2871 lines
76 KiB
Plaintext
2871 lines
76 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 image_RNM4[1];
|
|
.global .align 1 .b8 image_RNM5[1];
|
|
.global .align 1 .b8 image_RNM6[1];
|
|
.global .align 1 .b8 image_RNM7[1];
|
|
.global .align 1 .b8 image_RNM8[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 4 .b8 directDir[12];
|
|
.global .align 4 .b8 directColor[12];
|
|
.global .align 4 .f32 shadowSpread;
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .u32 ignoreNormal;
|
|
.global .align 4 .u32 lightCookie;
|
|
.global .align 16 .b8 lightTilingOffset[16];
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo9directDirE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11directColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12shadowSpreadE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12ignoreNormalE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo11lightCookieE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo17lightTilingOffsetE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename9directDirE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename11directColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename12shadowSpreadE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12ignoreNormalE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename11lightCookieE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename17lightTilingOffsetE[7] = {102, 108, 111, 97, 116, 52, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum9directDirE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11directColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12shadowSpreadE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12ignoreNormalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum11lightCookieE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum17lightTilingOffsetE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic9directDirE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11directColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12shadowSpreadE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12ignoreNormalE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic11lightCookieE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic17lightTilingOffsetE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation9directDirE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11directColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12shadowSpreadE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12ignoreNormalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation11lightCookieE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation17lightTilingOffsetE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[32];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<134>;
|
|
.reg .b16 %rs<298>;
|
|
.reg .f32 %f<959>;
|
|
.reg .b32 %r<460>;
|
|
.reg .b64 %rd<486>;
|
|
|
|
|
|
mov.u64 %rd485, __local_depot0;
|
|
cvta.local.u64 %SP, %rd485;
|
|
ld.global.v2.u32 {%r101, %r102}, [pixelID];
|
|
cvt.u64.u32 %rd22, %r101;
|
|
cvt.u64.u32 %rd23, %r102;
|
|
mov.u64 %rd26, uvnormal;
|
|
cvta.global.u64 %rd21, %rd26;
|
|
mov.u32 %r99, 2;
|
|
mov.u32 %r100, 4;
|
|
mov.u64 %rd25, 0;
|
|
// inline asm
|
|
call (%rd20), _rt_buffer_get_64, (%rd21, %r99, %r100, %rd22, %rd23, %rd25, %rd25);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd20];
|
|
shr.u32 %r105, %r1, 16;
|
|
cvt.u16.u32 %rs1, %r105;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r1;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p5, %rs14, 0;
|
|
mov.f32 %f927, 0f00000000;
|
|
mov.f32 %f928, %f927;
|
|
mov.f32 %f929, %f927;
|
|
@%p5 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd20+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f165, %rs17;
|
|
div.rn.f32 %f166, %f165, 0f437F0000;
|
|
fma.rn.f32 %f167, %f166, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f168, %rs15;
|
|
div.rn.f32 %f169, %f168, 0f437F0000;
|
|
fma.rn.f32 %f170, %f169, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f171, %rs12;
|
|
div.rn.f32 %f172, %f171, 0f437F0000;
|
|
fma.rn.f32 %f173, %f172, 0f40000000, 0fBF800000;
|
|
mul.f32 %f174, %f170, %f170;
|
|
fma.rn.f32 %f175, %f167, %f167, %f174;
|
|
fma.rn.f32 %f176, %f173, %f173, %f175;
|
|
sqrt.rn.f32 %f177, %f176;
|
|
rcp.rn.f32 %f178, %f177;
|
|
mul.f32 %f927, %f167, %f178;
|
|
mul.f32 %f928, %f170, %f178;
|
|
mul.f32 %f929, %f173, %f178;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r106, %r107}, [pixelID];
|
|
ld.global.v2.u32 {%r109, %r110}, [tileInfo];
|
|
add.s32 %r2, %r106, %r109;
|
|
add.s32 %r3, %r107, %r110;
|
|
setp.eq.f32 %p6, %f928, 0f00000000;
|
|
setp.eq.f32 %p7, %f927, 0f00000000;
|
|
and.pred %p8, %p7, %p6;
|
|
setp.eq.f32 %p9, %f929, 0f00000000;
|
|
and.pred %p10, %p8, %p9;
|
|
@%p10 bra BB0_136;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_136:
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
and.b32 %r349, %r459, 1;
|
|
setp.eq.b32 %p121, %r349, 1;
|
|
@!%p121 bra BB0_138;
|
|
bra.uni BB0_137;
|
|
|
|
BB0_137:
|
|
cvt.u64.u32 %rd269, %r2;
|
|
cvt.u64.u32 %rd270, %r3;
|
|
mov.u64 %rd273, image;
|
|
cvta.global.u64 %rd268, %rd273;
|
|
// inline asm
|
|
call (%rd267), _rt_buffer_get_64, (%rd268, %r99, %r100, %rd269, %rd270, %rd25, %rd25);
|
|
// inline asm
|
|
mov.u16 %rs165, 0;
|
|
st.v4.u8 [%rd267], {%rs165, %rs165, %rs165, %rs165};
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
|
|
BB0_138:
|
|
and.b32 %r352, %r459, 8;
|
|
setp.eq.s32 %p122, %r352, 0;
|
|
@%p122 bra BB0_140;
|
|
|
|
cvt.u64.u32 %rd277, %r3;
|
|
cvt.u64.u32 %rd276, %r2;
|
|
mov.u64 %rd280, image_Mask;
|
|
cvta.global.u64 %rd275, %rd280;
|
|
// inline asm
|
|
call (%rd274), _rt_buffer_get_64, (%rd275, %r99, %r99, %rd276, %rd277, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f813, 0f00000000;
|
|
cvt.rzi.u32.f32 %r355, %f813;
|
|
cvt.u16.u32 %rs166, %r355;
|
|
mov.u16 %rs167, 0;
|
|
st.v2.u8 [%rd274], {%rs166, %rs167};
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
|
|
BB0_140:
|
|
cvt.u64.u32 %rd18, %r2;
|
|
cvt.u64.u32 %rd19, %r3;
|
|
and.b32 %r356, %r459, 4;
|
|
setp.eq.s32 %p123, %r356, 0;
|
|
@%p123 bra BB0_144;
|
|
|
|
ld.global.u32 %r357, [additive];
|
|
setp.eq.s32 %p124, %r357, 0;
|
|
@%p124 bra BB0_143;
|
|
|
|
mov.u64 %rd293, image_HDR;
|
|
cvta.global.u64 %rd282, %rd293;
|
|
mov.u32 %r361, 8;
|
|
// inline asm
|
|
call (%rd281), _rt_buffer_get_64, (%rd282, %r99, %r361, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs174, %rs175, %rs176, %rs177}, [%rd281];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f814, %rs174;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f815, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f816, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd287), _rt_buffer_get_64, (%rd282, %r99, %r361, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f817, %f814, 0f00000000;
|
|
add.f32 %f818, %f815, 0f00000000;
|
|
add.f32 %f819, %f816, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f819;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f818;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f817;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs178, 0;
|
|
st.v4.u16 [%rd287], {%rs171, %rs172, %rs173, %rs178};
|
|
bra.uni BB0_144;
|
|
|
|
BB0_3:
|
|
ld.global.f32 %f9, [directDir+8];
|
|
ld.global.f32 %f8, [directDir+4];
|
|
ld.global.f32 %f7, [directDir];
|
|
mul.f32 %f180, %f927, %f7;
|
|
mul.f32 %f181, %f928, %f8;
|
|
neg.f32 %f182, %f181;
|
|
sub.f32 %f183, %f182, %f180;
|
|
mul.f32 %f184, %f929, %f9;
|
|
sub.f32 %f10, %f183, %f184;
|
|
ld.global.v2.u32 {%r118, %r119}, [pixelID];
|
|
cvt.u64.u32 %rd29, %r118;
|
|
cvt.u64.u32 %rd30, %r119;
|
|
mov.u64 %rd39, uvpos;
|
|
cvta.global.u64 %rd28, %rd39;
|
|
mov.u32 %r115, 12;
|
|
// inline asm
|
|
call (%rd27), _rt_buffer_get_64, (%rd28, %r99, %r115, %rd29, %rd30, %rd25, %rd25);
|
|
// inline asm
|
|
ld.f32 %f13, [%rd27+8];
|
|
ld.f32 %f12, [%rd27+4];
|
|
ld.f32 %f11, [%rd27];
|
|
mul.f32 %f185, %f11, 0f3456BF95;
|
|
mul.f32 %f186, %f12, 0f3456BF95;
|
|
mul.f32 %f187, %f13, 0f3456BF95;
|
|
abs.f32 %f188, %f927;
|
|
div.rn.f32 %f189, %f185, %f188;
|
|
abs.f32 %f190, %f928;
|
|
div.rn.f32 %f191, %f186, %f190;
|
|
abs.f32 %f192, %f929;
|
|
div.rn.f32 %f193, %f187, %f192;
|
|
abs.f32 %f194, %f189;
|
|
abs.f32 %f195, %f191;
|
|
abs.f32 %f196, %f193;
|
|
mov.f32 %f197, 0f38D1B717;
|
|
max.f32 %f198, %f194, %f197;
|
|
max.f32 %f199, %f195, %f197;
|
|
max.f32 %f200, %f196, %f197;
|
|
fma.rn.f32 %f14, %f927, %f198, %f11;
|
|
fma.rn.f32 %f15, %f928, %f199, %f12;
|
|
fma.rn.f32 %f16, %f929, %f200, %f13;
|
|
abs.f32 %f201, %f7;
|
|
abs.f32 %f202, %f9;
|
|
setp.gt.f32 %p11, %f201, %f202;
|
|
neg.f32 %f203, %f8;
|
|
neg.f32 %f204, %f9;
|
|
selp.f32 %f205, %f203, 0f00000000, %p11;
|
|
selp.f32 %f206, %f7, %f204, %p11;
|
|
selp.f32 %f207, 0f00000000, %f8, %p11;
|
|
mul.f32 %f208, %f206, %f206;
|
|
fma.rn.f32 %f209, %f205, %f205, %f208;
|
|
fma.rn.f32 %f210, %f207, %f207, %f209;
|
|
sqrt.rn.f32 %f211, %f210;
|
|
rcp.rn.f32 %f212, %f211;
|
|
mul.f32 %f17, %f205, %f212;
|
|
mul.f32 %f18, %f206, %f212;
|
|
mul.f32 %f19, %f207, %f212;
|
|
mul.f32 %f213, %f9, %f18;
|
|
mul.f32 %f214, %f8, %f19;
|
|
sub.f32 %f20, %f213, %f214;
|
|
mul.f32 %f215, %f7, %f19;
|
|
mul.f32 %f216, %f9, %f17;
|
|
sub.f32 %f21, %f215, %f216;
|
|
mul.f32 %f217, %f8, %f17;
|
|
mul.f32 %f218, %f7, %f18;
|
|
sub.f32 %f22, %f217, %f218;
|
|
ld.global.v2.u32 {%r122, %r123}, [pixelID];
|
|
cvt.u64.u32 %rd35, %r122;
|
|
cvt.u64.u32 %rd36, %r123;
|
|
mov.u64 %rd40, rnd_seeds;
|
|
