The Rules

SVP64 is designed around fundamental and inviolate RISC principles. This gives a uniformity and regularity to the ISA, making implementation straightforward, which was why RISC as a concept became popular.

  1. There are no actual Vector instructions: Scalar instructions are the sole exclusive bedrock.
  2. No scalar instruction ever deviates in its encoding or meaning just because it is prefixed (semantic caveats below)
  3. A hardware-level for-loop (the prefix) makes vector elements 100% synonymous with scalar instructions (the suffix)

How can a Vector ISA even exist when no actual Vector instructions are permitted to be added? It comes down to the strict RISC abstraction. First you start from a scalar instruction (32-bit). Second, the Prefixing is applied in the abstract to give the appearance and ultimately the same effect as if an explicit Vector instruction had also been added. Looking at the pseudocode of any Vector ISA (RVV, NEC SX Aurora, Cray) they always comprise (a) a for-loop around (b) element-based operations. It is perfectly reasonable and rational to separate (a) from (b) then find a powerful pre-existing Supercomputing-class ISA that qualifies for (b).

There are a few exceptional places where these rules get bent, and others where the rules take some explaining, and this page tracks them all.

The modification caveat in (2) above semantically exempts element width overrides, which still do not actually modify the meaning of the instruction: an add remains an add, even if its override makes it an 8-bit add rather than a 64-bit add. Even add-with-carry remains an add-with-carry: it's just that when elwidth=8 in the Prefix it's an 8-bit add-with-carry where the 9th bit becomes Carry-out (not the 65th bit). In other words, elwidth overrides definitely do not fundamentally alter the actual Scalar v3.0 ISA encoding itself. Consequently we can still, in the strictest semantic sense, not be breaking rule (2).

Likewise, other "modifications" such as saturation or Data-dependent Fail-First likewise are actually post-augmentation or post-analysis, and do not fundamentally change an add operation into a subtract for example, and under absolutely no circumstances do the actual 32-bit Scalar v3.0 operand field bits change or the number of operands change.

In an early Draft of SVP64, an experiment was attempted, to modify LD-immediate instructions to include a third RC register i.e. reinterpret the normal v3.0 32-bit instruction as a completely different encoding if SVP64-prefixed. It did not go well. The complexity that resulted in the decode phase was too great. The lesson was learned, the hard way: it would be infinitely preferable to add a 32-bit Scalar Load-with-Shift instruction first, which then inherently becomes Vectorised. Perhaps a future Power ISA spec will have this Load-with-Shift instruction: both ARM and x86 have it, because it saves greatly on instruction count in hot-loops.

The other reason for not adding an SVP64-Prefixed instruction without also having it as a Scalar un-prefixed instruction is that if the 32-bit encoding is ever allocated in a future revision of the Power ISA to a completely unrelated operation then how can a Vectorised version of that new instruction ever be added? The uniformity and RISC Abstraction is irreparably damaged. Bottom line here is that the fundamental RISC Principle is strictly adhered to, even though these are Advanced 64-bit Vector instructions. Advocates of the RISC Principle will appreciate the uniformity of SVP64 and the level of systematic abstraction kept between Prefix and Suffix.

Instruction Groups

The basic principle of SVP64 is the prefix, which contains mode as well as register augmentation and predicates. When thinking of instructions and Vectorising them, it is natural for arithmetic operations (ADD, OR) to be the first to spring to mind. Arithmetic instructions have registers, therefore augmentation applies, end of story, right?

Except, Load and Store deals also with Memory, not just registers. Power ISA has Condition Register Fields: how can element widths apply there? And branches: how can you have Saturation on something that does not return an arithmetic result? In short: there are actually four different categories (five including those for which Vectorisation makes no sense at all, such as sc or mtmsr). The categories are:

  • arithmetic/logical including floating-point
  • Load/Store
  • Condition Register Field operations
  • branch

Arithmetic

Arithmetic (known as "normal" mode) is where Scalar and Parallel Reduction can be done: Saturation as well, and two new innovative modes for Vector ISAs: data-dependent fail-first and predicate result. Reduction and Saturation are common to see in Vector ISAs: it is just that they are usually added as explicit instructions, and NEC SX Aurora has even more iterative instructions. In SVP64 these concepts are applied in the abstract general form, which takes some getting used to.

Reduction may, when applied to non-commutative instructions incorrectly, result in invalid results, but ultimately it is critical to think in terms of the "rules", that everything is Scalar instructions in strict Program Order. Reduction on non-commutative Scalar Operations is not prohibited: the strict Program Order allows the programmer to think through what would happen and thus potentially actually come up with legitimate use.

