2515 lines
70 KiB
Plaintext
2515 lines
70 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 1 .b8 rnd_seeds2[1];
|
|
.global .align 4 .u32 samples;
|
|
.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_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[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 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[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_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 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_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[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_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[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 16 .b8 __local_depot0[64];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<149>;
|
|
.reg .b16 %rs<159>;
|
|
.reg .f32 %f<1265>;
|
|
.reg .b32 %r<418>;
|
|
.reg .b64 %rd<329>;
|
|
|
|
|
|
mov.u64 %rd328, __local_depot0;
|
|
cvta.local.u64 %SP, %rd328;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
shl.b32 %r4, %r3, 1;
|
|
ld.global.v2.u32 {%r103, %r104}, [pixelID];
|
|
cvt.u64.u32 %rd24, %r103;
|
|
cvt.u64.u32 %rd25, %r104;
|
|
mov.u64 %rd28, uvnormal;
|
|
cvta.global.u64 %rd23, %rd28;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd27, 0;
|
|
// inline asm
|
|
call (%rd22), _rt_buffer_get_64, (%rd23, %r101, %r102, %rd24, %rd25, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd22];
|
|
shr.u32 %r107, %r5, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r5;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p7, %rs9, 0;
|
|
mov.f32 %f1126, 0f00000000;
|
|
mov.f32 %f1127, %f1126;
|
|
mov.f32 %f1128, %f1126;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd22+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f334, %rs12;
|
|
div.rn.f32 %f335, %f334, 0f437F0000;
|
|
fma.rn.f32 %f336, %f335, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f337, %rs10;
|
|
div.rn.f32 %f338, %f337, 0f437F0000;
|
|
fma.rn.f32 %f339, %f338, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f340, %rs7;
|
|
div.rn.f32 %f341, %f340, 0f437F0000;
|
|
fma.rn.f32 %f342, %f341, 0f40000000, 0fBF800000;
|
|
mul.f32 %f343, %f339, %f339;
|
|
fma.rn.f32 %f344, %f336, %f336, %f343;
|
|
fma.rn.f32 %f345, %f342, %f342, %f344;
|
|
sqrt.rn.f32 %f346, %f345;
|
|
rcp.rn.f32 %f347, %f346;
|
|
mul.f32 %f1126, %f336, %f347;
|
|
mul.f32 %f1127, %f339, %f347;
|
|
mul.f32 %f1128, %f342, %f347;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r108, %r109}, [pixelID];
|
|
ld.global.v2.u32 {%r111, %r112}, [tileInfo];
|
|
add.s32 %r6, %r108, %r111;
|
|
add.s32 %r7, %r109, %r112;
|
|
setp.eq.f32 %p8, %f1127, 0f00000000;
|
|
setp.eq.f32 %p9, %f1126, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1128, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_140;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_140:
|
|
ld.global.u32 %r417, [imageEnabled];
|
|
and.b32 %r340, %r417, 1;
|
|
setp.eq.b32 %p140, %r340, 1;
|
|
@!%p140 bra BB0_142;
|
|
bra.uni BB0_141;
|
|
|
|
BB0_141:
|
|
cvt.u64.u32 %rd206, %r6;
|
|
cvt.u64.u32 %rd207, %r7;
|
|
mov.u64 %rd210, image;
|
|
cvta.global.u64 %rd205, %rd210;
|
|
mov.u64 %rd209, 0;
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd205, %r101, %r102, %rd206, %rd207, %rd209, %rd209);
|
|
// inline asm
|
|
mov.u16 %rs91, 0;
|
|
st.v4.u8 [%rd204], {%rs91, %rs91, %rs91, %rs91};
|
|
ld.global.u32 %r417, [imageEnabled];
|
|
|
|
BB0_142:
|
|
and.b32 %r343, %r417, 8;
|
|
setp.eq.s32 %p141, %r343, 0;
|
|
@%p141 bra BB0_144;
|
|
|
|
cvt.u64.u32 %rd214, %r7;
|
|
cvt.u64.u32 %rd213, %r6;
|
|
mov.u64 %rd217, image_Mask;
|
|
cvta.global.u64 %rd212, %rd217;
|
|
mov.u64 %rd216, 0;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r101, %r101, %rd213, %rd214, %rd216, %rd216);
|
|
// inline asm
|
|
mov.f32 %f1083, 0f00000000;
|
|
cvt.rzi.u32.f32 %r346, %f1083;
|
|
cvt.u16.u32 %rs92, %r346;
|
|
mov.u16 %rs93, 0;
|
|
st.v2.u8 [%rd211], {%rs92, %rs93};
|
|
ld.global.u32 %r417, [imageEnabled];
|
|
|
|
BB0_144:
|
|
cvt.u64.u32 %rd20, %r6;
|
|
cvt.u64.u32 %rd21, %r7;
|
|
and.b32 %r347, %r417, 4;
|
|
setp.eq.s32 %p142, %r347, 0;
|
|
@%p142 bra BB0_148;
|
|
|
|
ld.global.u32 %r348, [additive];
|
|
setp.eq.s32 %p143, %r348, 0;
|
|
@%p143 bra BB0_147;
|
|
|
|
mov.u64 %rd230, image_HDR;
|
|
cvta.global.u64 %rd219, %rd230;
|
|
mov.u32 %r352, 8;
|
|
mov.u64 %rd229, 0;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r101, %r352, %rd20, %rd21, %rd229, %rd229);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd218];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1084, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1085, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1086, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd219, %r101, %r352, %rd20, %rd21, %rd229, %rd229);
|
|
// inline asm
|
|
add.f32 %f1087, %f1084, 0f00000000;
|
|
add.f32 %f1088, %f1085, 0f00000000;
|
|
add.f32 %f1089, %f1086, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1089;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1088;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f1087;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs104, 0;
|
|
st.v4.u16 [%rd224], {%rs97, %rs98, %rs99, %rs104};
|
|
bra.uni BB0_148;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd31, %r121;
|
|
cvt.u64.u32 %rd32, %r122;
|
|
mov.u64 %rd41, uvpos;
|
|
cvta.global.u64 %rd30, %rd41;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r101, %r117, %rd31, %rd32, %rd27, %rd27);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd29+8];
|
|
ld.f32 %f8, [%rd29+4];
|
|
ld.f32 %f7, [%rd29];
|
|
mul.f32 %f364, %f7, 0f3456BF95;
|
|
mul.f32 %f365, %f8, 0f3456BF95;
|
|
mul.f32 %f366, %f9, 0f3456BF95;
|
|
abs.f32 %f367, %f1126;
|
|
div.rn.f32 %f368, %f364, %f367;
|
|
abs.f32 %f369, %f1127;
|
|
div.rn.f32 %f370, %f365, %f369;
|
|
abs.f32 %f371, %f1128;
|
|
div.rn.f32 %f372, %f366, %f371;
|
|
abs.f32 %f373, %f368;
|
|
abs.f32 %f374, %f370;
|
|
abs.f32 %f375, %f372;
|
|
mov.f32 %f376, 0f38D1B717;
|
|
max.f32 %f377, %f373, %f376;
|
|
max.f32 %f378, %f374, %f376;
|
|
max.f32 %f379, %f375, %f376;
|
|
fma.rn.f32 %f10, %f1126, %f377, %f7;
|
|
fma.rn.f32 %f11, %f1127, %f378, %f8;
|
|
fma.rn.f32 %f12, %f1128, %f379, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd37, %r125;
|
|
cvt.u64.u32 %rd38, %r126;
|
|
mov.u64 %rd42, rnd_seeds;
|
|
cvta.global.u64 %rd36, %rd42;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd36, %r101, %r102, %rd37, %rd38, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r393, [%rd35];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r389, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f13, [lightPointSize];
|
|
mul.f32 %f14, %f10, 0f3456BF95;
|
|
mul.f32 %f15, %f11, 0f3456BF95;
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r389, 0;
|
|
abs.f32 %f519, %f15;
|
|
abs.f32 %f520, %f14;
|
|
max.f32 %f521, %f520, %f519;
|
|
abs.f32 %f522, %f16;
|
|
max.f32 %f523, %f521, %f522;
|
|
mov.f32 %f18, %f17;
|
|
mov.f32 %f19, %f17;
|
|
mov.f32 %f20, %f17;
|
|
mov.f32 %f21, %f17;
|
|
mov.f32 %f22, %f17;
|
|
mov.f32 %f23, %f17;
|
|
mov.f32 %f24, %f17;
|
|
mov.f32 %f25, %f17;
|
|
mov.f32 %f26, %f17;
|
|
mov.f32 %f27, %f17;
|
|
mov.f32 %f28, %f17;
|
|
mov.f32 %f29, %f17;
|
|
mov.f32 %f30, %f17;
|
|
mov.f32 %f31, %f17;
|
|
mov.f32 %f32, %f17;
|
|
|
|
BB0_6:
|
|
mul.lo.s32 %r10, %r389, 3;
|
|
cvt.s64.s32 %rd45, %r10;
|
|
mov.u64 %rd55, lightMeshBuffer;
|
|
cvta.global.u64 %rd44, %rd55;
|
|
mov.u32 %r130, 1;
|
|
mov.u64 %rd54, 0;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r130, %r117, %rd45, %rd54, %rd54, %rd54);
|
|
// inline asm
|
|
ld.f32 %f396, [%rd43];
|
|
sub.f32 %f397, %f396, %f7;
|
|
ld.f32 %f398, [%rd43+4];
|
|
sub.f32 %f399, %f398, %f8;
|
|
ld.f32 %f400, [%rd43+8];
|
|
sub.f32 %f401, %f400, %f9;
|
|
mul.f32 %f402, %f399, %f399;
|
|
fma.rn.f32 %f403, %f397, %f397, %f402;
|
|
fma.rn.f32 %f404, %f401, %f401, %f403;
|
|
sqrt.rn.f32 %f33, %f404;
|
|
rcp.rn.f32 %f405, %f33;
|
|
mul.f32 %f34, %f397, %f405;
|
|
mul.f32 %f35, %f399, %f405;
|
|
mul.f32 %f36, %f401, %f405;
|
|
mul.f32 %f406, %f33, %f33;
|
|
div.rn.f32 %f37, %f13, %f406;
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
|
cvt.u64.u32 %rd56, %r135;
|
|
cvt.u64.u32 %rd57, %r134;
|
|
and.b64 %rd51, %rd57, 255;
|
|
and.b64 %rd52, %rd56, 255;
|
|
mov.u64 %rd58, rnd_seeds2;
|
|
cvta.global.u64 %rd50, %rd58;
|
|
mov.u32 %r133, 3;
|
|
// inline asm
|
|
call (%rd49), _rt_buffer_get_64, (%rd50, %r101, %r133, %rd51, %rd52, %rd54, %rd54);
|
|
// inline asm
