1228 lines
40 KiB
C
1228 lines
40 KiB
C
const char * crml_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 .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 .f32 __cudaparm_kernel_pair_cut_coulsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_qqrd2e,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_g_ewald,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_denom_lj,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_cut_bothsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_cut_ljsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_cut_lj_innersq,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<91>;\n"
|
|
" .reg .u64 %rd<64>;\n"
|
|
" .reg .f32 %f<190>;\n"
|
|
" .reg .pred %p<23>;\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32500_33_non_const_sp_lj120[32];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32624_35_non_const_red_acc152[3072];\n"
|
|
" .loc 16 110 0\n"
|
|
"$LDWbegin_kernel_pair:\n"
|
|
" .loc 16 118 0\n"
|
|
" ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];\n"
|
|
" ldu.global.f32 %f1, [%rd1+0];\n"
|
|
" .loc 16 119 0\n"
|
|
" ld.global.f32 %f2, [%rd1+4];\n"
|
|
" .loc 16 120 0\n"
|
|
" ld.global.f32 %f3, [%rd1+8];\n"
|
|
" .loc 16 121 0\n"
|
|
" ld.global.f32 %f4, [%rd1+12];\n"
|
|
" st.shared.v4.f32 [__cuda___cuda_local_var_32500_33_non_const_sp_lj120+0], {%f1,%f2,%f3,%f4};\n"
|
|
" .loc 16 122 0\n"
|
|
" ld.global.f32 %f5, [%rd1+16];\n"
|
|
" .loc 16 123 0\n"
|
|
" ld.global.f32 %f6, [%rd1+20];\n"
|
|
" .loc 16 124 0\n"
|
|
" ld.global.f32 %f7, [%rd1+24];\n"
|
|
" .loc 16 125 0\n"
|
|
" ld.global.f32 %f8, [%rd1+28];\n"
|
|
" st.shared.v4.f32 [__cuda___cuda_local_var_32500_33_non_const_sp_lj120+16], {%f5,%f6,%f7,%f8};\n"
|
|
" .loc 16 135 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_23810;\n"
|
|
" .loc 16 139 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 16 141 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_24322;\n"
|
|
" .loc 16 147 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 16 148 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 16 149 0\n"
|
|
" mul.lo.s32 %r17, %r14, %r1;\n"
|
|
" bra.uni $Lt_0_24066;\n"
|
|
"$Lt_0_24322:\n"
|
|
" .loc 16 151 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 16 152 0\n"
|
|
" cvt.s64.s32 %rd20, %r13;\n"
|
|
" mul.wide.s32 %rd21, %r13, 4;\n"
|
|
" add.u64 %rd13, %rd19, %rd21;\n"
|
|
" .loc 16 153 0\n"
|
|
" mov.s32 %r17, %r1;\n"
|
|
" .loc 16 154 0\n"
|
|
" cvt.s64.s32 %rd22, %r6;\n"
|
|
" mul.wide.s32 %rd23, %r6, 4;\n"
|
|
" add.u64 %rd16, %rd19, %rd23;\n"
|
|
"$Lt_0_24066:\n"
|
|
" .loc 16 157 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 16 158 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 $Lt_0_34818;\n"
|
|
" cvt.s64.s32 %rd24, %r17;\n"
|
|
" ld.param.f32 %f34, [__cudaparm_kernel_pair_cut_bothsq];\n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
" mov.u64 %rd25, __cuda___cuda_local_var_32500_33_non_const_sp_lj120;\n"
|
|
"$Lt_0_25090:\n"
|
|
" .loc 16 162 0\n"
|
|
" ld.global.s32 %r33, [%rd16+0];\n"
|
|
" .loc 16 165 0\n"
|
|
" shr.s32 %r34, %r33, 30;\n"
|
|
" and.b32 %r35, %r34, 3;\n"
|
|
" cvt.s64.s32 %rd26, %r35;\n"
|
|
" mul.wide.s32 %rd27, %r35, 4;\n"
|
|
" add.u64 %rd28, %rd25, %rd27;\n"
|
|
" ld.shared.f32 %f40, [%rd28+0];\n"
|
|
" .loc 16 166 0\n"
|
|
" mov.f32 %f41, 0f3f800000; \n"
|
|
" ld.shared.f32 %f42, [%rd28+16];\n"
|
|
" sub.ftz.f32 %f43, %f41, %f42;\n"
|
|
" .loc 16 169 0\n"
|
|
" and.b32 %r36, %r33, 1073741823;\n"
|
|
" mov.u32 %r37, %r36;\n"
|
|
" mov.s32 %r38, 0;\n"
|
|
" mov.u32 %r39, %r38;\n"
|
|
" mov.s32 %r40, 0;\n"
|
|
" mov.u32 %r41, %r40;\n"
|
|
" mov.s32 %r42, 0;\n"
|
|
" mov.u32 %r43, %r42;\n"
|
|
" tex.1d.v4.f32.s32 {%f44,%f45,%f46,%f47},[pos_tex,{%r37,%r39,%r41,%r43}];\n"
|
|
" mov.f32 %f48, %f44;\n"
|
|
" mov.f32 %f49, %f45;\n"
|
|
" mov.f32 %f50, %f46;\n"
|
|
" mov.f32 %f51, %f47;\n"
|
|
" sub.ftz.f32 %f52, %f26, %f49;\n"
|
|
" sub.ftz.f32 %f53, %f25, %f48;\n"
|
|
" sub.ftz.f32 %f54, %f27, %f50;\n"
|
|
" mul.ftz.f32 %f55, %f52, %f52;\n"
|
|
" fma.rn.ftz.f32 %f56, %f53, %f53, %f55;\n"
|
|
" fma.rn.ftz.f32 %f57, %f54, %f54, %f56;\n"
|
|
" setp.lt.ftz.f32 %p4, %f57, %f34;\n"
|
|
" @!%p4 bra $Lt_0_28418;\n"
|
|
" ld.param.f32 %f58, [__cudaparm_kernel_pair_cut_ljsq];\n"
|
|
" setp.lt.ftz.f32 %p5, %f57, %f58;\n"
|
|
" rcp.approx.ftz.f32 %f59, %f57;\n"
|
|
" @!%p5 bra $Lt_0_26114;\n"
|
|
" .loc 16 184 0\n"
|
|
" mul.ftz.f32 %f60, %f59, %f59;\n"
|
|
" mul.ftz.f32 %f61, %f59, %f60;\n"
|
|
" mov.f32 %f62, %f61;\n"
|
|
" .loc 16 185 0\n"
|
|
" cvt.rzi.ftz.s32.f32 %r44, %f51;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r45, %f28;\n"
|
|
" ld.param.u64 %rd29, [__cudaparm_kernel_pair_lj1];\n"
|
|
" ld.param.s32 %r46, [__cudaparm_kernel_pair_lj_types];\n"
|
|
" mul.lo.s32 %r47, %r46, %r45;\n"
|
|
" add.s32 %r48, %r44, %r47;\n"
|
|
" cvt.s64.s32 %rd30, %r48;\n"
|
|
" mul.wide.s32 %rd31, %r48, 16;\n"
|
|
" add.u64 %rd32, %rd29, %rd31;\n"
|
|
" mul.ftz.f32 %f63, %f61, %f40;\n"
|
|
" ld.global.v2.f32 {%f64,%f65}, [%rd32+0];\n"
|
|
" mul.ftz.f32 %f66, %f64, %f61;\n"
