//
// 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/IRIDASAutoColorMatch.cu", 1405709992, 4314
	.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<181>;
	.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	%f26, %r2;
	add.ftz.f32 	%f1, %f26, 0f3F000000;
	cvt.rn.f32.s32	%f27, %r3;
	add.ftz.f32 	%f2, %f27, 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 	{%f28, %f29, %f30, %f31}, [%rd11];
	st.shared.v4.f32 	[%rd10], {%f28, %f29, %f30, %f31};

BB0_3:
	.loc 1 40 1
	bar.sync 	0;
	.loc 1 40 109
	// inline asm
	tex.2d.v4.f32.f32 {%f36, %f37, %f38, %f39}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f42, %f38;
	cvt.ftz.sat.f32.f32	%f43, %f37;
	cvt.ftz.sat.f32.f32	%f44, %f36;
	.loc 1 40 1
	mul.ftz.f32 	%f45, %f43, 0f3F140B78;
	fma.rn.ftz.f32 	%f46, %f42, 0f3EC31F8A, %f45;
	fma.rn.ftz.f32 	%f47, %f44, 0f3D24A8C1, %f46;
	mul.ftz.f32 	%f48, %f43, 0f3F397247;
	fma.rn.ftz.f32 	%f49, %f42, 0f3E496BBA, %f48;
	fma.rn.ftz.f32 	%f50, %f44, 0f3DA02752, %f49;
	mul.ftz.f32 	%f51, %f43, 0f3E03E426;
	fma.rn.ftz.f32 	%f52, %f42, 0f3CC56D5D, %f51;
	fma.rn.ftz.f32 	%f53, %f44, 0f3F582A99, %f52;
	.loc 1 40 1
	add.ftz.f32 	%f54, %f47, 0f3F800000;
	add.ftz.f32 	%f55, %f50, 0f3F800000;
	add.ftz.f32 	%f56, %f53, 0f3F800000;
	.loc 2 3559 10
	lg2.approx.ftz.f32 	%f57, %f54;
	lg2.approx.ftz.f32 	%f58, %f55;
	lg2.approx.ftz.f32 	%f59, %f56;
	.loc 1 40 1
	mul.ftz.f32 	%f60, %f57, 0f3E9A209B;
	mul.ftz.f32 	%f61, %f58, 0f3E9A209B;
	mul.ftz.f32 	%f62, %f59, 0f3E9A209B;
	.loc 1 40 1
	add.ftz.f32 	%f63, %f60, %f61;
	add.ftz.f32 	%f64, %f63, %f62;
	fma.rn.ftz.f32 	%f65, %f62, 0fC0000000, %f63;
	sub.ftz.f32 	%f66, %f60, %f61;
	.loc 1 40 1
	mul.ftz.f32 	%f7, %f64, 0f3F13CD3A;
	mul.ftz.f32 	%f8, %f65, 0f3ED105EC;
	mul.ftz.f32 	%f9, %f66, 0f3F3504F3;
	.loc 1 40 1
	ld.global.u32 	%r13, [%rd2];
	setp.eq.s32	%p5, %r13, 0;
	@%p5 bra 	BB0_5;

	.loc 1 40 1
	mul.ftz.f32 	%f67, %f37, 0f3F1645A2;
	fma.rn.ftz.f32 	%f68, %f38, 0f3E991687, %f67;
	fma.rn.ftz.f32 	%f69, %f36, 0f3DE978D5, %f68;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f70, %f69, 0fC0000000, 0f3F800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f71, %f70;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f72, %f69, 0f00000000, 0f3F800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f73, %f72;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f74, %f69, 0f40000000, 0fBF800000;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f75, %f74;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f76, %f71, 0fBF800000, %f73;
	fma.rn.ftz.f32 	%f77, %f75, 0fBF800000, %f76;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f78, %f77;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f79, %f80, %f81, %f82}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+48];
	ld.shared.v4.f32 	{%f84, %f85, %f86, %f87}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f89, %f7, %f84, %f79;
	fma.rn.ftz.f32 	%f92, %f8, %f85, %f80;
	fma.rn.ftz.f32 	%f95, %f9, %f86, %f81;
	.loc 1 40 1
	mul.ftz.f32 	%f96, %f71, %f89;
	mul.ftz.f32 	%f97, %f71, %f92;
	mul.ftz.f32 	%f98, %f71, %f95;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f99, %f100, %f101, %f102}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+64];
	ld.shared.v4.f32 	{%f104, %f105, %f106, %f107}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+16];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f109, %f7, %f104, %f99;
	fma.rn.ftz.f32 	%f112, %f8, %f105, %f100;
	fma.rn.ftz.f32 	%f115, %f9, %f106, %f101;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f116, %f78, %f109, %f96;
	fma.rn.ftz.f32 	%f117, %f78, %f112, %f97;
	fma.rn.ftz.f32 	%f118, %f78, %f115, %f98;
	.loc 1 40 1
	ld.shared.v4.f32 	{%f119, %f120, %f121, %f122}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+80];
	ld.shared.v4.f32 	{%f124, %f125, %f126, %f127}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+32];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f129, %f7, %f124, %f119;
	fma.rn.ftz.f32 	%f132, %f8, %f125, %f120;
	fma.rn.ftz.f32 	%f135, %f9, %f126, %f121;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f178, %f75, %f129, %f116;
	fma.rn.ftz.f32 	%f179, %f75, %f132, %f117;
	fma.rn.ftz.f32 	%f180, %f75, %f135, %f118;
	bra.uni 	BB0_6;

