summaryrefslogtreecommitdiff
path: root/VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx
diff options
context:
space:
mode:
Diffstat (limited to 'VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx')
-rw-r--r--VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx652
1 files changed, 652 insertions, 0 deletions
diff --git a/VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx
new file mode 100644
index 00000000..b1b2008d
--- /dev/null
+++ b/VRCSDK3Worlds/Assets/Editor/x64/Bakery/mulHalfByte.ptx
@@ -0,0 +1,652 @@
+//
+// 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 image2[1];
+.global .align 1 .b8 image3[1];
+.global .align 4 .u32 composeMode;
+.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_typeinfo11composeModeE[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_typename11composeModeE[4] = {105, 110, 116, 0};
+.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
+.global .align 4 .u32 _ZN21rti_internal_typeenum11composeModeE = 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_semantic11composeModeE[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_annotation11composeModeE[1];
+
+.visible .entry _Z6oxMainv(
+
+)
+{
+ .reg .pred %p<52>;
+ .reg .b16 %rs<12>;
+ .reg .f32 %f<334>;
+ .reg .b32 %r<79>;
+ .reg .b64 %rd<55>;
+
+
+ ld.global.v2.u32 {%r15, %r16}, [pixelID];
+ cvt.u64.u32 %rd3, %r15;
+ cvt.u64.u32 %rd4, %r16;
+ mov.u64 %rd43, image2;
+ cvta.global.u64 %rd2, %rd43;
+ mov.u32 %r13, 2;
+ mov.u32 %r8, 8;
+ mov.u64 %rd42, 0;
+ // inline asm
+ call (%rd1), _rt_buffer_get_64, (%rd2, %r13, %r8, %rd3, %rd4, %rd42, %rd42);
+ // inline asm
+ ld.u16 %rs4, [%rd1+6];
+ cvt.rn.f32.u16 %f1, %rs4;
+ ld.global.v2.u32 {%r19, %r20}, [pixelID];
+ cvt.u64.u32 %rd9, %r19;
+ cvt.u64.u32 %rd10, %r20;
+ // inline asm
+ call (%rd7), _rt_buffer_get_64, (%rd2, %r13, %r8, %rd9, %rd10, %rd42, %rd42);
+ // inline asm
+ ld.u16 %rs1, [%rd7];
+ // inline asm
+ { cvt.f32.f16 %f47, %rs1;}
+
+ // inline asm
+ ld.global.v2.u32 {%r23, %r24}, [pixelID];
+ cvt.u64.u32 %rd15, %r23;
+ cvt.u64.u32 %rd16, %r24;
+ // inline asm
+ call (%rd13), _rt_buffer_get_64, (%rd2, %r13, %r8, %rd15, %rd16, %rd42, %rd42);
+ // inline asm
+ ld.u16 %rs2, [%rd13+2];
+ // inline asm
+ { cvt.f32.f16 %f48, %rs2;}
+
+ // inline asm
+ ld.global.v2.u32 {%r27, %r28}, [pixelID];
+ cvt.u64.u32 %rd21, %r27;
+ cvt.u64.u32 %rd22, %r28;
+ // inline asm
