OpenPOWER

In the late 1980s IBM developed a POWER family of processors. This evolved to a specification known as the POWER ISA. In 2019 IBM made the POWER ISA Open source, to be looked after by the existing OpenPOWER Foundation. Here is a longer history of IBM POWER microprocessors. These IBM proprietary processors happen to implement what is now known as the POWER ISA. The names POWER8, POWER9, POWER10 etc. are product designations equivalent to Intel i5, i7, i9 etc. and are frequently conflated with versions of the POWER ISA (v2.07, v3.0c, v3.1b).

Libre-SOC is basing its Simple-V Vectorisation CPU extensions on POWER ISA, because it wants to be able to specify a machine that can be completely trusted, and because POWER, thanks to IBM's involvement, is designed for high performance.

See wikipedia page https://en.m.wikipedia.org/wiki/Power_ISA

very useful resource describing all assembly instructions https://www.ibm.com/docs/en/aix/7.1?topic=reference-instruction-set

Evaluation

EULA released! looks good. https://openpowerfoundation.org/final-draft-of-the-power-isa-eula-released/

Links

PowerPC Unit Tests

Summary

  • FP32 is converted to FP64. Requires SimpleV to be active.
  • FP16 needed
  • transcendental FP opcodes needed (sin, cos, atan2, root, log1p)
  • FCVT between 16/32/64 needed
  • c++11 atomics not very efficient
  • no 16/48/64 opcodes, needs a shuffle of opcodes. TODO investigate Power VLE
  • needs escape sequencing (ISAMUX/NS) - see isans letter

What we are NOT doing:

  • A processor that is fundamentally incompatible (noncompliant) with Power. (escape-sequencing requires and guarantees compatibility).
  • Opcode 4 Signal Processing (SPE)
  • Opcode 4 Vectors or Opcode 60 VSX (600+ additional instructions)
  • Avoidable legacy opcodes
  • SIMD. it's awful.

SimpleV

see sv. SimpleV: a "hardware for-loop" which involves type-casting (both) the register files to "a sequence of elements". The one instruction (an unmodified scalar instruction) is interpreted as a hardware for-loop that issues multiple internal instructions with sequentially-incrementing register numbers.

Thus it is completely unnecessary to add any vector opcodes - at all - saving hugely on both hardware and compiler development time when the concept is dropped on top of a pre-existing ISA.

Integer Overflow / Saturate

Typically used on vector operations (audio DSP), it makes no sense to have separate opcodes (Opcode 4 SPE). To be done instead as CSRs / vector-flags on standard arithmetic operations.

atomics

Single instruction on RV, and x86, but multiple on Power. Needs investigation, particularly as to why cache flush exists.

https://www.cl.cam.ac.uk/~pes20/cpp/cpp0xmappings.html

Hot loops contain significant instruction count, really need new c++11 atomics. To be proposed as new extension because other OpenPower members will need them too

FP16

Doesn't exist in Power, need to work out suitable opcodes, basically means duplicating the entire range of FP32/64 ops, symmetrically.

Usually done with a fmt field, 2 bit, last one is FP128

idea: rather than add dozens of new opcodes, add "repurposer" instructions that remap FP32 to 16/32/64/128 and FP64 likewise. can also be done as C instruction, only needs 4 bits to specify.

Escape Sequencing

aka "ISAMUX/NS". Absolutely critical, also to have official endorsement from OpenPower Foundation.

This will allow extending ISA (see ISAMUX/NS) in a clean fashion (including for and by OpenPower Foundation)

Branches in namespaces

Branches are fine as it is up to the compiler to decide whether to let the ISAMUX/NS/escape-sequence countdown run out.

This is all a software / compiler / ABI issue.

Function calls in namespaces

Storing and restoring the state of the page/subpage CSR should be done by the caller. Or, again, let the countdowns run out.

If certain alternative configs are expected, they are part of the function ABI which must be spec'd.

All of this is a software issue (compiler / ABI).

Compressed, 48, 64, VBLOCK

TODO investigate Power VLE (Freescale doc Ref 314-68105)

Under Esc Seq, move mulli, twi, tdi out of major OP000 then use the entire row, 2 bits instead of 3. greatly simplifies decoder.

  • OP 000-000 and 000-001 for 16 bit compressed, 11 bit instructions
  • OP 000-010 and 000-011 for 48 bit. 11 bits for SVP P48
  • OP 000-100 and 000-201 for 64 bit. 11 bits for SVP P64
  • OP 000-110 and 000-111 for VBLOCK. 11 bits available.

Note that this requires BE instruction encoding (separate from data BE/LE encoding). BE encoding always places the major opcode in the first 2 bytes of the raw (uninterpreted) sequential instruction byte stream.

Thus in BE-instruction-mode, the first 2 bytes may be analysed to detect whether the instruction is 16-bit Compressed, 48-bit SVP-P48, 64-bit SVP-64, variable-length VBLOCK, or plain 32-bit.

It is not possible to distinguish LE-encoded 32-bit instructions from LE-encoded 16-bit instructions because in LE-encoded 32-bit instructions, the opcode falls into:

  • bytes 2 and 3 of any given raw (uninterpreted) sequential instruction byte stream for a 32-bit instruction
  • bytes 0 and 1 for a 16-bit Compressed instruction
  • bytes 4 and 5 for a 48-bit SVP P48
  • bytes 6 and 7 for a 64-bit SVP P64

Clearly this is an impossible situation, therefore BE is the only option. Note: this is completely separate from BE/LE for data

Compressed 16

Further "escape-sequencing".

Only 11 bits available. Idea: have "pages" where one instruction selects the page number. It also specifies for how long that page is activated (terminated on a branch)

The length to be a maximum of 4 bits, where 0b1111 indicates "permanently active".

Perhaps split OP000-000 and OP000-001 so that 2 pages can be active.

Store activation length in a CSR.

2nd idea: 11 bits can be used for extremely common operations, then length-encoding page selection for further ops, using the full 16 bit range and an entirely new encoding scheme. 1 bit specifies which of 2 pages was selected?

3rd idea: "stack" mechanism. Allow subpages like a stack, to page in new pages.

3 bits for subpage number. 4 bits for length, gives 7 bits. 4x7 is 28, then 3 bits can be used to specify "stack depth".

Requirements are to have one instruction in each subpage which resets all the way back to PowerISA default. The other is a "back up stack by 1".