Branches

Branch is the one and only place where the Scalar (non-prefixed) operations differ from the Vector (element) instructions, as explained in a separate section. The RM bits can be used for other purposes because the Arithmetic modes make no sense at all for a Branch. Almost the entire SVP64 RM Field is interpreted differently from other Modes, in order to support a wide range of parallel boolean condition options which are expected of a Vector / GPU ISA. These save a considerable number of instructions in tight inner loop situations.

CR Field Ops

Condition Register Fields are 4-bit wide and consequently element-width overrides make absolutely no sense whatsoever. Therefore the elwidth override field bits can be used for other purposes when Vectorising CR Field instructions. Moreover, Rc=1 is completely invalid for CR operations such as crand: Rc=1 is for arithmetic operations, producing a "co-result" that goes into CR0 or CR1. Thus, the Arithmetic modes such as predicate-result make no sense, and neither does Saturation. All of these differences, which require quite a lot of logical reasoning and deduction, help explain why there is an entirely different CR ops Vectorisation Category.

A particularly strange quirk of CR-based Vector Operations is that the Scalar Power ISA CR Register is 32-bits, but actually comprises eight CR Fields, CR0-CR7. With each CR Field being four bits (EQ, LT, GT, SO) this makes up 32 bits, and therefore a CR operand referring to one bit of the CR will be 5 bits in length (BA, BT). However, some instructions refer to a CR Field (CR0-CR7) and consequently these operands (BF, BFA etc) are only 3-bits.

(It helps here to think of the top 3 bits of BA as referring to a CR Field, like BFA does, and the bottom 2 bits of BA referring to EQ/LT/GT/SO within that Field)

With SVP64 extending the number of CR Fields to 128, the number of 32-bit CR Registers extends to 16, in order to hold all 128 CR Fields (8 per CR Register). Then, it gets even more strange, when it comes to Vectorisation, which applies to the CR Field numbers. The hardware-for-loop for Rc=1 for example starts at CR0 for element 0, and moves to CR1 for element 1, and so on. The reason here is quite simple: each element result has to have its own CR Field co-result.

In other words, the element is the 4-bit CR Field, not the bits of the 32-bit CR Register, and not the CR Register (of which there are now 16). All quite logical, but a little mind-bending.

Load/Store

LOAD/STORE is another area that has different needs: this time it is down to limitations in Scalar LD/ST. Vector ISAs have Load/Store modes which simply make no sense in a RISC Scalar ISA: element-stride and unit-stride and the entire concept of a stride itself (a spacing between elements) has no place at all in a Scalar ISA. The problems come when trying to retrofit the concept of "Vector Elements" onto a Scalar ISA. Consequently it required a couple of bits (Modes) in the SVP64 RM Prefix to convey the stride mode, changing the Effective Address computation as a result. Interestingly, worth noting for Hardware designers: it did turn out to be possible to perform pre-multiplication of the D/DS Immediate by the stride amount, making it possible to avoid actually modifying the LD/ST Pipeline itself.

Other areas where LD/ST went quirky: element-width overrides especially when combined with Saturation, given that LD/ST operations have byte, halfword, word, dword and quad variants. The interaction between these widths as part of the actual operation, and the source and destination elwidth overrides, was particularly obtuse and hard to derive: some care and attention is advised, here, when reading the specification, especially on arithmetic loads (lbarx, lharx etc.)

Non-vectorised

The concept of a Vectorised halt (attn) makes no sense. There are never going to be a Vector of global MSRs (Machine Status Register). mtcr on the other hand is a grey area: mtspr is clearly Vectoriseable. Even td and tdi makes a strange type of sense to permit it to be Vectorised, because a sequence of comparisons could be Vectorised. Vectorised System Calls (sc) or tlbie and other Cache or Virtual Nemory Management instructions, these make no sense to Vectorise.

However, it is really quite important to not be tempted to conclude that just because these instructions are un-vectoriseable, the opcode space must be free for reiterpretation and use for other purposes. This would be a serious mistake because a future revision of the specification might retire the Scalar instruction, replace it with another. Again this comes down to being quite strict about the rules: only Scalar instructions get Vectorised: there are no actual explicit Vector instructions.

Summary

Where a traditional Vector ISA effectively duplicates the entirety of a Scalar ISA and then adds additional instructions which only make sense in a Vector Context, such as Vector Shuffle, SVP64 goes to considerable lengths to keep strictly to augmentation and embedding of an entire Scalar ISA's instructions into an abstract Vectorisation Context. That abstraction subdivides down into Categories appropriate for the type of operation (Branch, CRs, Memory, Arithmetic), and each Category has its own relevant but ultimately rational quirks.

