Raw pixel performance version 1.11 to 1.35 slowdown?

Below is a code sample that runs much slower on versions 1.35 and 1.34 compared to version 1.11.

Basically it uses ReadPixelFast and WritePixelFast on an image bufer and the canvas.

The numbers I get are
version 1.11 - 1900 millisecs
version 1.35 - 14500 millisecs

That's about 7x slower.

I thought perhaps it was the canvas operations, but commenting out the write to the canvas is still slower.
version 1.11 - 600 millisecs
version 1.35 - 900 millisecs

I suspect I'm doing something wrong but I don't know what.


; Simple Speed test of read and write pixels

Const EVENT_MOUSEDOWN = $201
Const EVENT_CLOSE = $803

Const IMAGE_WIDTH = 400
Const IMAGE_HEIGHT = 400

; Create a main window
window = CreateWindow ("Speed Test of raw pixel ops", 0, 0, ClientWidth(Desktop()), ClientHeight(Desktop()))

If window

; Create panel
panel = CreatePanel (0, 0, ClientWidth(window), ClientHeight(window), window, 1)
SetPanelColor panel, 64, 64, 64
SetGadgetLayout panel, 1, 1, 1, 1

; Create drawing canvas
canvas = CreateCanvas (0, 0, IMAGE_WIDTH, IMAGE_HEIGHT, panel)
SetGadgetLayout canvas,1,1,1,1

; Create image buffer
ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT)
SetBuffer ImageBuffer (ibuffer)

; Main event loop...
Repeat

; Wait for a window event...
e = WaitEvent ()

; Process the event...
If e = EVENT_MOUSEDOWN Then

LockBuffer CanvasBuffer(canvas)
LockBuffer ImageBuffer(ibuffer)

startTime = MilliSecs()

For count = 0 To 100
For y = 0 To IMAGE_HEIGHT
For x = 0 To IMAGE_WIDTH
rrgb = ReadPixelFast(x,y)
WritePixelFast x,y,rrgb

SetBuffer CanvasBuffer (canvas)
WritePixelFast x,y,rrgb
SetBuffer ImageBuffer (ibuffer)
Next
Next
Next

UnlockBuffer CanvasBuffer(canvas)
UnlockBuffer ImageBuffer(ibuffer)

endTime = MilliSecs()
Print endTime - startTime

FlipCanvas (canvas)
EndIf
If e = EVENT_CLOSE Then End
Forever

Else
Notify ("Failed to create window!", True)
End
EndIf

ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT)

Change the above line to either of the ones below, and re-run the test. I'll say no more.

Managed Style:
ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT,1,2)

SysRam Style:
ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT,1,4)

Enjoy! :)

I've given these suggestions a try. This first one resulted if very little change

Managed Style:
ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT,1,2)

around 14500 millisecs


This second suggestion slowed things down even more.
SysRam Style:
ibuffer = CreateImage (IMAGE_WIDTH, IMAGE_HEIGHT,1,4)

around 26000 millisecs


Any other suggestions? Has the management of canvases changed since 1.11?

Thanks
Ken

Personally I never use a canvas for reading/writing direct to. I only use SysRam images for that. It's far quicker. Try your example with the main write loop as this. (in both versions).

I would be interested to know the results.

; Process the event... 
If e = EVENT_MOUSEDOWN Then 

	SetBuffer ImageBuffer (ibuffer) 
	LockBuffer ImageBuffer(ibuffer) 
	
	startTime = MilliSecs() 
	
	For count = 0 To 100 
		For y = 0 To IMAGE_HEIGHT 
			For x = 0 To IMAGE_WIDTH 
				rrgb = ReadPixelFast(x,y) 
				WritePixelFast x,y,rrgb 
			Next 
		Next 
	Next 
	
	UnlockBuffer ImageBuffer(ibuffer) 
	SetBuffer CanvasBuffer(canvas)
	DrawBlock iBuffer,0,0
	
	endTime = MilliSecs() 
	Print endTime - startTime 
	
	FlipCanvas (canvas) 
EndIf 
If e = EVENT_CLOSE Then End 

Also, consider using LockedPixel()'s method. This again, is very fast.

HI,

The repeated use of SetBuffer inside the loop might be problem.

Use the 'buffer' options of ReadPixel/WritePixel to control which buffers are read/written and remove the SetBuffers.

Snarty,

Thanks for your feedback.

I've tried your version and the results are

version 1.11 - 510 millisecs
version 1.35 - 720 millisecs

Of course this version changes the test so that the canvas is not updated along the way. The reason I was writting the pixels directly to the canvas was that the information I may normally have to display is not very dense, and in earlier tests on 1.11 it was quicker to just update the pixels rather than the full dirty rect.

I also had previously tried the LockedPixel() method and I couldn't get this to perform as well as the readpixelfast(), of course my approach may not have been the best.


Mark,
Thanks for this suggestion, not doing the SetBuffer()'s in the loop shaved 4000 millisecs (or 25 percent).

Are canvases created using video ram?

Thanks
Ken