//
// Generated by NVIDIA NVVM Compiler
// Compiler built on Fri Jul 25 04:36:16 2014 (1406288176)
// Cuda compilation tools, release 6.5, V6.5.13
//

.version 4.1
.target sm_30
.address_size 64

.global .texref texture0_RECT;
.global .texref texture2_2D;
// ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local has been demoted

.visible .entry ShaderKernel_PrColorWheels(
	.param .u64 ShaderKernel_PrColorWheels_param_0,
	.param .u32 ShaderKernel_PrColorWheels_param_1,
	.param .u32 ShaderKernel_PrColorWheels_param_2,
	.param .u32 ShaderKernel_PrColorWheels_param_3,
	.param .u32 ShaderKernel_PrColorWheels_param_4,
	.param .u64 ShaderKernel_PrColorWheels_param_5,
	.param .u64 ShaderKernel_PrColorWheels_param_6,
	.param .u64 ShaderKernel_PrColorWheels_param_7,
	.param .u64 ShaderKernel_PrColorWheels_param_8
)
{
	.reg .pred 	%p<17>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<23>;
	.reg .f32 	%f<170>;
	.reg .s64 	%rd<24>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local[112];

	ld.param.u64 	%rd3, [ShaderKernel_PrColorWheels_param_0];
	ld.param.u32 	%r5, [ShaderKernel_PrColorWheels_param_1];
	ld.param.u32 	%r6, [ShaderKernel_PrColorWheels_param_2];
	ld.param.u32 	%r7, [ShaderKernel_PrColorWheels_param_3];
	ld.param.u32 	%r8, [ShaderKernel_PrColorWheels_param_4];
	ld.param.u64 	%rd4, [ShaderKernel_PrColorWheels_param_5];
	ld.param.u64 	%rd5, [ShaderKernel_PrColorWheels_param_8];
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r1, %tid.x;
	mad.lo.s32 	%r2, %r9, %r10, %r1;
	mov.u32 	%r11, %ntid.y;
	mov.u32 	%r12, %ctaid.y;
	mov.u32 	%r13, %tid.y;
	mad.lo.s32 	%r3, %r11, %r12, %r13;
	setp.lt.s32	%p1, %r2, %r7;
	setp.lt.s32	%p2, %r3, %r8;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB0_17;
	bra.uni 	BB0_1;

BB0_1:
	setp.gt.u32	%p4, %r1, 6;
	@%p4 bra 	BB0_3;

	cvta.to.global.u64 	%rd6, %rd4;
	mul.wide.u32 	%rd7, %r1, 16;
	mov.u64 	%rd8, ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local;
	add.s64 	%rd9, %rd8, %rd7;
	add.s64 	%rd10, %rd6, %rd7;
	ld.global.v4.f32 	{%f26, %f27, %f28, %f29}, [%rd10];
	st.shared.v4.f32 	[%rd9], {%f26, %f27, %f28, %f29};

BB0_3:
	cvta.to.global.u64 	%rd1, %rd5;
	cvt.rn.f32.s32	%f34, %r2;
	add.ftz.f32 	%f1, %f34, 0f3F000000;
	cvt.rn.f32.s32	%f35, %r3;
	add.ftz.f32 	%f2, %f35, 0f3F000000;
	bar.sync 	0;
	// inline asm
	tex.2d.v4.f32.f32 {%f36, %f37, %f38, %f39}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	ld.global.u32 	%r4, [%rd1];
	setp.eq.s32	%p5, %r4, 0;
	ld.shared.f32 	%f42, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+64];
	mul.ftz.f32 	%f43, %f42, %f38;
	mul.ftz.f32 	%f44, %f42, %f37;
	mul.ftz.f32 	%f45, %f42, %f36;
	selp.f32	%f7, %f38, %f43, %p5;
	selp.f32	%f8, %f37, %f44, %p5;
	selp.f32	%f9, %f36, %f45, %p5;
	setp.gtu.ftz.f32	%p6, %f7, 0f00000000;
	@%p6 bra 	BB0_5;

	mov.f32 	%f164, 0f3A000000;
	bra.uni 	BB0_6;

BB0_5:
	lg2.approx.ftz.f32 	%f47, %f7;
	mul.ftz.f32 	%f48, %f47, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f49, %f48;
	fma.rn.ftz.f32 	%f164, %f49, 0f3F7FF000, 0f3A000000;

BB0_6:
	setp.gtu.ftz.f32	%p7, %f8, 0f00000000;
	@%p7 bra 	BB0_8;

	mov.f32 	%f165, 0f3A000000;
	bra.uni 	BB0_9;

BB0_8:
	lg2.approx.ftz.f32 	%f51, %f8;
	mul.ftz.f32 	%f52, %f51, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f53, %f52;
	fma.rn.ftz.f32 	%f165, %f53, 0f3F7FF000, 0f3A000000;

BB0_9:
	setp.gtu.ftz.f32	%p8, %f9, 0f00000000;
	@%p8 bra 	BB0_11;

	mov.f32 	%f166, 0f3A000000;
	bra.uni 	BB0_12;

BB0_11:
	lg2.approx.ftz.f32 	%f55, %f9;
	mul.ftz.f32 	%f56, %f55, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f57, %f56;
	fma.rn.ftz.f32 	%f166, %f57, 0f3F7FF000, 0f3A000000;

