diff --git a/salsa20/benches/mod.rs b/salsa20/benches/mod.rs
index f07b023a..c33b40ba 100644
--- a/salsa20/benches/mod.rs
+++ b/salsa20/benches/mod.rs
@@ -1,9 +1,15 @@
 #![feature(test)]
+
+use cipher::{
+    Array,
+    consts::{U1, U4, U64},
+};
 extern crate test;
 
 cipher::stream_cipher_bench!(
     salsa20::Salsa8;
     salsa8_bench1_16b 16;
+    salsa8_bench1_64b 64;
     salsa8_bench2_256b 256;
     salsa8_bench3_1kib 1024;
     salsa8_bench4_16kib 16384;
@@ -12,6 +18,7 @@ cipher::stream_cipher_bench!(
 cipher::stream_cipher_bench!(
     salsa20::Salsa12;
     salsa12_bench1_16b 16;
+    salsa12_bench1_64b 64;
     salsa12_bench2_256b 256;
     salsa12_bench3_1kib 1024;
     salsa12_bench4_16kib 16384;
@@ -20,7 +27,50 @@ cipher::stream_cipher_bench!(
 cipher::stream_cipher_bench!(
     salsa20::Salsa20;
     salsa20_bench1_16b 16;
+    salsa20_bench1_64b 64;
     salsa20_bench2_256b 256;
     salsa20_bench3_1kib 1024;
     salsa20_bench4_16kib 16384;
 );
+
+#[bench]
+fn salsa8_bench1_ks_altn(b: &mut test::Bencher) {
+    use salsa20::SalsaChaining;
+    use std::hash::{BuildHasher, Hasher};
+
+    let seed = std::hash::RandomState::new().build_hasher().finish();
+
+    let mut buf: Array<[u32; 16], U1> = [[0u32; 16]].into();
+    buf[0][0] = seed as u32;
+    buf[0][1] = (seed >> 32) as u32;
+
+    b.iter(|| {
+        let mut cipher = salsa20::SalsaCore::<U4>::from_raw_state_cv(buf);
+        cipher.write_keystream_block_cv([&mut buf[0]].into());
+        test::black_box(&buf);
+    });
+
+    b.bytes = buf[0].len() as u64 * core::mem::size_of::<u32>() as u64;
+}
+
+#[bench]
+fn salsa8_bench1_ks(b: &mut test::Bencher) {
+    use cipher::StreamCipherCore;
+    use std::hash::{BuildHasher, Hasher};
+
+    let seed = std::hash::RandomState::new().build_hasher().finish();
+
+    let mut buf = [0u32; 16];
+    buf[0] = seed as u32;
+    buf[1] = (seed >> 32) as u32;
+
+    b.iter(|| {
+        let mut cipher = salsa20::SalsaCore::<U4>::from_raw_state(buf);
+        cipher.write_keystream_block(unsafe {
+            core::mem::transmute::<&mut [u32; 16], &mut Array<u8, U64>>(&mut buf)
+        });
+        test::black_box(&buf);
+    });
+
+    b.bytes = buf.len() as u64 * core::mem::size_of::<u32>() as u64;
+}
diff --git a/salsa20/src/backends.rs b/salsa20/src/backends.rs
index fbc9393c..4623d655 100644
--- a/salsa20/src/backends.rs
+++ b/salsa20/src/backends.rs
@@ -1 +1,30 @@
-pub(crate) mod soft;
+use cfg_if::cfg_if;
+
+cfg_if! {
+    if #[cfg(all(target_feature = "sse2", any(target_arch = "x86", target_arch = "x86_64")))] {
+        pub(crate) mod sse2;
+        pub(crate) type Backend<'a, R> = sse2::Backend<'a, R>;
+    } else {
+        pub(crate) mod soft;
+        pub(crate) type Backend<'a, R> = soft::Backend<'a, R>;
+    }
+}
+
+#[inline]
+#[allow(clippy::many_single_char_names)]
+pub(crate) fn quarter_round(
+    a: usize,
+    b: usize,
+    c: usize,
+    d: usize,
+    state: &mut [u32; crate::STATE_WORDS],
+) {
+    let a = crate::DATA_LAYOUT_INVERSE[a];
+    let b = crate::DATA_LAYOUT_INVERSE[b];
+    let c = crate::DATA_LAYOUT_INVERSE[c];
+    let d = crate::DATA_LAYOUT_INVERSE[d];
+    state[b] ^= state[a].wrapping_add(state[d]).rotate_left(7);
+    state[c] ^= state[b].wrapping_add(state[a]).rotate_left(9);
+    state[d] ^= state[c].wrapping_add(state[b]).rotate_left(13);
+    state[a] ^= state[d].wrapping_add(state[c]).rotate_left(18);
+}
diff --git a/salsa20/src/backends/soft.rs b/salsa20/src/backends/soft.rs
index caf2693f..7207020f 100644
--- a/salsa20/src/backends/soft.rs
+++ b/salsa20/src/backends/soft.rs
@@ -7,8 +7,16 @@ use cipher::{
     consts::{U1, U64},
 };
 
