//
// 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

.const .align 4 .b8 kRGB32f_To_601YPbPr[36] = {135, 22, 153, 62, 162, 69, 22, 63, 213, 120, 233, 61, 33, 201, 44, 190, 111, 155, 169, 190, 0, 0, 0, 63, 0, 0, 0, 63, 70, 94, 214, 190, 232, 134, 166, 189};
.const .align 4 .b8 k601YPbPr_To_RGB32f[36] = {0, 0, 128, 63, 0, 0, 0, 0, 188, 116, 179, 63, 0, 0, 128, 63, 152, 50, 176, 190, 158, 209, 54, 191, 0, 0, 128, 63, 229, 208, 226, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_601YCbCr[36] = {70, 246, 130, 66, 145, 141, 0, 67, 94, 186, 199, 65, 33, 48, 23, 194, 240, 103, 148, 194, 0, 0, 224, 66, 0, 0, 224, 66, 111, 146, 187, 194, 70, 182, 145, 193};
.const .align 4 .b8 k601YCbCr_To_RGB32f[36] = {37, 160, 149, 59, 0, 0, 0, 0, 182, 23, 205, 59, 37, 160, 149, 59, 40, 15, 201, 186, 156, 239, 80, 187, 37, 160, 149, 59, 236, 155, 1, 60, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_601YCbCr[36] = {219, 121, 131, 62, 152, 14, 1, 63, 18, 131, 200, 61, 174, 199, 23, 190, 238, 252, 148, 190, 197, 224, 224, 62, 197, 224, 224, 62, 217, 78, 188, 190, 174, 71, 146, 189};
.const .align 4 .b8 k601YCbCr_To_RGB8u[36] = {127, 10, 149, 63, 0, 0, 0, 0, 160, 74, 204, 63, 127, 10, 149, 63, 254, 148, 200, 190, 184, 30, 80, 191, 127, 10, 149, 63, 78, 26, 1, 64, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_601YCbCrFullRange[36] = {135, 22, 153, 62, 162, 69, 22, 63, 213, 120, 233, 61, 166, 27, 44, 190, 39, 241, 168, 190, 250, 254, 254, 62, 250, 254, 254, 62, 43, 135, 213, 190, 59, 223, 165, 189};
.const .align 4 .b8 k601YCbCrFullRange_To_RGB8u[36] = {0, 0, 128, 63, 0, 0, 0, 0, 72, 193, 178, 63, 0, 0, 128, 63, 143, 130, 175, 190, 225, 26, 54, 191, 0, 0, 128, 63, 20, 238, 225, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_601YCbCrFullRange[36] = {113, 125, 152, 66, 92, 175, 21, 67, 92, 143, 232, 65, 158, 111, 43, 194, 49, 72, 168, 194, 0, 0, 254, 66, 0, 0, 254, 66, 170, 177, 212, 194, 88, 57, 165, 193};
.const .align 4 .b8 k601YCbCrFullRange_To_RGB32f[36] = {129, 128, 128, 59, 0, 0, 0, 0, 188, 116, 179, 59, 129, 128, 128, 59, 194, 50, 176, 186, 179, 209, 54, 187, 129, 128, 128, 59, 229, 208, 226, 59, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_709YPbPr[36] = {208, 179, 89, 62, 89, 23, 55, 63, 152, 221, 147, 61, 186, 164, 234, 189, 210, 86, 197, 190, 0, 0, 0, 63, 0, 0, 0, 63, 190, 134, 232, 190, 16, 202, 59, 189};
.const .align 4 .b8 k709YPbPr_To_RGB32f[36] = {0, 0, 128, 63, 0, 0, 0, 0, 12, 147, 201, 63, 0, 0, 128, 63, 221, 209, 63, 190, 243, 173, 239, 190, 0, 0, 128, 63, 77, 132, 237, 63, 0, 0, 0, 0};
.const .align 4 .b8 kRGB32f_To_709YCbCr[36] = {106, 60, 58, 66, 6, 161, 28, 67, 244, 253, 124, 65, 223, 79, 205, 193, 8, 172, 172, 194, 0, 0, 224, 66, 0, 0, 224, 66, 195, 117, 203, 194, 236, 81, 36, 193};
.const .align 4 .b8 k709YCbCr_To_RGB32f[36] = {37, 160, 149, 59, 0, 0, 0, 0, 239, 94, 230, 59, 37, 160, 149, 59, 33, 57, 91, 186, 178, 245, 8, 187, 37, 160, 149, 59, 82, 185, 7, 60, 0, 0, 0, 0};
.const .align 4 .b8 k709YCbCrFullRange_To_RGB32f[36] = {131, 128, 128, 59, 0, 0, 0, 0, 28, 147, 201, 59, 131, 128, 128, 59, 61, 210, 63, 186, 248, 173, 239, 186, 131, 128, 128, 59, 82, 132, 237, 59, 0, 0, 0, 0};
.const .align 4 .b8 kRGB8u_To_709YCbCr[36] = {207, 247, 58, 62, 53, 62, 29, 63, 231, 251, 125, 61, 184, 30, 206, 189, 23, 89, 173, 190, 197, 224, 224, 62, 197, 224, 224, 62, 12, 66, 204, 190, 195, 245, 36, 189};
.const .align 4 .b8 k709YCbCr_To_RGB8u[36] = {127, 10, 149, 63, 0, 0, 0, 0, 147, 120, 229, 63, 127, 10, 149, 63, 53, 94, 90, 190, 205, 108, 8, 191, 127, 10, 149, 63, 154, 49, 7, 64, 0, 0, 0, 0};
.const .align 4 .b8 k709YCbCr_To_601YCbCr[36] = {0, 0, 128, 63, 23, 100, 203, 61, 1, 77, 68, 62, 0, 0, 0, 0, 18, 103, 125, 63, 10, 158, 226, 189, 0, 0, 0, 0, 61, 98, 148, 189, 249, 191, 123, 63};
.const .align 4 .b8 k601YCbCr_To_709YCbCr[36] = {0, 0, 128, 63, 122, 165, 236, 189, 179, 237, 84, 190, 0, 0, 0, 0, 204, 98, 130, 63, 216, 188, 234, 61, 0, 0, 0, 0, 74, 179, 153, 61, 234, 61, 131, 63};
.const .align 4 .b8 kYCbCrOffset[12] = {0, 0, 128, 65, 0, 0, 0, 67, 0, 0, 0, 67};
.const .align 4 .b8 kYCbCrFullRangeOffset[12] = {0, 0, 0, 0, 0, 0, 0, 67, 0, 0, 0, 67};
.const .align 4 .u32 kRandMax = 32767;

.visible .func  (.param .b32 func_retval0) _Z21CalculateThreeWayLumafffffffi(
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_0,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_1,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_2,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_3,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_4,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_5,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_6,
	.param .b32 _Z21CalculateThreeWayLumafffffffi_param_7
)
{
	.reg .pred 	%p<6>;
	.reg .s32 	%r<54>;
	.reg .f32 	%f<39>;


	ld.param.f32 	%f9, [_Z21CalculateThreeWayLumafffffffi_param_0];
	ld.param.f32 	%f10, [_Z21CalculateThreeWayLumafffffffi_param_1];
	ld.param.f32 	%f11, [_Z21CalculateThreeWayLumafffffffi_param_2];
	ld.param.f32 	%f14, [_Z21CalculateThreeWayLumafffffffi_param_3];
	ld.param.f32 	%f12, [_Z21CalculateThreeWayLumafffffffi_param_4];
	ld.param.f32 	%f13, [_Z21CalculateThreeWayLumafffffffi_param_5];
	ld.param.f32 	%f15, [_Z21CalculateThreeWayLumafffffffi_param_6];
	ld.param.u32 	%r1, [_Z21CalculateThreeWayLumafffffffi_param_7];
	sub.ftz.f32 	%f1, %f14, %f11;
	setp.neu.ftz.f32	%p1, %f1, 0f3F800000;
	sub.ftz.f32 	%f2, %f15, %f13;
	setp.neu.ftz.f32	%p2, %f2, 0f3F800000;
	or.pred  	%p3, %p1, %p2;
	mov.f32 	%f37, %f9;
	@!%p3 bra 	BB0_2;
	bra.uni 	BB0_1;

