//
// 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 "C:/ame2014Iupdate/releases/2014.03/external/adobe/Iridas/IRIDASLIB/GPU/IRIDASStereoColorMatch.cu", 1405709992, 4282
	.file	2 "c:\\ame2014iupdate\\releases\\2014.03\\external\\adobe\\mediacore\\external\\3rdparty\\nvidia\\cuda\\win\\include\\device_functions.h", 1405710127, 191626
.global .texref texture0_RECT;
// ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local has been demoted
.global .align 1 .b8 $str[11] = {95, 95, 67, 85, 68, 65, 95, 70, 84, 90, 0};

.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<175>;
	.reg .s64 	%rd<17>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local[96];

	ld.param.u64 	%rd5, [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 	%rd6, [ShaderKernel_IRIDASAutoColorMatch_param_5];
	ld.param.u64 	%rd7, [ShaderKernel_IRIDASAutoColorMatch_param_7];
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd2, %rd7;
	cvta.to.global.u64 	%rd3, %rd6;
	.loc 1 40 1
	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;
	.loc 1 40 1
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 40 1
	@!%p3 bra 	BB0_9;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 40 1
	cvt.rn.f32.s32	%f23, %r2;
	add.ftz.f32 	%f1, %f23, 0f3F000000;
	cvt.rn.f32.s32	%f24, %r3;
	add.ftz.f32 	%f2, %f24, 0f3F000000;
	.loc 1 40 1
	setp.gt.u32	%p4, %r1, 5;
	@%p4 bra 	BB0_3;

	.loc 1 40 1
	mul.wide.u32 	%rd8, %r1, 16;
	mov.u64 	%rd9, ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local;
	add.s64 	%rd10, %rd9, %rd8;
	add.s64 	%rd11, %rd3, %rd8;
	ld.global.v4.f32 	{%f25, %f26, %f27, %f28}, [%rd11];
	st.shared.v4.f32 	[%rd10], {%f25, %f26, %f27, %f28};

BB0_3:
	.loc 1 40 1
	bar.sync 	0;
	.loc 1 40 109
	// inline asm
	tex.2d.v4.f32.f32 {%f33, %f34, %f35, %f36}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f39, %f35;
	cvt.ftz.sat.f32.f32	%f40, %f34;
	cvt.ftz.sat.f32.f32	%f41, %f33;
	.loc 1 40 1
	mul.ftz.f32 	%f42, %f40, 0f3F140B78;
	fma.rn.ftz.f32 	%f43, %f39, 0f3EC31F8A, %f42;
	fma.rn.ftz.f32 	%f44, %f41, 0f3D24A8C1, %f43;
	mul.ftz.f32 	%f45, %f40, 0f3F397247;
	fma.rn.ftz.f32 	%f46, %f39, 0f3E496BBA, %f45;
	fma.rn.ftz.f32 	%f47, %f41, 0f3DA02752, %f46;
	mul.ftz.f32 	%f48, %f40, 0f3E03E426;
	fma.rn.ftz.f32 	%f49, %f39, 0f3CC56D5D, %f48;
	fma.rn.ftz.f32 	%f50, %f41, 0f3F582A99, %f49;
	.loc 1 40 1
	add.ftz.f32 	%f51, %f44, 0f3F800000;
	add.ftz.f32 	%f52, %f47, 0f3F800000;
	add.ftz.f32 	%f53, %f50, 0f3F800000;
	.loc 2 3559 10
	lg2.approx.ftz.f32 	%f54, %f51;
	lg2.approx.ftz.f32 	%f55, %f52;
	lg2.approx.ftz.f32 	%f56, %f53;
	.loc 1 40 1
	mul.ftz.f32 	%f57, %f54, 0f3E9A209B;
	mul.ftz.f32 	%f58, %f55, 0f3E9A209B;
	mul.ftz.f32 	%f59, %f56, 0f3E9A209B;
	.loc 1 40 1
	add.ftz.f32 	%f60, %f57, %f58;
	add.ftz.f32 	%f61, %f60, %f59;
	fma.rn.ftz.f32 	%f62, %f59, 0fC0000000, %f60;
	sub.ftz.f32 	%f63, %f57, %f58;
	.loc 1 40 1
	mul.ftz.f32 	%f4, %f61, 0f3F13CD3A;
	mul.ftz.f32 	%f5, %f62, 0f3ED105EC;
	mul.ftz.f32 	%f6, %f63, 0f3F3504F3;
	.loc 1 40 1
	ld.global.u32 	%r13, [%rd2];
	setp.eq.s32	%p5, %r13, 0;
	@%p5 bra 	BB0_5;

