//
// 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/IRIDASSimplePrimary.cu", 1405709992, 3721
	.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_IRIDASSimplePrimary$__cuda_local_var_170285_497_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_IRIDASSimplePrimary(
	.param .u64 ShaderKernel_IRIDASSimplePrimary_param_0,
	.param .u32 ShaderKernel_IRIDASSimplePrimary_param_1,
	.param .u32 ShaderKernel_IRIDASSimplePrimary_param_2,
	.param .u32 ShaderKernel_IRIDASSimplePrimary_param_3,
	.param .u32 ShaderKernel_IRIDASSimplePrimary_param_4,
	.param .u64 ShaderKernel_IRIDASSimplePrimary_param_5,
	.param .u64 ShaderKernel_IRIDASSimplePrimary_param_6,
	.param .u64 ShaderKernel_IRIDASSimplePrimary_param_7
)
{
	.reg .pred 	%p<15>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<17>;
	.reg .f32 	%f<155>;
	.reg .s64 	%rd<20>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local[128];

	ld.param.u64 	%rd5, [ShaderKernel_IRIDASSimplePrimary_param_0];
	ld.param.u32 	%r4, [ShaderKernel_IRIDASSimplePrimary_param_1];
	ld.param.u32 	%r5, [ShaderKernel_IRIDASSimplePrimary_param_2];
	ld.param.u32 	%r6, [ShaderKernel_IRIDASSimplePrimary_param_3];
	ld.param.u32 	%r7, [ShaderKernel_IRIDASSimplePrimary_param_4];
	ld.param.u64 	%rd7, [ShaderKernel_IRIDASSimplePrimary_param_5];
	ld.param.u64 	%rd6, [ShaderKernel_IRIDASSimplePrimary_param_7];
	cvta.to.global.u64 	%rd1, %rd6;
	cvta.to.global.u64 	%rd2, %rd7;
	.loc 1 47 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 47 1
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 47 1
	@!%p3 bra 	BB0_21;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 47 1
	cvt.rn.f32.s32	%f50, %r2;
	add.ftz.f32 	%f1, %f50, 0f3F000000;
	cvt.rn.f32.s32	%f51, %r3;
	add.ftz.f32 	%f2, %f51, 0f3F000000;
	.loc 1 47 1
	setp.gt.u32	%p4, %r1, 7;
	@%p4 bra 	BB0_3;

	.loc 1 47 1
	mul.wide.u32 	%rd8, %r1, 16;
	mov.u64 	%rd9, ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local;
	add.s64 	%rd10, %rd9, %rd8;
	add.s64 	%rd11, %rd2, %rd8;
	ld.global.v4.f32 	{%f52, %f53, %f54, %f55}, [%rd11];
	st.shared.v4.f32 	[%rd10], {%f52, %f53, %f54, %f55};

BB0_3:
	.loc 1 47 1
	bar.sync 	0;
	.loc 1 47 105
	// inline asm
	tex.2d.v4.f32.f32 {%f60, %f61, %f62, %f63}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 47 1
	ld.global.u32 	%r13, [%rd1+4];
	setp.eq.s32	%p5, %r13, 0;
	.loc 1 47 105
	mov.f32 	%f143, %f62;
	mov.f32 	%f144, %f61;
	mov.f32 	%f145, %f60;
	.loc 1 47 1
	@%p5 bra 	BB0_5;

	.loc 1 47 1
	ld.shared.v4.f32 	{%f66, %f67, %f68, %f69}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+64];
	.loc 1 47 1
	mul.ftz.f32 	%f71, %f61, %f67;
	fma.rn.ftz.f32 	%f73, %f62, %f66, %f71;
	fma.rn.ftz.f32 	%f75, %f60, %f68, %f73;
	.loc 1 47 1
	sub.ftz.f32 	%f76, %f62, %f75;
	sub.ftz.f32 	%f77, %f61, %f75;
	sub.ftz.f32 	%f78, %f60, %f75;
	.loc 1 47 1
	ld.shared.v4.f32 	{%f79, %f80, %f81, %f82}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+80];
	.loc 1 47 1
	fma.rn.ftz.f32 	%f143, %f76, %f79, %f75;
	fma.rn.ftz.f32 	%f144, %f77, %f80, %f75;
	fma.rn.ftz.f32 	%f145, %f78, %f81, %f75;

BB0_5:
	.loc 1 47 1
	ld.shared.v4.f32 	{%f86, %f87, %f88, %f89}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+16];
	.loc 1 47 1
	ld.shared.v4.f32 	{%f91, %f92, %f93, %f94}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local];
	.loc 1 47 1
	fma.rn.ftz.f32 	%f149, %f143, %f91, %f86;
	fma.rn.ftz.f32 	%f150, %f144, %f92, %f87;
	fma.rn.ftz.f32 	%f151, %f145, %f93, %f88;
	.loc 1 47 1
	ld.global.u32 	%r14, [%rd1];
	setp.eq.s32	%p6, %r14, 0;
	@%p6 bra 	BB0_16;

