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 | deg | 01 011 |Rc| gfadd |
65 | NN | RT | RA | RB | deg | 10 011 |Rc| gfinv |
66 | NN | RS | RA | RB | deg | 11 011 |Rc| gfmuli |
68 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31| name |
69 | -- | -- | --- | ----- | ---- | ----- |--| ------ |
70 | NN | RT | RA | imm | mask | 101 |1 | ternaryv |
72 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31| name |
73 | -- | -- | --- | --- |- |-----|----- | -----|--| -------|
74 | NN | BA | BB | BC |0 |imm | mask | 101 |0 | ternarycr |
78 | 0.5|6.10|11.15|16.20| 21.22 | 23 | 24....30 |31| name |
79 | -- | -- | --- | --- | ----- | -- | -------- |--| ---- |
80 | NN | RA | RB | | | 0 | 0000 110 |Rc| rsvd |
81 | NN | RA | RB | RC | itype | 1 | 0000 110 |Rc| xperm |
82 | NN | RA | RB | RC | itype | 0 | 0100 110 |Rc| minmax |
83 | NN | RA | RB | | | 1 | 0100 110 |Rc| rsvd |
84 | NN | RA | RB | sh | itype | SH | 1000 110 |Rc| bmopsi |
85 | NN | RA | RB | | | | 1100 110 |Rc| rsvd |
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 | | | 0 | 0001 110 |Rc| rsvd |
90 | NN | RA | RB | | | 0 | 0101 110 |Rc| rsvd |
91 | NN | RA | RB | RC | 00 | 0 | 0010 110 |Rc| gorc |
92 | NN | RA | RB | sh | 00 | SH | 1010 110 |Rc| gorci |
93 | NN | RA | RB | RC | 00 | 0 | 0110 110 |Rc| gorcw |
94 | NN | RA | RB | sh | 00 | 0 | 1110 110 |Rc| gorcwi |
95 | NN | RA | RB | RC | 00 | 1 | 1110 110 |Rc| bmator |
96 | NN | RA | RB | RC | 01 | 0 | 0010 110 |Rc| grev |
97 | NN | RA | RB | RC | 01 | 1 | 0010 110 |Rc| clmul |
98 | NN | RA | RB | sh | 01 | SH | 1010 110 |Rc| grevi |
99 | NN | RA | RB | RC | 01 | 0 | 0110 110 |Rc| grevw |
100 | NN | RA | RB | sh | 01 | 0 | 1110 110 |Rc| grevwi |
101 | NN | RA | RB | RC | 01 | 1 | 1110 110 |Rc| bmatxor |
102 | NN | RA | RB | RC | 10 | 0 | 0010 110 |Rc| shfl |
103 | NN | RA | RB | sh | 10 | SH | 1010 110 |Rc| shfli |
104 | NN | RA | RB | RC | 10 | 0 | 0110 110 |Rc| shflw |
105 | NN | RA | RB | RC | 10 | 0 | 1110 110 |Rc| bdep |
106 | NN | RA | RB | RC | 10 | 1 | 1110 110 |Rc| bext |
107 | NN | RA | RB | RC | 11 | 0 | 1110 110 |Rc| clmulr |
108 | NN | RA | RB | RC | 11 | 1 | 1110 110 |Rc| clmulh |
109 | NN | RA | RB | | | | NN11 110 |Rc| rsvd |
111 # bit to byte permute
113 similar to matrix permute in RV bitmanip, which has XOR and OR variants
117 b = VSR[VRB+32].dword[i].byte[k].bit[j]
118 VSR[VRT+32].dword[i].byte[j].bit[k] = b
122 vpdepd VRT,VRA,VRB, identical to RV bitmamip bdep
125 if VSR[VRB+32].dword[i].bit[63-m]=1 then do
126 result = VSR[VRA+32].dword[i].bit[63-k]
127 VSR[VRT+32].dword[i].bit[63-m] = result
133 uint_xlen_t bdep(uint_xlen_t RA, uint_xlen_t RB)
136 for (int i = 0, j = 0; i < XLEN; i++)
139 r |= uint_xlen_t(1) << i;
149 other way round: identical to RV bext
152 uint_xlen_t bext(uint_xlen_t RA, uint_xlen_t RB)
155 for (int i = 0, j = 0; i < XLEN; i++)
158 r |= uint_xlen_t(1) << j;
167 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).
169 signed/unsigned min/max gives more flexibility.