BB0_1:
	mov.u32 	%r2, %ctaid.x;
	mov.u32 	%r3, %ntid.x;
	mov.u32 	%r4, %ntid.y;
	mov.u32 	%r5, %ctaid.y;
	mov.u32 	%r6, %tid.y;
	mad.lo.s32 	%r7, %r4, %r5, %r6;
	mov.u32 	%r8, %tid.x;
	mad.lo.s32 	%r9, %r3, %r2, %r8;
	sub.s32 	%r10, %r9, %r7;
	add.s32 	%r11, %r10, -1;
	mov.u32 	%r12, 1;
	sub.s32 	%r13, %r12, %r10;
	add.s32 	%r14, %r7, %r13;
	add.s32 	%r15, %r14, -1;
	shl.b32 	%r16, %r11, 8;
	xor.b32  	%r17, %r15, %r16;
	mov.u32 	%r18, 2;
	sub.s32 	%r19, %r18, %r10;
	sub.s32 	%r20, %r19, %r17;
	shr.u32 	%r21, %r17, 13;
	xor.b32  	%r22, %r20, %r21;
	sub.s32 	%r23, %r11, %r17;
	sub.s32 	%r24, %r23, %r22;
	shr.u32 	%r25, %r22, 12;
	xor.b32  	%r26, %r24, %r25;
	sub.s32 	%r27, %r17, %r22;
	sub.s32 	%r28, %r27, %r26;
	shl.b32 	%r29, %r26, 16;
	xor.b32  	%r30, %r28, %r29;
	sub.s32 	%r31, %r22, %r26;
	sub.s32 	%r32, %r31, %r30;
	shr.u32 	%r33, %r30, 5;
	xor.b32  	%r34, %r32, %r33;
	sub.s32 	%r35, %r26, %r30;
	sub.s32 	%r36, %r35, %r34;
	shr.u32 	%r37, %r34, 3;
	xor.b32  	%r38, %r36, %r37;
	sub.s32 	%r39, %r30, %r34;
	sub.s32 	%r40, %r39, %r38;
	shl.b32 	%r41, %r38, 10;
	xor.b32  	%r42, %r40, %r41;
	sub.s32 	%r43, %r34, %r38;
	sub.s32 	%r44, %r43, %r42;
	shr.u32 	%r45, %r42, 15;
	xor.b32  	%r46, %r44, %r45;
	mad.lo.s32 	%r47, %r46, 1103515245, 12345;
	mad.lo.s32 	%r48, %r47, 1103515245, 12345;
	shr.u32 	%r49, %r47, 9;
	and.b32  	%r50, %r49, 32640;
	bfe.u32 	%r51, %r48, 16, 8;
	xor.b32  	%r52, %r51, %r50;
	cvt.rn.f32.s32	%f16, %r52;
	ld.const.u32 	%r53, [kRandMax];
	cvt.rn.f32.u32	%f17, %r53;
	mov.f32 	%f18, 0f40000000;
	div.approx.ftz.f32 	%f19, %f17, %f18;
	div.approx.ftz.f32 	%f20, %f16, %f19;
	add.ftz.f32 	%f21, %f20, 0fBF800000;
	fma.rn.ftz.f32 	%f3, %f21, 0f3B270D73, %f9;
	mov.f32 	%f37, %f3;

BB0_2:
	mov.f32 	%f4, %f37;
	sub.ftz.f32 	%f22, %f4, %f11;
	div.approx.ftz.f32 	%f23, %f22, %f1;
	mov.f32 	%f24, 0f00000000;
	max.ftz.f32 	%f5, %f24, %f23;
	setp.ltu.ftz.f32	%p4, %f5, 0f00000000;
	@%p4 bra 	BB0_4;

	lg2.approx.ftz.f32 	%f25, %f5;
	mul.ftz.f32 	%f26, %f25, %f12;
	ex2.approx.ftz.f32 	%f38, %f26;
	bra.uni 	BB0_5;

BB0_4:
	neg.ftz.f32 	%f27, %f5;
	lg2.approx.ftz.f32 	%f28, %f27;
	mul.ftz.f32 	%f29, %f28, %f12;
	ex2.approx.ftz.f32 	%f30, %f29;
	neg.ftz.f32 	%f38, %f30;

BB0_5:
	fma.rn.ftz.f32 	%f31, %f2, %f38, %f13;
	sub.ftz.f32 	%f32, %f4, %f31;
	fma.rn.ftz.f32 	%f33, %f32, 0f3F000000, %f31;
	setp.eq.s32	%p5, %r1, 0;
	selp.f32	%f34, %f31, %f33, %p5;
	sub.ftz.f32 	%f35, %f34, %f9;
	fma.rn.ftz.f32 	%f36, %f35, %f10, %f9;
	st.param.f32	[func_retval0+0], %f36;
	ret;
}

.visible .entry ThreeWayColorCorrector_MaskKernel(
	.param .u64 ThreeWayColorCorrector_MaskKernel_param_0,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_1,
	.param .u64 ThreeWayColorCorrector_MaskKernel_param_2,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_3,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_4,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_5,
	.param .u32 ThreeWayColorCorrector_MaskKernel_param_6
)
{
	.reg .pred 	%p<7>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<16>;
	.reg .f32 	%f<41>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd4, [ThreeWayColorCorrector_MaskKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_MaskKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_MaskKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_MaskKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_MaskKernel_param_4];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_MaskKernel_param_5];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_MaskKernel_param_6];
	mov.u32 	%r8, %ntid.x;
	mov.u32 	%r9, %ctaid.x;
	mov.u32 	%r10, %tid.x;
	mad.lo.s32 	%r1, %r8, %r9, %r10;
	mov.u32 	%r11, %ntid.y;
	mov.u32 	%r12, %ctaid.y;
	mov.u32 	%r13, %tid.y;
	mad.lo.s32 	%r2, %r11, %r12, %r13;
	setp.lt.s32	%p1, %r1, %r6;
	setp.lt.s32	%p2, %r2, %r7;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB1_10;
	bra.uni 	BB1_1;

BB1_1:
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd6, %rd4;
	mad.lo.s32 	%r14, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r14, 16;
	add.s64 	%rd2, %rd6, %rd7;
	mul.wide.s32 	%rd8, %r14, 8;
	add.s64 	%rd3, %rd6, %rd8;
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB1_3;

	ld.global.v4.f32 	{%f19, %f20, %f21, %f22}, [%rd2];
	mov.f32 	%f39, %f22;
	mov.f32 	%f3, %f21;
	mov.f32 	%f2, %f20;
	mov.f32 	%f1, %f19;
	bra.uni 	BB1_4;

BB1_3:
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f39, %temp;
	}

BB1_4:
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB1_6;

	mov.f32 	%f40, 0f3F800000;
	bra.uni 	BB1_7;

