//
// 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/MediaCore/GPUFoundation/Src/ImageProcessing/SingleChannelBlur.cu", 1399785311, 10576
.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, 147, 24, 206, 61, 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};
// HorizontalRecursiveGaussianGrayF32_kernel$__cuda_local_var_169757_333_non_const_smem has been demoted

.visible .entry VerticalRecursiveGaussianGRAYF32_kernel(
	.param .u64 VerticalRecursiveGaussianGRAYF32_kernel_param_0,
	.param .u64 VerticalRecursiveGaussianGRAYF32_kernel_param_1,
	.param .u32 VerticalRecursiveGaussianGRAYF32_kernel_param_2,
	.param .u32 VerticalRecursiveGaussianGRAYF32_kernel_param_3,
	.param .u32 VerticalRecursiveGaussianGRAYF32_kernel_param_4,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_5,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_6,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_7,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_8,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_9,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_10,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_11,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_12,
	.param .f32 VerticalRecursiveGaussianGRAYF32_kernel_param_13
)
{
	.reg .pred 	%p<6>;
	.reg .s32 	%r<29>;
	.reg .f32 	%f<54>;
	.reg .s64 	%rd<24>;


	ld.param.u64 	%rd9, [VerticalRecursiveGaussianGRAYF32_kernel_param_0];
	ld.param.u64 	%rd10, [VerticalRecursiveGaussianGRAYF32_kernel_param_1];
	ld.param.u32 	%r16, [VerticalRecursiveGaussianGRAYF32_kernel_param_2];
	ld.param.u32 	%r14, [VerticalRecursiveGaussianGRAYF32_kernel_param_3];
	ld.param.u32 	%r15, [VerticalRecursiveGaussianGRAYF32_kernel_param_4];
	ld.param.f32 	%f17, [VerticalRecursiveGaussianGRAYF32_kernel_param_5];
	ld.param.f32 	%f18, [VerticalRecursiveGaussianGRAYF32_kernel_param_6];
	ld.param.f32 	%f19, [VerticalRecursiveGaussianGRAYF32_kernel_param_7];
	ld.param.f32 	%f20, [VerticalRecursiveGaussianGRAYF32_kernel_param_8];
	ld.param.f32 	%f21, [VerticalRecursiveGaussianGRAYF32_kernel_param_9];
	ld.param.f32 	%f22, [VerticalRecursiveGaussianGRAYF32_kernel_param_10];
	ld.param.f32 	%f23, [VerticalRecursiveGaussianGRAYF32_kernel_param_11];
	ld.param.f32 	%f24, [VerticalRecursiveGaussianGRAYF32_kernel_param_12];
	ld.param.f32 	%f25, [VerticalRecursiveGaussianGRAYF32_kernel_param_13];
	.loc 1 25 1
	mov.u32 	%r17, %ctaid.x;
	mov.u32 	%r18, %ntid.x;
	mov.u32 	%r1, %tid.x;
	mad.lo.s32 	%r28, %r18, %r17, %r1;
	.loc 1 25 1
	setp.gt.s32	%p1, %r28, %r16;
	@%p1 bra 	BB0_8;

	cvta.to.global.u64 	%rd11, %rd10;
	mul.wide.s32 	%rd12, %r28, 4;
	add.s64 	%rd13, %rd11, %rd12;
	.loc 1 25 1
	ld.global.f32 	%f42, [%rd13];
	mul.ftz.f32 	%f41, %f42, %f25;
	.loc 1 25 1
	setp.gt.s32	%p2, %r14, 0;
	@%p2 bra 	BB0_3;

	mov.u32 	%r27, 0;
	.loc 1 25 1
	mov.f32 	%f40, %f41;
	bra.uni 	BB0_5;

BB0_3:
	.loc 1 25 1
	mul.lo.s32 	%r23, %r18, %r17;
	add.s32 	%r24, %r1, %r23;
	cvta.to.global.u64 	%rd14, %rd9;
	mul.wide.s32 	%rd15, %r24, 4;
	add.s64 	%rd23, %rd14, %rd15;
	.loc 1 25 1
	add.s32 	%r25, %r1, %r15;
	add.s32 	%r28, %r25, %r23;
	mul.wide.s32 	%rd17, %r28, 4;
	add.s64 	%rd22, %rd11, %rd17;
	mov.u32 	%r27, 1;
	mov.f32 	%f44, %f41;
	mov.f32 	%f43, %f41;