|
|
setp.eq.s32 %p15, %r3, 0;
|
|
@%p15 bra BB0_8;
|
|
|
|
ld.u8 %rs14, [%rd49];
|
|
cvt.rn.f32.u16 %f407, %rs14;
|
|
div.rn.f32 %f408, %f407, 0fC37F0000;
|
|
fma.rn.f32 %f409, %f408, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p17, %f37, %f409;
|
|
mov.pred %p148, -1;
|
|
@%p17 bra BB0_23;
|
|
|
|
BB0_8:
|
|
ld.global.f32 %f412, [lightInvCutoff];
|
|
mul.f32 %f38, %f33, %f412;
|
|
mov.f32 %f416, 0f40800000;
|
|
abs.f32 %f40, %f38;
|
|
setp.lt.f32 %p18, %f40, 0f00800000;
|
|
mul.f32 %f418, %f40, 0f4B800000;
|
|
selp.f32 %f419, 0fC3170000, 0fC2FE0000, %p18;
|
|
selp.f32 %f420, %f418, %f40, %p18;
|
|
mov.b32 %r138, %f420;
|
|
and.b32 %r139, %r138, 8388607;
|
|
or.b32 %r140, %r139, 1065353216;
|
|
mov.b32 %f421, %r140;
|
|
shr.u32 %r141, %r138, 23;
|
|
cvt.rn.f32.u32 %f422, %r141;
|
|
add.f32 %f423, %f419, %f422;
|
|
setp.gt.f32 %p19, %f421, 0f3FB504F3;
|
|
mul.f32 %f424, %f421, 0f3F000000;
|
|
add.f32 %f425, %f423, 0f3F800000;
|
|
selp.f32 %f426, %f424, %f421, %p19;
|
|
selp.f32 %f427, %f425, %f423, %p19;
|
|
add.f32 %f428, %f426, 0fBF800000;
|
|
add.f32 %f411, %f426, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f410,%f411;
|
|
// inline asm
|
|
add.f32 %f429, %f428, %f428;
|
|
mul.f32 %f430, %f410, %f429;
|
|
mul.f32 %f431, %f430, %f430;
|
|
mov.f32 %f432, 0f3C4CAF63;
|
|
mov.f32 %f433, 0f3B18F0FE;
|
|
fma.rn.f32 %f434, %f433, %f431, %f432;
|
|
mov.f32 %f435, 0f3DAAAABD;
|
|
fma.rn.f32 %f436, %f434, %f431, %f435;
|
|
mul.rn.f32 %f437, %f436, %f431;
|
|
mul.rn.f32 %f438, %f437, %f430;
|
|
sub.f32 %f439, %f428, %f430;
|
|
neg.f32 %f440, %f430;
|
|
add.f32 %f441, %f439, %f439;
|
|
fma.rn.f32 %f442, %f440, %f428, %f441;
|
|
mul.rn.f32 %f443, %f410, %f442;
|
|
add.f32 %f444, %f438, %f430;
|
|
sub.f32 %f445, %f430, %f444;
|
|
add.f32 %f446, %f438, %f445;
|
|
add.f32 %f447, %f443, %f446;
|
|
add.f32 %f448, %f444, %f447;
|
|
sub.f32 %f449, %f444, %f448;
|
|
add.f32 %f450, %f447, %f449;
|
|
mov.f32 %f451, 0f3F317200;
|
|
mul.rn.f32 %f452, %f427, %f451;
|
|
mov.f32 %f453, 0f35BFBE8E;
|
|
mul.rn.f32 %f454, %f427, %f453;
|
|
add.f32 %f455, %f452, %f448;
|
|
sub.f32 %f456, %f452, %f455;
|
|
add.f32 %f457, %f448, %f456;
|
|
add.f32 %f458, %f450, %f457;
|
|
add.f32 %f459, %f454, %f458;
|
|
add.f32 %f460, %f455, %f459;
|
|
sub.f32 %f461, %f455, %f460;
|
|
add.f32 %f462, %f459, %f461;
|
|
mul.rn.f32 %f463, %f416, %f460;
|
|
neg.f32 %f464, %f463;
|
|
fma.rn.f32 %f465, %f416, %f460, %f464;
|
|
fma.rn.f32 %f466, %f416, %f462, %f465;
|
|
mov.f32 %f467, 0f00000000;
|
|
fma.rn.f32 %f468, %f467, %f460, %f466;
|
|
add.rn.f32 %f469, %f463, %f468;
|
|
neg.f32 %f470, %f469;
|
|
add.rn.f32 %f471, %f463, %f470;
|
|
add.rn.f32 %f472, %f471, %f468;
|
|
mov.b32 %r142, %f469;
|
|
setp.eq.s32 %p20, %r142, 1118925336;
|
|
add.s32 %r143, %r142, -1;
|
|
mov.b32 %f473, %r143;
|
|
add.f32 %f474, %f472, 0f37000000;
|
|
selp.f32 %f475, %f473, %f469, %p20;
|
|
selp.f32 %f41, %f474, %f472, %p20;
|
|
mul.f32 %f476, %f475, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f477, %f476;
|
|
mov.f32 %f478, 0fBF317200;
|
|
fma.rn.f32 %f479, %f477, %f478, %f475;
|
|
mov.f32 %f480, 0fB5BFBE8E;
|
|
fma.rn.f32 %f481, %f477, %f480, %f479;
|
|
mul.f32 %f482, %f481, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f483, %f482;
|
|
add.f32 %f484, %f477, 0f00000000;
|
|
ex2.approx.f32 %f485, %f484;
|
|
mul.f32 %f486, %f483, %f485;
|
|
setp.lt.f32 %p21, %f475, 0fC2D20000;
|
|
selp.f32 %f487, 0f00000000, %f486, %p21;
|
|
setp.gt.f32 %p22, %f475, 0f42D20000;
|
|
selp.f32 %f1145, 0f7F800000, %f487, %p22;
|
|
setp.eq.f32 %p23, %f1145, 0f7F800000;
|
|
@%p23 bra BB0_10;
|
|
|
|
fma.rn.f32 %f1145, %f1145, %f41, %f1145;
|
|
|
|
BB0_10:
|
|
mov.f32 %f1124, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1123, %f1124;
|
|
add.f32 %f1122, %f1123, %f1123;
|
|
mov.f32 %f1121, 0f40800000;
|
|
sub.f32 %f1120, %f1121, %f1122;
|
|
abs.f32 %f1119, %f1120;
|
|
setp.lt.f32 %p24, %f38, 0f00000000;
|
|
setp.eq.f32 %p25, %f1119, 0f3F800000;
|
|
and.pred %p1, %p24, %p25;
|
|
mov.b32 %r144, %f1145;
|
|
xor.b32 %r145, %r144, -2147483648;
|
|
mov.b32 %f488, %r145;
|
|
selp.f32 %f1147, %f488, %f1145, %p1;
|
|
setp.eq.f32 %p26, %f38, 0f00000000;
|
|
@%p26 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f491, %f38, %f38;
|
|
selp.f32 %f1147, %f491, 0f00000000, %p25;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p27, %f38, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
|
|
mov.f32 %f1125, 0f40800000;
|
|
cvt.rzi.f32.f32 %f490, %f1125;
|
|
setp.neu.f32 %p28, %f490, 0f40800000;
|
|
selp.f32 %f1147, 0f7FFFFFFF, %f1147, %p28;
|
|
|
|
BB0_14:
|
|
add.f32 %f492, %f40, 0f40800000;
|
|
mov.b32 %r146, %f492;
|
|
setp.lt.s32 %p30, %r146, 2139095040;
|
|
@%p30 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p31, %f40, 0f7F800000;
|
|
@%p31 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f1147, %f38, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p32, %f40, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
|
|
selp.f32 %f1147, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u64 %rd320, 0;
|
|
mov.u32 %r384, 1;
|
|
mov.u64 %rd319, lightMeshBuffer;
|
|
cvta.global.u64 %rd318, %rd319;
|
|
mul.lo.s32 %r383, %r389, 3;
|
|
mov.f32 %f493, 0f3F800000;
|
|
sub.f32 %f494, %f493, %f1147;
|
|
setp.eq.f32 %p33, %f38, 0f3F800000;
|
|
selp.f32 %f495, 0f00000000, %f494, %p33;
|
|
cvt.sat.f32.f32 %f496, %f495;
|
|
mul.f32 %f497, %f37, %f496;
|
|
add.s32 %r149, %r383, 1;
|
|
cvt.s64.s32 %rd61, %r149;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd318, %r384, %r117, %rd61, %rd320, %rd320, %rd320);
|
|
// inline asm
|
|
ld.f32 %f498, [%rd59];
|
|
mul.f32 %f499, %f34, %f498;
|
|
ld.f32 %f500, [%rd59+4];
|
|
mul.f32 %f501, %f35, %f500;
|
|
neg.f32 %f502, %f501;
|
|
sub.f32 %f503, %f502, %f499;
|
|
ld.f32 %f504, [%rd59+8];
|
|
mul.f32 %f505, %f36, %f504;
|
|
sub.f32 %f506, %f503, %f505;
|
|
cvt.sat.f32.f32 %f507, %f506;
|
|
mul.f32 %f52, %f497, %f507;
|
|
mul.f32 %f508, %f1127, %f35;
|
|
fma.rn.f32 %f509, %f1126, %f34, %f508;
|
|
fma.rn.f32 %f53, %f1128, %f36, %f509;
|
|
setp.leu.f32 %p34, %f52, 0f3727C5AC;
|
|
@%p34 bra BB0_21;
|
|
|
|
mov.u64 %rd323, 0;
|
|
mov.u32 %r387, 1;
|
|
mov.u64 %rd322, lightMeshBuffer;
|
|
cvta.global.u64 %rd321, %rd322;
|
|
mul.lo.s32 %r386, %r389, 3;
|
|
cvt.sat.f32.f32 %f518, %f53;
|
|
add.u64 %rd66, %SP, 16;
|
|
cvta.to.local.u64 %rd73, %rd66;
|
|
max.f32 %f516, %f523, %f376;
|
|
sub.f32 %f517, %f33, %f516;
|
|
mov.u32 %r155, 1065353216;
|
|
st.local.u32 [%rd73+8], %r155;
|
|
st.local.u32 [%rd73+4], %r155;
|
|
st.local.u32 [%rd73], %r155;
|
|
ld.global.u32 %r150, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r150, %f10, %f11, %f12, %f34, %f35, %f36, %r387, %f516, %f517, %rd66, %r117);
|
|
// inline asm
|
|
add.s32 %r156, %r386, 2;
|
|
cvt.s64.s32 %rd69, %r156;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd321, %r387, %r117, %rd69, %rd323, %rd323, %rd323);
|
|
// inline asm
|
|
ld.local.f32 %f525, [%rd73];
|
|
ld.f32 %f526, [%rd67];
|
|
mul.f32 %f527, %f526, %f525;
|
|
ld.local.f32 %f528, [%rd73+4];
|
|
ld.f32 %f529, [%rd67+4];
|
|
mul.f32 %f530, %f529, %f528;
|
|
ld.local.f32 %f531, [%rd73+8];
|
|
ld.f32 %f532, [%rd67+8];
|
|
mul.f32 %f533, %f532, %f531;
|
|
mul.f32 %f534, %f52, 0f3EA2F983;
|
|
mul.f32 %f535, %f534, %f527;
|
|
mul.f32 %f536, %f534, %f530;
|
|
mul.f32 %f537, %f534, %f533;
|
|
fma.rn.f32 %f20, %f518, %f535, %f20;
|
|
fma.rn.f32 %f19, %f518, %f536, %f19;
|
|
fma.rn.f32 %f18, %f518, %f537, %f18;
|
|
fma.rn.f32 %f26, %f34, %f535, %f26;
|
|
fma.rn.f32 %f25, %f34, %f536, %f25;
|
|
fma.rn.f32 %f24, %f34, %f537, %f24;
|
|
fma.rn.f32 %f29, %f35, %f535, %f29;
|
|
fma.rn.f32 %f28, %f35, %f536, %f28;
|
|
fma.rn.f32 %f27, %f35, %f537, %f27;
|
|
fma.rn.f32 %f32, %f36, %f535, %f32;
|
|
fma.rn.f32 %f31, %f36, %f536, %f31;
|
|
fma.rn.f32 %f30, %f36, %f537, %f30;
|
|
add.f32 %f23, %f23, %f535;
|
|
add.f32 %f22, %f22, %f536;
|
|
add.f32 %f21, %f21, %f537;
|
|
add.f32 %f17, %f17, %f525;
|
|
|
|
BB0_21:
|
|
add.s32 %r389, %r389, 1;
|
|
setp.lt.s32 %p36, %r389, %r2;
|
|
@%p36 bra BB0_6;
|
|
bra.uni BB0_22;
|
|
|
|
BB0_4:
|
|
mov.f32 %f18, %f17;
|
|
mov.f32 %f19, %f17;
|
|
mov.f32 %f20, %f17;
|
|
mov.f32 %f21, %f17;
|
|
mov.f32 %f22, %f17;
|
|
mov.f32 %f23, %f17;
|
|
mov.f32 %f24, %f17;
|
|
mov.f32 %f25, %f17;
|
|
mov.f32 %f26, %f17;
|
|
mov.f32 %f27, %f17;
|
|
mov.f32 %f28, %f17;
|
|
mov.f32 %f29, %f17;
|
|
mov.f32 %f30, %f17;
|
|
mov.f32 %f31, %f17;
|
|
mov.f32 %f32, %f17;
|
|
|
|
BB0_22:
|
|
mov.pred %p148, %p13;
|
|
|
|
BB0_23:
|
|
cvt.rn.f32.s32 %f538, %r389;
|
|
