//
// Generated by NVIDIA NVVM Compiler
// Compiler built on Wed Jul 10 12:41:20 2013 (1373485280)
// Cuda compilation tools, release 5.5, V5.5.0
//

.version 3.2
.target sm_30
.address_size 64

	.file	1 "D:/singlebarrel/releases/2014.03/shared/adobe/MediaCore/Renderers/RendererGPU/Src/Effects/ThreeWayColorCorrector.cu", 1399785316, 12641
	.file	2 "d:\\singlebarrel\\releases\\2014.03\\shared\\adobe\\mediacore\\external\\3rdparty\\nvidia\\cuda\\win\\include\\device_functions.h", 1399785281, 191626
	.file	3 "d:\\singlebarrel\\releases\\2014.03\\shared\\adobe\\mediacore\\external\\3rdparty\\nvidia\\cuda\\win\\include\\math_functions.h", 1399785281, 404374
.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, 188, 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, 147, 24, 206, 61, 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 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 .u32 kRandMax = 32767;
.global .align 1 .b8 $str[11] = {95, 95, 67, 85, 68, 65, 95, 70, 84, 90, 0};

.visible .func  (.param .b32 func_retval0) _Z21CalculateThreeWayLumafffffffi(
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_0,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_1,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_2,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_3,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_4,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_5,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_6,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_7
)
{
	.reg .pred 	%p<6>;
	.reg .s32 	%r<55>;
	.reg .f32 	%f<39>;


	ld.param.f32 	%f9, [_Z21CalculateThreeWayLumafffffffi_param_0];
	ld.param.f32 	%f10, [_Z21CalculateThreeWayLumafffffffi_param_1];
	ld.param.f32 	%f11, [_Z21CalculateThreeWayLumafffffffi_param_2];
	ld.param.f32 	%f14, [_Z21CalculateThreeWayLumafffffffi_param_3];
	ld.param.f32 	%f12, [_Z21CalculateThreeWayLumafffffffi_param_4];
	ld.param.f32 	%f13, [_Z21CalculateThreeWayLumafffffffi_param_5];
	ld.param.f32 	%f15, [_Z21CalculateThreeWayLumafffffffi_param_6];
	ld.param.u32 	%r1, [_Z21CalculateThreeWayLumafffffffi_param_7];
	.loc 1 42 1
	sub.ftz.f32 	%f1, %f14, %f11;
	.loc 1 42 1
	setp.neu.ftz.f32	%p1, %f1, 0f3F800000;
	.loc 1 42 1
	sub.ftz.f32 	%f2, %f15, %f13;
	.loc 1 42 1
	setp.neu.ftz.f32	%p2, %f2, 0f3F800000;
	or.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	mov.f32 	%f37, %f9;
	@!%p3 bra 	BB0_2;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 42 1
	mov.u32 	%r2, %ctaid.x;
	mov.u32 	%r3, %ntid.x;
	mov.u32 	%r4, %ntid.y;
	mov.u32 	%r5, %ctaid.y;
	mov.u32 	%r6, %tid.y;
	mad.lo.s32 	%r7, %r4, %r5, %r6;
	mov.u32 	%r8, %tid.x;
	mad.lo.s32 	%r9, %r3, %r2, %r8;
	sub.s32 	%r10, %r9, %r7;
	add.s32 	%r11, %r10, -1;
	mov.u32 	%r12, 1;
	sub.s32 	%r13, %r12, %r10;
	.loc 1 42 1
	add.s32 	%r14, %r7, %r13;
	add.s32 	%r15, %r14, -1;
	shl.b32 	%r16, %r11, 8;
	xor.b32  	%r17, %r15, %r16;
	mov.u32 	%r18, 2;
	.loc 1 42 1
	sub.s32 	%r19, %r18, %r10;
	sub.s32 	%r20, %r19, %r17;
	shr.u32 	%r21, %r17, 13;
	xor.b32  	%r22, %r20, %r21;
	sub.s32 	%r23, %r11, %r17;
	sub.s32 	%r24, %r23, %r22;
	shr.u32 	%r25, %r22, 12;
	xor.b32  	%r26, %r24, %r25;
	sub.s32 	%r27, %r17, %r22;
	sub.s32 	%r28, %r27, %r26;
	shl.b32 	%r29, %r26, 16;
	xor.b32  	%r30, %r28, %r29;
	sub.s32 	%r31, %r22, %r26;
	sub.s32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 5;
	xor.b32  	%r34, %r32, %r33;
	sub.s32 	%r35, %r26, %r30;
	sub.s32 	%r36, %r35, %r34;
	shr.u32 	%r37, %r34, 3;
	xor.b32  	%r38, %r36, %r37;
	sub.s32 	%r39, %r30, %r34;
	sub.s32 	%r40, %r39, %r38;
	shl.b32 	%r41, %r38, 10;
	xor.b32  	%r42, %r40, %r41;
	sub.s32 	%r43, %r34, %r38;
	sub.s32 	%r44, %r43, %r42;
	shr.u32 	%r45, %r42, 15;
	xor.b32  	%r46, %r44, %r45;
	mad.lo.s32 	%r47, %r46, 1103515245, 12345;
	mad.lo.s32 	%r48, %r47, 1103515245, 12345;
	shr.u32 	%r49, %r47, 9;
	and.b32  	%r50, %r49, 32640;
	shr.u32 	%r51, %r48, 16;
	and.b32  	%r52, %r51, 255;
	xor.b32  	%r53, %r52, %r50;
	cvt.rn.f32.s32	%f16, %r53;
	ld.const.u32 	%r54, [kRandMax];
	cvt.rn.f32.u32	%f17, %r54;
	mov.f32 	%f18, 0f40000000;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f19, %f17, %f18;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f20, %f16, %f19;
	.loc 1 42 60
	add.ftz.f32 	%f21, %f20, 0fBF800000;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f3, %f21, 0f3B270D73, %f9;
	mov.f32 	%f37, %f3;

BB0_2:
	.loc 1 42 1
	mov.f32 	%f4, %f37;
	sub.ftz.f32 	%f22, %f4, %f11;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f23, %f22, %f1;
	mov.f32 	%f24, 0f00000000;
	.loc 2 2770 10
	max.ftz.f32 	%f5, %f24, %f23;
	setp.ltu.ftz.f32	%p4, %f5, 0f00000000;
	@%p4 bra 	BB0_4;

	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f25, %f5;
	mul.ftz.f32 	%f26, %f25, %f12;
	ex2.approx.ftz.f32 	%f38, %f26;
	bra.uni 	BB0_5;

BB0_4:
	neg.ftz.f32 	%f27, %f5;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f28, %f27;
	mul.ftz.f32 	%f29, %f28, %f12;
	ex2.approx.ftz.f32 	%f30, %f29;
	neg.ftz.f32 	%f38, %f30;

