--- /dev/null
+/* SPDX-License-Identifier: LGPL-2.1-only */
+/**
+ * GStreamer / NNStreamer tensor_converter subplugin, "Flexbuffer"
+ * Copyright (C) 2021 Gichan Jang <gichan2.jang@samsung.com>
+ */
+/**
+ * @file tensor_converter_flexbuf.cc
+ * @date 12 Mar 2021
+ * @brief NNStreamer tensor-converter subplugin, "flexbuffer",
+ * which converts flexbuffers byte stream to tensors.
+ * @see https://github.com/nnstreamer/nnstreamer
+ * @author Gichan Jang <gichan2.jang@samsung.com>
+ * @bug No known bugs except for NYI items
+ *
+ */
+/**
+ * SECTION:tensor_converter::flexbuf
+ * @see https://google.github.io/flatbuffers/flexbuffers.html
+ *
+ * tensor_converter::flexbuf converts flexbuffers to tensors stream..
+ * The output is always in the format of other/tensor or other/tensors.
+ *
+ * Binary format of the flexbuffers for tensors (default in nnstreamer).
+ * Each data is represented in `KEY : TYPE | <VALUE>` form.
+ *
+ * Map {
+ * "num_tensors" : UInt32 | <The number of tensors>
+ * "rate_n" : Int32 | <Framerate numerator>
+ * "rate_d" : Int32 | <Framerate denominator>
+ * "tensor_#": Vector | { String | <tensor name>,
+ * Int32 | <data type>,
+ * Vector | <tensor dimension>,
+ * Blob | <tensor data>
+ * }
+ * }
+ */
+
+#include <glib.h>
+#include <nnstreamer_log.h>
+#include <nnstreamer_plugin_api.h>
+#include <nnstreamer_plugin_api_converter.h>
+#include <flatbuffers/flexbuffers.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+void init_flxc (void) __attribute__ ((constructor));
+void fini_flxc (void) __attribute__ ((destructor));
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
+
+/** @brief tensor converter plugin's NNStreamerExternalConverter callback */
+static GstCaps *
+flxc_query_caps (const GstTensorsConfig *config)
+{
+ return gst_caps_from_string (GST_FLEXBUF_CAP_DEFAULT);
+}
+
+/** @brief tensor converter plugin's NNStreamerExternalConverter callback */
+static gboolean
+flxc_get_out_config (const GstCaps *in_cap, GstTensorsConfig *config)
+{
+ GstStructure *structure;
+ g_return_val_if_fail (config != NULL, FALSE);
+ gst_tensors_config_init (config);
+ g_return_val_if_fail (in_cap != NULL, FALSE);
+
+ structure = gst_caps_get_structure (in_cap, 0);
+ g_return_val_if_fail (structure != NULL, FALSE);
+
+ /* All tensor info should be updated later in chain function. */
+ config->info.info[0].type = _NNS_UINT8;
+ config->info.num_tensors = 1;
+ if (gst_tensor_parse_dimension ("1:1:1:1", config->info.info[0].dimension) == 0) {
+ ml_loge ("Failed to set initial dimension for subplugin");
+ return FALSE;
+ }
+
+ if (gst_structure_has_field (structure, "framerate")) {
+ gst_structure_get_fraction (structure, "framerate", &config->rate_n, &config->rate_d);
+ } else {
+ /* cannot get the framerate */
+ config->rate_n = 0;
+ config->rate_d = 1;
+ }
+ return TRUE;
+}
+
+/** @brief tensor converter plugin's NNStreamerExternalConverter callback
+ */
+static GstBuffer *
+flxc_convert (GstBuffer *in_buf, gsize *frame_size, guint *frames_in, GstTensorsConfig *config)
+{
+ GstBuffer *out_buf = NULL;
+ GstMemory *in_mem, *out_mem;
+ GstMapInfo in_info;
+ guint mem_size;
+ gpointer mem_data;
+
+ in_mem = gst_buffer_peek_memory (in_buf, 0);
+ if (gst_memory_map (in_mem, &in_info, GST_MAP_READ) == FALSE) {
+ ml_loge ("Cannot map input memory / tensor_converter::flexbuf.\n");
+ return NULL;
+ }
+
+ flexbuffers::Map tensors = flexbuffers::GetRoot (in_info.data, in_info.size).AsMap ();
+ config->info.num_tensors = tensors["num_tensors"].AsUInt32 ();
+
+ if (config->info.num_tensors > NNS_TENSOR_SIZE_LIMIT) {
+ nns_loge ("The number of tensors is limited to %d", NNS_TENSOR_SIZE_LIMIT);
+ goto done;
+ }
+ config->rate_n = tensors["rate_n"].AsInt32 ();
+ config->rate_d = tensors["rate_d"].AsInt32 ();
+ out_buf = gst_buffer_new ();
+ *frame_size = 0;
+ *frames_in = 1;
+
+ for (guint i = 0; i < config->info.num_tensors; i++) {
+ gchar * tensor_key = g_strdup_printf ("tensor_%d", i);
+ flexbuffers::Vector tensor = tensors[tensor_key].AsVector ();
+ config->info.info[i].name = g_strdup (tensor[0].AsString ().c_str ());
+ config->info.info[i].type = (tensor_type) tensor[1].AsInt32 ();
+
+ flexbuffers::TypedVector dim = tensor[2].AsTypedVector ();
+ for (guint j = 0; j < NNS_TENSOR_RANK_LIMIT; j++) {
+ config->info.info[i].dimension[j] = dim[j].AsInt32 ();
+ }
+ flexbuffers::Blob tensor_data = tensor[3].AsBlob ();
+ mem_size = tensor_data.size ();
+ *frame_size += mem_size;
+
+ /** @todo Consider removing memory copies for better performance. */
+ mem_data = g_memdup (tensor_data.data (), mem_size);
+
+ out_mem = gst_memory_new_wrapped (GST_MEMORY_FLAG_READONLY, mem_data,
+ mem_size, 0, mem_size, mem_data, g_free);
+
+ gst_buffer_append_memory (out_buf, out_mem);
+ g_free (tensor_key);
+ }
+
+ /** copy timestamps */
+ gst_buffer_copy_into (
+ out_buf, in_buf, (GstBufferCopyFlags)GST_BUFFER_COPY_METADATA, 0, -1);
+done:
+ gst_memory_unmap (in_mem, &in_info);
+
+ return out_buf;
+}
+
+static const gchar converter_subplugin_flexbuf[] = "flexbuf";
+
+/** @brief flexbuffer tensor converter sub-plugin NNStreamerExternalConverter instance */
+static NNStreamerExternalConverter flexBuf = {
+ .name = converter_subplugin_flexbuf,
+ .convert = flxc_convert,
+ .get_out_config = flxc_get_out_config,
+ .query_caps = flxc_query_caps
+};
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+/** @brief Initialize this object for tensor converter sub-plugin */
+void
+init_flxc (void)
+{
+ registerExternalConverter (&flexBuf);
+}
+
+/** @brief Destruct this object for tensor converter sub-plugin */
+void
+fini_flxc (void)
+{
+ unregisterExternalConverter (flexBuf.name);
+}
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
projects / platform / upstream / nnstreamer.git / commitdiff