//
// 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/LumaCorrector.cu", 1399785316, 13457
	.file	2 "D:\\singlebarrel\\releases\\2014.03\\shared\\adobe\\MediaCore\\Renderers\\RendererGPU\\Inc\\Effects/ColorCorrector_Common.h", 1399785316, 2534
	.file	3 "d:\\singlebarrel\\releases\\2014.03\\shared\\adobe\\mediacore\\external\\3rdparty\\nvidia\\cuda\\win\\include\\device_functions.h", 1399785281, 191626
.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};
.global .align 1 .b8 $str[11] = {95, 95, 67, 85, 68, 65, 95, 70, 84, 90, 0};

.visible .func  (.param .b32 func_retval0) _Z22CalculateShadowsWeightfff(
	.param .b32 _Z22CalculateShadowsWeightfff_param_0,
	.param .b32 _Z22CalculateShadowsWeightfff_param_1,
	.param .b32 _Z22CalculateShadowsWeightfff_param_2
)
{
	.reg .pred 	%p<3>;
	.reg .f32 	%f<11>;


	ld.param.f32 	%f3, [_Z22CalculateShadowsWeightfff_param_0];
	ld.param.f32 	%f4, [_Z22CalculateShadowsWeightfff_param_1];
	ld.param.f32 	%f5, [_Z22CalculateShadowsWeightfff_param_2];
	.loc 2 35 1
	setp.gtu.ftz.f32	%p1, %f3, %f4;
	@%p1 bra 	BB0_2;

	mov.f32 	%f10, 0f3F800000;
	bra.uni 	BB0_5;

BB0_2:
	.loc 2 35 1
	add.ftz.f32 	%f7, %f4, %f5;
	setp.gtu.ftz.f32	%p2, %f7, %f3;
	@%p2 bra 	BB0_4;

	mov.f32 	%f10, 0f00000000;
	bra.uni 	BB0_5;

BB0_4:
	.loc 2 35 1
	sub.ftz.f32 	%f9, %f3, %f4;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f10, %f9, %f5;

BB0_5:
	st.param.f32	[func_retval0+0], %f10;
	.loc 2 35 8
	ret;
}

.visible .func  (.param .b32 func_retval0) _Z25CalculateHighlightsWeightfff(
	.param .b32 _Z25CalculateHighlightsWeightfff_param_0,
	.param .b32 _Z25CalculateHighlightsWeightfff_param_1,
	.param .b32 _Z25CalculateHighlightsWeightfff_param_2
)
{
	.reg .pred 	%p<3>;
	.reg .f32 	%f<11>;


	ld.param.f32 	%f4, [_Z25CalculateHighlightsWeightfff_param_0];
	ld.param.f32 	%f5, [_Z25CalculateHighlightsWeightfff_param_1];
	ld.param.f32 	%f6, [_Z25CalculateHighlightsWeightfff_param_2];
	.loc 2 35 1
	sub.ftz.f32 	%f1, %f5, %f6;
	setp.leu.ftz.f32	%p1, %f1, %f4;
	@%p1 bra 	BB1_2;

	mov.f32 	%f10, 0f00000000;
	bra.uni 	BB1_5;

BB1_2:
	.loc 2 35 1
	setp.leu.ftz.f32	%p2, %f4, %f5;
	@%p2 bra 	BB1_4;

	mov.f32 	%f10, 0f3F800000;
	bra.uni 	BB1_5;

BB1_4:
	.loc 2 35 1
	sub.ftz.f32 	%f7, %f4, %f1;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f10, %f7, %f6;

BB1_5:
	st.param.f32	[func_retval0+0], %f10;
	.loc 2 35 8
	ret;
}

.visible .func  (.param .b32 func_retval0) _Z23ApplyBrightnessContrastffff(
	.param .b32 _Z23ApplyBrightnessContrastffff_param_0,
	.param .b32 _Z23ApplyBrightnessContrastffff_param_1,
	.param .b32 _Z23ApplyBrightnessContrastffff_param_2,
	.param .b32 _Z23ApplyBrightnessContrastffff_param_3
)
{
	.reg .pred 	%p<6>;
	.reg .f32 	%f<26>;


	ld.param.f32 	%f8, [_Z23ApplyBrightnessContrastffff_param_0];
	ld.param.f32 	%f9, [_Z23ApplyBrightnessContrastffff_param_1];
	ld.param.f32 	%f10, [_Z23ApplyBrightnessContrastffff_param_2];
	ld.param.f32 	%f11, [_Z23ApplyBrightnessContrastffff_param_3];
	.loc 1 41 1
	setp.ltu.ftz.f32	%p1, %f10, 0f00000000;
	@%p1 bra 	BB2_6;

	.loc 1 41 1
	sub.ftz.f32 	%f1, %f11, %f9;
	mov.f32 	%f12, 0f3F800000;
	.loc 1 41 1
	sub.ftz.f32 	%f2, %f12, %f10;
	.loc 1 41 1
	setp.eq.ftz.f32	%p2, %f2, 0f00000000;
	@%p2 bra 	BB2_5;

	setp.gtu.ftz.f32	%p3, %f1, %f8;
	.loc 1 41 1
	@%p3 bra 	BB2_4;

	.loc 1 41 1
	sub.ftz.f32 	%f13, %f8, %f1;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f14, %f13, %f2;
	.loc 1 41 95
	add.ftz.f32 	%f25, %f14, %f11;
	bra.uni 	BB2_7;

BB2_4:
	.loc 1 41 1
	sub.ftz.f32 	%f15, %f1, %f8;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f16, %f15, %f2;
	.loc 1 41 95
	sub.ftz.f32 	%f25, %f11, %f16;
	bra.uni 	BB2_7;

BB2_5:
	setp.gtu.ftz.f32	%p4, %f1, %f8;
	.loc 1 41 1
	selp.f32	%f25, 0f00000000, 0f3F800000, %p4;
	bra.uni 	BB2_7;

BB2_6:
	.loc 1 41 1
	add.ftz.f32 	%f17, %f11, %f9;
	.loc 1 41 1
	sub.ftz.f32 	%f18, %f8, %f11;
	.loc 1 41 1
	add.ftz.f32 	%f19, %f10, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f20, %f19, %f18;
	.loc 1 41 1
	sub.ftz.f32 	%f21, %f11, %f8;
	mul.ftz.f32 	%f22, %f19, %f21;
	.loc 1 41 1
	neg.ftz.f32 	%f23, %f22;
	setp.ltu.ftz.f32	%p5, %f8, %f11;
	selp.f32	%f24, %f23, %f20, %p5;
	add.ftz.f32 	%f25, %f17, %f24;

BB2_7:
	st.param.f32	[func_retval0+0], %f25;
	.loc 1 41 1
	ret;
}

.visible .func  (.param .b32 func_retval0) _Z13CalculateLumaffffffffffffffffffffffffffffff(
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_0,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_1,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_2,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_3,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_4,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_5,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_6,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_7,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_8,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_9,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_10,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_11,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_12,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_13,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_14,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_15,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_16,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_17,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_18,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_19,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_20,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_21,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_22,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_23,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_24,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_25,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_26,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_27,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_28,
	.param .b32 _Z13CalculateLumaffffffffffffffffffffffffffffff_param_29
)
{
	.reg .pred 	%p<19>;
	.reg .f32 	%f<164>;


	ld.param.f32 	%f163, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_0];
	ld.param.f32 	%f42, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_1];
	ld.param.f32 	%f43, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_2];
	ld.param.f32 	%f44, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_3];
	ld.param.f32 	%f45, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_4];
	ld.param.f32 	%f46, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_5];
	ld.param.f32 	%f47, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_6];
	ld.param.f32 	%f48, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_7];
	ld.param.f32 	%f49, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_8];
	ld.param.f32 	%f50, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_9];
	ld.param.f32 	%f51, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_10];
	ld.param.f32 	%f52, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_11];
	ld.param.f32 	%f53, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_12];
	ld.param.f32 	%f54, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_13];
	ld.param.f32 	%f55, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_14];
	ld.param.f32 	%f56, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_15];
	ld.param.f32 	%f57, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_16];
	ld.param.f32 	%f58, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_17];
	ld.param.f32 	%f59, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_18];
	ld.param.f32 	%f60, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_19];
	ld.param.f32 	%f61, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_20];
	ld.param.f32 	%f62, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_21];
	ld.param.f32 	%f63, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_22];
	ld.param.f32 	%f64, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_23];
	ld.param.f32 	%f65, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_24];
	ld.param.f32 	%f66, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_25];
	ld.param.f32 	%f67, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_26];
	ld.param.f32 	%f68, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_27];
	ld.param.f32 	%f69, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_28];
	ld.param.f32 	%f70, [_Z13CalculateLumaffffffffffffffffffffffffffffff_param_29];
	.loc 1 41 1
	setp.leu.ftz.f32	%p1, %f42, 0f3A83126F;
	@%p1 bra 	BB3_32;