BB0_5:
	fma.rn.ftz.f32 	%f31, %f2, %f38, %f13;
	.loc 1 42 1
	sub.ftz.f32 	%f32, %f4, %f31;
	fma.rn.ftz.f32 	%f33, %f32, 0f3F000000, %f31;
	.loc 1 42 1
	setp.eq.s32	%p5, %r1, 0;
	selp.f32	%f34, %f31, %f33, %p5;
	.loc 1 42 1
	sub.ftz.f32 	%f35, %f34, %f9;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f36, %f35, %f10, %f9;
	st.param.f32	[func_retval0+0], %f36;
	ret;
}

.visible .entry ThreeWayColorCorrector_MaskKernel(
	.param .u64 ThreeWayColorCorrector_MaskKernel_param_0,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_1,
	.param .u64 ThreeWayColorCorrector_MaskKernel_param_2,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_3,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_4,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_5,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_6
)
{
	.reg .pred 	%p<7>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<16>;
	.reg .f32 	%f<45>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd6, [ThreeWayColorCorrector_MaskKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_MaskKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_MaskKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_MaskKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_MaskKernel_param_4];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_MaskKernel_param_5];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_MaskKernel_param_6];
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd2, %rd6;
	.loc 1 42 1
	mov.u32 	%r8, %ntid.x;
	mov.u32 	%r9, %ctaid.x;
	mov.u32 	%r10, %tid.x;
	mad.lo.s32 	%r1, %r8, %r9, %r10;
	mov.u32 	%r11, %ntid.y;
	mov.u32 	%r12, %ctaid.y;
	mov.u32 	%r13, %tid.y;
	mad.lo.s32 	%r2, %r11, %r12, %r13;
	.loc 1 42 1
	setp.lt.s32	%p1, %r1, %r6;
	setp.lt.s32	%p2, %r2, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	@!%p3 bra 	BB1_10;
	bra.uni 	BB1_1;

BB1_1:
	.loc 1 42 1
	mad.lo.s32 	%r14, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r14, 16;
	add.s64 	%rd3, %rd2, %rd7;
	mul.wide.s32 	%rd8, %r14, 8;
	add.s64 	%rd4, %rd2, %rd8;
	.loc 1 42 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB1_3;

	ld.global.v4.f32 	{%f23, %f24, %f25, %f26}, [%rd3];
	mov.f32 	%f43, %f26;
	mov.f32 	%f3, %f25;
	mov.f32 	%f2, %f24;
	mov.f32 	%f1, %f23;
	bra.uni 	BB1_4;

BB1_3:
	.loc 1 42 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f43, %temp;
	}

BB1_4:
	.loc 1 42 1
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB1_6;

	mov.f32 	%f44, 0f3F800000;
	bra.uni 	BB1_7;

BB1_6:
	.loc 1 42 1
	mad.lo.s32 	%r15, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r15, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 42 1
	ld.global.f32 	%f44, [%rd10];

BB1_7:
	.loc 1 42 1
	ld.const.f32 	%f28, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f29, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f30, %f29, 0f00000000;
	fma.rn.ftz.f32 	%f31, %f44, %f28, %f30;
	ld.const.f32 	%f32, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f16, %f32, 0f00000000, %f31;
	ld.const.f32 	%f33, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f35, %f34, 0f00000000;
	fma.rn.ftz.f32 	%f36, %f44, %f33, %f35;
	ld.const.f32 	%f37, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f17, %f37, 0f00000000, %f36;
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f39, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f40, %f39, 0f00000000;
	fma.rn.ftz.f32 	%f41, %f44, %f38, %f40;
	ld.const.f32 	%f42, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f18, %f42, 0f00000000, %f41;
	.loc 1 42 1
	@%p4 bra 	BB1_9;

	.loc 1 42 1
	st.global.v4.f32 	[%rd3], {%f18, %f17, %f16, %f43};
	bra.uni 	BB1_10;

BB1_9:
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f18;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f17;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f16;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f43;
	mov.b16 	%rs12, %temp;
}
	.loc 1 42 231
	st.global.v4.u16 	[%rd4], {%rs9, %rs10, %rs11, %rs12};

BB1_10:
	.loc 1 42 2
	ret;
}

.visible .entry ThreeWayColorCorrector_TonalRangeKernel(
	.param .u64 ThreeWayColorCorrector_TonalRangeKernel_param_0,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_1,
	.param .u64 ThreeWayColorCorrector_TonalRangeKernel_param_2,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_3,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_4,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_5,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_6,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_7,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_8,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_9,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_10
)
{
	.reg .pred 	%p<21>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<82>;
	.reg .s64 	%rd<7>;


	ld.param.u64 	%rd4, [ThreeWayColorCorrector_TonalRangeKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_TonalRangeKernel_param_1];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_TonalRangeKernel_param_4];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_TonalRangeKernel_param_5];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_TonalRangeKernel_param_6];
	ld.param.f32 	%f27, [ThreeWayColorCorrector_TonalRangeKernel_param_7];
	ld.param.f32 	%f28, [ThreeWayColorCorrector_TonalRangeKernel_param_8];
	ld.param.f32 	%f29, [ThreeWayColorCorrector_TonalRangeKernel_param_9];
	ld.param.f32 	%f30, [ThreeWayColorCorrector_TonalRangeKernel_param_10];
	cvta.to.global.u64 	%rd1, %rd4;
	.loc 1 42 1
	mov.u32 	%r7, %ntid.x;
	mov.u32 	%r8, %ctaid.x;
	mov.u32 	%r9, %tid.x;
	mad.lo.s32 	%r1, %r7, %r8, %r9;
	mov.u32 	%r10, %ntid.y;
	mov.u32 	%r11, %ctaid.y;
	mov.u32 	%r12, %tid.y;
	mad.lo.s32 	%r2, %r10, %r11, %r12;
	.loc 1 42 1
	setp.lt.s32	%p1, %r1, %r5;
	setp.lt.s32	%p2, %r2, %r6;
	and.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	@!%p3 bra 	BB2_20;
	bra.uni 	BB2_1;

BB2_1:
	.loc 1 42 1
	mad.lo.s32 	%r13, %r2, %r3, %r1;
	mul.wide.s32 	%rd5, %r13, 16;
	add.s64 	%rd2, %rd1, %rd5;
	mul.wide.s32 	%rd6, %r13, 8;
	add.s64 	%rd3, %rd1, %rd6;
	.loc 1 42 1
	setp.eq.s32	%p4, %r4, 0;
	@%p4 bra 	BB2_3;

	ld.global.v4.f32 	{%f31, %f32, %f33, %f34}, [%rd2];
	mov.f32 	%f77, %f34;
	mov.f32 	%f76, %f33;
	mov.f32 	%f75, %f32;
	mov.f32 	%f74, %f31;
	bra.uni 	BB2_4;