BB1_6:
	mad.lo.s32 	%r15, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r15, 4;
	add.s64 	%rd10, %rd1, %rd9;
	ld.global.f32 	%f40, [%rd10];

BB1_7:
	ld.const.f32 	%f24, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f25, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f26, %f25, 0f00000000;
	fma.rn.ftz.f32 	%f27, %f40, %f24, %f26;
	ld.const.f32 	%f28, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f16, %f28, 0f00000000, %f27;
	ld.const.f32 	%f29, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f30, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f31, %f30, 0f00000000;
	fma.rn.ftz.f32 	%f32, %f40, %f29, %f31;
	ld.const.f32 	%f33, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f17, %f33, 0f00000000, %f32;
	ld.const.f32 	%f34, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f35, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f36, %f35, 0f00000000;
	fma.rn.ftz.f32 	%f37, %f40, %f34, %f36;
	ld.const.f32 	%f38, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f18, %f38, 0f00000000, %f37;
	@%p4 bra 	BB1_9;

	st.global.v4.f32 	[%rd2], {%f18, %f17, %f16, %f39};
	bra.uni 	BB1_10;

BB1_9:
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f39;
	mov.b16 	%rs9, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f16;
	mov.b16 	%rs10, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f17;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f18;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd3], {%rs12, %rs11, %rs10, %rs9};

BB1_10:
	ret;
}

.visible .entry ThreeWayColorCorrector_TonalRangeKernel(
	.param .u64 ThreeWayColorCorrector_TonalRangeKernel_param_0,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_1,
	.param .u64 ThreeWayColorCorrector_TonalRangeKernel_param_2,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_3,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_4,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_5,
	.param .u32 ThreeWayColorCorrector_TonalRangeKernel_param_6,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_7,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_8,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_9,
	.param .f32 ThreeWayColorCorrector_TonalRangeKernel_param_10
)
{
	.reg .pred 	%p<21>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<14>;
	.reg .f32 	%f<78>;
	.reg .s64 	%rd<7>;


	ld.param.u64 	%rd3, [ThreeWayColorCorrector_TonalRangeKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_TonalRangeKernel_param_1];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_TonalRangeKernel_param_4];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_TonalRangeKernel_param_5];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_TonalRangeKernel_param_6];
	ld.param.f32 	%f23, [ThreeWayColorCorrector_TonalRangeKernel_param_7];
	ld.param.f32 	%f24, [ThreeWayColorCorrector_TonalRangeKernel_param_8];
	ld.param.f32 	%f25, [ThreeWayColorCorrector_TonalRangeKernel_param_9];
	ld.param.f32 	%f26, [ThreeWayColorCorrector_TonalRangeKernel_param_10];
	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 	BB2_20;
	bra.uni 	BB2_1;

BB2_1:
	cvta.to.global.u64 	%rd4, %rd3;
	mad.lo.s32 	%r13, %r2, %r3, %r1;
	mul.wide.s32 	%rd5, %r13, 16;
	add.s64 	%rd1, %rd4, %rd5;
	mul.wide.s32 	%rd6, %r13, 8;
	add.s64 	%rd2, %rd4, %rd6;
	setp.eq.s32	%p4, %r4, 0;
	@%p4 bra 	BB2_3;

	ld.global.v4.f32 	{%f27, %f28, %f29, %f30}, [%rd1];
	mov.f32 	%f73, %f30;
	mov.f32 	%f72, %f29;
	mov.f32 	%f71, %f28;
	mov.f32 	%f70, %f27;
	bra.uni 	BB2_4;

BB2_3:
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd2];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f70, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f71, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f72, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f73, %temp;
	}

BB2_4:
	ld.const.f32 	%f31, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f32, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f33, %f71, %f32;
	fma.rn.ftz.f32 	%f34, %f72, %f31, %f33;
	ld.const.f32 	%f35, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f13, %f70, %f35, %f34;
	add.ftz.f32 	%f36, %f23, %f24;
	mov.f32 	%f37, 0f3F800000;
	min.ftz.f32 	%f14, %f37, %f36;
	sub.ftz.f32 	%f38, %f25, %f26;
	mov.f32 	%f39, 0f00000000;
	max.ftz.f32 	%f15, %f39, %f38;
	setp.gtu.ftz.f32	%p5, %f13, %f23;
	@%p5 bra 	BB2_6;

	mov.f32 	%f77, %f39;
	bra.uni 	BB2_17;

BB2_6:
	setp.eq.ftz.f32	%p6, %f24, 0f00000000;
	setp.le.ftz.f32	%p7, %f13, %f14;
	and.pred  	%p8, %p7, %p6;
	@!%p8 bra 	BB2_8;
	bra.uni 	BB2_7;

BB2_7:
	mov.f32 	%f77, %f39;
	bra.uni 	BB2_17;

BB2_8:
	setp.ltu.ftz.f32	%p9, %f13, %f25;
	@%p9 bra 	BB2_10;

	mov.f32 	%f77, %f37;
	bra.uni 	BB2_17;

BB2_10:
	setp.eq.ftz.f32	%p10, %f26, 0f00000000;
	setp.ge.ftz.f32	%p11, %f13, %f15;
	and.pred  	%p12, %p11, %p10;
	@!%p12 bra 	BB2_12;
	bra.uni 	BB2_11;

BB2_11:
	mov.f32 	%f77, %f37;
	bra.uni 	BB2_17;

BB2_12:
	setp.gt.ftz.f32	%p13, %f13, %f14;
	setp.lt.ftz.f32	%p14, %f13, %f15;
	and.pred  	%p15, %p13, %p14;
	mov.f32 	%f77, 0f3F000000;
	@!%p15 bra 	BB2_13;
	bra.uni 	BB2_17;

BB2_13:
	setp.gt.ftz.f32	%p16, %f13, %f23;
	setp.lt.ftz.f32	%p17, %f13, %f14;
	and.pred  	%p18, %p16, %p17;
	@%p18 bra 	BB2_16;

	setp.gt.ftz.f32	%p19, %f13, %f15;
	@%p19 bra 	BB2_15;
	bra.uni 	BB2_17;

BB2_15:
	sub.ftz.f32 	%f43, %f25, %f15;
	sub.ftz.f32 	%f44, %f13, %f15;
	div.approx.ftz.f32 	%f45, %f44, %f43;
	fma.rn.ftz.f32 	%f77, %f45, 0f3F000000, 0f3F000000;
	bra.uni 	BB2_17;

BB2_16:
	sub.ftz.f32 	%f46, %f14, %f23;
	sub.ftz.f32 	%f47, %f13, %f23;
	div.approx.ftz.f32 	%f48, %f47, %f46;
	mul.ftz.f32 	%f77, %f48, 0f3F000000;

BB2_17:
	ld.const.f32 	%f52, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f53, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f54, %f53, 0f00000000;
	fma.rn.ftz.f32 	%f55, %f77, %f52, %f54;
	ld.const.f32 	%f56, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f20, %f56, 0f00000000, %f55;
	ld.const.f32 	%f57, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f58, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f59, %f58, 0f00000000;
	fma.rn.ftz.f32 	%f60, %f77, %f57, %f59;
	ld.const.f32 	%f61, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f21, %f61, 0f00000000, %f60;
	ld.const.f32 	%f62, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f63, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f64, %f63, 0f00000000;
	fma.rn.ftz.f32 	%f65, %f77, %f62, %f64;
	ld.const.f32 	%f66, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f22, %f66, 0f00000000, %f65;
	@%p4 bra 	BB2_19;

	st.global.v4.f32 	[%rd1], {%f22, %f21, %f20, %f73};
	bra.uni 	BB2_20;