Abstraction between Prefix and Suffix

In the introduction paragraph, a great fuss was made emphasising that the Prefix is kept separate from the Suffix. The whole idea there is that a Multi-issue Decoder and subsequent pipelines would in no way have "back-propagation" of state that can only be determined far too late. This has been preserved, however there is a hiccup.

Examining the Power ISA 3.1 a 64-bit Prefix was introduced, EXT001. The encoding of the prefix has 6 bits that are dedicated to letting the hardware know what the remainder of the Prefix bits mean: how they are formatted, even without having to examine the Suffix to which they are applied.

SVP64 has such pressure on its 24-bit encoding that it was simply not possible to perform the same trick used by Power ISA 3.1 Prefixing. Therefore, rather unfortunately, it becomes necessary to perform a partial decoding of the v3.0 Suffix before the 24-bit SVP64 RM Fields may be identified. Fortunately this is straightforward, and does not rely on any outside state, and even more fortunately for a Multi-Issue Execution decoder, the length 32/64 is also easy to identify by looking for the EXT001 pattern. Once identified the 32/64 bits may be passed independently to multiple Decoders in parallel.

Predication

Predication is entirely missing from the Power ISA. Adding it would be a costly mistake because it cannot be retrofitted to an ISA without literally duplicating all instructions. Prefixing is about the only sane way to go.

CR Fields as predicate masks could be spread across multiple register file entries, making them costly to read in one hit. Therefore the possibility exists that an instruction element writing to a CR Field could overwrite the Predicate mask CR Vector during the middle of a for-loop.

Clearly this is bad, so don't do it. If there are potential issues they can be avoided by using the crweird instructions to get CR Field bits into an Integer GPR (r3, r10 or r30) and use that GPR as a Predicate mask instead.

Even in Vertical-First Mode, which is a single Scalar instruction executed with "offset" registers (in effect), the rule still applies: don't write to the same register being used as the predicate, it's UNDEFINED behaviour.

Single Predication

So named because there is a Twin Predication concept as well, Single Predication is also unlike other Vector ISAs because it allows zeroing on both the source and destination. This takes some explaining.

In Vector ISAs, there is a Predicate Mask, it applies to the destination only, and there is a choice of actions when a Predicate Mask bit is zero:

  • set the destination element to zero
  • skip that element operation entirely, leaving the destination unmodified

The problem comes if the underlying register file SRAM is say 64-bit wide write granularity but the Vector elements are say 8-bit wide. Some Vector ISAs strongly advocate Zeroing because to leave one single element at a small bitwidth in amongst other elements where the register file does not have the prerequisite access granularity is very expensive, requiring a Read-Modify-Write cycle to preserve the untouched elements. Putting zero into the destination avoids that Read.

This is technically very easy to solve: use a Register File that does in fact have the smallest element-level write-enable granularity. If the elements are 8 bit then allow 8-bit writes!

With that technical issue solved there is nothing in the way of choosing to support both zeroing and non-zeroing (skipping) at the ISA level: SV chooses to further support both on both the source and destination. This can result in the source and destination element indices getting "out-of-sync" even though the Predicate Mask is the same because the behaviour is different when zeros in the Predicate are encountered.

Twin Predication

Twin Predication is an entirely new concept not present in any commercial Vector ISA of the past forty years. To explain how normal Single-predication is applied in a standard Vector ISA:

  • Predication on the source of a LOAD instruction creates something called "Vector Compressed Load" (VCOMPRESS).
  • Predication on the destination of a STORE instruction creates something called "Vector Expanded Store" (VEXPAND).
  • SVP64 allows the two to be put back-to-back: one on source, one on destination.

The above allows a reader familiar with VCOMPRESS and VEXPAND to conceptualise what the effect of Twin Predication is, but it actually goes much further: in any twin-predicated instruction (extsw, fmv) it is possible to apply one predicate to the source register (compressing the source element array) and another completely separate predicate to the destination register, not just on Load/Stores but on arithmetic operations.

No other Vector ISA in the world has this back-to-back capability. All true Vector ISAs have Predicate Masks: it is an absolutely essential characteristic. However none of them have abstracted dual predicates out to the extent where this VCOMPRESS-VEXPAND effect is applicable in general to a wide range of arithmetic instructions, as well as Load/Store.

