//
// Generated by NVIDIA NVVM Compiler
// Compiler built on Fri Jul 25 04:36:16 2014 (1406288176)
// Cuda compilation tools, release 6.5, V6.5.13
//

.version 4.1
.target sm_30
.address_size 64

.global .texref texture0_RECT;
.global .texref texture2_2D;
// ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local has been demoted

.visible .entry ShaderKernel_PrColorBasicCorrection(
	.param .u64 ShaderKernel_PrColorBasicCorrection_param_0,
	.param .u32 ShaderKernel_PrColorBasicCorrection_param_1,
	.param .u32 ShaderKernel_PrColorBasicCorrection_param_2,
	.param .u32 ShaderKernel_PrColorBasicCorrection_param_3,
	.param .u32 ShaderKernel_PrColorBasicCorrection_param_4,
	.param .u64 ShaderKernel_PrColorBasicCorrection_param_5,
	.param .u64 ShaderKernel_PrColorBasicCorrection_param_6,
	.param .u64 ShaderKernel_PrColorBasicCorrection_param_7,
	.param .u64 ShaderKernel_PrColorBasicCorrection_param_8
)
{
	.reg .pred 	%p<15>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<26>;
	.reg .f32 	%f<160>;
	.reg .s64 	%rd<22>;
	// demoted variable
	.shared .align 16 .b8 ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local[48];

	ld.param.u64 	%rd3, [ShaderKernel_PrColorBasicCorrection_param_0];
	ld.param.u32 	%r4, [ShaderKernel_PrColorBasicCorrection_param_1];
	ld.param.u32 	%r5, [ShaderKernel_PrColorBasicCorrection_param_2];
	ld.param.u32 	%r6, [ShaderKernel_PrColorBasicCorrection_param_3];
	ld.param.u32 	%r7, [ShaderKernel_PrColorBasicCorrection_param_4];
	ld.param.u64 	%rd4, [ShaderKernel_PrColorBasicCorrection_param_5];
	ld.param.u64 	%rd5, [ShaderKernel_PrColorBasicCorrection_param_8];
	cvta.to.global.u64 	%rd1, %rd5;
	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;
	setp.lt.s32	%p1, %r2, %r6;
	setp.lt.s32	%p2, %r3, %r7;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB0_20;
	bra.uni 	BB0_1;

BB0_1:
	setp.gt.u32	%p4, %r1, 2;
	@%p4 bra 	BB0_3;

	cvta.to.global.u64 	%rd6, %rd4;
	mul.wide.u32 	%rd7, %r1, 16;
	mov.u64 	%rd8, ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local;
	add.s64 	%rd9, %rd8, %rd7;
	add.s64 	%rd10, %rd6, %rd7;
	ld.global.v4.f32 	{%f56, %f57, %f58, %f59}, [%rd10];
	st.shared.v4.f32 	[%rd9], {%f56, %f57, %f58, %f59};

BB0_3:
	cvt.rn.f32.s32	%f64, %r2;
	add.ftz.f32 	%f1, %f64, 0f3F000000;
	cvt.rn.f32.s32	%f65, %r3;
	add.ftz.f32 	%f2, %f65, 0f3F000000;
	bar.sync 	0;
	// inline asm
	tex.2d.v4.f32.f32 {%f66, %f67, %f68, %f69}, [texture0_RECT, {%f1, %f2}];
	// inline asm
	ld.global.u32 	%r13, [%rd1];
	setp.eq.s32	%p5, %r13, 0;
	mov.f32 	%f146, %f68;
	mov.f32 	%f147, %f67;
	mov.f32 	%f148, %f66;
	@%p5 bra 	BB0_5;

	ld.shared.v4.f32 	{%f72, %f73, %f74, %f75}, [ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local];
	mul.ftz.f32 	%f146, %f68, %f72;
	mul.ftz.f32 	%f147, %f67, %f73;
	mul.ftz.f32 	%f148, %f66, %f74;

BB0_5:
	abs.ftz.f32 	%f79, %f146;
	cvt.ftz.sat.f32.f32	%f154, %f79;
	abs.ftz.f32 	%f80, %f147;
	cvt.ftz.sat.f32.f32	%f155, %f80;
	abs.ftz.f32 	%f81, %f148;
	cvt.ftz.sat.f32.f32	%f156, %f81;
	sub.ftz.f32 	%f150, %f79, %f154;
	sub.ftz.f32 	%f151, %f80, %f155;
	sub.ftz.f32 	%f152, %f81, %f156;
	ld.global.u32 	%r14, [%rd1+4];
	setp.eq.s32	%p6, %r14, 0;
	@%p6 bra 	BB0_10;

	ld.shared.v4.f32 	{%f82, %f83, %f84, %f85}, [ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local+32];
	mul.ftz.f32 	%f87, %f155, %f83;
	fma.rn.ftz.f32 	%f89, %f154, %f82, %f87;
	fma.rn.ftz.f32 	%f22, %f156, %f84, %f89;
	setp.gtu.ftz.f32	%p7, %f22, 0f00000000;
	@%p7 bra 	BB0_8;

	mov.f32 	%f149, 0f3A000000;
	bra.uni 	BB0_9;

BB0_8:
	lg2.approx.ftz.f32 	%f92, %f22;
	mul.ftz.f32 	%f93, %f92, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f94, %f93;
	fma.rn.ftz.f32 	%f149, %f94, 0f3F7FF000, 0f3A000000;

