[Profile] Support layer manipulation in kernel profiling

This patch supports layer manipulation in kernel profiling.

Previously, the layer manipulation was supported in emulated
driver API using the mrpsim simulator.

In this patch, I added ModelProfiler class to provide common
interface used in both simulator and kernel driver.

Signed-off-by: Dongju Chae <dongju.chae@samsung.com>

diff --git a/src/core/meson.build b/src/core/meson.build
index 85e9e13..c2c808b 100644 (file)
--- a/src/core/meson.build
+++ b/src/core/meson.build
@@ -16,6 +16,7 @@ ne_core_sources = [
   'ne-mem.cc',
   'ne-data.cc',
   'ne-handler.cc',
+  'ne-profiler.cc',
   'ne-scheduler.cc',
   'ne-host-input-service.cc',
   'ne-hw-input-service.cc',

diff --git a/src/core/ne-handler.cc b/src/core/ne-handler.cc
index 1550b79..260fca9 100644 (file)
--- a/src/core/ne-handler.cc
+++ b/src/core/ne-handler.cc
@@ -35,11 +35,13 @@ HostHandler::HostHandler (Device *device)
     /* ignored as we don't use double buffering anymore, but for backward-compatibility */
     async_mode_ (NPUASYNC_WAIT)
 {
+  profiler_ = new ModelProfiler (device_->getDriverAPI ());
 }
 
 /** @brief host handler destructor */
 HostHandler::~HostHandler ()
 {
+  delete profiler_;
 }
 
 /**
@@ -133,19 +135,10 @@ HostHandler::getProfile (int task_id, npu_profile *profile)
     return -EINVAL;
   }
 
-  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 0;
+  return profiler_->getTaskProfile (task_id, profile);
 }
 
 /**
@@ -354,7 +347,12 @@ HostHandler::runAsync (uint32_t modelid, const input_buffers *input,
   }
 
   device_->setAsyncMode (mode);
-  return device_->run (NPUINPUT_HOST, model, input, cb, cb_data, sequence);
+
+  int task_id = device_->run (NPUINPUT_HOST, model, input, cb, cb_data, sequence);
+  if (task_id > 0)
+    profiler_->appendTask (task_id, model);
+
+  return task_id;
 }
 
 /**

diff --git a/src/core/ne-handler.h b/src/core/ne-handler.h
index 47b41ab..f9fa4f6 100644 (file)
--- a/src/core/ne-handler.h
+++ b/src/core/ne-handler.h
@@ -22,6 +22,7 @@
 #include "ne-scheduler.h"
 #include "ne-model.h"
 #include "ne-utils.h"
+#include "ne-profiler.h"
 
 class Device;
 /** @brief class def. of host handler */
@@ -73,6 +74,8 @@ class HostHandler {
     HostHandler (Device *device);
 
     Device *device_;  /**< dedicated device instance */
+    ModelProfiler *profiler_;
+
     ThreadSafeMap<uint32_t, Model> models_;
                       /**< registerd models */
     npu_async_mode async_mode_;

diff --git a/src/core/ne-model.h b/src/core/ne-model.h
index e60be65..dc2afef 100644 (file)
--- a/src/core/ne-model.h
+++ b/src/core/ne-model.h
@@ -347,6 +347,7 @@ class Model : public HWmem {
     uint64_t getInternalID () const { return internal_id_; }
     HWmem * getWeightData () const { return weight_data_; }
     HWmem * getProgramData () const { return program_data_; }
+    HWmem * getExtendedMetadata () const { return extended_meta_; }
 
     uint32_t getInputTensorNum () const;
     uint32_t getOutputTensorNum () const;

