2397 lines
70 KiB
Plaintext
2397 lines
70 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 uvtangent[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightRadius;
|
|
.global .align 4 .b8 lightPos[12];
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 4 .u32 ignoreNormal;
|
|
.global .align 4 .u32 lightCookie;
|
|
.global .align 4 .b8 lightMatrix[36];
|
|
.global .align 4 .f32 lightFOV;
|
|
.global .align 4 .f32 lightFalloffFakeDistanceMult;
|
|
.global .align 4 .f32 lightFalloffMinRadiusSq;
|
|
.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_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightRadiusE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8lightPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightCookieE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightMatrixE[8] = {82, 97, 121, 0, 36, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8lightFOVE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo28lightFalloffFakeDistanceMultE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo23lightFalloffMinRadiusSqE[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 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename11lightRadiusE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8lightPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename11lightCookieE[4] = {105, 110, 116, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename11lightMatrixE[10] = {77, 97, 116, 114, 105, 120, 51, 120, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8lightFOVE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename28lightFalloffFakeDistanceMultE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename23lightFalloffMinRadiusSqE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightRadiusE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8lightPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightCookieE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightMatrixE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8lightFOVE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum28lightFalloffFakeDistanceMultE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum23lightFalloffMinRadiusSqE = 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_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11lightRadiusE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8lightPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11lightCookieE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11lightMatrixE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8lightFOVE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic28lightFalloffFakeDistanceMultE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic23lightFalloffMinRadiusSqE[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_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11lightRadiusE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8lightPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11lightCookieE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11lightMatrixE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8lightFOVE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation28lightFalloffFakeDistanceMultE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation23lightFalloffMinRadiusSqE[1];
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[4];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<129>;
|
|
.reg .b16 %rs<213>;
|
|
.reg .f32 %f<1090>;
|
|
.reg .b32 %r<250>;
|
|
.reg .b64 %rd<334>;
|
|
|
|
|
|
mov.u64 %rd333, __local_depot0;
|
|
cvta.local.u64 %SP, %rd333;
|
|
ld.global.v2.u32 {%r33, %r34}, [pixelID];
|
|
cvt.u64.u32 %rd16, %r33;
|
|
cvt.u64.u32 %rd17, %r34;
|
|
mov.u64 %rd20, uvnormal;
|
|
cvta.global.u64 %rd15, %rd20;
|
|
mov.u32 %r31, 2;
|
|
mov.u32 %r32, 4;
|
|
mov.u64 %rd19, 0;
|
|
// inline asm
|
|
call (%rd14), _rt_buffer_get_64, (%rd15, %r31, %r32, %rd16, %rd17, %rd19, %rd19);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd14];
|
|
shr.u32 %r37, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r37;
|
|
and.b16 %rs10, %rs1, 255;
|
|
cvt.u16.u32 %rs11, %r1;
|
|
or.b16 %rs12, %rs11, %rs10;
|
|
setp.eq.s16 %p6, %rs12, 0;
|
|
mov.f32 %f1060, 0f00000000;
|
|
mov.f32 %f1061, %f1060;
|
|
mov.f32 %f1062, %f1060;
|
|
@%p6 bra BB0_2;
|
|
|
|
ld.u8 %rs13, [%rd14+1];
|
|
and.b16 %rs15, %rs11, 255;
|
|
cvt.rn.f32.u16 %f139, %rs15;
|
|
div.rn.f32 %f140, %f139, 0f437F0000;
|
|
fma.rn.f32 %f141, %f140, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f142, %rs13;
|
|
div.rn.f32 %f143, %f142, 0f437F0000;
|
|
fma.rn.f32 %f144, %f143, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f145, %rs10;
|
|
div.rn.f32 %f146, %f145, 0f437F0000;
|
|
fma.rn.f32 %f147, %f146, 0f40000000, 0fBF800000;
|
|
mul.f32 %f148, %f144, %f144;
|
|
fma.rn.f32 %f149, %f141, %f141, %f148;
|
|
fma.rn.f32 %f150, %f147, %f147, %f149;
|
|
sqrt.rn.f32 %f151, %f150;
|
|
rcp.rn.f32 %f152, %f151;
|
|
mul.f32 %f1060, %f141, %f152;
|
|
mul.f32 %f1061, %f144, %f152;
|
|
mul.f32 %f1062, %f147, %f152;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r38, %r39}, [pixelID];
|
|
ld.global.v2.u32 {%r41, %r42}, [tileInfo];
|
|
add.s32 %r2, %r38, %r41;
|
|
add.s32 %r3, %r39, %r42;
|
|
setp.eq.f32 %p7, %f1061, 0f00000000;
|
|
setp.eq.f32 %p8, %f1060, 0f00000000;
|
|
and.pred %p9, %p8, %p7;
|
|
setp.eq.f32 %p10, %f1062, 0f00000000;
|
|
and.pred %p11, %p9, %p10;
|
|
@%p11 bra BB0_104;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_104:
|
|
ld.global.u32 %r249, [imageEnabled];
|
|
and.b32 %r202, %r249, 1;
|
|
setp.eq.b32 %p122, %r202, 1;
|
|
@!%p122 bra BB0_106;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_105:
|
|
cvt.u64.u32 %rd241, %r2;
|
|
cvt.u64.u32 %rd242, %r3;
|
|
mov.u64 %rd245, image;
|
|
cvta.global.u64 %rd240, %rd245;
|
|
// inline asm
|
|
call (%rd239), _rt_buffer_get_64, (%rd240, %r31, %r32, %rd241, %rd242, %rd19, %rd19);
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u8 [%rd239], {%rs158, %rs158, %rs158, %rs158};
|
|
ld.global.u32 %r249, [imageEnabled];
|
|
|
|
BB0_106:
|
|
and.b32 %r205, %r249, 8;
|
|
setp.eq.s32 %p123, %r205, 0;
|
|
@%p123 bra BB0_108;
|
|
|
|
cvt.u64.u32 %rd249, %r3;
|
|
cvt.u64.u32 %rd248, %r2;
|
|
mov.u64 %rd252, image_Mask;
|
|
cvta.global.u64 %rd247, %rd252;
|
|
// inline asm
|
|
call (%rd246), _rt_buffer_get_64, (%rd247, %r31, %r31, %rd248, %rd249, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f963, 0f00000000;
|
|
cvt.rzi.u32.f32 %r208, %f963;
|
|
cvt.u16.u32 %rs159, %r208;
|
|
mov.u16 %rs160, 0;
|
|
st.v2.u8 [%rd246], {%rs159, %rs160};
|
|
ld.global.u32 %r249, [imageEnabled];
|
|
|
|
BB0_108:
|
|
cvt.u64.u32 %rd12, %r2;
|
|
cvt.u64.u32 %rd13, %r3;
|
|
and.b32 %r209, %r249, 4;
|
|
setp.eq.s32 %p124, %r209, 0;
|
|
@%p124 bra BB0_112;
|
|
|
|
ld.global.u32 %r210, [additive];
|
|
setp.eq.s32 %p125, %r210, 0;
|
|
@%p125 bra BB0_111;
|
|
|
|
mov.u64 %rd265, image_HDR;
|
|
cvta.global.u64 %rd254, %rd265;
|
|
mov.u32 %r214, 8;
|
|
// inline asm
|
|
call (%rd253), _rt_buffer_get_64, (%rd254, %r31, %r214, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs167, %rs168, %rs169, %rs170}, [%rd253];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f964, %rs167;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f965, %rs168;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f966, %rs169;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd259), _rt_buffer_get_64, (%rd254, %r31, %r214, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f967, %f964, 0f00000000;
|
|
add.f32 %f968, %f965, 0f00000000;
|
|
add.f32 %f969, %f966, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs166, %f969;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f968;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f967;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs171, 0;
|
|
st.v4.u16 [%rd259], {%rs164, %rs165, %rs166, %rs171};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r48, %r49}, [pixelID];
|
|
cvt.u64.u32 %rd23, %r48;
|
|
cvt.u64.u32 %rd24, %r49;
|
|
mov.u64 %rd27, uvpos;
|
|
cvta.global.u64 %rd22, %rd27;
|
|
mov.u32 %r47, 12;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd22, %r31, %r47, %rd23, %rd24, %rd19, %rd19);
|
|
// inline asm
|
|
ld.global.f32 %f155, [lightPos];
|
|
ld.f32 %f9, [%rd21+8];
|
|
ld.f32 %f8, [%rd21+4];
|
|
ld.f32 %f7, [%rd21];
|
|
sub.f32 %f156, %f155, %f7;
|
|
ld.global.f32 %f157, [lightPos+4];
|
|
sub.f32 %f158, %f157, %f8;
|
|
ld.global.f32 %f159, [lightPos+8];
|
|
sub.f32 %f160, %f159, %f9;
|
|
mul.f32 %f161, %f158, %f158;
|
|
fma.rn.f32 %f162, %f156, %f156, %f161;
|
|
fma.rn.f32 %f163, %f160, %f160, %f162;
|
|
sqrt.rn.f32 %f164, %f163;
|
|
rcp.rn.f32 %f165, %f164;
|
|