BB0_4:
	.loc 1 25 1
	mov.f32 	%f4, %f43;
	mov.f32 	%f34, %f42;
	mov.f32 	%f6, %f41;
	mov.f32 	%f43, %f44;
	mov.f32 	%f40, %f34;
	.loc 1 25 1
	mul.ftz.f32 	%f26, %f6, %f18;
	fma.rn.ftz.f32 	%f27, %f40, %f17, %f26;
	fma.rn.ftz.f32 	%f28, %f43, %f19, %f27;
	fma.rn.ftz.f32 	%f44, %f4, %f20, %f28;
	st.global.f32 	[%rd23], %f44;
	.loc 1 25 1
	setp.lt.s32	%p3, %r27, %r14;
	@%p3 bra 	BB0_9;

BB0_5:
	.loc 1 25 1
	mul.ftz.f32 	%f51, %f40, %f25;
	cvta.to.global.u64 	%rd6, %rd9;
	.loc 1 25 1
	setp.lt.s32	%p4, %r27, 1;
	@%p4 bra 	BB0_8;

	mov.f32 	%f50, %f51;
	mov.f32 	%f53, %f51;
	mov.f32 	%f52, %f51;

BB0_7:
	.loc 1 25 1
	mov.f32 	%f10, %f52;
	mov.f32 	%f12, %f50;
	mov.f32 	%f50, %f51;
	mov.f32 	%f52, %f53;
	mul.ftz.f32 	%f29, %f12, %f22;
	fma.rn.ftz.f32 	%f30, %f50, %f21, %f29;
	fma.rn.ftz.f32 	%f31, %f52, %f23, %f30;
	fma.rn.ftz.f32 	%f53, %f10, %f24, %f31;
	sub.s32 	%r28, %r28, %r15;
	mul.wide.s32 	%rd18, %r28, 4;
	add.s64 	%rd19, %rd11, %rd18;
	.loc 1 25 1
	ld.global.f32 	%f51, [%rd19];
	add.s64 	%rd20, %rd6, %rd18;
	.loc 1 25 1
	ld.global.f32 	%f32, [%rd20];
	add.ftz.f32 	%f33, %f32, %f53;
	st.global.f32 	[%rd20], %f33;
	.loc 1 25 55
	add.s32 	%r27, %r27, -1;
	.loc 1 25 1
	setp.gt.s32	%p5, %r27, 0;
	@%p5 bra 	BB0_7;

BB0_8:
	.loc 1 25 2
	ret;

BB0_9:
	.loc 1 25 1
	ld.global.f32 	%f42, [%rd22];
	.loc 1 25 1
	add.s32 	%r27, %r27, 1;
	shl.b32 	%r26, %r15, 2;
	cvt.s64.s32	%rd21, %r26;
	add.s64 	%rd23, %rd23, %rd21;
	.loc 1 25 1
	add.s32 	%r28, %r28, %r15;
	add.s64 	%rd22, %rd22, %rd21;
	.loc 1 25 1
	mov.f32 	%f41, %f40;
	bra.uni 	BB0_4;
}

.visible .entry HorizontalRecursiveGaussianGrayF32_kernel(
	.param .u64 HorizontalRecursiveGaussianGrayF32_kernel_param_0,
	.param .u64 HorizontalRecursiveGaussianGrayF32_kernel_param_1,
	.param .u32 HorizontalRecursiveGaussianGrayF32_kernel_param_2,
	.param .u32 HorizontalRecursiveGaussianGrayF32_kernel_param_3,
	.param .u32 HorizontalRecursiveGaussianGrayF32_kernel_param_4,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_5,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_6,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_7,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_8,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_9,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_10,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_11,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_12,
	.param .f32 HorizontalRecursiveGaussianGrayF32_kernel_param_13
)
{
	.reg .pred 	%p<59>;
	.reg .s32 	%r<185>;
	.reg .f32 	%f<202>;
	.reg .s64 	%rd<72>;
	// demoted variable
	.shared .align 4 .b8 HorizontalRecursiveGaussianGrayF32_kernel$__cuda_local_var_169757_333_non_const_smem[4224];

	ld.param.u64 	%rd28, [HorizontalRecursiveGaussianGrayF32_kernel_param_0];
	ld.param.u64 	%rd29, [HorizontalRecursiveGaussianGrayF32_kernel_param_1];
	ld.param.u32 	%r33, [HorizontalRecursiveGaussianGrayF32_kernel_param_2];
	ld.param.u32 	%r34, [HorizontalRecursiveGaussianGrayF32_kernel_param_3];
	ld.param.u32 	%r35, [HorizontalRecursiveGaussianGrayF32_kernel_param_4];
	ld.param.f32 	%f33, [HorizontalRecursiveGaussianGrayF32_kernel_param_5];
	ld.param.f32 	%f34, [HorizontalRecursiveGaussianGrayF32_kernel_param_6];
	ld.param.f32 	%f35, [HorizontalRecursiveGaussianGrayF32_kernel_param_7];
	ld.param.f32 	%f36, [HorizontalRecursiveGaussianGrayF32_kernel_param_8];
	ld.param.f32 	%f37, [HorizontalRecursiveGaussianGrayF32_kernel_param_9];
	ld.param.f32 	%f38, [HorizontalRecursiveGaussianGrayF32_kernel_param_10];
	ld.param.f32 	%f39, [HorizontalRecursiveGaussianGrayF32_kernel_param_11];
	ld.param.f32 	%f40, [HorizontalRecursiveGaussianGrayF32_kernel_param_12];
	ld.param.f32 	%f41, [HorizontalRecursiveGaussianGrayF32_kernel_param_13];
	cvta.to.global.u64 	%rd1, %rd29;
	.loc 1 119 1
	mov.u32 	%r36, %ctaid.y;
	shl.b32 	%r37, %r36, 5;
	mov.u32 	%r1, %tid.y;
	add.s32 	%r38, %r37, %r1;
	mul.lo.s32 	%r2, %r38, %r35;
	setp.eq.s32	%p7, %r1, 0;
	setp.gt.ftz.f32	%p8, %f41, 0f00000000;
	and.pred  	%p9, %p7, %p8;
	.loc 1 119 1
	@%p9 bra 	BB1_2;

	mov.f32 	%f168, 0f00000000;
	bra.uni 	BB1_3;

BB1_2:
	mul.wide.s32 	%rd30, %r2, 4;
	add.s64 	%rd31, %rd1, %rd30;
	.loc 1 119 1
	ld.global.f32 	%f43, [%rd31];
	mul.ftz.f32 	%f168, %f43, %f41;

BB1_3:
	.loc 1 119 1
	mov.f32 	%f165, %f168;
	setp.gt.s32	%p10, %r33, 0;
	@%p10 bra 	BB1_5;

	mov.u32 	%r183, 0;
	bra.uni 	BB1_34;

BB1_5:
	.loc 1 119 1
	mov.u32 	%r42, %tid.x;
	mad.lo.s32 	%r43, %r1, 33, %r42;
	mul.wide.s32 	%rd32, %r43, 4;
	mov.u64 	%rd33, HorizontalRecursiveGaussianGrayF32_kernel$__cuda_local_var_169757_333_non_const_smem;
	add.s64 	%rd2, %rd33, %rd32;
	mov.u32 	%r183, 0;
	mov.u32 	%r180, %r183;
	mov.f32 	%f166, %f165;
	mov.f32 	%f167, %f165;

BB1_6:
	.loc 1 119 1
	mov.f32 	%f164, %f167;
	mov.f32 	%f162, %f166;
	mov.f32 	%f160, %f165;
	add.s32 	%r47, %r1, %r37;
	mad.lo.s32 	%r5, %r47, %r35, %r42;
	.loc 1 119 1
	shl.b32 	%r49, %r180, 5;
	add.s32 	%r6, %r5, %r49;
	mul.wide.s32 	%rd35, %r6, 4;
	add.s64 	%rd3, %rd1, %rd35;
	.loc 1 119 1
	setp.ge.s32	%p11, %r47, %r34;
	@%p11 bra 	BB1_10;

	.loc 1 119 1
	add.s32 	%r51, %r42, %r183;
	setp.lt.s32	%p12, %r51, %r33;
	.loc 1 119 1
	@%p12 bra 	BB1_9;

	mov.u32 	%r52, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd2], %r52;
	bra.uni 	BB1_10;

BB1_9:
	.loc 1 119 1
	ld.global.f32 	%f44, [%rd3];
	st.shared.f32 	[%rd2], %f44;

BB1_10:
	.loc 1 119 1
	add.s32 	%r57, %r47, 8;
	shl.b32 	%r58, %r35, 3;
	add.s32 	%r7, %r6, %r58;
	mul.wide.s32 	%rd37, %r7, 4;
	add.s64 	%rd4, %rd1, %rd37;
	.loc 1 119 1
	setp.ge.s32	%p13, %r57, %r34;
	@%p13 bra 	BB1_14;

	.loc 1 119 1
	add.s32 	%r60, %r42, %r183;
	setp.lt.s32	%p14, %r60, %r33;
	.loc 1 119 1
	@%p14 bra 	BB1_13;

	mov.u32 	%r61, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd2+1056], %r61;
	bra.uni 	BB1_14;

