7 this extension amalgamates bitnanipulation 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. Vectorisation Context is provided by [[openpower/sv]].
9 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.
11 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.
15 minor opcode allocation
18 | ------ |--| --------- |
24 | 101 |0 | ternarycr |
30 | dest | src1 | subop | op |
31 | ---- | ---- | ----- | -------- |
32 | RT | RA | .. | bmatflip |
36 | dest | src1 | src2 | subop | op |
37 | ---- | ---- | ---- | ----- | -------- |
38 | RT | RA | RB | or | bmatflip |
39 | RT | RA | RB | xor | bmatflip |
40 | RT | RA | RB | bdep | dep/ext |
41 | RT | RA | RB | bext | dep/ext |
42 | RT | RA | RB | | grev |
43 | RT | RA | RB | | clmul* |
44 | RT | RA | RB | | gorc |
45 | RT | RA | RB | shuf | shuffle |
46 | RT | RA | RB | unshuf| shuffle |
47 | RT | RA | RB | width | xperm |
48 | RT | RA | RB | type | minmax |
59 | 0.5|6.10|11.15|16.20|21..25 | 26....30 |31| name |
60 | -- | -- | --- | --- | ----- | -------- |--| ------ |
61 | NN | RT | RA | RB | RC | mode 001 |Rc| ternary |
62 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc| ternaryi |
63 | NN | RS | RA | RB | RC | 00 011 |Rc| gfmul |
64 | NN | RS | RA | RB | RC | 01 011 |Rc| gfadd |
65 | NN | RT | RA | RB | deg | 10 011 |Rc| gfinv |
66 | NN | RS | RA | RB | deg | 11 011 |Rc| gfmuli |
67 | NN | RS | RA | RB | deg | 11 111 |Rc| gfaddi |
69 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31| name |
70 | -- | -- | --- | ----- | ---- | ----- |--| ------ |
71 | NN | RT | RA | imm | mask | 101 |1 | ternaryv |
73 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31| name |
74 | -- | -- | --- | --- |- |-----|----- | -----|--| -------|
75 | NN | BA | BB | BC |0 |imm | mask | 101 |0 | ternarycr |
79 | 0.5|6.10|11.15|16.20| 21.22 | 23 | 24....30 |31| name |
80 | -- | -- | --- | --- | ----- | -- | -------- |--| ---- |
81 | NN | RA | RB | | | 0 | 0000 110 |Rc| rsvd |
82 | NN | RA | RB | RC | itype | 1 | 0000 110 |Rc| xperm |
83 | NN | RA | RB | RC | itype | 0 | 0100 110 |Rc| minmax |
84 | NN | RA | RB | | | 1 | 0100 110 |Rc| rsvd |
85 | NN | RA | RB | sh | itype | SH | 1000 110 |Rc| bmopsi |
86 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
87 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
88 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
89 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
90 | NN | RA | RB | | | 0 | 0001 110 |Rc| rsvd |
91 | NN | RA | RB | | | 0 | 0101 110 |Rc| rsvd |
92 | NN | RA | RB | RC | 00 | 0 | 0010 110 |Rc| gorc |
93 | NN | RA | RB | sh | 00 | SH | 1010 110 |Rc| gorci |
94 | NN | RA | RB | RC | 00 | 0 | 0110 110 |Rc| gorcw |
95 | NN | RA | RB | sh | 00 | 0 | 1110 110 |Rc| gorcwi |
96 | NN | RA | RB | RC | 00 | 1 | 1110 110 |Rc| bmator |
