forked from lijiext/lammps
570 lines
18 KiB
C
570 lines
18 KiB
C
const char * lj_cut_gpu_kernel =
|
|
" .version 1.4\n"
|
|
" .target sm_13\n"
|
|
" .tex .u64 pos_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_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_nall,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_nbor_pitch)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<33>;\n"
|
|
" .reg .u64 %rd<36>;\n"
|
|
" .reg .f32 %f<86>;\n"
|
|
" .reg .pred %p<8>;\n"
|
|
" .shared .align 4 .b8 __cuda_sp_lj84[16];\n"
|
|
" .loc 14 87 0\n"
|
|
"$LBB1_kernel_pair:\n"
|
|
" .loc 14 91 0\n"
|
|
" ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];\n"
|
|
" ld.global.f32 %f1, [%rd1+0];\n"
|
|
" st.shared.f32 [__cuda_sp_lj84+0], %f1;\n"
|
|
" .loc 14 92 0\n"
|
|
" ld.global.f32 %f2, [%rd1+4];\n"
|
|
" st.shared.f32 [__cuda_sp_lj84+4], %f2;\n"
|
|
" .loc 14 93 0\n"
|
|
" ld.global.f32 %f3, [%rd1+8];\n"
|
|
" st.shared.f32 [__cuda_sp_lj84+8], %f3;\n"
|
|
" .loc 14 94 0\n"
|
|
" ld.global.f32 %f4, [%rd1+12];\n"
|
|
" st.shared.f32 [__cuda_sp_lj84+12], %f4;\n"
|
|
" cvt.s32.u16 %r1, %ctaid.x;\n"
|
|
" cvt.s32.u16 %r2, %ntid.x;\n"
|
|
" mul24.lo.s32 %r3, %r1, %r2;\n"
|
|
" cvt.u32.u16 %r4, %tid.x;\n"
|
|
" add.u32 %r5, %r3, %r4;\n"
|
|
" ld.param.s32 %r6, [__cudaparm_kernel_pair_inum];\n"
|
|
" setp.le.s32 %p1, %r6, %r5;\n"
|
|
" @%p1 bra $Lt_0_7938;\n"
|
|
" .loc 14 105 0\n"
|
|
" mov.f32 %f5, 0f00000000; \n"
|
|
" mov.f32 %f6, %f5;\n"
|
|
" mov.f32 %f7, 0f00000000; \n"
|
|
" mov.f32 %f8, %f7;\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"
|
|
" .loc 14 108 0\n"
|
|
" cvt.u64.s32 %rd2, %r5;\n"
|
|
" mul.lo.u64 %rd3, %rd2, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_dev_nbor];\n"
|
|
" add.u64 %rd5, %rd4, %rd3;\n"
|
|
" ld.global.s32 %r7, [%rd5+0];\n"
|
|
" .loc 14 110 0\n"
|
|
" ld.param.s32 %r8, [__cudaparm_kernel_pair_nbor_pitch];\n"
|
|
" cvt.u64.s32 %rd6, %r8;\n"
|
|
" mul.lo.u64 %rd7, %rd6, 4;\n"
|
|
" add.u64 %rd8, %rd5, %rd7;\n"
|
|
" ld.global.s32 %r9, [%rd8+0];\n"
|
|
" .loc 14 111 0\n"
|
|
" add.u64 %rd9, %rd8, %rd7;\n"
|
|
" mov.s64 %rd10, %rd9;\n"
|
|
" mov.s32 %r10, %r7;\n"
|
|
" mov.s32 %r11, 0;\n"
|
|
" mov.s32 %r12, 0;\n"
|
|
" mov.s32 %r13, 0;\n"
|
|
" tex.1d.v4.f32.s32 {%f17,%f18,%f19,%f20},[pos_tex,{%r10,%r11,%r12,%r13}];\n"