|
|
" sub.ftz.f32 %f67, %f66, %f65;\n"
|
|
" mul.ftz.f32 %f68, %f63, %f67;\n"
|
|
" ld.param.f32 %f69, [__cudaparm_kernel_pair_cut_lj_innersq];\n"
|
|
" setp.gt.ftz.f32 %p6, %f57, %f69;\n"
|
|
" @!%p6 bra $Lt_0_25858;\n"
|
|
" .loc 16 191 0\n"
|
|
" add.ftz.f32 %f70, %f57, %f57;\n"
|
|
" sub.ftz.f32 %f71, %f58, %f57;\n"
|
|
" add.ftz.f32 %f72, %f70, %f58;\n"
|
|
" mul.ftz.f32 %f73, %f71, %f71;\n"
|
|
" mov.f32 %f74, 0f40400000; \n"
|
|
" mul.ftz.f32 %f75, %f74, %f69;\n"
|
|
" sub.ftz.f32 %f76, %f72, %f75;\n"
|
|
" ld.param.f32 %f77, [__cudaparm_kernel_pair_denom_lj];\n"
|
|
" div.approx.ftz.f32 %f78, %f76, %f77;\n"
|
|
" mul.ftz.f32 %f79, %f73, %f78;\n"
|
|
" mov.f32 %f80, %f79;\n"
|
|
" .loc 16 194 0\n"
|
|
" mov.f32 %f81, 0f41400000; \n"
|
|
" mul.ftz.f32 %f82, %f57, %f81;\n"
|
|
" mul.ftz.f32 %f83, %f71, %f82;\n"
|
|
" sub.ftz.f32 %f84, %f57, %f69;\n"
|
|
" mul.ftz.f32 %f85, %f83, %f84;\n"
|
|
" div.approx.ftz.f32 %f86, %f85, %f77;\n"
|
|
" ld.global.v2.f32 {%f87,%f88}, [%rd32+8];\n"
|
|
" mul.ftz.f32 %f89, %f87, %f61;\n"
|
|
" sub.ftz.f32 %f90, %f89, %f88;\n"
|
|
" mul.ftz.f32 %f91, %f61, %f90;\n"
|
|
" mul.ftz.f32 %f92, %f86, %f91;\n"
|
|
" fma.rn.ftz.f32 %f68, %f68, %f79, %f92;\n"
|
|
" bra.uni $Lt_0_25858;\n"
|
|
"$Lt_0_26114:\n"
|
|
" .loc 16 197 0\n"
|
|
" mov.f32 %f68, 0f00000000; \n"
|
|
"$Lt_0_25858:\n"
|
|
" ld.param.f32 %f93, [__cudaparm_kernel_pair_cut_coulsq];\n"
|
|
" setp.gt.ftz.f32 %p7, %f93, %f57;\n"
|
|
" @!%p7 bra $Lt_0_27138;\n"
|
|
" .loc 16 204 0\n"
|
|
" sqrt.approx.ftz.f32 %f94, %f57;\n"
|
|
" ld.param.f32 %f95, [__cudaparm_kernel_pair_g_ewald];\n"
|
|
" mul.ftz.f32 %f96, %f95, %f94;\n"
|
|
" mul.ftz.f32 %f97, %f96, %f96;\n"
|
|
" mov.f32 %f98, 0f3f800000; \n"
|
|
" mov.f32 %f99, 0f3ea7ba05; \n"
|
|
" fma.rn.ftz.f32 %f100, %f99, %f96, %f98;\n"
|
|
" neg.ftz.f32 %f101, %f97;\n"
|
|
" rcp.approx.ftz.f32 %f102, %f100;\n"
|
|
" mov.f32 %f103, 0f3fb8aa3b; \n"
|
|
" mul.ftz.f32 %f104, %f101, %f103;\n"
|
|
" ex2.approx.ftz.f32 %f105, %f104;\n"
|
|
" mov.f32 %f106, 0f3e827906; \n"
|
|
" mov.f32 %f107, 0fbe91a98e; \n"
|
|
" mov.f32 %f108, 0f3fb5f0e3; \n"
|
|
" mov.f32 %f109, 0fbfba00e3; \n"
|
|
" mov.f32 %f110, 0f3f87dc22; \n"
|
|
" fma.rn.ftz.f32 %f111, %f110, %f102, %f109;\n"
|
|
" fma.rn.ftz.f32 %f112, %f102, %f111, %f108;\n"
|
|
" fma.rn.ftz.f32 %f113, %f102, %f112, %f107;\n"
|
|
" fma.rn.ftz.f32 %f114, %f102, %f113, %f106;\n"
|
|
" mul.ftz.f32 %f115, %f102, %f114;\n"
|
|
" mul.ftz.f32 %f116, %f105, %f115;\n"
|
|
" mov.f32 %f117, %f116;\n"
|
|
" .loc 16 205 0\n"
|
|
" mov.u32 %r49, %r36;\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 {%f118,%f119,%f120,%f121},[q_tex,{%r49,%r51,%r53,%r55}];\n"
|
|
" mov.f32 %f122, %f118;\n"
|
|
" ld.param.f32 %f123, [__cudaparm_kernel_pair_qqrd2e];\n"
|
|
" mul.ftz.f32 %f124, %f123, %f33;\n"
|
|
" mul.ftz.f32 %f125, %f124, %f122;\n"
|
|
" div.approx.ftz.f32 %f126, %f125, %f94;\n"
|
|
" mov.f32 %f127, %f126;\n"
|
|
" .loc 16 206 0\n"
|
|
" mov.f32 %f128, 0f3f906ebb; \n"
|
|
" mul.ftz.f32 %f129, %f96, %f128;\n"
|
|
" fma.rn.ftz.f32 %f130, %f105, %f129, %f116;\n"
|
|
" sub.ftz.f32 %f131, %f130, %f43;\n"
|
|
" mul.ftz.f32 %f132, %f126, %f131;\n"
|
|
" bra.uni $Lt_0_26882;\n"
|
|
"$Lt_0_27138:\n"
|
|
" .loc 16 208 0\n"
|
|
" mov.f32 %f132, 0f00000000; \n"
|
|
"$Lt_0_26882:\n"
|
|
" .loc 16 212 0\n"
|
|
" add.ftz.f32 %f133, %f132, %f68;\n"
|
|
" mul.ftz.f32 %f134, %f133, %f59;\n"
|
|
" fma.rn.ftz.f32 %f37, %f53, %f134, %f37;\n"
|
|
" .loc 16 213 0\n"
|
|
" fma.rn.ftz.f32 %f36, %f52, %f134, %f36;\n"
|
|
" .loc 16 214 0\n"
|
|
" fma.rn.ftz.f32 %f35, %f54, %f134, %f35;\n"
|
|
" ld.param.s32 %r56, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r57, 0;\n"
|
|
" setp.le.s32 %p8, %r56, %r57;\n"
|
|
" @%p8 bra $Lt_0_27906;\n"
|
|
" .loc 16 217 0\n"
|
|
" mov.f32 %f135, %f127;\n"
|
|
" mov.f32 %f136, %f117;\n"
|
|
" sub.ftz.f32 %f137, %f136, %f43;\n"
|
|
" fma.rn.ftz.f32 %f138, %f135, %f137, %f38;\n"
|
|
" selp.f32 %f38, %f138, %f38, %p7;\n"
|
|
" @!%p5 bra $Lt_0_27906;\n"
|
|
" .loc 16 220 0\n"
|
|
" cvt.rzi.ftz.s32.f32 %r58, %f51;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r59, %f28;\n"
|
|
" ld.param.u64 %rd33, [__cudaparm_kernel_pair_lj1];\n"
|
|
" ld.param.s32 %r60, [__cudaparm_kernel_pair_lj_types];\n"
|
|
" mul.lo.s32 %r61, %r60, %r59;\n"
|
|
" add.s32 %r62, %r58, %r61;\n"
|
|
" cvt.s64.s32 %rd34, %r62;\n"
|
|
" mul.wide.s32 %rd35, %r62, 16;\n"
|
|
" add.u64 %rd32, %rd33, %rd35;\n"
|
|
" mov.f32 %f139, %f62;\n"
|
|
" ld.global.v2.f32 {%f140,%f141}, [%rd32+8];\n"
|
|
" mul.ftz.f32 %f142, %f140, %f139;\n"
|
|
" sub.ftz.f32 %f143, %f142, %f141;\n"
|
|
" mul.ftz.f32 %f144, %f139, %f143;\n"
|
|
" mov.f32 %f145, %f80;\n"
|
|
" mul.ftz.f32 %f146, %f145, %f144;\n"
|
|
" ld.param.f32 %f147, [__cudaparm_kernel_pair_cut_lj_innersq];\n"
|
|
" setp.lt.ftz.f32 %p9, %f147, %f57;\n"
|
|
" selp.f32 %f148, %f146, %f144, %p9;\n"
|
|
" .loc 16 223 0\n"
|
|
" fma.rn.ftz.f32 %f39, %f40, %f148, %f39;\n"
|
|
"$Lt_0_27906:\n"
|
|
"$Lt_0_27394:\n"
|
|
