core/crypto/chacha20: Use 128-bit/256-bit SIMD

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
+// NONCE_SIZE is the ChaCha20 nonce size in bytes.
+NONCE_SIZE :: 12
+// XNONCE_SIZE is the XChaCha20 nonce size in bytes.
+XNONCE_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
+// nonce.
+//
+// WARNING: This ONLY handles ChaCha20.  XChaCha20 sub-key and nonce
+// derivation is expected to be handled by the caller, so that the
+// HChaCha call can be suitably accelerated.
+init :: proc "contextless" (ctx: ^Context, key, nonce: []byte, is_xchacha: bool) {
+	if len(key) != KEY_SIZE || len(nonce) != NONCE_SIZE {
+		intrinsics.trap()
+	}
+
+	k, n := key, nonce
+
+	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 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.
+
+	ERR_CTR_EXHAUSTED :: "crypto/chacha20: maximum (X)ChaCha20 keystream per nonce 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)
+		}
+	}
+}

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 nonce.
+	_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, nonce: []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(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 := _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)
+}

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 nonce.
+	_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, nonce: []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)(&nonce[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)
+}

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 nonce.
+	_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, nonce: []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, nonce)
+}

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, nonce: []byte) {
+	intrinsics.trap()
+}

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
+KEY_SIZE :: _chacha20.KEY_SIZE
 // NONCE_SIZE is the ChaCha20 nonce size in bytes.
-NONCE_SIZE :: 12
+NONCE_SIZE :: _chacha20.NONCE_SIZE
 // 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
+XNONCE_SIZE :: _chacha20.XNONCE_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) {
+init :: proc(ctx: ^Context, key, nonce: []byte, impl := Implementation.Simd256) {
 	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 {
+	if l := len(nonce); l != NONCE_SIZE && l != XNONCE_SIZE {
 		panic("crypto/chacha20: invalid (X)ChaCha20 nonce size")
 	}
 
 	k, n := key, nonce
 
-	// Derive the XChaCha20 subkey and sub-nonce via HChaCha20.
+	init_impl(ctx, impl)
+
 	is_xchacha := len(nonce) == XNONCE_SIZE
 	if is_xchacha {
-		sub_key := ctx._buffer[:KEY_SIZE]
-		_hchacha20(sub_key, k, n)
+		sub_nonce: [NONCE_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_nonce[4:], n[16:])
+		n = sub_nonce[:]
 	}
 
-	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)
 }

core/crypto/chacha20/chacha20_impl.odin

@@ -0,0 +1,52 @@
+package chacha20
+
+import "base:intrinsics"
+import "core:crypto/_chacha20/ref"
+import "core:crypto/_chacha20/simd128"
+import "core:crypto/_chacha20/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, nonce: []byte, impl: Implementation) {
+	switch impl {
+	case .Simd256:
+		simd256.hchacha20(dst, key, nonce)
+	case .Simd128:
+		simd128.hchacha20(dst, key, nonce)
+	case .Portable:
+		ref.hchacha20(dst, key, nonce)
+	}
+}

tests/core/crypto/test_core_crypto.odin

@@ -19,15 +19,36 @@ 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 = "file")
 _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 := make([dynamic]chacha20.Implementation, 0, 3)
+	defer delete(impls)
+	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)
+	}
+
+	for impl in impls {
+		test_chacha20_stream(t, impl)
+	}
+	test_chacha20poly1305(t) // TODO: Move into loop.
+}
+
+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)
 
@@ -64,7 +85,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 
 	derived_ciphertext: [114]byte
 	ctx: chacha20.Context = ---
-	chacha20.init(&ctx, key[:], nonce[:])
+	chacha20.init(&ctx, key[:], nonce[:], impl)
 	chacha20.seek(&ctx, 1) // The test vectors start the counter at 1.
 	chacha20.xor_bytes(&ctx, derived_ciphertext[:], plaintext[:])
 
@@ -109,7 +130,7 @@ test_chacha20 :: proc(t: ^testing.T) {
 	}
 	xciphertext_str := string(hex.encode(xciphertext[:], context.temp_allocator))
 
-	chacha20.init(&ctx, xkey[:], xnonce[:])
+	chacha20.init(&ctx, xkey[:], xnonce[:], impl)
 	chacha20.seek(&ctx, 1)
 	chacha20.xor_bytes(&ctx, derived_ciphertext[:], plaintext[:])
 
@@ -121,9 +142,40 @@ test_chacha20 :: proc(t: ^testing.T) {
 		xciphertext_str,
 		derived_ciphertext_str,
 	)
+
+	// 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.
+
+	tmp := make([]byte, 2048, context.temp_allocator)
+
+	mem.zero(&key, size_of(key))
+	mem.zero(&nonce, size_of(nonce))
+	chacha20.init(&ctx, key[:], nonce[:], impl)
+
+	h_ctx: sha2.Context_512
+	sha2.init_512_256(&h_ctx)
+
+	for i := 1; i <= 2048; i = i + 1 {
+		chacha20.keystream_bytes(&ctx, tmp[:i])
+		sha2.update(&h_ctx, tmp[:i])
+	}
+
+	digest: [32]byte
+	sha2.final(&h_ctx, digest[:])
+	digest_str := string(hex.encode(digest[:], context.temp_allocator))
+
+	expected_digest_str := "cfd6e949225b854fe04946491e6935ff05ff983d1554bc885bca0ec8082dd5b8"
+	testing.expectf(
+		t,
+		expected_digest_str == digest_str,
+		"Expected %s for keystream digest, but got %s instead",
+		expected_digest_str,
+		digest_str,
+	)
 }
 
-@(test)
 test_chacha20poly1305 :: proc(t: ^testing.T) {
 	plaintext := transmute([]byte)(_PLAINTEXT_SUNSCREEN_STR)