//
// 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/Invert.cu", 1399785316, 8272
	.file	2 "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};
.const .align 4 .b8 kRGB32f_To_YIQ[36] = {135, 22, 153, 62, 162, 69, 22, 63, 213, 120, 233, 61, 216, 128, 24, 63, 27, 133, 140, 190, 149, 124, 164, 190, 236, 135, 88, 62, 134, 200, 5, 191, 22, 77, 159, 62};
.const .align 4 .b8 kYIQ_To_RGB32f[36] = {0, 0, 128, 63, 20, 208, 116, 63, 219, 249, 30, 63, 0, 0, 128, 63, 177, 80, 139, 190, 2, 188, 37, 191, 0, 0, 128, 63, 45, 178, 141, 191, 85, 48, 218, 63};
.global .align 1 .b8 $str[11] = {95, 95, 67, 85, 68, 65, 95, 70, 84, 90, 0};

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z15ConvertRGBtoHSL8PixelRGB(
	.param .align 16 .b8 _Z15ConvertRGBtoHSL8PixelRGB_param_0[16]
)
{
	.reg .pred 	%p<16>;
	.reg .f32 	%f<37>;


	ld.param.f32 	%f10, [_Z15ConvertRGBtoHSL8PixelRGB_param_0+8];
	ld.param.f32 	%f11, [_Z15ConvertRGBtoHSL8PixelRGB_param_0];
	ld.param.f32 	%f12, [_Z15ConvertRGBtoHSL8PixelRGB_param_0+4];
	.loc 1 36 1
	sub.ftz.f32 	%f13, %f12, %f11;
	.loc 1 36 1
	setp.ge.ftz.f32	%p1, %f12, %f11;
	setp.ge.ftz.f32	%p2, %f12, %f10;
	and.pred  	%p3, %p2, %p1;
	.loc 1 36 1
	sub.ftz.f32 	%f14, %f11, %f10;
	.loc 1 36 1
	selp.f32	%f15, %f14, %f13, %p3;
	selp.f32	%f16, 0f40000000, 0f00000000, %p3;
	mov.f32 	%f17, 0f40000000;
	.loc 1 36 1
	selp.f32	%f18, %f12, %f10, %p3;
	.loc 1 36 1
	setp.ge.ftz.f32	%p4, %f11, %f12;
	setp.ge.ftz.f32	%p5, %f11, %f10;
	and.pred  	%p6, %p5, %p4;
	.loc 1 36 1
	sub.ftz.f32 	%f19, %f10, %f12;
	.loc 1 36 1
	selp.f32	%f1, %f19, %f15, %p6;
	selp.f32	%f2, 0f40800000, %f16, %p6;
	selp.f32	%f3, %f11, %f18, %p6;
	.loc 1 36 1
	setp.le.ftz.f32	%p7, %f12, %f11;
	setp.le.ftz.f32	%p8, %f12, %f10;
	and.pred  	%p9, %p8, %p7;
	.loc 1 36 1
	selp.f32	%f20, %f12, %f10, %p9;
	.loc 1 36 1
	setp.le.ftz.f32	%p10, %f11, %f12;
	setp.le.ftz.f32	%p11, %f11, %f10;
	and.pred  	%p12, %p11, %p10;
	.loc 1 36 1
	selp.f32	%f4, %f11, %f20, %p12;
	.loc 1 36 1
	add.ftz.f32 	%f21, %f3, %f4;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f5, %f21, %f17;
	.loc 1 36 1
	setp.gt.ftz.f32	%p13, %f5, 0f38D1B717;
	.loc 1 36 1
	mov.f32 	%f34, 0f00000000;
	mov.f32 	%f33, 0f00000000;
	.loc 1 36 1
	@%p13 bra 	BB0_2;

	mov.f32 	%f36, %f33;
	bra.uni 	BB0_5;

BB0_2:
	.loc 1 36 1
	sub.ftz.f32 	%f6, %f3, %f4;
	.loc 1 36 1
	setp.gt.ftz.f32	%p14, %f6, 0f38D1B717;
	@%p14 bra 	BB0_4;

	mov.f32 	%f36, %f33;
	bra.uni 	BB0_5;

