gsttensor_transform.c

Go to the documentation of this file.

 
#ifdef HAVE_CONFIG_H
#include <config.h>
#endif
 
#include <string.h>
#include <math.h>
#include <nnstreamer_log.h>
#include <nnstreamer_util.h>
#include "gsttensor_transform.h"
 
#ifdef HAVE_ORC
#include "nnstreamer-orc.h"
#endif
 
#ifndef DBG
#define DBG (!filter->silent)
#endif
 
GST_DEBUG_CATEGORY_STATIC (gst_tensor_transform_debug);
#define GST_CAT_DEFAULT gst_tensor_transform_debug
#define CAPS_STRING GST_TENSOR_CAP_DEFAULT ";" GST_TENSORS_CAP_MAKE ("{ static, flexible }")
#define REGEX_DIMCHG_OPTION "^([0-9]|1[0-5]):([0-9]|1[0-5])$"
#define REGEX_TYPECAST_OPTION "(^[u]?int(8|16|32|64)$|^float(16|32|64)$)"
#define REGEX_TRANSPOSE_OPTION "^(?:([0-2]):(?!.*\\1)){3}3$"
#define REGEX_STAND_OPTION "^(default|dc-average)(:([u]?int(8|16|32|64)|float(16|32|64)))?(,per-channel:(true|false))?$"
#define REGEX_CLAMP_OPTION "^((([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?))):"\
    "((([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?)))$"
#define REGEX_PADDING_OPTION "^((left|right|top|bottom|front|back):(\\d)(,)?)+(layout:(NCHW|NHWC))?$"
#define REGEX_ARITH_OPTION "^(typecast:([u]?int(8|16|32|64)|float(16|32|64)),)?"\
    "(per-channel:(false|true@[0-9]+),)?"\
    "(((add|mul|div)(:([-+]?[0-9]*\\.?[0-9]+([eE][-+]?[0-9]+)?))+(@[0-9]+)?)(,|))+$"
 
#define REGEX_ARITH_OPTION_TYPECAST "(typecast:([u]?int(8|16|32|64)|float(16|32|64)))"
 
#define NNS_TENSOR_TRANSPOSE_RANK_LIMIT (4)
 
#define NNS_TENSOR_PADDING_RANK_LIMIT (3)
 
enum
{
  PROP_0,
  PROP_SILENT,
  PROP_MODE,
  PROP_OPTION,
  PROP_ACCELERATION,
  PROP_APPLY,
  PROP_TRANSPOSE_RANK_LIMIT
};
 
#ifdef HAVE_ORC
#define DEFAULT_ACCELERATION TRUE
#else
#define DEFAULT_ACCELERATION FALSE
#endif
 
static const gchar *gst_tensor_transform_stand_string[] = {
  [STAND_DEFAULT] = "default",
  [STAND_DC_AVERAGE] = "dc-average",
  [STAND_END] = NULL
};
 
static const gchar *gst_tensor_transform_operator_string[] = {
  [GTT_OP_TYPECAST] = "typecast",
  [GTT_OP_ADD] = "add",
  [GTT_OP_MUL] = "mul",
  [GTT_OP_DIV] = "div",
  [GTT_OP_UNKNOWN] = NULL
};
 
static GstStaticPadTemplate sink_factory = GST_STATIC_PAD_TEMPLATE ("sink",
    GST_PAD_SINK,
    GST_PAD_ALWAYS,
    GST_STATIC_CAPS (CAPS_STRING));
 
static GstStaticPadTemplate src_factory = GST_STATIC_PAD_TEMPLATE ("src",
    GST_PAD_SRC,
    GST_PAD_ALWAYS,
    GST_STATIC_CAPS (CAPS_STRING));
 
#define gst_tensor_transform_parent_class parent_class
G_DEFINE_TYPE (GstTensorTransform, gst_tensor_transform,
    GST_TYPE_BASE_TRANSFORM);
 
/* GObject vmethod implementations */
static void gst_tensor_transform_set_property (GObject * object, guint prop_id,
    const GValue * value, GParamSpec * pspec);
static void gst_tensor_transform_get_property (GObject * object, guint prop_id,
    GValue * value, GParamSpec * pspec);
static void gst_tensor_transform_finalize (GObject * object);
 
/* GstBaseTransformer vmethod implementations */
static GstFlowReturn gst_tensor_transform_transform (GstBaseTransform * trans,
    GstBuffer * inbuf, GstBuffer * outbuf);
static GstCaps *gst_tensor_transform_transform_caps (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, GstCaps * filter);
static GstCaps *gst_tensor_transform_fixate_caps (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, GstCaps * othercaps);
static gboolean gst_tensor_transform_set_caps (GstBaseTransform * trans,
    GstCaps * incaps, GstCaps * outcaps);
static gboolean gst_tensor_transform_transform_size (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, gsize size,
    GstCaps * othercaps, gsize * othersize);
 
static gboolean gst_tensor_transform_convert_dimension (GstTensorTransform *
    filter, GstPadDirection direction, guint idx, const GstTensorInfo * in_info,
    GstTensorInfo * out_info);
 
#define GST_TYPE_TENSOR_TRANSFORM_MODE (gst_tensor_transform_mode_get_type ())
 
static GType
gst_tensor_transform_mode_get_type (void)
{
  static GType mode_type = 0;
 
  if (mode_type == 0) {
    static GEnumValue mode_types[] = {
      {GTT_DIMCHG, "Mode for changing tensor dimensions, "
            "option=FROM_DIM:TO_DIM (with a regex, " REGEX_DIMCHG_OPTION
            ", where NNS_TENSOR_RANK_LIMIT is 16)",
          "dimchg"},
      {GTT_TYPECAST, "Mode for casting type of tensor, "
            "option=" REGEX_TYPECAST_OPTION, "typecast"},
      {GTT_ARITHMETIC, "Mode for arithmetic operations with tensor, "
            "option=[typecast:TYPE,][per-channel:(false|true@DIM),]add|mul|div:NUMBER[@CH_IDX], ...",
          "arithmetic"},
      {GTT_TRANSPOSE, "Mode for transposing shape of tensor, "
            "option=D1\':D2\':D3\':D4 (fixed to 3)",
          "transpose"},
      {GTT_STAND, "Mode for statistical standardization of tensor, "
            "option=(default|dc-average)[:TYPE][,per-channel:(false|true)]",
          "stand"},
      {GTT_CLAMP, "Mode for clamping all elements of tensor into the range, "
            "option=CLAMP_MIN:CLAMP_MAX",
          "clamp"},
      {GTT_PADDING, "Mode for padding of tensor, "
            "option=left|right|top|bottom|front|back:NUMBER[,layout:(NCHW|NHWC)]",
          "padding"},
      {GTT_UNKNOWN, "Unknown or not-implemented-yet mode",
          "unknown"},
      {0, NULL, NULL},
    };
 
    mode_type = g_enum_register_static ("gtt_mode_type", mode_types);
  }
 
  return mode_type;
}
 
static void
gst_tensor_transform_class_init (GstTensorTransformClass * klass)
{
  GObjectClass *gobject_class;
  GstElementClass *gstelement_class;
  GstBaseTransformClass *trans_class;
 
  GST_DEBUG_CATEGORY_INIT (gst_tensor_transform_debug, "tensor_transform", 0,
      "Element to transforms tensor dimension or type");
 
  trans_class = (GstBaseTransformClass *) klass;
  gstelement_class = (GstElementClass *) trans_class;
  gobject_class = (GObjectClass *) gstelement_class;
 
  gobject_class->set_property = gst_tensor_transform_set_property;
  gobject_class->get_property = gst_tensor_transform_get_property;
  gobject_class->finalize = gst_tensor_transform_finalize;
 
  g_object_class_install_property (gobject_class, PROP_SILENT,
      g_param_spec_boolean ("silent", "Silent", "Produce verbose output ?",
          FALSE, G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_MODE,
      g_param_spec_enum ("mode", "Mode", "Mode used for transforming tensor",
          GST_TYPE_TENSOR_TRANSFORM_MODE, GTT_UNKNOWN,
          G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
  g_object_class_install_property (gobject_class, PROP_OPTION,
      g_param_spec_string ("option", "Option",
          "Option for the tensor transform mode ?", "", G_PARAM_READWRITE));
  g_object_class_install_property (gobject_class, PROP_ACCELERATION,
      g_param_spec_boolean ("acceleration", "Acceleration", "Orc acceleration",
          DEFAULT_ACCELERATION, G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
  g_object_class_install_property (gobject_class, PROP_APPLY,
      g_param_spec_string ("apply", "Apply", "Select tensors to apply, "
          "separated with ',' in case of multiple tensors. Default to apply all tensors.",
          "", G_PARAM_READWRITE | G_PARAM_STATIC_STRINGS));
  g_object_class_install_property (gobject_class, PROP_TRANSPOSE_RANK_LIMIT,
      g_param_spec_uint ("transpose-rank-limit", "Transpose rank limit",
          "The rank limit of transpose, which varies per version of nnstreamer and may be lower than the global rank limit if it is over 4.",
          0, NNS_TENSOR_RANK_LIMIT, NNS_TENSOR_TRANSPOSE_RANK_LIMIT,
          G_PARAM_READABLE | G_PARAM_STATIC_STRINGS));
 
  gst_element_class_set_details_simple (gstelement_class,
      "TensorTransform",
      "Filter/Tensor",
      "Transforms other/tensor dimensions for different models or frameworks",
      "MyungJoo Ham <myungjoo.ham@samsung.com>");
 
  gst_element_class_add_pad_template (gstelement_class,
      gst_static_pad_template_get (&src_factory));
  gst_element_class_add_pad_template (gstelement_class,
      gst_static_pad_template_get (&sink_factory));
  /* Refer: https://gstreamer.freedesktop.org/documentation/design/element-transform.html */
  trans_class->passthrough_on_same_caps = FALSE;
 