diff --git a/src/core/ne-profiler.cc b/src/core/ne-profiler.cc
new file mode 100644 (file)
index 0000000..3b6b1c0
--- /dev/null
+++ b/src/core/ne-profiler.cc
@@ -0,0 +1,175 @@
+/**
+ * Proprietary
+ * Copyright (C) 2021 Samsung Electronics
+ * Copyright (C) 2021 Dongju Chae <dongju.chae@samsung.com>
+ */
+/**
+ * @file ne-profiler.cc
+ * @date 25 Mar 2021
+ * @brief Model profiler for NPU Engine (NE) users.
+ * @author Dongju Chae <dongju.chae@samsung.com>
+ * @bug No known bugs except for NYI items
+ */
+
+#include "ne-profiler.h"
+
+ModelProfiler::ModelProfiler (const DriverAPI * api)
+  : api_ (api)
+{
+}
+
+ModelProfiler::~ModelProfiler ()
+{
+  profile_map_.clear ();
+}
+
+int
+ModelProfiler::appendTask (int task_id, const Model * model)
+{
+  ProfileData * data = new ProfileData (task_id, model);
+  return profile_map_.insert (task_id, data);
+}
+
+int
+ModelProfiler::getTaskProfile (int task_id, npu_profile *profile)
+{
+  ProfileData * data = profile_map_.find (task_id);
+  if (data == nullptr)
+    return -ENOENT;
+
+  const Model * model = data->getModel ();
+  if (model == nullptr)
+    return -EINVAL;
+
+  int status = api_->getProfile (task_id, profile);
+  if (status != 0)
+    return status;
+
+  HWmem * extended = model->getExtendedMetadata ();
+  if (extended != nullptr)
+    manipulateProfile (extended, profile);
+
+  profile_map_.remove (task_id);
+
+  return 0;
+}
+
+void
+ModelProfiler::manipulateProfile (HWmem * extended, npu_profile *profile)
+{
+  npubin_meta_profile *meta_profile =
+    reinterpret_cast <npubin_meta_profile *> (extended->getData ());
+  npu_profile_layer * new_layers =
+    new npu_profile_layer[meta_profile->node_entry_num + 1];
+
+  npu_profile_layer * unclassified =
+    &new_layers[meta_profile->node_entry_num];
+
+  snprintf (unclassified->name, NPU_OPNAME_MAX - 1, "%s", "Unclassified");
+  unclassified->name[NPU_OPNAME_MAX - 1] = '\x00';
+  unclassified->node_id = -1;
+
+  unclassified->running_cycles = 0;
+  unclassified->dram_read_bytes = 0;
+  unclassified->dram_write_bytes = 0;
+  unclassified->sram_read_bytes = 0;
+  unclassified->sram_write_bytes = 0;
+
+  /** 1) parsing node table */
+  std::unordered_map<uint32_t, npu_profile_layer *> node_table;
+  uint32_t pos = 0;
+
+  node_table.reserve (meta_profile->node_entry_num);
+
+  for (uint32_t i = 0; i < meta_profile->node_entry_num; i++) {
+    uint32_t id, length;
+
+    memcpy (&id, meta_profile->entry_data + pos, sizeof (uint32_t));
+    pos += sizeof (uint32_t);
+
+    memcpy (&length, meta_profile->entry_data + pos, sizeof (uint32_t));
+    pos += sizeof (uint32_t);
+
+    if (length == 0) {
+      std::cerr << "Zero length detected at ";
+      std::cerr << id << "th node" << std::endl;
+
+      delete [] new_layers;
+      return;
+    }
+
+    std::string name (meta_profile->entry_data + pos);
+    pos += length;
+
+    npu_profile_layer * layer = &new_layers[i];
+
+    snprintf (layer->name, NPU_OPNAME_MAX - 1, "%s", name.c_str ());
+    layer->name[NPU_OPNAME_MAX - 1] = '\x00';
+    layer->node_id = id;
+
+    layer->running_cycles = 0;
+    layer->dram_read_bytes = 0;
+    layer->dram_write_bytes = 0;
+    layer->sram_read_bytes = 0;
+    layer->sram_write_bytes = 0;
+
+    node_table.insert(std::make_pair(id, layer));
+  }
+
+  /** 2) parsing visa table */
+  pos = meta_profile->node_table_size;
+  for (uint32_t i = 0; i < meta_profile->visa_entry_num; i++) {
+    uint32_t id, node_num;
+
+    memcpy (&id, meta_profile->entry_data + pos, sizeof (uint32_t));
+    pos += sizeof (uint32_t);
+
+    memcpy (&node_num, meta_profile->entry_data + pos, sizeof (uint32_t));
+    pos += sizeof (uint32_t);
+
+    if (node_num > 0) {
+      uint32_t * node_ids = (uint32_t *) (meta_profile->entry_data + pos);
+
+      for (uint32_t j = 0; j < node_num; j++) {
+        uint32_t node_id = node_ids[j];
+        auto it = node_table.find (node_id);
+
+        if (it != node_table.end ()) {
+          npu_profile_layer * layer = it->second;
+
+          /** TODO: evenly divided to fused layers */
+          layer->running_cycles += profile->layers[i].running_cycles / node_num;
+          layer->dram_read_bytes += profile->layers[i].dram_read_bytes / node_num;
+          layer->dram_write_bytes += profile->layers[i].dram_write_bytes / node_num;
+          layer->sram_read_bytes += profile->layers[i].sram_read_bytes / node_num;
+          layer->sram_write_bytes += profile->layers[i].sram_write_bytes / node_num;
+          layer->visa_exec_seq = -1;
+        } else {
+          std::cerr << "Unable to find the node ID " << node_id << std::endl;
+        }
+      }
+    } else {
+      unclassified->running_cycles += profile->layers[i].running_cycles;
+      unclassified->dram_read_bytes += profile->layers[i].dram_read_bytes;
+      unclassified->dram_write_bytes += profile->layers[i].dram_write_bytes;
+      unclassified->sram_read_bytes += profile->layers[i].sram_read_bytes;
+      unclassified->sram_write_bytes += profile->layers[i].sram_write_bytes;
+      unclassified->visa_exec_seq = -1;
+    }
+
+    pos += sizeof (uint32_t) * node_num;
+  }
+
+  /** 3) profile data mapping */
+  size_t num_layers = node_table.size ();
+  if (num_layers > 0) {
+    delete [] profile->layers;
+
+    profile->layers = new_layers;
+    profile->num_layers = num_layers + 1;
+  } else {
+    delete [] new_layers;
+  }
+}
+
+