BB0_12:
	add.ftz.f32 	%f76, %f7, 0fBF800000;
	add.ftz.f32 	%f77, %f8, 0fBF800000;
	add.ftz.f32 	%f78, %f9, 0fBF800000;
	mov.f32 	%f79, 0f00000000;
	fma.rn.ftz.f32 	%f63, %f79, %f80, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f58, %f59, %f60, %f61}, [texture2_2D, {%f164, %f63}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f64, %f65, %f66, %f67}, [texture2_2D, {%f165, %f63}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f70, %f71, %f72, %f73}, [texture2_2D, {%f166, %f63}];
	// inline asm
	ld.shared.v4.f32 	{%f81, %f82, %f83, %f84}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local];
	ld.shared.v4.f32 	{%f85, %f86, %f87, %f88}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+16];
	fma.rn.ftz.f32 	%f91, %f7, %f81, %f85;
	fma.rn.ftz.f32 	%f94, %f8, %f82, %f86;
	fma.rn.ftz.f32 	%f97, %f9, %f83, %f87;
	ld.shared.v4.f32 	{%f98, %f99, %f100, %f101}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+32];
	ld.shared.v4.f32 	{%f102, %f103, %f104, %f105}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+48];
	fma.rn.ftz.f32 	%f108, %f7, %f98, %f102;
	fma.rn.ftz.f32 	%f111, %f8, %f99, %f103;
	fma.rn.ftz.f32 	%f114, %f9, %f100, %f104;
	setp.lt.ftz.f32	%p9, %f7, 0f00000000;
	selp.f32	%f115, %f91, %f58, %p9;
	setp.lt.ftz.f32	%p10, %f8, 0f00000000;
	selp.f32	%f116, %f94, %f65, %p10;
	setp.lt.ftz.f32	%p11, %f9, 0f00000000;
	selp.f32	%f117, %f97, %f72, %p11;
	setp.lt.ftz.f32	%p12, %f76, 0f00000000;
	selp.f32	%f167, %f115, %f108, %p12;
	setp.lt.ftz.f32	%p13, %f77, 0f00000000;
	selp.f32	%f168, %f116, %f111, %p13;
	setp.lt.ftz.f32	%p14, %f78, 0f00000000;
	selp.f32	%f169, %f117, %f114, %p14;
	@%p5 bra 	BB0_14;

	add.ftz.f32 	%f118, %f7, 0fBF4CCCCD;
	mul.ftz.f32 	%f119, %f118, 0f40A00000;
	cvt.ftz.sat.f32.f32	%f120, %f119;
	add.ftz.f32 	%f121, %f8, 0fBF4CCCCD;
	mul.ftz.f32 	%f122, %f121, 0f40A00000;
	cvt.ftz.sat.f32.f32	%f123, %f122;
	add.ftz.f32 	%f124, %f9, 0fBF4CCCCD;
	mul.ftz.f32 	%f125, %f124, 0f40A00000;
	cvt.ftz.sat.f32.f32	%f126, %f125;
	fma.rn.ftz.f32 	%f127, %f120, 0f3DCCCCCD, 0f3F4CCCCD;
	fma.rn.ftz.f32 	%f128, %f123, 0f3DCCCCCD, 0f3F4CCCCD;
	fma.rn.ftz.f32 	%f129, %f126, 0f3DCCCCCD, 0f3F4CCCCD;
	ld.shared.v4.f32 	{%f130, %f131, %f132, %f133}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+96];
	mul.ftz.f32 	%f135, %f127, %f130;
	mul.ftz.f32 	%f137, %f128, %f131;
	mul.ftz.f32 	%f139, %f129, %f132;
	sub.ftz.f32 	%f140, %f167, %f135;
	sub.ftz.f32 	%f141, %f168, %f137;
	sub.ftz.f32 	%f142, %f169, %f139;
	ld.shared.v4.f32 	{%f143, %f144, %f145, %f146}, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+80];
	fma.rn.ftz.f32 	%f148, %f140, %f143, %f135;
	fma.rn.ftz.f32 	%f150, %f141, %f144, %f137;
	fma.rn.ftz.f32 	%f152, %f142, %f145, %f139;
	mov.f32 	%f153, 0f3F800000;
	sub.ftz.f32 	%f154, %f153, %f120;
	mul.ftz.f32 	%f155, %f154, %f167;
	fma.rn.ftz.f32 	%f156, %f120, %f148, %f155;
	sub.ftz.f32 	%f157, %f153, %f123;
	mul.ftz.f32 	%f158, %f157, %f168;
	fma.rn.ftz.f32 	%f159, %f123, %f150, %f158;
	sub.ftz.f32 	%f160, %f153, %f126;
	mul.ftz.f32 	%f161, %f160, %f169;
	fma.rn.ftz.f32 	%f162, %f126, %f152, %f161;
	ld.shared.f32 	%f163, [ShaderKernel_PrColorWheels$__cuda_local_var_180686_584_non_const_p_local+68];
	mul.ftz.f32 	%f167, %f163, %f156;
	mul.ftz.f32 	%f168, %f163, %f159;
	mul.ftz.f32 	%f169, %f163, %f162;

BB0_14:
	mad.lo.s32 	%r22, %r3, %r5, %r2;
	cvt.s64.s32	%rd2, %r22;
	setp.eq.s32	%p16, %r6, 0;
	@%p16 bra 	BB0_16;

	cvta.to.global.u64 	%rd18, %rd3;
	shl.b64 	%rd19, %rd2, 4;
	add.s64 	%rd20, %rd18, %rd19;
	st.global.v4.f32 	[%rd20], {%f169, %f168, %f167, %f39};
	bra.uni 	BB0_17;

BB0_16:
	cvta.to.global.u64 	%rd21, %rd3;
	shl.b64 	%rd22, %rd2, 3;
	add.s64 	%rd23, %rd21, %rd22;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f39;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f167;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f168;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f169;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd23], {%rs4, %rs3, %rs2, %rs1};

BB0_17:
	ret;
}