BB1_13:
	.loc 1 119 1
	ld.global.f32 	%f45, [%rd4];
	st.shared.f32 	[%rd2+1056], %f45;

BB1_14:
	.loc 1 119 1
	add.s32 	%r66, %r47, 16;
	shl.b32 	%r69, %r35, 4;
	add.s32 	%r70, %r6, %r69;
	mul.wide.s32 	%rd39, %r70, 4;
	add.s64 	%rd5, %rd1, %rd39;
	.loc 1 119 1
	setp.ge.s32	%p15, %r66, %r34;
	@%p15 bra 	BB1_18;

	.loc 1 119 1
	add.s32 	%r72, %r42, %r183;
	setp.lt.s32	%p16, %r72, %r33;
	.loc 1 119 1
	@%p16 bra 	BB1_17;

	mov.u32 	%r73, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd2+2112], %r73;
	bra.uni 	BB1_18;

BB1_17:
	.loc 1 119 1
	ld.global.f32 	%f46, [%rd5];
	st.shared.f32 	[%rd2+2112], %f46;

BB1_18:
	.loc 1 119 1
	add.s32 	%r78, %r47, 24;
	mad.lo.s32 	%r8, %r35, 24, %r6;
	mul.wide.s32 	%rd41, %r8, 4;
	add.s64 	%rd6, %rd1, %rd41;
	.loc 1 119 1
	setp.ge.s32	%p17, %r78, %r34;
	@%p17 bra 	BB1_22;

	.loc 1 119 1
	add.s32 	%r80, %r42, %r183;
	setp.lt.s32	%p18, %r80, %r33;
	.loc 1 119 1
	@%p18 bra 	BB1_21;

	mov.u32 	%r81, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd2+3168], %r81;
	bra.uni 	BB1_22;

BB1_21:
	.loc 1 119 1
	ld.global.f32 	%f47, [%rd6];
	st.shared.f32 	[%rd2+3168], %f47;

BB1_22:
	mad.lo.s32 	%r82, %r42, 33, 4;
	mul.wide.s32 	%rd42, %r82, 4;
	add.s64 	%rd70, %rd33, %rd42;
	.loc 1 119 1
	bar.sync 	0;
	.loc 1 119 1
	@!%p7 bra 	BB1_25;
	bra.uni 	BB1_23;

BB1_23:
	mov.u32 	%r181, 0;
	mov.f32 	%f159, %f160;
	mov.f32 	%f161, %f162;
	mov.f32 	%f163, %f164;

BB1_24:
	.loc 1 119 1
	ld.shared.f32 	%f48, [%rd70+-16];
	mul.ftz.f32 	%f49, %f159, %f34;
	fma.rn.ftz.f32 	%f50, %f48, %f33, %f49;
	fma.rn.ftz.f32 	%f51, %f161, %f35, %f50;
	fma.rn.ftz.f32 	%f52, %f163, %f36, %f51;
	ld.shared.f32 	%f53, [%rd70+-12];
	ld.shared.f32 	%f54, [%rd70+-8];
	ld.shared.f32 	%f55, [%rd70+-4];
	st.shared.f32 	[%rd70+-16], %f52;
	mul.ftz.f32 	%f56, %f48, %f34;
	fma.rn.ftz.f32 	%f57, %f53, %f33, %f56;
	fma.rn.ftz.f32 	%f58, %f52, %f35, %f57;
	fma.rn.ftz.f32 	%f59, %f161, %f36, %f58;
	st.shared.f32 	[%rd70+-12], %f59;
	mul.ftz.f32 	%f60, %f53, %f34;
	fma.rn.ftz.f32 	%f61, %f54, %f33, %f60;
	fma.rn.ftz.f32 	%f62, %f59, %f35, %f61;
	fma.rn.ftz.f32 	%f63, %f52, %f36, %f62;
	st.shared.f32 	[%rd70+-8], %f63;
	mul.ftz.f32 	%f64, %f54, %f34;
	fma.rn.ftz.f32 	%f65, %f55, %f33, %f64;
	fma.rn.ftz.f32 	%f66, %f63, %f35, %f65;
	fma.rn.ftz.f32 	%f67, %f59, %f36, %f66;
	st.shared.f32 	[%rd70+-4], %f67;
	ld.shared.f32 	%f68, [%rd70];
	mul.ftz.f32 	%f69, %f55, %f34;
	fma.rn.ftz.f32 	%f70, %f68, %f33, %f69;
	fma.rn.ftz.f32 	%f71, %f67, %f35, %f70;
	fma.rn.ftz.f32 	%f72, %f63, %f36, %f71;
	ld.shared.f32 	%f73, [%rd70+4];
	ld.shared.f32 	%f74, [%rd70+8];
	ld.shared.f32 	%f160, [%rd70+12];
	st.shared.f32 	[%rd70], %f72;
	mul.ftz.f32 	%f75, %f68, %f34;
	fma.rn.ftz.f32 	%f76, %f73, %f33, %f75;
	fma.rn.ftz.f32 	%f77, %f72, %f35, %f76;
	fma.rn.ftz.f32 	%f78, %f67, %f36, %f77;
	st.shared.f32 	[%rd70+4], %f78;
	mul.ftz.f32 	%f79, %f73, %f34;
	fma.rn.ftz.f32 	%f80, %f74, %f33, %f79;
	fma.rn.ftz.f32 	%f81, %f78, %f35, %f80;
	fma.rn.ftz.f32 	%f164, %f72, %f36, %f81;
	st.shared.f32 	[%rd70+8], %f164;
	mul.ftz.f32 	%f82, %f74, %f34;
	fma.rn.ftz.f32 	%f83, %f160, %f33, %f82;
	fma.rn.ftz.f32 	%f84, %f164, %f35, %f83;
	fma.rn.ftz.f32 	%f162, %f78, %f36, %f84;
	st.shared.f32 	[%rd70+12], %f162;
	add.s64 	%rd70, %rd70, 32;
	.loc 1 119 18
	add.s32 	%r181, %r181, 8;
	.loc 1 119 1
	setp.ne.s32	%p19, %r181, 32;
	mov.f32 	%f163, %f164;
	mov.f32 	%f161, %f162;
	mov.f32 	%f159, %f160;
	@%p19 bra 	BB1_24;

