7 this extension amalgamates bitmanipulation primitives from many sources, including RISC-V bitmanip, Packed SIMD, AVX-512 and OpenPOWER VSX. Vectorisation and SIMD are removed: these are straight scalar (element) operations making them suitable for embedded applications.
8 Vectorisation Context is provided by [[openpower/sv]].
10 Scaoar variants of bitmanip oerations found in VSX are added so that VSX may be retired as "legacy" in the far future (10 to 20 years). Also, because VSX is hundreds of opcodes, requires 128 bit pathways, and is wholly unsuited to low power or embedded scenarios.
12 ternaryv is experimental and is the only operation that may be considered a "Packed SIMD". It is added as a variant of the already well-justified ternary operation (done in AVX512 as an immediate only) "because it looks fun". As it is based on the LUT4 concept it will allow accelerated emulation of FPGAs. Other vendors of ISAs are buying FPGA companies to achieve a similar objective.
14 general-purpose Galois Field operations are added so as to avoid huge opcode proliferation across many areas of Computer Science. however for convenience and also to avoid setup costs, some of the more common operations (clmul, crc32) are also added. The expectation is that these operations would all be covered by the same pipeline.
16 note that there are brownfield spaces below that could incorporate some of the set-before-first and other scalar operations listed in [[sv/vector_ops]],
17 the [[sv/av_opcodes]] and [[sv/cr_int_predication]]
21 minor opcode allocation
24 | ------ |--| --------- |
30 | 101 |0 | ternarycr |
36 | dest | src1 | subop | op |
37 | ---- | ---- | ----- | -------- |
38 | RT | RA | .. | bmatflip |
42 | dest | src1 | src2 | subop | op |
43 | ---- | ---- | ---- | ----- | -------- |
44 | RT | RA | RB | or | bmatflip |
45 | RT | RA | RB | xor | bmatflip |
46 | RT | RA | RB | bdep | dep/ext |
47 | RT | RA | RB | bext | dep/ext |
48 | RT | RA | RB | | grev |
49 | RT | RA | RB | | clmul* |
50 | RT | RA | RB | | gorc |
51 | RT | RA | RB | shuf | shuffle |
52 | RT | RA | RB | unshuf| shuffle |
53 | RT | RA | RB | width | xperm |
54 | RT | RA | RB | type | minmax |
65 | 0.5|6.10|11.15|16.20|21..25 | 26....30 |31| name |
66 | -- | -- | --- | --- | ----- | -------- |--| ------ |
67 | NN | RT | RA | RB | RC | mode 001 |Rc| ternary |
68 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc| ternaryi |
69 | NN | RS | RA | RB | RC | 00 011 |Rc| gfmul |
70 | NN | RS | RA | RB | RC | 01 011 |Rc| gfadd |
71 | NN | RT | RA | RB | deg | 10 011 |Rc| gfinv |
72 | NN | RS | RA | RB | deg | 11 011 |Rc| gfmuli |
73 | NN | RS | RA | RB | deg | 11 111 |Rc| gfaddi |
75 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31| name |
76 | -- | -- | --- | ----- | ---- | ----- |--| ------ |
77 | NN | RT | RA | imm | mask | 101 |1 | ternaryv |
79 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31| name |
80 | -- | -- | --- | --- |- |-----|----- | -----|--| -------|
81 | NN | BA | BB | BC |0 |imm | mask | 101 |0 | ternarycr |
85 | 0.5|6.10|11.15|16.20| 21.22 | 23 | 24....30 |31| name |
