USRP2 Block RAMs Requiring Empty Multipliers

GNU Radio
1

I’m trying to make some modifications to the FPGA of the USRP2 which
involves using more multipliers and I came across this error during
the Mapping phase of the compile:

ERROR:Place:665 - The design has 31 block-RAM components of which 13
block-RAM
components require the adjacent multiplier site to remain empty.
This is
because certain input pins of adjacent block-RAM and multiplier sites
share
routing ressources. In addition, the design has 28 multiplier
components.
Therefore, the design would require a total of 41 multiplier sites on
the
device. The current device has only 40 multiplier sites.

Looking through Xilinx’s support and Googling, I’ve found that the
512x36 BRAM and the multipliers share interconnects, so that you
cannot use both, you either need to use less multipliers, less block
RAMs, or reduce the input width of the BRAMs to 18 bits. I’ve tried
reducing the number of block RAMs used by changing the RAM_SIZE
parameter in u2_core to 16384 instead of 32768. This reduced the
number of block RAMs, but this did not change the number of block RAMs
that require an empty multiplier.

I’d like to know where the 512x36 BRAMs are that cause this problem
and if anyone has encountered this problem and found a workaround.
I’m currently using rev. 11641 from the 3.2 release branch and
building on Xilinx 9.2i

Thanks,
Kyle

2

Kyle Pearson wrote:

Looking through Xilinx’s support and Googling, I’ve found that the
512x36 BRAM and the multipliers share interconnects, so that you
cannot use both, you either need to use less multipliers, less block
RAMs, or reduce the input width of the BRAMs to 18 bits. I’ve tried
reducing the number of block RAMs used by changing the RAM_SIZE
parameter in u2_core to 16384 instead of 32768. This reduced the
number of block RAMs, but this did not change the number of block RAMs
that require an empty multiplier.

The block rams used for processor ram are only used at 8 bits wide, so
we can do byte writes.

I’d like to know where the 512x36 BRAMs are that cause this problem
and if anyone has encountered this problem and found a workaround.
I’m currently using rev. 11641 from the 3.2 release branch and
building on Xilinx 9.2i

The 36 bit wide ones are in the buffer pool and some of the fifos.
There isn’t much you can do about the buffer pool. If you aren’t using
the MIMO hardware, you can drop the serdes stuff out and that should get
rid of a couple of the FIFOs. There are a few other places you can
shave them. I would suggest you work from the latest git tree, however,
since the FIFOs have all changed, and this problem may be a little less
bad.

Matt