2921 lines
77 KiB
Plaintext
2921 lines
77 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[40];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<134>;
|
|
.reg .b16 %rs<298>;
|
|
.reg .f32 %f<992>;
|
|
.reg .b32 %r<460>;
|
|
.reg .b64 %rd<497>;
|
|
|
|
|
|
mov.u64 %rd496, __local_depot0;
|
|
cvta.local.u64 %SP, %rd496;
|
|
ld.global.v2.u32 {%r101, %r102}, [pixelID];
|
|
cvt.u64.u32 %rd23, %r101;
|
|
cvt.u64.u32 %rd24, %r102;
|
|
mov.u64 %rd27, uvnormal;
|
|
cvta.global.u64 %rd22, %rd27;
|
|
mov.u32 %r99, 2;
|
|
mov.u32 %r100, 4;
|
|
mov.u64 %rd26, 0;
|
|
// inline asm
|
|
call (%rd21), _rt_buffer_get_64, (%rd22, %r99, %r100, %rd23, %rd24, %rd26, %rd26);
|
|
// inline asm
|
|
ld.u32 %r1, [%rd21];
|
|
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 %f950, 0f00000000;
|
|
mov.f32 %f951, %f950;
|
|
mov.f32 %f952, %f950;
|
|
@%p5 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd21+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f179, %rs17;
|
|
div.rn.f32 %f180, %f179, 0f437F0000;
|
|
fma.rn.f32 %f181, %f180, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f182, %rs15;
|
|
div.rn.f32 %f183, %f182, 0f437F0000;
|
|
fma.rn.f32 %f184, %f183, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f185, %rs12;
|
|
div.rn.f32 %f186, %f185, 0f437F0000;
|
|
fma.rn.f32 %f187, %f186, 0f40000000, 0fBF800000;
|
|
mul.f32 %f188, %f184, %f184;
|
|
fma.rn.f32 %f189, %f181, %f181, %f188;
|
|
fma.rn.f32 %f190, %f187, %f187, %f189;
|
|
sqrt.rn.f32 %f191, %f190;
|
|
rcp.rn.f32 %f192, %f191;
|
|
mul.f32 %f950, %f181, %f192;
|
|
mul.f32 %f951, %f184, %f192;
|
|
mul.f32 %f952, %f187, %f192;
|
|
|
|
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, %f951, 0f00000000;
|
|
setp.eq.f32 %p7, %f950, 0f00000000;
|
|
and.pred %p8, %p7, %p6;
|
|
setp.eq.f32 %p9, %f952, 0f00000000;
|
|
and.pred %p10, %p8, %p9;
|
|
@%p10 bra BB0_137;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_137:
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
and.b32 %r349, %r459, 1;
|
|
setp.eq.b32 %p121, %r349, 1;
|
|
@!%p121 bra BB0_139;
|
|
bra.uni BB0_138;
|
|
|
|
BB0_138:
|
|
cvt.u64.u32 %rd280, %r2;
|
|
cvt.u64.u32 %rd281, %r3;
|
|
mov.u64 %rd284, image;
|
|
cvta.global.u64 %rd279, %rd284;
|
|
mov.u64 %rd283, 0;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r99, %r100, %rd280, %rd281, %rd283, %rd283);
|
|
// inline asm
|
|
mov.u16 %rs165, 0;
|
|
st.v4.u8 [%rd278], {%rs165, %rs165, %rs165, %rs165};
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
|
|
BB0_139:
|
|
and.b32 %r352, %r459, 8;
|
|
setp.eq.s32 %p122, %r352, 0;
|
|
@%p122 bra BB0_141;
|
|
|
|
cvt.u64.u32 %rd288, %r3;
|
|
cvt.u64.u32 %rd287, %r2;
|
|
mov.u64 %rd291, image_Mask;
|
|
cvta.global.u64 %rd286, %rd291;
|
|
mov.u64 %rd290, 0;
|
|
// inline asm
|
|
call (%rd285), _rt_buffer_get_64, (%rd286, %r99, %r99, %rd287, %rd288, %rd290, %rd290);
|
|
// inline asm
|
|
mov.f32 %f836, 0f00000000;
|
|
cvt.rzi.u32.f32 %r355, %f836;
|
|
cvt.u16.u32 %rs166, %r355;
|
|
mov.u16 %rs167, 0;
|
|
st.v2.u8 [%rd285], {%rs166, %rs167};
|
|
ld.global.u32 %r459, [imageEnabled];
|
|
|
|
BB0_141:
|
|
cvt.u64.u32 %rd19, %r2;
|
|
cvt.u64.u32 %rd20, %r3;
|
|
and.b32 %r356, %r459, 4;
|
|
setp.eq.s32 %p123, %r356, 0;
|
|
@%p123 bra BB0_145;
|
|
|
|
ld.global.u32 %r357, [additive];
|
|
setp.eq.s32 %p124, %r357, 0;
|
|
@%p124 bra BB0_144;
|
|
|
|
mov.u64 %rd304, image_HDR;
|
|
cvta.global.u64 %rd293, %rd304;
|
|
mov.u32 %r361, 8;
|
|
mov.u64 %rd303, 0;
|
|
// inline asm
|
|
call (%rd292), _rt_buffer_get_64, (%rd293, %r99, %r361, %rd19, %rd20, %rd303, %rd303);
|
|
// inline asm
|
|
ld.v4.u16 {%rs174, %rs175, %rs176, %rs177}, [%rd292];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f837, %rs174;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f838, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f839, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd293, %r99, %r361, %rd19, %rd20, %rd303, %rd303);
|
|
// inline asm
|
|
add.f32 %f840, %f837, 0f00000000;
|
|
add.f32 %f841, %f838, 0f00000000;
|
|
add.f32 %f842, %f839, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f842;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f841;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs171, %f840;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs178, 0;
|
|
st.v4.u16 [%rd298], {%rs171, %rs172, %rs173, %rs178};
|
|
bra.uni BB0_145;
|
|
|
|
BB0_3:
|
|
ld.global.f32 %f9, [directDir+8];
|
|
ld.global.f32 %f8, [directDir+4];
|
|
ld.global.f32 %f7, [directDir];
|
|
mul.f32 %f196, %f950, %f7;
|
|
mul.f32 %f197, %f951, %f8;
|
|
neg.f32 %f198, %f197;
|
|
sub.f32 %f199, %f198, %f196;
|
|
mul.f32 %f200, %f952, %f9;
|
|
sub.f32 %f10, %f199, %f200;
|
|
ld.global.v2.u32 {%r118, %r119}, [pixelID];
|
|
cvt.u64.u32 %rd30, %r118;
|
|
cvt.u64.u32 %rd31, %r119;
|
|
mov.u64 %rd40, uvpos;
|
|
cvta.global.u64 %rd29, %rd40;
|
|
mov.u32 %r115, 12;
|
|
// inline asm
|
|
call (%rd28), _rt_buffer_get_64, (%rd29, %r99, %r115, %rd30, %rd31, %rd26, %rd26);
|
|
// inline asm
|
|
ld.f32 %f13, [%rd28+8];
|
|
ld.f32 %f12, [%rd28+4];
|
|
ld.f32 %f11, [%rd28];
|
|
mul.f32 %f201, %f11, 0f3456BF95;
|
|
mul.f32 %f202, %f12, 0f3456BF95;
|
|
mul.f32 %f203, %f13, 0f3456BF95;
|
|
abs.f32 %f204, %f950;
|
|
div.rn.f32 %f205, %f201, %f204;
|
|
abs.f32 %f206, %f951;
|
|
div.rn.f32 %f207, %f202, %f206;
|
|
abs.f32 %f208, %f952;
|
|
div.rn.f32 %f209, %f203, %f208;
|
|
abs.f32 %f210, %f205;
|
|
abs.f32 %f211, %f207;
|
|
abs.f32 %f212, %f209;
|
|
mov.f32 %f213, 0f38D1B717;
|
|
max.f32 %f214, %f210, %f213;
|
|
max.f32 %f215, %f211, %f213;
|
|
max.f32 %f216, %f212, %f213;
|
|
fma.rn.f32 %f14, %f950, %f214, %f11;
|
|
fma.rn.f32 %f15, %f951, %f215, %f12;
|
|
fma.rn.f32 %f16, %f952, %f216, %f13;
|
|
abs.f32 %f217, %f7;
|
|
abs.f32 %f218, %f9;
|
|
setp.gt.f32 %p11, %f217, %f218;
|
|
neg.f32 %f219, %f8;
|
|
neg.f32 %f220, %f9;
|
|
selp.f32 %f221, %f219, 0f00000000, %p11;
|
|
selp.f32 %f222, %f7, %f220, %p11;
|
|
selp.f32 %f223, 0f00000000, %f8, %p11;
|
|
mul.f32 %f224, %f222, %f222;
|
|
fma.rn.f32 %f225, %f221, %f221, %f224;
|
|
fma.rn.f32 %f226, %f223, %f223, %f225;
|
|
sqrt.rn.f32 %f227, %f226;
|
|
rcp.rn.f32 %f228, %f227;
|
|
mul.f32 %f17, %f221, %f228;
|
|
mul.f32 %f18, %f222, %f228;
|
|
mul.f32 %f19, %f223, %f228;
|
|
mul.f32 %f229, %f9, %f18;
|
|
mul.f32 %f230, %f8, %f19;
|
|
sub.f32 %f20, %f229, %f230;
|
|
mul.f32 %f231, %f7, %f19;
|
|
mul.f32 %f232, %f9, %f17;
|
|
sub.f32 %f21, %f231, %f232;
|
|
mul.f32 %f233, %f8, %f17;
|
|
mul.f32 %f234, %f7, %f18;
|
|
sub.f32 %f22, %f233, %f234;
|
|
ld.global.v2.u32 {%r122, %r123}, [pixelID];
|
|
cvt.u64.u32 %rd36, %r122;
|
|
cvt.u64.u32 %rd37, %r123;
|
|
mov.u64 %rd41, rnd_seeds;
|
|
cvta.global.u64 %rd35, %rd41;
|
|
// inline asm
|
|
call (%rd34), _rt_buffer_get_64, (%rd35, %r99, %r100, %rd36, %rd37, %rd26, %rd26);
|
|
// inline asm
|
|
ld.global.u32 %r428, [samples];
|
|
mov.f32 %f971, 0f00000000;
|
|
setp.lt.s32 %p12, %r428, 1;
|
|
@%p12 bra BB0_4;
|
|
|
|
cvt.rn.f32.s32 %f238, %r428;
|
|
rcp.rn.f32 %f23, %f238;
|
|
ld.u32 %r454, [%rd34];
|
|
mul.f32 %f24, %f14, 0f3456BF95;
|
|
mul.f32 %f25, %f15, 0f3456BF95;
|
|
mul.f32 %f26, %f16, 0f3456BF95;
|
|
mov.f32 %f971, 0f00000000;
|
|
mov.u32 %r429, 0;
|
|
abs.f32 %f239, %f25;
|
|
abs.f32 %f240, %f24;
|
|
max.f32 %f241, %f240, %f239;
|
|
abs.f32 %f242, %f26;
|
|
max.f32 %f243, %f241, %f242;
|
|
mov.f32 %f972, %f971;
|
|
mov.f32 %f973, %f971;
|
|
|
|
BB0_6:
|
|
setp.lt.s32 %p13, %r428, 1;
|
|
@%p13 bra BB0_55;
|
|
|
|
cvt.rn.f32.s32 %f30, %r429;
|
|
max.f32 %f31, %f243, %f213;
|
|
mov.u32 %r431, 0;
|
|
|
|
BB0_8:
|
|
mad.lo.s32 %r128, %r454, 1664525, 1013904223;
|
|
and.b32 %r129, %r128, 16777215;
|
|
cvt.rn.f32.u32 %f245, %r129;
|
|
fma.rn.f32 %f246, %f245, 0f33800000, %f30;
|
|
mul.f32 %f247, %f23, %f246;
|
|
mad.lo.s32 %r454, %r128, 1664525, 1013904223;
|
|
and.b32 %r130, %r454, 16777215;
|
|
cvt.rn.f32.u32 %f248, %r130;
|
|
cvt.rn.f32.s32 %f249, %r431;
|
|
fma.rn.f32 %f250, %f248, 0f33800000, %f249;
|
|
mul.f32 %f251, %f23, %f250;
|
|
sqrt.rn.f32 %f35, %f247;
|
|
mul.f32 %f965, %f251, 0f40C90FDB;
|
|
abs.f32 %f37, %f965;
|
|
setp.neu.f32 %p14, %f37, 0f7F800000;
|
|
mov.f32 %f959, %f965;
|
|
@%p14 bra BB0_10;
|
|
|
|
mov.f32 %f252, 0f00000000;
|
|
mul.rn.f32 %f959, %f965, %f252;
|
|
|
|
BB0_10:
|
|
mul.f32 %f253, %f959, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r442, %f253;
|
|
cvt.rn.f32.s32 %f254, %r442;
|
|
neg.f32 %f255, %f254;
|
|
mov.f32 %f256, 0f3FC90FDA;
|
|
fma.rn.f32 %f257, %f255, %f256, %f959;
|
|
mov.f32 %f258, 0f33A22168;
|
|
fma.rn.f32 %f259, %f255, %f258, %f257;
|
|
mov.f32 %f260, 0f27C234C5;
|
|
fma.rn.f32 %f960, %f255, %f260, %f259;
|
|
abs.f32 %f261, %f959;
|
|
setp.leu.f32 %p15, %f261, 0f47CE4780;
|
|
@%p15 bra BB0_21;
|
|
|
|
add.u64 %rd43, %SP, 12;
|
|
cvta.to.local.u64 %rd3, %rd43;
|
|
mov.b32 %r13, %f959;
|
|
shr.u32 %r14, %r13, 23;
|
|
shl.b32 %r133, %r13, 8;
|
|
or.b32 %r15, %r133, -2147483648;
|
|
mov.u32 %r433, 0;
|
|
mov.u64 %rd493, 0;
|
|
mov.u64 %rd492, %rd3;
|
|
mov.u32 %r434, %r433;
|
|
|
|
BB0_12:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd44, %rd493, 2;
