A   k=      (#      r2Dec  6 202318:11:30HOST64sm_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	ColorHLSKernel
.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;

.visible .entry ColorHLSKernel(
	.param .u64 ColorHLSKernel_param_0,
	.param .u64 ColorHLSKernel_param_1,
	.param .u32 ColorHLSKernel_param_2,
	.param .u32 ColorHLSKernel_param_3,
	.param .u32 ColorHLSKernel_param_4,
	.param .u32 ColorHLSKernel_param_5,
	.param .u32 ColorHLSKernel_param_6,
	.param .u32 ColorHLSKernel_param_7,
	.param .u32 ColorHLSKernel_param_8,
	.param .align 8 .b8 ColorHLSKernel_param_9[8],
	.param .f32 ColorHLSKernel_param_10,
	.param .f32 ColorHLSKernel_param_11,
	.param .f32 ColorHLSKernel_param_12
)
{
	.reg .pred 	%p<43>;
	.reg .b16 	%rs<9>;
	.reg .f32 	%f<181>;
	.reg .b32 	%r<23>;
	.reg .b64 	%rd<15>;


	ld.param.u64 	%rd5, [ColorHLSKernel_param_0];
	ld.param.u64 	%rd6, [ColorHLSKernel_param_1];
	ld.param.u32 	%r5, [ColorHLSKernel_param_2];
	ld.param.u32 	%r6, [ColorHLSKernel_param_3];
	ld.param.u32 	%r7, [ColorHLSKernel_param_4];
	ld.param.u32 	%r8, [ColorHLSKernel_param_5];
	ld.param.u32 	%r9, [ColorHLSKernel_param_6];
	ld.param.u32 	%r12, [ColorHLSKernel_param_7];
	ld.param.u32 	%r13, [ColorHLSKernel_param_8];
	ld.param.f32 	%f72, [ColorHLSKernel_param_10];
	ld.param.f32 	%f73, [ColorHLSKernel_param_11];
	ld.param.f32 	%f74, [ColorHLSKernel_param_12];
	ld.param.u32 	%r11, [ColorHLSKernel_param_9+4];
	ld.param.u32 	%r10, [ColorHLSKernel_param_9];
	cvta.to.global.u64 	%rd1, %rd6;
	cvta.to.global.u64 	%rd2, %rd5;
	mov.u32 	%r14, %ntid.x;
	mov.u32 	%r15, %ctaid.x;
	mov.u32 	%r16, %tid.x;
	mad.lo.s32 	%r1, %r15, %r14, %r16;
	mov.u32 	%r17, %ntid.y;
	mov.u32 	%r18, %ctaid.y;
	mov.u32 	%r19, %tid.y;
	mad.lo.s32 	%r2, %r18, %r17, %r19;
	setp.ge.s32 	%p1, %r1, %r12;
	setp.ge.s32 	%p2, %r2, %r13;
	or.pred  	%p3, %p1, %p2;
	@%p3 bra 	$L__BB0_40;

	sub.s32 	%r3, %r1, %r10;
	sub.s32 	%r4, %r2, %r11;
	setp.ge.s32 	%p4, %r3, %r7;
	or.b32  	%r20, %r3, %r4;
	setp.lt.s32 	%p5, %r20, 0;
	or.pred  	%p6, %p4, %p5;
	setp.ge.s32 	%p7, %r4, %r8;
	mov.f32 	%f177, 0f00000000;
	or.pred  	%p8, %p7, %p6;
	mov.f32 	%f178, %f177;
	mov.f32 	%f179, %f177;
	mov.f32 	%f180, %f177;
	@%p8 bra 	$L__BB0_37;

	mad.lo.s32 	%r21, %r4, %r6, %r3;
	cvt.s64.s32 	%rd3, %r21;
	setp.eq.s32 	%p9, %r5, 0;
	@%p9 bra 	$L__BB0_4;

	shl.b64 	%rd7, %rd3, 4;
	add.s64 	%rd8, %rd2, %rd7;
	ld.global.v4.f32 	{%f161, %f162, %f163, %f164}, [%rd8];
	bra.uni 	$L__BB0_5;

$L__BB0_4:
	shl.b64 	%rd9, %rd3, 3;
	add.s64 	%rd10, %rd2, %rd9;
	ld.global.u16 	%rs1, [%rd10];
	ld.global.u16 	%rs2, [%rd10+2];
	ld.global.u16 	%rs3, [%rd10+4];
	ld.global.u16 	%rs4, [%rd10+6];
	// begin inline asm
	{  cvt.f32.f16 %f161, %rs1;}

