Similarity

37b69c080
37b69c080slarzt_:
37b69c080push r15
37b69c082push r14
37b69c084push r13
37b69c086push r12
37b69c088push rbp
37b69c089push rdi
37b69c08apush rsi
37b69c08bpush rbx
37b69c08csub rsp, 152
37b69c093mov rax, [rsp+288]
37b69c09bmov r14, [rsp+272]
37b69c0a3movsxd r12, [rax]
37b69c0a6mov [rsp+240], r8
37b69c0aemov rbx, rdx
37b69c0b1mov r13, r9
37b69c0b4mov r8l, 1
37b69c0bamov r9l, 1
37b69c0c0lea rdx, [rip+0xd98109]
37b69c0c7call lsame_
37b69c0cctest eax, eax
37b69c0ceje 0x37b69c330
37b69c0d4mov r9l, 1
37b69c0damov r8l, 1
37b69c0e0lea rdx, [rip+0xd980ea]
37b69c0e7mov rcx, rbx
37b69c0eacall lsame_
37b69c0eftest eax, eax
37b69c0f1je 0x37b69c1a0
37b69c0f7mov edx, [r13]
37b69c0fbtest edx, edx
37b69c0fdjle 0x37b69c1c7
37b69c103xor eax, eax
37b69c105test r12, r12
37b69c108movsxd rbx, edx
37b69c10bmov rsi, [rsp+280]
37b69c113cmovs r12, rax
37b69c117mov rax, [rsp+256]
37b69c11fmov rdi, r12
37b69c122mov rcx, r12
37b69c125lea rbp, [rax+4*rbx]
37b69c129imul rdi, rbx
37b69c12dnot rcx
37b69c130mov [rsp+120], rcx
37b69c135lea r15, [4*rcx]
37b69c13dadd rdi, rcx
37b69c140lea rax, [rdi+rbx]
37b69c144lea rsi, [rsi+4*rax]
37b69c148jmp 0x37b69c172
37b69c14a
37b69c14a 66 0f 1f 44 00 00 f..D..
37b69c150
37b69c150loc_37b69c150:
37b69c150cmp edx, ebx
37b69c152jg 0x37b69c1e0
37b69c158
37b69c158loc_37b69c158:
37b69c158movss [rsi], x0
37b69c15c
37b69c15cloc_37b69c15c:
37b69c15csub rbx, 1
37b69c160test ebx, ebx
37b69c162jle 0x37b69c1c7
37b69c164mov edx, [r13]
37b69c168sub rdi, r12
37b69c16bsub rbp, 4
37b69c16fadd rsi, r15
37b69c172
37b69c172loc_37b69c172:
37b69c172movss x0, [r14+4*rbx-4]
37b69c179pxor x1, x1
37b69c17dmov eax, ebx
37b69c17fucomiss x0, x1
37b69c182jp 0x37b69c150
37b69c184jne 0x37b69c150
37b69c186cmp edx, ebx
37b69c188jl 0x37b69c15c
37b69c18asub edx, ebx
37b69c18cmov rcx, rsi
37b69c18flea r8, [4*rdx+4]
37b69c197xor edx, edx
37b69c199call memset
37b69c19ejmp 0x37b69c15c
37b69c1a0
37b69c1a0loc_37b69c1a0:
37b69c1a0mov eax, 2
37b69c1a5
37b69c1a5loc_37b69c1a5:
37b69c1a5lea rdx, [rsp+140]
37b69c1admov r8l, 6
37b69c1b3lea rcx, [rip+0xd98018]
37b69c1bamov [rsp+140], eax
37b69c1c1call xerbla_
37b69c1c6nop
37b69c1c7
37b69c1c7loc_37b69c1c7:
37b69c1c7add rsp, 152
37b69c1cepop rbx
37b69c1cfpop rsi
37b69c1d0pop rdi
37b69c1d1pop rbp
37b69c1d2pop r12
37b69c1d4pop r13
37b69c1d6pop r14
37b69c1d8pop r15
37b69c1daret
37b69c1db
37b69c1db 0f 1f 44 00 00 ..D..
37b69c1e0
37b69c1e0loc_37b69c1e0:
37b69c1e0lea rcx, [rip+0xd98001]
37b69c1e7add eax, 1
37b69c1easub edx, ebx
37b69c1ecmov [rsp+88], 12
37b69c1f5mov [rsp+64], rcx
37b69c1famov rcx, [rsp+264]
37b69c202lea r9, [rsp+140]
37b69c20amov [rsp+136], edx
37b69c211lea rdx, [rdi+rax]
37b69c215xorps x0, [rip+0xd97fe4]
37b69c21cmov [rsp+112], rax
37b69c221mov rax, [rsp+280]
37b69c229mov [rsp+56], rcx
37b69c22elea rcx, [rbp-4]
37b69c232mov r8, [rsp+240]
37b69c23amov [rsp+48], rcx
37b69c23flea r11, [rax+4*rdx]
37b69c243mov rcx, [rsp+264]
37b69c24blea rax, [rip+0xd97f92]
37b69c252mov [rsp+80], rax
37b69c257lea rdx, [rsp+136]
37b69c25fmov [rsp+72], r11
37b69c264mov [rsp+40], rcx
37b69c269lea rcx, [rip+0xd97f68]
37b69c270mov [rsp+32], rbp
37b69c275movss [rsp+140], x0
37b69c27emov [rsp+104], r11
37b69c283mov [rsp+96], r9
37b69c288call sgemv_
37b69c28dlea rax, [rip+0xd97f50]
37b69c294mov edx, [r13]
37b69c298mov rcx, [rsp+280]
37b69c2a0mov r11, [rsp+104]
37b69c2a5mov [rsp+56], rax
37b69c2aalea r8, [rip+0xd97f3b]
37b69c2b1mov rax, [rsp+288]
37b69c2b9sub edx, ebx
37b69c2bbmov r9, [rsp+96]
37b69c2c0mov [rsp+80], 8
37b69c2c9mov [rsp+140], edx
37b69c2d0mov [rsp+40], rax
37b69c2d5mov rax, [rsp+112]
37b69c2damov [rsp+72], 12
37b69c2e3mov rdx, rax
37b69c2e6mov [rsp+48], r11
37b69c2ebmov [rsp+64], 5
37b69c2f4imul rdx, r12
37b69c2f8add rdx, [rsp+120]
37b69c2fdadd rax, rdx
37b69c300lea rdx, [rip+0xd97ed1]
37b69c307lea rax, [rcx+4*rax]
37b69c30blea rcx, [rip+0xd97ee2]
37b69c312mov [rsp+32], rax
37b69c317call strmv_
37b69c31cmovss x0, [r14+4*rbx-4]
37b69c323jmp 0x37b69c158
37b69c328
37b69c328 0f 1f 84 00 00 00 00 00 ........
37b69c330
37b69c330loc_37b69c330:
37b69c330mov eax, 1
37b69c335jmp 0x37b69c1a5
37b69c080

37b69c33a
37b69c33a // align 0x10

37b69c340
37b69c340slasy2_:
37b69c340push r15
37b69c342push r14
37b69c344push r13
37b69c346push r12
37b69c348push rbp
37b69c349push rdi
37b69c34apush rsi
37b69c34bpush rbx
37b69c34csub rsp, 376
37b69c353movups [rsp+272], x6
37b69c35bmovups [rsp+288], x7
37b69c363movups [rsp+304], x8
37b69c36cmovups [rsp+320], x9
37b69c375movups [rsp+336], x10
37b69c37emovups [rsp+352], x11
37b69c387mov rax, [rsp+496]
37b69c38fmov r15l, [r9]
37b69c392movsxd r14, [rax]
37b69c395mov rax, [rsp+512]
37b69c39dmov r12, rcx
37b69c3a0mov r10l, [rax]
37b69c3a3mov rbp, rdx
37b69c3a6mov r13, r9
37b69c3a9mov rax, [rsp+528]
37b69c3b1mov rdx, [rsp+552]
37b69c3b9mov rcx, [rsp+568]
37b69c3c1mov eax, [rax]
37b69c3c3movsxd rdi, [rdx]
37b69c3c6mov [rcx], 0
37b69c3cctest r15l, r15l
37b69c3cfje 0x37b69c3e1
37b69c3d1mov rcx, [rsp+480]
37b69c3d9mov r11l, [rcx]
37b69c3dctest r11l, r11l
37b69c3dfjne 0x37b69c430
37b69c3e1
37b69c3e1loc_37b69c3e1:
37b69c3e1movups x6, [rsp+272]
37b69c3e9movups x7, [rsp+288]
37b69c3f1movups x8, [rsp+304]
37b69c3famovups x9, [rsp+320]
37b69c403movups x10, [rsp+336]
37b69c40cmovups x11, [rsp+352]
37b69c415add rsp, 376
37b69c41cpop rbx
37b69c41dpop rsi
37b69c41epop rdi
37b69c41fpop rbp
37b69c420pop r12
37b69c422pop r13
37b69c424pop r14
37b69c426pop r15
37b69c428ret
37b69c429
37b69c429 0f 1f 80 00 00 00 00 .......
37b69c430
37b69c430loc_37b69c430:
37b69c430xor edx, edx
37b69c432test r14, r14
37b69c435mov [rsp+56], eax
37b69c439cmovs r14, rdx
37b69c43dtest rdi, rdi
37b69c440mov [rsp+64], r10l
37b69c445cmovs rdi, rdx
37b69c449mov edx, 1
37b69c44emov [rsp+464], r8
37b69c456mov rcx, r14
37b69c459not rcx
37b69c45cmov r15, rcx
37b69c45fmov rcx, rdi
37b69c462not rcx
37b69c465mov [rsp+72], rcx
37b69c46alea rcx, [rip+0xd97d9f]
37b69c471call slamch_
37b69c476mov edx, 1
37b69c47blea rcx, [rip+0xd97d8f]
37b69c482movaps x6, x0
37b69c485movaps x9, x0
37b69c489call slamch_
37b69c48emov ecx, [r13]
37b69c492pxor x1, x1
37b69c496mov rax, [rsp+480]
37b69c49emovaps x4, x0
37b69c4a1mov r8, [rsp+464]
37b69c4a9divss x4, x6
37b69c4admov edx, [rax]
37b69c4afmov rax, [rsp+504]
37b69c4b7cvtsi2ss x1, [r8]
37b69c4bcmovss x0, [rax]
37b69c4c0mov rax, [rsp+488]
37b69c4c8lea ecx, [rdx+2*rcx-2]
37b69c4ccmovss x2, [rax]
37b69c4d0movaps x8, x4
37b69c4d4cmp ecx, 3
37b69c4d7je 0x37b69ca20
37b69c4ddmovsxd rdx, [rsp+64]
37b69c4e2movsxd rax, [rsp+56]
37b69c4e7xor r8l, r8l
37b69c4eatest rdx, rdx
37b69c4edcmovs rdx, r8
37b69c4f1xor r10l, r10l
37b69c4f4test rax, rax
37b69c4f7cmovs rax, r10
37b69c4fbmov r8, rdx
37b69c4femov r10, rax
37b69c501not r8
37b69c504not r10
37b69c507mov [rsp+56], r10
37b69c50ccmp ecx, 4
37b69c50fje 0x37b69c5c0
37b69c515cmp ecx, 2
37b69c518je 0x37b69ccd0
37b69c51emulss x1, x0
37b69c522movss x3, [rip+0xd97d66]
37b69c52aaddss x2, x1
37b69c52emovaps x5, x2
37b69c531andps x5, x3
37b69c534comiss x4, x5
37b69c537jb 0x37b69c54d
37b69c539mov rax, [rsp+568]
37b69c541movaps x2, x4
37b69c544movaps x5, x4
37b69c547mov [rax], 1
37b69c54d
37b69c54dloc_37b69c54d:
37b69c54dmov rax, [rsp+536]
37b69c555movaps x4, x8
37b69c559mov [rax], 0x3f800000
37b69c55fmov rax, [rsp+520]
37b69c567movss x1, [rax]
37b69c56bmovaps x0, x1
37b69c56eandps x0, x3
37b69c571mulss x4, x0
37b69c575comiss x4, x5
37b69c578jbe 0x37b69c596
37b69c57amovss x4, [rip+0xd97cfe]
37b69c582mov rax, [rsp+536]
37b69c58adivss x4, x0
37b69c58emulss x1, x4
37b69c592movss [rax], x4
37b69c596
37b69c596loc_37b69c596:
37b69c596divss x1, x2
37b69c59amov rax, [rsp+544]
37b69c5a2movss [rax], x1
37b69c5a6mov rax, [rsp+560]
37b69c5aeandps x1, x3
37b69c5b1movss [rax], x1
37b69c5b5jmp 0x37b69c3e1
37b69c5ba
37b69c5ba 66 0f 1f 44 00 00 f..D..