	.loc 1 41 1
	setp.gt.ftz.f32	%p2, %f44, %f46;
	selp.f32	%f1, 0f40000000, 0f3F800000, %p2;
	.loc 2 35 1
	setp.gtu.ftz.f32	%p3, %f163, %f43;
	@%p3 bra 	BB3_3;

	mov.f32 	%f157, 0f3F800000;
	bra.uni 	BB3_6;

BB3_3:
	.loc 2 35 1
	add.ftz.f32 	%f72, %f43, %f44;
	setp.gtu.ftz.f32	%p4, %f72, %f163;
	@%p4 bra 	BB3_5;

	mov.f32 	%f157, 0f00000000;
	bra.uni 	BB3_6;

BB3_5:
	.loc 2 35 1
	sub.ftz.f32 	%f74, %f163, %f43;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f157, %f74, %f44;

BB3_6:
	.loc 2 35 1
	sub.ftz.f32 	%f4, %f45, %f46;
	setp.leu.ftz.f32	%p5, %f4, %f163;
	@%p5 bra 	BB3_8;

	mov.f32 	%f158, 0f00000000;
	bra.uni 	BB3_11;

BB3_8:
	.loc 2 35 1
	setp.leu.ftz.f32	%p6, %f163, %f45;
	@%p6 bra 	BB3_10;

	mov.f32 	%f158, 0f3F800000;
	bra.uni 	BB3_11;

BB3_10:
	.loc 2 35 1
	sub.ftz.f32 	%f75, %f163, %f4;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f158, %f75, %f46;

BB3_11:
	.loc 1 41 1
	sub.ftz.f32 	%f78, %f1, %f157;
	sub.ftz.f32 	%f79, %f78, %f158;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f80, %f79, %f1;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f81, %f157, %f1;
	.loc 1 41 1
	mul.ftz.f32 	%f82, %f81, %f65;
	mul.ftz.f32 	%f83, %f80, %f59;
	mul.ftz.f32 	%f84, %f83, %f42;
	fma.rn.ftz.f32 	%f85, %f82, %f42, %f84;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f86, %f158, %f1;
	.loc 1 41 1
	mul.ftz.f32 	%f87, %f86, %f53;
	fma.rn.ftz.f32 	%f7, %f87, %f42, %f85;
	mul.ftz.f32 	%f88, %f81, %f66;
	mul.ftz.f32 	%f89, %f80, %f60;
	mul.ftz.f32 	%f90, %f89, %f42;
	fma.rn.ftz.f32 	%f91, %f88, %f42, %f90;
	mul.ftz.f32 	%f92, %f86, %f54;
	fma.rn.ftz.f32 	%f8, %f92, %f42, %f91;
	mul.ftz.f32 	%f93, %f80, %f61;
	fma.rn.ftz.f32 	%f94, %f81, %f67, %f93;
	fma.rn.ftz.f32 	%f95, %f86, %f55, %f94;
	add.ftz.f32 	%f96, %f95, 0fBF000000;
	fma.rn.ftz.f32 	%f9, %f96, %f42, 0f3F000000;
	mul.ftz.f32 	%f97, %f80, %f62;
	fma.rn.ftz.f32 	%f98, %f81, %f68, %f97;
	fma.rn.ftz.f32 	%f99, %f86, %f56, %f98;
	add.ftz.f32 	%f100, %f99, 0fBF800000;
	fma.rn.ftz.f32 	%f10, %f100, %f42, 0f3F800000;
	mul.ftz.f32 	%f101, %f81, %f69;
	mul.ftz.f32 	%f102, %f80, %f63;
	mul.ftz.f32 	%f103, %f102, %f42;
	fma.rn.ftz.f32 	%f104, %f101, %f42, %f103;
	mul.ftz.f32 	%f105, %f86, %f57;
	fma.rn.ftz.f32 	%f11, %f105, %f42, %f104;
	mul.ftz.f32 	%f106, %f80, %f64;
	fma.rn.ftz.f32 	%f107, %f81, %f70, %f106;
	fma.rn.ftz.f32 	%f108, %f86, %f58, %f107;
	add.ftz.f32 	%f109, %f108, 0fBF800000;
	fma.rn.ftz.f32 	%f12, %f109, %f42, 0f3F800000;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p7, %f8, 0f00000000;
	@%p7 bra 	BB3_17;

	.loc 1 41 1
	sub.ftz.f32 	%f13, %f9, %f7;
	mov.f32 	%f110, 0f3F800000;
	.loc 1 41 1
	sub.ftz.f32 	%f14, %f110, %f8;
	.loc 1 41 1
	setp.eq.ftz.f32	%p8, %f14, 0f00000000;
	@%p8 bra 	BB3_16;

	setp.gtu.ftz.f32	%p9, %f13, %f163;
	.loc 1 41 1
	@%p9 bra 	BB3_15;

	.loc 1 41 1
	sub.ftz.f32 	%f111, %f163, %f13;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f112, %f111, %f14;
	.loc 1 41 95
	add.ftz.f32 	%f159, %f9, %f112;
	bra.uni 	BB3_18;

BB3_15:
	.loc 1 41 1
	sub.ftz.f32 	%f113, %f13, %f163;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f114, %f113, %f14;
	.loc 1 41 95
	sub.ftz.f32 	%f159, %f9, %f114;
	bra.uni 	BB3_18;

BB3_16:
	setp.gtu.ftz.f32	%p10, %f13, %f163;
	.loc 1 41 1
	selp.f32	%f159, 0f00000000, 0f3F800000, %p10;
	bra.uni 	BB3_18;

BB3_17:
	.loc 1 41 1
	add.ftz.f32 	%f115, %f9, %f7;
	.loc 1 41 1
	sub.ftz.f32 	%f116, %f163, %f9;
	.loc 1 41 1
	add.ftz.f32 	%f117, %f8, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f118, %f117, %f116;
	.loc 1 41 1
	sub.ftz.f32 	%f119, %f9, %f163;
	mul.ftz.f32 	%f120, %f117, %f119;
	.loc 1 41 1
	neg.ftz.f32 	%f121, %f120;
	setp.gtu.ftz.f32	%p11, %f9, %f163;
	selp.f32	%f122, %f121, %f118, %p11;
	add.ftz.f32 	%f159, %f115, %f122;

BB3_18:
	mov.f32 	%f123, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f20, %f123, %f10;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p12, %f159, 0f00000000;
	@%p12 bra 	BB3_20;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f124, %f159;
	mul.ftz.f32 	%f125, %f20, %f124;
	ex2.approx.ftz.f32 	%f160, %f125;
	bra.uni 	BB3_21;

BB3_20:
	.loc 1 41 216
	neg.ftz.f32 	%f126, %f159;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f127, %f126;
	mul.ftz.f32 	%f128, %f20, %f127;
	ex2.approx.ftz.f32 	%f129, %f128;
	.loc 1 41 243
	neg.ftz.f32 	%f160, %f129;

BB3_21:
	.loc 1 41 1
	add.ftz.f32 	%f130, %f160, %f11;
	mul.ftz.f32 	%f24, %f130, %f12;
	add.ftz.f32 	%f131, %f49, 0fBF000000;
	fma.rn.ftz.f32 	%f25, %f131, %f42, 0f3F000000;
	mul.ftz.f32 	%f26, %f47, %f42;
	mul.ftz.f32 	%f27, %f48, %f42;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p13, %f27, 0f00000000;
	@%p13 bra 	BB3_27;

	.loc 1 41 1
	sub.ftz.f32 	%f28, %f25, %f26;
	sub.ftz.f32 	%f29, %f123, %f27;
	.loc 1 41 1
	setp.eq.ftz.f32	%p14, %f29, 0f00000000;
	@%p14 bra 	BB3_26;

	setp.ltu.ftz.f32	%p15, %f24, %f28;
	.loc 1 41 1
	@%p15 bra 	BB3_25;

	.loc 1 41 1
	sub.ftz.f32 	%f133, %f24, %f28;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f134, %f133, %f29;
	.loc 1 41 95
	add.ftz.f32 	%f161, %f25, %f134;
	bra.uni 	BB3_28;

BB3_25:
	.loc 1 41 1
	sub.ftz.f32 	%f135, %f28, %f24;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f136, %f135, %f29;
	.loc 1 41 95
	sub.ftz.f32 	%f161, %f25, %f136;
	bra.uni 	BB3_28;

BB3_26:
	setp.ltu.ftz.f32	%p16, %f24, %f28;
	.loc 1 41 1
	selp.f32	%f161, 0f00000000, 0f3F800000, %p16;
	bra.uni 	BB3_28;