	// end inline asm
	// begin inline asm
	{  cvt.f32.f16 %f162, %rs2;}

	// end inline asm
	// begin inline asm
	{  cvt.f32.f16 %f163, %rs3;}

	// end inline asm
	// begin inline asm
	{  cvt.f32.f16 %f164, %rs4;}

	// end inline asm

$L__BB0_5:
	mov.f32 	%f87, 0f00000000;
	max.ftz.f32 	%f88, %f161, %f87;
	mov.f32 	%f89, 0f3F800000;
	min.ftz.f32 	%f178, %f88, %f89;
	max.ftz.f32 	%f90, %f162, %f87;
	min.ftz.f32 	%f177, %f90, %f89;
	max.ftz.f32 	%f91, %f163, %f87;
	min.ftz.f32 	%f179, %f91, %f89;
	setp.lt.ftz.f32 	%p10, %f73, 0f00000000;
	@%p10 bra 	$L__BB0_8;
	bra.uni 	$L__BB0_6;

$L__BB0_8:
	add.ftz.f32 	%f96, %f73, 0f3F800000;
	mul.ftz.f32 	%f179, %f96, %f179;
	mul.ftz.f32 	%f177, %f96, %f177;
	mul.ftz.f32 	%f178, %f96, %f178;
	bra.uni 	$L__BB0_9;

$L__BB0_6:
	setp.leu.ftz.f32 	%p11, %f73, 0f00000000;
	@%p11 bra 	$L__BB0_9;

	mov.f32 	%f92, 0f3F800000;
	sub.ftz.f32 	%f93, %f92, %f179;
	fma.rn.ftz.f32 	%f179, %f93, %f73, %f179;
	sub.ftz.f32 	%f94, %f92, %f177;
	fma.rn.ftz.f32 	%f177, %f94, %f73, %f177;
	sub.ftz.f32 	%f95, %f92, %f178;
	fma.rn.ftz.f32 	%f178, %f95, %f73, %f178;

$L__BB0_9:
	mov.f32 	%f169, 0f00000000;
	max.ftz.f32 	%f98, %f164, %f169;
	mov.f32 	%f99, 0f3F800000;
	min.ftz.f32 	%f180, %f98, %f99;
	setp.eq.ftz.f32 	%p12, %f74, 0f00000000;
	setp.eq.ftz.f32 	%p13, %f72, 0f00000000;
	and.pred  	%p14, %p13, %p12;
	@%p14 bra 	$L__BB0_37;

	max.ftz.f32 	%f102, %f179, %f177;
	max.ftz.f32 	%f26, %f102, %f178;
	min.ftz.f32 	%f103, %f179, %f177;
	min.ftz.f32 	%f27, %f103, %f178;
	add.ftz.f32 	%f28, %f26, %f27;
	mov.f32 	%f104, 0f40000000;
	div.approx.ftz.f32 	%f29, %f28, %f104;
	setp.eq.ftz.f32 	%p15, %f26, %f27;
	mov.f32 	%f170, 0fBF800000;
	@%p15 bra 	$L__BB0_17;

	setp.lt.ftz.f32 	%p16, %f29, 0f3F000000;
	sub.ftz.f32 	%f106, %f104, %f26;
	sub.ftz.f32 	%f107, %f106, %f27;
	selp.f32 	%f30, %f28, %f107, %p16;
	sub.ftz.f32 	%f31, %f26, %f27;
	setp.eq.ftz.f32 	%p17, %f26, %f179;
	@%p17 bra 	$L__BB0_15;
	bra.uni 	$L__BB0_12;

$L__BB0_15:
	sub.ftz.f32 	%f112, %f177, %f178;
	div.approx.ftz.f32 	%f168, %f112, %f31;
	bra.uni 	$L__BB0_16;

$L__BB0_12:
	setp.eq.ftz.f32 	%p18, %f26, %f177;
	@%p18 bra 	$L__BB0_14;
	bra.uni 	$L__BB0_13;

$L__BB0_14:
	sub.ftz.f32 	%f110, %f178, %f179;
	div.approx.ftz.f32 	%f111, %f110, %f31;
	add.ftz.f32 	%f168, %f111, 0f40000000;
	bra.uni 	$L__BB0_16;

$L__BB0_13:
	sub.ftz.f32 	%f108, %f179, %f177;
	div.approx.ftz.f32 	%f109, %f108, %f31;
	add.ftz.f32 	%f168, %f109, 0f40800000;

