Edit: 2018-04-28: there was a nasty bug in assignment statements that's fixed now. I've bumped the version number to reflect this, so it's now fastspin 3.7.1. I do encourage readers to check https://github.com/totalspectrum/spin2cpp/releases
for newer releases, I probably won't keep updating this thread now that the .cog.spin feature is in an "offical" release.
Edit: 2018-04-25: updated to yet another fastspin.zip that is the 3.7.0 release candidate. Added a number of optimizations and improved the handling of comments in the generated code.
Here are some thoughts I have on a mechanism to automatically convert Spin objects to PASM, along with an implementation. I'd love to get feedback on this.
Automatically Converting Spin To Pasm
Spin is the standard programming language for the Propeller. It's generally implemented by means of a bytecode compiler on the PC, producing binaries that are interpreted by a bytecode interpreter in the Propeller ROM. This works well -- Spin programs are compact, and perform reasonably well. But sometimes you need more performance than an interpreter can give.
The fastspin compiler can convert Spin programs to PASM automatically. Originally fastspin only worked on whole programs, which was fine if you wanted to speed up a small program. But PASM code is a lot bigger then bytecode, so converting a whole program to PASM increases its size quite a lot, and may not be feasible for larger programs.
The upcoming release of fastspin (3.7.0) will allow individual Spin objects to be converted to PASM and then integrated with regular Spin programs (compiled with the Propeller Tool, openspin, or other Spin bytecode compilers).
I've attached a prerelease version of 3.7.0 to this message, use it for now. After 3.7.0 is released, you can download fastspin.zip from https://github.com/totalspectrum/spin2cpp/releases
. (Make sure you get version 3.7.0 or later.)
Unzip it to a folder on your hard drive. You'll get two files: fastspin.exe (the program) and fastspin.md (the documentation). For convenience you may want to add the folder containing fastspin.exe to your system PATH; or, you can just copy fastspin.exe into the folder you're working in.
Using Fastspin to Convert to PASM
This is very easy. For example, to convert a Spin object Fibo.spin to a PASM object that can run in another COG, called Fibo.cog.spin, just do:
fastspin -w Fibo.spin
in a command line (Windows or Linux or Mac). This produces a file Fibo.cog.spin. You now can use "Fibo.cog" in place of "Fibo" and get the speed benefits.
The only other thing you have to do is add a call to the __cognew
method to actually start the object up. (We have to use the special __cognew
method because the Spin cognew function only supports executing bytecode methods.)
A complete example
if (n < 2)
return fibo(n-1) + fibo(n-2)
and here is fibodemo.spin, which uses it:
'' fibodemo.spin: demonstrate COGOBJ with a Fibonacci calculator
'' runs both bytecode and PASM versions of the fibo function
_clkmode = xtal1 + pll16x
_clkfreq = 80_000_000
PUB hello | e1, e2, i, n, n2
ser.start(31, 30, 0, 115200)
repeat i from 1 to 10
e1 := CNT
n := bytecode.fibo(i)
e1 := CNT - e1
e2 := CNT
n2 := pasm.fibo(i)
e2 := CNT - e2
ser.str(string(") = "))
ser.str(string(" bytecode time "))
ser.str(string(" / pasm time "))
ser.str(string(" cycles", 13, 10))
if ( n <> n2)
ser.str(string(" ERROR", 13, 10))
Note that you can mix regular Spin objects and COG objects freely, and in fact can use the original Fibo.spin alongside the converted Fibo.cog.spin.
method was automatically added by fastspin
when it translated blinker.spin from Spin to PASM. It does all the housekeeping involved with getting communication going between the Spin code and the PASM code running on another processor. Similarly there's a __cogstop
method which will stop the COG. You can also use the regular Spin cogstop
function, but this may leave the code in a state which makes it hard to start again, so __cogstop
To compile and run this on the command line I do:
fastspin -w Fibo.spin
propeller-load fibodemo.binary -r -t
You can also use Fibo.cog.spin with any Spin IDE or other Spin tools. It should be legal for all Spin compilers.
The output looks like:
fibo(1) = 1 bytecode time 2272 / pasm time 11536 cycles
fibo(2) = 1 bytecode time 6368 / pasm time 11536 cycles
fibo(3) = 2 bytecode time 10464 / pasm time 12640 cycles
fibo(4) = 3 bytecode time 18656 / pasm time 12640 cycles
fibo(5) = 5 bytecode time 30944 / pasm time 13744 cycles
fibo(6) = 8 bytecode time 51424 / pasm time 15952 cycles
fibo(7) = 13 bytecode time 84192 / pasm time 18160 cycles
fibo(8) = 21 bytecode time 137440 / pasm time 22576 cycles
fibo(9) = 34 bytecode time 223456 / pasm time 29200 cycles
fibo(10) = 55 bytecode time 362720 / pasm time 41344 cycles
Some things to note about the performance:
(1) There's a fixed overhead of nearly 10000 cycles for managing the inter-COG communication and getting the answer back from the remote COG. So for small calculations the PASM isn't worth it.
(2) Once we start getting slightly more complicated calculations the PASM speed quickly becomes apparent. In this case we can see that the PASM approaches 10x faster than regular Spin bytecode.
