diff --git a/ethosu/mlw_codec/mlw_codecmodule.c b/ethosu/mlw_codec/mlw_codecmodule.c
index 2c2fba2..1e13dd2 100644
--- a/ethosu/mlw_codec/mlw_codecmodule.c
+++ b/ethosu/mlw_codec/mlw_codecmodule.c
@@ -18,10 +18,137 @@
 
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
+#include <numpy/ndarrayobject.h>
 
 #include "mlw_decode.h"
 #include "mlw_encode.h"
 
+/* C extension wrapper for mlw_reorder_encode
+ *
+ * This method is exposed directly in python with the arguments with a
+ * prototype of the form:
+ *
+ * output = mlw_codec.reorder_encode(
+ *  ifm_ublock_depth,
+ *  ofm_ublock_depth,
+ *  input,
+ *  ofm_block_depth,
+ *  is_depthwise,
+ *  is_partkernel,
+ *  ifm_bitdepth,
+ *  decomp_h,
+ *  decomp_w,
+ *  verbose=0)
+ *
+ * output: bytearray
+ */
+
+static PyObject *
+method_reorder_encode (PyObject *self, PyObject *args)
+{
+    /* Object to hold the input integer list. */
+    int ifm_ublock_depth;
+    int ofm_ublock_depth;
+    PyObject *input_object;
+    int ofm_block_depth;
+    int is_depthwise;
+    int is_partkernel;
+    int ifm_bitdepth;
+    int decomp_h;
+    int decomp_w;
+
+    /* Object to hold the input verbosity integer, the verbose argument
+     * is optional so defaulted to 0.
+     */
+    int verbose = 0;
+
+    /* Arguments to the method are delivered as a tuple, unpack the
+     * tuple to get the individual arguments, note the second is
+     * optional.
+     */
+    if (!PyArg_ParseTuple(args, "iiOiiiiii|i",
+        &ifm_ublock_depth,
+        &ofm_ublock_depth,
+        &input_object,
+        &ofm_block_depth,
+        &is_depthwise,
+        &is_partkernel,
+        &ifm_bitdepth,
+        &decomp_h,
+        &decomp_w,
+        &verbose))
+        return NULL;
+
+    PyArrayObject* input_ndarray_object = PyArray_FROM_OTF(
+        input_object,
+        NPY_INT64,
+        NPY_ARRAY_ALIGNED);
+    if (input_ndarray_object == NULL)
+    {
+        return NULL;
+    }
+
+    if ((int)PyArray_NDIM(input_ndarray_object) < 4)
+    {
+        PyErr_SetString(PyExc_ValueError, "Invalid input shape");
+        return NULL;
+    }
+
+    int ofm_depth = (int)PyArray_DIM(input_ndarray_object, 0);
+    int kernel_height = (int)PyArray_DIM(input_ndarray_object, 1);
+    int kernel_width = (int)PyArray_DIM(input_ndarray_object, 2);
+    int ifm_depth = (int)PyArray_DIM(input_ndarray_object, 3);
+
+    int64_t* brick_weights = (int64_t*)PyArray_DATA(input_ndarray_object);
+    int brick_strides[4];
+    for (int i = 0; i < 4; i++)
+    {
+        brick_strides[i] = (int)PyArray_STRIDE(input_ndarray_object, i);
+    }
+    if ((unsigned)PyArray_ITEMSIZE(input_ndarray_object) != sizeof(int64_t))
+    {
+        PyErr_SetString(PyExc_ValueError, "Invalid input type");
+        return NULL;
+    }
+    uint8_t* output_buffer = NULL;
+    int padded_length;
+
+    int output_length = mlw_reorder_encode(
+        ifm_ublock_depth,
+        ofm_ublock_depth,
+        ofm_depth,
+        kernel_height,
+        kernel_width,
+        ifm_depth,
+        brick_strides,
+        brick_weights,
+        ofm_block_depth,
+        is_depthwise,
+        is_partkernel,
+        ifm_bitdepth,
+        decomp_h,
+        decomp_w,
+        &output_buffer,
+        &padded_length,
+        verbose);
+
+    if (output_buffer == NULL)
+    {
+        return PyErr_NoMemory();
+    }
+
+    PyObject *output_byte_array = PyByteArray_FromStringAndSize((char*)output_buffer, output_length);
+    PyObject *padded_length_obj = Py_BuildValue("i", padded_length);
+
+    /* Discard the output buffer */
+    mlw_free_outbuf(output_buffer);
+
+    PyObject* ret = PyTuple_Pack(2, output_byte_array, padded_length_obj);
+    Py_DECREF(output_byte_array);
+    Py_DECREF(padded_length_obj);
+    return ret;
+}
+
 /* C extension wrapper for mlw_encode
  *
  * This method is exposed directly in python with the arguments with a
@@ -63,6 +190,7 @@
    * for that purpose.
    */
   int16_t *input_buffer = (int16_t *) malloc(sizeof(int16_t *) * input_length);
+  uint8_t *output_buffer = NULL;
   if (input_buffer == NULL)
     return PyErr_NoMemory();
 