+ call (%rd19), _rt_buffer_get_64, (%rd2, %r13, %r8, %rd21, %rd22, %rd42, %rd42);
+ // inline asm
+ ld.u16 %rs3, [%rd19+4];
+ // inline asm
+ { cvt.f32.f16 %f49, %rs3;}
+
+ // inline asm
+ ld.global.v2.u32 {%r31, %r32}, [pixelID];
+ cvt.u64.u32 %rd27, %r31;
+ cvt.u64.u32 %rd28, %r32;
+ mov.u64 %rd44, image3;
+ cvta.global.u64 %rd26, %rd44;
+ mov.u32 %r14, 4;
+ // inline asm
+ call (%rd25), _rt_buffer_get_64, (%rd26, %r13, %r14, %rd27, %rd28, %rd42, %rd42);
+ // inline asm
+ ld.u8 %rs5, [%rd25];
+ cvt.rn.f32.u16 %f52, %rs5;
+ div.rn.f32 %f5, %f52, 0f437F0000;
+ ld.global.v2.u32 {%r35, %r36}, [pixelID];
+ cvt.u64.u32 %rd33, %r35;
+ cvt.u64.u32 %rd34, %r36;
+ // inline asm
+ call (%rd31), _rt_buffer_get_64, (%rd26, %r13, %r14, %rd33, %rd34, %rd42, %rd42);
+ // inline asm
+ ld.u8 %rs6, [%rd31+1];
+ cvt.rn.f32.u16 %f53, %rs6;
+ div.rn.f32 %f6, %f53, 0f437F0000;
+ ld.global.v2.u32 {%r39, %r40}, [pixelID];
+ cvt.u64.u32 %rd39, %r39;
+ cvt.u64.u32 %rd40, %r40;
+ // inline asm
+ call (%rd37), _rt_buffer_get_64, (%rd26, %r13, %r14, %rd39, %rd40, %rd42, %rd42);
+ // inline asm
+ ld.u8 %rs7, [%rd37+2];
+ cvt.rn.f32.u16 %f54, %rs7;
+ div.rn.f32 %f7, %f54, 0f437F0000;
+ abs.f32 %f9, %f5;
+ setp.lt.f32 %p4, %f9, 0f00800000;
+ mul.f32 %f58, %f9, 0f4B800000;
+ selp.f32 %f59, 0fC3170000, 0fC2FE0000, %p4;
+ selp.f32 %f60, %f58, %f9, %p4;
+ mov.b32 %r43, %f60;
+ and.b32 %r44, %r43, 8388607;
+ or.b32 %r45, %r44, 1065353216;
+ mov.b32 %f61, %r45;
+ shr.u32 %r46, %r43, 23;
+ cvt.rn.f32.u32 %f62, %r46;
+ add.f32 %f63, %f59, %f62;
+ setp.gt.f32 %p5, %f61, 0f3FB504F3;
+ mul.f32 %f64, %f61, 0f3F000000;
+ add.f32 %f65, %f63, 0f3F800000;
+ selp.f32 %f66, %f64, %f61, %p5;
+ selp.f32 %f67, %f65, %f63, %p5;
+ add.f32 %f68, %f66, 0fBF800000;
+ add.f32 %f51, %f66, 0f3F800000;
+ // inline asm
+ rcp.approx.ftz.f32 %f50,%f51;
+ // inline asm
+ add.f32 %f69, %f68, %f68;
+ mul.f32 %f70, %f50, %f69;
+ mul.f32 %f71, %f70, %f70;
+ mov.f32 %f72, 0f3C4CAF63;
+ mov.f32 %f73, 0f3B18F0FE;
+ fma.rn.f32 %f74, %f73, %f71, %f72;
+ mov.f32 %f75, 0f3DAAAABD;
+ fma.rn.f32 %f76, %f74, %f71, %f75;
+ mul.rn.f32 %f77, %f76, %f71;
+ mul.rn.f32 %f78, %f77, %f70;
+ sub.f32 %f79, %f68, %f70;
+ neg.f32 %f80, %f70;
+ add.f32 %f81, %f79, %f79;
+ fma.rn.f32 %f82, %f80, %f68, %f81;
+ mul.rn.f32 %f83, %f50, %f82;
+ add.f32 %f84, %f78, %f70;
+ sub.f32 %f85, %f70, %f84;
+ add.f32 %f86, %f78, %f85;
+ add.f32 %f87, %f83, %f86;
