//
// 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 "D:/singlebarrel/releases/2014.03/shared/adobe/Iridas/IRIDASLIB/GPU/fxDay2NiteH.cu", 1399785249, 4807
	.file	2 "d:\\singlebarrel\\releases\\2014.03\\shared\\adobe\\mediacore\\external\\3rdparty\\nvidia\\cuda\\win\\include\\device_functions.h", 1399785281, 191626
.global .texref texture0_RECT;
// ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_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_fxDay2NiteH(
	.param .u64 ShaderKernel_fxDay2NiteH_param_0,
	.param .u32 ShaderKernel_fxDay2NiteH_param_1,
	.param .u32 ShaderKernel_fxDay2NiteH_param_2,
	.param .u32 ShaderKernel_fxDay2NiteH_param_3,
	.param .u32 ShaderKernel_fxDay2NiteH_param_4,
	.param .u64 ShaderKernel_fxDay2NiteH_param_5,
	.param .u64 ShaderKernel_fxDay2NiteH_param_6
)
{
	.reg .pred 	%p<8>;
	.reg .s16 	%rs<4>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<233>;
	.reg .s64 	%rd<32>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local[96];

	ld.param.u64 	%rd4, [ShaderKernel_fxDay2NiteH_param_0];
	ld.param.u32 	%r4, [ShaderKernel_fxDay2NiteH_param_1];
	ld.param.u32 	%r5, [ShaderKernel_fxDay2NiteH_param_2];
	ld.param.u32 	%r6, [ShaderKernel_fxDay2NiteH_param_3];
	ld.param.u32 	%r7, [ShaderKernel_fxDay2NiteH_param_4];
	ld.param.u64 	%rd5, [ShaderKernel_fxDay2NiteH_param_5];
	cvta.to.global.u64 	%rd1, %rd4;
	cvta.to.global.u64 	%rd2, %rd5;
	.loc 1 33 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 33 1
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 33 1
	@!%p3 bra 	BB0_6;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 33 1
	cvt.rn.f32.s32	%f9, %r2;
	add.ftz.f32 	%f1, %f9, 0f3F000000;
	cvt.rn.f32.s32	%f10, %r3;
	add.ftz.f32 	%f2, %f10, 0f3F000000;
	.loc 1 33 1
	setp.gt.u32	%p4, %r1, 5;
	@%p4 bra 	BB0_3;

	.loc 1 33 1
	mul.wide.u32 	%rd6, %r1, 16;
	mov.u64 	%rd7, ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local;
	add.s64 	%rd8, %rd7, %rd6;
	add.s64 	%rd9, %rd2, %rd6;
	ld.global.v4.f32 	{%f11, %f12, %f13, %f14}, [%rd9];
	st.shared.v4.f32 	[%rd8], {%f11, %f12, %f13, %f14};

BB0_3:
	.loc 1 33 1
	bar.sync 	0;
	.loc 1 33 111
	// inline asm
	tex.2d.v4.f32.f32 {%f19, %f20, %f21, %f22}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 33 1
	add.ftz.f32 	%f72, %f2, 0f00000000;
	add.ftz.f32 	%f29, %f1, 0f3F800000;
	mov.f32 	%f73, 0f3F800000;
	.loc 1 33 111
	// inline asm
	tex.2d.v4.f32.f32 {%f25, %f26, %f27, %f28}, [texture0_RECT, {%f29, %f72}];
	// inline asm
	.loc 1 33 1
	ld.shared.v2.f32 	{%f74, %f75}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local];
	.loc 1 33 1
	fma.rn.ftz.f32 	%f77, %f27, %f74, %f21;
	fma.rn.ftz.f32 	%f78, %f26, %f74, %f20;
	fma.rn.ftz.f32 	%f79, %f25, %f74, %f19;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f81, %f27, %f75, %f21;
	fma.rn.ftz.f32 	%f82, %f26, %f75, %f20;