+use super::quarter_round;
+
 pub(crate) struct Backend<'a, R: Unsigned>(pub(crate) &'a mut SalsaCore<R>);
 
+impl<'a, R: Unsigned> From<&'a mut SalsaCore<R>> for Backend<'a, R> {
+    fn from(core: &'a mut SalsaCore<R>) -> Self {
+        Backend(core)
+    }
+}
+
 impl<R: Unsigned> BlockSizeUser for Backend<'_, R> {
     type BlockSize = U64;
 }
@@ -17,6 +25,17 @@ impl<R: Unsigned> ParBlocksSizeUser for Backend<'_, R> {
     type ParBlocksSize = U1;
 }
 
+impl<R: Unsigned> Backend<'_, R> {
+    #[inline(always)]
+    pub(crate) fn gen_ks_block_altn(&mut self, block: &mut [u32; STATE_WORDS]) {
+        let res = run_rounds::<R>(&self.0.state);
+
+        self.0.set_block_pos(self.0.get_block_pos() + 1);
+
+        block.copy_from_slice(&res);
+    }
+}
+
 impl<R: Unsigned> StreamCipherBackend for Backend<'_, R> {
     #[inline(always)]
     fn gen_ks_block(&mut self, block: &mut Block<Self>) {
@@ -24,27 +43,13 @@ impl<R: Unsigned> StreamCipherBackend for Backend<'_, R> {
 
         self.0.set_block_pos(self.0.get_block_pos() + 1);
 
-        for (chunk, val) in block.chunks_exact_mut(4).zip(res.iter()) {
-            chunk.copy_from_slice(&val.to_le_bytes());
+        for i in 0..16 {
+            block[i * 4..(i + 1) * 4]
+                .copy_from_slice(&res[crate::DATA_LAYOUT_INVERSE[i]].to_le_bytes());
         }
     }
 }
 
