Mini Kabibi Habibi
A ]� (~ M�AӨ�&�Oct 1 202004:12:00HOST64sm_35//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: UNKNOWN
// Unknown Toolkit Version
// Based on LLVM 3.4svn
//
.version 7.1
.target sm_35, texmode_independent
.address_size 64
// .globl imageWritei1D
.entry imageWritei1D(
.param .align 16 .b8 imageWritei1D_param_0[16],
.param .surfref imageWritei1D_param_1,
.param .u32 imageWritei1D_param_2,
.param .u32 imageWritei1D_param_3
)
{
.reg .pred %p<7>;
.reg .b32 %r<43>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r29, %r41, %r42, %r32}, [imageWritei1D_param_0];
ld.param.u32 %r28, [imageWritei1D_param_2];
ld.param.u32 %r33, [imageWritei1D_param_3];
mov.b32 %r34, %envreg3;
mov.u32 %r35, %ctaid.x;
mov.u32 %r36, %ntid.x;
mad.lo.s32 %r37, %r35, %r36, %r34;
mov.u32 %r38, %tid.x;
add.s32 %r1, %r37, %r38;
setp.ge.s32 %p1, %r1, %r33;
@%p1 bra BB0_15;
add.s32 %r2, %r1, %r28;
suq.channel_order.b32 %r3, [imageWritei1D_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB0_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB0_10;
bra.uni BB0_3;
BB0_10:
mov.u32 %r39, %r32;
bra.uni BB0_11;
BB0_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB0_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB0_8;
mov.u32 %r39, %r32;
mov.u32 %r40, %r29;
bra.uni BB0_14;
BB0_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB0_4;
bra.uni BB0_8;
BB0_4:
mov.u32 %r39, %r29;
mov.u32 %r40, %r32;
bra.uni BB0_12;
BB0_6:
mov.u32 %r39, %r42;
mov.u32 %r40, %r41;
mov.u32 %r41, %r29;
bra.uni BB0_13;
BB0_8:
mov.u32 %r39, %r29;
BB0_11:
mov.u32 %r40, %r41;
BB0_12:
mov.u32 %r41, %r42;
BB0_13:
mov.u32 %r42, %r32;
BB0_14:
sust.p.1d.v4.b32.trap [imageWritei1D_param_1, {%r2}], {%r39, %r40, %r41, %r42};
BB0_15:
ret;
}
// .globl imageWritei1D_arr
.entry imageWritei1D_arr(
.param .align 16 .b8 imageWritei1D_arr_param_0[16],
.param .surfref imageWritei1D_arr_param_1,
.param .u32 imageWritei1D_arr_param_2,
.param .u32 imageWritei1D_arr_param_3,
.param .u32 imageWritei1D_arr_param_4,
.param .u32 imageWritei1D_arr_param_5
)
{
.reg .pred %p<9>;
.reg .b32 %r<52>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r32, %r50, %r51, %r35}, [imageWritei1D_arr_param_0];
ld.param.u32 %r30, [imageWritei1D_arr_param_2];
ld.param.u32 %r36, [imageWritei1D_arr_param_3];
ld.param.u32 %r31, [imageWritei1D_arr_param_4];
ld.param.u32 %r37, [imageWritei1D_arr_param_5];
mov.b32 %r38, %envreg3;
mov.u32 %r39, %ctaid.x;
mov.u32 %r40, %ntid.x;
mad.lo.s32 %r41, %r39, %r40, %r38;
mov.u32 %r42, %tid.x;
add.s32 %r1, %r41, %r42;
mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ntid.y;
mov.b32 %r45, %envreg4;
mad.lo.s32 %r46, %r43, %r44, %r45;
mov.u32 %r47, %tid.y;
add.s32 %r2, %r46, %r47;
setp.ge.s32 %p1, %r1, %r36;
setp.ge.s32 %p2, %r2, %r37;
or.pred %p3, %p1, %p2;
@%p3 bra BB1_15;
add.s32 %r3, %r1, %r30;
add.s32 %r4, %r2, %r31;
suq.channel_order.b32 %r5, [imageWritei1D_arr_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB1_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB1_10;
bra.uni BB1_3;
BB1_10:
mov.u32 %r48, %r35;
bra.uni BB1_11;
BB1_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB1_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB1_8;
mov.u32 %r48, %r35;
mov.u32 %r49, %r32;
bra.uni BB1_14;
BB1_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB1_4;
bra.uni BB1_8;
BB1_4:
mov.u32 %r48, %r32;
mov.u32 %r49, %r35;
bra.uni BB1_12;
BB1_6:
mov.u32 %r48, %r51;
mov.u32 %r49, %r50;
mov.u32 %r50, %r32;
bra.uni BB1_13;
BB1_8:
mov.u32 %r48, %r32;
BB1_11:
mov.u32 %r49, %r50;
BB1_12:
mov.u32 %r50, %r51;
BB1_13:
mov.u32 %r51, %r35;
BB1_14:
sust.p.a1d.v4.b32.trap [imageWritei1D_arr_param_1, {%r4, %r3}], {%r48, %r49, %r50, %r51};
BB1_15:
ret;
}
// .globl imageWritei1D_buf
.entry imageWritei1D_buf(
.param .align 16 .b8 imageWritei1D_buf_param_0[16],
.param .surfref imageWritei1D_buf_param_1,
.param .u32 imageWritei1D_buf_param_2,
.param .u32 imageWritei1D_buf_param_3
)
{
.reg .pred %p<7>;
.reg .b32 %r<43>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r29, %r41, %r42, %r32}, [imageWritei1D_buf_param_0];
ld.param.u32 %r28, [imageWritei1D_buf_param_2];
ld.param.u32 %r33, [imageWritei1D_buf_param_3];
mov.b32 %r34, %envreg3;
mov.u32 %r35, %ctaid.x;
mov.u32 %r36, %ntid.x;
mad.lo.s32 %r37, %r35, %r36, %r34;
mov.u32 %r38, %tid.x;
add.s32 %r1, %r37, %r38;
setp.ge.s32 %p1, %r1, %r33;
@%p1 bra BB2_15;
add.s32 %r2, %r1, %r28;
suq.channel_order.b32 %r3, [imageWritei1D_buf_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB2_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB2_10;
bra.uni BB2_3;
BB2_10:
mov.u32 %r39, %r32;
bra.uni BB2_11;
BB2_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB2_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB2_8;
mov.u32 %r39, %r32;
mov.u32 %r40, %r29;
bra.uni BB2_14;
BB2_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB2_4;
bra.uni BB2_8;
BB2_4:
mov.u32 %r39, %r29;
mov.u32 %r40, %r32;
bra.uni BB2_12;
BB2_6:
mov.u32 %r39, %r42;
mov.u32 %r40, %r41;
mov.u32 %r41, %r29;
