xmos-ai-tools 1.3.2.dev80__py3-none-macosx_10_15_universal2.whl

xmos_ai_tools/runtime/include/lib_nn/src/asm/asm_constants.h

@@ -0,0 +1,41 @@
+// Copyright 2020-2021 XMOS LIMITED.
+// This Software is subject to the terms of the XMOS Public Licence: Version 1.
+#pragma once
+
+#ifndef __ASSEMBLER__
+
+#include <stdint.h>
+
+#include "xs3_vpu.h"
+
+typedef struct {
+  // Word offset = 0
+  int16_t vec_0x007F[VPU_INT8_ACC_PERIOD];
+  // Word offset = 8
+  int8_t vec_0x01[VPU_INT8_ACC_PERIOD];
+  // Word offset = 12
+  int16_t vec_0x0002[VPU_INT8_ACC_PERIOD];
+  // Word offset = 20
+  int8_t vec_0x80[VPU_INT8_EPV];
+  // Word offset = 28
+
+} vpu_constants_t;
+
+extern const vpu_constants_t vpu_vects;
+
+extern const uint32_t vpu_vect_zero[VPU_INT32_EPV];
+extern const int16_t vpu_vect_0x007F[VPU_INT16_EPV];
+extern const int8_t vpu_vect_0x01[VPU_INT8_EPV];
+extern const int8_t vpu_vect_0x02[VPU_INT8_EPV];
+extern const int8_t vpu_vect_0x80[VPU_INT8_EPV];
+
+#endif  // __ASSEMBLER__
+
+#define VPU_MODE_32BIT 0x0000
+#define VPU_MODE_16BIT 0x0100
+#define VPU_MODE_8BIT 0x0200
+
+#define VPU_VEC_0x007F (0)
+#define VPU_VEC_0x01 (8)
+#define VPU_VEC_0x0002 (12)
+#define VPU_VEC_0x80 (20)

xmos_ai_tools/runtime/include/lib_nn/src/asm/window_op_plan.h

@@ -0,0 +1,25 @@
+// Copyright 2020-2021 XMOS LIMITED.
+// This Software is subject to the terms of the XMOS Public Licence: Version 1.
+
+#ifndef WINDOW_OP_PLAN_H_
+#define WINDOW_OP_PLAN_H_
+
+#define WOP_OUTPUT_ROWS (0)
+#define WOP_OUTPUT_COLS (1)
+#define WOP_OUTPUT_CHANS (2)
+#define WOP_WINDOW_ROWS (3)
+#define WOP_WINDOW_COLS (4)
+#define WOP_START_STRIDE_X (5)
+#define WOP_START_STRIDE_Y (6)
+#define WOP_INNER_STRIDE_VERT (7)
+#define WOP_INNER_STRIDE_HORI (8)
+#define WOP_OUTER_STRIDE_VERT_X (9)
+#define WOP_OUTER_STRIDE_VERT_Y (10)
+#define WOP_OUTER_STRIDE_HORI_X (11)
+#define WOP_OUTER_STRIDE_HORI_Y (12)
+#define WOP_CHAN_STRIDE_X (13)
+#define WOP_CHAN_STRIDE_Y (14)
+
+#define WOP_WORDS (15)
+
+#endif  // WINDOW_OP_PLAN_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/api/fast_flash.h

