2553 lines
71 KiB
Plaintext
2553 lines
71 KiB
Plaintext
//
|
|
// Generated by NVIDIA NVVM Compiler
|
|
//
|
|
// Compiler Build ID: CL-23083092
|
|
// Cuda compilation tools, release 9.1, V9.1.85
|
|
// Based on LLVM 3.4svn
|
|
//
|
|
|
|
.version 6.1
|
|
.target sm_30
|
|
.address_size 64
|
|
|
|
// .globl _Z6oxMainv
|
|
.global .align 8 .b8 pixelID[8];
|
|
.global .align 8 .b8 resolution[8];
|
|
.global .align 4 .b8 normal[12];
|
|
.global .align 4 .b8 camPos[12];
|
|
.global .align 4 .b8 root[4];
|
|
.global .align 4 .u32 imageEnabled;
|
|
.global .texref lightmap;
|
|
.global .align 16 .b8 tileInfo[16];
|
|
.global .align 4 .u32 additive;
|
|
.global .align 1 .b8 image[1];
|
|
.global .align 1 .b8 image_HDR[1];
|
|
.global .align 1 .b8 image_HDR2[1];
|
|
.global .align 1 .b8 image_Mask[1];
|
|
.global .align 1 .b8 image_RNM0[1];
|
|
.global .align 1 .b8 image_RNM1[1];
|
|
.global .align 1 .b8 image_RNM2[1];
|
|
.global .align 1 .b8 image_RNM3[1];
|
|
.global .align 1 .b8 uvpos[1];
|
|
.global .align 1 .b8 uvnormal[1];
|
|
.global .align 1 .b8 lightMeshBuffer[1];
|
|
.global .align 4 .u32 lightMeshBufferSize;
|
|
.global .align 4 .f32 lightInvCutoff;
|
|
.global .align 4 .f32 lightPointSize;
|
|
.global .align 4 .b8 lightColor[12];
|
|
.global .align 1 .b8 rnd_seeds[1];
|
|
.global .align 1 .b8 rnd_seeds2[1];
|
|
.global .align 4 .u32 samples;
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7pixelIDE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10resolutionE[8] = {82, 97, 121, 0, 8, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6normalE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo6camPosE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo4rootE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo12imageEnabledE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8tileInfoE[8] = {82, 97, 121, 0, 16, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo8additiveE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo19lightMeshBufferSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightInvCutoffE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo14lightPointSizeE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo10lightColorE[8] = {82, 97, 121, 0, 12, 0, 0, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typeinfo7samplesE[8] = {82, 97, 121, 0, 4, 0, 0, 0};
|
|
.global .align 8 .u64 _ZN21rti_internal_register20reg_bitness_detectorE;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail0E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail1E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail2E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail3E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail4E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail5E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail6E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail7E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail8E;
|
|
.global .align 8 .u64 _ZN21rti_internal_register24reg_exception_64_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail0E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail1E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail2E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail3E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail4E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail5E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail6E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail7E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail8E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register21reg_exception_detail9E;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_xE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_yE;
|
|
.global .align 4 .u32 _ZN21rti_internal_register14reg_rayIndex_zE;
|
|
.global .align 8 .b8 _ZN21rti_internal_typename7pixelIDE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10resolutionE[6] = {117, 105, 110, 116, 50, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6normalE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename6camPosE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename4rootE[9] = {114, 116, 79, 98, 106, 101, 99, 116, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename12imageEnabledE[4] = {105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename8tileInfoE[6] = {117, 105, 110, 116, 52, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename8additiveE[4] = {105, 110, 116, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_typename19lightMeshBufferSizeE[13] = {117, 110, 115, 105, 103, 110, 101, 100, 32, 105, 110, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightInvCutoffE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename14lightPointSizeE[6] = {102, 108, 111, 97, 116, 0};
|
|
.global .align 8 .b8 _ZN21rti_internal_typename10lightColorE[7] = {102, 108, 111, 97, 116, 51, 0};
|
|
.global .align 4 .b8 _ZN21rti_internal_typename7samplesE[4] = {105, 110, 116, 0};
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7pixelIDE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10resolutionE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6normalE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum6camPosE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum4rootE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum12imageEnabledE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8tileInfoE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum8additiveE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum19lightMeshBufferSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightInvCutoffE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum14lightPointSizeE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum10lightColorE = 4919;
|
|
.global .align 4 .u32 _ZN21rti_internal_typeenum7samplesE = 4919;
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic7pixelIDE[14] = {114, 116, 76, 97, 117, 110, 99, 104, 73, 110, 100, 101, 120, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic10resolutionE[12] = {114, 116, 76, 97, 117, 110, 99, 104, 68, 105, 109, 0};
|
|
.global .align 16 .b8 _ZN21rti_internal_semantic6normalE[17] = {97, 116, 116, 114, 105, 98, 117, 116, 101, 32, 110, 111, 114, 109, 97, 108, 0};
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic6camPosE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic4rootE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic8additiveE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic10lightColorE[1];
|
|
.global .align 1 .b8 _ZN21rti_internal_semantic7samplesE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7pixelIDE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10resolutionE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6normalE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation6camPosE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation4rootE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation12imageEnabledE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8tileInfoE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation8additiveE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation19lightMeshBufferSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightInvCutoffE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation14lightPointSizeE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation10lightColorE[1];
|
|
.global .align 1 .b8 _ZN23rti_internal_annotation7samplesE[1];
|
|
.const .align 4 .b8 __cudart_i2opi_f[24] = {65, 144, 67, 60, 153, 149, 98, 219, 192, 221, 52, 245, 209, 87, 39, 252, 41, 21, 68, 78, 110, 131, 249, 162};
|
|
|
|
.visible .entry _Z6oxMainv(
|
|
|
|
)
|
|
{
|
|
.local .align 4 .b8 __local_depot0[44];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<150>;
|
|
.reg .b16 %rs<159>;
|
|
.reg .f32 %f<1321>;
|
|
.reg .b32 %r<422>;
|
|
.reg .b64 %rd<330>;
|
|
|
|
|
|
mov.u64 %rd329, __local_depot0;
|
|
cvta.local.u64 %SP, %rd329;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd24, %r105;
|
|
cvt.u64.u32 %rd25, %r106;
|
|
mov.u64 %rd28, uvnormal;
|
|
cvta.global.u64 %rd23, %rd28;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd27, 0;
|
|
// inline asm
|
|
call (%rd22), _rt_buffer_get_64, (%rd23, %r103, %r104, %rd24, %rd25, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd22];
|
|
shr.u32 %r109, %r4, 16;
|
|
cvt.u16.u32 %rs1, %r109;
|
|
and.b16 %rs7, %rs1, 255;
|
|
cvt.u16.u32 %rs8, %r4;
|
|
or.b16 %rs9, %rs8, %rs7;
|
|
setp.eq.s16 %p7, %rs9, 0;
|
|
mov.f32 %f1166, 0f00000000;
|
|
mov.f32 %f1167, %f1166;
|
|
mov.f32 %f1168, %f1166;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd22+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f358, %rs12;
|
|
div.rn.f32 %f359, %f358, 0f437F0000;
|
|
fma.rn.f32 %f360, %f359, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f361, %rs10;
|
|
div.rn.f32 %f362, %f361, 0f437F0000;
|
|
fma.rn.f32 %f363, %f362, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f364, %rs7;
|
|
div.rn.f32 %f365, %f364, 0f437F0000;
|
|
fma.rn.f32 %f366, %f365, 0f40000000, 0fBF800000;
|
|
mul.f32 %f367, %f363, %f363;
|
|
fma.rn.f32 %f368, %f360, %f360, %f367;
|
|
fma.rn.f32 %f369, %f366, %f366, %f368;
|
|
sqrt.rn.f32 %f370, %f369;
|
|
rcp.rn.f32 %f371, %f370;
|
|
mul.f32 %f1166, %f360, %f371;
|
|
mul.f32 %f1167, %f363, %f371;
|
|
mul.f32 %f1168, %f366, %f371;
|
|
|
|
BB0_2:
|
|
ld.global.v2.u32 {%r110, %r111}, [pixelID];
|
|
ld.global.v2.u32 {%r113, %r114}, [tileInfo];
|
|
add.s32 %r5, %r110, %r113;
|
|
add.s32 %r6, %r111, %r114;
|
|
setp.eq.f32 %p8, %f1167, 0f00000000;