97 | NN | RA | RB | RC | 01 | 0 | 0010 110 |Rc| grev |
98 | NN | RA | RB | RC | 01 | 1 | 0010 110 |Rc| clmul |
99 | NN | RA | RB | sh | 01 | SH | 1010 110 |Rc| grevi |
100 | NN | RA | RB | RC | 01 | 0 | 0110 110 |Rc| grevw |
101 | NN | RA | RB | sh | 01 | 0 | 1110 110 |Rc| grevwi |
102 | NN | RA | RB | RC | 01 | 1 | 1110 110 |Rc| bmatxor |
103 | NN | RA | RB | RC | 10 | 0 | 0010 110 |Rc| shfl |
104 | NN | RA | RB | sh | 10 | SH | 1010 110 |Rc| shfli |
105 | NN | RA | RB | RC | 10 | 0 | 0110 110 |Rc| shflw |
106 | NN | RA | RB | RC | 10 | 0 | 1110 110 |Rc| bdep |
107 | NN | RA | RB | RC | 10 | 1 | 1110 110 |Rc| bext |
108 | NN | RA | RB | RC | 11 | 0 | 1110 110 |Rc| clmulr |
109 | NN | RA | RB | RC | 11 | 1 | 1110 110 |Rc| clmulh |
110 | NN | RA | RB | | | | NN11 110 |Rc| rsvd |
112 # count leading/trailing zeros with mask
118 do i = 0 to 63 if((RB)i=1) then do
119 if((RS)i=1) then break end end count ← count + 1
123 # bit to byte permute
125 similar to matrix permute in RV bitmanip, which has XOR and OR variants
129 b = VSR[VRB+32].dword[i].byte[k].bit[j]
130 VSR[VRT+32].dword[i].byte[j].bit[k] = b
134 vpdepd VRT,VRA,VRB, identical to RV bitmamip bdep, found already in v3.1 p106
137 if VSR[VRB+32].dword[i].bit[63-m]=1 then do
138 result = VSR[VRA+32].dword[i].bit[63-k]
139 VSR[VRT+32].dword[i].bit[63-m] = result
145 uint_xlen_t bdep(uint_xlen_t RA, uint_xlen_t RB)
148 for (int i = 0, j = 0; i < XLEN; i++)
151 r |= uint_xlen_t(1) << i;
161 other way round: identical to RV bext, found in v3.1 p196
164 uint_xlen_t bext(uint_xlen_t RA, uint_xlen_t RB)
167 for (int i = 0, j = 0; i < XLEN; i++)
170 r |= uint_xlen_t(1) << j;
182 ptr0 ← 0 ptr1 ← 0 do i = 0 to 63 if((RB)i=0) then do
183 resultptr0 ← (RS)i end ptr0 ← ptr0 + 1
184 if((RB)63-i==1) then do
185 result63-ptr1 ← (RS)63-i end end ptr1 ← ptr1 + 1
191 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).
193 signed/unsigned min/max gives more flexibility.
196 uint_xlen_t min(uint_xlen_t rs1, uint_xlen_t rs2)
197 { return (int_xlen_t)rs1 < (int_xlen_t)rs2 ? rs1 : rs2;
199 uint_xlen_t max(uint_xlen_t rs1, uint_xlen_t rs2)
200 { return (int_xlen_t)rs1 > (int_xlen_t)rs2 ? rs1 : rs2;
202 uint_xlen_t minu(uint_xlen_t rs1, uint_xlen_t rs2)
203 { return rs1 < rs2 ? rs1 : rs2;
205 uint_xlen_t maxu(uint_xlen_t rs1, uint_xlen_t rs2)
206 { return rs1 > rs2 ? rs1 : rs2;
213 Similar to FPGA LUTs: for every bit perform a lookup into a table using an 8bit immediate, or in another register
215 | 0.5|6.10|11.15|16.20| 21..25| 26..30 |31|
216 | -- | -- | --- | --- | ----- | -------- |--|
217 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc|
220 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
221 RT[i] = (imm & (1<<idx)) != 0
223 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.
