* Copyright (C) 2020 Dongju Chae <dongju.chae@samsung.com>
*/
/**
- * @file ne-host-handler.cc
+ * @file ne-handler.cc
* @date 03 Apr 2020
- * @brief Implementation of APIs to access NPU from Host
+ * @brief Impelemetation of NPU Engine entrypoint that handles APIs from host
* @see https://code.sec.samsung.net/confluence/display/ODLC/2020+Overall+Software+Stack
* @author Dongju Chae <dongju.chae@samsung.com>
* @bug No known bugs except for NYI items
*/
+#include "ne-handler.h"
+#include "ne-data.h"
+
#include <npubinfmt.h>
#include <NPUdrvAPI.h>
#include <CommPlugin.h>
-#include "ne-utils.h"
-#include "ne-mem.h"
-#include "ne-scheduler.h"
-#include "ne-handler.h"
-
#include <string.h>
#include <assert.h>
#define TAG _N2
-#define INIT_HOST_HANDLER(handler, dev) \
- Device *tdev = static_cast <Device *> (dev); \
- if (tdev == nullptr) return -EINVAL; \
- HostHandler *handler = tdev->getHostHandler (); \
- if (handler == nullptr) return -EINVAL;
-
-/** @brief device class. it contains all related instances */
-class Device {
- public:
- /** @brief Factory method to create a trinity device dependong on dev type */
- static Device *createInstance (dev_type device_type, int device_id);
-
- /** @brief constructor of device */
- Device (dev_type type, int id, bool need_model = true)
- : comm_(CommPlugin::getCommPlugin()), type_ (type), id_ (id),
- need_model_ (true), mode_ (NPUASYNC_WAIT), initialized_ (ATOMIC_FLAG_INIT) {}
-
- /** @brief destructor of device */
- virtual ~Device () {}
-
- /** @brief initialization */
- int init () {
- if (!initialized_.test_and_set()) {
- /** create the corresponding driver API */
- api_ = DriverAPI::createDriverAPI (type_, id_);
- if (api_.get() == nullptr) {
- initialized_.clear();
- logerr (TAG, "Failed to create driver API\n");
- return -EINVAL;
- }
-
- handler_.reset (new HostHandler (this));
- scheduler_.reset (new Scheduler (api_.get()));
- mem_ = MemAllocator::createInstance (api_.get());
- }
-
- return 0;
- }
-
- HostHandler *getHostHandler () { return handler_.get(); }
- dev_type getType () { return type_; }
- int getID () { return id_; }
- bool needModel () { return need_model_; }
- void setAsyncMode (npu_async_mode mode) { mode_ = mode; }
-
- HWmem * allocMemory () { return mem_->allocMemory (); }
- void deallocMemory (int dmabuf_fd) { mem_->deallocMemory (dmabuf_fd); }
-
- /** it stops all requests in this device (choose wait or force) */
- int stop (bool force_stop) {
- Request *req = new Request (NPUINPUT_STOP);
- req->setForceStop (force_stop);
- return scheduler_->submitRequest (req);
- }
-
- virtual Model * registerModel (const generic_buffer *model) = 0;
- virtual int run (npu_input_opmode opmode, const Model *model,
- const input_buffers *input, npuOutputNotify cb, void *cb_data,
- uint64_t *sequence) = 0;
-
- protected:
- /** the device instance has ownership of all related components */
- std::unique_ptr<DriverAPI> api_; /**< device api */
- std::unique_ptr<MemAllocator> mem_; /**< memory allocator */
- std::unique_ptr<HostHandler> handler_; /**< host handler */
- std::unique_ptr<Scheduler> scheduler_; /**< scheduler */
-
- CommPlugin& comm_; /**< plugin communicator */
-
- dev_type type_; /**< device type */
- int id_; /**< device id */
- bool need_model_; /**< indicates whether the device needs model */
- npu_async_mode mode_; /**< async run mode */
-
- private:
- std::atomic_flag initialized_;
-};
-
-/** @brief Trinity Vision (TRIV) classs */
-class TrinityVision : public Device {
- public:
- TrinityVision (int id) : Device (NPUCOND_TRIV_CONN_SOCIP, id) {}
- ~TrinityVision () {}
-
- static size_t manipulateData (const Model *model, uint32_t idx, bool is_input,
- void *dst, void *src, size_t size);
-
- Model * registerModel (const generic_buffer *model_buf) {
- Model *model = mem_->allocModel ();
- if (model == nullptr) {
- logerr (TAG, "Failed to allocate model\n");
- return nullptr;
- }
-
- int status;
- if (model_buf->type == BUFFER_DMABUF) {
- model->setDmabuf (model_buf->dmabuf);
- model->setOffset (model_buf->offset);
- model->setSize (model_buf->size);
- } else {
- status = model->alloc (model_buf->size);
- if (status != 0) {
- logerr (TAG, "Failed to allocate model: %d\n", status);
- goto delete_exit;
- }
-
- status = comm_.extractGenericBuffer (model_buf, model->getData(), nullptr);
- if (status != 0) {
- logerr (TAG, "Failed to extract generic buffer: %d\n", status);
- goto delete_exit;
- }
- }
-
- status = model->setMetadata (model->getData());
- if (status != 0)
- goto delete_exit;
-
- model_config_t config;
- config.dmabuf_id = model->getDmabuf();
- config.program_size = model->getMetadata()->getProgramSize();
- config.program_offset_addr = model->getOffset() + model->getMetadata()->getMetaSize();
- config.weight_offset_addr = config.program_offset_addr + config.program_size;
-
- status = api_->setModel (&config);
- if (status != 0)
- goto delete_exit;
-
- return model;
-
-delete_exit:
- delete model;
- return nullptr;
- }
-
- Buffer * prepareInputBuffers (const Model *model, const input_buffers *input) {
- const Metadata *meta = model->getMetadata();
- const generic_buffer *first = &input->bufs[0];
-
- if (meta->getInputNum() != input->num_buffers)
- return nullptr;
-
- Buffer * buffer = mem_->allocBuffer ();
- if (buffer != nullptr) {
- if (first->type == BUFFER_DMABUF) {
- buffer->setDmabuf (first->dmabuf);
- buffer->setOffset (first->offset);
- buffer->setSize (meta->getBufferSize());
- } else {
- int status = buffer->alloc (meta->getBufferSize ());
- if (status != 0) {
- logerr (TAG, "Failed to allocate buffer: %d\n", status);
- delete buffer;
- return nullptr;
- }
- }
- }
-
- buffer->createTensors (meta);
- return buffer;
- }
-
- int run (npu_input_opmode opmode, const Model *model,
- const input_buffers *input, npuOutputNotify cb, void *cb_data,
- uint64_t *sequence) {
- if (opmode != NPUINPUT_HOST)
- return -EINVAL;
-
- Buffer *buffer = prepareInputBuffers (model, input);
- if (buffer == nullptr)
- return -EINVAL;
-
- for (uint32_t idx = 0; idx < input->num_buffers; idx++) {
- auto func = std::bind (TrinityVision::manipulateData, model, idx, true,
- std::placeholders::_1, std::placeholders::_2, std::placeholders::_3);
- int status = comm_.extractGenericBuffer (&input->bufs[idx],
- buffer->getInputTensor(idx)->getData(), func);
- if (status != 0) {
- logerr (TAG, "Failed to feed input buffer: %d\n", status);
- return status;
- }
- }
-
- /** this device uses CMA buffer */
-
- Request *req = new Request (opmode);
- req->setModel (model);
- req->setBuffer (buffer);
- req->setCallback (std::bind (&TrinityVision::callback, this, req, cb, cb_data));
-
- if (sequence)
- *sequence = req->getID();
-
- return scheduler_->submitRequest (req);
- }
-
- void callback (Request *req, npuOutputNotify cb, void *cb_data) {
- const Model *model = req->getModel ();
- Buffer *buffer = req->getBuffer ();
- output_buffers output = {
- .num_buffers = buffer->getOutputNum ()
- };
-
- for (uint32_t idx = 0; idx < output.num_buffers; idx++) {
- uint32_t output_tensor_size = model->getOutputTensorSize (idx);
-
- output.bufs[idx].type = BUFFER_MAPPED;
- output.bufs[idx].size = output_tensor_size;
- /** user needs to free this */
- output.bufs[idx].addr = malloc (output_tensor_size);
-
- auto func = std::bind (TrinityVision::manipulateData, model, idx, false,
- std::placeholders::_1, std::placeholders::_2, std::placeholders::_3);
- int status = comm_.insertGenericBuffer (buffer->getOutputTensor(idx)->getData(),
- &output.bufs[idx], func);
- if (status != 0) {