BB3_27:
	.loc 1 41 1
	add.ftz.f32 	%f137, %f25, %f26;
	.loc 1 41 1
	sub.ftz.f32 	%f138, %f24, %f25;
	.loc 1 41 1
	add.ftz.f32 	%f139, %f27, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f140, %f139, %f138;
	.loc 1 41 1
	sub.ftz.f32 	%f141, %f25, %f24;
	mul.ftz.f32 	%f142, %f139, %f141;
	.loc 1 41 1
	neg.ftz.f32 	%f143, %f142;
	setp.ltu.ftz.f32	%p17, %f24, %f25;
	selp.f32	%f144, %f143, %f140, %p17;
	add.ftz.f32 	%f161, %f137, %f144;

BB3_28:
	.loc 1 41 1
	add.ftz.f32 	%f145, %f50, 0fBF800000;
	fma.rn.ftz.f32 	%f146, %f145, %f42, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f35, %f123, %f146;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p18, %f161, 0f00000000;
	@%p18 bra 	BB3_30;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f148, %f161;
	mul.ftz.f32 	%f149, %f35, %f148;
	ex2.approx.ftz.f32 	%f162, %f149;
	bra.uni 	BB3_31;

BB3_30:
	.loc 1 41 219
	neg.ftz.f32 	%f150, %f161;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f151, %f150;
	mul.ftz.f32 	%f152, %f35, %f151;
	ex2.approx.ftz.f32 	%f153, %f152;
	.loc 1 41 246
	neg.ftz.f32 	%f162, %f153;

BB3_31:
	.loc 1 41 1
	fma.rn.ftz.f32 	%f154, %f51, %f42, %f162;
	add.ftz.f32 	%f155, %f52, 0fBF800000;
	fma.rn.ftz.f32 	%f156, %f155, %f42, 0f3F800000;
	mul.ftz.f32 	%f163, %f154, %f156;

BB3_32:
	st.param.f32	[func_retval0+0], %f163;
	.loc 1 41 1
	ret;
}

.visible .entry LumaCorrector_MaskKernel(
	.param .u64 LumaCorrector_MaskKernel_param_0,
	.param .u32 LumaCorrector_MaskKernel_param_1,
	.param .u64 LumaCorrector_MaskKernel_param_2,
	.param .u32 LumaCorrector_MaskKernel_param_3,
	.param .u32 LumaCorrector_MaskKernel_param_4,
	.param .u32 LumaCorrector_MaskKernel_param_5,
	.param .u32 LumaCorrector_MaskKernel_param_6
)
{
	.reg .pred 	%p<6>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<16>;
	.reg .f32 	%f<28>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd4, [LumaCorrector_MaskKernel_param_0];
	ld.param.u32 	%r4, [LumaCorrector_MaskKernel_param_1];
	ld.param.u64 	%rd3, [LumaCorrector_MaskKernel_param_2];
	ld.param.u32 	%r5, [LumaCorrector_MaskKernel_param_3];
	ld.param.u32 	%r6, [LumaCorrector_MaskKernel_param_4];
	ld.param.u32 	%r7, [LumaCorrector_MaskKernel_param_5];
	ld.param.u32 	%r8, [LumaCorrector_MaskKernel_param_6];
	cvta.to.global.u64 	%rd1, %rd3;
	cvta.to.global.u64 	%rd2, %rd4;
	.loc 1 41 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 41 1
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	.loc 1 41 1
	@!%p3 bra 	BB4_7;
	bra.uni 	BB4_1;

BB4_1:
	.loc 1 41 1
	setp.ne.s64	%p4, %rd3, 0;
	@%p4 bra 	BB4_3;

	mov.f32 	%f27, 0f3F800000;
	bra.uni 	BB4_4;

BB4_3:
	.loc 1 41 1
	mad.lo.s32 	%r15, %r2, %r5, %r1;
	mul.wide.s32 	%rd5, %r15, 4;
	add.s64 	%rd6, %rd1, %rd5;
	.loc 1 41 1
	ld.global.f32 	%f27, [%rd6];

BB4_4:
	.loc 1 41 1
	ld.const.f32 	%f11, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f12, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f13, %f12, 0f00000000;
	fma.rn.ftz.f32 	%f14, %f27, %f11, %f13;
	ld.const.f32 	%f15, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f3, %f15, 0f00000000, %f14;
	ld.const.f32 	%f16, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f17, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f18, %f17, 0f00000000;
	fma.rn.ftz.f32 	%f19, %f27, %f16, %f18;
	ld.const.f32 	%f20, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f4, %f20, 0f00000000, %f19;
	ld.const.f32 	%f21, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f22, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f23, %f22, 0f00000000;
	fma.rn.ftz.f32 	%f24, %f27, %f21, %f23;
	ld.const.f32 	%f25, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f5, %f25, 0f00000000, %f24;
	mov.f32 	%f9, 0f3F800000;
	.loc 1 41 1
	mad.lo.s32 	%r3, %r2, %r4, %r1;
	.loc 1 41 1
	setp.eq.s32	%p5, %r6, 0;
	@%p5 bra 	BB4_6;

	mul.wide.s32 	%rd7, %r3, 16;
	add.s64 	%rd8, %rd2, %rd7;
	.loc 1 41 1
	st.global.v4.f32 	[%rd8], {%f5, %f4, %f3, %f9};
	bra.uni 	BB4_7;

BB4_6:
	mul.wide.s32 	%rd9, %r3, 8;
	add.s64 	%rd10, %rd2, %rd9;
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f5;
	mov.b16 	%rs1, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f4;
	mov.b16 	%rs2, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f3;
	mov.b16 	%rs3, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f9;
	mov.b16 	%rs4, %temp;
}
	.loc 1 41 231
	st.global.v4.u16 	[%rd10], {%rs1, %rs2, %rs3, %rs4};

BB4_7:
	.loc 1 41 2
	ret;
}

.visible .entry LumaCorrector_TonalRangeKernel(
	.param .u64 LumaCorrector_TonalRangeKernel_param_0,
	.param .u32 LumaCorrector_TonalRangeKernel_param_1,
	.param .u64 LumaCorrector_TonalRangeKernel_param_2,
	.param .u32 LumaCorrector_TonalRangeKernel_param_3,
	.param .u32 LumaCorrector_TonalRangeKernel_param_4,
	.param .u32 LumaCorrector_TonalRangeKernel_param_5,
	.param .u32 LumaCorrector_TonalRangeKernel_param_6,
	.param .f32 LumaCorrector_TonalRangeKernel_param_7,
	.param .f32 LumaCorrector_TonalRangeKernel_param_8,
	.param .f32 LumaCorrector_TonalRangeKernel_param_9,
	.param .f32 LumaCorrector_TonalRangeKernel_param_10
)
{
	.reg .pred 	%p<14>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<82>;
	.reg .s64 	%rd<7>;


	ld.param.u64 	%rd4, [LumaCorrector_TonalRangeKernel_param_0];
	ld.param.u32 	%r3, [LumaCorrector_TonalRangeKernel_param_1];
	ld.param.u32 	%r4, [LumaCorrector_TonalRangeKernel_param_4];
	ld.param.u32 	%r5, [LumaCorrector_TonalRangeKernel_param_5];
	ld.param.u32 	%r6, [LumaCorrector_TonalRangeKernel_param_6];
	ld.param.f32 	%f32, [LumaCorrector_TonalRangeKernel_param_7];
	ld.param.f32 	%f33, [LumaCorrector_TonalRangeKernel_param_8];
	ld.param.f32 	%f34, [LumaCorrector_TonalRangeKernel_param_9];
	ld.param.f32 	%f35, [LumaCorrector_TonalRangeKernel_param_10];
	cvta.to.global.u64 	%rd1, %rd4;
	.loc 1 41 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 41 1
	setp.lt.s32	%p1, %r1, %r5;
	setp.lt.s32	%p2, %r2, %r6;
	and.pred  	%p3, %p1, %p2;
	.loc 1 41 1
	@!%p3 bra 	BB5_22;
	bra.uni 	BB5_1;

BB5_1:
	.loc 1 41 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 41 1
	setp.eq.s32	%p4, %r4, 0;
	@%p4 bra 	BB5_3;

	ld.global.v4.f32 	{%f36, %f37, %f38, %f39}, [%rd2];
	mov.f32 	%f78, %f39;
	mov.f32 	%f77, %f38;
	mov.f32 	%f76, %f37;
	mov.f32 	%f75, %f36;
	bra.uni 	BB5_4;

BB5_3:
	.loc 1 41 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	.loc 3 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f75, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f76, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f77, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f78, %temp;
	}

