//
// 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/fxDay2NiteV.cu", 1405709992, 4459
	.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;
.global .texref texture1_RECT;
// ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_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_fxDay2NiteV(
	.param .u64 ShaderKernel_fxDay2NiteV_param_0,
	.param .u32 ShaderKernel_fxDay2NiteV_param_1,
	.param .u32 ShaderKernel_fxDay2NiteV_param_2,
	.param .u32 ShaderKernel_fxDay2NiteV_param_3,
	.param .u32 ShaderKernel_fxDay2NiteV_param_4,
	.param .u64 ShaderKernel_fxDay2NiteV_param_5,
	.param .u64 ShaderKernel_fxDay2NiteV_param_6,
	.param .u64 ShaderKernel_fxDay2NiteV_param_7
)
{
	.reg .pred 	%p<8>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<199>;
	.reg .s64 	%rd<40>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local[112];

	ld.param.u64 	%rd4, [ShaderKernel_fxDay2NiteV_param_0];
	ld.param.u32 	%r4, [ShaderKernel_fxDay2NiteV_param_1];
	ld.param.u32 	%r5, [ShaderKernel_fxDay2NiteV_param_2];
	ld.param.u32 	%r6, [ShaderKernel_fxDay2NiteV_param_3];
	ld.param.u32 	%r7, [ShaderKernel_fxDay2NiteV_param_4];
	ld.param.u64 	%rd5, [ShaderKernel_fxDay2NiteV_param_5];
	cvta.to.global.u64 	%rd1, %rd4;
	cvta.to.global.u64 	%rd2, %rd5;
	.loc 1 34 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 34 1
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 34 1
	@!%p3 bra 	BB0_9;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 34 1
	cvt.rn.f32.s32	%f18, %r2;
	add.ftz.f32 	%f1, %f18, 0f3F000000;
	cvt.rn.f32.s32	%f19, %r3;
	add.ftz.f32 	%f2, %f19, 0f3F000000;
	.loc 1 34 1
	setp.gt.u32	%p4, %r1, 6;
	@%p4 bra 	BB0_3;

	.loc 1 34 1
	mul.wide.u32 	%rd6, %r1, 16;
	mov.u64 	%rd7, ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local;
	add.s64 	%rd8, %rd7, %rd6;
	add.s64 	%rd9, %rd2, %rd6;
	ld.global.v4.f32 	{%f20, %f21, %f22, %f23}, [%rd9];
	st.shared.v4.f32 	[%rd8], {%f20, %f21, %f22, %f23};

