2824 lines
82 KiB
Plaintext
2824 lines
82 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 rnd_seeds[1];
|
||
|
|
.global .align 4 .b8 directDir[12];
|
||
|
|
.global .align 4 .b8 directColor[12];
|
||
|
|
.global .align 4 .f32 shadowSpread;
|
||
|
|
.global .align 4 .u32 samples;
|
||
|
|
.global .align 4 .u32 ignoreNormal;
|
||
|
|
.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_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[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_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
||
|
|
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
||
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[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_typeenum9directDirE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
||
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 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_semantic9directDirE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[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_annotation9directDirE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
||
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[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<155>;
|
||
|
|
.reg .b16 %rs<177>;
|
||
|
|
.reg .f32 %f<1404>;
|
||
|
|
.reg .b32 %r<502>;
|
||
|
|
.reg .b64 %rd<326>;
|
||
|
|
|
||
|
|
|
||
|
|
mov.u64 %rd325, __local_depot0;
|
||
|
|
cvta.local.u64 %SP, %rd325;
|
||
|
|
ld.global.v2.u32 {%r109, %r110}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd24, %r109;
|
||
|
|
cvt.u64.u32 %rd25, %r110;
|
||
|
|
mov.u64 %rd28, uvnormal;
|
||
|
|
cvta.global.u64 %rd23, %rd28;
|
||
|
|
mov.u32 %r107, 2;
|
||
|
|
mov.u32 %r108, 4;
|
||
|
|
mov.u64 %rd27, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd22), _rt_buffer_get_64, (%rd23, %r107, %r108, %rd24, %rd25, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.u32 %r1, [%rd22];
|
||
|
|
shr.u32 %r113, %r1, 16;
|
||
|
|
cvt.u16.u32 %rs1, %r113;
|
||
|
|
and.b16 %rs15, %rs1, 255;
|
||
|
|
cvt.u16.u32 %rs16, %r1;
|
||
|
|
or.b16 %rs17, %rs16, %rs15;
|
||
|
|
setp.eq.s16 %p5, %rs17, 0;
|
||
|
|
mov.f32 %f1350, 0f00000000;
|
||
|
|
mov.f32 %f1351, %f1350;
|
||
|
|
mov.f32 %f1352, %f1350;
|
||
|
|
@%p5 bra BB0_2;
|
||
|
|
|
||
|
|
ld.u8 %rs18, [%rd22+1];
|
||
|
|
and.b16 %rs20, %rs16, 255;
|
||
|
|
cvt.rn.f32.u16 %f208, %rs20;
|
||
|
|
div.rn.f32 %f209, %f208, 0f437F0000;
|
||
|
|
fma.rn.f32 %f210, %f209, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f211, %rs18;
|
||
|
|
div.rn.f32 %f212, %f211, 0f437F0000;
|
||
|
|
fma.rn.f32 %f213, %f212, 0f40000000, 0fBF800000;
|
||
|
|
cvt.rn.f32.u16 %f214, %rs15;
|
||
|
|
div.rn.f32 %f215, %f214, 0f437F0000;
|
||
|
|
fma.rn.f32 %f216, %f215, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f217, %f213, %f213;
|
||
|
|
fma.rn.f32 %f218, %f210, %f210, %f217;
|
||
|
|
fma.rn.f32 %f219, %f216, %f216, %f218;
|
||
|
|
sqrt.rn.f32 %f220, %f219;
|
||
|
|
rcp.rn.f32 %f221, %f220;
|
||
|
|
mul.f32 %f1350, %f210, %f221;
|
||
|
|
mul.f32 %f1351, %f213, %f221;
|
||
|
|
mul.f32 %f1352, %f216, %f221;
|
||
|
|
|
||
|
|
BB0_2:
|
||
|
|
ld.global.v2.u32 {%r114, %r115}, [pixelID];
|
||
|
|
ld.global.v2.u32 {%r117, %r118}, [tileInfo];
|
||
|
|
add.s32 %r2, %r114, %r117;
|
||
|
|
add.s32 %r3, %r115, %r118;
|
||
|
|
setp.eq.f32 %p6, %f1351, 0f00000000;
|
||
|
|
setp.eq.f32 %p7, %f1350, 0f00000000;
|
||
|
|
and.pred %p8, %p7, %p6;
|
||
|
|
setp.eq.f32 %p9, %f1352, 0f00000000;
|
||
|
|
and.pred %p10, %p8, %p9;
|
||
|
|
@%p10 bra BB0_140;
|
||
|
|
bra.uni BB0_3;
|
||
|
|
|
||
|
|
BB0_140:
|
||
|
|
ld.global.u32 %r501, [imageEnabled];
|
||
|
|
and.b32 %r427, %r501, 1;
|
||
|
|
setp.eq.b32 %p147, %r427, 1;
|
||
|
|
@!%p147 bra BB0_142;
|
||
|
|
bra.uni BB0_141;
|
||
|
|
|
||
|
|
BB0_141:
|
||
|
|
cvt.u64.u32 %rd209, %r2;
|
||
|
|
cvt.u64.u32 %rd210, %r3;
|
||
|
|
mov.u64 %rd213, image;
|
||
|
|
cvta.global.u64 %rd208, %rd213;
|
||
|
|
mov.u64 %rd212, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd207), _rt_buffer_get_64, (%rd208, %r107, %r108, %rd209, %rd210, %rd212, %rd212);
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs101, 0;
|
||
|
|
st.v4.u8 [%rd207], {%rs101, %rs101, %rs101, %rs101};
|
||
|
|
ld.global.u32 %r501, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_142:
|
||
|
|
and.b32 %r430, %r501, 8;
|
||
|
|
setp.eq.s32 %p148, %r430, 0;
|
||
|
|
@%p148 bra BB0_144;
|
||
|
|
|
||
|
|
cvt.u64.u32 %rd217, %r3;
|
||
|
|
cvt.u64.u32 %rd216, %r2;
|
||
|
|
mov.u64 %rd220, image_Mask;
|
||
|
|
cvta.global.u64 %rd215, %rd220;
|
||
|
|
mov.u64 %rd219, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd214), _rt_buffer_get_64, (%rd215, %r107, %r107, %rd216, %rd217, %rd219, %rd219);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1312, 0f00000000;
|
||
|
|
cvt.rzi.u32.f32 %r433, %f1312;
|
||
|
|
cvt.u16.u32 %rs102, %r433;
|
||
|
|
mov.u16 %rs103, 0;
|
||
|
|
st.v2.u8 [%rd214], {%rs102, %rs103};
|
||
|
|
ld.global.u32 %r501, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_144:
|
||
|
|
cvt.u64.u32 %rd20, %r2;
|
||
|
|
cvt.u64.u32 %rd21, %r3;
|
||
|
|
and.b32 %r434, %r501, 4;
|
||
|
|
setp.eq.s32 %p149, %r434, 0;
|
||
|
|
@%p149 bra BB0_148;
|
||
|
|
|
||
|
|
ld.global.u32 %r435, [additive];
|
||
|
|
setp.eq.s32 %p150, %r435, 0;
|
||
|
|
@%p150 bra BB0_147;
|
||
|
|
|
||
|
|
mov.u64 %rd233, image_HDR;
|
||
|
|
cvta.global.u64 %rd222, %rd233;
|
||
|
|
mov.u32 %r439, 8;
|
||
|
|
mov.u64 %rd232, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd221), _rt_buffer_get_64, (%rd222, %r107, %r439, %rd20, %rd21, %rd232, %rd232);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs110, %rs111, %rs112, %rs113}, [%rd221];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1313, %rs110;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1314, %rs111;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1315, %rs112;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd227), _rt_buffer_get_64, (%rd222, %r107, %r439, %rd20, %rd21, %rd232, %rd232);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1316, %f1313, 0f00000000;
|
||
|
|
add.f32 %f1317, %f1314, 0f00000000;
|
||
|
|
add.f32 %f1318, %f1315, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs109, %f1318;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs108, %f1317;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs107, %f1316;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs114, 0;
|
||
|
|
st.v4.u16 [%rd227], {%rs107, %rs108, %rs109, %rs114};
|
||
|
|
bra.uni BB0_148;
|
||
|
|
|
||
|
|
BB0_3:
|
||
|
|
ld.global.f32 %f8, [directDir+4];
|
||
|
|
ld.global.f32 %f9, [directDir+8];
|
||
|
|
ld.global.f32 %f7, [directDir];
|
||
|
|
ld.global.v2.u32 {%r130, %r131}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd31, %r130;
|
||
|
|
cvt.u64.u32 %rd32, %r131;
|
||
|
|
mov.u64 %rd53, uvpos;
|
||
|
|
cvta.global.u64 %rd30, %rd53;
|
||
|
|
mov.u32 %r127, 12;
|
||
|
|
// inline asm
|
||
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd31, %rd32, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f224, [%rd29+8];
|
||
|
|
ld.f32 %f225, [%rd29+4];
|
||
|
|
ld.f32 %f226, [%rd29];
|
||
|
|
mul.f32 %f227, %f226, 0f3456BF95;
|
||
|
|
mul.f32 %f228, %f225, 0f3456BF95;
|
||
|
|
mul.f32 %f229, %f224, 0f3456BF95;
|
||
|
|
abs.f32 %f230, %f1350;
|
||
|
|
div.rn.f32 %f231, %f227, %f230;
|
||
|
|
abs.f32 %f232, %f1351;
|
||
|
|
div.rn.f32 %f233, %f228, %f232;
|
||
|
|
abs.f32 %f234, %f1352;
|
||
|
|
div.rn.f32 %f235, %f229, %f234;
|
||
|
|
abs.f32 %f236, %f231;
|
||
|
|
abs.f32 %f237, %f233;
|
||
|
|
abs.f32 %f238, %f235;
|
||
|
|
mov.f32 %f239, 0f38D1B717;
|
||
|
|
max.f32 %f240, %f236, %f239;
|
||
|
|
max.f32 %f241, %f237, %f239;
|
||
|
|
max.f32 %f242, %f238, %f239;
|
||
|
|
fma.rn.f32 %f10, %f1350, %f240, %f226;
|
||
|
|
fma.rn.f32 %f11, %f1351, %f241, %f225;
|
||
|
|
fma.rn.f32 %f12, %f1352, %f242, %f224;
|
||
|
|
setp.gt.f32 %p11, %f230, %f234;
|
||
|
|
neg.f32 %f243, %f1351;
|
||
|
|
selp.f32 %f244, %f243, 0f00000000, %p11;
|
||
|
|
neg.f32 %f245, %f1352;
|
||
|
|
selp.f32 %f246, %f1350, %f245, %p11;
|
||
|
|
selp.f32 %f247, 0f00000000, %f1351, %p11;
|
||
|
|
mul.f32 %f248, %f246, %f246;
|
||
|
|
fma.rn.f32 %f249, %f244, %f244, %f248;
|
||
|
|
fma.rn.f32 %f250, %f247, %f247, %f249;
|
||
|
|
sqrt.rn.f32 %f251, %f250;
|
||
|
|
rcp.rn.f32 %f252, %f251;
|
||
|
|
mul.f32 %f13, %f244, %f252;
|
||
|
|
mul.f32 %f14, %f246, %f252;
|
||
|
|
mul.f32 %f15, %f247, %f252;
|
||
|
|
mul.f32 %f253, %f1352, %f14;
|
||
|
|
mul.f32 %f254, %f1351, %f15;
|
||
|
|
sub.f32 %f16, %f253, %f254;
|
||
|
|
mul.f32 %f255, %f1350, %f15;
|
||
|
|
mul.f32 %f256, %f1352, %f13;
|
||
|
|
sub.f32 %f17, %f255, %f256;
|
||
|
|
mul.f32 %f257, %f1351, %f13;
|
||
|
|
mul.f32 %f258, %f1350, %f14;
|
||
|
|
sub.f32 %f18, %f257, %f258;
|
||
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd38, %r135;
|
||
|
|
add.s32 %r138, %r134, 1;
|
||
|
|
ld.global.u32 %r139, [resolution];
|
||
|
|
add.s32 %r140, %r139, -1;
|
||
|
|
min.u32 %r141, %r138, %r140;
|
||
|
|
cvt.u64.u32 %rd37, %r141;
|
||
|
|
// inline asm
|
||
|
|
call (%rd35), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd37, %rd38, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f259, [%rd35+8];
|
||
|
|
ld.f32 %f260, [%rd35+4];
|
||
|
|
ld.f32 %f261, [%rd35];
|
||
|
|
setp.eq.f32 %p12, %f261, 0f00000000;
|
||
|
|
setp.eq.f32 %p13, %f260, 0f00000000;
|
||
|
|
and.pred %p14, %p12, %p13;
|
||
|
|
setp.eq.f32 %p15, %f259, 0f00000000;
|
||
|
|
and.pred %p16, %p14, %p15;
|
||
|
|
selp.f32 %f262, %f226, %f261, %p16;
|
||
|
|
selp.f32 %f263, %f225, %f260, %p16;
|
||
|
|
selp.f32 %f264, %f224, %f259, %p16;
|
||
|
|
ld.global.v2.u32 {%r142, %r143}, [pixelID];
|
||
|
|
add.s32 %r146, %r143, 1;
|
||
|
|
ld.global.u32 %r147, [resolution+4];
|
||
|
|
add.s32 %r148, %r147, -1;
|
||
|
|
min.u32 %r149, %r146, %r148;
|
||
|
|
cvt.u64.u32 %rd43, %r142;
|
||
|
|
cvt.u64.u32 %rd44, %r149;
|
||
|
|
// inline asm
|
||
|
|
call (%rd41), _rt_buffer_get_64, (%rd30, %r107, %r127, %rd43, %rd44, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.f32 %f265, [%rd41+8];
|
||
|
|
ld.f32 %f266, [%rd41+4];
|
||
|
|
ld.f32 %f267, [%rd41];
|
||
|
|
setp.eq.f32 %p17, %f267, 0f00000000;
|
||
|
|
setp.eq.f32 %p18, %f266, 0f00000000;
|
||
|
|
and.pred %p19, %p17, %p18;
|
||
|
|
setp.eq.f32 %p20, %f265, 0f00000000;
|
||
|
|
and.pred %p21, %p19, %p20;
|
||
|
|
selp.f32 %f268, %f226, %f267, %p21;
|
||
|
|
selp.f32 %f269, %f225, %f266, %p21;
|
||
|
|
selp.f32 %f270, %f224, %f265, %p21;
|
||
|
|
sub.f32 %f271, %f262, %f226;
|
||
|
|
sub.f32 %f272, %f263, %f225;
|
||
|
|
sub.f32 %f273, %f264, %f224;
|
||
|
|
sub.f32 %f274, %f268, %f226;
|
||
|
|
sub.f32 %f275, %f269, %f225;
|
||
|
|
sub.f32 %f276, %f270, %f224;
|
||
|
|
abs.f32 %f277, %f271;
|
||
|
|
abs.f32 %f278, %f274;
|
||
|
|
max.f32 %f279, %f277, %f278;
|
||
|
|
abs.f32 %f280, %f272;
|
||
|
|
abs.f32 %f281, %f275;
|
||
|
|
max.f32 %f282, %f280, %f281;
|
||
|
|
abs.f32 %f283, %f273;
|
||
|
|
abs.f32 %f284, %f276;
|
||
|
|
max.f32 %f285, %f283, %f284;
|
||
|
|
max.f32 %f286, %f279, %f282;
|
||
|
|
max.f32 %f287, %f286, %f285;
|
||
|
|
setp.neu.f32 %p22, %f271, 0f00000000;
|
||
|
|
setp.neu.f32 %p23, %f272, 0f00000000;
|
||
|
|
or.pred %p24, %p22, %p23;
|
||
|
|
setp.neu.f32 %p25, %f273, 0f00000000;
|
||
|
|
setp.neu.f32 %p26, %f274, 0f00000000;
|
||
|
|
setp.neu.f32 %p27, %f275, 0f00000000;
|
||
|
|
or.pred %p28, %p26, %p27;
|
||
|
|
setp.neu.f32 %p29, %f276, 0f00000000;
|
||
|
|
or.pred %p30, %p28, %p29;
|
||
|
|
or.pred %p31, %p24, %p25;
|
||
|
|
and.pred %p32, %p31, %p30;
|
||
|
|
selp.f32 %f19, %f287, 0f00000000, %p32;
|
||
|
|
abs.f32 %f288, %f7;
|
||
|
|
abs.f32 %f289, %f9;
|
||
|
|
setp.gt.f32 %p33, %f288, %f289;
|
||
|
|
neg.f32 %f290, %f8;
|
||
|
|
neg.f32 %f291, %f9;
|
||
|
|
selp.f32 %f292, %f290, 0f00000000, %p33;
|
||
|
|
selp.f32 %f293, %f7, %f291, %p33;
|
||
|
|