86 | -- | -- | --- | --- | ----- | -- | -------- |--| ---- |
87 | NN | RA | RB | | | 0 | 0000 110 |Rc| rsvd |
88 | NN | RA | RB | RC | itype | 1 | 0000 110 |Rc| xperm |
89 | NN | RA | RB | RC | itype | 0 | 0100 110 |Rc| minmax |
90 | NN | RA | RB | | | 1 | 0100 110 |Rc| rsvd |
91 | NN | RA | RB | sh | itype | SH | 1000 110 |Rc| bmopsi |
92 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
93 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
94 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
95 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
96 | NN | RA | RB | | | 0 | 0001 110 |Rc| rsvd |
97 | NN | RA | RB | | | 0 | 0101 110 |Rc| rsvd |
98 | NN | RA | RB | RC | 00 | 0 | 0010 110 |Rc| gorc |
99 | NN | RA | RB | sh | 00 | SH | 1010 110 |Rc| gorci |
100 | NN | RA | RB | RC | 00 | 0 | 0110 110 |Rc| gorcw |
101 | NN | RA | RB | sh | 00 | 0 | 1110 110 |Rc| gorcwi |
102 | NN | RA | RB | RC | 00 | 1 | 1110 110 |Rc| bmator |
103 | NN | RA | RB | RC | 01 | 0 | 0010 110 |Rc| grev |
104 | NN | RA | RB | RC | 01 | 1 | 0010 110 |Rc| clmul |
105 | NN | RA | RB | sh | 01 | SH | 1010 110 |Rc| grevi |
106 | NN | RA | RB | RC | 01 | 0 | 0110 110 |Rc| grevw |
107 | NN | RA | RB | sh | 01 | 0 | 1110 110 |Rc| grevwi |
108 | NN | RA | RB | RC | 01 | 1 | 1110 110 |Rc| bmatxor |
109 | NN | RA | RB | RC | 10 | 0 | 0010 110 |Rc| shfl |
110 | NN | RA | RB | sh | 10 | SH | 1010 110 |Rc| shfli |
111 | NN | RA | RB | RC | 10 | 0 | 0110 110 |Rc| shflw |
112 | NN | RA | RB | RC | 10 | 0 | 1110 110 |Rc| bdep |
113 | NN | RA | RB | RC | 10 | 1 | 1110 110 |Rc| bext |
114 | NN | RA | RB | RC | 11 | 0 | 1110 110 |Rc| clmulr |
115 | NN | RA | RB | RC | 11 | 1 | 1110 110 |Rc| clmulh |
116 | NN | RA | RB | | | | NN11 110 |Rc| rsvd |
118 # count leading/trailing zeros with mask
124 do i = 0 to 63 if((RB)i=1) then do
125 if((RS)i=1) then break end end count ← count + 1
129 # bit to byte permute
131 similar to matrix permute in RV bitmanip, which has XOR and OR variants
135 b = VSR[VRB+32].dword[i].byte[k].bit[j]
136 VSR[VRT+32].dword[i].byte[j].bit[k] = b
140 vpdepd VRT,VRA,VRB, identical to RV bitmamip bdep, found already in v3.1 p106
143 if VSR[VRB+32].dword[i].bit[63-m]=1 then do
144 result = VSR[VRA+32].dword[i].bit[63-k]
145 VSR[VRT+32].dword[i].bit[63-m] = result
151 uint_xlen_t bdep(uint_xlen_t RA, uint_xlen_t RB)
154 for (int i = 0, j = 0; i < XLEN; i++)
157 r |= uint_xlen_t(1) << i;
167 other way round: identical to RV bext, found in v3.1 p196
170 uint_xlen_t bext(uint_xlen_t RA, uint_xlen_t RB)
173 for (int i = 0, j = 0; i < XLEN; i++)
176 r |= uint_xlen_t(1) << j;
195 if((RB)63-i==1) then do
196 result63-ptr1 = (RS)63-i
204 signed and unsigned min/max for integer. this is sort-of partly synthesiseable in [[sv/svp64]] with pred-result as long as the dest reg is one of the sources, but not both signed and unsigned. when the dest is also one of the srces and the mv fails due to the CR bittest failing this will only overwrite the dest where the src is greater (or less).
206 signed/unsigned min/max gives more flexibility.
