diff --git a/main.c b/main.c
index 168ffd7..4d642a7 100644
--- a/main.c
+++ b/main.c
@@ -1224,7 +1224,8 @@ redo:
 		rctx = tdat->rctx;
 		reset_dedupe_context(tdat->rctx);
 		rctx->cbuf = tdat->uncompressed_chunk;
-		dedupe_index_sz = dedupe_compress(tdat->rctx, tdat->cmp_seg, &(tdat->rbytes), 0, NULL);
+		dedupe_index_sz = dedupe_compress(tdat->rctx, tdat->cmp_seg, &(tdat->rbytes), 0,
+						  NULL, tdat->cksum_mt);
 		if (!rctx->valid) {
 			memcpy(tdat->uncompressed_chunk, tdat->cmp_seg, rbytes);
 			tdat->rbytes = rbytes;
diff --git a/rabin/rabin_dedup.c b/rabin/rabin_dedup.c
index 7612269..8ecdb85 100755
--- a/rabin/rabin_dedup.c
+++ b/rabin/rabin_dedup.c
@@ -76,6 +76,10 @@
 #	define	SSE_MODE		1
 #endif
 
+#if defined(_OPENMP)
+#include <omp.h>
+#endif
+
 #define	DELTA_EXTRA2_PCT(x) ((x) >> 1)
 #define	DELTA_EXTRA_PCT(x) (((x) >> 1) + ((x) >> 3))
 #define	DELTA_NORMAL_PCT(x) (((x) >> 1) + ((x) >> 2) + ((x) >> 3))
@@ -298,7 +302,8 @@ destroy_dedupe_context(dedupe_context_t *ctx)
  * from 4K-128K.
  */
 uint32_t
-dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, uint64_t *size, uint64_t offset, uint64_t *rabin_pos)
+dedupe_compress(dedupe_context_t *ctx, uchar_t *buf, uint64_t *size, uint64_t offset,
+		uint64_t *rabin_pos, int mt)
 {
 	uint64_t i, last_offset, j, ary_sz;
 	uint32_t blknum, window_pos;
@@ -569,14 +574,20 @@ process_blocks:
 		 * have a fast linear scan through the buffer.
 		 */
 		if (ctx->delta_flag) {
+#if defined(_OPENMP)
+#	pragma omp parallel for if (mt)
+#endif
 			for (i=0; i<blknum; i++) {
 				ctx->blocks[i]->hash = XXH32(buf1+ctx->blocks[i]->offset,
-								    ctx->blocks[i]->length, 0);
+								ctx->blocks[i]->length, 0);
 			}
 		} else {
+#if defined(_OPENMP)
+#	pragma omp parallel for if (mt)
+#endif
 			for (i=0; i<blknum; i++) {
 				ctx->blocks[i]->hash = XXH32(buf1+ctx->blocks[i]->offset,
-								    ctx->blocks[i]->length, 0);
+								ctx->blocks[i]->length, 0);
 				ctx->blocks[i]->similarity_hash = ctx->blocks[i]->hash;
 			}
 		}
diff --git a/rabin/rabin_dedup.h b/rabin/rabin_dedup.h
index e6d4bbb..1feae91 100644
--- a/rabin/rabin_dedup.h
+++ b/rabin/rabin_dedup.h
@@ -170,7 +170,7 @@ extern dedupe_context_t *create_dedupe_context(uint64_t chunksize, uint64_t real
 	int file_version, compress_op_t op);
 extern void destroy_dedupe_context(dedupe_context_t *ctx);
 extern unsigned int dedupe_compress(dedupe_context_t *ctx, unsigned char *buf, 
-	uint64_t *size, uint64_t offset, uint64_t *rabin_pos);
+	uint64_t *size, uint64_t offset, uint64_t *rabin_pos, int mt);
 extern void dedupe_decompress(dedupe_context_t *ctx, uchar_t *buf, uint64_t *size);
 extern void parse_dedupe_hdr(uchar_t *buf, unsigned int *blknum, uint64_t *dedupe_index_sz,
 		uint64_t *dedupe_data_sz, uint64_t *rabin_index_sz_cmp,
diff --git a/utils/utils.c b/utils/utils.c
index 795bd2f..a7ab363 100644
--- a/utils/utils.c
+++ b/utils/utils.c
@@ -239,7 +239,7 @@ Read_Adjusted(int fd, uchar_t *buf, uint64_t count, int64_t *rabin_count, void *
 			 * This call does not actually dedupe but finds the last rabin boundary
 			 * in the buf.
 			 */
-			dedupe_compress(rctx, buf, &rc, 0, &rbc);
+			dedupe_compress(rctx, buf, &rc, 0, &rbc, 0);
 			rcount = rc;
 			*rabin_count = rbc;
 		} else {