A   d&            N'z>Dec  3 202420:16:52HOST64sm_86//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-31833905
// Cuda compilation tools, release 11.8, V11.8.89
// Based on NVVM 7.0.1
//

.version 7.8
.target sm_50
.address_size 64

	// .globl	ShaderKernel_fxEmboss
.const .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135, 22, 153, 62, 162, 69, 22, 63, 213, 120, 233, 61, 33, 201, 44, 190, 111, 155, 169, 190, 0, 0, 0, 63, 0, 0, 0, 63, 70, 94, 214, 190, 232, 134, 166, 189};
.const .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0, 0, 128, 63, 0, 0, 0, 0, 188, 116, 179, 63, 0, 0, 128, 63, 152, 50, 176, 190, 158, 209, 54, 191, 0, 0, 128, 63, 229, 208, 226, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70, 246, 130, 66, 145, 141, 0, 67, 94, 186, 199, 65, 33, 48, 23, 194, 240, 103, 148, 194, 0, 0, 224, 66, 0, 0, 224, 66, 111, 146, 187, 194, 70, 182, 145, 193};
.const .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37, 160, 149, 59, 0, 0, 0, 0, 182, 23, 205, 59, 37, 160, 149, 59, 40, 15, 201, 186, 156, 239, 80, 187, 37, 160, 149, 59, 236, 155, 1, 60, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219, 121, 131, 62, 152, 14, 1, 63, 18, 131, 200, 61, 174, 199, 23, 190, 238, 252, 148, 190, 197, 224, 224, 62, 197, 224, 224, 62, 217, 78, 188, 190, 174, 71, 146, 189};
.const .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127, 10, 149, 63, 0, 0, 0, 0, 160, 74, 204, 63, 127, 10, 149, 63, 254, 148, 200, 190, 184, 30, 80, 191, 127, 10, 149, 63, 78, 26, 1, 64, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_601YCbCrFullRange[36] = {135, 22, 153, 62, 162, 69, 22, 63, 213, 120, 233, 61, 166, 27, 44, 190, 39, 241, 168, 190, 250, 254, 254, 62, 250, 254, 254, 62, 43, 135, 213, 190, 59, 223, 165, 189};
.const .align 4 .b8 k601YCbCrFullRange_To_RGB8u[36] = {0, 0, 128, 63, 0, 0, 0, 0, 72, 193, 178, 63, 0, 0, 128, 63, 143, 130, 175, 190, 225, 26, 54, 191, 0, 0, 128, 63, 20, 238, 225, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_601YCbCrFullRange[36] = {113, 125, 152, 66, 92, 175, 21, 67, 92, 143, 232, 65, 158, 111, 43, 194, 49, 72, 168, 194, 0, 0, 254, 66, 0, 0, 254, 66, 170, 177, 212, 194, 88, 57, 165, 193};
.const .align 4 .b8 k601YCbCrFullRange_To_RGB32f[36] = {129, 128, 128, 59, 0, 0, 0, 0, 189, 116, 179, 59, 129, 128, 128, 59, 194, 50, 176, 186, 179, 209, 54, 187, 129, 128, 128, 59, 229, 208, 226, 59, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_709YPbPr[36] = {208, 179, 89, 62, 89, 23, 55, 63, 152, 221, 147, 61, 186, 164, 234, 189, 210, 86, 197, 190, 0, 0, 0, 63, 0, 0, 0, 63, 190, 134, 232, 190, 16, 202, 59, 189};
.const .align 4 .b8 k709YPbPr_To_RGB32f[36] = {0, 0, 128, 63, 0, 0, 0, 0, 12, 147, 201, 63, 0, 0, 128, 63, 221, 209, 63, 190, 243, 173, 239, 190, 0, 0, 128, 63, 77, 132, 237, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_709YCbCr[36] = {106, 60, 58, 66, 6, 161, 28, 67, 244, 253, 124, 65, 223, 79, 205, 193, 8, 172, 172, 194, 0, 0, 224, 66, 0, 0, 224, 66, 195, 117, 203, 194, 236, 81, 36, 193};