|
|
mov.u64 %rd45, __cudart_i2opi_f;
|
|
add.s64 %rd46, %rd45, %rd44;
|
|
ld.const.u32 %r136, [%rd46];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r134, %r136, %r15, %r434;
|
|
madc.hi.u32 %r434, %r136, %r15, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd492], %r134;
|
|
add.s32 %r433, %r433, 1;
|
|
cvt.s64.s32 %rd493, %r433;
|
|
mul.wide.s32 %rd49, %r433, 4;
|
|
add.s64 %rd492, %rd3, %rd49;
|
|
setp.ne.s32 %p16, %r433, 6;
|
|
@%p16 bra BB0_12;
|
|
|
|
and.b32 %r139, %r14, 255;
|
|
add.s32 %r140, %r139, -128;
|
|
shr.u32 %r141, %r140, 5;
|
|
and.b32 %r20, %r13, -2147483648;
|
|
cvta.to.local.u64 %rd51, %rd43;
|
|
st.local.u32 [%rd51+24], %r434;
|
|
mov.u32 %r142, 6;
|
|
sub.s32 %r143, %r142, %r141;
|
|
mul.wide.s32 %rd52, %r143, 4;
|
|
add.s64 %rd8, %rd51, %rd52;
|
|
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_15;
|
|
|
|
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_15:
|
|
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_16;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_16:
|
|
mov.u32 %r438, %r20;
|
|
mov.u32 %r439, %r29;
|
|
bra.uni BB0_18;
|
|
|
|
BB0_17:
|
|
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_18:
|
|
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_20;
|
|
|
|
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_20:
|
|
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 %f960, %r175;
|
|
|
|
BB0_21:
|
|
mul.rn.f32 %f43, %f960, %f960;
|
|
add.s32 %r46, %r442, 1;
|
|
and.b32 %r47, %r46, 1;
|
|
setp.eq.s32 %p23, %r47, 0;
|
|
@%p23 bra BB0_23;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_23:
|
|
mov.f32 %f264, 0f3C08839E;
|
|
mov.f32 %f265, 0fB94CA1F9;
|
|
fma.rn.f32 %f961, %f265, %f43, %f264;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_22:
|
|
mov.f32 %f262, 0fBAB6061A;
|
|
mov.f32 %f263, 0f37CCF5CE;
|
|
fma.rn.f32 %f961, %f263, %f43, %f262;
|
|
|
|
BB0_24:
|
|
@%p23 bra BB0_26;
|
|
bra.uni BB0_25;
|
|
|
|
BB0_26:
|
|
mov.f32 %f269, 0fBE2AAAA3;
|
|
fma.rn.f32 %f270, %f961, %f43, %f269;
|
|
mov.f32 %f271, 0f00000000;
|
|
fma.rn.f32 %f962, %f270, %f43, %f271;
|
|
bra.uni BB0_27;
|
|
|
|
BB0_25:
|
|
mov.f32 %f266, 0f3D2AAAA5;
|
|
fma.rn.f32 %f267, %f961, %f43, %f266;
|
|
mov.f32 %f268, 0fBF000000;
|
|
fma.rn.f32 %f962, %f267, %f43, %f268;
|
|
|
|
BB0_27:
|
|
fma.rn.f32 %f963, %f962, %f960, %f960;
|
|
@%p23 bra BB0_29;
|
|
|
|
mov.f32 %f272, 0f3F800000;
|
|
fma.rn.f32 %f963, %f962, %f43, %f272;
|
|
|
|
BB0_29:
|
|
and.b32 %r176, %r46, 2;
|
|
setp.eq.s32 %p26, %r176, 0;
|
|
@%p26 bra BB0_31;
|
|
|
|
mov.f32 %f273, 0f00000000;
|
|
mov.f32 %f274, 0fBF800000;
|
|
fma.rn.f32 %f963, %f963, %f274, %f273;
|
|
|
|
BB0_31:
|
|
@%p14 bra BB0_33;
|
|
|
|
mov.f32 %f275, 0f00000000;
|
|
mul.rn.f32 %f965, %f965, %f275;
|
|
|
|
BB0_33:
|
|
mul.f32 %f276, %f965, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r452, %f276;
|
|
cvt.rn.f32.s32 %f277, %r452;
|
|
neg.f32 %f278, %f277;
|
|
fma.rn.f32 %f280, %f278, %f256, %f965;
|
|
fma.rn.f32 %f282, %f278, %f258, %f280;
|
|
fma.rn.f32 %f966, %f278, %f260, %f282;
|
|
abs.f32 %f284, %f965;
|
|
setp.leu.f32 %p28, %f284, 0f47CE4780;
|
|
@%p28 bra BB0_44;
|
|
|
|
add.u64 %rd54, %SP, 12;
|
|
cvta.to.local.u64 %rd9, %rd54;
|
|
mov.b32 %r49, %f965;
|
|
shr.u32 %r50, %r49, 23;
|
|
shl.b32 %r179, %r49, 8;
|
|
or.b32 %r51, %r179, -2147483648;
|
|
mov.u32 %r443, 0;
|
|
mov.u64 %rd494, %rd9;
|
|
mov.u64 %rd495, %rd26;
|
|
mov.u32 %r444, %r443;
|
|
|
|
BB0_35:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd55, %rd495, 2;
|
|
mov.u64 %rd56, __cudart_i2opi_f;
|
|
add.s64 %rd57, %rd56, %rd55;
|
|
ld.const.u32 %r182, [%rd57];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r180, %r182, %r51, %r444;
|
|
madc.hi.u32 %r444, %r182, %r51, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd494], %r180;
|
|
add.s32 %r443, %r443, 1;
|
|
cvt.s64.s32 %rd495, %r443;
|
|
mul.wide.s32 %rd58, %r443, 4;
|
|
add.s64 %rd494, %rd9, %rd58;
|
|
setp.ne.s32 %p29, %r443, 6;
|
|
@%p29 bra BB0_35;
|
|
|
|
and.b32 %r185, %r50, 255;
|
|
add.s32 %r186, %r185, -128;
|
|
shr.u32 %r187, %r186, 5;
|
|
and.b32 %r56, %r49, -2147483648;
|
|
cvta.to.local.u64 %rd60, %rd54;
|
|
st.local.u32 [%rd60+24], %r444;
|
|
mov.u32 %r188, 6;
|
|
sub.s32 %r189, %r188, %r187;
|
|
mul.wide.s32 %rd61, %r189, 4;
|
|
add.s64 %rd15, %rd60, %rd61;
|
|
ld.local.u32 %r445, [%rd15];
|
|
ld.local.u32 %r446, [%rd15+-4];
|
|
and.b32 %r59, %r50, 31;
|
|
setp.eq.s32 %p30, %r59, 0;
|
|
@%p30 bra BB0_38;
|
|
|
|
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, [%rd15+-8];
|
|
shr.u32 %r195, %r194, %r191;
|
|
shl.b32 %r196, %r446, %r59;
|
|
add.s32 %r446, %r195, %r196;
|
|
|
|
BB0_38:
|
|
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_39;
|
|
bra.uni BB0_40;
|
|
|
|
BB0_39:
|
|
mov.u32 %r448, %r56;
|
|
mov.u32 %r449, %r65;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_40:
|
|
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_41:
|
|
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_43;
|
|
|
|
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_43:
|
|
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 %f966, %r221;
|
|
|
|
BB0_44:
|
|
mul.rn.f32 %f60, %f966, %f966;
|
|
and.b32 %r82, %r452, 1;
|
|
setp.eq.s32 %p36, %r82, 0;
|
|
@%p36 bra BB0_46;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_46:
|
|
mov.f32 %f287, 0f3C08839E;
|
|
mov.f32 %f288, 0fB94CA1F9;
|
|
fma.rn.f32 %f967, %f288, %f60, %f287;
|
|
bra.uni BB0_47;
|
|
|
|
BB0_45:
|
|
mov.f32 %f285, 0fBAB6061A;
|
|
mov.f32 %f286, 0f37CCF5CE;
|
|
fma.rn.f32 %f967, %f286, %f60, %f285;
|
|
|
|
BB0_47:
|
|
@%p36 bra BB0_49;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_49:
|
|
mov.f32 %f292, 0fBE2AAAA3;
|
|
fma.rn.f32 %f293, %f967, %f60, %f292;
|
|
mov.f32 %f294, 0f00000000;
|
|
fma.rn.f32 %f968, %f293, %f60, %f294;
|
|
bra.uni BB0_50;
|
|
|
|
BB0_48:
|
|
mov.f32 %f289, 0f3D2AAAA5;
|
|
fma.rn.f32 %f290, %f967, %f60, %f289;
|
|
mov.f32 %f291, 0fBF000000;
|
|
fma.rn.f32 %f968, %f290, %f60, %f291;
|
|
|
|
BB0_50:
|
|
fma.rn.f32 %f969, %f968, %f966, %f966;
|
|
@%p36 bra BB0_52;
|
|
|
|
mov.f32 %f295, 0f3F800000;
|
|
fma.rn.f32 %f969, %f968, %f60, %f295;
|
|
|
|
BB0_52:
|
|
and.b32 %r222, %r452, 2;
|
|
setp.eq.s32 %p39, %r222, 0;
|
|
@%p39 bra BB0_54;
|
|
|
|
mov.f32 %f296, 0f00000000;
|
|
mov.f32 %f297, 0fBF800000;
|
|
fma.rn.f32 %f969, %f969, %f297, %f296;
|
|
|
|
BB0_54:
|
|
mul.f32 %f306, %f35, %f963;
|
|
add.u64 %rd62, %SP, 0;
|
|
cvta.to.local.u64 %rd63, %rd62;
|
|
mul.f32 %f307, %f306, %f306;
|
|
mov.f32 %f308, 0f3F800000;
|
|
sub.f32 %f309, %f308, %f307;
|
|
mul.f32 %f310, %f35, %f969;
|
|
mul.f32 %f311, %f310, %f310;
|
|
sub.f32 %f312, %f309, %f311;
|
|
mov.f32 %f313, 0f00000000;
|
|
max.f32 %f314, %f313, %f312;
|
|
sqrt.rn.f32 %f315, %f314;
|
|
mul.f32 %f316, %f17, %f310;
|
|
mul.f32 %f317, %f18, %f310;
|
|
mul.f32 %f318, %f19, %f310;
|
|
fma.rn.f32 %f319, %f20, %f306, %f316;
|
|
fma.rn.f32 %f320, %f21, %f306, %f317;
|
|
fma.rn.f32 %f321, %f22, %f306, %f318;
|
|
fma.rn.f32 %f322, %f7, %f315, %f319;
|
|
fma.rn.f32 %f323, %f8, %f315, %f320;
|
|
fma.rn.f32 %f324, %f9, %f315, %f321;
|
|
add.f32 %f325, %f7, %f322;
|
|
add.f32 %f326, %f8, %f323;
|
|
add.f32 %f327, %f9, %f324;
|
|
ld.global.f32 %f328, [shadowSpread];
|
|
mul.f32 %f329, %f328, %f325;
|
|
mul.f32 %f330, %f328, %f326;
|
|
mul.f32 %f331, %f328, %f327;
|
|
sub.f32 %f332, %f329, %f7;
|
|
sub.f32 %f333, %f330, %f8;
|
|
sub.f32 %f334, %f331, %f9;
|
|
mul.f32 %f335, %f333, %f333;
|
|
fma.rn.f32 %f336, %f332, %f332, %f335;
|
|
fma.rn.f32 %f337, %f334, %f334, %f336;
|
|
sqrt.rn.f32 %f338, %f337;
|
|
rcp.rn.f32 %f339, %f338;
|
|
mul.f32 %f301, %f339, %f332;
|
|
mul.f32 %f302, %f339, %f333;
|
|
mul.f32 %f303, %f339, %f334;
|
|
ld.global.u32 %r226, [imageEnabled];
|
|
and.b32 %r227, %r226, 32;
|
|
setp.eq.s32 %p40, %r227, 0;
|
|
selp.f32 %f340, 0f3F800000, 0f41200000, %p40;
|
|
mul.f32 %f304, %f340, %f31;
|
|
mov.u32 %r228, 1065353216;
|
|
st.local.u32 [%rd63+8], %r228;
|
|
st.local.u32 [%rd63+4], %r228;
|
|
st.local.u32 [%rd63], %r228;
|
|
ld.global.u32 %r223, [root];
|
|
mov.u32 %r224, 1;
|
|
mov.f32 %f305, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r223, %f14, %f15, %f16, %f301, %f302, %f303, %r224, %f304, %f305, %rd62, %r115);
|
|
// inline asm
|
|
ld.local.f32 %f341, [%rd63];
|
|
add.f32 %f973, %f973, %f341;
|
|
ld.local.f32 %f342, [%rd63+4];
|
|
add.f32 %f972, %f972, %f342;
|
|
ld.local.f32 %f343, [%rd63+8];
|
|
add.f32 %f971, %f971, %f343;
|
|
ld.global.u32 %r428, [samples];
|
|
add.s32 %r431, %r431, 1;
|
|
setp.lt.s32 %p41, %r431, %r428;
|
|
@%p41 bra BB0_8;
|
|
|
|
BB0_55:
|
|
add.s32 %r429, %r429, 1;
|
|
setp.lt.s32 %p42, %r429, %r428;
|
|
@%p42 bra BB0_6;
|
|
bra.uni BB0_56;
|
|
|
|
BB0_4:
|
|
mov.f32 %f972, %f971;
|
|
mov.f32 %f973, %f971;
|
|
|
|
BB0_56:
|
|
setp.eq.s32 %p43, %r428, 0;
|
|
mov.f32 %f977, 0f3F800000;
|
|
mov.f32 %f978, %f977;
|
|
mov.f32 %f979, %f977;
|
|
@%p43 bra BB0_58;
|
|
|
|
mul.lo.s32 %r229, %r428, %r428;
|
|
cvt.rn.f32.s32 %f347, %r229;
|
|
rcp.rn.f32 %f348, %f347;
|
|
mul.f32 %f977, %f973, %f348;
|
|