@@ -0,0 +1,47 @@
+#ifndef _FAST_FLASH_H_
+#define _FAST_FLASH_H_
+
+#include <quadflash.h>
+
+/** Fast flash library.
+ * Before calling any of the functions in here, lib_quad_flash must be initialised as normal by using
+ * fl_connectToDevice(qspi, flash_spec, n_flash_spec).
+ * After that, a call to fast_flash_init shall be made.
+ * After that, a sequence of calls to fast_flash_read can be made.
+ *
+ * The data partition must start with the following 32 bytes: **NOTE: REMOVE THE +4 in fast_flash_init**
+ *
+ *   0xff, 0x00, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f, 0x0f,
+ *   0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00,
+ *   0x31, 0xf7, 0xce, 0x08, 0x31, 0xf7, 0xce, 0x08,
+ *   0x9c, 0x63, 0x9c, 0x63, 0x9c, 0x63, 0x9c, 0x63
+ * 
+ * This pattern is designed to create maximum difficulties electrically and is used
+ * to calibrate the electrical settings. Note that this pattern must be nibble reversed
+ * before being written to flash; just like all other data.
+ * The rest of the data partition can be used as normal
+ */
+
+/** Function that initialises the fast_flash library
+ *
+ * \param      qspi        ports that connect to flash
+ *
+ * \returns    a negative value of -1..-5 if the window is too small (size 0..4)
+ *             zero if successful
+ */
+int fast_flash_init(fl_QSPIPorts &qspi);
+
+/** Function that reads a sequential set of bytes from memory.
+ * This function assumes that nibbles have been reversed ((x << 4) & 0xf0 | (x >> 4) & 0x0f)
+ * before the data was written to flash.
+ * Note that reading 32 bytes from offset 0 shall yield the special pattern above.
+ *
+ * \param      qspi        ports that connect to flash
+ * \param      addr        address in flash data segment
+ * \param      word_count  Number of words to read
+ * \param      read_data   array to store data in to.
+ * \param      c_out_data  optional channel end over which data is out() instead.
+ */
+void fast_flash_read(fl_QSPIPorts &qspi, unsigned addr, unsigned word_count, unsigned read_data[], chanend ?c_data_out);
+
+#endif