BB1_25:
	.loc 1 119 1
	mov.f32 	%f167, %f164;
	mov.f32 	%f166, %f162;
	mov.f32 	%f165, %f160;
	.loc 1 119 1
	add.s32 	%r84, %r42, %r183;
	setp.lt.s32	%p2, %r84, %r33;
	setp.lt.s32	%p3, %r47, %r34;
	.loc 1 119 1
	bar.sync 	0;
	cvta.to.global.u64 	%rd44, %rd28;
	mul.wide.s32 	%rd45, %r6, 4;
	add.s64 	%rd10, %rd44, %rd45;
	.loc 1 119 1
	not.pred 	%p20, %p3;
	not.pred 	%p21, %p2;
	or.pred  	%p22, %p20, %p21;
	@%p22 bra 	BB1_27;

	.loc 1 119 1
	ld.shared.f32 	%f85, [%rd2];
	st.global.f32 	[%rd10], %f85;

BB1_27:
	.loc 1 119 1
	shl.b32 	%r170, %r35, 3;
	add.s32 	%r169, %r6, %r170;
	mul.wide.s32 	%rd46, %r169, 4;
	add.s64 	%rd12, %rd44, %rd46;
	.loc 1 119 1
	setp.ge.s32	%p24, %r84, %r33;
	or.pred  	%p25, %p13, %p24;
	@%p25 bra 	BB1_29;

	.loc 1 119 1
	ld.shared.f32 	%f86, [%rd2+1056];
	st.global.f32 	[%rd12], %f86;

BB1_29:
	.loc 1 119 1
	shl.b32 	%r172, %r35, 4;
	add.s32 	%r171, %r6, %r172;
	mul.wide.s32 	%rd47, %r171, 4;
	add.s64 	%rd13, %rd44, %rd47;
	.loc 1 119 1
	or.pred  	%p28, %p15, %p24;
	@%p28 bra 	BB1_31;

	.loc 1 119 1
	ld.shared.f32 	%f87, [%rd2+2112];
	st.global.f32 	[%rd13], %f87;

BB1_31:
	.loc 1 119 1
	mad.lo.s32 	%r173, %r35, 24, %r6;
	mul.wide.s32 	%rd48, %r173, 4;
	add.s64 	%rd14, %rd44, %rd48;
	.loc 1 119 1
	or.pred  	%p31, %p17, %p24;
	@%p31 bra 	BB1_33;

	.loc 1 119 1
	ld.shared.f32 	%f88, [%rd2+3168];
	st.global.f32 	[%rd14], %f88;

BB1_33:
	.loc 1 119 1
	bar.sync 	0;
	.loc 1 119 1
	add.s32 	%r183, %r183, 32;
	.loc 1 119 1
	setp.lt.s32	%p32, %r183, %r33;
	add.s32 	%r180, %r180, 1;
	@%p32 bra 	BB1_6;