BB5_4:
	ld.const.f32 	%f40, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f41, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f42, %f76, %f41;
	fma.rn.ftz.f32 	%f43, %f77, %f40, %f42;
	ld.const.f32 	%f44, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f75, %f44, %f43;
	.loc 1 41 1
	setp.gt.ftz.f32	%p5, %f33, %f35;
	selp.f32	%f15, 0f40000000, 0f3F800000, %p5;
	.loc 2 35 1
	setp.gtu.ftz.f32	%p6, %f14, %f32;
	@%p6 bra 	BB5_6;

	mov.f32 	%f79, 0f3F800000;
	bra.uni 	BB5_9;

BB5_6:
	.loc 2 35 1
	add.ftz.f32 	%f46, %f32, %f33;
	setp.ltu.ftz.f32	%p7, %f14, %f46;
	@%p7 bra 	BB5_8;

	mov.f32 	%f79, 0f00000000;
	bra.uni 	BB5_9;

BB5_8:
	.loc 2 35 1
	sub.ftz.f32 	%f48, %f14, %f32;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f79, %f48, %f33;

BB5_9:
	.loc 2 35 1
	sub.ftz.f32 	%f18, %f34, %f35;
	setp.geu.ftz.f32	%p8, %f14, %f18;
	@%p8 bra 	BB5_11;

	mov.f32 	%f80, 0f00000000;
	bra.uni 	BB5_14;

BB5_11:
	.loc 2 35 1
	setp.leu.ftz.f32	%p9, %f14, %f34;
	@%p9 bra 	BB5_13;

	mov.f32 	%f80, 0f3F800000;
	bra.uni 	BB5_14;

BB5_13:
	.loc 2 35 1
	sub.ftz.f32 	%f49, %f14, %f18;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f80, %f49, %f35;

BB5_14:
	.loc 1 41 1
	sub.ftz.f32 	%f52, %f15, %f79;
	sub.ftz.f32 	%f53, %f52, %f80;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f21, %f53, %f15;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f22, %f80, %f15;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f54, %f79, %f15;
	.loc 1 41 1
	setp.neu.ftz.f32	%p10, %f54, 0f3F800000;
	@%p10 bra 	BB5_16;

	mov.f32 	%f81, 0f00000000;
	bra.uni 	BB5_19;

BB5_16:
	.loc 1 41 1
	setp.neu.ftz.f32	%p11, %f21, 0f3F800000;
	@%p11 bra 	BB5_18;

	mov.f32 	%f81, 0f3F000000;
	bra.uni 	BB5_19;

BB5_18:
	.loc 1 41 1
	setp.eq.ftz.f32	%p12, %f22, 0f3F800000;
	.loc 1 41 1
	fma.rn.ftz.f32 	%f55, %f21, 0f3F000000, 0f00000000;
	add.ftz.f32 	%f56, %f22, 0f00000000;
	add.ftz.f32 	%f57, %f55, %f56;
	.loc 1 41 1
	selp.f32	%f81, 0f3F800000, %f57, %p12;

BB5_19:
	.loc 1 41 1
	ld.const.f32 	%f60, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f61, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f62, %f61, 0f00000000;
	fma.rn.ftz.f32 	%f63, %f81, %f60, %f62;
	ld.const.f32 	%f64, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f25, %f64, 0f00000000, %f63;
	ld.const.f32 	%f65, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f66, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f67, %f66, 0f00000000;
	fma.rn.ftz.f32 	%f68, %f81, %f65, %f67;
	ld.const.f32 	%f69, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f26, %f69, 0f00000000, %f68;
	ld.const.f32 	%f70, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f71, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f72, %f71, 0f00000000;
	fma.rn.ftz.f32 	%f73, %f81, %f70, %f72;
	ld.const.f32 	%f74, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f27, %f74, 0f00000000, %f73;
	.loc 1 41 1
	@%p4 bra 	BB5_21;

	.loc 1 41 1
	st.global.v4.f32 	[%rd2], {%f27, %f26, %f25, %f78};
	bra.uni 	BB5_22;

BB5_21:
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f27;
	mov.b16 	%rs9, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f26;
	mov.b16 	%rs10, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f25;
	mov.b16 	%rs11, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f78;
	mov.b16 	%rs12, %temp;
}
	.loc 1 41 234
	st.global.v4.u16 	[%rd3], {%rs9, %rs10, %rs11, %rs12};

BB5_22:
	.loc 1 41 2
	ret;
}

.visible .entry LumaCorrector_LumaKernel(
	.param .u64 LumaCorrector_LumaKernel_param_0,
	.param .u32 LumaCorrector_LumaKernel_param_1,
	.param .u64 LumaCorrector_LumaKernel_param_2,
	.param .u32 LumaCorrector_LumaKernel_param_3,
	.param .u32 LumaCorrector_LumaKernel_param_4,
	.param .u32 LumaCorrector_LumaKernel_param_5,
	.param .u32 LumaCorrector_LumaKernel_param_6,
	.param .f32 LumaCorrector_LumaKernel_param_7,
	.param .f32 LumaCorrector_LumaKernel_param_8,
	.param .f32 LumaCorrector_LumaKernel_param_9,
	.param .f32 LumaCorrector_LumaKernel_param_10,
	.param .f32 LumaCorrector_LumaKernel_param_11,
	.param .f32 LumaCorrector_LumaKernel_param_12,
	.param .f32 LumaCorrector_LumaKernel_param_13,
	.param .f32 LumaCorrector_LumaKernel_param_14,
	.param .f32 LumaCorrector_LumaKernel_param_15,
	.param .f32 LumaCorrector_LumaKernel_param_16,
	.param .f32 LumaCorrector_LumaKernel_param_17,
	.param .f32 LumaCorrector_LumaKernel_param_18,
	.param .f32 LumaCorrector_LumaKernel_param_19,
	.param .f32 LumaCorrector_LumaKernel_param_20,
	.param .f32 LumaCorrector_LumaKernel_param_21,
	.param .f32 LumaCorrector_LumaKernel_param_22,
	.param .f32 LumaCorrector_LumaKernel_param_23,
	.param .f32 LumaCorrector_LumaKernel_param_24,
	.param .f32 LumaCorrector_LumaKernel_param_25,
	.param .f32 LumaCorrector_LumaKernel_param_26,
	.param .f32 LumaCorrector_LumaKernel_param_27,
	.param .f32 LumaCorrector_LumaKernel_param_28,
	.param .f32 LumaCorrector_LumaKernel_param_29,
	.param .f32 LumaCorrector_LumaKernel_param_30,
	.param .f32 LumaCorrector_LumaKernel_param_31,
	.param .f32 LumaCorrector_LumaKernel_param_32,
	.param .f32 LumaCorrector_LumaKernel_param_33,
	.param .f32 LumaCorrector_LumaKernel_param_34
)
{
	.reg .pred 	%p<25>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<16>;
	.reg .f32 	%f<214>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd6, [LumaCorrector_LumaKernel_param_0];
	ld.param.u32 	%r3, [LumaCorrector_LumaKernel_param_1];
	ld.param.u64 	%rd5, [LumaCorrector_LumaKernel_param_2];
	ld.param.u32 	%r4, [LumaCorrector_LumaKernel_param_3];
	ld.param.u32 	%r5, [LumaCorrector_LumaKernel_param_4];
	ld.param.u32 	%r6, [LumaCorrector_LumaKernel_param_5];
	ld.param.u32 	%r7, [LumaCorrector_LumaKernel_param_6];
	ld.param.f32 	%f63, [LumaCorrector_LumaKernel_param_7];
	ld.param.f32 	%f64, [LumaCorrector_LumaKernel_param_8];
	ld.param.f32 	%f65, [LumaCorrector_LumaKernel_param_9];
	ld.param.f32 	%f66, [LumaCorrector_LumaKernel_param_10];
	ld.param.f32 	%f67, [LumaCorrector_LumaKernel_param_11];
	ld.param.f32 	%f68, [LumaCorrector_LumaKernel_param_12];
	ld.param.f32 	%f69, [LumaCorrector_LumaKernel_param_13];
	ld.param.f32 	%f70, [LumaCorrector_LumaKernel_param_14];
	ld.param.f32 	%f71, [LumaCorrector_LumaKernel_param_15];
	ld.param.f32 	%f72, [LumaCorrector_LumaKernel_param_16];
	ld.param.f32 	%f73, [LumaCorrector_LumaKernel_param_17];
	ld.param.f32 	%f74, [LumaCorrector_LumaKernel_param_18];
	ld.param.f32 	%f75, [LumaCorrector_LumaKernel_param_19];
	ld.param.f32 	%f76, [LumaCorrector_LumaKernel_param_20];
	ld.param.f32 	%f77, [LumaCorrector_LumaKernel_param_21];
	ld.param.f32 	%f78, [LumaCorrector_LumaKernel_param_22];
	ld.param.f32 	%f79, [LumaCorrector_LumaKernel_param_23];
	ld.param.f32 	%f80, [LumaCorrector_LumaKernel_param_24];
	ld.param.f32 	%f81, [LumaCorrector_LumaKernel_param_25];
	ld.param.f32 	%f82, [LumaCorrector_LumaKernel_param_26];
	ld.param.f32 	%f83, [LumaCorrector_LumaKernel_param_27];
	ld.param.f32 	%f84, [LumaCorrector_LumaKernel_param_28];
	ld.param.f32 	%f85, [LumaCorrector_LumaKernel_param_29];
	ld.param.f32 	%f86, [LumaCorrector_LumaKernel_param_30];
	ld.param.f32 	%f87, [LumaCorrector_LumaKernel_param_31];
	ld.param.f32 	%f88, [LumaCorrector_LumaKernel_param_32];
	ld.param.f32 	%f89, [LumaCorrector_LumaKernel_param_33];
	ld.param.f32 	%f90, [LumaCorrector_LumaKernel_param_34];
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd2, %rd6;
	.loc 1 41 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 41 1
	setp.lt.s32	%p1, %r1, %r6;
	setp.lt.s32	%p2, %r2, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 41 1
	@!%p3 bra 	BB6_42;
	bra.uni 	BB6_1;

