Hi, i'm playing with parallel execution/multithreading and i've got below program:
proc spawn_collatz_p(cpu:int, ub:uint64):seq[(uint64,int)] =
var responses = newSeq[FlowVarBase](cpu)
parallel:
for i in 0..<cpu:
responses[i] = spawn list_collatz(uint64(i + 1), ub, cpu.uint64)
for i in 0..<cpu:
result = result & ^responses[i]
type mismatch: got <FlowVarBase>
but expected one of:
proc `^`[T](fv: FlowVar[T]): T
first type mismatch at position: 1
required type for fv: FlowVar[^.T]
but expression 'responses[i]' is of type: FlowVarBase
proc `^`[T](fv: FlowVar[ref T]): ref T
first type mismatch at position: 1
required type for fv: FlowVar[ref T]
but expression 'responses[i]' is of type: FlowVarBase
template `^`(x: int): BackwardsIndex
first type mismatch at position: 1
required type for x: int
but expression 'responses[i]' is of type: FlowVarBase
expression: ^responses[i]
For reference below code works, but yeah, not ideal if you want to dynamically set your number of threads ;-)
proc spawn_collatz_p2(cpu:int, ub:uint64):seq[(uint64,int)] =
parallel:
let c1 = spawn list_collatz(uint64(1), ub, cpu.uint64)
let c2 = spawn list_collatz(uint64(2), ub, cpu.uint64)
result = ^c1 & ^c2
if '^' wants a FlowVar, give it a FlowVar
proc spawn_collatz_p(cpu:int, ub:uint64):seq[(uint64,int)] =
var responses:seq[FlowVar[seq[(uint64,uint64)]]]
responses.setLen(cpu)
parallel:
for i in 0..<cpu:
responses[i] = spawn list_collatz(uint64(i + 1), ub, cpu.uint64)
for i in 0..<cpu:
result = result & ^responses[i]
Ofcourse!! Thank you, i mixed up parentheses with square brackets..
What i'm actually trying to do with this program is using al my cores/threads on my processor to minimize total time to finish. So i'm calculating all the lengths of collatz sequences (https://en.wikipedia.org/wiki/Collatz_conjecture) up to a certain upper bound. At the moment when i'm using 15 threads it takes about half an hour to calculate them upto 10^11 (i've got an 8 core/16 thread processor). Because you can calculate them seperatly per number an ideal candidate to test parallel execution. Would also be a nice challenge who could write the fastest program!
Well I did ;-)
So my Nim program, compiled with -d:danger --threads:on took 5.1 seconds to create a list with the 'running maximums' for the numbers up to 2*10^8 (see https://oeis.org/A006877)
What I also did is make a program compiled for my GPU (Nvidia RTX 2060) with futhark (https://futhark-lang.org) which is a Haskell like language which compiles down to C , CUDA or OpenCL. This could most probably be improved as well, I'm not an expert, but it took a staggering 1.1 seconds...!