[NPUMGR/Dummy] Enable NHWC format handling

This patch enables NHWC format handling.

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

diff --git a/src/core/ne-handler.cc b/src/core/ne-handler.cc
index f68b531..9cffcd9 100644 (file)
--- a/src/core/ne-handler.cc
+++ b/src/core/ne-handler.cc
@@ -275,7 +275,7 @@ HostHandler::getModel (uint32_t modelid) {
 /** @brief dummay callback for runSync. */
 class callbackSync {
  public:
-  callbackSync (output_buffers *output) : output_ (output), done_ (false) {}
+  callbackSync () : done_ (false) {}
 
   static void callback (output_buffers *output, int req_id, void *data) {
     callbackSync *sync = static_cast<callbackSync *> (data);
@@ -283,10 +283,6 @@ class callbackSync {
   }
 
   void callback (output_buffers *output, int req_id) {
-    if (output_ != nullptr && output != nullptr) {
-      /** just copy internal variables of output buffers */
-      memcpy (output_, output, sizeof (output_buffers));
-    }
     done_ = true;
     cv_.notify_one ();
   }
@@ -299,7 +295,6 @@ class callbackSync {
  private:
   std::mutex m_;
   std::condition_variable cv_;
-  output_buffers *output_;
   bool done_;
 };
 
@@ -334,7 +329,7 @@ HostHandler::runModel (uint32_t modelid, npu_infer_mode mode,
 
   switch (mode) {
     case NPU_INFER_BLOCKING: {
-      callbackSync sync (output);
+      callbackSync sync;
       req_id =
           device_->run (NPUINPUT_HOST, model, input, output,
                         callbackSync::callback, static_cast<void *> (&sync));
@@ -1002,6 +997,7 @@ TrinityVision2::run (npu_input_opmode opmode, const Model *model,
   Request *req = new Request (opmode);
   req->setModel (model);
   req->setInferData (segt);
+  req->setOutputBuffers (output);
   req->setCallback (
       std::bind (&TrinityVision2::callback, this, req, cb, cb_data));
 
@@ -1046,57 +1042,66 @@ TrinityVision2::callback (Request *req, npuOutputNotify cb, void *cb_data) {
   /** internal logic error */
   assert (segt != nullptr);
 
-  output_buffers output = {.num_buffers = segt->getNumOutputSegments ()};
+  output_buffers *output = req->getOutputBuffers ();
+  /* user didn't provide valid output buffers. So, pass internally allocated one */
+  if (output == nullptr) {
+    output = new output_buffers;
+    memset (output, '\x00', sizeof (output_buffers));
+  }
+  if (output->num_buffers == 0)
+    output->num_buffers = segt->getNumOutputSegments ();
 
-  for (uint32_t idx = 0; idx < output.num_buffers; idx++) {
+  /* perform data manipulation such as layout conversion and de-quantization */
+  for (uint32_t idx = 0; idx < output->num_buffers; idx++) {
     uint32_t output_tensor_size = model->getOutputTensorSize (idx);
     HWmem *output_segment = segt->getOutputSegment (idx);
 
-    if (output_segment->isExternal ()) {
-      output.bufs[idx].type = BUFFER_DMABUF;
-      output.bufs[idx].size = output_segment->getSize ();
-      output.bufs[idx].addr = output_segment->getData ();
-      output.bufs[idx].dmabuf = output_segment->getDmabuf ();
-      output.bufs[idx].offset = output_segment->getOffset ();
-    } else {
-      output.bufs[idx].type = BUFFER_MAPPED;
-      output.bufs[idx].size = output_tensor_size;
+    if (output->bufs[idx].type == BUFFER_DMABUF) {
+      /* it's external memory. can't do */
+      continue;
+    } else if (output->bufs[idx].addr == nullptr ||
+               output->bufs[idx].size == 0) {
+      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 (output.bufs[idx].addr == NULL) {
+      output->bufs[idx].addr = calloc (1, output_tensor_size);
+      if (output->bufs[idx].addr == NULL) {
         logerr (TAG, "Unable to allocate output buffer\n");
         break;
       }
+    }
 
