update dependencies

vendor/github.com/zeebo/blake3/internal/alg/hash/hash.go

@@ -0,0 +1,23 @@
+package hash
+
+import (
+	"github.com/zeebo/blake3/internal/alg/hash/hash_avx2"
+	"github.com/zeebo/blake3/internal/alg/hash/hash_pure"
+	"github.com/zeebo/blake3/internal/consts"
+)
+
+func HashF(input *[8192]byte, length, counter uint64, flags uint32, key *[8]uint32, out *[64]uint32, chain *[8]uint32) {
+	if consts.HasAVX2 && length > 2*consts.ChunkLen {
+		hash_avx2.HashF(input, length, counter, flags, key, out, chain)
+	} else {
+		hash_pure.HashF(input, length, counter, flags, key, out, chain)
+	}
+}
+
+func HashP(left, right *[64]uint32, flags uint32, key *[8]uint32, out *[64]uint32, n int) {
+	if consts.HasAVX2 && n >= 2 {
+		hash_avx2.HashP(left, right, flags, key, out, n)
+	} else {
+		hash_pure.HashP(left, right, flags, key, out, n)
+	}
+}

vendor/github.com/zeebo/blake3/internal/alg/hash/hash_avx2/impl_other.go

@@ -0,0 +1,13 @@
+// +build !amd64
+
+package hash_avx2
+
+import "github.com/zeebo/blake3/internal/alg/hash/hash_pure"
+
+func HashF(input *[8192]byte, length, counter uint64, flags uint32, key *[8]uint32, out *[64]uint32, chain *[8]uint32) {
+	hash_pure.HashF(input, length, counter, flags, key, out, chain)
+}
+
+func HashP(left, right *[64]uint32, flags uint32, key *[8]uint32, out *[64]uint32, n int) {
+	hash_pure.HashP(left, right, flags, key, out, n)
+}

vendor/github.com/zeebo/blake3/internal/alg/hash/hash_avx2/stubs.go

@@ -0,0 +1,9 @@
+// +build amd64
+
+package hash_avx2
+
+//go:noescape
+func HashF(input *[8192]byte, length, counter uint64, flags uint32, key *[8]uint32, out *[64]uint32, chain *[8]uint32)
+
+//go:noescape
+func HashP(left, right *[64]uint32, flags uint32, key *[8]uint32, out *[64]uint32, n int)

vendor/github.com/zeebo/blake3/internal/alg/hash/hash_pure/hashf.go

@@ -0,0 +1,56 @@
+package hash_pure
+
+import (
+	"unsafe"
+
+	"github.com/zeebo/blake3/internal/alg/compress"
+	"github.com/zeebo/blake3/internal/consts"
+	"github.com/zeebo/blake3/internal/utils"
+)
+
+func HashF(input *[8192]byte, length, counter uint64, flags uint32, key *[8]uint32, out *[64]uint32, chain *[8]uint32) {
+	var tmp [16]uint32
+
+	for i := uint64(0); consts.ChunkLen*i < length && i < 8; i++ {
+		bchain := *key
+		bflags := flags | consts.Flag_ChunkStart
+		start := consts.ChunkLen * i
+
+		for n := uint64(0); n < 16; n++ {
+			if n == 15 {
+				bflags |= consts.Flag_ChunkEnd
+			}
+			if start+64*n >= length {
+				break
+			}
+			if start+64+64*n >= length {
+				*chain = bchain
+			}
+
+			var blockPtr *[16]uint32
+			if consts.IsLittleEndian {
+				blockPtr = (*[16]uint32)(unsafe.Pointer(&input[consts.ChunkLen*i+consts.BlockLen*n]))
+			} else {
+				var block [16]uint32
+				utils.BytesToWords((*[64]uint8)(unsafe.Pointer(&input[consts.ChunkLen*i+consts.BlockLen*n])), &block)
+				blockPtr = &block
+			}
+
+			compress.Compress(&bchain, blockPtr, counter, consts.BlockLen, bflags, &tmp)
+
+			bchain = *(*[8]uint32)(unsafe.Pointer(&tmp[0]))
+			bflags = flags
+		}
+
+		out[i+0] = bchain[0]
+		out[i+8] = bchain[1]
+		out[i+16] = bchain[2]
+		out[i+24] = bchain[3]
+		out[i+32] = bchain[4]
+		out[i+40] = bchain[5]
+		out[i+48] = bchain[6]
+		out[i+56] = bchain[7]
+
+		counter++
+	}
+}

vendor/github.com/zeebo/blake3/internal/alg/hash/hash_pure/hashp.go

@@ -0,0 +1,38 @@
+package hash_pure
+
+import "github.com/zeebo/blake3/internal/alg/compress"
+
+func HashP(left, right *[64]uint32, flags uint32, key *[8]uint32, out *[64]uint32, n int) {
+	var tmp [16]uint32
+	var block [16]uint32
+
+	for i := 0; i < n && i < 8; i++ {
+		block[0] = left[i+0]
+		block[1] = left[i+8]
+		block[2] = left[i+16]
+		block[3] = left[i+24]
+		block[4] = left[i+32]
+		block[5] = left[i+40]
+		block[6] = left[i+48]
+		block[7] = left[i+56]
+		block[8] = right[i+0]
+		block[9] = right[i+8]
+		block[10] = right[i+16]
+		block[11] = right[i+24]
+		block[12] = right[i+32]
+		block[13] = right[i+40]
+		block[14] = right[i+48]
+		block[15] = right[i+56]
+
+		compress.Compress(key, &block, 0, 64, flags, &tmp)
+
+		out[i+0] = tmp[0]
+		out[i+8] = tmp[1]
+		out[i+16] = tmp[2]
+		out[i+24] = tmp[3]
+		out[i+32] = tmp[4]
+		out[i+40] = tmp[5]
+		out[i+48] = tmp[6]
+		out[i+56] = tmp[7]
+	}
+}