diff --git a/CMakeLists.txt b/CMakeLists.txt
index ccc952d..324a091 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,70 +1,109 @@
 cmake_minimum_required(VERSION 3.18)
 
 project(pulp-nnx
-	VERSION 0.3.0
-	DESCRIPTION "Kernel library for PULP-based NN accelerators."
-	LANGUAGES C)
+  VERSION 0.3.0
+  DESCRIPTION "Kernel library for PULP-based NN accelerators."
+  LANGUAGES C)
 
-add_library(pulp-nnx STATIC)
+add_library(pulp-nnx INTERFACE)
 
-target_sources(pulp-nnx PRIVATE util/pulp_nnx_util.c util/hwpe.c)
-target_include_directories(pulp-nnx PUBLIC inc util)
+add_library(pulp-nnx-hal STATIC)
+
+target_sources(pulp-nnx-hal PRIVATE util/pulp_nnx_util.c util/hwpe.c)
+target_include_directories(pulp-nnx-hal PUBLIC inc util)
 
 option(USE_NE16 "Use the NE16 accelerator.")
 option(USE_NEUREKA "Use the N-EUREKA accelerator.")
 option(USE_NEUREKA_V2 "Use the N-EUREKA v2 accelerator.")
+option(ENABLE_BSP "Enable the build of the BSP for your chosen accelerator. Requires the PULP-SDK.")
 
 if (NOT ${USE_NE16} AND NOT ${USE_NEUREKA} AND NOT ${USE_NEUREKA_V2})
-	message(FATAL_ERROR "[PULP-NNX] No accelerator in use. Please set an appropriate USE_<acc> option.")
+  message(FATAL_ERROR "[PULP-NNX] No accelerator in use. Please set an appropriate USE_<acc> option.")
+endif()
+
+if(${ENABLE_BSP})
+  add_library(pulp-nnx-bsp STATIC)
 endif()
 
 if (${USE_NE16})
-	message(STATUS "[PULP-NNX] Using the NE16 accelerator.")
-	target_sources(pulp-nnx
-		PRIVATE
-			ne16/bsp/ne16_pulp_bsp.c
-			ne16/hal/ne16.c
-			ne16/hal/ne16_task.c
-			src/pulp_nnx_ne16.c
-		)
-	target_include_directories(pulp-nnx
-		PUBLIC
-			ne16/bsp
-			ne16/hal
-			ne16/gvsoc
-		)
+  message(STATUS "[PULP-NNX] Using the NE16 accelerator.")
+  target_sources(pulp-nnx-hal
+    PRIVATE
+    ne16/hal/ne16.c
+    ne16/hal/ne16_task.c
+  )
+  target_include_directories(pulp-nnx-hal
+    PUBLIC
+    ne16/hal
+    ne16/gvsoc
+  )
+  if(${ENABLE_BSP})
+  target_sources(pulp-nnx-bsp
+    PRIVATE
+    ne16/bsp/ne16_pulp_bsp.c
+    src/pulp_nnx_ne16.c
+  )
+  target_include_directories(pulp-nnx-bsp
+    PUBLIC
+    ne16/bsp
+  )
+  endif()
+
 endif()
 
 if (${USE_NEUREKA})
-	message(STATUS "[PULP-NNX] Using the N-EUREKA accelerator.")
-	target_sources(pulp-nnx
-		PRIVATE
-			neureka/bsp/neureka_siracusa_bsp.c
-			neureka/hal/neureka.c
-			neureka/hal/neureka_task.c
-			src/pulp_nnx_neureka.c
-		)
-	target_include_directories(pulp-nnx
-		PUBLIC
-			neureka/bsp
-			neureka/hal
-			neureka/gvsoc
-		)
+  message(STATUS "[PULP-NNX] Using the N-EUREKA accelerator.")
+  target_sources(pulp-nnx-hal
+    PRIVATE
+    neureka/hal/neureka.c
+    neureka/hal/neureka_task.c
+  )
+  target_include_directories(pulp-nnx-hal
+    PUBLIC
+    neureka/hal
+    neureka/gvsoc
+  )
+  if(${ENABLE_BSP})
+  target_sources(pulp-nnx-bsp
+    PRIVATE
+    neureka/bsp/neureka_siracusa_bsp.c
+    src/pulp_nnx_neureka.c
+  )
+  target_include_directories(pulp-nnx-bsp
+    PUBLIC
+    neureka/bsp
+  )
+  endif()
+
 endif()
 
