simple_v_extension/open_3d_alliance_2019aug26.tex

   1 \documentclass[slidestop]{beamer}
   2 \usepackage{beamerthemesplit}
   3 \usepackage{graphics}
   4 \usepackage{pstricks}
   5
   6 \title{Open 3D Alliance RISC-V}
   7 \author{Luke Kenneth Casson Leighton}
   8
   9
  10 \begin{document}
  11
  12 \frame{
  13    \begin{center}
  14     \huge{Open 3D Alliance: RISC-V}\\
  15     \vspace{32pt}
  16     \Large{An open invitation to collaborate on 3D Graphics}\\
  17     \Large{Hardware and Software}\\
  18     \Large{for mobile, embedded, and innovative purposes}\\
  19     \vspace{24pt}
  20     \Large{With thanks to Pixilica, GoWin, and Western Digital}\\
  21     \vspace{16pt}
  22     \large{\today}
  23   \end{center}
  24 }
  25
  26
  27 \frame{\frametitle{Why collaborate?}
  28
  29  \begin{itemize}
  30    \item 3D is hard. It's also not the same as HPC\vspace{15pt}
  31    \item NVIDIA, AMD, Imagination - cannot meet "unusual" needs\vspace{15pt}
  32    \item Working together on flexible standards, everyone wins\vspace{15pt}
  33    \item Without collaboration: 10-20 man-years development\vspace{10pt}
  34    \item With collaboration: cross-verification (avoids mistakes)
  35   \end{itemize}
  36 }
  37
  38
  39 \frame{\frametitle{What is the goal?}
  40
  41  \begin{itemize}
  42    \item You get to decide! No, really!\vspace{12pt}
  43    \item Outlined here: some ideas and cost/time-saving approaches\vspace{12pt}
  44    \item Two new platforms: 3D "Embedded", 3D "UNIX"\vspace{12pt}
  45    \item Flexible optional extensions (Transcendentals, Vectors,\\
  46                  Texturisation, Pixel/Z-Buffers - all optional)\vspace{12pt}
  47    \item Good software support absolutely essential\\
  48              (basically, that means Vulkan)\vspace{15pt}
  49   \end{itemize}
  50 }
  51
  52 \frame{\frametitle{Libre RISC-V Team}
  53
  54  \begin{itemize}
  55    \item Small team, sponsored by Purism and the NLNet Foundation\vspace{8pt}
  56    \item Therefore, focus is on efficiency: leap-frogging ahead\\
  57              without requiring huge resources.\vspace{8pt}
  58    \item OpenGL API? Gallium3D / Vulkan is better\vspace{8pt}
  59    \item Gallium3D turns out to be a single-threaded interpreter\\
  60              (Vulkan is compiled, and can be parallelised)\vspace{8pt}
  61    \item Independent teams have provided OpenGL to Vulkan adaptors\vspace{8pt}
  62    \item Same approach on hardware: seek highest bang-per-buck\\
  63                  Save design time, save implementation time\vspace{8pt}
  64   \end{itemize}
  65 }
  66
  67 \frame{\frametitle{What (optional) things are needed?}
  68
  69  \begin{itemize}
  70    \item Vectorisation.  (SIMD? RVV? Other?)\vspace{12pt}
  71    \item Transcendentals (SIN, COS, EXP, LOG)\vspace{12pt}
  72    \item Texture opcodes, Pixel/Z-Buffers\vspace{12pt}
  73    \item Pixel conversion (YUV/RGB etc.)\vspace{12pt}
  74    \item Optional accuracy (embedded space needs less accuracy)\vspace{12pt}
  75    \item Options give implementors flexibility.  No imposition:\\
  76          imposition risks fragmentation (however, collaboration does\\
  77          need some hard easily-logically-justifiable rules)
  78   \end{itemize}
  79 }
  80
  81 \frame{\frametitle{What is essential (not really optional)}
  82
  83  \begin{itemize}
  84    \item The software, basically.  Anything other than Vulkan\\
  85              is a 10+ man-year effort
  86    \item Two new 3D "platforms".  Vulkan compliance has implications\\
  87              for hardware, and, with the API being public, interoperability\\
  88              (and Khronos Compliance - which is Trademarked) is critical.
