3372 lines
91 KiB
Plaintext
3372 lines
91 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 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<160>;
|
|
.reg .b16 %rs<299>;
|
|
.reg .f32 %f<1760>;
|
|
.reg .b32 %r<484>;
|
|
.reg .b64 %rd<520>;
|
|
|
|
|
|
mov.u64 %rd519, __local_depot0;
|
|
cvta.local.u64 %SP, %rd519;
|
|
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 %rd25, %r103;
|
|
cvt.u64.u32 %rd26, %r104;
|
|
mov.u64 %rd29, uvnormal;
|
|
cvta.global.u64 %rd24, %rd29;
|
|
mov.u32 %r101, 2;
|
|
mov.u32 %r102, 4;
|
|
mov.u64 %rd28, 0;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r101, %r102, %rd25, %rd26, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r5, [%rd23];
|
|
shr.u32 %r107, %r5, 16;
|
|
cvt.u16.u32 %rs1, %r107;
|
|
and.b16 %rs12, %rs1, 255;
|
|
cvt.u16.u32 %rs13, %r5;
|
|
or.b16 %rs14, %rs13, %rs12;
|
|
setp.eq.s16 %p7, %rs14, 0;
|
|
mov.f32 %f1510, 0f00000000;
|
|
mov.f32 %f1511, %f1510;
|
|
mov.f32 %f1512, %f1510;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs15, [%rd23+1];
|
|
and.b16 %rs17, %rs13, 255;
|
|
cvt.rn.f32.u16 %f543, %rs17;
|
|
div.rn.f32 %f544, %f543, 0f437F0000;
|
|
fma.rn.f32 %f545, %f544, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f546, %rs15;
|
|
div.rn.f32 %f547, %f546, 0f437F0000;
|
|
fma.rn.f32 %f548, %f547, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f549, %rs12;
|
|
div.rn.f32 %f550, %f549, 0f437F0000;
|
|
fma.rn.f32 %f551, %f550, 0f40000000, 0fBF800000;
|
|
mul.f32 %f552, %f548, %f548;
|
|
fma.rn.f32 %f553, %f545, %f545, %f552;
|
|
fma.rn.f32 %f554, %f551, %f551, %f553;
|
|
sqrt.rn.f32 %f555, %f554;
|
|
rcp.rn.f32 %f556, %f555;
|
|
mul.f32 %f1510, %f545, %f556;
|
|
mul.f32 %f1511, %f548, %f556;
|
|
mul.f32 %f1512, %f551, %f556;
|
|
|
|
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, %f1511, 0f00000000;
|
|
setp.eq.f32 %p9, %f1510, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1512, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_157;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_157:
|
|
ld.global.u32 %r483, [imageEnabled];
|
|
and.b32 %r372, %r483, 1;
|
|
setp.eq.b32 %p146, %r372, 1;
|
|
@!%p146 bra BB0_159;
|
|
bra.uni BB0_158;
|
|
|
|
BB0_158:
|
|
cvt.u64.u32 %rd300, %r6;
|
|
cvt.u64.u32 %rd301, %r7;
|
|
mov.u64 %rd304, image;
|
|
cvta.global.u64 %rd299, %rd304;
|
|
mov.u64 %rd303, 0;
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd299, %r101, %r102, %rd300, %rd301, %rd303, %rd303);
|
|
// inline asm
|
|
mov.u16 %rs166, 0;
|
|
st.v4.u8 [%rd298], {%rs166, %rs166, %rs166, %rs166};
|
|
ld.global.u32 %r483, [imageEnabled];
|
|
|
|
BB0_159:
|
|
and.b32 %r375, %r483, 8;
|
|
setp.eq.s32 %p147, %r375, 0;
|
|
@%p147 bra BB0_161;
|
|
|
|
cvt.u64.u32 %rd308, %r7;
|
|
cvt.u64.u32 %rd307, %r6;
|
|
mov.u64 %rd311, image_Mask;
|
|
cvta.global.u64 %rd306, %rd311;
|
|
mov.u64 %rd310, 0;
|
|
// inline asm
|
|
call (%rd305), _rt_buffer_get_64, (%rd306, %r101, %r101, %rd307, %rd308, %rd310, %rd310);
|
|
// inline asm
|
|
mov.f32 %f1432, 0f00000000;
|
|
cvt.rzi.u32.f32 %r378, %f1432;
|
|
cvt.u16.u32 %rs167, %r378;
|
|
mov.u16 %rs168, 0;
|
|
st.v2.u8 [%rd305], {%rs167, %rs168};
|
|
ld.global.u32 %r483, [imageEnabled];
|
|
|
|
BB0_161:
|
|
cvt.u64.u32 %rd21, %r6;
|
|
cvt.u64.u32 %rd22, %r7;
|
|
and.b32 %r379, %r483, 4;
|
|
setp.eq.s32 %p148, %r379, 0;
|
|
@%p148 bra BB0_165;
|
|
|
|
ld.global.u32 %r380, [additive];
|
|
setp.eq.s32 %p149, %r380, 0;
|
|
@%p149 bra BB0_164;
|
|
|
|
mov.u64 %rd324, image_HDR;
|
|
cvta.global.u64 %rd313, %rd324;
|
|
mov.u32 %r384, 8;
|
|
mov.u64 %rd323, 0;
|
|
// inline asm
|
|
call (%rd312), _rt_buffer_get_64, (%rd313, %r101, %r384, %rd21, %rd22, %rd323, %rd323);
|
|
// inline asm
|
|
ld.v4.u16 {%rs175, %rs176, %rs177, %rs178}, [%rd312];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1433, %rs175;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1434, %rs176;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1435, %rs177;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd318), _rt_buffer_get_64, (%rd313, %r101, %r384, %rd21, %rd22, %rd323, %rd323);
|
|
// inline asm
|
|
add.f32 %f1436, %f1433, 0f00000000;
|
|
add.f32 %f1437, %f1434, 0f00000000;
|
|
add.f32 %f1438, %f1435, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs174, %f1438;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs173, %f1437;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs172, %f1436;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs179, 0;
|
|
st.v4.u16 [%rd318], {%rs172, %rs173, %rs174, %rs179};
|
|
bra.uni BB0_165;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r121, %r122}, [pixelID];
|
|
cvt.u64.u32 %rd32, %r121;
|
|
cvt.u64.u32 %rd33, %r122;
|
|
mov.u64 %rd42, uvpos;
|
|
cvta.global.u64 %rd31, %rd42;
|
|
mov.u32 %r117, 12;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r101, %r117, %rd32, %rd33, %rd28, %rd28);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd30+8];
|
|
ld.f32 %f8, [%rd30+4];
|
|
ld.f32 %f7, [%rd30];
|
|
mul.f32 %f588, %f7, 0f3456BF95;
|
|
mul.f32 %f589, %f8, 0f3456BF95;
|
|
mul.f32 %f590, %f9, 0f3456BF95;
|
|
abs.f32 %f591, %f1510;
|
|
div.rn.f32 %f592, %f588, %f591;
|
|
abs.f32 %f593, %f1511;
|
|
div.rn.f32 %f594, %f589, %f593;
|
|
abs.f32 %f595, %f1512;
|
|
div.rn.f32 %f596, %f590, %f595;
|
|
abs.f32 %f597, %f592;
|
|
abs.f32 %f598, %f594;
|
|
abs.f32 %f599, %f596;
|
|
mov.f32 %f600, 0f38D1B717;
|
|
max.f32 %f601, %f597, %f600;
|
|
max.f32 %f602, %f598, %f600;
|
|
max.f32 %f603, %f599, %f600;
|
|
fma.rn.f32 %f10, %f1510, %f601, %f7;
|
|
fma.rn.f32 %f11, %f1511, %f602, %f8;
|
|
fma.rn.f32 %f12, %f1512, %f603, %f9;
|
|
ld.global.v2.u32 {%r125, %r126}, [pixelID];
|
|
cvt.u64.u32 %rd38, %r125;
|
|
cvt.u64.u32 %rd39, %r126;
|
|
mov.u64 %rd43, rnd_seeds;
|
|
cvta.global.u64 %rd37, %rd43;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r101, %r102, %rd38, %rd39, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r459, [%rd36];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f17, 0f00000000;
|
|
mov.u32 %r455, 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 %r455, 0;
|
|
abs.f32 %f758, %f15;
|
|
abs.f32 %f759, %f14;
|
|
max.f32 %f760, %f759, %f758;
|
|
abs.f32 %f761, %f16;
|
|
max.f32 %f762, %f760, %f761;
|
|
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;
|
|
mov.f32 %f33, %f17;
|
|
mov.f32 %f34, %f17;
|
|
mov.f32 %f35, %f17;
|
|
mov.f32 %f36, %f17;
|
|
mov.f32 %f37, %f17;
|
|
mov.f32 %f38, %f17;
|
|
mov.f32 %f39, %f17;
|
|
mov.f32 %f40, %f17;
|
|
mov.f32 %f41, %f17;
|
|
mov.f32 %f42, %f17;
|
|
mov.f32 %f43, %f17;
|
|
mov.f32 %f44, %f17;
|
|
mov.f32 %f45, %f17;
|
|
mov.f32 %f46, %f17;
|
|
mov.f32 %f47, %f17;
|
|
|
|
BB0_6:
|
|
shl.b32 %r10, %r455, 1;
|
|
cvt.s64.s32 %rd46, %r10;
|
|
mov.u64 %rd56, lightMeshBuffer;
|
|
cvta.global.u64 %rd45, %rd56;
|
|
mov.u32 %r130, 1;
|
|
mov.u64 %rd55, 0;
|
|
// inline asm
|
|
call (%rd44), _rt_buffer_get_64, (%rd45, %r130, %r117, %rd46, %rd55, %rd55, %rd55);
|
|
// inline asm
|
|
ld.f32 %f635, [%rd44];
|
|
sub.f32 %f636, %f635, %f7;
|
|
ld.f32 %f637, [%rd44+4];
|
|
sub.f32 %f638, %f637, %f8;
|
|
ld.f32 %f639, [%rd44+8];
|
|
sub.f32 %f640, %f639, %f9;
|
|
mul.f32 %f641, %f638, %f638;
|
|
fma.rn.f32 %f642, %f636, %f636, %f641;
|
|
fma.rn.f32 %f643, %f640, %f640, %f642;
|
|
sqrt.rn.f32 %f48, %f643;
|
|
rcp.rn.f32 %f644, %f48;
|
|
mul.f32 %f49, %f636, %f644;
|
|
mul.f32 %f50, %f638, %f644;
|
|
mul.f32 %f51, %f640, %f644;
|
|
mul.f32 %f645, %f48, %f48;
|
|
div.rn.f32 %f52, %f13, %f645;
|
|
ld.global.v2.u32 {%r134, %r135}, [pixelID];
|
|
cvt.u64.u32 %rd57, %r135;
|
|
cvt.u64.u32 %rd58, %r134;
|
|
and.b64 %rd52, %rd58, 255;
|
|
and.b64 %rd53, %rd57, 255;
|
|
mov.u64 %rd59, rnd_seeds2;
|
|
cvta.global.u64 %rd51, %rd59;
|
|
mov.u32 %r133, 3;
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd51, %r101, %r133, %rd52, %rd53, %rd55, %rd55);
|
|
// inline asm
|
|
setp.eq.s32 %p15, %r3, 0;
|
|
@%p15 bra BB0_8;
|
|
|
|
ld.u8 %rs19, [%rd50];
|
|
cvt.rn.f32.u16 %f646, %rs19;
|
|
div.rn.f32 %f647, %f646, 0fC37F0000;
|
|
fma.rn.f32 %f648, %f647, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p17, %f52, %f648;
|
|
mov.pred %p159, -1;
|
|
@%p17 bra BB0_23;
|
|
|
|
BB0_8:
|
|
ld.global.f32 %f651, [lightInvCutoff];
|
|
mul.f32 %f53, %f48, %f651;
|
|
mov.f32 %f655, 0f40800000;
|
|
abs.f32 %f55, %f53;
|
|
setp.lt.f32 %p18, %f55, 0f00800000;
|
|
mul.f32 %f657, %f55, 0f4B800000;
|
|
selp.f32 %f658, 0fC3170000, 0fC2FE0000, %p18;
|
|
selp.f32 %f659, %f657, %f55, %p18;
|
|
mov.b32 %r138, %f659;
|
|
and.b32 %r139, %r138, 8388607;
|
|
or.b32 %r140, %r139, 1065353216;
|
|
mov.b32 %f660, %r140;
|
|
shr.u32 %r141, %r138, 23;
|
|
cvt.rn.f32.u32 %f661, %r141;
|
|
add.f32 %f662, %f658, %f661;
|
|
setp.gt.f32 %p19, %f660, 0f3FB504F3;
|
|
mul.f32 %f663, %f660, 0f3F000000;
|
|
add.f32 %f664, %f662, 0f3F800000;
|
|
selp.f32 %f665, %f663, %f660, %p19;
|
|
selp.f32 %f666, %f664, %f662, %p19;
|
|
add.f32 %f667, %f665, 0fBF800000;
|
|
add.f32 %f650, %f665, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f649,%f650;
|
|
// inline asm
|
|
add.f32 %f668, %f667, %f667;
|
|
mul.f32 %f669, %f649, %f668;
|
|
mul.f32 %f670, %f669, %f669;
|
|
mov.f32 %f671, 0f3C4CAF63;
|
|
mov.f32 %f672, 0f3B18F0FE;
|
|
fma.rn.f32 %f673, %f672, %f670, %f671;
|
|
mov.f32 %f674, 0f3DAAAABD;
|
|
fma.rn.f32 %f675, %f673, %f670, %f674;
|
|
mul.rn.f32 %f676, %f675, %f670;
|
|
mul.rn.f32 %f677, %f676, %f669;
|
|
sub.f32 %f678, %f667, %f669;
|
|
neg.f32 %f679, %f669;
|
|
add.f32 %f680, %f678, %f678;
|
|
fma.rn.f32 %f681, %f679, %f667, %f680;
|
|
mul.rn.f32 %f682, %f649, %f681;
|
|
add.f32 %f683, %f677, %f669;
|
|
sub.f32 %f684, %f669, %f683;
|
|
add.f32 %f685, %f677, %f684;
|
|
add.f32 %f686, %f682, %f685;
|
|
add.f32 %f687, %f683, %f686;
|
|
sub.f32 %f688, %f683, %f687;
|
|
add.f32 %f689, %f686, %f688;
|
|
mov.f32 %f690, 0f3F317200;
|
|
mul.rn.f32 %f691, %f666, %f690;
|
|
mov.f32 %f692, 0f35BFBE8E;
|
|
mul.rn.f32 %f693, %f666, %f692;
|
|
add.f32 %f694, %f691, %f687;
|
|
sub.f32 %f695, %f691, %f694;
|
|
add.f32 %f696, %f687, %f695;
|
|
add.f32 %f697, %f689, %f696;
|
|
add.f32 %f698, %f693, %f697;
|
|
add.f32 %f699, %f694, %f698;
|
|
sub.f32 %f700, %f694, %f699;
|
|
add.f32 %f701, %f698, %f700;
|
|
mul.rn.f32 %f702, %f655, %f699;
|
|
neg.f32 %f703, %f702;
|
|
fma.rn.f32 %f704, %f655, %f699, %f703;
|
|
fma.rn.f32 %f705, %f655, %f701, %f704;
|
|
mov.f32 %f706, 0f00000000;
|
|
fma.rn.f32 %f707, %f706, %f699, %f705;
|
|
add.rn.f32 %f708, %f702, %f707;
|
|
neg.f32 %f709, %f708;
|
|
add.rn.f32 %f710, %f702, %f709;
|
|
add.rn.f32 %f711, %f710, %f707;
|
|
mov.b32 %r142, %f708;
|
|
setp.eq.s32 %p20, %r142, 1118925336;
|
|
add.s32 %r143, %r142, -1;
|
|
mov.b32 %f712, %r143;
|
|
add.f32 %f713, %f711, 0f37000000;
|
|
selp.f32 %f714, %f712, %f708, %p20;
|
|
selp.f32 %f56, %f713, %f711, %p20;
|
|
mul.f32 %f715, %f714, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f716, %f715;
|
|
mov.f32 %f717, 0fBF317200;
|
|
fma.rn.f32 %f718, %f716, %f717, %f714;
|
|
mov.f32 %f719, 0fB5BFBE8E;
|
|
fma.rn.f32 %f720, %f716, %f719, %f718;
|
|
mul.f32 %f721, %f720, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f722, %f721;
|
|
add.f32 %f723, %f716, 0f00000000;
|
|
ex2.approx.f32 %f724, %f723;
|
|
mul.f32 %f725, %f722, %f724;
|
|
setp.lt.f32 %p21, %f714, 0fC2D20000;
|
|
selp.f32 %f726, 0f00000000, %f725, %p21;
|
|
setp.gt.f32 %p22, %f714, 0f42D20000;
|
|
selp.f32 %f1544, 0f7F800000, %f726, %p22;
|
|
