//
// 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 inTexture;

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z12QuadCoverage6float2S_S_S_Pv(
	.param .align 8 .b8 _Z12QuadCoverage6float2S_S_S_Pv_param_0[8],
	.param .align 8 .b8 _Z12QuadCoverage6float2S_S_S_Pv_param_1[8],
	.param .align 8 .b8 _Z12QuadCoverage6float2S_S_S_Pv_param_2[8],
	.param .align 8 .b8 _Z12QuadCoverage6float2S_S_S_Pv_param_3[8],
	.param .b64 _Z12QuadCoverage6float2S_S_S_Pv_param_4
)
{
	.local .align 8 .b8 	__local_depot0[256];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<16>;
	.reg .s32 	%r<36>;
	.reg .f32 	%f<263>;
	.reg .s64 	%rd<40>;


	mov.u64 	%SPL, __local_depot0;
	ld.param.f32 	%f253, [_Z12QuadCoverage6float2S_S_S_Pv_param_2+4];
	ld.param.f32 	%f252, [_Z12QuadCoverage6float2S_S_S_Pv_param_2];
	add.u64 	%rd6, %SPL, 16;
	add.u64 	%rd7, %SPL, 240;
	add.u64 	%rd8, %SPL, 0;
	mov.u32 	%r34, 0;
	st.local.u32 	[%rd7], %r34;
	mov.u32 	%r16, 1;
	st.local.u32 	[%rd8], %r16;
	ld.param.f32 	%f257, [_Z12QuadCoverage6float2S_S_S_Pv_param_0+4];
	ld.param.f32 	%f256, [_Z12QuadCoverage6float2S_S_S_Pv_param_0];
	st.local.v2.f32 	[%rd6], {%f256, %f257};
	ld.param.f32 	%f255, [_Z12QuadCoverage6float2S_S_S_Pv_param_1+4];
	ld.param.f32 	%f254, [_Z12QuadCoverage6float2S_S_S_Pv_param_1];
	st.local.v2.f32 	[%rd6+8], {%f254, %f255};
	st.local.f32 	[%rd6+16], %f252;
	st.local.f32 	[%rd6+20], %f253;
	ld.param.f32 	%f251, [_Z12QuadCoverage6float2S_S_S_Pv_param_3+4];
	ld.param.f32 	%f250, [_Z12QuadCoverage6float2S_S_S_Pv_param_3];
	st.local.v2.f32 	[%rd6+24], {%f250, %f251};
	mov.f32 	%f262, 0f00000000;
	mov.u32 	%r33, %r34;
	mov.f32 	%f261, %f262;
	mov.f32 	%f260, %f262;
	mov.f32 	%f259, %f262;
	mov.f32 	%f258, %f262;

BB0_1:
	add.ftz.f32 	%f16, %f256, %f254;
	mul.ftz.f32 	%f17, %f16, 0f3F000000;
	add.ftz.f32 	%f20, %f257, %f255;
	mul.ftz.f32 	%f21, %f20, 0f3F000000;
	add.ftz.f32 	%f90, %f254, %f252;
	mul.ftz.f32 	%f23, %f90, 0f3F000000;
	add.ftz.f32 	%f91, %f255, %f253;
	mul.ftz.f32 	%f25, %f91, 0f3F000000;
	add.ftz.f32 	%f92, %f252, %f250;
	mul.ftz.f32 	%f27, %f92, 0f3F000000;
	add.ftz.f32 	%f93, %f253, %f251;
	mul.ftz.f32 	%f29, %f93, 0f3F000000;
	add.ftz.f32 	%f94, %f250, %f256;
	mul.ftz.f32 	%f30, %f94, 0f3F000000;
	add.ftz.f32 	%f95, %f251, %f257;
	mul.ftz.f32 	%f31, %f95, 0f3F000000;
	add.ftz.f32 	%f96, %f17, %f27;
	mul.ftz.f32 	%f32, %f96, 0f3F000000;
	add.ftz.f32 	%f97, %f21, %f29;
	mul.ftz.f32 	%f33, %f97, 0f3F000000;
	sub.ftz.f32 	%f98, %f30, %f23;
	sub.ftz.f32 	%f99, %f31, %f25;
	mul.ftz.f32 	%f100, %f99, %f99;
	fma.rn.ftz.f32 	%f34, %f98, %f98, %f100;
	setp.gt.ftz.f32	%p1, %f34, 0f3F800000;
	selp.u32	%r17, 1, 0, %p1;
	sub.ftz.f32 	%f101, %f27, %f17;
	sub.ftz.f32 	%f102, %f29, %f21;
	mul.ftz.f32 	%f103, %f102, %f102;
	fma.rn.ftz.f32 	%f35, %f101, %f101, %f103;
	setp.gt.ftz.f32	%p2, %f35, 0f3F800000;
	or.b32  	%r18, %r17, 2;
	selp.b32	%r3, %r18, %r17, %p2;
	setp.eq.s32	%p3, %r3, 0;
	cvt.s64.s32	%rd4, %r34;
	@%p3 bra 	BB0_16;

	shl.b64 	%rd9, %rd4, 2;
	add.s64 	%rd10, %rd7, %rd9;
	ld.local.u32 	%r4, [%rd10];
	setp.eq.s32	%p4, %r4, 2;
	@%p4 bra 	BB0_10;

	setp.lt.ftz.f32	%p5, %f34, 0f40800000;
	setp.lt.ftz.f32	%p6, %f35, 0f40800000;
	and.pred  	%p7, %p5, %p6;
	@%p7 bra 	BB0_10;

	setp.eq.s32	%p8, %r3, 3;
	mul.wide.s32 	%rd11, %r33, 32;
	add.s64 	%rd5, %rd6, %rd11;
	@%p8 bra 	BB0_8;

	setp.eq.s32	%p9, %r3, 2;
	st.local.v2.f32 	[%rd5], {%f256, %f257};
	@%p9 bra 	BB0_7;

	st.local.v2.f32 	[%rd5+8], {%f17, %f21};
	st.local.v2.f32 	[%rd5+16], {%f27, %f29};
	st.local.v2.f32 	[%rd5+24], {%f250, %f251};
	st.local.v2.f32 	[%rd5+32], {%f17, %f21};
	st.local.f32 	[%rd5+40], %f254;
	st.local.f32 	[%rd5+44], %f255;
	st.local.v2.f32 	[%rd5+48], {%f252, %f253};
	st.local.f32 	[%rd5+56], %f27;
	st.local.f32 	[%rd5+60], %f29;
	mov.u32 	%r35, 2;
	bra.uni 	BB0_9;

BB0_7:
	st.local.v2.f32 	[%rd5+8], {%f254, %f255};
	st.local.v2.f32 	[%rd5+16], {%f23, %f25};
	st.local.f32 	[%rd5+24], %f30;
	st.local.f32 	[%rd5+28], %f31;
	st.local.v2.f32 	[%rd5+32], {%f30, %f31};
	st.local.f32 	[%rd5+40], %f23;
	st.local.f32 	[%rd5+44], %f25;
	st.local.v2.f32 	[%rd5+48], {%f252, %f253};
	st.local.f32 	[%rd5+56], %f250;
	st.local.f32 	[%rd5+60], %f251;
	mov.u32 	%r35, 2;
	bra.uni 	BB0_9;

BB0_8:
	st.local.v2.f32 	[%rd5], {%f32, %f33};
	st.local.v2.f32 	[%rd5+8], {%f30, %f31};
	st.local.v2.f32 	[%rd5+16], {%f256, %f257};
	st.local.v2.f32 	[%rd5+24], {%f17, %f21};
	st.local.f32 	[%rd5+32], %f32;
	st.local.f32 	[%rd5+36], %f33;
	st.local.v2.f32 	[%rd5+40], {%f17, %f21};
	st.local.f32 	[%rd5+48], %f254;
	st.local.f32 	[%rd5+52], %f255;
	st.local.v2.f32 	[%rd5+56], {%f23, %f25};
	st.local.f32 	[%rd5+64], %f32;
	st.local.f32 	[%rd5+68], %f33;
	st.local.v2.f32 	[%rd5+72], {%f23, %f25};
	st.local.f32 	[%rd5+80], %f252;
	st.local.f32 	[%rd5+84], %f253;
	st.local.v2.f32 	[%rd5+88], {%f27, %f29};
	st.local.f32 	[%rd5+96], %f32;
	st.local.f32 	[%rd5+100], %f33;
	st.local.v2.f32 	[%rd5+104], {%f27, %f29};
	st.local.f32 	[%rd5+112], %f250;
	st.local.f32 	[%rd5+116], %f251;
	st.local.v2.f32 	[%rd5+120], {%f30, %f31};
	mov.u32 	%r35, 4;

BB0_9:
	add.s32 	%r22, %r33, %r35;
	add.s32 	%r33, %r22, -1;
	mul.wide.s32 	%rd12, %r34, 4;
	add.s64 	%rd13, %rd8, %rd12;
	ld.local.u32 	%r23, [%rd13];
	add.s32 	%r24, %r23, -1;
	st.local.u32 	[%rd13], %r24;
	setp.ne.s32	%p10, %r24, 0;
	selp.u32	%r25, 1, 0, %p10;
	add.s32 	%r34, %r25, %r34;
	mul.wide.s32 	%rd14, %r34, 4;
	add.s64 	%rd15, %rd7, %rd14;
	add.s32 	%r26, %r4, 1;
	st.local.u32 	[%rd15], %r26;
	add.s64 	%rd16, %rd8, %rd14;
	st.local.u32 	[%rd16], %r35;
	bra.uni 	BB0_17;

BB0_10:
	add.s32 	%r33, %r33, -1;
	setp.eq.s32	%p11, %r3, 3;
	@%p11 bra 	BB0_14;

	setp.eq.s32	%p12, %r3, 2;
	@%p12 bra 	BB0_13;

	add.ftz.f32 	%f116, %f256, %f17;
	add.ftz.f32 	%f117, %f116, %f27;
	add.ftz.f32 	%f118, %f117, %f250;
	mul.ftz.f32 	%f108, %f118, 0f3E800000;
	add.ftz.f32 	%f119, %f257, %f21;
	add.ftz.f32 	%f120, %f119, %f29;
	add.ftz.f32 	%f121, %f120, %f251;
	mul.ftz.f32 	%f109, %f121, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f104, %f105, %f106, %f107}, [inTexture, {%f108, %f109}];
	// inline asm
	add.ftz.f32 	%f122, %f259, %f104;
	add.ftz.f32 	%f123, %f258, %f105;
	add.ftz.f32 	%f124, %f260, %f106;
	add.ftz.f32 	%f125, %f261, %f107;
	add.ftz.f32 	%f126, %f17, %f254;
	add.ftz.f32 	%f127, %f126, %f252;
	add.ftz.f32 	%f128, %f127, %f27;
	mul.ftz.f32 	%f114, %f128, 0f3E800000;
	add.ftz.f32 	%f129, %f21, %f255;
	add.ftz.f32 	%f130, %f129, %f253;
	add.ftz.f32 	%f131, %f130, %f29;
	mul.ftz.f32 	%f115, %f131, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f110, %f111, %f112, %f113}, [inTexture, {%f114, %f115}];
	// inline asm
	add.ftz.f32 	%f259, %f122, %f110;
	add.ftz.f32 	%f258, %f123, %f111;
	add.ftz.f32 	%f260, %f124, %f112;
	add.ftz.f32 	%f261, %f125, %f113;
	add.ftz.f32 	%f262, %f262, 0f40000000;
	bra.uni 	BB0_15;

BB0_13:
	add.ftz.f32 	%f249, %f257, %f255;
	add.ftz.f32 	%f248, %f256, %f254;
	add.ftz.f32 	%f144, %f248, %f23;
	add.ftz.f32 	%f145, %f144, %f30;
	mul.ftz.f32 	%f136, %f145, 0f3E800000;
	add.ftz.f32 	%f146, %f249, %f25;
	add.ftz.f32 	%f147, %f146, %f31;
	mul.ftz.f32 	%f137, %f147, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f132, %f133, %f134, %f135}, [inTexture, {%f136, %f137}];
	// inline asm
	add.ftz.f32 	%f148, %f259, %f132;
	add.ftz.f32 	%f149, %f258, %f133;
	add.ftz.f32 	%f150, %f260, %f134;
	add.ftz.f32 	%f151, %f261, %f135;
	add.ftz.f32 	%f152, %f30, %f23;
	add.ftz.f32 	%f153, %f152, %f252;
	add.ftz.f32 	%f154, %f153, %f250;
	mul.ftz.f32 	%f142, %f154, 0f3E800000;
	add.ftz.f32 	%f155, %f31, %f25;
	add.ftz.f32 	%f156, %f155, %f253;
	add.ftz.f32 	%f157, %f156, %f251;
	mul.ftz.f32 	%f143, %f157, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f138, %f139, %f140, %f141}, [inTexture, {%f142, %f143}];
	// inline asm
	add.ftz.f32 	%f259, %f148, %f138;
	add.ftz.f32 	%f258, %f149, %f139;
	add.ftz.f32 	%f260, %f150, %f140;
	add.ftz.f32 	%f261, %f151, %f141;
	add.ftz.f32 	%f262, %f262, 0f40000000;
	bra.uni 	BB0_15;

BB0_14:
	add.ftz.f32 	%f262, %f262, 0f40800000;
	add.ftz.f32 	%f182, %f32, %f30;
	add.ftz.f32 	%f183, %f182, %f256;
	add.ftz.f32 	%f184, %f183, %f17;
	mul.ftz.f32 	%f162, %f184, 0f3E800000;
	add.ftz.f32 	%f185, %f33, %f31;
	add.ftz.f32 	%f186, %f185, %f257;
	add.ftz.f32 	%f187, %f186, %f21;
	mul.ftz.f32 	%f163, %f187, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f158, %f159, %f160, %f161}, [inTexture, {%f162, %f163}];
	// inline asm
	add.ftz.f32 	%f188, %f259, %f158;
	add.ftz.f32 	%f189, %f258, %f159;
	add.ftz.f32 	%f190, %f260, %f160;
	add.ftz.f32 	%f191, %f261, %f161;
	add.ftz.f32 	%f192, %f32, %f17;
	add.ftz.f32 	%f193, %f192, %f254;
	add.ftz.f32 	%f194, %f193, %f23;
	mul.ftz.f32 	%f168, %f194, 0f3E800000;
	add.ftz.f32 	%f195, %f33, %f21;
	add.ftz.f32 	%f196, %f195, %f255;
	add.ftz.f32 	%f197, %f196, %f25;
	mul.ftz.f32 	%f169, %f197, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f164, %f165, %f166, %f167}, [inTexture, {%f168, %f169}];
	// inline asm
	add.ftz.f32 	%f198, %f188, %f164;
	add.ftz.f32 	%f199, %f189, %f165;
	add.ftz.f32 	%f200, %f190, %f166;
	add.ftz.f32 	%f201, %f191, %f167;
	add.ftz.f32 	%f202, %f32, %f23;
	add.ftz.f32 	%f203, %f202, %f252;
	add.ftz.f32 	%f204, %f203, %f27;
	mul.ftz.f32 	%f174, %f204, 0f3E800000;
	add.ftz.f32 	%f205, %f33, %f25;
	add.ftz.f32 	%f206, %f205, %f253;
	add.ftz.f32 	%f207, %f206, %f29;
	mul.ftz.f32 	%f175, %f207, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f170, %f171, %f172, %f173}, [inTexture, {%f174, %f175}];
	// inline asm
	add.ftz.f32 	%f208, %f198, %f170;
	add.ftz.f32 	%f209, %f199, %f171;
	add.ftz.f32 	%f210, %f200, %f172;
	add.ftz.f32 	%f211, %f201, %f173;
	add.ftz.f32 	%f212, %f32, %f27;
	add.ftz.f32 	%f213, %f212, %f250;
	add.ftz.f32 	%f214, %f213, %f30;
	mul.ftz.f32 	%f180, %f214, 0f3E800000;
	add.ftz.f32 	%f215, %f33, %f29;
	add.ftz.f32 	%f216, %f215, %f251;
	add.ftz.f32 	%f217, %f216, %f31;
	mul.ftz.f32 	%f181, %f217, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f176, %f177, %f178, %f179}, [inTexture, {%f180, %f181}];
	// inline asm
	add.ftz.f32 	%f259, %f208, %f176;
	add.ftz.f32 	%f258, %f209, %f177;
	add.ftz.f32 	%f260, %f210, %f178;
	add.ftz.f32 	%f261, %f211, %f179;

BB0_15:
	mul.wide.s32 	%rd33, %r34, 4;
	add.s64 	%rd34, %rd8, %rd33;
	ld.local.u32 	%r27, [%rd34];
	add.s32 	%r28, %r27, -1;
	st.local.u32 	[%rd34], %r28;
	setp.eq.s32	%p13, %r28, 0;
	selp.b32	%r29, -1, 0, %p13;
	add.s32 	%r34, %r29, %r34;
	bra.uni 	BB0_17;

BB0_16:
	shl.b64 	%rd36, %rd4, 2;
	add.s64 	%rd37, %rd8, %rd36;
	add.s32 	%r33, %r33, -1;
	add.ftz.f32 	%f262, %f262, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f218, %f219, %f220, %f221}, [inTexture, {%f32, %f33}];
	// inline asm
	add.ftz.f32 	%f259, %f259, %f218;
	add.ftz.f32 	%f258, %f258, %f219;
	add.ftz.f32 	%f260, %f260, %f220;
	add.ftz.f32 	%f261, %f261, %f221;
	ld.local.u32 	%r30, [%rd37];
	add.s32 	%r31, %r30, -1;
	st.local.u32 	[%rd37], %r31;
	setp.eq.s32	%p14, %r31, 0;
	selp.b32	%r32, -1, 0, %p14;
	add.s32 	%r34, %r32, %r34;

BB0_17:
	setp.gt.s32	%p15, %r33, -1;
	@%p15 bra 	BB0_19;

	mov.f32 	%f224, 0f3F800000;
	div.approx.ftz.f32 	%f225, %f224, %f262;
	mul.ftz.f32 	%f226, %f259, %f225;
	mul.ftz.f32 	%f227, %f258, %f225;
	mul.ftz.f32 	%f228, %f260, %f225;
	mul.ftz.f32 	%f229, %f261, %f225;
	st.param.f32	[func_retval0+0], %f226;
	st.param.f32	[func_retval0+4], %f227;
	st.param.f32	[func_retval0+8], %f228;
	st.param.f32	[func_retval0+12], %f229;
	ret;

