• View
• Edit
• History
• Print

GR32Addons

This set of unit containt differents addons to GR32 more orless usefull, tested with GR32 version 1.8beta.

GR32_MathAddons.pas

This unit define some procedures and functions specially axed on Single float arithmetic. This include some trigonometric functions (such as Tan, ArcTan, ArcCos or Modulo functions) plus some basic mathematical functions (Power, Mean, Min, Max...) and some other function (such as Convex who can be used to tell if a polygon is convex or not (usefull for the TProjectiveTransformation)).

GR32_TransformsEx.pas and GR32_TransformsWarp.pas

Theses unit define some new transformations, most of then are taken from the Paul Bourke article : Image warping / distortion. See his web page for details, C implementation and screenshoot. Every distortions are based on a generic transformation TCustomWarpTransformation. this one transform the coordinate in an normal system, then X and Y are Single form -1 to 1, the final transformation can apply any function to the coordinates without state of the real Source and Dest rects.

GR32_TransformsEx.pas define a spacial transformation, TSphereTransformation, witch perform an sheprical projection of a planar image. The resulting sphere projection can be moved, scalled and rotated (with Longitude and Latitude coordinates system).

PerfsTransforms.zip is an modified version of the GR32 sample application TransformEx with this Transformations illustrated. (here is the Exe only archive).

GR32_RasterizerEx.pas

This unit define a really simple rasterizer class. It draw the bitmap in a "boxed" way, instead of define a color for each pixel of the image, a color value is caculated for a square of pixel (a property of the Rasterizer). The result is a weird image really pixelixed, but the resulting bitmap can be calculated more quickly than the whole image.

This can be usefull, if each pixel is the result of a complex transformation, to preview the final image. This rasterizer has been used in the above Transforms demo to accelerate the rendering during mouse dragging operations on the spherical projection.

GR32_Convolution.pas

This unit is a fused version of the Antti Lukats MMX implementation of kernel 3x3 convolution and the so called PixelPtr implementation of Patrick Quinn. It use 3 PixelPtr (one per line) and MMX operations when available, if there is not MMX or when the use of MMX operation is not profitable it use the PixelPtr implementation.

i.e. MMX does not propose a divide operator, so when the kernel has a factor (i.e. the matrix is not normalized) and when this factor is not a power of 2, it doas NOT use MMX operations at all, because switching from MMX to integer arithmetic, apply 4 div and switch back to MMX, is slower than doing the whole stuff with standart operations.

By the way, kernels normalmized or with a power of 2 normalisation factor represent most of the kernels... If you are using user's defined kernels you should (if you can) adjust the normalisation factor to be executed with an MMX version of the function. The following application is a benchmark between the diferents implementation.

Applied to an image 1600x1200, with a Pentium 4 :

• "Scanline" = 377ms
• "PixelPtr" = 205ms
• "MMX" = 90ms
• the newer "MMX/PixelPtr" = 32ms

You can test yourself the differents convolution with the Morpheus Convolution Benchmark (exe) ;-) (source files).

Some kernels have been to the list of Patrick (4 Emboss style matrixes). Moreover 5x5 and 7x7 convolutions have been implemented (the 7x7 version does not use MMX yet) plus two specially optimized convolutions : Gaussian Blur and Sharpen. Theses last one are 20% and 30% faster then the fastest generic 3x3 implementations.