setp.eq.f32 %p23, %f1544, 0f7F800000;
|
|
@%p23 bra BB0_10;
|
|
|
|
fma.rn.f32 %f1544, %f1544, %f56, %f1544;
|
|
|
|
BB0_10:
|
|
mov.f32 %f1508, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1507, %f1508;
|
|
add.f32 %f1506, %f1507, %f1507;
|
|
mov.f32 %f1505, 0f40800000;
|
|
sub.f32 %f1504, %f1505, %f1506;
|
|
abs.f32 %f1503, %f1504;
|
|
setp.lt.f32 %p24, %f53, 0f00000000;
|
|
setp.eq.f32 %p25, %f1503, 0f3F800000;
|
|
and.pred %p1, %p24, %p25;
|
|
mov.b32 %r144, %f1544;
|
|
xor.b32 %r145, %r144, -2147483648;
|
|
mov.b32 %f727, %r145;
|
|
selp.f32 %f1546, %f727, %f1544, %p1;
|
|
setp.eq.f32 %p26, %f53, 0f00000000;
|
|
@%p26 bra BB0_13;
|
|
bra.uni BB0_11;
|
|
|
|
BB0_13:
|
|
add.f32 %f730, %f53, %f53;
|
|
selp.f32 %f1546, %f730, 0f00000000, %p25;
|
|
bra.uni BB0_14;
|
|
|
|
BB0_11:
|
|
setp.geu.f32 %p27, %f53, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
|
|
mov.f32 %f1509, 0f40800000;
|
|
cvt.rzi.f32.f32 %f729, %f1509;
|
|
setp.neu.f32 %p28, %f729, 0f40800000;
|
|
selp.f32 %f1546, 0f7FFFFFFF, %f1546, %p28;
|
|
|
|
BB0_14:
|
|
add.f32 %f731, %f55, 0f40800000;
|
|
mov.b32 %r146, %f731;
|
|
setp.lt.s32 %p30, %r146, 2139095040;
|
|
@%p30 bra BB0_19;
|
|
|
|
setp.gtu.f32 %p31, %f55, 0f7F800000;
|
|
@%p31 bra BB0_18;
|
|
bra.uni BB0_16;
|
|
|
|
BB0_18:
|
|
add.f32 %f1546, %f53, 0f40800000;
|
|
bra.uni BB0_19;
|
|
|
|
BB0_16:
|
|
setp.neu.f32 %p32, %f55, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
|
|
selp.f32 %f1546, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_19:
|
|
mov.u64 %rd514, 0;
|
|
mov.u32 %r451, 1;
|
|
mov.u64 %rd513, lightMeshBuffer;
|
|
cvta.global.u64 %rd512, %rd513;
|
|
shl.b32 %r450, %r455, 1;
|
|
mov.f32 %f732, 0f3F800000;
|
|
sub.f32 %f733, %f732, %f1546;
|
|
setp.eq.f32 %p33, %f53, 0f3F800000;
|
|
selp.f32 %f734, 0f00000000, %f733, %p33;
|
|
cvt.sat.f32.f32 %f735, %f734;
|
|
mul.f32 %f736, %f52, %f735;
|
|
add.s32 %r149, %r450, 1;
|
|
cvt.s64.s32 %rd62, %r149;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd512, %r451, %r117, %rd62, %rd514, %rd514, %rd514);
|
|
// inline asm
|
|
ld.f32 %f737, [%rd60];
|
|
mul.f32 %f738, %f49, %f737;
|
|
ld.f32 %f739, [%rd60+4];
|
|
mul.f32 %f740, %f50, %f739;
|
|
neg.f32 %f741, %f740;
|
|
sub.f32 %f742, %f741, %f738;
|
|
ld.f32 %f743, [%rd60+8];
|
|
mul.f32 %f744, %f51, %f743;
|
|
sub.f32 %f745, %f742, %f744;
|
|
cvt.sat.f32.f32 %f746, %f745;
|
|
mul.f32 %f67, %f736, %f746;
|
|
mul.f32 %f747, %f1511, %f50;
|
|
fma.rn.f32 %f748, %f1510, %f49, %f747;
|
|
fma.rn.f32 %f68, %f1512, %f51, %f748;
|
|
setp.leu.f32 %p34, %f67, 0f3727C5AC;
|
|
@%p34 bra BB0_21;
|
|
|
|
mov.u32 %r453, 1;
|
|
cvt.sat.f32.f32 %f757, %f68;
|
|
add.u64 %rd67, %SP, 16;
|
|
cvta.to.local.u64 %rd68, %rd67;
|
|
max.f32 %f755, %f762, %f600;
|
|
sub.f32 %f756, %f48, %f755;
|
|
mov.u32 %r153, 1065353216;
|
|
st.local.u32 [%rd68+8], %r153;
|
|
st.local.u32 [%rd68+4], %r153;
|
|
st.local.u32 [%rd68], %r153;
|
|
ld.global.u32 %r150, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r150, %f10, %f11, %f12, %f49, %f50, %f51, %r453, %f755, %f756, %rd67, %r117);
|
|
// inline asm
|
|
ld.local.f32 %f764, [%rd68];
|
|
mul.f32 %f765, %f67, 0f3EA2F983;
|
|
mul.f32 %f766, %f765, %f764;
|
|
ld.local.f32 %f767, [%rd68+4];
|
|
mul.f32 %f768, %f765, %f767;
|
|
ld.local.f32 %f769, [%rd68+8];
|
|
mul.f32 %f770, %f765, %f769;
|
|
fma.rn.f32 %f20, %f757, %f766, %f20;
|
|
fma.rn.f32 %f19, %f757, %f768, %f19;
|
|
fma.rn.f32 %f18, %f757, %f770, %f18;
|
|
add.f32 %f23, %f23, %f766;
|
|
add.f32 %f22, %f22, %f768;
|
|
add.f32 %f21, %f21, %f770;
|
|
fma.rn.f32 %f26, %f50, %f766, %f26;
|
|
fma.rn.f32 %f25, %f50, %f768, %f25;
|
|
fma.rn.f32 %f24, %f50, %f770, %f24;
|
|
fma.rn.f32 %f29, %f51, %f766, %f29;
|
|
fma.rn.f32 %f28, %f51, %f768, %f28;
|
|
fma.rn.f32 %f27, %f51, %f770, %f27;
|
|
fma.rn.f32 %f32, %f49, %f766, %f32;
|
|
fma.rn.f32 %f31, %f49, %f768, %f31;
|
|
fma.rn.f32 %f30, %f49, %f770, %f30;
|
|
mul.f32 %f771, %f49, %f50;
|
|
fma.rn.f32 %f35, %f771, %f766, %f35;
|
|
fma.rn.f32 %f34, %f771, %f768, %f34;
|
|
fma.rn.f32 %f33, %f771, %f770, %f33;
|
|
mul.f32 %f772, %f50, %f51;
|
|
fma.rn.f32 %f38, %f772, %f766, %f38;
|
|
fma.rn.f32 %f37, %f772, %f768, %f37;
|
|
fma.rn.f32 %f36, %f772, %f770, %f36;
|
|
fma.rn.f32 %f773, %f51, %f51, 0fBEAAAAAB;
|
|
fma.rn.f32 %f41, %f773, %f766, %f41;
|
|
fma.rn.f32 %f40, %f773, %f768, %f40;
|
|
fma.rn.f32 %f39, %f773, %f770, %f39;
|
|
mul.f32 %f774, %f49, %f51;
|
|
fma.rn.f32 %f44, %f774, %f766, %f44;
|
|
fma.rn.f32 %f43, %f774, %f768, %f43;
|
|
fma.rn.f32 %f42, %f774, %f770, %f42;
|
|
mul.f32 %f775, %f50, %f50;
|
|
mul.f32 %f776, %f49, %f49;
|
|
sub.f32 %f777, %f776, %f775;
|
|
fma.rn.f32 %f47, %f777, %f766, %f47;
|
|
fma.rn.f32 %f46, %f777, %f768, %f46;
|
|
fma.rn.f32 %f45, %f777, %f770, %f45;
|
|
add.f32 %f17, %f17, %f764;
|
|
|
|
BB0_21:
|
|
add.s32 %r455, %r455, 1;
|
|
setp.lt.s32 %p36, %r455, %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;
|
|
mov.f32 %f33, %f17;
|
|
mov.f32 %f34, %f17;
|
|
mov.f32 %f35, %f17;
|
|
mov.f32 %f36, %f17;
|
|
mov.f32 %f37, %f17;
|
|
mov.f32 %f38, %f17;
|
|
mov.f32 %f39, %f17;
|
|
mov.f32 %f40, %f17;
|
|
mov.f32 %f41, %f17;
|
|
mov.f32 %f42, %f17;
|
|
mov.f32 %f43, %f17;
|
|
mov.f32 %f44, %f17;
|
|
mov.f32 %f45, %f17;
|
|
mov.f32 %f46, %f17;
|
|
mov.f32 %f47, %f17;
|
|
|
|
BB0_22:
|
|
mov.pred %p159, %p13;
|
|
|
|
BB0_23:
|
|
cvt.rn.f32.s32 %f778, %r455;
|
|
mov.f32 %f779, 0f3F800000;
|
|
max.f32 %f780, %f778, %f779;
|
|
rcp.rn.f32 %f781, %f780;
|
|
mul.f32 %f1744, %f20, %f781;
|
|
mul.f32 %f1745, %f19, %f781;
|
|
mul.f32 %f1746, %f18, %f781;
|
|
div.rn.f32 %f1747, %f17, %f780;
|
|
mul.f32 %f1741, %f23, %f781;
|
|
mul.f32 %f1742, %f22, %f781;
|
|
mul.f32 %f1743, %f21, %f781;
|
|
mul.f32 %f1738, %f26, %f781;
|
|
mul.f32 %f1739, %f25, %f781;
|
|
mul.f32 %f1740, %f24, %f781;
|
|
mul.f32 %f1735, %f29, %f781;
|
|
mul.f32 %f1736, %f28, %f781;
|
|
mul.f32 %f1737, %f27, %f781;
|
|
mul.f32 %f1732, %f32, %f781;
|
|
mul.f32 %f1733, %f31, %f781;
|
|
mul.f32 %f1734, %f30, %f781;
|
|
mul.f32 %f1729, %f35, %f781;
|
|
mul.f32 %f1730, %f34, %f781;
|
|
mul.f32 %f1731, %f33, %f781;
|
|
mul.f32 %f1726, %f38, %f781;
|
|
mul.f32 %f1727, %f37, %f781;
|
|
mul.f32 %f1728, %f36, %f781;
|
|
mul.f32 %f1723, %f41, %f781;
|
|
mul.f32 %f1724, %f40, %f781;
|
|
mul.f32 %f1725, %f39, %f781;
|
|
mul.f32 %f1720, %f44, %f781;
|
|
mul.f32 %f1721, %f43, %f781;
|
|
mul.f32 %f1722, %f42, %f781;
|
|
mul.f32 %f1717, %f47, %f781;
|
|
mul.f32 %f1718, %f46, %f781;
|
|
mul.f32 %f1719, %f45, %f781;
|
|
@!%p159 bra BB0_78;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_24:
|
|
mov.f32 %f1640, 0f00000000;
|
|
setp.lt.s32 %p37, %r3, 1;
|
|
mov.f32 %f1641, %f1640;
|
|
mov.f32 %f1642, %f1640;
|
|
mov.f32 %f1643, %f1640;
|
|
mov.f32 %f1644, %f1640;
|
|
mov.f32 %f1645, %f1640;
|
|
mov.f32 %f1646, %f1640;
|
|
mov.f32 %f1647, %f1640;
|
|
mov.f32 %f1648, %f1640;
|
|
mov.f32 %f1649, %f1640;
|
|
mov.f32 %f1650, %f1640;
|
|
mov.f32 %f1651, %f1640;
|
|
mov.f32 %f1652, %f1640;
|
|
mov.f32 %f1653, %f1640;
|
|
mov.f32 %f1654, %f1640;
|
|
mov.f32 %f1655, %f1640;
|
|
mov.f32 %f1656, %f1640;
|
|
mov.f32 %f1657, %f1640;
|
|
mov.f32 %f1658, %f1640;
|
|
mov.f32 %f1659, %f1640;
|
|
mov.f32 %f1660, %f1640;
|
|
mov.f32 %f1661, %f1640;
|
|
mov.f32 %f1662, %f1640;
|
|
mov.f32 %f1663, %f1640;
|
|
mov.f32 %f1664, %f1640;
|
|
mov.f32 %f1665, %f1640;
|
|
mov.f32 %f1666, %f1640;
|
|
mov.f32 %f1667, %f1640;
|
|
mov.f32 %f1668, %f1640;
|
|
mov.f32 %f1669, %f1640;
|
|
mov.f32 %f1670, %f1640;
|
|
@%p37 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f844, %r4;
|
|
rcp.rn.f32 %f193, %f844;
|
|
add.u64 %rd69, %SP, 28;
|
|
cvta.to.local.u64 %rd3, %rd69;
|
|
mul.f32 %f194, %f10, 0f3456BF95;
|
|
mul.f32 %f195, %f11, 0f3456BF95;
|
|
mul.f32 %f196, %f12, 0f3456BF95;
|
|
add.u64 %rd70, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd70;
|
|
mov.f32 %f1640, 0f00000000;
|
|
mov.u32 %r154, 0;
|
|
abs.f32 %f845, %f195;
|
|
abs.f32 %f846, %f194;
|
|
max.f32 %f847, %f846, %f845;
|
|
abs.f32 %f848, %f196;
|
|
max.f32 %f849, %f847, %f848;
|
|
mov.u32 %r456, %r154;
|
|
mov.f32 %f1641, %f1640;
|
|
mov.f32 %f1642, %f1640;
|
|
mov.f32 %f1643, %f1640;
|
|
mov.f32 %f1644, %f1640;
|
|
mov.f32 %f1645, %f1640;
|
|
mov.f32 %f1646, %f1640;
|
|
mov.f32 %f1647, %f1640;
|
|
mov.f32 %f1648, %f1640;
|
|
mov.f32 %f1649, %f1640;
|
|
mov.f32 %f1650, %f1640;
|
|
mov.f32 %f1651, %f1640;
|
|
mov.f32 %f1652, %f1640;
|
|
mov.f32 %f1653, %f1640;
|
|
mov.f32 %f1654, %f1640;
|
|
mov.f32 %f1655, %f1640;
|
|
mov.f32 %f1656, %f1640;
|
|
mov.f32 %f1657, %f1640;
|
|
mov.f32 %f1658, %f1640;
|
|
mov.f32 %f1659, %f1640;
|
|
mov.f32 %f1660, %f1640;
|
|
mov.f32 %f1661, %f1640;
|
|
mov.f32 %f1662, %f1640;
|
|
mov.f32 %f1663, %f1640;
|
|
mov.f32 %f1664, %f1640;
|
|
mov.f32 %f1665, %f1640;
|
|
mov.f32 %f1666, %f1640;
|
|
mov.f32 %f1667, %f1640;
|
|
mov.f32 %f1668, %f1640;
|
|
mov.f32 %f1669, %f1640;
|
|
mov.f32 %f1670, %f1640;
|
|
|
|
BB0_26:
|
|
cvt.rn.f32.s32 %f228, %r456;
|
|
max.f32 %f229, %f849, %f600;
|
|
mov.u32 %r458, %r154;
|
|
|
|
BB0_27:
|
|
mad.lo.s32 %r156, %r459, 1664525, 1013904223;
|
|
and.b32 %r157, %r156, 16777215;
|
|
cvt.rn.f32.u32 %f851, %r157;
|
|
fma.rn.f32 %f852, %f851, 0f33800000, %f228;
|
|
mul.f32 %f853, %f193, %f852;
|
|
mad.lo.s32 %r459, %r156, 1664525, 1013904223;
|
|
and.b32 %r158, %r459, 16777215;
|
|
cvt.rn.f32.u32 %f854, %r158;
|
|
cvt.rn.f32.s32 %f855, %r458;
|
|
fma.rn.f32 %f856, %f854, 0f33800000, %f855;
|
|
mul.f32 %f857, %f193, %f856;
|
|
fma.rn.f32 %f261, %f853, 0fC0000000, 0f3F800000;
|
|
mul.f32 %f262, %f261, %f261;
|
|
sub.f32 %f859, %f779, %f262;
|
|
mov.f32 %f1683, 0f00000000;
|
|
max.f32 %f861, %f1683, %f859;
|
|
sqrt.rn.f32 %f263, %f861;
|
|
mul.f32 %f1677, %f857, 0f40C90FDB;
|
|
abs.f32 %f265, %f1677;
|
|
setp.neu.f32 %p38, %f265, 0f7F800000;
|
|
mov.f32 %f1671, %f1677;
|
|
@%p38 bra BB0_29;
|
|
|
|
mov.f32 %f862, 0f00000000;
|
|
mul.rn.f32 %f1671, %f1677, %f862;
|
|
|
|
BB0_29:
|
|
mul.f32 %f863, %f1671, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r469, %f863;
|
|
cvt.rn.f32.s32 %f864, %r469;
|
|
neg.f32 %f865, %f864;
|
|
mov.f32 %f866, 0f3FC90FDA;
|
|
fma.rn.f32 %f867, %f865, %f866, %f1671;
|
|
mov.f32 %f868, 0f33A22168;
|
|
fma.rn.f32 %f869, %f865, %f868, %f867;
|
|
mov.f32 %f870, 0f27C234C5;
|
|
fma.rn.f32 %f1672, %f865, %f870, %f869;
|
|
abs.f32 %f871, %f1671;
|
|
setp.leu.f32 %p39, %f871, 0f47CE4780;
|
|
@%p39 bra BB0_40;
|
|
|
|
mov.b32 %r19, %f1671;
|
|
shr.u32 %r20, %r19, 23;
|
|
shl.b32 %r161, %r19, 8;
|
|
or.b32 %r21, %r161, -2147483648;
|
|
mov.u32 %r460, 0;
|
|
mov.u64 %rd516, 0;
|
|
mov.u64 %rd515, %rd3;
|
|
mov.u32 %r461, %r460;
|
|
|
|
BB0_31:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd72, %rd516, 2;
|
|
mov.u64 %rd73, __cudart_i2opi_f;
|
|
add.s64 %rd74, %rd73, %rd72;
|
|
ld.const.u32 %r164, [%rd74];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r162, %r164, %r21, %r461;
|
|
madc.hi.u32 %r461, %r164, %r21, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd515], %r162;
|
|
add.s32 %r460, %r460, 1;
|
|
cvt.s64.s32 %rd516, %r460;
|
|
mul.wide.s32 %rd75, %r460, 4;
|
|
add.s64 %rd515, %rd3, %rd75;
|
|
setp.ne.s32 %p40, %r460, 6;
|
|
@%p40 bra BB0_31;
|
|
|
|
and.b32 %r167, %r20, 255;
|
|
add.s32 %r168, %r167, -128;
|
|
shr.u32 %r169, %r168, 5;
|
|
and.b32 %r26, %r19, -2147483648;
|
|
st.local.u32 [%rd3+24], %r461;
|
|
mov.u32 %r170, 6;
|
|
sub.s32 %r171, %r170, %r169;