BB2_3:
	.loc 1 42 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f74, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f75, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f76, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f77, %temp;
	}

BB2_4:
	ld.const.f32 	%f35, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f36, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f37, %f75, %f36;
	fma.rn.ftz.f32 	%f38, %f76, %f35, %f37;
	ld.const.f32 	%f39, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f74, %f39, %f38;
	.loc 1 42 1
	add.ftz.f32 	%f40, %f27, %f28;
	mov.f32 	%f41, 0f3F800000;
	.loc 2 2765 10
	min.ftz.f32 	%f15, %f41, %f40;
	.loc 1 42 1
	sub.ftz.f32 	%f42, %f29, %f30;
	mov.f32 	%f43, 0f00000000;
	.loc 2 2770 10
	max.ftz.f32 	%f16, %f43, %f42;
	.loc 1 42 1
	setp.gtu.ftz.f32	%p5, %f14, %f27;
	@%p5 bra 	BB2_6;

	mov.f32 	%f81, %f43;
	bra.uni 	BB2_17;

BB2_6:
	.loc 1 42 1
	setp.eq.ftz.f32	%p6, %f28, 0f00000000;
	setp.le.ftz.f32	%p7, %f14, %f15;
	and.pred  	%p8, %p7, %p6;
	.loc 1 42 1
	@!%p8 bra 	BB2_8;
	bra.uni 	BB2_7;

BB2_7:
	mov.f32 	%f81, %f43;
	bra.uni 	BB2_17;

BB2_8:
	.loc 1 42 1
	setp.ltu.ftz.f32	%p9, %f14, %f29;
	@%p9 bra 	BB2_10;

	mov.f32 	%f81, %f41;
	bra.uni 	BB2_17;

BB2_10:
	.loc 1 42 1
	setp.eq.ftz.f32	%p10, %f30, 0f00000000;
	setp.ge.ftz.f32	%p11, %f14, %f16;
	and.pred  	%p12, %p11, %p10;
	.loc 1 42 1
	@!%p12 bra 	BB2_12;
	bra.uni 	BB2_11;

BB2_11:
	mov.f32 	%f81, %f41;
	bra.uni 	BB2_17;

BB2_12:
	.loc 1 42 1
	setp.gt.ftz.f32	%p13, %f14, %f15;
	setp.lt.ftz.f32	%p14, %f14, %f16;
	and.pred  	%p15, %p13, %p14;
	.loc 1 42 1
	mov.f32 	%f81, 0f3F000000;
	.loc 1 42 1
	@!%p15 bra 	BB2_13;
	bra.uni 	BB2_17;

BB2_13:
	.loc 1 42 1
	setp.gt.ftz.f32	%p16, %f14, %f27;
	setp.lt.ftz.f32	%p17, %f14, %f15;
	and.pred  	%p18, %p16, %p17;
	.loc 1 42 1
	@%p18 bra 	BB2_16;

	.loc 1 42 1
	setp.gt.ftz.f32	%p19, %f14, %f16;
	@%p19 bra 	BB2_15;
	bra.uni 	BB2_17;

BB2_15:
	.loc 1 42 1
	sub.ftz.f32 	%f47, %f29, %f16;
	sub.ftz.f32 	%f48, %f14, %f16;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f49, %f48, %f47;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f81, %f49, 0f3F000000, 0f3F000000;
	bra.uni 	BB2_17;

BB2_16:
	.loc 1 42 1
	sub.ftz.f32 	%f50, %f15, %f27;
	sub.ftz.f32 	%f51, %f14, %f27;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f52, %f51, %f50;
	.loc 1 42 1
	mul.ftz.f32 	%f81, %f52, 0f3F000000;

BB2_17:
	.loc 1 42 1
	ld.const.f32 	%f56, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f57, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f58, %f57, 0f00000000;
	fma.rn.ftz.f32 	%f59, %f81, %f56, %f58;
	ld.const.f32 	%f60, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f20, %f60, 0f00000000, %f59;
	ld.const.f32 	%f61, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f62, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f63, %f62, 0f00000000;
	fma.rn.ftz.f32 	%f64, %f81, %f61, %f63;
	ld.const.f32 	%f65, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f21, %f65, 0f00000000, %f64;
	ld.const.f32 	%f66, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f67, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f68, %f67, 0f00000000;
	fma.rn.ftz.f32 	%f69, %f81, %f66, %f68;
	ld.const.f32 	%f70, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f22, %f70, 0f00000000, %f69;
	.loc 1 42 1
	@%p4 bra 	BB2_19;

	.loc 1 42 1
	st.global.v4.f32 	[%rd2], {%f22, %f21, %f20, %f77};
	bra.uni 	BB2_20;

BB2_19:
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f22;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f21;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f20;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f77;
	mov.b16 	%rs12, %temp;
}
	.loc 1 42 241
	st.global.v4.u16 	[%rd3], {%rs9, %rs10, %rs11, %rs12};

BB2_20:
	.loc 1 42 2
	ret;
}

.visible .entry ThreeWayColorCorrector_LumaKernel(
	.param .u64 ThreeWayColorCorrector_LumaKernel_param_0,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_1,
	.param .u64 ThreeWayColorCorrector_LumaKernel_param_2,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_3,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_4,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_5,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_6,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_7,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_8,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_9,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_10,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_11,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_12
)
{
	.reg .pred 	%p<12>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<62>;
	.reg .f32 	%f<90>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd6, [ThreeWayColorCorrector_LumaKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_LumaKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_LumaKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_LumaKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_LumaKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_LumaKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_LumaKernel_param_6];
	ld.param.f32 	%f32, [ThreeWayColorCorrector_LumaKernel_param_7];
	ld.param.f32 	%f33, [ThreeWayColorCorrector_LumaKernel_param_8];
	ld.param.f32 	%f34, [ThreeWayColorCorrector_LumaKernel_param_9];
	ld.param.f32 	%f35, [ThreeWayColorCorrector_LumaKernel_param_10];
	ld.param.f32 	%f36, [ThreeWayColorCorrector_LumaKernel_param_11];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_LumaKernel_param_12];
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd2, %rd6;
	.loc 1 42 1
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	.loc 1 42 1
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	@!%p3 bra 	BB3_15;
	bra.uni 	BB3_1;

BB3_1:
	.loc 1 42 1
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd3, %rd2, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd4, %rd2, %rd8;
	.loc 1 42 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB3_3;

	ld.global.v4.f32 	{%f37, %f38, %f39, %f40}, [%rd3];
	mov.f32 	%f86, %f40;
	mov.f32 	%f85, %f39;
	mov.f32 	%f84, %f38;
	mov.f32 	%f83, %f37;
	bra.uni 	BB3_4;

BB3_3:
	.loc 1 42 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f83, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f84, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f85, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f86, %temp;
	}

