2423 lines
68 KiB
Plaintext
2423 lines
68 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 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 .b8 lightConeDir[12];
|
|
.global .align 8 .b8 lightCone[8];
|
|
.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_typeinfo12lightConeDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo9lightConeE[8] = {82, 97, 121, 0, 8, 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 8 .b8 _ZN21rti_internal_typename12lightConeDirE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename9lightConeE[7] = {102, 108, 111, 97, 116, 50, 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_typeenum12lightConeDirE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9lightConeE = 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_semantic12lightConeDirE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic9lightConeE[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_annotation12lightConeDirE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9lightConeE[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<139>;
|
|
.reg .b16 %rs<221>;
|
|
.reg .f32 %f<984>;
|
|
.reg .b32 %r<251>;
|
|
.reg .b64 %rd<369>;
|
|
|
|
|
|
mov.u64 %rd368, __local_depot0;
|
|
cvta.local.u64 %SP, %rd368;
|
|
ld.global.v2.u32 {%r28, %r29}, [pixelID];
|
|
cvt.u64.u32 %rd12, %r28;
|
|
cvt.u64.u32 %rd13, %r29;
|
|
mov.u64 %rd16, uvnormal;
|
|
cvta.global.u64 %rd11, %rd16;
|
|
mov.u32 %r26, 2;
|
|
mov.u32 %r27, 4;
|
|
mov.u64 %rd15, 0;
|
|
// inline asm
|
|
call (%rd10), _rt_buffer_get_64, (%rd11, %r26, %r27, %rd12, %rd13, %rd15, %rd15);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd10];
|
|
shr.u32 %r32, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r32;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r1;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p7, %rs14, 0;
|
|
mov.f32 %f957, 0f00000000;
|
|
mov.f32 %f958, %f957;
|
|
mov.f32 %f959, %f957;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd10+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f128, %rs17;
|
|
div.rn.f32 %f129, %f128, 0f437F0000;
|
|
fma.rn.f32 %f130, %f129, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f131, %rs15;
|
|
div.rn.f32 %f132, %f131, 0f437F0000;
|
|
fma.rn.f32 %f133, %f132, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f134, %rs12;
|
|
div.rn.f32 %f135, %f134, 0f437F0000;
|
|
fma.rn.f32 %f136, %f135, 0f40000000, 0fBF800000;
|
|
mul.f32 %f137, %f133, %f133;
|
|
fma.rn.f32 %f138, %f130, %f130, %f137;
|
|
fma.rn.f32 %f139, %f136, %f136, %f138;
|
|
sqrt.rn.f32 %f140, %f139;
|
|
rcp.rn.f32 %f141, %f140;
|
|
mul.f32 %f957, %f130, %f141;
|
|
mul.f32 %f958, %f133, %f141;
|
|
mul.f32 %f959, %f136, %f141;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r33, %r34}, [pixelID];
|
|
ld.global.v2.u32 {%r36, %r37}, [tileInfo];
|
|
add.s32 %r2, %r33, %r36;
|
|
add.s32 %r3, %r34, %r37;
|
|
setp.eq.f32 %p8, %f958, 0f00000000;
|
|
setp.eq.f32 %p9, %f957, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f959, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_116;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_116:
|
|
ld.global.u32 %r250, [imageEnabled];
|
|
and.b32 %r199, %r250, 1;
|
|
setp.eq.b32 %p131, %r199, 1;
|
|
@!%p131 bra BB0_118;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_117:
|
|
cvt.u64.u32 %rd256, %r2;
|
|
cvt.u64.u32 %rd257, %r3;
|
|
mov.u64 %rd260, image;
|
|
cvta.global.u64 %rd255, %rd260;
|
|
// inline asm
|
|
call (%rd254), _rt_buffer_get_64, (%rd255, %r26, %r27, %rd256, %rd257, %rd15, %rd15);
|
|
// inline asm
|
|
mov.u16 %rs153, 0;
|
|
st.v4.u8 [%rd254], {%rs153, %rs153, %rs153, %rs153};
|
|
ld.global.u32 %r250, [imageEnabled];
|
|
|
|
BB0_118:
|
|
and.b32 %r202, %r250, 8;
|
|
setp.eq.s32 %p132, %r202, 0;
|
|
@%p132 bra BB0_120;
|
|
|
|
cvt.u64.u32 %rd264, %r3;
|
|
cvt.u64.u32 %rd263, %r2;
|
|
mov.u64 %rd267, image_Mask;
|
|
cvta.global.u64 %rd262, %rd267;
|
|
// inline asm
|
|
call (%rd261), _rt_buffer_get_64, (%rd262, %r26, %r26, %rd263, %rd264, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f866, 0f00000000;
|
|
cvt.rzi.u32.f32 %r205, %f866;
|
|
cvt.u16.u32 %rs154, %r205;
|
|
mov.u16 %rs155, 0;
|
|
st.v2.u8 [%rd261], {%rs154, %rs155};
|
|
ld.global.u32 %r250, [imageEnabled];
|
|
|
|
BB0_120:
|
|
cvt.u64.u32 %rd8, %r2;
|
|
cvt.u64.u32 %rd9, %r3;
|
|
and.b32 %r206, %r250, 4;
|
|
setp.eq.s32 %p133, %r206, 0;
|
|
@%p133 bra BB0_124;
|
|
|
|
ld.global.u32 %r207, [additive];
|
|
setp.eq.s32 %p134, %r207, 0;
|
|
@%p134 bra BB0_123;
|
|
|
|
mov.u64 %rd280, image_HDR;
|
|
cvta.global.u64 %rd269, %rd280;
|
|
mov.u32 %r211, 8;
|
|
// inline asm
|
|
call (%rd268), _rt_buffer_get_64, (%rd269, %r26, %r211, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs162, %rs163, %rs164, %rs165}, [%rd268];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f867, %rs162;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f868, %rs163;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f869, %rs164;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd274), _rt_buffer_get_64, (%rd269, %r26, %r211, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f870, %f867, 0f00000000;
|
|
add.f32 %f871, %f868, 0f00000000;
|
|
add.f32 %f872, %f869, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs161, %f872;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f871;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f870;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs166, 0;
|
|
st.v4.u16 [%rd274], {%rs159, %rs160, %rs161, %rs166};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r43, %r44}, [pixelID];
|
|
cvt.u64.u32 %rd19, %r43;
|
|
cvt.u64.u32 %rd20, %r44;
|
|
mov.u64 %rd23, uvpos;
|
|
cvta.global.u64 %rd18, %rd23;
|
|
mov.u32 %r42, 12;
|
|
// inline asm
|
|
call (%rd17), _rt_buffer_get_64, (%rd18, %r26, %r42, %rd19, %rd20, %rd15, %rd15);
|
|
// inline asm
|
|
ld.global.f32 %f968, [lightPos];
|
|
ld.f32 %f12, [%rd17+8];
|
|
ld.f32 %f10, [%rd17+4];
|
|
ld.f32 %f8, [%rd17];
|
|
sub.f32 %f144, %f968, %f8;
|
|
ld.global.f32 %f967, [lightPos+4];
|
|
sub.f32 %f145, %f967, %f10;
|
|
ld.global.f32 %f966, [lightPos+8];
|
|
sub.f32 %f146, %f966, %f12;
|
|
mul.f32 %f147, %f145, %f145;
|
|
fma.rn.f32 %f148, %f144, %f144, %f147;
|
|
fma.rn.f32 %f149, %f146, %f146, %f148;
|
|
sqrt.rn.f32 %f150, %f149;
|
|
rcp.rn.f32 %f151, %f150;
|
|
mul.f32 %f13, %f144, %f151;
|
|
mul.f32 %f14, %f145, %f151;
|
|
mul.f32 %f15, %f146, %f151;
|
|
ld.global.f32 %f152, [lightFalloffFakeDistanceMult];
|
|
mul.f32 %f16, %f150, %f152;
|
|
ld.global.f32 %f153, [lightInvCutoff];
|
|
mul.f32 %f17, %f150, %f153;
|
|
mov.f32 %f157, 0f40800000;
|
|
abs.f32 %f19, %f17;
|
|
setp.lt.f32 %p13, %f19, 0f00800000;
|
|
mul.f32 %f159, %f19, 0f4B800000;
|
|
selp.f32 %f160, 0fC3170000, 0fC2FE0000, %p13;
|
|
selp.f32 %f161, %f159, %f19, %p13;
|
|
mov.b32 %r47, %f161;
|
|
and.b32 %r48, %r47, 8388607;
|
|
or.b32 %r49, %r48, 1065353216;
|
|
mov.b32 %f162, %r49;
|
|
shr.u32 %r50, %r47, 23;
|
|
cvt.rn.f32.u32 %f163, %r50;
|
|
add.f32 %f164, %f160, %f163;
|
|
setp.gt.f32 %p14, %f162, 0f3FB504F3;
|
|
mul.f32 %f165, %f162, 0f3F000000;
|
|
add.f32 %f166, %f164, 0f3F800000;
|
|
selp.f32 %f167, %f165, %f162, %p14;
|
|
selp.f32 %f168, %f166, %f164, %p14;
|
|
add.f32 %f169, %f167, 0fBF800000;
|
|
add.f32 %f143, %f167, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f142,%f143;
|
|
// inline asm
|
|
add.f32 %f170, %f169, %f169;
|
|
mul.f32 %f171, %f142, %f170;
|
|
mul.f32 %f172, %f171, %f171;
|
|
mov.f32 %f173, 0f3C4CAF63;
|
|
mov.f32 %f174, 0f3B18F0FE;
|
|
fma.rn.f32 %f175, %f174, %f172, %f173;
|
|
mov.f32 %f176, 0f3DAAAABD;
|
|
fma.rn.f32 %f177, %f175, %f172, %f176;
|
|
mul.rn.f32 %f178, %f177, %f172;
|
|
mul.rn.f32 %f179, %f178, %f171;
|
|
sub.f32 %f180, %f169, %f171;
|
|
neg.f32 %f181, %f171;
|
|
add.f32 %f182, %f180, %f180;
|
|
fma.rn.f32 %f183, %f181, %f169, %f182;
|
|
mul.rn.f32 %f184, %f142, %f183;
|
|
add.f32 %f185, %f179, %f171;
|
|
sub.f32 %f186, %f171, %f185;
|
|
add.f32 %f187, %f179, %f186;
|
|
add.f32 %f188, %f184, %f187;
|
|
add.f32 %f189, %f185, %f188;
|
|
sub.f32 %f190, %f185, %f189;
|
|
add.f32 %f191, %f188, %f190;
|
|
mov.f32 %f192, 0f3F317200;
|
|
mul.rn.f32 %f193, %f168, %f192;
|
|
mov.f32 %f194, 0f35BFBE8E;