+ add.f32 %f88, %f84, %f87;
+ sub.f32 %f89, %f84, %f88;
+ add.f32 %f90, %f87, %f89;
+ mov.f32 %f91, 0f3F317200;
+ mul.rn.f32 %f92, %f67, %f91;
+ mov.f32 %f93, 0f35BFBE8E;
+ mul.rn.f32 %f94, %f67, %f93;
+ add.f32 %f95, %f92, %f88;
+ sub.f32 %f96, %f92, %f95;
+ add.f32 %f97, %f88, %f96;
+ add.f32 %f98, %f90, %f97;
+ add.f32 %f99, %f94, %f98;
+ add.f32 %f100, %f95, %f99;
+ sub.f32 %f101, %f95, %f100;
+ add.f32 %f102, %f99, %f101;
+ mov.f32 %f103, 0f400CCCCD;
+ mul.rn.f32 %f104, %f103, %f100;
+ neg.f32 %f105, %f104;
+ fma.rn.f32 %f106, %f103, %f100, %f105;
+ fma.rn.f32 %f107, %f103, %f102, %f106;
+ mov.f32 %f108, 0f00000000;
+ fma.rn.f32 %f109, %f108, %f100, %f107;
+ add.rn.f32 %f110, %f104, %f109;
+ neg.f32 %f111, %f110;
+ add.rn.f32 %f112, %f104, %f111;
+ add.rn.f32 %f113, %f112, %f109;
+ mov.b32 %r47, %f110;
+ setp.eq.s32 %p6, %r47, 1118925336;
+ add.s32 %r48, %r47, -1;
+ mov.b32 %f114, %r48;
+ add.f32 %f115, %f113, 0f37000000;
+ selp.f32 %f116, %f114, %f110, %p6;
+ selp.f32 %f10, %f115, %f113, %p6;
+ mul.f32 %f117, %f116, 0f3FB8AA3B;
+ cvt.rzi.f32.f32 %f118, %f117;
+ mov.f32 %f119, 0fBF317200;
+ fma.rn.f32 %f120, %f118, %f119, %f116;
+ mov.f32 %f121, 0fB5BFBE8E;
+ fma.rn.f32 %f122, %f118, %f121, %f120;
+ mul.f32 %f123, %f122, 0f3FB8AA3B;
+ ex2.approx.ftz.f32 %f124, %f123;
+ add.f32 %f125, %f118, 0f00000000;
+ ex2.approx.f32 %f126, %f125;
+ mul.f32 %f127, %f124, %f126;
+ setp.lt.f32 %p7, %f116, 0fC2D20000;
+ selp.f32 %f128, 0f00000000, %f127, %p7;
+ setp.gt.f32 %p8, %f116, 0f42D20000;
+ selp.f32 %f325, 0f7F800000, %f128, %p8;
+ setp.eq.f32 %p9, %f325, 0f7F800000;
+ @%p9 bra BB0_2;
+
+ fma.rn.f32 %f325, %f325, %f10, %f325;
+
+BB0_2:
+ mov.f32 %f322, 0f3F8CCCCD;
+ cvt.rzi.f32.f32 %f321, %f322;
+ fma.rn.f32 %f320, %f321, 0fC0000000, 0f400CCCCD;
+ abs.f32 %f319, %f320;
+ setp.lt.f32 %p10, %f5, 0f00000000;
+ setp.eq.f32 %p11, %f319, 0f3F800000;
+ and.pred %p1, %p10, %p11;
+ mov.b32 %r49, %f325;
+ xor.b32 %r50, %r49, -2147483648;
+ mov.b32 %f129, %r50;
+ selp.f32 %f327, %f129, %f325, %p1;
+ setp.eq.f32 %p12, %f5, 0f00000000;
+ @%p12 bra BB0_5;
+ bra.uni BB0_3;
+
+BB0_5:
+ add.f32 %f132, %f5, %f5;
+ selp.f32 %f327, %f132, 0f00000000, %p11;
+ bra.uni BB0_6;
+
+BB0_3:
+ setp.geu.f32 %p13, %f5, 0f00000000;
+ @%p13 bra BB0_6;
+
+ mov.f32 %f324, 0f400CCCCD;
+ cvt.rzi.f32.f32 %f131, %f324;
+ setp.neu.f32 %p14, %f131, 0f400CCCCD;