diff --git a/src/core/ne-profiler.h b/src/core/ne-profiler.h
new file mode 100644 (file)
index 0000000..3e3bd97
--- /dev/null
+++ b/src/core/ne-profiler.h
@@ -0,0 +1,49 @@
+/**
+ * Proprietary
+ * Copyright (C) 2021 Samsung Electronics
+ * Copyright (C) 2021 Dongju Chae <dongju.chae@samsung.com>
+ */
+/**
+ * @file ne-profiler.h
+ * @date 25 Mar 2021
+ * @brief Model profiler for NPU Engine (NE) users.
+ * @author Dongju Chae <dongju.chae@samsung.com>
+ * @bug No known bugs except for NYI items
+ */
+
+#ifndef __NPU_ENGINE_PROFILER_H__
+#define __NPU_ENGINE_PROFILER_H__
+
+#include <typedef.h>
+#include "ne-model.h"
+
+class ProfileData
+{
+  public:
+    ProfileData (int task_id, const Model * model) :
+      task_id_ (task_id), model_ (model) {}
+
+    const Model * getModel () { return model_; }
+
+  private:
+    int task_id_;
+    const Model * model_;
+};
+
+class ModelProfiler
+{
+  public:
+    ModelProfiler (const DriverAPI * api);
+    ~ModelProfiler ();
+
+    int appendTask (int task_id, const Model * model);
+    int getTaskProfile (int task_id, npu_profile *profile);
+
+    void manipulateProfile (HWmem * extended, npu_profile *profile);
+
+  private:
+    const DriverAPI * api_;
+    ThreadSafeMap<int, ProfileData> profile_map_;
+};
+
+#endif

diff --git a/src/core/npu/NPUdrvAPI.h b/src/core/npu/NPUdrvAPI.h
index fbe813a..a251e63 100644 (file)
--- a/src/core/npu/NPUdrvAPI.h
+++ b/src/core/npu/NPUdrvAPI.h
@@ -108,6 +108,7 @@ class DriverAPI {
         void *addr, size_t size) const { return -EPERM; }
 #endif
 
