openpower/isans_letter.mdwn

   1 (Draft Status)
   2
   3 # Letter regarding ISAMUX / NS
   4
   5 This is a quick overview of the way that we would like to add changes
   6 that we are proposing to the PowerPC instruction set.  It is based on
   7 a Open Standardisation of the way that existing "mode switches",
   8 already found in the POWER instruction set, are added:
   9
  10 * FPSCR's "NI" bit, setting non-IEEE754 FP mode
  11 * MSR's "LE" bit (and associated HILE bit), setting little-endian mode
  12 * MSR's "SF" bit, setting either 32-bit or 64-bit mode
  13
  14 All of these are set by one instruction, that, once set, radically
  15 changes the entire behaviour and characteristics of subsequent instructions.
  16
  17 With these (and other) long-established precedents already in POWER,
  18 there is therefore essentially conceptually nothing new about what we
  19 propose: we simply seek that the process by which such "switching" is
  20 added is formalised and standardised, such that we (and others) have
  21 a clear, standards-non-disruptive, atomic and non-intrusive path to
  22 extend the POWER ISA.
  23
  24 # Summary of Libre-SOC Project
  25
  26 TODO brief summary of Libre-SOC project (hybrid CPU-GPU-VPU), thereby
  27 helping explain exactly why we need extensive augmentation of POWER ISA.
  28
  29 Basically it's because it is not a separate GPU-VPU, it's an *actual*
  30 CPU-GPU-VPU.  No separate GPU, because the CPU *is* the GPU.  No separate
  31 VPU, because the CPU *is* the GPU.  There is not even a separate "pipeline":
  32 the CPU pipelines *are* the GPU and VPU pipelines.
  33
  34 Closest equivalents include the ARC core (which has VPU extensions and
  35 3D extensions in the form of Broadcom's VideoCore IV) and the ICubeCorp
  36 IC3128.  Both are considered "hybrid" CPU-GPU-VPU processors.
  37
  38 With the project being Libre - not proprietary and secretive and never
  39 to be published, ever - it is no good having the extensions as "custom"
  40 because "custom" is specifically for the cases where the augmented
  41 toolchain is never, under any circumstances, published and made public by
  42 the proprietary company.  For business commercial reasons, Libre-SOC is
  43 the total opposite of this proprietary, secretive approach.
  44
  45 ## Overview
  46
  47 The PowerPC Instruction Set Architecture (ISA) is an abstract model of a
  48 computer. This is what programmers use when they write programs for the machine,
  49 even if indirectly via a compiler for a high level language. We must be
  50 conservative in how we add to the ISA to:
  51
  52 * not break existing programs
  53 * be mindful as to how others may wish to add to the ISA in the future
  54
  55 This document describes our strategy.
  56
  57
  58 ## ISA modes and escape sequences
  59
  60 New chips usually need to be able to run older (legacy) software that is
  61 incompatible with the latest and greatest ISA. Eg: 64 bit chip must be able to
  62 run older 16 bit and 32 bit software.
  63
  64 To enable backwards compatability the CPU will be set into 'legacy' mode. This
  65 is done with an ISA Mode switch, also known as ISA Muxing or ISA Namespaces.
  66
  67 The operating system is able to quickly change between 'modern' ISA mode and
  68 various legacy modes.
  69
  70 Another technique is an ISA escape-sequence. This is a type of mode that is
  71 only operational for a short time, unlike 32 or 64 bit which would be for the
  72 entire run of a program.
  73
  74
  75 ## What are we adding to the ISA
  76
  77 When high quality graphical display were developed the CPUs at the time were
  78 shown to not be able to run the display fast enough. The solution was the use of
  79 Graphics cards, these are specialised computers that are good at rendering
  80 pixels; often by doing the same thing in different parts of the screen at the
  81 same time (in parallel). These specialised computers are called Graphical
  82 Processing Units (GPUs).
  83
  84 The parallelism of some GPUs is thousands. This has led to GPUs being used to
  85 solve non graphical problems where high parallelism is useful.