mul.f32 %f10, %f156, %f165;
|
|
mul.f32 %f11, %f158, %f165;
|
|
mul.f32 %f12, %f160, %f165;
|
|
ld.global.f32 %f166, [lightFalloffFakeDistanceMult];
|
|
mul.f32 %f13, %f164, %f166;
|
|
ld.global.f32 %f167, [lightInvCutoff];
|
|
mul.f32 %f14, %f164, %f167;
|
|
mov.f32 %f171, 0f40800000;
|
|
abs.f32 %f16, %f14;
|
|
setp.lt.f32 %p12, %f16, 0f00800000;
|
|
mul.f32 %f173, %f16, 0f4B800000;
|
|
selp.f32 %f174, 0fC3170000, 0fC2FE0000, %p12;
|
|
selp.f32 %f175, %f173, %f16, %p12;
|
|
mov.b32 %r52, %f175;
|
|
and.b32 %r53, %r52, 8388607;
|
|
or.b32 %r54, %r53, 1065353216;
|
|
mov.b32 %f176, %r54;
|
|
shr.u32 %r55, %r52, 23;
|
|
cvt.rn.f32.u32 %f177, %r55;
|
|
add.f32 %f178, %f174, %f177;
|
|
setp.gt.f32 %p13, %f176, 0f3FB504F3;
|
|
mul.f32 %f179, %f176, 0f3F000000;
|
|
add.f32 %f180, %f178, 0f3F800000;
|
|
selp.f32 %f181, %f179, %f176, %p13;
|
|
selp.f32 %f182, %f180, %f178, %p13;
|
|
add.f32 %f183, %f181, 0fBF800000;
|
|
add.f32 %f154, %f181, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f153,%f154;
|
|
// inline asm
|
|
add.f32 %f184, %f183, %f183;
|
|
mul.f32 %f185, %f153, %f184;
|
|
mul.f32 %f186, %f185, %f185;
|
|
mov.f32 %f187, 0f3C4CAF63;
|
|
mov.f32 %f188, 0f3B18F0FE;
|
|
fma.rn.f32 %f189, %f188, %f186, %f187;
|
|
mov.f32 %f190, 0f3DAAAABD;
|
|
fma.rn.f32 %f191, %f189, %f186, %f190;
|
|
mul.rn.f32 %f192, %f191, %f186;
|
|
mul.rn.f32 %f193, %f192, %f185;
|
|
sub.f32 %f194, %f183, %f185;
|
|
neg.f32 %f195, %f185;
|
|
add.f32 %f196, %f194, %f194;
|
|
fma.rn.f32 %f197, %f195, %f183, %f196;
|
|
mul.rn.f32 %f198, %f153, %f197;
|
|
add.f32 %f199, %f193, %f185;
|
|
sub.f32 %f200, %f185, %f199;
|
|
add.f32 %f201, %f193, %f200;
|
|
add.f32 %f202, %f198, %f201;
|
|
add.f32 %f203, %f199, %f202;
|
|
sub.f32 %f204, %f199, %f203;
|
|
add.f32 %f205, %f202, %f204;
|
|
mov.f32 %f206, 0f3F317200;
|
|
mul.rn.f32 %f207, %f182, %f206;
|
|
mov.f32 %f208, 0f35BFBE8E;
|
|
mul.rn.f32 %f209, %f182, %f208;
|
|
add.f32 %f210, %f207, %f203;
|
|
sub.f32 %f211, %f207, %f210;
|
|
add.f32 %f212, %f203, %f211;
|
|
add.f32 %f213, %f205, %f212;
|
|
add.f32 %f214, %f209, %f213;
|
|
add.f32 %f215, %f210, %f214;
|
|
sub.f32 %f216, %f210, %f215;
|
|
add.f32 %f217, %f214, %f216;
|
|
mul.rn.f32 %f218, %f171, %f215;
|
|
neg.f32 %f219, %f218;
|
|
fma.rn.f32 %f220, %f171, %f215, %f219;
|
|
fma.rn.f32 %f221, %f171, %f217, %f220;
|
|
mov.f32 %f222, 0f00000000;
|
|
fma.rn.f32 %f223, %f222, %f215, %f221;
|
|
add.rn.f32 %f224, %f218, %f223;
|
|
neg.f32 %f225, %f224;
|
|
add.rn.f32 %f226, %f218, %f225;
|
|
add.rn.f32 %f227, %f226, %f223;
|
|
mov.b32 %r56, %f224;
|
|
setp.eq.s32 %p14, %r56, 1118925336;
|
|
add.s32 %r57, %r56, -1;
|
|
mov.b32 %f228, %r57;
|
|
add.f32 %f229, %f227, 0f37000000;
|
|
selp.f32 %f230, %f228, %f224, %p14;
|
|
selp.f32 %f17, %f229, %f227, %p14;
|
|
mul.f32 %f231, %f230, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f232, %f231;
|
|
mov.f32 %f233, 0fBF317200;
|
|
fma.rn.f32 %f234, %f232, %f233, %f230;
|
|
mov.f32 %f235, 0fB5BFBE8E;
|
|
fma.rn.f32 %f236, %f232, %f235, %f234;
|
|
mul.f32 %f237, %f236, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f238, %f237;
|
|
add.f32 %f239, %f232, 0f00000000;
|
|
ex2.approx.f32 %f240, %f239;
|
|
mul.f32 %f241, %f238, %f240;
|
|
setp.lt.f32 %p15, %f230, 0fC2D20000;
|
|
selp.f32 %f242, 0f00000000, %f241, %p15;
|
|
setp.gt.f32 %p16, %f230, 0f42D20000;
|
|
selp.f32 %f1063, 0f7F800000, %f242, %p16;
|
|
setp.eq.f32 %p17, %f1063, 0f7F800000;
|
|
@%p17 bra BB0_5;
|
|
|
|
fma.rn.f32 %f1063, %f1063, %f17, %f1063;
|
|
|
|
BB0_5:
|
|
mov.f32 %f997, 0f40000000;
|
|
cvt.rzi.f32.f32 %f996, %f997;
|
|
add.f32 %f995, %f996, %f996;
|
|
mov.f32 %f994, 0f40800000;
|
|
sub.f32 %f993, %f994, %f995;
|
|
abs.f32 %f992, %f993;
|
|
setp.lt.f32 %p18, %f14, 0f00000000;
|
|
setp.eq.f32 %p19, %f992, 0f3F800000;
|
|
and.pred %p1, %p18, %p19;
|
|
mov.b32 %r58, %f1063;
|
|
xor.b32 %r59, %r58, -2147483648;
|
|
mov.b32 %f243, %r59;
|
|
selp.f32 %f1065, %f243, %f1063, %p1;
|
|
setp.eq.f32 %p20, %f14, 0f00000000;
|
|
@%p20 bra BB0_8;
|
|
bra.uni BB0_6;
|
|
|
|
BB0_8:
|
|
add.f32 %f246, %f14, %f14;
|
|
selp.f32 %f1065, %f246, 0f00000000, %p19;
|
|
bra.uni BB0_9;
|
|
|
|
BB0_6:
|
|
setp.geu.f32 %p21, %f14, 0f00000000;
|
|
@%p21 bra BB0_9;
|
|
|
|
mov.f32 %f1031, 0f40800000;
|
|
cvt.rzi.f32.f32 %f245, %f1031;
|
|
setp.neu.f32 %p22, %f245, 0f40800000;
|
|
selp.f32 %f1065, 0f7FFFFFFF, %f1065, %p22;
|
|
|
|
BB0_9:
|
|
abs.f32 %f998, %f14;
|
|
add.f32 %f247, %f998, 0f40800000;
|
|
mov.b32 %r60, %f247;
|
|
setp.lt.s32 %p24, %r60, 2139095040;
|
|
@%p24 bra BB0_14;
|
|
|
|
abs.f32 %f1029, %f14;
|
|
setp.gtu.f32 %p25, %f1029, 0f7F800000;
|
|
@%p25 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f1065, %f14, 0f40800000;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
abs.f32 %f1030, %f14;
|
|
setp.neu.f32 %p26, %f1030, 0f7F800000;
|
|
@%p26 bra BB0_14;
|
|
|
|
selp.f32 %f1065, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_14:
|
|
mov.f32 %f999, 0f00000000;
|
|
mov.f32 %f1068, 0f3F800000;
|
|
sub.f32 %f257, %f1068, %f1065;
|
|
setp.eq.f32 %p27, %f14, 0f3F800000;
|
|
selp.f32 %f258, 0f00000000, %f257, %p27;
|
|
cvt.sat.f32.f32 %f259, %f258;
|
|
ld.global.f32 %f260, [lightFalloffMinRadiusSq];
|
|
fma.rn.f32 %f261, %f13, %f13, %f260;
|
|
div.rn.f32 %f28, %f259, %f261;
|
|
mul.f32 %f262, %f1061, %f11;
|
|
fma.rn.f32 %f263, %f1060, %f10, %f262;
|
|
fma.rn.f32 %f29, %f1062, %f12, %f263;
|
|
ld.global.u32 %r63, [imageEnabled];
|
|
and.b32 %r64, %r63, 32;
|
|
ld.global.u32 %r65, [ignoreNormal];
|
|
or.b32 %r4, %r64, %r65;
|
|
ld.global.f32 %f264, [lightMatrix+24];
|
|
mul.f32 %f265, %f10, %f264;
|
|
ld.global.f32 %f266, [lightMatrix+28];
|
|
mul.f32 %f267, %f11, %f266;
|
|
neg.f32 %f268, %f267;
|
|
sub.f32 %f269, %f268, %f265;
|
|
ld.global.f32 %f270, [lightMatrix+32];
|
|
mul.f32 %f271, %f12, %f270;
|
|
sub.f32 %f272, %f269, %f271;
|
|
setp.gt.f32 %p28, %f272, 0f00000000;
|
|
ld.global.f32 %f273, [lightMatrix];
|
|
mul.f32 %f274, %f273, %f10;
|
|
sub.f32 %f275, %f999, %f274;
|
|
ld.global.f32 %f276, [lightMatrix+4];
|
|
mul.f32 %f277, %f276, %f11;
|
|
sub.f32 %f278, %f275, %f277;
|
|
ld.global.f32 %f279, [lightMatrix+8];
|
|
mul.f32 %f280, %f279, %f12;
|
|
sub.f32 %f281, %f278, %f280;
|
|
selp.f32 %f282, 0f3F800000, 0f00000000, %p28;
|
|
ld.global.f32 %f283, [lightMatrix+12];
|
|
mul.f32 %f284, %f283, %f10;
|
|
sub.f32 %f285, %f999, %f284;
|
|
ld.global.f32 %f286, [lightMatrix+16];
|
|
mul.f32 %f287, %f286, %f11;
|
|
sub.f32 %f288, %f285, %f287;
|
|
ld.global.f32 %f289, [lightMatrix+20];
|
|
mul.f32 %f290, %f289, %f12;
|
|
sub.f32 %f291, %f288, %f290;
|
|
ld.global.u32 %r61, [lightCookie];
|
|
ld.global.f32 %f292, [lightFOV];
|
|
fma.rn.f32 %f293, %f281, %f292, 0f3F000000;
|
|
sub.f32 %f252, %f1068, %f293;
|
|
fma.rn.f32 %f253, %f291, %f292, 0f3F000000;
|
|
// inline asm
|
|
call (%f248, %f249, %f250, %f251), _rt_texture_get_f_id, (%r61, %r31, %f252, %f253, %f999, %f999);
|
|
// inline asm
|
|
max.f32 %f294, %f248, %f249;
|
|
max.f32 %f295, %f294, %f250;
|
|
mul.f32 %f296, %f282, %f295;
|
|
mul.f32 %f297, %f28, %f296;
|
|
ld.global.f32 %f298, [lightColor+4];
|
|
ld.global.f32 %f299, [lightColor];
|
|
max.f32 %f300, %f299, %f298;
|
|
ld.global.f32 %f301, [lightColor+8];
|
|
max.f32 %f302, %f300, %f301;
|
|
mul.f32 %f303, %f297, %f302;
|
|
setp.lt.f32 %p29, %f303, 0f3727C5AC;
|
|
@%p29 bra BB0_87;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_87:
|
|
ld.global.u32 %r247, [imageEnabled];
|
|
and.b32 %r166, %r247, 1;
|
|
setp.eq.b32 %p115, %r166, 1;
|
|
@!%p115 bra BB0_89;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_88:
|
|
cvt.u64.u32 %rd147, %r2;
|
|
cvt.u64.u32 %rd148, %r3;
|
|
mov.u64 %rd151, image;
|
|
cvta.global.u64 %rd146, %rd151;
|
|
// inline asm
|
|
call (%rd145), _rt_buffer_get_64, (%rd146, %r31, %r32, %rd147, %rd148, %rd19, %rd19);
|
|
// inline asm
|
|
mov.u16 %rs106, 1;
|
|
mov.u16 %rs107, 0;
|
|
st.v4.u8 [%rd145], {%rs107, %rs107, %rs107, %rs106};
|
|
ld.global.u32 %r247, [imageEnabled];
|
|
|
|
BB0_89:
|
|
and.b32 %r169, %r247, 8;
|
|
setp.eq.s32 %p116, %r169, 0;
|
|
@%p116 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd155, %r3;
|
|
cvt.u64.u32 %rd154, %r2;
|
|
mov.u64 %rd158, image_Mask;
|
|
cvta.global.u64 %rd153, %rd158;
|
|
// inline asm
|
|
call (%rd152), _rt_buffer_get_64, (%rd153, %r31, %r31, %rd154, %rd155, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f928, 0f00000000;
|
|
cvt.rzi.u32.f32 %r172, %f928;
|
|
cvt.u16.u32 %rs108, %r172;
|
|
mov.u16 %rs109, 255;
|
|
st.v2.u8 [%rd152], {%rs108, %rs109};
|
|
ld.global.u32 %r247, [imageEnabled];
|
|
|
|
BB0_91:
|
|
cvt.u64.u32 %rd10, %r2;
|
|
cvt.u64.u32 %rd11, %r3;
|
|
and.b32 %r173, %r247, 4;
|
|
setp.eq.s32 %p117, %r173, 0;