.const .align 4 .b8 k709YCbCr_To_RGB32f[36] = {37, 160, 149, 59, 0, 0, 0, 0, 239, 94, 230, 59, 37, 160, 149, 59, 33, 57, 91, 186, 178, 245, 8, 187, 37, 160, 149, 59, 82, 185, 7, 60, 0, 0, 0, 0};
.const .align 4 .b8 k709YCbCrFullRange_To_RGB32f[36] = {131, 128, 128, 59, 0, 0, 0, 0, 28, 147, 201, 59, 131, 128, 128, 59, 61, 210, 63, 186, 248, 173, 239, 186, 131, 128, 128, 59, 82, 132, 237, 59, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_709YCbCr[36] = {207, 247, 58, 62, 53, 62, 29, 63, 231, 251, 125, 61, 184, 30, 206, 189, 23, 89, 173, 190, 197, 224, 224, 62, 197, 224, 224, 62, 12, 66, 204, 190, 195, 245, 36, 189};
.const .align 4 .b8 k709YCbCr_To_RGB8u[36] = {127, 10, 149, 63, 0, 0, 0, 0, 147, 120, 229, 63, 127, 10, 149, 63, 53, 94, 90, 190, 205, 108, 8, 191, 127, 10, 149, 63, 154, 49, 7, 64, 0, 0, 0, 0};
.const .align 4 .b8 k709YCbCr_To_601YCbCr[36] = {0, 0, 128, 63, 23, 100, 203, 61, 1, 77, 68, 62, 0, 0, 0, 0, 18, 103, 125, 63, 10, 158, 226, 189, 0, 0, 0, 0, 61, 98, 148, 189, 249, 191, 123, 63};
.const .align 4 .b8 k601YCbCr_To_709YCbCr[36] = {0, 0, 128, 63, 122, 165, 236, 189, 179, 237, 84, 190, 0, 0, 0, 0, 204, 98, 130, 63, 216, 188, 234, 61, 0, 0, 0, 0, 74, 179, 153, 61, 234, 61, 131, 63};
.const .align 4 .b8 kZeroMatrix[36];
.const .align 4 .b8 kYCbCrOffset[12] = {0, 0, 128, 65, 0, 0, 0, 67, 0, 0, 0, 67};
.const .align 4 .b8 kYCbCrFullRangeOffset[12] = {0, 0, 0, 0, 0, 0, 0, 67, 0, 0, 0, 67};
.const .align 4 .f32 PQ_m1 = 0f3E232000;
.const .align 4 .f32 PQ_m1Inv = 0f40C8E06B;
.const .align 4 .f32 PQ_m2 = 0f429DB000;
.const .align 4 .f32 PQ_m2Inv = 0f3C4FCDAC;
.const .align 4 .f32 PQ_c1 = 0f3F560000;
.const .align 4 .f32 PQ_c2 = 0f4196D000;
.const .align 4 .f32 PQ_c3 = 0f41958000;
.const .align 4 .f32 Gamma1886 = 0f4019999A;
.const .align 4 .f32 PQ_Lpeak = 0f461C4000;
.const .align 4 .f32 PQ_a = 0f3F8CAC08;
.const .align 4 .f32 PQ_b = 0f426E1556;
.const .align 4 .f32 PQ_c = 0f39B033E5;
.const .align 4 .f32 PQ_s = 0f4385EB85;
.const .align 4 .f32 PQ_g = 0f3EE66666;
.const .align 4 .f32 scaleFD = 0f42C80000;
.const .align 4 .f32 HLG_a = 0f3E371FF0;
.const .align 4 .f32 HLG_b = 0f3E91C020;
.const .align 4 .f32 HLG_c = 0f3F0F564F;
.const .align 4 .f32 HLG_inva = 0f40B2F029;
.const .align 4 .f32 HLG_alpha = 0f41200000;
.const .align 4 .f32 HLG_invAlpha = 0f3DCCCCCD;
.const .align 4 .f32 HLG_gamma = 0f3F99999A;
.const .align 4 .f32 HLG_gammaM1 = 0f3E4CCCCD;
.const .align 4 .f32 HLG_gammaM1Dgamma = 0f3E2AAAAB;
.const .align 4 .f32 HLG_YR = 0f3E86809D;
.const .align 4 .f32 HLG_YG = 0f3F2D9168;
.const .align 4 .f32 HLG_YB = 0f3D72E48F;
.const .align 4 .f32 HLG_Lpeak = 0f447A0000;
.const .align 4 .f32 AppleLog_R0 = 0fBD670F18;
.const .align 4 .f32 AppleLog_Rt = 0f3C23D70A;
.const .align 4 .f32 AppleLog_c = 0f423D2601;
.const .align 4 .f32 AppleLog_beta = 0f3C1DF346;
.const .align 4 .f32 AppleLog_gamma = 0f3DAF1D23;
.const .align 4 .f32 AppleLog_delta = 0f3F3180A9;
.global .texref texture0_RECT;
// _ZZ30ShaderKernel_fxEmboss_DelegateP6float4PKS_Pvi17DevicePixelFormatii5uint2E7p_local has been demoted