|
|
setp.eq.f32 %p9, %f1166, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1168, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_142;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_142:
|
|
ld.global.u32 %r421, [imageEnabled];
|
|
and.b32 %r343, %r421, 1;
|
|
setp.eq.b32 %p141, %r343, 1;
|
|
@!%p141 bra BB0_144;
|
|
bra.uni BB0_143;
|
|
|
|
BB0_143:
|
|
cvt.u64.u32 %rd206, %r5;
|
|
cvt.u64.u32 %rd207, %r6;
|
|
mov.u64 %rd210, image;
|
|
cvta.global.u64 %rd205, %rd210;
|
|
mov.u64 %rd209, 0;
|
|
// inline asm
|
|
call (%rd204), _rt_buffer_get_64, (%rd205, %r103, %r104, %rd206, %rd207, %rd209, %rd209);
|
|
// inline asm
|
|
mov.u16 %rs91, 0;
|
|
st.v4.u8 [%rd204], {%rs91, %rs91, %rs91, %rs91};
|
|
ld.global.u32 %r421, [imageEnabled];
|
|
|
|
BB0_144:
|
|
and.b32 %r346, %r421, 8;
|
|
setp.eq.s32 %p142, %r346, 0;
|
|
@%p142 bra BB0_146;
|
|
|
|
cvt.u64.u32 %rd214, %r6;
|
|
cvt.u64.u32 %rd213, %r5;
|
|
mov.u64 %rd217, image_Mask;
|
|
cvta.global.u64 %rd212, %rd217;
|
|
mov.u64 %rd216, 0;
|
|
// inline asm
|
|
call (%rd211), _rt_buffer_get_64, (%rd212, %r103, %r103, %rd213, %rd214, %rd216, %rd216);
|
|
// inline asm
|
|
mov.f32 %f1121, 0f00000000;
|
|
cvt.rzi.u32.f32 %r349, %f1121;
|
|
cvt.u16.u32 %rs92, %r349;
|
|
mov.u16 %rs93, 0;
|
|
st.v2.u8 [%rd211], {%rs92, %rs93};
|
|
ld.global.u32 %r421, [imageEnabled];
|
|
|
|
BB0_146:
|
|
cvt.u64.u32 %rd20, %r5;
|
|
cvt.u64.u32 %rd21, %r6;
|
|
and.b32 %r350, %r421, 4;
|
|
setp.eq.s32 %p143, %r350, 0;
|
|
@%p143 bra BB0_150;
|
|
|
|
ld.global.u32 %r351, [additive];
|
|
setp.eq.s32 %p144, %r351, 0;
|
|
@%p144 bra BB0_149;
|
|
|
|
mov.u64 %rd230, image_HDR;
|
|
cvta.global.u64 %rd219, %rd230;
|
|
mov.u32 %r355, 8;
|
|
mov.u64 %rd229, 0;
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd219, %r103, %r355, %rd20, %rd21, %rd229, %rd229);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd218];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1122, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1123, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1124, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd224), _rt_buffer_get_64, (%rd219, %r103, %r355, %rd20, %rd21, %rd229, %rd229);
|
|
// inline asm
|
|
add.f32 %f1125, %f1122, 0f00000000;
|
|
add.f32 %f1126, %f1123, 0f00000000;
|
|
add.f32 %f1127, %f1124, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f1125;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs104, 0;
|
|
st.v4.u16 [%rd224], {%rs97, %rs98, %rs99, %rs104};
|
|
bra.uni BB0_150;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd31, %r123;
|
|
cvt.u64.u32 %rd32, %r124;
|
|
mov.u64 %rd41, uvpos;
|
|
cvta.global.u64 %rd30, %rd41;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd29), _rt_buffer_get_64, (%rd30, %r103, %r119, %rd31, %rd32, %rd27, %rd27);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd29+8];
|
|
ld.f32 %f8, [%rd29+4];
|
|
ld.f32 %f7, [%rd29];
|
|
mul.f32 %f388, %f7, 0f3456BF95;
|
|
mul.f32 %f389, %f8, 0f3456BF95;
|
|
mul.f32 %f390, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f1166;
|
|
div.rn.f32 %f391, %f388, %f10;
|
|
abs.f32 %f392, %f1167;
|
|
div.rn.f32 %f393, %f389, %f392;
|
|
abs.f32 %f11, %f1168;
|
|
div.rn.f32 %f394, %f390, %f11;
|
|
abs.f32 %f395, %f391;
|
|
abs.f32 %f396, %f393;
|
|
abs.f32 %f397, %f394;
|
|
mov.f32 %f398, 0f38D1B717;
|
|
max.f32 %f399, %f395, %f398;
|
|
max.f32 %f400, %f396, %f398;
|
|
max.f32 %f401, %f397, %f398;
|
|
fma.rn.f32 %f12, %f1166, %f399, %f7;
|
|
fma.rn.f32 %f13, %f1167, %f400, %f8;
|
|
fma.rn.f32 %f14, %f1168, %f401, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd37, %r127;
|
|
cvt.u64.u32 %rd38, %r128;
|
|
mov.u64 %rd42, rnd_seeds;
|
|
cvta.global.u64 %rd36, %rd42;
|
|
// inline asm
|
|
call (%rd35), _rt_buffer_get_64, (%rd36, %r103, %r104, %rd37, %rd38, %rd27, %rd27);
|
|
// inline asm
|
|
ld.u32 %r397, [%rd35];
|
|
mov.pred %p13, 0;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r9, 0;
|
|
setp.lt.s32 %p14, %r2, 1;
|
|
@%p14 bra BB0_4;
|
|
|
|
ld.global.f32 %f15, [lightPointSize];
|
|
mul.f32 %f16, %f12, 0f3456BF95;
|
|
mul.f32 %f17, %f13, 0f3456BF95;
|
|
mul.f32 %f18, %f14, 0f3456BF95;
|
|
mov.f32 %f19, 0f00000000;
|
|
mov.u32 %r390, 0;
|
|
abs.f32 %f541, %f17;
|
|
abs.f32 %f542, %f16;
|
|
max.f32 %f543, %f542, %f541;
|
|
abs.f32 %f544, %f18;
|
|
max.f32 %f545, %f543, %f544;
|
|
mov.u32 %r9, %r390;
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f23, %f19;
|
|
mov.f32 %f24, %f19;
|
|
mov.f32 %f25, %f19;
|
|
mov.f32 %f26, %f19;
|
|
mov.f32 %f27, %f19;
|
|
mov.f32 %f28, %f19;
|
|
mov.f32 %f29, %f19;
|
|
mov.f32 %f30, %f19;
|
|
mov.f32 %f31, %f19;
|
|
mov.f32 %f32, %f19;
|
|
mov.f32 %f33, %f19;
|
|
mov.f32 %f34, %f19;
|
|
|
|
BB0_6:
|
|
mul.lo.s32 %r10, %r390, 3;
|
|
cvt.s64.s32 %rd45, %r10;
|
|
mov.u64 %rd49, lightMeshBuffer;
|
|
cvta.global.u64 %rd44, %rd49;
|
|
mov.u32 %r133, 1;
|
|
mov.u64 %rd48, 0;
|
|
// inline asm
|
|
call (%rd43), _rt_buffer_get_64, (%rd44, %r133, %r119, %rd45, %rd48, %rd48, %rd48);
|
|
// inline asm
|
|
ld.f32 %f418, [%rd43];
|
|
sub.f32 %f419, %f418, %f7;
|
|
ld.f32 %f420, [%rd43+4];
|
|
sub.f32 %f421, %f420, %f8;
|
|
ld.f32 %f422, [%rd43+8];
|
|
sub.f32 %f423, %f422, %f9;
|
|
mul.f32 %f424, %f421, %f421;
|
|
fma.rn.f32 %f425, %f419, %f419, %f424;
|
|
fma.rn.f32 %f426, %f423, %f423, %f425;
|
|
sqrt.rn.f32 %f35, %f426;
|
|
rcp.rn.f32 %f427, %f35;
|
|
mul.f32 %f36, %f419, %f427;
|
|
mul.f32 %f37, %f421, %f427;
|
|
mul.f32 %f38, %f423, %f427;
|
|
mul.f32 %f428, %f1167, %f37;
|
|
fma.rn.f32 %f429, %f1166, %f36, %f428;
|
|
fma.rn.f32 %f39, %f1168, %f38, %f429;
|
|
setp.leu.f32 %p15, %f39, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
mov.u64 %rd321, 0;
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f430, %f35, %f35;
|
|
div.rn.f32 %f40, %f15, %f430;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd56, %r138;
|
|
cvt.u64.u32 %rd57, %r137;
|
|
and.b64 %rd52, %rd57, 255;
|
|
and.b64 %rd53, %rd56, 255;
|
|
mov.u64 %rd58, rnd_seeds2;
|
|
cvta.global.u64 %rd51, %rd58;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd50), _rt_buffer_get_64, (%rd51, %r103, %r136, %rd52, %rd53, %rd321, %rd321);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs14, [%rd50];
|
|
cvt.rn.f32.u16 %f431, %rs14;
|
|
div.rn.f32 %f432, %f431, 0fC37F0000;
|
|
fma.rn.f32 %f433, %f432, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f40, %f433;
|
|
mov.pred %p149, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f436, [lightInvCutoff];
|
|
mul.f32 %f41, %f35, %f436;
|
|
mov.f32 %f440, 0f40800000;
|
|
abs.f32 %f43, %f41;
|
|
setp.lt.f32 %p19, %f43, 0f00800000;
|
|
mul.f32 %f442, %f43, 0f4B800000;
|
|
selp.f32 %f443, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f444, %f442, %f43, %p19;
|
|
mov.b32 %r141, %f444;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f445, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f446, %r144;
|
|
add.f32 %f447, %f443, %f446;
|
|
setp.gt.f32 %p20, %f445, 0f3FB504F3;
|
|
mul.f32 %f448, %f445, 0f3F000000;
|
|
add.f32 %f449, %f447, 0f3F800000;
|
|
selp.f32 %f450, %f448, %f445, %p20;
|
|
selp.f32 %f451, %f449, %f447, %p20;
|
|
add.f32 %f452, %f450, 0fBF800000;
|
|
add.f32 %f435, %f450, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f434,%f435;
|
|
// inline asm
|
|
add.f32 %f453, %f452, %f452;
|
|
mul.f32 %f454, %f434, %f453;
|
|
mul.f32 %f455, %f454, %f454;
|
|
mov.f32 %f456, 0f3C4CAF63;
|
|
mov.f32 %f457, 0f3B18F0FE;
|
|
fma.rn.f32 %f458, %f457, %f455, %f456;
|
|
mov.f32 %f459, 0f3DAAAABD;
|
|
fma.rn.f32 %f460, %f458, %f455, %f459;
|
|
mul.rn.f32 %f461, %f460, %f455;
|
|
mul.rn.f32 %f462, %f461, %f454;
|
|
sub.f32 %f463, %f452, %f454;
|
|
neg.f32 %f464, %f454;
|
|
add.f32 %f465, %f463, %f463;
|
|
fma.rn.f32 %f466, %f464, %f452, %f465;
|
|
mul.rn.f32 %f467, %f434, %f466;
|
|
add.f32 %f468, %f462, %f454;
|
|
sub.f32 %f469, %f454, %f468;
|
|
add.f32 %f470, %f462, %f469;
|
|
add.f32 %f471, %f467, %f470;
|
|
add.f32 %f472, %f468, %f471;
|
|
sub.f32 %f473, %f468, %f472;
|
|
add.f32 %f474, %f471, %f473;
|
|
mov.f32 %f475, 0f3F317200;
|
|
mul.rn.f32 %f476, %f451, %f475;
|
|
mov.f32 %f477, 0f35BFBE8E;
|
|
mul.rn.f32 %f478, %f451, %f477;
|
|
add.f32 %f479, %f476, %f472;
|
|
sub.f32 %f480, %f476, %f479;
|
|
add.f32 %f481, %f472, %f480;
|
|
add.f32 %f482, %f474, %f481;
|
|
add.f32 %f483, %f478, %f482;
|
|
add.f32 %f484, %f479, %f483;
|
|
sub.f32 %f485, %f479, %f484;
|
|
add.f32 %f486, %f483, %f485;
|
|
mul.rn.f32 %f487, %f440, %f484;
|
|
neg.f32 %f488, %f487;
|
|
fma.rn.f32 %f489, %f440, %f484, %f488;
|
|
fma.rn.f32 %f490, %f440, %f486, %f489;
|
|
mov.f32 %f491, 0f00000000;
|
|
fma.rn.f32 %f492, %f491, %f484, %f490;
|
|
add.rn.f32 %f493, %f487, %f492;
|
|
neg.f32 %f494, %f493;
|
|
add.rn.f32 %f495, %f487, %f494;
|
|
add.rn.f32 %f496, %f495, %f492;
|
|
mov.b32 %r145, %f493;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f497, %r146;
|
|
add.f32 %f498, %f496, 0f37000000;
|
|
selp.f32 %f499, %f497, %f493, %p21;
|
|
selp.f32 %f44, %f498, %f496, %p21;
|
|
mul.f32 %f500, %f499, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f501, %f500;
|
|
mov.f32 %f502, 0fBF317200;
|
|
fma.rn.f32 %f503, %f501, %f502, %f499;
|
|
mov.f32 %f504, 0fB5BFBE8E;
|
|
fma.rn.f32 %f505, %f501, %f504, %f503;
|
|
mul.f32 %f506, %f505, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f507, %f506;
|
|
add.f32 %f508, %f501, 0f00000000;
|
|
ex2.approx.f32 %f509, %f508;
|
|
mul.f32 %f510, %f507, %f509;
|
|
setp.lt.f32 %p22, %f499, 0fC2D20000;
|
|
selp.f32 %f511, 0f00000000, %f510, %p22;
|
|
setp.gt.f32 %p23, %f499, 0f42D20000;
|
|
selp.f32 %f1185, 0f7F800000, %f511, %p23;