BB0_19:
	mul.wide.s32 	%rd38, %r33, 32;
	add.s64 	%rd39, %rd6, %rd38;
	ld.local.v2.f32 	{%f230, %f231}, [%rd39];
	ld.local.v2.f32 	{%f232, %f233}, [%rd39+8];
	ld.local.v2.f32 	{%f234, %f235}, [%rd39+16];
	ld.local.v2.f32 	{%f236, %f237}, [%rd39+24];
	mov.f32 	%f251, %f237;
	mov.f32 	%f250, %f236;
	mov.f32 	%f253, %f235;
	mov.f32 	%f252, %f234;
	mov.f32 	%f255, %f233;
	mov.f32 	%f254, %f232;
	mov.f32 	%f257, %f231;
	mov.f32 	%f256, %f230;
	bra.uni 	BB0_1;
}

.visible .func  (.param .align 16 .b8 func_retval0[16]) _Z10SampleQuad6float2S_S_S_Pv(
	.param .align 8 .b8 _Z10SampleQuad6float2S_S_S_Pv_param_0[8],
	.param .align 8 .b8 _Z10SampleQuad6float2S_S_S_Pv_param_1[8],
	.param .align 8 .b8 _Z10SampleQuad6float2S_S_S_Pv_param_2[8],
	.param .align 8 .b8 _Z10SampleQuad6float2S_S_S_Pv_param_3[8],
	.param .b64 _Z10SampleQuad6float2S_S_S_Pv_param_4
)
{
	.local .align 8 .b8 	__local_depot1[256];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<23>;
	.reg .s32 	%r<39>;
	.reg .f32 	%f<813>;
	.reg .s64 	%rd<126>;


	mov.u64 	%SPL, __local_depot1;
	cvta.local.u64 	%SP, %SPL;
	ld.param.f32 	%f5, [_Z10SampleQuad6float2S_S_S_Pv_param_0+4];
	ld.param.f32 	%f2, [_Z10SampleQuad6float2S_S_S_Pv_param_0];
	ld.param.f32 	%f4, [_Z10SampleQuad6float2S_S_S_Pv_param_1+4];
	ld.param.f32 	%f1, [_Z10SampleQuad6float2S_S_S_Pv_param_1];
	ld.param.f32 	%f9, [_Z10SampleQuad6float2S_S_S_Pv_param_2+4];
	ld.param.f32 	%f7, [_Z10SampleQuad6float2S_S_S_Pv_param_2];
	ld.param.f32 	%f13, [_Z10SampleQuad6float2S_S_S_Pv_param_3+4];
	ld.param.f32 	%f11, [_Z10SampleQuad6float2S_S_S_Pv_param_3];
	add.ftz.f32 	%f134, %f2, %f1;
	mul.ftz.f32 	%f3, %f134, 0f3F000000;
	add.ftz.f32 	%f135, %f5, %f4;
	mul.ftz.f32 	%f6, %f135, 0f3F000000;
	add.ftz.f32 	%f136, %f1, %f7;
	mul.ftz.f32 	%f8, %f136, 0f3F000000;
	add.ftz.f32 	%f137, %f4, %f9;
	mul.ftz.f32 	%f10, %f137, 0f3F000000;
	add.ftz.f32 	%f138, %f7, %f11;
	mul.ftz.f32 	%f12, %f138, 0f3F000000;
	add.ftz.f32 	%f139, %f9, %f13;
	mul.ftz.f32 	%f14, %f139, 0f3F000000;
	add.ftz.f32 	%f140, %f11, %f2;
	mul.ftz.f32 	%f15, %f140, 0f3F000000;
	add.ftz.f32 	%f141, %f13, %f5;
	mul.ftz.f32 	%f16, %f141, 0f3F000000;
	add.ftz.f32 	%f142, %f3, %f12;
	mul.ftz.f32 	%f17, %f142, 0f3F000000;
	add.ftz.f32 	%f143, %f6, %f14;
	mul.ftz.f32 	%f18, %f143, 0f3F000000;
	sub.ftz.f32 	%f144, %f15, %f8;
	sub.ftz.f32 	%f145, %f16, %f10;
	mul.ftz.f32 	%f146, %f145, %f145;
	fma.rn.ftz.f32 	%f19, %f144, %f144, %f146;
	setp.gt.ftz.f32	%p1, %f19, 0f3F8020C5;
	selp.u32	%r14, 1, 0, %p1;
	sub.ftz.f32 	%f147, %f3, %f12;
	sub.ftz.f32 	%f148, %f6, %f14;
	mul.ftz.f32 	%f149, %f148, %f148;
	fma.rn.ftz.f32 	%f20, %f147, %f147, %f149;
	setp.gt.ftz.f32	%p2, %f20, 0f3F8020C5;
	or.b32  	%r15, %r14, 2;
	selp.b32	%r16, %r15, %r14, %p2;
	setp.eq.s32	%p3, %r16, 0;
	mov.f32 	%f797, %f13;
	mov.f32 	%f796, %f11;
	mov.f32 	%f799, %f9;
	mov.f32 	%f798, %f7;
	mov.f32 	%f801, %f4;
	mov.f32 	%f800, %f1;
	mov.f32 	%f803, %f5;
	mov.f32 	%f802, %f2;
	@%p3 bra 	BB1_25;

	setp.gt.ftz.f32	%p4, %f19, 0f40800000;
	setp.gt.ftz.f32	%p5, %f20, 0f40800000;
	or.pred  	%p6, %p4, %p5;
	@%p6 bra 	BB1_5;

	add.ftz.f32 	%f174, %f17, %f3;
	add.ftz.f32 	%f175, %f174, %f2;
	add.ftz.f32 	%f176, %f175, %f15;
	mul.ftz.f32 	%f154, %f176, 0f3E800000;
	add.ftz.f32 	%f177, %f18, %f6;
	add.ftz.f32 	%f178, %f177, %f5;
	add.ftz.f32 	%f179, %f178, %f16;
	mul.ftz.f32 	%f155, %f179, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f150, %f151, %f152, %f153}, [inTexture, {%f154, %f155}];
	// inline asm
	add.ftz.f32 	%f180, %f174, %f1;
	add.ftz.f32 	%f181, %f180, %f8;
	mul.ftz.f32 	%f160, %f181, 0f3E800000;
	add.ftz.f32 	%f182, %f177, %f4;
	add.ftz.f32 	%f183, %f182, %f10;
	mul.ftz.f32 	%f161, %f183, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f156, %f157, %f158, %f159}, [inTexture, {%f160, %f161}];
	// inline asm
	add.ftz.f32 	%f184, %f150, %f156;
	add.ftz.f32 	%f185, %f151, %f157;
	add.ftz.f32 	%f186, %f152, %f158;
	add.ftz.f32 	%f187, %f153, %f159;
	add.ftz.f32 	%f188, %f17, %f12;
	add.ftz.f32 	%f189, %f188, %f7;
	add.ftz.f32 	%f190, %f189, %f8;
	mul.ftz.f32 	%f166, %f190, 0f3E800000;
	add.ftz.f32 	%f191, %f18, %f14;
	add.ftz.f32 	%f192, %f191, %f9;
	add.ftz.f32 	%f193, %f192, %f10;
	mul.ftz.f32 	%f167, %f193, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f162, %f163, %f164, %f165}, [inTexture, {%f166, %f167}];
	// inline asm
	add.ftz.f32 	%f194, %f184, %f162;
	add.ftz.f32 	%f195, %f185, %f163;
	add.ftz.f32 	%f196, %f186, %f164;
	add.ftz.f32 	%f197, %f187, %f165;
	add.ftz.f32 	%f198, %f188, %f11;
	add.ftz.f32 	%f199, %f198, %f15;
	mul.ftz.f32 	%f172, %f199, 0f3E800000;
	add.ftz.f32 	%f200, %f191, %f13;
	add.ftz.f32 	%f201, %f200, %f16;
	mul.ftz.f32 	%f173, %f201, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f168, %f169, %f170, %f171}, [inTexture, {%f172, %f173}];
	// inline asm
	add.ftz.f32 	%f202, %f194, %f168;
	add.ftz.f32 	%f203, %f195, %f169;
	add.ftz.f32 	%f204, %f196, %f170;
	add.ftz.f32 	%f205, %f197, %f171;
	mul.ftz.f32 	%f792, %f202, 0f3E800000;
	mul.ftz.f32 	%f793, %f203, 0f3E800000;
	mul.ftz.f32 	%f794, %f204, 0f3E800000;
	mul.ftz.f32 	%f795, %f205, 0f3E800000;
	max.ftz.f32 	%f25, %f19, %f20;
	setp.geu.ftz.f32	%p7, %f25, 0f3FB851EC;
	@%p7 bra 	BB1_4;

	add.ftz.f32 	%f302, %f17, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f303, %f302;
	add.ftz.f32 	%f304, %f18, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f305, %f304;
	add.ftz.f32 	%f306, %f303, 0fBF000000;
	add.ftz.f32 	%f307, %f305, 0fBF000000;
	add.ftz.f32 	%f258, %f306, 0f00000000;
	add.ftz.f32 	%f229, %f307, 0f00000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f206, %f207, %f208, %f209}, [inTexture, {%f258, %f229}];
	// inline asm
	sub.ftz.f32 	%f308, %f304, %f305;
	sub.ftz.f32 	%f309, %f302, %f303;
	fma.rn.ftz.f32 	%f310, %f308, 0fBF000000, 0f3F800000;
	mov.f32 	%f311, 0f3F800000;
	fma.rn.ftz.f32 	%f312, %f309, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f313, %f310, %f308, 0fBF000000;
	fma.rn.ftz.f32 	%f314, %f312, %f309, 0fBF000000;
	mul.ftz.f32 	%f315, %f313, %f308;
	mul.ftz.f32 	%f316, %f314, %f309;
	mul.ftz.f32 	%f317, %f316, %f315;
	add.ftz.f32 	%f288, %f306, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f212, %f213, %f214, %f215}, [inTexture, {%f288, %f229}];
	// inline asm
	fma.rn.ftz.f32 	%f318, %f309, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f319, %f318, %f309;
	fma.rn.ftz.f32 	%f320, %f319, %f309, 0f3F800000;
	mul.ftz.f32 	%f321, %f320, %f315;
	mul.ftz.f32 	%f322, %f212, %f321;
	mul.ftz.f32 	%f323, %f213, %f321;
	mul.ftz.f32 	%f324, %f214, %f321;
	mul.ftz.f32 	%f325, %f215, %f321;
	fma.rn.ftz.f32 	%f326, %f206, %f317, %f322;
	fma.rn.ftz.f32 	%f327, %f207, %f317, %f323;
	fma.rn.ftz.f32 	%f328, %f208, %f317, %f324;
	fma.rn.ftz.f32 	%f329, %f209, %f317, %f325;
	add.ftz.f32 	%f294, %f306, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f218, %f219, %f220, %f221}, [inTexture, {%f294, %f229}];
	// inline asm
	sub.ftz.f32 	%f330, %f311, %f309;
	fma.rn.ftz.f32 	%f331, %f330, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f332, %f331, %f330;
	fma.rn.ftz.f32 	%f333, %f332, %f330, 0f3F800000;
	mul.ftz.f32 	%f334, %f333, %f315;
	fma.rn.ftz.f32 	%f335, %f218, %f334, %f326;
	fma.rn.ftz.f32 	%f336, %f219, %f334, %f327;
	fma.rn.ftz.f32 	%f337, %f220, %f334, %f328;
	fma.rn.ftz.f32 	%f338, %f221, %f334, %f329;
	add.ftz.f32 	%f300, %f306, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f224, %f225, %f226, %f227}, [inTexture, {%f300, %f229}];
	// inline asm
	fma.rn.ftz.f32 	%f339, %f330, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f340, %f339, %f330, 0fBF000000;
	mul.ftz.f32 	%f341, %f340, %f330;
	mul.ftz.f32 	%f342, %f341, %f315;
	fma.rn.ftz.f32 	%f343, %f224, %f342, %f335;
	fma.rn.ftz.f32 	%f344, %f225, %f342, %f336;
	fma.rn.ftz.f32 	%f345, %f226, %f342, %f337;
	fma.rn.ftz.f32 	%f346, %f227, %f342, %f338;
	add.ftz.f32 	%f253, %f307, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f230, %f231, %f232, %f233}, [inTexture, {%f258, %f253}];
	// inline asm
	fma.rn.ftz.f32 	%f347, %f308, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f348, %f347, %f308;
	fma.rn.ftz.f32 	%f349, %f348, %f308, 0f3F800000;
	mul.ftz.f32 	%f350, %f316, %f349;
	fma.rn.ftz.f32 	%f351, %f230, %f350, %f343;
	fma.rn.ftz.f32 	%f352, %f231, %f350, %f344;
	fma.rn.ftz.f32 	%f353, %f232, %f350, %f345;
	fma.rn.ftz.f32 	%f354, %f233, %f350, %f346;
	// inline asm
	tex.2d.v4.f32.f32 {%f236, %f237, %f238, %f239}, [inTexture, {%f288, %f253}];
	// inline asm
	mul.ftz.f32 	%f355, %f320, %f349;
	fma.rn.ftz.f32 	%f356, %f236, %f355, %f351;
	fma.rn.ftz.f32 	%f357, %f237, %f355, %f352;
	fma.rn.ftz.f32 	%f358, %f238, %f355, %f353;
	fma.rn.ftz.f32 	%f359, %f239, %f355, %f354;
	// inline asm
	tex.2d.v4.f32.f32 {%f242, %f243, %f244, %f245}, [inTexture, {%f294, %f253}];
	// inline asm
	mul.ftz.f32 	%f360, %f333, %f349;
	fma.rn.ftz.f32 	%f361, %f242, %f360, %f356;
	fma.rn.ftz.f32 	%f362, %f243, %f360, %f357;
	fma.rn.ftz.f32 	%f363, %f244, %f360, %f358;
	fma.rn.ftz.f32 	%f364, %f245, %f360, %f359;
	// inline asm
	tex.2d.v4.f32.f32 {%f248, %f249, %f250, %f251}, [inTexture, {%f300, %f253}];
	// inline asm
	mul.ftz.f32 	%f365, %f341, %f349;
	fma.rn.ftz.f32 	%f366, %f248, %f365, %f361;
	fma.rn.ftz.f32 	%f367, %f249, %f365, %f362;
	fma.rn.ftz.f32 	%f368, %f250, %f365, %f363;
	fma.rn.ftz.f32 	%f369, %f251, %f365, %f364;
	add.ftz.f32 	%f277, %f307, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f254, %f255, %f256, %f257}, [inTexture, {%f258, %f277}];
	// inline asm
	sub.ftz.f32 	%f370, %f311, %f308;
	fma.rn.ftz.f32 	%f371, %f370, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f372, %f371, %f370;
	fma.rn.ftz.f32 	%f373, %f372, %f370, 0f3F800000;
	mul.ftz.f32 	%f374, %f316, %f373;
	fma.rn.ftz.f32 	%f375, %f254, %f374, %f366;
	fma.rn.ftz.f32 	%f376, %f255, %f374, %f367;
	fma.rn.ftz.f32 	%f377, %f256, %f374, %f368;
	fma.rn.ftz.f32 	%f378, %f257, %f374, %f369;
	// inline asm
	tex.2d.v4.f32.f32 {%f260, %f261, %f262, %f263}, [inTexture, {%f288, %f277}];
	// inline asm
	mul.ftz.f32 	%f379, %f320, %f373;
	fma.rn.ftz.f32 	%f380, %f260, %f379, %f375;
	fma.rn.ftz.f32 	%f381, %f261, %f379, %f376;
	fma.rn.ftz.f32 	%f382, %f262, %f379, %f377;
	fma.rn.ftz.f32 	%f383, %f263, %f379, %f378;
	// inline asm
	tex.2d.v4.f32.f32 {%f266, %f267, %f268, %f269}, [inTexture, {%f294, %f277}];
	// inline asm
	mul.ftz.f32 	%f384, %f333, %f373;
	fma.rn.ftz.f32 	%f385, %f266, %f384, %f380;
	fma.rn.ftz.f32 	%f386, %f267, %f384, %f381;
	fma.rn.ftz.f32 	%f387, %f268, %f384, %f382;
	fma.rn.ftz.f32 	%f388, %f269, %f384, %f383;
	// inline asm
	tex.2d.v4.f32.f32 {%f272, %f273, %f274, %f275}, [inTexture, {%f300, %f277}];
	// inline asm
	mul.ftz.f32 	%f389, %f341, %f373;
	fma.rn.ftz.f32 	%f390, %f272, %f389, %f385;
	fma.rn.ftz.f32 	%f391, %f273, %f389, %f386;
	fma.rn.ftz.f32 	%f392, %f274, %f389, %f387;
	fma.rn.ftz.f32 	%f393, %f275, %f389, %f388;
	add.ftz.f32 	%f282, %f306, 0f80000000;
	add.ftz.f32 	%f301, %f307, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f278, %f279, %f280, %f281}, [inTexture, {%f282, %f301}];
	// inline asm
	fma.rn.ftz.f32 	%f394, %f370, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f395, %f394, %f370, 0fBF000000;
	mul.ftz.f32 	%f396, %f395, %f370;
	mul.ftz.f32 	%f397, %f316, %f396;
	fma.rn.ftz.f32 	%f398, %f278, %f397, %f390;
	fma.rn.ftz.f32 	%f399, %f279, %f397, %f391;
	fma.rn.ftz.f32 	%f400, %f280, %f397, %f392;
	fma.rn.ftz.f32 	%f401, %f281, %f397, %f393;
	// inline asm
	tex.2d.v4.f32.f32 {%f284, %f285, %f286, %f287}, [inTexture, {%f288, %f301}];
	// inline asm
	mul.ftz.f32 	%f402, %f320, %f396;
	fma.rn.ftz.f32 	%f403, %f284, %f402, %f398;
	fma.rn.ftz.f32 	%f404, %f285, %f402, %f399;
	fma.rn.ftz.f32 	%f405, %f286, %f402, %f400;
	fma.rn.ftz.f32 	%f406, %f287, %f402, %f401;
	// inline asm
	tex.2d.v4.f32.f32 {%f290, %f291, %f292, %f293}, [inTexture, {%f294, %f301}];
	// inline asm
	mul.ftz.f32 	%f407, %f333, %f396;
	fma.rn.ftz.f32 	%f408, %f290, %f407, %f403;
	fma.rn.ftz.f32 	%f409, %f291, %f407, %f404;
	fma.rn.ftz.f32 	%f410, %f292, %f407, %f405;
	fma.rn.ftz.f32 	%f411, %f293, %f407, %f406;
	// inline asm
	tex.2d.v4.f32.f32 {%f296, %f297, %f298, %f299}, [inTexture, {%f300, %f301}];
	// inline asm
	add.ftz.f32 	%f412, %f25, 0fBF800000;
	mov.f32 	%f413, 0f3EE147B0;
	div.approx.ftz.f32 	%f414, %f412, %f413;
	mul.ftz.f32 	%f415, %f341, %f396;
	fma.rn.ftz.f32 	%f416, %f296, %f415, %f408;
	fma.rn.ftz.f32 	%f417, %f297, %f415, %f409;
	fma.rn.ftz.f32 	%f418, %f298, %f415, %f410;
	fma.rn.ftz.f32 	%f419, %f299, %f415, %f411;
	sub.ftz.f32 	%f420, %f792, %f416;
	fma.rn.ftz.f32 	%f792, %f414, %f420, %f416;
	sub.ftz.f32 	%f421, %f793, %f417;
	fma.rn.ftz.f32 	%f793, %f414, %f421, %f417;
	sub.ftz.f32 	%f422, %f794, %f418;
	fma.rn.ftz.f32 	%f794, %f414, %f422, %f418;
	sub.ftz.f32 	%f423, %f795, %f419;
	fma.rn.ftz.f32 	%f795, %f414, %f423, %f419;