.visible .entry ShaderKernel_fxEmboss(
	.param .u64 ShaderKernel_fxEmboss_param_0,
	.param .u64 ShaderKernel_fxEmboss_param_1,
	.param .u64 ShaderKernel_fxEmboss_param_2,
	.param .u32 ShaderKernel_fxEmboss_param_3,
	.param .u32 ShaderKernel_fxEmboss_param_4,
	.param .u32 ShaderKernel_fxEmboss_param_5,
	.param .u32 ShaderKernel_fxEmboss_param_6
)
{
	.reg .pred 	%p<7>;
	.reg .b16 	%rs<5>;
	.reg .f32 	%f<48>;
	.reg .b32 	%r<23>;
	.reg .b64 	%rd<11>;
	// demoted variable
	.shared .align 16 .b8 _ZZ30ShaderKernel_fxEmboss_DelegateP6float4PKS_Pvi17DevicePixelFormatii5uint2E7p_local[16];

	ld.param.u64 	%rd4, [ShaderKernel_fxEmboss_param_0];
	ld.param.u64 	%rd3, [ShaderKernel_fxEmboss_param_1];
	ld.param.u32 	%r4, [ShaderKernel_fxEmboss_param_3];
	ld.param.u32 	%r5, [ShaderKernel_fxEmboss_param_4];
	ld.param.u32 	%r6, [ShaderKernel_fxEmboss_param_5];
	ld.param.u32 	%r7, [ShaderKernel_fxEmboss_param_6];
	cvta.to.global.u64 	%rd1, %rd4;
	mov.u32 	%r1, %ntid.x;
	mov.u32 	%r8, %ctaid.x;
	mov.u32 	%r9, %tid.x;
	mad.lo.s32 	%r2, %r8, %r1, %r9;
	mov.u32 	%r10, %ntid.y;
	mov.u32 	%r11, %ctaid.y;
	mov.u32 	%r12, %tid.y;
	mad.lo.s32 	%r3, %r11, %r10, %r12;
	or.b32  	%r13, %r12, %r9;
	setp.ne.s32 	%p1, %r13, 0;
	@%p1 bra 	$L__BB0_3;

	setp.eq.s32 	%p2, %r1, 0;
	@%p2 bra 	$L__BB0_9;

	cvta.to.global.u64 	%rd5, %rd3;
	ld.global.v4.u32 	{%r14, %r15, %r16, %r17}, [%rd5];
	st.shared.v4.u32 	[_ZZ30ShaderKernel_fxEmboss_DelegateP6float4PKS_Pvi17DevicePixelFormatii5uint2E7p_local], {%r14, %r15, %r16, %r17};

