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{The Simon Sinek lowdown (Why, How, What)
}
29 \item Vectorisation needs to fit an implementor's needs:\\
30 RV32E, Embedded/Mobile, DSP, Servers and more.
\vspace{15pt
}
31 \item By implicitly marking INT/FP regs as "Vectorised",\\
32 everything else follows from there.
\vspace{15pt
}
33 \item A Standard Vector "API" with flexibility for implementors:\\
34 choice to optimise for area or performance as desired
\vspace{10pt
}
39 \frame{\frametitle{Why another Vector Extension?
}
42 \item RVV very heavy-duty (excellent for supercomputing)
\vspace{10pt
}
43 \item Simple-V abstracts parallelism (based on best of RVV)
\vspace{10pt
}
44 \item Graded levels: hardware, hybrid or traps
\vspace{10pt
}
45 \item Even Compressed instructions become vectorised
\vspace{10pt
}
47 What Simple-V is not:
\vspace{10pt
}
49 \item A full supercomputer-level Vector Proposal
\vspace{10pt
}
50 \item A replacement for RVV (designed to be augmented)
\vspace{10pt
}
55 \frame{\frametitle{Quick refresher on SIMD
}
58 \item SIMD very easy to implement (and very seductive)
\vspace{10pt
}
59 \item Parallelism is in the ALU
\vspace{10pt
}
60 \item Zero-to-Negligeable impact for rest of core
\vspace{10pt
}
62 Where SIMD Goes Wrong:
\vspace{10pt
}
64 \item See "SIMD instructions considered harmful"
65 https://www.sigarch.org/simd-instructions-considered-harmful
66 \item (Corner-cases alone are extremely complex)
\vspace{10pt
}
67 \item O($N^
{6}$) ISA opcode proliferation!
\vspace{10pt
}
71 \frame{\frametitle{Quick refresher on RVV
}
74 \item Extremely powerful (extensible to
256 registers)
\vspace{10pt
}
75 \item Supports polymorphism, several datatypes (inc. FP16)
\vspace{10pt
}
76 \item Requires a separate Register File
\vspace{10pt
}
77 \item Can be implemented as a separate pipeline
\vspace{10pt
}
79 However...
\vspace{10pt
}
81 \item 98 percent opcode duplication with rest of RV (CLIP)
\vspace{10pt
}
82 \item Extending RVV requires customisation
\vspace{10pt
}
87 \frame{\frametitle{How is Parallelism abstracted?
}
90 \item Almost all opcodes removed in favour of implicit "typing"
\vspace{10pt
}
91 \item Primarily at the Instruction issue phase (except SIMD)
\vspace{10pt
}
92 \item Standard (and future, and custom) opcodes now parallel
\vspace{10pt
}
96 \item LOAD/STORE (inc. C.LD and C.ST, LDX: everything)
\vspace{10pt
}
97 \item All ALU ops (soft / hybrid / full HW, on per-op basis)
\vspace{10pt
}
98 \item All branches become predication targets (C.FNE added)
\vspace{10pt
}
103 \frame{\frametitle{Implementation Options
}
106 \item Absolute minimum: Exceptions (if CSRs indicate "V", trap)
\vspace{10pt
}
107 \item Hardware loop, single-instruction issue
\vspace{10pt
}
108 \item Hardware loop, parallel (multi-instruction) issue
\vspace{10pt
}
109 \item Hardware loop, full parallel ALU (not recommended)
\vspace{10pt
}
113 \item 4 (or more?) options above may be deployed on per-op basis
114 \item Minimum MVL MUST be sufficient to cover regfile LD/ST
115 \item OoO may split off
4+ single-instructions at a time
120 \frame{\frametitle{How are SIMD Instructions Vectorised?
}
123 \item SIMD ALU(s) primarily unchanged
\vspace{10pt
}
124 \item Predication is added to each SIMD element (NO ZEROING!)
\vspace{10pt
}
125 \item End of Vector enables predication (NO ZEROING!)
\vspace{10pt
}
127 Considerations:
\vspace{10pt
}
129 \item Many SIMD ALUs possible (parallel execution)
\vspace{10pt
}
130 \item Very long SIMD ALUs could waste die area (short vectors)
\vspace{10pt
}
131 \item Implementor free to choose (API remains the same)
\vspace{10pt
}
134 % With multiple SIMD ALUs at for example 32-bit wide they can be used
135 % to either issue 64-bit or 128-bit or 256-bit wide SIMD operations
136 % or they can be used to cover several operations on totally different
137 % vectors / registers.