BB2_19:
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f73;
	mov.b16 	%rs9, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f20;
	mov.b16 	%rs10, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f21;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f22;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd2], {%rs12, %rs11, %rs10, %rs9};

BB2_20:
	ret;
}

.visible .entry ThreeWayColorCorrector_LumaKernel(
	.param .u64 ThreeWayColorCorrector_LumaKernel_param_0,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_1,
	.param .u64 ThreeWayColorCorrector_LumaKernel_param_2,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_3,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_4,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_5,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_6,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_7,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_8,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_9,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_10,
	.param .f32 ThreeWayColorCorrector_LumaKernel_param_11,
	.param .u32 ThreeWayColorCorrector_LumaKernel_param_12
)
{
	.reg .pred 	%p<12>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<61>;
	.reg .f32 	%f<86>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd4, [ThreeWayColorCorrector_LumaKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_LumaKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_LumaKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_LumaKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_LumaKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_LumaKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_LumaKernel_param_6];
	ld.param.f32 	%f28, [ThreeWayColorCorrector_LumaKernel_param_7];
	ld.param.f32 	%f29, [ThreeWayColorCorrector_LumaKernel_param_8];
	ld.param.f32 	%f30, [ThreeWayColorCorrector_LumaKernel_param_9];
	ld.param.f32 	%f31, [ThreeWayColorCorrector_LumaKernel_param_10];
	ld.param.f32 	%f32, [ThreeWayColorCorrector_LumaKernel_param_11];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_LumaKernel_param_12];
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB3_15;
	bra.uni 	BB3_1;

BB3_1:
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd6, %rd4;
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd2, %rd6, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd3, %rd6, %rd8;
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB3_3;

	ld.global.v4.f32 	{%f33, %f34, %f35, %f36}, [%rd2];
	mov.f32 	%f82, %f36;
	mov.f32 	%f81, %f35;
	mov.f32 	%f80, %f34;
	mov.f32 	%f79, %f33;
	bra.uni 	BB3_4;

BB3_3:
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f79, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f80, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f81, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f82, %temp;
	}

BB3_4:
	ld.const.f32 	%f37, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f38, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f39, %f80, %f38;
	fma.rn.ftz.f32 	%f40, %f81, %f37, %f39;
	ld.const.f32 	%f41, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f13, %f79, %f41, %f40;
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB3_6;

	mov.f32 	%f83, 0f3F800000;
	bra.uni 	BB3_7;

BB3_6:
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	ld.global.f32 	%f83, [%rd10];

BB3_7:
	sub.ftz.f32 	%f16, %f29, %f28;
	setp.neu.ftz.f32	%p6, %f16, 0f3F800000;
	sub.ftz.f32 	%f17, %f32, %f31;
	setp.neu.ftz.f32	%p7, %f17, 0f3F800000;
	or.pred  	%p8, %p6, %p7;
	mov.f32 	%f84, %f13;
	@!%p8 bra 	BB3_9;
	bra.uni 	BB3_8;

BB3_8:
	sub.s32 	%r17, %r1, %r2;
	add.s32 	%r18, %r17, -1;
	mov.u32 	%r19, 1;
	sub.s32 	%r20, %r19, %r17;
	add.s32 	%r21, %r2, %r20;
	add.s32 	%r22, %r21, -1;
	shl.b32 	%r23, %r18, 8;
	xor.b32  	%r24, %r22, %r23;
	mov.u32 	%r25, 2;
	sub.s32 	%r26, %r25, %r17;
	sub.s32 	%r27, %r26, %r24;
	shr.u32 	%r28, %r24, 13;
	xor.b32  	%r29, %r27, %r28;
	sub.s32 	%r30, %r18, %r24;
	sub.s32 	%r31, %r30, %r29;
	shr.u32 	%r32, %r29, 12;
	xor.b32  	%r33, %r31, %r32;
	sub.s32 	%r34, %r24, %r29;
	sub.s32 	%r35, %r34, %r33;
	shl.b32 	%r36, %r33, 16;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r29, %r33;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 5;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r33, %r37;
	sub.s32 	%r43, %r42, %r41;
	shr.u32 	%r44, %r41, 3;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shl.b32 	%r48, %r45, 10;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 15;
	xor.b32  	%r53, %r51, %r52;
	mad.lo.s32 	%r54, %r53, 1103515245, 12345;
	mad.lo.s32 	%r55, %r54, 1103515245, 12345;
	shr.u32 	%r56, %r54, 9;
	and.b32  	%r57, %r56, 32640;
	bfe.u32 	%r58, %r55, 16, 8;
	xor.b32  	%r59, %r58, %r57;
	cvt.rn.f32.s32	%f43, %r59;
	ld.const.u32 	%r60, [kRandMax];
	cvt.rn.f32.u32	%f44, %r60;
	mov.f32 	%f45, 0f40000000;
	div.approx.ftz.f32 	%f46, %f44, %f45;
	div.approx.ftz.f32 	%f47, %f43, %f46;
	add.ftz.f32 	%f48, %f47, 0fBF800000;
	fma.rn.ftz.f32 	%f18, %f48, 0f3B270D73, %f13;
	mov.f32 	%f84, %f18;

BB3_9:
	mov.f32 	%f19, %f84;
	sub.ftz.f32 	%f49, %f19, %f28;
	div.approx.ftz.f32 	%f50, %f49, %f16;
	mov.f32 	%f51, 0f00000000;
	max.ftz.f32 	%f21, %f51, %f50;
	setp.ltu.ftz.f32	%p9, %f21, 0f00000000;
	@%p9 bra 	BB3_11;

	lg2.approx.ftz.f32 	%f52, %f21;
	mul.ftz.f32 	%f53, %f52, %f30;
	ex2.approx.ftz.f32 	%f85, %f53;
	bra.uni 	BB3_12;

BB3_11:
	neg.ftz.f32 	%f54, %f21;
	lg2.approx.ftz.f32 	%f55, %f54;
	mul.ftz.f32 	%f56, %f55, %f30;
	ex2.approx.ftz.f32 	%f57, %f56;
	neg.ftz.f32 	%f85, %f57;

BB3_12:
	fma.rn.ftz.f32 	%f58, %f17, %f85, %f31;
	sub.ftz.f32 	%f59, %f19, %f58;
	fma.rn.ftz.f32 	%f60, %f59, 0f3F000000, %f58;
	setp.eq.s32	%p10, %r6, 0;
	selp.f32	%f61, %f58, %f60, %p10;
	sub.ftz.f32 	%f62, %f61, %f13;
	fma.rn.ftz.f32 	%f63, %f83, %f62, %f13;
	ld.const.f32 	%f64, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f65, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f66, %f65, 0f00000000;
	fma.rn.ftz.f32 	%f67, %f63, %f64, %f66;
	ld.const.f32 	%f68, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f25, %f68, 0f00000000, %f67;
	ld.const.f32 	%f69, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f70, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f71, %f70, 0f00000000;
	fma.rn.ftz.f32 	%f72, %f63, %f69, %f71;
	ld.const.f32 	%f73, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f26, %f73, 0f00000000, %f72;
	ld.const.f32 	%f74, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f75, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f76, %f75, 0f00000000;
	fma.rn.ftz.f32 	%f77, %f63, %f74, %f76;
	ld.const.f32 	%f78, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f27, %f78, 0f00000000, %f77;
	@%p4 bra 	BB3_14;

	st.global.v4.f32 	[%rd2], {%f27, %f26, %f25, %f82};
	bra.uni 	BB3_15;

