1 # RFC ls013 Min/Max GPR/FPR
3 * Funded by NLnet under the Privacy and Enhanced Trust Programme, EU
4 Horizon2020 Grant 825310, and NGI0 Entrust No 101069594
5 * <https://libre-soc.org/openpower/sv/rfc/ls013/>
6 * <https://git.openpower.foundation/isa/PowerISA/issues/TODO>
7 * <https://bugs.libre-soc.org/show_bug.cgi?id=1057>
19 **Books and Section affected**:
22 Book I Fixed-Point and Floating-Point Instructions
23 Appendix E Power ISA sorted by opcode
24 Appendix F Power ISA sorted by version
25 Appendix G Power ISA sorted by Compliancy Subset
26 Appendix H Power ISA sorted by mnemonic
35 **Submitter**: Luke Leighton (Libre-SOC)
37 **Requester**: Libre-SOC
39 **Impact on processor**:
42 Addition of new GPR-based and FPR-based instructions
45 **Impact on software**:
48 Requires support for new instructions in assembler, debuggers,
55 GPR, FPR, min, max, fmin, fmax
60 Minimum/Maximum are common operations that can take an astounding number of
61 operations to implement in software. Additionally, Vector Reduce-Min/Max are
62 common vector operations, and SVP64 Parallel Reduction needs a single Scalar
63 instruction in order to effectively implement Reduce-Min/Max.
65 **Notes and Observations**:
67 1. SVP64 REMAP Parallel Reduction needs a single Scalar instruction to
68 work with, for best effectiveness. With no SFFS minimum/maximum
69 instructions Simple-V min/max Parallel Reduction is severely compromised.
70 2. Once one FP min/max mode is implemented the rest are not much more hardware.
71 3. There exists similar instructions in VSX (not IEEE754-2019 though).
72 This is frequently used to justify not adding them. However SVP64/VSX may
73 have different meaning from SVP64/SFFS, so it is *really* crucial to have
74 SFFS ops even if "equivalent" to VSX in order for SVP64 to not be
75 compromised (non-orthogonal).
76 4. FP min/max are rather complex to implement in software, the most commonly
77 used FP max function `fmax` from glibc compiled for SFFS is an astounding
82 Add the following entries to:
84 * the Appendices of Book I
85 * Book I 3.3.9 Fixed-Point Arithmetic Instructions
86 * Book I 4.6.6.1 Floating-Point Elementary Arithmetic Instructions
87 * Book I 1.6.1 and 1.6.2
93 # Floating-Point Instructions
95 This group is to provide Floating-Point min/max, however with the 2019 version
96 of IEEE 754 there are now subtle differences. These are selectable with a
99 ## `FMM` -- Floating Min/Max Mode
101 <a id="fmm-floating-min-max-mode"></a>
103 <!-- hyphens in table determine width of columns for pandoc -- -->
104 | `FMM`| Extended Mnemonic | Origin | Semantics |
105 |------|-------------------------------|--------------------|--------------------------------------------|
106 | 0000 | fminnum08[s] FRT,FRA,FRB | IEEE 754-2008 | minNum(FRA,FRB) (1) |
107 | 0001 | fmin19[s] FRT,FRA,FRB | IEEE 754-2019 | minimum(FRA,FRB) |
108 | 0010 | fminnum19[s] FRT,FRA,FRB | IEEE 754-2019 | minimumNumber(FRA,FRB) |
109 | 0011 | fminc[s] FRT,FRA,FRB | x86 minss (4) | FRA\<FRB ? FRA:FRB |
110 | 0100 | fminmagnum08[s] FRT,FRA,FRB | IEEE 754-2008 (3) | mmmag(FRA,FRB,False,fminnum08) (2) |
111 | 0101 | fminmag19[s] FRT,FRA,FRB | IEEE 754-2019 | mmmag(FRA,FRB,False,fmin19) (2) |
112 | 0110 | fminmagnum19[s] FRT,FRA,FRB | IEEE 754-2019 | mmmag(FRA,FRB,False,fminnum19) (2) |
113 | 0111 | fminmagc[s] FRT,FRA,FRB | - | mmmag(FRA,FRB,False,fminc) (2) |
114 | 1000 | fmaxnum08[s] FRT,FRA,FRB | IEEE 754-2008 | maxNum(FRA,FRB) (1) |
115 | 1001 | fmax19[s] FRT,FRA,FRB | IEEE 754-2019 | maximum(FRA,FRB) |
116 | 1010 | fmaxnum19[s] FRT,FRA,FRB | IEEE 754-2019 | maximumNumber(FRA,FRB) |
117 | 1011 | fmaxc[s] FRT,FRA,FRB | x86 maxss (4) | FRA\>FRB ? FRA:FRB |
118 | 1100 | fmaxmagnum08[s] FRT,FRA,FRB | IEEE 754-2008 (3) | mmmag(FRA,FRB,True,fmaxnum08) (2) |
119 | 1101 | fmaxmag19[s] FRT,FRA,FRB | IEEE 754-2019 | mmmag(FRA,FRB,True,fmax19) (2) |
120 | 1110 | fmaxmagnum19[s] FRT,FRA,FRB | IEEE 754-2019 | mmmag(FRA,FRB,True,fmaxnum19) (2) |
121 | 1111 | fmaxmagc[s] FRT,FRA,FRB | - | mmmag(FRA,FRB,True,fmaxc) (2) |
123 Note (1): for the purposes of minNum/maxNum, -0.0 is defined to be less than
124 +0.0. This is left unspecified in IEEE 754-2008.