  /* Processing units */
  trans_class->transform = GST_DEBUG_FUNCPTR (gst_tensor_transform_transform);
 
  /* Negotiation units */
  trans_class->transform_caps =
      GST_DEBUG_FUNCPTR (gst_tensor_transform_transform_caps);
  trans_class->fixate_caps =
      GST_DEBUG_FUNCPTR (gst_tensor_transform_fixate_caps);
  trans_class->set_caps = GST_DEBUG_FUNCPTR (gst_tensor_transform_set_caps);
 
  /* Allocation units */
  trans_class->transform_size =
      GST_DEBUG_FUNCPTR (gst_tensor_transform_transform_size);
}
 
static void
gst_tensor_transform_init (GstTensorTransform * filter)
{
  filter->silent = TRUE;
  filter->mode = GTT_UNKNOWN;
  filter->option = NULL;
  filter->loaded = FALSE;
  filter->operators = NULL;
  filter->acceleration = DEFAULT_ACCELERATION;
  filter->apply = NULL;
 
  gst_tensors_config_init (&filter->in_config);
  gst_tensors_config_init (&filter->out_config);
}
 
static tensor_transform_operator
gst_tensor_transform_get_operator (const gchar * str)
{
  int index;
 
  index = find_key_strv (gst_tensor_transform_operator_string, str);
 
  return (index < 0) ? GTT_OP_UNKNOWN : index;
}
 
static tensor_transform_stand_mode
gst_tensor_transform_get_stand_mode (const gchar * str)
{
  int index;
 
  index = find_key_strv (gst_tensor_transform_stand_string, str);
 
  return (index < 0) ? STAND_END : index;
}
 
#ifndef FLOAT16_SUPPORT
 
static void
float16_not_supported (void)
{
  ml_loge
      ("Tensor_transform does not support float16 operators. Apply -Denable-float16=true for meson build option if your architecture support float16. Note that tensor-transform's float16 is adhoc and does NOT perform good (slow!).\n");
  g_assert (0);
}
#endif
 
#ifdef FLOAT16_SUPPORT
 
static void
refrain_from_heavy_op_on_float16 (gulong n)
{
  static int warned = 0;
  /* 1 million */
  if (n > 1000000) {
    if (warned)
      return;
    ml_logw
        ("Tensor_transform implementation for float16 does not support SIMD. Heavy tensor-transform operations of float16 is not recommended. Try to apply heavy ops with other types (e.g., float32) and convert it to float16 at the time when it's really needed.\n");
    warned = 1;
  }
}
 
#define _conv_to_f16(intype, o, i, n) \
  do { \
    float16 *op = (gpointer) (o); \
    intype *ip = (gpointer) (i); \
    gulong idx; \
    refrain_from_heavy_op_on_float16 (n); \
    for (idx = 0; idx < n; idx++) \
      *(op + idx) = (float16) *(ip + idx); \
  } while (0)
 
#define _conv_from_f16_action(n, op, ip, otypename) \
  do { \
    gulong idx; \
    for (idx = 0; idx < n; idx++) \
      *(op + idx) = (otypename) *(ip + idx); \
  } while (0)
 
#define _conv_from_f16(otype, o, i, n) \
  do { \
    float16 *ip = (gpointer) (i); \
    refrain_from_heavy_op_on_float16 (n); \
    switch (otype) { \
      case _NNS_INT32: { \
        int32_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, int32_t); \
        break; } \
      case _NNS_UINT32: {  \
        uint32_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, uint32_t); \
        break; } \
      case _NNS_INT16: {  \
        int16_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, int16_t); \
        break; } \
      case _NNS_UINT16: {  \
        uint16_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, uint16_t); \
        break; } \
      case _NNS_INT8: {  \
        int8_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, int8_t); \
        break; } \
      case _NNS_UINT8: {  \
        uint8_t *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, uint8_t); \
        break; } \
      case _NNS_FLOAT64: {  \
        double *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, double); \
        break; } \
      case _NNS_FLOAT32: {  \
        float *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, float); \
        break; } \
      case _NNS_FLOAT16: {  \
        float16 *op = (gpointer) (o); \
        _conv_from_f16_action (n, op, ip, float16); \
        break; } \
      default: GST_ERROR_OBJECT (filter, "Unsupported type %d", (otype)); g_assert (0); \
    } \
  } while (0)
 
#define _op_float16(i, n, v, op) \
  do { \
    gulong idx; \
    float16 *data_in = (float16 *) (i); \
    refrain_from_heavy_op_on_float16 (n); \
    switch (op) { \
      case GTT_OP_ADD: \
        for (idx = 0; idx < n; idx++) \
          data_in[idx] = data_in[idx] + (v); \
        break; \
      case GTT_OP_MUL: \
        for (idx = 0; idx < n; idx++) \
          data_in[idx] = data_in[idx] * (v); \
        break; \
      case GTT_OP_DIV: \
        for (idx = 0; idx < n; idx++) \
          data_in[idx] = data_in[idx] / (v); \
        break; \
      default: GST_ERROR_OBJECT (filter, "Unknown operator for float16: %d", op); break; \
    } \
  } while (0)
 
#else /* ! FLOAT16_SUPPORT */
#define _conv_to_f16(intype, o, i, n) do { float16_not_supported (); } while (0)
#define _conv_from_f16(otype, o, i, n) do { float16_not_supported (); } while (0)
#define _op_float16(i, n, v, op) do { float16_not_supported (); } while (0)
#endif /* FLOAT16_SUPPORT */
 
#ifdef HAVE_ORC
/* define macros for orc */
#define orc_func_conv(intype,outtype) nns_orc_conv_ 
#define orc_func_add(intype) nns_orc_add_c_ 
#define orc_func_mul(intype) nns_orc_mul_c_ 
#define orc_func_div(intype) nns_orc_div_c_ 
 
#define orc_typecast_to(i,o,n,intype,otype,intypename) do { \
    switch (otype) { \
      case _NNS_INT32: orc_func_conv (intype, s32) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_UINT32: orc_func_conv (intype, u32) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_INT16: orc_func_conv (intype, s16) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_UINT16: orc_func_conv (intype, u16) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_INT8: orc_func_conv (intype, s8) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_UINT8: orc_func_conv (intype, u8) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_FLOAT64: orc_func_conv (intype, f64) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_FLOAT32: orc_func_conv (intype, f32) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_INT64: orc_func_conv (intype, s64) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_UINT64: orc_func_conv (intype, u64) ((gpointer) o, (gpointer) i, n); break; \
      case _NNS_FLOAT16: _conv_to_f16 (intypename, o, i, n); break; \
      default: GST_ERROR_OBJECT (filter, "Unsupported output type %d", otype); g_assert (0); break; \
    } \
  } while (0)
 
#define orc_typecast(i,o,n,itype,otype) do { \
    switch (itype) { \
      case _NNS_INT32: orc_typecast_to (i, o, n, s32, otype, int32_t); break; \
      case _NNS_UINT32: orc_typecast_to (i, o, n, u32, otype, uint32_t); break; \
      case _NNS_INT16: orc_typecast_to (i, o, n, s16, otype, int16_t); break; \
      case _NNS_UINT16: orc_typecast_to (i, o, n, u16, otype, uint16_t); break; \
      case _NNS_INT8: orc_typecast_to (i, o, n, s8, otype, int8_t); break; \
      case _NNS_UINT8: orc_typecast_to (i, o, n, u8, otype, uint8_t); break; \
      case _NNS_FLOAT64: orc_typecast_to (i, o, n, f64, otype, double); break; \
      case _NNS_FLOAT32: orc_typecast_to (i, o, n, f32, otype, float); break; \
      case _NNS_INT64: orc_typecast_to (i, o, n, s64, otype, int64_t); break; \
      case _NNS_UINT64: orc_typecast_to (i, o, n, u64, otype, uint64_t); break; \
      case _NNS_FLOAT16: _conv_from_f16 (otype, o, i, n); break; \
      default: GST_ERROR_OBJECT (filter, "Unsupported input type %d", itype); g_assert (0); break; \
    } \
  } while (0)
 
#define orc_typesize(size, type) do { \
    switch (type) { \
      case _NNS_INT32: size = sizeof(int32_t); break; \
      case _NNS_UINT32: size = sizeof(uint32_t); break; \
      case _NNS_INT16: size = sizeof(int16_t); break; \
      case _NNS_UINT16: size = sizeof(uint16_t); break; \
      case _NNS_INT8: size = sizeof(int8_t); break; \
      case _NNS_UINT8: size = sizeof(uint8_t); break; \
      case _NNS_FLOAT64: size = sizeof(double); break; \
      case _NNS_FLOAT32: size = sizeof(float); break; \
      case _NNS_INT64: size = sizeof(int64_t); break; \
      case _NNS_UINT64: size = sizeof(uint64_t); break; \
      default: GST_ERROR_OBJECT (filter, "Unsupported type %d", type); g_assert (0); break; \
    } \
  } while (0)
 