BB3_14:
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f82;
	mov.b16 	%rs9, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f25;
	mov.b16 	%rs10, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f26;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f27;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd3], {%rs12, %rs11, %rs10, %rs9};

BB3_15:
	ret;
}

.visible .entry ThreeWayColorCorrector_CompositeKernel(
	.param .u64 ThreeWayColorCorrector_CompositeKernel_param_0,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_1,
	.param .u64 ThreeWayColorCorrector_CompositeKernel_param_2,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_3,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_4,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_5,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_6,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_7,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_8,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_9,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_10,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_11,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_12,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_13,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_14,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_15,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_16,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_17,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_18,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_19,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_20,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_21,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_22,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_23,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_24,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_25,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_26,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_27,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_28,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_29,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_30,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_31,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_32,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_33,
	.param .f32 ThreeWayColorCorrector_CompositeKernel_param_34,
	.param .u32 ThreeWayColorCorrector_CompositeKernel_param_35
)
{
	.reg .pred 	%p<27>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<87>;
	.reg .f32 	%f<204>;
	.reg .s64 	%rd<17>;


	ld.param.u64 	%rd4, [ThreeWayColorCorrector_CompositeKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_CompositeKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_CompositeKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_CompositeKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_CompositeKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_CompositeKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_CompositeKernel_param_6];
	ld.param.f32 	%f46, [ThreeWayColorCorrector_CompositeKernel_param_7];
	ld.param.f32 	%f47, [ThreeWayColorCorrector_CompositeKernel_param_8];
	ld.param.f32 	%f48, [ThreeWayColorCorrector_CompositeKernel_param_9];
	ld.param.f32 	%f49, [ThreeWayColorCorrector_CompositeKernel_param_10];
	ld.param.f32 	%f50, [ThreeWayColorCorrector_CompositeKernel_param_11];
	ld.param.f32 	%f51, [ThreeWayColorCorrector_CompositeKernel_param_12];
	ld.param.f32 	%f52, [ThreeWayColorCorrector_CompositeKernel_param_13];
	ld.param.f32 	%f53, [ThreeWayColorCorrector_CompositeKernel_param_14];
	ld.param.f32 	%f54, [ThreeWayColorCorrector_CompositeKernel_param_15];
	ld.param.f32 	%f55, [ThreeWayColorCorrector_CompositeKernel_param_16];
	ld.param.f32 	%f56, [ThreeWayColorCorrector_CompositeKernel_param_17];
	ld.param.f32 	%f57, [ThreeWayColorCorrector_CompositeKernel_param_18];
	ld.param.f32 	%f58, [ThreeWayColorCorrector_CompositeKernel_param_19];
	ld.param.f32 	%f59, [ThreeWayColorCorrector_CompositeKernel_param_20];
	ld.param.f32 	%f60, [ThreeWayColorCorrector_CompositeKernel_param_21];
	ld.param.f32 	%f61, [ThreeWayColorCorrector_CompositeKernel_param_22];
	ld.param.f32 	%f62, [ThreeWayColorCorrector_CompositeKernel_param_23];
	ld.param.f32 	%f63, [ThreeWayColorCorrector_CompositeKernel_param_24];
	ld.param.f32 	%f64, [ThreeWayColorCorrector_CompositeKernel_param_25];
	ld.param.f32 	%f65, [ThreeWayColorCorrector_CompositeKernel_param_26];
	ld.param.f32 	%f66, [ThreeWayColorCorrector_CompositeKernel_param_27];
	ld.param.f32 	%f67, [ThreeWayColorCorrector_CompositeKernel_param_28];
	ld.param.f32 	%f68, [ThreeWayColorCorrector_CompositeKernel_param_29];
	ld.param.f32 	%f69, [ThreeWayColorCorrector_CompositeKernel_param_30];
	ld.param.f32 	%f70, [ThreeWayColorCorrector_CompositeKernel_param_31];
	ld.param.f32 	%f71, [ThreeWayColorCorrector_CompositeKernel_param_32];
	ld.param.f32 	%f72, [ThreeWayColorCorrector_CompositeKernel_param_33];
	ld.param.f32 	%f73, [ThreeWayColorCorrector_CompositeKernel_param_34];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_CompositeKernel_param_35];
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB4_24;
	bra.uni 	BB4_1;

BB4_1:
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd6, %rd4;
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd2, %rd6, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd3, %rd6, %rd8;
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB4_3;

	ld.global.v4.f32 	{%f74, %f75, %f76, %f77}, [%rd2];
	mov.f32 	%f192, %f77;
	mov.f32 	%f191, %f76;
	mov.f32 	%f190, %f75;
	mov.f32 	%f189, %f74;
	bra.uni 	BB4_4;

BB4_3:
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f189, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f190, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f191, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f192, %temp;
	}

BB4_4:
	ld.const.f32 	%f78, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f79, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f80, %f190, %f79;
	fma.rn.ftz.f32 	%f81, %f191, %f78, %f80;
	ld.const.f32 	%f82, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f13, %f189, %f82, %f81;
	ld.const.f32 	%f83, [kRGB32f_To_601YPbPr+12];
	ld.const.f32 	%f84, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f85, %f190, %f84;
	fma.rn.ftz.f32 	%f86, %f191, %f83, %f85;
	ld.const.f32 	%f87, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f14, %f189, %f87, %f86;
	ld.const.f32 	%f88, [kRGB32f_To_601YPbPr+24];
	ld.const.f32 	%f89, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f90, %f190, %f89;
	fma.rn.ftz.f32 	%f91, %f191, %f88, %f90;
	ld.const.f32 	%f92, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f15, %f189, %f92, %f91;
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB4_6;

	mov.f32 	%f193, 0f3F800000;
	bra.uni 	BB4_7;

BB4_6:
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	ld.global.f32 	%f193, [%rd10];