|
|
" .loc 14 114 0\n"
|
|
" mov.f32 %f21, %f17;\n"
|
|
" mov.f32 %f22, %f18;\n"
|
|
" mov.f32 %f23, %f19;\n"
|
|
" mov.f32 %f24, %f20;\n"
|
|
" mul24.lo.s32 %r14, %r9, %r8;\n"
|
|
" cvt.s64.s32 %rd11, %r14;\n"
|
|
" mul.lo.u64 %rd12, %rd11, 4;\n"
|
|
" add.u64 %rd13, %rd9, %rd12;\n"
|
|
" ld.param.s32 %r15, [__cudaparm_kernel_pair_vflag];\n"
|
|
" ld.param.s32 %r16, [__cudaparm_kernel_pair_eflag];\n"
|
|
" setp.ge.u64 %p2, %rd9, %rd13;\n"
|
|
" mov.f32 %f25, 0f00000000; \n"
|
|
" mov.f32 %f26, 0f00000000; \n"
|
|
" mov.f32 %f27, 0f00000000; \n"
|
|
" mov.f32 %f28, 0f00000000; \n"
|
|
" @%p2 bra $Lt_0_12034;\n"
|
|
" mov.s32 %r17, 0;\n"
|
|
" setp.gt.s32 %p3, %r16, %r17;\n"
|
|
" mov.s32 %r18, 0;\n"
|
|
" setp.gt.s32 %p4, %r15, %r18;\n"
|
|
" cvt.rzi.s32.f32 %r19, %f24;\n"
|
|
" ld.param.s32 %r20, [__cudaparm_kernel_pair_lj_types];\n"
|
|
" mul.lo.s32 %r21, %r20, %r19;\n"
|
|
" ld.param.u64 %rd14, [__cudaparm_kernel_pair_lj1];\n"
|
|
" mov.u64 %rd15, __cuda_sp_lj84;\n"
|
|
"$Lt_0_8962:\n"
|
|
" .loc 14 120 0\n"
|
|
" ld.global.s32 %r22, [%rd10+0];\n"
|
|
" .loc 14 121 0\n"
|
|
" shr.s32 %r23, %r22, 30;\n"
|
|
" cvt.s64.s32 %rd16, %r23;\n"
|
|
" and.b64 %rd17, %rd16, 3;\n"
|
|
" mul.lo.u64 %rd18, %rd17, 4;\n"
|
|
" add.u64 %rd19, %rd15, %rd18;\n"
|
|
" ld.shared.f32 %f29, [%rd19+0];\n"
|
|
" and.b32 %r24, %r22, 1073741823;\n"
|
|
" mov.s32 %r25, 0;\n"
|
|
" mov.s32 %r26, 0;\n"
|
|
" mov.s32 %r27, 0;\n"
|
|
" tex.1d.v4.f32.s32 {%f30,%f31,%f32,%f33},[pos_tex,{%r24,%r25,%r26,%r27}];\n"
|
|
" .loc 14 124 0\n"
|
|
" mov.f32 %f34, %f30;\n"
|
|
" mov.f32 %f35, %f31;\n"
|
|
" mov.f32 %f36, %f32;\n"
|
|
" mov.f32 %f37, %f33;\n"
|
|
" cvt.rzi.s32.f32 %r28, %f37;\n"
|
|
" sub.f32 %f38, %f22, %f35;\n"
|
|
" sub.f32 %f39, %f21, %f34;\n"
|
|
" sub.f32 %f40, %f23, %f36;\n"
|
|
" mul.f32 %f41, %f38, %f38;\n"
|
|
" mad.f32 %f42, %f39, %f39, %f41;\n"
|
|
" mad.f32 %f43, %f40, %f40, %f42;\n"
|
|
" add.s32 %r29, %r28, %r21;\n"
|
|
" cvt.u64.s32 %rd20, %r29;\n"
|
|
" mul.lo.u64 %rd21, %rd20, 16;\n"
|
|
" add.u64 %rd22, %rd21, %rd14;\n"
|
|
" ld.global.f32 %f44, [%rd22+8];\n"
|
|
" setp.gt.f32 %p5, %f44, %f43;\n"
|
|
" @!%p5 bra $Lt_0_10242;\n"
|
|
" .loc 14 138 0\n"
|
|
" rcp.approx.f32 %f45, %f43;\n"
|
|
" mul.f32 %f46, %f45, %f45;\n"