209 uint_xlen_t min(uint_xlen_t rs1, uint_xlen_t rs2)
210 { return (int_xlen_t)rs1 < (int_xlen_t)rs2 ? rs1 : rs2;
212 uint_xlen_t max(uint_xlen_t rs1, uint_xlen_t rs2)
213 { return (int_xlen_t)rs1 > (int_xlen_t)rs2 ? rs1 : rs2;
215 uint_xlen_t minu(uint_xlen_t rs1, uint_xlen_t rs2)
216 { return rs1 < rs2 ? rs1 : rs2;
218 uint_xlen_t maxu(uint_xlen_t rs1, uint_xlen_t rs2)
219 { return rs1 > rs2 ? rs1 : rs2;
226 Similar to FPGA LUTs: for every bit perform a lookup into a table using an 8bit immediate, or in another register
228 | 0.5|6.10|11.15|16.20| 21..25| 26..30 |31|
229 | -- | -- | --- | --- | ----- | -------- |--|
230 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc|
233 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
234 RT[i] = (imm & (1<<idx)) != 0
236 bits 21..22 may be used to specify a mode, such as treating the whole integer zero/nonzero and putting 1/0 in the result, rather than bitwise test.
238 a 4 operand variant which becomes more along the lines of an FPGA:
240 | 0.5|6.10|11.15|16.20|21.25| 26...30 |31|
241 | -- | -- | --- | --- | --- | -------- |--|
242 | NN | RT | RA | RB | RC | mode 001 |Rc|
245 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
246 RT[i] = (RC & (1<<idx)) != 0
248 mode (2 bit) may be used to do inversion of ordering, similar to carryless mul,
251 also, another possible variant involving swizzle and vec4:
253 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31|
254 | -- | -- | --- | ----- | ---- | ----- |--|
255 | NN | RT | RA | imm | mask | 101 |1 |
258 idx = RA.x[i] << 2 | RA.y[i] << 1 | RA.z[i]
259 res = (imm & (1<<idx)) != 0
261 if mask[j]: RT[i+j*8] = res
263 another mode selection would be CRs not Ints.
265 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31|
266 | -- | -- | --- | --- |- |-----|----- | -----|--|
267 | NN | BA | BB | BC |0 |imm | mask | 101 |0 |
270 if not mask[i] continue
271 idx = crregs[BA][i] << 2 |
274 crregs[BA][i] = (imm & (1<<idx)) != 0
278 based on RV bitmanip singlebit set, instruction format similar to shift
279 [[isa/fixedshift]]. bmext is actually covered already (shift-with-mask rldicl but only immediate version).
280 however bitmask-invert is not, and set/clr are not covered, although they can use the same Shift ALU.
282 bmext (RB) version is not the same as rldicl because bmext is a right shift by RC, where rldicl is a left rotate. for the immediate version this does not matter, so a bmexti is not required.
283 bmrev however there is no direct equivalent and consequently a bmrevi is required.
285 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
286 | -- | -- | --- | --- | --- | ------- |--| ----- |
287 | NN | RT | RA | RB | RC | mode 010 |Rc| bm* |
288 | NN | RT | RA | RB | RC | 0 1 111 |Rc| bmrev |
292 uint_xlen_t bmset(RA, RB, sh)
294 int shamt = RB & (XLEN - 1);
296 return RA | (mask << shamt);
299 uint_xlen_t bmclr(RA, RB, sh)
301 int shamt = RB & (XLEN - 1);
303 return RA & ~(mask << shamt);
306 uint_xlen_t bminv(RA, RB, sh)
308 int shamt = RB & (XLEN - 1);
310 return RA ^ (mask << shamt);
313 uint_xlen_t bmext(RA, RB, sh)
315 int shamt = RB & (XLEN - 1);
317 return mask & (RA >> shamt);
321 bitmask extract with reverse. can be done by bitinverting all of RA and getting bits of RA from the opposite end.