" ld.param.s32 %r63, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r64, 0;\n"
|
|
" setp.le.s32 %p10, %r63, %r64;\n"
|
|
" @%p10 bra $Lt_0_28418;\n"
|
|
" .loc 16 227 0\n"
|
|
" mov.f32 %f149, %f10;\n"
|
|
" mul.ftz.f32 %f150, %f53, %f53;\n"
|
|
" fma.rn.ftz.f32 %f151, %f134, %f150, %f149;\n"
|
|
" mov.f32 %f10, %f151;\n"
|
|
" .loc 16 228 0\n"
|
|
" mov.f32 %f152, %f12;\n"
|
|
" fma.rn.ftz.f32 %f153, %f134, %f55, %f152;\n"
|
|
" mov.f32 %f12, %f153;\n"
|
|
" .loc 16 229 0\n"
|
|
" mov.f32 %f154, %f14;\n"
|
|
" mul.ftz.f32 %f155, %f54, %f54;\n"
|
|
" fma.rn.ftz.f32 %f156, %f134, %f155, %f154;\n"
|
|
" mov.f32 %f14, %f156;\n"
|
|
" .loc 16 230 0\n"
|
|
" mov.f32 %f157, %f16;\n"
|
|
" mul.ftz.f32 %f158, %f52, %f53;\n"
|
|
" fma.rn.ftz.f32 %f159, %f134, %f158, %f157;\n"
|
|
" mov.f32 %f16, %f159;\n"
|
|
" .loc 16 231 0\n"
|
|
" mov.f32 %f160, %f18;\n"
|
|
" mul.ftz.f32 %f161, %f53, %f54;\n"
|
|
" fma.rn.ftz.f32 %f162, %f134, %f161, %f160;\n"
|
|
" mov.f32 %f18, %f162;\n"
|
|
" .loc 16 232 0\n"
|
|
" mul.ftz.f32 %f163, %f52, %f54;\n"
|
|
" fma.rn.ftz.f32 %f19, %f134, %f163, %f19;\n"
|
|
" mov.f32 %f20, %f19;\n"
|
|
"$Lt_0_28418:\n"
|
|
"$Lt_0_25346:\n"
|
|
" .loc 16 161 0\n"
|
|
" mul.lo.u64 %rd36, %rd24, 4;\n"
|
|
" add.u64 %rd16, %rd16, %rd36;\n"
|
|
" setp.lt.u64 %p11, %rd16, %rd13;\n"
|
|
" @%p11 bra $Lt_0_25090;\n"
|
|
" bra.uni $Lt_0_23554;\n"
|
|
"$Lt_0_34818:\n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
" bra.uni $Lt_0_23554;\n"
|
|
"$Lt_0_23810:\n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" mov.f32 %f37, 0f00000000; \n"
|
|
" mov.f32 %f38, 0f00000000; \n"
|
|
" mov.f32 %f39, 0f00000000; \n"
|
|
"$Lt_0_23554:\n"
|
|
" mov.u32 %r65, 1;\n"
|
|
" setp.le.s32 %p12, %r1, %r65;\n"
|
|
" @%p12 bra $Lt_0_31234;\n"
|
|
" .loc 16 243 0\n"
|
|
" mov.u64 %rd37, __cuda___cuda_local_var_32624_35_non_const_red_acc152;\n"
|
|
" cvt.s64.s32 %rd38, %r2;\n"
|
|
" mul.wide.s32 %rd39, %r2, 4;\n"
|
|
" add.u64 %rd40, %rd37, %rd39;\n"
|
|
" mov.f32 %f164, %f37;\n"
|
|
" st.shared.f32 [%rd40+0], %f164;\n"
|
|
" .loc 16 244 0\n"
|
|
" mov.f32 %f165, %f36;\n"
|
|
" st.shared.f32 [%rd40+512], %f165;\n"
|
|
" .loc 16 245 0\n"
|
|
" mov.f32 %f166, %f35;\n"
|
|
" st.shared.f32 [%rd40+1024], %f166;\n"
|
|
" .loc 16 246 0\n"
|
|
" mov.f32 %f167, %f39;\n"
|
|
" st.shared.f32 [%rd40+1536], %f167;\n"
|
|
" .loc 16 247 0\n"
|
|
" mov.f32 %f168, %f38;\n"
|
|
" st.shared.f32 [%rd40+2048], %f168;\n"
|
|
" .loc 16 249 0\n"
|
|
" shr.s32 %r66, %r1, 31;\n"
|
|
" mov.s32 %r67, 1;\n"
|
|
" and.b32 %r68, %r66, %r67;\n"
|
|
" add.s32 %r69, %r68, %r1;\n"
|
|
" shr.s32 %r70, %r69, 1;\n"
|
|
" mov.s32 %r71, %r70;\n"
|
|
" mov.u32 %r72, 0;\n"
|
|
" setp.ne.u32 %p13, %r70, %r72;\n"
|
|
" @!%p13 bra $Lt_0_29698;\n"
|
|
"$Lt_0_30210:\n"
|
|
" setp.ge.u32 %p14, %r6, %r71;\n"
|
|
" @%p14 bra $Lt_0_30466;\n"
|
|
" .loc 16 252 0\n"
|
|
" add.u32 %r73, %r2, %r71;\n"
|
|
" cvt.u64.u32 %rd41, %r73;\n"
|
|
" mul.wide.u32 %rd42, %r73, 4;\n"
|
|
" add.u64 %rd43, %rd37, %rd42;\n"
|
|
" ld.shared.f32 %f169, [%rd43+0];\n"
|
|
" add.ftz.f32 %f164, %f169, %f164;\n"
|
|
" st.shared.f32 [%rd40+0], %f164;\n"
|
|
" ld.shared.f32 %f170, [%rd43+512];\n"
|
|
" add.ftz.f32 %f165, %f170, %f165;\n"
|
|
" st.shared.f32 [%rd40+512], %f165;\n"
|
|
" ld.shared.f32 %f171, [%rd43+1024];\n"
|
|
" add.ftz.f32 %f166, %f171, %f166;\n"
|
|
" st.shared.f32 [%rd40+1024], %f166;\n"
|
|
" ld.shared.f32 %f172, [%rd43+1536];\n"
|
|
" add.ftz.f32 %f167, %f172, %f167;\n"
|
|
" st.shared.f32 [%rd40+1536], %f167;\n"
|
|
" ld.shared.f32 %f173, [%rd43+2048];\n"
|
|
" add.ftz.f32 %f168, %f173, %f168;\n"
|
|
" st.shared.f32 [%rd40+2048], %f168;\n"
|
|
"$Lt_0_30466:\n"
|
|
" .loc 16 249 0\n"
|
|
" shr.u32 %r71, %r71, 1;\n"
|
|
" mov.u32 %r74, 0;\n"
|
|
" setp.ne.u32 %p15, %r71, %r74;\n"
|
|
" @%p15 bra $Lt_0_30210;\n"
|
|
"$Lt_0_29698:\n"
|
|
" .loc 16 256 0\n"
|
|
" mov.f32 %f37, %f164;\n"
|
|
" .loc 16 257 0\n"
|
|
" mov.f32 %f36, %f165;\n"
|
|
" .loc 16 258 0\n"
|
|
" mov.f32 %f35, %f166;\n"
|
|
" .loc 16 259 0\n"
|
|
" mov.f32 %f39, %f167;\n"
|
|
" .loc 16 260 0\n"
|
|
" mov.f32 %f38, %f168;\n"
|
|
" ld.param.s32 %r75, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r76, 0;\n"
|
|
" setp.le.s32 %p16, %r75, %r76;\n"
|
|
" @%p16 bra $Lt_0_31234;\n"
|
|
" .loc 16 264 0\n"
|
|
" mov.f32 %f164, %f10;\n"
|
|
" st.shared.f32 [%rd40+0], %f164;\n"
|
|
" mov.f32 %f165, %f12;\n"
|
|
" st.shared.f32 [%rd40+512], %f165;\n"
|
|
" mov.f32 %f166, %f14;\n"
|
|
" st.shared.f32 [%rd40+1024], %f166;\n"
|
|
" mov.f32 %f167, %f16;\n"
|
|
" st.shared.f32 [%rd40+1536], %f167;\n"
|
|
" mov.f32 %f168, %f18;\n"
|
|
" st.shared.f32 [%rd40+2048], %f168;\n"
|
|
" mov.f32 %f174, %f20;\n"
|
|
" st.shared.f32 [%rd40+2560], %f174;\n"
|
|
" .loc 16 266 0\n"
|
|
" mov.s32 %r77, %r70;\n"
|
|
" @!%p13 bra $Lt_0_31746;\n"
|
|
"$Lt_0_32258:\n"
|
|
" setp.ge.u32 %p17, %r6, %r77;\n"
|
|
" @%p17 bra $Lt_0_32514;\n"
|
|
" .loc 16 269 0\n"
|
|
" add.u32 %r78, %r2, %r77;\n"
|
|
" cvt.u64.u32 %rd44, %r78;\n"
|
|
" mul.wide.u32 %rd45, %r78, 4;\n"
|
|
" add.u64 %rd46, %rd37, %rd45;\n"
|
|