+    /** @brief get profile data for vISA instructions */
     virtual int getProfile (int task_id, npu_profile *profile) const { return -EPERM; }
 
     virtual int getStatApps (npu_stat_apps *stat) const { return -EPERM; }

diff --git a/src/core/npu/NPUdrvAPI_emul.cc b/src/core/npu/NPUdrvAPI_emul.cc
index e221009..993c9b4 100644 (file)
--- a/src/core/npu/NPUdrvAPI_emul.cc
+++ b/src/core/npu/NPUdrvAPI_emul.cc
@@ -30,8 +30,7 @@
 class EmulTask {
   public:
     EmulTask (int taskid) :
-      taskid_ (taskid), stop_ (false), first_run_ (false),
-      extended_dbuf_fd_ (0), extended_size_ (0) {}
+      taskid_ (taskid), stop_ (false), first_run_ (false) {}
 
     void run_emul (char *prog, char **segt, char *metadata,
         std::string cmd_path, std::string prof_path) {
@@ -57,19 +56,6 @@ class EmulTask {
       task_.join ();
     }
 
-    void set_extended (int dbuf_fd, size_t size) {
-      extended_dbuf_fd_ = dbuf_fd;
-      extended_size_ = size;
-    }
-
-    int get_extended_dbuf_fd () {
-      return extended_dbuf_fd_;
-    }
-
-    size_t get_extended_size () {
-      return extended_size_;
-    }
-
     bool get_profile (npu_profile *profile) {
       std::string path (DEFAULT_PROFILE_PATH);
       path += "/ne_profile." + std::to_string (taskid_) + ".rec";
@@ -191,9 +177,6 @@ class EmulTask {
     bool stop_;
     bool first_run_;
     std::thread task_;
-
-    int extended_dbuf_fd_;
-    size_t extended_size_;
 };
 
 /**
@@ -643,7 +626,6 @@ TrinityEmulAPI::runInput (input_config_t *input_config) const
         prog, segment_table, static_cast <char*> (elem_metadata->getAddr ()),
         cmd_path, prof_path);
 
-    task->set_extended (model->metadata_ext_dbuf_fd, model->metadata_ext_size);
     task->run (func);
     status = taskid;
 
@@ -692,139 +674,6 @@ TrinityEmulAPI::stop_target (int taskid) const
   return 0;
 }
 