+ selp.f32 %f327, 0f7FFFFFFF, %f327, %p14;
+
+BB0_6:
+ abs.f32 %f294, %f5;
+ add.f32 %f133, %f294, 0f400CCCCD;
+ mov.b32 %r51, %f133;
+ setp.lt.s32 %p16, %r51, 2139095040;
+ @%p16 bra BB0_11;
+
+ abs.f32 %f311, %f5;
+ setp.gtu.f32 %p17, %f311, 0f7F800000;
+ @%p17 bra BB0_10;
+ bra.uni BB0_8;
+
+BB0_10:
+ add.f32 %f327, %f5, 0f400CCCCD;
+ bra.uni BB0_11;
+
+BB0_8:
+ abs.f32 %f312, %f5;
+ setp.neu.f32 %p18, %f312, 0f7F800000;
+ @%p18 bra BB0_11;
+
+ selp.f32 %f327, 0fFF800000, 0f7F800000, %p1;
+
+BB0_11:
+ mov.f32 %f323, 0f400CCCCD;
+ mov.f32 %f302, 0fB5BFBE8E;
+ mov.f32 %f301, 0fBF317200;
+ mov.f32 %f300, 0f00000000;
+ mov.f32 %f299, 0f35BFBE8E;
+ mov.f32 %f298, 0f3F317200;
+ mov.f32 %f297, 0f3DAAAABD;
+ mov.f32 %f296, 0f3C4CAF63;
+ mov.f32 %f295, 0f3B18F0FE;
+ setp.eq.f32 %p19, %f5, 0f3F800000;
+ selp.f32 %f21, 0f3F800000, %f327, %p19;
+ abs.f32 %f22, %f6;
+ setp.lt.f32 %p20, %f22, 0f00800000;
+ mul.f32 %f136, %f22, 0f4B800000;
+ selp.f32 %f137, 0fC3170000, 0fC2FE0000, %p20;
+ selp.f32 %f138, %f136, %f22, %p20;
+ mov.b32 %r52, %f138;
+ and.b32 %r53, %r52, 8388607;
+ or.b32 %r54, %r53, 1065353216;
+ mov.b32 %f139, %r54;
+ shr.u32 %r55, %r52, 23;
+ cvt.rn.f32.u32 %f140, %r55;
+ add.f32 %f141, %f137, %f140;
+ setp.gt.f32 %p21, %f139, 0f3FB504F3;
+ mul.f32 %f142, %f139, 0f3F000000;
+ add.f32 %f143, %f141, 0f3F800000;
+ selp.f32 %f144, %f142, %f139, %p21;
+ selp.f32 %f145, %f143, %f141, %p21;
+ add.f32 %f146, %f144, 0fBF800000;
+ add.f32 %f135, %f144, 0f3F800000;
+ // inline asm
+ rcp.approx.ftz.f32 %f134,%f135;
+ // inline asm
+ add.f32 %f147, %f146, %f146;
+ mul.f32 %f148, %f134, %f147;
+ mul.f32 %f149, %f148, %f148;
+ fma.rn.f32 %f152, %f295, %f149, %f296;
+ fma.rn.f32 %f154, %f152, %f149, %f297;
+ mul.rn.f32 %f155, %f154, %f149;
+ mul.rn.f32 %f156, %f155, %f148;
+ sub.f32 %f157, %f146, %f148;
+ neg.f32 %f158, %f148;
+ add.f32 %f159, %f157, %f157;
+ fma.rn.f32 %f160, %f158, %f146, %f159;
+ mul.rn.f32 %f161, %f134, %f160;
+ add.f32 %f162, %f156, %f148;
+ sub.f32 %f163, %f148, %f162;
+ add.f32 %f164, %f156, %f163;
+ add.f32 %f165, %f161, %f164;
+ add.f32 %f166, %f162, %f165;
+ sub.f32 %f167, %f162, %f166;
+ add.f32 %f168, %f165, %f167;
+ mul.rn.f32 %f170, %f145, %f298;
+ mul.rn.f32 %f172, %f145, %f299;
+ add.f32 %f173, %f170, %f166;
+ sub.f32 %f174, %f170, %f173;
+ add.f32 %f175, %f166, %f174;
+ add.f32 %f176, %f168, %f175;
+ add.f32 %f177, %f172, %f176;