|
|
mul.rn.f32 %f195, %f168, %f194;
|
|
add.f32 %f196, %f193, %f189;
|
|
sub.f32 %f197, %f193, %f196;
|
|
add.f32 %f198, %f189, %f197;
|
|
add.f32 %f199, %f191, %f198;
|
|
add.f32 %f200, %f195, %f199;
|
|
add.f32 %f201, %f196, %f200;
|
|
sub.f32 %f202, %f196, %f201;
|
|
add.f32 %f203, %f200, %f202;
|
|
mul.rn.f32 %f204, %f157, %f201;
|
|
neg.f32 %f205, %f204;
|
|
fma.rn.f32 %f206, %f157, %f201, %f205;
|
|
fma.rn.f32 %f207, %f157, %f203, %f206;
|
|
mov.f32 %f208, 0f00000000;
|
|
fma.rn.f32 %f209, %f208, %f201, %f207;
|
|
add.rn.f32 %f210, %f204, %f209;
|
|
neg.f32 %f211, %f210;
|
|
add.rn.f32 %f212, %f204, %f211;
|
|
add.rn.f32 %f213, %f212, %f209;
|
|
mov.b32 %r51, %f210;
|
|
setp.eq.s32 %p15, %r51, 1118925336;
|
|
add.s32 %r52, %r51, -1;
|
|
mov.b32 %f214, %r52;
|
|
add.f32 %f215, %f213, 0f37000000;
|
|
selp.f32 %f216, %f214, %f210, %p15;
|
|
selp.f32 %f20, %f215, %f213, %p15;
|
|
mul.f32 %f217, %f216, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f218, %f217;
|
|
mov.f32 %f219, 0fBF317200;
|
|
fma.rn.f32 %f220, %f218, %f219, %f216;
|
|
mov.f32 %f221, 0fB5BFBE8E;
|
|
fma.rn.f32 %f222, %f218, %f221, %f220;
|
|
mul.f32 %f223, %f222, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f224, %f223;
|
|
add.f32 %f225, %f218, 0f00000000;
|
|
ex2.approx.f32 %f226, %f225;
|
|
mul.f32 %f227, %f224, %f226;
|
|
setp.lt.f32 %p16, %f216, 0fC2D20000;
|
|
selp.f32 %f228, 0f00000000, %f227, %p16;
|
|
setp.gt.f32 %p17, %f216, 0f42D20000;
|
|
selp.f32 %f960, 0f7F800000, %f228, %p17;
|
|
setp.eq.f32 %p18, %f960, 0f7F800000;
|
|
@%p18 bra BB0_5;
|
|
|
|
fma.rn.f32 %f960, %f960, %f20, %f960;
|
|
|
|
BB0_5:
|
|
mov.f32 %f907, 0f40000000;
|
|
cvt.rzi.f32.f32 %f906, %f907;
|
|
add.f32 %f905, %f906, %f906;
|
|
mov.f32 %f904, 0f40800000;
|
|
sub.f32 %f903, %f904, %f905;
|
|
abs.f32 %f902, %f903;
|
|
setp.lt.f32 %p19, %f17, 0f00000000;
|
|
setp.eq.f32 %p20, %f902, 0f3F800000;
|
|
and.pred %p1, %p19, %p20;
|
|
mov.b32 %r53, %f960;
|
|
xor.b32 %r54, %r53, -2147483648;
|
|
mov.b32 %f229, %r54;
|
|
selp.f32 %f962, %f229, %f960, %p1;
|
|
setp.eq.f32 %p21, %f17, 0f00000000;
|
|
@%p21 bra BB0_8;
|
|
bra.uni BB0_6;
|
|
|
|
BB0_8:
|
|
add.f32 %f232, %f17, %f17;
|
|
selp.f32 %f962, %f232, 0f00000000, %p20;
|
|
bra.uni BB0_9;
|
|
|
|
BB0_6:
|
|
setp.geu.f32 %p22, %f17, 0f00000000;
|
|
@%p22 bra BB0_9;
|
|
|
|
mov.f32 %f948, 0f40800000;
|
|
cvt.rzi.f32.f32 %f231, %f948;
|
|
setp.neu.f32 %p23, %f231, 0f40800000;
|
|
selp.f32 %f962, 0f7FFFFFFF, %f962, %p23;
|
|
|
|
BB0_9:
|
|
abs.f32 %f908, %f17;
|
|
add.f32 %f233, %f908, 0f40800000;
|
|
mov.b32 %r55, %f233;
|
|
setp.lt.s32 %p25, %r55, 2139095040;
|
|
@%p25 bra BB0_14;
|
|
|
|
abs.f32 %f946, %f17;
|
|
setp.gtu.f32 %p26, %f946, 0f7F800000;
|
|
@%p26 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f962, %f17, 0f40800000;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
abs.f32 %f947, %f17;
|
|
setp.neu.f32 %p27, %f947, 0f7F800000;
|
|
@%p27 bra BB0_14;
|
|
|
|
selp.f32 %f962, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_14:
|
|
mov.f32 %f916, 0fB5BFBE8E;
|
|
mov.f32 %f915, 0fBF317200;
|
|
mov.f32 %f914, 0f00000000;
|
|
mov.f32 %f913, 0f35BFBE8E;
|
|
mov.f32 %f912, 0f3F317200;
|
|
mov.f32 %f911, 0f3DAAAABD;
|
|
mov.f32 %f910, 0f3C4CAF63;
|
|
mov.f32 %f909, 0f3B18F0FE;
|
|
mov.f32 %f236, 0f3F800000;
|
|
sub.f32 %f237, %f236, %f962;
|
|
setp.eq.f32 %p28, %f17, 0f3F800000;
|
|
selp.f32 %f238, 0f00000000, %f237, %p28;
|
|
cvt.sat.f32.f32 %f239, %f238;
|
|
ld.global.f32 %f240, [lightFalloffMinRadiusSq];
|
|
fma.rn.f32 %f241, %f16, %f16, %f240;
|
|
div.rn.f32 %f31, %f239, %f241;
|
|
mul.f32 %f242, %f958, %f14;
|
|
fma.rn.f32 %f243, %f957, %f13, %f242;
|
|
fma.rn.f32 %f244, %f959, %f15, %f243;
|
|
ld.global.u32 %r248, [imageEnabled];
|
|
and.b32 %r56, %r248, 32;
|
|
ld.global.u32 %r57, [ignoreNormal];
|
|
or.b32 %r58, %r56, %r57;
|
|
setp.eq.s32 %p29, %r58, 0;
|
|
selp.f32 %f32, %f244, 0f3F800000, %p29;
|
|
fma.rn.f32 %f245, %f32, 0f3F000000, 0f3F000000;
|
|
cvt.sat.f32.f32 %f246, %f245;
|
|
add.f32 %f33, %f246, %f246;
|
|
mov.f32 %f250, 0f41A00000;
|
|
abs.f32 %f35, %f33;
|
|
setp.lt.f32 %p30, %f35, 0f00800000;
|
|
mul.f32 %f252, %f35, 0f4B800000;
|
|
selp.f32 %f253, 0fC3170000, 0fC2FE0000, %p30;
|
|
selp.f32 %f254, %f252, %f35, %p30;
|
|
mov.b32 %r59, %f254;
|
|
and.b32 %r60, %r59, 8388607;
|
|
or.b32 %r61, %r60, 1065353216;
|
|
mov.b32 %f255, %r61;
|
|
shr.u32 %r62, %r59, 23;
|
|
cvt.rn.f32.u32 %f256, %r62;
|
|
add.f32 %f257, %f253, %f256;
|
|
setp.gt.f32 %p31, %f255, 0f3FB504F3;
|
|
mul.f32 %f258, %f255, 0f3F000000;
|
|
add.f32 %f259, %f257, 0f3F800000;
|
|
selp.f32 %f260, %f258, %f255, %p31;
|
|
selp.f32 %f261, %f259, %f257, %p31;
|
|
add.f32 %f262, %f260, 0fBF800000;
|
|
add.f32 %f235, %f260, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f234,%f235;
|
|
// inline asm
|
|
add.f32 %f263, %f262, %f262;
|
|
mul.f32 %f264, %f234, %f263;
|
|
mul.f32 %f265, %f264, %f264;
|
|
fma.rn.f32 %f268, %f909, %f265, %f910;
|
|
fma.rn.f32 %f270, %f268, %f265, %f911;
|
|
mul.rn.f32 %f271, %f270, %f265;
|
|
mul.rn.f32 %f272, %f271, %f264;
|
|
sub.f32 %f273, %f262, %f264;
|
|
neg.f32 %f274, %f264;
|
|
add.f32 %f275, %f273, %f273;
|
|
fma.rn.f32 %f276, %f274, %f262, %f275;
|
|
mul.rn.f32 %f277, %f234, %f276;
|
|
add.f32 %f278, %f272, %f264;
|
|
sub.f32 %f279, %f264, %f278;
|
|
add.f32 %f280, %f272, %f279;
|
|
add.f32 %f281, %f277, %f280;
|
|
add.f32 %f282, %f278, %f281;
|
|
sub.f32 %f283, %f278, %f282;
|
|
add.f32 %f284, %f281, %f283;
|
|
mul.rn.f32 %f286, %f261, %f912;
|
|
mul.rn.f32 %f288, %f261, %f913;
|
|
add.f32 %f289, %f286, %f282;
|
|
sub.f32 %f290, %f286, %f289;
|
|
add.f32 %f291, %f282, %f290;
|
|
add.f32 %f292, %f284, %f291;
|
|
add.f32 %f293, %f288, %f292;
|
|
add.f32 %f294, %f289, %f293;
|
|
sub.f32 %f295, %f289, %f294;
|
|
add.f32 %f296, %f293, %f295;
|
|
mul.rn.f32 %f297, %f250, %f294;
|
|
neg.f32 %f298, %f297;
|
|
fma.rn.f32 %f299, %f250, %f294, %f298;
|
|
fma.rn.f32 %f300, %f250, %f296, %f299;
|
|
fma.rn.f32 %f302, %f914, %f294, %f300;
|
|
add.rn.f32 %f303, %f297, %f302;
|
|
neg.f32 %f304, %f303;
|
|
add.rn.f32 %f305, %f297, %f304;
|
|
add.rn.f32 %f306, %f305, %f302;
|
|
mov.b32 %r63, %f303;
|
|
setp.eq.s32 %p32, %r63, 1118925336;
|
|
add.s32 %r64, %r63, -1;
|
|
mov.b32 %f307, %r64;
|
|
add.f32 %f308, %f306, 0f37000000;
|
|
selp.f32 %f309, %f307, %f303, %p32;
|
|
selp.f32 %f36, %f308, %f306, %p32;
|
|
mul.f32 %f310, %f309, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f311, %f310;
|
|
fma.rn.f32 %f313, %f311, %f915, %f309;
|
|
fma.rn.f32 %f315, %f311, %f916, %f313;
|
|
mul.f32 %f316, %f315, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f317, %f316;
|
|
add.f32 %f318, %f311, 0f00000000;
|
|
ex2.approx.f32 %f319, %f318;
|
|
mul.f32 %f320, %f317, %f319;
|
|
setp.lt.f32 %p33, %f309, 0fC2D20000;
|
|
selp.f32 %f321, 0f00000000, %f320, %p33;
|
|
setp.gt.f32 %p34, %f309, 0f42D20000;
|
|
selp.f32 %f963, 0f7F800000, %f321, %p34;
|
|
setp.eq.f32 %p35, %f963, 0f7F800000;
|
|
@%p35 bra BB0_16;
|
|
|
|
fma.rn.f32 %f963, %f963, %f36, %f963;
|
|
|
|
BB0_16:
|
|
mov.f32 %f954, 0f41200000;
|
|
cvt.rzi.f32.f32 %f953, %f954;
|
|
add.f32 %f952, %f953, %f953;
|
|
mov.f32 %f951, 0f41A00000;
|
|
sub.f32 %f950, %f951, %f952;
|
|
abs.f32 %f949, %f950;
|
|
setp.lt.f32 %p36, %f33, 0f00000000;
|
|
setp.eq.f32 %p37, %f949, 0f3F800000;
|
|
and.pred %p2, %p36, %p37;
|
|
mov.b32 %r65, %f963;
|
|
xor.b32 %r66, %r65, -2147483648;
|
|
mov.b32 %f322, %r66;
|
|
selp.f32 %f965, %f322, %f963, %p2;
|
|
setp.eq.f32 %p38, %f33, 0f00000000;
|
|
@%p38 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f325, %f33, %f33;
|
|
selp.f32 %f965, %f325, 0f00000000, %p37;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.geu.f32 %p39, %f33, 0f00000000;
|
|
@%p39 bra BB0_20;
|
|
|
|
mov.f32 %f956, 0f41A00000;
|
|
cvt.rzi.f32.f32 %f324, %f956;
|
|
setp.neu.f32 %p40, %f324, 0f41A00000;
|
|
selp.f32 %f965, 0f7FFFFFFF, %f965, %p40;
|
|
|
|
BB0_20:
|
|
add.f32 %f326, %f35, 0f41A00000;
|
|
mov.b32 %r67, %f326;
|
|
setp.lt.s32 %p42, %r67, 2139095040;
|
|
@%p42 bra BB0_25;
|
|
|
|
setp.gtu.f32 %p43, %f35, 0f7F800000;
|
|
@%p43 bra BB0_24;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_24:
|
|