BB6_1:
	.loc 1 41 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 41 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB6_3;

	ld.global.v4.f32 	{%f91, %f92, %f93, %f94}, [%rd3];
	mov.f32 	%f4, %f94;
	mov.f32 	%f205, %f93;
	mov.f32 	%f204, %f92;
	mov.f32 	%f203, %f91;
	bra.uni 	BB6_4;

BB6_3:
	.loc 1 41 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 3 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f203, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f204, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f205, %temp;
	}

BB6_4:
	ld.const.f32 	%f95, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f96, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f97, %f204, %f96;
	fma.rn.ftz.f32 	%f98, %f205, %f95, %f97;
	ld.const.f32 	%f99, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f213, %f203, %f99, %f98;
	.loc 1 41 1
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB6_6;

	mov.f32 	%f206, 0f3F800000;
	bra.uni 	BB6_7;

BB6_6:
	.loc 1 41 1
	mad.lo.s32 	%r15, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r15, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 41 1
	ld.global.f32 	%f206, [%rd10];

BB6_7:
	.loc 1 41 1
	setp.leu.ftz.f32	%p6, %f206, 0f3A83126F;
	@%p6 bra 	BB6_39;

	.loc 1 41 1
	setp.gt.ftz.f32	%p7, %f64, %f66;
	selp.f32	%f16, 0f40000000, 0f3F800000, %p7;
	.loc 2 35 1
	setp.gtu.ftz.f32	%p8, %f213, %f63;
	@%p8 bra 	BB6_10;

	mov.f32 	%f207, 0f3F800000;
	bra.uni 	BB6_13;

BB6_10:
	.loc 2 35 1
	add.ftz.f32 	%f102, %f63, %f64;
	setp.ltu.ftz.f32	%p9, %f213, %f102;
	@%p9 bra 	BB6_12;

	mov.f32 	%f207, 0f00000000;
	bra.uni 	BB6_13;

BB6_12:
	.loc 2 35 1
	sub.ftz.f32 	%f104, %f213, %f63;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f207, %f104, %f64;

BB6_13:
	.loc 2 35 1
	sub.ftz.f32 	%f19, %f65, %f66;
	setp.geu.ftz.f32	%p10, %f213, %f19;
	@%p10 bra 	BB6_15;

	mov.f32 	%f208, 0f00000000;
	bra.uni 	BB6_18;

BB6_15:
	.loc 2 35 1
	setp.leu.ftz.f32	%p11, %f213, %f65;
	@%p11 bra 	BB6_17;

	mov.f32 	%f208, 0f3F800000;
	bra.uni 	BB6_18;

BB6_17:
	.loc 2 35 1
	sub.ftz.f32 	%f105, %f213, %f19;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f208, %f105, %f66;

BB6_18:
	.loc 1 41 1
	sub.ftz.f32 	%f108, %f16, %f207;
	sub.ftz.f32 	%f109, %f108, %f208;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f110, %f109, %f16;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f111, %f207, %f16;
	.loc 1 41 1
	mul.ftz.f32 	%f112, %f111, %f85;
	mul.ftz.f32 	%f113, %f110, %f79;
	mul.ftz.f32 	%f114, %f113, %f206;
	fma.rn.ftz.f32 	%f115, %f112, %f206, %f114;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f116, %f208, %f16;
	.loc 1 41 1
	mul.ftz.f32 	%f117, %f116, %f73;
	fma.rn.ftz.f32 	%f22, %f117, %f206, %f115;
	mul.ftz.f32 	%f118, %f111, %f86;
	mul.ftz.f32 	%f119, %f110, %f80;
	mul.ftz.f32 	%f120, %f119, %f206;
	fma.rn.ftz.f32 	%f121, %f118, %f206, %f120;
	mul.ftz.f32 	%f122, %f116, %f74;
	fma.rn.ftz.f32 	%f23, %f122, %f206, %f121;
	mul.ftz.f32 	%f123, %f110, %f81;
	fma.rn.ftz.f32 	%f124, %f111, %f87, %f123;
	fma.rn.ftz.f32 	%f125, %f116, %f75, %f124;
	add.ftz.f32 	%f126, %f125, 0fBF000000;
	fma.rn.ftz.f32 	%f24, %f206, %f126, 0f3F000000;
	mul.ftz.f32 	%f127, %f110, %f82;
	fma.rn.ftz.f32 	%f128, %f111, %f88, %f127;
	fma.rn.ftz.f32 	%f129, %f116, %f76, %f128;
	add.ftz.f32 	%f130, %f129, 0fBF800000;
	fma.rn.ftz.f32 	%f25, %f206, %f130, 0f3F800000;
	mul.ftz.f32 	%f131, %f111, %f89;
	mul.ftz.f32 	%f132, %f110, %f83;
	mul.ftz.f32 	%f133, %f132, %f206;
	fma.rn.ftz.f32 	%f134, %f131, %f206, %f133;
	mul.ftz.f32 	%f135, %f116, %f77;
	fma.rn.ftz.f32 	%f26, %f135, %f206, %f134;
	mul.ftz.f32 	%f136, %f110, %f84;
	fma.rn.ftz.f32 	%f137, %f111, %f90, %f136;
	fma.rn.ftz.f32 	%f138, %f116, %f78, %f137;
	add.ftz.f32 	%f139, %f138, 0fBF800000;
	fma.rn.ftz.f32 	%f27, %f206, %f139, 0f3F800000;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p12, %f23, 0f00000000;
	@%p12 bra 	BB6_24;

	.loc 1 41 1
	sub.ftz.f32 	%f28, %f24, %f22;
	mov.f32 	%f140, 0f3F800000;
	.loc 1 41 1
	sub.ftz.f32 	%f29, %f140, %f23;
	.loc 1 41 1
	setp.eq.ftz.f32	%p13, %f29, 0f00000000;
	@%p13 bra 	BB6_23;

	setp.ltu.ftz.f32	%p14, %f213, %f28;
	.loc 1 41 1
	@%p14 bra 	BB6_22;

	.loc 1 41 1
	sub.ftz.f32 	%f141, %f213, %f28;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f142, %f141, %f29;
	.loc 1 41 95
	add.ftz.f32 	%f209, %f24, %f142;
	bra.uni 	BB6_25;

BB6_22:
	.loc 1 41 1
	sub.ftz.f32 	%f143, %f28, %f213;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f144, %f143, %f29;
	.loc 1 41 95
	sub.ftz.f32 	%f209, %f24, %f144;
	bra.uni 	BB6_25;

BB6_23:
	setp.ltu.ftz.f32	%p15, %f213, %f28;
	.loc 1 41 1
	selp.f32	%f209, 0f00000000, 0f3F800000, %p15;
	bra.uni 	BB6_25;

BB6_24:
	.loc 1 41 1
	add.ftz.f32 	%f145, %f24, %f22;
	.loc 1 41 1
	sub.ftz.f32 	%f146, %f213, %f24;
	.loc 1 41 1
	add.ftz.f32 	%f147, %f23, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f148, %f147, %f146;
	.loc 1 41 1
	sub.ftz.f32 	%f149, %f24, %f213;
	mul.ftz.f32 	%f150, %f147, %f149;
	.loc 1 41 1
	neg.ftz.f32 	%f151, %f150;
	setp.ltu.ftz.f32	%p16, %f213, %f24;
	selp.f32	%f152, %f151, %f148, %p16;
	add.ftz.f32 	%f209, %f145, %f152;