+ add.f32 %f178, %f173, %f177;
+ sub.f32 %f179, %f173, %f178;
+ add.f32 %f180, %f177, %f179;
+ mul.rn.f32 %f182, %f323, %f178;
+ neg.f32 %f183, %f182;
+ fma.rn.f32 %f184, %f323, %f178, %f183;
+ fma.rn.f32 %f185, %f323, %f180, %f184;
+ fma.rn.f32 %f187, %f300, %f178, %f185;
+ add.rn.f32 %f188, %f182, %f187;
+ neg.f32 %f189, %f188;
+ add.rn.f32 %f190, %f182, %f189;
+ add.rn.f32 %f191, %f190, %f187;
+ mov.b32 %r56, %f188;
+ setp.eq.s32 %p22, %r56, 1118925336;
+ add.s32 %r57, %r56, -1;
+ mov.b32 %f192, %r57;
+ add.f32 %f193, %f191, 0f37000000;
+ selp.f32 %f194, %f192, %f188, %p22;
+ selp.f32 %f23, %f193, %f191, %p22;
+ mul.f32 %f195, %f194, 0f3FB8AA3B;
+ cvt.rzi.f32.f32 %f196, %f195;
+ fma.rn.f32 %f198, %f196, %f301, %f194;
+ fma.rn.f32 %f200, %f196, %f302, %f198;
+ mul.f32 %f201, %f200, 0f3FB8AA3B;
+ ex2.approx.ftz.f32 %f202, %f201;
+ add.f32 %f203, %f196, 0f00000000;
+ ex2.approx.f32 %f204, %f203;
+ mul.f32 %f205, %f202, %f204;
+ setp.lt.f32 %p23, %f194, 0fC2D20000;
+ selp.f32 %f206, 0f00000000, %f205, %p23;
+ setp.gt.f32 %p24, %f194, 0f42D20000;
+ selp.f32 %f328, 0f7F800000, %f206, %p24;
+ setp.eq.f32 %p25, %f328, 0f7F800000;
+ @%p25 bra BB0_13;
+
+ fma.rn.f32 %f328, %f328, %f23, %f328;
+
+BB0_13:
+ setp.lt.f32 %p26, %f6, 0f00000000;
+ and.pred %p2, %p26, %p11;
+ mov.b32 %r58, %f328;
+ xor.b32 %r59, %r58, -2147483648;
+ mov.b32 %f207, %r59;
+ selp.f32 %f330, %f207, %f328, %p2;
+ setp.eq.f32 %p28, %f6, 0f00000000;
+ @%p28 bra BB0_16;
+ bra.uni BB0_14;
+
+BB0_16:
+ add.f32 %f210, %f6, %f6;
+ selp.f32 %f330, %f210, 0f00000000, %p11;
+ bra.uni BB0_17;
+
+BB0_14:
+ setp.geu.f32 %p29, %f6, 0f00000000;
+ @%p29 bra BB0_17;
+
+ mov.f32 %f318, 0f400CCCCD;
+ cvt.rzi.f32.f32 %f209, %f318;
+ setp.neu.f32 %p30, %f209, 0f400CCCCD;
+ selp.f32 %f330, 0f7FFFFFFF, %f330, %p30;
+
+BB0_17:
+ abs.f32 %f313, %f6;
+ add.f32 %f211, %f313, 0f400CCCCD;
+ mov.b32 %r60, %f211;
+ setp.lt.s32 %p32, %r60, 2139095040;
+ @%p32 bra BB0_22;
+
+ abs.f32 %f316, %f6;
+ setp.gtu.f32 %p33, %f316, 0f7F800000;
+ @%p33 bra BB0_21;
+ bra.uni BB0_19;
+
+BB0_21:
+ add.f32 %f330, %f6, 0f400CCCCD;
+ bra.uni BB0_22;
+
+BB0_19:
+ abs.f32 %f317, %f6;
+ setp.neu.f32 %p34, %f317, 0f7F800000;
+ @%p34 bra BB0_22;
+
+ selp.f32 %f330, 0fFF800000, 0f7F800000, %p2;
+
+BB0_22:
+ mov.f32 %f314, 0f400CCCCD;
+ mov.f32 %f310, 0fB5BFBE8E;
+ mov.f32 %f309, 0fBF317200;
+ mov.f32 %f308, 0f00000000;
