RFC ls009 Simple-V REMAP Subsystem
- Funded by NLnet under the Privacy and Enhanced Trust Programme, EU Horizon2020 Grant 825310, and NGI0 Entrust No 101069594
- https://libre-soc.org/openpower/sv/rfc/ls009/
- https://bugs.libre-soc.org/show_bug.cgi?id=1060
- https://git.openpower.foundation/isa/PowerISA/issues/124
Severity: Major
Status: New
Date: 26 Mar 2023. v2 TODO
Target: v3.2B
Source: v3.0B
Books and Section affected:
Book I, new Zero-Overhead-Loop Chapter.
Appendix E Power ISA sorted by opcode
Appendix F Power ISA sorted by version
Appendix G Power ISA sorted by Compliancy Subset
Appendix H Power ISA sorted by mnemonic
Summary
svremap - Re-Mapping of Register Element Offsets
svindex - General-purpose setting of SHAPEs to be re-mapped
svshape - Hardware-level setting of SHAPEs for element re-mapping
svshape2 - Hardware-level setting of SHAPEs for element re-mapping (v2)
Submitter: Luke Leighton (Libre-SOC)
Requester: Libre-SOC
Impact on processor:
Addition of four new "Zero-Overhead-Loop-Control" DSP-style Vector-style
Management Instructions which provide advanced features such as Matrix
FFT DCT Hardware-Assist Schedules and general-purpose Index reordering.
Impact on software:
Requires support for new instructions in assembler, debuggers,
and related tools.
Keywords:
Cray Supercomputing, Vectorization, Zero-Overhead-Loop-Control (ZOLC),
Scalable Vectors, Multi-Issue Out-of-Order, Sequential Programming Model,
Digital Signal Processing (DSP)
Motivation
These REMAP Management instructions provide state-of-the-art advanced capabilities to dramatically decrease instruction count and power reduction whilst retaining unprecedented general-purpose capability and a standard Sequential Execution Model.
Notes and Observations:
- Although costly the alternatives in SIMD-paradigm software result in huge algorithmic complexity and associated power consumption increases. Loop-unrolling compilers are prevalent as is thousands to tens of thousands of instructions.
- Core inner kernels of Matrix DCT DFT FFT NTT are dramatically reduced to a handful of instructions.
- No REMAP instructions with the exception of Indexed rely on registers for the establishment of REMAP capability.
- Future EXT1xx variants and SVP64/VSX will dramatically expand the power and flexibility of REMAP.
- The Simple-V Compliancy Subsets make REMAP optional except in the Advanced Levels. Like v3.1 MMA it is not necessary for all hardware to implement REMAP.
Changes
Add the following entries to:
- the Appendices of SV Book
- Instructions of SV Book as a new Section
- SVI, SVM, SVM2, SVRM Form of Book I Section 1.6.1.6 and 1.6.2
\newpage{}
Rationale and background
In certain common algorithms there are clear patterns of behaviour which typically require inline loop-unrolled instructions comprising interleaved permute and arithmetic operations. REMAP was invented to split the permuting from the arithmetic, and to allow the Indexing to be done as a hardware Schedule.
A normal Vector Add:
for i in range(VL):
GPR[RT+i] <= GPR[RA+i] + GPR[RB+i];
A Hardware-assisted REMAP Vector Add:
for i in range(VL):
GPR[RT+remap1(i)] <= GPR[RA+remap2(i)] + GPR[RB+remap3(i)];
The result is a huge saving on register file accesses (no need to calculate Indices then use Permutation instructions), instruction count (Matrix Multiply up to 127 FMACs is 3 instructions), and programmer sanity.
Basic principle
The following illustrates why REMAP was added.
- normal vector element read/write of operands would be sequential (0 1 2 3 ....)
- this is not appropriate for (e.g.) Matrix multiply which requires accessing elements in alternative sequences (0 3 6 1 4 7 ...)
- normal Vector ISAs use either Indexed-MV or Indexed-LD/ST to "cope" with this. both are expensive (copy large vectors, spill through memory) and very few Packed SIMD ISAs cope with non-Power-2 (Duplicate-data inline-loop-unrolling is the costly solution)
- REMAP redefines the order of access according to set (Deterministic) "Schedules".
- Matrix Schedules are not at all restricted to power-of-two boundaries making it unnecessary to have for example specialised 3x4 transpose instructions of other Vector ISAs.
- DCT and FFT REMAP are RADIX-2 limited but this is the case in existing Packed/Predicated SIMD ISAs anyway (and Bluestein Convolution is typically deployed to solve that).
Only the most commonly-used algorithms in computer science have REMAP
support, due to the high cost in both the ISA and in hardware. For
arbitrary remapping the Indexed
REMAP may be used.
REMAP types
This section summarises the motivation for each REMAP Schedule and briefly goes over their characteristics and limitations. Further details on the Deterministic Precise-Interruptible algorithms used in these Schedules is found in the appendix.
Matrix (1D/2D/3D shaping)
Matrix Multiplication is a huge part of High-Performance Compute, and 3D. In many PackedSIMD as well as Scalable Vector ISAs, non-power-of-two Matrix sizes are a serious challenge. PackedSIMD ISAs, in order to cope with for example 3x4 Matrices, recommend rolling data-repetition and loop-unrolling. Aside from the cost of the load on the L1 I-Cache, the trick only works if one of the dimensions X or Y are power-two. Prime Numbers (5x7, 3x5) become deeply problematic to unroll.
Even traditional Scalable Vector ISAs have issues with Matrices, often having to perform data Transpose by pushing out through Memory and back (costly), or computing Transposition Indices (costly) then copying to another Vector (costly).
Matrix REMAP was thus designed to solve these issues by providing Hardware Assisted "Schedules" that can view what would otherwise be limited to a strictly linear Vector as instead being 2D (even 3D) in-place reordered. With both Transposition and non-power-two being supported the issues faced by other ISAs are mitigated.
Limitations of Matrix REMAP are that the Vector Length (VL) is currently restricted to 127: up to 127 FMAs (or other operation) may be performed in total. Also given that it is in-registers only at present some care has to be taken on regfile resource utilisation. However it is perfectly possible to utilise Matrix REMAP to perform the three inner-most "kernel" loops of the usual 6-level "Tiled" large Matrix Multiply, without the usual difficulties associated with SIMD.
Also the svshape
instruction only provides access to part of the
Matrix REMAP capability. Rotation and mirroring need to be done by
programming the SVSHAPE SPRs directly, which can take a lot more
instructions. Future versions of SVP64 will include EXT1xx prefixed
variants (psvshape
) which provide more comprehensive capacity and
mitigate the need to write direct to the SVSHAPE SPRs.
FFT/DCT Triple Loop
DCT and FFT are some of the most astonishingly used algorithms in Computer Science. Radar, Audio, Video, R.F. Baseband and dozens more. At least two DSPs, TMS320 and Hexagon, have VLIW instructions specially tailored to FFT.
An in-depth analysis showed that it is possible to do in-place in-register DCT and FFT as long as twin-result "butterfly" instructions are provided. These can be found in the svfparith page if performing IEEE754 FP transforms. (For fixed-point transforms, equivalent 3-in 2-out integer operations would be required, currently under development). These "butterfly" instructions avoid the need for a temporary register because the two array positions being overwritten will be "in-flight" in any In-Order or Out-of-Order micro-architecture.
DCT and FFT Schedules are currently limited to RADIX2 sizes and do not accept predicate masks. Given that it is common to perform recursive convolutions combining smaller Power-2 DCT/FFT to create larger DCT/FFTs in practice the RADIX2 limit is not a problem. A Bluestein convolution to compute arbitrary length is demonstrated by Project Nayuki
Indexed
The purpose of Indexing is to provide a generalised version of
Vector ISA "Permute" instructions, such as VSX vperm
. The
Indexing is abstracted out and may be applied to much more
than an element move/copy, and is not limited for example
to the number of bytes that can fit into a VSX register.
Indexing may be applied to LD/ST (even on Indexed LD/ST
instructions such as sv.lbzx
), arithmetic operations,
extsw: there is no artificial limit.
The original motivation for Indexed REMAP was to mitigate the need to add
an expensive mv.x
to the Scalar ISA, which was highly likely to be rejected as
a stand-alone instruction
(GPR(RT) <- GPR(GPR(RA))
). Usually a Vector ISA would add a non-conflicting
variant (as in VSX vperm
) but it is common to need to permute by source,
with the risk of conflict, that has to be resolved, for example, in AVX-512
with conflictd
.
Indexed REMAP is the "Get out of Jail Free" card which (for a price) allows any general-purpose arbitrary Element Permutation not covered by the other Hardware Schedules.
Parallel Reduction
Vector Reduce Mode issues a deterministic tree-reduction schedule to the underlying micro-architecture. Like Scalar reduction, the "Scalar Base" (Power ISA v3.0B) operation is leveraged, unmodified, to give the appearance and effect of Reduction. Parallel Reduction is not limited to Power-of-two but is limited as usual by the total number of element operations (127) as well as available register file size.
Parallel Reduction is normally done explicitly as a loop-unrolled operation, taking up a significant number of instructions. With REMAP it is just three instructions: two for setup and one Scalar Base.
Parallel Prefix Sum
This is a work-efficient Parallel Schedule that for example produces Trangular
or Factorial number sequences. Half of the Prefix Sum Schedule is near-identical
to Parallel Reduction. Whilst the Arithmetic mapreduce Mode (/mr
) may achieve the same
end-result, implementations may only implement Mapreduce in serial form (or give
the appearance to Programmers of the same). The Parallel Prefix Schedule is
required to be implemented in such a way that its Deterministic Schedule may be
parallelised. Like the Reduction Schedule it is 100% Deterministic and consequently
may be used with non-commutative operations.
Parallel Reduction combined with Parallel Prefix Sum can be combined to help perform AI non-linear interpolation in around twelve to fifteen instructions.
\newpage{}
Add the following to SV Book
REMAP
- https://bugs.libre-soc.org/show_bug.cgi?id=143 matrix multiply
- https://bugs.libre-soc.org/show_bug.cgi?id=867 add svindex
- https://bugs.libre-soc.org/show_bug.cgi?id=885 svindex in simulator
- https://bugs.libre-soc.org/show_bug.cgi?id=911 offset svshape option
- https://bugs.libre-soc.org/show_bug.cgi?id=864 parallel reduction
- https://bugs.libre-soc.org/show_bug.cgi?id=930 DCT/FFT "strides"
- https://bugs.libre-soc.org/show_bug.cgi?id=1155 bigmul (normal and carry-save)
- see appendix for examples and usage
- see propagation for a future way to apply REMAP
- discussion
REMAP is an advanced form of Vector "Structure Packing" that provides
hardware-level support for commonly-used nested loop patterns that would
otherwise require full inline loop unrolling. For more general reordering
an Indexed REMAP mode is available (a RISC-paradigm
abstracted analog to xxperm
).
REMAP allows the usual sequential vector loop 0..VL-1
to be "reshaped"
(re-mapped) from a linear form to a 2D or 3D transposed form, or "offset"
to permit arbitrary access to elements, independently on each
Vector src or dest register. Up to four separate independent REMAPs may be applied
to the registers of any instruction.
A normal Vector Add (no Element-width Overrides):
for i in range(VL):
GPR[RT+i] <= GPR[RA+i] + GPR[RB+i];
A Hardware-assisted REMAP Vector Add:
for i in range(VL):
GPR[RT+remap1(i)] <= GPR[RA+remap2(i)] + GPR[RB+remap3(i)];
Aside from Indexed REMAP this is entirely Hardware-accelerated reordering and consequently not costly in terms of register access for the Indices. It will however place a burden on Multi-Issue systems but no more than if the equivalent Scalar instructions were explicitly loop-unrolled without SVP64, and some advanced implementations may even find the Deterministic nature of the Scheduling to be easier on resources.
Hardware note: in its general form, REMAP is quite expensive to set up, and on some implementations may introduce latency, so should realistically be used only where it is worthwhile. Given that even with latency the fact that up to 127 operations can be Deterministically issued (from a single instruction) it should be clear that REMAP should not be dismissed for possible latency alone. Commonly-used patterns such as Matrix Multiply, DCT and FFT have helper instruction options which make REMAP easier to use.
There are five types of REMAP:
- Matrix, also known as 2D and 3D reshaping, can perform in-place Matrix transpose and rotate. The Shapes are set up for an "Outer Product" Matrix Multiply (a future variant may introduce Inner Product).
- FFT/DCT, with full triple-loop in-place support: limited to Power-2 RADIX
- Indexing, for any general-purpose reordering, also includes limited 2D reshaping as well as Element "offsetting".