selp.f32 %f294, 0f00000000, %f8, %p33;
|
||
|
|
mul.f32 %f295, %f293, %f293;
|
||
|
|
fma.rn.f32 %f296, %f292, %f292, %f295;
|
||
|
|
fma.rn.f32 %f297, %f294, %f294, %f296;
|
||
|
|
sqrt.rn.f32 %f298, %f297;
|
||
|
|
rcp.rn.f32 %f299, %f298;
|
||
|
|
mul.f32 %f20, %f292, %f299;
|
||
|
|
mul.f32 %f21, %f293, %f299;
|
||
|
|
mul.f32 %f22, %f294, %f299;
|
||
|
|
ld.global.v2.u32 {%r150, %r151}, [pixelID];
|
||
|
|
cvt.u64.u32 %rd49, %r150;
|
||
|
|
cvt.u64.u32 %rd50, %r151;
|
||
|
|
mov.u64 %rd54, rnd_seeds;
|
||
|
|
cvta.global.u64 %rd48, %rd54;
|
||
|
|
// inline asm
|
||
|
|
call (%rd47), _rt_buffer_get_64, (%rd48, %r107, %r108, %rd49, %rd50, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.global.u32 %r470, [samples];
|
||
|
|
mov.f32 %f1355, 0f00000000;
|
||
|
|
setp.lt.s32 %p34, %r470, 1;
|
||
|
|
@%p34 bra BB0_4;
|
||
|
|
|
||
|
|
cvt.rn.f32.s32 %f302, %r470;
|
||
|
|
rcp.rn.f32 %f23, %f302;
|
||
|
|
ld.u32 %r496, [%rd47];
|
||
|
|
mul.f32 %f24, %f10, 0f3456BF95;
|
||
|
|
mul.f32 %f25, %f11, 0f3456BF95;
|
||
|
|
mul.f32 %f26, %f12, 0f3456BF95;
|
||
|
|
mul.f32 %f303, %f7, %f21;
|
||
|
|
mul.f32 %f304, %f8, %f20;
|
||
|
|
sub.f32 %f27, %f304, %f303;
|
||
|
|
mul.f32 %f305, %f9, %f20;
|
||
|
|
mul.f32 %f306, %f7, %f22;
|
||
|
|
sub.f32 %f28, %f306, %f305;
|
||
|
|
mul.f32 %f307, %f8, %f22;
|
||
|
|
mul.f32 %f308, %f9, %f21;
|
||
|
|
sub.f32 %f29, %f308, %f307;
|
||
|
|
mov.f32 %f1355, 0f00000000;
|
||
|
|
mov.u32 %r471, 0;
|
||
|
|
abs.f32 %f309, %f25;
|
||
|
|
abs.f32 %f310, %f24;
|
||
|
|
max.f32 %f311, %f310, %f309;
|
||
|
|
abs.f32 %f312, %f26;
|
||
|
|
max.f32 %f313, %f311, %f312;
|
||
|
|
mov.f32 %f1356, %f1355;
|
||
|
|
|
||
|
|
BB0_6:
|
||
|
|
setp.lt.s32 %p35, %r470, 1;
|
||
|
|
@%p35 bra BB0_73;
|
||
|
|
|
||
|
|
mov.u32 %r473, 0;
|
||
|
|
|
||
|
|
BB0_8:
|
||
|
|
cvt.rn.f32.s32 %f1348, %r471;
|
||
|
|
mad.lo.s32 %r156, %r496, 1664525, 1013904223;
|
||
|
|
and.b32 %r157, %r156, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f315, %r157;
|
||
|
|
fma.rn.f32 %f316, %f315, 0f33800000, %f1348;
|
||
|
|
mul.f32 %f317, %f23, %f316;
|
||
|
|
mad.lo.s32 %r11, %r156, 1664525, 1013904223;
|
||
|
|
and.b32 %r158, %r11, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f318, %r158;
|
||
|
|
cvt.rn.f32.s32 %f319, %r473;
|
||
|
|
fma.rn.f32 %f320, %f318, 0f33800000, %f319;
|
||
|
|
mul.f32 %f321, %f23, %f320;
|
||
|
|
sqrt.rn.f32 %f36, %f317;
|
||
|
|
mul.f32 %f1363, %f321, 0f40C90FDB;
|
||
|
|
abs.f32 %f38, %f1363;
|
||
|
|
setp.neu.f32 %p36, %f38, 0f7F800000;
|
||
|
|
mov.f32 %f1357, %f1363;
|
||
|
|
@%p36 bra BB0_10;
|
||
|
|
|
||
|
|
mov.f32 %f322, 0f00000000;
|
||
|
|
mul.rn.f32 %f1357, %f1363, %f322;
|
||
|
|
|
||
|
|
BB0_10:
|
||
|
|
mul.f32 %f323, %f1357, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r484, %f323;
|
||
|
|
cvt.rn.f32.s32 %f324, %r484;
|
||
|
|
neg.f32 %f325, %f324;
|
||
|
|
mov.f32 %f326, 0f3FC90FDA;
|
||
|
|
fma.rn.f32 %f327, %f325, %f326, %f1357;
|
||
|
|
mov.f32 %f328, 0f33A22168;
|
||
|
|
fma.rn.f32 %f329, %f325, %f328, %f327;
|
||
|
|
mov.f32 %f330, 0f27C234C5;
|
||
|
|
fma.rn.f32 %f1358, %f325, %f330, %f329;
|
||
|
|
abs.f32 %f331, %f1357;
|
||
|
|
setp.leu.f32 %p37, %f331, 0f47CE4780;
|
||
|
|
@%p37 bra BB0_21;
|
||
|
|
|
||
|
|
add.u64 %rd56, %SP, 8;
|
||
|
|
cvta.to.local.u64 %rd3, %rd56;
|
||
|
|
mov.b32 %r13, %f1357;
|
||
|
|
shr.u32 %r14, %r13, 23;
|
||
|
|
shl.b32 %r161, %r13, 8;
|
||
|
|
or.b32 %r15, %r161, -2147483648;
|
||
|
|
mov.u32 %r475, 0;
|
||
|
|
mov.u64 %rd322, 0;
|
||
|
|
mov.u64 %rd321, %rd3;
|
||
|
|
mov.u32 %r476, %r475;
|
||
|
|
|
||
|
|
BB0_12:
|
||
|
|
.pragma "nounroll";
|
||
|
|
shl.b64 %rd57, %rd322, 2;
|
||
|
|
mov.u64 %rd58, __cudart_i2opi_f;
|
||
|
|
add.s64 %rd59, %rd58, %rd57;
|
||
|
|
ld.const.u32 %r164, [%rd59];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r162, %r164, %r15, %r476;
|
||
|
|
madc.hi.u32 %r476, %r164, %r15, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd321], %r162;
|
||
|
|
add.s32 %r475, %r475, 1;
|
||
|
|
cvt.s64.s32 %rd322, %r475;
|
||
|
|
mul.wide.s32 %rd62, %r475, 4;
|
||
|
|
add.s64 %rd321, %rd3, %rd62;
|
||
|
|
setp.ne.s32 %p38, %r475, 6;
|
||
|
|
@%p38 bra BB0_12;
|
||
|
|
|
||
|
|
and.b32 %r167, %r14, 255;
|
||
|
|
add.s32 %r168, %r167, -128;
|
||
|
|
shr.u32 %r169, %r168, 5;
|
||
|
|
and.b32 %r20, %r13, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd64, %rd56;
|
||
|
|
st.local.u32 [%rd64+24], %r476;
|
||
|
|
mov.u32 %r170, 6;
|
||
|
|
sub.s32 %r171, %r170, %r169;
|
||
|
|
mul.wide.s32 %rd65, %r171, 4;
|
||
|
|
add.s64 %rd8, %rd64, %rd65;
|
||
|
|
ld.local.u32 %r477, [%rd8];
|
||
|
|
ld.local.u32 %r478, [%rd8+-4];
|
||
|
|
and.b32 %r23, %r14, 31;
|
||
|
|
setp.eq.s32 %p39, %r23, 0;
|
||
|
|
@%p39 bra BB0_15;
|
||
|
|
|
||
|
|
mov.u32 %r172, 32;
|
||
|
|
sub.s32 %r173, %r172, %r23;
|
||
|
|
shr.u32 %r174, %r478, %r173;
|
||
|
|
shl.b32 %r175, %r477, %r23;
|
||
|
|
add.s32 %r477, %r174, %r175;
|
||
|
|
ld.local.u32 %r176, [%rd8+-8];
|
||
|
|
shr.u32 %r177, %r176, %r173;
|
||
|
|
shl.b32 %r178, %r478, %r23;
|
||
|
|
add.s32 %r478, %r177, %r178;
|
||
|
|
|
||
|
|
BB0_15:
|
||
|
|
shr.u32 %r179, %r478, 30;
|
||
|
|
shl.b32 %r180, %r477, 2;
|
||
|
|
add.s32 %r479, %r179, %r180;
|
||
|
|
shl.b32 %r29, %r478, 2;
|
||
|
|
shr.u32 %r181, %r479, 31;
|
||
|
|
shr.u32 %r182, %r477, 30;
|
||
|
|
add.s32 %r30, %r181, %r182;
|
||
|
|
setp.eq.s32 %p40, %r181, 0;
|
||
|
|
@%p40 bra BB0_16;
|
||
|
|
bra.uni BB0_17;
|
||
|
|
|
||
|
|
BB0_16:
|
||
|
|
mov.u32 %r480, %r20;
|
||
|
|
mov.u32 %r481, %r29;
|
||
|
|
bra.uni BB0_18;
|
||
|
|
|
||
|
|
BB0_17:
|
||
|
|
not.b32 %r183, %r479;
|
||
|
|
neg.s32 %r481, %r29;
|
||
|
|
setp.eq.s32 %p41, %r29, 0;
|
||
|
|
selp.u32 %r184, 1, 0, %p41;
|
||
|
|
add.s32 %r479, %r184, %r183;
|
||
|
|
xor.b32 %r480, %r20, -2147483648;
|
||
|
|
|
||
|
|
BB0_18:
|
||
|
|
clz.b32 %r483, %r479;
|
||
|
|
setp.eq.s32 %p42, %r483, 0;
|
||
|
|
shl.b32 %r185, %r479, %r483;
|
||
|
|
mov.u32 %r186, 32;
|
||
|
|
sub.s32 %r187, %r186, %r483;
|
||
|
|
shr.u32 %r188, %r481, %r187;
|
||
|
|
add.s32 %r189, %r188, %r185;
|
||
|
|
selp.b32 %r38, %r479, %r189, %p42;
|
||
|
|
mov.u32 %r190, -921707870;
|
||
|
|
mul.hi.u32 %r482, %r38, %r190;
|
||
|
|
setp.eq.s32 %p43, %r20, 0;
|
||
|
|
neg.s32 %r191, %r30;
|
||
|
|
selp.b32 %r484, %r30, %r191, %p43;
|
||
|
|
setp.lt.s32 %p44, %r482, 1;
|
||
|
|
@%p44 bra BB0_20;
|
||
|
|
|
||
|
|
mul.lo.s32 %r192, %r38, -921707870;
|
||
|
|
shr.u32 %r193, %r192, 31;
|
||
|
|
shl.b32 %r194, %r482, 1;
|
||
|
|
add.s32 %r482, %r193, %r194;
|
||
|
|
add.s32 %r483, %r483, 1;
|
||
|
|
|
||
|
|
BB0_20:
|
||
|
|
mov.u32 %r195, 126;
|
||
|
|
sub.s32 %r196, %r195, %r483;
|
||
|
|
shl.b32 %r197, %r196, 23;
|
||
|
|
add.s32 %r198, %r482, 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, %r480;
|
||
|
|
mov.b32 %f1358, %r203;
|
||
|
|
|
||
|
|
BB0_21:
|
||
|
|
mul.rn.f32 %f44, %f1358, %f1358;
|
||
|
|
add.s32 %r46, %r484, 1;
|
||
|
|
and.b32 %r47, %r46, 1;
|
||
|
|
setp.eq.s32 %p45, %r47, 0;
|
||
|
|
@%p45 bra BB0_23;
|
||
|
|
bra.uni BB0_22;
|
||
|
|
|
||
|
|
BB0_23:
|
||
|
|
mov.f32 %f334, 0f3C08839E;
|
||
|
|
mov.f32 %f335, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f1359, %f335, %f44, %f334;
|
||
|
|
bra.uni BB0_24;
|
||
|
|
|
||
|
|
BB0_22:
|
||
|
|
mov.f32 %f332, 0fBAB6061A;
|
||
|
|
mov.f32 %f333, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f1359, %f333, %f44, %f332;
|
||
|
|
|
||
|
|
BB0_24:
|
||
|
|
@%p45 bra BB0_26;
|
||
|
|
bra.uni BB0_25;
|
||
|
|
|
||
|
|
BB0_26:
|
||
|
|
mov.f32 %f339, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f340, %f1359, %f44, %f339;
|
||
|
|
mov.f32 %f341, 0f00000000;
|
||
|
|
fma.rn.f32 %f1360, %f340, %f44, %f341;
|
||
|
|
bra.uni BB0_27;
|
||
|
|
|
||
|
|
BB0_25:
|
||
|
|
mov.f32 %f336, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f337, %f1359, %f44, %f336;
|
||
|
|
mov.f32 %f338, 0fBF000000;
|
||
|
|
fma.rn.f32 %f1360, %f337, %f44, %f338;
|
||
|
|
|
||
|
|
BB0_27:
|
||
|
|
fma.rn.f32 %f1361, %f1360, %f1358, %f1358;
|
||
|
|
@%p45 bra BB0_29;
|
||
|
|
|
||
|
|
mov.f32 %f342, 0f3F800000;
|
||
|
|
fma.rn.f32 %f1361, %f1360, %f44, %f342;
|
||
|
|
|
||
|
|
BB0_29:
|
||
|
|
and.b32 %r204, %r46, 2;
|
||
|
|
setp.eq.s32 %p48, %r204, 0;
|
||
|
|
@%p48 bra BB0_31;
|
||
|
|
|
||
|
|
mov.f32 %f343, 0f00000000;
|
||
|
|
mov.f32 %f344, 0fBF800000;
|
||
|
|
fma.rn.f32 %f1361, %f1361, %f344, %f343;
|
||
|
|
|
||
|
|
BB0_31:
|
||
|
|
@%p36 bra BB0_33;
|
||
|
|
|
||
|
|
mov.f32 %f345, 0f00000000;
|
||
|
|
mul.rn.f32 %f1363, %f1363, %f345;
|
||
|
|
|
||
|
|
BB0_33:
|
||
|
|
mul.f32 %f346, %f1363, 0f3F22F983;
|
||
|
|
cvt.rni.s32.f32 %r494, %f346;
|
||
|
|
cvt.rn.f32.s32 %f347, %r494;
|
||
|
|
neg.f32 %f348, %f347;
|
||
|
|
fma.rn.f32 %f350, %f348, %f326, %f1363;
|
||
|
|
fma.rn.f32 %f352, %f348, %f328, %f350;
|
||
|
|
fma.rn.f32 %f1364, %f348, %f330, %f352;
|
||
|
|
abs.f32 %f354, %f1363;
|
||
|
|
setp.leu.f32 %p50, %f354, 0f47CE4780;
|
||
|
|
@%p50 bra BB0_44;
|
||
|
|
|
||
|
|
add.u64 %rd67, %SP, 8;
|
||
|
|
cvta.to.local.u64 %rd9, %rd67;
|
||
|
|
mov.b32 %r49, %f1363;
|
||
|
|
shr.u32 %r50, %r49, 23;
|
||
|
|
shl.b32 %r207, %r49, 8;
|
||
|
|
or.b32 %r51, %r207, -2147483648;
|
||
|
|
mov.u32 %r485, 0;
|
||
|
|
mov.u64 %rd323, %rd9;
|
||
|
|
mov.u64 %rd324, %rd27;
|
||
|
|
mov.u32 %r486, %r485;
|
||
|
|
|
||
|
|
BB0_35:
|
||
|
|
.pragma "nounroll";
|
||
|
|
shl.b64 %rd68, %rd324, 2;
|
||
|
|
mov.u64 %rd69, __cudart_i2opi_f;
|
||
|
|
add.s64 %rd70, %rd69, %rd68;
|
||
|
|
ld.const.u32 %r210, [%rd70];
|
||
|
|
// inline asm
|
||
|
|
{
|
||
|
|
mad.lo.cc.u32 %r208, %r210, %r51, %r486;
|
||
|
|
madc.hi.u32 %r486, %r210, %r51, 0;
|
||
|
|
}
|
||
|
|
// inline asm
|
||
|
|
st.local.u32 [%rd323], %r208;
|
||
|
|
add.s32 %r485, %r485, 1;
|
||
|
|
cvt.s64.s32 %rd324, %r485;
|
||
|
|
mul.wide.s32 %rd71, %r485, 4;
|
||
|
|
add.s64 %rd323, %rd9, %rd71;
|
||
|
|
setp.ne.s32 %p51, %r485, 6;
|
||
|
|
@%p51 bra BB0_35;
|
||
|
|
|
||
|
|
and.b32 %r213, %r50, 255;
|
||
|
|
add.s32 %r214, %r213, -128;
|
||
|
|
shr.u32 %r215, %r214, 5;
|
||
|
|
and.b32 %r56, %r49, -2147483648;
|
||
|
|
cvta.to.local.u64 %rd73, %rd67;
|
||
|
|
st.local.u32 [%rd73+24], %r486;
|
||
|
|
mov.u32 %r216, 6;
|
||
|
|
sub.s32 %r217, %r216, %r215;
|
||
|
|
mul.wide.s32 %rd74, %r217, 4;
|
||
|
|
add.s64 %rd15, %rd73, %rd74;
|
||
|
|
ld.local.u32 %r487, [%rd15];
|
||
|
|
ld.local.u32 %r488, [%rd15+-4];
|
||
|
|
and.b32 %r59, %r50, 31;
|
||
|
|
setp.eq.s32 %p52, %r59, 0;
|
||
|
|
@%p52 bra BB0_38;
|
||
|
|
|
||
|
|
mov.u32 %r218, 32;
|
||
|
|
sub.s32 %r219, %r218, %r59;
|
||
|
|
shr.u32 %r220, %r488, %r219;
|
||
|
|
shl.b32 %r221, %r487, %r59;
|
||
|
|
add.s32 %r487, %r220, %r221;
|
||
|
|
ld.local.u32 %r222, [%rd15+-8];
|
||
|
|
shr.u32 %r223, %r222, %r219;
|
||
|
|
shl.b32 %r224, %r488, %r59;
|
||
|
|
add.s32 %r488, %r223, %r224;
|
||
|
|
|
||
|
|
BB0_38:
|
||
|
|
shr.u32 %r225, %r488, 30;
|
||
|
|
shl.b32 %r226, %r487, 2;
|
||
|
|
add.s32 %r489, %r225, %r226;
|
||
|
|
shl.b32 %r65, %r488, 2;
|
||
|
|
shr.u32 %r227, %r489, 31;
|
||
|
|
shr.u32 %r228, %r487, 30;
|
||
|
|
add.s32 %r66, %r227, %r228;
|
||
|
|
setp.eq.s32 %p53, %r227, 0;
|
||
|
|
@%p53 bra BB0_39;
|
||
|
|
bra.uni BB0_40;
|
||
|
|
|
||
|
|
BB0_39:
|
||
|
|
mov.u32 %r490, %r56;
|
||
|
|
mov.u32 %r491, %r65;
|
||
|
|
bra.uni BB0_41;
|
||
|
|
|
||
|
|
BB0_40:
|
||
|
|
not.b32 %r229, %r489;
|
||
|
|
neg.s32 %r491, %r65;
|
||
|
|
setp.eq.s32 %p54, %r65, 0;
|
||
|
|
selp.u32 %r230, 1, 0, %p54;
|
||
|
|
add.s32 %r489, %r230, %r229;
|
||
|
|
xor.b32 %r490, %r56, -2147483648;
|
||
|
|
|
||
|
|
BB0_41:
|
||
|
|
clz.b32 %r493, %r489;
|
||
|
|
setp.eq.s32 %p55, %r493, 0;
|
||
|
|
shl.b32 %r231, %r489, %r493;
|
||
|
|
mov.u32 %r232, 32;
|
||
|
|
sub.s32 %r233, %r232, %r493;
|
||
|
|
shr.u32 %r234, %r491, %r233;
|
||
|
|
add.s32 %r235, %r234, %r231;
|
||
|
|
selp.b32 %r74, %r489, %r235, %p55;
|
||
|
|
mov.u32 %r236, -921707870;
