Merge pull request #4003 from Yawning/feature/crypto-improvements

core:crypto: Various improvements

27 files changed, 2253 insertions, 905 deletions

diff --git a/core/crypto/_aes/ct64/ghash.odin b/core/crypto/_aes/ct64/ghash.odin
index 21ac2ca97..a522a481a 100644
--- a/core/crypto/_aes/ct64/ghash.odin
+++ b/core/crypto/_aes/ct64/ghash.odin
@@ -80,8 +80,8 @@ ghash :: proc "contextless" (dst, key, data: []byte) {
 	h2 := h0 ~ h1
 	h2r := h0r ~ h1r
 
-	src: []byte
 	for l > 0 {
+		src: []byte = ---
 		if l >= _aes.GHASH_BLOCK_SIZE {
 			src = buf
 			buf = buf[_aes.GHASH_BLOCK_SIZE:]
diff --git a/core/crypto/_aes/hw_intel/api.odin b/core/crypto/_aes/hw_intel/api.odin
index 5cb5a68bb..1796bb093 100644
--- a/core/crypto/_aes/hw_intel/api.odin
+++ b/core/crypto/_aes/hw_intel/api.odin
@@ -3,7 +3,7 @@ package aes_hw_intel
 
 import "core:sys/info"
 
-// is_supporte returns true iff hardware accelerated AES
+// is_supported returns true iff hardware accelerated AES
 // is supported.
 is_supported :: proc "contextless" () -> bool {
 	features, ok := info.cpu_features.?
diff --git a/core/crypto/_aes/hw_intel/ghash.odin b/core/crypto/_aes/hw_intel/ghash.odin
index 9a5208523..d61e71b3a 100644
--- a/core/crypto/_aes/hw_intel/ghash.odin
+++ b/core/crypto/_aes/hw_intel/ghash.odin
@@ -25,7 +25,6 @@ package aes_hw_intel
 
 import "base:intrinsics"
 import "core:crypto/_aes"
-import "core:simd"
 import "core:simd/x86"
 
 @(private = "file")
@@ -58,14 +57,11 @@ GHASH_STRIDE_BYTES_HW :: GHASH_STRIDE_HW * _aes.GHASH_BLOCK_SIZE
 // chunks. We number chunks from 0 to 3 in left to right order.
 
 @(private = "file")
-byteswap_index := transmute(x86.__m128i)simd.i8x16{
-	// Note: simd.i8x16 is reverse order from x86._mm_set_epi8.
-	15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0,
-}
+_BYTESWAP_INDEX: x86.__m128i : { 0x08090a0b0c0d0e0f, 0x0001020304050607 }
 
 @(private = "file", require_results, enable_target_feature = "sse2,ssse3")
 byteswap :: #force_inline proc "contextless" (x: x86.__m128i) -> x86.__m128i {
-	return x86._mm_shuffle_epi8(x, byteswap_index)
+	return x86._mm_shuffle_epi8(x, _BYTESWAP_INDEX)
 }
 
 // From a 128-bit value kw, compute kx as the XOR of the two 64-bit
@@ -244,8 +240,8 @@ ghash :: proc "contextless" (dst, key, data: []byte) #no_bounds_check {
 	}
 
