Communications - October 2011

video frame into 16 × 16 macro-blocks and encodes each one
separately. Each block goes through five major functions:
IME finds the closest match for an image block versus a
previous reference image. While it is one of the most com-
pute intensive parts of the encoder, the basic algorithm
lends itself well to data-parallel architectures. On our base
CMP, IME takes up 56% of the total encoder execution time
and 52% of total energy.

2. 4. current h.264 implementations

The computationally intensive H.264 encoding algorithm poses a challenge for general-purpose processors, and is typically implemented as an ASIC. For example, T. C. Chen et al. implement a full-system H.264 encoder and demonstrate that real-time HD H.264 encoding is possible in hardware using relatively low power and area cost. 2

H.264 software optimizations exist, particularly for
motion estimation, which takes most of the encoding time.
For example, sparse search techniques speed performance
of IME and FME by up to 10×. 14, 25 Combining aggressive
algorithmic modifications with multiple cores and SSE
extensions leads to highly optimized H.264 encoders on
Intel processors. 3, 12

3. eXPeRimentaL methoDoLoGY

To understand what is needed to gain ASIC level efficiency, we use existing H.264 partitioning techniques, and modify the H.264 encoder reference code JM 8.611 to remove dependencies and allow mapping of the five major algorithmic blocks to the four-stage macro-block (MB) pipeline shown in Figure 1. This mapping exploits task level parallelism at the macro-block level and significantly reduces the inter-processor communication bandwidth requirements by sharing data between pipeline stages.

table 1. intel’s optimized h.264 encoder versus a 720p hD asic.

fPs area (mm2) energy/frame (mJ)

Intel (720 × 480 sD) 30 122 742 Intel (1280 × 720 hD) 11 122 2023 AsIC 30 8 4

The second row gives Intel’s sD data scaled to hD. AsIC data is scaled from 180 down to 90nm.

Read/Write to
main memory
figure 1. four stage macroblock partition of h.264. (a) Data
flow between stages. (b) how the pipeline works on different
macroblocks. iP includes Dct + Quant. ec is caBac.