2307 lines
65 KiB
Plaintext
2307 lines
65 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 .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<123>;
|
|
.reg .b16 %rs<221>;
|
|
.reg .f32 %f<897>;
|
|
.reg .b32 %r<244>;
|
|
.reg .b64 %rd<369>;
|
|
|
|
|
|
mov.u64 %rd368, __local_depot0;
|
|
cvta.local.u64 %SP, %rd368;
|
|
ld.global.v2.u32 {%r26, %r27}, [pixelID];
|
|
cvt.u64.u32 %rd12, %r26;
|
|
cvt.u64.u32 %rd13, %r27;
|
|
mov.u64 %rd16, uvnormal;
|
|
cvta.global.u64 %rd11, %rd16;
|
|
mov.u32 %r24, 2;
|
|
mov.u32 %r25, 4;
|
|
mov.u64 %rd15, 0;
|
|
// inline asm
|
|
call (%rd10), _rt_buffer_get_64, (%rd11, %r24, %r25, %rd12, %rd13, %rd15, %rd15);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd10];
|
|
shr.u32 %r30, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r30;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r1;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p6, %rs14, 0;
|
|
mov.f32 %f876, 0f00000000;
|
|
mov.f32 %f877, %f876;
|
|
mov.f32 %f878, %f876;
|
|
@%p6 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd10+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f106, %rs17;
|
|
div.rn.f32 %f107, %f106, 0f437F0000;
|
|
fma.rn.f32 %f108, %f107, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f109, %rs15;
|
|
div.rn.f32 %f110, %f109, 0f437F0000;
|
|
fma.rn.f32 %f111, %f110, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f112, %rs12;
|
|
div.rn.f32 %f113, %f112, 0f437F0000;
|
|
fma.rn.f32 %f114, %f113, 0f40000000, 0fBF800000;
|
|
mul.f32 %f115, %f111, %f111;
|
|
fma.rn.f32 %f116, %f108, %f108, %f115;
|
|
fma.rn.f32 %f117, %f114, %f114, %f116;
|
|
sqrt.rn.f32 %f118, %f117;
|
|
rcp.rn.f32 %f119, %f118;
|
|
mul.f32 %f876, %f108, %f119;
|
|
mul.f32 %f877, %f111, %f119;
|
|
mul.f32 %f878, %f114, %f119;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r31, %r32}, [pixelID];
|
|
ld.global.v2.u32 {%r34, %r35}, [tileInfo];
|
|
add.s32 %r2, %r31, %r34;
|
|
add.s32 %r3, %r32, %r35;
|
|
setp.eq.f32 %p7, %f877, 0f00000000;
|
|
setp.eq.f32 %p8, %f876, 0f00000000;
|
|
and.pred %p9, %p8, %p7;
|
|
setp.eq.f32 %p10, %f878, 0f00000000;
|
|
and.pred %p11, %p9, %p10;
|
|
@%p11 bra BB0_104;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_104:
|
|
ld.global.u32 %r243, [imageEnabled];
|
|
and.b32 %r192, %r243, 1;
|
|
setp.eq.b32 %p115, %r192, 1;
|
|
@!%p115 bra BB0_106;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_105:
|
|
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, %r24, %r25, %rd256, %rd257, %rd15, %rd15);
|
|
// inline asm
|
|
mov.u16 %rs153, 0;
|
|
st.v4.u8 [%rd254], {%rs153, %rs153, %rs153, %rs153};
|
|
ld.global.u32 %r243, [imageEnabled];
|
|
|
|
BB0_106:
|
|
and.b32 %r195, %r243, 8;
|
|
setp.eq.s32 %p116, %r195, 0;
|
|
@%p116 bra BB0_108;
|
|
|
|
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, %r24, %r24, %rd263, %rd264, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f792, 0f00000000;
|
|
cvt.rzi.u32.f32 %r198, %f792;
|
|
cvt.u16.u32 %rs154, %r198;
|
|
mov.u16 %rs155, 0;
|
|
st.v2.u8 [%rd261], {%rs154, %rs155};
|
|
ld.global.u32 %r243, [imageEnabled];
|
|
|
|
BB0_108:
|
|
cvt.u64.u32 %rd8, %r2;
|
|
cvt.u64.u32 %rd9, %r3;
|
|
and.b32 %r199, %r243, 4;
|
|
setp.eq.s32 %p117, %r199, 0;
|
|
@%p117 bra BB0_112;
|
|
|
|
ld.global.u32 %r200, [additive];
|
|
setp.eq.s32 %p118, %r200, 0;
|
|
@%p118 bra BB0_111;
|
|
|
|
mov.u64 %rd280, image_HDR;
|
|
cvta.global.u64 %rd269, %rd280;
|
|
mov.u32 %r204, 8;
|
|
// inline asm
|
|
call (%rd268), _rt_buffer_get_64, (%rd269, %r24, %r204, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs162, %rs163, %rs164, %rs165}, [%rd268];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f793, %rs162;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f794, %rs163;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f795, %rs164;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd274), _rt_buffer_get_64, (%rd269, %r24, %r204, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f796, %f793, 0f00000000;
|
|
add.f32 %f797, %f794, 0f00000000;
|
|
add.f32 %f798, %f795, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs161, %f798;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs160, %f797;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs159, %f796;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs166, 0;
|
|
st.v4.u16 [%rd274], {%rs159, %rs160, %rs161, %rs166};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r41, %r42}, [pixelID];
|
|
cvt.u64.u32 %rd19, %r41;
|
|
cvt.u64.u32 %rd20, %r42;
|
|
mov.u64 %rd23, uvpos;
|
|
cvta.global.u64 %rd18, %rd23;
|
|
mov.u32 %r40, 12;
|
|
// inline asm
|
|
call (%rd17), _rt_buffer_get_64, (%rd18, %r24, %r40, %rd19, %rd20, %rd15, %rd15);
|
|
// inline asm
|
|
ld.global.f32 %f122, [lightPos];
|
|
ld.f32 %f9, [%rd17+8];
|
|
ld.f32 %f8, [%rd17+4];
|
|
ld.f32 %f7, [%rd17];
|
|
sub.f32 %f123, %f122, %f7;
|
|
ld.global.f32 %f124, [lightPos+4];
|
|
sub.f32 %f125, %f124, %f8;
|
|
ld.global.f32 %f126, [lightPos+8];
|
|
sub.f32 %f127, %f126, %f9;
|
|
mul.f32 %f128, %f125, %f125;
|
|
fma.rn.f32 %f129, %f123, %f123, %f128;
|
|
fma.rn.f32 %f130, %f127, %f127, %f129;
|
|
sqrt.rn.f32 %f131, %f130;
|
|
rcp.rn.f32 %f132, %f131;
|
|
mul.f32 %f10, %f123, %f132;
|
|
mul.f32 %f11, %f125, %f132;
|
|
mul.f32 %f12, %f127, %f132;
|
|
ld.global.f32 %f133, [lightFalloffFakeDistanceMult];
|
|
mul.f32 %f13, %f131, %f133;
|
|
ld.global.f32 %f134, [lightInvCutoff];
|
|
mul.f32 %f14, %f131, %f134;
|
|
mov.f32 %f138, 0f40800000;
|
|
abs.f32 %f16, %f14;
|
|
setp.lt.f32 %p12, %f16, 0f00800000;
|
|
mul.f32 %f140, %f16, 0f4B800000;
|
|
selp.f32 %f141, 0fC3170000, 0fC2FE0000, %p12;
|
|
selp.f32 %f142, %f140, %f16, %p12;
|
|
mov.b32 %r45, %f142;
|
|
and.b32 %r46, %r45, 8388607;
|
|
or.b32 %r47, %r46, 1065353216;
|
|
mov.b32 %f143, %r47;
|
|
shr.u32 %r48, %r45, 23;
|
|
cvt.rn.f32.u32 %f144, %r48;
|
|
add.f32 %f145, %f141, %f144;
|
|
setp.gt.f32 %p13, %f143, 0f3FB504F3;
|
|
mul.f32 %f146, %f143, 0f3F000000;
|
|
add.f32 %f147, %f145, 0f3F800000;
|
|
selp.f32 %f148, %f146, %f143, %p13;
|
|
selp.f32 %f149, %f147, %f145, %p13;
|
|
add.f32 %f150, %f148, 0fBF800000;
|
|
add.f32 %f121, %f148, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f120,%f121;
|
|
// inline asm
|
|
add.f32 %f151, %f150, %f150;
|
|
mul.f32 %f152, %f120, %f151;
|
|
mul.f32 %f153, %f152, %f152;
|
|
mov.f32 %f154, 0f3C4CAF63;
|
|
mov.f32 %f155, 0f3B18F0FE;
|
|
fma.rn.f32 %f156, %f155, %f153, %f154;
|
|
mov.f32 %f157, 0f3DAAAABD;
|
|
fma.rn.f32 %f158, %f156, %f153, %f157;
|
|
mul.rn.f32 %f159, %f158, %f153;
|
|
mul.rn.f32 %f160, %f159, %f152;
|
|
sub.f32 %f161, %f150, %f152;
|
|
neg.f32 %f162, %f152;
|
|
add.f32 %f163, %f161, %f161;
|
|
fma.rn.f32 %f164, %f162, %f150, %f163;
|
|
mul.rn.f32 %f165, %f120, %f164;
|
|
add.f32 %f166, %f160, %f152;
|
|
sub.f32 %f167, %f152, %f166;
|
|
add.f32 %f168, %f160, %f167;
|
|
add.f32 %f169, %f165, %f168;
|
|
add.f32 %f170, %f166, %f169;
|
|
sub.f32 %f171, %f166, %f170;
|
|
add.f32 %f172, %f169, %f171;
|
|
mov.f32 %f173, 0f3F317200;
|
|
mul.rn.f32 %f174, %f149, %f173;
|
|
mov.f32 %f175, 0f35BFBE8E;
|
|
mul.rn.f32 %f176, %f149, %f175;
|
|
add.f32 %f177, %f174, %f170;
|
|
sub.f32 %f178, %f174, %f177;
|
|
add.f32 %f179, %f170, %f178;
|
|
add.f32 %f180, %f172, %f179;
|
|
add.f32 %f181, %f176, %f180;
|
|
add.f32 %f182, %f177, %f181;
|
|
sub.f32 %f183, %f177, %f182;
|
|
add.f32 %f184, %f181, %f183;
|
|
mul.rn.f32 %f185, %f138, %f182;
|
|
neg.f32 %f186, %f185;
|
|
fma.rn.f32 %f187, %f138, %f182, %f186;
|
|
fma.rn.f32 %f188, %f138, %f184, %f187;
|
|
mov.f32 %f189, 0f00000000;
|
|
