File lapack_564f5cd6dd718723530ccdd421a63387e76d5fcdb30f49d1f0425db7bc512296.dll
Function:
    37b69ea40
    37b69ea40slasyf_rook_:
    37b69ea40push r15
    37b69ea42push r14
    37b69ea44push r13
    37b69ea46push r12
    37b69ea48push rbp
    37b69ea49push rdi
    37b69ea4apush rsi
    37b69ea4bpush rbx
    37b69ea4csub rsp, 520
    37b69ea53movups [rsp+400], x6
    37b69ea5bmovups [rsp+416], x7
    37b69ea63movups [rsp+432], x8
    37b69ea6cmovups [rsp+448], x9
    37b69ea75movups [rsp+464], x10
    37b69ea7emovups [rsp+480], x11
    37b69ea87movups [rsp+496], x12
    37b69ea90mov rax, [rsp+632]
    37b69ea98mov r13, [rsp+624]
    37b69eaa0mov r12, [rsp+648]
    37b69eaa8movsxd rdi, [rax]
    37b69eaabxor eax, eax
    37b69eaadtest rdi, rdi
    37b69eab0mov rbx, rcx
    37b69eab3mov r15, rdx
    37b69eab6mov edx, 1
    37b69eabbcmovns rax, rdi
    37b69eabflea rcx, [rip+0xd9585a]
    37b69eac6mov [rsp+608], r8
    37b69eacemov r14, r9
    37b69ead1mov [rsp+144], rax
    37b69ead9not rax
    37b69eadcmov [rsp+152], rax
    37b69eae4mov rax, [rsp+656]
    37b69eaecmovsxd rdi, [rax]
    37b69eaefxor eax, eax
    37b69eaf1test rdi, rdi
    37b69eaf4cmovns rax, rdi
    37b69eaf8mov [rsp+176], rax
    37b69eb00not rax
    37b69eb03mov [rsp+184], rax
    37b69eb0bmov rax, [rsp+664]
    37b69eb13mov [rax], 0
    37b69eb19call slamch_
    37b69eb1emov r9l, 1
    37b69eb24mov r8l, 1
    37b69eb2amov rcx, rbx
    37b69eb2dlea rdx, [rip+0xd957ed]
    37b69eb34movaps x11, x0
    37b69eb38call lsame_
    37b69eb3dtest eax, eax
    37b69eb3fje 0x37b69ef10
    37b69eb45mov rax, [rsp+608]
    37b69eb4dmov ebp, [r15]
    37b69eb50mov edi, [rax]
    37b69eb52mov eax, ebp
    37b69eb54mov [rsp+372], ebp
    37b69eb5bsub eax, edi
    37b69eb5dadd eax, 1
    37b69eb60cmp ebp, eax
    37b69eb62jg 0x37b69eb6c
    37b69eb64cmp ebp, edi
    37b69eb66jg 0x37b6a0d19
    37b69eb6c
    37b69eb6cloc_37b69eb6c:
    37b69eb6clea rax, [rsp+372]
    37b69eb74mov [rsp+616], r14
    37b69eb7cmov edx, ebp
    37b69eb7epxor x7, x7
    37b69eb82mov [rsp+192], rax
    37b69eb8alea rax, [rsp+396]
    37b69eb92movss x6, [rip+0xd957c6]
    37b69eb9amov [rsp+136], rax
    37b69eba2lea rax, [rip+0xd95783]
    37b69eba9mov [rsp+256], rax
    37b69ebb1lea rax, [rip+0xd95780]
    37b69ebb8mov [rsp+264], rax
    37b69ebc0
    37b69ebc0loc_37b69ebc0:
    37b69ebc0test ebp, ebp
    37b69ebc2jle 0x37b69ed6c
    37b69ebc8movsxd rsi, edi
    37b69ebcbmov rax, [rsp+184]
    37b69ebd3imul rsi, [rsp+176]
    37b69ebdcmov [rsp+384], ebp
    37b69ebe3movsxd rbp, ebp
    37b69ebe6mov rcx, [rsp+192]
    37b69ebeeimul rbp, [rsp+144]
    37b69ebf7add rax, rsi
    37b69ebfalea r9, [r12+4*rax+4]
    37b69ebffmov [rsp+216], rax
    37b69ec07mov rax, [rsp+152]
    37b69ec0fmov [rsp+200], r9
    37b69ec17lea rax, [rax+rbp+1]
    37b69ec1clea rdx, [r13+4*rax]
    37b69ec21lea rax, [rip+0xd956fc]
    37b69ec28mov [rsp+32], rax
    37b69ec2dmov r8, rax
    37b69ec30call scopy_
    37b69ec35mov eax, [r15]
    37b69ec38mov ebx, [rsp+372]
    37b69ec3fcmp eax, ebx
    37b69ec41jg 0x37b6a0828
    37b69ec47
    37b69ec47loc_37b69ec47:
    37b69ec47mov rax, [rsp+216]
    37b69ec4fmovsxd rbp, ebx
    37b69ec52add rax, rbp
    37b69ec55movss x9, [r12+4*rax]
    37b69ec5bcmp ebx, 1
    37b69ec5ejg 0x37b69ff80
    37b69ec64ucomiss x9, x7
    37b69ec68jp 0x37b69ec70
    37b69ec6aje 0x37b69ff10
    37b69ec70
    37b69ec70loc_37b69ec70:
    37b69ec70mov rax, [rsp+144]
    37b69ec78mov [rsp+380], ebx
    37b69ec7fimul rax, rbp
    37b69ec83add rax, [rsp+152]
    37b69ec8bmov [rsp+208], rax
    37b69ec93
    37b69ec93loc_37b69ec93:
    37b69ec93lea rax, [rip+0xd9568a]
    37b69ec9amov rdx, [rsp+200]
    37b69eca2mov rcx, [rsp+192]
    37b69ecaalea r8, [rip+0xd95673]
    37b69ecb1mov [rsp+32], rax
    37b69ecb6mov rax, [rsp+208]
    37b69ecbelea r9, [r13+4*rax+4]
    37b69ecc3call scopy_
    37b69ecc8movsxd rsi, [rsp+372]
    37b69ecd0mov ebx, [rsp+380]
    37b69ecd7mov rax, rsi
    37b69ecdalea ebp, [rsi-1]
    37b69ecddcmp esi, 1
    37b69ece0jle 0x37b69ed30
    37b69ece2mov rcx, [rsp+144]
    37b69eceaimul rcx, rsi
    37b69eceeadd rcx, [rsp+152]
    37b69ecf6lea r8, [rcx+rsi]
    37b69ecfamovss x0, [r13+4*r8]
    37b69ed01movaps x1, x0
    37b69ed04andps x1, [rip+0xd95645]
    37b69ed0bcomiss x1, x11
    37b69ed0fjae 0x37b6a0c90
    37b69ed15pxor x1, x1
    37b69ed19ucomiss x0, x1
    37b69ed1cjp 0x37b6a0c27
    37b69ed22jne 0x37b6a0c27
    37b69ed28nop [rax+rax]
    37b69ed30
    37b69ed30loc_37b69ed30:
    37b69ed30mov rax, [rsp+640]
    37b69ed38mov [rax+4*rsi-4], ebx
    37b69ed3cmov rax, [rsp+608]
    37b69ed44mov ecx, [rax]
    37b69ed46
    37b69ed46loc_37b69ed46:
    37b69ed46mov edx, [r15]
    37b69ed49lea edi, [rcx+rbp]
    37b69ed4cmov [rsp+372], ebp
    37b69ed53mov eax, edx
    37b69ed55sub edi, edx
    37b69ed57sub eax, ecx
    37b69ed59add eax, 1
    37b69ed5ccmp eax, ebp
    37b69ed5ejl 0x37b69ebc0
    37b69ed64cmp ecx, edx
    37b69ed66jge 0x37b69ebc0
    37b69ed6c
    37b69ed6cloc_37b69ed6c:
    37b69ed6cmov r14, [rsp+616]
    37b69ed74sub edx, ebp
    37b69ed76
    37b69ed76loc_37b69ed76:
    37b69ed76mov rax, [rsp+632]
    37b69ed7eadd edi, 1
    37b69ed81mov [rsp+396], edx
    37b69ed88lea edx, [rbp+1]
    37b69ed8bmovsxd rdi, edi
    37b69ed8emovsxd rdx, edx
    37b69ed91mov [rsp+88], r13
    37b69ed96mov r9, [rsp+192]
    37b69ed9emov [rsp+96], rax
    37b69eda3mov rax, [rsp+264]
    37b69edablea rcx, [rip+0xd95593]
    37b69edb2lea r8, [rip+0xd95583]
    37b69edb9imul rdi, [rsp+176]
    37b69edc2mov [rsp+120], 9
    37b69edcbimul rdx, [rsp+144]
    37b69edd4mov [rsp+80], rax
    37b69edd9mov rax, [rsp+656]
    37b69ede1mov [rsp+112], 12
    37b69edeamov [rsp+104], 5
    37b69edf3mov [rsp+72], rax
    37b69edf8mov rax, [rsp+184]
    37b69ee00lea rax, [rax+rdi+1]
    37b69ee05lea rax, [r12+4*rax]
    37b69ee09mov [rsp+64], rax
    37b69ee0emov rax, [rsp+632]
    37b69ee16mov [rsp+56], rax
    37b69ee1bmov rax, [rsp+152]
    37b69ee23lea rax, [rax+rdx+1]
    37b69ee28mov rdx, [rsp+256]
    37b69ee30lea rax, [r13+4*rax]
    37b69ee35mov [rsp+48], rax
    37b69ee3alea rax, [rip+0xd954e7]
    37b69ee41mov [rsp+40], rax
    37b69ee46mov rax, [rsp+136]
    37b69ee4emov [rsp+32], rax
    37b69ee53call sgemmtr_
    37b69ee58mov esi, [rsp+372]
    37b69ee5fmov eax, [r15]
    37b69ee62mov r12, [rsp+632]
    37b69ee6amov [rsp+160], esi
    37b69ee71lea ebx, [rsi+1]
    37b69ee74mov ecx, eax
    37b69ee76mov rsi, [rsp+640]
    37b69ee7enop
    37b69ee80
    37b69ee80loc_37b69ee80:
    37b69ee80movsxd r8, ebx
    37b69ee83lea edi, [rbx+1]
    37b69ee86mov edx, [rsi+4*r8-4]
    37b69ee8btest edx, edx
    37b69ee8djs 0x37b69fe70
    37b69ee93mov [rsp+392], edi
    37b69ee9acmp edx, ebx
    37b69ee9cje 0x37b69fef7
    37b69eea2cmp ecx, edi
    37b69eea4jge 0x37b6a0a40
    37b69eeaa
    37b69eeaaloc_37b69eeaa:
    37b69eeaamov esi, [rsp+160]
    37b69eeb1mov eax, ecx
    37b69eeb3sub eax, esi
    37b69eeb5mov ebx, eax
    37b69eeb7
    37b69eeb7loc_37b69eeb7:
    37b69eeb7mov [r14], ebx
    37b69eebamovups x6, [rsp+400]
    37b69eec2movups x7, [rsp+416]
    37b69eecamovups x8, [rsp+432]
    37b69eed3movups x9, [rsp+448]
    37b69eedcmovups x10, [rsp+464]
    37b69eee5movups x11, [rsp+480]
    37b69eeeemovups x12, [rsp+496]
    37b69eef7add rsp, 520
    37b69eefepop rbx
    37b69eeffpop rsi
    37b69ef00pop rdi
    37b69ef01pop rbp
    37b69ef02pop r12
    37b69ef04pop r13
    37b69ef06pop r14
    37b69ef08pop r15
    37b69ef0aret
    37b69ef0b
    37b69ef0b 0f 1f 44 00 00 ..D..
    37b69ef10
    37b69ef10loc_37b69ef10:
    37b69ef10lea rax, [rsp+396]
    37b69ef18mov ebp, [r15]
    37b69ef1bmov [rsp+616], r14
    37b69ef23mov edi, 1
    37b69ef28mov [rsp+136], rax
    37b69ef30lea rax, [rip+0xd953f5]
    37b69ef37movss x7, [rip+0xd95421]
    37b69ef3fpxor x6, x6
    37b69ef43mov [rsp+256], rax
    37b69ef4blea rax, [rip+0xd953e6]
    37b69ef52mov edx, ebp
    37b69ef54mov [rsp+264], rax
    37b69ef5clea rax, [rsp+392]
    37b69ef64mov [rsp+336], rax
    37b69ef6cnop [rax]
    37b69ef70
    37b69ef70loc_37b69ef70:
    37b69ef70mov rax, [rsp+608]
    37b69ef78mov [rsp+372], edi
    37b69ef7fmov eax, [rax]
    37b69ef81cmp eax, edi
    37b69ef83jg 0x37b69ef8d
    37b69ef85cmp eax, edx
    37b69ef87jl 0x37b69fba0
    37b69ef8d
    37b69ef8dloc_37b69ef8d:
    37b69ef8dcmp edi, edx
    37b69ef8fjg 0x37b69fba0
    37b69ef95mov rax, [rsp+144]
    37b69ef9dmovsxd rbp, edi
    37b69efa0sub edx, edi
    37b69efa2mov rsi, [rsp+152]
    37b69efaamov rcx, [rsp+184]
    37b69efb2add edx, 1
    37b69efb5lea r14, [rbp-1]
    37b69efb9mov [rsp+312], rbp
    37b69efc1imul rax, rbp
    37b69efc5mov [rsp+396], edx
    37b69efccmov [rsp+208], r14
    37b69efd4mov [rsp+384], edi
    37b69efdbmov [rsp+304], rax
    37b69efe3add rax, rsi
    37b69efe6mov rsi, [rsp+176]
    37b69efeemov [rsp+168], rax
    37b69eff6add rax, rbp
    37b69eff9imul rsi, rbp
    37b69effdmov [rsp+272], rax
    37b69f005lea rax, [r13+4*rax]
    37b69f00amov rdx, rax
    37b69f00dmov [rsp+280], rax
    37b69f015lea rbx, [rsi+rcx]
    37b69f019lea rcx, [rbx+rbp]
    37b69f01dmov [rsp+248], rbx
    37b69f025mov [rsp+224], rcx
    37b69f02dlea rbx, [r12+4*rcx]
    37b69f031lea rcx, [rip+0xd952ec]
    37b69f038mov [rsp+32], rcx
    37b69f03dmov r8, rcx
    37b69f040mov r9, rbx
    37b69f043mov rcx, [rsp+136]
    37b69f04bmov [rsp+232], rbx
    37b69f053call scopy_
    37b69f058cmp edi, 1
    37b69f05bjle 0x37b69f0f8
    37b69f061mov eax, [r15]
    37b69f064mov [rsp+72], rbx
    37b69f069lea rdx, [rip+0xd952b4]
    37b69f070lea r9, [rip+0xd952b1]
    37b69f077mov rbx, [rsp+264]
    37b69f07fmov [rsp+80], rdx
    37b69f084sub eax, edi
    37b69f086mov r8, [rsp+136]
    37b69f08emov [rsp+88], 12
    37b69f097add eax, 1
    37b69f09amov [rsp+64], rbx
    37b69f09fmov rbx, [rsp+656]
    37b69f0a7mov [rsp+392], eax
    37b69f0aelea eax, [rdi-1]
    37b69f0b1mov rcx, [rsp+256]
    37b69f0b9mov [rsp+396], eax
    37b69f0c0lea rax, [4*r14]
    37b69f0c8mov [rsp+56], rbx
    37b69f0cdmov rbx, [rsp+632]
    37b69f0d5lea rdx, [r12+rax]
    37b69f0d9add rax, r13
    37b69f0dcmov [rsp+48], rdx
    37b69f0e1mov rdx, [rsp+336]
    37b69f0e9mov [rsp+40], rbx
    37b69f0eemov [rsp+32], rax
    37b69f0f3call sgemv_
    37b69f0f8
    37b69f0f8loc_37b69f0f8:
    37b69f0f8mov rax, [rsp+224]
    37b69f100lea ebx, [rdi+1]
    37b69f103mov [rsp+240], ebx
    37b69f10amovsxd rbx, ebx
    37b69f10dmovss x8, [r12+4*rax]
    37b69f113mov eax, [r15]
    37b69f116mov [rsp+216], rbx
    37b69f11ecmp eax, edi
    37b69f120jg 0x37b69f1a0
    37b69f122ucomiss x8, x6
    37b69f126jp 0x37b69f12e
    37b69f128je 0x37b69f220
    37b69f12e
    37b69f12eloc_37b69f12e:
    37b69f12emov r14l, edi
    37b69f131
    37b69f131loc_37b69f131:
    37b69f131sub eax, edi
    37b69f133mov rdx, [rsp+232]
    37b69f13bmov r9, [rsp+280]
    37b69f143add eax, 1
    37b69f146mov rcx, [rsp+136]
    37b69f14emov [rsp+396], eax
    37b69f155lea rax, [rip+0xd951c8]
    37b69f15cmov [rsp+32], rax
    37b69f161mov r8, rax
    37b69f164call scopy_
    37b69f169mov edx, [r15]
    37b69f16ccmp edx, edi
    37b69f16ejg 0x37b69fd78
    37b69f174
    37b69f174loc_37b69f174:
    37b69f174mov edi, [rsp+240]
    37b69f17bnop [rax+rax]
    37b69f180
    37b69f180loc_37b69f180:
    37b69f180mov rax, [rsp+640]
    37b69f188mov rsi, [rsp+208]
    37b69f190mov [rax+4*rsi], r14l
    37b69f194jmp 0x37b69ef70
    37b69f199
    37b69f199 0f 1f 80 00 00 00 00 .......
    37b69f1a0
    37b69f1a0loc_37b69f1a0:
    37b69f1a0sub eax, edi
    37b69f1a2mov rbx, [rsp+216]
    37b69f1aamov rcx, [rsp+136]
    37b69f1b2lea r8, [rip+0xd9516b]
    37b69f1b9mov [rsp+396], eax
    37b69f1c0mov rax, [rsp+248]
    37b69f1c8andps x8, [rip+0xd95180]
    37b69f1d0lea r9, [rax+rbx]
    37b69f1d4lea rdx, [r12+4*r9]
    37b69f1d8mov [rsp+192], r9
    37b69f1e0call isamax_
    37b69f1e5movaps x0, x8
    37b69f1e9mov r9, [rsp+192]
    37b69f1f1lea r14l, [rax+rdi]
    37b69f1f5mov rax, [rsp+248]
    37b69f1fdmovsxd rbx, r14l
    37b69f200add rax, rbx
    37b69f203movss x9, [r12+4*rax]
    37b69f209andps x9, [rip+0xd9513f]
    37b69f211maxss x0, x9
    37b69f216ucomiss x0, x6
    37b69f219jp 0x37b69f280
    37b69f21bjne 0x37b69f280
    37b69f21dmov eax, [r15]
    37b69f220
    37b69f220loc_37b69f220:
    37b69f220mov rsi, [rsp+664]
    37b69f228mov edx, [rsi]
    37b69f22atest edx, edx
    37b69f22cjne 0x37b69f230
    37b69f22emov [rsi], edi
    37b69f230
    37b69f230loc_37b69f230:
    37b69f230sub eax, edi
    37b69f232mov rdx, [rsp+232]
    37b69f23amov r14l, edi
    37b69f23dmov r9, [rsp+280]
    37b69f245add eax, 1
    37b69f248mov rcx, [rsp+136]
    37b69f250mov [rsp+396], eax
    37b69f257lea rax, [rip+0xd950c6]
    37b69f25emov [rsp+32], rax
    37b69f263mov r8, rax
    37b69f266call scopy_
    37b69f26bmov edi, [rsp+240]
    37b69f272mov edx, [r15]
    37b69f275jmp 0x37b69f180
    37b69f27a
    37b69f27a 66 0f 1f 44 00 00 f..D..