 	// Process 1 block at a time
-	src: []byte
 	for l > 0 {
+		src: []byte = ---
 		if l >= _aes.GHASH_BLOCK_SIZE {
 			src = buf
 			buf = buf[_aes.GHASH_BLOCK_SIZE:]
diff --git a/core/crypto/_chacha20/chacha20.odin b/core/crypto/_chacha20/chacha20.odin
new file mode 100644
index 000000000..a907209de
--- /dev/null
+++ b/core/crypto/_chacha20/chacha20.odin
@@ -0,0 +1,123 @@
+package _chacha20
+
+import "base:intrinsics"
+import "core:encoding/endian"
+import "core:math/bits"
+import "core:mem"
+
+// KEY_SIZE is the (X)ChaCha20 key size in bytes.
+KEY_SIZE :: 32
+// IV_SIZE is the ChaCha20 IV size in bytes.
+IV_SIZE :: 12
+// XIV_SIZE is the XChaCha20 IV size in bytes.
+XIV_SIZE :: 24
+
+// MAX_CTR_IETF is the maximum counter value for the IETF flavor ChaCha20.
+MAX_CTR_IETF :: 0xffffffff
+// BLOCK_SIZE is the (X)ChaCha20 block size in bytes.
+BLOCK_SIZE :: 64
+// STATE_SIZE_U32 is the (X)ChaCha20 state size in u32s.
+STATE_SIZE_U32 :: 16
+// Rounds is the (X)ChaCha20 round count.
+ROUNDS :: 20
+
+// SIGMA_0 is sigma[0:4].
+SIGMA_0: u32 : 0x61707865
+// SIGMA_1 is sigma[4:8].
+SIGMA_1: u32 : 0x3320646e
+// SIGMA_2 is sigma[8:12].
+SIGMA_2: u32 : 0x79622d32
+// SIGMA_3 is sigma[12:16].
+SIGMA_3: u32 : 0x6b206574
+
+// Context is a ChaCha20 or XChaCha20 instance.
+Context :: struct {
+	_s:              [STATE_SIZE_U32]u32,
+	_buffer:         [BLOCK_SIZE]byte,
+	_off:            int,
+	_is_ietf_flavor: bool,
+	_is_initialized: bool,
+}
+
+// init inititializes a Context for ChaCha20 with the provided key and
+// iv.
+//
+// WARNING: This ONLY handles ChaCha20.  XChaCha20 sub-key and IV
+// derivation is expected to be handled by the caller, so that the
+// HChaCha call can be suitably accelerated.
+init :: proc "contextless" (ctx: ^Context, key, iv: []byte, is_xchacha: bool) {
+	if len(key) != KEY_SIZE || len(iv) != IV_SIZE {
+		intrinsics.trap()
+	}
+
+	k, n := key, iv
+
+	ctx._s[0] = SIGMA_0
+	ctx._s[1] = SIGMA_1
+	ctx._s[2] = SIGMA_2
+	ctx._s[3] = SIGMA_3
+	ctx._s[4] = endian.unchecked_get_u32le(k[0:4])
+	ctx._s[5] = endian.unchecked_get_u32le(k[4:8])
+	ctx._s[6] = endian.unchecked_get_u32le(k[8:12])
+	ctx._s[7] = endian.unchecked_get_u32le(k[12:16])
+	ctx._s[8] = endian.unchecked_get_u32le(k[16:20])
+	ctx._s[9] = endian.unchecked_get_u32le(k[20:24])
+	ctx._s[10] = endian.unchecked_get_u32le(k[24:28])
+	ctx._s[11] = endian.unchecked_get_u32le(k[28:32])
+	ctx._s[12] = 0
+	ctx._s[13] = endian.unchecked_get_u32le(n[0:4])
+	ctx._s[14] = endian.unchecked_get_u32le(n[4:8])
+	ctx._s[15] = endian.unchecked_get_u32le(n[8:12])
+
+	ctx._off = BLOCK_SIZE
+	ctx._is_ietf_flavor = !is_xchacha
+	ctx._is_initialized = true
+}
+
+// seek seeks the (X)ChaCha20 stream counter to the specified block.
+seek :: proc(ctx: ^Context, block_nr: u64) {
+	assert(ctx._is_initialized)
+
+	if ctx._is_ietf_flavor {
+		if block_nr > MAX_CTR_IETF {
+			panic("crypto/chacha20: attempted to seek past maximum counter")
+		}
+	} else {
+		ctx._s[13] = u32(block_nr >> 32)
+	}
+	ctx._s[12] = u32(block_nr)
+	ctx._off = BLOCK_SIZE
+}
+
+// reset sanitizes the Context.  The Context must be re-initialized to
+// be used again.
+reset :: proc(ctx: ^Context) {
+	mem.zero_explicit(&ctx._s, size_of(ctx._s))
+	mem.zero_explicit(&ctx._buffer, size_of(ctx._buffer))
+
+	ctx._is_initialized = false
+}
+
+check_counter_limit :: proc(ctx: ^Context, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	//
+	// While all modern "standard" definitions of ChaCha20 use
+	// the IETF 32-bit counter, for XChaCha20 most common
+	// implementations allow for a 64-bit counter.
+	//
+	// Honestly, the answer here is "use a MRAE primitive", but
+	// go with "common" practice in the case of XChaCha20.
+
+	ERR_CTR_EXHAUSTED :: "crypto/chacha20: maximum (X)ChaCha20 keystream per IV reached"
+
+	if ctx._is_ietf_flavor {
+		if u64(ctx._s[12]) + u64(nr_blocks) > MAX_CTR_IETF {
+			panic(ERR_CTR_EXHAUSTED)
+		}
+	} else {
+		ctr := (u64(ctx._s[13]) << 32) | u64(ctx._s[12])
+		if _, carry := bits.add_u64(ctr, u64(nr_blocks), 0); carry != 0 {
+			panic(ERR_CTR_EXHAUSTED)
+		}
+	}
+}
diff --git a/core/crypto/_chacha20/ref/chacha20_ref.odin b/core/crypto/_chacha20/ref/chacha20_ref.odin
new file mode 100644
index 000000000..c111c1c76
--- /dev/null
+++ b/core/crypto/_chacha20/ref/chacha20_ref.odin
@@ -0,0 +1,360 @@
+package chacha20_ref
+
+import "core:crypto/_chacha20"
+import "core:encoding/endian"
+import "core:math/bits"
+
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst, src := dst, src
+	x := &ctx._s
+	for n := 0; n < nr_blocks; n = n + 1 {
+		x0, x1, x2, x3 :=
+			_chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3
+		x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15 :=
+			x[4], x[5], x[6], x[7], x[8], x[9], x[10], x[11], x[12], x[13], x[14], x[15]
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			// Even when forcing inlining manually inlining all of
+			// these is decently faster.
+
+			// quarterround(x, 0, 4, 8, 12)
+			x0 += x4
+			x12 ~= x0
+			x12 = bits.rotate_left32(x12, 16)
+			x8 += x12
+			x4 ~= x8
+			x4 = bits.rotate_left32(x4, 12)
+			x0 += x4
+			x12 ~= x0
+			x12 = bits.rotate_left32(x12, 8)
+			x8 += x12
+			x4 ~= x8
+			x4 = bits.rotate_left32(x4, 7)
+
+			// quarterround(x, 1, 5, 9, 13)
+			x1 += x5
+			x13 ~= x1
+			x13 = bits.rotate_left32(x13, 16)
+			x9 += x13
+			x5 ~= x9
+			x5 = bits.rotate_left32(x5, 12)
+			x1 += x5
+			x13 ~= x1
+			x13 = bits.rotate_left32(x13, 8)
+			x9 += x13
+			x5 ~= x9
+			x5 = bits.rotate_left32(x5, 7)
+
+			// quarterround(x, 2, 6, 10, 14)
+			x2 += x6
+			x14 ~= x2
+			x14 = bits.rotate_left32(x14, 16)
+			x10 += x14
+			x6 ~= x10
+			x6 = bits.rotate_left32(x6, 12)
+			x2 += x6
+			x14 ~= x2
+			x14 = bits.rotate_left32(x14, 8)
+			x10 += x14
+			x6 ~= x10
+			x6 = bits.rotate_left32(x6, 7)
+
+			// quarterround(x, 3, 7, 11, 15)
+			x3 += x7
+			x15 ~= x3
+			x15 = bits.rotate_left32(x15, 16)
+			x11 += x15
+			x7 ~= x11
+			x7 = bits.rotate_left32(x7, 12)
+			x3 += x7
+			x15 ~= x3
+			x15 = bits.rotate_left32(x15, 8)
+			x11 += x15
+			x7 ~= x11
+			x7 = bits.rotate_left32(x7, 7)
+
+			// quarterround(x, 0, 5, 10, 15)
+			x0 += x5
+			x15 ~= x0
+			x15 = bits.rotate_left32(x15, 16)
+			x10 += x15
+			x5 ~= x10
+			x5 = bits.rotate_left32(x5, 12)
+			x0 += x5
+			x15 ~= x0
+			x15 = bits.rotate_left32(x15, 8)
+			x10 += x15
+			x5 ~= x10
+			x5 = bits.rotate_left32(x5, 7)
+
+			// quarterround(x, 1, 6, 11, 12)
+			x1 += x6
+			x12 ~= x1
+			x12 = bits.rotate_left32(x12, 16)
+			x11 += x12
+			x6 ~= x11
+			x6 = bits.rotate_left32(x6, 12)
+			x1 += x6
+			x12 ~= x1
+			x12 = bits.rotate_left32(x12, 8)
+			x11 += x12
+			x6 ~= x11
+			x6 = bits.rotate_left32(x6, 7)
+
+			// quarterround(x, 2, 7, 8, 13)
+			x2 += x7
+			x13 ~= x2
+			x13 = bits.rotate_left32(x13, 16)
+			x8 += x13
+			x7 ~= x8
+			x7 = bits.rotate_left32(x7, 12)
+			x2 += x7
+			x13 ~= x2
+			x13 = bits.rotate_left32(x13, 8)
+			x8 += x13
+			x7 ~= x8
+			x7 = bits.rotate_left32(x7, 7)
+
+			// quarterround(x, 3, 4, 9, 14)
+			x3 += x4
+			x14 ~= x3
+			x14 = bits.rotate_left32(x14, 16)
+			x9 += x14
+			x4 ~= x9
+			x4 = bits.rotate_left32(x4, 12)
+			x3 += x4
+			x14 ~= x3
+			x14 = bits.rotate_left32(x14, 8)
+			x9 += x14
+			x4 ~= x9
+			x4 = bits.rotate_left32(x4, 7)
+		}
+
+		x0 += _chacha20.SIGMA_0
+		x1 += _chacha20.SIGMA_1
+		x2 += _chacha20.SIGMA_2
+		x3 += _chacha20.SIGMA_3
+		x4 += x[4]
+		x5 += x[5]
+		x6 += x[6]
+		x7 += x[7]
+		x8 += x[8]
+		x9 += x[9]
+		x10 += x[10]
+		x11 += x[11]
+		x12 += x[12]
+		x13 += x[13]
+		x14 += x[14]
+		x15 += x[15]
+
+		// - The caller(s) ensure that src/dst are valid.
+		// - The compiler knows if the target is picky about alignment.
+
+		#no_bounds_check {
+			if src != nil {
+				endian.unchecked_put_u32le(dst[0:4], endian.unchecked_get_u32le(src[0:4]) ~ x0)
+				endian.unchecked_put_u32le(dst[4:8], endian.unchecked_get_u32le(src[4:8]) ~ x1)
+				endian.unchecked_put_u32le(dst[8:12], endian.unchecked_get_u32le(src[8:12]) ~ x2)
+				endian.unchecked_put_u32le(dst[12:16], endian.unchecked_get_u32le(src[12:16]) ~ x3)
+				endian.unchecked_put_u32le(dst[16:20], endian.unchecked_get_u32le(src[16:20]) ~ x4)
+				endian.unchecked_put_u32le(dst[20:24], endian.unchecked_get_u32le(src[20:24]) ~ x5)
+				endian.unchecked_put_u32le(dst[24:28], endian.unchecked_get_u32le(src[24:28]) ~ x6)
+				endian.unchecked_put_u32le(dst[28:32], endian.unchecked_get_u32le(src[28:32]) ~ x7)
+				endian.unchecked_put_u32le(dst[32:36], endian.unchecked_get_u32le(src[32:36]) ~ x8)
+				endian.unchecked_put_u32le(dst[36:40], endian.unchecked_get_u32le(src[36:40]) ~ x9)
+				endian.unchecked_put_u32le(
+					dst[40:44],
+					endian.unchecked_get_u32le(src[40:44]) ~ x10,
+				)
+				endian.unchecked_put_u32le(
+					dst[44:48],
+					endian.unchecked_get_u32le(src[44:48]) ~ x11,
+				)
+				endian.unchecked_put_u32le(
+					dst[48:52],
+					endian.unchecked_get_u32le(src[48:52]) ~ x12,
+				)
+				endian.unchecked_put_u32le(
+					dst[52:56],
+					endian.unchecked_get_u32le(src[52:56]) ~ x13,
+				)
+				endian.unchecked_put_u32le(
+					dst[56:60],
+					endian.unchecked_get_u32le(src[56:60]) ~ x14,
+				)
+				endian.unchecked_put_u32le(
+					dst[60:64],
+					endian.unchecked_get_u32le(src[60:64]) ~ x15,
+				)
+				src = src[_chacha20.BLOCK_SIZE:]
+			} else {
+				endian.unchecked_put_u32le(dst[0:4], x0)
+				endian.unchecked_put_u32le(dst[4:8], x1)
+				endian.unchecked_put_u32le(dst[8:12], x2)
+				endian.unchecked_put_u32le(dst[12:16], x3)
+				endian.unchecked_put_u32le(dst[16:20], x4)
+				endian.unchecked_put_u32le(dst[20:24], x5)
+				endian.unchecked_put_u32le(dst[24:28], x6)
+				endian.unchecked_put_u32le(dst[28:32], x7)
+				endian.unchecked_put_u32le(dst[32:36], x8)
+				endian.unchecked_put_u32le(dst[36:40], x9)
+				endian.unchecked_put_u32le(dst[40:44], x10)
+				endian.unchecked_put_u32le(dst[44:48], x11)
+				endian.unchecked_put_u32le(dst[48:52], x12)
+				endian.unchecked_put_u32le(dst[52:56], x13)
+				endian.unchecked_put_u32le(dst[56:60], x14)
+				endian.unchecked_put_u32le(dst[60:64], x15)
+			}
+			dst = dst[_chacha20.BLOCK_SIZE:]
+		}
+
+		// Increment the counter.  Overflow checking is done upon
+		// entry into the routine, so a 64-bit increment safely
+		// covers both cases.
+		new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + 1
+		x[12] = u32(new_ctr)
+		x[13] = u32(new_ctr >> 32)
+	}
+}
+
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	x0, x1, x2, x3 := _chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3
+	x4 := endian.unchecked_get_u32le(key[0:4])
+	x5 := endian.unchecked_get_u32le(key[4:8])
+	x6 := endian.unchecked_get_u32le(key[8:12])
+	x7 := endian.unchecked_get_u32le(key[12:16])
+	x8 := endian.unchecked_get_u32le(key[16:20])
+	x9 := endian.unchecked_get_u32le(key[20:24])
+	x10 := endian.unchecked_get_u32le(key[24:28])
+	x11 := endian.unchecked_get_u32le(key[28:32])
+	x12 := endian.unchecked_get_u32le(iv[0:4])
+	x13 := endian.unchecked_get_u32le(iv[4:8])
+	x14 := endian.unchecked_get_u32le(iv[8:12])
+	x15 := endian.unchecked_get_u32le(iv[12:16])
+
+	for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+		// quarterround(x, 0, 4, 8, 12)
+		x0 += x4
+		x12 ~= x0
+		x12 = bits.rotate_left32(x12, 16)
+		x8 += x12
+		x4 ~= x8
+		x4 = bits.rotate_left32(x4, 12)
+		x0 += x4
+		x12 ~= x0
+		x12 = bits.rotate_left32(x12, 8)
+		x8 += x12
+		x4 ~= x8
+		x4 = bits.rotate_left32(x4, 7)
+
+		// quarterround(x, 1, 5, 9, 13)
+		x1 += x5
+		x13 ~= x1
+		x13 = bits.rotate_left32(x13, 16)
+		x9 += x13
+		x5 ~= x9
+		x5 = bits.rotate_left32(x5, 12)
+		x1 += x5
+		x13 ~= x1
+		x13 = bits.rotate_left32(x13, 8)
+		x9 += x13
+		x5 ~= x9
+		x5 = bits.rotate_left32(x5, 7)
+
+		// quarterround(x, 2, 6, 10, 14)
+		x2 += x6
+		x14 ~= x2
+		x14 = bits.rotate_left32(x14, 16)
+		x10 += x14
+		x6 ~= x10
+		x6 = bits.rotate_left32(x6, 12)
+		x2 += x6
+		x14 ~= x2
+		x14 = bits.rotate_left32(x14, 8)
+		x10 += x14
+		x6 ~= x10
+		x6 = bits.rotate_left32(x6, 7)
+
+		// quarterround(x, 3, 7, 11, 15)
+		x3 += x7
+		x15 ~= x3
+		x15 = bits.rotate_left32(x15, 16)
+		x11 += x15
+		x7 ~= x11
+		x7 = bits.rotate_left32(x7, 12)
+		x3 += x7
+		x15 ~= x3
+		x15 = bits.rotate_left32(x15, 8)
+		x11 += x15
+		x7 ~= x11
+		x7 = bits.rotate_left32(x7, 7)
+
+		// quarterround(x, 0, 5, 10, 15)
+		x0 += x5
+		x15 ~= x0
+		x15 = bits.rotate_left32(x15, 16)
+		x10 += x15
+		x5 ~= x10
+		x5 = bits.rotate_left32(x5, 12)
+		x0 += x5
+		x15 ~= x0
+		x15 = bits.rotate_left32(x15, 8)
+		x10 += x15
+		x5 ~= x10
+		x5 = bits.rotate_left32(x5, 7)
+
+		// quarterround(x, 1, 6, 11, 12)
+		x1 += x6
+		x12 ~= x1
+		x12 = bits.rotate_left32(x12, 16)
+		x11 += x12
+		x6 ~= x11
+		x6 = bits.rotate_left32(x6, 12)
+		x1 += x6
+		x12 ~= x1
+		x12 = bits.rotate_left32(x12, 8)
+		x11 += x12
+		x6 ~= x11
+		x6 = bits.rotate_left32(x6, 7)
+
+		// quarterround(x, 2, 7, 8, 13)
+		x2 += x7
+		x13 ~= x2
+		x13 = bits.rotate_left32(x13, 16)
+		x8 += x13
+		x7 ~= x8
+		x7 = bits.rotate_left32(x7, 12)
+		x2 += x7
+		x13 ~= x2
+		x13 = bits.rotate_left32(x13, 8)
+		x8 += x13
+		x7 ~= x8
+		x7 = bits.rotate_left32(x7, 7)
+
+		// quarterround(x, 3, 4, 9, 14)
+		x3 += x4
+		x14 ~= x3
+		x14 = bits.rotate_left32(x14, 16)
+		x9 += x14
+		x4 ~= x9
+		x4 = bits.rotate_left32(x4, 12)
+		x3 += x4
+		x14 ~= x3
+		x14 = bits.rotate_left32(x14, 8)
+		x9 += x14
+		x4 ~= x9
+		x4 = bits.rotate_left32(x4, 7)
+	}
+
+	endian.unchecked_put_u32le(dst[0:4], x0)
+	endian.unchecked_put_u32le(dst[4:8], x1)
+	endian.unchecked_put_u32le(dst[8:12], x2)
+	endian.unchecked_put_u32le(dst[12:16], x3)
+	endian.unchecked_put_u32le(dst[16:20], x12)
+	endian.unchecked_put_u32le(dst[20:24], x13)
+	endian.unchecked_put_u32le(dst[24:28], x14)
+	endian.unchecked_put_u32le(dst[28:32], x15)
+}
diff --git a/core/crypto/_chacha20/simd128/chacha20_simd128.odin b/core/crypto/_chacha20/simd128/chacha20_simd128.odin
new file mode 100644
index 000000000..4cab3c5e8
--- /dev/null
+++ b/core/crypto/_chacha20/simd128/chacha20_simd128.odin
@@ -0,0 +1,481 @@
+package chacha20_simd128
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+import "core:simd"
+import "core:sys/info"
+
+// Portable 128-bit `core:simd` implementation.
+//
+// This is loosely based on Ted Krovetz's public domain C intrinsic
+// implementation.
+//
+// This is written to perform adequately on any target that has "enough"
+// 128-bit vector registers, the current thought is that 4 blocks at at
+// time is reasonable for amd64, though Ted's code is more conservative.
+//
+// See:
+// supercop-20230530/crypto_stream/chacha20/krovetz/vec128
+
+// Ensure the compiler emits SIMD instructions.  This is a minimum, and
+// setting the microarchitecture at compile time will allow for better
+// code gen when applicable (eg: AVX).  This is somewhat redundant with
+// the default microarchitecture configurations.
+when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 {
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: "neon"
+} else when ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+	// Note: LLVM appears to be smart enough to use PSHUFB despite not
+	// explicitly using simd.u8x16 shuffles.
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: "sse2,ssse3"
+} else {
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: ""
+}
+
+@(private = "file")
+_ROT_7L: simd.u32x4 : {7, 7, 7, 7}
+@(private = "file")
+_ROT_7R: simd.u32x4 : {25, 25, 25, 25}
+@(private = "file")
+_ROT_12L: simd.u32x4 : {12, 12, 12, 12}
+@(private = "file")
+_ROT_12R: simd.u32x4 : {20, 20, 20, 20}
+@(private = "file")
+_ROT_8L: simd.u32x4 : {8, 8, 8, 8}
+@(private = "file")
+_ROT_8R: simd.u32x4 : {24, 24, 24, 24}
+@(private = "file")
+_ROT_16: simd.u32x4 : {16, 16, 16, 16}
+
+when ODIN_ENDIAN == .Big {
+	@(private = "file")
+	_increment_counter :: #force_inline proc "contextless" (ctx: ^Context) -> simd.u32x4 {
+		// In the Big Endian case, the low and high portions in the vector
+		// are flipped, so the 64-bit addition can't be done with a simple
+		// vector add.
+		x := &ctx._s
+
+		new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + 1
+		x[12] = u32(new_ctr)
+		x[13] = u32(new_ctr >> 32)
+
+		return intrinsics.unaligned_load(transmute(^simd.u32x4)&x[12])
+	}
+
+	// Convert the endian-ness of the components of a u32x4 vector, for
+	// the purposes of output.
+	@(private = "file")
+	_byteswap_u32x4 :: #force_inline proc "contextless" (v: simd.u32x4) -> simd.u32x4 {
+		return(
+			transmute(simd.u32x4)simd.shuffle(
+				transmute(simd.u8x16)v,
+				transmute(simd.u8x16)v,
+				3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
+			)
+		)
+	}
+} else {
+	@(private = "file")
+	_VEC_ONE: simd.u64x2 : {1, 0}
+}
+
+@(private = "file")
+_dq_round_simd128 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV1(b); c = ROTV2(c);  d = ROTV3(d);
+	v1 = simd.shuffle(v1, v1, 1, 2, 3, 0)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1)
+	v3 = simd.shuffle(v3, v3, 3, 0, 1, 2)
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV3(b); c = ROTV2(c); d = ROTV1(d);
+	v1 = simd.shuffle(v1, v1, 3, 0, 1, 2)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1)
+	v3 = simd.shuffle(v3, v3, 1, 2, 3, 0)
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_add_state_simd128 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3, s0, s1, s2, s3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.add(v0, s0)
+	v1 = simd.add(v1, s1)
+	v2 = simd.add(v2, s2)
+	v3 = simd.add(v3, s3)
+
+	when ODIN_ENDIAN == .Big {
+		v0 = _byteswap_u32x4(v0)
+		v1 = _byteswap_u32x4(v1)
+		v2 = _byteswap_u32x4(v2)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_xor_simd128 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x4,
+	v0, v1, v2, v3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x4)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x4)(src[1:])))
+	v2 = simd.bit_xor(v2, intrinsics.unaligned_load((^simd.u32x4)(src[2:])))
+	v3 = simd.bit_xor(v3, intrinsics.unaligned_load((^simd.u32x4)(src[3:])))
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_store_simd128 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x4,
+	v0, v1, v2, v3: simd.u32x4,
+) {
+	intrinsics.unaligned_store((^simd.u32x4)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[1:]), v1)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[2:]), v2)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[3:]), v3)
+}
+
+// is_performant returns true iff the target and current host both support
+// "enough" 128-bit SIMD to make this implementation performant.
+is_performant :: proc "contextless" () -> bool {
+	when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 || ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+		when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 {
+			req_features :: info.CPU_Features{.asimd}
+		} else when ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+			req_features :: info.CPU_Features{.sse2, .ssse3}
+		}
+
+		features, ok := info.cpu_features.?
+		if !ok {
+			return false
+		}
+
+		return features >= req_features
+	} else when ODIN_ARCH == .wasm64p32 || ODIN_ARCH == .wasm32 {
+		return intrinsics.has_target_feature("simd128")
+	} else {
+		return false
+	}
+}
+
+@(enable_target_feature = TARGET_SIMD_FEATURES)
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst_v := ([^]simd.u32x4)(raw_data(dst))
+	src_v := ([^]simd.u32x4)(raw_data(src))
+
+	x := &ctx._s
+	n := nr_blocks
+
+	// The state vector is an array of uint32s in native byte-order.
+	x_v := ([^]simd.u32x4)(raw_data(x))
+	s0 := intrinsics.unaligned_load((^simd.u32x4)(x_v[0:]))
+	s1 := intrinsics.unaligned_load((^simd.u32x4)(x_v[1:]))
+	s2 := intrinsics.unaligned_load((^simd.u32x4)(x_v[2:]))
+	s3 := intrinsics.unaligned_load((^simd.u32x4)(x_v[3:]))
+
+	// 8 blocks at a time.
+	//
+	// Note: This is only worth it on Aarch64.
+	when ODIN_ARCH == .arm64 {
+		for ; n >= 8; n = n - 8 {
+			v0, v1, v2, v3 := s0, s1, s2, s3
+
+			when ODIN_ENDIAN == .Little {
+				s7 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+			} else {
+				s7 := _increment_counter(ctx)
+			}
+			v4, v5, v6, v7 := s0, s1, s2, s7
+
+			when ODIN_ENDIAN == .Little {
+				s11 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s7, _VEC_ONE)
+			} else {
+				s11 := _increment_counter(ctx)
+			}
+			v8, v9, v10, v11 := s0, s1, s2, s11
+
+			when ODIN_ENDIAN == .Little {
+				s15 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s11, _VEC_ONE)
+			} else {
+				s15 := _increment_counter(ctx)
+			}
+			v12, v13, v14, v15 := s0, s1, s2, s15
+
+			when ODIN_ENDIAN == .Little {
+				s19 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s15, _VEC_ONE)
+			} else {
+				s19 := _increment_counter(ctx)
+			}
+
+			v16, v17, v18, v19 := s0, s1, s2, s19
+			when ODIN_ENDIAN == .Little {
+				s23 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s19, _VEC_ONE)
+			} else {
+				s23 := _increment_counter(ctx)
+			}
+
+			v20, v21, v22, v23 := s0, s1, s2, s23
+			when ODIN_ENDIAN == .Little {
+				s27 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s23, _VEC_ONE)
+			} else {
+				s27 := _increment_counter(ctx)
+			}
+
+			v24, v25, v26, v27 := s0, s1, s2, s27
+			when ODIN_ENDIAN == .Little {
+				s31 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s27, _VEC_ONE)
+			} else {
+				s31 := _increment_counter(ctx)
+			}
+			v28, v29, v30, v31 := s0, s1, s2, s31
+
+			for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+				v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+				v4, v5, v6, v7 = _dq_round_simd128(v4, v5, v6, v7)
+				v8, v9, v10, v11 = _dq_round_simd128(v8, v9, v10, v11)
+				v12, v13, v14, v15 = _dq_round_simd128(v12, v13, v14, v15)
+				v16, v17, v18, v19 = _dq_round_simd128(v16, v17, v18, v19)
+				v20, v21, v22, v23 = _dq_round_simd128(v20, v21, v22, v23)
+				v24, v25, v26, v27 = _dq_round_simd128(v24, v25, v26, v27)
+				v28, v29, v30, v31 = _dq_round_simd128(v28, v29, v30, v31)
+			}
+
+			v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+			v4, v5, v6, v7 = _add_state_simd128(v4, v5, v6, v7, s0, s1, s2, s7)
+			v8, v9, v10, v11 = _add_state_simd128(v8, v9, v10, v11, s0, s1, s2, s11)
+			v12, v13, v14, v15 = _add_state_simd128(v12, v13, v14, v15, s0, s1, s2, s15)
+			v16, v17, v18, v19 = _add_state_simd128(v16, v17, v18, v19, s0, s1, s2, s19)
+			v20, v21, v22, v23 = _add_state_simd128(v20, v21, v22, v23, s0, s1, s2, s23)
+			v24, v25, v26, v27 = _add_state_simd128(v24, v25, v26, v27, s0, s1, s2, s27)
+			v28, v29, v30, v31 = _add_state_simd128(v28, v29, v30, v31, s0, s1, s2, s31)
+
+			#no_bounds_check {
+				if src != nil {
+					v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+					v4, v5, v6, v7 = _xor_simd128(src_v[4:], v4, v5, v6, v7)
+					v8, v9, v10, v11 = _xor_simd128(src_v[8:], v8, v9, v10, v11)
+					v12, v13, v14, v15 = _xor_simd128(src_v[12:], v12, v13, v14, v15)
+					v16, v17, v18, v19 = _xor_simd128(src_v[16:], v16, v17, v18, v19)
+					v20, v21, v22, v23 = _xor_simd128(src_v[20:], v20, v21, v22, v23)
+					v24, v25, v26, v27 = _xor_simd128(src_v[24:], v24, v25, v26, v27)
+					v28, v29, v30, v31 = _xor_simd128(src_v[28:], v28, v29, v30, v31)
+					src_v = src_v[32:]
+				}
+
+				_store_simd128(dst_v, v0, v1, v2, v3)
