//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-35583870
// Cuda compilation tools, release 12.8, V12.8.93
// Based on NVVM 7.0.1
//

.version 8.7
.target sm_90
.address_size 64

	// .globl	r_8_32_2_32_2_2_2_4_4_24_2

.visible .entry r_8_32_2_32_2_2_2_4_4_24_2(
	.param .u64 r_8_32_2_32_2_2_2_4_4_24_2_param_0,
	.param .u64 r_8_32_2_32_2_2_2_4_4_24_2_param_1,
	.param .u64 r_8_32_2_32_2_2_2_4_4_24_2_param_2,
	.param .u64 r_8_32_2_32_2_2_2_4_4_24_2_param_3
)
.maxntid 64, 1, 1
{
	.reg .pred 	%p<2>;
	.reg .b16 	%rs<265>;
	.reg .f32 	%f<513>;
	.reg .b32 	%r<492>;
	.reg .b64 	%rd<25>;


	ld.param.u64 	%rd11, [r_8_32_2_32_2_2_2_4_4_24_2_param_1];
	ld.param.u64 	%rd12, [r_8_32_2_32_2_2_2_4_4_24_2_param_2];
	cvta.to.global.u64 	%rd23, %rd12;
	cvta.to.global.u64 	%rd24, %rd11;
	mov.u32 	%r4, %tid.x;
	shr.s32 	%r5, %r4, 2;
	shl.b32 	%r6, %r4, 1;
	and.b32  	%r7, %r6, 6;
	mov.u32 	%r8, %ctaid.y;
	mul.lo.s32 	%r9, %r8, 98304;
	or.b32  	%r10, %r7, %r9;
	mov.u32 	%r11, %tid.y;
	mad.lo.s32 	%r12, %r11, 12288, %r10;
	mul.lo.s32 	%r13, %r5, 768;
	add.s32 	%r14, %r12, %r13;
	add.s32 	%r15, %r14, 16;
	mul.wide.s32 	%rd3, %r15, 2;
	mov.u32 	%r16, %ctaid.x;
	mul.lo.s32 	%r17, %r16, 24576;
	or.b32  	%r18, %r7, %r17;
	add.s32 	%r19, %r18, %r13;
	add.s32 	%r20, %r19, 12288;
	mul.wide.s32 	%rd4, %r20, 2;
	mov.f32 	%f449, 0f00000000;
	mov.u32 	%r3, 0;
	mov.u32 	%r491, %r3;
	mov.f32 	%f450, %f449;
	mov.f32 	%f451, %f449;
	mov.f32 	%f452, %f449;
	mov.f32 	%f453, %f449;
	mov.f32 	%f454, %f449;
	mov.f32 	%f455, %f449;
	mov.f32 	%f456, %f449;
	mov.f32 	%f457, %f449;
	mov.f32 	%f458, %f449;
	mov.f32 	%f459, %f449;
	mov.f32 	%f460, %f449;
	mov.f32 	%f461, %f449;
	mov.f32 	%f462, %f449;
	mov.f32 	%f463, %f449;
	mov.f32 	%f464, %f449;
	mov.f32 	%f465, %f449;
	mov.f32 	%f466, %f449;
	mov.f32 	%f467, %f449;
	mov.f32 	%f468, %f449;
	mov.f32 	%f469, %f449;
	mov.f32 	%f470, %f449;
	mov.f32 	%f471, %f449;
	mov.f32 	%f472, %f449;
	mov.f32 	%f473, %f449;
	mov.f32 	%f474, %f449;
	mov.f32 	%f475, %f449;
	mov.f32 	%f476, %f449;
	mov.f32 	%f477, %f449;
	mov.f32 	%f478, %f449;
	mov.f32 	%f479, %f449;
	mov.f32 	%f480, %f449;
	mov.f32 	%f481, %f449;
	mov.f32 	%f482, %f449;
	mov.f32 	%f483, %f449;
	mov.f32 	%f484, %f449;
	mov.f32 	%f485, %f449;
	mov.f32 	%f486, %f449;
	mov.f32 	%f487, %f449;
	mov.f32 	%f488, %f449;
	mov.f32 	%f489, %f449;
	mov.f32 	%f490, %f449;
	mov.f32 	%f491, %f449;
	mov.f32 	%f492, %f449;
	mov.f32 	%f493, %f449;
	mov.f32 	%f494, %f449;
	mov.f32 	%f495, %f449;
	mov.f32 	%f496, %f449;
	mov.f32 	%f497, %f449;
	mov.f32 	%f498, %f449;
	mov.f32 	%f499, %f449;
	mov.f32 	%f500, %f449;
	mov.f32 	%f501, %f449;
	mov.f32 	%f502, %f449;
	mov.f32 	%f503, %f449;
	mov.f32 	%f504, %f449;
	mov.f32 	%f505, %f449;
	mov.f32 	%f506, %f449;
	mov.f32 	%f507, %f449;
	mov.f32 	%f508, %f449;
	mov.f32 	%f509, %f449;
	mov.f32 	%f510, %f449;
	mov.f32 	%f511, %f449;
	mov.f32 	%f512, %f449;

$L__BB0_1:
	add.s64 	%rd13, %rd24, %rd3;
	ld.global.u32 	%r25, [%rd13];
	ld.global.u32 	%r27, [%rd13+16];
	ld.global.u32 	%r81, [%rd13+49120];
	ld.global.u32 	%r83, [%rd13+49136];
	ld.global.u32 	%r137, [%rd13+49152];
	ld.global.u32 	%r139, [%rd13+49168];
	ld.global.u32 	%r82, [%rd13+61408];
	ld.global.u32 	%r84, [%rd13+61424];
	ld.global.u32 	%r138, [%rd13+61440];
	ld.global.u32 	%r140, [%rd13+61456];
	ld.global.u32 	%r193, [%rd13+98272];
	ld.global.u32 	%r195, [%rd13+98288];
	ld.global.u32 	%r249, [%rd13+98304];
	ld.global.u32 	%r251, [%rd13+98320];
	ld.global.u32 	%r194, [%rd13+110560];
	ld.global.u32 	%r196, [%rd13+110576];
	ld.global.u32 	%r250, [%rd13+110592];
	ld.global.u32 	%r252, [%rd13+110608];
	ld.global.u32 	%r418, [%rd13+12256];
	ld.global.u32 	%r420, [%rd13+12272];
	ld.global.u32 	%r26, [%rd13+12288];
	ld.global.u32 	%r28, [%rd13+12304];
	ld.global.u32 	%r305, [%rd13+147424];
	ld.global.u32 	%r307, [%rd13+147440];
	ld.global.u32 	%r361, [%rd13+147456];
	ld.global.u32 	%r363, [%rd13+147472];
	ld.global.u32 	%r306, [%rd13+159712];
	ld.global.u32 	%r308, [%rd13+159728];