@@ -84,20 +212,13 @@
     return NULL;
   }
 
-  /* We don't know the output length required, we guess worst case,
-   * the mlw_encode call will do a resize (downwards) anyway.
-   */
-  uint8_t *output_buffer = (uint8_t *) malloc(input_length);
-  if (output_buffer == NULL)
-    return PyErr_NoMemory();
-
   int output_length = mlw_encode(input_buffer, input_length, &output_buffer, verbose);
 
   PyObject *output_byte_array = PyByteArray_FromStringAndSize ((char *) output_buffer, output_length);
 
   /* Discard the temporary input and output buffers.  */
   free (input_buffer);
-  free (output_buffer);
+  mlw_free_outbuf(output_buffer);
 
   return output_byte_array;
 }
@@ -163,6 +284,7 @@
 static PyMethodDef mlw_methods[] = {
     {"decode", method_decode, METH_VARARGS, "Python interface for decode"},
     {"encode", method_encode, METH_VARARGS, "Python interface for encode"},
+    {"reorder_encode", method_reorder_encode, METH_VARARGS, "Python interface for reorder and encode"},
     {NULL, NULL, 0, NULL}
 };
 
@@ -177,6 +299,9 @@
     mlw_methods
 };
 
-PyMODINIT_FUNC PyInit_mlw_codec(void) {
-    return PyModule_Create(&mlw_codecmodule);
+PyMODINIT_FUNC PyInit_mlw_codec(void)
+{
+    PyObject* ret = PyModule_Create(&mlw_codecmodule);
+    import_array();
+    return ret;
 }
diff --git a/ethosu/mlw_codec/mlw_encode.c b/ethosu/mlw_codec/mlw_encode.c
index 04afa3e..62e8360 100644
--- a/ethosu/mlw_codec/mlw_encode.c
+++ b/ethosu/mlw_codec/mlw_encode.c
@@ -819,12 +819,13 @@
 // Encode the given weight stream
 //      inbuf       uncompressed 9bit signed weights
 //      inbuf_size  number of weights
-//      outbuf      compressed bitstream, buffer is malloced
+//      outbuf      compressed bitstream, buffer is malloced within this function
 //      verbose     if non-zero, printf log
 // Return value is the size in bytes of the compressed output
 // Return -1 if error
 int mlw_encode( int16_t *inbuf, int inbuf_size, uint8_t **outbuf, int verbose) {
     int i;
+#ifndef NDEBUG
     // Range check
     for(i=0; i<inbuf_size; i++) {
         if (inbuf[i]<-255 || inbuf[i]>255) {
@@ -832,8 +833,10 @@
             return -1;
         }
     }
+#endif
 
     int bitbuf_size = inbuf_size*2+1024;
+    assert(*outbuf == NULL);
     *outbuf = malloc( bitbuf_size );
 