	fma.rn.ftz.f32 	%f83, %f25, %f75, %f19;
	.loc 1 33 1
	add.ftz.f32 	%f35, %f1, 0fBF800000;
	.loc 1 33 111
	// inline asm
	tex.2d.v4.f32.f32 {%f31, %f32, %f33, %f34}, [texture0_RECT, {%f35, %f72}];
	// inline asm
	.loc 1 33 1
	fma.rn.ftz.f32 	%f84, %f33, %f74, %f77;
	fma.rn.ftz.f32 	%f85, %f32, %f74, %f78;
	fma.rn.ftz.f32 	%f86, %f31, %f74, %f79;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f87, %f33, %f75, %f81;
	fma.rn.ftz.f32 	%f88, %f32, %f75, %f82;
	fma.rn.ftz.f32 	%f89, %f31, %f75, %f83;
	.loc 1 33 1
	add.ftz.f32 	%f41, %f1, 0f40000000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f37, %f38, %f39, %f40}, [texture0_RECT, {%f41, %f72}];
	// inline asm
	.loc 1 33 1
	ld.shared.v2.f32 	{%f90, %f91}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local+16];
	.loc 1 33 1
	fma.rn.ftz.f32 	%f93, %f39, %f90, %f84;
	fma.rn.ftz.f32 	%f94, %f38, %f90, %f85;
	fma.rn.ftz.f32 	%f95, %f37, %f90, %f86;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f97, %f39, %f91, %f87;
	fma.rn.ftz.f32 	%f98, %f38, %f91, %f88;
	fma.rn.ftz.f32 	%f99, %f37, %f91, %f89;
	.loc 1 33 1
	add.ftz.f32 	%f47, %f1, 0fC0000000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f43, %f44, %f45, %f46}, [texture0_RECT, {%f47, %f72}];
	// inline asm
	.loc 1 33 1
	fma.rn.ftz.f32 	%f100, %f45, %f90, %f93;
	fma.rn.ftz.f32 	%f101, %f44, %f90, %f94;
	fma.rn.ftz.f32 	%f102, %f43, %f90, %f95;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f103, %f45, %f91, %f97;
	fma.rn.ftz.f32 	%f104, %f44, %f91, %f98;
	fma.rn.ftz.f32 	%f105, %f43, %f91, %f99;
	.loc 1 33 1
	add.ftz.f32 	%f53, %f1, 0f40400000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f49, %f50, %f51, %f52}, [texture0_RECT, {%f53, %f72}];
	// inline asm
	.loc 1 33 1
	ld.shared.v2.f32 	{%f106, %f107}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local+32];
	.loc 1 33 1
	fma.rn.ftz.f32 	%f109, %f51, %f106, %f100;
	fma.rn.ftz.f32 	%f110, %f50, %f106, %f101;
	fma.rn.ftz.f32 	%f111, %f49, %f106, %f102;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f113, %f51, %f107, %f103;
	fma.rn.ftz.f32 	%f114, %f50, %f107, %f104;
	fma.rn.ftz.f32 	%f115, %f49, %f107, %f105;
	.loc 1 33 1
	add.ftz.f32 	%f59, %f1, 0fC0400000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f55, %f56, %f57, %f58}, [texture0_RECT, {%f59, %f72}];
	// inline asm
	.loc 1 33 1
	fma.rn.ftz.f32 	%f116, %f57, %f106, %f109;
	fma.rn.ftz.f32 	%f117, %f56, %f106, %f110;
	fma.rn.ftz.f32 	%f118, %f55, %f106, %f111;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f119, %f57, %f107, %f113;
	fma.rn.ftz.f32 	%f120, %f56, %f107, %f114;
	fma.rn.ftz.f32 	%f121, %f55, %f107, %f115;
	.loc 1 33 1
	add.ftz.f32 	%f65, %f1, 0f40800000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f61, %f62, %f63, %f64}, [texture0_RECT, {%f65, %f72}];
	// inline asm
	.loc 1 33 1