	ld.global.u32 	%r362, [%rd13+159744];
	ld.global.u32 	%r364, [%rd13+159760];
	ld.global.u32 	%r419, [%rd13+-16];
	ld.global.u32 	%r417, [%rd13+-32];
	add.s64 	%rd14, %rd23, %rd4;
	ld.global.u32 	%r85, [%rd14];
	ld.global.u32 	%r86, [%rd14+16];
	ld.global.u32 	%r29, [%rd14+32];
	ld.global.u32 	%r30, [%rd14+48];
	ld.global.u32 	%r43, [%rd14+-24544];
	ld.global.u32 	%r99, [%rd14+12288];
	ld.global.u32 	%r100, [%rd14+12304];
	ld.global.u32 	%r57, [%rd14+12320];
	ld.global.u32 	%r58, [%rd14+12336];
	ld.global.u32 	%r44, [%rd14+-24528];
	ld.global.u32 	%r113, [%rd14+-12288];
	ld.global.u32 	%r114, [%rd14+-12272];
	ld.global.u32 	%r71, [%rd14+-12256];
	ld.global.u32 	%r72, [%rd14+-12240];
	ld.global.u32 	%r128, [%rd14+-24560];
	ld.global.u32 	%r127, [%rd14+-24576];
	mov.u32 	%r21, %r3;
	mov.u32 	%r22, %r3;
	mov.u32 	%r23, %r3;
	mov.u32 	%r24, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r21, %r22, %r23, %r24}, {%r25, %r26, %r27, %r28},{%r29, %r30}, {%r21, %r22, %r23, %r24};
	// end inline asm
	mov.b32 	%f193, %r21;
	mov.b32 	%f194, %r22;
	mov.b32 	%f195, %r23;
	mov.b32 	%f196, %r24;
	mov.u32 	%r35, %r3;
	mov.u32 	%r36, %r3;
	mov.u32 	%r37, %r3;
	mov.u32 	%r38, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r35, %r36, %r37, %r38}, {%r25, %r26, %r27, %r28},{%r43, %r44}, {%r35, %r36, %r37, %r38};
	// end inline asm
	mov.b32 	%f197, %r35;
	mov.b32 	%f198, %r36;
	mov.b32 	%f199, %r37;
	mov.b32 	%f200, %r38;
	mov.u32 	%r49, %r3;
	mov.u32 	%r50, %r3;
	mov.u32 	%r51, %r3;
	mov.u32 	%r52, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r49, %r50, %r51, %r52}, {%r25, %r26, %r27, %r28},{%r57, %r58}, {%r49, %r50, %r51, %r52};
	// end inline asm
	mov.b32 	%f201, %r49;
	mov.b32 	%f202, %r50;
	mov.b32 	%f203, %r51;
	mov.b32 	%f204, %r52;
	mov.u32 	%r63, %r3;
	mov.u32 	%r64, %r3;
	mov.u32 	%r65, %r3;
	mov.u32 	%r66, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r63, %r64, %r65, %r66}, {%r25, %r26, %r27, %r28},{%r71, %r72}, {%r63, %r64, %r65, %r66};
	// end inline asm
	mov.b32 	%f205, %r63;
	mov.b32 	%f206, %r64;
	mov.b32 	%f207, %r65;
	mov.b32 	%f208, %r66;
	mov.u32 	%r77, %r3;
	mov.u32 	%r78, %r3;
	mov.u32 	%r79, %r3;
	mov.u32 	%r80, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r77, %r78, %r79, %r80}, {%r81, %r82, %r83, %r84},{%r85, %r86}, {%r77, %r78, %r79, %r80};
	// end inline asm
	mov.b32 	%f209, %r77;
	mov.b32 	%f210, %r78;
	mov.b32 	%f211, %r79;
	mov.b32 	%f212, %r80;
	mov.u32 	%r91, %r3;
	mov.u32 	%r92, %r3;
	mov.u32 	%r93, %r3;
	mov.u32 	%r94, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r91, %r92, %r93, %r94}, {%r81, %r82, %r83, %r84},{%r99, %r100}, {%r91, %r92, %r93, %r94};
	// end inline asm
	mov.b32 	%f213, %r91;
	mov.b32 	%f214, %r92;
	mov.b32 	%f215, %r93;
	mov.b32 	%f216, %r94;
	mov.u32 	%r105, %r3;
	mov.u32 	%r106, %r3;
	mov.u32 	%r107, %r3;
	mov.u32 	%r108, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r105, %r106, %r107, %r108}, {%r81, %r82, %r83, %r84},{%r113, %r114}, {%r105, %r106, %r107, %r108};
	// end inline asm
	mov.b32 	%f217, %r105;
	mov.b32 	%f218, %r106;
	mov.b32 	%f219, %r107;
	mov.b32 	%f220, %r108;
	mov.u32 	%r119, %r3;
	mov.u32 	%r120, %r3;
	mov.u32 	%r121, %r3;
	mov.u32 	%r122, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r119, %r120, %r121, %r122}, {%r81, %r82, %r83, %r84},{%r127, %r128}, {%r119, %r120, %r121, %r122};
	// end inline asm
	mov.b32 	%f221, %r119;
	mov.b32 	%f222, %r120;
	mov.b32 	%f223, %r121;
	mov.b32 	%f224, %r122;
	mov.u32 	%r133, %r3;
	mov.u32 	%r134, %r3;
	mov.u32 	%r135, %r3;
	mov.u32 	%r136, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r133, %r134, %r135, %r136}, {%r137, %r138, %r139, %r140},{%r29, %r30}, {%r133, %r134, %r135, %r136};
	// end inline asm
	mov.b32 	%f225, %r133;
	mov.b32 	%f226, %r134;
	mov.b32 	%f227, %r135;
	mov.b32 	%f228, %r136;
	mov.u32 	%r147, %r3;
	mov.u32 	%r148, %r3;
	mov.u32 	%r149, %r3;