|
||
|
|
mul.hi.u32 %r492, %r74, %r236;
|
||
|
|
setp.eq.s32 %p56, %r56, 0;
|
||
|
|
neg.s32 %r237, %r66;
|
||
|
|
selp.b32 %r494, %r66, %r237, %p56;
|
||
|
|
setp.lt.s32 %p57, %r492, 1;
|
||
|
|
@%p57 bra BB0_43;
|
||
|
|
|
||
|
|
mul.lo.s32 %r238, %r74, -921707870;
|
||
|
|
shr.u32 %r239, %r238, 31;
|
||
|
|
shl.b32 %r240, %r492, 1;
|
||
|
|
add.s32 %r492, %r239, %r240;
|
||
|
|
add.s32 %r493, %r493, 1;
|
||
|
|
|
||
|
|
BB0_43:
|
||
|
|
mov.u32 %r241, 126;
|
||
|
|
sub.s32 %r242, %r241, %r493;
|
||
|
|
shl.b32 %r243, %r242, 23;
|
||
|
|
add.s32 %r244, %r492, 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, %r490;
|
||
|
|
mov.b32 %f1364, %r249;
|
||
|
|
|
||
|
|
BB0_44:
|
||
|
|
mul.rn.f32 %f61, %f1364, %f1364;
|
||
|
|
and.b32 %r82, %r494, 1;
|
||
|
|
setp.eq.s32 %p58, %r82, 0;
|
||
|
|
@%p58 bra BB0_46;
|
||
|
|
bra.uni BB0_45;
|
||
|
|
|
||
|
|
BB0_46:
|
||
|
|
mov.f32 %f357, 0f3C08839E;
|
||
|
|
mov.f32 %f358, 0fB94CA1F9;
|
||
|
|
fma.rn.f32 %f1365, %f358, %f61, %f357;
|
||
|
|
bra.uni BB0_47;
|
||
|
|
|
||
|
|
BB0_45:
|
||
|
|
mov.f32 %f355, 0fBAB6061A;
|
||
|
|
mov.f32 %f356, 0f37CCF5CE;
|
||
|
|
fma.rn.f32 %f1365, %f356, %f61, %f355;
|
||
|
|
|
||
|
|
BB0_47:
|
||
|
|
@%p58 bra BB0_49;
|
||
|
|
bra.uni BB0_48;
|
||
|
|
|
||
|
|
BB0_49:
|
||
|
|
mov.f32 %f362, 0fBE2AAAA3;
|
||
|
|
fma.rn.f32 %f363, %f1365, %f61, %f362;
|
||
|
|
mov.f32 %f364, 0f00000000;
|
||
|
|
fma.rn.f32 %f1366, %f363, %f61, %f364;
|
||
|
|
bra.uni BB0_50;
|
||
|
|
|
||
|
|
BB0_48:
|
||
|
|
mov.f32 %f359, 0f3D2AAAA5;
|
||
|
|
fma.rn.f32 %f360, %f1365, %f61, %f359;
|
||
|
|
mov.f32 %f361, 0fBF000000;
|
||
|
|
fma.rn.f32 %f1366, %f360, %f61, %f361;
|
||
|
|
|
||
|
|
BB0_50:
|
||
|
|
fma.rn.f32 %f1367, %f1366, %f1364, %f1364;
|
||
|
|
@%p58 bra BB0_52;
|
||
|
|
|
||
|
|
mov.f32 %f365, 0f3F800000;
|
||
|
|
fma.rn.f32 %f1367, %f1366, %f61, %f365;
|
||
|
|
|
||
|
|
BB0_52:
|
||
|
|
and.b32 %r250, %r494, 2;
|
||
|
|
setp.eq.s32 %p61, %r250, 0;
|
||
|
|
@%p61 bra BB0_54;
|
||
|
|
|
||
|
|
mov.f32 %f366, 0f00000000;
|
||
|
|
mov.f32 %f367, 0fBF800000;
|
||
|
|
fma.rn.f32 %f1367, %f1367, %f367, %f366;
|
||
|
|
|
||
|
|
BB0_54:
|
||
|
|
max.f32 %f1349, %f313, %f239;
|
||
|
|
mul.f32 %f376, %f36, %f1361;
|
||
|
|
add.u64 %rd75, %SP, 4;
|
||
|
|
cvta.to.local.u64 %rd16, %rd75;
|
||
|
|
mul.f32 %f377, %f376, %f376;
|
||
|
|
mov.f32 %f378, 0f3F800000;
|
||
|
|
sub.f32 %f379, %f378, %f377;
|
||
|
|
mul.f32 %f380, %f36, %f1367;
|
||
|
|
mul.f32 %f381, %f380, %f380;
|
||
|
|
sub.f32 %f382, %f379, %f381;
|
||
|
|
mov.f32 %f383, 0f00000000;
|
||
|
|
max.f32 %f384, %f383, %f382;
|
||
|
|
sqrt.rn.f32 %f385, %f384;
|
||
|
|
mul.f32 %f386, %f20, %f380;
|
||
|
|
mul.f32 %f387, %f21, %f380;
|
||
|
|
mul.f32 %f388, %f22, %f380;
|
||
|
|
fma.rn.f32 %f389, %f29, %f376, %f386;
|
||
|
|
fma.rn.f32 %f390, %f28, %f376, %f387;
|
||
|
|
fma.rn.f32 %f391, %f27, %f376, %f388;
|
||
|
|
fma.rn.f32 %f392, %f7, %f385, %f389;
|
||
|
|
fma.rn.f32 %f393, %f8, %f385, %f390;
|
||
|
|
fma.rn.f32 %f394, %f9, %f385, %f391;
|
||
|
|
add.f32 %f395, %f7, %f392;
|
||
|
|
add.f32 %f396, %f8, %f393;
|
||
|
|
add.f32 %f397, %f9, %f394;
|
||
|
|
ld.global.f32 %f398, [shadowSpread];
|
||
|
|
mul.f32 %f399, %f398, %f395;
|
||
|
|
mul.f32 %f400, %f398, %f396;
|
||
|
|
mul.f32 %f401, %f398, %f397;
|
||
|
|
sub.f32 %f402, %f399, %f7;
|
||
|
|
sub.f32 %f403, %f400, %f8;
|
||
|
|
sub.f32 %f404, %f401, %f9;
|
||
|
|
mul.f32 %f405, %f403, %f403;
|
||
|
|
fma.rn.f32 %f406, %f402, %f402, %f405;
|
||
|
|
fma.rn.f32 %f407, %f404, %f404, %f406;
|
||
|
|
sqrt.rn.f32 %f408, %f407;
|
||
|
|
rcp.rn.f32 %f409, %f408;
|
||
|
|
mul.f32 %f73, %f409, %f402;
|
||
|
|
mul.f32 %f74, %f409, %f403;
|
||
|
|
mul.f32 %f75, %f409, %f404;
|
||
|
|
ld.global.u32 %r254, [imageEnabled];
|
||
|
|
and.b32 %r255, %r254, 32;
|
||
|
|
setp.eq.s32 %p62, %r255, 0;
|
||
|
|
selp.f32 %f410, 0f3F800000, 0f41200000, %p62;
|
||
|
|
mul.f32 %f374, %f410, %f1349;
|
||
|
|
mad.lo.s32 %r256, %r11, 1664525, 1013904223;
|
||
|
|
and.b32 %r257, %r256, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f411, %r257;
|
||
|
|
mul.f32 %f412, %f411, 0f33800000;
|
||
|
|
fma.rn.f32 %f413, %f412, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r83, %r256, 1664525, 1013904223;
|
||
|
|
and.b32 %r258, %r83, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f414, %r258;
|
||
|
|
mul.f32 %f415, %f414, 0f33800000;
|
||
|
|
fma.rn.f32 %f416, %f415, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f417, %f13, %f416;
|
||
|
|
mul.f32 %f418, %f14, %f416;
|
||
|
|
mul.f32 %f419, %f15, %f416;
|
||
|
|
fma.rn.f32 %f420, %f16, %f413, %f417;
|
||
|
|
fma.rn.f32 %f421, %f17, %f413, %f418;
|
||
|
|
fma.rn.f32 %f422, %f18, %f413, %f419;
|
||
|
|
mul.f32 %f423, %f421, %f421;
|
||
|
|
fma.rn.f32 %f424, %f420, %f420, %f423;
|
||
|
|
fma.rn.f32 %f425, %f422, %f422, %f424;
|
||
|
|
sqrt.rn.f32 %f426, %f425;
|
||
|
|
rcp.rn.f32 %f427, %f426;
|
||
|
|
mul.f32 %f428, %f420, %f427;
|
||
|
|
mul.f32 %f429, %f421, %f427;
|
||
|
|
mul.f32 %f430, %f422, %f427;
|
||
|
|
mul.f32 %f431, %f1351, %f430;
|
||
|
|
mul.f32 %f432, %f1352, %f429;
|
||
|
|
sub.f32 %f433, %f431, %f432;
|
||
|
|
mul.f32 %f434, %f1352, %f428;
|
||
|
|
mul.f32 %f435, %f1350, %f430;
|
||
|
|
sub.f32 %f436, %f434, %f435;
|
||
|
|
mul.f32 %f437, %f1350, %f429;
|
||
|
|
mul.f32 %f438, %f1351, %f428;
|
||
|
|
sub.f32 %f439, %f437, %f438;
|
||
|
|
mul.f32 %f440, %f428, 0f3E24A490;
|
||
|
|
mul.f32 %f441, %f429, 0f3E24A490;
|
||
|
|
mul.f32 %f442, %f430, 0f3E24A490;
|
||
|
|
fma.rn.f32 %f443, %f19, %f440, %f10;
|
||
|
|
fma.rn.f32 %f444, %f19, %f441, %f11;
|
||
|
|
fma.rn.f32 %f445, %f19, %f442, %f12;
|
||
|
|
mul.f32 %f446, %f433, 0fBECECEDE;
|
||
|
|
mul.f32 %f447, %f436, 0fBECECEDE;
|
||
|
|
mul.f32 %f448, %f439, 0fBECECEDE;
|
||
|
|
fma.rn.f32 %f77, %f19, %f446, %f443;
|
||
|
|
fma.rn.f32 %f78, %f19, %f447, %f444;
|
||
|
|
fma.rn.f32 %f79, %f19, %f448, %f445;
|
||
|
|
sub.f32 %f449, %f77, %f10;
|
||
|
|
sub.f32 %f450, %f78, %f11;
|
||
|
|
sub.f32 %f451, %f79, %f12;
|
||
|
|
mul.f32 %f452, %f450, %f450;
|
||
|
|
fma.rn.f32 %f453, %f449, %f449, %f452;
|
||
|
|
fma.rn.f32 %f454, %f451, %f451, %f453;
|
||
|
|
sqrt.rn.f32 %f375, %f454;
|
||
|
|
rcp.rn.f32 %f455, %f375;
|
||
|
|
mul.f32 %f371, %f455, %f449;
|
||
|
|
mul.f32 %f372, %f455, %f450;
|
||
|
|
mul.f32 %f373, %f455, %f451;
|
||
|
|
mov.u32 %r259, 1065353216;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r251, [root];
|
||
|
|
mov.u32 %r252, 1;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f371, %f372, %f373, %r252, %f374, %f375, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f456, [%rd16];
|
||
|
|
mov.u16 %rs170, 0;
|
||
|
|
setp.neu.f32 %p63, %f456, 0f3F800000;
|
||
|
|
@%p63 bra BB0_56;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r260, [root];
|
||
|
|
mov.f32 %f464, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r260, %f77, %f78, %f79, %f73, %f74, %f75, %r252, %f374, %f464, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f465, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f465;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_56:
|
||
|
|
mad.lo.s32 %r267, %r83, 1664525, 1013904223;
|
||
|
|
and.b32 %r268, %r267, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f474, %r268;
|
||
|
|
mul.f32 %f475, %f474, 0f33800000;
|
||
|
|
fma.rn.f32 %f476, %f475, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r84, %r267, 1664525, 1013904223;
|
||
|
|
and.b32 %r269, %r84, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f477, %r269;
|
||
|
|
mul.f32 %f478, %f477, 0f33800000;
|
||
|
|
fma.rn.f32 %f479, %f478, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f480, %f13, %f479;
|
||
|
|
mul.f32 %f481, %f14, %f479;
|
||
|
|
mul.f32 %f482, %f15, %f479;
|
||
|
|
fma.rn.f32 %f483, %f16, %f476, %f480;
|
||
|
|
fma.rn.f32 %f484, %f17, %f476, %f481;
|
||
|
|
fma.rn.f32 %f485, %f18, %f476, %f482;
|
||
|
|
mul.f32 %f486, %f484, %f484;
|
||
|
|
fma.rn.f32 %f487, %f483, %f483, %f486;
|
||
|
|
fma.rn.f32 %f488, %f485, %f485, %f487;
|
||
|
|
sqrt.rn.f32 %f489, %f488;
|
||
|
|
rcp.rn.f32 %f490, %f489;
|
||
|
|
mul.f32 %f491, %f483, %f490;
|
||
|
|
mul.f32 %f492, %f484, %f490;
|
||
|
|
mul.f32 %f493, %f485, %f490;
|
||
|
|
mul.f32 %f494, %f1351, %f493;
|
||
|
|
mul.f32 %f495, %f1352, %f492;
|
||
|
|
sub.f32 %f496, %f494, %f495;
|
||
|
|
mul.f32 %f497, %f1352, %f491;
|
||
|
|
mul.f32 %f498, %f1350, %f493;
|
||
|
|
sub.f32 %f499, %f497, %f498;
|
||
|
|
mul.f32 %f500, %f1350, %f492;
|
||
|
|
mul.f32 %f501, %f1351, %f491;
|
||
|
|
sub.f32 %f502, %f500, %f501;
|
||
|
|
mul.f32 %f503, %f491, 0fBE24A490;
|
||
|
|
mul.f32 %f504, %f492, 0fBE24A490;
|
||
|
|
mul.f32 %f505, %f493, 0fBE24A490;
|
||
|
|
fma.rn.f32 %f506, %f19, %f503, %f10;
|
||
|
|
fma.rn.f32 %f507, %f19, %f504, %f11;
|
||
|
|
fma.rn.f32 %f508, %f19, %f505, %f12;
|
||
|
|
mul.f32 %f509, %f496, 0f3ECACABC;
|
||
|
|
mul.f32 %f510, %f499, 0f3ECACABC;
|
||
|
|
mul.f32 %f511, %f502, 0f3ECACABC;
|
||
|
|
fma.rn.f32 %f84, %f19, %f509, %f506;
|
||
|
|
fma.rn.f32 %f85, %f19, %f510, %f507;
|
||
|
|
fma.rn.f32 %f86, %f19, %f511, %f508;
|
||
|
|
sub.f32 %f512, %f84, %f10;
|
||
|
|
sub.f32 %f513, %f85, %f11;
|
||
|
|
sub.f32 %f514, %f86, %f12;
|
||
|
|
mul.f32 %f515, %f513, %f513;
|
||
|
|
fma.rn.f32 %f516, %f512, %f512, %f515;
|
||
|
|
fma.rn.f32 %f517, %f514, %f514, %f516;
|
||
|
|
sqrt.rn.f32 %f473, %f517;
|
||
|
|
rcp.rn.f32 %f518, %f473;
|
||
|
|
mul.f32 %f469, %f518, %f512;
|
||
|
|
mul.f32 %f470, %f518, %f513;
|
||
|
|
mul.f32 %f471, %f518, %f514;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r264, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r264, %f10, %f11, %f12, %f469, %f470, %f471, %r252, %f374, %f473, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f519, [%rd16];
|
||
|
|
setp.neu.f32 %p64, %f519, 0f3F800000;
|
||
|
|
@%p64 bra BB0_58;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r271, [root];
|
||
|
|
mov.f32 %f527, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r271, %f84, %f85, %f86, %f73, %f74, %f75, %r252, %f374, %f527, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f528, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f528;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_58:
|
||
|
|
mad.lo.s32 %r278, %r84, 1664525, 1013904223;
|
||
|
|
and.b32 %r279, %r278, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f537, %r279;
|
||
|
|
mul.f32 %f538, %f537, 0f33800000;
|
||
|
|
fma.rn.f32 %f539, %f538, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r85, %r278, 1664525, 1013904223;
|
||
|
|
and.b32 %r280, %r85, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f540, %r280;
|
||
|
|
mul.f32 %f541, %f540, 0f33800000;
|
||
|
|
fma.rn.f32 %f542, %f541, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f543, %f13, %f542;
|
||
|
|
mul.f32 %f544, %f14, %f542;
|
||
|
|
mul.f32 %f545, %f15, %f542;
|
||
|
|
fma.rn.f32 %f546, %f16, %f539, %f543;
|
||
|
|
fma.rn.f32 %f547, %f17, %f539, %f544;
|
||
|
|
fma.rn.f32 %f548, %f18, %f539, %f545;
|
||
|
|
mul.f32 %f549, %f547, %f547;
|
||
|
|
fma.rn.f32 %f550, %f546, %f546, %f549;
|
||
|
|
fma.rn.f32 %f551, %f548, %f548, %f550;
|
||
|
|
sqrt.rn.f32 %f552, %f551;
|
||
|
|
rcp.rn.f32 %f553, %f552;
|
||
|
|
mul.f32 %f554, %f546, %f553;
|
||
|
|
mul.f32 %f555, %f547, %f553;
|
||
|
|
mul.f32 %f556, %f548, %f553;
|
||
|
|
mul.f32 %f557, %f1351, %f556;
|
||
|
|
mul.f32 %f558, %f1352, %f555;
|
||
|
|
sub.f32 %f559, %f557, %f558;
|
||
|
|
mul.f32 %f560, %f1352, %f554;
|
||
|
|
mul.f32 %f561, %f1350, %f556;
|
||
|
|
sub.f32 %f562, %f560, %f561;
|
||
|
|
mul.f32 %f563, %f1350, %f555;
|
||
|
|
mul.f32 %f564, %f1351, %f554;
|
||
|
|
sub.f32 %f565, %f563, %f564;
|
||
|
|
mul.f32 %f566, %f554, 0f3F23A3A8;
|
||
|
|
mul.f32 %f567, %f555, 0f3F23A3A8;
|
||
|
|
mul.f32 %f568, %f556, 0f3F23A3A8;
|
||
|
|
fma.rn.f32 %f569, %f19, %f566, %f10;
|
||
|
|
fma.rn.f32 %f570, %f19, %f567, %f11;
|
||
|