- logerr (TAG, "Failed to return output buffer: %d\n", status);
- }
- }
-
- cb (&output, req->getID(), cb_data);
- }
-};
-
-/** @brief Trinity Vision2 (TRIV2) classs */
-class TrinityVision2 : public Device {
- public:
- TrinityVision2 (int id) : Device (NPUCOND_TRIV2_CONN_SOCIP, id) {}
- ~TrinityVision2 () {}
-
- static size_t manipulateData (const Model *model, uint32_t idx, bool is_input,
- void *dst, void *src, size_t size) {
- memcpy (dst, src, size);
- return size;
- }
-
- Model * registerModel (const generic_buffer *model_buf) {
- /** TODO: model's weight values are stored in segments */
- return nullptr;
- }
-
- int run (npu_input_opmode opmode, const Model *model,
- const input_buffers *input, npuOutputNotify cb, void *cb_data,
- uint64_t *sequence) {
- if (opmode != NPUINPUT_HOST && opmode != NPUINPUT_HW_RECURRING)
- return -EINVAL;
-
- /** this device uses segment table */
-
- Request *req = new Request (opmode);
- req->setModel (model);
-#if 0
- req->setSegmentTable (segt);
-#endif
- req->setCallback (std::bind (&TrinityVision2::callback, this, req, cb, cb_data));
-
- if (sequence)
- *sequence = req->getID();
-
- return scheduler_->submitRequest (req);
- }
-
- void callback (Request *req, npuOutputNotify cb, void *cb_data) {
- }
-};
-
-/** @brief Trinity Asr (TRIA) classs */
-class TrinityAsr : public Device {
- public:
- TrinityAsr (int id) : Device (NPUCOND_TRIA_CONN_SOCIP, id, false) {}
- ~TrinityAsr () {}
-
- static size_t manipulateData (const Model *model, uint32_t idx, bool is_input,
- void *dst, void *src, size_t size) {
- memcpy (dst, src, size);
- return size;
- }
-
- Model * registerModel (const generic_buffer *model_buf) { return nullptr; }
-
- int run (npu_input_opmode opmode, const Model *model,
- const input_buffers *input, npuOutputNotify cb, void *cb_data,
- uint64_t *sequence) {
- if (opmode != NPUINPUT_HOST)
- return -EINVAL;
-
- /** ASR does not require model and support only a single tensor */
- const generic_buffer *first_buf = &input->bufs[0];
- Buffer * buffer = mem_->allocBuffer ();
- int status;
- if (first_buf->type == BUFFER_DMABUF) {
- buffer->setDmabuf (first_buf->dmabuf);
- buffer->setOffset (first_buf->offset);
- buffer->setSize (first_buf->size);
- } else {
- status = buffer->alloc (first_buf->size);
- if (status != 0) {
- delete buffer;
- return status;
- }
- }
- buffer->createTensors ();
-
- status = comm_.extractGenericBuffer (first_buf,
- buffer->getInputTensor(0)->getData(), nullptr);
- if (status != 0)
- return status;
-
- Request *req = new Request (opmode);
- req->setBuffer (buffer);
- req->setCallback (std::bind (&TrinityAsr::callback, this, req, cb, cb_data));
-
- if (sequence)
- *sequence = req->getID();
-
- return scheduler_->submitRequest (req);
- }
-
- void callback (Request *req, npuOutputNotify cb, void *cb_data) {
- }
-};
-
-#ifdef ENABLE_MANIP
-
-#define do_quantized_memcpy(type) do {\
- idx = 0;\
- if (quant) {\
- while (idx < num_elems) {\
- val = ((type *) src)[idx];\
- val = val / _scale;\
- val += _zero_point;\
- val = (val > 255.0) ? 255.0 : 0.0;\
- ((uint8_t *) dst)[idx++] = (uint8_t) val;\
- }\
- } else {\
- while (idx < num_elems) {\
- val = *(uint8_t *) src;\
- val -= _zero_point;\
- val *= _scale;\
- ((type *) dst)[idx++] = (type) val;\
- dst = (void*)(((uint8_t *) dst) + data_size);\
- src = (void*)(((uint8_t *) src) + 1);\
- }\
- }\
- } while (0)
-
-/**
- * @brief memcpy during quantization
- */
-static void memcpy_with_quant (bool quant, data_type type, float scale, uint32_t zero_point,
- void *dst, const void *src, uint32_t num_elems)
+/** @brief host handler constructor */
+HostHandler::HostHandler (Device *device)
+ : device_(device),
+ /* ignored as we don't use double buffering anymore, but for backward-compatibility */
+ async_mode_ (NPUASYNC_WAIT)
{
- double _scale = (double) scale;
- double _zero_point = (double) zero_point;
- double val;
- uint32_t data_size = get_data_size (type);
- uint32_t idx;
+}
- switch (type) {
- case DATA_TYPE_INT8:
- do_quantized_memcpy (int8_t);
- break;
- case DATA_TYPE_UINT8:
- do_quantized_memcpy (uint8_t);
- break;
- case DATA_TYPE_INT16:
- do_quantized_memcpy (int16_t);
- break;
- case DATA_TYPE_UINT16:
- do_quantized_memcpy (uint16_t);
- break;
- case DATA_TYPE_INT32:
- do_quantized_memcpy (int32_t);
- break;
- case DATA_TYPE_UINT32:
- do_quantized_memcpy (uint32_t);
- break;
- case DATA_TYPE_INT64:
- do_quantized_memcpy (int64_t);
- break;
- case DATA_TYPE_UINT64:
- do_quantized_memcpy (uint64_t);
- break;
- case DATA_TYPE_FLOAT32:
- do_quantized_memcpy (float);
- break;
- case DATA_TYPE_FLOAT64:
- do_quantized_memcpy (double);
- break;
- default:
- logerr (TAG, "Unsupported datatype %d\n", type);
- }
+/** @brief host handler destructor */
+HostHandler::~HostHandler ()
+{
}
/**
- * @brief perform data manipulation
- * @param[in] model model instance
- * @param[in] idx tensor index
- * @param[in] is_input indicate it's input manipulation
- * @param[out] dst destination buffer
- * @param[in] src source buffer (feature map)
- * @param[in] size size to be copied
- * @return size of memory copy if no error, otherwise zero
- *
- * @note the input data format should be NHWC
- * @detail rules for the memory address of activations in NPU HW.
- * (https://code.sec.samsung.net/confluence/pages/viewpage.action?pageId=146491864)
- *
- * 1) Special case (depth == 3)
- * - addr(x,y,z) = addr(0,0,0) + (z) + 3 * (x + width * y)
- *
- * 2) Common case
- * - addr(x,y,z) = addr(0,0,0) + (z % MPA_L) + MPA_L * (x + width * (y + height * (z / MPA_L)))
- *
- * Thus, if depth is not a multiple of MPA_L (i.e., 64), zero padding is required
+ * @brief register model from generic buffer
+ * @param[in] model_buf model buffer
+ * @param[out] modelid model id
+ * @return 0 if no error. otherwise a negative errno
*/
-size_t
-TrinityVision::manipulateData (const Model *model, uint32_t idx, bool is_input,
- void *dst, void *src, size_t size)
+int
+HostHandler::registerModel (generic_buffer *model_buf, uint32_t *modelid)
{
- const Metadata *meta = model->getMetadata();
- const tensor_data_info* info;
- const uint32_t *dims;
- uint32_t zero_point;
- float scale;
-
- /** extract required information from the metadata */
- if (is_input) {
- if (idx >= meta->getInputNum()) {
- logerr (TAG, "Wrong information for input tensors in metadata\n");
- return 0;
- }
-
- info = model->getInputDataInfo (idx);
- dims = meta->getInputDims (idx);
- zero_point = meta->getInputQuantZero (idx);
- scale = meta->getInputQuantScale (idx);
- } else {
- if (idx >= meta->getOutputNum()) {
- logerr (TAG, "Wrong information for output tensors in metadata\n");
- return 0;
- }
-
- info = model->getOutputDataInfo (idx);
- dims = meta->getOutputDims (idx);
- zero_point = meta->getOutputQuantZero (idx);
- scale = meta->getOutputQuantScale (idx);
+ if (model_buf == nullptr || modelid == nullptr) {
+ logerr (TAG, "Invalid arguments given\n");
+ return -EINVAL;
}
- if (info == nullptr) {
- logerr (TAG, "Unmatched tensors info\n");
- return 0;
+ Model *model = nullptr;
+ int status = device_->setModel (model_buf, &model);
+ if (status != 0) {
+ logerr (TAG, "Failed to set model: %d\n", status);
+ return status;
}
- uint32_t batch = dims[0];
- uint32_t height = dims[1];
- uint32_t width = dims[2];
- uint32_t depth = dims[3];
-
- uint32_t data_size = get_data_size (info->type);
- if (data_size == 0) {
- logerr (TAG, "Invalid data size\n");
- return 0;
- }
+ assert (model != nullptr);
- bool need_quantization = false;
- /**
- * note that we assume DATA_TYPE_SRNPU is the smallest data type that we consider.