cvta.global.u64 %rd34, %rd40;
|
|
// inline asm
|
|
call (%rd33), _rt_buffer_get_64, (%rd34, %r99, %r100, %rd35, %rd36, %rd25, %rd25);
|
|
// inline asm
|
|
ld.global.u32 %r428, [samples];
|
|
mov.f32 %f944, 0f00000000;
|
|
setp.lt.s32 %p12, %r428, 1;
|
|
@%p12 bra BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f220, %r428;
|
|
rcp.rn.f32 %f23, %f220;
|
|
ld.u32 %r454, [%rd33];
|
|
mul.f32 %f24, %f14, 0f3456BF95;
|
|
mul.f32 %f25, %f15, 0f3456BF95;
|
|
mul.f32 %f26, %f16, 0f3456BF95;
|
|
mov.f32 %f944, 0f00000000;
|
|
mov.u32 %r429, 0;
|
|
abs.f32 %f221, %f25;
|
|
abs.f32 %f222, %f24;
|
|
max.f32 %f223, %f222, %f221;
|
|
abs.f32 %f224, %f26;
|
|
max.f32 %f225, %f223, %f224;
|
|
|
|
BB0_5:
|
|
setp.lt.s32 %p13, %r428, 1;
|
|
@%p13 bra BB0_54;
|
|
|
|
cvt.rn.f32.s32 %f28, %r429;
|
|
max.f32 %f29, %f225, %f197;
|
|
mov.u32 %r431, 0;
|
|
|
|
BB0_7:
|
|
mad.lo.s32 %r128, %r454, 1664525, 1013904223;
|
|
and.b32 %r129, %r128, 16777215;
|
|
cvt.rn.f32.u32 %f227, %r129;
|
|
fma.rn.f32 %f228, %f227, 0f33800000, %f28;
|
|
mul.f32 %f229, %f23, %f228;
|
|
mad.lo.s32 %r454, %r128, 1664525, 1013904223;
|
|
and.b32 %r130, %r454, 16777215;
|
|
cvt.rn.f32.u32 %f230, %r130;
|
|
cvt.rn.f32.s32 %f231, %r431;
|
|
fma.rn.f32 %f232, %f230, 0f33800000, %f231;
|
|
mul.f32 %f233, %f23, %f232;
|
|
sqrt.rn.f32 %f31, %f229;
|
|
mul.f32 %f938, %f233, 0f40C90FDB;
|
|
abs.f32 %f33, %f938;
|
|
setp.neu.f32 %p14, %f33, 0f7F800000;
|
|
mov.f32 %f932, %f938;
|
|
@%p14 bra BB0_9;
|
|
|
|
mov.f32 %f234, 0f00000000;
|
|
mul.rn.f32 %f932, %f938, %f234;
|
|
|
|
BB0_9:
|
|
mul.f32 %f235, %f932, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r442, %f235;
|
|
cvt.rn.f32.s32 %f236, %r442;
|
|
neg.f32 %f237, %f236;
|
|
mov.f32 %f238, 0f3FC90FDA;
|
|
fma.rn.f32 %f239, %f237, %f238, %f932;
|
|
mov.f32 %f240, 0f33A22168;
|
|
fma.rn.f32 %f241, %f237, %f240, %f239;
|
|
mov.f32 %f242, 0f27C234C5;
|
|
fma.rn.f32 %f933, %f237, %f242, %f241;
|
|
abs.f32 %f243, %f932;
|
|
setp.leu.f32 %p15, %f243, 0f47CE4780;
|
|
@%p15 bra BB0_20;
|
|
|
|
mov.b32 %r13, %f932;
|
|
shr.u32 %r14, %r13, 23;
|
|
shl.b32 %r133, %r13, 8;
|
|
or.b32 %r15, %r133, -2147483648;
|
|
add.u64 %rd42, %SP, 4;
|
|
cvta.to.local.u64 %rd482, %rd42;
|
|
mov.u32 %r434, 0;
|
|
mov.u64 %rd481, __cudart_i2opi_f;
|
|
mov.u32 %r433, -6;
|
|
|
|
BB0_11:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r136, [%rd481];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r134, %r136, %r15, %r434;
|
|
madc.hi.u32 %r434, %r136, %r15, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd482], %r134;
|
|
add.s64 %rd482, %rd482, 4;
|
|
add.s64 %rd481, %rd481, 4;
|
|
add.s32 %r433, %r433, 1;
|
|
setp.ne.s32 %p16, %r433, 0;
|
|
@%p16 bra BB0_11;
|
|
|
|
and.b32 %r139, %r14, 255;
|
|
add.s32 %r140, %r139, -128;
|
|
shr.u32 %r141, %r140, 5;
|
|
and.b32 %r20, %r13, -2147483648;
|
|
cvta.to.local.u64 %rd44, %rd42;
|
|
st.local.u32 [%rd44+24], %r434;
|
|
mov.u32 %r142, 6;
|
|
sub.s32 %r143, %r142, %r141;
|
|
mul.wide.s32 %rd45, %r143, 4;
|
|
add.s64 %rd8, %rd44, %rd45;
|
|
ld.local.u32 %r435, [%rd8];
|
|
ld.local.u32 %r436, [%rd8+-4];
|
|
and.b32 %r23, %r14, 31;
|
|
setp.eq.s32 %p17, %r23, 0;
|
|
@%p17 bra BB0_14;
|
|
|
|
mov.u32 %r144, 32;
|
|
sub.s32 %r145, %r144, %r23;
|
|
shr.u32 %r146, %r436, %r145;
|
|
shl.b32 %r147, %r435, %r23;
|
|
add.s32 %r435, %r146, %r147;
|
|
ld.local.u32 %r148, [%rd8+-8];
|
|
shr.u32 %r149, %r148, %r145;
|
|
shl.b32 %r150, %r436, %r23;
|
|
add.s32 %r436, %r149, %r150;
|
|
|
|
BB0_14:
|
|
shr.u32 %r151, %r436, 30;
|
|
shl.b32 %r152, %r435, 2;
|
|
add.s32 %r437, %r151, %r152;
|
|
shl.b32 %r29, %r436, 2;
|
|
shr.u32 %r153, %r437, 31;
|
|
shr.u32 %r154, %r435, 30;
|
|
add.s32 %r30, %r153, %r154;
|
|
setp.eq.s32 %p18, %r153, 0;
|
|
@%p18 bra BB0_15;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_15:
|
|
mov.u32 %r438, %r20;
|
|
mov.u32 %r439, %r29;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
not.b32 %r155, %r437;
|
|
neg.s32 %r439, %r29;
|
|
setp.eq.s32 %p19, %r29, 0;
|
|
selp.u32 %r156, 1, 0, %p19;
|
|
add.s32 %r437, %r156, %r155;
|
|
xor.b32 %r438, %r20, -2147483648;
|
|
|
|
BB0_17:
|
|
clz.b32 %r441, %r437;
|
|
setp.eq.s32 %p20, %r441, 0;
|
|
shl.b32 %r157, %r437, %r441;
|
|
mov.u32 %r158, 32;
|
|
sub.s32 %r159, %r158, %r441;
|
|
shr.u32 %r160, %r439, %r159;
|
|
add.s32 %r161, %r160, %r157;
|
|
selp.b32 %r38, %r437, %r161, %p20;
|
|
mov.u32 %r162, -921707870;
|
|
mul.hi.u32 %r440, %r38, %r162;
|
|
setp.eq.s32 %p21, %r20, 0;
|
|
neg.s32 %r163, %r30;
|
|
selp.b32 %r442, %r30, %r163, %p21;
|
|
setp.lt.s32 %p22, %r440, 1;
|
|
@%p22 bra BB0_19;
|
|
|
|
mul.lo.s32 %r164, %r38, -921707870;
|
|
shr.u32 %r165, %r164, 31;
|
|
shl.b32 %r166, %r440, 1;
|
|
add.s32 %r440, %r165, %r166;
|
|
add.s32 %r441, %r441, 1;
|
|
|
|
BB0_19:
|
|
mov.u32 %r167, 126;
|
|
sub.s32 %r168, %r167, %r441;
|
|
shl.b32 %r169, %r168, 23;
|
|
add.s32 %r170, %r440, 1;
|
|
shr.u32 %r171, %r170, 7;
|
|
add.s32 %r172, %r171, 1;
|
|
shr.u32 %r173, %r172, 1;
|
|
add.s32 %r174, %r173, %r169;
|
|
or.b32 %r175, %r174, %r438;
|
|
mov.b32 %f933, %r175;
|
|
|
|
BB0_20:
|
|
mul.rn.f32 %f39, %f933, %f933;
|
|
add.s32 %r46, %r442, 1;
|
|
and.b32 %r47, %r46, 1;
|
|
setp.eq.s32 %p23, %r47, 0;
|
|
@%p23 bra BB0_22;
|
|
bra.uni BB0_21;
|
|
|
|
BB0_22:
|
|
mov.f32 %f246, 0f3C08839E;
|
|
mov.f32 %f247, 0fB94CA1F9;
|
|
fma.rn.f32 %f934, %f247, %f39, %f246;
|
|
bra.uni BB0_23;
|
|
|
|
BB0_21:
|
|
mov.f32 %f244, 0fBAB6061A;
|
|
mov.f32 %f245, 0f37CCF5CE;
|
|
fma.rn.f32 %f934, %f245, %f39, %f244;
|
|
|
|
BB0_23:
|
|
@%p23 bra BB0_25;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_25:
|
|
mov.f32 %f251, 0fBE2AAAA3;
|
|
fma.rn.f32 %f252, %f934, %f39, %f251;
|
|
mov.f32 %f253, 0f00000000;
|
|
fma.rn.f32 %f935, %f252, %f39, %f253;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_24:
|
|
mov.f32 %f248, 0f3D2AAAA5;
|
|
fma.rn.f32 %f249, %f934, %f39, %f248;
|
|
mov.f32 %f250, 0fBF000000;
|
|
fma.rn.f32 %f935, %f249, %f39, %f250;
|
|
|
|
BB0_26:
|
|
fma.rn.f32 %f936, %f935, %f933, %f933;
|
|
@%p23 bra BB0_28;
|
|
|
|
mov.f32 %f254, 0f3F800000;
|
|
fma.rn.f32 %f936, %f935, %f39, %f254;
|
|
|
|
BB0_28:
|
|
and.b32 %r176, %r46, 2;
|
|
setp.eq.s32 %p26, %r176, 0;
|
|
@%p26 bra BB0_30;
|
|
|
|
mov.f32 %f255, 0f00000000;
|
|
mov.f32 %f256, 0fBF800000;
|
|
fma.rn.f32 %f936, %f936, %f256, %f255;
|
|
|
|
BB0_30:
|
|
@%p14 bra BB0_32;
|
|
|
|
mov.f32 %f257, 0f00000000;
|
|
mul.rn.f32 %f938, %f938, %f257;
|
|
|
|
BB0_32:
|
|
mul.f32 %f258, %f938, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r452, %f258;
|
|
cvt.rn.f32.s32 %f259, %r452;
|
|
neg.f32 %f260, %f259;
|
|
fma.rn.f32 %f262, %f260, %f238, %f938;
|
|
fma.rn.f32 %f264, %f260, %f240, %f262;
|
|
fma.rn.f32 %f939, %f260, %f242, %f264;
|
|
abs.f32 %f266, %f938;
|
|
setp.leu.f32 %p28, %f266, 0f47CE4780;
|
|
@%p28 bra BB0_43;
|
|
|
|
mov.b32 %r49, %f938;
|
|
shr.u32 %r50, %r49, 23;
|
|
shl.b32 %r179, %r49, 8;
|
|
or.b32 %r51, %r179, -2147483648;
|
|
add.u64 %rd47, %SP, 4;
|
|
cvta.to.local.u64 %rd484, %rd47;
|
|
mov.u32 %r444, 0;
|
|
mov.u64 %rd483, __cudart_i2opi_f;
|
|
mov.u32 %r443, -6;
|
|
|
|
BB0_34:
|
|
.pragma "nounroll";
|
|
ld.const.u32 %r182, [%rd483];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r180, %r182, %r51, %r444;
|
|
madc.hi.u32 %r444, %r182, %r51, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd484], %r180;
|
|
add.s64 %rd484, %rd484, 4;
|
|
add.s64 %rd483, %rd483, 4;
|
|
add.s32 %r443, %r443, 1;
|
|
setp.ne.s32 %p29, %r443, 0;
|
|
@%p29 bra BB0_34;
|
|
|
|
and.b32 %r185, %r50, 255;
|
|
add.s32 %r186, %r185, -128;
|
|
shr.u32 %r187, %r186, 5;
|
|
and.b32 %r56, %r49, -2147483648;
|
|
cvta.to.local.u64 %rd49, %rd47;
|
|
st.local.u32 [%rd49+24], %r444;
|
|
mov.u32 %r188, 6;
|
|
sub.s32 %r189, %r188, %r187;
|
|
mul.wide.s32 %rd50, %r189, 4;
|
|
add.s64 %rd14, %rd49, %rd50;
|
|
ld.local.u32 %r445, [%rd14];
|
|
ld.local.u32 %r446, [%rd14+-4];
|
|
and.b32 %r59, %r50, 31;
|
|
setp.eq.s32 %p30, %r59, 0;
|
|
@%p30 bra BB0_37;
|
|
|
|
mov.u32 %r190, 32;
|
|
sub.s32 %r191, %r190, %r59;
|
|
shr.u32 %r192, %r446, %r191;
|
|
shl.b32 %r193, %r445, %r59;
|
|
add.s32 %r445, %r192, %r193;
|
|
ld.local.u32 %r194, [%rd14+-8];
|
|
shr.u32 %r195, %r194, %r191;
|
|
shl.b32 %r196, %r446, %r59;
|
|
add.s32 %r446, %r195, %r196;
|
|
|
|
BB0_37:
|
|
shr.u32 %r197, %r446, 30;
|
|
shl.b32 %r198, %r445, 2;
|
|
add.s32 %r447, %r197, %r198;
|
|
shl.b32 %r65, %r446, 2;
|
|
shr.u32 %r199, %r447, 31;
|
|
shr.u32 %r200, %r445, 30;
|
|
add.s32 %r66, %r199, %r200;
|
|
setp.eq.s32 %p31, %r199, 0;
|
|
@%p31 bra BB0_38;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
mov.u32 %r448, %r56;
|
|
mov.u32 %r449, %r65;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
not.b32 %r201, %r447;
|
|
neg.s32 %r449, %r65;
|
|
setp.eq.s32 %p32, %r65, 0;
|
|
selp.u32 %r202, 1, 0, %p32;
|
|
add.s32 %r447, %r202, %r201;
|
|
xor.b32 %r448, %r56, -2147483648;
|
|
|
|
BB0_40:
|
|
clz.b32 %r451, %r447;
|
|
setp.eq.s32 %p33, %r451, 0;
|
|
shl.b32 %r203, %r447, %r451;
|
|
mov.u32 %r204, 32;
|
|
sub.s32 %r205, %r204, %r451;
|
|
shr.u32 %r206, %r449, %r205;
|
|
add.s32 %r207, %r206, %r203;
|
|
selp.b32 %r74, %r447, %r207, %p33;
|
|
mov.u32 %r208, -921707870;
|
|
mul.hi.u32 %r450, %r74, %r208;
|
|
setp.eq.s32 %p34, %r56, 0;
|
|
neg.s32 %r209, %r66;
|
|
selp.b32 %r452, %r66, %r209, %p34;
|
|
setp.lt.s32 %p35, %r450, 1;
|
|
@%p35 bra BB0_42;
|
|
|
|
mul.lo.s32 %r210, %r74, -921707870;
|
|
shr.u32 %r211, %r210, 31;
|
|
shl.b32 %r212, %r450, 1;
|
|
add.s32 %r450, %r211, %r212;
|
|
add.s32 %r451, %r451, 1;
|
|
|
|
BB0_42:
|
|
mov.u32 %r213, 126;
|
|