fma.rn.f32 %f190, %f189, %f182, %f188;
|
|
add.rn.f32 %f191, %f185, %f190;
|
|
neg.f32 %f192, %f191;
|
|
add.rn.f32 %f193, %f185, %f192;
|
|
add.rn.f32 %f194, %f193, %f190;
|
|
mov.b32 %r49, %f191;
|
|
setp.eq.s32 %p14, %r49, 1118925336;
|
|
add.s32 %r50, %r49, -1;
|
|
mov.b32 %f195, %r50;
|
|
add.f32 %f196, %f194, 0f37000000;
|
|
selp.f32 %f197, %f195, %f191, %p14;
|
|
selp.f32 %f17, %f196, %f194, %p14;
|
|
mul.f32 %f198, %f197, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f199, %f198;
|
|
mov.f32 %f200, 0fBF317200;
|
|
fma.rn.f32 %f201, %f199, %f200, %f197;
|
|
mov.f32 %f202, 0fB5BFBE8E;
|
|
fma.rn.f32 %f203, %f199, %f202, %f201;
|
|
mul.f32 %f204, %f203, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f205, %f204;
|
|
add.f32 %f206, %f199, 0f00000000;
|
|
ex2.approx.f32 %f207, %f206;
|
|
mul.f32 %f208, %f205, %f207;
|
|
setp.lt.f32 %p15, %f197, 0fC2D20000;
|
|
selp.f32 %f209, 0f00000000, %f208, %p15;
|
|
setp.gt.f32 %p16, %f197, 0f42D20000;
|
|
selp.f32 %f879, 0f7F800000, %f209, %p16;
|
|
setp.eq.f32 %p17, %f879, 0f7F800000;
|
|
@%p17 bra BB0_5;
|
|
|
|
fma.rn.f32 %f879, %f879, %f17, %f879;
|
|
|
|
BB0_5:
|
|
mov.f32 %f833, 0f40000000;
|
|
cvt.rzi.f32.f32 %f832, %f833;
|
|
add.f32 %f831, %f832, %f832;
|
|
mov.f32 %f830, 0f40800000;
|
|
sub.f32 %f829, %f830, %f831;
|
|
abs.f32 %f828, %f829;
|
|
setp.lt.f32 %p18, %f14, 0f00000000;
|
|
setp.eq.f32 %p19, %f828, 0f3F800000;
|
|
and.pred %p1, %p18, %p19;
|
|
mov.b32 %r51, %f879;
|
|
xor.b32 %r52, %r51, -2147483648;
|
|
mov.b32 %f210, %r52;
|
|
selp.f32 %f881, %f210, %f879, %p1;
|
|
setp.eq.f32 %p20, %f14, 0f00000000;
|
|
@%p20 bra BB0_8;
|
|
bra.uni BB0_6;
|
|
|
|
BB0_8:
|
|
add.f32 %f213, %f14, %f14;
|
|
selp.f32 %f881, %f213, 0f00000000, %p19;
|
|
bra.uni BB0_9;
|
|
|
|
BB0_6:
|
|
setp.geu.f32 %p21, %f14, 0f00000000;
|
|
@%p21 bra BB0_9;
|
|
|
|
mov.f32 %f867, 0f40800000;
|
|
cvt.rzi.f32.f32 %f212, %f867;
|
|
setp.neu.f32 %p22, %f212, 0f40800000;
|
|
selp.f32 %f881, 0f7FFFFFFF, %f881, %p22;
|
|
|
|
BB0_9:
|
|
abs.f32 %f834, %f14;
|
|
add.f32 %f214, %f834, 0f40800000;
|
|
mov.b32 %r53, %f214;
|
|
setp.lt.s32 %p24, %r53, 2139095040;
|
|
@%p24 bra BB0_14;
|
|
|
|
abs.f32 %f865, %f14;
|
|
setp.gtu.f32 %p25, %f865, 0f7F800000;
|
|
@%p25 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f881, %f14, 0f40800000;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
abs.f32 %f866, %f14;
|
|
setp.neu.f32 %p26, %f866, 0f7F800000;
|
|
@%p26 bra BB0_14;
|
|
|
|
selp.f32 %f881, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_14:
|
|
mov.f32 %f835, 0f00000000;
|
|
mov.f32 %f884, 0f3F800000;
|
|
sub.f32 %f224, %f884, %f881;
|
|
setp.eq.f32 %p27, %f14, 0f3F800000;
|
|
selp.f32 %f225, 0f00000000, %f224, %p27;
|
|
cvt.sat.f32.f32 %f226, %f225;
|
|
ld.global.f32 %f227, [lightFalloffMinRadiusSq];
|
|
fma.rn.f32 %f228, %f13, %f13, %f227;
|
|
div.rn.f32 %f28, %f226, %f228;
|
|
mul.f32 %f229, %f877, %f11;
|
|
fma.rn.f32 %f230, %f876, %f10, %f229;
|
|
fma.rn.f32 %f231, %f878, %f12, %f230;
|
|
ld.global.u32 %r56, [imageEnabled];
|
|
and.b32 %r57, %r56, 32;
|
|
ld.global.u32 %r58, [ignoreNormal];
|
|
or.b32 %r59, %r57, %r58;
|
|
setp.eq.s32 %p28, %r59, 0;
|
|
selp.f32 %f29, %f231, 0f3F800000, %p28;
|
|
ld.global.f32 %f232, [lightMatrix+24];
|
|
mul.f32 %f233, %f10, %f232;
|
|
ld.global.f32 %f234, [lightMatrix+28];
|
|
mul.f32 %f235, %f11, %f234;
|
|
neg.f32 %f236, %f235;
|
|
sub.f32 %f237, %f236, %f233;
|
|
ld.global.f32 %f238, [lightMatrix+32];
|
|
mul.f32 %f239, %f12, %f238;
|
|
sub.f32 %f240, %f237, %f239;
|
|
setp.gt.f32 %p29, %f240, 0f00000000;
|
|
ld.global.f32 %f241, [lightMatrix];
|
|
mul.f32 %f242, %f241, %f10;
|
|
sub.f32 %f243, %f835, %f242;
|
|
ld.global.f32 %f244, [lightMatrix+4];
|
|
mul.f32 %f245, %f244, %f11;
|
|
sub.f32 %f246, %f243, %f245;
|
|
ld.global.f32 %f247, [lightMatrix+8];
|
|
mul.f32 %f248, %f247, %f12;
|
|
sub.f32 %f249, %f246, %f248;
|
|
selp.f32 %f250, 0f3F800000, 0f00000000, %p29;
|
|
ld.global.f32 %f251, [lightMatrix+12];
|
|
mul.f32 %f252, %f251, %f10;
|
|
sub.f32 %f253, %f835, %f252;
|
|
ld.global.f32 %f254, [lightMatrix+16];
|
|
mul.f32 %f255, %f254, %f11;
|
|
sub.f32 %f256, %f253, %f255;
|
|
ld.global.f32 %f257, [lightMatrix+20];
|
|
mul.f32 %f258, %f257, %f12;
|
|
sub.f32 %f259, %f256, %f258;
|
|
ld.global.u32 %r54, [lightCookie];
|
|
ld.global.f32 %f260, [lightFOV];
|
|
fma.rn.f32 %f261, %f249, %f260, 0f3F000000;
|
|
sub.f32 %f219, %f884, %f261;
|
|
fma.rn.f32 %f220, %f259, %f260, 0f3F000000;
|
|
// inline asm
|
|
call (%f215, %f216, %f217, %f218), _rt_texture_get_f_id, (%r54, %r24, %f219, %f220, %f835, %f835);
|
|
// inline asm
|
|
max.f32 %f262, %f215, %f216;
|
|
max.f32 %f263, %f262, %f217;
|
|
mul.f32 %f264, %f250, %f263;
|
|
mul.f32 %f265, %f28, %f264;
|
|
ld.global.f32 %f266, [lightColor+4];
|
|
ld.global.f32 %f267, [lightColor];
|
|
max.f32 %f268, %f267, %f266;
|
|
ld.global.f32 %f269, [lightColor+8];
|
|
max.f32 %f270, %f268, %f269;
|
|
mul.f32 %f271, %f265, %f270;
|
|
setp.lt.f32 %p30, %f271, 0f3727C5AC;
|
|
@%p30 bra BB0_84;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_84:
|
|
ld.global.u32 %r241, [imageEnabled];
|
|
and.b32 %r149, %r241, 1;
|
|
setp.eq.b32 %p107, %r149, 1;
|
|
@!%p107 bra BB0_86;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_85:
|
|
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, %r24, %r25, %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 %r241, [imageEnabled];
|
|
|
|
BB0_86:
|
|
and.b32 %r152, %r241, 8;
|
|
setp.eq.s32 %p108, %r152, 0;
|
|
@%p108 bra BB0_88;
|
|
|
|
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, %r24, %r24, %rd149, %rd150, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f749, 0f00000000;
|
|
cvt.rzi.u32.f32 %r155, %f749;
|
|
cvt.u16.u32 %rs91, %r155;
|
|
mov.u16 %rs92, 255;
|
|
st.v2.u8 [%rd147], {%rs91, %rs92};
|
|
ld.global.u32 %r241, [imageEnabled];
|
|
|
|
BB0_88:
|
|
cvt.u64.u32 %rd6, %r2;
|
|
cvt.u64.u32 %rd7, %r3;
|
|
and.b32 %r156, %r241, 4;
|
|
setp.eq.s32 %p109, %r156, 0;
|
|
@%p109 bra BB0_92;
|
|
|
|
ld.global.u32 %r157, [additive];
|
|
setp.eq.s32 %p110, %r157, 0;
|
|
mov.f32 %f750, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f750;}
|
|
|
|
// inline asm
|
|
@%p110 bra BB0_91;
|
|
|
|
mov.u64 %rd166, image_HDR;
|
|
cvta.global.u64 %rd155, %rd166;
|
|
mov.u32 %r161, 8;
|
|
// inline asm
|
|
call (%rd154), _rt_buffer_get_64, (%rd155, %r24, %r161, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd154];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f751, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f752, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f753, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd160), _rt_buffer_get_64, (%rd155, %r24, %r161, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f754, %f751, 0f00000000;
|
|
add.f32 %f755, %f752, 0f00000000;
|
|
add.f32 %f756, %f753, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f756;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f755;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f754;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd160], {%rs97, %rs98, %rs99, %rs93};
|
|
bra.uni BB0_92;
|
|
|
|
BB0_15:
|
|
mov.f32 %f883, 0f00000000;
|
|
mul.f32 %f273, %f7, 0f3456BF95;
|
|
abs.f32 %f274, %f876;
|
|
div.rn.f32 %f275, %f273, %f274;
|
|
abs.f32 %f276, %f877;
|
|
mul.f32 %f277, %f8, 0f3456BF95;
|
|
div.rn.f32 %f278, %f277, %f276;
|
|
abs.f32 %f279, %f878;
|
|
mul.f32 %f280, %f9, 0f3456BF95;
|
|
div.rn.f32 %f281, %f280, %f279;
|
|
abs.f32 %f282, %f275;
|
|
abs.f32 %f283, %f278;
|
|
abs.f32 %f284, %f281;
|
