• State Resolved
  • Locked Locked
  • Replies 2 replies
  • Subscribers 54 subscribers
  • Views 854 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

Loop optimization

Kunal Lal
Prodigy 10 points

While compiling the below code with O3 compilation option with compiler version 7.3.5, I still see 6 NOPs in the assembly code for the loop.
void function(long long unsigned  * restrict x1,
                   long long unsigned  * restrict x2,
                   long long unsigned  * restrict multArr,
                   long long unsigned  * restrict y,
                   int nofLoops,
                   char shrVal) {
    #pragma MUST_ITERATE(6, 600, 6)
    for (int i=0; i < nofLoops; i++) {
        long long unsigned int x1_val = _amem8(x1++);
        long long unsigned int x2_val = _amem8(x2++);
        long long unsigned int mult1 = _amem8(multArr++);
        long long unsigned int mult2 = _amem8(multArr++);
        long long unsigned int prod1 = _lo128(_cmatmpy(mult1, _llto128(x2_val, x1_val)));
        long long unsigned int prod2 = _hi128(_cmatmpy(mult2, _llto128(x2_val, x1_val)));
        unsigned int res1 = _sshvr(_loll(prod1), shrVal);
        unsigned int res2 = _sshvr(_hill(prod1), shrVal);
        unsigned int res3 = _sshvr(_loll(prod2), shrVal);
        unsigned int res4 = _sshvr(_hill(prod2), shrVal);
        _amem8(y++) = _itoll(_spack2(res4, res3), _spack2(res2, res1));
    }
}
Loop unrolling does not improve cycle count. Can anyone help me on how to improve the cycle count for this loop? Below is the assembly code:
    1423                    ;*----------------------------------------------------------------------------*
    1424                    ;*   SOFTWARE PIPELINE INFORMATION
    1430                    ;*      Loop Unroll Multiple             : 2x
    1431                    ;*      Known Minimum Trip Count         : 3                    
    1432                    ;*      Known Maximum Trip Count         : 300                    
    1433                    ;*      Known Max Trip Count Factor      : 3
    1434                    ;*      Loop Carried Dependency Bound(^) : 0
    1435                    ;*      Unpartitioned Resource Bound     : 6
    1436                    ;*      Partitioned Resource Bound(*)    : 6
    1437                    ;*      Resource Partition:
    1438                    ;*                                A-side   B-side
    1439                    ;*      .L units                     0        0     
    1440                    ;*      .S units                     2        2     
    1441                    ;*      .D units                     5        5     
    1442                    ;*      .M units                     6*       6*    
    1443                    ;*      .X cross paths               2        2     
    1444                    ;*      .T address paths             5        5     
    1445                    ;*      Long read paths              0        0     
    1446                    ;*      Long write paths             0        0     
    1447                    ;*      Logical  ops (.LS)           6        6     (.L or .S unit)
    1448                    ;*      Addition ops (.LSD)          0        0     (.L or .S or .D unit)
    1449                    ;*      Bound(.L .S .LS)             4        4     
    1450                    ;*      Bound(.L .S .D .LS .LSD)     5        5     
    1451                    ;*
    1452                    ;*      Searching for software pipeline schedule at ...
    1453                    ;*         ii = 6  Did not find schedule
    1454                    ;*         ii = 7  Schedule found with 4 iterations in parallel
    1455                    ;*      Done