BB4_7:
	add.ftz.f32 	%f94, %f46, %f47;
	mov.f32 	%f95, 0f3F800000;
	min.ftz.f32 	%f18, %f95, %f94;
	sub.ftz.f32 	%f96, %f48, %f49;
	mov.f32 	%f97, 0f00000000;
	max.ftz.f32 	%f19, %f97, %f96;
	cos.approx.ftz.f32 	%f98, %f54;
	mul.ftz.f32 	%f99, %f14, %f98;
	sin.approx.ftz.f32 	%f100, %f54;
	mul.ftz.f32 	%f101, %f15, %f100;
	sub.ftz.f32 	%f102, %f99, %f101;
	mul.ftz.f32 	%f103, %f14, %f100;
	fma.rn.ftz.f32 	%f104, %f15, %f98, %f103;
	cos.approx.ftz.f32 	%f105, %f55;
	mul.ftz.f32 	%f106, %f14, %f105;
	sin.approx.ftz.f32 	%f107, %f55;
	mul.ftz.f32 	%f108, %f15, %f107;
	sub.ftz.f32 	%f109, %f106, %f108;
	mul.ftz.f32 	%f110, %f14, %f107;
	fma.rn.ftz.f32 	%f111, %f15, %f105, %f110;
	cos.approx.ftz.f32 	%f112, %f56;
	mul.ftz.f32 	%f113, %f14, %f112;
	sin.approx.ftz.f32 	%f114, %f56;
	mul.ftz.f32 	%f115, %f15, %f114;
	sub.ftz.f32 	%f116, %f113, %f115;
	mul.ftz.f32 	%f117, %f14, %f114;
	fma.rn.ftz.f32 	%f118, %f15, %f112, %f117;
	sub.ftz.f32 	%f119, %f60, %f102;
	fma.rn.ftz.f32 	%f120, %f119, %f62, %f102;
	sub.ftz.f32 	%f121, %f61, %f104;
	fma.rn.ftz.f32 	%f122, %f121, %f62, %f104;
	sub.ftz.f32 	%f123, %f63, %f109;
	fma.rn.ftz.f32 	%f124, %f123, %f65, %f109;
	sub.ftz.f32 	%f125, %f64, %f111;
	fma.rn.ftz.f32 	%f126, %f125, %f65, %f111;
	sub.ftz.f32 	%f127, %f66, %f116;
	fma.rn.ftz.f32 	%f128, %f127, %f68, %f116;
	sub.ftz.f32 	%f129, %f67, %f118;
	fma.rn.ftz.f32 	%f130, %f129, %f68, %f118;
	mul.ftz.f32 	%f20, %f120, %f51;
	mul.ftz.f32 	%f21, %f122, %f51;
	mul.ftz.f32 	%f22, %f124, %f52;
	mul.ftz.f32 	%f23, %f126, %f52;
	mul.ftz.f32 	%f24, %f128, %f53;
	mul.ftz.f32 	%f25, %f130, %f53;
	setp.le.ftz.f32	%p6, %f13, %f46;
	mov.f32 	%f197, %f21;
	mov.f32 	%f201, %f20;
	@%p6 bra 	BB4_16;

	setp.eq.ftz.f32	%p7, %f47, 0f00000000;
	setp.le.ftz.f32	%p8, %f13, %f18;
	and.pred  	%p9, %p8, %p7;
	mov.f32 	%f194, %f21;
	mov.f32 	%f197, %f194;
	mov.f32 	%f198, %f20;
	mov.f32 	%f201, %f198;
	@%p9 bra 	BB4_16;

	setp.ge.ftz.f32	%p10, %f13, %f48;
	mov.f32 	%f197, %f25;
	mov.f32 	%f201, %f24;
	@%p10 bra 	BB4_16;

	setp.eq.ftz.f32	%p11, %f49, 0f00000000;
	setp.ge.ftz.f32	%p12, %f13, %f19;
	and.pred  	%p13, %p12, %p11;
	mov.f32 	%f195, %f25;
	mov.f32 	%f197, %f195;
	mov.f32 	%f199, %f24;
	mov.f32 	%f201, %f199;
	@%p13 bra 	BB4_16;

	setp.gt.ftz.f32	%p14, %f13, %f18;
	setp.lt.ftz.f32	%p15, %f13, %f19;
	and.pred  	%p16, %p14, %p15;
	mov.f32 	%f197, %f23;
	mov.f32 	%f201, %f22;
	@%p16 bra 	BB4_16;

	setp.gt.ftz.f32	%p17, %f13, %f46;
	setp.lt.ftz.f32	%p18, %f13, %f18;
	and.pred  	%p19, %p17, %p18;
	@%p19 bra 	BB4_15;

	setp.leu.ftz.f32	%p20, %f13, %f19;
	mov.f32 	%f196, %f23;
	mov.f32 	%f197, %f196;
	mov.f32 	%f200, %f22;
	mov.f32 	%f201, %f200;
	@%p20 bra 	BB4_16;

	sub.ftz.f32 	%f131, %f48, %f19;
	sub.ftz.f32 	%f132, %f13, %f19;
	div.approx.ftz.f32 	%f133, %f132, %f131;
	sub.ftz.f32 	%f134, %f24, %f22;
	fma.rn.ftz.f32 	%f201, %f133, %f134, %f22;
	sub.ftz.f32 	%f135, %f25, %f23;
	fma.rn.ftz.f32 	%f197, %f133, %f135, %f23;
	bra.uni 	BB4_16;

BB4_15:
	sub.ftz.f32 	%f136, %f18, %f46;
	sub.ftz.f32 	%f137, %f13, %f46;
	div.approx.ftz.f32 	%f138, %f137, %f136;
	sub.ftz.f32 	%f139, %f22, %f20;
	fma.rn.ftz.f32 	%f201, %f138, %f139, %f20;
	sub.ftz.f32 	%f140, %f23, %f21;
	fma.rn.ftz.f32 	%f197, %f138, %f140, %f21;

BB4_16:
	sub.ftz.f32 	%f141, %f57, %f14;
	fma.rn.ftz.f32 	%f142, %f141, %f59, %f14;
	sub.ftz.f32 	%f143, %f58, %f15;
	fma.rn.ftz.f32 	%f144, %f143, %f59, %f15;
	mul.ftz.f32 	%f145, %f142, %f50;
	mul.ftz.f32 	%f146, %f144, %f50;
	sub.ftz.f32 	%f147, %f201, %f145;
	fma.rn.ftz.f32 	%f148, %f147, 0f3F000000, %f145;
	sub.ftz.f32 	%f149, %f197, %f146;
	fma.rn.ftz.f32 	%f150, %f149, 0f3F000000, %f146;
	sub.ftz.f32 	%f151, %f148, %f14;
	fma.rn.ftz.f32 	%f33, %f193, %f151, %f14;
	sub.ftz.f32 	%f152, %f150, %f15;
	fma.rn.ftz.f32 	%f34, %f193, %f152, %f15;
	sub.ftz.f32 	%f35, %f70, %f69;
	setp.neu.ftz.f32	%p21, %f35, 0f3F800000;
	sub.ftz.f32 	%f36, %f73, %f72;
	setp.neu.ftz.f32	%p22, %f36, 0f3F800000;
	or.pred  	%p23, %p21, %p22;
	mov.f32 	%f202, %f13;
	@!%p23 bra 	BB4_18;
	bra.uni 	BB4_17;

BB4_17:
	sub.s32 	%r25, %r1, %r2;
	add.s32 	%r26, %r25, -1;
	mov.u32 	%r27, 1;
	sub.s32 	%r28, %r27, %r25;
	add.s32 	%r29, %r2, %r28;
	add.s32 	%r30, %r29, -1;
	shl.b32 	%r31, %r26, 8;
	xor.b32  	%r32, %r30, %r31;
	mov.u32 	%r33, 2;
	sub.s32 	%r34, %r33, %r25;
	sub.s32 	%r35, %r34, %r32;
	shr.u32 	%r36, %r32, 13;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r26, %r32;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 12;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r32, %r37;
	sub.s32 	%r43, %r42, %r41;
	shl.b32 	%r44, %r41, 16;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shr.u32 	%r48, %r45, 5;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 3;
	xor.b32  	%r53, %r51, %r52;
	sub.s32 	%r54, %r45, %r49;
	sub.s32 	%r55, %r54, %r53;
	shl.b32 	%r56, %r53, 10;
	xor.b32  	%r57, %r55, %r56;
	sub.s32 	%r58, %r49, %r53;
	sub.s32 	%r59, %r58, %r57;
	shr.u32 	%r60, %r57, 15;
	xor.b32  	%r61, %r59, %r60;
	mad.lo.s32 	%r62, %r61, 1103515245, 12345;
	mad.lo.s32 	%r63, %r62, 1103515245, 12345;
	shr.u32 	%r64, %r62, 9;
	and.b32  	%r65, %r64, 32640;
	bfe.u32 	%r66, %r63, 16, 8;
	xor.b32  	%r67, %r66, %r65;
	cvt.rn.f32.s32	%f153, %r67;
	ld.const.u32 	%r68, [kRandMax];
	cvt.rn.f32.u32	%f154, %r68;
	mov.f32 	%f155, 0f40000000;
	div.approx.ftz.f32 	%f156, %f154, %f155;
	div.approx.ftz.f32 	%f157, %f153, %f156;
	add.ftz.f32 	%f158, %f157, 0fBF800000;
	fma.rn.ftz.f32 	%f37, %f158, 0f3B270D73, %f13;
	mov.f32 	%f202, %f37;