+				_store_simd128(dst_v[4:], v4, v5, v6, v7)
+				_store_simd128(dst_v[8:], v8, v9, v10, v11)
+				_store_simd128(dst_v[12:], v12, v13, v14, v15)
+				_store_simd128(dst_v[16:], v16, v17, v18, v19)
+				_store_simd128(dst_v[20:], v20, v21, v22, v23)
+				_store_simd128(dst_v[24:], v24, v25, v26, v27)
+				_store_simd128(dst_v[28:], v28, v29, v30, v31)
+				dst_v = dst_v[32:]
+			}
+
+			when ODIN_ENDIAN == .Little {
+				// s31 holds the most current counter, so `s3 = s31 + 1`.
+				s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s31, _VEC_ONE)
+			} else {
+				s3 = _increment_counter(ctx)
+			}
+		}
+	}
+
+	// 4 blocks at a time.
+	//
+	// Note: The i386 target lacks the required number of registers
+	// for this to be performant, so it is skipped.
+	when ODIN_ARCH != .i386 {
+		for ; n >= 4; n = n - 4 {
+			v0, v1, v2, v3 := s0, s1, s2, s3
+
+			when ODIN_ENDIAN == .Little {
+				s7 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+			} else {
+				s7 := _increment_counter(ctx)
+			}
+			v4, v5, v6, v7 := s0, s1, s2, s7
+
+			when ODIN_ENDIAN == .Little {
+				s11 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s7, _VEC_ONE)
+			} else {
+				s11 := _increment_counter(ctx)
+			}
+			v8, v9, v10, v11 := s0, s1, s2, s11
+
+			when ODIN_ENDIAN == .Little {
+				s15 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s11, _VEC_ONE)
+			} else {
+				s15 := _increment_counter(ctx)
+			}
+			v12, v13, v14, v15 := s0, s1, s2, s15
+
+			for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+				v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+				v4, v5, v6, v7 = _dq_round_simd128(v4, v5, v6, v7)
+				v8, v9, v10, v11 = _dq_round_simd128(v8, v9, v10, v11)
+				v12, v13, v14, v15 = _dq_round_simd128(v12, v13, v14, v15)
+			}
+
+			v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+			v4, v5, v6, v7 = _add_state_simd128(v4, v5, v6, v7, s0, s1, s2, s7)
+			v8, v9, v10, v11 = _add_state_simd128(v8, v9, v10, v11, s0, s1, s2, s11)
+			v12, v13, v14, v15 = _add_state_simd128(v12, v13, v14, v15, s0, s1, s2, s15)
+
+			#no_bounds_check {
+				if src != nil {
+					v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+					v4, v5, v6, v7 = _xor_simd128(src_v[4:], v4, v5, v6, v7)
+					v8, v9, v10, v11 = _xor_simd128(src_v[8:], v8, v9, v10, v11)
+					v12, v13, v14, v15 = _xor_simd128(src_v[12:], v12, v13, v14, v15)
+					src_v = src_v[16:]
+				}
+
+				_store_simd128(dst_v, v0, v1, v2, v3)
+				_store_simd128(dst_v[4:], v4, v5, v6, v7)
+				_store_simd128(dst_v[8:], v8, v9, v10, v11)
+				_store_simd128(dst_v[12:], v12, v13, v14, v15)
+				dst_v = dst_v[16:]
+			}
+
+			when ODIN_ENDIAN == .Little {
+				// s15 holds the most current counter, so `s3 = s15 + 1`.
+				s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s15, _VEC_ONE)
+			} else {
+				s3 = _increment_counter(ctx)
+			}
+		}
+	}
+
+	// 1 block at a time.
+	for ; n > 0; n = n - 1 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+		}
+		v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+				src_v = src_v[4:]
+			}
+
+			_store_simd128(dst_v, v0, v1, v2, v3)
+			dst_v = dst_v[4:]
+		}
+
+		// Increment the counter.  Overflow checking is done upon
+		// entry into the routine, so a 64-bit increment safely
+		// covers both cases.
+		when ODIN_ENDIAN == .Little {
+			s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+		} else {
+			s3 = _increment_counter(ctx)
+		}
+	}
+
+	when ODIN_ENDIAN == .Little {
+		// Write back the counter to the state.
+		intrinsics.unaligned_store((^simd.u32x4)(x_v[3:]), s3)
+	}
+}
+
+@(enable_target_feature = TARGET_SIMD_FEATURES)
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	v0 := simd.u32x4{_chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3}
+	v1 := intrinsics.unaligned_load((^simd.u32x4)(&key[0]))
+	v2 := intrinsics.unaligned_load((^simd.u32x4)(&key[16]))
+	v3 := intrinsics.unaligned_load((^simd.u32x4)(&iv[0]))
+
+	when ODIN_ENDIAN == .Big {
+		v1 = _byteswap_u32x4(v1)
+		v2 = _byteswap_u32x4(v2)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+		v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+	}
+
+	when ODIN_ENDIAN == .Big {
+		v0 = _byteswap_u32x4(v0)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	dst_v := ([^]simd.u32x4)(raw_data(dst))
+	intrinsics.unaligned_store((^simd.u32x4)(dst_v[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x4)(dst_v[1:]), v3)
+}
diff --git a/core/crypto/_chacha20/simd256/chacha20_simd256.odin b/core/crypto/_chacha20/simd256/chacha20_simd256.odin
new file mode 100644
index 000000000..10f2d75fe
--- /dev/null
+++ b/core/crypto/_chacha20/simd256/chacha20_simd256.odin
@@ -0,0 +1,319 @@
+//+build amd64
+package chacha20_simd256
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+import chacha_simd128 "core:crypto/_chacha20/simd128"
+import "core:simd"
+import "core:sys/info"
+
+// This is loosely based on Ted Krovetz's public domain C intrinsic
+// implementations.  While written using `core:simd`, this is currently
+// amd64 specific because we do not have a way to detect ARM SVE.
+//
+// See:
+// supercop-20230530/crypto_stream/chacha20/krovetz/vec128
+// supercop-20230530/crypto_stream/chacha20/krovetz/avx2
+
+#assert(ODIN_ENDIAN == .Little)
+
+@(private = "file")
+_ROT_7L: simd.u32x8 : {7, 7, 7, 7, 7, 7, 7, 7}
+@(private = "file")
+_ROT_7R: simd.u32x8 : {25, 25, 25, 25, 25, 25, 25, 25}
+@(private = "file")
+_ROT_12L: simd.u32x8 : {12, 12, 12, 12, 12, 12, 12, 12}
+@(private = "file")
+_ROT_12R: simd.u32x8 : {20, 20, 20, 20, 20, 20, 20, 20}
+@(private = "file")
+_ROT_8L: simd.u32x8 : {8, 8, 8, 8, 8, 8, 8, 8}
+@(private = "file")
+_ROT_8R: simd.u32x8 : {24, 24, 24, 24, 24, 24, 24, 24}
+@(private = "file")
+_ROT_16: simd.u32x8 : {16, 16, 16, 16, 16, 16, 16, 16}
+@(private = "file")
+_VEC_ZERO_ONE: simd.u64x4 : {0, 0, 1, 0}
+@(private = "file")
+_VEC_TWO: simd.u64x4 : {2, 0, 2, 0}
+
+// is_performant returns true iff the target and current host both support
+// "enough" SIMD to make this implementation performant.
+is_performant :: proc "contextless" () -> bool {
+	req_features :: info.CPU_Features{.avx, .avx2}
+
+	features, ok := info.cpu_features.?
+	if !ok {
+		return false
+	}
+
+	return features >= req_features
+}
+
+@(private = "file")
+_dq_round_simd256 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV1(b); c = ROTV2(c);  d = ROTV3(d);
+	v1 = simd.shuffle(v1, v1, 1, 2, 3, 0, 5, 6, 7, 4)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1, 6, 7, 4, 5)
+	v3 = simd.shuffle(v3, v3, 3, 0, 1, 2, 7, 4, 5, 6)
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV3(b); c = ROTV2(c); d = ROTV1(d);
+	v1 = simd.shuffle(v1, v1, 3, 0, 1, 2, 7, 4, 5, 6)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1, 6, 7, 4, 5)
+	v3 = simd.shuffle(v3, v3, 1, 2, 3, 0, 5, 6, 7, 4)
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_add_and_permute_state_simd256 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3, s0, s1, s2, s3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	t0 := simd.add(v0, s0)
+	t1 := simd.add(v1, s1)
+	t2 := simd.add(v2, s2)
+	t3 := simd.add(v3, s3)
+
+	// Big Endian would byteswap here.
+
+	// Each of v0 .. v3 has 128-bits of keystream for 2 separate blocks.
+	// permute the state such that (r0, r1) contains block 0, and (r2, r3)
+	// contains block 1.
+	r0 := simd.shuffle(t0, t1, 0, 1, 2, 3, 8, 9, 10, 11)
+	r2 := simd.shuffle(t0, t1, 4, 5, 6, 7, 12, 13, 14, 15)
+	r1 := simd.shuffle(t2, t3, 0, 1, 2, 3, 8, 9, 10, 11)
+	r3 := simd.shuffle(t2, t3, 4, 5, 6, 7, 12, 13, 14, 15)
+
+	return r0, r1, r2, r3
+}
+
+@(private = "file")
+_xor_simd256 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x8,
+	v0, v1, v2, v3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x8)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x8)(src[1:])))
+	v2 = simd.bit_xor(v2, intrinsics.unaligned_load((^simd.u32x8)(src[2:])))
+	v3 = simd.bit_xor(v3, intrinsics.unaligned_load((^simd.u32x8)(src[3:])))
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_xor_simd256_x1 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x8,
+	v0, v1: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1 := v0, v1
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x8)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x8)(src[1:])))
+
+	return v0, v1
+}
+
+@(private = "file")
+_store_simd256 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x8,
+	v0, v1, v2, v3: simd.u32x8,
+) {
+	intrinsics.unaligned_store((^simd.u32x8)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[1:]), v1)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[2:]), v2)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[3:]), v3)
+}
+
+@(private = "file")
+_store_simd256_x1 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x8,
+	v0, v1: simd.u32x8,
+) {
+	intrinsics.unaligned_store((^simd.u32x8)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[1:]), v1)
+}
+
+@(enable_target_feature = "sse2,ssse3,avx,avx2")
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst_v := ([^]simd.u32x8)(raw_data(dst))
+	src_v := ([^]simd.u32x8)(raw_data(src))
+
+	x := &ctx._s
+	n := nr_blocks
+
+	// The state vector is an array of uint32s in native byte-order.
+	// Setup s0 .. s3 such that each register stores 2 copies of the
+	// state.
+	x_v := ([^]simd.u32x4)(raw_data(x))
+	t0 := intrinsics.unaligned_load((^simd.u32x4)(x_v[0:]))
+	t1 := intrinsics.unaligned_load((^simd.u32x4)(x_v[1:]))
+	t2 := intrinsics.unaligned_load((^simd.u32x4)(x_v[2:]))
+	t3 := intrinsics.unaligned_load((^simd.u32x4)(x_v[3:]))
+	s0 := simd.swizzle(t0, 0, 1, 2, 3, 0, 1, 2, 3)
+	s1 := simd.swizzle(t1, 0, 1, 2, 3, 0, 1, 2, 3)
+	s2 := simd.swizzle(t2, 0, 1, 2, 3, 0, 1, 2, 3)
+	s3 := simd.swizzle(t3, 0, 1, 2, 3, 0, 1, 2, 3)
+
+	// Advance the counter in the 2nd copy of the state by one.
+	s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_ZERO_ONE)
+
+	// 8 blocks at a time.
+	for ; n >= 8; n = n - 8 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		s7 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_TWO)
+		v4, v5, v6, v7 := s0, s1, s2, s7
+
+		s11 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s7, _VEC_TWO)
+		v8, v9, v10, v11 := s0, s1, s2, s11
+
+		s15 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s11, _VEC_TWO)
+		v12, v13, v14, v15 := s0, s1, s2, s15
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd256(v0, v1, v2, v3)
+			v4, v5, v6, v7 = _dq_round_simd256(v4, v5, v6, v7)
+			v8, v9, v10, v11 = _dq_round_simd256(v8, v9, v10, v11)
+			v12, v13, v14, v15 = _dq_round_simd256(v12, v13, v14, v15)
+		}
+
+		v0, v1, v2, v3 = _add_and_permute_state_simd256(v0, v1, v2, v3, s0, s1, s2, s3)
+		v4, v5, v6, v7 = _add_and_permute_state_simd256(v4, v5, v6, v7, s0, s1, s2, s7)
+		v8, v9, v10, v11 = _add_and_permute_state_simd256(v8, v9, v10, v11, s0, s1, s2, s11)
+		v12, v13, v14, v15 = _add_and_permute_state_simd256(v12, v13, v14, v15, s0, s1, s2, s15)
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd256(src_v, v0, v1, v2, v3)
+				v4, v5, v6, v7 = _xor_simd256(src_v[4:], v4, v5, v6, v7)
+				v8, v9, v10, v11 = _xor_simd256(src_v[8:], v8, v9, v10, v11)
+				v12, v13, v14, v15 = _xor_simd256(src_v[12:], v12, v13, v14, v15)
+				src_v = src_v[16:]
+			}
+
+			_store_simd256(dst_v, v0, v1, v2, v3)
+			_store_simd256(dst_v[4:], v4, v5, v6, v7)
+			_store_simd256(dst_v[8:], v8, v9, v10, v11)
+			_store_simd256(dst_v[12:], v12, v13, v14, v15)
+			dst_v = dst_v[16:]
+		}
+
+		s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s15, _VEC_TWO)
+	}
+
+
+	// 2 (or 1) block at a time.
+	for ; n > 0; n = n - 2 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd256(v0, v1, v2, v3)
+		}
+		v0, v1, v2, v3 = _add_and_permute_state_simd256(v0, v1, v2, v3, s0, s1, s2, s3)
+
+		if n == 1 {
+			// Note: No need to advance src_v, dst_v, or increment the counter
+			// since this is guaranteed to be the final block.
+			#no_bounds_check {
+				if src != nil {
+					v0, v1 = _xor_simd256_x1(src_v, v0, v1)
+				}
+
+				_store_simd256_x1(dst_v, v0, v1)
+			}
+			break
+		}
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd256(src_v, v0, v1, v2, v3)
+				src_v = src_v[4:]
+			}
+
+			_store_simd256(dst_v, v0, v1, v2, v3)
+			dst_v = dst_v[4:]
+		}
+
+		s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_TWO)
+	}
+
+	// Write back the counter.  Doing it this way, saves having to
+	// pull out the correct counter value from s3.
+	new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + u64(nr_blocks)
+	ctx._s[12] = u32(new_ctr)
+	ctx._s[13] = u32(new_ctr >> 32)
+}
+
+@(enable_target_feature = "sse2,ssse3,avx")
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	// We can just enable AVX and call the simd128 code as going
+	// wider has 0 performance benefit, but VEX encoded instructions
+	// is nice.
+	#force_inline chacha_simd128.hchacha20(dst, key, iv)
+}
\ No newline at end of file
diff --git a/core/crypto/_chacha20/simd256/chacha20_simd256_stub.odin b/core/crypto/_chacha20/simd256/chacha20_simd256_stub.odin
new file mode 100644
index 000000000..039d6cb96
--- /dev/null
+++ b/core/crypto/_chacha20/simd256/chacha20_simd256_stub.odin
@@ -0,0 +1,17 @@
+//+build !amd64
+package chacha20_simd256
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+
+is_performant :: proc "contextless" () -> bool {
+	return false
+}
+
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	panic("crypto/chacha20: simd256 implementation unsupported")
+}
+
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	intrinsics.trap()
+}
\ No newline at end of file
diff --git a/core/crypto/aead/aead.odin b/core/crypto/aead/aead.odin
new file mode 100644
index 000000000..9b7d810e4
--- /dev/null
+++ b/core/crypto/aead/aead.odin
@@ -0,0 +1,36 @@
+package aead
+
+// seal_oneshot encrypts the plaintext and authenticates the aad and ciphertext,
+// with the provided algorithm, key, and iv, stores the output in dst and tag.
+//
+// dst and plaintext MUST alias exactly or not at all.
+seal_oneshot :: proc(algo: Algorithm, dst, tag, key, iv, aad, plaintext: []byte, impl: Implementation = nil) {
+	ctx: Context
+	init(&ctx, algo, key, impl)
+	defer reset(&ctx)
+	seal_ctx(&ctx, dst, tag, iv, aad, plaintext)
+}
+
+// open authenticates the aad and ciphertext, and decrypts the ciphertext,
+// with the provided algorithm, key, iv, and tag, and stores the output in dst,
+// returning true iff the authentication was successful.  If authentication
+// fails, the destination buffer will be zeroed.
+//
+// dst and plaintext MUST alias exactly or not at all.
+@(require_results)
+open_oneshot :: proc(algo: Algorithm, dst, key, iv, aad, ciphertext, tag: []byte, impl: Implementation = nil) -> bool {
+	ctx: Context
+	init(&ctx, algo, key, impl)
+	defer reset(&ctx)
+	return open_ctx(&ctx, dst, iv, aad, ciphertext, tag)
+}
+
+seal :: proc {
+	seal_ctx,
+	seal_oneshot,
+}
+
+open :: proc {
+	open_ctx,
+	open_oneshot,
+}
diff --git a/core/crypto/aead/doc.odin b/core/crypto/aead/doc.odin
new file mode 100644
index 000000000..388d31453
--- /dev/null
+++ b/core/crypto/aead/doc.odin
@@ -0,0 +1,58 @@
+/*
+package aead provides a generic interface to the supported Authenticated
+Encryption with Associated Data algorithms.
+
+Both a one-shot and context based interface are provided, with similar
+usage.  If multiple messages are to be sealed/opened via the same key,
+the context based interface may be more efficient, depending on the
+algorithm.
+
+WARNING: Reusing the same key + iv to seal (encrypt) multiple messages
+results in catastrophic loss of security for most algorithms.
+
+```odin
+package aead_example
+
+import "core:bytes"
+import "core:crypto"
+import "core:crypto/aead"
+
+main :: proc() {
+	algo := aead.Algorithm.XCHACHA20POLY1305
+
+	// The example added associated data, and plaintext.
+	aad_str := "Get your ass in gear boys."
+	pt_str := "They're immanetizing the Eschaton."
+
+	aad := transmute([]byte)aad_str
+	plaintext := transmute([]byte)pt_str
+	pt_len := len(plaintext)
+
+	// Generate a random key for the purposes of illustration.
+	key := make([]byte, aead.KEY_SIZES[algo])
+	defer delete(key)
+	crypto.rand_bytes(key)
+
+	// `ciphertext || tag`, is a common way data is transmitted, so
+	// demonstrate that.
+	buf := make([]byte, pt_len + aead.TAG_SIZES[algo])
+	defer delete(buf)
+	ciphertext, tag := buf[:pt_len], buf[pt_len:]
+
+	// Seal the AAD + Plaintext.
+	iv := make([]byte, aead.IV_SIZES[algo])
+	defer delete(iv)
+	crypto.rand_bytes(iv) // Random IVs are safe with XChaCha20-Poly1305.
+	aead.seal(algo, ciphertext, tag, key, iv, aad, plaintext)
+
+	// Open the AAD + Ciphertext.
+	opened_pt := buf[:pt_len]
+	if ok := aead.open(algo, opened_pt, key, iv, aad, ciphertext, tag); !ok {
+		panic("aead example: failed to open")
+	}
+
+	assert(bytes.equal(opened_pt, plaintext))
+}
+```
+*/
+package aead
\ No newline at end of file
diff --git a/core/crypto/aead/low_level.odin b/core/crypto/aead/low_level.odin
new file mode 100644
index 000000000..38a0c84ba
--- /dev/null
+++ b/core/crypto/aead/low_level.odin
@@ -0,0 +1,187 @@
+package aead
+
+import "core:crypto/aes"
+import "core:crypto/chacha20"
+import "core:crypto/chacha20poly1305"
+import "core:reflect"
+
+// Implementation is an AEAD implementation.  Most callers will not need
+// to use this as the package will automatically select the most performant
+// implementation available.
+Implementation :: union {
+	aes.Implementation,
+	chacha20.Implementation,
+}
+
+// MAX_TAG_SIZE is the maximum size tag that can be returned by any of the
+// Algorithms supported via this package.
+MAX_TAG_SIZE :: 16
+
+// Algorithm is the algorithm identifier associated with a given Context.
+Algorithm :: enum {
+	Invalid,
+	AES_GCM_128,
+	AES_GCM_192,
+	AES_GCM_256,
+	CHACHA20POLY1305,
+	XCHACHA20POLY1305,
+}
+
+// ALGORITM_NAMES is the Agorithm to algorithm name string.
+ALGORITHM_NAMES := [Algorithm]string {
+	.Invalid           = "Invalid",
+	.AES_GCM_128       = "AES-GCM-128",
+	.AES_GCM_192       = "AES-GCM-192",
+	.AES_GCM_256       = "AES-GCM-256",
+	.CHACHA20POLY1305  = "chacha20poly1305",
+	.XCHACHA20POLY1305 = "xchacha20poly1305",
+}
+
+// TAG_SIZES is the Algorithm to tag size in bytes.
+TAG_SIZES := [Algorithm]int {
+	.Invalid           = 0,
+	.AES_GCM_128       = aes.GCM_TAG_SIZE,
+	.AES_GCM_192       = aes.GCM_TAG_SIZE,
+	.AES_GCM_256       = aes.GCM_TAG_SIZE,
+	.CHACHA20POLY1305  = chacha20poly1305.TAG_SIZE,
+	.XCHACHA20POLY1305 = chacha20poly1305.TAG_SIZE,
+}
+
+// KEY_SIZES is the Algorithm to key size in bytes.
+KEY_SIZES := [Algorithm]int {
+	.Invalid           = 0,
+	.AES_GCM_128       = aes.KEY_SIZE_128,
+	.AES_GCM_192       = aes.KEY_SIZE_192,
+	.AES_GCM_256       = aes.KEY_SIZE_256,
+	.CHACHA20POLY1305  = chacha20poly1305.KEY_SIZE,
+	.XCHACHA20POLY1305 = chacha20poly1305.KEY_SIZE,
+}
+
+// IV_SIZES is the Algorithm to initialization vector size in bytes.
+//
+// Note: Some algorithms (such as AES-GCM) support variable IV sizes.
+IV_SIZES := [Algorithm]int {
+	.Invalid           = 0,
+	.AES_GCM_128       = aes.GCM_IV_SIZE,
+	.AES_GCM_192       = aes.GCM_IV_SIZE,
+	.AES_GCM_256       = aes.GCM_IV_SIZE,
+	.CHACHA20POLY1305  = chacha20poly1305.IV_SIZE,
+	.XCHACHA20POLY1305 = chacha20poly1305.XIV_SIZE,
+}
+
+// Context is a concrete instantiation of a specific AEAD algorithm.
+Context :: struct {
+	_algo: Algorithm,
+	_impl: union {
+		aes.Context_GCM,
+		chacha20poly1305.Context,
+	},
+}
+
+@(private)
+_IMPL_IDS := [Algorithm]typeid {
+	.Invalid           = nil,
+	.AES_GCM_128       = typeid_of(aes.Context_GCM),
+	.AES_GCM_192       = typeid_of(aes.Context_GCM),
+	.AES_GCM_256       = typeid_of(aes.Context_GCM),
+	.CHACHA20POLY1305  = typeid_of(chacha20poly1305.Context),
+	.XCHACHA20POLY1305 = typeid_of(chacha20poly1305.Context),
+}
+
+// init initializes a Context with a specific AEAD Algorithm.
+init :: proc(ctx: ^Context, algorithm: Algorithm, key: []byte, impl: Implementation = nil) {
+	if ctx._impl != nil {
+		reset(ctx)
+	}
+
+	if len(key) != KEY_SIZES[algorithm] {
+		panic("crypto/aead: invalid key size")
+	}
+
+	// Directly specialize the union by setting the type ID (save a copy).
+	reflect.set_union_variant_typeid(
+		ctx._impl,
+		_IMPL_IDS[algorithm],
+	)
+	switch algorithm {
+	case .AES_GCM_128, .AES_GCM_192, .AES_GCM_256:
+		impl_ := impl != nil ? impl.(aes.Implementation) : aes.DEFAULT_IMPLEMENTATION
+		aes.init_gcm(&ctx._impl.(aes.Context_GCM), key, impl_)
+	case .CHACHA20POLY1305:
+		impl_ := impl != nil ? impl.(chacha20.Implementation) : chacha20.DEFAULT_IMPLEMENTATION
+		chacha20poly1305.init(&ctx._impl.(chacha20poly1305.Context), key, impl_)
+	case .XCHACHA20POLY1305:
+		impl_ := impl != nil ? impl.(chacha20.Implementation) : chacha20.DEFAULT_IMPLEMENTATION
+		chacha20poly1305.init_xchacha(&ctx._impl.(chacha20poly1305.Context), key, impl_)
+	case .Invalid:
+		panic("crypto/aead: uninitialized algorithm")
+	case:
+		panic("crypto/aead: invalid algorithm")
+	}
+
+	ctx._algo = algorithm
+}
+
+// seal_ctx encrypts the plaintext and authenticates the aad and ciphertext,
+// with the provided Context and iv, stores the output in dst and tag.
+//
+// dst and plaintext MUST alias exactly or not at all.
+seal_ctx :: proc(ctx: ^Context, dst, tag, iv, aad, plaintext: []byte) {
+	switch &impl in ctx._impl {
+	case aes.Context_GCM:
+		aes.seal_gcm(&impl, dst, tag, iv, aad, plaintext)
+	case chacha20poly1305.Context:
+		chacha20poly1305.seal(&impl, dst, tag, iv, aad, plaintext)
+	case:
+		panic("crypto/aead: uninitialized algorithm")
+	}
+}
+
+// open_ctx authenticates the aad and ciphertext, and decrypts the ciphertext,
+// with the provided Context, iv, and tag, and stores the output in dst,
+// returning true iff the authentication was successful.  If authentication
+// fails, the destination buffer will be zeroed.
+//
+// dst and plaintext MUST alias exactly or not at all.
+@(require_results)
+open_ctx :: proc(ctx: ^Context, dst, iv, aad, ciphertext, tag: []byte) -> bool {
+	switch &impl in ctx._impl {
+	case aes.Context_GCM:
+		return aes.open_gcm(&impl, dst, iv, aad, ciphertext, tag)
+	case chacha20poly1305.Context:
+		return chacha20poly1305.open(&impl, dst, iv, aad, ciphertext, tag)
+	case:
+		panic("crypto/aead: uninitialized algorithm")
+	}
+}
+
+// reset sanitizes the Context.  The Context must be re-initialized to
+// be used again.
+reset :: proc(ctx: ^Context) {
+	switch &impl in ctx._impl {
+	case aes.Context_GCM:
+		aes.reset_gcm(&impl)
+	case chacha20poly1305.Context:
+		chacha20poly1305.reset(&impl)
+	case:
+		// Calling reset repeatedly is fine.
+	}
+
+	ctx._algo = .Invalid
+	ctx._impl = nil
+}
+
+// algorithm returns the Algorithm used by a Context instance.
+algorithm :: proc(ctx: ^Context) -> Algorithm {
+	return ctx._algo
+}
+
+// iv_size returns the IV size of a Context instance in bytes.
+iv_size :: proc(ctx: ^Context) -> int {
+	return IV_SIZES[ctx._algo]
+}
+
+// tag_size returns the tag size of a Context instance in bytes.
+tag_size :: proc(ctx: ^Context) -> int {
+	return TAG_SIZES[ctx._algo]
+}
diff --git a/core/crypto/aes/aes_ctr.odin b/core/crypto/aes/aes_ctr.odin
index 1c5fe31e8..20b75e57f 100644
--- a/core/crypto/aes/aes_ctr.odin
+++ b/core/crypto/aes/aes_ctr.odin
@@ -20,7 +20,7 @@ Context_CTR :: struct {
 }
 