|
|
setp.eq.f32 %p24, %f1185, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f1185, %f1185, %f44, %f1185;
|
|
|
|
BB0_11:
|
|
mov.f32 %f1162, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1161, %f1162;
|
|
add.f32 %f1160, %f1161, %f1161;
|
|
mov.f32 %f1159, 0f40800000;
|
|
sub.f32 %f1158, %f1159, %f1160;
|
|
abs.f32 %f1157, %f1158;
|
|
setp.lt.f32 %p25, %f41, 0f00000000;
|
|
setp.eq.f32 %p26, %f1157, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f1185;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f512, %r148;
|
|
selp.f32 %f1187, %f512, %f1185, %p1;
|
|
setp.eq.f32 %p27, %f41, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f515, %f41, %f41;
|
|
selp.f32 %f1187, %f515, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f41, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f1165, 0f40800000;
|
|
cvt.rzi.f32.f32 %f514, %f1165;
|
|
setp.neu.f32 %p29, %f514, 0f40800000;
|
|
selp.f32 %f1187, 0f7FFFFFFF, %f1187, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f516, %f43, 0f40800000;
|
|
mov.b32 %r149, %f516;
|
|
setp.lt.s32 %p31, %r149, 2139095040;
|
|
@%p31 bra BB0_20;
|
|
|
|
setp.gtu.f32 %p32, %f43, 0f7F800000;
|
|
@%p32 bra BB0_19;
|
|
bra.uni BB0_17;
|
|
|
|
BB0_19:
|
|
add.f32 %f1187, %f41, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f43, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f1187, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r387, 1;
|
|
mov.u64 %rd320, lightMeshBuffer;
|
|
cvta.global.u64 %rd319, %rd320;
|
|
mul.lo.s32 %r386, %r390, 3;
|
|
mov.u64 %rd318, 0;
|
|
mov.f32 %f517, 0f3F800000;
|
|
sub.f32 %f518, %f517, %f1187;
|
|
setp.eq.f32 %p34, %f41, 0f3F800000;
|
|
selp.f32 %f519, 0f00000000, %f518, %p34;
|
|
cvt.sat.f32.f32 %f520, %f519;
|
|
mul.f32 %f521, %f40, %f520;
|
|
add.s32 %r152, %r386, 1;
|
|
cvt.s64.s32 %rd61, %r152;
|
|
// inline asm
|
|
call (%rd59), _rt_buffer_get_64, (%rd319, %r387, %r119, %rd61, %rd318, %rd318, %rd318);
|
|
// inline asm
|
|
ld.f32 %f522, [%rd59];
|
|
mul.f32 %f523, %f36, %f522;
|
|
ld.f32 %f524, [%rd59+4];
|
|
mul.f32 %f525, %f37, %f524;
|
|
neg.f32 %f526, %f525;
|
|
sub.f32 %f527, %f526, %f523;
|
|
ld.f32 %f528, [%rd59+8];
|
|
mul.f32 %f529, %f38, %f528;
|
|
sub.f32 %f530, %f527, %f529;
|
|
cvt.sat.f32.f32 %f531, %f530;
|
|
mul.f32 %f55, %f521, %f531;
|
|
setp.leu.f32 %p35, %f55, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r389, 1;
|
|
mov.u64 %rd324, lightMeshBuffer;
|
|
cvta.global.u64 %rd323, %rd324;
|
|
mul.lo.s32 %r388, %r390, 3;
|
|
mov.u64 %rd322, 0;
|
|
cvt.sat.f32.f32 %f540, %f39;
|
|
add.u64 %rd66, %SP, 12;
|
|
cvta.to.local.u64 %rd73, %rd66;
|
|
max.f32 %f538, %f545, %f398;
|
|
sub.f32 %f539, %f35, %f538;
|
|
mov.u32 %r158, 1065353216;
|
|
st.local.u32 [%rd73], %r158;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f36, %f37, %f38, %r389, %f538, %f539, %rd66, %r104);
|
|
// inline asm
|
|
add.s32 %r159, %r388, 2;
|
|
cvt.s64.s32 %rd69, %r159;
|
|
// inline asm
|
|
call (%rd67), _rt_buffer_get_64, (%rd323, %r389, %r119, %rd69, %rd322, %rd322, %rd322);
|
|
// inline asm
|
|
ld.f32 %f547, [%rd67];
|
|
ld.local.f32 %f548, [%rd73];
|
|
mul.f32 %f549, %f548, %f547;
|
|
ld.f32 %f550, [%rd67+4];
|
|
mul.f32 %f551, %f548, %f550;
|
|
ld.f32 %f552, [%rd67+8];
|
|
mul.f32 %f553, %f548, %f552;
|
|
mul.f32 %f554, %f55, 0f3EA2F983;
|
|
mul.f32 %f555, %f554, %f549;
|
|
mul.f32 %f556, %f554, %f551;
|
|
mul.f32 %f557, %f554, %f553;
|
|
fma.rn.f32 %f22, %f540, %f555, %f22;
|
|
fma.rn.f32 %f21, %f540, %f556, %f21;
|
|
fma.rn.f32 %f20, %f540, %f557, %f20;
|
|
mul.f32 %f558, %f39, 0f40800000;
|
|
cvt.sat.f32.f32 %f559, %f558;
|
|
mul.f32 %f560, %f555, %f559;
|
|
mul.f32 %f561, %f556, %f559;
|
|
mul.f32 %f562, %f559, %f557;
|
|
fma.rn.f32 %f28, %f36, %f560, %f28;
|
|
fma.rn.f32 %f27, %f36, %f561, %f27;
|
|
fma.rn.f32 %f26, %f36, %f562, %f26;
|
|
fma.rn.f32 %f31, %f37, %f560, %f31;
|
|
fma.rn.f32 %f30, %f37, %f561, %f30;
|
|
fma.rn.f32 %f29, %f37, %f562, %f29;
|
|
fma.rn.f32 %f34, %f38, %f560, %f34;
|
|
fma.rn.f32 %f33, %f38, %f561, %f33;
|
|
fma.rn.f32 %f32, %f38, %f562, %f32;
|
|
add.f32 %f25, %f25, %f560;
|
|
add.f32 %f24, %f24, %f561;
|
|
add.f32 %f23, %f23, %f562;
|
|
add.f32 %f19, %f19, %f548;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r390, %r390, 1;
|
|
setp.lt.s32 %p37, %r390, %r2;
|
|
@%p37 bra BB0_6;
|
|
bra.uni BB0_24;
|
|
|
|
BB0_4:
|
|
mov.f32 %f20, %f19;
|
|
mov.f32 %f21, %f19;
|
|
mov.f32 %f22, %f19;
|
|
mov.f32 %f23, %f19;
|
|
mov.f32 %f24, %f19;
|
|
mov.f32 %f25, %f19;
|
|
mov.f32 %f26, %f19;
|
|
mov.f32 %f27, %f19;
|
|
mov.f32 %f28, %f19;
|
|
mov.f32 %f29, %f19;
|
|
mov.f32 %f30, %f19;
|
|
mov.f32 %f31, %f19;
|
|
mov.f32 %f32, %f19;
|
|
mov.f32 %f33, %f19;
|
|
mov.f32 %f34, %f19;
|
|
|
|
BB0_24:
|
|
mov.pred %p149, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f563, %r9;
|
|
mov.f32 %f564, 0f3F800000;
|
|
max.f32 %f565, %f563, %f564;
|
|
rcp.rn.f32 %f566, %f565;
|
|
mul.f32 %f1305, %f22, %f566;
|
|
mul.f32 %f1306, %f21, %f566;
|
|
mul.f32 %f1307, %f20, %f566;
|
|
div.rn.f32 %f1308, %f19, %f565;
|
|
mul.f32 %f1302, %f25, %f566;
|
|
mul.f32 %f1303, %f24, %f566;
|
|
mul.f32 %f1304, %f23, %f566;
|
|
mul.f32 %f1299, %f28, %f566;
|
|
mul.f32 %f1300, %f27, %f566;
|
|
mul.f32 %f1301, %f26, %f566;
|
|
mul.f32 %f1296, %f31, %f566;
|
|
mul.f32 %f1297, %f30, %f566;
|
|
mul.f32 %f1298, %f29, %f566;
|
|
mul.f32 %f1293, %f34, %f566;
|
|
mul.f32 %f1294, %f33, %f566;
|
|
mul.f32 %f1295, %f32, %f566;
|
|
@!%p149 bra BB0_78;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f1164, %f1168;
|
|
abs.f32 %f1163, %f1166;
|
|
setp.gt.f32 %p38, %f1163, %f1164;
|
|
neg.f32 %f582, %f1167;
|
|
selp.f32 %f583, %f582, 0f00000000, %p38;
|
|
neg.f32 %f584, %f1168;
|
|
selp.f32 %f585, %f1166, %f584, %p38;
|
|
selp.f32 %f586, 0f00000000, %f1167, %p38;
|
|
mul.f32 %f587, %f585, %f585;
|
|
fma.rn.f32 %f588, %f583, %f583, %f587;
|
|
fma.rn.f32 %f589, %f586, %f586, %f588;
|
|
sqrt.rn.f32 %f590, %f589;
|
|
rcp.rn.f32 %f591, %f590;
|
|
mul.f32 %f136, %f583, %f591;
|
|
mul.f32 %f137, %f585, %f591;
|
|
mul.f32 %f138, %f586, %f591;
|
|
mov.f32 %f581, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f1251, %f581;
|
|
mov.f32 %f1252, %f581;
|
|
mov.f32 %f1253, %f581;
|
|
mov.f32 %f1254, %f581;
|
|
mov.f32 %f1255, %f581;
|
|
mov.f32 %f1256, %f581;
|
|
mov.f32 %f1257, %f581;
|
|
mov.f32 %f1258, %f581;
|
|
mov.f32 %f1259, %f581;
|
|
mov.f32 %f1260, %f581;
|
|
mov.f32 %f1261, %f581;
|
|
mov.f32 %f1262, %f581;
|
|
mov.f32 %f1263, %f581;
|
|
mov.f32 %f1264, %f581;
|
|
mov.f32 %f1265, %f581;
|
|
@%p39 bra BB0_77;
|
|
|
|
cvt.rn.f32.s32 %f607, %r3;
|
|
rcp.rn.f32 %f139, %f607;
|
|
add.u64 %rd75, %SP, 16;
|
|
cvta.to.local.u64 %rd3, %rd75;
|
|
mul.f32 %f140, %f12, 0f3456BF95;
|
|
mul.f32 %f141, %f13, 0f3456BF95;
|
|
mul.f32 %f142, %f14, 0f3456BF95;
|
|
add.u64 %rd76, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd76;
|
|
mul.f32 %f608, %f1166, %f137;
|
|
mul.f32 %f609, %f1167, %f136;
|
|
sub.f32 %f143, %f609, %f608;
|
|
mul.f32 %f610, %f1168, %f136;
|
|
mul.f32 %f611, %f1166, %f138;
|
|
sub.f32 %f144, %f611, %f610;
|
|
mul.f32 %f612, %f1167, %f138;
|
|
mul.f32 %f613, %f1168, %f137;
|
|
sub.f32 %f145, %f613, %f612;
|
|
mov.f32 %f1251, 0f00000000;
|
|
mov.u32 %r160, 0;
|
|
abs.f32 %f614, %f141;
|
|
abs.f32 %f615, %f140;
|
|
max.f32 %f616, %f615, %f614;
|
|
abs.f32 %f617, %f142;
|
|
max.f32 %f618, %f616, %f617;
|
|
mov.u32 %r394, %r160;
|
|
mov.f32 %f1252, %f1251;
|
|
mov.f32 %f1253, %f1251;
|
|
mov.f32 %f1254, %f1251;
|
|
mov.f32 %f1255, %f1251;
|
|
mov.f32 %f1256, %f1251;
|
|
mov.f32 %f1257, %f1251;
|
|
mov.f32 %f1258, %f1251;
|
|
mov.f32 %f1259, %f1251;
|
|
mov.f32 %f1260, %f1251;
|
|
mov.f32 %f1261, %f1251;
|
|
mov.f32 %f1262, %f1251;
|
|
mov.f32 %f1263, %f1251;
|
|
mov.f32 %f1264, %f1251;
|
|
mov.f32 %f1265, %f1251;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f161, %r394;
|
|
max.f32 %f162, %f618, %f398;
|
|
mov.u32 %r396, %r160;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r162, %r397, 1664525, 1013904223;
|
|
and.b32 %r163, %r162, 16777215;
|
|
cvt.rn.f32.u32 %f620, %r163;
|
|
fma.rn.f32 %f621, %f620, 0f33800000, %f161;
|
|
mul.f32 %f178, %f139, %f621;
|
|
mad.lo.s32 %r397, %r162, 1664525, 1013904223;
|
|
and.b32 %r164, %r397, 16777215;
|
|
cvt.rn.f32.u32 %f622, %r164;
|
|
cvt.rn.f32.s32 %f623, %r396;
|
|
fma.rn.f32 %f624, %f622, 0f33800000, %f623;
|
|
mul.f32 %f625, %f139, %f624;
|
|
mul.f32 %f626, %f178, %f178;
|
|
sub.f32 %f628, %f564, %f626;
|
|
mov.f32 %f629, 0f00000000;
|
|
max.f32 %f630, %f629, %f628;
|
|
sqrt.rn.f32 %f179, %f630;
|
|
mul.f32 %f1272, %f625, 0f40C90FDB;
|
|
abs.f32 %f181, %f1272;
|
|
setp.neu.f32 %p40, %f181, 0f7F800000;
|
|
mov.f32 %f1266, %f1272;
|
|
@%p40 bra BB0_31;
|
|
|
|
mul.rn.f32 %f1266, %f1272, %f629;
|
|
|
|
BB0_31:
|
|
mul.f32 %f632, %f1266, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r407, %f632;
|
|
cvt.rn.f32.s32 %f633, %r407;
|
|
neg.f32 %f634, %f633;
|
|
mov.f32 %f635, 0f3FC90FDA;
|
|
fma.rn.f32 %f636, %f634, %f635, %f1266;
|
|
mov.f32 %f637, 0f33A22168;
|
|
fma.rn.f32 %f638, %f634, %f637, %f636;
|
|
mov.f32 %f639, 0f27C234C5;
|
|
fma.rn.f32 %f1267, %f634, %f639, %f638;
|
|
abs.f32 %f640, %f1266;
|
|
setp.leu.f32 %p41, %f640, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f1266;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r167, %r21, 8;
|
|
or.b32 %r23, %r167, -2147483648;
|
|
mov.u32 %r398, 0;
|
|
mov.u64 %rd326, 0;
|
|
mov.u64 %rd325, %rd3;
|
|
mov.u32 %r399, %r398;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd326, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r170, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r168, %r170, %r23, %r399;