+ mov.f32 %f307, 0f35BFBE8E;
+ mov.f32 %f306, 0f3F317200;
+ mov.f32 %f305, 0f3DAAAABD;
+ mov.f32 %f304, 0f3C4CAF63;
+ mov.f32 %f303, 0f3B18F0FE;
+ setp.eq.f32 %p35, %f6, 0f3F800000;
+ selp.f32 %f34, 0f3F800000, %f330, %p35;
+ abs.f32 %f35, %f7;
+ setp.lt.f32 %p36, %f35, 0f00800000;
+ mul.f32 %f214, %f35, 0f4B800000;
+ selp.f32 %f215, 0fC3170000, 0fC2FE0000, %p36;
+ selp.f32 %f216, %f214, %f35, %p36;
+ mov.b32 %r61, %f216;
+ and.b32 %r62, %r61, 8388607;
+ or.b32 %r63, %r62, 1065353216;
+ mov.b32 %f217, %r63;
+ shr.u32 %r64, %r61, 23;
+ cvt.rn.f32.u32 %f218, %r64;
+ add.f32 %f219, %f215, %f218;
+ setp.gt.f32 %p37, %f217, 0f3FB504F3;
+ mul.f32 %f220, %f217, 0f3F000000;
+ add.f32 %f221, %f219, 0f3F800000;
+ selp.f32 %f222, %f220, %f217, %p37;
+ selp.f32 %f223, %f221, %f219, %p37;
+ add.f32 %f224, %f222, 0fBF800000;
+ add.f32 %f213, %f222, 0f3F800000;
+ // inline asm
+ rcp.approx.ftz.f32 %f212,%f213;
+ // inline asm
+ add.f32 %f225, %f224, %f224;
+ mul.f32 %f226, %f212, %f225;
+ mul.f32 %f227, %f226, %f226;
+ fma.rn.f32 %f230, %f303, %f227, %f304;
+ fma.rn.f32 %f232, %f230, %f227, %f305;
+ mul.rn.f32 %f233, %f232, %f227;
+ mul.rn.f32 %f234, %f233, %f226;
+ sub.f32 %f235, %f224, %f226;
+ neg.f32 %f236, %f226;
+ add.f32 %f237, %f235, %f235;
+ fma.rn.f32 %f238, %f236, %f224, %f237;
+ mul.rn.f32 %f239, %f212, %f238;
+ add.f32 %f240, %f234, %f226;
+ sub.f32 %f241, %f226, %f240;
+ add.f32 %f242, %f234, %f241;
+ add.f32 %f243, %f239, %f242;
+ add.f32 %f244, %f240, %f243;
+ sub.f32 %f245, %f240, %f244;
+ add.f32 %f246, %f243, %f245;
+ mul.rn.f32 %f248, %f223, %f306;
+ mul.rn.f32 %f250, %f223, %f307;
+ add.f32 %f251, %f248, %f244;
+ sub.f32 %f252, %f248, %f251;
+ add.f32 %f253, %f244, %f252;
+ add.f32 %f254, %f246, %f253;
+ add.f32 %f255, %f250, %f254;
+ add.f32 %f256, %f251, %f255;
+ sub.f32 %f257, %f251, %f256;
+ add.f32 %f258, %f255, %f257;
+ mul.rn.f32 %f260, %f314, %f256;
+ neg.f32 %f261, %f260;
+ fma.rn.f32 %f262, %f314, %f256, %f261;
+ fma.rn.f32 %f263, %f314, %f258, %f262;
+ fma.rn.f32 %f265, %f308, %f256, %f263;
+ add.rn.f32 %f266, %f260, %f265;
+ neg.f32 %f267, %f266;
+ add.rn.f32 %f268, %f260, %f267;
+ add.rn.f32 %f269, %f268, %f265;
+ mov.b32 %r65, %f266;
+ setp.eq.s32 %p38, %r65, 1118925336;
+ add.s32 %r66, %r65, -1;
+ mov.b32 %f270, %r66;
+ add.f32 %f271, %f269, 0f37000000;
+ selp.f32 %f272, %f270, %f266, %p38;
+ selp.f32 %f36, %f271, %f269, %p38;
