saymrwulf/betrusted-curve25519-dalek-source
1
0
Fork 0
mirror of https://github.com/saymrwulf/betrusted-curve25519-dalek-source.git synced 2026-09-06 20:41:11 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add hardware acceleration to differential point add and double

Browse source 
hacky implementation, but it'll let us do a performance benchmark
This commit is contained in:
bunnie 2020-08-20 15:47:02 +08:00
parent a2835635e7
commit 9c3d34345c
4 changed files with 294 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
Cargo.toml
View file

@ -49,6 +49,8 @@ packed_simd = { version = "0.3", features = ["into_bits"], optional = true }
zeroize = { version = "1", default-features = false }
engine25519-as = { path="../engine25519-as", default-features = false, features = [] }
rand = { version = "0.7" }
volatile = "0.2.6"
pac = { path = "../../../../sim_support/rust/pac" }
[features]
nightly = ["subtle/nightly"]

133
src/field.rs
View file

@ -910,10 +910,143 @@ mod test {
}
}
fn test_diff_add_and_double(mut file: &mut File) {
use montgomery::ProjectivePoint;
// test cswap. three input registers: (r0, r1) to swap, (r2) to control swap, one output register (r31).
let num_src_regs = 5;
let reg_window = 0;
let num_tests = 8;
let loading_address = 0; // microcode loading address
let mcode = assemble_engine25519!(
start:
// test preamble
psa %20, %0
psa %21, %1
psa %22, %2
psa %23, %3
psa %24, %4
// P.U in %20
// P.W in %21
// Q.U in %22
// Q.W in %23
// affine_PmQ in %24
// %30 is the TRD scratch register
// %29 is the subtraction temporary value register
// let t0 = &P.U + &P.W;
add %0, %20, %21
trd %30, %0
sub %0, %0, %30
// let t1 = &P.U - &P.W;
sub %21, #3, %21 // negate &P.W using #FIELDPRIME (#3)
add %1, %20, %21
trd %30, %1
sub %1, %1, %30
// let t2 = &Q.U + &Q.W;
add %2, %22, %23
trd %30, %2
sub %2, %2, %30
// let t3 = &Q.U - &Q.W;
sub %23, #3, %23
add %3, %22, %23
trd %30, %3
sub %3, %3, %30
// let t4 = t0.square(); // (U_P + W_P)^2 = U_P^2 + 2 U_P W_P + W_P^2
mul %4, %0, %0
// let t5 = t1.square(); // (U_P - W_P)^2 = U_P^2 - 2 U_P W_P + W_P^2
mul %5, %1, %1
// let t6 = &t4 - &t5; // 4 U_P W_P
sub %29, #3, %5
add %6, %4, %29
trd %30, %6
sub %6, %6, %30
// let t7 = &t0 * &t3; // (U_P + W_P) (U_Q - W_Q) = U_P U_Q + W_P U_Q - U_P W_Q - W_P W_Q
mul %7, %0, %3
// let t8 = &t1 * &t2; // (U_P - W_P) (U_Q + W_Q) = U_P U_Q - W_P U_Q + U_P W_Q - W_P W_Q
mul %8, %1, %2
// let t9 = &t7 + &t8; // 2 (U_P U_Q - W_P W_Q)
add %9, %7, %8
trd %30, %9
sub %9, %9, %30
// let t10 = &t7 - &t8; // 2 (W_P U_Q - U_P W_Q)
sub %29, #3, %8
add %10, %7, %29
trd %30, %10
sub %10, %10, %30
// let t11 = t9.square(); // 4 (U_P U_Q - W_P W_Q)^2
mul %11, %9, %9
// let t12 = t10.square(); // 4 (W_P U_Q - U_P W_Q)^2
mul %12, %10, %10
// let t13 = &APLUS2_OVER_FOUR * &t6; // (A + 2) U_P U_Q
mul %13, #4, %6 // #4 is A+2/4
// let t14 = &t4 * &t5; // ((U_P + W_P)(U_P - W_P))^2 = (U_P^2 - W_P^2)^2
mul %14, %4, %5
// let t15 = &t13 + &t5; // (U_P - W_P)^2 + (A + 2) U_P W_P
add %15, %13, %5
trd %30, %15
sub %15, %15, %30
// let t16 = &t6 * &t15; // 4 (U_P W_P) ((U_P - W_P)^2 + (A + 2) U_P W_P)
mul %16, %6, %15
// let t17 = affine_PmQ * &t12; // U_D * 4 (W_P U_Q - U_P W_Q)^2
mul %17, %24, %12 // affine_PmQ loaded into %24
///// these can be eliminated down the road, but included for 1:1 algorithm correspodence to reference in early testing
// let t18 = t11; // W_D * 4 (U_P U_Q - W_P W_Q)^2
psa %18, %11
// P.U = t14; // U_{P'} = (U_P + W_P)^2 (U_P - W_P)^2
psa %20, %14
// P.W = t16; // W_{P'} = (4 U_P W_P) ((U_P - W_P)^2 + ((A + 2)/4) 4 U_P W_P)
psa %21, %16
// Q.U = t18; // U_{Q'} = W_D * 4 (U_P U_Q - W_P W_Q)^2
psa %22, %18
// Q.W = t17; // W_{Q'} = U_D * 4 (W_P U_Q - U_P W_Q)^2
psa %23, %17
// test postamble -- sum together the points to create a single composite test output
add %31, %20, %21
trd %30, %31
sub %31, %31, %30
add %31, %31, %22
trd %30, %31
sub %31, %31, %30
add %31, %31, %23
trd %30, %31
sub %31, %31, %30 // leave result in r31
fin // finish execution
); write_test_header(&mut file, loading_address, &mcode, num_src_regs, reg_window, num_tests);
use montgomery::differential_add_and_double;
// test vectors
for _ in 0..8 {
let pu = FieldElement::from_bytes(&rand::thread_rng().gen::<[u8; 32]>());
let pw = FieldElement::from_bytes(&rand::thread_rng().gen::<[u8; 32]>());
let qu = FieldElement::from_bytes(&rand::thread_rng().gen::<[u8; 32]>());
let qw = FieldElement::from_bytes(&rand::thread_rng().gen::<[u8; 32]>());
let pmq = FieldElement::from_bytes(&rand::thread_rng().gen::<[u8; 32]>());
write_helper(&mut file, pu);
write_helper(&mut file, pw);
write_helper(&mut file, qu);
write_helper(&mut file, qw);
write_helper(&mut file, pmq);
let mut P: ProjectivePoint = ProjectivePoint{U:pu, W:pw};
let mut Q: ProjectivePoint = ProjectivePoint{U:qu, W:qw};
differential_add_and_double(&mut P, &mut Q, &pmq);
write_helper(&mut file, &(&P.U + &P.W) + &(&Q.U + &Q.W));
}
}
test_add(&mut file);
test_loop(&mut file);
test_cswap(&mut file);
test_mul(&mut file);
test_diff_add_and_double(&mut file);
// end sequence
let _ = file.write(&(0xFFFF_FFFF as u32).to_le_bytes());

