %f539, %f546;
|
||
|
|
add.rn.f32 %f548, %f547, %f544;
|
||
|
|
mov.b32 %r262, %f545;
|
||
|
|
setp.eq.s32 %p71, %r262, 1118925336;
|
||
|
|
add.s32 %r263, %r262, -1;
|
||
|
|
mov.b32 %f549, %r263;
|
||
|
|
add.f32 %f550, %f548, 0f37000000;
|
||
|
|
selp.f32 %f551, %f549, %f545, %p71;
|
||
|
|
selp.f32 %f151, %f550, %f548, %p71;
|
||
|
|
mul.f32 %f552, %f551, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f553, %f552;
|
||
|
|
mov.f32 %f554, 0fBF317200;
|
||
|
|
fma.rn.f32 %f555, %f553, %f554, %f551;
|
||
|
|
mov.f32 %f556, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f557, %f553, %f556, %f555;
|
||
|
|
mul.f32 %f558, %f557, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f559, %f558;
|
||
|
|
add.f32 %f560, %f553, 0f00000000;
|
||
|
|
ex2.approx.f32 %f561, %f560;
|
||
|
|
mul.f32 %f562, %f559, %f561;
|
||
|
|
setp.lt.f32 %p72, %f551, 0fC2D20000;
|
||
|
|
selp.f32 %f563, 0f00000000, %f562, %p72;
|
||
|
|
setp.gt.f32 %p73, %f551, 0f42D20000;
|
||
|
|
selp.f32 %f976, 0f7F800000, %f563, %p73;
|
||
|
|
setp.eq.f32 %p74, %f976, 0f7F800000;
|
||
|
|
@%p74 bra BB0_79;
|
||
|
|
|
||
|
|
fma.rn.f32 %f976, %f976, %f151, %f976;
|
||
|
|
|
||
|
|
BB0_79:
|
||
|
|
setp.lt.f32 %p75, %f148, 0f00000000;
|
||
|
|
setp.eq.f32 %p76, %f149, 0f3F800000;
|
||
|
|
and.pred %p3, %p75, %p76;
|
||
|
|
mov.b32 %r264, %f976;
|
||
|
|
xor.b32 %r265, %r264, -2147483648;
|
||
|
|
mov.b32 %f564, %r265;
|
||
|
|
selp.f32 %f978, %f564, %f976, %p3;
|
||
|
|
setp.eq.f32 %p77, %f148, 0f00000000;
|
||
|
|
@%p77 bra BB0_82;
|
||
|
|
bra.uni BB0_80;
|
||
|
|
|
||
|
|
BB0_82:
|
||
|
|
add.f32 %f567, %f148, %f148;
|
||
|
|
selp.f32 %f978, %f567, 0f00000000, %p76;
|
||
|
|
bra.uni BB0_83;
|
||
|
|
|
||
|
|
BB0_147:
|
||
|
|
mov.u64 %rd220, image_HDR;
|
||
|
|
cvta.global.u64 %rd215, %rd220;
|
||
|
|
mov.u32 %r351, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd214), _rt_buffer_get_64, (%rd215, %r101, %r351, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f880, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs99, %f880;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs100, 0;
|
||
|
|
st.v4.u16 [%rd214], {%rs99, %rs99, %rs99, %rs100};
|
||
|
|
|
||
|
|
BB0_148:
|
||
|
|
ld.global.u32 %r352, [additive];
|
||
|
|
setp.eq.s32 %p144, %r352, 0;
|
||
|
|
@%p144 bra BB0_150;
|
||
|
|
|
||
|
|
mov.u64 %rd233, image_RNM0;
|
||
|
|
cvta.global.u64 %rd222, %rd233;
|
||
|
|
mov.u32 %r356, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd221), _rt_buffer_get_64, (%rd222, %r101, %r356, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs107, %rs108, %rs109, %rs110}, [%rd221];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f881, %rs107;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f882, %rs108;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f883, %rs109;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd227), _rt_buffer_get_64, (%rd222, %r101, %r356, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f884, %f881, 0f00000000;
|
||
|
|
add.f32 %f885, %f882, 0f00000000;
|
||
|
|
add.f32 %f886, %f883, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs106, %f886;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs105, %f885;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs104, %f884;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs111, 0;
|
||
|
|
st.v4.u16 [%rd227], {%rs104, %rs105, %rs106, %rs111};
|
||
|
|
bra.uni BB0_151;
|
||
|
|
|
||
|
|
BB0_150:
|
||
|
|
mov.u64 %rd240, image_RNM0;
|
||
|
|
cvta.global.u64 %rd235, %rd240;
|
||
|
|
mov.u32 %r358, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd234), _rt_buffer_get_64, (%rd235, %r101, %r358, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f887, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs112, %f887;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs113, 0;
|
||
|
|
st.v4.u16 [%rd234], {%rs112, %rs112, %rs112, %rs113};
|
||
|
|
|
||
|
|
BB0_151:
|
||
|
|
ld.global.u32 %r359, [additive];
|
||
|
|
setp.eq.s32 %p145, %r359, 0;
|
||
|
|
@%p145 bra BB0_153;
|
||
|
|
|
||
|
|
mov.u64 %rd253, image_RNM1;
|
||
|
|
cvta.global.u64 %rd242, %rd253;
|
||
|
|
mov.u32 %r363, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd241), _rt_buffer_get_64, (%rd242, %r101, %r363, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs120, %rs121, %rs122, %rs123}, [%rd241];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f888, %rs120;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f889, %rs121;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f890, %rs122;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd247), _rt_buffer_get_64, (%rd242, %r101, %r363, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f891, %f888, 0f00000000;