Another classic test. Here's a simple pin blinking object:
pin = 15
pausetime = 40_000_000
DIRA[pin] := 1
OUTA[pin] ^= 1
Tweak pin to be something that's actually wired up to an LED, and pausetime to the number of cycles you want to pause.
To use this in a program do something like:
'' blink an LED in another COG
'' COGSPIN version
_clkmode = xtal1 + pll16x
_clkfreq = 80_000_000
PUB demo | id
fds.start(31, 30, 0, 115200)
id := blink.__cognew
fds.str(string("blink running in cog "))
Convert blinker.spin to the PASM blinker.cog.spin as usual...
fastspin -w blinker.spin
...and then you can compile and run blinkdemo.spin
in the Propeller tool of your choice. In my case it's the command line:
Synchronous and Asynchronous Operation
fastspin -w blinker.spin
openspin -q blinkdemo.spin
propeller-load blinkdemo.binary -r -t
If you've been watching carefully you've noticed that the fibo demo got results back from the PASM COG (i.e. the Spin COG waited for the PASM COG to finish) but the blink demo did not (the blinking ran alongside the Spin COG). The first case is "synchronous" operation, and the second is "asynchronous". You may wonder how fastspin knew to wait in one case and not in the other. The answer is simple: if a method returns a value, fastspin adds code to make the Spin bytecode wait for the PASM's result. If the method never returns a value (and never assigns to the result
variable) then there's no wait, and the Spin bytecode can continue on its way.
You can use asynchronous operation to allow the Spin code to do work even while a calculation is in progress. The trick is to start the computation in a method that returns nothing, and then to provide a "getter" method that actually returns the result. For example:
'' multiply all N elements in an array by a number
'' and return the resulting sum of them
'' start the array scaling operation
'' we want this to run asynchronously,
'' so do NOT return the result, just store
'' it in the variable "answer"
PUB scaleArray(arrptr, n, scale) | i, t, sum
sum := 0
repeat i from 0 to n-1
t := long[arrptr][i] * scale
long[arrptr][i] := t
sum += t
answer := sum
'' here's the getter
Now on the Spin COG side we can launch the computation with the scaleArray
method, then do some work and come back later to get the answer with getAnswer
. If instead of answer := sum
we had done return sum
or result := sum
would have caused the Spin COG to wait for the result.
A More Real World Example
'' Trivial Serial port written in Spin
'' this is for a very simple serial port
'' (transmit only, and only on the default pin for now)
txpin = 30
bitcycles := clkfreq / baudrate
PUB tx(c) | val, nextcnt
OUTA[txpin] := 1
DIRA[txpin] := 1
val := (c | 256) << 1
nextcnt := CNT
waitcnt(nextcnt += bitcycles)
OUTA[txpin] := val
val >>= 1
PUB str(s) | c
REPEAT WHILE ((c := byte[s++]) <> 0)
This simple serial port, in the original Spin, works correctly and is fine for output at low speeds. But if you convert it to PASM then it will transmit at 921600 baud, and perhaps higher -- I haven't actually tried to see how high it will go.
Granted, we already have FullDuplexSerial.spin and a host of other serial ports. But this TrivialSerial is really easy to understand and to modify, so if you need to change something to communicate with a non-standard device, well, this would be easier than trying to tweak somebody else's PASM code
The Generated Code
The .cog.spin produced by fastspin is somewhat readable, and we can look in it to see how the compiler does. For example, the code generated for the tx
method of the trival serial driver looks like:
movs _Tx_ret, #doreturn
mov _Tx__mask_0001, imm_1073741824_
or OUTA, imm_1073741824_
or DIRA, imm_1073741824_
or arg1, #256
shl arg1, #1
mov _Tx_Val, arg1
mov _Tx_Nextcnt, CNT
mov _Tx__idx__0002, #10
rdlong Tx_tmp002_, objptr
add _Tx_Nextcnt, Tx_tmp002_
mov arg1, _Tx_Nextcnt
waitcnt arg1, #0
shr _Tx_Val, #1 wc
muxc OUTA, _Tx__mask_0001
djnz _Tx__idx__0002, #L__0021
which isn't too bad. We could improve it by keeping the bitcycles counter in a local variable, which would avoid the read from memory in the inner loop. Or we could tweak the generated assembly by hand.
Behind the Scenes
Some other things you will note in the .cog.spin file:
- The compiler guesses at the size of stack needed for the object and adds a constant __STACK_SIZE at the top of the file to say how big to make it (in longs). This is a conservative guess, and if memory is tight you may be able to reduce this size. Conversely, if you have recursive functions you may find you need more space and can increase it here. If you add a definition for __STACK_SIZE to your original .spin file, fastspin will notice it and use that value instead of the one it calculates. (__STACK_SIZE is basically the same as the Spin constant _STACK, but is used for PASM code instead of bytecode.)
- There's also a define for __MBOX_SIZE, giving the size of the mailbox used to communicate with the COG. This one is exact, don't try to change it.
- The first two items in the mailbox are a global lock and a function code. The global lock is to make sure only one Spin COG at a time uses the PASM COG. This feature hasn't been tested very much yet. The function code indicates what function the PASM COG is running. If it is 0 it means the PASM COG is idle and is able to accept new commands.