add.f32 %f965, %f33, 0f41A00000;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_22:
|
|
setp.neu.f32 %p44, %f35, 0f7F800000;
|
|
@%p44 bra BB0_25;
|
|
|
|
selp.f32 %f965, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_25:
|
|
setp.eq.f32 %p45, %f33, 0f3F800000;
|
|
selp.f32 %f327, 0f3F800000, %f965, %p45;
|
|
cvt.sat.f32.f32 %f328, %f327;
|
|
mul.f32 %f329, %f31, %f328;
|
|
mul.f32 %f330, %f32, 0f40800000;
|
|
cvt.sat.f32.f32 %f47, %f330;
|
|
ld.global.f32 %f331, [lightConeDir];
|
|
mul.f32 %f332, %f13, %f331;
|
|
ld.global.f32 %f333, [lightConeDir+4];
|
|
mul.f32 %f334, %f14, %f333;
|
|
neg.f32 %f335, %f334;
|
|
sub.f32 %f336, %f335, %f332;
|
|
ld.global.f32 %f337, [lightConeDir+8];
|
|
mul.f32 %f338, %f15, %f337;
|
|
sub.f32 %f339, %f336, %f338;
|
|
ld.global.v2.f32 {%f340, %f341}, [lightCone];
|
|
fma.rn.f32 %f344, %f340, %f339, %f341;
|
|
cvt.sat.f32.f32 %f345, %f344;
|
|
mul.f32 %f346, %f345, %f345;
|
|
mul.f32 %f48, %f329, %f346;
|
|
mul.f32 %f347, %f47, %f48;
|
|
ld.global.f32 %f348, [lightColor+4];
|
|
ld.global.f32 %f349, [lightColor];
|
|
max.f32 %f350, %f349, %f348;
|
|
ld.global.f32 %f351, [lightColor+8];
|
|
max.f32 %f352, %f350, %f351;
|
|
mul.f32 %f353, %f347, %f352;
|
|
setp.lt.f32 %p46, %f353, 0f3727C5AC;
|
|
@%p46 bra BB0_96;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_96:
|
|
and.b32 %r156, %r248, 1;
|
|
setp.eq.b32 %p123, %r156, 1;
|
|
@!%p123 bra BB0_98;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_97:
|
|
cvt.u64.u32 %rd142, %r2;
|
|
cvt.u64.u32 %rd143, %r3;
|
|
mov.u64 %rd146, image;
|
|
cvta.global.u64 %rd141, %rd146;
|
|
// inline asm
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r26, %r27, %rd142, %rd143, %rd15, %rd15);
|
|
// inline asm
|
|
mov.u16 %rs89, 1;
|
|
mov.u16 %rs90, 0;
|
|
st.v4.u8 [%rd140], {%rs90, %rs90, %rs90, %rs89};
|
|
ld.global.u32 %r248, [imageEnabled];
|
|
|
|
BB0_98:
|
|
and.b32 %r159, %r248, 8;
|
|
setp.eq.s32 %p124, %r159, 0;
|
|
@%p124 bra BB0_100;
|
|
|
|
cvt.u64.u32 %rd150, %r3;
|
|
cvt.u64.u32 %rd149, %r2;
|
|
mov.u64 %rd153, image_Mask;
|
|
cvta.global.u64 %rd148, %rd153;
|
|
// inline asm
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r26, %r26, %rd149, %rd150, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f823, 0f00000000;
|
|
cvt.rzi.u32.f32 %r162, %f823;
|
|
cvt.u16.u32 %rs91, %r162;
|
|
mov.u16 %rs92, 255;
|
|
st.v2.u8 [%rd147], {%rs91, %rs92};
|
|
ld.global.u32 %r248, [imageEnabled];
|
|
|
|
BB0_100:
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
and.b32 %r163, %r248, 4;
|
|
setp.eq.s32 %p125, %r163, 0;
|
|
@%p125 bra BB0_104;
|
|
|
|
ld.global.u32 %r164, [additive];
|
|
setp.eq.s32 %p126, %r164, 0;
|
|
mov.f32 %f824, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f824;}
|
|
|
|
// inline asm
|
|
@%p126 bra BB0_103;
|
|
|
|
mov.u64 %rd166, image_HDR;
|
|
cvta.global.u64 %rd155, %rd166;
|
|
mov.u32 %r168, 8;
|
|
// inline asm
|
|
call (%rd154), _rt_buffer_get_64, (%rd155, %r26, %r168, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd154];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f825, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f826, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f827, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd160), _rt_buffer_get_64, (%rd155, %r26, %r168, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f828, %f825, 0f00000000;
|
|
add.f32 %f829, %f826, 0f00000000;
|
|
add.f32 %f830, %f827, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f830;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f829;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f828;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd160], {%rs97, %rs98, %rs99, %rs93};
|
|
bra.uni BB0_104;
|
|
|
|
BB0_26:
|
|
mov.f32 %f969, 0f00000000;
|
|
mul.f32 %f355, %f8, 0f3456BF95;
|
|
abs.f32 %f356, %f957;
|
|
div.rn.f32 %f357, %f355, %f356;
|
|
abs.f32 %f358, %f958;
|
|
mul.f32 %f359, %f10, 0f3456BF95;
|
|
div.rn.f32 %f360, %f359, %f358;
|
|
abs.f32 %f361, %f959;
|
|
mul.f32 %f362, %f12, 0f3456BF95;
|
|
div.rn.f32 %f363, %f362, %f361;
|
|
abs.f32 %f364, %f357;
|
|
abs.f32 %f365, %f360;
|
|
abs.f32 %f366, %f363;
|
|
mov.f32 %f367, 0f38D1B717;
|
|
max.f32 %f368, %f364, %f367;
|
|
max.f32 %f369, %f365, %f367;
|
|
max.f32 %f370, %f366, %f367;
|
|
fma.rn.f32 %f49, %f957, %f368, %f8;
|
|
fma.rn.f32 %f50, %f958, %f369, %f10;
|
|
fma.rn.f32 %f51, %f959, %f370, %f12;
|
|
ld.global.u32 %r244, [samples];
|
|
setp.lt.s32 %p47, %r244, 1;
|
|
@%p47 bra BB0_29;
|
|
|
|
mul.f32 %f372, %f49, 0f3456BF95;
|
|
abs.f32 %f373, %f372;
|
|
mul.f32 %f374, %f50, 0f3456BF95;
|
|
abs.f32 %f375, %f374;
|
|
mul.f32 %f376, %f51, 0f3456BF95;
|
|
abs.f32 %f377, %f376;
|
|
max.f32 %f378, %f373, %f375;
|
|
max.f32 %f379, %f378, %f377;
|
|
max.f32 %f52, %f379, %f367;
|
|
add.u64 %rd24, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd24;
|
|
mov.f32 %f969, 0f00000000;
|
|
mov.u32 %r243, 0;
|
|
bra.uni BB0_28;
|
|
|
|
BB0_95:
|
|
ld.global.f32 %f968, [lightPos];
|
|
ld.global.f32 %f967, [lightPos+4];
|
|
ld.global.f32 %f966, [lightPos+8];
|
|
ld.global.u32 %r248, [imageEnabled];
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f389, %r243;
|
|
mul.f32 %f390, %f389, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f391, %f390;
|
|
sub.f32 %f392, %f390, %f391;
|
|
mul.f32 %f393, %f389, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f394, %f393;
|
|
sub.f32 %f395, %f393, %f394;
|
|
mul.f32 %f396, %f389, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f397, %f396;
|
|
sub.f32 %f398, %f396, %f397;
|
|
add.f32 %f399, %f395, 0f4199851F;
|
|
add.f32 %f400, %f398, 0f4199851F;
|
|
add.f32 %f401, %f392, 0f4199851F;
|
|
mul.f32 %f402, %f395, %f400;
|
|
fma.rn.f32 %f403, %f392, %f399, %f402;
|
|
fma.rn.f32 %f404, %f401, %f398, %f403;
|
|
add.f32 %f405, %f392, %f404;
|
|
add.f32 %f406, %f395, %f404;
|
|
add.f32 %f407, %f398, %f404;
|
|
add.f32 %f408, %f405, %f406;
|
|
mul.f32 %f409, %f407, %f408;
|
|
cvt.rmi.f32.f32 %f410, %f409;
|
|
sub.f32 %f411, %f409, %f410;
|
|
add.f32 %f412, %f405, %f407;
|
|
mul.f32 %f413, %f406, %f412;
|
|
cvt.rmi.f32.f32 %f414, %f413;
|
|
sub.f32 %f415, %f413, %f414;
|
|
add.f32 %f416, %f406, %f407;
|
|
mul.f32 %f417, %f405, %f416;
|
|
cvt.rmi.f32.f32 %f418, %f417;
|
|
sub.f32 %f419, %f417, %f418;
|
|
fma.rn.f32 %f420, %f411, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f421, %f415, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f422, %f419, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f423, [lightRadius];
|
|
fma.rn.f32 %f424, %f423, %f420, %f968;
|
|
fma.rn.f32 %f425, %f423, %f421, %f967;
|
|
fma.rn.f32 %f426, %f423, %f422, %f966;
|
|
sub.f32 %f427, %f424, %f8;
|
|
sub.f32 %f428, %f425, %f10;
|
|
sub.f32 %f429, %f426, %f12;
|
|
mul.f32 %f430, %f428, %f428;
|
|
fma.rn.f32 %f431, %f427, %f427, %f430;
|
|
fma.rn.f32 %f432, %f429, %f429, %f431;
|
|
sqrt.rn.f32 %f388, %f432;
|
|
rcp.rn.f32 %f433, %f388;
|
|
mul.f32 %f384, %f433, %f427;
|
|
mul.f32 %f385, %f433, %f428;
|
|
mul.f32 %f386, %f433, %f429;
|
|
and.b32 %r72, %r248, 32;
|
|
setp.eq.s32 %p48, %r72, 0;
|
|
selp.f32 %f434, 0f3F800000, 0f41200000, %p48;
|
|
mul.f32 %f387, %f434, %f52;
|
|
mov.u32 %r73, 1065353216;
|
|
st.local.u32 [%rd2], %r73;
|
|
ld.global.u32 %r69, [root];
|
|
mov.u32 %r70, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r69, %f49, %f50, %f51, %f384, %f385, %f386, %r70, %f387, %f388, %rd24, %r27);
|
|
// inline asm
|
|
ld.local.f32 %f435, [%rd2];
|
|
add.f32 %f969, %f969, %f435;
|
|
ld.global.u32 %r244, [samples];
|
|
add.s32 %r243, %r243, 1;
|
|
setp.lt.s32 %p49, %r243, %r244;
|
|
@%p49 bra BB0_95;
|
|
|
|
BB0_29:
|
|
mov.f32 %f971, 0f3F800000;
|
|
setp.eq.s32 %p50, %r244, 0;
|
|
@%p50 bra BB0_31;
|
|
|
|
cvt.rn.f32.s32 %f437, %r244;
|
|
div.rn.f32 %f971, %f969, %f437;
|
|
|
|
BB0_31:
|
|
cvt.sat.f32.f32 %f438, %f32;
|
|
mul.f32 %f439, %f48, %f438;
|
|
mul.f32 %f440, %f971, %f439;
|
|
ld.global.f32 %f441, [lightColor];
|
|
mul.f32 %f61, %f441, %f440;
|
|
ld.global.f32 %f442, [lightColor+4];
|
|
mul.f32 %f62, %f442, %f440;
|
|
ld.global.f32 %f443, [lightColor+8];
|
|
mul.f32 %f63, %f440, %f443;
|
|
ld.global.u32 %r246, [imageEnabled];
|
|
and.b32 %r74, %r246, 8;
|
|
setp.eq.s32 %p51, %r74, 0;
|