sub.s32 %r214, %r213, %r451;
|
|
shl.b32 %r215, %r214, 23;
|
|
add.s32 %r216, %r450, 1;
|
|
shr.u32 %r217, %r216, 7;
|
|
add.s32 %r218, %r217, 1;
|
|
shr.u32 %r219, %r218, 1;
|
|
add.s32 %r220, %r219, %r215;
|
|
or.b32 %r221, %r220, %r448;
|
|
mov.b32 %f939, %r221;
|
|
|
|
BB0_43:
|
|
mul.rn.f32 %f56, %f939, %f939;
|
|
and.b32 %r82, %r452, 1;
|
|
setp.eq.s32 %p36, %r82, 0;
|
|
@%p36 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f269, 0f3C08839E;
|
|
mov.f32 %f270, 0fB94CA1F9;
|
|
fma.rn.f32 %f940, %f270, %f56, %f269;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f267, 0fBAB6061A;
|
|
mov.f32 %f268, 0f37CCF5CE;
|
|
fma.rn.f32 %f940, %f268, %f56, %f267;
|
|
|
|
BB0_46:
|
|
@%p36 bra BB0_48;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_48:
|
|
mov.f32 %f274, 0fBE2AAAA3;
|
|
fma.rn.f32 %f275, %f940, %f56, %f274;
|
|
mov.f32 %f276, 0f00000000;
|
|
fma.rn.f32 %f941, %f275, %f56, %f276;
|
|
bra.uni BB0_49;
|
|
|
|
BB0_47:
|
|
mov.f32 %f271, 0f3D2AAAA5;
|
|
fma.rn.f32 %f272, %f940, %f56, %f271;
|
|
mov.f32 %f273, 0fBF000000;
|
|
fma.rn.f32 %f941, %f272, %f56, %f273;
|
|
|
|
BB0_49:
|
|
fma.rn.f32 %f942, %f941, %f939, %f939;
|
|
@%p36 bra BB0_51;
|
|
|
|
mov.f32 %f277, 0f3F800000;
|
|
fma.rn.f32 %f942, %f941, %f56, %f277;
|
|
|
|
BB0_51:
|
|
and.b32 %r222, %r452, 2;
|
|
setp.eq.s32 %p39, %r222, 0;
|
|
@%p39 bra BB0_53;
|
|
|
|
mov.f32 %f278, 0f00000000;
|
|
mov.f32 %f279, 0fBF800000;
|
|
fma.rn.f32 %f942, %f942, %f279, %f278;
|
|
|
|
BB0_53:
|
|
mul.f32 %f288, %f31, %f936;
|
|
add.u64 %rd51, %SP, 0;
|
|
cvta.to.local.u64 %rd52, %rd51;
|
|
mul.f32 %f289, %f288, %f288;
|
|
mov.f32 %f290, 0f3F800000;
|
|
sub.f32 %f291, %f290, %f289;
|
|
mul.f32 %f292, %f31, %f942;
|
|
mul.f32 %f293, %f292, %f292;
|
|
sub.f32 %f294, %f291, %f293;
|
|
mov.f32 %f295, 0f00000000;
|
|
max.f32 %f296, %f295, %f294;
|
|
sqrt.rn.f32 %f297, %f296;
|
|
mul.f32 %f298, %f17, %f292;
|
|
mul.f32 %f299, %f18, %f292;
|
|
mul.f32 %f300, %f19, %f292;
|
|
fma.rn.f32 %f301, %f20, %f288, %f298;
|
|
fma.rn.f32 %f302, %f21, %f288, %f299;
|
|
fma.rn.f32 %f303, %f22, %f288, %f300;
|
|
fma.rn.f32 %f304, %f7, %f297, %f301;
|
|
fma.rn.f32 %f305, %f8, %f297, %f302;
|
|
fma.rn.f32 %f306, %f9, %f297, %f303;
|
|
add.f32 %f307, %f7, %f304;
|
|
add.f32 %f308, %f8, %f305;
|
|
add.f32 %f309, %f9, %f306;
|
|
ld.global.f32 %f310, [shadowSpread];
|
|
mul.f32 %f311, %f310, %f307;
|
|
mul.f32 %f312, %f310, %f308;
|
|
mul.f32 %f313, %f310, %f309;
|
|
sub.f32 %f314, %f311, %f7;
|
|
sub.f32 %f315, %f312, %f8;
|
|
sub.f32 %f316, %f313, %f9;
|
|
mul.f32 %f317, %f315, %f315;
|
|
fma.rn.f32 %f318, %f314, %f314, %f317;
|
|
fma.rn.f32 %f319, %f316, %f316, %f318;
|
|
sqrt.rn.f32 %f320, %f319;
|
|
rcp.rn.f32 %f321, %f320;
|
|
mul.f32 %f283, %f321, %f314;
|
|
mul.f32 %f284, %f321, %f315;
|
|
mul.f32 %f285, %f321, %f316;
|
|
ld.global.u32 %r226, [imageEnabled];
|
|
and.b32 %r227, %r226, 32;
|
|
setp.eq.s32 %p40, %r227, 0;
|
|
selp.f32 %f322, 0f3F800000, 0f41200000, %p40;
|
|
mul.f32 %f286, %f322, %f29;
|
|
mov.u32 %r228, 1065353216;
|
|
st.local.u32 [%rd52], %r228;
|
|
ld.global.u32 %r223, [root];
|
|
mov.u32 %r224, 1;
|
|
mov.f32 %f287, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r223, %f14, %f15, %f16, %f283, %f284, %f285, %r224, %f286, %f287, %rd51, %r100);
|
|
// inline asm
|
|
ld.local.f32 %f323, [%rd52];
|
|
add.f32 %f944, %f944, %f323;
|
|
ld.global.u32 %r428, [samples];
|
|
add.s32 %r431, %r431, 1;
|
|
setp.lt.s32 %p41, %r431, %r428;
|
|
@%p41 bra BB0_7;
|
|
|
|
BB0_54:
|
|
add.s32 %r429, %r429, 1;
|
|
setp.lt.s32 %p42, %r429, %r428;
|
|
@%p42 bra BB0_5;
|
|
|
|
BB0_55:
|
|
setp.eq.s32 %p43, %r428, 0;
|
|
mov.f32 %f946, 0f3F800000;
|
|
@%p43 bra BB0_57;
|
|
|
|
mul.lo.s32 %r229, %r428, %r428;
|
|
cvt.rn.f32.s32 %f325, %r229;
|
|
div.rn.f32 %f946, %f944, %f325;
|
|
|
|
BB0_57:
|
|
mul.f32 %f334, %f12, %f21;
|
|
fma.rn.f32 %f335, %f11, %f20, %f334;
|
|
fma.rn.f32 %f336, %f13, %f22, %f335;
|
|
ld.global.v4.f32 {%f337, %f338, %f339, %f340}, [lightTilingOffset];
|
|
fma.rn.f32 %f330, %f336, %f337, %f339;
|
|
mul.f32 %f343, %f12, %f18;
|
|
fma.rn.f32 %f344, %f11, %f17, %f343;
|
|
fma.rn.f32 %f345, %f13, %f19, %f344;
|
|
fma.rn.f32 %f331, %f345, %f338, %f340;
|
|
ld.global.u32 %r230, [lightCookie];
|
|
mov.f32 %f333, 0f00000000;
|
|
// inline asm
|
|
call (%f326, %f327, %f328, %f329), _rt_texture_get_f_id, (%r230, %r99, %f330, %f331, %f333, %f333);
|
|
// inline asm
|
|
min.f32 %f348, %f326, %f327;
|
|
min.f32 %f349, %f348, %f328;
|
|
mul.f32 %f73, %f946, %f349;
|
|
ld.global.f32 %f350, [directColor];
|
|
mul.f32 %f351, %f946, %f350;
|
|
ld.global.f32 %f352, [directColor+4];
|
|
mul.f32 %f353, %f946, %f352;
|
|
ld.global.f32 %f354, [directColor+8];
|
|
mul.f32 %f355, %f946, %f354;
|
|
mul.f32 %f74, %f326, %f351;
|
|
mul.f32 %f75, %f327, %f353;
|
|
mul.f32 %f76, %f328, %f355;
|
|
cvt.sat.f32.f32 %f356, %f10;
|
|
mul.f32 %f77, %f74, %f356;
|
|
mul.f32 %f78, %f75, %f356;
|
|
mul.f32 %f79, %f356, %f76;
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
and.b32 %r232, %r457, 8;
|
|
setp.eq.s32 %p44, %r232, 0;
|
|
@%p44 bra BB0_70;
|
|
|
|
cvt.u64.u32 %rd55, %r2;
|
|
cvt.u64.u32 %rd56, %r3;
|
|
mov.u64 %rd59, image_Mask;
|
|
cvta.global.u64 %rd54, %rd59;
|
|
// inline asm
|
|
call (%rd53), _rt_buffer_get_64, (%rd54, %r99, %r99, %rd55, %rd56, %rd25, %rd25);
|
|
// inline asm
|
|
abs.f32 %f81, %f73;
|
|
setp.lt.f32 %p45, %f81, 0f00800000;
|
|
mul.f32 %f362, %f81, 0f4B800000;
|
|
selp.f32 %f363, 0fC3170000, 0fC2FE0000, %p45;
|
|
selp.f32 %f364, %f362, %f81, %p45;
|
|
mov.b32 %r235, %f364;
|
|
and.b32 %r236, %r235, 8388607;
|
|
or.b32 %r237, %r236, 1065353216;
|
|
mov.b32 %f365, %r237;
|
|
shr.u32 %r238, %r235, 23;
|
|
cvt.rn.f32.u32 %f366, %r238;
|
|
add.f32 %f367, %f363, %f366;
|
|
setp.gt.f32 %p46, %f365, 0f3FB504F3;
|
|
mul.f32 %f368, %f365, 0f3F000000;
|
|
add.f32 %f369, %f367, 0f3F800000;
|
|
selp.f32 %f370, %f368, %f365, %p46;
|
|
selp.f32 %f371, %f369, %f367, %p46;
|
|
add.f32 %f372, %f370, 0fBF800000;
|
|
add.f32 %f358, %f370, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f357,%f358;
|
|
// inline asm
|
|
add.f32 %f373, %f372, %f372;
|
|
mul.f32 %f374, %f357, %f373;
|
|
mul.f32 %f375, %f374, %f374;
|
|
mov.f32 %f376, 0f3C4CAF63;
|
|
mov.f32 %f377, 0f3B18F0FE;
|
|
fma.rn.f32 %f378, %f377, %f375, %f376;
|
|
mov.f32 %f379, 0f3DAAAABD;
|
|
fma.rn.f32 %f380, %f378, %f375, %f379;
|
|
mul.rn.f32 %f381, %f380, %f375;
|
|
mul.rn.f32 %f382, %f381, %f374;
|
|
sub.f32 %f383, %f372, %f374;
|
|
neg.f32 %f384, %f374;
|
|
add.f32 %f385, %f383, %f383;
|
|
fma.rn.f32 %f386, %f384, %f372, %f385;
|
|
mul.rn.f32 %f387, %f357, %f386;
|
|
add.f32 %f388, %f382, %f374;
|
|
sub.f32 %f389, %f374, %f388;
|
|
add.f32 %f390, %f382, %f389;
|
|
add.f32 %f391, %f387, %f390;
|
|
add.f32 %f392, %f388, %f391;
|
|
sub.f32 %f393, %f388, %f392;
|
|
add.f32 %f394, %f391, %f393;
|
|
mov.f32 %f395, 0f3F317200;
|
|
mul.rn.f32 %f396, %f371, %f395;
|
|
mov.f32 %f397, 0f35BFBE8E;
|
|
mul.rn.f32 %f398, %f371, %f397;
|
|
add.f32 %f399, %f396, %f392;
|
|
sub.f32 %f400, %f396, %f399;
|
|
add.f32 %f401, %f392, %f400;
|
|
add.f32 %f402, %f394, %f401;
|
|
add.f32 %f403, %f398, %f402;
|
|
add.f32 %f404, %f399, %f403;
|
|
sub.f32 %f405, %f399, %f404;
|
|
add.f32 %f406, %f403, %f405;
|
|
mov.f32 %f407, 0f3EE8BA2E;
|
|
mul.rn.f32 %f408, %f407, %f404;
|
|
neg.f32 %f409, %f408;
|
|
fma.rn.f32 %f410, %f407, %f404, %f409;
|
|
fma.rn.f32 %f411, %f407, %f406, %f410;
|
|
fma.rn.f32 %f413, %f333, %f404, %f411;
|
|
add.rn.f32 %f414, %f408, %f413;
|
|
neg.f32 %f415, %f414;
|
|
add.rn.f32 %f416, %f408, %f415;
|
|
add.rn.f32 %f417, %f416, %f413;
|
|
mov.b32 %r239, %f414;
|
|
setp.eq.s32 %p47, %r239, 1118925336;
|
|
add.s32 %r240, %r239, -1;
|
|
mov.b32 %f418, %r240;
|
|
add.f32 %f419, %f417, 0f37000000;
|
|
selp.f32 %f420, %f418, %f414, %p47;
|
|
selp.f32 %f82, %f419, %f417, %p47;
|
|
mul.f32 %f421, %f420, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f422, %f421;
|
|
mov.f32 %f423, 0fBF317200;
|
|
fma.rn.f32 %f424, %f422, %f423, %f420;
|
|
mov.f32 %f425, 0fB5BFBE8E;
|
|
fma.rn.f32 %f426, %f422, %f425, %f424;
|
|
mul.f32 %f427, %f426, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f428, %f427;
|
|
add.f32 %f429, %f422, 0f00000000;
|
|
ex2.approx.f32 %f430, %f429;
|
|
mul.f32 %f431, %f428, %f430;
|
|
setp.lt.f32 %p48, %f420, 0fC2D20000;
|
|
selp.f32 %f432, 0f00000000, %f431, %p48;
|
|
setp.gt.f32 %p49, %f420, 0f42D20000;
|
|
selp.f32 %f947, 0f7F800000, %f432, %p49;
|
|
setp.eq.f32 %p50, %f947, 0f7F800000;
|
|
@%p50 bra BB0_60;
|
|
|
|
fma.rn.f32 %f947, %f947, %f82, %f947;
|
|
|
|
BB0_60:
|
|
mov.f32 %f918, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f917, %f918;
|
|
fma.rn.f32 %f916, %f917, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f915, %f916;
|
|
setp.lt.f32 %p51, %f73, 0f00000000;
|
|
setp.eq.f32 %p52, %f915, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
mov.b32 %r241, %f947;
|
|
xor.b32 %r242, %r241, -2147483648;
|
|
mov.b32 %f433, %r242;
|
|
selp.f32 %f949, %f433, %f947, %p1;
|
|
setp.eq.f32 %p53, %f73, 0f00000000;
|
|
@%p53 bra BB0_63;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_63:
|
|
add.f32 %f436, %f73, %f73;
|
|
selp.f32 %f949, %f436, 0f00000000, %p52;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_143:
|
|
mov.u64 %rd300, image_HDR;
|
|
cvta.global.u64 %rd295, %rd300;
|
|
mov.u32 %r363, 8;
|
|
// inline asm
|
|
call (%rd294), _rt_buffer_get_64, (%rd295, %r99, %r363, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f820, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs179, %f820;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs180, 0;