- * Also, DATA_TYPE_SRNPU and uint8_t may be regarded as the same in the view of apps.
- */
- if (info->type != DATA_TYPE_SRNPU) {
- assert (data_size >= get_data_size (DATA_TYPE_SRNPU));
-
- if (data_size > get_data_size (DATA_TYPE_SRNPU) ||
- !(zero_point == DEFAULT_ZERO_POINT && scale == DEFAULT_SCALE))
- need_quantization = true;
+ status = models_.insert (model->getID(), model);
+ if (status != 0) {
+ logerr (TAG, "Failed to insert model id\n");
+ delete model;
+ return status;
}
- /** check data manipulation is required */
- if (depth != 3 && depth != 64 && info->layout != DATA_LAYOUT_SRNPU) {
- uint32_t MPA_L = DATA_GRANULARITY;
- uint32_t n, h, w, d;
- uint32_t std_offset; /* standard offset in NHWC data format */
- uint32_t npu_offset; /* npu offset in NPU HW data format*/
- uint32_t src_offset;
- uint32_t dst_offset;
- uint32_t slice_size;
-
- /* @todo we currently support only NHWC */
- if (info->layout != DATA_LAYOUT_NHWC) {
- logerr (TAG, "data manipulation is supported for NHWC only\n");
- return -EINVAL;
- }
+ *modelid = model->getID();
+ return 0;
+}
- for (n = 0; n < batch; n++) {
- for (h = 0; h < height; h++) {
- for (w = 0; w < width; w++) {
- for (d = 0; d < depth; d += MPA_L) {
- std_offset = d + depth * (w + width * (h + n * height));
- npu_offset = MPA_L * (w + width * (h + (n + d / MPA_L) * height));
- slice_size = (depth - d >= MPA_L) ? MPA_L : depth - d;
-
- if (is_input) {
- src_offset = std_offset * data_size;
- dst_offset = npu_offset;
- } else {
- src_offset = npu_offset;
- dst_offset = std_offset * data_size;
- }
-
- /* if depth is not a multiple of MPA_L, add zero paddings (not exact values) */
- if (need_quantization) {
- memcpy_with_quant (is_input, info->type, scale, zero_point,
- static_cast<char*>(dst) + dst_offset,
- static_cast<char*>(src) + src_offset,
- slice_size);
- } else {
- memcpy (
- static_cast<char*>(dst) + dst_offset,
- static_cast<char*>(src) + src_offset,
- slice_size);
- }
- }
- }
- }
- }
- } else if (need_quantization) {
- /** depth == 3 || depth == 64; special cases which can directly copy input tensor data */
- if (is_input)
- size = size / data_size;
+/**
+ * @brief remove the registered model
+ * @param[in] modelid model id
+ * @return 0 if no error. otherwise a negative errno
+ */
+int
+HostHandler::unregisterModel (uint32_t modelid)
+{
+ Model *model = models_.find (modelid);
+ if (model == nullptr)
+ return -ENOENT;
- memcpy_with_quant (is_input, info->type, scale, zero_point,
- dst, src, size);
- } else {
- memcpy (dst, src, size);
+ int status = device_->unsetModel (model);
+ if (status != 0) {
+ logerr (TAG, "Failed to unset model: %d\n", status);
+ return status;
}
- return 0;
+ return models_.remove (modelid);
}
-#else
-
-size_t
-TrinityVision::manipulateData (const Model *model, uint32_t idx, bool is_input,
- void *dst, void *src, size_t size)
+/**
+ * @brief remove all registered models
+ * @return 0
+ */
+int
+HostHandler::unregisterModels ()
{
- memcpy (dst, src, size);
- return size;
-}
+ std::function <bool (Model *)> functor =
+ [&] (Model *m) -> bool {
+ bool can_remove = true;
+ int status = device_->unsetModel (m);
+ if (status != 0) {
+ logwarn (TAG, "Failed to unset model: %d\n", status);
+ can_remove = false;
+ }
+ return can_remove;
+ };
-#endif
+ models_.for_each (functor);
+ return 0;
+}
/**
- * @brief create device instance depending on device type and id
- * @param[in] type device type
- * @param[in] id device id
- * @return device instance
+ * @brief Get the profile information from NPU
+ * @param[in] task_id The identifier for each inference
+ * @param[out] profile The profile instance
+ * @return 0 if no error, otherwise a negative errno.
*/
-Device *
-Device::createInstance (dev_type type, int id)
+int
+HostHandler::getProfile (int task_id, npu_profile *profile)
{
- Device *device = nullptr;
-
- switch (type & DEVICETYPE_MASK) {
- case DEVICETYPE_TRIV:
- device = new TrinityVision (id);
- break;
- case DEVICETYPE_TRIV2:
- device = new TrinityVision2 (id);
- break;
- case DEVICETYPE_TRIA:
- device = new TrinityAsr (id);
- break;
- default:
- break;
+ if (task_id < 0 || profile == nullptr) {
+ logerr (TAG, "Invalid parameter provided\n");
+ return -EINVAL;
}
- if (device != nullptr && device->init () != 0) {
- delete device;
- device = nullptr;
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+
+ profile->num_layers = 0;
+ profile->layers = nullptr;
+
+ int status = api->getProfile (task_id, profile);
+ if (status != 0) {
+ logerr (TAG, "Failed to get profile information: %d\n", status);
+ return status;
}
- return device;
+ return 0;
}
-/** @brief host handler constructor */
-HostHandler::HostHandler (Device *device)
- : device_(device)
+/**
+ * @brief get the stats for the latest apps of the target device
+ * @param[out] stat The list of app stat
+ * @note The caller has the responsibility to free the resources.
+ * This API is not working on the emulated envionment.
+ */
+int
+HostHandler::getStatApps (npu_stat_apps *stat)
{
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+
+ return api->getStatApps (stat);
}
-/** @brief host handler destructor */
-HostHandler::~HostHandler ()
+/**
+ * @brief get the stats for the latest tasks of the target app
+ * @param[in] appid The identifier of target app
+ * @param[out] stat The list of task stat
+ * @note The caller has the responsibility to free the resources.
+ * This API is not working on the emulated envionment.