 // init_ctr initializes a Context_CTR with the provided key and IV.
-init_ctr :: proc(ctx: ^Context_CTR, key, iv: []byte, impl := Implementation.Hardware) {
+init_ctr :: proc(ctx: ^Context_CTR, key, iv: []byte, impl := DEFAULT_IMPLEMENTATION) {
 	if len(iv) != CTR_IV_SIZE {
 		panic("crypto/aes: invalid CTR IV size")
 	}
@@ -47,7 +47,7 @@ xor_bytes_ctr :: proc(ctx: ^Context_CTR, dst, src: []byte) {
 		panic("crypto/aes: dst and src alias inexactly")
 	}
 
-	for remaining := len(src); remaining > 0; {
+	#no_bounds_check for remaining := len(src); remaining > 0; {
 		// Process multiple blocks at once
 		if ctx._off == BLOCK_SIZE {
 			if nr_blocks := remaining / BLOCK_SIZE; nr_blocks > 0 {
@@ -85,7 +85,7 @@ keystream_bytes_ctr :: proc(ctx: ^Context_CTR, dst: []byte) {
 	assert(ctx._is_initialized)
 
 	dst := dst
-	for remaining := len(dst); remaining > 0; {
+	#no_bounds_check for remaining := len(dst); remaining > 0; {
 		// Process multiple blocks at once
 		if ctx._off == BLOCK_SIZE {
 			if nr_blocks := remaining / BLOCK_SIZE; nr_blocks > 0 {
diff --git a/core/crypto/aes/aes_ecb.odin b/core/crypto/aes/aes_ecb.odin
index 498429e29..32476006c 100644
--- a/core/crypto/aes/aes_ecb.odin
+++ b/core/crypto/aes/aes_ecb.odin
@@ -12,7 +12,7 @@ Context_ECB :: struct {
 }
 
 // init_ecb initializes a Context_ECB with the provided key.
-init_ecb :: proc(ctx: ^Context_ECB, key: []byte, impl := Implementation.Hardware) {
+init_ecb :: proc(ctx: ^Context_ECB, key: []byte, impl := DEFAULT_IMPLEMENTATION) {
 	init_impl(&ctx._impl, key, impl)
 	ctx._is_initialized = true
 }
diff --git a/core/crypto/aes/aes_gcm.odin b/core/crypto/aes/aes_gcm.odin
index 25e0cc35b..8616821ce 100644
--- a/core/crypto/aes/aes_gcm.odin
+++ b/core/crypto/aes/aes_gcm.odin
@@ -7,10 +7,10 @@ import "core:crypto/_aes/ct64"
 import "core:encoding/endian"
 import "core:mem"
 
-// GCM_NONCE_SIZE is the default size of the GCM nonce in bytes.
-GCM_NONCE_SIZE :: 12
-// GCM_NONCE_SIZE_MAX is the maximum size of the GCM nonce in bytes.
-GCM_NONCE_SIZE_MAX :: 0x2000000000000000 // floor((2^64 - 1) / 8) bits
+// GCM_IV_SIZE is the default size of the GCM IV in bytes.
+GCM_IV_SIZE :: 12
+// GCM_IV_SIZE_MAX is the maximum size of the GCM IV in bytes.
+GCM_IV_SIZE_MAX :: 0x2000000000000000 // floor((2^64 - 1) / 8) bits
 // GCM_TAG_SIZE is the size of a GCM tag in bytes.
 GCM_TAG_SIZE :: _aes.GHASH_TAG_SIZE
 
@@ -26,19 +26,19 @@ Context_GCM :: struct {
 }
 
 // init_gcm initializes a Context_GCM with the provided key.
-init_gcm :: proc(ctx: ^Context_GCM, key: []byte, impl := Implementation.Hardware) {
+init_gcm :: proc(ctx: ^Context_GCM, key: []byte, impl := DEFAULT_IMPLEMENTATION) {
 	init_impl(&ctx._impl, key, impl)
 	ctx._is_initialized = true
 }
 
 // seal_gcm encrypts the plaintext and authenticates the aad and ciphertext,
-// with the provided Context_GCM and nonce, stores the output in dst and tag.
+// with the provided Context_GCM and iv, stores the output in dst and tag.
 //
 // dst and plaintext MUST alias exactly or not at all.
-seal_gcm :: proc(ctx: ^Context_GCM, dst, tag, nonce, aad, plaintext: []byte) {
+seal_gcm :: proc(ctx: ^Context_GCM, dst, tag, iv, aad, plaintext: []byte) {
 	assert(ctx._is_initialized)
 
-	gcm_validate_common_slice_sizes(tag, nonce, aad, plaintext)
+	gcm_validate_common_slice_sizes(tag, iv, aad, plaintext)
 	if len(dst) != len(plaintext) {
 		panic("crypto/aes: invalid destination ciphertext size")
 	}
@@ -47,7 +47,7 @@ seal_gcm :: proc(ctx: ^Context_GCM, dst, tag, nonce, aad, plaintext: []byte) {
 	}
 
 	if impl, is_hw := ctx._impl.(Context_Impl_Hardware); is_hw {
-		gcm_seal_hw(&impl, dst, tag, nonce, aad, plaintext)
+		gcm_seal_hw(&impl, dst, tag, iv, aad, plaintext)
 		return
 	}
 
@@ -55,7 +55,7 @@ seal_gcm :: proc(ctx: ^Context_GCM, dst, tag, nonce, aad, plaintext: []byte) {
 	j0: [_aes.GHASH_BLOCK_SIZE]byte
 	j0_enc: [_aes.GHASH_BLOCK_SIZE]byte
 	s: [_aes.GHASH_TAG_SIZE]byte
-	init_ghash_ct64(ctx, &h, &j0, &j0_enc, nonce)
+	init_ghash_ct64(ctx, &h, &j0, &j0_enc, iv)
 
 	// Note: Our GHASH implementation handles appending padding.
 	ct64.ghash(s[:], h[:], aad)
@@ -69,15 +69,16 @@ seal_gcm :: proc(ctx: ^Context_GCM, dst, tag, nonce, aad, plaintext: []byte) {
 }
 
 // open_gcm authenticates the aad and ciphertext, and decrypts the ciphertext,
-// with the provided Context_GCM, nonce, and tag, and stores the output in dst,
+// with the provided Context_GCM, iv, and tag, and stores the output in dst,
 // returning true iff the authentication was successful.  If authentication
 // fails, the destination buffer will be zeroed.
 //
 // dst and plaintext MUST alias exactly or not at all.
-open_gcm :: proc(ctx: ^Context_GCM, dst, nonce, aad, ciphertext, tag: []byte) -> bool {
+@(require_results)
+open_gcm :: proc(ctx: ^Context_GCM, dst, iv, aad, ciphertext, tag: []byte) -> bool {
 	assert(ctx._is_initialized)
 
-	gcm_validate_common_slice_sizes(tag, nonce, aad, ciphertext)
+	gcm_validate_common_slice_sizes(tag, iv, aad, ciphertext)
 	if len(dst) != len(ciphertext) {
 		panic("crypto/aes: invalid destination plaintext size")
 	}
@@ -86,14 +87,14 @@ open_gcm :: proc(ctx: ^Context_GCM, dst, nonce, aad, ciphertext, tag: []byte) ->
 	}
 
 	if impl, is_hw := ctx._impl.(Context_Impl_Hardware); is_hw {
-		return gcm_open_hw(&impl, dst, nonce, aad, ciphertext, tag)
+		return gcm_open_hw(&impl, dst, iv, aad, ciphertext, tag)
 	}
 
 	h: [_aes.GHASH_KEY_SIZE]byte
 	j0: [_aes.GHASH_BLOCK_SIZE]byte
 	j0_enc: [_aes.GHASH_BLOCK_SIZE]byte
 	s: [_aes.GHASH_TAG_SIZE]byte
-	init_ghash_ct64(ctx, &h, &j0, &j0_enc, nonce)
+	init_ghash_ct64(ctx, &h, &j0, &j0_enc, iv)
 
 	ct64.ghash(s[:], h[:], aad)
 	gctr_ct64(ctx, dst, &s, ciphertext, &h, &j0, false)
@@ -112,7 +113,7 @@ open_gcm :: proc(ctx: ^Context_GCM, dst, nonce, aad, ciphertext, tag: []byte) ->
 	return ok
 }
 
-// reset_ctr sanitizes the Context_GCM.  The Context_GCM must be
+// reset_gcm sanitizes the Context_GCM.  The Context_GCM must be
 // re-initialized to be used again.
 reset_gcm :: proc "contextless" (ctx: ^Context_GCM) {
 	reset_impl(&ctx._impl)
@@ -120,14 +121,14 @@ reset_gcm :: proc "contextless" (ctx: ^Context_GCM) {
 }
 
 @(private = "file")
-gcm_validate_common_slice_sizes :: proc(tag, nonce, aad, text: []byte) {
+gcm_validate_common_slice_sizes :: proc(tag, iv, aad, text: []byte) {
 	if len(tag) != GCM_TAG_SIZE {
 		panic("crypto/aes: invalid GCM tag size")
 	}
 
-	// The specification supports nonces in the range [1, 2^64) bits.
-	if l := len(nonce); l == 0 || u64(l) >= GCM_NONCE_SIZE_MAX {
-		panic("crypto/aes: invalid GCM nonce size")
+	// The specification supports IVs in the range [1, 2^64) bits.
+	if l := len(iv); l == 0 || u64(l) >= GCM_IV_SIZE_MAX {
+		panic("crypto/aes: invalid GCM IV size")
 	}
 
 	if aad_len := u64(len(aad)); aad_len > GCM_A_MAX {
@@ -144,7 +145,7 @@ init_ghash_ct64 :: proc(
 	h: ^[_aes.GHASH_KEY_SIZE]byte,
 	j0: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	j0_enc: ^[_aes.GHASH_BLOCK_SIZE]byte,
-	nonce: []byte,
+	iv: []byte,
 ) {
 	impl := &ctx._impl.(ct64.Context)
 
@@ -152,14 +153,14 @@ init_ghash_ct64 :: proc(
 	ct64.encrypt_block(impl, h[:], h[:])
 
 	// Define a block, J0, as follows:
-	if l := len(nonce); l == GCM_NONCE_SIZE {
+	if l := len(iv); l == GCM_IV_SIZE {
 		// if len(IV) = 96, then let J0 = IV || 0^31 || 1
-		copy(j0[:], nonce)
+		copy(j0[:], iv)
 		j0[_aes.GHASH_BLOCK_SIZE - 1] = 1
 	} else {
 		// If len(IV) != 96, then let s = 128 ceil(len(IV)/128) - len(IV),
 		// and let J0 = GHASHH(IV || 0^(s+64) || ceil(len(IV))^64).
-		ct64.ghash(j0[:], h[:], nonce)
+		ct64.ghash(j0[:], h[:], iv)
 
 		tmp: [_aes.GHASH_BLOCK_SIZE]byte
 		endian.unchecked_put_u64be(tmp[8:], u64(l) * 8)
@@ -197,7 +198,7 @@ gctr_ct64 :: proc(
 	s: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	src: []byte,
 	h: ^[_aes.GHASH_KEY_SIZE]byte,
-	nonce: ^[_aes.GHASH_BLOCK_SIZE]byte,
+	iv: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	is_seal: bool,
 ) #no_bounds_check {
 	ct64_inc_ctr32 := #force_inline proc "contextless" (dst: []byte, ctr: u32) -> u32 {
@@ -208,14 +209,14 @@ gctr_ct64 :: proc(
 	// Setup the counter blocks.
 	tmp, tmp2: [ct64.STRIDE][BLOCK_SIZE]byte = ---, ---
 	ctrs, blks: [ct64.STRIDE][]byte = ---, ---
-	ctr := endian.unchecked_get_u32be(nonce[GCM_NONCE_SIZE:]) + 1
+	ctr := endian.unchecked_get_u32be(iv[GCM_IV_SIZE:]) + 1
 	for i in 0 ..< ct64.STRIDE {
 		// Setup scratch space for the keystream.
 		blks[i] = tmp2[i][:]
 
 		// Pre-copy the IV to all the counter blocks.
 		ctrs[i] = tmp[i][:]
-		copy(ctrs[i], nonce[:GCM_NONCE_SIZE])
+		copy(ctrs[i], iv[:GCM_IV_SIZE])
 	}
 
 	impl := &ctx._impl.(ct64.Context)
diff --git a/core/crypto/aes/aes_gcm_hw_intel.odin b/core/crypto/aes/aes_gcm_hw_intel.odin
index 7d32d4d96..ffd8ed642 100644
--- a/core/crypto/aes/aes_gcm_hw_intel.odin
+++ b/core/crypto/aes/aes_gcm_hw_intel.odin
@@ -10,12 +10,12 @@ import "core:mem"
 import "core:simd/x86"
 
 @(private)
-gcm_seal_hw :: proc(ctx: ^Context_Impl_Hardware, dst, tag, nonce, aad, plaintext: []byte) {
+gcm_seal_hw :: proc(ctx: ^Context_Impl_Hardware, dst, tag, iv, aad, plaintext: []byte) {
 	h: [_aes.GHASH_KEY_SIZE]byte
 	j0: [_aes.GHASH_BLOCK_SIZE]byte
 	j0_enc: [_aes.GHASH_BLOCK_SIZE]byte
 	s: [_aes.GHASH_TAG_SIZE]byte
-	init_ghash_hw(ctx, &h, &j0, &j0_enc, nonce)
+	init_ghash_hw(ctx, &h, &j0, &j0_enc, iv)
 
 	// Note: Our GHASH implementation handles appending padding.
 	hw_intel.ghash(s[:], h[:], aad)
@@ -29,12 +29,12 @@ gcm_seal_hw :: proc(ctx: ^Context_Impl_Hardware, dst, tag, nonce, aad, plaintext
 }
 
 @(private)
-gcm_open_hw :: proc(ctx: ^Context_Impl_Hardware, dst, nonce, aad, ciphertext, tag: []byte) -> bool {
+gcm_open_hw :: proc(ctx: ^Context_Impl_Hardware, dst, iv, aad, ciphertext, tag: []byte) -> bool {
 	h: [_aes.GHASH_KEY_SIZE]byte
 	j0: [_aes.GHASH_BLOCK_SIZE]byte
 	j0_enc: [_aes.GHASH_BLOCK_SIZE]byte
 	s: [_aes.GHASH_TAG_SIZE]byte
-	init_ghash_hw(ctx, &h, &j0, &j0_enc, nonce)
+	init_ghash_hw(ctx, &h, &j0, &j0_enc, iv)
 
 	hw_intel.ghash(s[:], h[:], aad)
 	gctr_hw(ctx, dst, &s, ciphertext, &h, &j0, false)
@@ -59,20 +59,20 @@ init_ghash_hw :: proc(
 	h: ^[_aes.GHASH_KEY_SIZE]byte,
 	j0: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	j0_enc: ^[_aes.GHASH_BLOCK_SIZE]byte,
-	nonce: []byte,
+	iv: []byte,
 ) {
 	// 1. Let H = CIPH(k, 0^128)
 	encrypt_block_hw(ctx, h[:], h[:])
 
 	// Define a block, J0, as follows:
-	if l := len(nonce); l == GCM_NONCE_SIZE {
+	if l := len(iv); l == GCM_IV_SIZE {
 		// if len(IV) = 96, then let J0 = IV || 0^31 || 1
-		copy(j0[:], nonce)
+		copy(j0[:], iv)
 		j0[_aes.GHASH_BLOCK_SIZE - 1] = 1
 	} else {
 		// If len(IV) != 96, then let s = 128 ceil(len(IV)/128) - len(IV),
 		// and let J0 = GHASHH(IV || 0^(s+64) || ceil(len(IV))^64).
-		hw_intel.ghash(j0[:], h[:], nonce)
+		hw_intel.ghash(j0[:], h[:], iv)
 
 		tmp: [_aes.GHASH_BLOCK_SIZE]byte
 		endian.unchecked_put_u64be(tmp[8:], u64(l) * 8)
@@ -109,7 +109,7 @@ gctr_hw :: proc(
 	s: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	src: []byte,
 	h: ^[_aes.GHASH_KEY_SIZE]byte,
-	nonce: ^[_aes.GHASH_BLOCK_SIZE]byte,
+	iv: ^[_aes.GHASH_BLOCK_SIZE]byte,
 	is_seal: bool,
 ) #no_bounds_check {
 	sks: [15]x86.__m128i = ---
@@ -118,8 +118,8 @@ gctr_hw :: proc(
 	}
 
 	// Setup the counter block
-	ctr_blk := intrinsics.unaligned_load((^x86.__m128i)(nonce))
-	ctr := endian.unchecked_get_u32be(nonce[GCM_NONCE_SIZE:]) + 1
+	ctr_blk := intrinsics.unaligned_load((^x86.__m128i)(iv))
+	ctr := endian.unchecked_get_u32be(iv[GCM_IV_SIZE:]) + 1
 
 	src, dst := src, dst
 
diff --git a/core/crypto/aes/aes_impl.odin b/core/crypto/aes/aes_impl.odin
index 03747f1fb..f26874809 100644
--- a/core/crypto/aes/aes_impl.odin
+++ b/core/crypto/aes/aes_impl.odin
@@ -10,6 +10,10 @@ Context_Impl :: union {
 	Context_Impl_Hardware,
 }
 
+// DEFAULT_IMPLEMENTATION is the implementation that will be used by
+// default if possible.
+DEFAULT_IMPLEMENTATION :: Implementation.Hardware
+
 // Implementation is an AES implementation.  Most callers will not need
 // to use this as the package will automatically select the most performant
 // implementation available (See `is_hardware_accelerated()`).
diff --git a/core/crypto/aes/aes_impl_hw_gen.odin b/core/crypto/aes/aes_impl_hw_gen.odin
index 5361c6ef0..3557b1aae 100644
--- a/core/crypto/aes/aes_impl_hw_gen.odin
+++ b/core/crypto/aes/aes_impl_hw_gen.odin
@@ -34,11 +34,11 @@ ctr_blocks_hw :: proc(ctx: ^Context_CTR, dst, src: []byte, nr_blocks: int) {
 }
 
 @(private)
-gcm_seal_hw :: proc(ctx: ^Context_Impl_Hardware, dst, tag, nonce, aad, plaintext: []byte) {
+gcm_seal_hw :: proc(ctx: ^Context_Impl_Hardware, dst, tag, iv, aad, plaintext: []byte) {
 	panic(ERR_HW_NOT_SUPPORTED)
 }
 
 @(private)
-gcm_open_hw :: proc(ctx: ^Context_Impl_Hardware, dst, nonce, aad, ciphertext, tag: []byte) -> bool {
+gcm_open_hw :: proc(ctx: ^Context_Impl_Hardware, dst, iv, aad, ciphertext, tag: []byte) -> bool {
 	panic(ERR_HW_NOT_SUPPORTED)
 }
diff --git a/core/crypto/chacha20/chacha20.odin b/core/crypto/chacha20/chacha20.odin
index 73d3e1ea2..6d1a6bfc5 100644
--- a/core/crypto/chacha20/chacha20.odin
+++ b/core/crypto/chacha20/chacha20.odin
@@ -8,119 +8,66 @@ See:
 package chacha20
 
 import "core:bytes"
-import "core:encoding/endian"
-import "core:math/bits"
+import "core:crypto/_chacha20"
 import "core:mem"
 