bra.uni BB2_13;
BB2_8:
mov.u32 %r39, %r29;
BB2_11:
mov.u32 %r40, %r41;
BB2_12:
mov.u32 %r41, %r42;
BB2_13:
mov.u32 %r42, %r32;
BB2_14:
sust.p.1d_buffer.v4.b32.trap [imageWritei1D_buf_param_1, {%r2}], {%r39, %r40, %r41, %r42};
BB2_15:
ret;
}
// .globl imageWritef1D
.entry imageWritef1D(
.param .align 16 .b8 imageWritef1D_param_0[16],
.param .surfref imageWritef1D_param_1,
.param .u32 imageWritef1D_param_2,
.param .u32 imageWritef1D_param_3
)
{
.reg .pred %p<7>;
.reg .f32 %f<33>;
.reg .b32 %r<15>;
.reg .b64 %rd<2>;
ld.param.v4.f32 {%f25, %f31, %f32, %f28}, [imageWritef1D_param_0];
ld.param.u32 %r4, [imageWritef1D_param_2];
ld.param.u32 %r5, [imageWritef1D_param_3];
mov.b32 %r6, %envreg3;
mov.u32 %r7, %ctaid.x;
mov.u32 %r8, %ntid.x;
mad.lo.s32 %r9, %r7, %r8, %r6;
mov.u32 %r10, %tid.x;
add.s32 %r1, %r9, %r10;
setp.ge.s32 %p1, %r1, %r5;
@%p1 bra BB3_12;
add.s32 %r2, %r1, %r4;
suq.channel_order.b32 %r3, [imageWritef1D_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB3_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB3_10;
bra.uni BB3_3;
BB3_10:
mov.f32 %f30, 0f00000000;
mov.f32 %f29, %f28;
mov.f32 %f31, %f30;
mov.f32 %f32, %f30;
bra.uni BB3_11;
BB3_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB3_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB3_8;
mov.f32 %f29, %f28;
mov.f32 %f30, %f25;
bra.uni BB3_11;
BB3_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB3_4;
bra.uni BB3_8;
BB3_4:
mov.f32 %f31, 0f00000000;
mov.f32 %f29, %f25;
mov.f32 %f30, %f28;
mov.f32 %f32, %f31;
bra.uni BB3_11;
BB3_6:
mov.f32 %f29, %f32;
mov.f32 %f30, %f31;
mov.f32 %f31, %f25;
mov.f32 %f32, %f28;
bra.uni BB3_11;
BB3_8:
mov.f32 %f29, %f25;
mov.f32 %f30, %f31;
mov.f32 %f31, %f32;
mov.f32 %f32, %f28;
BB3_11:
mov.b32 %r11, %f29;
mov.b32 %r12, %f30;
mov.b32 %r13, %f31;
mov.b32 %r14, %f32;
sust.p.1d.v4.b32.trap [imageWritef1D_param_1, {%r2}], {%r11, %r12, %r13, %r14};
BB3_12:
ret;
}
// .globl imageWritef1D_arr
.entry imageWritef1D_arr(
.param .align 16 .b8 imageWritef1D_arr_param_0[16],
.param .surfref imageWritef1D_arr_param_1,
.param .u32 imageWritef1D_arr_param_2,
.param .u32 imageWritef1D_arr_param_3,
.param .u32 imageWritef1D_arr_param_4,
.param .u32 imageWritef1D_arr_param_5
)
{
.reg .pred %p<9>;
.reg .f32 %f<33>;
.reg .b32 %r<24>;
.reg .b64 %rd<2>;
ld.param.v4.f32 {%f25, %f31, %f32, %f28}, [imageWritef1D_arr_param_0];
ld.param.u32 %r6, [imageWritef1D_arr_param_2];
ld.param.u32 %r8, [imageWritef1D_arr_param_3];
ld.param.u32 %r7, [imageWritef1D_arr_param_4];
ld.param.u32 %r9, [imageWritef1D_arr_param_5];
mov.b32 %r10, %envreg3;
mov.u32 %r11, %ctaid.x;
mov.u32 %r12, %ntid.x;
mad.lo.s32 %r13, %r11, %r12, %r10;
mov.u32 %r14, %tid.x;
add.s32 %r1, %r13, %r14;
mov.u32 %r15, %ctaid.y;
mov.u32 %r16, %ntid.y;
mov.b32 %r17, %envreg4;
mad.lo.s32 %r18, %r15, %r16, %r17;
mov.u32 %r19, %tid.y;
add.s32 %r2, %r18, %r19;
setp.ge.s32 %p1, %r1, %r8;
setp.ge.s32 %p2, %r2, %r9;
or.pred %p3, %p1, %p2;
@%p3 bra BB4_12;
add.s32 %r3, %r1, %r6;
add.s32 %r4, %r2, %r7;
suq.channel_order.b32 %r5, [imageWritef1D_arr_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB4_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB4_10;
bra.uni BB4_3;
BB4_10:
mov.f32 %f30, 0f00000000;
mov.f32 %f29, %f28;
mov.f32 %f31, %f30;
mov.f32 %f32, %f30;
bra.uni BB4_11;
BB4_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB4_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB4_8;
mov.f32 %f29, %f28;
mov.f32 %f30, %f25;
bra.uni BB4_11;
BB4_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB4_4;
bra.uni BB4_8;
BB4_4:
mov.f32 %f31, 0f00000000;
mov.f32 %f29, %f25;
mov.f32 %f30, %f28;
mov.f32 %f32, %f31;
bra.uni BB4_11;
BB4_6:
mov.f32 %f29, %f32;
mov.f32 %f30, %f31;
mov.f32 %f31, %f25;
mov.f32 %f32, %f28;
bra.uni BB4_11;
BB4_8:
mov.f32 %f29, %f25;
mov.f32 %f30, %f31;
mov.f32 %f31, %f32;
mov.f32 %f32, %f28;
BB4_11:
mov.b32 %r20, %f29;
mov.b32 %r21, %f30;
mov.b32 %r22, %f31;
mov.b32 %r23, %f32;
sust.p.a1d.v4.b32.trap [imageWritef1D_arr_param_1, {%r4, %r3}], {%r20, %r21, %r22, %r23};
BB4_12:
ret;
}
// .globl imageWritef1D_buf
.entry imageWritef1D_buf(
.param .align 16 .b8 imageWritef1D_buf_param_0[16],
.param .surfref imageWritef1D_buf_param_1,
.param .u32 imageWritef1D_buf_param_2,
.param .u32 imageWritef1D_buf_param_3
)
{
.reg .pred %p<7>;
.reg .f32 %f<33>;
.reg .b32 %r<15>;
.reg .b64 %rd<2>;
ld.param.v4.f32 {%f25, %f31, %f32, %f28}, [imageWritef1D_buf_param_0];
ld.param.u32 %r4, [imageWritef1D_buf_param_2];
ld.param.u32 %r5, [imageWritef1D_buf_param_3];
mov.b32 %r6, %envreg3;
mov.u32 %r7, %ctaid.x;
mov.u32 %r8, %ntid.x;
mad.lo.s32 %r9, %r7, %r8, %r6;
mov.u32 %r10, %tid.x;
add.s32 %r1, %r9, %r10;
setp.ge.s32 %p1, %r1, %r5;
@%p1 bra BB5_12;
add.s32 %r2, %r1, %r4;
suq.channel_order.b32 %r3, [imageWritef1D_buf_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB5_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB5_10;
bra.uni BB5_3;
BB5_10:
mov.f32 %f30, 0f00000000;
mov.f32 %f29, %f28;
mov.f32 %f31, %f30;
mov.f32 %f32, %f30;
bra.uni BB5_11;
