News:

Masm32 SDK description, downloads and other helpful links
Message to All Guests

Main Menu

Algorithm that are. Hard to unroll and optimise breaks code?

Started by daydreamer, June 29, 2021, 08:44:34 PM

Previous topic - Next topic

daydreamer

I am working with the mod % way of testing primes
Going SIMT would be advantage, because spread division to multiple cores,because of lack of many execution units for division

So I start with usual c optimize techniques
Stepping 2 between odd numbers and restrict max prime test to sqrt(number), breaks code so numbers ending with 5 falls thru. Test
Unroll loop with or without SIMD?
That's a hard problem to solve,because uncertainly # of loops,before stop when find its not a prime
my none asm creations
https://masm32.com/board/index.php?topic=6937.msg74303#msg74303
I am an Invoker
"An Invoker is a mage who specializes in the manipulation of raw and elemental energies."
Like SIMD coding

hutch--

If you start off with the assumption that SSE is the technique to use, you are blocking most of the techniques that can be used. The idea is to design what you have to do then select the best technique to perform that task. Also give some thought to multi-threading different numeric ranges in parallel as this will improve your speed.

SSE is useful legacy technology but it remains to be seen if you can write that type of code for the task you have in mind.

jj2007

Quote from: hutch-- on June 29, 2021, 09:01:04 PMSSE is useful legacy technology

Some recent C compilers still haven't discovered that "legacy" technology :cool:

daydreamer

I tried double scalar mod test version, slow but works

You could do sieve in one thread,but interesting test would be 8 core cpu doing lots of divs,vs fewer cores cpu
my none asm creations
https://masm32.com/board/index.php?topic=6937.msg74303#msg74303
I am an Invoker
"An Invoker is a mage who specializes in the manipulation of raw and elemental energies."
Like SIMD coding

LiaoMi

Hi daydreamer,

https://stackoverflow.com/questions/17241028/finding-lists-of-prime-numbers-with-simd-sse-avx
https://stackoverflow.com/questions/8045807/do-i-get-a-performance-penalty-when-mixing-simd-instructions-and-multithreading

https://github.com/kozub/PrimalityMillerRabinTest-CUDA_GPU
https://github.com/MichaelBell/PrimesCUDA

https://www.openssl.org/docs/man1.1.0/man3/BN_is_prime_fasttest_ex.html
int BN_is_prime_fasttest_ex   (   const BIGNUM *    a,
int    checks,
BN_CTX *    ctx_passed,
int    do_trial_division,
BN_GENCB *    cb
)      

Fast isPrime() for n < 2^32 - https://www.mersenneforum.org/showthread.php?t=12209&page=1
Tables of pseudoprimes and related data - http://www.cecm.sfu.ca/Pseudoprimes/index-2-to-64.html

isPrime64Bit.c
mulmod(unsigned long long, unsigned long long, unsigned long long):                           # @mulmod(unsigned long long, unsigned long long, unsigned long long)
        mov     rcx, rdx
        mov     rax, rdi
        mul     rsi
        div     rcx
        mov     rcx, rdx

        mov     rax, rcx
        ret
powmod(unsigned long long, unsigned long long, unsigned long long):                           # @powmod(unsigned long long, unsigned long long, unsigned long long)
        test    rsi, rsi
        je      .LBB1_1
        mov     rcx, rdx
        mov     r8d, 1
        jmp     .LBB1_3
.LBB1_5:                                #   in Loop: Header=BB1_3 Depth=1
        mov     rax, rdi
        mul     rdi
        div     rcx
        mov     rdi, rdx

        shr     rsi
        je      .LBB1_6
.LBB1_3:                                # =>This Inner Loop Header: Depth=1
        test    sil, 1
        je      .LBB1_5
        mov     rax, r8
        mul     rdi
        div     rcx
        mov     r8, rdx