172 uint_xlen_t min(uint_xlen_t rs1, uint_xlen_t rs2)
173 { return (int_xlen_t)rs1 < (int_xlen_t)rs2 ? rs1 : rs2;
175 uint_xlen_t max(uint_xlen_t rs1, uint_xlen_t rs2)
176 { return (int_xlen_t)rs1 > (int_xlen_t)rs2 ? rs1 : rs2;
178 uint_xlen_t minu(uint_xlen_t rs1, uint_xlen_t rs2)
179 { return rs1 < rs2 ? rs1 : rs2;
181 uint_xlen_t maxu(uint_xlen_t rs1, uint_xlen_t rs2)
182 { return rs1 > rs2 ? rs1 : rs2;
189 Similar to FPGA LUTs: for every bit perform a lookup into a table using an 8bit immediate, or in another register
191 | 0.5|6.10|11.15|16.20| 21..25| 26..30 |31|
192 | -- | -- | --- | --- | ----- | -------- |--|
193 | NN | RT | RA | RB | im0-4 | im5-7 00 |Rc|
196 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
197 RT[i] = (imm & (1<<idx)) != 0
199 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.
201 a 4 operand variant which becomes more along the lines of an FPGA:
203 | 0.5|6.10|11.15|16.20|21.25| 26...30 |31|
204 | -- | -- | --- | --- | --- | -------- |--|
205 | NN | RT | RA | RB | RC | mode 001 |Rc|
208 idx = RT[i] << 2 | RA[i] << 1 | RB[i]
209 RT[i] = (RC & (1<<idx)) != 0
211 mode (2 bit) may be used to do inversion of ordering, similar to carryless mul,
214 also, another possible variant involving swizzle and vec4:
216 | 0.5|6.10|11.15| 16.23 |24.27 | 28.30 |31|
217 | -- | -- | --- | ----- | ---- | ----- |--|
218 | NN | RT | RA | imm | mask | 101 |1 |
221 idx = RA.x[i] << 2 | RA.y[i] << 1 | RA.z[i]
222 res = (imm & (1<<idx)) != 0
224 if mask[j]: RT[i+j*8] = res
226 another mode selection would be CRs not Ints.
228 | 0.5|6.8 | 9.11|12.14|15|16.23|24.27 | 28.30|31|
229 | -- | -- | --- | --- |- |-----|----- | -----|--|
230 | NN | BA | BB | BC |0 |imm | mask | 101 |0 |
233 if not mask[i] continue
234 idx = crregs[BA][i] << 2 |
237 crregs[BA][i] = (imm & (1<<idx)) != 0
241 based on RV bitmanip singlebit set, instruction format similar to shift
242 [[isa/fixedshift]]. bmext is actually covered already (shift-with-mask rldicl but only immediate version).
243 however bitmask-invert is not, and set/clr are not covered, although they can use the same Shift ALU.
245 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.
246 bmrev however there is no direct equivalent and consequently a bmrevi is required.
248 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
249 | -- | -- | --- | --- | --- | ------- |--| ----- |
250 | NN | RT | RA | RB | RC | mode 010 |Rc| bm* |
251 | NN | RT | RA | RB | RC | 0 1 111 |Rc| bmrev |
255 uint_xlen_t bmset(RA, RB, sh)
257 int shamt = RB & (XLEN - 1);
259 return RA | (mask << shamt);
262 uint_xlen_t bmclr(RA, RB, sh)
264 int shamt = RB & (XLEN - 1);
266 return RA & ~(mask << shamt);
269 uint_xlen_t bminv(RA, RB, sh)
271 int shamt = RB & (XLEN - 1);
273 return RA ^ (mask << shamt);
276 uint_xlen_t bmext(RA, RB, sh)
278 int shamt = RB & (XLEN - 1);
280 return mask & (RA >> shamt);
284 bitmask extract with reverse. can be done by bitinverting all of RA and getting bits of RA from the opposite end.