37b69c5c0
37b69c5c0loc_37b69c5c0:
37b69c5c0mov [rsp+128], rax
37b69c5c8lea rdx, [r8+2*rdx]
37b69c5ccmov rax, [rsp+504]
37b69c5d4movaps x4, x9
37b69c5d8lea r13, [rsp+160]
37b69c5e0lea rcx, [rip+0xd97c39]
37b69c5e7movsldup x7, x1
37b69c5ebmovss [rsp+112], x1
37b69c5f1lea r11, [rax+4*rdx]
37b69c5f5lea rdx, [r15+2*r14]
37b69c5f9mov rax, [rsp+488]
37b69c601mov [rsp+160], 0
37b69c60cmovss x3, [r11+4]
37b69c612movss x5, [r11+8]
37b69c618lea r14, [rsp+208]
37b69c620lea r8, [rip+0xd97bf5]
37b69c627lea r10, [rax+4*rdx]
37b69c62bmov r9, r14
37b69c62emov rdx, r13
37b69c631mov [rsp+96], r11
37b69c636movss x6, [r10+4]
37b69c63cunpcklps x5, x3
37b69c63fmovss x3, [r10+8]
37b69c645mov [rsp+80], r10
37b69c64aunpcklps x3, x6
37b69c64dmovlhps x3, x5
37b69c650movss x5, [rax+4]
37b69c655mov rax, [rsp+504]
37b69c65dandps x3, [rip+0xd97c4c]
37b69c664unpcklps x2, x5
37b69c667movss x5, [rax+4]
37b69c66clea rax, [rip+0xd97ba1]
37b69c673mov [rsp+32], rax
37b69c678unpcklps x5, x0
37b69c67bmovlhps x5, x2
37b69c67eandps x5, [rip+0xd97c2b]
37b69c685maxps x3, x5
37b69c688movaps x2, x3
37b69c68bmovhlps x3, x3
37b69c68emaxps x3, x2
37b69c691movaps x2, x3
37b69c694shufps x2, x3, 85
37b69c698maxps x2, x3
37b69c69bmulss x4, x2
37b69c69fmaxss x4, x8
37b69c6a4movss [rsp+64], x4
37b69c6aacall scopy_
37b69c6afmovss x1, [rsp+112]
37b69c6b5mov r11, [rsp+96]
37b69c6bamov rax, [rsp+488]
37b69c6c2mov r10, [rsp+80]
37b69c6c7mov ecx, [r12]
37b69c6cbmovss x4, [rsp+64]
37b69c6d1movss x3, [rax]
37b69c6d5mov rax, [rsp+504]
37b69c6ddtest ecx, ecx
37b69c6dfmovss x5, [rax]
37b69c6e3movaps x2, x3
37b69c6e6mov rax, [rsp+488]
37b69c6eemulss x5, x1
37b69c6f2mulss x1, [r11+8]
37b69c6f8addss x2, x5
37b69c6fcaddss x3, x1
37b69c700movss [rsp+208], x2
37b69c709movss x2, [r10+8]
37b69c70fmovss [rsp+248], x3
37b69c718addss x5, x2
37b69c71caddss x2, x1
37b69c720movss x1, [r10+4]
37b69c726movss [rsp+268], x2
37b69c72fmovss x2, [rax+4]
37b69c734mov rax, [rsp+128]
37b69c73cmovss [rsp+228], x5
37b69c745je 0x37b69ce15
37b69c74b
37b69c74bloc_37b69c74b:
37b69c74bmov rcx, [rsp+504]
37b69c753mov edx, [rbp]
37b69c756movss [rsp+224], x2
37b69c75fmovss [rsp+212], x1
37b69c768movss [rsp+264], x2
37b69c771movss x2, [rcx+4]
37b69c776movss [rsp+252], x1
37b69c77fmovss x1, [r11+4]
37b69c785test edx, edx
37b69c787jne 0x37b69ce02
37b69c78dunpcklps x1, x2
37b69c790movq x2, x7
37b69c794movq x1, x1
37b69c798mulps x1, x2
37b69c79b
37b69c79bloc_37b69c79b:
37b69c79bmov rcx, [rsp+56]
37b69c7a0movshdup x7, x1
37b69c7a4mov r9, r14
37b69c7a7movss [rsp+216], x1
37b69c7b0movlps [rsp+236], x1
37b69c7b8mov ebp, 3
37b69c7bdlea rax, [rcx+2*rax]
37b69c7c1mov [rsp+64], r14
37b69c7c6mov rcx, [rsp+520]
37b69c7cemov r12l, ebp
37b69c7d1mov [rsp+56], 4
37b69c7damovss x1, [rcx+4*rax+8]
37b69c7e0movss x2, [rcx+4*rax+4]
37b69c7e6mov [rsp+136], rdi
37b69c7eemov edi, 1
37b69c7f3movss x3, [rcx+4]
37b69c7f8mov [rsp+144], r14
37b69c800unpcklps x2, x1
37b69c803movss x1, [rcx]
37b69c807mov [rsp+152], r13
37b69c80flea r13, [rsp+232]
37b69c817mov [rsp+80], r14
37b69c81clea r14, [rsp+164]
37b69c824movss [rsp+260], x7
37b69c82dunpcklps x1, x3
37b69c830movss x3, [rip+0xd97a58]
37b69c838movlhps x1, x2
37b69c83bmovups [rsp+160], x1
37b69c843
37b69c843loc_37b69c843:
37b69c843lea rbp, [r9+4]
37b69c847mov r15, rdi
37b69c84apxor x2, x2
37b69c84emov rcx, rbp
37b69c851
37b69c851loc_37b69c851:
37b69c851mov edx, r15l
37b69c854mov rax, rdi
37b69c857
37b69c857loc_37b69c857:
37b69c857mov r8, rax
37b69c85amovaps x1, x2
37b69c85dshl r8, 4
37b69c861movss x0, [rcx+r8-20]
37b69c868andps x0, x3
37b69c86bcmpss x1, x0, 2
37b69c870comiss x0, x2
37b69c873andps x0, x1
37b69c876andnps x1, x2
37b69c879cmovae ebx, eax
37b69c87ccmovae esi, edx
37b69c87fmovaps x2, x1
37b69c882add rax, 1
37b69c886orps x2, x0
37b69c889cmp eax, 5
37b69c88cjne 0x37b69c857
37b69c88eadd r15, 1
37b69c892add rcx, 4
37b69c896cmp r15l, 5
37b69c89ajne 0x37b69c851
37b69c89cmov r15l, edi
37b69c89fcmp edi, esi
37b69c8a1je 0x37b69c90e
37b69c8a3mov rcx, [rsp+80]
37b69c8a8movsxd rax, esi
37b69c8abmovups [rsp+112], x3
37b69c8b0movss [rsp+128], x4
37b69c8b9sub rax, 1
37b69c8bdlea rdx, [rcx+4*rax]
37b69c8c1mov [rsp+96], rax
37b69c8c6lea rax, [rip+0xd9794b]
37b69c8cdmov [rsp+32], rax
37b69c8d2mov r8, rax
37b69c8d5mov rcx, rax
37b69c8d8call sswap_
37b69c8ddmov rax, [rsp+96]
37b69c8e2movss x0, [r14-4]
37b69c8e8movss x4, [rsp+128]
37b69c8f1movups x3, [rsp+112]
37b69c8f6movss x2, [rsp+4*rax+160]
37b69c8ffmovss [r14-4], x2
37b69c905movss [rsp+4*rax+160], x0
37b69c90e
37b69c90eloc_37b69c90e:
37b69c90ecmp r15l, ebx
37b69c911je 0x37b69c959
37b69c913lea rax, [rip+0xd978fa]
37b69c91amovsxd rdx, ebx
37b69c91dmov r9, [rsp+64]
37b69c922lea rcx, [rip+0xd978ef]
37b69c929mov [rsp+32], rax
37b69c92esub rdx, 1
37b69c932mov r8, rax
37b69c935shl rdx, 4
37b69c939add rdx, [rsp+80]
37b69c93emovss [rsp+112], x4
37b69c944movups [rsp+96], x3
37b69c949call sswap_
37b69c94emovss x4, [rsp+112]
37b69c954movups x3, [rsp+96]
37b69c959
37b69c959loc_37b69c959:
37b69c959movss x1, [r13-24]
37b69c95fmov rax, [rsp+56]
37b69c964movaps x2, x1
37b69c967mov [rsp+rax+172], ebx
37b69c96eandps x2, x3
37b69c971comiss x4, x2
37b69c974jbe 0x37b69c98d
37b69c976mov rax, [rsp+568]
37b69c97emovss [r13-24], x4
37b69c984movaps x1, x4
37b69c987mov [rax], 1
37b69c98d
37b69c98dloc_37b69c98d:
37b69c98dmov rdx, r13
37b69c990mov r9, r14
37b69c993mov rcx, -1
37b69c99alea r15, [rdi-5]
37b69c99e
37b69c99eloc_37b69c99e:
37b69c99emovss x0, [rdx-20]
37b69c9a3movss x2, [r14-4]
37b69c9a9divss x0, x1
37b69c9admovss x1, [r9]
37b69c9b2mulss x2, x0
37b69c9b6movss [rdx-20], x0
37b69c9bbmulss x0, [r13-8]
37b69c9c1subss x1, x2
37b69c9c5movss [r9], x1
37b69c9camovss x1, [rdx-4]
37b69c9cfsubss x1, x0
37b69c9d3movss [rdx-4], x1
37b69c9d8cmp r12l, 5
37b69c9dcje 0x37b69ce28
37b69c9e2mov r8l, r12l
37b69c9e5mov rax, rdx
37b69c9e8
37b69c9e8loc_37b69c9e8:
37b69c9e8movss x1, [rax+4*rcx+12]
37b69c9eemulss x1, [rdx-20]
37b69c9f3add rax, 16
37b69c9f7movss x0, [rax-4]
37b69c9fcsubss x0, x1
37b69ca00movss [rax-4], x0
37b69ca05cmp r8l, 4
37b69ca09je 0x37b69ce28
37b69ca0fmov r8l, 4
37b69ca15jmp 0x37b69c9e8
37b69ca17
37b69ca17 66 0f 1f 84 00 00 00 00 00 f........
