nlnet_2019_amdvlk_port.mdwn

   1 # NL.net proposal 2019-10-042
   2
   3 * [[questions]]
   4
   5 ## Project name
   6
   7 Port of AMDVLK/RADV 3D Driver to the Libre RISC-V SoC
   8
   9 ## Website / wiki
  10
  11 <https://libre-riscv.org/nlnet_2019_amdvlk_port>
  12
  13 Please be short and to the point in your answers; focus primarily on
  14 the what and how, not so much on the why. Add longer descriptions as
  15 attachments (see below). If English isn't your first language, don't
  16 worry - our reviewers don't care about spelling errors, only about
  17 great ideas. We apologise for the inconvenience of having to submit in
  18 English. On the up side, you can be as technical as you need to be (but
  19 you don't have to). Do stay concrete. Use plain text in your reply only,
  20 if you need any HTML to make your point please include this as attachment.
  21
  22 ## Abstract: Can you explain the whole project and its expected outcome(s).
  23
  24 The Libre RISCV SoC is being developed to provide a privacy-respecting
  25 modern processor, developed transparently and as libre to the bedrock
  26 as possible.  As a hybrid processor, it is intended to be both a CPU
  27 *and* a GPU.  GPUs are normally proprietary (and thus are perfect candidate
  28 attack vectors), as is the 3D driver software.
  29
  30 In January 2019, AMD released a fully-functioning libre-licensed Vulkan
  31 Driver for their Radeon GPUs.  A close examination of its source code
  32 shows that it would be relatively straightforward to replace the libraries
  33 that generate Radeon GPU assembly code with ones that generate assembly
  34 for the Libre RISC-V SoC, instead.
  35
  36 In addition, further investigation shows that RADV, the libre-licensed
  37 MESA 3D Driver, also supports SPIR-V (by way of conversion to MESA NIR),
  38 and, likewise, may be a good candidate for replacing Radeon with Libre
  39 RISC-V assembly.
  40
  41 Thus we intend to do exactly that: leverage the excellent work already
  42 done to create a libre-licensed commercial-grade Vulkan 3D driver that
  43 takes full advantage
  44 of the parallelism and Vectorisation in the hybrid Libre RISC-V SoC.
  45
  46 # Have you been involved with projects or organisations relevant to this project before? And if so, can you tell us a bit about your contributions?
  47
  48 Luke Leighton is an ethical technology specialist who has a consistent
  49 24-year track record of developing code in a real-time transparent
  50 (fully libre) fashion, and in managing Software Libre teams.  He is the
  51 lead developer on the Libre RISC-V SoC.
  52
  53 Jacob Lifshay is a software libre 3D expert who developed a Vulkan 3D
  54 software render engine under the GSoc2017 Programme.  He also developed
  55 his own libre-licensed 32-bit RISC-V processor, and has written an
  56 optimising javascript compiler.  Jacob is a valuable member of the team and is
  57 working on Kazan (https://salsa.debian.org/Kazan-team/kazan)
  58
  59 # Requested Amount
  60
  61 EUR 50,000.
  62
  63 # Explain what the requested budget will be used for?
  64
  65 After a thorough and comprehensive evaluation to see which will be the
  66 best to choose (RADV or AMDVLK), we are aiming for a multi-stage process,
  67 starting with the basics:
  68
  69 * The first stage is to remove AMD's "PAL" Library in AMDVLK, or the
  70   AMDGPU engine used in RADV, and replace it with a straightforward
  71   upstream port of the current LLVM JIT compiler, alongside a "support"
  72   library that will call OpenCL / OpenGL functions directly on the main
  73   processor.  This "effectively" turns the engine into a peer of google
  74   swiftshader (a "Software 3D Renderer") which will allow us to carry out
  75   rapid testing on stable x86 systems before moving on to the next stage.
  76 * The second stage is to confirm that the standard RISC-V LLVM JIT
  77   (which was recently upstreamed as of LLVM 9.0.0) is properly functional
  78   under an emulator or other RV64GC system.
  79 * The third phase will be to begin the iterative process, tying in closely
  80   with the work on Kazan, to experiment in both a software simulator
  81   as well as in FPGAs, with the addition of both Vectorisation as well
  82   as custom opcodes that will significantly improve performance as well
  83   as meet commercially-acceptable power-performance demands.
  84
  85 At the point where commercial power-performance requirements are met we may
  86 officially declare the project a "success".
  87
  88 # Does the project have other funding sources, both past and present?
  89
  90 The overall project has sponsorship from Purism as well as a prior grant
  91 from NLNet.  However that is for specifically covering the development
  92 of the RTL (the hardware source code), and for a "parallel" 3D Vulkan
  93 Driver effort, Kazan.
  94
  95 Kazan is to be written in Rust and its inclusion in the initial 2018
  96 proposal was *before* AMDVLK was released (Jan 2019).  AMDVLK is written
  97 in c++, is a significant way along (further than Kazan), however there
  98 is a different design focus in each that makes choosing one over the
  99 other not only difficult but potentially a costly mistake.
 100
 101 # Compare your own project with existing or historical efforts.