|
||
|
|
add.f32 %f892, %f889, 0f00000000;
|
||
|
|
add.f32 %f893, %f890, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs119, %f893;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs118, %f892;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs117, %f891;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs124, 0;
|
||
|
|
st.v4.u16 [%rd247], {%rs117, %rs118, %rs119, %rs124};
|
||
|
|
bra.uni BB0_154;
|
||
|
|
|
||
|
|
BB0_153:
|
||
|
|
mov.u64 %rd260, image_RNM1;
|
||
|
|
cvta.global.u64 %rd255, %rd260;
|
||
|
|
mov.u32 %r365, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd254), _rt_buffer_get_64, (%rd255, %r101, %r365, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f894, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs125, %f894;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs126, 0;
|
||
|
|
st.v4.u16 [%rd254], {%rs125, %rs125, %rs125, %rs126};
|
||
|
|
|
||
|
|
BB0_154:
|
||
|
|
ld.global.u32 %r366, [additive];
|
||
|
|
setp.eq.s32 %p146, %r366, 0;
|
||
|
|
@%p146 bra BB0_156;
|
||
|
|
|
||
|
|
mov.u64 %rd273, image_RNM2;
|
||
|
|
cvta.global.u64 %rd262, %rd273;
|
||
|
|
mov.u32 %r370, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd261), _rt_buffer_get_64, (%rd262, %r101, %r370, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs133, %rs134, %rs135, %rs136}, [%rd261];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f895, %rs133;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f896, %rs134;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f897, %rs135;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd267), _rt_buffer_get_64, (%rd262, %r101, %r370, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f898, %f895, 0f00000000;
|
||
|
|
add.f32 %f899, %f896, 0f00000000;
|
||
|
|
add.f32 %f900, %f897, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs132, %f900;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs131, %f899;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs130, %f898;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs137, 0;
|
||
|
|
st.v4.u16 [%rd267], {%rs130, %rs131, %rs132, %rs137};
|
||
|
|
bra.uni BB0_157;
|
||
|
|
|
||
|
|
BB0_156:
|
||
|
|
mov.u64 %rd280, image_RNM2;
|
||
|
|
cvta.global.u64 %rd275, %rd280;
|
||
|
|
mov.u32 %r372, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd274), _rt_buffer_get_64, (%rd275, %r101, %r372, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f901, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs138, %f901;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs139, 0;
|
||
|
|
st.v4.u16 [%rd274], {%rs138, %rs138, %rs138, %rs139};
|
||
|
|
|
||
|
|
BB0_157:
|
||
|
|
ld.global.u32 %r373, [additive];
|
||
|
|
setp.eq.s32 %p147, %r373, 0;
|
||
|
|
@%p147 bra BB0_159;
|
||
|
|
|
||
|
|
mov.u64 %rd293, image_RNM3;
|
||
|
|
cvta.global.u64 %rd282, %rd293;
|
||
|
|
mov.u32 %r377, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r101, %r377, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs146, %rs147, %rs148, %rs149}, [%rd281];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f902, %rs146;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f903, %rs147;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f904, %rs148;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd287), _rt_buffer_get_64, (%rd282, %r101, %r377, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f905, %f902, 0f00000000;
|
||
|
|
add.f32 %f906, %f903, 0f00000000;
|
||
|
|
add.f32 %f907, %f904, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs145, %f907;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs144, %f906;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs143, %f905;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs150, 0;
|
||
|
|
st.v4.u16 [%rd287], {%rs143, %rs144, %rs145, %rs150};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_159:
|
||
|
|
mov.u64 %rd300, image_RNM3;
|
||
|
|
cvta.global.u64 %rd295, %rd300;
|
||
|
|
mov.u32 %r379, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd294), _rt_buffer_get_64, (%rd295, %r101, %r379, %rd18, %rd19, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f908, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs151, %f908;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs152, 0;
|
||
|
|
st.v4.u16 [%rd294], {%rs151, %rs151, %rs151, %rs152};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_80:
|
||
|
|
setp.geu.f32 %p78, %f148, 0f00000000;
|
||
|
|
@%p78 bra BB0_83;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f566, %f538;
|
||
|
|
setp.neu.f32 %p79, %f566, 0f3EE8BA2E;
|
||
|
|
selp.f32 %f978, 0f7FFFFFFF, %f978, %p79;
|
||
|
|
|
||
|
|
BB0_83:
|
||
|
|
add.f32 %f568, %f150, 0f3EE8BA2E;
|
||
|
|
mov.b32 %r266, %f568;
|
||
|
|
setp.lt.s32 %p81, %r266, 2139095040;
|
||
|
|
@%p81 bra BB0_88;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p82, %f150, 0f7F800000;
|
||
|
|
@%p82 bra BB0_87;
|
||
|
|
bra.uni BB0_85;
|
||
|
|
|
||
|
|
BB0_87:
|
||
|
|
add.f32 %f978, %f148, 0f3EE8BA2E;
|
||
|
|
bra.uni BB0_88;
|
||
|
|
|
||
|
|
BB0_85:
|
||
|
|
setp.neu.f32 %p83, %f150, 0f7F800000;
|
||
|
|
@%p83 bra BB0_88;
|
||
|
|
|
||
|
|
selp.f32 %f978, 0fFF800000, 0f7F800000, %p3;
|
||
|
|
|
||
|
|
BB0_88:
|
||
|
|
mul.f32 %f569, %f978, 0f437F0000;
|
||
|
|
setp.eq.f32 %p84, %f148, 0f3F800000;
|
||
|
|
selp.f32 %f570, 0f437F0000, %f569, %p84;
|
||
|
|
cvt.rzi.u32.f32 %r267, %f570;
|
||
|
|
cvt.u16.u32 %rs15, %r267;
|
||
|
|
mov.u16 %rs16, 255;
|
||
|
|
st.v2.u8 [%rd73], {%rs15, %rs16};
|
||
|
|
ld.global.u32 %r410, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_89:
|
||
|
|
ld.global.f32 %f571, [lightColor];
|
||
|
|
mul.f32 %f162, %f970, %f571;
|
||
|
|
ld.global.f32 %f572, [lightColor+4];
|
||
|
|
mul.f32 %f163, %f970, %f572;
|
||
|
|
ld.global.f32 %f573, [lightColor+8];
|
||
|
|
mul.f32 %f164, %f970, %f573;
|
||
|
|
and.b32 %r268, %r410, 1;
|
||
|
|
setp.eq.b32 %p85, %r268, 1;
|
||
|
|
@!%p85 bra BB0_124;
|
||
|
|
bra.uni BB0_90;
|
||
|
|
|
||
|
|
BB0_90:
|
||
|
|
mov.f32 %f576, 0f3E666666;
|
||
|
|
cvt.rzi.f32.f32 %f577, %f576;
|
||
|
|
fma.rn.f32 %f578, %f577, 0fC0000000, 0f3EE66666;
|
||
|
|
abs.f32 %f165, %f578;
|
||
|
|
abs.f32 %f166, %f162;
|
||
|
|
setp.lt.f32 %p86, %f166, 0f00800000;
|
||
|
|
mul.f32 %f579, %f166, 0f4B800000;
|
||
|
|
selp.f32 %f580, 0fC3170000, 0fC2FE0000, %p86;
|
||
|
|
selp.f32 %f581, %f579, %f166, %p86;
|
||
|
|
mov.b32 %r269, %f581;
|
||
|
|
and.b32 %r270, %r269, 8388607;
|
||
|
|
or.b32 %r271, %r270, 1065353216;
|
||
|
|
mov.b32 %f582, %r271;
|
||
|
|
shr.u32 %r272, %r269, 23;
|
||
|
|
cvt.rn.f32.u32 %f583, %r272;
|
||
|
|
add.f32 %f584, %f580, %f583;
|
||
|
|
setp.gt.f32 %p87, %f582, 0f3FB504F3;
|
||
|
|
mul.f32 %f585, %f582, 0f3F000000;
|
||
|
|
add.f32 %f586, %f584, 0f3F800000;
|
||
|
|
selp.f32 %f587, %f585, %f582, %p87;
|
||
|
|
selp.f32 %f588, %f586, %f584, %p87;
|
||
|
|
add.f32 %f589, %f587, 0fBF800000;
|
||
|
|
add.f32 %f575, %f587, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f574,%f575;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f590, %f589, %f589;
|
||
|
|
mul.f32 %f591, %f574, %f590;
|
||
|
|
mul.f32 %f592, %f591, %f591;
|
||
|
|
mov.f32 %f593, 0f3C4CAF63;
|
||
|
|
mov.f32 %f594, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f595, %f594, %f592, %f593;
|
||
|
|
mov.f32 %f596, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f597, %f595, %f592, %f596;
|
||
|
|
mul.rn.f32 %f598, %f597, %f592;
|
||
|
|
mul.rn.f32 %f599, %f598, %f591;
|
||
|
|
sub.f32 %f600, %f589, %f591;
|
||
|
|
neg.f32 %f601, %f591;
|
||
|
|
add.f32 %f602, %f600, %f600;
|
||
|
|
fma.rn.f32 %f603, %f601, %f589, %f602;
|
||
|
|
mul.rn.f32 %f604, %f574, %f603;
|
||
|
|
add.f32 %f605, %f599, %f591;
|
||
|
|
sub.f32 %f606, %f591, %f605;
|
||
|
|
add.f32 %f607, %f599, %f606;
|
||
|
|
add.f32 %f608, %f604, %f607;
|
||
|
|
add.f32 %f609, %f605, %f608;
|
||
|
|
sub.f32 %f610, %f605, %f609;
|
||
|
|
add.f32 %f611, %f608, %f610;
|
||
|
|
mov.f32 %f612, 0f3F317200;
|
||
|
|
mul.rn.f32 %f613, %f588, %f612;
|
||
|
|
mov.f32 %f614, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f615, %f588, %f614;
|
||
|
|
add.f32 %f616, %f613, %f609;
|
||
|
|
sub.f32 %f617, %f613, %f616;
|
||
|
|
add.f32 %f618, %f609, %f617;
|
||
|
|
add.f32 %f619, %f611, %f618;
|
||
|
|
add.f32 %f620, %f615, %f619;
|
||
|
|
add.f32 %f621, %f616, %f620;
|
||
|
|
sub.f32 %f622, %f616, %f621;
|
||
|
|
add.f32 %f623, %f620, %f622;
|
||
|
|
mov.f32 %f624, 0f3EE66666;
|
||
|
|
mul.rn.f32 %f625, %f624, %f621;
|
||
|
|
neg.f32 %f626, %f625;
|
||
|
|
fma.rn.f32 %f627, %f624, %f621, %f626;
|
||
|
|
fma.rn.f32 %f628, %f624, %f623, %f627;
|
||
|
|
mov.f32 %f629, 0f00000000;
|
||
|
|
fma.rn.f32 %f630, %f629, %f621, %f628;
|
||
|
|
add.rn.f32 %f631, %f625, %f630;
|
||
|
|
neg.f32 %f632, %f631;
|
||
|
|
add.rn.f32 %f633, %f625, %f632;
|
||
|
|
add.rn.f32 %f634, %f633, %f630;
|
||
|
|
mov.b32 %r273, %f631;
|
||
|
|
setp.eq.s32 %p88, %r273, 1118925336;