 if (${USE_NEUREKA_V2})
-	message(STATUS "[PULP-NNX] Using the N-EUREKA v2 accelerator.")
-	target_sources(pulp-nnx
-		PRIVATE
-			neureka_v2/bsp/neureka_v2_siracusa_bsp.c
-			neureka_v2/hal/neureka_v2.c
-			neureka_v2/hal/neureka_v2_task.c
-			src/pulp_nnx_neureka_v2.c
-		)
-	target_include_directories(pulp-nnx
-		PUBLIC
-			neureka_v2/bsp
-			neureka_v2/hal
-			neureka_v2/gvsoc
-		)
+  message(STATUS "[PULP-NNX] Using the N-EUREKA v2 accelerator.")
+  target_sources(pulp-nnx-hal
+    PRIVATE
+    neureka_v2/hal/neureka_v2.c
+    neureka_v2/hal/neureka_v2_task.c
+  )
+  target_include_directories(pulp-nnx-hal
+    PUBLIC
+    neureka_v2/hal
+    neureka_v2/gvsoc
+  )
+  if(${ENABLE_BSP})
+  target_sources(pulp-nnx-bsp
+    PRIVATE
+    neureka_v2/bsp/neureka_v2_pulp_bsp.c
+    src/pulp_nnx_neureka_v2.c
+  )
+  target_include_directories(pulp-nnx-bsp
+    PUBLIC
+    neureka_v2/bsp
+  )
+  endif()
+
+endif()
+
+target_link_libraries(pulp-nnx INTERFACE pulp-nnx-hal)
+if(${ENABLE_BSP})
+  target_link_libraries(pulp-nnx INTERFACE pulp-nnx-bsp)
 endif()
diff --git a/inc/pulp_nnx_neureka.h b/inc/pulp_nnx_neureka.h
index fea4bb4..0ddef59 100644
--- a/inc/pulp_nnx_neureka.h
+++ b/inc/pulp_nnx_neureka.h
@@ -19,7 +19,6 @@
  */
 
 #include "neureka.h"
-#include "neureka_siracusa_bsp.h"
 #include "neureka_task.h"
 #include <stdint.h>
 
diff --git a/neureka/hal/neureka_task.c b/neureka/hal/neureka_task.c
index 9a311dc..18939b8 100644
--- a/neureka/hal/neureka_task.c
+++ b/neureka/hal/neureka_task.c
@@ -165,14 +165,17 @@ 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 / 8) * task->qw * 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;
diff --git a/neureka/hal/neureka_task_defs.h b/neureka/hal/neureka_task_defs.h
index fa08289..b48f78e 100644
--- a/neureka/hal/neureka_task_defs.h
+++ b/neureka/hal/neureka_task_defs.h
@@ -21,22 +21,38 @@
 #ifndef __NEUREKA_DEFS_H__
 #define __NEUREKA_DEFS_H__
 
-/* ARHITECTURE */
-
-#define NEUREKA_SUBTILE_INPUT_HEIGHT_1x1 (6)
-#define NEUREKA_SUBTILE_INPUT_WIDTH_1x1 (6)
+/* ARCHITECTURE */
+// The definitions wrapped in #ifndefs can be overwritten with compiler flags
+// for different parametrizations of the Neureka architecture
+#ifndef NNX_NEUREKA_PE_H
+#define NNX_NEUREKA_PE_H (6)
+#endif
+#ifndef NNX_NEUREKA_PE_W
+#define NNX_NEUREKA_PE_W (6)
+#endif
+#define NNX_NEUREKA_BANDWIDTH_1x1 (256)
+
+#ifndef NNX_NEUREKA_BANDWIDTH_3x3
+#define NNX_NEUREKA_BANDWIDTH_3x3 (256)
+#endif
+
+#define NEUREKA_SUBTILE_INPUT_HEIGHT_1x1 (NNX_NEUREKA_PE_H)
+#define NEUREKA_SUBTILE_INPUT_WIDTH_1x1 (NNX_NEUREKA_PE_W)
 #define NEUREKA_SUBTILE_INPUT_CHANNEL_1x1 (32)
 
-#define NEUREKA_SUBTILE_INPUT_HEIGHT_3x3 (8)
-#define NEUREKA_SUBTILE_INPUT_WIDTH_3x3 (8)
+#define NEUREKA_SUBTILE_INPUT_HEIGHT_3x3 (NNX_NEUREKA_PE_H + 2)
+#define NEUREKA_SUBTILE_INPUT_WIDTH_3x3 (NNX_NEUREKA_PE_W + 2)
+#ifndef NEUREKA_SUBTILE_INPUT_CHANNEL_3x3
 #define NEUREKA_SUBTILE_INPUT_CHANNEL_3x3 (28)
+#endif
 
-#define NEUREKA_SUBTILE_OUTPUT_HEIGHT (6)
-#define NEUREKA_SUBTILE_OUTPUT_WIDTH (6)
+#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 */
 
diff --git a/util/hwpe.c b/util/hwpe.c
index 0430081..1f99c1f 100644
--- a/util/hwpe.c
+++ b/util/hwpe.c
@@ -30,33 +30,34 @@
 #define HWPE_SWSYNC 6
 #define HWPE_TASK_REG_OFFSET 8
 
-inline void hwpe_reg_write(hwpe_dev_t *dev, int reg, uint32_t value) {
+inline void hwpe_reg_write(const hwpe_dev_t *dev, int reg, uint32_t value) {
   dev->base_addr[reg] = value;
 }
 
-inline uint32_t hwpe_reg_read(hwpe_dev_t *dev, int reg) {
+inline uint32_t hwpe_reg_read(const hwpe_dev_t *dev, int reg) {
   return dev->base_addr[reg];
 }
 