BB0_5:
	.loc 1 40 1
	ld.shared.v4.f32 	{%f136, %f137, %f138, %f139}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local+48];
	ld.shared.v4.f32 	{%f141, %f142, %f143, %f144}, [ShaderKernel_IRIDASAutoColorMatch$__cuda_local_var_170278_498_non_const_p_local];
	.loc 1 40 1
	fma.rn.ftz.f32 	%f178, %f7, %f141, %f136;
	fma.rn.ftz.f32 	%f179, %f8, %f142, %f137;
	fma.rn.ftz.f32 	%f180, %f9, %f143, %f138;

BB0_6:
	.loc 1 40 1
	mul.ftz.f32 	%f150, %f179, 0f3ED105EC;
	.loc 1 40 1
	fma.rn.ftz.f32 	%f151, %f178, 0f3F13CD3A, %f150;
	.loc 1 40 1
	mul.ftz.f32 	%f152, %f180, 0f3F3504F3;
	.loc 1 40 1
	add.ftz.f32 	%f153, %f151, %f152;
	fma.rn.ftz.f32 	%f154, %f152, 0fBF800000, %f151;
	mul.ftz.f32 	%f155, %f150, 0fC0000000;
	fma.rn.ftz.f32 	%f156, %f178, 0f3F13CD3A, %f155;
	fma.rn.ftz.f32 	%f157, %f152, 0f00000000, %f156;
	mov.f32 	%f158, 0f41200000;
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f159, %f158;
	mul.ftz.f32 	%f160, %f153, %f159;
	ex2.approx.ftz.f32 	%f161, %f160;
	mul.ftz.f32 	%f162, %f154, %f159;
	ex2.approx.ftz.f32 	%f163, %f162;
	mul.ftz.f32 	%f164, %f157, %f159;
	ex2.approx.ftz.f32 	%f165, %f164;
	.loc 1 40 1
	add.ftz.f32 	%f166, %f161, 0fBF800000;
	add.ftz.f32 	%f167, %f163, 0fBF800000;
	add.ftz.f32 	%f168, %f165, 0fBF800000;
	.loc 1 40 1
	mul.ftz.f32 	%f169, %f167, 0fC0659653;
	fma.rn.ftz.f32 	%f170, %f166, 0f408EF909, %f169;
	fma.rn.ftz.f32 	%f171, %f168, 0f3DF4538F, %f170;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f24, %f171;
	.loc 1 40 1
	mul.ftz.f32 	%f172, %f167, 0f401860AA;
	fma.rn.ftz.f32 	%f173, %f166, 0fBF9BFB16, %f172;
	fma.rn.ftz.f32 	%f174, %f168, 0fBE264C30, %f173;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f23, %f174;
	.loc 1 40 1
	mul.ftz.f32 	%f175, %f167, 0fBE79C0EC;
	fma.rn.ftz.f32 	%f176, %f166, 0f3D4B923A, %f175;
	fma.rn.ftz.f32 	%f177, %f168, 0f3F9A2D0E, %f176;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f22, %f177;
	.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], {%f22, %f23, %f24, %f39};
	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, %f22;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f23;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f24;
	mov.b16 	%rs3, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f39;
	mov.b16 	%rs4, %temp;
}
	.loc 1 40 231
	st.global.v4.u16 	[%rd16], {%rs1, %rs2, %rs3, %rs4};

BB0_9:
	.loc 1 40 2
	ret;
}