-#[inline]
-#[allow(clippy::many_single_char_names)]
-pub(crate) fn quarter_round(
-    a: usize,
-    b: usize,
-    c: usize,
-    d: usize,
-    state: &mut [u32; STATE_WORDS],
-) {
-    state[b] ^= state[a].wrapping_add(state[d]).rotate_left(7);
-    state[c] ^= state[b].wrapping_add(state[a]).rotate_left(9);
-    state[d] ^= state[c].wrapping_add(state[b]).rotate_left(13);
-    state[a] ^= state[d].wrapping_add(state[c]).rotate_left(18);
-}
-
 #[inline(always)]
 fn run_rounds<R: Unsigned>(state: &[u32; STATE_WORDS]) -> [u32; STATE_WORDS] {
     let mut res = *state;
diff --git a/salsa20/src/backends/sse2.rs b/salsa20/src/backends/sse2.rs
new file mode 100644
index 00000000..1276695a
--- /dev/null
+++ b/salsa20/src/backends/sse2.rs
@@ -0,0 +1,111 @@
+//! SSE2 backend for Salsa20.
+
+use crate::{Block, STATE_WORDS, SalsaCore, Unsigned};
+use cipher::{
+    BlockSizeUser, ParBlocksSizeUser, StreamCipherBackend, StreamCipherSeekCore,
+    consts::{U1, U64},
+};
+
+pub(crate) struct Backend<'a, R: Unsigned>(pub(crate) &'a mut SalsaCore<R>);
+
+impl<'a, R: Unsigned> From<&'a mut SalsaCore<R>> for Backend<'a, R> {
+    fn from(core: &'a mut SalsaCore<R>) -> Self {
+        Backend(core)
+    }
+}
+
+impl<R: Unsigned> BlockSizeUser for Backend<'_, R> {
+    type BlockSize = U64;
+}
+
+impl<R: Unsigned> ParBlocksSizeUser for Backend<'_, R> {
+    type ParBlocksSize = U1;
+}
+
+impl<R: Unsigned> Backend<'_, R> {
+    #[inline(always)]
+    pub(crate) fn gen_ks_block_altn(&mut self, block: &mut [u32; STATE_WORDS]) {
+        let res = run_rounds_sse2::<R>(&self.0.state);
+
+        self.0.set_block_pos(self.0.get_block_pos() + 1);
+
+        block.copy_from_slice(&res);
+    }
+}
+
+impl<R: Unsigned> StreamCipherBackend for Backend<'_, R> {
+    #[inline(always)]
+    fn gen_ks_block(&mut self, block: &mut Block<Self>) {
+        let res = run_rounds_sse2::<R>(&self.0.state);
+
+        self.0.set_block_pos(self.0.get_block_pos() + 1);
+
+        for i in 0..16 {
+            block[i * 4..(i + 1) * 4]
+                .copy_from_slice(&res[crate::DATA_LAYOUT_INVERSE[i]].to_le_bytes());
+        }
+    }
+}
+
+#[inline(always)]
+fn run_rounds_sse2<R: Unsigned>(state: &[u32; STATE_WORDS]) -> [u32; STATE_WORDS] {
+    use core::arch::x86_64::*;
+    unsafe {
+        let [a_save, b_save, d_save, c_save] = [
+            _mm_loadu_si128(state.as_ptr().add(0).cast()),
+            _mm_loadu_si128(state.as_ptr().add(4).cast()),
+            _mm_loadu_si128(state.as_ptr().add(8).cast()),
+            _mm_loadu_si128(state.as_ptr().add(12).cast()),
+        ];