|
||
|
|
add.s32 %r274, %r273, -1;
|
||
|
|
mov.b32 %f635, %r274;
|
||
|
|
add.f32 %f636, %f634, 0f37000000;
|
||
|
|
selp.f32 %f637, %f635, %f631, %p88;
|
||
|
|
selp.f32 %f167, %f636, %f634, %p88;
|
||
|
|
mul.f32 %f638, %f637, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f639, %f638;
|
||
|
|
mov.f32 %f640, 0fBF317200;
|
||
|
|
fma.rn.f32 %f641, %f639, %f640, %f637;
|
||
|
|
mov.f32 %f642, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f643, %f639, %f642, %f641;
|
||
|
|
mul.f32 %f644, %f643, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f645, %f644;
|
||
|
|
add.f32 %f646, %f639, 0f00000000;
|
||
|
|
ex2.approx.f32 %f647, %f646;
|
||
|
|
mul.f32 %f648, %f645, %f647;
|
||
|
|
setp.lt.f32 %p89, %f637, 0fC2D20000;
|
||
|
|
selp.f32 %f649, 0f00000000, %f648, %p89;
|
||
|
|
setp.gt.f32 %p90, %f637, 0f42D20000;
|
||
|
|
selp.f32 %f979, 0f7F800000, %f649, %p90;
|
||
|
|
setp.eq.f32 %p91, %f979, 0f7F800000;
|
||
|
|
@%p91 bra BB0_92;
|
||
|
|
|
||
|
|
fma.rn.f32 %f979, %f979, %f167, %f979;
|
||
|
|
|
||
|
|
BB0_92:
|
||
|
|
setp.lt.f32 %p92, %f162, 0f00000000;
|
||
|
|
setp.eq.f32 %p93, %f165, 0f3F800000;
|
||
|
|
and.pred %p4, %p92, %p93;
|
||
|
|
mov.b32 %r275, %f979;
|
||
|
|
xor.b32 %r276, %r275, -2147483648;
|
||
|
|
mov.b32 %f650, %r276;
|
||
|
|
selp.f32 %f981, %f650, %f979, %p4;
|
||
|
|
setp.eq.f32 %p94, %f162, 0f00000000;
|
||
|
|
@%p94 bra BB0_95;
|
||
|
|
bra.uni BB0_93;
|
||
|
|
|
||
|
|
BB0_95:
|
||
|
|
add.f32 %f653, %f162, %f162;
|
||
|
|
selp.f32 %f981, %f653, 0f00000000, %p93;
|
||
|
|
bra.uni BB0_96;
|
||
|
|
|
||
|
|
BB0_93:
|
||
|
|
setp.geu.f32 %p95, %f162, 0f00000000;
|
||
|
|
@%p95 bra BB0_96;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f652, %f624;
|
||
|
|
setp.neu.f32 %p96, %f652, 0f3EE66666;
|
||
|
|
selp.f32 %f981, 0f7FFFFFFF, %f981, %p96;
|
||
|
|
|
||
|
|
BB0_96:
|
||
|
|
add.f32 %f654, %f166, 0f3EE66666;
|
||
|
|
mov.b32 %r277, %f654;
|
||
|
|
setp.lt.s32 %p98, %r277, 2139095040;
|
||
|
|
@%p98 bra BB0_101;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p99, %f166, 0f7F800000;
|
||
|
|
@%p99 bra BB0_100;
|
||
|
|
bra.uni BB0_98;
|
||
|
|
|
||
|
|
BB0_100:
|
||
|
|
add.f32 %f981, %f162, 0f3EE66666;
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_98:
|
||
|
|
setp.neu.f32 %p100, %f166, 0f7F800000;
|
||
|
|
@%p100 bra BB0_101;
|
||
|
|
|
||
|
|
selp.f32 %f981, 0fFF800000, 0f7F800000, %p4;
|
||
|
|
|
||
|
|
BB0_101:
|
||
|
|
setp.eq.f32 %p101, %f162, 0f3F800000;
|
||
|
|
selp.f32 %f178, 0f3F800000, %f981, %p101;
|
||
|
|
abs.f32 %f179, %f163;
|
||
|
|
setp.lt.f32 %p102, %f179, 0f00800000;
|
||
|
|
mul.f32 %f657, %f179, 0f4B800000;
|
||
|
|
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p102;
|
||
|
|
selp.f32 %f659, %f657, %f179, %p102;
|
||
|
|
mov.b32 %r278, %f659;
|
||
|
|
and.b32 %r279, %r278, 8388607;
|
||
|
|
or.b32 %r280, %r279, 1065353216;
|
||
|
|
mov.b32 %f660, %r280;
|
||
|
|
shr.u32 %r281, %r278, 23;
|
||
|
|
cvt.rn.f32.u32 %f661, %r281;
|
||
|
|
add.f32 %f662, %f658, %f661;
|
||
|
|
setp.gt.f32 %p103, %f660, 0f3FB504F3;
|
||
|
|
mul.f32 %f663, %f660, 0f3F000000;
|
||
|
|
add.f32 %f664, %f662, 0f3F800000;
|
||
|
|
selp.f32 %f665, %f663, %f660, %p103;
|
||
|
|
selp.f32 %f666, %f664, %f662, %p103;
|
||
|
|
add.f32 %f667, %f665, 0fBF800000;
|
||
|
|
add.f32 %f656, %f665, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f655,%f656;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f668, %f667, %f667;
|
||
|
|
mul.f32 %f669, %f655, %f668;
|
||
|
|
mul.f32 %f670, %f669, %f669;
|
||
|
|
fma.rn.f32 %f673, %f594, %f670, %f593;
|
||
|
|
fma.rn.f32 %f675, %f673, %f670, %f596;
|
||
|
|
mul.rn.f32 %f676, %f675, %f670;
|
||
|
|
mul.rn.f32 %f677, %f676, %f669;
|
||
|
|
sub.f32 %f678, %f667, %f669;
|
||
|
|
neg.f32 %f679, %f669;
|
||
|
|
add.f32 %f680, %f678, %f678;
|
||
|
|
fma.rn.f32 %f681, %f679, %f667, %f680;
|
||
|
|
mul.rn.f32 %f682, %f655, %f681;
|
||
|
|
add.f32 %f683, %f677, %f669;
|
||
|
|
sub.f32 %f684, %f669, %f683;
|
||
|
|
add.f32 %f685, %f677, %f684;
|
||
|
|
add.f32 %f686, %f682, %f685;
|
||
|
|
add.f32 %f687, %f683, %f686;
|
||
|
|
sub.f32 %f688, %f683, %f687;
|
||
|
|
add.f32 %f689, %f686, %f688;
|
||
|
|
mul.rn.f32 %f691, %f666, %f612;
|
||
|
|
mul.rn.f32 %f693, %f666, %f614;
|
||
|
|
add.f32 %f694, %f691, %f687;
|
||
|
|
sub.f32 %f695, %f691, %f694;
|
||
|
|
add.f32 %f696, %f687, %f695;
|
||
|
|
add.f32 %f697, %f689, %f696;
|
||
|
|
add.f32 %f698, %f693, %f697;
|
||
|
|
add.f32 %f699, %f694, %f698;
|
||
|
|
sub.f32 %f700, %f694, %f699;
|