" ld.shared.f32 %f175, [%rd46+0];\n"
|
|
" add.ftz.f32 %f164, %f175, %f164;\n"
|
|
" st.shared.f32 [%rd40+0], %f164;\n"
|
|
" ld.shared.f32 %f176, [%rd46+512];\n"
|
|
" add.ftz.f32 %f165, %f176, %f165;\n"
|
|
" st.shared.f32 [%rd40+512], %f165;\n"
|
|
" ld.shared.f32 %f177, [%rd46+1024];\n"
|
|
" add.ftz.f32 %f166, %f177, %f166;\n"
|
|
" st.shared.f32 [%rd40+1024], %f166;\n"
|
|
" ld.shared.f32 %f178, [%rd46+1536];\n"
|
|
" add.ftz.f32 %f167, %f178, %f167;\n"
|
|
" st.shared.f32 [%rd40+1536], %f167;\n"
|
|
" ld.shared.f32 %f179, [%rd46+2048];\n"
|
|
" add.ftz.f32 %f168, %f179, %f168;\n"
|
|
" st.shared.f32 [%rd40+2048], %f168;\n"
|
|
" ld.shared.f32 %f180, [%rd46+2560];\n"
|
|
" add.ftz.f32 %f174, %f180, %f174;\n"
|
|
" st.shared.f32 [%rd40+2560], %f174;\n"
|
|
"$Lt_0_32514:\n"
|
|
" .loc 16 266 0\n"
|
|
" shr.u32 %r77, %r77, 1;\n"
|
|
" mov.u32 %r79, 0;\n"
|
|
" setp.ne.u32 %p18, %r77, %r79;\n"
|
|
" @%p18 bra $Lt_0_32258;\n"
|
|
"$Lt_0_31746:\n"
|
|
" .loc 16 274 0\n"
|
|
" mov.f32 %f10, %f164;\n"
|
|
" mov.f32 %f12, %f165;\n"
|
|
" mov.f32 %f14, %f166;\n"
|
|
" mov.f32 %f16, %f167;\n"
|
|
" mov.f32 %f18, %f168;\n"
|
|
" mov.f32 %f20, %f174;\n"
|
|
"$Lt_0_31234:\n"
|
|
"$Lt_0_29186:\n"
|
|
" selp.s32 %r80, 1, 0, %p1;\n"
|
|
" mov.s32 %r81, 0;\n"
|
|
" set.eq.u32.s32 %r82, %r6, %r81;\n"
|
|
" neg.s32 %r83, %r82;\n"
|
|
" and.b32 %r84, %r80, %r83;\n"
|
|
" mov.u32 %r85, 0;\n"
|
|
" setp.eq.s32 %p19, %r84, %r85;\n"
|
|
" @%p19 bra $Lt_0_33282;\n"
|
|
" .loc 16 280 0\n"
|
|
" cvt.s64.s32 %rd47, %r9;\n"
|
|
" ld.param.u64 %rd48, [__cudaparm_kernel_pair_engv];\n"
|
|
" mul.wide.s32 %rd49, %r9, 4;\n"
|
|
" add.u64 %rd50, %rd48, %rd49;\n"
|
|
" ld.param.s32 %r86, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r87, 0;\n"
|
|
" setp.le.s32 %p20, %r86, %r87;\n"
|
|
" @%p20 bra $Lt_0_33794;\n"
|
|
" .loc 16 282 0\n"
|
|
" st.global.f32 [%rd50+0], %f39;\n"
|
|
" .loc 16 283 0\n"
|
|
" cvt.s64.s32 %rd51, %r10;\n"
|
|
" mul.wide.s32 %rd52, %r10, 4;\n"
|
|
" add.u64 %rd53, %rd52, %rd50;\n"
|
|
" .loc 16 284 0\n"
|
|
" st.global.f32 [%rd53+0], %f38;\n"
|
|
" .loc 16 285 0\n"
|
|
" add.u64 %rd50, %rd52, %rd53;\n"
|
|
"$Lt_0_33794:\n"
|
|
" ld.param.s32 %r88, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r89, 0;\n"
|
|
" setp.le.s32 %p21, %r88, %r89;\n"
|
|
" @%p21 bra $Lt_0_34306;\n"
|
|
" .loc 16 289 0\n"
|
|
" mov.f32 %f181, %f10;\n"
|
|
" st.global.f32 [%rd50+0], %f181;\n"
|
|
" .loc 16 290 0\n"
|
|
" cvt.s64.s32 %rd54, %r10;\n"
|
|
" mul.wide.s32 %rd55, %r10, 4;\n"
|
|
" add.u64 %rd56, %rd55, %rd50;\n"
|
|
" .loc 16 289 0\n"
|
|
" mov.f32 %f182, %f12;\n"
|
|
" st.global.f32 [%rd56+0], %f182;\n"
|
|
" .loc 16 290 0\n"
|
|
" add.u64 %rd57, %rd55, %rd56;\n"
|
|
" .loc 16 289 0\n"
|
|
" mov.f32 %f183, %f14;\n"
|
|
" st.global.f32 [%rd57+0], %f183;\n"
|
|
" .loc 16 290 0\n"
|
|
" add.u64 %rd58, %rd55, %rd57;\n"
|
|
" .loc 16 289 0\n"
|
|
" mov.f32 %f184, %f16;\n"
|
|
" st.global.f32 [%rd58+0], %f184;\n"
|
|
" .loc 16 290 0\n"
|
|
" add.u64 %rd50, %rd55, %rd58;\n"
|
|
" .loc 16 289 0\n"
|
|
" mov.f32 %f185, %f18;\n"
|
|
" st.global.f32 [%rd50+0], %f185;\n"
|
|
" mov.f32 %f186, %f20;\n"
|
|
" add.u64 %rd59, %rd55, %rd50;\n"
|
|
" st.global.f32 [%rd59+0], %f186;\n"
|
|
"$Lt_0_34306:\n"
|
|
" .loc 16 293 0\n"
|
|
" ld.param.u64 %rd60, [__cudaparm_kernel_pair_ans];\n"
|
|
" mul.lo.u64 %rd61, %rd47, 16;\n"
|
|
" add.u64 %rd62, %rd60, %rd61;\n"
|
|
" mov.f32 %f187, %f188;\n"
|
|
" st.global.v4.f32 [%rd62+0], {%f37,%f36,%f35,%f187};\n"
|
|
"$Lt_0_33282:\n"
|
|
" .loc 16 295 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair:\n"
|
|
" }\n"
|
|
" .entry kernel_pair_fast (\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_x_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_ljd_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 .f32 __cudaparm_kernel_pair_fast_cut_coulsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_qqrd2e,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_g_ewald,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_denom_lj,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_cut_bothsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_cut_ljsq,\n"
|
|
" .param .f32 __cudaparm_kernel_pair_fast_cut_lj_innersq,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<86>;\n"
|
|
" .reg .u64 %rd<71>;\n"
|
|
" .reg .f32 %f<196>;\n"
|
|
" .reg .pred %p<25>;\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32697_33_non_const_sp_lj3336[32];\n"
|
|
" .shared .align 8 .b8 __cuda___cuda_local_var_32696_34_non_const_ljd3368[1024];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32826_35_non_const_red_acc4392[3072];\n"
|
|
" .loc 16 307 0\n"
|
|
"$LDWbegin_kernel_pair_fast:\n"
|
|
" cvt.s32.u32 %r1, %tid.x;\n"
|
|
" cvt.s64.s32 %rd1, %r1;\n"
|
|
" mov.u32 %r2, 7;\n"
|
|
" setp.gt.s32 %p1, %r1, %r2;\n"
|
|
" @%p1 bra $Lt_1_25090;\n"
|
|
" .loc 16 316 0\n"
|
|
" mov.u64 %rd2, __cuda___cuda_local_var_32697_33_non_const_sp_lj3336;\n"
|
|
" mul.lo.u64 %rd3, %rd1, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_fast_sp_lj_in];\n"
|
|