288 rev[0:msb] = ra[msb:0];
291 uint_xlen_t bmextrev(RA, RB, sh)
293 int shamt = (RB & (XLEN - 1));
294 shamt = (XLEN-1)-shamt; # shift other end
295 bra = bitreverse(RA) # swap LSB-MSB
297 return mask & (bra >> shamt);
301 | 0.5|6.10|11.15|16.20|21.26| 27..30 |31| name |
302 | -- | -- | --- | --- | --- | ------- |--| ------ |
303 | NN | RT | RA | RB | sh | 0 111 |Rc| bmrevi |
312 uint64_t grev64(uint64_t RA, uint64_t RB)
316 if (shamt & 1) x = ((x & 0x5555555555555555LL) << 1) |
317 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
318 if (shamt & 2) x = ((x & 0x3333333333333333LL) << 2) |
319 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
320 if (shamt & 4) x = ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
321 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
322 if (shamt & 8) x = ((x & 0x00FF00FF00FF00FFLL) << 8) |
323 ((x & 0xFF00FF00FF00FF00LL) >> 8);
324 if (shamt & 16) x = ((x & 0x0000FFFF0000FFFFLL) << 16) |
325 ((x & 0xFFFF0000FFFF0000LL) >> 16);
326 if (shamt & 32) x = ((x & 0x00000000FFFFFFFFLL) << 32) |
327 ((x & 0xFFFFFFFF00000000LL) >> 32);
333 # shuffle / unshuffle
338 uint32_t shfl32(uint32_t RA, uint32_t RB)
342 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
343 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
344 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
345 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
348 uint32_t unshfl32(uint32_t RA, uint32_t RB)
352 if (shamt & 1) x = shuffle32_stage(x, 0x44444444, 0x22222222, 1);
353 if (shamt & 2) x = shuffle32_stage(x, 0x30303030, 0x0c0c0c0c, 2);
354 if (shamt & 4) x = shuffle32_stage(x, 0x0f000f00, 0x00f000f0, 4);
355 if (shamt & 8) x = shuffle32_stage(x, 0x00ff0000, 0x0000ff00, 8);
359 uint64_t shuffle64_stage(uint64_t src, uint64_t maskL, uint64_t maskR, int N)
361 uint64_t x = src & ~(maskL | maskR);
362 x |= ((src << N) & maskL) | ((src >> N) & maskR);
365 uint64_t shfl64(uint64_t RA, uint64_t RB)
369 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
370 0x00000000ffff0000LL, 16);
371 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
372 0x0000ff000000ff00LL, 8);
373 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
374 0x00f000f000f000f0LL, 4);
375 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
376 0x0c0c0c0c0c0c0c0cLL, 2);
377 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
378 0x2222222222222222LL, 1);
381 uint64_t unshfl64(uint64_t RA, uint64_t RB)
385 if (shamt & 1) x = shuffle64_stage(x, 0x4444444444444444LL,
386 0x2222222222222222LL, 1);
387 if (shamt & 2) x = shuffle64_stage(x, 0x3030303030303030LL,
388 0x0c0c0c0c0c0c0c0cLL, 2);
389 if (shamt & 4) x = shuffle64_stage(x, 0x0f000f000f000f00LL,
390 0x00f000f000f000f0LL, 4);
391 if (shamt & 8) x = shuffle64_stage(x, 0x00ff000000ff0000LL,
392 0x0000ff000000ff00LL, 8);
393 if (shamt & 16) x = shuffle64_stage(x, 0x0000ffff00000000LL,
394 0x00000000ffff0000LL, 16);
404 uint_xlen_t xperm(uint_xlen_t RA, uint_xlen_t RB, int sz_log2)
407 uint_xlen_t sz = 1LL << sz_log2;
408 uint_xlen_t mask = (1LL << sz) - 1;
409 for (int i = 0; i < XLEN; i += sz) {
410 uint_xlen_t pos = ((RB >> i) & mask) << sz_log2;
412 r |= ((RA >> pos) & mask) << i;
416 uint_xlen_t xperm_n (uint_xlen_t RA, uint_xlen_t RB)
417 { return xperm(RA, RB, 2); }
418 uint_xlen_t xperm_b (uint_xlen_t RA, uint_xlen_t RB)
419 { return xperm(RA, RB, 3); }
420 uint_xlen_t xperm_h (uint_xlen_t RA, uint_xlen_t RB)
421 { return xperm(RA, RB, 4); }
422 uint_xlen_t xperm_w (uint_xlen_t RA, uint_xlen_t RB)
423 { return xperm(RA, RB, 5); }
431 uint32_t gorc32(uint32_t RA, uint32_t RB)
435 if (shamt & 1) x |= ((x & 0x55555555) << 1) | ((x & 0xAAAAAAAA) >> 1);
436 if (shamt & 2) x |= ((x & 0x33333333) << 2) | ((x & 0xCCCCCCCC) >> 2);
437 if (shamt & 4) x |= ((x & 0x0F0F0F0F) << 4) | ((x & 0xF0F0F0F0) >> 4);
438 if (shamt & 8) x |= ((x & 0x00FF00FF) << 8) | ((x & 0xFF00FF00) >> 8);