|
|
@%p117 bra BB0_95;
|
|
|
|
ld.global.u32 %r174, [additive];
|
|
setp.eq.s32 %p118, %r174, 0;
|
|
mov.f32 %f929, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f929;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_94;
|
|
|
|
mov.u64 %rd171, image_HDR;
|
|
cvta.global.u64 %rd160, %rd171;
|
|
mov.u32 %r178, 8;
|
|
// inline asm
|
|
call (%rd159), _rt_buffer_get_64, (%rd160, %r31, %r178, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs117, %rs118, %rs119, %rs120}, [%rd159];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f930, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f931, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f932, %rs119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd165), _rt_buffer_get_64, (%rd160, %r31, %r178, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f933, %f930, 0f00000000;
|
|
add.f32 %f934, %f931, 0f00000000;
|
|
add.f32 %f935, %f932, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f935;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f934;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f933;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd165], {%rs114, %rs115, %rs116, %rs110};
|
|
bra.uni BB0_95;
|
|
|
|
BB0_15:
|
|
mov.f32 %f1067, 0f00000000;
|
|
mul.f32 %f305, %f7, 0f3456BF95;
|
|
abs.f32 %f306, %f1060;
|
|
div.rn.f32 %f307, %f305, %f306;
|
|
abs.f32 %f308, %f1061;
|
|
mul.f32 %f309, %f8, 0f3456BF95;
|
|
div.rn.f32 %f310, %f309, %f308;
|
|
abs.f32 %f311, %f1062;
|
|
mul.f32 %f312, %f9, 0f3456BF95;
|
|
div.rn.f32 %f313, %f312, %f311;
|
|
abs.f32 %f314, %f307;
|
|
abs.f32 %f315, %f310;
|
|
abs.f32 %f316, %f313;
|
|
mov.f32 %f317, 0f38D1B717;
|
|
max.f32 %f318, %f314, %f317;
|
|
max.f32 %f319, %f315, %f317;
|
|
max.f32 %f320, %f316, %f317;
|
|
fma.rn.f32 %f33, %f1060, %f318, %f7;
|
|
fma.rn.f32 %f34, %f1061, %f319, %f8;
|
|
fma.rn.f32 %f35, %f1062, %f320, %f9;
|
|
ld.global.u32 %r243, [samples];
|
|
setp.lt.s32 %p30, %r243, 1;
|
|
@%p30 bra BB0_18;
|
|
|
|
mul.f32 %f322, %f33, 0f3456BF95;
|
|
abs.f32 %f323, %f322;
|
|
mul.f32 %f324, %f34, 0f3456BF95;
|
|
abs.f32 %f325, %f324;
|
|
mul.f32 %f326, %f35, 0f3456BF95;
|
|
abs.f32 %f327, %f326;
|
|
max.f32 %f328, %f323, %f325;
|
|
max.f32 %f329, %f328, %f327;
|
|
max.f32 %f36, %f329, %f317;
|
|
mov.f32 %f1067, 0f00000000;
|
|
mov.u32 %r242, 0;
|
|
|
|
BB0_17:
|
|
add.u64 %rd28, %SP, 0;
|
|
cvta.to.local.u64 %rd29, %rd28;
|
|
cvt.rn.f32.s32 %f339, %r242;
|
|
mul.f32 %f340, %f339, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f341, %f340;
|
|
sub.f32 %f342, %f340, %f341;
|
|
mul.f32 %f343, %f339, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f344, %f343;
|
|
sub.f32 %f345, %f343, %f344;
|
|
mul.f32 %f346, %f339, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f347, %f346;
|
|
sub.f32 %f348, %f346, %f347;
|
|
add.f32 %f349, %f345, 0f4199851F;
|
|
add.f32 %f350, %f348, 0f4199851F;
|
|
add.f32 %f351, %f342, 0f4199851F;
|
|
mul.f32 %f352, %f345, %f350;
|
|
fma.rn.f32 %f353, %f342, %f349, %f352;
|
|
fma.rn.f32 %f354, %f351, %f348, %f353;
|
|
add.f32 %f355, %f342, %f354;
|
|
add.f32 %f356, %f345, %f354;
|
|
add.f32 %f357, %f348, %f354;
|
|
add.f32 %f358, %f355, %f356;
|
|
mul.f32 %f359, %f357, %f358;
|
|
cvt.rmi.f32.f32 %f360, %f359;
|
|
sub.f32 %f361, %f359, %f360;
|
|
add.f32 %f362, %f355, %f357;
|
|
mul.f32 %f363, %f356, %f362;
|
|
cvt.rmi.f32.f32 %f364, %f363;
|
|
sub.f32 %f365, %f363, %f364;
|
|
add.f32 %f366, %f356, %f357;
|
|
mul.f32 %f367, %f355, %f366;
|
|
cvt.rmi.f32.f32 %f368, %f367;
|
|
sub.f32 %f369, %f367, %f368;
|
|
fma.rn.f32 %f370, %f361, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f371, %f365, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f372, %f369, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f373, [lightRadius];
|
|
ld.global.f32 %f374, [lightPos];
|
|
fma.rn.f32 %f375, %f373, %f370, %f374;
|
|
ld.global.f32 %f376, [lightPos+4];
|
|
fma.rn.f32 %f377, %f373, %f371, %f376;
|
|
ld.global.f32 %f378, [lightPos+8];
|
|
fma.rn.f32 %f379, %f373, %f372, %f378;
|
|
sub.f32 %f380, %f375, %f7;
|
|
sub.f32 %f381, %f377, %f8;
|
|
sub.f32 %f382, %f379, %f9;
|
|
mul.f32 %f383, %f381, %f381;
|
|
fma.rn.f32 %f384, %f380, %f380, %f383;
|
|
fma.rn.f32 %f385, %f382, %f382, %f384;
|
|
sqrt.rn.f32 %f338, %f385;
|
|
rcp.rn.f32 %f386, %f338;
|
|
mul.f32 %f334, %f386, %f380;
|
|
mul.f32 %f335, %f386, %f381;
|
|
mul.f32 %f336, %f386, %f382;
|
|
ld.global.u32 %r70, [imageEnabled];
|
|
and.b32 %r71, %r70, 32;
|
|
setp.eq.s32 %p31, %r71, 0;
|
|
selp.f32 %f387, 0f3F800000, 0f41200000, %p31;
|
|
mul.f32 %f337, %f387, %f36;
|
|
mov.u32 %r72, 1065353216;
|
|
st.local.u32 [%rd29], %r72;
|
|
ld.global.u32 %r67, [root];
|
|
mov.u32 %r68, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r67, %f33, %f34, %f35, %f334, %f335, %f336, %r68, %f337, %f338, %rd28, %r32);
|
|
// inline asm
|
|
ld.local.f32 %f388, [%rd29];
|
|
add.f32 %f1067, %f1067, %f388;
|
|
ld.global.u32 %r243, [samples];
|
|
add.s32 %r242, %r242, 1;
|
|
setp.lt.s32 %p32, %r242, %r243;
|
|
@%p32 bra BB0_17;
|
|
|
|
BB0_18:
|
|
setp.eq.s32 %p33, %r243, 0;
|
|
@%p33 bra BB0_20;
|
|
|
|
cvt.rn.f32.s32 %f390, %r243;
|
|
div.rn.f32 %f1068, %f1067, %f390;
|
|
|
|
BB0_20:
|
|
setp.eq.s32 %p34, %r4, 0;
|
|
selp.f32 %f391, %f29, 0f3F800000, %p34;
|
|
cvt.sat.f32.f32 %f392, %f391;
|
|
mul.f32 %f393, %f28, %f392;
|
|
mul.f32 %f394, %f1068, %f393;
|
|
ld.global.f32 %f395, [lightColor];
|
|
mul.f32 %f396, %f395, %f394;
|
|
ld.global.f32 %f397, [lightColor+4];
|
|
mul.f32 %f398, %f397, %f394;
|
|
ld.global.f32 %f399, [lightColor+8];
|
|
mul.f32 %f400, %f394, %f399;
|
|
mul.f32 %f42, %f248, %f396;
|
|
mul.f32 %f43, %f249, %f398;
|
|
mul.f32 %f44, %f250, %f400;
|
|
ld.global.u32 %r245, [imageEnabled];
|
|
and.b32 %r73, %r245, 8;
|
|
setp.eq.s32 %p35, %r73, 0;
|
|
@%p35 bra BB0_33;
|
|
|
|
mov.f32 %f1007, 0fB5BFBE8E;
|
|
mov.f32 %f1006, 0fBF317200;
|
|
mov.f32 %f1005, 0f35BFBE8E;
|
|
mov.f32 %f1004, 0f3F317200;
|
|
mov.f32 %f1003, 0f3DAAAABD;
|
|
mov.f32 %f1002, 0f3C4CAF63;
|
|
mov.f32 %f1001, 0f3B18F0FE;
|
|
cvt.u64.u32 %rd32, %r2;
|
|
cvt.u64.u32 %rd33, %r3;
|
|
mov.u64 %rd36, image_Mask;
|
|
cvta.global.u64 %rd31, %rd36;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r31, %r31, %rd32, %rd33, %rd19, %rd19);
|
|
// inline asm
|
|
abs.f32 %f46, %f1068;
|
|
setp.lt.f32 %p36, %f46, 0f00800000;
|
|
mul.f32 %f406, %f46, 0f4B800000;
|
|
selp.f32 %f407, 0fC3170000, 0fC2FE0000, %p36;
|
|
selp.f32 %f408, %f406, %f46, %p36;
|
|
mov.b32 %r76, %f408;
|
|
and.b32 %r77, %r76, 8388607;
|
|
or.b32 %r78, %r77, 1065353216;
|
|
mov.b32 %f409, %r78;
|
|
shr.u32 %r79, %r76, 23;
|
|
cvt.rn.f32.u32 %f410, %r79;
|
|
add.f32 %f411, %f407, %f410;
|
|
setp.gt.f32 %p37, %f409, 0f3FB504F3;
|
|
mul.f32 %f412, %f409, 0f3F000000;
|
|
add.f32 %f413, %f411, 0f3F800000;
|
|
selp.f32 %f414, %f412, %f409, %p37;
|
|
selp.f32 %f415, %f413, %f411, %p37;
|
|
add.f32 %f416, %f414, 0fBF800000;
|
|
add.f32 %f402, %f414, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f401,%f402;
|
|
// inline asm
|
|
add.f32 %f417, %f416, %f416;
|
|
mul.f32 %f418, %f401, %f417;
|
|
mul.f32 %f419, %f418, %f418;
|
|
fma.rn.f32 %f422, %f1001, %f419, %f1002;
|
|
fma.rn.f32 %f424, %f422, %f419, %f1003;
|
|
mul.rn.f32 %f425, %f424, %f419;
|
|
mul.rn.f32 %f426, %f425, %f418;
|
|
sub.f32 %f427, %f416, %f418;
|
|
neg.f32 %f428, %f418;
|
|
add.f32 %f429, %f427, %f427;
|
|
fma.rn.f32 %f430, %f428, %f416, %f429;
|
|
mul.rn.f32 %f431, %f401, %f430;
|
|
add.f32 %f432, %f426, %f418;
|
|
sub.f32 %f433, %f418, %f432;
|
|
add.f32 %f434, %f426, %f433;
|
|
add.f32 %f435, %f431, %f434;
|
|
add.f32 %f436, %f432, %f435;
|
|
sub.f32 %f437, %f432, %f436;
|
|
add.f32 %f438, %f435, %f437;
|
|
mul.rn.f32 %f440, %f415, %f1004;
|
|
mul.rn.f32 %f442, %f415, %f1005;
|
|
add.f32 %f443, %f440, %f436;
|
|
sub.f32 %f444, %f440, %f443;
|
|
add.f32 %f445, %f436, %f444;
|
|
add.f32 %f446, %f438, %f445;
|
|
add.f32 %f447, %f442, %f446;
|
|
add.f32 %f448, %f443, %f447;
|
|
sub.f32 %f449, %f443, %f448;
|
|
add.f32 %f450, %f447, %f449;
|
|
mov.f32 %f451, 0f3EE8BA2E;
|
|
mul.rn.f32 %f452, %f451, %f448;
|
|
neg.f32 %f453, %f452;
|
|
fma.rn.f32 %f454, %f451, %f448, %f453;
|
|
fma.rn.f32 %f455, %f451, %f450, %f454;
|
|
mov.f32 %f456, 0f00000000;
|
|
fma.rn.f32 %f457, %f456, %f448, %f455;
|
|
add.rn.f32 %f458, %f452, %f457;
|
|
neg.f32 %f459, %f458;
|
|
add.rn.f32 %f460, %f452, %f459;
|
|
add.rn.f32 %f461, %f460, %f457;
|
|
mov.b32 %r80, %f458;
|
|
setp.eq.s32 %p38, %r80, 1118925336;
|
|
add.s32 %r81, %r80, -1;
|
|
mov.b32 %f462, %r81;
|
|
add.f32 %f463, %f461, 0f37000000;
|
|
selp.f32 %f464, %f462, %f458, %p38;
|
|
selp.f32 %f47, %f463, %f461, %p38;
|
|