mov.f32 %f539, 0f3F800000;
|
|
max.f32 %f540, %f538, %f539;
|
|
rcp.rn.f32 %f541, %f540;
|
|
mul.f32 %f1249, %f20, %f541;
|
|
mul.f32 %f1250, %f19, %f541;
|
|
mul.f32 %f1251, %f18, %f541;
|
|
div.rn.f32 %f1252, %f17, %f540;
|
|
mul.f32 %f1246, %f23, %f541;
|
|
mul.f32 %f1247, %f22, %f541;
|
|
mul.f32 %f1248, %f21, %f541;
|
|
mul.f32 %f1243, %f26, %f541;
|
|
mul.f32 %f1244, %f25, %f541;
|
|
mul.f32 %f1245, %f24, %f541;
|
|
mul.f32 %f1240, %f29, %f541;
|
|
mul.f32 %f1241, %f28, %f541;
|
|
mul.f32 %f1242, %f27, %f541;
|
|
mul.f32 %f1237, %f32, %f541;
|
|
mul.f32 %f1238, %f31, %f541;
|
|
mul.f32 %f1239, %f30, %f541;
|
|
@!%p148 bra BB0_76;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_24:
|
|
mov.f32 %f556, 0f00000000;
|
|
setp.lt.s32 %p37, %r3, 1;
|
|
mov.f32 %f1195, %f556;
|
|
mov.f32 %f1196, %f556;
|
|
mov.f32 %f1197, %f556;
|
|
mov.f32 %f1198, %f556;
|
|
mov.f32 %f1199, %f556;
|
|
mov.f32 %f1200, %f556;
|
|
mov.f32 %f1201, %f556;
|
|
mov.f32 %f1202, %f556;
|
|
mov.f32 %f1203, %f556;
|
|
mov.f32 %f1204, %f556;
|
|
mov.f32 %f1205, %f556;
|
|
mov.f32 %f1206, %f556;
|
|
mov.f32 %f1207, %f556;
|
|
mov.f32 %f1208, %f556;
|
|
mov.f32 %f1209, %f556;
|
|
@%p37 bra BB0_75;
|
|
|
|
cvt.rn.f32.s32 %f572, %r4;
|
|
rcp.rn.f32 %f118, %f572;
|
|
add.u64 %rd75, %SP, 28;
|
|
cvta.to.local.u64 %rd3, %rd75;
|
|
mul.f32 %f119, %f10, 0f3456BF95;
|
|
mul.f32 %f120, %f11, 0f3456BF95;
|
|
mul.f32 %f121, %f12, 0f3456BF95;
|
|
add.u64 %rd76, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd76;
|
|
mov.f32 %f1195, 0f00000000;
|
|
mov.u32 %r157, 0;
|
|
abs.f32 %f573, %f120;
|
|
abs.f32 %f574, %f119;
|
|
max.f32 %f575, %f574, %f573;
|
|
abs.f32 %f576, %f121;
|
|
max.f32 %f577, %f575, %f576;
|
|
mov.u32 %r390, %r157;
|
|
mov.f32 %f1196, %f1195;
|
|
mov.f32 %f1197, %f1195;
|
|
mov.f32 %f1198, %f1195;
|
|
mov.f32 %f1199, %f1195;
|
|
mov.f32 %f1200, %f1195;
|
|
mov.f32 %f1201, %f1195;
|
|
mov.f32 %f1202, %f1195;
|
|
mov.f32 %f1203, %f1195;
|
|
mov.f32 %f1204, %f1195;
|
|
mov.f32 %f1205, %f1195;
|
|
mov.f32 %f1206, %f1195;
|
|
mov.f32 %f1207, %f1195;
|
|
mov.f32 %f1208, %f1195;
|
|
mov.f32 %f1209, %f1195;
|
|
|
|
BB0_26:
|
|
cvt.rn.f32.s32 %f137, %r390;
|
|
max.f32 %f138, %f577, %f376;
|
|
mov.u32 %r392, %r157;
|
|
|
|
BB0_27:
|
|
mad.lo.s32 %r159, %r393, 1664525, 1013904223;
|
|
and.b32 %r160, %r159, 16777215;
|
|
cvt.rn.f32.u32 %f579, %r160;
|
|
fma.rn.f32 %f580, %f579, 0f33800000, %f137;
|
|
mul.f32 %f581, %f118, %f580;
|
|
mad.lo.s32 %r393, %r159, 1664525, 1013904223;
|
|
and.b32 %r161, %r393, 16777215;
|
|
cvt.rn.f32.u32 %f582, %r161;
|
|
cvt.rn.f32.s32 %f583, %r392;
|
|
fma.rn.f32 %f584, %f582, 0f33800000, %f583;
|
|
mul.f32 %f585, %f118, %f584;
|
|
fma.rn.f32 %f154, %f581, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f586, %f154, %f154;
|
|
sub.f32 %f588, %f539, %f586;
|
|
mov.f32 %f589, 0f00000000;
|
|
max.f32 %f590, %f589, %f588;
|
|
sqrt.rn.f32 %f155, %f590;
|
|
mul.f32 %f1216, %f585, 0f40C90FDB;
|
|
abs.f32 %f157, %f1216;
|
|
setp.neu.f32 %p38, %f157, 0f7F800000;
|
|
mov.f32 %f1210, %f1216;
|
|
@%p38 bra BB0_29;
|
|
|
|
mul.rn.f32 %f1210, %f1216, %f589;
|
|
|
|
BB0_29:
|
|
mul.f32 %f592, %f1210, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r403, %f592;
|
|
cvt.rn.f32.s32 %f593, %r403;
|
|
neg.f32 %f594, %f593;
|
|
mov.f32 %f595, 0f3FC90FDA;
|
|
fma.rn.f32 %f596, %f594, %f595, %f1210;
|
|
mov.f32 %f597, 0f33A22168;
|
|
fma.rn.f32 %f598, %f594, %f597, %f596;
|
|
mov.f32 %f599, 0f27C234C5;
|
|
fma.rn.f32 %f1211, %f594, %f599, %f598;
|
|
abs.f32 %f600, %f1210;
|
|
setp.leu.f32 %p39, %f600, 0f47CE4780;
|
|
@%p39 bra BB0_40;
|
|
|
|
mov.b32 %r19, %f1210;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r164, %r19, 8;
|
|
or.b32 %r21, %r164, -2147483648;
|
|
mov.u32 %r394, 0;
|
|
mov.u64 %rd325, 0;
|
|
mov.u64 %rd324, %rd3;
|
|
mov.u32 %r395, %r394;
|
|
|
|
BB0_31:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd325, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r167, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r165, %r167, %r21, %r395;
|
|
madc.hi.u32 %r395, %r167, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd324], %r165;
|
|
add.s32 %r394, %r394, 1;
|
|
cvt.s64.s32 %rd325, %r394;
|
|
mul.wide.s32 %rd81, %r394, 4;
|
|
add.s64 %rd324, %rd3, %rd81;
|
|
setp.ne.s32 %p40, %r394, 6;
|
|
@%p40 bra BB0_31;
|
|
|
|
and.b32 %r170, %r20, 255;
|
|
add.s32 %r171, %r170, -128;
|
|
shr.u32 %r172, %r171, 5;
|
|
and.b32 %r26, %r19, -2147483648;
|
|
st.local.u32 [%rd3+24], %r395;
|
|
mov.u32 %r173, 6;
|
|
sub.s32 %r174, %r173, %r172;
|
|
mul.wide.s32 %rd82, %r174, 4;
|
|
add.s64 %rd10, %rd3, %rd82;
|
|
ld.local.u32 %r396, [%rd10];
|
|
ld.local.u32 %r397, [%rd10+-4];
|
|
and.b32 %r29, %r20, 31;
|
|
setp.eq.s32 %p41, %r29, 0;
|
|
@%p41 bra BB0_34;
|
|
|
|
mov.u32 %r175, 32;
|
|
sub.s32 %r176, %r175, %r29;
|
|
shr.u32 %r177, %r397, %r176;
|
|
shl.b32 %r178, %r396, %r29;
|
|
add.s32 %r396, %r177, %r178;
|
|
ld.local.u32 %r179, [%rd10+-8];
|
|
shr.u32 %r180, %r179, %r176;
|
|
shl.b32 %r181, %r397, %r29;
|
|
add.s32 %r397, %r180, %r181;
|
|
|
|
BB0_34:
|
|
shr.u32 %r182, %r397, 30;
|
|
shl.b32 %r183, %r396, 2;
|
|
add.s32 %r398, %r182, %r183;
|
|
shl.b32 %r35, %r397, 2;
|
|
shr.u32 %r184, %r398, 31;
|
|
shr.u32 %r185, %r396, 30;
|
|
add.s32 %r36, %r184, %r185;
|
|
setp.eq.s32 %p42, %r184, 0;
|
|
@%p42 bra BB0_35;
|
|
bra.uni BB0_36;
|
|
|
|
BB0_35:
|
|
mov.u32 %r399, %r26;
|
|
mov.u32 %r400, %r35;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
not.b32 %r186, %r398;
|
|
neg.s32 %r400, %r35;
|
|
setp.eq.s32 %p43, %r35, 0;
|
|
selp.u32 %r187, 1, 0, %p43;
|
|
add.s32 %r398, %r187, %r186;
|
|
xor.b32 %r399, %r26, -2147483648;
|
|
|
|
BB0_37:
|
|
clz.b32 %r402, %r398;
|
|
setp.eq.s32 %p44, %r402, 0;
|
|
shl.b32 %r188, %r398, %r402;
|
|
mov.u32 %r189, 32;
|
|
sub.s32 %r190, %r189, %r402;
|
|
shr.u32 %r191, %r400, %r190;
|
|
add.s32 %r192, %r191, %r188;
|
|
selp.b32 %r44, %r398, %r192, %p44;
|
|
mov.u32 %r193, -921707870;
|
|
mul.hi.u32 %r401, %r44, %r193;
|
|
setp.eq.s32 %p45, %r26, 0;
|
|
neg.s32 %r194, %r36;
|
|
selp.b32 %r403, %r36, %r194, %p45;
|
|
setp.lt.s32 %p46, %r401, 1;
|
|
@%p46 bra BB0_39;
|
|
|
|
mul.lo.s32 %r195, %r44, -921707870;
|
|
shr.u32 %r196, %r195, 31;
|
|
shl.b32 %r197, %r401, 1;
|
|
add.s32 %r401, %r196, %r197;
|
|
add.s32 %r402, %r402, 1;
|
|
|
|
BB0_39:
|
|
mov.u32 %r198, 126;
|
|
sub.s32 %r199, %r198, %r402;
|
|
shl.b32 %r200, %r199, 23;
|
|
add.s32 %r201, %r401, 1;
|
|
shr.u32 %r202, %r201, 7;
|
|
add.s32 %r203, %r202, 1;
|
|
shr.u32 %r204, %r203, 1;
|
|
add.s32 %r205, %r204, %r200;
|
|
or.b32 %r206, %r205, %r399;
|
|
mov.b32 %f1211, %r206;
|
|
|
|
BB0_40:
|
|
mul.rn.f32 %f163, %f1211, %f1211;
|
|
add.s32 %r52, %r403, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p47, %r53, 0;
|
|
@%p47 bra BB0_42;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_42:
|
|
mov.f32 %f603, 0f3C08839E;
|
|
mov.f32 %f604, 0fB94CA1F9;
|
|
fma.rn.f32 %f1212, %f604, %f163, %f603;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_41:
|
|
mov.f32 %f601, 0fBAB6061A;
|
|
mov.f32 %f602, 0f37CCF5CE;
|
|
fma.rn.f32 %f1212, %f602, %f163, %f601;
|
|
|
|
BB0_43:
|
|
@%p47 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f608, 0fBE2AAAA3;
|
|
fma.rn.f32 %f609, %f1212, %f163, %f608;
|
|
fma.rn.f32 %f1213, %f609, %f163, %f589;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f605, 0f3D2AAAA5;
|
|
fma.rn.f32 %f606, %f1212, %f163, %f605;
|
|
mov.f32 %f607, 0fBF000000;
|
|
fma.rn.f32 %f1213, %f606, %f163, %f607;
|
|
|
|
BB0_46:
|
|
fma.rn.f32 %f1214, %f1213, %f1211, %f1211;
|
|
@%p47 bra BB0_48;
|
|
|
|
fma.rn.f32 %f1214, %f1213, %f163, %f539;
|
|
|
|
BB0_48:
|
|
and.b32 %r207, %r52, 2;
|
|
setp.eq.s32 %p50, %r207, 0;
|
|
@%p50 bra BB0_50;
|
|
|
|
mov.f32 %f613, 0fBF800000;
|
|
fma.rn.f32 %f1214, %f1214, %f613, %f589;
|
|
|
|
BB0_50:
|
|
@%p38 bra BB0_52;
|
|
|
|
mul.rn.f32 %f1216, %f1216, %f589;
|
|
|
|
BB0_52:
|
|
mul.f32 %f615, %f1216, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r413, %f615;
|
|
cvt.rn.f32.s32 %f616, %r413;
|
|
neg.f32 %f617, %f616;
|
|
fma.rn.f32 %f619, %f617, %f595, %f1216;
|
|
fma.rn.f32 %f621, %f617, %f597, %f619;
|
|
fma.rn.f32 %f1217, %f617, %f599, %f621;
|
|
abs.f32 %f623, %f1216;
|
|
setp.leu.f32 %p52, %f623, 0f47CE4780;
|
|
@%p52 bra BB0_63;
|
|
|
|
mov.b32 %r55, %f1216;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r210, %r55, 8;
|
|
or.b32 %r57, %r210, -2147483648;
|
|
mov.u32 %r404, 0;
|
|
mov.u64 %rd326, %rd3;
|
|
mov.u64 %rd327, %rd27;
|
|
mov.u32 %r405, %r404;
|
|
|
|