  89    \item Respecting that standards are hard to get right\\
  90          (and that consequences of mistakes are severe:\\
  91          no opportunity for corrections after a freeze)
  92    \item Respecting that, for collaboration and interoperability,\\
  93          some things go into a standard that you might not "need"
  94    \item Mutually respectful open and fully transparent collaboration.\\
  95          No NDAs, no "closed forums".  We need the help of experts\\
  96          (such as Mitch Alsup) in this highly technical specialist area.
  97   \end{itemize}
  98 }
  99
 100 \frame{\frametitle{Why Two new Platforms?}
 101
 102  \begin{itemize}
 103    \item Unique pragmatic consequences of "Hybrid" CPU/GPU
 104    \item Embedded - no traps need be raised. Interoperability is\\
 105    impossible, software toolchain collaboration is incidental).
 106    \item UNIX - illegal instruction traps mandatory: software\\
 107    interoperability is mandatory and essential.
 108    \item 3D Embedded - failure to allow implementors the freedom\\
 109    to reduce FP accuracy automatically results in product failure\\
 110    (too many gates, too much power, equals end-user rejection).
 111    \item 3D UNIX - likewise.  Also: failure to comply with Khronos\\
 112    Specifications (then use "Vulkan") is a Trademark violation.
 113    \item Solution: allow software to select FP accuracy level\\
 114    \textbf{at runtime}. (UNIX Platform: IEEE754. 3D UNIX: Vulkan).\\
 115    \item HW: slow for IEEE754, fast for 3D. Product now competitive!
 116  \end{itemize}
 117 }
 118
 119 \frame{\frametitle{What has our team done already?}
 120
 121  \begin{itemize}
 122    \item Decided to go the "Hybrid" Route (Separate GPUs requires a\\
 123          full-blown RPC/IPC mechanism to transfer all 3D API calls\\
 124          to and from userspace memory to GPU memory... and back).
 125    \item Developed Simple-V (a "Parallelising" API)\\
 126                  (Simple-V is very hard to describe, because it is unique:\\
 127                  there is no common Computer Science terminology)
 128    \item Started on Kazan (a Vulkan SPIR-V to LLVM compiler)
 129    \item Started work on a highly flexible IEEE754 FPU
 130    \item Started work on a "Precise" CDC 6600 style OoO Engine,\\
 131              with help from Mitch Alsup, the designer of the M68000
 132    \item Variable-issue, predicated SIMD backend, Vector front-end\\
 133              "precise" exceptions, branch shadowing, much more
 134    \item All Libre-licensed and developed publicly and transparently.
 135   \end{itemize}
 136 }
 137
 138 \frame{\frametitle{Why Simple-V? Why not RVV?}
 139
 140  \begin{itemize}
 141    \item RVV is designed exclusively for supercomputing\\
 142          (RVV simply has not been designed with 3D in mind).\vspace{6pt}
 143    \item Like SIMD, RVV uses dedicated opcodes\\
 144              (google "SIMD considered harmful")\vspace{6pt}
 145    \item 98\% of FP opcodes are duplicated in RVV. Large portion\\
 146          of BitManip opcodes duplicated in predicate Masks\vspace{6pt}
 147    \item OP32 space is extremely precious: 48 and 64 bit opcode space\\
 148          comes with an inherent I-Cache power consumption penalty\vspace{6pt}
 149    \item Simple-V "prefixes" scalar opcodes (all of them)\\
 150              No need for any new "vector" opcodes (at all).\\
 151              Can therefore use the RVV major opcode for 3D\vspace{6pt}
 152   \end{itemize}
 153 }
 154
 155 \frame{\frametitle{Simple-V "Prefixing"}
 156
 157  \begin{itemize}
 158    \item SV "Prefix" does exactly that: takes RVC and OP32 opcodes\\
 159          and "prefixes" them with predication and a "vector" tag\vspace{8pt}
 160    \item Three prefix types: SV P32 (prefixed RVC), P48 and P64\vspace{8pt}
 161    \item Prefixed RVC takes 3 "Custom" OP32 opcodes.\\
 162              P48 takes standard OP32 scalar opcodes and "prefixes" them\\
 163              P64 adds additional vector context on top of P48\\