xmos_ai_tools/runtime/include/lib_tflite_micro/api/inference_engine.h

@@ -0,0 +1,218 @@
+// Copyright (c) 2020, XMOS Ltd, All rights reserved
+#ifndef INFERENCE_ENGINE_H_
+#define INFERENCE_ENGINE_H_
+
+#ifndef __XC__
+#define UNSAFE /**/
+#else
+#define UNSAFE unsafe
+#endif
+
+#if !defined(XTFLM_DISABLED)
+
+#if defined(__xtflm_conf_h_exists__)
+#include "xtflm_conf.h"
+#else
+
+#ifndef XTFLM_OPERATORS
+#define XTFLM_OPERATORS 10
+#endif
+
+#endif
+
+#include "tensorflow/lite/c/c_api_types.h"
+#include "xcore_config.h"
+
+#ifdef __cplusplus
+
+#include "tensorflow/lite/micro/micro_mutable_op_resolver.h"
+#include "xcore_device_memory.h"
+#include "xcore_error_reporter.h"
+#include "xcore_interpreter.h"
+#include "xcore_ops.h"
+#include "xcore_profiler.h"
+
+/** Structure that contains all the TensorFlowLite for Micro objects that must
+ * be allocated to create an interpreter. One of these structures has to be
+ * allocated for each inference engine. This structure contains C++ objects, and
+ * must therefore be allocated inside a C++ source file.
+ */
+struct tflite_micro_objects {
+  tflite_micro::micro::xcore::XCoreErrorReporter error_reporter;
+  tflite_micro::micro::xcore::XCoreProfiler xcore_profiler;
+  uint64_t interpreter_buffer
+      [(sizeof(tflite_micro::micro::xcore::XCoreInterpreter) + sizeof(uint64_t) - 1) /
+       sizeof(uint64_t)]; // This needs to be aligned on a double word boundary
+  tflite_micro::MicroMutableOpResolver<XTFLM_OPERATORS> resolver;
+
+  tflite_micro::micro::xcore::XCoreInterpreter *interpreter;
+  const tflite_micro::Model *model;
+};
+#endif
+
+#endif
+
+// Opaque definition for the C++ struct above, used in C.
+struct tflite_micro_objects;
+
+/** Structure that contains all the data needed to describe an inference engine
+ * This structure contains no C++, just standard C pointers and arrays.
+ */
+typedef struct inference_engine {
+  uint32_t *UNSAFE
+      memory_primary; ///< Pointer to space for tensor arena and optional model
+  uint32_t *UNSAFE
+      memory_secondary; ///< Pointer to secondary space. If null,
+                        // use the primary for model and tensor arena
+  uint32_t
+      outputs; ///< Number of output tensors, initialised on loading a model.
+  uint32_t inputs; ///< Number of input tensors, initialised on loading a model.
+  uint32_t *UNSAFE
+      output_buffers[NUM_OUTPUT_TENSORS]; ///< Pointers to output tensors.
+  uint32_t *UNSAFE
+      input_buffers[NUM_INPUT_TENSORS];      ///< Pointers to input tensors.
+  uint32_t output_sizes[NUM_OUTPUT_TENSORS]; ///< Size of each output tensor in
+                                             ///< bytes.
+  uint32_t
+      input_sizes[NUM_INPUT_TENSORS]; ///< Size of each input tensor in bytes.
+  uint32_t output_size; ///< Total size of all outputs - TODO: obsolete?
+  uint32_t input_size;  ///< Total size of all inputs - TODO: obsolete?
+  uint32_t
+      memory_primary_bytes; ///< Number of bytes available in primary memory
+  uint32_t
+      memory_secondary_bytes; ///< Number of bytes available in secondary memory
+  uint32_t
+      output_times_size; ///< Number of bytes available to store profiling data
+  uint32_t operators_size;       ///< ???
+  uint32_t *UNSAFE output_times; ///< pointer to profiling data, one per layer
+  struct tflite_micro_objects *UNSAFE
+      xtflm;             ///< Pointer to C++ XTFLM object - opaque to C
+                         // status for the engine to maintain
+  uint32_t haveModel;    ///< if 1: we have a model
+  uint32_t chainToNext;  ///< if 1: we are chained (could be implicit in c_push
+                         ///< being non-null?)
+  uint32_t acquireMode;  ///< if non zero we're acquiring data
+  uint32_t outputGpioEn; // TODO: should this be here? Possibly not
+  int8_t outputGpioThresh[AISRV_GPIO_LENGTH];
+  uint8_t outputGpioMode;
+  uint32_t debug_log_buffer[MAX_DEBUG_LOG_LENGTH /
+                            sizeof(uint32_t)]; ///< buffer for error messages
+  uint32_t arena_needed_bytes;                 ///< Total arena needed in bytes.
+  uint32_t num_threads;
+  struct xc_context_config_t xc_config;
+} inference_engine_t;
+
+#ifdef __cplusplus
+#ifndef XTFLM_DISABLED
+/** Function that initializes the inference_engine object, given a
+ * tflite_micro_objects object. This function has to be called from a C++ source
+ * files, and it performs the initialisation of the inference engine. This
+ * involves setting up basic pointers inside the IE object, nothing else.
+ *
+ * The function returns the operator-resolver, which must be be used to add all
+ * necessary operators to the inference engine. A typical calling sequence is as
+ * follows::
+ *
+ *    uint32_t data_ext[100000/sizeof(int)];
+ *    [...]
+ *        static struct tflite_micro_objects s0;
+ *        auto *resolver = inference_engine_initialize(ie,
+ *                                                     data_ext, 100000,
+ *                                                     nullptr,  0,
+ *                                                     &s0);
+ *        resolver->AddAdd();
+ *        resolver->AddConv2D();
+ *        resolver->AddCustom(tflite_micro::ops::micro::xcore::Conv2D_V2_OpCode,
+ *                   tflite_micro::ops::micro::xcore::Register_Conv2D_V2());
+ *    [...]
+ *
+ * Note that when tensorflow lite for micro is disabled this function will not
+ * exist as it depends on all and sundry in XTFLM.
+ *
+ * Note that the lifetime of all spaces passed to this function should be longer
+ * than the lifetime of the inference engine. Typically all spaces are declared
+ * globally.
+ *
+ * \param ie                  Pointer to an uninitialized inference engine
+ * object, allocated by the caller. \param memory_primary      Pointer to the
+ * space to be used for the tensor arena, allocated by the caller. If the fourth
+ * parameter is null, then this space will be used for both model and tensor
+ * arena. \param n_primary           Number of bytes available in primary memory
+ * \param memory_secondary    Pointer to the space to be used for the model
+ *                            allocated by the caller. If this parameter is
+ * null, then the primary memory will be used for both model and tensor arena.
+ * \param n_secondary         Number of bytes available in secondary memory
+ * \param xtflmo               C++ structure for storing the XTFLM data
+ * structures. Must be allocated by the caller.
+ *
+ */
+tflite_micro::MicroMutableOpResolver<XTFLM_OPERATORS> *inference_engine_initialize(
+    inference_engine_t *UNSAFE ie, uint32_t memory_primary[],
+    uint32_t n_memory_primary, uint32_t memory_secondary[],
+    uint32_t n_secondary, struct tflite_micro_objects *UNSAFE xtflmo);
+#endif
+extern "C" {
+#endif
+/** Function that unloads a model frmo the inference engine. This function
+ * must be called before before overwriting the model. For example, you
+ * have ran a model successfully, before you store a new model over the top
+ * of the model (from flash or anywhere else), you must first unload the
+ * model, then you can overwrite the model, and finally you can call the
+ * inference_engine_load_model to set up the new model.
+ *
+ * It is safe to call unload model if there is no model loaded.
+ *
+ * \param ie           pointer to inference engine.
+ */
+void inference_engine_unload_model(inference_engine_t *UNSAFE ie);
+
+/** Function that loads a model into the inference engine. The model must be
+ * stored in either of the two spaces passed to the inference_engine_initialize
+ * function above: either the dedicated data_model space or the space shared
+ * with the tensor_arena. This funciton assumes the model is in place already
+ * and will simply parse it, not copy it.
+ *
+ * \param ie           pointer to inference engine.
+ * \param model_bytes  Number of bytes in the model
+ * \param model_data   Pointer to the model (one of data_model or
+ * data_tensor_arena passed above) \param c_flash_or_tile      Optional channel to flash
+ * or tile server
+ *
+ * \returns            non zero indicates an error
+ */
+#ifdef __XC__
+    int inference_engine_load_model(inference_engine_t * UNSAFE ie, uint32_t model_bytes, uint32_t * UNSAFE model_data, chanend ?c_flash_or_tile);
+#else
+int inference_engine_load_model(inference_engine_t *UNSAFE ie,
+                                uint32_t model_bytes,
+                                uint32_t *UNSAFE model_data, void *weights_data_ptr);
+#endif
+
+    /** Function that invokes the inference engine. This function will create an
+     * extra four threads enabling a model to run in 5 threads.
+     *
+     * \param ie           pointer to inference engine.
+     */
+    int interp_invoke_par_5(inference_engine_t *ie);
+    int interp_invoke_par_4(inference_engine_t *ie);
+    int interp_invoke_par_3(inference_engine_t *ie);
+    int interp_invoke_par_2(inference_engine_t *ie);
+    int interp_invoke(inference_engine_t *ie);
+    TfLiteStatus interp_invoke_internal(inference_engine_t *ie);
+
+    /** Function that resets variable tensors.
+     * This should be called after invoking a model with stateful ops such as LSTM.
+     */
+    int interp_reset(inference_engine_t *ie);
+
+    /** Function that prints a summary of the time each operator took. This
+     * function uses printf - you may want to avoid calling it.
+     *
+     * \param ie           pointer to inference engine.
+     */
+    void print_profiler_summary(inference_engine_t *UNSAFE ie);
+#ifdef __cplusplus
+    };
+#endif
+
+#endif // INFERENCE_ENGINE_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/api/memory_parallel_transport.h

