1.1 --- /dev/null	Thu Jan 01 00:00:00 1970 +0000
     1.2 +++ b/documents/pygar-slides-handout.lyx	Wed May 12 02:25:34 2010 -0400
     1.3 @@ -0,0 +1,217 @@
     1.4 +#LyX 1.6.4 created this file. For more info see http://www.lyx.org/
     1.5 +\lyxformat 345
     1.6 +\begin_document
     1.7 +\begin_header
     1.8 +\textclass article
     1.9 +\use_default_options true
    1.10 +\language english
    1.11 +\inputencoding auto
    1.12 +\font_roman default
    1.13 +\font_sans default
    1.14 +\font_typewriter default
    1.15 +\font_default_family default
    1.16 +\font_sc false
    1.17 +\font_osf false
    1.18 +\font_sf_scale 100
    1.19 +\font_tt_scale 100
    1.20 +
    1.21 +\graphics default
    1.22 +\paperfontsize default
    1.23 +\use_hyperref false
    1.24 +\papersize default
    1.25 +\use_geometry false
    1.26 +\use_amsmath 1
    1.27 +\use_esint 1
    1.28 +\cite_engine basic
    1.29 +\use_bibtopic false
    1.30 +\paperorientation portrait
    1.31 +\secnumdepth 3
    1.32 +\tocdepth 3
    1.33 +\paragraph_separation indent
    1.34 +\defskip medskip
    1.35 +\quotes_language english
    1.36 +\papercolumns 1
    1.37 +\papersides 1
    1.38 +\paperpagestyle default
    1.39 +\tracking_changes false
    1.40 +\output_changes false
    1.41 +\author "" 
    1.42 +\author "" 
    1.43 +\end_header
    1.44 +
    1.45 +\begin_body
    1.46 +
    1.47 +\begin_layout Title
    1.48 +Pygar: Parallel Audio Processing
    1.49 +\end_layout
    1.50 +
    1.51 +\begin_layout Author
    1.52 +Laurel Pardue, Robert McIntyre
    1.53 +\end_layout
    1.54 +
    1.55 +\begin_layout Subsection*
    1.56 +Problem
    1.57 +\end_layout
    1.58 +
    1.59 +\begin_layout Standard
    1.60 +Music naturally comes in parallel sequences of samples called 
    1.61 +\emph on
    1.62 +voices
    1.63 +\emph default
    1.64 + (ex.
    1.65 + from multiple instruments).
    1.66 + Pure-software mixers are forced to pass these voices through the Von Neuman
    1.67 + bottleneck of a single processor, operating on these streams in series
    1.68 + and switching between each one quickly.
    1.69 + They are therefore naturally limited in the number of voices they can handle.
    1.70 + Worse, since the processing of each voice has to share the same processor,
    1.71 + too many voices at once can fully max out the processor and crash the system.
    1.72 + On typical laptop hardware and a high end software tool like ProTools,
    1.73 + this number is around 5.
    1.74 + Embedded devices have an even tougher time at meeting any sort of reasonable
    1.75 + timing requirements.
    1.76 +\end_layout
    1.77 +
    1.78 +\begin_layout Standard
    1.79 +[screenie Just 6 voices are enough to bring this session of ProTools to
    1.80 + it's knees.]
    1.81 +\end_layout
    1.82 +
    1.83 +\begin_layout Standard
    1.84 +We want the power of writing transforms for voices in a high level language
    1.85 + combined with a framework that applies these transforms to the voices in
    1.86 + parallel.
    1.87 +\end_layout
    1.88 +
    1.89 +\begin_layout Subsection*
    1.90 +Vision --- Pygar
    1.91 +\end_layout
    1.92 +
    1.93 +\begin_layout Standard
    1.94 +Our system addresses the limitations of pure software mixers.
    1.95 + It is a grid of SMIPS processors capped by a mixer.
    1.96 + The voices flow through the processors in parallel and are combined at
    1.97 + the final mixer into a single stream.
    1.98 + Each processor can be loaded with any arbitrary C program.
    1.99 +\end_layout
   1.100 +
   1.101 +\begin_layout Standard
   1.102 +\begin_inset Float figure
   1.103 +placement H
   1.104 +wide false
   1.105 +sideways false
   1.106 +status collapsed
   1.107 +
   1.108 +\begin_layout Plain Layout
   1.109 +\begin_inset Graphics
   1.110 +	filename ../../../../Pygar/documents/000402.png
   1.111 +	width 5in
   1.112 +
   1.113 +\end_inset
   1.114 +
   1.115 +
   1.116 +\begin_inset Caption
   1.117 +
   1.118 +\begin_layout Plain Layout
   1.119 +The audio data (“samples”) start in the memory, but are soon pulled into
   1.120 + action by the DMA (direct memory access).
