SLIDE 5
- 17 -
- L. Bauer, CES, KIT, 2014
ASIP Implementation (cont’d)
src: Tensilica, Inc.: “Xtensa LC Product Brief”
Provides noticeably improved
performance after targe- ting the ma- jor hot spots
However,
performance still not suf- ficient to achieve real- time require- ments
- More hot spots need to be
accelerated
I_ME MC_L TQ_PL
- 18 -
- L. Bauer, CES, KIT, 2014
ASIP Implementation (cont’d)
src: Tensilica, Inc.: “Xtensa LC Product Brief”
Scalability
problem when rather many hot- pots exist
all relevant hot spots are covered
MC_L CABAC
S_ME
CAVLC FM MAC H245_C Dec_ MB V34 mod I_ME TQ_PL
- 19 -
- L. Bauer, CES, KIT, 2014
ASIPs perform well when
1. rather few hot spots need to be accelerated and 2. those hot spots are well known in advance
ASIPs are less efficient when targeting rather many
hot spots
- All accelerators are provided statically (i.e. they require area
and consume power) even though typically just a few of them are needed at a certain time
ASIPs are less efficient when targeting unknown hot
spots
- Even for a given application it is not necessarily clear, which
parts of it are ‘hot’ during execution as this may depend on input data (as demonstrated in the following)
Summary of ASIP Implementation
- 20 -
- L. Bauer, CES, KIT, 2014
If MB_Type = P_MB
MC
Loop Over MB
Encoding Engine
Loop Over MB
ME: SA(T)D RD
MB-Type Decision (I or P) Mode Decision (for I or P)
Loop Over MB
IPRED DCT / Q DCT / HT / Q IDCT / IQ IDCT / IHT / IQ CAVLC
then else
MB Encoding Loop
In-Loop De- Blocking Filter
Iterates on MacroBlocks (MBs, i.e. 16x16 pixels) 2 different MB-types
different computational paths with different computational requirements
- I-MB (spatial prediction)
- P-MB (temporal prediction)
Example Application: H.264 video Encoder