|
mov.f32 %f285, 0f38D1B717;
|
|
max.f32 %f286, %f282, %f285;
|
|
max.f32 %f287, %f283, %f285;
|
|
max.f32 %f288, %f284, %f285;
|
|
fma.rn.f32 %f33, %f876, %f286, %f7;
|
|
fma.rn.f32 %f34, %f877, %f287, %f8;
|
|
fma.rn.f32 %f35, %f878, %f288, %f9;
|
|
ld.global.u32 %r237, [samples];
|
|
setp.lt.s32 %p31, %r237, 1;
|
|
@%p31 bra BB0_18;
|
|
|
|
mul.f32 %f290, %f33, 0f3456BF95;
|
|
abs.f32 %f291, %f290;
|
|
mul.f32 %f292, %f34, 0f3456BF95;
|
|
abs.f32 %f293, %f292;
|
|
mul.f32 %f294, %f35, 0f3456BF95;
|
|
abs.f32 %f295, %f294;
|
|
max.f32 %f296, %f291, %f293;
|
|
max.f32 %f297, %f296, %f295;
|
|
max.f32 %f36, %f297, %f285;
|
|
add.u64 %rd24, %SP, 0;
|
|
cvta.to.local.u64 %rd2, %rd24;
|
|
mov.f32 %f883, 0f00000000;
|
|
mov.u32 %r236, 0;
|
|
|
|
BB0_17:
|
|
cvt.rn.f32.s32 %f307, %r236;
|
|
mul.f32 %f308, %f307, 0f3DD32618;
|
|
cvt.rmi.f32.f32 %f309, %f308;
|
|
sub.f32 %f310, %f308, %f309;
|
|
mul.f32 %f311, %f307, 0f3DD2F1AA;
|
|
cvt.rmi.f32.f32 %f312, %f311;
|
|
sub.f32 %f313, %f311, %f312;
|
|
mul.f32 %f314, %f307, 0f3DC74539;
|
|
cvt.rmi.f32.f32 %f315, %f314;
|
|
sub.f32 %f316, %f314, %f315;
|
|
add.f32 %f317, %f313, 0f4199851F;
|
|
add.f32 %f318, %f316, 0f4199851F;
|
|
add.f32 %f319, %f310, 0f4199851F;
|
|
mul.f32 %f320, %f313, %f318;
|
|
fma.rn.f32 %f321, %f310, %f317, %f320;
|
|
fma.rn.f32 %f322, %f319, %f316, %f321;
|
|
add.f32 %f323, %f310, %f322;
|
|
add.f32 %f324, %f313, %f322;
|
|
add.f32 %f325, %f316, %f322;
|
|
add.f32 %f326, %f323, %f324;
|
|
mul.f32 %f327, %f325, %f326;
|
|
cvt.rmi.f32.f32 %f328, %f327;
|
|
sub.f32 %f329, %f327, %f328;
|
|
add.f32 %f330, %f323, %f325;
|
|
mul.f32 %f331, %f324, %f330;
|
|
cvt.rmi.f32.f32 %f332, %f331;
|
|
sub.f32 %f333, %f331, %f332;
|
|
add.f32 %f334, %f324, %f325;
|
|
mul.f32 %f335, %f323, %f334;
|
|
cvt.rmi.f32.f32 %f336, %f335;
|
|
sub.f32 %f337, %f335, %f336;
|
|
fma.rn.f32 %f338, %f329, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f339, %f333, 0f40000000, 0fBF800000;
|
|
fma.rn.f32 %f340, %f337, 0f40000000, 0fBF800000;
|
|
ld.global.f32 %f341, [lightRadius];
|
|
ld.global.f32 %f342, [lightPos];
|
|
fma.rn.f32 %f343, %f341, %f338, %f342;
|
|
ld.global.f32 %f344, [lightPos+4];
|
|
fma.rn.f32 %f345, %f341, %f339, %f344;
|
|
ld.global.f32 %f346, [lightPos+8];
|
|
fma.rn.f32 %f347, %f341, %f340, %f346;
|
|
sub.f32 %f348, %f343, %f7;
|
|
sub.f32 %f349, %f345, %f8;
|
|
sub.f32 %f350, %f347, %f9;
|
|
mul.f32 %f351, %f349, %f349;
|
|
fma.rn.f32 %f352, %f348, %f348, %f351;
|
|
fma.rn.f32 %f353, %f350, %f350, %f352;
|
|
sqrt.rn.f32 %f306, %f353;
|
|
rcp.rn.f32 %f354, %f306;
|
|
mul.f32 %f302, %f354, %f348;
|
|
mul.f32 %f303, %f354, %f349;
|
|
mul.f32 %f304, %f354, %f350;
|
|
ld.global.u32 %r64, [imageEnabled];
|
|
and.b32 %r65, %r64, 32;
|
|
setp.eq.s32 %p32, %r65, 0;
|
|
selp.f32 %f355, 0f3F800000, 0f41200000, %p32;
|
|
mul.f32 %f305, %f355, %f36;
|
|
mov.u32 %r66, 1065353216;
|
|
st.local.u32 [%rd2], %r66;
|
|
ld.global.u32 %r61, [root];
|
|
mov.u32 %r62, 1;
|
|
// inline asm
|
|
call _rt_trace_64, (%r61, %f33, %f34, %f35, %f302, %f303, %f304, %r62, %f305, %f306, %rd24, %r25);
|
|
// inline asm
|
|
ld.local.f32 %f356, [%rd2];
|
|
add.f32 %f883, %f883, %f356;
|
|
ld.global.u32 %r237, [samples];
|
|
add.s32 %r236, %r236, 1;
|
|
setp.lt.s32 %p33, %r236, %r237;
|
|
@%p33 bra BB0_17;
|
|
|
|
BB0_18:
|
|
setp.eq.s32 %p34, %r237, 0;
|
|
@%p34 bra BB0_20;
|
|
|
|
cvt.rn.f32.s32 %f358, %r237;
|
|
div.rn.f32 %f884, %f883, %f358;
|
|
|
|
BB0_20:
|
|
cvt.sat.f32.f32 %f359, %f29;
|
|
mul.f32 %f360, %f28, %f359;
|
|
mul.f32 %f361, %f884, %f360;
|
|
ld.global.f32 %f362, [lightColor];
|
|
mul.f32 %f363, %f362, %f361;
|
|
ld.global.f32 %f364, [lightColor+4];
|
|
mul.f32 %f365, %f364, %f361;
|
|
ld.global.f32 %f366, [lightColor+8];
|
|
mul.f32 %f367, %f361, %f366;
|
|
mul.f32 %f42, %f215, %f363;
|
|
mul.f32 %f43, %f216, %f365;
|
|
mul.f32 %f44, %f217, %f367;
|
|
ld.global.u32 %r239, [imageEnabled];
|
|
and.b32 %r67, %r239, 8;
|
|
setp.eq.s32 %p35, %r67, 0;
|
|
@%p35 bra BB0_33;
|
|
|
|
mov.f32 %f843, 0fB5BFBE8E;
|
|
mov.f32 %f842, 0fBF317200;
|
|
mov.f32 %f841, 0f35BFBE8E;
|
|
mov.f32 %f840, 0f3F317200;
|
|
mov.f32 %f839, 0f3DAAAABD;
|
|
mov.f32 %f838, 0f3C4CAF63;
|
|
mov.f32 %f837, 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, %r24, %r24, %rd28, %rd29, %rd15, %rd15);
|
|
// inline asm
|
|
abs.f32 %f46, %f884;
|
|
setp.lt.f32 %p36, %f46, 0f00800000;
|
|
mul.f32 %f373, %f46, 0f4B800000;
|
|
selp.f32 %f374, 0fC3170000, 0fC2FE0000, %p36;
|
|
selp.f32 %f375, %f373, %f46, %p36;
|
|
mov.b32 %r70, %f375;
|
|
and.b32 %r71, %r70, 8388607;
|
|
or.b32 %r72, %r71, 1065353216;
|
|
mov.b32 %f376, %r72;
|
|
shr.u32 %r73, %r70, 23;
|
|
cvt.rn.f32.u32 %f377, %r73;
|
|
add.f32 %f378, %f374, %f377;
|
|
setp.gt.f32 %p37, %f376, 0f3FB504F3;
|
|
mul.f32 %f379, %f376, 0f3F000000;
|
|
add.f32 %f380, %f378, 0f3F800000;
|
|
selp.f32 %f381, %f379, %f376, %p37;
|
|
selp.f32 %f382, %f380, %f378, %p37;
|
|
add.f32 %f383, %f381, 0fBF800000;
|
|
add.f32 %f369, %f381, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f368,%f369;
|
|
// inline asm
|
|
add.f32 %f384, %f383, %f383;
|
|
mul.f32 %f385, %f368, %f384;
|
|
mul.f32 %f386, %f385, %f385;
|
|
fma.rn.f32 %f389, %f837, %f386, %f838;
|
|
fma.rn.f32 %f391, %f389, %f386, %f839;
|
|
mul.rn.f32 %f392, %f391, %f386;
|
|
mul.rn.f32 %f393, %f392, %f385;
|
|
sub.f32 %f394, %f383, %f385;
|
|
neg.f32 %f395, %f385;
|
|
add.f32 %f396, %f394, %f394;
|
|
fma.rn.f32 %f397, %f395, %f383, %f396;
|
|
mul.rn.f32 %f398, %f368, %f397;
|
|
add.f32 %f399, %f393, %f385;
|
|
sub.f32 %f400, %f385, %f399;
|
|
add.f32 %f401, %f393, %f400;
|
|
add.f32 %f402, %f398, %f401;
|
|
add.f32 %f403, %f399, %f402;
|
|
sub.f32 %f404, %f399, %f403;
|
|
add.f32 %f405, %f402, %f404;
|
|
mul.rn.f32 %f407, %f382, %f840;
|
|
mul.rn.f32 %f409, %f382, %f841;
|
|
add.f32 %f410, %f407, %f403;
|
|
sub.f32 %f411, %f407, %f410;
|
|
add.f32 %f412, %f403, %f411;
|
|
add.f32 %f413, %f405, %f412;
|
|
add.f32 %f414, %f409, %f413;
|
|
add.f32 %f415, %f410, %f414;
|
|
sub.f32 %f416, %f410, %f415;
|
|
add.f32 %f417, %f414, %f416;
|
|
mov.f32 %f418, 0f3EE8BA2E;
|
|
mul.rn.f32 %f419, %f418, %f415;
|
|
neg.f32 %f420, %f419;
|
|
fma.rn.f32 %f421, %f418, %f415, %f420;
|
|
fma.rn.f32 %f422, %f418, %f417, %f421;
|
|
mov.f32 %f423, 0f00000000;
|
|
fma.rn.f32 %f424, %f423, %f415, %f422;
|
|
add.rn.f32 %f425, %f419, %f424;
|
|
neg.f32 %f426, %f425;
|
|
add.rn.f32 %f427, %f419, %f426;
|
|
add.rn.f32 %f428, %f427, %f424;
|
|
mov.b32 %r74, %f425;
|
|
setp.eq.s32 %p38, %r74, 1118925336;
|
|
add.s32 %r75, %r74, -1;
|
|
mov.b32 %f429, %r75;
|
|
add.f32 %f430, %f428, 0f37000000;
|
|
selp.f32 %f431, %f429, %f425, %p38;
|
|
selp.f32 %f47, %f430, %f428, %p38;
|
|
mul.f32 %f432, %f431, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f433, %f432;
|
|
fma.rn.f32 %f435, %f433, %f842, %f431;
|
|
fma.rn.f32 %f437, %f433, %f843, %f435;
|
|
mul.f32 %f438, %f437, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f439, %f438;
|
|
add.f32 %f440, %f433, 0f00000000;
|
|
ex2.approx.f32 %f441, %f440;
|
|
mul.f32 %f442, %f439, %f441;
|
|
setp.lt.f32 %p39, %f431, 0fC2D20000;
|
|
selp.f32 %f443, 0f00000000, %f442, %p39;
|
|
setp.gt.f32 %p40, %f431, 0f42D20000;
|
|
selp.f32 %f885, 0f7F800000, %f443, %p40;
|
|
setp.eq.f32 %p41, %f885, 0f7F800000;
|
|
@%p41 bra BB0_23;
|
|
|
|
fma.rn.f32 %f885, %f885, %f47, %f885;
|
|
|
|
BB0_23:
|
|
mov.f32 %f871, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f870, %f871;
|
|
fma.rn.f32 %f869, %f870, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f868, %f869;
|
|