	mov.u32 	%r150, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r147, %r148, %r149, %r150}, {%r137, %r138, %r139, %r140},{%r43, %r44}, {%r147, %r148, %r149, %r150};
	// end inline asm
	mov.b32 	%f229, %r147;
	mov.b32 	%f230, %r148;
	mov.b32 	%f231, %r149;
	mov.b32 	%f232, %r150;
	mov.u32 	%r161, %r3;
	mov.u32 	%r162, %r3;
	mov.u32 	%r163, %r3;
	mov.u32 	%r164, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r161, %r162, %r163, %r164}, {%r137, %r138, %r139, %r140},{%r57, %r58}, {%r161, %r162, %r163, %r164};
	// end inline asm
	mov.b32 	%f233, %r161;
	mov.b32 	%f234, %r162;
	mov.b32 	%f235, %r163;
	mov.b32 	%f236, %r164;
	mov.u32 	%r175, %r3;
	mov.u32 	%r176, %r3;
	mov.u32 	%r177, %r3;
	mov.u32 	%r178, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r175, %r176, %r177, %r178}, {%r137, %r138, %r139, %r140},{%r71, %r72}, {%r175, %r176, %r177, %r178};
	// end inline asm
	mov.b32 	%f237, %r175;
	mov.b32 	%f238, %r176;
	mov.b32 	%f239, %r177;
	mov.b32 	%f240, %r178;
	mov.u32 	%r189, %r3;
	mov.u32 	%r190, %r3;
	mov.u32 	%r191, %r3;
	mov.u32 	%r192, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r189, %r190, %r191, %r192}, {%r193, %r194, %r195, %r196},{%r85, %r86}, {%r189, %r190, %r191, %r192};
	// end inline asm
	mov.b32 	%f241, %r189;
	mov.b32 	%f242, %r190;
	mov.b32 	%f243, %r191;
	mov.b32 	%f244, %r192;
	mov.u32 	%r203, %r3;
	mov.u32 	%r204, %r3;
	mov.u32 	%r205, %r3;
	mov.u32 	%r206, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r203, %r204, %r205, %r206}, {%r193, %r194, %r195, %r196},{%r99, %r100}, {%r203, %r204, %r205, %r206};
	// end inline asm
	mov.b32 	%f245, %r203;
	mov.b32 	%f246, %r204;
	mov.b32 	%f247, %r205;
	mov.b32 	%f248, %r206;
	mov.u32 	%r217, %r3;
	mov.u32 	%r218, %r3;
	mov.u32 	%r219, %r3;
	mov.u32 	%r220, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r217, %r218, %r219, %r220}, {%r193, %r194, %r195, %r196},{%r113, %r114}, {%r217, %r218, %r219, %r220};
	// end inline asm
	mov.b32 	%f249, %r217;
	mov.b32 	%f250, %r218;
	mov.b32 	%f251, %r219;
	mov.b32 	%f252, %r220;
	mov.u32 	%r231, %r3;
	mov.u32 	%r232, %r3;
	mov.u32 	%r233, %r3;
	mov.u32 	%r234, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r231, %r232, %r233, %r234}, {%r193, %r194, %r195, %r196},{%r127, %r128}, {%r231, %r232, %r233, %r234};
	// end inline asm
	mov.b32 	%f253, %r231;
	mov.b32 	%f254, %r232;
	mov.b32 	%f255, %r233;
	mov.b32 	%f256, %r234;
	mov.u32 	%r245, %r3;
	mov.u32 	%r246, %r3;
	mov.u32 	%r247, %r3;
	mov.u32 	%r248, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r245, %r246, %r247, %r248}, {%r249, %r250, %r251, %r252},{%r29, %r30}, {%r245, %r246, %r247, %r248};
	// end inline asm
	mov.b32 	%f257, %r245;
	mov.b32 	%f258, %r246;
	mov.b32 	%f259, %r247;
	mov.b32 	%f260, %r248;
	mov.u32 	%r259, %r3;
	mov.u32 	%r260, %r3;
	mov.u32 	%r261, %r3;
	mov.u32 	%r262, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r259, %r260, %r261, %r262}, {%r249, %r250, %r251, %r252},{%r43, %r44}, {%r259, %r260, %r261, %r262};
	// end inline asm
	mov.b32 	%f261, %r259;
	mov.b32 	%f262, %r260;
	mov.b32 	%f263, %r261;
	mov.b32 	%f264, %r262;
	mov.u32 	%r273, %r3;
	mov.u32 	%r274, %r3;
	mov.u32 	%r275, %r3;
	mov.u32 	%r276, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r273, %r274, %r275, %r276}, {%r249, %r250, %r251, %r252},{%r57, %r58}, {%r273, %r274, %r275, %r276};
	// end inline asm
	mov.b32 	%f265, %r273;
	mov.b32 	%f266, %r274;
	mov.b32 	%f267, %r275;
	mov.b32 	%f268, %r276;
	mov.u32 	%r287, %r3;
	mov.u32 	%r288, %r3;
	mov.u32 	%r289, %r3;
	mov.u32 	%r290, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r287, %r288, %r289, %r290}, {%r249, %r250, %r251, %r252},{%r71, %r72}, {%r287, %r288, %r289, %r290};
	// end inline asm
	mov.b32 	%f269, %r287;
	mov.b32 	%f270, %r288;
	mov.b32 	%f271, %r289;
	mov.b32 	%f272, %r290;
	mov.u32 	%r301, %r3;
	mov.u32 	%r302, %r3;
	mov.u32 	%r303, %r3;
	mov.u32 	%r304, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r301, %r302, %r303, %r304}, {%r305, %r306, %r307, %r308},{%r85, %r86}, {%r301, %r302, %r303, %r304};
	// end inline asm