mul.f32 %f978, %f972, %f348;
|
|
mul.f32 %f979, %f971, %f348;
|
|
|
|
BB0_58:
|
|
mul.f32 %f357, %f12, %f21;
|
|
fma.rn.f32 %f358, %f11, %f20, %f357;
|
|
fma.rn.f32 %f359, %f13, %f22, %f358;
|
|
ld.global.v4.f32 {%f360, %f361, %f362, %f363}, [lightTilingOffset];
|
|
fma.rn.f32 %f353, %f359, %f360, %f362;
|
|
mul.f32 %f366, %f12, %f18;
|
|
fma.rn.f32 %f367, %f11, %f17, %f366;
|
|
fma.rn.f32 %f368, %f13, %f19, %f367;
|
|
fma.rn.f32 %f354, %f368, %f361, %f363;
|
|
ld.global.u32 %r230, [lightCookie];
|
|
mov.f32 %f356, 0f00000000;
|
|
// inline asm
|
|
call (%f349, %f350, %f351, %f352), _rt_texture_get_f_id, (%r230, %r99, %f353, %f354, %f356, %f356);
|
|
// inline asm
|
|
min.f32 %f371, %f349, %f350;
|
|
min.f32 %f372, %f371, %f351;
|
|
mul.f32 %f87, %f977, %f372;
|
|
ld.global.f32 %f373, [directColor];
|
|
mul.f32 %f374, %f977, %f373;
|
|
ld.global.f32 %f375, [directColor+4];
|
|
mul.f32 %f376, %f978, %f375;
|
|
ld.global.f32 %f377, [directColor+8];
|
|
mul.f32 %f378, %f979, %f377;
|
|
mul.f32 %f88, %f349, %f374;
|
|
mul.f32 %f89, %f350, %f376;
|
|
mul.f32 %f90, %f351, %f378;
|
|
cvt.sat.f32.f32 %f379, %f10;
|
|
mul.f32 %f91, %f88, %f379;
|
|
mul.f32 %f92, %f89, %f379;
|
|
mul.f32 %f93, %f379, %f90;
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
and.b32 %r232, %r457, 8;
|
|
setp.eq.s32 %p44, %r232, 0;
|
|
@%p44 bra BB0_71;
|
|
|
|
cvt.u64.u32 %rd66, %r2;
|
|
cvt.u64.u32 %rd67, %r3;
|
|
mov.u64 %rd70, image_Mask;
|
|
cvta.global.u64 %rd65, %rd70;
|
|
// inline asm
|
|
call (%rd64), _rt_buffer_get_64, (%rd65, %r99, %r99, %rd66, %rd67, %rd26, %rd26);
|
|
// inline asm
|
|
abs.f32 %f95, %f87;
|
|
setp.lt.f32 %p45, %f95, 0f00800000;
|
|
mul.f32 %f385, %f95, 0f4B800000;
|
|
selp.f32 %f386, 0fC3170000, 0fC2FE0000, %p45;
|
|
selp.f32 %f387, %f385, %f95, %p45;
|
|
mov.b32 %r235, %f387;
|
|
and.b32 %r236, %r235, 8388607;
|
|
or.b32 %r237, %r236, 1065353216;
|
|
mov.b32 %f388, %r237;
|
|
shr.u32 %r238, %r235, 23;
|
|
cvt.rn.f32.u32 %f389, %r238;
|
|
add.f32 %f390, %f386, %f389;
|
|
setp.gt.f32 %p46, %f388, 0f3FB504F3;
|
|
mul.f32 %f391, %f388, 0f3F000000;
|
|
add.f32 %f392, %f390, 0f3F800000;
|
|
selp.f32 %f393, %f391, %f388, %p46;
|
|
selp.f32 %f394, %f392, %f390, %p46;
|
|
add.f32 %f395, %f393, 0fBF800000;
|
|
add.f32 %f381, %f393, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f380,%f381;
|
|
// inline asm
|
|
add.f32 %f396, %f395, %f395;
|
|
mul.f32 %f397, %f380, %f396;
|
|
mul.f32 %f398, %f397, %f397;
|
|
mov.f32 %f399, 0f3C4CAF63;
|
|
mov.f32 %f400, 0f3B18F0FE;
|
|
fma.rn.f32 %f401, %f400, %f398, %f399;
|
|
mov.f32 %f402, 0f3DAAAABD;
|
|
fma.rn.f32 %f403, %f401, %f398, %f402;
|
|
mul.rn.f32 %f404, %f403, %f398;
|
|
mul.rn.f32 %f405, %f404, %f397;
|
|
sub.f32 %f406, %f395, %f397;
|
|
neg.f32 %f407, %f397;
|
|
add.f32 %f408, %f406, %f406;
|
|
fma.rn.f32 %f409, %f407, %f395, %f408;
|
|
mul.rn.f32 %f410, %f380, %f409;
|
|
add.f32 %f411, %f405, %f397;
|
|
sub.f32 %f412, %f397, %f411;
|
|
add.f32 %f413, %f405, %f412;
|
|
add.f32 %f414, %f410, %f413;
|
|
add.f32 %f415, %f411, %f414;
|
|
sub.f32 %f416, %f411, %f415;
|
|
add.f32 %f417, %f414, %f416;
|
|
mov.f32 %f418, 0f3F317200;
|
|
mul.rn.f32 %f419, %f394, %f418;
|
|
mov.f32 %f420, 0f35BFBE8E;
|
|
mul.rn.f32 %f421, %f394, %f420;
|
|
add.f32 %f422, %f419, %f415;
|
|
sub.f32 %f423, %f419, %f422;
|
|
add.f32 %f424, %f415, %f423;
|
|
add.f32 %f425, %f417, %f424;
|
|
add.f32 %f426, %f421, %f425;
|
|
add.f32 %f427, %f422, %f426;
|
|
sub.f32 %f428, %f422, %f427;
|
|
add.f32 %f429, %f426, %f428;
|
|
mov.f32 %f430, 0f3EE8BA2E;
|
|
mul.rn.f32 %f431, %f430, %f427;
|
|
neg.f32 %f432, %f431;
|
|
fma.rn.f32 %f433, %f430, %f427, %f432;
|
|
fma.rn.f32 %f434, %f430, %f429, %f433;
|
|
fma.rn.f32 %f436, %f356, %f427, %f434;
|
|
add.rn.f32 %f437, %f431, %f436;
|
|
neg.f32 %f438, %f437;
|
|
add.rn.f32 %f439, %f431, %f438;
|
|
add.rn.f32 %f440, %f439, %f436;
|
|
mov.b32 %r239, %f437;
|
|
setp.eq.s32 %p47, %r239, 1118925336;
|
|
add.s32 %r240, %r239, -1;
|
|
mov.b32 %f441, %r240;
|
|
add.f32 %f442, %f440, 0f37000000;
|
|
selp.f32 %f443, %f441, %f437, %p47;
|
|
selp.f32 %f96, %f442, %f440, %p47;
|
|
mul.f32 %f444, %f443, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f445, %f444;
|
|
mov.f32 %f446, 0fBF317200;
|
|
fma.rn.f32 %f447, %f445, %f446, %f443;
|
|
mov.f32 %f448, 0fB5BFBE8E;
|
|
fma.rn.f32 %f449, %f445, %f448, %f447;
|
|
mul.f32 %f450, %f449, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f451, %f450;
|
|
add.f32 %f452, %f445, 0f00000000;
|
|
ex2.approx.f32 %f453, %f452;
|
|
mul.f32 %f454, %f451, %f453;
|
|
setp.lt.f32 %p48, %f443, 0fC2D20000;
|
|
selp.f32 %f455, 0f00000000, %f454, %p48;
|
|
setp.gt.f32 %p49, %f443, 0f42D20000;
|
|
selp.f32 %f980, 0f7F800000, %f455, %p49;
|
|
setp.eq.f32 %p50, %f980, 0f7F800000;
|
|
@%p50 bra BB0_61;
|
|
|
|
fma.rn.f32 %f980, %f980, %f96, %f980;
|
|
|
|
BB0_61:
|
|
mov.f32 %f941, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f940, %f941;
|
|
fma.rn.f32 %f939, %f940, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f938, %f939;
|
|
setp.lt.f32 %p51, %f87, 0f00000000;
|
|
setp.eq.f32 %p52, %f938, 0f3F800000;
|
|
and.pred %p1, %p51, %p52;
|
|
mov.b32 %r241, %f980;
|
|
xor.b32 %r242, %r241, -2147483648;
|
|
mov.b32 %f456, %r242;
|
|
selp.f32 %f982, %f456, %f980, %p1;
|
|
setp.eq.f32 %p53, %f87, 0f00000000;
|
|
@%p53 bra BB0_64;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_64:
|
|
add.f32 %f459, %f87, %f87;
|
|
selp.f32 %f982, %f459, 0f00000000, %p52;
|
|
bra.uni BB0_65;
|
|
|
|
BB0_144:
|
|
mov.u64 %rd311, image_HDR;
|
|
cvta.global.u64 %rd306, %rd311;
|
|
mov.u32 %r363, 8;
|
|
mov.u64 %rd310, 0;
|
|
// inline asm
|
|
call (%rd305), _rt_buffer_get_64, (%rd306, %r99, %r363, %rd19, %rd20, %rd310, %rd310);
|
|
// inline asm
|
|
mov.f32 %f843, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs179, %f843;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs180, 0;
|
|
st.v4.u16 [%rd305], {%rs179, %rs179, %rs179, %rs180};
|
|
|
|
BB0_145:
|
|
ld.global.u32 %r364, [additive];
|
|
setp.eq.s32 %p125, %r364, 0;
|
|
@%p125 bra BB0_147;
|
|
|
|
mov.u64 %rd324, image_RNM0;
|
|
cvta.global.u64 %rd313, %rd324;
|
|
mov.u32 %r368, 8;
|
|
mov.u64 %rd323, 0;
|
|
// inline asm
|
|
call (%rd312), _rt_buffer_get_64, (%rd313, %r99, %r368, %rd19, %rd20, %rd323, %rd323);
|
|
// inline asm
|
|
ld.v4.u16 {%rs187, %rs188, %rs189, %rs190}, [%rd312];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f844, %rs187;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f845, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f846, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd318), _rt_buffer_get_64, (%rd313, %r99, %r368, %rd19, %rd20, %rd323, %rd323);
|
|
// inline asm
|
|
add.f32 %f847, %f844, 0f00000000;
|
|
add.f32 %f848, %f845, 0f00000000;
|
|
add.f32 %f849, %f846, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f849;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f848;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs184, %f847;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs191, 0;
|
|
st.v4.u16 [%rd318], {%rs184, %rs185, %rs186, %rs191};
|
|
bra.uni BB0_148;
|
|
|
|
BB0_147:
|
|
mov.u64 %rd331, image_RNM0;
|
|
cvta.global.u64 %rd326, %rd331;
|
|
mov.u32 %r370, 8;
|
|
mov.u64 %rd330, 0;
|
|
// inline asm
|
|
call (%rd325), _rt_buffer_get_64, (%rd326, %r99, %r370, %rd19, %rd20, %rd330, %rd330);
|
|
// inline asm
|
|
mov.f32 %f850, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs192, %f850;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs193, 0;
|
|
st.v4.u16 [%rd325], {%rs192, %rs192, %rs192, %rs193};
|
|
|
|
BB0_148:
|
|
ld.global.u32 %r371, [additive];
|
|
setp.eq.s32 %p126, %r371, 0;
|
|
@%p126 bra BB0_150;
|
|
|
|
mov.u64 %rd344, image_RNM1;
|
|
cvta.global.u64 %rd333, %rd344;
|
|
mov.u32 %r375, 8;
|
|
mov.u64 %rd343, 0;
|
|
// inline asm
|
|
call (%rd332), _rt_buffer_get_64, (%rd333, %r99, %r375, %rd19, %rd20, %rd343, %rd343);
|
|
// inline asm
|
|
ld.v4.u16 {%rs200, %rs201, %rs202, %rs203}, [%rd332];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f851, %rs200;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f852, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f853, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd338), _rt_buffer_get_64, (%rd333, %r99, %r375, %rd19, %rd20, %rd343, %rd343);
|
|
// inline asm
|
|
add.f32 %f854, %f851, 0f00000000;
|
|
add.f32 %f855, %f852, 0f00000000;
|
|
add.f32 %f856, %f853, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f856;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f855;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs197, %f854;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs204, 0;
|
|
st.v4.u16 [%rd338], {%rs197, %rs198, %rs199, %rs204};
|
|
bra.uni BB0_151;
|
|
|
|
BB0_150:
|
|
mov.u64 %rd351, image_RNM1;
|
|
cvta.global.u64 %rd346, %rd351;
|
|
mov.u32 %r377, 8;
|
|
mov.u64 %rd350, 0;
|
|
// inline asm
|
|
call (%rd345), _rt_buffer_get_64, (%rd346, %r99, %r377, %rd19, %rd20, %rd350, %rd350);
|
|
// inline asm
|
|