    37b69f280
    37b69f280loc_37b69f280:
    37b69f280movaps x0, x9
    37b69f284mulss x0, x7
    37b69f288comiss x0, x8
    37b69f28cjbe 0x37b6a0c80
    37b69f292mov rax, [rsp+184]
    37b69f29aadd rsi, [rsp+176]
    37b69f2a2mov [rsp+352], r9
    37b69f2aamov [rsp+360], rbp
    37b69f2b2lea rcx, [rsi+rax]
    37b69f2b6lea rax, [rcx+rbp]
    37b69f2bamov [rsp+200], rcx
    37b69f2c2lea rax, [r12+4*rax]
    37b69f2c6mov [rsp+192], rax
    37b69f2celea rax, [r13+4*rbp-4]
    37b69f2d3mov [rsp+344], rax
    37b69f2dbmov eax, [rsp+160]
    37b69f2e2mov [rsp+332], eax
    37b69f2e9lea eax, [rdi-1]
    37b69f2ecmov [rsp+328], eax
    37b69f2f3
    37b69f2f3loc_37b69f2f3:
    37b69f2f3mov rax, [rsp+168]
    37b69f2fbmov r9, [rsp+192]
    37b69f303mov ebp, r14l
    37b69f306mov r8, [rsp+632]
    37b69f30emov rcx, [rsp+136]
    37b69f316sub ebp, edi
    37b69f318add rax, rbx
    37b69f31bmov [rsp+396], ebp
    37b69f322lea rdx, [r13+4*rax]
    37b69f327mov [rsp+296], rax
    37b69f32flea rax, [rip+0xd94fee]
    37b69f336mov [rsp+32], rax
    37b69f33bmov [rsp+288], rdx
    37b69f343call scopy_
    37b69f348mov esi, [r15]
    37b69f34bmov rdx, [rsp+144]
    37b69f353mov rcx, [rsp+136]
    37b69f35bsub esi, r14l
    37b69f35eimul rdx, rbx
    37b69f362lea eax, [rsi+1]
    37b69f365mov rsi, [rsp+200]
    37b69f36dmov [rsp+396], eax
    37b69f374mov rax, [rsp+152]
    37b69f37cadd rsi, rbx
    37b69f37fadd rax, rdx
    37b69f382lea r9, [r12+4*rsi]
    37b69f386add rax, rbx
    37b69f389lea rdx, [r13+4*rax]
    37b69f38elea rax, [rip+0xd94f8f]
    37b69f395mov [rsp+32], rax
    37b69f39amov r8, rax
    37b69f39dmov [rsp+320], rdx
    37b69f3a5call scopy_
    37b69f3aacmp edi, 1
    37b69f3adjle 0x37b69f454
    37b69f3b3mov edx, [r15]
    37b69f3b6mov r8, [rsp+136]
    37b69f3bemov [rsp+88], 12
    37b69f3c7lea r9, [rip+0xd94f5a]
    37b69f3cemov rcx, [rsp+256]
    37b69f3d6sub edx, edi
    37b69f3d8lea eax, [rdx+1]
    37b69f3dbmov edx, [rsp+328]
    37b69f3e2mov [rsp+392], eax
    37b69f3e9lea rax, [rip+0xd94f34]
    37b69f3f0mov [rsp+80], rax
    37b69f3f5mov rax, [rsp+192]
    37b69f3fdmov [rsp+396], edx
    37b69f404mov rdx, [rsp+336]
    37b69f40cmov [rsp+72], rax
    37b69f411mov rax, [rsp+264]
    37b69f419mov [rsp+64], rax
    37b69f41emov rax, [rsp+656]
    37b69f426mov [rsp+56], rax
    37b69f42blea rax, [r12+4*rbx-4]
    37b69f430mov [rsp+48], rax
    37b69f435mov rax, [rsp+632]
    37b69f43dmov [rsp+40], rax
    37b69f442mov rax, [rsp+344]
    37b69f44amov [rsp+32], rax
    37b69f44fcall sgemv_
    37b69f454
    37b69f454loc_37b69f454:
    37b69f454cmp edi, r14l
    37b69f457jne 0x37b69f4a8
    37b69f459mov eax, [r15]
    37b69f45ccmp eax, edi
    37b69f45ejg 0x37b69fa40
    37b69f464sub eax, edi
    37b69f466mov r9, [rsp+232]
    37b69f46emov rdx, [rsp+192]
    37b69f476add eax, 1
    37b69f479mov rcx, [rsp+136]
    37b69f481mov [rsp+396], eax
    37b69f488lea rax, [rip+0xd94e95]
    37b69f48fmov [rsp+32], rax
    37b69f494mov r8, rax
    37b69f497call scopy_
    37b69f49c
    37b69f49cloc_37b69f49c:
    37b69f49cmov eax, [r15]
    37b69f49fjmp 0x37b69f131
    37b69f4a4
    37b69f4a4 0f 1f 40 00 ..@.
    37b69f4a8
    37b69f4a8loc_37b69f4a8:
    37b69f4a8mov rcx, [rsp+136]
    37b69f4b0mov rdx, [rsp+192]
    37b69f4b8lea r8, [rip+0xd94e65]
    37b69f4bfmov [rsp+396], ebp
    37b69f4c6call isamax_
    37b69f4cbmov ecx, [rsp+328]
    37b69f4d2add eax, ecx
    37b69f4d4mov [rsp+160], eax
    37b69f4dbcdqe
    37b69f4ddadd rax, [rsp+200]
    37b69f4e5movss x8, [r12+4*rax]
    37b69f4ebmov eax, [r15]
    37b69f4eeandps x8, [rip+0xd94e5a]
    37b69f4f6cmp eax, r14l
    37b69f4f9jg 0x37b69fab8
    37b69f4ff
    37b69f4ffloc_37b69f4ff:
    37b69f4ffmovaps x1, x8
    37b69f503movss x0, [r12+4*rsi]
    37b69f509andps x0, [rip+0xd94e40]
    37b69f510mulss x1, x7
    37b69f514comiss x1, x0
    37b69f517jbe 0x37b69f7c0
    37b69f51dmov eax, [rsp+384]
    37b69f524cmp eax, [rsp+160]
    37b69f52bje 0x37b69f598
    37b69f52dcomiss x9, x8
    37b69f531jae 0x37b69f598
    37b69f533mov eax, [r15]
    37b69f536mov r9, [rsp+232]
    37b69f53emov [rsp+384], r14l
    37b69f546movaps x9, x8
    37b69f54amov rdx, [rsp+192]
    37b69f552mov rcx, [rsp+136]
    37b69f55asub eax, edi
    37b69f55cadd eax, 1
    37b69f55fmov [rsp+396], eax
    37b69f566lea rax, [rip+0xd94db7]
    37b69f56dmov [rsp+32], rax
    37b69f572mov r8, rax
    37b69f575call scopy_
    37b69f57amovsxd rbx, [rsp+160]
    37b69f582mov [rsp+332], ebx
    37b69f589mov r14l, ebx
    37b69f58cjmp 0x37b69f2f3
    37b69f591
    37b69f591 0f 1f 80 00 00 00 00 .......
    37b69f598
    37b69f598loc_37b69f598:
    37b69f598mov edx, [rsp+240]
    37b69f59fmov esi, ebp
    37b69f5a1mov [rsp+380], r14l
    37b69f5a9mov r9, [rsp+352]
    37b69f5b1mov rcx, [rsp+200]
    37b69f5b9mov [rsp+376], edx
    37b69f5c0mov rbp, [rsp+360]
    37b69f5c8cmp eax, edi
    37b69f5caje 0x37b6a0aa8
    37b69f5d0mov edx, eax
    37b69f5d2movsxd rsi, eax
    37b69f5d5mov [rsp+192], eax
    37b69f5dcmov rax, [rsp+632]
    37b69f5e4sub edx, edi
    37b69f5e6mov rbp, [rsp+136]
    37b69f5eelea r8, [rip+0xd94d2f]
    37b69f5f5mov [rsp+396], edx
    37b69f5fcmov rdx, [rsp+168]
    37b69f604mov [rsp+32], rax
    37b69f609mov rcx, rbp
    37b69f60cadd rdx, rsi
    37b69f60flea rdi, [r13+4*rdx]
    37b69f614mov rdx, [rsp+280]
    37b69f61cmov r9, rdi
    37b69f61fcall scopy_
    37b69f624mov eax, [rsp+192]
    37b69f62bmov edx, [r15]
    37b69f62elea rcx, [rip+0xd94cef]
    37b69f635mov [rsp+32], rcx
    37b69f63amov r8, rcx
    37b69f63dmov rcx, rbp
    37b69f640sub edx, eax
    37b69f642mov [rsp+200], eax
    37b69f649add edx, 1
    37b69f64cmov [rsp+396], edx
    37b69f653mov rdx, [rsp+144]
    37b69f65bimul rdx, rsi
    37b69f65fadd rdx, [rsp+152]
    37b69f667add rdx, rsi
    37b69f66alea rsi, [4*rsi-4]
    37b69f672lea r9, [r13+4*rdx]
    37b69f677mov rdx, rdi
    37b69f67acall scopy_
    37b69f67fmov r11, [rsp+632]
    37b69f687lea r9, [r13+rsi]
    37b69f68cmov rcx, [rsp+208]
    37b69f694mov [rsp+32], r11
    37b69f699lea rdx, [r13+4*rcx]
    37b69f69emov r8, r11
    37b69f6a1lea rcx, [rsp+372]
    37b69f6a9call sswap_
    37b69f6aemov r11, [rsp+656]
    37b69f6b6lea r9, [r12+rsi]
    37b69f6bamovsxd rdi, [rsp+372]
    37b69f6c2mov [rsp+32], r11
    37b69f6c7lea rcx, [rdi-1]
    37b69f6cbmov r8, r11
    37b69f6celea rdx, [r12+4*rcx]
    37b69f6d2mov [rsp+208], rcx
    37b69f6dalea rcx, [rsp+376]
    37b69f6e2mov [rsp+168], edi
    37b69f6e9mov [rsp+312], rdi
    37b69f6f1mov [rsp+192], rcx
    37b69f6f9call sswap_
    37b69f6femov rcx, [rsp+176]
    37b69f706mov rax, rdi
    37b69f709mov r11, [rsp+184]
    37b69f711movsxd r10, [rsp+168]
    37b69f719mov rsi, [rsp+152]
    37b69f721imul rcx, rdi
    37b69f725movsxd rbp, [rsp+376]
    37b69f72dlea r9l, [r10+1]
    37b69f731movsxd r8, r9l
    37b69f734mov [rsp+240], r9l
    37b69f73cadd r11, rcx
    37b69f73fmov [rsp+216], r8
    37b69f747mov [rsp+248], r11
    37b69f74fadd r11, rdi
    37b69f752mov rdi, [rsp+144]
    37b69f75amov [rsp+224], r11
    37b69f762mov rdx, rdi
    37b69f765imul rdx, rax
    37b69f769lea r9, [rsi+rdx]
    37b69f76dadd rdx, rdi
    37b69f770add rax, r9
    37b69f773cmp ebp, r14l
    37b69f776mov [rsp+168], r9
    37b69f77elea r11, [rdx+rsi]
    37b69f782mov [rsp+272], rax
    37b69f78amov eax, [rsp+200]
    37b69f791jne 0x37b69fe20
    37b69f797mov rdi, [rsp+248]
    37b69f79fmov rsi, [rsp+216]
    37b69f7a7add rcx, [rsp+176]
    37b69f7afadd rcx, [rsp+184]
    37b69f7b7lea r9, [rdi+rsi]
    37b69f7bbjmp 0x37b69f9a0
    37b69f7c0
    37b69f7c0loc_37b69f7c0:
    37b69f7c0mov eax, [r15]
    37b69f7c3mov r9, [rsp+232]
    37b69f7cbmov [rsp+380], r14l
    37b69f7d3mov esi, ebp
    37b69f7d5mov rdx, [rsp+192]
    37b69f7ddmov rcx, [rsp+136]
    37b69f7e5sub eax, edi
    37b69f7e7add eax, 1
    37b69f7eamov [rsp+396], eax
    37b69f7f1lea rax, [rip+0xd94b2c]
    37b69f7f8mov [rsp+32], rax
    37b69f7fdmov r8, rax
    37b69f800call scopy_
    37b69f805mov [rsp+376], edi
    37b69f80ccmp edi, r14l
    37b69f80fje 0x37b69f49c
    37b69f815mov rcx, [rsp+152]
    37b69f81dmov rax, [rsp+304]
    37b69f825mov r10l, 1
    37b69f82badd rax, [rsp+144]
    37b69f833mov rbp, [rsp+208]
    37b69f83blea r11, [rax+rcx]
    37b69f83flea rcx, [rsp+376]
    37b69f847mov rdx, [rsp+272]
    37b69f84flea rax, [rbx+r11]
    37b69f853mov [rsp+192], rcx
    37b69f85blea r9, [r13+4*rax]
    37b69f860mov rax, [rsp+168]
    37b69f868
    37b69f868loc_37b69f868:
    37b69f868sub esi, 1
    37b69f86bmovss x0, [r13+4*rdx]
    37b69f872mov rcx, [rsp+296]
    37b69f87alea r8, [rip+0xd94aa3]
    37b69f881mov [rsp+396], esi
    37b69f888mov rsi, [rsp+216]
    37b69f890lea rbx, [4*rbx-4]
    37b69f898movss [r13+4*rcx], x0
    37b69f89fmov rcx, [rsp+136]
    37b69f8a7add rax, rsi
    37b69f8aamov [rsp+232], r11
    37b69f8b2lea rdx, [r13+4*rax]
    37b69f8b7mov rax, [rsp+632]
    37b69f8bfmov [rsp+200], r10l
    37b69f8c7mov [rsp+32], rax
    37b69f8cccall scopy_
    37b69f8d1mov eax, [r15]
    37b69f8d4mov r9, [rsp+320]
    37b69f8dcmov rdx, [rsp+288]
    37b69f8e4mov rcx, [rsp+136]
    37b69f8ecsub eax, r14l
    37b69f8efadd eax, 1
    37b69f8f2mov [rsp+396], eax
    37b69f8f9lea rax, [rip+0xd94a24]
    37b69f900mov [rsp+32], rax
    37b69f905mov r8, rax
    37b69f908call scopy_
    37b69f90dlea rdx, [r13+4*rbp]
    37b69f912lea r9, [r13+rbx]
    37b69f917mov rax, [rsp+632]
    37b69f91fmov rbp, [rsp+192]
    37b69f927mov [rsp+32], rax
    37b69f92cmov r8, rax
    37b69f92fmov rcx, rbp
    37b69f932call sswap_
    37b69f937movsxd rax, [rsp+376]
    37b69f93flea r9, [r12+rbx]
    37b69f943mov rcx, rbp
    37b69f946lea rdx, [r12+4*rax-4]
    37b69f94bmov rax, [rsp+656]
    37b69f953mov [rsp+32], rax
    37b69f958mov r8, rax
    37b69f95bcall sswap_
    37b69f960cmp [rsp+200], 1
    37b69f968je 0x37b6a0ce0
    37b69f96emov rbx, [rsp+248]
    37b69f976mov rcx, [rsp+176]
    37b69f97emovsxd r10, edi
    37b69f981mov eax, [rsp+384]
    37b69f988mov r11, [rsp+232]
    37b69f990imul rcx, rsi
    37b69f994lea r9, [rsi+rbx]
    37b69f998add rcx, [rsp+184]
    37b69f9a0
    37b69f9a0loc_37b69f9a0:
    37b69f9a0mov rdi, [rsp+216]
    37b69f9a8mov edx, [r15]
    37b69f9abmovss x2, [r12+4*r9]
    37b69f9b1lea r8, [rdi+rcx]
    37b69f9b5mov rdi, [rsp+224]
    37b69f9bdmovss x9, [r12+4*r8]
    37b69f9c3lea r8l, [rdx-1]
    37b69f9c7movss x10, [r12+4*rdi]
    37b69f9cdlea edi, [r10+2]
    37b69f9d1cmp r8l, r10l
    37b69f9d4jg 0x37b69fad0
    37b69f9da
    37b69f9daloc_37b69f9da:
    37b69f9damov rsi, [rsp+272]
    37b69f9e2mov rcx, [rsp+168]
    37b69f9eaneg eax
    37b69f9ecmov r10l, r14l
    37b69f9efmov rbx, [rsp+208]
    37b69f9f7neg r10l
    37b69f9famovss [r13+4*rsi], x10
    37b69fa01mov rsi, [rsp+216]
    37b69fa09add rcx, rsi
    37b69fa0cadd r11, rsi
    37b69fa0fmov rsi, [rsp+640]
    37b69fa17movss [r13+4*rcx], x2
    37b69fa1emov [rsi+4*rbx], eax
    37b69fa21mov rax, rsi
    37b69fa24mov rsi, [rsp+312]
    37b69fa2cmovss [r13+4*r11], x9
    37b69fa33mov [rax+4*rsi], r10l
    37b69fa37jmp 0x37b69ef70
    37b69fa3c
    37b69fa3c 0f 1f 40 00 ..@.
    37b69fa40
    37b69fa40loc_37b69fa40:
    37b69fa40pxor x10, x10
    37b69fa45
    37b69fa45loc_37b69fa45:
    37b69fa45sub eax, r14l
    37b69fa48mov rcx, [rsp+136]
    37b69fa50lea r8, [rip+0xd948cd]
    37b69fa57mov [rsp+396], eax
    37b69fa5elea eax, [r14+1]
    37b69fa62cdqe
    37b69fa64add rax, [rsp+200]
    37b69fa6clea rdx, [r12+4*rax]
    37b69fa70call isamax_
    37b69fa75lea edx, [rax+r14]
    37b69fa79movsxd rax, edx
    37b69fa7cadd rax, [rsp+200]
    37b69fa84movss x8, [r12+4*rax]
    37b69fa8aandps x8, [rip+0xd948be]
    37b69fa92mov eax, [rsp+332]
    37b69fa99comiss x8, x10
    37b69fa9dmaxss x8, x10
    37b69faa2cmova eax, edx
    37b69faa5mov [rsp+160], eax
    37b69faacjmp 0x37b69f4ff
    37b69fab1
    37b69fab1 0f 1f 80 00 00 00 00 .......
    37b69fab8
    37b69fab8loc_37b69fab8:
    37b69fab8mov ecx, [rsp+160]
    37b69fabfmovaps x10, x8
    37b69fac3mov [rsp+332], ecx
    37b69facajmp 0x37b69fa45
    37b69facf
    37b69facf 90 .
    37b69fad0
    37b69fad0loc_37b69fad0:
    37b69fad0cmp edx, edi
    37b69fad2jl 0x37b69f9da
    37b69fad8movaps x5, x9
    37b69fadcmovaps x4, x10
    37b69fae0lea r8l, [rdx-2]
    37b69fae4add rcx, r10
    37b69fae7divss x5, x2
    37b69faebmovss x3, [rip+0xd94871]
    37b69faf3sub r8l, r10l
    37b69faf6lea rsi, [r12+4*rcx]
    37b69fafalea r9, [4*r8+4]
    37b69fb02mov rcx, [rsp+168]
    37b69fb0amov r8, [rsp+248]
    37b69fb12add rcx, r10
    37b69fb15add r8, r10
    37b69fb18lea rbx, [r12+4*r8]
    37b69fb1clea r8, [r13+4*rcx]
    37b69fb21lea rcx, [r11+r10]
    37b69fb25lea rbp, [r13+4*rcx]
    37b69fb2axor ecx, ecx
    37b69fb2cdivss x4, x2
    37b69fb30movaps x1, x5
    37b69fb33mulss x1, x4
    37b69fb37subss x1, [rip+0xd94825]
    37b69fb3fdivss x3, x1
    37b69fb43nop
    37b69fb45nop [rax+rax]
    37b69fb50
    37b69fb50loc_37b69fb50:
    37b69fb50movss x8, [rbx+rcx+8]
    37b69fb57movss x0, [rsi+rcx+8]
    37b69fb5dmovaps x1, x8
    37b69fb61mulss x1, x5
    37b69fb65subss x1, x0
    37b69fb69mulss x0, x4
    37b69fb6dsubss x0, x8
    37b69fb72divss x1, x2
    37b69fb76divss x0, x2
    37b69fb7amulss x1, x3
    37b69fb7emovss [r8+rcx+8], x1
    37b69fb85mulss x0, x3
    37b69fb89movss [rbp+rcx+8], x0
    37b69fb8fadd rcx, 4
    37b69fb93cmp r9, rcx
    37b69fb96jne 0x37b69fb50
    37b69fb98jmp 0x37b69f9da
    37b69fb9d
    37b69fb9d 0f 1f 00 ...