|
|
madc.hi.u32 %r399, %r170, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd325], %r168;
|
|
add.s32 %r398, %r398, 1;
|
|
cvt.s64.s32 %rd326, %r398;
|
|
mul.wide.s32 %rd81, %r398, 4;
|
|
add.s64 %rd325, %rd3, %rd81;
|
|
setp.ne.s32 %p42, %r398, 6;
|
|
@%p42 bra BB0_33;
|
|
|
|
and.b32 %r173, %r22, 255;
|
|
add.s32 %r174, %r173, -128;
|
|
shr.u32 %r175, %r174, 5;
|
|
and.b32 %r28, %r21, -2147483648;
|
|
st.local.u32 [%rd3+24], %r399;
|
|
mov.u32 %r176, 6;
|
|
sub.s32 %r177, %r176, %r175;
|
|
mul.wide.s32 %rd82, %r177, 4;
|
|
add.s64 %rd10, %rd3, %rd82;
|
|
ld.local.u32 %r400, [%rd10];
|
|
ld.local.u32 %r401, [%rd10+-4];
|
|
and.b32 %r31, %r22, 31;
|
|
setp.eq.s32 %p43, %r31, 0;
|
|
@%p43 bra BB0_36;
|
|
|
|
mov.u32 %r178, 32;
|
|
sub.s32 %r179, %r178, %r31;
|
|
shr.u32 %r180, %r401, %r179;
|
|
shl.b32 %r181, %r400, %r31;
|
|
add.s32 %r400, %r180, %r181;
|
|
ld.local.u32 %r182, [%rd10+-8];
|
|
shr.u32 %r183, %r182, %r179;
|
|
shl.b32 %r184, %r401, %r31;
|
|
add.s32 %r401, %r183, %r184;
|
|
|
|
BB0_36:
|
|
shr.u32 %r185, %r401, 30;
|
|
shl.b32 %r186, %r400, 2;
|
|
add.s32 %r402, %r185, %r186;
|
|
shl.b32 %r37, %r401, 2;
|
|
shr.u32 %r187, %r402, 31;
|
|
shr.u32 %r188, %r400, 30;
|
|
add.s32 %r38, %r187, %r188;
|
|
setp.eq.s32 %p44, %r187, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r403, %r28;
|
|
mov.u32 %r404, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r189, %r402;
|
|
neg.s32 %r404, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r190, 1, 0, %p45;
|
|
add.s32 %r402, %r190, %r189;
|
|
xor.b32 %r403, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r406, %r402;
|
|
setp.eq.s32 %p46, %r406, 0;
|
|
shl.b32 %r191, %r402, %r406;
|
|
mov.u32 %r192, 32;
|
|
sub.s32 %r193, %r192, %r406;
|
|
shr.u32 %r194, %r404, %r193;
|
|
add.s32 %r195, %r194, %r191;
|
|
selp.b32 %r46, %r402, %r195, %p46;
|
|
mov.u32 %r196, -921707870;
|
|
mul.hi.u32 %r405, %r46, %r196;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r197, %r38;
|
|
selp.b32 %r407, %r38, %r197, %p47;
|
|
setp.lt.s32 %p48, %r405, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r198, %r46, -921707870;
|
|
shr.u32 %r199, %r198, 31;
|
|
shl.b32 %r200, %r405, 1;
|
|
add.s32 %r405, %r199, %r200;
|
|
add.s32 %r406, %r406, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r201, 126;
|
|
sub.s32 %r202, %r201, %r406;
|
|
shl.b32 %r203, %r202, 23;
|
|
add.s32 %r204, %r405, 1;
|
|
shr.u32 %r205, %r204, 7;
|
|
add.s32 %r206, %r205, 1;
|
|
shr.u32 %r207, %r206, 1;
|
|
add.s32 %r208, %r207, %r203;
|
|
or.b32 %r209, %r208, %r403;
|
|
mov.b32 %f1267, %r209;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f187, %f1267, %f1267;
|
|
add.s32 %r54, %r407, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f643, 0f3C08839E;
|
|
mov.f32 %f644, 0fB94CA1F9;
|
|
fma.rn.f32 %f1268, %f644, %f187, %f643;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f641, 0fBAB6061A;
|
|
mov.f32 %f642, 0f37CCF5CE;
|
|
fma.rn.f32 %f1268, %f642, %f187, %f641;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f648, 0fBE2AAAA3;
|
|
fma.rn.f32 %f649, %f1268, %f187, %f648;
|
|
fma.rn.f32 %f1269, %f649, %f187, %f629;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f645, 0f3D2AAAA5;
|
|
fma.rn.f32 %f646, %f1268, %f187, %f645;
|
|
mov.f32 %f647, 0fBF000000;
|
|
fma.rn.f32 %f1269, %f646, %f187, %f647;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f1270, %f1269, %f1267, %f1267;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f1270, %f1269, %f187, %f564;
|
|
|
|
BB0_50:
|
|
and.b32 %r210, %r54, 2;
|
|
setp.eq.s32 %p52, %r210, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f653, 0fBF800000;
|
|
fma.rn.f32 %f1270, %f1270, %f653, %f629;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mul.rn.f32 %f1272, %f1272, %f629;
|
|
|
|
BB0_54:
|
|
mul.f32 %f655, %f1272, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r417, %f655;
|
|
cvt.rn.f32.s32 %f656, %r417;
|
|
neg.f32 %f657, %f656;
|
|
fma.rn.f32 %f659, %f657, %f635, %f1272;
|
|
fma.rn.f32 %f661, %f657, %f637, %f659;
|
|
fma.rn.f32 %f1273, %f657, %f639, %f661;
|
|
abs.f32 %f663, %f1272;
|
|
setp.leu.f32 %p54, %f663, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f1272;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r213, %r57, 8;
|
|
or.b32 %r59, %r213, -2147483648;
|
|
mov.u32 %r408, 0;
|
|
mov.u64 %rd327, %rd3;
|
|
mov.u64 %rd328, %rd27;
|
|
mov.u32 %r409, %r408;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd84, %rd328, 2;
|
|
mov.u64 %rd85, __cudart_i2opi_f;
|
|
add.s64 %rd86, %rd85, %rd84;
|
|
ld.const.u32 %r216, [%rd86];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r214, %r216, %r59, %r409;
|
|
madc.hi.u32 %r409, %r216, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd327], %r214;
|
|
add.s32 %r408, %r408, 1;
|
|
cvt.s64.s32 %rd328, %r408;
|
|
mul.wide.s32 %rd87, %r408, 4;
|
|
add.s64 %rd327, %rd3, %rd87;
|
|
setp.ne.s32 %p55, %r408, 6;
|
|
@%p55 bra BB0_56;
|
|
|
|
and.b32 %r219, %r58, 255;
|
|
add.s32 %r220, %r219, -128;
|
|
shr.u32 %r221, %r220, 5;
|
|
and.b32 %r64, %r57, -2147483648;
|
|
st.local.u32 [%rd3+24], %r409;
|
|
mov.u32 %r222, 6;
|
|
sub.s32 %r223, %r222, %r221;
|
|
mul.wide.s32 %rd88, %r223, 4;
|
|
add.s64 %rd16, %rd3, %rd88;
|
|
ld.local.u32 %r410, [%rd16];
|
|
ld.local.u32 %r411, [%rd16+-4];
|
|
and.b32 %r67, %r58, 31;
|
|
setp.eq.s32 %p56, %r67, 0;
|
|
@%p56 bra BB0_59;
|
|
|
|
mov.u32 %r224, 32;
|
|
sub.s32 %r225, %r224, %r67;
|
|
shr.u32 %r226, %r411, %r225;
|
|
shl.b32 %r227, %r410, %r67;
|
|
add.s32 %r410, %r226, %r227;
|
|
ld.local.u32 %r228, [%rd16+-8];
|
|
shr.u32 %r229, %r228, %r225;
|
|
shl.b32 %r230, %r411, %r67;
|
|
add.s32 %r411, %r229, %r230;
|
|
|
|
BB0_59:
|
|
shr.u32 %r231, %r411, 30;
|
|
shl.b32 %r232, %r410, 2;
|
|
add.s32 %r412, %r231, %r232;
|
|
shl.b32 %r73, %r411, 2;
|
|
shr.u32 %r233, %r412, 31;
|
|
shr.u32 %r234, %r410, 30;
|
|
add.s32 %r74, %r233, %r234;
|
|
setp.eq.s32 %p57, %r233, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r413, %r64;
|
|
mov.u32 %r414, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r235, %r412;
|
|
neg.s32 %r414, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r236, 1, 0, %p58;
|
|
add.s32 %r412, %r236, %r235;
|
|
xor.b32 %r413, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r416, %r412;
|
|
setp.eq.s32 %p59, %r416, 0;
|
|
shl.b32 %r237, %r412, %r416;
|
|
mov.u32 %r238, 32;
|
|
sub.s32 %r239, %r238, %r416;
|
|
shr.u32 %r240, %r414, %r239;
|
|
add.s32 %r241, %r240, %r237;
|
|
selp.b32 %r82, %r412, %r241, %p59;
|
|
mov.u32 %r242, -921707870;
|
|
mul.hi.u32 %r415, %r82, %r242;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r243, %r74;
|
|
selp.b32 %r417, %r74, %r243, %p60;
|
|
setp.lt.s32 %p61, %r415, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r244, %r82, -921707870;
|
|
shr.u32 %r245, %r244, 31;
|
|
shl.b32 %r246, %r415, 1;
|
|
add.s32 %r415, %r245, %r246;
|
|
add.s32 %r416, %r416, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r247, 126;
|
|
sub.s32 %r248, %r247, %r416;
|
|
shl.b32 %r249, %r248, 23;
|
|
add.s32 %r250, %r415, 1;
|
|
shr.u32 %r251, %r250, 7;
|
|
add.s32 %r252, %r251, 1;
|
|
shr.u32 %r253, %r252, 1;
|
|
add.s32 %r254, %r253, %r249;
|
|
or.b32 %r255, %r254, %r413;
|
|
mov.b32 %f1273, %r255;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f204, %f1273, %f1273;
|
|
and.b32 %r90, %r417, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f666, 0f3C08839E;
|
|
mov.f32 %f667, 0fB94CA1F9;
|
|
fma.rn.f32 %f1274, %f667, %f204, %f666;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f664, 0fBAB6061A;
|
|
mov.f32 %f665, 0f37CCF5CE;
|
|
fma.rn.f32 %f1274, %f665, %f204, %f664;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f671, 0fBE2AAAA3;
|
|
fma.rn.f32 %f672, %f1274, %f204, %f671;
|
|
fma.rn.f32 %f1275, %f672, %f204, %f629;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f668, 0f3D2AAAA5;
|
|
fma.rn.f32 %f669, %f1274, %f204, %f668;
|
|
mov.f32 %f670, 0fBF000000;
|
|
fma.rn.f32 %f1275, %f669, %f204, %f670;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f1276, %f1275, %f1273, %f1273;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f1276, %f1275, %f204, %f564;
|
|
|
|
BB0_73:
|
|
and.b32 %r256, %r417, 2;
|
|
setp.eq.s32 %p65, %r256, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f676, 0fBF800000;
|
|
fma.rn.f32 %f1276, %f1276, %f676, %f629;
|
|
|
|
BB0_75:
|
|
mul.f32 %f685, %f179, %f1270;
|
|
mul.f32 %f686, %f179, %f1276;
|
|
mul.f32 %f687, %f136, %f686;
|
|
mul.f32 %f688, %f137, %f686;
|
|
mul.f32 %f689, %f138, %f686;
|
|
fma.rn.f32 %f690, %f145, %f685, %f687;
|
|
fma.rn.f32 %f691, %f144, %f685, %f688;
|
|
fma.rn.f32 %f692, %f143, %f685, %f689;
|
|
fma.rn.f32 %f680, %f1166, %f178, %f690;
|
|
fma.rn.f32 %f681, %f1167, %f178, %f691;
|
|
fma.rn.f32 %f682, %f1168, %f178, %f692;
|
|
mov.u32 %r258, 0;
|
|
st.local.u32 [%rd4+8], %r258;
|
|
st.local.u32 [%rd4+4], %r258;
|
|
st.local.u32 [%rd4], %r258;
|
|
ld.global.u32 %r257, [root];
|
|
mov.f32 %f684, 0f6C4ECB8F;
|
|
// inline asm
|
|
call _rt_trace_64, (%r257, %f12, %f13, %f14, %f680, %f681, %f682, %r258, %f162, %f684, %rd76, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f693, [%rd4];
|
|
max.f32 %f695, %f693, %f629;
|
|
ld.local.f32 %f696, [%rd4+4];
|
|
max.f32 %f697, %f696, %f629;
|
|
ld.local.f32 %f698, [%rd4+8];
|
|
max.f32 %f699, %f698, %f629;
|
|
fma.rn.f32 %f1259, %f680, %f695, %f1259;
|
|
fma.rn.f32 %f1258, %f680, %f697, %f1258;
|
|
fma.rn.f32 %f1257, %f680, %f699, %f1257;
|
|
fma.rn.f32 %f1262, %f681, %f695, %f1262;
|
|
fma.rn.f32 %f1261, %f681, %f697, %f1261;