BB4_18:
	mov.f32 	%f38, %f202;
	sub.ftz.f32 	%f159, %f38, %f69;
	div.approx.ftz.f32 	%f160, %f159, %f35;
	max.ftz.f32 	%f39, %f97, %f160;
	setp.ltu.ftz.f32	%p24, %f39, 0f00000000;
	@%p24 bra 	BB4_20;

	lg2.approx.ftz.f32 	%f162, %f39;
	mul.ftz.f32 	%f163, %f162, %f71;
	ex2.approx.ftz.f32 	%f203, %f163;
	bra.uni 	BB4_21;

BB4_20:
	neg.ftz.f32 	%f164, %f39;
	lg2.approx.ftz.f32 	%f165, %f164;
	mul.ftz.f32 	%f166, %f165, %f71;
	ex2.approx.ftz.f32 	%f167, %f166;
	neg.ftz.f32 	%f203, %f167;

BB4_21:
	fma.rn.ftz.f32 	%f168, %f36, %f203, %f72;
	sub.ftz.f32 	%f169, %f38, %f168;
	fma.rn.ftz.f32 	%f170, %f169, 0f3F000000, %f168;
	setp.eq.s32	%p25, %r6, 0;
	selp.f32	%f171, %f168, %f170, %p25;
	sub.ftz.f32 	%f172, %f171, %f13;
	fma.rn.ftz.f32 	%f173, %f193, %f172, %f13;
	ld.const.f32 	%f174, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f175, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f176, %f33, %f175;
	fma.rn.ftz.f32 	%f177, %f173, %f174, %f176;
	ld.const.f32 	%f178, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f43, %f34, %f178, %f177;
	ld.const.f32 	%f179, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f180, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f181, %f33, %f180;
	fma.rn.ftz.f32 	%f182, %f173, %f179, %f181;
	ld.const.f32 	%f183, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f44, %f34, %f183, %f182;
	ld.const.f32 	%f184, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f185, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f186, %f33, %f185;
	fma.rn.ftz.f32 	%f187, %f173, %f184, %f186;
	ld.const.f32 	%f188, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f45, %f34, %f188, %f187;
	@%p4 bra 	BB4_23;

	mul.wide.s32 	%rd12, %r15, 16;
	add.s64 	%rd13, %rd6, %rd12;
	st.global.v4.f32 	[%rd13], {%f45, %f44, %f43, %f192};
	bra.uni 	BB4_24;

BB4_23:
	mul.wide.s32 	%rd15, %r15, 8;
	add.s64 	%rd16, %rd6, %rd15;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f192;
	mov.b16 	%rs9, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f43;
	mov.b16 	%rs10, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f44;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f45;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd16], {%rs12, %rs11, %rs10, %rs9};

BB4_24:
	ret;
}

.visible .entry ThreeWayColorCorrector_MasterLevelsKernel(
	.param .u64 ThreeWayColorCorrector_MasterLevelsKernel_param_0,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_1,
	.param .u64 ThreeWayColorCorrector_MasterLevelsKernel_param_2,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_3,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_4,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_5,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_6,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_7,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_8,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_9,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_10,
	.param .f32 ThreeWayColorCorrector_MasterLevelsKernel_param_11,
	.param .u32 ThreeWayColorCorrector_MasterLevelsKernel_param_12
)
{
	.reg .pred 	%p<12>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<61>;
	.reg .f32 	%f<99>;
	.reg .s64 	%rd<11>;


	ld.param.u64 	%rd4, [ThreeWayColorCorrector_MasterLevelsKernel_param_0];
	ld.param.u32 	%r3, [ThreeWayColorCorrector_MasterLevelsKernel_param_1];
	ld.param.u64 	%rd5, [ThreeWayColorCorrector_MasterLevelsKernel_param_2];
	ld.param.u32 	%r4, [ThreeWayColorCorrector_MasterLevelsKernel_param_3];
	ld.param.u32 	%r5, [ThreeWayColorCorrector_MasterLevelsKernel_param_4];
	ld.param.u32 	%r7, [ThreeWayColorCorrector_MasterLevelsKernel_param_5];
	ld.param.u32 	%r8, [ThreeWayColorCorrector_MasterLevelsKernel_param_6];
	ld.param.f32 	%f29, [ThreeWayColorCorrector_MasterLevelsKernel_param_7];
	ld.param.f32 	%f30, [ThreeWayColorCorrector_MasterLevelsKernel_param_8];
	ld.param.f32 	%f31, [ThreeWayColorCorrector_MasterLevelsKernel_param_9];
	ld.param.f32 	%f32, [ThreeWayColorCorrector_MasterLevelsKernel_param_10];
	ld.param.f32 	%f33, [ThreeWayColorCorrector_MasterLevelsKernel_param_11];
	ld.param.u32 	%r6, [ThreeWayColorCorrector_MasterLevelsKernel_param_12];
	mov.u32 	%r9, %ntid.x;
	mov.u32 	%r10, %ctaid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r9, %r10, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	setp.lt.s32	%p1, %r1, %r7;
	setp.lt.s32	%p2, %r2, %r8;
	and.pred  	%p3, %p1, %p2;
	@!%p3 bra 	BB5_15;
	bra.uni 	BB5_1;

BB5_1:
	cvta.to.global.u64 	%rd1, %rd5;
	cvta.to.global.u64 	%rd6, %rd4;
	mad.lo.s32 	%r15, %r2, %r3, %r1;
	mul.wide.s32 	%rd7, %r15, 16;
	add.s64 	%rd2, %rd6, %rd7;
	mul.wide.s32 	%rd8, %r15, 8;
	add.s64 	%rd3, %rd6, %rd8;
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB5_3;

	ld.global.v4.f32 	{%f34, %f35, %f36, %f37}, [%rd2];
	mov.f32 	%f95, %f37;
	mov.f32 	%f94, %f36;
	mov.f32 	%f93, %f35;
	mov.f32 	%f92, %f34;
	bra.uni 	BB5_4;

BB5_3:
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd3];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f92, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f93, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f94, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f95, %temp;
	}

BB5_4:
	ld.const.f32 	%f38, [kRGB32f_To_601YPbPr];
	ld.const.f32 	%f39, [kRGB32f_To_601YPbPr+4];
	mul.ftz.f32 	%f40, %f93, %f39;
	fma.rn.ftz.f32 	%f41, %f94, %f38, %f40;
	ld.const.f32 	%f42, [kRGB32f_To_601YPbPr+8];
	fma.rn.ftz.f32 	%f14, %f92, %f42, %f41;
	ld.const.f32 	%f43, [kRGB32f_To_601YPbPr+12];
	ld.const.f32 	%f44, [kRGB32f_To_601YPbPr+16];
	mul.ftz.f32 	%f45, %f93, %f44;
	fma.rn.ftz.f32 	%f46, %f94, %f43, %f45;
	ld.const.f32 	%f47, [kRGB32f_To_601YPbPr+20];
	fma.rn.ftz.f32 	%f15, %f92, %f47, %f46;
	ld.const.f32 	%f48, [kRGB32f_To_601YPbPr+24];
	ld.const.f32 	%f49, [kRGB32f_To_601YPbPr+28];
	mul.ftz.f32 	%f50, %f93, %f49;
	fma.rn.ftz.f32 	%f51, %f94, %f48, %f50;
	ld.const.f32 	%f52, [kRGB32f_To_601YPbPr+32];
	fma.rn.ftz.f32 	%f16, %f92, %f52, %f51;
	setp.ne.s64	%p5, %rd5, 0;
	@%p5 bra 	BB5_6;

	mov.f32 	%f96, 0f3F800000;
	bra.uni 	BB5_7;

