948 lines
30 KiB
C
948 lines
30 KiB
C
const char * morse_gpu_kernel =
|
|
" .version 2.3\n"
|
|
" .target sm_20\n"
|
|
" .address_size 64\n"
|
|
" .global .texref pos_tex;\n"
|
|
" .entry kernel_pair (\n"
|
|
" .param .u64 __cudaparm_kernel_pair_x_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_mor1,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_mor2,\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 .s32 __cudaparm_kernel_pair_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<72>;\n"
|
|
" .reg .u64 %rd<63>;\n"
|
|
" .reg .f32 %f<104>;\n"
|
|
" .reg .f64 %fd<10>;\n"
|
|
" .reg .pred %p<19>;\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32497_33_non_const_sp_lj92[16];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32582_35_non_const_red_acc108[3072];\n"
|
|
" .loc 16 88 0\n"
|
|
"$LDWbegin_kernel_pair:\n"
|
|
" .loc 16 95 0\n"
|
|
" ld.param.u64 %rd1, [__cudaparm_kernel_pair_sp_lj_in];\n"
|
|
" ldu.global.f32 %f1, [%rd1+0];\n"
|
|
" .loc 16 96 0\n"
|
|
" ld.global.f32 %f2, [%rd1+4];\n"
|
|
" .loc 16 97 0\n"
|
|
" ld.global.f32 %f3, [%rd1+8];\n"
|
|
" .loc 16 98 0\n"
|
|
" ld.global.f32 %f4, [%rd1+12];\n"
|
|
" st.shared.v4.f32 [__cuda___cuda_local_var_32497_33_non_const_sp_lj92+0], {%f1,%f2,%f3,%f4};\n"
|
|
" .loc 16 107 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"
|
|
" 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_19202;\n"
|
|
" .loc 16 113 0\n"
|
|
" ld.param.s32 %r11, [__cudaparm_kernel_pair_nbor_pitch];\n"
|
|
" cvt.s64.s32 %rd2, %r11;\n"
|
|
" mul.wide.s32 %rd3, %r11, 4;\n"
|
|
" cvt.s64.s32 %rd4, %r9;\n"
|
|
" mul.wide.s32 %rd5, %r9, 4;\n"
|
|
" ld.param.u64 %rd6, [__cudaparm_kernel_pair_dev_nbor];\n"
|
|
" add.u64 %rd7, %rd5, %rd6;\n"
|
|
" add.u64 %rd8, %rd3, %rd7;\n"
|
|
" ld.global.s32 %r12, [%rd8+0];\n"
|
|
" add.u64 %rd9, %rd3, %rd8;\n"
|
|
" ld.param.u64 %rd10, [__cudaparm_kernel_pair_dev_packed];\n"
|
|
" setp.ne.u64 %p2, %rd10, %rd6;\n"
|
|
" @%p2 bra $Lt_0_19714;\n"
|
|
" .loc 16 119 0\n"
|
|
" cvt.s32.s64 %r13, %rd2;\n"
|
|
" mul.lo.s32 %r14, %r13, %r12;\n"
|
|
" cvt.s64.s32 %rd11, %r14;\n"
|
|
" mul.wide.s32 %rd12, %r14, 4;\n"
|
|
" add.u64 %rd13, %rd9, %rd12;\n"
|
|
" .loc 16 120 0\n"
|
|
" mul.lo.s32 %r15, %r6, %r13;\n"
|
|
" cvt.s64.s32 %rd14, %r15;\n"
|
|
" mul.wide.s32 %rd15, %r15, 4;\n"
|
|
" add.u64 %rd16, %rd9, %rd15;\n"
|
|
" .loc 16 121 0\n"
|
|
" mul.lo.s32 %r16, %r13, %r1;\n"
|
|
" bra.uni $Lt_0_19458;\n"
|
|
"$Lt_0_19714:\n"
|
|
" .loc 16 123 0\n"
|
|
" ld.global.s32 %r17, [%rd9+0];\n"
|
|
" cvt.s64.s32 %rd17, %r17;\n"
|
|
" mul.wide.s32 %rd18, %r17, 4;\n"
|
|
" add.u64 %rd19, %rd10, %rd18;\n"
|
|
" .loc 16 124 0\n"
|
|
" cvt.s64.s32 %rd20, %r12;\n"
|
|
" mul.wide.s32 %rd21, %r12, 4;\n"
|
|
" add.u64 %rd13, %rd19, %rd21;\n"
|
|
" .loc 16 125 0\n"
|
|
" mov.s32 %r16, %r1;\n"
|
|
" .loc 16 126 0\n"
|
|
" cvt.s64.s32 %rd22, %r6;\n"
|
|
" mul.wide.s32 %rd23, %r6, 4;\n"
|
|
" add.u64 %rd16, %rd19, %rd23;\n"
|
|
"$Lt_0_19458:\n"
|
|
" .loc 16 129 0\n"
|
|
" ld.global.s32 %r18, [%rd7+0];\n"
|
|
" mov.u32 %r19, %r18;\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 {%f17,%f18,%f19,%f20},[pos_tex,{%r19,%r21,%r23,%r25}];\n"
|
|
" mov.f32 %f21, %f17;\n"
|
|
" mov.f32 %f22, %f18;\n"
|
|
" mov.f32 %f23, %f19;\n"
|
|
" mov.f32 %f24, %f20;\n"
|
|
" setp.ge.u64 %p3, %rd16, %rd13;\n"
|
|
" @%p3 bra $Lt_0_28162;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r26, %f24;\n"
|
|
" cvt.s64.s32 %rd24, %r16;\n"
|
|
" ld.param.s32 %r27, [__cudaparm_kernel_pair_lj_types];\n"
|
|
" mul.lo.s32 %r28, %r27, %r26;\n"
|
|
" ld.param.u64 %rd25, [__cudaparm_kernel_pair_mor1];\n"
|
|
" mov.f32 %f25, 0f00000000; \n"
|
|
" mov.f32 %f26, 0f00000000; \n"
|
|
" mov.f32 %f27, 0f00000000; \n"
|
|
" mov.f32 %f28, 0f00000000; \n"
|
|
" mov.u64 %rd26, __cuda___cuda_local_var_32497_33_non_const_sp_lj92;\n"
|
|
"$Lt_0_20482:\n"
|
|
" .loc 16 135 0\n"
