1093 lines
35 KiB
C
1093 lines
35 KiB
C
const char * ljc_cut_gpu_kernel =
|
|
" .version 2.3\n"
|
|
" .target sm_20\n"
|
|
" .address_size 64\n"
|
|
" .global .texref pos_tex;\n"
|
|
" .global .texref q_tex;\n"
|
|
" .entry kernel_pair (\n"
|
|
" .param .u64 __cudaparm_kernel_pair_x_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_lj1,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_lj3,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_lj_types,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_sp_lj_in,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_dev_nbor,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_dev_packed,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_ans,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_engv,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_eflag,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_vflag,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_inum,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_nbor_pitch,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_q_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_cutsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_qqrd2e,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<86>;\n"
|
|
" .reg .u64 %rd<66>;\n"
|
|
" .reg .f32 %f<130>;\n"
|
|
" .reg .pred %p<21>;\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32498_33_non_const_sp_lj112[32];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32603_35_non_const_red_acc144[3072];\n"
|
|
" .loc 16 100 0\n"
|
|
"$LDWbegin_kernel_pair:\n"
|
|
" .loc 16 107 0\n"
|
|
" ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];\n"
|
|
" ldu.global.f32 %f1, [%rd1+0];\n"
|
|
" .loc 16 108 0\n"
|
|
" ld.global.f32 %f2, [%rd1+4];\n"
|
|
" .loc 16 109 0\n"
|
|
" ld.global.f32 %f3, [%rd1+8];\n"
|
|
" .loc 16 110 0\n"
|
|
" ld.global.f32 %f4, [%rd1+12];\n"
|
|
" st.shared.v4.f32 [__cuda___cuda_local_var_32498_33_non_const_sp_lj112+0], {%f1,%f2,%f3,%f4};\n"
|
|
" .loc 16 111 0\n"
|
|
" ld.global.f32 %f5, [%rd1+16];\n"
|
|
" .loc 16 112 0\n"
|
|
" ld.global.f32 %f6, [%rd1+20];\n"
|
|
" .loc 16 113 0\n"
|
|
" ld.global.f32 %f7, [%rd1+24];\n"
|
|
" .loc 16 114 0\n"
|
|
" ld.global.f32 %f8, [%rd1+28];\n"
|
|
" st.shared.v4.f32 [__cuda___cuda_local_var_32498_33_non_const_sp_lj112+16], {%f5,%f6,%f7,%f8};\n"
|
|
" .loc 16 124 0\n"
|
|
" mov.f32 %f9, 0f00000000; \n"
|
|
" mov.f32 %f10, %f9;\n"
|
|
" mov.f32 %f11, 0f00000000; \n"
|
|
" mov.f32 %f12, %f11;\n"
|
|
" mov.f32 %f13, 0f00000000; \n"
|
|
" mov.f32 %f14, %f13;\n"
|
|
" mov.f32 %f15, 0f00000000; \n"
|
|
" mov.f32 %f16, %f15;\n"
|
|
" mov.f32 %f17, 0f00000000; \n"
|
|
" mov.f32 %f18, %f17;\n"
|
|
" mov.f32 %f19, 0f00000000; \n"
|
|
" mov.f32 %f20, %f19;\n"
|
|
" ld.param.s32 %r1, [__cudaparm_kernel_pair_t_per_atom];\n"
|
|
" cvt.s32.u32 %r2, %tid.x;\n"
|
|
" div.s32 %r3, %r2, %r1;\n"
|
|
" cvt.s32.u32 %r4, %ntid.x;\n"
|
|
" div.s32 %r5, %r4, %r1;\n"
|
|
" rem.s32 %r6, %r2, %r1;\n"
|
|
" cvt.s32.u32 %r7, %ctaid.x;\n"
|
|
" mul.lo.s32 %r8, %r7, %r5;\n"
|
|
" add.s32 %r9, %r3, %r8;\n"
|
|
" ld.param.s32 %r10, [__cudaparm_kernel_pair_inum];\n"
|
|
" setp.lt.s32 %p1, %r9, %r10;\n"
|
|
" @!%p1 bra $Lt_0_21506;\n"
|
|
" .loc 16 128 0\n"
|
|
" cvt.s64.s32 %rd2, %r9;\n"
|
|
" mul.wide.s32 %rd3, %r9, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_dev_nbor];\n"
|
|
" add.u64 %rd5, %rd3, %rd4;\n"
|
|
" ld.global.s32 %r11, [%rd5+0];\n"
|
|
" .loc 16 130 0\n"
|
|
" ld.param.s32 %r12, [__cudaparm_kernel_pair_nbor_pitch];\n"
|
|
" cvt.s64.s32 %rd6, %r12;\n"
|
|
" mul.wide.s32 %rd7, %r12, 4;\n"
|
|
" add.u64 %rd8, %rd7, %rd5;\n"
|
|
" ld.global.s32 %r13, [%rd8+0];\n"
|
|
" add.u64 %rd9, %rd7, %rd8;\n"
|
|
" ld.param.u64 %rd10, [__cudaparm_kernel_pair_dev_packed];\n"
|
|
" setp.ne.u64 %p2, %rd10, %rd4;\n"
|
|
" @%p2 bra $Lt_0_22018;\n"
|
|
" .loc 16 136 0\n"
|
|
" cvt.s32.s64 %r14, %rd6;\n"
|
|
" mul.lo.s32 %r15, %r14, %r13;\n"
|
|
" cvt.s64.s32 %rd11, %r15;\n"
|
|
" mul.wide.s32 %rd12, %r15, 4;\n"
|
|
" add.u64 %rd13, %rd9, %rd12;\n"
|
|
" .loc 16 137 0\n"
|
|
" mul.lo.s32 %r16, %r6, %r14;\n"
|
|
" cvt.s64.s32 %rd14, %r16;\n"
|
|
" mul.wide.s32 %rd15, %r16, 4;\n"
|
|
" add.u64 %rd16, %rd9, %rd15;\n"
|
|
" .loc 16 138 0\n"
|
|
" mul.lo.s32 %r17, %r14, %r1;\n"
|
|
" bra.uni $Lt_0_21762;\n"
|
|
"$Lt_0_22018:\n"
|
|
" .loc 16 140 0\n"
|
|
" ld.global.s32 %r18, [%rd9+0];\n"
|
|
" cvt.s64.s32 %rd17, %r18;\n"
|
|
" mul.wide.s32 %rd18, %r18, 4;\n"
|
|
" add.u64 %rd19, %rd10, %rd18;\n"
|
|
" .loc 16 141 0\n"
|
|
" cvt.s64.s32 %rd20, %r13;\n"
|
|
" mul.wide.s32 %rd21, %r13, 4;\n"
|
|
" add.u64 %rd13, %rd19, %rd21;\n"
|
|
" .loc 16 142 0\n"
|
|
" mov.s32 %r17, %r1;\n"
|
|
" .loc 16 143 0\n"
|
|
" cvt.s64.s32 %rd22, %r6;\n"
|
|
" mul.wide.s32 %rd23, %r6, 4;\n"
|
|
" add.u64 %rd16, %rd19, %rd23;\n"
|
|
"$Lt_0_21762:\n"
|
|
" .loc 16 146 0\n"
|
|
