//
// 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_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local has been demoted

.visible .entry ShaderKernel_IRIDASAutoColorMatch(
	.param .u64 ShaderKernel_IRIDASAutoColorMatch_param_0,
	.param .u32 ShaderKernel_IRIDASAutoColorMatch_param_1,
	.param .u32 ShaderKernel_IRIDASAutoColorMatch_param_2,
	.param .u32 ShaderKernel_IRIDASAutoColorMatch_param_3,
	.param .u32 ShaderKernel_IRIDASAutoColorMatch_param_4,
	.param .u64 ShaderKernel_IRIDASAutoColorMatch_param_5,
	.param .u64 ShaderKernel_IRIDASAutoColorMatch_param_6,
	.param .u64 ShaderKernel_IRIDASAutoColorMatch_param_7
)
{
	.reg .pred 	%p<7>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<15>;
	.reg .f32 	%f<177>;
	.reg .s64 	%rd<17>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local[96];

	ld.param.u64 	%rd6, [ShaderKernel_IRIDASAutoColorMatch_param_0];
	ld.param.u32 	%r4, [ShaderKernel_IRIDASAutoColorMatch_param_1];
	ld.param.u32 	%r5, [ShaderKernel_IRIDASAutoColorMatch_param_2];
	ld.param.u32 	%r6, [ShaderKernel_IRIDASAutoColorMatch_param_3];
	ld.param.u32 	%r7, [ShaderKernel_IRIDASAutoColorMatch_param_4];
	ld.param.u64 	%rd4, [ShaderKernel_IRIDASAutoColorMatch_param_5];
	ld.param.u64 	%rd5, [ShaderKernel_IRIDASAutoColorMatch_param_7];
	cvta.to.global.u64 	%rd1, %rd6;
	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_9;
	bra.uni 	BB0_1;

BB0_1:
	setp.gt.u32	%p4, %r1, 5;
	@%p4 bra 	BB0_3;

	cvta.to.global.u64 	%rd7, %rd4;
	mul.wide.u32 	%rd8, %r1, 16;
	mov.u64 	%rd9, ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local;
	add.s64 	%rd10, %rd9, %rd8;
	add.s64 	%rd11, %rd7, %rd8;
	ld.global.v4.f32 	{%f22, %f23, %f24, %f25}, [%rd11];
	st.shared.v4.f32 	[%rd10], {%f22, %f23, %f24, %f25};

BB0_3:
	cvta.to.global.u64 	%rd2, %rd5;
	cvt.rn.f32.s32	%f30, %r2;
	add.ftz.f32 	%f1, %f30, 0f3F000000;
	cvt.rn.f32.s32	%f31, %r3;
	add.ftz.f32 	%f2, %f31, 0f3F000000;
	bar.sync 	0;
	// inline asm
	tex.2d.v4.f32.f32 {%f32, %f33, %f34, %f35}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	cvt.ftz.sat.f32.f32	%f38, %f34;
	cvt.ftz.sat.f32.f32	%f39, %f33;
	cvt.ftz.sat.f32.f32	%f40, %f32;
	mul.ftz.f32 	%f41, %f39, 0f3F140B78;
	fma.rn.ftz.f32 	%f42, %f38, 0f3EC31F8A, %f41;
	fma.rn.ftz.f32 	%f43, %f40, 0f3D24A8C1, %f42;
	mul.ftz.f32 	%f44, %f39, 0f3F397247;
	fma.rn.ftz.f32 	%f45, %f38, 0f3E496BBA, %f44;
	fma.rn.ftz.f32 	%f46, %f40, 0f3DA02752, %f45;
	mul.ftz.f32 	%f47, %f39, 0f3E03E426;
	fma.rn.ftz.f32 	%f48, %f38, 0f3CC56D5D, %f47;
	fma.rn.ftz.f32 	%f49, %f40, 0f3F582A99, %f48;
	add.ftz.f32 	%f50, %f43, 0f3F800000;
	add.ftz.f32 	%f51, %f46, 0f3F800000;
	add.ftz.f32 	%f52, %f49, 0f3F800000;
	lg2.approx.ftz.f32 	%f53, %f50;
	lg2.approx.ftz.f32 	%f54, %f51;
	lg2.approx.ftz.f32 	%f55, %f52;
	mul.ftz.f32 	%f56, %f53, 0f3E9A209B;
	mul.ftz.f32 	%f57, %f54, 0f3E9A209B;
	mul.ftz.f32 	%f58, %f55, 0f3E9A209B;
	add.ftz.f32 	%f59, %f56, %f57;
	add.ftz.f32 	%f60, %f59, %f58;
	fma.rn.ftz.f32 	%f61, %f58, 0fC0000000, %f59;
	sub.ftz.f32 	%f62, %f56, %f57;
	mul.ftz.f32 	%f7, %f60, 0f3F13CD3A;
	mul.ftz.f32 	%f8, %f61, 0f3ED105EC;
	mul.ftz.f32 	%f9, %f62, 0f3F3504F3;
	ld.global.u32 	%r13, [%rd2];
	setp.eq.s32	%p5, %r13, 0;
	@%p5 bra 	BB0_5;