225 a 4 operand variant which becomes more along the lines of an FPGA:
227 | 0.5|6.10|11.15|16.20|21.25| 26...30 |31|
228 | -- | -- | --- | --- | --- | -------- |--|
229 | NN | RT | RA | RB | RC | mode 001 |Rc|
232 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
233 RT[i] = (RC & (1<<idx)) != 0
235 mode (2 bit) may be used to do inversion of ordering, similar to carryless mul,
238 also, another possible variant involving swizzle and vec4:
240 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31|
241 | -- | -- | --- | ----- | ---- | ----- |--|
242 | NN | RT | RA | imm | mask | 101 |1 |
245 idx = RA.x[i] << 2 | RA.y[i] << 1 | RA.z[i]
246 res = (imm & (1<<idx)) != 0
248 if mask[j]: RT[i+j*8] = res
250 another mode selection would be CRs not Ints.
252 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31|
253 | -- | -- | --- | --- |- |-----|----- | -----|--|
254 | NN | BA | BB | BC |0 |imm | mask | 101 |0 |
257 if not mask[i] continue
258 idx = crregs[BA][i] << 2 |
261 crregs[BA][i] = (imm & (1<<idx)) != 0
265 based on RV bitmanip singlebit set, instruction format similar to shift
266 [[isa/fixedshift]]. bmext is actually covered already (shift-with-mask rldicl but only immediate version).
267 however bitmask-invert is not, and set/clr are not covered, although they can use the same Shift ALU.
269 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.
270 bmrev however there is no direct equivalent and consequently a bmrevi is required.
272 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
273 | -- | -- | --- | --- | --- | ------- |--| ----- |
274 | NN | RT | RA | RB | RC | mode 010 |Rc| bm* |
275 | NN | RT | RA | RB | RC | 0 1 111 |Rc| bmrev |
279 uint_xlen_t bmset(RA, RB, sh)
281 int shamt = RB & (XLEN - 1);
283 return RA | (mask << shamt);
286 uint_xlen_t bmclr(RA, RB, sh)
288 int shamt = RB & (XLEN - 1);
290 return RA & ~(mask << shamt);
293 uint_xlen_t bminv(RA, RB, sh)
295 int shamt = RB & (XLEN - 1);
297 return RA ^ (mask << shamt);
300 uint_xlen_t bmext(RA, RB, sh)
302 int shamt = RB & (XLEN - 1);
304 return mask & (RA >> shamt);
308 bitmask extract with reverse. can be done by bitinverting all of RA and getting bits of RA from the opposite end.
312 rev[0:msb] = ra[msb:0];
315 uint_xlen_t bmextrev(RA, RB, sh)
317 int shamt = (RB & (XLEN - 1));
318 shamt = (XLEN-1)-shamt; # shift other end
319 bra = bitreverse(RA) # swap LSB-MSB
321 return mask & (bra >> shamt);
325 | 0.5|6.10|11.15|16.20|21.26| 27..30 |31| name |
326 | -- | -- | --- | --- | --- | ------- |--| ------ |
327 | NN | RT | RA | RB | sh | 0 111 |Rc| bmrevi |
336 uint64_t grev64(uint64_t RA, uint64_t RB)
340 if (shamt & 1) x = ((x & 0x5555555555555555LL) << 1) |
341 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
342 if (shamt & 2) x = ((x & 0x3333333333333333LL) << 2) |
343 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
344 if (shamt & 4) x = ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
345 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
346 if (shamt & 8) x = ((x & 0x00FF00FF00FF00FFLL) << 8) |
347 ((x & 0xFF00FF00FF00FF00LL) >> 8);
348 if (shamt & 16) x = ((x & 0x0000FFFF0000FFFFLL) << 16) |
349 ((x & 0xFFFF0000FFFF0000LL) >> 16);
350 if (shamt & 32) x = ((x & 0x00000000FFFFFFFFLL) << 32) |
351 ((x & 0xFFFFFFFF00000000LL) >> 32);
357 # shuffle / unshuffle