- Parallel Reduction, for scheduling a sequence of operations in a Deterministic fashion, in a way that may be parallelised, to reduce a Vector down to a single value.
- Parallel Prefix Sum, implemented as a work-efficient Schedule, has several key Computer Science uses. Again Prefix Sum is 100% Deterministic.
Best implemented on top of a Multi-Issue Out-of-Order Micro-architecture, REMAP Schedules are 100% Deterministic including Indexing and are designed to be incorporated in between the Decode and Issue phases, directly into Register Hazard Management.
As long as the SVSHAPE SPRs
are not written to directly, Hardware may treat REMAP as 100%
Deterministic: all REMAP Management instructions take static
operands (no dynamic register operands)
with the exception of Indexed Mode, and even then
Architectural State is permitted to assume that the Indices
are cacheable from the point at which the svindex
instruction
is executed.
Further details on the Deterministic Precise-Interruptible algorithms used in these Schedules is found in the appendix.
Future specification note: future versions of the REMAP Management instructions will extend to EXT1xx Prefixed variants. This will overcome some of the limitations present in the 32-bit variants of the REMAP Management instructions that at present require direct writing to SVSHAPE0-3 SPRs. Additional REMAP Modes may also be introduced at that time.
Determining Register Hazards (hphint)
For high-performance (Multi-Issue, Out-of-Order) systems it is critical to be able to statically determine the extent of Vectors in order to allocate pre-emptive Hazard protection. The next task is to eliminate masked-out elements using predicate bits, freeing up the associated Hazards.
For non-REMAP situations VL
is sufficient to ascertain early
Hazard coverage, and with SVSTATE being a high priority cached
quantity at the same level of MSR and PC this is not a problem.
The problems come when REMAP is enabled. Indexed REMAP must instead
use MAXVL
as the earliest (simplest)
batch-level Hazard Reservation indicator (after taking element-width
overriding on the Index source into consideration),
but Matrix, FFT and Parallel Reduction must all use completely different
schemes. The reason is that VL is used to step through the total
number of operations, not the number of registers.
The "Saving Grace" is that all of the REMAP Schedules are 100% Deterministic.
Advance-notice Parallel computation and subsequent cacheing of all of these complex Deterministic REMAP Schedules is strongly recommended, thus allowing clear and precise multi-issue batched Hazard coverage to be deployed, even for Indexed Mode. This is only possible for Indexed due to the strict guidelines given to Programmers.
In short, there exists solutions to the problem of Hazard Management, with varying degrees of refinement possible at correspondingly increasing levels of complexity in hardware.
A reminder: when Rc=1 each result register (element) has an associated co-result CR Field (one per result element). Thus above when determining the Write-Hazards for result registers the corresponding Write-Hazards for the corresponding associated co-result CR Field must not be forgotten, including when Predication is used.
Horizontal-Parallelism Hint
To help further in reducing Hazards,
SVSTATE.hphint
is an indicator to hardware of how many elements are 100%
fully independent. Hardware is permitted to assume that groups of elements
up to hphint
in size need not have Register (or Memory) Hazards created
between them, including when hphint > VL
, which greatly aids simplification of
Multi-Issue implementations.
If care is not taken in setting hphint
correctly it may wreak havoc.
For example Matrix Outer Product relies on the innermost loop computations
being independent. If hphint
is set to greater than the Outer Product
depth then data corruption is guaranteed to occur.
Likewise on FFTs it is assumed that each layer of the RADIX2 triple-loop
is independent, but that there is strict inter-layer Register Hazards.
Therefore if hphint
is set to greater than the RADIX2 width of the FFT,
data corruption is guaranteed.
Thus the key message is that setting hphint
requires in-depth knowledge
of the REMAP Algorithm Schedules, given in the Appendix.
REMAP area of SVSTATE SPR
The following bits of the SVSTATE SPR are used for REMAP:
|32:33|34:35|36:37|38:39|40:41| 42:46 | 62 |
| -- | -- | -- | -- | -- | ----- | ------ |
|mi0 |mi1 |mi2 |mo0 |mo1 | SVme | RMpst |
mi0-2 and mo0-1 each select SVSHAPE0-3 to apply to a given register. mi0-2 apply to RA, RB, RC respectively, as input registers, and likewise mo0-1 apply to output registers (RT/FRT, RS/FRS) respectively. SVme is 5 bits (one for each of mi0-2/mo0-1) and indicates whether the SVSHAPE is actively applied or not, and if so, to which registers.
- bit 4 of SVme indicates if mi0 is applied to source RA / FRA / BA / BFA / RT / FRT
- bit 3 of SVme indicates if mi1 is applied to source RB / FRB / BB
- bit 2 of SVme indicates if mi2 is applied to source RC / FRC / BC
- bit 1 of SVme indicates if mo0 is applied to result RT / FRT / BT / BF
- bit 0 of SVme indicates if mo1 is applied to result Effective Address / FRS / RS (LD/ST-with-update has an implicit 2nd write register, RA)
The "persistence" bit if set will result in all Active REMAPs being applied indefinitely.
\newpage{}
svremap instruction
SVRM-Form:
0 | 6 | 11 | 13 | 15 | 17 | 19 | 21 | 22:25 | 26:31 |
---|---|---|---|---|---|---|---|---|---|
PO | SVme | mi0 | mi1 | mi2 | mo0 | mo1 | pst | rsvd | XO |
- svremap SVme,mi0,mi1,mi2,mo0,mo1,pst
Pseudo-code:
# registers RA RB RC RT EA/FRS SVSHAPE0-3 indices
SVSTATE[32:33] <- mi0
SVSTATE[34:35] <- mi1
SVSTATE[36:37] <- mi2
SVSTATE[38:39] <- mo0
SVSTATE[40:41] <- mo1
# enable bit for RA RB RC RT EA/FRS
SVSTATE[42:46] <- SVme
# persistence bit (applies to more than one instruction)
SVSTATE[62] <- pst
Special Registers Altered:
SVSTATE
svremap
establishes the connection between registers and SVSHAPE SPRs.
The bitmask SVme
determines which registers have a REMAP applied, and mi0-mo1
determine which shape is applied to an activated register. the pst
bit if
cleared indicated that the REMAP operation shall only apply to the immediately-following
instruction. If set then REMAP remains permanently enabled until such time as it is
explicitly disabled, either by setvl
setting a new MAXVL, or with another
svremap
instruction. svindex
and svshape2
are also capable of setting or
clearing persistence, as well as partially covering a subset of the capability of
svremap
to set register-to-SVSHAPE relationships.
Programmer's Note: applying non-persistent svremap
to an instruction that has
no REMAP enabled or is a Scalar operation will obviously have no effect but
the bits 32 to 46 will at least have been set in SVSTATE. This may prove useful
when using svindex
or svshape2
.
Hardware Architectural Note: when persistence is not set it is critically important
to treat the svremap
and the immediately-following SVP64 instruction as an
indivisible fused operation.
No state is stored in the SVSTATE SPR in order to allow continuation should an
Interrupt occur between the two instructions. Thus, Interrupts must be prohibited
from occurring or other workaround deployed. When persistence is set this issue
is moot.
It is critical to note that if persistence is clear then svremap
is the only way
to activate REMAP on any given (following) instruction. If persistence is set however then
all SVP64 instructions go through REMAP as long as SVme
is non-zero.
\newpage{}
SHAPE Remapping SPRs
There are four "shape" SPRs, SHAPE0-3, 32-bits in each,
which have the same format. It is possible to write directly to these
SPRs but it is recommended to use the Management instructions
svshape
, svshape2
or svindex
.
When SHAPE is set entirely to zeros, remapping is disabled: the register's elements are a linear (1D) vector.
0:5 | 6:11 | 12:17 | 18:20 | 21:23 | 24:27 | 28:29 | 30:31 | Mode |
---|---|---|---|---|---|---|---|---|
xdimsz | ydimsz | zdimsz | permute | invxyz | offset | skip | 0b00 | Matrix |
xdimsz | ydimsz | SVGPR | 11/ | sk1/invxy | offset | elwidth | 0b00 | Indexed |
xdimsz | mode | zdimsz | submode2 | invxyz | offset | submode | 0b01 | DCT/FFT |
rsvd | rsvd | xdimsz | rsvd | invxyz | offset | submode | 0b10 | Red/Sum |
0b11 | rsvd |
mode
(combined with permute
when mode=0b00
) sets different
general behaviours: straight matrix multiply, FFT, DCT, Reduction or Prefix-Sum.
- mode=0b00 with
permute != 0b110/0b111
sets straight Matrix Mode - mode=0b00 with
permute = 0b110/0b111
sets Indexed Mode - mode=0b01 sets "FFT/DCT" mode and activates submodes
- mode=0b10 sets "Parallel Reduction or Prefix-Sum" Schedules.
Architectural Resource Allocation note: the four SVSHAPE SPRs are best
allocated sequentially and contiguously in order that sv.mtspr
may
be used to manipulate (save/restore) them.
This is safe to do directly as long as SVSTATE.SVme=0
Parallel Reduction / Prefix-Sum Mode
Creates the Schedules for Parallel Tree Reduction and Prefix-Sum
- submode=0b00 selects the left operand index for Reduction
- submode=0b01 selects the right operand index for Reduction
- submode=0b10 selects the left operand index for Prefix-Sum
submode=0b11 selects the right operand index for Prefix-Sum
When bit 0 of
invxyz
is set, the order of the indices in the inner for-loop are reversed. This has the side-effect of placing the final reduced result in the last-predicated element. It also has the indirect side-effect of swapping the source registers: Left-operand index numbers will always exceed Right-operand indices. When clear, the reduced result will be in the first-predicated element, and Left-operand indices will always be less than Right-operand ones.- When bit 1 of
invxyz
is set, the order of the outer loop step is inverted: stepping begins at the nearest power-of two to half of the vector length and reduces by half each time. When clear the step will begin at 2 and double on each inner loop.
Parallel Prefix Sum
This is a work-efficient Parallel Schedule that for example produces Trangular
or Factorial number sequences. Half of the Prefix Sum Schedule is near-identical
to Parallel Reduction. Whilst the Arithmetic mapreduce Mode (/mr
) may achieve the same
end-result, implementations may only implement Mapreduce in serial form (or give
the appearance to Programmers of the same). The Parallel Prefix Schedule is
required to be implemented in such a way that its Deterministic Schedule may be
parallelised. Like the Reduction Schedule it is 100% Deterministic and consequently
may be used with non-commutative operations.
The Schedule Algorithm may be found in the appendix
Parallel Reduction
Vector Reduce Mode issues a deterministic tree-reduction schedule to the underlying micro-architecture. Like Scalar reduction, the "Scalar Base" (Power ISA v3.0B) operation is leveraged, unmodified, to give the appearance and effect of Reduction. Parallel Reduction is not limited to Power-of-two but is limited as usual by the total number of element operations (127) as well as available register file size.
In Horizontal-First Mode, Vector-result reduction requires the destination to be a Vector, which will be used to store intermediary results, in order to achieve a correct final result.
Given that the tree-reduction schedule is deterministic, Interrupts and exceptions can therefore also be precise. The final result will be in the first non-predicate-masked-out destination element, but due again to the deterministic schedule programmers may find uses for the intermediate results, even for non-commutative Defined Word-instruction operations. Additionally, because the intermediate results are always written out it is possible to service Precise Interrupts without affecting latency (a common limitation of Vector ISAs implementing explicit Parallel Reduction instructions, because their Architectural State cannot hold the partial results).
When Rc=1 a corresponding Vector of co-resultant CRs is also created. No special action is taken: the result and its CR Field are stored "as usual" exactly as all other SVP64 Rc=1 operations.
Note that the Schedule only makes sense on top of certain instructions:
X-Form with a Register Profile of RT,RA,RB
is fine because two sources
and the destination are all the same type. Like Scalar
Reduction, nothing is prohibited:
the results of execution on an unsuitable instruction may simply
not make sense. With care, even 3-input instructions (madd, fmadd, ternlogi)
may be used, and whilst it is down to the Programmer to walk through the
process the Programmer can be confident that the Parallel-Reduction is
guaranteed 100% Deterministic.
Critical to note regarding use of Parallel-Reduction REMAP is that,
exactly as with all REMAP Modes, the svshape
instruction requests
a certain Vector Length (number of elements to reduce) and then
sets VL and MAXVL at the number of operations needed to be
carried out. Thus, equally as importantly, like Matrix REMAP
the total number of operations
is restricted to 127. Any Parallel-Reduction requiring more operations
will need to be done manually in batches (hierarchical
recursive Reduction).