 164              \vspace{8pt}
 165    \item "Prefixing" is a bit like SIMD.  Vectors may be specified\\
 166              of length 2 to 4, elements may be "packed" into registers,\\
 167              opcode element widths over-ridden.\vspace{8pt}
 168    \item Convenient, but not very space-efficient (and VBLOCK is)\vspace{8pt}
 169   \end{itemize}
 170 }
 171
 172 \frame{\frametitle{VBLOCK Format}
 173
 174  \begin{itemize}
 175    \item Again: hard to describe.  It is a bit like VLIW (only not really)\\
 176              A "block" of instructions is "prefixed" with register "tags"\\
 177                  which give extra context to scalar instructions within the block
 178   \item Sub-blocks include: Vector Length, Swizzling, Vector/Width\\
 179          overrides, and predication.  All this is added to scalar opcodes!\\
 180          \textbf{There are NO vector opcodes} (and no need for any)
 181    \item In the "context", it goes like this: "if a register is used\\
 182          by a scalar opcode, and the register is listed in the "context",\\
 183          SV mode is "activated"
 184    \item "Activation" results in a hardware-level "for-loop" issuing\\
 185          \textbf{multiple} contiguous scalar operations (instead of just one).
 186    \item Implementors are free to implement the "loop" in any fashion\\
 187          they see fit.  SIMD, Multi-issue, single-execution: anything.
 188   \end{itemize}
 189 }
 190
 191 \frame{\frametitle{Other Standard Proposals}
 192
 193  \begin{itemize}
 194    \item Ztrans and Ztrig* - Transcendentals and Trigonometrics\\
 195          (optional so that Embedded implementors have some leeway)
 196    \item ISAMUX / ISANS - stops arguments over OP32 space\\
 197          (also allows clean "paging" of new opcodes into e.g. RVC)
 198    \item MV.SWIZZLE and MV.X - RV does not have a MV opcode.
 199    \item Zfacc - dynamic FP accuracy.  Needed for "fast" Vulkan native\\
 200              and to switch between fast 3D accuracy and IEEE754 modes.
 201    \item These - and more - need your input! 3D is hard!
 202    \item The key strategic premise: these are required as \textbf{public}\\
 203          standards, because the \textbf{software} is to be public.
 204    \item This is \textbf{not} understood by the RISC-V Foundation.\\
 205          ("custom" status not appropriate for high-profile mass-volume\\
 206          end-user APIs such as Vulkan).
 207   \end{itemize}
 208 }
 209
 210
 211 \frame{\frametitle{Summary}
 212
 213  \begin{itemize}
 214    \item 3D is hard (and pure Vectorisation gets you 25\% of \\
 215          commercially-acceptable performance).
 216    \item Layered optional extensions are going to be key to\\
 217          acceptance by a wide variety of 3D Alliance Members.
 218    \item With a custom specialised SPIR-V (Vulkan) Compiler\\
 219          being an absolutely critical strategic requirement,\\
 220          RVV and its associated compiler (still not developed)\\
 221          is of marginal value (no clear benefits, extra cost)
 222    \item Question everything!  Your input, and a willingness to\\
 223              take active responsibility for tasks that your Company\\
 224              is critically dependent on, are extremely important.
 225    \item Public and transparent Collaboration is key.  There is simply\\
 226          too much to do.
 227   \end{itemize}
 228 }
 229
 230
 231 \frame{
 232   \begin{center}
 233     {\Huge \vspace{20pt}
 234                    The end\vspace{20pt}\\
 235                    Thank you\vspace{20pt}\\
 236         }
 237   \end{center}
 238
 239   \begin{itemize}
 240         \item http://lists.libre-riscv.org/pipermail/libre-riscv-dev/
 241         \item http://libre-riscv.org/simple\_v\_extension/abridged\_spec/
 242         \item https://libre-riscv.org/ztrans\_proposal/
 243         \item https://libre-riscv.org/simple\_v\_extension/specification/mv.x/
 244   \end{itemize}
 245 }
 246
 247
 248 \end{document}