    1456                    ;*
    1457                    ;*      Loop will be splooped
    1458                    ;*      Collapsed epilog stages       : 0
    1459                    ;*      Collapsed prolog stages       : 0
    1460                    ;*      Minimum required memory pad   : 0 bytes
    1461                    ;*
    1462                    ;*      For further improvement on this loop, try option -mh32
    1463                    ;*
    1464                    ;*      Minimum safe trip count       : 1 (after unrolling)
    1465                    ;*----------------------------------------------------------------------------*
    1466 00000120           $C$L12:    ; PIPED LOOP PROLOG
    1468                    
    1469 00000120 0303a001             SPLOOPD 7       ;28               ; (P)
    1470 00000124 0d110059  ||         ADD     .L1     8,A4,A26
    1471 00000128 06e013a2  ||         MVC     .S2X    A24,ILC
    1472                    
    1473                    ;** --------------------------------------------------------------------------*
    1474 0000012c           $C$L13:    ; PIPED LOOP KERNEL
    1478 0000012c     2ce7             SPMASK          L1,L2
    1479 00000130 0d911059  ||         ADD     .L1X    8,B4,A27
    1480 0000012e     4e47  ||         MV      .L2     B4,B26
    1481 00000134 0b685765  ||         LDDW    .D1T1   *A26++(16),A23:A22 ; |213| (P) <0,0>
    1482 00000138 006457e6  ||         LDDW    .D2T2   *B25++(16),B1:B0  ; |213| (P) <0,0>
    1483                    
    1484 00000140 096857e7             LDDW    .D2T2   *B26++(16),B19:B18 ; |215| (P) <0,1>
    1485 00000144 0e6c5764  ||         LDDW    .D1T1   *A27++(16),A29:A28 ; |215| (P) <0,1>
    1487 00000148     2ce6             SPMASK          L2
    1488 0000014a     7dc7  ||         MV      .L2X    A3,B27
    1489                    
    1490 0000014c     2fe7             SPMASK          L1,S1,L2,S2
    1491 0000014e     9dc7  ||         MV      .L2X    A3,B28
    1492 00000150     cdc6  ||         MV      .L1     A3,A30
    1493 00000152     0dae  ||         ADD     .S1     8,A3,A3
    1494 00000154 0d800c52  ||         ADDK    .S2     24,B27
    1495                    
    1496 00000158 000f0001             SPMASK          L1,L2
    1497 00000160 0ca01fd9  ||         MV      .L1X    B8,A25
    1498 00000164 0c200fdb  ||         MV      .L2     B8,B24
    1499 00000168 020c9765  ||         LDDW    .D1T1   *A3++(32),A5:A4   ; |231| (P) <0,4>
    1500 0000016c 046c97e6  ||         LDDW    .D2T2   *B27++(32),B9:B8  ; |231| (P) <0,4>
    1501                    
    1502 00000170 00230001             SPMASK          S2
    1503 00000174 0e000852  ||         ADDK    .S2     16,B28
    1504                    
    1505 00000178 0a789765             LDDW    .D1T1   *A30++(32),A21:A20 ; |231| (P) <0,6>
    1506 0000017c 087097e7  ||         LDDW    .D2T2   *B28++(32),B17:B16 ; |231| (P) <0,6>
    1507 00000180 09002d9b  ||         DMV     .L2     B1,B0,B19:B18     ; |213| (P) <0,6>
    1508 00000184 025aed99  ||         DMV     .L1     A23,A22,A5:A4     ; |213| (P) <0,6>
    1509 00000188 044a6ef3  ||         DMV     .S2     B19,B18,B9:B8     ; |215| (P) <0,6>
    1510 0000018c 0473aef0  ||         DMV     .S1     A29,A28,A9:A8     ; |215| (P) <0,6>
    1511                    
    1512 00000190     2f66             SPMASK          S1,S2
    1513 00000194 0c1911a1  ||         ADD     .S1X    8,B6,A24
    1514 00000192     af4f  ||         MV      .S2     B6,B29
    1515 00000198 03212d9b  ||         DMV     .L2     B9,B8,B7:B6       ; |223| (P) <0,7>
    1516 000001a0 05212d98  ||         DMV     .L1     A9,A8,A11:A10     ; |223| (P) <0,7>
    1517                    
