Chain example updated to show speedup

Author: Rajat Sarkari

cpp_kernels/kernel_chain/details.rst

@@ -26,3 +26,20 @@ and simple kernel.
 
    #pragma HLS INTERFACE s_axilite port = return bundle = control
    #pragma HLS INTERFACE ap_ctrl_chain port = return bundle = control
+
+Following is the real log reported while running the design on U200
+platform with 10 iterations:
+
+::
+
+    --------------------------------------------------
+                    Performance Summary                  
+    --------------------------------------------------
+    -------------------------+-------------------------
+     Kernel(10 iterations)   |    Wall-Clock Time (s)  
+    -------------------------+-------------------------
+     krnl_chain_mmult        | 0.00305935              
+     krnl_simple_mmult       | 0.00534336              
+    -------------------------+-------------------------
+     Speedup:                | 1.74657	                
+    -------------------------+-------------------------

cpp_kernels/kernel_chain/src/host.cpp

@@ -18,7 +18,7 @@
 #include <array>
 #include <iostream>
 
-#define MAT_DIM 8
+#define MAT_DIM 32
 #define MAT_SIZE MAT_DIM* MAT_DIM
 #define NUM_TIMES 10
 ////////////////////UTILITY FUNCTION///////////////

cpp_kernels/kernel_chain/src/krnl_chain_mmult.cpp

@@ -45,16 +45,16 @@ void krnl_chain_mmult(int* a, int* b, int* c, int* d, int* output, int dim) {
     int tmp = dim;
     strm_ctrl_trans1.write(tmp);
 
-#pragma HLS STREAM variable = strm_ctrl_trans1 depth = 2
-#pragma HLS STREAM variable = strm_ctrl_trans2 depth = 2
-#pragma HLS STREAM variable = strm_ctrl_trans3 depth = 2
-#pragma HLS STREAM variable = strm_ctrl_trans4 depth = 2
-#pragma HLS STREAM variable = strm_ctrl_trans5 depth = 2
+#pragma HLS STREAM variable = strm_ctrl_trans1 depth = 1024
+#pragma HLS STREAM variable = strm_ctrl_trans2 depth = 1024
+#pragma HLS STREAM variable = strm_ctrl_trans3 depth = 1024
+#pragma HLS STREAM variable = strm_ctrl_trans4 depth = 1024
+#pragma HLS STREAM variable = strm_ctrl_trans5 depth = 1024
 
-#pragma HLS STREAM variable = strm_a depth = 64
-#pragma HLS STREAM variable = strm_b depth = 64
-#pragma HLS STREAM variable = strm_c depth = 64
-#pragma HLS STREAM variable = strm_d depth = 64
+#pragma HLS STREAM variable = strm_a depth = 4096
+#pragma HLS STREAM variable = strm_b depth = 4096
+#pragma HLS STREAM variable = strm_c depth = 4096
+#pragma HLS STREAM variable = strm_d depth = 4096
 
 #pragma HLS DATAFLOW
 

cpp_kernels/kernel_chain/src/krnl_mmult.hpp

@@ -16,13 +16,14 @@
 #ifndef __KRNL_MMULT__
 #define __KRNL_MMULT__
 #define DWIDTH 32
-#define MAT_DIM 8
+#define MAT_DIM 32
 // typedef ap_axis<DWIDTH, 0, 0, 0> pkt;
 typedef int pkt;
 // Template to avoid signature conflict in sw_em
 template <int DUMMY = 0>
 void mm2s(int* a, hls::stream<pkt>& strm_a, hls::stream<int>& strm_ctrl_trans1, hls::stream<int>& strm_ctrl_trans2) {
     int dim = strm_ctrl_trans1.read();
+    strm_ctrl_trans2.write(dim);
     int size = dim * dim;
 
 // Auto-pipeline is going to apply pipeline to this loop
@@ -33,7 +34,6 @@ void mm2s(int* a, hls::stream<pkt>& strm_a, hls::stream<int>& strm_ctrl_trans1,
 
         strm_a.write(p1);
     }
-    strm_ctrl_trans2.write(dim);
 }
 
 // Template to avoid signature conflict in sw_em
@@ -44,6 +44,7 @@ void mmult(hls::stream<pkt>& strm_a,
            hls::stream<pkt>& strm_out,
            hls::stream<int>& strm_ctrl_trans3) {
     int dim = strm_ctrl_trans2.read();
+    strm_ctrl_trans3.write(dim);
     int size = dim * dim;
 
     int buf_a[MAT_DIM][MAT_DIM];
@@ -96,7 +97,6 @@ void mmult(hls::stream<pkt>& strm_a,
 
         strm_out.write(temp);
     }
-    strm_ctrl_trans3.write(dim);
 }
 
 // Template to avoid signature conflict in sw_em