BB1_4:
	mov.f32 	%f812, %f795;
	mov.f32 	%f811, %f794;
	mov.f32 	%f810, %f793;
	mov.f32 	%f809, %f792;
	bra.uni 	BB1_26;

BB1_5:
	add.u64 	%rd43, %SP, 32;
	add.u64 	%rd44, %SP, 16;
	add.u64 	%rd45, %SP, 0;
	cvta.to.local.u64 	%rd46, %rd43;
	cvta.to.local.u64 	%rd47, %rd44;
	cvta.to.local.u64 	%rd48, %rd45;
	mov.u32 	%r37, 0;
	st.local.u32 	[%rd47], %r37;
	mov.u32 	%r19, 1;
	st.local.u32 	[%rd48], %r19;
	st.local.v2.f32 	[%rd46], {%f2, %f5};
	st.local.v2.f32 	[%rd46+8], {%f1, %f4};
	st.local.f32 	[%rd46+16], %f7;
	st.local.f32 	[%rd46+20], %f9;
	st.local.v2.f32 	[%rd46+24], {%f11, %f13};
	mov.f32 	%f808, 0f00000000;
	mov.u32 	%r36, %r37;
	mov.f32 	%f807, %f808;
	mov.f32 	%f806, %f808;
	mov.f32 	%f805, %f808;
	mov.f32 	%f804, %f808;

BB1_6:
	add.ftz.f32 	%f53, %f802, %f800;
	mul.ftz.f32 	%f54, %f53, 0f3F000000;
	add.ftz.f32 	%f57, %f803, %f801;
	mul.ftz.f32 	%f58, %f57, 0f3F000000;
	add.ftz.f32 	%f429, %f800, %f798;
	mul.ftz.f32 	%f60, %f429, 0f3F000000;
	add.ftz.f32 	%f430, %f801, %f799;
	mul.ftz.f32 	%f62, %f430, 0f3F000000;
	add.ftz.f32 	%f431, %f798, %f796;
	mul.ftz.f32 	%f64, %f431, 0f3F000000;
	add.ftz.f32 	%f432, %f799, %f797;
	mul.ftz.f32 	%f66, %f432, 0f3F000000;
	add.ftz.f32 	%f433, %f796, %f802;
	mul.ftz.f32 	%f67, %f433, 0f3F000000;
	add.ftz.f32 	%f434, %f797, %f803;
	mul.ftz.f32 	%f68, %f434, 0f3F000000;
	add.ftz.f32 	%f435, %f54, %f64;
	mul.ftz.f32 	%f69, %f435, 0f3F000000;
	add.ftz.f32 	%f436, %f58, %f66;
	mul.ftz.f32 	%f70, %f436, 0f3F000000;
	sub.ftz.f32 	%f437, %f67, %f60;
	sub.ftz.f32 	%f438, %f68, %f62;
	mul.ftz.f32 	%f439, %f438, %f438;
	fma.rn.ftz.f32 	%f71, %f437, %f437, %f439;
	setp.gt.ftz.f32	%p8, %f71, 0f3F800000;
	selp.u32	%r20, 1, 0, %p8;
	sub.ftz.f32 	%f440, %f64, %f54;
	sub.ftz.f32 	%f441, %f66, %f58;
	mul.ftz.f32 	%f442, %f441, %f441;
	fma.rn.ftz.f32 	%f72, %f440, %f440, %f442;
	setp.gt.ftz.f32	%p9, %f72, 0f3F800000;
	or.b32  	%r21, %r20, 2;
	selp.b32	%r3, %r21, %r20, %p9;
	setp.eq.s32	%p10, %r3, 0;
	cvt.s64.s32	%rd1, %r37;
	@%p10 bra 	BB1_21;

	shl.b64 	%rd51, %rd1, 2;
	add.s64 	%rd52, %rd47, %rd51;
	ld.local.u32 	%r4, [%rd52];
	setp.eq.s32	%p11, %r4, 2;
	@%p11 bra 	BB1_15;

	setp.lt.ftz.f32	%p12, %f71, 0f40800000;
	setp.lt.ftz.f32	%p13, %f72, 0f40800000;
	and.pred  	%p14, %p12, %p13;
	@%p14 bra 	BB1_15;

	setp.eq.s32	%p15, %r3, 3;
	mul.wide.s32 	%rd55, %r36, 32;
	add.s64 	%rd2, %rd46, %rd55;
	@%p15 bra 	BB1_13;

	setp.eq.s32	%p16, %r3, 2;
	st.local.v2.f32 	[%rd2], {%f802, %f803};
	@%p16 bra 	BB1_12;

	st.local.v2.f32 	[%rd2+8], {%f54, %f58};
	st.local.v2.f32 	[%rd2+16], {%f64, %f66};
	st.local.v2.f32 	[%rd2+24], {%f796, %f797};
	st.local.v2.f32 	[%rd2+32], {%f54, %f58};
	st.local.f32 	[%rd2+40], %f800;
	st.local.f32 	[%rd2+44], %f801;
	st.local.v2.f32 	[%rd2+48], {%f798, %f799};
	st.local.f32 	[%rd2+56], %f64;
	st.local.f32 	[%rd2+60], %f66;
	mov.u32 	%r38, 2;
	bra.uni 	BB1_14;

BB1_12:
	st.local.v2.f32 	[%rd2+8], {%f800, %f801};
	st.local.v2.f32 	[%rd2+16], {%f60, %f62};
	st.local.f32 	[%rd2+24], %f67;
	st.local.f32 	[%rd2+28], %f68;
	st.local.v2.f32 	[%rd2+32], {%f67, %f68};
	st.local.f32 	[%rd2+40], %f60;
	st.local.f32 	[%rd2+44], %f62;
	st.local.v2.f32 	[%rd2+48], {%f798, %f799};
	st.local.f32 	[%rd2+56], %f796;
	st.local.f32 	[%rd2+60], %f797;
	mov.u32 	%r38, 2;
	bra.uni 	BB1_14;

BB1_13:
	st.local.v2.f32 	[%rd2], {%f69, %f70};
	st.local.v2.f32 	[%rd2+8], {%f67, %f68};
	st.local.v2.f32 	[%rd2+16], {%f802, %f803};
	st.local.v2.f32 	[%rd2+24], {%f54, %f58};
	st.local.f32 	[%rd2+32], %f69;
	st.local.f32 	[%rd2+36], %f70;
	st.local.v2.f32 	[%rd2+40], {%f54, %f58};
	st.local.f32 	[%rd2+48], %f800;
	st.local.f32 	[%rd2+52], %f801;
	st.local.v2.f32 	[%rd2+56], {%f60, %f62};
	st.local.f32 	[%rd2+64], %f69;
	st.local.f32 	[%rd2+68], %f70;
	st.local.v2.f32 	[%rd2+72], {%f60, %f62};
	st.local.f32 	[%rd2+80], %f798;
	st.local.f32 	[%rd2+84], %f799;
	st.local.v2.f32 	[%rd2+88], {%f64, %f66};
	st.local.f32 	[%rd2+96], %f69;
	st.local.f32 	[%rd2+100], %f70;
	st.local.v2.f32 	[%rd2+104], {%f64, %f66};
	st.local.f32 	[%rd2+112], %f796;
	st.local.f32 	[%rd2+116], %f797;
	st.local.v2.f32 	[%rd2+120], {%f67, %f68};
	mov.u32 	%r38, 4;

BB1_14:
	add.s32 	%r25, %r36, %r38;
	add.s32 	%r36, %r25, -1;
	mul.wide.s32 	%rd58, %r37, 4;
	add.s64 	%rd59, %rd48, %rd58;
	ld.local.u32 	%r26, [%rd59];
	add.s32 	%r27, %r26, -1;
	st.local.u32 	[%rd59], %r27;
	setp.ne.s32	%p17, %r27, 0;
	selp.u32	%r28, 1, 0, %p17;
	add.s32 	%r37, %r28, %r37;
	mul.wide.s32 	%rd62, %r37, 4;
	add.s64 	%rd63, %rd47, %rd62;
	add.s32 	%r29, %r4, 1;
	st.local.u32 	[%rd63], %r29;
	add.s64 	%rd64, %rd48, %rd62;
	st.local.u32 	[%rd64], %r38;
	bra.uni 	BB1_22;

BB1_15:
	add.s32 	%r36, %r36, -1;
	setp.eq.s32	%p18, %r3, 3;
	@%p18 bra 	BB1_19;

	setp.eq.s32	%p19, %r3, 2;
	@%p19 bra 	BB1_18;

	add.ftz.f32 	%f455, %f802, %f54;
	add.ftz.f32 	%f456, %f455, %f64;
	add.ftz.f32 	%f457, %f456, %f796;
	mul.ftz.f32 	%f447, %f457, 0f3E800000;
	add.ftz.f32 	%f458, %f803, %f58;
	add.ftz.f32 	%f459, %f458, %f66;
	add.ftz.f32 	%f460, %f459, %f797;
	mul.ftz.f32 	%f448, %f460, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f443, %f444, %f445, %f446}, [inTexture, {%f447, %f448}];
	// inline asm
	add.ftz.f32 	%f461, %f805, %f443;
	add.ftz.f32 	%f462, %f804, %f444;
	add.ftz.f32 	%f463, %f806, %f445;
	add.ftz.f32 	%f464, %f807, %f446;
	add.ftz.f32 	%f465, %f54, %f800;
	add.ftz.f32 	%f466, %f465, %f798;
	add.ftz.f32 	%f467, %f466, %f64;
	mul.ftz.f32 	%f453, %f467, 0f3E800000;
	add.ftz.f32 	%f468, %f58, %f801;
	add.ftz.f32 	%f469, %f468, %f799;
	add.ftz.f32 	%f470, %f469, %f66;
	mul.ftz.f32 	%f454, %f470, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f449, %f450, %f451, %f452}, [inTexture, {%f453, %f454}];
	// inline asm
	add.ftz.f32 	%f805, %f461, %f449;
	add.ftz.f32 	%f804, %f462, %f450;
	add.ftz.f32 	%f806, %f463, %f451;
	add.ftz.f32 	%f807, %f464, %f452;
	add.ftz.f32 	%f808, %f808, 0f40000000;
	bra.uni 	BB1_20;

BB1_18:
	add.ftz.f32 	%f791, %f803, %f801;
	add.ftz.f32 	%f790, %f802, %f800;
	add.ftz.f32 	%f483, %f790, %f60;
	add.ftz.f32 	%f484, %f483, %f67;
	mul.ftz.f32 	%f475, %f484, 0f3E800000;
	add.ftz.f32 	%f485, %f791, %f62;
	add.ftz.f32 	%f486, %f485, %f68;
	mul.ftz.f32 	%f476, %f486, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f471, %f472, %f473, %f474}, [inTexture, {%f475, %f476}];
	// inline asm
	add.ftz.f32 	%f487, %f805, %f471;
	add.ftz.f32 	%f488, %f804, %f472;
	add.ftz.f32 	%f489, %f806, %f473;
	add.ftz.f32 	%f490, %f807, %f474;
	add.ftz.f32 	%f491, %f67, %f60;
	add.ftz.f32 	%f492, %f491, %f798;
	add.ftz.f32 	%f493, %f492, %f796;
	mul.ftz.f32 	%f481, %f493, 0f3E800000;
	add.ftz.f32 	%f494, %f68, %f62;
	add.ftz.f32 	%f495, %f494, %f799;
	add.ftz.f32 	%f496, %f495, %f797;
	mul.ftz.f32 	%f482, %f496, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f477, %f478, %f479, %f480}, [inTexture, {%f481, %f482}];
	// inline asm
	add.ftz.f32 	%f805, %f487, %f477;
	add.ftz.f32 	%f804, %f488, %f478;
	add.ftz.f32 	%f806, %f489, %f479;
	add.ftz.f32 	%f807, %f490, %f480;
	add.ftz.f32 	%f808, %f808, 0f40000000;
	bra.uni 	BB1_20;

BB1_19:
	add.ftz.f32 	%f808, %f808, 0f40800000;
	add.ftz.f32 	%f521, %f69, %f67;
	add.ftz.f32 	%f522, %f521, %f802;
	add.ftz.f32 	%f523, %f522, %f54;
	mul.ftz.f32 	%f501, %f523, 0f3E800000;
	add.ftz.f32 	%f524, %f70, %f68;
	add.ftz.f32 	%f525, %f524, %f803;
	add.ftz.f32 	%f526, %f525, %f58;
	mul.ftz.f32 	%f502, %f526, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f497, %f498, %f499, %f500}, [inTexture, {%f501, %f502}];
	// inline asm
	add.ftz.f32 	%f527, %f805, %f497;
	add.ftz.f32 	%f528, %f804, %f498;
	add.ftz.f32 	%f529, %f806, %f499;
	add.ftz.f32 	%f530, %f807, %f500;
	add.ftz.f32 	%f531, %f69, %f54;
	add.ftz.f32 	%f532, %f531, %f800;
	add.ftz.f32 	%f533, %f532, %f60;
	mul.ftz.f32 	%f507, %f533, 0f3E800000;
	add.ftz.f32 	%f534, %f70, %f58;
	add.ftz.f32 	%f535, %f534, %f801;
	add.ftz.f32 	%f536, %f535, %f62;
	mul.ftz.f32 	%f508, %f536, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f503, %f504, %f505, %f506}, [inTexture, {%f507, %f508}];
	// inline asm
	add.ftz.f32 	%f537, %f527, %f503;
	add.ftz.f32 	%f538, %f528, %f504;
	add.ftz.f32 	%f539, %f529, %f505;
	add.ftz.f32 	%f540, %f530, %f506;
	add.ftz.f32 	%f541, %f69, %f60;
	add.ftz.f32 	%f542, %f541, %f798;
	add.ftz.f32 	%f543, %f542, %f64;
	mul.ftz.f32 	%f513, %f543, 0f3E800000;
	add.ftz.f32 	%f544, %f70, %f62;
	add.ftz.f32 	%f545, %f544, %f799;
	add.ftz.f32 	%f546, %f545, %f66;
	mul.ftz.f32 	%f514, %f546, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f509, %f510, %f511, %f512}, [inTexture, {%f513, %f514}];
	// inline asm
	add.ftz.f32 	%f547, %f537, %f509;
	add.ftz.f32 	%f548, %f538, %f510;
	add.ftz.f32 	%f549, %f539, %f511;
	add.ftz.f32 	%f550, %f540, %f512;
	add.ftz.f32 	%f551, %f69, %f64;
	add.ftz.f32 	%f552, %f551, %f796;
	add.ftz.f32 	%f553, %f552, %f67;
	mul.ftz.f32 	%f519, %f553, 0f3E800000;
	add.ftz.f32 	%f554, %f70, %f66;
	add.ftz.f32 	%f555, %f554, %f797;
	add.ftz.f32 	%f556, %f555, %f68;
	mul.ftz.f32 	%f520, %f556, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f515, %f516, %f517, %f518}, [inTexture, {%f519, %f520}];
	// inline asm
	add.ftz.f32 	%f805, %f547, %f515;
	add.ftz.f32 	%f804, %f548, %f516;
	add.ftz.f32 	%f806, %f549, %f517;
	add.ftz.f32 	%f807, %f550, %f518;

BB1_20:
	mul.wide.s32 	%rd83, %r37, 4;
	add.s64 	%rd84, %rd48, %rd83;
	ld.local.u32 	%r30, [%rd84];
	add.s32 	%r31, %r30, -1;
	st.local.u32 	[%rd84], %r31;
	setp.eq.s32	%p20, %r31, 0;
	selp.b32	%r32, -1, 0, %p20;
	add.s32 	%r37, %r32, %r37;
	bra.uni 	BB1_22;

BB1_21:
	add.s32 	%r36, %r36, -1;
	add.ftz.f32 	%f808, %f808, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f557, %f558, %f559, %f560}, [inTexture, {%f69, %f70}];
	// inline asm
	add.ftz.f32 	%f805, %f805, %f557;
	add.ftz.f32 	%f804, %f804, %f558;
	add.ftz.f32 	%f806, %f806, %f559;
	add.ftz.f32 	%f807, %f807, %f560;
	mul.wide.s32 	%rd88, %r37, 4;
	add.s64 	%rd89, %rd48, %rd88;
	ld.local.u32 	%r33, [%rd89];
	add.s32 	%r34, %r33, -1;
	st.local.u32 	[%rd89], %r34;
	setp.eq.s32	%p21, %r34, 0;
	selp.b32	%r35, -1, 0, %p21;
	add.s32 	%r37, %r35, %r37;

BB1_22:
	setp.gt.s32	%p22, %r36, -1;
	@%p22 bra 	BB1_24;

	mov.f32 	%f563, 0f3F800000;
	div.approx.ftz.f32 	%f564, %f563, %f808;
	mul.ftz.f32 	%f809, %f805, %f564;
	mul.ftz.f32 	%f810, %f804, %f564;
	mul.ftz.f32 	%f811, %f806, %f564;
	mul.ftz.f32 	%f812, %f807, %f564;
	bra.uni 	BB1_26;