mov.f32 %f857, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs205, %f857;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs206, 0;
|
|
st.v4.u16 [%rd345], {%rs205, %rs205, %rs205, %rs206};
|
|
|
|
BB0_151:
|
|
ld.global.u32 %r378, [additive];
|
|
setp.eq.s32 %p127, %r378, 0;
|
|
@%p127 bra BB0_153;
|
|
|
|
mov.u64 %rd364, image_RNM2;
|
|
cvta.global.u64 %rd353, %rd364;
|
|
mov.u32 %r382, 8;
|
|
mov.u64 %rd363, 0;
|
|
// inline asm
|
|
call (%rd352), _rt_buffer_get_64, (%rd353, %r99, %r382, %rd19, %rd20, %rd363, %rd363);
|
|
// inline asm
|
|
ld.v4.u16 {%rs213, %rs214, %rs215, %rs216}, [%rd352];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f858, %rs213;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f859, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f860, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd358), _rt_buffer_get_64, (%rd353, %r99, %r382, %rd19, %rd20, %rd363, %rd363);
|
|
// inline asm
|
|
add.f32 %f861, %f858, 0f00000000;
|
|
add.f32 %f862, %f859, 0f00000000;
|
|
add.f32 %f863, %f860, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f863;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f862;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs210, %f861;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs217, 0;
|
|
st.v4.u16 [%rd358], {%rs210, %rs211, %rs212, %rs217};
|
|
bra.uni BB0_154;
|
|
|
|
BB0_153:
|
|
mov.u64 %rd371, image_RNM2;
|
|
cvta.global.u64 %rd366, %rd371;
|
|
mov.u32 %r384, 8;
|
|
mov.u64 %rd370, 0;
|
|
// inline asm
|
|
call (%rd365), _rt_buffer_get_64, (%rd366, %r99, %r384, %rd19, %rd20, %rd370, %rd370);
|
|
// inline asm
|
|
mov.f32 %f864, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs218, %f864;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs219, 0;
|
|
st.v4.u16 [%rd365], {%rs218, %rs218, %rs218, %rs219};
|
|
|
|
BB0_154:
|
|
ld.global.u32 %r385, [additive];
|
|
setp.eq.s32 %p128, %r385, 0;
|
|
@%p128 bra BB0_156;
|
|
|
|
mov.u64 %rd384, image_RNM3;
|
|
cvta.global.u64 %rd373, %rd384;
|
|
mov.u32 %r389, 8;
|
|
mov.u64 %rd383, 0;
|
|
// inline asm
|
|
call (%rd372), _rt_buffer_get_64, (%rd373, %r99, %r389, %rd19, %rd20, %rd383, %rd383);
|
|
// inline asm
|
|
ld.v4.u16 {%rs226, %rs227, %rs228, %rs229}, [%rd372];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f865, %rs226;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f866, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f867, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd378), _rt_buffer_get_64, (%rd373, %r99, %r389, %rd19, %rd20, %rd383, %rd383);
|
|
// inline asm
|
|
add.f32 %f868, %f865, 0f00000000;
|
|
add.f32 %f869, %f866, 0f00000000;
|
|
add.f32 %f870, %f867, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f870;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f869;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs223, %f868;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs230, 0;
|
|
st.v4.u16 [%rd378], {%rs223, %rs224, %rs225, %rs230};
|
|
bra.uni BB0_157;
|
|
|
|
BB0_156:
|
|
mov.u64 %rd391, image_RNM3;
|
|
cvta.global.u64 %rd386, %rd391;
|
|
mov.u32 %r391, 8;
|
|
mov.u64 %rd390, 0;
|
|
// inline asm
|
|
call (%rd385), _rt_buffer_get_64, (%rd386, %r99, %r391, %rd19, %rd20, %rd390, %rd390);
|
|
// inline asm
|
|
mov.f32 %f871, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs231, %f871;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs232, 0;
|
|
st.v4.u16 [%rd385], {%rs231, %rs231, %rs231, %rs232};
|
|
|
|
BB0_157:
|
|
ld.global.u32 %r392, [additive];
|
|
setp.eq.s32 %p129, %r392, 0;
|
|
@%p129 bra BB0_159;
|
|
|
|
mov.u64 %rd404, image_RNM4;
|
|
cvta.global.u64 %rd393, %rd404;
|
|
mov.u32 %r396, 8;
|
|
mov.u64 %rd403, 0;
|
|
// inline asm
|
|
call (%rd392), _rt_buffer_get_64, (%rd393, %r99, %r396, %rd19, %rd20, %rd403, %rd403);
|
|
// inline asm
|
|
ld.v4.u16 {%rs239, %rs240, %rs241, %rs242}, [%rd392];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f872, %rs239;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f873, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f874, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd398), _rt_buffer_get_64, (%rd393, %r99, %r396, %rd19, %rd20, %rd403, %rd403);
|
|
// inline asm
|
|
add.f32 %f875, %f872, 0f00000000;
|
|
add.f32 %f876, %f873, 0f00000000;
|
|
add.f32 %f877, %f874, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f877;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f876;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs236, %f875;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs243, 0;
|
|
st.v4.u16 [%rd398], {%rs236, %rs237, %rs238, %rs243};
|
|
bra.uni BB0_160;
|
|
|
|
BB0_159:
|
|
mov.u64 %rd411, image_RNM4;
|
|
cvta.global.u64 %rd406, %rd411;
|
|
mov.u32 %r398, 8;
|
|
mov.u64 %rd410, 0;
|
|
// inline asm
|
|
call (%rd405), _rt_buffer_get_64, (%rd406, %r99, %r398, %rd19, %rd20, %rd410, %rd410);
|
|
// inline asm
|
|
mov.f32 %f878, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs244, %f878;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs245, 0;
|
|
st.v4.u16 [%rd405], {%rs244, %rs244, %rs244, %rs245};
|
|
|
|
BB0_160:
|
|
ld.global.u32 %r399, [additive];
|
|
setp.eq.s32 %p130, %r399, 0;
|
|
@%p130 bra BB0_162;
|
|
|
|
mov.u64 %rd424, image_RNM5;
|
|
cvta.global.u64 %rd413, %rd424;
|
|
mov.u32 %r403, 8;
|
|
mov.u64 %rd423, 0;
|
|
// inline asm
|
|
call (%rd412), _rt_buffer_get_64, (%rd413, %r99, %r403, %rd19, %rd20, %rd423, %rd423);
|
|
// inline asm
|
|
ld.v4.u16 {%rs252, %rs253, %rs254, %rs255}, [%rd412];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f879, %rs252;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f880, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f881, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd418), _rt_buffer_get_64, (%rd413, %r99, %r403, %rd19, %rd20, %rd423, %rd423);
|
|
// inline asm
|
|
add.f32 %f882, %f879, 0f00000000;
|
|
add.f32 %f883, %f880, 0f00000000;
|
|
add.f32 %f884, %f881, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f884;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f883;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs249, %f882;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs256, 0;
|
|
st.v4.u16 [%rd418], {%rs249, %rs250, %rs251, %rs256};
|
|
bra.uni BB0_163;
|
|
|
|
BB0_162:
|
|
mov.u64 %rd431, image_RNM5;
|
|
cvta.global.u64 %rd426, %rd431;
|
|
mov.u32 %r405, 8;
|
|
mov.u64 %rd430, 0;
|
|
// inline asm
|
|
call (%rd425), _rt_buffer_get_64, (%rd426, %r99, %r405, %rd19, %rd20, %rd430, %rd430);
|
|
// inline asm
|
|
mov.f32 %f885, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs257, %f885;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs258, 0;
|
|
st.v4.u16 [%rd425], {%rs257, %rs257, %rs257, %rs258};
|
|
|
|
BB0_163:
|
|
ld.global.u32 %r406, [additive];
|
|
setp.eq.s32 %p131, %r406, 0;
|
|
@%p131 bra BB0_165;
|
|
|
|
mov.u64 %rd444, image_RNM6;
|
|
cvta.global.u64 %rd433, %rd444;
|
|
mov.u32 %r410, 8;
|
|
mov.u64 %rd443, 0;
|
|
// inline asm
|
|
call (%rd432), _rt_buffer_get_64, (%rd433, %r99, %r410, %rd19, %rd20, %rd443, %rd443);
|
|
// inline asm
|
|
ld.v4.u16 {%rs265, %rs266, %rs267, %rs268}, [%rd432];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f886, %rs265;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f887, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f888, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd438), _rt_buffer_get_64, (%rd433, %r99, %r410, %rd19, %rd20, %rd443, %rd443);
|
|
// inline asm
|
|
add.f32 %f889, %f886, 0f00000000;
|
|
add.f32 %f890, %f887, 0f00000000;
|
|
add.f32 %f891, %f888, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f891;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f890;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs262, %f889;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs269, 0;
|
|
st.v4.u16 [%rd438], {%rs262, %rs263, %rs264, %rs269};
|
|
bra.uni BB0_166;
|
|
|
|
BB0_165:
|
|
mov.u64 %rd451, image_RNM6;
|
|
cvta.global.u64 %rd446, %rd451;
|
|
mov.u32 %r412, 8;
|
|
mov.u64 %rd450, 0;
|
|
// inline asm
|
|
call (%rd445), _rt_buffer_get_64, (%rd446, %r99, %r412, %rd19, %rd20, %rd450, %rd450);
|
|
// inline asm
|
|
mov.f32 %f892, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs270, %f892;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs271, 0;
|
|
st.v4.u16 [%rd445], {%rs270, %rs270, %rs270, %rs271};
|
|
|
|
BB0_166:
|
|
ld.global.u32 %r413, [additive];
|
|
setp.eq.s32 %p132, %r413, 0;
|
|
@%p132 bra BB0_168;
|
|
|
|
mov.u64 %rd464, image_RNM7;
|
|
cvta.global.u64 %rd453, %rd464;
|
|
mov.u32 %r417, 8;
|
|
mov.u64 %rd463, 0;
|
|
// inline asm
|
|
call (%rd452), _rt_buffer_get_64, (%rd453, %r99, %r417, %rd19, %rd20, %rd463, %rd463);
|
|
// inline asm
|
|
ld.v4.u16 {%rs278, %rs279, %rs280, %rs281}, [%rd452];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f893, %rs278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f894, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f895, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd458), _rt_buffer_get_64, (%rd453, %r99, %r417, %rd19, %rd20, %rd463, %rd463);
|
|
// inline asm
|
|
add.f32 %f896, %f893, 0f00000000;