|
|
" mul.f32 %f47, %f45, %f46;\n"
|
|
" mul.f32 %f48, %f45, %f47;\n"
|
|
" ld.global.v2.f32 {%f49,%f50}, [%rd22+0];\n"
|
|
" mul.f32 %f51, %f49, %f47;\n"
|
|
" sub.f32 %f52, %f51, %f50;\n"
|
|
" mul.f32 %f53, %f48, %f52;\n"
|
|
" .loc 14 121 0\n"
|
|
" ld.shared.f32 %f29, [%rd19+0];\n"
|
|
" .loc 14 138 0\n"
|
|
" mul.f32 %f54, %f29, %f53;\n"
|
|
" .loc 14 140 0\n"
|
|
" mad.f32 %f27, %f39, %f54, %f27;\n"
|
|
" .loc 14 141 0\n"
|
|
" mad.f32 %f26, %f38, %f54, %f26;\n"
|
|
" .loc 14 142 0\n"
|
|
" mad.f32 %f25, %f40, %f54, %f25;\n"
|
|
" @!%p3 bra $Lt_0_9730;\n"
|
|
" .loc 14 146 0\n"
|
|
" ld.param.u64 %rd23, [__cudaparm_kernel_pair_lj3];\n"
|
|
" add.u64 %rd24, %rd23, %rd21;\n"
|
|
" ld.global.v4.f32 {%f55,%f56,%f57,_}, [%rd24+0];\n"
|
|
" mul.f32 %f58, %f55, %f47;\n"
|
|
" sub.f32 %f59, %f58, %f56;\n"
|
|
" mul.f32 %f60, %f47, %f59;\n"
|
|
" sub.f32 %f61, %f60, %f57;\n"
|
|
" .loc 14 121 0\n"
|
|
" ld.shared.f32 %f29, [%rd19+0];\n"
|
|
" .loc 14 146 0\n"
|
|
" mad.f32 %f28, %f29, %f61, %f28;\n"
|
|
"$Lt_0_9730:\n"
|
|
" @!%p4 bra $Lt_0_10242;\n"
|
|
" .loc 14 149 0\n"
|
|
" mov.f32 %f62, %f6;\n"
|
|
" mul.f32 %f63, %f39, %f39;\n"
|
|
" mad.f32 %f64, %f54, %f63, %f62;\n"
|
|
" mov.f32 %f6, %f64;\n"
|
|
" .loc 14 150 0\n"
|
|
" mov.f32 %f65, %f8;\n"
|
|
" mad.f32 %f66, %f54, %f41, %f65;\n"
|
|
" mov.f32 %f8, %f66;\n"
|
|
" .loc 14 151 0\n"
|
|
" mov.f32 %f67, %f10;\n"
|
|
" mul.f32 %f68, %f40, %f40;\n"
|
|
" mad.f32 %f69, %f54, %f68, %f67;\n"
|
|
" mov.f32 %f10, %f69;\n"
|
|
" .loc 14 152 0\n"
|
|
" mov.f32 %f70, %f12;\n"
|
|
" mul.f32 %f71, %f38, %f39;\n"
|
|
" mad.f32 %f72, %f54, %f71, %f70;\n"
|
|
" mov.f32 %f12, %f72;\n"
|
|
" .loc 14 153 0\n"
|
|
" mov.f32 %f73, %f14;\n"
|
|
" mul.f32 %f74, %f39, %f40;\n"
|
|
" mad.f32 %f75, %f54, %f74, %f73;\n"
|
|
" mov.f32 %f14, %f75;\n"
|
|
" .loc 14 154 0\n"
|
|
" mul.f32 %f76, %f38, %f40;\n"
|
|
" mad.f32 %f15, %f54, %f76, %f15;\n"
|
|
" mov.f32 %f77, %f15;\n"
|
|
"$Lt_0_10242:\n"
|
|
"$Lt_0_9218:\n"
|
|
" .loc 14 118 0\n"
|
|
" add.u64 %rd10, %rd7, %rd10;\n"
|
|
" setp.gt.u64 %p6, %rd13, %rd10;\n"
|
|
" @%p6 bra $Lt_0_8962;\n"
|
|
" bra.uni $Lt_0_8450;\n"
|
|
"$Lt_0_12034:\n"
|
|
" mov.s32 %r30, 0;\n"
|
|
" setp.gt.s32 %p3, %r16, %r30;\n"
|
|
" mov.s32 %r31, 0;\n"
|
|
" setp.gt.s32 %p4, %r15, %r31;\n"
