//
// 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/fxBloomV.cu", 1399785249, 4672
	.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;
.global .texref texture1_RECT;
// ShaderKernel_fxBloomV$__cuda_local_var_170276_633_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_fxBloomV(
	.param .u64 ShaderKernel_fxBloomV_param_0,
	.param .u32 ShaderKernel_fxBloomV_param_1,
	.param .u32 ShaderKernel_fxBloomV_param_2,
	.param .u32 ShaderKernel_fxBloomV_param_3,
	.param .u32 ShaderKernel_fxBloomV_param_4,
	.param .u64 ShaderKernel_fxBloomV_param_5,
	.param .u64 ShaderKernel_fxBloomV_param_6,
	.param .u64 ShaderKernel_fxBloomV_param_7
)
{
	.reg .pred 	%p<6>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<278>;
	.reg .s64 	%rd<46>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local[176];

	ld.param.u64 	%rd4, [ShaderKernel_fxBloomV_param_0];
	ld.param.u32 	%r4, [ShaderKernel_fxBloomV_param_1];
	ld.param.u32 	%r5, [ShaderKernel_fxBloomV_param_2];
	ld.param.u32 	%r6, [ShaderKernel_fxBloomV_param_3];
	ld.param.u32 	%r7, [ShaderKernel_fxBloomV_param_4];
	ld.param.u64 	%rd5, [ShaderKernel_fxBloomV_param_5];
	cvta.to.global.u64 	%rd1, %rd4;
	cvta.to.global.u64 	%rd2, %rd5;
	.loc 1 38 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 38 1
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	.loc 1 38 1
	@!%p3 bra 	BB0_6;
	bra.uni 	BB0_1;

BB0_1:
	.loc 1 38 1
	cvt.rn.f32.s32	%f11, %r2;
	add.ftz.f32 	%f1, %f11, 0f3F000000;
	cvt.rn.f32.s32	%f12, %r3;
	add.ftz.f32 	%f2, %f12, 0f3F000000;
	.loc 1 38 1
	setp.gt.u32	%p4, %r1, 10;
	@%p4 bra 	BB0_3;

	.loc 1 38 1
	mul.wide.u32 	%rd6, %r1, 16;
	mov.u64 	%rd7, ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local;
	add.s64 	%rd8, %rd7, %rd6;
	add.s64 	%rd9, %rd2, %rd6;
	ld.global.v4.f32 	{%f13, %f14, %f15, %f16}, [%rd9];
	st.shared.v4.f32 	[%rd8], {%f13, %f14, %f15, %f16};