|
@%p51 bra BB0_44;
|
|
|
|
mov.f32 %f924, 0fB5BFBE8E;
|
|
mov.f32 %f923, 0fBF317200;
|
|
mov.f32 %f922, 0f35BFBE8E;
|
|
mov.f32 %f921, 0f3F317200;
|
|
mov.f32 %f920, 0f3DAAAABD;
|
|
mov.f32 %f919, 0f3C4CAF63;
|
|
mov.f32 %f918, 0f3B18F0FE;
|
|
cvt.u64.u32 %rd28, %r2;
|
|
cvt.u64.u32 %rd29, %r3;
|
|
mov.u64 %rd32, image_Mask;
|
|
cvta.global.u64 %rd27, %rd32;
|
|
// inline asm
|
|
call (%rd26), _rt_buffer_get_64, (%rd27, %r26, %r26, %rd28, %rd29, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f446, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f447, %f446;
|
|
fma.rn.f32 %f448, %f447, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f64, %f448;
|
|
abs.f32 %f65, %f971;
|
|
setp.lt.f32 %p52, %f65, 0f00800000;
|
|
mul.f32 %f449, %f65, 0f4B800000;
|
|
selp.f32 %f450, 0fC3170000, 0fC2FE0000, %p52;
|
|
selp.f32 %f451, %f449, %f65, %p52;
|
|
mov.b32 %r77, %f451;
|
|
and.b32 %r78, %r77, 8388607;
|
|
or.b32 %r79, %r78, 1065353216;
|
|
mov.b32 %f452, %r79;
|
|
shr.u32 %r80, %r77, 23;
|
|
cvt.rn.f32.u32 %f453, %r80;
|
|
add.f32 %f454, %f450, %f453;
|
|
setp.gt.f32 %p53, %f452, 0f3FB504F3;
|
|
mul.f32 %f455, %f452, 0f3F000000;
|
|
add.f32 %f456, %f454, 0f3F800000;
|
|
selp.f32 %f457, %f455, %f452, %p53;
|
|
selp.f32 %f458, %f456, %f454, %p53;
|
|
add.f32 %f459, %f457, 0fBF800000;
|
|
add.f32 %f445, %f457, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f444,%f445;
|
|
// inline asm
|
|
add.f32 %f460, %f459, %f459;
|
|
mul.f32 %f461, %f444, %f460;
|
|
mul.f32 %f462, %f461, %f461;
|
|
fma.rn.f32 %f465, %f918, %f462, %f919;
|
|
fma.rn.f32 %f467, %f465, %f462, %f920;
|
|
mul.rn.f32 %f468, %f467, %f462;
|
|
mul.rn.f32 %f469, %f468, %f461;
|
|
sub.f32 %f470, %f459, %f461;
|
|
neg.f32 %f471, %f461;
|
|
add.f32 %f472, %f470, %f470;
|
|
fma.rn.f32 %f473, %f471, %f459, %f472;
|
|
mul.rn.f32 %f474, %f444, %f473;
|
|
add.f32 %f475, %f469, %f461;
|
|
sub.f32 %f476, %f461, %f475;
|
|
add.f32 %f477, %f469, %f476;
|
|
add.f32 %f478, %f474, %f477;
|
|
add.f32 %f479, %f475, %f478;
|
|
sub.f32 %f480, %f475, %f479;
|
|
add.f32 %f481, %f478, %f480;
|
|
mul.rn.f32 %f483, %f458, %f921;
|
|
mul.rn.f32 %f485, %f458, %f922;
|
|
add.f32 %f486, %f483, %f479;
|
|
sub.f32 %f487, %f483, %f486;
|
|
add.f32 %f488, %f479, %f487;
|
|
add.f32 %f489, %f481, %f488;
|
|
add.f32 %f490, %f485, %f489;
|
|
add.f32 %f491, %f486, %f490;
|
|
sub.f32 %f492, %f486, %f491;
|
|
add.f32 %f493, %f490, %f492;
|
|
mov.f32 %f494, 0f3EE8BA2E;
|
|
mul.rn.f32 %f495, %f494, %f491;
|
|
neg.f32 %f496, %f495;
|
|
fma.rn.f32 %f497, %f494, %f491, %f496;
|
|
fma.rn.f32 %f498, %f494, %f493, %f497;
|
|
mov.f32 %f499, 0f00000000;
|
|
fma.rn.f32 %f500, %f499, %f491, %f498;
|
|
add.rn.f32 %f501, %f495, %f500;
|
|
neg.f32 %f502, %f501;
|
|
add.rn.f32 %f503, %f495, %f502;
|
|
add.rn.f32 %f504, %f503, %f500;
|
|
mov.b32 %r81, %f501;
|
|
setp.eq.s32 %p54, %r81, 1118925336;
|
|
add.s32 %r82, %r81, -1;
|
|
mov.b32 %f505, %r82;
|
|
add.f32 %f506, %f504, 0f37000000;
|
|
selp.f32 %f507, %f505, %f501, %p54;
|
|
selp.f32 %f66, %f506, %f504, %p54;
|
|
mul.f32 %f508, %f507, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f509, %f508;
|
|
fma.rn.f32 %f511, %f509, %f923, %f507;
|
|
fma.rn.f32 %f513, %f509, %f924, %f511;
|
|
mul.f32 %f514, %f513, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f515, %f514;
|
|
add.f32 %f516, %f509, 0f00000000;
|
|
ex2.approx.f32 %f517, %f516;
|
|
mul.f32 %f518, %f515, %f517;
|
|
setp.lt.f32 %p55, %f507, 0fC2D20000;
|
|
selp.f32 %f519, 0f00000000, %f518, %p55;
|
|
setp.gt.f32 %p56, %f507, 0f42D20000;
|
|
selp.f32 %f972, 0f7F800000, %f519, %p56;
|
|
setp.eq.f32 %p57, %f972, 0f7F800000;
|
|
@%p57 bra BB0_34;
|
|
|
|
fma.rn.f32 %f972, %f972, %f66, %f972;
|
|
|
|
BB0_34:
|
|
setp.lt.f32 %p58, %f971, 0f00000000;
|
|
setp.eq.f32 %p59, %f64, 0f3F800000;
|
|
and.pred %p3, %p58, %p59;
|
|
mov.b32 %r83, %f972;
|
|
xor.b32 %r84, %r83, -2147483648;
|
|
mov.b32 %f520, %r84;
|
|
selp.f32 %f974, %f520, %f972, %p3;
|
|
setp.eq.f32 %p60, %f971, 0f00000000;
|
|
@%p60 bra BB0_37;
|
|
bra.uni BB0_35;
|
|
|
|
BB0_37:
|
|
add.f32 %f523, %f971, %f971;
|
|
selp.f32 %f974, %f523, 0f00000000, %p59;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_123:
|
|
mov.u64 %rd287, image_HDR;
|
|
cvta.global.u64 %rd282, %rd287;
|
|
mov.u32 %r213, 8;
|
|
// inline asm
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r26, %r213, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f873, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f873;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs168, 0;
|
|
st.v4.u16 [%rd281], {%rs167, %rs167, %rs167, %rs168};
|
|
|
|
BB0_124:
|
|
ld.global.u32 %r214, [additive];
|
|
setp.eq.s32 %p135, %r214, 0;
|
|
@%p135 bra BB0_126;
|
|
|
|
mov.u64 %rd300, image_RNM0;
|
|
cvta.global.u64 %rd289, %rd300;
|
|
mov.u32 %r218, 8;
|
|
// inline asm
|
|
call (%rd288), _rt_buffer_get_64, (%rd289, %r26, %r218, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs175, %rs176, %rs177, %rs178}, [%rd288];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f874, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f875, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f876, %rs177;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd294), _rt_buffer_get_64, (%rd289, %r26, %r218, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f877, %f874, 0f00000000;
|
|
add.f32 %f878, %f875, 0f00000000;
|
|
add.f32 %f879, %f876, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f879;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f878;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f877;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs179, 0;
|
|
st.v4.u16 [%rd294], {%rs172, %rs173, %rs174, %rs179};
|
|
bra.uni BB0_127;
|
|
|
|
BB0_126:
|
|
mov.u64 %rd307, image_RNM0;
|
|
cvta.global.u64 %rd302, %rd307;
|
|
mov.u32 %r220, 8;
|
|
// inline asm
|
|
call (%rd301), _rt_buffer_get_64, (%rd302, %r26, %r220, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f880, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f880;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs181, 0;
|
|
st.v4.u16 [%rd301], {%rs180, %rs180, %rs180, %rs181};
|
|
|
|
BB0_127:
|
|
ld.global.u32 %r221, [additive];
|
|
setp.eq.s32 %p136, %r221, 0;
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd320, image_RNM1;
|
|
cvta.global.u64 %rd309, %rd320;
|
|
mov.u32 %r225, 8;
|
|
// inline asm
|
|
call (%rd308), _rt_buffer_get_64, (%rd309, %r26, %r225, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs188, %rs189, %rs190, %rs191}, [%rd308];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f881, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f882, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f883, %rs190;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd314), _rt_buffer_get_64, (%rd309, %r26, %r225, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f884, %f881, 0f00000000;
|
|
add.f32 %f885, %f882, 0f00000000;
|
|
add.f32 %f886, %f883, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs187, %f886;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f885;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f884;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs192, 0;
|
|
st.v4.u16 [%rd314], {%rs185, %rs186, %rs187, %rs192};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd327, image_RNM1;
|
|
cvta.global.u64 %rd322, %rd327;
|
|
mov.u32 %r227, 8;
|
|
// inline asm
|
|
call (%rd321), _rt_buffer_get_64, (%rd322, %r26, %r227, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f887, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs193, %f887;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs194, 0;
|
|
st.v4.u16 [%rd321], {%rs193, %rs193, %rs193, %rs194};
|
|
|
|
BB0_130:
|
|
ld.global.u32 %r228, [additive];
|
|
setp.eq.s32 %p137, %r228, 0;
|
|
@%p137 bra BB0_132;
|
|
|
|
mov.u64 %rd340, image_RNM2;
|
|
cvta.global.u64 %rd329, %rd340;
|
|
mov.u32 %r232, 8;
|
|
// inline asm
|
|
call (%rd328), _rt_buffer_get_64, (%rd329, %r26, %r232, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs201, %rs202, %rs203, %rs204}, [%rd328];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f888, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f889, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f890, %rs203;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd334), _rt_buffer_get_64, (%rd329, %r26, %r232, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f891, %f888, 0f00000000;