+ */
+int
+HostHandler::getStatTasks (int appid, npu_stat_tasks *stat)
{
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+
+ return api->getStatTasks (appid, stat);
}
/**
- * @brief register model from generic buffer
- * @param[in] model_buf model buffer
- * @param[out] modelid model id
- * @return 0 if no error. otherwise a negative errno
+ * @brief Get the driver API level of opened NPU device
+ * @param[out] level driver API level
+ * @return 0 if no error, otherwise a negative errno
*/
int
-HostHandler::registerModel (generic_buffer *model_buf, uint32_t *modelid)
+HostHandler::getAPILevel (uint32_t *level)
{
- Model *model = device_->registerModel (model_buf);
- if (model == nullptr) {
- logerr (TAG, "Failed to register model\n");
- return -EINVAL;
- }
-
- int status = models_.insert (model->getID(), model);
- if (status != 0) {
- logerr (TAG, "Failed to insert model id\n");
- delete model;
- return status;
- }
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
- *modelid = model->getID();
- return 0;
+ return api->getAPILevel (level);
}
/**
- * @brief remove the registered model
- * @param[in] modelid model id
- * @return 0 if no error. otherwise a negative errno
+ * @brief Get the TOPS of the opened NPU device
+ * @param[in] dev the NPU device handle
+ * @param[out] tops npu tops
+ * @return 0 if no error, otherwise a negative errno
+ * @note this does not support for emulated devices
*/
int
-HostHandler::unregisterModel (uint32_t modelid)
+HostHandler::getTops (uint32_t *tops)
{
- return models_.remove (modelid);
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+
+ return api->getTops (tops);
}
/**
- * @brief remove all registered models
- * @return 0
+ * @brief Get the DSP DSPM size of the opened NPU device
+ * @param[in] dev the NPU device handle
+ * @param[out] dspm dspm size
+ * @return 0 if no error, otherwise a negative errno
+ * @note this does not support for emulated devices
*/
int
-HostHandler::unregisterModels ()
+HostHandler::getDspmSize (uint32_t *dspm)
{
- models_.clear ();
- return 0;
-}
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+ return api->getDspmSize (dspm);
+}
/**
* @brief Set the data layout for input/output tensors
* @param[in] modelid The ID of model whose layouts are set
if (model == nullptr)
return -ENOENT;
- model->setDataInfo (in, out);
-
- return 0;
+ return model->setDataInfo (in, out);
}
/**
}
void callback (output_buffers *output, uint64_t sequence) {
- /** just copy internal variables of output buffers */
- memcpy (output_, output, sizeof (output_buffers));
+ if (output_ != nullptr && output != nullptr) {
+ /** just copy internal variables of output buffers */
+ memcpy (output_, output, sizeof (output_buffers));
+ }
done_ = true;
cv_.notify_one ();
}
* @param[in] modelid The model to be inferred.
* @param[in] input The input data to be inferred.
* @param[out] output The output result.
- * @return @c 0 if no error. otherwise a negative error value
+ * @return @c positive id if no error. otherwise a negative error value
*/
int
HostHandler::runSync (uint32_t modelid, const input_buffers *input,
callbackSync sync (output);
int status = runAsync (modelid, input, callbackSync::callback,
static_cast <void*> (&sync), NPUASYNC_DROP_OLD, nullptr);
- if (status == 0) {
+ if (status > 0) {
/** sync needs to wait callback */
sync.wait ();
}
* @param[in] cb_data The data given as a parameter to the runNPU_async call.
* @param[in] mode Configures how this operation works.
* @param[out] sequence The sequence number returned with runNPU_async.
- * @return @c 0 if no error. otherwise a negative error value
+ * @return @c positive id if no error. otherwise a negative error value
*/
int
HostHandler::runAsync (uint32_t modelid, const input_buffers *input,
return -ENOENT;
}
+ /* check the given model before running */
+ if (model != nullptr && !model->finalize ()) {
+ logerr (TAG, "Failed to finalize the model. Please see the log messages\n");
+ return -EINVAL;
+ }
+
device_->setAsyncMode (mode);
return device_->run (NPUINPUT_HOST, model, input, cb, cb_data, sequence);
}
/**
+ * @brief Let NPU accept input frames from its internal source continuously
+ * @param[in] modelid The model to be inferred.
+ * @param[in] opmode NPU has different opmode with auto-inputs. Choose one.
+ * @param[in] hw_dev The target device feeding input data
+ * @return @c 0 if no error. otherwise a negative error value
+ */
+int
+HostHandler::runInternal (uint32_t modelid, npu_input_opmode opmode,
+ std::string hw_dev)
+{
+ Model *model = nullptr;
+
+ if (device_->needModel()) {
+ model = getModel (modelid);
+ if (model == nullptr)
+ return -ENOENT;
+ }
+
+ /* check the given model before running */
+ if (model != nullptr && !model->finalize ()) {
+ logerr (TAG, "Failed to finalize the model. Please see the log messages\n");
+ return -EINVAL;
+ }
+
+ return device_->runInternal (opmode, model, hw_dev);
+}
+
+/**
+ * @brief Stop the request with the given id
+ * @param[in] dev The NPU device handle
+ * @param[in] id The request id
+ * @return @c 0 if no error. otherwise a negative error value
+ */
+int
+HostHandler::stopInternal (int id)
+{
+ if (id <= 0) {
+ logerr (TAG, "Unable to stop this request with id (%d)\n", id);
+ return -EINVAL;
+ }
+
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
+
+ return api->stop_target (id);
+}
+
+/**
* @brief get number of available devices
* @param[in] type device type
* @return number of devices
}
*dev = device;
- /** This is just for backward-compatility; we don't guarantee its corresness */
- latest_dev_ = *dev;
return 0;
}
if (buffer == NULL)
return -EINVAL;
+ if (buffer->size == 0) {
+ logerr (TAG, "Invalid size\n");
+ return -EINVAL;
+ }
+
if (buffer->size > UINT32_MAX) {
logerr (TAG, "Don't support such a large size");
return -ENOMEM;
}
- if (buffer->type == BUFFER_FILE) {
- /* nothing to do */
- if (buffer->filepath == nullptr)
+ switch (buffer->type) {
+ case BUFFER_FILE:
+ /* nothing to do */
+ if (buffer->filepath == nullptr)
+ return -EINVAL;
+ break;
+ case BUFFER_MAPPED:
+ {
+ /* now, npu-engine always provides dmabuf-based allocation */
+ void *addr = nullptr;
+ int dmabuf = device_->allocMemory (buffer->size, &addr);
+ if (dmabuf < 0)
+ return dmabuf;
+
+ buffer->dmabuf = dmabuf;
+ buffer->offset = 0;
+ buffer->addr = addr;
+ } break;
+ default:
return -EINVAL;
- } else {
- /* now, npu-engine always provides dmabuf-based allocation */
- HWmem *hwmem = device_->allocMemory ();
- if (hwmem == nullptr || hwmem->alloc (buffer->size) < 0)
- return -ENOMEM;
-
- buffer->dmabuf = hwmem->getDmabuf();
- buffer->offset = hwmem->getOffset();
- buffer->addr = hwmem->getData();
}
+
return 0;
}
if (buffer == NULL)
return -EINVAL;
- if (buffer->type != BUFFER_FILE)
- device_->deallocMemory (buffer->dmabuf);
-
- return 0;
-}
-
-/**
- * @brief allocate multiple generic buffers (just for users)
- * @param[out] buffers multi-buffer instance
- * @return 0 if no error. otherwise a negative errno
- */
-int
-HostHandler::allocGenericBuffer (generic_buffers *buffers)
-{
- if (buffers == NULL || buffers->num_buffers < 1)
- return -EINVAL;
-
- buffer_types type = buffers->bufs[0].type;
- if (type == BUFFER_FILE)
- return 0;
-
- uint64_t total_size = 0;
- for (uint32_t idx = 0; idx < buffers->num_buffers; idx++)
- total_size += buffers->bufs[idx].size;
-
- uint64_t first_size = buffers->bufs[0].size;
- buffers->bufs[0].size = total_size;
- int status = allocGenericBuffer (&buffers->bufs[0]);
- if (status != 0)
- return status;
-
- uint64_t offset = first_size;
- for (uint32_t idx = 1; idx < buffers->num_buffers; idx++) {
- buffers->bufs[idx].dmabuf = buffers->bufs[0].dmabuf;
- buffers->bufs[idx].offset = buffers->bufs[0].offset + offset;
- buffers->bufs[idx].type = type;
-
- offset += buffers->bufs[idx].size;
+ switch (buffer->type) {
+ case BUFFER_FILE:
+ /** always true cuz nothing to do */
+ break;
+ case BUFFER_MAPPED:
+ return device_->deallocMemory (buffer->dmabuf, buffer->size, buffer->addr);
+ default:
+ return -EINVAL;
}
return 0;
}
/**
- * @brief deallocate multiple generic buffers (just for users)
- * @param[in] buffers multi-buffer instance
- * @return 0 if no error. otherwise a negative errno
- */
-int
-HostHandler::deallocGenericBuffer (generic_buffers *buffers)
-{
- if (buffers == NULL || buffers->num_buffers < 1)
- return -EINVAL;
-
- return deallocGenericBuffer (&buffers->bufs[0]);
-}
-
-/** just for backward-compatability */
-npudev_h HostHandler::latest_dev_ = nullptr;
-
-/** implementation of libnpuhost APIs */
-
-/**
- * @brief Returns the number of available NPU devices.
- * @return @c The number of NPU devices.
- * @retval 0 if no NPU devices available. if positive (number of NPUs) if NPU devices available. otherwise, a negative error value.