BB0_4:
	.loc 2 3606 10
	div.approx.ftz.f32 	%f26, %f1, %f6;
	.loc 1 36 1
	add.ftz.f32 	%f27, %f26, %f2;
	.loc 1 36 1
	setp.lt.ftz.f32	%p15, %f27, 0f00000000;
	.loc 1 36 1
	add.ftz.f32 	%f28, %f27, 0f40C00000;
	mov.f32 	%f29, 0f40C00000;
	.loc 1 36 1
	selp.f32	%f30, %f28, %f27, %p15;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f34, %f30, %f29;
	mov.f32 	%f36, %f6;

BB0_5:
	st.param.f32	[func_retval0+0], %f34;
	st.param.f32	[func_retval0+4], %f36;
	st.param.f32	[func_retval0+8], %f5;
	st.param.f32	[func_retval0+12], %f31;
	.loc 1 36 1
	ret;
}

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z15ConvertHSLtoRGB8PixelHSL(
	.param .align 16 .b8 _Z15ConvertHSLtoRGB8PixelHSL_param_0[16]
)
{
	.reg .pred 	%p<10>;
	.reg .s32 	%r<2>;
	.reg .f32 	%f<54>;


	ld.param.f32 	%f1, [_Z15ConvertHSLtoRGB8PixelHSL_param_0];
	ld.param.f32 	%f10, [_Z15ConvertHSLtoRGB8PixelHSL_param_0+4];
	ld.param.f32 	%f2, [_Z15ConvertHSLtoRGB8PixelHSL_param_0+8];
	.loc 1 36 1
	setp.gtu.ftz.f32	%p1, %f2, 0f3F000000;
	add.ftz.f32 	%f11, %f10, 0f3F800000;
	mul.ftz.f32 	%f12, %f2, %f11;
	add.ftz.f32 	%f13, %f2, %f10;
	mul.ftz.f32 	%f14, %f2, %f10;
	sub.ftz.f32 	%f15, %f13, %f14;
	selp.f32	%f3, %f15, %f12, %p1;
	.loc 1 36 1
	setp.gt.ftz.f32	%p2, %f3, 0f00000000;
	.loc 1 36 1
	mov.f32 	%f31, 0f00000000;
	mov.f32 	%f32, 0f00000000;
	mov.f32 	%f33, 0f00000000;
	.loc 1 36 1
	@%p2 bra 	BB1_2;

	mov.f32 	%f53, %f31;
	mov.f32 	%f52, %f32;
	mov.f32 	%f50, %f33;
	mov.f32 	%f51, %f50;
	bra.uni 	BB1_10;

BB1_2:
	.loc 1 36 1
	add.ftz.f32 	%f19, %f2, %f2;
	sub.ftz.f32 	%f4, %f19, %f3;
	sub.ftz.f32 	%f20, %f3, %f4;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f21, %f20, %f3;
	.loc 1 36 1
	mul.ftz.f32 	%f22, %f1, 0f40C00000;
	.loc 2 3251 10
	cvt.rzi.ftz.s32.f32	%r1, %f22;
	.loc 1 36 1
	cvt.rn.f32.s32	%f23, %r1;
	sub.ftz.f32 	%f24, %f22, %f23;
	mul.ftz.f32 	%f25, %f3, %f21;
	mul.ftz.f32 	%f26, %f25, %f24;
	add.ftz.f32 	%f5, %f4, %f26;
	sub.ftz.f32 	%f6, %f3, %f26;
	.loc 1 36 1
	setp.gt.s32	%p3, %r1, 2;
	@%p3 bra 	BB1_6;

	setp.eq.s32	%p7, %r1, 0;
	mov.f32 	%f43, %f3;
	mov.f32 	%f48, %f5;
	mov.f32 	%f51, %f43;
	mov.f32 	%f52, %f48;
	mov.f32 	%f53, %f4;
	@%p7 bra 	BB1_10;

	setp.eq.s32	%p8, %r1, 1;
	mov.f32 	%f39, %f4;
	mov.f32 	%f40, %f6;
	mov.f32 	%f44, %f3;
	mov.f32 	%f51, %f40;
	mov.f32 	%f52, %f44;
	mov.f32 	%f53, %f39;
	@%p8 bra 	BB1_10;

	setp.eq.s32	%p9, %r1, 2;
	mov.f32 	%f35, %f4;
	mov.f32 	%f45, %f3;
	mov.f32 	%f51, %f35;
	mov.f32 	%f52, %f45;
	mov.f32 	%f53, %f5;
	@%p9 bra 	BB1_10;
	bra.uni 	BB1_9;