|
|
add.f32 %f892, %f889, 0f00000000;
|
|
add.f32 %f893, %f890, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs200, %f893;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f892;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f891;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs205, 0;
|
|
st.v4.u16 [%rd334], {%rs198, %rs199, %rs200, %rs205};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd347, image_RNM2;
|
|
cvta.global.u64 %rd342, %rd347;
|
|
mov.u32 %r234, 8;
|
|
// inline asm
|
|
call (%rd341), _rt_buffer_get_64, (%rd342, %r26, %r234, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f894, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f894;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs207, 0;
|
|
st.v4.u16 [%rd341], {%rs206, %rs206, %rs206, %rs207};
|
|
|
|
BB0_133:
|
|
ld.global.u32 %r235, [additive];
|
|
setp.eq.s32 %p138, %r235, 0;
|
|
@%p138 bra BB0_135;
|
|
|
|
mov.u64 %rd360, image_RNM3;
|
|
cvta.global.u64 %rd349, %rd360;
|
|
mov.u32 %r239, 8;
|
|
// inline asm
|
|
call (%rd348), _rt_buffer_get_64, (%rd349, %r26, %r239, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs214, %rs215, %rs216, %rs217}, [%rd348];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f895, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f896, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f897, %rs216;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd354), _rt_buffer_get_64, (%rd349, %r26, %r239, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f898, %f895, 0f00000000;
|
|
add.f32 %f899, %f896, 0f00000000;
|
|
add.f32 %f900, %f897, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs213, %f900;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f899;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f898;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs218, 0;
|
|
st.v4.u16 [%rd354], {%rs211, %rs212, %rs213, %rs218};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd367, image_RNM3;
|
|
cvta.global.u64 %rd362, %rd367;
|
|
mov.u32 %r241, 8;
|
|
// inline asm
|
|
call (%rd361), _rt_buffer_get_64, (%rd362, %r26, %r241, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f901, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs219, %f901;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs220, 0;
|
|
st.v4.u16 [%rd361], {%rs219, %rs219, %rs219, %rs220};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_103:
|
|
mov.u64 %rd173, image_HDR;
|
|
cvta.global.u64 %rd168, %rd173;
|
|
mov.u32 %r170, 8;
|
|
// inline asm
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r26, %r170, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f831, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f831;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd167], {%rs104, %rs104, %rs104, %rs93};
|
|
|
|
BB0_104:
|
|
ld.global.u32 %r171, [additive];
|
|
setp.eq.s32 %p127, %r171, 0;
|
|
mov.f32 %f832, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f832;}
|
|
|
|
// inline asm
|
|
@%p127 bra BB0_106;
|
|
|
|
mov.u64 %rd186, image_RNM0;
|
|
cvta.global.u64 %rd175, %rd186;
|
|
mov.u32 %r175, 8;
|
|
// inline asm
|
|
call (%rd174), _rt_buffer_get_64, (%rd175, %r26, %r175, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd174];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f833, %rs112;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f834, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f835, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd180), _rt_buffer_get_64, (%rd175, %r26, %r175, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f836, %f833, 0f00000000;
|
|
add.f32 %f837, %f834, 0f00000000;
|
|
add.f32 %f838, %f835, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f838;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f837;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f836;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd180], {%rs109, %rs110, %rs111, %rs105};
|
|
bra.uni BB0_107;
|
|
|
|
BB0_106:
|
|
mov.u64 %rd193, image_RNM0;
|
|
cvta.global.u64 %rd188, %rd193;
|
|
mov.u32 %r177, 8;
|
|
// inline asm
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r26, %r177, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f839, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f839;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd187], {%rs116, %rs116, %rs116, %rs105};
|
|
|
|
BB0_107:
|
|
ld.global.u32 %r178, [additive];
|
|
setp.eq.s32 %p128, %r178, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f832;}
|
|
|
|
// inline asm
|
|
@%p128 bra BB0_109;
|
|
|
|
mov.u64 %rd206, image_RNM1;
|
|
cvta.global.u64 %rd195, %rd206;
|
|
mov.u32 %r182, 8;
|
|
// inline asm
|
|
call (%rd194), _rt_buffer_get_64, (%rd195, %r26, %r182, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs124, %rs125, %rs126, %rs127}, [%rd194];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f841, %rs124;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f842, %rs125;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f843, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd200), _rt_buffer_get_64, (%rd195, %r26, %r182, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f844, %f841, 0f00000000;
|
|
add.f32 %f845, %f842, 0f00000000;
|
|
add.f32 %f846, %f843, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f846;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f845;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f844;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd200], {%rs121, %rs122, %rs123, %rs117};
|
|
bra.uni BB0_110;
|
|
|
|
BB0_109:
|
|
mov.u64 %rd213, image_RNM1;
|
|
cvta.global.u64 %rd208, %rd213;
|
|
mov.u32 %r184, 8;
|
|
// inline asm
|
|
call (%rd207), _rt_buffer_get_64, (%rd208, %r26, %r184, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f847, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f847;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd207], {%rs128, %rs128, %rs128, %rs117};
|
|
|
|
BB0_110:
|
|
ld.global.u32 %r185, [additive];
|
|
setp.eq.s32 %p129, %r185, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f832;}
|
|
|
|
// inline asm
|
|
@%p129 bra BB0_112;
|
|
|
|
mov.u64 %rd226, image_RNM2;
|
|
cvta.global.u64 %rd215, %rd226;
|
|
mov.u32 %r189, 8;
|
|
// inline asm
|
|
call (%rd214), _rt_buffer_get_64, (%rd215, %r26, %r189, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd214];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs136;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f850, %rs137;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd220), _rt_buffer_get_64, (%rd215, %r26, %r189, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f852, %f849, 0f00000000;
|
|
add.f32 %f853, %f850, 0f00000000;
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f854;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f853;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f852;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd220], {%rs133, %rs134, %rs135, %rs129};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_112:
|
|
mov.u64 %rd233, image_RNM2;
|
|
cvta.global.u64 %rd228, %rd233;
|
|
mov.u32 %r191, 8;
|
|
// inline asm
|
|
call (%rd227), _rt_buffer_get_64, (%rd228, %r26, %r191, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f855, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f855;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd227], {%rs140, %rs140, %rs140, %rs129};
|
|
|
|
BB0_113:
|
|
ld.global.u32 %r192, [additive];
|
|
setp.eq.s32 %p130, %r192, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f832;}
|
|
|
|
// inline asm
|
|
@%p130 bra BB0_115;
|
|
|
|
mov.u64 %rd246, image_RNM3;
|
|
cvta.global.u64 %rd235, %rd246;
|
|
mov.u32 %r196, 8;
|
|
// inline asm
|
|
call (%rd234), _rt_buffer_get_64, (%rd235, %r26, %r196, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs148, %rs149, %rs150, %rs151}, [%rd234];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f857, %rs148;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f859, %rs150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd240), _rt_buffer_get_64, (%rd235, %r26, %r196, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f860, %f857, 0f00000000;
