[[!toc]]
SVP64 is designed around fundamental and inviolate RISC principles.
-This gives a uniformity and regularity to the ISA which was why RISC
-as a concept became popular. It is just that nobody has ever considered
-applying the RISC concept to a *Vector* ISA before.
+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 makes vector elements 100% synonymous
- with scalar instructions (the suffix)
+3. A hardware-level for-loop (the prefix) makes vector elements
+ 100% synonymous with scalar instructions (the suffix)
+4. Exactly as with Scalar RISC ISAs, the uniformity does produce
+ "holes" in the encoding or some strange combinations.
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.
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 to a completely unrelated operation
+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
**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 can be done: Saturation as well, and a new innovative
+modes for Vector ISAs: data-dependent fail-first.
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
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.
+instructions (as explained in a separate section) although
+a case could be made for the perspective that they are identical,
+but the defaults for new parameters in the Scalar case makes branch
+identical to Power ISA v3.1 Scalar branches.
+
The
RM bits can be used for other purposes because the Arithmetic modes
make no sense at all for a Branch.
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.
+a "co-result" that goes into CR0 or CR1. Thus, Saturation makes no sense.
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.
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.
+and attention is advised, here, when reading the specification,
+especially on arithmetic loads (lbarx, lharx etc.)
**Non-vectorised**
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
+just because these instructions are un-vectoriseable, the Prefix 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.
+might *retire* the Scalar instruction, and, worse, 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.
and each Category has its own relevant but
ultimately rational quirks.
-# Single Predication
+# 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
is the same because the behaviour is different when zeros in the
Predicate are encountered.
-# Twin Predication
+## 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
REMAP added in its place. Indexed REMAP comes with its own quirks,
solving the Hazard problem, described in a later section.
+# REMAP and other reordering
+
+There are several places in Simple-V which apply some sort of reordering
+schedule to elements. srcstep and dststep do not themselves reorder:
+they continue to march in sequence (VL-1 downto 0 in the case of reverse-gear)
+
+It is perfectly legal to apply Parallel-Reduction on top of any type
+of REMAP, for example, and it is possible to apply Pack/Unpack on a
+REMAP as well.
+
+The order of application of REMAP combined with Parallel-Reduction
+should be logically obvious: REMAP has to come first because otherwise
+how can the Parallel-Reduction perform a tree-walk?
+
+Pack/Unpack on the other hand is best implemented as applying first,
+because it is applied
+as the inversion of the for-loops which generate the steps and substeps.
+REMAP then applies to the src/dst-step indices (never to the subvl
+step indices: that is SWIZZLE's job).
+
+It's all perfectly logical, just a lot going on.
+
# Branch-Conditional
[[sv/branches]] are a very special exception to the rule that there
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
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.
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 circumstances.
+
+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 IEEE754 FP8
+format, so these (`sv.fadds/ew=8`) should be avoided.
+
+# Word frequently becomes "half"
+
+Again, related to "Single" becoming "half of element width", unless there
+are compelling reasons the same trick applies to Scalar GPR operations.
+With the pseudocode being "XLEN//2" then of course if XLEN=8 the operation
+becomes a 4-bit one.
+
+Similarly byte operations which use "XLEN//8" when XLEN=8 actually become
+single-bit operations, which is very useful with `sv.extsb/w=8`
+for example. This instruction copies the LSB of each byte in a sequence of bytes,
+and expands it to all 8 bits in each result byte.
+
+# Vertical-First and Subvectors
+
+Documented in the [[sv/setvl]] page, Vertical-First goes through
+elements second instructions first and requires an explicit
+[[sv/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: [[sv/mv.swizzle]] and [[sv/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 started out as
+[[sv/mv.vec]] but became its own distinct Mode over time.
+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
+that becomes outermost. A bit of thought shows that this is
+a 2D "Transpose" where Dimension X is VL and Dimension Y is SUBVL.
+However *both* source *and* destination may be independently
+"Transposed", which makes no sense at all until the fact that
+Swizzle can have a *different SUBVL* is taken into account.
+
+Basically Pack/Unpack covers everything that VSX `vpkpx` and
+other ops can do, and then some: Saturation included, for arithmetic ops.
