A         h      ̗hVwDec  3 202420:16:52HOST64sm_86//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-25769353
// Cuda compilation tools, release 10.1, V10.1.105
// Based on LLVM 3.4svn
//

.version 6.4
.target sm_30
.address_size 64

	// .globl	SharpenKernel
.global .texref inSrcTexture;

.visible .entry SharpenKernel(
	.param .u64 SharpenKernel_param_0,
	.param .u64 SharpenKernel_param_1,
	.param .u32 SharpenKernel_param_2,
	.param .u32 SharpenKernel_param_3,
	.param .u32 SharpenKernel_param_4,
	.param .u32 SharpenKernel_param_5,
	.param .f32 SharpenKernel_param_6,
	.param .f32 SharpenKernel_param_7,
	.param .u32 SharpenKernel_param_8
)
{
	.reg .pred 	%p<10>;
	.reg .b16 	%rs<5>;
	.reg .f32 	%f<152>;
	.reg .b32 	%r<29>;
	.reg .b64 	%rd<9>;


	ld.param.u64 	%rd2, [SharpenKernel_param_1];
	ld.param.u32 	%r4, [SharpenKernel_param_2];
	ld.param.u32 	%r5, [SharpenKernel_param_3];
	ld.param.u32 	%r6, [SharpenKernel_param_4];
	ld.param.u32 	%r7, [SharpenKernel_param_5];
	ld.param.f32 	%f73, [SharpenKernel_param_6];
	ld.param.f32 	%f74, [SharpenKernel_param_7];
	ld.param.u32 	%r8, [SharpenKernel_param_8];
	mov.u32 	%r9, %ctaid.x;
	mov.u32 	%r10, %ntid.x;
	mov.u32 	%r11, %tid.x;
	mad.lo.s32 	%r1, %r10, %r9, %r11;
	mov.u32 	%r12, %ntid.y;
	mov.u32 	%r13, %ctaid.y;
	mov.u32 	%r14, %tid.y;
	mad.lo.s32 	%r2, %r12, %r13, %r14;
	setp.ge.s32	%p1, %r1, %r6;
	setp.ge.s32	%p2, %r2, %r7;
	or.pred  	%p3, %p1, %p2;
	@%p3 bra 	BB0_15;

	cvt.rn.f32.s32	%f1, %r1;
	add.ftz.f32 	%f2, %f1, 0f3F000000;
	cvt.rn.f32.s32	%f3, %r2;
	add.ftz.f32 	%f4, %f3, 0f3F000000;
	tex.2d.v4.f32.f32	{%f5, %f6, %f7, %f8}, [inSrcTexture, {%f2, %f4}];
	setp.eq.s32	%p4, %r8, 0;
	@%p4 bra 	BB0_3;

	add.ftz.f32 	%f75, %f3, 0fBF000000;
	tex.2d.v4.f32.f32	{%f148, %f149, %f150, %f151}, [inSrcTexture, {%f2, %f75}];
	add.ftz.f32 	%f76, %f1, 0fBF000000;
	tex.2d.v4.f32.f32	{%f144, %f145, %f146, %f147}, [inSrcTexture, {%f76, %f4}];
	add.ftz.f32 	%f77, %f1, 0f3FC00000;
	tex.2d.v4.f32.f32	{%f140, %f141, %f142, %f143}, [inSrcTexture, {%f77, %f4}];
	add.ftz.f32 	%f78, %f3, 0f3FC00000;
	tex.2d.v4.f32.f32	{%f136, %f137, %f138, %f139}, [inSrcTexture, {%f2, %f78}];
	bra.uni 	BB0_12;

BB0_3:
	mov.f32 	%f136, 0f00000000;
	setp.lt.s32	%p5, %r2, 1;
	mov.f32 	%f148, %f136;
	mov.f32 	%f149, %f136;
	mov.f32 	%f150, %f136;
	mov.f32 	%f151, %f136;
	@%p5 bra 	BB0_5;

	add.ftz.f32 	%f83, %f3, 0fBF000000;
	tex.2d.v4.f32.f32	{%f148, %f149, %f150, %f151}, [inSrcTexture, {%f2, %f83}];

BB0_5:
	setp.lt.s32	%p6, %r1, 1;
	mov.f32 	%f144, %f136;
	mov.f32 	%f145, %f136;
	mov.f32 	%f146, %f136;
	mov.f32 	%f147, %f136;
	@%p6 bra 	BB0_7;

	add.ftz.f32 	%f88, %f1, 0fBF000000;
	tex.2d.v4.f32.f32	{%f144, %f145, %f146, %f147}, [inSrcTexture, {%f88, %f4}];

BB0_7:
	add.s32 	%r15, %r6, -1;
	setp.ge.s32	%p7, %r1, %r15;
	mov.f32 	%f140, %f136;
	mov.f32 	%f141, %f136;
	mov.f32 	%f142, %f136;
	mov.f32 	%f143, %f136;
	@%p7 bra 	BB0_9;

	add.ftz.f32 	%f93, %f1, 0f3FC00000;
	tex.2d.v4.f32.f32	{%f140, %f141, %f142, %f143}, [inSrcTexture, {%f93, %f4}];