BB0_3:
	.loc 1 34 1
	bar.sync 	0;
	.loc 1 34 111
	// inline asm
	tex.2d.v4.f32.f32 {%f28, %f29, %f30, %f31}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 34 1
	add.ftz.f32 	%f80, %f1, 0f00000000;
	add.ftz.f32 	%f39, %f2, 0fBF800000;
	.loc 1 34 111
	// inline asm
	tex.2d.v4.f32.f32 {%f34, %f35, %f36, %f37}, [texture0_RECT, {%f80, %f39}];
	// inline asm
	.loc 1 34 1
	ld.shared.v2.f32 	{%f82, %f83}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local];
	.loc 1 34 1
	fma.rn.ftz.f32 	%f85, %f36, %f82, %f30;
	fma.rn.ftz.f32 	%f87, %f35, %f83, %f29;
	.loc 1 34 1
	add.ftz.f32 	%f45, %f2, 0f3F800000;
	mov.f32 	%f88, 0f3F800000;
	.loc 1 34 111
	// inline asm
	tex.2d.v4.f32.f32 {%f40, %f41, %f42, %f43}, [texture0_RECT, {%f80, %f45}];
	// inline asm
	.loc 1 34 1
	fma.rn.ftz.f32 	%f89, %f42, %f82, %f85;
	fma.rn.ftz.f32 	%f90, %f41, %f83, %f87;
	.loc 1 34 1
	add.ftz.f32 	%f51, %f2, 0fC0000000;
	.loc 1 34 113
	// inline asm
	tex.2d.v4.f32.f32 {%f46, %f47, %f48, %f49}, [texture0_RECT, {%f80, %f51}];
	// inline asm
	.loc 1 34 1
	ld.shared.v2.f32 	{%f91, %f92}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+16];
	.loc 1 34 1
	fma.rn.ftz.f32 	%f94, %f48, %f91, %f89;
	fma.rn.ftz.f32 	%f96, %f47, %f92, %f90;
	.loc 1 34 1
	add.ftz.f32 	%f57, %f2, 0f40000000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f52, %f53, %f54, %f55}, [texture0_RECT, {%f80, %f57}];
	// inline asm
	.loc 1 34 1
	fma.rn.ftz.f32 	%f97, %f54, %f91, %f94;
	fma.rn.ftz.f32 	%f98, %f53, %f92, %f96;
	.loc 1 34 1
	add.ftz.f32 	%f63, %f2, 0fC0400000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f58, %f59, %f60, %f61}, [texture0_RECT, {%f80, %f63}];
	// inline asm
	.loc 1 34 1
	ld.shared.v2.f32 	{%f99, %f100}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+32];
	.loc 1 34 1
	fma.rn.ftz.f32 	%f102, %f60, %f99, %f97;
	fma.rn.ftz.f32 	%f104, %f59, %f100, %f98;
	.loc 1 34 1
	add.ftz.f32 	%f69, %f2, 0f40400000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f64, %f65, %f66, %f67}, [texture0_RECT, {%f80, %f69}];
	// inline asm
	.loc 1 34 1
	fma.rn.ftz.f32 	%f105, %f66, %f99, %f102;
	fma.rn.ftz.f32 	%f106, %f65, %f100, %f104;
	.loc 1 34 1
	add.ftz.f32 	%f75, %f2, 0fC0800000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f70, %f71, %f72, %f73}, [texture0_RECT, {%f80, %f75}];
	// inline asm
	.loc 1 34 1
	ld.shared.v2.f32 	{%f107, %f108}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+48];
	.loc 1 34 1
	fma.rn.ftz.f32 	%f110, %f72, %f107, %f105;
	fma.rn.ftz.f32 	%f112, %f71, %f108, %f106;
	.loc 1 34 1
	add.ftz.f32 	%f81, %f2, 0f40800000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f76, %f77, %f78, %f79}, [texture0_RECT, {%f80, %f81}];
	// inline asm
	.loc 1 34 1
	fma.rn.ftz.f32 	%f113, %f78, %f107, %f110;
	fma.rn.ftz.f32 	%f114, %f77, %f108, %f112;
	.loc 1 34 1
	ld.shared.v2.f32 	{%f115, %f116}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+64];
	.loc 1 34 1
	mul.ftz.f32 	%f4, %f113, %f115;
	mul.ftz.f32 	%f5, %f114, %f116;
	.loc 1 34 1
	sub.ftz.f32 	%f119, %f4, %f5;
	.loc 2 2910 10
	div.rn.ftz.f32 	%f6, %f88, %f119;
	.loc 1 34 1
	setp.gt.ftz.f32	%p5, %f119, 0f3A83126F;
	selp.f32	%f7, %f119, 0f3A83126F, %p5;
	setp.gtu.ftz.f32	%p6, %f7, 0f00000000;
	@%p6 bra 	BB0_5;

	mov.f32 	%f198, 0f00000000;
	bra.uni 	BB0_6;

BB0_5:
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f121, %f7;
	mul.ftz.f32 	%f122, %f121, 0f3F4CCCCD;
	ex2.approx.ftz.f32 	%f198, %f122;

