2545 lines
71 KiB
Plaintext
2545 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 16 .b8 __local_depot0[64];
|
|
.reg .b64 %SP;
|
|
.reg .b64 %SPL;
|
|
.reg .pred %p<152>;
|
|
.reg .b16 %rs<159>;
|
|
.reg .f32 %f<1332>;
|
|
.reg .b32 %r<419>;
|
|
.reg .b64 %rd<321>;
|
|
|
|
|
|
mov.u64 %rd320, __local_depot0;
|
|
cvta.local.u64 %SP, %rd320;
|
|
ld.global.u32 %r1, [lightMeshBufferSize];
|
|
abs.s32 %r2, %r1;
|
|
ld.global.u32 %r3, [samples];
|
|
ld.global.v2.u32 {%r105, %r106}, [pixelID];
|
|
cvt.u64.u32 %rd25, %r105;
|
|
cvt.u64.u32 %rd26, %r106;
|
|
mov.u64 %rd29, uvnormal;
|
|
cvta.global.u64 %rd24, %rd29;
|
|
mov.u32 %r103, 2;
|
|
mov.u32 %r104, 4;
|
|
mov.u64 %rd28, 0;
|
|
// inline asm
|
|
call (%rd23), _rt_buffer_get_64, (%rd24, %r103, %r104, %rd25, %rd26, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r4, [%rd23];
|
|
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 %f1171, 0f00000000;
|
|
mov.f32 %f1172, %f1171;
|
|
mov.f32 %f1173, %f1171;
|
|
@%p7 bra BB0_2;
|
|
|
|
ld.u8 %rs10, [%rd23+1];
|
|
and.b16 %rs12, %rs8, 255;
|
|
cvt.rn.f32.u16 %f375, %rs12;
|
|
div.rn.f32 %f376, %f375, 0f437F0000;
|
|
fma.rn.f32 %f377, %f376, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f378, %rs10;
|
|
div.rn.f32 %f379, %f378, 0f437F0000;
|
|
fma.rn.f32 %f380, %f379, 0f40000000, 0fBF800000;
|
|
cvt.rn.f32.u16 %f381, %rs7;
|
|
div.rn.f32 %f382, %f381, 0f437F0000;
|
|
fma.rn.f32 %f383, %f382, 0f40000000, 0fBF800000;
|
|
mul.f32 %f384, %f380, %f380;
|
|
fma.rn.f32 %f385, %f377, %f377, %f384;
|
|
fma.rn.f32 %f386, %f383, %f383, %f385;
|
|
sqrt.rn.f32 %f387, %f386;
|
|
rcp.rn.f32 %f388, %f387;
|
|
mul.f32 %f1171, %f377, %f388;
|
|
mul.f32 %f1172, %f380, %f388;
|
|
mul.f32 %f1173, %f383, %f388;
|
|
|
|
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, %f1172, 0f00000000;
|
|
setp.eq.f32 %p9, %f1171, 0f00000000;
|
|
and.pred %p10, %p9, %p8;
|
|
setp.eq.f32 %p11, %f1173, 0f00000000;
|
|
and.pred %p12, %p10, %p11;
|
|
@%p12 bra BB0_144;
|
|
bra.uni BB0_3;
|
|
|
|
BB0_144:
|
|
ld.global.u32 %r418, [imageEnabled];
|
|
and.b32 %r340, %r418, 1;
|
|
setp.eq.b32 %p143, %r340, 1;
|
|
@!%p143 bra BB0_146;
|
|
bra.uni BB0_145;
|
|
|
|
BB0_145:
|
|
cvt.u64.u32 %rd200, %r5;
|
|
cvt.u64.u32 %rd201, %r6;
|
|
mov.u64 %rd204, image;
|
|
cvta.global.u64 %rd199, %rd204;
|
|
mov.u64 %rd203, 0;
|
|
// inline asm
|
|
call (%rd198), _rt_buffer_get_64, (%rd199, %r103, %r104, %rd200, %rd201, %rd203, %rd203);
|
|
// inline asm
|
|
mov.u16 %rs91, 0;
|
|
st.v4.u8 [%rd198], {%rs91, %rs91, %rs91, %rs91};
|
|
ld.global.u32 %r418, [imageEnabled];
|
|
|
|
BB0_146:
|
|
and.b32 %r343, %r418, 8;
|
|
setp.eq.s32 %p144, %r343, 0;
|
|
@%p144 bra BB0_148;
|
|
|
|
cvt.u64.u32 %rd208, %r6;
|
|
cvt.u64.u32 %rd207, %r5;
|
|
mov.u64 %rd211, image_Mask;
|
|
cvta.global.u64 %rd206, %rd211;
|
|
mov.u64 %rd210, 0;
|
|
// inline asm
|
|
call (%rd205), _rt_buffer_get_64, (%rd206, %r103, %r103, %rd207, %rd208, %rd210, %rd210);
|
|
// inline asm
|
|
mov.f32 %f1126, 0f00000000;
|
|
cvt.rzi.u32.f32 %r346, %f1126;
|
|
cvt.u16.u32 %rs92, %r346;
|
|
mov.u16 %rs93, 0;
|
|
st.v2.u8 [%rd205], {%rs92, %rs93};
|
|
ld.global.u32 %r418, [imageEnabled];
|
|
|
|
BB0_148:
|
|
cvt.u64.u32 %rd21, %r5;
|
|
cvt.u64.u32 %rd22, %r6;
|
|
and.b32 %r347, %r418, 4;
|
|
setp.eq.s32 %p145, %r347, 0;
|
|
@%p145 bra BB0_152;
|
|
|
|
ld.global.u32 %r348, [additive];
|
|
setp.eq.s32 %p146, %r348, 0;
|
|
@%p146 bra BB0_151;
|
|
|
|
mov.u64 %rd224, image_HDR;
|
|
cvta.global.u64 %rd213, %rd224;
|
|
mov.u32 %r352, 8;
|
|
mov.u64 %rd223, 0;
|
|
// inline asm
|
|
call (%rd212), _rt_buffer_get_64, (%rd213, %r103, %r352, %rd21, %rd22, %rd223, %rd223);
|
|
// inline asm
|
|
ld.v4.u16 {%rs100, %rs101, %rs102, %rs103}, [%rd212];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1127, %rs100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1128, %rs101;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1129, %rs102;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd218), _rt_buffer_get_64, (%rd213, %r103, %r352, %rd21, %rd22, %rd223, %rd223);
|
|
// inline asm
|
|
add.f32 %f1130, %f1127, 0f00000000;
|
|
add.f32 %f1131, %f1128, 0f00000000;
|
|
add.f32 %f1132, %f1129, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs99, %f1132;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs98, %f1131;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs97, %f1130;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs104, 0;
|
|
st.v4.u16 [%rd218], {%rs97, %rs98, %rs99, %rs104};
|
|
bra.uni BB0_152;
|
|
|
|
BB0_3:
|
|
ld.global.v2.u32 {%r123, %r124}, [pixelID];
|
|
cvt.u64.u32 %rd32, %r123;
|
|
cvt.u64.u32 %rd33, %r124;
|
|
mov.u64 %rd42, uvpos;
|
|
cvta.global.u64 %rd31, %rd42;
|
|
mov.u32 %r119, 12;
|
|
// inline asm
|
|
call (%rd30), _rt_buffer_get_64, (%rd31, %r103, %r119, %rd32, %rd33, %rd28, %rd28);
|
|
// inline asm
|
|
ld.f32 %f9, [%rd30+8];
|
|
ld.f32 %f8, [%rd30+4];
|
|
ld.f32 %f7, [%rd30];
|
|
mul.f32 %f405, %f7, 0f3456BF95;
|
|
mul.f32 %f406, %f8, 0f3456BF95;
|
|
mul.f32 %f407, %f9, 0f3456BF95;
|
|
abs.f32 %f10, %f1171;
|
|
div.rn.f32 %f408, %f405, %f10;
|
|
abs.f32 %f409, %f1172;
|
|
div.rn.f32 %f410, %f406, %f409;
|
|
abs.f32 %f11, %f1173;
|
|
div.rn.f32 %f411, %f407, %f11;
|
|
abs.f32 %f412, %f408;
|
|
abs.f32 %f413, %f410;
|
|
abs.f32 %f414, %f411;
|
|
mov.f32 %f415, 0f38D1B717;
|
|
max.f32 %f416, %f412, %f415;
|
|
max.f32 %f417, %f413, %f415;
|
|
max.f32 %f418, %f414, %f415;
|
|
fma.rn.f32 %f12, %f1171, %f416, %f7;
|
|
fma.rn.f32 %f13, %f1172, %f417, %f8;
|
|
fma.rn.f32 %f14, %f1173, %f418, %f9;
|
|
ld.global.v2.u32 {%r127, %r128}, [pixelID];
|
|
cvt.u64.u32 %rd38, %r127;
|
|
cvt.u64.u32 %rd39, %r128;
|
|
mov.u64 %rd43, rnd_seeds;
|
|
cvta.global.u64 %rd37, %rd43;
|
|
// inline asm
|
|
call (%rd36), _rt_buffer_get_64, (%rd37, %r103, %r104, %rd38, %rd39, %rd28, %rd28);
|
|
// inline asm
|
|
ld.u32 %r394, [%rd36];
|
|
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 %r387, 0;
|
|
abs.f32 %f558, %f17;
|
|
abs.f32 %f559, %f16;
|
|
max.f32 %f560, %f559, %f558;
|
|
abs.f32 %f561, %f18;
|
|
max.f32 %f562, %f560, %f561;
|
|
mov.u32 %r9, %r387;
|
|
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:
|
|
shl.b32 %r10, %r387, 1;
|
|
cvt.s64.s32 %rd46, %r10;
|
|
mov.u64 %rd50, lightMeshBuffer;
|
|
cvta.global.u64 %rd45, %rd50;
|
|
mov.u32 %r133, 1;
|
|
mov.u64 %rd49, 0;
|
|
// inline asm
|
|
call (%rd44), _rt_buffer_get_64, (%rd45, %r133, %r119, %rd46, %rd49, %rd49, %rd49);
|
|
// inline asm
|
|
ld.f32 %f435, [%rd44];
|
|
sub.f32 %f436, %f435, %f7;
|
|
ld.f32 %f437, [%rd44+4];
|
|
sub.f32 %f438, %f437, %f8;
|
|
ld.f32 %f439, [%rd44+8];
|
|
sub.f32 %f440, %f439, %f9;
|
|
mul.f32 %f441, %f438, %f438;
|
|
fma.rn.f32 %f442, %f436, %f436, %f441;
|
|
fma.rn.f32 %f443, %f440, %f440, %f442;
|
|
sqrt.rn.f32 %f35, %f443;
|
|
rcp.rn.f32 %f444, %f35;
|
|
mul.f32 %f36, %f436, %f444;
|
|
mul.f32 %f37, %f438, %f444;
|
|
mul.f32 %f38, %f440, %f444;
|
|
mul.f32 %f445, %f1172, %f37;
|
|
fma.rn.f32 %f446, %f1171, %f36, %f445;
|
|
fma.rn.f32 %f39, %f1173, %f38, %f446;
|
|
setp.leu.f32 %p15, %f39, 0f00000000;
|
|
@%p15 bra BB0_23;
|
|
|
|
mov.u64 %rd315, 0;
|
|
setp.eq.s32 %p16, %r3, 0;
|
|
mul.f32 %f447, %f35, %f35;
|
|
div.rn.f32 %f40, %f15, %f447;
|
|
ld.global.v2.u32 {%r137, %r138}, [pixelID];
|
|
cvt.u64.u32 %rd57, %r138;
|
|
cvt.u64.u32 %rd58, %r137;
|
|
and.b64 %rd53, %rd58, 255;
|
|
and.b64 %rd54, %rd57, 255;
|
|
mov.u64 %rd59, rnd_seeds2;
|
|
cvta.global.u64 %rd52, %rd59;
|
|
mov.u32 %r136, 3;
|
|
// inline asm
|
|
call (%rd51), _rt_buffer_get_64, (%rd52, %r103, %r136, %rd53, %rd54, %rd315, %rd315);
|
|
// inline asm
|
|
@%p16 bra BB0_9;
|
|
|
|
ld.u8 %rs14, [%rd51];
|
|
cvt.rn.f32.u16 %f448, %rs14;
|
|
div.rn.f32 %f449, %f448, 0fC37F0000;
|
|
fma.rn.f32 %f450, %f449, 0f3F333333, 0f3F800000;
|
|
setp.gt.f32 %p18, %f40, %f450;
|
|
mov.pred %p151, -1;
|
|
@%p18 bra BB0_25;
|
|
|
|
BB0_9:
|
|
ld.global.f32 %f453, [lightInvCutoff];
|
|
mul.f32 %f41, %f35, %f453;
|
|
mov.f32 %f457, 0f40800000;
|
|
abs.f32 %f43, %f41;
|
|
setp.lt.f32 %p19, %f43, 0f00800000;
|
|
mul.f32 %f459, %f43, 0f4B800000;
|
|
selp.f32 %f460, 0fC3170000, 0fC2FE0000, %p19;
|
|
selp.f32 %f461, %f459, %f43, %p19;
|
|
mov.b32 %r141, %f461;
|
|
and.b32 %r142, %r141, 8388607;
|
|
or.b32 %r143, %r142, 1065353216;
|
|
mov.b32 %f462, %r143;
|
|
shr.u32 %r144, %r141, 23;
|
|
cvt.rn.f32.u32 %f463, %r144;
|
|
add.f32 %f464, %f460, %f463;
|
|
setp.gt.f32 %p20, %f462, 0f3FB504F3;
|
|
mul.f32 %f465, %f462, 0f3F000000;
|
|
add.f32 %f466, %f464, 0f3F800000;
|
|
selp.f32 %f467, %f465, %f462, %p20;
|
|