|
|
" ld.global.s32 %r29, [%rd16+0];\n"
|
|
" .loc 16 136 0\n"
|
|
" shr.s32 %r30, %r29, 30;\n"
|
|
" and.b32 %r31, %r30, 3;\n"
|
|
" cvt.s64.s32 %rd27, %r31;\n"
|
|
" mul.wide.s32 %rd28, %r31, 4;\n"
|
|
" add.u64 %rd29, %rd26, %rd28;\n"
|
|
" ld.shared.f32 %f29, [%rd29+0];\n"
|
|
" .loc 16 139 0\n"
|
|
" and.b32 %r32, %r29, 1073741823;\n"
|
|
" mov.u32 %r33, %r32;\n"
|
|
" mov.s32 %r34, 0;\n"
|
|
" mov.u32 %r35, %r34;\n"
|
|
" mov.s32 %r36, 0;\n"
|
|
" mov.u32 %r37, %r36;\n"
|
|
" mov.s32 %r38, 0;\n"
|
|
" mov.u32 %r39, %r38;\n"
|
|
" tex.1d.v4.f32.s32 {%f30,%f31,%f32,%f33},[pos_tex,{%r33,%r35,%r37,%r39}];\n"
|
|
" mov.f32 %f34, %f30;\n"
|
|
" mov.f32 %f35, %f31;\n"
|
|
" mov.f32 %f36, %f32;\n"
|
|
" mov.f32 %f37, %f33;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r40, %f37;\n"
|
|
" sub.ftz.f32 %f38, %f22, %f35;\n"
|
|
" sub.ftz.f32 %f39, %f21, %f34;\n"
|
|
" sub.ftz.f32 %f40, %f23, %f36;\n"
|
|
" mul.ftz.f32 %f41, %f38, %f38;\n"
|
|
" fma.rn.ftz.f32 %f42, %f39, %f39, %f41;\n"
|
|
" add.s32 %r41, %r40, %r28;\n"
|
|
" cvt.s64.s32 %rd30, %r41;\n"
|
|
" fma.rn.ftz.f32 %f43, %f40, %f40, %f42;\n"
|
|
" mul.wide.s32 %rd31, %r41, 16;\n"
|
|
" add.u64 %rd32, %rd25, %rd31;\n"
|
|
" ld.global.f32 %f44, [%rd32+0];\n"
|
|
" setp.gt.ftz.f32 %p4, %f44, %f43;\n"
|
|
" @!%p4 bra $Lt_0_21762;\n"
|
|
" .loc 16 152 0\n"
|
|
" sqrt.approx.ftz.f32 %f45, %f43;\n"
|
|
" ld.global.v4.f32 {_,%f46,%f47,%f48}, [%rd32+0];\n"
|
|
" sub.ftz.f32 %f49, %f45, %f47;\n"
|
|
" mul.ftz.f32 %f50, %f48, %f49;\n"
|
|
" neg.ftz.f32 %f51, %f50;\n"
|
|
" .loc 16 154 0\n"
|
|
" mov.f32 %f52, 0f3fb8aa3b; \n"
|
|
" mul.ftz.f32 %f53, %f51, %f52;\n"
|
|
" ex2.approx.ftz.f32 %f54, %f53;\n"
|
|
" mul.ftz.f32 %f55, %f54, %f54;\n"
|
|
" sub.ftz.f32 %f56, %f55, %f54;\n"
|
|
" mul.ftz.f32 %f57, %f46, %f56;\n"
|
|
" .loc 16 156 0\n"
|
|
" div.approx.ftz.f32 %f58, %f57, %f45;\n"
|
|
" mul.ftz.f32 %f59, %f58, %f29;\n"
|
|
" fma.rn.ftz.f32 %f27, %f39, %f59, %f27;\n"
|
|
" .loc 16 157 0\n"
|
|
" fma.rn.ftz.f32 %f26, %f38, %f59, %f26;\n"
|
|
" .loc 16 158 0\n"
|
|
" fma.rn.ftz.f32 %f25, %f40, %f59, %f25;\n"
|
|
" ld.param.s32 %r42, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r43, 0;\n"
|
|
" setp.le.s32 %p5, %r42, %r43;\n"
|
|
" @%p5 bra $Lt_0_21250;\n"
|
|
" .loc 16 162 0\n"
|
|
" cvt.ftz.f64.f32 %fd1, %f54;\n"
|
|
" ld.param.u64 %rd33, [__cudaparm_kernel_pair_mor2];\n"
|
|
" mul.lo.u64 %rd34, %rd30, 8;\n"
|
|
" add.u64 %rd35, %rd33, %rd34;\n"
|
|
" ld.global.v2.f32 {%f60,%f61}, [%rd35+0];\n"
|
|
" cvt.ftz.f64.f32 %fd2, %f61;\n"
|
|
" cvt.ftz.f64.f32 %fd3, %f60;\n"
|
|
" mul.ftz.f32 %f62, %f54, %f54;\n"
|
|
" cvt.ftz.f64.f32 %fd4, %f62;\n"
|
|
" add.f64 %fd5, %fd1, %fd1;\n"
|
|
" sub.f64 %fd6, %fd4, %fd5;\n"
|
|
" mul.f64 %fd7, %fd3, %fd6;\n"
|
|
" sub.f64 %fd8, %fd7, %fd2;\n"
|
|
" cvt.rn.ftz.f32.f64 %f63, %fd8;\n"
|
|
" fma.rn.ftz.f32 %f28, %f29, %f63, %f28;\n"
|
|
"$Lt_0_21250:\n"
|
|
" ld.param.s32 %r44, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r45, 0;\n"
|
|
" setp.le.s32 %p6, %r44, %r45;\n"
|
|
" @%p6 bra $Lt_0_21762;\n"
|
|
" .loc 16 165 0\n"
|
|
" mov.f32 %f64, %f6;\n"
|
|
" mul.ftz.f32 %f65, %f39, %f39;\n"
|
|
" fma.rn.ftz.f32 %f66, %f59, %f65, %f64;\n"
|
|
" mov.f32 %f6, %f66;\n"
|
|
" .loc 16 166 0\n"
|
|
" mov.f32 %f67, %f8;\n"
|
|
" fma.rn.ftz.f32 %f68, %f59, %f41, %f67;\n"
|
|
" mov.f32 %f8, %f68;\n"
|
|
" .loc 16 167 0\n"
|
|
" mov.f32 %f69, %f10;\n"
|
|
" mul.ftz.f32 %f70, %f40, %f40;\n"
|
|
" fma.rn.ftz.f32 %f71, %f59, %f70, %f69;\n"
|
|
" mov.f32 %f10, %f71;\n"
|
|
" .loc 16 168 0\n"
|
|
" mov.f32 %f72, %f12;\n"
|
|
" mul.ftz.f32 %f73, %f38, %f39;\n"
|
|
" fma.rn.ftz.f32 %f74, %f59, %f73, %f72;\n"
|
|
" mov.f32 %f12, %f74;\n"
|
|
" .loc 16 169 0\n"
|
|
" mov.f32 %f75, %f14;\n"
|
|
" mul.ftz.f32 %f76, %f39, %f40;\n"
|
|
" fma.rn.ftz.f32 %f77, %f59, %f76, %f75;\n"
|
|
" mov.f32 %f14, %f77;\n"
|
|