2
src/lib.rs
View file

@ -62,6 +62,8 @@ pub(crate) mod macros;
#[macro_use]
extern crate engine25519_as;
extern crate rand;
extern crate volatile;
extern crate pac;
//------------------------------------------------------------------------
// curve25519-dalek public modules

159
src/montgomery.rs
View file

@ -160,7 +160,7 @@ impl MontgomeryPoint {
/// \\( \mathbb P(\mathbb F\_p) \\), which we identify with the Kummer
/// line of the Montgomery curve.
#[derive(Copy, Clone, Debug)]
struct ProjectivePoint {
pub struct ProjectivePoint {
pub U: FieldElement,
pub W: FieldElement,
}
@ -220,7 +220,7 @@ impl ProjectivePoint {
/// $$
/// (U\_Q : W\_Q) \gets u(P + Q).
/// $$
fn differential_add_and_double(
pub fn differential_add_and_double(
P: &mut ProjectivePoint,
Q: &mut ProjectivePoint,
affine_PmQ: &FieldElement,
@ -260,6 +260,161 @@ fn differential_add_and_double(
Q.W = t17; // W_{Q'} = U_D * 4 (W_P U_Q - U_P W_Q)^2
}
pub fn differential_add_and_double_hw(
P: &mut ProjectivePoint,
Q: &mut ProjectivePoint,
affine_PmQ: &FieldElement,
) {
use volatile::Volatile;
let p = unsafe { pac::Peripherals::steal() };
let mcode = assemble_engine25519!(
start:
// P.U in %20
// P.W in %21
// Q.U in %22
// Q.W in %23
// affine_PmQ in %24
// %30 is the TRD scratch register
// %29 is the subtraction temporary value register
// let t0 = &P.U + &P.W;
add %0, %20, %21
trd %30, %0
sub %0, %0, %30
// let t1 = &P.U - &P.W;
sub %21, #3, %21 // negate &P.W using #FIELDPRIME (#3)
add %1, %20, %21
trd %30, %1
sub %1, %1, %30
// let t2 = &Q.U + &Q.W;
add %2, %22, %23
trd %30, %2
sub %2, %2, %30
// let t3 = &Q.U - &Q.W;
sub %23, #3, %23
add %3, %22, %23
trd %30, %3
sub %3, %3, %30
// let t4 = t0.square(); // (U_P + W_P)^2 = U_P^2 + 2 U_P W_P + W_P^2
mul %4, %0, %0
// let t5 = t1.square(); // (U_P - W_P)^2 = U_P^2 - 2 U_P W_P + W_P^2
mul %5, %1, %1
// let t6 = &t4 - &t5; // 4 U_P W_P
sub %29, #3, %5
add %6, %4, %29
trd %30, %6
sub %6, %6, %30
// let t7 = &t0 * &t3; // (U_P + W_P) (U_Q - W_Q) = U_P U_Q + W_P U_Q - U_P W_Q - W_P W_Q
mul %7, %0, %3
// let t8 = &t1 * &t2; // (U_P - W_P) (U_Q + W_Q) = U_P U_Q - W_P U_Q + U_P W_Q - W_P W_Q
mul %8, %1, %2
// let t9 = &t7 + &t8; // 2 (U_P U_Q - W_P W_Q)
add %9, %7, %8
trd %30, %9
sub %9, %9, %30
// let t10 = &t7 - &t8; // 2 (W_P U_Q - U_P W_Q)
sub %29, #3, %8
add %10, %7, %29
trd %30, %10
sub %10, %10, %30
// let t11 = t9.square(); // 4 (U_P U_Q - W_P W_Q)^2
mul %11, %9, %9
// let t12 = t10.square(); // 4 (W_P U_Q - U_P W_Q)^2
mul %12, %10, %10
// let t13 = &APLUS2_OVER_FOUR * &t6; // (A + 2) U_P U_Q
mul %13, #4, %6 // #4 is A+2/4
// let t14 = &t4 * &t5; // ((U_P + W_P)(U_P - W_P))^2 = (U_P^2 - W_P^2)^2
mul %14, %4, %5