" mov.u32 %r19, %r11;\n"
|
|
" mov.s32 %r20, 0;\n"
|
|
" mov.u32 %r21, %r20;\n"
|
|
" mov.s32 %r22, 0;\n"
|
|
" mov.u32 %r23, %r22;\n"
|
|
" mov.s32 %r24, 0;\n"
|
|
" mov.u32 %r25, %r24;\n"
|
|
" tex.1d.v4.f32.s32 {%f21,%f22,%f23,%f24},[pos_tex,{%r19,%r21,%r23,%r25}];\n"
|
|
" mov.f32 %f25, %f21;\n"
|
|
" mov.f32 %f26, %f22;\n"
|
|
" mov.f32 %f27, %f23;\n"
|
|
" mov.f32 %f28, %f24;\n"
|
|
" .loc 16 147 0\n"
|
|
" mov.u32 %r26, %r11;\n"
|
|
" mov.s32 %r27, 0;\n"
|
|
" mov.u32 %r28, %r27;\n"
|
|
" mov.s32 %r29, 0;\n"
|
|
" mov.u32 %r30, %r29;\n"
|
|
" mov.s32 %r31, 0;\n"
|
|
" mov.u32 %r32, %r31;\n"
|
|
" tex.1d.v4.f32.s32 {%f29,%f30,%f31,%f32},[q_tex,{%r26,%r28,%r30,%r32}];\n"
|
|
" mov.f32 %f33, %f29;\n"
|
|
" setp.ge.u64 %p3, %rd16, %rd13;\n"
|
|
" @%p3 bra $Lt_0_32002;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r33, %f28;\n"
|
|
" cvt.s64.s32 %rd24, %r17;\n"
|
|
" ld.param.s32 %r34, [__cudaparm_kernel_pair_lj_types];\n"
|
|
" mul.lo.s32 %r35, %r34, %r33;\n"
|
|
" ld.param.u64 %rd25, [__cudaparm_kernel_pair_cutsq];\n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.u64 %rd26, __cuda___cuda_local_var_32498_33_non_const_sp_lj112;\n"
|
|
"$Lt_0_22786:\n"
|
|
" .loc 16 151 0\n"
|
|
" ld.global.s32 %r36, [%rd16+0];\n"
|
|
" .loc 16 154 0\n"
|
|
" shr.s32 %r37, %r36, 30;\n"
|
|
" and.b32 %r38, %r37, 3;\n"
|
|
" cvt.s64.s32 %rd27, %r38;\n"
|
|
" mul.wide.s32 %rd28, %r38, 4;\n"
|
|
" add.u64 %rd29, %rd26, %rd28;\n"
|
|
" ld.shared.f32 %f39, [%rd29+0];\n"
|
|
" .loc 16 158 0\n"
|
|
" and.b32 %r39, %r36, 1073741823;\n"
|
|
" mov.u32 %r40, %r39;\n"
|
|
" mov.s32 %r41, 0;\n"
|
|
" mov.u32 %r42, %r41;\n"
|
|
" mov.s32 %r43, 0;\n"
|
|
" mov.u32 %r44, %r43;\n"
|
|
" mov.s32 %r45, 0;\n"
|
|
" mov.u32 %r46, %r45;\n"
|
|
" tex.1d.v4.f32.s32 {%f40,%f41,%f42,%f43},[pos_tex,{%r40,%r42,%r44,%r46}];\n"
|
|
" mov.f32 %f44, %f40;\n"
|
|
" mov.f32 %f45, %f41;\n"
|
|
" mov.f32 %f46, %f42;\n"
|
|
" mov.f32 %f47, %f43;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r47, %f47;\n"
|
|
" sub.ftz.f32 %f48, %f26, %f45;\n"
|
|
" sub.ftz.f32 %f49, %f25, %f44;\n"
|
|
" sub.ftz.f32 %f50, %f27, %f46;\n"
|
|
" mul.ftz.f32 %f51, %f48, %f48;\n"
|
|
" fma.rn.ftz.f32 %f52, %f49, %f49, %f51;\n"
|
|
" add.s32 %r48, %r47, %r35;\n"
|
|
" cvt.s64.s32 %rd30, %r48;\n"
|
|
" fma.rn.ftz.f32 %f53, %f50, %f50, %f52;\n"
|
|
" mul.wide.s32 %rd31, %r48, 4;\n"
|
|
" add.u64 %rd32, %rd25, %rd31;\n"
|
|
" ld.global.f32 %f54, [%rd32+0];\n"
|
|
" setp.gt.ftz.f32 %p4, %f54, %f53;\n"
|
|
" @!%p4 bra $Lt_0_25602;\n"
|
|
" mul.lo.u64 %rd33, %rd30, 16;\n"
|
|
" rcp.approx.ftz.f32 %f55, %f53;\n"
|
|
" ld.param.u64 %rd34, [__cudaparm_kernel_pair_lj1];\n"
|
|
" add.u64 %rd35, %rd34, %rd33;\n"
|
|
" ld.global.f32 %f56, [%rd35+8];\n"
|
|
" setp.lt.ftz.f32 %p5, %f53, %f56;\n"
|
|
" @!%p5 bra $Lt_0_23810;\n"
|
|
" .loc 16 173 0\n"
|
|
" mul.ftz.f32 %f57, %f55, %f55;\n"
|
|
" mul.ftz.f32 %f58, %f55, %f57;\n"
|
|
" mov.f32 %f59, %f58;\n"
|
|
" .loc 16 174 0\n"
|
|
" mul.ftz.f32 %f60, %f58, %f39;\n"
|
|
" ld.global.v2.f32 {%f61,%f62}, [%rd35+0];\n"
|
|
" mul.ftz.f32 %f63, %f61, %f58;\n"
|
|
" sub.ftz.f32 %f64, %f63, %f62;\n"
|
|
" mul.ftz.f32 %f65, %f60, %f64;\n"
|
|
" bra.uni $Lt_0_23554;\n"
|
|
"$Lt_0_23810:\n"
|
|
" .loc 16 176 0\n"
|
|
" mov.f32 %f65, 0f00000000; \n"
|
|
"$Lt_0_23554:\n"
|
|
" ld.global.f32 %f66, [%rd35+12];\n"
|
|
" setp.gt.ftz.f32 %p6, %f66, %f53;\n"
|
|
" @!%p6 bra $Lt_0_24322;\n"
|
|
" .loc 16 179 0\n"
|
|
" mov.u32 %r49, %r39;\n"
|
|
" mov.s32 %r50, 0;\n"
|
|
" mov.u32 %r51, %r50;\n"
|
|
" mov.s32 %r52, 0;\n"
|
|
" mov.u32 %r53, %r52;\n"
|
|
" mov.s32 %r54, 0;\n"
|
|
" mov.u32 %r55, %r54;\n"
|
|
" tex.1d.v4.f32.s32 {%f67,%f68,%f69,%f70},[q_tex,{%r49,%r51,%r53,%r55}];\n"
|
|
" mov.f32 %f71, %f67;\n"
|
|
" ld.shared.f32 %f72, [%rd29+16];\n"
|
|
" ld.param.f32 %f73, [__cudaparm_kernel_pair_qqrd2e];\n"
|
|
" mul.ftz.f32 %f74, %f73, %f33;\n"
|
|
" mul.ftz.f32 %f75, %f71, %f74;\n"
|
|
" sqrt.approx.ftz.f32 %f76, %f55;\n"
|
|
" mul.ftz.f32 %f77, %f75, %f76;\n"
|
|
" mul.ftz.f32 %f78, %f72, %f77;\n"
|
|
" bra.uni $Lt_0_24066;\n"
|
|
"$Lt_0_24322:\n"
|
|
" .loc 16 181 0\n"
|
|
" mov.f32 %f78, 0f00000000; \n"
|
|
"$Lt_0_24066:\n"
|
|
" .loc 16 185 0\n"
|
|
" add.ftz.f32 %f79, %f78, %f65;\n"
|
|
" mul.ftz.f32 %f80, %f79, %f55;\n"
|
|
" fma.rn.ftz.f32 %f36, %f49, %f80, %f36;\n"
|
|
" .loc 16 186 0\n"
|
|
" fma.rn.ftz.f32 %f35, %f48, %f80, %f35;\n"
|
|
" .loc 16 187 0\n"
|
|
" fma.rn.ftz.f32 %f34, %f50, %f80, %f34;\n"
|
|
" ld.param.s32 %r56, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r57, 0;\n"
|
|
" setp.le.s32 %p7, %r56, %r57;\n"
|
|
" @%p7 bra $Lt_0_25090;\n"
|
|
" .loc 16 190 0\n"
|
|
" add.ftz.f32 %f37, %f78, %f37;\n"
|
|
" @!%p5 bra $Lt_0_25090;\n"
|
|
" .loc 16 193 0\n"
|
|
" ld.param.u64 %rd36, [__cudaparm_kernel_pair_lj3];\n"
|
|
" add.u64 %rd37, %rd36, %rd33;\n"
|
|
" mov.f32 %f81, %f59;\n"
