support for back2back aha test with sparse tile pipelining

Author: kalhankoul96

copy_formatted.py

@@ -83,7 +83,7 @@ def find_tensor(tensor_list, tile_id):
 
 #os.chdir("SPARSE_TESTS")
 
-if app_name == "matmul_ijk" and mode_ns == False:
+if "matmul_ijk" in app_name and mode_ns == False:
     for b in b_tensors:
         for c in c_tensors:
             tile_str = str(app_name) + "-" + str(test) + "_tile" + str(tile)
@@ -180,7 +180,7 @@ def find_tensor(tensor_list, tile_id):
                 tile = tile + 1
 
 
-if app_name == "matmul_ijk" and mode_ns == True:
+if "matmul_ijk" in app_name and mode_ns == True:
     for b in b_ns_tensors:
         for c in c_ns_tensors:
             # mode 2 == mode 0, mode 3 == mode 1

tests/test_app/Makefile

@@ -144,7 +144,7 @@ compile:
 
 .PHONY: run
 run:
-	@rm -rf *.txt
+	@rm -rf *.txt; rm -rf ../../SPARSE_TESTS/tensor*
 	$(RUN) $(DUMP_ARGS) $(RUN_ARGS) $(APP_ARGS)
 
 .PHONY: sim

tests/test_app/lib/map.c

@@ -57,41 +57,53 @@ void update_bs_configuration(struct BitstreamInfo *bs_info) {
 int glb_map(void *kernel_, int dpr_enabled) {
     struct KernelInfo *kernel = kernel_;
     int num_groups = kernel->num_groups;
-    printf("number of groups: %d\n", num_groups);
-
-    // This is just greedy algorithm to schedule applications
-    // TODO: Need a better way to schedule kernels
-    int group_start = -1;
-    for (int i = 0; i < monitor.num_groups; i++) {
-        int success = 0;
-        for (int j = 0; j < num_groups; j++) {
-            // if the number of groups required exceeds the hardware resources, then
-            // fail
-            if ((i + j) >= monitor.num_groups) {
-                success = -1;
-                break;
-            }
-            if (monitor.groups[i + j] != 0) {
-                break;
-            }
-            if (j == (num_groups - 1)) {
-                group_start = i;
-                success = 1;
-            }
-        }
-        if (success != 0) break;
-    }
-    // no available group
-    if (group_start == -1) {
+    printf("number of groups: %d\n",  kernel->num_groups);
+    printf("number of inputs: %d\n",  kernel->num_inputs);
+    printf("number of outputs: %d\n",  kernel->num_outputs);
+
+    printf("monitor.num_groups: %d\n", monitor.num_groups);
+
+    if (num_groups > monitor.num_groups) {
         printf("Application does not fit on array. Possible error CGRA too small, applications overlapping\n");
         return 0;
     }
 
-    for (int i = group_start; i < group_start + num_groups; i++) {
-        monitor.groups[i] = 1;
-    }
+    int group_start;
+    // DPR needs to check if enough array space
+    if (dpr_enabled == 1) {
+        // This is just greedy algorithm to schedule applications
+        // TODO: Need a better way to schedule kernels
+        group_start = -1;
+        for (int i = 0; i < monitor.num_groups; i++) {
+            int success = 0;
+            for (int j = 0; j < num_groups; j++) {
+                // if the number of groups required exceeds the hardware resources, then
+                // fail
+                if ((i + j) >= monitor.num_groups) {
+                    success = -1;
+                    break;
+                }
+                if (monitor.groups[i + j] != 0) {
+                    break;
+                }
+                if (j == (num_groups - 1)) {
+                    group_start = i;
+                    success = 1;
+                }
+            }
+            if (success != 0) break;
+        }
+        // no available group
+        if (group_start == -1) {
+            printf("Combination of Applications does not fit on array. Possible error CGRA too small, applications overlapping\n");
+            return 0;
+        }
+
+        for (int i = group_start; i < group_start + num_groups; i++) {
+            monitor.groups[i] = 1;
+        }
     // DPR is not enabled so just clear the group start
-    if(dpr_enabled == 0){
+    } else {
         group_start = 0;
     }
     kernel->group_start = group_start;

tests/test_app/lib/parser.h

@@ -7,9 +7,9 @@
 #define MAX_NUM_IO 16
 #define MAX_NUM_IO_TILES 16
 #define BUFFER_SIZE 1024
-#define MAX_NUM_KERNEL 16
+#define MAX_NUM_KERNEL 128
 #define MAX_JSON_FIELDS 2048
-#define MAX_CONFIG 40
+#define MAX_CONFIG 8192
 
 #define GET_BS_INFO(info) struct BitstreamInfo *bs_info = (struct BitstreamInfo *)info
 #define GET_KERNEL_INFO(info) struct KernelInfo *kernel_info = (struct KernelInfo *)info

tests/test_app/map.c

@@ -8,7 +8,7 @@
 import "DPI-C" function int initialize_monitor(int num_cols);
 
 program garnet_test #(
-    parameter int MAX_NUM_APPS = 3
+    parameter int MAX_NUM_APPS = 1000
 ) (
     input logic clk,
     reset,