|
|
add.f32 %f897, %f894, 0f00000000;
|
|
add.f32 %f898, %f895, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f898;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f897;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs275, %f896;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs282, 0;
|
|
st.v4.u16 [%rd458], {%rs275, %rs276, %rs277, %rs282};
|
|
bra.uni BB0_169;
|
|
|
|
BB0_168:
|
|
mov.u64 %rd471, image_RNM7;
|
|
cvta.global.u64 %rd466, %rd471;
|
|
mov.u32 %r419, 8;
|
|
mov.u64 %rd470, 0;
|
|
// inline asm
|
|
call (%rd465), _rt_buffer_get_64, (%rd466, %r99, %r419, %rd19, %rd20, %rd470, %rd470);
|
|
// inline asm
|
|
mov.f32 %f899, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs283, %f899;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs284, 0;
|
|
st.v4.u16 [%rd465], {%rs283, %rs283, %rs283, %rs284};
|
|
|
|
BB0_169:
|
|
ld.global.u32 %r420, [additive];
|
|
setp.eq.s32 %p133, %r420, 0;
|
|
@%p133 bra BB0_171;
|
|
|
|
mov.u64 %rd484, image_RNM8;
|
|
cvta.global.u64 %rd473, %rd484;
|
|
mov.u32 %r424, 8;
|
|
mov.u64 %rd483, 0;
|
|
// inline asm
|
|
call (%rd472), _rt_buffer_get_64, (%rd473, %r99, %r424, %rd19, %rd20, %rd483, %rd483);
|
|
// inline asm
|
|
ld.v4.u16 {%rs291, %rs292, %rs293, %rs294}, [%rd472];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f900, %rs291;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f901, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f902, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd478), _rt_buffer_get_64, (%rd473, %r99, %r424, %rd19, %rd20, %rd483, %rd483);
|
|
// inline asm
|
|
add.f32 %f903, %f900, 0f00000000;
|
|
add.f32 %f904, %f901, 0f00000000;
|
|
add.f32 %f905, %f902, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f905;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f904;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs288, %f903;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs295, 0;
|
|
st.v4.u16 [%rd478], {%rs288, %rs289, %rs290, %rs295};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_171:
|
|
mov.u64 %rd491, image_RNM8;
|
|
cvta.global.u64 %rd486, %rd491;
|
|
mov.u32 %r426, 8;
|
|
mov.u64 %rd490, 0;
|
|
// inline asm
|
|
call (%rd485), _rt_buffer_get_64, (%rd486, %r99, %r426, %rd19, %rd20, %rd490, %rd490);
|
|
// inline asm
|
|
mov.f32 %f906, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs296, %f906;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs297, 0;
|
|
st.v4.u16 [%rd485], {%rs296, %rs296, %rs296, %rs297};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_62:
|
|
setp.geu.f32 %p54, %f87, 0f00000000;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.f32 %f946, 0f3EE8BA2E;
|
|
cvt.rzi.f32.f32 %f458, %f946;
|
|
setp.neu.f32 %p55, %f458, 0f3EE8BA2E;
|
|
selp.f32 %f982, 0f7FFFFFFF, %f982, %p55;
|
|
|
|
BB0_65:
|
|
abs.f32 %f942, %f87;
|
|
add.f32 %f460, %f942, 0f3EE8BA2E;
|
|
mov.b32 %r243, %f460;
|
|
setp.lt.s32 %p57, %r243, 2139095040;
|
|
@%p57 bra BB0_70;
|
|
|
|
abs.f32 %f944, %f87;
|
|
setp.gtu.f32 %p58, %f944, 0f7F800000;
|
|
@%p58 bra BB0_69;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_69:
|
|
add.f32 %f982, %f87, 0f3EE8BA2E;
|
|
bra.uni BB0_70;
|
|
|
|
BB0_67:
|
|
abs.f32 %f945, %f87;
|
|
setp.neu.f32 %p59, %f945, 0f7F800000;
|
|
@%p59 bra BB0_70;
|
|
|
|
selp.f32 %f982, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_70:
|
|
mul.f32 %f461, %f982, 0f437F0000;
|
|
setp.eq.f32 %p60, %f87, 0f3F800000;
|
|
selp.f32 %f462, 0f437F0000, %f461, %p60;
|
|
cvt.rzi.u32.f32 %r244, %f462;
|
|
cvt.u16.u32 %rs19, %r244;
|
|
mov.u16 %rs20, 255;
|
|
st.v2.u8 [%rd64], {%rs19, %rs20};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_71:
|
|
and.b32 %r245, %r457, 1;
|
|
setp.eq.b32 %p61, %r245, 1;
|
|
@!%p61 bra BB0_106;
|
|
bra.uni BB0_72;
|
|
|
|
BB0_72:
|
|
mov.f32 %f943, 0f00000000;
|
|
abs.f32 %f108, %f91;
|
|
setp.lt.f32 %p62, %f108, 0f00800000;
|
|
mul.f32 %f468, %f108, 0f4B800000;
|
|
selp.f32 %f469, 0fC3170000, 0fC2FE0000, %p62;
|
|
selp.f32 %f470, %f468, %f108, %p62;
|
|
mov.b32 %r246, %f470;
|
|
and.b32 %r247, %r246, 8388607;
|
|
or.b32 %r248, %r247, 1065353216;
|
|
mov.b32 %f471, %r248;
|
|
shr.u32 %r249, %r246, 23;
|
|
cvt.rn.f32.u32 %f472, %r249;
|
|
add.f32 %f473, %f469, %f472;
|
|
setp.gt.f32 %p63, %f471, 0f3FB504F3;
|
|
mul.f32 %f474, %f471, 0f3F000000;
|
|
add.f32 %f475, %f473, 0f3F800000;
|
|
selp.f32 %f476, %f474, %f471, %p63;
|
|
selp.f32 %f477, %f475, %f473, %p63;
|
|
add.f32 %f478, %f476, 0fBF800000;
|
|
add.f32 %f464, %f476, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f463,%f464;
|
|
// inline asm
|
|
add.f32 %f479, %f478, %f478;
|
|
mul.f32 %f480, %f463, %f479;
|
|
mul.f32 %f481, %f480, %f480;
|
|
mov.f32 %f482, 0f3C4CAF63;
|
|
mov.f32 %f483, 0f3B18F0FE;
|
|
fma.rn.f32 %f484, %f483, %f481, %f482;
|
|
mov.f32 %f485, 0f3DAAAABD;
|
|
fma.rn.f32 %f486, %f484, %f481, %f485;
|
|
mul.rn.f32 %f487, %f486, %f481;
|
|
mul.rn.f32 %f488, %f487, %f480;
|
|
sub.f32 %f489, %f478, %f480;
|
|
neg.f32 %f490, %f480;
|
|
add.f32 %f491, %f489, %f489;
|
|
fma.rn.f32 %f492, %f490, %f478, %f491;
|
|
mul.rn.f32 %f493, %f463, %f492;
|
|
add.f32 %f494, %f488, %f480;
|
|
sub.f32 %f495, %f480, %f494;
|
|
add.f32 %f496, %f488, %f495;
|
|
add.f32 %f497, %f493, %f496;
|
|
add.f32 %f498, %f494, %f497;
|
|
sub.f32 %f499, %f494, %f498;
|
|
add.f32 %f500, %f497, %f499;
|
|
mov.f32 %f501, 0f3F317200;
|
|
mul.rn.f32 %f502, %f477, %f501;
|
|
mov.f32 %f503, 0f35BFBE8E;
|
|
mul.rn.f32 %f504, %f477, %f503;
|
|
add.f32 %f505, %f502, %f498;
|
|
sub.f32 %f506, %f502, %f505;
|
|
add.f32 %f507, %f498, %f506;
|
|
add.f32 %f508, %f500, %f507;
|
|
add.f32 %f509, %f504, %f508;
|
|
add.f32 %f510, %f505, %f509;
|
|
sub.f32 %f511, %f505, %f510;
|
|
add.f32 %f512, %f509, %f511;
|
|
mov.f32 %f513, 0f3EE66666;
|
|
mul.rn.f32 %f514, %f513, %f510;
|
|
neg.f32 %f515, %f514;
|
|
fma.rn.f32 %f516, %f513, %f510, %f515;
|
|
fma.rn.f32 %f517, %f513, %f512, %f516;
|
|
fma.rn.f32 %f519, %f943, %f510, %f517;
|
|
add.rn.f32 %f520, %f514, %f519;
|
|
neg.f32 %f521, %f520;
|
|
add.rn.f32 %f522, %f514, %f521;
|
|
add.rn.f32 %f523, %f522, %f519;
|
|
mov.b32 %r250, %f520;
|
|
setp.eq.s32 %p64, %r250, 1118925336;
|
|
add.s32 %r251, %r250, -1;
|
|
mov.b32 %f524, %r251;
|
|
add.f32 %f525, %f523, 0f37000000;
|
|
selp.f32 %f526, %f524, %f520, %p64;
|
|
selp.f32 %f109, %f525, %f523, %p64;
|
|
mul.f32 %f527, %f526, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f528, %f527;
|
|
mov.f32 %f529, 0fBF317200;
|
|
fma.rn.f32 %f530, %f528, %f529, %f526;
|
|
mov.f32 %f531, 0fB5BFBE8E;
|
|
fma.rn.f32 %f532, %f528, %f531, %f530;
|
|
mul.f32 %f533, %f532, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f534, %f533;
|
|
add.f32 %f535, %f528, 0f00000000;
|
|
ex2.approx.f32 %f536, %f535;
|
|
mul.f32 %f537, %f534, %f536;
|
|
setp.lt.f32 %p65, %f526, 0fC2D20000;
|
|
selp.f32 %f538, 0f00000000, %f537, %p65;
|
|
setp.gt.f32 %p66, %f526, 0f42D20000;
|
|
selp.f32 %f983, 0f7F800000, %f538, %p66;
|
|
setp.eq.f32 %p67, %f983, 0f7F800000;
|
|
@%p67 bra BB0_74;
|
|
|
|
fma.rn.f32 %f983, %f983, %f109, %f983;
|
|
|
|
BB0_74:
|
|
mov.f32 %f910, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f909, %f910;
|
|
fma.rn.f32 %f908, %f909, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f907, %f908;
|
|
setp.lt.f32 %p68, %f91, 0f00000000;
|
|
setp.eq.f32 %p69, %f907, 0f3F800000;
|
|
and.pred %p2, %p68, %p69;
|
|
mov.b32 %r252, %f983;
|
|
xor.b32 %r253, %r252, -2147483648;
|
|
mov.b32 %f539, %r253;
|
|
selp.f32 %f985, %f539, %f983, %p2;
|
|
setp.eq.f32 %p70, %f91, 0f00000000;
|
|
@%p70 bra BB0_77;
|
|
bra.uni BB0_75;
|
|
|
|
BB0_77:
|
|
add.f32 %f542, %f91, %f91;
|
|
selp.f32 %f985, %f542, 0f00000000, %p69;
|
|
bra.uni BB0_78;
|
|
|
|
BB0_75:
|
|
setp.geu.f32 %p71, %f91, 0f00000000;
|
|
@%p71 bra BB0_78;
|
|
|
|
mov.f32 %f934, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f541, %f934;
|
|
setp.neu.f32 %p72, %f541, 0f3EE66666;
|
|
selp.f32 %f985, 0f7FFFFFFF, %f985, %p72;
|
|
|
|
BB0_78:
|
|
abs.f32 %f911, %f91;
|
|
add.f32 %f543, %f911, 0f3EE66666;
|
|
mov.b32 %r254, %f543;
|
|
setp.lt.s32 %p74, %r254, 2139095040;
|
|
@%p74 bra BB0_83;
|
|
|
|
abs.f32 %f932, %f91;
|
|
setp.gtu.f32 %p75, %f932, 0f7F800000;
|
|
@%p75 bra BB0_82;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_82:
|
|
add.f32 %f985, %f91, 0f3EE66666;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_80:
|
|
abs.f32 %f933, %f91;
|
|
setp.neu.f32 %p76, %f933, 0f7F800000;
|
|
@%p76 bra BB0_83;
|
|
|
|
selp.f32 %f985, 0fFF800000, 0f7F800000, %p2;
|
|
|
|
BB0_83:
|
|
mov.f32 %f920, 0fB5BFBE8E;
|
|
mov.f32 %f919, 0fBF317200;
|
|
mov.f32 %f918, 0f35BFBE8E;
|
|
mov.f32 %f917, 0f3F317200;
|
|
mov.f32 %f916, 0f3DAAAABD;
|
|
mov.f32 %f915, 0f3C4CAF63;
|
|
mov.f32 %f914, 0f3B18F0FE;
|
|
mov.f32 %f913, 0f3EE66666;
|
|
mov.f32 %f912, 0f00000000;
|
|
setp.eq.f32 %p77, %f91, 0f3F800000;
|
|
selp.f32 %f120, 0f3F800000, %f985, %p77;
|
|
abs.f32 %f121, %f92;
|
|
setp.lt.f32 %p78, %f121, 0f00800000;
|
|
mul.f32 %f546, %f121, 0f4B800000;
|
|
selp.f32 %f547, 0fC3170000, 0fC2FE0000, %p78;
|
|
selp.f32 %f548, %f546, %f121, %p78;