|
|
fma.rn.f32 %f1260, %f681, %f699, %f1260;
|
|
fma.rn.f32 %f1265, %f682, %f695, %f1265;
|
|
fma.rn.f32 %f1264, %f682, %f697, %f1264;
|
|
fma.rn.f32 %f1263, %f682, %f699, %f1263;
|
|
add.f32 %f1256, %f1256, %f695;
|
|
add.f32 %f1255, %f1255, %f697;
|
|
add.f32 %f1254, %f1254, %f699;
|
|
mul.f32 %f700, %f1167, %f681;
|
|
fma.rn.f32 %f701, %f1166, %f680, %f700;
|
|
fma.rn.f32 %f702, %f1168, %f682, %f701;
|
|
cvt.sat.f32.f32 %f703, %f702;
|
|
fma.rn.f32 %f1253, %f695, %f703, %f1253;
|
|
fma.rn.f32 %f1252, %f697, %f703, %f1252;
|
|
fma.rn.f32 %f1251, %f699, %f703, %f1251;
|
|
add.s32 %r396, %r396, 1;
|
|
setp.lt.s32 %p66, %r396, %r3;
|
|
@%p66 bra BB0_29;
|
|
|
|
add.s32 %r394, %r394, 1;
|
|
setp.lt.s32 %p67, %r394, %r3;
|
|
@%p67 bra BB0_28;
|
|
|
|
BB0_77:
|
|
mul.lo.s32 %r260, %r3, %r3;
|
|
cvt.rn.f32.s32 %f704, %r260;
|
|
rcp.rn.f32 %f705, %f704;
|
|
mul.f32 %f706, %f1253, %f705;
|
|
mul.f32 %f707, %f1252, %f705;
|
|
mul.f32 %f708, %f1251, %f705;
|
|
div.rn.f32 %f1308, %f581, %f704;
|
|
mul.f32 %f1302, %f1256, %f705;
|
|
mul.f32 %f1303, %f1255, %f705;
|
|
mul.f32 %f1304, %f1254, %f705;
|
|
mul.f32 %f1299, %f1259, %f705;
|
|
mul.f32 %f1300, %f1258, %f705;
|
|
mul.f32 %f1301, %f1257, %f705;
|
|
mul.f32 %f1296, %f1262, %f705;
|
|
mul.f32 %f1297, %f1261, %f705;
|
|
mul.f32 %f1298, %f1260, %f705;
|
|
mul.f32 %f1293, %f1265, %f705;
|
|
mul.f32 %f1294, %f1264, %f705;
|
|
mul.f32 %f1295, %f1263, %f705;
|
|
fma.rn.f32 %f1305, %f1253, %f705, %f706;
|
|
fma.rn.f32 %f1306, %f1252, %f705, %f707;
|
|
fma.rn.f32 %f1307, %f1251, %f705, %f708;
|
|
|
|
BB0_78:
|
|
setp.gt.s32 %p68, %r1, -1;
|
|
selp.f32 %f278, %f1308, %f1305, %p68;
|
|
ld.global.u32 %r419, [imageEnabled];
|
|
and.b32 %r261, %r419, 8;
|
|
setp.eq.s32 %p69, %r261, 0;
|
|
@%p69 bra BB0_91;
|
|
|
|
cvt.u64.u32 %rd92, %r5;
|
|
cvt.u64.u32 %rd93, %r6;
|
|
mov.u64 %rd96, image_Mask;
|
|
cvta.global.u64 %rd91, %rd96;
|
|
// inline asm
|
|
call (%rd90), _rt_buffer_get_64, (%rd91, %r103, %r103, %rd92, %rd93, %rd27, %rd27);
|
|
// inline asm
|
|
mov.f32 %f712, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f713, %f712;
|
|
fma.rn.f32 %f714, %f713, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f279, %f714;
|
|
abs.f32 %f280, %f278;
|
|
setp.lt.f32 %p70, %f280, 0f00800000;
|
|
mul.f32 %f715, %f280, 0f4B800000;
|
|
selp.f32 %f716, 0fC3170000, 0fC2FE0000, %p70;
|
|
selp.f32 %f717, %f715, %f280, %p70;
|
|
mov.b32 %r264, %f717;
|
|
and.b32 %r265, %r264, 8388607;
|
|
or.b32 %r266, %r265, 1065353216;
|
|
mov.b32 %f718, %r266;
|
|
shr.u32 %r267, %r264, 23;
|
|
cvt.rn.f32.u32 %f719, %r267;
|
|
add.f32 %f720, %f716, %f719;
|
|
setp.gt.f32 %p71, %f718, 0f3FB504F3;
|
|
mul.f32 %f721, %f718, 0f3F000000;
|
|
add.f32 %f722, %f720, 0f3F800000;
|
|
selp.f32 %f723, %f721, %f718, %p71;
|
|
selp.f32 %f724, %f722, %f720, %p71;
|
|
add.f32 %f725, %f723, 0fBF800000;
|
|
add.f32 %f711, %f723, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f710,%f711;
|
|
// inline asm
|
|
add.f32 %f726, %f725, %f725;
|
|
mul.f32 %f727, %f710, %f726;
|
|
mul.f32 %f728, %f727, %f727;
|
|
mov.f32 %f729, 0f3C4CAF63;
|
|
mov.f32 %f730, 0f3B18F0FE;
|
|
fma.rn.f32 %f731, %f730, %f728, %f729;
|
|
mov.f32 %f732, 0f3DAAAABD;
|
|
fma.rn.f32 %f733, %f731, %f728, %f732;
|
|
mul.rn.f32 %f734, %f733, %f728;
|
|
mul.rn.f32 %f735, %f734, %f727;
|
|
sub.f32 %f736, %f725, %f727;
|
|
neg.f32 %f737, %f727;
|
|
add.f32 %f738, %f736, %f736;
|
|
fma.rn.f32 %f739, %f737, %f725, %f738;
|
|
mul.rn.f32 %f740, %f710, %f739;
|
|
add.f32 %f741, %f735, %f727;
|
|
sub.f32 %f742, %f727, %f741;
|
|
add.f32 %f743, %f735, %f742;
|
|
add.f32 %f744, %f740, %f743;
|
|
add.f32 %f745, %f741, %f744;
|
|
sub.f32 %f746, %f741, %f745;
|
|
add.f32 %f747, %f744, %f746;
|
|
mov.f32 %f748, 0f3F317200;
|
|
mul.rn.f32 %f749, %f724, %f748;
|
|
mov.f32 %f750, 0f35BFBE8E;
|
|
mul.rn.f32 %f751, %f724, %f750;
|
|
add.f32 %f752, %f749, %f745;
|
|
sub.f32 %f753, %f749, %f752;
|
|
add.f32 %f754, %f745, %f753;
|
|
add.f32 %f755, %f747, %f754;
|
|
add.f32 %f756, %f751, %f755;
|
|
add.f32 %f757, %f752, %f756;
|
|
sub.f32 %f758, %f752, %f757;
|
|
add.f32 %f759, %f756, %f758;
|
|
mov.f32 %f760, 0f3EE8BA2E;
|
|
mul.rn.f32 %f761, %f760, %f757;
|
|
neg.f32 %f762, %f761;
|
|
fma.rn.f32 %f763, %f760, %f757, %f762;
|
|
fma.rn.f32 %f764, %f760, %f759, %f763;
|
|
mov.f32 %f765, 0f00000000;
|
|
fma.rn.f32 %f766, %f765, %f757, %f764;
|
|
add.rn.f32 %f767, %f761, %f766;
|
|
neg.f32 %f768, %f767;
|
|
add.rn.f32 %f769, %f761, %f768;
|
|
add.rn.f32 %f770, %f769, %f766;
|
|
mov.b32 %r268, %f767;
|
|
setp.eq.s32 %p72, %r268, 1118925336;
|
|
add.s32 %r269, %r268, -1;
|
|
mov.b32 %f771, %r269;
|
|
add.f32 %f772, %f770, 0f37000000;
|
|
selp.f32 %f773, %f771, %f767, %p72;
|
|
selp.f32 %f281, %f772, %f770, %p72;
|
|
mul.f32 %f774, %f773, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f775, %f774;
|
|
mov.f32 %f776, 0fBF317200;
|
|
fma.rn.f32 %f777, %f775, %f776, %f773;
|
|
mov.f32 %f778, 0fB5BFBE8E;
|
|
fma.rn.f32 %f779, %f775, %f778, %f777;
|
|
mul.f32 %f780, %f779, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f781, %f780;
|
|
add.f32 %f782, %f775, 0f00000000;
|
|
ex2.approx.f32 %f783, %f782;
|
|
mul.f32 %f784, %f781, %f783;
|
|
setp.lt.f32 %p73, %f773, 0fC2D20000;
|
|
selp.f32 %f785, 0f00000000, %f784, %p73;
|
|
setp.gt.f32 %p74, %f773, 0f42D20000;
|
|
selp.f32 %f1309, 0f7F800000, %f785, %p74;
|
|
setp.eq.f32 %p75, %f1309, 0f7F800000;
|
|
@%p75 bra BB0_81;
|
|
|
|
fma.rn.f32 %f1309, %f1309, %f281, %f1309;
|
|
|
|
BB0_81:
|
|
setp.lt.f32 %p76, %f278, 0f00000000;
|
|
setp.eq.f32 %p77, %f279, 0f3F800000;
|
|
and.pred %p3, %p76, %p77;
|
|
mov.b32 %r270, %f1309;
|
|
xor.b32 %r271, %r270, -2147483648;
|
|
mov.b32 %f786, %r271;
|
|
selp.f32 %f1311, %f786, %f1309, %p3;
|
|
setp.eq.f32 %p78, %f278, 0f00000000;
|
|
@%p78 bra BB0_84;
|
|
bra.uni BB0_82;
|
|
|
|
BB0_84:
|
|
add.f32 %f789, %f278, %f278;
|
|
selp.f32 %f1311, %f789, 0f00000000, %p77;
|
|
bra.uni BB0_85;
|
|
|
|
BB0_149:
|
|
mov.u64 %rd237, image_HDR;
|
|
cvta.global.u64 %rd232, %rd237;
|
|
mov.u32 %r357, 8;
|
|
mov.u64 %rd236, 0;
|
|
// inline asm
|
|
call (%rd231), _rt_buffer_get_64, (%rd232, %r103, %r357, %rd20, %rd21, %rd236, %rd236);
|
|
// inline asm
|
|
mov.f32 %f1128, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1128;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd231], {%rs105, %rs105, %rs105, %rs106};
|
|
|
|
BB0_150:
|
|
ld.global.u32 %r358, [additive];
|
|
setp.eq.s32 %p145, %r358, 0;
|
|
@%p145 bra BB0_152;
|
|
|
|
mov.u64 %rd250, image_RNM0;
|
|
cvta.global.u64 %rd239, %rd250;
|
|
mov.u32 %r362, 8;
|
|
mov.u64 %rd249, 0;
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd239, %r103, %r362, %rd20, %rd21, %rd249, %rd249);
|
|
// inline asm
|
|
ld.v4.u16 {%rs113, %rs114, %rs115, %rs116}, [%rd238];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1129, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1130, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1131, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd244), _rt_buffer_get_64, (%rd239, %r103, %r362, %rd20, %rd21, %rd249, %rd249);
|
|
// inline asm
|
|
add.f32 %f1132, %f1129, 0f00000000;
|
|
add.f32 %f1133, %f1130, 0f00000000;
|
|
add.f32 %f1134, %f1131, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1134;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1133;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1132;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs117, 0;
|
|
st.v4.u16 [%rd244], {%rs110, %rs111, %rs112, %rs117};
|
|
bra.uni BB0_153;
|
|
|
|
BB0_152:
|
|
mov.u64 %rd257, image_RNM0;
|
|
cvta.global.u64 %rd252, %rd257;
|
|
mov.u32 %r364, 8;
|
|
mov.u64 %rd256, 0;
|
|
// inline asm
|
|
call (%rd251), _rt_buffer_get_64, (%rd252, %r103, %r364, %rd20, %rd21, %rd256, %rd256);
|
|
// inline asm
|
|
mov.f32 %f1135, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1135;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd251], {%rs118, %rs118, %rs118, %rs119};
|
|
|
|
BB0_153:
|
|
ld.global.u32 %r365, [additive];
|
|
setp.eq.s32 %p146, %r365, 0;
|
|
@%p146 bra BB0_155;
|
|
|
|
mov.u64 %rd270, image_RNM1;
|
|
cvta.global.u64 %rd259, %rd270;
|
|
mov.u32 %r369, 8;
|
|
mov.u64 %rd269, 0;
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd259, %r103, %r369, %rd20, %rd21, %rd269, %rd269);
|
|
// inline asm
|
|
ld.v4.u16 {%rs126, %rs127, %rs128, %rs129}, [%rd258];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1136, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1137, %rs127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1138, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd264), _rt_buffer_get_64, (%rd259, %r103, %r369, %rd20, %rd21, %rd269, %rd269);
|
|
// inline asm
|
|
add.f32 %f1139, %f1136, 0f00000000;
|
|
add.f32 %f1140, %f1137, 0f00000000;
|
|
add.f32 %f1141, %f1138, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1139;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs130, 0;
|
|
st.v4.u16 [%rd264], {%rs123, %rs124, %rs125, %rs130};
|
|
bra.uni BB0_156;
|
|
|
|
BB0_155:
|
|
mov.u64 %rd277, image_RNM1;
|
|
cvta.global.u64 %rd272, %rd277;
|
|
mov.u32 %r371, 8;
|
|
mov.u64 %rd276, 0;
|
|
// inline asm
|
|
call (%rd271), _rt_buffer_get_64, (%rd272, %r103, %r371, %rd20, %rd21, %rd276, %rd276);
|
|
// inline asm
|
|
mov.f32 %f1142, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1142;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd271], {%rs131, %rs131, %rs131, %rs132};
|
|
|
|
BB0_156:
|
|
ld.global.u32 %r372, [additive];
|
|
setp.eq.s32 %p147, %r372, 0;
|
|
@%p147 bra BB0_158;
|
|
|
|
mov.u64 %rd290, image_RNM2;
|
|