BB1_24:
	mul.wide.s32 	%rd92, %r36, 32;
	add.s64 	%rd93, %rd46, %rd92;
	ld.local.v2.f32 	{%f565, %f566}, [%rd93];
	ld.local.v2.f32 	{%f567, %f568}, [%rd93+8];
	ld.local.v2.f32 	{%f569, %f570}, [%rd93+16];
	ld.local.v2.f32 	{%f571, %f572}, [%rd93+24];
	mov.f32 	%f797, %f572;
	mov.f32 	%f796, %f571;
	mov.f32 	%f799, %f570;
	mov.f32 	%f798, %f569;
	mov.f32 	%f801, %f568;
	mov.f32 	%f800, %f567;
	mov.f32 	%f803, %f566;
	mov.f32 	%f802, %f565;
	bra.uni 	BB1_6;

BB1_25:
	add.ftz.f32 	%f669, %f17, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f670, %f669;
	add.ftz.f32 	%f671, %f18, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f672, %f671;
	add.ftz.f32 	%f673, %f670, 0fBF000000;
	add.ftz.f32 	%f674, %f672, 0fBF000000;
	add.ftz.f32 	%f625, %f673, 0f00000000;
	add.ftz.f32 	%f596, %f674, 0f00000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f573, %f574, %f575, %f576}, [inTexture, {%f625, %f596}];
	// inline asm
	sub.ftz.f32 	%f675, %f671, %f672;
	sub.ftz.f32 	%f676, %f669, %f670;
	fma.rn.ftz.f32 	%f677, %f675, 0fBF000000, 0f3F800000;
	mov.f32 	%f678, 0f3F800000;
	fma.rn.ftz.f32 	%f679, %f676, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f680, %f677, %f675, 0fBF000000;
	fma.rn.ftz.f32 	%f681, %f679, %f676, 0fBF000000;
	mul.ftz.f32 	%f682, %f680, %f675;
	mul.ftz.f32 	%f683, %f681, %f676;
	mul.ftz.f32 	%f684, %f683, %f682;
	add.ftz.f32 	%f655, %f673, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f579, %f580, %f581, %f582}, [inTexture, {%f655, %f596}];
	// inline asm
	fma.rn.ftz.f32 	%f685, %f676, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f686, %f685, %f676;
	fma.rn.ftz.f32 	%f687, %f686, %f676, 0f3F800000;
	mul.ftz.f32 	%f688, %f687, %f682;
	mul.ftz.f32 	%f689, %f579, %f688;
	mul.ftz.f32 	%f690, %f580, %f688;
	mul.ftz.f32 	%f691, %f581, %f688;
	mul.ftz.f32 	%f692, %f582, %f688;
	fma.rn.ftz.f32 	%f693, %f573, %f684, %f689;
	fma.rn.ftz.f32 	%f694, %f574, %f684, %f690;
	fma.rn.ftz.f32 	%f695, %f575, %f684, %f691;
	fma.rn.ftz.f32 	%f696, %f576, %f684, %f692;
	add.ftz.f32 	%f661, %f673, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f585, %f586, %f587, %f588}, [inTexture, {%f661, %f596}];
	// inline asm
	sub.ftz.f32 	%f697, %f678, %f676;
	fma.rn.ftz.f32 	%f698, %f697, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f699, %f698, %f697;
	fma.rn.ftz.f32 	%f700, %f699, %f697, 0f3F800000;
	mul.ftz.f32 	%f701, %f700, %f682;
	fma.rn.ftz.f32 	%f702, %f585, %f701, %f693;
	fma.rn.ftz.f32 	%f703, %f586, %f701, %f694;
	fma.rn.ftz.f32 	%f704, %f587, %f701, %f695;
	fma.rn.ftz.f32 	%f705, %f588, %f701, %f696;
	add.ftz.f32 	%f667, %f673, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f591, %f592, %f593, %f594}, [inTexture, {%f667, %f596}];
	// inline asm
	fma.rn.ftz.f32 	%f706, %f697, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f707, %f706, %f697, 0fBF000000;
	mul.ftz.f32 	%f708, %f707, %f697;
	mul.ftz.f32 	%f709, %f708, %f682;
	fma.rn.ftz.f32 	%f710, %f591, %f709, %f702;
	fma.rn.ftz.f32 	%f711, %f592, %f709, %f703;
	fma.rn.ftz.f32 	%f712, %f593, %f709, %f704;
	fma.rn.ftz.f32 	%f713, %f594, %f709, %f705;
	add.ftz.f32 	%f620, %f674, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f597, %f598, %f599, %f600}, [inTexture, {%f625, %f620}];
	// inline asm
	fma.rn.ftz.f32 	%f714, %f675, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f715, %f714, %f675;
	fma.rn.ftz.f32 	%f716, %f715, %f675, 0f3F800000;
	mul.ftz.f32 	%f717, %f683, %f716;
	fma.rn.ftz.f32 	%f718, %f597, %f717, %f710;
	fma.rn.ftz.f32 	%f719, %f598, %f717, %f711;
	fma.rn.ftz.f32 	%f720, %f599, %f717, %f712;
	fma.rn.ftz.f32 	%f721, %f600, %f717, %f713;
	// inline asm
	tex.2d.v4.f32.f32 {%f603, %f604, %f605, %f606}, [inTexture, {%f655, %f620}];
	// inline asm
	mul.ftz.f32 	%f722, %f687, %f716;
	fma.rn.ftz.f32 	%f723, %f603, %f722, %f718;
	fma.rn.ftz.f32 	%f724, %f604, %f722, %f719;
	fma.rn.ftz.f32 	%f725, %f605, %f722, %f720;
	fma.rn.ftz.f32 	%f726, %f606, %f722, %f721;
	// inline asm
	tex.2d.v4.f32.f32 {%f609, %f610, %f611, %f612}, [inTexture, {%f661, %f620}];
	// inline asm
	mul.ftz.f32 	%f727, %f700, %f716;
	fma.rn.ftz.f32 	%f728, %f609, %f727, %f723;
	fma.rn.ftz.f32 	%f729, %f610, %f727, %f724;
	fma.rn.ftz.f32 	%f730, %f611, %f727, %f725;
	fma.rn.ftz.f32 	%f731, %f612, %f727, %f726;
	// inline asm
	tex.2d.v4.f32.f32 {%f615, %f616, %f617, %f618}, [inTexture, {%f667, %f620}];
	// inline asm
	mul.ftz.f32 	%f732, %f708, %f716;
	fma.rn.ftz.f32 	%f733, %f615, %f732, %f728;
	fma.rn.ftz.f32 	%f734, %f616, %f732, %f729;
	fma.rn.ftz.f32 	%f735, %f617, %f732, %f730;
	fma.rn.ftz.f32 	%f736, %f618, %f732, %f731;
	add.ftz.f32 	%f644, %f674, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f621, %f622, %f623, %f624}, [inTexture, {%f625, %f644}];
	// inline asm
	sub.ftz.f32 	%f737, %f678, %f675;
	fma.rn.ftz.f32 	%f738, %f737, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f739, %f738, %f737;
	fma.rn.ftz.f32 	%f740, %f739, %f737, 0f3F800000;
	mul.ftz.f32 	%f741, %f683, %f740;
	fma.rn.ftz.f32 	%f742, %f621, %f741, %f733;
	fma.rn.ftz.f32 	%f743, %f622, %f741, %f734;
	fma.rn.ftz.f32 	%f744, %f623, %f741, %f735;
	fma.rn.ftz.f32 	%f745, %f624, %f741, %f736;
	// inline asm
	tex.2d.v4.f32.f32 {%f627, %f628, %f629, %f630}, [inTexture, {%f655, %f644}];
	// inline asm
	mul.ftz.f32 	%f746, %f687, %f740;
	fma.rn.ftz.f32 	%f747, %f627, %f746, %f742;
	fma.rn.ftz.f32 	%f748, %f628, %f746, %f743;
	fma.rn.ftz.f32 	%f749, %f629, %f746, %f744;
	fma.rn.ftz.f32 	%f750, %f630, %f746, %f745;
	// inline asm
	tex.2d.v4.f32.f32 {%f633, %f634, %f635, %f636}, [inTexture, {%f661, %f644}];
	// inline asm
	mul.ftz.f32 	%f751, %f700, %f740;
	fma.rn.ftz.f32 	%f752, %f633, %f751, %f747;
	fma.rn.ftz.f32 	%f753, %f634, %f751, %f748;
	fma.rn.ftz.f32 	%f754, %f635, %f751, %f749;
	fma.rn.ftz.f32 	%f755, %f636, %f751, %f750;
	// inline asm
	tex.2d.v4.f32.f32 {%f639, %f640, %f641, %f642}, [inTexture, {%f667, %f644}];
	// inline asm
	mul.ftz.f32 	%f756, %f708, %f740;
	fma.rn.ftz.f32 	%f757, %f639, %f756, %f752;
	fma.rn.ftz.f32 	%f758, %f640, %f756, %f753;
	fma.rn.ftz.f32 	%f759, %f641, %f756, %f754;
	fma.rn.ftz.f32 	%f760, %f642, %f756, %f755;
	add.ftz.f32 	%f649, %f673, 0f80000000;
	add.ftz.f32 	%f668, %f674, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f645, %f646, %f647, %f648}, [inTexture, {%f649, %f668}];
	// inline asm
	fma.rn.ftz.f32 	%f761, %f737, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f762, %f761, %f737, 0fBF000000;
	mul.ftz.f32 	%f763, %f762, %f737;
	mul.ftz.f32 	%f764, %f683, %f763;
	fma.rn.ftz.f32 	%f765, %f645, %f764, %f757;
	fma.rn.ftz.f32 	%f766, %f646, %f764, %f758;
	fma.rn.ftz.f32 	%f767, %f647, %f764, %f759;
	fma.rn.ftz.f32 	%f768, %f648, %f764, %f760;
	// inline asm
	tex.2d.v4.f32.f32 {%f651, %f652, %f653, %f654}, [inTexture, {%f655, %f668}];
	// inline asm
	mul.ftz.f32 	%f769, %f687, %f763;
	fma.rn.ftz.f32 	%f770, %f651, %f769, %f765;
	fma.rn.ftz.f32 	%f771, %f652, %f769, %f766;
	fma.rn.ftz.f32 	%f772, %f653, %f769, %f767;
	fma.rn.ftz.f32 	%f773, %f654, %f769, %f768;
	// inline asm
	tex.2d.v4.f32.f32 {%f657, %f658, %f659, %f660}, [inTexture, {%f661, %f668}];
	// inline asm
	mul.ftz.f32 	%f774, %f700, %f763;
	fma.rn.ftz.f32 	%f775, %f657, %f774, %f770;
	fma.rn.ftz.f32 	%f776, %f658, %f774, %f771;
	fma.rn.ftz.f32 	%f777, %f659, %f774, %f772;
	fma.rn.ftz.f32 	%f778, %f660, %f774, %f773;
	// inline asm
	tex.2d.v4.f32.f32 {%f663, %f664, %f665, %f666}, [inTexture, {%f667, %f668}];
	// inline asm
	mul.ftz.f32 	%f779, %f708, %f763;
	fma.rn.ftz.f32 	%f809, %f663, %f779, %f775;
	fma.rn.ftz.f32 	%f810, %f664, %f779, %f776;
	fma.rn.ftz.f32 	%f811, %f665, %f779, %f777;
	fma.rn.ftz.f32 	%f812, %f666, %f779, %f778;

BB1_26:
	st.param.f32	[func_retval0+0], %f809;
	st.param.f32	[func_retval0+4], %f810;
	st.param.f32	[func_retval0+8], %f811;
	st.param.f32	[func_retval0+12], %f812;
	ret;
}

.visible .entry cuda_kernel_renderquad(
	.param .u64 cuda_kernel_renderquad_param_0,
	.param .u32 cuda_kernel_renderquad_param_1,
	.param .u32 cuda_kernel_renderquad_param_2,
	.param .f32 cuda_kernel_renderquad_param_3,
	.param .f32 cuda_kernel_renderquad_param_4,
	.param .u32 cuda_kernel_renderquad_param_5,
	.param .u32 cuda_kernel_renderquad_param_6,
	.param .u32 cuda_kernel_renderquad_param_7,
	.param .align 16 .b8 cuda_kernel_renderquad_param_8[16],
	.param .align 16 .b8 cuda_kernel_renderquad_param_9[16],
	.param .align 16 .b8 cuda_kernel_renderquad_param_10[16],
	.param .align 16 .b8 cuda_kernel_renderquad_param_11[16],
	.param .align 8 .b8 cuda_kernel_renderquad_param_12[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_13[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_14[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_15[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_16[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_17[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_18[8],
	.param .align 8 .b8 cuda_kernel_renderquad_param_19[8],
	.param .u32 cuda_kernel_renderquad_param_20,
	.param .u64 cuda_kernel_renderquad_param_21
)
{
	.local .align 8 .b8 	__local_depot2[256];
	.reg .b64 	%SP;
	.reg .b64 	%SPL;
	.reg .pred 	%p<35>;
	.reg .s16 	%rs<6>;
	.reg .s32 	%r<75>;
	.reg .f32 	%f<1176>;
	.reg .s64 	%rd<138>;


	mov.u64 	%SPL, __local_depot2;
	cvta.local.u64 	%SP, %SPL;
	ld.param.u64 	%rd5, [cuda_kernel_renderquad_param_0];
	ld.param.u32 	%r19, [cuda_kernel_renderquad_param_1];
	ld.param.u32 	%r20, [cuda_kernel_renderquad_param_2];
	ld.param.f32 	%f179, [cuda_kernel_renderquad_param_3];
	ld.param.f32 	%f180, [cuda_kernel_renderquad_param_4];
	ld.param.u32 	%r16, [cuda_kernel_renderquad_param_6];
	ld.param.u32 	%r17, [cuda_kernel_renderquad_param_7];
	ld.param.f32 	%f3, [cuda_kernel_renderquad_param_8+8];
	ld.param.f32 	%f2, [cuda_kernel_renderquad_param_8+4];
	ld.param.f32 	%f1, [cuda_kernel_renderquad_param_8];
	ld.param.f32 	%f6, [cuda_kernel_renderquad_param_9+8];
	ld.param.f32 	%f5, [cuda_kernel_renderquad_param_9+4];
	ld.param.f32 	%f4, [cuda_kernel_renderquad_param_9];
	ld.param.f32 	%f9, [cuda_kernel_renderquad_param_10+8];
	ld.param.f32 	%f8, [cuda_kernel_renderquad_param_10+4];
	ld.param.f32 	%f7, [cuda_kernel_renderquad_param_10];
	ld.param.f32 	%f12, [cuda_kernel_renderquad_param_11+8];
	ld.param.f32 	%f11, [cuda_kernel_renderquad_param_11+4];
	ld.param.f32 	%f10, [cuda_kernel_renderquad_param_11];
	ld.param.f32 	%f198, [cuda_kernel_renderquad_param_12+4];
	ld.param.f32 	%f197, [cuda_kernel_renderquad_param_12];
	ld.param.f32 	%f200, [cuda_kernel_renderquad_param_13+4];
	ld.param.f32 	%f199, [cuda_kernel_renderquad_param_13];
	ld.param.f32 	%f202, [cuda_kernel_renderquad_param_14+4];
	ld.param.f32 	%f201, [cuda_kernel_renderquad_param_14];
	ld.param.f32 	%f204, [cuda_kernel_renderquad_param_15+4];
	ld.param.f32 	%f203, [cuda_kernel_renderquad_param_15];
	ld.param.f32 	%f206, [cuda_kernel_renderquad_param_16+4];
	ld.param.f32 	%f205, [cuda_kernel_renderquad_param_16];
	ld.param.f32 	%f208, [cuda_kernel_renderquad_param_17+4];
	ld.param.f32 	%f207, [cuda_kernel_renderquad_param_17];
	ld.param.f32 	%f210, [cuda_kernel_renderquad_param_18+4];
	ld.param.f32 	%f209, [cuda_kernel_renderquad_param_18];
	ld.param.f32 	%f212, [cuda_kernel_renderquad_param_19+4];
	ld.param.f32 	%f211, [cuda_kernel_renderquad_param_19];
	mov.u32 	%r21, %ntid.x;
	mov.u32 	%r22, %ctaid.x;
	mov.u32 	%r23, %tid.x;
	mad.lo.s32 	%r1, %r21, %r22, %r23;
	mov.u32 	%r24, %ntid.y;
	mov.u32 	%r25, %ctaid.y;
	mov.u32 	%r26, %tid.y;
	mad.lo.s32 	%r2, %r24, %r25, %r26;
	setp.ge.s32	%p1, %r2, %r20;
	setp.ge.s32	%p2, %r1, %r19;
	or.pred  	%p3, %p2, %p1;
	@%p3 bra 	BB2_50;