setp.lt.f32 %p42, %f884, 0f00000000;
|
|
setp.eq.f32 %p43, %f868, 0f3F800000;
|
|
and.pred %p2, %p42, %p43;
|
|
mov.b32 %r76, %f885;
|
|
xor.b32 %r77, %r76, -2147483648;
|
|
mov.b32 %f444, %r77;
|
|
selp.f32 %f887, %f444, %f885, %p2;
|
|
setp.eq.f32 %p44, %f884, 0f00000000;
|
|
@%p44 bra BB0_26;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_26:
|
|
add.f32 %f447, %f884, %f884;
|
|
selp.f32 %f887, %f447, 0f00000000, %p43;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_111:
|
|
mov.u64 %rd287, image_HDR;
|
|
cvta.global.u64 %rd282, %rd287;
|
|
mov.u32 %r206, 8;
|
|
// inline asm
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r24, %r206, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f799, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs167, %f799;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs168, 0;
|
|
st.v4.u16 [%rd281], {%rs167, %rs167, %rs167, %rs168};
|
|
|
|
BB0_112:
|
|
ld.global.u32 %r207, [additive];
|
|
setp.eq.s32 %p119, %r207, 0;
|
|
@%p119 bra BB0_114;
|
|
|
|
mov.u64 %rd300, image_RNM0;
|
|
cvta.global.u64 %rd289, %rd300;
|
|
mov.u32 %r211, 8;
|
|
// inline asm
|
|
call (%rd288), _rt_buffer_get_64, (%rd289, %r24, %r211, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs175, %rs176, %rs177, %rs178}, [%rd288];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f800, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f801, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f802, %rs177;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd294), _rt_buffer_get_64, (%rd289, %r24, %r211, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f803, %f800, 0f00000000;
|
|
add.f32 %f804, %f801, 0f00000000;
|
|
add.f32 %f805, %f802, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f805;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f804;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f803;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs179, 0;
|
|
st.v4.u16 [%rd294], {%rs172, %rs173, %rs174, %rs179};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
mov.u64 %rd307, image_RNM0;
|
|
cvta.global.u64 %rd302, %rd307;
|
|
mov.u32 %r213, 8;
|
|
// inline asm
|
|
call (%rd301), _rt_buffer_get_64, (%rd302, %r24, %r213, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f806, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f806;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs181, 0;
|
|
st.v4.u16 [%rd301], {%rs180, %rs180, %rs180, %rs181};
|
|
|
|
BB0_115:
|
|
ld.global.u32 %r214, [additive];
|
|
setp.eq.s32 %p120, %r214, 0;
|
|
@%p120 bra BB0_117;
|
|
|
|
mov.u64 %rd320, image_RNM1;
|
|
cvta.global.u64 %rd309, %rd320;
|
|
mov.u32 %r218, 8;
|
|
// inline asm
|
|
call (%rd308), _rt_buffer_get_64, (%rd309, %r24, %r218, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs188, %rs189, %rs190, %rs191}, [%rd308];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f807, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f808, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f809, %rs190;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd314), _rt_buffer_get_64, (%rd309, %r24, %r218, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f810, %f807, 0f00000000;
|
|
add.f32 %f811, %f808, 0f00000000;
|
|
add.f32 %f812, %f809, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs187, %f812;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f811;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f810;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs192, 0;
|
|
st.v4.u16 [%rd314], {%rs185, %rs186, %rs187, %rs192};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
mov.u64 %rd327, image_RNM1;
|
|
cvta.global.u64 %rd322, %rd327;
|
|
mov.u32 %r220, 8;
|
|
// inline asm
|
|
call (%rd321), _rt_buffer_get_64, (%rd322, %r24, %r220, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f813, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs193, %f813;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs194, 0;
|
|
st.v4.u16 [%rd321], {%rs193, %rs193, %rs193, %rs194};
|
|
|
|
BB0_118:
|
|
ld.global.u32 %r221, [additive];
|
|
setp.eq.s32 %p121, %r221, 0;
|
|
@%p121 bra BB0_120;
|
|
|
|
mov.u64 %rd340, image_RNM2;
|
|
cvta.global.u64 %rd329, %rd340;
|
|
mov.u32 %r225, 8;
|
|
// inline asm
|
|
call (%rd328), _rt_buffer_get_64, (%rd329, %r24, %r225, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs201, %rs202, %rs203, %rs204}, [%rd328];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f814, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f815, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f816, %rs203;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd334), _rt_buffer_get_64, (%rd329, %r24, %r225, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f817, %f814, 0f00000000;
|
|
add.f32 %f818, %f815, 0f00000000;
|
|
add.f32 %f819, %f816, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs200, %f819;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f818;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f817;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs205, 0;
|
|
st.v4.u16 [%rd334], {%rs198, %rs199, %rs200, %rs205};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_120:
|
|
mov.u64 %rd347, image_RNM2;
|
|
cvta.global.u64 %rd342, %rd347;
|
|
mov.u32 %r227, 8;
|
|
// inline asm
|
|
call (%rd341), _rt_buffer_get_64, (%rd342, %r24, %r227, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f820, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f820;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs207, 0;
|
|
st.v4.u16 [%rd341], {%rs206, %rs206, %rs206, %rs207};
|
|
|
|
BB0_121:
|
|
ld.global.u32 %r228, [additive];
|
|
setp.eq.s32 %p122, %r228, 0;
|
|
@%p122 bra BB0_123;
|
|
|
|
mov.u64 %rd360, image_RNM3;
|
|
cvta.global.u64 %rd349, %rd360;
|
|
mov.u32 %r232, 8;
|
|
// inline asm
|
|
call (%rd348), _rt_buffer_get_64, (%rd349, %r24, %r232, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs214, %rs215, %rs216, %rs217}, [%rd348];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f821, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f822, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f823, %rs216;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd354), _rt_buffer_get_64, (%rd349, %r24, %r232, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f824, %f821, 0f00000000;
|
|
add.f32 %f825, %f822, 0f00000000;
|
|
add.f32 %f826, %f823, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs213, %f826;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f825;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f824;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs218, 0;
|
|
st.v4.u16 [%rd354], {%rs211, %rs212, %rs213, %rs218};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_123:
|
|
mov.u64 %rd367, image_RNM3;
|
|
cvta.global.u64 %rd362, %rd367;
|
|
mov.u32 %r234, 8;
|
|
// inline asm
|
|
call (%rd361), _rt_buffer_get_64, (%rd362, %r24, %r234, %rd8, %rd9, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f827, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs219, %f827;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs220, 0;
|
|
st.v4.u16 [%rd361], {%rs219, %rs219, %rs219, %rs220};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_91:
|
|
mov.u64 %rd173, image_HDR;
|
|
cvta.global.u64 %rd168, %rd173;
|
|
mov.u32 %r163, 8;
|
|
// inline asm
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r24, %r163, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f757, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs104, %f757;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd167], {%rs104, %rs104, %rs104, %rs93};