    37b69fba0
    37b69fba0loc_37b69fba0:
    37b69fba0mov ebp, edx
    37b69fba2mov rdx, [rsp+144]
    37b69fbaamovsxd rax, edi
    37b69fbadlea ebx, [rdi-1]
    37b69fbb0sub ebp, edi
    37b69fbb2lea rcx, [4*rax-4]
    37b69fbbamov rdi, [rsp+632]
    37b69fbc2lea r8, [rip+0xd94773]
    37b69fbc9imul rdx, rax
    37b69fbcdadd ebp, 1
    37b69fbd0add rdx, [rsp+152]
    37b69fbd8mov [rsp+396], ebx
    37b69fbdfadd rax, rdx
    37b69fbe2mov [rsp+392], ebp
    37b69fbe9mov rbp, [rsp+336]
    37b69fbf1lea rax, [r13+4*rax]
    37b69fbf6mov [rsp+96], rdi
    37b69fbfbmov rdx, [rsp+256]
    37b69fc03mov [rsp+88], rax
    37b69fc08mov r9, rbp
    37b69fc0bmov rax, [rsp+264]
    37b69fc13mov [rsp+120], 9
    37b69fc1cmov r14, [rsp+616]
    37b69fc24mov [rsp+80], rax
    37b69fc29mov rax, [rsp+656]
    37b69fc31mov [rsp+112], 12
    37b69fc3amov [rsp+72], rax
    37b69fc3flea rax, [r12+rcx]
    37b69fc43add rcx, r13
    37b69fc46mov [rsp+64], rax
    37b69fc4bmov rax, [rsp+632]
    37b69fc53mov [rsp+48], rcx
    37b69fc58lea rcx, [rip+0xd946eb]
    37b69fc5fmov [rsp+56], rax
    37b69fc64lea rax, [rip+0xd946bd]
    37b69fc6bmov [rsp+40], rax
    37b69fc70mov rax, [rsp+136]
    37b69fc78mov [rsp+104], 5
    37b69fc81mov [rsp+32], rax
    37b69fc86call sgemmtr_
    37b69fc8bmov rdi, [rsp+632]
    37b69fc93mov rsi, [rsp+640]
    37b69fc9bmov eax, ebx
    37b69fc9djmp 0x37b69fd35
    37b69fca2
    37b69fca2 66 0f 1f 44 00 00 f..D..
    37b69fca8
    37b69fca8loc_37b69fca8:
    37b69fca8mov r8l, eax
    37b69fcabxor r15l, r15l
    37b69fcaemov r12l, 1
    37b69fcb4lea ecx, [r8-1]
    37b69fcb8mov [rsp+392], ecx
    37b69fcbfcmp edx, eax
    37b69fcc1je 0x37b69fd6f
    37b69fcc7
    37b69fcc7loc_37b69fcc7:
    37b69fcc7test ecx, ecx
    37b69fcc9jle 0x37b69fd6f
    37b69fccfmov [rsp+32], rdi
    37b69fcd4sub edx, 1
    37b69fcd7mov r8, rdi
    37b69fcdalea r9, [r13+4*r9]
    37b69fcdfmovsxd rdx, edx
    37b69fce2mov rcx, rbp
    37b69fce5lea rdx, [r13+4*rdx]
    37b69fceacall sswap_
    37b69fcefmovsxd rax, [rsp+392]
    37b69fcf7lea r8l, [rax+1]
    37b69fcfb
    37b69fcfbloc_37b69fcfb:
    37b69fcfbcmp r12l, r8l
    37b69fcfeje 0x37b69fd2d
    37b69fd00test r15l, r15l
    37b69fd03je 0x37b69fd2d
    37b69fd05mov [rsp+32], rdi
    37b69fd0asub r12l, 1
    37b69fd0emov r8, rdi
    37b69fd11mov rcx, rbp
    37b69fd14movsxd r12, r12l
    37b69fd17lea r9, [r13+4*rax]
    37b69fd1clea rdx, [r13+4*r12]
    37b69fd21call sswap_
    37b69fd26mov eax, [rsp+392]
    37b69fd2d
    37b69fd2dloc_37b69fd2d:
    37b69fd2dtest eax, eax
    37b69fd2fjle 0x37b69eeb7
    37b69fd35
    37b69fd35loc_37b69fd35:
    37b69fd35movsxd rcx, eax
    37b69fd38lea r9, [rcx-1]
    37b69fd3cmov edx, [rsi+4*r9]
    37b69fd40test edx, edx
    37b69fd42jns 0x37b69fca8
    37b69fd48mov r12l, [rsi+4*rcx-8]
    37b69fd4dlea r8l, [rax-1]
    37b69fd51neg edx
    37b69fd53mov r15l, 1
    37b69fd59lea ecx, [r8-1]
    37b69fd5dmov [rsp+392], ecx
    37b69fd64neg r12l
    37b69fd67cmp edx, eax
    37b69fd69jne 0x37b69fcc7
    37b69fd6f
    37b69fd6floc_37b69fd6f:
    37b69fd6fmovsxd rax, ecx
    37b69fd72jmp 0x37b69fcfb
    37b69fd74
    37b69fd74 0f 1f 40 00 ..@.
    37b69fd78
    37b69fd78loc_37b69fd78:
    37b69fd78mov rax, [rsp+272]
    37b69fd80movss x0, [r13+4*rax]
    37b69fd87movaps x1, x0
    37b69fd8aandps x1, [rip+0xd945bf]
    37b69fd91comiss x1, x11
    37b69fd95jae 0x37b6a0bd0
    37b69fd9bpxor x1, x1
    37b69fd9fucomiss x0, x1
    37b69fda2jp 0x37b69fdaa
    37b69fda4je 0x37b69f174
    37b69fdaa
    37b69fdaaloc_37b69fdaa:
    37b69fdaamov eax, [rsp+240]
    37b69fdb1cmp edx, eax
    37b69fdb3jl 0x37b69f174
    37b69fdb9mov r8, [rsp+216]
    37b69fdc1add r8, [rsp+168]
    37b69fdc9mov edi, eax
    37b69fdcbmovss x1, [r13+4*r8]
    37b69fdd2divss x1, x0
    37b69fdd6movss [r13+4*r8], x1
    37b69fdddcmp edx, eax
    37b69fddfjle 0x37b69f174
    37b69fde5lea ecx, [rdx-1]
    37b69fde8lea rax, [r13+4*r8+4]
    37b69fdedsub ecx, edi
    37b69fdefadd rcx, r8
    37b69fdf2mov r8, [rsp+272]
    37b69fdfalea rcx, [r13+4*rcx+8]
    37b69fdffnop
    37b69fe00
    37b69fe00loc_37b69fe00:
    37b69fe00movss x0, [rax]
    37b69fe04add rax, 4
    37b69fe08divss x0, [r13+4*r8]
    37b69fe0fmovss [rax-4], x0
    37b69fe14cmp rcx, rax
    37b69fe17jne 0x37b69fe00
    37b69fe19jmp 0x37b69f174
    37b69fe1e
    37b69fe1e 66 90 f.
    37b69fe20
    37b69fe20loc_37b69fe20:
    37b69fe20mov rax, rdi
    37b69fe23lea rdi, [rbx+r9]
    37b69fe27lea rdx, [rbp+r9]
    37b69fe2cimul rax, rbp
    37b69fe30mov [rsp+296], rdi
    37b69fe38sub rbp, 1
    37b69fe3cadd rax, rsi
    37b69fe3fmov esi, r14l
    37b69fe42lea rcx, [rbx+rax]
    37b69fe46sub esi, r10l
    37b69fe49lea rdi, [r13+4*rcx]
    37b69fe4elea rcx, [rbx+r11]
    37b69fe52mov [rsp+288], rdi
    37b69fe5alea r9, [r13+4*rcx]
    37b69fe5fmov edi, r10l
    37b69fe62mov r10l, 2
    37b69fe68jmp 0x37b69f868
    37b69fe6d
    37b69fe6d 0f 1f 00 ...
    37b69fe70
    37b69fe70loc_37b69fe70:
    37b69fe70mov ebp, [rsi+4*r8]
    37b69fe74xor eax, eax
    37b69fe76neg edx
    37b69fe78lea r9l, [rbx+2]
    37b69fe7cmov [rsp+392], r9l
    37b69fe84neg ebp
    37b69fe86cmp ebp, edi
    37b69fe88setne al
    37b69fe8bcmp edx, ebx
    37b69fe8dje 0x37b6a0968
    37b69fe93cmp r9l, ecx
    37b69fe96jle 0x37b6a0b58
    37b69fe9ccmp ebp, edi
    37b69fe9eje 0x37b69eeaa
    37b69fea4mov ebx, edi
    37b69fea6mov edi, r9l
    37b69fea9
    37b69fea9loc_37b69fea9:
    37b69fea9sub ecx, edi
    37b69feabmov [rsp+32], r12
    37b69feb0movsxd rax, edi
    37b69feb3movsxd rbx, ebx
    37b69feb6imul rax, [rsp+144]
    37b69febfadd ecx, 1
    37b69fec2movsxd rbp, ebp
    37b69fec5mov r8, r12
    37b69fec8add rax, [rsp+152]
    37b69fed0mov [rsp+396], ecx
    37b69fed7add rbx, rax
    37b69fedamov rcx, [rsp+136]
    37b69fee2add rbp, rax
    37b69fee5lea rdx, [r13+4*rbp]
    37b69feealea r9, [r13+4*rbx]
    37b69feefcall sswap_
    37b69fef4mov ecx, [r15]
    37b69fef7
    37b69fef7loc_37b69fef7:
    37b69fef7cmp edi, ecx
    37b69fef9jg 0x37b69eeaa
    37b69feffmov ebx, [rsp+392]
    37b69ff06jmp 0x37b69ee80
    37b69ff0b
    37b69ff0b 0f 1f 44 00 00 ..D..
    37b69ff10
    37b69ff10loc_37b69ff10:
    37b69ff10mov rax, [rsp+664]
    37b69ff18mov ecx, [rax]
    37b69ff1atest ecx, ecx
    37b69ff1cjne 0x37b69ff28
    37b69ff1emov rax, [rsp+664]
    37b69ff26mov [rax], ebx
    37b69ff28
    37b69ff28loc_37b69ff28:
    37b69ff28mov rax, [rsp+144]
    37b69ff30mov rdi, [rsp+152]
    37b69ff38mov rdx, [rsp+200]
    37b69ff40mov rcx, [rsp+192]
    37b69ff48imul rax, rbp
    37b69ff4clea rax, [rdi+rax+1]
    37b69ff51lea rdi, [rip+0xd943cc]
    37b69ff58mov [rsp+32], rdi
    37b69ff5dlea r9, [r13+4*rax]
    37b69ff62mov r8, rdi
    37b69ff65call scopy_
    37b69ff6amovsxd rsi, [rsp+372]
    37b69ff72mov rbp, rsi
    37b69ff75sub ebp, 1
    37b69ff78jmp 0x37b69ed30
    37b69ff7d
    37b69ff7d 0f 1f 00 ...
    37b69ff80
    37b69ff80loc_37b69ff80:
    37b69ff80mov rdx, [rsp+200]
    37b69ff88mov rcx, [rsp+136]
    37b69ff90lea eax, [rbx-1]
    37b69ff93lea r8, [rip+0xd9438a]
    37b69ff9aandps x9, [rip+0xd943ae]
    37b69ffa2mov [rsp+396], eax
    37b69ffa9call isamax_
    37b69ffaemov rcx, [rsp+216]
    37b69ffb6mov [rsp+368], eax
    37b69ffbdcdqe
    37b69ffbfmovaps x0, x9
    37b69ffc3lea rdx, [rax+rcx]
    37b69ffc7movss x10, [r12+4*rdx]
    37b69ffcdandps x10, [rip+0xd9437b]
    37b69ffd5maxss x0, x10
    37b69ffdaucomiss x0, x7
    37b69ffddjp 0x37b69ffe5
    37b69ffdfje 0x37b69ff10
    37b69ffe5
    37b69ffe5loc_37b69ffe5:
    37b69ffe5movaps x0, x10
    37b69ffe9mulss x0, x6
    37b69ffedcomiss x0, x9
    37b69fff1jbe 0x37b69ec70
    37b69fff7mov rcx, [rsp+176]
    37b69ffffadd edi, 1
    37b6a0002mov rdx, [rsp+184]
    37b6a000amovsxd rdi, edi
    37b6a000dimul rdi, rcx
    37b6a0011sub rsi, rcx
    37b6a0014add rsi, rdx
    37b6a0017mov [rsp+168], rsi
    37b6a001flea rsi, [r12+4*rsi+4]
    37b6a0024add rdi, rdx
    37b6a0027mov [rsp+208], rsi
    37b6a002fmov [rsp+232], rdi
    37b6a0037lea rdi, [rsp+368]
    37b6a003fmov [rsp+224], rdi
    37b6a0047
    37b6a0047loc_37b6a0047:
    37b6a0047imul rax, [rsp+144]
    37b6a0050mov rcx, [rsp+152]
    37b6a0058lea rdi, [rip+0xd942c5]
    37b6a005fmov [rsp+32], rdi
    37b6a0064mov r9, [rsp+208]
    37b6a006cmov r8, rdi
    37b6a006flea rax, [rcx+rax+1]
    37b6a0074mov rcx, [rsp+224]
    37b6a007clea rdx, [r13+4*rax]
    37b6a0081call scopy_
    37b6a0086movsxd rbp, [rsp+368]
    37b6a008emov edx, ebx
    37b6a0090mov r8, [rsp+632]
    37b6a0098mov rcx, [rsp+136]
    37b6a00a0sub edx, ebp
    37b6a00a2lea edi, [rbp+1]
    37b6a00a5mov r14, rbp
    37b6a00a8mov [rsp+396], edx
    37b6a00afmov rdx, [rsp+168]
    37b6a00b7movsxd rdi, edi
    37b6a00balea rax, [rdx+rdi]
    37b6a00belea rsi, [r12+4*rax]
    37b6a00c2mov rax, [rsp+144]
    37b6a00camov r9, rsi
    37b6a00cdimul rax, rdi
    37b6a00d1add rax, [rsp+152]
    37b6a00d9add rax, rbp
    37b6a00dclea rdx, [r13+4*rax]
    37b6a00e1lea rax, [rip+0xd9423c]
    37b6a00e8mov [rsp+32], rax
    37b6a00edmov [rsp+240], rdx
    37b6a00f5call scopy_
    37b6a00famov eax, [r15]
    37b6a00fdcmp eax, ebx
    37b6a00ffjg 0x37b6a0768
    37b6a0105
    37b6a0105loc_37b6a0105:
    37b6a0105cmp r14l, ebx
    37b6a0108jne 0x37b6a0428
    37b6a010ecmp r14l, 1
    37b6a0112jg 0x37b6a08f8
    37b6a0118mov [rsp+248], rbp
    37b6a0120mov esi, r14l
    37b6a0123
    37b6a0123loc_37b6a0123:
    37b6a0123lea rax, [rip+0xd941fa]
    37b6a012amov rcx, [rsp+192]
    37b6a0132mov r9, [rsp+200]
    37b6a013amov [rsp+380], esi
    37b6a0141mov [rsp+32], rax
    37b6a0146mov rdx, [rsp+208]
    37b6a014emov r8, rax
    37b6a0151call scopy_
    37b6a0156mov rax, [rsp+608]
    37b6a015emovsxd rcx, [rsp+372]
    37b6a0166mov ebp, [rax]
    37b6a0168mov rax, [rsp+144]
    37b6a0170mov [rsp+232], rcx
    37b6a0178mov r14, rcx
    37b6a017bimul rax, rcx
    37b6a017fadd rax, [rsp+152]
    37b6a0187add ebp, ecx
    37b6a0189sub ebp, [r15]
    37b6a018cmov [rsp+208], rax
    37b6a0194cmp esi, ecx
    37b6a0196je 0x37b69ec93
    37b6a019cmovsxd rbp, ebp
    37b6a019flea r11l, [rcx-1]
    37b6a01a3mov ebx, ecx
    37b6a01a5mov r10l, 1
    37b6a01abmov [rsp+224], rcx
    37b6a01b3imul rbp, [rsp+176]
    37b6a01bcadd rbp, [rsp+184]
    37b6a01c4
    37b6a01c4loc_37b6a01c4:
    37b6a01c4mov r9, [rsp+208]
    37b6a01ccmov rcx, [rsp+232]
    37b6a01d4add rdi, rax
    37b6a01d7mov [rsp+280], r10l
    37b6a01dfmov r10, [rsp+248]
    37b6a01e7mov [rsp+272], rax
    37b6a01eflea r8, [rip+0xd9412e]
    37b6a01f6lea rdx, [r9+rcx]
    37b6a01famov rax, [rsp+632]
    37b6a0202mov rcx, [rsp+136]
    37b6a020amov [rsp+288], r11l
    37b6a0212movss x0, [r13+4*rdx]
    37b6a0219lea rdx, [r10+r9]
    37b6a021dmov r9, [rsp+240]
    37b6a0225movss [r13+4*rdx], x0
    37b6a022cmov edx, r11l
    37b6a022fmov [rsp+32], rax
    37b6a0234sub edx, esi
    37b6a0236mov [rsp+396], edx
    37b6a023dlea rdx, [r13+4*rdi]
    37b6a0242call scopy_
    37b6a0247mov r10, [rsp+248]
    37b6a024fimul r10, [rsp+144]
    37b6a0258lea rcx, [rsp+380]
    37b6a0260mov rax, [rsp+152]
    37b6a0268lea r8, [rax+r10+1]
    37b6a026dmov rax, [rsp+272]
    37b6a0275lea r9, [r13+4*r8]
    37b6a027alea rdx, [r13+4*rax+4]
    37b6a027fmov [rsp+240], rax
    37b6a0287lea rax, [rip+0xd94096]
    37b6a028emov [rsp+32], rax
    37b6a0293mov r8, rax
    37b6a0296call scopy_
    37b6a029bmov edx, [r15]
    37b6a029emovsxd rdi, [rsp+380]
    37b6a02a6mov rax, [rsp+240]
    37b6a02aesub edx, r14l
    37b6a02b1mov rsi, rdi
    37b6a02b4lea rcx, [rdi+rax]
    37b6a02b8add edx, 1
    37b6a02bbadd rax, [rsp+232]
    37b6a02c3add rdi, rbp
    37b6a02c6mov [rsp+396], edx
    37b6a02cdlea rdx, [r13+4*rax]
    37b6a02d2mov rax, [rsp+632]
    37b6a02dalea r9, [r13+4*rcx]
    37b6a02dfmov rcx, [rsp+136]
    37b6a02e7mov [rsp+32], rax
    37b6a02ecmov r8, rax
    37b6a02efcall sswap_
    37b6a02f4mov eax, [r15]
    37b6a02f7lea r9, [r12+4*rdi]
    37b6a02fbadd rbp, [rsp+232]
    37b6a0303mov rcx, [rsp+136]
    37b6a030blea rdx, [r12+4*rbp]
    37b6a030fsub eax, r14l
    37b6a0312add eax, 1
    37b6a0315mov [rsp+396], eax
    37b6a031cmov rax, [rsp+656]
    37b6a0324mov [rsp+32], rax
    37b6a0329mov r8, rax
    37b6a032ccall sswap_
    37b6a0331cmp [rsp+280], 1
    37b6a0339je 0x37b69ec93
    37b6a033fmov rax, [rsp+224]
    37b6a0347add rax, [rsp+208]
    37b6a034fmov [rsp+272], rax
    37b6a0357mov rax, [rsp+608]
    37b6a035fmovsxd rdx, [rsp+372]
    37b6a0367movsxd r10, [rsp+288]
    37b6a036fmovd x8, [rsp+384]
    37b6a0379mov ecx, [rax]
    37b6a037bmov rbp, rdx
    37b6a037e
    37b6a037eloc_37b6a037e:
    37b6a037emovd x0, esi
    37b6a0382mov rax, [rsp+224]
    37b6a038amov rsi, [rsp+216]
    37b6a0392pxor x3, x3
    37b6a0396punpckldq x0, x8
    37b6a039bmov r8, [rsp+144]
    37b6a03a3add rax, rsi
    37b6a03a6psubd x3, x0
    37b6a03aamovss x8, [r12+4*rax]
    37b6a03b0mov rax, [rsp+168]
    37b6a03b8imul r8, r10
    37b6a03bcadd r8, [rsp+152]
    37b6a03c4add rax, r10
    37b6a03c7movss x12, [r12+4*rax]
    37b6a03cdlea rax, [rsi+r10]
    37b6a03d1movss x1, [r12+4*rax]
    37b6a03d7cmp ebx, 2
    37b6a03dajg 0x37b6a0980
    37b6a03e0
    37b6a03e0loc_37b6a03e0:
    37b6a03e0mov rax, [rsp+208]
    37b6a03e8add r8, r10
    37b6a03ebsub ebp, 2
    37b6a03eemovss [r13+4*r8], x12
    37b6a03f5add rax, r10
    37b6a03f8movss [r13+4*rax], x1
    37b6a03ffmov rax, [rsp+272]
    37b6a0407movss [r13+4*rax], x8
    37b6a040emov rax, [rsp+640]
    37b6a0416movq [rax+4*rdx-8], x3
    37b6a041cjmp 0x37b69ed46
    37b6a0421
    37b6a0421 0f 1f 80 00 00 00 00 .......