||
|
|
add.f32 %f701, %f698, %f700;
|
||
|
|
mul.rn.f32 %f703, %f624, %f699;
|
||
|
|
neg.f32 %f704, %f703;
|
||
|
|
fma.rn.f32 %f705, %f624, %f699, %f704;
|
||
|
|
fma.rn.f32 %f706, %f624, %f701, %f705;
|
||
|
|
fma.rn.f32 %f708, %f629, %f699, %f706;
|
||
|
|
add.rn.f32 %f709, %f703, %f708;
|
||
|
|
neg.f32 %f710, %f709;
|
||
|
|
add.rn.f32 %f711, %f703, %f710;
|
||
|
|
add.rn.f32 %f712, %f711, %f708;
|
||
|
|
mov.b32 %r282, %f709;
|
||
|
|
setp.eq.s32 %p104, %r282, 1118925336;
|
||
|
|
add.s32 %r283, %r282, -1;
|
||
|
|
mov.b32 %f713, %r283;
|
||
|
|
add.f32 %f714, %f712, 0f37000000;
|
||
|
|
selp.f32 %f715, %f713, %f709, %p104;
|
||
|
|
selp.f32 %f180, %f714, %f712, %p104;
|
||
|
|
mul.f32 %f716, %f715, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f717, %f716;
|
||
|
|
fma.rn.f32 %f719, %f717, %f640, %f715;
|
||
|
|
fma.rn.f32 %f721, %f717, %f642, %f719;
|
||
|
|
mul.f32 %f722, %f721, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f723, %f722;
|
||
|
|
add.f32 %f724, %f717, 0f00000000;
|
||
|
|
ex2.approx.f32 %f725, %f724;
|
||
|
|
mul.f32 %f726, %f723, %f725;
|
||
|
|
setp.lt.f32 %p105, %f715, 0fC2D20000;
|
||
|
|
selp.f32 %f727, 0f00000000, %f726, %p105;
|
||
|
|
setp.gt.f32 %p106, %f715, 0f42D20000;
|
||
|
|
selp.f32 %f982, 0f7F800000, %f727, %p106;
|
||
|
|
setp.eq.f32 %p107, %f982, 0f7F800000;
|
||
|
|
@%p107 bra BB0_103;
|
||
|
|
|
||
|
|
fma.rn.f32 %f982, %f982, %f180, %f982;
|
||
|
|
|
||
|
|
BB0_103:
|
||
|
|
setp.lt.f32 %p108, %f163, 0f00000000;
|
||
|
|
and.pred %p5, %p108, %p93;
|
||
|
|
mov.b32 %r284, %f982;
|
||
|
|
xor.b32 %r285, %r284, -2147483648;
|
||
|
|
mov.b32 %f728, %r285;
|
||
|
|
selp.f32 %f984, %f728, %f982, %p5;
|
||
|
|
setp.eq.f32 %p110, %f163, 0f00000000;
|
||
|
|
@%p110 bra BB0_106;
|
||
|
|
bra.uni BB0_104;
|
||
|
|
|
||
|
|
BB0_106:
|
||
|
|
add.f32 %f731, %f163, %f163;
|
||
|
|
selp.f32 %f984, %f731, 0f00000000, %p93;
|
||
|
|
bra.uni BB0_107;
|
||
|
|
|
||
|
|
BB0_104:
|
||
|
|
setp.geu.f32 %p111, %f163, 0f00000000;
|
||
|
|
@%p111 bra BB0_107;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f730, %f624;
|
||
|
|
setp.neu.f32 %p112, %f730, 0f3EE66666;
|
||
|
|
selp.f32 %f984, 0f7FFFFFFF, %f984, %p112;
|
||
|
|
|
||
|
|
BB0_107:
|
||
|
|
add.f32 %f732, %f179, 0f3EE66666;
|
||
|
|
mov.b32 %r286, %f732;
|
||
|
|
setp.lt.s32 %p114, %r286, 2139095040;
|
||
|
|
@%p114 bra BB0_112;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p115, %f179, 0f7F800000;
|
||
|
|
@%p115 bra BB0_111;
|
||
|
|
bra.uni BB0_109;
|
||
|
|
|
||
|
|
BB0_111:
|
||
|
|
add.f32 %f984, %f163, 0f3EE66666;
|
||
|
|
bra.uni BB0_112;
|
||
|
|
|
||
|
|
BB0_109:
|
||
|
|
setp.neu.f32 %p116, %f179, 0f7F800000;
|
||
|
|
@%p116 bra BB0_112;
|
||
|
|
|
||
|
|
selp.f32 %f984, 0fFF800000, 0f7F800000, %p5;
|
||
|
|
|
||
|
|
BB0_112:
|
||
|
|
setp.eq.f32 %p117, %f163, 0f3F800000;
|
||
|
|
selp.f32 %f191, 0f3F800000, %f984, %p117;
|
||
|
|
abs.f32 %f192, %f164;
|
||
|
|
setp.lt.f32 %p118, %f192, 0f00800000;
|
||
|
|
mul.f32 %f735, %f192, 0f4B800000;
|
||
|
|
selp.f32 %f736, 0fC3170000, 0fC2FE0000, %p118;
|
||
|
|
selp.f32 %f737, %f735, %f192, %p118;
|
||
|
|
mov.b32 %r287, %f737;
|
||
|
|
and.b32 %r288, %r287, 8388607;
|
||
|
|
or.b32 %r289, %r288, 1065353216;
|
||
|
|
mov.b32 %f738, %r289;
|
||
|
|
shr.u32 %r290, %r287, 23;
|
||
|
|
cvt.rn.f32.u32 %f739, %r290;
|
||
|
|
add.f32 %f740, %f736, %f739;
|
||
|
|
setp.gt.f32 %p119, %f738, 0f3FB504F3;
|
||
|
|
mul.f32 %f741, %f738, 0f3F000000;
|
||
|
|
add.f32 %f742, %f740, 0f3F800000;
|
||
|
|
selp.f32 %f743, %f741, %f738, %p119;
|
||
|
|
selp.f32 %f744, %f742, %f740, %p119;
|
||
|
|
add.f32 %f745, %f743, 0fBF800000;
|
||
|
|
add.f32 %f734, %f743, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f733,%f734;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f746, %f745, %f745;
|
||
|
|
mul.f32 %f747, %f733, %f746;
|
||
|
|
mul.f32 %f748, %f747, %f747;
|
||
|
|
fma.rn.f32 %f751, %f594, %f748, %f593;
|
||
|
|
fma.rn.f32 %f753, %f751, %f748, %f596;
|
||
|
|
mul.rn.f32 %f754, %f753, %f748;
|
||
|
|
mul.rn.f32 %f755, %f754, %f747;
|
||
|
|
sub.f32 %f756, %f745, %f747;
|
||
|
|
neg.f32 %f757, %f747;
|
||
|
|
add.f32 %f758, %f756, %f756;
|
||
|
|
fma.rn.f32 %f759, %f757, %f745, %f758;
|
||
|
|
mul.rn.f32 %f760, %f733, %f759;
|
||
|
|
add.f32 %f761, %f755, %f747;
|
||
|
|
sub.f32 %f762, %f747, %f761;
|
||
|
|
add.f32 %f763, %f755, %f762;
|
||
|
|
add.f32 %f764, %f760, %f763;
|
||
|
|
add.f32 %f765, %f761, %f764;
|
||
|
|
sub.f32 %f766, %f761, %f765;
|
||
|
|
add.f32 %f767, %f764, %f766;
|
||