|
|
|
|
BB0_92:
|
|
ld.global.u32 %r164, [additive];
|
|
setp.eq.s32 %p111, %r164, 0;
|
|
mov.f32 %f758, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f758;}
|
|
|
|
// inline asm
|
|
@%p111 bra BB0_94;
|
|
|
|
mov.u64 %rd186, image_RNM0;
|
|
cvta.global.u64 %rd175, %rd186;
|
|
mov.u32 %r168, 8;
|
|
// inline asm
|
|
call (%rd174), _rt_buffer_get_64, (%rd175, %r24, %r168, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs112, %rs113, %rs114, %rs115}, [%rd174];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f759, %rs112;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f760, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f761, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd180), _rt_buffer_get_64, (%rd175, %r24, %r168, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f762, %f759, 0f00000000;
|
|
add.f32 %f763, %f760, 0f00000000;
|
|
add.f32 %f764, %f761, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f764;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f763;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f762;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd180], {%rs109, %rs110, %rs111, %rs105};
|
|
bra.uni BB0_95;
|
|
|
|
BB0_94:
|
|
mov.u64 %rd193, image_RNM0;
|
|
cvta.global.u64 %rd188, %rd193;
|
|
mov.u32 %r170, 8;
|
|
// inline asm
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r24, %r170, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f765, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f765;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd187], {%rs116, %rs116, %rs116, %rs105};
|
|
|
|
BB0_95:
|
|
ld.global.u32 %r171, [additive];
|
|
setp.eq.s32 %p112, %r171, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs117, %f758;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_97;
|
|
|
|
mov.u64 %rd206, image_RNM1;
|
|
cvta.global.u64 %rd195, %rd206;
|
|
mov.u32 %r175, 8;
|
|
// inline asm
|
|
call (%rd194), _rt_buffer_get_64, (%rd195, %r24, %r175, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs124, %rs125, %rs126, %rs127}, [%rd194];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f767, %rs124;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f768, %rs125;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f769, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd200), _rt_buffer_get_64, (%rd195, %r24, %r175, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f770, %f767, 0f00000000;
|
|
add.f32 %f771, %f768, 0f00000000;
|
|
add.f32 %f772, %f769, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f772;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f771;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f770;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd200], {%rs121, %rs122, %rs123, %rs117};
|
|
bra.uni BB0_98;
|
|
|
|
BB0_97:
|
|
mov.u64 %rd213, image_RNM1;
|
|
cvta.global.u64 %rd208, %rd213;
|
|
mov.u32 %r177, 8;
|
|
// inline asm
|
|
call (%rd207), _rt_buffer_get_64, (%rd208, %r24, %r177, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f773, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f773;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd207], {%rs128, %rs128, %rs128, %rs117};
|
|
|
|
BB0_98:
|
|
ld.global.u32 %r178, [additive];
|
|
setp.eq.s32 %p113, %r178, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f758;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_100;
|
|
|
|
mov.u64 %rd226, image_RNM2;
|
|
cvta.global.u64 %rd215, %rd226;
|
|
mov.u32 %r182, 8;
|
|
// inline asm
|
|
call (%rd214), _rt_buffer_get_64, (%rd215, %r24, %r182, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs136, %rs137, %rs138, %rs139}, [%rd214];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f775, %rs136;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f776, %rs137;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f777, %rs138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd220), _rt_buffer_get_64, (%rd215, %r24, %r182, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f778, %f775, 0f00000000;
|
|
add.f32 %f779, %f776, 0f00000000;
|
|
add.f32 %f780, %f777, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f780;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f779;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs133, %f778;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd220], {%rs133, %rs134, %rs135, %rs129};
|
|
bra.uni BB0_101;
|
|
|
|
BB0_100:
|
|
mov.u64 %rd233, image_RNM2;
|
|
cvta.global.u64 %rd228, %rd233;
|
|
mov.u32 %r184, 8;
|
|
// inline asm
|
|
call (%rd227), _rt_buffer_get_64, (%rd228, %r24, %r184, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f781, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs140, %f781;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd227], {%rs140, %rs140, %rs140, %rs129};
|
|
|
|
BB0_101:
|
|
ld.global.u32 %r185, [additive];
|
|
setp.eq.s32 %p114, %r185, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f758;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_103;
|
|
|
|
mov.u64 %rd246, image_RNM3;
|
|
cvta.global.u64 %rd235, %rd246;
|
|
mov.u32 %r189, 8;
|
|
// inline asm
|
|
call (%rd234), _rt_buffer_get_64, (%rd235, %r24, %r189, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs148, %rs149, %rs150, %rs151}, [%rd234];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f783, %rs148;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f784, %rs149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f785, %rs150;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd240), _rt_buffer_get_64, (%rd235, %r24, %r189, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f786, %f783, 0f00000000;
|
|
add.f32 %f787, %f784, 0f00000000;
|
|
add.f32 %f788, %f785, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs147, %f788;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs146, %f787;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs145, %f786;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd240], {%rs145, %rs146, %rs147, %rs141};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_103:
|
|
mov.u64 %rd253, image_RNM3;
|
|
cvta.global.u64 %rd248, %rd253;
|
|
mov.u32 %r191, 8;
|
|
// inline asm
|
|
call (%rd247), _rt_buffer_get_64, (%rd248, %r24, %r191, %rd6, %rd7, %rd15, %rd15);
|
|
// inline asm
|
|
mov.f32 %f789, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f789;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd247], {%rs152, %rs152, %rs152, %rs141};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_24:
|
|
setp.geu.f32 %p45, %f884, 0f00000000;
|
|
@%p45 bra BB0_27;
|
|
|
|
mov.f32 %f875, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f446, %f875;
|
|
setp.neu.f32 %p46, %f446, 0f3EE8BA2E;
|
|
selp.f32 %f887, 0f7FFFFFFF, %f887, %p46;
|
|
|
|
BB0_27:
|
|
abs.f32 %f872, %f884;
|
|
add.f32 %f448, %f872, 0f3EE8BA2E;
|
|
mov.b32 %r78, %f448;
|
|
setp.lt.s32 %p48, %r78, 2139095040;
|
|
@%p48 bra BB0_32;
|
|
|
|
abs.f32 %f873, %f884;
|
|
setp.gtu.f32 %p49, %f873, 0f7F800000;
|
|
@%p49 bra BB0_31;
|
|
bra.uni BB0_29;
|
|
|
|
BB0_31:
|
|
add.f32 %f887, %f884, 0f3EE8BA2E;
|
|
bra.uni BB0_32;
|
|
|
|
BB0_29:
|
|
abs.f32 %f874, %f884;
|
|
setp.neu.f32 %p50, %f874, 0f7F800000;
|
|
@%p50 bra BB0_32;
|
|
|
|
selp.f32 %f887, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_32:
|
|
mul.f32 %f449, %f887, 0f437F0000;
|
|
setp.eq.f32 %p51, %f884, 0f3F800000;
|
|
selp.f32 %f450, 0f437F0000, %f449, %p51;
|
|
cvt.rzi.u32.f32 %r79, %f450;
|
|
cvt.u16.u32 %rs19, %r79;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd26], {%rs19, %rs20};