BB5_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB5_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB5_8;
mov.f32 %f29, %f28;
mov.f32 %f30, %f25;
bra.uni BB5_11;
BB5_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB5_4;
bra.uni BB5_8;
BB5_4:
mov.f32 %f31, 0f00000000;
mov.f32 %f29, %f25;
mov.f32 %f30, %f28;
mov.f32 %f32, %f31;
bra.uni BB5_11;
BB5_6:
mov.f32 %f29, %f32;
mov.f32 %f30, %f31;
mov.f32 %f31, %f25;
mov.f32 %f32, %f28;
bra.uni BB5_11;
BB5_8:
mov.f32 %f29, %f25;
mov.f32 %f30, %f31;
mov.f32 %f31, %f32;
mov.f32 %f32, %f28;
BB5_11:
mov.b32 %r11, %f29;
mov.b32 %r12, %f30;
mov.b32 %r13, %f31;
mov.b32 %r14, %f32;
sust.p.1d_buffer.v4.b32.trap [imageWritef1D_buf_param_1, {%r2}], {%r11, %r12, %r13, %r14};
BB5_12:
ret;
}
// .globl imageWriteui1D
.entry imageWriteui1D(
.param .align 16 .b8 imageWriteui1D_param_0[16],
.param .surfref imageWriteui1D_param_1,
.param .u32 imageWriteui1D_param_2,
.param .u32 imageWriteui1D_param_3
)
{
.reg .pred %p<7>;
.reg .b32 %r<43>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r29, %r41, %r42, %r32}, [imageWriteui1D_param_0];
ld.param.u32 %r28, [imageWriteui1D_param_2];
ld.param.u32 %r33, [imageWriteui1D_param_3];
mov.b32 %r34, %envreg3;
mov.u32 %r35, %ctaid.x;
mov.u32 %r36, %ntid.x;
mad.lo.s32 %r37, %r35, %r36, %r34;
mov.u32 %r38, %tid.x;
add.s32 %r1, %r37, %r38;
setp.ge.s32 %p1, %r1, %r33;
@%p1 bra BB6_15;
add.s32 %r2, %r1, %r28;
suq.channel_order.b32 %r3, [imageWriteui1D_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB6_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB6_10;
bra.uni BB6_3;
BB6_10:
mov.u32 %r39, %r32;
bra.uni BB6_11;
BB6_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB6_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB6_8;
mov.u32 %r39, %r32;
mov.u32 %r40, %r29;
bra.uni BB6_14;
BB6_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB6_4;
bra.uni BB6_8;
BB6_4:
mov.u32 %r39, %r29;
mov.u32 %r40, %r32;
bra.uni BB6_12;
BB6_6:
mov.u32 %r39, %r42;
mov.u32 %r40, %r41;
mov.u32 %r41, %r29;
bra.uni BB6_13;
BB6_8:
mov.u32 %r39, %r29;
BB6_11:
mov.u32 %r40, %r41;
BB6_12:
mov.u32 %r41, %r42;
BB6_13:
mov.u32 %r42, %r32;
BB6_14:
sust.p.1d.v4.b32.trap [imageWriteui1D_param_1, {%r2}], {%r39, %r40, %r41, %r42};
BB6_15:
ret;
}
// .globl imageWriteui1D_arr
.entry imageWriteui1D_arr(
.param .align 16 .b8 imageWriteui1D_arr_param_0[16],
.param .surfref imageWriteui1D_arr_param_1,
.param .u32 imageWriteui1D_arr_param_2,
.param .u32 imageWriteui1D_arr_param_3,
.param .u32 imageWriteui1D_arr_param_4,
.param .u32 imageWriteui1D_arr_param_5
)
{
.reg .pred %p<9>;
.reg .b32 %r<52>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r32, %r50, %r51, %r35}, [imageWriteui1D_arr_param_0];
ld.param.u32 %r30, [imageWriteui1D_arr_param_2];
ld.param.u32 %r36, [imageWriteui1D_arr_param_3];
ld.param.u32 %r31, [imageWriteui1D_arr_param_4];
ld.param.u32 %r37, [imageWriteui1D_arr_param_5];
mov.b32 %r38, %envreg3;
mov.u32 %r39, %ctaid.x;
mov.u32 %r40, %ntid.x;
mad.lo.s32 %r41, %r39, %r40, %r38;
mov.u32 %r42, %tid.x;
add.s32 %r1, %r41, %r42;
mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ntid.y;
mov.b32 %r45, %envreg4;
mad.lo.s32 %r46, %r43, %r44, %r45;
mov.u32 %r47, %tid.y;
add.s32 %r2, %r46, %r47;
setp.ge.s32 %p1, %r1, %r36;
setp.ge.s32 %p2, %r2, %r37;
or.pred %p3, %p1, %p2;
@%p3 bra BB7_15;
add.s32 %r3, %r1, %r30;
add.s32 %r4, %r2, %r31;
suq.channel_order.b32 %r5, [imageWriteui1D_arr_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB7_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB7_10;
bra.uni BB7_3;
BB7_10:
mov.u32 %r48, %r35;
bra.uni BB7_11;
BB7_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB7_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB7_8;
mov.u32 %r48, %r35;
mov.u32 %r49, %r32;
bra.uni BB7_14;
BB7_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB7_4;
bra.uni BB7_8;
BB7_4:
mov.u32 %r48, %r32;
mov.u32 %r49, %r35;
bra.uni BB7_12;
BB7_6:
mov.u32 %r48, %r51;
mov.u32 %r49, %r50;
mov.u32 %r50, %r32;
bra.uni BB7_13;
BB7_8:
mov.u32 %r48, %r32;
BB7_11:
mov.u32 %r49, %r50;
BB7_12:
mov.u32 %r50, %r51;
BB7_13:
mov.u32 %r51, %r35;
BB7_14:
sust.p.a1d.v4.b32.trap [imageWriteui1D_arr_param_1, {%r4, %r3}], {%r48, %r49, %r50, %r51};
BB7_15:
ret;
}
// .globl imageWriteui1D_buf
.entry imageWriteui1D_buf(
.param .align 16 .b8 imageWriteui1D_buf_param_0[16],
.param .surfref imageWriteui1D_buf_param_1,
.param .u32 imageWriteui1D_buf_param_2,
.param .u32 imageWriteui1D_buf_param_3
)
{
.reg .pred %p<7>;
.reg .b32 %r<43>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r29, %r41, %r42, %r32}, [imageWriteui1D_buf_param_0];
ld.param.u32 %r28, [imageWriteui1D_buf_param_2];
ld.param.u32 %r33, [imageWriteui1D_buf_param_3];
mov.b32 %r34, %envreg3;
mov.u32 %r35, %ctaid.x;
mov.u32 %r36, %ntid.x;
mad.lo.s32 %r37, %r35, %r36, %r34;
mov.u32 %r38, %tid.x;
add.s32 %r1, %r37, %r38;
setp.ge.s32 %p1, %r1, %r33;
@%p1 bra BB8_15;
add.s32 %r2, %r1, %r28;
suq.channel_order.b32 %r3, [imageWriteui1D_buf_param_1];
setp.gt.s32 %p2, %r3, 4277;