   1.121 + The DMA sends the samples to a chain of 0 or more soft-cores, where they
   1.122 + are transformed according to the soft-cores’ algorithms.
   1.123 + After running the gauntlet of soft-cores, the samples flow first to a buffering
   1.124 + FIFO, and finally to a mixer, which sends the samples off to be played
   1.125 + by speakers or stored in a file.
   1.126 + 
   1.127 +\end_layout
   1.128 +
   1.129 +\end_inset
   1.130 +
   1.131 +
   1.132 +\end_layout
   1.133 +
   1.134 +\begin_layout Plain Layout
   1.135 +
   1.136 +\end_layout
   1.137 +
   1.138 +\end_inset
   1.139 +
   1.140 +
   1.141 +\end_layout
   1.142 +
   1.143 +\begin_layout Subsection*
   1.144 +Steps
   1.145 +\end_layout
   1.146 +
   1.147 +\begin_layout Standard
   1.148 +The difficult part of this project is managing code reuse.
   1.149 + We need three things for success.
   1.150 + 
   1.151 +\end_layout
   1.152 +
   1.153 +\begin_layout Itemize
   1.154 +SMIPS processor -- Easy.
   1.155 + Just use the Lab 5 processors.
   1.156 +\end_layout
   1.157 +
   1.158 +\begin_layout Itemize
   1.159 +Some way to program the processors
   1.160 +\end_layout
   1.161 +
   1.162 +\begin_layout Itemize
   1.163 +DMA (Direct Memory Access) to load voices into the processors.
   1.164 +\end_layout
   1.165 +
   1.166 +\begin_layout Standard
   1.167 +We use ScratchPad to load code into the processors.
   1.168 + ScratchPad is an Intel module which implements a cache hierarchy.
   1.169 + The hierarchy reaches all the way from RAM created on the FPGA to on-chip
   1.170 + DRAM to RAM on the host computer all the way to the Hard Disk of the host
   1.171 + computer.
   1.172 + The first time a processor tries to access one of its instructions, the
   1.173 + cache goes all the way to the hard disk of the host computer to retrieve
   1.174 + the data.
   1.175 + Subsequent attempts to access this data only go as far as the on-chip DRAM.
   1.176 + Each processor has its own ScratchPad and thus can be programmed independently.
   1.177 + The ScratchPad abstraction allows each processor to run a program of any
   1.178 + size.
   1.179 +\end_layout
   1.180 +
   1.181 +\begin_layout Standard
   1.182 +Music access is achieved through RRR, another Intel abstraction which allows
   1.183 + us to treat the hard disk of the host computer as if it were a normal FIFO.
   1.184 +\end_layout
   1.185 +
   1.186 +\begin_layout Subsection*
   1.187 +News
   1.188 +\end_layout
   1.189 +
   1.190 +\begin_layout Standard
   1.191 +We have run our system with 12 sample voices and various combinations of
   1.192 + simple c voice processing programs and the results have been better than
   1.193 + software implementations.
   1.194 + Significantly, increasing the number of voices does not increase the processing
   1.195 + load since each voice is processed in parallel.
   1.196 +\end_layout
   1.197 +
   1.198 +\begin_layout Subsection*
   1.199 +Contributions
   1.200 +\end_layout
   1.201 +
   1.202 +\begin_layout Itemize
   1.203 +Implemented Pygar, a system for quick parallel processing of audio.
   1.204 + 
   1.205 +\end_layout
   1.206 +
   1.207 +\begin_layout Itemize
   1.208 +Implemented 4 basic algorithms which serve as components for this system
   1.209 + (identity, bit-shift, volume-change, and delay) 
   1.210 +\end_layout
   1.211 +
   1.212 +\begin_layout Itemize
   1.213 +Demonstrated Pygar out-performs software-only systems.
   1.214 + Pure-software systems have a limit of around 6 voices, while our system
   1.215 + achieves 12 voices in parallel with no architecturally imposed limit on
   1.216 + the number of voices.
   1.217 +\end_layout
   1.218 +
   1.219 +\end_body
   1.220 +\end_document