|
|
" ld.global.v4.f32 {%f82,%f83,%f84,_}, [%rd37+0];\n"
|
|
" mul.ftz.f32 %f85, %f82, %f81;\n"
|
|
" sub.ftz.f32 %f86, %f85, %f83;\n"
|
|
" mul.ftz.f32 %f87, %f81, %f86;\n"
|
|
" sub.ftz.f32 %f88, %f87, %f84;\n"
|
|
" fma.rn.ftz.f32 %f38, %f39, %f88, %f38;\n"
|
|
"$Lt_0_25090:\n"
|
|
"$Lt_0_24578:\n"
|
|
" ld.param.s32 %r58, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r59, 0;\n"
|
|
" setp.le.s32 %p8, %r58, %r59;\n"
|
|
" @%p8 bra $Lt_0_25602;\n"
|
|
" .loc 16 197 0\n"
|
|
" mov.f32 %f89, %f10;\n"
|
|
" mul.ftz.f32 %f90, %f49, %f49;\n"
|
|
" fma.rn.ftz.f32 %f91, %f80, %f90, %f89;\n"
|
|
" mov.f32 %f10, %f91;\n"
|
|
" .loc 16 198 0\n"
|
|
" mov.f32 %f92, %f12;\n"
|
|
" fma.rn.ftz.f32 %f93, %f80, %f51, %f92;\n"
|
|
" mov.f32 %f12, %f93;\n"
|
|
" .loc 16 199 0\n"
|
|
" mov.f32 %f94, %f14;\n"
|
|
" mul.ftz.f32 %f95, %f50, %f50;\n"
|
|
" fma.rn.ftz.f32 %f96, %f80, %f95, %f94;\n"
|
|
" mov.f32 %f14, %f96;\n"
|
|
" .loc 16 200 0\n"
|
|
" mov.f32 %f97, %f16;\n"
|
|
" mul.ftz.f32 %f98, %f48, %f49;\n"
|
|
" fma.rn.ftz.f32 %f99, %f80, %f98, %f97;\n"
|
|
" mov.f32 %f16, %f99;\n"
|
|
" .loc 16 201 0\n"
|
|
" mov.f32 %f100, %f18;\n"
|
|
" mul.ftz.f32 %f101, %f49, %f50;\n"
|
|
" fma.rn.ftz.f32 %f102, %f80, %f101, %f100;\n"
|
|
" mov.f32 %f18, %f102;\n"
|
|
" .loc 16 202 0\n"
|
|
" mul.ftz.f32 %f103, %f48, %f50;\n"
|
|
" fma.rn.ftz.f32 %f19, %f80, %f103, %f19;\n"
|
|
" mov.f32 %f20, %f19;\n"
|
|
"$Lt_0_25602:\n"
|
|
"$Lt_0_23042:\n"
|
|
" .loc 16 150 0\n"
|
|
" mul.lo.u64 %rd38, %rd24, 4;\n"
|
|
" add.u64 %rd16, %rd16, %rd38;\n"
|
|
" setp.lt.u64 %p9, %rd16, %rd13;\n"
|
|
" @%p9 bra $Lt_0_22786;\n"
|
|
" bra.uni $Lt_0_21250;\n"
|
|
"$Lt_0_32002:\n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" bra.uni $Lt_0_21250;\n"
|
|
"$Lt_0_21506:\n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
"$Lt_0_21250:\n"
|
|
" mov.u32 %r60, 1;\n"
|
|
" setp.le.s32 %p10, %r1, %r60;\n"
|
|
" @%p10 bra $Lt_0_28418;\n"
|
|
" .loc 16 213 0\n"
|
|
" mov.u64 %rd39, __cuda___cuda_local_var_32603_35_non_const_red_acc144;\n"
|
|
" cvt.s64.s32 %rd40, %r2;\n"
|
|
" mul.wide.s32 %rd41, %r2, 4;\n"
|
|
" add.u64 %rd42, %rd39, %rd41;\n"
|
|
" mov.f32 %f104, %f36;\n"
|
|
" st.shared.f32 [%rd42+0], %f104;\n"
|
|
" .loc 16 214 0\n"
|
|
" mov.f32 %f105, %f35;\n"
|
|
" st.shared.f32 [%rd42+512], %f105;\n"
|
|
" .loc 16 215 0\n"
|
|
" mov.f32 %f106, %f34;\n"
|
|
" st.shared.f32 [%rd42+1024], %f106;\n"
|
|
" .loc 16 216 0\n"
|
|
" mov.f32 %f107, %f38;\n"
|
|
" st.shared.f32 [%rd42+1536], %f107;\n"
|
|
" .loc 16 217 0\n"
|
|
" mov.f32 %f108, %f37;\n"
|
|
" st.shared.f32 [%rd42+2048], %f108;\n"
|
|
" .loc 16 219 0\n"
|
|
" shr.s32 %r61, %r1, 31;\n"
|
|
" mov.s32 %r62, 1;\n"
|
|
" and.b32 %r63, %r61, %r62;\n"
|
|
" add.s32 %r64, %r63, %r1;\n"
|
|
" shr.s32 %r65, %r64, 1;\n"
|
|
" mov.s32 %r66, %r65;\n"
|
|
" mov.u32 %r67, 0;\n"
|
|
" setp.ne.u32 %p11, %r65, %r67;\n"
|
|
" @!%p11 bra $Lt_0_26882;\n"
|
|
"$Lt_0_27394:\n"
|
|
" setp.ge.u32 %p12, %r6, %r66;\n"
|
|
" @%p12 bra $Lt_0_27650;\n"
|
|
" .loc 16 222 0\n"
|
|
" add.u32 %r68, %r2, %r66;\n"
|
|
" cvt.u64.u32 %rd43, %r68;\n"
|
|
" mul.wide.u32 %rd44, %r68, 4;\n"
|
|
" add.u64 %rd45, %rd39, %rd44;\n"
|
|
" ld.shared.f32 %f109, [%rd45+0];\n"
|
|
" add.ftz.f32 %f104, %f109, %f104;\n"
|
|
" st.shared.f32 [%rd42+0], %f104;\n"
|
|
" ld.shared.f32 %f110, [%rd45+512];\n"
|
|
" add.ftz.f32 %f105, %f110, %f105;\n"
|
|
" st.shared.f32 [%rd42+512], %f105;\n"
|
|
" ld.shared.f32 %f111, [%rd45+1024];\n"
|
|
" add.ftz.f32 %f106, %f111, %f106;\n"
|
|
" st.shared.f32 [%rd42+1024], %f106;\n"
|
|
" ld.shared.f32 %f112, [%rd45+1536];\n"
|
|
" add.ftz.f32 %f107, %f112, %f107;\n"
|
|
" st.shared.f32 [%rd42+1536], %f107;\n"
|
|
" ld.shared.f32 %f113, [%rd45+2048];\n"
|
|
" add.ftz.f32 %f108, %f113, %f108;\n"
|
|
" st.shared.f32 [%rd42+2048], %f108;\n"
|
|
"$Lt_0_27650:\n"
|
|
" .loc 16 219 0\n"
|
|
" shr.u32 %r66, %r66, 1;\n"
|
|
" mov.u32 %r69, 0;\n"
|
|
" setp.ne.u32 %p13, %r66, %r69;\n"
|
|
" @%p13 bra $Lt_0_27394;\n"
|
|
"$Lt_0_26882:\n"
|
|
" .loc 16 226 0\n"
|
|
" mov.f32 %f36, %f104;\n"
|
|
" .loc 16 227 0\n"
|
|
" mov.f32 %f35, %f105;\n"
|
|
" .loc 16 228 0\n"
|
|
" mov.f32 %f34, %f106;\n"
|
|
" .loc 16 229 0\n"
|
|
" mov.f32 %f38, %f107;\n"
|
|
" .loc 16 230 0\n"
|
|
" mov.f32 %f37, %f108;\n"
|
|
" ld.param.s32 %r70, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r71, 0;\n"
|
|
" setp.le.s32 %p14, %r70, %r71;\n"
|
|
" @%p14 bra $Lt_0_28418;\n"
|
|
" .loc 16 234 0\n"
|
|
" mov.f32 %f104, %f10;\n"
|
|
" st.shared.f32 [%rd42+0], %f104;\n"
|
|
" mov.f32 %f105, %f12;\n"
|
|
" st.shared.f32 [%rd42+512], %f105;\n"
|
|
" mov.f32 %f106, %f14;\n"
|
|
" st.shared.f32 [%rd42+1024], %f106;\n"
|
|
" mov.f32 %f107, %f16;\n"
|
|
" st.shared.f32 [%rd42+1536], %f107;\n"
|
|
" mov.f32 %f108, %f18;\n"
|
|
" st.shared.f32 [%rd42+2048], %f108;\n"
|
|
" mov.f32 %f114, %f20;\n"
|
|
