//
// 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<171>;
	.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 	{%f19, %f20, %f21, %f22}, [%rd11];
	st.shared.v4.f32 	[%rd10], {%f19, %f20, %f21, %f22};

BB0_3:
	cvta.to.global.u64 	%rd2, %rd5;
	cvt.rn.f32.s32	%f27, %r2;
	add.ftz.f32 	%f1, %f27, 0f3F000000;
	cvt.rn.f32.s32	%f28, %r3;
	add.ftz.f32 	%f2, %f28, 0f3F000000;
	bar.sync 	0;
	// inline asm
	tex.2d.v4.f32.f32 {%f29, %f30, %f31, %f32}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	cvt.ftz.sat.f32.f32	%f35, %f31;
	cvt.ftz.sat.f32.f32	%f36, %f30;
	cvt.ftz.sat.f32.f32	%f37, %f29;
	mul.ftz.f32 	%f38, %f36, 0f3F140B78;
	fma.rn.ftz.f32 	%f39, %f35, 0f3EC31F8A, %f38;
	fma.rn.ftz.f32 	%f40, %f37, 0f3D24A8C1, %f39;
	mul.ftz.f32 	%f41, %f36, 0f3F397247;
	fma.rn.ftz.f32 	%f42, %f35, 0f3E496BBA, %f41;
	fma.rn.ftz.f32 	%f43, %f37, 0f3DA02752, %f42;
	mul.ftz.f32 	%f44, %f36, 0f3E03E426;
	fma.rn.ftz.f32 	%f45, %f35, 0f3CC56D5D, %f44;
	fma.rn.ftz.f32 	%f46, %f37, 0f3F582A99, %f45;
	add.ftz.f32 	%f47, %f40, 0f3F800000;
	add.ftz.f32 	%f48, %f43, 0f3F800000;
	add.ftz.f32 	%f49, %f46, 0f3F800000;
	lg2.approx.ftz.f32 	%f50, %f47;
	lg2.approx.ftz.f32 	%f51, %f48;
	lg2.approx.ftz.f32 	%f52, %f49;
	mul.ftz.f32 	%f53, %f50, 0f3E9A209B;
	mul.ftz.f32 	%f54, %f51, 0f3E9A209B;
	mul.ftz.f32 	%f55, %f52, 0f3E9A209B;
	add.ftz.f32 	%f56, %f53, %f54;
	add.ftz.f32 	%f57, %f56, %f55;
	fma.rn.ftz.f32 	%f58, %f55, 0fC0000000, %f56;
	sub.ftz.f32 	%f59, %f53, %f54;
	mul.ftz.f32 	%f4, %f57, 0f3F13CD3A;
	mul.ftz.f32 	%f5, %f58, 0f3ED105EC;
	mul.ftz.f32 	%f6, %f59, 0f3F3504F3;
	ld.global.u32 	%r13, [%rd2];
	setp.eq.s32	%p5, %r13, 0;
	@%p5 bra 	BB0_5;

	fma.rn.ftz.f32 	%f60, %f4, 0fC0000000, 0f3F800000;
	cvt.ftz.sat.f32.f32	%f61, %f60;
	fma.rn.ftz.f32 	%f62, %f4, 0f00000000, 0f3F800000;
	cvt.ftz.sat.f32.f32	%f63, %f62;
	fma.rn.ftz.f32 	%f64, %f4, 0f40000000, 0fBF800000;
	cvt.ftz.sat.f32.f32	%f65, %f64;
	fma.rn.ftz.f32 	%f66, %f61, 0fBF800000, %f63;
	fma.rn.ftz.f32 	%f67, %f65, 0fBF800000, %f66;
	cvt.ftz.sat.f32.f32	%f68, %f67;
	ld.shared.v4.f32 	{%f69, %f70, %f71, %f72}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local];
	ld.shared.v4.f32 	{%f73, %f74, %f75, %f76}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+48];
	fma.rn.ftz.f32 	%f79, %f4, %f69, %f73;
	fma.rn.ftz.f32 	%f82, %f5, %f70, %f74;
	fma.rn.ftz.f32 	%f85, %f6, %f71, %f75;
	mul.ftz.f32 	%f86, %f61, %f79;
	mul.ftz.f32 	%f87, %f61, %f82;
	mul.ftz.f32 	%f88, %f61, %f85;
	ld.shared.v4.f32 	{%f89, %f90, %f91, %f92}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+16];
	ld.shared.v4.f32 	{%f93, %f94, %f95, %f96}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+64];
	fma.rn.ftz.f32 	%f99, %f4, %f89, %f93;
	fma.rn.ftz.f32 	%f102, %f5, %f90, %f94;
	fma.rn.ftz.f32 	%f105, %f6, %f91, %f95;
	fma.rn.ftz.f32 	%f106, %f68, %f99, %f86;
	fma.rn.ftz.f32 	%f107, %f68, %f102, %f87;
	fma.rn.ftz.f32 	%f108, %f68, %f105, %f88;
	ld.shared.v4.f32 	{%f109, %f110, %f111, %f112}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+32];
	ld.shared.v4.f32 	{%f113, %f114, %f115, %f116}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+80];
	fma.rn.ftz.f32 	%f119, %f4, %f109, %f113;
	fma.rn.ftz.f32 	%f122, %f5, %f110, %f114;
	fma.rn.ftz.f32 	%f125, %f6, %f111, %f115;
	fma.rn.ftz.f32 	%f168, %f65, %f119, %f106;
	fma.rn.ftz.f32 	%f169, %f65, %f122, %f107;
	fma.rn.ftz.f32 	%f170, %f65, %f125, %f108;
	bra.uni 	BB0_6;

