added simpler version sDTW

CMakeLists.txt

@@ -15,7 +15,7 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O0 -g -fsanitize=address -fno-inline -D
 # "-fsanitize=address" flag was used to check the stack smashing with Google Address Sanitizer. Use this flag with CLang and 
 # run the program to check. 
 
-set(DP_HLS_HOME "/home/centos/workspace/banding/DP-HLS")
+set(DP_HLS_HOME "/home/centos/workspace/DP-HLS")
 
 set(EXECUTABLE_TARGETS
     baseline_local_linear

kernels/sdtw/kernel_sdtw.cpp

@@ -14,9 +14,9 @@ void SDTW::InitializeScores(
     // Because the cell scores are non-negative (because the distance is L1 distance and accumulates), thus initialize
     // them to all 0 means exactly the same to initialize the first column in actual score matrix to be the distance
     // between q_i and r_0. min(diag, left, always have 0).
-    for (int i = 0; i < MAX_QUERY_LENGTH; i++)
+    for (int i = 0; i < MAX_REFERENCE_LENGTH; i++)
     {
-        init_col_scr[i][0] = 0;
+        init_row_scr[i][0] = 0;
     }
 
     // doesn't need to initialize the initial reference since no upper cell dependencey.
@@ -31,8 +31,10 @@ void SDTW::PE::Compute(char_t local_query_val,
                        score_vec_t &write_score,
                        tbp_t &write_traceback)
 {
+    // The paper mentions there is no reference deletion, which means that the there is no left dependency but only diag and up. 
+    // The RTL code uses left and diagonal is because possible the query and reference is transposed. 
     // find max from diagonal and left
-    write_score[0] = (diag_prev[0] < left_prev[0] ? diag_prev[0] : left_prev[0]) + abs(local_query_val - local_reference_val);
+    write_score[0] = (diag_prev[0] < up_prev[0] ? diag_prev[0] : up_prev[0]) + abs(local_query_val - local_reference_val);
 }
 
 void SDTW::UpdatePEMaximum(
@@ -46,10 +48,12 @@ void SDTW::UpdatePEMaximum(
     // Like SF, only do when QueryLength is multiple of PE_NUM, thus only let last PE findmax. 
     if (predicate[PE_NUM-1])
     {
-        if (scores[PE_NUM][LAYER_MAXIMIUM] > max[PE_NUM-1].score)
+        if (scores[PE_NUM][LAYER_MAXIMIUM] > max[PE_NUM-1].score && chunk_row_offset + PE_NUM == query_len)
         {
             // NOTE: If we just care about the score but doesn't care where does the score come from, then we doesn't need to update p_cols and ck index as well. 
             max[PE_NUM-1].score = scores[PE_NUM][LAYER_MAXIMIUM];
+            max[PE_NUM-1].p_col = p_cols;
+            max[PE_NUM-1].ck = ck_idx;
         }
     }
 }

kernels/sdtw/params.h

@@ -37,7 +37,7 @@
 // const int sf_score_width = 10 + std::ceil(std::log2(MAX_REFERENCE_LENGTH));
 
 typedef ap_uint<8> char_t;  // Sequence Alphabet
-typedef ap_uint<32> type_t;  // FIXME
+typedef ap_uint<64> type_t;  // FIXME
 typedef short idx_t;  // Indexing Type, could be much less than 32. ap_uint<8>
 typedef ap_uint<2> tbp_t;  // Traceback Pointer Type
 
@@ -49,11 +49,11 @@ typedef ap_uint<2> tbp_t;  // Traceback Pointer Type
 
 // Define Zero Value
 #define zero_fp ((type_t)0)
-#define ZERO_CHAR (char_t({0,0}))
+#define ZERO_CHAR (char_t(0))
 
 // Defien upper and lower bound for score type, aka type_t
 #define INF 16777216
-#define NINF -16777216
+#define NINF 0  // Because all the scores are non-negative, it make sense only if the NINF is 0. Assign negative score here result in underflow for type_t and incorrect alignment score. 
 
 // Legacy Debugger Configuration
 #define DEBUG_OUTPUT_PATH "/home/[email protected]/DP-HLS-Debug/global_affine/"

testbench/test_csim_sdtw.cpp

@@ -112,10 +112,6 @@ int main(){
         }
     }
 
-
-    // Allocate traceback streams
-    tbr_t tb_streams[MAX_REFERENCE_LENGTH + MAX_QUERY_LENGTH][N_BLOCKS];
-
     // initialize traceback starting coordinates
     idx_t tb_is[N_BLOCKS];
     idx_t tb_js[N_BLOCKS];
@@ -167,19 +163,10 @@ int main(){
         tb_query_lengths[i] = (int) tb_is[i];
         tb_reference_lengths[i] = (int) tb_js[i];
     }
-    tbr_t tb_streams_host[N_BLOCKS][MAX_REFERENCE_LENGTH + MAX_QUERY_LENGTH];
-    HostUtils::IO::SwitchDimension(tb_streams, tb_streams_host);
-    kernel_alignments = HostUtils::Sequence::ReconstructTracebackBlocks<tbr_t, N_BLOCKS, MAX_QUERY_LENGTH, MAX_REFERENCE_LENGTH>(
-        query_string_blocks, reference_string_blocks,
-        tb_query_lengths, tb_reference_lengths, 
-        tb_streams_host);
-
-    // Print kernel 0 traceback
+
+    // Print kernel alignment scores
     for (int i = 0; i < N_BLOCKS; i++) {
         cout << "Kernel " << i << " Traceback" << endl;
-        cout << "Kernel   Aligned Query    : " << kernel_alignments[0]["query"] << endl;
-        cout << "Kernel   Aligned Reference: " << kernel_alignments[0]["reference"] << endl;
+        cout << "Alignment Scores: " << scores[i] << endl;
     }
-
-
 }