BB0_6:
	.loc 1 34 1
	mul.ftz.f32 	%f147, %f6, %f198;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f148, %f147;
	.loc 1 34 1
	sub.ftz.f32 	%f150, %f88, %f148;
	mul.ftz.f32 	%f151, %f150, %f5;
	fma.rn.ftz.f32 	%f152, %f148, %f4, %f151;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f123, %f124, %f125, %f126}, [texture1_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 34 1
	add.ftz.f32 	%f134, %f2, 0f00000000;
	add.ftz.f32 	%f133, %f1, 0fBF800000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f129, %f130, %f131, %f132}, [texture1_RECT, {%f133, %f134}];
	// inline asm
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f135, %f136, %f137, %f138}, [texture1_RECT, {%f80, %f45}];
	// inline asm
	.loc 1 34 1
	add.ftz.f32 	%f145, %f1, 0f3F800000;
	.loc 1 34 119
	// inline asm
	tex.2d.v4.f32.f32 {%f141, %f142, %f143, %f144}, [texture1_RECT, {%f145, %f39}];
	// inline asm
	.loc 1 34 1
	add.ftz.f32 	%f153, %f131, %f125;
	add.ftz.f32 	%f154, %f130, %f124;
	add.ftz.f32 	%f155, %f129, %f123;
	.loc 1 34 1
	add.ftz.f32 	%f156, %f153, %f137;
	add.ftz.f32 	%f157, %f154, %f136;
	add.ftz.f32 	%f158, %f155, %f135;
	.loc 1 34 1
	add.ftz.f32 	%f159, %f156, %f143;
	add.ftz.f32 	%f160, %f157, %f142;
	add.ftz.f32 	%f161, %f158, %f141;
	.loc 1 34 1
	ld.shared.v4.f32 	{%f162, %f163, %f164, %f165}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+80];
	.loc 1 34 1
	mul.ftz.f32 	%f167, %f152, %f162;
	mul.ftz.f32 	%f169, %f152, %f163;
	mul.ftz.f32 	%f171, %f152, %f164;
	mul.ftz.f32 	%f173, %f152, %f165;
	.loc 1 34 1
	mul.ftz.f32 	%f174, %f160, 0f3E1645A2;
	fma.rn.ftz.f32 	%f175, %f159, 0f3D991687, %f174;
	fma.rn.ftz.f32 	%f176, %f161, 0f3CE978D5, %f175;
	.loc 1 34 1
	ld.shared.v4.f32 	{%f177, %f178, %f179, %f180}, [ShaderKernel_fxDay2NiteV$__cuda_local_var_170272_636_non_const_p_local+96];
	.loc 1 34 1
	add.ftz.f32 	%f182, %f176, %f177;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f183, %f182;
	.loc 1 34 1
	add.ftz.f32 	%f185, %f176, %f178;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f186, %f185;
	.loc 1 34 1
	add.ftz.f32 	%f188, %f176, %f179;
	.loc 2 2820 10
	cvt.ftz.sat.f32.f32	%f189, %f188;
	.loc 1 34 1
	sub.ftz.f32 	%f190, %f88, %f183;
	mul.ftz.f32 	%f191, %f190, %f167;
	fma.rn.ftz.f32 	%f10, %f183, %f125, %f191;
	sub.ftz.f32 	%f192, %f88, %f186;
	mul.ftz.f32 	%f193, %f192, %f169;
	fma.rn.ftz.f32 	%f11, %f186, %f124, %f193;
	sub.ftz.f32 	%f194, %f88, %f189;
	mul.ftz.f32 	%f195, %f194, %f171;
	fma.rn.ftz.f32 	%f12, %f189, %f123, %f195;
	sub.ftz.f32 	%f196, %f88, 0f3F800000;
	mul.ftz.f32 	%f197, %f196, %f173;
	fma.rn.ftz.f32 	%f13, %f126, 0f3F800000, %f197;
	.loc 1 34 1
	mad.lo.s32 	%r13, %r3, %r4, %r2;
	.loc 1 34 1
	cvt.s64.s32	%rd3, %r13;
	.loc 1 34 1
	setp.eq.s32	%p7, %r5, 0;
	@%p7 bra 	BB0_8;

	.loc 1 34 1
	shl.b64 	%rd36, %rd3, 4;
	add.s64 	%rd37, %rd1, %rd36;
	st.global.v4.f32 	[%rd37], {%f12, %f11, %f10, %f13};
	bra.uni 	BB0_9;

BB0_8:
	.loc 1 34 1
	shl.b64 	%rd38, %rd3, 3;
	add.s64 	%rd39, %rd1, %rd38;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f12;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f11;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f10;
	mov.b16 	%rs3, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f13;
	mov.b16 	%rs4, %temp;
}
	.loc 1 34 241
	st.global.v4.u16 	[%rd39], {%rs1, %rs2, %rs3, %rs4};

BB0_9:
	.loc 1 34 2
	ret;
}


