//
// 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;

.visible .entry ShaderKernel_PrColorToneCurves(
	.param .u64 ShaderKernel_PrColorToneCurves_param_0,
	.param .u32 ShaderKernel_PrColorToneCurves_param_1,
	.param .u32 ShaderKernel_PrColorToneCurves_param_2,
	.param .u32 ShaderKernel_PrColorToneCurves_param_3,
	.param .u32 ShaderKernel_PrColorToneCurves_param_4,
	.param .u64 ShaderKernel_PrColorToneCurves_param_5,
	.param .u64 ShaderKernel_PrColorToneCurves_param_6,
	.param .u64 ShaderKernel_PrColorToneCurves_param_7,
	.param .u64 ShaderKernel_PrColorToneCurves_param_8
)
{
	.reg .pred 	%p<23>;
	.reg .s16 	%rs<5>;
	.reg .s32 	%r<24>;
	.reg .f32 	%f<148>;
	.reg .s64 	%rd<25>;


	ld.param.u64 	%rd2, [ShaderKernel_PrColorToneCurves_param_0];
	ld.param.u32 	%r3, [ShaderKernel_PrColorToneCurves_param_1];
	ld.param.u32 	%r4, [ShaderKernel_PrColorToneCurves_param_2];
	ld.param.u32 	%r5, [ShaderKernel_PrColorToneCurves_param_3];
	ld.param.u32 	%r6, [ShaderKernel_PrColorToneCurves_param_4];
	ld.param.u64 	%rd3, [ShaderKernel_PrColorToneCurves_param_8];
	mov.u32 	%r7, %ntid.x;
	mov.u32 	%r8, %ctaid.x;
	mov.u32 	%r9, %tid.x;
	mad.lo.s32 	%r1, %r7, %r8, %r9;
	mov.u32 	%r10, %ntid.y;
	mov.u32 	%r11, %ctaid.y;
	mov.u32 	%r12, %tid.y;
	mad.lo.s32 	%r2, %r10, %r11, %r12;
	setp.lt.s32	%p1, %r1, %r5;
	setp.lt.s32	%p2, %r2, %r6;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB0_26;
	bra.uni 	BB0_1;

BB0_1:
	cvta.to.global.u64 	%rd5, %rd3;
	cvt.rn.f32.s32	%f47, %r1;
	add.ftz.f32 	%f45, %f47, 0f3F000000;
	cvt.rn.f32.s32	%f48, %r2;
	add.ftz.f32 	%f46, %f48, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f41, %f42, %f43, %f44}, [texture0_RECT, {%f45, %f46}];
	// inline asm
	cvt.ftz.sat.f32.f32	%f6, %f43;
	cvt.ftz.sat.f32.f32	%f8, %f42;
	cvt.ftz.sat.f32.f32	%f10, %f41;
	ld.global.u32 	%r13, [%rd5];
	setp.eq.s32	%p4, %r13, 0;
	mov.f32 	%f141, %f6;
	mov.f32 	%f144, %f8;
	mov.f32 	%f147, %f10;
	@%p4 bra 	BB0_12;

	setp.gtu.ftz.f32	%p5, %f6, 0f00000000;
	@%p5 bra 	BB0_4;

	mov.f32 	%f133, 0f3A000000;
	bra.uni 	BB0_5;

BB0_4:
	lg2.approx.ftz.f32 	%f50, %f6;
	mul.ftz.f32 	%f51, %f50, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f52, %f51;
	fma.rn.ftz.f32 	%f133, %f52, 0f3F7FF000, 0f3A000000;

BB0_5:
	setp.gtu.ftz.f32	%p6, %f8, 0f00000000;
	@%p6 bra 	BB0_7;

	mov.f32 	%f134, 0f3A000000;
	bra.uni 	BB0_8;

BB0_7:
	lg2.approx.ftz.f32 	%f54, %f8;
	mul.ftz.f32 	%f55, %f54, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f56, %f55;
	fma.rn.ftz.f32 	%f134, %f56, 0f3F7FF000, 0f3A000000;

BB0_8:
	setp.gtu.ftz.f32	%p7, %f10, 0f00000000;
	@%p7 bra 	BB0_10;

	mov.f32 	%f135, 0f3A000000;
	bra.uni 	BB0_11;

BB0_10:
	lg2.approx.ftz.f32 	%f58, %f10;
	mul.ftz.f32 	%f59, %f58, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f60, %f59;
	fma.rn.ftz.f32 	%f135, %f60, 0f3F7FF000, 0f3A000000;