BB1_6:
	setp.eq.s32	%p4, %r1, 3;
	mov.f32 	%f36, %f4;
	mov.f32 	%f41, %f6;
	mov.f32 	%f46, %f3;
	mov.f32 	%f51, %f36;
	mov.f32 	%f52, %f41;
	mov.f32 	%f53, %f46;
	@%p4 bra 	BB1_10;

	setp.eq.s32	%p5, %r1, 4;
	mov.f32 	%f37, %f4;
	mov.f32 	%f47, %f5;
	mov.f32 	%f51, %f47;
	mov.f32 	%f52, %f37;
	mov.f32 	%f53, %f3;
	@%p5 bra 	BB1_10;

	setp.ne.s32	%p6, %r1, 5;
	mov.f32 	%f38, %f4;
	mov.f32 	%f42, %f3;
	mov.f32 	%f51, %f42;
	mov.f32 	%f52, %f38;
	mov.f32 	%f53, %f6;
	@%p6 bra 	BB1_9;
	bra.uni 	BB1_10;

BB1_9:
	mov.f32 	%f51, %f33;
	mov.f32 	%f52, %f32;
	mov.f32 	%f53, %f31;

BB1_10:
	.loc 1 36 1
	mov.f32 	%f9, %f51;
	st.param.f32	[func_retval0+0], %f53;
	st.param.f32	[func_retval0+4], %f52;
	st.param.f32	[func_retval0+8], %f9;
	st.param.f32	[func_retval0+12], %f33;
	ret;
}

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z15ConvertRGBtoYIQ8PixelRGB(
	.param .align 16 .b8 _Z15ConvertRGBtoYIQ8PixelRGB_param_0[16]
)
{
	.reg .f32 	%f<23>;


	ld.param.f32 	%f1, [_Z15ConvertRGBtoYIQ8PixelRGB_param_0+12];
	ld.param.f32 	%f2, [_Z15ConvertRGBtoYIQ8PixelRGB_param_0];
	ld.param.f32 	%f3, [_Z15ConvertRGBtoYIQ8PixelRGB_param_0+8];
	ld.param.f32 	%f4, [_Z15ConvertRGBtoYIQ8PixelRGB_param_0+4];
	.loc 1 36 1
	ld.const.f32 	%f5, [kRGB32f_To_YIQ];
	ld.const.f32 	%f6, [kRGB32f_To_YIQ+4];
	mul.ftz.f32 	%f7, %f4, %f6;
	fma.rn.ftz.f32 	%f8, %f3, %f5, %f7;
	ld.const.f32 	%f9, [kRGB32f_To_YIQ+8];
	fma.rn.ftz.f32 	%f10, %f2, %f9, %f8;
	ld.const.f32 	%f11, [kRGB32f_To_YIQ+12];
	ld.const.f32 	%f12, [kRGB32f_To_YIQ+16];
	mul.ftz.f32 	%f13, %f4, %f12;
	fma.rn.ftz.f32 	%f14, %f3, %f11, %f13;
	ld.const.f32 	%f15, [kRGB32f_To_YIQ+20];
	fma.rn.ftz.f32 	%f16, %f2, %f15, %f14;
	ld.const.f32 	%f17, [kRGB32f_To_YIQ+24];
	ld.const.f32 	%f18, [kRGB32f_To_YIQ+28];
	mul.ftz.f32 	%f19, %f4, %f18;
	fma.rn.ftz.f32 	%f20, %f3, %f17, %f19;
	ld.const.f32 	%f21, [kRGB32f_To_YIQ+32];
	fma.rn.ftz.f32 	%f22, %f2, %f21, %f20;
	st.param.f32	[func_retval0+0], %f1;
	st.param.f32	[func_retval0+4], %f10;
	st.param.f32	[func_retval0+8], %f16;
	st.param.f32	[func_retval0+12], %f22;
	ret;
}