BB5_6:
	mad.lo.s32 	%r16, %r2, %r4, %r1;
	mul.wide.s32 	%rd9, %r16, 4;
	add.s64 	%rd10, %rd1, %rd9;
	ld.global.f32 	%f96, [%rd10];

BB5_7:
	sub.ftz.f32 	%f54, %f30, %f29;
	setp.neu.ftz.f32	%p6, %f54, 0f3F800000;
	sub.ftz.f32 	%f19, %f33, %f32;
	setp.neu.ftz.f32	%p7, %f19, 0f3F800000;
	or.pred  	%p8, %p6, %p7;
	mov.f32 	%f97, %f14;
	@!%p8 bra 	BB5_9;
	bra.uni 	BB5_8;

BB5_8:
	sub.s32 	%r17, %r1, %r2;
	add.s32 	%r18, %r17, -1;
	mov.u32 	%r19, 1;
	sub.s32 	%r20, %r19, %r17;
	add.s32 	%r21, %r2, %r20;
	add.s32 	%r22, %r21, -1;
	shl.b32 	%r23, %r18, 8;
	xor.b32  	%r24, %r22, %r23;
	mov.u32 	%r25, 2;
	sub.s32 	%r26, %r25, %r17;
	sub.s32 	%r27, %r26, %r24;
	shr.u32 	%r28, %r24, 13;
	xor.b32  	%r29, %r27, %r28;
	sub.s32 	%r30, %r18, %r24;
	sub.s32 	%r31, %r30, %r29;
	shr.u32 	%r32, %r29, 12;
	xor.b32  	%r33, %r31, %r32;
	sub.s32 	%r34, %r24, %r29;
	sub.s32 	%r35, %r34, %r33;
	shl.b32 	%r36, %r33, 16;
	xor.b32  	%r37, %r35, %r36;
	sub.s32 	%r38, %r29, %r33;
	sub.s32 	%r39, %r38, %r37;
	shr.u32 	%r40, %r37, 5;
	xor.b32  	%r41, %r39, %r40;
	sub.s32 	%r42, %r33, %r37;
	sub.s32 	%r43, %r42, %r41;
	shr.u32 	%r44, %r41, 3;
	xor.b32  	%r45, %r43, %r44;
	sub.s32 	%r46, %r37, %r41;
	sub.s32 	%r47, %r46, %r45;
	shl.b32 	%r48, %r45, 10;
	xor.b32  	%r49, %r47, %r48;
	sub.s32 	%r50, %r41, %r45;
	sub.s32 	%r51, %r50, %r49;
	shr.u32 	%r52, %r49, 15;
	xor.b32  	%r53, %r51, %r52;
	mad.lo.s32 	%r54, %r53, 1103515245, 12345;
	mad.lo.s32 	%r55, %r54, 1103515245, 12345;
	shr.u32 	%r56, %r54, 9;
	and.b32  	%r57, %r56, 32640;
	bfe.u32 	%r58, %r55, 16, 8;
	xor.b32  	%r59, %r58, %r57;
	cvt.rn.f32.s32	%f55, %r59;
	ld.const.u32 	%r60, [kRandMax];
	cvt.rn.f32.u32	%f56, %r60;
	mov.f32 	%f57, 0f40000000;
	div.approx.ftz.f32 	%f58, %f56, %f57;
	div.approx.ftz.f32 	%f59, %f55, %f58;
	add.ftz.f32 	%f60, %f59, 0fBF800000;
	fma.rn.ftz.f32 	%f20, %f60, 0f3B270D73, %f14;
	mov.f32 	%f97, %f20;

BB5_9:
	mov.f32 	%f21, %f97;
	sub.ftz.f32 	%f62, %f21, %f29;
	div.approx.ftz.f32 	%f63, %f62, %f54;
	mov.f32 	%f64, 0f00000000;
	max.ftz.f32 	%f22, %f64, %f63;
	setp.ltu.ftz.f32	%p9, %f22, 0f00000000;
	@%p9 bra 	BB5_11;

	lg2.approx.ftz.f32 	%f65, %f22;
	mul.ftz.f32 	%f66, %f65, %f31;
	ex2.approx.ftz.f32 	%f98, %f66;
	bra.uni 	BB5_12;

BB5_11:
	neg.ftz.f32 	%f67, %f22;
	lg2.approx.ftz.f32 	%f68, %f67;
	mul.ftz.f32 	%f69, %f68, %f31;
	ex2.approx.ftz.f32 	%f70, %f69;
	neg.ftz.f32 	%f98, %f70;

BB5_12:
	fma.rn.ftz.f32 	%f71, %f19, %f98, %f32;
	sub.ftz.f32 	%f72, %f21, %f71;
	fma.rn.ftz.f32 	%f73, %f72, 0f3F000000, %f71;
	setp.eq.s32	%p10, %r6, 0;
	selp.f32	%f74, %f71, %f73, %p10;
	sub.ftz.f32 	%f75, %f74, %f14;
	fma.rn.ftz.f32 	%f76, %f96, %f75, %f14;
	ld.const.f32 	%f77, [k601YPbPr_To_RGB32f];
	ld.const.f32 	%f78, [k601YPbPr_To_RGB32f+4];
	mul.ftz.f32 	%f79, %f15, %f78;
	fma.rn.ftz.f32 	%f80, %f76, %f77, %f79;
	ld.const.f32 	%f81, [k601YPbPr_To_RGB32f+8];
	fma.rn.ftz.f32 	%f26, %f16, %f81, %f80;
	ld.const.f32 	%f82, [k601YPbPr_To_RGB32f+12];
	ld.const.f32 	%f83, [k601YPbPr_To_RGB32f+16];
	mul.ftz.f32 	%f84, %f15, %f83;
	fma.rn.ftz.f32 	%f85, %f76, %f82, %f84;
	ld.const.f32 	%f86, [k601YPbPr_To_RGB32f+20];
	fma.rn.ftz.f32 	%f27, %f16, %f86, %f85;
	ld.const.f32 	%f87, [k601YPbPr_To_RGB32f+24];
	ld.const.f32 	%f88, [k601YPbPr_To_RGB32f+28];
	mul.ftz.f32 	%f89, %f15, %f88;
	fma.rn.ftz.f32 	%f90, %f76, %f87, %f89;
	ld.const.f32 	%f91, [k601YPbPr_To_RGB32f+32];
	fma.rn.ftz.f32 	%f28, %f16, %f91, %f90;
	@%p4 bra 	BB5_14;

	st.global.v4.f32 	[%rd2], {%f28, %f27, %f26, %f95};
	bra.uni 	BB5_15;

BB5_14:
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f95;
	mov.b16 	%rs9, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f26;
	mov.b16 	%rs10, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f27;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f28;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd3], {%rs12, %rs11, %rs10, %rs9};

BB5_15:
	ret;
}