    37b6a0428
    37b6a0428loc_37b6a0428:
    37b6a0428mov rcx, [rsp+136]
    37b6a0430mov eax, ebx
    37b6a0432mov rdx, rsi
    37b6a0435lea r8, [rip+0xd93ee8]
    37b6a043csub eax, r14l
    37b6a043fmov [rsp+396], eax
    37b6a0446call isamax_
    37b6a044blea esi, [r14+rax]
    37b6a044fmovsxd rax, esi
    37b6a0452add rax, [rsp+168]
    37b6a045amovss x9, [r12+4*rax]
    37b6a0460andps x9, [rip+0xd93ee8]
    37b6a0468cmp r14l, 1
    37b6a046cjg 0x37b6a0958
    37b6a0472
    37b6a0472loc_37b6a0472:
    37b6a0472movaps x1, x9
    37b6a0476mov rax, [rsp+168]
    37b6a047emulss x1, x6
    37b6a0482add rax, rbp
    37b6a0485movss x0, [r12+4*rax]
    37b6a048bandps x0, [rip+0xd93ebe]
    37b6a0492comiss x1, x0
    37b6a0495jbe 0x37b6a0bb0
    37b6a049bmov r8l, [rsp+384]
    37b6a04a3cmp r8l, esi
    37b6a04a6je 0x37b6a0508
    37b6a04a8comiss x10, x9
    37b6a04acjae 0x37b6a0508
    37b6a04aelea rax, [rip+0xd93e6f]
    37b6a04b5mov r9, [rsp+200]
    37b6a04bdmov rdx, [rsp+208]
    37b6a04c5mov [rsp+384], r14l
    37b6a04cdmov [rsp+32], rax
    37b6a04d2mov r8, rax
    37b6a04d5movaps x10, x9
    37b6a04d9mov rcx, [rsp+192]
    37b6a04e1mov [rsp+368], esi
    37b6a04e8call scopy_
    37b6a04edmov [rsp+160], esi
    37b6a04f4mov ebx, [rsp+372]
    37b6a04fbmovsxd rax, esi
    37b6a04fejmp 0x37b6a0047
    37b6a0503
    37b6a0503 0f 1f 44 00 00 ..D..
    37b6a0508
    37b6a0508loc_37b6a0508:
    37b6a0508mov eax, esi
    37b6a050amov esi, r14l
    37b6a050dlea r11l, [rbx-1]
    37b6a0511movsxd r10, ebx
    37b6a0514mov r14l, eax
    37b6a0517mov rax, [rsp+608]
    37b6a051fmov rdx, [rsp+144]
    37b6a0527mov [rsp+248], rbp
    37b6a052fmov r9, [rsp+152]
    37b6a0537mov [rsp+224], r10
    37b6a053fmov ecx, [rax]
    37b6a0541mov [rsp+380], esi
    37b6a0548lea eax, [r11+rcx]
    37b6a054csub eax, [r15]
    37b6a054fmovsxd rbp, eax
    37b6a0552mov rax, rdx
    37b6a0555imul rax, r10
    37b6a0559lea r10, [r9+rax]
    37b6a055dsub rax, rdx
    37b6a0560mov [rsp+208], r10
    37b6a0568movsxd r10, r11l
    37b6a056badd rax, r9
    37b6a056emov [rsp+232], r10
    37b6a0576cmp r8l, ebx
    37b6a0579je 0x37b6a0b20
    37b6a057fmov edx, ebx
    37b6a0581lea ecx, [r8+1]
    37b6a0585mov [rsp+280], r11l
    37b6a058dmov r11, [rsp+152]
    37b6a0595sub edx, r8l
    37b6a0598movsxd rcx, ecx
    37b6a059bmov [rsp+296], r10
    37b6a05a3mov [rsp+396], edx
    37b6a05aamov rdx, [rsp+144]
    37b6a05b2mov [rsp+288], rax
    37b6a05baimul rdx, rcx
    37b6a05beadd rcx, [rsp+208]
    37b6a05c6lea r9, [rdx+r11]
    37b6a05camov [rsp+160], rdx
    37b6a05d2movsxd rdx, r8l
    37b6a05d5mov r8, [rsp+632]
    37b6a05ddadd rdx, r9
    37b6a05e0mov [rsp+32], r8
    37b6a05e5lea r9, [r13+4*rdx]
    37b6a05ealea rdx, [r13+4*rcx]
    37b6a05efmov rcx, [rsp+136]
    37b6a05f7lea r8, [rip+0xd93d26]
    37b6a05fecall scopy_
    37b6a0603mov r10, [rsp+152]
    37b6a060bmov rdx, [rsp+160]
    37b6a0613lea rcx, [rsp+384]
    37b6a061bsub rdx, [rsp+144]
    37b6a0623mov rax, [rsp+208]
    37b6a062blea r8, [r10+rdx+1]
    37b6a0630lea r10, [rip+0xd93ced]
    37b6a0637mov [rsp+32], r10
    37b6a063clea rdx, [r13+4*rax+4]
    37b6a0641lea r9, [r13+4*r8]
    37b6a0646mov r8, r10
    37b6a0649call scopy_
    37b6a064emov edx, [r15]
    37b6a0651mov rax, [rsp+208]
    37b6a0659movd x8, [rsp+384]
    37b6a0663sub edx, ebx
    37b6a0665add edx, 1
    37b6a0668mov [rsp+396], edx
    37b6a066fmovd edx, x8
    37b6a0674movsxd r10, edx
    37b6a0677mov rdx, [rsp+224]
    37b6a067flea rcx, [r10+rax]
    37b6a0683mov [rsp+160], r10
    37b6a068blea r11, [rdx+rax]
    37b6a068fmov rax, [rsp+632]
    37b6a0697lea r9, [r13+4*rcx]
    37b6a069cmov rcx, [rsp+136]
    37b6a06a4lea rdx, [r13+4*r11]
    37b6a06a9mov [rsp+272], r11
    37b6a06b1mov [rsp+32], rax
    37b6a06b6mov r8, rax
    37b6a06b9call sswap_
    37b6a06bemov eax, [r15]
    37b6a06c1sub eax, [rsp+280]
    37b6a06c8lea edx, [rax+1]
    37b6a06cbmov r10, [rsp+160]
    37b6a06d3mov rax, [rsp+656]
    37b6a06dbimul rbp, [rsp+176]
    37b6a06e4mov [rsp+396], edx
    37b6a06ebadd rbp, [rsp+184]
    37b6a06f3mov rdx, [rsp+224]
    37b6a06fbmov [rsp+32], rax
    37b6a0700mov r8, rax
    37b6a0703lea rcx, [r10+rbp]
    37b6a0707add rdx, rbp
    37b6a070alea r9, [r12+4*rcx]
    37b6a070emov rcx, [rsp+136]
    37b6a0716lea rdx, [r12+4*rdx]
    37b6a071acall sswap_
    37b6a071fmov r11l, [rsp+280]
    37b6a0727mov rax, [rsp+288]
    37b6a072fcmp esi, r11l
    37b6a0732jne 0x37b6a0b90
    37b6a0738mov rax, [rsp+608]
    37b6a0740mov [rsp+160], r14l
    37b6a0748mov ebp, ebx
    37b6a074amov rdx, [rsp+224]
    37b6a0752mov r10, [rsp+296]
    37b6a075amov ecx, [rax]
    37b6a075cjmp 0x37b6a037e
    37b6a0761
    37b6a0761 0f 1f 80 00 00 00 00 .......
    37b6a0768
    37b6a0768loc_37b6a0768:
    37b6a0768sub eax, ebx
    37b6a076aadd ebx, 1
    37b6a076dmov [rsp+88], 12
    37b6a0776mov r8, [rsp+136]
    37b6a077emov [rsp+396], eax
    37b6a0785lea rax, [rip+0xd93b98]
    37b6a078cmovsxd rbx, ebx
    37b6a078fmov rdx, [rsp+192]
    37b6a0797mov [rsp+80], rax
    37b6a079cmov rax, [rsp+208]
    37b6a07a4lea r9, [rip+0xd93b7d]
    37b6a07abimul rbx, [rsp+144]
    37b6a07b4mov rcx, [rsp+256]
    37b6a07bcmov [rsp+72], rax
    37b6a07c1mov rax, [rsp+264]
    37b6a07c9mov [rsp+64], rax
    37b6a07cemov rax, [rsp+656]
    37b6a07d6mov [rsp+56], rax
    37b6a07dbmov rax, [rsp+232]
    37b6a07e3add rax, rbp
    37b6a07e6lea rax, [r12+4*rax]
    37b6a07eamov [rsp+48], rax
    37b6a07efmov rax, [rsp+632]
    37b6a07f7mov [rsp+40], rax
    37b6a07fcmov rax, [rsp+152]
    37b6a0804lea rax, [rax+rbx+1]
    37b6a0809lea rax, [r13+4*rax]
    37b6a080emov [rsp+32], rax
    37b6a0813call sgemv_
    37b6a0818mov ebx, [rsp+372]
    37b6a081fjmp 0x37b6a0105
    37b6a0824
    37b6a0824 0f 1f 40 00 ..@.
    37b6a0828
    37b6a0828loc_37b6a0828:
    37b6a0828sub eax, ebx
    37b6a082amovsxd rdx, ebx
    37b6a082dadd ebx, 1
    37b6a0830mov [rsp+88], 12
    37b6a0839mov [rsp+396], eax
    37b6a0840lea rax, [rip+0xd93add]
    37b6a0847movsxd rbx, ebx
    37b6a084amov r8, [rsp+136]
    37b6a0852mov [rsp+80], rax
    37b6a0857mov rax, [rsp+200]
    37b6a085flea r9, [rip+0xd93ac2]
    37b6a0866imul rbx, [rsp+144]
    37b6a086fmov rcx, [rsp+256]
    37b6a0877mov [rsp+72], rax
    37b6a087cmov rax, [rsp+264]
    37b6a0884mov [rsp+64], rax
    37b6a0889mov rax, [rsp+656]
    37b6a0891mov [rsp+56], rax
    37b6a0896mov rax, [rsp+176]
    37b6a089eadd rax, rsi
    37b6a08a1add rax, [rsp+184]
    37b6a08a9add rax, rdx
    37b6a08acmov rdx, [rsp+192]
    37b6a08b4lea rax, [r12+4*rax]
    37b6a08b8mov [rsp+48], rax
    37b6a08bdmov rax, [rsp+632]
    37b6a08c5mov [rsp+40], rax
    37b6a08camov rax, [rsp+152]
    37b6a08d2lea rax, [rax+rbx+1]
    37b6a08d7lea rax, [r13+4*rax]
    37b6a08dcmov [rsp+32], rax
    37b6a08e1call sgemv_
    37b6a08e6mov ebx, [rsp+372]
    37b6a08edjmp 0x37b69ec47
    37b6a08f2
    37b6a08f2 66 0f 1f 44 00 00 f..D..
    37b6a08f8
    37b6a08f8loc_37b6a08f8:
    37b6a08f8pxor x8, x8
    37b6a08fd
    37b6a08fdloc_37b6a08fd:
    37b6a08fdmov rdx, [rsp+208]
    37b6a0905mov rcx, [rsp+136]
    37b6a090dlea eax, [r14-1]
    37b6a0911lea r8, [rip+0xd93a0c]
    37b6a0918mov [rsp+396], eax
    37b6a091fcall isamax_
    37b6a0924mov esi, eax
    37b6a0926cdqe
    37b6a0928add rax, [rsp+168]
    37b6a0930movss x9, [r12+4*rax]
    37b6a0936andps x9, [rip+0xd93a12]
    37b6a093ecomiss x9, x8
    37b6a0942maxss x9, x8
    37b6a0947cmovbe esi, [rsp+160]
    37b6a094fjmp 0x37b6a0472
    37b6a0954
    37b6a0954 0f 1f 40 00 ..@.
    37b6a0958
    37b6a0958loc_37b6a0958:
    37b6a0958mov [rsp+160], esi
    37b6a095fmovaps x8, x9
    37b6a0963jmp 0x37b6a08fd
    37b6a0965
    37b6a0965 0f 1f 00 ...
    37b6a0968
    37b6a0968loc_37b6a0968:
    37b6a0968mov ebx, edi
    37b6a096amov edi, r9l
    37b6a096d
    37b6a096dloc_37b6a096d:
    37b6a096dtest eax, eax
    37b6a096fjne 0x37b69fea9
    37b6a0975jmp 0x37b69fef7
    37b6a097a
    37b6a097a 66 0f 1f 44 00 00 f..D..
    37b6a0980
    37b6a0980loc_37b6a0980:
    37b6a0980movaps x5, x8
    37b6a0984movaps x4, x12
    37b6a0988lea r11, [r13+4*r8]
    37b6a098dsub ebx, 1
    37b6a0990divss x5, x1
    37b6a0994movss x2, [rip+0xd939c8]
    37b6a099cmov rax, [rsp+168]
    37b6a09a4lea rsi, [r12+4*rax]
    37b6a09a8mov rax, [rsp+216]
    37b6a09b0lea rdi, [r12+4*rax]
    37b6a09b4mov rax, [rsp+208]
    37b6a09bclea r9, [r13+4*rax]
    37b6a09c1mov eax, 1
    37b6a09c6divss x4, x1
    37b6a09camovaps x10, x5
    37b6a09cemulss x10, x4
    37b6a09d3subss x10, [rip+0xd93988]
    37b6a09dcdivss x2, x10
    37b6a09e1nop [rax]
    37b6a09e5nop [rax+rax]
    37b6a09f0
    37b6a09f0loc_37b6a09f0:
    37b6a09f0movss x9, [rsi+4*rax]
    37b6a09f6movss x0, [rdi+4*rax]
    37b6a09fbmovaps x10, x9
    37b6a09ffmulss x10, x5
    37b6a0a04subss x10, x0
    37b6a0a09mulss x0, x4
    37b6a0a0dsubss x0, x9
    37b6a0a12divss x10, x1
    37b6a0a17divss x0, x1
    37b6a0a1bmulss x10, x2
    37b6a0a20movss [r11+4*rax], x10
    37b6a0a26mulss x0, x2
    37b6a0a2amovss [r9+4*rax], x0
    37b6a0a30add rax, 1
    37b6a0a34cmp rax, rbx
    37b6a0a37jne 0x37b6a09f0
    37b6a0a39jmp 0x37b6a03e0
    37b6a0a3e
    37b6a0a3e 66 90 f.
    37b6a0a40
    37b6a0a40loc_37b6a0a40:
    37b6a0a40xor eax, eax
    37b6a0a42mov ebp, 1
    37b6a0a47
    37b6a0a47loc_37b6a0a47:
    37b6a0a47sub ecx, edi
    37b6a0a49mov [rsp+32], r12
    37b6a0a4emovsxd rdx, edx
    37b6a0a51add ecx, 1
    37b6a0a54mov [rsp+168], eax
    37b6a0a5bmov [rsp+396], ecx
    37b6a0a62movsxd rcx, edi
    37b6a0a65imul rcx, [rsp+144]
    37b6a0a6eadd rcx, [rsp+152]
    37b6a0a76add r8, rcx
    37b6a0a79add rdx, rcx
    37b6a0a7cmov rcx, [rsp+136]
    37b6a0a84lea r9, [r13+4*r8]
    37b6a0a89lea rdx, [r13+4*rdx]
    37b6a0a8emov r8, r12
    37b6a0a91call sswap_
    37b6a0a96mov ecx, [r15]
    37b6a0a99mov eax, [rsp+168]
    37b6a0aa0jmp 0x37b6a096d
    37b6a0aa5
    37b6a0aa5 0f 1f 00 ...
    37b6a0aa8
    37b6a0aa8loc_37b6a0aa8:
    37b6a0aa8mov r11, [rsp+152]
    37b6a0ab0mov rdx, [rsp+304]
    37b6a0ab8movsxd r10, edi
    37b6a0abbadd rdx, [rsp+144]
    37b6a0ac3add r11, rdx
    37b6a0ac6cmp r14l, [rsp+240]
    37b6a0aceje 0x37b69f9a0
    37b6a0ad4mov rax, [rsp+216]
    37b6a0adcmov rcx, [rsp+168]
    37b6a0ae4mov r10l, 2
    37b6a0aealea rdx, [rax+rcx]
    37b6a0aeelea rax, [rbx+r11]
    37b6a0af2lea rax, [r13+4*rax]
    37b6a0af7lea rcx, [rsp+376]
    37b6a0affmov [rsp+288], rax
    37b6a0b07mov r9, rax
    37b6a0b0amov rax, r11
    37b6a0b0dmov [rsp+192], rcx
    37b6a0b15jmp 0x37b69f868
    37b6a0b1a
    37b6a0b1a 66 0f 1f 44 00 00 f..D..
    37b6a0b20
    37b6a0b20loc_37b6a0b20:
    37b6a0b20cmp esi, r11l
    37b6a0b23jne 0x37b6a0b68
    37b6a0b25mov rdx, [rsp+224]
    37b6a0b2dmov rax, [rsp+208]
    37b6a0b35mov ebp, ebx
    37b6a0b37movd x8, ebx
    37b6a0b3cmov [rsp+160], r14l
    37b6a0b44add rax, rdx
    37b6a0b47mov [rsp+272], rax
    37b6a0b4fjmp 0x37b6a037e
    37b6a0b54
    37b6a0b54 0f 1f 40 00 ..@.
    37b6a0b58
    37b6a0b58loc_37b6a0b58:
    37b6a0b58mov ebx, edi
    37b6a0b5amov edi, r9l
    37b6a0b5djmp 0x37b6a0a47
    37b6a0b62
    37b6a0b62 66 0f 1f 44 00 00 f..D..
    37b6a0b68
    37b6a0b68loc_37b6a0b68:
    37b6a0b68imul rbp, [rsp+176]
    37b6a0b71mov [rsp+160], r14l
    37b6a0b79mov r10l, 2
    37b6a0b7fmov r14l, r11l
    37b6a0b82add rbp, [rsp+184]
    37b6a0b8ajmp 0x37b6a01c4
    37b6a0b8f
    37b6a0b8f 90 .
    37b6a0b90
    37b6a0b90loc_37b6a0b90:
    37b6a0b90mov [rsp+160], r14l
    37b6a0b98mov r10l, 2
    37b6a0b9emov r14l, r11l
    37b6a0ba1jmp 0x37b6a01c4
    37b6a0ba6
    37b6a0ba6 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b6a0bb0
    37b6a0bb0loc_37b6a0bb0:
    37b6a0bb0mov eax, esi
    37b6a0bb2mov [rsp+248], rbp
    37b6a0bbamov esi, r14l
    37b6a0bbdmov [rsp+160], eax
    37b6a0bc4jmp 0x37b6a0123
    37b6a0bc9
    37b6a0bc9 0f 1f 80 00 00 00 00 .......
    37b6a0bd0
    37b6a0bd0loc_37b6a0bd0:
    37b6a0bd0movss x1, [rip+0xd9378c]
    37b6a0bd8mov rax, [rsp+168]
    37b6a0be0sub edx, edi
    37b6a0be2lea r9, [rip+0xd9373b]
    37b6a0be9add rax, [rsp+216]
    37b6a0bf1mov rcx, [rsp+136]
    37b6a0bf9mov [rsp+396], edx
    37b6a0c00lea rdx, [rsp+388]
    37b6a0c08divss x1, x0
    37b6a0c0clea r8, [r13+4*rax]
    37b6a0c11movss [rsp+388], x1
    37b6a0c1acall sscal_
    37b6a0c1fmov edx, [r15]
    37b6a0c22jmp 0x37b69f174
    37b6a0c27
    37b6a0c27loc_37b6a0c27:
    37b6a0c27lea rdx, [rcx+1]
    37b6a0c2bmovss x1, [r13+4*rdx]
    37b6a0c32divss x1, x0
    37b6a0c36movss [r13+4*rdx], x1
    37b6a0c3dcmp eax, 2
    37b6a0c40je 0x37b69ed30
    37b6a0c46sub eax, 3
    37b6a0c49lea rdx, [r13+4*rcx+8]
    37b6a0c4eadd rax, rcx
    37b6a0c51lea rax, [r13+4*rax+12]
    37b6a0c56nop [rax+rax]
    37b6a0c60
    37b6a0c60loc_37b6a0c60:
    37b6a0c60movss x0, [rdx]
    37b6a0c64add rdx, 4
    37b6a0c68divss x0, [r13+4*r8]
    37b6a0c6fmovss [rdx-4], x0
    37b6a0c74cmp rdx, rax
    37b6a0c77jne 0x37b6a0c60
    37b6a0c79jmp 0x37b69ed30
    37b6a0c7e
    37b6a0c7e 66 90 f.
    37b6a0c80
    37b6a0c80loc_37b6a0c80:
    37b6a0c80mov eax, [r15]
    37b6a0c83jmp 0x37b69f12e
    37b6a0c88
    37b6a0c88 0f 1f 84 00 00 00 00 00 ........