|
|
add.f32 %f861, %f858, 0f00000000;
|
|
add.f32 %f862, %f859, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs147, %f862;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f861;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f860;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd240], {%rs145, %rs146, %rs147, %rs141};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_115:
|
|
mov.u64 %rd253, image_RNM3;
|
|
cvta.global.u64 %rd248, %rd253;
|
|
mov.u32 %r198, 8;
|
|
// inline asm
|
|
call (%rd247), _rt_buffer_get_64, (%rd248, %r26, %r198, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f863, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f863;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd247], {%rs152, %rs152, %rs152, %rs141};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_35:
|
|
setp.geu.f32 %p61, %f971, 0f00000000;
|
|
@%p61 bra BB0_38;
|
|
|
|
cvt.rzi.f32.f32 %f522, %f494;
|
|
setp.neu.f32 %p62, %f522, 0f3EE8BA2E;
|
|
selp.f32 %f974, 0f7FFFFFFF, %f974, %p62;
|
|
|
|
BB0_38:
|
|
add.f32 %f524, %f65, 0f3EE8BA2E;
|
|
mov.b32 %r85, %f524;
|
|
setp.lt.s32 %p64, %r85, 2139095040;
|
|
@%p64 bra BB0_43;
|
|
|
|
setp.gtu.f32 %p65, %f65, 0f7F800000;
|
|
@%p65 bra BB0_42;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_42:
|
|
add.f32 %f974, %f971, 0f3EE8BA2E;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_40:
|
|
setp.neu.f32 %p66, %f65, 0f7F800000;
|
|
@%p66 bra BB0_43;
|
|
|
|
selp.f32 %f974, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_43:
|
|
mul.f32 %f525, %f974, 0f437F0000;
|
|
setp.eq.f32 %p67, %f971, 0f3F800000;
|
|
selp.f32 %f526, 0f437F0000, %f525, %p67;
|
|
cvt.rzi.u32.f32 %r86, %f526;
|
|
cvt.u16.u32 %rs19, %r86;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd26], {%rs19, %rs20};
|
|
ld.global.u32 %r246, [imageEnabled];
|
|
|
|
BB0_44:
|
|
and.b32 %r87, %r246, 1;
|
|
setp.eq.b32 %p68, %r87, 1;
|
|
@!%p68 bra BB0_79;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_45:
|
|
mov.f32 %f931, 0fB5BFBE8E;
|
|
mov.f32 %f930, 0fBF317200;
|
|
mov.f32 %f929, 0f35BFBE8E;
|
|
mov.f32 %f928, 0f3F317200;
|
|
mov.f32 %f927, 0f3DAAAABD;
|
|
mov.f32 %f926, 0f3C4CAF63;
|
|
mov.f32 %f925, 0f3B18F0FE;
|
|
mov.f32 %f529, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f530, %f529;
|
|
fma.rn.f32 %f531, %f530, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f77, %f531;
|
|
abs.f32 %f78, %f61;
|
|
setp.lt.f32 %p69, %f78, 0f00800000;
|
|
mul.f32 %f532, %f78, 0f4B800000;
|
|
selp.f32 %f533, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f534, %f532, %f78, %p69;
|
|
mov.b32 %r88, %f534;
|
|
and.b32 %r89, %r88, 8388607;
|
|
or.b32 %r90, %r89, 1065353216;
|
|
mov.b32 %f535, %r90;
|
|
shr.u32 %r91, %r88, 23;
|
|
cvt.rn.f32.u32 %f536, %r91;
|
|
add.f32 %f537, %f533, %f536;
|
|
setp.gt.f32 %p70, %f535, 0f3FB504F3;
|
|
mul.f32 %f538, %f535, 0f3F000000;
|
|
add.f32 %f539, %f537, 0f3F800000;
|
|
selp.f32 %f540, %f538, %f535, %p70;
|
|
selp.f32 %f541, %f539, %f537, %p70;
|
|
add.f32 %f542, %f540, 0fBF800000;
|
|
add.f32 %f528, %f540, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f527,%f528;
|
|
// inline asm
|
|
add.f32 %f543, %f542, %f542;
|
|
mul.f32 %f544, %f527, %f543;
|
|
mul.f32 %f545, %f544, %f544;
|
|
fma.rn.f32 %f548, %f925, %f545, %f926;
|
|
fma.rn.f32 %f550, %f548, %f545, %f927;
|
|
mul.rn.f32 %f551, %f550, %f545;
|
|
mul.rn.f32 %f552, %f551, %f544;
|
|
sub.f32 %f553, %f542, %f544;
|
|
neg.f32 %f554, %f544;
|
|
add.f32 %f555, %f553, %f553;
|
|
fma.rn.f32 %f556, %f554, %f542, %f555;
|
|
mul.rn.f32 %f557, %f527, %f556;
|
|
add.f32 %f558, %f552, %f544;
|
|
sub.f32 %f559, %f544, %f558;
|
|
add.f32 %f560, %f552, %f559;
|
|
add.f32 %f561, %f557, %f560;
|
|
add.f32 %f562, %f558, %f561;
|
|
sub.f32 %f563, %f558, %f562;
|
|
add.f32 %f564, %f561, %f563;
|
|
mul.rn.f32 %f566, %f541, %f928;
|
|
mul.rn.f32 %f568, %f541, %f929;
|
|
add.f32 %f569, %f566, %f562;
|
|
sub.f32 %f570, %f566, %f569;
|
|
add.f32 %f571, %f562, %f570;
|
|
add.f32 %f572, %f564, %f571;
|
|
add.f32 %f573, %f568, %f572;
|
|
add.f32 %f574, %f569, %f573;
|
|
sub.f32 %f575, %f569, %f574;
|
|
add.f32 %f576, %f573, %f575;
|
|
mov.f32 %f577, 0f3EE66666;
|
|
mul.rn.f32 %f578, %f577, %f574;
|
|
neg.f32 %f579, %f578;
|
|
fma.rn.f32 %f580, %f577, %f574, %f579;
|
|
fma.rn.f32 %f581, %f577, %f576, %f580;
|
|
mov.f32 %f582, 0f00000000;
|
|
fma.rn.f32 %f583, %f582, %f574, %f581;
|
|
add.rn.f32 %f584, %f578, %f583;
|
|
neg.f32 %f585, %f584;
|
|
add.rn.f32 %f586, %f578, %f585;
|
|
add.rn.f32 %f587, %f586, %f583;
|
|
mov.b32 %r92, %f584;
|
|
setp.eq.s32 %p71, %r92, 1118925336;
|
|
add.s32 %r93, %r92, -1;
|
|
mov.b32 %f588, %r93;
|
|
add.f32 %f589, %f587, 0f37000000;
|
|
selp.f32 %f590, %f588, %f584, %p71;
|
|
selp.f32 %f79, %f589, %f587, %p71;
|
|
mul.f32 %f591, %f590, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f592, %f591;
|
|
fma.rn.f32 %f594, %f592, %f930, %f590;
|
|
fma.rn.f32 %f596, %f592, %f931, %f594;
|
|
mul.f32 %f597, %f596, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f598, %f597;
|
|
add.f32 %f599, %f592, 0f00000000;
|
|
ex2.approx.f32 %f600, %f599;
|
|
mul.f32 %f601, %f598, %f600;
|
|
setp.lt.f32 %p72, %f590, 0fC2D20000;
|
|
selp.f32 %f602, 0f00000000, %f601, %p72;
|
|
setp.gt.f32 %p73, %f590, 0f42D20000;
|
|
selp.f32 %f975, 0f7F800000, %f602, %p73;
|
|
setp.eq.f32 %p74, %f975, 0f7F800000;
|
|
@%p74 bra BB0_47;
|
|
|
|
fma.rn.f32 %f975, %f975, %f79, %f975;
|
|
|
|
BB0_47:
|
|
setp.lt.f32 %p75, %f61, 0f00000000;
|
|
setp.eq.f32 %p76, %f77, 0f3F800000;
|
|
and.pred %p4, %p75, %p76;
|
|
mov.b32 %r94, %f975;
|
|
xor.b32 %r95, %r94, -2147483648;
|
|
mov.b32 %f603, %r95;
|
|
selp.f32 %f977, %f603, %f975, %p4;
|
|
setp.eq.f32 %p77, %f61, 0f00000000;
|
|
@%p77 bra BB0_50;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_50:
|
|
add.f32 %f606, %f61, %f61;
|
|
selp.f32 %f977, %f606, 0f00000000, %p76;
|
|
bra.uni BB0_51;
|
|
|
|
BB0_48:
|
|
setp.geu.f32 %p78, %f61, 0f00000000;
|
|
@%p78 bra BB0_51;
|
|
|
|
cvt.rzi.f32.f32 %f605, %f577;
|
|
setp.neu.f32 %p79, %f605, 0f3EE66666;
|
|
selp.f32 %f977, 0f7FFFFFFF, %f977, %p79;
|
|
|
|
BB0_51:
|
|
add.f32 %f607, %f78, 0f3EE66666;
|
|
mov.b32 %r96, %f607;
|
|
setp.lt.s32 %p81, %r96, 2139095040;
|
|
@%p81 bra BB0_56;
|
|
|
|
setp.gtu.f32 %p82, %f78, 0f7F800000;
|
|
@%p82 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
add.f32 %f977, %f61, 0f3EE66666;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_53:
|
|
setp.neu.f32 %p83, %f78, 0f7F800000;
|
|
@%p83 bra BB0_56;
|
|
|
|
selp.f32 %f977, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_56:
|
|
mov.f32 %f938, 0fB5BFBE8E;
|
|
mov.f32 %f937, 0fBF317200;
|
|
mov.f32 %f936, 0f35BFBE8E;
|
|
mov.f32 %f935, 0f3F317200;
|
|
mov.f32 %f934, 0f3DAAAABD;
|
|
mov.f32 %f933, 0f3C4CAF63;
|
|
mov.f32 %f932, 0f3B18F0FE;
|
|
setp.eq.f32 %p84, %f61, 0f3F800000;
|
|
selp.f32 %f90, 0f3F800000, %f977, %p84;
|
|
abs.f32 %f91, %f62;
|
|
setp.lt.f32 %p85, %f91, 0f00800000;
|
|
mul.f32 %f610, %f91, 0f4B800000;
|
|
selp.f32 %f611, 0fC3170000, 0fC2FE0000, %p85;
|
|
selp.f32 %f612, %f610, %f91, %p85;
|
|
mov.b32 %r97, %f612;
|
|
and.b32 %r98, %r97, 8388607;
|
|
or.b32 %r99, %r98, 1065353216;
|
|
mov.b32 %f613, %r99;
|
|
shr.u32 %r100, %r97, 23;
|
|
cvt.rn.f32.u32 %f614, %r100;
|
|
add.f32 %f615, %f611, %f614;
|
|
setp.gt.f32 %p86, %f613, 0f3FB504F3;
|
|
mul.f32 %f616, %f613, 0f3F000000;
|
|
add.f32 %f617, %f615, 0f3F800000;
|
|
selp.f32 %f618, %f616, %f613, %p86;
|
|
selp.f32 %f619, %f617, %f615, %p86;
|
|
add.f32 %f620, %f618, 0fBF800000;
|
|
add.f32 %f609, %f618, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f608,%f609;
|
|
// inline asm
|
|
add.f32 %f621, %f620, %f620;
|
|
mul.f32 %f622, %f608, %f621;
|
|
mul.f32 %f623, %f622, %f622;
|
|
fma.rn.f32 %f626, %f932, %f623, %f933;
|
|
fma.rn.f32 %f628, %f626, %f623, %f934;
|
|
mul.rn.f32 %f629, %f628, %f623;
|
|
mul.rn.f32 %f630, %f629, %f622;
|
|
sub.f32 %f631, %f620, %f622;
|
|
neg.f32 %f632, %f622;
|
|
add.f32 %f633, %f631, %f631;
|
|