It is however important to note that not all instructions can be Twin Predicated (2P): some remain only Single Predicated (1P), as is normally found in other Vector ISAs. Arithmetic operations with four registers (3-in, 1-out, VA-Form for example) are Single. The reason is that there just wasn't enough space in the 24-bits of the SVP64 Prefix. Consequently, when using a given instruction, it is necessary to look up in the ISA Tables whether it is 1P or 2P. caveat emptor!

Also worth a special mention: all Load/Store operations are Twin-Predicated. The underlying key to understanding:

  • one Predicate effectively applies to the Array of Memory Addresses,
  • the other Predicate effectively applies to the Array of Memory Data.

CR weird instructions

cr int predication is by far the biggest violator of the SVP64 rules, for good reasons. Transfers between Vectors of CR Fields and Integers for use as predicates is very awkward without them.

Normally, element width overrides allow the element width to be specified as 8, 16, 32 or default (64) bit. With CR weird instructions producing or consuming either 1 bit or 4 bit elements (in effect) some adaptation was required. When this perspective is taken (that results or sources are 1 or 4 bits) the weirdness starts to make sense, because the "elements", such as they are, are still packed sequentially.

From a hardware implementation perspective however they will need special handling as far as Hazard Dependencies are concerned, due to nonconformance (bit-level management)

mv.x (vector permute)

mv.x aka GPR(RT) = GPR(GPR(RA)) is so horrendous in terms of Register Hazard Management that its addition to any Scalar ISA is anathematic. In a Traditional Vector ISA however, where the indices are isolated behind a single Vector Hazard, there is no problem at all. sv.mv.x is also fraught, precisely because it sits on top of a Standard Scalar register paradigm, not a Vector ISA with separate and distinct Vector registers.

To help partly solve this, sv.mv.x would have had to have been made relative:

for i in range(VL):
    GPR(RT+i) = GPR(RT+MIN(GPR(RA+i), VL))

The reason for doing so is that MAXVL or VL may be used to limit the number of Register Hazards that need to be raised to a fixed quantity, at Issue time.

mv.x itself would still have to be added as a Scalar instruction, but the behaviour of sv.mv.x would have to be different from that Scalar version.

Normally, Scalar Instructions have a good justification for being added as Scalar instructions on their own merit. mv.x is the polar opposite, and in the end, the idea was thrown out, and Indexed REMAP added in its place. Indexed REMAP comes with its own quirks, solving the Hazard problem, described in a later section.

Branch-Conditional

branches are a very special exception to the rule that there shall be no deviation from the corresponding Scalar instruction. This because of the tight integration with looping and the application of Boolean Logic manipulation needed for Parallel operations (predicate mask usage). This results in an extremely important observation that scalar identity behaviour is violated: the SV Prefixed variant of branch is not the same operation as the unprefixed 32-bit scalar version.

One key difference is that LR is only updated if certain additional conditions are met, whereas Scalar bclrl for example unconditionally overwrites LR.

Another is that the Vectorised Branch-Conditional instructions are the only ones where there are side-effects on predication when skipping is enabled. This is so as to be able to use CTR to count down masked-out elements.

Well over 500 Vectorised branch instructions exist in SVP64 due to the number of options available: close integration and interaction with the base Scalar Branch was unavoidable in order to create Conditional Branching suitable for parallel 3D / CUDA GPU workloads.

Saturation

The application of Saturation as a retro-fit to a Scalar ISA is challenging. It does help that within the SFFS Compliancy subset there are no Saturated operations at all: they are only added in VSX.

Saturation does not inherently change the instruction itself: it does however come with some fundamental implications, when applied. For example: a Floating-Point operation that would normally raise an exception will no longer do so, instead setting the CR1.SO Flag. Another quirky example: signed operations which produce a negative result will be truncated to zero if Unsigned Saturation is requested.

One very important aspect for implementors is that the operation in effect has to be considered to be performed at infinite precision, followed by saturation detection. In practice this does not actually require infinite precision hardware! Two 8-bit integers being added can only ever overflow into a 9-bit result.

Overall some care and consideration needs to be applied.

Fail-First

Fail-First (both the Load/Store and Data-Dependent variants) is worthy of a special mention in its own right. Where VL is normally forward-looking and may be part of a pre-decode phase in a (simplified) pipelined architecture with no Read-after-Write Hazards, Fail-First changes that because at any point during the execution of the element-level instructions, one of those elements may not only terminate further continuation of the hardware-for-looping but also effect a change of VL:

for i in range(VL):
    result = element_operation(GPR(RA+i), GPR(RB+i))
    if test(result):
        VL = i
        break

This is not exactly a violation of SVP64 Rules, more of a breakage of user expectations, particularly for LD/ST where exceptions would normally be expected to be raised, Fail-First provides for avoidance of those exceptions.