@@ -0,0 +1,52 @@
+#include <stdint.h>
+#include "thread_call.h"
+
+#ifdef __XC__
+#include <xs1.h>
+#else
+#include <xcore/chanend.h>
+#endif
+
+#ifdef __XC__
+
+extern void memory_parallel_receive(chanend c, uint32_t data[], uint32_t bytes);
+extern void memory_parallel_receive_thread_call(chanend c, uint32_t data[], uint32_t bytes, thread_info_t &ptr);
+extern void memory_parallel_send(chanend c, uint32_t data[], uint32_t bytes);
+
+#else
+
+/** Function that receives a block of data.
+ * The number of bytes must be a multiple of 4.
+ * This function creates three threads and three channel ends in order to
+ * make full use of the bandwidth of the switch.
+ * 
+ * \param c        channel end to the sender
+ * \param data     pointer where data must be stored
+ * \param bytes    number of bytes that will be received.
+ */
+extern void memory_parallel_receive(chanend_t c, uint32_t *data, uint32_t bytes);
+
+/** Function that receives a block of data.
+ * The number of bytes must be a multiple of 4.
+ * This function assumes that at least three threads have been created by the
+ * thread_call library and will use those together with three fresh channel
+ * ends in order to make full use of the bandwidth of the switch.
+ * 
+ * \param c        channel end to the sender
+ * \param data     pointer where data must be stored
+ * \param bytes    number of bytes that will be received.
+ */
+extern void memory_parallel_receive_thread_call(chanend_t c, uint32_t *data, uint32_t bytes, thread_info_t *ptr);
+
+/** Function that sends a block of data.
+ * The number of bytes must be a multiple of 4.
+ * This function creates three threads and three channel ends in order to
+ * make full use of the bandwidth of the switch.
+ * 
+ * \param c        channel end to the receiver
+ * \param data     pointer where data must be loaded frmo
+ * \param bytes    number of bytes that will be sent.
+ */
+extern void memory_parallel_send(chanend_t c, uint32_t *data, uint32_t bytes);
+
+#endif