#define orc_operator_func(i,n,v,opfunc,op) do { \
    switch ((v)->type) { \
      case _NNS_INT32: opfunc (s32) ((gpointer) i, (v)->data._int32_t, n); break; \
      case _NNS_UINT32: opfunc (u32) ((gpointer) i, (v)->data._uint32_t, n); break; \
      case _NNS_INT16: opfunc (s16) ((gpointer) i, (v)->data._int16_t, n); break; \
      case _NNS_UINT16: opfunc (u16) ((gpointer) i, (v)->data._uint16_t, n); break; \
      case _NNS_INT8: opfunc (s8) ((gpointer) i, (v)->data._int8_t, n); break; \
      case _NNS_UINT8: opfunc (u8) ((gpointer) i, (v)->data._uint8_t, n); break; \
      case _NNS_FLOAT64: opfunc (f64) ((gpointer) i, (v)->data._double, n); break; \
      case _NNS_FLOAT32: opfunc (f32) ((gpointer) i, (v)->data._float, n); break; \
      case _NNS_INT64: opfunc (s64) ((gpointer) i, (v)->data._int64_t, n); break; \
      case _NNS_UINT64: opfunc (u64) ((gpointer) i, (v)->data._uint64_t, n); break; \
      case _NNS_FLOAT16: _op_float16 (i, n, (v)->data._float16, op); break; \
      default: GST_ERROR_OBJECT (filter, "Unsupported type %d", (v)->type); g_assert (0); break; \
    } \
  } while (0)
 
#define orc_operator_div_loop(i,n,val,typename) do { \
    gsize idx_div; \
    typename *data_in = (typename *) (i); \
    for (idx_div = 0; idx_div < (n); ++idx_div) { \
      data_in[idx_div] = data_in[idx_div] / (val); \
    } \
  } while (0)
 
#define orc_operator(i,n,v,op) do { \
    switch (op) { \
      case GTT_OP_ADD: orc_operator_func (i, n, v, orc_func_add, op); break; \
      case GTT_OP_MUL: orc_operator_func (i, n, v, orc_func_mul, op); break; \
      case GTT_OP_DIV: \
        switch ((v)->type) { \
          case _NNS_INT32: orc_operator_div_loop (i, n, (v)->data._int32_t, int32_t); break; \
          case _NNS_UINT32: orc_operator_div_loop (i, n, (v)->data._uint32_t, uint32_t); break; \
          case _NNS_INT16: orc_operator_div_loop (i, n, (v)->data._int16_t, int16_t); break; \
          case _NNS_UINT16: orc_operator_div_loop (i, n, (v)->data._uint16_t, uint16_t); break; \
          case _NNS_INT8: orc_operator_div_loop (i, n, (v)->data._int8_t, int8_t); break; \
          case _NNS_UINT8: orc_operator_div_loop (i, n, (v)->data._uint8_t, uint8_t); break; \
          case _NNS_FLOAT64: orc_func_div (f64) ((gpointer) i, (v)->data._double, n); break; \
          case _NNS_FLOAT32: orc_func_div (f32) ((gpointer) i, (v)->data._float, n); break; \
          case _NNS_INT64: orc_operator_div_loop (i, n, (v)->data._int64_t, int64_t); break; \
          case _NNS_UINT64: orc_operator_div_loop (i, n, (v)->data._uint64_t, uint64_t); break; \
          case _NNS_FLOAT16: _op_float16 (i, n, (v)->data._float16, op); break; \
          default: GST_ERROR_OBJECT (filter, "Unsupported type %d", (v)->type); g_assert (0); break; \
        } \
        break; \
      default: GST_ERROR_OBJECT (filter, "Unknown operator %d", op); break; \
    } \
  } while (0)
#endif /* HAVE_ORC */
 
#define handle_operator(d,v,oper,vtype) do { \
    switch (oper) { \
      case GTT_OP_ADD: \
        (d)->data._
        break; \
      case GTT_OP_MUL: \
        (d)->data._
        break; \
      case GTT_OP_DIV: \
        if ((v)->data._
          GST_ERROR_OBJECT (filter, "Invalid state, denominator is 0."); \
          return FALSE; \
        } \
        (d)->data._
        break; \
      default: \
        GST_ERROR_OBJECT (filter, "Unknown operator %d", oper); \
        return FALSE; \
    } \
  } while (0)
 
static gboolean
gst_tensor_transform_do_operator (GstTensorTransform * filter,
    tensor_data_s * desc, const tensor_data_s * val,
    tensor_transform_operator op)
{
  g_return_val_if_fail (desc != NULL, FALSE);
  g_return_val_if_fail (val != NULL, FALSE);
  g_return_val_if_fail (desc->type == val->type, FALSE);
 
  switch (desc->type) {
    case _NNS_INT32:
      handle_operator (desc, val, op, int32_t);
      break;
    case _NNS_UINT32:
      handle_operator (desc, val, op, uint32_t);
      break;
    case _NNS_INT16:
      handle_operator (desc, val, op, int16_t);
      break;
    case _NNS_UINT16:
      handle_operator (desc, val, op, uint16_t);
      break;
    case _NNS_INT8:
      handle_operator (desc, val, op, int8_t);
      break;
    case _NNS_UINT8:
      handle_operator (desc, val, op, uint8_t);
      break;
    case _NNS_FLOAT64:
      handle_operator (desc, val, op, double);
      break;
    case _NNS_FLOAT32:
      handle_operator (desc, val, op, float);
      break;
    case _NNS_FLOAT16:
#ifdef FLOAT16_SUPPORT
      handle_operator (desc, val, op, float16);
#else
      float16_not_supported ();
#endif
      break;
    case _NNS_INT64:
      handle_operator (desc, val, op, int64_t);
      break;
    case _NNS_UINT64:
      handle_operator (desc, val, op, uint64_t);
      break;
    default:
      GST_ERROR_OBJECT (filter, "Unknown tensor type %d", desc->type);
      return FALSE;
  }
 
  return TRUE;
}
 
static gboolean
gst_tensor_transform_set_option_data (GstTensorTransform * filter)
{
  gchar *filter_name;
  gboolean ret = FALSE;
 
  if (filter->mode == GTT_UNKNOWN || filter->option == NULL)
    return TRUE;
 
  filter_name = gst_object_get_name ((GstObject *) filter);
 
  switch (filter->mode) {
    case GTT_DIMCHG:
    {
      gchar **strv = NULL;
 
      if (!g_regex_match_simple (REGEX_DIMCHG_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: dimchg: \'%s\' is not valid option string: it should be in the form of IDX_DIM_FROM:IDX_DIM_TO: with a regex, "
            REGEX_DIMCHG_OPTION "\n", filter_name, filter->option);
        break;
      }
 
      strv = g_strsplit (filter->option, ":", 2);
 
      filter->data_dimchg.from = (int) g_ascii_strtoll (strv[0], NULL, 10);
      filter->data_dimchg.to = (int) g_ascii_strtoll (strv[1], NULL, 10);
      ret = filter->loaded = TRUE;
      g_strfreev (strv);
      break;
    }
    case GTT_TYPECAST:
    {
      if (g_regex_match_simple (REGEX_TYPECAST_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        filter->data_typecast.to = gst_tensor_get_type (filter->option);
        ret = filter->loaded = TRUE;
      } else {
        ml_loge
            ("%s: typecast: \'%s\' is not valid data type for tensor: data type of tensor should be one of %s\n",
            filter_name, filter->option, GST_TENSOR_TYPE_ALL);
      }
      break;
    }
    case GTT_ARITHMETIC:
    {
      gchar *str_option;
      gchar **str_operators;
      gchar **str_op;
      tensor_transform_operator_s *op_s;
      guint i, num_operators, num_op;
      GRegex *regex_option_tc;
 
      filter->data_arithmetic.out_type = _NNS_END;
      filter->data_arithmetic.per_channel_arith = FALSE;
 
      if (filter->operators) {
        GST_WARNING_OBJECT (filter,
            "There exists pre-defined operators (total %d), now reset these.",
            g_slist_length (filter->operators));
 
        g_slist_free_full (filter->operators, g_free);
        filter->operators = NULL;
      }
 
      regex_option_tc = g_regex_new (REGEX_ARITH_OPTION_TYPECAST,
          G_REGEX_CASELESS, 0, 0);
 
      if (!regex_option_tc) {
        GST_ERROR_OBJECT (filter,
            "arithmetic: failed to create a GRegex structure for %s\n",
            REGEX_ARITH_OPTION_TYPECAST);
        break;
      }
 
      if (g_regex_match_full (regex_option_tc, filter->option, -1,
              1, 0, NULL, NULL)) {
        str_option = g_regex_replace (regex_option_tc, filter->option, -1, 1,
            "", 0, 0);
        ml_loge
            ("%s: arithmetic: [typecast:TYPE,] should be located at the first to prevent memory re-allocation: typecast(s) in the middle of \'%s\' will be ignored\n",
            filter_name, filter->option);
      } else {
        str_option = g_strdup (filter->option);
      }
      g_regex_unref (regex_option_tc);
 
      if (!g_regex_match_simple (REGEX_ARITH_OPTION, str_option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: arithmetic: \'%s\' is not valid option string: it should be in the form of [typecast:TYPE,][per-channel:(false|true@DIM),]add|mul|div:NUMBER[@CH_IDX]..., ...\n",
            filter_name, str_option);
        g_free (str_option);
        break;
      }
      str_operators = g_strsplit (str_option, ",", -1);
      num_operators = g_strv_length (str_operators);
 
      for (i = 0; i < num_operators; ++i) {
        str_op = g_strsplit (str_operators[i], ":", -1);
        num_op = g_strv_length (str_op);
 
        if (str_op[0]) {
          gchar **values = g_strsplit (str_op[1], "@", -1);
          guint num_values = g_strv_length (values);
 