126 Note (2): mmmag(x, y, cmp, fallback) is defined as:
129 def mmmag(x, y, is_max, fallback):
138 # equal magnitudes, or NaN input(s)
139 return fallback(x, y)
142 Note (3): TODO: icr if IEEE 754-2008 has min/maxMagNum like IEEE 754-2019's
143 minimum/maximumMagnitudeNumber
145 Note (4) or Win32's min macro
151 ## Floating Minimum/Maximum MM-form
153 * fminmax FRT, FRA, FRB, FMM
154 * fminmax. FRT, FRA, FRB, FMM
157 |0 |6 |11 |16 |21 |25 |31 |
158 | PO | FRT | FRA | FRB | FMM | XO | Rc |
165 abs_a <- 0b0 || a[1:63]
166 abs_b <- 0b0 || b[1:63]
167 a_is_nan <- abs_a >u 0x7FF0_0000_0000_0000
168 a_is_snan <- a_is_nan & (a[12] = 0)
169 b_is_nan <- abs_b >u 0x7FF0_0000_0000_0000
170 b_is_snan <- b_is_nan & (b[12] = 0)
171 any_snan <- a_is_snan | b_is_snan
176 if a_is_nan | b_is_nan then
177 if FMM[2:3] = 0b00 then # min/maxnum08
178 if a_is_snan then result <- a_quieted
179 else if b_is_snan then result <- b_quieted
180 else if a_is_nan & b_is_nan then result <- a_quieted
181 else if a_is_nan then result <- b
183 if FMM[2:3] = 0b01 then # min/max19
184 if a_is_nan then result <- a_quieted
185 else result <- b_quieted
186 if FMM[2:3] = 0b10 then # min/maxnum19
187 if a_is_nan & b_is_nan then result <- a_quieted
188 else if a_is_nan then result <- b
190 if FMM[2:3] = 0b11 then # min/maxc
195 if FMM[1] then # min/maxmag
196 if abs_a != abs_b then
199 if FMM[2:3] = 0b11 then # min/maxc
200 if abs_a = 0 then cmp_l[0:63] <- 0
201 if abs_b = 0 then cmp_r[0:63] <- 0
203 # swap cmp_* so comparison goes the other way
204 cmp_l, cmp_r <- cmp_r, cmp_l
206 if cmp_r[0] = 0 then result <- a
207 else if cmp_l >u cmp_r then
208 # IEEE 754 is sign-magnitude,
209 # so bigger magnitude negative is smaller
212 else if cmp_r[0] = 1 then result <- b
213 else if cmp_l <u cmp_r then result <- a
215 if any_snan then SetFX(FPSCR.VXSNAN)
216 if (FPSCR.VE = 0) | ¬any_snan then (FRT) <- result
219 Compute the minimum/maximum of FRA and FRB, according to FMM, and store the
222 Special Registers altered:
231 see [`FMM` -- Floating Min/Max Mode](#fmm-floating-min-max-mode)
235 # Fixed-Point Instructions
237 These are signed and unsigned, min or max. SVP64 Prefixing defines Saturation
238 semantics therefore Saturated variants of these instructions need not be proposed.