-inline void hwpe_task_reg_write(hwpe_dev_t *dev, int reg, uint32_t value) {
+inline void hwpe_task_reg_write(const hwpe_dev_t *dev, int reg,
+                                uint32_t value) {
   hwpe_reg_write(dev, HWPE_TASK_REG_OFFSET + reg, value);
 }
 
-inline uint32_t hwpe_task_reg_read(hwpe_dev_t *dev, int reg) {
+inline uint32_t hwpe_task_reg_read(const hwpe_dev_t *dev, int reg) {
   return hwpe_reg_read(dev, HWPE_TASK_REG_OFFSET + reg);
 }
 
-void hwpe_soft_clear(hwpe_dev_t *dev) {
+void hwpe_soft_clear(const hwpe_dev_t *dev) {
   hwpe_reg_write(dev, HWPE_SOFT_CLEAR, 0);
   for (volatile int i = 0; i < 10; i++)
     ;
 }
 
-uint32_t hwpe_task_queue_status(hwpe_dev_t *dev) {
+uint32_t hwpe_task_queue_status(const hwpe_dev_t *dev) {
   return hwpe_reg_read(dev, HWPE_STATUS);
 }
 
-int hwpe_task_queue_acquire_task(hwpe_dev_t *dev, uint8_t *id) {
+int hwpe_task_queue_acquire_task(const hwpe_dev_t *dev, uint8_t *id) {
   uint32_t read_value = (int32_t)hwpe_reg_read(dev, HWPE_ACQUIRE);
   if (read_value >= 256) {
     return 1;
@@ -66,20 +67,21 @@ int hwpe_task_queue_acquire_task(hwpe_dev_t *dev, uint8_t *id) {
   }
 }
 
-void hwpe_task_queue_write_task(hwpe_dev_t *dev, uint32_t *data, int len) {
+void hwpe_task_queue_write_task(const hwpe_dev_t *dev, uint32_t *data,
+                                int len) {
   for (int i = 0; i < len; i++) {
     hwpe_task_reg_write(dev, i, data[i]);
   }
 }
 
-void hwpe_task_queue_release_and_run(hwpe_dev_t *dev) {
+void hwpe_task_queue_release_and_run(const hwpe_dev_t *dev) {
   hwpe_reg_write(dev, HWPE_TRIGGER, 0);
 }
 
-void hwpe_task_queue_release(hwpe_dev_t *dev) {
+void hwpe_task_queue_release(const hwpe_dev_t *dev) {
   hwpe_reg_write(dev, HWPE_TRIGGER, 1);
 }
 
-uint8_t hwpe_last_task_id(hwpe_dev_t *dev) {
+uint8_t hwpe_last_task_id(const hwpe_dev_t *dev) {
   return (uint8_t)hwpe_reg_read(dev, HWPE_RUNNING_JOB);
 }
diff --git a/util/hwpe.h b/util/hwpe.h
index 52bf912..7b7f65c 100644
--- a/util/hwpe.h
+++ b/util/hwpe.h
@@ -28,16 +28,16 @@ typedef struct hwpe_dev_t {
   volatile uint32_t *base_addr;
 } hwpe_dev_t;
 
-void hwpe_reg_write(hwpe_dev_t *dev, int reg, uint32_t value);
-uint32_t hwpe_reg_read(hwpe_dev_t *dev, int reg);
-void hwpe_task_reg_write(hwpe_dev_t *dev, int reg, uint32_t value);
-uint32_t hwpe_task_reg_read(hwpe_dev_t *dev, int reg);
-void hwpe_soft_clear(hwpe_dev_t *dev);
-uint32_t hwpe_task_queue_status(hwpe_dev_t *dev);
-int hwpe_task_queue_acquire_task(hwpe_dev_t *dev, uint8_t *id);
-void hwpe_task_queue_write_task(hwpe_dev_t *dev, uint32_t *data, int len);
-void hwpe_task_queue_release_and_run(hwpe_dev_t *dev);
-void hwpe_task_queue_release(hwpe_dev_t *dev);
-uint8_t hwpe_last_task_id(hwpe_dev_t *dev);
+void hwpe_reg_write(const hwpe_dev_t *dev, int reg, uint32_t value);
+uint32_t hwpe_reg_read(const hwpe_dev_t *dev, int reg);
+void hwpe_task_reg_write(const hwpe_dev_t *dev, int reg, uint32_t value);
+uint32_t hwpe_task_reg_read(const hwpe_dev_t *dev, int reg);
+void hwpe_soft_clear(const hwpe_dev_t *dev);
+uint32_t hwpe_task_queue_status(const hwpe_dev_t *dev);
+int hwpe_task_queue_acquire_task(const hwpe_dev_t *dev, uint8_t *id);
+void hwpe_task_queue_write_task(const hwpe_dev_t *dev, uint32_t *data, int len);
+void hwpe_task_queue_release_and_run(const hwpe_dev_t *dev);
+void hwpe_task_queue_release(const hwpe_dev_t *dev);
+uint8_t hwpe_last_task_id(const hwpe_dev_t *dev);
 
 #endif // !__HWPE_H__