fe_generic.mx raw

   1  // Copyright (c) 2017 The Go Authors. All rights reserved.
   2  // Use of this source code is governed by a BSD-style
   3  // license that can be found in the LICENSE file.
   4  
   5  package field
   6  
   7  import "math/bits"
   8  
   9  // uint128 holds a 128-bit number as two 64-bit limbs, for use with the
  10  // bits.Mul64 and bits.Add64 intrinsics.
  11  type uint128 struct {
  12  	lo, hi uint64
  13  }
  14  
  15  // mul returns a * b.
  16  func mul(a, b uint64) uint128 {
  17  	hi, lo := bits.Mul64(a, b)
  18  	return uint128{lo, hi}
  19  }
  20  
  21  // addMul returns v + a * b.
  22  func addMul(v uint128, a, b uint64) uint128 {
  23  	hi, lo := bits.Mul64(a, b)
  24  	lo, c := bits.Add64(lo, v.lo, 0)
  25  	hi, _ = bits.Add64(hi, v.hi, c)
  26  	return uint128{lo, hi}
  27  }
  28  
  29  // mul19 returns v * 19.
  30  func mul19(v uint64) uint64 {
  31  	// Using this approach seems to yield better optimizations than *19.
  32  	return v + (v+v<<3)<<1
  33  }
  34  
  35  // addMul19 returns v + 19 * a * b, where a and b are at most 52 bits.
  36  func addMul19(v uint128, a, b uint64) uint128 {
  37  	hi, lo := bits.Mul64(mul19(a), b)
  38  	lo, c := bits.Add64(lo, v.lo, 0)
  39  	hi, _ = bits.Add64(hi, v.hi, c)
  40  	return uint128{lo, hi}
  41  }
  42  
  43  // addMul38 returns v + 38 * a * b, where a and b are at most 52 bits.
  44  func addMul38(v uint128, a, b uint64) uint128 {
  45  	hi, lo := bits.Mul64(mul19(a), b*2)
  46  	lo, c := bits.Add64(lo, v.lo, 0)
  47  	hi, _ = bits.Add64(hi, v.hi, c)
  48  	return uint128{lo, hi}
  49  }
  50  
  51  // shiftRightBy51 returns a >> 51. a is assumed to be at most 115 bits.
  52  func shiftRightBy51(a uint128) uint64 {
  53  	return (a.hi << (64 - 51)) | (a.lo >> 51)
  54  }
  55  
  56  func feMulGeneric(v, a, b *Element) {
  57  	a0 := a.l0
  58  	a1 := a.l1
  59  	a2 := a.l2
  60  	a3 := a.l3
  61  	a4 := a.l4
  62  
  63  	b0 := b.l0
  64  	b1 := b.l1
  65  	b2 := b.l2
  66  	b3 := b.l3
  67  	b4 := b.l4
  68  
  69  	// Limb multiplication works like pen-and-paper columnar multiplication, but
  70  	// with 51-bit limbs instead of digits.
  71  	//
  72  	//                          a4   a3   a2   a1   a0  x
  73  	//                          b4   b3   b2   b1   b0  =
  74  	//                         ------------------------
  75  	//                        a4b0 a3b0 a2b0 a1b0 a0b0  +
  76  	//                   a4b1 a3b1 a2b1 a1b1 a0b1       +
  77  	//              a4b2 a3b2 a2b2 a1b2 a0b2            +
  78  	//         a4b3 a3b3 a2b3 a1b3 a0b3                 +
  79  	//    a4b4 a3b4 a2b4 a1b4 a0b4                      =
  80  	//   ----------------------------------------------
  81  	//      r8   r7   r6   r5   r4   r3   r2   r1   r0
  82  	//
  83  	// We can then use the reduction identity (a * 2²⁵⁵ + b = a * 19 + b) to
  84  	// reduce the limbs that would overflow 255 bits. r5 * 2²⁵⁵ becomes 19 * r5,
  85  	// r6 * 2³⁰⁶ becomes 19 * r6 * 2⁵¹, etc.
  86  	//
  87  	// Reduction can be carried out simultaneously to multiplication. For
  88  	// example, we do not compute r5: whenever the result of a multiplication
  89  	// belongs to r5, like a1b4, we multiply it by 19 and add the result to r0.
  90  	//
  91  	//            a4b0    a3b0    a2b0    a1b0    a0b0  +
  92  	//            a3b1    a2b1    a1b1    a0b1 19×a4b1  +
  93  	//            a2b2    a1b2    a0b2 19×a4b2 19×a3b2  +
  94  	//            a1b3    a0b3 19×a4b3 19×a3b3 19×a2b3  +
  95  	//            a0b4 19×a4b4 19×a3b4 19×a2b4 19×a1b4  =
  96  	//           --------------------------------------
  97  	//              r4      r3      r2      r1      r0
  98  	//
  99  	// Finally we add up the columns into wide, overlapping limbs.