@%p2 bra BB8_5;
setp.eq.s32 %p5, %r3, 4273;
@%p5 bra BB8_10;
bra.uni BB8_3;
BB8_10:
mov.u32 %r39, %r32;
bra.uni BB8_11;
BB8_5:
setp.eq.s32 %p3, %r3, 4278;
@%p3 bra BB8_6;
setp.ne.s32 %p4, %r3, 4279;
@%p4 bra BB8_8;
mov.u32 %r39, %r32;
mov.u32 %r40, %r29;
bra.uni BB8_14;
BB8_3:
setp.eq.s32 %p6, %r3, 4275;
@%p6 bra BB8_4;
bra.uni BB8_8;
BB8_4:
mov.u32 %r39, %r29;
mov.u32 %r40, %r32;
bra.uni BB8_12;
BB8_6:
mov.u32 %r39, %r42;
mov.u32 %r40, %r41;
mov.u32 %r41, %r29;
bra.uni BB8_13;
BB8_8:
mov.u32 %r39, %r29;
BB8_11:
mov.u32 %r40, %r41;
BB8_12:
mov.u32 %r41, %r42;
BB8_13:
mov.u32 %r42, %r32;
BB8_14:
sust.p.1d_buffer.v4.b32.trap [imageWriteui1D_buf_param_1, {%r2}], {%r39, %r40, %r41, %r42};
BB8_15:
ret;
}
// .globl imageWritei2D
.entry imageWritei2D(
.param .align 16 .b8 imageWritei2D_param_0[16],
.param .surfref imageWritei2D_param_1,
.param .u32 imageWritei2D_param_2,
.param .u32 imageWritei2D_param_3,
.param .u32 imageWritei2D_param_4,
.param .u32 imageWritei2D_param_5
)
{
.reg .pred %p<9>;
.reg .b32 %r<52>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r32, %r50, %r51, %r35}, [imageWritei2D_param_0];
ld.param.u32 %r30, [imageWritei2D_param_2];
ld.param.u32 %r36, [imageWritei2D_param_3];
ld.param.u32 %r31, [imageWritei2D_param_4];
ld.param.u32 %r37, [imageWritei2D_param_5];
mov.b32 %r38, %envreg3;
mov.u32 %r39, %ctaid.x;
mov.u32 %r40, %ntid.x;
mad.lo.s32 %r41, %r39, %r40, %r38;
mov.u32 %r42, %tid.x;
add.s32 %r1, %r41, %r42;
mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ntid.y;
mov.b32 %r45, %envreg4;
mad.lo.s32 %r46, %r43, %r44, %r45;
mov.u32 %r47, %tid.y;
add.s32 %r2, %r46, %r47;
setp.ge.s32 %p1, %r1, %r36;
setp.ge.s32 %p2, %r2, %r37;
or.pred %p3, %p1, %p2;
@%p3 bra BB9_15;
add.s32 %r3, %r1, %r30;
add.s32 %r4, %r2, %r31;
suq.channel_order.b32 %r5, [imageWritei2D_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB9_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB9_10;
bra.uni BB9_3;
BB9_10:
mov.u32 %r48, %r35;
bra.uni BB9_11;
BB9_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB9_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB9_8;
mov.u32 %r48, %r35;
mov.u32 %r49, %r32;
bra.uni BB9_14;
BB9_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB9_4;
bra.uni BB9_8;
BB9_4:
mov.u32 %r48, %r32;
mov.u32 %r49, %r35;
bra.uni BB9_12;
BB9_6:
mov.u32 %r48, %r51;
mov.u32 %r49, %r50;
mov.u32 %r50, %r32;
bra.uni BB9_13;
BB9_8:
mov.u32 %r48, %r32;
BB9_11:
mov.u32 %r49, %r50;
BB9_12:
mov.u32 %r50, %r51;
BB9_13:
mov.u32 %r51, %r35;
BB9_14:
sust.p.2d.v4.b32.trap [imageWritei2D_param_1, {%r3, %r4}], {%r48, %r49, %r50, %r51};
BB9_15:
ret;
}
// .globl imageWritei2D_arr
.entry imageWritei2D_arr(
.param .align 16 .b8 imageWritei2D_arr_param_0[16],
.param .surfref imageWritei2D_arr_param_1,
.param .u32 imageWritei2D_arr_param_2,
.param .u32 imageWritei2D_arr_param_3,
.param .u32 imageWritei2D_arr_param_4,
.param .u32 imageWritei2D_arr_param_5,
.param .u32 imageWritei2D_arr_param_6,
.param .u32 imageWritei2D_arr_param_7
)
{
.reg .pred %p<11>;
.reg .b32 %r<61>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r35, %r59, %r60, %r38}, [imageWritei2D_arr_param_0];
ld.param.u32 %r32, [imageWritei2D_arr_param_2];
ld.param.u32 %r39, [imageWritei2D_arr_param_3];
ld.param.u32 %r33, [imageWritei2D_arr_param_4];
ld.param.u32 %r40, [imageWritei2D_arr_param_5];
ld.param.u32 %r34, [imageWritei2D_arr_param_6];
ld.param.u32 %r41, [imageWritei2D_arr_param_7];
mov.b32 %r42, %envreg3;
mov.u32 %r43, %ctaid.x;
mov.u32 %r44, %ntid.x;
mad.lo.s32 %r45, %r43, %r44, %r42;
mov.u32 %r46, %tid.x;
add.s32 %r1, %r45, %r46;
mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ntid.y;
mov.b32 %r49, %envreg4;
mad.lo.s32 %r50, %r47, %r48, %r49;
mov.u32 %r51, %tid.y;
add.s32 %r2, %r50, %r51;
mov.u32 %r52, %ctaid.z;
mov.u32 %r53, %ntid.z;
mov.b32 %r54, %envreg5;
mad.lo.s32 %r55, %r52, %r53, %r54;
mov.u32 %r56, %tid.z;
add.s32 %r3, %r55, %r56;
setp.ge.s32 %p1, %r1, %r39;
setp.ge.s32 %p2, %r2, %r40;
or.pred %p3, %p1, %p2;
setp.ge.s32 %p4, %r3, %r41;
or.pred %p5, %p3, %p4;
@%p5 bra BB10_15;
add.s32 %r4, %r1, %r32;
add.s32 %r5, %r2, %r33;
add.s32 %r6, %r3, %r34;
suq.channel_order.b32 %r7, [imageWritei2D_arr_param_1];
setp.gt.s32 %p6, %r7, 4277;
@%p6 bra BB10_5;
setp.eq.s32 %p9, %r7, 4273;
@%p9 bra BB10_10;
bra.uni BB10_3;
BB10_10:
mov.u32 %r57, %r38;
bra.uni BB10_11;
BB10_5:
setp.eq.s32 %p7, %r7, 4278;
@%p7 bra BB10_6;
setp.ne.s32 %p8, %r7, 4279;
@%p8 bra BB10_8;
mov.u32 %r57, %r38;
mov.u32 %r58, %r35;
bra.uni BB10_14;
BB10_3:
setp.eq.s32 %p10, %r7, 4275;
@%p10 bra BB10_4;
bra.uni BB10_8;
BB10_4:
mov.u32 %r57, %r35;
mov.u32 %r58, %r38;
bra.uni BB10_12;
BB10_6:
mov.u32 %r57, %r60;
mov.u32 %r58, %r59;
mov.u32 %r59, %r35;
bra.uni BB10_13;
BB10_8:
mov.u32 %r57, %r35;
BB10_11:
mov.u32 %r58, %r59;
BB10_12:
mov.u32 %r59, %r60;
BB10_13:
mov.u32 %r60, %r38;
BB10_14:
sust.p.a2d.v4.b32.trap [imageWritei2D_arr_param_1, {%r6, %r4, %r5, %r5}], {%r57, %r58, %r59, %r60};
BB10_15:
ret;
}
// .globl imageWritef2D
.entry imageWritef2D(