|
fma.rn.f32 %f571, %f19, %f568, %f12;
|
||
|
|
mul.f32 %f572, %f559, 0f3E24A490;
|
||
|
|
mul.f32 %f573, %f562, 0f3E24A490;
|
||
|
|
mul.f32 %f574, %f565, 0f3E24A490;
|
||
|
|
fma.rn.f32 %f91, %f19, %f572, %f569;
|
||
|
|
fma.rn.f32 %f92, %f19, %f573, %f570;
|
||
|
|
fma.rn.f32 %f93, %f19, %f574, %f571;
|
||
|
|
sub.f32 %f575, %f91, %f10;
|
||
|
|
sub.f32 %f576, %f92, %f11;
|
||
|
|
sub.f32 %f577, %f93, %f12;
|
||
|
|
mul.f32 %f578, %f576, %f576;
|
||
|
|
fma.rn.f32 %f579, %f575, %f575, %f578;
|
||
|
|
fma.rn.f32 %f580, %f577, %f577, %f579;
|
||
|
|
sqrt.rn.f32 %f536, %f580;
|
||
|
|
rcp.rn.f32 %f581, %f536;
|
||
|
|
mul.f32 %f532, %f581, %f575;
|
||
|
|
mul.f32 %f533, %f581, %f576;
|
||
|
|
mul.f32 %f534, %f581, %f577;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r275, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r275, %f10, %f11, %f12, %f532, %f533, %f534, %r252, %f374, %f536, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f582, [%rd16];
|
||
|
|
setp.neu.f32 %p65, %f582, 0f3F800000;
|
||
|
|
@%p65 bra BB0_60;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r282, [root];
|
||
|
|
mov.f32 %f590, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r282, %f91, %f92, %f93, %f73, %f74, %f75, %r252, %f374, %f590, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f591, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f591;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_60:
|
||
|
|
mad.lo.s32 %r289, %r85, 1664525, 1013904223;
|
||
|
|
and.b32 %r290, %r289, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f600, %r290;
|
||
|
|
mul.f32 %f601, %f600, 0f33800000;
|
||
|
|
fma.rn.f32 %f602, %f601, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r86, %r289, 1664525, 1013904223;
|
||
|
|
and.b32 %r291, %r86, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f603, %r291;
|
||
|
|
mul.f32 %f604, %f603, 0f33800000;
|
||
|
|
fma.rn.f32 %f605, %f604, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f606, %f13, %f605;
|
||
|
|
mul.f32 %f607, %f14, %f605;
|
||
|
|
mul.f32 %f608, %f15, %f605;
|
||
|
|
fma.rn.f32 %f609, %f16, %f602, %f606;
|
||
|
|
fma.rn.f32 %f610, %f17, %f602, %f607;
|
||
|
|
fma.rn.f32 %f611, %f18, %f602, %f608;
|
||
|
|
mul.f32 %f612, %f610, %f610;
|
||
|
|
fma.rn.f32 %f613, %f609, %f609, %f612;
|
||
|
|
fma.rn.f32 %f614, %f611, %f611, %f613;
|
||
|
|
sqrt.rn.f32 %f615, %f614;
|
||
|
|
rcp.rn.f32 %f616, %f615;
|
||
|
|
mul.f32 %f617, %f609, %f616;
|
||
|
|
mul.f32 %f618, %f610, %f616;
|
||
|
|
mul.f32 %f619, %f611, %f616;
|
||
|
|
mul.f32 %f620, %f1351, %f619;
|
||
|
|
mul.f32 %f621, %f1352, %f618;
|
||
|
|
sub.f32 %f622, %f620, %f621;
|
||
|
|
mul.f32 %f623, %f1352, %f617;
|
||
|
|
mul.f32 %f624, %f1350, %f619;
|
||
|
|
sub.f32 %f625, %f623, %f624;
|
||
|
|
mul.f32 %f626, %f1350, %f618;
|
||
|
|
mul.f32 %f627, %f1351, %f617;
|
||
|
|
sub.f32 %f628, %f626, %f627;
|
||
|
|
mul.f32 %f629, %f617, 0fBECECEDE;
|
||
|
|
mul.f32 %f630, %f618, 0fBECECEDE;
|
||
|
|
mul.f32 %f631, %f619, 0fBECECEDE;
|
||
|
|
fma.rn.f32 %f632, %f19, %f629, %f10;
|
||
|
|
fma.rn.f32 %f633, %f19, %f630, %f11;
|
||
|
|
fma.rn.f32 %f634, %f19, %f631, %f12;
|
||
|
|
mul.f32 %f635, %f622, 0fBF23A3A9;
|
||
|
|
mul.f32 %f636, %f625, 0fBF23A3A9;
|
||
|
|
mul.f32 %f637, %f628, 0fBF23A3A9;
|
||
|
|
fma.rn.f32 %f98, %f19, %f635, %f632;
|
||
|
|
fma.rn.f32 %f99, %f19, %f636, %f633;
|
||
|
|
fma.rn.f32 %f100, %f19, %f637, %f634;
|
||
|
|
sub.f32 %f638, %f98, %f10;
|
||
|
|
sub.f32 %f639, %f99, %f11;
|
||
|
|
sub.f32 %f640, %f100, %f12;
|
||
|
|
mul.f32 %f641, %f639, %f639;
|
||
|
|
fma.rn.f32 %f642, %f638, %f638, %f641;
|
||
|
|
fma.rn.f32 %f643, %f640, %f640, %f642;
|
||
|
|
sqrt.rn.f32 %f599, %f643;
|
||
|
|
rcp.rn.f32 %f644, %f599;
|
||
|
|
mul.f32 %f595, %f644, %f638;
|
||
|
|
mul.f32 %f596, %f644, %f639;
|
||
|
|
mul.f32 %f597, %f644, %f640;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r286, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r286, %f10, %f11, %f12, %f595, %f596, %f597, %r252, %f374, %f599, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f645, [%rd16];
|
||
|
|
setp.neu.f32 %p66, %f645, 0f3F800000;
|
||
|
|
@%p66 bra BB0_62;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r293, [root];
|
||
|
|
mov.f32 %f653, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r293, %f98, %f99, %f100, %f73, %f74, %f75, %r252, %f374, %f653, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f654, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f654;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_62:
|
||
|
|
mad.lo.s32 %r300, %r86, 1664525, 1013904223;
|
||
|
|
and.b32 %r301, %r300, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f663, %r301;
|
||
|
|
mul.f32 %f664, %f663, 0f33800000;
|
||
|
|
fma.rn.f32 %f665, %f664, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r87, %r300, 1664525, 1013904223;
|
||
|
|
and.b32 %r302, %r87, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f666, %r302;
|
||
|
|
mul.f32 %f667, %f666, 0f33800000;
|
||
|
|
fma.rn.f32 %f668, %f667, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f669, %f13, %f668;
|
||
|
|
mul.f32 %f670, %f14, %f668;
|
||
|
|
mul.f32 %f671, %f15, %f668;
|
||
|
|
fma.rn.f32 %f672, %f16, %f665, %f669;
|
||
|
|
fma.rn.f32 %f673, %f17, %f665, %f670;
|
||
|
|
fma.rn.f32 %f674, %f18, %f665, %f671;
|
||
|
|
mul.f32 %f675, %f673, %f673;
|
||
|
|
fma.rn.f32 %f676, %f672, %f672, %f675;
|
||
|
|
fma.rn.f32 %f677, %f674, %f674, %f676;
|
||
|
|
sqrt.rn.f32 %f678, %f677;
|
||
|
|
rcp.rn.f32 %f679, %f678;
|
||
|
|
mul.f32 %f680, %f672, %f679;
|
||
|
|
mul.f32 %f681, %f673, %f679;
|
||
|
|
mul.f32 %f682, %f674, %f679;
|
||
|
|
mul.f32 %f683, %f1351, %f682;
|
||
|
|
mul.f32 %f684, %f1352, %f681;
|
||
|
|
sub.f32 %f685, %f683, %f684;
|
||
|
|
mul.f32 %f686, %f1352, %f680;
|
||
|
|
mul.f32 %f687, %f1350, %f682;
|
||
|
|
sub.f32 %f688, %f686, %f687;
|
||
|
|
mul.f32 %f689, %f1350, %f681;
|
||
|
|
mul.f32 %f690, %f1351, %f680;
|
||
|
|
sub.f32 %f691, %f689, %f690;
|
||
|
|
mul.f32 %f692, %f680, 0fBF23A3A9;
|
||
|
|
mul.f32 %f693, %f681, 0fBF23A3A9;
|
||
|
|
mul.f32 %f694, %f682, 0fBF23A3A9;
|
||
|
|
fma.rn.f32 %f695, %f19, %f692, %f10;
|
||
|
|
fma.rn.f32 %f696, %f19, %f693, %f11;
|
||
|
|
fma.rn.f32 %f697, %f19, %f694, %f12;
|
||
|
|
mul.f32 %f698, %f685, 0f3F23A3A8;
|
||
|
|
mul.f32 %f699, %f688, 0f3F23A3A8;
|
||
|
|
mul.f32 %f700, %f691, 0f3F23A3A8;
|
||
|
|
fma.rn.f32 %f105, %f19, %f698, %f695;
|
||
|
|
fma.rn.f32 %f106, %f19, %f699, %f696;
|
||
|
|
fma.rn.f32 %f107, %f19, %f700, %f697;
|
||
|
|
sub.f32 %f701, %f105, %f10;
|
||
|
|
sub.f32 %f702, %f106, %f11;
|
||
|
|
sub.f32 %f703, %f107, %f12;
|
||
|
|
mul.f32 %f704, %f702, %f702;
|
||
|
|
fma.rn.f32 %f705, %f701, %f701, %f704;
|
||
|
|
fma.rn.f32 %f706, %f703, %f703, %f705;
|
||
|
|
sqrt.rn.f32 %f662, %f706;
|
||
|
|
rcp.rn.f32 %f707, %f662;
|
||
|
|
mul.f32 %f658, %f707, %f701;
|
||
|
|
mul.f32 %f659, %f707, %f702;
|
||
|
|
mul.f32 %f660, %f707, %f703;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r297, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r297, %f10, %f11, %f12, %f658, %f659, %f660, %r252, %f374, %f662, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f708, [%rd16];
|
||
|
|
setp.neu.f32 %p67, %f708, 0f3F800000;
|
||
|
|
@%p67 bra BB0_64;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r304, [root];
|
||
|
|
mov.f32 %f716, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r304, %f105, %f106, %f107, %f73, %f74, %f75, %r252, %f374, %f716, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f717, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f717;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_64:
|
||
|
|
mad.lo.s32 %r311, %r87, 1664525, 1013904223;
|
||
|
|
and.b32 %r312, %r311, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f726, %r312;
|
||
|
|
mul.f32 %f727, %f726, 0f33800000;
|
||
|
|
fma.rn.f32 %f728, %f727, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r88, %r311, 1664525, 1013904223;
|
||
|
|
and.b32 %r313, %r88, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f729, %r313;
|
||
|
|
mul.f32 %f730, %f729, 0f33800000;
|
||
|
|
fma.rn.f32 %f731, %f730, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f732, %f13, %f731;
|
||
|
|
mul.f32 %f733, %f14, %f731;
|
||
|
|
mul.f32 %f734, %f15, %f731;
|
||
|
|
fma.rn.f32 %f735, %f16, %f728, %f732;
|
||
|
|
fma.rn.f32 %f736, %f17, %f728, %f733;
|
||
|
|
fma.rn.f32 %f737, %f18, %f728, %f734;
|
||
|
|
mul.f32 %f738, %f736, %f736;
|
||
|
|
fma.rn.f32 %f739, %f735, %f735, %f738;
|
||
|
|
fma.rn.f32 %f740, %f737, %f737, %f739;
|
||
|
|
sqrt.rn.f32 %f741, %f740;
|
||
|
|
rcp.rn.f32 %f742, %f741;
|
||
|
|
mul.f32 %f743, %f735, %f742;
|
||
|
|
mul.f32 %f744, %f736, %f742;
|
||
|
|
mul.f32 %f745, %f737, %f742;
|
||
|
|
mul.f32 %f746, %f1351, %f745;
|
||
|
|
mul.f32 %f747, %f1352, %f744;
|
||
|
|
sub.f32 %f748, %f746, %f747;
|
||
|
|
mul.f32 %f749, %f1352, %f743;
|
||
|
|
mul.f32 %f750, %f1350, %f745;
|
||
|
|
sub.f32 %f751, %f749, %f750;
|
||
|
|
mul.f32 %f752, %f1350, %f744;
|
||
|
|
mul.f32 %f753, %f1351, %f743;
|
||
|
|
sub.f32 %f754, %f752, %f753;
|
||
|
|
mul.f32 %f755, %f743, 0fBF61E1E3;
|
||
|
|
mul.f32 %f756, %f744, 0fBF61E1E3;
|
||
|
|
mul.f32 %f757, %f745, 0fBF61E1E3;
|
||
|
|
fma.rn.f32 %f758, %f19, %f755, %f10;
|
||
|
|
fma.rn.f32 %f759, %f19, %f756, %f11;
|
||
|
|
fma.rn.f32 %f760, %f19, %f757, %f12;
|
||
|
|
mul.f32 %f761, %f748, 0fBE24A490;
|
||
|
|
mul.f32 %f762, %f751, 0fBE24A490;
|
||
|
|
mul.f32 %f763, %f754, 0fBE24A490;
|
||
|
|
fma.rn.f32 %f112, %f19, %f761, %f758;
|
||
|
|
fma.rn.f32 %f113, %f19, %f762, %f759;
|
||
|
|
fma.rn.f32 %f114, %f19, %f763, %f760;
|
||
|
|
sub.f32 %f764, %f112, %f10;
|
||
|
|
sub.f32 %f765, %f113, %f11;
|
||
|
|
sub.f32 %f766, %f114, %f12;
|
||
|
|
mul.f32 %f767, %f765, %f765;
|
||
|
|
fma.rn.f32 %f768, %f764, %f764, %f767;
|
||
|
|
fma.rn.f32 %f769, %f766, %f766, %f768;
|
||
|
|
sqrt.rn.f32 %f725, %f769;
|
||
|
|
rcp.rn.f32 %f770, %f725;
|
||
|
|
mul.f32 %f721, %f770, %f764;
|
||
|
|
mul.f32 %f722, %f770, %f765;
|
||
|
|
mul.f32 %f723, %f770, %f766;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r308, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r308, %f10, %f11, %f12, %f721, %f722, %f723, %r252, %f374, %f725, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f771, [%rd16];
|
||
|
|
setp.neu.f32 %p68, %f771, 0f3F800000;
|
||
|
|
@%p68 bra BB0_66;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r315, [root];
|
||
|
|
mov.f32 %f779, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r315, %f112, %f113, %f114, %f73, %f74, %f75, %r252, %f374, %f779, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f780, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f780;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_66:
|
||
|
|
mad.lo.s32 %r322, %r88, 1664525, 1013904223;
|
||
|
|
and.b32 %r323, %r322, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f789, %r323;
|
||
|
|
mul.f32 %f790, %f789, 0f33800000;
|
||
|
|
fma.rn.f32 %f791, %f790, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r89, %r322, 1664525, 1013904223;
|
||
|
|
and.b32 %r324, %r89, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f792, %r324;
|
||
|
|
mul.f32 %f793, %f792, 0f33800000;
|
||
|
|
fma.rn.f32 %f794, %f793, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f795, %f13, %f794;
|
||
|
|
mul.f32 %f796, %f14, %f794;
|
||
|
|
mul.f32 %f797, %f15, %f794;
|
||
|
|
fma.rn.f32 %f798, %f16, %f791, %f795;
|
||
|
|
fma.rn.f32 %f799, %f17, %f791, %f796;
|
||
|
|
fma.rn.f32 %f800, %f18, %f791, %f797;
|
||
|
|
mul.f32 %f801, %f799, %f799;
|
||
|
|
fma.rn.f32 %f802, %f798, %f798, %f801;
|
||
|
|
fma.rn.f32 %f803, %f800, %f800, %f802;
|
||
|
|
sqrt.rn.f32 %f804, %f803;
|
||
|
|
rcp.rn.f32 %f805, %f804;
|
||
|
|
mul.f32 %f806, %f798, %f805;
|
||
|
|
mul.f32 %f807, %f799, %f805;
|
||
|
|
mul.f32 %f808, %f800, %f805;
|
||
|
|
mul.f32 %f809, %f1351, %f808;