|
|
mov.b32 %r255, %f548;
|
|
and.b32 %r256, %r255, 8388607;
|
|
or.b32 %r257, %r256, 1065353216;
|
|
mov.b32 %f549, %r257;
|
|
shr.u32 %r258, %r255, 23;
|
|
cvt.rn.f32.u32 %f550, %r258;
|
|
add.f32 %f551, %f547, %f550;
|
|
setp.gt.f32 %p79, %f549, 0f3FB504F3;
|
|
mul.f32 %f552, %f549, 0f3F000000;
|
|
add.f32 %f553, %f551, 0f3F800000;
|
|
selp.f32 %f554, %f552, %f549, %p79;
|
|
selp.f32 %f555, %f553, %f551, %p79;
|
|
add.f32 %f556, %f554, 0fBF800000;
|
|
add.f32 %f545, %f554, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f544,%f545;
|
|
// inline asm
|
|
add.f32 %f557, %f556, %f556;
|
|
mul.f32 %f558, %f544, %f557;
|
|
mul.f32 %f559, %f558, %f558;
|
|
fma.rn.f32 %f562, %f914, %f559, %f915;
|
|
fma.rn.f32 %f564, %f562, %f559, %f916;
|
|
mul.rn.f32 %f565, %f564, %f559;
|
|
mul.rn.f32 %f566, %f565, %f558;
|
|
sub.f32 %f567, %f556, %f558;
|
|
neg.f32 %f568, %f558;
|
|
add.f32 %f569, %f567, %f567;
|
|
fma.rn.f32 %f570, %f568, %f556, %f569;
|
|
mul.rn.f32 %f571, %f544, %f570;
|
|
add.f32 %f572, %f566, %f558;
|
|
sub.f32 %f573, %f558, %f572;
|
|
add.f32 %f574, %f566, %f573;
|
|
add.f32 %f575, %f571, %f574;
|
|
add.f32 %f576, %f572, %f575;
|
|
sub.f32 %f577, %f572, %f576;
|
|
add.f32 %f578, %f575, %f577;
|
|
mul.rn.f32 %f580, %f555, %f917;
|
|
mul.rn.f32 %f582, %f555, %f918;
|
|
add.f32 %f583, %f580, %f576;
|
|
sub.f32 %f584, %f580, %f583;
|
|
add.f32 %f585, %f576, %f584;
|
|
add.f32 %f586, %f578, %f585;
|
|
add.f32 %f587, %f582, %f586;
|
|
add.f32 %f588, %f583, %f587;
|
|
sub.f32 %f589, %f583, %f588;
|
|
add.f32 %f590, %f587, %f589;
|
|
mul.rn.f32 %f592, %f913, %f588;
|
|
neg.f32 %f593, %f592;
|
|
fma.rn.f32 %f594, %f913, %f588, %f593;
|
|
fma.rn.f32 %f595, %f913, %f590, %f594;
|
|
fma.rn.f32 %f597, %f912, %f588, %f595;
|
|
add.rn.f32 %f598, %f592, %f597;
|
|
neg.f32 %f599, %f598;
|
|
add.rn.f32 %f600, %f592, %f599;
|
|
add.rn.f32 %f601, %f600, %f597;
|
|
mov.b32 %r259, %f598;
|
|
setp.eq.s32 %p80, %r259, 1118925336;
|
|
add.s32 %r260, %r259, -1;
|
|
mov.b32 %f602, %r260;
|
|
add.f32 %f603, %f601, 0f37000000;
|
|
selp.f32 %f604, %f602, %f598, %p80;
|
|
selp.f32 %f122, %f603, %f601, %p80;
|
|
mul.f32 %f605, %f604, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f606, %f605;
|
|
fma.rn.f32 %f608, %f606, %f919, %f604;
|
|
fma.rn.f32 %f610, %f606, %f920, %f608;
|
|
mul.f32 %f611, %f610, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f612, %f611;
|
|
add.f32 %f613, %f606, 0f00000000;
|
|
ex2.approx.f32 %f614, %f613;
|
|
mul.f32 %f615, %f612, %f614;
|
|
setp.lt.f32 %p81, %f604, 0fC2D20000;
|
|
selp.f32 %f616, 0f00000000, %f615, %p81;
|
|
setp.gt.f32 %p82, %f604, 0f42D20000;
|
|
selp.f32 %f986, 0f7F800000, %f616, %p82;
|
|
setp.eq.f32 %p83, %f986, 0f7F800000;
|
|
@%p83 bra BB0_85;
|
|
|
|
fma.rn.f32 %f986, %f986, %f122, %f986;
|
|
|
|
BB0_85:
|
|
setp.lt.f32 %p84, %f92, 0f00000000;
|
|
and.pred %p3, %p84, %p69;
|
|
mov.b32 %r261, %f986;
|
|
xor.b32 %r262, %r261, -2147483648;
|
|
mov.b32 %f617, %r262;
|
|
selp.f32 %f988, %f617, %f986, %p3;
|
|
setp.eq.f32 %p86, %f92, 0f00000000;
|
|
@%p86 bra BB0_88;
|
|
bra.uni BB0_86;
|
|
|
|
BB0_88:
|
|
add.f32 %f620, %f92, %f92;
|
|
selp.f32 %f988, %f620, 0f00000000, %p69;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_86:
|
|
setp.geu.f32 %p87, %f92, 0f00000000;
|
|
@%p87 bra BB0_89;
|
|
|
|
mov.f32 %f931, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f619, %f931;
|
|
setp.neu.f32 %p88, %f619, 0f3EE66666;
|
|
selp.f32 %f988, 0f7FFFFFFF, %f988, %p88;
|
|
|
|
BB0_89:
|
|
abs.f32 %f935, %f92;
|
|
add.f32 %f621, %f935, 0f3EE66666;
|
|
mov.b32 %r263, %f621;
|
|
setp.lt.s32 %p90, %r263, 2139095040;
|
|
@%p90 bra BB0_94;
|
|
|
|
abs.f32 %f936, %f92;
|
|
setp.gtu.f32 %p91, %f936, 0f7F800000;
|
|
@%p91 bra BB0_93;
|
|
bra.uni BB0_91;
|
|
|
|
BB0_93:
|
|
add.f32 %f988, %f92, 0f3EE66666;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_91:
|
|
abs.f32 %f937, %f92;
|
|
setp.neu.f32 %p92, %f937, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
selp.f32 %f988, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_94:
|
|
mov.f32 %f929, 0fB5BFBE8E;
|
|
mov.f32 %f928, 0fBF317200;
|
|
mov.f32 %f927, 0f35BFBE8E;
|
|
mov.f32 %f926, 0f3F317200;
|
|
mov.f32 %f925, 0f3DAAAABD;
|
|
mov.f32 %f924, 0f3C4CAF63;
|
|
mov.f32 %f923, 0f3B18F0FE;
|
|
mov.f32 %f922, 0f3EE66666;
|
|
mov.f32 %f921, 0f00000000;
|
|
setp.eq.f32 %p93, %f92, 0f3F800000;
|
|
selp.f32 %f133, 0f3F800000, %f988, %p93;
|
|
abs.f32 %f134, %f93;
|
|
setp.lt.f32 %p94, %f134, 0f00800000;
|
|
mul.f32 %f624, %f134, 0f4B800000;
|
|
selp.f32 %f625, 0fC3170000, 0fC2FE0000, %p94;
|
|
selp.f32 %f626, %f624, %f134, %p94;
|
|
mov.b32 %r264, %f626;
|
|
and.b32 %r265, %r264, 8388607;
|
|
or.b32 %r266, %r265, 1065353216;
|
|
mov.b32 %f627, %r266;
|
|
shr.u32 %r267, %r264, 23;
|
|
cvt.rn.f32.u32 %f628, %r267;
|
|
add.f32 %f629, %f625, %f628;
|
|
setp.gt.f32 %p95, %f627, 0f3FB504F3;
|
|
mul.f32 %f630, %f627, 0f3F000000;
|
|
add.f32 %f631, %f629, 0f3F800000;
|
|
selp.f32 %f632, %f630, %f627, %p95;
|
|
selp.f32 %f633, %f631, %f629, %p95;
|
|
add.f32 %f634, %f632, 0fBF800000;
|
|
add.f32 %f623, %f632, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f622,%f623;
|
|
// inline asm
|
|
add.f32 %f635, %f634, %f634;
|
|
mul.f32 %f636, %f622, %f635;
|
|
mul.f32 %f637, %f636, %f636;
|
|
fma.rn.f32 %f640, %f923, %f637, %f924;
|
|
fma.rn.f32 %f642, %f640, %f637, %f925;
|
|
mul.rn.f32 %f643, %f642, %f637;
|
|
mul.rn.f32 %f644, %f643, %f636;
|
|
sub.f32 %f645, %f634, %f636;
|
|
neg.f32 %f646, %f636;
|
|
add.f32 %f647, %f645, %f645;
|
|
fma.rn.f32 %f648, %f646, %f634, %f647;
|
|
mul.rn.f32 %f649, %f622, %f648;
|
|
add.f32 %f650, %f644, %f636;
|
|
sub.f32 %f651, %f636, %f650;
|
|
add.f32 %f652, %f644, %f651;
|
|
add.f32 %f653, %f649, %f652;
|
|
add.f32 %f654, %f650, %f653;
|
|
sub.f32 %f655, %f650, %f654;
|
|
add.f32 %f656, %f653, %f655;
|
|
mul.rn.f32 %f658, %f633, %f926;
|
|
mul.rn.f32 %f660, %f633, %f927;
|
|
add.f32 %f661, %f658, %f654;
|
|
sub.f32 %f662, %f658, %f661;
|
|
add.f32 %f663, %f654, %f662;
|
|
add.f32 %f664, %f656, %f663;
|
|
add.f32 %f665, %f660, %f664;
|
|
add.f32 %f666, %f661, %f665;
|
|
sub.f32 %f667, %f661, %f666;
|
|
add.f32 %f668, %f665, %f667;
|
|
mul.rn.f32 %f670, %f922, %f666;
|
|
neg.f32 %f671, %f670;
|
|
fma.rn.f32 %f672, %f922, %f666, %f671;
|
|
fma.rn.f32 %f673, %f922, %f668, %f672;
|
|
fma.rn.f32 %f675, %f921, %f666, %f673;
|
|
add.rn.f32 %f676, %f670, %f675;
|
|
neg.f32 %f677, %f676;
|
|
add.rn.f32 %f678, %f670, %f677;
|
|
add.rn.f32 %f679, %f678, %f675;
|
|
mov.b32 %r268, %f676;
|
|
setp.eq.s32 %p96, %r268, 1118925336;
|
|
add.s32 %r269, %r268, -1;
|
|
mov.b32 %f680, %r269;
|
|
add.f32 %f681, %f679, 0f37000000;
|
|
selp.f32 %f682, %f680, %f676, %p96;
|
|
selp.f32 %f135, %f681, %f679, %p96;
|
|
mul.f32 %f683, %f682, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f684, %f683;
|
|
fma.rn.f32 %f686, %f684, %f928, %f682;
|
|
fma.rn.f32 %f688, %f684, %f929, %f686;
|
|
mul.f32 %f689, %f688, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f690, %f689;
|
|
add.f32 %f691, %f684, 0f00000000;
|
|
ex2.approx.f32 %f692, %f691;
|
|
mul.f32 %f693, %f690, %f692;
|
|
setp.lt.f32 %p97, %f682, 0fC2D20000;
|
|
selp.f32 %f694, 0f00000000, %f693, %p97;
|
|
setp.gt.f32 %p98, %f682, 0f42D20000;
|
|
selp.f32 %f989, 0f7F800000, %f694, %p98;
|
|
setp.eq.f32 %p99, %f989, 0f7F800000;
|
|
@%p99 bra BB0_96;
|
|
|
|
fma.rn.f32 %f989, %f989, %f135, %f989;
|
|
|
|
BB0_96:
|
|
setp.lt.f32 %p100, %f93, 0f00000000;
|
|
and.pred %p4, %p100, %p69;
|
|
mov.b32 %r270, %f989;
|
|
xor.b32 %r271, %r270, -2147483648;
|
|
mov.b32 %f695, %r271;
|
|
selp.f32 %f991, %f695, %f989, %p4;
|
|
setp.eq.f32 %p102, %f93, 0f00000000;
|
|
@%p102 bra BB0_99;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_99:
|
|
add.f32 %f698, %f93, %f93;
|
|
selp.f32 %f991, %f698, 0f00000000, %p69;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_97:
|
|
setp.geu.f32 %p103, %f93, 0f00000000;
|
|
@%p103 bra BB0_100;
|
|
|
|
mov.f32 %f930, 0f3EE66666;
|
|
cvt.rzi.f32.f32 %f697, %f930;
|
|
setp.neu.f32 %p104, %f697, 0f3EE66666;
|
|
selp.f32 %f991, 0f7FFFFFFF, %f991, %p104;
|
|
|
|
BB0_100:
|
|
abs.f32 %f947, %f93;
|
|
add.f32 %f699, %f947, 0f3EE66666;
|
|
mov.b32 %r272, %f699;
|
|
setp.lt.s32 %p106, %r272, 2139095040;
|
|
@%p106 bra BB0_105;
|
|
|
|
abs.f32 %f948, %f93;
|
|
setp.gtu.f32 %p107, %f948, 0f7F800000;
|
|
@%p107 bra BB0_104;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_104:
|
|
add.f32 %f991, %f93, 0f3EE66666;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_102:
|
|
abs.f32 %f949, %f93;
|
|
setp.neu.f32 %p108, %f949, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
selp.f32 %f991, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_105:
|
|
mov.u32 %r427, 4;
|
|
setp.eq.f32 %p109, %f93, 0f3F800000;
|
|
selp.f32 %f700, 0f3F800000, %f991, %p109;
|
|
cvt.u64.u32 %rd74, %r3;
|
|
cvt.u64.u32 %rd73, %r2;
|
|
mov.u64 %rd77, image;
|
|
cvta.global.u64 %rd72, %rd77;
|
|
// inline asm
|
|
call (%rd71), _rt_buffer_get_64, (%rd72, %r99, %r427, %rd73, %rd74, %rd26, %rd26);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f701, %f700;