 // KEY_SIZE is the (X)ChaCha20 key size in bytes.
-KEY_SIZE :: 32
-// NONCE_SIZE is the ChaCha20 nonce size in bytes.
-NONCE_SIZE :: 12
-// XNONCE_SIZE is the XChaCha20 nonce size in bytes.
-XNONCE_SIZE :: 24
-
-@(private)
-_MAX_CTR_IETF :: 0xffffffff
-
-@(private)
-_BLOCK_SIZE :: 64
-@(private)
-_STATE_SIZE_U32 :: 16
-@(private)
-_ROUNDS :: 20
-
-@(private)
-_SIGMA_0: u32 : 0x61707865
-@(private)
-_SIGMA_1: u32 : 0x3320646e
-@(private)
-_SIGMA_2: u32 : 0x79622d32
-@(private)
-_SIGMA_3: u32 : 0x6b206574
+KEY_SIZE :: _chacha20.KEY_SIZE
+// IV_SIZE is the ChaCha20 IV size in bytes.
+IV_SIZE :: _chacha20.IV_SIZE
+// XIV_SIZE is the XChaCha20 IV size in bytes.
+XIV_SIZE :: _chacha20.XIV_SIZE
 
 // Context is a ChaCha20 or XChaCha20 instance.
 Context :: struct {
-	_s:              [_STATE_SIZE_U32]u32,
-	_buffer:         [_BLOCK_SIZE]byte,
-	_off:            int,
-	_is_ietf_flavor: bool,
-	_is_initialized: bool,
+	_state: _chacha20.Context,
+	_impl:  Implementation,
 }
 
 // init inititializes a Context for ChaCha20 or XChaCha20 with the provided
-// key and nonce.
-init :: proc(ctx: ^Context, key, nonce: []byte) {
+// key and iv.
+init :: proc(ctx: ^Context, key, iv: []byte, impl := DEFAULT_IMPLEMENTATION) {
 	if len(key) != KEY_SIZE {
-		panic("crypto/chacha20: invalid ChaCha20 key size")
+		panic("crypto/chacha20: invalid (X)ChaCha20 key size")
 	}
-	if n_len := len(nonce); n_len != NONCE_SIZE && n_len != XNONCE_SIZE {
-		panic("crypto/chacha20: invalid (X)ChaCha20 nonce size")
+	if l := len(iv); l != IV_SIZE && l != XIV_SIZE {
+		panic("crypto/chacha20: invalid (X)ChaCha20 IV size")
 	}
 
-	k, n := key, nonce
+	k, n := key, iv
 
-	// Derive the XChaCha20 subkey and sub-nonce via HChaCha20.
-	is_xchacha := len(nonce) == XNONCE_SIZE
+	init_impl(ctx, impl)
+
+	is_xchacha := len(iv) == XIV_SIZE
 	if is_xchacha {
-		sub_key := ctx._buffer[:KEY_SIZE]
-		_hchacha20(sub_key, k, n)
+		sub_iv: [IV_SIZE]byte
+		sub_key := ctx._state._buffer[:KEY_SIZE]
+		hchacha20(sub_key, k, n, ctx._impl)
 		k = sub_key
-		n = n[16:24]
+		copy(sub_iv[4:], n[16:])
+		n = sub_iv[:]
 	}
 
-	ctx._s[0] = _SIGMA_0
-	ctx._s[1] = _SIGMA_1
-	ctx._s[2] = _SIGMA_2
-	ctx._s[3] = _SIGMA_3
-	ctx._s[4] = endian.unchecked_get_u32le(k[0:4])
-	ctx._s[5] = endian.unchecked_get_u32le(k[4:8])
-	ctx._s[6] = endian.unchecked_get_u32le(k[8:12])
-	ctx._s[7] = endian.unchecked_get_u32le(k[12:16])
-	ctx._s[8] = endian.unchecked_get_u32le(k[16:20])
-	ctx._s[9] = endian.unchecked_get_u32le(k[20:24])
-	ctx._s[10] = endian.unchecked_get_u32le(k[24:28])
-	ctx._s[11] = endian.unchecked_get_u32le(k[28:32])
-	ctx._s[12] = 0
-	if !is_xchacha {
-		ctx._s[13] = endian.unchecked_get_u32le(n[0:4])
-		ctx._s[14] = endian.unchecked_get_u32le(n[4:8])
-		ctx._s[15] = endian.unchecked_get_u32le(n[8:12])
-	} else {
-		ctx._s[13] = 0
-		ctx._s[14] = endian.unchecked_get_u32le(n[0:4])
-		ctx._s[15] = endian.unchecked_get_u32le(n[4:8])
+	_chacha20.init(&ctx._state, k, n, is_xchacha)
 
+	if is_xchacha {
 		// The sub-key is stored in the keystream buffer.  While
 		// this will be overwritten in most circumstances, explicitly
 		// clear it out early.
-		mem.zero_explicit(&ctx._buffer, KEY_SIZE)
+		mem.zero_explicit(&ctx._state._buffer, KEY_SIZE)
 	}
-
-	ctx._off = _BLOCK_SIZE
-	ctx._is_ietf_flavor = !is_xchacha
-	ctx._is_initialized = true
 }
 
 // seek seeks the (X)ChaCha20 stream counter to the specified block.
 seek :: proc(ctx: ^Context, block_nr: u64) {
-	assert(ctx._is_initialized)
-
-	if ctx._is_ietf_flavor {
-		if block_nr > _MAX_CTR_IETF {
-			panic("crypto/chacha20: attempted to seek past maximum counter")
-		}
-	} else {
-		ctx._s[13] = u32(block_nr >> 32)
-	}
-	ctx._s[12] = u32(block_nr)
-	ctx._off = _BLOCK_SIZE
+	_chacha20.seek(&ctx._state, block_nr)
 }
 
 // xor_bytes XORs each byte in src with bytes taken from the (X)ChaCha20
 // keystream, and writes the resulting output to dst.  Dst and src MUST
 // alias exactly or not at all.
 xor_bytes :: proc(ctx: ^Context, dst, src: []byte) {
-	assert(ctx._is_initialized)
+	assert(ctx._state._is_initialized)
 
 	src, dst := src, dst
 	if dst_len := len(dst); dst_len < len(src) {
@@ -131,12 +78,13 @@ xor_bytes :: proc(ctx: ^Context, dst, src: []byte) {
 		panic("crypto/chacha20: dst and src alias inexactly")
 	}
 
-	for remaining := len(src); remaining > 0; {
+	st := &ctx._state
+	#no_bounds_check for remaining := len(src); remaining > 0; {
 		// Process multiple blocks at once
-		if ctx._off == _BLOCK_SIZE {
-			if nr_blocks := remaining / _BLOCK_SIZE; nr_blocks > 0 {
-				direct_bytes := nr_blocks * _BLOCK_SIZE
-				_do_blocks(ctx, dst, src, nr_blocks)
+		if st._off == _chacha20.BLOCK_SIZE {
+			if nr_blocks := remaining / _chacha20.BLOCK_SIZE; nr_blocks > 0 {
+				direct_bytes := nr_blocks * _chacha20.BLOCK_SIZE
+				stream_blocks(ctx, dst, src, nr_blocks)
 				remaining -= direct_bytes
 				if remaining == 0 {
 					return
@@ -147,17 +95,17 @@ xor_bytes :: proc(ctx: ^Context, dst, src: []byte) {
 
 			// If there is a partial block, generate and buffer 1 block
 			// worth of keystream.
-			_do_blocks(ctx, ctx._buffer[:], nil, 1)
-			ctx._off = 0
+			stream_blocks(ctx, st._buffer[:], nil, 1)
+			st._off = 0
 		}
 
 		// Process partial blocks from the buffered keystream.
-		to_xor := min(_BLOCK_SIZE - ctx._off, remaining)
-		buffered_keystream := ctx._buffer[ctx._off:]
+		to_xor := min(_chacha20.BLOCK_SIZE - st._off, remaining)
+		buffered_keystream := st._buffer[st._off:]
 		for i := 0; i < to_xor; i = i + 1 {
 			dst[i] = buffered_keystream[i] ~ src[i]
 		}
-		ctx._off += to_xor
+		st._off += to_xor
 		dst = dst[to_xor:]
 		src = src[to_xor:]
 		remaining -= to_xor
@@ -166,15 +114,15 @@ xor_bytes :: proc(ctx: ^Context, dst, src: []byte) {
 
 // keystream_bytes fills dst with the raw (X)ChaCha20 keystream output.
 keystream_bytes :: proc(ctx: ^Context, dst: []byte) {
-	assert(ctx._is_initialized)
+	assert(ctx._state._is_initialized)
 
-	dst := dst
-	for remaining := len(dst); remaining > 0; {
+	dst, st := dst, &ctx._state
+	#no_bounds_check for remaining := len(dst); remaining > 0; {
 		// Process multiple blocks at once
-		if ctx._off == _BLOCK_SIZE {
-			if nr_blocks := remaining / _BLOCK_SIZE; nr_blocks > 0 {
-				direct_bytes := nr_blocks * _BLOCK_SIZE
-				_do_blocks(ctx, dst, nil, nr_blocks)
+		if st._off == _chacha20.BLOCK_SIZE {
+			if nr_blocks := remaining / _chacha20.BLOCK_SIZE; nr_blocks > 0 {
+				direct_bytes := nr_blocks * _chacha20.BLOCK_SIZE
+				stream_blocks(ctx, dst, nil, nr_blocks)
 				remaining -= direct_bytes
 				if remaining == 0 {
 					return
@@ -184,15 +132,15 @@ keystream_bytes :: proc(ctx: ^Context, dst: []byte) {
 
 			// If there is a partial block, generate and buffer 1 block
 			// worth of keystream.
-			_do_blocks(ctx, ctx._buffer[:], nil, 1)
-			ctx._off = 0
+			stream_blocks(ctx, st._buffer[:], nil, 1)
+			st._off = 0
 		}
 
 		// Process partial blocks from the buffered keystream.
-		to_copy := min(_BLOCK_SIZE - ctx._off, remaining)
-		buffered_keystream := ctx._buffer[ctx._off:]
+		to_copy := min(_chacha20.BLOCK_SIZE - st._off, remaining)
+		buffered_keystream := st._buffer[st._off:]
 		copy(dst[:to_copy], buffered_keystream[:to_copy])
-		ctx._off += to_copy
+		st._off += to_copy
 		dst = dst[to_copy:]
 		remaining -= to_copy
 	}
@@ -201,366 +149,5 @@ keystream_bytes :: proc(ctx: ^Context, dst: []byte) {
 // reset sanitizes the Context.  The Context must be re-initialized to
 // be used again.
 reset :: proc(ctx: ^Context) {
-	mem.zero_explicit(&ctx._s, size_of(ctx._s))
-	mem.zero_explicit(&ctx._buffer, size_of(ctx._buffer))
-
-	ctx._is_initialized = false
-}
-
-@(private)
-_do_blocks :: proc(ctx: ^Context, dst, src: []byte, nr_blocks: int) {
-	// Enforce the maximum consumed keystream per nonce.
-	//
-	// While all modern "standard" definitions of ChaCha20 use
-	// the IETF 32-bit counter, for XChaCha20 most common
-	// implementations allow for a 64-bit counter.
-	//
-	// Honestly, the answer here is "use a MRAE primitive", but
-	// go with common practice in the case of XChaCha20.
-	if ctx._is_ietf_flavor {
-		if u64(ctx._s[12]) + u64(nr_blocks) > 0xffffffff {
-			panic("crypto/chacha20: maximum ChaCha20 keystream per nonce reached")
-		}
-	} else {
-		ctr := (u64(ctx._s[13]) << 32) | u64(ctx._s[12])
-		if _, carry := bits.add_u64(ctr, u64(nr_blocks), 0); carry != 0 {
-			panic("crypto/chacha20: maximum XChaCha20 keystream per nonce reached")
-		}
-	}
-
-	dst, src := dst, src
-	x := &ctx._s
-	for n := 0; n < nr_blocks; n = n + 1 {
-		x0, x1, x2, x3 := _SIGMA_0, _SIGMA_1, _SIGMA_2, _SIGMA_3
-		x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15 := x[4], x[5], x[6], x[7], x[8], x[9], x[10], x[11], x[12], x[13], x[14], x[15]
-
-		for i := _ROUNDS; i > 0; i = i - 2 {
-			// Even when forcing inlining manually inlining all of
-			// these is decently faster.
-
-			// quarterround(x, 0, 4, 8, 12)
-			x0 += x4
-			x12 ~= x0
-			x12 = bits.rotate_left32(x12, 16)
-			x8 += x12
-			x4 ~= x8
-			x4 = bits.rotate_left32(x4, 12)
-			x0 += x4
-			x12 ~= x0
-			x12 = bits.rotate_left32(x12, 8)
-			x8 += x12
-			x4 ~= x8
-			x4 = bits.rotate_left32(x4, 7)
-
-			// quarterround(x, 1, 5, 9, 13)
-			x1 += x5
-			x13 ~= x1
-			x13 = bits.rotate_left32(x13, 16)
-			x9 += x13
-			x5 ~= x9
-			x5 = bits.rotate_left32(x5, 12)
-			x1 += x5
-			x13 ~= x1
-			x13 = bits.rotate_left32(x13, 8)
-			x9 += x13
-			x5 ~= x9
-			x5 = bits.rotate_left32(x5, 7)
-
-			// quarterround(x, 2, 6, 10, 14)
-			x2 += x6
-			x14 ~= x2
-			x14 = bits.rotate_left32(x14, 16)
-			x10 += x14
-			x6 ~= x10
-			x6 = bits.rotate_left32(x6, 12)
-			x2 += x6
-			x14 ~= x2
-			x14 = bits.rotate_left32(x14, 8)
-			x10 += x14
-			x6 ~= x10
-			x6 = bits.rotate_left32(x6, 7)
-
-			// quarterround(x, 3, 7, 11, 15)
-			x3 += x7
-			x15 ~= x3
-			x15 = bits.rotate_left32(x15, 16)
-			x11 += x15
-			x7 ~= x11
-			x7 = bits.rotate_left32(x7, 12)
-			x3 += x7
-			x15 ~= x3
-			x15 = bits.rotate_left32(x15, 8)
-			x11 += x15
-			x7 ~= x11
-			x7 = bits.rotate_left32(x7, 7)
-
-			// quarterround(x, 0, 5, 10, 15)
-			x0 += x5
-			x15 ~= x0
-			x15 = bits.rotate_left32(x15, 16)
-			x10 += x15
-			x5 ~= x10
-			x5 = bits.rotate_left32(x5, 12)
-			x0 += x5
-			x15 ~= x0
-			x15 = bits.rotate_left32(x15, 8)
-			x10 += x15
-			x5 ~= x10
-			x5 = bits.rotate_left32(x5, 7)
-
-			// quarterround(x, 1, 6, 11, 12)
-			x1 += x6
-			x12 ~= x1
-			x12 = bits.rotate_left32(x12, 16)
-			x11 += x12
-			x6 ~= x11
-			x6 = bits.rotate_left32(x6, 12)
-			x1 += x6
-			x12 ~= x1
-			x12 = bits.rotate_left32(x12, 8)
-			x11 += x12
-			x6 ~= x11
-			x6 = bits.rotate_left32(x6, 7)
-
-			// quarterround(x, 2, 7, 8, 13)
-			x2 += x7
-			x13 ~= x2
-			x13 = bits.rotate_left32(x13, 16)
-			x8 += x13
-			x7 ~= x8
-			x7 = bits.rotate_left32(x7, 12)
-			x2 += x7
-			x13 ~= x2
-			x13 = bits.rotate_left32(x13, 8)
-			x8 += x13
-			x7 ~= x8
-			x7 = bits.rotate_left32(x7, 7)
-
-			// quarterround(x, 3, 4, 9, 14)
-			x3 += x4
-			x14 ~= x3
-			x14 = bits.rotate_left32(x14, 16)
-			x9 += x14
-			x4 ~= x9
-			x4 = bits.rotate_left32(x4, 12)
-			x3 += x4
-			x14 ~= x3
-			x14 = bits.rotate_left32(x14, 8)
-			x9 += x14
-			x4 ~= x9
-			x4 = bits.rotate_left32(x4, 7)
-		}
-
-		x0 += _SIGMA_0
-		x1 += _SIGMA_1
-		x2 += _SIGMA_2
-		x3 += _SIGMA_3
-		x4 += x[4]
-		x5 += x[5]
-		x6 += x[6]
-		x7 += x[7]
-		x8 += x[8]
-		x9 += x[9]
-		x10 += x[10]
-		x11 += x[11]
-		x12 += x[12]
-		x13 += x[13]
-		x14 += x[14]
-		x15 += x[15]
-
-		// While the "correct" answer to getting more performance out of
-		// this is "use vector operations", support for that is currently
-		// a work in progress/to be designed.
-		//
-		// In the meantime:
-		// - The caller(s) ensure that src/dst are valid.
-		// - The compiler knows if the target is picky about alignment.
-
-		#no_bounds_check {
-			if src != nil {
-				endian.unchecked_put_u32le(dst[0:4], endian.unchecked_get_u32le(src[0:4]) ~ x0)
-				endian.unchecked_put_u32le(dst[4:8], endian.unchecked_get_u32le(src[4:8]) ~ x1)
-				endian.unchecked_put_u32le(dst[8:12], endian.unchecked_get_u32le(src[8:12]) ~ x2)
-				endian.unchecked_put_u32le(dst[12:16], endian.unchecked_get_u32le(src[12:16]) ~ x3)
-				endian.unchecked_put_u32le(dst[16:20], endian.unchecked_get_u32le(src[16:20]) ~ x4)
-				endian.unchecked_put_u32le(dst[20:24], endian.unchecked_get_u32le(src[20:24]) ~ x5)
-				endian.unchecked_put_u32le(dst[24:28], endian.unchecked_get_u32le(src[24:28]) ~ x6)
-				endian.unchecked_put_u32le(dst[28:32], endian.unchecked_get_u32le(src[28:32]) ~ x7)
-				endian.unchecked_put_u32le(dst[32:36], endian.unchecked_get_u32le(src[32:36]) ~ x8)
-				endian.unchecked_put_u32le(dst[36:40], endian.unchecked_get_u32le(src[36:40]) ~ x9)
-				endian.unchecked_put_u32le(dst[40:44], endian.unchecked_get_u32le(src[40:44]) ~ x10)
-				endian.unchecked_put_u32le(dst[44:48], endian.unchecked_get_u32le(src[44:48]) ~ x11)
-				endian.unchecked_put_u32le(dst[48:52], endian.unchecked_get_u32le(src[48:52]) ~ x12)
-				endian.unchecked_put_u32le(dst[52:56], endian.unchecked_get_u32le(src[52:56]) ~ x13)
-				endian.unchecked_put_u32le(dst[56:60], endian.unchecked_get_u32le(src[56:60]) ~ x14)
-				endian.unchecked_put_u32le(dst[60:64], endian.unchecked_get_u32le(src[60:64]) ~ x15)
-				src = src[_BLOCK_SIZE:]
-			} else {
-				endian.unchecked_put_u32le(dst[0:4], x0)
-				endian.unchecked_put_u32le(dst[4:8], x1)
-				endian.unchecked_put_u32le(dst[8:12], x2)
-				endian.unchecked_put_u32le(dst[12:16], x3)
-				endian.unchecked_put_u32le(dst[16:20], x4)
-				endian.unchecked_put_u32le(dst[20:24], x5)
-				endian.unchecked_put_u32le(dst[24:28], x6)
-				endian.unchecked_put_u32le(dst[28:32], x7)
-				endian.unchecked_put_u32le(dst[32:36], x8)
-				endian.unchecked_put_u32le(dst[36:40], x9)
-				endian.unchecked_put_u32le(dst[40:44], x10)
-				endian.unchecked_put_u32le(dst[44:48], x11)
-				endian.unchecked_put_u32le(dst[48:52], x12)
-				endian.unchecked_put_u32le(dst[52:56], x13)
-				endian.unchecked_put_u32le(dst[56:60], x14)
-				endian.unchecked_put_u32le(dst[60:64], x15)
-			}
-			dst = dst[_BLOCK_SIZE:]
-		}
-
-		// Increment the counter.  Overflow checking is done upon
-		// entry into the routine, so a 64-bit increment safely
-		// covers both cases.
-		new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + 1
-		x[12] = u32(new_ctr)
-		x[13] = u32(new_ctr >> 32)
-	}
-}
-
-@(private)
-_hchacha20 :: proc "contextless" (dst, key, nonce: []byte) {
-	x0, x1, x2, x3 := _SIGMA_0, _SIGMA_1, _SIGMA_2, _SIGMA_3
-	x4 := endian.unchecked_get_u32le(key[0:4])
-	x5 := endian.unchecked_get_u32le(key[4:8])
-	x6 := endian.unchecked_get_u32le(key[8:12])
-	x7 := endian.unchecked_get_u32le(key[12:16])
-	x8 := endian.unchecked_get_u32le(key[16:20])
-	x9 := endian.unchecked_get_u32le(key[20:24])
-	x10 := endian.unchecked_get_u32le(key[24:28])
-	x11 := endian.unchecked_get_u32le(key[28:32])
-	x12 := endian.unchecked_get_u32le(nonce[0:4])
-	x13 := endian.unchecked_get_u32le(nonce[4:8])
-	x14 := endian.unchecked_get_u32le(nonce[8:12])
-	x15 := endian.unchecked_get_u32le(nonce[12:16])
-
-	for i := _ROUNDS; i > 0; i = i - 2 {
-		// quarterround(x, 0, 4, 8, 12)
-		x0 += x4
-		x12 ~= x0
-		x12 = bits.rotate_left32(x12, 16)
-		x8 += x12
-		x4 ~= x8
-		x4 = bits.rotate_left32(x4, 12)
-		x0 += x4
-		x12 ~= x0
-		x12 = bits.rotate_left32(x12, 8)
-		x8 += x12
-		x4 ~= x8
-		x4 = bits.rotate_left32(x4, 7)
-
-		// quarterround(x, 1, 5, 9, 13)
-		x1 += x5
-		x13 ~= x1
-		x13 = bits.rotate_left32(x13, 16)
-		x9 += x13
-		x5 ~= x9
-		x5 = bits.rotate_left32(x5, 12)
-		x1 += x5
-		x13 ~= x1
-		x13 = bits.rotate_left32(x13, 8)
-		x9 += x13
-		x5 ~= x9
-		x5 = bits.rotate_left32(x5, 7)
-
-		// quarterround(x, 2, 6, 10, 14)
-		x2 += x6
-		x14 ~= x2
-		x14 = bits.rotate_left32(x14, 16)
-		x10 += x14
-		x6 ~= x10
-		x6 = bits.rotate_left32(x6, 12)
-		x2 += x6
-		x14 ~= x2
-		x14 = bits.rotate_left32(x14, 8)
-		x10 += x14
-		x6 ~= x10
-		x6 = bits.rotate_left32(x6, 7)
-
-		// quarterround(x, 3, 7, 11, 15)
-		x3 += x7
-		x15 ~= x3
-		x15 = bits.rotate_left32(x15, 16)
-		x11 += x15
-		x7 ~= x11
-		x7 = bits.rotate_left32(x7, 12)
-		x3 += x7
-		x15 ~= x3
-		x15 = bits.rotate_left32(x15, 8)
-		x11 += x15
-		x7 ~= x11
-		x7 = bits.rotate_left32(x7, 7)
-
-		// quarterround(x, 0, 5, 10, 15)
-		x0 += x5
-		x15 ~= x0
-		x15 = bits.rotate_left32(x15, 16)
-		x10 += x15
-		x5 ~= x10
-		x5 = bits.rotate_left32(x5, 12)
-		x0 += x5
-		x15 ~= x0
-		x15 = bits.rotate_left32(x15, 8)
-		x10 += x15
-		x5 ~= x10
-		x5 = bits.rotate_left32(x5, 7)
-
-		// quarterround(x, 1, 6, 11, 12)
-		x1 += x6
-		x12 ~= x1
-		x12 = bits.rotate_left32(x12, 16)
-		x11 += x12
-		x6 ~= x11
-		x6 = bits.rotate_left32(x6, 12)
-		x1 += x6
-		x12 ~= x1
-		x12 = bits.rotate_left32(x12, 8)
-		x11 += x12
-		x6 ~= x11
-		x6 = bits.rotate_left32(x6, 7)
-
-		// quarterround(x, 2, 7, 8, 13)
-		x2 += x7
-		x13 ~= x2
-		x13 = bits.rotate_left32(x13, 16)
-		x8 += x13
-		x7 ~= x8
-		x7 = bits.rotate_left32(x7, 12)
-		x2 += x7
-		x13 ~= x2
-		x13 = bits.rotate_left32(x13, 8)
-		x8 += x13
-		x7 ~= x8
-		x7 = bits.rotate_left32(x7, 7)
-
-		// quarterround(x, 3, 4, 9, 14)
-		x3 += x4
-		x14 ~= x3
-		x14 = bits.rotate_left32(x14, 16)
-		x9 += x14
-		x4 ~= x9
-		x4 = bits.rotate_left32(x4, 12)
-		x3 += x4
-		x14 ~= x3
-		x14 = bits.rotate_left32(x14, 8)
-		x9 += x14
-		x4 ~= x9
-		x4 = bits.rotate_left32(x4, 7)
-	}
-
-	endian.unchecked_put_u32le(dst[0:4], x0)
-	endian.unchecked_put_u32le(dst[4:8], x1)
-	endian.unchecked_put_u32le(dst[8:12], x2)
-	endian.unchecked_put_u32le(dst[12:16], x3)
-	endian.unchecked_put_u32le(dst[16:20], x12)
-	endian.unchecked_put_u32le(dst[20:24], x13)
-	endian.unchecked_put_u32le(dst[24:28], x14)
-	endian.unchecked_put_u32le(dst[28:32], x15)
+	_chacha20.reset(&ctx._state)
 }
diff --git a/core/crypto/chacha20/chacha20_impl.odin b/core/crypto/chacha20/chacha20_impl.odin
new file mode 100644
index 000000000..be2ee06b4
--- /dev/null
+++ b/core/crypto/chacha20/chacha20_impl.odin
@@ -0,0 +1,56 @@
+package chacha20
+
+import "base:intrinsics"
+import "core:crypto/_chacha20/ref"
+import "core:crypto/_chacha20/simd128"
+import "core:crypto/_chacha20/simd256"
+
+// DEFAULT_IMPLEMENTATION is the implementation that will be used by
+// default if possible.
+DEFAULT_IMPLEMENTATION :: Implementation.Simd256
+
+// Implementation is a ChaCha20 implementation.  Most callers will not need
+// to use this as the package will automatically select the most performant
+// implementation available.
+Implementation :: enum {
+	Portable,
+	Simd128,
+	Simd256,
+}
+
+@(private)
+init_impl :: proc(ctx: ^Context, impl: Implementation) {
+	impl := impl
+	if impl == .Simd256 && !simd256.is_performant() {
+			impl = .Simd128
+	}
+	if impl == .Simd128 && !simd128.is_performant() {
+		impl = .Portable
+	}
+
+	ctx._impl = impl
+}
+
+@(private)
+stream_blocks :: proc(ctx: ^Context, dst, src: []byte, nr_blocks: int) {
+	switch ctx._impl {
+	case .Simd256:
+		simd256.stream_blocks(&ctx._state, dst, src, nr_blocks)
+	case .Simd128:
+		simd128.stream_blocks(&ctx._state, dst, src, nr_blocks)
+	case .Portable:
+		ref.stream_blocks(&ctx._state, dst, src, nr_blocks)
+	}
+}
+
+@(private)
+hchacha20 :: proc "contextless" (dst, key, iv: []byte, impl: Implementation) {
+	switch impl {
+	case .Simd256:
+		simd256.hchacha20(dst, key, iv)
+	case .Simd128:
+		simd128.hchacha20(dst, key, iv)
+	case .Portable:
+		ref.hchacha20(dst, key, iv)
+	}
+}
diff --git a/core/crypto/chacha20poly1305/chacha20poly1305.odin b/core/crypto/chacha20poly1305/chacha20poly1305.odin
index 7fc112d0d..e2cd35a7e 100644
--- a/core/crypto/chacha20poly1305/chacha20poly1305.odin
+++ b/core/crypto/chacha20poly1305/chacha20poly1305.odin
@@ -1,9 +1,11 @@
 /*
-package chacha20poly1305 implements the AEAD_CHACHA20_POLY1305 Authenticated
-Encryption with Additional Data algorithm.
+package chacha20poly1305 implements the AEAD_CHACHA20_POLY1305 and
+AEAD_XChaCha20_Poly1305 Authenticated Encryption with Additional Data
+algorithms.
 