|
||
|
|
mul.f32 %f810, %f1352, %f807;
|
||
|
|
sub.f32 %f811, %f809, %f810;
|
||
|
|
mul.f32 %f812, %f1352, %f806;
|
||
|
|
mul.f32 %f813, %f1350, %f808;
|
||
|
|
sub.f32 %f814, %f812, %f813;
|
||
|
|
mul.f32 %f815, %f1350, %f807;
|
||
|
|
mul.f32 %f816, %f1351, %f806;
|
||
|
|
sub.f32 %f817, %f815, %f816;
|
||
|
|
mul.f32 %f818, %f806, 0f3ECACABC;
|
||
|
|
mul.f32 %f819, %f807, 0f3ECACABC;
|
||
|
|
mul.f32 %f820, %f808, 0f3ECACABC;
|
||
|
|
fma.rn.f32 %f821, %f19, %f818, %f10;
|
||
|
|
fma.rn.f32 %f822, %f19, %f819, %f11;
|
||
|
|
fma.rn.f32 %f823, %f19, %f820, %f12;
|
||
|
|
mul.f32 %f824, %f811, 0f3F61E1D2;
|
||
|
|
mul.f32 %f825, %f814, 0f3F61E1D2;
|
||
|
|
mul.f32 %f826, %f817, 0f3F61E1D2;
|
||
|
|
fma.rn.f32 %f119, %f19, %f824, %f821;
|
||
|
|
fma.rn.f32 %f120, %f19, %f825, %f822;
|
||
|
|
fma.rn.f32 %f121, %f19, %f826, %f823;
|
||
|
|
sub.f32 %f827, %f119, %f10;
|
||
|
|
sub.f32 %f828, %f120, %f11;
|
||
|
|
sub.f32 %f829, %f121, %f12;
|
||
|
|
mul.f32 %f830, %f828, %f828;
|
||
|
|
fma.rn.f32 %f831, %f827, %f827, %f830;
|
||
|
|
fma.rn.f32 %f832, %f829, %f829, %f831;
|
||
|
|
sqrt.rn.f32 %f788, %f832;
|
||
|
|
rcp.rn.f32 %f833, %f788;
|
||
|
|
mul.f32 %f784, %f833, %f827;
|
||
|
|
mul.f32 %f785, %f833, %f828;
|
||
|
|
mul.f32 %f786, %f833, %f829;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r319, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r319, %f10, %f11, %f12, %f784, %f785, %f786, %r252, %f374, %f788, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f834, [%rd16];
|
||
|
|
setp.neu.f32 %p69, %f834, 0f3F800000;
|
||
|
|
@%p69 bra BB0_68;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r326, [root];
|
||
|
|
mov.f32 %f842, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r326, %f119, %f120, %f121, %f73, %f74, %f75, %r252, %f374, %f842, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f843, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f843;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_68:
|
||
|
|
mad.lo.s32 %r333, %r89, 1664525, 1013904223;
|
||
|
|
and.b32 %r334, %r333, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f852, %r334;
|
||
|
|
mul.f32 %f853, %f852, 0f33800000;
|
||
|
|
fma.rn.f32 %f854, %f853, 0f40000000, 0fBF800000;
|
||
|
|
mad.lo.s32 %r496, %r333, 1664525, 1013904223;
|
||
|
|
and.b32 %r335, %r496, 16777215;
|
||
|
|
cvt.rn.f32.u32 %f855, %r335;
|
||
|
|
mul.f32 %f856, %f855, 0f33800000;
|
||
|
|
fma.rn.f32 %f857, %f856, 0f40000000, 0fBF800000;
|
||
|
|
mul.f32 %f858, %f13, %f857;
|
||
|
|
mul.f32 %f859, %f14, %f857;
|
||
|
|
mul.f32 %f860, %f15, %f857;
|
||
|
|
fma.rn.f32 %f861, %f16, %f854, %f858;
|
||
|
|
fma.rn.f32 %f862, %f17, %f854, %f859;
|
||
|
|
fma.rn.f32 %f863, %f18, %f854, %f860;
|
||
|
|
mul.f32 %f864, %f862, %f862;
|
||
|
|
fma.rn.f32 %f865, %f861, %f861, %f864;
|
||
|
|
fma.rn.f32 %f866, %f863, %f863, %f865;
|
||
|
|
sqrt.rn.f32 %f867, %f866;
|
||
|
|
rcp.rn.f32 %f868, %f867;
|
||
|
|
mul.f32 %f869, %f861, %f868;
|
||
|
|
mul.f32 %f870, %f862, %f868;
|
||
|
|
mul.f32 %f871, %f863, %f868;
|
||
|
|
mul.f32 %f872, %f1351, %f871;
|
||
|
|
mul.f32 %f873, %f1352, %f870;
|
||
|
|
sub.f32 %f874, %f872, %f873;
|
||
|
|
mul.f32 %f875, %f1352, %f869;
|
||
|
|
mul.f32 %f876, %f1350, %f871;
|
||
|
|
sub.f32 %f877, %f875, %f876;
|
||
|
|
mul.f32 %f878, %f1350, %f870;
|
||
|
|
mul.f32 %f879, %f1351, %f869;
|
||
|
|
sub.f32 %f880, %f878, %f879;
|
||
|
|
mul.f32 %f881, %f869, 0f3F61E1D2;
|
||
|
|
mul.f32 %f882, %f870, 0f3F61E1D2;
|
||
|
|
mul.f32 %f883, %f871, 0f3F61E1D2;
|
||
|
|
fma.rn.f32 %f884, %f19, %f881, %f10;
|
||
|
|
fma.rn.f32 %f885, %f19, %f882, %f11;
|
||
|
|
fma.rn.f32 %f886, %f19, %f883, %f12;
|
||
|
|
mul.f32 %f887, %f874, 0fBF61E1E3;
|
||
|
|
mul.f32 %f888, %f877, 0fBF61E1E3;
|
||
|
|
mul.f32 %f889, %f880, 0fBF61E1E3;
|
||
|
|
fma.rn.f32 %f126, %f19, %f887, %f884;
|
||
|
|
fma.rn.f32 %f127, %f19, %f888, %f885;
|
||
|
|
fma.rn.f32 %f128, %f19, %f889, %f886;
|
||
|
|
sub.f32 %f890, %f126, %f10;
|
||
|
|
sub.f32 %f891, %f127, %f11;
|
||
|
|
sub.f32 %f892, %f128, %f12;
|
||
|
|
mul.f32 %f893, %f891, %f891;
|
||
|
|
fma.rn.f32 %f894, %f890, %f890, %f893;
|
||
|
|
fma.rn.f32 %f895, %f892, %f892, %f894;
|
||
|
|
sqrt.rn.f32 %f851, %f895;
|
||
|
|
rcp.rn.f32 %f896, %f851;
|
||
|
|
mul.f32 %f847, %f896, %f890;
|
||
|
|
mul.f32 %f848, %f896, %f891;
|
||
|
|
mul.f32 %f849, %f896, %f892;
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r330, [root];
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r330, %f10, %f11, %f12, %f847, %f848, %f849, %r252, %f374, %f851, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f897, [%rd16];
|
||
|
|
setp.neu.f32 %p70, %f897, 0f3F800000;
|
||
|
|
@%p70 bra BB0_70;
|
||
|
|
|
||
|
|
st.local.u32 [%rd16], %r259;
|
||
|
|
ld.global.u32 %r337, [root];
|
||
|
|
mov.f32 %f905, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r337, %f126, %f127, %f128, %f73, %f74, %f75, %r252, %f374, %f905, %rd75, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f906, [%rd16];
|
||
|
|
add.f32 %f1356, %f1356, %f906;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
mov.u16 %rs170, 1;
|
||
|
|
|
||
|
|
BB0_70:
|
||
|
|
setp.ne.s16 %p71, %rs170, 0;
|
||
|
|
@%p71 bra BB0_72;
|
||
|
|
|
||
|
|
add.u64 %rd91, %SP, 0;
|
||
|
|
cvta.to.local.u64 %rd92, %rd91;
|
||
|
|
st.local.u32 [%rd92], %r259;
|
||
|
|
ld.global.u32 %r341, [root];
|
||
|
|
mov.f32 %f914, 0f6C4ECB8F;
|
||
|
|
// inline asm
|
||
|
|
call _rt_trace_64, (%r341, %f10, %f11, %f12, %f73, %f74, %f75, %r252, %f374, %f914, %rd91, %r108);
|
||
|
|
// inline asm
|
||
|
|
ld.local.f32 %f915, [%rd92];
|
||
|
|
add.f32 %f1356, %f1356, %f915;
|
||
|
|
add.f32 %f1355, %f1355, 0f3F800000;
|
||
|
|
|
||
|
|
BB0_72:
|
||
|
|
ld.global.u32 %r470, [samples];
|
||
|
|
add.s32 %r473, %r473, 1;
|
||
|
|
setp.lt.s32 %p72, %r473, %r470;
|
||
|
|
@%p72 bra BB0_8;
|
||
|
|
|
||
|
|
BB0_73:
|
||
|
|
add.s32 %r471, %r471, 1;
|
||
|
|
setp.lt.s32 %p73, %r471, %r470;
|
||
|
|
@%p73 bra BB0_6;
|
||
|
|
bra.uni BB0_74;
|
||
|
|
|
||
|
|
BB0_4:
|
||
|
|
mov.f32 %f1356, %f1355;
|
||
|
|
|
||
|
|
BB0_74:
|
||
|
|
mul.f32 %f917, %f1352, %f9;
|
||
|
|
mul.f32 %f918, %f1351, %f8;
|
||
|
|
neg.f32 %f919, %f918;
|
||
|
|
mul.f32 %f920, %f1350, %f7;
|
||
|
|
sub.f32 %f921, %f919, %f920;
|
||
|
|
sub.f32 %f141, %f921, %f917;
|
||
|
|
setp.eq.s32 %p74, %r470, 0;
|
||
|
|
mov.f32 %f1391, 0f3F800000;
|
||
|
|
@%p74 bra BB0_76;
|
||
|
|
|
||
|
|
mov.f32 %f922, 0f3F800000;
|
||
|
|
max.f32 %f923, %f1355, %f922;
|
||
|
|
div.rn.f32 %f1391, %f1356, %f923;
|
||
|
|
|
||
|
|
BB0_76:
|
||
|
|
ld.global.f32 %f924, [directColor];
|
||
|
|
mul.f32 %f925, %f1391, %f924;
|
||
|
|
ld.global.f32 %f926, [directColor+4];
|
||
|
|
mul.f32 %f927, %f1391, %f926;
|
||
|
|
ld.global.f32 %f928, [directColor+8];
|
||
|
|
mul.f32 %f929, %f1391, %f928;
|
||
|
|
cvt.sat.f32.f32 %f930, %f141;
|
||
|
|
mul.f32 %f144, %f925, %f930;
|
||
|
|
mul.f32 %f145, %f927, %f930;
|
||
|
|
mul.f32 %f146, %f929, %f930;
|
||
|
|
mul.f32 %f931, %f141, 0f40800000;
|
||
|
|
cvt.sat.f32.f32 %f932, %f931;
|
||
|
|
mul.f32 %f933, %f925, %f932;
|
||
|
|
mul.f32 %f934, %f927, %f932;
|
||
|
|
mul.f32 %f935, %f929, %f932;
|
||
|
|
mul.f32 %f147, %f933, 0f3E800000;
|
||
|
|
mul.f32 %f148, %f934, 0f3E800000;
|
||
|
|
mul.f32 %f149, %f935, 0f3E800000;
|
||
|
|
ld.global.u32 %r499, [imageEnabled];
|
||
|
|
and.b32 %r345, %r499, 8;
|
||
|
|
setp.eq.s32 %p75, %r345, 0;
|
||
|
|
@%p75 bra BB0_89;
|
||
|
|
|
||
|
|
cvt.u64.u32 %rd95, %r2;
|
||
|
|
cvt.u64.u32 %rd96, %r3;
|
||
|
|
mov.u64 %rd99, image_Mask;
|
||
|
|
cvta.global.u64 %rd94, %rd99;
|
||
|
|
// inline asm
|
||
|
|
call (%rd93), _rt_buffer_get_64, (%rd94, %r107, %r107, %rd95, %rd96, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f938, 0f3E68BA2E;
|
||
|
|
cvt.rzi.f32.f32 %f939, %f938;
|
||
|
|
fma.rn.f32 %f940, %f939, 0fC0000000, 0f3EE8BA2E;
|
||
|
|
abs.f32 %f150, %f940;
|
||
|
|
abs.f32 %f151, %f1391;
|
||
|
|
setp.lt.f32 %p76, %f151, 0f00800000;
|
||
|
|
mul.f32 %f941, %f151, 0f4B800000;
|
||
|
|
selp.f32 %f942, 0fC3170000, 0fC2FE0000, %p76;
|
||
|
|
selp.f32 %f943, %f941, %f151, %p76;
|
||
|
|
mov.b32 %r348, %f943;
|
||
|
|
and.b32 %r349, %r348, 8388607;
|
||
|
|
or.b32 %r350, %r349, 1065353216;
|
||
|
|
mov.b32 %f944, %r350;
|
||
|
|
shr.u32 %r351, %r348, 23;
|
||
|
|
cvt.rn.f32.u32 %f945, %r351;
|
||
|
|
add.f32 %f946, %f942, %f945;
|
||
|
|
setp.gt.f32 %p77, %f944, 0f3FB504F3;
|
||
|
|
mul.f32 %f947, %f944, 0f3F000000;
|
||
|
|
add.f32 %f948, %f946, 0f3F800000;
|
||
|
|
selp.f32 %f949, %f947, %f944, %p77;
|
||
|
|
selp.f32 %f950, %f948, %f946, %p77;
|
||
|
|
add.f32 %f951, %f949, 0fBF800000;
|
||
|
|
add.f32 %f937, %f949, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f936,%f937;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f952, %f951, %f951;
|
||
|
|
mul.f32 %f953, %f936, %f952;
|
||
|
|
mul.f32 %f954, %f953, %f953;
|
||
|
|
mov.f32 %f955, 0f3C4CAF63;
|
||
|
|
mov.f32 %f956, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f957, %f956, %f954, %f955;
|
||
|
|
mov.f32 %f958, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f959, %f957, %f954, %f958;
|
||
|
|
mul.rn.f32 %f960, %f959, %f954;
|
||
|
|
mul.rn.f32 %f961, %f960, %f953;
|
||
|
|
sub.f32 %f962, %f951, %f953;
|
||
|
|
neg.f32 %f963, %f953;
|
||
|
|
add.f32 %f964, %f962, %f962;
|
||
|
|
fma.rn.f32 %f965, %f963, %f951, %f964;
|
||
|
|
mul.rn.f32 %f966, %f936, %f965;
|
||
|
|
add.f32 %f967, %f961, %f953;
|
||
|
|
sub.f32 %f968, %f953, %f967;
|
||
|
|
add.f32 %f969, %f961, %f968;
|
||
|
|
add.f32 %f970, %f966, %f969;
|
||
|
|
add.f32 %f971, %f967, %f970;
|
||
|
|
sub.f32 %f972, %f967, %f971;
|
||
|
|
add.f32 %f973, %f970, %f972;
|
||
|
|
mov.f32 %f974, 0f3F317200;
|
||
|
|
mul.rn.f32 %f975, %f950, %f974;
|
||
|
|
mov.f32 %f976, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f977, %f950, %f976;
|
||
|
|
add.f32 %f978, %f975, %f971;
|
||
|
|
sub.f32 %f979, %f975, %f978;
|
||
|
|
add.f32 %f980, %f971, %f979;
|
||
|
|
add.f32 %f981, %f973, %f980;
|
||
|
|
add.f32 %f982, %f977, %f981;
|
||
|
|
add.f32 %f983, %f978, %f982;
|
||
|
|
sub.f32 %f984, %f978, %f983;
|
||
|
|
add.f32 %f985, %f982, %f984;
|
||
|
|
mov.f32 %f986, 0f3EE8BA2E;
|
||
|
|
mul.rn.f32 %f987, %f986, %f983;
|
||
|
|
neg.f32 %f988, %f987;
|
||
|
|
fma.rn.f32 %f989, %f986, %f983, %f988;
|
||
|
|
fma.rn.f32 %f990, %f986, %f985, %f989;
|
||
|
|
mov.f32 %f991, 0f00000000;
|
||
|
|
fma.rn.f32 %f992, %f991, %f983, %f990;
|
||
|
|
add.rn.f32 %f993, %f987, %f992;
|
||
|
|
neg.f32 %f994, %f993;
|
||
|
|
add.rn.f32 %f995, %f987, %f994;
|
||
|
|
add.rn.f32 %f996, %f995, %f992;
|
||
|
|
mov.b32 %r352, %f993;
|
||
|
|
setp.eq.s32 %p78, %r352, 1118925336;
|
||
|
|
add.s32 %r353, %r352, -1;
|
||
|
|
mov.b32 %f997, %r353;
|
||
|
|
add.f32 %f998, %f996, 0f37000000;
|
||
|
|
selp.f32 %f999, %f997, %f993, %p78;
|
||
|
|
selp.f32 %f152, %f998, %f996, %p78;
|
||
|
|
mul.f32 %f1000, %f999, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f1001, %f1000;
|
||
|
|
mov.f32 %f1002, 0fBF317200;
|
||
|
|
fma.rn.f32 %f1003, %f1001, %f1002, %f999;
|
||
|
|
mov.f32 %f1004, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f1005, %f1001, %f1004, %f1003;
|
||
|
|
mul.f32 %f1006, %f1005, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f1007, %f1006;
|
||
|
|
add.f32 %f1008, %f1001, 0f00000000;
|
||
|
|
ex2.approx.f32 %f1009, %f1008;
|
||
|
|
mul.f32 %f1010, %f1007, %f1009;
|
||
|
|
setp.lt.f32 %p79, %f999, 0fC2D20000;
|
||
|
|
selp.f32 %f1011, 0f00000000, %f1010, %p79;
|
||
|
|
setp.gt.f32 %p80, %f999, 0f42D20000;
|
||
|
|
selp.f32 %f1392, 0f7F800000, %f1011, %p80;
|
||
|
|
setp.eq.f32 %p81, %f1392, 0f7F800000;
|
||
|
|
@%p81 bra BB0_79;
|
||
|
|
|
||
|
|
fma.rn.f32 %f1392, %f1392, %f152, %f1392;
|
||
|
|
|
||
|
|
BB0_79:
|
||
|
|
setp.lt.f32 %p82, %f1391, 0f00000000;
|
||
|
|
setp.eq.f32 %p83, %f150, 0f3F800000;
|
||
|
|
and.pred %p1, %p82, %p83;
|
||
|
|