|
|
ld.global.u32 %r239, [imageEnabled];
|
|
|
|
BB0_33:
|
|
and.b32 %r80, %r239, 1;
|
|
setp.eq.b32 %p52, %r80, 1;
|
|
@!%p52 bra BB0_68;
|
|
bra.uni BB0_34;
|
|
|
|
BB0_34:
|
|
mov.f32 %f850, 0fB5BFBE8E;
|
|
mov.f32 %f849, 0fBF317200;
|
|
mov.f32 %f848, 0f35BFBE8E;
|
|
mov.f32 %f847, 0f3F317200;
|
|
mov.f32 %f846, 0f3DAAAABD;
|
|
mov.f32 %f845, 0f3C4CAF63;
|
|
mov.f32 %f844, 0f3B18F0FE;
|
|
mov.f32 %f453, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f454, %f453;
|
|
fma.rn.f32 %f455, %f454, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f58, %f455;
|
|
abs.f32 %f59, %f42;
|
|
setp.lt.f32 %p53, %f59, 0f00800000;
|
|
mul.f32 %f456, %f59, 0f4B800000;
|
|
selp.f32 %f457, 0fC3170000, 0fC2FE0000, %p53;
|
|
selp.f32 %f458, %f456, %f59, %p53;
|
|
mov.b32 %r81, %f458;
|
|
and.b32 %r82, %r81, 8388607;
|
|
or.b32 %r83, %r82, 1065353216;
|
|
mov.b32 %f459, %r83;
|
|
shr.u32 %r84, %r81, 23;
|
|
cvt.rn.f32.u32 %f460, %r84;
|
|
add.f32 %f461, %f457, %f460;
|
|
setp.gt.f32 %p54, %f459, 0f3FB504F3;
|
|
mul.f32 %f462, %f459, 0f3F000000;
|
|
add.f32 %f463, %f461, 0f3F800000;
|
|
selp.f32 %f464, %f462, %f459, %p54;
|
|
selp.f32 %f465, %f463, %f461, %p54;
|
|
add.f32 %f466, %f464, 0fBF800000;
|
|
add.f32 %f452, %f464, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f451,%f452;
|
|
// inline asm
|
|
add.f32 %f467, %f466, %f466;
|
|
mul.f32 %f468, %f451, %f467;
|
|
mul.f32 %f469, %f468, %f468;
|
|
fma.rn.f32 %f472, %f844, %f469, %f845;
|
|
fma.rn.f32 %f474, %f472, %f469, %f846;
|
|
mul.rn.f32 %f475, %f474, %f469;
|
|
mul.rn.f32 %f476, %f475, %f468;
|
|
sub.f32 %f477, %f466, %f468;
|
|
neg.f32 %f478, %f468;
|
|
add.f32 %f479, %f477, %f477;
|
|
fma.rn.f32 %f480, %f478, %f466, %f479;
|
|
mul.rn.f32 %f481, %f451, %f480;
|
|
add.f32 %f482, %f476, %f468;
|
|
sub.f32 %f483, %f468, %f482;
|
|
add.f32 %f484, %f476, %f483;
|
|
add.f32 %f485, %f481, %f484;
|
|
add.f32 %f486, %f482, %f485;
|
|
sub.f32 %f487, %f482, %f486;
|
|
add.f32 %f488, %f485, %f487;
|
|
mul.rn.f32 %f490, %f465, %f847;
|
|
mul.rn.f32 %f492, %f465, %f848;
|
|
add.f32 %f493, %f490, %f486;
|
|
sub.f32 %f494, %f490, %f493;
|
|
add.f32 %f495, %f486, %f494;
|
|
add.f32 %f496, %f488, %f495;
|
|
add.f32 %f497, %f492, %f496;
|
|
add.f32 %f498, %f493, %f497;
|
|
sub.f32 %f499, %f493, %f498;
|
|
add.f32 %f500, %f497, %f499;
|
|
mov.f32 %f501, 0f3EE66666;
|
|
mul.rn.f32 %f502, %f501, %f498;
|
|
neg.f32 %f503, %f502;
|
|
fma.rn.f32 %f504, %f501, %f498, %f503;
|
|
fma.rn.f32 %f505, %f501, %f500, %f504;
|
|
mov.f32 %f506, 0f00000000;
|
|
fma.rn.f32 %f507, %f506, %f498, %f505;
|
|
add.rn.f32 %f508, %f502, %f507;
|
|
neg.f32 %f509, %f508;
|
|
add.rn.f32 %f510, %f502, %f509;
|
|
add.rn.f32 %f511, %f510, %f507;
|
|
mov.b32 %r85, %f508;
|
|
setp.eq.s32 %p55, %r85, 1118925336;
|
|
add.s32 %r86, %r85, -1;
|
|
mov.b32 %f512, %r86;
|
|
add.f32 %f513, %f511, 0f37000000;
|
|
selp.f32 %f514, %f512, %f508, %p55;
|
|
selp.f32 %f60, %f513, %f511, %p55;
|
|
mul.f32 %f515, %f514, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f516, %f515;
|
|
fma.rn.f32 %f518, %f516, %f849, %f514;
|
|
fma.rn.f32 %f520, %f516, %f850, %f518;
|
|
mul.f32 %f521, %f520, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f522, %f521;
|
|
add.f32 %f523, %f516, 0f00000000;
|
|
ex2.approx.f32 %f524, %f523;
|
|
mul.f32 %f525, %f522, %f524;
|
|
setp.lt.f32 %p56, %f514, 0fC2D20000;
|
|
selp.f32 %f526, 0f00000000, %f525, %p56;
|
|
setp.gt.f32 %p57, %f514, 0f42D20000;
|
|
selp.f32 %f888, 0f7F800000, %f526, %p57;
|
|
setp.eq.f32 %p58, %f888, 0f7F800000;
|
|
@%p58 bra BB0_36;
|
|
|
|
fma.rn.f32 %f888, %f888, %f60, %f888;
|
|
|
|
BB0_36:
|
|
setp.lt.f32 %p59, %f42, 0f00000000;
|
|
setp.eq.f32 %p60, %f58, 0f3F800000;
|
|
and.pred %p3, %p59, %p60;
|
|
mov.b32 %r87, %f888;
|
|
xor.b32 %r88, %r87, -2147483648;
|
|
mov.b32 %f527, %r88;
|
|
selp.f32 %f890, %f527, %f888, %p3;
|
|
setp.eq.f32 %p61, %f42, 0f00000000;
|
|
@%p61 bra BB0_39;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_39:
|
|
add.f32 %f530, %f42, %f42;
|
|
selp.f32 %f890, %f530, 0f00000000, %p60;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_37:
|
|
setp.geu.f32 %p62, %f42, 0f00000000;
|
|
@%p62 bra BB0_40;
|
|
|
|
cvt.rzi.f32.f32 %f529, %f501;
|
|
setp.neu.f32 %p63, %f529, 0f3EE66666;
|
|
selp.f32 %f890, 0f7FFFFFFF, %f890, %p63;
|
|
|
|
BB0_40:
|
|
add.f32 %f531, %f59, 0f3EE66666;
|
|
mov.b32 %r89, %f531;
|
|
setp.lt.s32 %p65, %r89, 2139095040;
|
|
@%p65 bra BB0_45;
|
|
|
|
setp.gtu.f32 %p66, %f59, 0f7F800000;
|
|
@%p66 bra BB0_44;
|
|
bra.uni BB0_42;
|
|
|
|
BB0_44:
|
|
add.f32 %f890, %f42, 0f3EE66666;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_42:
|
|
setp.neu.f32 %p67, %f59, 0f7F800000;
|
|
@%p67 bra BB0_45;
|
|
|
|
selp.f32 %f890, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_45:
|
|
mov.f32 %f857, 0fB5BFBE8E;
|
|
mov.f32 %f856, 0fBF317200;
|
|
mov.f32 %f855, 0f35BFBE8E;
|
|
mov.f32 %f854, 0f3F317200;
|
|
mov.f32 %f853, 0f3DAAAABD;
|
|
mov.f32 %f852, 0f3C4CAF63;
|
|
mov.f32 %f851, 0f3B18F0FE;
|
|
setp.eq.f32 %p68, %f42, 0f3F800000;
|
|
selp.f32 %f71, 0f3F800000, %f890, %p68;
|
|
abs.f32 %f72, %f43;
|
|
setp.lt.f32 %p69, %f72, 0f00800000;
|
|
mul.f32 %f534, %f72, 0f4B800000;
|
|
selp.f32 %f535, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f536, %f534, %f72, %p69;
|
|
mov.b32 %r90, %f536;
|
|
and.b32 %r91, %r90, 8388607;
|
|
or.b32 %r92, %r91, 1065353216;
|
|
mov.b32 %f537, %r92;
|
|
shr.u32 %r93, %r90, 23;
|
|
cvt.rn.f32.u32 %f538, %r93;
|
|
add.f32 %f539, %f535, %f538;
|
|
setp.gt.f32 %p70, %f537, 0f3FB504F3;
|
|
mul.f32 %f540, %f537, 0f3F000000;
|
|
add.f32 %f541, %f539, 0f3F800000;
|
|
selp.f32 %f542, %f540, %f537, %p70;
|
|
selp.f32 %f543, %f541, %f539, %p70;
|
|
add.f32 %f544, %f542, 0fBF800000;
|
|
add.f32 %f533, %f542, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f532,%f533;
|
|
// inline asm
|
|
add.f32 %f545, %f544, %f544;
|
|
mul.f32 %f546, %f532, %f545;
|
|
mul.f32 %f547, %f546, %f546;
|
|
fma.rn.f32 %f550, %f851, %f547, %f852;
|
|
fma.rn.f32 %f552, %f550, %f547, %f853;
|
|
mul.rn.f32 %f553, %f552, %f547;
|
|
mul.rn.f32 %f554, %f553, %f546;
|
|
sub.f32 %f555, %f544, %f546;
|
|
neg.f32 %f556, %f546;
|
|
add.f32 %f557, %f555, %f555;
|
|
fma.rn.f32 %f558, %f556, %f544, %f557;
|
|
mul.rn.f32 %f559, %f532, %f558;
|
|
add.f32 %f560, %f554, %f546;
|
|
sub.f32 %f561, %f546, %f560;
|
|
add.f32 %f562, %f554, %f561;
|
|
add.f32 %f563, %f559, %f562;
|
|
add.f32 %f564, %f560, %f563;
|
|
sub.f32 %f565, %f560, %f564;
|
|
add.f32 %f566, %f563, %f565;
|
|
mul.rn.f32 %f568, %f543, %f854;
|
|
mul.rn.f32 %f570, %f543, %f855;
|
|
add.f32 %f571, %f568, %f564;
|
|
sub.f32 %f572, %f568, %f571;
|
|
add.f32 %f573, %f564, %f572;
|
|
add.f32 %f574, %f566, %f573;
|
|
add.f32 %f575, %f570, %f574;
|
|
add.f32 %f576, %f571, %f575;
|
|
sub.f32 %f577, %f571, %f576;
|
|
add.f32 %f578, %f575, %f577;
|
|
mul.rn.f32 %f580, %f501, %f576;
|
|
neg.f32 %f581, %f580;
|
|
fma.rn.f32 %f582, %f501, %f576, %f581;
|
|
fma.rn.f32 %f583, %f501, %f578, %f582;
|
|
fma.rn.f32 %f585, %f506, %f576, %f583;
|
|
add.rn.f32 %f586, %f580, %f585;
|
|
neg.f32 %f587, %f586;
|
|
add.rn.f32 %f588, %f580, %f587;
|
|
add.rn.f32 %f589, %f588, %f585;
|
|
mov.b32 %r94, %f586;
|
|
setp.eq.s32 %p71, %r94, 1118925336;
|
|
add.s32 %r95, %r94, -1;
|
|
mov.b32 %f590, %r95;
|
|
add.f32 %f591, %f589, 0f37000000;
|
|
selp.f32 %f592, %f590, %f586, %p71;
|
|
selp.f32 %f73, %f591, %f589, %p71;
|
|
mul.f32 %f593, %f592, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f594, %f593;
|
|
fma.rn.f32 %f596, %f594, %f856, %f592;