|
|
mul.wide.s32 %rd76, %r171, 4;
|
|
add.s64 %rd10, %rd3, %rd76;
|
|
ld.local.u32 %r462, [%rd10];
|
|
ld.local.u32 %r463, [%rd10+-4];
|
|
and.b32 %r29, %r20, 31;
|
|
setp.eq.s32 %p41, %r29, 0;
|
|
@%p41 bra BB0_34;
|
|
|
|
mov.u32 %r172, 32;
|
|
sub.s32 %r173, %r172, %r29;
|
|
shr.u32 %r174, %r463, %r173;
|
|
shl.b32 %r175, %r462, %r29;
|
|
add.s32 %r462, %r174, %r175;
|
|
ld.local.u32 %r176, [%rd10+-8];
|
|
shr.u32 %r177, %r176, %r173;
|
|
shl.b32 %r178, %r463, %r29;
|
|
add.s32 %r463, %r177, %r178;
|
|
|
|
BB0_34:
|
|
shr.u32 %r179, %r463, 30;
|
|
shl.b32 %r180, %r462, 2;
|
|
add.s32 %r464, %r179, %r180;
|
|
shl.b32 %r35, %r463, 2;
|
|
shr.u32 %r181, %r464, 31;
|
|
shr.u32 %r182, %r462, 30;
|
|
add.s32 %r36, %r181, %r182;
|
|
setp.eq.s32 %p42, %r181, 0;
|
|
@%p42 bra BB0_35;
|
|
bra.uni BB0_36;
|
|
|
|
BB0_35:
|
|
mov.u32 %r465, %r26;
|
|
mov.u32 %r466, %r35;
|
|
bra.uni BB0_37;
|
|
|
|
BB0_36:
|
|
not.b32 %r183, %r464;
|
|
neg.s32 %r466, %r35;
|
|
setp.eq.s32 %p43, %r35, 0;
|
|
selp.u32 %r184, 1, 0, %p43;
|
|
add.s32 %r464, %r184, %r183;
|
|
xor.b32 %r465, %r26, -2147483648;
|
|
|
|
BB0_37:
|
|
clz.b32 %r468, %r464;
|
|
setp.eq.s32 %p44, %r468, 0;
|
|
shl.b32 %r185, %r464, %r468;
|
|
mov.u32 %r186, 32;
|
|
sub.s32 %r187, %r186, %r468;
|
|
shr.u32 %r188, %r466, %r187;
|
|
add.s32 %r189, %r188, %r185;
|
|
selp.b32 %r44, %r464, %r189, %p44;
|
|
mov.u32 %r190, -921707870;
|
|
mul.hi.u32 %r467, %r44, %r190;
|
|
setp.eq.s32 %p45, %r26, 0;
|
|
neg.s32 %r191, %r36;
|
|
selp.b32 %r469, %r36, %r191, %p45;
|
|
setp.lt.s32 %p46, %r467, 1;
|
|
@%p46 bra BB0_39;
|
|
|
|
mul.lo.s32 %r192, %r44, -921707870;
|
|
shr.u32 %r193, %r192, 31;
|
|
shl.b32 %r194, %r467, 1;
|
|
add.s32 %r467, %r193, %r194;
|
|
add.s32 %r468, %r468, 1;
|
|
|
|
BB0_39:
|
|
mov.u32 %r195, 126;
|
|
sub.s32 %r196, %r195, %r468;
|
|
shl.b32 %r197, %r196, 23;
|
|
add.s32 %r198, %r467, 1;
|
|
shr.u32 %r199, %r198, 7;
|
|
add.s32 %r200, %r199, 1;
|
|
shr.u32 %r201, %r200, 1;
|
|
add.s32 %r202, %r201, %r197;
|
|
or.b32 %r203, %r202, %r465;
|
|
mov.b32 %f1672, %r203;
|
|
|
|
BB0_40:
|
|
mul.rn.f32 %f271, %f1672, %f1672;
|
|
add.s32 %r52, %r469, 1;
|
|
and.b32 %r53, %r52, 1;
|
|
setp.eq.s32 %p47, %r53, 0;
|
|
@%p47 bra BB0_42;
|
|
bra.uni BB0_41;
|
|
|
|
BB0_42:
|
|
mov.f32 %f874, 0f3C08839E;
|
|
mov.f32 %f875, 0fB94CA1F9;
|
|
fma.rn.f32 %f1673, %f875, %f271, %f874;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_41:
|
|
mov.f32 %f872, 0fBAB6061A;
|
|
mov.f32 %f873, 0f37CCF5CE;
|
|
fma.rn.f32 %f1673, %f873, %f271, %f872;
|
|
|
|
BB0_43:
|
|
@%p47 bra BB0_45;
|
|
bra.uni BB0_44;
|
|
|
|
BB0_45:
|
|
mov.f32 %f879, 0fBE2AAAA3;
|
|
fma.rn.f32 %f880, %f1673, %f271, %f879;
|
|
mov.f32 %f881, 0f00000000;
|
|
fma.rn.f32 %f1674, %f880, %f271, %f881;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_44:
|
|
mov.f32 %f876, 0f3D2AAAA5;
|
|
fma.rn.f32 %f877, %f1673, %f271, %f876;
|
|
mov.f32 %f878, 0fBF000000;
|
|
fma.rn.f32 %f1674, %f877, %f271, %f878;
|
|
|
|
BB0_46:
|
|
fma.rn.f32 %f1675, %f1674, %f1672, %f1672;
|
|
@%p47 bra BB0_48;
|
|
|
|
fma.rn.f32 %f1675, %f1674, %f271, %f779;
|
|
|
|
BB0_48:
|
|
and.b32 %r204, %r52, 2;
|
|
setp.eq.s32 %p50, %r204, 0;
|
|
@%p50 bra BB0_50;
|
|
|
|
mov.f32 %f884, 0fBF800000;
|
|
fma.rn.f32 %f1675, %f1675, %f884, %f1683;
|
|
|
|
BB0_50:
|
|
mul.f32 %f283, %f263, %f1675;
|
|
@%p38 bra BB0_52;
|
|
|
|
mov.f32 %f885, 0f00000000;
|
|
mul.rn.f32 %f1677, %f1677, %f885;
|
|
|
|
BB0_52:
|
|
mul.f32 %f886, %f1677, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r479, %f886;
|
|
cvt.rn.f32.s32 %f887, %r479;
|
|
neg.f32 %f888, %f887;
|
|
fma.rn.f32 %f890, %f888, %f866, %f1677;
|
|
fma.rn.f32 %f892, %f888, %f868, %f890;
|
|
fma.rn.f32 %f1678, %f888, %f870, %f892;
|
|
abs.f32 %f894, %f1677;
|
|
setp.leu.f32 %p52, %f894, 0f47CE4780;
|
|
@%p52 bra BB0_63;
|
|
|
|
mov.b32 %r55, %f1677;
|
|
shr.u32 %r56, %r55, 23;
|
|
shl.b32 %r207, %r55, 8;
|
|
or.b32 %r57, %r207, -2147483648;
|
|
mov.u32 %r470, 0;
|
|
mov.u64 %rd517, %rd3;
|
|
mov.u64 %rd518, %rd28;
|
|
mov.u32 %r471, %r470;
|
|
|
|
BB0_54:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd518, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r210, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r208, %r210, %r57, %r471;
|
|
madc.hi.u32 %r471, %r210, %r57, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd517], %r208;
|
|
add.s32 %r470, %r470, 1;
|
|
cvt.s64.s32 %rd518, %r470;
|
|
mul.wide.s32 %rd81, %r470, 4;
|
|
add.s64 %rd517, %rd3, %rd81;
|
|
setp.ne.s32 %p53, %r470, 6;
|
|
@%p53 bra BB0_54;
|
|
|
|
and.b32 %r213, %r56, 255;
|
|
add.s32 %r214, %r213, -128;
|
|
shr.u32 %r215, %r214, 5;
|
|
and.b32 %r62, %r55, -2147483648;
|
|
st.local.u32 [%rd3+24], %r471;
|
|
mov.u32 %r216, 6;
|
|
sub.s32 %r217, %r216, %r215;
|
|
mul.wide.s32 %rd82, %r217, 4;
|
|
add.s64 %rd16, %rd3, %rd82;
|
|
ld.local.u32 %r472, [%rd16];
|
|
ld.local.u32 %r473, [%rd16+-4];
|
|
and.b32 %r65, %r56, 31;
|
|
setp.eq.s32 %p54, %r65, 0;
|
|
@%p54 bra BB0_57;
|
|
|
|
mov.u32 %r218, 32;
|
|
sub.s32 %r219, %r218, %r65;
|
|
shr.u32 %r220, %r473, %r219;
|
|
shl.b32 %r221, %r472, %r65;
|
|
add.s32 %r472, %r220, %r221;
|
|
ld.local.u32 %r222, [%rd16+-8];
|
|
shr.u32 %r223, %r222, %r219;
|
|
shl.b32 %r224, %r473, %r65;
|
|
add.s32 %r473, %r223, %r224;
|
|
|
|
BB0_57:
|
|
shr.u32 %r225, %r473, 30;
|
|
shl.b32 %r226, %r472, 2;
|
|
add.s32 %r474, %r225, %r226;
|
|
shl.b32 %r71, %r473, 2;
|
|
shr.u32 %r227, %r474, 31;
|
|
shr.u32 %r228, %r472, 30;
|
|
add.s32 %r72, %r227, %r228;
|
|
setp.eq.s32 %p55, %r227, 0;
|
|
@%p55 bra BB0_58;
|
|
bra.uni BB0_59;
|
|
|
|
BB0_58:
|
|
mov.u32 %r475, %r62;
|
|
mov.u32 %r476, %r71;
|
|
bra.uni BB0_60;
|
|
|
|
BB0_59:
|
|
not.b32 %r229, %r474;
|
|
neg.s32 %r476, %r71;
|
|
setp.eq.s32 %p56, %r71, 0;
|
|
selp.u32 %r230, 1, 0, %p56;
|
|
add.s32 %r474, %r230, %r229;
|
|
xor.b32 %r475, %r62, -2147483648;
|
|
|
|
BB0_60:
|
|
clz.b32 %r478, %r474;
|
|
setp.eq.s32 %p57, %r478, 0;
|
|
shl.b32 %r231, %r474, %r478;
|
|
mov.u32 %r232, 32;
|
|
sub.s32 %r233, %r232, %r478;
|
|
shr.u32 %r234, %r476, %r233;
|
|
add.s32 %r235, %r234, %r231;
|
|
selp.b32 %r80, %r474, %r235, %p57;
|
|
mov.u32 %r236, -921707870;
|
|
mul.hi.u32 %r477, %r80, %r236;
|
|
setp.eq.s32 %p58, %r62, 0;
|
|
neg.s32 %r237, %r72;
|
|
selp.b32 %r479, %r72, %r237, %p58;
|
|
setp.lt.s32 %p59, %r477, 1;
|
|
@%p59 bra BB0_62;
|
|
|
|
mul.lo.s32 %r238, %r80, -921707870;
|
|
shr.u32 %r239, %r238, 31;
|
|
shl.b32 %r240, %r477, 1;
|
|
add.s32 %r477, %r239, %r240;
|
|
add.s32 %r478, %r478, 1;
|
|
|
|
BB0_62:
|
|
mov.u32 %r241, 126;
|
|
sub.s32 %r242, %r241, %r478;
|
|
shl.b32 %r243, %r242, 23;
|
|
add.s32 %r244, %r477, 1;
|
|
shr.u32 %r245, %r244, 7;
|
|
add.s32 %r246, %r245, 1;
|
|
shr.u32 %r247, %r246, 1;
|
|
add.s32 %r248, %r247, %r243;
|
|
or.b32 %r249, %r248, %r475;
|
|
mov.b32 %f1678, %r249;
|
|
|
|
BB0_63:
|
|
mul.rn.f32 %f289, %f1678, %f1678;
|
|
and.b32 %r88, %r479, 1;
|
|
setp.eq.s32 %p60, %r88, 0;
|
|
@%p60 bra BB0_65;
|
|
bra.uni BB0_64;
|
|
|
|
BB0_65:
|
|
mov.f32 %f897, 0f3C08839E;
|
|
mov.f32 %f898, 0fB94CA1F9;
|
|
fma.rn.f32 %f1679, %f898, %f289, %f897;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_64:
|
|
mov.f32 %f895, 0fBAB6061A;
|
|
mov.f32 %f896, 0f37CCF5CE;
|
|
fma.rn.f32 %f1679, %f896, %f289, %f895;
|
|
|
|
BB0_66:
|
|
@%p60 bra BB0_68;
|
|
bra.uni BB0_67;
|
|
|
|
BB0_68:
|
|
mov.f32 %f902, 0fBE2AAAA3;
|
|
fma.rn.f32 %f903, %f1679, %f289, %f902;
|
|
mov.f32 %f904, 0f00000000;
|
|
fma.rn.f32 %f1680, %f903, %f289, %f904;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_67:
|
|
mov.f32 %f899, 0f3D2AAAA5;
|
|
fma.rn.f32 %f900, %f1679, %f289, %f899;
|
|
mov.f32 %f901, 0fBF000000;
|
|
fma.rn.f32 %f1680, %f900, %f289, %f901;
|
|
|
|
BB0_69:
|
|
fma.rn.f32 %f1681, %f1680, %f1678, %f1678;
|
|
@%p60 bra BB0_71;
|
|
|
|
fma.rn.f32 %f1681, %f1680, %f289, %f779;
|
|
|
|
BB0_71:
|
|
and.b32 %r250, %r479, 2;
|
|
setp.eq.s32 %p63, %r250, 0;
|
|
@%p63 bra BB0_73;
|
|
|
|
mov.f32 %f907, 0fBF800000;
|
|
fma.rn.f32 %f1681, %f1681, %f907, %f1683;
|
|
|
|
BB0_73:
|
|
mov.f32 %f919, 0f7F7FFFFF;
|
|
st.local.v4.f32 [%rd4], {%f779, %f779, %f779, %f919};
|
|
ld.global.u32 %r251, [root];
|
|
mul.f32 %f912, %f263, %f1681;
|
|
mov.u32 %r252, 0;
|
|
mov.f32 %f915, 0f6C4ECB8F;
|
|
mov.u32 %r253, 16;
|
|
// inline asm
|
|
call _rt_trace_64, (%r251, %f10, %f11, %f12, %f283, %f912, %f261, %r252, %f229, %f915, %rd70, %r253);
|
|
// inline asm
|
|
ld.local.f32 %f302, [%rd4+12];
|
|
setp.ge.f32 %p64, %f302, 0f00000000;
|
|
mov.f32 %f1684, %f1683;
|
|
mov.f32 %f1685, %f1683;
|
|
@%p64 bra BB0_75;
|
|
|
|
ld.local.v4.f32 {%f1685, %f1684, %f1683, %f924}, [%rd4];
|
|
|
|
BB0_75:
|
|
setp.ltu.f32 %p65, %f302, 0f00000000;
|
|
selp.f32 %f925, 0f3F800000, 0f00000000, %p65;
|
|
add.f32 %f1640, %f1640, %f925;
|
|
add.f32 %f1646, %f1646, %f1685;
|
|
add.f32 %f1645, %f1645, %f1684;
|
|
add.f32 %f1644, %f1644, %f1683;
|
|
fma.rn.f32 %f1649, %f912, %f1685, %f1649;
|
|
fma.rn.f32 %f1648, %f912, %f1684, %f1648;
|
|
fma.rn.f32 %f1647, %f912, %f1683, %f1647;
|
|
fma.rn.f32 %f1652, %f261, %f1685, %f1652;
|
|
fma.rn.f32 %f1651, %f261, %f1684, %f1651;
|
|
fma.rn.f32 %f1650, %f261, %f1683, %f1650;
|
|
fma.rn.f32 %f1655, %f283, %f1685, %f1655;
|
|
fma.rn.f32 %f1654, %f283, %f1684, %f1654;
|
|
fma.rn.f32 %f1653, %f283, %f1683, %f1653;
|
|
mul.f32 %f926, %f283, %f912;
|
|
fma.rn.f32 %f1658, %f926, %f1685, %f1658;
|
|
fma.rn.f32 %f1657, %f926, %f1684, %f1657;
|
|
fma.rn.f32 %f1656, %f926, %f1683, %f1656;
|
|
mul.f32 %f927, %f261, %f912;
|
|
fma.rn.f32 %f1661, %f927, %f1685, %f1661;
|
|
fma.rn.f32 %f1660, %f927, %f1684, %f1660;
|
|
fma.rn.f32 %f1659, %f927, %f1683, %f1659;
|
|
add.f32 %f928, %f262, 0fBEAAAAAB;
|
|
fma.rn.f32 %f1664, %f928, %f1685, %f1664;
|
|
fma.rn.f32 %f1663, %f928, %f1684, %f1663;
|
|
fma.rn.f32 %f1662, %f928, %f1683, %f1662;
|
|
mul.f32 %f929, %f261, %f283;
|
|
fma.rn.f32 %f1667, %f929, %f1685, %f1667;
|
|
fma.rn.f32 %f1666, %f929, %f1684, %f1666;
|
|
fma.rn.f32 %f1665, %f929, %f1683, %f1665;
|
|
mul.f32 %f930, %f912, %f912;
|
|
mul.f32 %f931, %f283, %f283;
|
|
sub.f32 %f932, %f931, %f930;
|
|
fma.rn.f32 %f1670, %f932, %f1685, %f1670;
|
|
fma.rn.f32 %f1669, %f932, %f1684, %f1669;
|
|
fma.rn.f32 %f1668, %f932, %f1683, %f1668;
|
|
mul.f32 %f933, %f1511, %f912;
|
|
fma.rn.f32 %f934, %f1510, %f283, %f933;
|
|
fma.rn.f32 %f935, %f1512, %f261, %f934;
|
|
cvt.sat.f32.f32 %f936, %f935;
|
|
fma.rn.f32 %f1643, %f1685, %f936, %f1643;
|
|
fma.rn.f32 %f1642, %f1684, %f936, %f1642;
|
|
fma.rn.f32 %f1641, %f1683, %f936, %f1641;
|
|
add.s32 %r458, %r458, 1;
|
|
setp.lt.s32 %p66, %r458, %r4;
|
|
@%p66 bra BB0_27;
|
|
|
|
add.s32 %r456, %r456, 1;
|
|
setp.lt.s32 %p67, %r456, %r4;
|
|
@%p67 bra BB0_26;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r254, %r4, %r4;
|
|
cvt.rn.f32.s32 %f937, %r254;
|
|
rcp.rn.f32 %f938, %f937;
|
|
mul.f32 %f939, %f1643, %f938;
|
|
mul.f32 %f940, %f1642, %f938;
|
|
mul.f32 %f941, %f1641, %f938;
|
|
div.rn.f32 %f1747, %f1640, %f937;
|
|
mul.f32 %f1741, %f1646, %f938;
|
|
mul.f32 %f1742, %f1645, %f938;
|
|
mul.f32 %f1743, %f1644, %f938;
|
|
mul.f32 %f1738, %f1649, %f938;
|
|
mul.f32 %f1739, %f1648, %f938;
|
|
mul.f32 %f1740, %f1647, %f938;
|
|
mul.f32 %f1735, %f1652, %f938;
|
|
mul.f32 %f1736, %f1651, %f938;
|
|
mul.f32 %f1737, %f1650, %f938;
|
|
mul.f32 %f1732, %f1655, %f938;
|
|
mul.f32 %f1733, %f1654, %f938;
|
|
mul.f32 %f1734, %f1653, %f938;
|
|