	cvta.to.global.u64 	%rd6, %rd5;
	cvt.rn.f32.s32	%f13, %r1;
	add.ftz.f32 	%f213, %f13, 0f3F000000;
	cvt.rn.f32.s32	%f14, %r2;
	add.ftz.f32 	%f214, %f14, 0f3F000000;
	sub.ftz.f32 	%f215, %f198, %f214;
	mul.ftz.f32 	%f216, %f205, %f215;
	sub.ftz.f32 	%f217, %f197, %f213;
	mul.ftz.f32 	%f218, %f206, %f217;
	sub.ftz.f32 	%f219, %f216, %f218;
	add.ftz.f32 	%f220, %f219, 0f3F000000;
	cvt.ftz.sat.f32.f32	%f221, %f220;
	mov.f32 	%f222, 0f3F800000;
	sub.ftz.f32 	%f223, %f222, %f221;
	sub.ftz.f32 	%f224, %f200, %f214;
	mul.ftz.f32 	%f225, %f207, %f224;
	sub.ftz.f32 	%f226, %f199, %f213;
	mul.ftz.f32 	%f227, %f208, %f226;
	sub.ftz.f32 	%f228, %f225, %f227;
	add.ftz.f32 	%f229, %f228, 0f3F000000;
	cvt.ftz.sat.f32.f32	%f230, %f229;
	sub.ftz.f32 	%f231, %f222, %f230;
	mul.ftz.f32 	%f232, %f223, %f231;
	sub.ftz.f32 	%f233, %f202, %f214;
	mul.ftz.f32 	%f234, %f209, %f233;
	sub.ftz.f32 	%f235, %f201, %f213;
	mul.ftz.f32 	%f236, %f210, %f235;
	sub.ftz.f32 	%f237, %f234, %f236;
	add.ftz.f32 	%f238, %f237, 0f3F000000;
	cvt.ftz.sat.f32.f32	%f239, %f238;
	sub.ftz.f32 	%f240, %f222, %f239;
	mul.ftz.f32 	%f241, %f232, %f240;
	sub.ftz.f32 	%f242, %f204, %f214;
	mul.ftz.f32 	%f243, %f211, %f242;
	sub.ftz.f32 	%f244, %f203, %f213;
	mul.ftz.f32 	%f245, %f212, %f244;
	sub.ftz.f32 	%f246, %f243, %f245;
	add.ftz.f32 	%f247, %f246, 0f3F000000;
	cvt.ftz.sat.f32.f32	%f248, %f247;
	sub.ftz.f32 	%f249, %f222, %f248;
	mul.ftz.f32 	%f15, %f241, %f249;
	setp.eq.ftz.f32	%p4, %f15, 0f00000000;
	mad.lo.s32 	%r27, %r2, %r16, %r1;
	mul.wide.s32 	%rd7, %r27, 16;
	add.s64 	%rd1, %rd6, %rd7;
	mul.wide.s32 	%rd8, %r27, 8;
	add.s64 	%rd2, %rd6, %rd8;
	@%p4 bra 	BB2_47;

	sub.ftz.f32 	%f250, %f13, %f10;
	sub.ftz.f32 	%f251, %f14, %f11;
	mul.ftz.f32 	%f252, %f251, %f251;
	fma.rn.ftz.f32 	%f253, %f250, %f250, %f252;
	mov.f32 	%f254, 0f00000000;
	sub.ftz.f32 	%f255, %f254, %f12;
	fma.rn.ftz.f32 	%f256, %f255, %f255, %f253;
	rsqrt.approx.ftz.f32 	%f257, %f256;
	mul.ftz.f32 	%f258, %f250, %f257;
	mul.ftz.f32 	%f259, %f251, %f257;
	mul.ftz.f32 	%f260, %f255, %f257;
	mul.ftz.f32 	%f261, %f259, %f9;
	mul.ftz.f32 	%f262, %f8, %f260;
	sub.ftz.f32 	%f263, %f261, %f262;
	mul.ftz.f32 	%f264, %f260, %f7;
	mul.ftz.f32 	%f265, %f9, %f258;
	sub.ftz.f32 	%f266, %f264, %f265;
	mul.ftz.f32 	%f267, %f258, %f8;
	mul.ftz.f32 	%f268, %f7, %f259;
	sub.ftz.f32 	%f269, %f267, %f268;
	mul.ftz.f32 	%f270, %f5, %f266;
	fma.rn.ftz.f32 	%f271, %f4, %f263, %f270;
	fma.rn.ftz.f32 	%f272, %f6, %f269, %f271;
	rcp.approx.ftz.f32 	%f273, %f272;
	sub.ftz.f32 	%f274, %f10, %f1;
	sub.ftz.f32 	%f275, %f11, %f2;
	mul.ftz.f32 	%f276, %f275, %f266;
	fma.rn.ftz.f32 	%f277, %f274, %f263, %f276;
	sub.ftz.f32 	%f278, %f12, %f3;
	fma.rn.ftz.f32 	%f279, %f278, %f269, %f277;
	mul.ftz.f32 	%f280, %f279, %f273;
	mul.ftz.f32 	%f281, %f275, %f6;
	mul.ftz.f32 	%f282, %f5, %f278;
	sub.ftz.f32 	%f283, %f281, %f282;
	mul.ftz.f32 	%f284, %f278, %f4;
	mul.ftz.f32 	%f285, %f6, %f274;
	sub.ftz.f32 	%f286, %f284, %f285;
	mul.ftz.f32 	%f287, %f274, %f5;
	mul.ftz.f32 	%f288, %f4, %f275;
	sub.ftz.f32 	%f289, %f287, %f288;
	mul.ftz.f32 	%f290, %f259, %f286;
	fma.rn.ftz.f32 	%f291, %f258, %f283, %f290;
	fma.rn.ftz.f32 	%f292, %f260, %f289, %f291;
	mul.ftz.f32 	%f293, %f292, %f273;
	add.s32 	%r28, %r1, 1;
	cvt.rn.f32.s32	%f294, %r28;
	sub.ftz.f32 	%f295, %f294, %f10;
	fma.rn.ftz.f32 	%f296, %f295, %f295, %f252;
	fma.rn.ftz.f32 	%f297, %f255, %f255, %f296;
	rsqrt.approx.ftz.f32 	%f298, %f297;
	mul.ftz.f32 	%f299, %f295, %f298;
	mul.ftz.f32 	%f300, %f251, %f298;
	mul.ftz.f32 	%f301, %f255, %f298;
	mul.ftz.f32 	%f302, %f300, %f9;
	mul.ftz.f32 	%f303, %f8, %f301;
	sub.ftz.f32 	%f304, %f302, %f303;
	mul.ftz.f32 	%f305, %f301, %f7;
	mul.ftz.f32 	%f306, %f9, %f299;
	sub.ftz.f32 	%f307, %f305, %f306;
	mul.ftz.f32 	%f308, %f299, %f8;
	mul.ftz.f32 	%f309, %f7, %f300;
	sub.ftz.f32 	%f310, %f308, %f309;
	mul.ftz.f32 	%f311, %f5, %f307;
	fma.rn.ftz.f32 	%f312, %f4, %f304, %f311;
	fma.rn.ftz.f32 	%f313, %f6, %f310, %f312;
	rcp.approx.ftz.f32 	%f314, %f313;
	mul.ftz.f32 	%f315, %f275, %f307;
	fma.rn.ftz.f32 	%f316, %f274, %f304, %f315;
	fma.rn.ftz.f32 	%f317, %f278, %f310, %f316;
	mul.ftz.f32 	%f318, %f317, %f314;
	mul.ftz.f32 	%f319, %f300, %f286;
	fma.rn.ftz.f32 	%f320, %f299, %f283, %f319;
	fma.rn.ftz.f32 	%f321, %f301, %f289, %f320;
	mul.ftz.f32 	%f322, %f321, %f314;
	add.s32 	%r29, %r2, 1;
	cvt.rn.f32.s32	%f323, %r29;
	sub.ftz.f32 	%f324, %f323, %f11;
	mul.ftz.f32 	%f325, %f324, %f324;
	fma.rn.ftz.f32 	%f326, %f295, %f295, %f325;
	fma.rn.ftz.f32 	%f327, %f255, %f255, %f326;
	rsqrt.approx.ftz.f32 	%f328, %f327;
	mul.ftz.f32 	%f329, %f295, %f328;
	mul.ftz.f32 	%f330, %f324, %f328;
	mul.ftz.f32 	%f331, %f255, %f328;
	mul.ftz.f32 	%f332, %f330, %f9;
	mul.ftz.f32 	%f333, %f8, %f331;
	sub.ftz.f32 	%f334, %f332, %f333;
	mul.ftz.f32 	%f335, %f331, %f7;
	mul.ftz.f32 	%f336, %f9, %f329;
	sub.ftz.f32 	%f337, %f335, %f336;
	mul.ftz.f32 	%f338, %f329, %f8;
	mul.ftz.f32 	%f339, %f7, %f330;
	sub.ftz.f32 	%f340, %f338, %f339;
	mul.ftz.f32 	%f341, %f5, %f337;
	fma.rn.ftz.f32 	%f342, %f4, %f334, %f341;
	fma.rn.ftz.f32 	%f343, %f6, %f340, %f342;
	rcp.approx.ftz.f32 	%f344, %f343;
	mul.ftz.f32 	%f345, %f275, %f337;
	fma.rn.ftz.f32 	%f346, %f274, %f334, %f345;
	fma.rn.ftz.f32 	%f347, %f278, %f340, %f346;
	mul.ftz.f32 	%f348, %f347, %f344;
	mul.ftz.f32 	%f349, %f330, %f286;
	fma.rn.ftz.f32 	%f350, %f329, %f283, %f349;
	fma.rn.ftz.f32 	%f351, %f331, %f289, %f350;
	mul.ftz.f32 	%f352, %f351, %f344;
	fma.rn.ftz.f32 	%f353, %f250, %f250, %f325;
	fma.rn.ftz.f32 	%f354, %f255, %f255, %f353;
	rsqrt.approx.ftz.f32 	%f355, %f354;
	mul.ftz.f32 	%f356, %f250, %f355;
	mul.ftz.f32 	%f357, %f324, %f355;
	mul.ftz.f32 	%f358, %f255, %f355;
	mul.ftz.f32 	%f359, %f357, %f9;
	mul.ftz.f32 	%f360, %f8, %f358;
	sub.ftz.f32 	%f361, %f359, %f360;
	mul.ftz.f32 	%f362, %f358, %f7;
	mul.ftz.f32 	%f363, %f9, %f356;
	sub.ftz.f32 	%f364, %f362, %f363;
	mul.ftz.f32 	%f365, %f356, %f8;
	mul.ftz.f32 	%f366, %f7, %f357;
	sub.ftz.f32 	%f367, %f365, %f366;
	mul.ftz.f32 	%f368, %f5, %f364;
	fma.rn.ftz.f32 	%f369, %f4, %f361, %f368;
	fma.rn.ftz.f32 	%f370, %f6, %f367, %f369;
	rcp.approx.ftz.f32 	%f371, %f370;
	mul.ftz.f32 	%f372, %f275, %f364;
	fma.rn.ftz.f32 	%f373, %f274, %f361, %f372;
	fma.rn.ftz.f32 	%f374, %f278, %f367, %f373;
	mul.ftz.f32 	%f375, %f374, %f371;
	mul.ftz.f32 	%f376, %f357, %f286;
	fma.rn.ftz.f32 	%f377, %f356, %f283, %f376;
	fma.rn.ftz.f32 	%f378, %f358, %f289, %f377;
	mul.ftz.f32 	%f379, %f378, %f371;
	mul.ftz.f32 	%f16, %f280, %f179;
	mul.ftz.f32 	%f17, %f293, %f180;
	mul.ftz.f32 	%f1156, %f318, %f179;
	mul.ftz.f32 	%f1157, %f322, %f180;
	mul.ftz.f32 	%f1154, %f348, %f179;
	mul.ftz.f32 	%f1155, %f352, %f180;
	mul.ftz.f32 	%f1152, %f375, %f179;
	mul.ftz.f32 	%f1153, %f379, %f180;
	add.ftz.f32 	%f380, %f16, %f1156;
	mul.ftz.f32 	%f30, %f380, 0f3F000000;
	add.ftz.f32 	%f381, %f17, %f1157;
	mul.ftz.f32 	%f31, %f381, 0f3F000000;
	add.ftz.f32 	%f382, %f1156, %f1154;
	mul.ftz.f32 	%f32, %f382, 0f3F000000;
	add.ftz.f32 	%f383, %f1157, %f1155;
	mul.ftz.f32 	%f33, %f383, 0f3F000000;
	add.ftz.f32 	%f384, %f1154, %f1152;
	mul.ftz.f32 	%f34, %f384, 0f3F000000;
	add.ftz.f32 	%f385, %f1155, %f1153;
	mul.ftz.f32 	%f35, %f385, 0f3F000000;
	add.ftz.f32 	%f386, %f1152, %f16;
	mul.ftz.f32 	%f36, %f386, 0f3F000000;
	add.ftz.f32 	%f387, %f1153, %f17;
	mul.ftz.f32 	%f37, %f387, 0f3F000000;
	add.ftz.f32 	%f388, %f30, %f34;
	mul.ftz.f32 	%f38, %f388, 0f3F000000;
	add.ftz.f32 	%f389, %f31, %f35;
	mul.ftz.f32 	%f39, %f389, 0f3F000000;
	sub.ftz.f32 	%f390, %f36, %f32;
	sub.ftz.f32 	%f391, %f37, %f33;
	mul.ftz.f32 	%f392, %f391, %f391;
	fma.rn.ftz.f32 	%f40, %f390, %f390, %f392;
	setp.gt.ftz.f32	%p5, %f40, 0f3F8020C5;
	selp.u32	%r30, 1, 0, %p5;
	sub.ftz.f32 	%f393, %f30, %f34;
	sub.ftz.f32 	%f394, %f31, %f35;
	mul.ftz.f32 	%f395, %f394, %f394;
	fma.rn.ftz.f32 	%f41, %f393, %f393, %f395;
	setp.gt.ftz.f32	%p6, %f41, 0f3F8020C5;
	or.b32  	%r31, %r30, 2;
	selp.b32	%r32, %r31, %r30, %p6;
	setp.eq.s32	%p7, %r32, 0;
	@%p7 bra 	BB2_27;

	setp.gt.ftz.f32	%p8, %f40, 0f40800000;
	setp.gt.ftz.f32	%p9, %f41, 0f40800000;
	or.pred  	%p10, %p8, %p9;
	@%p10 bra 	BB2_7;

	add.ftz.f32 	%f420, %f38, %f30;
	add.ftz.f32 	%f421, %f420, %f16;
	add.ftz.f32 	%f422, %f421, %f36;
	mul.ftz.f32 	%f400, %f422, 0f3E800000;
	add.ftz.f32 	%f423, %f39, %f31;
	add.ftz.f32 	%f424, %f423, %f17;
	add.ftz.f32 	%f425, %f424, %f37;
	mul.ftz.f32 	%f401, %f425, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f396, %f397, %f398, %f399}, [inTexture, {%f400, %f401}];
	// inline asm
	add.ftz.f32 	%f426, %f420, %f1156;
	add.ftz.f32 	%f427, %f426, %f32;
	mul.ftz.f32 	%f406, %f427, 0f3E800000;
	add.ftz.f32 	%f428, %f423, %f1157;
	add.ftz.f32 	%f429, %f428, %f33;
	mul.ftz.f32 	%f407, %f429, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f402, %f403, %f404, %f405}, [inTexture, {%f406, %f407}];
	// inline asm
	add.ftz.f32 	%f430, %f396, %f402;
	add.ftz.f32 	%f431, %f397, %f403;
	add.ftz.f32 	%f432, %f398, %f404;
	add.ftz.f32 	%f433, %f399, %f405;
	add.ftz.f32 	%f434, %f38, %f34;
	add.ftz.f32 	%f435, %f434, %f1154;
	add.ftz.f32 	%f436, %f435, %f32;
	mul.ftz.f32 	%f412, %f436, 0f3E800000;
	add.ftz.f32 	%f437, %f39, %f35;
	add.ftz.f32 	%f438, %f437, %f1155;
	add.ftz.f32 	%f439, %f438, %f33;
	mul.ftz.f32 	%f413, %f439, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f408, %f409, %f410, %f411}, [inTexture, {%f412, %f413}];
	// inline asm
	add.ftz.f32 	%f440, %f430, %f408;
	add.ftz.f32 	%f441, %f431, %f409;
	add.ftz.f32 	%f442, %f432, %f410;
	add.ftz.f32 	%f443, %f433, %f411;
	add.ftz.f32 	%f444, %f434, %f1152;
	add.ftz.f32 	%f445, %f444, %f36;
	mul.ftz.f32 	%f418, %f445, 0f3E800000;
	add.ftz.f32 	%f446, %f437, %f1153;
	add.ftz.f32 	%f447, %f446, %f37;
	mul.ftz.f32 	%f419, %f447, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f414, %f415, %f416, %f417}, [inTexture, {%f418, %f419}];
	// inline asm
	add.ftz.f32 	%f448, %f440, %f414;
	add.ftz.f32 	%f449, %f441, %f415;
	add.ftz.f32 	%f450, %f442, %f416;
	add.ftz.f32 	%f451, %f443, %f417;
	mul.ftz.f32 	%f1148, %f448, 0f3E800000;
	mul.ftz.f32 	%f1149, %f449, 0f3E800000;
	mul.ftz.f32 	%f1150, %f450, 0f3E800000;
	mul.ftz.f32 	%f1151, %f451, 0f3E800000;
	max.ftz.f32 	%f46, %f40, %f41;
	setp.geu.ftz.f32	%p11, %f46, 0f3FB851EC;
	@%p11 bra 	BB2_6;

