//
// 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;
// ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local has been demoted

.visible .entry ShaderKernel_PrColorHighlightRecovery(
	.param .u64 ShaderKernel_PrColorHighlightRecovery_param_0,
	.param .u32 ShaderKernel_PrColorHighlightRecovery_param_1,
	.param .u32 ShaderKernel_PrColorHighlightRecovery_param_2,
	.param .u32 ShaderKernel_PrColorHighlightRecovery_param_3,
	.param .u32 ShaderKernel_PrColorHighlightRecovery_param_4,
	.param .u64 ShaderKernel_PrColorHighlightRecovery_param_5,
	.param .u64 ShaderKernel_PrColorHighlightRecovery_param_6,
	.param .u64 ShaderKernel_PrColorHighlightRecovery_param_7
)
{
	.reg .pred 	%p<30>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<15>;
	.reg .f32 	%f<140>;
	.reg .s64 	%rd<18>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local[32];

	ld.param.u64 	%rd3, [ShaderKernel_PrColorHighlightRecovery_param_0];
	ld.param.u32 	%r4, [ShaderKernel_PrColorHighlightRecovery_param_1];
	ld.param.u32 	%r5, [ShaderKernel_PrColorHighlightRecovery_param_2];
	ld.param.u32 	%r6, [ShaderKernel_PrColorHighlightRecovery_param_3];
	ld.param.u32 	%r7, [ShaderKernel_PrColorHighlightRecovery_param_4];
	ld.param.u64 	%rd4, [ShaderKernel_PrColorHighlightRecovery_param_5];
	ld.param.u64 	%rd5, [ShaderKernel_PrColorHighlightRecovery_param_7];
	mov.u32 	%r8, %ntid.x;
	mov.u32 	%r9, %ctaid.x;
	mov.u32 	%r1, %tid.x;
	mad.lo.s32 	%r2, %r8, %r9, %r1;
	mov.u32 	%r10, %ntid.y;
	mov.u32 	%r11, %ctaid.y;
	mov.u32 	%r12, %tid.y;
	mad.lo.s32 	%r3, %r10, %r11, %r12;
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB0_18;
	bra.uni 	BB0_1;

BB0_1:
	setp.gt.u32	%p4, %r1, 1;
	@%p4 bra 	BB0_3;

	cvta.to.global.u64 	%rd6, %rd4;
	mul.wide.u32 	%rd7, %r1, 16;
	mov.u64 	%rd8, ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local;
	add.s64 	%rd9, %rd8, %rd7;
	add.s64 	%rd10, %rd6, %rd7;
	ld.global.v4.f32 	{%f39, %f40, %f41, %f42}, [%rd10];
	st.shared.v4.f32 	[%rd9], {%f39, %f40, %f41, %f42};

BB0_3:
	cvta.to.global.u64 	%rd1, %rd5;
	cvt.rn.f32.s32	%f47, %r2;
	add.ftz.f32 	%f1, %f47, 0f3F000000;
	cvt.rn.f32.s32	%f48, %r3;
	add.ftz.f32 	%f2, %f48, 0f3F000000;
	bar.sync 	0;
	// inline asm
	tex.2d.v4.f32.f32 {%f49, %f50, %f51, %f52}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	abs.ftz.f32 	%f55, %f51;
	abs.ftz.f32 	%f56, %f50;
	abs.ftz.f32 	%f57, %f49;
	cvt.ftz.sat.f32.f32	%f7, %f55;
	cvt.ftz.sat.f32.f32	%f8, %f56;
	cvt.ftz.sat.f32.f32	%f9, %f57;
	sub.ftz.f32 	%f10, %f55, %f7;
	sub.ftz.f32 	%f11, %f56, %f8;
	sub.ftz.f32 	%f12, %f57, %f9;
	ld.global.u32 	%r13, [%rd1];
	setp.eq.s32	%p5, %r13, 0;
	setp.gt.ftz.f32	%p6, %f7, %f8;
	selp.f32	%f58, %f8, %f7, %p6;
	setp.gt.ftz.f32	%p7, %f58, %f9;
	selp.f32	%f59, %f9, %f58, %p7;
	cvt.ftz.sat.f32.f32	%f13, %f59;
	@%p5 bra 	BB0_5;