325 rev[0:msb] = ra[msb:0];
328 uint_xlen_t bmextrev(RA, RB, sh)
330 int shamt = (RB & (XLEN - 1));
331 shamt = (XLEN-1)-shamt; # shift other end
332 bra = bitreverse(RA) # swap LSB-MSB
334 return mask & (bra >> shamt);
338 | 0.5|6.10|11.15|16.20|21.26| 27..30 |31| name |
339 | -- | -- | --- | --- | --- | ------- |--| ------ |
340 | NN | RT | RA | RB | sh | 0 111 |Rc| bmrevi |
349 uint64_t grev64(uint64_t RA, uint64_t RB)
353 if (shamt & 1) x = ((x & 0x5555555555555555LL) << 1) |
354 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
355 if (shamt & 2) x = ((x & 0x3333333333333333LL) << 2) |
356 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
357 if (shamt & 4) x = ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
358 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
359 if (shamt & 8) x = ((x & 0x00FF00FF00FF00FFLL) << 8) |
360 ((x & 0xFF00FF00FF00FF00LL) >> 8);
361 if (shamt & 16) x = ((x & 0x0000FFFF0000FFFFLL) << 16) |
362 ((x & 0xFFFF0000FFFF0000LL) >> 16);
363 if (shamt & 32) x = ((x & 0x00000000FFFFFFFFLL) << 32) |
364 ((x & 0xFFFFFFFF00000000LL) >> 32);
370 # shuffle / unshuffle
375 uint32_t shfl32(uint32_t RA, uint32_t RB)
379 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
380 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
381 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
382 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
385 uint32_t unshfl32(uint32_t RA, uint32_t RB)
389 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
390 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
391 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
392 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
396 uint64_t shuffle64_stage(uint64_t src, uint64_t maskL, uint64_t maskR, int N)
398 uint64_t x = src & ~(maskL | maskR);
399 x |= ((src << N) & maskL) | ((src >> N) & maskR);
402 uint64_t shfl64(uint64_t RA, uint64_t RB)
406 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
407 0x00000000ffff0000LL, 16);
408 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
409 0x0000ff000000ff00LL, 8);
410 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
411 0x00f000f000f000f0LL, 4);
412 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
413 0x0c0c0c0c0c0c0c0cLL, 2);
414 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
415 0x2222222222222222LL, 1);
418 uint64_t unshfl64(uint64_t RA, uint64_t RB)
422 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
423 0x2222222222222222LL, 1);
424 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
425 0x0c0c0c0c0c0c0c0cLL, 2);
426 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
427 0x00f000f000f000f0LL, 4);
428 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
429 0x0000ff000000ff00LL, 8);
430 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
431 0x00000000ffff0000LL, 16);
441 uint_xlen_t xperm(uint_xlen_t RA, uint_xlen_t RB, int sz_log2)
444 uint_xlen_t sz = 1LL << sz_log2;
445 uint_xlen_t mask = (1LL << sz) - 1;
446 for (int i = 0; i < XLEN; i += sz) {
447 uint_xlen_t pos = ((RB >> i) & mask) << sz_log2;
449 r |= ((RA >> pos) & mask) << i;
453 uint_xlen_t xperm_n (uint_xlen_t RA, uint_xlen_t RB)
454 { return xperm(RA, RB, 2); }
455 uint_xlen_t xperm_b (uint_xlen_t RA, uint_xlen_t RB)
456 { return xperm(RA, RB, 3); }
457 uint_xlen_t xperm_h (uint_xlen_t RA, uint_xlen_t RB)
458 { return xperm(RA, RB, 4); }
459 uint_xlen_t xperm_w (uint_xlen_t RA, uint_xlen_t RB)
460 { return xperm(RA, RB, 5); }
468 uint32_t gorc32(uint32_t RA, uint32_t RB)
472 if (shamt & 1) x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
473 if (shamt & 2) x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
474 if (shamt & 4) x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
475 if (shamt & 8) x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
476 if (shamt & 16) x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
479 uint64_t gorc64(uint64_t RA, uint64_t RB)