+        let [mut a, mut b, mut c, mut d] = [a_save, b_save, c_save, d_save];
+
+        macro_rules! mm_rol_epi32x {
+            ($w:expr, $amt:literal) => {{
+                let w = $w;
+                _mm_xor_si128(_mm_slli_epi32(w, $amt), _mm_srli_epi32(w, 32 - $amt))
+            }};
+        }
+
+        macro_rules! quarter_xmmwords {
+            ($a:expr, $b:expr, $c:expr, $d:expr) => {
+                $b = _mm_xor_si128($b, mm_rol_epi32x!(_mm_add_epi32($a, $d), 7));
+                $c = _mm_xor_si128($c, mm_rol_epi32x!(_mm_add_epi32($b, $a), 9));
+                $d = _mm_xor_si128($d, mm_rol_epi32x!(_mm_add_epi32($c, $b), 13));
+                $a = _mm_xor_si128($a, mm_rol_epi32x!(_mm_add_epi32($d, $c), 18));
+            };
+        }
+
+        for _ in 0..R::USIZE {
+            quarter_xmmwords!(a, b, c, d);
+
+            // a stays in place
+            // b = left shuffle d by 1 element
+            d = _mm_shuffle_epi32(d, 0b00_11_10_01);
+            // c = left shuffle c by 2 elements
+            c = _mm_shuffle_epi32(c, 0b01_00_11_10);
+            // d = left shuffle b by 3 elements
+            b = _mm_shuffle_epi32(b, 0b10_01_00_11);
+
+            (b, d) = (d, b);
+
+            quarter_xmmwords!(a, b, c, d);
+
+            // a stays in place
+            // b = left shuffle d by 1 element
+            d = _mm_shuffle_epi32(d, 0b00_11_10_01);
+            // c = left shuffle c by 2 elements
+            c = _mm_shuffle_epi32(c, 0b01_00_11_10);
+            // d = left shuffle b by 3 elements
+            b = _mm_shuffle_epi32(b, 0b10_01_00_11);
+
+            (b, d) = (d, b);
+        }
+
+        let mut res = [0u32; STATE_WORDS];
+        _mm_storeu_si128(res.as_mut_ptr().add(0).cast(), _mm_add_epi32(a, a_save));
+        _mm_storeu_si128(res.as_mut_ptr().add(4).cast(), _mm_add_epi32(b, b_save));
+        _mm_storeu_si128(res.as_mut_ptr().add(8).cast(), _mm_add_epi32(d, d_save));
+        _mm_storeu_si128(res.as_mut_ptr().add(12).cast(), _mm_add_epi32(c, c_save));
+        res
+    }
+}
diff --git a/salsa20/src/lib.rs b/salsa20/src/lib.rs
index c75246f0..9adbfcd2 100644
--- a/salsa20/src/lib.rs
+++ b/salsa20/src/lib.rs
@@ -78,8 +78,8 @@ pub use cipher;
 use cipher::{
     Block, BlockSizeUser, IvSizeUser, KeyIvInit, KeySizeUser, StreamCipherClosure,
     StreamCipherCore, StreamCipherCoreWrapper, StreamCipherSeekCore,
-    array::{Array, typenum::Unsigned},
-    consts::{U4, U6, U8, U10, U24, U32, U64},
+    array::{Array, ArraySize, typenum::Unsigned},
+    consts::{U1, U4, U6, U8, U10, U24, U32, U64},
 };
 use core::marker::PhantomData;
 