BB0_54:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd84, %rd327, 2;
|
|
mov.u64 %rd85, __cudart_i2opi_f;
|
|
add.s64 %rd86, %rd85, %rd84;
|
|
ld.const.u32 %r213, [%rd86];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r211, %r213, %r57, %r405;
|
|
madc.hi.u32 %r405, %r213, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd326], %r211;
|
|
add.s32 %r404, %r404, 1;
|
|
cvt.s64.s32 %rd327, %r404;
|
|
mul.wide.s32 %rd87, %r404, 4;
|
|
add.s64 %rd326, %rd3, %rd87;
|
|
setp.ne.s32 %p53, %r404, 6;
|
|
@%p53 bra BB0_54;
|
|
|
|
and.b32 %r216, %r56, 255;
|
|
add.s32 %r217, %r216, -128;
|
|
shr.u32 %r218, %r217, 5;
|
|
and.b32 %r62, %r55, -2147483648;
|
|
st.local.u32 [%rd3+24], %r405;
|
|
mov.u32 %r219, 6;
|
|
sub.s32 %r220, %r219, %r218;
|
|
mul.wide.s32 %rd88, %r220, 4;
|
|
add.s64 %rd16, %rd3, %rd88;
|
|
ld.local.u32 %r406, [%rd16];
|
|
ld.local.u32 %r407, [%rd16+-4];
|
|
and.b32 %r65, %r56, 31;
|
|
setp.eq.s32 %p54, %r65, 0;
|
|
@%p54 bra BB0_57;
|
|
|
|
mov.u32 %r221, 32;
|
|
sub.s32 %r222, %r221, %r65;
|
|
shr.u32 %r223, %r407, %r222;
|
|
shl.b32 %r224, %r406, %r65;
|
|
add.s32 %r406, %r223, %r224;
|
|
ld.local.u32 %r225, [%rd16+-8];
|
|
shr.u32 %r226, %r225, %r222;
|
|
shl.b32 %r227, %r407, %r65;
|
|
add.s32 %r407, %r226, %r227;
|
|
|
|
BB0_57:
|
|
shr.u32 %r228, %r407, 30;
|
|
shl.b32 %r229, %r406, 2;
|
|
add.s32 %r408, %r228, %r229;
|
|
shl.b32 %r71, %r407, 2;
|
|
shr.u32 %r230, %r408, 31;
|
|
shr.u32 %r231, %r406, 30;
|
|
add.s32 %r72, %r230, %r231;
|
|
setp.eq.s32 %p55, %r230, 0;
|
|
@%p55 bra BB0_58;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_58:
|
|
mov.u32 %r409, %r62;
|
|
mov.u32 %r410, %r71;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
not.b32 %r232, %r408;
|
|
neg.s32 %r410, %r71;
|
|
setp.eq.s32 %p56, %r71, 0;
|
|
selp.u32 %r233, 1, 0, %p56;
|
|
add.s32 %r408, %r233, %r232;
|
|
xor.b32 %r409, %r62, -2147483648;
|
|
|
|
BB0_60:
|
|
clz.b32 %r412, %r408;
|
|
setp.eq.s32 %p57, %r412, 0;
|
|
shl.b32 %r234, %r408, %r412;
|
|
mov.u32 %r235, 32;
|
|
sub.s32 %r236, %r235, %r412;
|
|
shr.u32 %r237, %r410, %r236;
|
|
add.s32 %r238, %r237, %r234;
|
|
selp.b32 %r80, %r408, %r238, %p57;
|
|
mov.u32 %r239, -921707870;
|
|
mul.hi.u32 %r411, %r80, %r239;
|
|
setp.eq.s32 %p58, %r62, 0;
|
|
neg.s32 %r240, %r72;
|
|
selp.b32 %r413, %r72, %r240, %p58;
|
|
setp.lt.s32 %p59, %r411, 1;
|
|
@%p59 bra BB0_62;
|
|
|
|
mul.lo.s32 %r241, %r80, -921707870;
|
|
shr.u32 %r242, %r241, 31;
|
|
shl.b32 %r243, %r411, 1;
|
|
add.s32 %r411, %r242, %r243;
|
|
add.s32 %r412, %r412, 1;
|
|
|
|
BB0_62:
|
|
mov.u32 %r244, 126;
|
|
sub.s32 %r245, %r244, %r412;
|
|
shl.b32 %r246, %r245, 23;
|
|
add.s32 %r247, %r411, 1;
|
|
shr.u32 %r248, %r247, 7;
|
|
add.s32 %r249, %r248, 1;
|
|
shr.u32 %r250, %r249, 1;
|
|
add.s32 %r251, %r250, %r246;
|
|
or.b32 %r252, %r251, %r409;
|
|
mov.b32 %f1217, %r252;
|
|
|
|
BB0_63:
|
|
mul.rn.f32 %f180, %f1217, %f1217;
|
|
and.b32 %r88, %r413, 1;
|
|
setp.eq.s32 %p60, %r88, 0;
|
|
@%p60 bra BB0_65;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_65:
|
|
mov.f32 %f626, 0f3C08839E;
|
|
mov.f32 %f627, 0fB94CA1F9;
|
|
fma.rn.f32 %f1218, %f627, %f180, %f626;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_64:
|
|
mov.f32 %f624, 0fBAB6061A;
|
|
mov.f32 %f625, 0f37CCF5CE;
|
|
fma.rn.f32 %f1218, %f625, %f180, %f624;
|
|
|
|
BB0_66:
|
|
@%p60 bra BB0_68;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_68:
|
|
mov.f32 %f631, 0fBE2AAAA3;
|
|
fma.rn.f32 %f632, %f1218, %f180, %f631;
|
|
fma.rn.f32 %f1219, %f632, %f180, %f589;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_67:
|
|
mov.f32 %f628, 0f3D2AAAA5;
|
|
fma.rn.f32 %f629, %f1218, %f180, %f628;
|
|
mov.f32 %f630, 0fBF000000;
|
|
fma.rn.f32 %f1219, %f629, %f180, %f630;
|
|
|
|
BB0_69:
|
|
fma.rn.f32 %f1220, %f1219, %f1217, %f1217;
|
|
@%p60 bra BB0_71;
|
|
|
|
fma.rn.f32 %f1220, %f1219, %f180, %f539;
|
|
|
|
BB0_71:
|
|
and.b32 %r253, %r413, 2;
|
|
setp.eq.s32 %p63, %r253, 0;
|
|
@%p63 bra BB0_73;
|
|
|
|
mov.f32 %f636, 0fBF800000;
|
|
fma.rn.f32 %f1220, %f1220, %f636, %f589;
|
|
|
|
BB0_73:
|
|
mul.f32 %f640, %f155, %f1214;
|
|
mov.f32 %f645, 0f7F7FFFFF;
|
|
st.local.v4.f32 [%rd4], {%f539, %f539, %f539, %f645};
|
|
ld.global.u32 %r254, [root];
|
|
mul.f32 %f641, %f155, %f1220;
|
|
mov.u32 %r255, 0;
|
|
mov.f32 %f644, 0f6C4ECB8F;
|
|
mov.u32 %r256, 16;
|
|
// inline asm
|
|
call _rt_trace_64, (%r254, %f10, %f11, %f12, %f640, %f641, %f154, %r255, %f138, %f644, %rd76, %r256);
|
|
// inline asm
|
|
ld.local.v4.f32 {%f647, %f648, %f649, %f650}, [%rd4];
|
|
max.f32 %f653, %f647, %f589;
|
|
max.f32 %f655, %f648, %f589;
|
|
max.f32 %f657, %f649, %f589;
|
|
setp.ltu.f32 %p64, %f650, 0f00000000;
|
|
selp.f32 %f659, %f657, 0f00000000, %p64;
|
|
selp.f32 %f660, %f655, 0f00000000, %p64;
|
|
selp.f32 %f661, %f653, 0f00000000, %p64;
|
|
fma.rn.f32 %f1203, %f640, %f661, %f1203;
|
|
fma.rn.f32 %f1202, %f640, %f660, %f1202;
|
|
fma.rn.f32 %f1201, %f640, %f659, %f1201;
|
|
fma.rn.f32 %f1206, %f641, %f661, %f1206;
|
|
fma.rn.f32 %f1205, %f641, %f660, %f1205;
|
|
fma.rn.f32 %f1204, %f641, %f659, %f1204;
|
|
fma.rn.f32 %f1209, %f154, %f661, %f1209;
|
|
fma.rn.f32 %f1208, %f154, %f660, %f1208;
|
|
fma.rn.f32 %f1207, %f154, %f659, %f1207;
|
|
add.f32 %f1200, %f1200, %f661;
|
|
add.f32 %f1199, %f1199, %f660;
|
|
add.f32 %f1198, %f1198, %f659;
|
|
mul.f32 %f662, %f1127, %f641;
|
|
fma.rn.f32 %f663, %f1126, %f640, %f662;
|
|
fma.rn.f32 %f664, %f1128, %f154, %f663;
|
|
cvt.sat.f32.f32 %f665, %f664;
|
|
fma.rn.f32 %f1197, %f665, %f661, %f1197;
|
|
fma.rn.f32 %f1196, %f665, %f660, %f1196;
|
|
fma.rn.f32 %f1195, %f665, %f659, %f1195;
|
|
add.s32 %r392, %r392, 1;
|
|
setp.lt.s32 %p65, %r392, %r4;
|
|
@%p65 bra BB0_27;
|
|
|
|
add.s32 %r390, %r390, 1;
|
|
setp.lt.s32 %p66, %r390, %r4;
|
|
@%p66 bra BB0_26;
|
|
|
|
BB0_75:
|
|
mul.lo.s32 %r257, %r4, %r4;
|
|
cvt.rn.f32.s32 %f666, %r257;
|
|
rcp.rn.f32 %f667, %f666;
|
|
mul.f32 %f668, %f1197, %f667;
|
|
mul.f32 %f669, %f1196, %f667;
|
|
mul.f32 %f670, %f1195, %f667;
|
|
div.rn.f32 %f1252, %f556, %f666;
|
|
mul.f32 %f1246, %f1200, %f667;
|
|
mul.f32 %f1247, %f1199, %f667;
|
|
mul.f32 %f1248, %f1198, %f667;
|
|
mul.f32 %f1243, %f1203, %f667;
|
|
mul.f32 %f1244, %f1202, %f667;
|
|
mul.f32 %f1245, %f1201, %f667;
|
|
mul.f32 %f1240, %f1206, %f667;
|
|
mul.f32 %f1241, %f1205, %f667;
|
|
mul.f32 %f1242, %f1204, %f667;
|
|
mul.f32 %f1237, %f1209, %f667;
|
|
mul.f32 %f1238, %f1208, %f667;
|
|
mul.f32 %f1239, %f1207, %f667;
|
|
fma.rn.f32 %f1249, %f1197, %f667, %f668;
|
|
fma.rn.f32 %f1250, %f1196, %f667, %f669;
|
|
fma.rn.f32 %f1251, %f1195, %f667, %f670;
|
|
|
|
BB0_76:
|
|
setp.gt.s32 %p67, %r1, -1;
|
|
selp.f32 %f254, %f1252, %f1249, %p67;
|
|
ld.global.u32 %r415, [imageEnabled];
|
|
and.b32 %r258, %r415, 8;
|
|
setp.eq.s32 %p68, %r258, 0;
|
|
@%p68 bra BB0_89;
|
|
|
|
cvt.u64.u32 %rd92, %r6;
|
|
cvt.u64.u32 %rd93, %r7;
|
|
mov.u64 %rd96, image_Mask;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r101, %r101, %rd92, %rd93, %rd27, %rd27);
|
|
// inline asm
|
|
mov.f32 %f674, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f675, %f674;
|
|
fma.rn.f32 %f676, %f675, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f255, %f676;
|
|
abs.f32 %f256, %f254;
|
|
setp.lt.f32 %p69, %f256, 0f00800000;
|
|
mul.f32 %f677, %f256, 0f4B800000;
|
|
selp.f32 %f678, 0fC3170000, 0fC2FE0000, %p69;
|
|
selp.f32 %f679, %f677, %f256, %p69;
|
|
mov.b32 %r261, %f679;
|
|
and.b32 %r262, %r261, 8388607;
|
|
or.b32 %r263, %r262, 1065353216;
|
|
mov.b32 %f680, %r263;
|
|
shr.u32 %r264, %r261, 23;
|
|
cvt.rn.f32.u32 %f681, %r264;
|
|
add.f32 %f682, %f678, %f681;
|
|
setp.gt.f32 %p70, %f680, 0f3FB504F3;
|
|
mul.f32 %f683, %f680, 0f3F000000;
|
|
add.f32 %f684, %f682, 0f3F800000;
|
|
selp.f32 %f685, %f683, %f680, %p70;
|
|
selp.f32 %f686, %f684, %f682, %p70;
|
|
add.f32 %f687, %f685, 0fBF800000;
|
|
add.f32 %f673, %f685, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f672,%f673;
|
|
// inline asm
|
|
add.f32 %f688, %f687, %f687;
|
|
mul.f32 %f689, %f672, %f688;
|
|
mul.f32 %f690, %f689, %f689;
|
|
mov.f32 %f691, 0f3C4CAF63;
|
|
mov.f32 %f692, 0f3B18F0FE;
|
|
fma.rn.f32 %f693, %f692, %f690, %f691;
|
|
mov.f32 %f694, 0f3DAAAABD;
|
|
fma.rn.f32 %f695, %f693, %f690, %f694;
|
|
mul.rn.f32 %f696, %f695, %f690;
|
|
mul.rn.f32 %f697, %f696, %f689;
|
|
sub.f32 %f698, %f687, %f689;
|
|
neg.f32 %f699, %f689;
|
|
add.f32 %f700, %f698, %f698;
|
|
fma.rn.f32 %f701, %f699, %f687, %f700;
|
|
mul.rn.f32 %f702, %f672, %f701;
|
|
add.f32 %f703, %f697, %f689;
|
|
sub.f32 %f704, %f689, %f703;
|
|
add.f32 %f705, %f697, %f704;
|
|