" .loc 16 170 0\n"
|
|
" mul.ftz.f32 %f78, %f38, %f40;\n"
|
|
" fma.rn.ftz.f32 %f15, %f59, %f78, %f15;\n"
|
|
" mov.f32 %f16, %f15;\n"
|
|
"$Lt_0_21762:\n"
|
|
"$Lt_0_20738:\n"
|
|
" .loc 16 133 0\n"
|
|
" mul.lo.u64 %rd36, %rd24, 4;\n"
|
|
" add.u64 %rd16, %rd16, %rd36;\n"
|
|
" setp.lt.u64 %p7, %rd16, %rd13;\n"
|
|
" @%p7 bra $Lt_0_20482;\n"
|
|
" bra.uni $Lt_0_18946;\n"
|
|
"$Lt_0_28162:\n"
|
|
" mov.f32 %f25, 0f00000000; \n"
|
|
" mov.f32 %f26, 0f00000000; \n"
|
|
" mov.f32 %f27, 0f00000000; \n"
|
|
" mov.f32 %f28, 0f00000000; \n"
|
|
" bra.uni $Lt_0_18946;\n"
|
|
"$Lt_0_19202:\n"
|
|
" mov.f32 %f25, 0f00000000; \n"
|
|
" mov.f32 %f26, 0f00000000; \n"
|
|
" mov.f32 %f27, 0f00000000; \n"
|
|
" mov.f32 %f28, 0f00000000; \n"
|
|
"$Lt_0_18946:\n"
|
|
" mov.u32 %r46, 1;\n"
|
|
" setp.le.s32 %p8, %r1, %r46;\n"
|
|
" @%p8 bra $Lt_0_24578;\n"
|
|
" .loc 16 181 0\n"
|
|
" mov.u64 %rd37, __cuda___cuda_local_var_32582_35_non_const_red_acc108;\n"
|
|
" cvt.s64.s32 %rd38, %r2;\n"
|
|
" mul.wide.s32 %rd39, %r2, 4;\n"
|
|
" add.u64 %rd40, %rd37, %rd39;\n"
|
|
" mov.f32 %f79, %f27;\n"
|
|
" st.shared.f32 [%rd40+0], %f79;\n"
|
|
" .loc 16 182 0\n"
|
|
" mov.f32 %f80, %f26;\n"
|
|
" st.shared.f32 [%rd40+512], %f80;\n"
|
|
" .loc 16 183 0\n"
|
|
" mov.f32 %f81, %f25;\n"
|
|
" st.shared.f32 [%rd40+1024], %f81;\n"
|
|
" .loc 16 184 0\n"
|
|
" mov.f32 %f82, %f28;\n"
|
|
" st.shared.f32 [%rd40+1536], %f82;\n"
|
|
" .loc 16 186 0\n"
|
|
" shr.s32 %r47, %r1, 31;\n"
|
|
" mov.s32 %r48, 1;\n"
|
|
" and.b32 %r49, %r47, %r48;\n"
|
|
" add.s32 %r50, %r49, %r1;\n"
|
|
" shr.s32 %r51, %r50, 1;\n"
|
|
" mov.s32 %r52, %r51;\n"
|
|
" mov.u32 %r53, 0;\n"
|
|
" setp.ne.u32 %p9, %r51, %r53;\n"
|
|
" @!%p9 bra $Lt_0_23042;\n"
|
|
"$Lt_0_23554:\n"
|
|
" setp.ge.u32 %p10, %r6, %r52;\n"
|
|
" @%p10 bra $Lt_0_23810;\n"
|
|
" .loc 16 189 0\n"
|
|
" add.u32 %r54, %r2, %r52;\n"
|
|
" cvt.u64.u32 %rd41, %r54;\n"
|
|
" mul.wide.u32 %rd42, %r54, 4;\n"
|
|
" add.u64 %rd43, %rd37, %rd42;\n"
|
|
" ld.shared.f32 %f83, [%rd43+0];\n"
|
|
" add.ftz.f32 %f79, %f83, %f79;\n"
|
|
" st.shared.f32 [%rd40+0], %f79;\n"
|
|
" ld.shared.f32 %f84, [%rd43+512];\n"
|
|
" add.ftz.f32 %f80, %f84, %f80;\n"
|
|
" st.shared.f32 [%rd40+512], %f80;\n"
|
|
" ld.shared.f32 %f85, [%rd43+1024];\n"
|
|
" add.ftz.f32 %f81, %f85, %f81;\n"
|
|
" st.shared.f32 [%rd40+1024], %f81;\n"
|
|
" ld.shared.f32 %f86, [%rd43+1536];\n"
|
|
" add.ftz.f32 %f82, %f86, %f82;\n"
|
|
" st.shared.f32 [%rd40+1536], %f82;\n"
|
|
"$Lt_0_23810:\n"
|
|
" .loc 16 186 0\n"
|
|
" shr.u32 %r52, %r52, 1;\n"
|
|
" mov.u32 %r55, 0;\n"
|
|
" setp.ne.u32 %p11, %r52, %r55;\n"
|
|
" @%p11 bra $Lt_0_23554;\n"
|
|
"$Lt_0_23042:\n"
|
|
" .loc 16 193 0\n"
|
|
" mov.f32 %f27, %f79;\n"
|
|
" .loc 16 194 0\n"
|
|
" mov.f32 %f26, %f80;\n"
|
|
" .loc 16 195 0\n"
|
|
" mov.f32 %f25, %f81;\n"
|
|
" .loc 16 196 0\n"
|
|
" mov.f32 %f28, %f82;\n"
|
|
" ld.param.s32 %r56, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r57, 0;\n"
|
|
" setp.le.s32 %p12, %r56, %r57;\n"
|
|
" @%p12 bra $Lt_0_24578;\n"
|
|
" .loc 16 200 0\n"
|
|
" mov.f32 %f79, %f6;\n"
|
|
" st.shared.f32 [%rd40+0], %f79;\n"
|
|
" mov.f32 %f80, %f8;\n"
|
|
" st.shared.f32 [%rd40+512], %f80;\n"
|
|
" mov.f32 %f81, %f10;\n"
|
|
" st.shared.f32 [%rd40+1024], %f81;\n"
|
|
" mov.f32 %f82, %f12;\n"
|
|
" st.shared.f32 [%rd40+1536], %f82;\n"
|
|
" mov.f32 %f87, %f14;\n"
|
|
" st.shared.f32 [%rd40+2048], %f87;\n"
|
|
" mov.f32 %f88, %f16;\n"
|
|
" st.shared.f32 [%rd40+2560], %f88;\n"
|
|
" .loc 16 202 0\n"
|
|
" mov.s32 %r58, %r51;\n"
|
|
" @!%p9 bra $Lt_0_25090;\n"
|
|
"$Lt_0_25602:\n"
|
|
" setp.ge.u32 %p13, %r6, %r58;\n"
|
|
" @%p13 bra $Lt_0_25858;\n"
|
|
" .loc 16 205 0\n"
|
|
" add.u32 %r59, %r2, %r58;\n"
|
|
" cvt.u64.u32 %rd44, %r59;\n"
|
|
" mul.wide.u32 %rd45, %r59, 4;\n"
|
|
" add.u64 %rd46, %rd37, %rd45;\n"
|
|
" ld.shared.f32 %f89, [%rd46+0];\n"
|
|
