SpinalHDL/SpinalHDL

SpinalHDL core

korbin

@korbin

def streamReadSyncRegistered[T2 <: Data](cmd: Stream[UInt], linkedData: T2, crossClock:Boolean = false) : Stream[ReadRetLinked[T,T2]] = {
    val ret = Stream(new ReadRetLinked(mem.wordType, linkedData))

    val retData = mem.readSync(cmd.payload, cmd.ready, clockCrossing = crossClock)
    val retLinked = RegNextWhen(linkedData, cmd.ready)

    cmd.ready := ret.ready

    ret.valid := RegNext(RegNext(cmd.valid))
    ret.value := RegNext(retData)
    ret.linked := RegNext(retLinked)
    ret
  }

i cannot for the life of me figure out how to remove the bubble from this extension to Mem.scala.

BRAM requires an additional register stage to be absorbed for better timing in Xilinx devices, so RegNext(retData) should work, but then the handshake logic breaks

Dolu1990

@Dolu1990

Ahh you want to have it with the bram output register enabled ?

korbin

@korbin

yes

exactly

Dolu1990

@Dolu1990

Do you have a clock enable on that output register in the FPGA ?

korbin

@korbin

i am not sure if they allow clock enables or not - i believe so

it may be different for block ram / ultraram

Dolu1990

@Dolu1990

If yes, i would say, just use the regular streamReadSync, and add a returnStream.stage()

korbin

@korbin

i was doing that, they don't get absorbed

Dolu1990

@Dolu1990

damned

korbin

@korbin

in that case, i do not think clock enables are allowed

Dolu1990

@Dolu1990

Sadness of my life

hmm

what's about using a low latency fifo on the output ? with "almostfull" style logic to halt the cmd stream ?

korbin

@korbin

there are a number of places where there is register absorbing in xilinx FPGAs like this - BRAM, URAM, DSP, etc

the verilog needs to look like:

always@(posedge clk) begin
       ram_out <= ram[addra];
       dout <= ram_out;
end

it feels like there should be a way to skid buffer this somehow but i am just not familiar enough yet with Stream

Dolu1990

@Dolu1990

val haltIt = Bool()
val cmdHalted = cmd.haltWhen(haltIt)
val rsp = mem.streamReadSync(cmdHalted).queueLowLatency(size=2)
haltIt := rsp.isStall()

Something in that kind ?

(but not exactly)

More like :

val haltIt = Bool()
val cmdHalted = cmd.haltWhen(haltIt).toFlow
val rsp = mem.flowReadSync(cmdHalted).toStream.queueLowLatency(size=2)
haltIt := rsp.isStall()

(flowReadSync isn't a thing in the lib)

korbin

@korbin

yeah this makes sense, thanks

Dolu1990

@Dolu1990

Ahhh missing one thing

mem.flowReadSync(cmdHalted).stage.toStream.

sebastien-riou

@sebastien-riou

question to scala gurus:
we have elegant way to concatenate "reads": myBits_24bits := bits_8bits_1 ## bits_8bits_2 ## bits_8bits_3
do we have an elegant way to concatenate "writes" ?
I tried a ## b ## c = d but it does not seems to work, at least when a,b,c are registers

Dolu1990

@Dolu1990

(a, b, c) := d should work

@sebastien-riou

as long as d is Bits

saahm

@saahm

will that translate to the verilog equivalent? like {a,b,c} = sameVal[idx] where sameVal[idx] is same width as a, b and c together?

Dolu1990

@Dolu1990

yes, but with one assigment per destination

saahm

@saahm

I dont know if thats 'wanted' behavior, but components with the same name in different packages cause the same simWorkspace subdirectory to be used, is that wanted behavior?

Dolu1990

@Dolu1990

in SpinalSim or SpinalFormal ?

saahm

@saahm

SpinalSim

I only had it happen once and I didnt try reproducing it yet tho

Dolu1990

@Dolu1990

It isn't wanted behaviour

From a single run ?

or between multiple run ?

saahm

@saahm

So I have like

main/
|--foo
|    |--fooComp.scala
|    |--fooSim.scala
|
|--bar
     |--fooComp.scala
     |--fooSim.scala

in the the comp and sims are in different packages (also with the package statement at the head of the file). when I run the bar/fooSim it creates simWorkspace/fooComp, after that I run foo/fooSim and it overwrites the simWorkspace/fooComp from what I saw

Dolu1990

@Dolu1990

Ahh

yes it is expected

It would avoid the override if you were running them both in the same runtime

saahm

@saahm

like locking the directory?

Dolu1990

@Dolu1990

Yes

saahm

@saahm

ah ok

Dolu1990

@Dolu1990

and second one to come will be served with xxxx_1 as workspace

saahm

@saahm

when they are executed at the same time

Dolu1990

@Dolu1990

yes

saahm

@saahm

quite an "undocumented" feature xD

Where communities thrive

People

Repo info

Activity