+
+# LD/ST with zero-immediate vs mapreduce mode
+
+LD/ST operations with a zero immediate effectively means that on a
+Vector operation the element index to offset the memory location is
+multiplied by zero. Thus, a sequence of LD operations will load from
+the exact same address, and likewise STs to the exact same address.
+
+Ordinarily this would make absolutely no sense whatsoever, except
+that Power ISA has cache-inhibited LD/STs (Power ISA v.1, Book III,
+1.6.1, p1033), for accessing memory-mapped
+peripherals and other crucial uses. Thus, *despite not being a mapreduce mode*,
+zero-immediates cause multiple hits on the same element.
+
+Mapreduce mode is not actually mapreduce at all: it is
+a relaxation of the normal rule where if the destination is a Scalar the
+Vector for-looping is not terminated on first write to the destination.
+Instead, the developer is expected to exploit the strict Program Order,
+make one of the sources the same as that Scalar destination, effectively
+making that Scalar register an "Accumulator", thus creating the *appearance*
+(effect) of Simple-V having a mapreduce capability, when in fact it is
+more of an artefact.
+
+LD/ST zero-immediate has similar quirky overwriting as the "mapreduce"
+mode, but actually requires the registers to be Vectors. It is simply
+a mathematical artefact of multiplying by zero, which happens to be
+useful for cache-inhibited operations.
+
+# Limited space in LD/ST Mode
+
+As pointed out in the [[sv/ldst]] page there is limited space in only
+5 mode bits to fully express all potential modes of operation.
+
+* LD/ST Immediate has no individual control over src/dest zeroing,
+ whereas LD/ST Indexed does.
+* Post-Increment is not possible with Saturation or Data-Dependent Fail-First
+* Element-Strided LD/ST Indexed is not possible with Data-Dependent Fail-First.
+
+Also, the LD/ST Indexed Mode can be element-strided (RB as
+a Scalar, times
+the element index), or, if that is not enough,
+although potentially costly it is possible to
+use `svstep` to compute a Vector RB sequence of
+Indices, then activate either `sz` or `dz` as required, as a workaround
+for LDST Immediate only having `zz`.
+
+Simple-V is powerful but it cannot do everything! There is just not
+enough space and so some compromises had to be made.
+
+# sv.mtcr on entire 64-bit Condition Register
+
+Normally, CR operations are either bit-based (where the element numbering actually
+applies to the CR Field) or field-based in which case the elements are still
+fields. The `sv.mtcr` and other instructions are actually full 64-bit Condition
+*Register* operations and are therefore qualified as Normal/Arithmetic not
+CRops.
+
+This is to save on both Vector Length (VL of 16 is sufficient) as well as
+complexity in the Hazard Management when context-switching CR fields, as the
+entire batch of 128 CR Fields may be transferred to 8 GPRs with a VL of 16
+and elwidth overriding of 32. Truncation is sufficent, dropping the top 32 bits
+of the Condition Register(s) which are always zero anyway.
+
+# Separate Scalar and Vector Condition Register files
+
+As explained in the introduction [[sv/svp64]] and [[sv/cr_ops]]
+Scalar Power ISA lacks "Conditional Execution" present in ARM
+Scalar ISA of several decades. When Vectorised the fact that
+Rc=1 Vector results can immediately be used as a Predicate Mask
+back into the following instruction can result in large latency
+unless "Vector Chaining" is used in the Micro-Architecture.
+
+But that aside is not the main problem faced by the introduction
+of Simple-V to the Power ISA: it's that the existing implementations
+(IBM) don't have "Conditional Execution" and to add it to their
+existing designs would be too disruptive a first step.
+
+A compromise is to wipe blank certain entries in the Register Dependency
+Matrices by prohibiting some operations involving the two groups
+of CR Fields: those that fall into the existing Scalar 32-bit CR
+(fields CR0-CR7) and those that fall into the newly-introduced
+CR Fields, CR8-CR127.
+
+This will drive compiler writers nuts, and give assembler writers headaches,
+but it gives IBM the opportunity to implement SVP64 without massive
+disruption. They can add an entirely new Vector CR register file,
+new pipelines etc safe in the knowledge that existing Scalar HDL
+needs no modification.
projects / libreriscv.git / blobdiff