" st.shared.f32 [%rd42+2560], %f114;\n"
|
|
" .loc 16 236 0\n"
|
|
" mov.s32 %r72, %r65;\n"
|
|
" @!%p11 bra $Lt_0_28930;\n"
|
|
"$Lt_0_29442:\n"
|
|
" setp.ge.u32 %p15, %r6, %r72;\n"
|
|
" @%p15 bra $Lt_0_29698;\n"
|
|
" .loc 16 239 0\n"
|
|
" add.u32 %r73, %r2, %r72;\n"
|
|
" cvt.u64.u32 %rd46, %r73;\n"
|
|
" mul.wide.u32 %rd47, %r73, 4;\n"
|
|
" add.u64 %rd48, %rd39, %rd47;\n"
|
|
" ld.shared.f32 %f115, [%rd48+0];\n"
|
|
" add.ftz.f32 %f104, %f115, %f104;\n"
|
|
" st.shared.f32 [%rd42+0], %f104;\n"
|
|
" ld.shared.f32 %f116, [%rd48+512];\n"
|
|
" add.ftz.f32 %f105, %f116, %f105;\n"
|
|
" st.shared.f32 [%rd42+512], %f105;\n"
|
|
" ld.shared.f32 %f117, [%rd48+1024];\n"
|
|
" add.ftz.f32 %f106, %f117, %f106;\n"
|
|
" st.shared.f32 [%rd42+1024], %f106;\n"
|
|
" ld.shared.f32 %f118, [%rd48+1536];\n"
|
|
" add.ftz.f32 %f107, %f118, %f107;\n"
|
|
" st.shared.f32 [%rd42+1536], %f107;\n"
|
|
" ld.shared.f32 %f119, [%rd48+2048];\n"
|
|
" add.ftz.f32 %f108, %f119, %f108;\n"
|
|
" st.shared.f32 [%rd42+2048], %f108;\n"
|
|
" ld.shared.f32 %f120, [%rd48+2560];\n"
|
|
" add.ftz.f32 %f114, %f120, %f114;\n"
|
|
" st.shared.f32 [%rd42+2560], %f114;\n"
|
|
"$Lt_0_29698:\n"
|
|
" .loc 16 236 0\n"
|
|
" shr.u32 %r72, %r72, 1;\n"
|
|
" mov.u32 %r74, 0;\n"
|
|
" setp.ne.u32 %p16, %r72, %r74;\n"
|
|
" @%p16 bra $Lt_0_29442;\n"
|
|
"$Lt_0_28930:\n"
|
|
" .loc 16 244 0\n"
|
|
" mov.f32 %f10, %f104;\n"
|
|
" mov.f32 %f12, %f105;\n"
|
|
" mov.f32 %f14, %f106;\n"
|
|
" mov.f32 %f16, %f107;\n"
|
|
" mov.f32 %f18, %f108;\n"
|
|
" mov.f32 %f20, %f114;\n"
|
|
"$Lt_0_28418:\n"
|
|
"$Lt_0_26370:\n"
|
|
" selp.s32 %r75, 1, 0, %p1;\n"
|
|
" mov.s32 %r76, 0;\n"
|
|
" set.eq.u32.s32 %r77, %r6, %r76;\n"
|
|
" neg.s32 %r78, %r77;\n"
|
|
" and.b32 %r79, %r75, %r78;\n"
|
|
" mov.u32 %r80, 0;\n"
|
|
" setp.eq.s32 %p17, %r79, %r80;\n"
|
|
" @%p17 bra $Lt_0_30466;\n"
|
|
" .loc 16 250 0\n"
|
|
" cvt.s64.s32 %rd49, %r9;\n"
|
|
" ld.param.u64 %rd50, [__cudaparm_kernel_pair_engv];\n"
|
|
" mul.wide.s32 %rd51, %r9, 4;\n"
|
|
" add.u64 %rd52, %rd50, %rd51;\n"
|
|
" ld.param.s32 %r81, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r82, 0;\n"
|
|
" setp.le.s32 %p18, %r81, %r82;\n"
|
|
" @%p18 bra $Lt_0_30978;\n"
|
|
" .loc 16 252 0\n"
|
|
" st.global.f32 [%rd52+0], %f38;\n"
|
|
" .loc 16 253 0\n"
|
|
" cvt.s64.s32 %rd53, %r10;\n"
|
|
" mul.wide.s32 %rd54, %r10, 4;\n"
|
|
" add.u64 %rd55, %rd54, %rd52;\n"
|
|
" .loc 16 254 0\n"
|
|
" st.global.f32 [%rd55+0], %f37;\n"
|
|
" .loc 16 255 0\n"
|
|
" add.u64 %rd52, %rd54, %rd55;\n"
|
|
"$Lt_0_30978:\n"
|
|
" ld.param.s32 %r83, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r84, 0;\n"
|
|
" setp.le.s32 %p19, %r83, %r84;\n"
|
|
" @%p19 bra $Lt_0_31490;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f121, %f10;\n"
|
|
" st.global.f32 [%rd52+0], %f121;\n"
|
|
" .loc 16 260 0\n"
|
|
" cvt.s64.s32 %rd56, %r10;\n"
|
|
" mul.wide.s32 %rd57, %r10, 4;\n"
|
|
" add.u64 %rd58, %rd57, %rd52;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f122, %f12;\n"
|
|
" st.global.f32 [%rd58+0], %f122;\n"
|
|
" .loc 16 260 0\n"
|
|
" add.u64 %rd59, %rd57, %rd58;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f123, %f14;\n"
|
|
" st.global.f32 [%rd59+0], %f123;\n"
|
|
" .loc 16 260 0\n"
|
|
" add.u64 %rd60, %rd57, %rd59;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f124, %f16;\n"
|
|
" st.global.f32 [%rd60+0], %f124;\n"
|
|
" .loc 16 260 0\n"
|
|
" add.u64 %rd52, %rd57, %rd60;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f125, %f18;\n"
|
|
" st.global.f32 [%rd52+0], %f125;\n"
|
|
" mov.f32 %f126, %f20;\n"
|
|
" add.u64 %rd61, %rd57, %rd52;\n"
|
|
" st.global.f32 [%rd61+0], %f126;\n"
|
|
"$Lt_0_31490:\n"
|
|
" .loc 16 263 0\n"
|
|
" ld.param.u64 %rd62, [__cudaparm_kernel_pair_ans];\n"
|
|
" mul.lo.u64 %rd63, %rd49, 16;\n"
|
|
" add.u64 %rd64, %rd62, %rd63;\n"
|
|
" mov.f32 %f127, %f128;\n"
|
|
" st.global.v4.f32 [%rd64+0], {%f36,%f35,%f34,%f127};\n"
|
|
"$Lt_0_30466:\n"
|
|
" .loc 16 265 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair:\n"
|
|
" }\n"
|
|
" .entry kernel_pair_fast (\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_x_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_lj1_in,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_lj3_in,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_sp_lj_in,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_dev_nbor,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_dev_packed,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_ans,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_engv,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_eflag,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_vflag,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_inum,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_nbor_pitch,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_q_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast__cutsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_qqrd2e,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<88>;\n"
