Parallel OCB decryption

1 files changed, 34 insertions, 10 deletions

diff --git a/src/filters/modes/ocb/ocb.cpp b/src/filters/modes/ocb/ocb.cpp
index 24f777b73..5750e3da2 100644
--- a/src/filters/modes/ocb/ocb.cpp
+++ b/src/filters/modes/ocb/ocb.cpp
@@ -307,24 +307,48 @@ void OCB_Decryption::buffered_block(const byte input[], size_t input_length)
 
    const size_t blocks = input_length / BS;
 
-   const L_computer& L = *m_L;
+   const size_t par_bytes = m_cipher->parallel_bytes();
+
+   BOTAN_ASSERT(par_bytes % BS == 0, "Cipher is parallel in full blocks");
+
+   const size_t par_blocks = par_bytes / BS;
+
+   const L_computer& L = *m_L; // convenient name
 
-   secure_vector<byte> ptext_buf(BS);
+   secure_vector<byte> ptext_buf(par_bytes);
+   secure_vector<byte> csum_accum(par_bytes);
+   secure_vector<byte> offsets(par_bytes);
+
+   size_t blocks_left = blocks;
 
-   for(size_t i = 0; i != blocks; ++i)
+   while(blocks_left)
       {
-      // could run in parallel
+      const size_t to_proc = std::min(blocks_left, par_blocks);
+      const size_t proc_bytes = to_proc * BS;
 
-      m_offset ^= L(ctz(++m_block_index));
+      for(size_t i = 0; i != to_proc; ++i)
+         {
+         m_offset ^= L(ctz(++m_block_index));
+         copy_mem(&offsets[BS*i], &m_offset[0], BS);
+         }
+
+      copy_mem(&ptext_buf[0], &input[0], proc_bytes);
 
-      ptext_buf = m_offset;
-      xor_buf(&ptext_buf[0], &input[BS*i], BS);
+      ptext_buf ^= offsets;
       m_cipher->decrypt(ptext_buf);
-      ptext_buf ^= m_offset;
+      ptext_buf ^= offsets;
+
+      xor_buf(&csum_accum[0], &ptext_buf[0], proc_bytes);
 
-      send(ptext_buf);
-      m_checksum ^= ptext_buf;
+      send(ptext_buf, proc_bytes);
+
+      input += proc_bytes;
+      blocks_left -= to_proc;
       }
+
+   // fold into checksum
+   for(size_t i = 0; i != csum_accum.size(); ++i)
+      m_checksum[i % BS] ^= csum_accum[i];
    }
 
 void OCB_Decryption::buffered_final(const byte input[], size_t input_length)