.param .align 16 .b8 imageWritef2D_param_0[16],
.param .surfref imageWritef2D_param_1,
.param .u32 imageWritef2D_param_2,
.param .u32 imageWritef2D_param_3,
.param .u32 imageWritef2D_param_4,
.param .u32 imageWritef2D_param_5
)
{
.reg .pred %p<9>;
.reg .f32 %f<33>;
.reg .b32 %r<24>;
.reg .b64 %rd<2>;
ld.param.v4.f32 {%f25, %f31, %f32, %f28}, [imageWritef2D_param_0];
ld.param.u32 %r6, [imageWritef2D_param_2];
ld.param.u32 %r8, [imageWritef2D_param_3];
ld.param.u32 %r7, [imageWritef2D_param_4];
ld.param.u32 %r9, [imageWritef2D_param_5];
mov.b32 %r10, %envreg3;
mov.u32 %r11, %ctaid.x;
mov.u32 %r12, %ntid.x;
mad.lo.s32 %r13, %r11, %r12, %r10;
mov.u32 %r14, %tid.x;
add.s32 %r1, %r13, %r14;
mov.u32 %r15, %ctaid.y;
mov.u32 %r16, %ntid.y;
mov.b32 %r17, %envreg4;
mad.lo.s32 %r18, %r15, %r16, %r17;
mov.u32 %r19, %tid.y;
add.s32 %r2, %r18, %r19;
setp.ge.s32 %p1, %r1, %r8;
setp.ge.s32 %p2, %r2, %r9;
or.pred %p3, %p1, %p2;
@%p3 bra BB11_12;
add.s32 %r3, %r1, %r6;
add.s32 %r4, %r2, %r7;
suq.channel_order.b32 %r5, [imageWritef2D_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB11_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB11_10;
bra.uni BB11_3;
BB11_10:
mov.f32 %f30, 0f00000000;
mov.f32 %f29, %f28;
mov.f32 %f31, %f30;
mov.f32 %f32, %f30;
bra.uni BB11_11;
BB11_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB11_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB11_8;
mov.f32 %f29, %f28;
mov.f32 %f30, %f25;
bra.uni BB11_11;
BB11_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB11_4;
bra.uni BB11_8;
BB11_4:
mov.f32 %f31, 0f00000000;
mov.f32 %f29, %f25;
mov.f32 %f30, %f28;
mov.f32 %f32, %f31;
bra.uni BB11_11;
BB11_6:
mov.f32 %f29, %f32;
mov.f32 %f30, %f31;
mov.f32 %f31, %f25;
mov.f32 %f32, %f28;
bra.uni BB11_11;
BB11_8:
mov.f32 %f29, %f25;
mov.f32 %f30, %f31;
mov.f32 %f31, %f32;
mov.f32 %f32, %f28;
BB11_11:
mov.b32 %r20, %f29;
mov.b32 %r21, %f30;
mov.b32 %r22, %f31;
mov.b32 %r23, %f32;
sust.p.2d.v4.b32.trap [imageWritef2D_param_1, {%r3, %r4}], {%r20, %r21, %r22, %r23};
BB11_12:
ret;
}
// .globl imageWritef2D_arr
.entry imageWritef2D_arr(
.param .align 16 .b8 imageWritef2D_arr_param_0[16],
.param .surfref imageWritef2D_arr_param_1,
.param .u32 imageWritef2D_arr_param_2,
.param .u32 imageWritef2D_arr_param_3,
.param .u32 imageWritef2D_arr_param_4,
.param .u32 imageWritef2D_arr_param_5,
.param .u32 imageWritef2D_arr_param_6,
.param .u32 imageWritef2D_arr_param_7
)
{
.reg .pred %p<11>;
.reg .f32 %f<33>;
.reg .b32 %r<33>;
.reg .b64 %rd<2>;
ld.param.v4.f32 {%f25, %f31, %f32, %f28}, [imageWritef2D_arr_param_0];
ld.param.u32 %r8, [imageWritef2D_arr_param_2];
ld.param.u32 %r11, [imageWritef2D_arr_param_3];
ld.param.u32 %r9, [imageWritef2D_arr_param_4];
ld.param.u32 %r12, [imageWritef2D_arr_param_5];
ld.param.u32 %r10, [imageWritef2D_arr_param_6];
ld.param.u32 %r13, [imageWritef2D_arr_param_7];
mov.b32 %r14, %envreg3;
mov.u32 %r15, %ctaid.x;
mov.u32 %r16, %ntid.x;
mad.lo.s32 %r17, %r15, %r16, %r14;
mov.u32 %r18, %tid.x;
add.s32 %r1, %r17, %r18;
mov.u32 %r19, %ctaid.y;
mov.u32 %r20, %ntid.y;
mov.b32 %r21, %envreg4;
mad.lo.s32 %r22, %r19, %r20, %r21;
mov.u32 %r23, %tid.y;
add.s32 %r2, %r22, %r23;
mov.u32 %r24, %ctaid.z;
mov.u32 %r25, %ntid.z;
mov.b32 %r26, %envreg5;
mad.lo.s32 %r27, %r24, %r25, %r26;
mov.u32 %r28, %tid.z;
add.s32 %r3, %r27, %r28;
setp.ge.s32 %p1, %r1, %r11;
setp.ge.s32 %p2, %r2, %r12;
or.pred %p3, %p1, %p2;
setp.ge.s32 %p4, %r3, %r13;
or.pred %p5, %p3, %p4;
@%p5 bra BB12_12;
add.s32 %r4, %r1, %r8;
add.s32 %r5, %r2, %r9;
add.s32 %r6, %r3, %r10;
suq.channel_order.b32 %r7, [imageWritef2D_arr_param_1];
setp.gt.s32 %p6, %r7, 4277;
@%p6 bra BB12_5;
setp.eq.s32 %p9, %r7, 4273;
@%p9 bra BB12_10;
bra.uni BB12_3;
BB12_10:
mov.f32 %f30, 0f00000000;
mov.f32 %f29, %f28;
mov.f32 %f31, %f30;
mov.f32 %f32, %f30;
bra.uni BB12_11;
BB12_5:
setp.eq.s32 %p7, %r7, 4278;
@%p7 bra BB12_6;
setp.ne.s32 %p8, %r7, 4279;
@%p8 bra BB12_8;
mov.f32 %f29, %f28;
mov.f32 %f30, %f25;
bra.uni BB12_11;
BB12_3:
setp.eq.s32 %p10, %r7, 4275;
@%p10 bra BB12_4;
bra.uni BB12_8;
BB12_4:
mov.f32 %f31, 0f00000000;
mov.f32 %f29, %f25;
mov.f32 %f30, %f28;
mov.f32 %f32, %f31;
bra.uni BB12_11;
BB12_6:
mov.f32 %f29, %f32;
mov.f32 %f30, %f31;
mov.f32 %f31, %f25;
mov.f32 %f32, %f28;
bra.uni BB12_11;
BB12_8:
mov.f32 %f29, %f25;
mov.f32 %f30, %f31;
mov.f32 %f31, %f32;
mov.f32 %f32, %f28;
BB12_11:
mov.b32 %r29, %f29;
mov.b32 %r30, %f30;
mov.b32 %r31, %f31;
mov.b32 %r32, %f32;
sust.p.a2d.v4.b32.trap [imageWritef2D_arr_param_1, {%r6, %r4, %r5, %r5}], {%r29, %r30, %r31, %r32};
BB12_12:
ret;
}
// .globl imageWriteui2D
.entry imageWriteui2D(
.param .align 16 .b8 imageWriteui2D_param_0[16],
.param .surfref imageWriteui2D_param_1,
.param .u32 imageWriteui2D_param_2,
.param .u32 imageWriteui2D_param_3,
.param .u32 imageWriteui2D_param_4,
.param .u32 imageWriteui2D_param_5
)
{
.reg .pred %p<9>;
.reg .b32 %r<52>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r32, %r50, %r51, %r35}, [imageWriteui2D_param_0];
