Problem with HPI interface on C6713

We have a product currently in design.  We are using the HPI to transfer data directly into the DSP's memory from a host PC.  We've had a working HPI for several weeks until a couple days ago.  We cannot determine what went wrong to break things.  HPI reads (host reading data from DSP memory) still work flawlessly but HPI writes (host writing data to DSP memory) exhibits the behavior as shown in the below example.  This behavior is solid and repeatable.

For this example lets say we are trying to write data to DSP address 0x80000000 which in our case is SRAM but this behavior is observed at all addresses.   We'll also assume that this example starts with the first HPI access after power-up (thus the garbage at the beginning).  Finally, we are observing the actual value of the memory using Code Composer v4 and a JTAG Emulator.  The table below shows a sequence of HPI writes and the actual memory after each write.  These writes are done individually so there are many seconds in between each write.

It is obvious that the data is being buffered internally and this buffer is not correctly being flushed.  It seems that once the buffer is full (I would expect it to be 8 dwords according to documentation but it seems to be 6 so maybe once it is 75% full) each new write results in one dword is being flushed and written to memory.  Whatever it is internally (EDMA?) that is responsible for flushing this buffer has stopped working.

The board is still in heavy development so we have been tweaking both hardware and software.  We are not sure what we changed or broke to create this behavior.  Any clues or directional pointers would be much appreciative.

David Clark | Senior Software Engineer | C&H Technologies, Inc. | www.chtech.com | 512.733.2621