mul.f32 %f1729, %f1658, %f938;
|
|
mul.f32 %f1730, %f1657, %f938;
|
|
mul.f32 %f1731, %f1656, %f938;
|
|
mul.f32 %f1726, %f1661, %f938;
|
|
mul.f32 %f1727, %f1660, %f938;
|
|
mul.f32 %f1728, %f1659, %f938;
|
|
mul.f32 %f1723, %f1664, %f938;
|
|
mul.f32 %f1724, %f1663, %f938;
|
|
mul.f32 %f1725, %f1662, %f938;
|
|
mul.f32 %f1720, %f1667, %f938;
|
|
mul.f32 %f1721, %f1666, %f938;
|
|
mul.f32 %f1722, %f1665, %f938;
|
|
mul.f32 %f1717, %f1670, %f938;
|
|
mul.f32 %f1718, %f1669, %f938;
|
|
mul.f32 %f1719, %f1668, %f938;
|
|
fma.rn.f32 %f1744, %f1643, %f938, %f939;
|
|
fma.rn.f32 %f1745, %f1642, %f938, %f940;
|
|
fma.rn.f32 %f1746, %f1641, %f938, %f941;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p68, %r1, -1;
|
|
selp.f32 %f433, %f1747, %f1744, %p68;
|
|
ld.global.u32 %r481, [imageEnabled];
|
|
and.b32 %r255, %r481, 8;
|
|
setp.eq.s32 %p69, %r255, 0;
|
|
@%p69 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd86, %r6;
|
|
cvt.u64.u32 %rd87, %r7;
|
|
mov.u64 %rd90, image_Mask;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r101, %r101, %rd86, %rd87, %rd28, %rd28);
|
|
// inline asm
|
|
mov.f32 %f944, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f945, %f944;
|
|
fma.rn.f32 %f946, %f945, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f434, %f946;
|
|
abs.f32 %f435, %f433;
|
|
setp.lt.f32 %p70, %f435, 0f00800000;
|
|
mul.f32 %f947, %f435, 0f4B800000;
|
|
selp.f32 %f948, 0fC3170000, 0fC2FE0000, %p70;
|
|
selp.f32 %f949, %f947, %f435, %p70;
|
|
mov.b32 %r258, %f949;
|
|
and.b32 %r259, %r258, 8388607;
|
|
or.b32 %r260, %r259, 1065353216;
|
|
mov.b32 %f950, %r260;
|
|
shr.u32 %r261, %r258, 23;
|
|
cvt.rn.f32.u32 %f951, %r261;
|
|
add.f32 %f952, %f948, %f951;
|
|
setp.gt.f32 %p71, %f950, 0f3FB504F3;
|
|
mul.f32 %f953, %f950, 0f3F000000;
|
|
add.f32 %f954, %f952, 0f3F800000;
|
|
selp.f32 %f955, %f953, %f950, %p71;
|
|
selp.f32 %f956, %f954, %f952, %p71;
|
|
add.f32 %f957, %f955, 0fBF800000;
|
|
add.f32 %f943, %f955, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f942,%f943;
|
|
// inline asm
|
|
add.f32 %f958, %f957, %f957;
|
|
mul.f32 %f959, %f942, %f958;
|
|
mul.f32 %f960, %f959, %f959;
|
|
mov.f32 %f961, 0f3C4CAF63;
|
|
mov.f32 %f962, 0f3B18F0FE;
|
|
fma.rn.f32 %f963, %f962, %f960, %f961;
|
|
mov.f32 %f964, 0f3DAAAABD;
|
|
fma.rn.f32 %f965, %f963, %f960, %f964;
|
|
mul.rn.f32 %f966, %f965, %f960;
|
|
mul.rn.f32 %f967, %f966, %f959;
|
|
sub.f32 %f968, %f957, %f959;
|
|
neg.f32 %f969, %f959;
|
|
add.f32 %f970, %f968, %f968;
|
|
fma.rn.f32 %f971, %f969, %f957, %f970;
|
|
mul.rn.f32 %f972, %f942, %f971;
|
|
add.f32 %f973, %f967, %f959;
|
|
sub.f32 %f974, %f959, %f973;
|
|
add.f32 %f975, %f967, %f974;
|
|
add.f32 %f976, %f972, %f975;
|
|
add.f32 %f977, %f973, %f976;
|
|
sub.f32 %f978, %f973, %f977;
|
|
add.f32 %f979, %f976, %f978;
|
|
mov.f32 %f980, 0f3F317200;
|
|
mul.rn.f32 %f981, %f956, %f980;
|
|
mov.f32 %f982, 0f35BFBE8E;
|
|
mul.rn.f32 %f983, %f956, %f982;
|
|
add.f32 %f984, %f981, %f977;
|
|
sub.f32 %f985, %f981, %f984;
|
|
add.f32 %f986, %f977, %f985;
|
|
add.f32 %f987, %f979, %f986;
|
|
add.f32 %f988, %f983, %f987;
|
|
add.f32 %f989, %f984, %f988;
|
|
sub.f32 %f990, %f984, %f989;
|
|
add.f32 %f991, %f988, %f990;
|
|
mov.f32 %f992, 0f3EE8BA2E;
|
|
mul.rn.f32 %f993, %f992, %f989;
|
|
neg.f32 %f994, %f993;
|
|
fma.rn.f32 %f995, %f992, %f989, %f994;
|
|
fma.rn.f32 %f996, %f992, %f991, %f995;
|
|
mov.f32 %f997, 0f00000000;
|
|
fma.rn.f32 %f998, %f997, %f989, %f996;
|
|
add.rn.f32 %f999, %f993, %f998;
|
|
neg.f32 %f1000, %f999;
|
|
add.rn.f32 %f1001, %f993, %f1000;
|
|
add.rn.f32 %f1002, %f1001, %f998;
|
|
mov.b32 %r262, %f999;
|
|
setp.eq.s32 %p72, %r262, 1118925336;
|
|
add.s32 %r263, %r262, -1;
|
|
mov.b32 %f1003, %r263;
|
|
add.f32 %f1004, %f1002, 0f37000000;
|
|
selp.f32 %f1005, %f1003, %f999, %p72;
|
|
selp.f32 %f436, %f1004, %f1002, %p72;
|
|
mul.f32 %f1006, %f1005, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1007, %f1006;
|
|
mov.f32 %f1008, 0fBF317200;
|
|
fma.rn.f32 %f1009, %f1007, %f1008, %f1005;
|
|
mov.f32 %f1010, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1011, %f1007, %f1010, %f1009;
|
|
mul.f32 %f1012, %f1011, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1013, %f1012;
|
|
add.f32 %f1014, %f1007, 0f00000000;
|
|
ex2.approx.f32 %f1015, %f1014;
|
|
mul.f32 %f1016, %f1013, %f1015;
|
|
setp.lt.f32 %p73, %f1005, 0fC2D20000;
|
|
selp.f32 %f1017, 0f00000000, %f1016, %p73;
|
|
setp.gt.f32 %p74, %f1005, 0f42D20000;
|
|
selp.f32 %f1748, 0f7F800000, %f1017, %p74;
|
|
setp.eq.f32 %p75, %f1748, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
|
|
fma.rn.f32 %f1748, %f1748, %f436, %f1748;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p76, %f433, 0f00000000;
|
|
setp.eq.f32 %p77, %f434, 0f3F800000;
|
|
and.pred %p3, %p76, %p77;
|
|
mov.b32 %r264, %f1748;
|
|
xor.b32 %r265, %r264, -2147483648;
|
|
mov.b32 %f1018, %r265;
|
|
selp.f32 %f1750, %f1018, %f1748, %p3;
|
|
setp.eq.f32 %p78, %f433, 0f00000000;
|
|
@%p78 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f1021, %f433, %f433;
|
|
selp.f32 %f1750, %f1021, 0f00000000, %p77;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_164:
|
|
mov.u64 %rd331, image_HDR;
|
|
cvta.global.u64 %rd326, %rd331;
|
|
mov.u32 %r386, 8;
|
|
mov.u64 %rd330, 0;
|
|
// inline asm
|
|
call (%rd325), _rt_buffer_get_64, (%rd326, %r101, %r386, %rd21, %rd22, %rd330, %rd330);
|
|
// inline asm
|
|
mov.f32 %f1439, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs180, %f1439;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs181, 0;
|
|
st.v4.u16 [%rd325], {%rs180, %rs180, %rs180, %rs181};
|
|
|
|
BB0_165:
|
|
ld.global.u32 %r387, [additive];
|
|
setp.eq.s32 %p150, %r387, 0;
|
|
@%p150 bra BB0_167;
|
|
|
|
mov.u64 %rd344, image_RNM0;
|
|
cvta.global.u64 %rd333, %rd344;
|
|
mov.u32 %r391, 8;
|
|
mov.u64 %rd343, 0;
|
|
// inline asm
|
|
call (%rd332), _rt_buffer_get_64, (%rd333, %r101, %r391, %rd21, %rd22, %rd343, %rd343);
|
|
// inline asm
|
|
ld.v4.u16 {%rs188, %rs189, %rs190, %rs191}, [%rd332];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1440, %rs188;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1441, %rs189;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1442, %rs190;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd338), _rt_buffer_get_64, (%rd333, %r101, %r391, %rd21, %rd22, %rd343, %rd343);
|
|
// inline asm
|
|
add.f32 %f1443, %f1440, 0f00000000;
|
|
add.f32 %f1444, %f1441, 0f00000000;
|
|
add.f32 %f1445, %f1442, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs187, %f1445;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs186, %f1444;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs185, %f1443;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs192, 0;
|
|
st.v4.u16 [%rd338], {%rs185, %rs186, %rs187, %rs192};
|
|
bra.uni BB0_168;
|
|
|
|
BB0_167:
|
|
mov.u64 %rd351, image_RNM0;
|
|
cvta.global.u64 %rd346, %rd351;
|
|
mov.u32 %r393, 8;
|
|
mov.u64 %rd350, 0;
|
|
// inline asm
|
|
call (%rd345), _rt_buffer_get_64, (%rd346, %r101, %r393, %rd21, %rd22, %rd350, %rd350);
|
|
// inline asm
|
|
mov.f32 %f1446, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs193, %f1446;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs194, 0;
|
|
st.v4.u16 [%rd345], {%rs193, %rs193, %rs193, %rs194};
|
|
|
|
BB0_168:
|
|
ld.global.u32 %r394, [additive];
|
|
setp.eq.s32 %p151, %r394, 0;
|
|
@%p151 bra BB0_170;
|
|
|
|
mov.u64 %rd364, image_RNM1;
|
|
cvta.global.u64 %rd353, %rd364;
|
|
mov.u32 %r398, 8;
|
|
mov.u64 %rd363, 0;
|
|
// inline asm
|
|
call (%rd352), _rt_buffer_get_64, (%rd353, %r101, %r398, %rd21, %rd22, %rd363, %rd363);
|
|
// inline asm
|
|
ld.v4.u16 {%rs201, %rs202, %rs203, %rs204}, [%rd352];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1447, %rs201;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1448, %rs202;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1449, %rs203;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd358), _rt_buffer_get_64, (%rd353, %r101, %r398, %rd21, %rd22, %rd363, %rd363);
|
|
// inline asm
|
|
add.f32 %f1450, %f1447, 0f00000000;
|
|
add.f32 %f1451, %f1448, 0f00000000;
|
|
add.f32 %f1452, %f1449, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs200, %f1452;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs199, %f1451;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs198, %f1450;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs205, 0;
|
|
st.v4.u16 [%rd358], {%rs198, %rs199, %rs200, %rs205};
|
|
bra.uni BB0_171;
|
|
|
|
BB0_170:
|
|
mov.u64 %rd371, image_RNM1;
|
|
cvta.global.u64 %rd366, %rd371;
|
|
mov.u32 %r400, 8;
|
|
mov.u64 %rd370, 0;
|
|
// inline asm
|
|
call (%rd365), _rt_buffer_get_64, (%rd366, %r101, %r400, %rd21, %rd22, %rd370, %rd370);
|
|
// inline asm
|
|
mov.f32 %f1453, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs206, %f1453;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs207, 0;
|
|
st.v4.u16 [%rd365], {%rs206, %rs206, %rs206, %rs207};
|
|
|
|
BB0_171:
|
|
ld.global.u32 %r401, [additive];
|
|
setp.eq.s32 %p152, %r401, 0;
|
|
@%p152 bra BB0_173;
|
|
|
|
mov.u64 %rd384, image_RNM2;
|
|
cvta.global.u64 %rd373, %rd384;
|
|
mov.u32 %r405, 8;
|
|
mov.u64 %rd383, 0;
|
|
// inline asm
|
|
call (%rd372), _rt_buffer_get_64, (%rd373, %r101, %r405, %rd21, %rd22, %rd383, %rd383);
|
|
// inline asm
|
|
ld.v4.u16 {%rs214, %rs215, %rs216, %rs217}, [%rd372];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1454, %rs214;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1455, %rs215;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1456, %rs216;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd378), _rt_buffer_get_64, (%rd373, %r101, %r405, %rd21, %rd22, %rd383, %rd383);
|
|
// inline asm
|
|
add.f32 %f1457, %f1454, 0f00000000;
|
|
add.f32 %f1458, %f1455, 0f00000000;
|
|
add.f32 %f1459, %f1456, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs213, %f1459;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs212, %f1458;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs211, %f1457;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs218, 0;
|
|
st.v4.u16 [%rd378], {%rs211, %rs212, %rs213, %rs218};
|
|
bra.uni BB0_174;
|
|
|
|
BB0_173:
|
|
mov.u64 %rd391, image_RNM2;
|
|
cvta.global.u64 %rd386, %rd391;
|
|
mov.u32 %r407, 8;
|
|
mov.u64 %rd390, 0;
|
|
// inline asm
|
|
call (%rd385), _rt_buffer_get_64, (%rd386, %r101, %r407, %rd21, %rd22, %rd390, %rd390);
|
|
// inline asm
|
|
mov.f32 %f1460, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs219, %f1460;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs220, 0;
|
|
st.v4.u16 [%rd385], {%rs219, %rs219, %rs219, %rs220};
|
|
|
|
BB0_174:
|
|
ld.global.u32 %r408, [additive];
|
|
setp.eq.s32 %p153, %r408, 0;
|
|
@%p153 bra BB0_176;
|
|
|
|
mov.u64 %rd404, image_RNM3;
|
|
cvta.global.u64 %rd393, %rd404;
|
|
mov.u32 %r412, 8;
|
|
mov.u64 %rd403, 0;
|
|
// inline asm
|
|
call (%rd392), _rt_buffer_get_64, (%rd393, %r101, %r412, %rd21, %rd22, %rd403, %rd403);
|
|
// inline asm
|
|
ld.v4.u16 {%rs227, %rs228, %rs229, %rs230}, [%rd392];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1461, %rs227;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1462, %rs228;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1463, %rs229;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd398), _rt_buffer_get_64, (%rd393, %r101, %r412, %rd21, %rd22, %rd403, %rd403);
|
|