|
|
mul.rn.f32 %f769, %f744, %f612;
|
||
|
|
mul.rn.f32 %f771, %f744, %f614;
|
||
|
|
add.f32 %f772, %f769, %f765;
|
||
|
|
sub.f32 %f773, %f769, %f772;
|
||
|
|
add.f32 %f774, %f765, %f773;
|
||
|
|
add.f32 %f775, %f767, %f774;
|
||
|
|
add.f32 %f776, %f771, %f775;
|
||
|
|
add.f32 %f777, %f772, %f776;
|
||
|
|
sub.f32 %f778, %f772, %f777;
|
||
|
|
add.f32 %f779, %f776, %f778;
|
||
|
|
mul.rn.f32 %f781, %f624, %f777;
|
||
|
|
neg.f32 %f782, %f781;
|
||
|
|
fma.rn.f32 %f783, %f624, %f777, %f782;
|
||
|
|
fma.rn.f32 %f784, %f624, %f779, %f783;
|
||
|
|
fma.rn.f32 %f786, %f629, %f777, %f784;
|
||
|
|
add.rn.f32 %f787, %f781, %f786;
|
||
|
|
neg.f32 %f788, %f787;
|
||
|
|
add.rn.f32 %f789, %f781, %f788;
|
||
|
|
add.rn.f32 %f790, %f789, %f786;
|
||
|
|
mov.b32 %r291, %f787;
|
||
|
|
setp.eq.s32 %p120, %r291, 1118925336;
|
||
|
|
add.s32 %r292, %r291, -1;
|
||
|
|
mov.b32 %f791, %r292;
|
||
|
|
add.f32 %f792, %f790, 0f37000000;
|
||
|
|
selp.f32 %f793, %f791, %f787, %p120;
|
||
|
|
selp.f32 %f193, %f792, %f790, %p120;
|
||
|
|
mul.f32 %f794, %f793, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f795, %f794;
|
||
|
|
fma.rn.f32 %f797, %f795, %f640, %f793;
|
||
|
|
fma.rn.f32 %f799, %f795, %f642, %f797;
|
||
|
|
mul.f32 %f800, %f799, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f801, %f800;
|
||
|
|
add.f32 %f802, %f795, 0f00000000;
|
||
|
|
ex2.approx.f32 %f803, %f802;
|
||
|
|
mul.f32 %f804, %f801, %f803;
|
||
|
|
setp.lt.f32 %p121, %f793, 0fC2D20000;
|
||
|
|
selp.f32 %f805, 0f00000000, %f804, %p121;
|
||
|
|
setp.gt.f32 %p122, %f793, 0f42D20000;
|
||
|
|
selp.f32 %f985, 0f7F800000, %f805, %p122;
|
||
|
|
setp.eq.f32 %p123, %f985, 0f7F800000;
|
||
|
|
@%p123 bra BB0_114;
|
||
|
|
|
||
|
|
fma.rn.f32 %f985, %f985, %f193, %f985;
|
||
|
|
|
||
|
|
BB0_114:
|
||
|
|
setp.lt.f32 %p124, %f164, 0f00000000;
|
||
|
|
and.pred %p6, %p124, %p93;
|
||
|
|
mov.b32 %r293, %f985;
|
||
|
|
xor.b32 %r294, %r293, -2147483648;
|
||
|
|
mov.b32 %f806, %r294;
|
||
|
|
selp.f32 %f987, %f806, %f985, %p6;
|
||
|
|
setp.eq.f32 %p126, %f164, 0f00000000;
|
||
|
|
@%p126 bra BB0_117;
|
||
|
|
bra.uni BB0_115;
|
||
|
|
|
||
|
|
BB0_117:
|
||
|
|
add.f32 %f809, %f164, %f164;
|
||
|
|
selp.f32 %f987, %f809, 0f00000000, %p93;
|
||
|
|
bra.uni BB0_118;
|
||
|
|
|
||
|
|
BB0_115:
|
||
|
|
setp.geu.f32 %p127, %f164, 0f00000000;
|
||
|
|
@%p127 bra BB0_118;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f808, %f624;
|
||
|
|
setp.neu.f32 %p128, %f808, 0f3EE66666;
|
||
|
|
selp.f32 %f987, 0f7FFFFFFF, %f987, %p128;
|
||
|
|
|
||
|
|
BB0_118:
|
||
|
|
add.f32 %f810, %f192, 0f3EE66666;
|
||
|
|
mov.b32 %r295, %f810;
|
||
|
|
setp.lt.s32 %p130, %r295, 2139095040;
|
||
|
|
@%p130 bra BB0_123;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p131, %f192, 0f7F800000;
|
||
|
|
@%p131 bra BB0_122;
|
||
|
|
bra.uni BB0_120;
|
||
|
|
|
||
|
|
BB0_122:
|
||
|
|
add.f32 %f987, %f164, 0f3EE66666;
|
||
|
|
bra.uni BB0_123;
|
||
|
|
|
||
|
|
BB0_120:
|
||
|
|
setp.neu.f32 %p132, %f192, 0f7F800000;
|
||
|
|
@%p132 bra BB0_123;
|
||
|
|
|
||
|
|
selp.f32 %f987, 0fFF800000, 0f7F800000, %p6;
|
||
|
|
|
||
|
|
BB0_123:
|
||
|
|
setp.eq.f32 %p133, %f164, 0f3F800000;
|
||
|
|
selp.f32 %f811, 0f3F800000, %f987, %p133;
|
||
|
|
cvt.u64.u32 %rd83, %r7;
|
||
|
|
cvt.u64.u32 %rd82, %r6;
|
||
|
|
mov.u64 %rd86, image;
|
||
|
|
cvta.global.u64 %rd81, %rd86;
|
||
|
|
// inline asm
|
||
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r101, %r102, %rd82, %rd83, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
cvt.sat.f32.f32 %f812, %f811;
|
||
|
|
mul.f32 %f813, %f812, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r298, %f813;
|
||
|
|
cvt.sat.f32.f32 %f814, %f191;
|
||
|
|
mul.f32 %f815, %f814, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r299, %f815;
|
||
|
|
cvt.sat.f32.f32 %f816, %f178;
|
||
|
|
mul.f32 %f817, %f816, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r300, %f817;
|
||
|
|
cvt.u16.u32 %rs17, %r298;
|
||
|
|
cvt.u16.u32 %rs18, %r300;
|
||
|
|
cvt.u16.u32 %rs19, %r299;
|
||
|
|
mov.u16 %rs20, 255;
|
||
|
|
st.v4.u8 [%rd80], {%rs17, %rs19, %rs18, %rs20};
|
||
|
|
ld.global.u32 %r410, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_124:
|
||
|
|
cvt.u64.u32 %rd16, %r6;
|
||
|
|
cvt.u64.u32 %rd17, %r7;
|
||
|
|
and.b32 %r301, %r410, 4;
|
||
|
|
setp.eq.s32 %p134, %r301, 0;
|
||
|
|
@%p134 bra BB0_128;
|
||
|
|
|
||
|
|
ld.global.u32 %r302, [additive];
|
||
|
|
setp.eq.s32 %p135, %r302, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs21, %f392;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p135 bra BB0_127;