|
|
st.v4.u16 [%rd294], {%rs179, %rs179, %rs179, %rs180};
|
|
|
|
BB0_144:
|
|
ld.global.u32 %r364, [additive];
|
|
setp.eq.s32 %p125, %r364, 0;
|
|
@%p125 bra BB0_146;
|
|
|
|
mov.u64 %rd313, image_RNM0;
|
|
cvta.global.u64 %rd302, %rd313;
|
|
mov.u32 %r368, 8;
|
|
// inline asm
|
|
call (%rd301), _rt_buffer_get_64, (%rd302, %r99, %r368, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs187, %rs188, %rs189, %rs190}, [%rd301];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f821, %rs187;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f822, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f823, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd307), _rt_buffer_get_64, (%rd302, %r99, %r368, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f824, %f821, 0f00000000;
|
|
add.f32 %f825, %f822, 0f00000000;
|
|
add.f32 %f826, %f823, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f826;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f825;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs184, %f824;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs191, 0;
|
|
st.v4.u16 [%rd307], {%rs184, %rs185, %rs186, %rs191};
|
|
bra.uni BB0_147;
|
|
|
|
BB0_146:
|
|
mov.u64 %rd320, image_RNM0;
|
|
cvta.global.u64 %rd315, %rd320;
|
|
mov.u32 %r370, 8;
|
|
// inline asm
|
|
call (%rd314), _rt_buffer_get_64, (%rd315, %r99, %r370, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f827, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f827;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs193, 0;
|
|
st.v4.u16 [%rd314], {%rs192, %rs192, %rs192, %rs193};
|
|
|
|
BB0_147:
|
|
ld.global.u32 %r371, [additive];
|
|
setp.eq.s32 %p126, %r371, 0;
|
|
@%p126 bra BB0_149;
|
|
|
|
mov.u64 %rd333, image_RNM1;
|
|
cvta.global.u64 %rd322, %rd333;
|
|
mov.u32 %r375, 8;
|
|
// inline asm
|
|
call (%rd321), _rt_buffer_get_64, (%rd322, %r99, %r375, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs200, %rs201, %rs202, %rs203}, [%rd321];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f828, %rs200;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f829, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f830, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd327), _rt_buffer_get_64, (%rd322, %r99, %r375, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f831, %f828, 0f00000000;
|
|
add.f32 %f832, %f829, 0f00000000;
|
|
add.f32 %f833, %f830, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f833;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f832;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs197, %f831;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs204, 0;
|
|
st.v4.u16 [%rd327], {%rs197, %rs198, %rs199, %rs204};
|
|
bra.uni BB0_150;
|
|
|
|
BB0_149:
|
|
mov.u64 %rd340, image_RNM1;
|
|
cvta.global.u64 %rd335, %rd340;
|
|
mov.u32 %r377, 8;
|
|
// inline asm
|
|
call (%rd334), _rt_buffer_get_64, (%rd335, %r99, %r377, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f834, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f834;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs206, 0;
|
|
st.v4.u16 [%rd334], {%rs205, %rs205, %rs205, %rs206};
|
|
|
|
BB0_150:
|
|
ld.global.u32 %r378, [additive];
|
|
setp.eq.s32 %p127, %r378, 0;
|
|
@%p127 bra BB0_152;
|
|
|
|
mov.u64 %rd353, image_RNM2;
|
|
cvta.global.u64 %rd342, %rd353;
|
|
mov.u32 %r382, 8;
|
|
// inline asm
|
|
call (%rd341), _rt_buffer_get_64, (%rd342, %r99, %r382, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs213, %rs214, %rs215, %rs216}, [%rd341];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f835, %rs213;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f836, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f837, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd347), _rt_buffer_get_64, (%rd342, %r99, %r382, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f838, %f835, 0f00000000;
|
|
add.f32 %f839, %f836, 0f00000000;
|
|
add.f32 %f840, %f837, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f840;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f839;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs210, %f838;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs217, 0;
|
|
st.v4.u16 [%rd347], {%rs210, %rs211, %rs212, %rs217};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd360, image_RNM2;
|
|
cvta.global.u64 %rd355, %rd360;
|
|
mov.u32 %r384, 8;
|
|
// inline asm
|
|
call (%rd354), _rt_buffer_get_64, (%rd355, %r99, %r384, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f841, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs218, %f841;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs219, 0;
|
|
st.v4.u16 [%rd354], {%rs218, %rs218, %rs218, %rs219};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r385, [additive];
|
|
setp.eq.s32 %p128, %r385, 0;
|
|
@%p128 bra BB0_155;
|
|
|
|
mov.u64 %rd373, image_RNM3;
|
|
cvta.global.u64 %rd362, %rd373;
|
|
mov.u32 %r389, 8;
|
|
// inline asm
|
|
call (%rd361), _rt_buffer_get_64, (%rd362, %r99, %r389, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs226, %rs227, %rs228, %rs229}, [%rd361];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f842, %rs226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f843, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f844, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd367), _rt_buffer_get_64, (%rd362, %r99, %r389, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f845, %f842, 0f00000000;
|
|
add.f32 %f846, %f843, 0f00000000;
|
|
add.f32 %f847, %f844, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f847;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f846;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs223, %f845;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs230, 0;
|
|
st.v4.u16 [%rd367], {%rs223, %rs224, %rs225, %rs230};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd380, image_RNM3;
|
|
cvta.global.u64 %rd375, %rd380;
|
|
mov.u32 %r391, 8;
|
|
// inline asm
|
|
call (%rd374), _rt_buffer_get_64, (%rd375, %r99, %r391, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f848, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs231, %f848;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs232, 0;
|
|
st.v4.u16 [%rd374], {%rs231, %rs231, %rs231, %rs232};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r392, [additive];
|
|
setp.eq.s32 %p129, %r392, 0;
|
|
@%p129 bra BB0_158;
|
|
|
|
mov.u64 %rd393, image_RNM4;
|
|
cvta.global.u64 %rd382, %rd393;
|
|
mov.u32 %r396, 8;
|
|
// inline asm
|
|
call (%rd381), _rt_buffer_get_64, (%rd382, %r99, %r396, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs239, %rs240, %rs241, %rs242}, [%rd381];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f849, %rs239;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f850, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd387), _rt_buffer_get_64, (%rd382, %r99, %r396, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f852, %f849, 0f00000000;
|
|
add.f32 %f853, %f850, 0f00000000;
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f854;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f853;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs236, %f852;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs243, 0;
|
|
st.v4.u16 [%rd387], {%rs236, %rs237, %rs238, %rs243};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd400, image_RNM4;
|
|
cvta.global.u64 %rd395, %rd400;
|
|
mov.u32 %r398, 8;
|
|
// inline asm
|
|
call (%rd394), _rt_buffer_get_64, (%rd395, %r99, %r398, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f855, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs244, %f855;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs245, 0;
|
|
st.v4.u16 [%rd394], {%rs244, %rs244, %rs244, %rs245};
|
|
|
|
BB0_159:
|
|
ld.global.u32 %r399, [additive];
|
|
setp.eq.s32 %p130, %r399, 0;
|
|
@%p130 bra BB0_161;
|
|
|
|
mov.u64 %rd413, image_RNM5;
|
|
cvta.global.u64 %rd402, %rd413;
|
|
mov.u32 %r403, 8;
|
|
// inline asm
|
|
call (%rd401), _rt_buffer_get_64, (%rd402, %r99, %r403, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs252, %rs253, %rs254, %rs255}, [%rd401];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f856, %rs252;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f857, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd407), _rt_buffer_get_64, (%rd402, %r99, %r403, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f859, %f856, 0f00000000;
|
|
add.f32 %f860, %f857, 0f00000000;
|
|
add.f32 %f861, %f858, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f861;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f860;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs249, %f859;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs256, 0;
|
|
st.v4.u16 [%rd407], {%rs249, %rs250, %rs251, %rs256};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_161:
|
|
mov.u64 %rd420, image_RNM5;
|
|
cvta.global.u64 %rd415, %rd420;
|
|
mov.u32 %r405, 8;
|
|
// inline asm
|
|
call (%rd414), _rt_buffer_get_64, (%rd415, %r99, %r405, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f862, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs257, %f862;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs258, 0;
|
|
st.v4.u16 [%rd414], {%rs257, %rs257, %rs257, %rs258};
|
|
|
|
BB0_162:
|
|
ld.global.u32 %r406, [additive];
|
|
setp.eq.s32 %p131, %r406, 0;
|
|
@%p131 bra BB0_164;
|
|
|
|
mov.u64 %rd433, image_RNM6;
|
|
cvta.global.u64 %rd422, %rd433;
|
|
mov.u32 %r410, 8;
|
|
// inline asm
|
|