add.f32 %f706, %f702, %f705;
|
|
add.f32 %f707, %f703, %f706;
|
|
sub.f32 %f708, %f703, %f707;
|
|
add.f32 %f709, %f706, %f708;
|
|
mov.f32 %f710, 0f3F317200;
|
|
mul.rn.f32 %f711, %f686, %f710;
|
|
mov.f32 %f712, 0f35BFBE8E;
|
|
mul.rn.f32 %f713, %f686, %f712;
|
|
add.f32 %f714, %f711, %f707;
|
|
sub.f32 %f715, %f711, %f714;
|
|
add.f32 %f716, %f707, %f715;
|
|
add.f32 %f717, %f709, %f716;
|
|
add.f32 %f718, %f713, %f717;
|
|
add.f32 %f719, %f714, %f718;
|
|
sub.f32 %f720, %f714, %f719;
|
|
add.f32 %f721, %f718, %f720;
|
|
mov.f32 %f722, 0f3EE8BA2E;
|
|
mul.rn.f32 %f723, %f722, %f719;
|
|
neg.f32 %f724, %f723;
|
|
fma.rn.f32 %f725, %f722, %f719, %f724;
|
|
fma.rn.f32 %f726, %f722, %f721, %f725;
|
|
mov.f32 %f727, 0f00000000;
|
|
fma.rn.f32 %f728, %f727, %f719, %f726;
|
|
add.rn.f32 %f729, %f723, %f728;
|
|
neg.f32 %f730, %f729;
|
|
add.rn.f32 %f731, %f723, %f730;
|
|
add.rn.f32 %f732, %f731, %f728;
|
|
mov.b32 %r265, %f729;
|
|
setp.eq.s32 %p71, %r265, 1118925336;
|
|
add.s32 %r266, %r265, -1;
|
|
mov.b32 %f733, %r266;
|
|
add.f32 %f734, %f732, 0f37000000;
|
|
selp.f32 %f735, %f733, %f729, %p71;
|
|
selp.f32 %f257, %f734, %f732, %p71;
|
|
mul.f32 %f736, %f735, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f737, %f736;
|
|
mov.f32 %f738, 0fBF317200;
|
|
fma.rn.f32 %f739, %f737, %f738, %f735;
|
|
mov.f32 %f740, 0fB5BFBE8E;
|
|
fma.rn.f32 %f741, %f737, %f740, %f739;
|
|
mul.f32 %f742, %f741, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f743, %f742;
|
|
add.f32 %f744, %f737, 0f00000000;
|
|
ex2.approx.f32 %f745, %f744;
|
|
mul.f32 %f746, %f743, %f745;
|
|
setp.lt.f32 %p72, %f735, 0fC2D20000;
|
|
selp.f32 %f747, 0f00000000, %f746, %p72;
|
|
setp.gt.f32 %p73, %f735, 0f42D20000;
|
|
selp.f32 %f1253, 0f7F800000, %f747, %p73;
|
|
setp.eq.f32 %p74, %f1253, 0f7F800000;
|
|
@%p74 bra BB0_79;
|
|
|
|
fma.rn.f32 %f1253, %f1253, %f257, %f1253;
|
|
|
|
BB0_79:
|
|
setp.lt.f32 %p75, %f254, 0f00000000;
|
|
setp.eq.f32 %p76, %f255, 0f3F800000;
|
|
and.pred %p3, %p75, %p76;
|
|
mov.b32 %r267, %f1253;
|
|
xor.b32 %r268, %r267, -2147483648;
|
|
mov.b32 %f748, %r268;
|
|
selp.f32 %f1255, %f748, %f1253, %p3;
|
|
setp.eq.f32 %p77, %f254, 0f00000000;
|
|
@%p77 bra BB0_82;
|
|
bra.uni BB0_80;
|
|
|
|
BB0_82:
|
|
add.f32 %f751, %f254, %f254;
|
|
selp.f32 %f1255, %f751, 0f00000000, %p76;
|
|
bra.uni BB0_83;
|
|
|
|
BB0_147:
|
|
mov.u64 %rd237, image_HDR;
|
|
cvta.global.u64 %rd232, %rd237;
|
|
mov.u32 %r354, 8;
|
|
mov.u64 %rd236, 0;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r101, %r354, %rd20, %rd21, %rd236, %rd236);
|
|
// inline asm
|
|
mov.f32 %f1090, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1090;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd231], {%rs105, %rs105, %rs105, %rs106};
|
|
|
|
BB0_148:
|
|
ld.global.u32 %r355, [additive];
|
|
setp.eq.s32 %p144, %r355, 0;
|
|
@%p144 bra BB0_150;
|
|
|
|
mov.u64 %rd250, image_RNM0;
|
|
cvta.global.u64 %rd239, %rd250;
|
|
mov.u32 %r359, 8;
|
|
mov.u64 %rd249, 0;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r101, %r359, %rd20, %rd21, %rd249, %rd249);
|
|
// inline asm
|
|
ld.v4.u16 {%rs113, %rs114, %rs115, %rs116}, [%rd238];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1091, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1092, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1093, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd239, %r101, %r359, %rd20, %rd21, %rd249, %rd249);
|
|
// inline asm
|
|
add.f32 %f1094, %f1091, 0f00000000;
|
|
add.f32 %f1095, %f1092, 0f00000000;
|
|
add.f32 %f1096, %f1093, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1096;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1095;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1094;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs117, 0;
|
|
st.v4.u16 [%rd244], {%rs110, %rs111, %rs112, %rs117};
|
|
bra.uni BB0_151;
|
|
|
|
BB0_150:
|
|
mov.u64 %rd257, image_RNM0;
|
|
cvta.global.u64 %rd252, %rd257;
|
|
mov.u32 %r361, 8;
|
|
mov.u64 %rd256, 0;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r101, %r361, %rd20, %rd21, %rd256, %rd256);
|
|
// inline asm
|
|
mov.f32 %f1097, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1097;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd251], {%rs118, %rs118, %rs118, %rs119};
|
|
|
|
BB0_151:
|
|
ld.global.u32 %r362, [additive];
|
|
setp.eq.s32 %p145, %r362, 0;
|
|
@%p145 bra BB0_153;
|
|
|
|
mov.u64 %rd270, image_RNM1;
|
|
cvta.global.u64 %rd259, %rd270;
|
|
mov.u32 %r366, 8;
|
|
mov.u64 %rd269, 0;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r101, %r366, %rd20, %rd21, %rd269, %rd269);
|
|
// inline asm
|
|
ld.v4.u16 {%rs126, %rs127, %rs128, %rs129}, [%rd258];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1098, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1099, %rs127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1100, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd259, %r101, %r366, %rd20, %rd21, %rd269, %rd269);
|
|
// inline asm
|
|
add.f32 %f1101, %f1098, 0f00000000;
|
|
add.f32 %f1102, %f1099, 0f00000000;
|
|
add.f32 %f1103, %f1100, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1101;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs130, 0;
|
|
st.v4.u16 [%rd264], {%rs123, %rs124, %rs125, %rs130};
|
|
bra.uni BB0_154;
|
|
|
|
BB0_153:
|
|
mov.u64 %rd277, image_RNM1;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r368, 8;
|
|
mov.u64 %rd276, 0;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r101, %r368, %rd20, %rd21, %rd276, %rd276);
|
|
// inline asm
|
|
mov.f32 %f1104, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1104;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd271], {%rs131, %rs131, %rs131, %rs132};
|
|
|
|
BB0_154:
|
|
ld.global.u32 %r369, [additive];
|
|
setp.eq.s32 %p146, %r369, 0;
|
|
@%p146 bra BB0_156;
|
|
|
|
mov.u64 %rd290, image_RNM2;
|
|
cvta.global.u64 %rd279, %rd290;
|
|
mov.u32 %r373, 8;
|
|
mov.u64 %rd289, 0;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r101, %r373, %rd20, %rd21, %rd289, %rd289);
|
|
// inline asm
|
|
ld.v4.u16 {%rs139, %rs140, %rs141, %rs142}, [%rd278];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1105, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1106, %rs140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1107, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd284), _rt_buffer_get_64, (%rd279, %r101, %r373, %rd20, %rd21, %rd289, %rd289);
|
|
// inline asm
|
|
add.f32 %f1108, %f1105, 0f00000000;
|
|
add.f32 %f1109, %f1106, 0f00000000;
|
|
add.f32 %f1110, %f1107, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1110;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1109;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1108;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs143, 0;
|
|
st.v4.u16 [%rd284], {%rs136, %rs137, %rs138, %rs143};
|
|
bra.uni BB0_157;
|
|
|
|
BB0_156:
|
|
mov.u64 %rd297, image_RNM2;
|
|
cvta.global.u64 %rd292, %rd297;
|
|
mov.u32 %r375, 8;
|
|
mov.u64 %rd296, 0;
|
|
// inline asm
|
|
call (%rd291), _rt_buffer_get_64, (%rd292, %r101, %r375, %rd20, %rd21, %rd296, %rd296);
|
|
// inline asm
|
|
mov.f32 %f1111, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1111;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd291], {%rs144, %rs144, %rs144, %rs145};
|
|
|
|
BB0_157:
|
|
ld.global.u32 %r376, [additive];
|
|
setp.eq.s32 %p147, %r376, 0;
|
|
@%p147 bra BB0_159;
|
|
|
|
mov.u64 %rd310, image_RNM3;
|
|
cvta.global.u64 %rd299, %rd310;
|
|
mov.u32 %r380, 8;
|
|
mov.u64 %rd309, 0;
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd299, %r101, %r380, %rd20, %rd21, %rd309, %rd309);
|
|
// inline asm
|
|
ld.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd298];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1112, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1113, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1114, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd304), _rt_buffer_get_64, (%rd299, %r101, %r380, %rd20, %rd21, %rd309, %rd309);
|
|
// inline asm
|
|
add.f32 %f1115, %f1112, 0f00000000;
|
|
add.f32 %f1116, %f1113, 0f00000000;
|
|
add.f32 %f1117, %f1114, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1116;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1115;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs156, 0;
|
|
st.v4.u16 [%rd304], {%rs149, %rs150, %rs151, %rs156};
|
|
bra.uni BB0_160;
|
|
|
|
BB0_159:
|
|
mov.u64 %rd317, image_RNM3;
|
|
cvta.global.u64 %rd312, %rd317;
|
|
mov.u32 %r382, 8;
|
|
mov.u64 %rd316, 0;