selp.f32 %f468, %f466, %f464, %p20;
|
|
add.f32 %f469, %f467, 0fBF800000;
|
|
add.f32 %f452, %f467, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f451,%f452;
|
|
// inline asm
|
|
add.f32 %f470, %f469, %f469;
|
|
mul.f32 %f471, %f451, %f470;
|
|
mul.f32 %f472, %f471, %f471;
|
|
mov.f32 %f473, 0f3C4CAF63;
|
|
mov.f32 %f474, 0f3B18F0FE;
|
|
fma.rn.f32 %f475, %f474, %f472, %f473;
|
|
mov.f32 %f476, 0f3DAAAABD;
|
|
fma.rn.f32 %f477, %f475, %f472, %f476;
|
|
mul.rn.f32 %f478, %f477, %f472;
|
|
mul.rn.f32 %f479, %f478, %f471;
|
|
sub.f32 %f480, %f469, %f471;
|
|
neg.f32 %f481, %f471;
|
|
add.f32 %f482, %f480, %f480;
|
|
fma.rn.f32 %f483, %f481, %f469, %f482;
|
|
mul.rn.f32 %f484, %f451, %f483;
|
|
add.f32 %f485, %f479, %f471;
|
|
sub.f32 %f486, %f471, %f485;
|
|
add.f32 %f487, %f479, %f486;
|
|
add.f32 %f488, %f484, %f487;
|
|
add.f32 %f489, %f485, %f488;
|
|
sub.f32 %f490, %f485, %f489;
|
|
add.f32 %f491, %f488, %f490;
|
|
mov.f32 %f492, 0f3F317200;
|
|
mul.rn.f32 %f493, %f468, %f492;
|
|
mov.f32 %f494, 0f35BFBE8E;
|
|
mul.rn.f32 %f495, %f468, %f494;
|
|
add.f32 %f496, %f493, %f489;
|
|
sub.f32 %f497, %f493, %f496;
|
|
add.f32 %f498, %f489, %f497;
|
|
add.f32 %f499, %f491, %f498;
|
|
add.f32 %f500, %f495, %f499;
|
|
add.f32 %f501, %f496, %f500;
|
|
sub.f32 %f502, %f496, %f501;
|
|
add.f32 %f503, %f500, %f502;
|
|
mul.rn.f32 %f504, %f457, %f501;
|
|
neg.f32 %f505, %f504;
|
|
fma.rn.f32 %f506, %f457, %f501, %f505;
|
|
fma.rn.f32 %f507, %f457, %f503, %f506;
|
|
mov.f32 %f508, 0f00000000;
|
|
fma.rn.f32 %f509, %f508, %f501, %f507;
|
|
add.rn.f32 %f510, %f504, %f509;
|
|
neg.f32 %f511, %f510;
|
|
add.rn.f32 %f512, %f504, %f511;
|
|
add.rn.f32 %f513, %f512, %f509;
|
|
mov.b32 %r145, %f510;
|
|
setp.eq.s32 %p21, %r145, 1118925336;
|
|
add.s32 %r146, %r145, -1;
|
|
mov.b32 %f514, %r146;
|
|
add.f32 %f515, %f513, 0f37000000;
|
|
selp.f32 %f516, %f514, %f510, %p21;
|
|
selp.f32 %f44, %f515, %f513, %p21;
|
|
mul.f32 %f517, %f516, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f518, %f517;
|
|
mov.f32 %f519, 0fBF317200;
|
|
fma.rn.f32 %f520, %f518, %f519, %f516;
|
|
mov.f32 %f521, 0fB5BFBE8E;
|
|
fma.rn.f32 %f522, %f518, %f521, %f520;
|
|
mul.f32 %f523, %f522, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f524, %f523;
|
|
add.f32 %f525, %f518, 0f00000000;
|
|
ex2.approx.f32 %f526, %f525;
|
|
mul.f32 %f527, %f524, %f526;
|
|
setp.lt.f32 %p22, %f516, 0fC2D20000;
|
|
selp.f32 %f528, 0f00000000, %f527, %p22;
|
|
setp.gt.f32 %p23, %f516, 0f42D20000;
|
|
selp.f32 %f1190, 0f7F800000, %f528, %p23;
|
|
setp.eq.f32 %p24, %f1190, 0f7F800000;
|
|
@%p24 bra BB0_11;
|
|
|
|
fma.rn.f32 %f1190, %f1190, %f44, %f1190;
|
|
|
|
BB0_11:
|
|
mov.f32 %f1167, 0f40000000;
|
|
cvt.rzi.f32.f32 %f1166, %f1167;
|
|
add.f32 %f1165, %f1166, %f1166;
|
|
mov.f32 %f1164, 0f40800000;
|
|
sub.f32 %f1163, %f1164, %f1165;
|
|
abs.f32 %f1162, %f1163;
|
|
setp.lt.f32 %p25, %f41, 0f00000000;
|
|
setp.eq.f32 %p26, %f1162, 0f3F800000;
|
|
and.pred %p1, %p25, %p26;
|
|
mov.b32 %r147, %f1190;
|
|
xor.b32 %r148, %r147, -2147483648;
|
|
mov.b32 %f529, %r148;
|
|
selp.f32 %f1192, %f529, %f1190, %p1;
|
|
setp.eq.f32 %p27, %f41, 0f00000000;
|
|
@%p27 bra BB0_14;
|
|
bra.uni BB0_12;
|
|
|
|
BB0_14:
|
|
add.f32 %f532, %f41, %f41;
|
|
selp.f32 %f1192, %f532, 0f00000000, %p26;
|
|
bra.uni BB0_15;
|
|
|
|
BB0_12:
|
|
setp.geu.f32 %p28, %f41, 0f00000000;
|
|
@%p28 bra BB0_15;
|
|
|
|
mov.f32 %f1170, 0f40800000;
|
|
cvt.rzi.f32.f32 %f531, %f1170;
|
|
setp.neu.f32 %p29, %f531, 0f40800000;
|
|
selp.f32 %f1192, 0f7FFFFFFF, %f1192, %p29;
|
|
|
|
BB0_15:
|
|
add.f32 %f533, %f43, 0f40800000;
|
|
mov.b32 %r149, %f533;
|
|
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 %f1192, %f41, 0f40800000;
|
|
bra.uni BB0_20;
|
|
|
|
BB0_17:
|
|
setp.neu.f32 %p33, %f43, 0f7F800000;
|
|
@%p33 bra BB0_20;
|
|
|
|
selp.f32 %f1192, 0fFF800000, 0f7F800000, %p1;
|
|
|
|
BB0_20:
|
|
mov.u32 %r384, 1;
|
|
mov.u64 %rd314, lightMeshBuffer;
|
|
cvta.global.u64 %rd313, %rd314;
|
|
shl.b32 %r383, %r387, 1;
|
|
mov.u64 %rd312, 0;
|
|
mov.f32 %f534, 0f3F800000;
|
|
sub.f32 %f535, %f534, %f1192;
|
|
setp.eq.f32 %p34, %f41, 0f3F800000;
|
|
selp.f32 %f536, 0f00000000, %f535, %p34;
|
|
cvt.sat.f32.f32 %f537, %f536;
|
|
mul.f32 %f538, %f40, %f537;
|
|
add.s32 %r152, %r383, 1;
|
|
cvt.s64.s32 %rd62, %r152;
|
|
// inline asm
|
|
call (%rd60), _rt_buffer_get_64, (%rd313, %r384, %r119, %rd62, %rd312, %rd312, %rd312);
|
|
// inline asm
|
|
ld.f32 %f539, [%rd60];
|
|
mul.f32 %f540, %f36, %f539;
|
|
ld.f32 %f541, [%rd60+4];
|
|
mul.f32 %f542, %f37, %f541;
|
|
neg.f32 %f543, %f542;
|
|
sub.f32 %f544, %f543, %f540;
|
|
ld.f32 %f545, [%rd60+8];
|
|
mul.f32 %f546, %f38, %f545;
|
|
sub.f32 %f547, %f544, %f546;
|
|
cvt.sat.f32.f32 %f548, %f547;
|
|
mul.f32 %f55, %f538, %f548;
|
|
setp.leu.f32 %p35, %f55, 0f3727C5AC;
|
|
@%p35 bra BB0_22;
|
|
|
|
mov.u32 %r386, 1;
|
|
cvt.sat.f32.f32 %f557, %f39;
|
|
add.u64 %rd67, %SP, 16;
|
|
cvta.to.local.u64 %rd68, %rd67;
|
|
max.f32 %f555, %f562, %f415;
|
|
sub.f32 %f556, %f35, %f555;
|
|
mov.u32 %r156, 1065353216;
|
|
st.local.u32 [%rd68+8], %r156;
|
|
st.local.u32 [%rd68+4], %r156;
|
|
st.local.u32 [%rd68], %r156;
|
|
ld.global.u32 %r153, [root];
|
|
// inline asm
|
|
call _rt_trace_64, (%r153, %f12, %f13, %f14, %f36, %f37, %f38, %r386, %f555, %f556, %rd67, %r119);
|
|
// inline asm
|
|
ld.local.f32 %f564, [%rd68];
|
|
mul.f32 %f565, %f55, 0f3EA2F983;
|
|
mul.f32 %f566, %f565, %f564;
|
|
ld.local.f32 %f567, [%rd68+4];
|
|
mul.f32 %f568, %f565, %f567;
|
|
ld.local.f32 %f569, [%rd68+8];
|
|
mul.f32 %f570, %f565, %f569;
|
|
fma.rn.f32 %f22, %f557, %f566, %f22;
|
|
fma.rn.f32 %f21, %f557, %f568, %f21;
|
|
fma.rn.f32 %f20, %f557, %f570, %f20;
|
|
mul.f32 %f571, %f39, 0f40800000;
|
|
cvt.sat.f32.f32 %f572, %f571;
|
|
mul.f32 %f573, %f566, %f572;
|
|
mul.f32 %f574, %f568, %f572;
|
|
mul.f32 %f575, %f570, %f572;
|
|
fma.rn.f32 %f28, %f36, %f573, %f28;
|
|
fma.rn.f32 %f27, %f36, %f574, %f27;
|
|
fma.rn.f32 %f26, %f36, %f575, %f26;
|
|
fma.rn.f32 %f31, %f37, %f573, %f31;
|
|
fma.rn.f32 %f30, %f37, %f574, %f30;
|
|
fma.rn.f32 %f29, %f37, %f575, %f29;
|
|
fma.rn.f32 %f34, %f38, %f573, %f34;
|
|
fma.rn.f32 %f33, %f38, %f574, %f33;
|
|
fma.rn.f32 %f32, %f38, %f575, %f32;
|
|
add.f32 %f25, %f25, %f573;
|
|
add.f32 %f24, %f24, %f574;
|
|
add.f32 %f23, %f23, %f575;
|
|
add.f32 %f19, %f19, %f564;
|
|
|
|
BB0_22:
|
|
add.s32 %r9, %r9, 1;
|
|
|
|
BB0_23:
|
|
add.s32 %r387, %r387, 1;
|
|
setp.lt.s32 %p37, %r387, %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 %p151, %p13;
|
|
|
|
BB0_25:
|
|
cvt.rn.f32.s32 %f576, %r9;
|
|
mov.f32 %f577, 0f3F800000;
|
|
max.f32 %f578, %f576, %f577;
|
|
rcp.rn.f32 %f579, %f578;
|
|
mul.f32 %f1316, %f22, %f579;
|
|
mul.f32 %f1317, %f21, %f579;
|
|
mul.f32 %f1318, %f20, %f579;
|
|
div.rn.f32 %f1319, %f19, %f578;
|
|
mul.f32 %f1313, %f25, %f579;
|
|
mul.f32 %f1314, %f24, %f579;
|
|
mul.f32 %f1315, %f23, %f579;
|
|
mul.f32 %f1310, %f28, %f579;
|
|
mul.f32 %f1311, %f27, %f579;
|
|
mul.f32 %f1312, %f26, %f579;
|
|
mul.f32 %f1307, %f31, %f579;
|
|
mul.f32 %f1308, %f30, %f579;
|
|
mul.f32 %f1309, %f29, %f579;
|
|
mul.f32 %f1304, %f34, %f579;
|
|
mul.f32 %f1305, %f33, %f579;
|
|
mul.f32 %f1306, %f32, %f579;
|
|
@!%p151 bra BB0_80;
|
|
bra.uni BB0_26;
|
|
|
|
BB0_26:
|
|
abs.f32 %f1169, %f1173;
|
|
abs.f32 %f1168, %f1171;
|
|
setp.gt.f32 %p38, %f1168, %f1169;
|
|
neg.f32 %f596, %f1172;
|
|
selp.f32 %f597, %f596, 0f00000000, %p38;
|
|
neg.f32 %f598, %f1173;
|
|
selp.f32 %f599, %f1171, %f598, %p38;
|
|
selp.f32 %f600, 0f00000000, %f1172, %p38;
|
|
mul.f32 %f601, %f599, %f599;
|
|
fma.rn.f32 %f602, %f597, %f597, %f601;
|
|
fma.rn.f32 %f603, %f600, %f600, %f602;
|
|
sqrt.rn.f32 %f604, %f603;
|
|
rcp.rn.f32 %f605, %f604;
|
|
mul.f32 %f136, %f597, %f605;
|
|
mul.f32 %f137, %f599, %f605;
|
|
mul.f32 %f138, %f600, %f605;
|
|
mov.f32 %f1257, 0f00000000;
|
|
setp.lt.s32 %p39, %r3, 1;
|
|
mov.f32 %f1258, %f1257;
|
|
mov.f32 %f1259, %f1257;
|
|
mov.f32 %f1260, %f1257;
|
|
mov.f32 %f1261, %f1257;
|
|
mov.f32 %f1262, %f1257;
|
|
mov.f32 %f1263, %f1257;
|
|
mov.f32 %f1264, %f1257;
|
|
mov.f32 %f1265, %f1257;
|
|
mov.f32 %f1266, %f1257;
|
|
mov.f32 %f1267, %f1257;
|
|
mov.f32 %f1268, %f1257;
|
|
mov.f32 %f1269, %f1257;
|
|
mov.f32 %f1270, %f1257;
|
|
mov.f32 %f1271, %f1257;
|
|
mov.f32 %f1272, %f1257;
|
|
@%p39 bra BB0_79;
|
|
|
|
cvt.rn.f32.s32 %f622, %r3;
|
|
rcp.rn.f32 %f139, %f622;
|
|
add.u64 %rd69, %SP, 28;
|
|
cvta.to.local.u64 %rd3, %rd69;
|
|
mul.f32 %f140, %f12, 0f3456BF95;
|
|
mul.f32 %f141, %f13, 0f3456BF95;
|
|
mul.f32 %f142, %f14, 0f3456BF95;
|
|