" add.u64 %rd5, %rd4, %rd3;\n"
|
|
" ld.global.f32 %f1, [%rd5+0];\n"
|
|
" add.u64 %rd6, %rd3, %rd2;\n"
|
|
" st.shared.f32 [%rd6+0], %f1;\n"
|
|
"$Lt_1_25090:\n"
|
|
" mov.u64 %rd2, __cuda___cuda_local_var_32697_33_non_const_sp_lj3336;\n"
|
|
" mov.u64 %rd7, __cuda___cuda_local_var_32696_34_non_const_ljd3368;\n"
|
|
" .loc 16 317 0\n"
|
|
" mul.lo.u64 %rd8, %rd1, 8;\n"
|
|
" ld.param.u64 %rd9, [__cudaparm_kernel_pair_fast_ljd_in];\n"
|
|
" add.u64 %rd10, %rd9, %rd8;\n"
|
|
" add.u64 %rd11, %rd8, %rd7;\n"
|
|
" ld.global.v2.f32 {%f2,%f3}, [%rd10+0];\n"
|
|
" st.shared.v2.f32 [%rd11+0], {%f2,%f3};\n"
|
|
" add.s32 %r3, %r1, 128;\n"
|
|
" mov.u32 %r4, 127;\n"
|
|
" setp.gt.s32 %p2, %r3, %r4;\n"
|
|
" @%p2 bra $Lt_1_25602;\n"
|
|
" ld.global.v2.f32 {%f4,%f5}, [%rd10+1024];\n"
|
|
" st.shared.v2.f32 [%rd11+1024], {%f4,%f5};\n"
|
|
"$Lt_1_25602:\n"
|
|
" .loc 16 329 0\n"
|
|
" mov.f32 %f6, 0f00000000; \n"
|
|
" mov.f32 %f7, %f6;\n"
|
|
" mov.f32 %f8, 0f00000000; \n"
|
|
" mov.f32 %f9, %f8;\n"
|
|
" mov.f32 %f10, 0f00000000; \n"
|
|
" mov.f32 %f11, %f10;\n"
|
|
" mov.f32 %f12, 0f00000000; \n"
|
|
" mov.f32 %f13, %f12;\n"
|
|
" mov.f32 %f14, 0f00000000; \n"
|
|
" mov.f32 %f15, %f14;\n"
|
|
" mov.f32 %f16, 0f00000000; \n"
|
|
" mov.f32 %f17, %f16;\n"
|
|
" .loc 16 331 0\n"
|
|
" bar.sync 0;\n"
|
|
" ld.param.s32 %r5, [__cudaparm_kernel_pair_fast_t_per_atom];\n"
|
|
" div.s32 %r6, %r1, %r5;\n"
|
|
" cvt.s32.u32 %r7, %ntid.x;\n"
|
|
" div.s32 %r8, %r7, %r5;\n"
|
|
" rem.s32 %r9, %r1, %r5;\n"
|
|
" cvt.s32.u32 %r10, %ctaid.x;\n"
|
|
" mul.lo.s32 %r11, %r10, %r8;\n"
|
|
" add.s32 %r12, %r6, %r11;\n"
|
|
" ld.param.s32 %r13, [__cudaparm_kernel_pair_fast_inum];\n"
|
|
" setp.lt.s32 %p3, %r12, %r13;\n"
|
|
" @!%p3 bra $Lt_1_26370;\n"
|
|
" .loc 16 335 0\n"
|
|
" cvt.s64.s32 %rd12, %r12;\n"
|
|
" mul.wide.s32 %rd13, %r12, 4;\n"
|
|
" ld.param.u64 %rd14, [__cudaparm_kernel_pair_fast_dev_nbor];\n"
|
|
" add.u64 %rd15, %rd13, %rd14;\n"
|
|
" ld.global.s32 %r14, [%rd15+0];\n"
|
|
" .loc 16 337 0\n"
|
|
" ld.param.s32 %r15, [__cudaparm_kernel_pair_fast_nbor_pitch];\n"
|
|
" cvt.s64.s32 %rd16, %r15;\n"
|
|
" mul.wide.s32 %rd17, %r15, 4;\n"
|
|
" add.u64 %rd18, %rd17, %rd15;\n"
|
|
" ld.global.s32 %r16, [%rd18+0];\n"
|
|
" add.u64 %rd19, %rd17, %rd18;\n"
|
|
" ld.param.u64 %rd20, [__cudaparm_kernel_pair_fast_dev_packed];\n"
|
|
" setp.ne.u64 %p4, %rd20, %rd14;\n"
|
|
" @%p4 bra $Lt_1_26882;\n"
|
|
" .loc 16 343 0\n"
|
|
" cvt.s32.s64 %r17, %rd16;\n"
|
|
" mul.lo.s32 %r18, %r17, %r16;\n"
|
|
" cvt.s64.s32 %rd21, %r18;\n"
|
|
" mul.wide.s32 %rd22, %r18, 4;\n"
|
|
" add.u64 %rd23, %rd19, %rd22;\n"
|
|
" .loc 16 344 0\n"
|
|
" mul.lo.s32 %r19, %r9, %r17;\n"
|
|
" cvt.s64.s32 %rd24, %r19;\n"
|
|
" mul.wide.s32 %rd25, %r19, 4;\n"
|
|
" add.u64 %rd26, %rd19, %rd25;\n"
|
|
" .loc 16 345 0\n"
|
|
" mul.lo.s32 %r20, %r17, %r5;\n"
|
|
" bra.uni $Lt_1_26626;\n"
|
|
"$Lt_1_26882:\n"
|
|
" .loc 16 347 0\n"
|
|
" ld.global.s32 %r21, [%rd19+0];\n"
|
|
" cvt.s64.s32 %rd27, %r21;\n"
|
|
" mul.wide.s32 %rd28, %r21, 4;\n"
|
|
" add.u64 %rd29, %rd20, %rd28;\n"
|
|
" .loc 16 348 0\n"
|
|
" cvt.s64.s32 %rd30, %r16;\n"
|
|
" mul.wide.s32 %rd31, %r16, 4;\n"
|
|
" add.u64 %rd23, %rd29, %rd31;\n"
|
|
" .loc 16 349 0\n"
|
|
" mov.s32 %r20, %r5;\n"
|
|
" .loc 16 350 0\n"
|
|
" cvt.s64.s32 %rd32, %r9;\n"
|
|
" mul.wide.s32 %rd33, %r9, 4;\n"
|
|
" add.u64 %rd26, %rd29, %rd33;\n"
|
|
"$Lt_1_26626:\n"
|
|
" .loc 16 353 0\n"
|
|
" mov.u32 %r22, %r14;\n"
|
|
" mov.s32 %r23, 0;\n"
|
|
" mov.u32 %r24, %r23;\n"
|
|
" mov.s32 %r25, 0;\n"
|
|
" mov.u32 %r26, %r25;\n"
|
|
" mov.s32 %r27, 0;\n"
|
|
" mov.u32 %r28, %r27;\n"
|
|
" tex.1d.v4.f32.s32 {%f18,%f19,%f20,%f21},[pos_tex,{%r22,%r24,%r26,%r28}];\n"
|
|
" mov.f32 %f22, %f18;\n"
|
|
" mov.f32 %f23, %f19;\n"
|
|
" mov.f32 %f24, %f20;\n"
|
|
" mov.f32 %f25, %f21;\n"
|
|
" .loc 16 354 0\n"
|
|
" mov.u32 %r29, %r14;\n"
|
|
" mov.s32 %r30, 0;\n"
|
|
" mov.u32 %r31, %r30;\n"
|
|
" mov.s32 %r32, 0;\n"
|
|
" mov.u32 %r33, %r32;\n"
|
|
" mov.s32 %r34, 0;\n"
|
|
" mov.u32 %r35, %r34;\n"
|
|
" tex.1d.v4.f32.s32 {%f26,%f27,%f28,%f29},[q_tex,{%r29,%r31,%r33,%r35}];\n"
|
|
" mov.f32 %f30, %f26;\n"
|
|
" setp.ge.u64 %p5, %rd26, %rd23;\n"
|
|
" @%p5 bra $Lt_1_37378;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r36, %f25;\n"
|
|
" cvt.s64.s32 %rd34, %r20;\n"
|
|
" ld.param.f32 %f31, [__cudaparm_kernel_pair_fast_cut_bothsq];\n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
" mov.f32 %f33, 0f00000000; \n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
"$Lt_1_27650:\n"
|
|
" .loc 16 358 0\n"
|
|
" ld.global.s32 %r37, [%rd26+0];\n"
|
|
" .loc 16 361 0\n"
|
|
" shr.s32 %r38, %r37, 30;\n"
|
|
" and.b32 %r39, %r38, 3;\n"
|
|
" cvt.s64.s32 %rd35, %r39;\n"
|
|
" mul.wide.s32 %rd36, %r39, 4;\n"
|
|
" add.u64 %rd37, %rd2, %rd36;\n"
|
|
" ld.shared.f32 %f37, [%rd37+0];\n"
|
|
" .loc 16 362 0\n"
|
|
" mov.f32 %f38, 0f3f800000; \n"
|
|
" ld.shared.f32 %f39, [%rd37+16];\n"
|
|
" sub.ftz.f32 %f40, %f38, %f39;\n"
|
|
" .loc 16 365 0\n"
|
|
" and.b32 %r40, %r37, 1073741823;\n"
|
|
