Gaussian blur - "fast" method: large performance improvements

Relates to #279

This new gaussian blur design takes advantage of the perf improvements in f010eec and extends them to the vertical component as well.  (Actually, we cheat and perform the vertical blurs by rotating the image 90 degrees, applying horizontal blurs, then rotating it back - this provides a larger perf boost, and it ensures that vertical edge behavior is 100% identical to horizontal edge behavior.)

Edge pixel handling is now identical to PD's more precise IIR-based technique, with edge pixels being "extended" instead of "clamped".  This produces closer results to Photoshop, for better or worse.

I've also further reduced memory requirements of the fast function (alongside some truly ugly object aliasing, lol), and performance of PD's fast gaussian is now back on par with Paint.NET.

Forms/Effects_Blur_GaussianBlur.frm

@@ -145,12 +145,8 @@ Public Sub GaussianBlurFilter(ByVal effectParams As String, Optional ByVal toPre
 
         '3 iteration box blur
         Case 0
-            Dim srcDIB As pdDIB
-            Set srcDIB = New pdDIB
-            srcDIB.CreateFromExistingDIB workingDIB
-            Filters_Layers.CreateApproximateGaussianBlurDIB gRadius, srcDIB, workingDIB, 3, toPreview
-            Set srcDIB = Nothing
-
+            Filters_Layers.CreateApproximateGaussianBlurDIB gRadius, workingDIB, workingDIB, 3, toPreview
+
         'IIR Gaussian estimation
         Case Else
             Filters_Area.GaussianBlur_IIRImplementation workingDIB, gRadius, 3, toPreview

Modules/Filters_Area.bas

@@ -624,6 +624,10 @@ Public Function GaussianBlur_IIRImplementation(ByRef srcDIB As pdDIB, ByVal radi
 
     Next y
 
+    'TODO: profile a potentially faster approach: rotate all arrays 90 degrees, horizontally blur *those*,
+    ' then rotate back.  On an integer-based box blur, this is much faster thanks to greatly improved
+    ' cache behavior.
+
     'Now repeat all the above steps, but filtering vertically along each column, instead
     If (Not g_cancelCurrentAction) Then
 

Modules/Filters_Layers.bas

@@ -1382,11 +1382,15 @@ Public Function CreateApproximateGaussianBlurDIB(ByVal equivalentGaussianRadius
     'To keep processing quick, only update the progress bar when absolutely necessary.  This function calculates that value
     ' based on the size of the area to be processed.
     Dim progBarCheck As Long
-    If (modifyProgBarMax = -1) Then modifyProgBarMax = srcDIB.GetDIBHeight * (numIterations * 2)
+    If (modifyProgBarMax = -1) Then modifyProgBarMax = srcDIB.GetDIBHeight * numIterations + srcDIB.GetDIBWidth * numIterations
     If (Not suppressMessages) Then SetProgBarMax modifyProgBarMax
-
     progBarCheck = ProgressBars.FindBestProgBarValue()
 
+    'Cache image dimensions as well
+    Dim origSrcWidth As Long, origSrcHeight As Long
+    origSrcWidth = srcDIB.GetDIBWidth
+    origSrcHeight = srcDIB.GetDIBHeight
+
     'Gaussian convolution can be (swiftly!) approximated using a piece-wise quadratic convolution kernel.
     ' Said another way, repeating a box blur 3x can produce an end result that's ~97% identical to a
     ' true Gaussian kernel (per the central-limit theorem).
@@ -1428,21 +1432,68 @@ Public Function CreateApproximateGaussianBlurDIB(ByVal equivalentGaussianRadius
     Dim gaussDIB As pdDIB
     Set gaussDIB = New pdDIB
     gaussDIB.CreateFromExistingDIB srcDIB
-    dstDIB.CreateFromExistingDIB gaussDIB
 
-    'Iterate a box blur, switching between the gauss and destination DIBs as we go
+    If (dstDIB Is Nothing) Then Set dstDIB = New pdDIB
+    dstDIB.CreateFromExistingDIB srcDIB
+
+    'First we're going to apply *all* horizontal blurs to the image.
     For i = 0 To numIterations - 1
-        If (CreateHorizontalBlurDIB(radiiTable(i), radiiTable(i), dstDIB, gaussDIB, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (gaussDIB.GetDIBHeight * i) + (gaussDIB.GetDIBHeight * i)) > 0) Then
-            If (CreateVerticalBlurDIB(radiiTable(i), radiiTable(i), gaussDIB, dstDIB, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (gaussDIB.GetDIBHeight * (i + 1)) + (gaussDIB.GetDIBHeight * i)) = 0) Then Exit For
+
+        'On even-numbered passes, draw from source to destination; on odd-numbered passes, draw from destination to source.
+        If ((i And 1) = 0) Then
+            If (CreateHorizontalBlurDIB(radiiTable(i), radiiTable(i), gaussDIB, dstDIB, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (origSrcHeight * i)) = 0) Then Exit For
         Else
-            Exit For
+            If (CreateHorizontalBlurDIB(radiiTable(i), radiiTable(i), dstDIB, gaussDIB, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (origSrcHeight * i)) = 0) Then Exit For
         End If
+
     Next i
 