     // Analyse input data to find palette re-programming points
@@ -882,3 +885,265 @@
     if (outbuf)
         free(outbuf);
 }
+
+static int round_up_divide(int num, int den)
+{
+    return (num + den - 1) / den;
+}
+
+static int round_up(int num, int den)
+{
+    return round_up_divide(num, den) * den;
+}
+
+static int get_weight_cnt(
+    int ifm_ublock_depth,
+    int ofm_ublock_depth,
+    int ofm_depth,
+    int kernel_height,
+    int kernel_width,
+    int ifm_depth,
+    int ofm_block_depth,
+    int is_depthwise,
+    int is_partkernel,
+    int ifm_bitdepth,
+    int decomp_h,
+    int decomp_w)
+{
+    int ifm_block_depth = is_partkernel || ifm_bitdepth == 16 ? 16 : 32;
+    int subkernel_elements = decomp_w * decomp_h;
+    if (is_partkernel)
+    {
+        if (ifm_bitdepth == 16 && subkernel_elements % 2 != 0)
+        {
+            subkernel_elements = round_up(subkernel_elements, 2);
+        }
+        else if (ifm_bitdepth == 8 && subkernel_elements % 4 != 0)
+        {
+            subkernel_elements = round_up(subkernel_elements, 4);
+        }
+    }
+    else if (is_depthwise)
+    {
+        subkernel_elements = round_up(subkernel_elements, 4);
+    }
+    int clipped_ifm_block_depth = is_depthwise ? ifm_ublock_depth : ifm_block_depth;
+    int ifm_block_depth_outer = is_partkernel ? clipped_ifm_block_depth : 1;
+    int ifm_block_depth_inner = is_partkernel ? 1 : clipped_ifm_block_depth;
+
+    int input_length = 1;
+    input_length *= is_depthwise ? 1 : ifm_ublock_depth;
+    input_length *= ofm_ublock_depth;
+    input_length *= round_up_divide(ifm_block_depth_inner, ifm_ublock_depth);
+    input_length *= subkernel_elements;
+    input_length *= round_up_divide(ofm_block_depth, ofm_ublock_depth);
+    input_length *= round_up_divide(ifm_block_depth_outer, ifm_ublock_depth);
+    input_length *= round_up_divide(kernel_width, decomp_w);
+    input_length *= round_up_divide(kernel_height, decomp_h);
+    input_length *= round_up_divide(is_depthwise ? 1 : ifm_depth, ifm_block_depth);
+    input_length *= round_up_divide(ofm_depth, ofm_block_depth);
+
+    return input_length;
+}
+
+struct brick_buf_s
+{
+    uint8_t* buf;
+    int* strides;
+};
+typedef struct brick_buf_s brick_buf_t;
+
+static int16_t get_brick_weight(brick_buf_t* buf, int ofm_z, int wy, int wx, int ifm_z)
+{
+    uint8_t* p = buf->buf;
+
+    p += ofm_z * buf->strides[0];
+    p += wy * buf->strides[1];
+    p += wx * buf->strides[2];
+    p += ifm_z * buf->strides[3];
+
+    return *(int16_t*)p;
+}
+
+static int reorder(
+    int ifm_ublock_depth,
+    int ofm_ublock_depth,
+    int ofm_depth,
+    int kernel_height,
+    int kernel_width,
+    int ifm_depth,
+    int* strides,
+    void* inbuf,
+    int ofm_block_depth,
+    int is_depthwise,
+    int is_partkernel,
+    int ifm_bitdepth,
+    int decomp_h,
+    int decomp_w,
+    int16_t* weights)
+{
+    brick_buf_t brick_buf;
+    brick_buf.buf = inbuf;
+    brick_buf.strides = strides;
+
+    int ifm_block_depth = is_partkernel || ifm_bitdepth == 16 ? 16 : 32;
+    int weight_cnt = 0;
+    for (int ofm_block_z = 0; ofm_block_z < ofm_depth; ofm_block_z += ofm_block_depth)
+    {
+        int clipped_ofm_block_depth = min(ofm_block_depth, ofm_depth - ofm_block_z);
+        // IFM blocks required for the brick
+        for (int ifm_block_z = 0; ifm_block_z < (is_depthwise ? 1 : ifm_depth); ifm_block_z += ifm_block_depth)
+        {
+            int clipped_ifm_block_depth;
+            if (is_depthwise)
+            {
+                clipped_ifm_block_depth = ifm_ublock_depth;
+            }
+            else
+            {
+                clipped_ifm_block_depth = is_partkernel ?
+                    min(ifm_block_depth, ifm_depth - ifm_block_z) : ifm_block_depth;
+            }
+            // Weight decomposition
+            // Subkernel Splitting  (H)
+            for (int subkernel_y = 0; subkernel_y < kernel_height; subkernel_y += decomp_h)
+            {
+                int sub_height = min(kernel_height - subkernel_y, decomp_h);
+                // Subkernel splitting (W)
+                for (int subkernel_x = 0; subkernel_x < kernel_width; subkernel_x += decomp_w)
+                {
+                    int sub_width = min(kernel_width - subkernel_x, decomp_w);
+                    int subkernel_elements = sub_width * sub_height;
+                    // Part kernel first works across the kernel H/W and needs padding
+                    if (is_partkernel)
+                    {
+                        if (ifm_bitdepth == 16 && subkernel_elements % 2 != 0)
+                        {
+                            subkernel_elements = round_up(subkernel_elements, 2);
+                        }
+                        else if (ifm_bitdepth == 8 && subkernel_elements % 4 != 0)
+                        {
+                            subkernel_elements = round_up(subkernel_elements, 4);
+                        }
+                    }
+                    else if (is_depthwise)
+                    {
+                        subkernel_elements = round_up(subkernel_elements, 4);
