Minor update, there's no need to update Aj inside of systolic loop.

1 files changed, 8 insertions, 3 deletions

diff --git a/modexp_fpga_systolic.cpp b/modexp_fpga_systolic.cpp
index 9dce130..93c6111 100644
--- a/modexp_fpga_systolic.cpp
+++ b/modexp_fpga_systolic.cpp
@@ -42,6 +42,9 @@
 #include "modexp_fpga_model_pe.h"
 
 
+#include <stdio.h>
+
+
 //----------------------------------------------------------------
 void multiply_systolic(const FPGA_WORD *A, const FPGA_WORD *B, FPGA_WORD *P, size_t len_ab, size_t len_p)
 //----------------------------------------------------------------
@@ -76,15 +79,17 @@ void multiply_systolic(const FPGA_WORD *A, const FPGA_WORD *B, FPGA_WORD *P, siz
 			// reset word index
 		j_index = 0;
 
-			// scan chunks of A
+			// current word of A
+		Aj = (i < len_ab) ? A[i] : 0;
+
+			// scan chunks of B
 		for (k=0; k<num_systolic_cycles; k++)
 		{
 				// simulate how systolic array would work
 			for (j=0; j<SYSTOLIC_NUM_WORDS; j++)
 			{
-					// current words of B and A
+					// current word of B
 				Bj = (j_index < len_ab) ? B[j_index] : 0;
-				Aj = (i       < len_ab) ? A[i]       : 0;
 
 					// Pj = Aj * Bj
 				pe_mul(Aj, Bj, t[k][j], c_in[k][j], &s[k][j], &c_out[k][j]);