[software] Fix systolic software parametrization

Author: sermazz

software/apps/systolic/conv_xqueue/main.c

@@ -56,7 +56,7 @@ void print_matrix(int32_t const *matrix, uint32_t num_rows,
 int main() {
   uint32_t core_id = mempool_get_core_id();
   uint32_t num_cores = mempool_get_core_count();
-  uint32_t tile_id = core_id / 4;
+  uint32_t tile_id = core_id / NUM_CORES_PER_TILE;
 
   // Initialize synchronization variables
   mempool_barrier_init(core_id);
@@ -66,8 +66,8 @@ int main() {
 
   // Allocate tile and core maps
   if (core_id == 0) {
-    tile_map = (uint32_t *)simple_malloc(num_cores * 4);
-    core_map = (uint32_t *)simple_malloc(num_cores * 4);
+    tile_map = (uint32_t *)simple_malloc(num_cores * sizeof(uint32_t));
+    core_map = (uint32_t *)simple_malloc(num_cores * sizeof(uint32_t));
   }
 
   // Wait for all cores

software/apps/systolic/matmul_xqueue/main.c

@@ -142,10 +142,10 @@ int main() {
     printf("> Initialize\n");
 
     // Print out tile mapping
-    //print_matrix((int32_t *)tile_mapping, SYSTOLIC_SIZE, SYSTOLIC_SIZE);
+    // print_matrix((int32_t *)tile_mapping, SYSTOLIC_SIZE, SYSTOLIC_SIZE);
 
     // Print out core mapping
-    //print_matrix((int32_t *)core_mapping, SYSTOLIC_SIZE, SYSTOLIC_SIZE);
+    // print_matrix((int32_t *)core_mapping, SYSTOLIC_SIZE, SYSTOLIC_SIZE);
 
     // Initialize systolic array
     systolic_init(tile_mapping, core_mapping);

software/runtime/runtime.h

@@ -11,10 +11,12 @@
 #include <stddef.h>
 #include <stdint.h>
 
-#define NUM_BANKS_PER_TILE NUM_CORES_PER_TILE * BANKING_FACTOR
+#define NUM_BANKS_PER_TILE NUM_CORES_PER_TILE *BANKING_FACTOR
 
 #define SIZEOF_INT32_T 4 // hack: result of sizeof(int32_t)
-#if ((NUM_CORES * SEQ_MEM_SIZE) < (NUM_CORES * STACK_SIZE + NUM_CORES * BANKING_FACTOR * SIZEOF_INT32_T * XQUEUE_SIZE))
+#if ((NUM_CORES * SEQ_MEM_SIZE) <                                              \
+     (NUM_CORES * STACK_SIZE +                                                 \
+      NUM_CORES * BANKING_FACTOR * SIZEOF_INT32_T * XQUEUE_SIZE))
 #error Sequential memory required for stack and Xqueues is bigger than the available one
 #endif
 

software/runtime/systolic/conv_xqueue.h

@@ -11,15 +11,17 @@
 /* TODO DESCRIPTION
  * TODO: LIMITATION NUM_COLS_Y >= 2 <=> NUM_COLS >= 4
  * TODO: COMPLETELY FIXED TO KERNEL SIZE OF 3
- *
- *
- *
- *
  */
 
+// FIXME: Does not work with GCC -O3 optimization as only a limited number of
+// outstanding queue ops are supported
+
 #include "alloc.h"
 #include "printf.h"
 
+// Queue dump macro
+dump(queue, 16);
+
 // Array of queue ptrs in row-major order (concatenated kernels)
 int32_t *queues_x_0[NUM_CORES];
 int32_t *queues_x_1[NUM_CORES];
@@ -49,23 +51,21 @@ void systolic_init(uint32_t const *tile_map, uint32_t const *core_map) {
   for (uint32_t i = 0; i < NUM_CORES; ++i) {
     tile_id = tile_map[i];
     core_id = core_map[i];
-    tile_offset = tile_id * 4 * SEQ_MEM_SIZE / 4;
-    core_offset = core_id % 4 * 4;
+    tile_offset = tile_id * NUM_CORES_PER_TILE * SEQ_MEM_SIZE / sizeof(int32_t);
+    core_offset = core_id % NUM_CORES_PER_TILE * 4;
     queues_x_0[i] = &__seq_start + tile_offset + core_offset + 0;
     queues_x_1[i] = &__seq_start + tile_offset + core_offset + 1;
   }
 
   // Print out queue addresses
-  // printf("queues_x_0\n");
+  // printf("\n[QUEUE] queues_x_0\n");
   // for (uint32_t i = 0; i < NUM_CORES; ++i) {
-  //   printf("%5d ", queues_x_0[i]);
-  // }
-  // printf("\n");
-  // printf("queues_x_1\n");
+  //  dump_queue((uint32_t)(queues_x_0[i]));
+  //}
+  // printf("\n[QUEUE] queues_x_1\n");
   // for (uint32_t i = 0; i < NUM_CORES; ++i) {
-  //   printf("%5d ", queues_x_1[i]);
-  // }
-  // printf("\n");
+  //  dump_queue((uint32_t)(queues_x_1[i]));
+  //}
 }
 
 void systolic_conv_front(const uint32_t num_rows, const uint32_t num_cols,
@@ -894,6 +894,19 @@ void systolic_conv_end(const uint32_t kernel_id, const uint32_t num_rows,
     // ----------
     // POPULATE 0
     // ----------
+
+    curr_x[2] = X[row * num_cols + 0];
+    asm volatile(
+        "q.pop.w %[curr_x_1], 0(%[queue_1]) \n\t"
+        "q.pop.w %[curr_x_0], 0(%[queue_0]) \n\t"
+        "p.mac %[acc_y_2],  \n\t"
+        : [curr_x_0] "=r"(curr_x[0]), [curr_x_1] "=r"(curr_x[1]),
+          [acc_y_2] "+r"(acc_y[2])
+        : [queue_0] "r"(queue_prev_x_0), [queue_1] "r"(queue_prev_x_1),
+          [curr_x_2] "r"(curr_x[2])[weight_0_0] "r"(weights[0][0]),
+          [weight_1_0] "r"(weights[1][0]), [weight_2_0] "r"(weights[2][0])
+        : "memory");
+
     // Pop and load x vector
     queue_pop(queue_prev_x_1, &curr_x[1]);
     curr_x[2] = X[row * num_cols + 0];