BB3_4:
	ld.const.f32 	%f41, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f42, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f43, %f84, %f42;
	fma.rn.ftz.f32 	%f44, %f85, %f41, %f43;
	ld.const.f32 	%f45, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f83, %f45, %f44;
	.loc 1 42 1
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB3_6;

	mov.f32 	%f87, 0f3F800000;
	bra.uni 	BB3_7;

BB3_6:
	.loc 1 42 1
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 42 1
	ld.global.f32 	%f87, [%rd10];

BB3_7:
	.loc 1 42 1
	sub.ftz.f32 	%f17, %f33, %f32;
	.loc 1 42 1
	setp.neu.ftz.f32	%p6, %f17, 0f3F800000;
	.loc 1 42 1
	sub.ftz.f32 	%f18, %f36, %f35;
	.loc 1 42 1
	setp.neu.ftz.f32	%p7, %f18, 0f3F800000;
	or.pred  	%p8, %p6, %p7;
	.loc 1 42 1
	mov.f32 	%f88, %f14;
	@!%p8 bra 	BB3_9;
	bra.uni 	BB3_8;

BB3_8:
	.loc 1 42 1
	sub.s32 	%r17, %r1, %r2;
	add.s32 	%r18, %r17, -1;
	mov.u32 	%r19, 1;
	sub.s32 	%r20, %r19, %r17;
	.loc 1 42 1
	add.s32 	%r21, %r2, %r20;
	add.s32 	%r22, %r21, -1;
	shl.b32 	%r23, %r18, 8;
	xor.b32  	%r24, %r22, %r23;
	mov.u32 	%r25, 2;
	.loc 1 42 1
	sub.s32 	%r26, %r25, %r17;
	sub.s32 	%r27, %r26, %r24;
	shr.u32 	%r28, %r24, 13;
	xor.b32  	%r29, %r27, %r28;
	sub.s32 	%r30, %r18, %r24;
	sub.s32 	%r31, %r30, %r29;
	shr.u32 	%r32, %r29, 12;
	xor.b32  	%r33, %r31, %r32;
	sub.s32 	%r34, %r24, %r29;
	sub.s32 	%r35, %r34, %r33;
	shl.b32 	%r36, %r33, 16;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r29, %r33;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 5;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r33, %r37;
	sub.s32 	%r43, %r42, %r41;
	shr.u32 	%r44, %r41, 3;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shl.b32 	%r48, %r45, 10;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 15;
	xor.b32  	%r53, %r51, %r52;
	mad.lo.s32 	%r54, %r53, 1103515245, 12345;
	mad.lo.s32 	%r55, %r54, 1103515245, 12345;
	shr.u32 	%r56, %r54, 9;
	and.b32  	%r57, %r56, 32640;
	shr.u32 	%r58, %r55, 16;
	and.b32  	%r59, %r58, 255;
	xor.b32  	%r60, %r59, %r57;
	cvt.rn.f32.s32	%f47, %r60;
	ld.const.u32 	%r61, [kRandMax];
	cvt.rn.f32.u32	%f48, %r61;
	mov.f32 	%f49, 0f40000000;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f50, %f48, %f49;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f51, %f47, %f50;
	.loc 1 42 60
	add.ftz.f32 	%f52, %f51, 0fBF800000;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f19, %f52, 0f3B270D73, %f14;
	mov.f32 	%f88, %f19;

BB3_9:
	.loc 1 42 1
	mov.f32 	%f20, %f88;
	sub.ftz.f32 	%f53, %f20, %f32;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f54, %f53, %f17;
	mov.f32 	%f55, 0f00000000;
	.loc 2 2770 10
	max.ftz.f32 	%f21, %f55, %f54;
	setp.ltu.ftz.f32	%p9, %f21, 0f00000000;
	@%p9 bra 	BB3_11;

	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f56, %f21;
	mul.ftz.f32 	%f57, %f56, %f34;
	ex2.approx.ftz.f32 	%f89, %f57;
	bra.uni 	BB3_12;

BB3_11:
	neg.ftz.f32 	%f58, %f21;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f59, %f58;
	mul.ftz.f32 	%f60, %f59, %f34;
	ex2.approx.ftz.f32 	%f61, %f60;
	neg.ftz.f32 	%f89, %f61;

BB3_12:
	fma.rn.ftz.f32 	%f62, %f18, %f89, %f35;
	.loc 1 42 1
	sub.ftz.f32 	%f63, %f20, %f62;
	fma.rn.ftz.f32 	%f64, %f63, 0f3F000000, %f62;
	.loc 1 42 1
	setp.eq.s32	%p10, %r6, 0;
	selp.f32	%f65, %f62, %f64, %p10;
	.loc 1 42 1
	sub.ftz.f32 	%f66, %f65, %f14;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f67, %f87, %f66, %f14;
	.loc 1 42 1
	ld.const.f32 	%f68, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f69, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f70, %f69, 0f00000000;
	fma.rn.ftz.f32 	%f71, %f67, %f68, %f70;
	ld.const.f32 	%f72, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f25, %f72, 0f00000000, %f71;
	ld.const.f32 	%f73, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f74, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f75, %f74, 0f00000000;
	fma.rn.ftz.f32 	%f76, %f67, %f73, %f75;
	ld.const.f32 	%f77, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f26, %f77, 0f00000000, %f76;
	ld.const.f32 	%f78, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f79, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f80, %f79, 0f00000000;
	fma.rn.ftz.f32 	%f81, %f67, %f78, %f80;
	ld.const.f32 	%f82, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f27, %f82, 0f00000000, %f81;
	.loc 1 42 1
	@%p4 bra 	BB3_14;

	.loc 1 42 1
	st.global.v4.f32 	[%rd3], {%f27, %f26, %f25, %f86};
	bra.uni 	BB3_15;

BB3_14:
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f27;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f26;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f25;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f86;
	mov.b16 	%rs12, %temp;
}
	.loc 1 42 241
	st.global.v4.u16 	[%rd4], {%rs9, %rs10, %rs11, %rs12};