mul.f32 %f465, %f464, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f466, %f465;
|
|
fma.rn.f32 %f468, %f466, %f1006, %f464;
|
|
fma.rn.f32 %f470, %f466, %f1007, %f468;
|
|
mul.f32 %f471, %f470, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f472, %f471;
|
|
add.f32 %f473, %f466, 0f00000000;
|
|
ex2.approx.f32 %f474, %f473;
|
|
mul.f32 %f475, %f472, %f474;
|
|
setp.lt.f32 %p39, %f464, 0fC2D20000;
|
|
selp.f32 %f476, 0f00000000, %f475, %p39;
|
|
setp.gt.f32 %p40, %f464, 0f42D20000;
|
|
selp.f32 %f1069, 0f7F800000, %f476, %p40;
|
|
setp.eq.f32 %p41, %f1069, 0f7F800000;
|
|
@%p41 bra BB0_23;
|
|
|
|
fma.rn.f32 %f1069, %f1069, %f47, %f1069;
|
|
|
|
BB0_23:
|
|
mov.f32 %f1035, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f1034, %f1035;
|
|
fma.rn.f32 %f1033, %f1034, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f1032, %f1033;
|
|
setp.lt.f32 %p42, %f1068, 0f00000000;
|
|
setp.eq.f32 %p43, %f1032, 0f3F800000;
|
|
and.pred %p2, %p42, %p43;
|
|
mov.b32 %r82, %f1069;
|
|
xor.b32 %r83, %r82, -2147483648;
|
|
mov.b32 %f477, %r83;
|
|
selp.f32 %f1071, %f477, %f1069, %p2;
|
|
setp.eq.f32 %p44, %f1068, 0f00000000;
|
|
@%p44 bra BB0_26;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_26:
|
|
add.f32 %f480, %f1068, %f1068;
|
|
selp.f32 %f1071, %f480, 0f00000000, %p43;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_111:
|
|
mov.u64 %rd272, image_HDR;
|
|
cvta.global.u64 %rd267, %rd272;
|
|
mov.u32 %r216, 8;
|
|
// inline asm
|
|
call (%rd266), _rt_buffer_get_64, (%rd267, %r31, %r216, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f970, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f970;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs173, 0;
|
|
st.v4.u16 [%rd266], {%rs172, %rs172, %rs172, %rs173};
|
|
|
|
BB0_112:
|
|
ld.global.u32 %r217, [additive];
|
|
setp.eq.s32 %p126, %r217, 0;
|
|
@%p126 bra BB0_114;
|
|
|
|
mov.u64 %rd285, image_RNM0;
|
|
cvta.global.u64 %rd274, %rd285;
|
|
mov.u32 %r221, 8;
|
|
// inline asm
|
|
call (%rd273), _rt_buffer_get_64, (%rd274, %r31, %r221, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs180, %rs181, %rs182, %rs183}, [%rd273];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f971, %rs180;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f972, %rs181;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f973, %rs182;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd279), _rt_buffer_get_64, (%rd274, %r31, %r221, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f974, %f971, 0f00000000;
|
|
add.f32 %f975, %f972, 0f00000000;
|
|
add.f32 %f976, %f973, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs179, %f976;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs178, %f975;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs177, %f974;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs184, 0;
|
|
st.v4.u16 [%rd279], {%rs177, %rs178, %rs179, %rs184};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
mov.u64 %rd292, image_RNM0;
|
|
cvta.global.u64 %rd287, %rd292;
|
|
mov.u32 %r223, 8;
|
|
// inline asm
|
|
call (%rd286), _rt_buffer_get_64, (%rd287, %r31, %r223, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f977, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f977;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs186, 0;
|
|
st.v4.u16 [%rd286], {%rs185, %rs185, %rs185, %rs186};
|
|
|
|
BB0_115:
|
|
ld.global.u32 %r224, [additive];
|
|
setp.eq.s32 %p127, %r224, 0;
|
|
@%p127 bra BB0_117;
|
|
|
|
mov.u64 %rd305, image_RNM1;
|
|
cvta.global.u64 %rd294, %rd305;
|
|
mov.u32 %r228, 8;
|
|
// inline asm
|
|
call (%rd293), _rt_buffer_get_64, (%rd294, %r31, %r228, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs193, %rs194, %rs195, %rs196}, [%rd293];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f978, %rs193;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f979, %rs194;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f980, %rs195;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd299), _rt_buffer_get_64, (%rd294, %r31, %r228, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f981, %f978, 0f00000000;
|
|
add.f32 %f982, %f979, 0f00000000;
|
|
add.f32 %f983, %f980, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f983;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs191, %f982;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs190, %f981;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs197, 0;
|
|
st.v4.u16 [%rd299], {%rs190, %rs191, %rs192, %rs197};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
mov.u64 %rd312, image_RNM1;
|
|
cvta.global.u64 %rd307, %rd312;
|
|
mov.u32 %r230, 8;
|
|
// inline asm
|
|
call (%rd306), _rt_buffer_get_64, (%rd307, %r31, %r230, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f984, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f984;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs199, 0;
|
|
st.v4.u16 [%rd306], {%rs198, %rs198, %rs198, %rs199};
|
|
|
|
BB0_118:
|
|
ld.global.u32 %r231, [additive];
|
|
setp.eq.s32 %p128, %r231, 0;
|
|
@%p128 bra BB0_120;
|
|
|
|
mov.u64 %rd325, image_RNM2;
|
|
cvta.global.u64 %rd314, %rd325;
|
|
mov.u32 %r235, 8;
|
|
// inline asm
|
|
call (%rd313), _rt_buffer_get_64, (%rd314, %r31, %r235, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs206, %rs207, %rs208, %rs209}, [%rd313];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f985, %rs206;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f986, %rs207;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f987, %rs208;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd319), _rt_buffer_get_64, (%rd314, %r31, %r235, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f988, %f985, 0f00000000;
|
|
add.f32 %f989, %f986, 0f00000000;
|
|
add.f32 %f990, %f987, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f990;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs204, %f989;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs203, %f988;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs210, 0;
|
|
st.v4.u16 [%rd319], {%rs203, %rs204, %rs205, %rs210};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_120:
|
|
mov.u64 %rd332, image_RNM2;
|
|
cvta.global.u64 %rd327, %rd332;
|
|
mov.u32 %r237, 8;
|
|
// inline asm
|
|
call (%rd326), _rt_buffer_get_64, (%rd327, %r31, %r237, %rd12, %rd13, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f991, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f991;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs212, 0;
|
|
st.v4.u16 [%rd326], {%rs211, %rs211, %rs211, %rs212};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_94:
|
|
mov.u64 %rd178, image_HDR;
|
|
cvta.global.u64 %rd173, %rd178;
|
|
mov.u32 %r180, 8;
|
|
// inline asm
|
|
call (%rd172), _rt_buffer_get_64, (%rd173, %r31, %r180, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f936, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f936;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd172], {%rs121, %rs121, %rs121, %rs110};
|
|
|
|
BB0_95:
|
|
ld.global.u32 %r181, [additive];
|
|
setp.eq.s32 %p119, %r181, 0;
|
|
mov.f32 %f937, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f937;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_97;
|
|
|
|
mov.u64 %rd191, image_RNM0;
|
|
cvta.global.u64 %rd180, %rd191;
|
|
mov.u32 %r185, 8;
|
|
// inline asm
|
|
call (%rd179), _rt_buffer_get_64, (%rd180, %r31, %r185, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs129, %rs130, %rs131, %rs132}, [%rd179];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f938, %rs129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f939, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f940, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd185), _rt_buffer_get_64, (%rd180, %r31, %r185, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f941, %f938, 0f00000000;
|
|
add.f32 %f942, %f939, 0f00000000;
|
|
add.f32 %f943, %f940, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f943;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f942;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs126, %f941;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd185], {%rs126, %rs127, %rs128, %rs122};
|
|
bra.uni BB0_98;
|
|
|
|
BB0_97:
|
|
mov.u64 %rd198, image_RNM0;
|
|
cvta.global.u64 %rd193, %rd198;
|
|
mov.u32 %r187, 8;
|
|
// inline asm
|
|
call (%rd192), _rt_buffer_get_64, (%rd193, %r31, %r187, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f944, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f944;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd192], {%rs133, %rs133, %rs133, %rs122};
|
|
|
|
BB0_98:
|
|
ld.global.u32 %r188, [additive];
|
|
setp.eq.s32 %p120, %r188, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f937;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_100;