	mov.b32 	%f273, %r301;
	mov.b32 	%f274, %r302;
	mov.b32 	%f275, %r303;
	mov.b32 	%f276, %r304;
	mov.u32 	%r315, %r3;
	mov.u32 	%r316, %r3;
	mov.u32 	%r317, %r3;
	mov.u32 	%r318, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r315, %r316, %r317, %r318}, {%r305, %r306, %r307, %r308},{%r99, %r100}, {%r315, %r316, %r317, %r318};
	// end inline asm
	mov.b32 	%f277, %r315;
	mov.b32 	%f278, %r316;
	mov.b32 	%f279, %r317;
	mov.b32 	%f280, %r318;
	mov.u32 	%r329, %r3;
	mov.u32 	%r330, %r3;
	mov.u32 	%r331, %r3;
	mov.u32 	%r332, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r329, %r330, %r331, %r332}, {%r305, %r306, %r307, %r308},{%r113, %r114}, {%r329, %r330, %r331, %r332};
	// end inline asm
	mov.b32 	%f281, %r329;
	mov.b32 	%f282, %r330;
	mov.b32 	%f283, %r331;
	mov.b32 	%f284, %r332;
	mov.u32 	%r343, %r3;
	mov.u32 	%r344, %r3;
	mov.u32 	%r345, %r3;
	mov.u32 	%r346, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r343, %r344, %r345, %r346}, {%r305, %r306, %r307, %r308},{%r127, %r128}, {%r343, %r344, %r345, %r346};
	// end inline asm
	mov.b32 	%f285, %r343;
	mov.b32 	%f286, %r344;
	mov.b32 	%f287, %r345;
	mov.b32 	%f288, %r346;
	mov.u32 	%r357, %r3;
	mov.u32 	%r358, %r3;
	mov.u32 	%r359, %r3;
	mov.u32 	%r360, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r357, %r358, %r359, %r360}, {%r361, %r362, %r363, %r364},{%r29, %r30}, {%r357, %r358, %r359, %r360};
	// end inline asm
	mov.b32 	%f289, %r357;
	mov.b32 	%f290, %r358;
	mov.b32 	%f291, %r359;
	mov.b32 	%f292, %r360;
	mov.u32 	%r371, %r3;
	mov.u32 	%r372, %r3;
	mov.u32 	%r373, %r3;
	mov.u32 	%r374, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r371, %r372, %r373, %r374}, {%r361, %r362, %r363, %r364},{%r43, %r44}, {%r371, %r372, %r373, %r374};
	// end inline asm
	mov.b32 	%f293, %r371;
	mov.b32 	%f294, %r372;
	mov.b32 	%f295, %r373;
	mov.b32 	%f296, %r374;
	mov.u32 	%r385, %r3;
	mov.u32 	%r386, %r3;
	mov.u32 	%r387, %r3;
	mov.u32 	%r388, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r385, %r386, %r387, %r388}, {%r361, %r362, %r363, %r364},{%r57, %r58}, {%r385, %r386, %r387, %r388};
	// end inline asm
	mov.b32 	%f297, %r385;
	mov.b32 	%f298, %r386;
	mov.b32 	%f299, %r387;
	mov.b32 	%f300, %r388;
	mov.u32 	%r399, %r3;
	mov.u32 	%r400, %r3;
	mov.u32 	%r401, %r3;
	mov.u32 	%r402, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r399, %r400, %r401, %r402}, {%r361, %r362, %r363, %r364},{%r71, %r72}, {%r399, %r400, %r401, %r402};
	// end inline asm
	mov.b32 	%f301, %r399;
	mov.b32 	%f302, %r400;
	mov.b32 	%f303, %r401;
	mov.b32 	%f304, %r402;
	mov.u32 	%r413, %r3;
	mov.u32 	%r414, %r3;
	mov.u32 	%r415, %r3;
	mov.u32 	%r416, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r413, %r414, %r415, %r416}, {%r417, %r418, %r419, %r420},{%r85, %r86}, {%r413, %r414, %r415, %r416};
	// end inline asm
	mov.b32 	%f305, %r413;
	mov.b32 	%f306, %r414;
	mov.b32 	%f307, %r415;
	mov.b32 	%f308, %r416;
	mov.u32 	%r427, %r3;
	mov.u32 	%r428, %r3;
	mov.u32 	%r429, %r3;
	mov.u32 	%r430, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r427, %r428, %r429, %r430}, {%r417, %r418, %r419, %r420},{%r99, %r100}, {%r427, %r428, %r429, %r430};
	// end inline asm
	mov.b32 	%f309, %r427;
	mov.b32 	%f310, %r428;
	mov.b32 	%f311, %r429;
	mov.b32 	%f312, %r430;
	mov.u32 	%r441, %r3;
	mov.u32 	%r442, %r3;
	mov.u32 	%r443, %r3;
	mov.u32 	%r444, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r441, %r442, %r443, %r444}, {%r417, %r418, %r419, %r420},{%r113, %r114}, {%r441, %r442, %r443, %r444};
	// end inline asm
	mov.b32 	%f313, %r441;
	mov.b32 	%f314, %r442;
	mov.b32 	%f315, %r443;
	mov.b32 	%f316, %r444;
	mov.u32 	%r455, %r3;
	mov.u32 	%r456, %r3;
	mov.u32 	%r457, %r3;
	mov.u32 	%r458, %r3;
	// begin inline asm
	mma.sync.aligned.m16n8k16.row.col.f32.f16.f16.f32{%r455, %r456, %r457, %r458}, {%r417, %r418, %r419, %r420},{%r127, %r128}, {%r455, %r456, %r457, %r458};
	// end inline asm
	mov.b32 	%f317, %r455;
	mov.b32 	%f318, %r456;
	mov.b32 	%f319, %r457;