BB3_15:
	.loc 1 42 2
	ret;
}

.visible .entry ThreeWayColorCorrector_CompositeKernel(
	.param .u64 ThreeWayColorCorrector_CompositeKernel_param_0,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_1,
	.param .u64 ThreeWayColorCorrector_CompositeKernel_param_2,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_3,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_4,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_5,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_6,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_7,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_8,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_9,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_10,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_11,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_12,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_13,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_14,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_15,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_16,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_17,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_18,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_19,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_20,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_21,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_22,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_23,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_24,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_25,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_26,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_27,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_28,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_29,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_30,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_31,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_32,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_33,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_34,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_35
)
{
	.reg .pred 	%p<27>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<88>;
	.reg .f32 	%f<208>;
	.reg .s64 	%rd<17>;


	ld.param.u64 	%rd5, [ThreeWayColorCorrector_CompositeKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_CompositeKernel_param_1];
	ld.param.u64 	%rd6, [ThreeWayColorCorrector_CompositeKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_CompositeKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_CompositeKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_CompositeKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_CompositeKernel_param_6];
	ld.param.f32 	%f52, [ThreeWayColorCorrector_CompositeKernel_param_7];
	ld.param.f32 	%f53, [ThreeWayColorCorrector_CompositeKernel_param_8];
	ld.param.f32 	%f54, [ThreeWayColorCorrector_CompositeKernel_param_9];
	ld.param.f32 	%f55, [ThreeWayColorCorrector_CompositeKernel_param_10];
	ld.param.f32 	%f56, [ThreeWayColorCorrector_CompositeKernel_param_11];
	ld.param.f32 	%f57, [ThreeWayColorCorrector_CompositeKernel_param_12];
	ld.param.f32 	%f58, [ThreeWayColorCorrector_CompositeKernel_param_13];
	ld.param.f32 	%f59, [ThreeWayColorCorrector_CompositeKernel_param_14];
	ld.param.f32 	%f60, [ThreeWayColorCorrector_CompositeKernel_param_15];
	ld.param.f32 	%f61, [ThreeWayColorCorrector_CompositeKernel_param_16];
	ld.param.f32 	%f62, [ThreeWayColorCorrector_CompositeKernel_param_17];
	ld.param.f32 	%f63, [ThreeWayColorCorrector_CompositeKernel_param_18];
	ld.param.f32 	%f64, [ThreeWayColorCorrector_CompositeKernel_param_19];
	ld.param.f32 	%f65, [ThreeWayColorCorrector_CompositeKernel_param_20];
	ld.param.f32 	%f66, [ThreeWayColorCorrector_CompositeKernel_param_21];
	ld.param.f32 	%f67, [ThreeWayColorCorrector_CompositeKernel_param_22];
	ld.param.f32 	%f68, [ThreeWayColorCorrector_CompositeKernel_param_23];
	ld.param.f32 	%f69, [ThreeWayColorCorrector_CompositeKernel_param_24];
	ld.param.f32 	%f70, [ThreeWayColorCorrector_CompositeKernel_param_25];
	ld.param.f32 	%f71, [ThreeWayColorCorrector_CompositeKernel_param_26];
	ld.param.f32 	%f72, [ThreeWayColorCorrector_CompositeKernel_param_27];
	ld.param.f32 	%f73, [ThreeWayColorCorrector_CompositeKernel_param_28];
	ld.param.f32 	%f74, [ThreeWayColorCorrector_CompositeKernel_param_29];
	ld.param.f32 	%f75, [ThreeWayColorCorrector_CompositeKernel_param_30];
	ld.param.f32 	%f76, [ThreeWayColorCorrector_CompositeKernel_param_31];
	ld.param.f32 	%f77, [ThreeWayColorCorrector_CompositeKernel_param_32];
	ld.param.f32 	%f78, [ThreeWayColorCorrector_CompositeKernel_param_33];
	ld.param.f32 	%f79, [ThreeWayColorCorrector_CompositeKernel_param_34];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_CompositeKernel_param_35];
	cvta.to.global.u64 	%rd1, %rd6;
	cvta.to.global.u64 	%rd2, %rd5;
	.loc 1 42 1
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	.loc 1 42 1
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	@!%p3 bra 	BB4_24;
	bra.uni 	BB4_1;

BB4_1:
	.loc 1 42 1
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd3, %rd2, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd4, %rd2, %rd8;
	.loc 1 42 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB4_3;

	ld.global.v4.f32 	{%f80, %f81, %f82, %f83}, [%rd3];
	mov.f32 	%f196, %f83;
	mov.f32 	%f195, %f82;
	mov.f32 	%f194, %f81;
	mov.f32 	%f193, %f80;
	bra.uni 	BB4_4;

BB4_3:
	.loc 1 42 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f193, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f194, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f195, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f196, %temp;
	}

BB4_4:
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f85, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f86, %f194, %f85;
	fma.rn.ftz.f32 	%f87, %f195, %f84, %f86;
	ld.const.f32 	%f88, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f193, %f88, %f87;
	ld.const.f32 	%f89, [kRGB32f_To_601YPbPr+12];
	ld.const.f32 	%f90, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f91, %f194, %f90;
	fma.rn.ftz.f32 	%f92, %f195, %f89, %f91;
	ld.const.f32 	%f93, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f15, %f193, %f93, %f92;
	ld.const.f32 	%f94, [kRGB32f_To_601YPbPr+24];
	ld.const.f32 	%f95, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f96, %f194, %f95;
	fma.rn.ftz.f32 	%f97, %f195, %f94, %f96;
	ld.const.f32 	%f98, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f16, %f193, %f98, %f97;
	.loc 1 42 1
	setp.ne.s64	%p5, %rd6, 0;
	@%p5 bra 	BB4_6;

	mov.f32 	%f197, 0f3F800000;
	bra.uni 	BB4_7;

BB4_6:
	.loc 1 42 1
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 42 1
	ld.global.f32 	%f197, [%rd10];