    37b6a0c90
    37b6a0c90loc_37b6a0c90:
    37b6a0c90movss x1, [rip+0xd936cc]
    37b6a0c98lea r8, [r13+4*rcx+4]
    37b6a0c9dmov rcx, [rsp+136]
    37b6a0ca5lea rdx, [rsp+388]
    37b6a0cadlea r9, [rip+0xd93670]
    37b6a0cb4mov [rsp+396], ebp
    37b6a0cbbdivss x1, x0
    37b6a0cbfmovss [rsp+388], x1
    37b6a0cc8call sscal_
    37b6a0ccdmov ebx, [rsp+380]
    37b6a0cd4jmp 0x37b69ed30
    37b6a0cd9
    37b6a0cd9 0f 1f 80 00 00 00 00 .......
    37b6a0ce0
    37b6a0ce0loc_37b6a0ce0:
    37b6a0ce0mov rsi, [rsp+272]
    37b6a0ce8mov eax, [r15]
    37b6a0cebmov r14l, [rsp+380]
    37b6a0cf3lea rsi, [r13+4*rsi]
    37b6a0cf8mov [rsp+280], rsi
    37b6a0d00mov rsi, [rsp+224]
    37b6a0d08lea rsi, [r12+4*rsi]
    37b6a0d0cmov [rsp+232], rsi
    37b6a0d14jmp 0x37b69f131
    37b6a0d19
    37b6a0d19loc_37b6a0d19:
    37b6a0d19lea rax, [rsp+372]
    37b6a0d21xor edx, edx
    37b6a0d23mov [rsp+192], rax
    37b6a0d2blea rax, [rsp+396]
    37b6a0d33mov [rsp+136], rax
    37b6a0d3blea rax, [rip+0xd935ea]
    37b6a0d42mov [rsp+256], rax
    37b6a0d4alea rax, [rip+0xd935e7]
    37b6a0d51mov [rsp+264], rax
    37b6a0d59jmp 0x37b69ed76
    37b69ea40
    37b6a0d5e
    37b6a0d5e // align 0x10
    37b6a0d60
    37b6a0d60slasyf_rk_:
    37b6a0d60push r15
    37b6a0d62push r14
    37b6a0d64push r13
    37b6a0d66push r12
    37b6a0d68push rbp
    37b6a0d69push rdi
    37b6a0d6apush rsi
    37b6a0d6bpush rbx
    37b6a0d6csub rsp, 520
    37b6a0d73movups [rsp+400], x6
    37b6a0d7bmovups [rsp+416], x7
    37b6a0d83movups [rsp+432], x8
    37b6a0d8cmovups [rsp+448], x9
    37b6a0d95movups [rsp+464], x10
    37b6a0d9emovups [rsp+480], x11
    37b6a0da7movups [rsp+496], x12
    37b6a0db0mov rax, [rsp+632]
    37b6a0db8mov r14, [rsp+624]
    37b6a0dc0mov r13, [rsp+656]
    37b6a0dc8mov rbx, rcx
    37b6a0dcbmovsxd rcx, [rax]
    37b6a0dcexor eax, eax
    37b6a0dd0mov r15, rdx
    37b6a0dd3mov edx, 1
    37b6a0dd8mov [rsp+608], r8
    37b6a0de0mov rdi, r9
    37b6a0de3test rcx, rcx
    37b6a0de6cmovns rax, rcx
    37b6a0dealea rcx, [rip+0xd9357f]
    37b6a0df1mov [rsp+144], rax
    37b6a0df9not rax
    37b6a0dfcmov [rsp+136], rax
    37b6a0e04mov rax, [rsp+664]
    37b6a0e0cmovsxd rsi, [rax]
    37b6a0e0fxor eax, eax
    37b6a0e11test rsi, rsi
    37b6a0e14cmovns rax, rsi
    37b6a0e18mov [rsp+160], rax
    37b6a0e20not rax
    37b6a0e23mov [rsp+168], rax
    37b6a0e2bmov rax, [rsp+672]
    37b6a0e33mov [rax], 0
    37b6a0e39call slamch_
    37b6a0e3emov r9l, 1
    37b6a0e44mov r8l, 1
    37b6a0e4amov rcx, rbx
    37b6a0e4dlea rdx, [rip+0xd9351d]
    37b6a0e54movaps x11, x0
    37b6a0e58call lsame_
    37b6a0e5dmov ebp, [r15]
    37b6a0e60mov r12l, ebp
    37b6a0e63test eax, eax
    37b6a0e65je 0x37b6a1d50
    37b6a0e6bmov rax, [rsp+608]
    37b6a0e73mov [rsp+372], ebp
    37b6a0e7amov esi, [rax]
    37b6a0e7cmov rax, [rsp+640]
    37b6a0e84mov [rax], 0
    37b6a0e8amov eax, ebp
    37b6a0e8csub eax, esi
    37b6a0e8eadd eax, 1
    37b6a0e91cmp eax, ebp
    37b6a0e93jl 0x37b6a0e9d
    37b6a0e95cmp ebp, esi
    37b6a0e97jg 0x37b6a2e9d
    37b6a0e9d
    37b6a0e9dloc_37b6a0e9d:
    37b6a0e9dlea rax, [rsp+372]
    37b6a0ea5mov [rsp+616], rdi
    37b6a0eadmov edx, r12l
    37b6a0eb0pxor x7, x7
    37b6a0eb4mov [rsp+176], rax
    37b6a0ebclea rax, [rsp+396]
    37b6a0ec4movss x6, [rip+0xd934e4]
    37b6a0eccmov [rsp+128], rax
    37b6a0ed4lea rax, [rip+0xd934a1]
    37b6a0edbmov [rsp+240], rax
    37b6a0ee3lea rax, [rip+0xd9349e]
    37b6a0eeamov [rsp+248], rax
    37b6a0ef2nop [rax]
    37b6a0ef5nop [rax+rax]
    37b6a0f00
    37b6a0f00loc_37b6a0f00:
    37b6a0f00test ebp, ebp
    37b6a0f02jle 0x37b6a10aa
    37b6a0f08movsxd rdi, esi
    37b6a0f0bmov rax, [rsp+168]
    37b6a0f13imul rdi, [rsp+160]
    37b6a0f1cmov [rsp+384], ebp
    37b6a0f23movsxd rbp, ebp
    37b6a0f26mov rcx, [rsp+176]
    37b6a0f2eimul rbp, [rsp+144]
    37b6a0f37add rax, rdi
    37b6a0f3alea r9, [r13+4*rax+4]
    37b6a0f3fmov [rsp+200], rax
    37b6a0f47mov rax, [rsp+136]
    37b6a0f4fmov [rsp+184], r9
    37b6a0f57lea rax, [rax+rbp+1]
    37b6a0f5clea rdx, [r14+4*rax]
    37b6a0f60lea rax, [rip+0xd9340d]
    37b6a0f67mov [rsp+32], rax
    37b6a0f6cmov r8, rax
    37b6a0f6fcall scopy_
    37b6a0f74mov eax, [r15]
    37b6a0f77mov ebx, [rsp+372]
    37b6a0f7ecmp eax, ebx
    37b6a0f80jg 0x37b6a1b50
    37b6a0f86
    37b6a0f86loc_37b6a0f86:
    37b6a0f86mov rax, [rsp+200]
    37b6a0f8emovsxd rbp, ebx
    37b6a0f91add rax, rbp
    37b6a0f94movss x9, [r13+4*rax]
    37b6a0f9bcmp ebx, 1
    37b6a0f9ejg 0x37b6a1278
    37b6a0fa4ucomiss x9, x7
    37b6a0fa8jp 0x37b6a0fb0
    37b6a0faaje 0x37b6a11f8
    37b6a0fb0
    37b6a0fb0loc_37b6a0fb0:
    37b6a0fb0mov rax, [rsp+144]
    37b6a0fb8imul rax, rbp
    37b6a0fbcadd rax, [rsp+136]
    37b6a0fc4mov [rsp+192], rax
    37b6a0fcc
    37b6a0fccloc_37b6a0fcc:
    37b6a0fcclea rax, [rip+0xd933a1]
    37b6a0fd3mov rcx, [rsp+176]
    37b6a0fdbmov rdx, [rsp+184]
    37b6a0fe3lea r8, [rip+0xd9338a]
    37b6a0feamov [rsp+32], rax
    37b6a0fefmov rax, [rsp+192]
    37b6a0ff7lea r9, [r14+4*rax+4]
    37b6a0ffccall scopy_
    37b6a1001movsxd rax, [rsp+372]
    37b6a1009mov rcx, rax
    37b6a100clea ebp, [rax-1]
    37b6a100flea rsi, [rax-1]
    37b6a1013cmp eax, 1
    37b6a1016jle 0x37b6a106f
    37b6a1018mov rdx, [rsp+144]
    37b6a1020imul rdx, rax
    37b6a1024add rdx, [rsp+136]
    37b6a102clea r8, [rdx+rax]
    37b6a1030movss x0, [r14+4*r8]
    37b6a1036movaps x1, x0
    37b6a1039andps x1, [rip+0xd93360]
    37b6a1040comiss x1, x11
    37b6a1044jae 0x37b6a2da0
    37b6a104apxor x1, x1
    37b6a104eucomiss x0, x1
    37b6a1051jp 0x37b6a2d48
    37b6a1057jne 0x37b6a2d48
    37b6a105dnop [rax]
    37b6a1060
    37b6a1060loc_37b6a1060:
    37b6a1060mov rax, [rsp+640]
    37b6a1068mov [rax+4*rsi], 0
    37b6a106f
    37b6a106floc_37b6a106f:
    37b6a106fmov rax, [rsp+648]
    37b6a1077mov [rax+4*rsi], ebx
    37b6a107amov rax, [rsp+608]
    37b6a1082mov ecx, [rax]
    37b6a1084
    37b6a1084loc_37b6a1084:
    37b6a1084mov edx, [r15]
    37b6a1087lea esi, [rcx+rbp]
    37b6a108amov [rsp+372], ebp
    37b6a1091mov eax, edx
    37b6a1093sub esi, edx
    37b6a1095sub eax, ecx
    37b6a1097add eax, 1
    37b6a109acmp eax, ebp
    37b6a109cjl 0x37b6a0f00
    37b6a10a2cmp ecx, edx
    37b6a10a4jge 0x37b6a0f00
    37b6a10aa
    37b6a10aaloc_37b6a10aa:
    37b6a10aamov rdi, [rsp+616]
    37b6a10b2sub edx, ebp
    37b6a10b4
    37b6a10b4loc_37b6a10b4:
    37b6a10b4mov rax, [rsp+632]
    37b6a10bcadd esi, 1
    37b6a10bfmov [rsp+396], edx
    37b6a10c6lea edx, [rbp+1]
    37b6a10c9movsxd rsi, esi
    37b6a10ccmovsxd rdx, edx
    37b6a10cfmov [rsp+88], r14
    37b6a10d4mov r9, [rsp+176]
    37b6a10dcmov [rsp+96], rax
    37b6a10e1mov rax, [rsp+248]
    37b6a10e9lea r8, [rip+0xd9329c]
    37b6a10f0lea rcx, [rip+0xd9329e]
    37b6a10f7imul rsi, [rsp+160]
    37b6a1100mov [rsp+120], 9
    37b6a1109imul rdx, [rsp+144]
    37b6a1112mov [rsp+80], rax
    37b6a1117mov rax, [rsp+664]
    37b6a111fmov [rsp+112], 12
    37b6a1128mov [rsp+104], 5
    37b6a1131mov [rsp+72], rax
    37b6a1136mov rax, [rsp+168]
    37b6a113elea rax, [rax+rsi+1]
    37b6a1143lea rax, [r13+4*rax]
    37b6a1148mov [rsp+64], rax
    37b6a114dmov rax, [rsp+632]
    37b6a1155mov [rsp+56], rax
    37b6a115amov rax, [rsp+136]
    37b6a1162lea rax, [rax+rdx+1]
    37b6a1167mov rdx, [rsp+240]
    37b6a116flea rax, [r14+4*rax]
    37b6a1173mov [rsp+48], rax
    37b6a1178lea rax, [rip+0xd931f9]
    37b6a117fmov [rsp+40], rax
    37b6a1184mov rax, [rsp+128]
    37b6a118cmov [rsp+32], rax
    37b6a1191call sgemmtr_
    37b6a1196mov ebx, [r15]
    37b6a1199sub ebx, [rsp+372]
    37b6a11a0
    37b6a11a0loc_37b6a11a0:
    37b6a11a0mov [rdi], ebx
    37b6a11a2movups x6, [rsp+400]
    37b6a11aamovups x7, [rsp+416]
    37b6a11b2movups x8, [rsp+432]
    37b6a11bbmovups x9, [rsp+448]
    37b6a11c4movups x10, [rsp+464]
    37b6a11cdmovups x11, [rsp+480]
    37b6a11d6movups x12, [rsp+496]
    37b6a11dfadd rsp, 520
    37b6a11e6pop rbx
    37b6a11e7pop rsi
    37b6a11e8pop rdi
    37b6a11e9pop rbp
    37b6a11eapop r12
    37b6a11ecpop r13
    37b6a11eepop r14
    37b6a11f0pop r15
    37b6a11f2ret
    37b6a11f3
    37b6a11f3 0f 1f 44 00 00 ..D..
    37b6a11f8
    37b6a11f8loc_37b6a11f8:
    37b6a11f8mov rax, [rsp+672]
    37b6a1200mov ecx, [rax]
    37b6a1202test ecx, ecx
    37b6a1204jne 0x37b6a1210
    37b6a1206mov rax, [rsp+672]
    37b6a120emov [rax], ebx
    37b6a1210
    37b6a1210loc_37b6a1210:
    37b6a1210mov rax, [rsp+144]
    37b6a1218mov rdi, [rsp+136]
    37b6a1220mov rdx, [rsp+184]
    37b6a1228mov rcx, [rsp+176]
    37b6a1230imul rax, rbp
    37b6a1234lea rax, [rdi+rax+1]
    37b6a1239lea rdi, [rip+0xd93134]
    37b6a1240mov [rsp+32], rdi
    37b6a1245lea r9, [r14+4*rax]
    37b6a1249mov r8, rdi
    37b6a124ccall scopy_
    37b6a1251movsxd rsi, [rsp+372]
    37b6a1259mov rax, rsi
    37b6a125csub rsi, 1
    37b6a1260lea ebp, [rax-1]
    37b6a1263cmp eax, 1
    37b6a1266jg 0x37b6a1060
    37b6a126cjmp 0x37b6a106f
    37b6a1271
    37b6a1271 0f 1f 80 00 00 00 00 .......
    37b6a1278
    37b6a1278loc_37b6a1278:
    37b6a1278mov rdx, [rsp+184]
    37b6a1280mov rcx, [rsp+128]
    37b6a1288lea eax, [rbx-1]
    37b6a128blea r8, [rip+0xd930e2]
    37b6a1292andps x9, [rip+0xd93106]
    37b6a129amov [rsp+396], eax
    37b6a12a1call isamax_
    37b6a12a6mov rcx, [rsp+200]
    37b6a12aemov [rsp+392], eax
    37b6a12b5cdqe
    37b6a12b7movaps x0, x9
    37b6a12bblea rdx, [rax+rcx]
    37b6a12bfmovss x10, [r13+4*rdx]
    37b6a12c6andps x10, [rip+0xd930d2]
    37b6a12cemaxss x0, x10
    37b6a12d3ucomiss x0, x7
    37b6a12d6jp 0x37b6a12de
    37b6a12d8je 0x37b6a11f8
    37b6a12de
    37b6a12deloc_37b6a12de:
    37b6a12demovaps x0, x10
    37b6a12e2mulss x0, x6
    37b6a12e6comiss x0, x9
    37b6a12eajbe 0x37b6a0fb0
    37b6a12f0mov rcx, [rsp+160]
    37b6a12f8add esi, 1
    37b6a12fbmov rdx, [rsp+168]
    37b6a1303movsxd rsi, esi
    37b6a1306imul rsi, rcx
    37b6a130asub rdi, rcx
    37b6a130dadd rdi, rdx
    37b6a1310mov [rsp+152], rdi
    37b6a1318lea rdi, [r13+4*rdi+4]
    37b6a131dmov [rsp+192], rdi
    37b6a1325lea rdi, [rsi+rdx]
    37b6a1329mov [rsp+224], rdi
    37b6a1331lea rdi, [rsp+392]
    37b6a1339mov [rsp+208], rdi
    37b6a1341
    37b6a1341loc_37b6a1341:
    37b6a1341imul rax, [rsp+144]
    37b6a134amov rcx, [rsp+136]
    37b6a1352lea rsi, [rip+0xd9301b]
    37b6a1359mov [rsp+32], rsi
    37b6a135emov r9, [rsp+192]
    37b6a1366mov r8, rsi
    37b6a1369lea rax, [rcx+rax+1]
    37b6a136emov rcx, [rsp+208]
    37b6a1376lea rdx, [r14+4*rax]
    37b6a137acall scopy_
    37b6a137fmovsxd rbp, [rsp+392]
    37b6a1387mov edx, ebx
    37b6a1389mov r8, [rsp+632]
    37b6a1391mov rcx, [rsp+128]
    37b6a1399sub edx, ebp
    37b6a139blea edi, [rbp+1]
    37b6a139emov r12, rbp
    37b6a13a1mov [rsp+396], edx
    37b6a13a8mov rdx, [rsp+152]
    37b6a13b0movsxd rdi, edi
    37b6a13b3lea rax, [rdx+rdi]
    37b6a13b7lea rsi, [r13+4*rax]
    37b6a13bcmov rax, [rsp+144]
    37b6a13c4mov r9, rsi
    37b6a13c7imul rax, rdi
    37b6a13cbadd rax, [rsp+136]
    37b6a13d3add rax, rbp
    37b6a13d6lea rdx, [r14+4*rax]
    37b6a13dalea rax, [rip+0xd92f93]
    37b6a13e1mov [rsp+32], rax
    37b6a13e6mov [rsp+232], rdx
    37b6a13eecall scopy_
    37b6a13f3mov eax, [r15]
    37b6a13f6cmp eax, ebx
    37b6a13f8jg 0x37b6a1a90
    37b6a13fe
    37b6a13feloc_37b6a13fe:
    37b6a13fecmp r12l, ebx
    37b6a1401jne 0x37b6a1748
    37b6a1407cmp r12l, 1
    37b6a140bjg 0x37b6a1c20
    37b6a1411mov [rsp+256], rbp
    37b6a1419mov esi, r12l
    37b6a141c
    37b6a141cloc_37b6a141c:
    37b6a141clea rax, [rip+0xd92f51]
    37b6a1423mov r9, [rsp+184]
    37b6a142bmov rdx, [rsp+192]
    37b6a1433mov [rsp+380], esi
    37b6a143amov [rsp+32], rax
    37b6a143fmov rcx, [rsp+176]
    37b6a1447mov r8, rax
    37b6a144acall scopy_
    37b6a144fmov rax, [rsp+608]
    37b6a1457movsxd r11, [rsp+372]
    37b6a145fmov ebp, [rax]
    37b6a1461mov rax, [rsp+144]
    37b6a1469mov r12, r11
    37b6a146cimul rax, r11
    37b6a1470add rax, [rsp+136]
    37b6a1478add ebp, r11l
    37b6a147bsub ebp, [r15]
    37b6a147emov [rsp+192], rax
    37b6a1486cmp esi, r11l
    37b6a1489je 0x37b6a2e20
    37b6a148fmov [rsp+208], r11
    37b6a1497lea ebx, [r11-1]
    37b6a149bmovsxd rbp, ebp
    37b6a149emov r10l, 1
    37b6a14a4imul rbp, [rsp+160]
    37b6a14admov [rsp+224], ebx
    37b6a14b4mov ebx, r11l
    37b6a14b7add rbp, [rsp+168]
    37b6a14bf
    37b6a14bfloc_37b6a14bf:
    37b6a14bfmov [rsp+280], r10l
    37b6a14c7mov r10, [rsp+192]
    37b6a14cfadd rdi, rax
    37b6a14d2lea r8, [rip+0xd92e9b]
    37b6a14d9mov rcx, [rsp+256]
    37b6a14e1mov [rsp+264], rax
    37b6a14e9lea rdx, [r10+r11]
    37b6a14edmov rax, [rsp+632]
    37b6a14f5mov r9, [rsp+232]
    37b6a14fdmov [rsp+272], r11
    37b6a1505movss x0, [r14+4*rdx]
    37b6a150blea rdx, [rcx+r10]
    37b6a150fmov rcx, [rsp+128]
    37b6a1517movss [r14+4*rdx], x0
    37b6a151dmov edx, [rsp+224]
    37b6a1524mov [rsp+32], rax
    37b6a1529sub edx, esi
    37b6a152bmov [rsp+396], edx
    37b6a1532lea rdx, [r14+4*rdi]
    37b6a1536call scopy_
    37b6a153bmov rdx, [rsp+256]
    37b6a1543imul rdx, [rsp+144]
    37b6a154clea rcx, [rsp+380]
    37b6a1554mov rax, [rsp+136]
    37b6a155clea r8, [rax+rdx+1]
    37b6a1561mov rax, [rsp+264]
    37b6a1569lea r9, [r14+4*r8]
    37b6a156dlea rdx, [r14+4*rax+4]
    37b6a1572mov [rsp+232], rax
    37b6a157alea rax, [rip+0xd92df3]
    37b6a1581mov [rsp+32], rax
    37b6a1586mov r8, rax
    37b6a1589call scopy_
    37b6a158emov eax, [r15]
    37b6a1591movsxd rdi, [rsp+380]
    37b6a1599mov r11, [rsp+272]
    37b6a15a1sub eax, r12l
    37b6a15a4mov rsi, rdi
    37b6a15a7lea edx, [rax+1]
    37b6a15aamov rax, [rsp+232]
    37b6a15b2mov [rsp+232], r11
    37b6a15bamov [rsp+396], edx
    37b6a15c1lea rcx, [rdi+rax]
    37b6a15c5add rax, r11
    37b6a15c8add rdi, rbp
    37b6a15cblea rdx, [r14+4*rax]
    37b6a15cfmov rax, [rsp+632]
    37b6a15d7lea r9, [r14+4*rcx]
    37b6a15dbmov rcx, [rsp+128]
    37b6a15e3mov [rsp+32], rax
    37b6a15e8mov r8, rax
    37b6a15ebcall sswap_
    37b6a15f0mov eax, [r15]
    37b6a15f3lea r9, [r13+4*rdi]
    37b6a15f8mov r11, [rsp+232]
    37b6a1600mov rcx, [rsp+128]
    37b6a1608sub eax, r12l
    37b6a160badd r11, rbp
    37b6a160eadd eax, 1
    37b6a1611lea rdx, [r13+4*r11]
    37b6a1616mov [rsp+396], eax
    37b6a161dmov rax, [rsp+664]
    37b6a1625mov [rsp+32], rax
    37b6a162amov r8, rax
    37b6a162dcall sswap_
    37b6a1632cmp [rsp+280], 1
    37b6a163aje 0x37b6a2e20
    37b6a1640mov rax, [rsp+208]
    37b6a1648add rax, [rsp+192]
    37b6a1650mov [rsp+264], rax
    37b6a1658mov rax, [rsp+608]
    37b6a1660movsxd rdi, [rsp+372]
    37b6a1668movsxd r10, [rsp+224]
    37b6a1670movd x8, [rsp+384]
    37b6a167amov ecx, [rax]
    37b6a167cmov rbp, rdi
    37b6a167f
    37b6a167floc_37b6a167f:
    37b6a167fmovd x0, esi
    37b6a1683mov rax, [rsp+208]
    37b6a168bmov rsi, [rsp+200]
    37b6a1693pxor x3, x3
    37b6a1697punpckldq x0, x8
    37b6a169cmov rdx, [rsp+144]
    37b6a16a4add rax, rsi
    37b6a16a7psubd x3, x0
    37b6a16abmovss x8, [r13+4*rax]
    37b6a16b2mov rax, [rsp+152]
    37b6a16baimul rdx, r10
    37b6a16beadd rdx, [rsp+136]
    37b6a16c6add rax, r10
    37b6a16c9movss x12, [r13+4*rax]
    37b6a16d0lea rax, [rsi+r10]
    37b6a16d4movss x1, [r13+4*rax]
    37b6a16dbcmp ebx, 2
    37b6a16dejg 0x37b6a1c90
    37b6a16e4
    37b6a16e4loc_37b6a16e4:
    37b6a16e4mov rax, [rsp+192]
    37b6a16ecadd rdx, r10
    37b6a16efmov rsi, [rsp+208]
    37b6a16f7sub ebp, 2
    37b6a16famovss [r14+4*rdx], x12
    37b6a1700add rax, r10
    37b6a1703mov [r14+4*rax], 0
    37b6a170bmov rax, [rsp+264]
    37b6a1713movss [r14+4*rax], x8
    37b6a1719mov rax, [rsp+640]
    37b6a1721movss [rax+4*rsi-4], x1
    37b6a1727mov [rax+4*r10-4], 0
    37b6a1730mov rax, [rsp+648]
    37b6a1738movq [rax+4*rdi-8], x3
    37b6a173ejmp 0x37b6a1084
    37b6a1743
    37b6a1743 0f 1f 44 00 00 ..D..
