Revert one of the previous fixes -- we should always iterate over the nonzeroes of x, since that is the vector with guaranteed the fewest number of nonzeroes. Instead, the error was in the computation of the mask -- it used y instead of x. Special note that the other fix WAS correct-- so there were two bugs in total. Also a minor code style fix

Author: anyzelman

include/graphblas/nonblocking/blas1.hpp

@@ -10398,6 +10398,9 @@ namespace grb {
 	// internal namespace for implementation of grb::dot
 	namespace internal {
 
+		/**
+		 * Invariant: x should have fewer nonzeroes than y
+		 */
 		template<
 			Descriptor descr,
 #ifdef GRB_BOOLEAN_DISPATCHER
@@ -10432,6 +10435,7 @@ namespace grb {
 #else
 			(void) upper_bound;
 #endif
+			assert( local_x.nonzeroes() <= local_y.nonzeroes() );
 
 			// get raw alias
 			const InputType1 * __restrict__ a = internal::getRaw( x );
@@ -10450,8 +10454,7 @@ namespace grb {
 
 					// prepare registers
 					for( size_t k = 0; k < AnyOp::blocksize; ++k, ++i ) {
-						mask[ k ] = already_dense_input_x ||
-							local_x.assigned( already_dense_input_y ? i : local_y.index( i ) );
+						mask[ k ] = already_dense_input_x || local_y.assigned(local_x.index( i ));
 					}
 
 					// rewind
@@ -10463,7 +10466,7 @@ namespace grb {
 							xx[ k ] = static_cast< typename AnyOp::D1 >(
 								a[ ( already_dense_input_x ? i : local_x.index( i ) ) + lower_bound ] );
 							yy[ k ] = static_cast< typename AnyOp::D2 >(
-								b[ ( already_dense_input_y ? i : local_y.index( i ) ) + lower_bound ] );
+								b[ ( already_dense_input_x ? i : local_x.index( i ) ) + lower_bound ] );
 						}
 					}
 
@@ -10658,7 +10661,9 @@ namespace grb {
 										already_dense_input_y, already_dense_input_x,
 										array_reduced[ thread_id ],
 										lower_bound, upper_bound,
-										local_y, local_x, x, y, local_y_nz,
+										local_y, local_x,
+										x, y,
+										local_y_nz,
 										addMonoid, anyOp
 									);
 							}