BB6_25:
	mov.f32 	%f153, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f35, %f153, %f25;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p17, %f209, 0f00000000;
	@%p17 bra 	BB6_27;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f154, %f209;
	mul.ftz.f32 	%f155, %f35, %f154;
	ex2.approx.ftz.f32 	%f210, %f155;
	bra.uni 	BB6_28;

BB6_27:
	.loc 1 41 216
	neg.ftz.f32 	%f156, %f209;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f157, %f156;
	mul.ftz.f32 	%f158, %f35, %f157;
	ex2.approx.ftz.f32 	%f159, %f158;
	.loc 1 41 243
	neg.ftz.f32 	%f210, %f159;

BB6_28:
	.loc 1 41 1
	add.ftz.f32 	%f160, %f210, %f26;
	mul.ftz.f32 	%f39, %f160, %f27;
	add.ftz.f32 	%f161, %f69, 0fBF000000;
	fma.rn.ftz.f32 	%f40, %f206, %f161, 0f3F000000;
	mul.ftz.f32 	%f41, %f206, %f67;
	mul.ftz.f32 	%f42, %f206, %f68;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p18, %f42, 0f00000000;
	@%p18 bra 	BB6_34;

	.loc 1 41 1
	sub.ftz.f32 	%f43, %f40, %f41;
	sub.ftz.f32 	%f44, %f153, %f42;
	.loc 1 41 1
	setp.eq.ftz.f32	%p19, %f44, 0f00000000;
	@%p19 bra 	BB6_33;

	setp.ltu.ftz.f32	%p20, %f39, %f43;
	.loc 1 41 1
	@%p20 bra 	BB6_32;

	.loc 1 41 1
	sub.ftz.f32 	%f163, %f39, %f43;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f164, %f163, %f44;
	.loc 1 41 95
	add.ftz.f32 	%f211, %f40, %f164;
	bra.uni 	BB6_35;

BB6_32:
	.loc 1 41 1
	sub.ftz.f32 	%f165, %f43, %f39;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f166, %f165, %f44;
	.loc 1 41 95
	sub.ftz.f32 	%f211, %f40, %f166;
	bra.uni 	BB6_35;

BB6_33:
	setp.ltu.ftz.f32	%p21, %f39, %f43;
	.loc 1 41 1
	selp.f32	%f211, 0f00000000, 0f3F800000, %p21;
	bra.uni 	BB6_35;

BB6_34:
	.loc 1 41 1
	add.ftz.f32 	%f167, %f40, %f41;
	.loc 1 41 1
	sub.ftz.f32 	%f168, %f39, %f40;
	.loc 1 41 1
	add.ftz.f32 	%f169, %f42, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f170, %f169, %f168;
	.loc 1 41 1
	sub.ftz.f32 	%f171, %f40, %f39;
	mul.ftz.f32 	%f172, %f169, %f171;
	.loc 1 41 1
	neg.ftz.f32 	%f173, %f172;
	setp.ltu.ftz.f32	%p22, %f39, %f40;
	selp.f32	%f174, %f173, %f170, %p22;
	add.ftz.f32 	%f211, %f167, %f174;

BB6_35:
	.loc 1 41 1
	add.ftz.f32 	%f175, %f70, 0fBF800000;
	fma.rn.ftz.f32 	%f176, %f206, %f175, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f50, %f153, %f176;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p23, %f211, 0f00000000;
	@%p23 bra 	BB6_37;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f178, %f211;
	mul.ftz.f32 	%f179, %f50, %f178;
	ex2.approx.ftz.f32 	%f212, %f179;
	bra.uni 	BB6_38;

BB6_37:
	.loc 1 41 219
	neg.ftz.f32 	%f180, %f211;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f181, %f180;
	mul.ftz.f32 	%f182, %f50, %f181;
	ex2.approx.ftz.f32 	%f183, %f182;
	.loc 1 41 246
	neg.ftz.f32 	%f212, %f183;

BB6_38:
	.loc 1 41 1
	fma.rn.ftz.f32 	%f184, %f206, %f71, %f212;
	add.ftz.f32 	%f185, %f72, 0fBF800000;
	fma.rn.ftz.f32 	%f186, %f206, %f185, 0f3F800000;
	mul.ftz.f32 	%f213, %f184, %f186;

BB6_39:
	.loc 1 41 1
	ld.const.f32 	%f187, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f188, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f189, %f188, 0f00000000;
	fma.rn.ftz.f32 	%f190, %f213, %f187, %f189;
	ld.const.f32 	%f191, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f56, %f191, 0f00000000, %f190;
	ld.const.f32 	%f192, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f193, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f194, %f193, 0f00000000;
	fma.rn.ftz.f32 	%f195, %f213, %f192, %f194;
	ld.const.f32 	%f196, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f57, %f196, 0f00000000, %f195;
	ld.const.f32 	%f197, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f198, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f199, %f198, 0f00000000;
	fma.rn.ftz.f32 	%f200, %f213, %f197, %f199;
	ld.const.f32 	%f201, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f58, %f201, 0f00000000, %f200;
	mov.f32 	%f62, 0f3F800000;
	.loc 1 41 1
	@%p4 bra 	BB6_41;

	.loc 1 41 1
	st.global.v4.f32 	[%rd3], {%f58, %f57, %f56, %f62};
	bra.uni 	BB6_42;

BB6_41:
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f58;
	mov.b16 	%rs9, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f57;
	mov.b16 	%rs10, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f56;
	mov.b16 	%rs11, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f62;
	mov.b16 	%rs12, %temp;
}
	.loc 1 41 241
	st.global.v4.u16 	[%rd4], {%rs9, %rs10, %rs11, %rs12};

BB6_42:
	.loc 1 41 2
	ret;
}

.visible .entry LumaCorrector_CompositeKernel(
	.param .u64 LumaCorrector_CompositeKernel_param_0,
	.param .u32 LumaCorrector_CompositeKernel_param_1,
	.param .u64 LumaCorrector_CompositeKernel_param_2,
	.param .u32 LumaCorrector_CompositeKernel_param_3,
	.param .u32 LumaCorrector_CompositeKernel_param_4,
	.param .u32 LumaCorrector_CompositeKernel_param_5,
	.param .u32 LumaCorrector_CompositeKernel_param_6,
	.param .f32 LumaCorrector_CompositeKernel_param_7,
	.param .f32 LumaCorrector_CompositeKernel_param_8,
	.param .f32 LumaCorrector_CompositeKernel_param_9,
	.param .f32 LumaCorrector_CompositeKernel_param_10,
	.param .f32 LumaCorrector_CompositeKernel_param_11,
	.param .f32 LumaCorrector_CompositeKernel_param_12,
	.param .f32 LumaCorrector_CompositeKernel_param_13,
	.param .f32 LumaCorrector_CompositeKernel_param_14,
	.param .f32 LumaCorrector_CompositeKernel_param_15,
	.param .f32 LumaCorrector_CompositeKernel_param_16,
	.param .f32 LumaCorrector_CompositeKernel_param_17,
	.param .f32 LumaCorrector_CompositeKernel_param_18,
	.param .f32 LumaCorrector_CompositeKernel_param_19,
	.param .f32 LumaCorrector_CompositeKernel_param_20,
	.param .f32 LumaCorrector_CompositeKernel_param_21,
	.param .f32 LumaCorrector_CompositeKernel_param_22,
	.param .f32 LumaCorrector_CompositeKernel_param_23,
	.param .f32 LumaCorrector_CompositeKernel_param_24,
	.param .f32 LumaCorrector_CompositeKernel_param_25,
	.param .f32 LumaCorrector_CompositeKernel_param_26,
	.param .f32 LumaCorrector_CompositeKernel_param_27,
	.param .f32 LumaCorrector_CompositeKernel_param_28,
	.param .f32 LumaCorrector_CompositeKernel_param_29,
	.param .f32 LumaCorrector_CompositeKernel_param_30,
	.param .f32 LumaCorrector_CompositeKernel_param_31,
	.param .f32 LumaCorrector_CompositeKernel_param_32,
	.param .f32 LumaCorrector_CompositeKernel_param_33,
	.param .f32 LumaCorrector_CompositeKernel_param_34
)
{
	.reg .pred 	%p<25>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<34>;
	.reg .f32 	%f<227>;
	.reg .s64 	%rd<17>;