Also important to note is that the Deterministic Schedule is arranged so that some implementations may parallelise it (as long as doing so respects Program Order and Register Hazards). Performance (speed) of any given implementation is neither strictly defined or guaranteed. As with the Vulkan(tm) Specification, strict compliance is paramount whilst performance is at the discretion of Implementors.
Parallel-Reduction with Predication
To avoid breaking the strict RISC-paradigm, keeping the Issue-Schedule completely separate from the actual element-level (scalar) operations, Move operations are not included in the Schedule. This means that the Schedule leaves the final (scalar) result in the first-non-masked element of the Vector used. With the predicate mask being dynamic (but deterministic) at a superficial glance it seems this result could be anywhere.
If that result is needed to be moved to a (single) scalar register
then a follow-up sv.mv/sm=predicate rt, *ra
instruction will be
needed to get it, where the predicate is the exact same predicate used
in the prior Parallel-Reduction instruction.
- If there was only a single bit in the predicate then the result will not have moved or been altered from the source vector prior to the Reduction
- If there was more than one bit the result will be in the first element with a predicate bit set.
In either case the result is in the element with the first bit set in the predicate mask. Thus, no move/copy within the Reduction itself was needed.
Programmer's Note: For some hardware implementations the vector-to-scalar copy may be a slow operation, as may the Predicated Parallel Reduction itself. It may be better to perform a pre-copy of the values, compressing them (VREDUCE-style) into a contiguous block, which will guarantee that the result goes into the very first element of the destination vector, in which case clearly no follow-up predicated vector-to-scalar MV operation is needed. A VREDUCE effect is achieved by setting just a source predicate mask on Twin-Predicated operations.
Usage conditions
The simplest usage is to perform an overwrite, specifying all three register operands the same.
svshape parallelreduce, 6
sv.add *8, *8, *8
The Reduction Schedule will issue the Parallel Tree Reduction spanning registers 8 through 13, by adjusting the offsets to RT, RA and RB as necessary (see "Parallel Reduction algorithm" in a later section).
A non-overwrite is possible as well but just as with the overwrite version, only those destination elements necessary for storing intermediary computations will be written to: the remaining elements will not be overwritten and will not be zero'd.
svshape parallelreduce, 6
sv.add *0, *8, *8
However it is critical to note that if the source and destination are not the same then the trick of using a follow-up vector-scalar MV will not work.
Sub-Vector Horizontal Reduction
To achieve Sub-Vector Horizontal Reduction, Pack/Unpack should be enabled,
which will turn the Schedule around such that issuing of the Scalar
Defined Word-instructions is done with SUBVL looping as the inner loop not the
outer loop. Rc=1 with Sub-Vectors (SUBVL=2,3,4) is UNDEFINED
behaviour.
Programmer's Note: Overwrite Parallel Reduction with Sub-Vectors will clearly result in data corruption. It may be best to perform a Pack/Unpack Transposing copy of the data first
FFT/DCT mode
submode2=0 is for FFT. For FFT submode the following schedules may be selected:
- submode=0b00 selects the
j
offset of the innermost for-loop of Tukey-Cooley - submode=0b10 selects the
j+halfsize
offset of the innermost for-loop of Tukey-Cooley - submode=0b11 selects the
k
of exptable (which coefficient)
When submode2 is 1 or 2, for DCT inner butterfly submode the following schedules may be selected. When submode2 is 1, additional bit-reversing is also performed.
- submode=0b00 selects the
j
offset of the innermost for-loop, in-place - submode=0b010 selects the
j+halfsize
offset of the innermost for-loop, in reverse-order, in-place - submode=0b10 selects the
ci
count of the innermost for-loop, useful for calculating the cosine coefficient - submode=0b11 selects the
size
offset of the outermost for-loop, useful for the cosine coefficientcos(ci + 0.5) * pi / size
When submode2 is 3 or 4, for DCT outer butterfly submode the following schedules may be selected. When submode is 3, additional bit-reversing is also performed.
- submode=0b00 selects the
j
offset of the innermost for-loop, - submode=0b01 selects the
j+1
offset of the innermost for-loop,
zdimsz
is used as an in-place "Stride", particularly useful for
column-based in-place DCT/FFT.
Matrix Mode
In Matrix Mode, skip allows dimensions to be skipped from being included
in the resultant output index. This allows sequences to be repeated:
0 0 0 1 1 1 2 2 2 ...
or in the case of skip=0b11 this results in
modulo 0 1 2 0 1 2 ...
- skip=0b00 indicates no dimensions to be skipped
- skip=0b01 sets "skip 1st dimension"
- skip=0b10 sets "skip 2nd dimension"
- skip=0b11 sets "skip 3rd dimension"
invxyz will invert the start index of each of x, y or z. If invxyz[0] is zero then x-dimensional counting begins from 0 and increments, otherwise it begins from xdimsz-1 and iterates down to zero. Likewise for y and z.
offset will have the effect of offsetting the result by offset
elements:
for i in 0..VL-1:
GPR(RT + remap(i) + SVSHAPE.offset) = ....
This appears redundant because the register RT could simply be changed by a compiler, until element width overrides are introduced. Also bear in mind that unlike a static compiler SVSHAPE.offset may be set dynamically at runtime.
xdimsz, ydimsz and zdimsz are offset by 1, such that a value of 0 indicates that the array dimensionality for that dimension is 1. any dimension not intended to be used must have its value set to 0 (dimensionality of 1). A value of xdimsz=2 would indicate that in the first dimension there are 3 elements in the array. For example, to create a 2D array X,Y of dimensionality X=3 and Y=2, set xdimsz=2, ydimsz=1 and zdimsz=0
The format of the array is therefore as follows:
array[xdimsz+1][ydimsz+1][zdimsz+1]
However whilst illustrative of the dimensionality, that does not take the "permute" setting into account. "permute" may be any one of six values (0-5, with values of 6 and 7 indicating "Indexed" Mode). The table below shows how the permutation dimensionality order works:
permute | order | array format |
---|---|---|
000 | 0,1,2 | (xdim+1)(ydim+1)(zdim+1) |
001 | 0,2,1 | (xdim+1)(zdim+1)(ydim+1) |
010 | 1,0,2 | (ydim+1)(xdim+1)(zdim+1) |
011 | 1,2,0 | (ydim+1)(zdim+1)(xdim+1) |
100 | 2,0,1 | (zdim+1)(xdim+1)(ydim+1) |
101 | 2,1,0 | (zdim+1)(ydim+1)(xdim+1) |
110 | 0,1 | Indexed (xdim+1)(ydim+1) |
111 | 1,0 | Indexed (ydim+1)(xdim+1) |
In other words, the "permute" option changes the order in which nested for-loops over the array would be done. See executable python reference code for further details.
Note: permute=0b110 and permute=0b111 enable Indexed REMAP Mode, described below
With all these options it is possible to support in-place transpose, in-place rotate, Matrix Multiply and Convolutions, without being limited to Power-of-Two dimension sizes.
Limitations and caveats
Limitations of Matrix REMAP are that the Vector Length (VL) is currently restricted to 127: up to 127 FMAs (or other operation) may be performed in total. Also given that it is in-registers only at present some care has to be taken on regfile resource utilisation. However it is perfectly possible to utilise Matrix REMAP to perform the three inner-most "kernel" loops of the usual 6-level "Tiled" large Matrix Multiply, without the usual difficulties associated with SIMD.
Also the svshape
instruction only provides access to part of the
Matrix REMAP capability. Rotation and mirroring need to be done by
programming the SVSHAPE SPRs directly, which can take a lot more
instructions. Future versions of SVP64 will
provide more comprehensive capacity and
mitigate the need to write direct to the SVSHAPE SPRs.
Additionally there is not yet a way to set Matrix sizes from registers
with svshape
: this was an intentional decision to simplify Hardware, that
may be corrected in a future version of SVP64. The limitation may presently
be overcome by direct programming of the SVSHAPE SPRs.
Hardware Architectural note: with the Scheduling applying as a Phase between Decode and Issue in a Deterministic fashion the Register Hazards may be easily computed and a standard Out-of-Order Micro-Architecture exploited to good effect. Even an In-Order system may observe that for large Outer Product Schedules there will be no stalls, but if the Matrices are particularly small size an In-Order system would have to stall, just as it would if the operations were loop-unrolled without Simple-V. Thus: regardless of the Micro-Architecture the Hardware Engineer should first consider how best to process the exact same equivalent loop-unrolled instruction stream. Once solved Matrix REMAP will fit naturally.
Indexed Mode
Indexed Mode activates reading of the element indices from the GPR and includes optional limited 2D reordering. In its simplest form (without elwidth overrides or other modes):
def index_remap(i):
return GPR((SVSHAPE.SVGPR<<1)+i) + SVSHAPE.offset
for i in 0..VL-1:
element_result = ....
GPR(RT + indexed_remap(i)) = element_result
With element-width overrides included, and using the pseudocode from the SVP64 appendix elwidth section this becomes:
def index_remap(i):
svreg = SVSHAPE.SVGPR << 1
srcwid = elwid_to_bitwidth(SVSHAPE.elwid)
offs = SVSHAPE.offset
return get_polymorphed_reg(svreg, srcwid, i) + offs
for i in 0..VL-1:
element_result = ....
rt_idx = indexed_remap(i)
set_polymorphed_reg(RT, destwid, rt_idx, element_result)
Matrix-style reordering still applies to the indices, except limited to up to 2 Dimensions (X,Y). Ordering is therefore limited to (X,Y) or (Y,X) for in-place Transposition. Only one dimension may optionally be skipped. Inversion of either X or Y or both is possible (2D mirroring). Pseudocode for Indexed Mode (including elwidth overrides) may be written in terms of Matrix Mode, specifically purposed to ensure that the 3rd dimension (Z) has no effect:
def index_remap(ISHAPE, i):
MSHAPE.skip = 0b0 || ISHAPE.sk1
MSHAPE.invxyz = 0b0 || ISHAPE.invxy
MSHAPE.xdimsz = ISHAPE.xdimsz
MSHAPE.ydimsz = ISHAPE.ydimsz
MSHAPE.zdimsz = 0 # disabled
if ISHAPE.permute = 0b110 # 0,1
MSHAPE.permute = 0b000 # 0,1,2
if ISHAPE.permute = 0b111 # 1,0
MSHAPE.permute = 0b010 # 1,0,2
el_idx = remap_matrix(MSHAPE, i)
svreg = ISHAPE.SVGPR << 1
srcwid = elwid_to_bitwidth(ISHAPE.elwid)
offs = ISHAPE.offset
return get_polymorphed_reg(svreg, srcwid, el_idx) + offs
The most important observation above is that the Matrix-style remapping occurs first and the Index lookup second. Thus it becomes possible to perform in-place Transpose of Indices which may have been costly to set up or costly to duplicate (waste register file space). In other words: it is fine for two or more SVSHAPEs to simultaneously use the same Indices (use the same GPRs), even if one SVSHAPE has different 2D dimensions and ordering from the others.
Caveats and Limitations
The purpose of Indexing is to provide a generalised version of
Vector ISA "Permute" instructions, such as VSX vperm
. The
Indexing is abstracted out and may be applied to much more
than an element move/copy, and is not limited for example
to the number of bytes that can fit into a VSX register.
Indexing may be applied to LD/ST (even on Indexed LD/ST
instructions such as sv.lbzx
), arithmetic operations,
extsw: there is no artificial limit.
The only major caveat is that the registers to be used as Indices must not be modified by any instruction after Indexed Mode is established, and neither must MAXVL be altered. Additionally, no register used as an Index may exceed MAXVL-1.
Failure to observe
these conditions results in UNDEFINED
behaviour.
These conditions allow a Read-After-Write (RAW) Hazard to be created on
the entire range of Indices to be subsequently used, but a corresponding
Write-After-Read Hazard by any instruction that modifies the Indices
does not have to be created. Given the large number of registers
involved in Indexing this is a huge resource saving and reduction
in micro-architectural complexity. MAXVL is likewise
included in the RAW Hazards because it is involved in calculating
how many registers are to be considered Indices.
With these Hazard Mitigations in place, high-performance implementations
may read-cache the Indices at the point where a given svindex
instruction
is called (or SVSHAPE SPRs - and MAXVL - directly altered) by issuing
background GPR register file reads whilst other instructions are being
issued and executed.
Indexed REMAP does not prevent conflicts (overlapping destinations), which on a superficial analysis may be perceived to be a problem, until it is recalled that, firstly, Simple-V is designed specifically to require Program Order to be respected, and that Matrix, DCT and FFT all already critically depend on overlapping Reads/Writes: Matrix uses overlapping registers as accumulators. Thus the Register Hazard Management needed by Indexed REMAP has to be in place anyway.