BB1_34:
	.loc 1 119 1
	@%p9 bra 	BB1_36;

	mov.f32 	%f201, 0f00000000;
	bra.uni 	BB1_37;

BB1_36:
	.loc 1 119 1
	add.s32 	%r174, %r37, %r1;
	ld.param.f32 	%f143, [HorizontalRecursiveGaussianGrayF32_kernel_param_13];
	.loc 1 119 1
	mad.lo.s32 	%r100, %r174, %r35, %r33;
	add.s32 	%r101, %r100, -1;
	mul.wide.s32 	%rd50, %r101, 4;
	add.s64 	%rd51, %rd1, %rd50;
	.loc 1 119 1
	ld.global.f32 	%f90, [%rd51];
	mul.ftz.f32 	%f201, %f90, %f143;

BB1_37:
	.loc 1 119 1
	mov.f32 	%f197, %f201;
	setp.lt.s32	%p36, %r183, 1;
	@%p36 bra 	BB1_67;

	.loc 1 119 1
	mov.u32 	%r103, %tid.x;
	mad.lo.s32 	%r105, %r1, 33, %r103;
	.loc 1 119 1
	add.s32 	%r108, %r1, %r37;
	add.s32 	%r109, %r103, %r183;
	mad.lo.s32 	%r19, %r108, %r35, %r109;
	mul.wide.s32 	%rd52, %r105, 4;
	mov.u64 	%rd53, HorizontalRecursiveGaussianGrayF32_kernel$__cuda_local_var_169757_333_non_const_smem;
	add.s64 	%rd15, %rd53, %rd52;
	mov.u32 	%r182, 0;
	mov.f32 	%f198, %f197;
	mov.f32 	%f199, %f197;
	mov.f32 	%f200, %f197;

BB1_39:
	.loc 1 119 1
	mov.f32 	%f196, %f200;
	mov.f32 	%f194, %f199;
	mov.f32 	%f192, %f198;
	mov.f32 	%f190, %f197;
	mad.lo.s32 	%r22, %r182, -32, %r19;
	mul.wide.s32 	%rd55, %r22, 4;
	add.s64 	%rd16, %rd1, %rd55;
	.loc 1 119 1
	add.s32 	%r183, %r183, -32;
	.loc 1 119 1
	setp.ge.s32	%p37, %r108, %r34;
	@%p37 bra 	BB1_43;

	.loc 1 119 1
	add.s32 	%r115, %r103, %r183;
	setp.lt.s32	%p38, %r115, %r33;
	.loc 1 119 1
	@%p38 bra 	BB1_42;

	mov.u32 	%r116, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd15], %r116;
	bra.uni 	BB1_43;

BB1_42:
	.loc 1 119 1
	ld.global.f32 	%f91, [%rd16+-128];
	st.shared.f32 	[%rd15], %f91;

BB1_43:
	.loc 1 119 1
	add.s32 	%r121, %r108, 8;
	shl.b32 	%r122, %r35, 3;
	add.s32 	%r123, %r22, %r122;
	.loc 1 119 1
	add.s32 	%r25, %r123, -32;
	mul.wide.s32 	%rd57, %r123, 4;
	add.s64 	%rd17, %rd1, %rd57;
	.loc 1 119 1
	setp.ge.s32	%p39, %r121, %r34;
	@%p39 bra 	BB1_47;

	.loc 1 119 1
	add.s32 	%r125, %r103, %r183;
	setp.lt.s32	%p40, %r125, %r33;
	.loc 1 119 1
	@%p40 bra 	BB1_46;

	mov.u32 	%r126, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd15+1056], %r126;
	bra.uni 	BB1_47;

BB1_46:
	.loc 1 119 1
	ld.global.f32 	%f92, [%rd17+-128];
	st.shared.f32 	[%rd15+1056], %f92;

BB1_47:
	.loc 1 119 1
	add.s32 	%r131, %r108, 16;
	shl.b32 	%r132, %r35, 4;
	add.s32 	%r133, %r22, %r132;
	.loc 1 119 1
	add.s32 	%r26, %r133, -32;
	mul.wide.s32 	%rd59, %r133, 4;
	add.s64 	%rd18, %rd1, %rd59;
	.loc 1 119 1
	setp.ge.s32	%p41, %r131, %r34;
	@%p41 bra 	BB1_51;