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z15ConvertYIQtoRGB6float4(
	.param .align 16 .b8 _Z15ConvertYIQtoRGB6float4_param_0[16]
)
{
	.reg .f32 	%f<23>;


	ld.param.f32 	%f1, [_Z15ConvertYIQtoRGB6float4_param_0];
	ld.param.f32 	%f2, [_Z15ConvertYIQtoRGB6float4_param_0+12];
	ld.param.f32 	%f3, [_Z15ConvertYIQtoRGB6float4_param_0+4];
	ld.param.f32 	%f4, [_Z15ConvertYIQtoRGB6float4_param_0+8];
	.loc 1 36 1
	ld.const.f32 	%f5, [kYIQ_To_RGB32f];
	ld.const.f32 	%f6, [kYIQ_To_RGB32f+4];
	mul.ftz.f32 	%f7, %f4, %f6;
	fma.rn.ftz.f32 	%f8, %f3, %f5, %f7;
	ld.const.f32 	%f9, [kYIQ_To_RGB32f+8];
	fma.rn.ftz.f32 	%f10, %f2, %f9, %f8;
	ld.const.f32 	%f11, [kYIQ_To_RGB32f+12];
	ld.const.f32 	%f12, [kYIQ_To_RGB32f+16];
	mul.ftz.f32 	%f13, %f4, %f12;
	fma.rn.ftz.f32 	%f14, %f3, %f11, %f13;
	ld.const.f32 	%f15, [kYIQ_To_RGB32f+20];
	fma.rn.ftz.f32 	%f16, %f2, %f15, %f14;
	ld.const.f32 	%f17, [kYIQ_To_RGB32f+24];
	ld.const.f32 	%f18, [kYIQ_To_RGB32f+28];
	mul.ftz.f32 	%f19, %f4, %f18;
	fma.rn.ftz.f32 	%f20, %f3, %f17, %f19;
	ld.const.f32 	%f21, [kYIQ_To_RGB32f+32];
	fma.rn.ftz.f32 	%f22, %f2, %f21, %f20;
	st.param.f32	[func_retval0+0], %f22;
	st.param.f32	[func_retval0+4], %f16;
	st.param.f32	[func_retval0+8], %f10;
	st.param.f32	[func_retval0+12], %f1;
	ret;
}

.visible .entry InvertKernel(
	.param .u32 InvertKernel_param_0,
	.param .u64 InvertKernel_param_1,
	.param .u32 InvertKernel_param_2,
	.param .u32 InvertKernel_param_3,
	.param .u32 InvertKernel_param_4,
	.param .u64 InvertKernel_param_5,
	.param .u32 InvertKernel_param_6,
	.param .u32 InvertKernel_param_7,
	.param .f32 InvertKernel_param_8
)
{
	.reg .pred 	%p<47>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<24>;
	.reg .f32 	%f<222>;
	.reg .s64 	%rd<14>;


	ld.param.u32 	%r6, [InvertKernel_param_0];
	ld.param.u64 	%rd2, [InvertKernel_param_1];
	ld.param.u32 	%r10, [InvertKernel_param_2];
	ld.param.u32 	%r11, [InvertKernel_param_3];
	ld.param.u32 	%r7, [InvertKernel_param_4];
	ld.param.u64 	%rd3, [InvertKernel_param_5];
	ld.param.u32 	%r8, [InvertKernel_param_6];
	ld.param.u32 	%r9, [InvertKernel_param_7];
	ld.param.f32 	%f78, [InvertKernel_param_8];
	cvta.to.global.u64 	%rd1, %rd3;
	.loc 1 36 1
	mov.u32 	%r12, %ntid.x;
	mov.u32 	%r13, %ctaid.x;
	mov.u32 	%r14, %tid.x;
	mad.lo.s32 	%r1, %r12, %r13, %r14;
	mov.u32 	%r15, %ntid.y;
	mov.u32 	%r16, %ctaid.y;
	mov.u32 	%r17, %tid.y;
	mad.lo.s32 	%r2, %r15, %r16, %r17;
	.loc 1 36 1
	setp.ge.s32	%p1, %r2, %r11;
	setp.ge.s32	%p2, %r1, %r10;
	or.pred  	%p3, %p2, %p1;
	.loc 1 36 1
	@%p3 bra 	BB4_51;

	.loc 1 36 1
	mad.lo.s32 	%r3, %r2, %r8, %r1;
	setp.eq.s32	%p4, %r6, 0;
	@%p4 bra 	BB4_3;

	mul.wide.s32 	%rd4, %r3, 16;
	add.s64 	%rd5, %rd1, %rd4;
	ld.global.v4.f32 	{%f79, %f80, %f81, %f82}, [%rd5];
	mov.f32 	%f186, %f82;
	mov.f32 	%f185, %f81;
	mov.f32 	%f184, %f80;
	mov.f32 	%f183, %f79;
	bra.uni 	BB4_4;