add.u64 %rd70, %SP, 0;
|
|
cvta.to.local.u64 %rd4, %rd70;
|
|
mul.f32 %f623, %f1171, %f137;
|
|
mul.f32 %f624, %f1172, %f136;
|
|
sub.f32 %f143, %f624, %f623;
|
|
mul.f32 %f625, %f1173, %f136;
|
|
mul.f32 %f626, %f1171, %f138;
|
|
sub.f32 %f144, %f626, %f625;
|
|
mul.f32 %f627, %f1172, %f138;
|
|
mul.f32 %f628, %f1173, %f137;
|
|
sub.f32 %f145, %f628, %f627;
|
|
mov.f32 %f1257, 0f00000000;
|
|
mov.u32 %r157, 0;
|
|
abs.f32 %f629, %f141;
|
|
abs.f32 %f630, %f140;
|
|
max.f32 %f631, %f630, %f629;
|
|
abs.f32 %f632, %f142;
|
|
max.f32 %f633, %f631, %f632;
|
|
mov.u32 %r391, %r157;
|
|
mov.f32 %f1258, %f1257;
|
|
mov.f32 %f1259, %f1257;
|
|
mov.f32 %f1260, %f1257;
|
|
mov.f32 %f1261, %f1257;
|
|
mov.f32 %f1262, %f1257;
|
|
mov.f32 %f1263, %f1257;
|
|
mov.f32 %f1264, %f1257;
|
|
mov.f32 %f1265, %f1257;
|
|
mov.f32 %f1266, %f1257;
|
|
mov.f32 %f1267, %f1257;
|
|
mov.f32 %f1268, %f1257;
|
|
mov.f32 %f1269, %f1257;
|
|
mov.f32 %f1270, %f1257;
|
|
mov.f32 %f1271, %f1257;
|
|
mov.f32 %f1272, %f1257;
|
|
|
|
BB0_28:
|
|
cvt.rn.f32.s32 %f162, %r391;
|
|
max.f32 %f163, %f633, %f415;
|
|
mov.u32 %r393, %r157;
|
|
|
|
BB0_29:
|
|
mad.lo.s32 %r159, %r394, 1664525, 1013904223;
|
|
and.b32 %r160, %r159, 16777215;
|
|
cvt.rn.f32.u32 %f635, %r160;
|
|
fma.rn.f32 %f636, %f635, 0f33800000, %f162;
|
|
mul.f32 %f180, %f139, %f636;
|
|
mad.lo.s32 %r394, %r159, 1664525, 1013904223;
|
|
and.b32 %r161, %r394, 16777215;
|
|
cvt.rn.f32.u32 %f637, %r161;
|
|
cvt.rn.f32.s32 %f638, %r393;
|
|
fma.rn.f32 %f639, %f637, 0f33800000, %f638;
|
|
mul.f32 %f640, %f139, %f639;
|
|
mul.f32 %f641, %f180, %f180;
|
|
sub.f32 %f643, %f577, %f641;
|
|
mov.f32 %f1285, 0f00000000;
|
|
max.f32 %f645, %f1285, %f643;
|
|
sqrt.rn.f32 %f181, %f645;
|
|
mul.f32 %f1279, %f640, 0f40C90FDB;
|
|
abs.f32 %f183, %f1279;
|
|
setp.neu.f32 %p40, %f183, 0f7F800000;
|
|
mov.f32 %f1273, %f1279;
|
|
@%p40 bra BB0_31;
|
|
|
|
mov.f32 %f646, 0f00000000;
|
|
mul.rn.f32 %f1273, %f1279, %f646;
|
|
|
|
BB0_31:
|
|
mul.f32 %f647, %f1273, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r404, %f647;
|
|
cvt.rn.f32.s32 %f648, %r404;
|
|
neg.f32 %f649, %f648;
|
|
mov.f32 %f650, 0f3FC90FDA;
|
|
fma.rn.f32 %f651, %f649, %f650, %f1273;
|
|
mov.f32 %f652, 0f33A22168;
|
|
fma.rn.f32 %f653, %f649, %f652, %f651;
|
|
mov.f32 %f654, 0f27C234C5;
|
|
fma.rn.f32 %f1274, %f649, %f654, %f653;
|
|
abs.f32 %f655, %f1273;
|
|
setp.leu.f32 %p41, %f655, 0f47CE4780;
|
|
@%p41 bra BB0_42;
|
|
|
|
mov.b32 %r21, %f1273;
|
|
shr.u32 %r22, %r21, 23;
|
|
shl.b32 %r164, %r21, 8;
|
|
or.b32 %r23, %r164, -2147483648;
|
|
mov.u32 %r395, 0;
|
|
mov.u64 %rd317, 0;
|
|
mov.u64 %rd316, %rd3;
|
|
mov.u32 %r396, %r395;
|
|
|
|
BB0_33:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd72, %rd317, 2;
|
|
mov.u64 %rd73, __cudart_i2opi_f;
|
|
add.s64 %rd74, %rd73, %rd72;
|
|
ld.const.u32 %r167, [%rd74];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r165, %r167, %r23, %r396;
|
|
madc.hi.u32 %r396, %r167, %r23, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd316], %r165;
|
|
add.s32 %r395, %r395, 1;
|
|
cvt.s64.s32 %rd317, %r395;
|
|
mul.wide.s32 %rd75, %r395, 4;
|
|
add.s64 %rd316, %rd3, %rd75;
|
|
setp.ne.s32 %p42, %r395, 6;
|
|
@%p42 bra BB0_33;
|
|
|
|
and.b32 %r170, %r22, 255;
|
|
add.s32 %r171, %r170, -128;
|
|
shr.u32 %r172, %r171, 5;
|
|
and.b32 %r28, %r21, -2147483648;
|
|
st.local.u32 [%rd3+24], %r396;
|
|
mov.u32 %r173, 6;
|
|
sub.s32 %r174, %r173, %r172;
|
|
mul.wide.s32 %rd76, %r174, 4;
|
|
add.s64 %rd10, %rd3, %rd76;
|
|
ld.local.u32 %r397, [%rd10];
|
|
ld.local.u32 %r398, [%rd10+-4];
|
|
and.b32 %r31, %r22, 31;
|
|
setp.eq.s32 %p43, %r31, 0;
|
|
@%p43 bra BB0_36;
|
|
|
|
mov.u32 %r175, 32;
|
|
sub.s32 %r176, %r175, %r31;
|
|
shr.u32 %r177, %r398, %r176;
|
|
shl.b32 %r178, %r397, %r31;
|
|
add.s32 %r397, %r177, %r178;
|
|
ld.local.u32 %r179, [%rd10+-8];
|
|
shr.u32 %r180, %r179, %r176;
|
|
shl.b32 %r181, %r398, %r31;
|
|
add.s32 %r398, %r180, %r181;
|
|
|
|
BB0_36:
|
|
shr.u32 %r182, %r398, 30;
|
|
shl.b32 %r183, %r397, 2;
|
|
add.s32 %r399, %r182, %r183;
|
|
shl.b32 %r37, %r398, 2;
|
|
shr.u32 %r184, %r399, 31;
|
|
shr.u32 %r185, %r397, 30;
|
|
add.s32 %r38, %r184, %r185;
|
|
setp.eq.s32 %p44, %r184, 0;
|
|
@%p44 bra BB0_37;
|
|
bra.uni BB0_38;
|
|
|
|
BB0_37:
|
|
mov.u32 %r400, %r28;
|
|
mov.u32 %r401, %r37;
|
|
bra.uni BB0_39;
|
|
|
|
BB0_38:
|
|
not.b32 %r186, %r399;
|
|
neg.s32 %r401, %r37;
|
|
setp.eq.s32 %p45, %r37, 0;
|
|
selp.u32 %r187, 1, 0, %p45;
|
|
add.s32 %r399, %r187, %r186;
|
|
xor.b32 %r400, %r28, -2147483648;
|
|
|
|
BB0_39:
|
|
clz.b32 %r403, %r399;
|
|
setp.eq.s32 %p46, %r403, 0;
|
|
shl.b32 %r188, %r399, %r403;
|
|
mov.u32 %r189, 32;
|
|
sub.s32 %r190, %r189, %r403;
|
|
shr.u32 %r191, %r401, %r190;
|
|
add.s32 %r192, %r191, %r188;
|
|
selp.b32 %r46, %r399, %r192, %p46;
|
|
mov.u32 %r193, -921707870;
|
|
mul.hi.u32 %r402, %r46, %r193;
|
|
setp.eq.s32 %p47, %r28, 0;
|
|
neg.s32 %r194, %r38;
|
|
selp.b32 %r404, %r38, %r194, %p47;
|
|
setp.lt.s32 %p48, %r402, 1;
|
|
@%p48 bra BB0_41;
|
|
|
|
mul.lo.s32 %r195, %r46, -921707870;
|
|
shr.u32 %r196, %r195, 31;
|
|
shl.b32 %r197, %r402, 1;
|
|
add.s32 %r402, %r196, %r197;
|
|
add.s32 %r403, %r403, 1;
|
|
|
|
BB0_41:
|
|
mov.u32 %r198, 126;
|
|
sub.s32 %r199, %r198, %r403;
|
|
shl.b32 %r200, %r199, 23;
|
|
add.s32 %r201, %r402, 1;
|
|
shr.u32 %r202, %r201, 7;
|
|
add.s32 %r203, %r202, 1;
|
|
shr.u32 %r204, %r203, 1;
|
|
add.s32 %r205, %r204, %r200;
|
|
or.b32 %r206, %r205, %r400;
|
|
mov.b32 %f1274, %r206;
|
|
|
|
BB0_42:
|
|
mul.rn.f32 %f189, %f1274, %f1274;
|
|
add.s32 %r54, %r404, 1;
|
|
and.b32 %r55, %r54, 1;
|
|
setp.eq.s32 %p49, %r55, 0;
|
|
@%p49 bra BB0_44;
|
|
bra.uni BB0_43;
|
|
|
|
BB0_44:
|
|
mov.f32 %f658, 0f3C08839E;
|
|
mov.f32 %f659, 0fB94CA1F9;
|
|
fma.rn.f32 %f1275, %f659, %f189, %f658;
|
|
bra.uni BB0_45;
|
|
|
|
BB0_43:
|
|
mov.f32 %f656, 0fBAB6061A;
|
|
mov.f32 %f657, 0f37CCF5CE;
|
|
fma.rn.f32 %f1275, %f657, %f189, %f656;
|
|
|
|
BB0_45:
|
|
@%p49 bra BB0_47;
|
|
bra.uni BB0_46;
|
|
|
|
BB0_47:
|
|
mov.f32 %f663, 0fBE2AAAA3;
|
|
fma.rn.f32 %f664, %f1275, %f189, %f663;
|
|
mov.f32 %f665, 0f00000000;
|
|
fma.rn.f32 %f1276, %f664, %f189, %f665;
|
|
bra.uni BB0_48;
|
|
|
|
BB0_46:
|
|
mov.f32 %f660, 0f3D2AAAA5;
|
|
fma.rn.f32 %f661, %f1275, %f189, %f660;
|
|
mov.f32 %f662, 0fBF000000;
|
|
fma.rn.f32 %f1276, %f661, %f189, %f662;
|
|
|
|
BB0_48:
|
|
fma.rn.f32 %f1277, %f1276, %f1274, %f1274;
|
|
@%p49 bra BB0_50;
|
|
|
|
fma.rn.f32 %f1277, %f1276, %f189, %f577;
|
|
|
|
BB0_50:
|
|
and.b32 %r207, %r54, 2;
|
|
setp.eq.s32 %p52, %r207, 0;
|
|
@%p52 bra BB0_52;
|
|
|
|
mov.f32 %f668, 0fBF800000;
|
|
fma.rn.f32 %f1277, %f1277, %f668, %f1285;
|
|
|
|
BB0_52:
|
|
@%p40 bra BB0_54;
|
|
|
|
mov.f32 %f669, 0f00000000;
|
|
mul.rn.f32 %f1279, %f1279, %f669;
|
|
|
|
BB0_54:
|
|
mul.f32 %f670, %f1279, 0f3F22F983;
|
|
cvt.rni.s32.f32 %r414, %f670;
|
|
cvt.rn.f32.s32 %f671, %r414;
|
|
neg.f32 %f672, %f671;
|
|
fma.rn.f32 %f674, %f672, %f650, %f1279;
|
|
fma.rn.f32 %f676, %f672, %f652, %f674;
|
|
fma.rn.f32 %f1280, %f672, %f654, %f676;
|
|
abs.f32 %f678, %f1279;
|
|
setp.leu.f32 %p54, %f678, 0f47CE4780;
|
|
@%p54 bra BB0_65;
|
|
|
|
mov.b32 %r57, %f1279;
|
|
shr.u32 %r58, %r57, 23;
|
|
shl.b32 %r210, %r57, 8;
|
|
or.b32 %r59, %r210, -2147483648;
|
|
mov.u32 %r405, 0;
|
|
mov.u64 %rd318, %rd3;
|
|
mov.u64 %rd319, %rd28;
|
|
mov.u32 %r406, %r405;
|
|
|
|
BB0_56:
|
|
.pragma "nounroll";
|
|
shl.b64 %rd78, %rd319, 2;
|
|
mov.u64 %rd79, __cudart_i2opi_f;
|
|
add.s64 %rd80, %rd79, %rd78;
|
|
ld.const.u32 %r213, [%rd80];
|
|
// inline asm
|
|
{
|
|
mad.lo.cc.u32 %r211, %r213, %r59, %r406;
|
|
madc.hi.u32 %r406, %r213, %r59, 0;
|
|
}
|
|
// inline asm
|
|
st.local.u32 [%rd318], %r211;
|
|
add.s32 %r405, %r405, 1;
|
|
cvt.s64.s32 %rd319, %r405;
|
|
mul.wide.s32 %rd81, %r405, 4;
|
|
add.s64 %rd318, %rd3, %rd81;
|
|
setp.ne.s32 %p55, %r405, 6;
|
|
@%p55 bra BB0_56;
|
|
|
|
and.b32 %r216, %r58, 255;
|
|
add.s32 %r217, %r216, -128;
|
|
shr.u32 %r218, %r217, 5;
|
|
and.b32 %r64, %r57, -2147483648;
|
|
st.local.u32 [%rd3+24], %r406;
|
|
mov.u32 %r219, 6;
|
|
sub.s32 %r220, %r219, %r218;
|
|
mul.wide.s32 %rd82, %r220, 4;
|
|
add.s64 %rd16, %rd3, %rd82;
|
|
ld.local.u32 %r407, [%rd16];
|
|
ld.local.u32 %r408, [%rd16+-4];
|
|
and.b32 %r67, %r58, 31;
|
|
setp.eq.s32 %p56, %r67, 0;
|
|
@%p56 bra BB0_59;
|
|
|
|
mov.u32 %r221, 32;
|
|
sub.s32 %r222, %r221, %r67;
|
|
shr.u32 %r223, %r408, %r222;
|
|
shl.b32 %r224, %r407, %r67;
|
|
add.s32 %r407, %r223, %r224;
|
|
ld.local.u32 %r225, [%rd16+-8];
|
|
shr.u32 %r226, %r225, %r222;
|
|