|
|
|
|
mov.u64 %rd211, image_RNM1;
|
|
cvta.global.u64 %rd200, %rd211;
|
|
mov.u32 %r192, 8;
|
|
// inline asm
|
|
call (%rd199), _rt_buffer_get_64, (%rd200, %r31, %r192, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs141, %rs142, %rs143, %rs144}, [%rd199];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f946, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f947, %rs142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f948, %rs143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd205), _rt_buffer_get_64, (%rd200, %r31, %r192, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f949, %f946, 0f00000000;
|
|
add.f32 %f950, %f947, 0f00000000;
|
|
add.f32 %f951, %f948, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f951;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs139, %f950;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f949;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd205], {%rs138, %rs139, %rs140, %rs134};
|
|
bra.uni BB0_101;
|
|
|
|
BB0_100:
|
|
mov.u64 %rd218, image_RNM1;
|
|
cvta.global.u64 %rd213, %rd218;
|
|
mov.u32 %r194, 8;
|
|
// inline asm
|
|
call (%rd212), _rt_buffer_get_64, (%rd213, %r31, %r194, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f952, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f952;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd212], {%rs145, %rs145, %rs145, %rs134};
|
|
|
|
BB0_101:
|
|
ld.global.u32 %r195, [additive];
|
|
setp.eq.s32 %p121, %r195, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f937;}
|
|
|
|
// inline asm
|
|
@%p121 bra BB0_103;
|
|
|
|
mov.u64 %rd231, image_RNM2;
|
|
cvta.global.u64 %rd220, %rd231;
|
|
mov.u32 %r199, 8;
|
|
// inline asm
|
|
call (%rd219), _rt_buffer_get_64, (%rd220, %r31, %r199, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs153, %rs154, %rs155, %rs156}, [%rd219];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f954, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f955, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f956, %rs155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd225), _rt_buffer_get_64, (%rd220, %r31, %r199, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f957, %f954, 0f00000000;
|
|
add.f32 %f958, %f955, 0f00000000;
|
|
add.f32 %f959, %f956, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f959;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f958;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f957;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd225], {%rs150, %rs151, %rs152, %rs146};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_103:
|
|
mov.u64 %rd238, image_RNM2;
|
|
cvta.global.u64 %rd233, %rd238;
|
|
mov.u32 %r201, 8;
|
|
// inline asm
|
|
call (%rd232), _rt_buffer_get_64, (%rd233, %r31, %r201, %rd10, %rd11, %rd19, %rd19);
|
|
// inline asm
|
|
mov.f32 %f960, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f960;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd232], {%rs157, %rs157, %rs157, %rs146};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_24:
|
|
setp.geu.f32 %p45, %f1068, 0f00000000;
|
|
@%p45 bra BB0_27;
|
|
|
|
mov.f32 %f1039, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f479, %f1039;
|
|
setp.neu.f32 %p46, %f479, 0f3EE8BA2E;
|
|
selp.f32 %f1071, 0f7FFFFFFF, %f1071, %p46;
|
|
|
|
BB0_27:
|
|
abs.f32 %f1036, %f1068;
|
|
add.f32 %f481, %f1036, 0f3EE8BA2E;
|
|
mov.b32 %r84, %f481;
|
|
setp.lt.s32 %p48, %r84, 2139095040;
|
|
@%p48 bra BB0_32;
|
|
|
|
abs.f32 %f1037, %f1068;
|
|
setp.gtu.f32 %p49, %f1037, 0f7F800000;
|
|
@%p49 bra BB0_31;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_31:
|
|
add.f32 %f1071, %f1068, 0f3EE8BA2E;
|
|
bra.uni BB0_32;
|
|
|
|
BB0_29:
|
|
abs.f32 %f1038, %f1068;
|
|
setp.neu.f32 %p50, %f1038, 0f7F800000;
|
|
@%p50 bra BB0_32;
|
|
|
|
selp.f32 %f1071, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_32:
|
|
mul.f32 %f482, %f1071, 0f437F0000;
|
|
setp.eq.f32 %p51, %f1068, 0f3F800000;
|
|
selp.f32 %f483, 0f437F0000, %f482, %p51;
|
|
cvt.rzi.u32.f32 %r85, %f483;
|
|
cvt.u16.u32 %rs17, %r85;
|
|
mov.u16 %rs18, 255;
|
|
st.v2.u8 [%rd30], {%rs17, %rs18};
|
|
ld.global.u32 %r245, [imageEnabled];
|
|
|
|
BB0_33:
|
|
and.b32 %r86, %r245, 1;
|
|
setp.eq.b32 %p52, %r86, 1;
|
|
@!%p52 bra BB0_68;
|
|
bra.uni BB0_34;
|
|
|
|
BB0_34:
|
|
mov.f32 %f1014, 0fB5BFBE8E;
|
|
mov.f32 %f1013, 0fBF317200;
|
|
mov.f32 %f1012, 0f35BFBE8E;
|
|
mov.f32 %f1011, 0f3F317200;
|
|
mov.f32 %f1010, 0f3DAAAABD;
|
|
mov.f32 %f1009, 0f3C4CAF63;
|
|
mov.f32 %f1008, 0f3B18F0FE;
|
|
abs.f32 %f59, %f42;
|
|
setp.lt.f32 %p53, %f59, 0f00800000;
|
|
mul.f32 %f489, %f59, 0f4B800000;
|
|
selp.f32 %f490, 0fC3170000, 0fC2FE0000, %p53;
|
|
selp.f32 %f491, %f489, %f59, %p53;
|
|
mov.b32 %r87, %f491;
|
|
and.b32 %r88, %r87, 8388607;
|
|
or.b32 %r89, %r88, 1065353216;
|
|
mov.b32 %f492, %r89;
|
|
shr.u32 %r90, %r87, 23;
|
|
cvt.rn.f32.u32 %f493, %r90;
|
|
add.f32 %f494, %f490, %f493;
|
|
setp.gt.f32 %p54, %f492, 0f3FB504F3;
|
|
mul.f32 %f495, %f492, 0f3F000000;
|
|
add.f32 %f496, %f494, 0f3F800000;
|
|
selp.f32 %f497, %f495, %f492, %p54;
|
|
selp.f32 %f498, %f496, %f494, %p54;
|
|
add.f32 %f499, %f497, 0fBF800000;
|
|
add.f32 %f485, %f497, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f484,%f485;
|
|
// inline asm
|
|
add.f32 %f500, %f499, %f499;
|
|
mul.f32 %f501, %f484, %f500;
|
|
mul.f32 %f502, %f501, %f501;
|
|
fma.rn.f32 %f505, %f1008, %f502, %f1009;
|
|
fma.rn.f32 %f507, %f505, %f502, %f1010;
|
|
mul.rn.f32 %f508, %f507, %f502;
|
|
mul.rn.f32 %f509, %f508, %f501;
|
|
sub.f32 %f510, %f499, %f501;
|
|
neg.f32 %f511, %f501;
|
|
add.f32 %f512, %f510, %f510;
|
|
fma.rn.f32 %f513, %f511, %f499, %f512;
|
|
mul.rn.f32 %f514, %f484, %f513;
|
|
add.f32 %f515, %f509, %f501;
|
|
sub.f32 %f516, %f501, %f515;
|
|
add.f32 %f517, %f509, %f516;
|
|
add.f32 %f518, %f514, %f517;
|
|
add.f32 %f519, %f515, %f518;
|
|
sub.f32 %f520, %f515, %f519;
|
|
add.f32 %f521, %f518, %f520;
|
|
mul.rn.f32 %f523, %f498, %f1011;
|
|
mul.rn.f32 %f525, %f498, %f1012;
|
|
add.f32 %f526, %f523, %f519;
|
|
sub.f32 %f527, %f523, %f526;
|
|
add.f32 %f528, %f519, %f527;
|
|
add.f32 %f529, %f521, %f528;
|
|
add.f32 %f530, %f525, %f529;
|
|
add.f32 %f531, %f526, %f530;
|
|
sub.f32 %f532, %f526, %f531;
|
|
add.f32 %f533, %f530, %f532;
|
|
mov.f32 %f534, 0f3EE66666;
|
|
mul.rn.f32 %f535, %f534, %f531;
|
|
neg.f32 %f536, %f535;
|
|
fma.rn.f32 %f537, %f534, %f531, %f536;
|
|
fma.rn.f32 %f538, %f534, %f533, %f537;
|
|
mov.f32 %f539, 0f00000000;
|
|
fma.rn.f32 %f540, %f539, %f531, %f538;
|
|
add.rn.f32 %f541, %f535, %f540;
|
|
neg.f32 %f542, %f541;
|
|
add.rn.f32 %f543, %f535, %f542;
|
|
add.rn.f32 %f544, %f543, %f540;
|
|
mov.b32 %r91, %f541;
|
|
setp.eq.s32 %p55, %r91, 1118925336;
|
|
add.s32 %r92, %r91, -1;
|
|
mov.b32 %f545, %r92;
|
|
add.f32 %f546, %f544, 0f37000000;
|
|
selp.f32 %f547, %f545, %f541, %p55;
|
|
selp.f32 %f60, %f546, %f544, %p55;
|
|
mul.f32 %f548, %f547, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f549, %f548;
|
|
fma.rn.f32 %f551, %f549, %f1013, %f547;
|
|
fma.rn.f32 %f553, %f549, %f1014, %f551;
|
|
mul.f32 %f554, %f553, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f555, %f554;
|
|
add.f32 %f556, %f549, 0f00000000;
|
|
ex2.approx.f32 %f557, %f556;
|
|
mul.f32 %f558, %f555, %f557;
|
|
setp.lt.f32 %p56, %f547, 0fC2D20000;
|
|
selp.f32 %f559, 0f00000000, %f558, %p56;
|
|
setp.gt.f32 %p57, %f547, 0f42D20000;
|
|
selp.f32 %f1072, 0f7F800000, %f559, %p57;
|
|
setp.eq.f32 %p58, %f1072, 0f7F800000;
|
|
@%p58 bra BB0_36;
|
|
|
|
fma.rn.f32 %f1072, %f1072, %f60, %f1072;
|
|
|
|
BB0_36:
|
|
mov.f32 %f1043, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1042, %f1043;
|
|
fma.rn.f32 %f1041, %f1042, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f1040, %f1041;
|
|
setp.lt.f32 %p59, %f42, 0f00000000;
|
|
setp.eq.f32 %p60, %f1040, 0f3F800000;
|
|
and.pred %p3, %p59, %p60;
|
|
mov.b32 %r93, %f1072;
|
|
xor.b32 %r94, %r93, -2147483648;
|
|
mov.b32 %f560, %r94;
|
|
selp.f32 %f1074, %f560, %f1072, %p3;
|
|
setp.eq.f32 %p61, %f42, 0f00000000;
|
|
@%p61 bra BB0_39;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_39:
|
|
add.f32 %f563, %f42, %f42;
|
|
selp.f32 %f1074, %f563, 0f00000000, %p60;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_37:
|
|
setp.geu.f32 %p62, %f42, 0f00000000;
|
|
@%p62 bra BB0_40;
|
|
|
|
mov.f32 %f1053, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f562, %f1053;
|
|
setp.neu.f32 %p63, %f562, 0f3EE66666;
|
|
selp.f32 %f1074, 0f7FFFFFFF, %f1074, %p63;
|
|
|
|
BB0_40:
|
|
abs.f32 %f1044, %f42;
|
|
add.f32 %f564, %f1044, 0f3EE66666;
|
|
mov.b32 %r95, %f564;
|
|
setp.lt.s32 %p65, %r95, 2139095040;
|
|