	mov.b32 	%f320, %r458;
	add.f32 	%f321, %f489, %f317;
	add.f32 	%f489, %f321, %f197;
	add.f32 	%f322, %f490, %f318;
	add.f32 	%f490, %f322, %f198;
	add.f32 	%f323, %f499, %f315;
	add.f32 	%f499, %f323, %f207;
	add.f32 	%f324, %f500, %f316;
	add.f32 	%f500, %f324, %f208;
	add.f32 	%f325, %f501, %f307;
	add.f32 	%f501, %f325, %f195;
	add.f32 	%f326, %f502, %f308;
	add.f32 	%f502, %f326, %f196;
	add.f32 	%f327, %f503, %f311;
	add.f32 	%f503, %f327, %f203;
	add.f32 	%f328, %f504, %f312;
	add.f32 	%f504, %f328, %f204;
	add.f32 	%f329, %f505, %f221;
	add.f32 	%f505, %f329, %f229;
	add.f32 	%f330, %f506, %f222;
	add.f32 	%f506, %f330, %f230;
	add.f32 	%f331, %f507, %f217;
	add.f32 	%f507, %f331, %f237;
	add.f32 	%f332, %f508, %f218;
	add.f32 	%f508, %f332, %f238;
	add.f32 	%f333, %f491, %f313;
	add.f32 	%f491, %f333, %f205;
	add.f32 	%f334, %f509, %f209;
	add.f32 	%f509, %f334, %f225;
	add.f32 	%f335, %f510, %f210;
	add.f32 	%f510, %f335, %f226;
	add.f32 	%f336, %f511, %f213;
	add.f32 	%f511, %f336, %f233;
	add.f32 	%f337, %f512, %f214;
	add.f32 	%f512, %f337, %f234;
	add.f32 	%f338, %f488, %f223;
	add.f32 	%f488, %f338, %f231;
	add.f32 	%f339, %f487, %f224;
	add.f32 	%f487, %f339, %f232;
	add.f32 	%f340, %f486, %f219;
	add.f32 	%f486, %f340, %f239;
	add.f32 	%f341, %f485, %f220;
	add.f32 	%f485, %f341, %f240;
	add.f32 	%f342, %f484, %f211;
	add.f32 	%f484, %f342, %f227;
	add.f32 	%f343, %f483, %f212;
	add.f32 	%f483, %f343, %f228;
	add.f32 	%f344, %f492, %f314;
	add.f32 	%f492, %f344, %f206;
	add.f32 	%f345, %f482, %f215;
	add.f32 	%f482, %f345, %f235;
	add.f32 	%f346, %f481, %f216;
	add.f32 	%f481, %f346, %f236;
	add.f32 	%f347, %f480, %f253;
	add.f32 	%f480, %f347, %f261;
	add.f32 	%f348, %f479, %f254;
	add.f32 	%f479, %f348, %f262;
	add.f32 	%f349, %f478, %f249;
	add.f32 	%f478, %f349, %f269;
	add.f32 	%f350, %f477, %f250;
	add.f32 	%f477, %f350, %f270;
	add.f32 	%f351, %f476, %f241;
	add.f32 	%f476, %f351, %f257;
	add.f32 	%f352, %f475, %f242;
	add.f32 	%f475, %f352, %f258;
	add.f32 	%f353, %f474, %f245;
	add.f32 	%f474, %f353, %f265;
	add.f32 	%f354, %f473, %f246;
	add.f32 	%f473, %f354, %f266;
	add.f32 	%f355, %f493, %f305;
	add.f32 	%f493, %f355, %f193;
	add.f32 	%f356, %f472, %f255;
	add.f32 	%f472, %f356, %f263;
	add.f32 	%f357, %f471, %f256;
	add.f32 	%f471, %f357, %f264;
	add.f32 	%f358, %f470, %f251;
	add.f32 	%f470, %f358, %f271;
	add.f32 	%f359, %f469, %f252;
	add.f32 	%f469, %f359, %f272;
	add.f32 	%f360, %f468, %f243;
	add.f32 	%f468, %f360, %f259;
	add.f32 	%f361, %f467, %f244;
	add.f32 	%f467, %f361, %f260;
	add.f32 	%f362, %f466, %f247;
	add.f32 	%f466, %f362, %f267;
	add.f32 	%f363, %f465, %f248;
	add.f32 	%f465, %f363, %f268;
	add.f32 	%f364, %f464, %f285;
	add.f32 	%f464, %f364, %f293;
	add.f32 	%f365, %f463, %f286;
	add.f32 	%f463, %f365, %f294;
	add.f32 	%f366, %f494, %f306;
	add.f32 	%f494, %f366, %f194;
	add.f32 	%f367, %f462, %f281;
	add.f32 	%f462, %f367, %f301;
	add.f32 	%f368, %f461, %f282;
	add.f32 	%f461, %f368, %f302;
	add.f32 	%f369, %f460, %f273;
	add.f32 	%f460, %f369, %f289;
	add.f32 	%f370, %f459, %f274;
	add.f32 	%f459, %f370, %f290;
	add.f32 	%f371, %f458, %f277;
	add.f32 	%f458, %f371, %f297;
	add.f32 	%f372, %f457, %f278;
	add.f32 	%f457, %f372, %f298;
	add.f32 	%f373, %f456, %f287;
	add.f32 	%f456, %f373, %f295;
	add.f32 	%f374, %f455, %f288;
	add.f32 	%f455, %f374, %f296;
	add.f32 	%f375, %f454, %f283;
	add.f32 	%f454, %f375, %f303;
	add.f32 	%f376, %f453, %f284;
	add.f32 	%f453, %f376, %f304;
	add.f32 	%f377, %f495, %f309;
	add.f32 	%f495, %f377, %f201;
	add.f32 	%f378, %f452, %f275;
	add.f32 	%f452, %f378, %f291;
	add.f32 	%f379, %f451, %f276;
	add.f32 	%f451, %f379, %f292;
	add.f32 	%f380, %f450, %f279;
	add.f32 	%f450, %f380, %f299;
	add.f32 	%f381, %f449, %f280;
	add.f32 	%f449, %f381, %f300;
	add.f32 	%f382, %f496, %f310;
	add.f32 	%f496, %f382, %f202;
	add.f32 	%f383, %f497, %f319;
	add.f32 	%f497, %f383, %f199;
	add.f32 	%f384, %f498, %f320;