	ld.param.u64 	%rd5, [LumaCorrector_CompositeKernel_param_0];
	ld.param.u32 	%r3, [LumaCorrector_CompositeKernel_param_1];
	ld.param.u64 	%rd6, [LumaCorrector_CompositeKernel_param_2];
	ld.param.u32 	%r4, [LumaCorrector_CompositeKernel_param_3];
	ld.param.u32 	%r5, [LumaCorrector_CompositeKernel_param_4];
	ld.param.u32 	%r6, [LumaCorrector_CompositeKernel_param_5];
	ld.param.u32 	%r7, [LumaCorrector_CompositeKernel_param_6];
	ld.param.f32 	%f66, [LumaCorrector_CompositeKernel_param_7];
	ld.param.f32 	%f67, [LumaCorrector_CompositeKernel_param_8];
	ld.param.f32 	%f68, [LumaCorrector_CompositeKernel_param_9];
	ld.param.f32 	%f69, [LumaCorrector_CompositeKernel_param_10];
	ld.param.f32 	%f70, [LumaCorrector_CompositeKernel_param_11];
	ld.param.f32 	%f71, [LumaCorrector_CompositeKernel_param_12];
	ld.param.f32 	%f72, [LumaCorrector_CompositeKernel_param_13];
	ld.param.f32 	%f73, [LumaCorrector_CompositeKernel_param_14];
	ld.param.f32 	%f74, [LumaCorrector_CompositeKernel_param_15];
	ld.param.f32 	%f75, [LumaCorrector_CompositeKernel_param_16];
	ld.param.f32 	%f76, [LumaCorrector_CompositeKernel_param_17];
	ld.param.f32 	%f77, [LumaCorrector_CompositeKernel_param_18];
	ld.param.f32 	%f78, [LumaCorrector_CompositeKernel_param_19];
	ld.param.f32 	%f79, [LumaCorrector_CompositeKernel_param_20];
	ld.param.f32 	%f80, [LumaCorrector_CompositeKernel_param_21];
	ld.param.f32 	%f81, [LumaCorrector_CompositeKernel_param_22];
	ld.param.f32 	%f82, [LumaCorrector_CompositeKernel_param_23];
	ld.param.f32 	%f83, [LumaCorrector_CompositeKernel_param_24];
	ld.param.f32 	%f84, [LumaCorrector_CompositeKernel_param_25];
	ld.param.f32 	%f85, [LumaCorrector_CompositeKernel_param_26];
	ld.param.f32 	%f86, [LumaCorrector_CompositeKernel_param_27];
	ld.param.f32 	%f87, [LumaCorrector_CompositeKernel_param_28];
	ld.param.f32 	%f88, [LumaCorrector_CompositeKernel_param_29];
	ld.param.f32 	%f89, [LumaCorrector_CompositeKernel_param_30];
	ld.param.f32 	%f90, [LumaCorrector_CompositeKernel_param_31];
	ld.param.f32 	%f91, [LumaCorrector_CompositeKernel_param_32];
	ld.param.f32 	%f92, [LumaCorrector_CompositeKernel_param_33];
	ld.param.f32 	%f93, [LumaCorrector_CompositeKernel_param_34];
	cvta.to.global.u64 	%rd1, %rd6;
	cvta.to.global.u64 	%rd2, %rd5;
	.loc 1 41 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 41 1
	setp.lt.s32	%p1, %r1, %r6;
	setp.lt.s32	%p2, %r2, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 41 1
	@!%p3 bra 	BB7_42;
	bra.uni 	BB7_1;

BB7_1:
	.loc 1 41 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 41 1
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB7_3;

	ld.global.v4.f32 	{%f94, %f95, %f96, %f97}, [%rd3];
	mov.f32 	%f218, %f97;
	mov.f32 	%f217, %f96;
	mov.f32 	%f216, %f95;
	mov.f32 	%f215, %f94;
	bra.uni 	BB7_4;

BB7_3:
	.loc 1 41 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd4];
	.loc 3 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f215, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f216, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f217, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f218, %temp;
	}

BB7_4:
	ld.const.f32 	%f98, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f99, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f100, %f216, %f99;
	fma.rn.ftz.f32 	%f101, %f217, %f98, %f100;
	ld.const.f32 	%f102, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f226, %f215, %f102, %f101;
	ld.const.f32 	%f103, [kRGB32f_To_601YPbPr+12];
	ld.const.f32 	%f104, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f105, %f216, %f104;
	fma.rn.ftz.f32 	%f106, %f217, %f103, %f105;
	ld.const.f32 	%f107, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f15, %f215, %f107, %f106;
	ld.const.f32 	%f108, [kRGB32f_To_601YPbPr+24];
	ld.const.f32 	%f109, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f110, %f216, %f109;
	fma.rn.ftz.f32 	%f111, %f217, %f108, %f110;
	ld.const.f32 	%f112, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f16, %f215, %f112, %f111;
	.loc 1 41 1
	setp.ne.s64	%p5, %rd6, 0;
	@%p5 bra 	BB7_6;

	mov.f32 	%f219, 0f3F800000;
	bra.uni 	BB7_7;

BB7_6:
	.loc 1 41 1
	mad.lo.s32 	%r15, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r15, 4;
	add.s64 	%rd10, %rd1, %rd9;
	.loc 1 41 1
	ld.global.f32 	%f219, [%rd10];

BB7_7:
	.loc 1 41 1
	setp.leu.ftz.f32	%p6, %f219, 0f3A83126F;
	@%p6 bra 	BB7_39;

	.loc 1 41 1
	setp.gt.ftz.f32	%p7, %f67, %f69;
	selp.f32	%f19, 0f40000000, 0f3F800000, %p7;
	.loc 2 35 1
	setp.gtu.ftz.f32	%p8, %f226, %f66;
	@%p8 bra 	BB7_10;

	mov.f32 	%f220, 0f3F800000;
	bra.uni 	BB7_13;

BB7_10:
	.loc 2 35 1
	add.ftz.f32 	%f115, %f66, %f67;
	setp.ltu.ftz.f32	%p9, %f226, %f115;
	@%p9 bra 	BB7_12;

	mov.f32 	%f220, 0f00000000;
	bra.uni 	BB7_13;

BB7_12:
	.loc 2 35 1
	sub.ftz.f32 	%f117, %f226, %f66;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f220, %f117, %f67;

BB7_13:
	.loc 2 35 1
	sub.ftz.f32 	%f22, %f68, %f69;
	setp.geu.ftz.f32	%p10, %f226, %f22;
	@%p10 bra 	BB7_15;

	mov.f32 	%f221, 0f00000000;
	bra.uni 	BB7_18;

BB7_15:
	.loc 2 35 1
	setp.leu.ftz.f32	%p11, %f226, %f68;
	@%p11 bra 	BB7_17;

	mov.f32 	%f221, 0f3F800000;
	bra.uni 	BB7_18;

BB7_17:
	.loc 2 35 1
	sub.ftz.f32 	%f118, %f226, %f22;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f221, %f118, %f69;

BB7_18:
	.loc 1 41 1
	sub.ftz.f32 	%f121, %f19, %f220;
	sub.ftz.f32 	%f122, %f121, %f221;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f123, %f122, %f19;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f124, %f220, %f19;
	.loc 1 41 1
	mul.ftz.f32 	%f125, %f124, %f88;
	mul.ftz.f32 	%f126, %f123, %f82;
	mul.ftz.f32 	%f127, %f126, %f219;
	fma.rn.ftz.f32 	%f128, %f125, %f219, %f127;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f129, %f221, %f19;
	.loc 1 41 1
	mul.ftz.f32 	%f130, %f129, %f76;
	fma.rn.ftz.f32 	%f25, %f130, %f219, %f128;
	mul.ftz.f32 	%f131, %f124, %f89;
	mul.ftz.f32 	%f132, %f123, %f83;
	mul.ftz.f32 	%f133, %f132, %f219;
	fma.rn.ftz.f32 	%f134, %f131, %f219, %f133;
	mul.ftz.f32 	%f135, %f129, %f77;
	fma.rn.ftz.f32 	%f26, %f135, %f219, %f134;
	mul.ftz.f32 	%f136, %f123, %f84;
	fma.rn.ftz.f32 	%f137, %f124, %f90, %f136;
	fma.rn.ftz.f32 	%f138, %f129, %f78, %f137;
	add.ftz.f32 	%f139, %f138, 0fBF000000;
	fma.rn.ftz.f32 	%f27, %f219, %f139, 0f3F000000;
	mul.ftz.f32 	%f140, %f123, %f85;
	fma.rn.ftz.f32 	%f141, %f124, %f91, %f140;
	fma.rn.ftz.f32 	%f142, %f129, %f79, %f141;
	add.ftz.f32 	%f143, %f142, 0fBF800000;
	fma.rn.ftz.f32 	%f28, %f219, %f143, 0f3F800000;
	mul.ftz.f32 	%f144, %f124, %f92;
	mul.ftz.f32 	%f145, %f123, %f86;
	mul.ftz.f32 	%f146, %f145, %f219;
	fma.rn.ftz.f32 	%f147, %f144, %f219, %f146;
	mul.ftz.f32 	%f148, %f129, %f80;
	fma.rn.ftz.f32 	%f29, %f148, %f219, %f147;
	mul.ftz.f32 	%f149, %f123, %f87;
	fma.rn.ftz.f32 	%f150, %f124, %f93, %f149;
	fma.rn.ftz.f32 	%f151, %f129, %f81, %f150;
	add.ftz.f32 	%f152, %f151, 0fBF800000;
	fma.rn.ftz.f32 	%f30, %f219, %f152, 0f3F800000;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p12, %f26, 0f00000000;
	@%p12 bra 	BB7_24;