- * @note the caller should call putNPUdevice() to release the device handle
- */
-int getnumNPUdeviceByType (dev_type type)
-{
- return HostHandler::getNumDevices (type);
-}
-
-/**
- * @brief Returns the handle of the chosen NPU devices.
- * @param[out] dev The NPU device handle
- * @param[in] id The NPU id to get the handle. 0 <= id < getnumNPUdeviceByType().
- * @return @c 0 if no error. otherwise a negative error value
- * @note the caller should call putNPUdevice() to release the device handle
- */
-int getNPUdeviceByType (npudev_h *dev, dev_type type, uint32_t id)
-{
- return HostHandler::getDevice (dev, type, id);
-}
-
-/**
- * @brief Returns the handle of an NPU device meeting the condition
- * @param[out] dev The NPU device handle
- * @param[in] cond The condition for device search.
- * @return @c 0 if no error. otherwise a negative error value
- * @note the caller should call putNPUdevice() to release the device handle
- * @note it's not supported yet
- */
-int getNPUdeviceByCondition(npudev_h *dev, const npucondition *cond)
-{
- /** not implmeneted yet */
- return getNPUdeviceByType (dev, NPUCOND_TRIV_CONN_SOCIP, 0);
-}
-
-/**
- * @brief release the NPU device instance obtained by getDevice ()
- * @param[in] dev the NPU device handle
- */
-void putNPUdevice (npudev_h dev)
-{
- if (dev != nullptr)
- delete static_cast<Device *> (dev);
-}
-
-/**
- * @brief Send the NN model to NPU.
- * @param[in] dev The NPU device handle
- * @param[in] modelfile The filepath to the compiled NPU NN model in any buffer_type
- * @param[out] modelid The modelid allocated for this instance of NN model.
- * @return @c 0 if no error. otherwise a negative error value
- *
- * @detail For ASR devices, which do not accept models, but have models
- * embedded in devices, you do not need to call register and
- * register calls for ASR are ignored.
- *
- * @todo Add a variation: in-memory model register.
- */
-int registerNPUmodel (npudev_h dev, generic_buffer *modelfile, uint32_t *modelid)
-{
- INIT_HOST_HANDLER (host_handler, dev);
-
- return host_handler->registerModel (modelfile, modelid);
-}
-
-/**
- * @brief Remove the NN model from NPU
- * @param[in] dev The NPU device handle
- * @param[in] modelid The model to be removed from the NPU.
- * @return @c 0 if no error. otherwise a negative error value
- * @detail This may incur some latency with memory compatcion.
+ * @brief allocate multiple generic buffers (just for users)
+ * @param[out] buffers multi-buffer instance
+ * @return 0 if no error. otherwise a negative errno
*/
-int unregisterNPUmodel(npudev_h dev, uint32_t modelid)
+int
+HostHandler::allocGenericBuffer (generic_buffers *buffers)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ uint32_t idx;
+ int status = 0;
- return host_handler->unregisterModel (modelid);
-}
+ if (buffers == NULL || buffers->num_buffers < 1)
+ return -EINVAL;
-/**
- * @brief Remove all NN models from NPU
- * @param[in] dev The NPU device handle
- * @return @c 0 if no error. otherwise a negative error value
- */
-int unregisterNPUmodel_all(npudev_h dev)
-{
- INIT_HOST_HANDLER (host_handler, dev);
+ for (idx = 0; idx < buffers->num_buffers; idx++) {
+ status = allocGenericBuffer (&buffers->bufs[idx]);
+ if (status != 0)
+ goto free_buffer;
+ }
+
+ return 0;
+
+free_buffer:
+ while (idx != 0) {
+ deallocGenericBuffer (&buffers->bufs[--idx]);
+ }
- return host_handler->unregisterModels ();
+ return status;
}
/**
- * @brief [OPTIONAL] Set the data layout for input/output tensors
- * @param[in] dev The NPU device handle
- * @param[in] modelid The ID of model whose layouts are set
- * @param[in] info_in the layout/type info for input tensors
- * @param[in] info_out the layout/type info for output tensors
- * @return @c 0 if no error. otherwise a negative error value
- * @note if this function is not called, default layout/type will be used.
+ * @brief deallocate multiple generic buffers (just for users)
+ * @param[in] buffers multi-buffer instance
+ * @return 0 if no error. otherwise a negative errno
*/
-int setNPU_dataInfo(npudev_h dev, uint32_t modelid,
- tensors_data_info *info_in, tensors_data_info *info_out)
+int
+HostHandler::deallocGenericBuffer (generic_buffers *buffers)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ if (buffers == NULL || buffers->num_buffers < 1)
+ return -EINVAL;
+
+ for (uint32_t idx = 0; idx < buffers->num_buffers; idx++)
+ deallocGenericBuffer (&buffers->bufs[idx]);
+ buffers->num_buffers = 0;
- return host_handler->setDataInfo (modelid, info_in, info_out);
+ return 0;
}
/**
- * @brief [OPTIONAL] Set the inference constraint for next NPU inferences
- * @param[in] dev The NPU device handle
- * @param[in] modelid The target model id
- * @param[in] constraint inference constraint (e.g., timeout, priority)
- * @return @c 0 if no error. otherwise a negative error value
- * @note If this function is not called, default values are used.
+ * @brief get the current memory status
+ * @param[out] alloc_total The size of allocated memory until now
+ * @param[out] free_total The size of freed memory until now
+ * @return 0 if no error. otherwise a negatice error value
*/
-int setNPU_constraint(npudev_h dev, uint32_t modelid, npuConstraint constraint)
+int
+HostHandler::getMemoryStatus (size_t *alloc_total, size_t *free_total)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ /** API is always set in initialize () */
+ const DriverAPI * api = device_->getDriverAPI ();
+ assert (api != nullptr);
- return host_handler->setConstraint (modelid, constraint);
+ return api->getMemoryStatus (alloc_total, free_total);
}
/**
- * @brief Execute inference. Wait (block) until the output is available.
- * @param[in] dev The NPU device handle
- * @param[in] modelid The model to be inferred.
- * @param[in] input The input data to be inferred.
- * @param[out] output The output result. The caller MUST allocate appropriately before calling this.
- * @return @c 0 if no error. otherwise a negative error value
- *
- * @detail This is a syntactic sugar of runNPU_async().
- * CAUTION: There is a memcpy for the output buffer.
+ * @brief Get the current device status to be used
+ * @param[out] status the device status
+ * @param[out] num_requests the number of running requests (or pending)
+ * @return 0 if no error, otherwise a negative errno.
*/
-int runNPU_sync(npudev_h dev, uint32_t modelid, const input_buffers *input,
- output_buffers *output)
+int
+HostHandler::getDeviceStatus (npu_status *status, uint32_t *num_requests)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ /** API is always set in initialize () */
+ const DriverAPI * api = device_->getDriverAPI ();
+
+ if (!api)
+ return -EINVAL;
+
+ device_state_t state = api->isReady ();
+ if (state == device_state_t::STATE_READY) {
+ *num_requests = api->numRequests ();
+ if (*num_requests > 0)
+ *status = NPU_READY;
+ else
+ *status = NPU_IDLE;
+ } else {
+ *num_requests = 0;
+ *status = NPU_ERROR;
+ }
- return host_handler->runSync (modelid, input, output);
+ return 0;
}
-/**
- * @brief Invoke NPU inference. Unblocking call.
- * @param[in] dev The NPU device handle
- * @param[in] modelid The model to be inferred.
- * @param[in] input The input data to be inferred.
- * @param[in] cb The output buffer handler.
- * @param[out] sequence The sequence number returned with runNPU_async.
- * @param[in] data The data given as a parameter to the runNPU_async call.
- * @param[in] mode Configures how this operation works.
- * @return @c 0 if no error. otherwise a negative error value
- */
-int runNPU_async(npudev_h dev, uint32_t modelid, const input_buffers *input,
- npuOutputNotify cb, uint64_t *sequence, void *data,
- npu_async_mode mode)
-{
- INIT_HOST_HANDLER (host_handler, dev);
+/** implement methods of Device class */
- return host_handler->runAsync (modelid, input, cb, data, mode, sequence);
+/** @brief constructor of device */
+Device::Device (dev_type type, int id, bool need_model)
+ : comm_ (CommPlugin::getCommPlugin()), type_ (type), id_ (id), need_model_ (true),
+ mode_ (NPUASYNC_WAIT), initialized_ (false), atomic_flag_ (ATOMIC_FLAG_INIT)
+{
}
/**
- * @brief Allocate a buffer for NPU model with the requested buffer type.
