shakti/m_class/libre_3d_gpu.mdwn

   1 # Libre 3D GPU Requirements
   2
   3 ## GPU capabilities
   4
   5 Based on GC800 the following would be acceptable performance (as would
   6 Mali-400):
   7
   8 * 35 million triangles/sec
   9 * 325 milllion pixels/sec
  10 * 6 GFLOPS
  11
  12 ## GPU size and power
  13
  14 * Basically the power requirement should be at or below around 1 watt
  15   in 40nm. Beyond 1 watt it becomes... difficult.
  16 * Size is not particularly critical as such but should not be insane.
  17
  18 Based on GC800 the following would be acceptable area in 40nm:
  19
  20 * 1.9mm^2 synthesis area
  21 * 2.5mm^2 silicon area.
  22
  23 So here's a table showing embedded cores:
  24
  25 <https://www.cnx-software.com/2013/01/19/gpus-comparison-arm-mali-vs-vivante-gcxxx-vs-powervr-sgx-vs-nvidia-geforce-ulp/>
  26
  27 Silicon area corresponds *ROUGHLY* with power usage, but PLEASE do
  28 not take that as absolute, because if you read Jeff's Nyuzi 2016 paper
  29 you'll see that getting data through the L1/L2 cache barrier is by far
  30 and above the biggest eater of power.
  31
  32 Note lower down that the numbers for Mali-400 are for the *4* core
  33 version - Mali-400 (MP4) - where Jeff and I compared Mali-400 SINGLE CORE
  34 and discovered that Nyuzi, if 4 parallel Nyuzi cores were put
  35 together, would reach only 25% of Mali-400's performance (in about the
  36 same silicon area).
  37
  38 ## Other
  39
  40 * The deadline is about 12-18 months.
  41 * It is highly recommended to use Gallium3D for the software stack
  42   (see below if deciding whether to use Nyuzi or RISC-V or other)
  43 * Software must be licensed under LGPLv2+ or BSD/MIT.
  44 * Hardware (RTL) must be licensed under BSD or MIT with no
  45   "NON-COMMERCIAL" CLAUSES.
  46 * Any proposals will be competing against Vivante GC800 (using Etnaviv driver).
  47 * The GPU is integrated (like Mali-400). So all that the GPU needs
  48   to do is write to an area of memory (framebuffer or area of the
  49   framebuffer). The SoC - which in this case has a RISC-V core and has
  50   peripherals such as the LCD controller - will take care of the rest.
  51 * In this arcitecture, the GPU, the CPU and the peripherals are all on
  52   the same AXI4 shared memory bus. They all have access to the same shared
  53   DDR3/DDR4 RAM. So as a result the GPU will use AXI4 to write directly
  54   to the framebuffer and the rest will be handle by SoC.
  55 * The job must be done by a team that shows sufficient expertise to
  56   reduce the risk.
  57
  58 ## Notes
  59
  60 * The deadline is really tight. If an FPGA (or simulation) plus the basics
  61   of the software driver are at least prototyped by then it *might* be ok.
  62 * If using Nyuzi as the basis it *might* be possible to begin the
  63   software port in parallel because Jeff went to the trouble of writing
  64   a cycle-accurate simulation.
  65 * I *suspect* it will result in less work to use Gallium3D than, for
  66   example, writing an entire OpenGL stack from scratch.
  67 * A *demo* should run on an FPGA as an initial. The FPGA is not a priority
  68   for assessment, but it would be *nice* if it could fit into a ZC706.
  69 * Also if there is parallel hardware obviously it would be nice to be able
  70   to demonstrate parallelism to the maximum extend possible. But again,
  71   being reasonable, if the GPU is so big that only a single core can fit
  72   into even a large FPGA then for an initial demo that would be fine.
  73 * Note that no other licenses are acceptable for the hardware: all GPL
  74   licenses (GPL, AGPL, LGPL) are out.  GPL (all revisions v2, v3, v2+, v3+)
  75   are out for software, with the exception of the LGPL (v2+ or v3+ acceptable).
  76
  77 ## Design decisions and considerations
  78
  79 Whilst Nyuzi has a big advantage in that it has simuations and also a
  80 llvm port and so on, if utilised for this particular RISC-V chip it would
  81 mean needing to write a "memory shim" between the general-purpose Nyuzi
  82 core and the main processor, i.e. all the shader info, state etc. needs
  83 synchronisation hardware (and software). That could significantly
  84 complicate design, especially of software.
  85
  86 Whilst i *recommended* Gallium3D there is actually another possible approach:
  87
  88 A RISC-V multi-core design which accelerates *software*
  89 rendering... including potentially utilising the fact that Gallium3D
  90 has a *software* (LLVM) renderer:
  91
  92 <https://mesa3d.org/llvmpipe.html>
  93
  94 The general aim of this approach is *not* to have the complexity of
  95 transferring significant amounts of data structures to and from disparate
  96 cores (one Nyuzi, one RISC-V) but to STAY WITHIN THE RISC-V ARCHITECTURE
  97 and simply compile Mesa3D (for RISC-V), gallium3d-llvm (for RISC-V),
  98 modifying llvm for RISC-V to do the heavy-lifting instead.