240 ## `MMM` -- Integer Min/Max Mode
242 <a id="mmm-integer-min-max-mode"></a>
244 * bit 0: set if word variant else dword
245 * bit 1: set if signed else unsigned
246 * bit 2: set if max else min
248 | `MMM` | Extended Mnemonic | Semantics |
249 |-------|-------------------|----------------------------------------------|
250 | 000 | `minu RT,RA,RB` | `(uint64_t)RA < (uint64_t)RB ? RA : RB` |
251 | 001 | `maxu RT,RA,RB` | `(uint64_t)RA > (uint64_t)RB ? RA : RB` |
252 | 010 | `mins RT,RA,RB` | ` (int64_t)RA < (int64_t)RB ? RA : RB` |
253 | 011 | `maxs RT,RA,RB` | ` (int64_t)RA > (int64_t)RB ? RA : RB` |
254 | 100 | `minuw RT,RA,RB` | `(uint32_t)RA < (uint32_t)RB ? RA : RB` |
255 | 101 | `maxuw RT,RA,RB` | `(uint32_t)RA > (uint32_t)RB ? RA : RB` |
256 | 110 | `minsw RT,RA,RB` | ` (int32_t)RA < (int32_t)RB ? RA : RB` |
257 | 111 | `maxsw RT,RA,RB` | ` (int32_t)RA > (int32_t)RB ? RA : RB` |
259 ## Minimum/Maximum MM-Form
261 * minmax RT, RA, RB, MMM
262 * minmax. RT, RA, RB, MMM
265 |0 |6 |11 |16 |21 |24 |25 |31 |
266 | PO | RT | RA | RB | MMM | / | XO | Rc |
272 if MMM[0] then # word mode
273 # shift left by XLEN/2 to make the dword comparison
274 # do word comparison of the original inputs
275 a <- a[XLEN/2:XLEN-1] || [0] * XLEN/2
276 b <- b[XLEN/2:XLEN-1] || [0] * XLEN/2
277 if MMM[1] then # signed mode
278 # invert sign bits to make the unsigned comparison
279 # do signed comparison of the original inputs
282 # if Rc = 1 then store the result of comparing a and b to CR0
285 CR0 <- 0b100 || XER.SO
287 CR0 <- 0b001 || XER.SO
289 CR0 <- 0b010 || XER.SO
290 if MMM[2] then # max mode
291 # swap a and b to make the less than comparison do
292 # greater than comparison of the original inputs
296 # store the entire selected source (even in word mode)
297 # if Rc = 1 then store the result of comparing a and b to CR0
298 if a <u b then RT <- (RA|0)
302 Compute the integer minimum/maximum according to `MMM` of `(RA|0)` and `(RB)`
303 and store the result in `RT`.
305 Special Registers altered:
313 see [`MMM` -- Integer Min/Max Mode](#mmm-integer-min-max-mode)
319 # Instruction Formats
321 Add the following entries to Book I 1.6.1 Word Instruction Formats:
326 |0 |6 |11 |16 |21 |24 |25 |31 |
327 | PO | FRT | FRA | FRB | FMM | XO | Rc |
328 | PO | RT | RA | RB | MMM | / | XO | Rc |
331 Add the following new fields to Book I 1.6.2 Word Instruction Fields:
335 Field used to specify minimum/maximum mode for fminmax.
340 Field used to specify minimum/maximum mode for integer minmax.
345 Add `MM` to the `Formats:` list for all of `FRT`, `FRA`, `FRB`, `XO (25:30)`,
346 `Rc`, `RT`, `RA` and `RB`.
354 Appendix E Power ISA sorted by opcode
355 Appendix F Power ISA sorted by version
356 Appendix G Power ISA sorted by Compliancy Subset
357 Appendix H Power ISA sorted by mnemonic
359 | Form | Book | Page | Version | Mnemonic | Description |
360 |------|------|------|---------|----------|-------------|
361 | MM | I | # | 3.2B | fminmax | Floating Minimum/Maximum |
362 | MM | I | # | 3.2B | minmax | Minimum/Maximum |
364 ## fmax instruction count
366 32 instructions are required in SFFS to emulate fmax.
372 inline uint64_t asuint64(double f) {
380 inline int issignaling(double v) {
381 // copied from glibc:
382 // https://github.com/bminor/glibc/blob/e2756903/sysdeps/ieee754/dbl-64/math_config.h#L101
383 uint64_t ix = asuint64(v);
384 return 2 * (ix ^ 0x0008000000000000) > 2 * 0x7ff8000000000000ULL;
387 double fmax(double x, double y) {
388 // copied from glibc:
389 // https://github.com/bminor/glibc/blob/e2756903/math/s_fmax_template.c
390 if(__builtin_isgreaterequal(x, y))
392 else if(__builtin_isless(x, y))
394 else if(issignaling(x) || issignaling(y))
397 return __builtin_isnan(y) ? x : y;
404 fmax(double, double):
442 .byte 0,9,0,0,0,0,0,0