call (%rd421), _rt_buffer_get_64, (%rd422, %r99, %r410, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs265, %rs266, %rs267, %rs268}, [%rd421];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f863, %rs265;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f864, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f865, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd427), _rt_buffer_get_64, (%rd422, %r99, %r410, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f866, %f863, 0f00000000;
|
|
add.f32 %f867, %f864, 0f00000000;
|
|
add.f32 %f868, %f865, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f868;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f867;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs262, %f866;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs269, 0;
|
|
st.v4.u16 [%rd427], {%rs262, %rs263, %rs264, %rs269};
|
|
bra.uni BB0_165;
|
|
|
|
BB0_164:
|
|
mov.u64 %rd440, image_RNM6;
|
|
cvta.global.u64 %rd435, %rd440;
|
|
mov.u32 %r412, 8;
|
|
// inline asm
|
|
call (%rd434), _rt_buffer_get_64, (%rd435, %r99, %r412, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f869, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs270, %f869;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs271, 0;
|
|
st.v4.u16 [%rd434], {%rs270, %rs270, %rs270, %rs271};
|
|
|
|
BB0_165:
|
|
ld.global.u32 %r413, [additive];
|
|
setp.eq.s32 %p132, %r413, 0;
|
|
@%p132 bra BB0_167;
|
|
|
|
mov.u64 %rd453, image_RNM7;
|
|
cvta.global.u64 %rd442, %rd453;
|
|
mov.u32 %r417, 8;
|
|
// inline asm
|
|
call (%rd441), _rt_buffer_get_64, (%rd442, %r99, %r417, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs278, %rs279, %rs280, %rs281}, [%rd441];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f870, %rs278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f871, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f872, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd447), _rt_buffer_get_64, (%rd442, %r99, %r417, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f873, %f870, 0f00000000;
|
|
add.f32 %f874, %f871, 0f00000000;
|
|
add.f32 %f875, %f872, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f875;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f874;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs275, %f873;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs282, 0;
|
|
st.v4.u16 [%rd447], {%rs275, %rs276, %rs277, %rs282};
|
|
bra.uni BB0_168;
|
|
|
|
BB0_167:
|
|
mov.u64 %rd460, image_RNM7;
|
|
cvta.global.u64 %rd455, %rd460;
|
|
mov.u32 %r419, 8;
|
|
// inline asm
|
|
call (%rd454), _rt_buffer_get_64, (%rd455, %r99, %r419, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f876, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs283, %f876;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs284, 0;
|
|
st.v4.u16 [%rd454], {%rs283, %rs283, %rs283, %rs284};
|
|
|
|
BB0_168:
|
|
ld.global.u32 %r420, [additive];
|
|
setp.eq.s32 %p133, %r420, 0;
|
|
@%p133 bra BB0_170;
|
|
|
|
mov.u64 %rd473, image_RNM8;
|
|
cvta.global.u64 %rd462, %rd473;
|
|
mov.u32 %r424, 8;
|
|
// inline asm
|
|
call (%rd461), _rt_buffer_get_64, (%rd462, %r99, %r424, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs291, %rs292, %rs293, %rs294}, [%rd461];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f877, %rs291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f878, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f879, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd467), _rt_buffer_get_64, (%rd462, %r99, %r424, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f880, %f877, 0f00000000;
|
|
add.f32 %f881, %f878, 0f00000000;
|
|
add.f32 %f882, %f879, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f882;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f881;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs288, %f880;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs295, 0;
|
|
st.v4.u16 [%rd467], {%rs288, %rs289, %rs290, %rs295};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_170:
|
|
mov.u64 %rd480, image_RNM8;
|
|
cvta.global.u64 %rd475, %rd480;
|
|
mov.u32 %r426, 8;
|
|
// inline asm
|
|
call (%rd474), _rt_buffer_get_64, (%rd475, %r99, %r426, %rd18, %rd19, %rd25, %rd25);
|
|
// inline asm
|
|
mov.f32 %f883, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs296, %f883;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs297, 0;
|
|
st.v4.u16 [%rd474], {%rs296, %rs296, %rs296, %rs297};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_61:
|
|
setp.geu.f32 %p54, %f73, 0f00000000;
|
|
@%p54 bra BB0_64;
|
|
|
|
mov.f32 %f923, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f435, %f923;
|
|
setp.neu.f32 %p55, %f435, 0f3EE8BA2E;
|
|
selp.f32 %f949, 0f7FFFFFFF, %f949, %p55;
|
|
|
|
BB0_64:
|
|
abs.f32 %f919, %f73;
|
|
add.f32 %f437, %f919, 0f3EE8BA2E;
|
|
mov.b32 %r243, %f437;
|
|
setp.lt.s32 %p57, %r243, 2139095040;
|
|
@%p57 bra BB0_69;
|
|
|
|
abs.f32 %f921, %f73;
|
|
setp.gtu.f32 %p58, %f921, 0f7F800000;
|
|
@%p58 bra BB0_68;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_68:
|
|
add.f32 %f949, %f73, 0f3EE8BA2E;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_66:
|
|
abs.f32 %f922, %f73;
|
|
setp.neu.f32 %p59, %f922, 0f7F800000;
|
|
@%p59 bra BB0_69;
|
|
|
|
selp.f32 %f949, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_69:
|
|
mul.f32 %f438, %f949, 0f437F0000;
|
|
setp.eq.f32 %p60, %f73, 0f3F800000;
|
|
selp.f32 %f439, 0f437F0000, %f438, %p60;
|
|
cvt.rzi.u32.f32 %r244, %f439;
|
|
cvt.u16.u32 %rs19, %r244;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd53], {%rs19, %rs20};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_70:
|
|
and.b32 %r245, %r457, 1;
|
|
setp.eq.b32 %p61, %r245, 1;
|
|
@!%p61 bra BB0_105;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_71:
|
|
mov.f32 %f920, 0f00000000;
|
|
abs.f32 %f94, %f77;
|
|
setp.lt.f32 %p62, %f94, 0f00800000;
|
|
mul.f32 %f445, %f94, 0f4B800000;
|
|
selp.f32 %f446, 0fC3170000, 0fC2FE0000, %p62;
|
|
selp.f32 %f447, %f445, %f94, %p62;
|
|
mov.b32 %r246, %f447;
|
|
and.b32 %r247, %r246, 8388607;
|
|
or.b32 %r248, %r247, 1065353216;
|
|
mov.b32 %f448, %r248;
|
|
shr.u32 %r249, %r246, 23;
|
|
cvt.rn.f32.u32 %f449, %r249;
|
|
add.f32 %f450, %f446, %f449;
|
|
setp.gt.f32 %p63, %f448, 0f3FB504F3;
|
|
mul.f32 %f451, %f448, 0f3F000000;
|
|
add.f32 %f452, %f450, 0f3F800000;
|
|
selp.f32 %f453, %f451, %f448, %p63;
|
|
selp.f32 %f454, %f452, %f450, %p63;
|
|
add.f32 %f455, %f453, 0fBF800000;
|
|
add.f32 %f441, %f453, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f440,%f441;
|
|
// inline asm
|
|
add.f32 %f456, %f455, %f455;
|
|
mul.f32 %f457, %f440, %f456;
|
|
mul.f32 %f458, %f457, %f457;
|
|
mov.f32 %f459, 0f3C4CAF63;
|
|
mov.f32 %f460, 0f3B18F0FE;
|
|
fma.rn.f32 %f461, %f460, %f458, %f459;
|
|
mov.f32 %f462, 0f3DAAAABD;
|
|
fma.rn.f32 %f463, %f461, %f458, %f462;
|
|
mul.rn.f32 %f464, %f463, %f458;
|
|
mul.rn.f32 %f465, %f464, %f457;
|
|
sub.f32 %f466, %f455, %f457;
|
|
neg.f32 %f467, %f457;
|
|
add.f32 %f468, %f466, %f466;
|
|
fma.rn.f32 %f469, %f467, %f455, %f468;
|
|
mul.rn.f32 %f470, %f440, %f469;
|
|
add.f32 %f471, %f465, %f457;
|
|
sub.f32 %f472, %f457, %f471;
|
|
add.f32 %f473, %f465, %f472;
|
|
add.f32 %f474, %f470, %f473;
|
|
add.f32 %f475, %f471, %f474;
|
|
sub.f32 %f476, %f471, %f475;
|
|
add.f32 %f477, %f474, %f476;
|
|
mov.f32 %f478, 0f3F317200;
|
|
mul.rn.f32 %f479, %f454, %f478;
|
|
mov.f32 %f480, 0f35BFBE8E;
|
|
mul.rn.f32 %f481, %f454, %f480;
|
|
add.f32 %f482, %f479, %f475;
|
|
sub.f32 %f483, %f479, %f482;
|
|
add.f32 %f484, %f475, %f483;
|
|
add.f32 %f485, %f477, %f484;
|
|
add.f32 %f486, %f481, %f485;
|
|
add.f32 %f487, %f482, %f486;
|
|
sub.f32 %f488, %f482, %f487;
|
|
add.f32 %f489, %f486, %f488;
|
|
mov.f32 %f490, 0f3EE66666;
|
|
mul.rn.f32 %f491, %f490, %f487;
|
|
neg.f32 %f492, %f491;
|
|
fma.rn.f32 %f493, %f490, %f487, %f492;
|
|
fma.rn.f32 %f494, %f490, %f489, %f493;
|
|
fma.rn.f32 %f496, %f920, %f487, %f494;
|
|
add.rn.f32 %f497, %f491, %f496;
|
|
neg.f32 %f498, %f497;
|
|
add.rn.f32 %f499, %f491, %f498;
|
|
add.rn.f32 %f500, %f499, %f496;
|
|
mov.b32 %r250, %f497;
|
|
setp.eq.s32 %p64, %r250, 1118925336;
|
|
add.s32 %r251, %r250, -1;
|
|
mov.b32 %f501, %r251;
|
|
add.f32 %f502, %f500, 0f37000000;
|
|
selp.f32 %f503, %f501, %f497, %p64;
|
|
selp.f32 %f95, %f502, %f500, %p64;
|
|
mul.f32 %f504, %f503, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f505, %f504;
|
|
mov.f32 %f506, 0fBF317200;
|
|
fma.rn.f32 %f507, %f505, %f506, %f503;
|
|
mov.f32 %f508, 0fB5BFBE8E;
|
|
fma.rn.f32 %f509, %f505, %f508, %f507;
|
|
mul.f32 %f510, %f509, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f511, %f510;
|
|
add.f32 %f512, %f505, 0f00000000;
|
|
ex2.approx.f32 %f513, %f512;
|
|
mul.f32 %f514, %f511, %f513;
|
|
setp.lt.f32 %p65, %f503, 0fC2D20000;
|
|
selp.f32 %f515, 0f00000000, %f514, %p65;
|
|
setp.gt.f32 %p66, %f503, 0f42D20000;
|
|
selp.f32 %f950, 0f7F800000, %f515, %p66;
|
|
setp.eq.f32 %p67, %f950, 0f7F800000;
|
|
@%p67 bra BB0_73;
|
|
|
|
fma.rn.f32 %f950, %f950, %f95, %f950;
|
|
|
|
BB0_73:
|
|
mov.f32 %f887, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f886, %f887;
|
|
fma.rn.f32 %f885, %f886, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f884, %f885;
|
|
setp.lt.f32 %p68, %f77, 0f00000000;
|
|
setp.eq.f32 %p69, %f884, 0f3F800000;
|
|
and.pred %p2, %p68, %p69;
|
|
mov.b32 %r252, %f950;
|
|
xor.b32 %r253, %r252, -2147483648;
|
|
mov.b32 %f516, %r253;
|
|
selp.f32 %f952, %f516, %f950, %p2;
|
|
setp.eq.f32 %p70, %f77, 0f00000000;
|
|
@%p70 bra BB0_76;
|
|
bra.uni BB0_74;
|
|
|
|
BB0_76:
|
|
