同步32位GPIF写入的意外开销

我使用GPIF接口在同步32位模式，在100 MHz，写缓冲器进入FX3。我每次爆裂1024字节。我正在使用CyfxSLIFtiFNC的固件的修改版本。
缓冲区填充后的开销有问题…
让我们使用4个缓冲区，每个字节2048个。我预计，当固件启动时，我可以以每秒近400兆字节的速度填充这些缓冲区。但是没有，有显著的开销。我可以填补一个缓冲区（2048字节）在100兆赫，没有问题。在缓冲区被填充后，线程的满标志变为低约700纳秒。这比我预期的要长得多。
手动或自动DMA之间的切换没有区别。
在满/空标志或部分标志之间切换没有区别。在专用标志和当前线程标志之间切换没有区别。
改变缓冲区大小或缓冲区数量没有区别。
如果我把PCLK频率降到50 MHz，全标志开销改变到大约1100纳秒。（记住，它是100纳秒时钟的700纳秒）。显然，PCLK频率对开销时间的影响最大。
这里发生了什么事？
谢谢您。



     以下为原文
  I am using the GPIF interface in synchronous 32-bit mode, at 100 MHz, writing buffers into the FX3.  I burst 1024 bytes at a time.  I am using a modified version of the cyfxslfifosync example firmware.
         

    Something is wrong with the overhead after a buffer is filled...     

    Let’s use 4 buffers, 2048 bytes each.  I expect, when the firmware starts, that I can fill these buffers at nearly 400 megabytes per second.  But no, there is significant overhead.  I can fill up a buffer (2048 bytes) at 100 MHz, no problems.  After a buffer is filled, the thread’s full-flag goes low for about 700 nanoseconds.  This is much longer than I would expect.     

    Switching between manual or automatic DMA makes no difference.     

    Switching between full/empty flags or partial flags makes no difference.  Switching between dedicated flags and current-thread flags makes no difference.     

    Changing buffer size or number of buffers makes no difference.     

    If I drop the PCLK frequency down to 50 MHz, the full-flag overhead changes to about 1100 ns.  (Remember, it was 700 nanoseconds for the 100 Mhz clock).  Clearly the PCLK frequency has most influence over the overhead time.     

    What is going on here?     

    Thank you.