| 1 |
//go:generate go run src.go -out ../../f1600x4_amd64.s -stubs ../../f1600x4stubs_amd64.go -pkg keccakf1600 |
| 2 |
|
| 3 |
// AVX2 fourway parallelized Keccak-f[1600]. |
| 4 |
|
| 5 |
package main |
| 6 |
|
| 7 |
import ( |
| 8 |
. "github.com/mmcloughlin/avo/build" // nolint:stylecheck,golint |
| 9 |
. "github.com/mmcloughlin/avo/operand" // nolint:stylecheck,golint |
| 10 |
) |
| 11 |
|
| 12 |
// nolint:funlen |
| 13 |
func main() { |
| 14 |
ConstraintExpr("amd64,!purego") |
| 15 |
|
| 16 |
// Must be called on 32 byte aligned memory. |
| 17 |
TEXT("f1600x4AVX2", NOSPLIT, "func(state *uint64, rc *[24]uint64, turbo bool)") |
| 18 |
|
| 19 |
Pragma("noescape") |
| 20 |
|
| 21 |
statePtr := Load(Param("state"), GP64()) |
| 22 |
state := func(offset int) Op { |
| 23 |
return Mem{Base: statePtr, Disp: 32 * offset} |
| 24 |
} |
| 25 |
|
| 26 |
rcPtr := Load(Param("rc"), GP64()) |
| 27 |
|
| 28 |
// We use the same approach as the normal Keccak-f[1600] implementation |
| 29 |
// (in the internal/sha3 package): we group four rounds together into a |
| 30 |
// super round. Thus we have six super rounds. |
| 31 |
superRound := GP64() |
| 32 |
MOVQ(U64(6), superRound) // count down. |
| 33 |
|
| 34 |
turbo := Load(Param("turbo"), GP64()) |
| 35 |
TESTQ(turbo, turbo) |
| 36 |
JZ(LabelRef("loop")) |
| 37 |
|
| 38 |
MOVQ(U64(3), superRound) // Skip 3 * 4 = 12 rounds |
| 39 |
ADDQ(Imm(8*12), rcPtr) |
| 40 |
|
| 41 |
// XXX Because our AVX2 is significantly larger, it might better not |
| 42 |
// to group four rounds together, but simply loop over the rounds |
| 43 |
// themselves. |
| 44 |
|
| 45 |
Label("loop") |
| 46 |
|
| 47 |
for r := 0; r < 4; r++ { |
| 48 |
// Compute parities: p[i] = a[i] ^ a[i + 5] ^ ... ^ a[i + 20]. |
| 49 |
p := []Op{YMM(), YMM(), YMM(), YMM(), YMM()} |
| 50 |
for i := 0; i < 5; i++ { |
| 51 |
VMOVDQA(state(i), p[i]) |
| 52 |
} |
| 53 |
for j := 1; j < 5; j++ { |
| 54 |
for i := 0; i < 5; i++ { |
| 55 |
VPXOR(state(5*j+i), p[i], p[i]) |
| 56 |
} |
| 57 |
} |
| 58 |
|
| 59 |
// Rotate and xor parities: d[i] = rotate_left(p[i+1], 1) ^ p[i-1] |
| 60 |
t := []Op{YMM(), YMM(), YMM(), YMM(), YMM()} |
| 61 |
d := []Op{YMM(), YMM(), YMM(), YMM(), YMM()} |
| 62 |
for i := 0; i < 5; i++ { |
| 63 |
VPSLLQ(U8(1), p[(i+1)%5], t[i]) |
| 64 |
} |
| 65 |
for i := 0; i < 5; i++ { |
| 66 |
VPSRLQ(U8(63), p[(i+1)%5], d[i]) |
| 67 |
} |
| 68 |
for i := 0; i < 5; i++ { |