mov.b32 %r354, %f1392;
|
||
|
|
xor.b32 %r355, %r354, -2147483648;
|
||
|
|
mov.b32 %f1012, %r355;
|
||
|
|
selp.f32 %f1394, %f1012, %f1392, %p1;
|
||
|
|
setp.eq.f32 %p84, %f1391, 0f00000000;
|
||
|
|
@%p84 bra BB0_82;
|
||
|
|
bra.uni BB0_80;
|
||
|
|
|
||
|
|
BB0_82:
|
||
|
|
add.f32 %f1015, %f1391, %f1391;
|
||
|
|
selp.f32 %f1394, %f1015, 0f00000000, %p83;
|
||
|
|
bra.uni BB0_83;
|
||
|
|
|
||
|
|
BB0_147:
|
||
|
|
mov.u64 %rd240, image_HDR;
|
||
|
|
cvta.global.u64 %rd235, %rd240;
|
||
|
|
mov.u32 %r441, 8;
|
||
|
|
mov.u64 %rd239, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd234), _rt_buffer_get_64, (%rd235, %r107, %r441, %rd20, %rd21, %rd239, %rd239);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1319, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs115, %f1319;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs116, 0;
|
||
|
|
st.v4.u16 [%rd234], {%rs115, %rs115, %rs115, %rs116};
|
||
|
|
|
||
|
|
BB0_148:
|
||
|
|
ld.global.u32 %r442, [additive];
|
||
|
|
setp.eq.s32 %p151, %r442, 0;
|
||
|
|
@%p151 bra BB0_150;
|
||
|
|
|
||
|
|
mov.u64 %rd253, image_RNM0;
|
||
|
|
cvta.global.u64 %rd242, %rd253;
|
||
|
|
mov.u32 %r446, 8;
|
||
|
|
mov.u64 %rd252, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd241), _rt_buffer_get_64, (%rd242, %r107, %r446, %rd20, %rd21, %rd252, %rd252);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs123, %rs124, %rs125, %rs126}, [%rd241];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1320, %rs123;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1321, %rs124;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1322, %rs125;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd247), _rt_buffer_get_64, (%rd242, %r107, %r446, %rd20, %rd21, %rd252, %rd252);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1323, %f1320, 0f00000000;
|
||
|
|
add.f32 %f1324, %f1321, 0f00000000;
|
||
|
|
add.f32 %f1325, %f1322, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs122, %f1325;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs121, %f1324;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs120, %f1323;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs127, 0;
|
||
|
|
st.v4.u16 [%rd247], {%rs120, %rs121, %rs122, %rs127};
|
||
|
|
bra.uni BB0_151;
|
||
|
|
|
||
|
|
BB0_150:
|
||
|
|
mov.u64 %rd260, image_RNM0;
|
||
|
|
cvta.global.u64 %rd255, %rd260;
|
||
|
|
mov.u32 %r448, 8;
|
||
|
|
mov.u64 %rd259, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd254), _rt_buffer_get_64, (%rd255, %r107, %r448, %rd20, %rd21, %rd259, %rd259);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1326, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs128, %f1326;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs129, 0;
|
||
|
|
st.v4.u16 [%rd254], {%rs128, %rs128, %rs128, %rs129};
|
||
|
|
|
||
|
|
BB0_151:
|
||
|
|
ld.global.u32 %r449, [additive];
|
||
|
|
setp.eq.s32 %p152, %r449, 0;
|
||
|
|
@%p152 bra BB0_153;
|
||
|
|
|
||
|
|
mov.u64 %rd273, image_RNM1;
|
||
|
|
cvta.global.u64 %rd262, %rd273;
|
||
|
|
mov.u32 %r453, 8;
|
||
|
|
mov.u64 %rd272, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd261), _rt_buffer_get_64, (%rd262, %r107, %r453, %rd20, %rd21, %rd272, %rd272);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd261];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1327, %rs136;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1328, %rs137;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1329, %rs138;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd267), _rt_buffer_get_64, (%rd262, %r107, %r453, %rd20, %rd21, %rd272, %rd272);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1330, %f1327, 0f00000000;
|
||
|
|
add.f32 %f1331, %f1328, 0f00000000;
|
||
|
|
add.f32 %f1332, %f1329, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs135, %f1332;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs134, %f1331;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs133, %f1330;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs140, 0;
|
||
|
|
st.v4.u16 [%rd267], {%rs133, %rs134, %rs135, %rs140};
|
||
|
|
bra.uni BB0_154;
|
||
|
|
|
||
|
|
BB0_153:
|
||
|
|
mov.u64 %rd280, image_RNM1;
|
||
|
|
cvta.global.u64 %rd275, %rd280;
|
||
|
|
mov.u32 %r455, 8;
|
||
|
|
mov.u64 %rd279, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd274), _rt_buffer_get_64, (%rd275, %r107, %r455, %rd20, %rd21, %rd279, %rd279);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1333, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs141, %f1333;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs142, 0;
|
||
|
|
st.v4.u16 [%rd274], {%rs141, %rs141, %rs141, %rs142};
|
||
|
|
|
||
|
|
BB0_154:
|
||
|
|
ld.global.u32 %r456, [additive];
|
||
|
|
setp.eq.s32 %p153, %r456, 0;
|
||
|
|
@%p153 bra BB0_156;
|
||
|
|
|
||
|
|
mov.u64 %rd293, image_RNM2;
|
||
|
|
cvta.global.u64 %rd282, %rd293;
|
||
|
|
mov.u32 %r460, 8;
|
||
|
|
mov.u64 %rd292, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r107, %r460, %rd20, %rd21, %rd292, %rd292);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs149, %rs150, %rs151, %rs152}, [%rd281];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1334, %rs149;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1335, %rs150;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1336, %rs151;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd287), _rt_buffer_get_64, (%rd282, %r107, %r460, %rd20, %rd21, %rd292, %rd292);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1337, %f1334, 0f00000000;
|
||
|
|
add.f32 %f1338, %f1335, 0f00000000;
|
||
|
|
add.f32 %f1339, %f1336, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs148, %f1339;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs147, %f1338;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs146, %f1337;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs153, 0;
|
||
|
|
st.v4.u16 [%rd287], {%rs146, %rs147, %rs148, %rs153};
|
||
|
|
bra.uni BB0_157;
|
||
|
|
|
||
|
|
BB0_156:
|
||
|
|
mov.u64 %rd300, image_RNM2;
|
||
|
|
cvta.global.u64 %rd295, %rd300;
|
||
|
|
mov.u32 %r462, 8;
|
||
|
|
mov.u64 %rd299, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd294), _rt_buffer_get_64, (%rd295, %r107, %r462, %rd20, %rd21, %rd299, %rd299);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1340, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs154, %f1340;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs155, 0;
|
||
|
|
st.v4.u16 [%rd294], {%rs154, %rs154, %rs154, %rs155};
|
||
|
|
|
||
|
|
BB0_157:
|
||
|
|
ld.global.u32 %r463, [additive];
|
||
|
|
setp.eq.s32 %p154, %r463, 0;
|
||
|
|
@%p154 bra BB0_159;
|
||
|
|
|
||
|
|
mov.u64 %rd313, image_RNM3;
|
||
|
|
cvta.global.u64 %rd302, %rd313;
|
||
|
|
mov.u32 %r467, 8;
|
||
|
|
mov.u64 %rd312, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd301), _rt_buffer_get_64, (%rd302, %r107, %r467, %rd20, %rd21, %rd312, %rd312);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs162, %rs163, %rs164, %rs165}, [%rd301];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1341, %rs162;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1342, %rs163;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1343, %rs164;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd307), _rt_buffer_get_64, (%rd302, %r107, %r467, %rd20, %rd21, %rd312, %rd312);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1344, %f1341, 0f00000000;
|
||
|
|
add.f32 %f1345, %f1342, 0f00000000;
|
||
|
|
add.f32 %f1346, %f1343, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs161, %f1346;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs160, %f1345;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs159, %f1344;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs166, 0;
|
||
|
|
st.v4.u16 [%rd307], {%rs159, %rs160, %rs161, %rs166};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_159:
|
||
|
|
mov.u64 %rd320, image_RNM3;
|
||
|
|
cvta.global.u64 %rd315, %rd320;
|
||
|
|
mov.u32 %r469, 8;
|
||
|
|
mov.u64 %rd319, 0;
|
||
|
|
// inline asm
|
||
|
|
call (%rd314), _rt_buffer_get_64, (%rd315, %r107, %r469, %rd20, %rd21, %rd319, %rd319);
|
||
|
|
// inline asm
|
||
|
|
mov.f32 %f1347, 0f00000000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs167, %f1347;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
mov.u16 %rs168, 0;
|
||
|
|
st.v4.u16 [%rd314], {%rs167, %rs167, %rs167, %rs168};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_80:
|
||
|
|
setp.geu.f32 %p85, %f1391, 0f00000000;
|
||
|
|
@%p85 bra BB0_83;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f1014, %f986;
|
||
|
|
setp.neu.f32 %p86, %f1014, 0f3EE8BA2E;
|
||
|
|
selp.f32 %f1394, 0f7FFFFFFF, %f1394, %p86;
|
||
|
|
|
||
|
|
BB0_83:
|
||
|
|
add.f32 %f1016, %f151, 0f3EE8BA2E;
|
||
|
|
mov.b32 %r356, %f1016;
|
||
|
|
setp.lt.s32 %p88, %r356, 2139095040;
|
||
|
|
@%p88 bra BB0_88;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p89, %f151, 0f7F800000;
|
||
|
|
@%p89 bra BB0_87;
|
||
|
|
bra.uni BB0_85;
|
||
|
|
|
||
|
|
BB0_87:
|
||
|
|
add.f32 %f1394, %f1391, 0f3EE8BA2E;
|
||
|
|
bra.uni BB0_88;
|
||
|
|
|
||
|
|
BB0_85:
|
||
|
|
setp.neu.f32 %p90, %f151, 0f7F800000;
|
||
|
|
@%p90 bra BB0_88;
|
||
|
|
|
||
|
|
selp.f32 %f1394, 0fFF800000, 0f7F800000, %p1;
|
||
|
|
|
||
|
|
BB0_88:
|
||
|
|
mul.f32 %f1017, %f1394, 0f437F0000;
|
||
|
|
setp.eq.f32 %p91, %f1391, 0f3F800000;
|
||
|
|
selp.f32 %f1018, 0f437F0000, %f1017, %p91;
|
||
|
|
cvt.rzi.u32.f32 %r357, %f1018;
|
||
|
|
cvt.u16.u32 %rs31, %r357;
|
||
|
|
mov.u16 %rs32, 255;
|
||
|
|
st.v2.u8 [%rd93], {%rs31, %rs32};
|
||
|
|
ld.global.u32 %r499, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_89:
|
||
|
|
and.b32 %r358, %r499, 1;
|
||
|
|
setp.eq.b32 %p92, %r358, 1;
|
||
|
|
@!%p92 bra BB0_124;
|
||
|
|
bra.uni BB0_90;
|
||
|
|
|
||
|
|
BB0_90:
|
||
|
|
mov.f32 %f1021, 0f3E666666;
|
||
|
|
cvt.rzi.f32.f32 %f1022, %f1021;
|
||
|
|
fma.rn.f32 %f1023, %f1022, 0fC0000000, 0f3EE66666;
|
||
|
|
abs.f32 %f163, %f1023;
|
||
|
|
abs.f32 %f164, %f144;
|
||
|
|
setp.lt.f32 %p93, %f164, 0f00800000;
|
||
|
|
mul.f32 %f1024, %f164, 0f4B800000;
|
||
|
|
selp.f32 %f1025, 0fC3170000, 0fC2FE0000, %p93;
|
||
|
|
selp.f32 %f1026, %f1024, %f164, %p93;
|
||
|
|
mov.b32 %r359, %f1026;
|
||
|
|
and.b32 %r360, %r359, 8388607;
|
||
|
|
or.b32 %r361, %r360, 1065353216;
|
||
|
|
mov.b32 %f1027, %r361;
|
||
|
|
shr.u32 %r362, %r359, 23;
|
||
|
|
cvt.rn.f32.u32 %f1028, %r362;
|
||
|
|
add.f32 %f1029, %f1025, %f1028;
|
||
|
|
setp.gt.f32 %p94, %f1027, 0f3FB504F3;
|
||
|
|
mul.f32 %f1030, %f1027, 0f3F000000;
|
||
|
|
add.f32 %f1031, %f1029, 0f3F800000;
|
||
|
|
selp.f32 %f1032, %f1030, %f1027, %p94;
|
||
|
|
selp.f32 %f1033, %f1031, %f1029, %p94;
|
||
|
|
add.f32 %f1034, %f1032, 0fBF800000;
|
||
|
|
add.f32 %f1020, %f1032, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f1019,%f1020;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1035, %f1034, %f1034;
|
||
|
|
mul.f32 %f1036, %f1019, %f1035;
|
||
|
|
mul.f32 %f1037, %f1036, %f1036;
|
||
|
|
mov.f32 %f1038, 0f3C4CAF63;
|
||
|
|
mov.f32 %f1039, 0f3B18F0FE;
|
||
|
|
fma.rn.f32 %f1040, %f1039, %f1037, %f1038;
|
||
|
|
mov.f32 %f1041, 0f3DAAAABD;
|
||
|
|
fma.rn.f32 %f1042, %f1040, %f1037, %f1041;
|
||
|
|
mul.rn.f32 %f1043, %f1042, %f1037;
|
||
|
|
mul.rn.f32 %f1044, %f1043, %f1036;
|
||
|
|
sub.f32 %f1045, %f1034, %f1036;
|
||
|
|
neg.f32 %f1046, %f1036;
|
||
|
|
add.f32 %f1047, %f1045, %f1045;
|
||
|
|
fma.rn.f32 %f1048, %f1046, %f1034, %f1047;
|
||
|
|
mul.rn.f32 %f1049, %f1019, %f1048;
|
||
|
|
add.f32 %f1050, %f1044, %f1036;
|
||
|
|
sub.f32 %f1051, %f1036, %f1050;
|
||
|
|
add.f32 %f1052, %f1044, %f1051;
|
||
|
|
add.f32 %f1053, %f1049, %f1052;
|
||
|
|
add.f32 %f1054, %f1050, %f1053;
|
||
|
|
sub.f32 %f1055, %f1050, %f1054;
|
||
|
|
add.f32 %f1056, %f1053, %f1055;
|
||
|
|
mov.f32 %f1057, 0f3F317200;
|
||
|
|
mul.rn.f32 %f1058, %f1033, %f1057;
|
||
|
|