|
|
"$Lt_0_8450:\n"
|
|
" .loc 14 161 0\n"
|
|
" ld.param.u64 %rd25, [__cudaparm_kernel_pair_engv];\n"
|
|
" add.u64 %rd26, %rd25, %rd3;\n"
|
|
" @!%p3 bra $Lt_0_11010;\n"
|
|
" .loc 14 163 0\n"
|
|
" st.global.f32 [%rd26+0], %f28;\n"
|
|
" .loc 14 164 0\n"
|
|
" cvt.u64.s32 %rd27, %r6;\n"
|
|
" mul.lo.u64 %rd28, %rd27, 4;\n"
|
|
" add.u64 %rd26, %rd26, %rd28;\n"
|
|
"$Lt_0_11010:\n"
|
|
" @!%p4 bra $Lt_0_11522;\n"
|
|
" .loc 14 168 0\n"
|
|
" mov.f32 %f78, %f6;\n"
|
|
" st.global.f32 [%rd26+0], %f78;\n"
|
|
" .loc 14 169 0\n"
|
|
" cvt.u64.s32 %rd29, %r6;\n"
|
|
" mul.lo.u64 %rd30, %rd29, 4;\n"
|
|
" add.u64 %rd26, %rd30, %rd26;\n"
|
|
" .loc 14 168 0\n"
|
|
" mov.f32 %f79, %f8;\n"
|
|
" st.global.f32 [%rd26+0], %f79;\n"
|
|
" .loc 14 169 0\n"
|
|
" add.u64 %rd26, %rd30, %rd26;\n"
|
|
" .loc 14 168 0\n"
|
|
" mov.f32 %f80, %f10;\n"
|
|
" st.global.f32 [%rd26+0], %f80;\n"
|
|
" .loc 14 169 0\n"
|
|
" add.u64 %rd26, %rd30, %rd26;\n"
|
|
" .loc 14 168 0\n"
|
|
" mov.f32 %f81, %f12;\n"
|
|
" st.global.f32 [%rd26+0], %f81;\n"
|
|
" .loc 14 169 0\n"
|
|
" add.u64 %rd26, %rd30, %rd26;\n"
|
|
" .loc 14 168 0\n"
|
|
" mov.f32 %f82, %f14;\n"
|
|
" st.global.f32 [%rd26+0], %f82;\n"
|
|
" add.u64 %rd31, %rd30, %rd26;\n"
|
|
" st.global.f32 [%rd31+0], %f15;\n"
|
|
"$Lt_0_11522:\n"
|
|
" .loc 14 172 0\n"
|
|
" ld.param.u64 %rd32, [__cudaparm_kernel_pair_ans];\n"
|
|
" mul.lo.u64 %rd33, %rd2, 16;\n"
|
|
" add.u64 %rd34, %rd32, %rd33;\n"
|
|
" mov.f32 %f83, %f84;\n"
|
|
" st.global.v4.f32 [%rd34+0], {%f27,%f26,%f25,%f83};\n"
|
|
"$Lt_0_7938:\n"
|
|
" .loc 14 174 0\n"
|
|
" exit;\n"
|
|
"$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_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_nall,\n"
|
|
" .param .s32 __cudaparm_kernel_pair_fast_nbor_pitch)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<36>;\n"
|
|
" .reg .u64 %rd<48>;\n"
|
|
" .reg .f32 %f<93>;\n"
|
|
" .reg .pred %p<11>;\n"
|
|
" .shared .align 4 .b8 __cuda_sp_lj180[16];\n"
|
|
" .shared .align 16 .b8 __cuda_lj1208[1024];\n"
|
|
" .shared .align 16 .b8 __cuda_lj31232[1024];\n"
|
|
" .loc 14 181 0\n"
|
|
"$LBB1_kernel_pair_fast:\n"
|
|
" cvt.s32.u16 %r1, %tid.x;\n"
|
|
" mov.u32 %r2, 3;\n"
|
|
" setp.gt.s32 %p1, %r1, %r2;\n"
|
|
" @%p1 bra $Lt_1_10242;\n"
|
|
" .loc 14 188 0\n"
