Avalon DMA problems

C

cruzin

Guest
Hi,

I have a very simple Stratix SOPC system: An Avalon DMA block whose
read and write masters both connect to a MegaRAM. The DMA control bus
is mastered by Nios.

My DMA blocks registers are set up as:

np_dmareadaddress = 0x10000; // base of MegaRAM
np_dmawriteaddress = 0x11000; // 1024 words in
np_dmalength = 32; // xfer 32 bytes
np_dmacontrol = 0x8C; // do 32-bit xfer, end when length=0 and GO!

I verified the registers by reading them back, and the status reg
indicates operation just as I would expect.

When I simulate this system in Active-HDL everything works as
expected. On actual hardware, unfortunately, I must always run the DMA
operation twice to get the right data into MegaRAM - the first time
always writes data to the destination address that is from the
PREVIOUS time that I ran the DMA operation. It always write 4 words of
old data; after these 4 words, the written data is current and
correct.

The only way I can rationalize this is that the FIFO inside the DMA
block is writing data to the destination before the source data has
been read into it.

My questions are:
- Can I use the Avalon DMA block to DMA to/from MegaRAMs? Or do the
source and/or destination resources have to be streaming slaves? (ie.
can a streaming master such as DMA work with non-streaming slaves such
as MegaRAMs?
- Why does simulation give me the correct results? Does it have to do
with me running on Stratix, whereas Active-HDL may assume I'm running
on Apex, etc?
- Is there a bug with the DMA block, or am I doing something stupid?
 
FYI,

The Altera DMA block has bugs if "Use ESB/EABs" is selected in its GUI
on Stratix. I don't know about other architectures. The default option
works properly in this scenario. I'm guessing ESBs have different
timing than TriMatrix and therefore FIFO draining does not work
properly.

-- Pete

Hi,

I have a very simple Stratix SOPC system: An Avalon DMA block whose
read and write masters both connect to a MegaRAM. The DMA control bus
is mastered by Nios.

My DMA blocks registers are set up as:

np_dmareadaddress = 0x10000; // base of MegaRAM
np_dmawriteaddress = 0x11000; // 1024 words in
np_dmalength = 32; // xfer 32 bytes
np_dmacontrol = 0x8C; // do 32-bit xfer, end when length=0 and GO!

I verified the registers by reading them back, and the status reg
indicates operation just as I would expect.

When I simulate this system in Active-HDL everything works as
expected. On actual hardware, unfortunately, I must always run the DMA
operation twice to get the right data into MegaRAM - the first time
always writes data to the destination address that is from the
PREVIOUS time that I ran the DMA operation. It always write 4 words of
old data; after these 4 words, the written data is current and
correct.

The only way I can rationalize this is that the FIFO inside the DMA
block is writing data to the destination before the source data has
been read into it.

My questions are:
- Can I use the Avalon DMA block to DMA to/from MegaRAMs? Or do the
source and/or destination resources have to be streaming slaves? (ie.
can a streaming master such as DMA work with non-streaming slaves such
as MegaRAMs?
- Why does simulation give me the correct results? Does it have to do
with me running on Stratix, whereas Active-HDL may assume I'm running
on Apex, etc?
- Is there a bug with the DMA block, or am I doing something stupid?
Click to expand...
 
You must log in or register to reply here.

Welcome to EDABoard.com

Sponsor

Back
Top