|
|
" .reg .u64 %rd<82>;\n"
|
|
" .reg .f32 %f<134>;\n"
|
|
" .reg .pred %p<24>;\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32676_33_non_const_sp_lj3320[32];\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32673_34_non_const_lj13360[1936];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32675_33_non_const_cutsq5296[484];\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32674_34_non_const_lj35792[1936];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32783_35_non_const_red_acc7728[3072];\n"
|
|
" .loc 16 275 0\n"
|
|
"$LDWbegin_kernel_pair_fast:\n"
|
|
" cvt.s32.u32 %r1, %tid.x;\n"
|
|
" mov.u32 %r2, 7;\n"
|
|
" setp.gt.s32 %p1, %r1, %r2;\n"
|
|
" @%p1 bra $Lt_1_23554;\n"
|
|
" .loc 16 286 0\n"
|
|
" mov.u64 %rd1, __cuda___cuda_local_var_32676_33_non_const_sp_lj3320;\n"
|
|
" cvt.s64.s32 %rd2, %r1;\n"
|
|
" mul.wide.s32 %rd3, %r1, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_fast_sp_lj_in];\n"
|
|
" add.u64 %rd5, %rd4, %rd3;\n"
|
|
" ld.global.f32 %f1, [%rd5+0];\n"
|
|
" add.u64 %rd6, %rd3, %rd1;\n"
|
|
" st.shared.f32 [%rd6+0], %f1;\n"
|
|
"$Lt_1_23554:\n"
|
|
" mov.u64 %rd1, __cuda___cuda_local_var_32676_33_non_const_sp_lj3320;\n"
|
|
" mov.u32 %r3, 120;\n"
|
|
" setp.gt.s32 %p2, %r1, %r3;\n"
|
|
" @%p2 bra $Lt_1_24066;\n"
|
|
" .loc 16 288 0\n"
|
|
" mov.u64 %rd7, __cuda___cuda_local_var_32673_34_non_const_lj13360;\n"
|
|
" mov.u64 %rd8, __cuda___cuda_local_var_32675_33_non_const_cutsq5296;\n"
|
|
" cvt.s64.s32 %rd9, %r1;\n"
|
|
" mul.wide.s32 %rd10, %r1, 16;\n"
|
|
" ld.param.u64 %rd11, [__cudaparm_kernel_pair_fast_lj1_in];\n"
|
|
" add.u64 %rd12, %rd11, %rd10;\n"
|
|
" add.u64 %rd13, %rd10, %rd7;\n"
|
|
" ld.global.v4.f32 {%f2,%f3,%f4,%f5}, [%rd12+0];\n"
|
|
" st.shared.v4.f32 [%rd13+0], {%f2,%f3,%f4,%f5};\n"
|
|
" .loc 16 289 0\n"
|
|
" mul.wide.s32 %rd14, %r1, 4;\n"
|
|
" ld.param.u64 %rd15, [__cudaparm_kernel_pair_fast__cutsq];\n"
|
|
" add.u64 %rd16, %rd15, %rd14;\n"
|
|
" ld.global.f32 %f6, [%rd16+0];\n"
|
|
" add.u64 %rd17, %rd14, %rd8;\n"
|
|
" st.shared.f32 [%rd17+0], %f6;\n"
|
|
" ld.param.s32 %r4, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r5, 0;\n"
|
|
" setp.le.s32 %p3, %r4, %r5;\n"
|
|
" @%p3 bra $Lt_1_24578;\n"
|
|
" .loc 16 291 0\n"
|
|
" mov.u64 %rd18, __cuda___cuda_local_var_32674_34_non_const_lj35792;\n"
|
|
" ld.param.u64 %rd19, [__cudaparm_kernel_pair_fast_lj3_in];\n"
|
|
" add.u64 %rd20, %rd19, %rd10;\n"
|
|
" add.u64 %rd21, %rd10, %rd18;\n"
|
|
" ld.global.v4.f32 {%f7,%f8,%f9,%f10}, [%rd20+0];\n"
|
|
" st.shared.v4.f32 [%rd21+0], {%f7,%f8,%f9,%f10};\n"
|
|
"$Lt_1_24578:\n"
|
|
" mov.u64 %rd18, __cuda___cuda_local_var_32674_34_non_const_lj35792;\n"
|
|
"$Lt_1_24066:\n"
|
|
" mov.u64 %rd18, __cuda___cuda_local_var_32674_34_non_const_lj35792;\n"
|
|
" mov.u64 %rd7, __cuda___cuda_local_var_32673_34_non_const_lj13360;\n"
|
|
" mov.u64 %rd8, __cuda___cuda_local_var_32675_33_non_const_cutsq5296;\n"
|
|
" .loc 16 302 0\n"
|
|
" mov.f32 %f11, 0f00000000; \n"
|
|
" mov.f32 %f12, %f11;\n"
|
|
" mov.f32 %f13, 0f00000000; \n"
|
|
" mov.f32 %f14, %f13;\n"
|
|
" mov.f32 %f15, 0f00000000; \n"
|
|
" mov.f32 %f16, %f15;\n"
|
|
" mov.f32 %f17, 0f00000000; \n"
|
|
" mov.f32 %f18, %f17;\n"
|
|
" mov.f32 %f19, 0f00000000; \n"
|
|
" mov.f32 %f20, %f19;\n"
|
|
" mov.f32 %f21, 0f00000000; \n"
|
|
" mov.f32 %f22, %f21;\n"
|
|
" .loc 16 304 0\n"
|
|
" bar.sync 0;\n"
|
|
" ld.param.s32 %r6, [__cudaparm_kernel_pair_fast_t_per_atom];\n"
|
|
" div.s32 %r7, %r1, %r6;\n"
|
|
" cvt.s32.u32 %r8, %ntid.x;\n"
|
|
" div.s32 %r9, %r8, %r6;\n"
|
|
" rem.s32 %r10, %r1, %r6;\n"
|
|
" cvt.s32.u32 %r11, %ctaid.x;\n"
|
|
" mul.lo.s32 %r12, %r11, %r9;\n"
|
|
" add.s32 %r13, %r7, %r12;\n"
|
|
" ld.param.s32 %r14, [__cudaparm_kernel_pair_fast_inum];\n"
|
|
" setp.lt.s32 %p4, %r13, %r14;\n"
|
|
" @!%p4 bra $Lt_1_25346;\n"
|
|
" .loc 16 308 0\n"
|
|
" cvt.s64.s32 %rd22, %r13;\n"
|
|
" mul.wide.s32 %rd23, %r13, 4;\n"
|
|
" ld.param.u64 %rd24, [__cudaparm_kernel_pair_fast_dev_nbor];\n"
|
|
" add.u64 %rd25, %rd23, %rd24;\n"
|
|
" ld.global.s32 %r15, [%rd25+0];\n"
|
|
" .loc 16 310 0\n"
|
|
" ld.param.s32 %r16, [__cudaparm_kernel_pair_fast_nbor_pitch];\n"
|
|
" cvt.s64.s32 %rd26, %r16;\n"
|
|
" mul.wide.s32 %rd27, %r16, 4;\n"
|
|
" add.u64 %rd28, %rd27, %rd25;\n"
|
|
" ld.global.s32 %r17, [%rd28+0];\n"
|
|
" add.u64 %rd29, %rd27, %rd28;\n"
|
|
" ld.param.u64 %rd30, [__cudaparm_kernel_pair_fast_dev_packed];\n"
|
|
" setp.ne.u64 %p5, %rd30, %rd24;\n"
|
|
" @%p5 bra $Lt_1_25858;\n"
|
|
" .loc 16 316 0\n"
|
|
" cvt.s32.s64 %r18, %rd26;\n"
|
|
" mul.lo.s32 %r19, %r18, %r17;\n"
|
|
" cvt.s64.s32 %rd31, %r19;\n"
|
|
" mul.wide.s32 %rd32, %r19, 4;\n"
|
|