@%p65 bra BB0_45;
|
|
|
|
abs.f32 %f1051, %f42;
|
|
setp.gtu.f32 %p66, %f1051, 0f7F800000;
|
|
@%p66 bra BB0_44;
|
|
bra.uni BB0_42;
|
|
|
|
BB0_44:
|
|
add.f32 %f1074, %f42, 0f3EE66666;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_42:
|
|
abs.f32 %f1052, %f42;
|
|
setp.neu.f32 %p67, %f1052, 0f7F800000;
|
|
@%p67 bra BB0_45;
|
|
|
|
selp.f32 %f1074, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_45:
|
|
mov.f32 %f1046, 0f00000000;
|
|
mov.f32 %f1045, 0f3EE66666;
|
|
mov.f32 %f1021, 0fB5BFBE8E;
|
|
mov.f32 %f1020, 0fBF317200;
|
|
mov.f32 %f1019, 0f35BFBE8E;
|
|
mov.f32 %f1018, 0f3F317200;
|
|
mov.f32 %f1017, 0f3DAAAABD;
|
|
mov.f32 %f1016, 0f3C4CAF63;
|
|
mov.f32 %f1015, 0f3B18F0FE;
|
|
setp.eq.f32 %p68, %f42, 0f3F800000;
|
|
selp.f32 %f71, 0f3F800000, %f1074, %p68;
|
|
abs.f32 %f72, %f43;
|
|
setp.lt.f32 %p69, %f72, 0f00800000;
|
|
mul.f32 %f567, %f72, 0f4B800000;
|
|
selp.f32 %f568, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f569, %f567, %f72, %p69;
|
|
mov.b32 %r96, %f569;
|
|
and.b32 %r97, %r96, 8388607;
|
|
or.b32 %r98, %r97, 1065353216;
|
|
mov.b32 %f570, %r98;
|
|
shr.u32 %r99, %r96, 23;
|
|
cvt.rn.f32.u32 %f571, %r99;
|
|
add.f32 %f572, %f568, %f571;
|
|
setp.gt.f32 %p70, %f570, 0f3FB504F3;
|
|
mul.f32 %f573, %f570, 0f3F000000;
|
|
add.f32 %f574, %f572, 0f3F800000;
|
|
selp.f32 %f575, %f573, %f570, %p70;
|
|
selp.f32 %f576, %f574, %f572, %p70;
|
|
add.f32 %f577, %f575, 0fBF800000;
|
|
add.f32 %f566, %f575, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f565,%f566;
|
|
// inline asm
|
|
add.f32 %f578, %f577, %f577;
|
|
mul.f32 %f579, %f565, %f578;
|
|
mul.f32 %f580, %f579, %f579;
|
|
fma.rn.f32 %f583, %f1015, %f580, %f1016;
|
|
fma.rn.f32 %f585, %f583, %f580, %f1017;
|
|
mul.rn.f32 %f586, %f585, %f580;
|
|
mul.rn.f32 %f587, %f586, %f579;
|
|
sub.f32 %f588, %f577, %f579;
|
|
neg.f32 %f589, %f579;
|
|
add.f32 %f590, %f588, %f588;
|
|
fma.rn.f32 %f591, %f589, %f577, %f590;
|
|
mul.rn.f32 %f592, %f565, %f591;
|
|
add.f32 %f593, %f587, %f579;
|
|
sub.f32 %f594, %f579, %f593;
|
|
add.f32 %f595, %f587, %f594;
|
|
add.f32 %f596, %f592, %f595;
|
|
add.f32 %f597, %f593, %f596;
|
|
sub.f32 %f598, %f593, %f597;
|
|
add.f32 %f599, %f596, %f598;
|
|
mul.rn.f32 %f601, %f576, %f1018;
|
|
mul.rn.f32 %f603, %f576, %f1019;
|
|
add.f32 %f604, %f601, %f597;
|
|
sub.f32 %f605, %f601, %f604;
|
|
add.f32 %f606, %f597, %f605;
|
|
add.f32 %f607, %f599, %f606;
|
|
add.f32 %f608, %f603, %f607;
|
|
add.f32 %f609, %f604, %f608;
|
|
sub.f32 %f610, %f604, %f609;
|
|
add.f32 %f611, %f608, %f610;
|
|
mul.rn.f32 %f613, %f1045, %f609;
|
|
neg.f32 %f614, %f613;
|
|
fma.rn.f32 %f615, %f1045, %f609, %f614;
|
|
fma.rn.f32 %f616, %f1045, %f611, %f615;
|
|
fma.rn.f32 %f618, %f1046, %f609, %f616;
|
|
add.rn.f32 %f619, %f613, %f618;
|
|
neg.f32 %f620, %f619;
|
|
add.rn.f32 %f621, %f613, %f620;
|
|
add.rn.f32 %f622, %f621, %f618;
|
|
mov.b32 %r100, %f619;
|
|
setp.eq.s32 %p71, %r100, 1118925336;
|
|
add.s32 %r101, %r100, -1;
|
|
mov.b32 %f623, %r101;
|
|
add.f32 %f624, %f622, 0f37000000;
|
|
selp.f32 %f625, %f623, %f619, %p71;
|
|
selp.f32 %f73, %f624, %f622, %p71;
|
|
mul.f32 %f626, %f625, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f627, %f626;
|
|
fma.rn.f32 %f629, %f627, %f1020, %f625;
|
|
fma.rn.f32 %f631, %f627, %f1021, %f629;
|
|
mul.f32 %f632, %f631, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f633, %f632;
|
|
add.f32 %f634, %f627, 0f00000000;
|
|
ex2.approx.f32 %f635, %f634;
|
|
mul.f32 %f636, %f633, %f635;
|
|
setp.lt.f32 %p72, %f625, 0fC2D20000;
|
|
selp.f32 %f637, 0f00000000, %f636, %p72;
|
|
setp.gt.f32 %p73, %f625, 0f42D20000;
|
|
selp.f32 %f1075, 0f7F800000, %f637, %p73;
|
|
setp.eq.f32 %p74, %f1075, 0f7F800000;
|
|
@%p74 bra BB0_47;
|
|
|
|
fma.rn.f32 %f1075, %f1075, %f73, %f1075;
|
|
|
|
BB0_47:
|
|
setp.lt.f32 %p75, %f43, 0f00000000;
|
|
and.pred %p4, %p75, %p60;
|
|
mov.b32 %r102, %f1075;
|
|
xor.b32 %r103, %r102, -2147483648;
|
|
mov.b32 %f638, %r103;
|
|
selp.f32 %f1077, %f638, %f1075, %p4;
|
|
setp.eq.f32 %p77, %f43, 0f00000000;
|
|
@%p77 bra BB0_50;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_50:
|
|
add.f32 %f641, %f43, %f43;
|
|
selp.f32 %f1077, %f641, 0f00000000, %p60;
|
|
bra.uni BB0_51;
|
|
|
|
BB0_48:
|
|
setp.geu.f32 %p78, %f43, 0f00000000;
|
|
@%p78 bra BB0_51;
|
|
|
|
mov.f32 %f1050, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f640, %f1050;
|
|
setp.neu.f32 %p79, %f640, 0f3EE66666;
|
|
selp.f32 %f1077, 0f7FFFFFFF, %f1077, %p79;
|
|
|
|
BB0_51:
|
|
abs.f32 %f1054, %f43;
|
|
add.f32 %f642, %f1054, 0f3EE66666;
|
|
mov.b32 %r104, %f642;
|
|
setp.lt.s32 %p81, %r104, 2139095040;
|
|
@%p81 bra BB0_56;
|
|
|
|
abs.f32 %f1055, %f43;
|
|
setp.gtu.f32 %p82, %f1055, 0f7F800000;
|
|
@%p82 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
add.f32 %f1077, %f43, 0f3EE66666;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_53:
|
|
abs.f32 %f1056, %f43;
|
|
setp.neu.f32 %p83, %f1056, 0f7F800000;
|
|
@%p83 bra BB0_56;
|
|
|
|
selp.f32 %f1077, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_56:
|
|
mov.f32 %f1048, 0f00000000;
|
|
mov.f32 %f1047, 0f3EE66666;
|
|
mov.f32 %f1028, 0fB5BFBE8E;
|
|
mov.f32 %f1027, 0fBF317200;
|
|
mov.f32 %f1026, 0f35BFBE8E;
|
|
mov.f32 %f1025, 0f3F317200;
|
|
mov.f32 %f1024, 0f3DAAAABD;
|
|
mov.f32 %f1023, 0f3C4CAF63;
|
|
mov.f32 %f1022, 0f3B18F0FE;
|
|
setp.eq.f32 %p84, %f43, 0f3F800000;
|
|
selp.f32 %f84, 0f3F800000, %f1077, %p84;
|
|
abs.f32 %f85, %f44;
|
|
setp.lt.f32 %p85, %f85, 0f00800000;
|
|
mul.f32 %f645, %f85, 0f4B800000;
|
|
selp.f32 %f646, 0fC3170000, 0fC2FE0000, %p85;
|
|
selp.f32 %f647, %f645, %f85, %p85;
|
|
mov.b32 %r105, %f647;
|
|
and.b32 %r106, %r105, 8388607;
|
|
or.b32 %r107, %r106, 1065353216;
|
|
mov.b32 %f648, %r107;
|
|
shr.u32 %r108, %r105, 23;
|
|
cvt.rn.f32.u32 %f649, %r108;
|
|
add.f32 %f650, %f646, %f649;
|
|
setp.gt.f32 %p86, %f648, 0f3FB504F3;
|
|
mul.f32 %f651, %f648, 0f3F000000;
|
|
add.f32 %f652, %f650, 0f3F800000;
|
|
selp.f32 %f653, %f651, %f648, %p86;
|
|
selp.f32 %f654, %f652, %f650, %p86;
|
|
add.f32 %f655, %f653, 0fBF800000;
|
|
add.f32 %f644, %f653, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f643,%f644;
|
|
// inline asm
|
|
add.f32 %f656, %f655, %f655;
|
|
mul.f32 %f657, %f643, %f656;
|
|
mul.f32 %f658, %f657, %f657;
|
|
fma.rn.f32 %f661, %f1022, %f658, %f1023;
|
|
fma.rn.f32 %f663, %f661, %f658, %f1024;
|
|
mul.rn.f32 %f664, %f663, %f658;
|
|
mul.rn.f32 %f665, %f664, %f657;
|
|
sub.f32 %f666, %f655, %f657;
|
|
neg.f32 %f667, %f657;
|
|
add.f32 %f668, %f666, %f666;
|
|
fma.rn.f32 %f669, %f667, %f655, %f668;
|
|
mul.rn.f32 %f670, %f643, %f669;
|
|
add.f32 %f671, %f665, %f657;
|
|
sub.f32 %f672, %f657, %f671;
|
|
add.f32 %f673, %f665, %f672;
|
|
add.f32 %f674, %f670, %f673;
|
|
add.f32 %f675, %f671, %f674;
|
|
sub.f32 %f676, %f671, %f675;
|
|
add.f32 %f677, %f674, %f676;
|
|
mul.rn.f32 %f679, %f654, %f1025;
|
|
mul.rn.f32 %f681, %f654, %f1026;
|
|
add.f32 %f682, %f679, %f675;
|
|
sub.f32 %f683, %f679, %f682;
|
|
add.f32 %f684, %f675, %f683;
|
|
add.f32 %f685, %f677, %f684;
|
|
add.f32 %f686, %f681, %f685;
|
|
add.f32 %f687, %f682, %f686;
|
|
sub.f32 %f688, %f682, %f687;
|
|
add.f32 %f689, %f686, %f688;
|
|
mul.rn.f32 %f691, %f1047, %f687;
|
|
neg.f32 %f692, %f691;
|
|
fma.rn.f32 %f693, %f1047, %f687, %f692;
|
|
fma.rn.f32 %f694, %f1047, %f689, %f693;
|
|
fma.rn.f32 %f696, %f1048, %f687, %f694;
|
|
add.rn.f32 %f697, %f691, %f696;
|
|
neg.f32 %f698, %f697;
|
|
add.rn.f32 %f699, %f691, %f698;
|
|
add.rn.f32 %f700, %f699, %f696;
|
|
mov.b32 %r109, %f697;
|
|
setp.eq.s32 %p87, %r109, 1118925336;
|
|
add.s32 %r110, %r109, -1;
|
|
mov.b32 %f701, %r110;
|
|
add.f32 %f702, %f700, 0f37000000;
|
|
selp.f32 %f703, %f701, %f697, %p87;
|
|
selp.f32 %f86, %f702, %f700, %p87;
|
|
mul.f32 %f704, %f703, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f705, %f704;
|
|
fma.rn.f32 %f707, %f705, %f1027, %f703;
|
|
fma.rn.f32 %f709, %f705, %f1028, %f707;
|
|
mul.f32 %f710, %f709, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f711, %f710;
|
|
add.f32 %f712, %f705, 0f00000000;
|
|
ex2.approx.f32 %f713, %f712;
|
|
mul.f32 %f714, %f711, %f713;
|
|
setp.lt.f32 %p88, %f703, 0fC2D20000;
|
|
selp.f32 %f715, 0f00000000, %f714, %p88;
|
|
setp.gt.f32 %p89, %f703, 0f42D20000;
|
|
selp.f32 %f1078, 0f7F800000, %f715, %p89;
|
|
setp.eq.f32 %p90, %f1078, 0f7F800000;
|
|
@%p90 bra BB0_58;
|
|
|
|
fma.rn.f32 %f1078, %f1078, %f86, %f1078;
|
|
|
|
BB0_58:
|
|
setp.lt.f32 %p91, %f44, 0f00000000;
|
|
and.pred %p5, %p91, %p60;
|
|
mov.b32 %r111, %f1078;
|
|
xor.b32 %r112, %r111, -2147483648;
|
|
mov.b32 %f716, %r112;
|
|
selp.f32 %f1080, %f716, %f1078, %p5;
|
|
setp.eq.f32 %p93, %f44, 0f00000000;
|
|
