integrate with c code

4 files changed, 59 insertions, 14 deletions

diff --git a/c/blake3.c b/c/blake3.c
index 9becead..8c95d1c 100644
--- a/c/blake3.c
+++ b/c/blake3.c
@@ -91,21 +91,23 @@ INLINE void output_root_bytes(const output_t *self, uint64_t seek, uint8_t *out,
   uint64_t output_block_counter = seek / 64;
   size_t offset_within_block = seek % 64;
   uint8_t wide_buf[64];
-  while (out_len > 0) {
-    blake3_compress_xof(self->input_cv, self->block, self->block_len,
-                        output_block_counter, self->flags | ROOT, wide_buf);
-    size_t available_bytes = 64 - offset_within_block;
-    size_t memcpy_len;
-    if (out_len > available_bytes) {
-      memcpy_len = available_bytes;
-    } else {
-      memcpy_len = out_len;
-    }
-    memcpy(out, wide_buf + offset_within_block, memcpy_len);
-    out += memcpy_len;
-    out_len -= memcpy_len;
+  if(offset_within_block) {
+    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);
+    const size_t available_bytes = 64 - offset_within_block;
+    const size_t bytes = out_len > available_bytes ? available_bytes : out_len;
+    memcpy(out, wide_buf + offset_within_block, bytes);
+    out += bytes;
+    out_len -= bytes;
     output_block_counter += 1;
-    offset_within_block = 0;
+  }
+  if(out_len / 64)
+    blake3_xof_many(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, out, out_len / 64);
+  output_block_counter += out_len / 64;
+  out += out_len & -64;
+  out_len -= out_len & -64;
+  if(out_len) {
+    blake3_compress_xof(self->input_cv, self->block, self->block_len, output_block_counter, self->flags | ROOT, wide_buf);
+    memcpy(out, wide_buf, out_len);
   }
 }
 
diff --git a/c/blake3_dispatch.c b/c/blake3_dispatch.c
index c9abc13..725f7ee 100644
--- a/c/blake3_dispatch.c
+++ b/c/blake3_dispatch.c
@@ -223,6 +223,23 @@ void blake3_compress_xof(const uint32_t cv[8],
   blake3_compress_xof_portable(cv, block, block_len, counter, flags, out);
 }
 
+
+void blake3_xof_many(const uint32_t cv[8],
+                     const uint8_t block[BLAKE3_BLOCK_LEN],
+                     uint8_t block_len, uint64_t counter, uint8_t flags,
+                     uint8_t out[64], size_t outblocks) {
+#if defined(IS_X86)
+  const enum cpu_feature features = get_cpu_features();
+#if !defined(BLAKE3_NO_AVX512)
+  if (features & AVX512VL) {
+    blake3_xof_many_avx512(cv, block, block_len, counter, flags, out, outblocks);
+    return;
+  }
+#endif
+#endif
+  blake3_xof_many_portable(cv, block, block_len, counter, flags, out, outblocks);
+}
+
 void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
                       size_t blocks, const uint32_t key[8], uint64_t counter,
                       bool increment_counter, uint8_t flags,
diff --git a/c/blake3_impl.h b/c/blake3_impl.h
index 98611c3..e652f86 100644
--- a/c/blake3_impl.h
+++ b/c/blake3_impl.h
@@ -191,6 +191,11 @@ void blake3_compress_xof(const uint32_t cv[8],
                          uint8_t block_len, uint64_t counter, uint8_t flags,
                          uint8_t out[64]);
 
+void blake3_xof_many(const uint32_t cv[8],
+                     const uint8_t block[BLAKE3_BLOCK_LEN],
+                     uint8_t block_len, uint64_t counter, uint8_t flags,
+                     uint8_t out[64], size_t outblocks);
+
 void blake3_hash_many(const uint8_t *const *inputs, size_t num_inputs,
                       size_t blocks, const uint32_t key[8], uint64_t counter,
                       bool increment_counter, uint8_t flags,
@@ -210,6 +215,11 @@ void blake3_compress_xof_portable(const uint32_t cv[8],
                                   uint8_t block_len, uint64_t counter,
                                   uint8_t flags, uint8_t out[64]);
 
+void blake3_xof_many_portable(const uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter, uint8_t flags,
+                              uint8_t out[64], size_t outblocks);
+
 void blake3_hash_many_portable(const uint8_t *const *inputs, size_t num_inputs,
                                size_t blocks, const uint32_t key[8],
                                uint64_t counter, bool increment_counter,
@@ -265,6 +275,12 @@ void blake3_compress_xof_avx512(const uint32_t cv[8],
                                 uint8_t block_len, uint64_t counter,
                                 uint8_t flags, uint8_t out[64]);
 
+
+void blake3_xof_many_avx512(const uint32_t cv[8],
+                            const uint8_t block[BLAKE3_BLOCK_LEN],
+                            uint8_t block_len, uint64_t counter, uint8_t flags,
+                            uint8_t out[64], size_t outblocks);
+
 void blake3_hash_many_avx512(const uint8_t *const *inputs, size_t num_inputs,
                              size_t blocks, const uint32_t key[8],
                              uint64_t counter, bool increment_counter,
diff --git a/c/blake3_portable.c b/c/blake3_portable.c
index 062dd1b..73e7f3e 100644
--- a/c/blake3_portable.c
+++ b/c/blake3_portable.c
@@ -122,6 +122,16 @@ void blake3_compress_xof_portable(const uint32_t cv[8],
   store32(&out[15 * 4], state[15] ^ cv[7]);
 }
 
+void blake3_xof_many_portable(const uint32_t cv[8],
+                              const uint8_t block[BLAKE3_BLOCK_LEN],
+                              uint8_t block_len, uint64_t counter, uint8_t flags,
+                              uint8_t out[BLAKE3_BLOCK_LEN], size_t outblocks)
+{
+  for(size_t i = 0; i < outblocks; ++i) {
+    blake3_compress_xof_portable(cv, block, block_len, counter + i, flags, out + 64*i);
+  }
+}
+
 INLINE void hash_one_portable(const uint8_t *input, size_t blocks,
                               const uint32_t key[8], uint64_t counter,
                               uint8_t flags, uint8_t flags_start,