$L__BB0_3:
	bar.sync 	0;
	setp.ge.s32 	%p3, %r3, %r7;
	setp.ge.s32 	%p4, %r2, %r6;
	or.pred  	%p5, %p4, %p3;
	@%p5 bra 	$L__BB0_7;

	cvt.rn.f32.u32 	%f3, %r2;
	add.ftz.f32 	%f4, %f3, 0f3F000000;
	cvt.rn.f32.u32 	%f5, %r3;
	add.ftz.f32 	%f6, %f5, 0f3F000000;
	add.ftz.f32 	%f7, %f4, 0f00000000;
	add.ftz.f32 	%f8, %f6, 0fBF800000;
	add.ftz.f32 	%f9, %f4, 0fBF800000;
	add.ftz.f32 	%f10, %f6, 0f00000000;
	tex.2d.v4.f32.f32 	{%f11, %f12, %f13, %f1}, [texture0_RECT, {%f4, %f6}];
	mul.ftz.f32 	%f14, %f13, 0f40400000;
	mul.ftz.f32 	%f15, %f12, 0f40400000;
	mul.ftz.f32 	%f16, %f11, 0f40400000;
	tex.2d.v4.f32.f32 	{%f17, %f18, %f19, %f20}, [texture0_RECT, {%f7, %f8}];
	mul.ftz.f32 	%f21, %f19, 0f3FC00000;
	sub.ftz.f32 	%f22, %f21, %f14;
	mul.ftz.f32 	%f23, %f18, 0f3FC00000;
	sub.ftz.f32 	%f24, %f23, %f15;
	mul.ftz.f32 	%f25, %f17, 0f3FC00000;
	sub.ftz.f32 	%f26, %f25, %f16;
	tex.2d.v4.f32.f32 	{%f27, %f28, %f29, %f30}, [texture0_RECT, {%f9, %f10}];
	fma.rn.ftz.f32 	%f31, %f29, 0f3FC00000, %f22;
	fma.rn.ftz.f32 	%f32, %f28, 0f3FC00000, %f24;
	fma.rn.ftz.f32 	%f33, %f27, 0f3FC00000, %f26;
	ld.shared.v4.f32 	{%f34, %f35, %f36, %f37}, [_ZZ30ShaderKernel_fxEmboss_DelegateP6float4PKS_Pvi17DevicePixelFormatii5uint2E7p_local];
	mul.ftz.f32 	%f41, %f35, %f32;
	fma.rn.ftz.f32 	%f42, %f34, %f31, %f41;
	fma.rn.ftz.f32 	%f43, %f36, %f33, %f42;
	add.ftz.f32 	%f2, %f43, 0f3EAAAAAB;
	mad.lo.s32 	%r22, %r3, %r4, %r2;
	cvt.s64.s32 	%rd2, %r22;
	setp.eq.s32 	%p6, %r5, 0;
	@%p6 bra 	$L__BB0_6;

	shl.b64 	%rd7, %rd2, 4;
	add.s64 	%rd8, %rd1, %rd7;
	st.global.v4.f32 	[%rd8], {%f2, %f2, %f2, %f1};
	bra.uni 	$L__BB0_7;

$L__BB0_9:
	bra.uni 	$L__BB0_9;

$L__BB0_6:
	// begin inline asm
	{  cvt.rn.f16.f32 %rs1, %f2;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs2, %f2;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs3, %f2;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs4, %f1;}

	// end inline asm
	shl.b64 	%rd9, %rd2, 3;
	add.s64 	%rd10, %rd1, %rd9;
	st.global.u16 	[%rd10], %rs1;
	st.global.u16 	[%rd10+2], %rs2;
	st.global.u16 	[%rd10+4], %rs3;
	st.global.u16 	[%rd10+6], %rs4;

$L__BB0_7:
	ret;