	add.f32 	%f498, %f384, %f200;
	add.s64 	%rd24, %rd24, 64;
	add.s64 	%rd23, %rd23, 64;
	add.s32 	%r491, %r491, 1;
	setp.ne.s32 	%p1, %r491, 24;
	@%p1 bra 	$L__BB0_1;

	mov.u32 	%r490, %tid.x;
	shl.b32 	%r489, %r490, 1;
	and.b32  	%r488, %r489, 6;
	ld.param.u64 	%rd22, [r_8_32_2_32_2_2_2_4_4_24_2_param_0];
	mov.u32 	%r487, %tid.y;
	mov.u32 	%r486, %ctaid.y;
	mov.u32 	%r485, %tid.x;
	mov.u32 	%r484, %ctaid.x;
	ld.param.u64 	%rd21, [r_8_32_2_32_2_2_2_4_4_24_2_param_3];
	cvta.to.global.u64 	%rd15, %rd21;
	shl.b32 	%r470, %r484, 5;
	or.b32  	%r474, %r488, %r470;
	mul.wide.s32 	%rd16, %r474, 2;
	add.s64 	%rd17, %rd15, %rd16;
	ld.global.v2.u16 	{%rs257, %rs258}, [%rd17+32];
	ld.global.v2.u16 	{%rs259, %rs260}, [%rd17+48];
	ld.global.v2.u16 	{%rs261, %rs262}, [%rd17+16];
	ld.global.v2.u16 	{%rs263, %rs264}, [%rd17];
	shl.b32 	%r475, %r485, 8;
	and.b32  	%r476, %r475, -1024;
	shl.b32 	%r478, %r486, 17;
	shl.b32 	%r480, %r487, 14;
	add.s32 	%r481, %r480, %r478;
	add.s32 	%r482, %r481, %r476;
	add.s32 	%r483, %r482, %r474;
	// begin inline asm
	{  cvt.rn.f16.f32 %rs1, %f456;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs5, %f455;}

	// end inline asm
	cvta.to.global.u64 	%rd18, %rd22;
	mul.wide.s32 	%rd19, %r483, 2;
	add.s64 	%rd20, %rd18, %rd19;
	// begin inline asm
	{add.f16 %rs6,%rs5,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs2,%rs1,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+212992], {%rs2, %rs6};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs9, %f454;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs13, %f453;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs14,%rs13,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs10,%rs9,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+213008], {%rs10, %rs14};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs17, %f452;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs21, %f451;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs22,%rs21,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs18,%rs17,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+213024], {%rs18, %rs22};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs25, %f450;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs29, %f449;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs30,%rs29,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs26,%rs25,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+213040], {%rs26, %rs30};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs33, %f493;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs37, %f494;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs38,%rs37,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs34,%rs33,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+32], {%rs34, %rs38};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs41, %f495;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs45, %f496;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs46,%rs45,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs42,%rs41,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+48], {%rs42, %rs46};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs49, %f505;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs53, %f506;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs54,%rs53,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs50,%rs49,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+65536], {%rs50, %rs54};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs57, %f507;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs61, %f508;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs62,%rs61,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs58,%rs57,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+65552], {%rs58, %rs62};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs65, %f509;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs69, %f510;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs70,%rs69,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs66,%rs65,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+65568], {%rs66, %rs70};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs73, %f511;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs77, %f512;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs78,%rs77,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs74,%rs73,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+65584], {%rs74, %rs78};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs81, %f488;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs85, %f487;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs86,%rs85,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs82,%rs81,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+81920], {%rs82, %rs86};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs89, %f486;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs93, %f485;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs94,%rs93,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs90,%rs89,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+81936], {%rs90, %rs94};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs97, %f484;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs101, %f483;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs102,%rs101,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs98,%rs97,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+81952], {%rs98, %rs102};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs105, %f482;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs109, %f481;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs110,%rs109,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs106,%rs105,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+81968], {%rs106, %rs110};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs113, %f480;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs117, %f479;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs118,%rs117,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs114,%rs113,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+131072], {%rs114, %rs118};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs121, %f478;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs125, %f477;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs126,%rs125,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs122,%rs121,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+131088], {%rs122, %rs126};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs129, %f476;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs133, %f475;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs134,%rs133,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs130,%rs129,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+131104], {%rs130, %rs134};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs137, %f474;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs141, %f473;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs142,%rs141,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs138,%rs137,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+131120], {%rs138, %rs142};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs145, %f472;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs149, %f471;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs150,%rs149,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs146,%rs145,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+147456], {%rs146, %rs150};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs153, %f470;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs157, %f469;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs158,%rs157,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs154,%rs153,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+147472], {%rs154, %rs158};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs161, %f468;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs165, %f467;