37b69ca20
37b69ca20loc_37b69ca20:
37b69ca20mov rax, [rsp+488]
37b69ca28lea rcx, [r15+2*r14]
37b69ca2cmovaps x5, x2
37b69ca2fmovss x3, [rip+0xd97859]
37b69ca37mov rbx, [rsp+488]
37b69ca3fmulss x1, x0
37b69ca43andps x5, x3
37b69ca46mov r8l, [r12]
37b69ca4amov edx, [rax+4*rcx+4]
37b69ca4emov eax, [rax+4]
37b69ca51movss x6, [rbx+4*rcx+8]
37b69ca57movd x4, eax
37b69ca5bmovd x7, edx
37b69ca5fandps x4, x3
37b69ca62andps x7, x3
37b69ca65maxss x5, x4
37b69ca69movaps x4, x0
37b69ca6caddss x2, x1
37b69ca70andps x4, x3
37b69ca73addss x1, x6
37b69ca77movss [rsp+192], x2
37b69ca80maxss x5, x4
37b69ca84movaps x4, x6
37b69ca87movss [rsp+204], x1
37b69ca90andps x4, x3
37b69ca93maxss x4, x7
37b69ca97maxss x5, x4
37b69ca9bmovaps x4, x9
37b69ca9fmulss x4, x5
37b69caa3maxss x4, x8
37b69caa8test r8l, r8l
37b69caabjne 0x37b69cdf3
37b69cab1
37b69cab1loc_37b69cab1:
37b69cab1movd x0, eax
37b69cab5mov rax, [rsp+520]
37b69cabdmovd x5, edx
37b69cac1unpcklps x0, x5
37b69cac4movss x6, [rax]
37b69cac8movss x2, [rax+4]
37b69cacdmovlps [rsp+196], x0
37b69cad5movaps x11, x6
37b69cad9
37b69cad9loc_37b69cad9:
37b69cad9lea rdx, [rsp+192]
37b69cae1lea r8, [rip+0xd9772c]
37b69cae8movups [rsp+80], x3
37b69caedlea rcx, [rip+0xd97724]
37b69caf4movss [rsp+64], x2
37b69cafamovss [rsp+56], x4
37b69cb00call isamax_
37b69cb05movups x3, [rsp+80]
37b69cb0amovss x4, [rsp+56]
37b69cb10cdqe
37b69cb12movss x2, [rsp+64]
37b69cb18sub rax, 1
37b69cb1cmovss x5, [rsp+4*rax+192]
37b69cb25movaps x1, x5
37b69cb28andps x1, x3
37b69cb2bcomiss x4, x1
37b69cb2ejb 0x37b69cb41
37b69cb30mov rsi, [rsp+568]
37b69cb38movaps x5, x4
37b69cb3bmov [rsi], 1
37b69cb41
37b69cb41loc_37b69cb41:
37b69cb41lea rdx, [rip+0xd97728]
37b69cb48lea rcx, [rip+0xd976e1]
37b69cb4fmovsxd rdx, [rdx+4*rax]
37b69cb53movss x7, [rsp+4*rdx+188]
37b69cb5clea rdx, [rip+0xd976fd]
37b69cb63movsxd rdx, [rdx+4*rax]
37b69cb67movaps x10, x7
37b69cb6bmovss x1, [rsp+4*rdx+188]
37b69cb74lea rdx, [rip+0xd976d5]
37b69cb7bmovsxd rdx, [rdx+4*rax]
37b69cb7fdivss x1, x5
37b69cb83movss x0, [rsp+4*rdx+188]
37b69cb8clea rdx, [rip+0xd976ad]
37b69cb93mov edx, [rdx+4*rax]
37b69cb96mov eax, [rcx+4*rax]
37b69cb99mulss x7, x1
37b69cb9dmovaps x6, x7
37b69cba0movaps x7, x0
37b69cba3subss x7, x6
37b69cba7movaps x9, x7
37b69cbabandps x7, x3
37b69cbaecomiss x4, x7
37b69cbb1movaps x0, x7
37b69cbb4jb 0x37b69cbce
37b69cbb6mov rbx, [rsp+568]
37b69cbbemovaps x0, x4
37b69cbc1movaps x9, x4
37b69cbc5andps x0, x3
37b69cbc8mov [rbx], 1
37b69cbce
37b69cbceloc_37b69cbce:
37b69cbcemovaps x4, x1
37b69cbd1test eax, eax
37b69cbd3je 0x37b69ccb8
37b69cbd9mulss x4, x2
37b69cbddmovaps x1, x11
37b69cbe1subss x1, x4
37b69cbe5
37b69cbe5loc_37b69cbe5:
37b69cbe5movaps x6, x8
37b69cbe9movaps x4, x1
37b69cbecaddss x6, x8
37b69cbf1andps x4, x3
37b69cbf4movaps x8, x2
37b69cbf8andps x8, x3
37b69cbfcmovaps x11, x8
37b69cc00movaps x7, x6
37b69cc03mulss x7, x4
37b69cc07comiss x7, x0
37b69cc0aja 0x37b69cc20
37b69cc0cmulss x6, x8
37b69cc11movaps x7, x5
37b69cc14andps x7, x3
37b69cc17comiss x6, x7
37b69cc1ajbe 0x37b69cde6
37b69cc20
37b69cc20loc_37b69cc20:
37b69cc20movaps x0, x11
37b69cc24maxss x0, x4
37b69cc28movss x4, [rip+0xd97670]
37b69cc30divss x4, x0
37b69cc34mulss x2, x4
37b69cc38mulss x1, x4
37b69cc3c
37b69cc3cloc_37b69cc3c:
37b69cc3cdivss x2, x5
37b69cc40mov rax, [rsp+536]
37b69cc48movss [rax], x4
37b69cc4cdivss x1, x9
37b69cc51movaps x0, x2
37b69cc54movaps x2, x10
37b69cc58divss x2, x5
37b69cc5cmulss x2, x1
37b69cc60subss x0, x2
37b69cc64test edx, edx
37b69cc66jne 0x37b69cc71
37b69cc68movaps x2, x1
37b69cc6bmovaps x1, x0
37b69cc6emovaps x0, x2
37b69cc71
37b69cc71loc_37b69cc71:
37b69cc71mov rax, [rsp+544]
37b69cc79movaps x2, x0
37b69cc7candps x2, x3
37b69cc7fmovss [rax], x1
37b69cc83cmp [r13], 1
37b69cc88je 0x37b69cda8
37b69cc8emov rax, [rsp+544]
37b69cc96andps x1, x3
37b69cc99maxss x1, x2
37b69cc9dmovss [rax+4], x0
37b69cca2mov rax, [rsp+560]
37b69ccaamovss [rax], x1
37b69ccaejmp 0x37b69c3e1
37b69ccb3
37b69ccb3 0f 1f 44 00 00 ..D..
37b69ccb8
37b69ccb8loc_37b69ccb8:
37b69ccb8mulss x4, x11
37b69ccbdsubss x2, x4
37b69ccc1movaps x1, x2
37b69ccc4movaps x2, x11
37b69ccc8jmp 0x37b69cbe5
37b69cccd
37b69cccd 0f 1f 00 ...
37b69ccd0
37b69ccd0loc_37b69ccd0:
37b69ccd0mov rsi, [rsp+504]
37b69ccd8movaps x5, x0
37b69ccdblea rdx, [r8+2*rdx]
37b69ccdfmov r9l, [rbp]
37b69cce3movss x3, [rip+0xd975a5]
37b69ccebandps x5, x3
37b69cceemovss x7, [rsi+4]
37b69ccf3mov ecx, [rsi+4*rdx+4]
37b69ccf7mov edx, [rsi+4*rdx+8]
37b69ccfbmovaps x4, x7
37b69ccfemovd x6, ecx
37b69cd02andps x4, x3
37b69cd05andps x6, x3
37b69cd08maxss x5, x4
37b69cd0cmovaps x4, x2
37b69cd0fandps x4, x3
37b69cd12maxss x5, x4
37b69cd16movd x4, edx
37b69cd1aandps x4, x3
37b69cd1dmaxss x4, x6
37b69cd21maxss x5, x4
37b69cd25movaps x4, x9
37b69cd29mulss x4, x5
37b69cd2dmovaps x5, x0
37b69cd30mulss x5, x1
37b69cd34maxss x4, x8
37b69cd39addss x5, x2
37b69cd3dmovss [rsp+192], x5
37b69cd46movd x5, edx
37b69cd4amulss x5, x1
37b69cd4eaddss x5, x2
37b69cd52movd x2, ecx
37b69cd56mulss x2, x1
37b69cd5amulss x1, x7
37b69cd5emovss [rsp+204], x5
37b69cd67test r9l, r9l
37b69cd6aje 0x37b69cddb
37b69cd6c
37b69cd6cloc_37b69cd6c:
37b69cd6cmov rsi, [rsp+520]
37b69cd74mov rbx, [rsp+56]
37b69cd79movss [rsp+200], x2
37b69cd82movss [rsp+196], x1
37b69cd8bmovss x6, [rsi]
37b69cd8flea rax, [rbx+2*rax+1]
37b69cd94movss x2, [rsi+4*rax]
37b69cd99movaps x11, x6
37b69cd9djmp 0x37b69cad9
37b69cda2
37b69cda2 66 0f 1f 44 00 00 f..D..
37b69cda8
37b69cda8loc_37b69cda8:
37b69cda8mov rax, [rsp+72]
37b69cdadmov rsi, [rsp+544]
37b69cdb5lea rax, [rax+2*rdi+1]
37b69cdbamovss [rsi+4*rax], x0
37b69cdbfmovss x0, [rsi]
37b69cdc3mov rax, [rsp+560]
37b69cdcbandps x0, x3
37b69cdceaddss x0, x2
37b69cdd2movss [rax], x0
37b69cdd6jmp 0x37b69c3e1
37b69cddb
37b69cddbloc_37b69cddb:
37b69cddbmovaps x5, x1
37b69cddemovaps x1, x2
37b69cde1movaps x2, x5
37b69cde4jmp 0x37b69cd6c
37b69cde6
37b69cde6loc_37b69cde6:
37b69cde6movss x4, [rip+0xd97492]
37b69cdeejmp 0x37b69cc3c
37b69cdf3
37b69cdf3loc_37b69cdf3:
37b69cdf3movd x1, eax
37b69cdf7mov eax, edx
37b69cdf9movd edx, x1
37b69cdfdjmp 0x37b69cab1
37b69ce02
37b69ce02loc_37b69ce02:
37b69ce02unpcklps x2, x1
37b69ce05movq x3, x7
37b69ce09movq x1, x2
37b69ce0dmulps x1, x3
37b69ce10jmp 0x37b69c79b
37b69ce15
37b69ce15loc_37b69ce15:
37b69ce15movaps x3, x1
37b69ce18movaps x1, x2
37b69ce1bmovaps x2, x3
37b69ce1ejmp 0x37b69c74b
37b69ce23
37b69ce23 0f 1f 44 00 00 ..D..
37b69ce28
37b69ce28loc_37b69ce28:
37b69ce28sub rcx, 1
37b69ce2ccmp r15, rcx
37b69ce2fje 0x37b69ce48
37b69ce31movss x1, [r13-24]
37b69ce37add r9, 4
37b69ce3badd rdx, 4
37b69ce3fjmp 0x37b69c99e
37b69ce44
37b69ce44 0f 1f 40 00 ..@.
37b69ce48
37b69ce48loc_37b69ce48:
37b69ce48add rdi, 1
37b69ce4cadd r12l, 1
37b69ce50add r14, 4
37b69ce54add r13, 20
37b69ce58add [rsp+64], 16
37b69ce5ecmp rdi, 4
37b69ce62je 0x37b69ce78
37b69ce64add [rsp+56], 4
37b69ce6amov r9, rbp
37b69ce6djmp 0x37b69c843
37b69ce72
37b69ce72 66 0f 1f 44 00 00 f..D..