- * @param[in] dev The NPU device handle
- * @param[in/out] Buffer the buffer pointer where memory is allocated.
- * @return 0 if no error, otherwise a negative errno.
+ * @brief create device instance depending on device type and id
+ * @param[in] type device type
+ * @param[in] id device id
+ * @return device instance
*/
-int allocNPU_modelBuffer (npudev_h dev, generic_buffer *buffer)
+Device *
+Device::createInstance (dev_type type, int id)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ Device *device = nullptr;
+
+ switch (type & DEVICETYPE_MASK) {
+ case DEVICETYPE_TRIV2:
+ device = new TrinityVision2 (id);
+ break;
+ case DEVICETYPE_DEPR:
+ logwarn (TAG, "You're trying to open deprecated devices..\n");
+ break;
+ default:
+ break;
+ }
+
+ if (device != nullptr && device->init () != 0) {
+ delete device;
+ device = nullptr;
+ }
- return host_handler->allocGenericBuffer (buffer);
+ return device;
}
/**
- * @brief Free the buffer and remove the address mapping.
- * @param[in] dev The NPU device handle
- * @param[in] buffer the model buffer
- * @return 0 if no error, otherwise a negative errno.
+ * @brief device initialization
+ * @return 0 if no error, otherwise a negative errno
+ * @note Init failures come from createDriverAPI() only.
*/
-int cleanNPU_modelBuffer (npudev_h dev, generic_buffer *buffer)
+int
+Device::init ()
{
- INIT_HOST_HANDLER (host_handler, dev);
+ /** should be initilizaed only once */
+ if (!atomic_flag_.test_and_set()) {
+ /** create the corresponding driver API */
+ api_ = DriverAPI::createDriverAPI (type_, id_);
+ if (api_.get() == nullptr) {
+ atomic_flag_.clear();
+ logerr (TAG, "Failed to create driver API\n");
+ return -EINVAL;
+ }
+
+ handler_.reset (new HostHandler (this));
+ scheduler_.reset (new Scheduler (api_.get()));
+ mem_ = MemAllocator::createInstance (api_.get());
+
+ initialized_ = true; /** c++11 does not provide test() of atomic flag */
+ }
- return host_handler->deallocGenericBuffer (buffer);
+ return 0;
}
/**
- * @brief Allocate a buffer for NPU input with the requested buffer type.
- * @param[in] dev The NPU device handle
- * @param[in/out] Buffer the buffer pointer where memory is allocated.
- * @return 0 if no error, otherwise a negative errno.
- * @note please utilize allocInputBuffers() for multiple input tensors because subsequent
- * calls of allocInputBuffer() don't gurantee contiguous allocations between them.
+ * @brief stop all requests from this device
+ * @param[in] force_stop indicate the schedduler waits until to handle previous requests
+ * @return 0 if no error, otherwise a negative errno
*/
-int allocNPU_inputBuffer (npudev_h dev, generic_buffer *buffer)
+int
+Device::stop (bool force_stop)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
- return host_handler->allocGenericBuffer (buffer);
+ Request *req = new Request (NPUINPUT_STOP);
+ req->setForceStop (force_stop);
+ return scheduler_->submitRequest (req);
}
/**
- * @brief Free the buffer and remove the address mapping.
- * @param[in] dev The NPU device handle
- * @param[in] buffer the input buffer
- * @return 0 if no error, otherwise a negative errno.
+ * @brief allocate generic memory buffer
+ * @param[in] size the size to allocate
+ * @param[out] addr the mapped address
+ * @return dmabuf fd if no error, otherwise a negative errno
*/
-int cleanNPU_inputBuffer (npudev_h dev, generic_buffer *buffer)
+int
+Device::allocMemory (size_t size, void **addr)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
+
+ if (size == 0 || addr == nullptr) {
+ logerr (TAG, "Invalid arguments\n");
+ return -EINVAL;
+ }
- return host_handler->deallocGenericBuffer (buffer);
+ return mem_->allocMemory (size, addr);
}
/**
- * @brief Allocate input buffers, which have multiple instances of generic_buffer
- * @param[in] dev The NPU device handle
- * @param[in/out] input input buffers.
- * @return 0 if no error, otherwise a negative errno.
- * @note it reuses allocInputBuffer().
- * @details in case of BUFFER_DMABUF, this function can be used to gurantee physically-contiguous
- * memory mapping for multiple tensors (in a single inference, not batch size).
+ * @brief deallocate generic memory buffer
+ * @param[in] dmabuf_fd dmabuf file descriptor
+ * @param[in] size buffer size
+ * @param[in] addr mapped addr
+ * @return 0 if no error, otherwise a negative errno
*/
-int allocNPU_inputBuffers (npudev_h dev, input_buffers * input)
+int
+Device::deallocMemory (int dmabuf_fd, size_t size, void * addr)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
+
+ if (dmabuf_fd < 0 || size == 0 || addr == nullptr) {
+ logerr (TAG, "Invalid arguments\n");
+ return -EINVAL;
+ }
- return host_handler->allocGenericBuffer (input);
+ return mem_->deallocMemory (dmabuf_fd, size, addr);
}
/**
- * @brief Free input buffers allocated by allocInputBuffers().
- * @param[in] dev The NPU device handle
- * @param[in/out] input input buffers.
- * @note it reuses cleanInputbuffer().
- * @return 0 if no error, otherwise a negative errno.
+ * @brief extract the segment table instance from input generic buffers
+ * @param[in] model the model instance
+ * @param[in] input the input generic buffers
+ * @param[in] output the output generic buffers
+ * @return the segment table instance
*/
-int cleanNPU_inputBuffers (npudev_h dev, input_buffers * input)
+SegmentTable *
+TrinityVision2::prepareSegmentTable (const Model *model, const input_buffers *input,
+ const output_buffers *output)
{
- INIT_HOST_HANDLER (host_handler, dev);
+ const Metadata *meta = model->getMetadata ();
+ if (meta == nullptr || (input != nullptr &&
+ meta->getInputNum() != input->num_buffers)) {
+ logerr (TAG, "Invalid metadata info provided\n");
+ return nullptr;
+ }
+
+ SegmentTable * segt = mem_->allocSegmentTable (new HWmemDevice);
+ int status = segt->alloc ();
+ if (status != 0) {
+ logerr (TAG, "Failed to allocate segment table: %d\n", status);
+ goto delete_segt;
+ }
+
+ status = segt->createSegments (model, input, output);
+ if (status != 0) {
+ logerr (TAG, "Failed to create segments: %d\n", status);
+ goto delete_segt;
+ }
- return host_handler->deallocGenericBuffer (input);
+ return segt;
+
+delete_segt:
+ delete segt;
+ return nullptr;
}
/**
- * @brief Get metadata for NPU model
- * @param[in] model The path of model binary file
- * @param[in] need_extra whether you want to extract the extra data in metadata
- * @return the metadata structure to be filled if no error, otherwise nullptr
- *
- * @note For most npu-engine users, the extra data is not useful because it will be
- * used for second-party users (e.g., compiler, simulator).
- * Also, the caller needs to free the metadata.