" add.ftz.f32 %f79, %f89, %f79;\n"
|
|
" st.shared.f32 [%rd40+0], %f79;\n"
|
|
" ld.shared.f32 %f90, [%rd46+512];\n"
|
|
" add.ftz.f32 %f80, %f90, %f80;\n"
|
|
" st.shared.f32 [%rd40+512], %f80;\n"
|
|
" ld.shared.f32 %f91, [%rd46+1024];\n"
|
|
" add.ftz.f32 %f81, %f91, %f81;\n"
|
|
" st.shared.f32 [%rd40+1024], %f81;\n"
|
|
" ld.shared.f32 %f92, [%rd46+1536];\n"
|
|
" add.ftz.f32 %f82, %f92, %f82;\n"
|
|
" st.shared.f32 [%rd40+1536], %f82;\n"
|
|
" ld.shared.f32 %f93, [%rd46+2048];\n"
|
|
" add.ftz.f32 %f87, %f93, %f87;\n"
|
|
" st.shared.f32 [%rd40+2048], %f87;\n"
|
|
" ld.shared.f32 %f94, [%rd46+2560];\n"
|
|
" add.ftz.f32 %f88, %f94, %f88;\n"
|
|
" st.shared.f32 [%rd40+2560], %f88;\n"
|
|
"$Lt_0_25858:\n"
|
|
" .loc 16 202 0\n"
|
|
" shr.u32 %r58, %r58, 1;\n"
|
|
" mov.u32 %r60, 0;\n"
|
|
" setp.ne.u32 %p14, %r58, %r60;\n"
|
|
" @%p14 bra $Lt_0_25602;\n"
|
|
"$Lt_0_25090:\n"
|
|
" .loc 16 210 0\n"
|
|
" mov.f32 %f6, %f79;\n"
|
|
" mov.f32 %f8, %f80;\n"
|
|
" mov.f32 %f10, %f81;\n"
|
|
" mov.f32 %f12, %f82;\n"
|
|
" mov.f32 %f14, %f87;\n"
|
|
" mov.f32 %f16, %f88;\n"
|
|
"$Lt_0_24578:\n"
|
|
"$Lt_0_22530:\n"
|
|
" selp.s32 %r61, 1, 0, %p1;\n"
|
|
" mov.s32 %r62, 0;\n"
|
|
" set.eq.u32.s32 %r63, %r6, %r62;\n"
|
|
" neg.s32 %r64, %r63;\n"
|
|
" and.b32 %r65, %r61, %r64;\n"
|
|
" mov.u32 %r66, 0;\n"
|
|
" setp.eq.s32 %p15, %r65, %r66;\n"
|
|
" @%p15 bra $Lt_0_26626;\n"
|
|
" .loc 16 216 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 %r67, [__cudaparm_kernel_pair_eflag];\n"
|
|
" mov.u32 %r68, 0;\n"
|
|
" setp.le.s32 %p16, %r67, %r68;\n"
|
|
" @%p16 bra $Lt_0_27138;\n"
|
|
" .loc 16 218 0\n"
|
|
" st.global.f32 [%rd50+0], %f28;\n"
|
|
" .loc 16 219 0\n"
|
|
" cvt.s64.s32 %rd51, %r10;\n"
|
|
" mul.wide.s32 %rd52, %r10, 4;\n"
|
|
" add.u64 %rd50, %rd50, %rd52;\n"
|
|
"$Lt_0_27138:\n"
|
|
" ld.param.s32 %r69, [__cudaparm_kernel_pair_vflag];\n"
|
|
" mov.u32 %r70, 0;\n"
|
|
" setp.le.s32 %p17, %r69, %r70;\n"
|
|
" @%p17 bra $Lt_0_27650;\n"
|
|
" .loc 16 223 0\n"
|
|
" mov.f32 %f95, %f6;\n"
|
|
" st.global.f32 [%rd50+0], %f95;\n"
|
|
" .loc 16 224 0\n"
|
|
" cvt.s64.s32 %rd53, %r10;\n"
|
|
" mul.wide.s32 %rd54, %r10, 4;\n"
|
|
" add.u64 %rd55, %rd54, %rd50;\n"
|
|
" .loc 16 223 0\n"
|
|
" mov.f32 %f96, %f8;\n"
|
|
" st.global.f32 [%rd55+0], %f96;\n"
|
|
" .loc 16 224 0\n"
|
|
" add.u64 %rd56, %rd54, %rd55;\n"
|
|
" .loc 16 223 0\n"
|
|
" mov.f32 %f97, %f10;\n"
|
|
" st.global.f32 [%rd56+0], %f97;\n"
|
|
" .loc 16 224 0\n"
|
|
" add.u64 %rd57, %rd54, %rd56;\n"
|
|
" .loc 16 223 0\n"
|
|
" mov.f32 %f98, %f12;\n"
|
|
" st.global.f32 [%rd57+0], %f98;\n"
|
|
" .loc 16 224 0\n"
|
|
" add.u64 %rd50, %rd54, %rd57;\n"
|
|
" .loc 16 223 0\n"
|
|
" mov.f32 %f99, %f14;\n"
|
|
" st.global.f32 [%rd50+0], %f99;\n"
|
|
" mov.f32 %f100, %f16;\n"
|
|
" add.u64 %rd58, %rd54, %rd50;\n"
|
|
" st.global.f32 [%rd58+0], %f100;\n"
|
|
"$Lt_0_27650:\n"
|
|
" .loc 16 227 0\n"
|
|
" ld.param.u64 %rd59, [__cudaparm_kernel_pair_ans];\n"
|
|
" mul.lo.u64 %rd60, %rd47, 16;\n"
|
|
" add.u64 %rd61, %rd59, %rd60;\n"
|
|
" mov.f32 %f101, %f102;\n"
|
|
" st.global.v4.f32 [%rd61+0], {%f27,%f26,%f25,%f101};\n"
|
|
"$Lt_0_26626:\n"
|
|
" .loc 16 229 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair:\n"
|
|
" }\n"
|
|
" .entry kernel_pair_fast (\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_x_,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_mor1_in,\n"
|
|
" .param .u64 __cudaparm_kernel_pair_fast_mor2_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 .s32 __cudaparm_kernel_pair_fast_t_per_atom)\n"
|
|
" {\n"
|
|
" .reg .u32 %r<74>;\n"
|
|
" .reg .u64 %rd<76>;\n"
|
|
" .reg .f32 %f<110>;\n"
|
|
" .reg .pred %p<22>;\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32648_33_non_const_sp_lj3268[16];\n"
|
|
" .shared .align 16 .b8 __cuda___cuda_local_var_32646_34_non_const_mor13296[1936];\n"
|
|