	mov.f32 	%f1145, 0f3F800000;
	add.ftz.f32 	%f548, %f38, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f549, %f548;
	add.ftz.f32 	%f550, %f39, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f551, %f550;
	add.ftz.f32 	%f552, %f549, 0fBF000000;
	add.ftz.f32 	%f553, %f551, 0fBF000000;
	add.ftz.f32 	%f504, %f552, 0f00000000;
	add.ftz.f32 	%f475, %f553, 0f00000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f452, %f453, %f454, %f455}, [inTexture, {%f504, %f475}];
	// inline asm
	sub.ftz.f32 	%f554, %f550, %f551;
	sub.ftz.f32 	%f555, %f548, %f549;
	fma.rn.ftz.f32 	%f556, %f554, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f558, %f555, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f559, %f556, %f554, 0fBF000000;
	fma.rn.ftz.f32 	%f560, %f558, %f555, 0fBF000000;
	mul.ftz.f32 	%f561, %f559, %f554;
	mul.ftz.f32 	%f562, %f560, %f555;
	mul.ftz.f32 	%f563, %f562, %f561;
	add.ftz.f32 	%f534, %f552, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f458, %f459, %f460, %f461}, [inTexture, {%f534, %f475}];
	// inline asm
	fma.rn.ftz.f32 	%f564, %f555, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f565, %f564, %f555;
	fma.rn.ftz.f32 	%f566, %f565, %f555, 0f3F800000;
	mul.ftz.f32 	%f567, %f566, %f561;
	mul.ftz.f32 	%f568, %f458, %f567;
	mul.ftz.f32 	%f569, %f459, %f567;
	mul.ftz.f32 	%f570, %f460, %f567;
	mul.ftz.f32 	%f571, %f461, %f567;
	fma.rn.ftz.f32 	%f572, %f452, %f563, %f568;
	fma.rn.ftz.f32 	%f573, %f453, %f563, %f569;
	fma.rn.ftz.f32 	%f574, %f454, %f563, %f570;
	fma.rn.ftz.f32 	%f575, %f455, %f563, %f571;
	add.ftz.f32 	%f540, %f552, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f464, %f465, %f466, %f467}, [inTexture, {%f540, %f475}];
	// inline asm
	sub.ftz.f32 	%f576, %f1145, %f555;
	fma.rn.ftz.f32 	%f577, %f576, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f578, %f577, %f576;
	fma.rn.ftz.f32 	%f579, %f578, %f576, 0f3F800000;
	mul.ftz.f32 	%f580, %f579, %f561;
	fma.rn.ftz.f32 	%f581, %f464, %f580, %f572;
	fma.rn.ftz.f32 	%f582, %f465, %f580, %f573;
	fma.rn.ftz.f32 	%f583, %f466, %f580, %f574;
	fma.rn.ftz.f32 	%f584, %f467, %f580, %f575;
	add.ftz.f32 	%f546, %f552, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f470, %f471, %f472, %f473}, [inTexture, {%f546, %f475}];
	// inline asm
	fma.rn.ftz.f32 	%f585, %f576, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f586, %f585, %f576, 0fBF000000;
	mul.ftz.f32 	%f587, %f586, %f576;
	mul.ftz.f32 	%f588, %f587, %f561;
	fma.rn.ftz.f32 	%f589, %f470, %f588, %f581;
	fma.rn.ftz.f32 	%f590, %f471, %f588, %f582;
	fma.rn.ftz.f32 	%f591, %f472, %f588, %f583;
	fma.rn.ftz.f32 	%f592, %f473, %f588, %f584;
	add.ftz.f32 	%f499, %f553, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f476, %f477, %f478, %f479}, [inTexture, {%f504, %f499}];
	// inline asm
	fma.rn.ftz.f32 	%f593, %f554, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f594, %f593, %f554;
	fma.rn.ftz.f32 	%f595, %f594, %f554, 0f3F800000;
	mul.ftz.f32 	%f596, %f562, %f595;
	fma.rn.ftz.f32 	%f597, %f476, %f596, %f589;
	fma.rn.ftz.f32 	%f598, %f477, %f596, %f590;
	fma.rn.ftz.f32 	%f599, %f478, %f596, %f591;
	fma.rn.ftz.f32 	%f600, %f479, %f596, %f592;
	// inline asm
	tex.2d.v4.f32.f32 {%f482, %f483, %f484, %f485}, [inTexture, {%f534, %f499}];
	// inline asm
	mul.ftz.f32 	%f601, %f566, %f595;
	fma.rn.ftz.f32 	%f602, %f482, %f601, %f597;
	fma.rn.ftz.f32 	%f603, %f483, %f601, %f598;
	fma.rn.ftz.f32 	%f604, %f484, %f601, %f599;
	fma.rn.ftz.f32 	%f605, %f485, %f601, %f600;
	// inline asm
	tex.2d.v4.f32.f32 {%f488, %f489, %f490, %f491}, [inTexture, {%f540, %f499}];
	// inline asm
	mul.ftz.f32 	%f606, %f579, %f595;
	fma.rn.ftz.f32 	%f607, %f488, %f606, %f602;
	fma.rn.ftz.f32 	%f608, %f489, %f606, %f603;
	fma.rn.ftz.f32 	%f609, %f490, %f606, %f604;
	fma.rn.ftz.f32 	%f610, %f491, %f606, %f605;
	// inline asm
	tex.2d.v4.f32.f32 {%f494, %f495, %f496, %f497}, [inTexture, {%f546, %f499}];
	// inline asm
	mul.ftz.f32 	%f611, %f587, %f595;
	fma.rn.ftz.f32 	%f612, %f494, %f611, %f607;
	fma.rn.ftz.f32 	%f613, %f495, %f611, %f608;
	fma.rn.ftz.f32 	%f614, %f496, %f611, %f609;
	fma.rn.ftz.f32 	%f615, %f497, %f611, %f610;
	add.ftz.f32 	%f523, %f553, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f500, %f501, %f502, %f503}, [inTexture, {%f504, %f523}];
	// inline asm
	sub.ftz.f32 	%f616, %f1145, %f554;
	fma.rn.ftz.f32 	%f617, %f616, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f618, %f617, %f616;
	fma.rn.ftz.f32 	%f619, %f618, %f616, 0f3F800000;
	mul.ftz.f32 	%f620, %f562, %f619;
	fma.rn.ftz.f32 	%f621, %f500, %f620, %f612;
	fma.rn.ftz.f32 	%f622, %f501, %f620, %f613;
	fma.rn.ftz.f32 	%f623, %f502, %f620, %f614;
	fma.rn.ftz.f32 	%f624, %f503, %f620, %f615;
	// inline asm
	tex.2d.v4.f32.f32 {%f506, %f507, %f508, %f509}, [inTexture, {%f534, %f523}];
	// inline asm
	mul.ftz.f32 	%f625, %f566, %f619;
	fma.rn.ftz.f32 	%f626, %f506, %f625, %f621;
	fma.rn.ftz.f32 	%f627, %f507, %f625, %f622;
	fma.rn.ftz.f32 	%f628, %f508, %f625, %f623;
	fma.rn.ftz.f32 	%f629, %f509, %f625, %f624;
	// inline asm
	tex.2d.v4.f32.f32 {%f512, %f513, %f514, %f515}, [inTexture, {%f540, %f523}];
	// inline asm
	mul.ftz.f32 	%f630, %f579, %f619;
	fma.rn.ftz.f32 	%f631, %f512, %f630, %f626;
	fma.rn.ftz.f32 	%f632, %f513, %f630, %f627;
	fma.rn.ftz.f32 	%f633, %f514, %f630, %f628;
	fma.rn.ftz.f32 	%f634, %f515, %f630, %f629;
	// inline asm
	tex.2d.v4.f32.f32 {%f518, %f519, %f520, %f521}, [inTexture, {%f546, %f523}];
	// inline asm
	mul.ftz.f32 	%f635, %f587, %f619;
	fma.rn.ftz.f32 	%f636, %f518, %f635, %f631;
	fma.rn.ftz.f32 	%f637, %f519, %f635, %f632;
	fma.rn.ftz.f32 	%f638, %f520, %f635, %f633;
	fma.rn.ftz.f32 	%f639, %f521, %f635, %f634;
	add.ftz.f32 	%f528, %f552, 0f80000000;
	add.ftz.f32 	%f547, %f553, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f524, %f525, %f526, %f527}, [inTexture, {%f528, %f547}];
	// inline asm
	fma.rn.ftz.f32 	%f640, %f616, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f641, %f640, %f616, 0fBF000000;
	mul.ftz.f32 	%f642, %f641, %f616;
	mul.ftz.f32 	%f643, %f562, %f642;
	fma.rn.ftz.f32 	%f644, %f524, %f643, %f636;
	fma.rn.ftz.f32 	%f645, %f525, %f643, %f637;
	fma.rn.ftz.f32 	%f646, %f526, %f643, %f638;
	fma.rn.ftz.f32 	%f647, %f527, %f643, %f639;
	// inline asm
	tex.2d.v4.f32.f32 {%f530, %f531, %f532, %f533}, [inTexture, {%f534, %f547}];
	// inline asm
	mul.ftz.f32 	%f648, %f566, %f642;
	fma.rn.ftz.f32 	%f649, %f530, %f648, %f644;
	fma.rn.ftz.f32 	%f650, %f531, %f648, %f645;
	fma.rn.ftz.f32 	%f651, %f532, %f648, %f646;
	fma.rn.ftz.f32 	%f652, %f533, %f648, %f647;
	// inline asm
	tex.2d.v4.f32.f32 {%f536, %f537, %f538, %f539}, [inTexture, {%f540, %f547}];
	// inline asm
	mul.ftz.f32 	%f653, %f579, %f642;
	fma.rn.ftz.f32 	%f654, %f536, %f653, %f649;
	fma.rn.ftz.f32 	%f655, %f537, %f653, %f650;
	fma.rn.ftz.f32 	%f656, %f538, %f653, %f651;
	fma.rn.ftz.f32 	%f657, %f539, %f653, %f652;
	// inline asm
	tex.2d.v4.f32.f32 {%f542, %f543, %f544, %f545}, [inTexture, {%f546, %f547}];
	// inline asm
	add.ftz.f32 	%f658, %f46, 0fBF800000;
	mov.f32 	%f659, 0f3EE147B0;
	div.approx.ftz.f32 	%f660, %f658, %f659;
	mul.ftz.f32 	%f661, %f587, %f642;
	fma.rn.ftz.f32 	%f662, %f542, %f661, %f654;
	fma.rn.ftz.f32 	%f663, %f543, %f661, %f655;
	fma.rn.ftz.f32 	%f664, %f544, %f661, %f656;
	fma.rn.ftz.f32 	%f665, %f545, %f661, %f657;
	sub.ftz.f32 	%f666, %f1148, %f662;
	fma.rn.ftz.f32 	%f1148, %f660, %f666, %f662;
	sub.ftz.f32 	%f667, %f1149, %f663;
	fma.rn.ftz.f32 	%f1149, %f660, %f667, %f663;
	sub.ftz.f32 	%f668, %f1150, %f664;
	fma.rn.ftz.f32 	%f1150, %f660, %f668, %f664;
	sub.ftz.f32 	%f669, %f1151, %f665;
	fma.rn.ftz.f32 	%f1151, %f660, %f669, %f665;

BB2_6:
	mov.f32 	%f1168, %f1151;
	mov.f32 	%f1167, %f1150;
	mov.f32 	%f1166, %f1149;
	mov.f32 	%f1165, %f1148;
	bra.uni 	BB2_28;

BB2_7:
	mov.f32 	%f1164, 0f00000000;
	add.u64 	%rd49, %SP, 32;
	add.u64 	%rd50, %SP, 16;
	add.u64 	%rd51, %SP, 0;
	cvta.to.local.u64 	%rd52, %rd49;
	cvta.to.local.u64 	%rd53, %rd50;
	cvta.to.local.u64 	%rd54, %rd51;
	mov.u32 	%r73, 0;
	st.local.u32 	[%rd53], %r73;
	mov.u32 	%r35, 1;
	st.local.u32 	[%rd54], %r35;
	st.local.v2.f32 	[%rd52], {%f16, %f17};
	st.local.v2.f32 	[%rd52+8], {%f1156, %f1157};
	st.local.f32 	[%rd52+16], %f1154;
	st.local.f32 	[%rd52+20], %f1155;
	st.local.v2.f32 	[%rd52+24], {%f1152, %f1153};
	mov.f32 	%f1159, %f17;
	mov.f32 	%f1158, %f16;
	mov.u32 	%r72, %r73;
	mov.f32 	%f1163, %f1164;
	mov.f32 	%f1162, %f1164;
	mov.f32 	%f1161, %f1164;
	mov.f32 	%f1160, %f1164;

BB2_8:
	add.ftz.f32 	%f76, %f1158, %f1156;
	mul.ftz.f32 	%f77, %f76, 0f3F000000;
	add.ftz.f32 	%f80, %f1159, %f1157;
	mul.ftz.f32 	%f81, %f80, 0f3F000000;
	add.ftz.f32 	%f675, %f1156, %f1154;
	mul.ftz.f32 	%f83, %f675, 0f3F000000;
	add.ftz.f32 	%f676, %f1157, %f1155;
	mul.ftz.f32 	%f85, %f676, 0f3F000000;
	add.ftz.f32 	%f677, %f1154, %f1152;
	mul.ftz.f32 	%f87, %f677, 0f3F000000;
	add.ftz.f32 	%f678, %f1155, %f1153;
	mul.ftz.f32 	%f89, %f678, 0f3F000000;
	add.ftz.f32 	%f679, %f1152, %f1158;
	mul.ftz.f32 	%f90, %f679, 0f3F000000;
	add.ftz.f32 	%f680, %f1153, %f1159;
	mul.ftz.f32 	%f91, %f680, 0f3F000000;
	fma.rn.ftz.f32 	%f681, %f76, 0f3F000000, %f87;
	mul.ftz.f32 	%f92, %f681, 0f3F000000;
	fma.rn.ftz.f32 	%f682, %f80, 0f3F000000, %f89;
	mul.ftz.f32 	%f93, %f682, 0f3F000000;
	sub.ftz.f32 	%f683, %f90, %f83;
	sub.ftz.f32 	%f684, %f91, %f85;
	mul.ftz.f32 	%f685, %f684, %f684;
	fma.rn.ftz.f32 	%f94, %f683, %f683, %f685;
	setp.gt.ftz.f32	%p12, %f94, 0f3F800000;
	selp.u32	%r36, 1, 0, %p12;
	sub.ftz.f32 	%f686, %f87, %f77;
	sub.ftz.f32 	%f687, %f89, %f81;
	mul.ftz.f32 	%f688, %f687, %f687;
	fma.rn.ftz.f32 	%f95, %f686, %f686, %f688;
	setp.gt.ftz.f32	%p13, %f95, 0f3F800000;
	or.b32  	%r37, %r36, 2;
	selp.b32	%r5, %r37, %r36, %p13;
	setp.eq.s32	%p14, %r5, 0;
	cvt.s64.s32	%rd3, %r73;
	@%p14 bra 	BB2_23;

	shl.b64 	%rd57, %rd3, 2;
	add.s64 	%rd58, %rd53, %rd57;
	ld.local.u32 	%r6, [%rd58];
	setp.eq.s32	%p15, %r6, 2;
	@%p15 bra 	BB2_17;

	setp.lt.ftz.f32	%p16, %f94, 0f40800000;
	setp.lt.ftz.f32	%p17, %f95, 0f40800000;
	and.pred  	%p18, %p16, %p17;
	@%p18 bra 	BB2_17;

	setp.eq.s32	%p19, %r5, 3;
	mul.wide.s32 	%rd61, %r72, 32;
	add.s64 	%rd4, %rd52, %rd61;
	@%p19 bra 	BB2_15;

	setp.eq.s32	%p20, %r5, 2;
	st.local.v2.f32 	[%rd4], {%f1158, %f1159};
	@%p20 bra 	BB2_14;

	st.local.v2.f32 	[%rd4+8], {%f77, %f81};
	st.local.v2.f32 	[%rd4+16], {%f87, %f89};
	st.local.v2.f32 	[%rd4+24], {%f1152, %f1153};
	st.local.v2.f32 	[%rd4+32], {%f77, %f81};
	st.local.f32 	[%rd4+40], %f1156;
	st.local.f32 	[%rd4+44], %f1157;
	st.local.v2.f32 	[%rd4+48], {%f1154, %f1155};
	st.local.f32 	[%rd4+56], %f87;
	st.local.f32 	[%rd4+60], %f89;
	mov.u32 	%r74, 2;
	bra.uni 	BB2_16;

BB2_14:
	st.local.v2.f32 	[%rd4+8], {%f1156, %f1157};
	st.local.v2.f32 	[%rd4+16], {%f83, %f85};
	st.local.f32 	[%rd4+24], %f90;
	st.local.f32 	[%rd4+28], %f91;
	st.local.v2.f32 	[%rd4+32], {%f90, %f91};
	st.local.f32 	[%rd4+40], %f83;
	st.local.f32 	[%rd4+44], %f85;
	st.local.v2.f32 	[%rd4+48], {%f1154, %f1155};
	st.local.f32 	[%rd4+56], %f1152;
	st.local.f32 	[%rd4+60], %f1153;
	mov.u32 	%r74, 2;
	bra.uni 	BB2_16;

BB2_15:
	st.local.v2.f32 	[%rd4], {%f92, %f93};
	st.local.v2.f32 	[%rd4+8], {%f90, %f91};
	st.local.v2.f32 	[%rd4+16], {%f1158, %f1159};
	st.local.v2.f32 	[%rd4+24], {%f77, %f81};
	st.local.f32 	[%rd4+32], %f92;
	st.local.f32 	[%rd4+36], %f93;
	st.local.v2.f32 	[%rd4+40], {%f77, %f81};
	st.local.f32 	[%rd4+48], %f1156;
	st.local.f32 	[%rd4+52], %f1157;
	st.local.v2.f32 	[%rd4+56], {%f83, %f85};
	st.local.f32 	[%rd4+64], %f92;
	st.local.f32 	[%rd4+68], %f93;
	st.local.v2.f32 	[%rd4+72], {%f83, %f85};
	st.local.f32 	[%rd4+80], %f1154;
	st.local.f32 	[%rd4+84], %f1155;
	st.local.v2.f32 	[%rd4+88], {%f87, %f89};
	st.local.f32 	[%rd4+96], %f92;
	st.local.f32 	[%rd4+100], %f93;
	st.local.v2.f32 	[%rd4+104], {%f87, %f89};
	st.local.f32 	[%rd4+112], %f1152;
	st.local.f32 	[%rd4+116], %f1153;
	st.local.v2.f32 	[%rd4+120], {%f90, %f91};
	mov.u32 	%r74, 4;

BB2_16:
	add.s32 	%r41, %r72, %r74;
	add.s32 	%r72, %r41, -1;
	mul.wide.s32 	%rd64, %r73, 4;
	add.s64 	%rd65, %rd54, %rd64;
	ld.local.u32 	%r42, [%rd65];
	add.s32 	%r43, %r42, -1;
	st.local.u32 	[%rd65], %r43;
	setp.ne.s32	%p21, %r43, 0;
	selp.u32	%r44, 1, 0, %p21;
	add.s32 	%r73, %r44, %r73;
	mul.wide.s32 	%rd68, %r73, 4;
	add.s64 	%rd69, %rd53, %rd68;
	add.s32 	%r45, %r6, 1;
	st.local.u32 	[%rd69], %r45;
	add.s64 	%rd70, %rd54, %rd68;
	st.local.u32 	[%rd70], %r74;
	bra.uni 	BB2_24;