 100  
 101  	// r0 = a0×b0 + 19×(a1×b4 + a2×b3 + a3×b2 + a4×b1)
 102  	r0 := mul(a0, b0)
 103  	r0 = addMul19(r0, a1, b4)
 104  	r0 = addMul19(r0, a2, b3)
 105  	r0 = addMul19(r0, a3, b2)
 106  	r0 = addMul19(r0, a4, b1)
 107  
 108  	// r1 = a0×b1 + a1×b0 + 19×(a2×b4 + a3×b3 + a4×b2)
 109  	r1 := mul(a0, b1)
 110  	r1 = addMul(r1, a1, b0)
 111  	r1 = addMul19(r1, a2, b4)
 112  	r1 = addMul19(r1, a3, b3)
 113  	r1 = addMul19(r1, a4, b2)
 114  
 115  	// r2 = a0×b2 + a1×b1 + a2×b0 + 19×(a3×b4 + a4×b3)
 116  	r2 := mul(a0, b2)
 117  	r2 = addMul(r2, a1, b1)
 118  	r2 = addMul(r2, a2, b0)
 119  	r2 = addMul19(r2, a3, b4)
 120  	r2 = addMul19(r2, a4, b3)
 121  
 122  	// r3 = a0×b3 + a1×b2 + a2×b1 + a3×b0 + 19×a4×b4
 123  	r3 := mul(a0, b3)
 124  	r3 = addMul(r3, a1, b2)
 125  	r3 = addMul(r3, a2, b1)
 126  	r3 = addMul(r3, a3, b0)
 127  	r3 = addMul19(r3, a4, b4)
 128  
 129  	// r4 = a0×b4 + a1×b3 + a2×b2 + a3×b1 + a4×b0
 130  	r4 := mul(a0, b4)
 131  	r4 = addMul(r4, a1, b3)
 132  	r4 = addMul(r4, a2, b2)
 133  	r4 = addMul(r4, a3, b1)
 134  	r4 = addMul(r4, a4, b0)
 135  
 136  	// After the multiplication, we need to reduce (carry) the five coefficients
 137  	// to obtain a result with limbs that are at most slightly larger than 2⁵¹,
 138  	// to respect the Element invariant.
 139  	//
 140  	// Overall, the reduction works the same as carryPropagate, except with
 141  	// wider inputs: we take the carry for each coefficient by shifting it right
 142  	// by 51, and add it to the limb above it. The top carry is multiplied by 19
 143  	// according to the reduction identity and added to the lowest limb.
 144  	//
 145  	// The largest coefficient (r0) will be at most 111 bits, which guarantees
 146  	// that all carries are at most 111 - 51 = 60 bits, which fits in a uint64.
 147  	//
 148  	//     r0 = a0×b0 + 19×(a1×b4 + a2×b3 + a3×b2 + a4×b1)
 149  	//     r0 < 2⁵²×2⁵² + 19×(2⁵²×2⁵² + 2⁵²×2⁵² + 2⁵²×2⁵² + 2⁵²×2⁵²)
 150  	//     r0 < (1 + 19 × 4) × 2⁵² × 2⁵²
 151  	//     r0 < 2⁷ × 2⁵² × 2⁵²
 152  	//     r0 < 2¹¹¹
 153  	//
 154  	// Moreover, the top coefficient (r4) is at most 107 bits, so c4 is at most
 155  	// 56 bits, and c4 * 19 is at most 61 bits, which again fits in a uint64 and
 156  	// allows us to easily apply the reduction identity.
 157  	//
 158  	//     r4 = a0×b4 + a1×b3 + a2×b2 + a3×b1 + a4×b0
 159  	//     r4 < 5 × 2⁵² × 2⁵²
 160  	//     r4 < 2¹⁰⁷
 161  	//
 162  
 163  	c0 := shiftRightBy51(r0)
 164  	c1 := shiftRightBy51(r1)
 165  	c2 := shiftRightBy51(r2)
 166  	c3 := shiftRightBy51(r3)
 167  	c4 := shiftRightBy51(r4)
 168  
 169  	rr0 := r0.lo&maskLow51Bits + mul19(c4)
 170  	rr1 := r1.lo&maskLow51Bits + c0
 171  	rr2 := r2.lo&maskLow51Bits + c1
 172  	rr3 := r3.lo&maskLow51Bits + c2
 173  	rr4 := r4.lo&maskLow51Bits + c3
 174  
 175  	// Now all coefficients fit into 64-bit registers but are still too large to
 176  	// be passed around as an Element. We therefore do one last carry chain,
 177  	// where the carries will be small enough to fit in the wiggle room above 2⁵¹.
 178  
 179  	v.l0 = rr0&maskLow51Bits + mul19(rr4>>51)
 180  	v.l1 = rr1&maskLow51Bits + rr0>>51
 181  	v.l2 = rr2&maskLow51Bits + rr1>>51
 182  	v.l3 = rr3&maskLow51Bits + rr2>>51
 183  	v.l4 = rr4&maskLow51Bits + rr3>>51
 184  }
 185  
 186  func feSquareGeneric(v, a *Element) {
 187  	l0 := a.l0
 188  	l1 := a.l1
 189  	l2 := a.l2
 190  	l3 := a.l3
 191  	l4 := a.l4
 192  
 193  	// Squaring works precisely like multiplication above, but thanks to its
 194  	// symmetry we get to group a few terms together.