37b69ce78
37b69ce78loc_37b69ce78:
37b69ce78movss x5, [rsp+268]
37b69ce81mov rdi, [rsp+136]
37b69ce89mov rax, [rsp+144]
37b69ce91mov r13, [rsp+152]
37b69ce99movaps x1, x5
37b69ce9cmov r14, [rsp+80]
37b69cea1andps x1, x3
37b69cea4comiss x4, x1
37b69cea7jbe 0x37b69cec3
37b69cea9mov rbx, [rsp+568]
37b69ceb1movaps x5, x4
37b69ceb4movss [rsp+268], x4
37b69cebdmov [rbx], 1
37b69cec3
37b69cec3loc_37b69cec3:
37b69cec3movss x7, [rip+0xd973f5]
37b69cecbmovss x4, [rip+0xd973ad]
37b69ced3movss x1, [rsp+160]
37b69cedcmov rsi, [rsp+536]
37b69cee4mulss x7, x8
37b69cee9movups x6, [rsp+160]
37b69cef1movss x0, [rsp+208]
37b69cefaandps x1, x3
37b69cefdmovups x2, [rip+0xd973ac]
37b69cf04movss [rsi], x4
37b69cf08andps x0, x3
37b69cf0bandps x2, x6
37b69cf0emulss x1, x7
37b69cf12comiss x1, x0
37b69cf15ja 0x37b69cf70
37b69cf17movss x0, [rsp+228]
37b69cf20movaps x1, x2
37b69cf23shufps x1, x2, 85
37b69cf27mulss x1, x7
37b69cf2bandps x0, x3
37b69cf2ecomiss x1, x0
37b69cf31ja 0x37b69cf70
37b69cf33movss x1, [rsp+248]
37b69cf3cmovaps x0, x2
37b69cf3funpckhps x0, x2
37b69cf42mulss x0, x7
37b69cf46andps x1, x3
37b69cf49comiss x0, x1
37b69cf4cja 0x37b69cf70
37b69cf4emovaps x0, x2
37b69cf51andps x5, x3
37b69cf54shufps x0, x2, -1
37b69cf58mulss x7, x0
37b69cf5ccomiss x7, x5
37b69cf5fjbe 0x37b69cfaa
37b69cf61nop [rax]
37b69cf65nop [rax+rax]
37b69cf70
37b69cf70loc_37b69cf70:
37b69cf70movaps x0, x2
37b69cf73mov rbx, [rsp+536]
37b69cf7bmovhlps x0, x2
37b69cf7emaxps x0, x2
37b69cf81movaps x1, x0
37b69cf84shufps x1, x0, 85
37b69cf88maxps x1, x0
37b69cf8bmovss x0, [rip+0xd97331]
37b69cf93divss x0, x1
37b69cf97movss [rbx], x0
37b69cf9bshufps x0, x0, 0
37b69cf9fmulps x0, x6
37b69cfa2movups [rsp+160], x0
37b69cfaa
37b69cfaaloc_37b69cfaa:
37b69cfaalea r9, [rsp+192]
37b69cfb2lea rcx, [r14+16]
37b69cfb6mov r10l, 5
37b69cfbcmov r8, r9
37b69cfbf
37b69cfbfloc_37b69cfbf:
37b69cfbfmov rdx, r10
37b69cfc2movaps x2, x4
37b69cfc5divss x2, [rax+60]
37b69cfcasub r10, 1
37b69cfcemovss x1, [r13+4*rdx-8]
37b69cfd5mulss x1, x2
37b69cfd9movss [r8+12], x1
37b69cfdfcmp edx, 5
37b69cfe2je 0x37b69d012
37b69cfe4
37b69cfe4loc_37b69cfe4:
37b69cfe4mov r11, rdx
37b69cfe7shl r11, 4
37b69cfebmovss x0, [rcx+r11-20]
37b69cff2mulss x0, x2
37b69cff6mulss x0, [rsp+4*rdx+188]
37b69cfffadd rdx, 1
37b69d003subss x1, x0
37b69d007movss [r8+12], x1
37b69d00dcmp edx, 5
37b69d010jne 0x37b69cfe4
37b69d012
37b69d012loc_37b69d012:
37b69d012sub rax, 20
37b69d016sub r8, 4
37b69d01asub rcx, 4
37b69d01ecmp r10, 1
37b69d022jne 0x37b69cfbf
37b69d024lea rcx, [rsp+176]
37b69d02cmov edx, 3
37b69d031
37b69d031loc_37b69d031:
37b69d031movsxd rax, [rcx+8]
37b69d035cmp eax, edx
37b69d037je 0x37b69d05b
37b69d039sub rax, 1
37b69d03dmovss x0, [r9+8]
37b69d043movss x1, [rsp+4*rax+192]
37b69d04cmovss [r9+8], x1
37b69d052movss [rsp+4*rax+192], x0
37b69d05b
37b69d05bloc_37b69d05b:
37b69d05bsub rcx, 4
37b69d05fsub r9, 4
37b69d063sub edx, 1
37b69d066jne 0x37b69d031
37b69d068mov rax, [rsp+192]
37b69d070mov rbx, [rsp+544]
37b69d078movss x0, [rsp+192]
37b69d081mov rdx, [rsp+200]
37b69d089movss x1, [rsp+200]
37b69d092mov [rbx], rax
37b69d095andps x0, x3
37b69d098mov rax, [rsp+72]
37b69d09dmovss x2, [rsp+204]
37b69d0a6andps x1, x3
37b69d0a9addss x0, x1
37b69d0adandps x2, x3
37b69d0b0lea rax, [rax+2*rdi+1]
37b69d0b5movss x1, [rsp+196]
37b69d0bemov [rbx+4*rax], rdx
37b69d0c2mov rax, [rsp+560]
37b69d0caandps x1, x3
37b69d0cdaddss x1, x2
37b69d0d1maxss x0, x1
37b69d0d5movss [rax], x0
37b69d0d9jmp 0x37b69c3e1
37b69c340

37b69d0de
37b69d0de // align 0x40

37b69d100
37b69d100slasyf_:
37b69d100push r15
37b69d102push r14
37b69d104push r13
37b69d106push r12
37b69d108push rbp
37b69d109push rdi
37b69d10apush rsi
37b69d10bpush rbx
37b69d10csub rsp, 472
37b69d113movups [rsp+384], x6
37b69d11bmovups [rsp+400], x7
37b69d123movups [rsp+416], x8
37b69d12cmovups [rsp+432], x9
37b69d135movups [rsp+448], x10
37b69d13emov rax, [rsp+584]
37b69d146mov r14, [rsp+576]
37b69d14emov r15, [rsp+600]
37b69d156movsxd rbx, [rax]
37b69d159xor eax, eax
37b69d15btest rbx, rbx
37b69d15emov [rsp+560], r8
37b69d166mov r13, rdx
37b69d169mov r12, r9
37b69d16ccmovns rax, rbx
37b69d170mov r9l, 1
37b69d176mov r8l, 1
37b69d17clea rdx, [rip+0xd9714d]
37b69d183mov [rsp+152], rax
37b69d18bnot rax
37b69d18emov [rsp+144], rax
37b69d196mov rax, [rsp+608]
37b69d19emovsxd rbx, [rax]
37b69d1a1xor eax, eax
37b69d1a3test rbx, rbx
37b69d1a6cmovns rax, rbx
37b69d1aamov [rsp+160], rax
37b69d1b2not rax
37b69d1b5mov [rsp+168], rax
37b69d1bdmov rax, [rsp+616]
37b69d1c5mov [rax], 0
37b69d1cbcall lsame_
37b69d1d0test eax, eax
37b69d1d2je 0x37b69d540
37b69d1d8mov rax, [rsp+560]
37b69d1e0mov edi, [r13]
37b69d1e4mov esi, [rax]
37b69d1e6mov eax, edi
37b69d1e8mov [rsp+376], edi
37b69d1efsub eax, esi
37b69d1f1add eax, 1
37b69d1f4cmp edi, eax
37b69d1f6jg 0x37b69d200
37b69d1f8cmp edi, esi
37b69d1fajg 0x37b69e9dd
37b69d200
37b69d200loc_37b69d200:
37b69d200lea rax, [rsp+376]
37b69d208mov edx, edi
37b69d20amov [rsp+568], r12
37b69d212movss x7, [rip+0xd970f6]
37b69d21amov [rsp+184], rax
37b69d222lea rax, [rsp+380]
37b69d22amovss x8, [rip+0xd970e1]
37b69d233mov [rsp+136], rax
37b69d23blea rax, [rip+0xd9709a]
37b69d242mov [rsp+224], rax
37b69d24alea rax, [rip+0xd97097]
37b69d251mov [rsp+232], rax
37b69d259nop [rax]
37b69d260
37b69d260loc_37b69d260:
37b69d260test edi, edi
37b69d262jle 0x37b69d3a4
37b69d268movsxd rsi, esi
37b69d26bmov rax, [rsp+168]
37b69d273imul rsi, [rsp+160]
37b69d27cmovsxd rdi, edi
37b69d27fimul rdi, [rsp+152]
37b69d288mov rcx, [rsp+184]
37b69d290lea rbp, [rsi+rax]
37b69d294mov rax, [rsp+144]
37b69d29clea r9, [r15+4*rbp+4]
37b69d2a1lea rax, [rax+rdi+1]
37b69d2a6mov [rsp+176], r9
37b69d2aelea rdx, [r14+4*rax]
37b69d2b2lea rax, [rip+0xd9701b]
37b69d2b9mov [rsp+32], rax
37b69d2bemov r8, rax
37b69d2c1call scopy_
37b69d2c6mov eax, [r13]
37b69d2camov ebx, [rsp+376]
37b69d2d1cmp eax, ebx
37b69d2d3jg 0x37b69de20
37b69d2d9
37b69d2d9loc_37b69d2d9:
37b69d2d9movsxd r12, ebx
37b69d2dcpxor x9, x9
37b69d2e1lea rax, [rbp+r12]
37b69d2e6movss x6, [r15+4*rax]
37b69d2ecandps x6, [rip+0xd9700d]
37b69d2f3cmp ebx, 1
37b69d2f6jle 0x37b69d338
37b69d2f8mov rdx, [rsp+176]
37b69d300mov rcx, [rsp+136]
37b69d308lea eax, [rbx-1]
37b69d30blea r8, [rip+0xd96fc2]
37b69d312mov [rsp+380], eax
37b69d319call isamax_
37b69d31emov [rsp+364], eax
37b69d325cdqe
37b69d327add rax, rbp
37b69d32amovss x9, [r15+4*rax]
37b69d330andps x9, [rip+0xd96fc8]
37b69d338
37b69d338loc_37b69d338:
37b69d338movaps x1, x9
37b69d33cpxor x0, x0
37b69d340maxss x1, x6
37b69d344ucomiss x1, x0
37b69d347jp 0x37b69dd68
37b69d34djne 0x37b69dd68
37b69d353mov rax, [rsp+616]
37b69d35blea edi, [rbx-1]
37b69d35emov edx, [rax]
37b69d360test edx, edx
37b69d362jne 0x37b69d366
37b69d364mov [rax], ebx
37b69d366
37b69d366loc_37b69d366:
37b69d366mov rax, [rsp+592]
37b69d36emov [rax+4*r12-4], ebx
37b69d373
37b69d373loc_37b69d373:
37b69d373mov rax, [rsp+560]
37b69d37bmov edx, [r13]
37b69d37fmov [rsp+376], edi
37b69d386mov ecx, [rax]
37b69d388mov eax, edx
37b69d38asub eax, ecx
37b69d38clea esi, [rcx+rdi]
37b69d38fadd eax, 1
37b69d392sub esi, edx
37b69d394cmp eax, edi
37b69d396jl 0x37b69d260
37b69d39ccmp ecx, edx
37b69d39ejge 0x37b69d260
37b69d3a4
37b69d3a4loc_37b69d3a4:
37b69d3a4mov r12, [rsp+568]
37b69d3acsub edx, edi
37b69d3ae
37b69d3aeloc_37b69d3ae:
37b69d3aemov rax, [rsp+584]
37b69d3b6add esi, 1
37b69d3b9add edi, 1
37b69d3bcmov [rsp+88], r14
37b69d3c1movsxd rsi, esi
37b69d3c4movsxd rdi, edi
37b69d3c7mov [rsp+380], edx
37b69d3cemov r9, [rsp+184]
37b69d3d6mov [rsp+96], rax
37b69d3dbmov rax, [rsp+232]
37b69d3e3lea rcx, [rip+0xd96f0b]
37b69d3ealea r8, [rip+0xd96efb]
37b69d3f1imul rsi, [rsp+160]
37b69d3famov rdx, [rsp+224]
37b69d402mov [rsp+120], 9
37b69d40bimul rdi, [rsp+152]
37b69d414mov [rsp+80], rax
37b69d419mov rax, [rsp+608]
37b69d421mov [rsp+112], 12
37b69d42amov [rsp+104], 5
37b69d433mov [rsp+72], rax
37b69d438mov rax, [rsp+168]
37b69d440lea rax, [rax+rsi+1]
37b69d445lea rax, [r15+4*rax]
37b69d449mov r15, [rsp+144]
37b69d451mov [rsp+64], rax
37b69d456mov rax, [rsp+584]
37b69d45emov [rsp+56], rax
37b69d463lea rax, [r15+rdi+1]
37b69d468lea rax, [r14+4*rax]
37b69d46cmov [rsp+48], rax
37b69d471lea rax, [rip+0xd96e60]
37b69d478mov [rsp+40], rax
37b69d47dmov rax, [rsp+136]
37b69d485mov [rsp+32], rax
37b69d48acall sgemmtr_
37b69d48fmov ebx, [rsp+376]
37b69d496mov edx, [r13]
37b69d49amov rbp, [rsp+592]
37b69d4a2lea ecx, [rbx+1]
37b69d4a5mov esi, ebx
37b69d4a7mov rbx, r13
37b69d4aamov r13, r12
37b69d4admov r12, r14
37b69d4b0mov r14, r15
37b69d4b3mov r15, [rsp+584]
37b69d4bbnop [rax+rax]
37b69d4c0
37b69d4c0loc_37b69d4c0:
37b69d4c0movsxd r8, ecx
37b69d4c3lea edi, [rcx+1]
37b69d4c6mov eax, [rbp+4*r8-4]
37b69d4cbtest eax, eax
37b69d4cdmov r9l, eax
37b69d4d0cmovns edi, ecx
37b69d4d3neg r9l
37b69d4d6test eax, eax
37b69d4d8cmovs eax, r9l
37b69d4dcadd edi, 1
37b69d4dfcmp eax, ecx
37b69d4e1je 0x37b69dd52
37b69d4e7cmp edi, edx
37b69d4e9jle 0x37b69dd10
37b69d4ef
37b69d4efloc_37b69d4ef:
37b69d4efmov ebx, esi
37b69d4f1mov esi, edx
37b69d4f3mov r12, r13
37b69d4f6sub esi, ebx
37b69d4f8
37b69d4f8loc_37b69d4f8:
37b69d4f8mov [r12], esi
37b69d4fcmovups x6, [rsp+384]
37b69d504movups x7, [rsp+400]
37b69d50cmovups x8, [rsp+416]
37b69d515movups x9, [rsp+432]
37b69d51emovups x10, [rsp+448]
37b69d527add rsp, 472
37b69d52epop rbx
37b69d52fpop rsi
37b69d530pop rdi
37b69d531pop rbp
37b69d532pop r12
37b69d534pop r13
37b69d536pop r14
37b69d538pop r15
37b69d53aret
37b69d53b
37b69d53b 0f 1f 44 00 00 ..D..