  86
  87 **break**
  88
  89 # Letter regarding ISAMUX / NS
  90
  91 Hardware-level dynamic ISA Muxing (also known as ISA Namespaces and ISA
  92 escape-sequencing) is commonly used in instruction sets, in an arbitrary
  93 and ad-hoc fashion, added often on an on-demand basis.  Examples include:
  94
  95 * Setting a SPR to switch the meaning of certain opcodes for Little-Endian /
  96   Big-Endian behaviour (present in POWER and SPARC)
  97 * Setting a SPR to provide "backwards-compatibility" for features from
  98   older versions of an ISA (such as changing to new ratified versions of
  99   the IEEE754 standard)
 100
 101 (These we term "ISA Muxing" because, ultimately, they are extra bits
 102 (or change existing bits) in the actual instruction decoder phase,
 103 which involves "MUXes" to switch them on and off).
 104
 105 The Libre-SOC team, developing a hybrid CPU-VPU-GPU, needs to add
 106 significantly and strategically to the POWER ISA to support, for example,
 107 Khronos Vulkan IEEE754 Conformance, whilst *at the same time being able
 108 to run full POWER9 compliant instructions*.
 109
 110 There is absolutely no way that we are going to duplicate the
 111 entire FP opcode set as a custom extension to POWER, just to add a
 112 literally-identical suite of FP opcodes that are compliant with the
 113 Khronos Conformance Suites: this would be a significant and irresponsible
 114 use of opcode space.
 115
 116 In addition, as this processor is likely to be used for parallel
 117 compute purposes in high-efficiency environments, we also need to add
 118 FP16 support.  Again: there is no way that we are going to add *triple*
 119 duplicated opcodes to POWER, given that the opcodes needed are absolutely
 120 identical to those that already exist, apart from the FP bitwidth (32
 121 / 64).
 122
 123 There are several other strategically critical uses to which we would
 124 like to put such a scheme (related to power consumption and reducing
 125 throughput bottlenecks needed for heavy-computation workloads in GPU
 126 and VPU scenarios).
 127
 128 In addition, the scheme has several other key advantages over other ISA
 129 "extending" ideas (such as extending the general ISA POWER space to
 130 64 bit) in that, unlike 64 bit opcodes, its judicious and careful use
 131 does not require large increases in I-Cache size because all opcodes,
 132 ultimately, remain 32-bit.  The scheme also allows future *official*
 133 POWER extensions to the ISA - managed by the OpenPOWER Foundation -
 134 to be strategically managed in a controlled, long-term, non-damaging
 135 way to the reputation and stability of OpenPOWER.
 136
 137 Therefore we advocate being able to set "ISAMUX/NS" mode-switching bits
 138 that, like the *existing* LE/BE mode-switching bits, change the behaviour
 139 of *existing* opcodes to an alternative "meaning" (followed by another
 140 mode-switch that returns them to their original meaning. Note: to reduce
 141 binary code-size, alternative schemes include setting a countdown which,
 142 when it expires, automatically disables the requested mode-switch)
 143
 144 Note also that to ensure that kernels and hypervisors are not impacted
 145 by userspace ISAMUX/NS mode-switching, it is critical that Supervisor
 146 and Hypervisor modes have their own completely separate ISAMUX/NS SPRs
 147 (imagine a userspace application setting the LE/BE bit on a global basis,
 148 or setting a global IEEE754 FP Standards compatibility flag).
 149
 150 Further, that Supervisor / Hypervisor modes have access to and control
 151 over userspace ISAMUX/NS SPRs (without themselves being affected by
 152 setting *of* userspace ISAMUX/NS SPRs), in order to be able to correctly
 153 context-switch userspace applications to their correct (former) running
 154 state.
 155
 156 Given the number of mode-switch bits that we anticipate using, we advocate
 157 that such a scheme be formalised, and that the OpenPOWER Foundation be
 158 the "atomic arbiter" similar to IANA and JEDEC in the formal allocation
 159 of mode-switch bits to OpenPOWER implementors.
 160
 161 We envisage that some of these bits will be unary, some will be binary,
 162 some will be allocated for exclusive use by the OpenPOWER Foundation,
 163 some allocated to OpenPOWER Members (by the OpenPOWER Foundation),
 164 and some reserved for "custom and experimentation usage".
 165
 166 (This latter - custom experimentation - to be explicitly documented
 167 that upstream compiler and toolchain support will never, under any
 168 circumstances be accepted by the OpenPOWER Foundation, and that this be
 169 enforced through the EULA and through Trademark law).
 170
 171
 172 However as we are quite new to POWER 3.0B (1300+ page PDF), we do
 173 appreciate that such a formal scheme may already be present in POWER9
 174 3.0B, that we have simply overlooked.
 175