ld.param.u32 %r30, [imageWriteui2D_param_2];
ld.param.u32 %r36, [imageWriteui2D_param_3];
ld.param.u32 %r31, [imageWriteui2D_param_4];
ld.param.u32 %r37, [imageWriteui2D_param_5];
mov.b32 %r38, %envreg3;
mov.u32 %r39, %ctaid.x;
mov.u32 %r40, %ntid.x;
mad.lo.s32 %r41, %r39, %r40, %r38;
mov.u32 %r42, %tid.x;
add.s32 %r1, %r41, %r42;
mov.u32 %r43, %ctaid.y;
mov.u32 %r44, %ntid.y;
mov.b32 %r45, %envreg4;
mad.lo.s32 %r46, %r43, %r44, %r45;
mov.u32 %r47, %tid.y;
add.s32 %r2, %r46, %r47;
setp.ge.s32 %p1, %r1, %r36;
setp.ge.s32 %p2, %r2, %r37;
or.pred %p3, %p1, %p2;
@%p3 bra BB13_15;
add.s32 %r3, %r1, %r30;
add.s32 %r4, %r2, %r31;
suq.channel_order.b32 %r5, [imageWriteui2D_param_1];
setp.gt.s32 %p4, %r5, 4277;
@%p4 bra BB13_5;
setp.eq.s32 %p7, %r5, 4273;
@%p7 bra BB13_10;
bra.uni BB13_3;
BB13_10:
mov.u32 %r48, %r35;
bra.uni BB13_11;
BB13_5:
setp.eq.s32 %p5, %r5, 4278;
@%p5 bra BB13_6;
setp.ne.s32 %p6, %r5, 4279;
@%p6 bra BB13_8;
mov.u32 %r48, %r35;
mov.u32 %r49, %r32;
bra.uni BB13_14;
BB13_3:
setp.eq.s32 %p8, %r5, 4275;
@%p8 bra BB13_4;
bra.uni BB13_8;
BB13_4:
mov.u32 %r48, %r32;
mov.u32 %r49, %r35;
bra.uni BB13_12;
BB13_6:
mov.u32 %r48, %r51;
mov.u32 %r49, %r50;
mov.u32 %r50, %r32;
bra.uni BB13_13;
BB13_8:
mov.u32 %r48, %r32;
BB13_11:
mov.u32 %r49, %r50;
BB13_12:
mov.u32 %r50, %r51;
BB13_13:
mov.u32 %r51, %r35;
BB13_14:
sust.p.2d.v4.b32.trap [imageWriteui2D_param_1, {%r3, %r4}], {%r48, %r49, %r50, %r51};
BB13_15:
ret;
}
// .globl imageWriteui2D_arr
.entry imageWriteui2D_arr(
.param .align 16 .b8 imageWriteui2D_arr_param_0[16],
.param .surfref imageWriteui2D_arr_param_1,
.param .u32 imageWriteui2D_arr_param_2,
.param .u32 imageWriteui2D_arr_param_3,
.param .u32 imageWriteui2D_arr_param_4,
.param .u32 imageWriteui2D_arr_param_5,
.param .u32 imageWriteui2D_arr_param_6,
.param .u32 imageWriteui2D_arr_param_7
)
{
.reg .pred %p<11>;
.reg .b32 %r<61>;
.reg .b64 %rd<2>;
ld.param.v4.u32 {%r35, %r59, %r60, %r38}, [imageWriteui2D_arr_param_0];
ld.param.u32 %r32, [imageWriteui2D_arr_param_2];
ld.param.u32 %r39, [imageWriteui2D_arr_param_3];
ld.param.u32 %r33, [imageWriteui2D_arr_param_4];
ld.param.u32 %r40, [imageWriteui2D_arr_param_5];
ld.param.u32 %r34, [imageWriteui2D_arr_param_6];
ld.param.u32 %r41, [imageWriteui2D_arr_param_7];
mov.b32 %r42, %envreg3;
mov.u32 %r43, %ctaid.x;
mov.u32 %r44, %ntid.x;
mad.lo.s32 %r45, %r43, %r44, %r42;
mov.u32 %r46, %tid.x;
add.s32 %r1, %r45, %r46;
mov.u32 %r47, %ctaid.y;
mov.u32 %r48, %ntid.y;
mov.b32 %r49, %envreg4;
mad.lo.s32 %r50, %r47, %r48, %r49;
mov.u32 %r51, %tid.y;
add.s32 %r2, %r50, %r51;
mov.u32 %r52, %ctaid.z;
mov.u32 %r53, %ntid.z;
mov.b32 %r54, %envreg5;
mad.lo.s32 %r55, %r52, %r53, %r54;
mov.u32 %r56, %tid.z;
add.s32 %r3, %r55, %r56;
setp.ge.s32 %p1, %r1, %r39;
setp.ge.s32 %p2, %r2, %r40;
or.pred %p3, %p1, %p2;
setp.ge.s32 %p4, %r3, %r41;
or.pred %p5, %p3, %p4;
@%p5 bra BB14_15;
add.s32 %r4, %r1, %r32;
add.s32 %r5, %r2, %r33;
add.s32 %r6, %r3, %r34;
suq.channel_order.b32 %r7, [imageWriteui2D_arr_param_1];
setp.gt.s32 %p6, %r7, 4277;
@%p6 bra BB14_5;
setp.eq.s32 %p9, %r7, 4273;
@%p9 bra BB14_10;
bra.uni BB14_3;
BB14_10:
mov.u32 %r57, %r38;
bra.uni BB14_11;
BB14_5:
setp.eq.s32 %p7, %r7, 4278;
@%p7 bra BB14_6;
setp.ne.s32 %p8, %r7, 4279;
@%p8 bra BB14_8;
mov.u32 %r57, %r38;
mov.u32 %r58, %r35;
bra.uni BB14_14;
BB14_3:
setp.eq.s32 %p10, %r7, 4275;
@%p10 bra BB14_4;
bra.uni BB14_8;
BB14_4:
mov.u32 %r57, %r35;
mov.u32 %r58, %r38;
bra.uni BB14_12;
BB14_6:
mov.u32 %r57, %r60;
mov.u32 %r58, %r59;
mov.u32 %r59, %r35;
bra.uni BB14_13;
BB14_8:
mov.u32 %r57, %r35;
BB14_11:
mov.u32 %r58, %r59;
BB14_12:
mov.u32 %r59, %r60;
BB14_13:
mov.u32 %r60, %r38;
BB14_14:
sust.p.a2d.v4.b32.trap [imageWriteui2D_arr_param_1, {%r6, %r4, %r5, %r5}], {%r57, %r58, %r59, %r60};
BB14_15:
ret;
}
// .globl bufferFill
.entry bufferFill(
.param .u64 .ptr .global .align 1 bufferFill_param_0,
.param .align 4 .b8 bufferFill_param_1[128],
.param .u32 bufferFill_param_2,
.param .u32 bufferFill_param_3,
.param .u32 bufferFill_param_4
)
{
.reg .pred %p<13>;
.reg .b16 %rs<3>;
.reg .b32 %r<50>;
.reg .b64 %rd<19>;
ld.param.u64 %rd1, [bufferFill_param_0];
ld.param.u32 %r34, [bufferFill_param_2];
ld.param.u32 %r36, [bufferFill_param_3];
ld.param.u32 %r35, [bufferFill_param_4];
mov.u64 %rd2, bufferFill_param_1;
ld.param.u32 %r1, [%rd2];
ld.param.u32 %r2, [%rd2+4];
ld.param.u32 %r3, [%rd2+8];
ld.param.u32 %r4, [%rd2+12];
ld.param.u32 %r5, [%rd2+16];
ld.param.u32 %r6, [%rd2+20];
ld.param.u32 %r7, [%rd2+24];
ld.param.u32 %r8, [%rd2+28];
ld.param.u32 %r9, [%rd2+32];
ld.param.u32 %r10, [%rd2+36];
ld.param.u32 %r11, [%rd2+40];
ld.param.u32 %r12, [%rd2+44];
ld.param.u32 %r13, [%rd2+48];
ld.param.u32 %r14, [%rd2+52];
ld.param.u32 %r15, [%rd2+56];
ld.param.u32 %r16, [%rd2+60];
ld.param.u32 %r17, [%rd2+64];
ld.param.u32 %r18, [%rd2+68];
ld.param.u32 %r19, [%rd2+72];
ld.param.u32 %r20, [%rd2+76];
ld.param.u32 %r21, [%rd2+80];
ld.param.u32 %r22, [%rd2+84];
ld.param.u32 %r23, [%rd2+88];
ld.param.u32 %r24, [%rd2+92];
ld.param.u32 %r25, [%rd2+96];
