[julia-users] why so much faster with a global than a local vec?

[Scott Jones]

In the local case, it is having to allocate, populate, and then at some point 
garbage collect that array.
That is why using a const is much faster.

Re: [julia-users] why so much faster with a global than a local vec?

[Stefan Karpinski]

The first one creates a new array object every time it runs. It's a fairly
sophisticated analysis to prove that the array never escapes and is never
modified and can therefore be reused.

On Fri, Jul 3, 2015 at 2:21 PM, Jeffrey Sarnoff jeffrey.sarn...@gmail.com
wrote:

 I wrote a range limited isa_leapyear(year) good in 1800..2200.

 Your look is appreciated; I do not understand this:

 The first version runs 50x more slowly than the second version.
 The first is better practice, is there a way to make it behave?

 They differ only in placing the bit table (local vs const global).
 I ran both through code_typed(), the function code is the same.

 this is the way of the snail:

 function is_proximal_leapyear(year::Int)
   LeapYearBitTbl =
   [ 0xfdfe, 0x, 0xf7ff, 0x000f ]

   y400 = convert(Uint32, year-1800)
   y100 = y400  2
   mask = one(Uint32)  (y100  0x000f)
   indx = 1 +  (y100   5)
   ((LeapYearBitTbl[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
 end

 this way is lovely fast:

 const LeapYearBitTable =
   [ 0xfdfe, 0x, 0xf7ff, 0x000f ];

 function isa_proximal_leapyear(year::Int)
   y400 = convert(Uint32, year-1800)
   y100 = y400  2
   mask = one(Uint32)  (y100  0x000f)
   indx = 1 +  (y100   5)
   ((LeapYearBitTable[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
 end

[julia-users] why so much faster with a global than a local vec?

[Jeffrey Sarnoff]

I wrote a range limited isa_leapyear(year) good in 1800..2200.

Your look is appreciated; I do not understand this:

The first version runs 50x more slowly than the second version.
The first is better practice, is there a way to make it behave?

They differ only in placing the bit table (local vs const global).
I ran both through code_typed(), the function code is the same.

this is the way of the snail:

function is_proximal_leapyear(year::Int)
  LeapYearBitTbl =  
  [ 0xfdfe, 0x, 0xf7ff, 0x000f ]

  y400 = convert(Uint32, year-1800)
  y100 = y400  2
  mask = one(Uint32)  (y100  0x000f)
  indx = 1 +  (y100   5)
  ((LeapYearBitTbl[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
end

this way is lovely fast:

const LeapYearBitTable = 
  [ 0xfdfe, 0x, 0xf7ff, 0x000f ];

function isa_proximal_leapyear(year::Int)
  y400 = convert(Uint32, year-1800)
  y100 = y400  2
  mask = one(Uint32)  (y100  0x000f)
  indx = 1 +  (y100   5)
  ((LeapYearBitTable[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
end

Re: [julia-users] why so much faster with a global than a local vec?

[Yichao Yu]

On Fri, Jul 3, 2015 at 8:37 AM, Stefan Karpinski ste...@karpinski.org wrote:
 The first one creates a new array object every time it runs. It's a fairly
 sophisticated analysis to prove that the array never escapes and is never
 modified and can therefore be reused.

Also, in general, a global constant is faster than a local variable.


 On Fri, Jul 3, 2015 at 2:21 PM, Jeffrey Sarnoff jeffrey.sarn...@gmail.com
 wrote:

 I wrote a range limited isa_leapyear(year) good in 1800..2200.

 Your look is appreciated; I do not understand this:

 The first version runs 50x more slowly than the second version.
 The first is better practice, is there a way to make it behave?

 They differ only in placing the bit table (local vs const global).
 I ran both through code_typed(), the function code is the same.

 this is the way of the snail:

 function is_proximal_leapyear(year::Int)
   LeapYearBitTbl =
   [ 0xfdfe, 0x, 0xf7ff, 0x000f ]

   y400 = convert(Uint32, year-1800)
   y100 = y400  2
   mask = one(Uint32)  (y100  0x000f)
   indx = 1 +  (y100   5)
   ((LeapYearBitTbl[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
 end

 this way is lovely fast:

 const LeapYearBitTable =
   [ 0xfdfe, 0x, 0xf7ff, 0x000f ];

 function isa_proximal_leapyear(year::Int)
   y400 = convert(Uint32, year-1800)
   y100 = y400  2
   mask = one(Uint32)  (y100  0x000f)
   indx = 1 +  (y100   5)
   ((LeapYearBitTable[indx] $ mask) + (y400  0x0003)) === zero(Uint32)
 end