BB4_3:
	mul.wide.s32 	%rd6, %r3, 8;
	add.s64 	%rd7, %rd1, %rd6;
	.loc 1 36 1
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd7];
	.loc 2 3518 10
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f183, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f184, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f185, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f186, %temp;
	}

BB4_4:
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f216, %f185;
	cvt.ftz.sat.f32.f32	%f217, %f184;
	cvt.ftz.sat.f32.f32	%f218, %f183;
	cvt.ftz.sat.f32.f32	%f198, %f186;
	.loc 1 36 1
	setp.eq.s32	%p5, %r9, 0;
	@%p5 bra 	BB4_45;

	setp.eq.s32	%p6, %r9, 1;
	@%p6 bra 	BB4_44;

	setp.eq.s32	%p7, %r9, 2;
	@%p7 bra 	BB4_43;

	setp.eq.s32	%p8, %r9, 3;
	@%p8 bra 	BB4_42;

	add.s32 	%r18, %r9, -5;
	setp.lt.u32	%p9, %r18, 4;
	@%p9 bra 	BB4_21;

	add.s32 	%r19, %r9, -10;
	setp.lt.u32	%p10, %r19, 4;
	@%p10 bra 	BB4_12;

	setp.ne.s32	%p11, %r9, 15;
	mov.f32 	%f219, %f86;
	mov.f32 	%f220, %f87;
	mov.f32 	%f221, %f88;
	.loc 1 36 1
	@%p11 bra 	BB4_46;

	mov.f32 	%f92, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f93, %f92, %f198;
	sub.ftz.f32 	%f94, %f92, %f78;
	mul.ftz.f32 	%f95, %f94, %f93;
	fma.rn.ftz.f32 	%f198, %f198, %f78, %f95;
	mov.f32 	%f219, %f96;
	mov.f32 	%f220, %f97;
	mov.f32 	%f221, %f98;
	bra.uni 	BB4_46;

BB4_12:
	.loc 1 36 1
	ld.const.f32 	%f99, [kRGB32f_To_YIQ];
	ld.const.f32 	%f100, [kRGB32f_To_YIQ+4];
	mul.ftz.f32 	%f101, %f217, %f100;
	fma.rn.ftz.f32 	%f102, %f216, %f99, %f101;
	ld.const.f32 	%f103, [kRGB32f_To_YIQ+8];
	fma.rn.ftz.f32 	%f187, %f218, %f103, %f102;
	ld.const.f32 	%f104, [kRGB32f_To_YIQ+12];
	ld.const.f32 	%f105, [kRGB32f_To_YIQ+16];
	mul.ftz.f32 	%f106, %f217, %f105;
	fma.rn.ftz.f32 	%f107, %f216, %f104, %f106;
	ld.const.f32 	%f108, [kRGB32f_To_YIQ+20];
	fma.rn.ftz.f32 	%f188, %f218, %f108, %f107;
	ld.const.f32 	%f109, [kRGB32f_To_YIQ+24];
	ld.const.f32 	%f110, [kRGB32f_To_YIQ+28];
	mul.ftz.f32 	%f111, %f217, %f110;
	fma.rn.ftz.f32 	%f112, %f216, %f109, %f111;
	ld.const.f32 	%f113, [kRGB32f_To_YIQ+32];
	fma.rn.ftz.f32 	%f189, %f218, %f113, %f112;
	.loc 1 36 1
	setp.eq.s32	%p12, %r9, 10;
	@%p12 bra 	BB4_18;

	setp.eq.s32	%p13, %r9, 11;
	@%p13 bra 	BB4_17;

	setp.eq.s32	%p14, %r9, 12;
	@%p14 bra 	BB4_16;

	setp.ne.s32	%p15, %r9, 13;
	@%p15 bra 	BB4_20;
	bra.uni 	BB4_19;

BB4_16:
	.loc 1 36 1
	neg.ftz.f32 	%f188, %f188;
	bra.uni 	BB4_20;

BB4_17:
	mov.f32 	%f114, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f187, %f114, %f187;
	bra.uni 	BB4_20;