| 69 |
VPOR(t[i], d[i], d[i]) |
| 70 |
} |
| 71 |
for i := 0; i < 5; i++ { |
| 72 |
VPXOR(d[i], p[(i+4)%5], d[i]) |
| 73 |
} |
| 74 |
|
| 75 |
// Rotation to use |
| 76 |
rot := func(i, g int) int { |
| 77 |
table := [][]int{ |
| 78 |
{0, 24, 18, 6, 12}, |
| 79 |
{7, 23, 2, 9, 22}, |
| 80 |
{1, 3, 17, 16, 20}, |
| 81 |
{13, 8, 4, 5, 15}, |
| 82 |
{19, 10, 21, 14, 11}, |
| 83 |
} |
| 84 |
t := table[g][i] |
| 85 |
return ((t + 1) * t / 2) % 64 // t'th triangle number |
| 86 |
} |
| 87 |
|
| 88 |
// Index into d to use |
| 89 |
di := func(i, g int) int { |
| 90 |
return (3*g + i) % 5 |
| 91 |
} |
| 92 |
|
| 93 |
// Index into state to use |
| 94 |
si := func(i, g, r int) int { |
| 95 |
n := []int{6, 16, 11, 1}[r] |
| 96 |
m := []int{10, 20, 15, 5}[r] |
| 97 |
return (i*n + m*g) % 25 |
| 98 |
} |
| 99 |
|
| 100 |
for g := 0; g < 5; g++ { |
| 101 |
s := []Op{YMM(), YMM(), YMM(), YMM(), YMM()} |
| 102 |
|
| 103 |
// Load the right five words from the state and XOR d into them. |
| 104 |
for i := 0; i < 5; i++ { |
| 105 |
VPXOR(state(si(di(i, g), g, r)), d[di(i, g)], s[i]) |
| 106 |
} |
| 107 |
|
| 108 |
// Rotate each s[i] by the appropriate amount |
| 109 |
for i := 0; i < 5; i++ { |
| 110 |
if rot(i, g) != 0 { |
| 111 |
VPSLLQ(U8(rot(i, g)), s[i], t[i]) |
| 112 |
} |
| 113 |
} |
| 114 |
for i := 0; i < 5; i++ { |
| 115 |
if rot(i, g) != 0 { |
| 116 |
VPSRLQ(U8(64-rot(i, g)), s[i], s[i]) |
| 117 |
} |
| 118 |
} |
| 119 |
for i := 0; i < 5; i++ { |
| 120 |
if rot(i, g) != 0 { |
| 121 |
VPOR(t[i], s[i], s[i]) |
| 122 |
} |
| 123 |
} |
| 124 |
|
| 125 |
// Compute the new words s[i] ^ (s[i+2] & ~s[i+1]) |
| 126 |
for i := 0; i < 5; i++ { |
| 127 |
VPANDN(s[(i+2)%5], s[(i+1)%5], t[i]) |
| 128 |
} |
| 129 |
for i := 0; i < 5; i++ { |
| 130 |
VPXOR(s[i], t[i], t[i]) |
| 131 |
} |
| 132 |
|
| 133 |
// Round constant |
| 134 |
if g == 0 { |
| 135 |
// Note that we move rcPtr by 8*4 bytes after each superround. |
| 136 |
rc := YMM() |
| 137 |
VPBROADCASTQ(Mem{Base: rcPtr, Disp: r * 8}, rc) |
| 138 |
VPXOR(rc, t[0], t[0]) |
| 139 |
} |
| 140 |
|
| 141 |
// Store back into state |
| 142 |
for i := 0; i < 5; i++ { |
| 143 |
VMOVDQA(t[i], state(si(i, g, r))) |
| 144 |
} |
| 145 |
} |
| 146 |
} |
| 147 |
|
| 148 |
ADDQ(Imm(8*4), rcPtr) |
| 149 |
SUBQ(U32(1), superRound) |
| 150 |
JNZ(LabelRef("loop")) |
| 151 |
|
| 152 |
RET() |
| 153 |
|
| 154 |
Generate() |
| 155 |
} |