 See:
 - https://www.rfc-editor.org/rfc/rfc8439
+- https://datatracker.ietf.org/doc/html/draft-arciszewski-xchacha-03
 */
 package chacha20poly1305
 
@@ -15,8 +17,10 @@ import "core:mem"
 
 // KEY_SIZE is the chacha20poly1305 key size in bytes.
 KEY_SIZE :: chacha20.KEY_SIZE
-// NONCE_SIZE is the chacha20poly1305 nonce size in bytes.
-NONCE_SIZE :: chacha20.NONCE_SIZE
+// IV_SIZE is the chacha20poly1305 IV size in bytes.
+IV_SIZE :: chacha20.IV_SIZE
+// XIV_SIZE is the xchacha20poly1305 IV size in bytes.
+XIV_SIZE :: chacha20.XIV_SIZE
 // TAG_SIZE is the chacha20poly1305 tag size in bytes.
 TAG_SIZE :: poly1305.TAG_SIZE
 
@@ -24,15 +28,13 @@ TAG_SIZE :: poly1305.TAG_SIZE
 _P_MAX :: 64 * 0xffffffff // 64 * (2^32-1)
 
 @(private)
-_validate_common_slice_sizes :: proc (tag, key, nonce, aad, text: []byte) {
+_validate_common_slice_sizes :: proc (tag, iv, aad, text: []byte, is_xchacha: bool) {
 	if len(tag) != TAG_SIZE {
 		panic("crypto/chacha20poly1305: invalid destination tag size")
 	}
-	if len(key) != KEY_SIZE {
-		panic("crypto/chacha20poly1305: invalid key size")
-	}
-	if len(nonce) != NONCE_SIZE {
-		panic("crypto/chacha20poly1305: invalid nonce size")
+	expected_iv_len := is_xchacha ? XIV_SIZE : IV_SIZE
+	if len(iv) != expected_iv_len {
+		panic("crypto/chacha20poly1305: invalid IV size")
 	}
 
 	#assert(size_of(int) == 8 || size_of(int) <= 4)
@@ -59,18 +61,52 @@ _update_mac_pad16 :: #force_inline proc (ctx: ^poly1305.Context, x_len: int) {
 	}
 }
 
-// encrypt encrypts the plaintext and authenticates the aad and ciphertext,
-// with the provided key and nonce, stores the output in ciphertext and tag.
-encrypt :: proc (ciphertext, tag, key, nonce, aad, plaintext: []byte) {
-	_validate_common_slice_sizes(tag, key, nonce, aad, plaintext)
+// Context is a keyed (X)Chacha20Poly1305 instance.
+Context :: struct {
+	_key:            [KEY_SIZE]byte,
+	_impl:           chacha20.Implementation,
+	_is_xchacha:     bool,
+	_is_initialized: bool,
+}
+
+// init initializes a Context with the provided key, for AEAD_CHACHA20_POLY1305.
+init :: proc(ctx: ^Context, key: []byte, impl := chacha20.DEFAULT_IMPLEMENTATION) {
+	if len(key) != KEY_SIZE {
+		panic("crypto/chacha20poly1305: invalid key size")
+	}
+
+	copy(ctx._key[:], key)
+	ctx._impl = impl
+	ctx._is_xchacha = false
+	ctx._is_initialized = true
+}
+
+// init_xchacha initializes a Context with the provided key, for
+// AEAD_XChaCha20_Poly1305.
+//
+// Note: While there are multiple definitions of XChaCha20-Poly1305
+// this sticks to the IETF draft and uses a 32-bit counter.
+init_xchacha :: proc(ctx: ^Context, key: []byte, impl := chacha20.DEFAULT_IMPLEMENTATION) {
+	init(ctx, key, impl)
+	ctx._is_xchacha = true
+}
+
+// seal encrypts the plaintext and authenticates the aad and ciphertext,
+// with the provided Context and iv, stores the output in dst and tag.
+//
+// dst and plaintext MUST alias exactly or not at all.
+seal :: proc(ctx: ^Context, dst, tag, iv, aad, plaintext: []byte) {
+	ciphertext := dst
+	_validate_common_slice_sizes(tag, iv, aad, plaintext, ctx._is_xchacha)
 	if len(ciphertext) != len(plaintext) {
 		panic("crypto/chacha20poly1305: invalid destination ciphertext size")
 	}
 
 	stream_ctx: chacha20.Context = ---
-	chacha20.init(&stream_ctx, key, nonce)
+	chacha20.init(&stream_ctx, ctx._key[:],iv, ctx._impl)
+	stream_ctx._state._is_ietf_flavor = true
 
-	// otk = poly1305_key_gen(key, nonce)
+	// otk = poly1305_key_gen(key, iv)
 	otk: [poly1305.KEY_SIZE]byte = ---
 	chacha20.keystream_bytes(&stream_ctx, otk[:])
 	mac_ctx: poly1305.Context = ---
@@ -87,7 +123,7 @@ encrypt :: proc (ciphertext, tag, key, nonce, aad, plaintext: []byte) {
 	poly1305.update(&mac_ctx, aad)
 	_update_mac_pad16(&mac_ctx, aad_len)
 
-	// ciphertext = chacha20_encrypt(key, 1, nonce, plaintext)
+	// ciphertext = chacha20_encrypt(key, 1, iv, plaintext)
 	chacha20.seek(&stream_ctx, 1)
 	chacha20.xor_bytes(&stream_ctx, ciphertext, plaintext)
 	chacha20.reset(&stream_ctx) // Don't need the stream context anymore.
@@ -107,13 +143,16 @@ encrypt :: proc (ciphertext, tag, key, nonce, aad, plaintext: []byte) {
 	poly1305.final(&mac_ctx, tag) // Implicitly sanitizes context.
 }
 
-// decrypt authenticates the aad and ciphertext, and decrypts the ciphertext,
-// with the provided key, nonce, and tag, and stores the output in plaintext,
-// returning true iff the authentication was successful.
+// open authenticates the aad and ciphertext, and decrypts the ciphertext,
+// with the provided Context, iv, and tag, and stores the output in dst,
+// returning true iff the authentication was successful.  If authentication
+// fails, the destination buffer will be zeroed.
 //
-// If authentication fails, the destination plaintext buffer will be zeroed.
-decrypt :: proc (plaintext, tag, key, nonce, aad, ciphertext: []byte) -> bool {
-	_validate_common_slice_sizes(tag, key, nonce, aad, ciphertext)
+// dst and plaintext MUST alias exactly or not at all.
+@(require_results)
+open :: proc(ctx: ^Context, dst, iv, aad, ciphertext, tag: []byte) -> bool {
+	plaintext := dst
+	_validate_common_slice_sizes(tag, iv, aad, ciphertext, ctx._is_xchacha)
 	if len(ciphertext) != len(plaintext) {
 		panic("crypto/chacha20poly1305: invalid destination plaintext size")
 	}
@@ -123,9 +162,10 @@ decrypt :: proc (plaintext, tag, key, nonce, aad, ciphertext: []byte) -> bool {
 	// points where needed.
 
 	stream_ctx: chacha20.Context = ---
-	chacha20.init(&stream_ctx, key, nonce)
+	chacha20.init(&stream_ctx, ctx._key[:], iv, ctx._impl)
+	stream_ctx._state._is_ietf_flavor = true
 
-	// otk = poly1305_key_gen(key, nonce)
+	// otk = poly1305_key_gen(key, iv)
 	otk: [poly1305.KEY_SIZE]byte = ---
 	chacha20.keystream_bytes(&stream_ctx, otk[:])
 	defer chacha20.reset(&stream_ctx)
@@ -160,9 +200,17 @@ decrypt :: proc (plaintext, tag, key, nonce, aad, ciphertext: []byte) -> bool {
 		return false
 	}
 
-	// plaintext = chacha20_decrypt(key, 1, nonce, ciphertext)
+	// plaintext = chacha20_decrypt(key, 1, iv, ciphertext)
 	chacha20.seek(&stream_ctx, 1)
 	chacha20.xor_bytes(&stream_ctx, plaintext, ciphertext)
 
 	return true
 }
+
+// reset sanitizes the Context.  The Context must be
+// re-initialized to be used again.
+reset :: proc "contextless" (ctx: ^Context) {
+	mem.zero_explicit(&ctx._key, len(ctx._key))
+	ctx._is_xchacha = false
+	ctx._is_initialized = false
+}
diff --git a/core/crypto/ed25519/ed25519.odin b/core/crypto/ed25519/ed25519.odin
index 86da35669..5584b06f7 100644
--- a/core/crypto/ed25519/ed25519.odin
+++ b/core/crypto/ed25519/ed25519.odin
@@ -21,7 +21,7 @@ PUBLIC_KEY_SIZE :: 32
 SIGNATURE_SIZE :: 64
 
 @(private)
-NONCE_SIZE :: 32
+HDIGEST2_SIZE :: 32
 
 // Private_Key is an Ed25519 private key.
 Private_Key :: struct {
@@ -33,7 +33,7 @@ Private_Key :: struct {
 	// See: https://github.com/MystenLabs/ed25519-unsafe-libs
 	_b:              [PRIVATE_KEY_SIZE]byte,
 	_s:              grp.Scalar,
-	_nonce:          [NONCE_SIZE]byte,
+	_hdigest2:       [HDIGEST2_SIZE]byte,
 	_pub_key:        Public_Key,
 	_is_initialized: bool,
 }
@@ -63,7 +63,7 @@ private_key_set_bytes :: proc(priv_key: ^Private_Key, b: []byte) -> bool {
 	sha2.final(&ctx, h_bytes[:])
 
 	copy(priv_key._b[:], b)
-	copy(priv_key._nonce[:], h_bytes[32:])
+	copy(priv_key._hdigest2[:], h_bytes[32:])
 	grp.sc_set_bytes_rfc8032(&priv_key._s, h_bytes[:32])
 
 	// Derive the corresponding public key.
@@ -116,7 +116,7 @@ sign :: proc(priv_key: ^Private_Key, msg, sig: []byte) {
 	ctx: sha2.Context_512 = ---
 	digest_bytes: [sha2.DIGEST_SIZE_512]byte = ---
 	sha2.init_512(&ctx)
-	sha2.update(&ctx, priv_key._nonce[:])
+	sha2.update(&ctx, priv_key._hdigest2[:])
 	sha2.update(&ctx, msg)
 	sha2.final(&ctx, digest_bytes[:])
 
diff --git a/core/crypto/hash/hash.odin b/core/crypto/hash/hash.odin
index e4b3d4be1..f7671270a 100644
--- a/core/crypto/hash/hash.odin
+++ b/core/crypto/hash/hash.odin
@@ -28,20 +28,26 @@ hash_bytes :: proc(algorithm: Algorithm, data: []byte, allocator := context.allo
 
 // hash_string_to_buffer will hash the given input and assign the
 // computed digest to the third parameter.  It requires that the
-// destination buffer is at least as big as the digest size.
-hash_string_to_buffer :: proc(algorithm: Algorithm, data: string, hash: []byte) {
-	hash_bytes_to_buffer(algorithm, transmute([]byte)(data), hash)
+// destination buffer is at least as big as the digest size.  The
+// provided destination buffer is returned to match the behavior of
+// `hash_string`.
+hash_string_to_buffer :: proc(algorithm: Algorithm, data: string, hash: []byte) -> []byte {
+	return hash_bytes_to_buffer(algorithm, transmute([]byte)(data), hash)
 }
 
 // hash_bytes_to_buffer will hash the given input and write the
 // computed digest into the third parameter.  It requires that the
-// destination buffer is at least as big as the digest size.
-hash_bytes_to_buffer :: proc(algorithm: Algorithm, data, hash: []byte) {
+// destination buffer is at least as big as the digest size.  The
+// provided destination buffer is returned to match the behavior of
+// `hash_bytes`.
+hash_bytes_to_buffer :: proc(algorithm: Algorithm, data, hash: []byte) -> []byte {
 	ctx: Context
 
 	init(&ctx, algorithm)
 	update(&ctx, data)
 	final(&ctx, hash)
+
+	return hash
 }
 