BB0_3:
	.loc 1 38 1
	bar.sync 	0;
	.loc 1 38 105
	// inline asm
	tex.2d.v4.f32.f32 {%f21, %f22, %f23, %f24}, [texture1_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 38 111
	// inline asm
	tex.2d.v4.f32.f32 {%f27, %f28, %f29, %f30}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	.loc 1 38 1
	add.ftz.f32 	%f115, %f1, 0f00000000;
	add.ftz.f32 	%f38, %f2, 0fBF800000;
	.loc 1 38 111
	// inline asm
	tex.2d.v4.f32.f32 {%f33, %f34, %f35, %f36}, [texture0_RECT, {%f115, %f38}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f117, %f118, %f119, %f120}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f122, %f35, %f117, %f29;
	fma.rn.ftz.f32 	%f124, %f34, %f118, %f28;
	fma.rn.ftz.f32 	%f126, %f33, %f119, %f27;
	fma.rn.ftz.f32 	%f128, %f36, %f120, %f30;
	.loc 1 38 1
	add.ftz.f32 	%f44, %f2, 0f3F800000;
	mov.f32 	%f129, 0f3F800000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f39, %f40, %f41, %f42}, [texture0_RECT, {%f115, %f44}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f130, %f41, %f117, %f122;
	fma.rn.ftz.f32 	%f131, %f40, %f118, %f124;
	fma.rn.ftz.f32 	%f132, %f39, %f119, %f126;
	fma.rn.ftz.f32 	%f133, %f42, %f120, %f128;
	.loc 1 38 1
	add.ftz.f32 	%f50, %f2, 0fC0000000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f45, %f46, %f47, %f48}, [texture0_RECT, {%f115, %f50}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f134, %f135, %f136, %f137}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+16];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f139, %f47, %f134, %f130;
	fma.rn.ftz.f32 	%f141, %f46, %f135, %f131;
	fma.rn.ftz.f32 	%f143, %f45, %f136, %f132;
	fma.rn.ftz.f32 	%f145, %f48, %f137, %f133;
	.loc 1 38 1
	add.ftz.f32 	%f56, %f2, 0f40000000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f51, %f52, %f53, %f54}, [texture0_RECT, {%f115, %f56}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f146, %f53, %f134, %f139;
	fma.rn.ftz.f32 	%f147, %f52, %f135, %f141;
	fma.rn.ftz.f32 	%f148, %f51, %f136, %f143;
	fma.rn.ftz.f32 	%f149, %f54, %f137, %f145;
	.loc 1 38 1
	add.ftz.f32 	%f62, %f2, 0fC0400000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f57, %f58, %f59, %f60}, [texture0_RECT, {%f115, %f62}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f150, %f151, %f152, %f153}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+32];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f155, %f59, %f150, %f146;
	fma.rn.ftz.f32 	%f157, %f58, %f151, %f147;
	fma.rn.ftz.f32 	%f159, %f57, %f152, %f148;
	fma.rn.ftz.f32 	%f161, %f60, %f153, %f149;
	.loc 1 38 1
	add.ftz.f32 	%f68, %f2, 0f40400000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f63, %f64, %f65, %f66}, [texture0_RECT, {%f115, %f68}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f162, %f65, %f150, %f155;
	fma.rn.ftz.f32 	%f163, %f64, %f151, %f157;
	fma.rn.ftz.f32 	%f164, %f63, %f152, %f159;
	fma.rn.ftz.f32 	%f165, %f66, %f153, %f161;
	.loc 1 38 1
	add.ftz.f32 	%f74, %f2, 0fC0800000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f69, %f70, %f71, %f72}, [texture0_RECT, {%f115, %f74}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f166, %f167, %f168, %f169}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+48];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f171, %f71, %f166, %f162;
	fma.rn.ftz.f32 	%f173, %f70, %f167, %f163;
	fma.rn.ftz.f32 	%f175, %f69, %f168, %f164;
	fma.rn.ftz.f32 	%f177, %f72, %f169, %f165;
	.loc 1 38 1
	add.ftz.f32 	%f80, %f2, 0f40800000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f75, %f76, %f77, %f78}, [texture0_RECT, {%f115, %f80}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f178, %f77, %f166, %f171;
	fma.rn.ftz.f32 	%f179, %f76, %f167, %f173;
	fma.rn.ftz.f32 	%f180, %f75, %f168, %f175;
	fma.rn.ftz.f32 	%f181, %f78, %f169, %f177;
	.loc 1 38 1
	add.ftz.f32 	%f86, %f2, 0fC0A00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f81, %f82, %f83, %f84}, [texture0_RECT, {%f115, %f86}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f182, %f183, %f184, %f185}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+64];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f187, %f83, %f182, %f178;
	fma.rn.ftz.f32 	%f189, %f82, %f183, %f179;
	fma.rn.ftz.f32 	%f191, %f81, %f184, %f180;
	fma.rn.ftz.f32 	%f193, %f84, %f185, %f181;