	.loc 1 41 1
	sub.ftz.f32 	%f31, %f27, %f25;
	mov.f32 	%f153, 0f3F800000;
	.loc 1 41 1
	sub.ftz.f32 	%f32, %f153, %f26;
	.loc 1 41 1
	setp.eq.ftz.f32	%p13, %f32, 0f00000000;
	@%p13 bra 	BB7_23;

	setp.ltu.ftz.f32	%p14, %f226, %f31;
	.loc 1 41 1
	@%p14 bra 	BB7_22;

	.loc 1 41 1
	sub.ftz.f32 	%f154, %f226, %f31;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f155, %f154, %f32;
	.loc 1 41 95
	add.ftz.f32 	%f222, %f27, %f155;
	bra.uni 	BB7_25;

BB7_22:
	.loc 1 41 1
	sub.ftz.f32 	%f156, %f31, %f226;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f157, %f156, %f32;
	.loc 1 41 95
	sub.ftz.f32 	%f222, %f27, %f157;
	bra.uni 	BB7_25;

BB7_23:
	setp.ltu.ftz.f32	%p15, %f226, %f31;
	.loc 1 41 1
	selp.f32	%f222, 0f00000000, 0f3F800000, %p15;
	bra.uni 	BB7_25;

BB7_24:
	.loc 1 41 1
	add.ftz.f32 	%f158, %f27, %f25;
	.loc 1 41 1
	sub.ftz.f32 	%f159, %f226, %f27;
	.loc 1 41 1
	add.ftz.f32 	%f160, %f26, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f161, %f160, %f159;
	.loc 1 41 1
	sub.ftz.f32 	%f162, %f27, %f226;
	mul.ftz.f32 	%f163, %f160, %f162;
	.loc 1 41 1
	neg.ftz.f32 	%f164, %f163;
	setp.ltu.ftz.f32	%p16, %f226, %f27;
	selp.f32	%f165, %f164, %f161, %p16;
	add.ftz.f32 	%f222, %f158, %f165;

BB7_25:
	mov.f32 	%f166, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f38, %f166, %f28;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p17, %f222, 0f00000000;
	@%p17 bra 	BB7_27;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f167, %f222;
	mul.ftz.f32 	%f168, %f38, %f167;
	ex2.approx.ftz.f32 	%f223, %f168;
	bra.uni 	BB7_28;

BB7_27:
	.loc 1 41 216
	neg.ftz.f32 	%f169, %f222;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f170, %f169;
	mul.ftz.f32 	%f171, %f38, %f170;
	ex2.approx.ftz.f32 	%f172, %f171;
	.loc 1 41 243
	neg.ftz.f32 	%f223, %f172;

BB7_28:
	.loc 1 41 1
	add.ftz.f32 	%f173, %f223, %f29;
	mul.ftz.f32 	%f42, %f173, %f30;
	add.ftz.f32 	%f174, %f72, 0fBF000000;
	fma.rn.ftz.f32 	%f43, %f219, %f174, 0f3F000000;
	mul.ftz.f32 	%f44, %f219, %f70;
	mul.ftz.f32 	%f45, %f219, %f71;
	.loc 1 41 1
	setp.ltu.ftz.f32	%p18, %f45, 0f00000000;
	@%p18 bra 	BB7_34;

	.loc 1 41 1
	sub.ftz.f32 	%f46, %f43, %f44;
	sub.ftz.f32 	%f47, %f166, %f45;
	.loc 1 41 1
	setp.eq.ftz.f32	%p19, %f47, 0f00000000;
	@%p19 bra 	BB7_33;

	setp.ltu.ftz.f32	%p20, %f42, %f46;
	.loc 1 41 1
	@%p20 bra 	BB7_32;

	.loc 1 41 1
	sub.ftz.f32 	%f176, %f42, %f46;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f177, %f176, %f47;
	.loc 1 41 95
	add.ftz.f32 	%f224, %f43, %f177;
	bra.uni 	BB7_35;

BB7_32:
	.loc 1 41 1
	sub.ftz.f32 	%f178, %f46, %f42;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f179, %f178, %f47;
	.loc 1 41 95
	sub.ftz.f32 	%f224, %f43, %f179;
	bra.uni 	BB7_35;

BB7_33:
	setp.ltu.ftz.f32	%p21, %f42, %f46;
	.loc 1 41 1
	selp.f32	%f224, 0f00000000, 0f3F800000, %p21;
	bra.uni 	BB7_35;

BB7_34:
	.loc 1 41 1
	add.ftz.f32 	%f180, %f43, %f44;
	.loc 1 41 1
	sub.ftz.f32 	%f181, %f42, %f43;
	.loc 1 41 1
	add.ftz.f32 	%f182, %f45, 0f3F800000;
	.loc 1 41 1
	mul.ftz.f32 	%f183, %f182, %f181;
	.loc 1 41 1
	sub.ftz.f32 	%f184, %f43, %f42;
	mul.ftz.f32 	%f185, %f182, %f184;
	.loc 1 41 1
	neg.ftz.f32 	%f186, %f185;
	setp.ltu.ftz.f32	%p22, %f42, %f43;
	selp.f32	%f187, %f186, %f183, %p22;
	add.ftz.f32 	%f224, %f180, %f187;

BB7_35:
	.loc 1 41 1
	add.ftz.f32 	%f188, %f73, 0fBF800000;
	fma.rn.ftz.f32 	%f189, %f219, %f188, 0f3F800000;
	.loc 3 3606 10
	div.approx.ftz.f32 	%f53, %f166, %f189;
	.loc 1 41 122
	setp.ltu.ftz.f32	%p23, %f224, 0f00000000;
	@%p23 bra 	BB7_37;

	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f191, %f224;
	mul.ftz.f32 	%f192, %f53, %f191;
	ex2.approx.ftz.f32 	%f225, %f192;
	bra.uni 	BB7_38;

BB7_37:
	.loc 1 41 219
	neg.ftz.f32 	%f193, %f224;
	.loc 3 3600 10
	lg2.approx.ftz.f32 	%f194, %f193;
	mul.ftz.f32 	%f195, %f53, %f194;
	ex2.approx.ftz.f32 	%f196, %f195;
	.loc 1 41 246
	neg.ftz.f32 	%f225, %f196;

BB7_38:
	.loc 1 41 1
	fma.rn.ftz.f32 	%f197, %f219, %f74, %f225;
	add.ftz.f32 	%f198, %f75, 0fBF800000;
	fma.rn.ftz.f32 	%f199, %f219, %f198, 0f3F800000;
	mul.ftz.f32 	%f226, %f197, %f199;

BB7_39:
	.loc 1 41 1
	ld.const.f32 	%f200, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f201, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f202, %f15, %f201;
	fma.rn.ftz.f32 	%f203, %f226, %f200, %f202;
	ld.const.f32 	%f204, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f59, %f16, %f204, %f203;
	ld.const.f32 	%f205, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f206, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f207, %f15, %f206;
	fma.rn.ftz.f32 	%f208, %f226, %f205, %f207;
	ld.const.f32 	%f209, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f60, %f16, %f209, %f208;
	ld.const.f32 	%f210, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f211, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f212, %f15, %f211;
	fma.rn.ftz.f32 	%f213, %f226, %f210, %f212;
	ld.const.f32 	%f214, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f61, %f16, %f214, %f213;
	.loc 1 41 1
	@%p4 bra 	BB7_41;

	mul.wide.s32 	%rd12, %r14, 16;
	add.s64 	%rd13, %rd2, %rd12;
	.loc 1 41 1
	st.global.v4.f32 	[%rd13], {%f61, %f60, %f59, %f218};
	bra.uni 	BB7_42;

BB7_41:
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f61;
	mov.b16 	%rs9, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f60;
	mov.b16 	%rs10, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f59;
	mov.b16 	%rs11, %temp;
}
	.loc 3 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f218;
	mov.b16 	%rs12, %temp;
}
	mul.wide.s32 	%rd15, %r14, 8;
	add.s64 	%rd16, %rd2, %rd15;
	.loc 1 41 241
	st.global.v4.u16 	[%rd16], {%rs9, %rs10, %rs11, %rs12};

BB7_42:
	.loc 1 41 2
	ret;
}