" add.u64 %rd33, %rd29, %rd32;\n"
|
|
" .loc 16 317 0\n"
|
|
" mul.lo.s32 %r20, %r10, %r18;\n"
|
|
" cvt.s64.s32 %rd34, %r20;\n"
|
|
" mul.wide.s32 %rd35, %r20, 4;\n"
|
|
" add.u64 %rd36, %rd29, %rd35;\n"
|
|
" .loc 16 318 0\n"
|
|
" mul.lo.s32 %r21, %r18, %r6;\n"
|
|
" bra.uni $Lt_1_25602;\n"
|
|
"$Lt_1_25858:\n"
|
|
" .loc 16 320 0\n"
|
|
" ld.global.s32 %r22, [%rd29+0];\n"
|
|
" cvt.s64.s32 %rd37, %r22;\n"
|
|
" mul.wide.s32 %rd38, %r22, 4;\n"
|
|
" add.u64 %rd39, %rd30, %rd38;\n"
|
|
" .loc 16 321 0\n"
|
|
" cvt.s64.s32 %rd40, %r17;\n"
|
|
" mul.wide.s32 %rd41, %r17, 4;\n"
|
|
" add.u64 %rd33, %rd39, %rd41;\n"
|
|
" .loc 16 322 0\n"
|
|
" mov.s32 %r21, %r6;\n"
|
|
" .loc 16 323 0\n"
|
|
" cvt.s64.s32 %rd42, %r10;\n"
|
|
" mul.wide.s32 %rd43, %r10, 4;\n"
|
|
" add.u64 %rd36, %rd39, %rd43;\n"
|
|
"$Lt_1_25602:\n"
|
|
" .loc 16 326 0\n"
|
|
" mov.u32 %r23, %r15;\n"
|
|
" mov.s32 %r24, 0;\n"
|
|
" mov.u32 %r25, %r24;\n"
|
|
" mov.s32 %r26, 0;\n"
|
|
" mov.u32 %r27, %r26;\n"
|
|
" mov.s32 %r28, 0;\n"
|
|
" mov.u32 %r29, %r28;\n"
|
|
" tex.1d.v4.f32.s32 {%f23,%f24,%f25,%f26},[pos_tex,{%r23,%r25,%r27,%r29}];\n"
|
|
" mov.f32 %f27, %f23;\n"
|
|
" mov.f32 %f28, %f24;\n"
|
|
" mov.f32 %f29, %f25;\n"
|
|
" mov.f32 %f30, %f26;\n"
|
|
" .loc 16 327 0\n"
|
|
" mov.u32 %r30, %r15;\n"
|
|
" mov.s32 %r31, 0;\n"
|
|
" mov.u32 %r32, %r31;\n"
|
|
" mov.s32 %r33, 0;\n"
|
|
" mov.u32 %r34, %r33;\n"
|
|
" mov.s32 %r35, 0;\n"
|
|
" mov.u32 %r36, %r35;\n"
|
|
" tex.1d.v4.f32.s32 {%f31,%f32,%f33,%f34},[q_tex,{%r30,%r32,%r34,%r36}];\n"
|
|
" mov.f32 %f35, %f31;\n"
|
|
" setp.ge.u64 %p6, %rd36, %rd33;\n"
|
|
" @%p6 bra $Lt_1_35842;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r37, %f30;\n"
|
|
" cvt.s64.s32 %rd44, %r21;\n"
|
|
" mul.lo.s32 %r38, %r37, 11;\n"
|
|
" cvt.rn.f32.s32 %f36, %r38;\n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
" mov.f32 %f40, 0f00000000; \n"
|
|
" mov.f32 %f41, 0f00000000; \n"
|
|
"$Lt_1_26626:\n"
|
|
" .loc 16 332 0\n"
|
|
" ld.global.s32 %r39, [%rd36+0];\n"
|
|
" .loc 16 335 0\n"
|
|
" shr.s32 %r40, %r39, 30;\n"
|
|
" and.b32 %r41, %r40, 3;\n"
|
|
" cvt.s64.s32 %rd45, %r41;\n"
|
|
" mul.wide.s32 %rd46, %r41, 4;\n"
|
|
" add.u64 %rd47, %rd1, %rd46;\n"
|
|
" ld.shared.f32 %f42, [%rd47+0];\n"
|
|
" .loc 16 339 0\n"
|
|
" and.b32 %r42, %r39, 1073741823;\n"
|
|
" mov.u32 %r43, %r42;\n"
|
|
" mov.s32 %r44, 0;\n"
|
|
" mov.u32 %r45, %r44;\n"
|
|
" mov.s32 %r46, 0;\n"
|
|
" mov.u32 %r47, %r46;\n"
|
|
" mov.s32 %r48, 0;\n"
|
|
" mov.u32 %r49, %r48;\n"
|
|
" tex.1d.v4.f32.s32 {%f43,%f44,%f45,%f46},[pos_tex,{%r43,%r45,%r47,%r49}];\n"
|
|
" mov.f32 %f47, %f43;\n"
|
|
" mov.f32 %f48, %f44;\n"
|
|
" mov.f32 %f49, %f45;\n"
|
|
" mov.f32 %f50, %f46;\n"
|
|
" sub.ftz.f32 %f51, %f28, %f48;\n"
|
|
" sub.ftz.f32 %f52, %f27, %f47;\n"
|
|
" sub.ftz.f32 %f53, %f29, %f49;\n"
|
|
" mul.ftz.f32 %f54, %f51, %f51;\n"
|
|
" fma.rn.ftz.f32 %f55, %f52, %f52, %f54;\n"
|
|
" fma.rn.ftz.f32 %f56, %f53, %f53, %f55;\n"
|
|
" add.ftz.f32 %f57, %f36, %f50;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r50, %f57;\n"
|
|
" cvt.s64.s32 %rd48, %r50;\n"
|
|
" mul.wide.s32 %rd49, %r50, 4;\n"
|
|
" add.u64 %rd50, %rd8, %rd49;\n"
|
|
" ld.shared.f32 %f58, [%rd50+0];\n"
|
|
" setp.gt.ftz.f32 %p7, %f58, %f56;\n"
|
|
" @!%p7 bra $Lt_1_29442;\n"
|
|
" rcp.approx.ftz.f32 %f59, %f56;\n"
|
|
" mul.lo.u64 %rd51, %rd48, 16;\n"
|
|
" add.u64 %rd52, %rd51, %rd7;\n"
|
|
" ld.shared.f32 %f60, [%rd52+8];\n"
|
|
" setp.lt.ftz.f32 %p8, %f56, %f60;\n"
|
|
" @!%p8 bra $Lt_1_27650;\n"
|
|
" .loc 16 353 0\n"
|
|
" mul.ftz.f32 %f61, %f59, %f59;\n"
|
|
" mul.ftz.f32 %f62, %f59, %f61;\n"
|
|
" mov.f32 %f63, %f62;\n"
|
|
" .loc 16 354 0\n"
|
|
" mul.ftz.f32 %f64, %f62, %f42;\n"
|
|
" ld.shared.v2.f32 {%f65,%f66}, [%rd52+0];\n"
|
|
" mul.ftz.f32 %f67, %f65, %f62;\n"
|
|
" sub.ftz.f32 %f68, %f67, %f66;\n"
|
|
" mul.ftz.f32 %f69, %f64, %f68;\n"
|
|
" bra.uni $Lt_1_27394;\n"
|
|
"$Lt_1_27650:\n"
|
|
" .loc 16 356 0\n"
|
|
" mov.f32 %f69, 0f00000000; \n"
|
|
"$Lt_1_27394:\n"
|
|
" ld.shared.f32 %f70, [%rd52+12];\n"
|
|
" setp.gt.ftz.f32 %p9, %f70, %f56;\n"
|
|
" @!%p9 bra $Lt_1_28162;\n"
|
|
" .loc 16 359 0\n"
|
|
" mov.u32 %r51, %r42;\n"
|
|
" mov.s32 %r52, 0;\n"
|
|
" mov.u32 %r53, %r52;\n"
|
|
" mov.s32 %r54, 0;\n"
|
|
" mov.u32 %r55, %r54;\n"
|
|
" mov.s32 %r56, 0;\n"
|
|
" mov.u32 %r57, %r56;\n"
|
|
" tex.1d.v4.f32.s32 {%f71,%f72,%f73,%f74},[q_tex,{%r51,%r53,%r55,%r57}];\n"
|
|
" mov.f32 %f75, %f71;\n"
|
|
" ld.shared.f32 %f76, [%rd47+16];\n"
|
|
" ld.param.f32 %f77, [__cudaparm_kernel_pair_fast_qqrd2e];\n"
|
|
" mul.ftz.f32 %f78, %f77, %f35;\n"
|
|
" mul.ftz.f32 %f79, %f75, %f78;\n"
|
|
" sqrt.approx.ftz.f32 %f80, %f59;\n"
|
|
" mul.ftz.f32 %f81, %f79, %f80;\n"
|
|
" mul.ftz.f32 %f82, %f76, %f81;\n"
|
|
" bra.uni $Lt_1_27906;\n"
|
|
"$Lt_1_28162:\n"
|
|
" .loc 16 361 0\n"
|
|
" mov.f32 %f82, 0f00000000; \n"
|
|