BB4_7:
	add.ftz.f32 	%f100, %f52, %f53;
	mov.f32 	%f101, 0f3F800000;
	.loc 2 2765 10
	min.ftz.f32 	%f19, %f101, %f100;
	.loc 1 42 1
	sub.ftz.f32 	%f102, %f54, %f55;
	mov.f32 	%f103, 0f00000000;
	.loc 2 2770 10
	max.ftz.f32 	%f20, %f103, %f102;
	.loc 3 8366 10
	cos.approx.ftz.f32 	%f104, %f60;
	.loc 1 42 1
	mul.ftz.f32 	%f105, %f15, %f104;
	.loc 3 8357 10
	sin.approx.ftz.f32 	%f106, %f60;
	.loc 1 42 1
	mul.ftz.f32 	%f107, %f16, %f106;
	sub.ftz.f32 	%f108, %f105, %f107;
	mul.ftz.f32 	%f109, %f15, %f106;
	fma.rn.ftz.f32 	%f110, %f16, %f104, %f109;
	.loc 3 8366 10
	cos.approx.ftz.f32 	%f111, %f61;
	.loc 1 42 1
	mul.ftz.f32 	%f112, %f15, %f111;
	.loc 3 8357 10
	sin.approx.ftz.f32 	%f113, %f61;
	.loc 1 42 1
	mul.ftz.f32 	%f114, %f16, %f113;
	sub.ftz.f32 	%f115, %f112, %f114;
	mul.ftz.f32 	%f116, %f15, %f113;
	fma.rn.ftz.f32 	%f117, %f16, %f111, %f116;
	.loc 3 8366 10
	cos.approx.ftz.f32 	%f118, %f62;
	.loc 1 42 1
	mul.ftz.f32 	%f119, %f15, %f118;
	.loc 3 8357 10
	sin.approx.ftz.f32 	%f120, %f62;
	.loc 1 42 1
	mul.ftz.f32 	%f121, %f16, %f120;
	sub.ftz.f32 	%f122, %f119, %f121;
	mul.ftz.f32 	%f123, %f15, %f120;
	fma.rn.ftz.f32 	%f124, %f16, %f118, %f123;
	sub.ftz.f32 	%f125, %f63, %f15;
	fma.rn.ftz.f32 	%f126, %f125, %f65, %f15;
	sub.ftz.f32 	%f127, %f64, %f16;
	fma.rn.ftz.f32 	%f128, %f127, %f65, %f16;
	sub.ftz.f32 	%f129, %f66, %f108;
	fma.rn.ftz.f32 	%f130, %f129, %f68, %f108;
	sub.ftz.f32 	%f131, %f67, %f110;
	fma.rn.ftz.f32 	%f132, %f131, %f68, %f110;
	sub.ftz.f32 	%f133, %f69, %f115;
	fma.rn.ftz.f32 	%f134, %f133, %f71, %f115;
	sub.ftz.f32 	%f135, %f70, %f117;
	fma.rn.ftz.f32 	%f136, %f135, %f71, %f117;
	sub.ftz.f32 	%f137, %f72, %f122;
	fma.rn.ftz.f32 	%f138, %f137, %f74, %f122;
	sub.ftz.f32 	%f139, %f73, %f124;
	fma.rn.ftz.f32 	%f140, %f139, %f74, %f124;
	mul.ftz.f32 	%f21, %f126, %f56;
	mul.ftz.f32 	%f22, %f128, %f56;
	mul.ftz.f32 	%f23, %f130, %f57;
	mul.ftz.f32 	%f24, %f132, %f57;
	mul.ftz.f32 	%f25, %f134, %f58;
	mul.ftz.f32 	%f26, %f136, %f58;
	mul.ftz.f32 	%f27, %f138, %f59;
	mul.ftz.f32 	%f28, %f140, %f59;
	.loc 1 42 1
	setp.le.ftz.f32	%p6, %f14, %f52;
	mov.f32 	%f201, %f24;
	mov.f32 	%f205, %f23;
	@%p6 bra 	BB4_16;

	setp.eq.ftz.f32	%p7, %f53, 0f00000000;
	setp.le.ftz.f32	%p8, %f14, %f19;
	and.pred  	%p9, %p8, %p7;
	.loc 1 42 1
	mov.f32 	%f198, %f24;
	mov.f32 	%f201, %f198;
	mov.f32 	%f202, %f23;
	mov.f32 	%f205, %f202;
	@%p9 bra 	BB4_16;

	.loc 1 42 1
	setp.ge.ftz.f32	%p10, %f14, %f54;
	mov.f32 	%f201, %f28;
	mov.f32 	%f205, %f27;
	@%p10 bra 	BB4_16;

	setp.eq.ftz.f32	%p11, %f55, 0f00000000;
	setp.ge.ftz.f32	%p12, %f14, %f20;
	and.pred  	%p13, %p12, %p11;
	.loc 1 42 1
	mov.f32 	%f199, %f28;
	mov.f32 	%f201, %f199;
	mov.f32 	%f203, %f27;
	mov.f32 	%f205, %f203;
	@%p13 bra 	BB4_16;

	.loc 1 42 1
	setp.gt.ftz.f32	%p14, %f14, %f19;
	setp.lt.ftz.f32	%p15, %f14, %f20;
	and.pred  	%p16, %p14, %p15;
	.loc 1 42 1
	mov.f32 	%f201, %f26;
	mov.f32 	%f205, %f25;
	@%p16 bra 	BB4_16;

	.loc 1 42 1
	setp.gt.ftz.f32	%p17, %f14, %f52;
	setp.lt.ftz.f32	%p18, %f14, %f19;
	and.pred  	%p19, %p17, %p18;
	.loc 1 42 1
	@%p19 bra 	BB4_15;

	.loc 1 42 1
	setp.leu.ftz.f32	%p20, %f14, %f20;
	mov.f32 	%f200, %f26;
	mov.f32 	%f201, %f200;
	mov.f32 	%f204, %f25;
	mov.f32 	%f205, %f204;
	@%p20 bra 	BB4_16;

	.loc 1 42 1
	sub.ftz.f32 	%f141, %f54, %f20;
	sub.ftz.f32 	%f142, %f14, %f20;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f143, %f142, %f141;
	.loc 1 42 1
	sub.ftz.f32 	%f144, %f27, %f25;
	fma.rn.ftz.f32 	%f205, %f143, %f144, %f25;
	sub.ftz.f32 	%f145, %f28, %f26;
	fma.rn.ftz.f32 	%f201, %f143, %f145, %f26;
	bra.uni 	BB4_16;

BB4_15:
	.loc 1 42 1
	sub.ftz.f32 	%f146, %f19, %f52;
	sub.ftz.f32 	%f147, %f14, %f52;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f148, %f147, %f146;
	.loc 1 42 1
	sub.ftz.f32 	%f149, %f25, %f23;
	fma.rn.ftz.f32 	%f205, %f148, %f149, %f23;
	sub.ftz.f32 	%f150, %f26, %f24;
	fma.rn.ftz.f32 	%f201, %f148, %f150, %f24;

BB4_16:
	.loc 1 42 1
	sub.ftz.f32 	%f151, %f205, %f21;
	fma.rn.ftz.f32 	%f152, %f151, 0f3F000000, %f21;
	sub.ftz.f32 	%f153, %f201, %f22;
	fma.rn.ftz.f32 	%f154, %f153, 0f3F000000, %f22;
	sub.ftz.f32 	%f155, %f152, %f15;
	fma.rn.ftz.f32 	%f35, %f197, %f155, %f15;
	sub.ftz.f32 	%f156, %f154, %f16;
	fma.rn.ftz.f32 	%f36, %f197, %f156, %f16;
	.loc 1 42 1
	sub.ftz.f32 	%f37, %f76, %f75;
	.loc 1 42 1
	setp.neu.ftz.f32	%p21, %f37, 0f3F800000;
	.loc 1 42 1
	sub.ftz.f32 	%f38, %f79, %f78;
	.loc 1 42 1
	setp.neu.ftz.f32	%p22, %f38, 0f3F800000;
	or.pred  	%p23, %p21, %p22;
	.loc 1 42 1
	mov.f32 	%f206, %f14;
	@!%p23 bra 	BB4_18;
	bra.uni 	BB4_17;