fma.rn.f32 %f634, %f632, %f620, %f633;
|
|
mul.rn.f32 %f635, %f608, %f634;
|
|
add.f32 %f636, %f630, %f622;
|
|
sub.f32 %f637, %f622, %f636;
|
|
add.f32 %f638, %f630, %f637;
|
|
add.f32 %f639, %f635, %f638;
|
|
add.f32 %f640, %f636, %f639;
|
|
sub.f32 %f641, %f636, %f640;
|
|
add.f32 %f642, %f639, %f641;
|
|
mul.rn.f32 %f644, %f619, %f935;
|
|
mul.rn.f32 %f646, %f619, %f936;
|
|
add.f32 %f647, %f644, %f640;
|
|
sub.f32 %f648, %f644, %f647;
|
|
add.f32 %f649, %f640, %f648;
|
|
add.f32 %f650, %f642, %f649;
|
|
add.f32 %f651, %f646, %f650;
|
|
add.f32 %f652, %f647, %f651;
|
|
sub.f32 %f653, %f647, %f652;
|
|
add.f32 %f654, %f651, %f653;
|
|
mul.rn.f32 %f656, %f577, %f652;
|
|
neg.f32 %f657, %f656;
|
|
fma.rn.f32 %f658, %f577, %f652, %f657;
|
|
fma.rn.f32 %f659, %f577, %f654, %f658;
|
|
fma.rn.f32 %f661, %f582, %f652, %f659;
|
|
add.rn.f32 %f662, %f656, %f661;
|
|
neg.f32 %f663, %f662;
|
|
add.rn.f32 %f664, %f656, %f663;
|
|
add.rn.f32 %f665, %f664, %f661;
|
|
mov.b32 %r101, %f662;
|
|
setp.eq.s32 %p87, %r101, 1118925336;
|
|
add.s32 %r102, %r101, -1;
|
|
mov.b32 %f666, %r102;
|
|
add.f32 %f667, %f665, 0f37000000;
|
|
selp.f32 %f668, %f666, %f662, %p87;
|
|
selp.f32 %f92, %f667, %f665, %p87;
|
|
mul.f32 %f669, %f668, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f670, %f669;
|
|
fma.rn.f32 %f672, %f670, %f937, %f668;
|
|
fma.rn.f32 %f674, %f670, %f938, %f672;
|
|
mul.f32 %f675, %f674, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f676, %f675;
|
|
add.f32 %f677, %f670, 0f00000000;
|
|
ex2.approx.f32 %f678, %f677;
|
|
mul.f32 %f679, %f676, %f678;
|
|
setp.lt.f32 %p88, %f668, 0fC2D20000;
|
|
selp.f32 %f680, 0f00000000, %f679, %p88;
|
|
setp.gt.f32 %p89, %f668, 0f42D20000;
|
|
selp.f32 %f978, 0f7F800000, %f680, %p89;
|
|
setp.eq.f32 %p90, %f978, 0f7F800000;
|
|
@%p90 bra BB0_58;
|
|
|
|
fma.rn.f32 %f978, %f978, %f92, %f978;
|
|
|
|
BB0_58:
|
|
setp.lt.f32 %p91, %f62, 0f00000000;
|
|
and.pred %p5, %p91, %p76;
|
|
mov.b32 %r103, %f978;
|
|
xor.b32 %r104, %r103, -2147483648;
|
|
mov.b32 %f681, %r104;
|
|
selp.f32 %f980, %f681, %f978, %p5;
|
|
setp.eq.f32 %p93, %f62, 0f00000000;
|
|
@%p93 bra BB0_61;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_61:
|
|
add.f32 %f684, %f62, %f62;
|
|
selp.f32 %f980, %f684, 0f00000000, %p76;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_59:
|
|
setp.geu.f32 %p94, %f62, 0f00000000;
|
|
@%p94 bra BB0_62;
|
|
|
|
cvt.rzi.f32.f32 %f683, %f577;
|
|
setp.neu.f32 %p95, %f683, 0f3EE66666;
|
|
selp.f32 %f980, 0f7FFFFFFF, %f980, %p95;
|
|
|
|
BB0_62:
|
|
add.f32 %f685, %f91, 0f3EE66666;
|
|
mov.b32 %r105, %f685;
|
|
setp.lt.s32 %p97, %r105, 2139095040;
|
|
@%p97 bra BB0_67;
|
|
|
|
setp.gtu.f32 %p98, %f91, 0f7F800000;
|
|
@%p98 bra BB0_66;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_66:
|
|
add.f32 %f980, %f62, 0f3EE66666;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_64:
|
|
setp.neu.f32 %p99, %f91, 0f7F800000;
|
|
@%p99 bra BB0_67;
|
|
|
|
selp.f32 %f980, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_67:
|
|
mov.f32 %f945, 0fB5BFBE8E;
|
|
mov.f32 %f944, 0fBF317200;
|
|
mov.f32 %f943, 0f35BFBE8E;
|
|
mov.f32 %f942, 0f3F317200;
|
|
mov.f32 %f941, 0f3DAAAABD;
|
|
mov.f32 %f940, 0f3C4CAF63;
|
|
mov.f32 %f939, 0f3B18F0FE;
|
|
setp.eq.f32 %p100, %f62, 0f3F800000;
|
|
selp.f32 %f103, 0f3F800000, %f980, %p100;
|
|
abs.f32 %f104, %f63;
|
|
setp.lt.f32 %p101, %f104, 0f00800000;
|
|
mul.f32 %f688, %f104, 0f4B800000;
|
|
selp.f32 %f689, 0fC3170000, 0fC2FE0000, %p101;
|
|
selp.f32 %f690, %f688, %f104, %p101;
|
|
mov.b32 %r106, %f690;
|
|
and.b32 %r107, %r106, 8388607;
|
|
or.b32 %r108, %r107, 1065353216;
|
|
mov.b32 %f691, %r108;
|
|
shr.u32 %r109, %r106, 23;
|
|
cvt.rn.f32.u32 %f692, %r109;
|
|
add.f32 %f693, %f689, %f692;
|
|
setp.gt.f32 %p102, %f691, 0f3FB504F3;
|
|
mul.f32 %f694, %f691, 0f3F000000;
|
|
add.f32 %f695, %f693, 0f3F800000;
|
|
selp.f32 %f696, %f694, %f691, %p102;
|
|
selp.f32 %f697, %f695, %f693, %p102;
|
|
add.f32 %f698, %f696, 0fBF800000;
|
|
add.f32 %f687, %f696, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f686,%f687;
|
|
// inline asm
|
|
add.f32 %f699, %f698, %f698;
|
|
mul.f32 %f700, %f686, %f699;
|
|
mul.f32 %f701, %f700, %f700;
|
|
fma.rn.f32 %f704, %f939, %f701, %f940;
|
|
fma.rn.f32 %f706, %f704, %f701, %f941;
|
|
mul.rn.f32 %f707, %f706, %f701;
|
|
mul.rn.f32 %f708, %f707, %f700;
|
|
sub.f32 %f709, %f698, %f700;
|
|
neg.f32 %f710, %f700;
|
|
add.f32 %f711, %f709, %f709;
|
|
fma.rn.f32 %f712, %f710, %f698, %f711;
|
|
mul.rn.f32 %f713, %f686, %f712;
|
|
add.f32 %f714, %f708, %f700;
|
|
sub.f32 %f715, %f700, %f714;
|
|
add.f32 %f716, %f708, %f715;
|
|
add.f32 %f717, %f713, %f716;
|
|
add.f32 %f718, %f714, %f717;
|
|
sub.f32 %f719, %f714, %f718;
|
|
add.f32 %f720, %f717, %f719;
|
|
mul.rn.f32 %f722, %f697, %f942;
|
|
mul.rn.f32 %f724, %f697, %f943;
|
|
add.f32 %f725, %f722, %f718;
|
|
sub.f32 %f726, %f722, %f725;
|
|
add.f32 %f727, %f718, %f726;
|
|
add.f32 %f728, %f720, %f727;
|
|
add.f32 %f729, %f724, %f728;
|
|
add.f32 %f730, %f725, %f729;
|
|
sub.f32 %f731, %f725, %f730;
|
|
add.f32 %f732, %f729, %f731;
|
|
mul.rn.f32 %f734, %f577, %f730;
|
|
neg.f32 %f735, %f734;
|
|
fma.rn.f32 %f736, %f577, %f730, %f735;
|
|
fma.rn.f32 %f737, %f577, %f732, %f736;
|
|
fma.rn.f32 %f739, %f582, %f730, %f737;
|
|
add.rn.f32 %f740, %f734, %f739;
|
|
neg.f32 %f741, %f740;
|
|
add.rn.f32 %f742, %f734, %f741;
|
|
add.rn.f32 %f743, %f742, %f739;
|
|
mov.b32 %r110, %f740;
|
|
setp.eq.s32 %p103, %r110, 1118925336;
|
|
add.s32 %r111, %r110, -1;
|
|
mov.b32 %f744, %r111;
|
|
add.f32 %f745, %f743, 0f37000000;
|
|
selp.f32 %f746, %f744, %f740, %p103;
|
|
selp.f32 %f105, %f745, %f743, %p103;
|
|
mul.f32 %f747, %f746, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f748, %f747;
|
|
fma.rn.f32 %f750, %f748, %f944, %f746;
|
|
fma.rn.f32 %f752, %f748, %f945, %f750;
|
|
mul.f32 %f753, %f752, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f754, %f753;
|
|
add.f32 %f755, %f748, 0f00000000;
|
|
ex2.approx.f32 %f756, %f755;
|
|
mul.f32 %f757, %f754, %f756;
|
|
setp.lt.f32 %p104, %f746, 0fC2D20000;
|
|
selp.f32 %f758, 0f00000000, %f757, %p104;
|
|
setp.gt.f32 %p105, %f746, 0f42D20000;
|
|
selp.f32 %f981, 0f7F800000, %f758, %p105;
|
|
setp.eq.f32 %p106, %f981, 0f7F800000;
|
|
@%p106 bra BB0_69;
|
|
|
|
fma.rn.f32 %f981, %f981, %f105, %f981;
|
|
|
|
BB0_69:
|
|
setp.lt.f32 %p107, %f63, 0f00000000;
|
|
and.pred %p6, %p107, %p76;
|
|
mov.b32 %r112, %f981;
|
|
xor.b32 %r113, %r112, -2147483648;
|
|
mov.b32 %f759, %r113;
|
|
selp.f32 %f983, %f759, %f981, %p6;
|
|
setp.eq.f32 %p109, %f63, 0f00000000;
|
|
@%p109 bra BB0_72;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_72:
|
|
add.f32 %f762, %f63, %f63;
|
|
selp.f32 %f983, %f762, 0f00000000, %p76;
|
|
bra.uni BB0_73;
|
|
|
|
BB0_70:
|
|
setp.geu.f32 %p110, %f63, 0f00000000;
|
|
@%p110 bra BB0_73;
|
|
|
|
cvt.rzi.f32.f32 %f761, %f577;
|
|
setp.neu.f32 %p111, %f761, 0f3EE66666;
|
|
selp.f32 %f983, 0f7FFFFFFF, %f983, %p111;
|
|
|
|
BB0_73:
|
|
add.f32 %f763, %f104, 0f3EE66666;
|
|
mov.b32 %r114, %f763;
|
|
setp.lt.s32 %p113, %r114, 2139095040;
|
|
@%p113 bra BB0_78;
|
|
|
|
setp.gtu.f32 %p114, %f104, 0f7F800000;
|
|
@%p114 bra BB0_77;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_77:
|
|
add.f32 %f983, %f63, 0f3EE66666;
|
|
bra.uni BB0_78;
|
|
|
|
BB0_75:
|
|
setp.neu.f32 %p115, %f104, 0f7F800000;
|
|
@%p115 bra BB0_78;
|
|
|
|
selp.f32 %f983, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_78:
|
|
setp.eq.f32 %p116, %f63, 0f3F800000;
|
|
selp.f32 %f764, 0f3F800000, %f983, %p116;
|
|
cvt.u64.u32 %rd36, %r3;
|
|
cvt.u64.u32 %rd35, %r2;
|
|
mov.u64 %rd39, image;
|
|
cvta.global.u64 %rd34, %rd39;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r26, %r27, %rd35, %rd36, %rd15, %rd15);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f765, %f764;
|
|
mul.f32 %f766, %f765, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r117, %f766;
|
|
cvt.sat.f32.f32 %f767, %f103;
|
|
mul.f32 %f768, %f767, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r118, %f768;
|
|
cvt.sat.f32.f32 %f769, %f90;
|
|
mul.f32 %f770, %f769, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r119, %f770;
|
|
cvt.u16.u32 %rs21, %r117;
|
|