    37b6a1748
    37b6a1748loc_37b6a1748:
    37b6a1748mov rcx, [rsp+128]
    37b6a1750mov eax, ebx
    37b6a1752mov rdx, rsi
    37b6a1755lea r8, [rip+0xd92c18]
    37b6a175csub eax, r12l
    37b6a175fmov [rsp+396], eax
    37b6a1766call isamax_
    37b6a176blea esi, [r12+rax]
    37b6a176fmovsxd rax, esi
    37b6a1772add rax, [rsp+152]
    37b6a177amovss x9, [r13+4*rax]
    37b6a1781andps x9, [rip+0xd92c17]
    37b6a1789cmp r12l, 1
    37b6a178djg 0x37b6a1c80
    37b6a1793
    37b6a1793loc_37b6a1793:
    37b6a1793movaps x1, x9
    37b6a1797mov rax, [rsp+152]
    37b6a179fmulss x1, x6
    37b6a17a3add rax, rbp
    37b6a17a6movss x0, [r13+4*rax]
    37b6a17adandps x0, [rip+0xd92bec]
    37b6a17b4comiss x1, x0
    37b6a17b7jbe 0x37b6a2cd0
    37b6a17bdmov r8l, [rsp+384]
    37b6a17c5cmp r8l, esi
    37b6a17c8je 0x37b6a1828
    37b6a17cacomiss x10, x9
    37b6a17cejae 0x37b6a1828
    37b6a17d0lea rax, [rip+0xd92b9d]
    37b6a17d7mov r9, [rsp+184]
    37b6a17dfmov rdx, [rsp+192]
    37b6a17e7mov [rsp+384], r12l
    37b6a17efmov [rsp+32], rax
    37b6a17f4mov r8, rax
    37b6a17f7movaps x10, x9
    37b6a17fbmov rcx, [rsp+176]
    37b6a1803mov [rsp+392], esi
    37b6a180acall scopy_
    37b6a180fmov [rsp+216], esi
    37b6a1816mov ebx, [rsp+372]
    37b6a181dmovsxd rax, esi
    37b6a1820jmp 0x37b6a1341
    37b6a1825
    37b6a1825 0f 1f 00 ...
    37b6a1828
    37b6a1828loc_37b6a1828:
    37b6a1828mov eax, esi
    37b6a182amov esi, r12l
    37b6a182dlea edx, [rbx-1]
    37b6a1830mov r9, [rsp+144]
    37b6a1838mov r12l, eax
    37b6a183bmov rax, [rsp+608]
    37b6a1843movsxd r11, ebx
    37b6a1846mov [rsp+256], rbp
    37b6a184emov r10, [rsp+136]
    37b6a1856mov [rsp+208], r11
    37b6a185emov ecx, [rax]
    37b6a1860mov [rsp+380], esi
    37b6a1867mov [rsp+224], edx
    37b6a186elea eax, [rdx+rcx]
    37b6a1871sub eax, [r15]
    37b6a1874movsxd rbp, eax
    37b6a1877mov rax, r9
    37b6a187aimul rax, r11
    37b6a187elea r11, [r10+rax]
    37b6a1882movsxd r10, edx
    37b6a1885sub rax, r9
    37b6a1888add rax, [rsp+136]
    37b6a1890mov [rsp+192], r11
    37b6a1898mov r11, r10
    37b6a189bcmp r8l, ebx
    37b6a189eje 0x37b6a2bb8
    37b6a18a4mov edx, ebx
    37b6a18a6lea ecx, [r8+1]
    37b6a18aamov r11, [rsp+136]
    37b6a18b2mov [rsp+288], r10
    37b6a18basub edx, r8l
    37b6a18bdmovsxd rcx, ecx
    37b6a18c0mov [rsp+280], rax
    37b6a18c8mov [rsp+396], edx
    37b6a18cfmov rdx, [rsp+144]
    37b6a18d7mov [rsp+272], r10
    37b6a18dfimul rdx, rcx
    37b6a18e3add rcx, [rsp+192]
    37b6a18eblea r9, [rdx+r11]
    37b6a18efmov [rsp+216], rdx
    37b6a18f7movsxd rdx, r8l
    37b6a18famov r8, [rsp+632]
    37b6a1902add rdx, r9
    37b6a1905mov [rsp+32], r8
    37b6a190alea r9, [r14+4*rdx]
    37b6a190elea rdx, [r14+4*rcx]
    37b6a1912mov rcx, [rsp+128]
    37b6a191alea r8, [rip+0xd92a53]
    37b6a1921call scopy_
    37b6a1926mov r10, [rsp+136]
    37b6a192emov rdx, [rsp+216]
    37b6a1936lea rcx, [rsp+384]
    37b6a193esub rdx, [rsp+144]
    37b6a1946mov rax, [rsp+192]
    37b6a194elea r8, [r10+rdx+1]
    37b6a1953lea r10, [rip+0xd92a1a]
    37b6a195amov [rsp+32], r10
    37b6a195flea rdx, [r14+4*rax+4]
    37b6a1964lea r9, [r14+4*r8]
    37b6a1968mov r8, r10
    37b6a196bcall scopy_
    37b6a1970mov edx, [r15]
    37b6a1973mov rax, [rsp+192]
    37b6a197bmovd x8, [rsp+384]
    37b6a1985sub edx, ebx
    37b6a1987add edx, 1
    37b6a198amov [rsp+396], edx
    37b6a1991movd edx, x8
    37b6a1996movsxd r10, edx
    37b6a1999mov rdx, [rsp+208]
    37b6a19a1lea rcx, [r10+rax]
    37b6a19a5mov [rsp+216], r10
    37b6a19adlea r11, [rdx+rax]
    37b6a19b1mov rax, [rsp+632]
    37b6a19b9lea r9, [r14+4*rcx]
    37b6a19bdmov rcx, [rsp+128]
    37b6a19c5lea rdx, [r14+4*r11]
    37b6a19c9mov [rsp+264], r11
    37b6a19d1mov [rsp+32], rax
    37b6a19d6mov r8, rax
    37b6a19d9call sswap_
    37b6a19demov eax, [r15]
    37b6a19e1sub eax, [rsp+224]
    37b6a19e8lea edx, [rax+1]
    37b6a19ebmov r10, [rsp+216]
    37b6a19f3mov rax, [rsp+664]
    37b6a19fbimul rbp, [rsp+160]
    37b6a1a04mov [rsp+396], edx
    37b6a1a0badd rbp, [rsp+168]
    37b6a1a13mov rdx, [rsp+208]
    37b6a1a1bmov [rsp+32], rax
    37b6a1a20mov r8, rax
    37b6a1a23lea rcx, [r10+rbp]
    37b6a1a27add rdx, rbp
    37b6a1a2alea r9, [r13+4*rcx]
    37b6a1a2fmov rcx, [rsp+128]
    37b6a1a37lea rdx, [r13+4*rdx]
    37b6a1a3ccall sswap_
    37b6a1a41cmp esi, [rsp+224]
    37b6a1a48mov r11, [rsp+272]
    37b6a1a50mov rax, [rsp+280]
    37b6a1a58jne 0x37b6a2cb0
    37b6a1a5emov rax, [rsp+608]
    37b6a1a66mov [rsp+216], r12l
    37b6a1a6emov ebp, ebx
    37b6a1a70mov rdi, [rsp+208]
    37b6a1a78mov r10, [rsp+288]
    37b6a1a80mov ecx, [rax]
    37b6a1a82jmp 0x37b6a167f
    37b6a1a87
    37b6a1a87 66 0f 1f 84 00 00 00 00 00 f........
    37b6a1a90
    37b6a1a90loc_37b6a1a90:
    37b6a1a90sub eax, ebx
    37b6a1a92add ebx, 1
    37b6a1a95mov [rsp+88], 12
    37b6a1a9emov r8, [rsp+128]
    37b6a1aa6mov [rsp+396], eax
    37b6a1aadlea rax, [rip+0xd928c0]
    37b6a1ab4movsxd rbx, ebx
    37b6a1ab7mov rdx, [rsp+176]
    37b6a1abfmov [rsp+80], rax
    37b6a1ac4mov rax, [rsp+192]
    37b6a1acclea r9, [rip+0xd928a5]
    37b6a1ad3imul rbx, [rsp+144]
    37b6a1adcmov rcx, [rsp+240]
    37b6a1ae4mov [rsp+72], rax
    37b6a1ae9mov rax, [rsp+248]
    37b6a1af1mov [rsp+64], rax
    37b6a1af6mov rax, [rsp+664]
    37b6a1afemov [rsp+56], rax
    37b6a1b03mov rax, [rsp+224]
    37b6a1b0badd rax, rbp
    37b6a1b0elea rax, [r13+4*rax]
    37b6a1b13mov [rsp+48], rax
    37b6a1b18mov rax, [rsp+632]
    37b6a1b20mov [rsp+40], rax
    37b6a1b25mov rax, [rsp+136]
    37b6a1b2dlea rax, [rax+rbx+1]
    37b6a1b32lea rax, [r14+4*rax]
    37b6a1b36mov [rsp+32], rax
    37b6a1b3bcall sgemv_
    37b6a1b40mov ebx, [rsp+372]
    37b6a1b47jmp 0x37b6a13fe
    37b6a1b4c
    37b6a1b4c 0f 1f 40 00 ..@.
    37b6a1b50
    37b6a1b50loc_37b6a1b50:
    37b6a1b50sub eax, ebx
    37b6a1b52movsxd rdx, ebx
    37b6a1b55add ebx, 1
    37b6a1b58mov [rsp+88], 12
    37b6a1b61mov [rsp+396], eax
    37b6a1b68lea rax, [rip+0xd92805]
    37b6a1b6fmovsxd rbx, ebx
    37b6a1b72mov r8, [rsp+128]
    37b6a1b7amov [rsp+80], rax
    37b6a1b7fmov rax, [rsp+184]
    37b6a1b87lea r9, [rip+0xd927ea]
    37b6a1b8eimul rbx, [rsp+144]
    37b6a1b97mov rcx, [rsp+240]
    37b6a1b9fmov [rsp+72], rax
    37b6a1ba4mov rax, [rsp+248]
    37b6a1bacmov [rsp+64], rax
    37b6a1bb1mov rax, [rsp+664]
    37b6a1bb9mov [rsp+56], rax
    37b6a1bbemov rax, [rsp+160]
    37b6a1bc6add rax, rdi
    37b6a1bc9add rax, [rsp+168]
    37b6a1bd1add rax, rdx
    37b6a1bd4mov rdx, [rsp+176]
    37b6a1bdclea rax, [r13+4*rax]
    37b6a1be1mov [rsp+48], rax
    37b6a1be6mov rax, [rsp+632]
    37b6a1beemov [rsp+40], rax
    37b6a1bf3mov rax, [rsp+136]
    37b6a1bfblea rax, [rax+rbx+1]
    37b6a1c00lea rax, [r14+4*rax]
    37b6a1c04mov [rsp+32], rax
    37b6a1c09call sgemv_
    37b6a1c0emov ebx, [rsp+372]
    37b6a1c15jmp 0x37b6a0f86
    37b6a1c1a
    37b6a1c1a 66 0f 1f 44 00 00 f..D..
    37b6a1c20
    37b6a1c20loc_37b6a1c20:
    37b6a1c20pxor x8, x8
    37b6a1c25
    37b6a1c25loc_37b6a1c25:
    37b6a1c25mov rdx, [rsp+192]
    37b6a1c2dmov rcx, [rsp+128]
    37b6a1c35lea eax, [r12-1]
    37b6a1c3alea r8, [rip+0xd92733]
    37b6a1c41mov [rsp+396], eax
    37b6a1c48call isamax_
    37b6a1c4dmov esi, eax
    37b6a1c4fcdqe
    37b6a1c51add rax, [rsp+152]
    37b6a1c59movss x9, [r13+4*rax]
    37b6a1c60andps x9, [rip+0xd92738]
    37b6a1c68comiss x9, x8
    37b6a1c6cmaxss x9, x8
    37b6a1c71cmovbe esi, [rsp+216]
    37b6a1c79jmp 0x37b6a1793
    37b6a1c7e
    37b6a1c7e 66 90 f.
    37b6a1c80
    37b6a1c80loc_37b6a1c80:
    37b6a1c80mov [rsp+216], esi
    37b6a1c87movaps x8, x9
    37b6a1c8bjmp 0x37b6a1c25
    37b6a1c8d
    37b6a1c8d 0f 1f 00 ...
    37b6a1c90
    37b6a1c90loc_37b6a1c90:
    37b6a1c90movaps x5, x8
    37b6a1c94movaps x4, x12
    37b6a1c98lea r9, [r14+4*rdx]
    37b6a1c9csub ebx, 1
    37b6a1c9fdivss x5, x1
    37b6a1ca3movss x2, [rip+0xd92709]
    37b6a1cabmov rax, [rsp+152]
    37b6a1cb3lea rsi, [r13+4*rax]
    37b6a1cb8mov rax, [rsp+200]
    37b6a1cc0lea r11, [r13+4*rax]
    37b6a1cc5mov rax, [rsp+192]
    37b6a1ccdlea r8, [r14+4*rax]
    37b6a1cd1mov eax, 1
    37b6a1cd6divss x4, x1
    37b6a1cdamovaps x10, x5
    37b6a1cdemulss x10, x4
    37b6a1ce3subss x10, [rip+0xd926c8]
    37b6a1cecdivss x2, x10
    37b6a1cf1nop [rax]
    37b6a1cf5nop [rax+rax]
    37b6a1d00
    37b6a1d00loc_37b6a1d00:
    37b6a1d00movss x9, [rsi+4*rax]
    37b6a1d06movss x0, [r11+4*rax]
    37b6a1d0cmovaps x10, x9
    37b6a1d10mulss x10, x5
    37b6a1d15subss x10, x0
    37b6a1d1amulss x0, x4
    37b6a1d1esubss x0, x9
    37b6a1d23divss x10, x1
    37b6a1d28divss x0, x1
    37b6a1d2cmulss x10, x2
    37b6a1d31movss [r9+4*rax], x10
    37b6a1d37mulss x0, x2
    37b6a1d3bmovss [r8+4*rax], x0
    37b6a1d41add rax, 1
    37b6a1d45cmp rax, rbx
    37b6a1d48jne 0x37b6a1d00
    37b6a1d4ajmp 0x37b6a16e4
    37b6a1d4f
    37b6a1d4f 90 .