$L__BB0_16:
	mul.ftz.f32 	%f113, %f168, 0f42700000;
	setp.lt.ftz.f32 	%p19, %f113, 0f00000000;
	add.ftz.f32 	%f114, %f113, 0f43B40000;
	mov.f32 	%f115, 0f43B40000;
	selp.f32 	%f116, %f114, %f113, %p19;
	div.approx.ftz.f32 	%f170, %f116, %f115;
	div.approx.ftz.f32 	%f169, %f31, %f30;

$L__BB0_17:
	setp.eq.ftz.f32 	%p20, %f170, 0fBF800000;
	setp.le.ftz.f32 	%p21, %f169, 0f3B7F0000;
	or.pred  	%p22, %p21, %p20;
	@%p22 bra 	$L__BB0_37;

	add.ftz.f32 	%f117, %f170, %f72;
	setp.ltu.ftz.f32 	%p23, %f117, 0f3F800000;
	mov.f32 	%f118, 0f3F800000;
	add.ftz.f32 	%f119, %f117, 0fBF800000;
	selp.f32 	%f40, %f117, %f119, %p23;
	sub.ftz.f32 	%f120, %f118, %f169;
	setp.gt.ftz.f32 	%p24, %f74, 0f00000000;
	selp.f32 	%f121, %f120, %f169, %p24;
	fma.rn.ftz.f32 	%f41, %f121, %f74, %f169;
	setp.eq.ftz.f32 	%p25, %f41, 0f00000000;
	@%p25 bra 	$L__BB0_35;
	bra.uni 	$L__BB0_19;

$L__BB0_35:
	mov.f32 	%f178, %f29;
	mov.f32 	%f177, %f29;
	mov.f32 	%f179, %f29;
	bra.uni 	$L__BB0_36;

$L__BB0_19:
	setp.lt.ftz.f32 	%p26, %f29, 0f3F000000;
	add.ftz.f32 	%f122, %f41, 0f3F800000;
	mul.ftz.f32 	%f123, %f29, %f122;
	mul.ftz.f32 	%f124, %f29, %f41;
	add.ftz.f32 	%f125, %f29, %f41;
	sub.ftz.f32 	%f126, %f125, %f124;
	selp.f32 	%f42, %f123, %f126, %p26;
	add.ftz.f32 	%f127, %f29, %f29;
	sub.ftz.f32 	%f43, %f127, %f42;
	add.ftz.f32 	%f128, %f40, 0f3EAAAAAB;
	setp.lt.ftz.f32 	%p27, %f128, 0f00000000;
	add.ftz.f32 	%f129, %f128, 0f3F800000;
	selp.f32 	%f130, %f129, %f128, %p27;
	setp.gt.ftz.f32 	%p28, %f130, 0f3F800000;
	add.ftz.f32 	%f131, %f130, 0fBF800000;
	selp.f32 	%f44, %f131, %f130, %p28;
	setp.lt.ftz.f32 	%p29, %f44, 0f3E2AAAAB;
	@%p29 bra 	$L__BB0_23;
	bra.uni 	$L__BB0_20;

$L__BB0_23:
	sub.ftz.f32 	%f136, %f42, %f43;
	mul.ftz.f32 	%f137, %f136, 0f40C00000;
	fma.rn.ftz.f32 	%f179, %f44, %f137, %f43;
	bra.uni 	$L__BB0_24;

$L__BB0_20:
	setp.lt.ftz.f32 	%p30, %f44, 0f3F000000;
	mov.f32 	%f179, %f42;
	@%p30 bra 	$L__BB0_24;

	setp.geu.ftz.f32 	%p31, %f44, 0f3F2AAAAB;
	mov.f32 	%f179, %f43;
	@%p31 bra 	$L__BB0_24;

	sub.ftz.f32 	%f132, %f42, %f43;
	mov.f32 	%f133, 0f3F2AAAAB;
	sub.ftz.f32 	%f134, %f133, %f44;
	mul.ftz.f32 	%f135, %f132, %f134;
	fma.rn.ftz.f32 	%f179, %f135, 0f40C00000, %f43;

$L__BB0_24:
	add.ftz.f32 	%f138, %f40, 0f3F800000;
	setp.lt.ftz.f32 	%p32, %f40, 0f00000000;
	selp.f32 	%f139, %f138, %f40, %p32;
	setp.gt.ftz.f32 	%p33, %f139, 0f3F800000;
	add.ftz.f32 	%f140, %f139, 0fBF800000;
	selp.f32 	%f48, %f140, %f139, %p33;
	setp.lt.ftz.f32 	%p34, %f48, 0f3E2AAAAB;
	@%p34 bra 	$L__BB0_28;
	bra.uni 	$L__BB0_25;