xmos_ai_tools/runtime/include/lib_tflite_micro/api/version.h

@@ -0,0 +1,13 @@
+// Copyright (c) 2020, XMOS Ltd, All rights reserved
+#ifndef XCORE_VERSION_H_
+#define XCORE_VERSION_H_
+
+namespace lib_tflite_micro {
+
+static const unsigned major_version = 0;
+static const unsigned minor_version = 6;
+static const unsigned patch_version = 0;
+
+} // namespace lib_tflite_micro
+
+#endif // XCORE_VERSION_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_config.h

@@ -0,0 +1,17 @@
+#ifndef XCORE_CONFIG_H_
+#define XCORE_CONFIG_H_
+
+#include "../src/thread_call.h"
+
+struct xc_context_config_t {
+  // This is the thread count specified in the compiler.
+  // It's used by lookup op, beta float ops etc to split up work
+  // in the Prepare phase.
+  // Conv ops have their own thread count as the thread work is
+  // calculated in the compiler.
+  int model_thread_count;
+  thread_info_t thread_info;
+  void *UNSAFE weights_data_ptr; // DDR ptr or channel to flash/tile server.
+};
+
+#endif // XCORE_CONFIG_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_device_memory.h

@@ -0,0 +1,62 @@
+// Copyright (c) 2020, XMOS Ltd, All rights reserved
+#ifndef XCORE_DEVICE_MEMORY_H_
+#define XCORE_DEVICE_MEMORY_H_
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifdef XCORE
+#ifdef _TIME_H_
+#define _clock_defined
+#endif
+#include <xcore/thread.h>
+
+#define STRINGIFY(NAME) #NAME
+#define GET_STACKWORDS(DEST, NAME)                                             \
+  asm("ldc %[__dest], " STRINGIFY(NAME) ".nstackwords" : [__dest] "=r"(DEST))
+#define GET_STACKSIZE(DEST, NAME)                                              \
+  {                                                                            \
+    size_t _stack_words;                                                       \
+    asm("ldc %[__dest], " STRINGIFY(NAME) ".nstackwords"                       \
+        : [__dest] "=r"(_stack_words));                                        \
+    DEST = (_stack_words + 2) * 4;                                             \
+  }
+#define IS_RAM(a) (((uintptr_t)a >= 0x80000) && ((uintptr_t)a <= 0x100000))
+#define IS_NOT_RAM(a) ((uintptr_t)a > 0x100000)
+#define IS_EXTMEM(a)                                                           \
+  (((uintptr_t)a >= 0x10000000) && (((uintptr_t)a <= 0x20000000)))
+#define IS_SWMEM(a)                                                            \
+  (((uintptr_t)a >= 0x40000000) && (((uintptr_t)a <= 0x80000000)))
+
+#ifdef USE_SWMEM
+#ifndef USE_QSPI_SWMEM_DEV
+void swmem_setup();
+#else
+#include <xcore/chanend.h>
+void swmem_setup(chanend_t ctrl_swmem_c);
+#endif // USE_QSPI_SWMEM_DEV
+#endif // USE_SWMEM
+
+void swmem_handler(void *ignored);
+void swmem_teardown();
+
+#else // not XCORE
+
+#define GET_STACKSIZE(DEST, NAME) DEST = 0
+#define GET_STACKWORDS(DEST, NAME) DEST = 0
+#define IS_RAM(a) (1)
+#define IS_NOT_RAM(a) (0)
+
+#endif // XCORE
+
+void memload(void *dest, void *src, size_t size);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif // XCORE_DEVICE_MEMORY_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/api/xcore_shared_config.h