@%p93 bra BB0_61;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_61:
|
|
add.f32 %f719, %f44, %f44;
|
|
selp.f32 %f1080, %f719, 0f00000000, %p60;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_59:
|
|
setp.geu.f32 %p94, %f44, 0f00000000;
|
|
@%p94 bra BB0_62;
|
|
|
|
mov.f32 %f1049, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f718, %f1049;
|
|
setp.neu.f32 %p95, %f718, 0f3EE66666;
|
|
selp.f32 %f1080, 0f7FFFFFFF, %f1080, %p95;
|
|
|
|
BB0_62:
|
|
abs.f32 %f1057, %f44;
|
|
add.f32 %f720, %f1057, 0f3EE66666;
|
|
mov.b32 %r113, %f720;
|
|
setp.lt.s32 %p97, %r113, 2139095040;
|
|
@%p97 bra BB0_67;
|
|
|
|
abs.f32 %f1058, %f44;
|
|
setp.gtu.f32 %p98, %f1058, 0f7F800000;
|
|
@%p98 bra BB0_66;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_66:
|
|
add.f32 %f1080, %f44, 0f3EE66666;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_64:
|
|
abs.f32 %f1059, %f44;
|
|
setp.neu.f32 %p99, %f1059, 0f7F800000;
|
|
@%p99 bra BB0_67;
|
|
|
|
selp.f32 %f1080, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_67:
|
|
mov.u32 %r238, 4;
|
|
setp.eq.f32 %p100, %f44, 0f3F800000;
|
|
selp.f32 %f721, 0f3F800000, %f1080, %p100;
|
|
cvt.u64.u32 %rd40, %r3;
|
|
cvt.u64.u32 %rd39, %r2;
|
|
mov.u64 %rd43, image;
|
|
cvta.global.u64 %rd38, %rd43;
|
|
// inline asm
|
|
call (%rd37), _rt_buffer_get_64, (%rd38, %r31, %r238, %rd39, %rd40, %rd19, %rd19);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f722, %f721;
|
|
mul.f32 %f723, %f722, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r116, %f723;
|
|
cvt.sat.f32.f32 %f724, %f84;
|
|
mul.f32 %f725, %f724, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r117, %f725;
|
|
cvt.sat.f32.f32 %f726, %f71;
|
|
mul.f32 %f727, %f726, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r118, %f727;
|
|
cvt.u16.u32 %rs19, %r116;
|
|
cvt.u16.u32 %rs20, %r118;
|
|
cvt.u16.u32 %rs21, %r117;
|
|
mov.u16 %rs22, 255;
|
|
st.v4.u8 [%rd37], {%rs19, %rs21, %rs20, %rs22};
|
|
ld.global.u32 %r245, [imageEnabled];
|
|
|
|
BB0_68:
|
|
and.b32 %r119, %r245, 4;
|
|
setp.eq.s32 %p101, %r119, 0;
|
|
@%p101 bra BB0_72;
|
|
|
|
ld.global.u32 %r120, [additive];
|
|
setp.eq.s32 %p102, %r120, 0;
|
|
cvt.u64.u32 %rd3, %r2;
|
|
cvt.u64.u32 %rd4, %r3;
|
|
mov.f32 %f728, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs23, %f728;}
|
|
|
|
// inline asm
|
|
@%p102 bra BB0_71;
|
|
|
|
mov.u64 %rd56, image_HDR;
|
|
cvta.global.u64 %rd45, %rd56;
|
|
mov.u32 %r124, 8;
|
|
// inline asm
|
|
call (%rd44), _rt_buffer_get_64, (%rd45, %r31, %r124, %rd3, %rd4, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs30, %rs31, %rs32, %rs33}, [%rd44];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f729, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f730, %rs31;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f731, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd45, %r31, %r124, %rd3, %rd4, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f732, %f42, %f729;
|
|
add.f32 %f733, %f43, %f730;
|
|
add.f32 %f734, %f44, %f731;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f734;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs28, %f733;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f732;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd50], {%rs27, %rs28, %rs29, %rs23};
|
|
bra.uni BB0_72;
|
|
|
|
BB0_71:
|
|
mov.u64 %rd63, image_HDR;
|
|
cvta.global.u64 %rd58, %rd63;
|
|
mov.u32 %r126, 8;
|
|
// inline asm
|
|
call (%rd57), _rt_buffer_get_64, (%rd58, %r31, %r126, %rd3, %rd4, %rd19, %rd19);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f42;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd57], {%rs34, %rs35, %rs36, %rs23};
|
|
|
|
BB0_72:
|
|
mov.u32 %r239, 4;
|
|
mul.f32 %f741, %f28, %f1068;
|
|
ld.global.f32 %f742, [lightColor];
|
|
mul.f32 %f743, %f741, %f742;
|
|
ld.global.f32 %f744, [lightColor+4];
|
|
mul.f32 %f745, %f741, %f744;
|
|
ld.global.f32 %f746, [lightColor+8];
|
|
mul.f32 %f747, %f741, %f746;
|
|
mul.f32 %f97, %f248, %f743;
|
|
mul.f32 %f98, %f249, %f745;
|
|
mul.f32 %f99, %f250, %f747;
|
|
ld.global.v2.u32 {%r129, %r130}, [pixelID];
|
|
cvt.u64.u32 %rd66, %r129;
|
|
cvt.u64.u32 %rd67, %r130;
|
|
mov.u64 %rd70, uvtangent;
|
|
cvta.global.u64 %rd65, %rd70;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r31, %r239, %rd66, %rd67, %rd19, %rd19);
|
|
// inline asm
|
|
ld.u32 %r15, [%rd64];
|
|
shr.u32 %r16, %r15, 16;
|
|
cvt.u16.u32 %rs37, %r16;
|
|
and.b16 %rs38, %rs37, 255;
|
|
cvt.u16.u32 %rs39, %r15;
|
|
or.b16 %rs40, %rs39, %rs38;
|
|
setp.eq.s16 %p103, %rs40, 0;
|
|
mov.f32 %f1087, 0f00000000;
|
|
mov.f32 %f1081, %f1087;
|
|
mov.f32 %f1082, %f1087;
|
|
mov.f32 %f1083, %f1087;
|
|
@%p103 bra BB0_74;
|
|
|
|
ld.u8 %rs41, [%rd64+1];
|
|
and.b16 %rs43, %rs39, 255;
|
|
cvt.rn.f32.u16 %f748, %rs43;
|
|
div.rn.f32 %f749, %f748, 0f437F0000;
|
|
fma.rn.f32 %f750, %f749, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f751, %rs41;
|
|
div.rn.f32 %f752, %f751, 0f437F0000;
|
|
fma.rn.f32 %f753, %f752, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f754, %rs38;
|
|
div.rn.f32 %f755, %f754, 0f437F0000;
|
|
fma.rn.f32 %f756, %f755, 0f40000000, 0fBF800000;
|
|
mul.f32 %f757, %f753, %f753;
|
|
fma.rn.f32 %f758, %f750, %f750, %f757;
|
|
fma.rn.f32 %f759, %f756, %f756, %f758;
|
|
sqrt.rn.f32 %f760, %f759;
|
|
rcp.rn.f32 %f761, %f760;
|
|
mul.f32 %f1081, %f750, %f761;
|
|
mul.f32 %f1082, %f753, %f761;
|
|
mul.f32 %f1083, %f756, %f761;
|
|
|
|
BB0_74:
|
|
mov.u32 %r240, 4;
|
|
mul.f32 %f765, %f1062, %f1082;
|
|
mul.f32 %f766, %f1061, %f1083;
|
|
sub.f32 %f767, %f766, %f765;
|
|
mul.f32 %f768, %f1060, %f1083;
|
|
mul.f32 %f769, %f1062, %f1081;
|
|
sub.f32 %f770, %f769, %f768;
|
|
mul.f32 %f771, %f1061, %f1081;
|
|
mul.f32 %f772, %f1060, %f1082;
|
|
sub.f32 %f773, %f772, %f771;
|
|
setp.lt.u32 %p104, %r15, 16777216;
|
|
selp.f32 %f774, 0fBF800000, 0f3F800000, %p104;
|
|
mul.f32 %f775, %f767, %f774;
|
|
mul.f32 %f776, %f770, %f774;
|
|
mul.f32 %f777, %f773, %f774;
|
|
mul.f32 %f778, %f775, 0f00000000;
|
|
mul.f32 %f779, %f776, 0f00000000;
|
|
mul.f32 %f780, %f777, 0f00000000;
|
|
fma.rn.f32 %f781, %f1081, 0f3F5105EC, %f778;
|
|
fma.rn.f32 %f782, %f1082, 0f3F5105EC, %f779;
|
|
fma.rn.f32 %f783, %f1083, 0f3F5105EC, %f780;
|
|
mul.f32 %f106, %f1060, 0f3F13CD3A;
|
|
add.f32 %f107, %f106, %f781;
|
|
mul.f32 %f108, %f1061, 0f3F13CD3A;
|
|
add.f32 %f109, %f108, %f782;
|
|
mul.f32 %f110, %f1062, 0f3F13CD3A;
|
|
add.f32 %f111, %f110, %f783;
|
|
ld.global.v2.u32 {%r135, %r136}, [pixelID];
|
|
cvt.u64.u32 %rd73, %r135;
|
|
cvt.u64.u32 %rd74, %r136;
|
|
// inline asm
|
|
call (%rd71), _rt_buffer_get_64, (%rd65, %r31, %r240, %rd73, %rd74, %rd19, %rd19);
|
|
// inline asm
|
|
ld.u32 %r17, [%rd71];
|
|
shr.u32 %r18, %r17, 16;
|
|
cvt.u16.u32 %rs46, %r18;
|
|
and.b16 %rs47, %rs46, 255;
|
|
cvt.u16.u32 %rs48, %r17;
|
|
or.b16 %rs49, %rs48, %rs47;
|
|
setp.eq.s16 %p105, %rs49, 0;
|
|
mov.f32 %f1084, %f1087;
|
|
mov.f32 %f1085, %f1087;
|
|
mov.f32 %f1086, %f1087;
|
|
@%p105 bra BB0_76;
|
|
|
|
ld.u8 %rs50, [%rd71+1];
|
|
and.b16 %rs52, %rs48, 255;
|
|
cvt.rn.f32.u16 %f784, %rs52;
|
|
div.rn.f32 %f785, %f784, 0f437F0000;
|
|
fma.rn.f32 %f786, %f785, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f787, %rs50;
|
|
div.rn.f32 %f788, %f787, 0f437F0000;
|
|
fma.rn.f32 %f789, %f788, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f790, %rs47;
|
|
div.rn.f32 %f791, %f790, 0f437F0000;
|
|
fma.rn.f32 %f792, %f791, 0f40000000, 0fBF800000;
|
|
mul.f32 %f793, %f789, %f789;
|
|
fma.rn.f32 %f794, %f786, %f786, %f793;
|
|
fma.rn.f32 %f795, %f792, %f792, %f794;
|
|
sqrt.rn.f32 %f796, %f795;
|
|
rcp.rn.f32 %f797, %f796;
|
|
mul.f32 %f1084, %f786, %f797;
|
|
mul.f32 %f1085, %f789, %f797;
|
|
mul.f32 %f1086, %f792, %f797;
|
|
|
|
BB0_76:
|
|
mov.u32 %r241, 4;
|
|
mul.f32 %f801, %f1062, %f1085;
|
|
mul.f32 %f802, %f1061, %f1086;
|
|
sub.f32 %f803, %f802, %f801;
|
|
mul.f32 %f804, %f1060, %f1086;
|
|
mul.f32 %f805, %f1062, %f1084;
|
|
sub.f32 %f806, %f805, %f804;
|
|
mul.f32 %f807, %f1061, %f1084;
|
|
mul.f32 %f808, %f1060, %f1085;
|
|
sub.f32 %f809, %f808, %f807;
|
|
setp.lt.u32 %p106, %r17, 16777216;
|
|
selp.f32 %f810, 0fBF800000, 0f3F800000, %p106;
|
|
mul.f32 %f811, %f803, %f810;
|
|
mul.f32 %f812, %f806, %f810;
|
|
mul.f32 %f813, %f809, %f810;
|
|
mul.f32 %f814, %f811, 0f3F3504F3;
|
|
mul.f32 %f815, %f812, 0f3F3504F3;
|
|
mul.f32 %f816, %f813, 0f3F3504F3;
|
|
fma.rn.f32 %f817, %f1084, 0fBED105EC, %f814;
|
|
fma.rn.f32 %f818, %f1085, 0fBED105EC, %f815;
|
|
fma.rn.f32 %f819, %f1086, 0fBED105EC, %f816;
|
|
add.f32 %f118, %f106, %f817;
|
|
add.f32 %f119, %f108, %f818;
|
|
add.f32 %f120, %f110, %f819;
|
|
ld.global.v2.u32 {%r141, %r142}, [pixelID];
|
|