          /* check whether per-channel */
          if (g_ascii_strcasecmp (str_op[0], "per-channel") == 0) {
            if (num_values > 1 && g_ascii_strcasecmp (values[0], "true") == 0) {
              ml_logi
                  ("Set per-channel for arithmetic and assume that %s-th dim is the channel",
                  values[1]);
              filter->data_arithmetic.per_channel_arith = TRUE;
              filter->data_arithmetic.ch_dim =
                  (guint) g_ascii_strtoull (values[1], NULL, 10);
            }
 
            g_strfreev (values);
            g_strfreev (str_op);
            continue;
          }
 
          op_s = g_new0 (tensor_transform_operator_s, 1);
          g_assert (op_s);
 
          op_s->op = gst_tensor_transform_get_operator (str_op[0]);
          op_s->applying_ch = -1;       /* -1 means applying to all channels */
          switch (op_s->op) {
            case GTT_OP_TYPECAST:
              if (num_op > 1 && str_op[1]) {
                op_s->value.type = gst_tensor_get_type (values[0]);
                filter->data_arithmetic.out_type = op_s->value.type;
              } else {
                GST_WARNING_OBJECT (filter, "Invalid option for typecast %s",
                    str_operators[i]);
                op_s->op = GTT_OP_UNKNOWN;
              }
              break;
            case GTT_OP_ADD:
            case GTT_OP_MUL:
            case GTT_OP_DIV:
              if (num_op > 1 && str_op[1]) {
                /* get operand */
                if (strchr (values[0], '.') || strchr (values[0], 'e') ||
                    strchr (values[0], 'E')) {
                  double val;
 
                  val = g_ascii_strtod (values[0], NULL);
                  gst_tensor_data_set (&op_s->value, _NNS_FLOAT64, &val);
                } else {
                  int64_t val;
 
                  val = g_ascii_strtoll (values[0], NULL, 10);
                  gst_tensor_data_set (&op_s->value, _NNS_INT64, &val);
                }
 
                if (filter->data_arithmetic.per_channel_arith && num_values > 1) {
                  op_s->applying_ch = g_ascii_strtoll (values[1], NULL, 10);
                }
 
              } else {
                GST_WARNING_OBJECT (filter,
                    "Invalid option for arithmetic %s", str_operators[i]);
                op_s->op = GTT_OP_UNKNOWN;
              }
              break;
            default:
              GST_WARNING_OBJECT (filter, "Unknown operator %s", str_op[0]);
              break;
          }
 
          /* append operator */
          if (op_s->op != GTT_OP_UNKNOWN) {
            filter->operators = g_slist_append (filter->operators, op_s);
          } else {
            g_free (op_s);
          }
 
          g_strfreev (values);
        } else {
          GST_WARNING_OBJECT (filter, "Invalid option %s", str_operators[i]);
        }
 
        g_strfreev (str_op);
      }
 
      ret = filter->loaded = (filter->operators != NULL);
      g_strfreev (str_operators);
      g_free (str_option);
      break;
    }
    case GTT_TRANSPOSE:
    {
      int i;
      gchar **strv = NULL;
 
      if (!g_regex_match_simple (REGEX_TRANSPOSE_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: transpose: \'%s\' is not valid option string: it should be in the form of NEW_IDX_DIM0:NEW_IDX_DIM1:NEW_IDX_DIM2:3 (Now transpose mode's rank is fixed to 3. Note that the index of the last dim is always fixed to 3)\n",
            filter_name, filter->option);
        break;
      }
 
      strv = g_strsplit (filter->option, ":", NNS_TENSOR_TRANSPOSE_RANK_LIMIT);
      for (i = 0; i < NNS_TENSOR_TRANSPOSE_RANK_LIMIT; i++) {
        filter->data_transpose.trans_order[i] =
            (uint8_t) g_ascii_strtoull (strv[i], NULL, 10);
      }
 
      ret = filter->loaded = TRUE;
      g_strfreev (strv);
      break;
    }
    case GTT_STAND:
    {
      gchar **options = NULL;
      guint i, num_options;
 
      if (!g_regex_match_simple (REGEX_STAND_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: stand: \'%s\' is not a valid option string: it should be in the form of (default|dc-average)[:TYPE][,per-channel:(false|true)]\n",
            filter_name, filter->option);
        break;
      }
 
      filter->data_stand.out_type = _NNS_END;
      filter->data_stand.per_channel = FALSE;
 
      options = g_strsplit (filter->option, ",", -1);
      num_options = g_strv_length (options);
 
      for (i = 0; i < num_options; i++) {
        gchar **strv = g_strsplit (options[i], ":", -1);
 
        if (g_ascii_strcasecmp (strv[0], "default") == 0 ||
            g_ascii_strcasecmp (strv[0], "dc-average") == 0) {
          filter->data_stand.mode =
              gst_tensor_transform_get_stand_mode (strv[0]);
          if (g_strv_length (strv) > 1)
            filter->data_stand.out_type = gst_tensor_get_type (strv[1]);
        } else if (g_ascii_strcasecmp (strv[0], "per-channel") == 0) {
          if (g_strv_length (strv) > 1 &&
              g_ascii_strcasecmp (strv[1], "true") == 0)
            filter->data_stand.per_channel = TRUE;
        } else {
          filter->data_stand.mode = STAND_END;
          ml_logw ("Unknown option for stand mode: %s", strv[0]);
        }
 
        g_strfreev (strv);
      }
 
      g_strfreev (options);
      ret = filter->loaded = TRUE;
      break;
    }
    case GTT_CLAMP:
    {
      gchar **strv = NULL;
 
      if (!g_regex_match_simple (REGEX_CLAMP_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: clamp: \'%s\' is not valid option string: it should be in the form of [CLAMP_MIN:CLAMP_MAX]\n",
            filter_name, filter->option);
        break;
      }
 
      strv = g_strsplit (filter->option, ":", 2);
 
      filter->data_clamp.min = g_ascii_strtod (strv[0], NULL);
      if (errno == ERANGE) {
        ml_loge ("%s: clamp: CLAMP_MIN value has an invalid range\n",
            filter_name);
        g_strfreev (strv);
        break;
      }
      filter->data_clamp.max = g_ascii_strtod (strv[1], NULL);
      if (errno == ERANGE) {
        ml_loge ("%s: clamp: CLAMP_MAX value has an invalid range\n",
            filter_name);
        g_strfreev (strv);
        break;
      }
 
      g_strfreev (strv);
 
      if (filter->data_clamp.min > filter->data_clamp.max) {
        ml_loge ("%s: clamp: CLAMP_MIN is larger than CLAMP_MAX\n",
            filter_name);
        break;
      }
 
      ret = filter->loaded = TRUE;
      break;
    }
    case GTT_PADDING:
    {
      gchar **options = NULL;
      guint i, num_options;
 
      if (!g_regex_match_simple (REGEX_PADDING_OPTION, filter->option,
              G_REGEX_CASELESS, 0)) {
        ml_loge
            ("%s: padding: \'%s\' is not valid option string: it should be in the form of left|right|top|bottom|front|back:PADDING,[layout:(NCHW|NHWC)]\n",
            filter_name, filter->option);
        break;
      }
 
      for (i = 0; i < NNS_TENSOR_RANK_LIMIT; i++)
        filter->data_padding.pad[i] = 0;
      filter->data_padding.layout = _NNS_LAYOUT_ANY;
 
      options = g_strsplit (filter->option, ",", -1);
      num_options = g_strv_length (options);
 
      for (i = 0; i < num_options; i++) {
        gchar **strv = g_strsplit (options[i], ":", 2);
        if (g_ascii_strcasecmp (strv[0], "left") == 0) {
          filter->data_padding.pad[PADDING_LEFT] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "right") == 0) {
          filter->data_padding.pad[PADDING_RIGHT] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "top") == 0) {
          filter->data_padding.pad[PADDING_TOP] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "bottom") == 0) {
          filter->data_padding.pad[PADDING_BOTTOM] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "front") == 0) {
          filter->data_padding.pad[PADDING_FRONT] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "back") == 0) {
          filter->data_padding.pad[PADDING_BACK] =
              (guint) g_ascii_strtoull (strv[1], NULL, 10);
        } else if (g_ascii_strcasecmp (strv[0], "layout") == 0) {
          if (g_ascii_strcasecmp (strv[1], "NHWC") == 0)
            filter->data_padding.layout = _NNS_LAYOUT_NHWC;
          else
            filter->data_padding.layout = _NNS_LAYOUT_NCHW;
        } else {
          ml_logw ("Unknown option for padding mode: %s", strv[0]);
        }
        g_strfreev (strv);
      }
      g_strfreev (options);
 
      if (filter->data_padding.layout == _NNS_LAYOUT_NHWC) {
        guint prev_left = filter->data_padding.pad[PADDING_LEFT],
            prev_right = filter->data_padding.pad[PADDING_RIGHT];
        filter->data_padding.pad[PADDING_LEFT] =
            filter->data_padding.pad[PADDING_FRONT];
        filter->data_padding.pad[PADDING_RIGHT] =
            filter->data_padding.pad[PADDING_BACK];
        filter->data_padding.pad[PADDING_FRONT] = prev_left;
        filter->data_padding.pad[PADDING_BACK] = prev_right;
      }
 
      ret = filter->loaded = TRUE;
      break;
    }
    default:
      GST_ERROR_OBJECT (filter, "Cannot identify mode\n");
      ret = FALSE;
  }
 
  g_free (filter_name);
  return ret;
}
 
static void
gst_tensor_transform_set_property (GObject * object, guint prop_id,
    const GValue * value, GParamSpec * pspec)
{
  GstTensorTransform *filter = GST_TENSOR_TRANSFORM (object);
 