@@ -0,0 +1,31 @@
+// Copyright (c) 2020, XMOS Ltd, All rights reserved
+#ifndef XCORE_SHARED_CONFIG_H_
+#define XCORE_SHARED_CONFIG_H_
+
+namespace shared_config {
+
+// This string is used as a key to store the shared config
+// between xformer and lib_tflite_micro in the flatbuffer
+constexpr char xcoreMetadataName[] = "xcoreSharedConfig";
+
+struct xcore_metadata {
+  // Dummy variable for aligning the metadata struct to 16 bytes
+  // We cannot use alignas(16) yet in xcore
+  int32_t padding[2];
+  // Versions of libraries used to build the model
+  int32_t lib_nn_major_version;
+  int32_t lib_nn_minor_version;
+  int32_t lib_nn_patch_version;
+  int32_t lib_tflite_micro_major_version;
+  int32_t lib_tflite_micro_minor_version;
+  int32_t lib_tflite_micro_patch_version;
+  int32_t xformer_major_version;
+  int32_t xformer_minor_version;
+  int32_t xformer_patch_version;
+  // Number of threads required from the runtime to execute the model
+  int32_t required_thread_count;
+};
+
+} // namespace shared_config
+
+#endif // XCORE_SHARED_CONFIG_H_

xmos_ai_tools/runtime/include/lib_tflite_micro/src/tflite-xcore-kernels/conv2d_float.h