|
|
// inline asm
|
|
call (%rd311), _rt_buffer_get_64, (%rd312, %r101, %r382, %rd20, %rd21, %rd316, %rd316);
|
|
// inline asm
|
|
mov.f32 %f1118, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1118;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd311], {%rs157, %rs157, %rs157, %rs158};
|
|
bra.uni BB0_160;
|
|
|
|
BB0_80:
|
|
setp.geu.f32 %p78, %f254, 0f00000000;
|
|
@%p78 bra BB0_83;
|
|
|
|
cvt.rzi.f32.f32 %f750, %f722;
|
|
setp.neu.f32 %p79, %f750, 0f3EE8BA2E;
|
|
selp.f32 %f1255, 0f7FFFFFFF, %f1255, %p79;
|
|
|
|
BB0_83:
|
|
add.f32 %f752, %f256, 0f3EE8BA2E;
|
|
mov.b32 %r269, %f752;
|
|
setp.lt.s32 %p81, %r269, 2139095040;
|
|
@%p81 bra BB0_88;
|
|
|
|
setp.gtu.f32 %p82, %f256, 0f7F800000;
|
|
@%p82 bra BB0_87;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_87:
|
|
add.f32 %f1255, %f254, 0f3EE8BA2E;
|
|
bra.uni BB0_88;
|
|
|
|
BB0_85:
|
|
setp.neu.f32 %p83, %f256, 0f7F800000;
|
|
@%p83 bra BB0_88;
|
|
|
|
selp.f32 %f1255, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_88:
|
|
mul.f32 %f753, %f1255, 0f437F0000;
|
|
setp.eq.f32 %p84, %f254, 0f3F800000;
|
|
selp.f32 %f754, 0f437F0000, %f753, %p84;
|
|
cvt.rzi.u32.f32 %r270, %f754;
|
|
cvt.u16.u32 %rs15, %r270;
|
|
mov.u16 %rs16, 255;
|
|
st.v2.u8 [%rd90], {%rs15, %rs16};
|
|
ld.global.u32 %r415, [imageEnabled];
|
|
|
|
BB0_89:
|
|
ld.global.f32 %f755, [lightColor];
|
|
mul.f32 %f268, %f1249, %f755;
|
|
ld.global.f32 %f756, [lightColor+4];
|
|
mul.f32 %f269, %f1250, %f756;
|
|
ld.global.f32 %f757, [lightColor+8];
|
|
mul.f32 %f270, %f1251, %f757;
|
|
and.b32 %r271, %r415, 1;
|
|
setp.eq.b32 %p85, %r271, 1;
|
|
@!%p85 bra BB0_124;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_90:
|
|
mov.f32 %f760, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f761, %f760;
|
|
fma.rn.f32 %f762, %f761, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f271, %f762;
|
|
abs.f32 %f272, %f268;
|
|
setp.lt.f32 %p86, %f272, 0f00800000;
|
|
mul.f32 %f763, %f272, 0f4B800000;
|
|
selp.f32 %f764, 0fC3170000, 0fC2FE0000, %p86;
|
|
selp.f32 %f765, %f763, %f272, %p86;
|
|
mov.b32 %r272, %f765;
|
|
and.b32 %r273, %r272, 8388607;
|
|
or.b32 %r274, %r273, 1065353216;
|
|
mov.b32 %f766, %r274;
|
|
shr.u32 %r275, %r272, 23;
|
|
cvt.rn.f32.u32 %f767, %r275;
|
|
add.f32 %f768, %f764, %f767;
|
|
setp.gt.f32 %p87, %f766, 0f3FB504F3;
|
|
mul.f32 %f769, %f766, 0f3F000000;
|
|
add.f32 %f770, %f768, 0f3F800000;
|
|
selp.f32 %f771, %f769, %f766, %p87;
|
|
selp.f32 %f772, %f770, %f768, %p87;
|
|
add.f32 %f773, %f771, 0fBF800000;
|
|
add.f32 %f759, %f771, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f758,%f759;
|
|
// inline asm
|
|
add.f32 %f774, %f773, %f773;
|
|
mul.f32 %f775, %f758, %f774;
|
|
mul.f32 %f776, %f775, %f775;
|
|
mov.f32 %f777, 0f3C4CAF63;
|
|
mov.f32 %f778, 0f3B18F0FE;
|
|
fma.rn.f32 %f779, %f778, %f776, %f777;
|
|
mov.f32 %f780, 0f3DAAAABD;
|
|
fma.rn.f32 %f781, %f779, %f776, %f780;
|
|
mul.rn.f32 %f782, %f781, %f776;
|
|
mul.rn.f32 %f783, %f782, %f775;
|
|
sub.f32 %f784, %f773, %f775;
|
|
neg.f32 %f785, %f775;
|
|
add.f32 %f786, %f784, %f784;
|
|
fma.rn.f32 %f787, %f785, %f773, %f786;
|
|
mul.rn.f32 %f788, %f758, %f787;
|
|
add.f32 %f789, %f783, %f775;
|
|
sub.f32 %f790, %f775, %f789;
|
|
add.f32 %f791, %f783, %f790;
|
|
add.f32 %f792, %f788, %f791;
|
|
add.f32 %f793, %f789, %f792;
|
|
sub.f32 %f794, %f789, %f793;
|
|
add.f32 %f795, %f792, %f794;
|
|
mov.f32 %f796, 0f3F317200;
|
|
mul.rn.f32 %f797, %f772, %f796;
|
|
mov.f32 %f798, 0f35BFBE8E;
|
|
mul.rn.f32 %f799, %f772, %f798;
|
|
add.f32 %f800, %f797, %f793;
|
|
sub.f32 %f801, %f797, %f800;
|
|
add.f32 %f802, %f793, %f801;
|
|
add.f32 %f803, %f795, %f802;
|
|
add.f32 %f804, %f799, %f803;
|
|
add.f32 %f805, %f800, %f804;
|
|
sub.f32 %f806, %f800, %f805;
|
|
add.f32 %f807, %f804, %f806;
|
|
mov.f32 %f808, 0f3EE66666;
|
|
mul.rn.f32 %f809, %f808, %f805;
|
|
neg.f32 %f810, %f809;
|
|
fma.rn.f32 %f811, %f808, %f805, %f810;
|
|
fma.rn.f32 %f812, %f808, %f807, %f811;
|
|
mov.f32 %f813, 0f00000000;
|
|
fma.rn.f32 %f814, %f813, %f805, %f812;
|
|
add.rn.f32 %f815, %f809, %f814;
|
|
neg.f32 %f816, %f815;
|
|
add.rn.f32 %f817, %f809, %f816;
|
|
add.rn.f32 %f818, %f817, %f814;
|
|
mov.b32 %r276, %f815;
|
|
setp.eq.s32 %p88, %r276, 1118925336;
|
|
add.s32 %r277, %r276, -1;
|
|
mov.b32 %f819, %r277;
|
|
add.f32 %f820, %f818, 0f37000000;
|
|
selp.f32 %f821, %f819, %f815, %p88;
|
|
selp.f32 %f273, %f820, %f818, %p88;
|
|
mul.f32 %f822, %f821, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f823, %f822;
|
|
mov.f32 %f824, 0fBF317200;
|
|
fma.rn.f32 %f825, %f823, %f824, %f821;
|
|
mov.f32 %f826, 0fB5BFBE8E;
|
|
fma.rn.f32 %f827, %f823, %f826, %f825;
|
|
mul.f32 %f828, %f827, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f829, %f828;
|
|
add.f32 %f830, %f823, 0f00000000;
|
|
ex2.approx.f32 %f831, %f830;
|
|
mul.f32 %f832, %f829, %f831;
|
|
setp.lt.f32 %p89, %f821, 0fC2D20000;
|
|
selp.f32 %f833, 0f00000000, %f832, %p89;
|
|
setp.gt.f32 %p90, %f821, 0f42D20000;
|
|
selp.f32 %f1256, 0f7F800000, %f833, %p90;
|
|
setp.eq.f32 %p91, %f1256, 0f7F800000;
|
|
@%p91 bra BB0_92;
|
|
|
|
fma.rn.f32 %f1256, %f1256, %f273, %f1256;
|
|
|
|
BB0_92:
|
|
setp.lt.f32 %p92, %f268, 0f00000000;
|
|
setp.eq.f32 %p93, %f271, 0f3F800000;
|
|
and.pred %p4, %p92, %p93;
|
|
mov.b32 %r278, %f1256;
|
|
xor.b32 %r279, %r278, -2147483648;
|
|
mov.b32 %f834, %r279;
|
|
selp.f32 %f1258, %f834, %f1256, %p4;
|
|
setp.eq.f32 %p94, %f268, 0f00000000;
|
|
@%p94 bra BB0_95;
|
|
bra.uni BB0_93;
|
|
|
|
BB0_95:
|
|
add.f32 %f837, %f268, %f268;
|
|
selp.f32 %f1258, %f837, 0f00000000, %p93;
|
|
bra.uni BB0_96;
|
|
|
|
BB0_93:
|
|
setp.geu.f32 %p95, %f268, 0f00000000;
|
|
@%p95 bra BB0_96;
|
|
|
|
cvt.rzi.f32.f32 %f836, %f808;
|
|
setp.neu.f32 %p96, %f836, 0f3EE66666;
|
|
selp.f32 %f1258, 0f7FFFFFFF, %f1258, %p96;
|
|
|
|
BB0_96:
|
|
add.f32 %f838, %f272, 0f3EE66666;
|
|
mov.b32 %r280, %f838;
|
|
setp.lt.s32 %p98, %r280, 2139095040;
|
|
@%p98 bra BB0_101;
|
|
|
|
setp.gtu.f32 %p99, %f272, 0f7F800000;
|
|
@%p99 bra BB0_100;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_100:
|
|
add.f32 %f1258, %f268, 0f3EE66666;
|
|
bra.uni BB0_101;
|
|
|
|
BB0_98:
|
|
setp.neu.f32 %p100, %f272, 0f7F800000;
|
|
@%p100 bra BB0_101;
|
|
|
|
selp.f32 %f1258, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_101:
|
|
setp.eq.f32 %p101, %f268, 0f3F800000;
|
|
selp.f32 %f284, 0f3F800000, %f1258, %p101;
|
|
abs.f32 %f285, %f269;
|
|
setp.lt.f32 %p102, %f285, 0f00800000;
|
|
mul.f32 %f841, %f285, 0f4B800000;
|
|
selp.f32 %f842, 0fC3170000, 0fC2FE0000, %p102;
|
|
selp.f32 %f843, %f841, %f285, %p102;
|
|
mov.b32 %r281, %f843;
|
|
and.b32 %r282, %r281, 8388607;
|
|
or.b32 %r283, %r282, 1065353216;
|
|
mov.b32 %f844, %r283;
|
|
shr.u32 %r284, %r281, 23;
|
|
cvt.rn.f32.u32 %f845, %r284;
|
|
add.f32 %f846, %f842, %f845;
|
|
setp.gt.f32 %p103, %f844, 0f3FB504F3;
|
|
mul.f32 %f847, %f844, 0f3F000000;
|
|
add.f32 %f848, %f846, 0f3F800000;
|
|
selp.f32 %f849, %f847, %f844, %p103;
|
|
selp.f32 %f850, %f848, %f846, %p103;
|
|
add.f32 %f851, %f849, 0fBF800000;
|
|
add.f32 %f840, %f849, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f839,%f840;
|
|
// inline asm
|
|
add.f32 %f852, %f851, %f851;
|
|
mul.f32 %f853, %f839, %f852;
|
|
mul.f32 %f854, %f853, %f853;
|
|
fma.rn.f32 %f857, %f778, %f854, %f777;
|
|
fma.rn.f32 %f859, %f857, %f854, %f780;
|
|
mul.rn.f32 %f860, %f859, %f854;
|
|
mul.rn.f32 %f861, %f860, %f853;
|
|
sub.f32 %f862, %f851, %f853;
|
|
neg.f32 %f863, %f853;
|
|
add.f32 %f864, %f862, %f862;
|
|
fma.rn.f32 %f865, %f863, %f851, %f864;
|
|
mul.rn.f32 %f866, %f839, %f865;
|
|
add.f32 %f867, %f861, %f853;
|
|
sub.f32 %f868, %f853, %f867;
|
|
add.f32 %f869, %f861, %f868;
|
|
add.f32 %f870, %f866, %f869;
|
|
add.f32 %f871, %f867, %f870;
|
|
sub.f32 %f872, %f867, %f871;
|
|
add.f32 %f873, %f870, %f872;
|
|
mul.rn.f32 %f875, %f850, %f796;
|
|
mul.rn.f32 %f877, %f850, %f798;
|
|
add.f32 %f878, %f875, %f871;
|
|
sub.f32 %f879, %f875, %f878;
|
|
add.f32 %f880, %f871, %f879;
|
|
add.f32 %f881, %f873, %f880;
|
|
add.f32 %f882, %f877, %f881;
|
|
add.f32 %f883, %f878, %f882;
|
|
sub.f32 %f884, %f878, %f883;
|
|
add.f32 %f885, %f882, %f884;
|
|
mul.rn.f32 %f887, %f808, %f883;
|
|
neg.f32 %f888, %f887;
|
|
fma.rn.f32 %f889, %f808, %f883, %f888;
|
|
fma.rn.f32 %f890, %f808, %f885, %f889;
|
|
fma.rn.f32 %f892, %f813, %f883, %f890;
|
|
add.rn.f32 %f893, %f887, %f892;
|
|
neg.f32 %f894, %f893;
|
|
add.rn.f32 %f895, %f887, %f894;
|
|
add.rn.f32 %f896, %f895, %f892;
|
|
mov.b32 %r285, %f893;
|
|
setp.eq.s32 %p104, %r285, 1118925336;