|
|
" mov.u64 %rd1, __cuda_sp_lj180;\n"
|
|
" cvt.u64.s32 %rd2, %r1;\n"
|
|
" mul.lo.u64 %rd3, %rd2, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_fast_sp_lj_in];\n"
|
|
" add.u64 %rd5, %rd4, %rd3;\n"
|
|
" ld.global.f32 %f1, [%rd5+0];\n"
|
|
" add.u64 %rd6, %rd3, %rd1;\n"
|
|
" st.shared.f32 [%rd6+0], %f1;\n"
|
|
"$Lt_1_10242:\n"
|
|
" mov.u64 %rd1, __cuda_sp_lj180;\n"
|
|
" mov.u32 %r3, 63;\n"
|
|
" setp.gt.s32 %p2, %r1, %r3;\n"
|
|
" @%p2 bra $Lt_1_10754;\n"
|
|
" .loc 14 190 0\n"
|
|
" mov.u64 %rd7, __cuda_lj1208;\n"
|
|
" cvt.u64.s32 %rd8, %r1;\n"
|
|
" mul.lo.u64 %rd9, %rd8, 16;\n"
|
|
" ld.param.u64 %rd10, [__cudaparm_kernel_pair_fast_lj1_in];\n"
|
|
" add.u64 %rd11, %rd10, %rd9;\n"
|
|
" add.u64 %rd12, %rd9, %rd7;\n"
|
|
" ld.global.v4.f32 {%f2,%f3,%f4,%f5}, [%rd11+0];\n"
|
|
" st.shared.f32 [%rd12+0], %f2;\n"
|
|
" st.shared.f32 [%rd12+4], %f3;\n"
|
|
" st.shared.f32 [%rd12+8], %f4;\n"
|
|
" st.shared.f32 [%rd12+12], %f5;\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 $Lt_1_11266;\n"
|
|
" .loc 14 192 0\n"
|
|
" mov.u64 %rd13, __cuda_lj31232;\n"
|
|
" ld.param.u64 %rd14, [__cudaparm_kernel_pair_fast_lj3_in];\n"
|
|
" add.u64 %rd15, %rd14, %rd9;\n"
|
|
" add.u64 %rd16, %rd9, %rd13;\n"
|
|
" ld.global.v4.f32 {%f6,%f7,%f8,%f9}, [%rd15+0];\n"
|
|
" st.shared.f32 [%rd16+0], %f6;\n"
|
|
" st.shared.f32 [%rd16+4], %f7;\n"
|
|
" st.shared.f32 [%rd16+8], %f8;\n"
|
|
" st.shared.f32 [%rd16+12], %f9;\n"
|
|
"$Lt_1_11266:\n"
|
|
" mov.u64 %rd13, __cuda_lj31232;\n"
|
|
"$Lt_1_10754:\n"
|
|
" mov.u64 %rd13, __cuda_lj31232;\n"
|
|
" mov.u64 %rd7, __cuda_lj1208;\n"
|
|
" .loc 14 195 0\n"
|
|
" bar.sync 0;\n"
|
|
" cvt.s32.u16 %r6, %ctaid.x;\n"
|
|
" cvt.s32.u16 %r7, %ntid.x;\n"
|
|
" mul24.lo.s32 %r8, %r6, %r7;\n"
|
|
" add.s32 %r9, %r8, %r1;\n"
|
|
" ld.param.s32 %r10, [__cudaparm_kernel_pair_fast_inum];\n"
|
|
" setp.ge.s32 %p4, %r9, %r10;\n"
|
|
" @%p4 bra $Lt_1_11778;\n"
|
|
" .loc 14 206 0\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"
|
|
" mov.f32 %f18, 0f00000000; \n"
|
|
" mov.f32 %f19, %f18;\n"
|
|
" mov.f32 %f20, 0f00000000; \n"
|
|
" mov.f32 %f21, %f20;\n"
|
|
" .loc 14 209 0\n"
|
|
" cvt.u64.s32 %rd17, %r9;\n"
|
|
" mul.lo.u64 %rd18, %rd17, 4;\n"
|
|
" ld.param.u64 %rd19, [__cudaparm_kernel_pair_fast_dev_nbor];\n"
