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 # bit to byte permute
114 similar to matrix permute in RV bitmanip, which has XOR and OR variants
118 b = VSR[VRB+32].dword[i].byte[k].bit[j]
119 VSR[VRT+32].dword[i].byte[j].bit[k] = b
123 vpdepd VRT,VRA,VRB, identical to RV bitmamip bdep, found already in v3.1 p106
126 if VSR[VRB+32].dword[i].bit[63-m]=1 then do
127 result = VSR[VRA+32].dword[i].bit[63-k]
128 VSR[VRT+32].dword[i].bit[63-m] = result
134 uint_xlen_t bdep(uint_xlen_t RA, uint_xlen_t RB)
137 for (int i = 0, j = 0; i < XLEN; i++)
140 r |= uint_xlen_t(1) << i;
150 other way round: identical to RV bext, found in v3.1 p196
153 uint_xlen_t bext(uint_xlen_t RA, uint_xlen_t RB)
156 for (int i = 0, j = 0; i < XLEN; i++)
159 r |= uint_xlen_t(1) << j;
171 ptr0 ← 0 ptr1 ← 0 do i = 0 to 63 if((RB)i=0) then do
172 resultptr0 ← (RS)i end ptr0 ← ptr0 + 1
173 if((RB)63-i==1) then do
174 result63-ptr1 ← (RS)63-i end end ptr1 ← ptr1 + 1
180 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).
182 signed/unsigned min/max gives more flexibility.
185 uint_xlen_t min(uint_xlen_t rs1, uint_xlen_t rs2)
186 { return (int_xlen_t)rs1 < (int_xlen_t)rs2 ? rs1 : rs2;
188 uint_xlen_t max(uint_xlen_t rs1, uint_xlen_t rs2)
189 { return (int_xlen_t)rs1 > (int_xlen_t)rs2 ? rs1 : rs2;
191 uint_xlen_t minu(uint_xlen_t rs1, uint_xlen_t rs2)
192 { return rs1 < rs2 ? rs1 : rs2;
194 uint_xlen_t maxu(uint_xlen_t rs1, uint_xlen_t rs2)
195 { return rs1 > rs2 ? rs1 : rs2;
202 Similar to FPGA LUTs: for every bit perform a lookup into a table using an 8bit immediate, or in another register
204 | 0.5|6.10|11.15|16.20| 21..25| 26..30 |31|
205 | -- | -- | --- | --- | ----- | -------- |--|
206 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc|
209 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
210 RT[i] = (imm & (1<<idx)) != 0
212 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.
214 a 4 operand variant which becomes more along the lines of an FPGA:
216 | 0.5|6.10|11.15|16.20|21.25| 26...30 |31|
217 | -- | -- | --- | --- | --- | -------- |--|
218 | NN | RT | RA | RB | RC | mode 001 |Rc|
221 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
222 RT[i] = (RC & (1<<idx)) != 0
224 mode (2 bit) may be used to do inversion of ordering, similar to carryless mul,
227 also, another possible variant involving swizzle and vec4:
229 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31|
230 | -- | -- | --- | ----- | ---- | ----- |--|
231 | NN | RT | RA | imm | mask | 101 |1 |
234 idx = RA.x[i] << 2 | RA.y[i] << 1 | RA.z[i]
235 res = (imm & (1<<idx)) != 0
237 if mask[j]: RT[i+j*8] = res
239 another mode selection would be CRs not Ints.
241 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31|
242 | -- | -- | --- | --- |- |-----|----- | -----|--|
243 | NN | BA | BB | BC |0 |imm | mask | 101 |0 |
246 if not mask[i] continue
247 idx = crregs[BA][i] << 2 |
250 crregs[BA][i] = (imm & (1<<idx)) != 0
254 based on RV bitmanip singlebit set, instruction format similar to shift
255 [[isa/fixedshift]]. bmext is actually covered already (shift-with-mask rldicl but only immediate version).
256 however bitmask-invert is not, and set/clr are not covered, although they can use the same Shift ALU.
258 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.
259 bmrev however there is no direct equivalent and consequently a bmrevi is required.
261 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
262 | -- | -- | --- | --- | --- | ------- |--| ----- |
263 | NN | RT | RA | RB | RC | mode 010 |Rc| bm* |
264 | NN | RT | RA | RB | RC | 0 1 111 |Rc| bmrev |
268 uint_xlen_t bmset(RA, RB, sh)
270 int shamt = RB & (XLEN - 1);
272 return RA | (mask << shamt);
275 uint_xlen_t bmclr(RA, RB, sh)
277 int shamt = RB & (XLEN - 1);
279 return RA & ~(mask << shamt);
282 uint_xlen_t bminv(RA, RB, sh)
284 int shamt = RB & (XLEN - 1);
286 return RA ^ (mask << shamt);
289 uint_xlen_t bmext(RA, RB, sh)
291 int shamt = RB & (XLEN - 1);
293 return mask & (RA >> shamt);
297 bitmask extract with reverse. can be done by bitinverting all of RA and getting bits of RA from the opposite end.