-    'Erase the temporary DIB and exit
-    gaussDIB.EraseDIB
-    Set gaussDIB = Nothing
+    'We're now going to do something weird; we're going to rotate the source image 90 degrees, then blur *that*.
+    ' This seems crazy, but it's actually faster as it greatly improves CPU caching by placing blurred pixels
+    ' contiguously in-memory.
 
+    'Also, forgive the ugly aliasing nonsense - this is basically a cheap way to ensure that both even and odd
+    ' iteration counts are handled with a single copy operation.
+    Dim tmpSrc As pdDIB, tmpDst As pdDIB
+    If ((numIterations And 1) = 0) Then
+        Set tmpSrc = gaussDIB
+        Set tmpDst = dstDIB
+    Else
+        Set tmpSrc = dstDIB
+        Set tmpDst = gaussDIB
+    End If
+
+    tmpDst.CreateBlank origSrcHeight, origSrcWidth, 32, 0, 0
+    GDI_Plus.GDIPlusRotateFlipDIB tmpSrc, tmpDst, GP_RF_90FlipNone
+    tmpSrc.CreateFromExistingDIB tmpDst
+
+    'We're now going to apply *all* vertical blurs to the image
+    For i = 0 To numIterations - 1
+
+        'On even-numbered passes, draw from source to destination; on odd-numbered passes, draw from destination to source.
+        If ((i And 1) = 0) Then
+            If (CreateHorizontalBlurDIB(radiiTable(i), radiiTable(i), tmpSrc, tmpDst, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (origSrcHeight * numIterations) + (origSrcWidth * i)) = 0) Then Exit For
+        Else
+            If (CreateHorizontalBlurDIB(radiiTable(i), radiiTable(i), tmpDst, tmpSrc, suppressMessages, modifyProgBarMax, modifyProgBarOffset + (origSrcHeight * numIterations) + (origSrcWidth * i)) = 0) Then Exit For
+        End If
+
+    Next i
+
+    'Copy the source into the destination, using a similarly ridiculous aliasing approach
+    Dim finalSrc As pdDIB, finalDst As pdDIB
+    If ((numIterations And 1) = 0) Then
+        Set finalSrc = tmpSrc
+        Set finalDst = tmpDst
+    Else
+        Set finalSrc = tmpDst
+        Set finalDst = tmpSrc
+    End If
+
+    finalDst.CreateBlank origSrcWidth, origSrcHeight, 32, 0, 0
+    GDI_Plus.GDIPlusRotateFlipDIB finalSrc, finalDst, GP_RF_270FlipNone
+    If (finalDst.GetDIBDC <> dstDIB.GetDIBDC) Then dstDIB.CreateFromExistingDIB finalDst
+
+    'Based on global cancellation state, return success/failure
     If g_cancelCurrentAction Then CreateApproximateGaussianBlurDIB = 0 Else CreateApproximateGaussianBlurDIB = 1
 
 End Function
@@ -1946,11 +1997,7 @@ Public Function CreateHorizontalBlurDIB(ByVal lRadius As Long, ByVal rRadius As
     ' based on the size of the area to be processed.
     Dim progBarCheck As Long
     If (Not suppressMessages) Then
-        If modifyProgBarMax = -1 Then
-            SetProgBarMax finalY
-        Else
-            SetProgBarMax modifyProgBarMax
-        End If
+        If (modifyProgBarMax = -1) Then SetProgBarMax finalY Else SetProgBarMax modifyProgBarMax
         progBarCheck = ProgressBars.FindBestProgBarValue()
     End If
 

PhotoDemon.vbp

@@ -442,7 +442,7 @@ Description="PhotoDemon Photo Editor"
 CompatibleMode="0"
 MajorVer=7
 MinorVer=1
-RevisionVer=1541
+RevisionVer=1544
 AutoIncrementVer=1
 ServerSupportFiles=0
 VersionComments="Copyright 2000-2019 Tanner Helland -  photodemon.org"