// inline asm
|
|
add.f32 %f1464, %f1461, 0f00000000;
|
|
add.f32 %f1465, %f1462, 0f00000000;
|
|
add.f32 %f1466, %f1463, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs226, %f1466;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs225, %f1465;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs224, %f1464;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs231, 0;
|
|
st.v4.u16 [%rd398], {%rs224, %rs225, %rs226, %rs231};
|
|
bra.uni BB0_177;
|
|
|
|
BB0_176:
|
|
mov.u64 %rd411, image_RNM3;
|
|
cvta.global.u64 %rd406, %rd411;
|
|
mov.u32 %r414, 8;
|
|
mov.u64 %rd410, 0;
|
|
// inline asm
|
|
call (%rd405), _rt_buffer_get_64, (%rd406, %r101, %r414, %rd21, %rd22, %rd410, %rd410);
|
|
// inline asm
|
|
mov.f32 %f1467, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs232, %f1467;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs233, 0;
|
|
st.v4.u16 [%rd405], {%rs232, %rs232, %rs232, %rs233};
|
|
|
|
BB0_177:
|
|
ld.global.u32 %r415, [additive];
|
|
setp.eq.s32 %p154, %r415, 0;
|
|
@%p154 bra BB0_179;
|
|
|
|
mov.u64 %rd424, image_RNM4;
|
|
cvta.global.u64 %rd413, %rd424;
|
|
mov.u32 %r419, 8;
|
|
mov.u64 %rd423, 0;
|
|
// inline asm
|
|
call (%rd412), _rt_buffer_get_64, (%rd413, %r101, %r419, %rd21, %rd22, %rd423, %rd423);
|
|
// inline asm
|
|
ld.v4.u16 {%rs240, %rs241, %rs242, %rs243}, [%rd412];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1468, %rs240;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1469, %rs241;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1470, %rs242;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd418), _rt_buffer_get_64, (%rd413, %r101, %r419, %rd21, %rd22, %rd423, %rd423);
|
|
// inline asm
|
|
add.f32 %f1471, %f1468, 0f00000000;
|
|
add.f32 %f1472, %f1469, 0f00000000;
|
|
add.f32 %f1473, %f1470, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs239, %f1473;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs238, %f1472;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs237, %f1471;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs244, 0;
|
|
st.v4.u16 [%rd418], {%rs237, %rs238, %rs239, %rs244};
|
|
bra.uni BB0_180;
|
|
|
|
BB0_179:
|
|
mov.u64 %rd431, image_RNM4;
|
|
cvta.global.u64 %rd426, %rd431;
|
|
mov.u32 %r421, 8;
|
|
mov.u64 %rd430, 0;
|
|
// inline asm
|
|
call (%rd425), _rt_buffer_get_64, (%rd426, %r101, %r421, %rd21, %rd22, %rd430, %rd430);
|
|
// inline asm
|
|
mov.f32 %f1474, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs245, %f1474;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs246, 0;
|
|
st.v4.u16 [%rd425], {%rs245, %rs245, %rs245, %rs246};
|
|
|
|
BB0_180:
|
|
ld.global.u32 %r422, [additive];
|
|
setp.eq.s32 %p155, %r422, 0;
|
|
@%p155 bra BB0_182;
|
|
|
|
mov.u64 %rd444, image_RNM5;
|
|
cvta.global.u64 %rd433, %rd444;
|
|
mov.u32 %r426, 8;
|
|
mov.u64 %rd443, 0;
|
|
// inline asm
|
|
call (%rd432), _rt_buffer_get_64, (%rd433, %r101, %r426, %rd21, %rd22, %rd443, %rd443);
|
|
// inline asm
|
|
ld.v4.u16 {%rs253, %rs254, %rs255, %rs256}, [%rd432];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1475, %rs253;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1476, %rs254;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1477, %rs255;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd438), _rt_buffer_get_64, (%rd433, %r101, %r426, %rd21, %rd22, %rd443, %rd443);
|
|
// inline asm
|
|
add.f32 %f1478, %f1475, 0f00000000;
|
|
add.f32 %f1479, %f1476, 0f00000000;
|
|
add.f32 %f1480, %f1477, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs252, %f1480;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs251, %f1479;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs250, %f1478;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs257, 0;
|
|
st.v4.u16 [%rd438], {%rs250, %rs251, %rs252, %rs257};
|
|
bra.uni BB0_183;
|
|
|
|
BB0_182:
|
|
mov.u64 %rd451, image_RNM5;
|
|
cvta.global.u64 %rd446, %rd451;
|
|
mov.u32 %r428, 8;
|
|
mov.u64 %rd450, 0;
|
|
// inline asm
|
|
call (%rd445), _rt_buffer_get_64, (%rd446, %r101, %r428, %rd21, %rd22, %rd450, %rd450);
|
|
// inline asm
|
|
mov.f32 %f1481, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs258, %f1481;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs259, 0;
|
|
st.v4.u16 [%rd445], {%rs258, %rs258, %rs258, %rs259};
|
|
|
|
BB0_183:
|
|
ld.global.u32 %r429, [additive];
|
|
setp.eq.s32 %p156, %r429, 0;
|
|
@%p156 bra BB0_185;
|
|
|
|
mov.u64 %rd464, image_RNM6;
|
|
cvta.global.u64 %rd453, %rd464;
|
|
mov.u32 %r433, 8;
|
|
mov.u64 %rd463, 0;
|
|
// inline asm
|
|
call (%rd452), _rt_buffer_get_64, (%rd453, %r101, %r433, %rd21, %rd22, %rd463, %rd463);
|
|
// inline asm
|
|
ld.v4.u16 {%rs266, %rs267, %rs268, %rs269}, [%rd452];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1482, %rs266;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1483, %rs267;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1484, %rs268;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd458), _rt_buffer_get_64, (%rd453, %r101, %r433, %rd21, %rd22, %rd463, %rd463);
|
|
// inline asm
|
|
add.f32 %f1485, %f1482, 0f00000000;
|
|
add.f32 %f1486, %f1483, 0f00000000;
|
|
add.f32 %f1487, %f1484, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs265, %f1487;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs264, %f1486;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs263, %f1485;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs270, 0;
|
|
st.v4.u16 [%rd458], {%rs263, %rs264, %rs265, %rs270};
|
|
bra.uni BB0_186;
|
|
|
|
BB0_185:
|
|
mov.u64 %rd471, image_RNM6;
|
|
cvta.global.u64 %rd466, %rd471;
|
|
mov.u32 %r435, 8;
|
|
mov.u64 %rd470, 0;
|
|
// inline asm
|
|
call (%rd465), _rt_buffer_get_64, (%rd466, %r101, %r435, %rd21, %rd22, %rd470, %rd470);
|
|
// inline asm
|
|
mov.f32 %f1488, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs271, %f1488;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs272, 0;
|
|
st.v4.u16 [%rd465], {%rs271, %rs271, %rs271, %rs272};
|
|
|
|
BB0_186:
|
|
ld.global.u32 %r436, [additive];
|
|
setp.eq.s32 %p157, %r436, 0;
|
|
@%p157 bra BB0_188;
|
|
|
|
mov.u64 %rd484, image_RNM7;
|
|
cvta.global.u64 %rd473, %rd484;
|
|
mov.u32 %r440, 8;
|
|
mov.u64 %rd483, 0;
|
|
// inline asm
|
|
call (%rd472), _rt_buffer_get_64, (%rd473, %r101, %r440, %rd21, %rd22, %rd483, %rd483);
|
|
// inline asm
|
|
ld.v4.u16 {%rs279, %rs280, %rs281, %rs282}, [%rd472];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1489, %rs279;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1490, %rs280;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1491, %rs281;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd478), _rt_buffer_get_64, (%rd473, %r101, %r440, %rd21, %rd22, %rd483, %rd483);
|
|
// inline asm
|
|
add.f32 %f1492, %f1489, 0f00000000;
|
|
add.f32 %f1493, %f1490, 0f00000000;
|
|
add.f32 %f1494, %f1491, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs278, %f1494;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs277, %f1493;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs276, %f1492;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs283, 0;
|
|
st.v4.u16 [%rd478], {%rs276, %rs277, %rs278, %rs283};
|
|
bra.uni BB0_189;
|
|
|
|
BB0_188:
|
|
mov.u64 %rd491, image_RNM7;
|
|
cvta.global.u64 %rd486, %rd491;
|
|
mov.u32 %r442, 8;
|
|
mov.u64 %rd490, 0;
|
|
// inline asm
|
|
call (%rd485), _rt_buffer_get_64, (%rd486, %r101, %r442, %rd21, %rd22, %rd490, %rd490);
|
|
// inline asm
|
|
mov.f32 %f1495, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs284, %f1495;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs285, 0;
|
|
st.v4.u16 [%rd485], {%rs284, %rs284, %rs284, %rs285};
|
|
|
|
BB0_189:
|
|
ld.global.u32 %r443, [additive];
|
|
setp.eq.s32 %p158, %r443, 0;
|
|
@%p158 bra BB0_191;
|
|
|
|
mov.u64 %rd504, image_RNM8;
|
|
cvta.global.u64 %rd493, %rd504;
|
|
mov.u32 %r447, 8;
|
|
mov.u64 %rd503, 0;
|
|
// inline asm
|
|
call (%rd492), _rt_buffer_get_64, (%rd493, %r101, %r447, %rd21, %rd22, %rd503, %rd503);
|
|
// inline asm
|
|
ld.v4.u16 {%rs292, %rs293, %rs294, %rs295}, [%rd492];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1496, %rs292;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1497, %rs293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1498, %rs294;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd498), _rt_buffer_get_64, (%rd493, %r101, %r447, %rd21, %rd22, %rd503, %rd503);
|
|
// inline asm
|
|
add.f32 %f1499, %f1496, 0f00000000;
|
|
add.f32 %f1500, %f1497, 0f00000000;
|
|
add.f32 %f1501, %f1498, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs291, %f1501;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs290, %f1500;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs289, %f1499;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs296, 0;
|
|
st.v4.u16 [%rd498], {%rs289, %rs290, %rs291, %rs296};
|
|
bra.uni BB0_192;
|
|
|
|
BB0_191:
|
|
mov.u64 %rd511, image_RNM8;
|
|
cvta.global.u64 %rd506, %rd511;
|
|
mov.u32 %r449, 8;
|
|
mov.u64 %rd510, 0;
|
|
// inline asm
|
|
call (%rd505), _rt_buffer_get_64, (%rd506, %r101, %r449, %rd21, %rd22, %rd510, %rd510);
|
|
// inline asm
|
|
mov.f32 %f1502, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs297, %f1502;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs298, 0;
|
|
st.v4.u16 [%rd505], {%rs297, %rs297, %rs297, %rs298};
|
|
bra.uni BB0_192;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p79, %f433, 0f00000000;
|
|
@%p79 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f1020, %f992;
|
|
setp.neu.f32 %p80, %f1020, 0f3EE8BA2E;
|
|
selp.f32 %f1750, 0f7FFFFFFF, %f1750, %p80;
|
|
|
|
BB0_85:
|
|
add.f32 %f1022, %f435, 0f3EE8BA2E;
|
|
mov.b32 %r266, %f1022;
|
|
setp.lt.s32 %p82, %r266, 2139095040;
|
|
@%p82 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p83, %f435, 0f7F800000;
|
|
@%p83 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f1750, %f433, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p84, %f435, 0f7F800000;
|
|
@%p84 bra BB0_90;
|
|
|
|
selp.f32 %f1750, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f1023, %f1750, 0f437F0000;
|
|
setp.eq.f32 %p85, %f433, 0f3F800000;
|
|
selp.f32 %f1024, 0f437F0000, %f1023, %p85;
|
|
cvt.rzi.u32.f32 %r267, %f1024;
|
|
cvt.u16.u32 %rs20, %r267;
|
|
mov.u16 %rs21, 255;
|
|
st.v2.u8 [%rd84], {%rs20, %rs21};
|
|
ld.global.u32 %r481, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f1025, [lightColor];
|
|
mul.f32 %f447, %f1744, %f1025;
|
|
ld.global.f32 %f1026, [lightColor+4];
|
|
mul.f32 %f448, %f1745, %f1026;
|
|
ld.global.f32 %f1027, [lightColor+8];
|
|
mul.f32 %f449, %f1746, %f1027;
|
|
and.b32 %r268, %r481, 1;
|
|
setp.eq.b32 %p86, %r268, 1;
|
|
@!%p86 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f1030, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f1031, %f1030;
|
|
fma.rn.f32 %f1032, %f1031, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f450, %f1032;
|
|
abs.f32 %f451, %f447;
|
|
setp.lt.f32 %p87, %f451, 0f00800000;
|
|
mul.f32 %f1033, %f451, 0f4B800000;
|
|
selp.f32 %f1034, 0fC3170000, 0fC2FE0000, %p87;
|
|
selp.f32 %f1035, %f1033, %f451, %p87;
|
|
mov.b32 %r269, %f1035;
|
|
and.b32 %r270, %r269, 8388607;
|
|
or.b32 %r271, %r270, 1065353216;
|
|
mov.b32 %f1036, %r271;
|
|
shr.u32 %r272, %r269, 23;
|
|
cvt.rn.f32.u32 %f1037, %r272;
|
|
add.f32 %f1038, %f1034, %f1037;
|
|
setp.gt.f32 %p88, %f1036, 0f3FB504F3;
|
|
mul.f32 %f1039, %f1036, 0f3F000000;
|
|
