Really dumb division bug fixed.

All tests now pass except summation, which fails to meet tolerance.

Author: obilaniu

src/gpuarray_reduction.c

@@ -2386,10 +2386,11 @@ static void       reduxGenSrcAppendPhase1       (GpuReduction*        gr){
 	"        if(misalignL && doFinish && LID_0 < D){\n"
 	"            SETREDUXSTATE(accV, accI, wdL[(GID_0+0)*D+LID_0], waL[(GID_0+0)*D+LID_0]);\n"
 	"            \n"
-	"            for(k=-1;                /* Starting with the first block to our left... */\n"
-	"                (start      +0)/B == /* Is our write target the same as that of */\n"
-	"                (start+k*V+V-1)/B;   /* the target k blocks to our left? */\n"
-	"                k--){                /* Try moving one more to the left. */\n"
+	"                           /* vvv-- NOTA BENE: The +B hack is REALLY NECESSARY, since C division is rounding to zero: (-1)/B == (B-1)/B for B>1. */\n"
+	"            for(k=-1;                  /* Starting with the first block to our left... */\n"
+	"                (start        +B)/B == /* Is our write target the same as that of */\n"
+	"                (start+k*V+V-1+B)/B;   /* the target k blocks to our left? */\n"
+	"                k--){                  /* Try moving one more to the left. */\n"
 	"                REDUX(accV, accI, wdR[(GID_0+k)*D+LID_0], waR[(GID_0+k)*D+LID_0]);\n"
 	"            }\n"
 	"            \n");