  switch (prop_id) {
    case PROP_SILENT:
      filter->silent = g_value_get_boolean (value);
      break;
    case PROP_MODE:
      filter->mode = g_value_get_enum (value);
      gst_tensor_transform_set_option_data (filter);
      break;
    case PROP_OPTION:
    {
      gchar *backup_option = filter->option;
      filter->option = g_value_dup_string (value);
      if (gst_tensor_transform_set_option_data (filter)) {
        silent_debug (filter, "Option = %s --> %s\n", backup_option,
            filter->option);
        g_free (backup_option);
      } else {
        /* ERROR! Revert the change! */
        g_free (filter->option);
        filter->option = backup_option;
        gst_tensor_transform_set_option_data (filter);
      }
      break;
    }
    case PROP_ACCELERATION:
#ifdef HAVE_ORC
      filter->acceleration = g_value_get_boolean (value);
      silent_debug (filter, "acceleration = %d\n", filter->acceleration);
#else
      GST_WARNING_OBJECT (filter, "Orc acceleration is not supported");
      filter->acceleration = FALSE;
#endif
      break;
    case PROP_APPLY:
    {
      gint64 val;
      const gchar *param = g_value_get_string (value);
      gchar **strv = g_strsplit_set (param, ",", -1);
      guint i, num = g_strv_length (strv);
      gchar *endptr = NULL;
 
      for (i = 0; i < num; i++) {
        errno = 0;
        val = g_ascii_strtoll (strv[i], &endptr, 10);
        if (errno == ERANGE || errno == EINVAL || (endptr == strv[i])) {
          ml_loge ("Cannot convert string %s to a gint64 value", strv[i]);
        }
        filter->apply = g_list_append (filter->apply, GINT_TO_POINTER (val));
      }
      g_strfreev (strv);
      break;
    }
    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}
 
static void
gst_tensor_transform_get_property (GObject * object, guint prop_id,
    GValue * value, GParamSpec * pspec)
{
  GstTensorTransform *filter = GST_TENSOR_TRANSFORM (object);
 
  switch (prop_id) {
    case PROP_SILENT:
      g_value_set_boolean (value, filter->silent);
      break;
    case PROP_MODE:
      g_value_set_enum (value, filter->mode);
      break;
    case PROP_OPTION:
      g_value_set_string (value, filter->option);
      break;
    case PROP_ACCELERATION:
      g_value_set_boolean (value, filter->acceleration);
      break;
    case PROP_APPLY:
    {
      GList *list;
      gchar *p;
      GPtrArray *arr;
      gchar **strings;
 
      if (filter->apply == NULL) {
        g_value_set_string (value, "");
        return;
      }
 
      arr = g_ptr_array_new ();
      for (list = filter->apply; list != NULL; list = list->next) {
        g_ptr_array_add (arr, g_strdup_printf ("%i",
                GPOINTER_TO_INT (list->data)));
      }
      g_ptr_array_add (arr, NULL);
      strings = (gchar **) g_ptr_array_free (arr, FALSE);
      p = g_strjoinv (",", strings);
 
      g_strfreev (strings);
      g_value_take_string (value, p);
      break;
    }
    case PROP_TRANSPOSE_RANK_LIMIT:
      g_value_set_uint (value, NNS_TENSOR_TRANSPOSE_RANK_LIMIT);
      break;
    default:
      G_OBJECT_WARN_INVALID_PROPERTY_ID (object, prop_id, pspec);
      break;
  }
}
 
static void
gst_tensor_transform_finalize (GObject * object)
{
  GstTensorTransform *filter;
 
  filter = GST_TENSOR_TRANSFORM (object);
 
  if (filter->option) {
    g_free (filter->option);
    filter->option = NULL;
  }
 
  if (filter->operators) {
    g_slist_free_full (filter->operators, g_free);
    filter->operators = NULL;
  }
 
  if (filter->apply) {
    g_list_free (filter->apply);
    filter->apply = NULL;
  }
 
  G_OBJECT_CLASS (parent_class)->finalize (object);
}
 
static GstFlowReturn
gst_tensor_transform_dimchg (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  uint32_t *fromDim = in_info->dimension;
  uint32_t *toDim = out_info->dimension;
  unsigned int from = filter->data_dimchg.from;
  unsigned int to = filter->data_dimchg.to;
  unsigned int i, j, k;
  unsigned int loopLimit = 1;
  gsize loopBlockSize, copyblocksize, copyblocklimit;
 
  if (from == to) {
    nns_memcpy (outptr, inptr, gst_tensor_info_get_size (in_info));
    GST_WARNING_OBJECT (filter,
        "Calling tensor_transform with high memcpy overhead WITHOUT any effects! Check your stream whether you really need tensor_transform.\n");
    return GST_FLOW_OK;
  }
 
  g_assert (from < NNS_TENSOR_RANK_LIMIT);
  g_assert (to < NNS_TENSOR_RANK_LIMIT);
  g_assert (fromDim[from] == toDim[to]);
 
  loopBlockSize = copyblocksize = gst_tensor_get_element_size (in_info->type);
  copyblocklimit = 1;
 
  if (from < to) {
    for (i = NNS_TENSOR_RANK_LIMIT - 1; i > to; i--) {
      if (toDim[i] == 0)
        continue;
      loopLimit *= toDim[i];
    }
 
    for (i = 0; i < to; i++) {
      if (toDim[i] == 0)
        break;
      loopBlockSize *= toDim[i];
    }
 
    for (i = 0; i < from; i++) {
      if (fromDim[i] == 0)
        break;
      copyblocksize *= fromDim[i];
    }
    for (i = 0; i < to; i++) {
      if (toDim[i] == 0)
        break;
      copyblocklimit *= toDim[i];
    }
 
    for (i = 0; i < loopLimit; i++) {
      /* [i1][i2][...][iN][b][...] i = i1 x i2 x ... x iN */
      uint8_t *destptr = outptr + loopBlockSize * toDim[to] * i;
      const uint8_t *srcptr = inptr + loopBlockSize * toDim[to] * i;
 
      for (j = 0; j < toDim[to]; j++) {
        uint8_t *j_destptr = destptr + loopBlockSize * j;
        for (k = 0; k < copyblocklimit; k++) {
          nns_memcpy (j_destptr + copyblocksize * k,
              srcptr + k * copyblocksize * toDim[to] + j * copyblocksize,
              copyblocksize);
        }
      }
    }
  } else {
    ml_loge
        ("tensor-transform/dimchg operation is not permitted if from >= to.\n");
    return GST_FLOW_ERROR;
  }
 
  return GST_FLOW_OK;
}
 
static GstFlowReturn
gst_tensor_transform_typecast (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  gulong i, num;
  gsize in_element_size, out_element_size;
 
  num = gst_tensor_get_element_count (in_info->dimension);
 
#ifdef HAVE_ORC
  if (filter->acceleration) {
    orc_typecast (inptr, outptr, num, in_info->type, out_info->type);
    return GST_FLOW_OK;
  }
#endif
 
  in_element_size = gst_tensor_get_element_size (in_info->type);
  out_element_size = gst_tensor_get_element_size (out_info->type);
 
  for (i = 0; i < num; ++i) {
    gst_tensor_data_raw_typecast (
        (gpointer) (inptr + in_element_size * i), in_info->type,
        (gpointer) (outptr + out_element_size * i), out_info->type);
  }
 
  return GST_FLOW_OK;
}
 
static GstFlowReturn
gst_tensor_transform_arithmetic (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  gulong i, num, j, ch;
  gsize in_element_size, out_element_size;
 
  GSList *walk;
  tensor_transform_operator_s *op_s;
  tensor_data_s value;
 
  num = gst_tensor_get_element_count (in_info->dimension);
 
#ifdef HAVE_ORC
  if (filter->acceleration) {
    walk = filter->operators;
    orc_typecast (inptr, outptr, num, in_info->type, out_info->type);
 
    if (!filter->data_arithmetic.per_channel_arith) {
      while (walk) {
        op_s = (tensor_transform_operator_s *) walk->data;
 
        if (op_s->op != GTT_OP_TYPECAST) {
          gst_tensor_data_typecast (&op_s->value, out_info->type);
          orc_operator (outptr, num, &op_s->value, op_s->op);
        }
 
        walk = g_slist_next (walk);
      }
    } else {
      gsize typesize = 0;
      guint ch_dim = filter->data_arithmetic.ch_dim;
      gsize ch_offset, ch_size = 1;
      uint8_t *tmp_outptr = NULL;
 
      for (i = 0; i < ch_dim; ++i) {
        ch_size *= in_info->dimension[i];
      }
      ch_offset = ch_size * in_info->dimension[ch_dim];
      orc_typesize (typesize, out_info->type);
 
      while (walk) {
        op_s = (tensor_transform_operator_s *) walk->data;
        if (op_s->op == GTT_OP_TYPECAST) {
          walk = g_slist_next (walk);
          continue;
        }
 
        if (op_s->applying_ch == -1) {
          gst_tensor_data_typecast (&op_s->value, out_info->type);
          orc_operator (outptr, num, &op_s->value, op_s->op);
        } else {
          for (i = 0; i < num / ch_offset; ++i) {
            tmp_outptr =
                outptr + (ch_size * op_s->applying_ch +
                ch_offset * i) * typesize;
            gst_tensor_data_typecast (&op_s->value, out_info->type);
            orc_operator (tmp_outptr, ch_size, &op_s->value, op_s->op);
          }
        }
        walk = g_slist_next (walk);
      }
    }
    return GST_FLOW_OK;
  }
#endif
 
  in_element_size = gst_tensor_get_element_size (in_info->type);
  out_element_size = gst_tensor_get_element_size (out_info->type);
 