add.f32 %f1040, %f1038, 0f3F800000;
|
|
selp.f32 %f1041, %f1039, %f1036, %p88;
|
|
selp.f32 %f1042, %f1040, %f1038, %p88;
|
|
add.f32 %f1043, %f1041, 0fBF800000;
|
|
add.f32 %f1029, %f1041, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1028,%f1029;
|
|
// inline asm
|
|
add.f32 %f1044, %f1043, %f1043;
|
|
mul.f32 %f1045, %f1028, %f1044;
|
|
mul.f32 %f1046, %f1045, %f1045;
|
|
mov.f32 %f1047, 0f3C4CAF63;
|
|
mov.f32 %f1048, 0f3B18F0FE;
|
|
fma.rn.f32 %f1049, %f1048, %f1046, %f1047;
|
|
mov.f32 %f1050, 0f3DAAAABD;
|
|
fma.rn.f32 %f1051, %f1049, %f1046, %f1050;
|
|
mul.rn.f32 %f1052, %f1051, %f1046;
|
|
mul.rn.f32 %f1053, %f1052, %f1045;
|
|
sub.f32 %f1054, %f1043, %f1045;
|
|
neg.f32 %f1055, %f1045;
|
|
add.f32 %f1056, %f1054, %f1054;
|
|
fma.rn.f32 %f1057, %f1055, %f1043, %f1056;
|
|
mul.rn.f32 %f1058, %f1028, %f1057;
|
|
add.f32 %f1059, %f1053, %f1045;
|
|
sub.f32 %f1060, %f1045, %f1059;
|
|
add.f32 %f1061, %f1053, %f1060;
|
|
add.f32 %f1062, %f1058, %f1061;
|
|
add.f32 %f1063, %f1059, %f1062;
|
|
sub.f32 %f1064, %f1059, %f1063;
|
|
add.f32 %f1065, %f1062, %f1064;
|
|
mov.f32 %f1066, 0f3F317200;
|
|
mul.rn.f32 %f1067, %f1042, %f1066;
|
|
mov.f32 %f1068, 0f35BFBE8E;
|
|
mul.rn.f32 %f1069, %f1042, %f1068;
|
|
add.f32 %f1070, %f1067, %f1063;
|
|
sub.f32 %f1071, %f1067, %f1070;
|
|
add.f32 %f1072, %f1063, %f1071;
|
|
add.f32 %f1073, %f1065, %f1072;
|
|
add.f32 %f1074, %f1069, %f1073;
|
|
add.f32 %f1075, %f1070, %f1074;
|
|
sub.f32 %f1076, %f1070, %f1075;
|
|
add.f32 %f1077, %f1074, %f1076;
|
|
mov.f32 %f1078, 0f3EE66666;
|
|
mul.rn.f32 %f1079, %f1078, %f1075;
|
|
neg.f32 %f1080, %f1079;
|
|
fma.rn.f32 %f1081, %f1078, %f1075, %f1080;
|
|
fma.rn.f32 %f1082, %f1078, %f1077, %f1081;
|
|
mov.f32 %f1083, 0f00000000;
|
|
fma.rn.f32 %f1084, %f1083, %f1075, %f1082;
|
|
add.rn.f32 %f1085, %f1079, %f1084;
|
|
neg.f32 %f1086, %f1085;
|
|
add.rn.f32 %f1087, %f1079, %f1086;
|
|
add.rn.f32 %f1088, %f1087, %f1084;
|
|
mov.b32 %r273, %f1085;
|
|
setp.eq.s32 %p89, %r273, 1118925336;
|
|
add.s32 %r274, %r273, -1;
|
|
mov.b32 %f1089, %r274;
|
|
add.f32 %f1090, %f1088, 0f37000000;
|
|
selp.f32 %f1091, %f1089, %f1085, %p89;
|
|
selp.f32 %f452, %f1090, %f1088, %p89;
|
|
mul.f32 %f1092, %f1091, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1093, %f1092;
|
|
mov.f32 %f1094, 0fBF317200;
|
|
fma.rn.f32 %f1095, %f1093, %f1094, %f1091;
|
|
mov.f32 %f1096, 0fB5BFBE8E;
|
|
fma.rn.f32 %f1097, %f1093, %f1096, %f1095;
|
|
mul.f32 %f1098, %f1097, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1099, %f1098;
|
|
add.f32 %f1100, %f1093, 0f00000000;
|
|
ex2.approx.f32 %f1101, %f1100;
|
|
mul.f32 %f1102, %f1099, %f1101;
|
|
setp.lt.f32 %p90, %f1091, 0fC2D20000;
|
|
selp.f32 %f1103, 0f00000000, %f1102, %p90;
|
|
setp.gt.f32 %p91, %f1091, 0f42D20000;
|
|
selp.f32 %f1751, 0f7F800000, %f1103, %p91;
|
|
setp.eq.f32 %p92, %f1751, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
fma.rn.f32 %f1751, %f1751, %f452, %f1751;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p93, %f447, 0f00000000;
|
|
setp.eq.f32 %p94, %f450, 0f3F800000;
|
|
and.pred %p4, %p93, %p94;
|
|
mov.b32 %r275, %f1751;
|
|
xor.b32 %r276, %r275, -2147483648;
|
|
mov.b32 %f1104, %r276;
|
|
selp.f32 %f1753, %f1104, %f1751, %p4;
|
|
setp.eq.f32 %p95, %f447, 0f00000000;
|
|
@%p95 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f1107, %f447, %f447;
|
|
selp.f32 %f1753, %f1107, 0f00000000, %p94;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p96, %f447, 0f00000000;
|
|
@%p96 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f1106, %f1078;
|
|
setp.neu.f32 %p97, %f1106, 0f3EE66666;
|
|
selp.f32 %f1753, 0f7FFFFFFF, %f1753, %p97;
|
|
|
|
BB0_98:
|
|
add.f32 %f1108, %f451, 0f3EE66666;
|
|
mov.b32 %r277, %f1108;
|
|
setp.lt.s32 %p99, %r277, 2139095040;
|
|
@%p99 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p100, %f451, 0f7F800000;
|
|
@%p100 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f1753, %f447, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p101, %f451, 0f7F800000;
|
|
@%p101 bra BB0_103;
|
|
|
|
selp.f32 %f1753, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p102, %f447, 0f3F800000;
|
|
selp.f32 %f463, 0f3F800000, %f1753, %p102;
|
|
abs.f32 %f464, %f448;
|
|
setp.lt.f32 %p103, %f464, 0f00800000;
|
|
mul.f32 %f1111, %f464, 0f4B800000;
|
|
selp.f32 %f1112, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f1113, %f1111, %f464, %p103;
|
|
mov.b32 %r278, %f1113;
|
|
and.b32 %r279, %r278, 8388607;
|
|
or.b32 %r280, %r279, 1065353216;
|
|
mov.b32 %f1114, %r280;
|
|
shr.u32 %r281, %r278, 23;
|
|
cvt.rn.f32.u32 %f1115, %r281;
|
|
add.f32 %f1116, %f1112, %f1115;
|
|
setp.gt.f32 %p104, %f1114, 0f3FB504F3;
|
|
mul.f32 %f1117, %f1114, 0f3F000000;
|
|
add.f32 %f1118, %f1116, 0f3F800000;
|
|
selp.f32 %f1119, %f1117, %f1114, %p104;
|
|
selp.f32 %f1120, %f1118, %f1116, %p104;
|
|
add.f32 %f1121, %f1119, 0fBF800000;
|
|
add.f32 %f1110, %f1119, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1109,%f1110;
|
|
// inline asm
|
|
add.f32 %f1122, %f1121, %f1121;
|
|
mul.f32 %f1123, %f1109, %f1122;
|
|
mul.f32 %f1124, %f1123, %f1123;
|
|
fma.rn.f32 %f1127, %f1048, %f1124, %f1047;
|
|
fma.rn.f32 %f1129, %f1127, %f1124, %f1050;
|
|
mul.rn.f32 %f1130, %f1129, %f1124;
|
|
mul.rn.f32 %f1131, %f1130, %f1123;
|
|
sub.f32 %f1132, %f1121, %f1123;
|
|
neg.f32 %f1133, %f1123;
|
|
add.f32 %f1134, %f1132, %f1132;
|
|
fma.rn.f32 %f1135, %f1133, %f1121, %f1134;
|
|
mul.rn.f32 %f1136, %f1109, %f1135;
|
|
add.f32 %f1137, %f1131, %f1123;
|
|
sub.f32 %f1138, %f1123, %f1137;
|
|
add.f32 %f1139, %f1131, %f1138;
|
|
add.f32 %f1140, %f1136, %f1139;
|
|
add.f32 %f1141, %f1137, %f1140;
|
|
sub.f32 %f1142, %f1137, %f1141;
|
|
add.f32 %f1143, %f1140, %f1142;
|
|
mul.rn.f32 %f1145, %f1120, %f1066;
|
|
mul.rn.f32 %f1147, %f1120, %f1068;
|
|
add.f32 %f1148, %f1145, %f1141;
|
|
sub.f32 %f1149, %f1145, %f1148;
|
|
add.f32 %f1150, %f1141, %f1149;
|
|
add.f32 %f1151, %f1143, %f1150;
|
|
add.f32 %f1152, %f1147, %f1151;
|
|
add.f32 %f1153, %f1148, %f1152;
|
|
sub.f32 %f1154, %f1148, %f1153;
|
|
add.f32 %f1155, %f1152, %f1154;
|
|
mul.rn.f32 %f1157, %f1078, %f1153;
|
|
neg.f32 %f1158, %f1157;
|
|
fma.rn.f32 %f1159, %f1078, %f1153, %f1158;
|
|
fma.rn.f32 %f1160, %f1078, %f1155, %f1159;
|
|
fma.rn.f32 %f1162, %f1083, %f1153, %f1160;
|
|
add.rn.f32 %f1163, %f1157, %f1162;
|
|
neg.f32 %f1164, %f1163;
|
|
add.rn.f32 %f1165, %f1157, %f1164;
|
|
add.rn.f32 %f1166, %f1165, %f1162;
|
|
mov.b32 %r282, %f1163;
|
|
setp.eq.s32 %p105, %r282, 1118925336;
|
|
add.s32 %r283, %r282, -1;
|
|
mov.b32 %f1167, %r283;
|
|
add.f32 %f1168, %f1166, 0f37000000;
|
|
selp.f32 %f1169, %f1167, %f1163, %p105;
|
|
selp.f32 %f465, %f1168, %f1166, %p105;
|
|
mul.f32 %f1170, %f1169, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1171, %f1170;
|
|
fma.rn.f32 %f1173, %f1171, %f1094, %f1169;
|
|
fma.rn.f32 %f1175, %f1171, %f1096, %f1173;
|
|
mul.f32 %f1176, %f1175, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1177, %f1176;
|
|
add.f32 %f1178, %f1171, 0f00000000;
|
|
ex2.approx.f32 %f1179, %f1178;
|
|
mul.f32 %f1180, %f1177, %f1179;
|
|
setp.lt.f32 %p106, %f1169, 0fC2D20000;
|
|
selp.f32 %f1181, 0f00000000, %f1180, %p106;
|
|
setp.gt.f32 %p107, %f1169, 0f42D20000;
|
|
selp.f32 %f1754, 0f7F800000, %f1181, %p107;
|
|
setp.eq.f32 %p108, %f1754, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
fma.rn.f32 %f1754, %f1754, %f465, %f1754;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p109, %f448, 0f00000000;
|
|
and.pred %p5, %p109, %p94;
|
|
mov.b32 %r284, %f1754;
|
|
xor.b32 %r285, %r284, -2147483648;
|
|
mov.b32 %f1182, %r285;
|
|
selp.f32 %f1756, %f1182, %f1754, %p5;
|
|
setp.eq.f32 %p111, %f448, 0f00000000;
|
|
@%p111 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f1185, %f448, %f448;
|
|
selp.f32 %f1756, %f1185, 0f00000000, %p94;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p112, %f448, 0f00000000;
|
|
@%p112 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f1184, %f1078;
|
|
setp.neu.f32 %p113, %f1184, 0f3EE66666;
|
|
selp.f32 %f1756, 0f7FFFFFFF, %f1756, %p113;
|
|
|
|
BB0_109:
|
|
add.f32 %f1186, %f464, 0f3EE66666;
|
|
mov.b32 %r286, %f1186;
|
|
setp.lt.s32 %p115, %r286, 2139095040;
|
|
@%p115 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p116, %f464, 0f7F800000;
|
|
@%p116 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f1756, %f448, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p117, %f464, 0f7F800000;
|
|
@%p117 bra BB0_114;
|
|
|
|
selp.f32 %f1756, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p118, %f448, 0f3F800000;
|
|
selp.f32 %f476, 0f3F800000, %f1756, %p118;
|
|
abs.f32 %f477, %f449;
|
|
setp.lt.f32 %p119, %f477, 0f00800000;
|
|
mul.f32 %f1189, %f477, 0f4B800000;
|
|
selp.f32 %f1190, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f1191, %f1189, %f477, %p119;
|
|
mov.b32 %r287, %f1191;
|
|
and.b32 %r288, %r287, 8388607;
|
|
or.b32 %r289, %r288, 1065353216;
|
|
mov.b32 %f1192, %r289;
|
|
shr.u32 %r290, %r287, 23;
|
|
cvt.rn.f32.u32 %f1193, %r290;
|
|
add.f32 %f1194, %f1190, %f1193;
|
|
setp.gt.f32 %p120, %f1192, 0f3FB504F3;
|
|
mul.f32 %f1195, %f1192, 0f3F000000;
|
|
add.f32 %f1196, %f1194, 0f3F800000;
|
|
selp.f32 %f1197, %f1195, %f1192, %p120;
|
|
selp.f32 %f1198, %f1196, %f1194, %p120;
|
|
add.f32 %f1199, %f1197, 0fBF800000;
|
|
add.f32 %f1188, %f1197, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f1187,%f1188;
|
|
// inline asm
|
|
add.f32 %f1200, %f1199, %f1199;
|
|
mul.f32 %f1201, %f1187, %f1200;
|
|
mul.f32 %f1202, %f1201, %f1201;
|
|
fma.rn.f32 %f1205, %f1048, %f1202, %f1047;
|
|
fma.rn.f32 %f1207, %f1205, %f1202, %f1050;
|
|
mul.rn.f32 %f1208, %f1207, %f1202;
|
|
mul.rn.f32 %f1209, %f1208, %f1201;
|
|
sub.f32 %f1210, %f1199, %f1201;
|
|
neg.f32 %f1211, %f1201;
|
|
add.f32 %f1212, %f1210, %f1210;
|
|
fma.rn.f32 %f1213, %f1211, %f1199, %f1212;
|
|
mul.rn.f32 %f1214, %f1187, %f1213;
|
|
add.f32 %f1215, %f1209, %f1201;
|
|
sub.f32 %f1216, %f1201, %f1215;
|
|
add.f32 %f1217, %f1209, %f1216;
|
|
add.f32 %f1218, %f1214, %f1217;
|
|
add.f32 %f1219, %f1215, %f1218;
|
|
sub.f32 %f1220, %f1215, %f1219;
|
|
add.f32 %f1221, %f1218, %f1220;
|
|
mul.rn.f32 %f1223, %f1198, %f1066;
|
|
mul.rn.f32 %f1225, %f1198, %f1068;
|
|
add.f32 %f1226, %f1223, %f1219;
|
|
sub.f32 %f1227, %f1223, %f1226;
|
|
add.f32 %f1228, %f1219, %f1227;
|
|
add.f32 %f1229, %f1221, %f1228;
|
|
add.f32 %f1230, %f1225, %f1229;
|
|
add.f32 %f1231, %f1226, %f1230;
|
|
sub.f32 %f1232, %f1226, %f1231;
|
|
add.f32 %f1233, %f1230, %f1232;
|
|
mul.rn.f32 %f1235, %f1078, %f1231;
|
|
neg.f32 %f1236, %f1235;
|
|
fma.rn.f32 %f1237, %f1078, %f1231, %f1236;
|
|
fma.rn.f32 %f1238, %f1078, %f1233, %f1237;
|
|
fma.rn.f32 %f1240, %f1083, %f1231, %f1238;
|
|
add.rn.f32 %f1241, %f1235, %f1240;
|
|
neg.f32 %f1242, %f1241;
|
|
add.rn.f32 %f1243, %f1235, %f1242;
|
|
add.rn.f32 %f1244, %f1243, %f1240;
|
|
mov.b32 %r291, %f1241;
|
|
setp.eq.s32 %p121, %r291, 1118925336;
|
|
add.s32 %r292, %r291, -1;
|
|
mov.b32 %f1245, %r292;
|
|
add.f32 %f1246, %f1244, 0f37000000;
|
|
selp.f32 %f1247, %f1245, %f1241, %p121;
|
|
selp.f32 %f478, %f1246, %f1244, %p121;
|
|
mul.f32 %f1248, %f1247, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1249, %f1248;
|
|
fma.rn.f32 %f1251, %f1249, %f1094, %f1247;
|
|
fma.rn.f32 %f1253, %f1249, %f1096, %f1251;
|
|
mul.f32 %f1254, %f1253, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1255, %f1254;