cvta.global.u64 %rd279, %rd290;
|
|
mov.u32 %r376, 8;
|
|
mov.u64 %rd289, 0;
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd279, %r103, %r376, %rd20, %rd21, %rd289, %rd289);
|
|
// inline asm
|
|
ld.v4.u16 {%rs139, %rs140, %rs141, %rs142}, [%rd278];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1143, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1144, %rs140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1145, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd284), _rt_buffer_get_64, (%rd279, %r103, %r376, %rd20, %rd21, %rd289, %rd289);
|
|
// inline asm
|
|
add.f32 %f1146, %f1143, 0f00000000;
|
|
add.f32 %f1147, %f1144, 0f00000000;
|
|
add.f32 %f1148, %f1145, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1148;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1147;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1146;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs143, 0;
|
|
st.v4.u16 [%rd284], {%rs136, %rs137, %rs138, %rs143};
|
|
bra.uni BB0_159;
|
|
|
|
BB0_158:
|
|
mov.u64 %rd297, image_RNM2;
|
|
cvta.global.u64 %rd292, %rd297;
|
|
mov.u32 %r378, 8;
|
|
mov.u64 %rd296, 0;
|
|
// inline asm
|
|
call (%rd291), _rt_buffer_get_64, (%rd292, %r103, %r378, %rd20, %rd21, %rd296, %rd296);
|
|
// inline asm
|
|
mov.f32 %f1149, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1149;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd291], {%rs144, %rs144, %rs144, %rs145};
|
|
|
|
BB0_159:
|
|
ld.global.u32 %r379, [additive];
|
|
setp.eq.s32 %p148, %r379, 0;
|
|
@%p148 bra BB0_161;
|
|
|
|
mov.u64 %rd310, image_RNM3;
|
|
cvta.global.u64 %rd299, %rd310;
|
|
mov.u32 %r383, 8;
|
|
mov.u64 %rd309, 0;
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd299, %r103, %r383, %rd20, %rd21, %rd309, %rd309);
|
|
// inline asm
|
|
ld.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd298];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1150, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1151, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1152, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd304), _rt_buffer_get_64, (%rd299, %r103, %r383, %rd20, %rd21, %rd309, %rd309);
|
|
// inline asm
|
|
add.f32 %f1153, %f1150, 0f00000000;
|
|
add.f32 %f1154, %f1151, 0f00000000;
|
|
add.f32 %f1155, %f1152, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1155;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1153;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs156, 0;
|
|
st.v4.u16 [%rd304], {%rs149, %rs150, %rs151, %rs156};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_161:
|
|
mov.u64 %rd317, image_RNM3;
|
|
cvta.global.u64 %rd312, %rd317;
|
|
mov.u32 %r385, 8;
|
|
mov.u64 %rd316, 0;
|
|
// inline asm
|
|
call (%rd311), _rt_buffer_get_64, (%rd312, %r103, %r385, %rd20, %rd21, %rd316, %rd316);
|
|
// inline asm
|
|
mov.f32 %f1156, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1156;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd311], {%rs157, %rs157, %rs157, %rs158};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_82:
|
|
setp.geu.f32 %p79, %f278, 0f00000000;
|
|
@%p79 bra BB0_85;
|
|
|
|
cvt.rzi.f32.f32 %f788, %f760;
|
|
setp.neu.f32 %p80, %f788, 0f3EE8BA2E;
|
|
selp.f32 %f1311, 0f7FFFFFFF, %f1311, %p80;
|
|
|
|
BB0_85:
|
|
add.f32 %f790, %f280, 0f3EE8BA2E;
|
|
mov.b32 %r272, %f790;
|
|
setp.lt.s32 %p82, %r272, 2139095040;
|
|
@%p82 bra BB0_90;
|
|
|
|
setp.gtu.f32 %p83, %f280, 0f7F800000;
|
|
@%p83 bra BB0_89;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_89:
|
|
add.f32 %f1311, %f278, 0f3EE8BA2E;
|
|
bra.uni BB0_90;
|
|
|
|
BB0_87:
|
|
setp.neu.f32 %p84, %f280, 0f7F800000;
|
|
@%p84 bra BB0_90;
|
|
|
|
selp.f32 %f1311, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_90:
|
|
mul.f32 %f791, %f1311, 0f437F0000;
|
|
setp.eq.f32 %p85, %f278, 0f3F800000;
|
|
selp.f32 %f792, 0f437F0000, %f791, %p85;
|
|
cvt.rzi.u32.f32 %r273, %f792;
|
|
cvt.u16.u32 %rs15, %r273;
|
|
mov.u16 %rs16, 255;
|
|
st.v2.u8 [%rd90], {%rs15, %rs16};
|
|
ld.global.u32 %r419, [imageEnabled];
|
|
|
|
BB0_91:
|
|
ld.global.f32 %f793, [lightColor];
|
|
mul.f32 %f292, %f1305, %f793;
|
|
ld.global.f32 %f794, [lightColor+4];
|
|
mul.f32 %f293, %f1306, %f794;
|
|
ld.global.f32 %f795, [lightColor+8];
|
|
mul.f32 %f294, %f1307, %f795;
|
|
and.b32 %r274, %r419, 1;
|
|
setp.eq.b32 %p86, %r274, 1;
|
|
@!%p86 bra BB0_126;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_92:
|
|
mov.f32 %f798, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f799, %f798;
|
|
fma.rn.f32 %f800, %f799, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f295, %f800;
|
|
abs.f32 %f296, %f292;
|
|
setp.lt.f32 %p87, %f296, 0f00800000;
|
|
mul.f32 %f801, %f296, 0f4B800000;
|
|
selp.f32 %f802, 0fC3170000, 0fC2FE0000, %p87;
|
|
selp.f32 %f803, %f801, %f296, %p87;
|
|
mov.b32 %r275, %f803;
|
|
and.b32 %r276, %r275, 8388607;
|
|
or.b32 %r277, %r276, 1065353216;
|
|
mov.b32 %f804, %r277;
|
|
shr.u32 %r278, %r275, 23;
|
|
cvt.rn.f32.u32 %f805, %r278;
|
|
add.f32 %f806, %f802, %f805;
|
|
setp.gt.f32 %p88, %f804, 0f3FB504F3;
|
|
mul.f32 %f807, %f804, 0f3F000000;
|
|
add.f32 %f808, %f806, 0f3F800000;
|
|
selp.f32 %f809, %f807, %f804, %p88;
|
|
selp.f32 %f810, %f808, %f806, %p88;
|
|
add.f32 %f811, %f809, 0fBF800000;
|
|
add.f32 %f797, %f809, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f796,%f797;
|
|
// inline asm
|
|
add.f32 %f812, %f811, %f811;
|
|
mul.f32 %f813, %f796, %f812;
|
|
mul.f32 %f814, %f813, %f813;
|
|
mov.f32 %f815, 0f3C4CAF63;
|
|
mov.f32 %f816, 0f3B18F0FE;
|
|
fma.rn.f32 %f817, %f816, %f814, %f815;
|
|
mov.f32 %f818, 0f3DAAAABD;
|
|
fma.rn.f32 %f819, %f817, %f814, %f818;
|
|
mul.rn.f32 %f820, %f819, %f814;
|
|
mul.rn.f32 %f821, %f820, %f813;
|
|
sub.f32 %f822, %f811, %f813;
|
|
neg.f32 %f823, %f813;
|
|
add.f32 %f824, %f822, %f822;
|
|
fma.rn.f32 %f825, %f823, %f811, %f824;
|
|
mul.rn.f32 %f826, %f796, %f825;
|
|
add.f32 %f827, %f821, %f813;
|
|
sub.f32 %f828, %f813, %f827;
|
|
add.f32 %f829, %f821, %f828;
|
|
add.f32 %f830, %f826, %f829;
|
|
add.f32 %f831, %f827, %f830;
|
|
sub.f32 %f832, %f827, %f831;
|
|
add.f32 %f833, %f830, %f832;
|
|
mov.f32 %f834, 0f3F317200;
|
|
mul.rn.f32 %f835, %f810, %f834;
|
|
mov.f32 %f836, 0f35BFBE8E;
|
|
mul.rn.f32 %f837, %f810, %f836;
|
|
add.f32 %f838, %f835, %f831;
|
|
sub.f32 %f839, %f835, %f838;
|
|
add.f32 %f840, %f831, %f839;
|
|
add.f32 %f841, %f833, %f840;
|
|
add.f32 %f842, %f837, %f841;
|
|
add.f32 %f843, %f838, %f842;
|
|
sub.f32 %f844, %f838, %f843;
|
|
add.f32 %f845, %f842, %f844;
|
|
mov.f32 %f846, 0f3EE66666;
|
|
mul.rn.f32 %f847, %f846, %f843;
|
|
neg.f32 %f848, %f847;
|
|
fma.rn.f32 %f849, %f846, %f843, %f848;
|
|
fma.rn.f32 %f850, %f846, %f845, %f849;
|
|
mov.f32 %f851, 0f00000000;
|
|
fma.rn.f32 %f852, %f851, %f843, %f850;
|
|
add.rn.f32 %f853, %f847, %f852;
|
|
neg.f32 %f854, %f853;
|
|
add.rn.f32 %f855, %f847, %f854;
|
|
add.rn.f32 %f856, %f855, %f852;
|
|
mov.b32 %r279, %f853;
|
|
setp.eq.s32 %p89, %r279, 1118925336;
|
|
add.s32 %r280, %r279, -1;
|
|
mov.b32 %f857, %r280;
|
|
add.f32 %f858, %f856, 0f37000000;
|
|
selp.f32 %f859, %f857, %f853, %p89;
|
|
selp.f32 %f297, %f858, %f856, %p89;
|
|
mul.f32 %f860, %f859, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f861, %f860;
|
|
mov.f32 %f862, 0fBF317200;
|
|
fma.rn.f32 %f863, %f861, %f862, %f859;
|
|
mov.f32 %f864, 0fB5BFBE8E;
|
|
fma.rn.f32 %f865, %f861, %f864, %f863;
|
|
mul.f32 %f866, %f865, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f867, %f866;
|
|
add.f32 %f868, %f861, 0f00000000;
|
|
ex2.approx.f32 %f869, %f868;
|
|
mul.f32 %f870, %f867, %f869;
|
|
setp.lt.f32 %p90, %f859, 0fC2D20000;
|
|
selp.f32 %f871, 0f00000000, %f870, %p90;
|
|
setp.gt.f32 %p91, %f859, 0f42D20000;
|
|
selp.f32 %f1312, 0f7F800000, %f871, %p91;
|
|
setp.eq.f32 %p92, %f1312, 0f7F800000;
|
|
@%p92 bra BB0_94;
|
|
|
|
fma.rn.f32 %f1312, %f1312, %f297, %f1312;
|
|
|
|
BB0_94:
|
|
setp.lt.f32 %p93, %f292, 0f00000000;
|
|
setp.eq.f32 %p94, %f295, 0f3F800000;
|
|
and.pred %p4, %p93, %p94;
|
|
mov.b32 %r281, %f1312;
|
|
xor.b32 %r282, %r281, -2147483648;
|
|
mov.b32 %f872, %r282;
|
|
selp.f32 %f1314, %f872, %f1312, %p4;
|
|
setp.eq.f32 %p95, %f292, 0f00000000;
|
|
@%p95 bra BB0_97;
|
|
bra.uni BB0_95;
|
|
|
|
BB0_97:
|
|
add.f32 %f875, %f292, %f292;
|
|
selp.f32 %f1314, %f875, 0f00000000, %p94;
|
|
bra.uni BB0_98;
|
|
|
|
BB0_95:
|
|
setp.geu.f32 %p96, %f292, 0f00000000;
|
|
@%p96 bra BB0_98;
|
|
|
|
cvt.rzi.f32.f32 %f874, %f846;
|
|
setp.neu.f32 %p97, %f874, 0f3EE66666;
|
|
selp.f32 %f1314, 0f7FFFFFFF, %f1314, %p97;
|
|
|
|
BB0_98:
|
|
add.f32 %f876, %f296, 0f3EE66666;
|
|
mov.b32 %r283, %f876;
|
|
setp.lt.s32 %p99, %r283, 2139095040;
|
|
@%p99 bra BB0_103;
|
|
|
|
setp.gtu.f32 %p100, %f296, 0f7F800000;
|
|
@%p100 bra BB0_102;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_102:
|
|
add.f32 %f1314, %f292, 0f3EE66666;
|
|
bra.uni BB0_103;
|
|
|
|
BB0_100:
|
|
setp.neu.f32 %p101, %f296, 0f7F800000;
|
|
@%p101 bra BB0_103;
|
|
|
|
selp.f32 %f1314, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_103:
|
|
setp.eq.f32 %p102, %f292, 0f3F800000;
|
|
selp.f32 %f308, 0f3F800000, %f1314, %p102;
|
|
abs.f32 %f309, %f293;
|
|
setp.lt.f32 %p103, %f309, 0f00800000;
|
|
mul.f32 %f879, %f309, 0f4B800000;
|
|
selp.f32 %f880, 0fC3170000, 0fC2FE0000, %p103;
|
|
selp.f32 %f881, %f879, %f309, %p103;
|
|
mov.b32 %r284, %f881;
|
|
and.b32 %r285, %r284, 8388607;
|
|
or.b32 %r286, %r285, 1065353216;
|
|
mov.b32 %f882, %r286;
|
|
shr.u32 %r287, %r284, 23;
|
|
cvt.rn.f32.u32 %f883, %r287;
|
|
add.f32 %f884, %f880, %f883;
|
|
setp.gt.f32 %p104, %f882, 0f3FB504F3;