  /* per-channel */
  if (filter->data_arithmetic.per_channel_arith) {
    guint ch_dim = filter->data_arithmetic.ch_dim;
    gsize ch_offset, ch_size = 1;
    for (i = 0; i < ch_dim; ++i) {
      ch_size *= in_info->dimension[i];
    }
    ch_offset = ch_size * in_info->dimension[ch_dim];
 
    for (i = 0; i < num / ch_offset; ++i) {
      for (ch = 0; ch < in_info->dimension[ch_dim]; ++ch) {
        for (j = 0; j < ch_size; ++j) {
          gulong data_idx = (i * ch_offset) + (ch * ch_size) + j;
          gst_tensor_data_set (&value, in_info->type,
              (gpointer) (inptr + in_element_size * data_idx));
 
          walk = filter->operators;
          while (walk) {
            op_s = (tensor_transform_operator_s *) walk->data;
            switch (op_s->op) {
              case GTT_OP_TYPECAST:
                gst_tensor_data_typecast (&value, op_s->value.type);
                break;
              case GTT_OP_ADD:
              case GTT_OP_MUL:
              case GTT_OP_DIV:
              {
                gst_tensor_data_typecast (&op_s->value, value.type);
 
                if (op_s->applying_ch == (int) ch || op_s->applying_ch == -1) {
                  gst_tensor_transform_do_operator (filter, &value,
                      &op_s->value, op_s->op);
                }
                break;
              }
              default:
                g_assert (0);
                return GST_FLOW_ERROR;
            }
 
            walk = g_slist_next (walk);
          }
 
          /* set output value */
          g_assert (out_info->type == value.type);
          gst_tensor_data_get (&value, outptr + out_element_size * data_idx);
        }
      }
    }
 
    return GST_FLOW_OK;
  }
 
  for (i = 0; i < num; ++i) {
    /* init value with input tensor type */
    gst_tensor_data_set (&value, in_info->type,
        (gpointer) (inptr + in_element_size * i));
 
    walk = filter->operators;
    while (walk) {
      op_s = (tensor_transform_operator_s *) walk->data;
 
      switch (op_s->op) {
        case GTT_OP_TYPECAST:
          gst_tensor_data_typecast (&value, op_s->value.type);
          break;
        case GTT_OP_ADD:
        case GTT_OP_MUL:
        case GTT_OP_DIV:
          gst_tensor_data_typecast (&op_s->value, value.type);
          gst_tensor_transform_do_operator (filter, &value, &op_s->value,
              op_s->op);
          break;
        default:
          g_assert (0);
          return GST_FLOW_ERROR;
      }
 
      walk = g_slist_next (walk);
    }
 
    /* set output value */
    g_assert (out_info->type == value.type);
    gst_tensor_data_get (&value, outptr + out_element_size * i);
  }
 
  return GST_FLOW_OK;
}
 
#define transposeloop(cl,ck,cj,ci,sl,sk,sj,si,typesize) do { \
    size_t i, j, k, l;                                  \
    int inidx = 0, outidx=0;                            \
    for(cl=0;cl<sl;cl++)                      \
      for(ci=0;ci<si;ci++)                    \
        for(cj=0;cj<sj;cj++)                  \
          for(ck=0;ck<sk;ck++){               \
            const uint8_t *_in; \
            uint8_t *_out; \
            outidx = si*sj*sk*cl + sj*sk*ci + sk*cj + ck; \
            inidx = SK*SJ*SI*l + SJ*SI*k + SI*j + i; \
            _in = inptr + inidx * typesize; \
            _out = outptr + outidx * typesize; \
            nns_memcpy(_out, _in, typesize); \
      }                                                      \
  } while(0);
 
static GstFlowReturn
gst_tensor_transform_transpose (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  int i, from, to;
  gboolean checkdim = FALSE;
  uint32_t *fromDim = in_info->dimension;
  gsize type_size = gst_tensor_get_element_size (in_info->type);
  gsize indexI, indexJ, SL, SI, SJ, SK;
  UNUSED (out_info);
 
  for (i = 0; i < NNS_TENSOR_TRANSPOSE_RANK_LIMIT; i++) {
    from = i;
    to = filter->data_transpose.trans_order[i];
    if (from != to) {
      checkdim = TRUE;
      break;
    }
  }
 
  if (!checkdim) {
    nns_memcpy (outptr, inptr, gst_tensor_info_get_size (in_info));
    GST_WARNING_OBJECT (filter,
        "Calling tensor_transform with high memcpy overhead WITHOUT any effects!");
    return GST_FLOW_OK;
  }
 
  indexI = filter->data_transpose.trans_order[0];
  indexJ = filter->data_transpose.trans_order[1];
  SL = fromDim[3] > 0 ? fromDim[3] : 1;
  SI = fromDim[0] > 0 ? fromDim[0] : 1;
  SJ = fromDim[1] > 0 ? fromDim[1] : 1;
  SK = fromDim[2] > 0 ? fromDim[2] : 1;
 
  switch (indexI) {
    case 0:
      if (indexJ == 1) {
        transposeloop (l, i, j, k, SL, SI, SJ, SK, type_size);
      } else {
        transposeloop (l, i, k, j, SL, SI, SK, SJ, type_size);
      }
      break;
    case 1:
      if (indexJ == 0) {
        transposeloop (l, j, i, k, SL, SJ, SI, SK, type_size);
      } else {
        transposeloop (l, j, k, i, SL, SJ, SK, SI, type_size);
      }
      break;
    case 2:
      if (indexJ == 0) {
        transposeloop (l, k, i, j, SL, SK, SI, SJ, type_size);
      } else {
        transposeloop (l, k, j, i, SL, SK, SJ, SI, type_size);
      }
      break;
  }
 
  return GST_FLOW_OK;
}
 
static GstFlowReturn
gst_tensor_transform_stand (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  GstFlowReturn ret = GST_FLOW_OK;
  gsize in_element_size, out_element_size, data_size, ch_size;
  gulong i, num, data_idx, ch;
  gdouble tmp, *average, *std;
 
  in_element_size = gst_tensor_get_element_size (in_info->type);
  out_element_size = gst_tensor_get_element_size (out_info->type);
  num = gst_tensor_get_element_count (in_info->dimension);
 
  data_size = gst_tensor_info_get_size (in_info);
  ch_size = in_info->dimension[0];
 
  /* calc average and std */
  average = std = NULL;
  if (filter->data_stand.per_channel) {
    gst_tensor_data_raw_average_per_channel ((gpointer) inptr, data_size,
        in_info->type, in_info->dimension, &average);
    /* calculate std only for default mode */
    if (filter->data_stand.mode == STAND_DEFAULT)
      gst_tensor_data_raw_std_per_channel ((gpointer) inptr, data_size,
          in_info->type, in_info->dimension, average, &std);
  } else {
    gst_tensor_data_raw_average ((gpointer) inptr, data_size,
        in_info->type, &average);
    /* calculate std only for default mode */
    if (filter->data_stand.mode == STAND_DEFAULT)
      gst_tensor_data_raw_std ((gpointer) inptr, data_size, in_info->type,
          average, &std);
  }
 
  switch (filter->data_stand.mode) {
    case STAND_DEFAULT:
    {
      if (!filter->data_stand.per_channel) {
        for (i = 0; i < num; i++) {
          data_idx = in_element_size * i;
          gst_tensor_data_raw_typecast ((gpointer) (inptr + data_idx),
              in_info->type, &tmp, _NNS_FLOAT64);
 
          tmp = fabs ((tmp - *average) / *std);
 
          data_idx = out_element_size * i;
          gst_tensor_data_raw_typecast (&tmp, _NNS_FLOAT64,
              (gpointer) (outptr + data_idx), out_info->type);
        }
      } else {
        for (ch = 0; ch < ch_size; ++ch) {
          for (i = 0; i < num / ch_size; i++) {
            data_idx = in_element_size * ((i * ch_size) + ch);
            gst_tensor_data_raw_typecast ((gpointer) (inptr + data_idx),
                in_info->type, &tmp, _NNS_FLOAT64);
 
            tmp = fabs ((tmp - average[ch]) / std[ch]);
 
            data_idx = out_element_size * ((i * ch_size) + ch);
            gst_tensor_data_raw_typecast (&tmp, _NNS_FLOAT64,
                (gpointer) (outptr + data_idx), out_info->type);
          }
        }
      }
      break;
    }
    case STAND_DC_AVERAGE:
    {
      if (!filter->data_stand.per_channel) {
        for (i = 0; i < num; i++) {
          data_idx = in_element_size * i;
          gst_tensor_data_raw_typecast ((gpointer) (inptr + data_idx),
              in_info->type, &tmp, _NNS_FLOAT64);
 
          tmp -= *average;
 
          data_idx = out_element_size * i;
          gst_tensor_data_raw_typecast (&tmp, _NNS_FLOAT64,
              (gpointer) (outptr + data_idx), out_info->type);
        }
      } else {
        for (ch = 0; ch < ch_size; ++ch) {
          for (i = 0; i < num / ch_size; i++) {
            data_idx = in_element_size * ((i * ch_size) + ch);
            gst_tensor_data_raw_typecast ((gpointer) (inptr + data_idx),
                in_info->type, &tmp, _NNS_FLOAT64);
 
            tmp -= average[ch];
 
            data_idx = out_element_size * ((i * ch_size) + ch);
            gst_tensor_data_raw_typecast (&tmp, _NNS_FLOAT64,
                (gpointer) (outptr + data_idx), out_info->type);
          }
        }
      }
      break;
    }
    default:
      GST_ERROR_OBJECT (filter, "Cannot identify mode\n");
      ret = GST_FLOW_ERROR;
  }
 
  g_free (average);
  g_free (std);
 
  return ret;
}
 
static GstFlowReturn
gst_tensor_transform_clamp (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info,
    const uint8_t * inptr, uint8_t * outptr)
{
  gsize in_element_size, out_element_size;
  gulong i, num, data_idx;
  gdouble tmp;
 
  in_element_size = gst_tensor_get_element_size (in_info->type);
  out_element_size = gst_tensor_get_element_size (out_info->type);
  num = gst_tensor_get_element_count (in_info->dimension);
 