|
|
add.f32 %f1256, %f1249, 0f00000000;
|
|
ex2.approx.f32 %f1257, %f1256;
|
|
mul.f32 %f1258, %f1255, %f1257;
|
|
setp.lt.f32 %p122, %f1247, 0fC2D20000;
|
|
selp.f32 %f1259, 0f00000000, %f1258, %p122;
|
|
setp.gt.f32 %p123, %f1247, 0f42D20000;
|
|
selp.f32 %f1757, 0f7F800000, %f1259, %p123;
|
|
setp.eq.f32 %p124, %f1757, 0f7F800000;
|
|
@%p124 bra BB0_116;
|
|
|
|
fma.rn.f32 %f1757, %f1757, %f478, %f1757;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p125, %f449, 0f00000000;
|
|
and.pred %p6, %p125, %p94;
|
|
mov.b32 %r293, %f1757;
|
|
xor.b32 %r294, %r293, -2147483648;
|
|
mov.b32 %f1260, %r294;
|
|
selp.f32 %f1759, %f1260, %f1757, %p6;
|
|
setp.eq.f32 %p127, %f449, 0f00000000;
|
|
@%p127 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f1263, %f449, %f449;
|
|
selp.f32 %f1759, %f1263, 0f00000000, %p94;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p128, %f449, 0f00000000;
|
|
@%p128 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f1262, %f1078;
|
|
setp.neu.f32 %p129, %f1262, 0f3EE66666;
|
|
selp.f32 %f1759, 0f7FFFFFFF, %f1759, %p129;
|
|
|
|
BB0_120:
|
|
add.f32 %f1264, %f477, 0f3EE66666;
|
|
mov.b32 %r295, %f1264;
|
|
setp.lt.s32 %p131, %r295, 2139095040;
|
|
@%p131 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p132, %f477, 0f7F800000;
|
|
@%p132 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f1759, %f449, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p133, %f477, 0f7F800000;
|
|
@%p133 bra BB0_125;
|
|
|
|
selp.f32 %f1759, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
mov.u32 %r452, 4;
|
|
setp.eq.f32 %p134, %f449, 0f3F800000;
|
|
selp.f32 %f1265, 0f3F800000, %f1759, %p134;
|
|
cvt.u64.u32 %rd94, %r7;
|
|
cvt.u64.u32 %rd93, %r6;
|
|
mov.u64 %rd97, image;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r101, %r452, %rd93, %rd94, %rd28, %rd28);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1266, %f1265;
|
|
mul.f32 %f1267, %f1266, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r298, %f1267;
|
|
cvt.sat.f32.f32 %f1268, %f476;
|
|
mul.f32 %f1269, %f1268, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r299, %f1269;
|
|
cvt.sat.f32.f32 %f1270, %f463;
|
|
mul.f32 %f1271, %f1270, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r300, %f1271;
|
|
cvt.u16.u32 %rs22, %r298;
|
|
cvt.u16.u32 %rs23, %r300;
|
|
cvt.u16.u32 %rs24, %r299;
|
|
mov.u16 %rs25, 255;
|
|
st.v4.u8 [%rd91], {%rs22, %rs24, %rs23, %rs25};
|
|
ld.global.u32 %r481, [imageEnabled];
|
|
|
|
BB0_126:
|
|
cvt.u64.u32 %rd19, %r6;
|
|
cvt.u64.u32 %rd20, %r7;
|
|
and.b32 %r301, %r481, 4;
|
|
setp.eq.s32 %p135, %r301, 0;
|
|
@%p135 bra BB0_130;
|
|
|
|
ld.global.u32 %r302, [additive];
|
|
setp.eq.s32 %p136, %r302, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f779;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r306, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r101, %r306, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs33, %rs34, %rs35, %rs36}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1273, %rs33;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1274, %rs34;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1275, %rs35;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r101, %r306, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1276, %f447, %f1273;
|
|
add.f32 %f1277, %f448, %f1274;
|
|
add.f32 %f1278, %f449, %f1275;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f1278;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs31, %f1277;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs30, %f1276;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs30, %rs31, %rs32, %rs26};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r308, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r101, %r308, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f449;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs38, %f448;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs37, %f447;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs37, %rs38, %rs39, %rs26};
|
|
|
|
BB0_130:
|
|
selp.f32 %f1283, 0f40800000, 0f3F800000, %p159;
|
|
mul.f32 %f1284, %f1283, %f1741;
|
|
mul.f32 %f1285, %f1283, %f1742;
|
|
mul.f32 %f1286, %f1283, %f1743;
|
|
mul.f32 %f489, %f1283, %f1738;
|
|
mul.f32 %f490, %f1283, %f1739;
|
|
mul.f32 %f491, %f1283, %f1740;
|
|
mul.f32 %f492, %f1283, %f1735;
|
|
mul.f32 %f493, %f1283, %f1736;
|
|
mul.f32 %f494, %f1283, %f1737;
|
|
mul.f32 %f495, %f1283, %f1732;
|
|
mul.f32 %f496, %f1283, %f1733;
|
|
mul.f32 %f497, %f1283, %f1734;
|
|
mul.f32 %f498, %f1283, %f1729;
|
|
mul.f32 %f499, %f1283, %f1730;
|
|
mul.f32 %f500, %f1283, %f1731;
|
|
mul.f32 %f501, %f1283, %f1726;
|
|
mul.f32 %f502, %f1283, %f1727;
|
|
mul.f32 %f503, %f1283, %f1728;
|
|
mul.f32 %f504, %f1283, %f1723;
|
|
mul.f32 %f505, %f1283, %f1724;
|
|
mul.f32 %f506, %f1283, %f1725;
|
|
mul.f32 %f507, %f1283, %f1720;
|
|
mul.f32 %f508, %f1283, %f1721;
|
|
mul.f32 %f509, %f1283, %f1722;
|
|
mul.f32 %f510, %f1283, %f1717;
|
|
mul.f32 %f511, %f1283, %f1718;
|
|
mul.f32 %f512, %f1283, %f1719;
|
|
ld.global.f32 %f1287, [lightColor];
|
|
mul.f32 %f1288, %f1284, %f1287;
|
|
ld.global.f32 %f1289, [lightColor+4];
|
|
mul.f32 %f1290, %f1285, %f1289;
|
|
ld.global.f32 %f1291, [lightColor+8];
|
|
mul.f32 %f1292, %f1286, %f1291;
|
|
mul.f32 %f513, %f1288, 0f3F558750;
|
|
mul.f32 %f514, %f1290, 0f3F558750;
|
|
mul.f32 %f515, %f1292, 0f3F558750;
|
|
ld.global.u32 %r309, [additive];
|
|
setp.eq.s32 %p137, %r309, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f779;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_132;
|
|
|
|
mov.u64 %rd130, image_RNM0;
|
|
cvta.global.u64 %rd119, %rd130;
|
|
mov.u32 %r313, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r101, %r313, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs47, %rs48, %rs49, %rs50}, [%rd118];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1293, %rs47;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1294, %rs48;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1295, %rs49;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd119, %r101, %r313, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1296, %f513, %f1293;
|
|
add.f32 %f1297, %f514, %f1294;
|
|
add.f32 %f1298, %f515, %f1295;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f1298;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs45, %f1297;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs44, %f1296;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs44, %rs45, %rs46, %rs40};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd137, image_RNM0;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u32 %r315, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r101, %r315, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f515;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs52, %f514;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs51, %f513;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs51, %rs52, %rs53, %rs40};
|
|
|
|
BB0_133:
|
|
ld.global.f32 %f1303, [lightColor];
|
|
mul.f32 %f1304, %f489, %f1303;
|
|
ld.global.f32 %f1305, [lightColor+4];
|
|
mul.f32 %f1306, %f490, %f1305;
|
|
ld.global.f32 %f1307, [lightColor+8];
|
|
mul.f32 %f1308, %f491, %f1307;
|
|
mul.f32 %f516, %f1304, 0f3FB8EBD1;
|
|
mul.f32 %f517, %f1306, 0f3FB8EBD1;
|
|
mul.f32 %f518, %f1308, 0f3FB8EBD1;
|
|
ld.global.u32 %r316, [additive];
|
|
setp.eq.s32 %p138, %r316, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f779;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_135;
|
|
|
|
mov.u64 %rd150, image_RNM1;
|
|
cvta.global.u64 %rd139, %rd150;
|
|
mov.u32 %r320, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r101, %r320, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs61, %rs62, %rs63, %rs64}, [%rd138];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1309, %rs61;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1310, %rs62;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1311, %rs63;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd139, %r101, %r320, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1312, %f516, %f1309;
|
|
add.f32 %f1313, %f517, %f1310;
|
|
add.f32 %f1314, %f518, %f1311;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f1314;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs59, %f1313;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs58, %f1312;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs58, %rs59, %rs60, %rs54};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd157, image_RNM1;
|
|
cvta.global.u64 %rd152, %rd157;
|
|
mov.u32 %r322, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r101, %r322, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f518;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs66, %f517;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs65, %f516;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd151], {%rs65, %rs66, %rs67, %rs54};
|
|
|
|
BB0_136:
|
|
ld.global.f32 %f1319, [lightColor];
|
|
mul.f32 %f1320, %f492, %f1319;
|
|
ld.global.f32 %f1321, [lightColor+4];
|
|
mul.f32 %f1322, %f493, %f1321;
|
|
ld.global.f32 %f1323, [lightColor+8];
|
|
mul.f32 %f1324, %f494, %f1323;
|
|
mul.f32 %f519, %f1320, 0f3FB8EBD1;
|
|
mul.f32 %f520, %f1322, 0f3FB8EBD1;
|
|
mul.f32 %f521, %f1324, 0f3FB8EBD1;
|
|
ld.global.u32 %r323, [additive];
|
|
setp.eq.s32 %p139, %r323, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f779;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_138;
|
|
|
|
mov.u64 %rd170, image_RNM2;
|
|
cvta.global.u64 %rd159, %rd170;
|
|
mov.u32 %r327, 8;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r101, %r327, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs75, %rs76, %rs77, %rs78}, [%rd158];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1325, %rs75;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1326, %rs76;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1327, %rs77;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd159, %r101, %r327, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1328, %f519, %f1325;
|
|
add.f32 %f1329, %f520, %f1326;
|
|
add.f32 %f1330, %f521, %f1327;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f1330;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs73, %f1329;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs72, %f1328;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs72, %rs73, %rs74, %rs68};
|
|
bra.uni BB0_139;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd177, image_RNM2;
|
|
cvta.global.u64 %rd172, %rd177;
|
|
mov.u32 %r329, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r101, %r329, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f521;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs80, %f520;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs79, %f519;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd171], {%rs79, %rs80, %rs81, %rs68};
|
|
|
|
BB0_139:
|
|
ld.global.f32 %f1335, [lightColor];
|
|
mul.f32 %f1336, %f495, %f1335;
|
|
ld.global.f32 %f1337, [lightColor+4];
|
|
mul.f32 %f1338, %f496, %f1337;
|
|
ld.global.f32 %f1339, [lightColor+8];
|
|
mul.f32 %f1340, %f497, %f1339;
|
|
mul.f32 %f522, %f1336, 0f3FB8EBD1;
|
|
mul.f32 %f523, %f1338, 0f3FB8EBD1;
|
|
mul.f32 %f524, %f1340, 0f3FB8EBD1;
|
|
ld.global.u32 %r330, [additive];
|
|
setp.eq.s32 %p140, %r330, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f779;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_141;