        jmp     .LBB1_5
.LBB1_6:
        mov     rax, r8
        ret
.LBB1_1:
        mov     eax, 1
        ret
miller_rabin(unsigned long long, int*, int):                   # @miller_rabin(unsigned long long, int*, int)
        push    rbp
        push    r15
        push    r14
        push    r12
        push    rbx
        lea     r8, [rdi - 1]
        xor     r12d, r12d
        mov     r11, r8
        test    r8b, 1
        jne     .LBB2_3
        xor     r12d, r12d
        mov     rax, r8
        mov     r11, r8
.LBB2_2:                                # =>This Inner Loop Header: Depth=1
        add     r12d, 1
        shr     r11
        test    al, 2
        mov     rax, r11
        je      .LBB2_2
.LBB2_3:
        mov     r9d, 1
        test    edx, edx
        jle     .LBB2_28
        test    r11, r11
        je      .LBB2_28
        mov     r10d, edx
        xor     r15d, r15d
        test    r12d, r12d
        je      .LBB2_8
        xor     r9d, r9d
        jmp     .LBB2_17
.LBB2_7:                                #   in Loop: Header=BB2_8 Depth=1
        add     r15, 1
        cmp     r15, r10
        je      .LBB2_29
.LBB2_8:                                # =>This Loop Header: Depth=1
        movsxd  r14, dword ptr [rsi + 4*r15]
        mov     ebx, 1
        mov     rbp, r14
        mov     rcx, r11
        jmp     .LBB2_10
.LBB2_9:                                #   in Loop: Header=BB2_10 Depth=2
        mov     rax, rbp
        mul     rbp
        div     rdi
        mov     rbp, rdx

        shr     rcx
        je      .LBB2_12
.LBB2_10:                               #   Parent Loop BB2_8 Depth=1
        test    cl, 1
        je      .LBB2_9
        mov     rax, rbx
        mul     rbp
        div     rdi
        mov     rbx, rdx

        jmp     .LBB2_9
.LBB2_12:                               #   in Loop: Header=BB2_8 Depth=1
        cmp     rbx, 1
        je      .LBB2_7
        cmp     rbx, r8
        je      .LBB2_7
        test    r14d, r14d
        je      .LBB2_7
        xor     r9d, r9d
        jmp     .LBB2_29
.LBB2_16:                               #   in Loop: Header=BB2_17 Depth=1
        add     r15, 1
        cmp     r15, r10
        je      .LBB2_28
.LBB2_17:                               # =>This Loop Header: Depth=1
        movsxd  r14, dword ptr [rsi + 4*r15]
        mov     ebx, 1
        mov     rcx, r14
        mov     rbp, r11
        jmp     .LBB2_19
.LBB2_18:                               #   in Loop: Header=BB2_19 Depth=2
        mov     rax, rcx
        mul     rcx
        div     rdi
        mov     rcx, rdx

        shr     rbp
        je      .LBB2_21
.LBB2_19:                               #   Parent Loop BB2_17 Depth=1
        test    bpl, 1
        je      .LBB2_18
        mov     rax, rbx
        mul     rcx
        div     rdi
        mov     rbx, rdx

        jmp     .LBB2_18
.LBB2_21:                               #   in Loop: Header=BB2_17 Depth=1
        cmp     rbx, 1
        je      .LBB2_16
        mov     ecx, r12d
        cmp     rbx, r8
        je      .LBB2_16
.LBB2_23:                               #   Parent Loop BB2_17 Depth=1
        mov     rax, rbx
        mul     rbx
        div     rdi
        mov     rbx, rdx