37b69d540
37b69d540loc_37b69d540:
37b69d540lea rbx, [rsp+376]
37b69d548mov eax, [r13]
37b69d54cmov [rsp+568], r12
37b69d554mov ebp, 1
37b69d559mov [rsp+184], rbx
37b69d561lea rbx, [rsp+380]
37b69d569movss x7, [rip+0xd96d9f]
37b69d571pxor x6, x6
37b69d575mov [rsp+136], rbx
37b69d57dlea rbx, [rip+0xd96d58]
37b69d584mov [rsp+224], rbx
37b69d58clea rbx, [rip+0xd96d55]
37b69d593mov [rsp+232], rbx
37b69d59bnop [rax+rax]
37b69d5a0
37b69d5a0loc_37b69d5a0:
37b69d5a0mov rbx, [rsp+560]
37b69d5a8mov edx, [rbx]
37b69d5aacmp edx, ebp
37b69d5acjg 0x37b69d5c0
37b69d5aecmp edx, eax
37b69d5b0jl 0x37b69d7e4
37b69d5b6nop [rax+rax]
37b69d5c0
37b69d5c0loc_37b69d5c0:
37b69d5c0cmp ebp, eax
37b69d5c2jg 0x37b69d7e4
37b69d5c8mov rsi, [rsp+152]
37b69d5d0movsxd rbx, ebp
37b69d5d3mov rdi, [rsp+144]
37b69d5dblea r11, [rip+0xd96cf2]
37b69d5e2sub eax, ebp
37b69d5e4mov [rsp+32], r11
37b69d5e9mov r8, r11
37b69d5ecpxor x9, x9
37b69d5f1imul rsi, rbx
37b69d5f5mov r12, [rsp+160]
37b69d5fdadd eax, 1
37b69d600mov [rsp+380], eax
37b69d607imul r12, rbx
37b69d60badd rdi, rsi
37b69d60emov [rsp+296], rsi
37b69d616lea rsi, [rbx+rdi]
37b69d61amov [rsp+264], rdi
37b69d622lea edi, [rbp-1]
37b69d625lea rdx, [r14+4*rsi]
37b69d629mov [rsp+256], rsi
37b69d631lea rsi, [rbx-1]
37b69d635mov [rsp+192], edi
37b69d63clea rdi, [4*rsi]
37b69d644mov [rsp+208], rsi
37b69d64cmov rsi, [rsp+168]
37b69d654lea rcx, [r15+rdi]
37b69d658mov [rsp+216], rcx
37b69d660lea rcx, [r14+rdi]
37b69d664add rsi, r12
37b69d667mov [rsp+200], rcx
37b69d66fmov rcx, [rsp+136]
37b69d677mov [rsp+248], rsi
37b69d67fadd rsi, rbx
37b69d682lea rax, [r15+4*rsi]
37b69d686mov [rsp+288], rdx
37b69d68emov r9, rax
37b69d691mov [rsp+176], rax
37b69d699mov [rsp+280], rdi
37b69d6a1call scopy_
37b69d6a6mov r10l, [r13]
37b69d6aamov edx, [rsp+192]
37b69d6b1lea r11, [rip+0xd96c1c]
37b69d6b8mov [rsp+80], r11
37b69d6bdmov r11, [rsp+232]
37b69d6c5lea r9, [rip+0xd96c0c]
37b69d6ccsub r10l, ebp
37b69d6cfmov [rsp+380], edx
37b69d6d6mov rdx, [rsp+216]
37b69d6delea eax, [r10+1]
37b69d6e2mov r10, [rsp+608]
37b69d6eamov [rsp+88], 12
37b69d6f3mov [rsp+376], eax
37b69d6famov rax, [rsp+176]
37b69d702mov [rsp+48], rdx
37b69d707mov rdx, [rsp+584]
37b69d70fmov [rsp+72], rax
37b69d714mov rax, [rsp+200]
37b69d71cmov [rsp+40], rdx
37b69d721mov rcx, [rsp+224]
37b69d729mov [rsp+32], rax
37b69d72emov r8, [rsp+136]
37b69d736mov [rsp+64], r11
37b69d73bmov rdx, [rsp+184]
37b69d743mov [rsp+56], r10
37b69d748call sgemv_
37b69d74dlea ecx, [rbp+1]
37b69d750mov eax, [r13]
37b69d754movss x8, [r15+4*rsi]
37b69d75amov [rsp+304], ecx
37b69d761movsxd rcx, ecx
37b69d764andps x8, [rip+0xd96b94]
37b69d76cmov [rsp+216], rcx
37b69d774cmp eax, ebp
37b69d776jg 0x37b69d9b0
37b69d77c
37b69d77cloc_37b69d77c:
37b69d77cmovaps x0, x9
37b69d780maxss x0, x8
37b69d785ucomiss x0, x6
37b69d788jp 0x37b69d940
37b69d78ejne 0x37b69d940
37b69d794mov rax, [rsp+616]
37b69d79cmov eax, [rax]
37b69d79etest eax, eax
37b69d7a0jne 0x37b69d7ac
37b69d7a2mov rax, [rsp+616]
37b69d7aamov [rax], ebp
37b69d7ac
37b69d7acloc_37b69d7ac:
37b69d7acmov eax, [r13]
37b69d7b0
37b69d7b0loc_37b69d7b0:
37b69d7b0mov rbx, [rsp+592]
37b69d7b8mov rdi, [rsp+208]
37b69d7c0mov [rbx+4*rdi], ebp
37b69d7c3mov rbx, [rsp+560]
37b69d7cbmov ebp, [rsp+304]
37b69d7d2mov edx, [rbx]
37b69d7d4cmp edx, ebp
37b69d7d6jg 0x37b69d5c0
37b69d7dccmp edx, eax
37b69d7dejge 0x37b69d5c0
37b69d7e4
37b69d7e4loc_37b69d7e4:
37b69d7e4sub eax, ebp
37b69d7e6mov rdx, [rsp+152]
37b69d7eemov rdi, [rsp+136]
37b69d7f6lea esi, [rbp-1]
37b69d7f9add eax, 1
37b69d7fcmov rbx, [rsp+584]
37b69d804mov r9, [rsp+184]
37b69d80clea r8, [rip+0xd96ad9]
37b69d813mov [rsp+376], eax
37b69d81amovsxd rax, ebp
37b69d81dmov r12, [rsp+568]
37b69d825imul rdx, rax
37b69d829add rdx, [rsp+144]
37b69d831mov [rsp+32], rdi
37b69d836lea rcx, [4*rax-4]
37b69d83eadd rax, rdx
37b69d841mov [rsp+96], rbx
37b69d846mov rdx, [rsp+224]
37b69d84elea rax, [r14+4*rax]
37b69d852mov [rsp+380], esi
37b69d859mov [rsp+88], rax
37b69d85emov rax, [rsp+232]
37b69d866mov [rsp+120], 9
37b69d86fmov [rsp+80], rax
37b69d874mov rax, [rsp+608]
37b69d87cmov [rsp+112], 12
37b69d885mov [rsp+72], rax
37b69d88alea rax, [r15+rcx]
37b69d88eadd rcx, r14
37b69d891mov [rsp+64], rax
37b69d896mov rax, [rsp+584]
37b69d89emov [rsp+48], rcx
37b69d8a3lea rcx, [rip+0xd96a50]
37b69d8aamov [rsp+56], rax
37b69d8aflea rax, [rip+0xd96a22]
37b69d8b6mov [rsp+40], rax
37b69d8bbmov [rsp+104], 5
37b69d8c4call sgemmtr_
37b69d8c9mov rbx, [rsp+584]
37b69d8d1mov rbp, [rsp+592]
37b69d8d9mov ecx, esi
37b69d8dbjmp 0x37b69d911
37b69d8dd
37b69d8dd 0f 1f 00 ...
37b69d8e0
37b69d8e0loc_37b69d8e0:
37b69d8e0test edx, edx
37b69d8e2jle 0x37b69d93b
37b69d8e4mov [rsp+32], rbx
37b69d8e9sub eax, 1
37b69d8eclea r9, [r14+4*r8]
37b69d8f0mov rcx, rdi
37b69d8f3cdqe
37b69d8f5mov r8, rbx
37b69d8f8lea rdx, [r14+4*rax]
37b69d8fccall sswap_
37b69d901mov ecx, [rsp+380]
37b69d908
37b69d908loc_37b69d908:
37b69d908cmp ecx, 1
37b69d90bjle 0x37b69d4f8
37b69d911
37b69d911loc_37b69d911:
37b69d911lea edx, [rcx-1]
37b69d914movsxd r8, edx
37b69d917mov eax, [rbp+4*r8]
37b69d91ctest eax, eax
37b69d91emov r9l, eax
37b69d921cmovns edx, ecx
37b69d924neg r9l
37b69d927test eax, eax
37b69d929cmovs eax, r9l
37b69d92dsub edx, 1
37b69d930mov [rsp+380], edx
37b69d937cmp eax, ecx
37b69d939jne 0x37b69d8e0
37b69d93b
37b69d93bloc_37b69d93b:
37b69d93bmov ecx, edx
37b69d93djmp 0x37b69d908
37b69d93f
37b69d93f 90 .
37b69d940
37b69d940loc_37b69d940:
37b69d940movaps x10, x9
37b69d944mulss x10, x7
37b69d949comiss x8, x10
37b69d94djb 0x37b69da08
37b69d953
37b69d953loc_37b69d953:
37b69d953mov [rsp+240], ebp
37b69d95a
37b69d95aloc_37b69d95a:
37b69d95amov eax, [r13]
37b69d95emov r9, [rsp+288]
37b69d966mov rdx, [rsp+176]
37b69d96emov rcx, [rsp+136]
37b69d976sub eax, ebp
37b69d978add eax, 1
37b69d97bmov [rsp+380], eax
37b69d982lea rax, [rip+0xd9694b]
37b69d989mov [rsp+32], rax
37b69d98emov r8, rax
37b69d991call scopy_
37b69d996mov eax, [r13]
37b69d99acmp eax, ebp
37b69d99cjg 0x37b69e3d0
37b69d9a2mov ebp, [rsp+240]
37b69d9a9jmp 0x37b69d7b0
37b69d9ae
37b69d9ae 66 90 f.