shl.b32 %r227, %r408, %r67;
|
|
add.s32 %r408, %r226, %r227;
|
|
|
|
BB0_59:
|
|
shr.u32 %r228, %r408, 30;
|
|
shl.b32 %r229, %r407, 2;
|
|
add.s32 %r409, %r228, %r229;
|
|
shl.b32 %r73, %r408, 2;
|
|
shr.u32 %r230, %r409, 31;
|
|
shr.u32 %r231, %r407, 30;
|
|
add.s32 %r74, %r230, %r231;
|
|
setp.eq.s32 %p57, %r230, 0;
|
|
@%p57 bra BB0_60;
|
|
bra.uni BB0_61;
|
|
|
|
BB0_60:
|
|
mov.u32 %r410, %r64;
|
|
mov.u32 %r411, %r73;
|
|
bra.uni BB0_62;
|
|
|
|
BB0_61:
|
|
not.b32 %r232, %r409;
|
|
neg.s32 %r411, %r73;
|
|
setp.eq.s32 %p58, %r73, 0;
|
|
selp.u32 %r233, 1, 0, %p58;
|
|
add.s32 %r409, %r233, %r232;
|
|
xor.b32 %r410, %r64, -2147483648;
|
|
|
|
BB0_62:
|
|
clz.b32 %r413, %r409;
|
|
setp.eq.s32 %p59, %r413, 0;
|
|
shl.b32 %r234, %r409, %r413;
|
|
mov.u32 %r235, 32;
|
|
sub.s32 %r236, %r235, %r413;
|
|
shr.u32 %r237, %r411, %r236;
|
|
add.s32 %r238, %r237, %r234;
|
|
selp.b32 %r82, %r409, %r238, %p59;
|
|
mov.u32 %r239, -921707870;
|
|
mul.hi.u32 %r412, %r82, %r239;
|
|
setp.eq.s32 %p60, %r64, 0;
|
|
neg.s32 %r240, %r74;
|
|
selp.b32 %r414, %r74, %r240, %p60;
|
|
setp.lt.s32 %p61, %r412, 1;
|
|
@%p61 bra BB0_64;
|
|
|
|
mul.lo.s32 %r241, %r82, -921707870;
|
|
shr.u32 %r242, %r241, 31;
|
|
shl.b32 %r243, %r412, 1;
|
|
add.s32 %r412, %r242, %r243;
|
|
add.s32 %r413, %r413, 1;
|
|
|
|
BB0_64:
|
|
mov.u32 %r244, 126;
|
|
sub.s32 %r245, %r244, %r413;
|
|
shl.b32 %r246, %r245, 23;
|
|
add.s32 %r247, %r412, 1;
|
|
shr.u32 %r248, %r247, 7;
|
|
add.s32 %r249, %r248, 1;
|
|
shr.u32 %r250, %r249, 1;
|
|
add.s32 %r251, %r250, %r246;
|
|
or.b32 %r252, %r251, %r410;
|
|
mov.b32 %f1280, %r252;
|
|
|
|
BB0_65:
|
|
mul.rn.f32 %f206, %f1280, %f1280;
|
|
and.b32 %r90, %r414, 1;
|
|
setp.eq.s32 %p62, %r90, 0;
|
|
@%p62 bra BB0_67;
|
|
bra.uni BB0_66;
|
|
|
|
BB0_67:
|
|
mov.f32 %f681, 0f3C08839E;
|
|
mov.f32 %f682, 0fB94CA1F9;
|
|
fma.rn.f32 %f1281, %f682, %f206, %f681;
|
|
bra.uni BB0_68;
|
|
|
|
BB0_66:
|
|
mov.f32 %f679, 0fBAB6061A;
|
|
mov.f32 %f680, 0f37CCF5CE;
|
|
fma.rn.f32 %f1281, %f680, %f206, %f679;
|
|
|
|
BB0_68:
|
|
@%p62 bra BB0_70;
|
|
bra.uni BB0_69;
|
|
|
|
BB0_70:
|
|
mov.f32 %f686, 0fBE2AAAA3;
|
|
fma.rn.f32 %f687, %f1281, %f206, %f686;
|
|
mov.f32 %f688, 0f00000000;
|
|
fma.rn.f32 %f1282, %f687, %f206, %f688;
|
|
bra.uni BB0_71;
|
|
|
|
BB0_69:
|
|
mov.f32 %f683, 0f3D2AAAA5;
|
|
fma.rn.f32 %f684, %f1281, %f206, %f683;
|
|
mov.f32 %f685, 0fBF000000;
|
|
fma.rn.f32 %f1282, %f684, %f206, %f685;
|
|
|
|
BB0_71:
|
|
fma.rn.f32 %f1283, %f1282, %f1280, %f1280;
|
|
@%p62 bra BB0_73;
|
|
|
|
fma.rn.f32 %f1283, %f1282, %f206, %f577;
|
|
|
|
BB0_73:
|
|
and.b32 %r253, %r414, 2;
|
|
setp.eq.s32 %p65, %r253, 0;
|
|
@%p65 bra BB0_75;
|
|
|
|
mov.f32 %f691, 0fBF800000;
|
|
fma.rn.f32 %f1283, %f1283, %f691, %f1285;
|
|
|
|
BB0_75:
|
|
mul.f32 %f703, %f181, %f1277;
|
|
mul.f32 %f704, %f181, %f1283;
|
|
mul.f32 %f705, %f136, %f704;
|
|
mul.f32 %f706, %f137, %f704;
|
|
mul.f32 %f707, %f138, %f704;
|
|
fma.rn.f32 %f708, %f145, %f703, %f705;
|
|
fma.rn.f32 %f709, %f144, %f703, %f706;
|
|
fma.rn.f32 %f710, %f143, %f703, %f707;
|
|
fma.rn.f32 %f218, %f1171, %f180, %f708;
|
|
fma.rn.f32 %f219, %f1172, %f180, %f709;
|
|
fma.rn.f32 %f220, %f1173, %f180, %f710;
|
|
mov.f32 %f711, 0f7F7FFFFF;
|
|
st.local.v4.f32 [%rd4], {%f577, %f577, %f577, %f711};
|
|
ld.global.u32 %r254, [root];
|
|
mov.u32 %r255, 0;
|
|
mov.f32 %f699, 0f6C4ECB8F;
|
|
mov.u32 %r256, 16;
|
|
// inline asm
|
|
call _rt_trace_64, (%r254, %f12, %f13, %f14, %f218, %f219, %f220, %r255, %f163, %f699, %rd70, %r256);
|
|
// inline asm
|
|
ld.local.f32 %f221, [%rd4+12];
|
|
setp.ge.f32 %p66, %f221, 0f00000000;
|
|
mov.f32 %f1286, %f1285;
|
|
mov.f32 %f1287, %f1285;
|
|
@%p66 bra BB0_77;
|
|
|
|
ld.local.v4.f32 {%f1287, %f1286, %f1285, %f716}, [%rd4];
|
|
|
|
BB0_77:
|
|
setp.ltu.f32 %p67, %f221, 0f00000000;
|
|
selp.f32 %f717, 0f3F800000, 0f00000000, %p67;
|
|
add.f32 %f1257, %f1257, %f717;
|
|
fma.rn.f32 %f1266, %f218, %f1287, %f1266;
|
|
fma.rn.f32 %f1265, %f218, %f1286, %f1265;
|
|
fma.rn.f32 %f1264, %f218, %f1285, %f1264;
|
|
fma.rn.f32 %f1269, %f219, %f1287, %f1269;
|
|
fma.rn.f32 %f1268, %f219, %f1286, %f1268;
|
|
fma.rn.f32 %f1267, %f219, %f1285, %f1267;
|
|
fma.rn.f32 %f1272, %f220, %f1287, %f1272;
|
|
fma.rn.f32 %f1271, %f220, %f1286, %f1271;
|
|
fma.rn.f32 %f1270, %f220, %f1285, %f1270;
|
|
add.f32 %f1263, %f1263, %f1287;
|
|
add.f32 %f1262, %f1262, %f1286;
|
|
add.f32 %f1261, %f1261, %f1285;
|
|
mul.f32 %f718, %f1172, %f219;
|
|
fma.rn.f32 %f719, %f1171, %f218, %f718;
|
|
fma.rn.f32 %f720, %f1173, %f220, %f719;
|
|
cvt.sat.f32.f32 %f721, %f720;
|
|
fma.rn.f32 %f1260, %f1287, %f721, %f1260;
|
|
fma.rn.f32 %f1259, %f1286, %f721, %f1259;
|
|
fma.rn.f32 %f1258, %f1285, %f721, %f1258;
|
|
add.s32 %r393, %r393, 1;
|
|
setp.lt.s32 %p68, %r393, %r3;
|
|
@%p68 bra BB0_29;
|
|
|
|
add.s32 %r391, %r391, 1;
|
|
setp.lt.s32 %p69, %r391, %r3;
|
|
@%p69 bra BB0_28;
|
|
|
|
BB0_79:
|
|
mul.lo.s32 %r257, %r3, %r3;
|
|
cvt.rn.f32.s32 %f722, %r257;
|
|
rcp.rn.f32 %f723, %f722;
|
|
mul.f32 %f724, %f1260, %f723;
|
|
mul.f32 %f725, %f1259, %f723;
|
|
mul.f32 %f726, %f1258, %f723;
|
|
div.rn.f32 %f1319, %f1257, %f722;
|
|
mul.f32 %f1313, %f1263, %f723;
|
|
mul.f32 %f1314, %f1262, %f723;
|
|
mul.f32 %f1315, %f1261, %f723;
|
|
mul.f32 %f1310, %f1266, %f723;
|
|
mul.f32 %f1311, %f1265, %f723;
|
|
mul.f32 %f1312, %f1264, %f723;
|
|
mul.f32 %f1307, %f1269, %f723;
|
|
mul.f32 %f1308, %f1268, %f723;
|
|
mul.f32 %f1309, %f1267, %f723;
|
|
mul.f32 %f1304, %f1272, %f723;
|
|
mul.f32 %f1305, %f1271, %f723;
|
|
mul.f32 %f1306, %f1270, %f723;
|
|
fma.rn.f32 %f1316, %f1260, %f723, %f724;
|
|
fma.rn.f32 %f1317, %f1259, %f723, %f725;
|
|
fma.rn.f32 %f1318, %f1258, %f723, %f726;
|
|
|
|
BB0_80:
|
|
setp.gt.s32 %p70, %r1, -1;
|
|
selp.f32 %f292, %f1319, %f1316, %p70;
|
|
ld.global.u32 %r416, [imageEnabled];
|
|
and.b32 %r258, %r416, 8;
|
|
setp.eq.s32 %p71, %r258, 0;
|
|
@%p71 bra BB0_93;
|
|
|
|
cvt.u64.u32 %rd86, %r5;
|
|
cvt.u64.u32 %rd87, %r6;
|
|
mov.u64 %rd90, image_Mask;
|
|
cvta.global.u64 %rd85, %rd90;
|
|
// inline asm
|
|
call (%rd84), _rt_buffer_get_64, (%rd85, %r103, %r103, %rd86, %rd87, %rd28, %rd28);
|
|
// inline asm
|
|
mov.f32 %f729, 0f3E68BA2E;
|
|
cvt.rzi.f32.f32 %f730, %f729;
|
|
fma.rn.f32 %f731, %f730, 0fC0000000, 0f3EE8BA2E;
|
|
abs.f32 %f293, %f731;
|
|
abs.f32 %f294, %f292;
|
|
setp.lt.f32 %p72, %f294, 0f00800000;
|
|
mul.f32 %f732, %f294, 0f4B800000;
|
|
selp.f32 %f733, 0fC3170000, 0fC2FE0000, %p72;
|
|
selp.f32 %f734, %f732, %f294, %p72;
|
|
mov.b32 %r261, %f734;
|
|
and.b32 %r262, %r261, 8388607;
|
|
or.b32 %r263, %r262, 1065353216;
|
|
mov.b32 %f735, %r263;
|
|
shr.u32 %r264, %r261, 23;
|
|
cvt.rn.f32.u32 %f736, %r264;
|
|
add.f32 %f737, %f733, %f736;
|
|
setp.gt.f32 %p73, %f735, 0f3FB504F3;
|
|
mul.f32 %f738, %f735, 0f3F000000;
|
|
add.f32 %f739, %f737, 0f3F800000;
|
|
selp.f32 %f740, %f738, %f735, %p73;
|
|
selp.f32 %f741, %f739, %f737, %p73;
|
|
add.f32 %f742, %f740, 0fBF800000;
|
|
add.f32 %f728, %f740, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f727,%f728;
|
|
// inline asm
|
|
add.f32 %f743, %f742, %f742;
|
|
mul.f32 %f744, %f727, %f743;
|
|
mul.f32 %f745, %f744, %f744;
|
|
mov.f32 %f746, 0f3C4CAF63;
|
|
mov.f32 %f747, 0f3B18F0FE;
|
|
fma.rn.f32 %f748, %f747, %f745, %f746;
|
|
mov.f32 %f749, 0f3DAAAABD;
|
|
fma.rn.f32 %f750, %f748, %f745, %f749;
|
|
mul.rn.f32 %f751, %f750, %f745;
|
|
mul.rn.f32 %f752, %f751, %f744;
|
|
sub.f32 %f753, %f742, %f744;
|
|
neg.f32 %f754, %f744;
|
|
add.f32 %f755, %f753, %f753;
|
|
fma.rn.f32 %f756, %f754, %f742, %f755;
|
|
mul.rn.f32 %f757, %f727, %f756;
|
|
add.f32 %f758, %f752, %f744;
|
|
sub.f32 %f759, %f744, %f758;
|
|
add.f32 %f760, %f752, %f759;
|
|
add.f32 %f761, %f757, %f760;
|
|
add.f32 %f762, %f758, %f761;
|
|
sub.f32 %f763, %f758, %f762;
|
|
add.f32 %f764, %f761, %f763;
|
|
mov.f32 %f765, 0f3F317200;
|
|
mul.rn.f32 %f766, %f741, %f765;
|
|
mov.f32 %f767, 0f35BFBE8E;
|
|
mul.rn.f32 %f768, %f741, %f767;
|
|
add.f32 %f769, %f766, %f762;
|
|
sub.f32 %f770, %f766, %f769;
|
|
add.f32 %f771, %f762, %f770;
|
|
add.f32 %f772, %f764, %f771;
|
|
add.f32 %f773, %f768, %f772;
|
|
add.f32 %f774, %f769, %f773;
|
|
sub.f32 %f775, %f769, %f774;
|
|
add.f32 %f776, %f773, %f775;
|
|
mov.f32 %f777, 0f3EE8BA2E;
|
|
mul.rn.f32 %f778, %f777, %f774;
|
|
neg.f32 %f779, %f778;
|
|
fma.rn.f32 %f780, %f777, %f774, %f779;
|
|
fma.rn.f32 %f781, %f777, %f776, %f780;
|
|
mov.f32 %f782, 0f00000000;
|
|
fma.rn.f32 %f783, %f782, %f774, %f781;
|
|
add.rn.f32 %f784, %f778, %f783;
|
|
neg.f32 %f785, %f784;
|
|
add.rn.f32 %f786, %f778, %f785;