BB0_9:
	add.s32 	%r20, %r7, -1;
	setp.ge.s32	%p8, %r2, %r20;
	@%p8 bra 	BB0_10;

	add.ftz.f32 	%f99, %f3, 0f3FC00000;
	tex.2d.v4.f32.f32	{%f136, %f137, %f138, %f139}, [inSrcTexture, {%f2, %f99}];
	bra.uni 	BB0_12;

BB0_10:
	mov.f32 	%f137, %f136;
	mov.f32 	%f138, %f136;
	mov.f32 	%f139, %f136;

BB0_12:
	add.ftz.f32 	%f100, %f144, %f148;
	add.ftz.f32 	%f101, %f140, %f100;
	add.ftz.f32 	%f102, %f145, %f149;
	add.ftz.f32 	%f103, %f141, %f102;
	add.ftz.f32 	%f104, %f146, %f150;
	add.ftz.f32 	%f105, %f142, %f104;
	add.ftz.f32 	%f106, %f147, %f151;
	add.ftz.f32 	%f107, %f143, %f106;
	add.ftz.f32 	%f108, %f136, %f101;
	add.ftz.f32 	%f109, %f137, %f103;
	add.ftz.f32 	%f110, %f138, %f105;
	add.ftz.f32 	%f111, %f139, %f107;
	mul.ftz.f32 	%f112, %f108, %f74;
	mul.ftz.f32 	%f113, %f109, %f74;
	mul.ftz.f32 	%f114, %f110, %f74;
	mul.ftz.f32 	%f115, %f111, %f74;
	fma.rn.ftz.f32 	%f69, %f5, %f73, %f112;
	fma.rn.ftz.f32 	%f70, %f6, %f73, %f113;
	fma.rn.ftz.f32 	%f71, %f7, %f73, %f114;
	fma.rn.ftz.f32 	%f116, %f8, %f73, %f115;
	mov.f32 	%f117, 0f00000000;
	max.ftz.f32 	%f118, %f116, %f117;
	mov.f32 	%f119, 0f3F800000;
	min.ftz.f32 	%f72, %f118, %f119;
	mad.lo.s32 	%r3, %r2, %r4, %r1;
	setp.eq.s32	%p9, %r5, 0;
	@%p9 bra 	BB0_14;

	cvta.to.global.u64 	%rd3, %rd2;
	mul.wide.s32 	%rd4, %r3, 16;
	add.s64 	%rd5, %rd3, %rd4;
	st.global.v4.f32 	[%rd5], {%f69, %f70, %f71, %f72};
	bra.uni 	BB0_15;

BB0_14:
	cvta.to.global.u64 	%rd6, %rd2;
	// inline asm
	{  cvt.rn.f16.f32 %rs1, %f69;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs2, %f70;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs3, %f71;}

	// inline asm
	// inline asm
	{  cvt.rn.f16.f32 %rs4, %f72;}

	// inline asm
	mul.wide.s32 	%rd7, %r3, 8;
	add.s64 	%rd8, %rd6, %rd7;
	st.global.u16 	[%rd8], %rs1;
	st.global.u16 	[%rd8+2], %rs2;
	st.global.u16 	[%rd8+4], %rs3;
	st.global.u16 	[%rd8+6], %rs4;

BB0_15:
	ret;
}


  ELF3                                V @ 8  @    .shstrtab .strtab .symtab .symtab_shndx .nv.info .text.SharpenKernel .nv.info.SharpenKernel .nv.shared.SharpenKernel .nv.constant0.SharpenKernel .rel.nv.constant0.SharpenKernel .debug_frame .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action  .shstrtab .strtab .symtab .symtab_shndx .nv.info .text.SharpenKernel .nv.info.SharpenKernel .nv.shared.SharpenKernel .rel.nv.constant0.SharpenKernel .nv.constant0.SharpenKernel .debug_frame .rel.debug_frame .rela.debug_frame .nv.callgraph .nv.prototype .nv.rel.action SharpenKernel inSrcTexture                            2                                                                                                                                                $        |( ((   4                                 $   ( T         /    "                   7    5  
    `, ,       (         $                                                         !           !        @                    s          % 6            D                                                                                                                                                                                                                                                                                                                                                                                                                                             z  
       y      &   ( y      "   ( y      %   h y       !   b $z     z  _  pb  $z      / z  ^  pfp   M	          z  b  pR  Er         Er         x         !t
   ?     !t   ?     r 
       !        !       r        !  ?     r         !  ?     r 
        `{
c 0 B `c 0 B `c 0 B `c 0 D `c 0 B $z  \    G	        z  ^   
    x    pb x?  x    pb x?  r        x         r 
       x         |    pb  x         |    pb  x         r 
        x         x         !
        r         !  ?    !        !  ?     `c 0 B `c 0 B `
c 0 B `c 0 B 2         8         !r
       !r        z  ]  pR  !r        !r        !r

        !r        !r        !r        !r

        !r        !r        !r         z

 a    A    z a    A             z a    A    z a    A   #z `  
    #z `      #z
 `      #z		 `      %  Z    !r        z  F   
    	      M	          x         >r	        >r

        %v  Z    x 2v        x
2v       y      y      y 
    y     My          Gy    y            y            y            y            y            y            y            y            y            y            y            y            y            y            y                                                                                                  @                                                          M      (                                                   x                                                       8      p                              )      p                      $                             F      p@                                                      p                                                        p                                                      	   @                                                   	   @                                                v      B                                                 2                                       "                                                                                                                                                                