|
|
add.s32 %r286, %r285, -1;
|
|
mov.b32 %f897, %r286;
|
|
add.f32 %f898, %f896, 0f37000000;
|
|
selp.f32 %f899, %f897, %f893, %p104;
|
|
selp.f32 %f286, %f898, %f896, %p104;
|
|
mul.f32 %f900, %f899, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f901, %f900;
|
|
fma.rn.f32 %f903, %f901, %f824, %f899;
|
|
fma.rn.f32 %f905, %f901, %f826, %f903;
|
|
mul.f32 %f906, %f905, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f907, %f906;
|
|
add.f32 %f908, %f901, 0f00000000;
|
|
ex2.approx.f32 %f909, %f908;
|
|
mul.f32 %f910, %f907, %f909;
|
|
setp.lt.f32 %p105, %f899, 0fC2D20000;
|
|
selp.f32 %f911, 0f00000000, %f910, %p105;
|
|
setp.gt.f32 %p106, %f899, 0f42D20000;
|
|
selp.f32 %f1259, 0f7F800000, %f911, %p106;
|
|
setp.eq.f32 %p107, %f1259, 0f7F800000;
|
|
@%p107 bra BB0_103;
|
|
|
|
fma.rn.f32 %f1259, %f1259, %f286, %f1259;
|
|
|
|
BB0_103:
|
|
setp.lt.f32 %p108, %f269, 0f00000000;
|
|
and.pred %p5, %p108, %p93;
|
|
mov.b32 %r287, %f1259;
|
|
xor.b32 %r288, %r287, -2147483648;
|
|
mov.b32 %f912, %r288;
|
|
selp.f32 %f1261, %f912, %f1259, %p5;
|
|
setp.eq.f32 %p110, %f269, 0f00000000;
|
|
@%p110 bra BB0_106;
|
|
bra.uni BB0_104;
|
|
|
|
BB0_106:
|
|
add.f32 %f915, %f269, %f269;
|
|
selp.f32 %f1261, %f915, 0f00000000, %p93;
|
|
bra.uni BB0_107;
|
|
|
|
BB0_104:
|
|
setp.geu.f32 %p111, %f269, 0f00000000;
|
|
@%p111 bra BB0_107;
|
|
|
|
cvt.rzi.f32.f32 %f914, %f808;
|
|
setp.neu.f32 %p112, %f914, 0f3EE66666;
|
|
selp.f32 %f1261, 0f7FFFFFFF, %f1261, %p112;
|
|
|
|
BB0_107:
|
|
add.f32 %f916, %f285, 0f3EE66666;
|
|
mov.b32 %r289, %f916;
|
|
setp.lt.s32 %p114, %r289, 2139095040;
|
|
@%p114 bra BB0_112;
|
|
|
|
setp.gtu.f32 %p115, %f285, 0f7F800000;
|
|
@%p115 bra BB0_111;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_111:
|
|
add.f32 %f1261, %f269, 0f3EE66666;
|
|
bra.uni BB0_112;
|
|
|
|
BB0_109:
|
|
setp.neu.f32 %p116, %f285, 0f7F800000;
|
|
@%p116 bra BB0_112;
|
|
|
|
selp.f32 %f1261, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_112:
|
|
setp.eq.f32 %p117, %f269, 0f3F800000;
|
|
selp.f32 %f297, 0f3F800000, %f1261, %p117;
|
|
abs.f32 %f298, %f270;
|
|
setp.lt.f32 %p118, %f298, 0f00800000;
|
|
mul.f32 %f919, %f298, 0f4B800000;
|
|
selp.f32 %f920, 0fC3170000, 0fC2FE0000, %p118;
|
|
selp.f32 %f921, %f919, %f298, %p118;
|
|
mov.b32 %r290, %f921;
|
|
and.b32 %r291, %r290, 8388607;
|
|
or.b32 %r292, %r291, 1065353216;
|
|
mov.b32 %f922, %r292;
|
|
shr.u32 %r293, %r290, 23;
|
|
cvt.rn.f32.u32 %f923, %r293;
|
|
add.f32 %f924, %f920, %f923;
|
|
setp.gt.f32 %p119, %f922, 0f3FB504F3;
|
|
mul.f32 %f925, %f922, 0f3F000000;
|
|
add.f32 %f926, %f924, 0f3F800000;
|
|
selp.f32 %f927, %f925, %f922, %p119;
|
|
selp.f32 %f928, %f926, %f924, %p119;
|
|
add.f32 %f929, %f927, 0fBF800000;
|
|
add.f32 %f918, %f927, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f917,%f918;
|
|
// inline asm
|
|
add.f32 %f930, %f929, %f929;
|
|
mul.f32 %f931, %f917, %f930;
|
|
mul.f32 %f932, %f931, %f931;
|
|
fma.rn.f32 %f935, %f778, %f932, %f777;
|
|
fma.rn.f32 %f937, %f935, %f932, %f780;
|
|
mul.rn.f32 %f938, %f937, %f932;
|
|
mul.rn.f32 %f939, %f938, %f931;
|
|
sub.f32 %f940, %f929, %f931;
|
|
neg.f32 %f941, %f931;
|
|
add.f32 %f942, %f940, %f940;
|
|
fma.rn.f32 %f943, %f941, %f929, %f942;
|
|
mul.rn.f32 %f944, %f917, %f943;
|
|
add.f32 %f945, %f939, %f931;
|
|
sub.f32 %f946, %f931, %f945;
|
|
add.f32 %f947, %f939, %f946;
|
|
add.f32 %f948, %f944, %f947;
|
|
add.f32 %f949, %f945, %f948;
|
|
sub.f32 %f950, %f945, %f949;
|
|
add.f32 %f951, %f948, %f950;
|
|
mul.rn.f32 %f953, %f928, %f796;
|
|
mul.rn.f32 %f955, %f928, %f798;
|
|
add.f32 %f956, %f953, %f949;
|
|
sub.f32 %f957, %f953, %f956;
|
|
add.f32 %f958, %f949, %f957;
|
|
add.f32 %f959, %f951, %f958;
|
|
add.f32 %f960, %f955, %f959;
|
|
add.f32 %f961, %f956, %f960;
|
|
sub.f32 %f962, %f956, %f961;
|
|
add.f32 %f963, %f960, %f962;
|
|
mul.rn.f32 %f965, %f808, %f961;
|
|
neg.f32 %f966, %f965;
|
|
fma.rn.f32 %f967, %f808, %f961, %f966;
|
|
fma.rn.f32 %f968, %f808, %f963, %f967;
|
|
fma.rn.f32 %f970, %f813, %f961, %f968;
|
|
add.rn.f32 %f971, %f965, %f970;
|
|
neg.f32 %f972, %f971;
|
|
add.rn.f32 %f973, %f965, %f972;
|
|
add.rn.f32 %f974, %f973, %f970;
|
|
mov.b32 %r294, %f971;
|
|
setp.eq.s32 %p120, %r294, 1118925336;
|
|
add.s32 %r295, %r294, -1;
|
|
mov.b32 %f975, %r295;
|
|
add.f32 %f976, %f974, 0f37000000;
|
|
selp.f32 %f977, %f975, %f971, %p120;
|
|
selp.f32 %f299, %f976, %f974, %p120;
|
|
mul.f32 %f978, %f977, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f979, %f978;
|
|
fma.rn.f32 %f981, %f979, %f824, %f977;
|
|
fma.rn.f32 %f983, %f979, %f826, %f981;
|
|
mul.f32 %f984, %f983, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f985, %f984;
|
|
add.f32 %f986, %f979, 0f00000000;
|
|
ex2.approx.f32 %f987, %f986;
|
|
mul.f32 %f988, %f985, %f987;
|
|
setp.lt.f32 %p121, %f977, 0fC2D20000;
|
|
selp.f32 %f989, 0f00000000, %f988, %p121;
|
|
setp.gt.f32 %p122, %f977, 0f42D20000;
|
|
selp.f32 %f1262, 0f7F800000, %f989, %p122;
|
|
setp.eq.f32 %p123, %f1262, 0f7F800000;
|
|
@%p123 bra BB0_114;
|
|
|
|
fma.rn.f32 %f1262, %f1262, %f299, %f1262;
|
|
|
|
BB0_114:
|
|
setp.lt.f32 %p124, %f270, 0f00000000;
|
|
and.pred %p6, %p124, %p93;
|
|
mov.b32 %r296, %f1262;
|
|
xor.b32 %r297, %r296, -2147483648;
|
|
mov.b32 %f990, %r297;
|
|
selp.f32 %f1264, %f990, %f1262, %p6;
|
|
setp.eq.f32 %p126, %f270, 0f00000000;
|
|
@%p126 bra BB0_117;
|
|
bra.uni BB0_115;
|
|
|
|
BB0_117:
|
|
add.f32 %f993, %f270, %f270;
|
|
selp.f32 %f1264, %f993, 0f00000000, %p93;
|
|
bra.uni BB0_118;
|
|
|
|
BB0_115:
|
|
setp.geu.f32 %p127, %f270, 0f00000000;
|
|
@%p127 bra BB0_118;
|
|
|
|
cvt.rzi.f32.f32 %f992, %f808;
|
|
setp.neu.f32 %p128, %f992, 0f3EE66666;
|
|
selp.f32 %f1264, 0f7FFFFFFF, %f1264, %p128;
|
|
|
|
BB0_118:
|
|
add.f32 %f994, %f298, 0f3EE66666;
|
|
mov.b32 %r298, %f994;
|
|
setp.lt.s32 %p130, %r298, 2139095040;
|
|
@%p130 bra BB0_123;
|
|
|
|
setp.gtu.f32 %p131, %f298, 0f7F800000;
|
|
@%p131 bra BB0_122;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_122:
|
|
add.f32 %f1264, %f270, 0f3EE66666;
|
|
bra.uni BB0_123;
|
|
|
|
BB0_120:
|
|
setp.neu.f32 %p132, %f298, 0f7F800000;
|
|
@%p132 bra BB0_123;
|
|
|
|
selp.f32 %f1264, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_123:
|
|
mov.u32 %r385, 4;
|
|
setp.eq.f32 %p133, %f270, 0f3F800000;
|
|
selp.f32 %f995, 0f3F800000, %f1264, %p133;
|
|
cvt.u64.u32 %rd100, %r7;
|
|
cvt.u64.u32 %rd99, %r6;
|
|
mov.u64 %rd103, image;
|
|
cvta.global.u64 %rd98, %rd103;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r101, %r385, %rd99, %rd100, %rd27, %rd27);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f996, %f995;
|
|
mul.f32 %f997, %f996, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r301, %f997;
|
|
cvt.sat.f32.f32 %f998, %f297;
|
|
mul.f32 %f999, %f998, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r302, %f999;
|
|
cvt.sat.f32.f32 %f1000, %f284;
|
|
mul.f32 %f1001, %f1000, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r303, %f1001;
|
|
cvt.u16.u32 %rs17, %r301;
|
|
cvt.u16.u32 %rs18, %r303;
|
|
cvt.u16.u32 %rs19, %r302;
|
|
mov.u16 %rs20, 255;
|
|
st.v4.u8 [%rd97], {%rs17, %rs19, %rs18, %rs20};
|
|
ld.global.u32 %r415, [imageEnabled];
|
|
|
|
BB0_124:
|
|
cvt.u64.u32 %rd18, %r6;
|
|
cvt.u64.u32 %rd19, %r7;
|
|
and.b32 %r304, %r415, 4;
|
|
setp.eq.s32 %p134, %r304, 0;
|
|
@%p134 bra BB0_128;
|
|
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p135, %r305, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f539;}
|
|
|
|
// inline asm
|
|
@%p135 bra BB0_127;
|
|
|
|
mov.u64 %rd116, image_HDR;
|
|
cvta.global.u64 %rd105, %rd116;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd105, %r101, %r309, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs28, %rs29, %rs30, %rs31}, [%rd104];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1003, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1004, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1005, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd105, %r101, %r309, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1006, %f268, %f1003;
|
|
add.f32 %f1007, %f269, %f1004;
|
|
add.f32 %f1008, %f270, %f1005;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1008;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f1007;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f1006;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs25, %rs26, %rs27, %rs21};
|
|
bra.uni BB0_128;
|
|
|
|
BB0_127:
|
|
mov.u64 %rd123, image_HDR;
|
|