362 uint32_t shfl32(uint32_t RA, uint32_t RB)
366 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
367 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
368 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
369 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
372 uint32_t unshfl32(uint32_t RA, uint32_t RB)
376 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
377 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
378 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
379 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
383 uint64_t shuffle64_stage(uint64_t src, uint64_t maskL, uint64_t maskR, int N)
385 uint64_t x = src & ~(maskL | maskR);
386 x |= ((src << N) & maskL) | ((src >> N) & maskR);
389 uint64_t shfl64(uint64_t RA, uint64_t RB)
393 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
394 0x00000000ffff0000LL, 16);
395 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
396 0x0000ff000000ff00LL, 8);
397 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
398 0x00f000f000f000f0LL, 4);
399 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
400 0x0c0c0c0c0c0c0c0cLL, 2);
401 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
402 0x2222222222222222LL, 1);
405 uint64_t unshfl64(uint64_t RA, uint64_t RB)
409 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
410 0x2222222222222222LL, 1);
411 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
412 0x0c0c0c0c0c0c0c0cLL, 2);
413 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
414 0x00f000f000f000f0LL, 4);
415 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
416 0x0000ff000000ff00LL, 8);
417 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
418 0x00000000ffff0000LL, 16);
428 uint_xlen_t xperm(uint_xlen_t RA, uint_xlen_t RB, int sz_log2)
431 uint_xlen_t sz = 1LL << sz_log2;
432 uint_xlen_t mask = (1LL << sz) - 1;
433 for (int i = 0; i < XLEN; i += sz) {
434 uint_xlen_t pos = ((RB >> i) & mask) << sz_log2;
436 r |= ((RA >> pos) & mask) << i;
440 uint_xlen_t xperm_n (uint_xlen_t RA, uint_xlen_t RB)
441 { return xperm(RA, RB, 2); }
442 uint_xlen_t xperm_b (uint_xlen_t RA, uint_xlen_t RB)
443 { return xperm(RA, RB, 3); }
444 uint_xlen_t xperm_h (uint_xlen_t RA, uint_xlen_t RB)
445 { return xperm(RA, RB, 4); }
446 uint_xlen_t xperm_w (uint_xlen_t RA, uint_xlen_t RB)
447 { return xperm(RA, RB, 5); }
455 uint32_t gorc32(uint32_t RA, uint32_t RB)
459 if (shamt & 1) x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
460 if (shamt & 2) x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
461 if (shamt & 4) x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
462 if (shamt & 8) x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
463 if (shamt & 16) x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
466 uint64_t gorc64(uint64_t RA, uint64_t RB)
470 if (shamt & 1) x |= ((x & 0x5555555555555555LL) << 1) |
471 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
472 if (shamt & 2) x |= ((x & 0x3333333333333333LL) << 2) |
473 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
474 if (shamt & 4) x |= ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
475 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
476 if (shamt & 8) x |= ((x & 0x00FF00FF00FF00FFLL) << 8) |
477 ((x & 0xFF00FF00FF00FF00LL) >> 8);
478 if (shamt & 16) x |= ((x & 0x0000FFFF0000FFFFLL) << 16) |
479 ((x & 0xFFFF0000FFFF0000LL) >> 16);