|
|
mul.f32 %f885, %f882, 0f3F000000;
|
|
add.f32 %f886, %f884, 0f3F800000;
|
|
selp.f32 %f887, %f885, %f882, %p104;
|
|
selp.f32 %f888, %f886, %f884, %p104;
|
|
add.f32 %f889, %f887, 0fBF800000;
|
|
add.f32 %f878, %f887, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f877,%f878;
|
|
// inline asm
|
|
add.f32 %f890, %f889, %f889;
|
|
mul.f32 %f891, %f877, %f890;
|
|
mul.f32 %f892, %f891, %f891;
|
|
fma.rn.f32 %f895, %f816, %f892, %f815;
|
|
fma.rn.f32 %f897, %f895, %f892, %f818;
|
|
mul.rn.f32 %f898, %f897, %f892;
|
|
mul.rn.f32 %f899, %f898, %f891;
|
|
sub.f32 %f900, %f889, %f891;
|
|
neg.f32 %f901, %f891;
|
|
add.f32 %f902, %f900, %f900;
|
|
fma.rn.f32 %f903, %f901, %f889, %f902;
|
|
mul.rn.f32 %f904, %f877, %f903;
|
|
add.f32 %f905, %f899, %f891;
|
|
sub.f32 %f906, %f891, %f905;
|
|
add.f32 %f907, %f899, %f906;
|
|
add.f32 %f908, %f904, %f907;
|
|
add.f32 %f909, %f905, %f908;
|
|
sub.f32 %f910, %f905, %f909;
|
|
add.f32 %f911, %f908, %f910;
|
|
mul.rn.f32 %f913, %f888, %f834;
|
|
mul.rn.f32 %f915, %f888, %f836;
|
|
add.f32 %f916, %f913, %f909;
|
|
sub.f32 %f917, %f913, %f916;
|
|
add.f32 %f918, %f909, %f917;
|
|
add.f32 %f919, %f911, %f918;
|
|
add.f32 %f920, %f915, %f919;
|
|
add.f32 %f921, %f916, %f920;
|
|
sub.f32 %f922, %f916, %f921;
|
|
add.f32 %f923, %f920, %f922;
|
|
mul.rn.f32 %f925, %f846, %f921;
|
|
neg.f32 %f926, %f925;
|
|
fma.rn.f32 %f927, %f846, %f921, %f926;
|
|
fma.rn.f32 %f928, %f846, %f923, %f927;
|
|
fma.rn.f32 %f930, %f851, %f921, %f928;
|
|
add.rn.f32 %f931, %f925, %f930;
|
|
neg.f32 %f932, %f931;
|
|
add.rn.f32 %f933, %f925, %f932;
|
|
add.rn.f32 %f934, %f933, %f930;
|
|
mov.b32 %r288, %f931;
|
|
setp.eq.s32 %p105, %r288, 1118925336;
|
|
add.s32 %r289, %r288, -1;
|
|
mov.b32 %f935, %r289;
|
|
add.f32 %f936, %f934, 0f37000000;
|
|
selp.f32 %f937, %f935, %f931, %p105;
|
|
selp.f32 %f310, %f936, %f934, %p105;
|
|
mul.f32 %f938, %f937, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f939, %f938;
|
|
fma.rn.f32 %f941, %f939, %f862, %f937;
|
|
fma.rn.f32 %f943, %f939, %f864, %f941;
|
|
mul.f32 %f944, %f943, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f945, %f944;
|
|
add.f32 %f946, %f939, 0f00000000;
|
|
ex2.approx.f32 %f947, %f946;
|
|
mul.f32 %f948, %f945, %f947;
|
|
setp.lt.f32 %p106, %f937, 0fC2D20000;
|
|
selp.f32 %f949, 0f00000000, %f948, %p106;
|
|
setp.gt.f32 %p107, %f937, 0f42D20000;
|
|
selp.f32 %f1315, 0f7F800000, %f949, %p107;
|
|
setp.eq.f32 %p108, %f1315, 0f7F800000;
|
|
@%p108 bra BB0_105;
|
|
|
|
fma.rn.f32 %f1315, %f1315, %f310, %f1315;
|
|
|
|
BB0_105:
|
|
setp.lt.f32 %p109, %f293, 0f00000000;
|
|
and.pred %p5, %p109, %p94;
|
|
mov.b32 %r290, %f1315;
|
|
xor.b32 %r291, %r290, -2147483648;
|
|
mov.b32 %f950, %r291;
|
|
selp.f32 %f1317, %f950, %f1315, %p5;
|
|
setp.eq.f32 %p111, %f293, 0f00000000;
|
|
@%p111 bra BB0_108;
|
|
bra.uni BB0_106;
|
|
|
|
BB0_108:
|
|
add.f32 %f953, %f293, %f293;
|
|
selp.f32 %f1317, %f953, 0f00000000, %p94;
|
|
bra.uni BB0_109;
|
|
|
|
BB0_106:
|
|
setp.geu.f32 %p112, %f293, 0f00000000;
|
|
@%p112 bra BB0_109;
|
|
|
|
cvt.rzi.f32.f32 %f952, %f846;
|
|
setp.neu.f32 %p113, %f952, 0f3EE66666;
|
|
selp.f32 %f1317, 0f7FFFFFFF, %f1317, %p113;
|
|
|
|
BB0_109:
|
|
add.f32 %f954, %f309, 0f3EE66666;
|
|
mov.b32 %r292, %f954;
|
|
setp.lt.s32 %p115, %r292, 2139095040;
|
|
@%p115 bra BB0_114;
|
|
|
|
setp.gtu.f32 %p116, %f309, 0f7F800000;
|
|
@%p116 bra BB0_113;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_113:
|
|
add.f32 %f1317, %f293, 0f3EE66666;
|
|
bra.uni BB0_114;
|
|
|
|
BB0_111:
|
|
setp.neu.f32 %p117, %f309, 0f7F800000;
|
|
@%p117 bra BB0_114;
|
|
|
|
selp.f32 %f1317, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_114:
|
|
setp.eq.f32 %p118, %f293, 0f3F800000;
|
|
selp.f32 %f321, 0f3F800000, %f1317, %p118;
|
|
abs.f32 %f322, %f294;
|
|
setp.lt.f32 %p119, %f322, 0f00800000;
|
|
mul.f32 %f957, %f322, 0f4B800000;
|
|
selp.f32 %f958, 0fC3170000, 0fC2FE0000, %p119;
|
|
selp.f32 %f959, %f957, %f322, %p119;
|
|
mov.b32 %r293, %f959;
|
|
and.b32 %r294, %r293, 8388607;
|
|
or.b32 %r295, %r294, 1065353216;
|
|
mov.b32 %f960, %r295;
|
|
shr.u32 %r296, %r293, 23;
|
|
cvt.rn.f32.u32 %f961, %r296;
|
|
add.f32 %f962, %f958, %f961;
|
|
setp.gt.f32 %p120, %f960, 0f3FB504F3;
|
|
mul.f32 %f963, %f960, 0f3F000000;
|
|
add.f32 %f964, %f962, 0f3F800000;
|
|
selp.f32 %f965, %f963, %f960, %p120;
|
|
selp.f32 %f966, %f964, %f962, %p120;
|
|
add.f32 %f967, %f965, 0fBF800000;
|
|
add.f32 %f956, %f965, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f955,%f956;
|
|
// inline asm
|
|
add.f32 %f968, %f967, %f967;
|
|
mul.f32 %f969, %f955, %f968;
|
|
mul.f32 %f970, %f969, %f969;
|
|
fma.rn.f32 %f973, %f816, %f970, %f815;
|
|
fma.rn.f32 %f975, %f973, %f970, %f818;
|
|
mul.rn.f32 %f976, %f975, %f970;
|
|
mul.rn.f32 %f977, %f976, %f969;
|
|
sub.f32 %f978, %f967, %f969;
|
|
neg.f32 %f979, %f969;
|
|
add.f32 %f980, %f978, %f978;
|
|
fma.rn.f32 %f981, %f979, %f967, %f980;
|
|
mul.rn.f32 %f982, %f955, %f981;
|
|
add.f32 %f983, %f977, %f969;
|
|
sub.f32 %f984, %f969, %f983;
|
|
add.f32 %f985, %f977, %f984;
|
|
add.f32 %f986, %f982, %f985;
|
|
add.f32 %f987, %f983, %f986;
|
|
sub.f32 %f988, %f983, %f987;
|
|
add.f32 %f989, %f986, %f988;
|
|
mul.rn.f32 %f991, %f966, %f834;
|
|
mul.rn.f32 %f993, %f966, %f836;
|
|
add.f32 %f994, %f991, %f987;
|
|
sub.f32 %f995, %f991, %f994;
|
|
add.f32 %f996, %f987, %f995;
|
|
add.f32 %f997, %f989, %f996;
|
|
add.f32 %f998, %f993, %f997;
|
|
add.f32 %f999, %f994, %f998;
|
|
sub.f32 %f1000, %f994, %f999;
|
|
add.f32 %f1001, %f998, %f1000;
|
|
mul.rn.f32 %f1003, %f846, %f999;
|
|
neg.f32 %f1004, %f1003;
|
|
fma.rn.f32 %f1005, %f846, %f999, %f1004;
|
|
fma.rn.f32 %f1006, %f846, %f1001, %f1005;
|
|
fma.rn.f32 %f1008, %f851, %f999, %f1006;
|
|
add.rn.f32 %f1009, %f1003, %f1008;
|
|
neg.f32 %f1010, %f1009;
|
|
add.rn.f32 %f1011, %f1003, %f1010;
|
|
add.rn.f32 %f1012, %f1011, %f1008;
|
|
mov.b32 %r297, %f1009;
|
|
setp.eq.s32 %p121, %r297, 1118925336;
|
|
add.s32 %r298, %r297, -1;
|
|
mov.b32 %f1013, %r298;
|
|
add.f32 %f1014, %f1012, 0f37000000;
|
|
selp.f32 %f1015, %f1013, %f1009, %p121;
|
|
selp.f32 %f323, %f1014, %f1012, %p121;
|
|
mul.f32 %f1016, %f1015, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1017, %f1016;
|
|
fma.rn.f32 %f1019, %f1017, %f862, %f1015;
|
|
fma.rn.f32 %f1021, %f1017, %f864, %f1019;
|
|
mul.f32 %f1022, %f1021, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1023, %f1022;
|
|
add.f32 %f1024, %f1017, 0f00000000;
|
|
ex2.approx.f32 %f1025, %f1024;
|
|
mul.f32 %f1026, %f1023, %f1025;
|
|
setp.lt.f32 %p122, %f1015, 0fC2D20000;
|
|
selp.f32 %f1027, 0f00000000, %f1026, %p122;
|
|
setp.gt.f32 %p123, %f1015, 0f42D20000;
|
|
selp.f32 %f1318, 0f7F800000, %f1027, %p123;
|
|
setp.eq.f32 %p124, %f1318, 0f7F800000;
|
|
@%p124 bra BB0_116;
|
|
|
|
fma.rn.f32 %f1318, %f1318, %f323, %f1318;
|
|
|
|
BB0_116:
|
|
setp.lt.f32 %p125, %f294, 0f00000000;
|
|
and.pred %p6, %p125, %p94;
|
|
mov.b32 %r299, %f1318;
|
|
xor.b32 %r300, %r299, -2147483648;
|
|
mov.b32 %f1028, %r300;
|
|
selp.f32 %f1320, %f1028, %f1318, %p6;
|
|
setp.eq.f32 %p127, %f294, 0f00000000;
|
|
@%p127 bra BB0_119;
|
|
bra.uni BB0_117;
|
|
|
|
BB0_119:
|
|
add.f32 %f1031, %f294, %f294;
|
|
selp.f32 %f1320, %f1031, 0f00000000, %p94;
|
|
bra.uni BB0_120;
|
|
|
|
BB0_117:
|
|
setp.geu.f32 %p128, %f294, 0f00000000;
|
|
@%p128 bra BB0_120;
|
|
|
|
cvt.rzi.f32.f32 %f1030, %f846;
|
|
setp.neu.f32 %p129, %f1030, 0f3EE66666;
|
|
selp.f32 %f1320, 0f7FFFFFFF, %f1320, %p129;
|
|
|
|
BB0_120:
|
|
add.f32 %f1032, %f322, 0f3EE66666;
|
|
mov.b32 %r301, %f1032;
|
|
setp.lt.s32 %p131, %r301, 2139095040;
|
|
@%p131 bra BB0_125;
|
|
|
|
setp.gtu.f32 %p132, %f322, 0f7F800000;
|
|
@%p132 bra BB0_124;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_124:
|
|
add.f32 %f1320, %f294, 0f3EE66666;
|
|
bra.uni BB0_125;
|
|
|
|
BB0_122:
|
|
setp.neu.f32 %p133, %f322, 0f7F800000;
|
|
@%p133 bra BB0_125;
|
|
|
|
selp.f32 %f1320, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_125:
|
|
setp.eq.f32 %p134, %f294, 0f3F800000;
|
|
selp.f32 %f1033, 0f3F800000, %f1320, %p134;
|
|
cvt.u64.u32 %rd100, %r6;
|
|
cvt.u64.u32 %rd99, %r5;
|
|
mov.u64 %rd103, image;
|
|
cvta.global.u64 %rd98, %rd103;
|
|
// inline asm
|
|
call (%rd97), _rt_buffer_get_64, (%rd98, %r103, %r104, %rd99, %rd100, %rd27, %rd27);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1034, %f1033;
|
|
mul.f32 %f1035, %f1034, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r304, %f1035;
|
|
cvt.sat.f32.f32 %f1036, %f321;
|
|
mul.f32 %f1037, %f1036, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r305, %f1037;
|
|
cvt.sat.f32.f32 %f1038, %f308;
|
|
mul.f32 %f1039, %f1038, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r306, %f1039;
|
|
cvt.u16.u32 %rs17, %r304;
|
|
cvt.u16.u32 %rs18, %r306;
|
|
cvt.u16.u32 %rs19, %r305;
|
|
mov.u16 %rs20, 255;
|
|
st.v4.u8 [%rd97], {%rs17, %rs19, %rs18, %rs20};
|
|
ld.global.u32 %r419, [imageEnabled];
|
|
|
|
BB0_126:
|
|
cvt.u64.u32 %rd18, %r5;
|
|
cvt.u64.u32 %rd19, %r6;
|
|
and.b32 %r307, %r419, 4;
|
|
setp.eq.s32 %p135, %r307, 0;
|
|
@%p135 bra BB0_130;
|
|
|
|
ld.global.u32 %r308, [additive];
|
|
setp.eq.s32 %p136, %r308, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f564;}
|
|
|
|
// inline asm
|
|
@%p136 bra BB0_129;
|
|
|
|
mov.u64 %rd116, image_HDR;
|
|
cvta.global.u64 %rd105, %rd116;
|
|
mov.u32 %r312, 8;