 195  	//
 196  	//                          l4   l3   l2   l1   l0  x
 197  	//                          l4   l3   l2   l1   l0  =
 198  	//                         ------------------------
 199  	//                        l4l0 l3l0 l2l0 l1l0 l0l0  +
 200  	//                   l4l1 l3l1 l2l1 l1l1 l0l1       +
 201  	//              l4l2 l3l2 l2l2 l1l2 l0l2            +
 202  	//         l4l3 l3l3 l2l3 l1l3 l0l3                 +
 203  	//    l4l4 l3l4 l2l4 l1l4 l0l4                      =
 204  	//   ----------------------------------------------
 205  	//      r8   r7   r6   r5   r4   r3   r2   r1   r0
 206  	//
 207  	//            l4l0    l3l0    l2l0    l1l0    l0l0  +
 208  	//            l3l1    l2l1    l1l1    l0l1 19×l4l1  +
 209  	//            l2l2    l1l2    l0l2 19×l4l2 19×l3l2  +
 210  	//            l1l3    l0l3 19×l4l3 19×l3l3 19×l2l3  +
 211  	//            l0l4 19×l4l4 19×l3l4 19×l2l4 19×l1l4  =
 212  	//           --------------------------------------
 213  	//              r4      r3      r2      r1      r0
 214  
 215  	// r0 = l0×l0 + 19×(l1×l4 + l2×l3 + l3×l2 + l4×l1) = l0×l0 + 19×2×(l1×l4 + l2×l3)
 216  	r0 := mul(l0, l0)
 217  	r0 = addMul38(r0, l1, l4)
 218  	r0 = addMul38(r0, l2, l3)
 219  
 220  	// r1 = l0×l1 + l1×l0 + 19×(l2×l4 + l3×l3 + l4×l2) = 2×l0×l1 + 19×2×l2×l4 + 19×l3×l3
 221  	r1 := mul(l0*2, l1)
 222  	r1 = addMul38(r1, l2, l4)
 223  	r1 = addMul19(r1, l3, l3)
 224  
 225  	// r2 = l0×l2 + l1×l1 + l2×l0 + 19×(l3×l4 + l4×l3) = 2×l0×l2 + l1×l1 + 19×2×l3×l4
 226  	r2 := mul(l0*2, l2)
 227  	r2 = addMul(r2, l1, l1)
 228  	r2 = addMul38(r2, l3, l4)
 229  
 230  	// r3 = l0×l3 + l1×l2 + l2×l1 + l3×l0 + 19×l4×l4 = 2×l0×l3 + 2×l1×l2 + 19×l4×l4
 231  	r3 := mul(l0*2, l3)
 232  	r3 = addMul(r3, l1*2, l2)
 233  	r3 = addMul19(r3, l4, l4)
 234  
 235  	// r4 = l0×l4 + l1×l3 + l2×l2 + l3×l1 + l4×l0 = 2×l0×l4 + 2×l1×l3 + l2×l2
 236  	r4 := mul(l0*2, l4)
 237  	r4 = addMul(r4, l1*2, l3)
 238  	r4 = addMul(r4, l2, l2)
 239  
 240  	c0 := shiftRightBy51(r0)
 241  	c1 := shiftRightBy51(r1)
 242  	c2 := shiftRightBy51(r2)
 243  	c3 := shiftRightBy51(r3)
 244  	c4 := shiftRightBy51(r4)
 245  
 246  	rr0 := r0.lo&maskLow51Bits + mul19(c4)
 247  	rr1 := r1.lo&maskLow51Bits + c0
 248  	rr2 := r2.lo&maskLow51Bits + c1
 249  	rr3 := r3.lo&maskLow51Bits + c2
 250  	rr4 := r4.lo&maskLow51Bits + c3
 251  
 252  	v.l0 = rr0&maskLow51Bits + mul19(rr4>>51)
 253  	v.l1 = rr1&maskLow51Bits + rr0>>51
 254  	v.l2 = rr2&maskLow51Bits + rr1>>51
 255  	v.l3 = rr3&maskLow51Bits + rr2>>51
 256  	v.l4 = rr4&maskLow51Bits + rr3>>51
 257  }
 258  
 259  // carryPropagate brings the limbs below 52 bits by applying the reduction
 260  // identity (a * 2²⁵⁵ + b = a * 19 + b) to the l4 carry.
 261  func (v *Element) carryPropagate() *Element {
 262  	// (l4>>51) is at most 64 - 51 = 13 bits, so (l4>>51)*19 is at most 18 bits, and
 263  	// the final l0 will be at most 52 bits. Similarly for the rest.
 264  	l0 := v.l0
 265  	v.l0 = v.l0&maskLow51Bits + mul19(v.l4>>51)
 266  	v.l4 = v.l4&maskLow51Bits + v.l3>>51
 267  	v.l3 = v.l3&maskLow51Bits + v.l2>>51
 268  	v.l2 = v.l2&maskLow51Bits + v.l1>>51
 269  	v.l1 = v.l1&maskLow51Bits + l0>>51
 270  
 271  	return v
 272  }
 273