perf: improve parallelism of SortBotMerge

sources/ftypes.h

@@ -971,16 +971,28 @@ typedef int (*TFUN1)(UBYTE *,int);
 #define BUCKETTERMINATED 3
 #define BUCKETRELEASED 4
 
-#define NUMBEROFBLOCKSINSORT 10
-#define MINIMUMNUMBEROFTERMS 10
-
 #define BUCKETDOINGTERM 1
 #define BUCKETASSIGNED -1
 #define BUCKETTOBERELEASED -2
 #define BUCKETPREPARINGTERM 0
 
 #define BUCKETDOINGTERMS 0
 #define BUCKETDOINGBRACKET 1
+
+/*
+	Sortblock config
+*/
+// The number of blocks in the ring-buffer for data transfer between
+// workers or sortbots and the master thread or other sortbots. With
+// sortbots, each thread has half of this number.
+#define NUMBEROFBLOCKSINSORT 10
+// The minimum number of terms which must fit in a block, i.e. they
+// must be at least this number times MaxTermSize.
+#define MINIMUMNUMBEROFTERMS 2
+// The minimum number of terms which will be written in a block,
+// before potentially yielding the block to a waiting reading thread.
+#define MINWRITENUMBEROFTERMS 10
+
 #endif
 
 /*

sources/setfile.c

@@ -389,13 +389,13 @@ int RecalcSetups(void)
 	AM.MaxTer *= sizeof(WORD);
 #ifdef WITHSORTBOTS
 	if ( AM.totalnumberofthreads-1 > 2 ) {
-		minimumsize = (2*(AM.totalnumberofthreads-1)-2)*(AM.MaxTer+
+		minimumsize = (2*(AM.totalnumberofthreads-1)-2)*(
 			NUMBEROFBLOCKSINSORT*MINIMUMNUMBEROFTERMS/2*AM.MaxTer);
 	}
 	else
 #endif
 	{
-		minimumsize = (AM.totalnumberofthreads-1)*(AM.MaxTer+
+		minimumsize = (AM.totalnumberofthreads-1)*(
 			NUMBEROFBLOCKSINSORT*MINIMUMNUMBEROFTERMS*AM.MaxTer);
 	}
 	if ( totalsize < minimumsize ) {

sources/structs.h

@@ -1184,6 +1184,7 @@ typedef struct SoRtBlOcK {
     WORD    **MasterStart;
     WORD    **MasterFill;
     WORD    **MasterStop;
+    LONG    *BlockTerms;
     int     MasterNumBlocks;
     int     MasterBlock;
     int     FillBlock;