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd105, %r103, %r312, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs28, %rs29, %rs30, %rs31}, [%rd104];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1041, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1042, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1043, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd110), _rt_buffer_get_64, (%rd105, %r103, %r312, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1044, %f292, %f1041;
|
|
add.f32 %f1045, %f293, %f1042;
|
|
add.f32 %f1046, %f294, %f1043;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1046;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f1045;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f1044;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd110], {%rs25, %rs26, %rs27, %rs21};
|
|
bra.uni BB0_130;
|
|
|
|
BB0_129:
|
|
mov.u64 %rd123, image_HDR;
|
|
cvta.global.u64 %rd118, %rd123;
|
|
mov.u32 %r314, 8;
|
|
// inline asm
|
|
call (%rd117), _rt_buffer_get_64, (%rd118, %r103, %r314, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f294;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f293;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f292;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd117], {%rs32, %rs33, %rs34, %rs21};
|
|
|
|
BB0_130:
|
|
selp.f32 %f1051, 0f3F000000, 0f3E800000, %p149;
|
|
mul.f32 %f1052, %f1051, %f1302;
|
|
mul.f32 %f1053, %f1051, %f1303;
|
|
mul.f32 %f1054, %f1051, %f1304;
|
|
mul.f32 %f1055, %f1051, %f1299;
|
|
mul.f32 %f1056, %f1051, %f1300;
|
|
mul.f32 %f1057, %f1051, %f1301;
|
|
mul.f32 %f1058, %f1051, %f1296;
|
|
mul.f32 %f1059, %f1051, %f1297;
|
|
mul.f32 %f1060, %f1051, %f1298;
|
|
mul.f32 %f1061, %f1051, %f1293;
|
|
mul.f32 %f1062, %f1051, %f1294;
|
|
mul.f32 %f1063, %f1051, %f1295;
|
|
ld.global.f32 %f1064, [lightColor];
|
|
mul.f32 %f334, %f1052, %f1064;
|
|
ld.global.f32 %f1065, [lightColor+4];
|
|
mul.f32 %f335, %f1053, %f1065;
|
|
ld.global.f32 %f1066, [lightColor+8];
|
|
mul.f32 %f336, %f1054, %f1066;
|
|
mul.f32 %f337, %f1055, %f1064;
|
|
mul.f32 %f338, %f1056, %f1065;
|
|
mul.f32 %f339, %f1057, %f1066;
|
|
mul.f32 %f340, %f1058, %f1064;
|
|
mul.f32 %f341, %f1059, %f1065;
|
|
mul.f32 %f342, %f1060, %f1066;
|
|
mul.f32 %f343, %f1061, %f1064;
|
|
mul.f32 %f344, %f1062, %f1065;
|
|
mul.f32 %f345, %f1063, %f1066;
|
|
ld.global.u32 %r315, [additive];
|
|
setp.eq.s32 %p137, %r315, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f564;}
|
|
|
|
// inline asm
|
|
@%p137 bra BB0_132;
|
|
|
|
mov.u64 %rd136, image_RNM0;
|
|
cvta.global.u64 %rd125, %rd136;
|
|
mov.u32 %r319, 8;
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd125, %r103, %r319, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd124];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1067, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1068, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1069, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd130), _rt_buffer_get_64, (%rd125, %r103, %r319, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1070, %f334, %f1067;
|
|
add.f32 %f1071, %f335, %f1068;
|
|
add.f32 %f1072, %f336, %f1069;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1072;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1071;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f1070;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd130], {%rs39, %rs40, %rs41, %rs35};
|
|
bra.uni BB0_133;
|
|
|
|
BB0_132:
|
|
mov.u64 %rd143, image_RNM0;
|
|
cvta.global.u64 %rd138, %rd143;
|
|
mov.u32 %r321, 8;
|
|
// inline asm
|
|
call (%rd137), _rt_buffer_get_64, (%rd138, %r103, %r321, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f336;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f335;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f334;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd137], {%rs46, %rs47, %rs48, %rs35};
|
|
|
|
BB0_133:
|
|
mov.f32 %f1077, 0f34000000;
|
|
max.f32 %f1078, %f334, %f1077;
|
|
div.rn.f32 %f1079, %f337, %f1078;
|
|
max.f32 %f1080, %f335, %f1077;
|
|
div.rn.f32 %f1081, %f338, %f1080;
|
|
max.f32 %f1082, %f336, %f1077;
|
|
div.rn.f32 %f1083, %f339, %f1082;
|
|
fma.rn.f32 %f346, %f1079, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f347, %f1081, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f348, %f1083, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1084, %f340, %f1078;
|
|
div.rn.f32 %f1085, %f341, %f1080;
|
|
div.rn.f32 %f1086, %f342, %f1082;
|
|
fma.rn.f32 %f349, %f1084, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f350, %f1085, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f351, %f1086, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1087, %f343, %f1078;
|
|
div.rn.f32 %f1088, %f344, %f1080;
|
|
div.rn.f32 %f1089, %f345, %f1082;
|
|
fma.rn.f32 %f352, %f1087, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f353, %f1088, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f354, %f1089, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r322, [additive];
|
|
setp.eq.s32 %p138, %r322, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f564;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_135;
|
|
|
|
mov.u64 %rd156, image_RNM1;
|
|
cvta.global.u64 %rd145, %rd156;
|
|
mov.u32 %r326, 8;
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd145, %r103, %r326, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd144];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1090, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1091, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1092, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd150), _rt_buffer_get_64, (%rd145, %r103, %r326, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1093, %f346, %f1090;
|
|
add.f32 %f1094, %f347, %f1091;
|
|
add.f32 %f1095, %f348, %f1092;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1095;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1094;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1093;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd150], {%rs53, %rs54, %rs55, %rs49};
|
|
bra.uni BB0_136;
|
|
|
|
BB0_135:
|
|
mov.u64 %rd163, image_RNM1;
|
|
cvta.global.u64 %rd158, %rd163;
|
|
mov.u32 %r328, 8;
|
|
// inline asm
|
|
call (%rd157), _rt_buffer_get_64, (%rd158, %r103, %r328, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f348;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f347;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f346;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd157], {%rs60, %rs61, %rs62, %rs49};
|
|
|
|
BB0_136:
|
|
ld.global.u32 %r329, [additive];
|
|
setp.eq.s32 %p139, %r329, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f564;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_138;
|
|
|
|
mov.u64 %rd176, image_RNM2;
|
|
cvta.global.u64 %rd165, %rd176;
|
|
mov.u32 %r333, 8;
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd165, %r103, %r333, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd164];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1100, %rs70;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1101, %rs71;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1102, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd170), _rt_buffer_get_64, (%rd165, %r103, %r333, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1103, %f349, %f1100;
|
|
add.f32 %f1104, %f350, %f1101;
|
|
add.f32 %f1105, %f351, %f1102;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1105;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1104;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f1103;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd170], {%rs67, %rs68, %rs69, %rs63};
|
|
bra.uni BB0_139;
|
|
|
|
BB0_138:
|
|
mov.u64 %rd183, image_RNM2;
|
|
cvta.global.u64 %rd178, %rd183;
|
|
mov.u32 %r335, 8;
|
|
// inline asm
|
|
call (%rd177), _rt_buffer_get_64, (%rd178, %r103, %r335, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f351;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f350;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f349;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd177], {%rs74, %rs75, %rs76, %rs63};
|
|
|
|
BB0_139:
|
|
ld.global.u32 %r336, [additive];
|
|
setp.eq.s32 %p140, %r336, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f564;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_141;
|
|
|
|
mov.u64 %rd196, image_RNM3;
|
|
cvta.global.u64 %rd185, %rd196;
|
|
mov.u32 %r340, 8;
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd185, %r103, %r340, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd184];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1110, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1111, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1112, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd190), _rt_buffer_get_64, (%rd185, %r103, %r340, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
add.f32 %f1113, %f352, %f1110;
|
|
add.f32 %f1114, %f353, %f1111;
|
|
add.f32 %f1115, %f354, %f1112;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f1114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1113;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd190], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_162;
|
|
|
|
BB0_141:
|
|
mov.u64 %rd203, image_RNM3;
|
|
cvta.global.u64 %rd198, %rd203;
|
|
mov.u32 %r342, 8;
|
|
// inline asm
|
|
call (%rd197), _rt_buffer_get_64, (%rd198, %r103, %r342, %rd18, %rd19, %rd27, %rd27);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f354;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f353;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f352;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd197], {%rs88, %rs89, %rs90, %rs77};
|
|
|
|
BB0_162:
|
|
ret;
|
|
}
|
|
|
|
|