BB4_18:
	mov.f32 	%f115, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f187, %f115, %f187;
	neg.ftz.f32 	%f188, %f188;

BB4_19:
	neg.ftz.f32 	%f189, %f189;

BB4_20:
	.loc 1 36 1
	ld.const.f32 	%f116, [kYIQ_To_RGB32f];
	ld.const.f32 	%f117, [kYIQ_To_RGB32f+4];
	mul.ftz.f32 	%f118, %f188, %f117;
	fma.rn.ftz.f32 	%f119, %f187, %f116, %f118;
	ld.const.f32 	%f120, [kYIQ_To_RGB32f+8];
	fma.rn.ftz.f32 	%f219, %f189, %f120, %f119;
	ld.const.f32 	%f121, [kYIQ_To_RGB32f+12];
	ld.const.f32 	%f122, [kYIQ_To_RGB32f+16];
	mul.ftz.f32 	%f123, %f188, %f122;
	fma.rn.ftz.f32 	%f124, %f187, %f121, %f123;
	ld.const.f32 	%f125, [kYIQ_To_RGB32f+20];
	fma.rn.ftz.f32 	%f220, %f189, %f125, %f124;
	ld.const.f32 	%f126, [kYIQ_To_RGB32f+24];
	ld.const.f32 	%f127, [kYIQ_To_RGB32f+28];
	mul.ftz.f32 	%f128, %f188, %f127;
	fma.rn.ftz.f32 	%f129, %f187, %f126, %f128;
	ld.const.f32 	%f130, [kYIQ_To_RGB32f+32];
	fma.rn.ftz.f32 	%f221, %f189, %f130, %f129;
	bra.uni 	BB4_46;

BB4_21:
	.loc 1 36 1
	sub.ftz.f32 	%f131, %f217, %f218;
	setp.ge.ftz.f32	%p16, %f217, %f216;
	.loc 1 36 1
	setp.ge.ftz.f32	%p17, %f217, %f218;
	and.pred  	%p18, %p16, %p17;
	.loc 1 36 1
	sub.ftz.f32 	%f132, %f218, %f216;
	.loc 1 36 1
	selp.f32	%f133, %f132, %f131, %p18;
	selp.f32	%f134, 0f40000000, 0f00000000, %p18;
	mov.f32 	%f135, 0f40000000;
	.loc 1 36 1
	selp.f32	%f136, %f217, %f216, %p18;
	setp.ge.ftz.f32	%p19, %f218, %f216;
	.loc 1 36 1
	setp.ge.ftz.f32	%p20, %f218, %f217;
	and.pred  	%p21, %p19, %p20;
	.loc 1 36 1
	sub.ftz.f32 	%f137, %f216, %f217;
	.loc 1 36 1
	selp.f32	%f33, %f137, %f133, %p21;
	selp.f32	%f34, 0f40800000, %f134, %p21;
	selp.f32	%f35, %f218, %f136, %p21;
	setp.le.ftz.f32	%p22, %f217, %f216;
	.loc 1 36 1
	setp.le.ftz.f32	%p23, %f217, %f218;
	and.pred  	%p24, %p22, %p23;
	.loc 1 36 1
	selp.f32	%f138, %f217, %f216, %p24;
	setp.le.ftz.f32	%p25, %f218, %f216;
	.loc 1 36 1
	setp.le.ftz.f32	%p26, %f218, %f217;
	and.pred  	%p27, %p25, %p26;
	.loc 1 36 1
	selp.f32	%f36, %f218, %f138, %p27;
	.loc 1 36 1
	add.ftz.f32 	%f139, %f35, %f36;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f197, %f139, %f135;
	.loc 1 36 1
	setp.gt.ftz.f32	%p28, %f197, 0f38D1B717;
	.loc 1 36 1
	mov.f32 	%f181, 0f00000000;
	mov.f32 	%f190, 0f00000000;
	.loc 1 36 1
	@%p28 bra 	BB4_23;

	mov.f32 	%f196, %f181;
	bra.uni 	BB4_26;

BB4_23:
	.loc 1 36 1
	sub.ftz.f32 	%f38, %f35, %f36;
	.loc 1 36 1
	setp.gt.ftz.f32	%p29, %f38, 0f38D1B717;
	@%p29 bra 	BB4_25;

	mov.f32 	%f196, %f181;
	bra.uni 	BB4_26;