" mov.u32 %r41, %r40;\n"
|
|
" mov.s32 %r42, 0;\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"
|
|
" tex.1d.v4.f32.s32 {%f41,%f42,%f43,%f44},[pos_tex,{%r41,%r43,%r45,%r47}];\n"
|
|
" mov.f32 %f45, %f41;\n"
|
|
" mov.f32 %f46, %f42;\n"
|
|
" mov.f32 %f47, %f43;\n"
|
|
" mov.f32 %f48, %f44;\n"
|
|
" sub.ftz.f32 %f49, %f23, %f46;\n"
|
|
" sub.ftz.f32 %f50, %f22, %f45;\n"
|
|
" sub.ftz.f32 %f51, %f24, %f47;\n"
|
|
" mul.ftz.f32 %f52, %f49, %f49;\n"
|
|
" fma.rn.ftz.f32 %f53, %f50, %f50, %f52;\n"
|
|
" fma.rn.ftz.f32 %f54, %f51, %f51, %f53;\n"
|
|
" setp.lt.ftz.f32 %p6, %f54, %f31;\n"
|
|
" @!%p6 bra $Lt_1_30978;\n"
|
|
" ld.param.f32 %f55, [__cudaparm_kernel_pair_fast_cut_ljsq];\n"
|
|
" setp.lt.ftz.f32 %p7, %f54, %f55;\n"
|
|
" rcp.approx.ftz.f32 %f56, %f54;\n"
|
|
" @!%p7 bra $Lt_1_28674;\n"
|
|
" .loc 16 380 0\n"
|
|
" cvt.rzi.ftz.s32.f32 %r48, %f48;\n"
|
|
" cvt.s64.s32 %rd38, %r36;\n"
|
|
" mul.wide.s32 %rd39, %r36, 8;\n"
|
|
" add.u64 %rd40, %rd7, %rd39;\n"
|
|
" cvt.s64.s32 %rd41, %r48;\n"
|
|
" mul.wide.s32 %rd42, %r48, 8;\n"
|
|
" add.u64 %rd43, %rd7, %rd42;\n"
|
|
" ld.shared.v2.f32 {%f57,%f58}, [%rd40+0];\n"
|
|
" ld.shared.v2.f32 {%f59,%f60}, [%rd43+0];\n"
|
|
" mul.ftz.f32 %f61, %f57, %f59;\n"
|
|
" .loc 16 381 0\n"
|
|
" add.ftz.f32 %f62, %f58, %f60;\n"
|
|
" mov.f32 %f63, 0f3f000000; \n"
|
|
" mul.ftz.f32 %f64, %f62, %f63;\n"
|
|
" .loc 16 385 0\n"
|
|
" mul.ftz.f32 %f65, %f64, %f64;\n"
|
|
" sqrt.approx.ftz.f32 %f66, %f61;\n"
|
|
" mov.f32 %f67, 0f40800000; \n"
|
|
" mul.ftz.f32 %f68, %f66, %f67;\n"
|
|
" mul.ftz.f32 %f69, %f65, %f56;\n"
|
|
" mul.ftz.f32 %f70, %f69, %f69;\n"
|
|
" mul.ftz.f32 %f71, %f69, %f70;\n"
|
|
" mul.ftz.f32 %f72, %f68, %f71;\n"
|
|
" mov.f32 %f73, %f72;\n"
|
|
" .loc 16 386 0\n"
|
|
" mul.ftz.f32 %f74, %f71, %f72;\n"
|
|
" mov.f32 %f75, %f74;\n"
|
|
" .loc 16 387 0\n"
|
|
" mov.f32 %f76, 0f40c00000; \n"
|
|
" mul.ftz.f32 %f77, %f72, %f76;\n"
|
|
" mov.f32 %f78, 0f41400000; \n"
|
|
" mul.ftz.f32 %f79, %f78, %f74;\n"
|
|
" sub.ftz.f32 %f80, %f79, %f77;\n"
|
|
" mul.ftz.f32 %f81, %f37, %f80;\n"
|
|
" ld.param.f32 %f82, [__cudaparm_kernel_pair_fast_cut_lj_innersq];\n"
|
|
" setp.gt.ftz.f32 %p8, %f54, %f82;\n"
|
|
" @!%p8 bra $Lt_1_28418;\n"
|
|
" .loc 16 393 0\n"
|
|
" add.ftz.f32 %f83, %f54, %f54;\n"
|
|
" sub.ftz.f32 %f84, %f55, %f54;\n"
|
|
" add.ftz.f32 %f85, %f83, %f55;\n"
|
|
" mul.ftz.f32 %f86, %f84, %f84;\n"
|
|
" mov.f32 %f87, 0f40400000; \n"
|
|
" mul.ftz.f32 %f88, %f87, %f82;\n"
|
|
" sub.ftz.f32 %f89, %f85, %f88;\n"
|
|
" ld.param.f32 %f90, [__cudaparm_kernel_pair_fast_denom_lj];\n"
|
|
" div.approx.ftz.f32 %f91, %f89, %f90;\n"
|
|
" mul.ftz.f32 %f92, %f86, %f91;\n"
|
|
" mov.f32 %f93, %f92;\n"
|
|
" .loc 16 396 0\n"
|
|
" mov.f32 %f94, 0f41400000; \n"
|
|
" mul.ftz.f32 %f95, %f54, %f94;\n"
|
|
" mul.ftz.f32 %f96, %f84, %f95;\n"
|
|
" sub.ftz.f32 %f97, %f54, %f82;\n"
|
|
" mul.ftz.f32 %f98, %f96, %f97;\n"
|
|
" div.approx.ftz.f32 %f99, %f98, %f90;\n"
|
|
" sub.ftz.f32 %f100, %f74, %f72;\n"
|
|
" mul.ftz.f32 %f101, %f99, %f100;\n"
|
|
" fma.rn.ftz.f32 %f81, %f81, %f92, %f101;\n"
|
|
" bra.uni $Lt_1_28418;\n"
|
|
"$Lt_1_28674:\n"
|
|
" .loc 16 399 0\n"
|
|
" mov.f32 %f81, 0f00000000; \n"
|
|
"$Lt_1_28418:\n"
|
|
" ld.param.f32 %f102, [__cudaparm_kernel_pair_fast_cut_coulsq];\n"
|
|
" setp.gt.ftz.f32 %p9, %f102, %f54;\n"
|
|
" @!%p9 bra $Lt_1_29698;\n"
|
|
" .loc 16 406 0\n"
|
|
" sqrt.approx.ftz.f32 %f103, %f54;\n"
|
|
" ld.param.f32 %f104, [__cudaparm_kernel_pair_fast_g_ewald];\n"
|
|
" mul.ftz.f32 %f105, %f104, %f103;\n"
|
|
" mul.ftz.f32 %f106, %f105, %f105;\n"
|
|
" mov.f32 %f107, 0f3f800000; \n"
|
|
" mov.f32 %f108, 0f3ea7ba05; \n"
|
|
" fma.rn.ftz.f32 %f109, %f108, %f105, %f107;\n"
|
|
" neg.ftz.f32 %f110, %f106;\n"
|
|
" rcp.approx.ftz.f32 %f111, %f109;\n"
|
|
" mov.f32 %f112, 0f3fb8aa3b; \n"
|
|
" mul.ftz.f32 %f113, %f110, %f112;\n"
|
|
" ex2.approx.ftz.f32 %f114, %f113;\n"
|
|
" mov.f32 %f115, 0f3e827906; \n"
|
|
" mov.f32 %f116, 0fbe91a98e; \n"
|
|
" mov.f32 %f117, 0f3fb5f0e3; \n"
|
|
" mov.f32 %f118, 0fbfba00e3; \n"
|
|
" mov.f32 %f119, 0f3f87dc22; \n"
|
|
" fma.rn.ftz.f32 %f120, %f119, %f111, %f118;\n"
|
|
" fma.rn.ftz.f32 %f121, %f111, %f120, %f117;\n"
|
|
" fma.rn.ftz.f32 %f122, %f111, %f121, %f116;\n"
|
|
" fma.rn.ftz.f32 %f123, %f111, %f122, %f115;\n"
|
|
" mul.ftz.f32 %f124, %f111, %f123;\n"
|
|
" mul.ftz.f32 %f125, %f114, %f124;\n"
|
|
" mov.f32 %f126, %f125;\n"
|
|
" .loc 16 407 0\n"
|
|
" mov.u32 %r49, %r40;\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 {%f127,%f128,%f129,%f130},[q_tex,{%r49,%r51,%r53,%r55}];\n"
|
|
" mov.f32 %f131, %f127;\n"
|
|
" ld.param.f32 %f132, [__cudaparm_kernel_pair_fast_qqrd2e];\n"
|
|
" mul.ftz.f32 %f133, %f132, %f30;\n"
|
|
" mul.ftz.f32 %f134, %f133, %f131;\n"
|
|
" div.approx.ftz.f32 %f135, %f134, %f103;\n"
|
|
" mov.f32 %f136, %f135;\n"
|
|
" .loc 16 408 0\n"
|
|
" mov.f32 %f137, 0f3f906ebb; \n"
|
|
" mul.ftz.f32 %f138, %f105, %f137;\n"
|
|
" fma.rn.ftz.f32 %f139, %f114, %f138, %f125;\n"
|
|
" sub.ftz.f32 %f140, %f139, %f40;\n"
|
|