+                    }
+                    int ifm_block_depth_outer = is_partkernel ? clipped_ifm_block_depth : 1;
+                    int ifm_block_depth_inner = is_partkernel ? 1 : clipped_ifm_block_depth;
+                    for (int ifm_ublk_outer = 0; ifm_ublk_outer < ifm_block_depth_outer; ifm_ublk_outer += ifm_ublock_depth)
+                    {
+                        // OFM Ublocks in OFM-block over depth
+                        for (int ofm_ublk = 0; ofm_ublk < clipped_ofm_block_depth; ofm_ublk += ofm_ublock_depth)
+                        {
+                            // HW Kernel element traversal - cannot be a H/W loop due to element
+                            // padding requirement on depthwise/part-kernel configurations
+                            for (int element = 0; element < subkernel_elements; element++)
+                            {
+                                int kx = element % sub_width;
+                                int ky = element / sub_width;
+                                // IFM Ublocks in IFM-block over depth (only 1 ublock if depthwise)
+                                // In case of part-kernel-first IFM Ublock traversal have already been handled
+                                // and this loop is ignored.
+                                for (int ifm_ublk_inner = 0; ifm_ublk_inner < ifm_block_depth_inner; ifm_ublk_inner += ifm_ublock_depth)
+                                {
+                                    // Feed OFM ublock elements
+                                    for (int ofm_ublock_z = 0; ofm_ublock_z < ofm_ublock_depth; ofm_ublock_z++)
+                                    {
+                                        // Source IFM ublock elements (only 1 element deep if depthwise)
+                                        for (int ifm_ublock_z = 0; ifm_ublock_z < (is_depthwise ? 1 : ifm_ublock_depth); ifm_ublock_z++)
+                                        {
+                                            // Source position within the current subkernel
+                                            int wx = subkernel_x + kx;
+                                            int wy = subkernel_y + ky;
+                                            // Source IFM/OFM slices
+                                            int ifm_ublk = ifm_ublk_inner + ifm_ublk_outer;
+                                            int ifm_z = ifm_block_z + ifm_ublk + ifm_ublock_z;
+                                            int ofm_z = ofm_block_z + ofm_ublk + ofm_ublock_z;
+                                            if ((ifm_z < ifm_depth) && (ofm_z < ofm_depth) && (ky < sub_height))
+                                            {
+                                                weights[weight_cnt] = get_brick_weight(&brick_buf, ofm_z, wy, wx, ifm_z);
+                                            }
+                                            weight_cnt++;
+                                        }
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    return weight_cnt;
+}
+
+// Reorder and encode the given weight stream
+// Return value is the size in bytes of the compressed output
+// Return -1 if error
+int mlw_reorder_encode(
+    int ifm_ublock_depth,
+    int ofm_ublock_depth,
+    int ofm_depth,
+    int kernel_height,
+    int kernel_width,
+    int ifm_depth,
+    int* brick_strides,
+    void* inbuf,
+    int ofm_block_depth,
+    int is_depthwise,
+    int is_partkernel,
+    int ifm_bitdepth,
+    int decomp_h,
+    int decomp_w,
+    uint8_t **outbuf, // *outbuf must be freed by caller
+    int* padded_length,
+    int verbose)
+{
+    /* Get an upper bound of the weight count */
+    int input_length = get_weight_cnt(
+        ifm_ublock_depth,
+        ofm_ublock_depth,
+        ofm_depth,
+        kernel_height,
+        kernel_width,
+        ifm_depth,
+        ofm_block_depth,
+        is_depthwise,
+        is_partkernel,
+        ifm_bitdepth,
+        decomp_h,
+        decomp_w);
+
+    int16_t* weights = (int16_t*)calloc(input_length, sizeof(int16_t));
+    if (weights == NULL)
+    {
+        return 0;
+    }
+
+    /* Reorder weights and update input_length */
+    input_length = reorder(
+        ifm_ublock_depth,
+        ofm_ublock_depth,
+        ofm_depth,
+        kernel_height,
+        kernel_width,
+        ifm_depth,
+        brick_strides,
+        inbuf,
+        ofm_block_depth,
+        is_depthwise,
+        is_partkernel,
+        ifm_bitdepth,
+        decomp_h,
+        decomp_w,
+        weights);
+
+    int output_length = mlw_encode(weights, input_length, outbuf, verbose);
+    free(weights);
+    *padded_length = input_length;
+
+    return output_length;
+}
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 7ce237c..9a1d5a1 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -15,7 +15,6 @@
 # limitations under the License.
 # Description:
 # Compresses and pads the weigths. It also calculates the scales and packs with the biases.
-import math
 from collections import namedtuple
 from typing import Tuple
 