mov.f32 %f1059, 0f35BFBE8E;
|
||
|
|
mul.rn.f32 %f1060, %f1033, %f1059;
|
||
|
|
add.f32 %f1061, %f1058, %f1054;
|
||
|
|
sub.f32 %f1062, %f1058, %f1061;
|
||
|
|
add.f32 %f1063, %f1054, %f1062;
|
||
|
|
add.f32 %f1064, %f1056, %f1063;
|
||
|
|
add.f32 %f1065, %f1060, %f1064;
|
||
|
|
add.f32 %f1066, %f1061, %f1065;
|
||
|
|
sub.f32 %f1067, %f1061, %f1066;
|
||
|
|
add.f32 %f1068, %f1065, %f1067;
|
||
|
|
mov.f32 %f1069, 0f3EE66666;
|
||
|
|
mul.rn.f32 %f1070, %f1069, %f1066;
|
||
|
|
neg.f32 %f1071, %f1070;
|
||
|
|
fma.rn.f32 %f1072, %f1069, %f1066, %f1071;
|
||
|
|
fma.rn.f32 %f1073, %f1069, %f1068, %f1072;
|
||
|
|
mov.f32 %f1074, 0f00000000;
|
||
|
|
fma.rn.f32 %f1075, %f1074, %f1066, %f1073;
|
||
|
|
add.rn.f32 %f1076, %f1070, %f1075;
|
||
|
|
neg.f32 %f1077, %f1076;
|
||
|
|
add.rn.f32 %f1078, %f1070, %f1077;
|
||
|
|
add.rn.f32 %f1079, %f1078, %f1075;
|
||
|
|
mov.b32 %r363, %f1076;
|
||
|
|
setp.eq.s32 %p95, %r363, 1118925336;
|
||
|
|
add.s32 %r364, %r363, -1;
|
||
|
|
mov.b32 %f1080, %r364;
|
||
|
|
add.f32 %f1081, %f1079, 0f37000000;
|
||
|
|
selp.f32 %f1082, %f1080, %f1076, %p95;
|
||
|
|
selp.f32 %f165, %f1081, %f1079, %p95;
|
||
|
|
mul.f32 %f1083, %f1082, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f1084, %f1083;
|
||
|
|
mov.f32 %f1085, 0fBF317200;
|
||
|
|
fma.rn.f32 %f1086, %f1084, %f1085, %f1082;
|
||
|
|
mov.f32 %f1087, 0fB5BFBE8E;
|
||
|
|
fma.rn.f32 %f1088, %f1084, %f1087, %f1086;
|
||
|
|
mul.f32 %f1089, %f1088, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f1090, %f1089;
|
||
|
|
add.f32 %f1091, %f1084, 0f00000000;
|
||
|
|
ex2.approx.f32 %f1092, %f1091;
|
||
|
|
mul.f32 %f1093, %f1090, %f1092;
|
||
|
|
setp.lt.f32 %p96, %f1082, 0fC2D20000;
|
||
|
|
selp.f32 %f1094, 0f00000000, %f1093, %p96;
|
||
|
|
setp.gt.f32 %p97, %f1082, 0f42D20000;
|
||
|
|
selp.f32 %f1395, 0f7F800000, %f1094, %p97;
|
||
|
|
setp.eq.f32 %p98, %f1395, 0f7F800000;
|
||
|
|
@%p98 bra BB0_92;
|
||
|
|
|
||
|
|
fma.rn.f32 %f1395, %f1395, %f165, %f1395;
|
||
|
|
|
||
|
|
BB0_92:
|
||
|
|
setp.lt.f32 %p99, %f144, 0f00000000;
|
||
|
|
setp.eq.f32 %p100, %f163, 0f3F800000;
|
||
|
|
and.pred %p2, %p99, %p100;
|
||
|
|
mov.b32 %r365, %f1395;
|
||
|
|
xor.b32 %r366, %r365, -2147483648;
|
||
|
|
mov.b32 %f1095, %r366;
|
||
|
|
selp.f32 %f1397, %f1095, %f1395, %p2;
|
||
|
|
setp.eq.f32 %p101, %f144, 0f00000000;
|
||
|
|
@%p101 bra BB0_95;
|
||
|
|
bra.uni BB0_93;
|
||
|
|
|
||
|
|
BB0_95:
|
||
|
|
add.f32 %f1098, %f144, %f144;
|
||
|
|
selp.f32 %f1397, %f1098, 0f00000000, %p100;
|
||
|
|
bra.uni BB0_96;
|
||
|
|
|
||
|
|
BB0_93:
|
||
|
|
setp.geu.f32 %p102, %f144, 0f00000000;
|
||
|
|
@%p102 bra BB0_96;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f1097, %f1069;
|
||
|
|
setp.neu.f32 %p103, %f1097, 0f3EE66666;
|
||
|
|
selp.f32 %f1397, 0f7FFFFFFF, %f1397, %p103;
|
||
|
|
|
||
|
|
BB0_96:
|
||
|
|
add.f32 %f1099, %f164, 0f3EE66666;
|
||
|
|
mov.b32 %r367, %f1099;
|
||
|
|
setp.lt.s32 %p105, %r367, 2139095040;
|
||
|
|
@%p105 bra BB0_101;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p106, %f164, 0f7F800000;
|
||
|
|
@%p106 bra BB0_100;
|
||
|
|
bra.uni BB0_98;
|
||
|
|
|
||
|
|
BB0_100:
|
||
|
|
add.f32 %f1397, %f144, 0f3EE66666;
|
||
|
|
bra.uni BB0_101;
|
||
|
|
|
||
|
|
BB0_98:
|
||
|
|
setp.neu.f32 %p107, %f164, 0f7F800000;
|
||
|
|
@%p107 bra BB0_101;
|
||
|
|
|
||
|
|
selp.f32 %f1397, 0fFF800000, 0f7F800000, %p2;
|
||
|
|
|
||
|
|
BB0_101:
|
||
|
|
setp.eq.f32 %p108, %f144, 0f3F800000;
|
||
|
|
selp.f32 %f176, 0f3F800000, %f1397, %p108;
|
||
|
|
abs.f32 %f177, %f145;
|
||
|
|
setp.lt.f32 %p109, %f177, 0f00800000;
|
||
|
|
mul.f32 %f1102, %f177, 0f4B800000;
|
||
|
|
selp.f32 %f1103, 0fC3170000, 0fC2FE0000, %p109;
|
||
|
|
selp.f32 %f1104, %f1102, %f177, %p109;
|
||
|
|
mov.b32 %r368, %f1104;
|
||
|
|
and.b32 %r369, %r368, 8388607;
|
||
|
|
or.b32 %r370, %r369, 1065353216;
|
||
|
|
mov.b32 %f1105, %r370;
|
||
|
|
shr.u32 %r371, %r368, 23;
|
||
|
|
cvt.rn.f32.u32 %f1106, %r371;
|
||
|
|
add.f32 %f1107, %f1103, %f1106;
|
||
|
|
setp.gt.f32 %p110, %f1105, 0f3FB504F3;
|
||
|
|
mul.f32 %f1108, %f1105, 0f3F000000;
|
||
|
|
add.f32 %f1109, %f1107, 0f3F800000;
|
||
|
|
selp.f32 %f1110, %f1108, %f1105, %p110;
|
||
|
|
selp.f32 %f1111, %f1109, %f1107, %p110;
|
||
|
|
add.f32 %f1112, %f1110, 0fBF800000;
|
||
|
|
add.f32 %f1101, %f1110, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f1100,%f1101;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1113, %f1112, %f1112;
|
||
|
|
mul.f32 %f1114, %f1100, %f1113;
|
||
|
|
mul.f32 %f1115, %f1114, %f1114;
|
||
|
|
fma.rn.f32 %f1118, %f1039, %f1115, %f1038;
|
||
|
|
fma.rn.f32 %f1120, %f1118, %f1115, %f1041;
|
||
|
|
mul.rn.f32 %f1121, %f1120, %f1115;
|
||
|
|
mul.rn.f32 %f1122, %f1121, %f1114;
|
||
|
|
sub.f32 %f1123, %f1112, %f1114;
|
||
|
|
neg.f32 %f1124, %f1114;
|
||
|
|
add.f32 %f1125, %f1123, %f1123;
|
||
|
|
fma.rn.f32 %f1126, %f1124, %f1112, %f1125;
|
||
|
|
mul.rn.f32 %f1127, %f1100, %f1126;
|
||
|
|
add.f32 %f1128, %f1122, %f1114;
|
||
|
|
sub.f32 %f1129, %f1114, %f1128;
|
||
|
|
add.f32 %f1130, %f1122, %f1129;
|
||
|
|
add.f32 %f1131, %f1127, %f1130;
|
||
|
|
add.f32 %f1132, %f1128, %f1131;
|
||
|
|
sub.f32 %f1133, %f1128, %f1132;
|
||
|
|
add.f32 %f1134, %f1131, %f1133;
|
||
|
|
mul.rn.f32 %f1136, %f1111, %f1057;
|
||
|
|
mul.rn.f32 %f1138, %f1111, %f1059;
|
||
|
|
add.f32 %f1139, %f1136, %f1132;
|
||
|
|
sub.f32 %f1140, %f1136, %f1139;
|
||
|
|
add.f32 %f1141, %f1132, %f1140;
|
||
|
|
add.f32 %f1142, %f1134, %f1141;
|
||
|
|
add.f32 %f1143, %f1138, %f1142;
|
||
|
|
add.f32 %f1144, %f1139, %f1143;
|
||
|
|
sub.f32 %f1145, %f1139, %f1144;
|
||
|
|
add.f32 %f1146, %f1143, %f1145;
|
||
|
|
mul.rn.f32 %f1148, %f1069, %f1144;
|
||
|
|
neg.f32 %f1149, %f1148;
|
||
|
|
fma.rn.f32 %f1150, %f1069, %f1144, %f1149;
|
||
|
|
fma.rn.f32 %f1151, %f1069, %f1146, %f1150;
|
||
|
|
fma.rn.f32 %f1153, %f1074, %f1144, %f1151;
|
||
|
|
add.rn.f32 %f1154, %f1148, %f1153;
|
||
|
|
neg.f32 %f1155, %f1154;
|
||
|
|
add.rn.f32 %f1156, %f1148, %f1155;
|
||
|
|
add.rn.f32 %f1157, %f1156, %f1153;
|
||
|
|
mov.b32 %r372, %f1154;
|
||
|
|
setp.eq.s32 %p111, %r372, 1118925336;
|
||
|
|
add.s32 %r373, %r372, -1;
|
||
|
|
mov.b32 %f1158, %r373;
|
||
|
|
add.f32 %f1159, %f1157, 0f37000000;
|
||
|
|
selp.f32 %f1160, %f1158, %f1154, %p111;
|
||
|
|
selp.f32 %f178, %f1159, %f1157, %p111;
|
||
|
|
mul.f32 %f1161, %f1160, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f1162, %f1161;
|
||
|
|
fma.rn.f32 %f1164, %f1162, %f1085, %f1160;
|
||
|
|
fma.rn.f32 %f1166, %f1162, %f1087, %f1164;
|
||
|
|
mul.f32 %f1167, %f1166, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f1168, %f1167;
|
||
|
|
add.f32 %f1169, %f1162, 0f00000000;
|
||
|
|
ex2.approx.f32 %f1170, %f1169;
|
||
|
|
mul.f32 %f1171, %f1168, %f1170;
|
||
|
|
setp.lt.f32 %p112, %f1160, 0fC2D20000;
|
||
|
|
selp.f32 %f1172, 0f00000000, %f1171, %p112;
|
||
|
|
setp.gt.f32 %p113, %f1160, 0f42D20000;
|
||
|
|
selp.f32 %f1398, 0f7F800000, %f1172, %p113;
|
||
|
|
setp.eq.f32 %p114, %f1398, 0f7F800000;
|
||
|
|
@%p114 bra BB0_103;
|
||
|
|
|
||
|
|
fma.rn.f32 %f1398, %f1398, %f178, %f1398;
|
||
|
|
|
||
|
|
BB0_103:
|
||
|
|
setp.lt.f32 %p115, %f145, 0f00000000;
|
||
|
|
and.pred %p3, %p115, %p100;
|
||
|
|
mov.b32 %r374, %f1398;
|
||
|
|
xor.b32 %r375, %r374, -2147483648;
|
||
|
|
mov.b32 %f1173, %r375;
|
||
|
|
selp.f32 %f1400, %f1173, %f1398, %p3;
|
||
|
|
setp.eq.f32 %p117, %f145, 0f00000000;
|
||
|
|
@%p117 bra BB0_106;
|
||
|
|
bra.uni BB0_104;
|
||
|
|
|
||
|
|
BB0_106:
|
||
|
|
add.f32 %f1176, %f145, %f145;
|
||
|
|
selp.f32 %f1400, %f1176, 0f00000000, %p100;
|
||
|
|
bra.uni BB0_107;
|
||
|
|
|
||
|
|
BB0_104:
|
||
|
|
setp.geu.f32 %p118, %f145, 0f00000000;
|
||
|
|
@%p118 bra BB0_107;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f1175, %f1069;
|
||
|
|
setp.neu.f32 %p119, %f1175, 0f3EE66666;
|
||
|
|
selp.f32 %f1400, 0f7FFFFFFF, %f1400, %p119;
|
||
|
|
|
||
|
|
BB0_107:
|
||
|
|
add.f32 %f1177, %f177, 0f3EE66666;
|
||
|
|
mov.b32 %r376, %f1177;
|
||
|
|
setp.lt.s32 %p121, %r376, 2139095040;
|
||
|
|
@%p121 bra BB0_112;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p122, %f177, 0f7F800000;
|
||
|
|
@%p122 bra BB0_111;
|
||
|
|
bra.uni BB0_109;
|
||
|
|
|
||
|
|
BB0_111:
|
||
|
|
add.f32 %f1400, %f145, 0f3EE66666;
|
||
|
|
bra.uni BB0_112;
|
||
|
|
|
||
|
|
BB0_109:
|
||
|
|
setp.neu.f32 %p123, %f177, 0f7F800000;
|
||
|
|
@%p123 bra BB0_112;
|
||
|
|
|
||
|
|
selp.f32 %f1400, 0fFF800000, 0f7F800000, %p3;
|
||
|
|
|
||
|
|
BB0_112:
|
||
|
|
setp.eq.f32 %p124, %f145, 0f3F800000;
|
||
|
|
selp.f32 %f189, 0f3F800000, %f1400, %p124;
|
||
|
|
abs.f32 %f190, %f146;
|
||
|
|
setp.lt.f32 %p125, %f190, 0f00800000;
|
||
|
|
mul.f32 %f1180, %f190, 0f4B800000;
|
||
|
|
selp.f32 %f1181, 0fC3170000, 0fC2FE0000, %p125;
|
||
|
|
selp.f32 %f1182, %f1180, %f190, %p125;
|
||
|
|
mov.b32 %r377, %f1182;
|
||
|
|
and.b32 %r378, %r377, 8388607;
|
||
|
|
or.b32 %r379, %r378, 1065353216;
|
||
|
|
mov.b32 %f1183, %r379;
|
||
|
|
shr.u32 %r380, %r377, 23;
|
||
|
|
cvt.rn.f32.u32 %f1184, %r380;
|
||
|
|
add.f32 %f1185, %f1181, %f1184;
|
||
|
|
setp.gt.f32 %p126, %f1183, 0f3FB504F3;
|
||
|
|
mul.f32 %f1186, %f1183, 0f3F000000;
|
||
|
|
add.f32 %f1187, %f1185, 0f3F800000;
|
||
|
|
selp.f32 %f1188, %f1186, %f1183, %p126;
|
||
|
|
selp.f32 %f1189, %f1187, %f1185, %p126;
|
||
|
|
add.f32 %f1190, %f1188, 0fBF800000;
|
||
|
|
add.f32 %f1179, %f1188, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
rcp.approx.ftz.f32 %f1178,%f1179;
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1191, %f1190, %f1190;
|
||
|
|
mul.f32 %f1192, %f1178, %f1191;
|
||
|
|
mul.f32 %f1193, %f1192, %f1192;
|
||
|
|
fma.rn.f32 %f1196, %f1039, %f1193, %f1038;
|
||
|
|
fma.rn.f32 %f1198, %f1196, %f1193, %f1041;
|
||
|
|
mul.rn.f32 %f1199, %f1198, %f1193;
|
||
|
|
mul.rn.f32 %f1200, %f1199, %f1192;
|
||
|
|
sub.f32 %f1201, %f1190, %f1192;
|
||
|
|
neg.f32 %f1202, %f1192;
|
||
|
|
add.f32 %f1203, %f1201, %f1201;
|
||
|
|
fma.rn.f32 %f1204, %f1202, %f1190, %f1203;
|
||
|
|
mul.rn.f32 %f1205, %f1178, %f1204;
|
||
|
|
add.f32 %f1206, %f1200, %f1192;
|
||
|
|
sub.f32 %f1207, %f1192, %f1206;
|
||
|
|
add.f32 %f1208, %f1200, %f1207;
|
||
|
|
add.f32 %f1209, %f1205, %f1208;
|
||
|
|
add.f32 %f1210, %f1206, %f1209;
|
||
|
|
sub.f32 %f1211, %f1206, %f1210;
|
||
|
|
add.f32 %f1212, %f1209, %f1211;
|
||
|
|
mul.rn.f32 %f1214, %f1189, %f1057;
|
||
|
|
mul.rn.f32 %f1216, %f1189, %f1059;
|
||
|
|
add.f32 %f1217, %f1214, %f1210;
|
||
|
|
sub.f32 %f1218, %f1214, %f1217;
|
||
|
|
add.f32 %f1219, %f1210, %f1218;
|
||
|
|
add.f32 %f1220, %f1212, %f1219;
|
||
|
|
add.f32 %f1221, %f1216, %f1220;
|
||
|
|
add.f32 %f1222, %f1217, %f1221;
|
||
|
|
sub.f32 %f1223, %f1217, %f1222;
|
||
|
|
add.f32 %f1224, %f1221, %f1223;
|
||
|
|
mul.rn.f32 %f1226, %f1069, %f1222;
|
||
|
|
neg.f32 %f1227, %f1226;
|
||
|
|
fma.rn.f32 %f1228, %f1069, %f1222, %f1227;
|
||
|
|
fma.rn.f32 %f1229, %f1069, %f1224, %f1228;
|
||
|
|
fma.rn.f32 %f1231, %f1074, %f1222, %f1229;
|
||
|
|
add.rn.f32 %f1232, %f1226, %f1231;
|
||
|
|
neg.f32 %f1233, %f1232;
|
||
|
|
add.rn.f32 %f1234, %f1226, %f1233;
|
||
|
|
add.rn.f32 %f1235, %f1234, %f1231;
|
||
|
|
mov.b32 %r381, %f1232;
|
||
|
|
setp.eq.s32 %p127, %r381, 1118925336;
|
||
|
|
add.s32 %r382, %r381, -1;
|
||
|
|
mov.b32 %f1236, %r382;
|
||
|
|
add.f32 %f1237, %f1235, 0f37000000;
|
||
|
|
selp.f32 %f1238, %f1236, %f1232, %p127;
|
||
|
|
selp.f32 %f191, %f1237, %f1235, %p127;
|
||
|
|
mul.f32 %f1239, %f1238, 0f3FB8AA3B;
|
||
|
|
cvt.rzi.f32.f32 %f1240, %f1239;
|
||
|
|
fma.rn.f32 %f1242, %f1240, %f1085, %f1238;
|
||
|
|
fma.rn.f32 %f1244, %f1240, %f1087, %f1242;
|
||
|
|
mul.f32 %f1245, %f1244, 0f3FB8AA3B;
|
||
|
|
ex2.approx.ftz.f32 %f1246, %f1245;
|
||
|
|
add.f32 %f1247, %f1240, 0f00000000;
|
||
|
|
ex2.approx.f32 %f1248, %f1247;
|
||
|
|
mul.f32 %f1249, %f1246, %f1248;
|
||
|
|