37b69d9b0
37b69d9b0loc_37b69d9b0:
37b69d9b0sub eax, ebp
37b69d9b2lea r8, [rip+0xd9691b]
37b69d9b9mov [rsp+380], eax
37b69d9c0mov rax, [rsp+248]
37b69d9c8add rax, rcx
37b69d9cbmov rcx, [rsp+136]
37b69d9d3lea rdx, [r15+4*rax]
37b69d9d7call isamax_
37b69d9dcadd eax, ebp
37b69d9demov [rsp+364], eax
37b69d9e5cdqe
37b69d9e7add rax, [rsp+248]
37b69d9efmovss x9, [r15+4*rax]
37b69d9f5andps x9, [rip+0xd96903]
37b69d9fdjmp 0x37b69d77c
37b69da02
37b69da02 66 0f 1f 44 00 00 f..D..
37b69da08
37b69da08loc_37b69da08:
37b69da08mov eax, [rsp+364]
37b69da0fmov rdx, [rsp+168]
37b69da17add r12, [rsp+160]
37b69da1fmov r8, [rsp+264]
37b69da27add rdx, r12
37b69da2amovsxd r12, eax
37b69da2dmov [rsp+240], eax
37b69da34mov ecx, eax
37b69da36lea rax, [r12+r8]
37b69da3asub ecx, ebp
37b69da3clea r8, [rip+0xd96891]
37b69da43add rbx, rdx
37b69da46mov [rsp+32], r8
37b69da4blea rbx, [r15+4*rbx]
37b69da4fmov r8, [rsp+584]
37b69da57mov [rsp+308], ecx
37b69da5emov r9, rbx
37b69da61mov [rsp+380], ecx
37b69da68mov rcx, [rsp+136]
37b69da70mov [rsp+272], rdx
37b69da78lea rdx, [r14+4*rax]
37b69da7ccall scopy_
37b69da81mov r9l, [r13]
37b69da85sub r9l, [rsp+240]
37b69da8dlea r8, [rip+0xd96840]
37b69da94lea eax, [r9+1]
37b69da98mov r10, [rsp+272]
37b69daa0mov [rsp+32], r8
37b69daa5mov [rsp+380], eax
37b69daacmov rax, [rsp+152]
37b69dab4lea r9, [r10+r12]
37b69dab8mov rcx, [rsp+136]
37b69dac0imul rax, r12
37b69dac4add rax, [rsp+144]
37b69daccmov [rsp+336], r9
37b69dad4lea r9, [r15+4*r9]
37b69dad8lea r10, [r12+rax]
37b69dadcmov [rsp+344], rax
37b69dae4lea rdx, [r14+4*r10]
37b69dae8mov [rsp+320], r10
37b69daf0call scopy_
37b69daf5lea r8, [rip+0xd967d8]
37b69dafcmov eax, [r13]
37b69db00mov r9l, [rsp+192]
37b69db08mov [rsp+80], r8
37b69db0dmov r8, [rsp+232]
37b69db15sub eax, ebp
37b69db17mov [rsp+72], rbx
37b69db1cmov rdx, [rsp+184]
37b69db24mov [rsp+64], r8
37b69db29mov r8, [rsp+608]
37b69db31add eax, 1
37b69db34mov [rsp+376], eax
37b69db3blea rax, [4*r12-4]
37b69db43mov rcx, [rsp+224]
37b69db4bmov [rsp+56], r8
37b69db50mov r8, [rsp+584]
37b69db58mov [rsp+328], rax
37b69db60add rax, r15
37b69db63mov [rsp+40], r8
37b69db68mov r8, [rsp+200]
37b69db70mov [rsp+48], rax
37b69db75mov [rsp+32], r8
37b69db7amov r8, [rsp+136]
37b69db82mov [rsp+88], 12
37b69db8bmov [rsp+380], r9l
37b69db93lea r9, [rip+0xd9673e]
37b69db9amov [rsp+312], rax
37b69dba2call sgemv_
37b69dba7mov r11l, [rsp+308]
37b69dbafmov rcx, [rsp+136]
37b69dbb7mov rdx, rbx
37b69dbbalea r8, [rip+0xd96713]
37b69dbc1mov [rsp+380], r11l
37b69dbc9call isamax_
37b69dbceadd eax, [rsp+192]
37b69dbd5mov r10, [rsp+320]
37b69dbddcdqe
37b69dbdfadd rax, [rsp+272]
37b69dbe7movss x0, [r15+4*rax]
37b69dbedmov eax, [r13]
37b69dbf1cmp [rsp+240], eax
37b69dbf8andps x0, [rip+0xd96701]
37b69dbffjl 0x37b69e348
37b69dc05
37b69dc05loc_37b69dc05:
37b69dc05divss x9, x0
37b69dc0amulss x9, x10
37b69dc0fcomiss x8, x9
37b69dc13jae 0x37b69d953
37b69dc19mov rcx, [rsp+144]
37b69dc21mov rax, [rsp+296]
37b69dc29mulss x0, x7
37b69dc2dadd rax, [rsp+152]
37b69dc35add rcx, rax
37b69dc38mov [rsp+320], rcx
37b69dc40add rcx, [rsp+216]
37b69dc48mov [rsp+200], rcx
37b69dc50mov rcx, [rsp+336]
37b69dc58movss x1, [r15+4*rcx]
37b69dc5eandps x1, [rip+0xd9669b]
37b69dc65comiss x1, x0
37b69dc68jae 0x37b69e928
37b69dc6emov ebx, [rsp+304]
37b69dc75mov edx, [rsp+240]
37b69dc7clea ecx, [rbp+2]
37b69dc7fmov [rsp+368], ebx
37b69dc86cmp edx, ebx
37b69dc88jne 0x37b69e440
37b69dc8enop
37b69dc90
37b69dc90loc_37b69dc90:
37b69dc90mov rdx, [rsp+216]
37b69dc98mov rbx, [rsp+272]
37b69dca0mov eax, [r13]
37b69dca4movq x1, [r15+4*rsi]
37b69dcaaadd rdx, rbx
37b69dcadmovss x9, [r15+4*rdx]
37b69dcb3lea edx, [rax-1]
37b69dcb6cmp edx, ebp
37b69dcb8jg 0x37b69e0b0
37b69dcbe
37b69dcbeloc_37b69dcbe:
37b69dcbemov rbx, [rsp+256]
37b69dcc6mov rdi, [rsp+208]
37b69dccemov ebp, ecx
37b69dcd0movlps [r14+4*rbx], x1
37b69dcd5mov rbx, [rsp+200]
37b69dcddmovss [r14+4*rbx], x9
37b69dce3mov ebx, [rsp+240]
37b69dceaneg ebx
37b69dcecmovd x0, ebx
37b69dcf0mov rbx, [rsp+592]
37b69dcf8pshufd x0, x0, -32
37b69dcfdmovq [rbx+4*rdi], x0
37b69dd02jmp 0x37b69d5a0
37b69dd07
37b69dd07 66 0f 1f 84 00 00 00 00 00 f........
37b69dd10
37b69dd10loc_37b69dd10:
37b69dd10sub edx, edi
37b69dd12mov [rsp+32], r15
37b69dd17cdqe
37b69dd19mov rcx, [rsp+136]
37b69dd21add edx, 1
37b69dd24mov [rsp+380], edx
37b69dd2bmovsxd rdx, edi
37b69dd2eimul rdx, [rsp+152]
37b69dd37add rdx, r14
37b69dd3aadd r8, rdx
37b69dd3dadd rax, rdx
37b69dd40lea rdx, [r12+4*rax]
37b69dd44lea r9, [r12+4*r8]
37b69dd48mov r8, r15
37b69dd4bcall sswap_
37b69dd50mov edx, [rbx]
37b69dd52
37b69dd52loc_37b69dd52:
37b69dd52cmp edi, edx
37b69dd54jge 0x37b69d4ef
37b69dd5amov ecx, edi
37b69dd5cjmp 0x37b69d4c0
37b69dd61
37b69dd61 0f 1f 80 00 00 00 00 .......
37b69dd68
37b69dd68loc_37b69dd68:
37b69dd68movaps x10, x9
37b69dd6cmulss x10, x7
37b69dd71comiss x6, x10
37b69dd75jb 0x37b69def0
37b69dd7bimul r12, [rsp+152]
37b69dd84mov rax, [rsp+144]
37b69dd8clea rdx, [r12+rax]
37b69dd90
37b69dd90loc_37b69dd90:
37b69dd90lea rax, [rip+0xd9653d]
37b69dd97lea r9, [r14+4*rdx+4]
37b69dd9cmov rcx, [rsp+184]
37b69dda4mov rdx, [rsp+176]
37b69ddacmov [rsp+32], rax
37b69ddb1mov r8, rax
37b69ddb4call scopy_
37b69ddb9movaps x0, x8
37b69ddbdlea r9, [rip+0xd96510]
37b69ddc4movsxd r12, [rsp+376]
37b69ddccmov rax, [rsp+152]
37b69ddd4mov rcx, [rsp+136]
37b69dddcmov rdi, r12
37b69dddfimul rax, r12
37b69dde3add rax, [rsp+144]
37b69ddebsub edi, 1
37b69ddeelea rdx, [r12+rax]
37b69ddf2lea r8, [r14+4*rax+4]
37b69ddf7mov [rsp+380], edi
37b69ddfedivss x0, [r14+4*rdx]
37b69de04lea rdx, [rsp+372]
37b69de0cmovss [rsp+372], x0
37b69de15call sscal_
37b69de1ajmp 0x37b69d366
37b69de1f
37b69de1f 90 .
37b69de20
37b69de20loc_37b69de20:
37b69de20sub eax, ebx
37b69de22movsxd rdx, ebx
37b69de25add ebx, 1
37b69de28mov [rsp+88], 12
37b69de31mov [rsp+380], eax
37b69de38lea rax, [rip+0xd96495]
37b69de3fmovsxd rbx, ebx
37b69de42mov r8, [rsp+136]
37b69de4amov [rsp+80], rax
37b69de4fmov rax, [rsp+176]
37b69de57lea r9, [rip+0xd9647a]
37b69de5eimul rbx, [rsp+152]
37b69de67mov rcx, [rsp+224]
37b69de6fmov [rsp+72], rax
37b69de74mov rax, [rsp+232]
37b69de7cmov [rsp+64], rax
37b69de81mov rax, [rsp+608]
37b69de89mov [rsp+56], rax
37b69de8emov rax, [rsp+160]
37b69de96add rax, rsi
37b69de99add rax, [rsp+168]
37b69dea1add rax, rdx
37b69dea4mov rdx, [rsp+184]
37b69deaclea rax, [r15+4*rax]
37b69deb0mov [rsp+48], rax
37b69deb5mov rax, [rsp+584]
37b69debdmov [rsp+40], rax
37b69dec2mov rax, [rsp+144]
37b69decalea rax, [rax+rbx+1]
37b69decflea rax, [r14+4*rax]
37b69ded3mov [rsp+32], rax
37b69ded8call sgemv_
37b69deddmov ebx, [rsp+376]
37b69dee4jmp 0x37b69d2d9
37b69dee9
37b69dee9 0f 1f 80 00 00 00 00 .......