$L__BB0_28:
	sub.ftz.f32 	%f145, %f42, %f43;
	mul.ftz.f32 	%f146, %f145, 0f40C00000;
	fma.rn.ftz.f32 	%f177, %f48, %f146, %f43;
	bra.uni 	$L__BB0_29;

$L__BB0_25:
	setp.lt.ftz.f32 	%p35, %f48, 0f3F000000;
	mov.f32 	%f177, %f42;
	@%p35 bra 	$L__BB0_29;

	setp.geu.ftz.f32 	%p36, %f48, 0f3F2AAAAB;
	mov.f32 	%f177, %f43;
	@%p36 bra 	$L__BB0_29;

	sub.ftz.f32 	%f141, %f42, %f43;
	mov.f32 	%f142, 0f3F2AAAAB;
	sub.ftz.f32 	%f143, %f142, %f48;
	mul.ftz.f32 	%f144, %f141, %f143;
	fma.rn.ftz.f32 	%f177, %f144, 0f40C00000, %f43;

$L__BB0_29:
	add.ftz.f32 	%f147, %f40, 0fBEAAAAAB;
	setp.lt.ftz.f32 	%p37, %f147, 0f00000000;
	add.ftz.f32 	%f148, %f147, 0f3F800000;
	selp.f32 	%f149, %f148, %f147, %p37;
	setp.gt.ftz.f32 	%p38, %f149, 0f3F800000;
	add.ftz.f32 	%f150, %f149, 0fBF800000;
	selp.f32 	%f52, %f150, %f149, %p38;
	setp.lt.ftz.f32 	%p39, %f52, 0f3E2AAAAB;
	@%p39 bra 	$L__BB0_33;
	bra.uni 	$L__BB0_30;

$L__BB0_33:
	sub.ftz.f32 	%f155, %f42, %f43;
	mul.ftz.f32 	%f156, %f155, 0f40C00000;
	fma.rn.ftz.f32 	%f178, %f52, %f156, %f43;
	bra.uni 	$L__BB0_36;

$L__BB0_30:
	setp.lt.ftz.f32 	%p40, %f52, 0f3F000000;
	mov.f32 	%f178, %f42;
	@%p40 bra 	$L__BB0_36;

	setp.geu.ftz.f32 	%p41, %f52, 0f3F2AAAAB;
	mov.f32 	%f178, %f43;
	@%p41 bra 	$L__BB0_36;

	sub.ftz.f32 	%f151, %f42, %f43;
	mov.f32 	%f152, 0f3F2AAAAB;
	sub.ftz.f32 	%f153, %f152, %f52;
	mul.ftz.f32 	%f154, %f151, %f153;
	fma.rn.ftz.f32 	%f178, %f154, 0f40C00000, %f43;

$L__BB0_36:
	mov.f32 	%f180, %f164;

$L__BB0_37:
	mad.lo.s32 	%r22, %r2, %r9, %r1;
	cvt.s64.s32 	%rd4, %r22;
	setp.eq.s32 	%p42, %r5, 0;
	@%p42 bra 	$L__BB0_39;

	shl.b64 	%rd11, %rd4, 4;
	add.s64 	%rd12, %rd1, %rd11;
	st.global.v4.f32 	[%rd12], {%f178, %f177, %f179, %f180};
	bra.uni 	$L__BB0_40;

$L__BB0_39:
	// begin inline asm
	{  cvt.rn.f16.f32 %rs5, %f178;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs6, %f177;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs7, %f179;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs8, %f180;}

	// end inline asm
	shl.b64 	%rd13, %rd4, 3;
	add.s64 	%rd14, %rd1, %rd13;
	st.global.u16 	[%rd14], %rs5;
	st.global.u16 	[%rd14+2], %rs6;
	st.global.u16 	[%rd14+4], %rs7;
	st.global.u16 	[%rd14+6], %rs8;

$L__BB0_40:
	ret;