- *
- * @note the caller needs to free the metadata
+ * @brief implementation of TRIV2's setModel ()
+ * @param[in] model_buf the model generic buffer
+ * @param[out] model the model instance
+ * @return 0 if no error, otherwise a negative errno
*/
-npubin_meta * getNPUmodel_metadata (const char *model, bool need_extra)
+int
+TrinityVision2::setModel (const generic_buffer *model_buf, Model ** model_ptr)
{
- npubin_meta *meta;
- FILE *fp;
- size_t ret;
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
- if (!model)
- return nullptr;
+ if (model_buf == nullptr || model_ptr == nullptr)
+ return -EINVAL;
- fp = fopen (model, "rb");
- if (!fp) {
- logerr (TAG, "Failed to open the model binary: %d\n", -errno);
- return nullptr;
- }
+ Model *model;
+ int status;
- meta = (npubin_meta *) malloc (NPUBIN_META_SIZE);
- if (!meta) {
- logerr (TAG, "Failed to allocate metadata\n");
- goto exit_err;
- }
+ switch (model_buf->type) {
+ case BUFFER_FILE:
+ case BUFFER_MAPPED:
+ model = mem_->allocModel (new HWmemDevice);
+ if (model == nullptr) {
+ logerr (TAG, "Failed to allocate model\n");
+ return -ENOMEM;
+ }
- ret = fread (meta, 1, NPUBIN_META_SIZE, fp);
- if (ret != NPUBIN_META_SIZE) {
- logerr (TAG, "Failed to read the metadata\n");
- goto exit_free;
- }
+ status = model->alloc (NPUBIN_META_SIZE);
+ if (status != 0) {
+ logerr (TAG, "Failed to allocate model: %d\n", status);
+ goto delete_exit;
+ }
- if (!CHECK_NPUBIN (meta->magiccode)) {
- logerr (TAG, "Invalid metadata provided\n");
- goto exit_free;
+ status = comm_.extractGenericBuffer (model_buf, model->getData(), nullptr,
+ 0, NPUBIN_META_SIZE);
+ if (status != 0) {
+ logerr (TAG, "Failed to extract generic buffer: %d\n", status);
+ goto delete_exit;
+ }
+ break;
+ default:
+ return -EINVAL;
}
- if (need_extra && NPUBIN_META_EXTRA (meta->magiccode) > 0) {
- npubin_meta *new_meta;
+ status = model->setMetadata (model->getData());
+ if (status != 0)
+ goto delete_exit;
+
+ /** allocate program (optional; NOP) */
+ if (model->getMetadata()->getProgramSize() > 0) {
+ HWmem * hwmem_prog = new HWmem (new HWmemDevice);
+ hwmem_prog->setDriverAPI (api_.get());
+ hwmem_prog->setContiguous (true);
+
+ model->setProgramData (hwmem_prog);
+
+ status = hwmem_prog->alloc (model->getMetadata()->getProgramSize());
+ if (status != 0) {
+ logerr (TAG, "Failed to allocate program\n");
+ goto delete_exit;
+ }
- new_meta = (npubin_meta *) realloc (meta, NPUBIN_META_TOTAL_SIZE(meta->magiccode));
- if (!new_meta) {
- logerr (TAG, "Failed to allocate extra metadata\n");
- goto exit_free;
+ status = comm_.extractGenericBuffer (model_buf, hwmem_prog->getData(), nullptr,
+ model->getMetadata()->getMetaSize(),
+ model->getMetadata()->getProgramSize());
+ if (status != 0) {
+ logerr (TAG, "Failed to extract generic buffer: %d\n", status);
+ goto delete_exit;
}
- ret = fread (new_meta->reserved_extra, 1, NPUBIN_META_EXTRA_SIZE (meta->magiccode), fp);
- if (ret != NPUBIN_META_EXTRA_SIZE (meta->magiccode)) {
- logerr (TAG, "Invalid extra metadata provided\n");
- free (new_meta);
- goto exit_err;
+ /** register this model to the driver */
+ model_config_t config;
+ config.version = model->getMetadata()->getVersion ();
+ config.dbuf_fd = hwmem_prog->getDmabuf ();
+ config.program_size = hwmem_prog->getSize ();
+ config.program_offset_addr = 0;
+ config.metadata_dbuf_fd = model->getDmabuf ();
+
+ /** for metadata extended section */
+ size_t extended_size = model->getMetadata()->getMetaExtendedSize();
+ if (extended_size > 0) {
+ HWmem * hwmem_extended = new HWmem (new HWmemDevice);
+ hwmem_extended->setDriverAPI (api_.get ());
+
+ model->setExtendedMetadata (hwmem_extended);
+
+ status = hwmem_extended->alloc (extended_size);
+ if (status != 0) {
+ logerr (TAG, "Failed to allocate extended metadata: %d\n", status);
+ goto delete_exit;
+ }
+
+ config.metadata_ext_dbuf_fd = hwmem_extended->getDmabuf ();
+ config.metadata_ext_size = extended_size;
+
+ status = comm_.extractGenericBuffer (model_buf, hwmem_extended->getData (),
+ nullptr, NPUBIN_META_SIZE, extended_size);
+ if (status != 0) {
+ logerr (TAG, "Failed to extract generic buffer: %d\n", status);
+ goto delete_exit;
+ }
+ } else {
+ config.metadata_ext_dbuf_fd = -1;
+ config.metadata_ext_size = 0;
}
- meta = new_meta;
+ status = api_->registerModel (&config, model->getMetadata()->getNPUVersion());
+ if (status != 0)
+ goto delete_exit;
+
+ model->setInternalID(config.id);
}
- fclose (fp);
+ /** allocate weight (optional) */
+ if (model->getMetadata()->getWeightSize() > 0) {
+ HWmem * hwmem_weight = new HWmem (new HWmemDevice);
+ hwmem_weight->setDriverAPI (api_.get());
- return meta;
+ model->setWeightData (hwmem_weight);
-exit_free:
- free (meta);
-exit_err:
- fclose (fp);
+ status = hwmem_weight->alloc (model->getMetadata()->getWeightSize());
+ if (status != 0) {
+ logerr (TAG, "Failed to allocate program\n");
+ goto delete_exit;
+ }
- return nullptr;
-}
+ status = comm_.extractGenericBuffer (model_buf, hwmem_weight->getData(), nullptr,
+ model->getMetadata()->getMetaSize() + model->getMetadata()->getProgramSize(),
+ model->getMetadata()->getWeightSize());
+ if (status != 0) {
+ logerr (TAG, "Failed to extract generic buffer: %d\n", status);
+ goto delete_exit;
+ }
+ }
-/** deprecated buffer APIs; please use the above APIs */
+ *model_ptr = model;
+ return status;
-/**
- * @brief Returns the number of NPU devices (TRIV).
- */
-int getnumNPUdevice (void)
-{
- logwarn (TAG, "deprecated. Please use getnumNPUdeviceByType ()\n");
- return getnumNPUdeviceByType (NPUCOND_TRIV_CONN_SOCIP);
+delete_exit:
+ delete model;
+ return status;
}
/**
- * @brief Returns the list of ASR devices (TRIA)
+ * @brief implementation of TRIV2's unsetModel ()
+ * @param[in] model the model instance
+ * @return 0 if no error, otherwise a negative errno
*/
-int getnumASRdevice (void)
+int
+TrinityVision2::unsetModel (Model * model)
{
- logwarn (TAG, "deprecated. Please use getnumNPUdeviceByType ()\n");
- return getnumNPUdeviceByType (NPUCOND_TRIA_CONN_SOCIP);
-}
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
-/**
- * @brief Returns the handle of the chosen TRIV device.
- */
-int getNPUdevice (npudev_h *dev, uint32_t id)
-{
- logwarn (TAG, "deprecated. Please use getNPUdeviceByType ()\n");
- return getNPUdeviceByType (dev, NPUCOND_TRIV_CONN_SOCIP, id);
-}
+ if (model == nullptr) {
+ logerr (TAG, "Invalid model instance\n");
+ return -EINVAL;
+ }
-/**
- * @brief Returns the handle of the chosen TRIA device.