301 rev[0:msb] = ra[msb:0];
304 uint_xlen_t bmextrev(RA, RB, sh)
306 int shamt = (RB & (XLEN - 1));
307 shamt = (XLEN-1)-shamt; # shift other end
308 bra = bitreverse(RA) # swap LSB-MSB
310 return mask & (bra >> shamt);
314 | 0.5|6.10|11.15|16.20|21.26| 27..30 |31| name |
315 | -- | -- | --- | --- | --- | ------- |--| ------ |
316 | NN | RT | RA | RB | sh | 0 111 |Rc| bmrevi |
325 uint64_t grev64(uint64_t RA, uint64_t RB)
329 if (shamt & 1) x = ((x & 0x5555555555555555LL) << 1) |
330 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
331 if (shamt & 2) x = ((x & 0x3333333333333333LL) << 2) |
332 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
333 if (shamt & 4) x = ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
334 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
335 if (shamt & 8) x = ((x & 0x00FF00FF00FF00FFLL) << 8) |
336 ((x & 0xFF00FF00FF00FF00LL) >> 8);
337 if (shamt & 16) x = ((x & 0x0000FFFF0000FFFFLL) << 16) |
338 ((x & 0xFFFF0000FFFF0000LL) >> 16);
339 if (shamt & 32) x = ((x & 0x00000000FFFFFFFFLL) << 32) |
340 ((x & 0xFFFFFFFF00000000LL) >> 32);
346 # shuffle / unshuffle
351 uint32_t shfl32(uint32_t RA, uint32_t RB)
355 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
356 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
357 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
358 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
361 uint32_t unshfl32(uint32_t RA, uint32_t RB)
365 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
366 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
367 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
368 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
372 uint64_t shuffle64_stage(uint64_t src, uint64_t maskL, uint64_t maskR, int N)
374 uint64_t x = src & ~(maskL | maskR);
375 x |= ((src << N) & maskL) | ((src >> N) & maskR);
378 uint64_t shfl64(uint64_t RA, uint64_t RB)
382 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
383 0x00000000ffff0000LL, 16);
384 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
385 0x0000ff000000ff00LL, 8);
386 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
387 0x00f000f000f000f0LL, 4);
388 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
389 0x0c0c0c0c0c0c0c0cLL, 2);
390 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
391 0x2222222222222222LL, 1);
394 uint64_t unshfl64(uint64_t RA, uint64_t RB)
398 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
399 0x2222222222222222LL, 1);
400 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
401 0x0c0c0c0c0c0c0c0cLL, 2);
402 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
403 0x00f000f000f000f0LL, 4);
404 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
405 0x0000ff000000ff00LL, 8);
406 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
407 0x00000000ffff0000LL, 16);
417 uint_xlen_t xperm(uint_xlen_t RA, uint_xlen_t RB, int sz_log2)
420 uint_xlen_t sz = 1LL << sz_log2;
421 uint_xlen_t mask = (1LL << sz) - 1;
422 for (int i = 0; i < XLEN; i += sz) {
423 uint_xlen_t pos = ((RB >> i) & mask) << sz_log2;
425 r |= ((RA >> pos) & mask) << i;
429 uint_xlen_t xperm_n (uint_xlen_t RA, uint_xlen_t RB)
430 { return xperm(RA, RB, 2); }
431 uint_xlen_t xperm_b (uint_xlen_t RA, uint_xlen_t RB)
432 { return xperm(RA, RB, 3); }
433 uint_xlen_t xperm_h (uint_xlen_t RA, uint_xlen_t RB)
434 { return xperm(RA, RB, 4); }
435 uint_xlen_t xperm_w (uint_xlen_t RA, uint_xlen_t RB)
436 { return xperm(RA, RB, 5); }
444 uint32_t gorc32(uint32_t RA, uint32_t RB)
448 if (shamt & 1) x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
449 if (shamt & 2) x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
450 if (shamt & 4) x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
451 if (shamt & 8) x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
452 if (shamt & 16) x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
455 uint64_t gorc64(uint64_t RA, uint64_t RB)