37b69def0
37b69def0loc_37b69def0:
37b69def0mov rax, [rsp+168]
37b69def8mov r10, rsi
37b69defbsub r10, [rsp+160]
37b69df03lea rcx, [rsp+364]
37b69df0bmov rdx, [rsp+144]
37b69df13mov [rsp+216], r10
37b69df1blea rdi, [r10+rax]
37b69df1fmovsxd rax, [rsp+364]
37b69df27lea r10, [rip+0xd963a6]
37b69df2emov [rsp+32], r10
37b69df33lea r9, [r15+4*rdi+4]
37b69df38mov r8, r10
37b69df3bimul rax, [rsp+152]
37b69df44mov [rsp+248], r9
37b69df4clea rax, [rdx+rax+1]
37b69df51lea rdx, [r14+4*rax]
37b69df55call scopy_
37b69df5amov r12l, [rsp+364]
37b69df62lea r10, [rip+0xd9636b]
37b69df69mov eax, ebx
37b69df6bmov [rsp+32], r10
37b69df70mov r8, [rsp+584]
37b69df78lea r11l, [r12+1]
37b69df7dsub eax, r12l
37b69df80movsxd rdx, r11l
37b69df83mov [rsp+240], r11l
37b69df8bmovsxd r11, r12l
37b69df8elea rcx, [rdi+rdx]
37b69df92mov [rsp+380], eax
37b69df99imul rdx, [rsp+152]
37b69dfa2add rdx, [rsp+144]
37b69dfaalea r9, [r15+4*rcx]
37b69dfaemov [rsp+208], eax
37b69dfb5add rdx, r11
37b69dfb8mov rcx, [rsp+136]
37b69dfc0mov [rsp+200], r9
37b69dfc8lea rdx, [r14+4*rdx]
37b69dfccmov [rsp+192], r11
37b69dfd4call scopy_
37b69dfd9mov edx, [r13]
37b69dfddmov eax, [rsp+208]
37b69dfe4mov r10, [rsp+216]
37b69dfeccmp edx, ebx
37b69dfeejg 0x37b69e270
37b69dff4
37b69dff4loc_37b69dff4:
37b69dff4mov rdx, [rsp+200]
37b69dffcmov rcx, [rsp+136]
37b69e004lea r8, [rip+0xd962c9]
37b69e00bmov [rsp+380], eax
37b69e012call isamax_
37b69e017add eax, r12l
37b69e01acdqe
37b69e01cadd rax, rdi
37b69e01fmovss x0, [r15+4*rax]
37b69e025andps x0, [rip+0xd962d4]
37b69e02ccmp r12l, 1
37b69e030jle 0x37b69e082
37b69e032mov rdx, [rsp+248]
37b69e03amov rcx, [rsp+136]
37b69e042lea eax, [r12-1]
37b69e047lea r8, [rip+0xd96286]
37b69e04emovss [rsp+200], x0
37b69e057mov [rsp+380], eax
37b69e05ecall isamax_
37b69e063movss x0, [rsp+200]
37b69e06ccdqe
37b69e06eadd rax, rdi
37b69e071movss x1, [r15+4*rax]
37b69e077andps x1, [rip+0xd96282]
37b69e07emaxss x0, x1
37b69e082
37b69e082loc_37b69e082:
37b69e082divss x9, x0
37b69e087mulss x9, x10
37b69e08ccomiss x6, x9
37b69e090jb 0x37b69e188
37b69e096movsxd rdx, ebx
37b69e099imul rdx, [rsp+152]
37b69e0a2add rdx, [rsp+144]
37b69e0aajmp 0x37b69dd90
37b69e0af
37b69e0af 90 .
37b69e0b0
37b69e0b0loc_37b69e0b0:
37b69e0b0cmp eax, ecx
37b69e0b2jl 0x37b69dcbe
37b69e0b8movshdup x0, x1
37b69e0bcmovaps x5, x9
37b69e0c0movaps x4, x1
37b69e0c3movsxd rdx, ecx
37b69e0c6movss x3, [rip+0xd96246]
37b69e0celea r11, [rbx+rdx]
37b69e0d2mov r9l, eax
37b69e0d5mov r8, [rsp+248]
37b69e0ddmov rbx, [rsp+264]
37b69e0e5mov rdi, [rsp+320]
37b69e0edsub r9l, ecx
37b69e0f0lea r11, [r15+4*r11]
37b69e0f4add r8, rdx
37b69e0f7add r9, 1
37b69e0fbdivss x5, x0
37b69e0ffadd rbx, rdx
37b69e102lea r8, [r15+4*r8]
37b69e106add rdx, rdi
37b69e109lea rsi, [r14+4*rdx]
37b69e10dlea rbx, [r14+4*rbx]
37b69e111xor edx, edx
37b69e113divss x4, x0
37b69e117movaps x8, x5
37b69e11bmulss x8, x4
37b69e120subss x8, [rip+0xd961eb]
37b69e129divss x3, x8
37b69e12edivss x3, x0
37b69e132nop [rax]
37b69e135nop [rax+rax]
37b69e140
37b69e140loc_37b69e140:
37b69e140movss x2, [r8+4*rdx]
37b69e146movss x0, [r11+4*rdx]
37b69e14cmovaps x8, x2
37b69e150mulss x8, x5
37b69e155subss x8, x0
37b69e15amulss x0, x4
37b69e15emulss x8, x3
37b69e163subss x0, x2
37b69e167mulss x0, x3
37b69e16bmovss [rbx+4*rdx], x8
37b69e171movss [rsi+4*rdx], x0
37b69e176add rdx, 1
37b69e17acmp r9, rdx
37b69e17djne 0x37b69e140
37b69e17fjmp 0x37b69dcbe
37b69e184
37b69e184 0f 1f 40 00 ..@.
37b69e188
37b69e188loc_37b69e188:
37b69e188mulss x0, x7
37b69e18cadd rdi, [rsp+192]
37b69e194movss x1, [r15+4*rdi]
37b69e19aandps x1, [rip+0xd9615f]
37b69e1a1comiss x1, x0
37b69e1a4jae 0x37b69e8c8
37b69e1aalea r10l, [rbx-1]
37b69e1aemovsxd rax, ebx
37b69e1b1mov [rsp+200], r10l
37b69e1b9cmp r10l, r12l
37b69e1bcjne 0x37b69e578
37b69e1c2mov rdx, [rsp+152]
37b69e1caimul rdx, rax
37b69e1ceadd rdx, [rsp+144]
37b69e1d6nop [rax+rax]
37b69e1e0
37b69e1e0loc_37b69e1e0:
37b69e1e0movsxd r8, r10l
37b69e1e3lea rcx, [rbp+rax]
37b69e1e8lea edi, [rbx-2]
37b69e1ebsub rsi, [rsp+160]
37b69e1f3add rsi, [rsp+168]
37b69e1fbmovss x6, [r15+4*rcx]
37b69e201mov r9, [rsp+152]
37b69e209lea rcx, [rsi+r8]
37b69e20dmovss x9, [r15+4*rcx]
37b69e213lea rcx, [rbp+r8]
37b69e218imul r9, r8
37b69e21cmovss x1, [r15+4*rcx]
37b69e222add r9, [rsp+144]
37b69e22acmp ebx, 2
37b69e22djg 0x37b69e750
37b69e233
37b69e233loc_37b69e233:
37b69e233neg r12l
37b69e236mov rbx, [rsp+592]
37b69e23eadd r9, r8
37b69e241add r8, rdx
37b69e244movd x0, r12l
37b69e249add rdx, rax
37b69e24cmovss [r14+4*r9], x9
37b69e252pshufd x0, x0, -32
37b69e257movss [r14+4*r8], x1
37b69e25dmovq [rbx+4*rax-8], x0
37b69e263movss [r14+4*rdx], x6
37b69e269jmp 0x37b69d373
37b69e26e
37b69e26e 66 90 f.
37b69e270
37b69e270loc_37b69e270:
37b69e270lea rax, [rip+0xd9605d]
37b69e277sub edx, ebx
37b69e279add ebx, 1
37b69e27cmov [rsp+88], 12
37b69e285mov [rsp+80], rax
37b69e28amov rax, [rsp+248]
37b69e292movsxd rbx, ebx
37b69e295lea r9, [rip+0xd9603c]
37b69e29cimul rbx, [rsp+152]
37b69e2a5mov [rsp+380], edx
37b69e2acmov [rsp+72], rax
37b69e2b1mov rax, [rsp+232]
37b69e2b9mov r8, [rsp+136]
37b69e2c1mov rdx, [rsp+184]
37b69e2c9mov [rsp+64], rax
37b69e2cemov rax, [rsp+608]
37b69e2d6mov rcx, [rsp+224]
37b69e2demov [rsp+56], rax
37b69e2e3mov rax, [rsp+160]
37b69e2eblea rax, [r10+2*rax]
37b69e2efadd rax, [rsp+168]
37b69e2f7add rax, [rsp+192]
37b69e2fflea rax, [r15+4*rax]
37b69e303mov [rsp+48], rax
37b69e308mov rax, [rsp+584]
37b69e310mov [rsp+40], rax
37b69e315mov rax, [rsp+144]
37b69e31dlea rax, [rax+rbx+1]
37b69e322lea rax, [r14+4*rax]
37b69e326mov [rsp+32], rax
37b69e32bcall sgemv_
37b69e330mov ebx, [rsp+376]
37b69e337mov eax, ebx
37b69e339sub eax, r12l
37b69e33cjmp 0x37b69dff4
37b69e341
37b69e341 0f 1f 80 00 00 00 00 .......
37b69e348
37b69e348loc_37b69e348:
37b69e348mov ecx, [rsp+240]
37b69e34flea r8, [rip+0xd95f7e]
37b69e356mov [rsp+200], r10
37b69e35emovss [rsp+320], x0
37b69e367sub eax, ecx
37b69e369mov [rsp+380], eax
37b69e370lea eax, [rcx+1]
37b69e373mov rcx, [rsp+136]
37b69e37bcdqe
37b69e37dadd rax, [rsp+272]
37b69e385lea rdx, [r15+4*rax]
37b69e389call isamax_
37b69e38eadd eax, [rsp+240]
37b69e395movss x0, [rsp+320]
37b69e39ecdqe
37b69e3a0add rax, [rsp+272]
37b69e3a8mov r10, [rsp+200]
37b69e3b0movss x1, [r15+4*rax]
37b69e3b6andps x1, [rip+0xd95f43]
37b69e3bdmaxss x0, x1
37b69e3c1jmp 0x37b69dc05
37b69e3c6
37b69e3c6 66 2e 0f 1f 84 00 00 00 00 00 f.........
37b69e3d0
37b69e3d0loc_37b69e3d0:
37b69e3d0sub eax, ebp
37b69e3d2mov rbx, [rsp+256]
37b69e3damov rcx, [rsp+136]
37b69e3e2lea rdx, [rsp+372]
37b69e3eamov [rsp+380], eax
37b69e3f1mov rax, [rsp+216]
37b69e3f9lea r9, [rip+0xd95ed4]
37b69e400add rax, [rsp+264]
37b69e408movss x0, [rip+0xd95f04]
37b69e410divss x0, [r14+4*rbx]
37b69e416movss [rsp+372], x0
37b69e41flea r8, [r14+4*rax]
37b69e423call sscal_
37b69e428mov eax, [r13]
37b69e42cmov ebp, [rsp+240]
37b69e433jmp 0x37b69d7b0
37b69e438
37b69e438 0f 1f 84 00 00 00 00 00 ........