BB4_17:
	.loc 1 42 1
	sub.s32 	%r25, %r1, %r2;
	add.s32 	%r26, %r25, -1;
	mov.u32 	%r27, 1;
	sub.s32 	%r28, %r27, %r25;
	.loc 1 42 1
	add.s32 	%r29, %r2, %r28;
	add.s32 	%r30, %r29, -1;
	shl.b32 	%r31, %r26, 8;
	xor.b32  	%r32, %r30, %r31;
	mov.u32 	%r33, 2;
	.loc 1 42 1
	sub.s32 	%r34, %r33, %r25;
	sub.s32 	%r35, %r34, %r32;
	shr.u32 	%r36, %r32, 13;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r26, %r32;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 12;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r32, %r37;
	sub.s32 	%r43, %r42, %r41;
	shl.b32 	%r44, %r41, 16;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shr.u32 	%r48, %r45, 5;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 3;
	xor.b32  	%r53, %r51, %r52;
	sub.s32 	%r54, %r45, %r49;
	sub.s32 	%r55, %r54, %r53;
	shl.b32 	%r56, %r53, 10;
	xor.b32  	%r57, %r55, %r56;
	sub.s32 	%r58, %r49, %r53;
	sub.s32 	%r59, %r58, %r57;
	shr.u32 	%r60, %r57, 15;
	xor.b32  	%r61, %r59, %r60;
	mad.lo.s32 	%r62, %r61, 1103515245, 12345;
	mad.lo.s32 	%r63, %r62, 1103515245, 12345;
	shr.u32 	%r64, %r62, 9;
	and.b32  	%r65, %r64, 32640;
	shr.u32 	%r66, %r63, 16;
	and.b32  	%r67, %r66, 255;
	xor.b32  	%r68, %r67, %r65;
	cvt.rn.f32.s32	%f157, %r68;
	ld.const.u32 	%r69, [kRandMax];
	cvt.rn.f32.u32	%f158, %r69;
	mov.f32 	%f159, 0f40000000;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f160, %f158, %f159;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f161, %f157, %f160;
	.loc 1 42 60
	add.ftz.f32 	%f162, %f161, 0fBF800000;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f39, %f162, 0f3B270D73, %f14;
	mov.f32 	%f206, %f39;

BB4_18:
	.loc 1 42 1
	mov.f32 	%f40, %f206;
	sub.ftz.f32 	%f163, %f40, %f75;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f164, %f163, %f37;
	.loc 2 2770 10
	max.ftz.f32 	%f41, %f103, %f164;
	setp.ltu.ftz.f32	%p24, %f41, 0f00000000;
	@%p24 bra 	BB4_20;

	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f166, %f41;
	mul.ftz.f32 	%f167, %f166, %f77;
	ex2.approx.ftz.f32 	%f207, %f167;
	bra.uni 	BB4_21;

BB4_20:
	neg.ftz.f32 	%f168, %f41;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f169, %f168;
	mul.ftz.f32 	%f170, %f169, %f77;
	ex2.approx.ftz.f32 	%f171, %f170;
	neg.ftz.f32 	%f207, %f171;

BB4_21:
	fma.rn.ftz.f32 	%f172, %f38, %f207, %f78;
	.loc 1 42 1
	sub.ftz.f32 	%f173, %f40, %f172;
	fma.rn.ftz.f32 	%f174, %f173, 0f3F000000, %f172;
	.loc 1 42 1
	setp.eq.s32	%p25, %r6, 0;
	selp.f32	%f175, %f172, %f174, %p25;
	.loc 1 42 1
	sub.ftz.f32 	%f176, %f175, %f14;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f177, %f197, %f176, %f14;
	.loc 1 42 1
	ld.const.f32 	%f178, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f179, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f180, %f35, %f179;
	fma.rn.ftz.f32 	%f181, %f177, %f178, %f180;
	ld.const.f32 	%f182, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f45, %f36, %f182, %f181;
	ld.const.f32 	%f183, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f184, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f185, %f35, %f184;
	fma.rn.ftz.f32 	%f186, %f177, %f183, %f185;
	ld.const.f32 	%f187, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f46, %f36, %f187, %f186;
	ld.const.f32 	%f188, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f189, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f190, %f35, %f189;
	fma.rn.ftz.f32 	%f191, %f177, %f188, %f190;
	ld.const.f32 	%f192, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f47, %f36, %f192, %f191;
	.loc 1 42 1
	@%p4 bra 	BB4_23;

	mul.wide.s32 	%rd12, %r15, 16;
	add.s64 	%rd13, %rd2, %rd12;
	.loc 1 42 1
	st.global.v4.f32 	[%rd13], {%f47, %f46, %f45, %f196};
	bra.uni 	BB4_24;

BB4_23:
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f47;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f46;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f45;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f196;
	mov.b16 	%rs12, %temp;
}
	mul.wide.s32 	%rd15, %r15, 8;
	add.s64 	%rd16, %rd2, %rd15;
	.loc 1 42 241
	st.global.v4.u16 	[%rd16], {%rs9, %rs10, %rs11, %rs12};

BB4_24:
	.loc 1 42 2
	ret;
}

.visible .entry ThreeWayColorCorrector_MasterLevelsKernel(
	.param .u64 ThreeWayColorCorrector_MasterLevelsKernel_param_0,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_1,
	.param .u64 ThreeWayColorCorrector_MasterLevelsKernel_param_2,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_3,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_4,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_5,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_6,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_7,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_8,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_9,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_10,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_11,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_12
)
{
	.reg .pred 	%p<12>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<62>;
	.reg .f32 	%f<103>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd6, [ThreeWayColorCorrector_MasterLevelsKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_MasterLevelsKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_MasterLevelsKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_MasterLevelsKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_MasterLevelsKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_MasterLevelsKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_MasterLevelsKernel_param_6];
	ld.param.f32 	%f33, [ThreeWayColorCorrector_MasterLevelsKernel_param_7];
	ld.param.f32 	%f34, [ThreeWayColorCorrector_MasterLevelsKernel_param_8];
	ld.param.f32 	%f35, [ThreeWayColorCorrector_MasterLevelsKernel_param_9];
	ld.param.f32 	%f36, [ThreeWayColorCorrector_MasterLevelsKernel_param_10];
	ld.param.f32 	%f37, [ThreeWayColorCorrector_MasterLevelsKernel_param_11];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_MasterLevelsKernel_param_12];
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd2, %rd6;
	.loc 1 42 1
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	.loc 1 42 1
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	.loc 1 42 1
	@!%p3 bra 	BB5_15;
	bra.uni 	BB5_1;

BB5_1:
	.loc 1 42 1
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd3, %rd2, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd4, %rd2, %rd8;
	.loc 1 42 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB5_3;

	ld.global.v4.f32 	{%f38, %f39, %f40, %f41}, [%rd3];
	mov.f32 	%f99, %f41;
	mov.f32 	%f98, %f40;
	mov.f32 	%f97, %f39;
	mov.f32 	%f96, %f38;
	bra.uni 	BB5_4;