Programmer's Note: hphint
may be used to help hardware identify
parallelism opportunities but it is critical to remember that the
groupings are by FLOOR(step/MAXVL)
not FLOOR(REMAP(step)/MAXVL)
.
The cost compared to Matrix and other REMAPs (and Pack/Unpack) is clearly that of the additional reading of the GPRs to be used as Indices, plus the setup cost associated with creating those same Indices. If any Deterministic REMAP can cover the required task, clearly it is adviseable to use it instead.
Programmer's note: some algorithms may require skipping of Indices exceeding
VL-1, not MAXVL-1. This may be achieved programmatically by performing
an sv.cmp *BF,*RA,RB
where RA is the same GPRs used in the Indexed REMAP,
and RB contains the value of VL returned from setvl
. The resultant
CR Fields may then be used as Predicate Masks to exclude those operations
with an Index exceeding VL-1.
\newpage{}
svshape instruction
SVM-Form
svshape SVxd,SVyd,SVzd,SVRM,vf
0:5 | 6:10 | 11:15 | 16:20 | 21:24 | 25 | 26:31 | name |
---|---|---|---|---|---|---|---|
PO | SVxd | SVyd | SVzd | SVRM | vf | XO | svshape |
See appendix for svshape
pseudocode
Special Registers Altered:
SVSTATE, SVSHAPE0-3
svshape
is a convenience instruction that reduces instruction
count for common usage patterns, particularly Matrix, DCT and FFT. It sets up
(overwrites) all required SVSHAPE SPRs and also modifies SVSTATE
including VL and MAXVL. Using svshape
therefore does not also
require setvl
.
Fields:
- SVxd - SV REMAP "xdim" (X-dimension)
- SVyd - SV REMAP "ydim" (Y-dimension, sometimes used for sub-mode selection)
- SVzd - SV REMAP "zdim" (Z-dimension)
- SVRM - SV REMAP Mode (0b00000 for Matrix, 0b00001 for FFT etc.)
- vf - sets "Vertical-First" mode
- XO - standard 6-bit XO field
Note: SVxd, SVyz and SVzd are all stored "off-by-one". In the assembler
mnemonic the values 1-32
are stored in binary as 0b00000..0b11111
There are 12 REMAP Modes (2 Modes are RESERVED for svshape2
, 2 Modes
are RESERVED)
SVRM | Remap Mode description |
---|---|
0b0000 | Matrix 1/2/3D |
0b0001 | FFT Butterfly |
0b0010 | reserved for Matrix Outer Product |
0b0011 | DCT Outer butterfly |
0b0100 | DCT Inner butterfly, on-the-fly (Vertical-First Mode) |
0b0101 | DCT COS table index generation |
0b0110 | DCT half-swap |
0b0111 | Parallel Reduction and Prefix Sum |
0b1000 | reserved for svshape2 |
0b1001 | reserved for svshape2 |
0b1010 | reserved |
0b1011 | iDCT Outer butterfly |
0b1100 | iDCT Inner butterfly, on-the-fly (Vertical-First Mode) |
0b1101 | iDCT COS table index generation |
0b1110 | iDCT half-swap |
0b1111 | FFT half-swap |
Examples showing how all of these Modes operate exists in the online SVP64 unit tests. Explaining these Modes further in detail is beyond the scope of this document.
In Indexed Mode, there are only 5 bits available to specify the GPR
to use, out of 128 GPRs (7 bit numbering). Therefore, only the top
5 bits are given in the SVxd
field: the bottom two implicit bits
will be zero (SVxd || 0b00
).
svshape
has limited applicability due to being a 32-bit instruction.
The full capability of SVSHAPE SPRs may be accessed by directly writing
to SVSHAPE0-3 with mtspr
. Circumstances include Matrices with dimensions
larger than 32, and in-place Transpose. Potentially a future
instruction may extend the capability here.
Programmer's Note: Parallel Reduction Mode is selected by setting SVRM=7,SVyd=1
.
Prefix Sum Mode is selected by setting SVRM=7,SVyd=3
:
# Vector length of 8.
svshape 8, 3, 1, 0x7, 0
# activate SVSHAPE0 (prefix-sum lhs) for RA
# activate SVSHAPE1 (prefix-sum rhs) for RT and RB
svremap 7, 0, 1, 0, 1, 0, 0
sv.add *10, *10, *10
Architectural Resource Allocation note: the SVRM field is carefully
crafted to allocate two Modes, corresponding to bits 21-23 within the
instruction being set to the value 0b100
, to svshape2
(not
svshape
). These two Modes are
considered "RESERVED" within the context of svshape
but it is
absolutely critical to allocate the exact same pattern in XO for
both instructions in bits 26-31.
\newpage{}
svindex instruction
SVI-Form
0:5 | 6:10 | 11:15 | 16:20 | 21:25 | 26:31 | Form |
---|---|---|---|---|---|---|
PO | SVG | rmm | SVd | ew/yx/mm/sk | XO | SVI-Form |
- svindex SVG,rmm,SVd,ew,SVyx,mm,sk
See appendix for svindex
pseudocode
Special Registers Altered:
SVSTATE, SVSHAPE0-3
svindex
is a convenience instruction that reduces instruction count
for Indexed REMAP Mode. It sets up (overwrites) all required SVSHAPE
SPRs and unlike svshape
can modify the REMAP area of the SVSTATE
SPR as well, including setting persistence. The relevant SPRs may
be directly programmed with mtspr
however it is laborious to do so:
svindex saves instructions covering much of Indexed REMAP capability.
Fields:
- SVd - SV REMAP x/y dim
- rmm - REMAP mask: sets remap mi0-2/mo0-1 and SVSHAPEs, controlled by mm
- ew - sets element width override on the Indices
- SVG - GPR SVG<<2 to be used for Indexing
- yx - 2D reordering to be used if yx=1
- mm - mask mode. determines how
rmm
is interpreted. - sk - Dimension skipping enabled
Note: SVd, like SVxd, SVyz and SVzd of svshape
, are all stored
"off-by-one". In the assembler
mnemonic the values 1-32
are stored in binary as 0b00000..0b11111
.
Note: when yx=1,sk=0
the second dimension is calculated as
CEIL(MAXVL/SVd)
.
When mm=0
:
rmm
, like REMAP.SVme, has bit 0 correspond to mi0, bit 1 to mi1, bit 2 to mi2, bit 3 to mo0 and bit 4 to mi1- all SVSHAPEs and the REMAP parts of SVSHAPE are first reset (initialised to zero)
- for each bit set in the 5-bit
rmm
, in order, the first as-yet-unset SVSHAPE will be updated with the other operands in the instruction, and the REMAP SPR set. - If all 5 bits of
rmm
are set then both mi0 and mo1 use SVSHAPE0. - SVSTATE persistence bit is cleared
- No other alterations to SVSTATE are carried out
Example 1: if rmm=0b00110 then SVSHAPE0 and SVSHAPE1 are set up, and the REMAP SPR set so that mi1 uses SVSHAPE0 and mi2 uses mi2. REMAP.SVme is also set to 0b00110, REMAP.mi1=0 (SVSHAPE0) and REMAP.mi2=1 (SVSHAPE1)
Example 2: if rmm=0b10001 then again SVSHAPE0 and SVSHAPE1 are set up, but the REMAP SPR is set so that mi0 uses SVSHAPE0 and mo1 uses SVSHAPE1. REMAP.SVme=0b10001, REMAP.mi0=0, REMAP.mo1=1
Rough algorithmic form:
marray = [mi0, mi1, mi2, mo0, mo1]
idx = 0
for bit = 0 to 4:
if not rmm[bit]: continue
setup(SVSHAPE[idx])
SVSTATE{marray[bit]} = idx
idx = (idx+1) modulo 4
When mm=1
:
- bits 0-2 (MSB0 numbering) of
rmm
indicate an index selecting mi0-mo1 - bits 3-4 (MSB0 numbering) of
rmm
indicate which SVSHAPE 0-3 shall be updated - only the selected SVSHAPE is overwritten
- only the relevant bits in the REMAP area of SVSTATE are updated
- REMAP persistence bit is set.
Example 1: if rmm
=0b01110 then bits 0-2 (MSB0) are 0b011 and
bits 3-4 are 0b10. thus, mo0 is selected and SVSHAPE2
to be updated. REMAP.SVme[3] will be set high and REMAP.mo0
set to 2 (SVSHAPE2).
Example 2: if rmm
=0b10011 then bits 0-2 (MSB0) are 0b100 and
bits 3-4 are 0b11. thus, mo1 is selected and SVSHAPE3
to be updated. REMAP.SVme[4] will be set high and REMAP.mo1
set to 3 (SVSHAPE3).
Rough algorithmic form:
marray = [mi0, mi1, mi2, mo0, mo1]
bit = rmm[0:2]
idx = rmm[3:4]
setup(SVSHAPE[idx])
SVSTATE{marray[bit]} = idx
SVSTATE.pst = 1
In essence, mm=0
is intended for use to set as much of the
REMAP State SPRs as practical with a single instruction,
whilst mm=1
is intended to be a little more refined.
Usage guidelines
- Disable 2D mapping: to only perform Indexing without
reordering use
SVd=1,sk=0,yx=0
(or set SVd to a value larger or equal to VL) - Modulo 1D mapping: to perform Indexing cycling through the
first N Indices use
SVd=N,sk=0,yx=0
whereVL>N
. There is no requirement to set VL equal to a multiple of N. - Modulo 2D transposed:
SVd=M,sk=0,yx=1
, setsxdim=M,ydim=CEIL(MAXVL/M)
.
Beyond these mappings it becomes necessary to write directly to the SVSTATE SPRs manually.
\newpage{}
svshape2 (offset-priority)
SVM2-Form
0:5 | 6:9 | 10 | 11:15 | 16:20 | 21:24 | 25 | 26:31 | Form |
---|---|---|---|---|---|---|---|---|
PO | offs | yx | rmm | SVd | 100/mm | sk | XO | SVM2-Form |
- svshape2 offs,yx,rmm,SVd,sk,mm
See appendix for svshape2
pseudocode
Special Registers Altered:
SVSTATE, SVSHAPE0-3
svshape2
is an additional convenience instruction that prioritises
setting SVSHAPE.offset
. Its primary purpose is for use when
element-width overrides are used. It has identical capabilities to svindex
in terms of both options (skip, etc.) and ability to activate REMAP
(rmm, mask mode) but unlike svindex
it does not set GPR REMAP:
only a 1D or 2D svshape
, and
unlike svshape
it can set an arbitrary SVSHAPE.offset
immediate.
One of the limitations of Simple-V is that Vector elements start on the boundary
of the Scalar regfile, which is fine when element-width overrides are not
needed. If the starting point of a Vector with smaller elwidths must begin
in the middle of a register, normally there would be no way to do so except
through costly LD/ST. SVSHAPE.offset
caters for this scenario and svshape2
makes it easier to access.
Operand Fields:
- offs (4 bits) - unsigned offset
- yx (1 bit) - swap XY to YX
- SVd dimension size
- rmm REMAP mask
- mm mask mode
- sk (1 bit) skips 1st dimension if set
Dimensions are calculated exactly as svindex
. rmm
and
mm
are as per svindex
.
Programmer's Note: offsets for svshape2
may be specified in the range
0-15. Given that the principle of Simple-V is to fit on top of
byte-addressable register files and that GPR and FPR are 64-bit (8 bytes)
it should be clear that the offset may, when elwidth=8
, begin an
element-level operation starting element zero at any arbitrary byte.
On cursory examination attempting to go beyond the range 0-7 seems
unnecessary given that the next GPR or FPR is an
alias for an offset in the range 8-15. Thus by simply increasing
the starting Vector point of the operation to the next register it
can be seen that the offset of 0-7 would be sufficient. Unfortunately
however some operations are EXTRA2-encoded it is not possible
to increase the GPR/FPR register number by one, because EXTRA2-encoding
of GPR/FPR Vector numbers are restricted to even numbering.
For CR Fields the EXTRA2 encoding is even more sparse.
The additional offset range (8-15) helps overcome these limitations.
Hardware Implementor's note: with the offsets only being immediates and with register numbering being entirely immediate as well it is possible to correctly compute Register Hazards without requiring reading the contents of any SPRs. If however there are instructions that have directly written to the SVSTATE or SVSHAPE SPRs and those instructions are still in-flight then this position is clearly invalid. This is why Programmers are strongly discouraged from directly writing to these SPRs.
Architectural Resource Allocation note: this instruction shares
the space of svshape
. Therefore it is critical that the two
instructions, svshape
and svshape2
have the exact same XO
in bits 26 thru 31. It is also critical that for svshape2
,
bit 21 of XO is a 1, bit 22 of XO is a 0, and bit 23 of XO is a 0.