  for (i = 0; i < num; ++i) {
    data_idx = in_element_size * i;
    gst_tensor_data_raw_typecast ((gpointer) (inptr + data_idx), in_info->type,
        &tmp, _NNS_FLOAT64);
 
    tmp = CLAMP (tmp, filter->data_clamp.min, filter->data_clamp.max);
 
    data_idx = out_element_size * i;
    gst_tensor_data_raw_typecast (&tmp, _NNS_FLOAT64, outptr + data_idx,
        out_info->type);
  }
 
  return GST_FLOW_OK;
}
 
static GstFlowReturn
gst_tensor_transform_padding (GstTensorTransform * filter,
    GstTensorInfo * in_info, GstTensorInfo * out_info, const uint8_t * inptr,
    uint8_t * outptr)
{
  gsize element_size, in_loop_size, out_loop_size, copy_block_size;
  guint i, j, k, left, top, front, loop_limit = 1;
  element_size = gst_tensor_get_element_size (in_info->type);
 
  in_loop_size = (gsize) in_info->dimension[2] * in_info->dimension[1]
      * in_info->dimension[0] * element_size;
  out_loop_size =(gsize) out_info->dimension[2] * out_info->dimension[1]
      * out_info->dimension[0] * element_size;
  copy_block_size = in_info->dimension[0] * element_size;
 
  for (i = NNS_TENSOR_PADDING_RANK_LIMIT; i < NNS_TENSOR_RANK_LIMIT; i++) {
    if (in_info->dimension[i] == 0)
      break;
    loop_limit *= in_info->dimension[i];
  }
 
  left = filter->data_padding.pad[PADDING_LEFT];
  top = filter->data_padding.pad[PADDING_TOP];
  front = filter->data_padding.pad[PADDING_FRONT];
 
  memset (outptr, 0, out_loop_size * loop_limit);
 
  for (i = 0; i < loop_limit; i++)
    for (j = 0; j < in_info->dimension[2]; j++)
      for (k = 0; k < in_info->dimension[1]; k++) {
        guint in_idx = j * in_info->dimension[1] * in_info->dimension[0]
            + k * in_info->dimension[0];
        guint out_idx = j * out_info->dimension[1] * out_info->dimension[0]
            + k * out_info->dimension[0];
 
        out_idx += left + top * out_info->dimension[0]
            + front * out_info->dimension[1] * out_info->dimension[0];
 
        memcpy (outptr + out_idx * element_size + out_loop_size * i,
            inptr + in_idx * element_size + in_loop_size * i, copy_block_size);
      }
 
  return GST_FLOW_OK;
}
 
static GstFlowReturn
gst_tensor_transform_transform (GstBaseTransform * trans,
    GstBuffer * inbuf, GstBuffer * outbuf)
{
  GstTensorTransform *filter;
  GstTensorInfo *in_info, *out_info;
  GstFlowReturn res = GST_FLOW_ERROR;
  GstMemory *in_mem[NNS_TENSOR_SIZE_LIMIT] = { 0, };
  GstMemory *out_mem[NNS_TENSOR_SIZE_LIMIT] = { 0, };
  GstMapInfo in_map[NNS_TENSOR_SIZE_LIMIT];
  GstMapInfo out_map[NNS_TENSOR_SIZE_LIMIT];
  uint8_t *inptr, *outptr;
  guint i, num_tensors, num_mems;
  gsize buf_size, hsize;
  GstTensorMetaInfo meta;
  GstTensorInfo in_flex_info, out_flex_info;
  gboolean in_flexible, out_flexible;
 
  filter = GST_TENSOR_TRANSFORM_CAST (trans);
 
  g_return_val_if_fail (filter->loaded, GST_FLOW_ERROR);
  inbuf = gst_tensor_buffer_from_config (inbuf, &filter->in_config);
 
  in_flexible =
      gst_tensor_pad_caps_is_flexible (GST_BASE_TRANSFORM_SINK_PAD (trans));
  out_flexible =
      gst_tensor_pad_caps_is_flexible (GST_BASE_TRANSFORM_SRC_PAD (trans));
 
  num_mems = gst_tensor_buffer_get_count (inbuf);
  if (in_flexible) {
    num_tensors = num_mems;
    g_return_val_if_fail (out_flexible, GST_FLOW_ERROR);
  } else {
    num_tensors = filter->in_config.info.num_tensors;
    g_return_val_if_fail (num_mems == num_tensors, GST_FLOW_ERROR);
  }
 
  for (i = 0; i < num_tensors; i++) {
    in_info = gst_tensors_info_get_nth_info (&filter->in_config.info, i);
    out_info = gst_tensors_info_get_nth_info (&filter->out_config.info, i);
 
    if (filter->apply && !g_list_find (filter->apply, GINT_TO_POINTER (i))) {
      GstMemory *mem = gst_tensor_buffer_get_nth_memory (inbuf, i);
 
      if (!in_flexible && out_flexible) {
        GstMemory *old = mem;
 
        /* append meta */
        gst_tensor_info_convert_to_meta (out_info, &meta);
        mem = gst_tensor_meta_info_append_header (&meta, old);
        gst_memory_unref (old);
      }
 
      gst_tensor_buffer_append_memory (outbuf, mem, out_info);
      continue;
    }
 
    /* parse input buffer */
    in_mem[i] = gst_tensor_buffer_get_nth_memory (inbuf, i);
    if (!gst_memory_map (in_mem[i], &in_map[i], GST_MAP_READ)) {
      ml_loge ("Cannot map input buffer to gst-buf at tensor-transform.\n");
      res = GST_FLOW_ERROR;
      goto done;
    }
    inptr = in_map[i].data;
 
    if (in_flexible) {
      in_info = &in_flex_info;
      out_info = &out_flex_info;
 
      gst_tensor_meta_info_parse_header (&meta, inptr);
      if (!gst_tensor_meta_info_convert (&meta, in_info)) {
        res = GST_FLOW_ERROR;
        goto done;
      }
 
      gst_tensor_transform_convert_dimension (filter, GST_PAD_SINK,
          i, in_info, out_info);
 
      hsize = gst_tensor_meta_info_get_header_size (&meta);
      inptr += hsize;
    }
 
    /* prepare output buffer */
    buf_size = gst_tensor_info_get_size (out_info);
    if (out_flexible) {
      gst_tensor_info_convert_to_meta (out_info, &meta);
      hsize = gst_tensor_meta_info_get_header_size (&meta);
      buf_size += hsize;
    }
 
    out_mem[i] = gst_allocator_alloc (NULL, buf_size, NULL);
    gst_tensor_buffer_append_memory (outbuf, out_mem[i], out_info);
 
    if (!gst_memory_map (out_mem[i], &out_map[i], GST_MAP_WRITE)) {
      ml_loge ("Cannot map output buffer to gst-buf at tensor-transform.\n");
      res = GST_FLOW_ERROR;
      goto done;
    }
    outptr = out_map[i].data;
 
    if (out_flexible) {
      gst_tensor_meta_info_update_header (&meta, outptr);
      outptr += hsize;
    }
 
    switch (filter->mode) {
      case GTT_DIMCHG:
        res = gst_tensor_transform_dimchg (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_TYPECAST:
        res = gst_tensor_transform_typecast (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_ARITHMETIC:
        res = gst_tensor_transform_arithmetic (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_TRANSPOSE:
        res = gst_tensor_transform_transpose (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_STAND:
        res = gst_tensor_transform_stand (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_CLAMP:
        res = gst_tensor_transform_clamp (filter, in_info, out_info,
            inptr, outptr);
        break;
      case GTT_PADDING:
        res = gst_tensor_transform_padding (filter, in_info, out_info,
            inptr, outptr);
        break;
      default:
        ml_loge ("Not supported tensor transform mode");
        res = GST_FLOW_NOT_SUPPORTED;
        goto done;
    }
  }
 
done:
  for (i = 0; i < num_tensors; i++) {
    if (in_mem[i]) {
      gst_memory_unmap (in_mem[i], &in_map[i]);
      gst_memory_unref (in_mem[i]);
    }
    if (out_mem[i])
      gst_memory_unmap (out_mem[i], &out_map[i]);
  }
 
  return res;
}
 
static gboolean
gst_tensor_transform_read_caps (GstTensorTransform * filter,
    const GstCaps * caps, GstTensorsConfig * config)
{
  GstStructure *structure;
  g_return_val_if_fail (config != NULL, FALSE);
 
  structure = gst_caps_get_structure (caps, 0);
 
  if (!gst_tensors_config_from_structure (config, structure)) {
    GST_WARNING_OBJECT (filter, "caps is not tensor %s\n",
        gst_structure_get_name (structure));
    return FALSE;
  }
 
  return gst_tensors_config_validate (config);
}
 
static gboolean
gst_tensor_transform_convert_dimension (GstTensorTransform * filter,
    GstPadDirection direction, guint idx, const GstTensorInfo * in_info,
    GstTensorInfo * out_info)
{
  guint i;
 