|
|
mul.f32 %f702, %f701, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r275, %f702;
|
|
cvt.sat.f32.f32 %f703, %f133;
|
|
mul.f32 %f704, %f703, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r276, %f704;
|
|
cvt.sat.f32.f32 %f705, %f120;
|
|
mul.f32 %f706, %f705, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r277, %f706;
|
|
cvt.u16.u32 %rs21, %r275;
|
|
cvt.u16.u32 %rs22, %r277;
|
|
cvt.u16.u32 %rs23, %r276;
|
|
mov.u16 %rs24, 255;
|
|
st.v4.u8 [%rd71], {%rs21, %rs23, %rs22, %rs24};
|
|
ld.global.u32 %r457, [imageEnabled];
|
|
|
|
BB0_106:
|
|
cvt.u64.u32 %rd17, %r2;
|
|
cvt.u64.u32 %rd18, %r3;
|
|
and.b32 %r278, %r457, 4;
|
|
setp.eq.s32 %p110, %r278, 0;
|
|
@%p110 bra BB0_110;
|
|
|
|
ld.global.u32 %r279, [additive];
|
|
setp.eq.s32 %p111, %r279, 0;
|
|
mov.f32 %f707, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f707;}
|
|
|
|
// inline asm
|
|
@%p111 bra BB0_109;
|
|
|
|
mov.u64 %rd90, image_HDR;
|
|
cvta.global.u64 %rd79, %rd90;
|
|
mov.u32 %r283, 8;
|
|
// inline asm
|
|
call (%rd78), _rt_buffer_get_64, (%rd79, %r99, %r283, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs32, %rs33, %rs34, %rs35}, [%rd78];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f708, %rs32;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f709, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f710, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd79, %r99, %r283, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f711, %f91, %f708;
|
|
add.f32 %f712, %f92, %f709;
|
|
add.f32 %f713, %f93, %f710;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f713;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f712;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs29, %f711;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd84], {%rs29, %rs30, %rs31, %rs25};
|
|
bra.uni BB0_110;
|
|
|
|
BB0_109:
|
|
mov.u64 %rd97, image_HDR;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
mov.u32 %r285, 8;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r99, %r285, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f93;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f92;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs36, %f91;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd91], {%rs36, %rs37, %rs38, %rs25};
|
|
|
|
BB0_110:
|
|
mul.f32 %f146, %f88, 0f403D3C19;
|
|
mul.f32 %f147, %f146, 0f3E906EBB;
|
|
mul.f32 %f148, %f89, 0f403D3C19;
|
|
mul.f32 %f149, %f148, 0f3E906EBB;
|
|
mul.f32 %f150, %f90, 0f403D3C19;
|
|
mul.f32 %f151, %f150, 0f3E906EBB;
|
|
ld.global.u32 %r286, [additive];
|
|
setp.eq.s32 %p112, %r286, 0;
|
|
mov.f32 %f717, 0f3F800000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f717;}
|
|
|
|
// inline asm
|
|
@%p112 bra BB0_112;
|
|
|
|
mov.u64 %rd110, image_RNM0;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r290, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r99, %r290, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs46, %rs47, %rs48, %rs49}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f718, %rs46;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f719, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f720, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r99, %r290, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f721, %f147, %f718;
|
|
add.f32 %f722, %f149, %f719;
|
|
add.f32 %f723, %f151, %f720;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f723;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f722;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs43, %f721;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs43, %rs44, %rs45, %rs39};
|
|
bra.uni BB0_113;
|
|
|
|
BB0_112:
|
|
mov.u64 %rd117, image_RNM0;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r292, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r99, %r292, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f151;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f149;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs50, %f147;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs50, %rs51, %rs52, %rs39};
|
|
|
|
BB0_113:
|
|
ld.global.f32 %f728, [directDir+4];
|
|
mul.f32 %f729, %f728, 0fBEFA2A1C;
|
|
mul.f32 %f152, %f146, %f729;
|
|
mul.f32 %f153, %f148, %f729;
|
|
mul.f32 %f154, %f150, %f729;
|
|
ld.global.u32 %r293, [additive];
|
|
setp.eq.s32 %p113, %r293, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f717;}
|
|
|
|
// inline asm
|
|
@%p113 bra BB0_115;
|
|
|
|
mov.u64 %rd130, image_RNM1;
|
|
cvta.global.u64 %rd119, %rd130;
|
|
mov.u32 %r297, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r99, %r297, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs60, %rs61, %rs62, %rs63}, [%rd118];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f730, %rs60;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f731, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f732, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd119, %r99, %r297, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f733, %f152, %f730;
|
|
add.f32 %f734, %f153, %f731;
|
|
add.f32 %f735, %f154, %f732;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f735;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f734;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs57, %f733;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs57, %rs58, %rs59, %rs53};
|
|
bra.uni BB0_116;
|
|
|
|
BB0_115:
|
|
mov.u64 %rd137, image_RNM1;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u32 %r299, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r99, %r299, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs64, %f152;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs64, %rs65, %rs66, %rs53};
|
|
|
|
BB0_116:
|
|
ld.global.f32 %f740, [directDir+8];
|
|
mul.f32 %f741, %f740, 0fBEFA2A1C;
|
|
mul.f32 %f155, %f146, %f741;
|
|
mul.f32 %f156, %f148, %f741;
|
|
mul.f32 %f157, %f150, %f741;
|
|
ld.global.u32 %r300, [additive];
|
|
setp.eq.s32 %p114, %r300, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f717;}
|
|
|
|
// inline asm
|
|
@%p114 bra BB0_118;
|
|
|
|
mov.u64 %rd150, image_RNM2;
|
|
cvta.global.u64 %rd139, %rd150;
|
|
mov.u32 %r304, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r99, %r304, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs74, %rs75, %rs76, %rs77}, [%rd138];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f742, %rs74;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f743, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f744, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd139, %r99, %r304, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f745, %f155, %f742;
|
|
add.f32 %f746, %f156, %f743;
|
|
add.f32 %f747, %f157, %f744;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f747;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f746;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs71, %f745;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs71, %rs72, %rs73, %rs67};
|
|
bra.uni BB0_119;
|
|
|
|
BB0_118:
|
|
mov.u64 %rd157, image_RNM2;
|
|
cvta.global.u64 %rd152, %rd157;
|
|
mov.u32 %r306, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r99, %r306, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f157;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f156;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs78, %f155;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd151], {%rs78, %rs79, %rs80, %rs67};
|
|
|
|
BB0_119:
|
|
ld.global.f32 %f752, [directDir];
|
|
mul.f32 %f753, %f752, 0fBEFA2A1C;
|
|
mul.f32 %f158, %f146, %f753;
|
|
mul.f32 %f159, %f148, %f753;
|
|
mul.f32 %f160, %f150, %f753;
|
|
ld.global.u32 %r307, [additive];
|
|
setp.eq.s32 %p115, %r307, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f717;}
|
|
|
|
// inline asm
|
|
@%p115 bra BB0_121;
|
|
|
|
mov.u64 %rd170, image_RNM3;
|
|
cvta.global.u64 %rd159, %rd170;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r99, %r311, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs88, %rs89, %rs90, %rs91}, [%rd158];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f754, %rs88;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f755, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f756, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd159, %r99, %r311, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f757, %f158, %f754;
|
|
add.f32 %f758, %f159, %f755;
|
|
add.f32 %f759, %f160, %f756;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f759;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f758;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs85, %f757;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs85, %rs86, %rs87, %rs81};
|
|
bra.uni BB0_122;
|
|
|
|
BB0_121:
|
|
mov.u64 %rd177, image_RNM3;
|
|
cvta.global.u64 %rd172, %rd177;
|
|
mov.u32 %r313, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r99, %r313, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs92, %f158;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd171], {%rs92, %rs93, %rs94, %rs81};
|
|
|
|
BB0_122:
|
|
ld.global.f32 %f764, [directDir+4];