459 if (shamt & 1) x |= ((x & 0x5555555555555555LL) << 1) |
460 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
461 if (shamt & 2) x |= ((x & 0x3333333333333333LL) << 2) |
462 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
463 if (shamt & 4) x |= ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
464 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
465 if (shamt & 8) x |= ((x & 0x00FF00FF00FF00FFLL) << 8) |
466 ((x & 0xFF00FF00FF00FF00LL) >> 8);
467 if (shamt & 16) x |= ((x & 0x0000FFFF0000FFFFLL) << 16) |
468 ((x & 0xFFFF0000FFFF0000LL) >> 16);
469 if (shamt & 32) x |= ((x & 0x00000000FFFFFFFFLL) << 32) |
470 ((x & 0xFFFFFFFF00000000LL) >> 32);
478 based on RV bitmanip, covered by ternary bitops
481 uint_xlen_t cmix(uint_xlen_t RA, uint_xlen_t RB, uint_xlen_t RC) {
482 return (RA & RB) | (RC & ~RB);
489 see https://en.wikipedia.org/wiki/CLMUL_instruction_set
492 uint_xlen_t clmul(uint_xlen_t RA, uint_xlen_t RB)
495 for (int i = 0; i < XLEN; i++)
500 uint_xlen_t clmulh(uint_xlen_t RA, uint_xlen_t RB)
503 for (int i = 1; i < XLEN; i++)
508 uint_xlen_t clmulr(uint_xlen_t RA, uint_xlen_t RB)
511 for (int i = 0; i < XLEN; i++)
513 x ^= RA >> (XLEN-i-1);
519 see <https://courses.csail.mit.edu/6.857/2016/files/ffield.py>
523 this requires 3 parameters and a "degree"
525 RT = GFMUL(RA, RB, gfdegree, modulo=RC)
527 realistically with the degree also needing to be an immediate it should be brought down to an overwrite version:
529 RS = GFMUL(RS, RA, gfdegree, modulo=RB)
530 RS = GFMUL(RS, RA, gfdegree=RC, modulo=RB)
532 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31|
533 | -- | -- | --- | --- | --- | ------- |--|
534 | NN | RS | RA | RB | deg | 00 011 |Rc|
535 | NN | RS | RA | RB | RC | 11 011 |Rc|
537 where the SimpleV variant may override RS-as-src differently from RS-as-dest
542 from functools import reduce
544 # constants used in the multGF2 function
545 mask1 = mask2 = polyred = None
547 def setGF2(degree, irPoly):
548 """Define parameters of binary finite field GF(2^m)/g(x)
549 - degree: extension degree of binary field
550 - irPoly: coefficients of irreducible polynomial g(x)
553 """Convert an integer into a polynomial"""
554 return [(sInt >> i) & 1
555 for i in reversed(range(sInt.bit_length()))]
557 global mask1, mask2, polyred
558 mask1 = mask2 = 1 << degree
560 polyred = reduce(lambda x, y: (x << 1) + y, i2P(irPoly)[1:])
563 """Multiply two polynomials in GF(2^m)/g(x)"""
574 if __name__ == "__main__":
576 # Define binary field GF(2^3)/x^3 + x + 1
579 # Evaluate the product (x^2 + x + 1)(x^2 + 1)
580 print("{:02x}".format(multGF2(0b111, 0b101)))
582 # Define binary field GF(2^8)/x^8 + x^4 + x^3 + x + 1
583 # (used in the Advanced Encryption Standard-AES)
584 setGF2(8, 0b100011011)
586 # Evaluate the product (x^7)(x^7 + x + 1)
587 print("{:02x}".format(multGF2(0b10000000, 0b10000011)))
591 RS = GFADDI(RS, RA|0, gfdegree, modulo=RB)
592 RS = GFADD(RS, RA|0, gfdegree=RC, modulo=RB)
594 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
595 | -- | -- | --- | --- | --- | ------- |--| ----- |
596 | NN | RS | RA | RB | deg | 0 1 011 |Rc| gfaddi |
597 | NN | RS | RA | RB | RC | 1 1 111 |Rc| gfadd |
599 GFMOD is a pseudo-op where RA=0
612 def gf_invert(a, mod=0x1B) :
627 a %= 256 # Emulating 8-bit overflow
628 g1 %= 256 # Emulating 8-bit overflow
630 j = gf_degree(a) - gf_degree(v)
638 uint64_t bmatflip(uint64_t RA)
646 uint64_t bmatxor(uint64_t RA, uint64_t RB)
649 uint64_t RBt = bmatflip(RB);
650 uint8_t u[8]; // rows of RA
651 uint8_t v[8]; // cols of RB
652 for (int i = 0; i < 8; i++) {
657 for (int i = 0; i < 64; i++) {
658 if (pcnt(u[i / 8] & v[i % 8]) & 1)
663 uint64_t bmator(uint64_t RA, uint64_t RB)
666 uint64_t RBt = bmatflip(RB);
667 uint8_t u[8]; // rows of RA
668 uint8_t v[8]; // cols of RB
669 for (int i = 0; i < 8; i++) {
674 for (int i = 0; i < 64; i++) {
675 if ((u[i / 8] & v[i % 8]) != 0)