@@ -0,0 +1,155 @@
+#ifndef _conv2d_float_h_
+#define _conv2d_float_h_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/** Function that calculates a fully connected.
+ *
+ * @param  outputs              pointer to the output data, the output data will
+ * be stored as an array [out_features]
+ * @param  inputs               pointer to the input data, the input data must
+ * be stored as an array [input_features]
+ * @param  kernels              pointer to the kernels, the kernels
+ *                              must be stored as an array
+ *                              [out_features][input_features]
+ * @param  out_features         dimension 1 of the output array
+ * @param  input_features       dimension 1 of the input array
+ * @returns                     number of MACCs
+ */
+extern int xc_fc_float_ref(float *outputs, float *inputs, float *kernels,
+                           int out_features, int input_features);
+
+/** Optimized function that calculates a fully connected.
+ *
+ * @param  outputs              pointer to the output data, the output data will
+ * be stored as an array [out_features]
+ * @param  inputs               pointer to the input data, the input data must
+ * be stored as an array [input_features]
+ * @param  kernels              pointer to the kernels, the kernels
+ *                              must be stored as an array
+ *                              [out_features][input_features]
+ * @param  out_features         dimension 1 of the output array
+ * @param  input_features       dimension 1 of the input array
+ * @param  out_f_start          output features to start at
+ * @param  out_f_end            output features to end at plus one
+ * @returns                     number of MACCs
+ */
+extern int xc_fc_float_opt(float *outputs, float *inputs, float *kernels,
+                           int out_features, int input_features, int out_f_start,
+                           int out_f_end);
+
+/** Function that calculates a convolution with a 5x2 filter with stride 2
+ * over dimension 2 of a tensor over a tensor.
+ *
+ * @param  outputs     pointer to the output data, the output data will be
+ *                     stored as an array [out_w][out_h][out_depth]
+ * @param  inputs      pointer to the input data, the input data must be
+ *                     stored as an array [input_w][input_h][input_depth]
+ * @param  kernels     pointer to the kernels, the kernels
+ *                     must be stored as an array
+ *                     [out_depth][5][2][depth]
+ * @param  bias        pointer to the biases, the bias must be stored as an
+ * array [out_depth]
+ * @param  out_w       dimension 2 of the output array
+ * @param  out_h       dimension 1 of the output array
+ * @param  out_depth   dimension 3 of the output array
+ * @param  input_w     dimension 2 of the input array
+ * @param  input_h     dimension 1 of the input array
+ * @param  input_depth dimension 3 of the input array
+ * @returns            number of MACCs
+ */
+extern int xc_conv2d_float_kw5xh2_stride_w3_ref(float *outputs, float *inputs,
+                                                float *kernels, float *biases,
+                                                int out_w, int out_h,
+                                                int out_depth, int input_w,
+                                                int input_h, int input_depth);
+
+/** Optimised function that calculates a convolution with a 5x2 filter with
+ * stride 2 over dimension 2 of a tensor. For parallel usage, supply
+ * multiple invocations with different values of out_depth_start and
+ * out_depth_end so that the whole output depth is covered between all of
+ * them.
+ *
+ * @param  outputs     pointer to the output data, the output data will be
+ *                     stored as an array [out_w][out_h][out_depth]
+ * @param  inputs      pointer to the input data, the input data must be
+ *                     stored as an array [input_w][input_h][input_depth]
+ * @param  kernels     pointer to the kernels, the kernels
+ *                     must be stored as an array
+ *                     [out_depth][5][2][depth]
+ * @param  bias        pointer to the biases, the bias must be stored as an
+ * array [out_depth]
+ * @param  out_w       dimension 2 of the output array
+ * @param  out_h       dimension 1 of the output array
+ * @param  out_depth   dimension 3 of the output array
+ * @param  input_w     dimension 2 of the input array
+ * @param  input_h     dimension 1 of the input array
+ * @param  input_depth dimension 3 of the input array
+ * @param  out_depth_start output depth to start at
+ * @param  out_depth_end output depth to end at plus one
+ */
+extern void xc_conv2d_float_kw5xh2_stride_w3_opt(
+    float *outputs, float *inputs, float *kernels, float *biases, int out_w,
+    int out_h, int out_depth, int input_w, int input_h, int input_depth,
+    int out_depth_start, int out_depth_end);
+
+/** Function that calculates a transposed convolution with a 5x2 filter with
+ * stride 3 in dimension 2 over a tensor.
+ *
+ * @param  outputs     pointer to the output data, the output data will be
+ *                     stored as an array [out_w][out_h][out_depth]
+ * @param  inputs      pointer to the input data, the input data must be
+ *                     stored as an array [input_w][input_h][input_depth]
+ * @param  kernels     pointer to the kernels, the kernels
+ *                     must be stored as an array
+ *                     [out_depth][5][2][depth]
+ * @param  bias        pointer to the biases, the bias must be stored as an
+ * array [out_depth]
+ * @param  out_w       dimension 2 of the output array
+ * @param  out_h       dimension 1 of the output array
+ * @param  out_depth   dimension 3 of the output array
+ * @param  input_w     dimension 2 of the input array
+ * @param  input_h     dimension 1 of the input array
+ * @param  input_depth dimension 3 of the input array
+ * @returns            number of MACCs
+ */
+extern int xc_transpose_conv2d_float_kw5xh2_stride_h3_ref(
+    float *outputs, float *inputs, float *kernels, float *biases, int out_w,
+    int out_h, int out_depth, int input_w, int input_h, int input_depth);
+
+/** Optimised function that calculates a transposed convolution with a 5x2
+ * filter with stride 3 over dimension 2 of a tensor. For parallel usage,
+ * supply multiple invocations with different values of out_depth_start and
+ * out_depth_end so that the whole output depth is covered between all of
+ * them.
+ *
+ * @param  outputs     pointer to the output data, the output data will be
+ *                     stored as an array [out_w][out_h][out_depth]
+ * @param  inputs      pointer to the input data, the input data must be
+ *                     stored as an array [input_w][input_h][input_depth]
+ * @param  kernels     pointer to the kernels, the kernels
+ *                     must be stored as an array
+ *                     [out_depth][5][2][depth]
+ * @param  bias        pointer to the biases, the bias must be stored as an
+ * array [out_depth]
+ * @param  out_w       dimension 2 of the output array
+ * @param  out_h       dimension 1 of the output array
+ * @param  out_depth   dimension 3 of the output array
+ * @param  input_w     dimension 2 of the input array
+ * @param  input_h     dimension 1 of the input array
+ * @param  input_depth dimension 3 of the input array
+ * @param  out_depth_start output depth to start at
+ * @param  out_depth_end output depth to end at plus one
+ */
+extern void xc_transpose_conv2d_float_kw5xh2_stride_h3_opt(
+    float *outputs, float *inputs, float *kernels, float *biases, int out_w,
+    int out_h, int out_depth, int input_w, int input_h, int input_depth,
+    int out_depth_start, int out_depth_end);
+
+#ifdef __cplusplus
+};
+#endif
+
+#endif