483 if (shamt & 1) x |= ((x & 0x5555555555555555LL) << 1) |
484 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
485 if (shamt & 2) x |= ((x & 0x3333333333333333LL) << 2) |
486 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
487 if (shamt & 4) x |= ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
488 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
489 if (shamt & 8) x |= ((x & 0x00FF00FF00FF00FFLL) << 8) |
490 ((x & 0xFF00FF00FF00FF00LL) >> 8);
491 if (shamt & 16) x |= ((x & 0x0000FFFF0000FFFFLL) << 16) |
492 ((x & 0xFFFF0000FFFF0000LL) >> 16);
493 if (shamt & 32) x |= ((x & 0x00000000FFFFFFFFLL) << 32) |
494 ((x & 0xFFFFFFFF00000000LL) >> 32);
502 based on RV bitmanip, covered by ternary bitops
505 uint_xlen_t cmix(uint_xlen_t RA, uint_xlen_t RB, uint_xlen_t RC) {
506 return (RA & RB) | (RC & ~RB);
513 see https://en.wikipedia.org/wiki/CLMUL_instruction_set
516 uint_xlen_t clmul(uint_xlen_t RA, uint_xlen_t RB)
519 for (int i = 0; i < XLEN; i++)
524 uint_xlen_t clmulh(uint_xlen_t RA, uint_xlen_t RB)
527 for (int i = 1; i < XLEN; i++)
532 uint_xlen_t clmulr(uint_xlen_t RA, uint_xlen_t RB)
535 for (int i = 0; i < XLEN; i++)
537 x ^= RA >> (XLEN-i-1);
543 see <https://courses.csail.mit.edu/6.857/2016/files/ffield.py>
547 this requires 3 parameters and a "degree"
549 RT = GFMUL(RA, RB, gfdegree, modulo=RC)
551 realistically with the degree also needing to be an immediate it should be brought down to an overwrite version:
553 RS = GFMUL(RS, RA, gfdegree, modulo=RB)
554 RS = GFMUL(RS, RA, gfdegree=RC, modulo=RB)
556 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31|
557 | -- | -- | --- | --- | --- | ------- |--|
558 | NN | RS | RA | RB | deg | 00 011 |Rc|
559 | NN | RS | RA | RB | RC | 11 011 |Rc|
561 where the SimpleV variant may override RS-as-src differently from RS-as-dest
566 from functools import reduce
568 # constants used in the multGF2 function
569 mask1 = mask2 = polyred = None
571 def setGF2(degree, irPoly):
572 """Define parameters of binary finite field GF(2^m)/g(x)
573 - degree: extension degree of binary field
574 - irPoly: coefficients of irreducible polynomial g(x)
577 """Convert an integer into a polynomial"""
578 return [(sInt >> i) & 1
579 for i in reversed(range(sInt.bit_length()))]
581 global mask1, mask2, polyred
582 mask1 = mask2 = 1 << degree
584 polyred = reduce(lambda x, y: (x << 1) + y, i2P(irPoly)[1:])
587 """Multiply two polynomials in GF(2^m)/g(x)"""
598 if __name__ == "__main__":
600 # Define binary field GF(2^3)/x^3 + x + 1
603 # Evaluate the product (x^2 + x + 1)(x^2 + 1)
604 print("{:02x}".format(multGF2(0b111, 0b101)))
606 # Define binary field GF(2^8)/x^8 + x^4 + x^3 + x + 1
607 # (used in the Advanced Encryption Standard-AES)
608 setGF2(8, 0b100011011)
610 # Evaluate the product (x^7)(x^7 + x + 1)
611 print("{:02x}".format(multGF2(0b10000000, 0b10000011)))
615 RS = GFADDI(RS, RA|0, gfdegree, modulo=RB)
616 RS = GFADD(RS, RA|0, gfdegree=RC, modulo=RB)
618 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
619 | -- | -- | --- | --- | --- | ------- |--| ----- |
620 | NN | RS | RA | RB | deg | 0 1 011 |Rc| gfaddi |
621 | NN | RS | RA | RB | RC | 1 1 111 |Rc| gfadd |
623 GFMOD is a pseudo-op where RA=0
636 def gf_invert(a, mod=0x1B) :
651 a %= 256 # Emulating 8-bit overflow
652 g1 %= 256 # Emulating 8-bit overflow
654 j = gf_degree(a) - gf_degree(v)
662 uint64_t bmatflip(uint64_t RA)
670 uint64_t bmatxor(uint64_t RA, uint64_t RB)
673 uint64_t RBt = bmatflip(RB);
674 uint8_t u[8]; // rows of RA
675 uint8_t v[8]; // cols of RB
676 for (int i = 0; i < 8; i++) {
681 for (int i = 0; i < 64; i++) {
682 if (pcnt(u[i / 8] & v[i % 8]) & 1)
687 uint64_t bmator(uint64_t RA, uint64_t RB)
690 uint64_t RBt = bmatflip(RB);
691 uint8_t u[8]; // rows of RA
692 uint8_t v[8]; // cols of RB
693 for (int i = 0; i < 8; i++) {
698 for (int i = 0; i < 64; i++) {
699 if ((u[i / 8] & v[i % 8]) != 0)