add.f32 %f519, %f77, %f77;
|
|
selp.f32 %f952, %f519, 0f00000000, %p69;
|
|
bra.uni BB0_77;
|
|
|
|
BB0_74:
|
|
setp.geu.f32 %p71, %f77, 0f00000000;
|
|
@%p71 bra BB0_77;
|
|
|
|
mov.f32 %f911, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f518, %f911;
|
|
setp.neu.f32 %p72, %f518, 0f3EE66666;
|
|
selp.f32 %f952, 0f7FFFFFFF, %f952, %p72;
|
|
|
|
BB0_77:
|
|
abs.f32 %f888, %f77;
|
|
add.f32 %f520, %f888, 0f3EE66666;
|
|
mov.b32 %r254, %f520;
|
|
setp.lt.s32 %p74, %r254, 2139095040;
|
|
@%p74 bra BB0_82;
|
|
|
|
abs.f32 %f909, %f77;
|
|
setp.gtu.f32 %p75, %f909, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
bra.uni BB0_79;
|
|
|
|
BB0_81:
|
|
add.f32 %f952, %f77, 0f3EE66666;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_79:
|
|
abs.f32 %f910, %f77;
|
|
setp.neu.f32 %p76, %f910, 0f7F800000;
|
|
@%p76 bra BB0_82;
|
|
|
|
selp.f32 %f952, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_82:
|
|
mov.f32 %f897, 0fB5BFBE8E;
|
|
mov.f32 %f896, 0fBF317200;
|
|
mov.f32 %f895, 0f35BFBE8E;
|
|
mov.f32 %f894, 0f3F317200;
|
|
mov.f32 %f893, 0f3DAAAABD;
|
|
mov.f32 %f892, 0f3C4CAF63;
|
|
mov.f32 %f891, 0f3B18F0FE;
|
|
mov.f32 %f890, 0f3EE66666;
|
|
mov.f32 %f889, 0f00000000;
|
|
setp.eq.f32 %p77, %f77, 0f3F800000;
|
|
selp.f32 %f106, 0f3F800000, %f952, %p77;
|
|
abs.f32 %f107, %f78;
|
|
setp.lt.f32 %p78, %f107, 0f00800000;
|
|
mul.f32 %f523, %f107, 0f4B800000;
|
|
selp.f32 %f524, 0fC3170000, 0fC2FE0000, %p78;
|
|
selp.f32 %f525, %f523, %f107, %p78;
|
|
mov.b32 %r255, %f525;
|
|
and.b32 %r256, %r255, 8388607;
|
|
or.b32 %r257, %r256, 1065353216;
|
|
mov.b32 %f526, %r257;
|
|
shr.u32 %r258, %r255, 23;
|
|
cvt.rn.f32.u32 %f527, %r258;
|
|
add.f32 %f528, %f524, %f527;
|
|
setp.gt.f32 %p79, %f526, 0f3FB504F3;
|
|
mul.f32 %f529, %f526, 0f3F000000;
|
|
add.f32 %f530, %f528, 0f3F800000;
|
|
selp.f32 %f531, %f529, %f526, %p79;
|
|
selp.f32 %f532, %f530, %f528, %p79;
|
|
add.f32 %f533, %f531, 0fBF800000;
|
|
add.f32 %f522, %f531, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f521,%f522;
|
|
// inline asm
|
|
add.f32 %f534, %f533, %f533;
|
|
mul.f32 %f535, %f521, %f534;
|
|
mul.f32 %f536, %f535, %f535;
|
|
fma.rn.f32 %f539, %f891, %f536, %f892;
|
|
fma.rn.f32 %f541, %f539, %f536, %f893;
|
|
mul.rn.f32 %f542, %f541, %f536;
|
|
mul.rn.f32 %f543, %f542, %f535;
|
|
sub.f32 %f544, %f533, %f535;
|
|
neg.f32 %f545, %f535;
|
|
add.f32 %f546, %f544, %f544;
|
|
fma.rn.f32 %f547, %f545, %f533, %f546;
|
|
mul.rn.f32 %f548, %f521, %f547;
|
|
add.f32 %f549, %f543, %f535;
|
|
sub.f32 %f550, %f535, %f549;
|
|
add.f32 %f551, %f543, %f550;
|
|
add.f32 %f552, %f548, %f551;
|
|
add.f32 %f553, %f549, %f552;
|
|
sub.f32 %f554, %f549, %f553;
|
|
add.f32 %f555, %f552, %f554;
|
|
mul.rn.f32 %f557, %f532, %f894;
|
|
mul.rn.f32 %f559, %f532, %f895;
|
|
add.f32 %f560, %f557, %f553;
|
|
sub.f32 %f561, %f557, %f560;
|
|
add.f32 %f562, %f553, %f561;
|
|
add.f32 %f563, %f555, %f562;
|
|
add.f32 %f564, %f559, %f563;
|
|
add.f32 %f565, %f560, %f564;
|
|
sub.f32 %f566, %f560, %f565;
|
|
add.f32 %f567, %f564, %f566;
|
|
mul.rn.f32 %f569, %f890, %f565;
|
|
neg.f32 %f570, %f569;
|
|
fma.rn.f32 %f571, %f890, %f565, %f570;
|
|
fma.rn.f32 %f572, %f890, %f567, %f571;
|
|
fma.rn.f32 %f574, %f889, %f565, %f572;
|
|
add.rn.f32 %f575, %f569, %f574;
|
|
neg.f32 %f576, %f575;
|
|
add.rn.f32 %f577, %f569, %f576;
|
|
add.rn.f32 %f578, %f577, %f574;
|
|
mov.b32 %r259, %f575;
|
|
setp.eq.s32 %p80, %r259, 1118925336;
|
|
add.s32 %r260, %r259, -1;
|
|
mov.b32 %f579, %r260;
|
|
add.f32 %f580, %f578, 0f37000000;
|
|
selp.f32 %f581, %f579, %f575, %p80;
|
|
selp.f32 %f108, %f580, %f578, %p80;
|
|
mul.f32 %f582, %f581, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f583, %f582;
|
|
fma.rn.f32 %f585, %f583, %f896, %f581;
|
|
fma.rn.f32 %f587, %f583, %f897, %f585;
|
|
mul.f32 %f588, %f587, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f589, %f588;
|
|
add.f32 %f590, %f583, 0f00000000;
|
|
ex2.approx.f32 %f591, %f590;
|
|
mul.f32 %f592, %f589, %f591;
|
|
setp.lt.f32 %p81, %f581, 0fC2D20000;
|
|
selp.f32 %f593, 0f00000000, %f592, %p81;
|
|
setp.gt.f32 %p82, %f581, 0f42D20000;
|
|
selp.f32 %f953, 0f7F800000, %f593, %p82;
|
|
setp.eq.f32 %p83, %f953, 0f7F800000;
|
|
@%p83 bra BB0_84;
|
|
|
|
fma.rn.f32 %f953, %f953, %f108, %f953;
|
|
|
|
BB0_84:
|
|
setp.lt.f32 %p84, %f78, 0f00000000;
|
|
and.pred %p3, %p84, %p69;
|
|
mov.b32 %r261, %f953;
|
|
xor.b32 %r262, %r261, -2147483648;
|
|
mov.b32 %f594, %r262;
|
|
selp.f32 %f955, %f594, %f953, %p3;
|
|
setp.eq.f32 %p86, %f78, 0f00000000;
|
|
@%p86 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f597, %f78, %f78;
|
|
selp.f32 %f955, %f597, 0f00000000, %p69;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
setp.geu.f32 %p87, %f78, 0f00000000;
|
|
@%p87 bra BB0_88;
|
|
|
|
mov.f32 %f908, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f596, %f908;
|
|
setp.neu.f32 %p88, %f596, 0f3EE66666;
|
|
selp.f32 %f955, 0f7FFFFFFF, %f955, %p88;
|
|
|
|
BB0_88:
|
|
abs.f32 %f912, %f78;
|
|
add.f32 %f598, %f912, 0f3EE66666;
|
|
mov.b32 %r263, %f598;
|
|
setp.lt.s32 %p90, %r263, 2139095040;
|
|
@%p90 bra BB0_93;
|
|
|
|
abs.f32 %f913, %f78;
|
|
setp.gtu.f32 %p91, %f913, 0f7F800000;
|
|
@%p91 bra BB0_92;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_92:
|
|
add.f32 %f955, %f78, 0f3EE66666;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_90:
|
|
abs.f32 %f914, %f78;
|
|
setp.neu.f32 %p92, %f914, 0f7F800000;
|
|
@%p92 bra BB0_93;
|
|
|
|
selp.f32 %f955, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_93:
|
|
mov.f32 %f906, 0fB5BFBE8E;
|
|
mov.f32 %f905, 0fBF317200;
|
|
mov.f32 %f904, 0f35BFBE8E;
|
|
mov.f32 %f903, 0f3F317200;
|
|
mov.f32 %f902, 0f3DAAAABD;
|
|
mov.f32 %f901, 0f3C4CAF63;
|
|
mov.f32 %f900, 0f3B18F0FE;
|
|
mov.f32 %f899, 0f3EE66666;
|
|
mov.f32 %f898, 0f00000000;
|
|
setp.eq.f32 %p93, %f78, 0f3F800000;
|
|
selp.f32 %f119, 0f3F800000, %f955, %p93;
|
|
abs.f32 %f120, %f79;
|
|
setp.lt.f32 %p94, %f120, 0f00800000;
|
|
mul.f32 %f601, %f120, 0f4B800000;
|
|
selp.f32 %f602, 0fC3170000, 0fC2FE0000, %p94;
|
|
selp.f32 %f603, %f601, %f120, %p94;
|
|
mov.b32 %r264, %f603;
|
|
and.b32 %r265, %r264, 8388607;
|
|
or.b32 %r266, %r265, 1065353216;
|
|
mov.b32 %f604, %r266;
|
|
shr.u32 %r267, %r264, 23;
|
|
cvt.rn.f32.u32 %f605, %r267;
|
|
add.f32 %f606, %f602, %f605;
|
|
setp.gt.f32 %p95, %f604, 0f3FB504F3;
|
|
mul.f32 %f607, %f604, 0f3F000000;
|
|
add.f32 %f608, %f606, 0f3F800000;
|
|
selp.f32 %f609, %f607, %f604, %p95;
|
|
selp.f32 %f610, %f608, %f606, %p95;
|
|
add.f32 %f611, %f609, 0fBF800000;
|
|
add.f32 %f600, %f609, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f599,%f600;
|
|
// inline asm
|
|
add.f32 %f612, %f611, %f611;
|
|
mul.f32 %f613, %f599, %f612;
|
|
mul.f32 %f614, %f613, %f613;
|
|
fma.rn.f32 %f617, %f900, %f614, %f901;
|
|
fma.rn.f32 %f619, %f617, %f614, %f902;
|
|
mul.rn.f32 %f620, %f619, %f614;
|
|
mul.rn.f32 %f621, %f620, %f613;
|
|
sub.f32 %f622, %f611, %f613;
|
|
neg.f32 %f623, %f613;
|
|
add.f32 %f624, %f622, %f622;
|
|
fma.rn.f32 %f625, %f623, %f611, %f624;
|
|
mul.rn.f32 %f626, %f599, %f625;
|
|
add.f32 %f627, %f621, %f613;
|
|
sub.f32 %f628, %f613, %f627;
|
|
add.f32 %f629, %f621, %f628;
|
|
add.f32 %f630, %f626, %f629;
|
|
add.f32 %f631, %f627, %f630;
|
|
sub.f32 %f632, %f627, %f631;
|
|
add.f32 %f633, %f630, %f632;
|
|
mul.rn.f32 %f635, %f610, %f903;
|
|
mul.rn.f32 %f637, %f610, %f904;
|
|
add.f32 %f638, %f635, %f631;
|
|
sub.f32 %f639, %f635, %f638;
|
|
add.f32 %f640, %f631, %f639;
|
|
add.f32 %f641, %f633, %f640;
|
|
add.f32 %f642, %f637, %f641;
|
|
add.f32 %f643, %f638, %f642;
|
|
sub.f32 %f644, %f638, %f643;
|
|
add.f32 %f645, %f642, %f644;
|
|
mul.rn.f32 %f647, %f899, %f643;
|
|
neg.f32 %f648, %f647;
|
|
fma.rn.f32 %f649, %f899, %f643, %f648;
|
|
fma.rn.f32 %f650, %f899, %f645, %f649;
|
|
fma.rn.f32 %f652, %f898, %f643, %f650;
|
|
add.rn.f32 %f653, %f647, %f652;
|
|
neg.f32 %f654, %f653;
|
|
add.rn.f32 %f655, %f647, %f654;
|
|
add.rn.f32 %f656, %f655, %f652;
|
|
mov.b32 %r268, %f653;
|
|
setp.eq.s32 %p96, %r268, 1118925336;
|
|
add.s32 %r269, %r268, -1;
|
|
mov.b32 %f657, %r269;
|
|
add.f32 %f658, %f656, 0f37000000;
|
|
selp.f32 %f659, %f657, %f653, %p96;
|
|
selp.f32 %f121, %f658, %f656, %p96;
|
|
mul.f32 %f660, %f659, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f661, %f660;
|
|
fma.rn.f32 %f663, %f661, %f905, %f659;
|
|
fma.rn.f32 %f665, %f661, %f906, %f663;
|
|
mul.f32 %f666, %f665, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f667, %f666;
|
|
add.f32 %f668, %f661, 0f00000000;
|
|
ex2.approx.f32 %f669, %f668;
|
|
mul.f32 %f670, %f667, %f669;
|
|
setp.lt.f32 %p97, %f659, 0fC2D20000;
|
|
selp.f32 %f671, 0f00000000, %f670, %p97;
|
|
setp.gt.f32 %p98, %f659, 0f42D20000;
|
|
selp.f32 %f956, 0f7F800000, %f671, %p98;
|
|
setp.eq.f32 %p99, %f956, 0f7F800000;
|
|
@%p99 bra BB0_95;
|
|
|
|
fma.rn.f32 %f956, %f956, %f121, %f956;
|
|
|
|
BB0_95:
|
|
setp.lt.f32 %p100, %f79, 0f00000000;
|
|
and.pred %p4, %p100, %p69;
|
|
mov.b32 %r270, %f956;
|
|
xor.b32 %r271, %r270, -2147483648;
|
|
mov.b32 %f672, %r271;
|
|
selp.f32 %f958, %f672, %f956, %p4;
|
|
setp.eq.f32 %p102, %f79, 0f00000000;
|
|
@%p102 bra BB0_98;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_98:
|
|
add.f32 %f675, %f79, %f79;
|
|
selp.f32 %f958, %f675, 0f00000000, %p69;
|
|
bra.uni BB0_99;
|
|
|
|
BB0_96:
|
|
setp.geu.f32 %p103, %f79, 0f00000000;
|
|
@%p103 bra BB0_99;
|
|
|
|
mov.f32 %f907, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f674, %f907;
|
|
setp.neu.f32 %p104, %f674, 0f3EE66666;
|
|
selp.f32 %f958, 0f7FFFFFFF, %f958, %p104;