	.loc 1 119 1
	add.s32 	%r135, %r103, %r183;
	setp.lt.s32	%p42, %r135, %r33;
	.loc 1 119 1
	@%p42 bra 	BB1_50;

	mov.u32 	%r136, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd15+2112], %r136;
	bra.uni 	BB1_51;

BB1_50:
	.loc 1 119 1
	ld.global.f32 	%f93, [%rd18+-128];
	st.shared.f32 	[%rd15+2112], %f93;

BB1_51:
	.loc 1 119 1
	add.s32 	%r141, %r108, 24;
	mad.lo.s32 	%r142, %r35, 24, %r22;
	mul.wide.s32 	%rd61, %r142, 4;
	add.s64 	%rd19, %rd1, %rd61;
	.loc 1 119 1
	setp.ge.s32	%p43, %r141, %r34;
	@%p43 bra 	BB1_55;

	.loc 1 119 1
	add.s32 	%r144, %r103, %r183;
	setp.lt.s32	%p44, %r144, %r33;
	.loc 1 119 1
	@%p44 bra 	BB1_54;

	mov.u32 	%r145, 0;
	.loc 1 119 1
	st.shared.u32 	[%rd15+3168], %r145;
	bra.uni 	BB1_55;

BB1_54:
	.loc 1 119 1
	ld.global.f32 	%f94, [%rd19+-128];
	st.shared.f32 	[%rd15+3168], %f94;

BB1_55:
	mad.lo.s32 	%r147, %r103, 33, 31;
	mul.wide.s32 	%rd62, %r147, 4;
	add.s64 	%rd71, %rd53, %rd62;
	.loc 1 119 1
	bar.sync 	0;
	.loc 1 119 1
	@!%p7 bra 	BB1_58;
	bra.uni 	BB1_56;

BB1_56:
	mov.u32 	%r184, 0;
	mov.f32 	%f189, %f190;
	mov.f32 	%f191, %f192;
	mov.f32 	%f193, %f194;
	mov.f32 	%f195, %f196;

BB1_57:
	.loc 1 119 1
	mul.ftz.f32 	%f95, %f191, %f38;
	fma.rn.ftz.f32 	%f96, %f189, %f37, %f95;
	fma.rn.ftz.f32 	%f97, %f193, %f39, %f96;
	fma.rn.ftz.f32 	%f98, %f195, %f40, %f97;
	ld.shared.f32 	%f99, [%rd71];
	ld.shared.f32 	%f100, [%rd71+-4];
	ld.shared.f32 	%f101, [%rd71+-8];
	ld.shared.f32 	%f102, [%rd71+-12];
	st.shared.f32 	[%rd71], %f98;
	mul.ftz.f32 	%f103, %f189, %f38;
	fma.rn.ftz.f32 	%f104, %f99, %f37, %f103;
	fma.rn.ftz.f32 	%f105, %f98, %f39, %f104;
	fma.rn.ftz.f32 	%f106, %f193, %f40, %f105;
	st.shared.f32 	[%rd71+-4], %f106;
	mul.ftz.f32 	%f107, %f99, %f38;
	fma.rn.ftz.f32 	%f108, %f100, %f37, %f107;
	fma.rn.ftz.f32 	%f109, %f106, %f39, %f108;
	fma.rn.ftz.f32 	%f110, %f98, %f40, %f109;
	st.shared.f32 	[%rd71+-8], %f110;
	mul.ftz.f32 	%f111, %f100, %f38;
	fma.rn.ftz.f32 	%f112, %f101, %f37, %f111;
	fma.rn.ftz.f32 	%f113, %f110, %f39, %f112;
	fma.rn.ftz.f32 	%f114, %f106, %f40, %f113;
	st.shared.f32 	[%rd71+-12], %f114;
	mul.ftz.f32 	%f115, %f101, %f38;
	fma.rn.ftz.f32 	%f116, %f102, %f37, %f115;
	fma.rn.ftz.f32 	%f117, %f114, %f39, %f116;
	fma.rn.ftz.f32 	%f118, %f110, %f40, %f117;
	ld.shared.f32 	%f119, [%rd71+-16];
	ld.shared.f32 	%f120, [%rd71+-20];
	ld.shared.f32 	%f192, [%rd71+-24];
	ld.shared.f32 	%f190, [%rd71+-28];
	st.shared.f32 	[%rd71+-16], %f118;
	mul.ftz.f32 	%f121, %f102, %f38;
	fma.rn.ftz.f32 	%f122, %f119, %f37, %f121;
	fma.rn.ftz.f32 	%f123, %f118, %f39, %f122;
	fma.rn.ftz.f32 	%f124, %f114, %f40, %f123;
	st.shared.f32 	[%rd71+-20], %f124;
	mul.ftz.f32 	%f125, %f119, %f38;
	fma.rn.ftz.f32 	%f126, %f120, %f37, %f125;
	fma.rn.ftz.f32 	%f127, %f124, %f39, %f126;
	fma.rn.ftz.f32 	%f196, %f118, %f40, %f127;
	st.shared.f32 	[%rd71+-24], %f196;
	mul.ftz.f32 	%f128, %f120, %f38;
	fma.rn.ftz.f32 	%f129, %f192, %f37, %f128;
	fma.rn.ftz.f32 	%f130, %f196, %f39, %f129;
	fma.rn.ftz.f32 	%f194, %f124, %f40, %f130;
	st.shared.f32 	[%rd71+-28], %f194;
	add.s64 	%rd71, %rd71, -32;
	.loc 1 119 1
	add.s32 	%r184, %r184, 8;
	setp.ne.s32	%p45, %r184, 32;
	mov.f32 	%f195, %f196;
	mov.f32 	%f193, %f194;
	mov.f32 	%f191, %f192;
	mov.f32 	%f189, %f190;
	@%p45 bra 	BB1_57;