|
|
fma.rn.f32 %f598, %f594, %f857, %f596;
|
|
mul.f32 %f599, %f598, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f600, %f599;
|
|
add.f32 %f601, %f594, 0f00000000;
|
|
ex2.approx.f32 %f602, %f601;
|
|
mul.f32 %f603, %f600, %f602;
|
|
setp.lt.f32 %p72, %f592, 0fC2D20000;
|
|
selp.f32 %f604, 0f00000000, %f603, %p72;
|
|
setp.gt.f32 %p73, %f592, 0f42D20000;
|
|
selp.f32 %f891, 0f7F800000, %f604, %p73;
|
|
setp.eq.f32 %p74, %f891, 0f7F800000;
|
|
@%p74 bra BB0_47;
|
|
|
|
fma.rn.f32 %f891, %f891, %f73, %f891;
|
|
|
|
BB0_47:
|
|
setp.lt.f32 %p75, %f43, 0f00000000;
|
|
and.pred %p4, %p75, %p60;
|
|
mov.b32 %r96, %f891;
|
|
xor.b32 %r97, %r96, -2147483648;
|
|
mov.b32 %f605, %r97;
|
|
selp.f32 %f893, %f605, %f891, %p4;
|
|
setp.eq.f32 %p77, %f43, 0f00000000;
|
|
@%p77 bra BB0_50;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_50:
|
|
add.f32 %f608, %f43, %f43;
|
|
selp.f32 %f893, %f608, 0f00000000, %p60;
|
|
bra.uni BB0_51;
|
|
|
|
BB0_48:
|
|
setp.geu.f32 %p78, %f43, 0f00000000;
|
|
@%p78 bra BB0_51;
|
|
|
|
cvt.rzi.f32.f32 %f607, %f501;
|
|
setp.neu.f32 %p79, %f607, 0f3EE66666;
|
|
selp.f32 %f893, 0f7FFFFFFF, %f893, %p79;
|
|
|
|
BB0_51:
|
|
add.f32 %f609, %f72, 0f3EE66666;
|
|
mov.b32 %r98, %f609;
|
|
setp.lt.s32 %p81, %r98, 2139095040;
|
|
@%p81 bra BB0_56;
|
|
|
|
setp.gtu.f32 %p82, %f72, 0f7F800000;
|
|
@%p82 bra BB0_55;
|
|
bra.uni BB0_53;
|
|
|
|
BB0_55:
|
|
add.f32 %f893, %f43, 0f3EE66666;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_53:
|
|
setp.neu.f32 %p83, %f72, 0f7F800000;
|
|
@%p83 bra BB0_56;
|
|
|
|
selp.f32 %f893, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_56:
|
|
mov.f32 %f864, 0fB5BFBE8E;
|
|
mov.f32 %f863, 0fBF317200;
|
|
mov.f32 %f862, 0f35BFBE8E;
|
|
mov.f32 %f861, 0f3F317200;
|
|
mov.f32 %f860, 0f3DAAAABD;
|
|
mov.f32 %f859, 0f3C4CAF63;
|
|
mov.f32 %f858, 0f3B18F0FE;
|
|
setp.eq.f32 %p84, %f43, 0f3F800000;
|
|
selp.f32 %f84, 0f3F800000, %f893, %p84;
|
|
abs.f32 %f85, %f44;
|
|
setp.lt.f32 %p85, %f85, 0f00800000;
|
|
mul.f32 %f612, %f85, 0f4B800000;
|
|
selp.f32 %f613, 0fC3170000, 0fC2FE0000, %p85;
|
|
selp.f32 %f614, %f612, %f85, %p85;
|
|
mov.b32 %r99, %f614;
|
|
and.b32 %r100, %r99, 8388607;
|
|
or.b32 %r101, %r100, 1065353216;
|
|
mov.b32 %f615, %r101;
|
|
shr.u32 %r102, %r99, 23;
|
|
cvt.rn.f32.u32 %f616, %r102;
|
|
add.f32 %f617, %f613, %f616;
|
|
setp.gt.f32 %p86, %f615, 0f3FB504F3;
|
|
mul.f32 %f618, %f615, 0f3F000000;
|
|
add.f32 %f619, %f617, 0f3F800000;
|
|
selp.f32 %f620, %f618, %f615, %p86;
|
|
selp.f32 %f621, %f619, %f617, %p86;
|
|
add.f32 %f622, %f620, 0fBF800000;
|
|
add.f32 %f611, %f620, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f610,%f611;
|
|
// inline asm
|
|
add.f32 %f623, %f622, %f622;
|
|
mul.f32 %f624, %f610, %f623;
|
|
mul.f32 %f625, %f624, %f624;
|
|
fma.rn.f32 %f628, %f858, %f625, %f859;
|
|
fma.rn.f32 %f630, %f628, %f625, %f860;
|
|
mul.rn.f32 %f631, %f630, %f625;
|
|
mul.rn.f32 %f632, %f631, %f624;
|
|
sub.f32 %f633, %f622, %f624;
|
|
neg.f32 %f634, %f624;
|
|
add.f32 %f635, %f633, %f633;
|
|
fma.rn.f32 %f636, %f634, %f622, %f635;
|
|
mul.rn.f32 %f637, %f610, %f636;
|
|
add.f32 %f638, %f632, %f624;
|
|
sub.f32 %f639, %f624, %f638;
|
|
add.f32 %f640, %f632, %f639;
|
|
add.f32 %f641, %f637, %f640;
|
|
add.f32 %f642, %f638, %f641;
|
|
sub.f32 %f643, %f638, %f642;
|
|
add.f32 %f644, %f641, %f643;
|
|
mul.rn.f32 %f646, %f621, %f861;
|
|
mul.rn.f32 %f648, %f621, %f862;
|
|
add.f32 %f649, %f646, %f642;
|
|
sub.f32 %f650, %f646, %f649;
|
|
add.f32 %f651, %f642, %f650;
|
|
add.f32 %f652, %f644, %f651;
|
|
add.f32 %f653, %f648, %f652;
|
|
add.f32 %f654, %f649, %f653;
|
|
sub.f32 %f655, %f649, %f654;
|
|
add.f32 %f656, %f653, %f655;
|
|
mul.rn.f32 %f658, %f501, %f654;
|
|
neg.f32 %f659, %f658;
|
|
fma.rn.f32 %f660, %f501, %f654, %f659;
|
|
fma.rn.f32 %f661, %f501, %f656, %f660;
|
|
fma.rn.f32 %f663, %f506, %f654, %f661;
|
|
add.rn.f32 %f664, %f658, %f663;
|
|
neg.f32 %f665, %f664;
|
|
add.rn.f32 %f666, %f658, %f665;
|
|
add.rn.f32 %f667, %f666, %f663;
|
|
mov.b32 %r103, %f664;
|
|
setp.eq.s32 %p87, %r103, 1118925336;
|
|
add.s32 %r104, %r103, -1;
|
|
mov.b32 %f668, %r104;
|
|
add.f32 %f669, %f667, 0f37000000;
|
|
selp.f32 %f670, %f668, %f664, %p87;
|
|
selp.f32 %f86, %f669, %f667, %p87;
|
|
mul.f32 %f671, %f670, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f672, %f671;
|
|
fma.rn.f32 %f674, %f672, %f863, %f670;
|
|
fma.rn.f32 %f676, %f672, %f864, %f674;
|
|
mul.f32 %f677, %f676, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f678, %f677;
|
|
add.f32 %f679, %f672, 0f00000000;
|
|
ex2.approx.f32 %f680, %f679;
|
|
mul.f32 %f681, %f678, %f680;
|
|
setp.lt.f32 %p88, %f670, 0fC2D20000;
|
|
selp.f32 %f682, 0f00000000, %f681, %p88;
|
|
setp.gt.f32 %p89, %f670, 0f42D20000;
|
|
selp.f32 %f894, 0f7F800000, %f682, %p89;
|
|
setp.eq.f32 %p90, %f894, 0f7F800000;
|
|
@%p90 bra BB0_58;
|
|
|
|
fma.rn.f32 %f894, %f894, %f86, %f894;
|
|
|
|
BB0_58:
|
|
setp.lt.f32 %p91, %f44, 0f00000000;
|
|
and.pred %p5, %p91, %p60;
|
|
mov.b32 %r105, %f894;
|
|
xor.b32 %r106, %r105, -2147483648;
|
|
mov.b32 %f683, %r106;
|
|
selp.f32 %f896, %f683, %f894, %p5;
|
|
setp.eq.f32 %p93, %f44, 0f00000000;
|
|
@%p93 bra BB0_61;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_61:
|
|
add.f32 %f686, %f44, %f44;
|
|
selp.f32 %f896, %f686, 0f00000000, %p60;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_59:
|
|
setp.geu.f32 %p94, %f44, 0f00000000;
|
|
@%p94 bra BB0_62;
|
|
|
|
cvt.rzi.f32.f32 %f685, %f501;
|
|
setp.neu.f32 %p95, %f685, 0f3EE66666;
|
|
selp.f32 %f896, 0f7FFFFFFF, %f896, %p95;
|
|
|
|
BB0_62:
|
|
add.f32 %f687, %f85, 0f3EE66666;
|
|
mov.b32 %r107, %f687;
|
|
setp.lt.s32 %p97, %r107, 2139095040;
|
|
@%p97 bra BB0_67;
|
|
|
|
setp.gtu.f32 %p98, %f85, 0f7F800000;
|
|
@%p98 bra BB0_66;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_66:
|
|
add.f32 %f896, %f44, 0f3EE66666;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_64:
|
|
setp.neu.f32 %p99, %f85, 0f7F800000;
|
|
@%p99 bra BB0_67;
|
|
|
|
selp.f32 %f896, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_67:
|
|
mov.u32 %r235, 4;
|
|
setp.eq.f32 %p100, %f44, 0f3F800000;
|
|
selp.f32 %f688, 0f3F800000, %f896, %p100;
|
|
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, %r24, %r235, %rd35, %rd36, %rd15, %rd15);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f689, %f688;
|
|
mul.f32 %f690, %f689, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r110, %f690;
|
|
cvt.sat.f32.f32 %f691, %f84;
|
|
mul.f32 %f692, %f691, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r111, %f692;
|
|
cvt.sat.f32.f32 %f693, %f71;
|
|
mul.f32 %f694, %f693, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r112, %f694;
|
|
cvt.u16.u32 %rs21, %r110;
|
|
cvt.u16.u32 %rs22, %r112;
|
|
cvt.u16.u32 %rs23, %r111;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd33], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r239, [imageEnabled];
|
|
|
|
BB0_68:
|
|
cvt.u64.u32 %rd4, %r2;
|
|
cvt.u64.u32 %rd5, %r3;
|
|
and.b32 %r113, %r239, 4;
|
|
setp.eq.s32 %p101, %r113, 0;
|
|
@%p101 bra BB0_72;
|
|
|
|
ld.global.u32 %r114, [additive];
|
|
setp.eq.s32 %p102, %r114, 0;
|
|
mov.f32 %f695, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f695;}
|
|
|
|
// inline asm
|
|
@%p102 bra BB0_71;
|
|
|
|
mov.u64 %rd52, image_HDR;
|
|
cvta.global.u64 %rd41, %rd52;
|
|
mov.u32 %r118, 8;
|
|
// inline asm
|
|