setp.lt.f32 %p128, %f1238, 0fC2D20000;
|
||
|
|
selp.f32 %f1250, 0f00000000, %f1249, %p128;
|
||
|
|
setp.gt.f32 %p129, %f1238, 0f42D20000;
|
||
|
|
selp.f32 %f1401, 0f7F800000, %f1250, %p129;
|
||
|
|
setp.eq.f32 %p130, %f1401, 0f7F800000;
|
||
|
|
@%p130 bra BB0_114;
|
||
|
|
|
||
|
|
fma.rn.f32 %f1401, %f1401, %f191, %f1401;
|
||
|
|
|
||
|
|
BB0_114:
|
||
|
|
setp.lt.f32 %p131, %f146, 0f00000000;
|
||
|
|
and.pred %p4, %p131, %p100;
|
||
|
|
mov.b32 %r383, %f1401;
|
||
|
|
xor.b32 %r384, %r383, -2147483648;
|
||
|
|
mov.b32 %f1251, %r384;
|
||
|
|
selp.f32 %f1403, %f1251, %f1401, %p4;
|
||
|
|
setp.eq.f32 %p133, %f146, 0f00000000;
|
||
|
|
@%p133 bra BB0_117;
|
||
|
|
bra.uni BB0_115;
|
||
|
|
|
||
|
|
BB0_117:
|
||
|
|
add.f32 %f1254, %f146, %f146;
|
||
|
|
selp.f32 %f1403, %f1254, 0f00000000, %p100;
|
||
|
|
bra.uni BB0_118;
|
||
|
|
|
||
|
|
BB0_115:
|
||
|
|
setp.geu.f32 %p134, %f146, 0f00000000;
|
||
|
|
@%p134 bra BB0_118;
|
||
|
|
|
||
|
|
cvt.rzi.f32.f32 %f1253, %f1069;
|
||
|
|
setp.neu.f32 %p135, %f1253, 0f3EE66666;
|
||
|
|
selp.f32 %f1403, 0f7FFFFFFF, %f1403, %p135;
|
||
|
|
|
||
|
|
BB0_118:
|
||
|
|
add.f32 %f1255, %f190, 0f3EE66666;
|
||
|
|
mov.b32 %r385, %f1255;
|
||
|
|
setp.lt.s32 %p137, %r385, 2139095040;
|
||
|
|
@%p137 bra BB0_123;
|
||
|
|
|
||
|
|
setp.gtu.f32 %p138, %f190, 0f7F800000;
|
||
|
|
@%p138 bra BB0_122;
|
||
|
|
bra.uni BB0_120;
|
||
|
|
|
||
|
|
BB0_122:
|
||
|
|
add.f32 %f1403, %f146, 0f3EE66666;
|
||
|
|
bra.uni BB0_123;
|
||
|
|
|
||
|
|
BB0_120:
|
||
|
|
setp.neu.f32 %p139, %f190, 0f7F800000;
|
||
|
|
@%p139 bra BB0_123;
|
||
|
|
|
||
|
|
selp.f32 %f1403, 0fFF800000, 0f7F800000, %p4;
|
||
|
|
|
||
|
|
BB0_123:
|
||
|
|
setp.eq.f32 %p140, %f146, 0f3F800000;
|
||
|
|
selp.f32 %f1256, 0f3F800000, %f1403, %p140;
|
||
|
|
cvt.u64.u32 %rd103, %r3;
|
||
|
|
cvt.u64.u32 %rd102, %r2;
|
||
|
|
mov.u64 %rd106, image;
|
||
|
|
cvta.global.u64 %rd101, %rd106;
|
||
|
|
// inline asm
|
||
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r107, %r108, %rd102, %rd103, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
cvt.sat.f32.f32 %f1257, %f1256;
|
||
|
|
mul.f32 %f1258, %f1257, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r388, %f1258;
|
||
|
|
cvt.sat.f32.f32 %f1259, %f189;
|
||
|
|
mul.f32 %f1260, %f1259, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r389, %f1260;
|
||
|
|
cvt.sat.f32.f32 %f1261, %f176;
|
||
|
|
mul.f32 %f1262, %f1261, 0f437FFD71;
|
||
|
|
cvt.rzi.u32.f32 %r390, %f1262;
|
||
|
|
cvt.u16.u32 %rs33, %r388;
|
||
|
|
cvt.u16.u32 %rs34, %r390;
|
||
|
|
cvt.u16.u32 %rs35, %r389;
|
||
|
|
mov.u16 %rs36, 255;
|
||
|
|
st.v4.u8 [%rd100], {%rs33, %rs35, %rs34, %rs36};
|
||
|
|
ld.global.u32 %r499, [imageEnabled];
|
||
|
|
|
||
|
|
BB0_124:
|
||
|
|
cvt.u64.u32 %rd18, %r2;
|
||
|
|
cvt.u64.u32 %rd19, %r3;
|
||
|
|
and.b32 %r391, %r499, 4;
|
||
|
|
setp.eq.s32 %p141, %r391, 0;
|
||
|
|
@%p141 bra BB0_128;
|
||
|
|
|
||
|
|
ld.global.u32 %r392, [additive];
|
||
|
|
setp.eq.s32 %p142, %r392, 0;
|
||
|
|
mov.f32 %f1263, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs37, %f1263;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p142 bra BB0_127;
|
||
|
|
|
||
|
|
mov.u64 %rd119, image_HDR;
|
||
|
|
cvta.global.u64 %rd108, %rd119;
|
||
|
|
mov.u32 %r396, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r107, %r396, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs44, %rs45, %rs46, %rs47}, [%rd107];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1264, %rs44;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1265, %rs45;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1266, %rs46;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd113), _rt_buffer_get_64, (%rd108, %r107, %r396, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1267, %f144, %f1264;
|
||
|
|
add.f32 %f1268, %f145, %f1265;
|
||
|
|
add.f32 %f1269, %f146, %f1266;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs43, %f1269;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs42, %f1268;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs41, %f1267;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd113], {%rs41, %rs42, %rs43, %rs37};
|
||
|
|
bra.uni BB0_128;
|
||
|
|
|
||
|
|
BB0_127:
|
||
|
|
mov.u64 %rd126, image_HDR;
|
||
|
|
cvta.global.u64 %rd121, %rd126;
|
||
|
|
mov.u32 %r398, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r107, %r398, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs50, %f146;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs49, %f145;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs48, %f144;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd120], {%rs48, %rs49, %rs50, %rs37};
|
||
|
|
|
||
|
|
BB0_128:
|
||
|
|
ld.global.u32 %r399, [additive];
|
||
|
|
setp.eq.s32 %p143, %r399, 0;
|
||
|
|
mov.f32 %f1273, 0f3F800000;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs51, %f1273;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p143 bra BB0_130;
|
||
|
|
|
||
|
|
mov.u64 %rd139, image_RNM0;
|
||
|
|
cvta.global.u64 %rd128, %rd139;
|
||
|
|
mov.u32 %r403, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd127), _rt_buffer_get_64, (%rd128, %r107, %r403, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs58, %rs59, %rs60, %rs61}, [%rd127];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1274, %rs58;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1275, %rs59;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1276, %rs60;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd133), _rt_buffer_get_64, (%rd128, %r107, %r403, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1277, %f147, %f1274;
|
||
|
|
add.f32 %f1278, %f148, %f1275;
|
||
|
|
add.f32 %f1279, %f149, %f1276;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs57, %f1279;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs56, %f1278;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs55, %f1277;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd133], {%rs55, %rs56, %rs57, %rs51};
|
||
|
|
bra.uni BB0_131;
|
||
|
|
|
||
|
|
BB0_130:
|
||
|
|
mov.u64 %rd146, image_RNM0;
|
||
|
|
cvta.global.u64 %rd141, %rd146;
|
||
|
|
mov.u32 %r405, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r107, %r405, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs64, %f149;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs63, %f148;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs62, %f147;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd140], {%rs62, %rs63, %rs64, %rs51};
|
||
|
|
|
||
|
|
BB0_131:
|
||
|
|
ld.global.f32 %f1284, [directDir];
|
||
|
|
fma.rn.f32 %f202, %f1284, 0fBF000000, 0f3F000000;
|
||
|
|
ld.global.f32 %f1285, [directDir+4];
|
||
|
|
fma.rn.f32 %f203, %f1285, 0fBF000000, 0f3F000000;
|
||
|
|
ld.global.f32 %f1286, [directDir+8];
|
||
|
|
fma.rn.f32 %f204, %f1286, 0fBF000000, 0f3F000000;
|
||
|
|
ld.global.u32 %r406, [additive];
|
||
|
|
setp.eq.s32 %p144, %r406, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs65, %f1273;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p144 bra BB0_133;
|
||
|
|
|
||
|
|
mov.u64 %rd159, image_RNM1;
|
||
|
|
cvta.global.u64 %rd148, %rd159;
|
||
|
|
mov.u32 %r410, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r107, %r410, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd147];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1287, %rs72;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1288, %rs73;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1289, %rs74;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd153), _rt_buffer_get_64, (%rd148, %r107, %r410, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1290, %f202, %f1287;
|
||
|
|
add.f32 %f1291, %f202, %f1288;
|
||
|
|
add.f32 %f1292, %f202, %f1289;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs71, %f1292;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs70, %f1291;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs69, %f1290;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd153], {%rs69, %rs70, %rs71, %rs65};
|
||
|
|
bra.uni BB0_134;
|
||
|
|
|
||
|
|
BB0_133:
|
||
|
|
mov.u64 %rd166, image_RNM1;
|
||
|
|
cvta.global.u64 %rd161, %rd166;
|
||
|
|
mov.u32 %r412, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd160), _rt_buffer_get_64, (%rd161, %r107, %r412, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs76, %f202;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd160], {%rs76, %rs76, %rs76, %rs65};
|
||
|
|
|
||
|
|
BB0_134:
|
||
|
|
ld.global.u32 %r413, [additive];
|
||
|
|
setp.eq.s32 %p145, %r413, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs77, %f1273;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p145 bra BB0_136;
|
||
|
|
|
||
|
|
mov.u64 %rd179, image_RNM2;
|
||
|
|
cvta.global.u64 %rd168, %rd179;
|
||
|
|
mov.u32 %r417, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r107, %r417, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd167];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1295, %rs84;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1296, %rs85;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1297, %rs86;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd173), _rt_buffer_get_64, (%rd168, %r107, %r417, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1298, %f203, %f1295;
|
||
|
|
add.f32 %f1299, %f203, %f1296;
|
||
|
|
add.f32 %f1300, %f203, %f1297;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs83, %f1300;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs82, %f1299;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs81, %f1298;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd173], {%rs81, %rs82, %rs83, %rs77};
|
||
|
|
bra.uni BB0_137;
|
||
|
|
|
||
|
|
BB0_136:
|
||
|
|
mov.u64 %rd186, image_RNM2;
|
||
|
|
cvta.global.u64 %rd181, %rd186;
|
||
|
|
mov.u32 %r419, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd180), _rt_buffer_get_64, (%rd181, %r107, %r419, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs88, %f203;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd180], {%rs88, %rs88, %rs88, %rs77};
|
||
|
|
|
||
|
|
BB0_137:
|
||
|
|
ld.global.u32 %r420, [additive];
|
||
|
|
setp.eq.s32 %p146, %r420, 0;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs89, %f1273;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
@%p146 bra BB0_139;
|
||
|
|
|
||
|
|
mov.u64 %rd199, image_RNM3;
|
||
|
|
cvta.global.u64 %rd188, %rd199;
|
||
|
|
mov.u32 %r424, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r107, %r424, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
ld.v4.u16 {%rs96, %rs97, %rs98, %rs99}, [%rd187];
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1303, %rs96;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1304, %rs97;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.f32.f16 %f1305, %rs98;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
call (%rd193), _rt_buffer_get_64, (%rd188, %r107, %r424, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
add.f32 %f1306, %f204, %f1303;
|
||
|
|
add.f32 %f1307, %f204, %f1304;
|
||
|
|
add.f32 %f1308, %f204, %f1305;
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs95, %f1308;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs94, %f1307;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs93, %f1306;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd193], {%rs93, %rs94, %rs95, %rs89};
|
||
|
|
bra.uni BB0_160;
|
||
|
|
|
||
|
|
BB0_139:
|
||
|
|
mov.u64 %rd206, image_RNM3;
|
||
|
|
cvta.global.u64 %rd201, %rd206;
|
||
|
|
mov.u32 %r426, 8;
|
||
|
|
// inline asm
|
||
|
|
call (%rd200), _rt_buffer_get_64, (%rd201, %r107, %r426, %rd18, %rd19, %rd27, %rd27);
|
||
|
|
// inline asm
|
||
|
|
// inline asm
|
||
|
|
{ cvt.rn.f16.f32 %rs100, %f204;}
|
||
|
|
|
||
|
|
// inline asm
|
||
|
|
st.v4.u16 [%rd200], {%rs100, %rs100, %rs100, %rs89};
|
||
|
|
|
||
|
|
BB0_160:
|
||
|
|
ret;
|
||
|
|
}
|
||
|
|
|
||
|
|
|