	mul.ftz.f32 	%f63, %f33, 0f3F1645A2;
	fma.rn.ftz.f32 	%f64, %f34, 0f3E991687, %f63;
	fma.rn.ftz.f32 	%f65, %f32, 0f3DE978D5, %f64;
	fma.rn.ftz.f32 	%f66, %f65, 0fC0000000, 0f3F800000;
	cvt.ftz.sat.f32.f32	%f67, %f66;
	fma.rn.ftz.f32 	%f68, %f65, 0f00000000, 0f3F800000;
	cvt.ftz.sat.f32.f32	%f69, %f68;
	fma.rn.ftz.f32 	%f70, %f65, 0f40000000, 0fBF800000;
	cvt.ftz.sat.f32.f32	%f71, %f70;
	fma.rn.ftz.f32 	%f72, %f67, 0fBF800000, %f69;
	fma.rn.ftz.f32 	%f73, %f71, 0fBF800000, %f72;
	cvt.ftz.sat.f32.f32	%f74, %f73;
	ld.shared.v4.f32 	{%f75, %f76, %f77, %f78}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local];
	ld.shared.v4.f32 	{%f79, %f80, %f81, %f82}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+48];
	fma.rn.ftz.f32 	%f85, %f7, %f75, %f79;
	fma.rn.ftz.f32 	%f88, %f8, %f76, %f80;
	fma.rn.ftz.f32 	%f91, %f9, %f77, %f81;
	mul.ftz.f32 	%f92, %f67, %f85;
	mul.ftz.f32 	%f93, %f67, %f88;
	mul.ftz.f32 	%f94, %f67, %f91;
	ld.shared.v4.f32 	{%f95, %f96, %f97, %f98}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+16];
	ld.shared.v4.f32 	{%f99, %f100, %f101, %f102}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+64];
	fma.rn.ftz.f32 	%f105, %f7, %f95, %f99;
	fma.rn.ftz.f32 	%f108, %f8, %f96, %f100;
	fma.rn.ftz.f32 	%f111, %f9, %f97, %f101;
	fma.rn.ftz.f32 	%f112, %f74, %f105, %f92;
	fma.rn.ftz.f32 	%f113, %f74, %f108, %f93;
	fma.rn.ftz.f32 	%f114, %f74, %f111, %f94;
	ld.shared.v4.f32 	{%f115, %f116, %f117, %f118}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+32];
	ld.shared.v4.f32 	{%f119, %f120, %f121, %f122}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+80];
	fma.rn.ftz.f32 	%f125, %f7, %f115, %f119;
	fma.rn.ftz.f32 	%f128, %f8, %f116, %f120;
	fma.rn.ftz.f32 	%f131, %f9, %f117, %f121;
	fma.rn.ftz.f32 	%f174, %f71, %f125, %f112;
	fma.rn.ftz.f32 	%f175, %f71, %f128, %f113;
	fma.rn.ftz.f32 	%f176, %f71, %f131, %f114;
	bra.uni 	BB0_6;