	.loc 1 38 1
	add.ftz.f32 	%f92, %f2, 0f40A00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f87, %f88, %f89, %f90}, [texture0_RECT, {%f115, %f92}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f194, %f89, %f182, %f187;
	fma.rn.ftz.f32 	%f195, %f88, %f183, %f189;
	fma.rn.ftz.f32 	%f196, %f87, %f184, %f191;
	fma.rn.ftz.f32 	%f197, %f90, %f185, %f193;
	.loc 1 38 1
	add.ftz.f32 	%f98, %f2, 0fC0C00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f93, %f94, %f95, %f96}, [texture0_RECT, {%f115, %f98}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f198, %f199, %f200, %f201}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+80];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f203, %f95, %f198, %f194;
	fma.rn.ftz.f32 	%f205, %f94, %f199, %f195;
	fma.rn.ftz.f32 	%f207, %f93, %f200, %f196;
	fma.rn.ftz.f32 	%f209, %f96, %f201, %f197;
	.loc 1 38 1
	add.ftz.f32 	%f104, %f2, 0f40C00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f99, %f100, %f101, %f102}, [texture0_RECT, {%f115, %f104}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f210, %f101, %f198, %f203;
	fma.rn.ftz.f32 	%f211, %f100, %f199, %f205;
	fma.rn.ftz.f32 	%f212, %f99, %f200, %f207;
	fma.rn.ftz.f32 	%f213, %f102, %f201, %f209;
	.loc 1 38 1
	add.ftz.f32 	%f110, %f2, 0fC0E00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f105, %f106, %f107, %f108}, [texture0_RECT, {%f115, %f110}];
	// inline asm
	.loc 1 38 1
	ld.shared.v4.f32 	{%f214, %f215, %f216, %f217}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+96];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f219, %f107, %f214, %f210;
	fma.rn.ftz.f32 	%f221, %f106, %f215, %f211;
	fma.rn.ftz.f32 	%f223, %f105, %f216, %f212;
	fma.rn.ftz.f32 	%f225, %f108, %f217, %f213;
	.loc 1 38 1
	add.ftz.f32 	%f116, %f2, 0f40E00000;
	.loc 1 38 119
	// inline asm
	tex.2d.v4.f32.f32 {%f111, %f112, %f113, %f114}, [texture0_RECT, {%f115, %f116}];
	// inline asm
	.loc 1 38 1
	fma.rn.ftz.f32 	%f226, %f113, %f214, %f219;
	fma.rn.ftz.f32 	%f227, %f112, %f215, %f221;
	fma.rn.ftz.f32 	%f228, %f111, %f216, %f223;
	fma.rn.ftz.f32 	%f229, %f114, %f217, %f225;
	.loc 1 38 1
	ld.shared.v4.f32 	{%f230, %f231, %f232, %f233}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+128];
	ld.shared.v4.f32 	{%f235, %f236, %f237, %f238}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+112];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f240, %f226, %f235, %f230;
	fma.rn.ftz.f32 	%f243, %f227, %f236, %f231;
	fma.rn.ftz.f32 	%f246, %f228, %f237, %f232;
	fma.rn.ftz.f32 	%f249, %f229, %f238, %f233;
	.loc 1 38 1
	ld.shared.v4.f32 	{%f250, %f251, %f252, %f253}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+144];
	.loc 1 38 1
	fma.rn.ftz.f32 	%f255, %f240, %f250, %f23;
	fma.rn.ftz.f32 	%f257, %f243, %f251, %f22;
	fma.rn.ftz.f32 	%f259, %f246, %f252, %f21;
	fma.rn.ftz.f32 	%f261, %f249, %f253, %f24;
	.loc 1 38 1
	ld.shared.v4.f32 	{%f262, %f263, %f264, %f265}, [ShaderKernel_fxBloomV$__cuda_local_var_170276_633_non_const_p_local+160];
	.loc 1 38 1
	sub.ftz.f32 	%f267, %f129, %f262;
	mul.ftz.f32 	%f268, %f267, %f23;
	fma.rn.ftz.f32 	%f3, %f262, %f255, %f268;
	sub.ftz.f32 	%f270, %f129, %f263;
	mul.ftz.f32 	%f271, %f270, %f22;
	fma.rn.ftz.f32 	%f4, %f263, %f257, %f271;
	sub.ftz.f32 	%f273, %f129, %f264;
	mul.ftz.f32 	%f274, %f273, %f21;
	fma.rn.ftz.f32 	%f5, %f264, %f259, %f274;
	sub.ftz.f32 	%f276, %f129, %f265;
	mul.ftz.f32 	%f277, %f276, %f24;
	fma.rn.ftz.f32 	%f6, %f265, %f261, %f277;
	.loc 1 38 1
	mad.lo.s32 	%r13, %r3, %r4, %r2;
	.loc 1 38 1
	cvt.s64.s32	%rd3, %r13;
	.loc 1 38 1
	setp.eq.s32	%p5, %r5, 0;
	@%p5 bra 	BB0_5;

	.loc 1 38 1
	shl.b64 	%rd42, %rd3, 4;
	add.s64 	%rd43, %rd1, %rd42;
	st.global.v4.f32 	[%rd43], {%f5, %f4, %f3, %f6};
	bra.uni 	BB0_6;

BB0_5:
	.loc 1 38 1
	shl.b64 	%rd44, %rd3, 3;
	add.s64 	%rd45, %rd1, %rd44;
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f5;
	mov.b16 	%rs1, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f4;
	mov.b16 	%rs2, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f3;
	mov.b16 	%rs3, %temp;
}
	.loc 2 3513 10
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f6;
	mov.b16 	%rs4, %temp;
}
	.loc 1 38 241
	st.global.v4.u16 	[%rd45], {%rs1, %rs2, %rs3, %rs4};

BB0_6:
	.loc 1 38 2
	ret;
}