BB0_5:
	ld.shared.v4.f32 	{%f126, %f127, %f128, %f129}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local];
	ld.shared.v4.f32 	{%f130, %f131, %f132, %f133}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_180685_498_non_const_p_local+48];
	fma.rn.ftz.f32 	%f168, %f4, %f126, %f130;
	fma.rn.ftz.f32 	%f169, %f5, %f127, %f131;
	fma.rn.ftz.f32 	%f170, %f6, %f128, %f132;

BB0_6:
	mul.ftz.f32 	%f140, %f169, 0f3ED105EC;
	fma.rn.ftz.f32 	%f141, %f168, 0f3F13CD3A, %f140;
	mul.ftz.f32 	%f142, %f170, 0f3F3504F3;
	add.ftz.f32 	%f143, %f141, %f142;
	fma.rn.ftz.f32 	%f144, %f142, 0fBF800000, %f141;
	mul.ftz.f32 	%f145, %f140, 0fC0000000;
	fma.rn.ftz.f32 	%f146, %f168, 0f3F13CD3A, %f145;
	fma.rn.ftz.f32 	%f147, %f142, 0f00000000, %f146;
	mov.f32 	%f148, 0f41200000;
	lg2.approx.ftz.f32 	%f149, %f148;
	mul.ftz.f32 	%f150, %f149, %f143;
	ex2.approx.ftz.f32 	%f151, %f150;
	mul.ftz.f32 	%f152, %f149, %f144;
	ex2.approx.ftz.f32 	%f153, %f152;
	mul.ftz.f32 	%f154, %f149, %f147;
	ex2.approx.ftz.f32 	%f155, %f154;
	add.ftz.f32 	%f156, %f151, 0fBF800000;
	add.ftz.f32 	%f157, %f153, 0fBF800000;
	add.ftz.f32 	%f158, %f155, 0fBF800000;
	mul.ftz.f32 	%f159, %f157, 0fC0659653;
	fma.rn.ftz.f32 	%f160, %f156, 0f408EF909, %f159;
	fma.rn.ftz.f32 	%f161, %f158, 0f3DF4538F, %f160;
	cvt.ftz.sat.f32.f32	%f16, %f161;
	mul.ftz.f32 	%f162, %f157, 0f401860AA;
	fma.rn.ftz.f32 	%f163, %f156, 0fBF9BFB16, %f162;
	fma.rn.ftz.f32 	%f164, %f158, 0fBE264C30, %f163;
	cvt.ftz.sat.f32.f32	%f17, %f164;
	mul.ftz.f32 	%f165, %f157, 0fBE79C0EC;
	fma.rn.ftz.f32 	%f166, %f156, 0f3D4B923A, %f165;
	fma.rn.ftz.f32 	%f167, %f158, 0f3F9A2D0E, %f166;
	cvt.ftz.sat.f32.f32	%f18, %f167;
	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], {%f18, %f17, %f16, %f32};
	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, %f32;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f16;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f17;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f18;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd16], {%rs4, %rs3, %rs2, %rs1};

BB0_9:
	ret;
}