ld.param.u32 %r26, [%rd2+100];
ld.param.u32 %r27, [%rd2+104];
ld.param.u32 %r28, [%rd2+108];
ld.param.u32 %r29, [%rd2+112];
ld.param.u32 %r30, [%rd2+116];
ld.param.u32 %r31, [%rd2+120];
ld.param.u32 %r32, [%rd2+124];
mov.u32 %r37, %ctaid.x;
mov.u32 %r38, %ntid.x;
mov.b32 %r39, %envreg3;
mad.lo.s32 %r40, %r37, %r38, %r39;
mov.u32 %r41, %tid.x;
add.s32 %r33, %r40, %r41;
setp.ge.s32 %p1, %r33, %r36;
@%p1 bra BB15_20;
setp.gt.s32 %p2, %r34, 15;
@%p2 bra BB15_9;
setp.gt.s32 %p8, %r34, 3;
@%p8 bra BB15_6;
setp.eq.s32 %p11, %r34, 1;
@%p11 bra BB15_19;
bra.uni BB15_4;
BB15_19:
add.s32 %r49, %r33, %r35;
cvt.s64.s32 %rd17, %r49;
add.s64 %rd18, %rd1, %rd17;
st.global.u8 [%rd18], %r1;
bra.uni BB15_20;
BB15_9:
setp.gt.s32 %p3, %r34, 63;
@%p3 bra BB15_13;
setp.eq.s32 %p6, %r34, 16;
@%p6 bra BB15_17;
bra.uni BB15_11;
BB15_17:
add.s32 %r45, %r33, %r35;
mul.wide.s32 %rd9, %r45, 16;
add.s64 %rd10, %rd1, %rd9;
st.global.u32 [%rd10], %r1;
st.global.u32 [%rd10+4], %r2;
st.global.u32 [%rd10+8], %r3;
st.global.u32 [%rd10+12], %r4;
bra.uni BB15_20;
BB15_6:
setp.eq.s32 %p9, %r34, 4;
@%p9 bra BB15_18;
bra.uni BB15_7;
BB15_18:
add.s32 %r47, %r33, %r35;
mul.wide.s32 %rd13, %r47, 4;
add.s64 %rd14, %rd1, %rd13;
st.global.u32 [%rd14], %r1;
bra.uni BB15_20;
BB15_13:
setp.eq.s32 %p4, %r34, 64;
@%p4 bra BB15_16;
bra.uni BB15_14;
BB15_16:
add.s32 %r43, %r33, %r35;
mul.wide.s32 %rd5, %r43, 64;
add.s64 %rd6, %rd1, %rd5;
st.global.u32 [%rd6], %r1;
st.global.u32 [%rd6+4], %r2;
st.global.u32 [%rd6+8], %r3;
st.global.u32 [%rd6+12], %r4;
st.global.u32 [%rd6+16], %r5;
st.global.u32 [%rd6+20], %r6;
st.global.u32 [%rd6+24], %r7;
st.global.u32 [%rd6+28], %r8;
st.global.u32 [%rd6+32], %r9;
st.global.u32 [%rd6+36], %r10;
st.global.u32 [%rd6+40], %r11;
st.global.u32 [%rd6+44], %r12;
st.global.u32 [%rd6+48], %r13;
st.global.u32 [%rd6+52], %r14;
st.global.u32 [%rd6+56], %r15;
st.global.u32 [%rd6+60], %r16;
bra.uni BB15_20;
BB15_4:
setp.eq.s32 %p12, %r34, 2;
@%p12 bra BB15_5;
bra.uni BB15_20;
BB15_5:
mov.b32 {%rs1, %rs2}, %r1;
add.s32 %r48, %r33, %r35;
mul.wide.s32 %rd15, %r48, 2;
add.s64 %rd16, %rd1, %rd15;
st.global.u16 [%rd16], %rs1;
bra.uni BB15_20;
BB15_11:
setp.eq.s32 %p7, %r34, 32;
@%p7 bra BB15_12;
bra.uni BB15_20;
BB15_12:
add.s32 %r44, %r33, %r35;
mul.wide.s32 %rd7, %r44, 32;
add.s64 %rd8, %rd1, %rd7;
st.global.u32 [%rd8], %r1;
st.global.u32 [%rd8+4], %r2;
st.global.u32 [%rd8+8], %r3;
st.global.u32 [%rd8+12], %r4;
st.global.u32 [%rd8+16], %r5;
st.global.u32 [%rd8+20], %r6;
st.global.u32 [%rd8+24], %r7;
st.global.u32 [%rd8+28], %r8;
bra.uni BB15_20;
BB15_7:
setp.eq.s32 %p10, %r34, 8;
@%p10 bra BB15_8;
bra.uni BB15_20;
BB15_8:
add.s32 %r46, %r33, %r35;
mul.wide.s32 %rd11, %r46, 8;
add.s64 %rd12, %rd1, %rd11;
st.global.u32 [%rd12], %r1;
st.global.u32 [%rd12+4], %r2;
bra.uni BB15_20;
BB15_14:
setp.ne.s32 %p5, %r34, 128;
@%p5 bra BB15_20;
add.s32 %r42, %r33, %r35;
mul.wide.s32 %rd3, %r42, 128;
add.s64 %rd4, %rd1, %rd3;
st.global.u32 [%rd4], %r1;
st.global.u32 [%rd4+4], %r2;
st.global.u32 [%rd4+8], %r3;
st.global.u32 [%rd4+12], %r4;
st.global.u32 [%rd4+16], %r5;
st.global.u32 [%rd4+20], %r6;
st.global.u32 [%rd4+24], %r7;
st.global.u32 [%rd4+28], %r8;
st.global.u32 [%rd4+32], %r9;
st.global.u32 [%rd4+36], %r10;
st.global.u32 [%rd4+40], %r11;
st.global.u32 [%rd4+44], %r12;
st.global.u32 [%rd4+48], %r13;
st.global.u32 [%rd4+52], %r14;
st.global.u32 [%rd4+56], %r15;
st.global.u32 [%rd4+60], %r16;
st.global.u32 [%rd4+64], %r17;
st.global.u32 [%rd4+68], %r18;
st.global.u32 [%rd4+72], %r19;
st.global.u32 [%rd4+76], %r20;
st.global.u32 [%rd4+80], %r21;
st.global.u32 [%rd4+84], %r22;
st.global.u32 [%rd4+88], %r23;
st.global.u32 [%rd4+92], %r24;
st.global.u32 [%rd4+96], %r25;
st.global.u32 [%rd4+100], %r26;
st.global.u32 [%rd4+104], %r27;
st.global.u32 [%rd4+108], %r28;
st.global.u32 [%rd4+112], %r29;
st.global.u32 [%rd4+116], %r30;
st.global.u32 [%rd4+120], %r31;
st.global.u32 [%rd4+124], %r32;
BB15_20:
ret;
}
ELF3 � o �} @l ## @ 8 @ E .shstrtab .strtab .symtab .symtab_shndx .nv.info .text.bufferFill .nv.info.bufferFill .nv.shared.bufferFill .nv.constant0.bufferFill .text.imageWriteui2D_arr .nv.info.imageWriteui2D_arr .nv.shared.imageWriteui2D_arr .nv.constant0.imageWriteui2D_arr .text.imageWriteui2D .nv.info.imageWriteui2D .nv.shared.imageWriteui2D .nv.constant0.imageWriteui2D .text.imageWritef2D_arr .nv.info.imageWritef2D_arr .nv.shared.imageWritef2D_arr .nv.constant0.imageWritef2D_arr .text.imageWritef2D .nv.info.imageWritef2D .nv.shared.imageWritef2D .nv.constant0.imageWritef2D .text.imageWritei2D_arr .nv.info.imageWritei2D_arr .nv.shared.imageWritei2D_arr .nv.constant0.imageWritei2D_arr .text.imageWritei2D .nv.info.imageWritei2D .nv.shared.imageWritei2D .nv.constant0.imageWritei2D .text.imageWriteui1D_buf .nv.info.imageWriteui1D_buf .nv.shared.imageWriteui1D_buf .nv.constant0.imageWriteui1D_buf .text.imageWriteui1D_arr .nv.info.imageWriteui1D_arr .nv.shared.imageWriteui1D_arr .nv.constant0.imageWriteui1D_arr .text.imageWriteui1D .nv.info.imageWriteui1D .nv.shared.imageWriteui1D .nv.constant0.imageWriteui1D .text.imageWritef1D_buf .nv.info.imageWritef1D_buf .nv.shared.imageWritef1D_buf .nv.constant0.imageWritef1D_buf .text.imageWritef1D_arr .nv.info.imageWritef1D_arr .nv.shared.imageWritef1D_arr .nv.constant0.imageWritef1D_arr .text.imageWritef1D .nv.info.imageWritef1D .nv.shared.imageWritef1D .nv.constant0.imageWritef1D .text.imageWritei1D_buf .nv.info.imageWritei1D_buf .nv.shared.imageWritei1D_buf .nv.constant0.imageWritei1D_buf .text.imageWritei1D_arr .nv.info.imageWritei1D_arr .nv.shared.imageWritei1D_arr .nv.constant0.imageWritei1D_arr .text.imageWritei1D .nv.info.imageWritei1D .nv.shared.imageWritei1D .nv.constant0.imageWritei1D .rel.nv.constant0.imageWriteui2D_arr .rel.nv.constant0.imageWriteui2D .rel.nv.constant0.imageWritef2D_arr .rel.nv.constant0.imageWritef2D .rel.nv.constant0.imageWritei2D_arr .rel.nv.constant0.imageWritei2D .rel.nv.constant0.imageWriteui1D_buf .rel.nv.constant0.imageWriteui1D_arr .rel.nv.constant0.imageWriteui1D .rel.nv.constant0.imageWritef1D_buf .rel.nv.constant0.imageWritef1D_arr .rel.nv.constant0.imageWritef1D .rel.nv.constant0.imageWritei1D_buf .rel.nv.constant0.imageWritei1D_arr .rel.nv.constant0.imageWritei1D .nv.rel.action .shstrtab .strtab .symtab .symtab_shndx .nv.info bufferFill .text.bufferFill .nv.info.bufferFill .nv.shared.bufferFill .nv.constant0.bufferFill _param imageWriteui2D_arr .text.imageWriteui2D_arr .nv.info.imageWriteui2D_arr .nv.shared.imageWriteui2D_arr imageWriteui2D_arr_param_1 .nv.constant0.imageWriteui2D_arr $BINDLESS$imageWriteui2D_arr$imageWriteui2D_arr_param_1 imageWriteui2D .text.imageWriteui2D .nv.info.imageWriteui2D .nv.shared.imageWriteui2D imageWriteui2D_param_1 .nv.constant0.imageWriteui2D $BINDLESS$imageWriteui2D$imageWriteui2D_param_1 imageWritef2D_arr .text.imageWritef2D_arr .nv.info.imageWritef2D_arr .nv.shared.imageWritef2D_arr imageWritef2D_arr_param_1 .nv.constant0.imageWritef2D_arr $BINDLESS$imageWritef2D_arr$imageWritef2D_arr_param_1 imageWritef2D .text.imageWritef2D .nv.info.imageWritef2D .nv.shared.imageWritef2D imageWritef2D_param_1 .nv.constant0.imageWritef2D $BINDLESS$imageWritef2D$imageWritef2D_param_1 imageWritei2D_arr .text.imageWritei2D_arr .nv.info.imageWritei2D_arr .nv.shared.imageWritei2D_arr imageWritei2D_arr_param_1 .nv.constant0.imageWritei2D_arr $BINDLESS$imageWritei2D_arr$imageWritei2D_arr_param_1 imageWritei2D .text.imageWritei2D .nv.info.imageWritei2D .nv.shared.imageWritei2D imageWritei2D_param_1 .nv.constant0.imageWritei2D $BINDLESS$imageWritei2D$imageWritei2D_param_1 imageWriteui1D_buf .text.imageWriteui1D_buf .nv.info.imageWriteui1D_buf .nv.shared.imageWriteui1D_buf imageWriteui1D_buf_param_1 .nv.constant0.imageWriteui1D_buf $BINDLESS$imageWriteui1D_buf$imageWriteui1D_buf_param_1 imageWriteui1D_arr .text.imageWriteui1D_arr .nv.info.imageWriteui1D_arr .nv.shared.imageWriteui1D_arr imageWriteui1D_arr_param_1 .nv.constant0.imageWriteui1D_arr $BINDLESS$imageWriteui1D_arr$imageWriteui1D_arr_param_1 imageWriteui1D .text.imageWriteui1D .nv.info.imageWriteui1D .nv.shared.imageWriteui1D imageWriteui1D_param_1 .nv.constant0.imageWriteui1D $BINDLESS$imageWriteui1D$imageWriteui1D_param_1 imageWritef1D_buf .text.imageWritef1D_buf .nv.info.imageWritef1D_buf .nv.shared.imageWritef1D_buf imageWritef1D_buf_param_1 .nv.constant0.imageWritef1D_buf $BINDLESS$imageWritef1D_buf$imageWritef1D_buf_param_1 imageWritef1D_arr .text.imageWritef1D_arr .nv.info.imageWritef1D_arr .nv.shared.imageWritef1D_arr imageWritef1D_arr_param_1 .nv.constant0.imageWritef1D_arr $BINDLESS$imageWritef1D_arr$imageWritef1D_arr_param_1 imageWritef1D .text.imageWritef1D .nv.info.imageWritef1D .nv.shared.imageWritef1D imageWritef1D_param_1 .nv.constant0.imageWritef1D $BINDLESS$imageWritef1D$imageWritef1D_param_1 imageWritei1D_buf .text.imageWritei1D_buf .nv.info.imageWritei1D_buf .nv.shared.imageWritei1D_buf imageWritei1D_buf_param_1 .nv.constant0.imageWritei1D_buf $BINDLESS$imageWritei1D_buf$imageWritei1D_buf_param_1 imageWritei1D_arr .text.imageWritei1D_arr .nv.info.imageWritei1D_arr .nv.shared.imageWritei1D_arr imageWritei1D_arr_param_1 .nv.constant0.imageWritei1D_arr $BINDLESS$imageWritei1D_arr$imageWritei1D_arr_param_1 imageWritei1D .text.imageWritei1D .nv.info.imageWritei1D .nv.shared.imageWritei1D imageWritei1D_param_1 .nv.constant0.imageWritei1D $BINDLESS$imageWritei1D$imageWritei1D_param_1 .rel.nv.constant0.imageWriteui2D_arr .rel.nv.constant0.imageWriteui2D .rel.nv.constant0.imageWritef2D_arr .rel.nv.constant0.imageWritef2D .rel.nv.constant0.imageWritei2D_arr .rel.nv.constant0.imageWritei2D .rel.nv.constant0.imageWriteui1D_buf .rel.nv.constant0.imageWriteui1D_arr .rel.nv.constant0.imageWriteui1D .rel.nv.constant0.imageWritef1D_buf .rel.nv.constant0.imageWritef1D_arr .rel.nv.constant0.imageWritef1D .rel.nv.constant0.imageWritei1D_buf .rel.nv.constant0.imageWritei1D_arr .rel.nv.constant0.imageWritei1D .nv.rel.action 2 5 @ = 5 x % � 6 @ � 6 � 6 &