"$Lt_1_27906:\n"
|
|
" .loc 16 365 0\n"
|
|
" add.ftz.f32 %f83, %f82, %f69;\n"
|
|
" mul.ftz.f32 %f84, %f83, %f59;\n"
|
|
" fma.rn.ftz.f32 %f39, %f52, %f84, %f39;\n"
|
|
" .loc 16 366 0\n"
|
|
" fma.rn.ftz.f32 %f38, %f51, %f84, %f38;\n"
|
|
" .loc 16 367 0\n"
|
|
" fma.rn.ftz.f32 %f37, %f53, %f84, %f37;\n"
|
|
" ld.param.s32 %r58, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r59, 0;\n"
|
|
" setp.le.s32 %p10, %r58, %r59;\n"
|
|
" @%p10 bra $Lt_1_28930;\n"
|
|
" .loc 16 370 0\n"
|
|
" add.ftz.f32 %f40, %f82, %f40;\n"
|
|
" @!%p8 bra $Lt_1_28930;\n"
|
|
" .loc 16 372 0\n"
|
|
" add.u64 %rd53, %rd51, %rd18;\n"
|
|
" mov.f32 %f85, %f63;\n"
|
|
" ld.shared.v4.f32 {%f86,%f87,%f88,_}, [%rd53+0];\n"
|
|
" mul.ftz.f32 %f89, %f86, %f85;\n"
|
|
" sub.ftz.f32 %f90, %f89, %f87;\n"
|
|
" mul.ftz.f32 %f91, %f85, %f90;\n"
|
|
" .loc 16 373 0\n"
|
|
" sub.ftz.f32 %f92, %f91, %f88;\n"
|
|
" fma.rn.ftz.f32 %f41, %f42, %f92, %f41;\n"
|
|
"$Lt_1_28930:\n"
|
|
"$Lt_1_28418:\n"
|
|
" ld.param.s32 %r60, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r61, 0;\n"
|
|
" setp.le.s32 %p11, %r60, %r61;\n"
|
|
" @%p11 bra $Lt_1_29442;\n"
|
|
" .loc 16 377 0\n"
|
|
" mov.f32 %f93, %f12;\n"
|
|
" mul.ftz.f32 %f94, %f52, %f52;\n"
|
|
" fma.rn.ftz.f32 %f95, %f84, %f94, %f93;\n"
|
|
" mov.f32 %f12, %f95;\n"
|
|
" .loc 16 378 0\n"
|
|
" mov.f32 %f96, %f14;\n"
|
|
" fma.rn.ftz.f32 %f97, %f84, %f54, %f96;\n"
|
|
" mov.f32 %f14, %f97;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f98, %f16;\n"
|
|
" mul.ftz.f32 %f99, %f53, %f53;\n"
|
|
" fma.rn.ftz.f32 %f100, %f84, %f99, %f98;\n"
|
|
" mov.f32 %f16, %f100;\n"
|
|
" .loc 16 380 0\n"
|
|
" mov.f32 %f101, %f18;\n"
|
|
" mul.ftz.f32 %f102, %f51, %f52;\n"
|
|
" fma.rn.ftz.f32 %f103, %f84, %f102, %f101;\n"
|
|
" mov.f32 %f18, %f103;\n"
|
|
" .loc 16 381 0\n"
|
|
" mov.f32 %f104, %f20;\n"
|
|
" mul.ftz.f32 %f105, %f52, %f53;\n"
|
|
" fma.rn.ftz.f32 %f106, %f84, %f105, %f104;\n"
|
|
" mov.f32 %f20, %f106;\n"
|
|
" .loc 16 382 0\n"
|
|
" mul.ftz.f32 %f107, %f51, %f53;\n"
|
|
" fma.rn.ftz.f32 %f21, %f84, %f107, %f21;\n"
|
|
" mov.f32 %f22, %f21;\n"
|
|
"$Lt_1_29442:\n"
|
|
"$Lt_1_26882:\n"
|
|
" .loc 16 331 0\n"
|
|
" mul.lo.u64 %rd54, %rd44, 4;\n"
|
|
" add.u64 %rd36, %rd36, %rd54;\n"
|
|
" setp.lt.u64 %p12, %rd36, %rd33;\n"
|
|
" @%p12 bra $Lt_1_26626;\n"
|
|
" bra.uni $Lt_1_25090;\n"
|
|
"$Lt_1_35842:\n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
" mov.f32 %f40, 0f00000000; \n"
|
|
" mov.f32 %f41, 0f00000000; \n"
|
|
" bra.uni $Lt_1_25090;\n"
|
|
"$Lt_1_25346:\n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
" mov.f32 %f40, 0f00000000; \n"
|
|
" mov.f32 %f41, 0f00000000; \n"
|
|
"$Lt_1_25090:\n"
|
|
" mov.u32 %r62, 1;\n"
|
|
" setp.le.s32 %p13, %r6, %r62;\n"
|
|
" @%p13 bra $Lt_1_32258;\n"
|
|
" .loc 16 393 0\n"
|
|
" mov.u64 %rd55, __cuda___cuda_local_var_32783_35_non_const_red_acc7728;\n"
|
|
" cvt.s64.s32 %rd56, %r1;\n"
|
|
" mul.wide.s32 %rd57, %r1, 4;\n"
|
|
" add.u64 %rd58, %rd55, %rd57;\n"
|
|
" mov.f32 %f108, %f39;\n"
|
|
" st.shared.f32 [%rd58+0], %f108;\n"
|
|
" .loc 16 394 0\n"
|
|
" mov.f32 %f109, %f38;\n"
|
|
" st.shared.f32 [%rd58+512], %f109;\n"
|
|
" .loc 16 395 0\n"
|
|
" mov.f32 %f110, %f37;\n"
|
|
" st.shared.f32 [%rd58+1024], %f110;\n"
|
|
" .loc 16 396 0\n"
|
|
" mov.f32 %f111, %f41;\n"
|
|
" st.shared.f32 [%rd58+1536], %f111;\n"
|
|
" .loc 16 397 0\n"
|
|
" mov.f32 %f112, %f40;\n"
|
|
" st.shared.f32 [%rd58+2048], %f112;\n"
|
|
" .loc 16 399 0\n"
|
|
" shr.s32 %r63, %r6, 31;\n"
|
|
" mov.s32 %r64, 1;\n"
|
|
" and.b32 %r65, %r63, %r64;\n"
|
|
" add.s32 %r66, %r65, %r6;\n"
|
|
" shr.s32 %r67, %r66, 1;\n"
|
|
" mov.s32 %r68, %r67;\n"
|
|
" mov.u32 %r69, 0;\n"
|
|
" setp.ne.u32 %p14, %r67, %r69;\n"
|
|
" @!%p14 bra $Lt_1_30722;\n"
|
|
"$Lt_1_31234:\n"
|
|
" setp.ge.u32 %p15, %r10, %r68;\n"
|
|
" @%p15 bra $Lt_1_31490;\n"
|
|
" .loc 16 402 0\n"
|
|
" add.u32 %r70, %r1, %r68;\n"
|
|
" cvt.u64.u32 %rd59, %r70;\n"
|
|
" mul.wide.u32 %rd60, %r70, 4;\n"
|
|
" add.u64 %rd61, %rd55, %rd60;\n"
|
|
" ld.shared.f32 %f113, [%rd61+0];\n"
|
|
" add.ftz.f32 %f108, %f113, %f108;\n"
|
|
" st.shared.f32 [%rd58+0], %f108;\n"
|
|
" ld.shared.f32 %f114, [%rd61+512];\n"
|
|
" add.ftz.f32 %f109, %f114, %f109;\n"
|
|
" st.shared.f32 [%rd58+512], %f109;\n"
|
|
" ld.shared.f32 %f115, [%rd61+1024];\n"
|
|
" add.ftz.f32 %f110, %f115, %f110;\n"
|
|
" st.shared.f32 [%rd58+1024], %f110;\n"
|
|
" ld.shared.f32 %f116, [%rd61+1536];\n"
|
|
" add.ftz.f32 %f111, %f116, %f111;\n"
|
|
" st.shared.f32 [%rd58+1536], %f111;\n"
|
|
" ld.shared.f32 %f117, [%rd61+2048];\n"
|
|
" add.ftz.f32 %f112, %f117, %f112;\n"
|
|
" st.shared.f32 [%rd58+2048], %f112;\n"
|
|
"$Lt_1_31490:\n"
|
|
" .loc 16 399 0\n"
|
|
" shr.u32 %r68, %r68, 1;\n"
|
|
" mov.u32 %r71, 0;\n"
|
|
" setp.ne.u32 %p16, %r68, %r71;\n"
|
|
" @%p16 bra $Lt_1_31234;\n"
|
|
"$Lt_1_30722:\n"
|
|