call (%rd40), _rt_buffer_get_64, (%rd41, %r24, %r118, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd40];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f696, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f697, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f698, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd46), _rt_buffer_get_64, (%rd41, %r24, %r118, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f699, %f42, %f696;
|
|
add.f32 %f700, %f43, %f697;
|
|
add.f32 %f701, %f44, %f698;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f701;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f700;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f699;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd46], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_72;
|
|
|
|
BB0_71:
|
|
mov.u64 %rd59, image_HDR;
|
|
cvta.global.u64 %rd54, %rd59;
|
|
mov.u32 %r120, 8;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd54, %r24, %r120, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f42;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd53], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_72:
|
|
mul.f32 %f706, %f28, 0f3E800000;
|
|
mul.f32 %f707, %f706, %f884;
|
|
ld.global.f32 %f708, [lightColor];
|
|
mul.f32 %f709, %f707, %f708;
|
|
ld.global.f32 %f710, [lightColor+4];
|
|
mul.f32 %f711, %f707, %f710;
|
|
ld.global.f32 %f712, [lightColor+8];
|
|
mul.f32 %f713, %f707, %f712;
|
|
mul.f32 %f97, %f215, %f709;
|
|
mul.f32 %f98, %f216, %f711;
|
|
mul.f32 %f99, %f217, %f713;
|
|
ld.global.u32 %r121, [additive];
|
|
setp.eq.s32 %p103, %r121, 0;
|
|
mov.f32 %f705, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f705;}
|
|
|
|
// inline asm
|
|
@%p103 bra BB0_74;
|
|
|
|
mov.u64 %rd72, image_RNM0;
|
|
cvta.global.u64 %rd61, %rd72;
|
|
mov.u32 %r125, 8;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd61, %r24, %r125, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd60];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f714, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f715, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f716, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd66), _rt_buffer_get_64, (%rd61, %r24, %r125, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f717, %f97, %f714;
|
|
add.f32 %f718, %f98, %f715;
|
|
add.f32 %f719, %f99, %f716;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f719;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f718;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f717;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd66], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_75;
|
|
|
|
BB0_74:
|
|
mov.u64 %rd79, image_RNM0;
|
|
cvta.global.u64 %rd74, %rd79;
|
|
mov.u32 %r127, 8;
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd74, %r24, %r127, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f99;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f98;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f97;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd73], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_75:
|
|
fma.rn.f32 %f100, %f10, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f101, %f11, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f102, %f12, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r128, [additive];
|
|
setp.eq.s32 %p104, %r128, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f705;}
|
|
|
|
// inline asm
|
|
@%p104 bra BB0_77;
|
|
|
|
mov.u64 %rd92, image_RNM1;
|
|
cvta.global.u64 %rd81, %rd92;
|
|
mov.u32 %r132, 8;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r24, %r132, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd80];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f724, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f725, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f726, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd86), _rt_buffer_get_64, (%rd81, %r24, %r132, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f727, %f100, %f724;
|
|
add.f32 %f728, %f100, %f725;
|
|
add.f32 %f729, %f100, %f726;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f729;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f728;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f727;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd86], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_78;
|
|
|
|
BB0_77:
|
|
mov.u64 %rd99, image_RNM1;
|
|
cvta.global.u64 %rd94, %rd99;
|
|
mov.u32 %r134, 8;
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd94, %r24, %r134, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f100;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs64, %rs64, %rs64, %rs53};
|
|
|
|
BB0_78:
|
|
ld.global.u32 %r135, [additive];
|
|
setp.eq.s32 %p105, %r135, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f705;}
|
|
|
|
// inline asm
|
|
@%p105 bra BB0_80;
|
|
|
|
mov.u64 %rd112, image_RNM2;
|
|
cvta.global.u64 %rd101, %rd112;
|
|
mov.u32 %r139, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r24, %r139, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs72, %rs73, %rs74, %rs75}, [%rd100];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f732, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f733, %rs73;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f734, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd106), _rt_buffer_get_64, (%rd101, %r24, %r139, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f735, %f101, %f732;
|
|
add.f32 %f736, %f101, %f733;
|
|
add.f32 %f737, %f101, %f734;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f737;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs70, %f736;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f735;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd106], {%rs69, %rs70, %rs71, %rs65};
|
|
bra.uni BB0_81;
|
|
|
|
BB0_80:
|
|
mov.u64 %rd119, image_RNM2;
|
|
cvta.global.u64 %rd114, %rd119;
|
|
mov.u32 %r141, 8;
|
|
// inline asm
|
|
call (%rd113), _rt_buffer_get_64, (%rd114, %r24, %r141, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f101;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd113], {%rs76, %rs76, %rs76, %rs65};
|
|
|
|
BB0_81:
|
|
ld.global.u32 %r142, [additive];
|
|
setp.eq.s32 %p106, %r142, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f705;}
|
|
|
|
// inline asm
|
|
@%p106 bra BB0_83;
|
|
|
|
mov.u64 %rd132, image_RNM3;
|
|
cvta.global.u64 %rd121, %rd132;
|
|
mov.u32 %r146, 8;
|
|
// inline asm
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r24, %r146, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd120];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f740, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f741, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f742, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd126), _rt_buffer_get_64, (%rd121, %r24, %r146, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
add.f32 %f743, %f102, %f740;
|
|
add.f32 %f744, %f102, %f741;
|
|
add.f32 %f745, %f102, %f742;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f745;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f744;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f743;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd126], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_83:
|
|
mov.u64 %rd139, image_RNM3;
|
|
cvta.global.u64 %rd134, %rd139;
|
|
mov.u32 %r148, 8;
|
|
// inline asm
|
|
call (%rd133), _rt_buffer_get_64, (%rd134, %r24, %r148, %rd4, %rd5, %rd15, %rd15);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f102;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd133], {%rs88, %rs88, %rs88, %rs77};
|
|
|
|
BB0_124:
|
|
ret;
|
|
}
|
|
|
|
|