|
|
add.rn.f32 %f787, %f786, %f783;
|
|
mov.b32 %r265, %f784;
|
|
setp.eq.s32 %p74, %r265, 1118925336;
|
|
add.s32 %r266, %r265, -1;
|
|
mov.b32 %f788, %r266;
|
|
add.f32 %f789, %f787, 0f37000000;
|
|
selp.f32 %f790, %f788, %f784, %p74;
|
|
selp.f32 %f295, %f789, %f787, %p74;
|
|
mul.f32 %f791, %f790, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f792, %f791;
|
|
mov.f32 %f793, 0fBF317200;
|
|
fma.rn.f32 %f794, %f792, %f793, %f790;
|
|
mov.f32 %f795, 0fB5BFBE8E;
|
|
fma.rn.f32 %f796, %f792, %f795, %f794;
|
|
mul.f32 %f797, %f796, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f798, %f797;
|
|
add.f32 %f799, %f792, 0f00000000;
|
|
ex2.approx.f32 %f800, %f799;
|
|
mul.f32 %f801, %f798, %f800;
|
|
setp.lt.f32 %p75, %f790, 0fC2D20000;
|
|
selp.f32 %f802, 0f00000000, %f801, %p75;
|
|
setp.gt.f32 %p76, %f790, 0f42D20000;
|
|
selp.f32 %f1320, 0f7F800000, %f802, %p76;
|
|
setp.eq.f32 %p77, %f1320, 0f7F800000;
|
|
@%p77 bra BB0_83;
|
|
|
|
fma.rn.f32 %f1320, %f1320, %f295, %f1320;
|
|
|
|
BB0_83:
|
|
setp.lt.f32 %p78, %f292, 0f00000000;
|
|
setp.eq.f32 %p79, %f293, 0f3F800000;
|
|
and.pred %p3, %p78, %p79;
|
|
mov.b32 %r267, %f1320;
|
|
xor.b32 %r268, %r267, -2147483648;
|
|
mov.b32 %f803, %r268;
|
|
selp.f32 %f1322, %f803, %f1320, %p3;
|
|
setp.eq.f32 %p80, %f292, 0f00000000;
|
|
@%p80 bra BB0_86;
|
|
bra.uni BB0_84;
|
|
|
|
BB0_86:
|
|
add.f32 %f806, %f292, %f292;
|
|
selp.f32 %f1322, %f806, 0f00000000, %p79;
|
|
bra.uni BB0_87;
|
|
|
|
BB0_151:
|
|
mov.u64 %rd231, image_HDR;
|
|
cvta.global.u64 %rd226, %rd231;
|
|
mov.u32 %r354, 8;
|
|
mov.u64 %rd230, 0;
|
|
// inline asm
|
|
call (%rd225), _rt_buffer_get_64, (%rd226, %r103, %r354, %rd21, %rd22, %rd230, %rd230);
|
|
// inline asm
|
|
mov.f32 %f1133, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs105, %f1133;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs106, 0;
|
|
st.v4.u16 [%rd225], {%rs105, %rs105, %rs105, %rs106};
|
|
|
|
BB0_152:
|
|
ld.global.u32 %r355, [additive];
|
|
setp.eq.s32 %p147, %r355, 0;
|
|
@%p147 bra BB0_154;
|
|
|
|
mov.u64 %rd244, image_RNM0;
|
|
cvta.global.u64 %rd233, %rd244;
|
|
mov.u32 %r359, 8;
|
|
mov.u64 %rd243, 0;
|
|
// inline asm
|
|
call (%rd232), _rt_buffer_get_64, (%rd233, %r103, %r359, %rd21, %rd22, %rd243, %rd243);
|
|
// inline asm
|
|
ld.v4.u16 {%rs113, %rs114, %rs115, %rs116}, [%rd232];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1134, %rs113;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1135, %rs114;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1136, %rs115;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd238), _rt_buffer_get_64, (%rd233, %r103, %r359, %rd21, %rd22, %rd243, %rd243);
|
|
// inline asm
|
|
add.f32 %f1137, %f1134, 0f00000000;
|
|
add.f32 %f1138, %f1135, 0f00000000;
|
|
add.f32 %f1139, %f1136, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs112, %f1139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs111, %f1138;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs110, %f1137;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs117, 0;
|
|
st.v4.u16 [%rd238], {%rs110, %rs111, %rs112, %rs117};
|
|
bra.uni BB0_155;
|
|
|
|
BB0_154:
|
|
mov.u64 %rd251, image_RNM0;
|
|
cvta.global.u64 %rd246, %rd251;
|
|
mov.u32 %r361, 8;
|
|
mov.u64 %rd250, 0;
|
|
// inline asm
|
|
call (%rd245), _rt_buffer_get_64, (%rd246, %r103, %r361, %rd21, %rd22, %rd250, %rd250);
|
|
// inline asm
|
|
mov.f32 %f1140, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs118, %f1140;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs119, 0;
|
|
st.v4.u16 [%rd245], {%rs118, %rs118, %rs118, %rs119};
|
|
|
|
BB0_155:
|
|
ld.global.u32 %r362, [additive];
|
|
setp.eq.s32 %p148, %r362, 0;
|
|
@%p148 bra BB0_157;
|
|
|
|
mov.u64 %rd264, image_RNM1;
|
|
cvta.global.u64 %rd253, %rd264;
|
|
mov.u32 %r366, 8;
|
|
mov.u64 %rd263, 0;
|
|
// inline asm
|
|
call (%rd252), _rt_buffer_get_64, (%rd253, %r103, %r366, %rd21, %rd22, %rd263, %rd263);
|
|
// inline asm
|
|
ld.v4.u16 {%rs126, %rs127, %rs128, %rs129}, [%rd252];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1141, %rs126;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1142, %rs127;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1143, %rs128;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd258), _rt_buffer_get_64, (%rd253, %r103, %r366, %rd21, %rd22, %rd263, %rd263);
|
|
// inline asm
|
|
add.f32 %f1144, %f1141, 0f00000000;
|
|
add.f32 %f1145, %f1142, 0f00000000;
|
|
add.f32 %f1146, %f1143, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs125, %f1146;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs124, %f1145;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs123, %f1144;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs130, 0;
|
|
st.v4.u16 [%rd258], {%rs123, %rs124, %rs125, %rs130};
|
|
bra.uni BB0_158;
|
|
|
|
BB0_157:
|
|
mov.u64 %rd271, image_RNM1;
|
|
cvta.global.u64 %rd266, %rd271;
|
|
mov.u32 %r368, 8;
|
|
mov.u64 %rd270, 0;
|
|
// inline asm
|
|
call (%rd265), _rt_buffer_get_64, (%rd266, %r103, %r368, %rd21, %rd22, %rd270, %rd270);
|
|
// inline asm
|
|
mov.f32 %f1147, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs131, %f1147;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs132, 0;
|
|
st.v4.u16 [%rd265], {%rs131, %rs131, %rs131, %rs132};
|
|
|
|
BB0_158:
|
|
ld.global.u32 %r369, [additive];
|
|
setp.eq.s32 %p149, %r369, 0;
|
|
@%p149 bra BB0_160;
|
|
|
|
mov.u64 %rd284, image_RNM2;
|
|
cvta.global.u64 %rd273, %rd284;
|
|
mov.u32 %r373, 8;
|
|
mov.u64 %rd283, 0;
|
|
// inline asm
|
|
call (%rd272), _rt_buffer_get_64, (%rd273, %r103, %r373, %rd21, %rd22, %rd283, %rd283);
|
|
// inline asm
|
|
ld.v4.u16 {%rs139, %rs140, %rs141, %rs142}, [%rd272];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1148, %rs139;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1149, %rs140;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1150, %rs141;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd278), _rt_buffer_get_64, (%rd273, %r103, %r373, %rd21, %rd22, %rd283, %rd283);
|
|
// inline asm
|
|
add.f32 %f1151, %f1148, 0f00000000;
|
|
add.f32 %f1152, %f1149, 0f00000000;
|
|
add.f32 %f1153, %f1150, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs138, %f1153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs137, %f1152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs136, %f1151;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs143, 0;
|
|
st.v4.u16 [%rd278], {%rs136, %rs137, %rs138, %rs143};
|
|
bra.uni BB0_161;
|
|
|
|
BB0_160:
|
|
mov.u64 %rd291, image_RNM2;
|
|
cvta.global.u64 %rd286, %rd291;
|
|
mov.u32 %r375, 8;
|
|
mov.u64 %rd290, 0;
|
|
// inline asm
|
|
call (%rd285), _rt_buffer_get_64, (%rd286, %r103, %r375, %rd21, %rd22, %rd290, %rd290);
|
|
// inline asm
|
|
mov.f32 %f1154, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs144, %f1154;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs145, 0;
|
|
st.v4.u16 [%rd285], {%rs144, %rs144, %rs144, %rs145};
|
|
|
|
BB0_161:
|
|
ld.global.u32 %r376, [additive];
|
|
setp.eq.s32 %p150, %r376, 0;
|
|
@%p150 bra BB0_163;
|
|
|
|
mov.u64 %rd304, image_RNM3;
|
|
cvta.global.u64 %rd293, %rd304;
|
|
mov.u32 %r380, 8;
|
|
mov.u64 %rd303, 0;
|
|
// inline asm
|
|
call (%rd292), _rt_buffer_get_64, (%rd293, %r103, %r380, %rd21, %rd22, %rd303, %rd303);
|
|
// inline asm
|
|
ld.v4.u16 {%rs152, %rs153, %rs154, %rs155}, [%rd292];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1155, %rs152;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1156, %rs153;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1157, %rs154;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd298), _rt_buffer_get_64, (%rd293, %r103, %r380, %rd21, %rd22, %rd303, %rd303);
|
|
// inline asm
|
|
add.f32 %f1158, %f1155, 0f00000000;
|
|
add.f32 %f1159, %f1156, 0f00000000;
|
|
add.f32 %f1160, %f1157, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs151, %f1160;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs150, %f1159;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs149, %f1158;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs156, 0;
|
|
st.v4.u16 [%rd298], {%rs149, %rs150, %rs151, %rs156};
|
|
bra.uni BB0_164;
|
|
|
|
BB0_163:
|
|
mov.u64 %rd311, image_RNM3;
|
|
cvta.global.u64 %rd306, %rd311;
|
|
mov.u32 %r382, 8;
|
|
mov.u64 %rd310, 0;
|
|
// inline asm
|
|
call (%rd305), _rt_buffer_get_64, (%rd306, %r103, %r382, %rd21, %rd22, %rd310, %rd310);
|
|
// inline asm
|
|
mov.f32 %f1161, 0f00000000;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs157, %f1161;}
|
|
|
|
// inline asm
|
|
mov.u16 %rs158, 0;