BB0_11:
	mov.f32 	%f78, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f61, %f62, %f63, %f64}, [texture2_2D, {%f133, %f78}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f67, %f68, %f69, %f70}, [texture2_2D, {%f134, %f78}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f73, %f74, %f75, %f76}, [texture2_2D, {%f135, %f78}];
	// inline asm
	setp.gt.ftz.f32	%p8, %f64, %f70;
	selp.f32	%f79, %f70, %f64, %p8;
	setp.gt.ftz.f32	%p9, %f79, %f76;
	selp.f32	%f80, %f76, %f79, %p9;
	selp.f32	%f81, %f64, %f70, %p8;
	setp.gt.ftz.f32	%p10, %f81, %f76;
	selp.f32	%f82, %f81, %f76, %p10;
	setp.gt.ftz.f32	%p11, %f6, %f8;
	selp.f32	%f83, %f6, %f8, %p11;
	setp.gt.ftz.f32	%p12, %f83, %f10;
	selp.f32	%f84, %f83, %f10, %p12;
	selp.f32	%f85, %f8, %f6, %p11;
	setp.gt.ftz.f32	%p13, %f85, %f10;
	selp.f32	%f86, %f10, %f85, %p13;
	sub.ftz.f32 	%f87, %f84, %f86;
	sub.ftz.f32 	%f88, %f82, %f80;
	setp.gt.ftz.f32	%p14, %f87, 0f2EDBE6FF;
	selp.f32	%f89, %f87, 0f2EDBE6FF, %p14;
	mov.f32 	%f90, 0f3F800000;
	div.rn.ftz.f32 	%f91, %f90, %f89;
	mul.ftz.f32 	%f92, %f91, %f88;
	sub.ftz.f32 	%f93, %f6, %f86;
	sub.ftz.f32 	%f94, %f8, %f86;
	sub.ftz.f32 	%f95, %f10, %f86;
	fma.rn.ftz.f32 	%f17, %f92, %f93, %f80;
	fma.rn.ftz.f32 	%f18, %f92, %f94, %f80;
	fma.rn.ftz.f32 	%f19, %f92, %f95, %f80;
	mov.f32 	%f141, %f17;
	mov.f32 	%f144, %f18;
	mov.f32 	%f147, %f19;

BB0_12:
	mov.f32 	%f145, %f147;
	mov.f32 	%f146, %f145;
	mov.f32 	%f142, %f144;
	mov.f32 	%f143, %f142;
	mov.f32 	%f139, %f141;
	mov.f32 	%f140, %f139;
	sub.ftz.f32 	%f23, %f43, %f6;
	sub.ftz.f32 	%f24, %f42, %f8;
	sub.ftz.f32 	%f25, %f41, %f10;
	ld.global.u32 	%r14, [%rd5+4];
	setp.eq.s32	%p15, %r14, 0;
	@%p15 bra 	BB0_23;

	setp.gtu.ftz.f32	%p16, %f140, 0f00000000;
	@%p16 bra 	BB0_15;

	mov.f32 	%f136, 0f3A000000;
	bra.uni 	BB0_16;

BB0_15:
	lg2.approx.ftz.f32 	%f97, %f140;
	mul.ftz.f32 	%f98, %f97, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f99, %f98;
	fma.rn.ftz.f32 	%f136, %f99, 0f3F7FF000, 0f3A000000;

BB0_16:
	setp.gtu.ftz.f32	%p17, %f143, 0f00000000;
	@%p17 bra 	BB0_18;

	mov.f32 	%f137, 0f3A000000;
	bra.uni 	BB0_19;

BB0_18:
	lg2.approx.ftz.f32 	%f101, %f143;
	mul.ftz.f32 	%f102, %f101, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f103, %f102;
	fma.rn.ftz.f32 	%f137, %f103, 0f3F7FF000, 0f3A000000;

BB0_19:
	setp.gtu.ftz.f32	%p18, %f146, 0f00000000;
	@%p18 bra 	BB0_21;

	mov.f32 	%f138, 0f3A000000;
	bra.uni 	BB0_22;

BB0_21:
	lg2.approx.ftz.f32 	%f105, %f146;
	mul.ftz.f32 	%f106, %f105, 0f3EE8BA2F;
	ex2.approx.ftz.f32 	%f107, %f106;
	fma.rn.ftz.f32 	%f138, %f107, 0f3F7FF000, 0f3A000000;

BB0_22:
	mov.f32 	%f125, 0f3F000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f108, %f109, %f110, %f111}, [texture2_2D, {%f136, %f125}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f114, %f115, %f116, %f117}, [texture2_2D, {%f137, %f125}];
	// inline asm
	// inline asm
	tex.2d.v4.f32.f32 {%f120, %f121, %f122, %f123}, [texture2_2D, {%f138, %f125}];
	// inline asm
	mov.f32 	%f146, %f122;
	mov.f32 	%f143, %f115;
	mov.f32 	%f140, %f108;

BB0_23:
	setp.lt.ftz.f32	%p19, %f41, 0f00000000;
	setp.lt.ftz.f32	%p20, %f42, 0f00000000;
	setp.lt.ftz.f32	%p21, %f43, 0f00000000;
	mov.f32 	%f126, 0f3F800000;
	sub.ftz.f32 	%f127, %f126, %f140;
	selp.f32	%f128, %f127, %f140, %p21;
	fma.rn.ftz.f32 	%f38, %f23, %f128, %f140;
	sub.ftz.f32 	%f129, %f126, %f143;
	selp.f32	%f130, %f129, %f143, %p20;
	fma.rn.ftz.f32 	%f39, %f24, %f130, %f143;
	sub.ftz.f32 	%f131, %f126, %f146;
	selp.f32	%f132, %f131, %f146, %p19;
	fma.rn.ftz.f32 	%f40, %f25, %f132, %f146;
	mad.lo.s32 	%r23, %r2, %r3, %r1;
	cvt.s64.s32	%rd1, %r23;
	setp.eq.s32	%p22, %r4, 0;
	@%p22 bra 	BB0_25;

	cvta.to.global.u64 	%rd19, %rd2;
	shl.b64 	%rd20, %rd1, 4;
	add.s64 	%rd21, %rd19, %rd20;
	st.global.v4.f32 	[%rd21], {%f40, %f39, %f38, %f44};
	bra.uni 	BB0_26;

BB0_25:
	cvta.to.global.u64 	%rd22, %rd2;
	shl.b64 	%rd23, %rd1, 3;
	add.s64 	%rd24, %rd22, %rd23;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f38;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f39;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f40;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f44;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd24], {%rs3, %rs2, %rs1, %rs4};

BB0_26:
	ret;
}