|
|
ld.global.f32 %f765, [directDir];
|
|
mul.f32 %f766, %f765, %f764;
|
|
mul.f32 %f767, %f766, 0f3F8BD8A1;
|
|
mul.f32 %f161, %f146, %f767;
|
|
mul.f32 %f162, %f148, %f767;
|
|
mul.f32 %f163, %f150, %f767;
|
|
ld.global.u32 %r314, [additive];
|
|
setp.eq.s32 %p116, %r314, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f717;}
|
|
|
|
// inline asm
|
|
@%p116 bra BB0_124;
|
|
|
|
mov.u64 %rd190, image_RNM4;
|
|
cvta.global.u64 %rd179, %rd190;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd178), _rt_buffer_get_64, (%rd179, %r99, %r318, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs102, %rs103, %rs104, %rs105}, [%rd178];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f768, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f769, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f770, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd179, %r99, %r318, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f771, %f161, %f768;
|
|
add.f32 %f772, %f162, %f769;
|
|
add.f32 %f773, %f163, %f770;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f773;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f772;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f771;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs99, %rs100, %rs101, %rs95};
|
|
bra.uni BB0_125;
|
|
|
|
BB0_124:
|
|
mov.u64 %rd197, image_RNM4;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u32 %r320, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r99, %r320, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f163;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f162;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs106, %f161;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd191], {%rs106, %rs107, %rs108, %rs95};
|
|
|
|
BB0_125:
|
|
ld.global.f32 %f778, [directDir+4];
|
|
ld.global.f32 %f779, [directDir+8];
|
|
mul.f32 %f780, %f778, %f779;
|
|
mul.f32 %f781, %f780, 0f3F8BD8A1;
|
|
mul.f32 %f164, %f146, %f781;
|
|
mul.f32 %f165, %f148, %f781;
|
|
mul.f32 %f166, %f150, %f781;
|
|
ld.global.u32 %r321, [additive];
|
|
setp.eq.s32 %p117, %r321, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f717;}
|
|
|
|
// inline asm
|
|
@%p117 bra BB0_127;
|
|
|
|
mov.u64 %rd210, image_RNM5;
|
|
cvta.global.u64 %rd199, %rd210;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r99, %r325, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs116, %rs117, %rs118, %rs119}, [%rd198];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f782, %rs116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f783, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f784, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd199, %r99, %r325, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f785, %f164, %f782;
|
|
add.f32 %f786, %f165, %f783;
|
|
add.f32 %f787, %f166, %f784;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f787;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f786;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs113, %f785;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd204], {%rs113, %rs114, %rs115, %rs109};
|
|
bra.uni BB0_128;
|
|
|
|
BB0_127:
|
|
mov.u64 %rd217, image_RNM5;
|
|
cvta.global.u64 %rd212, %rd217;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r99, %r327, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f166;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f165;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs120, %f164;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd211], {%rs120, %rs121, %rs122, %rs109};
|
|
|
|
BB0_128:
|
|
ld.global.f32 %f792, [directDir+8];
|
|
fma.rn.f32 %f793, %f792, %f792, 0fBEAAAAAB;
|
|
mul.f32 %f794, %f793, 0f3F723881;
|
|
mul.f32 %f167, %f146, %f794;
|
|
mul.f32 %f168, %f148, %f794;
|
|
mul.f32 %f169, %f150, %f794;
|
|
ld.global.u32 %r328, [additive];
|
|
setp.eq.s32 %p118, %r328, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f717;}
|
|
|
|
// inline asm
|
|
@%p118 bra BB0_130;
|
|
|
|
mov.u64 %rd230, image_RNM6;
|
|
cvta.global.u64 %rd219, %rd230;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r99, %r332, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs130, %rs131, %rs132, %rs133}, [%rd218];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f795, %rs130;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f796, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f797, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd219, %r99, %r332, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f798, %f167, %f795;
|
|
add.f32 %f799, %f168, %f796;
|
|
add.f32 %f800, %f169, %f797;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f800;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f799;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs127, %f798;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd224], {%rs127, %rs128, %rs129, %rs123};
|
|
bra.uni BB0_131;
|
|
|
|
BB0_130:
|
|
mov.u64 %rd237, image_RNM6;
|
|
cvta.global.u64 %rd232, %rd237;
|
|
mov.u32 %r334, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r99, %r334, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f169;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f168;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs134, %f167;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd231], {%rs134, %rs135, %rs136, %rs123};
|
|
|
|
BB0_131:
|
|
ld.global.f32 %f805, [directDir+8];
|
|
ld.global.f32 %f806, [directDir];
|
|
mul.f32 %f807, %f806, %f805;
|
|
mul.f32 %f808, %f807, 0f3F8BD8A1;
|
|
mul.f32 %f170, %f146, %f808;
|
|
mul.f32 %f171, %f148, %f808;
|
|
mul.f32 %f172, %f150, %f808;
|
|
ld.global.u32 %r335, [additive];
|
|
setp.eq.s32 %p119, %r335, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f717;}
|
|
|
|
// inline asm
|
|
@%p119 bra BB0_133;
|
|
|
|
mov.u64 %rd250, image_RNM7;
|
|
cvta.global.u64 %rd239, %rd250;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r99, %r339, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs144, %rs145, %rs146, %rs147}, [%rd238];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f809, %rs144;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f810, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f811, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd239, %r99, %r339, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f812, %f170, %f809;
|
|
add.f32 %f813, %f171, %f810;
|
|
add.f32 %f814, %f172, %f811;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f814;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f813;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs141, %f812;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd244], {%rs141, %rs142, %rs143, %rs137};
|
|
bra.uni BB0_134;
|
|
|
|
BB0_133:
|
|
mov.u64 %rd257, image_RNM7;
|
|
cvta.global.u64 %rd252, %rd257;
|
|
mov.u32 %r341, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r99, %r341, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f172;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f171;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs148, %f170;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd251], {%rs148, %rs149, %rs150, %rs137};
|
|
|
|
BB0_134:
|
|
ld.global.f32 %f819, [directDir];
|
|
mul.f32 %f820, %f819, %f819;
|
|
ld.global.f32 %f821, [directDir+4];
|
|
mul.f32 %f822, %f821, %f821;
|
|
sub.f32 %f823, %f820, %f822;
|
|
mul.f32 %f824, %f823, 0f3F0BD8A1;
|
|
mul.f32 %f173, %f146, %f824;
|
|
mul.f32 %f174, %f148, %f824;
|
|
mul.f32 %f175, %f150, %f824;
|
|
ld.global.u32 %r342, [additive];
|
|
setp.eq.s32 %p120, %r342, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f717;}
|
|
|
|
// inline asm
|
|
@%p120 bra BB0_136;
|
|
|
|
mov.u64 %rd270, image_RNM8;
|
|
cvta.global.u64 %rd259, %rd270;
|
|
mov.u32 %r346, 8;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r99, %r346, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
ld.v4.u16 {%rs158, %rs159, %rs160, %rs161}, [%rd258];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f825, %rs158;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f826, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f827, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd259, %r99, %r346, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
add.f32 %f828, %f173, %f825;
|
|
add.f32 %f829, %f174, %f826;
|
|
add.f32 %f830, %f175, %f827;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f830;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f829;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs155, %f828;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd264], {%rs155, %rs156, %rs157, %rs151};
|
|
bra.uni BB0_172;
|
|
|
|
BB0_136:
|
|
mov.u64 %rd277, image_RNM8;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r348, 8;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r99, %r348, %rd17, %rd18, %rd26, %rd26);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f174;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs162, %f173;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd271], {%rs162, %rs163, %rs164, %rs151};
|
|
|
|
BB0_172:
|
|
ret;
|
|
}
|
|
|
|
|