Matrox Helios Series Installation And Hardware Reference page 90

90 Chapter 4: Matrox Helios hardware reference
MAC unit
To maximize efficiency on all image widths (so that all PEs are fully utilized), the
PA can split the PE linear array into shorter segments, each responsible for a
different horizontal strip (set of rows) of the source image. The segments can be
a power of two in length, from 32 to 256 pixels.
The main function of the MAC unit is to perform one (and only one at a time)
of the following operations:
• Convolution.
• Grayscale morphology.
• Binary morphology.
• 3×3 connectivity mapping (for fast binary thinning).
• Vector product (for color conversion).
Point-to-point operations don't use the MAC unit (except as a pass through when
they use all four sources); they only use the ALU.
The MAC unit is capable of performing a single 16-bit by 16-bit, two 8-bit by
16-bit, or four 8-bit by 8-bit multiplies with 40-bit accumulation per cycle for
convolution operations. In the latter case, this represents 42 GigaMACs/sec. The
40-bit accumulator guarantees no overflow situation for a 16 by 16 kernel with
16-bit coefficients and data. In addition, the PA architecture allows symmetrical
kernels to be processed up to four times faster. A kernel is symmetric if one half
of it (either horizontally or vertically) is a mirror image of the other half.
The MAC unit is also able to perform up to four minimum or maximum
operations per cycle for grayscale morphology operations.
Only one source can be fed to the MAC unit, and the supported data types are
1-bit for binary morphology and 3×3 connectivity mapping, and 8- and 16-bit
for the other operations. In single pixel mode, the MAC unit produces a single
result to be transferred to the ALU per output pixel. However, depending on the
data type, dual or quad pixel modes can be used, which produce, respectively, two
or four results to be transferred to the ALU at a time. If needed, the MAC unit