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs166,%rs165,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs162,%rs161,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+147488], {%rs162, %rs166};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs169, %f466;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs173, %f465;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs174,%rs173,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs170,%rs169,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+147504], {%rs170, %rs174};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs177, %f491;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs181, %f492;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs182,%rs181,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs178,%rs177,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+16], {%rs178, %rs182};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs185, %f497;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs189, %f498;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs190,%rs189,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs186,%rs185,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+16384], {%rs186, %rs190};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs193, %f499;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs197, %f500;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs198,%rs197,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs194,%rs193,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+16400], {%rs194, %rs198};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs201, %f501;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs205, %f502;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs206,%rs205,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs202,%rs201,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+16416], {%rs202, %rs206};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs209, %f503;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs213, %f504;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs214,%rs213,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs210,%rs209,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+16432], {%rs210, %rs214};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs217, %f464;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs221, %f463;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs222,%rs221,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs218,%rs217,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+196608], {%rs218, %rs222};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs225, %f462;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs229, %f461;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs230,%rs229,%rs262;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs226,%rs225,%rs261;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+196624], {%rs226, %rs230};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs233, %f460;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs237, %f459;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs238,%rs237,%rs258;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs234,%rs233,%rs257;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+196640], {%rs234, %rs238};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs241, %f458;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs245, %f457;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs246,%rs245,%rs260;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs242,%rs241,%rs259;
}
	// end inline asm
	st.global.v2.u16 	[%rd20+196656], {%rs242, %rs246};
	// begin inline asm
	{  cvt.rn.f16.f32 %rs249, %f489;}

	// end inline asm
	// begin inline asm
	{  cvt.rn.f16.f32 %rs253, %f490;}

	// end inline asm
	// begin inline asm
	{add.f16 %rs254,%rs253,%rs264;
}
	// end inline asm
	// begin inline asm
	{add.f16 %rs250,%rs249,%rs263;
}
	// end inline asm
	st.global.v2.u16 	[%rd20], {%rs250, %rs254};
	ret;

}