|
|
" add.u64 %rd20, %rd19, %rd18;\n"
|
|
" ld.global.s32 %r11, [%rd20+0];\n"
|
|
" .loc 14 211 0\n"
|
|
" ld.param.s32 %r12, [__cudaparm_kernel_pair_fast_nbor_pitch];\n"
|
|
" cvt.u64.s32 %rd21, %r12;\n"
|
|
" mul.lo.u64 %rd22, %rd21, 4;\n"
|
|
" add.u64 %rd23, %rd20, %rd22;\n"
|
|
" ld.global.s32 %r13, [%rd23+0];\n"
|
|
" .loc 14 212 0\n"
|
|
" add.u64 %rd24, %rd23, %rd22;\n"
|
|
" mov.s64 %rd25, %rd24;\n"
|
|
" mov.s32 %r14, %r11;\n"
|
|
" mov.s32 %r15, 0;\n"
|
|
" mov.s32 %r16, 0;\n"
|
|
" mov.s32 %r17, 0;\n"
|
|
" tex.1d.v4.f32.s32 {%f22,%f23,%f24,%f25},[pos_tex,{%r14,%r15,%r16,%r17}];\n"
|
|
" .loc 14 215 0\n"
|
|
" mov.f32 %f26, %f22;\n"
|
|
" mov.f32 %f27, %f23;\n"
|
|
" mov.f32 %f28, %f24;\n"
|
|
" mov.f32 %f29, %f25;\n"
|
|
" mul24.lo.s32 %r18, %r13, %r12;\n"
|
|
" cvt.s64.s32 %rd26, %r18;\n"
|
|
" mul.lo.u64 %rd27, %rd26, 4;\n"
|
|
" add.u64 %rd28, %rd24, %rd27;\n"
|
|
" ld.param.s32 %r19, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" ld.param.s32 %r20, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" setp.ge.u64 %p5, %rd24, %rd28;\n"
|
|
" mov.f32 %f30, 0f00000000; \n"
|
|
" mov.f32 %f31, 0f00000000; \n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
" mov.f32 %f33, 0f00000000; \n"
|
|
" @%p5 bra $Lt_1_15874;\n"
|
|
" mov.s32 %r21, 0;\n"
|
|
" setp.gt.s32 %p6, %r20, %r21;\n"
|
|
" mov.s32 %r22, 0;\n"
|
|
" setp.gt.s32 %p7, %r19, %r22;\n"
|
|
" cvt.rzi.s32.f32 %r23, %f29;\n"
|
|
" mov.s32 %r24, 8;\n"
|
|
" mul24.lo.s32 %r25, %r24, %r23;\n"
|
|
" cvt.rn.f32.s32 %f34, %r25;\n"
|
|
"$Lt_1_12802:\n"
|
|
" .loc 14 222 0\n"
|
|
" ld.global.s32 %r26, [%rd25+0];\n"
|
|
" .loc 14 223 0\n"
|
|
" shr.s32 %r27, %r26, 30;\n"
|
|
" cvt.s64.s32 %rd29, %r27;\n"
|
|
" and.b64 %rd30, %rd29, 3;\n"
|
|
" mul.lo.u64 %rd31, %rd30, 4;\n"
|
|
" add.u64 %rd32, %rd1, %rd31;\n"
|
|
" ld.shared.f32 %f35, [%rd32+0];\n"
|
|
" and.b32 %r28, %r26, 1073741823;\n"
|
|
" mov.s32 %r29, 0;\n"
|
|
" mov.s32 %r30, 0;\n"
|
|
" mov.s32 %r31, 0;\n"
|
|
" tex.1d.v4.f32.s32 {%f36,%f37,%f38,%f39},[pos_tex,{%r28,%r29,%r30,%r31}];\n"
|
|
" .loc 14 226 0\n"
|
|
" mov.f32 %f40, %f36;\n"
|
|
" mov.f32 %f41, %f37;\n"
|
|
" mov.f32 %f42, %f38;\n"
|
|
" mov.f32 %f43, %f39;\n"
|
|
" sub.f32 %f44, %f27, %f41;\n"
|
|
" sub.f32 %f45, %f26, %f40;\n"
|
|