}

  ELF3         {           "            V2 @ 8  @    .shstrtab .strtab .symtab .symtab_shndx .nv.uft.entry .nv.info .text.ColorHLSKernel .nv.info.ColorHLSKernel .nv.shared.ColorHLSKernel .nv.constant3 .nv.constant0.ColorHLSKernel .debug_frame .rel.debug_frame .rela.debug_frame .nv.rel.action  .shstrtab .strtab .symtab .symtab_shndx .nv.uft.entry .nv.info ColorHLSKernel .text.ColorHLSKernel .nv.info.ColorHLSKernel .nv.shared.ColorHLSKernel .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 .nv.constant0.ColorHLSKernel _param .debug_frame .rel.debug_frame .rela.debug_frame .nv.rel.action                              O                         	                     	         $           	 $       $           	 H       $           	 l       $           	        $          	        $          	        $       6   	        $       R   	        $       o   	 D      $          	 h      $          	       $          	       $          	       $          	       $          	       $          	 @      $          	 d      $       5   	       $       K   	       $       W   	              d   	              z   	                 	                 	                 	                 	                 	                 	                  	                 	                 	                 	                 	                 	                 	                 	                  	 $                	 (                	 ,                	 0                	 4                	 8                	 <                	 @             &   	 D             8   	 H             ?   	 L             F   	 P             M   	 T             W   
                 {                                        @                   $        |( ((   4                                 $   (          / 8      # 8        8        8       7 {   5  
 5   `D D       @         <        
 8        	 0  !       (         $                                                         !           !      P          s          % 6D          8   >E?x=!,o   ?   ?F^־膦  ?    t?  ?26  ??    FB C^A!0g  B  BoF%;    ;%;(ɺP%;<    y>?=>>NG
?    J?
?ȾP
?N@    >E?x=,'>>+վ;ߥ  ?    H?  ?6  ??    q}B\C\Ao+1H  B  BX9;    t;;26;;    гY>Y7?ݓ=Vž   ?   ?;  ?    ?  ??  ?M?    j<:BC|AO  B  BuQ$%;    ^;%;!9[%;R<    ;    ;;=?ﺃ;R;    :>5>?}=νY>>B̾$
?    x?
?5^Zl
?1@      ?d=MD>    g}?
    =b{?  ?z콳T    b?ؼ=    J==?                                      A   C   C       C   C  #>k@ BO<  V? ЖA A@ @F?VnB39Cff>  B7> >OV?)@   A=?L>*>>h-?r=  zD                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                        z  
       y       &   ( y      "   ( y      %   h y      !   b $z       z   b  pb  $z     / z  a  pfp   M	          z d   z  F   
    z  e   Ey        x         x
         r     r    pb  z  ^  pfp  z  _  pfp   G	  
      z  \  pR  $z ]    x      

 X     X   
 Y  $   Y    	
                (     h 	     z  g     0    A   ď 0    A   0    A   !r       O 0	    A    !r	        !r
        G         z  g     G         !t
  ?    !t	  ?    !t  ?    #z
 g  
    #z	 g  	    #z g      Gy  P       z  g       !t  ?      r

     A   r		     A   r     A   z  h     !r        z  f     r         G        	r	
      Ey       	r	
       	r      	r       x        r      !r         x   ?  A   G        r 
     !t   @   x    ?   Ey        !r       !       G  p       r 	      s        6 !	
      !
       #  @   #   @    Gy  0       s        " !r	        r     A   Ay          x  pB  A   s        & r       !  C      r     A    xa6;  A   Ay         x      x   ;   x      r   G	        !t	  ?    z  h     !v f       Ey P      r		         x   ? `  #z	 h      r       !       G        !t>     x    ?  !t  ?     !r       Ey       r        r     A  !r        #      !  ?     !r       x   ? @  !       x *>   G         x    ?   r
         G         x *?   r
         G	         !t	*?    !r        r	     A   #x
  @    Gy  0       !r	        x		  @  A   #r
	       Ay         !t  ?    r       Ey @      r        x   ? @  !       x *>   G         x    ?   r	         G         x *?   r	         G	         !t*?    !r	        r	     A   #x	  @    Gy  0       !r        x	  @  A   #r		       Ay         !t     r       !  ?     x   ? @  !       x *>   G         x    ?   r         G         x *?   r         G	         !r       !t*?     r     A   #x  @    Gy  p       !r        x  @  A   #r       Gy  0       r        r	        r
         Ay         r         Ay          z  \  pR  $z   `    x         Z       [  $         M          z  Z   >r	        >r

        z  [     x 2v        x
2v       y      y      y 
    y     My          Gy    y            y            y            y            y            y            y            y            y            y            y            y            y            y                                                                                                  @                                                           1                                                               X         8                                       H
      p                              7      p                
      0                             U      p                
                                      p                                                      	                                                                                X                                                  h                                  @                                     8                       "                                                                       p      p                   "                                           