	selp.f32	%f60, %f7, %f8, %p6;
	setp.gt.ftz.f32	%p9, %f60, %f9;
	selp.f32	%f61, %f60, %f9, %p9;
	cvt.ftz.sat.f32.f32	%f62, %f61;
	ld.shared.v4.f32 	{%f63, %f64, %f65, %f66}, [ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local+16];
	mul.ftz.f32 	%f68, %f8, %f64;
	fma.rn.ftz.f32 	%f70, %f7, %f63, %f68;
	fma.rn.ftz.f32 	%f72, %f9, %f65, %f70;
	cvt.ftz.sat.f32.f32	%f73, %f72;
	sub.ftz.f32 	%f76, %f73, %f59;
	sub.ftz.f32 	%f77, %f73, %f13;
	sub.ftz.f32 	%f78, %f61, %f73;
	sub.ftz.f32 	%f79, %f62, %f73;
	setp.gt.ftz.f32	%p11, %f78, 0f2EDBE6FF;
	selp.f32	%f80, %f78, 0f2EDBE6FF, %p11;
	mov.f32 	%f81, 0f3F800000;
	div.rn.ftz.f32 	%f82, %f81, %f80;
	setp.gt.ftz.f32	%p12, %f82, 0f00000000;
	selp.f32	%f83, %f82, 0f00000000, %p12;
	mul.ftz.f32 	%f84, %f83, %f79;
	setp.gt.ftz.f32	%p13, %f76, 0f2EDBE6FF;
	selp.f32	%f85, %f76, 0f2EDBE6FF, %p13;
	div.rn.ftz.f32 	%f86, %f81, %f85;
	setp.gt.ftz.f32	%p14, %f86, 0f00000000;
	selp.f32	%f87, %f86, 0f00000000, %p14;
	mul.ftz.f32 	%f88, %f87, %f77;
	setp.gt.ftz.f32	%p15, %f88, %f84;
	selp.f32	%f89, %f84, %f88, %p15;
	sub.ftz.f32 	%f90, %f7, %f73;
	sub.ftz.f32 	%f91, %f8, %f73;
	sub.ftz.f32 	%f92, %f9, %f73;
	fma.rn.ftz.f32 	%f134, %f90, %f89, %f73;
	fma.rn.ftz.f32 	%f135, %f91, %f89, %f73;
	fma.rn.ftz.f32 	%f136, %f92, %f89, %f73;
	bra.uni 	BB0_6;

BB0_5:
	ld.shared.v4.f32 	{%f93, %f94, %f95, %f96}, [ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local+16];
	mul.ftz.f32 	%f98, %f8, %f94;
	fma.rn.ftz.f32 	%f100, %f7, %f93, %f98;
	fma.rn.ftz.f32 	%f102, %f9, %f95, %f100;
	cvt.ftz.sat.f32.f32	%f103, %f102;
	sub.ftz.f32 	%f106, %f103, %f59;
	sub.ftz.f32 	%f107, %f103, %f13;
	setp.gt.ftz.f32	%p18, %f106, 0f2EDBE6FF;
	selp.f32	%f108, %f106, 0f2EDBE6FF, %p18;
	mov.f32 	%f109, 0f3F800000;
	div.rn.ftz.f32 	%f110, %f109, %f108;
	setp.gt.ftz.f32	%p19, %f110, 0f00000000;
	selp.f32	%f111, %f110, 0f00000000, %p19;
	mul.ftz.f32 	%f112, %f111, %f107;
	sub.ftz.f32 	%f113, %f7, %f103;
	sub.ftz.f32 	%f114, %f8, %f103;
	sub.ftz.f32 	%f115, %f9, %f103;
	fma.rn.ftz.f32 	%f134, %f113, %f112, %f103;
	fma.rn.ftz.f32 	%f135, %f114, %f112, %f103;
	fma.rn.ftz.f32 	%f136, %f115, %f112, %f103;