|
||
|
|
|
||
|
|
mov.u64 %rd99, image_HDR;
|
||
|
|
cvta.global.u64 %rd88, %rd99;
|
||
|
|
mov.u32 %r306, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r101, %r306, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs28, %rs29, %rs30, %rs31}, [%rd87];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f819, %rs28;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f820, %rs29;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f821, %rs30;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r101, %r306, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f822, %f162, %f819;
|
||
|
|
add.f32 %f823, %f163, %f820;
|
||
|
|
add.f32 %f824, %f164, %f821;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs27, %f824;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs26, %f823;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs25, %f822;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd93], {%rs25, %rs26, %rs27, %rs21};
|
||
|
|
bra.uni BB0_128;
|
||
|
|
|
||
|
|
BB0_127:
|
||
|
|
mov.u64 %rd106, image_HDR;
|
||
|
|
cvta.global.u64 %rd101, %rd106;
|
||
|
|
mov.u32 %r308, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r101, %r308, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs34, %f164;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs33, %f163;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs32, %f162;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd100], {%rs32, %rs33, %rs34, %rs21};
|
||
|
|
|
||
|
|
BB0_128:
|
||
|
|
selp.f32 %f829, 0f3F800000, 0f3E800000, %p148;
|
||
|
|
mul.f32 %f204, %f829, %f971;
|
||
|
|
mul.f32 %f205, %f829, %f972;
|
||
|
|
mul.f32 %f206, %f829, %f973;
|
||
|
|
mul.f32 %f207, %f829, %f974;
|
||
|
|
ld.global.f32 %f830, [lightColor];
|
||
|
|
mul.f32 %f208, %f204, %f830;
|
||
|
|
ld.global.f32 %f831, [lightColor+4];
|
||
|
|
mul.f32 %f209, %f204, %f831;
|
||
|
|
ld.global.f32 %f832, [lightColor+8];
|
||
|
|
mul.f32 %f210, %f204, %f832;
|
||
|
|
ld.global.u32 %r309, [additive];
|
||
|
|
setp.eq.s32 %p136, %r309, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs35, %f392;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p136 bra BB0_130;
|
||
|
|
|
||
|
|
mov.u64 %rd119, image_RNM0;
|
||
|
|
cvta.global.u64 %rd108, %rd119;
|
||
|
|
mov.u32 %r313, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r101, %r313, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd107];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f833, %rs42;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f834, %rs43;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f835, %rs44;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd113), _rt_buffer_get_64, (%rd108, %r101, %r313, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f836, %f208, %f833;
|
||
|
|
add.f32 %f837, %f209, %f834;
|
||
|
|
add.f32 %f838, %f210, %f835;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs41, %f838;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs40, %f837;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs39, %f836;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd113], {%rs39, %rs40, %rs41, %rs35};
|
||
|
|
bra.uni BB0_131;
|
||
|
|
|
||
|
|
BB0_130:
|
||
|
|
mov.u64 %rd126, image_RNM0;
|
||
|
|
cvta.global.u64 %rd121, %rd126;
|
||
|
|
mov.u32 %r315, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r101, %r315, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs48, %f210;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs47, %f209;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs46, %f208;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd120], {%rs46, %rs47, %rs48, %rs35};
|
||
|
|
|
||
|
|
BB0_131:
|
||
|
|
mov.f32 %f843, 0f34000000;
|
||
|
|
max.f32 %f844, %f204, %f843;
|
||
|
|
div.rn.f32 %f845, %f205, %f844;
|
||
|
|
fma.rn.f32 %f211, %f845, 0f3F000000, 0f3F000000;
|
||
|
|
div.rn.f32 %f846, %f206, %f844;
|
||
|
|
fma.rn.f32 %f212, %f846, 0f3F000000, 0f3F000000;
|
||
|
|
div.rn.f32 %f847, %f207, %f844;
|
||
|
|
fma.rn.f32 %f213, %f847, 0f3F000000, 0f3F000000;
|
||
|
|
ld.global.u32 %r316, [additive];
|
||
|
|
setp.eq.s32 %p137, %r316, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs49, %f392;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p137 bra BB0_133;
|
||
|
|
|
||
|
|
mov.u64 %rd139, image_RNM1;
|
||
|
|
cvta.global.u64 %rd128, %rd139;
|
||
|
|