\newpage{}
Forms
Add SVI, SVM, SVM2, SVRM
to XO (26:31)
Field in Book I, 1.6.2
Add the following to Book I, 1.6.1, SVI-Form
|0 |6 |11 |16 |21 |23 |24|25|26 31|
| PO | SVG|rmm | SVd |ew |SVyx|mm|sk| XO |
Add the following to Book I, 1.6.1, SVM-Form
|0 |6 |11 |16 |21 |25 |26 |31 |
| PO | SVxd | SVyd | SVzd | SVrm |vf | XO |
Add the following to Book I, 1.6.1, SVM2-Form
|0 |6 |10 |11 |16 |21 |24|25 |26 |31 |
| PO | SVo |SVyx| rmm | SVd |XO |mm|sk | XO |
Add the following to Book I, 1.6.1, SVRM-Form
|0 |6 |11 |13 |15 |17 |19 |21 |22 |26 |31 |
| PO | SVme |mi0 | mi1 | mi2 | mo0 | mo1 |pst |/// | XO |
Add the following to Book I, 1.6.2
mi0 (11:12)
Field used in REMAP to select the SVSHAPE for 1st input register
Formats: SVRM
mi1 (13:14)
Field used in REMAP to select the SVSHAPE for 2nd input register
Formats: SVRM
mi2 (15:16)
Field used in REMAP to select the SVSHAPE for 3rd input register
Formats: SVRM
mm (24)
Field used to specify the meaning of the rmm field for SVI-Form
and SVM2-Form
Formats: SVI, SVM2
mo0 (17:18)
Field used in REMAP to select the SVSHAPE for 1st output register
Formats: SVRM
mo1 (19:20)
Field used in REMAP to select the SVSHAPE for 2nd output register
Formats: SVRM
pst (21)
Field used in REMAP to indicate "persistence" mode (REMAP
continues to apply to multiple instructions)
Formats: SVRM
rmm (11:15)
REMAP Mode field for SVI-Form and SVM2-Form
Formats: SVI, SVM2
sk (25)
Field used to specify dimensional skipping in svindex
Formats: SVI, SVM2
SVd (16:20)
Immediate field used to specify the size of the REMAP dimension
in the svindex and svshape2 instructions
Formats: SVI, SVM2
SVDS (16:29)
Immediate field used to specify a 9-bit signed
two's complement integer which is concatenated
on the right with 0b00 and sign-extended to 64 bits.
Formats: SVDS
SVG (6:10)
Field used to specify a GPR to be used as a
source for indexing.
Formats: SVI
SVi (16:22)
Simple-V immediate field for setting VL or MVL
Formats: SVL
SVme (6:10)
Simple-V "REMAP" map-enable bits (0-4)
Formats: SVRM
SVo (6:9)
Field used by the svshape2 instruction as an offset
Formats: SVM2
SVrm (21:24)
Simple-V "REMAP" Mode
Formats: SVM
SVxd (6:10)
Simple-V "REMAP" x-dimension size
Formats: SVM
SVyd (11:15)
Simple-V "REMAP" y-dimension size
Formats: SVM
SVzd (16:20)
Simple-V "REMAP" z-dimension size
Formats: SVM
XO (21:23,26:31)
Extended opcode field. Note that bit 21 must be 1, 22 and 23
must be zero, and bits 26-31 must be exactly the same as
used for svshape.
Formats: SVM2
Appendices
Appendix E Power ISA sorted by opcode
Appendix F Power ISA sorted by version
Appendix G Power ISA sorted by Compliancy Subset
Appendix H Power ISA sorted by mnemonic
Form | Book | Page | Version | mnemonic | Description |
---|---|---|---|---|---|
SVRM | I | # | 3.0B | svremap | REMAP enabling instruction |
SVM | I | # | 3.0B | svshape | REMAP shape instruction |
SVM2 | I | # | 3.0B | svshape2 | REMAP shape instruction (2) |
SVI | I | # | 3.0B | svindex | REMAP General-purpose Indexing |
REMAP Matrix pseudocode
The algorithm below shows how REMAP works more clearly, and may be executed as a python program:
# Finite State Machine version of the REMAP system. much more likely
# to end up being actually used in actual hardware
# up to three dimensions permitted
xdim = 3
ydim = 2
zdim = 1
VL = xdim * ydim * zdim # set total (can repeat, e.g. VL=x*y*z*4)
lims = [xdim, ydim, zdim]
idxs = [0,0,0] # starting indices
applydim = [1, 1] # apply lower dims
order = [1,0,2] # experiment with different permutations, here
offset = 0 # experiment with different offsetet, here
invxyz = [0,1,0] # inversion allowed
# pre-prepare the index state: run for "offset" times before
# actually starting. this algorithm can also be used for re-entrancy
# if exceptions occur and a REMAP has to be started from where the
# interrupt left off.
for idx in range(offset):
for i in range(3):
idxs[order[i]] = idxs[order[i]] + 1
if (idxs[order[i]] != lims[order[i]]):
break
idxs[order[i]] = 0
break_count = 0 # for pretty-printing
for idx in range(VL):
ix = [0] * 3
for i in range(3):
ix[i] = idxs[i]
if invxyz[i]:
ix[i] = lims[i] - 1 - ix[i]
new_idx = ix[2]
if applydim[1]:
new_idx = new_idx * ydim + ix[1]
if applydim[0]:
new_idx = new_idx * xdim + ix[0]
print ("%d->%d" % (idx, new_idx)),
break_count += 1
if break_count == lims[order[0]]:
print ("")
break_count = 0
# this is the exact same thing as the pre-preparation stage
# above. step 1: count up to the limit of the current dimension
# step 2: if limit reached, zero it, and allow the *next* dimension
# to increment. repeat for 3 dimensions.
for i in range(3):
idxs[order[i]] = idxs[order[i]] + 1
if (idxs[order[i]] != lims[order[i]]):
break
idxs[order[i]] = 0
An easier-to-read version (using python iterators) is given in a later section of this Appendix.
Each element index from the for-loop 0..VL-1
is run through the above algorithm to work out the actual element
index, instead. Given that there are four possible SHAPE entries, up to
four separate registers in any given operation may be simultaneously
remapped:
function op_add(RT, RA, RB) # add not VADD!
for (i=0,id=0,irs1=0,irs2=0; i < VL; i++)
SVSTATE.srcstep = i # save context
if (predval & 1<<i) # predication mask
GPR[RT+remap1(id)] <= GPR[RA+remap2(irs1)] +
GPR[RB+remap3(irs2)];
if (!RT.isvector) break;
if (RT.isvector) { id += 1; }
if (RA.isvector) { irs1 += 1; }
if (RB.isvector) { irs2 += 1; }
By changing remappings, 2D matrices may be transposed "in-place" for one operation, followed by setting a different permutation order without having to move the values in the registers to or from memory.
Note that:
- Over-running the register file clearly has to be detected and an illegal instruction exception thrown
- When non-default elwidths are set, the exact same algorithm still applies (i.e. it offsets polymorphic elements within registers rather than entire registers).
- If permute option 000 is utilised, the actual order of the reindexing does not change. However, modulo MVL still occurs which will result in repeated operations (use with caution).
- If two or more dimensions are set to zero, the actual order does not change!
- The above algorithm is pseudo-code only. Actual implementations will need to take into account the fact that the element for-looping must be re-entrant, due to the possibility of exceptions occurring. See SVSTATE SPR, which records the current element index. Continuing after return from an interrupt may introduce latency due to re-computation of the remapped offsets.
- Twin-predicated operations require two separate and distinct
element offsets. The above pseudo-code algorithm will be applied
separately and independently to each, should each of the two
operands be remapped. This even includes unit-strided LD/ST
and other operations
in that category, where in that case it will be the address offset
that is remapped:
EA <- (RA) + immediate * REMAP(elementoffset)
. - Offset is especially useful, on its own, for accessing elements within the middle of a register. Without offsets, it is necessary to either use a predicated MV, skipping the first elements, or performing a LOAD/STORE cycle to memory. With offsets, the data does not have to be moved.
- Setting the total elements (xdim+1) times (ydim+1) times (zdim+1) to less than MVL is perfectly legal, albeit very obscure. It permits entries to be regularly presented to operands more than once, thus allowing the same underlying registers to act as an accumulator of multiple vector or matrix operations, for example.
- Note especially that Program Order must still be respected even when overlaps occur that read or write the same register elements including polymorphic ones
Clearly here some considerable care needs to be taken as the remapping could hypothetically create arithmetic operations that target the exact same underlying registers, resulting in data corruption due to pipeline overlaps. Out-of-order / Superscalar micro-architectures with register-renaming will have an easier time dealing with this than DSP-style SIMD micro-architectures.
4x4 Matrix to vec4 Multiply (4x4 by 1x4)
The following settings will allow a 4x4 matrix (starting at f8), expressed as a sequence of 16 numbers first by row then by column, to be multiplied by a vector of length 4 (starting at f0), using a single FMAC instruction.
- SHAPE0: xdim=4, ydim=4, permute=yx, applied to f0
- SHAPE1: xdim=4, ydim=1, permute=xy, applied to f4
- VL=16, f4=vec, f0=vec, f8=vec
- FMAC f4, f0, f8, f4
The permutation on SHAPE0 will use f0 as a vec4 source. On the first four iterations through the hardware loop, the REMAPed index will not increment. On the second four, the index will increase by one. Likewise on each subsequent group of four.
The permutation on SHAPE1 will increment f4 continuously cycling through f4-f7 every four iterations of the hardware loop.
At the same time, VL will, because there is no SHAPE on f8, increment straight sequentially through the 16 values f8-f23 in the Matrix. The equivalent sequence thus is issued:
fmac f4, f0, f8, f4
fmac f5, f0, f9, f5
fmac f6, f0, f10, f6
fmac f7, f0, f11, f7
fmac f4, f1, f12, f4
fmac f5, f1, f13, f5
fmac f6, f1, f14, f6
fmac f7, f1, f15, f7
fmac f4, f2, f16, f4
fmac f5, f2, f17, f5
fmac f6, f2, f18, f6
fmac f7, f2, f19, f7
fmac f4, f3, f20, f4
fmac f5, f3, f21, f5
fmac f6, f3, f22, f6
fmac f7, f3, f23, f7
Hardware should easily pipeline the above FMACs and as long as each FMAC completes in 4 cycles or less there should be 100% sustained throughput, from the one single Vector FMAC.
The only other instruction required is to ensure that f4-f7 are initialised (usually to zero) however obviously if used as part of some other computation, which is frequently the case, then clearly the zeroing is not needed.
REMAP FFT, DFT, NTT
The algorithm from a later section of this Appendix shows how FFT REMAP works, and it may be executed as a standalone python3 program. The executable code is designed to illustrate how a hardware implementation may generate Indices which are completely independent of the Execution of element-level operations, even for something as complex as a Triple-loop Tukey-Cooley Schedule. A comprehensive demo and test suite may be found here including Complex Number FFT which deploys Vertical-First Mode on top of the REMAP Schedules.
Other uses include more than DFT and NTT: as abstracted RISC-paradigm the Schedules are not restricted in any way or tied to any particular instruction. If the programmer can find any algorithm which has identical triple nesting then the FFT Schedule may be used even there.
svshape pseudocode
# for convenience, VL to be calculated and stored in SVSTATE
vlen <- [0] * 7
mscale[0:5] <- 0b000001 # for scaling MAXVL
itercount[0:6] <- [0] * 7
SVSTATE[0:31] <- [0] * 32
# only overwrite REMAP if "persistence" is zero
if (SVSTATE[62] = 0b0) then
SVSTATE[32:33] <- 0b00
SVSTATE[34:35] <- 0b00
SVSTATE[36:37] <- 0b00
SVSTATE[38:39] <- 0b00
SVSTATE[40:41] <- 0b00
SVSTATE[42:46] <- 0b00000
SVSTATE[62] <- 0b0
SVSTATE[63] <- 0b0
# clear out all SVSHAPEs
SVSHAPE0[0:31] <- [0] * 32
SVSHAPE1[0:31] <- [0] * 32
SVSHAPE2[0:31] <- [0] * 32
SVSHAPE3[0:31] <- [0] * 32
# set schedule up for multiply
if (SVrm = 0b0000) then
# VL in Matrix Multiply is xd*yd*zd
xd <- (0b00 || SVxd) + 1
yd <- (0b00 || SVyd) + 1
zd <- (0b00 || SVzd) + 1
n <- xd * yd * zd
vlen[0:6] <- n[14:20]
# set up template in SVSHAPE0, then copy to 1-3
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[6:11] <- (0b0 || SVyd) # ydim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim
SVSHAPE0[28:29] <- 0b11 # skip z
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31]
SVSHAPE2[0:31] <- SVSHAPE0[0:31]
SVSHAPE3[0:31] <- SVSHAPE0[0:31]
# set up FRA
SVSHAPE1[18:20] <- 0b001 # permute x,z,y
SVSHAPE1[28:29] <- 0b01 # skip z
# FRC
SVSHAPE2[18:20] <- 0b001 # permute x,z,y
SVSHAPE2[28:29] <- 0b11 # skip y
# set schedule up for FFT butterfly
if (SVrm = 0b0001) then
# calculate O(N log2 N)
n <- [0] * 3
do while n < 5
if SVxd[4-n] = 0 then
leave
n <- n + 1
n <- ((0b0 || SVxd) + 1) * n
vlen[0:6] <- n[1:7]
# set up template in SVSHAPE0, then copy to 1-3
# for FRA and FRT
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D FFT)
mscale <- (0b0 || SVzd) + 1
SVSHAPE0[30:31] <- 0b01 # Butterfly mode
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31]
SVSHAPE2[0:31] <- SVSHAPE0[0:31]
# set up FRB and FRS
SVSHAPE1[28:29] <- 0b01 # j+halfstep schedule
# FRC (coefficients)
SVSHAPE2[28:29] <- 0b10 # k schedule
# set schedule up for (i)DCT Inner butterfly
# SVrm Mode 4 (Mode 12 for iDCT) is for on-the-fly (Vertical-First Mode)
if ((SVrm = 0b0100) |
(SVrm = 0b1100)) then
# calculate O(N log2 N)
n <- [0] * 3
do while n < 5
if SVxd[4-n] = 0 then
leave
n <- n + 1
n <- ((0b0 || SVxd) + 1) * n
vlen[0:6] <- n[1:7]
# set up template in SVSHAPE0, then copy to 1-3
# set up FRB and FRS
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D DCT)
mscale <- (0b0 || SVzd) + 1
if (SVrm = 0b1100) then
SVSHAPE0[30:31] <- 0b11 # iDCT mode
SVSHAPE0[18:20] <- 0b011 # iDCT Inner Butterfly sub-mode
else
SVSHAPE0[30:31] <- 0b01 # DCT mode
SVSHAPE0[18:20] <- 0b001 # DCT Inner Butterfly sub-mode
SVSHAPE0[21:23] <- 0b001 # "inverse" on outer loop
SVSHAPE0[6:11] <- 0b000011 # (i)DCT Inner Butterfly mode 4
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31]
SVSHAPE2[0:31] <- SVSHAPE0[0:31]
if (SVrm != 0b0100) & (SVrm != 0b1100) then
SVSHAPE3[0:31] <- SVSHAPE0[0:31]
# for FRA and FRT
SVSHAPE0[28:29] <- 0b01 # j+halfstep schedule
# for cos coefficient
SVSHAPE2[28:29] <- 0b10 # ci (k for mode 4) schedule
SVSHAPE2[12:17] <- 0b000000 # reset costable "striding" to 1
if (SVrm != 0b0100) & (SVrm != 0b1100) then
SVSHAPE3[28:29] <- 0b11 # size schedule
# set schedule up for (i)DCT Outer butterfly
if (SVrm = 0b0011) | (SVrm = 0b1011) then
# calculate O(N log2 N) number of outer butterfly overlapping adds
vlen[0:6] <- [0] * 7
n <- 0b000
size <- 0b0000001
itercount[0:6] <- (0b00 || SVxd) + 0b0000001
itercount[0:6] <- (0b0 || itercount[0:5])
do while n < 5
if SVxd[4-n] = 0 then
leave
n <- n + 1
count <- (itercount - 0b0000001) * size
vlen[0:6] <- vlen + count[7:13]
size[0:6] <- (size[1:6] || 0b0)
itercount[0:6] <- (0b0 || itercount[0:5])
# set up template in SVSHAPE0, then copy to 1-3
# set up FRB and FRS
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D DCT)
mscale <- (0b0 || SVzd) + 1
if (SVrm = 0b1011) then
SVSHAPE0[30:31] <- 0b11 # iDCT mode
SVSHAPE0[18:20] <- 0b011 # iDCT Outer Butterfly sub-mode
SVSHAPE0[21:23] <- 0b101 # "inverse" on outer and inner loop
else
SVSHAPE0[30:31] <- 0b01 # DCT mode
SVSHAPE0[18:20] <- 0b100 # DCT Outer Butterfly sub-mode
SVSHAPE0[6:11] <- 0b000010 # DCT Butterfly mode
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31] # j+halfstep schedule
SVSHAPE2[0:31] <- SVSHAPE0[0:31] # costable coefficients
# for FRA and FRT
SVSHAPE1[28:29] <- 0b01 # j+halfstep schedule
# reset costable "striding" to 1
SVSHAPE2[12:17] <- 0b000000
# set schedule up for DCT COS table generation
if (SVrm = 0b0101) | (SVrm = 0b1101) then
# calculate O(N log2 N)
vlen[0:6] <- [0] * 7
itercount[0:6] <- (0b00 || SVxd) + 0b0000001
itercount[0:6] <- (0b0 || itercount[0:5])
n <- [0] * 3
do while n < 5
if SVxd[4-n] = 0 then
leave
n <- n + 1
vlen[0:6] <- vlen + itercount
itercount[0:6] <- (0b0 || itercount[0:5])
# set up template in SVSHAPE0, then copy to 1-3
# set up FRB and FRS
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D DCT)
mscale <- (0b0 || SVzd) + 1
SVSHAPE0[30:31] <- 0b01 # DCT/FFT mode
SVSHAPE0[6:11] <- 0b000100 # DCT Inner Butterfly COS-gen mode
if (SVrm = 0b0101) then
SVSHAPE0[21:23] <- 0b001 # "inverse" on outer loop for DCT
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31]
SVSHAPE2[0:31] <- SVSHAPE0[0:31]
# for cos coefficient
SVSHAPE1[28:29] <- 0b10 # ci schedule
SVSHAPE2[28:29] <- 0b11 # size schedule
# set schedule up for iDCT / DCT inverse of half-swapped ordering
if (SVrm = 0b0110) | (SVrm = 0b1110) | (SVrm = 0b1111) then
vlen[0:6] <- (0b00 || SVxd) + 0b0000001
# set up template in SVSHAPE0
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D DCT)
mscale <- (0b0 || SVzd) + 1
if (SVrm = 0b1110) then
SVSHAPE0[18:20] <- 0b001 # DCT opposite half-swap
if (SVrm = 0b1111) then
SVSHAPE0[30:31] <- 0b01 # FFT mode
else
SVSHAPE0[30:31] <- 0b11 # DCT mode
SVSHAPE0[6:11] <- 0b000101 # DCT "half-swap" mode
# set schedule up for parallel reduction
if (SVrm = 0b0111) then
# calculate the total number of operations (brute-force)
vlen[0:6] <- [0] * 7
itercount[0:6] <- (0b00 || SVxd) + 0b0000001
step[0:6] <- 0b0000001
i[0:6] <- 0b0000000
do while step <u itercount
newstep <- step[1:6] || 0b0
j[0:6] <- 0b0000000
do while (j+step <u itercount)
j <- j + newstep
i <- i + 1
step <- newstep
# VL in Parallel-Reduce is the number of operations
vlen[0:6] <- i
# set up template in SVSHAPE0, then copy to 1. only 2 needed
SVSHAPE0[0:5] <- (0b0 || SVxd) # xdim
SVSHAPE0[12:17] <- (0b0 || SVzd) # zdim - "striding" (2D DCT)
mscale <- (0b0 || SVzd) + 1
SVSHAPE0[30:31] <- 0b10 # parallel reduce submode
# copy
SVSHAPE1[0:31] <- SVSHAPE0[0:31]
# set up right operand (left operand 28:29 is zero)
SVSHAPE1[28:29] <- 0b01 # right operand
# set VL, MVL and Vertical-First
m[0:12] <- vlen * mscale
maxvl[0:6] <- m[6:12]
SVSTATE[0:6] <- maxvl # MAVXL
SVSTATE[7:13] <- vlen # VL
SVSTATE[63] <- vf
svindex pseudocode
# based on nearest MAXVL compute other dimension
MVL <- SVSTATE[0:6]
d <- [0] * 6
dim <- SVd+1
do while d*dim <u ([0]*4 || MVL)
d <- d + 1
# set up template, then copy once location identified
shape <- [0]*32
shape[30:31] <- 0b00 # mode
if SVyx = 0 then
shape[18:20] <- 0b110 # indexed xd/yd
shape[0:5] <- (0b0 || SVd) # xdim
if sk = 0 then shape[6:11] <- 0 # ydim
else shape[6:11] <- 0b111111 # ydim max
else
shape[18:20] <- 0b111 # indexed yd/xd
if sk = 1 then shape[6:11] <- 0 # ydim
else shape[6:11] <- d-1 # ydim max
shape[0:5] <- (0b0 || SVd) # ydim
shape[12:17] <- (0b0 || SVG) # SVGPR
shape[28:29] <- ew # element-width override
shape[21] <- sk # skip 1st dimension
# select the mode for updating SVSHAPEs
SVSTATE[62] <- mm # set or clear persistence
if mm = 0 then
# clear out all SVSHAPEs first
SVSHAPE0[0:31] <- [0] * 32
SVSHAPE1[0:31] <- [0] * 32
SVSHAPE2[0:31] <- [0] * 32
SVSHAPE3[0:31] <- [0] * 32
SVSTATE[32:41] <- [0] * 10 # clear REMAP.mi/o
SVSTATE[42:46] <- rmm # rmm exactly REMAP.SVme
idx <- 0
for bit = 0 to 4
if rmm[4-bit] then
# activate requested shape
if idx = 0 then SVSHAPE0 <- shape
if idx = 1 then SVSHAPE1 <- shape
if idx = 2 then SVSHAPE2 <- shape
if idx = 3 then SVSHAPE3 <- shape
SVSTATE[bit*2+32:bit*2+33] <- idx
# increment shape index, modulo 4
if idx = 3 then idx <- 0
else idx <- idx + 1
else
# refined SVSHAPE/REMAP update mode
bit <- rmm[0:2]
idx <- rmm[3:4]
if idx = 0 then SVSHAPE0 <- shape
if idx = 1 then SVSHAPE1 <- shape
if idx = 2 then SVSHAPE2 <- shape
if idx = 3 then SVSHAPE3 <- shape
SVSTATE[bit*2+32:bit*2+33] <- idx
SVSTATE[46-bit] <- 1
svshape2 pseudocode
# based on nearest MAXVL compute other dimension
MVL <- SVSTATE[0:6]
d <- [0] * 6
dim <- SVd+1
do while d*dim <u ([0]*4 || MVL)
d <- d + 1
# set up template, then copy once location identified
shape <- [0]*32
shape[30:31] <- 0b00 # mode
shape[0:5] <- (0b0 || SVd) # x/ydim
if SVyx = 0 then
shape[18:20] <- 0b000 # ordering xd/yd(/zd)
if sk = 0 then shape[6:11] <- 0 # ydim
else shape[6:11] <- 0b111111 # ydim max
else
shape[18:20] <- 0b010 # ordering yd/xd(/zd)
if sk = 1 then shape[6:11] <- 0 # ydim
else shape[6:11] <- d-1 # ydim max
# offset (the prime purpose of this instruction)
shape[24:27] <- SVo # offset
if sk = 1 then shape[28:29] <- 0b01 # skip 1st dimension
else shape[28:29] <- 0b00 # no skipping
# select the mode for updating SVSHAPEs
SVSTATE[62] <- mm # set or clear persistence
if mm = 0 then
# clear out all SVSHAPEs first
SVSHAPE0[0:31] <- [0] * 32
SVSHAPE1[0:31] <- [0] * 32
SVSHAPE2[0:31] <- [0] * 32
SVSHAPE3[0:31] <- [0] * 32
SVSTATE[32:41] <- [0] * 10 # clear REMAP.mi/o
SVSTATE[42:46] <- rmm # rmm exactly REMAP.SVme
idx <- 0
for bit = 0 to 4
if rmm[4-bit] then
# activate requested shape
if idx = 0 then SVSHAPE0 <- shape
if idx = 1 then SVSHAPE1 <- shape
if idx = 2 then SVSHAPE2 <- shape
if idx = 3 then SVSHAPE3 <- shape
SVSTATE[bit*2+32:bit*2+33] <- idx
# increment shape index, modulo 4
if idx = 3 then idx <- 0
else idx <- idx + 1
else
# refined SVSHAPE/REMAP update mode
bit <- rmm[0:2]
idx <- rmm[3:4]
if idx = 0 then SVSHAPE0 <- shape
if idx = 1 then SVSHAPE1 <- shape
if idx = 2 then SVSHAPE2 <- shape
if idx = 3 then SVSHAPE3 <- shape
SVSTATE[bit*2+32:bit*2+33] <- idx
SVSTATE[46-bit] <- 1
Example Matrix Usage
svshape
to set the type of reordering to be applied to an otherwise usual0..VL-1
hardware for-loopsvremap
to set which registers a given reordering is to apply to (RA, RT etc)sv.{instruction}
where any Vectorized register marked bysvremap
will have its ordering REMAPPED according to the schedule set bysvshape
.
The following illustrative example multiplies a 3x4 and a 5x3 matrix to create a 5x4 result:
svshape 5,4,3,0,0 # Outer Product 5x4 by 4x3
svremap 15,1,2,3,0,0,0,0 # link Schedule to registers
sv.fmadds *0,*32,*64,*0 # 60 FMACs get executed here
- svshape sets up the four SVSHAPE SPRS for a Matrix Schedule
- svremap activates four out of five registers RA RB RC RT RS (15)
- svremap requests:
- RA to use SVSHAPE1
- RB to use SVSHAPE2
- RC to use SVSHAPE3
- RT to use SVSHAPE0
- RS Remapping to not be activated
- sv.fmadds has vectors RT=0, RA=32, RB=64, RC=0
- With REMAP being active each register's element index is independently transformed using the specified SHAPEs.
Thus the Vector Loop is arranged such that the use of the multiply-and-accumulate instruction executes precisely the required Schedule to perform an in-place in-registers Outer Product Matrix Multiply with no need to perform additional Transpose or register copy instructions. The example above may be executed as a unit test and demo, here
\newpage{}
REMAP pseudocode
Written in python3 the following stand-alone executable source code is the Canonical
Specification for each REMAP. Vectors of "loopends" are returned when Rc=1
in Vectors of CR Fields on sv.svstep.
, or in Vertical-First Mode
a single CR Field (CR0) on svstep.
. The SVSTATE.srcstep
or SVSTATE.dststep
sequential
offset is put through each algorithm to determine the actual Element Offset.
Alternative implementations producing different ordering
is prohibited as software will be critically relying on these Deterministic Schedules.
REMAP 2D/3D Matrix
The following stand-alone executable source code is the Canonical Specification for Matrix (2D/3D) REMAP. Hardware implementations are achievable with simple cascading counter-and-compares.
# python "yield" can be iterated. use this to make it clear how
# the indices are generated by using natural-looking nested loops
def iterate_indices(SVSHAPE):
# get indices to iterate over, in the required order
xd = SVSHAPE.lims[0]
yd = SVSHAPE.lims[1]
zd = SVSHAPE.lims[2]
# create lists of indices to iterate over in each dimension
x_r = list(range(xd))
y_r = list(range(yd))
z_r = list(range(zd))
# invert the indices if needed
if SVSHAPE.invxyz[0]: x_r.reverse()
if SVSHAPE.invxyz[1]: y_r.reverse()
if SVSHAPE.invxyz[2]: z_r.reverse()
# start an infinite (wrapping) loop
step = 0 # track src/dst step
while True:
for z in z_r: # loop over 1st order dimension
z_end = z == z_r[-1]
for y in y_r: # loop over 2nd order dimension
y_end = y == y_r[-1]
for x in x_r: # loop over 3rd order dimension
x_end = x == x_r[-1]
# ok work out which order to construct things in.
# start by creating a list of tuples of the dimension
# and its limit
vals = [(SVSHAPE.lims[0], x, "x"),
(SVSHAPE.lims[1], y, "y"),
(SVSHAPE.lims[2], z, "z")
]
# now select those by order. this allows us to
# create schedules for [z][x], [x][y], or [y][z]
# for matrix multiply.
vals = [vals[SVSHAPE.order[0]],
vals[SVSHAPE.order[1]],
vals[SVSHAPE.order[2]]
]
# ok now we can construct the result, using bits of
# "order" to say which ones get stacked on
result = 0
mult = 1
for i in range(3):
lim, idx, dbg = vals[i]
# some of the dimensions can be "skipped". the order
# was actually selected above on all 3 dimensions,
# e.g. [z][x][y] or [y][z][x]. "skip" allows one of
# those to be knocked out
if SVSHAPE.skip == i+1: continue
idx *= mult # shifts up by previous dimension(s)
result += idx # adds on this dimension
mult *= lim # for the next dimension
loopends = (x_end |
((y_end and x_end)<<1) |
((y_end and x_end and z_end)<<2))
yield result + SVSHAPE.offset, loopends
step += 1
def demo():
# set the dimension sizes here
xdim = 3
ydim = 2
zdim = 4
# set total (can repeat, e.g. VL=x*y*z*4)
VL = xdim * ydim * zdim
# set up an SVSHAPE
class SVSHAPE:
pass
SVSHAPE0 = SVSHAPE()
SVSHAPE0.lims = [xdim, ydim, zdim]
SVSHAPE0.order = [1,0,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b00
SVSHAPE0.skip = 0b00
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
# enumerate over the iterator function, getting new indices
for idx, (new_idx, end) in enumerate(iterate_indices(SVSHAPE0)):
if idx >= VL:
break
print ("%d->%d" % (idx, new_idx), "end", bin(end)[2:])
# run the demo
if __name__ == '__main__':
demo()
REMAP Parallel Reduction pseudocode
The python3 program below is stand-alone executable and is the Canonical Specification for Parallel Reduction REMAP. The Algorithm below is not limited to RADIX2 sizes, and Predicate sources, unlike in Matrix REMAP, apply to the Element Indices after REMAP has been applied, not before. MV operations are not required: the algorithm tracks positions of elements that would normally be moved and when applying an Element Reduction Operation sources the operands from their last-known (tracked) position.
# a "yield" version of the Parallel Reduction REMAP algorithm.
# the algorithm is in-place. it does not perform "MV" operations.
# instead, where a masked-out value *should* be read from is tracked
def iterate_indices(SVSHAPE, pred=None):
# get indices to iterate over, in the required order
xd = SVSHAPE.lims[0]
# create lists of indices to iterate over in each dimension
ix = list(range(xd))
# invert the indices if needed
if SVSHAPE.invxyz[0]: ix.reverse()
# start a loop from the lowest step
step = 1
steps = []
while step < xd:
step *= 2
steps.append(step)
# invert the indices if needed
if SVSHAPE.invxyz[1]: steps.reverse()
for step in steps:
stepend = (step == steps[-1]) # note end of steps
idxs = list(range(0, xd, step))
results = []
for i in idxs:
other = i + step // 2
ci = ix[i]
oi = ix[other] if other < xd else None
other_pred = other < xd and (pred is None or pred[oi])
if (pred is None or pred[ci]) and other_pred:
if SVSHAPE.skip == 0b00: # submode 00
result = ci
elif SVSHAPE.skip == 0b01: # submode 01
result = oi
results.append([result + SVSHAPE.offset, 0])
elif other_pred:
ix[i] = oi
if results:
results[-1][1] = (stepend<<1) | 1 # notify end of loops
yield from results
def demo():
# set the dimension sizes here
xdim = 9
# set up an SVSHAPE
class SVSHAPE:
pass
SVSHAPE0 = SVSHAPE()
SVSHAPE0.lims = [xdim, 0, 0]
SVSHAPE0.order = [0,1,2]
SVSHAPE0.mode = 0b10
SVSHAPE0.skip = 0b00
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
SVSHAPE1 = SVSHAPE()
SVSHAPE1.lims = [xdim, 0, 0]
SVSHAPE1.order = [0,1,2]
SVSHAPE1.mode = 0b10
SVSHAPE1.skip = 0b01
SVSHAPE1.offset = 0 # experiment with different offset, here
SVSHAPE1.invxyz = [0,0,0] # inversion if desired
# enumerate over the iterator function, getting new indices
shapes = list(iterate_indices(SVSHAPE0)), \
list(iterate_indices(SVSHAPE1))
for idx in range(len(shapes[0])):
l = shapes[0][idx]
r = shapes[1][idx]
(l_idx, lend) = l
(r_idx, rend) = r
print ("%d->%d:%d" % (idx, l_idx, r_idx),
"end", bin(lend)[2:], bin(rend)[2:])
# run the demo
if __name__ == '__main__':
demo()
REMAP FFT pseudocode
The FFT REMAP is RADIX2 only.
# a "yield" version of the REMAP algorithm, for FFT Tukey-Cooley schedules
# original code for the FFT Tukey-Cooley schedule:
# https://www.nayuki.io/res/free-small-fft-in-multiple-languages/fft.py
"""
# Radix-2 decimation-in-time FFT (real, not complex)
size = 2
while size <= n:
halfsize = size // 2
tablestep = n // size
for i in range(0, n, size):
k = 0
for j in range(i, i + halfsize):
jh = j+halfsize
jl = j
temp1 = vec[jh] * exptable[k]
temp2 = vec[jl]
vec[jh] = temp2 - temp1
vec[jl] = temp2 + temp1
k += tablestep
size *= 2
"""
# python "yield" can be iterated. use this to make it clear how
# the indices are generated by using natural-looking nested loops
def iterate_butterfly_indices(SVSHAPE):
# get indices to iterate over, in the required order
n = SVSHAPE.lims[0]
stride = SVSHAPE.lims[2] # stride-multiplier on reg access
# creating lists of indices to iterate over in each dimension
# has to be done dynamically, because it depends on the size
# first, the size-based loop (which can be done statically)
x_r = []
size = 2
while size <= n:
x_r.append(size)
size *= 2
# invert order if requested
if SVSHAPE.invxyz[0]: x_r.reverse()
if len(x_r) == 0:
return
# start an infinite (wrapping) loop
skip = 0
while True:
for size in x_r: # loop over 3rd order dimension (size)
x_end = size == x_r[-1]
# y_r schedule depends on size
halfsize = size // 2
tablestep = n // size
y_r = []
for i in range(0, n, size):
y_r.append(i)
# invert if requested
if SVSHAPE.invxyz[1]: y_r.reverse()
for i in y_r: # loop over 2nd order dimension
y_end = i == y_r[-1]
k_r = []
j_r = []
k = 0
for j in range(i, i+halfsize):
k_r.append(k)
j_r.append(j)
k += tablestep
# invert if requested
if SVSHAPE.invxyz[2]: k_r.reverse()
if SVSHAPE.invxyz[2]: j_r.reverse()
for j, k in zip(j_r, k_r): # loop over 1st order dimension
z_end = j == j_r[-1]
# now depending on MODE return the index
if SVSHAPE.skip == 0b00:
result = j # for vec[j]
elif SVSHAPE.skip == 0b01:
result = j + halfsize # for vec[j+halfsize]
elif SVSHAPE.skip == 0b10:
result = k # for exptable[k]
loopends = (z_end |
((y_end and z_end)<<1) |
((y_end and x_end and z_end)<<2))
yield (result * stride) + SVSHAPE.offset, loopends
def demo():
# set the dimension sizes here
xdim = 8
ydim = 0 # not needed
zdim = 1 # stride must be set to 1
# set total. err don't know how to calculate how many there are...
# do it manually for now
VL = 0
size = 2
n = xdim
while size <= n:
halfsize = size // 2
tablestep = n // size
for i in range(0, n, size):
for j in range(i, i + halfsize):
VL += 1
size *= 2
# set up an SVSHAPE
class SVSHAPE:
pass
# j schedule
SVSHAPE0 = SVSHAPE()
SVSHAPE0.lims = [xdim, ydim, zdim]
SVSHAPE0.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b01
SVSHAPE0.skip = 0b00
SVSHAPE0.offset = 0 # experiment with different offset, here
SVSHAPE0.invxyz = [0,0,0] # inversion if desired
# j+halfstep schedule
SVSHAPE1 = SVSHAPE()
SVSHAPE1.lims = [xdim, ydim, zdim]
SVSHAPE1.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b01
SVSHAPE1.skip = 0b01
SVSHAPE1.offset = 0 # experiment with different offset, here
SVSHAPE1.invxyz = [0,0,0] # inversion if desired
# k schedule
SVSHAPE2 = SVSHAPE()
SVSHAPE2.lims = [xdim, ydim, zdim]
SVSHAPE2.order = [0,1,2] # experiment with different permutations, here
SVSHAPE0.mode = 0b01
SVSHAPE2.skip = 0b10
SVSHAPE2.offset = 0 # experiment with different offset, here
SVSHAPE2.invxyz = [0,0,0] # inversion if desired
# enumerate over the iterator function, getting new indices
schedule = []
for idx, (jl, jh, k) in enumerate(zip(iterate_butterfly_indices(SVSHAPE0),
iterate_butterfly_indices(SVSHAPE1),
iterate_butterfly_indices(SVSHAPE2))):
if idx >= VL:
break
schedule.append((jl, jh, k))
# ok now pretty-print the results, with some debug output
size = 2
idx = 0
while size <= n:
halfsize = size // 2
tablestep = n // size
print ("size %d halfsize %d tablestep %d" % \
(size, halfsize, tablestep))
for i in range(0, n, size):
prefix = "i %d\t" % i
k = 0
for j in range(i, i + halfsize):
(jl, je), (jh, he), (ks, ke) = schedule[idx]
print (" %-3d\t%s j=%-2d jh=%-2d k=%-2d -> "
"j[jl=%-2d] j[jh=%-2d] ex[k=%d]" % \
(idx, prefix, j, j+halfsize, k,
jl, jh, ks,
),
"end", bin(je)[2:], bin(je)[2:], bin(ke)[2:])
k += tablestep
idx += 1
size *= 2
# run the demo
if __name__ == '__main__':
demo()
DCT REMAP
DCT REMAP is RADIX2 only. Convolutions may be applied as usual to create non-RADIX2 DCT. Combined with appropriate Twin-butterfly instructions the algorithm below (written in python3), becomes part of an in-place in-registers Vectorized DCT. The algorithms work by loading data such that as the nested loops progress the result is sorted into correct sequential order.
# DCT "REMAP" scheduler to create an in-place iterative DCT.
#
# bits of the integer 'val' of width 'width' are reversed
def reverse_bits(val, width):
result = 0
for _ in range(width):
result = (result << 1) | (val & 1)
val >>= 1
return result
# iterative version of [recursively-applied] half-reversing
# turns out this is Gray-Encoding.
def halfrev2(vec, pre_rev=True):
res = []
for i in range(len(vec)):
if pre_rev:
res.append(vec[i ^ (i>>1)])
else:
ri = i
bl = i.bit_length()
for ji in range(1, bl):
ri ^= (i >> ji)
res.append(vec[ri])
return res
def iterate_dct_inner_halfswap_loadstore(SVSHAPE):
# get indices to iterate over, in the required order
n = SVSHAPE.lims[0]
mode = SVSHAPE.lims[1]
stride = SVSHAPE.lims[2]
# reference list for not needing to do data-swaps, just swap what
# *indices* are referenced (two levels of indirection at the moment)
# pre-reverse the data-swap list so that it *ends up* in the order 0123..
ji = list(range(n))
levels = n.bit_length() - 1
ri = [reverse_bits(i, levels) for i in range(n)]
if SVSHAPE.mode == 0b01: # FFT, bitrev only
ji = [ji[ri[i]] for i in range(n)]
elif SVSHAPE.submode2 == 0b001:
ji = [ji[ri[i]] for i in range(n)]
ji = halfrev2(ji, True)
else:
ji = halfrev2(ji, False)
ji = [ji[ri[i]] for i in range(n)]
# invert order if requested
if SVSHAPE.invxyz[0]:
ji.reverse()
for i, jl in enumerate(ji):
y_end = jl == ji[-1]
yield jl * stride, (0b111 if y_end else 0b000)
def iterate_dct_inner_costable_indices(SVSHAPE):
# get indices to iterate over, in the required order
n = SVSHAPE.lims[0]
mode = SVSHAPE.lims[1]
stride = SVSHAPE.lims[2]
# creating lists of indices to iterate over in each dimension
# has to be done dynamically, because it depends on the size
# first, the size-based loop (which can be done statically)
x_r = []
size = 2
while size <= n:
x_r.append(size)
size *= 2
# invert order if requested
if SVSHAPE.invxyz[0]:
x_r.reverse()
if len(x_r) == 0:
return
# start an infinite (wrapping) loop
skip = 0
z_end = 1 # doesn't exist in this, only 2 loops
k = 0
while True:
for size in x_r: # loop over 3rd order dimension (size)
x_end = size == x_r[-1]
# y_r schedule depends on size
halfsize = size // 2
y_r = []
for i in range(0, n, size):
y_r.append(i)
# invert if requested
if SVSHAPE.invxyz[1]: y_r.reverse()
# two lists of half-range indices, e.g. j 0123, jr 7654
j = list(range(0, halfsize))
# invert if requested
if SVSHAPE.invxyz[2]: j_r.reverse()
# loop over 1st order dimension
for ci, jl in enumerate(j):
y_end = jl == j[-1]
# now depending on MODE return the index. inner butterfly
if SVSHAPE.skip == 0b00: # in [0b00, 0b10]:
result = k # offset into COS table
elif SVSHAPE.skip == 0b10: #
result = ci # coefficient helper
elif SVSHAPE.skip == 0b11: #
result = size # coefficient helper
loopends = (z_end |
((y_end and z_end)<<1) |
((y_end and x_end and z_end)<<2))
yield (result * stride) + SVSHAPE.offset, loopends
k += 1
def iterate_dct_inner_butterfly_indices(SVSHAPE):
# get indices to iterate over, in the required order
n = SVSHAPE.lims[0]
mode = SVSHAPE.lims[1]
stride = SVSHAPE.lims[2]
# creating lists of indices to iterate over in each dimension
# has to be done dynamically, because it depends on the size
# first, the size-based loop (which can be done statically)
x_r = []
size = 2
while size <= n:
x_r.append(size)
size *= 2
# invert order if requested
if SVSHAPE.invxyz[0]:
x_r.reverse()
if len(x_r) == 0:
return
# reference (read/write) the in-place data in *reverse-bit-order*
ri = list(range(n))
if SVSHAPE.submode2 == 0b01:
levels = n.bit_length() - 1
ri = [ri[reverse_bits(i, levels)] for i in range(n)]
# reference list for not needing to do data-swaps, just swap what
# *indices* are referenced (two levels of indirection at the moment)
# pre-reverse the data-swap list so that it *ends up* in the order 0123..
ji = list(range(n))
inplace_mode = True
if inplace_mode and SVSHAPE.submode2 == 0b01:
ji = halfrev2(ji, True)
if inplace_mode and SVSHAPE.submode2 == 0b11:
ji = halfrev2(ji, False)
# start an infinite (wrapping) loop
while True:
k = 0
k_start = 0
for size in x_r: # loop over 3rd order dimension (size)
x_end = size == x_r[-1]
# y_r schedule depends on size
halfsize = size // 2
y_r = []
for i in range(0, n, size):
y_r.append(i)
# invert if requested
if SVSHAPE.invxyz[1]: y_r.reverse()
for i in y_r: # loop over 2nd order dimension
y_end = i == y_r[-1]
# two lists of half-range indices, e.g. j 0123, jr 7654
j = list(range(i, i + halfsize))
jr = list(range(i+halfsize, i + size))
jr.reverse()
# invert if requested
if SVSHAPE.invxyz[2]:
j.reverse()
jr.reverse()
hz2 = halfsize // 2 # zero stops reversing 1-item lists
# loop over 1st order dimension
k = k_start
for ci, (jl, jh) in enumerate(zip(j, jr)):
z_end = jl == j[-1]
# now depending on MODE return the index. inner butterfly
if SVSHAPE.skip == 0b00: # in [0b00, 0b10]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jl]] # lower half
else:
result = ri[ji[jl]] # lower half
elif SVSHAPE.skip == 0b01: # in [0b01, 0b11]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jl+halfsize]] # upper half
else:
result = ri[ji[jh]] # upper half
elif mode == 4:
# COS table pre-generated mode
if SVSHAPE.skip == 0b10: #
result = k # cos table offset
else: # mode 2
# COS table generated on-demand ("Vertical-First") mode
if SVSHAPE.skip == 0b10: #
result = ci # coefficient helper
elif SVSHAPE.skip == 0b11: #
result = size # coefficient helper
loopends = (z_end |
((y_end and z_end)<<1) |
((y_end and x_end and z_end)<<2))
yield (result * stride) + SVSHAPE.offset, loopends
k += 1
# now in-place swap
if inplace_mode:
for ci, (jl, jh) in enumerate(zip(j[:hz2], jr[:hz2])):
jlh = jl+halfsize
tmp1, tmp2 = ji[jlh], ji[jh]
ji[jlh], ji[jh] = tmp2, tmp1
# new k_start point for cos tables( runs inside x_r loop NOT i loop)
k_start += halfsize
# python "yield" can be iterated. use this to make it clear how
# the indices are generated by using natural-looking nested loops
def iterate_dct_outer_butterfly_indices(SVSHAPE):
# get indices to iterate over, in the required order
n = SVSHAPE.lims[0]
mode = SVSHAPE.lims[1]
stride = SVSHAPE.lims[2]
# creating lists of indices to iterate over in each dimension
# has to be done dynamically, because it depends on the size
# first, the size-based loop (which can be done statically)
x_r = []
size = n // 2
while size >= 2:
x_r.append(size)
size //= 2
# invert order if requested
if SVSHAPE.invxyz[0]:
x_r.reverse()
if len(x_r) == 0:
return
# I-DCT, reference (read/write) the in-place data in *reverse-bit-order*
ri = list(range(n))
if SVSHAPE.submode2 in [0b11, 0b01]:
levels = n.bit_length() - 1
ri = [ri[reverse_bits(i, levels)] for i in range(n)]
# reference list for not needing to do data-swaps, just swap what
# *indices* are referenced (two levels of indirection at the moment)
# pre-reverse the data-swap list so that it *ends up* in the order 0123..
ji = list(range(n))
inplace_mode = False # need the space... SVSHAPE.skip in [0b10, 0b11]
if SVSHAPE.submode2 == 0b11:
ji = halfrev2(ji, False)
# start an infinite (wrapping) loop
while True:
k = 0
k_start = 0
for size in x_r: # loop over 3rd order dimension (size)
halfsize = size//2
x_end = size == x_r[-1]
y_r = list(range(0, halfsize))
# invert if requested
if SVSHAPE.invxyz[1]: y_r.reverse()
for i in y_r: # loop over 2nd order dimension
y_end = i == y_r[-1]
# one list to create iterative-sum schedule
jr = list(range(i+halfsize, i+n-halfsize, size))
# invert if requested
if SVSHAPE.invxyz[2]: jr.reverse()
hz2 = halfsize // 2 # zero stops reversing 1-item lists
k = k_start
for ci, jh in enumerate(jr): # loop over 1st order dimension
z_end = jh == jr[-1]
if mode == 4:
# COS table pre-generated mode
if SVSHAPE.skip == 0b00: # in [0b00, 0b10]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jh]] # upper half
else:
result = ri[ji[jh]] # lower half
elif SVSHAPE.skip == 0b01: # in [0b01, 0b11]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jh+size]] # upper half
else:
result = ri[ji[jh+size]] # upper half
elif SVSHAPE.skip == 0b10: #
result = k # cos table offset
else:
# COS table generated on-demand ("Vertical-First") mode
if SVSHAPE.skip == 0b00: # in [0b00, 0b10]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jh]] # lower half
else:
result = ri[ji[jh]] # lower half
elif SVSHAPE.skip == 0b01: # in [0b01, 0b11]:
if SVSHAPE.submode2 == 0b11: # iDCT
result = ji[ri[jh+size]] # upper half
else:
result = ri[ji[jh+size]] # upper half
elif SVSHAPE.skip == 0b10: #
result = ci # coefficient helper
elif SVSHAPE.skip == 0b11: #
result = size # coefficient helper
loopends = (z_end |
((y_end and z_end)<<1) |
((y_end and x_end and z_end)<<2))
yield (result * stride) + SVSHAPE.offset, loopends
k += 1
# new k_start point for cos tables( runs inside x_r loop NOT i loop)
k_start += halfsize
REMAP selector
Selecting which REMAP Schedule to use is shown by the pseudocode below. Each SVSHAPE 0-3 goes through this selection process.
if SVSHAPEn.mode == 0b00: iterate_fn = iterate_indices
if SVSHAPEn.mode == 0b10: iterate_fn = iterate_preduce_indices
if SVSHAPEn.mode in [0b01, 0b11]:
# further sub-selection
if SVSHAPEn.ydimsz == 1: iterate_fn = iterate_butterfly_indices
if SVSHAPEn.ydimsz == 2: iterate_fn = iterate_dct_inner_butterfly_indices
if SVSHAPEn.ydimsz == 3: iterate_fn = iterate_dct_outer_butterfly_indices
if SVSHAPEn.ydimsz in [5, 13]: iterate_fn = iterate_dct_inner_costable_indices
if SVSHAPEn.ydimsz in [6, 14, 15]: iterate_fn = iterate_dct_inner_halfswap_loadstore