BB0_6:
	fma.rn.ftz.f32 	%f116, %f10, %f134, %f134;
	fma.rn.ftz.f32 	%f117, %f11, %f135, %f135;
	fma.rn.ftz.f32 	%f118, %f12, %f136, %f136;
	setp.lt.ftz.f32	%p20, %f51, 0f00000000;
	selp.f32	%f119, 0fBF800000, 0f3F800000, %p20;
	mul.ftz.f32 	%f23, %f116, %f119;
	setp.lt.ftz.f32	%p21, %f50, 0f00000000;
	selp.f32	%f120, 0fBF800000, 0f3F800000, %p21;
	mul.ftz.f32 	%f24, %f117, %f120;
	setp.lt.ftz.f32	%p22, %f49, 0f00000000;
	selp.f32	%f121, 0fBF800000, 0f3F800000, %p22;
	mul.ftz.f32 	%f25, %f118, %f121;
	abs.ftz.f32 	%f26, %f23;
	abs.ftz.f32 	%f27, %f24;
	abs.ftz.f32 	%f28, %f25;
	ld.shared.f32 	%f29, [ShaderKernel_PrColorHighlightRecovery$__cuda_local_var_180681_502_non_const_p_local];
	setp.gtu.ftz.f32	%p23, %f26, 0f00000000;
	@%p23 bra 	BB0_8;

	mov.f32 	%f137, 0f00000000;
	bra.uni 	BB0_9;

BB0_8:
	lg2.approx.ftz.f32 	%f123, %f26;
	mul.ftz.f32 	%f124, %f123, %f29;
	ex2.approx.ftz.f32 	%f137, %f124;

BB0_9:
	setp.gtu.ftz.f32	%p24, %f27, 0f00000000;
	@%p24 bra 	BB0_11;

	mov.f32 	%f138, 0f00000000;
	bra.uni 	BB0_12;

BB0_11:
	lg2.approx.ftz.f32 	%f126, %f27;
	mul.ftz.f32 	%f127, %f126, %f29;
	ex2.approx.ftz.f32 	%f138, %f127;

BB0_12:
	setp.gtu.ftz.f32	%p25, %f28, 0f00000000;
	@%p25 bra 	BB0_14;

	mov.f32 	%f139, 0f00000000;
	bra.uni 	BB0_15;

BB0_14:
	lg2.approx.ftz.f32 	%f129, %f28;
	mul.ftz.f32 	%f130, %f129, %f29;
	ex2.approx.ftz.f32 	%f139, %f130;

BB0_15:
	setp.lt.ftz.f32	%p26, %f23, 0f00000000;
	selp.f32	%f131, 0fBF800000, 0f3F800000, %p26;
	setp.lt.ftz.f32	%p27, %f24, 0f00000000;
	selp.f32	%f132, 0fBF800000, 0f3F800000, %p27;
	setp.lt.ftz.f32	%p28, %f25, 0f00000000;
	selp.f32	%f133, 0fBF800000, 0f3F800000, %p28;
	mul.ftz.f32 	%f36, %f139, %f133;
	mul.ftz.f32 	%f37, %f138, %f132;
	mul.ftz.f32 	%f38, %f137, %f131;
	mad.lo.s32 	%r14, %r3, %r4, %r2;
	cvt.s64.s32	%rd2, %r14;
	setp.eq.s32	%p29, %r5, 0;
	@%p29 bra 	BB0_17;

	cvta.to.global.u64 	%rd12, %rd3;
	shl.b64 	%rd13, %rd2, 4;
	add.s64 	%rd14, %rd12, %rd13;
	st.global.v4.f32 	[%rd14], {%f36, %f37, %f38, %f52};
	bra.uni 	BB0_18;

BB0_17:
	cvta.to.global.u64 	%rd15, %rd3;
	shl.b64 	%rd16, %rd2, 3;
	add.s64 	%rd17, %rd15, %rd16;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f38;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f37;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f36;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f52;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd17], {%rs3, %rs2, %rs1, %rs4};

BB0_18:
	ret;
}