BB5_3:
	.loc 1 42 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f96, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f97, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f98, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f99, %temp;
	}

BB5_4:
	ld.const.f32 	%f42, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f43, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f44, %f97, %f43;
	fma.rn.ftz.f32 	%f45, %f98, %f42, %f44;
	ld.const.f32 	%f46, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f96, %f46, %f45;
	ld.const.f32 	%f47, [kRGB32f_To_601YPbPr+12];
	ld.const.f32 	%f48, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f49, %f97, %f48;
	fma.rn.ftz.f32 	%f50, %f98, %f47, %f49;
	ld.const.f32 	%f51, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f15, %f96, %f51, %f50;
	ld.const.f32 	%f52, [kRGB32f_To_601YPbPr+24];
	ld.const.f32 	%f53, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f54, %f97, %f53;
	fma.rn.ftz.f32 	%f55, %f98, %f52, %f54;
	ld.const.f32 	%f56, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f16, %f96, %f56, %f55;
	.loc 1 42 1
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB5_6;

	mov.f32 	%f100, 0f3F800000;
	bra.uni 	BB5_7;

BB5_6:
	.loc 1 42 1
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 42 1
	ld.global.f32 	%f100, [%rd10];

BB5_7:
	.loc 1 42 1
	sub.ftz.f32 	%f58, %f34, %f33;
	.loc 1 42 1
	setp.neu.ftz.f32	%p6, %f58, 0f3F800000;
	.loc 1 42 1
	sub.ftz.f32 	%f19, %f37, %f36;
	.loc 1 42 1
	setp.neu.ftz.f32	%p7, %f19, 0f3F800000;
	or.pred  	%p8, %p6, %p7;
	.loc 1 42 1
	mov.f32 	%f101, %f14;
	@!%p8 bra 	BB5_9;
	bra.uni 	BB5_8;

BB5_8:
	.loc 1 42 1
	sub.s32 	%r17, %r1, %r2;
	add.s32 	%r18, %r17, -1;
	mov.u32 	%r19, 1;
	sub.s32 	%r20, %r19, %r17;
	.loc 1 42 1
	add.s32 	%r21, %r2, %r20;
	add.s32 	%r22, %r21, -1;
	shl.b32 	%r23, %r18, 8;
	xor.b32  	%r24, %r22, %r23;
	mov.u32 	%r25, 2;
	.loc 1 42 1
	sub.s32 	%r26, %r25, %r17;
	sub.s32 	%r27, %r26, %r24;
	shr.u32 	%r28, %r24, 13;
	xor.b32  	%r29, %r27, %r28;
	sub.s32 	%r30, %r18, %r24;
	sub.s32 	%r31, %r30, %r29;
	shr.u32 	%r32, %r29, 12;
	xor.b32  	%r33, %r31, %r32;
	sub.s32 	%r34, %r24, %r29;
	sub.s32 	%r35, %r34, %r33;
	shl.b32 	%r36, %r33, 16;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r29, %r33;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 5;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r33, %r37;
	sub.s32 	%r43, %r42, %r41;
	shr.u32 	%r44, %r41, 3;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shl.b32 	%r48, %r45, 10;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 15;
	xor.b32  	%r53, %r51, %r52;
	mad.lo.s32 	%r54, %r53, 1103515245, 12345;
	mad.lo.s32 	%r55, %r54, 1103515245, 12345;
	shr.u32 	%r56, %r54, 9;
	and.b32  	%r57, %r56, 32640;
	shr.u32 	%r58, %r55, 16;
	and.b32  	%r59, %r58, 255;
	xor.b32  	%r60, %r59, %r57;
	cvt.rn.f32.s32	%f59, %r60;
	ld.const.u32 	%r61, [kRandMax];
	cvt.rn.f32.u32	%f60, %r61;
	mov.f32 	%f61, 0f40000000;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f62, %f60, %f61;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f63, %f59, %f62;
	.loc 1 42 60
	add.ftz.f32 	%f64, %f63, 0fBF800000;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f20, %f64, 0f3B270D73, %f14;
	mov.f32 	%f101, %f20;

BB5_9:
	.loc 1 42 1
	mov.f32 	%f21, %f101;
	sub.ftz.f32 	%f66, %f21, %f33;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f67, %f66, %f58;
	mov.f32 	%f68, 0f00000000;
	.loc 2 2770 10
	max.ftz.f32 	%f22, %f68, %f67;
	setp.ltu.ftz.f32	%p9, %f22, 0f00000000;
	@%p9 bra 	BB5_11;

	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f69, %f22;
	mul.ftz.f32 	%f70, %f69, %f35;
	ex2.approx.ftz.f32 	%f102, %f70;
	bra.uni 	BB5_12;

BB5_11:
	neg.ftz.f32 	%f71, %f22;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f72, %f71;
	mul.ftz.f32 	%f73, %f72, %f35;
	ex2.approx.ftz.f32 	%f74, %f73;
	neg.ftz.f32 	%f102, %f74;

BB5_12:
	fma.rn.ftz.f32 	%f75, %f19, %f102, %f36;
	.loc 1 42 1
	sub.ftz.f32 	%f76, %f21, %f75;
	fma.rn.ftz.f32 	%f77, %f76, 0f3F000000, %f75;
	.loc 1 42 1
	setp.eq.s32	%p10, %r6, 0;
	selp.f32	%f78, %f75, %f77, %p10;
	.loc 1 42 1
	sub.ftz.f32 	%f79, %f78, %f14;
	.loc 1 42 1
	fma.rn.ftz.f32 	%f80, %f100, %f79, %f14;
	.loc 1 42 1
	ld.const.f32 	%f81, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f82, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f83, %f15, %f82;
	fma.rn.ftz.f32 	%f84, %f80, %f81, %f83;
	ld.const.f32 	%f85, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f26, %f16, %f85, %f84;
	ld.const.f32 	%f86, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f88, %f15, %f87;
	fma.rn.ftz.f32 	%f89, %f80, %f86, %f88;
	ld.const.f32 	%f90, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f27, %f16, %f90, %f89;
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f92, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f93, %f15, %f92;
	fma.rn.ftz.f32 	%f94, %f80, %f91, %f93;
	ld.const.f32 	%f95, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f28, %f16, %f95, %f94;
	.loc 1 42 1
	@%p4 bra 	BB5_14;

	.loc 1 42 1
	st.global.v4.f32 	[%rd3], {%f28, %f27, %f26, %f99};
	bra.uni 	BB5_15;

BB5_14:
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f28;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f27;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f26;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f99;
	mov.b16 	%rs12, %temp;
}
	.loc 1 42 241
	st.global.v4.u16 	[%rd4], {%rs9, %rs10, %rs11, %rs12};

BB5_15:
	.loc 1 42 2
	ret;
}