" .shared .align 8 .b8 __cuda___cuda_local_var_32647_34_non_const_mor25232[968];\n"
|
|
" .shared .align 4 .b8 __cuda___cuda_local_var_32738_35_non_const_red_acc6200[3072];\n"
|
|
" .loc 16 237 0\n"
|
|
"$LDWbegin_kernel_pair_fast:\n"
|
|
" cvt.s32.u32 %r1, %tid.x;\n"
|
|
" mov.u32 %r2, 3;\n"
|
|
" setp.gt.s32 %p1, %r1, %r2;\n"
|
|
" @%p1 bra $Lt_1_21250;\n"
|
|
" .loc 16 247 0\n"
|
|
" mov.u64 %rd1, __cuda___cuda_local_var_32648_33_non_const_sp_lj3268;\n"
|
|
" cvt.s64.s32 %rd2, %r1;\n"
|
|
" mul.wide.s32 %rd3, %r1, 4;\n"
|
|
" ld.param.u64 %rd4, [__cudaparm_kernel_pair_fast_sp_lj_in];\n"
|
|
" add.u64 %rd5, %rd4, %rd3;\n"
|
|
" ld.global.f32 %f1, [%rd5+0];\n"
|
|
" add.u64 %rd6, %rd3, %rd1;\n"
|
|
" st.shared.f32 [%rd6+0], %f1;\n"
|
|
"$Lt_1_21250:\n"
|
|
" mov.u64 %rd1, __cuda___cuda_local_var_32648_33_non_const_sp_lj3268;\n"
|
|
" mov.u32 %r3, 120;\n"
|
|
" setp.gt.s32 %p2, %r1, %r3;\n"
|
|
" @%p2 bra $Lt_1_21762;\n"
|
|
" .loc 16 249 0\n"
|
|
" mov.u64 %rd7, __cuda___cuda_local_var_32646_34_non_const_mor13296;\n"
|
|
" cvt.s64.s32 %rd8, %r1;\n"
|
|
" mul.wide.s32 %rd9, %r1, 16;\n"
|
|
" ld.param.u64 %rd10, [__cudaparm_kernel_pair_fast_mor1_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.v4.f32 [%rd12+0], {%f2,%f3,%f4,%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 $Lt_1_22274;\n"
|
|
" .loc 16 251 0\n"
|
|
" mov.u64 %rd13, __cuda___cuda_local_var_32647_34_non_const_mor25232;\n"
|
|
" mul.lo.u64 %rd14, %rd8, 8;\n"
|
|
" ld.param.u64 %rd15, [__cudaparm_kernel_pair_fast_mor2_in];\n"
|
|
" add.u64 %rd16, %rd15, %rd14;\n"
|
|
" add.u64 %rd17, %rd14, %rd13;\n"
|
|
" ld.global.v2.f32 {%f6,%f7}, [%rd16+0];\n"
|
|
" st.shared.v2.f32 [%rd17+0], {%f6,%f7};\n"
|
|
"$Lt_1_22274:\n"
|
|
" mov.u64 %rd13, __cuda___cuda_local_var_32647_34_non_const_mor25232;\n"
|
|
"$Lt_1_21762:\n"
|
|
" mov.u64 %rd7, __cuda___cuda_local_var_32646_34_non_const_mor13296;\n"
|
|
" mov.u64 %rd13, __cuda___cuda_local_var_32647_34_non_const_mor25232;\n"
|
|
" .loc 16 261 0\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"
|
|
" mov.f32 %f18, 0f00000000; \n"
|
|
" mov.f32 %f19, %f18;\n"
|
|
" .loc 16 263 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_23042;\n"
|
|
" .loc 16 269 0\n"
|
|
" ld.param.s32 %r15, [__cudaparm_kernel_pair_fast_nbor_pitch];\n"
|
|
" cvt.s64.s32 %rd18, %r15;\n"
|
|
" mul.wide.s32 %rd19, %r15, 4;\n"
|
|
" cvt.s64.s32 %rd20, %r13;\n"
|
|
" mul.wide.s32 %rd21, %r13, 4;\n"
|
|
" ld.param.u64 %rd22, [__cudaparm_kernel_pair_fast_dev_nbor];\n"
|
|
" add.u64 %rd23, %rd21, %rd22;\n"
|
|
" add.u64 %rd24, %rd19, %rd23;\n"
|
|
" ld.global.s32 %r16, [%rd24+0];\n"
|
|
" add.u64 %rd25, %rd19, %rd24;\n"
|
|
" ld.param.u64 %rd26, [__cudaparm_kernel_pair_fast_dev_packed];\n"
|
|
" setp.ne.u64 %p5, %rd26, %rd22;\n"
|
|
" @%p5 bra $Lt_1_23554;\n"
|
|
" .loc 16 275 0\n"
|
|
" cvt.s32.s64 %r17, %rd18;\n"
|
|
" mul.lo.s32 %r18, %r17, %r16;\n"
|
|
" cvt.s64.s32 %rd27, %r18;\n"
|
|
" mul.wide.s32 %rd28, %r18, 4;\n"
|
|
" add.u64 %rd29, %rd25, %rd28;\n"
|
|
" .loc 16 276 0\n"
|
|
" mul.lo.s32 %r19, %r10, %r17;\n"
|
|
" cvt.s64.s32 %rd30, %r19;\n"
|
|
" mul.wide.s32 %rd31, %r19, 4;\n"
|
|
" add.u64 %rd32, %rd25, %rd31;\n"
|
|
" .loc 16 277 0\n"
|
|
" mul.lo.s32 %r20, %r17, %r6;\n"
|
|
" bra.uni $Lt_1_23298;\n"
|
|
"$Lt_1_23554:\n"
|
|
" .loc 16 279 0\n"
|
|
" ld.global.s32 %r21, [%rd25+0];\n"
|
|
" cvt.s64.s32 %rd33, %r21;\n"
|
|
" mul.wide.s32 %rd34, %r21, 4;\n"
|
|
" add.u64 %rd35, %rd26, %rd34;\n"
|
|
" .loc 16 280 0\n"
|
|
" cvt.s64.s32 %rd36, %r16;\n"
|
|
" mul.wide.s32 %rd37, %r16, 4;\n"
|
|
" add.u64 %rd29, %rd35, %rd37;\n"
|
|
" .loc 16 281 0\n"
|
|
" mov.s32 %r20, %r6;\n"
|
|
" .loc 16 282 0\n"
|
|
" cvt.s64.s32 %rd38, %r10;\n"
|
|
" mul.wide.s32 %rd39, %r10, 4;\n"
|
|
" add.u64 %rd32, %rd35, %rd39;\n"
|
|
"$Lt_1_23298:\n"
|
|
" .loc 16 285 0\n"
|
|
" ld.global.s32 %r22, [%rd23+0];\n"
|
|