" sub.f32 %f46, %f28, %f42;\n"
|
|
" mul.f32 %f47, %f44, %f44;\n"
|
|
" mad.f32 %f48, %f45, %f45, %f47;\n"
|
|
" mad.f32 %f49, %f46, %f46, %f48;\n"
|
|
" add.f32 %f50, %f34, %f43;\n"
|
|
" cvt.rzi.s32.f32 %r32, %f50;\n"
|
|
" cvt.u64.s32 %rd33, %r32;\n"
|
|
" mul.lo.u64 %rd34, %rd33, 16;\n"
|
|
" add.u64 %rd35, %rd34, %rd7;\n"
|
|
" ld.shared.f32 %f51, [%rd35+8];\n"
|
|
" setp.gt.f32 %p8, %f51, %f49;\n"
|
|
" @!%p8 bra $Lt_1_14082;\n"
|
|
" .loc 14 238 0\n"
|
|
" rcp.approx.f32 %f52, %f49;\n"
|
|
" mul.f32 %f53, %f52, %f52;\n"
|
|
" mul.f32 %f54, %f52, %f53;\n"
|
|
" .loc 14 223 0\n"
|
|
" ld.shared.f32 %f35, [%rd32+0];\n"
|
|
" .loc 14 238 0\n"
|
|
" mul.f32 %f55, %f52, %f35;\n"
|
|
" mul.f32 %f56, %f54, %f55;\n"
|
|
" ld.shared.f32 %f57, [%rd35+4];\n"
|
|
" ld.shared.f32 %f58, [%rd35+0];\n"
|
|
" mul.f32 %f59, %f58, %f54;\n"
|
|
" sub.f32 %f60, %f59, %f57;\n"
|
|
" mul.f32 %f61, %f56, %f60;\n"
|
|
" .loc 14 240 0\n"
|
|
" mad.f32 %f32, %f45, %f61, %f32;\n"
|
|
" .loc 14 241 0\n"
|
|
" mad.f32 %f31, %f44, %f61, %f31;\n"
|
|
" .loc 14 242 0\n"
|
|
" mad.f32 %f30, %f46, %f61, %f30;\n"
|
|
" @!%p6 bra $Lt_1_13570;\n"
|
|
" .loc 14 245 0\n"
|
|
" add.u64 %rd36, %rd34, %rd13;\n"
|
|
" ld.shared.f32 %f62, [%rd36+4];\n"
|
|
" ld.shared.f32 %f63, [%rd36+0];\n"
|
|
" mul.f32 %f64, %f63, %f54;\n"
|
|
" sub.f32 %f65, %f64, %f62;\n"
|
|
" mul.f32 %f66, %f54, %f65;\n"
|
|
" .loc 14 246 0\n"
|
|
" ld.shared.f32 %f67, [%rd36+8];\n"
|
|
" sub.f32 %f68, %f66, %f67;\n"
|
|
" .loc 14 223 0\n"
|
|
" ld.shared.f32 %f35, [%rd32+0];\n"
|
|
" .loc 14 246 0\n"
|
|
" mad.f32 %f33, %f35, %f68, %f33;\n"
|
|
"$Lt_1_13570:\n"
|
|
" @!%p7 bra $Lt_1_14082;\n"
|
|
" .loc 14 249 0\n"
|
|
" mov.f32 %f69, %f11;\n"
|
|
" mul.f32 %f70, %f45, %f45;\n"
|
|
" mad.f32 %f71, %f61, %f70, %f69;\n"
|
|
" mov.f32 %f11, %f71;\n"
|
|
" .loc 14 250 0\n"
|
|
" mov.f32 %f72, %f13;\n"
|
|
" mad.f32 %f73, %f61, %f47, %f72;\n"
|
|
" mov.f32 %f13, %f73;\n"
|
|
" .loc 14 251 0\n"
|
|
" mov.f32 %f74, %f15;\n"
|
|
" mul.f32 %f75, %f46, %f46;\n"
|
|
" mad.f32 %f76, %f61, %f75, %f74;\n"
|
|
" mov.f32 %f15, %f76;\n"
|
|
" .loc 14 252 0\n"
|
|
" mov.f32 %f77, %f17;\n"
|
|
" mul.f32 %f78, %f44, %f45;\n"
|
|
" mad.f32 %f79, %f61, %f78, %f77;\n"
|
|
" mov.f32 %f17, %f79;\n"
|
|
" .loc 14 253 0\n"
|
|