" mul.ftz.f32 %f141, %f135, %f140;\n"
|
|
" bra.uni $Lt_1_29442;\n"
|
|
"$Lt_1_29698:\n"
|
|
" .loc 16 410 0\n"
|
|
" mov.f32 %f141, 0f00000000; \n"
|
|
"$Lt_1_29442:\n"
|
|
" .loc 16 414 0\n"
|
|
" add.ftz.f32 %f142, %f141, %f81;\n"
|
|
" mul.ftz.f32 %f143, %f142, %f56;\n"
|
|
" fma.rn.ftz.f32 %f34, %f50, %f143, %f34;\n"
|
|
" .loc 16 415 0\n"
|
|
" fma.rn.ftz.f32 %f33, %f49, %f143, %f33;\n"
|
|
" .loc 16 416 0\n"
|
|
" fma.rn.ftz.f32 %f32, %f51, %f143, %f32;\n"
|
|
" ld.param.s32 %r56, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r57, 0;\n"
|
|
" setp.le.s32 %p10, %r56, %r57;\n"
|
|
" @%p10 bra $Lt_1_30466;\n"
|
|
" .loc 16 419 0\n"
|
|
" mov.f32 %f144, %f136;\n"
|
|
" mov.f32 %f145, %f126;\n"
|
|
" sub.ftz.f32 %f146, %f145, %f40;\n"
|
|
" fma.rn.ftz.f32 %f147, %f144, %f146, %f35;\n"
|
|
" selp.f32 %f35, %f147, %f35, %p9;\n"
|
|
" @!%p7 bra $Lt_1_30466;\n"
|
|
" .loc 16 425 0\n"
|
|
" mov.f32 %f148, %f75;\n"
|
|
" mov.f32 %f149, %f73;\n"
|
|
" sub.ftz.f32 %f150, %f148, %f149;\n"
|
|
" mov.f32 %f151, %f93;\n"
|
|
" mul.ftz.f32 %f152, %f151, %f150;\n"
|
|
" ld.param.f32 %f153, [__cudaparm_kernel_pair_fast_cut_lj_innersq];\n"
|
|
" setp.lt.ftz.f32 %p11, %f153, %f54;\n"
|
|
" selp.f32 %f154, %f152, %f150, %p11;\n"
|
|
" fma.rn.ftz.f32 %f36, %f37, %f154, %f36;\n"
|
|
"$Lt_1_30466:\n"
|
|
"$Lt_1_29954:\n"
|
|
" ld.param.s32 %r58, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r59, 0;\n"
|
|
" setp.le.s32 %p12, %r58, %r59;\n"
|
|
" @%p12 bra $Lt_1_30978;\n"
|
|
" .loc 16 429 0\n"
|
|
" mov.f32 %f155, %f7;\n"
|
|
" mul.ftz.f32 %f156, %f50, %f50;\n"
|
|
" fma.rn.ftz.f32 %f157, %f143, %f156, %f155;\n"
|
|
" mov.f32 %f7, %f157;\n"
|
|
" .loc 16 430 0\n"
|
|
" mov.f32 %f158, %f9;\n"
|
|
" fma.rn.ftz.f32 %f159, %f143, %f52, %f158;\n"
|
|
" mov.f32 %f9, %f159;\n"
|
|
" .loc 16 431 0\n"
|
|
" mov.f32 %f160, %f11;\n"
|
|
" mul.ftz.f32 %f161, %f51, %f51;\n"
|
|
" fma.rn.ftz.f32 %f162, %f143, %f161, %f160;\n"
|
|
" mov.f32 %f11, %f162;\n"
|
|
" .loc 16 432 0\n"
|
|
" mov.f32 %f163, %f13;\n"
|
|
" mul.ftz.f32 %f164, %f49, %f50;\n"
|
|
" fma.rn.ftz.f32 %f165, %f143, %f164, %f163;\n"
|
|
" mov.f32 %f13, %f165;\n"
|
|
" .loc 16 433 0\n"
|
|
" mov.f32 %f166, %f15;\n"
|
|
" mul.ftz.f32 %f167, %f50, %f51;\n"
|
|
" fma.rn.ftz.f32 %f168, %f143, %f167, %f166;\n"
|
|
" mov.f32 %f15, %f168;\n"
|
|
" .loc 16 434 0\n"
|
|
" mul.ftz.f32 %f169, %f49, %f51;\n"
|
|
" fma.rn.ftz.f32 %f16, %f143, %f169, %f16;\n"
|
|
" mov.f32 %f17, %f16;\n"
|
|
"$Lt_1_30978:\n"
|
|
"$Lt_1_27906:\n"
|
|
" .loc 16 357 0\n"
|
|
" mul.lo.u64 %rd44, %rd34, 4;\n"
|
|
" add.u64 %rd26, %rd26, %rd44;\n"
|
|
" setp.lt.u64 %p13, %rd26, %rd23;\n"
|
|
" @%p13 bra $Lt_1_27650;\n"
|
|
" bra.uni $Lt_1_26114;\n"
|
|
"$Lt_1_37378:\n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
" mov.f32 %f33, 0f00000000; \n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
" bra.uni $Lt_1_26114;\n"
|
|
"$Lt_1_26370:\n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
" mov.f32 %f33, 0f00000000; \n"
|
|
" mov.f32 %f34, 0f00000000; \n"
|
|
" mov.f32 %f35, 0f00000000; \n"
|
|
" mov.f32 %f36, 0f00000000; \n"
|
|
"$Lt_1_26114:\n"
|
|
" mov.u32 %r60, 1;\n"
|
|
" setp.le.s32 %p14, %r5, %r60;\n"
|
|
" @%p14 bra $Lt_1_33794;\n"
|
|
" .loc 16 445 0\n"
|
|
" mov.u64 %rd45, __cuda___cuda_local_var_32826_35_non_const_red_acc4392;\n"
|
|
" mul.lo.u64 %rd46, %rd1, 4;\n"
|
|
" add.u64 %rd47, %rd45, %rd46;\n"
|
|
" mov.f32 %f170, %f34;\n"
|
|
" st.shared.f32 [%rd47+0], %f170;\n"
|
|
" .loc 16 446 0\n"
|
|
" mov.f32 %f171, %f33;\n"
|
|
" st.shared.f32 [%rd47+512], %f171;\n"
|
|
" .loc 16 447 0\n"
|
|
" mov.f32 %f172, %f32;\n"
|
|
" st.shared.f32 [%rd47+1024], %f172;\n"
|
|
" .loc 16 448 0\n"
|
|
" mov.f32 %f173, %f36;\n"
|
|
" st.shared.f32 [%rd47+1536], %f173;\n"
|
|
" .loc 16 449 0\n"
|
|
" mov.f32 %f174, %f35;\n"
|
|
" st.shared.f32 [%rd47+2048], %f174;\n"
|
|
" .loc 16 451 0\n"
|
|
" shr.s32 %r61, %r5, 31;\n"
|
|
" mov.s32 %r62, 1;\n"
|
|
" and.b32 %r63, %r61, %r62;\n"
|
|
" add.s32 %r64, %r63, %r5;\n"
|
|
" shr.s32 %r65, %r64, 1;\n"
|
|
" mov.s32 %r66, %r65;\n"
|
|
" mov.u32 %r67, 0;\n"
|
|
" setp.ne.u32 %p15, %r65, %r67;\n"
|
|
" @!%p15 bra $Lt_1_32258;\n"
|
|
"$Lt_1_32770:\n"
|
|
" setp.ge.u32 %p16, %r9, %r66;\n"
|
|
" @%p16 bra $Lt_1_33026;\n"
|
|
" .loc 16 454 0\n"
|
|
" add.u32 %r68, %r1, %r66;\n"
|
|
" cvt.u64.u32 %rd48, %r68;\n"
|
|
" mul.wide.u32 %rd49, %r68, 4;\n"
|
|
" add.u64 %rd50, %rd45, %rd49;\n"
|
|
" ld.shared.f32 %f175, [%rd50+0];\n"
|
|
" add.ftz.f32 %f170, %f175, %f170;\n"
|
|
" st.shared.f32 [%rd47+0], %f170;\n"
|
|
" ld.shared.f32 %f176, [%rd50+512];\n"
|
|
" add.ftz.f32 %f171, %f176, %f171;\n"
|
|
" st.shared.f32 [%rd47+512], %f171;\n"
|
|
" ld.shared.f32 %f177, [%rd50+1024];\n"
|
|
" add.ftz.f32 %f172, %f177, %f172;\n"
|
|
" st.shared.f32 [%rd47+1024], %f172;\n"
|
|
" ld.shared.f32 %f178, [%rd50+1536];\n"
|
|
" add.ftz.f32 %f173, %f178, %f173;\n"
|
|
" st.shared.f32 [%rd47+1536], %f173;\n"
|
|
" ld.shared.f32 %f179, [%rd50+2048];\n"
|
|
" add.ftz.f32 %f174, %f179, %f174;\n"
|
|
" st.shared.f32 [%rd47+2048], %f174;\n"