" mov.u32 %r23, %r22;\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 {%f20,%f21,%f22,%f23},[pos_tex,{%r23,%r25,%r27,%r29}];\n"
|
|
" mov.f32 %f24, %f20;\n"
|
|
" mov.f32 %f25, %f21;\n"
|
|
" mov.f32 %f26, %f22;\n"
|
|
" mov.f32 %f27, %f23;\n"
|
|
" setp.ge.u64 %p6, %rd32, %rd29;\n"
|
|
" @%p6 bra $Lt_1_32002;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r30, %f27;\n"
|
|
" cvt.s64.s32 %rd40, %r20;\n"
|
|
" mul.lo.s32 %r31, %r30, 11;\n"
|
|
" cvt.rn.f32.s32 %f28, %r31;\n"
|
|
" mov.f32 %f29, 0f00000000; \n"
|
|
" mov.f32 %f30, 0f00000000; \n"
|
|
" mov.f32 %f31, 0f00000000; \n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
"$Lt_1_24322:\n"
|
|
" .loc 16 292 0\n"
|
|
" ld.global.s32 %r32, [%rd32+0];\n"
|
|
" .loc 16 293 0\n"
|
|
" shr.s32 %r33, %r32, 30;\n"
|
|
" and.b32 %r34, %r33, 3;\n"
|
|
" cvt.s64.s32 %rd41, %r34;\n"
|
|
" mul.wide.s32 %rd42, %r34, 4;\n"
|
|
" add.u64 %rd43, %rd1, %rd42;\n"
|
|
" ld.shared.f32 %f33, [%rd43+0];\n"
|
|
" .loc 16 296 0\n"
|
|
" and.b32 %r35, %r32, 1073741823;\n"
|
|
" mov.u32 %r36, %r35;\n"
|
|
" mov.s32 %r37, 0;\n"
|
|
" mov.u32 %r38, %r37;\n"
|
|
" mov.s32 %r39, 0;\n"
|
|
" mov.u32 %r40, %r39;\n"
|
|
" mov.s32 %r41, 0;\n"
|
|
" mov.u32 %r42, %r41;\n"
|
|
" tex.1d.v4.f32.s32 {%f34,%f35,%f36,%f37},[pos_tex,{%r36,%r38,%r40,%r42}];\n"
|
|
" mov.f32 %f38, %f34;\n"
|
|
" mov.f32 %f39, %f35;\n"
|
|
" mov.f32 %f40, %f36;\n"
|
|
" mov.f32 %f41, %f37;\n"
|
|
" sub.ftz.f32 %f42, %f25, %f39;\n"
|
|
" sub.ftz.f32 %f43, %f24, %f38;\n"
|
|
" sub.ftz.f32 %f44, %f26, %f40;\n"
|
|
" mul.ftz.f32 %f45, %f42, %f42;\n"
|
|
" fma.rn.ftz.f32 %f46, %f43, %f43, %f45;\n"
|
|
" fma.rn.ftz.f32 %f47, %f44, %f44, %f46;\n"
|
|
" add.ftz.f32 %f48, %f28, %f41;\n"
|
|
" cvt.rzi.ftz.s32.f32 %r43, %f48;\n"
|
|
" cvt.s64.s32 %rd44, %r43;\n"
|
|
" mul.wide.s32 %rd45, %r43, 16;\n"
|
|
" add.u64 %rd46, %rd7, %rd45;\n"
|
|
" ld.shared.f32 %f49, [%rd46+0];\n"
|
|
" setp.gt.ftz.f32 %p7, %f49, %f47;\n"
|
|
" @!%p7 bra $Lt_1_25602;\n"
|
|
" .loc 16 307 0\n"
|
|
" sqrt.approx.ftz.f32 %f50, %f47;\n"
|
|
" ld.shared.v4.f32 {_,%f51,%f52,%f53}, [%rd46+0];\n"
|
|
" sub.ftz.f32 %f54, %f50, %f52;\n"
|
|
" .loc 16 308 0\n"
|
|
" mul.ftz.f32 %f55, %f53, %f54;\n"
|
|
" neg.ftz.f32 %f56, %f55;\n"
|
|
" .loc 16 310 0\n"
|
|
" mov.f32 %f57, 0f3fb8aa3b; \n"
|
|
" mul.ftz.f32 %f58, %f56, %f57;\n"
|
|
" ex2.approx.ftz.f32 %f59, %f58;\n"
|
|
" mul.ftz.f32 %f60, %f59, %f59;\n"
|
|
" sub.ftz.f32 %f61, %f60, %f59;\n"
|
|
" mul.ftz.f32 %f62, %f51, %f61;\n"
|
|
" .loc 16 312 0\n"
|
|
" div.approx.ftz.f32 %f63, %f62, %f50;\n"
|
|
" mul.ftz.f32 %f64, %f63, %f33;\n"
|
|
" fma.rn.ftz.f32 %f31, %f43, %f64, %f31;\n"
|
|
" .loc 16 313 0\n"
|
|
" fma.rn.ftz.f32 %f30, %f42, %f64, %f30;\n"
|
|
" .loc 16 314 0\n"
|
|
" fma.rn.ftz.f32 %f29, %f44, %f64, %f29;\n"
|
|
" ld.param.s32 %r44, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r45, 0;\n"
|
|
" setp.le.s32 %p8, %r44, %r45;\n"
|
|
" @%p8 bra $Lt_1_25090;\n"
|
|
" .loc 16 317 0\n"
|
|
" mul.lo.u64 %rd47, %rd44, 8;\n"
|
|
" add.u64 %rd48, %rd13, %rd47;\n"
|
|
" ld.shared.v2.f32 {%f65,%f66}, [%rd48+0];\n"
|
|
" sub.ftz.f32 %f67, %f61, %f59;\n"
|
|
" mul.ftz.f32 %f68, %f65, %f67;\n"
|
|
" sub.ftz.f32 %f69, %f68, %f66;\n"
|
|
" .loc 16 318 0\n"
|
|
" fma.rn.ftz.f32 %f32, %f33, %f69, %f32;\n"
|
|
"$Lt_1_25090:\n"
|
|
" ld.param.s32 %r46, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r47, 0;\n"
|
|
" setp.le.s32 %p9, %r46, %r47;\n"
|
|
" @%p9 bra $Lt_1_25602;\n"
|
|
" .loc 16 321 0\n"
|
|
" mov.f32 %f70, %f9;\n"
|
|
" mul.ftz.f32 %f71, %f43, %f43;\n"
|
|
" fma.rn.ftz.f32 %f72, %f64, %f71, %f70;\n"
|
|
" mov.f32 %f9, %f72;\n"
|
|
" .loc 16 322 0\n"
|
|
" mov.f32 %f73, %f11;\n"
|
|
" fma.rn.ftz.f32 %f74, %f64, %f45, %f73;\n"
|
|
" mov.f32 %f11, %f74;\n"
|
|
" .loc 16 323 0\n"
|
|
" mov.f32 %f75, %f13;\n"
|
|
" mul.ftz.f32 %f76, %f44, %f44;\n"
|
|
" fma.rn.ftz.f32 %f77, %f64, %f76, %f75;\n"