|
|
st.v4.u16 [%rd305], {%rs157, %rs157, %rs157, %rs158};
|
|
bra.uni BB0_164;
|
|
|
|
BB0_84:
|
|
setp.geu.f32 %p81, %f292, 0f00000000;
|
|
@%p81 bra BB0_87;
|
|
|
|
cvt.rzi.f32.f32 %f805, %f777;
|
|
setp.neu.f32 %p82, %f805, 0f3EE8BA2E;
|
|
selp.f32 %f1322, 0f7FFFFFFF, %f1322, %p82;
|
|
|
|
BB0_87:
|
|
add.f32 %f807, %f294, 0f3EE8BA2E;
|
|
mov.b32 %r269, %f807;
|
|
setp.lt.s32 %p84, %r269, 2139095040;
|
|
@%p84 bra BB0_92;
|
|
|
|
setp.gtu.f32 %p85, %f294, 0f7F800000;
|
|
@%p85 bra BB0_91;
|
|
bra.uni BB0_89;
|
|
|
|
BB0_91:
|
|
add.f32 %f1322, %f292, 0f3EE8BA2E;
|
|
bra.uni BB0_92;
|
|
|
|
BB0_89:
|
|
setp.neu.f32 %p86, %f294, 0f7F800000;
|
|
@%p86 bra BB0_92;
|
|
|
|
selp.f32 %f1322, 0fFF800000, 0f7F800000, %p3;
|
|
|
|
BB0_92:
|
|
mul.f32 %f808, %f1322, 0f437F0000;
|
|
setp.eq.f32 %p87, %f292, 0f3F800000;
|
|
selp.f32 %f809, 0f437F0000, %f808, %p87;
|
|
cvt.rzi.u32.f32 %r270, %f809;
|
|
cvt.u16.u32 %rs15, %r270;
|
|
mov.u16 %rs16, 255;
|
|
st.v2.u8 [%rd84], {%rs15, %rs16};
|
|
ld.global.u32 %r416, [imageEnabled];
|
|
|
|
BB0_93:
|
|
ld.global.f32 %f810, [lightColor];
|
|
mul.f32 %f306, %f1316, %f810;
|
|
ld.global.f32 %f811, [lightColor+4];
|
|
mul.f32 %f307, %f1317, %f811;
|
|
ld.global.f32 %f812, [lightColor+8];
|
|
mul.f32 %f308, %f1318, %f812;
|
|
and.b32 %r271, %r416, 1;
|
|
setp.eq.b32 %p88, %r271, 1;
|
|
@!%p88 bra BB0_128;
|
|
bra.uni BB0_94;
|
|
|
|
BB0_94:
|
|
mov.f32 %f815, 0f3E666666;
|
|
cvt.rzi.f32.f32 %f816, %f815;
|
|
fma.rn.f32 %f817, %f816, 0fC0000000, 0f3EE66666;
|
|
abs.f32 %f309, %f817;
|
|
abs.f32 %f310, %f306;
|
|
setp.lt.f32 %p89, %f310, 0f00800000;
|
|
mul.f32 %f818, %f310, 0f4B800000;
|
|
selp.f32 %f819, 0fC3170000, 0fC2FE0000, %p89;
|
|
selp.f32 %f820, %f818, %f310, %p89;
|
|
mov.b32 %r272, %f820;
|
|
and.b32 %r273, %r272, 8388607;
|
|
or.b32 %r274, %r273, 1065353216;
|
|
mov.b32 %f821, %r274;
|
|
shr.u32 %r275, %r272, 23;
|
|
cvt.rn.f32.u32 %f822, %r275;
|
|
add.f32 %f823, %f819, %f822;
|
|
setp.gt.f32 %p90, %f821, 0f3FB504F3;
|
|
mul.f32 %f824, %f821, 0f3F000000;
|
|
add.f32 %f825, %f823, 0f3F800000;
|
|
selp.f32 %f826, %f824, %f821, %p90;
|
|
selp.f32 %f827, %f825, %f823, %p90;
|
|
add.f32 %f828, %f826, 0fBF800000;
|
|
add.f32 %f814, %f826, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f813,%f814;
|
|
// inline asm
|
|
add.f32 %f829, %f828, %f828;
|
|
mul.f32 %f830, %f813, %f829;
|
|
mul.f32 %f831, %f830, %f830;
|
|
mov.f32 %f832, 0f3C4CAF63;
|
|
mov.f32 %f833, 0f3B18F0FE;
|
|
fma.rn.f32 %f834, %f833, %f831, %f832;
|
|
mov.f32 %f835, 0f3DAAAABD;
|
|
fma.rn.f32 %f836, %f834, %f831, %f835;
|
|
mul.rn.f32 %f837, %f836, %f831;
|
|
mul.rn.f32 %f838, %f837, %f830;
|
|
sub.f32 %f839, %f828, %f830;
|
|
neg.f32 %f840, %f830;
|
|
add.f32 %f841, %f839, %f839;
|
|
fma.rn.f32 %f842, %f840, %f828, %f841;
|
|
mul.rn.f32 %f843, %f813, %f842;
|
|
add.f32 %f844, %f838, %f830;
|
|
sub.f32 %f845, %f830, %f844;
|
|
add.f32 %f846, %f838, %f845;
|
|
add.f32 %f847, %f843, %f846;
|
|
add.f32 %f848, %f844, %f847;
|
|
sub.f32 %f849, %f844, %f848;
|
|
add.f32 %f850, %f847, %f849;
|
|
mov.f32 %f851, 0f3F317200;
|
|
mul.rn.f32 %f852, %f827, %f851;
|
|
mov.f32 %f853, 0f35BFBE8E;
|
|
mul.rn.f32 %f854, %f827, %f853;
|
|
add.f32 %f855, %f852, %f848;
|
|
sub.f32 %f856, %f852, %f855;
|
|
add.f32 %f857, %f848, %f856;
|
|
add.f32 %f858, %f850, %f857;
|
|
add.f32 %f859, %f854, %f858;
|
|
add.f32 %f860, %f855, %f859;
|
|
sub.f32 %f861, %f855, %f860;
|
|
add.f32 %f862, %f859, %f861;
|
|
mov.f32 %f863, 0f3EE66666;
|
|
mul.rn.f32 %f864, %f863, %f860;
|
|
neg.f32 %f865, %f864;
|
|
fma.rn.f32 %f866, %f863, %f860, %f865;
|
|
fma.rn.f32 %f867, %f863, %f862, %f866;
|
|
mov.f32 %f868, 0f00000000;
|
|
fma.rn.f32 %f869, %f868, %f860, %f867;
|
|
add.rn.f32 %f870, %f864, %f869;
|
|
neg.f32 %f871, %f870;
|
|
add.rn.f32 %f872, %f864, %f871;
|
|
add.rn.f32 %f873, %f872, %f869;
|
|
mov.b32 %r276, %f870;
|
|
setp.eq.s32 %p91, %r276, 1118925336;
|
|
add.s32 %r277, %r276, -1;
|
|
mov.b32 %f874, %r277;
|
|
add.f32 %f875, %f873, 0f37000000;
|
|
selp.f32 %f876, %f874, %f870, %p91;
|
|
selp.f32 %f311, %f875, %f873, %p91;
|
|
mul.f32 %f877, %f876, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f878, %f877;
|
|
mov.f32 %f879, 0fBF317200;
|
|
fma.rn.f32 %f880, %f878, %f879, %f876;
|
|
mov.f32 %f881, 0fB5BFBE8E;
|
|
fma.rn.f32 %f882, %f878, %f881, %f880;
|
|
mul.f32 %f883, %f882, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f884, %f883;
|
|
add.f32 %f885, %f878, 0f00000000;
|
|
ex2.approx.f32 %f886, %f885;
|
|
mul.f32 %f887, %f884, %f886;
|
|
setp.lt.f32 %p92, %f876, 0fC2D20000;
|
|
selp.f32 %f888, 0f00000000, %f887, %p92;
|
|
setp.gt.f32 %p93, %f876, 0f42D20000;
|
|
selp.f32 %f1323, 0f7F800000, %f888, %p93;
|
|
setp.eq.f32 %p94, %f1323, 0f7F800000;
|
|
@%p94 bra BB0_96;
|
|
|
|
fma.rn.f32 %f1323, %f1323, %f311, %f1323;
|
|
|
|
BB0_96:
|
|
setp.lt.f32 %p95, %f306, 0f00000000;
|
|
setp.eq.f32 %p96, %f309, 0f3F800000;
|
|
and.pred %p4, %p95, %p96;
|
|
mov.b32 %r278, %f1323;
|
|
xor.b32 %r279, %r278, -2147483648;
|
|
mov.b32 %f889, %r279;
|
|
selp.f32 %f1325, %f889, %f1323, %p4;
|
|
setp.eq.f32 %p97, %f306, 0f00000000;
|
|
@%p97 bra BB0_99;
|
|
bra.uni BB0_97;
|
|
|
|
BB0_99:
|
|
add.f32 %f892, %f306, %f306;
|
|
selp.f32 %f1325, %f892, 0f00000000, %p96;
|
|
bra.uni BB0_100;
|
|
|
|
BB0_97:
|
|
setp.geu.f32 %p98, %f306, 0f00000000;
|
|
@%p98 bra BB0_100;
|
|
|
|
cvt.rzi.f32.f32 %f891, %f863;
|
|
setp.neu.f32 %p99, %f891, 0f3EE66666;
|
|
selp.f32 %f1325, 0f7FFFFFFF, %f1325, %p99;
|
|
|
|
BB0_100:
|
|
add.f32 %f893, %f310, 0f3EE66666;
|
|
mov.b32 %r280, %f893;
|
|
setp.lt.s32 %p101, %r280, 2139095040;
|
|
@%p101 bra BB0_105;
|
|
|
|
setp.gtu.f32 %p102, %f310, 0f7F800000;
|
|
@%p102 bra BB0_104;
|
|
bra.uni BB0_102;
|
|
|
|
BB0_104:
|
|
add.f32 %f1325, %f306, 0f3EE66666;
|
|
bra.uni BB0_105;
|
|
|
|
BB0_102:
|
|
setp.neu.f32 %p103, %f310, 0f7F800000;
|
|
@%p103 bra BB0_105;
|
|
|
|
selp.f32 %f1325, 0fFF800000, 0f7F800000, %p4;
|
|
|
|
BB0_105:
|
|
setp.eq.f32 %p104, %f306, 0f3F800000;
|
|
selp.f32 %f322, 0f3F800000, %f1325, %p104;
|
|
abs.f32 %f323, %f307;
|
|
setp.lt.f32 %p105, %f323, 0f00800000;
|
|
mul.f32 %f896, %f323, 0f4B800000;
|
|
selp.f32 %f897, 0fC3170000, 0fC2FE0000, %p105;
|
|
selp.f32 %f898, %f896, %f323, %p105;
|
|
mov.b32 %r281, %f898;
|
|
and.b32 %r282, %r281, 8388607;
|
|
or.b32 %r283, %r282, 1065353216;
|
|
mov.b32 %f899, %r283;
|
|
shr.u32 %r284, %r281, 23;
|
|
cvt.rn.f32.u32 %f900, %r284;
|
|
add.f32 %f901, %f897, %f900;
|
|
setp.gt.f32 %p106, %f899, 0f3FB504F3;
|
|
mul.f32 %f902, %f899, 0f3F000000;
|
|
add.f32 %f903, %f901, 0f3F800000;
|
|
selp.f32 %f904, %f902, %f899, %p106;
|
|
selp.f32 %f905, %f903, %f901, %p106;
|
|
add.f32 %f906, %f904, 0fBF800000;
|
|
add.f32 %f895, %f904, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f894,%f895;
|
|
// inline asm
|
|
add.f32 %f907, %f906, %f906;
|
|
mul.f32 %f908, %f894, %f907;
|
|
mul.f32 %f909, %f908, %f908;
|
|
fma.rn.f32 %f912, %f833, %f909, %f832;
|
|
fma.rn.f32 %f914, %f912, %f909, %f835;
|
|
mul.rn.f32 %f915, %f914, %f909;
|
|
mul.rn.f32 %f916, %f915, %f908;
|
|
sub.f32 %f917, %f906, %f908;
|
|
neg.f32 %f918, %f908;
|
|
add.f32 %f919, %f917, %f917;
|
|
fma.rn.f32 %f920, %f918, %f906, %f919;
|
|
mul.rn.f32 %f921, %f894, %f920;
|
|
add.f32 %f922, %f916, %f908;
|
|
sub.f32 %f923, %f908, %f922;
|
|
add.f32 %f924, %f916, %f923;
|
|
add.f32 %f925, %f921, %f924;
|
|
add.f32 %f926, %f922, %f925;
|
|
sub.f32 %f927, %f922, %f926;
|
|
add.f32 %f928, %f925, %f927;
|
|
mul.rn.f32 %f930, %f905, %f851;
|
|
mul.rn.f32 %f932, %f905, %f853;
|
|
add.f32 %f933, %f930, %f926;
|
|
sub.f32 %f934, %f930, %f933;
|
|
add.f32 %f935, %f926, %f934;
|
|
add.f32 %f936, %f928, %f935;
|
|
add.f32 %f937, %f932, %f936;
|
|
add.f32 %f938, %f933, %f937;
|
|
sub.f32 %f939, %f933, %f938;
|
|
add.f32 %f940, %f937, %f939;
|
|
mul.rn.f32 %f942, %f863, %f938;
|
|
neg.f32 %f943, %f942;
|
|
fma.rn.f32 %f944, %f863, %f938, %f943;
|
|
fma.rn.f32 %f945, %f863, %f940, %f944;
|
|
fma.rn.f32 %f947, %f868, %f938, %f945;
|
|
add.rn.f32 %f948, %f942, %f947;
|
|
neg.f32 %f949, %f948;
|
|
add.rn.f32 %f950, %f942, %f949;
|
|
add.rn.f32 %f951, %f950, %f947;
|
|
mov.b32 %r285, %f948;
|
|
setp.eq.s32 %p107, %r285, 1118925336;
|
|
add.s32 %r286, %r285, -1;
|
|
mov.b32 %f952, %r286;
|
|
add.f32 %f953, %f951, 0f37000000;
|
|
selp.f32 %f954, %f952, %f948, %p107;
|
|
selp.f32 %f324, %f953, %f951, %p107;
|
|
mul.f32 %f955, %f954, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f956, %f955;
|
|
fma.rn.f32 %f958, %f956, %f879, %f954;
|
|
fma.rn.f32 %f960, %f956, %f881, %f958;
|
|