|
|
"$Lt_1_33026:\n"
|
|
" .loc 16 451 0\n"
|
|
" shr.u32 %r66, %r66, 1;\n"
|
|
" mov.u32 %r69, 0;\n"
|
|
" setp.ne.u32 %p17, %r66, %r69;\n"
|
|
" @%p17 bra $Lt_1_32770;\n"
|
|
"$Lt_1_32258:\n"
|
|
" .loc 16 458 0\n"
|
|
" mov.f32 %f34, %f170;\n"
|
|
" .loc 16 459 0\n"
|
|
" mov.f32 %f33, %f171;\n"
|
|
" .loc 16 460 0\n"
|
|
" mov.f32 %f32, %f172;\n"
|
|
" .loc 16 461 0\n"
|
|
" mov.f32 %f36, %f173;\n"
|
|
" .loc 16 462 0\n"
|
|
" mov.f32 %f35, %f174;\n"
|
|
" ld.param.s32 %r70, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r71, 0;\n"
|
|
" setp.le.s32 %p18, %r70, %r71;\n"
|
|
" @%p18 bra $Lt_1_33794;\n"
|
|
" .loc 16 466 0\n"
|
|
" mov.f32 %f170, %f7;\n"
|
|
" st.shared.f32 [%rd47+0], %f170;\n"
|
|
" mov.f32 %f171, %f9;\n"
|
|
" st.shared.f32 [%rd47+512], %f171;\n"
|
|
" mov.f32 %f172, %f11;\n"
|
|
" st.shared.f32 [%rd47+1024], %f172;\n"
|
|
" mov.f32 %f173, %f13;\n"
|
|
" st.shared.f32 [%rd47+1536], %f173;\n"
|
|
" mov.f32 %f174, %f15;\n"
|
|
" st.shared.f32 [%rd47+2048], %f174;\n"
|
|
" mov.f32 %f180, %f17;\n"
|
|
" st.shared.f32 [%rd47+2560], %f180;\n"
|
|
" .loc 16 468 0\n"
|
|
" mov.s32 %r72, %r65;\n"
|
|
" @!%p15 bra $Lt_1_34306;\n"
|
|
"$Lt_1_34818:\n"
|
|
" setp.ge.u32 %p19, %r9, %r72;\n"
|
|
" @%p19 bra $Lt_1_35074;\n"
|
|
" .loc 16 471 0\n"
|
|
" add.u32 %r73, %r1, %r72;\n"
|
|
" cvt.u64.u32 %rd51, %r73;\n"
|
|
" mul.wide.u32 %rd52, %r73, 4;\n"
|
|
" add.u64 %rd53, %rd45, %rd52;\n"
|
|
" ld.shared.f32 %f181, [%rd53+0];\n"
|
|
" add.ftz.f32 %f170, %f181, %f170;\n"
|
|
" st.shared.f32 [%rd47+0], %f170;\n"
|
|
" ld.shared.f32 %f182, [%rd53+512];\n"
|
|
" add.ftz.f32 %f171, %f182, %f171;\n"
|
|
" st.shared.f32 [%rd47+512], %f171;\n"
|
|
" ld.shared.f32 %f183, [%rd53+1024];\n"
|
|
" add.ftz.f32 %f172, %f183, %f172;\n"
|
|
" st.shared.f32 [%rd47+1024], %f172;\n"
|
|
" ld.shared.f32 %f184, [%rd53+1536];\n"
|
|
" add.ftz.f32 %f173, %f184, %f173;\n"
|
|
" st.shared.f32 [%rd47+1536], %f173;\n"
|
|
" ld.shared.f32 %f185, [%rd53+2048];\n"
|
|
" add.ftz.f32 %f174, %f185, %f174;\n"
|
|
" st.shared.f32 [%rd47+2048], %f174;\n"
|
|
" ld.shared.f32 %f186, [%rd53+2560];\n"
|
|
" add.ftz.f32 %f180, %f186, %f180;\n"
|
|
" st.shared.f32 [%rd47+2560], %f180;\n"
|
|
"$Lt_1_35074:\n"
|
|
" .loc 16 468 0\n"
|
|
" shr.u32 %r72, %r72, 1;\n"
|
|
" mov.u32 %r74, 0;\n"
|
|
" setp.ne.u32 %p20, %r72, %r74;\n"
|
|
" @%p20 bra $Lt_1_34818;\n"
|
|
"$Lt_1_34306:\n"
|
|
" .loc 16 476 0\n"
|
|
" mov.f32 %f7, %f170;\n"
|
|
" mov.f32 %f9, %f171;\n"
|
|
" mov.f32 %f11, %f172;\n"
|
|
" mov.f32 %f13, %f173;\n"
|
|
" mov.f32 %f15, %f174;\n"
|
|
" mov.f32 %f17, %f180;\n"
|
|
"$Lt_1_33794:\n"
|
|
"$Lt_1_31746:\n"
|
|
" selp.s32 %r75, 1, 0, %p3;\n"
|
|
" mov.s32 %r76, 0;\n"
|
|
" set.eq.u32.s32 %r77, %r9, %r76;\n"
|
|
" neg.s32 %r78, %r77;\n"
|
|
" and.b32 %r79, %r75, %r78;\n"
|
|
" mov.u32 %r80, 0;\n"
|
|
" setp.eq.s32 %p21, %r79, %r80;\n"
|
|
" @%p21 bra $Lt_1_35842;\n"
|
|
" .loc 16 482 0\n"
|
|
" cvt.s64.s32 %rd54, %r12;\n"
|
|
" ld.param.u64 %rd55, [__cudaparm_kernel_pair_fast_engv];\n"
|
|
" mul.wide.s32 %rd56, %r12, 4;\n"
|
|
" add.u64 %rd57, %rd55, %rd56;\n"
|
|
" ld.param.s32 %r81, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r82, 0;\n"
|
|
" setp.le.s32 %p22, %r81, %r82;\n"
|
|
" @%p22 bra $Lt_1_36354;\n"
|
|
" .loc 16 484 0\n"
|
|
" st.global.f32 [%rd57+0], %f36;\n"
|
|
" .loc 16 485 0\n"
|
|
" cvt.s64.s32 %rd58, %r13;\n"
|
|
" mul.wide.s32 %rd59, %r13, 4;\n"
|
|
" add.u64 %rd60, %rd59, %rd57;\n"
|
|
" .loc 16 486 0\n"
|
|
" st.global.f32 [%rd60+0], %f35;\n"
|
|
" .loc 16 487 0\n"
|
|
" add.u64 %rd57, %rd59, %rd60;\n"
|
|
"$Lt_1_36354:\n"
|
|
" ld.param.s32 %r83, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r84, 0;\n"
|
|
" setp.le.s32 %p23, %r83, %r84;\n"
|
|
" @%p23 bra $Lt_1_36866;\n"
|
|
" .loc 16 491 0\n"
|
|
" mov.f32 %f187, %f7;\n"
|
|
" st.global.f32 [%rd57+0], %f187;\n"
|
|
" .loc 16 492 0\n"
|
|
" cvt.s64.s32 %rd61, %r13;\n"
|
|
" mul.wide.s32 %rd62, %r13, 4;\n"
|
|
" add.u64 %rd63, %rd62, %rd57;\n"
|
|
" .loc 16 491 0\n"
|
|
" mov.f32 %f188, %f9;\n"
|
|
" st.global.f32 [%rd63+0], %f188;\n"
|
|
" .loc 16 492 0\n"
|
|
" add.u64 %rd64, %rd62, %rd63;\n"
|
|
" .loc 16 491 0\n"
|
|
" mov.f32 %f189, %f11;\n"
|
|
" st.global.f32 [%rd64+0], %f189;\n"
|
|
" .loc 16 492 0\n"
|
|
" add.u64 %rd65, %rd62, %rd64;\n"
|
|
" .loc 16 491 0\n"
|
|
" mov.f32 %f190, %f13;\n"
|
|
" st.global.f32 [%rd65+0], %f190;\n"
|
|
" .loc 16 492 0\n"
|
|
" add.u64 %rd57, %rd62, %rd65;\n"
|
|
" .loc 16 491 0\n"
|
|
" mov.f32 %f191, %f15;\n"
|
|
" st.global.f32 [%rd57+0], %f191;\n"
|
|
" mov.f32 %f192, %f17;\n"
|
|
" add.u64 %rd66, %rd62, %rd57;\n"
|
|
" st.global.f32 [%rd66+0], %f192;\n"
|
|
"$Lt_1_36866:\n"
|
|
" .loc 16 495 0\n"
|
|
" ld.param.u64 %rd67, [__cudaparm_kernel_pair_fast_ans];\n"
|
|
" mul.lo.u64 %rd68, %rd54, 16;\n"
|
|
" add.u64 %rd69, %rd67, %rd68;\n"
|
|
" mov.f32 %f193, %f194;\n"
|
|
" st.global.v4.f32 [%rd69+0], {%f34,%f33,%f32,%f193};\n"
|
|
"$Lt_1_35842:\n"
|
|
" .loc 16 497 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair_fast:\n"
|
|
" }\n"
|
|
;
|