|
|
" mov.f32 %f13, %f77;\n"
|
|
" .loc 16 324 0\n"
|
|
" mov.f32 %f78, %f15;\n"
|
|
" mul.ftz.f32 %f79, %f42, %f43;\n"
|
|
" fma.rn.ftz.f32 %f80, %f64, %f79, %f78;\n"
|
|
" mov.f32 %f15, %f80;\n"
|
|
" .loc 16 325 0\n"
|
|
" mov.f32 %f81, %f17;\n"
|
|
" mul.ftz.f32 %f82, %f43, %f44;\n"
|
|
" fma.rn.ftz.f32 %f83, %f64, %f82, %f81;\n"
|
|
" mov.f32 %f17, %f83;\n"
|
|
" .loc 16 326 0\n"
|
|
" mul.ftz.f32 %f84, %f42, %f44;\n"
|
|
" fma.rn.ftz.f32 %f18, %f64, %f84, %f18;\n"
|
|
" mov.f32 %f19, %f18;\n"
|
|
"$Lt_1_25602:\n"
|
|
"$Lt_1_24578:\n"
|
|
" .loc 16 290 0\n"
|
|
" mul.lo.u64 %rd49, %rd40, 4;\n"
|
|
" add.u64 %rd32, %rd32, %rd49;\n"
|
|
" setp.lt.u64 %p10, %rd32, %rd29;\n"
|
|
" @%p10 bra $Lt_1_24322;\n"
|
|
" bra.uni $Lt_1_22786;\n"
|
|
"$Lt_1_32002:\n"
|
|
" mov.f32 %f29, 0f00000000; \n"
|
|
" mov.f32 %f30, 0f00000000; \n"
|
|
" mov.f32 %f31, 0f00000000; \n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
" bra.uni $Lt_1_22786;\n"
|
|
"$Lt_1_23042:\n"
|
|
" mov.f32 %f29, 0f00000000; \n"
|
|
" mov.f32 %f30, 0f00000000; \n"
|
|
" mov.f32 %f31, 0f00000000; \n"
|
|
" mov.f32 %f32, 0f00000000; \n"
|
|
"$Lt_1_22786:\n"
|
|
" mov.u32 %r48, 1;\n"
|
|
" setp.le.s32 %p11, %r6, %r48;\n"
|
|
" @%p11 bra $Lt_1_28418;\n"
|
|
" .loc 16 337 0\n"
|
|
" mov.u64 %rd50, __cuda___cuda_local_var_32738_35_non_const_red_acc6200;\n"
|
|
" cvt.s64.s32 %rd51, %r1;\n"
|
|
" mul.wide.s32 %rd52, %r1, 4;\n"
|
|
" add.u64 %rd53, %rd50, %rd52;\n"
|
|
" mov.f32 %f85, %f31;\n"
|
|
" st.shared.f32 [%rd53+0], %f85;\n"
|
|
" .loc 16 338 0\n"
|
|
" mov.f32 %f86, %f30;\n"
|
|
" st.shared.f32 [%rd53+512], %f86;\n"
|
|
" .loc 16 339 0\n"
|
|
" mov.f32 %f87, %f29;\n"
|
|
" st.shared.f32 [%rd53+1024], %f87;\n"
|
|
" .loc 16 340 0\n"
|
|
" mov.f32 %f88, %f32;\n"
|
|
" st.shared.f32 [%rd53+1536], %f88;\n"
|
|
" .loc 16 342 0\n"
|
|
" shr.s32 %r49, %r6, 31;\n"
|
|
" mov.s32 %r50, 1;\n"
|
|
" and.b32 %r51, %r49, %r50;\n"
|
|
" add.s32 %r52, %r51, %r6;\n"
|
|
" shr.s32 %r53, %r52, 1;\n"
|
|
" mov.s32 %r54, %r53;\n"
|
|
" mov.u32 %r55, 0;\n"
|
|
" setp.ne.u32 %p12, %r53, %r55;\n"
|
|
" @!%p12 bra $Lt_1_26882;\n"
|
|
"$Lt_1_27394:\n"
|
|
" setp.ge.u32 %p13, %r10, %r54;\n"
|
|
" @%p13 bra $Lt_1_27650;\n"
|
|
" .loc 16 345 0\n"
|
|
" add.u32 %r56, %r1, %r54;\n"
|
|
" cvt.u64.u32 %rd54, %r56;\n"
|
|
" mul.wide.u32 %rd55, %r56, 4;\n"
|
|
" add.u64 %rd56, %rd50, %rd55;\n"
|
|
" ld.shared.f32 %f89, [%rd56+0];\n"
|
|
" add.ftz.f32 %f85, %f89, %f85;\n"
|
|
" st.shared.f32 [%rd53+0], %f85;\n"
|
|
" ld.shared.f32 %f90, [%rd56+512];\n"
|
|
" add.ftz.f32 %f86, %f90, %f86;\n"
|
|
" st.shared.f32 [%rd53+512], %f86;\n"
|
|
" ld.shared.f32 %f91, [%rd56+1024];\n"
|
|
" add.ftz.f32 %f87, %f91, %f87;\n"
|
|
" st.shared.f32 [%rd53+1024], %f87;\n"
|
|
" ld.shared.f32 %f92, [%rd56+1536];\n"
|
|
" add.ftz.f32 %f88, %f92, %f88;\n"
|
|
" st.shared.f32 [%rd53+1536], %f88;\n"
|
|
"$Lt_1_27650:\n"
|
|
" .loc 16 342 0\n"
|
|
" shr.u32 %r54, %r54, 1;\n"
|
|
" mov.u32 %r57, 0;\n"
|
|
" setp.ne.u32 %p14, %r54, %r57;\n"
|
|
" @%p14 bra $Lt_1_27394;\n"
|
|
"$Lt_1_26882:\n"
|
|
" .loc 16 349 0\n"
|
|
" mov.f32 %f31, %f85;\n"
|
|
" .loc 16 350 0\n"
|
|
" mov.f32 %f30, %f86;\n"
|
|
" .loc 16 351 0\n"
|
|
" mov.f32 %f29, %f87;\n"
|
|
" .loc 16 352 0\n"
|
|
" mov.f32 %f32, %f88;\n"
|
|
" ld.param.s32 %r58, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r59, 0;\n"
|
|
" setp.le.s32 %p15, %r58, %r59;\n"
|
|
" @%p15 bra $Lt_1_28418;\n"
|
|
" .loc 16 356 0\n"
|
|
" mov.f32 %f85, %f9;\n"
|
|
" st.shared.f32 [%rd53+0], %f85;\n"
|
|
" mov.f32 %f86, %f11;\n"
|
|
" st.shared.f32 [%rd53+512], %f86;\n"
|
|
" mov.f32 %f87, %f13;\n"
|
|
" st.shared.f32 [%rd53+1024], %f87;\n"
|
|
" mov.f32 %f88, %f15;\n"
|
|
" st.shared.f32 [%rd53+1536], %f88;\n"
|
|
" mov.f32 %f93, %f17;\n"
|
|
" st.shared.f32 [%rd53+2048], %f93;\n"
|
|
" mov.f32 %f94, %f19;\n"
|
|
" st.shared.f32 [%rd53+2560], %f94;\n"