// let t15 = &t13 + &t5; // (U_P - W_P)^2 + (A + 2) U_P W_P
add %15, %13, %5
trd %30, %15
sub %15, %15, %30
// let t16 = &t6 * &t15; // 4 (U_P W_P) ((U_P - W_P)^2 + (A + 2) U_P W_P)
mul %16, %6, %15
// let t17 = affine_PmQ * &t12; // U_D * 4 (W_P U_Q - U_P W_Q)^2
mul %17, %24, %12 // affine_PmQ loaded into %24
///// these can be eliminated down the road, but included for 1:1 algorithm correspodence to reference in early testing
// let t18 = t11; // W_D * 4 (U_P U_Q - W_P W_Q)^2
psa %18, %11
// P.U = t14; // U_{P'} = (U_P + W_P)^2 (U_P - W_P)^2
psa %20, %14
// P.W = t16; // W_{P'} = (4 U_P W_P) ((U_P - W_P)^2 + ((A + 2)/4) 4 U_P W_P)
psa %21, %16
// Q.U = t18; // U_{Q'} = W_D * 4 (U_P U_Q - W_P W_Q)^2
psa %22, %18
// Q.W = t17; // W_{Q'} = U_D * 4 (W_P U_Q - U_P W_Q)^2
psa %23, %17
fin // finish execution
);
let microcode_ptr: *mut u32 = 0xe002_0000 as *mut u32;
let microcode = microcode_ptr as *mut Volatile<u32>;
// copy the microcode in -- later on we can optimize this so it's only done once?
for i in 0..mcode.len() as usize {
unsafe { (*(microcode.add(i))).write( mcode[i] ); }
}
// setup the engine microcode parameters
unsafe{
p.ENGINE.window.write(|w| w.bits(0));
p.ENGINE.mpstart.write(|w| w.bits(0));
p.ENGINE.mplen.write(|w| w.bits(mcode.len() as u32));
}
fn copy_to_rf(bytes: [u8; 32], register: usize) {
let rf_ptr: *mut u32 = 0xe003_0000 as *mut u32;
let rf = rf_ptr as *mut Volatile<u32>;
for word in 0..8 {
let mut temp: [u8; 4] = [0; 4];
for i in 0..4 {
temp[i] = bytes[word*4 + i];
}
unsafe { (*( rf.add( (register * 8 + word) as usize )) ).write( u32::from_le_bytes(temp) ); }
}
}
// P.U in %20
// P.W in %21
// Q.U in %22
// Q.W in %23
// affine_PmQ in %24
copy_to_rf(P.U.to_bytes(), 20);
copy_to_rf(P.W.to_bytes(), 21);
copy_to_rf(Q.U.to_bytes(), 22);
copy_to_rf(Q.W.to_bytes(), 23);
copy_to_rf(affine_PmQ.to_bytes(), 24);
// start the run
p.ENGINE.control.write(|w| w.go().set_bit());
loop {
let status = p.ENGINE.status.read().bits();
if (status & 1) == 0 {
break;
}
}
fn copy_from_rf(register: usize) -> [u8; 32] {
let rf_ptr: *mut u32 = 0xe003_0000 as *mut u32;
let rf = rf_ptr as *mut Volatile<u32>;
let mut bytes: [u8; 32] = [0; 32];
for word in 0..8 {
unsafe{
let value: u32 = (*( rf.add( (register * 8 + word) as usize ))).read();
let b = value.to_le_bytes();
for i in 0..4 {
bytes[word*4 + i] = b[i];
}
}
}
bytes
}
P.U = FieldElement::from_bytes(&copy_from_rf(20));
P.W = FieldElement::from_bytes(&copy_from_rf(21));
Q.U = FieldElement::from_bytes(&copy_from_rf(22));
Q.W = FieldElement::from_bytes(&copy_from_rf(23));
}
define_mul_assign_variants!(LHS = MontgomeryPoint, RHS = Scalar);
define_mul_variants!(LHS = MontgomeryPoint, RHS = Scalar, Output = MontgomeryPoint);