" .loc 16 406 0\n"
|
|
" mov.f32 %f39, %f108;\n"
|
|
" .loc 16 407 0\n"
|
|
" mov.f32 %f38, %f109;\n"
|
|
" .loc 16 408 0\n"
|
|
" mov.f32 %f37, %f110;\n"
|
|
" .loc 16 409 0\n"
|
|
" mov.f32 %f41, %f111;\n"
|
|
" .loc 16 410 0\n"
|
|
" mov.f32 %f40, %f112;\n"
|
|
" ld.param.s32 %r72, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r73, 0;\n"
|
|
" setp.le.s32 %p17, %r72, %r73;\n"
|
|
" @%p17 bra $Lt_1_32258;\n"
|
|
" .loc 16 414 0\n"
|
|
" mov.f32 %f108, %f12;\n"
|
|
" st.shared.f32 [%rd58+0], %f108;\n"
|
|
" mov.f32 %f109, %f14;\n"
|
|
" st.shared.f32 [%rd58+512], %f109;\n"
|
|
" mov.f32 %f110, %f16;\n"
|
|
" st.shared.f32 [%rd58+1024], %f110;\n"
|
|
" mov.f32 %f111, %f18;\n"
|
|
" st.shared.f32 [%rd58+1536], %f111;\n"
|
|
" mov.f32 %f112, %f20;\n"
|
|
" st.shared.f32 [%rd58+2048], %f112;\n"
|
|
" mov.f32 %f118, %f22;\n"
|
|
" st.shared.f32 [%rd58+2560], %f118;\n"
|
|
" .loc 16 416 0\n"
|
|
" mov.s32 %r74, %r67;\n"
|
|
" @!%p14 bra $Lt_1_32770;\n"
|
|
"$Lt_1_33282:\n"
|
|
" setp.ge.u32 %p18, %r10, %r74;\n"
|
|
" @%p18 bra $Lt_1_33538;\n"
|
|
" .loc 16 419 0\n"
|
|
" add.u32 %r75, %r1, %r74;\n"
|
|
" cvt.u64.u32 %rd62, %r75;\n"
|
|
" mul.wide.u32 %rd63, %r75, 4;\n"
|
|
" add.u64 %rd64, %rd55, %rd63;\n"
|
|
" ld.shared.f32 %f119, [%rd64+0];\n"
|
|
" add.ftz.f32 %f108, %f119, %f108;\n"
|
|
" st.shared.f32 [%rd58+0], %f108;\n"
|
|
" ld.shared.f32 %f120, [%rd64+512];\n"
|
|
" add.ftz.f32 %f109, %f120, %f109;\n"
|
|
" st.shared.f32 [%rd58+512], %f109;\n"
|
|
" ld.shared.f32 %f121, [%rd64+1024];\n"
|
|
" add.ftz.f32 %f110, %f121, %f110;\n"
|
|
" st.shared.f32 [%rd58+1024], %f110;\n"
|
|
" ld.shared.f32 %f122, [%rd64+1536];\n"
|
|
" add.ftz.f32 %f111, %f122, %f111;\n"
|
|
" st.shared.f32 [%rd58+1536], %f111;\n"
|
|
" ld.shared.f32 %f123, [%rd64+2048];\n"
|
|
" add.ftz.f32 %f112, %f123, %f112;\n"
|
|
" st.shared.f32 [%rd58+2048], %f112;\n"
|
|
" ld.shared.f32 %f124, [%rd64+2560];\n"
|
|
" add.ftz.f32 %f118, %f124, %f118;\n"
|
|
" st.shared.f32 [%rd58+2560], %f118;\n"
|
|
"$Lt_1_33538:\n"
|
|
" .loc 16 416 0\n"
|
|
" shr.u32 %r74, %r74, 1;\n"
|
|
" mov.u32 %r76, 0;\n"
|
|
" setp.ne.u32 %p19, %r74, %r76;\n"
|
|
" @%p19 bra $Lt_1_33282;\n"
|
|
"$Lt_1_32770:\n"
|
|
" .loc 16 424 0\n"
|
|
" mov.f32 %f12, %f108;\n"
|
|
" mov.f32 %f14, %f109;\n"
|
|
" mov.f32 %f16, %f110;\n"
|
|
" mov.f32 %f18, %f111;\n"
|
|
" mov.f32 %f20, %f112;\n"
|
|
" mov.f32 %f22, %f118;\n"
|
|
"$Lt_1_32258:\n"
|
|
"$Lt_1_30210:\n"
|
|
" selp.s32 %r77, 1, 0, %p4;\n"
|
|
" mov.s32 %r78, 0;\n"
|
|
" set.eq.u32.s32 %r79, %r10, %r78;\n"
|
|
" neg.s32 %r80, %r79;\n"
|
|
" and.b32 %r81, %r77, %r80;\n"
|
|
" mov.u32 %r82, 0;\n"
|
|
" setp.eq.s32 %p20, %r81, %r82;\n"
|
|
" @%p20 bra $Lt_1_34306;\n"
|
|
" .loc 16 430 0\n"
|
|
" cvt.s64.s32 %rd65, %r13;\n"
|
|
" ld.param.u64 %rd66, [__cudaparm_kernel_pair_fast_engv];\n"
|
|
" mul.wide.s32 %rd67, %r13, 4;\n"
|
|
" add.u64 %rd68, %rd66, %rd67;\n"
|
|
" ld.param.s32 %r83, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r84, 0;\n"
|
|
" setp.le.s32 %p21, %r83, %r84;\n"
|
|
" @%p21 bra $Lt_1_34818;\n"
|
|
" .loc 16 432 0\n"
|
|
" st.global.f32 [%rd68+0], %f41;\n"
|
|
" .loc 16 433 0\n"
|
|
" cvt.s64.s32 %rd69, %r14;\n"
|
|
" mul.wide.s32 %rd70, %r14, 4;\n"
|
|
" add.u64 %rd71, %rd70, %rd68;\n"
|
|
" .loc 16 434 0\n"
|
|
" st.global.f32 [%rd71+0], %f40;\n"
|
|
" .loc 16 435 0\n"
|
|
" add.u64 %rd68, %rd70, %rd71;\n"
|
|
"$Lt_1_34818:\n"
|
|
" ld.param.s32 %r85, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r86, 0;\n"
|
|
" setp.le.s32 %p22, %r85, %r86;\n"
|
|
" @%p22 bra $Lt_1_35330;\n"
|
|
" .loc 16 439 0\n"
|
|
" mov.f32 %f125, %f12;\n"
|
|
" st.global.f32 [%rd68+0], %f125;\n"
|
|
" .loc 16 440 0\n"
|
|
" cvt.s64.s32 %rd72, %r14;\n"
|
|
" mul.wide.s32 %rd73, %r14, 4;\n"
|
|
" add.u64 %rd74, %rd73, %rd68;\n"
|
|
" .loc 16 439 0\n"
|
|
" mov.f32 %f126, %f14;\n"
|
|
" st.global.f32 [%rd74+0], %f126;\n"
|
|
" .loc 16 440 0\n"
|
|
" add.u64 %rd75, %rd73, %rd74;\n"
|
|
" .loc 16 439 0\n"
|
|
" mov.f32 %f127, %f16;\n"
|
|
" st.global.f32 [%rd75+0], %f127;\n"
|
|
" .loc 16 440 0\n"
|
|
" add.u64 %rd76, %rd73, %rd75;\n"
|
|
" .loc 16 439 0\n"
|
|
" mov.f32 %f128, %f18;\n"
|
|
" st.global.f32 [%rd76+0], %f128;\n"
|
|
" .loc 16 440 0\n"
|
|
" add.u64 %rd68, %rd73, %rd76;\n"
|
|
" .loc 16 439 0\n"
|
|
" mov.f32 %f129, %f20;\n"
|
|
" st.global.f32 [%rd68+0], %f129;\n"
|
|
" mov.f32 %f130, %f22;\n"
|
|
" add.u64 %rd77, %rd73, %rd68;\n"
|
|
" st.global.f32 [%rd77+0], %f130;\n"
|
|
"$Lt_1_35330:\n"
|
|
" .loc 16 443 0\n"
|
|
" ld.param.u64 %rd78, [__cudaparm_kernel_pair_fast_ans];\n"
|
|
" mul.lo.u64 %rd79, %rd65, 16;\n"
|
|
" add.u64 %rd80, %rd78, %rd79;\n"
|
|
" mov.f32 %f131, %f132;\n"
|
|
" st.global.v4.f32 [%rd80+0], {%f39,%f38,%f37,%f131};\n"
|
|
"$Lt_1_34306:\n"
|
|
" .loc 16 445 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair_fast:\n"
|
|
" }\n"
|
|
;
|