BB2_17:
	add.s32 	%r72, %r72, -1;
	setp.eq.s32	%p22, %r5, 3;
	@%p22 bra 	BB2_21;

	setp.eq.s32	%p23, %r5, 2;
	@%p23 bra 	BB2_20;

	add.ftz.f32 	%f701, %f1158, %f77;
	add.ftz.f32 	%f702, %f701, %f87;
	add.ftz.f32 	%f703, %f702, %f1152;
	mul.ftz.f32 	%f693, %f703, 0f3E800000;
	add.ftz.f32 	%f704, %f1159, %f81;
	add.ftz.f32 	%f705, %f704, %f89;
	add.ftz.f32 	%f706, %f705, %f1153;
	mul.ftz.f32 	%f694, %f706, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f689, %f690, %f691, %f692}, [inTexture, {%f693, %f694}];
	// inline asm
	add.ftz.f32 	%f707, %f1161, %f689;
	add.ftz.f32 	%f708, %f1160, %f690;
	add.ftz.f32 	%f709, %f1162, %f691;
	add.ftz.f32 	%f710, %f1163, %f692;
	add.ftz.f32 	%f711, %f77, %f1156;
	add.ftz.f32 	%f712, %f711, %f1154;
	add.ftz.f32 	%f713, %f712, %f87;
	mul.ftz.f32 	%f699, %f713, 0f3E800000;
	add.ftz.f32 	%f714, %f81, %f1157;
	add.ftz.f32 	%f715, %f714, %f1155;
	add.ftz.f32 	%f716, %f715, %f89;
	mul.ftz.f32 	%f700, %f716, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f695, %f696, %f697, %f698}, [inTexture, {%f699, %f700}];
	// inline asm
	add.ftz.f32 	%f1161, %f707, %f695;
	add.ftz.f32 	%f1160, %f708, %f696;
	add.ftz.f32 	%f1162, %f709, %f697;
	add.ftz.f32 	%f1163, %f710, %f698;
	add.ftz.f32 	%f1164, %f1164, 0f40000000;
	bra.uni 	BB2_22;

BB2_20:
	add.ftz.f32 	%f1144, %f1159, %f1157;
	add.ftz.f32 	%f1143, %f1158, %f1156;
	add.ftz.f32 	%f729, %f1143, %f83;
	add.ftz.f32 	%f730, %f729, %f90;
	mul.ftz.f32 	%f721, %f730, 0f3E800000;
	add.ftz.f32 	%f731, %f1144, %f85;
	add.ftz.f32 	%f732, %f731, %f91;
	mul.ftz.f32 	%f722, %f732, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f717, %f718, %f719, %f720}, [inTexture, {%f721, %f722}];
	// inline asm
	add.ftz.f32 	%f733, %f1161, %f717;
	add.ftz.f32 	%f734, %f1160, %f718;
	add.ftz.f32 	%f735, %f1162, %f719;
	add.ftz.f32 	%f736, %f1163, %f720;
	add.ftz.f32 	%f737, %f90, %f83;
	add.ftz.f32 	%f738, %f737, %f1154;
	add.ftz.f32 	%f739, %f738, %f1152;
	mul.ftz.f32 	%f727, %f739, 0f3E800000;
	add.ftz.f32 	%f740, %f91, %f85;
	add.ftz.f32 	%f741, %f740, %f1155;
	add.ftz.f32 	%f742, %f741, %f1153;
	mul.ftz.f32 	%f728, %f742, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f723, %f724, %f725, %f726}, [inTexture, {%f727, %f728}];
	// inline asm
	add.ftz.f32 	%f1161, %f733, %f723;
	add.ftz.f32 	%f1160, %f734, %f724;
	add.ftz.f32 	%f1162, %f735, %f725;
	add.ftz.f32 	%f1163, %f736, %f726;
	add.ftz.f32 	%f1164, %f1164, 0f40000000;
	bra.uni 	BB2_22;

BB2_21:
	add.ftz.f32 	%f1164, %f1164, 0f40800000;
	add.ftz.f32 	%f767, %f92, %f90;
	add.ftz.f32 	%f768, %f767, %f1158;
	add.ftz.f32 	%f769, %f768, %f77;
	mul.ftz.f32 	%f747, %f769, 0f3E800000;
	add.ftz.f32 	%f770, %f93, %f91;
	add.ftz.f32 	%f771, %f770, %f1159;
	add.ftz.f32 	%f772, %f771, %f81;
	mul.ftz.f32 	%f748, %f772, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f743, %f744, %f745, %f746}, [inTexture, {%f747, %f748}];
	// inline asm
	add.ftz.f32 	%f773, %f1161, %f743;
	add.ftz.f32 	%f774, %f1160, %f744;
	add.ftz.f32 	%f775, %f1162, %f745;
	add.ftz.f32 	%f776, %f1163, %f746;
	add.ftz.f32 	%f777, %f92, %f77;
	add.ftz.f32 	%f778, %f777, %f1156;
	add.ftz.f32 	%f779, %f778, %f83;
	mul.ftz.f32 	%f753, %f779, 0f3E800000;
	add.ftz.f32 	%f780, %f93, %f81;
	add.ftz.f32 	%f781, %f780, %f1157;
	add.ftz.f32 	%f782, %f781, %f85;
	mul.ftz.f32 	%f754, %f782, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f749, %f750, %f751, %f752}, [inTexture, {%f753, %f754}];
	// inline asm
	add.ftz.f32 	%f783, %f773, %f749;
	add.ftz.f32 	%f784, %f774, %f750;
	add.ftz.f32 	%f785, %f775, %f751;
	add.ftz.f32 	%f786, %f776, %f752;
	add.ftz.f32 	%f787, %f92, %f83;
	add.ftz.f32 	%f788, %f787, %f1154;
	add.ftz.f32 	%f789, %f788, %f87;
	mul.ftz.f32 	%f759, %f789, 0f3E800000;
	add.ftz.f32 	%f790, %f93, %f85;
	add.ftz.f32 	%f791, %f790, %f1155;
	add.ftz.f32 	%f792, %f791, %f89;
	mul.ftz.f32 	%f760, %f792, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f755, %f756, %f757, %f758}, [inTexture, {%f759, %f760}];
	// inline asm
	add.ftz.f32 	%f793, %f783, %f755;
	add.ftz.f32 	%f794, %f784, %f756;
	add.ftz.f32 	%f795, %f785, %f757;
	add.ftz.f32 	%f796, %f786, %f758;
	add.ftz.f32 	%f797, %f92, %f87;
	add.ftz.f32 	%f798, %f797, %f1152;
	add.ftz.f32 	%f799, %f798, %f90;
	mul.ftz.f32 	%f765, %f799, 0f3E800000;
	add.ftz.f32 	%f800, %f93, %f89;
	add.ftz.f32 	%f801, %f800, %f1153;
	add.ftz.f32 	%f802, %f801, %f91;
	mul.ftz.f32 	%f766, %f802, 0f3E800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f761, %f762, %f763, %f764}, [inTexture, {%f765, %f766}];
	// inline asm
	add.ftz.f32 	%f1161, %f793, %f761;
	add.ftz.f32 	%f1160, %f794, %f762;
	add.ftz.f32 	%f1162, %f795, %f763;
	add.ftz.f32 	%f1163, %f796, %f764;

BB2_22:
	mul.wide.s32 	%rd89, %r73, 4;
	add.s64 	%rd90, %rd54, %rd89;
	ld.local.u32 	%r46, [%rd90];
	add.s32 	%r47, %r46, -1;
	st.local.u32 	[%rd90], %r47;
	setp.eq.s32	%p24, %r47, 0;
	selp.b32	%r48, -1, 0, %p24;
	add.s32 	%r73, %r48, %r73;
	bra.uni 	BB2_24;

BB2_23:
	add.s32 	%r72, %r72, -1;
	add.ftz.f32 	%f1164, %f1164, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f803, %f804, %f805, %f806}, [inTexture, {%f92, %f93}];
	// inline asm
	add.ftz.f32 	%f1161, %f1161, %f803;
	add.ftz.f32 	%f1160, %f1160, %f804;
	add.ftz.f32 	%f1162, %f1162, %f805;
	add.ftz.f32 	%f1163, %f1163, %f806;
	mul.wide.s32 	%rd94, %r73, 4;
	add.s64 	%rd95, %rd54, %rd94;
	ld.local.u32 	%r49, [%rd95];
	add.s32 	%r50, %r49, -1;
	st.local.u32 	[%rd95], %r50;
	setp.eq.s32	%p25, %r50, 0;
	selp.b32	%r51, -1, 0, %p25;
	add.s32 	%r73, %r51, %r73;

BB2_24:
	setp.gt.s32	%p26, %r72, -1;
	@%p26 bra 	BB2_26;

	mov.f32 	%f1129, 0f3F800000;
	div.approx.ftz.f32 	%f810, %f1129, %f1164;
	mul.ftz.f32 	%f1165, %f1161, %f810;
	mul.ftz.f32 	%f1166, %f1160, %f810;
	mul.ftz.f32 	%f1167, %f1162, %f810;
	mul.ftz.f32 	%f1168, %f1163, %f810;
	bra.uni 	BB2_28;

BB2_26:
	mul.wide.s32 	%rd98, %r72, 32;
	add.s64 	%rd99, %rd52, %rd98;
	ld.local.v2.f32 	{%f811, %f812}, [%rd99];
	ld.local.v2.f32 	{%f813, %f814}, [%rd99+8];
	ld.local.v2.f32 	{%f815, %f816}, [%rd99+16];
	ld.local.v2.f32 	{%f817, %f818}, [%rd99+24];
	mov.f32 	%f1153, %f818;
	mov.f32 	%f1152, %f817;
	mov.f32 	%f1155, %f816;
	mov.f32 	%f1154, %f815;
	mov.f32 	%f1157, %f814;
	mov.f32 	%f1156, %f813;
	mov.f32 	%f1159, %f812;
	mov.f32 	%f1158, %f811;
	bra.uni 	BB2_8;

BB2_27:
	mov.f32 	%f1147, 0f3F800000;
	add.ftz.f32 	%f915, %f38, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f916, %f915;
	add.ftz.f32 	%f917, %f39, 0fBF000000;
	cvt.rmi.ftz.f32.f32	%f918, %f917;
	add.ftz.f32 	%f919, %f916, 0fBF000000;
	add.ftz.f32 	%f920, %f918, 0fBF000000;
	add.ftz.f32 	%f871, %f919, 0f00000000;
	add.ftz.f32 	%f842, %f920, 0f00000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f819, %f820, %f821, %f822}, [inTexture, {%f871, %f842}];
	// inline asm
	sub.ftz.f32 	%f921, %f917, %f918;
	sub.ftz.f32 	%f922, %f915, %f916;
	fma.rn.ftz.f32 	%f923, %f921, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f925, %f922, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f926, %f923, %f921, 0fBF000000;
	fma.rn.ftz.f32 	%f927, %f925, %f922, 0fBF000000;
	mul.ftz.f32 	%f928, %f926, %f921;
	mul.ftz.f32 	%f929, %f927, %f922;
	mul.ftz.f32 	%f930, %f929, %f928;
	add.ftz.f32 	%f901, %f919, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f825, %f826, %f827, %f828}, [inTexture, {%f901, %f842}];
	// inline asm
	fma.rn.ftz.f32 	%f931, %f922, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f932, %f931, %f922;
	fma.rn.ftz.f32 	%f933, %f932, %f922, 0f3F800000;
	mul.ftz.f32 	%f934, %f933, %f928;
	mul.ftz.f32 	%f935, %f825, %f934;
	mul.ftz.f32 	%f936, %f826, %f934;
	mul.ftz.f32 	%f937, %f827, %f934;
	mul.ftz.f32 	%f938, %f828, %f934;
	fma.rn.ftz.f32 	%f939, %f819, %f930, %f935;
	fma.rn.ftz.f32 	%f940, %f820, %f930, %f936;
	fma.rn.ftz.f32 	%f941, %f821, %f930, %f937;
	fma.rn.ftz.f32 	%f942, %f822, %f930, %f938;
	add.ftz.f32 	%f907, %f919, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f831, %f832, %f833, %f834}, [inTexture, {%f907, %f842}];
	// inline asm
	sub.ftz.f32 	%f943, %f1147, %f922;
	fma.rn.ftz.f32 	%f944, %f943, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f945, %f944, %f943;
	fma.rn.ftz.f32 	%f946, %f945, %f943, 0f3F800000;
	mul.ftz.f32 	%f947, %f946, %f928;
	fma.rn.ftz.f32 	%f948, %f831, %f947, %f939;
	fma.rn.ftz.f32 	%f949, %f832, %f947, %f940;
	fma.rn.ftz.f32 	%f950, %f833, %f947, %f941;
	fma.rn.ftz.f32 	%f951, %f834, %f947, %f942;
	add.ftz.f32 	%f913, %f919, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f837, %f838, %f839, %f840}, [inTexture, {%f913, %f842}];
	// inline asm
	fma.rn.ftz.f32 	%f952, %f943, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f953, %f952, %f943, 0fBF000000;
	mul.ftz.f32 	%f954, %f953, %f943;
	mul.ftz.f32 	%f955, %f954, %f928;
	fma.rn.ftz.f32 	%f956, %f837, %f955, %f948;
	fma.rn.ftz.f32 	%f957, %f838, %f955, %f949;
	fma.rn.ftz.f32 	%f958, %f839, %f955, %f950;
	fma.rn.ftz.f32 	%f959, %f840, %f955, %f951;
	add.ftz.f32 	%f866, %f920, 0f3F800000;
	// inline asm
	tex.2d.v4.f32.f32 {%f843, %f844, %f845, %f846}, [inTexture, {%f871, %f866}];
	// inline asm
	fma.rn.ftz.f32 	%f960, %f921, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f961, %f960, %f921;
	fma.rn.ftz.f32 	%f962, %f961, %f921, 0f3F800000;
	mul.ftz.f32 	%f963, %f929, %f962;
	fma.rn.ftz.f32 	%f964, %f843, %f963, %f956;
	fma.rn.ftz.f32 	%f965, %f844, %f963, %f957;
	fma.rn.ftz.f32 	%f966, %f845, %f963, %f958;
	fma.rn.ftz.f32 	%f967, %f846, %f963, %f959;
	// inline asm
	tex.2d.v4.f32.f32 {%f849, %f850, %f851, %f852}, [inTexture, {%f901, %f866}];
	// inline asm
	mul.ftz.f32 	%f968, %f933, %f962;
	fma.rn.ftz.f32 	%f969, %f849, %f968, %f964;
	fma.rn.ftz.f32 	%f970, %f850, %f968, %f965;
	fma.rn.ftz.f32 	%f971, %f851, %f968, %f966;
	fma.rn.ftz.f32 	%f972, %f852, %f968, %f967;
	// inline asm
	tex.2d.v4.f32.f32 {%f855, %f856, %f857, %f858}, [inTexture, {%f907, %f866}];
	// inline asm
	mul.ftz.f32 	%f973, %f946, %f962;
	fma.rn.ftz.f32 	%f974, %f855, %f973, %f969;
	fma.rn.ftz.f32 	%f975, %f856, %f973, %f970;
	fma.rn.ftz.f32 	%f976, %f857, %f973, %f971;
	fma.rn.ftz.f32 	%f977, %f858, %f973, %f972;
	// inline asm
	tex.2d.v4.f32.f32 {%f861, %f862, %f863, %f864}, [inTexture, {%f913, %f866}];
	// inline asm
	mul.ftz.f32 	%f978, %f954, %f962;
	fma.rn.ftz.f32 	%f979, %f861, %f978, %f974;
	fma.rn.ftz.f32 	%f980, %f862, %f978, %f975;
	fma.rn.ftz.f32 	%f981, %f863, %f978, %f976;
	fma.rn.ftz.f32 	%f982, %f864, %f978, %f977;
	add.ftz.f32 	%f890, %f920, 0f40000000;
	// inline asm
	tex.2d.v4.f32.f32 {%f867, %f868, %f869, %f870}, [inTexture, {%f871, %f890}];
	// inline asm
	sub.ftz.f32 	%f983, %f1147, %f921;
	fma.rn.ftz.f32 	%f984, %f983, 0f3FC00000, 0fC0200000;
	mul.ftz.f32 	%f985, %f984, %f983;
	fma.rn.ftz.f32 	%f986, %f985, %f983, 0f3F800000;
	mul.ftz.f32 	%f987, %f929, %f986;
	fma.rn.ftz.f32 	%f988, %f867, %f987, %f979;
	fma.rn.ftz.f32 	%f989, %f868, %f987, %f980;
	fma.rn.ftz.f32 	%f990, %f869, %f987, %f981;
	fma.rn.ftz.f32 	%f991, %f870, %f987, %f982;
	// inline asm
	tex.2d.v4.f32.f32 {%f873, %f874, %f875, %f876}, [inTexture, {%f901, %f890}];
	// inline asm
	mul.ftz.f32 	%f992, %f933, %f986;
	fma.rn.ftz.f32 	%f993, %f873, %f992, %f988;
	fma.rn.ftz.f32 	%f994, %f874, %f992, %f989;
	fma.rn.ftz.f32 	%f995, %f875, %f992, %f990;
	fma.rn.ftz.f32 	%f996, %f876, %f992, %f991;
	// inline asm
	tex.2d.v4.f32.f32 {%f879, %f880, %f881, %f882}, [inTexture, {%f907, %f890}];
	// inline asm
	mul.ftz.f32 	%f997, %f946, %f986;
	fma.rn.ftz.f32 	%f998, %f879, %f997, %f993;
	fma.rn.ftz.f32 	%f999, %f880, %f997, %f994;
	fma.rn.ftz.f32 	%f1000, %f881, %f997, %f995;
	fma.rn.ftz.f32 	%f1001, %f882, %f997, %f996;
	// inline asm
	tex.2d.v4.f32.f32 {%f885, %f886, %f887, %f888}, [inTexture, {%f913, %f890}];
	// inline asm
	mul.ftz.f32 	%f1002, %f954, %f986;
	fma.rn.ftz.f32 	%f1003, %f885, %f1002, %f998;
	fma.rn.ftz.f32 	%f1004, %f886, %f1002, %f999;
	fma.rn.ftz.f32 	%f1005, %f887, %f1002, %f1000;
	fma.rn.ftz.f32 	%f1006, %f888, %f1002, %f1001;
	add.ftz.f32 	%f895, %f919, 0f80000000;
	add.ftz.f32 	%f914, %f920, 0f40400000;
	// inline asm
	tex.2d.v4.f32.f32 {%f891, %f892, %f893, %f894}, [inTexture, {%f895, %f914}];
	// inline asm
	fma.rn.ftz.f32 	%f1007, %f983, 0fBF000000, 0f3F800000;
	fma.rn.ftz.f32 	%f1008, %f1007, %f983, 0fBF000000;
	mul.ftz.f32 	%f1009, %f1008, %f983;
	mul.ftz.f32 	%f1010, %f929, %f1009;
	fma.rn.ftz.f32 	%f1011, %f891, %f1010, %f1003;
	fma.rn.ftz.f32 	%f1012, %f892, %f1010, %f1004;
	fma.rn.ftz.f32 	%f1013, %f893, %f1010, %f1005;
	fma.rn.ftz.f32 	%f1014, %f894, %f1010, %f1006;
	// inline asm
	tex.2d.v4.f32.f32 {%f897, %f898, %f899, %f900}, [inTexture, {%f901, %f914}];
	// inline asm
	mul.ftz.f32 	%f1015, %f933, %f1009;
	fma.rn.ftz.f32 	%f1016, %f897, %f1015, %f1011;
	fma.rn.ftz.f32 	%f1017, %f898, %f1015, %f1012;
	fma.rn.ftz.f32 	%f1018, %f899, %f1015, %f1013;
	fma.rn.ftz.f32 	%f1019, %f900, %f1015, %f1014;
	// inline asm
	tex.2d.v4.f32.f32 {%f903, %f904, %f905, %f906}, [inTexture, {%f907, %f914}];
	// inline asm
	mul.ftz.f32 	%f1020, %f946, %f1009;
	fma.rn.ftz.f32 	%f1021, %f903, %f1020, %f1016;
	fma.rn.ftz.f32 	%f1022, %f904, %f1020, %f1017;
	fma.rn.ftz.f32 	%f1023, %f905, %f1020, %f1018;
	fma.rn.ftz.f32 	%f1024, %f906, %f1020, %f1019;
	// inline asm
	tex.2d.v4.f32.f32 {%f909, %f910, %f911, %f912}, [inTexture, {%f913, %f914}];
	// inline asm
	mul.ftz.f32 	%f1025, %f954, %f1009;
	fma.rn.ftz.f32 	%f1165, %f909, %f1025, %f1021;
	fma.rn.ftz.f32 	%f1166, %f910, %f1025, %f1022;
	fma.rn.ftz.f32 	%f1167, %f911, %f1025, %f1023;
	fma.rn.ftz.f32 	%f1168, %f912, %f1025, %f1024;