mov.u32 %r320, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd127), _rt_buffer_get_64, (%rd128, %r101, %r320, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd127];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f848, %rs56;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f849, %rs57;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f850, %rs58;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd133), _rt_buffer_get_64, (%rd128, %r101, %r320, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f851, %f211, %f848;
|
||
|
|
add.f32 %f852, %f211, %f849;
|
||
|
|
add.f32 %f853, %f211, %f850;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs55, %f853;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs54, %f852;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs53, %f851;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd133], {%rs53, %rs54, %rs55, %rs49};
|
||
|
|
bra.uni BB0_134;
|
||
|
|
|
||
|
|
BB0_133:
|
||
|
|
mov.u64 %rd146, image_RNM1;
|
||
|
|
cvta.global.u64 %rd141, %rd146;
|
||
|
|
mov.u32 %r322, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r101, %r322, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs60, %f211;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd140], {%rs60, %rs60, %rs60, %rs49};
|
||
|
|
|
||
|
|
BB0_134:
|
||
|
|
ld.global.u32 %r323, [additive];
|
||
|
|
setp.eq.s32 %p138, %r323, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs61, %f392;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p138 bra BB0_136;
|
||
|
|
|
||
|
|
mov.u64 %rd159, image_RNM2;
|
||
|
|
cvta.global.u64 %rd148, %rd159;
|
||
|
|
mov.u32 %r327, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r101, %r327, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs68, %rs69, %rs70, %rs71}, [%rd147];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f856, %rs68;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f857, %rs69;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f858, %rs70;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd153), _rt_buffer_get_64, (%rd148, %r101, %r327, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f859, %f212, %f856;
|
||
|
|
add.f32 %f860, %f212, %f857;
|
||
|
|
add.f32 %f861, %f212, %f858;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs67, %f861;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs66, %f860;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs65, %f859;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd153], {%rs65, %rs66, %rs67, %rs61};
|
||
|
|
bra.uni BB0_137;
|
||
|
|
|
||
|
|
BB0_136:
|
||
|
|
mov.u64 %rd166, image_RNM2;
|
||
|
|
cvta.global.u64 %rd161, %rd166;
|
||
|
|
mov.u32 %r329, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd160), _rt_buffer_get_64, (%rd161, %r101, %r329, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs72, %f212;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd160], {%rs72, %rs72, %rs72, %rs61};
|
||
|
|
|
||
|
|
BB0_137:
|
||
|
|
ld.global.u32 %r330, [additive];
|
||
|
|
setp.eq.s32 %p139, %r330, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs73, %f392;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p139 bra BB0_139;
|
||
|
|
|
||
|
|
mov.u64 %rd179, image_RNM3;
|
||
|
|
cvta.global.u64 %rd168, %rd179;
|
||
|
|
mov.u32 %r334, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r101, %r334, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs80, %rs81, %rs82, %rs83}, [%rd167];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f864, %rs80;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f865, %rs81;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f866, %rs82;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd173), _rt_buffer_get_64, (%rd168, %r101, %r334, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f867, %f213, %f864;
|
||
|
|
add.f32 %f868, %f213, %f865;
|
||
|
|
add.f32 %f869, %f213, %f866;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs79, %f869;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs78, %f868;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs77, %f867;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd173], {%rs77, %rs78, %rs79, %rs73};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_139:
|
||
|
|
mov.u64 %rd186, image_RNM3;
|
||
|
|
cvta.global.u64 %rd181, %rd186;
|
||
|
|
mov.u32 %r336, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd180), _rt_buffer_get_64, (%rd181, %r101, %r336, %rd16, %rd17, %rd25, %rd25);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs84, %f213;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd180], {%rs84, %rs84, %rs84, %rs73};
|
||
|
|
|
||
|
|
BB0_160:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|