BB1_58:
	.loc 1 119 1
	mov.f32 	%f200, %f196;
	mov.f32 	%f199, %f194;
	mov.f32 	%f198, %f192;
	mov.f32 	%f197, %f190;
	.loc 1 119 1
	add.s32 	%r151, %r103, %r183;
	setp.lt.s32	%p5, %r151, %r33;
	setp.lt.s32	%p6, %r108, %r34;
	.loc 1 119 1
	bar.sync 	0;
	mad.lo.s32 	%r176, %r182, -32, %r19;
	.loc 1 119 1
	add.s32 	%r175, %r176, -32;
	cvta.to.global.u64 	%rd64, %rd28;
	mul.wide.s32 	%rd65, %r175, 4;
	add.s64 	%rd23, %rd64, %rd65;
	.loc 1 119 1
	not.pred 	%p46, %p6;
	not.pred 	%p47, %p5;
	or.pred  	%p48, %p46, %p47;
	@%p48 bra 	BB1_60;

	.loc 1 119 1
	ld.shared.f32 	%f131, [%rd15];
	ld.global.f32 	%f132, [%rd23];
	add.ftz.f32 	%f133, %f132, %f131;
	st.global.f32 	[%rd23], %f133;

BB1_60:
	mul.wide.s32 	%rd67, %r25, 4;
	add.s64 	%rd24, %rd64, %rd67;
	.loc 1 119 1
	setp.ge.s32	%p50, %r151, %r33;
	or.pred  	%p51, %p39, %p50;
	@%p51 bra 	BB1_62;

	.loc 1 119 1
	ld.global.f32 	%f134, [%rd24];
	ld.shared.f32 	%f135, [%rd15+1056];
	add.ftz.f32 	%f136, %f134, %f135;
	st.global.f32 	[%rd24], %f136;

BB1_62:
	mul.wide.s32 	%rd68, %r26, 4;
	add.s64 	%rd26, %rd64, %rd68;
	.loc 1 119 1
	or.pred  	%p54, %p41, %p50;
	@%p54 bra 	BB1_64;

	.loc 1 119 1
	ld.global.f32 	%f137, [%rd26];
	ld.shared.f32 	%f138, [%rd15+2112];
	add.ftz.f32 	%f139, %f137, %f138;
	st.global.f32 	[%rd26], %f139;

BB1_64:
	mad.lo.s32 	%r179, %r182, -32, %r19;
	mad.lo.s32 	%r178, %r35, 24, %r179;
	.loc 1 119 1
	add.s32 	%r177, %r178, -32;
	mul.wide.s32 	%rd69, %r177, 4;
	add.s64 	%rd27, %rd64, %rd69;
	.loc 1 119 1
	or.pred  	%p57, %p43, %p50;
	@%p57 bra 	BB1_66;

	.loc 1 119 1
	ld.global.f32 	%f140, [%rd27];
	ld.shared.f32 	%f141, [%rd15+3168];
	add.ftz.f32 	%f142, %f140, %f141;
	st.global.f32 	[%rd27], %f142;

BB1_66:
	.loc 1 119 1
	bar.sync 	0;
	.loc 1 119 1
	add.s32 	%r182, %r182, 1;
	setp.gt.s32	%p58, %r183, 0;
	@%p58 bra 	BB1_39;

BB1_67:
	.loc 1 119 2
	ret;
}