BB4_25:
	.loc 2 3606 10
	div.approx.ftz.f32 	%f144, %f33, %f38;
	.loc 1 36 1
	add.ftz.f32 	%f145, %f144, %f34;
	.loc 1 36 1
	setp.lt.ftz.f32	%p30, %f145, 0f00000000;
	.loc 1 36 1
	add.ftz.f32 	%f146, %f145, 0f40C00000;
	mov.f32 	%f147, 0f40C00000;
	.loc 1 36 1
	selp.f32	%f148, %f146, %f145, %p30;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f190, %f148, %f147;
	mov.f32 	%f196, %f38;

BB4_26:
	.loc 1 36 1
	mov.f32 	%f195, %f196;
	.loc 1 36 1
	setp.eq.s32	%p31, %r9, 5;
	@%p31 bra 	BB4_33;

	setp.eq.s32	%p32, %r9, 6;
	@%p32 bra 	BB4_32;

	setp.eq.s32	%p33, %r9, 7;
	@%p33 bra 	BB4_31;

	setp.ne.s32	%p34, %r9, 8;
	@%p34 bra 	BB4_34;

	mov.f32 	%f149, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f195, %f149, %f195;
	bra.uni 	BB4_34;

BB4_31:
	mov.f32 	%f150, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f197, %f150, %f197;
	bra.uni 	BB4_34;

BB4_32:
	mov.f32 	%f151, 0f3F000000;
	.loc 1 36 1
	sub.ftz.f32 	%f152, %f151, %f190;
	.loc 1 36 1
	setp.lt.ftz.f32	%p35, %f152, 0f00000000;
	.loc 1 36 1
	add.ftz.f32 	%f153, %f152, 0f3F800000;
	.loc 1 36 1
	selp.f32	%f190, %f153, %f152, %p35;
	bra.uni 	BB4_34;

BB4_33:
	mov.f32 	%f154, 0f3F000000;
	.loc 1 36 1
	sub.ftz.f32 	%f155, %f154, %f190;
	.loc 1 36 1
	setp.lt.ftz.f32	%p36, %f155, 0f00000000;
	.loc 1 36 1
	add.ftz.f32 	%f156, %f155, 0f3F800000;
	mov.f32 	%f157, 0f3F800000;
	.loc 1 36 1
	selp.f32	%f190, %f156, %f155, %p36;
	.loc 1 36 1
	sub.ftz.f32 	%f197, %f157, %f197;
	sub.ftz.f32 	%f195, %f157, %f195;

BB4_34:
	.loc 1 36 1
	add.ftz.f32 	%f158, %f195, 0f3F800000;
	mul.ftz.f32 	%f159, %f197, %f158;
	mul.ftz.f32 	%f160, %f197, %f195;
	add.ftz.f32 	%f161, %f197, %f195;
	sub.ftz.f32 	%f162, %f161, %f160;
	setp.gtu.ftz.f32	%p37, %f197, 0f3F000000;
	selp.f32	%f51, %f162, %f159, %p37;
	.loc 1 36 1
	setp.gt.ftz.f32	%p38, %f51, 0f00000000;
	@%p38 bra 	BB4_36;

	mov.f32 	%f219, 0f00000000;
	mov.f32 	%f220, %f219;
	mov.f32 	%f221, %f219;
	bra.uni 	BB4_46;

BB4_36:
	.loc 1 36 1
	add.ftz.f32 	%f166, %f197, %f197;
	sub.ftz.f32 	%f52, %f166, %f51;
	sub.ftz.f32 	%f167, %f51, %f52;
	.loc 2 3606 10
	div.approx.ftz.f32 	%f168, %f167, %f51;
	.loc 1 36 1
	mul.ftz.f32 	%f169, %f190, 0f40C00000;
	.loc 2 3251 10
	cvt.rzi.ftz.s32.f32	%r4, %f169;
	.loc 1 36 1
	cvt.rn.f32.s32	%f170, %r4;
	sub.ftz.f32 	%f171, %f169, %f170;
	mul.ftz.f32 	%f172, %f51, %f168;
	mul.ftz.f32 	%f173, %f172, %f171;
	add.ftz.f32 	%f53, %f52, %f173;
	sub.ftz.f32 	%f54, %f51, %f173;
	.loc 1 36 1
	setp.eq.s32	%p39, %r4, 0;
	mov.f32 	%f204, %f53;
	mov.f32 	%f219, %f51;
	mov.f32 	%f220, %f204;
	mov.f32 	%f221, %f52;
	@%p39 bra 	BB4_46;