|
|
|
|
mov.u64 %rd190, image_RNM3;
|
|
cvta.global.u64 %rd179, %rd190;
|
|
mov.u32 %r334, 8;
|
|
// inline asm
|
|
call (%rd178), _rt_buffer_get_64, (%rd179, %r101, %r334, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs89, %rs90, %rs91, %rs92}, [%rd178];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1341, %rs89;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1342, %rs90;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1343, %rs91;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd179, %r101, %r334, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1344, %f522, %f1341;
|
|
add.f32 %f1345, %f523, %f1342;
|
|
add.f32 %f1346, %f524, %f1343;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f1346;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs87, %f1345;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs86, %f1344;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs86, %rs87, %rs88, %rs82};
|
|
bra.uni BB0_142;
|
|
|
|
BB0_141:
|
|
mov.u64 %rd197, image_RNM3;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u32 %r336, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r101, %r336, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs95, %f524;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs94, %f523;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs93, %f522;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd191], {%rs93, %rs94, %rs95, %rs82};
|
|
|
|
BB0_142:
|
|
ld.global.f32 %f1351, [lightColor];
|
|
mul.f32 %f1352, %f498, %f1351;
|
|
ld.global.f32 %f1353, [lightColor+4];
|
|
mul.f32 %f1354, %f499, %f1353;
|
|
ld.global.f32 %f1355, [lightColor+8];
|
|
mul.f32 %f1356, %f500, %f1355;
|
|
mul.f32 %f525, %f1352, 0f404EBF87;
|
|
mul.f32 %f526, %f1354, 0f404EBF87;
|
|
mul.f32 %f527, %f1356, 0f404EBF87;
|
|
ld.global.u32 %r337, [additive];
|
|
setp.eq.s32 %p141, %r337, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs96, %f779;}
|
|
|
|
// inline asm
|
|
@%p141 bra BB0_144;
|
|
|
|
mov.u64 %rd210, image_RNM4;
|
|
cvta.global.u64 %rd199, %rd210;
|
|
mov.u32 %r341, 8;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r101, %r341, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs103, %rs104, %rs105, %rs106}, [%rd198];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1357, %rs103;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1358, %rs104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1359, %rs105;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd199, %r101, %r341, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1360, %f525, %f1357;
|
|
add.f32 %f1361, %f526, %f1358;
|
|
add.f32 %f1362, %f527, %f1359;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs102, %f1362;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs101, %f1361;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs100, %f1360;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd204], {%rs100, %rs101, %rs102, %rs96};
|
|
bra.uni BB0_145;
|
|
|
|
BB0_144:
|
|
mov.u64 %rd217, image_RNM4;
|
|
cvta.global.u64 %rd212, %rd217;
|
|
mov.u32 %r343, 8;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r101, %r343, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs109, %f527;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs108, %f526;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs107, %f525;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd211], {%rs107, %rs108, %rs109, %rs96};
|
|
|
|
BB0_145:
|
|
ld.global.f32 %f1367, [lightColor];
|
|
mul.f32 %f1368, %f501, %f1367;
|
|
ld.global.f32 %f1369, [lightColor+4];
|
|
mul.f32 %f1370, %f502, %f1369;
|
|
ld.global.f32 %f1371, [lightColor+8];
|
|
mul.f32 %f1372, %f503, %f1371;
|
|
mul.f32 %f528, %f1368, 0f404EBF87;
|
|
mul.f32 %f529, %f1370, 0f404EBF87;
|
|
mul.f32 %f530, %f1372, 0f404EBF87;
|
|
ld.global.u32 %r344, [additive];
|
|
setp.eq.s32 %p142, %r344, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f779;}
|
|
|
|
// inline asm
|
|
@%p142 bra BB0_147;
|
|
|
|
mov.u64 %rd230, image_RNM5;
|
|
cvta.global.u64 %rd219, %rd230;
|
|
mov.u32 %r348, 8;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r101, %r348, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs117, %rs118, %rs119, %rs120}, [%rd218];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1373, %rs117;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1374, %rs118;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1375, %rs119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd219, %r101, %r348, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1376, %f528, %f1373;
|
|
add.f32 %f1377, %f529, %f1374;
|
|
add.f32 %f1378, %f530, %f1375;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs116, %f1378;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs115, %f1377;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs114, %f1376;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd224], {%rs114, %rs115, %rs116, %rs110};
|
|
bra.uni BB0_148;
|
|
|
|
BB0_147:
|
|
mov.u64 %rd237, image_RNM5;
|
|
cvta.global.u64 %rd232, %rd237;
|
|
mov.u32 %r350, 8;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r101, %r350, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f530;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs122, %f529;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs121, %f528;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd231], {%rs121, %rs122, %rs123, %rs110};
|
|
|
|
BB0_148:
|
|
ld.global.f32 %f1383, [lightColor];
|
|
mul.f32 %f1384, %f504, %f1383;
|
|
ld.global.f32 %f1385, [lightColor+4];
|
|
mul.f32 %f1386, %f505, %f1385;
|
|
ld.global.f32 %f1387, [lightColor+8];
|
|
mul.f32 %f1388, %f506, %f1387;
|
|
mul.f32 %f531, %f1384, 0f40330C94;
|
|
mul.f32 %f532, %f1386, 0f40330C94;
|
|
mul.f32 %f533, %f1388, 0f40330C94;
|
|
ld.global.u32 %r351, [additive];
|
|
setp.eq.s32 %p143, %r351, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f779;}
|
|
|
|
// inline asm
|
|
@%p143 bra BB0_150;
|
|
|
|
mov.u64 %rd250, image_RNM6;
|
|
cvta.global.u64 %rd239, %rd250;
|
|
mov.u32 %r355, 8;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r101, %r355, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs131, %rs132, %rs133, %rs134}, [%rd238];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1389, %rs131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1390, %rs132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1391, %rs133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd239, %r101, %r355, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1392, %f531, %f1389;
|
|
add.f32 %f1393, %f532, %f1390;
|
|
add.f32 %f1394, %f533, %f1391;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs130, %f1394;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs129, %f1393;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs128, %f1392;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd244], {%rs128, %rs129, %rs130, %rs124};
|
|
bra.uni BB0_151;
|
|
|
|
BB0_150:
|
|
mov.u64 %rd257, image_RNM6;
|
|
cvta.global.u64 %rd252, %rd257;
|
|
mov.u32 %r357, 8;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r101, %r357, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f533;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f532;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs135, %f531;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd251], {%rs135, %rs136, %rs137, %rs124};
|
|
|
|
BB0_151:
|
|
ld.global.f32 %f1399, [lightColor];
|
|
mul.f32 %f1400, %f507, %f1399;
|
|
ld.global.f32 %f1401, [lightColor+4];
|
|
mul.f32 %f1402, %f508, %f1401;
|
|
ld.global.f32 %f1403, [lightColor+8];
|
|
mul.f32 %f1404, %f509, %f1403;
|
|
mul.f32 %f534, %f1400, 0f404EBF87;
|
|
mul.f32 %f535, %f1402, 0f404EBF87;
|
|
mul.f32 %f536, %f1404, 0f404EBF87;
|
|
ld.global.u32 %r358, [additive];
|
|
setp.eq.s32 %p144, %r358, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f779;}
|
|
|
|
// inline asm
|
|
@%p144 bra BB0_153;
|
|
|
|
mov.u64 %rd270, image_RNM7;
|
|
cvta.global.u64 %rd259, %rd270;
|
|
mov.u32 %r362, 8;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r101, %r362, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs145, %rs146, %rs147, %rs148}, [%rd258];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1405, %rs145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1406, %rs146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1407, %rs147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd259, %r101, %r362, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1408, %f534, %f1405;
|
|
add.f32 %f1409, %f535, %f1406;
|
|
add.f32 %f1410, %f536, %f1407;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1410;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs143, %f1409;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs142, %f1408;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd264], {%rs142, %rs143, %rs144, %rs138};
|
|
bra.uni BB0_154;
|
|
|
|
BB0_153:
|
|
mov.u64 %rd277, image_RNM7;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r364, 8;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r101, %r364, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f536;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f535;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f534;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd271], {%rs149, %rs150, %rs151, %rs138};
|
|
|
|
BB0_154:
|
|
ld.global.f32 %f1415, [lightColor];
|
|
mul.f32 %f1416, %f510, %f1415;
|
|
ld.global.f32 %f1417, [lightColor+4];
|
|
mul.f32 %f1418, %f511, %f1417;
|
|
ld.global.f32 %f1419, [lightColor+8];
|
|
mul.f32 %f1420, %f512, %f1419;
|
|
mul.f32 %f537, %f1416, 0f3FCEBF87;
|
|
mul.f32 %f538, %f1418, 0f3FCEBF87;
|
|
mul.f32 %f539, %f1420, 0f3FCEBF87;
|
|
ld.global.u32 %r365, [additive];
|
|
setp.eq.s32 %p145, %r365, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs152, %f779;}
|
|
|
|
// inline asm
|
|
@%p145 bra BB0_156;
|
|
|
|
mov.u64 %rd290, image_RNM8;
|
|
cvta.global.u64 %rd279, %rd290;
|
|
mov.u32 %r369, 8;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r101, %r369, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs159, %rs160, %rs161, %rs162}, [%rd278];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1421, %rs159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1422, %rs160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1423, %rs161;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd284), _rt_buffer_get_64, (%rd279, %r101, %r369, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1424, %f537, %f1421;
|
|
add.f32 %f1425, %f538, %f1422;
|
|
add.f32 %f1426, %f539, %f1423;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs158, %f1426;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1425;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs156, %f1424;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd284], {%rs156, %rs157, %rs158, %rs152};
|
|
bra.uni BB0_192;
|
|
|
|
BB0_156:
|
|
mov.u64 %rd297, image_RNM8;
|
|
cvta.global.u64 %rd292, %rd297;
|
|
mov.u32 %r371, 8;
|
|
// inline asm
|
|
call (%rd291), _rt_buffer_get_64, (%rd292, %r101, %r371, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs165, %f539;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs164, %f538;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs163, %f537;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd291], {%rs163, %rs164, %rs165, %rs152};
|
|
|
|
BB0_192:
|
|
ret;
|
|
}
|
|
|
|
|