@@ -118,7 +118,27 @@ const STATE_WORDS: usize = 16;
 /// State initialization constant ("expand 32-byte k")
 const CONSTANTS: [u32; 4] = [0x6170_7865, 0x3320_646e, 0x7962_2d32, 0x6b20_6574];
 
+const DATA_LAYOUT: [usize; 16] = [0, 5, 10, 15, 4, 9, 14, 3, 12, 1, 6, 11, 8, 13, 2, 7];
+
+const DATA_LAYOUT_INVERSE: [usize; 16] = {
+    let mut index = [0; 16];
+    let mut i = 0;
+    while i < 16 {
+        let mut inverse = 0;
+        while inverse < 16 {
+            if DATA_LAYOUT[inverse] == i {
+                index[i] = inverse;
+                break;
+            }
+            inverse += 1;
+        }
+        i += 1;
+    }
+    index
+};
+
 /// The Salsa20 core function.
+#[repr(transparent)]
 pub struct SalsaCore<R: Unsigned> {
     /// Internal state of the core function
     state: [u32; STATE_WORDS],
@@ -126,6 +146,56 @@ pub struct SalsaCore<R: Unsigned> {
     rounds: PhantomData<R>,
 }
 
+#[expect(unused)]
+const STATIC_ASSERT_CORE_IS_64_BYTES: [(); size_of::<SalsaCore<U10>>()] = [(); 64];
+
+/// Salsa20 chaining operations.
+pub trait SalsaChaining: BlockSizeUser<BlockSize = U64> {
+    /// Number of lanes
+    type LaneCount: ArraySize;
+
+    /// Permutation table for shuffling the natural order state into the internal order.
+    const ALTN_DATA_LAYOUT: [usize; STATE_WORDS];
+
+    /// Inverse permutation table.
+    const INVERSE_ALTN_DATA_LAYOUT: [usize; STATE_WORDS] = {
+        let mut index = [0; 16];
+        let mut i = 0;
+        while i < 16 {
+            let mut inverse = 0;
+            while inverse < 16 {
+                if Self::ALTN_DATA_LAYOUT[inverse] == i {
+                    index[i] = inverse;
+                    break;
+                }
+                inverse += 1;
+            }
+            i += 1;
+        }
+        index
+    };
+
+    /// Shuffle the state into the internal data layout.
+    fn shuffle_state_into_altn(state: &mut [u32; STATE_WORDS]) {
+        for i in 0..STATE_WORDS {
+            state[i] = state[Self::ALTN_DATA_LAYOUT[i]];
+        }
+    }
+
+    /// Shuffle the state from the internal data layout.
+    fn shuffle_state_from_altn(state: &mut [u32; STATE_WORDS]) {
+        for i in 0..STATE_WORDS {
+            state[i] = state[Self::INVERSE_ALTN_DATA_LAYOUT[i]];
+        }
+    }
+
+    /// Instantiate new Salsa core from raw state in internal order.
+    fn from_raw_state_cv(state: Array<[u32; STATE_WORDS], Self::LaneCount>) -> Self;
+
+    /// Generate keystream block in internal order.
+    fn write_keystream_block_cv(&mut self, block: Array<&mut [u32; STATE_WORDS], Self::LaneCount>);
+}
+
 impl<R: Unsigned> SalsaCore<R> {
     /// Create new Salsa core from raw state.
     ///
@@ -133,12 +203,40 @@ impl<R: Unsigned> SalsaCore<R> {
     /// Other users generally should not use this method.
     pub fn from_raw_state(state: [u32; STATE_WORDS]) -> Self {
         Self {
-            state,
+            state: core::array::from_fn(|i| state[DATA_LAYOUT[i]]),
             rounds: PhantomData,
         }
     }
 }
 
+impl<R: Unsigned> SalsaChaining for SalsaCore<R> {
+    type LaneCount = U1;
+
+    const ALTN_DATA_LAYOUT: [usize; STATE_WORDS] = DATA_LAYOUT;
+
+    /// Create new Salsa core from raw state with alternative data layout.
+    ///
+    /// This method is mainly intended for the `scrypt` crate.
+    /// Other users generally should not use this method.
+    fn from_raw_state_cv(state: Array<[u32; STATE_WORDS], Self::LaneCount>) -> Self {
+        Self {
+            state: state[0],
+            rounds: PhantomData,
+        }
+    }
+
+    /// Generate keystream block with alternative data layout.
+    ///
+    /// This method is used to generate keystream blocks with alternative data layout.
+    fn write_keystream_block_cv(
+        &mut self,
+        mut block: Array<&mut [u32; STATE_WORDS], Self::LaneCount>,
+    ) {
+        let mut backend = backends::Backend::<'_, R>::from(self);
+        backend.gen_ks_block_altn(block[0]);
+    }
+}
+
 impl<R: Unsigned> KeySizeUser for SalsaCore<R> {
     type KeySize = U32;
 }
@@ -177,7 +275,7 @@ impl<R: Unsigned> KeyIvInit for SalsaCore<R> {
         state[15] = CONSTANTS[3];
 
         Self {
-            state,
+            state: core::array::from_fn(|i| state[DATA_LAYOUT[i]]),
             rounds: PhantomData,
         }
     }
@@ -190,7 +288,7 @@ impl<R: Unsigned> StreamCipherCore for SalsaCore<R> {
         rem.try_into().ok()
     }
     fn process_with_backend(&mut self, f: impl StreamCipherClosure<BlockSize = Self::BlockSize>) {
-        f.call(&mut backends::soft::Backend(self));
+        f.call(&mut backends::Backend::<'_, R>::from(self));
     }
 }
 