	ld.shared.v2.f32 	{%f122, %f123}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local+48];
	.loc 1 33 1
	fma.rn.ftz.f32 	%f125, %f63, %f122, %f116;
	fma.rn.ftz.f32 	%f126, %f62, %f122, %f117;
	fma.rn.ftz.f32 	%f127, %f61, %f122, %f118;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f129, %f63, %f123, %f119;
	fma.rn.ftz.f32 	%f130, %f62, %f123, %f120;
	fma.rn.ftz.f32 	%f131, %f61, %f123, %f121;
	.loc 1 33 1
	add.ftz.f32 	%f71, %f1, 0fC0800000;
	.loc 1 33 119
	// inline asm
	tex.2d.v4.f32.f32 {%f67, %f68, %f69, %f70}, [texture0_RECT, {%f71, %f72}];
	// inline asm
	.loc 1 33 1
	fma.rn.ftz.f32 	%f132, %f69, %f122, %f125;
	fma.rn.ftz.f32 	%f133, %f68, %f122, %f126;
	fma.rn.ftz.f32 	%f134, %f67, %f122, %f127;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f135, %f69, %f123, %f129;
	fma.rn.ftz.f32 	%f136, %f68, %f123, %f130;
	fma.rn.ftz.f32 	%f137, %f67, %f123, %f131;
	.loc 1 33 1
	ld.shared.v2.f32 	{%f138, %f139}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local+64];
	.loc 1 33 1
	mul.ftz.f32 	%f141, %f138, %f132;
	mul.ftz.f32 	%f142, %f138, %f133;
	mul.ftz.f32 	%f143, %f138, %f134;
	.loc 1 33 1
	mul.ftz.f32 	%f145, %f139, %f135;
	mul.ftz.f32 	%f146, %f139, %f136;
	mul.ftz.f32 	%f147, %f139, %f137;
	.loc 1 33 1
	mul.ftz.f32 	%f148, %f142, 0f3F000000;
	fma.rn.ftz.f32 	%f149, %f141, 0f3F000000, %f148;
	fma.rn.ftz.f32 	%f150, %f143, 0f3F000000, %f149;
	mul.ftz.f32 	%f151, %f146, 0f3F000000;
	fma.rn.ftz.f32 	%f152, %f145, 0f3F000000, %f151;
	fma.rn.ftz.f32 	%f153, %f147, 0f3F000000, %f152;
	.loc 1 33 1
	mul.ftz.f32 	%f154, %f141, %f150;
	mul.ftz.f32 	%f155, %f142, %f150;
	mul.ftz.f32 	%f156, %f143, %f150;
	.loc 1 33 1
	mul.ftz.f32 	%f157, %f145, %f153;
	mul.ftz.f32 	%f158, %f146, %f153;
	mul.ftz.f32 	%f159, %f147, %f153;
	.loc 1 33 1
	mul.ftz.f32 	%f160, %f150, %f150;
	mul.ftz.f32 	%f161, %f153, %f153;
	.loc 1 33 1
	sub.ftz.f32 	%f162, %f154, %f160;
	sub.ftz.f32 	%f163, %f155, %f160;
	sub.ftz.f32 	%f164, %f156, %f160;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f165, %f162, 0f3F400000, %f160;
	fma.rn.ftz.f32 	%f166, %f163, 0f3F400000, %f160;
	fma.rn.ftz.f32 	%f167, %f164, 0f3F400000, %f160;
	.loc 1 33 1
	add.ftz.f32 	%f168, %f165, %f165;
	add.ftz.f32 	%f169, %f166, %f166;
	add.ftz.f32 	%f170, %f167, %f167;
	.loc 1 33 1
	ld.shared.v4.f32 	{%f171, %f172, %f173, %f174}, [ShaderKernel_fxDay2NiteH$__cuda_local_var_170271_471_non_const_p_local+80];
	.loc 1 33 1
	sub.ftz.f32 	%f176, %f73, %f171;
	mul.ftz.f32 	%f177, %f176, %f141;
	fma.rn.ftz.f32 	%f178, %f171, %f168, %f177;
	sub.ftz.f32 	%f180, %f73, %f172;
	mul.ftz.f32 	%f181, %f180, %f142;
	fma.rn.ftz.f32 	%f182, %f172, %f169, %f181;
	sub.ftz.f32 	%f184, %f73, %f173;
	mul.ftz.f32 	%f185, %f184, %f143;
	fma.rn.ftz.f32 	%f186, %f173, %f170, %f185;
	.loc 1 33 1
	sub.ftz.f32 	%f187, %f157, %f161;