@@ -93,18 +92,20 @@
 
     ifm_ublock = ArchitectureFeatures.accelerator_configs[accelerator].ifm_ublock
     ofm_ublock = ArchitectureFeatures.accelerator_configs[accelerator].ofm_ublock
-    raw_stream = generate_brick(
-        ifm_ublock=ifm_ublock,
-        ofm_ublock=ofm_ublock,
-        brick_weights=weights_volume,
-        ofm_block_depth=ofm_block_depth,
-        is_depthwise=is_depthwise,
-        is_partkernel=block_traversal == NpuBlockTraversal.PART_KERNEL_FIRST,
-        ifm_bitdepth=ifm_bitdepth,
-        dilation=dilation_xy,
+    decomp_h = ArchitectureFeatures.SubKernelMax.height // dilation_xy[0]
+    decomp_w = ArchitectureFeatures.SubKernelMax.width // dilation_xy[1]
+
+    return mlw_codec.reorder_encode(
+        ifm_ublock.depth,
+        ofm_ublock.depth,
+        weights_volume,
+        ofm_block_depth,
+        is_depthwise,
+        block_traversal == NpuBlockTraversal.PART_KERNEL_FIRST,
+        ifm_bitdepth,
+        decomp_h,
+        decomp_w,
     )
-    encoded_stream = encode(raw_stream)
-    return encoded_stream, len(raw_stream)
 
 
 def encode_bias(bias: np.int64, scale: int, shift: int):
@@ -180,111 +181,6 @@
         self.cache[wcc] = (tens_clone, unencoded_size)
 