 // hash_stream will incrementally fully consume a stream, and return the
diff --git a/examples/all/all_main.odin b/examples/all/all_main.odin
index d92a6b8c4..62b891352 100644
--- a/examples/all/all_main.odin
+++ b/examples/all/all_main.odin
@@ -25,6 +25,7 @@ import rbtree           "core:container/rbtree"
 import topological_sort "core:container/topological_sort"
 
 import crypto           "core:crypto"
+import aead             "core:crypto/aead"
 import aes              "core:crypto/aes"
 import blake2b          "core:crypto/blake2b"
 import blake2s          "core:crypto/blake2s"
@@ -164,6 +165,7 @@ _ :: rbtree
 _ :: topological_sort
 _ :: crypto
 _ :: crypto_hash
+_ :: aead
 _ :: aes
 _ :: blake2b
 _ :: blake2s
diff --git a/tests/benchmark/crypto/benchmark_crypto.odin b/tests/benchmark/crypto/benchmark_crypto.odin
index b2ac4bca3..66c9f89d3 100644
--- a/tests/benchmark/crypto/benchmark_crypto.odin
+++ b/tests/benchmark/crypto/benchmark_crypto.odin
@@ -279,13 +279,13 @@ _benchmark_chacha20 :: proc(
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 	}
-	nonce := [chacha20.NONCE_SIZE]byte {
+	iv := [chacha20.IV_SIZE]byte {
 		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 		0x00, 0x00, 0x00, 0x00,
 	}
 
 	ctx: chacha20.Context = ---
-	chacha20.init(&ctx, key[:], nonce[:])
+	chacha20.init(&ctx, key[:], iv[:])
 
 	for _ in 0 ..= options.rounds {
 		chacha20.xor_bytes(&ctx, buf, buf)
@@ -334,15 +334,18 @@ _benchmark_chacha20poly1305 :: proc(
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 	}
-	nonce := [chacha20.NONCE_SIZE]byte {
+	iv := [chacha20.IV_SIZE]byte {
 		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 		0x00, 0x00, 0x00, 0x00,
 	}
 
+	ctx: chacha20poly1305.Context = ---
+	chacha20poly1305.init(&ctx, key[:]) // Basically 0 overhead.
+
 	tag: [chacha20poly1305.TAG_SIZE]byte = ---
 
 	for _ in 0 ..= options.rounds {
-		chacha20poly1305.encrypt(buf, tag[:], key[:], nonce[:], nil, buf)
+		chacha20poly1305.seal(&ctx, buf, tag[:], iv[:], nil, buf)
 	}
 	options.count = options.rounds
 	options.processed = options.rounds * options.bytes
@@ -363,13 +366,13 @@ _benchmark_aes256_ctr :: proc(
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 		0xde, 0xad, 0xbe, 0xef, 0xde, 0xad, 0xbe, 0xef,
 	}
-	nonce := [aes.CTR_IV_SIZE]byte {
+	iv := [aes.CTR_IV_SIZE]byte {
 		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
 	}
 
 	ctx: aes.Context_CTR = ---
-	aes.init_ctr(&ctx, key[:], nonce[:])
+	aes.init_ctr(&ctx, key[:], iv[:])
 
 	for _ in 0 ..= options.rounds {
 		aes.xor_bytes_ctr(&ctx, buf, buf)
@@ -386,13 +389,13 @@ _benchmark_aes256_gcm :: proc(
 	err: time.Benchmark_Error,
 ) {
 	buf := options.input
-	nonce: [aes.GCM_NONCE_SIZE]byte
+	iv: [aes.GCM_IV_SIZE]byte
 	tag: [aes.GCM_TAG_SIZE]byte = ---
 
 	ctx := transmute(^aes.Context_GCM)context.user_ptr
 
 	for _ in 0 ..= options.rounds {
-		aes.seal_gcm(ctx, buf, tag[:], nonce[:], nil, buf)
+		aes.seal_gcm(ctx, buf, tag[:], iv[:], nil, buf)
 	}
 	options.count = options.rounds
 	options.processed = options.rounds * options.bytes
diff --git a/tests/core/crypto/test_core_crypto.odin b/tests/core/crypto/test_core_crypto.odin
index f3f76646b..b3eb6e041 100644
--- a/tests/core/crypto/test_core_crypto.odin
+++ b/tests/core/crypto/test_core_crypto.odin
@@ -19,15 +19,39 @@ import "base:runtime"
 import "core:log"
 
 import "core:crypto"
+import chacha_simd128 "core:crypto/_chacha20/simd128"
+import chacha_simd256 "core:crypto/_chacha20/simd256"
 import "core:crypto/chacha20"
-import "core:crypto/chacha20poly1305"
+import "core:crypto/sha2"
 
+@(private)
 _PLAINTEXT_SUNSCREEN_STR := "Ladies and Gentlemen of the class of '99: If I could offer you only one tip for the future, sunscreen would be it."
 
 @(test)
 test_chacha20 :: proc(t: ^testing.T) {
 	runtime.DEFAULT_TEMP_ALLOCATOR_TEMP_GUARD()
 
+	impls := supported_chacha_impls()
+
+	for impl in impls {
+		test_chacha20_stream(t, impl)
+	}
+}
+
+supported_chacha_impls :: proc() -> [dynamic]chacha20.Implementation {
+	impls := make([dynamic]chacha20.Implementation, 0, 3, context.temp_allocator)
+	append(&impls, chacha20.Implementation.Portable)
+	if chacha_simd128.is_performant() {
+		append(&impls, chacha20.Implementation.Simd128)
+	}
+	if chacha_simd256.is_performant() {
+		append(&impls, chacha20.Implementation.Simd256)
+	}
+
+	return impls
+}
+
+test_chacha20_stream :: proc(t: ^testing.T, impl: chacha20.Implementation) {
 	// Test cases taken from RFC 8439, and draft-irtf-cfrg-xchacha-03
 	plaintext := transmute([]byte)(_PLAINTEXT_SUNSCREEN_STR)
 
@@ -38,7 +62,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 		0x18, 0x19, 0x1a, 0x1b, 0x1c, 0x1d, 0x1e, 0x1f,
 	}
 
-	nonce := [chacha20.NONCE_SIZE]byte {
+	iv := [chacha20.IV_SIZE]byte {
 		0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x4a,
 		0x00, 0x00, 0x00, 0x00,
 	}
@@ -64,7 +88,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 
 	derived_ciphertext: [114]byte
 	ctx: chacha20.Context = ---
-	chacha20.init(&ctx, key[:], nonce[:])
+	chacha20.init(&ctx, key[:], iv[:], impl)
 	chacha20.seek(&ctx, 1) // The test vectors start the counter at 1.
 	chacha20.xor_bytes(&ctx, derived_ciphertext[:], plaintext[:])
 
@@ -72,7 +96,8 @@ test_chacha20 :: proc(t: ^testing.T) {
 	testing.expectf(
 		t,
 		derived_ciphertext_str == ciphertext_str,
-		"Expected %s for xor_bytes(plaintext_str), but got %s instead",
+		"chacha20/%v: Expected %s for xor_bytes(plaintext_str), but got %s instead",
+		impl,
 		ciphertext_str,
 		derived_ciphertext_str,
 	)
@@ -84,7 +109,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 		0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f,
 	}
 
-	xnonce := [chacha20.XNONCE_SIZE]byte {
+	xiv := [chacha20.XIV_SIZE]byte {
 		0x40, 0x41, 0x42, 0x43, 0x44, 0x45, 0x46, 0x47,
 		0x48, 0x49, 0x4a, 0x4b, 0x4c, 0x4d, 0x4e, 0x4f,
 		0x50, 0x51, 0x52, 0x53, 0x54, 0x55, 0x56, 0x57,
@@ -109,7 +134,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 	}
 	xciphertext_str := string(hex.encode(xciphertext[:], context.temp_allocator))
 
-	chacha20.init(&ctx, xkey[:], xnonce[:])
+	chacha20.init(&ctx, xkey[:], xiv[:], impl)
 	chacha20.seek(&ctx, 1)
 	chacha20.xor_bytes(&ctx, derived_ciphertext[:], plaintext[:])
 
@@ -117,128 +142,44 @@ test_chacha20 :: proc(t: ^testing.T) {
 	testing.expectf(
 		t,
 		derived_ciphertext_str == xciphertext_str,
-		"Expected %s for xor_bytes(plaintext_str), but got %s instead",
+		"chacha20/%v: Expected %s for xor_bytes(plaintext_str), but got %s instead",
+		impl,
 		xciphertext_str,
 		derived_ciphertext_str,
 	)
-}
 
-@(test)
-test_chacha20poly1305 :: proc(t: ^testing.T) {
-	plaintext := transmute([]byte)(_PLAINTEXT_SUNSCREEN_STR)
-
-	aad := [12]byte {
-		0x50, 0x51, 0x52, 0x53, 0xc0, 0xc1, 0xc2, 0xc3,
-		0xc4, 0xc5, 0xc6, 0xc7,
-	}
+	// Incrementally read 1, 2, 3, ..., 2048 bytes of keystream, and
+	// compare the SHA-512/256 digest with a known value.  Results
+	// and testcase taken from a known good implementation by the
+	// same author as the Odin test case.
 
-	key := [chacha20poly1305.KEY_SIZE]byte {
-		0x80, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
-		0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
-		0x90, 0x91, 0x92, 0x93, 0x94, 0x95, 0x96, 0x97,
-		0x98, 0x99, 0x9a, 0x9b, 0x9c, 0x9d, 0x9e, 0x9f,
-	}
+	tmp := make([]byte, 2048, context.temp_allocator)
 
-	nonce := [chacha20poly1305.NONCE_SIZE]byte {
-		0x07, 0x00, 0x00, 0x00, 0x40, 0x41, 0x42, 0x43,
-		0x44, 0x45, 0x46, 0x47,
-	}
+	mem.zero(&key, size_of(key))
+	mem.zero(&iv, size_of(iv))
+	chacha20.init(&ctx, key[:], iv[:], impl)
 
-	ciphertext := [114]byte {
-		0xd3, 0x1a, 0x8d, 0x34, 0x64, 0x8e, 0x60, 0xdb,
-		0x7b, 0x86, 0xaf, 0xbc, 0x53, 0xef, 0x7e, 0xc2,
-		0xa4, 0xad, 0xed, 0x51, 0x29, 0x6e, 0x08, 0xfe,
-		0xa9, 0xe2, 0xb5, 0xa7, 0x36, 0xee, 0x62, 0xd6,
-		0x3d, 0xbe, 0xa4, 0x5e, 0x8c, 0xa9, 0x67, 0x12,
-		0x82, 0xfa, 0xfb, 0x69, 0xda, 0x92, 0x72, 0x8b,
-		0x1a, 0x71, 0xde, 0x0a, 0x9e, 0x06, 0x0b, 0x29,
-		0x05, 0xd6, 0xa5, 0xb6, 0x7e, 0xcd, 0x3b, 0x36,
-		0x92, 0xdd, 0xbd, 0x7f, 0x2d, 0x77, 0x8b, 0x8c,
-		0x98, 0x03, 0xae, 0xe3, 0x28, 0x09, 0x1b, 0x58,
-		0xfa, 0xb3, 0x24, 0xe4, 0xfa, 0xd6, 0x75, 0x94,
-		0x55, 0x85, 0x80, 0x8b, 0x48, 0x31, 0xd7, 0xbc,
-		0x3f, 0xf4, 0xde, 0xf0, 0x8e, 0x4b, 0x7a, 0x9d,
-		0xe5, 0x76, 0xd2, 0x65, 0x86, 0xce, 0xc6, 0x4b,
-		0x61, 0x16,
-	}
-	ciphertext_str := string(hex.encode(ciphertext[:], context.temp_allocator))
+	h_ctx: sha2.Context_512
+	sha2.init_512_256(&h_ctx)
 
-	tag := [chacha20poly1305.TAG_SIZE]byte {
-		0x1a, 0xe1, 0x0b, 0x59, 0x4f, 0x09, 0xe2, 0x6a,
-		0x7e, 0x90, 0x2e, 0xcb, 0xd0, 0x60, 0x06, 0x91,
+	for i := 1; i <= 2048; i = i + 1 {
+		chacha20.keystream_bytes(&ctx, tmp[:i])
+		sha2.update(&h_ctx, tmp[:i])
 	}
-	tag_str := string(hex.encode(tag[:], context.temp_allocator))
-
-	derived_tag: [chacha20poly1305.TAG_SIZE]byte
-	derived_ciphertext: [114]byte
 
-	chacha20poly1305.encrypt(
-		derived_ciphertext[:],
-		derived_tag[:],
-		key[:],
-		nonce[:],
-		aad[:],
-		plaintext,
-	)
+	digest: [32]byte
+	sha2.final(&h_ctx, digest[:])
+	digest_str := string(hex.encode(digest[:], context.temp_allocator))
 
-	derived_ciphertext_str := string(hex.encode(derived_ciphertext[:], context.temp_allocator))
+	expected_digest_str := "cfd6e949225b854fe04946491e6935ff05ff983d1554bc885bca0ec8082dd5b8"
 	testing.expectf(
 		t,
-		derived_ciphertext_str == ciphertext_str,
-		"Expected ciphertext %s for encrypt(aad, plaintext), but got %s instead",
-		ciphertext_str,
-		derived_ciphertext_str,
-	)
-
-	derived_tag_str := string(hex.encode(derived_tag[:], context.temp_allocator))
-	testing.expectf(
-		t,
-		derived_tag_str == tag_str,
-		"Expected tag %s for encrypt(aad, plaintext), but got %s instead",
-		tag_str,
-		derived_tag_str,
-	)
-
-	derived_plaintext: [114]byte
-	ok := chacha20poly1305.decrypt(
-		derived_plaintext[:],
-		tag[:],
-		key[:],
-		nonce[:],
-		aad[:],
-		ciphertext[:],
-	)
-	derived_plaintext_str := string(derived_plaintext[:])
-	testing.expect(t, ok, "Expected true for decrypt(tag, aad, ciphertext)")
-	testing.expectf(
-		t,
-		derived_plaintext_str == _PLAINTEXT_SUNSCREEN_STR,
-		"Expected plaintext %s for decrypt(tag, aad, ciphertext), but got %s instead",
-		_PLAINTEXT_SUNSCREEN_STR,
-		derived_plaintext_str,
-	)
-
-	derived_ciphertext[0] ~= 0xa5
-	ok = chacha20poly1305.decrypt(
-		derived_plaintext[:],
-		tag[:],
-		key[:],
-		nonce[:],
-		aad[:],
-		derived_ciphertext[:],
-	)
-	testing.expect(t, !ok, "Expected false for decrypt(tag, aad, corrupted_ciphertext)")
-
-	aad[0] ~= 0xa5
-	ok = chacha20poly1305.decrypt(
-		derived_plaintext[:],
-		tag[:],
-		key[:],
-		nonce[:],
-		aad[:],
-		ciphertext[:],
+		expected_digest_str == digest_str,
+		"chacha20/%v: Expected %s for keystream digest, but got %s instead",
+		impl,
+		expected_digest_str,
+		digest_str,
 	)
-	testing.expect(t, !ok, "Expected false for decrypt(tag, corrupted_aad, ciphertext)")
 }
 