}

  ELF3                                 V2 @ 8  @    .shstrtab .strtab .symtab .symtab_shndx .nv.info .text.ShaderKernel_fxEmboss .nv.info.ShaderKernel_fxEmboss .nv.shared.ShaderKernel_fxEmboss .nv.constant3 .nv.constant0.ShaderKernel_fxEmboss .rel.nv.constant0.ShaderKernel_fxEmboss .debug_frame .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action  .shstrtab .strtab .symtab .symtab_shndx .nv.info .text.ShaderKernel_fxEmboss .nv.info.ShaderKernel_fxEmboss .nv.shared.ShaderKernel_fxEmboss .nv.constant3 kRGB32f_To_601YPbPr k601YPbPr_To_RGB32f kRGB32f_To_601YCbCr k601YCbCr_To_RGB32f kRGB8u_To_601YCbCr k601YCbCr_To_RGB8u kRGB8u_To_601YCbCrFullRange k601YCbCrFullRange_To_RGB8u kRGB32f_To_601YCbCrFullRange k601YCbCrFullRange_To_RGB32f kRGB32f_To_709YPbPr k709YPbPr_To_RGB32f kRGB32f_To_709YCbCr k709YCbCr_To_RGB32f k709YCbCrFullRange_To_RGB32f kRGB8u_To_709YCbCr k709YCbCr_To_RGB8u k709YCbCr_To_601YCbCr k601YCbCr_To_709YCbCr kZeroMatrix kYCbCrOffset kYCbCrFullRangeOffset PQ_m1 PQ_m1Inv PQ_m2 PQ_m2Inv PQ_c1 PQ_c2 PQ_c3 Gamma1886 PQ_Lpeak PQ_a PQ_b PQ_c PQ_s PQ_g scaleFD HLG_a HLG_b HLG_c HLG_inva HLG_alpha HLG_invAlpha HLG_gamma HLG_gammaM1 HLG_gammaM1Dgamma HLG_YR HLG_YG HLG_YB HLG_Lpeak AppleLog_R0 AppleLog_Rt AppleLog_c AppleLog_beta AppleLog_gamma AppleLog_delta .rel.nv.constant0.ShaderKernel_fxEmboss .nv.constant0.ShaderKernel_fxEmboss .debug_frame .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action ShaderKernel_fxEmboss texture0_RECT                             2                     m                                                       $            $       $            H       $            l       $                   $                   $                  $       .           $       J           $       g    D      $           h      $                 $                 $                 $                 $                 $           @      $           d      $       -          $       C          $       O                 \                 r                 x                                                                                                                                                                                                                                                    $                 (                 ,                 0                 4                 8                 <                 @                 D             0    H             7    L             >    P             E    T             O    X             [    \             g    `             r    d                 h                 l                                                                         6                    E                  [                     $        |( ((   4                                 $   (           / @       @        @       7    5  
 <   `( (       $                                       !         !           !    D      @    0                           s          % 6         A   D          @   >E?x=!,o   ?   ?F^־膦  ?    t?  ?26  ??    FB C^A!0g  B  BoF%;    ;%;(ɺP%;<    y>?=>>NG
?    J?
?ȾP
?N@    >E?x=,'>>+վ;ߥ  ?    H?  ?6  ??    q}B\C\Ao+1H  B  BX9;    t;;26;;    гY>Y7?ݓ=Vž   ?   ?;  ?    ?  ??  ?M?    j<:BC|AO  B  BuQ$%;    ^;%;!9[%;R<    ;    ;;=?ﺃ;R;    :>5>?}=νY>>B̾$
?    x?
?5^Zl
?1@      ?d=MD>    g}?
    =b{?  ?z콳T    b?ؼ=    J==?                                      A   C   C       C   C  #>k@ BO<  V? ЖA A@ @F?VnB39Cff>  B7> >OV?)@   A=?L>*>>h-?r=  zDg
#<&=BF<#=1?                                                                                                                                                                                                                                                                                                                                                                                                                                                                                z  
       y      !   " z  F   
    y      "   ( y       %   h y      &    r     $z       / $z    O G	  `       z     pR  G        z  Z       z  [       y      s        A Hy      {           z  a  pb  z   `  pfp   M	          Er          Er         x         !t   ?     !t   ?     !r
        !t       !t       !r	        `{b 0 F `{

b 0 B `{b 0 B y        " z  _  pR  $z  ^     x        x  @@  A   x  @@  A    x  @@  A   #x  ?   #x  ?   #x
  ?   r         #x	  ?    #x  ?    #x  ?     r     A   #r       #r       
  X     !t>     
  Y  $   r	        r
         	      M	          z  X   >r        z  Y     x v       xv      y      x2v       y      y     y     My          Gy    Gy    y            y            y            y            y            y            y            y            y            y                                                                                                  @       C                                                         i                                                         0         @                                              p                              )      p                      $                             N      p@                                                     p                l                                   4    p                                                      	   @                                                   	   @                                                                           p                                   B               0                                  2                                     @                 m      C                                                                                                                                @
      @
                                                                                                                 