 
-def encode(weight_stream):
-    if len(weight_stream) == 0:
-        return []
-    assert np.amin(weight_stream) >= -255
-    assert np.amax(weight_stream) <= 255
-
-    # Encode flattened signed weight stream
-    compressed = mlw_codec.encode(weight_stream)
-
-    # pad with 0xFF as needed so the length of the weight stream
-    # is a multiple of 16
-
-    while (len(compressed) % 16) != 0:
-        compressed.append(0xFF)
-
-    return compressed
-
-
-def generate_brick(
-    ifm_ublock, ofm_ublock, brick_weights, ofm_block_depth, is_depthwise, is_partkernel, ifm_bitdepth, dilation
-):
-
-    decomp_h = ArchitectureFeatures.SubKernelMax.height // dilation[0]
-    decomp_w = ArchitectureFeatures.SubKernelMax.width // dilation[1]
-    # Expect weights formatted OHWI
-    ofm_depth = brick_weights.shape[-4]
-    ifm_depth = brick_weights.shape[-1]
-    kernel_width = brick_weights.shape[-2]
-    kernel_height = brick_weights.shape[-3]
-    # IFM block depth
-    if is_partkernel or (ifm_bitdepth == 16):
-        # IFM block depth is always 16 for part-kernel-first
-        ifm_block_depth = 16
-    elif ifm_bitdepth == 8:
-        ifm_block_depth = 32
-    else:
-        assert False
-
-    stream = []
-
-    # Top level striping - OFM blocks in the entire brick's depth
-    for ofm_block_z in range(0, ofm_depth, ofm_block_depth):
-        clipped_ofm_block_depth = min(ofm_block_depth, ofm_depth - ofm_block_z)
-        # IFM blocks required for the brick
-        for ifm_block_z in range(0, (1 if is_depthwise else ifm_depth), ifm_block_depth):
-            if is_depthwise:
-                clipped_ifm_block_depth = ifm_ublock.depth
-            else:
-                clipped_ifm_block_depth = (
-                    min(ifm_block_depth, ifm_depth - ifm_block_z) if is_partkernel else ifm_block_depth
-                )
-            # Weight decomposition
-            # Subkernel Splitting  (H)
-            for subkernel_y in range(0, kernel_height, decomp_h):
-                sub_height = min(kernel_height - subkernel_y, decomp_h)
-                # Subkernel splitting (W)
-                for subkernel_x in range(0, kernel_width, decomp_w):
-                    sub_width = min(kernel_width - subkernel_x, decomp_w)
-                    subkernel_elements = sub_width * sub_height
-                    # Part kernel first works across the kernel H/W and needs padding
-                    if is_partkernel:
-                        if ifm_bitdepth == 16 and subkernel_elements % 2 != 0:
-                            subkernel_elements = int(math.ceil(subkernel_elements / 2) * 2)
-                        elif ifm_bitdepth == 8 and subkernel_elements % 4 != 0:
-                            subkernel_elements = int(math.ceil(subkernel_elements / 4) * 4)
-
-                    # Depthwise Conv requires multiple of 4 kernel elements in its weight block
-                    # this is different from normal conv which is considered "weights depth-first"
-                    elif is_depthwise:
-                        subkernel_elements = int(math.ceil(subkernel_elements / 4.0) * 4)
-
-                    ifm_block_depth_outer = clipped_ifm_block_depth if is_partkernel else 1
-                    ifm_block_depth_inner = 1 if is_partkernel else clipped_ifm_block_depth
-                    # IFM Ublocks in IFM-block over depth for part-kernel-first mode
-                    # For depth-first IFM Ublocks are traversed after subkernel elements so this loop is ignored.
-                    for ifm_ublk_outer in range(0, ifm_block_depth_outer, ifm_ublock.depth):
-                        # OFM Ublocks in OFM-block over depth
-                        for ofm_ublk in range(0, clipped_ofm_block_depth, ofm_ublock.depth):
-                            # HW Kernel element traversal - cannot be a H/W loop due to element
-                            # padding requirement on depthwise/part-kernel configurations
-                            for element in range(subkernel_elements):
-                                kx = element % sub_width
-                                ky = element // sub_width
-                                # IFM Ublocks in IFM-block over depth (only 1 ublock if depthwise)
-                                # In case of part-kernel-first IFM Ublock traversal have already been handled
-                                # and this loop is ignored.
-                                for ifm_ublk_inner in range(0, ifm_block_depth_inner, ifm_ublock.depth):
-                                    # Feed OFM ublock elements
-                                    for ofm_ublock_z in range(ofm_ublock.depth):
-                                        # Source IFM ublock elements (only 1 element deep if depthwise)
-                                        for ifm_ublock_z in range(1 if is_depthwise else ifm_ublock.depth):
-                                            # Source position within the current subkernel
-                                            wx = subkernel_x + kx
-                                            wy = subkernel_y + ky
-                                            # Source IFM/OFM slices
-                                            ifm_ublk = ifm_ublk_inner + ifm_ublk_outer
-                                            ifm_z = ifm_block_z + ifm_ublk + ifm_ublock_z
-                                            ofm_z = ofm_block_z + ofm_ublk + ofm_ublock_z
-                                            if (ifm_z >= ifm_depth) or (ofm_z >= ofm_depth) or (ky >= sub_height):
-                                                stream.append(0)
-                                            else:
-                                                stream.append(brick_weights[ofm_z][wy][wx][ifm_z])
-    return stream
-
-
 def core_deinterleave(hwio, core, ncores):
     # Put weights back into OHWI
     ohwi = np.transpose(hwio, (3, 0, 1, 2))
diff --git a/setup.py b/setup.py
index 8cf6195..d213743 100644
--- a/setup.py
+++ b/setup.py
@@ -18,6 +18,7 @@
 import os
 import re
 
+import numpy as np
 from setuptools import Extension
 from setuptools import find_namespace_packages
 from setuptools import setup
@@ -42,6 +43,7 @@
 mlw_module = Extension(
     "ethosu.mlw_codec",
     ["ethosu/mlw_codec/mlw_encode.c", "ethosu/mlw_codec/mlw_decode.c", "ethosu/mlw_codec/mlw_codecmodule.c"],
+    include_dirs=[np.get_include()],
 )
 
 setup(