 @(test)
diff --git a/tests/core/crypto/test_core_crypto_aead.odin b/tests/core/crypto/test_core_crypto_aead.odin
new file mode 100644
index 000000000..90eedc0b2
--- /dev/null
+++ b/tests/core/crypto/test_core_crypto_aead.odin
@@ -0,0 +1,339 @@
+package test_core_crypto
+
+import "base:runtime"
+import "core:crypto/aead"
+import "core:encoding/hex"
+import "core:testing"
+
+@(test)
+test_aead :: proc(t: ^testing.T) {
+	runtime.DEFAULT_TEMP_ALLOCATOR_TEMP_GUARD()
+
+	aes_impls := make([dynamic]aead.Implementation, context.temp_allocator)
+	for impl in supported_aes_impls() {
+		append(&aes_impls, impl)
+	}
+	chacha_impls := make([dynamic]aead.Implementation, context.temp_allocator)
+	for impl in supported_chacha_impls() {
+		append(&chacha_impls, impl)
+	}
+	impls := [aead.Algorithm][dynamic]aead.Implementation{
+		.Invalid           = nil,
+		.AES_GCM_128       = aes_impls,
+		.AES_GCM_192       = aes_impls,
+		.AES_GCM_256       = aes_impls,
+		.CHACHA20POLY1305  = chacha_impls,
+		.XCHACHA20POLY1305 = chacha_impls,
+	}
+
+	test_vectors := []struct{
+		algo: aead.Algorithm,
+		key: string,
+		iv: string,
+		aad: string,
+		plaintext: string,
+		ciphertext: string,
+		tag: string,
+	} {
+		// AES-GCM
+		// - https://csrc.nist.rip/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-revised-spec.pdf
+		//
+		// Note: NIST did a reorg of their site, so the source of the test vectors
+		// is only available from an archive.
+		{
+			.AES_GCM_128,
+			"00000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"",
+			"",
+			"58e2fccefa7e3061367f1d57a4e7455a",
+		},
+		{
+			.AES_GCM_128,
+			"00000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"00000000000000000000000000000000",
+			"0388dace60b6a392f328c2b971b2fe78",
+			"ab6e47d42cec13bdf53a67b21257bddf",
+		},
+		{
+			.AES_GCM_128,
+			"feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbaddecaf888",
+			"",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
+			"42831ec2217774244b7221b784d0d49ce3aa212f2c02a4e035c17e2329aca12e21d514b25466931c7d8f6a5aac84aa051ba30b396a0aac973d58e091473f5985",
+			"4d5c2af327cd64a62cf35abd2ba6fab4",
+		},
+		{
+			.AES_GCM_128,
+			"feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbaddecaf888",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"42831ec2217774244b7221b784d0d49ce3aa212f2c02a4e035c17e2329aca12e21d514b25466931c7d8f6a5aac84aa051ba30b396a0aac973d58e091",
+			"5bc94fbc3221a5db94fae95ae7121a47",
+		},
+		{
+			.AES_GCM_128,
+			"feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbad",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"61353b4c2806934a777ff51fa22a4755699b2a714fcdc6f83766e5f97b6c742373806900e49f24b22b097544d4896b424989b5e1ebac0f07c23f4598",
+			"3612d2e79e3b0785561be14aaca2fccb",
+		},
+		{
+			.AES_GCM_128,
+			"feffe9928665731c6d6a8f9467308308",
+			"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"8ce24998625615b603a033aca13fb894be9112a5c3a211a8ba262a3cca7e2ca701e4a9a4fba43c90ccdcb281d48c7c6fd62875d2aca417034c34aee5",
+			"619cc5aefffe0bfa462af43c1699d050",
+		},
+		{
+			.AES_GCM_192,
+			"000000000000000000000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"",
+			"",
+			"cd33b28ac773f74ba00ed1f312572435",
+		},
+		{
+			.AES_GCM_192,
+			"000000000000000000000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"00000000000000000000000000000000",
+			"98e7247c07f0fe411c267e4384b0f600",
+			"2ff58d80033927ab8ef4d4587514f0fb",
+		},
+		{
+			.AES_GCM_192,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
+			"cafebabefacedbaddecaf888",
+			"",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
+			"3980ca0b3c00e841eb06fac4872a2757859e1ceaa6efd984628593b40ca1e19c7d773d00c144c525ac619d18c84a3f4718e2448b2fe324d9ccda2710acade256",
+			"9924a7c8587336bfb118024db8674a14",
+		},
+		{
+			.AES_GCM_192,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
+			"cafebabefacedbaddecaf888",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"3980ca0b3c00e841eb06fac4872a2757859e1ceaa6efd984628593b40ca1e19c7d773d00c144c525ac619d18c84a3f4718e2448b2fe324d9ccda2710",
+			"2519498e80f1478f37ba55bd6d27618c",
+		},
+		{
+			.AES_GCM_192,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
+			"cafebabefacedbad",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"0f10f599ae14a154ed24b36e25324db8c566632ef2bbb34f8347280fc4507057fddc29df9a471f75c66541d4d4dad1c9e93a19a58e8b473fa0f062f7",
+			"65dcc57fcf623a24094fcca40d3533f8",
+		},
+		{
+			.AES_GCM_192,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
+			"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"d27e88681ce3243c4830165a8fdcf9ff1de9a1d8e6b447ef6ef7b79828666e4581e79012af34ddd9e2f037589b292db3e67c036745fa22e7e9b7373b",
+			"dcf566ff291c25bbb8568fc3d376a6d9",
+		},
+		{
+			.AES_GCM_256,
+			"0000000000000000000000000000000000000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"",
+			"",
+			"530f8afbc74536b9a963b4f1c4cb738b",
+		},
+		{
+			.AES_GCM_256,
+			"0000000000000000000000000000000000000000000000000000000000000000",
+			"000000000000000000000000",
+			"",
+			"00000000000000000000000000000000",
+			"cea7403d4d606b6e074ec5d3baf39d18",
+			"d0d1c8a799996bf0265b98b5d48ab919",
+		},
+		{
+			.AES_GCM_256,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbaddecaf888",
+			"",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
+			"522dc1f099567d07f47f37a32a84427d643a8cdcbfe5c0c97598a2bd2555d1aa8cb08e48590dbb3da7b08b1056828838c5f61e6393ba7a0abcc9f662898015ad",
+			"b094dac5d93471bdec1a502270e3cc6c",
+		},
+		{
+			.AES_GCM_256,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbaddecaf888",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"522dc1f099567d07f47f37a32a84427d643a8cdcbfe5c0c97598a2bd2555d1aa8cb08e48590dbb3da7b08b1056828838c5f61e6393ba7a0abcc9f662",
+			"76fc6ece0f4e1768cddf8853bb2d551b",
+		},
+		{
+			.AES_GCM_256,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
+			"cafebabefacedbad",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"c3762df1ca787d32ae47c13bf19844cbaf1ae14d0b976afac52ff7d79bba9de0feb582d33934a4f0954cc2363bc73f7862ac430e64abe499f47c9b1f",
+			"3a337dbf46a792c45e454913fe2ea8f2",
+		},
+		{
+			.AES_GCM_256,
+			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
+			"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
+			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
+			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
+			"5a8def2f0c9e53f1f75d7853659e2a20eeb2b22aafde6419a058ab4f6f746bf40fc0c3b780f244452da3ebf1c5d82cdea2418997200ef82e44ae7e3f",
+			"a44a8266ee1c8eb0c8b5d4cf5ae9f19a",
+		},
+		// Chacha20-Poly1305
+		// https://www.rfc-editor.org/rfc/rfc8439
+		{
+			.CHACHA20POLY1305,
+			"808182838485868788898a8b8c8d8e8f909192939495969798999a9b9c9d9e9f",
+			"070000004041424344454647",
+			"50515253c0c1c2c3c4c5c6c7",
+			string(hex.encode(transmute([]byte)(_PLAINTEXT_SUNSCREEN_STR), context.temp_allocator)),
+			"d31a8d34648e60db7b86afbc53ef7ec2a4aded51296e08fea9e2b5a736ee62d63dbea45e8ca9671282fafb69da92728b1a71de0a9e060b2905d6a5b67ecd3b3692ddbd7f2d778b8c9803aee328091b58fab324e4fad675945585808b4831d7bc3ff4def08e4b7a9de576d26586cec64b6116",
+			"1ae10b594f09e26a7e902ecbd0600691",
+		},
+		// XChaCha20-Poly1305-IETF
+		// - https://datatracker.ietf.org/doc/html/draft-arciszewski-xchacha-03
+		{
+			.XCHACHA20POLY1305,
+			"808182838485868788898a8b8c8d8e8f909192939495969798999a9b9c9d9e9f",
+			"404142434445464748494a4b4c4d4e4f5051525354555657",
+			"50515253c0c1c2c3c4c5c6c7",
+			"4c616469657320616e642047656e746c656d656e206f662074686520636c617373206f66202739393a204966204920636f756c64206f6666657220796f75206f6e6c79206f6e652074697020666f7220746865206675747572652c2073756e73637265656e20776f756c642062652069742e",
+			"bd6d179d3e83d43b9576579493c0e939572a1700252bfaccbed2902c21396cbb731c7f1b0b4aa6440bf3a82f4eda7e39ae64c6708c54c216cb96b72e1213b4522f8c9ba40db5d945b11b69b982c1bb9e3f3fac2bc369488f76b2383565d3fff921f9664c97637da9768812f615c68b13b52e",
+			"c0875924c1c7987947deafd8780acf49",
+		},
+	}
+	for v, _ in test_vectors {
+		algo_name := aead.ALGORITHM_NAMES[v.algo]
+
+		key, _ := hex.decode(transmute([]byte)(v.key), context.temp_allocator)
+		iv, _ := hex.decode(transmute([]byte)(v.iv), context.temp_allocator)
+		aad, _ := hex.decode(transmute([]byte)(v.aad), context.temp_allocator)
+		plaintext, _ := hex.decode(transmute([]byte)(v.plaintext), context.temp_allocator)
+		ciphertext, _ := hex.decode(transmute([]byte)(v.ciphertext), context.temp_allocator)
+		tag, _ := hex.decode(transmute([]byte)(v.tag), context.temp_allocator)
+
+		tag_ := make([]byte, len(tag), context.temp_allocator)
+		dst := make([]byte, len(ciphertext), context.temp_allocator)
+
+		ctx: aead.Context
+		for impl in impls[v.algo] {
+			aead.init(&ctx, v.algo, key, impl)
+
+			aead.seal(&ctx, dst, tag_, iv, aad, plaintext)
+			dst_str := string(hex.encode(dst, context.temp_allocator))
+			tag_str := string(hex.encode(tag_, context.temp_allocator))
+			testing.expectf(
+				t,
+				dst_str == v.ciphertext && tag_str == v.tag,
+				"%s/%v: Expected: (%s, %s) for seal_ctx(%s, %s, %s, %s), but got (%s, %s) instead",
+				algo_name,
+				impl,
+				v.ciphertext,
+				v.tag,
+				v.key,
+				v.iv,
+				v.aad,
+				v.plaintext,
+				dst_str,
+				tag_str,
+			)
+
+			aead.seal(v.algo, dst, tag_, key, iv, aad, plaintext, impl)
+			dst_str = string(hex.encode(dst, context.temp_allocator))
+			tag_str = string(hex.encode(tag_, context.temp_allocator))
+			testing.expectf(
+				t,
+				dst_str == v.ciphertext && tag_str == v.tag,
+				"%s/%v: Expected: (%s, %s) for seal_oneshot(%s, %s, %s, %s), but got (%s, %s) instead",
+				algo_name,
+				impl,
+				v.ciphertext,
+				v.tag,
+				v.key,
+				v.iv,
+				v.aad,
+				v.plaintext,
+				dst_str,
+				tag_str,
+			)
+
+			ok := aead.open(&ctx, dst, iv, aad, ciphertext, tag)
+			dst_str = string(hex.encode(dst, context.temp_allocator))
+			testing.expectf(
+				t,
+				ok && dst_str == v.plaintext,
+				"%s/%v: Expected: (%s, true) for open_ctx(%s, %s, %s, %s, %s), but got (%s, %v) instead",
+				algo_name,
+				impl,
+				v.plaintext,
+				v.key,
+				v.iv,
+				v.aad,
+				v.ciphertext,
+				v.tag,
+				dst_str,
+				ok,
+			)
+
+			ok = aead.open(v.algo, dst, key, iv, aad, ciphertext, tag, impl)
+			dst_str = string(hex.encode(dst, context.temp_allocator))
+			testing.expectf(
+				t,
+				ok && dst_str == v.plaintext,
+				"%s/%v: Expected: (%s, true) for open_oneshot(%s, %s, %s, %s, %s), but got (%s, %v) instead",
+				algo_name,
+				impl,
+				v.plaintext,
+				v.key,
+				v.iv,
+				v.aad,
+				v.ciphertext,
+				v.tag,
+				dst_str,
+				ok,
+			)
+
+			tag_[0] ~= 0xa5
+			ok = aead.open(&ctx, dst, iv, aad, ciphertext, tag_)
+			testing.expectf(t, !ok, "%s/%v: Expected false for open(bad_tag, aad, ciphertext)", algo_name, impl)
+
+			if len(dst) > 0 {
+				copy(dst, ciphertext[:])
+				dst[0] ~= 0xa5
+				ok = aead.open(&ctx, dst, iv, aad, dst, tag)
+				testing.expectf(t, !ok, "%s/%v: Expected false for open(tag, aad, bad_ciphertext)", algo_name, impl)
+			}
+
+			if len(aad) > 0 {
+				aad_ := make([]byte, len(aad), context.temp_allocator)
+				copy(aad_, aad)
+				aad_[0] ~= 0xa5
+				ok = aead.open(&ctx, dst, iv, aad_, ciphertext, tag)
+				testing.expectf(t, !ok, "%s/%v: Expected false for open(tag, bad_aad, ciphertext)", algo_name, impl)
+			}
+		}
+	}
+}
diff --git a/tests/core/crypto/test_core_crypto_aes.odin b/tests/core/crypto/test_core_crypto_aes.odin
index c2fa2835c..b68b30976 100644
--- a/tests/core/crypto/test_core_crypto_aes.odin
+++ b/tests/core/crypto/test_core_crypto_aes.odin
@@ -12,20 +12,24 @@ import "core:crypto/sha2"
 test_aes :: proc(t: ^testing.T) {
 	runtime.DEFAULT_TEMP_ALLOCATOR_TEMP_GUARD()
 
-	impls := make([dynamic]aes.Implementation, 0, 2)
-	defer delete(impls)
-	append(&impls, aes.Implementation.Portable)
-	if aes.is_hardware_accelerated() {
-		append(&impls, aes.Implementation.Hardware)
-	}
+	impls := supported_aes_impls()
 
 	for impl in impls {
 		test_aes_ecb(t, impl)
 		test_aes_ctr(t, impl)
-		test_aes_gcm(t, impl)
 	}
 }
 
+supported_aes_impls :: proc() -> [dynamic]aes.Implementation {
+	impls := make([dynamic]aes.Implementation, 0, 2, context.temp_allocator)
+	append(&impls, aes.Implementation.Portable)
+	if aes.is_hardware_accelerated() {
+		append(&impls, aes.Implementation.Hardware)
+	}
+
+	return impls
+}
+
 test_aes_ecb :: proc(t: ^testing.T, impl: aes.Implementation) {
 	log.debugf("Testing AES-ECB/%v", impl)
 
@@ -197,13 +201,13 @@ test_aes_ctr :: proc(t: ^testing.T, impl: aes.Implementation) {
 
 	ctx: aes.Context_CTR
 	key: [aes.KEY_SIZE_256]byte
-	nonce: [aes.CTR_IV_SIZE]byte
-	aes.init_ctr(&ctx, key[:], nonce[:], impl)
+	iv: [aes.CTR_IV_SIZE]byte
+	aes.init_ctr(&ctx, key[:], iv[:], impl)
 
 	h_ctx: sha2.Context_512
 	sha2.init_512_256(&h_ctx)
 
-	for i := 1; i < 2048; i = i + 1 {
+	for i := 1; i <= 2048; i = i + 1 {
 		aes.keystream_bytes_ctr(&ctx, tmp[:i])
 		sha2.update(&h_ctx, tmp[:i])
 	}
@@ -212,7 +216,7 @@ test_aes_ctr :: proc(t: ^testing.T, impl: aes.Implementation) {
 	sha2.final(&h_ctx, digest[:])
 	digest_str := string(hex.encode(digest[:], context.temp_allocator))
 
-	expected_digest_str := "d4445343afeb9d1237f95b10d00358aed4c1d7d57c9fe480cd0afb5e2ffd448c"
+	expected_digest_str := "b5ba4e7d6e3d1ff2bb54387fc1528573a6b351610ce7bcc80b00da089f4b1bf0"
 	testing.expectf(
 		t,
 		expected_digest_str == digest_str,
@@ -222,223 +226,3 @@ test_aes_ctr :: proc(t: ^testing.T, impl: aes.Implementation) {
 		digest_str,
 	)
 }
-
-test_aes_gcm :: proc(t: ^testing.T, impl: aes.Implementation) {
-	log.debugf("Testing AES-GCM/%v", impl)
-
-	// NIST did a reorg of their site, so the source of the test vectors
-	// is only available from an archive.  The commented out tests are
-	// for non-96-bit IVs which our implementation does not support.
-	//
-	// https://csrc.nist.rip/groups/ST/toolkit/BCM/documents/proposedmodes/gcm/gcm-revised-spec.pdf
-	test_vectors := []struct {
-		key: string,
-		iv: string,
-		aad: string,
-		plaintext: string,
-		ciphertext: string,
-		tag: string,
-	} {
-		{
-			"00000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"",
-			"",
-			"58e2fccefa7e3061367f1d57a4e7455a",
-		},
-		{
-			"00000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"00000000000000000000000000000000",
-			"0388dace60b6a392f328c2b971b2fe78",
-			"ab6e47d42cec13bdf53a67b21257bddf",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308",
-			"cafebabefacedbaddecaf888",
-			"",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
-			"42831ec2217774244b7221b784d0d49ce3aa212f2c02a4e035c17e2329aca12e21d514b25466931c7d8f6a5aac84aa051ba30b396a0aac973d58e091473f5985",
-			"4d5c2af327cd64a62cf35abd2ba6fab4",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308",
-			"cafebabefacedbaddecaf888",
-			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-			"42831ec2217774244b7221b784d0d49ce3aa212f2c02a4e035c17e2329aca12e21d514b25466931c7d8f6a5aac84aa051ba30b396a0aac973d58e091",
-			"5bc94fbc3221a5db94fae95ae7121a47",
-		},
-		/*
-			{
-				"feffe9928665731c6d6a8f9467308308",
-				"cafebabefacedbad",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"61353b4c2806934a777ff51fa22a4755699b2a714fcdc6f83766e5f97b6c742373806900e49f24b22b097544d4896b424989b5e1ebac0f07c23f4598",
-				"3612d2e79e3b0785561be14aaca2fccb",
-			},
-			{
-				"feffe9928665731c6d6a8f9467308308",
-				"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"8ce24998625615b603a033aca13fb894be9112a5c3a211a8ba262a3cca7e2ca701e4a9a4fba43c90ccdcb281d48c7c6fd62875d2aca417034c34aee5",
-				"619cc5aefffe0bfa462af43c1699d050",
-			},
-		*/
-		{
-			"000000000000000000000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"",
-			"",
-			"cd33b28ac773f74ba00ed1f312572435",
-		},
-		{
-			"000000000000000000000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"00000000000000000000000000000000",
-			"98e7247c07f0fe411c267e4384b0f600",
-			"2ff58d80033927ab8ef4d4587514f0fb",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
-			"cafebabefacedbaddecaf888",
-			"",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
-			"3980ca0b3c00e841eb06fac4872a2757859e1ceaa6efd984628593b40ca1e19c7d773d00c144c525ac619d18c84a3f4718e2448b2fe324d9ccda2710acade256",
-			"9924a7c8587336bfb118024db8674a14",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308feffe9928665731c",
-			"cafebabefacedbaddecaf888",
-			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-			"3980ca0b3c00e841eb06fac4872a2757859e1ceaa6efd984628593b40ca1e19c7d773d00c144c525ac619d18c84a3f4718e2448b2fe324d9ccda2710",
-			"2519498e80f1478f37ba55bd6d27618c",
-		},
-		/*
-			{
-				"feffe9928665731c6d6a8f9467308308feffe9928665731c",
-				"cafebabefacedbad",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"0f10f599ae14a154ed24b36e25324db8c566632ef2bbb34f8347280fc4507057fddc29df9a471f75c66541d4d4dad1c9e93a19a58e8b473fa0f062f7",
-				"65dcc57fcf623a24094fcca40d3533f8",
-			},
-			{
-				"feffe9928665731c6d6a8f9467308308feffe9928665731c",
-				"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"d27e88681ce3243c4830165a8fdcf9ff1de9a1d8e6b447ef6ef7b79828666e4581e79012af34ddd9e2f037589b292db3e67c036745fa22e7e9b7373b",
-				"dcf566ff291c25bbb8568fc3d376a6d9",
-			},
-		*/
-		{
-			"0000000000000000000000000000000000000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"",
-			"",
-			"530f8afbc74536b9a963b4f1c4cb738b",
-		},
-		{
-			"0000000000000000000000000000000000000000000000000000000000000000",
-			"000000000000000000000000",
-			"",
-			"00000000000000000000000000000000",
-			"cea7403d4d606b6e074ec5d3baf39d18",
-			"d0d1c8a799996bf0265b98b5d48ab919",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
-			"cafebabefacedbaddecaf888",
-			"",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b391aafd255",
-			"522dc1f099567d07f47f37a32a84427d643a8cdcbfe5c0c97598a2bd2555d1aa8cb08e48590dbb3da7b08b1056828838c5f61e6393ba7a0abcc9f662898015ad",
-			"b094dac5d93471bdec1a502270e3cc6c",
-		},
-		{
-			"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
-			"cafebabefacedbaddecaf888",
-			"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-			"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-			"522dc1f099567d07f47f37a32a84427d643a8cdcbfe5c0c97598a2bd2555d1aa8cb08e48590dbb3da7b08b1056828838c5f61e6393ba7a0abcc9f662",
-			"76fc6ece0f4e1768cddf8853bb2d551b",
-		},
-		/*
-			{
-				"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
-				"cafebabefacedbad",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"c3762df1ca787d32ae47c13bf19844cbaf1ae14d0b976afac52ff7d79bba9de0feb582d33934a4f0954cc2363bc73f7862ac430e64abe499f47c9b1f",
-				"3a337dbf46a792c45e454913fe2ea8f2",
-			},
-			{
-				"feffe9928665731c6d6a8f9467308308feffe9928665731c6d6a8f9467308308",
-				"9313225df88406e555909c5aff5269aa6a7a9538534f7da1e4c303d2a318a728c3c0c95156809539fcf0e2429a6b525416aedbf5a0de6a57a637b39b",
-				"feedfacedeadbeeffeedfacedeadbeefabaddad2",
-				"d9313225f88406e5a55909c5aff5269a86a7a9531534f7da2e4c303d8a318a721c3c0c95956809532fcf0e2449a6b525b16aedf5aa0de657ba637b39",
-				"5a8def2f0c9e53f1f75d7853659e2a20eeb2b22aafde6419a058ab4f6f746bf40fc0c3b780f244452da3ebf1c5d82cdea2418997200ef82e44ae7e3f",
-				"a44a8266ee1c8eb0c8b5d4cf5ae9f19a",
-			},
-		*/
-	}
-	for v, _ in test_vectors {
-		key, _ := hex.decode(transmute([]byte)(v.key), context.temp_allocator)
-		iv, _ := hex.decode(transmute([]byte)(v.iv), context.temp_allocator)
-		aad, _ := hex.decode(transmute([]byte)(v.aad), context.temp_allocator)
-		plaintext, _ := hex.decode(transmute([]byte)(v.plaintext), context.temp_allocator)
-		ciphertext, _ := hex.decode(transmute([]byte)(v.ciphertext), context.temp_allocator)
-		tag, _ := hex.decode(transmute([]byte)(v.tag), context.temp_allocator)
-
-		tag_ := make([]byte, len(tag), context.temp_allocator)
-		dst := make([]byte, len(ciphertext), context.temp_allocator)
-
-		ctx: aes.Context_GCM
-		aes.init_gcm(&ctx, key, impl)
-
-		aes.seal_gcm(&ctx, dst, tag_, iv, aad, plaintext)
-		dst_str := string(hex.encode(dst[:], context.temp_allocator))
-		tag_str := string(hex.encode(tag_[:], context.temp_allocator))
-
-		testing.expectf(
-			t,
-			dst_str == v.ciphertext && tag_str == v.tag,
-			"AES-GCM/%v: Expected: (%s, %s) for seal(%s, %s, %s, %s), but got (%s, %s) instead",
-			impl,
-			v.ciphertext,
-			v.tag,
-			v.key,
-			v.iv,
-			v.aad,
-			v.plaintext,
-			dst_str,
-			tag_str,
-		)
-
-		ok := aes.open_gcm(&ctx, dst, iv, aad, ciphertext, tag)
-		dst_str = string(hex.encode(dst[:], context.temp_allocator))
-
-		testing.expectf(
-			t,
-			ok && dst_str == v.plaintext,
-			"AES-GCM/%v: Expected: (%s, true) for open(%s, %s, %s, %s, %s), but got (%s, %v) instead",
-			impl,
-			v.plaintext,
-			v.key,
-			v.iv,
-			v.aad,
-			v.ciphertext,
-			v.tag,
-			dst_str,
-			ok,
-		)
-	}
-}