        cmp     rbx, 1
        je      .LBB2_29
        cmp     rbx, r8
        je      .LBB2_16
        add     ecx, -1
        jne     .LBB2_23
        test    r14d, r14d
        je      .LBB2_16
        jmp     .LBB2_29
.LBB2_28:
        mov     r9d, 1
.LBB2_29:
        mov     eax, r9d
        pop     rbx
        pop     r12
        pop     r14
        pop     r15
        pop     rbp
        ret
isPrime64Bit(unsigned long long):                      # @isPrime64Bit(unsigned long long)
        sub     rsp, 24
        movabs  rax, 1395864371202
        mov     qword ptr [rsp], rax
        mov     dword ptr [rsp + 8], 9375
        movabs  rax, 3141592653589793239
        imul    rax, rdi
        shr     rax, 40
        and     eax, 8188
        mov     eax, dword ptr [rax + SPRPlookup]
        mov     dword ptr [rsp + 12], eax
        cmp     rdi, 13
        ja      .LBB3_2
        mov     eax, 1
        mov     ecx, 8236
        bt      rcx, rdi
        jae     .LBB3_2
.LBB3_5:
        add     rsp, 24
        ret
.LBB3_2:
        xor     eax, eax
        cmp     rdi, 5
        jb      .LBB3_5
        mov     ecx, edi
        and     ecx, 1
        test    rcx, rcx
        je      .LBB3_5
        mov     rsi, rsp
        mov     edx, 4
        call    miller_rabin(unsigned long long, int*, int)
        add     rsp, 24
        ret
main:                                   # @main
        push    rbp
        push    r15
        push    r14
        push    r13
        push    r12
        push    rbx
        sub     rsp, 24
        mov     r15, rsi
        cmp     edi, 2
        jge     .LBB4_1
        mov     rsi, qword ptr [r15]
        mov     edi, offset .L.str
        xor     eax, eax
        call    printf
.LBB4_9:
        xor     eax, eax
        add     rsp, 24
        pop     rbx
        pop     r12
        pop     r13
        pop     r14
        pop     r15
        pop     rbp
        ret
.LBB4_1:
        mov     r12d, edi
        mov     ebx, 1
        movabs  rbp, 1395864371202
        movabs  r14, 3141592653589793239
        jmp     .LBB4_2
.LBB4_7:                                #   in Loop: Header=BB4_2 Depth=1
        mov     edx, offset .L.str.3
.LBB4_8:                                #   in Loop: Header=BB4_2 Depth=1
        mov     edi, offset .L.str.1
        mov     rsi, r13
        xor     eax, eax
        call    printf
        add     rbx, 1
        cmp     r12, rbx
        je      .LBB4_9
.LBB4_2:                                # =>This Inner Loop Header: Depth=1
        mov     rdi, qword ptr [r15 + 8*rbx]
        xor     esi, esi
        mov     edx, 10
        call    strtoull
        mov     r13, rax
        mov     qword ptr [rsp], rbp
        mov     dword ptr [rsp + 8], 9375
        imul    rax, r14
        shr     rax, 40
        and     eax, 8188
        mov     eax, dword ptr [rax + SPRPlookup]
        mov     dword ptr [rsp + 12], eax
        cmp     r13, 13
        ja      .LBB4_4
        mov     eax, 8236
        bt      rax, r13
        jae     .LBB4_4
        mov     edx, offset .L.str.2
        jmp     .LBB4_8
.LBB4_4:                                #   in Loop: Header=BB4_2 Depth=1
        cmp     r13, 5
        jb      .LBB4_7
        mov     eax, r13d
        and     eax, 1
        test    rax, rax
        je      .LBB4_7
        mov     rdi, r13
        mov     rsi, rsp
        mov     edx, 4
        call    miller_rabin(unsigned long long, int*, int)
        mov     edx, offset .L.str.2
        test    eax, eax
        jne     .LBB4_8
        jmp     .LBB4_7
SPRPlookup:
        .long   99                              # 0x63
        .long   447                             # 0x1bf
        .long   426                             # 0x1aa
..........

.L.str:
        .asciz  "Usage: %s <num>...\n"

.L.str.1:
        .asciz  "%llu: %s\n"

.L.str.2:
        .asciz  "true"

.L.str.3:
        .asciz  "false"


LiaoMi

Or

isPrime64Bit.c
clang -C -O3 -mavx2 -fno-slp-vectorize -fno-vectorize -mllvm -unroll-count=20

daydreamer

my none asm creations
https://masm32.com/board/index.php?topic=6937.msg74303#msg74303
I am an Invoker
"An Invoker is a mage who specializes in the manipulation of raw and elemental energies."
Like SIMD coding

daydreamer

I could maybe try AVX version
seen output is the performance hog,so 1 lowpriority thread taking care of IO and others are x numbers of highpriority threads,x number of cores?would that be good solution?


my none asm creations
https://masm32.com/board/index.php?topic=6937.msg74303#msg74303
I am an Invoker
"An Invoker is a mage who specializes in the manipulation of raw and elemental energies."
Like SIMD coding

daydreamer

Quote from: daydreamer on July 06, 2021, 11:30:20 PM
I could maybe try AVX version
seen output is the performance hog,so 1 lowpriority thread taking care of IO and others are x numbers of highpriority threads,x number of cores?would that be good solution?
SSE unrolled version between 0 and 15ms for 65k numbers,for 8800000 numbers,156ms

now as exercise making avx calculator,maybe SIMD ascii2double and double2ascii would be faster ?,well at least its an exercise
my none asm creations
https://masm32.com/board/index.php?topic=6937.msg74303#msg74303
I am an Invoker
"An Invoker is a mage who specializes in the manipulation of raw and elemental energies."
Like SIMD coding