+ mul.f32 %f273, %f272, 0f3FB8AA3B;
+ cvt.rzi.f32.f32 %f274, %f273;
+ fma.rn.f32 %f276, %f274, %f309, %f272;
+ fma.rn.f32 %f278, %f274, %f310, %f276;
+ mul.f32 %f279, %f278, 0f3FB8AA3B;
+ ex2.approx.ftz.f32 %f280, %f279;
+ add.f32 %f281, %f274, 0f00000000;
+ ex2.approx.f32 %f282, %f281;
+ mul.f32 %f283, %f280, %f282;
+ setp.lt.f32 %p39, %f272, 0fC2D20000;
+ selp.f32 %f284, 0f00000000, %f283, %p39;
+ setp.gt.f32 %p40, %f272, 0f42D20000;
+ selp.f32 %f331, 0f7F800000, %f284, %p40;
+ setp.eq.f32 %p41, %f331, 0f7F800000;
+ @%p41 bra BB0_24;
+
+ fma.rn.f32 %f331, %f331, %f36, %f331;
+
+BB0_24:
+ setp.lt.f32 %p42, %f7, 0f00000000;
+ and.pred %p3, %p42, %p11;
+ mov.b32 %r67, %f331;
+ xor.b32 %r68, %r67, -2147483648;
+ mov.b32 %f285, %r68;
+ selp.f32 %f333, %f285, %f331, %p3;
+ setp.eq.f32 %p44, %f7, 0f00000000;
+ @%p44 bra BB0_27;
+ bra.uni BB0_25;
+
+BB0_27:
+ add.f32 %f288, %f7, %f7;
+ selp.f32 %f333, %f288, 0f00000000, %p11;
+ bra.uni BB0_28;
+
+BB0_25:
+ setp.geu.f32 %p45, %f7, 0f00000000;
+ @%p45 bra BB0_28;
+
+ mov.f32 %f315, 0f400CCCCD;
+ cvt.rzi.f32.f32 %f287, %f315;
+ setp.neu.f32 %p46, %f287, 0f400CCCCD;
+ selp.f32 %f333, 0f7FFFFFFF, %f333, %p46;
+
+BB0_28:
+ add.f32 %f289, %f35, 0f400CCCCD;
+ mov.b32 %r69, %f289;
+ setp.lt.s32 %p48, %r69, 2139095040;
+ @%p48 bra BB0_33;
+
+ setp.gtu.f32 %p49, %f35, 0f7F800000;
+ @%p49 bra BB0_32;
+ bra.uni BB0_30;
+
+BB0_32:
+ add.f32 %f333, %f7, 0f400CCCCD;
+ bra.uni BB0_33;
+
+BB0_30:
+ setp.neu.f32 %p50, %f35, 0f7F800000;
+ @%p50 bra BB0_33;
+
+ selp.f32 %f333, 0fFF800000, 0f7F800000, %p3;
+
+BB0_33:
+ mov.u64 %rd54, 0;
+ mov.u32 %r78, 8;
+ mov.u32 %r77, 2;
+ mov.u64 %rd53, image2;
+ cvta.global.u64 %rd52, %rd53;
+ setp.eq.f32 %p51, %f7, 0f3F800000;
+ selp.f32 %f293, 0f3F800000, %f333, %p51;
+ mul.f32 %f292, %f49, %f293;
+ ld.global.v2.u32 {%r72, %r73}, [pixelID];
+ cvt.u64.u32 %rd47, %r72;
+ cvt.u64.u32 %rd48, %r73;
+ // inline asm
+ call (%rd45), _rt_buffer_get_64, (%rd52, %r77, %r78, %rd47, %rd48, %rd54, %rd54);
+ // inline asm
+ mul.f32 %f290, %f47, %f21;
+ mul.f32 %f291, %f48, %f34;
+ cvt.rzi.u32.f32 %r76, %f1;
+ // inline asm
+ { cvt.rn.f16.f32 %rs10, %f292;}
+
+ // inline asm
+ // inline asm
+ { cvt.rn.f16.f32 %rs9, %f291;}
+
+ // inline asm
+ // inline asm
+ { cvt.rn.f16.f32 %rs8, %f290;}
+
+ // inline asm
+ cvt.u16.u32 %rs11, %r76;
+ st.v4.u16 [%rd45], {%rs8, %rs9, %rs10, %rs11};
+ ret;
+}
+
+