" mov.f32 %f80, %f19;\n"
|
|
" mul.f32 %f81, %f45, %f46;\n"
|
|
" mad.f32 %f82, %f61, %f81, %f80;\n"
|
|
" mov.f32 %f19, %f82;\n"
|
|
" .loc 14 254 0\n"
|
|
" mul.f32 %f83, %f44, %f46;\n"
|
|
" mad.f32 %f20, %f61, %f83, %f20;\n"
|
|
" mov.f32 %f84, %f20;\n"
|
|
"$Lt_1_14082:\n"
|
|
"$Lt_1_13058:\n"
|
|
" .loc 14 220 0\n"
|
|
" add.u64 %rd25, %rd22, %rd25;\n"
|
|
" setp.gt.u64 %p9, %rd28, %rd25;\n"
|
|
" @%p9 bra $Lt_1_12802;\n"
|
|
" bra.uni $Lt_1_12290;\n"
|
|
"$Lt_1_15874:\n"
|
|
" mov.s32 %r33, 0;\n"
|
|
" setp.gt.s32 %p6, %r20, %r33;\n"
|
|
" mov.s32 %r34, 0;\n"
|
|
" setp.gt.s32 %p7, %r19, %r34;\n"
|
|
"$Lt_1_12290:\n"
|
|
" .loc 14 261 0\n"
|
|
" ld.param.u64 %rd37, [__cudaparm_kernel_pair_fast_engv];\n"
|
|
" add.u64 %rd38, %rd37, %rd18;\n"
|
|
" @!%p6 bra $Lt_1_14850;\n"
|
|
" .loc 14 263 0\n"
|
|
" st.global.f32 [%rd38+0], %f33;\n"
|
|
" .loc 14 264 0\n"
|
|
" cvt.u64.s32 %rd39, %r10;\n"
|
|
" mul.lo.u64 %rd40, %rd39, 4;\n"
|
|
" add.u64 %rd38, %rd38, %rd40;\n"
|
|
"$Lt_1_14850:\n"
|
|
" @!%p7 bra $Lt_1_15362;\n"
|
|
" .loc 14 268 0\n"
|
|
" mov.f32 %f85, %f11;\n"
|
|
" st.global.f32 [%rd38+0], %f85;\n"
|
|
" .loc 14 269 0\n"
|
|
" cvt.u64.s32 %rd41, %r10;\n"
|
|
" mul.lo.u64 %rd42, %rd41, 4;\n"
|
|
" add.u64 %rd38, %rd42, %rd38;\n"
|
|
" .loc 14 268 0\n"
|
|
" mov.f32 %f86, %f13;\n"
|
|
" st.global.f32 [%rd38+0], %f86;\n"
|
|
" .loc 14 269 0\n"
|
|
" add.u64 %rd38, %rd42, %rd38;\n"
|
|
" .loc 14 268 0\n"
|
|
" mov.f32 %f87, %f15;\n"
|
|
" st.global.f32 [%rd38+0], %f87;\n"
|
|
" .loc 14 269 0\n"
|
|
" add.u64 %rd38, %rd42, %rd38;\n"
|
|
" .loc 14 268 0\n"
|
|
" mov.f32 %f88, %f17;\n"
|
|
" st.global.f32 [%rd38+0], %f88;\n"
|
|
" .loc 14 269 0\n"
|
|
" add.u64 %rd38, %rd42, %rd38;\n"
|
|
" .loc 14 268 0\n"
|
|
" mov.f32 %f89, %f19;\n"
|
|
" st.global.f32 [%rd38+0], %f89;\n"
|
|
" add.u64 %rd43, %rd42, %rd38;\n"
|
|
" st.global.f32 [%rd43+0], %f20;\n"
|
|
"$Lt_1_15362:\n"
|
|
" .loc 14 272 0\n"
|
|
" ld.param.u64 %rd44, [__cudaparm_kernel_pair_fast_ans];\n"
|
|
" mul.lo.u64 %rd45, %rd17, 16;\n"
|
|
" add.u64 %rd46, %rd44, %rd45;\n"
|
|
" mov.f32 %f90, %f91;\n"
|
|
" st.global.v4.f32 [%rd46+0], {%f32,%f31,%f30,%f90};\n"
|
|
"$Lt_1_11778:\n"
|
|
" .loc 14 274 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair_fast:\n"
|
|
" }\n"
|
|
;
|