For Hardware implementers, a standard Out-of-Order micro-architecture allows for Cancellation of speculatively-executed elements that extended beyond the Vector Truncation point. In-order systems will have a slightly harder time and may choose to execute one element only at a time, reducing performance as a result.

OE=1

The hardware cost of Sticky Overflow in a parallel environment is immense. The SFFS Compliancy Level is permitted optionally to support XER.SO. Therefore the decision is made to make it mandatory not to support XER.SO. However, CR.SO is supported such that when Rc=1 is set the CR.SO flag will contain only the overflow of the current instruction, rather than being actually "sticky". Hardware Out-of-Order designers will recognise and appreciate that the Hazards are reduced to Read-After-Write (RAW) and that the WAR Hazard is removed.

This is sort-of a quirk and sort-of not, because the option to support XER.SO is already optional from the SFFS Compliancy Level.

Indexed REMAP and CR Field Predication Hazards

Normal Vector ISAs and those Packed SIMD ISAs inspired by them have Vector "Permute" or "Shuffle" instructions. These provide a Vector of indices whereby another Vector is reordered (permuted, shuffled) according to the indices. Register Hazard Managent here is trivial because there are three registers: indices source vector, elements source vector to be shuffled, result vector.

For SVP64 which is based on top of a Scalar Register File paradigm, combined with the hard requirement to respect full Register Hazard Management as if element instructions were actual Scalar instructions, the addition of a Vector permute instruction under these strict conditions would result in a catastrophic reduction in performance, due to having to consider Read-after-Write and Write-after-Read Hazards at the element level.

A little leniency and rule-bending is therefore required.

Rather than add explicit Vector permute instructions, the "Indexing" has been separated out into a REMAP Schedule. When an Indexed REMAP is requested, it is assumed (required, of software) that subsequent instructions intending to use those indices will not attempt to modify the indices. It is Software that must consider them to be read-only.

This simple relaxation of the rules releases Hardware from having the horrendous job of dynamically detecting Write-after-Read Hazards on a huge range of registers.

A similar Hazard problem exists for CR Field Predicates, in Vertical-First Mode. Instructions could modify CR Fields currently being used as Predicate Masks: detecting this is so horrendous for hardware resource utilisation and hardware complexity that, again, the decision is made to relax these constraints and for Software to take that into account.

Floating-Point "Single" becomes "Half"

In several places in the Power ISA there are operations that are on 32-bit quantities in 64-bit registers. The best example is FP which has 64-bit operations (fadd) and 32-bit operations (fadds or FP Add "single"). Element-width overrides it would seem to be unnecessary, under these circunstances.

However, it is not possible for fadds to fit two elements into 64-bit: that breaks the simplicity of SVP64. Bear in mind that the FP32 bits are spread out across a 64 bit register in FP64 format. The solution here was to consider the "s" at the end of each instruction to mean "half of the element's width". Thus, sv.fadds/ew=32 actually stores an FP16 spread out across the 32 bits of an element, in FP32 format, where sv.fadd/ew=32 stores a full FP32 result into the full 32 bits.

Where this breaks down is when attempting to do half-width on BF16 or FP16 operations: there does not exist a BF8 or an IEE754 FP8 format, so these should be avoided.

Vertical-First and Subvectors

Documented in the setvl page, Vertical-First goes through elements second instructions first and requires an explicit svstep instruction to move to the next element, (whereas Horizontal-First loops through elements in full first before moving on to the next instruction): Subvectors are considered "elements" in Vertical-First Mode.

This is conceptually quite easy to keep in mind that a Vertical-First instruction does one element at a time, and when SUBVL is set, that "element" in essence becomes a vec2/3/4.

Swizzle and Pack/Unpack

These are both so weird it's best to just read the pages in full and pay attention: mv.swizzle and mv.vec. Swizzle Moves only engage with vec2/3/4, reordering the copying of the sub-vector elements (including allowing repeats and skips) based on an immediate supplied by the instruction. The fun comes when Pack/Unpack are enabled, and it is really important to be aware how the Arrays of vec2/3/4 become re-ordered and swizzled at the same time.

Pack/Unpack applies to mv.vec as well however the uniform relationship and the fact that the source and destination subvector length must be the same (vec2/3/4) makes things slightly easier to understand. The main thing to keep in mind about Pack/Unpack is that it engages a swap of the ordering of the VL-SUBVL nested for-loops, in exactly the same way that Matrix REMAP can do. When Pack or Unpack is enabled it is the SUBVL for-loop thar becomes outermost.