139 \frame{\frametitle{What's the deal / juice / score?
}
142 \item Standard Register File(s) overloaded with "vector span"
\vspace{10pt
}
143 \item Element width and type concepts remain same as RVV
\vspace{10pt
}
144 \item CSRs are key-value tables (overlaps allowed)
\vspace{10pt
}
146 Key differences from RVV:
\vspace{10pt
}
148 \item Predication in INT regs as a BIT field (max VL=XLEN)
\vspace{10pt
}
149 \item Minimum VL must be Num Regs -
1 (all regs single LD/ST)
\vspace{10pt
}
150 \item NO ZEROING: non-predicated elements are skipped
\vspace{10pt
}
155 \frame{\frametitle{Why are overlaps allowed in Regfiles?
}
158 \item Same register(s) can have multiple "interpretations"
\vspace{10pt
}
159 \item xBitManip plus SIMD plus xBitManip = Hi/Lo bitops
\vspace{10pt
}
160 \item (
32-bit GREV plus
4-wide
32-bit SIMD plus
32-bit GREV)
\vspace{10pt
}
161 \item Same register(s) can be offset (no need for VSLIDE)
\vspace{10pt
}
165 \item xBitManip reduces O($N^
{6}$) SIMD down to O($N^
{3}$)
\vspace{10pt
}
166 \item Hi-Performance: Macro-op fusion (more pipeline stages?)
\vspace{10pt
}
171 \frame{\frametitle{Why no Zeroing (place zeros in non-predicated elements)?
}
174 \item Zeroing is an implementation optimisation favouring OoO
\vspace{10pt
}
175 \item Simple implementations may skip non-predicated operations
\vspace{10pt
}
176 \item Simple implementations explicitly have to destroy data
\vspace{10pt
}
177 \item Complex implementations may use reg-renames to save power
\vspace{10pt
}
179 Considerations:
\vspace{10pt
}
181 \item Complex not really impacted, Simple impacted a LOT
182 \item Overlapping "Vectors" may issue overlapping ops
183 \item Please don't use Vectors for "security" (use Sec-Ext)
186 % with overlapping "vectors" - bearing in mind that "vectors" are
187 % just a remap onto the standard register file, if the top bits of
188 % predication are zero, and there happens to be a second vector
189 % that uses some of the same register file that happens to be
190 % predicated out, the second vector op may be issued *at the same time*
191 % if there are available parallel ALUs to do so.
194 \frame{\frametitle{Predication key-value CSR store
}
197 \item key is int regfile number or FP regfile number (
1 bit)
\vspace{10pt
}
198 \item register to be predicated if referred to (
5 bits, key)
\vspace{10pt
}
199 \item register to store actual predication in (
5 bits, value)
\vspace{10pt
}
200 \item predication is inverted (
1 bit)
\vspace{10pt
}
204 \item Table should be expanded out for high-speed implementations
205 \item Multiple "keys" (and values) theoretically permitted
206 \item RVV rules about deleting higher-indexed CSRs followed
211 \begin{frame
}[fragile
]
212 \frametitle{ADD pseudocode (or trap, or actual hardware loop)
}
215 function op_add(rd, rs1, rs2, predr) # add not VADD!
216 int i, id=
0, irs1=
0, irs2=
0;
217 for (i=
0; i < MIN(VL, vectorlen
[rd
]); i++)
218 if (ireg
[predr
] &
1<<i) # predication uses intregs
219 ireg
[rd+id
] <= ireg
[rs1+irs1
] + ireg
[rs2+irs2
];
220 if (reg_is_vectorised
[rd
]) \
{ id +=
1; \
}
221 if (reg_is_vectorised
[rs1
]) \
{ irs1 +=
1; \
}
222 if (reg_is_vectorised
[rs2
]) \
{ irs2 +=
1; \
}
226 \item SIMD slightly more complex (case above is elwidth = default)
227 \item Scalar-scalar and scalar-vector and vector-vector now all in one
228 \item OoO may choose to push ADDs into instr. queue (v. busy!)