cvta.global.u64 %rd118, %rd123;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r101, %r311, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f270;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f269;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f268;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd117], {%rs32, %rs33, %rs34, %rs21};
|
|
|
|
BB0_128:
|
|
selp.f32 %f1013, 0f3F800000, 0f3E800000, %p148;
|
|
mul.f32 %f1014, %f1013, %f1246;
|
|
mul.f32 %f1015, %f1013, %f1247;
|
|
mul.f32 %f1016, %f1013, %f1248;
|
|
mul.f32 %f1017, %f1013, %f1243;
|
|
mul.f32 %f1018, %f1013, %f1244;
|
|
mul.f32 %f1019, %f1013, %f1245;
|
|
mul.f32 %f1020, %f1013, %f1240;
|
|
mul.f32 %f1021, %f1013, %f1241;
|
|
mul.f32 %f1022, %f1013, %f1242;
|
|
mul.f32 %f1023, %f1013, %f1237;
|
|
mul.f32 %f1024, %f1013, %f1238;
|
|
mul.f32 %f1025, %f1013, %f1239;
|
|
ld.global.f32 %f1026, [lightColor];
|
|
mul.f32 %f310, %f1014, %f1026;
|
|
ld.global.f32 %f1027, [lightColor+4];
|
|
mul.f32 %f311, %f1015, %f1027;
|
|
ld.global.f32 %f1028, [lightColor+8];
|
|
mul.f32 %f312, %f1016, %f1028;
|
|
mul.f32 %f313, %f1017, %f1026;
|
|
mul.f32 %f314, %f1018, %f1027;
|
|
mul.f32 %f315, %f1019, %f1028;
|
|
mul.f32 %f316, %f1020, %f1026;
|
|
mul.f32 %f317, %f1021, %f1027;
|
|
mul.f32 %f318, %f1022, %f1028;
|
|
mul.f32 %f319, %f1023, %f1026;
|
|
mul.f32 %f320, %f1024, %f1027;
|
|
mul.f32 %f321, %f1025, %f1028;
|
|
ld.global.u32 %r312, [additive];
|
|
setp.eq.s32 %p136, %r312, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f539;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_130;
|
|
|
|
mov.u64 %rd136, image_RNM0;
|
|
cvta.global.u64 %rd125, %rd136;
|
|
mov.u32 %r316, 8;
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd125, %r101, %r316, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd124];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1029, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1030, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1031, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd125, %r101, %r316, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1032, %f310, %f1029;
|
|
add.f32 %f1033, %f311, %f1030;
|
|
add.f32 %f1034, %f312, %f1031;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1034;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1033;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f1032;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs39, %rs40, %rs41, %rs35};
|
|
bra.uni BB0_131;
|
|
|
|
BB0_130:
|
|
mov.u64 %rd143, image_RNM0;
|
|
cvta.global.u64 %rd138, %rd143;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r101, %r318, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f312;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f311;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f310;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd137], {%rs46, %rs47, %rs48, %rs35};
|
|
|
|
BB0_131:
|
|
mov.f32 %f1039, 0f34000000;
|
|
max.f32 %f1040, %f310, %f1039;
|
|
div.rn.f32 %f1041, %f313, %f1040;
|
|
max.f32 %f1042, %f311, %f1039;
|
|
div.rn.f32 %f1043, %f314, %f1042;
|
|
max.f32 %f1044, %f312, %f1039;
|
|
div.rn.f32 %f1045, %f315, %f1044;
|
|
fma.rn.f32 %f322, %f1041, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f323, %f1043, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f324, %f1045, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1046, %f316, %f1040;
|
|
div.rn.f32 %f1047, %f317, %f1042;
|
|
div.rn.f32 %f1048, %f318, %f1044;
|
|
fma.rn.f32 %f325, %f1046, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f326, %f1047, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f327, %f1048, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1049, %f319, %f1040;
|
|
div.rn.f32 %f1050, %f320, %f1042;
|
|
div.rn.f32 %f1051, %f321, %f1044;
|
|
fma.rn.f32 %f328, %f1049, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f329, %f1050, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f330, %f1051, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r319, [additive];
|
|
setp.eq.s32 %p137, %r319, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f539;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_133;
|
|
|
|
mov.u64 %rd156, image_RNM1;
|
|
cvta.global.u64 %rd145, %rd156;
|
|
mov.u32 %r323, 8;
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd145, %r101, %r323, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd144];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1052, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1053, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1054, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd145, %r101, %r323, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1055, %f322, %f1052;
|
|
add.f32 %f1056, %f323, %f1053;
|
|
add.f32 %f1057, %f324, %f1054;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1057;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1056;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1055;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs53, %rs54, %rs55, %rs49};
|
|
bra.uni BB0_134;
|
|
|
|
BB0_133:
|
|
mov.u64 %rd163, image_RNM1;
|
|
cvta.global.u64 %rd158, %rd163;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd158, %r101, %r325, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f324;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f323;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f322;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd157], {%rs60, %rs61, %rs62, %rs49};
|
|
|
|
BB0_134:
|
|
ld.global.u32 %r326, [additive];
|
|
setp.eq.s32 %p138, %r326, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f539;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_136;
|
|
|
|
mov.u64 %rd176, image_RNM2;
|
|
cvta.global.u64 %rd165, %rd176;
|
|
mov.u32 %r330, 8;
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd165, %r101, %r330, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd164];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1062, %rs70;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1063, %rs71;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1064, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd165, %r101, %r330, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1065, %f325, %f1062;
|
|
add.f32 %f1066, %f326, %f1063;
|
|
add.f32 %f1067, %f327, %f1064;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1067;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1066;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f1065;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd170], {%rs67, %rs68, %rs69, %rs63};
|
|
bra.uni BB0_137;
|
|
|
|
BB0_136:
|
|
mov.u64 %rd183, image_RNM2;
|
|
cvta.global.u64 %rd178, %rd183;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r101, %r332, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f327;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f326;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f325;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd177], {%rs74, %rs75, %rs76, %rs63};
|
|
|
|
BB0_137:
|
|
ld.global.u32 %r333, [additive];
|
|
setp.eq.s32 %p139, %r333, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f539;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_139;
|
|
|
|
mov.u64 %rd196, image_RNM3;
|
|
cvta.global.u64 %rd185, %rd196;
|
|
mov.u32 %r337, 8;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r101, %r337, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd184];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1072, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1073, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1074, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd190), _rt_buffer_get_64, (%rd185, %r101, %r337, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1075, %f328, %f1072;
|
|
add.f32 %f1076, %f329, %f1073;
|
|
add.f32 %f1077, %f330, %f1074;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1077;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f1076;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1075;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd190], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_160;
|
|
|
|
BB0_139:
|
|
mov.u64 %rd203, image_RNM3;
|
|
cvta.global.u64 %rd198, %rd203;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd198, %r101, %r339, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f330;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f329;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f328;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd197], {%rs88, %rs89, %rs90, %rs77};
|
|
|
|
BB0_160:
|
|
ret;
|
|
}
|
|
|
|
|