@@ -199,13 +297,14 @@ impl<R: Unsigned> StreamCipherSeekCore for SalsaCore<R> {
 
     #[inline(always)]
     fn get_block_pos(&self) -> u64 {
-        (self.state[8] as u64) + ((self.state[9] as u64) << 32)
+        (self.state[DATA_LAYOUT_INVERSE[8]] as u64)
+            + ((self.state[DATA_LAYOUT_INVERSE[9]] as u64) << 32)
     }
 
     #[inline(always)]
     fn set_block_pos(&mut self, pos: u64) {
-        self.state[8] = (pos & 0xffff_ffff) as u32;
-        self.state[9] = ((pos >> 32) & 0xffff_ffff) as u32;
+        self.state[DATA_LAYOUT_INVERSE[8]] = (pos & 0xffff_ffff) as u32;
+        self.state[DATA_LAYOUT_INVERSE[9]] = ((pos >> 32) & 0xffff_ffff) as u32;
     }
 }
 
diff --git a/salsa20/src/xsalsa.rs b/salsa20/src/xsalsa.rs
index 9f84cef5..aa30349b 100644
--- a/salsa20/src/xsalsa.rs
+++ b/salsa20/src/xsalsa.rs
@@ -8,7 +8,7 @@ use cipher::{
     consts::{U4, U6, U10, U16, U24, U32, U64},
 };
 
-use crate::backends::soft::quarter_round;
+use crate::backends::quarter_round;
 
 #[cfg(feature = "zeroize")]
 use cipher::zeroize::ZeroizeOnDrop;
@@ -90,28 +90,32 @@ impl<R: Unsigned> ZeroizeOnDrop for XSalsaCore<R> {}
 ///
 /// It produces 256-bits of output suitable for use as a Salsa20 key
 pub fn hsalsa<R: Unsigned>(key: &Key, input: &Array<u8, U16>) -> Array<u8, U32> {
+    const KEY_IDX: [usize; 8] = [0, 5, 10, 15, 6, 7, 8, 9];
+
     #[inline(always)]
     fn to_u32(chunk: &[u8]) -> u32 {
         u32::from_le_bytes(chunk.try_into().unwrap())
     }
 
-    let mut state = [0u32; 16];
-    state[0] = CONSTANTS[0];
-    state[1..5]
+    let mut t = [0u32; 16];
+    t[0] = CONSTANTS[0];
+    t[1..5]
         .iter_mut()
         .zip(key[0..16].chunks_exact(4))
         .for_each(|(v, chunk)| *v = to_u32(chunk));
-    state[5] = CONSTANTS[1];
-    state[6..10]
+    t[5] = CONSTANTS[1];
+    t[6..10]
         .iter_mut()
         .zip(input.chunks_exact(4))
         .for_each(|(v, chunk)| *v = to_u32(chunk));
-    state[10] = CONSTANTS[2];
-    state[11..15]
+    t[10] = CONSTANTS[2];
+    t[11..15]
         .iter_mut()
         .zip(key[16..].chunks_exact(4))
         .for_each(|(v, chunk)| *v = to_u32(chunk));
-    state[15] = CONSTANTS[3];
+    t[15] = CONSTANTS[3];
+
+    let mut state = core::array::from_fn(|i| t[crate::DATA_LAYOUT[i]]);
 
     // 20 rounds consisting of 10 column rounds and 10 diagonal rounds
     for _ in 0..R::USIZE {
@@ -129,10 +133,10 @@ pub fn hsalsa<R: Unsigned>(key: &Key, input: &Array<u8, U16>) -> Array<u8, U32>
     }
 
     let mut output = Array::default();
-    let key_idx: [usize; 8] = [0, 5, 10, 15, 6, 7, 8, 9];
 
-    for (i, chunk) in output.chunks_exact_mut(4).enumerate() {
-        chunk.copy_from_slice(&state[key_idx[i]].to_le_bytes());
+    for i in 0..8 {
+        output[i * 4..(i + 1) * 4]
+            .copy_from_slice(&state[crate::DATA_LAYOUT_INVERSE[KEY_IDX[i]]].to_le_bytes());
     }
 
     output