232 \begin{frame
}[fragile
]
233 \frametitle{Predication-Branch (or trap, or actual hardware loop)
}
236 s1 = vectorlen
[src1
] >
1;
237 s2 = vectorlen
[src2
] >
1;
238 for (int i =
0; i < VL; ++i)
239 preg
[rs3
] |=
1 << cmp(s1 ? reg
[src1+i
] : reg
[src1
],
240 s2 ? reg
[src2+i
] : reg
[src2
]);
244 \item SIMD slightly more complex (case above is elwidth = default)
245 \item If s1 and s2 both scalars, Standard branch occurs
246 \item Predication stored in integer regfile as a bitfield
247 \item Scalar-vector and vector-vector supported
251 \begin{frame
}[fragile
]
252 \frametitle{LD/LD.S/LD.X (or trap, or actual hardware loop)
}
255 if (unit-strided) stride = elsize;
256 else stride = areg
[as2
]; // constant-strided
257 for (int i =
0; i < VL; ++i)
258 if (preg_enabled
[rd
] && (
[!
]preg
[rd
] &
1<<i))
259 for (int j =
0; j < seglen+
1; j++)
260 if (vectorised
[rs2
]) offs = vreg
[rs2
][i
]
261 else offs = i*(seglen+
1)*stride;
262 vreg
[rd+j
][i
] = mem
[sreg
[base
] + offs + j*stride
]
266 \item Again: SIMD slightly more complex
267 \item rs2 vectorised taken to implicitly indicate LD.X
272 \frame{\frametitle{Opcodes, compared to RVV
}
275 \item All integer and FP opcodes all removed (no CLIP!)
\vspace{10pt
}
276 \item VMPOP, VFIRST etc. all removed (use xBitManip)
\vspace{10pt
}
277 \item VSLIDE, VEXTRACT, VINSERT removed (using regfile)
\vspace{10pt
}
278 \item VSETVL, VGETVL, VSELECT stay
\vspace{10pt
}
279 \item Issue: VCLIP is not in RV* (add with custom ext?)
\vspace{10pt
}
280 \item Vector (or scalar-vector) use C.MV (MV is a pseudo-op)
\vspace{10pt
}
281 \item VMERGE: twin predicated C.MVs (one inverted. macro-op'd)
\vspace{10pt
}
286 \frame{\frametitle{Under consideration
}
289 \item Can VSELECT be removed? (it's really complex)
\vspace{10pt
}
290 \item Can CLIP be done as a CSR (mode, like elwidth)
\vspace{10pt
}
291 \item SIMD saturation (etc.) also set as a mode?
\vspace{10pt
}
292 \item 8/
16-bit ops is it worthwhile adding a "start offset"? \\
293 (a bit like misaligned addressing... for registers)
\vspace{10pt
}
298 \frame{\frametitle{Summary
}
301 \item Designed for simplicity (graded levels of complexity)
\vspace{10pt
}
302 \item Fits RISC-V ethos: do more with less
\vspace{10pt
}
303 \item Reduces SIMD ISA proliferation by
3-
4 orders of magnitude \\
304 (without SIMD downsides or sacrificing speed trade-off)
\vspace{10pt
}
305 \item Covers
98\% of RVV, allows RVV to fit "on top"
\vspace{10pt
}
306 \item Huge range of implementor freedom and flexibility
\vspace{10pt
}
307 \item Not designed for supercomputing (that's RVV), designed for
308 in between: DSPs, RV32E, Embedded
3D GPUs etc.
\vspace{10pt
}
313 \frame{\frametitle{slide
}
318 Considerations:
\vspace{10pt
}
325 \frame{\frametitle{Including a plot
}
327 % \includegraphics[height=2in]{dental.ps}\\
328 {\bf \red Dental trajectories for
27 children:
}
332 \frame{\frametitle{Creating .pdf slides in WinEdt
}
335 \item LaTeX
[Shift-Control-L
]\vspace{10pt
}
336 \item dvi2pdf
[click the button
]\vspace{24pt
}
338 To print
4 slides per page in acrobat click
\vspace{10pt
}
340 \item File/print/properties
\vspace{10pt
}
341 \item Change ``pages per sheet'' to
4\vspace{10pt
}