Accelerator Design Example

1
Accelerator Design Example
2
Video Accelerator

• Objective
• Reduce the number of pixels that must be
  transmitted/stored

3
Problem Statement

• Design a motion estimation engine (accelerator)
  for a PC to enable experimenting with video
  processing techniques

4
Algorithm

• A variety exist Wolf chose
• block motion estimation
• Basis
• describe current frame in terms of differences
  between it and a previous one

5
Current frame divided into MACROBLOCKS
typically 16x16
6

• For every MACROBLOCK in the current frame,
  objective is to find the region in the previous
  frame that MOST CLOSELY matches the MACROBLOCK
• Limit search to a SEARCH AREA
• Measure sum of absolute differences in
  intensity over the MACROBLOCK
• MOTION VECTOR location of minimum

7
Measure S absM(i,j) S(i-ox,j-oy)
M(i,j) intensity of MACROBLOCVK at i,j S(i,j)
intensity of search region at i,j ox , oy
offset from MACROBLOCK center to search area
8
Algorithm in C
9

• If MBSIZE 16 SEARCHSIZE 8
• nops (16x16)x(17x17) 73,984
• Must be performed on every MACROBLOCK of every
  frame
• For one video format (CIF)
• frame size 352x288
• -gt 22x18 macroblocks

10
Requirements Table
11
Preliminary design decisions

• Purpose is experimentation ? one of a kind
• Hence -- use FPGA vs. ASIC
• lower performance a tradeoff

12
Specifications C program is essentially the spec
13
Specs - continued
14
Sequence diagram for compute-mv() operation of
class Motion-estimator
Correction - reverse second two arrows
15
Architecture

• Large memory
• MACROBLOCK 16x16 256
• SEARCH AREA (88188)2 1089 pixels
• mistake in text should be (881)2 - 289
• ? Text conclusion -memory external to the FPGA
  too large for memory on FPGA no true today
  particularly with correction

16
Possible Architecture
Two memories Sixteen PEs perform abs calculation
on a pair of pixels Comparator chooses smallest
17
In steady state, the schedule fetches one pixel
from the macroblock and two pixels from the
search area per clock cycle Computes 16
correlations between macroblock and search area
simultaneously
18
Architecture refinement
19
System Testing

• Can use images not video for testing because
  the design is only a motion estimator, not a
  video compressor