439 if (shamt & 16) x |= ((x & 0x0000FFFF) << 16) | ((x & 0xFFFF0000) >> 16);
442 uint64_t gorc64(uint64_t RA, uint64_t RB)
446 if (shamt & 1) x |= ((x & 0x5555555555555555LL) << 1) |
447 ((x & 0xAAAAAAAAAAAAAAAALL) >> 1);
448 if (shamt & 2) x |= ((x & 0x3333333333333333LL) << 2) |
449 ((x & 0xCCCCCCCCCCCCCCCCLL) >> 2);
450 if (shamt & 4) x |= ((x & 0x0F0F0F0F0F0F0F0FLL) << 4) |
451 ((x & 0xF0F0F0F0F0F0F0F0LL) >> 4);
452 if (shamt & 8) x |= ((x & 0x00FF00FF00FF00FFLL) << 8) |
453 ((x & 0xFF00FF00FF00FF00LL) >> 8);
454 if (shamt & 16) x |= ((x & 0x0000FFFF0000FFFFLL) << 16) |
455 ((x & 0xFFFF0000FFFF0000LL) >> 16);
456 if (shamt & 32) x |= ((x & 0x00000000FFFFFFFFLL) << 32) |
457 ((x & 0xFFFFFFFF00000000LL) >> 32);
465 based on RV bitmanip, covered by ternary bitops
468 uint_xlen_t cmix(uint_xlen_t RA, uint_xlen_t RB, uint_xlen_t RC) {
469 return (RA & RB) | (RC & ~RB);
476 see https://en.wikipedia.org/wiki/CLMUL_instruction_set
479 uint_xlen_t clmul(uint_xlen_t RA, uint_xlen_t RB)
482 for (int i = 0; i < XLEN; i++)
487 uint_xlen_t clmulh(uint_xlen_t RA, uint_xlen_t RB)
490 for (int i = 1; i < XLEN; i++)
495 uint_xlen_t clmulr(uint_xlen_t RA, uint_xlen_t RB)
498 for (int i = 0; i < XLEN; i++)
500 x ^= RA >> (XLEN-i-1);
506 see <https://courses.csail.mit.edu/6.857/2016/files/ffield.py>
510 this requires 3 parameters and a "degree"
512 RT = GFMUL(RA, RB, gfdegree, modulo=RC)
514 realistically with the degree also needing to be an immediate it should be brought down to an overwrite version:
516 RS = GFMUL(RS, RA, gfdegree, modulo=RB)
517 RS = GFMUL(RS, RA, gfdegree=RC, modulo=RB)
519 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31|
520 | -- | -- | --- | --- | --- | ------- |--|
521 | NN | RS | RA | RB | deg | 00 011 |Rc|
522 | NN | RS | RA | RB | RC | 11 011 |Rc|
524 where the SimpleV variant may override RS-as-src differently from RS-as-dest
529 from functools import reduce
531 # constants used in the multGF2 function
532 mask1 = mask2 = polyred = None
534 def setGF2(degree, irPoly):
535 """Define parameters of binary finite field GF(2^m)/g(x)
536 - degree: extension degree of binary field
537 - irPoly: coefficients of irreducible polynomial g(x)
540 """Convert an integer into a polynomial"""
541 return [(sInt >> i) & 1
542 for i in reversed(range(sInt.bit_length()))]
544 global mask1, mask2, polyred
545 mask1 = mask2 = 1 << degree
547 polyred = reduce(lambda x, y: (x << 1) + y, i2P(irPoly)[1:])
550 """Multiply two polynomials in GF(2^m)/g(x)"""
561 if __name__ == "__main__":
563 # Define binary field GF(2^3)/x^3 + x + 1
566 # Evaluate the product (x^2 + x + 1)(x^2 + 1)
567 print("{:02x}".format(multGF2(0b111, 0b101)))
569 # Define binary field GF(2^8)/x^8 + x^4 + x^3 + x + 1
570 # (used in the Advanced Encryption Standard-AES)
571 setGF2(8, 0b100011011)
573 # Evaluate the product (x^7)(x^7 + x + 1)
574 print("{:02x}".format(multGF2(0b10000000, 0b10000011)))
578 RS = GFADD(RS, RA|0, gfdegree, modulo=RB)
579 RS = GFADDI(RS, RA|0, gfdegree=RC, modulo=RB)
581 | 0.5|6.10|11.15|16.20|21.25| 26..30 |31| name |
582 | -- | -- | --- | --- | --- | ------- |--| ----- |
583 | NN | RS | RA | RB | RC | 0 1 011 |Rc| gfadd |
584 | NN | RS | RA | RB | RC | 1 1 111 |Rc| gfaddi |
586 GFMOD is a pseudo-op where RA=0
599 def gf_invert(a, mod=0x1B) :
614 a %= 256 # Emulating 8-bit overflow
615 g1 %= 256 # Emulating 8-bit overflow
617 j = gf_degree(a) - gf_degree(v)
625 uint64_t bmatflip(uint64_t RA)
633 uint64_t bmatxor(uint64_t RA, uint64_t RB)
636 uint64_t RBt = bmatflip(RB);
637 uint8_t u[8]; // rows of RA
638 uint8_t v[8]; // cols of RB
639 for (int i = 0; i < 8; i++) {
644 for (int i = 0; i < 64; i++) {
645 if (pcnt(u[i / 8] & v[i % 8]) & 1)
650 uint64_t bmator(uint64_t RA, uint64_t RB)
653 uint64_t RBt = bmatflip(RB);
654 uint8_t u[8]; // rows of RA
655 uint8_t v[8]; // cols of RB
656 for (int i = 0; i < 8; i++) {
661 for (int i = 0; i < 64; i++) {
662 if ((u[i / 8] & v[i % 8]) != 0)