Table of Contents
Advertisement
ALU
can also pass the original source pixel (the centre pixel of the neighborhood) to
the ALU along with the accumulated results. Therefore, a maximum of five values
can be transferred from the MAC unit to the ALU for each pixel.
The maximum number of kernel elements is 1024 if they are 16-bit, 2048 if they
are 8-bit, or 8192 for binary morphology (stored as 2 bits per element, where one
bit is a don't care mask). This represents, for example, a 32x32 16-bit kernel, a
45x45 8-bit kernel, or a 90x90 1-bit kernel. However, symmetrical 8- or 16-bit
kernels can be larger than this because not all kernel elements need to be used in
this case.
The PA can perform a convolution with up to four kernels at the same time, subject
to the following restrictions:
• All kernels must have the same size, centre, data type, and symmetry.
• All kernels are applied to the same source image.
The PA can perform grayscale morphology with up to two kernels (structuring
elements), but there are some differences from convolution with multiple kernels.
The rules for grayscale morphology with two kernels are:
• Both kernels must have the same size, centre, and symmetry.
• Both kernels are applied to the same source image.
• The first kernel always performs a MIN operation (erosion).
• The second kernel always performs a MAX operation (dilation).
Multiple kernels are not supported for binary morphology.
The ALU is 40 bits wide and can execute a wide variety of arithmetic and logical
operations. It can be programmed to execute a sequence of up to 256 instructions
per pixel at one instruction per clock cycle. This allows complex sequences of I/O
bound operations to be combined into a single processing pass, reducing the
amount of memory accesses and further improving performance.
Matrox Oasis
91
Advertisement
Table of Contents