  /* copy input info first, then update output info */
  gst_tensor_info_copy (out_info, in_info);
 
  if (filter->apply && !g_list_find (filter->apply, GINT_TO_POINTER (idx)))
    return TRUE;
 
  switch (filter->mode) {
    case GTT_DIMCHG:
    {
      unsigned int from = filter->data_dimchg.from;
      unsigned int to = filter->data_dimchg.to;
 
      if (direction == GST_PAD_SINK) {
        for (i = 0; i < NNS_TENSOR_RANK_LIMIT; i++) {
          if ((i < from && i < to) || (i > from && i > to) || from == to) {
            out_info->dimension[i] = in_info->dimension[i];
          } else if (i == to) {
            out_info->dimension[i] = in_info->dimension[from];
          } else if (from > to) {
            g_assert (i > 0 && i > to);
            out_info->dimension[i] = in_info->dimension[i - 1];
          } else {
            g_assert (i < to && i < (NNS_TENSOR_RANK_LIMIT - 1));
            out_info->dimension[i] = in_info->dimension[i + 1];
          }
        }
      } else {
        for (i = 0; i < NNS_TENSOR_RANK_LIMIT; i++) {
          if ((i < from && i < to) || (i > from && i > to) || from == to) {
            out_info->dimension[i] = in_info->dimension[i];
          } else if (i == from) {
            out_info->dimension[i] = in_info->dimension[to];
          } else if (from > to) {
            g_assert (i < from && i < (NNS_TENSOR_RANK_LIMIT - 1));
            out_info->dimension[i] = in_info->dimension[i + 1];
          } else {
            g_assert (i > 0 && i > from);
            out_info->dimension[i] = in_info->dimension[i - 1];
          }
        }
      }
      break;
    }
    case GTT_TYPECAST:
      if (direction == GST_PAD_SINK) {
        out_info->type = filter->data_typecast.to;
      } else {
        /* cannot get the incoming data type on sink pad */
        out_info->type = _NNS_END;
      }
      break;
 
    case GTT_ARITHMETIC:
      /* check arith mode option has typecast operator */
      if (filter->data_arithmetic.out_type != _NNS_END) {
        if (direction == GST_PAD_SINK) {
          out_info->type = filter->data_arithmetic.out_type;
        } else {
          /* cannot get the incoming data type on sink pad */
          out_info->type = _NNS_END;
        }
      }
      break;
 
    case GTT_TRANSPOSE:
      if (direction == GST_PAD_SINK) {
        for (i = 0; i < NNS_TENSOR_TRANSPOSE_RANK_LIMIT; i++) {
          out_info->dimension[i] =
              in_info->dimension[filter->data_transpose.trans_order[i]];
        }
      } else {
        for (i = 0; i < NNS_TENSOR_TRANSPOSE_RANK_LIMIT; i++) {
          g_assert (filter->data_transpose.trans_order[i] <
              NNS_TENSOR_RANK_LIMIT);
          out_info->dimension[filter->data_transpose.trans_order[i]] =
              in_info->dimension[i];
        }
      }
      break;
 
    case GTT_STAND:
      if (direction == GST_PAD_SINK) {
        if (filter->data_stand.out_type != _NNS_END)
          out_info->type = filter->data_stand.out_type;
      } else {
        /* cannot get the incoming data type on sink pad */
        out_info->type = _NNS_END;
      }
      break;
 
    case GTT_CLAMP:
      /* same tensors info, do nothing. */
      break;
 
    case GTT_PADDING:
      if (direction == GST_PAD_SINK) {
        out_info->dimension[0] +=
            filter->data_padding.pad[PADDING_LEFT] +
            filter->data_padding.pad[PADDING_RIGHT];
        out_info->dimension[1] +=
            filter->data_padding.pad[PADDING_TOP] +
            filter->data_padding.pad[PADDING_BOTTOM];
        out_info->dimension[2] +=
            filter->data_padding.pad[PADDING_FRONT] +
            filter->data_padding.pad[PADDING_BACK];
      }
      break;
    default:
      return FALSE;
  }
 
  return TRUE;
}
 
static GstCaps *
gst_tensor_transform_transform_caps (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, GstCaps * filtercap)
{
  GstTensorTransform *filter;
  GstCaps *result = NULL;
  GstStructure *structure;
  guint i, j;
 
  filter = GST_TENSOR_TRANSFORM_CAST (trans);
 
  silent_debug (filter, "Calling TransformCaps, direction = %d\n", direction);
  silent_debug_caps (filter, caps, "from");
  silent_debug_caps (filter, filtercap, "filter");
 
  result = gst_caps_new_empty ();
  for (i = 0; i < gst_caps_get_size (caps); i++) {
    GstTensorsConfig in_config, out_config;
    GstTensorInfo *in_info, *out_info;
    gboolean is_types_not_fixed = FALSE;
    GstCaps *result_aux = gst_caps_new_empty ();
 
    gst_tensors_config_init (&out_config);
 
    structure = gst_caps_get_structure (caps, i);
    gst_tensors_config_from_structure (&in_config, structure);
 
    if (gst_tensors_config_is_flexible (&in_config)) {
      /* output caps is also flexible */
      out_config.info.format = _NNS_TENSOR_FORMAT_FLEXIBLE;
    } else {
      for (j = 0; j < in_config.info.num_tensors; j++) {
        in_info = gst_tensors_info_get_nth_info (&in_config.info, j);
        out_info = gst_tensors_info_get_nth_info (&out_config.info, j);
 
        gst_tensor_transform_convert_dimension (filter, direction,
            j, in_info, out_info);
        if (out_info->type == _NNS_END) {
          /* types cannot be specified */
          is_types_not_fixed = TRUE;
        }
      }
    }
 
    out_config.rate_d = in_config.rate_d;
    out_config.rate_n = in_config.rate_n;
    out_config.info.num_tensors = in_config.info.num_tensors;
 
    if (gst_structure_has_name (structure, NNS_MIMETYPE_TENSOR)) {
      gst_caps_append (result_aux, gst_tensor_caps_from_config (&out_config));
    } else {
      gst_caps_append (result_aux, gst_tensors_caps_from_config (&out_config));
 
      /* remove `types` field from caps */
      if (is_types_not_fixed) {
        GstStructure *s = gst_caps_get_structure (result_aux, 0);
        gst_structure_remove_field (s, "types");
      }
    }
 
    gst_caps_append (result, result_aux);
 
    gst_tensors_config_free (&in_config);
    gst_tensors_config_free (&out_config);
  }
 
  if (filtercap && gst_caps_get_size (filtercap) > 0) {
    GstCaps *intersection;
 
    intersection =
        gst_caps_intersect_full (result, filtercap, GST_CAPS_INTERSECT_FIRST);
 
    gst_caps_unref (result);
    result = intersection;
  }
 
  silent_debug_caps (filter, result, "to");
  return result;
}
 
static GstCaps *
gst_tensor_transform_fixate_caps (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, GstCaps * othercaps)
{
  GstTensorTransform *filter;
  GstCaps *result;
 
  filter = GST_TENSOR_TRANSFORM_CAST (trans);
 
  silent_debug (filter, "Calling FixateCaps, direction = %d\n", direction);
  silent_debug_caps (filter, caps, "caps");
  silent_debug_caps (filter, othercaps, "othercaps");
 
  result =
      gst_tensor_transform_transform_caps (trans, direction, caps, othercaps);
  gst_caps_unref (othercaps);
 
  result = gst_caps_make_writable (result);
  result = gst_caps_fixate (result);
 
  silent_debug_caps (filter, result, "result");
  return result;
}
 
static gboolean
gst_tensor_transform_set_caps (GstBaseTransform * trans,
    GstCaps * incaps, GstCaps * outcaps)
{
  GstTensorTransform *filter;
  GstTensorsConfig in_config, out_config;
  GstTensorsConfig config;
  GstTensorInfo *in_info, *out_info;
  gboolean in_flexible, out_flexible;
  gboolean allowed = FALSE;
  guint i;
 
  filter = GST_TENSOR_TRANSFORM_CAST (trans);
 
  silent_debug (filter, "Calling SetCaps\n");
  silent_debug_caps (filter, incaps, "incaps");
  silent_debug_caps (filter, outcaps, "outcaps");
 
  if (!gst_tensor_transform_read_caps (filter, incaps, &in_config)) {
    GST_ERROR_OBJECT (filter, "Cannot read cap of incaps\n");
    goto error;
  }
 
  if (!gst_tensor_transform_read_caps (filter, outcaps, &out_config)) {
    GST_ERROR_OBJECT (filter, "Cannot read cap of outcaps\n");
    goto error;
  }
 
  in_flexible = gst_tensors_config_is_flexible (&in_config);
  out_flexible = gst_tensors_config_is_flexible (&out_config);
 
  /* compare type and dimension */
  gst_tensors_config_init (&config);
  config.info.format = out_config.info.format;
 
  config.rate_n = in_config.rate_n;
  config.rate_d = in_config.rate_d;
  config.info.num_tensors = in_config.info.num_tensors;
 
  if (!in_flexible) {
    for (i = 0; i < in_config.info.num_tensors; i++) {
      in_info = gst_tensors_info_get_nth_info (&in_config.info, i);
      out_info = gst_tensors_info_get_nth_info (&config.info, i);
 
      if (!gst_tensor_transform_convert_dimension (filter, GST_PAD_SINK,
              i, in_info, out_info)) {
        GST_ERROR_OBJECT (filter,
            "Tensor info is not matched with given properties.");
        goto error;
      }
    }
  }
 
  if (out_flexible) {
    GST_INFO_OBJECT (filter, "Output tensor is flexible.");
 
    /* set output configuration if input is static */
    if (!in_flexible)
      out_config = config;
  } else if (!gst_tensors_config_is_equal (&out_config, &config)) {
    GST_ERROR_OBJECT (filter,
        "Tensor info is not matched with given properties.\n");
    goto error;
  }
 
  /* set in/out tensor info */
  filter->in_config = in_config;
  filter->out_config = out_config;
  allowed = TRUE;
 
error:
  if (!allowed)
    GST_ERROR_OBJECT (filter, "Set Caps Failed!\n");
 
  return allowed;
}
 
static gboolean
gst_tensor_transform_transform_size (GstBaseTransform * trans,
    GstPadDirection direction, GstCaps * caps, gsize size, GstCaps * othercaps,
    gsize * othersize)
{
  UNUSED (trans);
  UNUSED (direction);
  UNUSED (caps);
  UNUSED (size);
  UNUSED (othercaps);
  *othersize = 0;
 
  return TRUE;
}