mul.f32 %f961, %f960, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f962, %f961;
|
|
add.f32 %f963, %f956, 0f00000000;
|
|
ex2.approx.f32 %f964, %f963;
|
|
mul.f32 %f965, %f962, %f964;
|
|
setp.lt.f32 %p108, %f954, 0fC2D20000;
|
|
selp.f32 %f966, 0f00000000, %f965, %p108;
|
|
setp.gt.f32 %p109, %f954, 0f42D20000;
|
|
selp.f32 %f1326, 0f7F800000, %f966, %p109;
|
|
setp.eq.f32 %p110, %f1326, 0f7F800000;
|
|
@%p110 bra BB0_107;
|
|
|
|
fma.rn.f32 %f1326, %f1326, %f324, %f1326;
|
|
|
|
BB0_107:
|
|
setp.lt.f32 %p111, %f307, 0f00000000;
|
|
and.pred %p5, %p111, %p96;
|
|
mov.b32 %r287, %f1326;
|
|
xor.b32 %r288, %r287, -2147483648;
|
|
mov.b32 %f967, %r288;
|
|
selp.f32 %f1328, %f967, %f1326, %p5;
|
|
setp.eq.f32 %p113, %f307, 0f00000000;
|
|
@%p113 bra BB0_110;
|
|
bra.uni BB0_108;
|
|
|
|
BB0_110:
|
|
add.f32 %f970, %f307, %f307;
|
|
selp.f32 %f1328, %f970, 0f00000000, %p96;
|
|
bra.uni BB0_111;
|
|
|
|
BB0_108:
|
|
setp.geu.f32 %p114, %f307, 0f00000000;
|
|
@%p114 bra BB0_111;
|
|
|
|
cvt.rzi.f32.f32 %f969, %f863;
|
|
setp.neu.f32 %p115, %f969, 0f3EE66666;
|
|
selp.f32 %f1328, 0f7FFFFFFF, %f1328, %p115;
|
|
|
|
BB0_111:
|
|
add.f32 %f971, %f323, 0f3EE66666;
|
|
mov.b32 %r289, %f971;
|
|
setp.lt.s32 %p117, %r289, 2139095040;
|
|
@%p117 bra BB0_116;
|
|
|
|
setp.gtu.f32 %p118, %f323, 0f7F800000;
|
|
@%p118 bra BB0_115;
|
|
bra.uni BB0_113;
|
|
|
|
BB0_115:
|
|
add.f32 %f1328, %f307, 0f3EE66666;
|
|
bra.uni BB0_116;
|
|
|
|
BB0_113:
|
|
setp.neu.f32 %p119, %f323, 0f7F800000;
|
|
@%p119 bra BB0_116;
|
|
|
|
selp.f32 %f1328, 0fFF800000, 0f7F800000, %p5;
|
|
|
|
BB0_116:
|
|
setp.eq.f32 %p120, %f307, 0f3F800000;
|
|
selp.f32 %f335, 0f3F800000, %f1328, %p120;
|
|
abs.f32 %f336, %f308;
|
|
setp.lt.f32 %p121, %f336, 0f00800000;
|
|
mul.f32 %f974, %f336, 0f4B800000;
|
|
selp.f32 %f975, 0fC3170000, 0fC2FE0000, %p121;
|
|
selp.f32 %f976, %f974, %f336, %p121;
|
|
mov.b32 %r290, %f976;
|
|
and.b32 %r291, %r290, 8388607;
|
|
or.b32 %r292, %r291, 1065353216;
|
|
mov.b32 %f977, %r292;
|
|
shr.u32 %r293, %r290, 23;
|
|
cvt.rn.f32.u32 %f978, %r293;
|
|
add.f32 %f979, %f975, %f978;
|
|
setp.gt.f32 %p122, %f977, 0f3FB504F3;
|
|
mul.f32 %f980, %f977, 0f3F000000;
|
|
add.f32 %f981, %f979, 0f3F800000;
|
|
selp.f32 %f982, %f980, %f977, %p122;
|
|
selp.f32 %f983, %f981, %f979, %p122;
|
|
add.f32 %f984, %f982, 0fBF800000;
|
|
add.f32 %f973, %f982, 0f3F800000;
|
|
// inline asm
|
|
rcp.approx.ftz.f32 %f972,%f973;
|
|
// inline asm
|
|
add.f32 %f985, %f984, %f984;
|
|
mul.f32 %f986, %f972, %f985;
|
|
mul.f32 %f987, %f986, %f986;
|
|
fma.rn.f32 %f990, %f833, %f987, %f832;
|
|
fma.rn.f32 %f992, %f990, %f987, %f835;
|
|
mul.rn.f32 %f993, %f992, %f987;
|
|
mul.rn.f32 %f994, %f993, %f986;
|
|
sub.f32 %f995, %f984, %f986;
|
|
neg.f32 %f996, %f986;
|
|
add.f32 %f997, %f995, %f995;
|
|
fma.rn.f32 %f998, %f996, %f984, %f997;
|
|
mul.rn.f32 %f999, %f972, %f998;
|
|
add.f32 %f1000, %f994, %f986;
|
|
sub.f32 %f1001, %f986, %f1000;
|
|
add.f32 %f1002, %f994, %f1001;
|
|
add.f32 %f1003, %f999, %f1002;
|
|
add.f32 %f1004, %f1000, %f1003;
|
|
sub.f32 %f1005, %f1000, %f1004;
|
|
add.f32 %f1006, %f1003, %f1005;
|
|
mul.rn.f32 %f1008, %f983, %f851;
|
|
mul.rn.f32 %f1010, %f983, %f853;
|
|
add.f32 %f1011, %f1008, %f1004;
|
|
sub.f32 %f1012, %f1008, %f1011;
|
|
add.f32 %f1013, %f1004, %f1012;
|
|
add.f32 %f1014, %f1006, %f1013;
|
|
add.f32 %f1015, %f1010, %f1014;
|
|
add.f32 %f1016, %f1011, %f1015;
|
|
sub.f32 %f1017, %f1011, %f1016;
|
|
add.f32 %f1018, %f1015, %f1017;
|
|
mul.rn.f32 %f1020, %f863, %f1016;
|
|
neg.f32 %f1021, %f1020;
|
|
fma.rn.f32 %f1022, %f863, %f1016, %f1021;
|
|
fma.rn.f32 %f1023, %f863, %f1018, %f1022;
|
|
fma.rn.f32 %f1025, %f868, %f1016, %f1023;
|
|
add.rn.f32 %f1026, %f1020, %f1025;
|
|
neg.f32 %f1027, %f1026;
|
|
add.rn.f32 %f1028, %f1020, %f1027;
|
|
add.rn.f32 %f1029, %f1028, %f1025;
|
|
mov.b32 %r294, %f1026;
|
|
setp.eq.s32 %p123, %r294, 1118925336;
|
|
add.s32 %r295, %r294, -1;
|
|
mov.b32 %f1030, %r295;
|
|
add.f32 %f1031, %f1029, 0f37000000;
|
|
selp.f32 %f1032, %f1030, %f1026, %p123;
|
|
selp.f32 %f337, %f1031, %f1029, %p123;
|
|
mul.f32 %f1033, %f1032, 0f3FB8AA3B;
|
|
cvt.rzi.f32.f32 %f1034, %f1033;
|
|
fma.rn.f32 %f1036, %f1034, %f879, %f1032;
|
|
fma.rn.f32 %f1038, %f1034, %f881, %f1036;
|
|
mul.f32 %f1039, %f1038, 0f3FB8AA3B;
|
|
ex2.approx.ftz.f32 %f1040, %f1039;
|
|
add.f32 %f1041, %f1034, 0f00000000;
|
|
ex2.approx.f32 %f1042, %f1041;
|
|
mul.f32 %f1043, %f1040, %f1042;
|
|
setp.lt.f32 %p124, %f1032, 0fC2D20000;
|
|
selp.f32 %f1044, 0f00000000, %f1043, %p124;
|
|
setp.gt.f32 %p125, %f1032, 0f42D20000;
|
|
selp.f32 %f1329, 0f7F800000, %f1044, %p125;
|
|
setp.eq.f32 %p126, %f1329, 0f7F800000;
|
|
@%p126 bra BB0_118;
|
|
|
|
fma.rn.f32 %f1329, %f1329, %f337, %f1329;
|
|
|
|
BB0_118:
|
|
setp.lt.f32 %p127, %f308, 0f00000000;
|
|
and.pred %p6, %p127, %p96;
|
|
mov.b32 %r296, %f1329;
|
|
xor.b32 %r297, %r296, -2147483648;
|
|
mov.b32 %f1045, %r297;
|
|
selp.f32 %f1331, %f1045, %f1329, %p6;
|
|
setp.eq.f32 %p129, %f308, 0f00000000;
|
|
@%p129 bra BB0_121;
|
|
bra.uni BB0_119;
|
|
|
|
BB0_121:
|
|
add.f32 %f1048, %f308, %f308;
|
|
selp.f32 %f1331, %f1048, 0f00000000, %p96;
|
|
bra.uni BB0_122;
|
|
|
|
BB0_119:
|
|
setp.geu.f32 %p130, %f308, 0f00000000;
|
|
@%p130 bra BB0_122;
|
|
|
|
cvt.rzi.f32.f32 %f1047, %f863;
|
|
setp.neu.f32 %p131, %f1047, 0f3EE66666;
|
|
selp.f32 %f1331, 0f7FFFFFFF, %f1331, %p131;
|
|
|
|
BB0_122:
|
|
add.f32 %f1049, %f336, 0f3EE66666;
|
|
mov.b32 %r298, %f1049;
|
|
setp.lt.s32 %p133, %r298, 2139095040;
|
|
@%p133 bra BB0_127;
|
|
|
|
setp.gtu.f32 %p134, %f336, 0f7F800000;
|
|
@%p134 bra BB0_126;
|
|
bra.uni BB0_124;
|
|
|
|
BB0_126:
|
|
add.f32 %f1331, %f308, 0f3EE66666;
|
|
bra.uni BB0_127;
|
|
|
|
BB0_124:
|
|
setp.neu.f32 %p135, %f336, 0f7F800000;
|
|
@%p135 bra BB0_127;
|
|
|
|
selp.f32 %f1331, 0fFF800000, 0f7F800000, %p6;
|
|
|
|
BB0_127:
|
|
mov.u32 %r385, 4;
|
|
setp.eq.f32 %p136, %f308, 0f3F800000;
|
|
selp.f32 %f1050, 0f3F800000, %f1331, %p136;
|
|
cvt.u64.u32 %rd94, %r6;
|
|
cvt.u64.u32 %rd93, %r5;
|
|
mov.u64 %rd97, image;
|
|
cvta.global.u64 %rd92, %rd97;
|
|
// inline asm
|
|
call (%rd91), _rt_buffer_get_64, (%rd92, %r103, %r385, %rd93, %rd94, %rd28, %rd28);
|
|
// inline asm
|
|
cvt.sat.f32.f32 %f1051, %f1050;
|
|
mul.f32 %f1052, %f1051, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r301, %f1052;
|
|
cvt.sat.f32.f32 %f1053, %f335;
|
|
mul.f32 %f1054, %f1053, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r302, %f1054;
|
|
cvt.sat.f32.f32 %f1055, %f322;
|
|
mul.f32 %f1056, %f1055, 0f437FFD71;
|
|
cvt.rzi.u32.f32 %r303, %f1056;
|
|
cvt.u16.u32 %rs17, %r301;
|
|
cvt.u16.u32 %rs18, %r303;
|
|
cvt.u16.u32 %rs19, %r302;
|
|
mov.u16 %rs20, 255;
|
|
st.v4.u8 [%rd91], {%rs17, %rs19, %rs18, %rs20};
|
|
ld.global.u32 %r416, [imageEnabled];
|
|
|
|
BB0_128:
|
|
cvt.u64.u32 %rd19, %r5;
|
|
cvt.u64.u32 %rd20, %r6;
|
|
and.b32 %r304, %r416, 4;
|
|
setp.eq.s32 %p137, %r304, 0;
|
|
@%p137 bra BB0_132;
|
|
|
|
ld.global.u32 %r305, [additive];
|
|
setp.eq.s32 %p138, %r305, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs21, %f577;}
|
|
|
|
// inline asm
|
|
@%p138 bra BB0_131;
|
|
|
|
mov.u64 %rd110, image_HDR;
|
|
cvta.global.u64 %rd99, %rd110;
|
|
mov.u32 %r309, 8;
|
|
// inline asm
|
|
call (%rd98), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs28, %rs29, %rs30, %rs31}, [%rd98];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1058, %rs28;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1059, %rs29;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1060, %rs30;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd104), _rt_buffer_get_64, (%rd99, %r103, %r309, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1061, %f306, %f1058;
|
|
add.f32 %f1062, %f307, %f1059;
|
|
add.f32 %f1063, %f308, %f1060;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs27, %f1063;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs26, %f1062;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs25, %f1061;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd104], {%rs25, %rs26, %rs27, %rs21};
|
|
bra.uni BB0_132;
|
|
|
|
BB0_131:
|
|
mov.u64 %rd117, image_HDR;
|
|
cvta.global.u64 %rd112, %rd117;
|
|
mov.u32 %r311, 8;
|
|
// inline asm
|
|
call (%rd111), _rt_buffer_get_64, (%rd112, %r103, %r311, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs34, %f308;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs33, %f307;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs32, %f306;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd111], {%rs32, %rs33, %rs34, %rs21};