cvt.u64.u32 %rd80, %r141;
|
|
cvt.u64.u32 %rd81, %r142;
|
|
// inline asm
|
|
call (%rd78), _rt_buffer_get_64, (%rd65, %r31, %r241, %rd80, %rd81, %rd19, %rd19);
|
|
// inline asm
|
|
ld.u32 %r19, [%rd78];
|
|
shr.u32 %r20, %r19, 16;
|
|
cvt.u16.u32 %rs55, %r20;
|
|
and.b16 %rs56, %rs55, 255;
|
|
cvt.u16.u32 %rs57, %r19;
|
|
or.b16 %rs58, %rs57, %rs56;
|
|
setp.eq.s16 %p107, %rs58, 0;
|
|
mov.f32 %f1088, %f1087;
|
|
mov.f32 %f1089, %f1087;
|
|
@%p107 bra BB0_78;
|
|
|
|
ld.u8 %rs59, [%rd78+1];
|
|
and.b16 %rs61, %rs57, 255;
|
|
cvt.rn.f32.u16 %f820, %rs61;
|
|
div.rn.f32 %f821, %f820, 0f437F0000;
|
|
fma.rn.f32 %f822, %f821, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f823, %rs59;
|
|
div.rn.f32 %f824, %f823, 0f437F0000;
|
|
fma.rn.f32 %f825, %f824, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f826, %rs56;
|
|
div.rn.f32 %f827, %f826, 0f437F0000;
|
|
fma.rn.f32 %f828, %f827, 0f40000000, 0fBF800000;
|
|
mul.f32 %f829, %f825, %f825;
|
|
fma.rn.f32 %f830, %f822, %f822, %f829;
|
|
fma.rn.f32 %f831, %f828, %f828, %f830;
|
|
sqrt.rn.f32 %f832, %f831;
|
|
rcp.rn.f32 %f833, %f832;
|
|
mul.f32 %f1087, %f822, %f833;
|
|
mul.f32 %f1088, %f825, %f833;
|
|
mul.f32 %f1089, %f828, %f833;
|
|
|
|
BB0_78:
|
|
mul.f32 %f835, %f1062, %f1088;
|
|
mul.f32 %f836, %f1061, %f1089;
|
|
sub.f32 %f837, %f836, %f835;
|
|
mul.f32 %f838, %f1060, %f1089;
|
|
mul.f32 %f839, %f1062, %f1087;
|
|
sub.f32 %f840, %f839, %f838;
|
|
mul.f32 %f841, %f1061, %f1087;
|
|
mul.f32 %f842, %f1060, %f1088;
|
|
sub.f32 %f843, %f842, %f841;
|
|
setp.lt.u32 %p108, %r19, 16777216;
|
|
selp.f32 %f844, 0fBF800000, 0f3F800000, %p108;
|
|
mul.f32 %f845, %f837, %f844;
|
|
mul.f32 %f846, %f840, %f844;
|
|
mul.f32 %f847, %f843, %f844;
|
|
mul.f32 %f848, %f845, 0fBF3504F3;
|
|
mul.f32 %f849, %f846, 0fBF3504F3;
|
|
mul.f32 %f850, %f847, 0fBF3504F3;
|
|
fma.rn.f32 %f851, %f1087, 0fBED105EC, %f848;
|
|
fma.rn.f32 %f852, %f1088, 0fBED105EC, %f849;
|
|
fma.rn.f32 %f853, %f1089, 0fBED105EC, %f850;
|
|
add.f32 %f854, %f106, %f851;
|
|
add.f32 %f855, %f108, %f852;
|
|
add.f32 %f856, %f110, %f853;
|
|
mul.f32 %f857, %f11, %f109;
|
|
fma.rn.f32 %f858, %f10, %f107, %f857;
|
|
fma.rn.f32 %f859, %f12, %f111, %f858;
|
|
cvt.sat.f32.f32 %f860, %f859;
|
|
mul.f32 %f861, %f97, %f860;
|
|
mul.f32 %f862, %f98, %f860;
|
|
mul.f32 %f863, %f99, %f860;
|
|
mul.f32 %f864, %f11, %f119;
|
|
fma.rn.f32 %f865, %f10, %f118, %f864;
|
|
fma.rn.f32 %f866, %f12, %f120, %f865;
|
|
cvt.sat.f32.f32 %f867, %f866;
|
|
mul.f32 %f868, %f97, %f867;
|
|
mul.f32 %f869, %f98, %f867;
|
|
mul.f32 %f870, %f99, %f867;
|
|
mul.f32 %f871, %f11, %f855;
|
|
fma.rn.f32 %f872, %f10, %f854, %f871;
|
|
fma.rn.f32 %f873, %f12, %f856, %f872;
|
|
cvt.sat.f32.f32 %f874, %f873;
|
|
mul.f32 %f875, %f97, %f874;
|
|
mul.f32 %f876, %f98, %f874;
|
|
mul.f32 %f877, %f99, %f874;
|
|
cvt.sat.f32.f32 %f878, %f29;
|
|
mul.f32 %f879, %f97, %f878;
|
|
mul.f32 %f880, %f98, %f878;
|
|
mul.f32 %f881, %f99, %f878;
|
|
add.f32 %f882, %f861, %f868;
|
|
add.f32 %f883, %f862, %f869;
|
|
add.f32 %f884, %f863, %f870;
|
|
add.f32 %f885, %f882, %f875;
|
|
add.f32 %f886, %f883, %f876;
|
|
add.f32 %f887, %f884, %f877;
|
|
mul.f32 %f888, %f885, 0f3F13CD3A;
|
|
mul.f32 %f889, %f886, 0f3F13CD3A;
|
|
mul.f32 %f890, %f887, 0f3F13CD3A;
|
|
div.rn.f32 %f891, %f879, %f888;
|
|
div.rn.f32 %f892, %f880, %f889;
|
|
div.rn.f32 %f893, %f881, %f890;
|
|
setp.eq.f32 %p109, %f879, 0f00000000;
|
|
selp.f32 %f894, 0f00000000, %f891, %p109;
|
|
setp.eq.f32 %p110, %f880, 0f00000000;
|
|
selp.f32 %f895, 0f00000000, %f892, %p110;
|
|
setp.eq.f32 %p111, %f881, 0f00000000;
|
|
selp.f32 %f896, 0f00000000, %f893, %p111;
|
|
mul.f32 %f127, %f861, %f894;
|
|
mul.f32 %f128, %f862, %f895;
|
|
mul.f32 %f129, %f863, %f896;
|
|
mul.f32 %f130, %f868, %f894;
|
|
mul.f32 %f131, %f869, %f895;
|
|
mul.f32 %f132, %f870, %f896;
|
|
mul.f32 %f133, %f875, %f894;
|
|
mul.f32 %f134, %f876, %f895;
|
|
mul.f32 %f135, %f877, %f896;
|
|
ld.global.u32 %r145, [additive];
|
|
setp.eq.s32 %p112, %r145, 0;
|
|
cvt.u64.u32 %rd8, %r2;
|
|
cvt.u64.u32 %rd9, %r3;
|
|
mov.f32 %f834, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f834;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_80;
|
|
|
|
mov.u64 %rd97, image_RNM0;
|
|
cvta.global.u64 %rd86, %rd97;
|
|
mov.u32 %r149, 8;
|
|
// inline asm
|
|
call (%rd85), _rt_buffer_get_64, (%rd86, %r31, %r149, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs71, %rs72, %rs73, %rs74}, [%rd85];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f897, %rs71;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f898, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f899, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd86, %r31, %r149, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f900, %f127, %f897;
|
|
add.f32 %f901, %f128, %f898;
|
|
add.f32 %f902, %f129, %f899;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f902;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f901;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f900;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd91], {%rs68, %rs69, %rs70, %rs64};
|
|
bra.uni BB0_81;
|
|
|
|
BB0_80:
|
|
mov.u64 %rd104, image_RNM0;
|
|
cvta.global.u64 %rd99, %rd104;
|
|
mov.u32 %r151, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r31, %r151, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f129;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f127;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd98], {%rs75, %rs76, %rs77, %rs64};
|
|
|
|
BB0_81:
|
|
ld.global.u32 %r152, [additive];
|
|
setp.eq.s32 %p113, %r152, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f834;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_83;
|
|
|
|
mov.u64 %rd117, image_RNM1;
|
|
cvta.global.u64 %rd106, %rd117;
|
|
mov.u32 %r156, 8;
|
|
// inline asm
|
|
call (%rd105), _rt_buffer_get_64, (%rd106, %r31, %r156, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs85, %rs86, %rs87, %rs88}, [%rd105];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f907, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f908, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f909, %rs87;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd106, %r31, %r156, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f910, %f130, %f907;
|
|
add.f32 %f911, %f131, %f908;
|
|
add.f32 %f912, %f132, %f909;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs84, %f912;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f911;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f910;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs82, %rs83, %rs84, %rs78};
|
|
bra.uni BB0_84;
|
|
|
|
BB0_83:
|
|
mov.u64 %rd124, image_RNM1;
|
|
cvta.global.u64 %rd119, %rd124;
|
|
mov.u32 %r158, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r31, %r158, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs91, %f132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f130;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd118], {%rs89, %rs90, %rs91, %rs78};
|
|
|
|
BB0_84:
|
|
ld.global.u32 %r159, [additive];
|
|
setp.eq.s32 %p114, %r159, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f834;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_86;
|
|
|
|
mov.u64 %rd137, image_RNM2;
|
|
cvta.global.u64 %rd126, %rd137;
|
|
mov.u32 %r163, 8;
|
|
// inline asm
|
|
call (%rd125), _rt_buffer_get_64, (%rd126, %r31, %r163, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
ld.v4.u16 {%rs99, %rs100, %rs101, %rs102}, [%rd125];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f917, %rs99;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f918, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f919, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd126, %r31, %r163, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
add.f32 %f920, %f133, %f917;
|
|
add.f32 %f921, %f134, %f918;
|
|
add.f32 %f922, %f135, %f919;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f922;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f921;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f920;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs96, %rs97, %rs98, %rs92};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_86:
|
|
mov.u64 %rd144, image_RNM2;
|
|
cvta.global.u64 %rd139, %rd144;
|
|
mov.u32 %r165, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r31, %r165, %rd8, %rd9, %rd19, %rd19);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f135;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f134;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs103, %f133;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd138], {%rs103, %rs104, %rs105, %rs92};
|
|
|
|
BB0_121:
|
|
ret;
|
|
}
|
|
|
|
|