37b69e440
37b69e440loc_37b69e440:
37b69e440sub edx, ebx
37b69e442lea rbx, [rdi+8]
37b69e446add rax, [rsp+152]
37b69e44emovsxd rcx, ecx
37b69e451mov [rsp+296], rbx
37b69e459lea rbx, [rdi+4]
37b69e45dadd rax, [rsp+144]
37b69e465mov edi, 2
37b69e46amov [rsp+308], edx
37b69e471mov rdx, [rsp+200]
37b69e479mov [rsp+280], rbx
37b69e481mov rbx, [rsp+320]
37b69e489
37b69e489loc_37b69e489:
37b69e489movss x0, [r14+4*rdx]
37b69e48fadd r12, rax
37b69e492mov edx, [rsp+308]
37b69e499add rcx, rbx
37b69e49cmov rax, [rsp+584]
37b69e4a4lea r9, [r14+4*r12]
37b69e4a8lea r8, [rip+0xd95e25]
37b69e4afmovss [r14+4*r10], x0
37b69e4b5sub edx, 1
37b69e4b8mov [rsp+32], rax
37b69e4bdmov [rsp+380], edx
37b69e4c4lea rdx, [r14+4*rcx]
37b69e4c8mov rcx, [rsp+136]
37b69e4d0call scopy_
37b69e4d5mov eax, [r13]
37b69e4d9cmp [rsp+240], eax
37b69e4e0jl 0x37b69e808
37b69e4e6
37b69e4e6loc_37b69e4e6:
37b69e4e6cmp ebp, 1
37b69e4e9jle 0x37b69e532
37b69e4ebmov eax, [rsp+192]
37b69e4f2mov rdx, [rsp+280]
37b69e4famov r8, [rsp+584]
37b69e502mov rcx, [rsp+136]
37b69e50amov [rsp+380], eax
37b69e511mov rax, [rsp+584]
37b69e519add rdx, r14
37b69e51cmov [rsp+32], rax
37b69e521mov rax, [rsp+328]
37b69e529lea r9, [r14+rax]
37b69e52dcall sswap_
37b69e532
37b69e532loc_37b69e532:
37b69e532mov rax, [rsp+296]
37b69e53amov r9, [rsp+312]
37b69e542lea rcx, [rsp+368]
37b69e54alea rdx, [r15+rax-4]
37b69e54fmov rax, [rsp+608]
37b69e557mov [rsp+32], rax
37b69e55cmov r8, rax
37b69e55fcall sswap_
37b69e564cmp edi, 1
37b69e567je 0x37b69d95a
37b69e56dlea ecx, [rbp+2]
37b69e570jmp 0x37b69dc90
37b69e575
37b69e575 0f 1f 00 ...
37b69e578
37b69e578loc_37b69e578:
37b69e578mov [rsp+304], 2
37b69e583movsxd rdi, [rsp+200]
37b69e58blea ecx, [rbx-2]
37b69e58e
37b69e58eloc_37b69e58e:
37b69e58emov [rsp+216], rax
37b69e596mov rax, [rsp+560]
37b69e59esub ecx, r12l
37b69e5a1mov edx, [rsp+200]
37b69e5a8mov r9, [rsp+192]
37b69e5b0mov [rsp+380], ecx
37b69e5b7add edx, [rax]
37b69e5b9mov rax, [rsp+152]
37b69e5c1sub edx, [r13]
37b69e5c5mov r11, [rsp+144]
37b69e5cdmov r10, rax
37b69e5d0mov [rsp+208], edx
37b69e5d7mov rdx, r9
37b69e5damovsxd rcx, [rsp+240]
37b69e5e2imul r10, rdi
37b69e5e6imul rdx, rax
37b69e5eaadd r10, r11
37b69e5edlea r8, [rdi+r10]
37b69e5f1add r11, rdx
37b69e5f4add rdx, rax
37b69e5f7mov rax, [rsp+584]
37b69e5ffmovss x0, [r14+4*r8]
37b69e605lea r8, [r9+r11]
37b69e609add rdx, [rsp+144]
37b69e611add rcx, r10
37b69e614add rdx, r9
37b69e617mov [rsp+248], r11
37b69e61fmovss [r14+4*r8], x0
37b69e625lea r9, [r14+4*rdx]
37b69e629lea rdx, [r14+4*rcx]
37b69e62dmov rcx, [rsp+136]
37b69e635mov [rsp+32], rax
37b69e63alea r8, [rip+0xd95c93]
37b69e641mov [rsp+240], r10
37b69e649call scopy_
37b69e64ecmp r12l, 1
37b69e652mov rax, [rsp+216]
37b69e65ajle 0x37b69e6a6
37b69e65clea rax, [rip+0xd95c71]
37b69e663mov r10, [rsp+240]
37b69e66bmov r11, [rsp+248]
37b69e673lea edx, [r12-1]
37b69e678mov [rsp+32], rax
37b69e67dmov rcx, [rsp+136]
37b69e685mov r8, rax
37b69e688mov [rsp+380], edx
37b69e68flea r9, [r14+4*r11+4]
37b69e694lea rdx, [r14+4*r10+4]
37b69e699call scopy_
37b69e69emov rax, [rsp+216]
37b69e6a6
37b69e6a6loc_37b69e6a6:
37b69e6a6mov edx, [r13]
37b69e6aacmp edx, ebx
37b69e6acjg 0x37b69e858
37b69e6b2
37b69e6b2loc_37b69e6b2:
37b69e6b2sub edx, [rsp+200]
37b69e6b9mov [rsp+216], rax
37b69e6c1mov rcx, [rsp+192]
37b69e6c9mov rax, [rsp+608]
37b69e6d1add edx, 1
37b69e6d4mov [rsp+380], edx
37b69e6dbmovsxd rdx, [rsp+208]
37b69e6e3imul rdx, [rsp+160]
37b69e6ecadd rdx, [rsp+168]
37b69e6f4mov [rsp+32], rax
37b69e6f9mov r8, rax
37b69e6fcadd rcx, rdx
37b69e6ffadd rdi, rdx
37b69e702lea r9, [r15+4*rcx]
37b69e706mov rcx, [rsp+136]
37b69e70elea rdx, [r15+4*rdi]
37b69e712call sswap_
37b69e717mov rdx, [rsp+152]
37b69e71fmov rax, [rsp+216]
37b69e727imul rdx, rax
37b69e72badd rdx, [rsp+144]
37b69e733cmp [rsp+304], 1
37b69e73bje 0x37b69ea22
37b69e741lea r10l, [rbx-1]
37b69e745jmp 0x37b69e1e0
37b69e74a
37b69e74a 66 0f 1f 44 00 00 f..D..
37b69e750
37b69e750loc_37b69e750:
37b69e750movaps x5, x6
37b69e753movaps x4, x9
37b69e757movaps x3, x8
37b69e75bmov r10l, r10l
37b69e75edivss x5, x1
37b69e762lea rsi, [r15+4*rsi]
37b69e766lea rbp, [r15+4*rbp]
37b69e76amov ecx, 1
37b69e76flea rbx, [r14+4*r9]
37b69e773lea r11, [r14+4*rdx]
37b69e777divss x4, x1
37b69e77bmovaps x0, x5
37b69e77emulss x0, x4
37b69e782subss x0, x8
37b69e787divss x3, x0
37b69e78bdivss x3, x1
37b69e78fnop [rax+rax]
37b69e794nop [rax+rax]
37b69e79fnop [rax+rax]
37b69e7aanop [rax+rax]
37b69e7b5nop [rax+rax]
37b69e7c0
37b69e7c0loc_37b69e7c0:
37b69e7c0movss x2, [rsi+4*rcx]
37b69e7c5movss x0, [rbp+4*rcx]
37b69e7cbmovaps x10, x2
37b69e7cfmulss x10, x5
37b69e7d4subss x10, x0
37b69e7d9mulss x0, x4
37b69e7ddmulss x10, x3
37b69e7e2subss x0, x2
37b69e7e6mulss x0, x3
37b69e7eamovss [rbx+4*rcx], x10
37b69e7f0movss [r11+4*rcx], x0
37b69e7f6add rcx, 1
37b69e7facmp rcx, r10
37b69e7fdjne 0x37b69e7c0
37b69e7ffjmp 0x37b69e233
37b69e804
37b69e804 0f 1f 40 00 ..@.
37b69e808
37b69e808loc_37b69e808:
37b69e808mov ecx, [rsp+240]
37b69e80fsub eax, ecx
37b69e811lea edx, [rcx+1]
37b69e814mov rcx, [rsp+136]
37b69e81cmov [rsp+380], eax
37b69e823mov rax, [rsp+344]
37b69e82bmovsxd rdx, edx
37b69e82eadd rax, rdx
37b69e831add rdx, rbx
37b69e834lea rbx, [rip+0xd95a99]
37b69e83bmov [rsp+32], rbx
37b69e840lea rdx, [r14+4*rdx]
37b69e844lea r9, [r14+4*rax]
37b69e848mov r8, rbx
37b69e84bcall scopy_
37b69e850jmp 0x37b69e4e6
37b69e855
37b69e855 0f 1f 00 ...
37b69e858
37b69e858loc_37b69e858:
37b69e858sub edx, ebx
37b69e85amov [rsp+216], rax
37b69e862mov rax, [rsp+192]
37b69e86amov [rsp+380], edx
37b69e871lea edx, [rbx+1]
37b69e874movsxd rdx, edx
37b69e877imul rdx, [rsp+152]
37b69e880add rdx, [rsp+144]
37b69e888lea rcx, [rax+rdx]
37b69e88cmov rax, [rsp+584]
37b69e894add rdx, rdi
37b69e897lea r9, [r14+4*rcx]
37b69e89bmov rcx, [rsp+136]
37b69e8a3lea rdx, [r14+4*rdx]
37b69e8a7mov [rsp+32], rax
37b69e8acmov r8, rax
37b69e8afcall sswap_
37b69e8b4mov edx, [r13]
37b69e8b8mov rax, [rsp+216]
37b69e8c0jmp 0x37b69e6b2
37b69e8c5
37b69e8c5 0f 1f 00 ...
37b69e8c8
37b69e8c8loc_37b69e8c8:
37b69e8c8lea rax, [rip+0xd95a05]
37b69e8cfmov r9, [rsp+176]
37b69e8d7mov rdx, [rsp+248]
37b69e8dfmov [rsp+32], rax
37b69e8e4mov rcx, [rsp+184]
37b69e8ecmov r8, rax
37b69e8efcall scopy_
37b69e8f4movsxd rdi, [rsp+376]
37b69e8fcmov [rsp+200], edi
37b69e903mov rbx, rdi
37b69e906mov rax, rdi
37b69e909cmp r12l, edi
37b69e90cje 0x37b69e9c0
37b69e912mov [rsp+304], 1
37b69e91dlea ecx, [rdi-1]
37b69e920jmp 0x37b69e58e
37b69e925
37b69e925 0f 1f 00 ...
37b69e928
37b69e928loc_37b69e928:
37b69e928mov eax, [r13]
37b69e92cmov r9, [rsp+176]
37b69e934mov rdx, rbx
37b69e937mov [rsp+336], r10
37b69e93fmov rcx, [rsp+136]
37b69e947sub eax, ebp
37b69e949add eax, 1
37b69e94cmov [rsp+380], eax
37b69e953lea rax, [rip+0xd9597a]
37b69e95amov [rsp+32], rax
37b69e95fmov r8, rax
37b69e962call scopy_
37b69e967mov [rsp+368], ebp
37b69e96ecmp [rsp+240], ebp
37b69e975je 0x37b69d953
37b69e97blea rax, [rdi+4]
37b69e97fmov rcx, [rsp+216]
37b69e987mov rdx, [rsp+256]
37b69e98fmov edi, 1
37b69e994mov [rsp+296], rax
37b69e99cmov rbx, [rsp+264]
37b69e9a4mov rax, [rsp+320]
37b69e9acmov r10, [rsp+336]
37b69e9b4jmp 0x37b69e489
37b69e9b9
37b69e9b9 0f 1f 80 00 00 00 00 .......
37b69e9c0
37b69e9c0loc_37b69e9c0:
37b69e9c0imul rdi, [rsp+152]
37b69e9c9mov rax, [rsp+144]
37b69e9d1mov ebx, r12l
37b69e9d4lea rdx, [rdi+rax]
37b69e9d8jmp 0x37b69dd90
37b69e9dd
37b69e9ddloc_37b69e9dd:
37b69e9ddlea rax, [rsp+376]
37b69e9e5xor edx, edx
37b69e9e7mov [rsp+184], rax
37b69e9eflea rax, [rsp+380]
37b69e9f7mov [rsp+136], rax
37b69e9fflea rax, [rip+0xd958d6]
37b69ea06mov [rsp+224], rax
37b69ea0elea rax, [rip+0xd958d3]
37b69ea15mov [rsp+232], rax
37b69ea1djmp 0x37b69d3ae
37b69ea22
37b69ea22loc_37b69ea22:
37b69ea22mov ebx, r12l
37b69ea25jmp 0x37b69dd90
37b69d100