Page 1 of 2 12 LastLast
Results 1 to 10 of 16
Like Tree8Likes

Thread: Memory bandwidth

777
  1. [translate]    #1
    Moderator
    Join Date
    Jun 2010
    Location
    Warsaw, Poland
    Posts
    2,828

    Default Memory bandwidth

    I was having problems with poor texturing and pixel generation performance and this led me to a discovery showing that we've got some memory bandwidth problem.

    I have found a memory benchmark for ARM somewhere and here is the result: (tested with CPU forced to 800 MHz)
    Code:
    # ./bench -c
    *** Memory Write Throughput (in MB/s) ***
    method           uncached  write alloc       cached  cached + wa
    STRB                20.13       111.72        21.09       703.36
    STR                 80.51       116.05        96.58      2360.09
    STM4               296.82       151.88       362.11      4626.08
    STM8               472.76       151.83       483.36      4627.50
    STRD               161.36       137.92       166.89      3869.30
    
    *** Uncached Memory Read/Copy Throughput (in MB/s) ***
    method    PLD=0   PLD=1   PLD=2   PLD=4   PLD=8  PLD=16  PLD=32  PLD=64 PLD=128
    LDRB     162.24  144.11  258.62  274.32  269.98  263.23  260.21  259.76  258.64
    LDR      193.90  193.66  310.99  351.18  375.03  367.31  360.99  358.13  355.26
    LDM      204.18  193.68  306.79  353.04  384.39  381.31  392.85  388.84  385.59
    LDRD     204.21  203.42  305.01  351.89  383.71  391.90  384.32  376.56  373.59
    CPY_B     19.77   19.60   19.82   20.84   20.51   20.44   20.34   20.32   20.18
    CPY_Bwa   72.03   70.54   75.73   64.99   62.53   63.87   68.30   67.75   65.36
    CPY_R     76.16   75.08   75.27   75.62   70.64   68.59   67.59   64.94   66.95
    CPY_Rwa   82.43   78.56   88.69   85.83   82.24   87.70   87.13   87.24   86.89
    CPY_M    172.33  165.57  218.40  273.56  249.96  234.81  226.51  212.76  209.11
    CPY_Mwa   87.58   82.60   88.01   89.85   88.74   89.15   88.99   89.15   88.97
    CPY_D    143.90  139.86  142.20  148.45  132.86  126.41  123.19  121.28  120.09
    CPY_Dwa   85.68   82.83  100.70  101.00  102.80  104.21  104.64  105.03  104.63
    
    *** Cached Memory Read/Copy Throughput (in MB/s) ***
    method    PLD=0   PLD=1   PLD=2   PLD=4   PLD=8  PLD=16  PLD=32  PLD=64 PLD=128
    LDRB     679.48  684.08  606.73  622.53  634.39  641.44  644.26  646.50  646.85
    LDR     2358.92 2147.46 1534.96 1727.22 1847.36 1909.69 1942.35 1960.75 1970.66
    LDM     3896.89 3881.15 2329.95 2860.00 3092.73 3215.32 3290.66 3324.51 3339.91
    LDRD    3883.67 3341.83 1738.20 1960.19 2425.67 2521.37 2573.97 2600.14 2610.03
    CPY_B    361.39  355.13  333.20  338.04  342.50  342.95  345.25  345.59  345.30
    CPY_Bwa  363.01  357.77  335.12  340.67  344.09  345.78  346.47  347.72  346.73
    CPY_R   1311.97 1243.69 1011.94 1053.65 1089.78 1110.73 1119.19 1123.18 1127.48
    CPY_Rwa 1314.46 1251.09 1013.02 1058.00 1093.04 1113.29 1124.37 1126.51 1132.56
    CPY_M   2363.95 1975.07 1521.41 1865.39 2015.65 2086.81 2117.08 2135.36 2143.50
    CPY_Mwa 2154.73 1827.76 1430.19 1760.06 1862.13 1915.94 1848.40 1962.59 1966.80
    CPY_D   2361.09 2148.24 1531.07 1727.52 1844.28 1907.03 1941.78 1959.55 1964.37
    CPY_Dwa 2358.54 2141.83 1531.46 1637.18 1725.21 1770.53 1794.15 1811.14 1818.27
    For comparison take a look at some results of a budget ARMv5 based PXA168:
    memspeed on Aspenite — ARM, OMAP, Xscale Linux Kernel

    The most alarming are cached writes and uncached reads, which look really slow.

    I have attached sources and binary of the benchmark I used:
    memspeed.tar.gz
    Last edited by tom3q; 10-10-2010 at 03:56 AM.

  2. [translate]    #2
    Senior Member
    Join Date
    Mar 2010
    Location
    The Netherlands
    Posts
    192

    Default

    Caching doesn't even have effect....
    So this is hardware i suppose, or is this fixable by software?

  3. [translate]    #3
    Moderator
    Join Date
    Jun 2010
    Location
    Warsaw, Poland
    Posts
    2,828

    Default

    Quote Originally Posted by Double View Post
    Caching doesn't even have effect....
    So this is hardware i suppose, or is this fixable by software?
    No idea. Maybe it's just a matter of some wrong register setting. I'm still trying to find the cause.

  4. [translate]    #4
    Senior Member
    Join Date
    Jun 2010
    Location
    Kiev
    Posts
    761

    Default

    tom, have you tried antibytes kernels with memcontweak?
    Eclair + SamDroid turbo by antibyte, QBanin kernel, system -> ext2, data -> ext4 with /data/data ext2 loop

  5. [translate]    #5
    Moderator
    Join Date
    Jun 2010
    Location
    Warsaw, Poland
    Posts
    2,828

    Default

    Quote Originally Posted by Yuna View Post
    tom, have you tried antibytes kernels with memcontweak?
    Not really the kernel, but I've tried setting the register manually. Nothing changes.

  6. [translate]    #6
    Moderator
    Join Date
    Jun 2010
    Location
    Warsaw, Poland
    Posts
    2,828

    Default

    It would be nice if someone run the benchmark on his phone and put the results here to check whether it's just an issue with my phone/firmware/kernel or something more global.

  7. [translate]    #7
    Member
    Join Date
    Apr 2010
    Location
    Munich
    Posts
    43

    Default

    I got almost the same results as you. The exact values aren't worth to post.

    EDIT: I've FE Beta 7 running at the moment.

  8. [translate]    #8
    Moderator
    Join Date
    Jun 2010
    Location
    Warsaw, Poland
    Posts
    2,828

    Default

    Quote Originally Posted by Lucky Luke View Post
    I got almost the same results as you. The exact values aren't worth to post.
    OK. Could you provide some details about firmware/kernel/etc. you use?

  9. [translate]    #9
    Member
    Join Date
    Apr 2010
    Location
    Munich
    Posts
    43

    Default

    As edited into my earlier post I use a Samdroid Kitchen v 1.3.1 cook & Antibyte's FE Beta 7.

  10. [translate]    #10
    Senior Member
    Join Date
    Jun 2010
    Location
    Kiev
    Posts
    761

    Default

    have the same values too - so we have a slow ram, or a slow controller for a fast ram?
    Eclair + SamDroid turbo by antibyte, QBanin kernel, system -> ext2, data -> ext4 with /data/data ext2 loop

Page 1 of 2 12 LastLast

Thread Information

Users Browsing this Thread

There are currently 1 users browsing this thread. (0 members and 1 guests)

Bookmarks

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •