1 \documentclass[slidestop
]{beamer
}
2 \usepackage{beamerthemesplit
}
6 \title{Simple-V RISC-V Extension for Vectorisation and SIMD
}
7 \author{Luke Kenneth Casson Leighton
}
14 \huge{Simple-V RISC-V Extension for Vectors and SIMD
}\\
16 \Large{Flexible Vectorisation
}\\
17 \Large{(aka not so Simple-V?)
}\\
19 \Large{[proposed for
] Chennai
9th RISC-V Workshop
}\\
26 \frame{\frametitle{Credits and Acknowledgements
}
29 \item The Designers of RISC-V
\vspace{15pt
}
30 \item The RVV Working Group and contributors
\vspace{15pt
}
31 \item Allen Baum, Jacob Bachmeyer, Xan Phung, Chuanhua Chang,\\
32 Guy Lemurieux, Jonathan Neuschafer, Roger Brussee,
33 and others
\vspace{15pt
}
34 \item ISA-Dev Group Members
\vspace{10pt
}
39 \frame{\frametitle{Quick refresher on SIMD
}
42 \item SIMD very easy to implement (and very seductive)
\vspace{10pt
}
43 \item Parallelism is in the ALU
\vspace{10pt
}
44 \item Zero-to-Negligeable impact for rest of core
\vspace{10pt
}
46 Where SIMD Goes Wrong:
\vspace{10pt
}
48 \item See "SIMD instructions considered harmful"
49 https://www.sigarch.org/simd-instructions-considered-harmful
50 \item Corner-cases alone are extremely complex.\\
51 Hardware is easy, but software is hell.
52 \item O($N^
{6}$) ISA opcode proliferation!\\
53 opcode, elwidth, veclen, src1-src2-dest hi/lo
57 \frame{\frametitle{Quick refresher on RVV
}
60 \item Extremely powerful (extensible to
256 registers)
\vspace{10pt
}
61 \item Supports polymorphism, several datatypes (inc. FP16)
\vspace{10pt
}
62 \item Requires a separate Register File (
32 w/ext to
256)
\vspace{10pt
}
63 \item Implemented as a separate pipeline (no impact on scalar)
\vspace{10pt
}
65 However...
\vspace{10pt
}
67 \item 98 percent opcode duplication with rest of RV (CLIP)
68 \item Extending RVV requires customisation not just of h/w:\\
69 gcc and s/w also need customisation (and maintenance)
74 \frame{\frametitle{The Simon Sinek lowdown (Why, How, What)
}
78 Implementors need flexibility in vectorisation to optimise for
79 area or performance depending on the scope:
80 embedded DSP, Mobile GPU's, Server CPU's and more.
\vspace{4pt
}\\
81 Compilers also need flexibility in vectorisation to optimise for cost
82 of pipeline setup, amount of state to context switch
83 and software portability
\vspace{4pt
}
85 By marking INT/FP regs as "Vectorised" and
86 adding a level of indirection,
87 SV expresses how existing instructions should act
88 on
[contiguous
] blocks of registers, in parallel.
\vspace{4pt
}
90 Simple-V is an "API" that implicitly extends
91 existing (scalar) instructions with explicit parallelisation
92 (i.e. SV is actually about parallelism NOT vectors per se)
97 \frame{\frametitle{What's the value of SV? Why adopt it even in non-V?
}
100 \item memcpy becomes much smaller (higher bang-per-buck)
101 \item context-switch (LOAD/STORE multiple):
1-
2 instructions
102 \item Compressed instrs further reduces I-cache (etc.)
103 \item Greatly-reduced I-cache load (and less reads)
104 \item Amazingly, SIMD becomes (more) tolerable\\
105 (corner-cases for setup and teardown are gone)
109 \item It's not just about Vectors: it's about instruction effectiveness
110 \item Anything that makes SIMD tolerable has to be a good thing
111 \item Anything implementor is not interested in HW-optimising,\\
112 let it fall through to exceptions (implement as a trap).
117 \frame{\frametitle{How does Simple-V relate to RVV? What's different?
}
120 \item RVV very heavy-duty (excellent for supercomputing)
\vspace{10pt
}
121 \item Simple-V abstracts parallelism (based on best of RVV)
\vspace{10pt
}
122 \item Graded levels: hardware, hybrid or traps (fit impl. need)
\vspace{10pt
}
123 \item Even Compressed become vectorised (RVV can't)
\vspace{10pt
}
125 What Simple-V is not:
\vspace{10pt
}
127 \item A full supercomputer-level Vector Proposal
128 \item A replacement for RVV (SV is designed to be over-ridden\\
129 by - or augmented to become, or just be replaced by - RVV)
134 \frame{\frametitle{How is Parallelism abstracted in Simple-V?
}
137 \item Register "typing" turns any op into an implicit Vector op:\\
138 registers are reinterpreted through a level of indirection
139 \item Primarily at the Instruction issue phase (except SIMD)\\
140 Note: it's ok to pass predication through to ALU (like SIMD)
141 \item Standard (and future, and custom) opcodes now parallel
\vspace{10pt
}
143 Note: EVERYTHING is parallelised:
145 \item All LOAD/STORE (inc. Compressed, Int/FP versions)
146 \item All ALU ops (soft / hybrid / full HW, on per-op basis)
147 \item All branches become predication targets (C.FNE added?)
148 \item C.MV of particular interest (s/v, v/v, v/s)
149 \item FCVT, FMV, FSGNJ etc. very similar to C.MV
154 \frame{\frametitle{Implementation Options
}
157 \item Absolute minimum: Exceptions (if CSRs indicate "V", trap)
158 \item Hardware loop, single-instruction issue\\
159 (Do / Don't send through predication to ALU)
160 \item Hardware loop, parallel (multi-instruction) issue\\
161 (Do / Don't send through predication to ALU)
162 \item Hardware loop, full parallel ALU (not recommended)
166 \item 4 (or more?) options above may be deployed on per-op basis
167 \item SIMD always sends predication bits through to ALU
168 \item Minimum MVL MUST be sufficient to cover regfile LD/ST
169 \item Instr. FIFO may repeatedly split off N scalar ops at a time
172 % Instr. FIFO may need its own slide. Basically, the vectorised op
173 % gets pushed into the FIFO, where it is then "processed". Processing
174 % will remove the first set of ops from its vector numbering (taking
175 % predication into account) and shoving them **BACK** into the FIFO,
176 % but MODIFYING the remaining "vectorised" op, subtracting the now
177 % scalar ops from it.
179 \frame{\frametitle{Predicated
8-parallel ADD:
1-wide ALU
}
181 \includegraphics[height=
2.5in
]{padd9_alu1.png
}\\
182 {\bf \red Predicated adds are shuffled down:
6 cycles in total
}
187 \frame{\frametitle{Predicated
8-parallel ADD:
4-wide ALU
}
189 \includegraphics[height=
2.5in
]{padd9_alu4.png
}\\
190 {\bf \red Predicated adds are shuffled down:
4 in
1st cycle,
2 in
2nd
}
195 \frame{\frametitle{Predicated
8-parallel ADD:
3 phase FIFO expansion
}
197 \includegraphics[height=
2.5in
]{padd9_fifo.png
}\\
198 {\bf \red First cycle takes first four
1s; second takes the rest
}
203 \frame{\frametitle{How are SIMD Instructions Vectorised?
}
206 \item SIMD ALU(s) primarily unchanged
\vspace{6pt
}
207 \item Predication is added to each SIMD element
\vspace{6pt
}
208 \item Predication bits sent in groups to the ALU
\vspace{6pt
}
209 \item End of Vector enables (additional) predication
\vspace{10pt
}
211 Considerations:
\vspace{4pt
}
213 \item Many SIMD ALUs possible (parallel execution)
214 \item Implementor free to choose (API remains the same)
215 \item Unused ALU units wasted, but s/w DRASTICALLY simpler
216 \item Very long SIMD ALUs could waste significant die area
219 % With multiple SIMD ALUs at for example 32-bit wide they can be used
220 % to either issue 64-bit or 128-bit or 256-bit wide SIMD operations
221 % or they can be used to cover several operations on totally different
222 % vectors / registers.
224 \frame{\frametitle{Predicated
9-parallel SIMD ADD
}
226 \includegraphics[height=
2.5in
]{padd9_simd.png
}\\
227 {\bf \red 4-wide
8-bit SIMD,
4 bits of predicate passed to ALU
}
232 \frame{\frametitle{What's the deal / juice / score?
}
235 \item Standard Register File(s) overloaded with CSR "reg is vector"\\
236 (see pseudocode slides for examples)
237 \item Element width (and type?) concepts remain same as RVV\\
238 (CSRs are used to "interpret" elements in registers)
239 \item CSRs are key-value tables (overlaps allowed)
\vspace{10pt
}
241 Key differences from RVV:
\vspace{10pt
}
243 \item Predication in INT regs as a BIT field (max VL=XLEN)
244 \item Minimum VL must be Num Regs -
1 (all regs single LD/ST)
245 \item SV may condense sparse Vecs: RVV lets ALU do predication
246 \item Choice to Zero or skip non-predicated elements
251 \begin{frame
}[fragile
]
252 \frametitle{ADD pseudocode (or trap, or actual hardware loop)
}
255 function op_add(rd, rs1, rs2, predr) # add not VADD!
256 Â int i, id=
0, irs1=
0, irs2=
0;
257 Â for (i =
0; i < VL; i++)
258 Â if (ireg
[predr
] &
1<<i) # predication uses intregs
259 Â Â ireg
[rd+id
] <= ireg
[rs1+irs1
] + ireg
[rs2+irs2
];
260 if (reg_is_vectorised
[rd
])Â \
{ id +=
1; \
}
261 if (reg_is_vectorised
[rs1
])Â \
{ irs1 +=
1; \
}
262 if (reg_is_vectorised
[rs2
])Â \
{ irs2 +=
1; \
}
266 \item SIMD slightly more complex (case above is elwidth = default)
267 \item Scalar-scalar and scalar-vector and vector-vector now all in one
268 \item OoO may choose to push ADDs into instr. queue (v. busy!)
272 % yes it really *is* ADD not VADD. that's the entire point of
273 % this proposal, that *standard* operations are overloaded to
274 % become vectorised-on-demand
277 \begin{frame
}[fragile
]
278 \frametitle{Predication-Branch (or trap, or actual hardware loop)
}
281 s1 = reg_is_vectorised(src1);
282 s2 = reg_is_vectorised(src2);
283 if (!s2 && !s1) goto branch;
284 for (int i =
0; i < VL; ++i)
285 if cmp(s1 ? reg
[src1+i
] : reg
[src1
],
286 s2 ? reg
[src2+i
] : reg
[src2
])
291 \item SIMD slightly more complex (case above is elwidth = default)
292 \item If s1 and s2 both scalars, Standard branch occurs
293 \item Predication stored in integer regfile as a bitfield
294 \item Scalar-vector and vector-vector supported
298 \begin{frame
}[fragile
]
299 \frametitle{VLD/VLD.S/VLD.X (or trap, or actual hardware loop)
}
302 if (unit-strided) stride = elsize;
303 else stride = areg
[as2
]; // constant-strided
304 for (int i =
0; i < VL; ++i)
305 if (preg_enabled
[rd
] && (
[!
]preg
[rd
] &
1<<i))
306 for (int j =
0; j < seglen+
1; j++)
307 if (reg_is_vectorised
[rs2
]) offs = vreg
[rs2+i
]
308 else offs = i*(seglen+
1)*stride;
309 vreg
[rd+j
][i
] = mem
[sreg
[base
] + offs + j*stride
]
313 \item Again: elwidth != default slightly more complex
314 \item rs2 vectorised taken to implicitly indicate VLD.X
319 \frame{\frametitle{Why are overlaps allowed in Regfiles?
}
322 \item Same register(s) can have multiple "interpretations"
323 \item Set "real" register (scalar) without needing to set/unset CSRs.
324 \item xBitManip plus SIMD plus xBitManip = Hi/Lo bitops
325 \item (
32-bit GREV plus
4x8-bit SIMD plus
32-bit GREV:\\
326 GREV @ VL=N,wid=
32; SIMD @ VL=Nx4,wid=
8)
327 \item RGB
565 (video): BEXTW plus
4x8-bit SIMD plus BDEPW\\
328 (BEXT/BDEP @ VL=N,wid=
32; SIMD @ VL=Nx4,wid=
8)
329 \item Same register(s) can be offset (no need for VSLIDE)
\vspace{6pt
}
333 \item xBitManip reduces O($N^
{6}$) SIMD down to O($N^
{3}$)
334 \item Hi-Performance: Macro-op fusion (more pipeline stages?)
339 \frame{\frametitle{To Zero or not to place zeros in non-predicated elements?
}
342 \item Zeroing is an implementation optimisation favouring OoO
343 \item Simple implementations may skip non-predicated operations
344 \item Simple implementations explicitly have to destroy data
345 \item Complex implementations may use reg-renames to save power\\
346 Zeroing on predication chains makes optimisation harder
347 \item Compromise: REQUIRE both (specified in predication CSRs).
351 \item Complex not really impacted, simple impacted a LOT\\
352 with Zeroing... however it's useful (memzero)
353 \item Non-zero'd overlapping "Vectors" may issue overlapping ops\\
354 (
2nd op's predicated elements slot in
1st's non-predicated ops)
355 \item Please don't use Vectors for "security" (use Sec-Ext)
358 % with overlapping "vectors" - bearing in mind that "vectors" are
359 % just a remap onto the standard register file, if the top bits of
360 % predication are zero, and there happens to be a second vector
361 % that uses some of the same register file that happens to be
362 % predicated out, the second vector op may be issued *at the same time*
363 % if there are available parallel ALUs to do so.
366 \frame{\frametitle{Predication key-value CSR store
}
369 \item key is int regfile number or FP regfile number (
1 bit)
\vspace{6pt
}
370 \item register to be predicated if referred to (
5 bits, key)
\vspace{6pt
}
371 \item register to store actual predication in (
5 bits, value)
\vspace{6pt
}
372 \item predication is inverted Y/N (
1 bit)
\vspace{6pt
}
373 \item non-predicated elements are to be zero'd Y/N (
1 bit)
\vspace{6pt
}
377 \item Table should be expanded out for high-speed implementations
378 \item Multiple "keys" (and values) theoretically permitted
379 \item RVV rules about deleting higher-indexed CSRs followed
384 \begin{frame
}[fragile
]
385 \frametitle{Predication key-value CSR table decoding pseudocode
}
388 struct pred fp_pred
[32];
389 struct pred int_pred
[32];
391 for (i =
0; i <
16; i++) //
16 CSRs?
392 tb = int
\_pred if CSRpred
[i
].type ==
0 else fp
\_pred
393 idx = CSRpred
[i
].regidx
394 tb
[idx
].zero = CSRpred
[i
].zero
395 tb
[idx
].inv = CSRpred
[i
].inv
396 tb
[idx
].predidx = CSRpred
[i
].predidx
397 tb
[idx
].enabled = true
401 \item All
64 (int and FP) Entries zero'd before setting
402 \item Might be a bit complex to set up (TBD)
408 \frame{\frametitle{Register key-value CSR store
}
411 \item key is int regfile number or FP regfile number (
1 bit)
\vspace{6pt
}
412 \item treated as vector if referred to in op (
5 bits, key)
\vspace{6pt
}
413 \item starting register to actually be used (
5 bits, value)
\vspace{6pt
}
414 \item element bitwidth: default/
8/
16/
32/
64/rsvd (
3 bits)
\vspace{6pt
}
415 \item element type: still under consideration
\vspace{6pt
}
419 \item Same notes apply (previous slide) as for predication CSR table
420 \item Level of indirection has implications for pipeline latency
425 \frame{\frametitle{Register key-value CSR pseudocode
}
433 \frame{\frametitle{C.MV extremely flexible!
}
436 \item scalar-to-vector (w/ no pred): VSPLAT
437 \item scalar-to-vector (w/ dest-pred): Sparse VSPLAT
438 \item scalar-to-vector (w/
1-bit dest-pred): VINSERT
439 \item vector-to-scalar (w/
[1-bit?
] src-pred): VEXTRACT
440 \item vector-to-vector (w/ no pred): Vector Copy
441 \item vector-to-vector (w/ src pred): Vector Gather
442 \item vector-to-vector (w/ dest pred): Vector Scatter
443 \item vector-to-vector (w/ src \& dest pred): Vector Gather/Scatter
448 \item Surprisingly powerful!
449 \item Same arrangement for FVCT, FMV, FSGNJ etc.
454 \frame{\frametitle{Opcodes, compared to RVV
}
457 \item All integer and FP opcodes all removed (no CLIP!)
\vspace{8pt
}
458 \item VMPOP, VFIRST etc. all removed (use xBitManip)
\vspace{8pt
}
459 \item VSLIDE removed (use regfile overlaps)
\vspace{8pt
}
460 \item C.MV covers VEXTRACT VINSERT and VSPLAT (and more)
\vspace{8pt
}
461 \item VSETVL, VGETVL, VSELECT stay
\vspace{8pt
}
462 \item Issue: VCLIP is not in RV* (add with custom ext?)
\vspace{8pt
}
463 \item Vector (or scalar-vector) use C.MV (MV is a pseudo-op)
\vspace{8pt
}
464 \item VMERGE: twin predicated C.MVs (one inverted. macro-op'd)
\vspace{8pt
}
469 \frame{\frametitle{Under consideration
}
472 \item Is C.FNE actually needed? Should it be added if it is?
473 \item Element type implies polymorphism. Should it be in SV?
474 \item Should use of registers be allowed to "wrap" (x30 x31 x1 x2)?
475 \item Is detection of all-scalar ops ok (without slowing pipeline)?
476 \item Can VSELECT be removed? (it's really complex)
477 \item Can CLIP be done as a CSR (mode, like elwidth)
478 \item SIMD saturation (etc.) also set as a mode?
479 \item 8/
16-bit ops is it worthwhile adding a "start offset"? \\
480 (a bit like misaligned addressing... for registers)\\
481 or just use predication to skip start?
486 \frame{\frametitle{What's the downside(s) of SV?
}
488 \item EVERY register operation is inherently parallelised\\
489 (scalar ops are just vectors of length
1)
\vspace{4pt
}
490 \item An extra pipeline phase is pretty much essential\\
491 for fast low-latency implementations
\vspace{4pt
}
492 \item Assuming an instruction FIFO, N ops could be taken off\\
493 of a parallel op per cycle (avoids filling entire FIFO;\\
494 also is less work per cycle: lower complexity / latency)
\vspace{4pt
}
495 \item With zeroing off, skipping non-predicated elements is hard:\\
496 it is however an optimisation (and could be skipped).
\vspace{4pt
}
497 \item Setting up the Register/Predication tables (interpreting the\\
498 CSR key-value stores) might be a bit complex to optimise
499 (any change to a CSR key-value entry needs to redo the table)
504 \frame{\frametitle{Is this OK (low latency)? Detect scalar-ops (only)
}
506 \includegraphics[height=
2.5in
]{scalardetect.png
}\\
507 {\bf \red Detect when all registers are scalar for a given op
}
512 \frame{\frametitle{TODO (break into separate slides)
}
515 \item Then explain why this proposal is a good way to \\
516 abstract parallelism\\
517 (hopefully also explaining how \\
518 a good compiler can make clever use of this increase parallelism\\
519 Then explain how this can be implemented (at instruction\\
520 issue time???) with\\
521 implementation options, and what these "cost".\\
522 Finally give examples that show simple usage that compares\\
531 \frame{\frametitle{Summary
}
534 \item Actually about parallelism, not Vectors (or SIMD) per se
535 \item Designed for flexibility (graded levels of complexity)
536 \item Huge range of implementor freedom
537 \item Fits RISC-V ethos: achieve more with less
538 \item Reduces SIMD ISA proliferation by
3-
4 orders of magnitude \\
539 (without SIMD downsides or sacrificing speed trade-off)
540 \item Covers
98\% of RVV, allows RVV to fit "on top"
541 \item Not designed for supercomputing (that's RVV), designed for
542 in between: DSPs, RV32E, Embedded
3D GPUs etc.
543 \item Not specifically designed for Vectorisation: designed to\\
544 reduce code size (increase efficiency, just
550 \frame{\frametitle{slide
}
555 Considerations:
\vspace{10pt
}
564 {\Huge \red The end
\vspace{20pt
}\\