|
|
|
|
BB0_99:
|
|
abs.f32 %f924, %f79;
|
|
add.f32 %f676, %f924, 0f3EE66666;
|
|
mov.b32 %r272, %f676;
|
|
setp.lt.s32 %p106, %r272, 2139095040;
|
|
@%p106 bra BB0_104;
|
|
|
|
abs.f32 %f925, %f79;
|
|
setp.gtu.f32 %p107, %f925, 0f7F800000;
|
|
@%p107 bra BB0_103;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_103:
|
|
add.f32 %f958, %f79, 0f3EE66666;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_101:
|
|
abs.f32 %f926, %f79;
|
|
setp.neu.f32 %p108, %f926, 0f7F800000;
|
|
@%p108 bra BB0_104;
|
|
|
|
selp.f32 %f958, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_104:
|
|
mov.u32 %r427, 4;
|
|
setp.eq.f32 %p109, %f79, 0f3F800000;
|
|
selp.f32 %f677, 0f3F800000, %f958, %p109;
|
|
cvt.u64.u32 %rd63, %r3;
|
|
cvt.u64.u32 %rd62, %r2;
|
|
mov.u64 %rd66, image;
|
|
cvta.global.u64 %rd61, %rd66;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd61, %r99, %r427, %rd62, %rd63, %rd25, %rd25);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f678, %f677;
|
|
mul.f32 %f679, %f678, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r275, %f679;
|
|
cvt.sat.f32.f32 %f680, %f119;
|
|
mul.f32 %f681, %f680, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r276, %f681;
|
|
cvt.sat.f32.f32 %f682, %f106;
|
|
mul.f32 %f683, %f682, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r277, %f683;
|
|
cvt.u16.u32 %rs21, %r275;
|
|
cvt.u16.u32 %rs22, %r277;
|
|
cvt.u16.u32 %rs23, %r276;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd60], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_105:
|
|
cvt.u64.u32 %rd16, %r2;
|
|
cvt.u64.u32 %rd17, %r3;
|
|
and.b32 %r278, %r457, 4;
|
|
setp.eq.s32 %p110, %r278, 0;
|
|
@%p110 bra BB0_109;
|
|
|
|
ld.global.u32 %r279, [additive];
|
|
setp.eq.s32 %p111, %r279, 0;
|
|
mov.f32 %f684, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f684;}
|
|
|
|
// inline asm
|
|
@%p111 bra BB0_108;
|
|
|
|
mov.u64 %rd79, image_HDR;
|
|
cvta.global.u64 %rd68, %rd79;
|
|
mov.u32 %r283, 8;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd68, %r99, %r283, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd67];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f685, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f686, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f687, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd73), _rt_buffer_get_64, (%rd68, %r99, %r283, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f688, %f77, %f685;
|
|
add.f32 %f689, %f78, %f686;
|
|
add.f32 %f690, %f79, %f687;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f690;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f689;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f688;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd73], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_109;
|
|
|
|
BB0_108:
|
|
mov.u64 %rd86, image_HDR;
|
|
cvta.global.u64 %rd81, %rd86;
|
|
mov.u32 %r285, 8;
|
|
// inline asm
|
|
call (%rd80), _rt_buffer_get_64, (%rd81, %r99, %r285, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f79;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f78;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f77;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd80], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_109:
|
|
mul.f32 %f132, %f74, 0f403D3C19;
|
|
mul.f32 %f133, %f132, 0f3E906EBB;
|
|
mul.f32 %f134, %f75, 0f403D3C19;
|
|
mul.f32 %f135, %f134, 0f3E906EBB;
|
|
mul.f32 %f136, %f76, 0f403D3C19;
|
|
mul.f32 %f137, %f136, 0f3E906EBB;
|
|
ld.global.u32 %r286, [additive];
|
|
setp.eq.s32 %p112, %r286, 0;
|
|
mov.f32 %f694, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f694;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_111;
|
|
|
|
mov.u64 %rd99, image_RNM0;
|
|
cvta.global.u64 %rd88, %rd99;
|
|
mov.u32 %r290, 8;
|
|
// inline asm
|
|
call (%rd87), _rt_buffer_get_64, (%rd88, %r99, %r290, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd87];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f695, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f696, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f697, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd93), _rt_buffer_get_64, (%rd88, %r99, %r290, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f698, %f133, %f695;
|
|
add.f32 %f699, %f135, %f696;
|
|
add.f32 %f700, %f137, %f697;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f700;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f699;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f698;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd93], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_112;
|
|
|
|
BB0_111:
|
|
mov.u64 %rd106, image_RNM0;
|
|
cvta.global.u64 %rd101, %rd106;
|
|
mov.u32 %r292, 8;
|
|
// inline asm
|
|
call (%rd100), _rt_buffer_get_64, (%rd101, %r99, %r292, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f137;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f135;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f133;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd100], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_112:
|
|
ld.global.f32 %f705, [directDir+4];
|
|
mul.f32 %f706, %f705, 0fBEFA2A1C;
|
|
mul.f32 %f138, %f132, %f706;
|
|
mul.f32 %f139, %f134, %f706;
|
|
mul.f32 %f140, %f136, %f706;
|
|
ld.global.u32 %r293, [additive];
|
|
setp.eq.s32 %p113, %r293, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f694;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_114;
|
|
|
|
mov.u64 %rd119, image_RNM1;
|
|
cvta.global.u64 %rd108, %rd119;
|
|
mov.u32 %r297, 8;
|
|
// inline asm
|
|
call (%rd107), _rt_buffer_get_64, (%rd108, %r99, %r297, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd107];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f707, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f708, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f709, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd113), _rt_buffer_get_64, (%rd108, %r99, %r297, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f710, %f138, %f707;
|
|
add.f32 %f711, %f139, %f708;
|
|
add.f32 %f712, %f140, %f709;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f712;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f711;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f710;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd113], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_115;
|
|
|
|
BB0_114:
|
|
mov.u64 %rd126, image_RNM1;
|
|
cvta.global.u64 %rd121, %rd126;
|
|
mov.u32 %r299, 8;
|
|
// inline asm
|
|
call (%rd120), _rt_buffer_get_64, (%rd121, %r99, %r299, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f138;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd120], {%rs64, %rs65, %rs66, %rs53};
|
|
|
|
BB0_115:
|
|
ld.global.f32 %f717, [directDir+8];
|
|
mul.f32 %f718, %f717, 0fBEFA2A1C;
|
|
mul.f32 %f141, %f132, %f718;
|
|
mul.f32 %f142, %f134, %f718;
|
|
mul.f32 %f143, %f136, %f718;
|
|
ld.global.u32 %r300, [additive];
|
|
setp.eq.s32 %p114, %r300, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f694;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_117;
|
|
|
|
mov.u64 %rd139, image_RNM2;
|
|
cvta.global.u64 %rd128, %rd139;
|
|
mov.u32 %r304, 8;
|
|
// inline asm
|
|
call (%rd127), _rt_buffer_get_64, (%rd128, %r99, %r304, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd127];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f719, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f720, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f721, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd133), _rt_buffer_get_64, (%rd128, %r99, %r304, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f722, %f141, %f719;
|
|
add.f32 %f723, %f142, %f720;
|
|
add.f32 %f724, %f143, %f721;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f724;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f723;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f722;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd133], {%rs71, %rs72, %rs73, %rs67};
|
|
bra.uni BB0_118;
|
|
|
|
BB0_117:
|
|
mov.u64 %rd146, image_RNM2;
|
|
cvta.global.u64 %rd141, %rd146;
|
|
mov.u32 %r306, 8;
|
|
// inline asm
|
|
call (%rd140), _rt_buffer_get_64, (%rd141, %r99, %r306, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f143;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f142;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f141;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd140], {%rs78, %rs79, %rs80, %rs67};
|
|
|
|
BB0_118:
|
|
ld.global.f32 %f729, [directDir];
|
|
mul.f32 %f730, %f729, 0fBEFA2A1C;
|
|
mul.f32 %f144, %f132, %f730;
|
|
mul.f32 %f145, %f134, %f730;
|
|
mul.f32 %f146, %f136, %f730;
|
|
ld.global.u32 %r307, [additive];
|
|
setp.eq.s32 %p115, %r307, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f694;}
|
|
|
|
// inline asm
|
|
@%p115 bra BB0_120;
|
|
|
|
mov.u64 %rd159, image_RNM3;
|
|
cvta.global.u64 %rd148, %rd159;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd147), _rt_buffer_get_64, (%rd148, %r99, %r311, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd147];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f731, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f732, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f733, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd153), _rt_buffer_get_64, (%rd148, %r99, %r311, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f734, %f144, %f731;
|
|
add.f32 %f735, %f145, %f732;
|
|