    37b6a1d50
    37b6a1d50loc_37b6a1d50:
    37b6a1d50mov rbx, [rsp+640]
    37b6a1d58movsxd rax, ebp
    37b6a1d5bpxor x6, x6
    37b6a1d5fmov [rsp+616], rdi
    37b6a1d67movss x7, [rip+0xd92641]
    37b6a1d6fmov ebp, 1
    37b6a1d74mov [rbx+4*rax-4], 0
    37b6a1d7clea rax, [rsp+396]
    37b6a1d84mov [rsp+128], rax
    37b6a1d8clea rax, [rip+0xd925e9]
    37b6a1d93mov [rsp+240], rax
    37b6a1d9blea rax, [rip+0xd925e6]
    37b6a1da2mov [rsp+248], rax
    37b6a1daalea rax, [rsp+392]
    37b6a1db2mov [rsp+328], rax
    37b6a1dbanop [rax+rax]
    37b6a1dc0
    37b6a1dc0loc_37b6a1dc0:
    37b6a1dc0mov rax, [rsp+608]
    37b6a1dc8mov [rsp+372], ebp
    37b6a1dcfmov eax, [rax]
    37b6a1dd1cmp eax, ebp
    37b6a1dd3jg 0x37b6a1dde
    37b6a1dd5cmp eax, r12l
    37b6a1dd8jl 0x37b6a2a60
    37b6a1dde
    37b6a1ddeloc_37b6a1dde:
    37b6a1ddecmp ebp, r12l
    37b6a1de1jg 0x37b6a2a60
    37b6a1de7mov rax, [rsp+144]
    37b6a1defmovsxd rsi, ebp
    37b6a1df2sub r12l, ebp
    37b6a1df5lea rdi, [rip+0xd92578]
    37b6a1dfcadd r12l, 1
    37b6a1e00mov [rsp+32], rdi
    37b6a1e05lea rbx, [rsi-1]
    37b6a1e09mov r8, rdi
    37b6a1e0cimul rax, rsi
    37b6a1e10mov rcx, [rsp+136]
    37b6a1e18mov [rsp+396], r12l
    37b6a1e20mov [rsp+304], rsi
    37b6a1e28mov [rsp+192], rbx
    37b6a1e30mov [rsp+296], rax
    37b6a1e38add rax, rcx
    37b6a1e3bmov rcx, [rsp+128]
    37b6a1e43mov [rsp+152], rax
    37b6a1e4badd rax, rsi
    37b6a1e4elea rdx, [r14+4*rax]
    37b6a1e52mov [rsp+256], rax
    37b6a1e5amov rax, [rsp+160]
    37b6a1e62mov [rsp+272], rdx
    37b6a1e6aimul rax, rsi
    37b6a1e6emov [rsp+384], ebp
    37b6a1e75mov [rsp+176], rax
    37b6a1e7dadd rax, [rsp+168]
    37b6a1e85mov [rsp+232], rax
    37b6a1e8dadd rax, rsi
    37b6a1e90lea r12, [r13+4*rax]
    37b6a1e95mov [rsp+208], rax
    37b6a1e9dmov r9, r12
    37b6a1ea0call scopy_
    37b6a1ea5cmp ebp, 1
    37b6a1ea8jle 0x37b6a1f3f
    37b6a1eaemov eax, [r15]
    37b6a1eb1mov [rsp+80], rdi
    37b6a1eb6lea r9, [rip+0xd924bb]
    37b6a1ebdmov rdi, [rsp+248]
    37b6a1ec5mov [rsp+72], r12
    37b6a1ecasub eax, ebp
    37b6a1eccmov r8, [rsp+128]
    37b6a1ed4mov [rsp+88], 12
    37b6a1eddadd eax, 1
    37b6a1ee0mov [rsp+64], rdi
    37b6a1ee5mov rdi, [rsp+664]
    37b6a1eedmov [rsp+392], eax
    37b6a1ef4lea eax, [rbp-1]
    37b6a1ef7mov rcx, [rsp+240]
    37b6a1effmov [rsp+396], eax
    37b6a1f06lea rax, [4*rbx]
    37b6a1f0emov [rsp+56], rdi
    37b6a1f13mov rdi, [rsp+632]
    37b6a1f1blea rdx, [r13+rax]
    37b6a1f20add rax, r14
    37b6a1f23mov [rsp+48], rdx
    37b6a1f28mov rdx, [rsp+328]
    37b6a1f30mov [rsp+40], rdi
    37b6a1f35mov [rsp+32], rax
    37b6a1f3acall sgemv_
    37b6a1f3f
    37b6a1f3floc_37b6a1f3f:
    37b6a1f3fmov rax, [rsp+208]
    37b6a1f47lea edi, [rbp+1]
    37b6a1f4amovsxd rcx, edi
    37b6a1f4dmov [rsp+264], edi
    37b6a1f54movss x8, [r13+4*rax]
    37b6a1f5bmov eax, [r15]
    37b6a1f5emov [rsp+200], rcx
    37b6a1f66cmp eax, ebp
    37b6a1f68jg 0x37b6a2060
    37b6a1f6eucomiss x8, x6
    37b6a1f72jp 0x37b6a2010
    37b6a1f78jne 0x37b6a2010
    37b6a1f7e
    37b6a1f7eloc_37b6a1f7e:
    37b6a1f7emov rbx, [rsp+672]
    37b6a1f86mov edx, [rbx]
    37b6a1f88test edx, edx
    37b6a1f8ajne 0x37b6a1f8e
    37b6a1f8cmov [rbx], ebp
    37b6a1f8e
    37b6a1f8eloc_37b6a1f8e:
    37b6a1f8esub eax, ebp
    37b6a1f90mov r9, [rsp+272]
    37b6a1f98mov rcx, [rsp+128]
    37b6a1fa0mov rdx, r12
    37b6a1fa3add eax, 1
    37b6a1fa6mov [rsp+396], eax
    37b6a1fadlea rax, [rip+0xd923c0]
    37b6a1fb4mov [rsp+32], rax
    37b6a1fb9mov r8, rax
    37b6a1fbccall scopy_
    37b6a1fc1cmp [r15], ebp
    37b6a1fc4jle 0x37b6a1fdd
    37b6a1fc6mov rax, [rsp+640]
    37b6a1fcemov rdi, [rsp+192]
    37b6a1fd6mov [rax+4*rdi], 0
    37b6a1fdd
    37b6a1fddloc_37b6a1fdd:
    37b6a1fddmov edi, ebp
    37b6a1fdfmov r12l, [r15]
    37b6a1fe2mov ebp, [rsp+264]
    37b6a1fe9
    37b6a1fe9loc_37b6a1fe9:
    37b6a1fe9mov rax, [rsp+648]
    37b6a1ff1mov rbx, [rsp+192]
    37b6a1ff9mov [rax+4*rbx], edi
    37b6a1ffcjmp 0x37b6a1dc0
    37b6a2001
    37b6a2001 0f 1f 80 00 00 00 00 .......
    37b6a2008
    37b6a2008loc_37b6a2008:
    37b6a2008mov eax, [r15]
    37b6a200bnop [rax+rax]
    37b6a2010
    37b6a2010loc_37b6a2010:
    37b6a2010mov edi, ebp
    37b6a2012
    37b6a2012loc_37b6a2012:
    37b6a2012sub eax, ebp
    37b6a2014mov r9, [rsp+272]
    37b6a201cmov rcx, [rsp+128]
    37b6a2024mov rdx, r12
    37b6a2027add eax, 1
    37b6a202amov [rsp+396], eax
    37b6a2031lea rax, [rip+0xd9233c]
    37b6a2038mov [rsp+32], rax
    37b6a203dmov r8, rax
    37b6a2040call scopy_
    37b6a2045mov r12l, [r15]
    37b6a2048cmp r12l, ebp
    37b6a204bjg 0x37b6a2bf8
    37b6a2051
    37b6a2051loc_37b6a2051:
    37b6a2051mov ebp, [rsp+264]
    37b6a2058jmp 0x37b6a1fe9
    37b6a205a
    37b6a205a 66 0f 1f 44 00 00 f..D..
    37b6a2060
    37b6a2060loc_37b6a2060:
    37b6a2060sub eax, ebp
    37b6a2062mov rdi, [rsp+200]
    37b6a206amov rcx, [rsp+128]
    37b6a2072lea r8, [rip+0xd922fb]
    37b6a2079mov [rsp+396], eax
    37b6a2080mov rax, [rsp+232]
    37b6a2088andps x8, [rip+0xd92310]
    37b6a2090lea r9, [rax+rdi]
    37b6a2094lea rdx, [r13+4*r9]
    37b6a2099mov [rsp+184], r9
    37b6a20a1call isamax_
    37b6a20a6movaps x0, x8
    37b6a20aamov r9, [rsp+184]
    37b6a20b2lea edi, [rax+rbp]
    37b6a20b5mov rax, [rsp+232]
    37b6a20bdmovsxd rbx, edi
    37b6a20c0add rax, rbx
    37b6a20c3movss x9, [r13+4*rax]
    37b6a20caandps x9, [rip+0xd922ce]
    37b6a20d2maxss x0, x9
    37b6a20d7ucomiss x0, x6
    37b6a20dajp 0x37b6a20f0
    37b6a20dcjne 0x37b6a20f0
    37b6a20demov eax, [r15]
    37b6a20e1jmp 0x37b6a1f7e
    37b6a20e6
    37b6a20e6 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b6a20f0
    37b6a20f0loc_37b6a20f0:
    37b6a20f0movaps x0, x9
    37b6a20f4mulss x0, x7
    37b6a20f8comiss x0, x8
    37b6a20fcjbe 0x37b6a2008
    37b6a2102mov rcx, [rsp+168]
    37b6a210amov rax, [rsp+176]
    37b6a2112mov [rsp+352], r9
    37b6a211aadd rax, [rsp+160]
    37b6a2122mov [rsp+360], rsi
    37b6a212alea rdx, [rax+rcx]
    37b6a212elea ecx, [rbp-1]
    37b6a2131mov [rsp+336], r12
    37b6a2139lea rax, [rdx+rsi]
    37b6a213dmov [rsp+184], rdx
    37b6a2145lea rax, [r13+4*rax]
    37b6a214amov [rsp+176], rax
    37b6a2152lea rax, [r14+4*rsi-4]
    37b6a2157mov [rsp+344], rax
    37b6a215fmov eax, [rsp+216]
    37b6a2166mov [rsp+216], ecx
    37b6a216dmov [rsp+324], eax
    37b6a2174mov r12l, eax
    37b6a2177
    37b6a2177loc_37b6a2177:
    37b6a2177mov eax, edi
    37b6a2179mov r9, [rsp+176]
    37b6a2181mov r8, [rsp+632]
    37b6a2189sub eax, ebp
    37b6a218bmov rcx, [rsp+128]
    37b6a2193mov [rsp+224], eax
    37b6a219amov [rsp+396], eax
    37b6a21a1mov rax, [rsp+152]
    37b6a21a9add rax, rbx
    37b6a21aclea rdx, [r14+4*rax]
    37b6a21b0mov [rsp+288], rax
    37b6a21b8lea rax, [rip+0xd921b5]
    37b6a21bfmov [rsp+32], rax
    37b6a21c4mov [rsp+280], rdx
    37b6a21cccall scopy_
    37b6a21d1mov esi, [r15]
    37b6a21d4mov rdx, [rsp+144]
    37b6a21dcmov rcx, [rsp+128]
    37b6a21e4sub esi, edi
    37b6a21e6imul rdx, rbx
    37b6a21ealea eax, [rsi+1]
    37b6a21edmov rsi, [rsp+184]
    37b6a21f5mov [rsp+396], eax
    37b6a21fcmov rax, [rsp+136]
    37b6a2204add rsi, rbx
    37b6a2207add rax, rdx
    37b6a220alea r9, [r13+4*rsi]
    37b6a220fadd rax, rbx
    37b6a2212lea rdx, [r14+4*rax]
    37b6a2216lea rax, [rip+0xd92157]
    37b6a221dmov [rsp+32], rax
    37b6a2222mov r8, rax
    37b6a2225mov [rsp+312], rdx
    37b6a222dcall scopy_
    37b6a2232cmp ebp, 1
    37b6a2235jle 0x37b6a22dc
    37b6a223bmov edx, [r15]
    37b6a223emov r8, [rsp+128]
    37b6a2246mov [rsp+88], 12
    37b6a224flea r9, [rip+0xd92122]
    37b6a2256mov rcx, [rsp+240]
    37b6a225esub edx, ebp
    37b6a2260lea eax, [rdx+1]
    37b6a2263mov edx, [rsp+216]
    37b6a226amov [rsp+392], eax
    37b6a2271lea rax, [rip+0xd920fc]
    37b6a2278mov [rsp+80], rax
    37b6a227dmov rax, [rsp+176]
    37b6a2285mov [rsp+396], edx
    37b6a228cmov rdx, [rsp+328]
    37b6a2294mov [rsp+72], rax
    37b6a2299mov rax, [rsp+248]
    37b6a22a1mov [rsp+64], rax
    37b6a22a6mov rax, [rsp+664]
    37b6a22aemov [rsp+56], rax
    37b6a22b3lea rax, [r13+4*rbx-4]
    37b6a22b8mov [rsp+48], rax
    37b6a22bdmov rax, [rsp+632]
    37b6a22c5mov [rsp+40], rax
    37b6a22camov rax, [rsp+344]
    37b6a22d2mov [rsp+32], rax
    37b6a22d7call sgemv_
    37b6a22dc
    37b6a22dcloc_37b6a22dc:
    37b6a22dccmp ebp, edi
    37b6a22dejne 0x37b6a2340
    37b6a22e0mov eax, [r15]
    37b6a22e3cmp eax, ebp
    37b6a22e5jg 0x37b6a2908
    37b6a22ebsub eax, ebp
    37b6a22edmov [rsp+216], r12l
    37b6a22f5mov r12, [rsp+336]
    37b6a22fdadd eax, 1
    37b6a2300mov rdx, [rsp+176]
    37b6a2308mov rcx, [rsp+128]
    37b6a2310mov [rsp+396], eax
    37b6a2317lea rax, [rip+0xd92056]
    37b6a231emov r9, r12
    37b6a2321mov [rsp+32], rax
    37b6a2326mov r8, rax
    37b6a2329call scopy_
    37b6a232e
    37b6a232eloc_37b6a232e:
    37b6a232emov eax, [r15]
    37b6a2331jmp 0x37b6a2012
    37b6a2336
    37b6a2336 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b6a2340
    37b6a2340loc_37b6a2340:
    37b6a2340mov eax, [rsp+224]
    37b6a2347mov rcx, [rsp+128]
    37b6a234flea r8, [rip+0xd9201e]
    37b6a2356mov rdx, [rsp+176]
    37b6a235emov [rsp+396], eax
    37b6a2365call isamax_
    37b6a236amov ecx, [rsp+216]
    37b6a2371lea r12l, [rcx+rax]
    37b6a2375movsxd rax, r12l
    37b6a2378add rax, [rsp+184]
    37b6a2380movss x8, [r13+4*rax]
    37b6a2387mov eax, [r15]
    37b6a238aandps x8, [rip+0xd9200e]
    37b6a2392cmp eax, edi
    37b6a2394jg 0x37b6a2978
    37b6a239a
    37b6a239aloc_37b6a239a:
    37b6a239amovaps x1, x8
    37b6a239emovss x0, [r13+4*rsi]
    37b6a23a5andps x0, [rip+0xd91ff4]
    37b6a23acmulss x1, x7
    37b6a23b0comiss x1, x0
    37b6a23b3jbe 0x37b6a2858
    37b6a23b9mov eax, [rsp+384]
    37b6a23c0cmp eax, r12l
    37b6a23c3je 0x37b6a2428
    37b6a23c5comiss x9, x8
    37b6a23c9jae 0x37b6a2428
    37b6a23cbmov eax, [r15]
    37b6a23cemov r9, [rsp+336]
    37b6a23d6mov [rsp+384], edi
    37b6a23ddmovaps x9, x8
    37b6a23e1mov rdx, [rsp+176]
    37b6a23e9mov rcx, [rsp+128]
    37b6a23f1mov edi, r12l
    37b6a23f4movsxd rbx, r12l
    37b6a23f7sub eax, ebp
    37b6a23f9add eax, 1
    37b6a23fcmov [rsp+396], eax
    37b6a2403lea rax, [rip+0xd91f6a]
    37b6a240amov [rsp+32], rax
    37b6a240fmov r8, rax
    37b6a2412call scopy_
    37b6a2417mov [rsp+324], r12l
    37b6a241fjmp 0x37b6a2177
    37b6a2424
    37b6a2424 0f 1f 40 00 ..@.
    37b6a2428
    37b6a2428loc_37b6a2428:
    37b6a2428mov ecx, [rsp+264]
    37b6a242fmov [rsp+216], r12l
    37b6a2437mov [rsp+380], edi
    37b6a243emov r9, [rsp+352]
    37b6a2446mov [rsp+376], ecx
    37b6a244dmov rdx, [rsp+184]
    37b6a2455mov rsi, [rsp+360]
    37b6a245dcmp eax, ebp
    37b6a245fje 0x37b6a2b50
    37b6a2465mov edx, eax
    37b6a2467movsxd rsi, eax
    37b6a246amov [rsp+176], eax
    37b6a2471mov rax, [rsp+632]
    37b6a2479sub edx, ebp
    37b6a247bmov r12, [rsp+128]
    37b6a2483lea r8, [rip+0xd91eea]
    37b6a248amov [rsp+396], edx
    37b6a2491mov rdx, [rsp+152]
    37b6a2499mov [rsp+32], rax
    37b6a249emov rcx, r12
    37b6a24a1add rdx, rsi
    37b6a24a4lea rbp, [r14+4*rdx]
    37b6a24a8mov rdx, [rsp+272]
    37b6a24b0mov r9, rbp
    37b6a24b3call scopy_
    37b6a24b8mov edx, [r15]
    37b6a24bbsub edx, [rsp+176]
    37b6a24c2lea rcx, [rip+0xd91eab]
    37b6a24c9add edx, 1
    37b6a24ccmov [rsp+32], rcx
    37b6a24d1mov r8, rcx
    37b6a24d4mov rcx, r12
    37b6a24d7mov [rsp+396], edx
    37b6a24demov rdx, [rsp+144]
    37b6a24e6lea r12, [rsp+376]
    37b6a24eeimul rdx, rsi
    37b6a24f2add rdx, [rsp+136]
    37b6a24faadd rdx, rsi
    37b6a24fdlea rsi, [4*rsi-4]
    37b6a2505lea r9, [r14+4*rdx]
    37b6a2509mov rdx, rbp
    37b6a250ccall scopy_
    37b6a2511mov r11, [rsp+632]
    37b6a2519lea r9, [r14+rsi]
    37b6a251dmov rcx, [rsp+192]
    37b6a2525mov [rsp+32], r11
    37b6a252amov r8, r11
    37b6a252dlea rdx, [r14+4*rcx]
    37b6a2531lea rcx, [rsp+372]
    37b6a2539call sswap_
    37b6a253emovsxd rbp, [rsp+372]
    37b6a2546lea r9, [r13+rsi]
    37b6a254blea rcx, [rbp-1]
    37b6a254fmov [rsp+152], ebp
    37b6a2556lea rdx, [r13+4*rcx]
    37b6a255bmov [rsp+192], rcx
    37b6a2563mov rcx, [rsp+664]
    37b6a256bmov [rsp+304], rbp
    37b6a2573mov [rsp+32], rcx
    37b6a2578mov r8, rcx
    37b6a257bmov rcx, r12
    37b6a257ecall sswap_
    37b6a2583mov rsi, [rsp+160]
    37b6a258bmov r9, [rsp+144]
    37b6a2593mov rax, rbp
    37b6a2596movsxd r10, [rsp+152]
    37b6a259emov ecx, [rsp+376]
    37b6a25a5imul rsi, rbp
    37b6a25a9mov r8, r9
    37b6a25acimul r8, rbp
    37b6a25b0lea r11l, [r10+1]
    37b6a25b4mov [rsp+264], r11l
    37b6a25bcmovsxd r11, r11l
    37b6a25bfmov rdx, rsi
    37b6a25c2mov rsi, [rsp+168]
    37b6a25camov [rsp+200], r11
    37b6a25d2add rsi, rdx
    37b6a25d5mov [rsp+232], rsi
    37b6a25ddadd rsi, rbp
    37b6a25e0mov rbp, [rsp+136]
    37b6a25e8mov [rsp+208], rsi
    37b6a25f0lea rsi, [rbp+r8]
    37b6a25f5add r8, r9
    37b6a25f8add rax, rsi
    37b6a25fbcmp ecx, edi
    37b6a25fdmov [rsp+152], rsi
    37b6a2605lea r11, [r8+rbp]
    37b6a2609mov [rsp+256], rax
    37b6a2611mov eax, [rsp+176]
    37b6a2618je 0x37b6a2c50
    37b6a261emovsxd rdx, ecx
    37b6a2621lea rcx, [rbx+rsi]
    37b6a2625lea rax, [rdx+rsi]
    37b6a2629lea rsi, [rdx-1]
    37b6a262dmov [rsp+288], rcx
    37b6a2635mov ecx, edi
    37b6a2637imul rdx, r9
    37b6a263bsub ecx, r10l
    37b6a263emov [rsp+224], ecx
    37b6a2645add rdx, rbp
    37b6a2648mov ebp, r10l
    37b6a264bmov r10l, 2
    37b6a2651lea rcx, [rbx+rdx]
    37b6a2655lea rcx, [r14+4*rcx]
    37b6a2659mov [rsp+280], rcx
    37b6a2661lea rcx, [rbx+r11]
    37b6a2665lea r9, [r14+4*rcx]
    37b6a2669nop [rax]
    37b6a2670
    37b6a2670loc_37b6a2670:
    37b6a2670movss x0, [r14+4*rax]
    37b6a2676mov rcx, [rsp+632]
    37b6a267elea r8, [rip+0xd91cef]
    37b6a2685mov [rsp+184], r11
    37b6a268dmov rax, [rsp+288]
    37b6a2695add rdx, [rsp+200]
    37b6a269dmov [rsp+176], r10l
    37b6a26a5lea rbx, [4*rbx-4]
    37b6a26adlea rdx, [r14+4*rdx]
    37b6a26b1movss [r14+4*rax], x0
    37b6a26b7mov eax, [rsp+224]
    37b6a26bemov [rsp+32], rcx
    37b6a26c3mov rcx, [rsp+128]
    37b6a26cbsub eax, 1
    37b6a26cemov [rsp+396], eax
    37b6a26d5call scopy_
    37b6a26damov eax, [r15]
    37b6a26ddlea rdx, [rip+0xd91c90]
    37b6a26e4mov r9, [rsp+312]
    37b6a26ecmov [rsp+32], rdx
    37b6a26f1mov r8, rdx
    37b6a26f4mov rcx, [rsp+128]
    37b6a26fcmov rdx, [rsp+280]
    37b6a2704sub eax, edi
    37b6a2706add eax, 1
    37b6a2709mov [rsp+396], eax
    37b6a2710call scopy_
    37b6a2715lea rdx, [r14+4*rsi]
    37b6a2719lea r9, [r14+rbx]
    37b6a271dmov rcx, [rsp+632]
    37b6a2725mov [rsp+32], rcx
    37b6a272amov r8, rcx
    37b6a272dmov rcx, r12
    37b6a2730call sswap_
    37b6a2735mov rsi, [rsp+664]
    37b6a273dlea r9, [r13+rbx]
    37b6a2742mov rcx, r12
    37b6a2745movsxd rax, [rsp+376]
    37b6a274dmov [rsp+32], rsi
    37b6a2752mov r8, rsi
    37b6a2755lea rdx, [r13+4*rax-4]
    37b6a275acall sswap_
    37b6a275fcmp [rsp+176], 1
    37b6a2767je 0x37b6a2de8
    37b6a276dmov rbx, [rsp+200]
    37b6a2775mov rcx, [rsp+232]
    37b6a277dmovsxd r10, ebp
    37b6a2780mov rdx, [rsp+160]
    37b6a2788mov eax, [rsp+384]
    37b6a278fmov r11, [rsp+184]
    37b6a2797lea r9, [rbx+rcx]
    37b6a279bimul rdx, rbx
    37b6a279fadd rdx, [rsp+168]
    37b6a27a7
    37b6a27a7loc_37b6a27a7:
    37b6a27a7mov rcx, [rsp+200]
    37b6a27afmov r12l, [r15]
    37b6a27b2lea ebp, [r10+2]
    37b6a27b6mov rbx, [rsp+208]
    37b6a27bemovss x2, [r13+4*r9]
    37b6a27c5add rcx, rdx
    37b6a27c8movss x9, [r13+4*rcx]
    37b6a27cflea ecx, [r12-1]
    37b6a27d4movss x10, [r13+4*rbx]
    37b6a27dbcmp ecx, r10l
    37b6a27dejg 0x37b6a2990
    37b6a27e4
    37b6a27e4loc_37b6a27e4:
    37b6a27e4mov rbx, [rsp+200]
    37b6a27ecneg eax
    37b6a27eeneg edi
    37b6a27f0mov rdx, [rsp+152]
    37b6a27f8mov rsi, [rsp+256]
    37b6a2800mov rcx, [rsp+640]
    37b6a2808add rdx, rbx
    37b6a280bmovss [r14+4*rsi], x10
    37b6a2811mov rsi, rcx
    37b6a2814mov [r14+4*rdx], 0
    37b6a281clea rdx, [rbx+r11]
    37b6a2820mov rbx, [rsp+192]
    37b6a2828movss [r14+4*rdx], x9
    37b6a282emovss [rcx+4*rbx], x2
    37b6a2833mov rcx, [rsp+304]
    37b6a283bmov [rsi+4*rcx], 0
    37b6a2842mov rsi, [rsp+648]
    37b6a284amov [rsi+4*rbx], eax
    37b6a284dmov [rsi+4*rcx], edi
    37b6a2850jmp 0x37b6a1dc0
    37b6a2855
    37b6a2855 0f 1f 00 ...