BB0_9:
	fma.rn.ftz.f32 	%f100, %f22, 0f00000000, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f95, %f96, %f97, %f98}, [texture2_2D, {%f149, %f100}];
	// inline asm
	fma.rn.ftz.f32 	%f101, %f95, %f154, %f96;
	fma.rn.ftz.f32 	%f102, %f95, %f155, %f96;
	fma.rn.ftz.f32 	%f103, %f95, %f156, %f96;
	fma.rn.ftz.f32 	%f104, %f150, %f101, %f101;
	fma.rn.ftz.f32 	%f105, %f151, %f102, %f102;
	fma.rn.ftz.f32 	%f106, %f152, %f103, %f103;
	cvt.ftz.sat.f32.f32	%f154, %f104;
	cvt.ftz.sat.f32.f32	%f155, %f105;
	cvt.ftz.sat.f32.f32	%f156, %f106;
	sub.ftz.f32 	%f150, %f104, %f154;
	sub.ftz.f32 	%f151, %f105, %f155;
	sub.ftz.f32 	%f152, %f106, %f156;

BB0_10:
	ld.global.u32 	%r15, [%rd1+8];
	setp.eq.s32	%p8, %r15, 0;
	@%p8 bra 	BB0_15;

	ld.shared.v4.f32 	{%f107, %f108, %f109, %f110}, [ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local+32];
	mul.ftz.f32 	%f112, %f155, %f108;
	fma.rn.ftz.f32 	%f114, %f154, %f107, %f112;
	fma.rn.ftz.f32 	%f37, %f156, %f109, %f114;
	setp.gtu.ftz.f32	%p9, %f37, 0f00000000;
	@%p9 bra 	BB0_13;

	mov.f32 	%f153, 0f3A000000;
	bra.uni 	BB0_14;

BB0_13:
	lg2.approx.ftz.f32 	%f117, %f37;
	mul.ftz.f32 	%f118, %f117, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f119, %f118;
	fma.rn.ftz.f32 	%f153, %f119, 0f3F7FF000, 0f3A000000;

BB0_14:
	fma.rn.ftz.f32 	%f125, %f37, 0f00000000, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f120, %f121, %f122, %f123}, [texture2_2D, {%f153, %f125}];
	// inline asm
	fma.rn.ftz.f32 	%f154, %f122, %f154, %f123;
	fma.rn.ftz.f32 	%f155, %f122, %f155, %f123;
	fma.rn.ftz.f32 	%f156, %f122, %f156, %f123;

BB0_15:
	setp.lt.ftz.f32	%p10, %f146, 0f00000000;
	selp.f32	%f126, 0fBF800000, 0f3F800000, %p10;
	setp.lt.ftz.f32	%p11, %f147, 0f00000000;
	selp.f32	%f127, 0fBF800000, 0f3F800000, %p11;
	setp.lt.ftz.f32	%p12, %f148, 0f00000000;
	selp.f32	%f128, 0fBF800000, 0f3F800000, %p12;
	fma.rn.ftz.f32 	%f129, %f150, %f154, %f154;
	fma.rn.ftz.f32 	%f130, %f151, %f155, %f155;
	fma.rn.ftz.f32 	%f131, %f152, %f156, %f156;
	mul.ftz.f32 	%f157, %f129, %f126;
	mul.ftz.f32 	%f158, %f130, %f127;
	mul.ftz.f32 	%f159, %f131, %f128;
	ld.global.u32 	%r16, [%rd1+12];
	setp.eq.s32	%p13, %r16, 0;
	@%p13 bra 	BB0_17;

	ld.shared.v4.f32 	{%f132, %f133, %f134, %f135}, [ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local+32];
	mul.ftz.f32 	%f137, %f158, %f133;
	fma.rn.ftz.f32 	%f139, %f157, %f132, %f137;
	fma.rn.ftz.f32 	%f141, %f159, %f134, %f139;
	sub.ftz.f32 	%f142, %f157, %f141;
	sub.ftz.f32 	%f143, %f158, %f141;
	sub.ftz.f32 	%f144, %f159, %f141;
	ld.shared.f32 	%f145, [ShaderKernel_PrColorBasicCorrection$__cuda_local_var_180688_593_non_const_p_local+16];
	fma.rn.ftz.f32 	%f157, %f145, %f142, %f141;
	fma.rn.ftz.f32 	%f158, %f145, %f143, %f141;
	fma.rn.ftz.f32 	%f159, %f145, %f144, %f141;

BB0_17:
	mad.lo.s32 	%r25, %r3, %r4, %r2;
	cvt.s64.s32	%rd2, %r25;
	setp.eq.s32	%p14, %r5, 0;
	@%p14 bra 	BB0_19;

	cvta.to.global.u64 	%rd16, %rd3;
	shl.b64 	%rd17, %rd2, 4;
	add.s64 	%rd18, %rd16, %rd17;
	st.global.v4.f32 	[%rd18], {%f159, %f158, %f157, %f69};
	bra.uni 	BB0_20;

BB0_19:
	cvta.to.global.u64 	%rd19, %rd3;
	shl.b64 	%rd20, %rd2, 3;
	add.s64 	%rd21, %rd19, %rd20;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f69;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f157;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f158;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f159;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd21], {%rs4, %rs3, %rs2, %rs1};

BB0_20:
	ret;
}