add.f32 %f736, %f146, %f733;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f736;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f735;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f734;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd153], {%rs85, %rs86, %rs87, %rs81};
|
|
bra.uni BB0_121;
|
|
|
|
BB0_120:
|
|
mov.u64 %rd166, image_RNM3;
|
|
cvta.global.u64 %rd161, %rd166;
|
|
mov.u32 %r313, 8;
|
|
// inline asm
|
|
call (%rd160), _rt_buffer_get_64, (%rd161, %r99, %r313, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f144;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd160], {%rs92, %rs93, %rs94, %rs81};
|
|
|
|
BB0_121:
|
|
ld.global.f32 %f741, [directDir+4];
|
|
ld.global.f32 %f742, [directDir];
|
|
mul.f32 %f743, %f742, %f741;
|
|
mul.f32 %f744, %f743, 0f3F8BD8A1;
|
|
mul.f32 %f147, %f132, %f744;
|
|
mul.f32 %f148, %f134, %f744;
|
|
mul.f32 %f149, %f136, %f744;
|
|
ld.global.u32 %r314, [additive];
|
|
setp.eq.s32 %p116, %r314, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f694;}
|
|
|
|
// inline asm
|
|
@%p116 bra BB0_123;
|
|
|
|
mov.u64 %rd179, image_RNM4;
|
|
cvta.global.u64 %rd168, %rd179;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd167), _rt_buffer_get_64, (%rd168, %r99, %r318, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd167];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f745, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f746, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f747, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd173), _rt_buffer_get_64, (%rd168, %r99, %r318, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f748, %f147, %f745;
|
|
add.f32 %f749, %f148, %f746;
|
|
add.f32 %f750, %f149, %f747;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f750;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f749;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f748;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd173], {%rs99, %rs100, %rs101, %rs95};
|
|
bra.uni BB0_124;
|
|
|
|
BB0_123:
|
|
mov.u64 %rd186, image_RNM4;
|
|
cvta.global.u64 %rd181, %rd186;
|
|
mov.u32 %r320, 8;
|
|
// inline asm
|
|
call (%rd180), _rt_buffer_get_64, (%rd181, %r99, %r320, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f148;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f147;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd180], {%rs106, %rs107, %rs108, %rs95};
|
|
|
|
BB0_124:
|
|
ld.global.f32 %f755, [directDir+4];
|
|
ld.global.f32 %f756, [directDir+8];
|
|
mul.f32 %f757, %f755, %f756;
|
|
mul.f32 %f758, %f757, 0f3F8BD8A1;
|
|
mul.f32 %f150, %f132, %f758;
|
|
mul.f32 %f151, %f134, %f758;
|
|
mul.f32 %f152, %f136, %f758;
|
|
ld.global.u32 %r321, [additive];
|
|
setp.eq.s32 %p117, %r321, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f694;}
|
|
|
|
// inline asm
|
|
@%p117 bra BB0_126;
|
|
|
|
mov.u64 %rd199, image_RNM5;
|
|
cvta.global.u64 %rd188, %rd199;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd187), _rt_buffer_get_64, (%rd188, %r99, %r325, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd187];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f759, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f760, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f761, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd193), _rt_buffer_get_64, (%rd188, %r99, %r325, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f762, %f150, %f759;
|
|
add.f32 %f763, %f151, %f760;
|
|
add.f32 %f764, %f152, %f761;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f764;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f763;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f762;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd193], {%rs113, %rs114, %rs115, %rs109};
|
|
bra.uni BB0_127;
|
|
|
|
BB0_126:
|
|
mov.u64 %rd206, image_RNM5;
|
|
cvta.global.u64 %rd201, %rd206;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd200), _rt_buffer_get_64, (%rd201, %r99, %r327, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f150;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd200], {%rs120, %rs121, %rs122, %rs109};
|
|
|
|
BB0_127:
|
|
ld.global.f32 %f769, [directDir+8];
|
|
fma.rn.f32 %f770, %f769, %f769, 0fBEAAAAAB;
|
|
mul.f32 %f771, %f770, 0f3F723881;
|
|
mul.f32 %f153, %f132, %f771;
|
|
mul.f32 %f154, %f134, %f771;
|
|
mul.f32 %f155, %f136, %f771;
|
|
ld.global.u32 %r328, [additive];
|
|
setp.eq.s32 %p118, %r328, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f694;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_129;
|
|
|
|
mov.u64 %rd219, image_RNM6;
|
|
cvta.global.u64 %rd208, %rd219;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd207), _rt_buffer_get_64, (%rd208, %r99, %r332, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs130, %rs131, %rs132, %rs133}, [%rd207];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f772, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f773, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f774, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd213), _rt_buffer_get_64, (%rd208, %r99, %r332, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f775, %f153, %f772;
|
|
add.f32 %f776, %f154, %f773;
|
|
add.f32 %f777, %f155, %f774;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f777;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f776;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f775;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd213], {%rs127, %rs128, %rs129, %rs123};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd226, image_RNM6;
|
|
cvta.global.u64 %rd221, %rd226;
|
|
mov.u32 %r334, 8;
|
|
// inline asm
|
|
call (%rd220), _rt_buffer_get_64, (%rd221, %r99, %r334, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f153;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd220], {%rs134, %rs135, %rs136, %rs123};
|
|
|
|
BB0_130:
|
|
ld.global.f32 %f782, [directDir+8];
|
|
ld.global.f32 %f783, [directDir];
|
|
mul.f32 %f784, %f783, %f782;
|
|
mul.f32 %f785, %f784, 0f3F8BD8A1;
|
|
mul.f32 %f156, %f132, %f785;
|
|
mul.f32 %f157, %f134, %f785;
|
|
mul.f32 %f158, %f136, %f785;
|
|
ld.global.u32 %r335, [additive];
|
|
setp.eq.s32 %p119, %r335, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f694;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_132;
|
|
|
|
mov.u64 %rd239, image_RNM7;
|
|
cvta.global.u64 %rd228, %rd239;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd227), _rt_buffer_get_64, (%rd228, %r99, %r339, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs144, %rs145, %rs146, %rs147}, [%rd227];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f786, %rs144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f787, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f788, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd233), _rt_buffer_get_64, (%rd228, %r99, %r339, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f789, %f156, %f786;
|
|
add.f32 %f790, %f157, %f787;
|
|
add.f32 %f791, %f158, %f788;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f791;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f790;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f789;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd233], {%rs141, %rs142, %rs143, %rs137};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd246, image_RNM7;
|
|
cvta.global.u64 %rd241, %rd246;
|
|
mov.u32 %r341, 8;
|
|
// inline asm
|
|
call (%rd240), _rt_buffer_get_64, (%rd241, %r99, %r341, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f157;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f156;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd240], {%rs148, %rs149, %rs150, %rs137};
|
|
|
|
BB0_133:
|
|
ld.global.f32 %f796, [directDir];
|
|
mul.f32 %f797, %f796, %f796;
|
|
ld.global.f32 %f798, [directDir+4];
|
|
mul.f32 %f799, %f798, %f798;
|
|
sub.f32 %f800, %f797, %f799;
|
|
mul.f32 %f801, %f800, 0f3F0BD8A1;
|
|
mul.f32 %f159, %f132, %f801;
|
|
mul.f32 %f160, %f134, %f801;
|
|
mul.f32 %f161, %f136, %f801;
|
|
ld.global.u32 %r342, [additive];
|
|
setp.eq.s32 %p120, %r342, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f694;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_135;
|
|
|
|
mov.u64 %rd259, image_RNM8;
|
|
cvta.global.u64 %rd248, %rd259;
|
|
mov.u32 %r346, 8;
|
|
// inline asm
|
|
call (%rd247), _rt_buffer_get_64, (%rd248, %r99, %r346, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
ld.v4.u16 {%rs158, %rs159, %rs160, %rs161}, [%rd247];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f802, %rs158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f803, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f804, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd253), _rt_buffer_get_64, (%rd248, %r99, %r346, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
add.f32 %f805, %f159, %f802;
|
|
add.f32 %f806, %f160, %f803;
|
|
add.f32 %f807, %f161, %f804;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f807;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f806;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f805;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd253], {%rs155, %rs156, %rs157, %rs151};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd266, image_RNM8;
|
|
cvta.global.u64 %rd261, %rd266;
|
|
mov.u32 %r348, 8;
|
|
// inline asm
|
|
call (%rd260), _rt_buffer_get_64, (%rd261, %r99, %r348, %rd16, %rd17, %rd25, %rd25);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f161;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f159;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd260], {%rs162, %rs163, %rs164, %rs151};
|
|
|
|
BB0_171:
|
|
ret;
|
|
}
|
|
|
|
|