	setp.eq.s32	%p40, %r4, 1;
	mov.f32 	%f202, %f52;
	mov.f32 	%f207, %f51;
	mov.f32 	%f219, %f54;
	mov.f32 	%f220, %f207;
	mov.f32 	%f221, %f202;
	@%p40 bra 	BB4_46;

	setp.eq.s32	%p41, %r4, 2;
	mov.f32 	%f219, %f52;
	mov.f32 	%f206, %f51;
	mov.f32 	%f220, %f206;
	mov.f32 	%f221, %f53;
	@%p41 bra 	BB4_46;

	setp.eq.s32	%p42, %r4, 3;
	mov.f32 	%f219, %f52;
	mov.f32 	%f220, %f54;
	mov.f32 	%f221, %f51;
	@%p42 bra 	BB4_46;

	setp.eq.s32	%p43, %r4, 4;
	mov.f32 	%f201, %f52;
	mov.f32 	%f219, %f53;
	mov.f32 	%f208, %f51;
	mov.f32 	%f220, %f201;
	mov.f32 	%f221, %f208;
	@%p43 bra 	BB4_46;

	setp.eq.s32	%p44, %r4, 5;
	.loc 1 36 1
	selp.f32	%f219, %f51, 0f00000000, %p44;
	selp.f32	%f56, %f52, 0f00000000, %p44;
	selp.f32	%f221, %f54, 0f00000000, %p44;
	mov.f32 	%f220, %f56;
	bra.uni 	BB4_46;

BB4_42:
	mov.f32 	%f174, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f221, %f174, %f218;
	mov.f32 	%f220, %f217;
	mov.f32 	%f219, %f216;
	bra.uni 	BB4_46;

BB4_43:
	mov.f32 	%f175, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f220, %f175, %f217;
	mov.f32 	%f221, %f218;
	mov.f32 	%f219, %f216;
	bra.uni 	BB4_46;

BB4_44:
	mov.f32 	%f176, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f219, %f176, %f216;
	mov.f32 	%f221, %f218;
	mov.f32 	%f220, %f217;
	bra.uni 	BB4_46;

BB4_45:
	mov.f32 	%f177, 0f3F800000;
	.loc 1 36 1
	sub.ftz.f32 	%f219, %f177, %f216;
	sub.ftz.f32 	%f220, %f177, %f217;
	sub.ftz.f32 	%f221, %f177, %f218;

BB4_46:
	.loc 1 36 1
	setp.eq.s32	%p45, %r9, 15;
	@%p45 bra 	BB4_48;

	.loc 1 36 1
	sub.ftz.f32 	%f178, %f216, %f219;
	fma.rn.ftz.f32 	%f216, %f178, %f78, %f219;
	sub.ftz.f32 	%f179, %f217, %f220;
	fma.rn.ftz.f32 	%f217, %f179, %f78, %f220;
	sub.ftz.f32 	%f180, %f218, %f221;
	fma.rn.ftz.f32 	%f218, %f180, %f78, %f221;

BB4_48:
	.loc 1 36 1
	mad.lo.s32 	%r5, %r2, %r7, %r1;
	.loc 1 36 1
	@%p4 bra 	BB4_50;

	cvta.to.global.u64 	%rd8, %rd2;
	mul.wide.s32 	%rd9, %r5, 16;
	add.s64 	%rd10, %rd8, %rd9;
	.loc 1 36 1
	st.global.v4.f32 	[%rd10], {%f218, %f217, %f216, %f198};
	bra.uni 	BB4_51;

BB4_50:
	cvta.to.global.u64 	%rd11, %rd2;
	mul.wide.s32 	%rd12, %r5, 8;
	add.s64 	%rd13, %rd11, %rd12;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f218;
	mov.b16 	%rs9, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f217;
	mov.b16 	%rs10, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f216;
	mov.b16 	%rs11, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f198;
	mov.b16 	%rs12, %temp;
}
	.loc 1 36 232
	st.global.v4.u16 	[%rd13], {%rs9, %rs10, %rs11, %rs12};

BB4_51:
	.loc 1 36 2
	ret;
}