BB0_5:
	ld.shared.v4.f32 	{%f132, %f133, %f134, %f135}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local];
	ld.shared.v4.f32 	{%f136, %f137, %f138, %f139}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+48];
	fma.rn.ftz.f32 	%f174, %f7, %f132, %f136;
	fma.rn.ftz.f32 	%f175, %f8, %f133, %f137;
	fma.rn.ftz.f32 	%f176, %f9, %f134, %f138;

BB0_6:
	mul.ftz.f32 	%f146, %f175, 0f3ED105EC;
	fma.rn.ftz.f32 	%f147, %f174, 0f3F13CD3A, %f146;
	mul.ftz.f32 	%f148, %f176, 0f3F3504F3;
	add.ftz.f32 	%f149, %f147, %f148;
	fma.rn.ftz.f32 	%f150, %f148, 0fBF800000, %f147;
	mul.ftz.f32 	%f151, %f146, 0fC0000000;
	fma.rn.ftz.f32 	%f152, %f174, 0f3F13CD3A, %f151;
	fma.rn.ftz.f32 	%f153, %f148, 0f00000000, %f152;
	mov.f32 	%f154, 0f41200000;
	lg2.approx.ftz.f32 	%f155, %f154;
	mul.ftz.f32 	%f156, %f155, %f149;
	ex2.approx.ftz.f32 	%f157, %f156;
	mul.ftz.f32 	%f158, %f155, %f150;
	ex2.approx.ftz.f32 	%f159, %f158;
	mul.ftz.f32 	%f160, %f155, %f153;
	ex2.approx.ftz.f32 	%f161, %f160;
	add.ftz.f32 	%f162, %f157, 0fBF800000;
	add.ftz.f32 	%f163, %f159, 0fBF800000;
	add.ftz.f32 	%f164, %f161, 0fBF800000;
	mul.ftz.f32 	%f165, %f163, 0fC0659653;
	fma.rn.ftz.f32 	%f166, %f162, 0f408EF909, %f165;
	fma.rn.ftz.f32 	%f167, %f164, 0f3DF4538F, %f166;
	cvt.ftz.sat.f32.f32	%f19, %f167;
	mul.ftz.f32 	%f168, %f163, 0f401860AA;
	fma.rn.ftz.f32 	%f169, %f162, 0fBF9BFB16, %f168;
	fma.rn.ftz.f32 	%f170, %f164, 0fBE264C30, %f169;
	cvt.ftz.sat.f32.f32	%f20, %f170;
	mul.ftz.f32 	%f171, %f163, 0fBE79C0EC;
	fma.rn.ftz.f32 	%f172, %f162, 0f3D4B923A, %f171;
	fma.rn.ftz.f32 	%f173, %f164, 0f3F9A2D0E, %f172;
	cvt.ftz.sat.f32.f32	%f21, %f173;
	mad.lo.s32 	%r14, %r3, %r4, %r2;
	cvt.s64.s32	%rd3, %r14;
	setp.eq.s32	%p6, %r5, 0;
	@%p6 bra 	BB0_8;

	shl.b64 	%rd13, %rd3, 4;
	add.s64 	%rd14, %rd1, %rd13;
	st.global.v4.f32 	[%rd14], {%f21, %f20, %f19, %f35};
	bra.uni 	BB0_9;

BB0_8:
	shl.b64 	%rd15, %rd3, 3;
	add.s64 	%rd16, %rd1, %rd15;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f35;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f19;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f20;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f21;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd16], {%rs4, %rs3, %rs2, %rs1};

BB0_9:
	ret;
}