480 if (shamt & 32) x |= ((x & 0x00000000FFFFFFFFLL) << 32) |
481 ((x & 0xFFFFFFFF00000000LL) >> 32);
489 based on RV bitmanip, covered by ternary bitops
492 uint_xlen_t cmix(uint_xlen_t RA, uint_xlen_t RB, uint_xlen_t RC) {
493 return (RA & RB) | (RC & ~RB);
500 see https://en.wikipedia.org/wiki/CLMUL_instruction_set
503 uint_xlen_t clmul(uint_xlen_t RA, uint_xlen_t RB)
506 for (int i = 0; i < XLEN; i++)
511 uint_xlen_t clmulh(uint_xlen_t RA, uint_xlen_t RB)
514 for (int i = 1; i < XLEN; i++)
519 uint_xlen_t clmulr(uint_xlen_t RA, uint_xlen_t RB)
522 for (int i = 0; i < XLEN; i++)
524 x ^= RA >> (XLEN-i-1);
530 see <https://courses.csail.mit.edu/6.857/2016/files/ffield.py>
534 this requires 3 parameters and a "degree"
536 RT = GFMUL(RA, RB, gfdegree, modulo=RC)
538 realistically with the degree also needing to be an immediate it should be brought down to an overwrite version:
540 RS = GFMUL(RS, RA, gfdegree, modulo=RB)
541 RS = GFMUL(RS, RA, gfdegree=RC, modulo=RB)
543 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31|
544 | -- | -- | --- | --- | --- | ------- |--|
545 | NN | RS | RA | RB | deg | 00 011 |Rc|
546 | NN | RS | RA | RB | RC | 11 011 |Rc|
548 where the SimpleV variant may override RS-as-src differently from RS-as-dest
553 from functools import reduce
555 # constants used in the multGF2 function
556 mask1 = mask2 = polyred = None
558 def setGF2(degree, irPoly):
559 """Define parameters of binary finite field GF(2^m)/g(x)
560 - degree: extension degree of binary field
561 - irPoly: coefficients of irreducible polynomial g(x)
564 """Convert an integer into a polynomial"""
565 return [(sInt >> i) & 1
566 for i in reversed(range(sInt.bit_length()))]
568 global mask1, mask2, polyred
569 mask1 = mask2 = 1 << degree
571 polyred = reduce(lambda x, y: (x << 1) + y, i2P(irPoly)[1:])
574 """Multiply two polynomials in GF(2^m)/g(x)"""
585 if __name__ == "__main__":
587 # Define binary field GF(2^3)/x^3 + x + 1
590 # Evaluate the product (x^2 + x + 1)(x^2 + 1)
591 print("{:02x}".format(multGF2(0b111, 0b101)))
593 # Define binary field GF(2^8)/x^8 + x^4 + x^3 + x + 1
594 # (used in the Advanced Encryption Standard-AES)
595 setGF2(8, 0b100011011)
597 # Evaluate the product (x^7)(x^7 + x + 1)
598 print("{:02x}".format(multGF2(0b10000000, 0b10000011)))
602 RS = GFADDI(RS, RA|0, gfdegree, modulo=RB)
603 RS = GFADD(RS, RA|0, gfdegree=RC, modulo=RB)
605 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
606 | -- | -- | --- | --- | --- | ------- |--| ----- |
607 | NN | RS | RA | RB | deg | 0 1 011 |Rc| gfaddi |
608 | NN | RS | RA | RB | RC | 1 1 111 |Rc| gfadd |
610 GFMOD is a pseudo-op where RA=0
623 def gf_invert(a, mod=0x1B) :
638 a %= 256 # Emulating 8-bit overflow
639 g1 %= 256 # Emulating 8-bit overflow
641 j = gf_degree(a) - gf_degree(v)
649 uint64_t bmatflip(uint64_t RA)
657 uint64_t bmatxor(uint64_t RA, uint64_t RB)
660 uint64_t RBt = bmatflip(RB);
661 uint8_t u[8]; // rows of RA
662 uint8_t v[8]; // cols of RB
663 for (int i = 0; i < 8; i++) {
668 for (int i = 0; i < 64; i++) {
669 if (pcnt(u[i / 8] & v[i % 8]) & 1)
674 uint64_t bmator(uint64_t RA, uint64_t RB)
677 uint64_t RBt = bmatflip(RB);
678 uint8_t u[8]; // rows of RA
679 uint8_t v[8]; // cols of RB
680 for (int i = 0; i < 8; i++) {
685 for (int i = 0; i < 64; i++) {
686 if ((u[i / 8] & v[i % 8]) != 0)