	.loc 1 47 1
	setp.lt.ftz.f32	%p7, %f149, 0f00000000;
	selp.f32	%f16, 0fBF800000, 0f3F800000, %p7;
	setp.lt.ftz.f32	%p8, %f150, 0f00000000;
	selp.f32	%f17, 0fBF800000, 0f3F800000, %p8;
	setp.lt.ftz.f32	%p9, %f151, 0f00000000;
	selp.f32	%f18, 0fBF800000, 0f3F800000, %p9;
	.loc 2 2750 10
	abs.ftz.f32 	%f19, %f150;
	abs.ftz.f32 	%f20, %f151;
	.loc 1 47 1
	ld.shared.f32 	%f21, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+112];
	.loc 2 2750 10
	abs.ftz.f32 	%f22, %f149;
	.loc 1 47 1
	setp.gtu.ftz.f32	%p10, %f22, 0f00000000;
	@%p10 bra 	BB0_8;

	mov.f32 	%f146, 0f00000000;
	bra.uni 	BB0_9;

BB0_8:
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f101, %f22;
	mul.ftz.f32 	%f102, %f21, %f101;
	ex2.approx.ftz.f32 	%f146, %f102;

BB0_9:
	.loc 1 47 1
	ld.shared.f32 	%f25, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+116];
	setp.gtu.ftz.f32	%p11, %f19, 0f00000000;
	@%p11 bra 	BB0_11;

	mov.f32 	%f147, 0f00000000;
	bra.uni 	BB0_12;

BB0_11:
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f104, %f19;
	mul.ftz.f32 	%f105, %f25, %f104;
	ex2.approx.ftz.f32 	%f147, %f105;

BB0_12:
	.loc 1 47 1
	ld.shared.f32 	%f28, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+120];
	setp.gtu.ftz.f32	%p12, %f20, 0f00000000;
	@%p12 bra 	BB0_14;

	mov.f32 	%f148, 0f00000000;
	bra.uni 	BB0_15;

BB0_14:
	.loc 2 3600 10
	lg2.approx.ftz.f32 	%f107, %f20;
	mul.ftz.f32 	%f108, %f28, %f107;
	ex2.approx.ftz.f32 	%f148, %f108;

BB0_15:
	.loc 1 47 1
	mul.ftz.f32 	%f149, %f146, %f16;
	mul.ftz.f32 	%f150, %f147, %f17;
	mul.ftz.f32 	%f151, %f148, %f18;

BB0_16:
	.loc 1 47 1
	ld.shared.v4.f32 	{%f109, %f110, %f111, %f112}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+48];
	.loc 1 47 1
	ld.shared.v4.f32 	{%f114, %f115, %f116, %f117}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+32];
	.loc 1 47 1
	fma.rn.ftz.f32 	%f152, %f149, %f114, %f109;
	fma.rn.ftz.f32 	%f153, %f150, %f115, %f110;
	fma.rn.ftz.f32 	%f154, %f151, %f116, %f111;
	.loc 1 47 1
	ld.global.u32 	%r15, [%rd1+8];
	setp.eq.s32	%p13, %r15, 0;
	@%p13 bra 	BB0_18;

	.loc 1 47 1
	ld.shared.v4.f32 	{%f123, %f124, %f125, %f126}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+64];
	.loc 1 47 1
	mul.ftz.f32 	%f128, %f153, %f124;
	fma.rn.ftz.f32 	%f130, %f152, %f123, %f128;
	fma.rn.ftz.f32 	%f132, %f154, %f125, %f130;
	.loc 1 47 1
	sub.ftz.f32 	%f133, %f152, %f132;
	sub.ftz.f32 	%f134, %f153, %f132;
	sub.ftz.f32 	%f135, %f154, %f132;
	.loc 1 47 1
	ld.shared.v4.f32 	{%f136, %f137, %f138, %f139}, [ShaderKernel_IRIDASSimplePrimary$__cuda_local_var_170285_497_non_const_p_local+96];
	.loc 1 47 1
	fma.rn.ftz.f32 	%f152, %f133, %f136, %f132;
	fma.rn.ftz.f32 	%f153, %f134, %f137, %f132;
	fma.rn.ftz.f32 	%f154, %f135, %f138, %f132;

BB0_18:
	.loc 1 47 1
	mad.lo.s32 	%r16, %r3, %r4, %r2;
	.loc 1 47 1
	cvt.s64.s32	%rd4, %r16;
	.loc 1 47 1
	setp.eq.s32	%p14, %r5, 0;
	@%p14 bra 	BB0_20;

	cvta.to.global.u64 	%rd14, %rd5;
	.loc 1 47 1
	shl.b64 	%rd15, %rd4, 4;
	add.s64 	%rd16, %rd14, %rd15;
	st.global.v4.f32 	[%rd16], {%f154, %f153, %f152, %f63};
	bra.uni 	BB0_21;

BB0_20:
	cvta.to.global.u64 	%rd17, %rd5;
	.loc 1 47 1
	shl.b64 	%rd18, %rd4, 3;
	add.s64 	%rd19, %rd17, %rd18;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f154;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f153;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f152;
	mov.b16 	%rs3, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f63;
	mov.b16 	%rs4, %temp;
}
	.loc 1 47 231
	st.global.v4.u16 	[%rd19], {%rs1, %rs2, %rs3, %rs4};

BB0_21:
	.loc 1 47 2
	ret;
}