  99
 100 Then it just becomes a matter of adding Vector/SIMD/Parallelization
 101 extensions to RISC-V, and adding support in LLVM for the same:
 102
 103 <https://lists.llvm.org/pipermail/llvm-dev/2018-April/122517.html>
 104
 105 So if considering to base the design on RISC-V, that means turning RISC-V
 106 into a vector processor. Now, whilst Hwacha has been located (finally),
 107 it's a design that is specifically targetted at supercomputers. I have
 108 been taking an alternative approach to vectorisation which is more about
 109 *parallelization* than it is about *vectorization*.
 110
 111 It would be great for Simple-V to be given consideration for
 112 implementation as the abstraction "API" of Simple-V would greatly simplify
 113 the addition process of Custom features such as fixed-function pixel
 114 conversion and rasterization instructions (if those are chosen to be
 115 added) and so on. Bear in mind that a high-speed clock rate is NOT a
 116 good idea for GPUs (power being a square law), multi-core parallelism
 117 and longer SIMD/vectors are much better to consider, instead.
 118
 119 The PDF/slides on Simple-V is here:
 120
 121 <http://hands.com/~lkcl/simple_v_chennai_2018.pdf>
 122
 123 And the assessment, design and implementation is being done here:
 124
 125 <http://libre-riscv.org/simple_v_extension/>
 126
 127 ----
 128
 129 My feeling on this is therefore that the following approach is one which involve minimal work:
 130
 131 * Investigate the ChiselGPU code to see if it can be leveraged (an
 132   "image" added instead of straight ARGB color).
 133 * OR... add sufficient fixed-function 3D instructions (plus a memory
 134   scratch area) to RISC-V to do the equivalent job.
 135 * Implement the Simple-V RISC-V "parallelism" extension (which can
 136   parallelize xBitManip *and* the above-suggested 3D fixed-function
 137   instructions).
 138 * Wait for RISC-V LLVM to have vectorization support added to it.
 139 * MODIFY the resultant RISC-V LLVM code so that it supports Simple-V.
 140 * Grab the gallium3d-llvm source code and hit the "compile" button.
 141 * Grab the *standard* Mesa3D library, tell it to use the gallium3d-llvm library and hit the "compile" button.
 142 * see what happens.
 143
 144 Now, interestingly, if spike is thrown into the mix there (as a
 145 cycle-accurate RISC-V simulator) it should be perfectly well possible to
 146 get an idea of where performance of the above would need optimization,
 147 just like Jeff did with the Nyuzi paper.
 148
 149 He focussed on specific algorithms and checked the assembly code, and
 150 worked out how many instruction cycles per pixel were needed, which is
 151 an invaluable measure.
 152
 153 As I mention in the above page, one of the problems with doing a
 154 completely separate engine (Nyuzi is actually a general-purpose RISC-based
 155 vector processor) is that when it comes to using it, you need to transfer
 156 all the "state" data structures from the main core over to the GPU's core.
 157
 158 ... But if the main core is RISC-V *and the GPU is RISC-V as well*
 159 and they are SMP cores then transferring the state is a simple matter of
 160 doing a context-switch... or if *all* cores have vector and 3D instruction
 161 extensions, a context-switch is not needed at all.
 162
 163 Will that approach work? Honestly I have absolutely no idea, but it
 164 would be a fascinating and extremely ambitious research project.
 165
 166 Can we get people to fund it?  Yeah I do.  there's a lot of buzz about
 167 RISC-V, and a lot of buzz can be created about a libre 3D GPU. If that
 168 same GPU happens to be good at doing crypto-currency mining there will be
 169 a LOT more attention paid, particularly given that people have noticed
 170 that relying on proprietary GPUs and CPUs to manage billions of dollars
 171 worth of crypto-currency, when the NSA is *known* to have blackmailed
 172 intel into putting a spying back-door co-processor in to x86, and that
 173 it miiight not be a good idea to trust proprietary hardware:
 174
 175 <http://libreboot.org/faq#intelme>
 176
 177 ## Q & A
 178
 179 > Q:
 180 >
 181 > Do you need a team with good CVs? What about if the team shows you
 182 > an acceptable FPGA prototype? I’m talking about a team of students
 183 > which do not have big industrial CVs but they know how to handle this
 184 > job (just like RocketChip or MIAOW or etc…).
 185
 186 A:
 187
 188 That would be fantastic as it would demonstrate not only competence but
 189 also commitment. And will have taken out the "risk" of being "unknown",
 190 entirely. So that works perfectly for me :) .
 191
 192 > Q:
 193 >
 194 > Is there any guarantee that there would be a sponsorship for the GPU?
 195
 196 A:
 197
 198 Please please let's be absolutely clear:
 199
 200 I can put the *business case* to the anonymous sponsor to *consider*
 201 sponsoring a libre GPU, *only* and purely on the basis of a *commercial*
 202 decision based on cost and risk analysis, comparing against the best
 203 alternative option which is USD $250,000 for a one-time proprietary
 204 license for Vivante GC800 using etnaviv. So as a result we need to be
 205 *really clear* that *there is no "guaranteed sponsorship"*.  this is a
 206 pure commercial *business* assessment.
 207
 208 However, it just so happens that there's quite a lot of people who are
 209 pissed at how things go in the 3D embedded space. That can be leveraged,
 210 by way of a crowd-funding campaign, to invite people to help, put money
 211 behind this that has *nothing to do with the libre-riscv anonymous
 212 sponsor*.