-void
-TrinityEmulAPI::manipulateProfile (EmulTask *task, npu_profile *profile) const
-{
-  int dbuf_fd = task->get_extended_dbuf_fd ();
-  size_t size = task->get_extended_size ();
-
-  EmulElement *elem = elem_map_.find (dbuf_fd);
-  if (elem == nullptr || elem->getAddr () == nullptr) {
-    std::cerr << "No available extended metadata" << std::endl;
-    return;
-  }
-
-  if (elem->getSize () != size) {
-    std::cerr << "Extended metadata size mismatch: ";
-    std::cerr << elem->getSize () << " vs. " << size << std::endl;
-    return;
-  }
-
-  npubin_meta_profile *meta_profile =
-         static_cast<npubin_meta_profile *> (elem->getAddr ());
-  npu_profile_layer * new_layers =
-    new npu_profile_layer[meta_profile->node_entry_num + 1];
-
-  npu_profile_layer * unclassified =
-    &new_layers[meta_profile->node_entry_num];
-
-  snprintf (unclassified->name, NPU_OPNAME_MAX - 1, "%s", "Unclassified");
-  unclassified->name[NPU_OPNAME_MAX - 1] = '\x00';
-  unclassified->node_id = -1;
-
-  unclassified->running_cycles = 0;
-  unclassified->dram_read_bytes = 0;
-  unclassified->dram_write_bytes = 0;
-  unclassified->sram_read_bytes = 0;
-  unclassified->sram_write_bytes = 0;
-
-  /** 1) parsing node table */
-  std::unordered_map<uint32_t, npu_profile_layer *> node_table;
-  uint32_t pos = 0;
-
-  node_table.reserve (meta_profile->node_entry_num);
-
-  for (uint32_t i = 0; i < meta_profile->node_entry_num; i++) {
-    uint32_t id, length;
-
-    memcpy (&id, meta_profile->entry_data + pos, sizeof (uint32_t));
-    pos += sizeof (uint32_t);
-
-    memcpy (&length, meta_profile->entry_data + pos, sizeof (uint32_t));
-    pos += sizeof (uint32_t);
-
-    if (length == 0) {
-      std::cerr << "Zero length detected at ";
-      std::cerr << id << "th node" << std::endl;
-
-      delete [] new_layers;
-      return;
-    }
-
-    std::string name (meta_profile->entry_data + pos);
-    pos += length;
-
-    npu_profile_layer * layer = &new_layers[i];
-
-    snprintf (layer->name, NPU_OPNAME_MAX - 1, "%s", name.c_str ());
-    layer->name[NPU_OPNAME_MAX - 1] = '\x00';
-    layer->node_id = id;
-
-    layer->running_cycles = 0;
-    layer->dram_read_bytes = 0;
-    layer->dram_write_bytes = 0;
-    layer->sram_read_bytes = 0;
-    layer->sram_write_bytes = 0;
-
-    node_table.insert(std::make_pair(id, layer));
-  }
-
-  /** 2) parsing visa table */
-  pos = meta_profile->node_table_size;
-  for (uint32_t i = 0; i < meta_profile->visa_entry_num; i++) {
-    uint32_t id, node_num;
-
-    memcpy (&id, meta_profile->entry_data + pos, sizeof (uint32_t));
-    pos += sizeof (uint32_t);
-
-    memcpy (&node_num, meta_profile->entry_data + pos, sizeof (uint32_t));
-    pos += sizeof (uint32_t);
-
-    if (node_num > 0) {
-      uint32_t * node_ids = (uint32_t *) (meta_profile->entry_data + pos);
-
-      for (uint32_t j = 0; j < node_num; j++) {
-        uint32_t node_id = node_ids[j];
-        auto it = node_table.find (node_id);
-
-        if (it != node_table.end ()) {
-          npu_profile_layer * layer = it->second;
-
-          /** TODO: evenly divided to fused layers */
-          layer->running_cycles += profile->layers[i].running_cycles / node_num;
-          layer->dram_read_bytes += profile->layers[i].dram_read_bytes / node_num;
-          layer->dram_write_bytes += profile->layers[i].dram_write_bytes / node_num;
-          layer->sram_read_bytes += profile->layers[i].sram_read_bytes / node_num;
-          layer->sram_write_bytes += profile->layers[i].sram_write_bytes / node_num;
-          layer->visa_exec_seq = -1;
-        } else {
-          std::cerr << "Unable to find the node ID " << node_id << std::endl;
-        }
-      }
-    } else {
-      unclassified->running_cycles += profile->layers[i].running_cycles;
-      unclassified->dram_read_bytes += profile->layers[i].dram_read_bytes;
-      unclassified->dram_write_bytes += profile->layers[i].dram_write_bytes;
-      unclassified->sram_read_bytes += profile->layers[i].sram_read_bytes;
-      unclassified->sram_write_bytes += profile->layers[i].sram_write_bytes;
-      unclassified->visa_exec_seq = -1;
-    }
-
-    pos += sizeof (uint32_t) * node_num;
-  }
-
-  /** 3) profile data mapping */
-  size_t num_layers = node_table.size ();
-  if (node_table.size () > 0) {
-    delete [] profile->layers;
-
-    profile->layers = new_layers;
-    profile->num_layers = num_layers + 1;
-  } else {
-    delete [] new_layers;
-  }
-}
-
 int
 TrinityEmulAPI::getProfile (int taskid, npu_profile *profile) const
 {
@@ -838,9 +687,6 @@ TrinityEmulAPI::getProfile (int taskid, npu_profile *profile) const
   if (!task->get_profile (profile))
     return -EINVAL;
 
-  if (task->get_extended_size () != 0)
-    manipulateProfile (task, profile);
-
   task_map_.remove (taskid);
 
   return 0;

diff --git a/src/core/npu/NPUdrvAPI_triv2.cc b/src/core/npu/NPUdrvAPI_triv2.cc
index 175e0b3..eccbc5a 100644 (file)
--- a/src/core/npu/NPUdrvAPI_triv2.cc
+++ b/src/core/npu/NPUdrvAPI_triv2.cc
@@ -473,8 +473,6 @@ TrinityVision2API::getProfile (int task_id, npu_profile *profile) const
     return -errno;
   }
 
-  /** TODO: manipulate the profiling info later (i.e., per-visa to per-layer) */
-
   return 0;
 }