fix neureka bandwidth in 3x3 mode

Author: FrancescoConti

neureka/hal/neureka_task.c

@@ -166,14 +166,16 @@ void neureka_task_set_strides(neureka_task_t *task, const uint32_t k_in,
                                           .d2 = h_out_stride};
   task->data.cfg.output_stride = output_stride;
 
-  task->data.cfg.weights_stride.d0 = NEUREKA_WEIGHT_BANDWIDTH_BYTES;
   if (task->kernel_shape == 1) { // 1x1
+    task->data.cfg.weights_stride.d0 = NEUREKA_WEIGHT_BANDWIDTH_BYTES_1x1;
     task->data.cfg.weights_stride.d1 =
-        NEUREKA_WEIGHT_BANDWIDTH_BYTES * num_k_in;
+        NEUREKA_WEIGHT_BANDWIDTH_BYTES_1x1 * num_k_in;
   } else if (!task->depthwise) { // 3x3
+    task->data.cfg.weights_stride.d0 = NEUREKA_WEIGHT_BANDWIDTH_BYTES_3x3;
     task->data.cfg.weights_stride.d1 =
-        NEUREKA_WEIGHT_BANDWIDTH_BYTES * task->qw * num_k_in;
+        NEUREKA_WEIGHT_BANDWIDTH_BYTES_3x3 * task->qw * num_k_in;
   } else { // 3x3 depthwise
+    task->data.cfg.weights_stride.d0 = NEUREKA_WEIGHT_BANDWIDTH_BYTES_3x3;
     task->data.cfg.weights_stride.d1 = 0;
   }
   task->data.cfg.weights_stride.d2 = 0;

neureka/hal/neureka_task_defs.h

@@ -23,8 +23,10 @@
 
 /* ARCHITECTURE */
 
-#define NNX_NEUREKA_PE_H (6)
-#define NNX_NEUREKA_PE_W (6)
+#define NNX_NEUREKA_PE_H (4)
+#define NNX_NEUREKA_PE_W (4)
+#define NNX_NEUREKA_BANDWIDTH_1x1 (256)
+#define NNX_NEUREKA_BANDWIDTH_3x3 (288)
 
 #define NEUREKA_SUBTILE_INPUT_HEIGHT_1x1 (NNX_NEUREKA_PE_H)
 #define NEUREKA_SUBTILE_INPUT_WIDTH_1x1  (NNX_NEUREKA_PE_W)
@@ -34,12 +36,13 @@
 #define NEUREKA_SUBTILE_INPUT_WIDTH_3x3  (NNX_NEUREKA_PE_W+2)
 #define NEUREKA_SUBTILE_INPUT_CHANNEL_3x3 (32)
 
-#define NEUREKA_SUBTILE_OUTPUT_HEIGHT (4)
-#define NEUREKA_SUBTILE_OUTPUT_WIDTH (4)
+#define NEUREKA_SUBTILE_OUTPUT_HEIGHT (NNX_NEUREKA_PE_H)
+#define NEUREKA_SUBTILE_OUTPUT_WIDTH (NNX_NEUREKA_PE_W)
 #define NEUREKA_SUBTILE_OUTPUT_CHANNEL (32)
 
 #define NEUREKA_OUTPUT_BANDWIDTH_BYTES (32)
-#define NEUREKA_WEIGHT_BANDWIDTH_BYTES (32)
+#define NEUREKA_WEIGHT_BANDWIDTH_BYTES_1x1 (NNX_NEUREKA_BANDWIDTH_1x1/8)
+#define NEUREKA_WEIGHT_BANDWIDTH_BYTES_3x3 (NNX_NEUREKA_BANDWIDTH_3x3/8)
 
 /* TASK REGISTERS */
 

test/NeurekaMemoryLayout.py

@@ -78,7 +78,6 @@ def weightEncode(
             weight = weight.reshape(-1, height * width * cinSubtile)
             # Pad only the last dimension to weight bandwidth size
             # (-1, Weight Bandwidth)
-            print("DEBUG", weight.shape)
             weight = np.pad(
                 weight,
                 ((0, 0), (0, NeurekaMemoryLayout._WEIGHT_BANDWIDTH_3x3 - weight.shape[-1])),