|
|
|
|
BB0_132:
|
|
selp.f32 %f1068, 0f3F000000, 0f3E800000, %p151;
|
|
mul.f32 %f348, %f1068, %f1313;
|
|
mul.f32 %f349, %f1068, %f1314;
|
|
mul.f32 %f350, %f1068, %f1315;
|
|
mul.f32 %f351, %f1068, %f1310;
|
|
mul.f32 %f352, %f1068, %f1311;
|
|
mul.f32 %f353, %f1068, %f1312;
|
|
mul.f32 %f354, %f1068, %f1307;
|
|
mul.f32 %f355, %f1068, %f1308;
|
|
mul.f32 %f356, %f1068, %f1309;
|
|
mul.f32 %f357, %f1068, %f1304;
|
|
mul.f32 %f358, %f1068, %f1305;
|
|
mul.f32 %f359, %f1068, %f1306;
|
|
ld.global.f32 %f1069, [lightColor];
|
|
mul.f32 %f360, %f348, %f1069;
|
|
ld.global.f32 %f1070, [lightColor+4];
|
|
mul.f32 %f361, %f349, %f1070;
|
|
ld.global.f32 %f1071, [lightColor+8];
|
|
mul.f32 %f362, %f350, %f1071;
|
|
ld.global.u32 %r312, [additive];
|
|
setp.eq.s32 %p139, %r312, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs35, %f577;}
|
|
|
|
// inline asm
|
|
@%p139 bra BB0_134;
|
|
|
|
mov.u64 %rd130, image_RNM0;
|
|
cvta.global.u64 %rd119, %rd130;
|
|
mov.u32 %r316, 8;
|
|
// inline asm
|
|
call (%rd118), _rt_buffer_get_64, (%rd119, %r103, %r316, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs42, %rs43, %rs44, %rs45}, [%rd118];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1072, %rs42;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1073, %rs43;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1074, %rs44;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd124), _rt_buffer_get_64, (%rd119, %r103, %r316, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1075, %f360, %f1072;
|
|
add.f32 %f1076, %f361, %f1073;
|
|
add.f32 %f1077, %f362, %f1074;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs41, %f1077;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs40, %f1076;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs39, %f1075;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd124], {%rs39, %rs40, %rs41, %rs35};
|
|
bra.uni BB0_135;
|
|
|
|
BB0_134:
|
|
mov.u64 %rd137, image_RNM0;
|
|
cvta.global.u64 %rd132, %rd137;
|
|
mov.u32 %r318, 8;
|
|
// inline asm
|
|
call (%rd131), _rt_buffer_get_64, (%rd132, %r103, %r318, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs48, %f362;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs47, %f361;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs46, %f360;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd131], {%rs46, %rs47, %rs48, %rs35};
|
|
|
|
BB0_135:
|
|
mov.f32 %f1082, 0f34000000;
|
|
max.f32 %f1083, %f348, %f1082;
|
|
div.rn.f32 %f1084, %f351, %f1083;
|
|
max.f32 %f1085, %f349, %f1082;
|
|
div.rn.f32 %f1086, %f352, %f1085;
|
|
max.f32 %f1087, %f350, %f1082;
|
|
div.rn.f32 %f1088, %f353, %f1087;
|
|
fma.rn.f32 %f363, %f1084, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f364, %f1086, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f365, %f1088, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1089, %f354, %f1083;
|
|
div.rn.f32 %f1090, %f355, %f1085;
|
|
div.rn.f32 %f1091, %f356, %f1087;
|
|
fma.rn.f32 %f366, %f1089, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f367, %f1090, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f368, %f1091, 0f3F000000, 0f3F000000;
|
|
div.rn.f32 %f1092, %f357, %f1083;
|
|
div.rn.f32 %f1093, %f358, %f1085;
|
|
div.rn.f32 %f1094, %f359, %f1087;
|
|
fma.rn.f32 %f369, %f1092, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f370, %f1093, 0f3F000000, 0f3F000000;
|
|
fma.rn.f32 %f371, %f1094, 0f3F000000, 0f3F000000;
|
|
ld.global.u32 %r319, [additive];
|
|
setp.eq.s32 %p140, %r319, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs49, %f577;}
|
|
|
|
// inline asm
|
|
@%p140 bra BB0_137;
|
|
|
|
mov.u64 %rd150, image_RNM1;
|
|
cvta.global.u64 %rd139, %rd150;
|
|
mov.u32 %r323, 8;
|
|
// inline asm
|
|
call (%rd138), _rt_buffer_get_64, (%rd139, %r103, %r323, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs56, %rs57, %rs58, %rs59}, [%rd138];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1095, %rs56;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1096, %rs57;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1097, %rs58;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd144), _rt_buffer_get_64, (%rd139, %r103, %r323, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1098, %f363, %f1095;
|
|
add.f32 %f1099, %f364, %f1096;
|
|
add.f32 %f1100, %f365, %f1097;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs55, %f1100;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs54, %f1099;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs53, %f1098;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd144], {%rs53, %rs54, %rs55, %rs49};
|
|
bra.uni BB0_138;
|
|
|
|
BB0_137:
|
|
mov.u64 %rd157, image_RNM1;
|
|
cvta.global.u64 %rd152, %rd157;
|
|
mov.u32 %r325, 8;
|
|
// inline asm
|
|
call (%rd151), _rt_buffer_get_64, (%rd152, %r103, %r325, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs62, %f365;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs61, %f364;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs60, %f363;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd151], {%rs60, %rs61, %rs62, %rs49};
|
|
|
|
BB0_138:
|
|
ld.global.u32 %r326, [additive];
|
|
setp.eq.s32 %p141, %r326, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs63, %f577;}
|
|
|
|
// inline asm
|
|
@%p141 bra BB0_140;
|
|
|
|
mov.u64 %rd170, image_RNM2;
|
|
cvta.global.u64 %rd159, %rd170;
|
|
mov.u32 %r330, 8;
|
|
// inline asm
|
|
call (%rd158), _rt_buffer_get_64, (%rd159, %r103, %r330, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs70, %rs71, %rs72, %rs73}, [%rd158];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1105, %rs70;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1106, %rs71;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1107, %rs72;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd164), _rt_buffer_get_64, (%rd159, %r103, %r330, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1108, %f366, %f1105;
|
|
add.f32 %f1109, %f367, %f1106;
|
|
add.f32 %f1110, %f368, %f1107;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs69, %f1110;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs68, %f1109;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs67, %f1108;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd164], {%rs67, %rs68, %rs69, %rs63};
|
|
bra.uni BB0_141;
|
|
|
|
BB0_140:
|
|
mov.u64 %rd177, image_RNM2;
|
|
cvta.global.u64 %rd172, %rd177;
|
|
mov.u32 %r332, 8;
|
|
// inline asm
|
|
call (%rd171), _rt_buffer_get_64, (%rd172, %r103, %r332, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs76, %f368;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs75, %f367;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs74, %f366;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd171], {%rs74, %rs75, %rs76, %rs63};
|
|
|
|
BB0_141:
|
|
ld.global.u32 %r333, [additive];
|
|
setp.eq.s32 %p142, %r333, 0;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs77, %f577;}
|
|
|
|
// inline asm
|
|
@%p142 bra BB0_143;
|
|
|
|
mov.u64 %rd190, image_RNM3;
|
|
cvta.global.u64 %rd179, %rd190;
|
|
mov.u32 %r337, 8;
|
|
// inline asm
|
|
call (%rd178), _rt_buffer_get_64, (%rd179, %r103, %r337, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
ld.v4.u16 {%rs84, %rs85, %rs86, %rs87}, [%rd178];
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1115, %rs84;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1116, %rs85;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.f32.f16 %f1117, %rs86;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
call (%rd184), _rt_buffer_get_64, (%rd179, %r103, %r337, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
add.f32 %f1118, %f369, %f1115;
|
|
add.f32 %f1119, %f370, %f1116;
|
|
add.f32 %f1120, %f371, %f1117;
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs83, %f1120;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs82, %f1119;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs81, %f1118;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd184], {%rs81, %rs82, %rs83, %rs77};
|
|
bra.uni BB0_164;
|
|
|
|
BB0_143:
|
|
mov.u64 %rd197, image_RNM3;
|
|
cvta.global.u64 %rd192, %rd197;
|
|
mov.u32 %r339, 8;
|
|
// inline asm
|
|
call (%rd191), _rt_buffer_get_64, (%rd192, %r103, %r339, %rd19, %rd20, %rd28, %rd28);
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs90, %f371;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs89, %f370;}
|
|
|
|
// inline asm
|
|
// inline asm
|
|
{ cvt.rn.f16.f32 %rs88, %f369;}
|
|
|
|
// inline asm
|
|
st.v4.u16 [%rd191], {%rs88, %rs89, %rs90, %rs77};
|
|
|
|
BB0_164:
|
|
ret;
|
|
}
|
|
|
|
|