BB2_28:
	ld.param.u32 	%r70, [cuda_kernel_renderquad_param_20];
	mul.ftz.f32 	%f1170, %f1165, %f15;
	mul.ftz.f32 	%f1171, %f1166, %f15;
	mul.ftz.f32 	%f1172, %f1167, %f15;
	mul.ftz.f32 	%f152, %f1168, %f15;
	setp.eq.s32	%p27, %r70, 0;
	@%p27 bra 	BB2_33;

	ld.param.f32 	%f1141, [cuda_kernel_renderquad_param_8+8];
	ld.param.f32 	%f1140, [cuda_kernel_renderquad_param_8+4];
	ld.param.f32 	%f1139, [cuda_kernel_renderquad_param_8];
	ld.param.f32 	%f1138, [cuda_kernel_renderquad_param_9+8];
	ld.param.f32 	%f1137, [cuda_kernel_renderquad_param_9];
	ld.param.f32 	%f1136, [cuda_kernel_renderquad_param_9+4];
	ld.param.f32 	%f1135, [cuda_kernel_renderquad_param_10];
	ld.param.f32 	%f1134, [cuda_kernel_renderquad_param_10+4];
	ld.param.f32 	%f1133, [cuda_kernel_renderquad_param_10+8];
	ld.param.f32 	%f1132, [cuda_kernel_renderquad_param_11+8];
	ld.param.f32 	%f1131, [cuda_kernel_renderquad_param_11+4];
	ld.param.f32 	%f1130, [cuda_kernel_renderquad_param_11];
	mul.ftz.f32 	%f1026, %f1136, %f1136;
	fma.rn.ftz.f32 	%f1027, %f1137, %f1137, %f1026;
	fma.rn.ftz.f32 	%f1028, %f1138, %f1138, %f1027;
	rsqrt.approx.ftz.f32 	%f1029, %f1028;
	mul.ftz.f32 	%f1030, %f1137, %f1029;
	mul.ftz.f32 	%f1031, %f1136, %f1029;
	mul.ftz.f32 	%f1032, %f1138, %f1029;
	mul.ftz.f32 	%f1033, %f1134, %f1134;
	fma.rn.ftz.f32 	%f1034, %f1135, %f1135, %f1033;
	fma.rn.ftz.f32 	%f1035, %f1133, %f1133, %f1034;
	rsqrt.approx.ftz.f32 	%f1036, %f1035;
	mul.ftz.f32 	%f1037, %f1135, %f1036;
	mul.ftz.f32 	%f1038, %f1134, %f1036;
	mul.ftz.f32 	%f1039, %f1133, %f1036;
	mul.ftz.f32 	%f1040, %f1031, %f1039;
	mul.ftz.f32 	%f1041, %f1038, %f1032;
	sub.ftz.f32 	%f1042, %f1040, %f1041;
	mul.ftz.f32 	%f1043, %f1032, %f1037;
	mul.ftz.f32 	%f1044, %f1039, %f1030;
	sub.ftz.f32 	%f1045, %f1043, %f1044;
	mul.ftz.f32 	%f1046, %f1030, %f1038;
	mul.ftz.f32 	%f1047, %f1037, %f1031;
	sub.ftz.f32 	%f1048, %f1046, %f1047;
	fma.rn.ftz.f32 	%f1049, %f1030, %f16, %f1139;
	fma.rn.ftz.f32 	%f1050, %f1037, %f17, %f1049;
	fma.rn.ftz.f32 	%f1051, %f1031, %f16, %f1140;
	fma.rn.ftz.f32 	%f1052, %f1038, %f17, %f1051;
	fma.rn.ftz.f32 	%f1053, %f1032, %f16, %f1141;
	fma.rn.ftz.f32 	%f1054, %f1039, %f17, %f1053;
	add.ftz.f32 	%f1055, %f1050, 0f42C80000;
	add.ftz.f32 	%f1056, %f1052, 0f42C80000;
	add.ftz.f32 	%f1057, %f1054, 0f42C80000;
	sub.ftz.f32 	%f1058, %f1050, %f1130;
	sub.ftz.f32 	%f1059, %f1052, %f1131;
	sub.ftz.f32 	%f1060, %f1054, %f1132;
	mul.ftz.f32 	%f1061, %f1056, %f1056;
	fma.rn.ftz.f32 	%f1062, %f1055, %f1055, %f1061;
	fma.rn.ftz.f32 	%f1063, %f1057, %f1057, %f1062;
	rsqrt.approx.ftz.f32 	%f1064, %f1063;
	mul.ftz.f32 	%f1065, %f1055, %f1064;
	mul.ftz.f32 	%f1066, %f1056, %f1064;
	mul.ftz.f32 	%f1067, %f1057, %f1064;
	mul.ftz.f32 	%f1068, %f1059, %f1059;
	fma.rn.ftz.f32 	%f1069, %f1058, %f1058, %f1068;
	fma.rn.ftz.f32 	%f1070, %f1060, %f1060, %f1069;
	rsqrt.approx.ftz.f32 	%f1071, %f1070;
	mul.ftz.f32 	%f1072, %f1058, %f1071;
	mul.ftz.f32 	%f1073, %f1059, %f1071;
	mul.ftz.f32 	%f1074, %f1060, %f1071;
	mul.ftz.f32 	%f1075, %f1045, %f1066;
	fma.rn.ftz.f32 	%f1076, %f1042, %f1065, %f1075;
	fma.rn.ftz.f32 	%f1077, %f1048, %f1067, %f1076;
	add.ftz.f32 	%f1078, %f1077, %f1077;
	mul.ftz.f32 	%f1079, %f1042, %f1078;
	mul.ftz.f32 	%f1080, %f1045, %f1078;
	mul.ftz.f32 	%f1081, %f1048, %f1078;
	sub.ftz.f32 	%f1082, %f1079, %f1065;
	sub.ftz.f32 	%f1083, %f1080, %f1066;
	sub.ftz.f32 	%f1084, %f1081, %f1067;
	mul.ftz.f32 	%f1085, %f1083, %f1073;
	fma.rn.ftz.f32 	%f1086, %f1082, %f1072, %f1085;
	fma.rn.ftz.f32 	%f1087, %f1084, %f1074, %f1086;
	mov.f32 	%f1088, 0f00000000;
	max.ftz.f32 	%f153, %f1087, %f1088;
	setp.ltu.ftz.f32	%p28, %f153, 0f00000000;
	@%p28 bra 	BB2_31;

	lg2.approx.ftz.f32 	%f1089, %f153;
	mul.ftz.f32 	%f1090, %f1089, 0f418CCCCD;
	ex2.approx.ftz.f32 	%f1169, %f1090;
	bra.uni 	BB2_32;

BB2_31:
	neg.ftz.f32 	%f1091, %f153;
	lg2.approx.ftz.f32 	%f1092, %f1091;
	mul.ftz.f32 	%f1093, %f1092, 0f418CCCCD;
	ex2.approx.ftz.f32 	%f1094, %f1093;
	neg.ftz.f32 	%f1169, %f1094;

BB2_32:
	fma.rn.ftz.f32 	%f1170, %f1169, %f152, %f1170;
	fma.rn.ftz.f32 	%f1171, %f1169, %f152, %f1171;
	fma.rn.ftz.f32 	%f1172, %f1169, %f152, %f1172;

BB2_33:
	cvt.ftz.sat.f32.f32	%f163, %f152;
	setp.neu.ftz.f32	%p29, %f163, 0f00000000;
	@%p29 bra 	BB2_35;

	mov.f32 	%f1175, 0f00000000;
	mov.f32 	%f1174, %f1175;
	mov.f32 	%f1173, %f1175;
	bra.uni 	BB2_44;

BB2_35:
	mov.f32 	%f1142, 0f3F800000;
	div.approx.ftz.f32 	%f164, %f1142, %f163;
	mul.ftz.f32 	%f165, %f1170, %f164;
	setp.ltu.ftz.f32	%p30, %f165, 0f00000000;
	@%p30 bra 	BB2_37;

	lg2.approx.ftz.f32 	%f1096, %f165;
	mul.ftz.f32 	%f1097, %f1096, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1173, %f1097;
	bra.uni 	BB2_38;

BB2_37:
	neg.ftz.f32 	%f1098, %f165;
	lg2.approx.ftz.f32 	%f1099, %f1098;
	mul.ftz.f32 	%f1100, %f1099, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1101, %f1100;
	neg.ftz.f32 	%f1173, %f1101;

BB2_38:
	mul.ftz.f32 	%f169, %f1171, %f164;
	setp.ltu.ftz.f32	%p31, %f169, 0f00000000;
	@%p31 bra 	BB2_40;

	lg2.approx.ftz.f32 	%f1102, %f169;
	mul.ftz.f32 	%f1103, %f1102, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1174, %f1103;
	bra.uni 	BB2_41;

BB2_40:
	neg.ftz.f32 	%f1104, %f169;
	lg2.approx.ftz.f32 	%f1105, %f1104;
	mul.ftz.f32 	%f1106, %f1105, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1107, %f1106;
	neg.ftz.f32 	%f1174, %f1107;

BB2_41:
	mul.ftz.f32 	%f173, %f1172, %f164;
	setp.ltu.ftz.f32	%p32, %f173, 0f00000000;
	@%p32 bra 	BB2_43;

	lg2.approx.ftz.f32 	%f1108, %f173;
	mul.ftz.f32 	%f1109, %f1108, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1175, %f1109;
	bra.uni 	BB2_44;

BB2_43:
	neg.ftz.f32 	%f1110, %f173;
	lg2.approx.ftz.f32 	%f1111, %f1110;
	mul.ftz.f32 	%f1112, %f1111, 0f3EE8BA2E;
	ex2.approx.ftz.f32 	%f1113, %f1112;
	neg.ftz.f32 	%f1175, %f1113;

BB2_44:
	ld.param.u32 	%r71, [cuda_kernel_renderquad_param_7];
	setp.eq.s32	%p33, %r71, 0;
	@%p33 bra 	BB2_46;

	mul.wide.s32 	%rd133, %r27, 16;
	add.s64 	%rd134, %rd6, %rd133;
	st.global.v4.f32 	[%rd134], {%f1173, %f1174, %f1175, %f163};
	bra.uni 	BB2_50;

BB2_46:
	mul.wide.s32 	%rd136, %r27, 8;
	add.s64 	%rd137, %rd6, %rd136;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f163;
	mov.b16 	%rs1, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f1175;
	mov.b16 	%rs2, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f1174;
	mov.b16 	%rs3, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f1173;
	mov.b16 	%rs4, %temp;
}
	st.global.v4.u16 	[%rd137], {%rs4, %rs3, %rs2, %rs1};
	bra.uni 	BB2_50;

BB2_47:
	setp.eq.s32	%p34, %r17, 0;
	@%p34 bra 	BB2_49;

	mov.f32 	%f1117, 0f00000000;
	st.global.v4.f32 	[%rd1], {%f1117, %f1117, %f1117, %f1117};
	bra.uni 	BB2_50;

BB2_49:
	mov.f32 	%f1118, 0f00000000;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f1118;
	mov.b16 	%rs5, %temp;
}
	st.global.v4.u16 	[%rd2], {%rs5, %rs5, %rs5, %rs5};

BB2_50:
	ret;
}

.visible .entry cuda_kernel_prepsource(
	.param .u64 cuda_kernel_prepsource_param_0,
	.param .u64 cuda_kernel_prepsource_param_1,
	.param .u32 cuda_kernel_prepsource_param_2,
	.param .u32 cuda_kernel_prepsource_param_3,
	.param .u32 cuda_kernel_prepsource_param_4,
	.param .u32 cuda_kernel_prepsource_param_5,
	.param .u32 cuda_kernel_prepsource_param_6
)
{
	.reg .pred 	%p<9>;
	.reg .s16 	%rs<13>;
	.reg .s32 	%r<16>;
	.reg .f32 	%f<58>;
	.reg .s64 	%rd<15>;


	ld.param.u64 	%rd5, [cuda_kernel_prepsource_param_0];
	ld.param.u64 	%rd6, [cuda_kernel_prepsource_param_1];
	ld.param.u32 	%r6, [cuda_kernel_prepsource_param_2];
	ld.param.u32 	%r7, [cuda_kernel_prepsource_param_3];
	ld.param.u32 	%r3, [cuda_kernel_prepsource_param_4];
	ld.param.u32 	%r4, [cuda_kernel_prepsource_param_5];
	ld.param.u32 	%r5, [cuda_kernel_prepsource_param_6];
	cvta.to.global.u64 	%rd1, %rd6;
	cvta.to.global.u64 	%rd2, %rd5;
	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 	BB3_16;
	bra.uni 	BB3_1;

BB3_1:
	mad.lo.s32 	%r14, %r2, %r3, %r1;
	cvt.s64.s32	%rd3, %r14;
	setp.eq.s32	%p4, %r5, 0;
	@%p4 bra 	BB3_3;

	shl.b64 	%rd7, %rd3, 4;
	add.s64 	%rd8, %rd2, %rd7;
	ld.global.v4.f32 	{%f29, %f30, %f31, %f32}, [%rd8];
	mov.f32 	%f54, %f32;
	mov.f32 	%f53, %f31;
	mov.f32 	%f52, %f30;
	mov.f32 	%f51, %f29;
	bra.uni 	BB3_4;

BB3_3:
	shl.b64 	%rd9, %rd3, 3;
	add.s64 	%rd10, %rd2, %rd9;
	ld.global.v4.u16 	{%rs1, %rs2, %rs3, %rs4}, [%rd10];
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs1;
	cvt.f32.f16 	%f51, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs2;
	cvt.f32.f16 	%f52, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs3;
	cvt.f32.f16 	%f53, %temp;
	}
	{
	.reg .b16 %temp;
	mov.b16 	%temp, %rs4;
	cvt.f32.f16 	%f54, %temp;
	}

BB3_4:
	cvt.ftz.sat.f32.f32	%f15, %f54;
	setp.ltu.ftz.f32	%p5, %f53, 0f00000000;
	@%p5 bra 	BB3_6;

	lg2.approx.ftz.f32 	%f33, %f53;
	mul.ftz.f32 	%f34, %f33, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f55, %f34;
	bra.uni 	BB3_7;

BB3_6:
	neg.ftz.f32 	%f35, %f53;
	lg2.approx.ftz.f32 	%f36, %f35;
	mul.ftz.f32 	%f37, %f36, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f38, %f37;
	neg.ftz.f32 	%f55, %f38;

BB3_7:
	mul.ftz.f32 	%f20, %f55, %f15;
	setp.ltu.ftz.f32	%p6, %f52, 0f00000000;
	@%p6 bra 	BB3_9;

	lg2.approx.ftz.f32 	%f39, %f52;
	mul.ftz.f32 	%f40, %f39, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f56, %f40;
	bra.uni 	BB3_10;

BB3_9:
	neg.ftz.f32 	%f41, %f52;
	lg2.approx.ftz.f32 	%f42, %f41;
	mul.ftz.f32 	%f43, %f42, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f44, %f43;
	neg.ftz.f32 	%f56, %f44;

BB3_10:
	mul.ftz.f32 	%f24, %f56, %f15;
	setp.ltu.ftz.f32	%p7, %f51, 0f00000000;
	@%p7 bra 	BB3_12;

	lg2.approx.ftz.f32 	%f45, %f51;
	mul.ftz.f32 	%f46, %f45, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f57, %f46;
	bra.uni 	BB3_13;

BB3_12:
	neg.ftz.f32 	%f47, %f51;
	lg2.approx.ftz.f32 	%f48, %f47;
	mul.ftz.f32 	%f49, %f48, 0f400CCCCD;
	ex2.approx.ftz.f32 	%f50, %f49;
	neg.ftz.f32 	%f57, %f50;

BB3_13:
	mul.ftz.f32 	%f28, %f57, %f15;
	mad.lo.s32 	%r15, %r2, %r4, %r1;
	cvt.s64.s32	%rd4, %r15;
	@%p4 bra 	BB3_15;

	shl.b64 	%rd11, %rd4, 4;
	add.s64 	%rd12, %rd1, %rd11;
	st.global.v4.f32 	[%rd12], {%f28, %f24, %f20, %f15};
	bra.uni 	BB3_16;

BB3_15:
	shl.b64 	%rd13, %rd4, 3;
	add.s64 	%rd14, %rd1, %rd13;
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f15;
	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, %f24;
	mov.b16 	%rs11, %temp;
}
	{
	.reg .b16 %temp;
	cvt.rn.ftz.f16.f32 	%temp, %f28;
	mov.b16 	%rs12, %temp;
}
	st.global.v4.u16 	[%rd14], {%rs12, %rs11, %rs10, %rs9};

BB3_16:
	ret;
}