- */
-int getASRdevice (npudev_h *dev, uint32_t id)
-{
- logwarn (TAG, "deprecated. Please use getNPUdeviceByType ()\n");
- return getNPUdeviceByType (dev, NPUCOND_TRIA_CONN_SOCIP, id);
-}
+ if (model->getMetadata()->getProgramSize() > 0)
+ return api_->deregisterModel (model->getInternalID ());
-/** @brief deprecated */
-int allocModelBuffer (generic_buffer *buffer)
-{
- logwarn (TAG, "deprecated. Please use allocNPU_modelBuffer\n");
- return allocNPU_modelBuffer (HostHandler::getLatestDevice(), buffer);
+ return 0;
}
-/** @brief deprecated */
-int cleanModelBuffer (generic_buffer *buffer)
+/** @brief implementation of TRIV2's run() */
+int
+TrinityVision2::run (npu_input_opmode opmode, const Model *model,
+ const input_buffers *input, npuOutputNotify cb, void *cb_data,
+ uint64_t *sequence)
{
- logwarn (TAG, "deprecated. Please use cleanNPU_modelBuffer\n");
- return allocNPU_modelBuffer (HostHandler::getLatestDevice(), buffer);
-}
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
-/** @brief deprecated */
-int allocInputBuffer (generic_buffer *buffer)
-{
- logwarn (TAG, "deprecated. Please use allocNPU_inputBuffer\n");
- return allocNPU_inputBuffer (HostHandler::getLatestDevice(), buffer);
-}
+ if (opmode != NPUINPUT_HOST)
+ return -EINVAL;
-/** @brief deprecated */
-int cleanInputBuffer (generic_buffer *buffer)
-{
- logwarn (TAG, "deprecated. Please use cleanNPU_inputBuffer\n");
- return cleanNPU_inputBuffer (HostHandler::getLatestDevice(), buffer);
+ if (input == nullptr || input->num_buffers == 0 || model == nullptr)
+ return -EINVAL;
+
+ const_cast<Model *>(model)->updateDataInfo ();
+
+ /** this device uses segment table */
+ SegmentTable * segt = prepareSegmentTable (model, input);
+ if (segt == nullptr) {
+ logerr (TAG, "Failed to create segment table instance\n");
+ return -EINVAL;
+ }
+
+ /** extract input data */
+ for (uint32_t idx = 0; idx < input->num_buffers; idx++) {
+ if (!segt->getInputSegment(idx)->isExternal ()) {
+ uint32_t seg_offset = segt->getInputSegmentOffset(idx);
+ auto func = std::bind (TrinityVision2::manipulateData, model, idx, true,
+ std::placeholders::_1, std::placeholders::_2, std::placeholders::_3);
+ int status = comm_.extractGenericBuffer (
+ &input->bufs[idx],
+ segt->getInputSegment(idx)->getData() + seg_offset,
+ func);
+ if (status != 0) {
+ logerr (TAG, "Failed to feed input segment: %d\n", status);
+ return status;
+ }
+ }
+ }
+
+ Request *req = new Request (opmode);
+ req->setModel (model);
+ req->setInferData (segt);
+ req->setCallback (std::bind (&TrinityVision2::callback, this, req, cb, cb_data));
+
+ if (sequence && req->getID () > 0) {
+ *sequence = (uint32_t) req->getID ();
+ }
+
+ return scheduler_->submitRequest (req);
}
-/** @brief deprecated */
-int allocInputBuffers (input_buffers * input)
+/** @brief implementation of TRIV2's runInternal() */
+int
+TrinityVision2::runInternal (npu_input_opmode opmode, const Model *model,
+ std::string hw_dev)
{
- logwarn (TAG, "deprecated. Please use allocNPU_inputBuffers\n");
- return allocNPU_inputBuffers (HostHandler::getLatestDevice(), input);
+ if (!initialized ()) {
+ logerr (TAG, "Uninitialized device; should use libnpuhost APIs\n");
+ return -EPERM;
+ }
+
+ if (opmode != NPUINPUT_HW_RECURRING)
+ return -EINVAL;
+
+ /** this device uses segment table */
+ SegmentTable * segt = prepareSegmentTable (model, nullptr, nullptr);
+ if (segt == nullptr) {
+ logerr (TAG, "Failed to create segment table instance\n");
+ return -EINVAL;
+ }
+
+ Request *req = new Request (opmode);
+ req->setModel (model);
+ req->setInferData (segt);
+ req->setHwDevice (hw_dev);
+
+ return scheduler_->submitRequest (req);
}
-/** @brief deprecated */
-int cleanInputBuffers (input_buffers * input)
+/** @brief callback of TRIV2 request */
+void
+TrinityVision2::callback (Request *req, npuOutputNotify cb, void *cb_data)
{
- logwarn (TAG, "deprecated. Please use cleanNPU_inputBuffers\n");
- return cleanNPU_inputBuffers (HostHandler::getLatestDevice(), input);
+ if (cb == nullptr)
+ return;
+
+ const Model *model = req->getModel ();
+ SegmentTable *segt = dynamic_cast<SegmentTable *> (req->getInferData ());
+ /** internal logic error */
+ assert (segt != nullptr);
+
+ output_buffers output = {
+ .num_buffers = segt->getNumOutputSegments ()
+ };
+
+ for (uint32_t idx = 0; idx < output.num_buffers; idx++) {
+ uint32_t output_tensor_size = model->getOutputTensorSize (idx);
+
+ output.bufs[idx].type = BUFFER_MAPPED;
+ output.bufs[idx].size = output_tensor_size;
+ /** user needs to free this */
+ output.bufs[idx].addr = calloc (1, output_tensor_size);
+
+#if defined(ENABLE_FPGA_WORKAROUND)
+ api_->fpga_memcpy (
+ segt->getOutputSegment(idx)->getDmabuf(),
+ segt->getOutputSegmentOffset(idx),
+ output.bufs[idx].addr,
+ output.bufs[idx].size);
+#else
+ auto func = std::bind (TrinityVision2::manipulateData, model, idx, false,
+ std::placeholders::_1, std::placeholders::_2, std::placeholders::_3);
+ int status = comm_.insertGenericBuffer (
+ segt->getOutputSegment(idx)->getData() + segt->getOutputSegmentOffset(idx),
+ &output.bufs[idx], func);
+
+ if (status != 0) {
+ logerr (TAG, "Failed to return output buffer: %d\n", status);
+ }
+#endif
+ }
+
+ cb (&output, req->getID(), cb_data);
+
+ delete segt;
}
-/** @brief deprecated */
-int allocNPUBuffer (uint64_t size, buffer_types type,
- const char * filepath, generic_buffer *buffer)
+/** Implement data manipulation (each device may have different impl.) */
+
+#ifdef ENABLE_MANIP
+/**
+ * @brief perform data manipulation
+ * @param[in] model model instance
+ * @param[in] idx tensor index
+ * @param[in] is_input indicate it's input manipulation
+ * @param[out] dst destination buffer
+ * @param[in] src source buffer (feature map)
+ * @param[in] size size to be copied
+ * @return size of memory copy if no error, otherwise zero
+ *
+ * @note the input data format should be NHWC
+ *
+ * @detail Feature map data in TRIV2, (x, y, z) = (width, height, depth)
+ *
+ * 1) Image input (depth == 1 or depth == 3)
+ * Addr(x,y,z) = Addr(0,0,0) + z + depth * x + ymod * y
+ *
+ * 2) Common cases
+ * Addr(x,y,z) = Addr(0,0,0) + (z % 64) + (64 * x) + ymod * y + zmod * (z / 64)
+ */
+size_t
+TrinityVision2::manipulateData (const Model *model, uint32_t idx, bool is_input,
+ void *dst, void *src, size_t size)
{
- if (buffer) {
- buffer->size = size;
- buffer->type = type;
- buffer->filepath = filepath;
+ const Metadata *meta = model->getMetadata ();
+ DataConverter converter (is_input);
+
+ converter.setData (src, dst, size);
+ converter.setTops (meta->getTops ());
+ if (is_input) {
+ const tensor_data_info* info = model->getInputDataInfo (idx);
+ if (info == nullptr)
+ return 0;
+
+ converter.setDataLayout (info->layout, DATA_LAYOUT_TRIV2);
+ converter.setDataType (info->type, meta->getInputQuantType (idx));
+ converter.setDataDims (meta->getInputDims (idx));
+ converter.setQuantZero (meta->getInputQuantZero (idx));
+ converter.setQuantScale (meta->getInputQuantScale (idx));
+ } else {
+ const tensor_data_info* info = model->getOutputDataInfo (idx);
+ if (info == nullptr)
+ return 0;
+
+ converter.setDataLayout (DATA_LAYOUT_TRIV2, info->layout);
+ converter.setDataType (meta->getOutputQuantType (idx), info->type);
+ converter.setDataDims (meta->getOutputDims (idx));
+ converter.setQuantZero (meta->getOutputQuantZero (idx));
+ converter.setQuantScale (meta->getOutputQuantScale (idx));
}
- logwarn (TAG, "deprecated. Please use allocNPU_* APIs\n");
- return allocModelBuffer (buffer);
+ return converter.perform ();
}
-/** @brief deprecated */
-int cleanNPUBuffer (generic_buffer * buffer)
+#else
+
+size_t
+TrinityVision2::manipulateData (const Model *model, uint32_t idx, bool is_input,
+ void *dst, void *src, size_t size)
{
- logwarn (TAG, "deprecated. Please use cleanNPU_* APIs\n");
- return cleanModelBuffer (buffer);
+ memcpy (dst, src, size);
+ return size;
}
+
+#endif
projects / platform / adaptation / npu / trix-engine.git / blobdiff