cvt.u16.u32 %rs22, %r119;
|
|
cvt.u16.u32 %rs23, %r118;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd33], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r246, [imageEnabled];
|
|
|
|
BB0_79:
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
and.b32 %r120, %r246, 4;
|
|
setp.eq.s32 %p117, %r120, 0;
|
|
@%p117 bra BB0_83;
|
|
|
|
ld.global.u32 %r121, [additive];
|
|
setp.eq.s32 %p118, %r121, 0;
|
|
mov.f32 %f771, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f771;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_82;
|
|
|
|
mov.u64 %rd52, image_HDR;
|
|
cvta.global.u64 %rd41, %rd52;
|
|
mov.u32 %r125, 8;
|
|
// inline asm
|
|
call (%rd40), _rt_buffer_get_64, (%rd41, %r26, %r125, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd40];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f772, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f773, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f774, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd46), _rt_buffer_get_64, (%rd41, %r26, %r125, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f775, %f61, %f772;
|
|
add.f32 %f776, %f62, %f773;
|
|
add.f32 %f777, %f63, %f774;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f777;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f776;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f775;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd46], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_83;
|
|
|
|
BB0_82:
|
|
mov.u64 %rd59, image_HDR;
|
|
cvta.global.u64 %rd54, %rd59;
|
|
mov.u32 %r127, 8;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd54, %r26, %r127, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f63;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f61;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd53], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_83:
|
|
mul.f32 %f782, %f48, 0f3E800000;
|
|
mul.f32 %f783, %f782, %f971;
|
|
mul.f32 %f784, %f47, %f783;
|
|
ld.global.f32 %f785, [lightColor];
|
|
mul.f32 %f116, %f784, %f785;
|
|
ld.global.f32 %f786, [lightColor+4];
|
|
mul.f32 %f117, %f784, %f786;
|
|
ld.global.f32 %f787, [lightColor+8];
|
|
mul.f32 %f118, %f784, %f787;
|
|
ld.global.u32 %r128, [additive];
|
|
setp.eq.s32 %p119, %r128, 0;
|
|
mov.f32 %f781, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f781;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_85;
|
|
|
|
mov.u64 %rd72, image_RNM0;
|
|
cvta.global.u64 %rd61, %rd72;
|
|
mov.u32 %r132, 8;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd61, %r26, %r132, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd60];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f788, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f789, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f790, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd66), _rt_buffer_get_64, (%rd61, %r26, %r132, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f791, %f116, %f788;
|
|
add.f32 %f792, %f117, %f789;
|
|
add.f32 %f793, %f118, %f790;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f793;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f792;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f791;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd66], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_86;
|
|
|
|
BB0_85:
|
|
mov.u64 %rd79, image_RNM0;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
mov.u32 %r134, 8;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r26, %r134, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f116;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd73], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_86:
|
|
fma.rn.f32 %f119, %f13, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f120, %f14, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f121, %f15, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r135, [additive];
|
|
setp.eq.s32 %p120, %r135, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f781;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_88;
|
|
|
|
mov.u64 %rd92, image_RNM1;
|
|
cvta.global.u64 %rd81, %rd92;
|
|
mov.u32 %r139, 8;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r26, %r139, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd80];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f798, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f799, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f800, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd86), _rt_buffer_get_64, (%rd81, %r26, %r139, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f801, %f119, %f798;
|
|
add.f32 %f802, %f119, %f799;
|
|
add.f32 %f803, %f119, %f800;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f803;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f802;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f801;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd86], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_89;
|
|
|
|
BB0_88:
|
|
mov.u64 %rd99, image_RNM1;
|
|
cvta.global.u64 %rd94, %rd99;
|
|
mov.u32 %r141, 8;
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd94, %r26, %r141, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f119;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs64, %rs64, %rs64, %rs53};
|
|
|
|
BB0_89:
|
|
ld.global.u32 %r142, [additive];
|
|
setp.eq.s32 %p121, %r142, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f781;}
|
|
|
|
// inline asm
|
|
@%p121 bra BB0_91;
|
|
|
|
mov.u64 %rd112, image_RNM2;
|
|
cvta.global.u64 %rd101, %rd112;
|
|
mov.u32 %r146, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r26, %r146, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd100];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f806, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f807, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f808, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd106), _rt_buffer_get_64, (%rd101, %r26, %r146, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f809, %f120, %f806;
|
|
add.f32 %f810, %f120, %f807;
|
|
add.f32 %f811, %f120, %f808;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f811;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f810;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f809;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd106], {%rs69, %rs70, %rs71, %rs65};
|
|
bra.uni BB0_92;
|
|
|
|
BB0_91:
|
|
mov.u64 %rd119, image_RNM2;
|
|
cvta.global.u64 %rd114, %rd119;
|
|
mov.u32 %r148, 8;
|
|
// inline asm
|
|
call (%rd113), _rt_buffer_get_64, (%rd114, %r26, %r148, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f120;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd113], {%rs76, %rs76, %rs76, %rs65};
|
|
|
|
BB0_92:
|
|
ld.global.u32 %r149, [additive];
|
|
setp.eq.s32 %p122, %r149, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f781;}
|
|
|
|
// inline asm
|
|
@%p122 bra BB0_94;
|
|
|
|
mov.u64 %rd132, image_RNM3;
|
|
cvta.global.u64 %rd121, %rd132;
|
|
mov.u32 %r153, 8;
|
|
// inline asm
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r26, %r153, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd120];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f814, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f815, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f816, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd126), _rt_buffer_get_64, (%rd121, %r26, %r153, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f817, %f121, %f814;
|
|
add.f32 %f818, %f121, %f815;
|
|
add.f32 %f819, %f121, %f816;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f819;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f818;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f817;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd126], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_94:
|
|
mov.u64 %rd139, image_RNM3;
|
|
cvta.global.u64 %rd134, %rd139;
|
|
mov.u32 %r155, 8;
|
|
// inline asm
|
|
call (%rd133), _rt_buffer_get_64, (%rd134, %r26, %r155, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f121;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd133], {%rs88, %rs88, %rs88, %rs77};
|
|
|
|
BB0_136:
|
|
ret;
|
|
}
|
|
|
|
|