	sub.ftz.f32 	%f188, %f158, %f161;
	sub.ftz.f32 	%f189, %f159, %f161;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f190, %f187, 0f3F400000, %f161;
	fma.rn.ftz.f32 	%f191, %f188, 0f3F400000, %f161;
	fma.rn.ftz.f32 	%f192, %f189, 0f3F400000, %f161;
	.loc 1 33 1
	add.ftz.f32 	%f193, %f190, %f190;
	add.ftz.f32 	%f194, %f191, %f191;
	add.ftz.f32 	%f195, %f192, %f192;
	.loc 1 33 1
	mul.ftz.f32 	%f196, %f176, %f145;
	fma.rn.ftz.f32 	%f197, %f171, %f193, %f196;
	mul.ftz.f32 	%f198, %f180, %f146;
	fma.rn.ftz.f32 	%f199, %f172, %f194, %f198;
	mul.ftz.f32 	%f200, %f184, %f147;
	fma.rn.ftz.f32 	%f201, %f173, %f195, %f200;
	.loc 1 33 1
	mul.ftz.f32 	%f202, %f182, 0f3EA617C2;
	fma.rn.ftz.f32 	%f203, %f178, 0f3F03D07D, %f202;
	fma.rn.ftz.f32 	%f204, %f186, 0f3E248E8A, %f203;
	mul.ftz.f32 	%f205, %f182, 0f3F2B9F56;
	fma.rn.ftz.f32 	%f206, %f178, 0f3E87E282, %f205;
	fma.rn.ftz.f32 	%f207, %f186, 0f3D837B4A, %f206;
	mul.ftz.f32 	%f208, %f182, 0f3DFF9724;
	fma.rn.ftz.f32 	%f209, %f178, 0f3CCB295F, %f208;
	fma.rn.ftz.f32 	%f210, %f186, 0f3F59B3D0, %f209;
	mul.ftz.f32 	%f211, %f199, 0f3EA617C2;
	fma.rn.ftz.f32 	%f212, %f197, 0f3F03D07D, %f211;
	fma.rn.ftz.f32 	%f213, %f201, 0f3E248E8A, %f212;
	mul.ftz.f32 	%f214, %f199, 0f3F2B9F56;
	fma.rn.ftz.f32 	%f215, %f197, 0f3E87E282, %f214;
	fma.rn.ftz.f32 	%f216, %f201, 0f3D837B4A, %f215;
	mul.ftz.f32 	%f217, %f199, 0f3DFF9724;
	fma.rn.ftz.f32 	%f218, %f197, 0f3CCB295F, %f217;
	fma.rn.ftz.f32 	%f219, %f201, 0f3F59B3D0, %f218;
	.loc 1 33 1
	setp.gt.ftz.f32	%p5, %f204, 0f3A83126F;
	selp.f32	%f220, %f204, 0f3A83126F, %p5;
	setp.gt.ftz.f32	%p6, %f213, 0f3A83126F;
	selp.f32	%f221, %f213, 0f3A83126F, %p6;
	.loc 1 33 1
	add.ftz.f32 	%f222, %f207, %f210;
	add.ftz.f32 	%f223, %f216, %f219;
	.loc 2 2910 10
	div.rn.ftz.f32 	%f224, %f73, %f220;
	div.rn.ftz.f32 	%f225, %f73, %f221;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f226, %f224, %f222, 0f3F800000;
	fma.rn.ftz.f32 	%f227, %f225, %f223, 0f3F800000;
	.loc 1 33 1
	fma.rn.ftz.f32 	%f228, %f226, 0f3FAA3D71, 0fBFD70A3D;
	fma.rn.ftz.f32 	%f229, %f227, 0f3FAA3D71, 0fBFD70A3D;
	.loc 1 33 1
	mul.ftz.f32 	%f7, %f228, %f207;
	mul.ftz.f32 	%f6, %f229, %f216;
	.loc 1 33 1
	mad.lo.s32 	%r13, %r3, %r4, %r2;
	.loc 1 33 1
	cvt.s64.s32	%rd3, %r13;
	.loc 1 33 1
	setp.eq.s32	%p7, %r5, 0;
	@%p7 bra 	BB0_5;

	.loc 1 33 1
	shl.b64 	%rd28, %rd3, 4;
	add.s64 	%rd29, %rd1, %rd28;
	st.global.v4.f32 	[%rd29], {%f231, %f6, %f7, %f230};
	bra.uni 	BB0_6;

BB0_5:
	.loc 1 33 1
	shl.b64 	%rd30, %rd3, 3;
	add.s64 	%rd31, %rd1, %rd30;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f232;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f6;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f7;
	mov.b16 	%rs3, %temp;
}
	.loc 1 33 241
	st.global.v4.u16 	[%rd31], {%rs1, %rs2, %rs3, %rs1};

BB0_6:
	.loc 1 33 2
	ret;
}