    1518 000001a4 0410ad99             DMV     .L1     A5,A4,A9:A8       ; |223| (P) <0,8>
    1519 000001a8 024a6d9a  ||         DMV     .L2     B19,B18,B5:B4     ; |223| (P) <0,8>
    1520                    
    1521 000001ac 0318045b             DADD    .L2     0,B7:B6,B7:B6     ; |231| (P) <0,9>
    1522 000001b0 02100863  ||         DADD    .S2     0,B5:B4,B5:B4     ; |231| (P) <0,9>
    1523 000001b4 09280459  ||         DADD    .L1     0,A11:A10,A19:A18 ; |231| (P) <0,9>
    1524 000001b8 08200861  ||         DADD    .S1     0,A9:A8,A17:A16   ; |231| (P) <0,9>
    1525 000001bc 12211201  ||         CMATMPY .M1X    B9:B8,A11:A10:A9:A8,A7:A6:A5:A4 ; |231| (P) <0,9>
    1526 000001c0 1a109202  ||         CMATMPY .M2X    A5:A4,B7:B6:B5:B4,B23:B22:B21:B20 ; |231| (P) <0,9>
    1527                    
    1528 000001c4 00006000             NOP             4
    1529                    
    1530 000001c8 14421201             CMATMPY .M1X    B17:B16,A19:A18:A17:A16,A11:A10:A9:A8 ; |231| (P) <0,14>
    1531 000001cc 14509202  ||         CMATMPY .M2X    A21:A20,B7:B6:B5:B4,B11:B10:B9:B8 ; |231| (P) <0,14>
    1532                    
    1533 000001d0 0fdf06b3             SSHVR   .M2     B23,B24,B31       ; |231| (P) <0,15>
    1534 000001d4 0f9f26b0  ||         SSHVR   .M1     A7,A25,A31        ; |231| (P) <0,15>
    1535                    
    1536 000001d8 00000000             NOP             1
    1537                    
    1538 000001dc 001b26b1             SSHVR   .M1     A6,A25,A0         ; |231| (P) <0,17>
    1539 000001e0 0f5b06b2  ||         SSHVR   .M2     B22,B24,B30       ; |231| (P) <0,17>
    1541 000001e4 00000000             NOP             1
    1542                    
    1543 000001e8 0b2726b1             SSHVR   .M1     A9,A25,A22        ; |231| (P) <0,19>
    1544 000001ec 0a2306b2  ||         SSHVR   .M2     B8,B24,B20        ; |231| (P) <0,19>
    1545                    
    1546 000001f0 042326b1             SSHVR   .M1     A8,A25,A8         ; |231| (P) <0,20>
    1547 000001f4 022706b2  ||         SSHVR   .M2     B9,B24,B4         ; |231| (P) <0,20>
    1548                    
    1549 000001f8 08fbecb3             SPACK2  .S2     B31,B30,B17       ; |231| <0,21>
    1550 000001fc 0883ecb0  ||         SPACK2  .S1     A31,A0,A17        ; |231| <0,21>
    1551                    
    1552 00000200 0822ccb1             SPACK2  .S1     A22,A8,A16        ; |231| <0,22>
    1553 00000204 08508cb2  ||         SPACK2  .S2     B4,B20,B16        ; |231| <0,22>
    1556                    
    1557 00000208 04c34001             SPKERNEL 2,3
    1558 0000020c 087457c7  ||         STDW    .D2T2   B17:B16,*B29++(16) ; |231| <0,23>
    1559 00000210 08605744  ||         STDW    .D1T1   A17:A16,*A24++(16) ; |231| <0,23>
    1560                    
    1562                    ;** --------------------------------------------------------------------------*
    1563 00000214           $C$L14:    ; PIPED LOOP EPILOG
    1564                    ;** --------------------------------------------------------------------------*

  • 0
    TI__Guru* 81335 points

    Hi Kunal,

    Moved this post to TI Compiler forum for faster response.

    Thank you.

  • 0
    TI__Guru* 84285 points

    The NOPS are probably not significant.  They are there to handle the delay slots of some of the instructions on the M units, not because nothing could be scheduled there.  Your iteration interval is bound from below at 6 by the fact you have 12 M unit instructions; there's not a lot of room left for scheduling, so it's entirely possible this is the best possible result.  Unrolling is not likely to improve the loop.  Perhaps it can be improved, but I don't see any obvious avenues.