|
|
" .loc 16 358 0\n"
|
|
" mov.s32 %r60, %r53;\n"
|
|
" @!%p12 bra $Lt_1_28930;\n"
|
|
"$Lt_1_29442:\n"
|
|
" setp.ge.u32 %p16, %r10, %r60;\n"
|
|
" @%p16 bra $Lt_1_29698;\n"
|
|
" .loc 16 361 0\n"
|
|
" add.u32 %r61, %r1, %r60;\n"
|
|
" cvt.u64.u32 %rd57, %r61;\n"
|
|
" mul.wide.u32 %rd58, %r61, 4;\n"
|
|
" add.u64 %rd59, %rd50, %rd58;\n"
|
|
" ld.shared.f32 %f95, [%rd59+0];\n"
|
|
" add.ftz.f32 %f85, %f95, %f85;\n"
|
|
" st.shared.f32 [%rd53+0], %f85;\n"
|
|
" ld.shared.f32 %f96, [%rd59+512];\n"
|
|
" add.ftz.f32 %f86, %f96, %f86;\n"
|
|
" st.shared.f32 [%rd53+512], %f86;\n"
|
|
" ld.shared.f32 %f97, [%rd59+1024];\n"
|
|
" add.ftz.f32 %f87, %f97, %f87;\n"
|
|
" st.shared.f32 [%rd53+1024], %f87;\n"
|
|
" ld.shared.f32 %f98, [%rd59+1536];\n"
|
|
" add.ftz.f32 %f88, %f98, %f88;\n"
|
|
" st.shared.f32 [%rd53+1536], %f88;\n"
|
|
" ld.shared.f32 %f99, [%rd59+2048];\n"
|
|
" add.ftz.f32 %f93, %f99, %f93;\n"
|
|
" st.shared.f32 [%rd53+2048], %f93;\n"
|
|
" ld.shared.f32 %f100, [%rd59+2560];\n"
|
|
" add.ftz.f32 %f94, %f100, %f94;\n"
|
|
" st.shared.f32 [%rd53+2560], %f94;\n"
|
|
"$Lt_1_29698:\n"
|
|
" .loc 16 358 0\n"
|
|
" shr.u32 %r60, %r60, 1;\n"
|
|
" mov.u32 %r62, 0;\n"
|
|
" setp.ne.u32 %p17, %r60, %r62;\n"
|
|
" @%p17 bra $Lt_1_29442;\n"
|
|
"$Lt_1_28930:\n"
|
|
" .loc 16 366 0\n"
|
|
" mov.f32 %f9, %f85;\n"
|
|
" mov.f32 %f11, %f86;\n"
|
|
" mov.f32 %f13, %f87;\n"
|
|
" mov.f32 %f15, %f88;\n"
|
|
" mov.f32 %f17, %f93;\n"
|
|
" mov.f32 %f19, %f94;\n"
|
|
"$Lt_1_28418:\n"
|
|
"$Lt_1_26370:\n"
|
|
" selp.s32 %r63, 1, 0, %p4;\n"
|
|
" mov.s32 %r64, 0;\n"
|
|
" set.eq.u32.s32 %r65, %r10, %r64;\n"
|
|
" neg.s32 %r66, %r65;\n"
|
|
" and.b32 %r67, %r63, %r66;\n"
|
|
" mov.u32 %r68, 0;\n"
|
|
" setp.eq.s32 %p18, %r67, %r68;\n"
|
|
" @%p18 bra $Lt_1_30466;\n"
|
|
" .loc 16 372 0\n"
|
|
" cvt.s64.s32 %rd60, %r13;\n"
|
|
" ld.param.u64 %rd61, [__cudaparm_kernel_pair_fast_engv];\n"
|
|
" mul.wide.s32 %rd62, %r13, 4;\n"
|
|
" add.u64 %rd63, %rd61, %rd62;\n"
|
|
" ld.param.s32 %r69, [__cudaparm_kernel_pair_fast_eflag];\n"
|
|
" mov.u32 %r70, 0;\n"
|
|
" setp.le.s32 %p19, %r69, %r70;\n"
|
|
" @%p19 bra $Lt_1_30978;\n"
|
|
" .loc 16 374 0\n"
|
|
" st.global.f32 [%rd63+0], %f32;\n"
|
|
" .loc 16 375 0\n"
|
|
" cvt.s64.s32 %rd64, %r14;\n"
|
|
" mul.wide.s32 %rd65, %r14, 4;\n"
|
|
" add.u64 %rd63, %rd63, %rd65;\n"
|
|
"$Lt_1_30978:\n"
|
|
" ld.param.s32 %r71, [__cudaparm_kernel_pair_fast_vflag];\n"
|
|
" mov.u32 %r72, 0;\n"
|
|
" setp.le.s32 %p20, %r71, %r72;\n"
|
|
" @%p20 bra $Lt_1_31490;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f101, %f9;\n"
|
|
" st.global.f32 [%rd63+0], %f101;\n"
|
|
" .loc 16 380 0\n"
|
|
" cvt.s64.s32 %rd66, %r14;\n"
|
|
" mul.wide.s32 %rd67, %r14, 4;\n"
|
|
" add.u64 %rd68, %rd67, %rd63;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f102, %f11;\n"
|
|
" st.global.f32 [%rd68+0], %f102;\n"
|
|
" .loc 16 380 0\n"
|
|
" add.u64 %rd69, %rd67, %rd68;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f103, %f13;\n"
|
|
" st.global.f32 [%rd69+0], %f103;\n"
|
|
" .loc 16 380 0\n"
|
|
" add.u64 %rd70, %rd67, %rd69;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f104, %f15;\n"
|
|
" st.global.f32 [%rd70+0], %f104;\n"
|
|
" .loc 16 380 0\n"
|
|
" add.u64 %rd63, %rd67, %rd70;\n"
|
|
" .loc 16 379 0\n"
|
|
" mov.f32 %f105, %f17;\n"
|
|
" st.global.f32 [%rd63+0], %f105;\n"
|
|
" mov.f32 %f106, %f19;\n"
|
|
" add.u64 %rd71, %rd67, %rd63;\n"
|
|
" st.global.f32 [%rd71+0], %f106;\n"
|
|
"$Lt_1_31490:\n"
|
|
" .loc 16 383 0\n"
|
|
" ld.param.u64 %rd72, [__cudaparm_kernel_pair_fast_ans];\n"
|
|
" mul.lo.u64 %rd73, %rd60, 16;\n"
|
|
" add.u64 %rd74, %rd72, %rd73;\n"
|
|
" mov.f32 %f107, %f108;\n"
|
|
" st.global.v4.f32 [%rd74+0], {%f31,%f30,%f29,%f107};\n"
|
|
"$Lt_1_30466:\n"
|
|
" .loc 16 385 0\n"
|
|
" exit;\n"
|
|
"$LDWend_kernel_pair_fast:\n"
|
|
" }\n"
|
|
;
|