	.loc 1 40 1
	fma.rn.ftz.f32 	%f64, %f4, 0fC0000000, 0f3F800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f65, %f64;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f66, %f4, 0f00000000, 0f3F800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f67, %f66;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f68, %f4, 0f40000000, 0fBF800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f69, %f68;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f70, %f65, 0fBF800000, %f67;
	fma.rn.ftz.f32 	%f71, %f69, 0fBF800000, %f70;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f72, %f71;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f73, %f74, %f75, %f76}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+48];
	ld.shared.v4.f32 	{%f78, %f79, %f80, %f81}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f83, %f4, %f78, %f73;
	fma.rn.ftz.f32 	%f86, %f5, %f79, %f74;
	fma.rn.ftz.f32 	%f89, %f6, %f80, %f75;
	.loc 1 40 1
	mul.ftz.f32 	%f90, %f65, %f83;
	mul.ftz.f32 	%f91, %f65, %f86;
	mul.ftz.f32 	%f92, %f65, %f89;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f93, %f94, %f95, %f96}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+64];
	ld.shared.v4.f32 	{%f98, %f99, %f100, %f101}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+16];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f103, %f4, %f98, %f93;
	fma.rn.ftz.f32 	%f106, %f5, %f99, %f94;
	fma.rn.ftz.f32 	%f109, %f6, %f100, %f95;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f110, %f72, %f103, %f90;
	fma.rn.ftz.f32 	%f111, %f72, %f106, %f91;
	fma.rn.ftz.f32 	%f112, %f72, %f109, %f92;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f113, %f114, %f115, %f116}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+80];
	ld.shared.v4.f32 	{%f118, %f119, %f120, %f121}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+32];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f123, %f4, %f118, %f113;
	fma.rn.ftz.f32 	%f126, %f5, %f119, %f114;
	fma.rn.ftz.f32 	%f129, %f6, %f120, %f115;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f172, %f69, %f123, %f110;
	fma.rn.ftz.f32 	%f173, %f69, %f126, %f111;
	fma.rn.ftz.f32 	%f174, %f69, %f129, %f112;
	bra.uni 	BB0_6;

BB0_5:
	.loc 1 40 1
	ld.shared.v4.f32 	{%f130, %f131, %f132, %f133}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+48];
	ld.shared.v4.f32 	{%f135, %f136, %f137, %f138}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f172, %f4, %f135, %f130;
	fma.rn.ftz.f32 	%f173, %f5, %f136, %f131;
	fma.rn.ftz.f32 	%f174, %f6, %f137, %f132;

BB0_6:
	.loc 1 40 1
	mul.ftz.f32 	%f144, %f173, 0f3ED105EC;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f145, %f172, 0f3F13CD3A, %f144;
	.loc 1 40 1
	mul.ftz.f32 	%f146, %f174, 0f3F3504F3;
	.loc 1 40 1
	add.ftz.f32 	%f147, %f145, %f146;
	fma.rn.ftz.f32 	%f148, %f146, 0fBF800000, %f145;
	mul.ftz.f32 	%f149, %f144, 0fC0000000;
	fma.rn.ftz.f32 	%f150, %f172, 0f3F13CD3A, %f149;
	fma.rn.ftz.f32 	%f151, %f146, 0f00000000, %f150;
	mov.f32 	%f152, 0f41200000;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f153, %f152;
	mul.ftz.f32 	%f154, %f147, %f153;
	ex2.approx.ftz.f32 	%f155, %f154;
	mul.ftz.f32 	%f156, %f148, %f153;
	ex2.approx.ftz.f32 	%f157, %f156;
	mul.ftz.f32 	%f158, %f151, %f153;
	ex2.approx.ftz.f32 	%f159, %f158;
	.loc 1 40 1
	add.ftz.f32 	%f160, %f155, 0fBF800000;
	add.ftz.f32 	%f161, %f157, 0fBF800000;
	add.ftz.f32 	%f162, %f159, 0fBF800000;
	.loc 1 40 1
	mul.ftz.f32 	%f163, %f161, 0fC0659653;
	fma.rn.ftz.f32 	%f164, %f160, 0f408EF909, %f163;
	fma.rn.ftz.f32 	%f165, %f162, 0f3DF4538F, %f164;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f21, %f165;
	.loc 1 40 1
	mul.ftz.f32 	%f166, %f161, 0f401860AA;
	fma.rn.ftz.f32 	%f167, %f160, 0fBF9BFB16, %f166;
	fma.rn.ftz.f32 	%f168, %f162, 0fBE264C30, %f167;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f20, %f168;
	.loc 1 40 1
	mul.ftz.f32 	%f169, %f161, 0fBE79C0EC;
	fma.rn.ftz.f32 	%f170, %f160, 0f3D4B923A, %f169;
	fma.rn.ftz.f32 	%f171, %f162, 0f3F9A2D0E, %f170;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f19, %f171;
	.loc 1 40 1
	mad.lo.s32 	%r14, %r3, %r4, %r2;
	.loc 1 40 1
	cvt.s64.s32	%rd4, %r14;
	.loc 1 40 1
	setp.eq.s32	%p6, %r5, 0;
	@%p6 bra 	BB0_8;

	.loc 1 40 1
	shl.b64 	%rd13, %rd4, 4;
	add.s64 	%rd14, %rd1, %rd13;
	st.global.v4.f32 	[%rd14], {%f19, %f20, %f21, %f36};
	bra.uni 	BB0_9;

BB0_8:
	.loc 1 40 1
	shl.b64 	%rd15, %rd4, 3;
	add.s64 	%rd16, %rd1, %rd15;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f19;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f20;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f21;
	mov.b16 	%rs3, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f36;
	mov.b16 	%rs4, %temp;
}
	.loc 1 40 231
	st.global.v4.u16 	[%rd16], {%rs1, %rs2, %rs3, %rs4};

BB0_9:
	.loc 1 40 2
	ret;
}