    37b6a2858
    37b6a2858loc_37b6a2858:
    37b6a2858mov eax, [r15]
    37b6a285bmov [rsp+216], r12l
    37b6a2863mov r12, [rsp+336]
    37b6a286bmov rdx, [rsp+176]
    37b6a2873mov [rsp+380], edi
    37b6a287asub eax, ebp
    37b6a287cmov rcx, [rsp+128]
    37b6a2884add eax, 1
    37b6a2887mov r9, r12
    37b6a288amov [rsp+396], eax
    37b6a2891lea rax, [rip+0xd91adc]
    37b6a2898mov [rsp+32], rax
    37b6a289dmov r8, rax
    37b6a28a0call scopy_
    37b6a28a5mov [rsp+376], ebp
    37b6a28accmp ebp, edi
    37b6a28aeje 0x37b6a232e
    37b6a28b4mov rsi, [rsp+136]
    37b6a28bcmov rax, [rsp+296]
    37b6a28c4mov r10l, 1
    37b6a28calea r12, [rsp+376]
    37b6a28d2add rax, [rsp+144]
    37b6a28damov rdx, [rsp+152]
    37b6a28e2lea r11, [rax+rsi]
    37b6a28e6mov rsi, [rsp+192]
    37b6a28eelea rax, [rbx+r11]
    37b6a28f2lea r9, [r14+4*rax]
    37b6a28f6mov rax, [rsp+256]
    37b6a28fejmp 0x37b6a2670
    37b6a2903
    37b6a2903 0f 1f 44 00 00 ..D..
    37b6a2908
    37b6a2908loc_37b6a2908:
    37b6a2908pxor x10, x10
    37b6a290d
    37b6a290dloc_37b6a290d:
    37b6a290dsub eax, edi
    37b6a290fmov rcx, [rsp+128]
    37b6a2917lea r8, [rip+0xd91a56]
    37b6a291emov [rsp+396], eax
    37b6a2925lea eax, [rdi+1]
    37b6a2928cdqe
    37b6a292aadd rax, [rsp+184]
    37b6a2932lea rdx, [r13+4*rax]
    37b6a2937call isamax_
    37b6a293clea r12l, [rax+rdi]
    37b6a2940movsxd rax, r12l
    37b6a2943add rax, [rsp+184]
    37b6a294bmovss x8, [r13+4*rax]
    37b6a2952andps x8, [rip+0xd91a46]
    37b6a295acomiss x8, x10
    37b6a295emaxss x8, x10
    37b6a2963cmovbe r12l, [rsp+324]
    37b6a296cjmp 0x37b6a239a
    37b6a2971
    37b6a2971 0f 1f 80 00 00 00 00 .......
    37b6a2978
    37b6a2978loc_37b6a2978:
    37b6a2978mov [rsp+324], r12l
    37b6a2980movaps x10, x8
    37b6a2984jmp 0x37b6a290d
    37b6a2986
    37b6a2986 66 2e 0f 1f 84 00 00 00 00 00 f.........
    37b6a2990
    37b6a2990loc_37b6a2990:
    37b6a2990cmp r12l, ebp
    37b6a2993jl 0x37b6a27e4
    37b6a2999movaps x5, x9
    37b6a299dmovaps x4, x10
    37b6a29a1lea ecx, [r12-2]
    37b6a29a6add rdx, r10
    37b6a29a9divss x5, x2
    37b6a29admovss x3, [rip+0xd919ff]
    37b6a29b5sub ecx, r10l
    37b6a29b8lea rbx, [r13+4*rdx]
    37b6a29bdmov rsi, [rsp+152]
    37b6a29c5lea r8, [4*rcx+4]
    37b6a29cdmov rcx, [rsp+232]
    37b6a29d5lea rdx, [rsi+r10]
    37b6a29d9add rcx, r10
    37b6a29dclea r9, [r13+4*rcx]
    37b6a29e1lea rcx, [r14+4*rdx]
    37b6a29e5lea rdx, [r11+r10]
    37b6a29e9lea rsi, [r14+4*rdx]
    37b6a29edxor edx, edx
    37b6a29efdivss x4, x2
    37b6a29f3movaps x1, x5
    37b6a29f6mulss x1, x4
    37b6a29fasubss x1, [rip+0xd919b2]
    37b6a2a02divss x3, x1
    37b6a2a06nop [rax+rax]
    37b6a2a10
    37b6a2a10loc_37b6a2a10:
    37b6a2a10movss x8, [r9+rdx+8]
    37b6a2a17movss x0, [rbx+rdx+8]
    37b6a2a1dmovaps x1, x8
    37b6a2a21mulss x1, x5
    37b6a2a25subss x1, x0
    37b6a2a29mulss x0, x4
    37b6a2a2dsubss x0, x8
    37b6a2a32divss x1, x2
    37b6a2a36divss x0, x2
    37b6a2a3amulss x1, x3
    37b6a2a3emovss [rcx+rdx+8], x1
    37b6a2a44mulss x0, x3
    37b6a2a48movss [rsi+rdx+8], x0
    37b6a2a4eadd rdx, 4
    37b6a2a52cmp r8, rdx
    37b6a2a55jne 0x37b6a2a10
    37b6a2a57jmp 0x37b6a27e4
    37b6a2a5c
    37b6a2a5c 0f 1f 40 00 ..@.
    37b6a2a60
    37b6a2a60loc_37b6a2a60:
    37b6a2a60mov rdx, [rsp+144]
    37b6a2a68movsxd rax, ebp
    37b6a2a6bsub r12l, ebp
    37b6a2a6elea ebx, [rbp-1]
    37b6a2a71lea rcx, [4*rax-4]
    37b6a2a79mov rsi, [rsp+632]
    37b6a2a81add r12l, 1
    37b6a2a85mov r9, [rsp+328]
    37b6a2a8dimul rdx, rax
    37b6a2a91add rdx, [rsp+136]
    37b6a2a99lea r8, [rip+0xd918ec]
    37b6a2aa0mov [rsp+392], r12l
    37b6a2aa8add rax, rdx
    37b6a2aabmov [rsp+396], ebx
    37b6a2ab2mov rdx, [rsp+240]
    37b6a2abalea rax, [r14+4*rax]
    37b6a2abemov [rsp+96], rsi
    37b6a2ac3mov rdi, [rsp+616]
    37b6a2acbmov [rsp+88], rax
    37b6a2ad0mov rax, [rsp+248]
    37b6a2ad8mov [rsp+120], 9
    37b6a2ae1mov [rsp+80], rax
    37b6a2ae6mov rax, [rsp+664]
    37b6a2aeemov [rsp+112], 12
    37b6a2af7mov [rsp+72], rax
    37b6a2afclea rax, [r13+rcx]
    37b6a2b01mov [rsp+64], rax
    37b6a2b06mov rax, [rsp+632]
    37b6a2b0emov [rsp+104], 5
    37b6a2b17mov [rsp+56], rax
    37b6a2b1clea rax, [r14+rcx]
    37b6a2b20lea rcx, [rip+0xd91873]
    37b6a2b27mov [rsp+48], rax
    37b6a2b2clea rax, [rip+0xd91845]
    37b6a2b33mov [rsp+40], rax
    37b6a2b38mov rax, [rsp+128]
    37b6a2b40mov [rsp+32], rax
    37b6a2b45call sgemmtr_
    37b6a2b4ajmp 0x37b6a11a0
    37b6a2b4f
    37b6a2b4f 90 .
    37b6a2b50
    37b6a2b50loc_37b6a2b50:
    37b6a2b50mov r11, [rsp+136]
    37b6a2b58mov rcx, [rsp+296]
    37b6a2b60movsxd r10, ebp
    37b6a2b63add rcx, [rsp+144]
    37b6a2b6badd r11, rcx
    37b6a2b6ecmp edi, [rsp+264]
    37b6a2b75je 0x37b6a27a7
    37b6a2b7blea rdx, [rbx+r11]
    37b6a2b7fmov rax, [rsp+200]
    37b6a2b87mov r10l, 2
    37b6a2b8dadd rax, [rsp+152]
    37b6a2b95lea rcx, [r14+4*rdx]
    37b6a2b99lea r12, [rsp+376]
    37b6a2ba1mov rdx, r11
    37b6a2ba4mov [rsp+280], rcx
    37b6a2bacmov r9, rcx
    37b6a2bafjmp 0x37b6a2670
    37b6a2bb4
    37b6a2bb4 0f 1f 40 00 ..@.
    37b6a2bb8
    37b6a2bb8loc_37b6a2bb8:
    37b6a2bb8cmp esi, [rsp+224]
    37b6a2bbfjne 0x37b6a2c80
    37b6a2bc5mov rdi, [rsp+208]
    37b6a2bcdmov rax, [rsp+192]
    37b6a2bd5mov ebp, ebx
    37b6a2bd7movd x8, ebx
    37b6a2bdcmov [rsp+216], r12l
    37b6a2be4add rax, rdi
    37b6a2be7mov [rsp+264], rax
    37b6a2befjmp 0x37b6a167f
    37b6a2bf4
    37b6a2bf4 0f 1f 40 00 ..@.
    37b6a2bf8
    37b6a2bf8loc_37b6a2bf8:
    37b6a2bf8mov rax, [rsp+256]
    37b6a2c00movss x0, [r14+4*rax]
    37b6a2c06movaps x1, x0
    37b6a2c09andps x1, [rip+0xd91790]
    37b6a2c10comiss x1, x11
    37b6a2c14jae 0x37b6a2cf0
    37b6a2c1apxor x1, x1
    37b6a2c1eucomiss x0, x1
    37b6a2c21jp 0x37b6a2e27
    37b6a2c27jne 0x37b6a2e27
    37b6a2c2d
    37b6a2c2dloc_37b6a2c2d:
    37b6a2c2dmov rax, [rsp+640]
    37b6a2c35mov rbx, [rsp+192]
    37b6a2c3dmov [rax+4*rbx], 0
    37b6a2c44jmp 0x37b6a2051
    37b6a2c49
    37b6a2c49 0f 1f 80 00 00 00 00 .......
    37b6a2c50
    37b6a2c50loc_37b6a2c50:
    37b6a2c50mov rbx, [rsp+232]
    37b6a2c58mov rsi, [rsp+200]
    37b6a2c60add rdx, [rsp+160]
    37b6a2c68add rdx, [rsp+168]
    37b6a2c70lea r9, [rbx+rsi]
    37b6a2c74jmp 0x37b6a27a7
    37b6a2c79
    37b6a2c79 0f 1f 80 00 00 00 00 .......
    37b6a2c80
    37b6a2c80loc_37b6a2c80:
    37b6a2c80imul rbp, [rsp+160]
    37b6a2c89mov [rsp+216], r12l
    37b6a2c91mov r10l, 2
    37b6a2c97mov r12l, [rsp+224]
    37b6a2c9fadd rbp, [rsp+168]
    37b6a2ca7jmp 0x37b6a14bf
    37b6a2cac
    37b6a2cac 0f 1f 40 00 ..@.
    37b6a2cb0
    37b6a2cb0loc_37b6a2cb0:
    37b6a2cb0mov [rsp+216], r12l
    37b6a2cb8mov r10l, 2
    37b6a2cbemov r12l, [rsp+224]
    37b6a2cc6jmp 0x37b6a14bf
    37b6a2ccb
    37b6a2ccb 0f 1f 44 00 00 ..D..
    37b6a2cd0
    37b6a2cd0loc_37b6a2cd0:
    37b6a2cd0mov eax, esi
    37b6a2cd2mov [rsp+256], rbp
    37b6a2cdamov esi, r12l
    37b6a2cddmov [rsp+216], eax
    37b6a2ce4jmp 0x37b6a141c
    37b6a2ce9
    37b6a2ce9 0f 1f 80 00 00 00 00 .......
    37b6a2cf0
    37b6a2cf0loc_37b6a2cf0:
    37b6a2cf0movss x1, [rip+0xd916bc]
    37b6a2cf8mov rax, [rsp+152]
    37b6a2d00sub r12l, ebp
    37b6a2d03lea rdx, [rsp+388]
    37b6a2d0badd rax, [rsp+200]
    37b6a2d13mov rcx, [rsp+128]
    37b6a2d1blea r9, [rip+0xd91652]
    37b6a2d22mov [rsp+396], r12l
    37b6a2d2adivss x1, x0
    37b6a2d2elea r8, [r14+4*rax]
    37b6a2d32movss [rsp+388], x1
    37b6a2d3bcall sscal_
    37b6a2d40mov r12l, [r15]
    37b6a2d43jmp 0x37b6a2c2d
    37b6a2d48
    37b6a2d48loc_37b6a2d48:
    37b6a2d48lea rax, [rdx+1]
    37b6a2d4cmovss x1, [r14+4*rax]
    37b6a2d52divss x1, x0
    37b6a2d56movss [r14+4*rax], x1
    37b6a2d5ccmp ecx, 2
    37b6a2d5fje 0x37b6a1060
    37b6a2d65sub ecx, 3
    37b6a2d68lea rax, [r14+4*rdx+8]
    37b6a2d6dadd rdx, rcx
    37b6a2d70lea rdx, [r14+4*rdx+12]
    37b6a2d75nop [rax+rax]
    37b6a2d80
    37b6a2d80loc_37b6a2d80:
    37b6a2d80movss x0, [rax]
    37b6a2d84add rax, 4
    37b6a2d88divss x0, [r14+4*r8]
    37b6a2d8emovss [rax-4], x0
    37b6a2d93cmp rax, rdx
    37b6a2d96jne 0x37b6a2d80
    37b6a2d98jmp 0x37b6a1060
    37b6a2d9d
    37b6a2d9d 0f 1f 00 ...
    37b6a2da0
    37b6a2da0loc_37b6a2da0:
    37b6a2da0movss x1, [rip+0xd9160c]
    37b6a2da8mov rcx, [rsp+128]
    37b6a2db0lea r8, [r14+4*rdx+4]
    37b6a2db5lea r9, [rip+0xd915b8]
    37b6a2dbclea rdx, [rsp+388]
    37b6a2dc4mov [rsp+396], ebp
    37b6a2dcbdivss x1, x0
    37b6a2dcfmovss [rsp+388], x1
    37b6a2dd8call sscal_
    37b6a2dddjmp 0x37b6a1060
    37b6a2de2
    37b6a2de2 66 0f 1f 44 00 00 f..D..
    37b6a2de8
    37b6a2de8loc_37b6a2de8:
    37b6a2de8mov rbx, [rsp+256]
    37b6a2df0mov eax, [r15]
    37b6a2df3mov edi, [rsp+380]
    37b6a2dfalea rbx, [r14+4*rbx]
    37b6a2dfemov [rsp+272], rbx
    37b6a2e06mov rbx, [rsp+208]
    37b6a2e0elea r12, [r13+4*rbx]
    37b6a2e13jmp 0x37b6a2012
    37b6a2e18
    37b6a2e18 0f 1f 84 00 00 00 00 00 ........
    37b6a2e20
    37b6a2e20loc_37b6a2e20:
    37b6a2e20mov ebx, esi
    37b6a2e22jmp 0x37b6a0fcc
    37b6a2e27
    37b6a2e27loc_37b6a2e27:
    37b6a2e27mov eax, [rsp+264]
    37b6a2e2ecmp r12l, eax
    37b6a2e31jl 0x37b6a2c2d
    37b6a2e37mov rcx, [rsp+200]
    37b6a2e3fadd rcx, [rsp+152]
    37b6a2e47mov ebx, eax
    37b6a2e49movss x1, [r14+4*rcx]
    37b6a2e4fdivss x1, x0
    37b6a2e53movss [r14+4*rcx], x1
    37b6a2e59cmp r12l, eax
    37b6a2e5cjle 0x37b6a2c2d
    37b6a2e62lea edx, [r12-1]
    37b6a2e67lea rax, [r14+4*rcx+4]
    37b6a2e6csub edx, ebx
    37b6a2e6eadd rdx, rcx
    37b6a2e71mov rcx, [rsp+256]
    37b6a2e79lea rdx, [r14+4*rdx+8]
    37b6a2e7enop
    37b6a2e80
    37b6a2e80loc_37b6a2e80:
    37b6a2e80movss x0, [rax]
    37b6a2e84add rax, 4
    37b6a2e88divss x0, [r14+4*rcx]
    37b6a2e8emovss [rax-4], x0
    37b6a2e93cmp rdx, rax
    37b6a2e96jne 0x37b6a2e80
    37b6a2e98jmp 0x37b6a2c2d
    37b6a2e9d
    37b6a2e9dloc_37b6a2e9d:
    37b6a2e9dlea rax, [rsp+372]
    37b6a2ea5xor edx, edx
    37b6a2ea7mov [rsp+176], rax
    37b6a2eaflea rax, [rsp+396]
    37b6a2eb7mov [rsp+128], rax
    37b6a2ebflea rax, [rip+0xd914b6]
    37b6a2ec6mov [rsp+240], rax
    37b6a2ecelea rax, [rip+0xd914b3]
    37b6a2ed5mov [rsp+248], rax
    37b6a2eddjmp 0x37b6a10b4
    37b6a0d60