 102
 103 Nyuzi is a Software-based 3D Engine that has an LLVM port.  The problem
 104 is that it has deliberately been designed to be a software-only
 105 Vector Processor.  As such, with no custom accelerated opcodes
 106 dedicated to 3D, its power-performance metric is a whopping 25% that of
 107 commercially-acceptable 3D GPUs.  It also has no actual 3D Vulkan Driver:
 108 the developers focussed only on the "core algorithms" as part of an
 109 (extremely useful) academic exercise, only.
 110
 111 Google's swiftshader is a software-based 3D Driver/Engine that is compatible
 112 with at least one version of Vulkan.  On the face of it, this would be a
 113 perfect match for the Libre RISC-V SoC due to it being a hybrid CPU / GPU.
 114 The problem is that swiftshader was designed - from the ground up - never
 115 to have Vectorisation or any form of accelerated hardware beyond SIMD
 116 (NEON, Altivec, SSE/AVX).  As Nyuzi clearly shows, this approach is
 117 known to give a massive 400% power penalty.  Not only that, but our
 118 additions would not be welcome due to the primary focus of swiftshader
 119 being on non-hardware-accelerated, non-custom processors.
 120
 121 RADV is the free software competitor to AMDVLK.  It takes a different
 122 route: converting SPIR-V to NIR (New Internal Representation) which will
 123 need close evaluation to ensure that it's directly suited to Vector
 124 Processing.  Like AMDVLK, it does not directly support RISC-V: it was
 125 purely intended to support Radeon GPUs.
 126
 127 Our initial proposal - Kazan - is much more interesting to discern and
 128 compare against.  Kazan is being specifically designed so that the
 129 SPIR-V compiler is capable of fully supporting "full-function vectorisation".
 130 LLVM IR does *NOT* normally support this (which is why SPIR-V was created
 131 by the Khronos Group in the first place).  However, AMDVLK, which is a
 132 hard fork of LLVM, has had its LLVM-IR specifically modified to support
 133 both full-function vectorisation, predication, and in addition, texturisation,
 134 such that this information may be "carried" down to the Radeon assembly level,
 135 through the PAL library.
 136
 137 Standard LLVM does *not* support this full-function vectorisation
 138 capability: it is typically left up to any given assembly-level
 139 converter (such as the RISC-V Vector Engine) to "opportunistically"
 140 turn non-vectorised programs *into* vectorised ones, whereas AMDVLK
 141 *explicitly* carries this very same information.
 142
 143 Kazan on the other hand intends to perform explicit Vectorisation
 144 code-transformations in a different location: inside the SPIR-V compiler
 145 itself.  This key radical technical difference is why we seek to explore
 146 the alternative approach taken by AMD side-by-side with that of Kazan,
 147 because it is just not possible to predict in advance which would be "better".
 148
 149 ## What are significant technical challenges you expect to solve during the project, if any?
 150
 151 This is compiler technology, which is traditionally viewed as particularly
 152 challenging.  We are slightly fortunate in that much of the pieces of
 153 the puzzle already exist: AMDVLK, RADV, the upstreamed acceptance of
 154 RISC-V LLVM 9.0.0 being the key ones.
 155
 156 Whilst we know *technically* what they did and why they did it, the key
 157 challenge will be to unravel what exact changes AMD made which caused
 158 them to have to "fork" LLVM several years back, to keep track of their
 159 efforts to introduce "mainline" LLVM patches on an ongoing piecemeal
 160 basis, and at the same time *add our own assembler back-end* into the
 161 same fast-moving target.
 162
 163 Whereas with RADV it is upstreamed in MESA, and has much wider community
 164 support, it will need very careful detailed evaluation to ensure that it meets
 165 the needs of the Libre RISC-V Vector Engine.
 166
 167 ## Describe the ecosystem of the project, and how you will engage with relevant actors and promote the outcomes?
 168
 169 As mentioned in the 2018 submission, the Libre RISC-V
 170 SoC has a full set of resources for Libre Project Management and development:
 171 mailing list, bugtracker, git repository and wiki - all listed here:
 172 <https://libre-riscv.org/>
 173
 174 In addition, we have a Crowdsupply page
 175 <https://www.crowdsupply.com/libre-risc-v/m-class> which provides a public
 176 gateway, and heise.de, reddit, phoronix, slashdot and other locations have
 177 all picked up the story.  The list is updated and maintained here:
 178 <https://libre-riscv.org/3d_gpu/>
 179
 180 # Extra info to be submitted
 181
 182 * <http://libre-riscv.org/3d_gpu/>
 183 * <https://nlnet.nl/project/Libre-RISCV/>
 184 * <https://github.com/GPUOpen-Drivers/AMDVLK>
 185 * <https://github.com/google/swiftshader/>
 186 * <https://salsa.debian.org/Kazan-team/kazan>
 187 * <https://github.com/mesa3d/mesa/tree/master/src/amd/vulkan>
 188
 189 # Management Summary
 190
 191 The Libre SoC Project core is funded from an initial 2018 proposal. This includes a 3D Driver, called Kazan, and its purpose is to provide a Vulkan compliant hybrid hardware-software API.  Given the complex nature of 3D driver development, and because Kazan is a novel approach (written in rust, for security reasons) a second oroposal was submitted to develop a Mesa3D driver (in c++). A second more traditional (c++) 3D Driver allows for increased transparency and collaboration on this ambitious project.