-      auto func = std::bind (TrinityVision2::manipulateData, model, idx, false,
-                             std::placeholders::_1, std::placeholders::_2,
-                             std::placeholders::_3);
-      void *dst;
+    auto func = std::bind (TrinityVision2::manipulateData, model, idx, false,
+                           std::placeholders::_1, std::placeholders::_2,
+                           std::placeholders::_3);
+    void *dst;
 #ifdef __FPGA__
-      /* this is fpga workaround codes for syncing output data */
-      dst = calloc (1, output_tensor_size);
-      if (dst == NULL) {
-        logerr (TAG, "Unable to allocate FPGA temp buffer\n");
-        break;
-      }
-      api_->fpga_memcpy (output_segment->getDmabuf (),
-                         segt->getOutputSegmentOffset (idx), dst,
-                         output_tensor_size);
+    /* this is fpga workaround codes for syncing output data */
+    dst = calloc (1, output_tensor_size);
+    if (dst == NULL) {
+      logerr (TAG, "Unable to allocate FPGA temp buffer\n");
+      break;
+    }
+    api_->fpga_memcpy (output_segment->getDmabuf (),
+                       segt->getOutputSegmentOffset (idx), dst,
+                       output_tensor_size);
 #else
-      dst = output_segment->getData () + segt->getOutputSegmentOffset (idx);
+    dst = output_segment->getData () + segt->getOutputSegmentOffset (idx);
 #endif
-      int status = comm_.insertGenericBuffer (dst, &output.bufs[idx], func);
-      if (status != 0) {
-        logerr (TAG, "Failed to return output buffer: %d\n", status);
-      }
+    int status = comm_.insertGenericBuffer (dst, &output->bufs[idx], func);
+    if (status != 0) {
+      logerr (TAG, "Failed to return output buffer: %d\n", status);
+    }
+
 #ifdef __FPGA__
-      free (dst);
+    free (dst);
 #endif
-    }
   }
 
-  cb (&output, req->getID (), cb_data);
+  cb (output, req->getID (), cb_data);
 
+  if (req->getOutputBuffers () != output)
+    delete output;
   delete segt;
 }
 

diff --git a/src/core/ne-scheduler.cc b/src/core/ne-scheduler.cc
index 8e3d21b..db3fd75 100644 (file)
--- a/src/core/ne-scheduler.cc
+++ b/src/core/ne-scheduler.cc
@@ -27,7 +27,8 @@ Request::Request (npu_input_opmode opmode)
       stopped_ (false),
       model_ (nullptr),
       data_ (nullptr),
-      cb_ (nullptr) {
+      cb_ (nullptr),
+      out_bufs_ (nullptr) {
   request_id_ = Request::global_request_id_.fetch_add (1);
 }
 

diff --git a/src/core/ne-scheduler.h b/src/core/ne-scheduler.h
index 0e58a39..2f6b3f4 100644 (file)
--- a/src/core/ne-scheduler.h
+++ b/src/core/ne-scheduler.h
@@ -53,6 +53,9 @@ class Request {
   void setHwDevice (std::string hw_dev) { hw_dev_ = hw_dev; }
   std::string getHwDevice () { return hw_dev_; }
 
+  void setOutputBuffers (output_buffers *out_bufs) { out_bufs_ = out_bufs; }
+  output_buffers *getOutputBuffers () { return out_bufs_; }
+
  private:
   static std::atomic<int> global_request_id_;
   int request_id_; /**< request id */
@@ -64,8 +67,9 @@ class Request {
   const Model *model_; /**< model of the request */
   HWmem *data_;        /**< inference data of the request */
 
-  outputCallback cb_;  /**< request callback */
-  std::string hw_dev_; /**< HW device path */
+  outputCallback cb_;        /**< request callback */
+  output_buffers *out_bufs_; /**< output buffers */
+  std::string hw_dev_;       /**< HW device path */
 };
 
 /** @brief class def. of scheduler to handle requests */

diff --git a/tests/apptests/npumgr/dummy/npumgr_api.h b/tests/apptests/npumgr/dummy/npumgr_api.h
index 6fae8e7..c17dcf8 100644 (file)
--- a/tests/apptests/npumgr/dummy/npumgr_api.h
+++ b/tests/apptests/npumgr/dummy/npumgr_api.h
@@ -165,7 +165,6 @@ typedef enum _npumgr_tensor_qnt_type {
 typedef enum _npumgr_tensor_fmt {
        NPUMGR_TENSOR_FMT_NCHW = 0,
        NPUMGR_TENSOR_FMT_NHWC,
-       NPUMGR_TENSOR_FMT_TRIV2,
 
        NPUMGR_TENSOR_FMT_MAX
 }npumgr_tensor_fmt_t;

diff --git a/tests/apptests/npumgr/dummy/npumgr_triv2.cc b/tests/apptests/npumgr/dummy/npumgr_triv2.cc
index e8b473d..c6cb0cf 100644 (file)
--- a/tests/apptests/npumgr/dummy/npumgr_triv2.cc
+++ b/tests/apptests/npumgr/dummy/npumgr_triv2.cc
@@ -90,9 +90,6 @@ convert_layout (const npumgr_tensor_fmt_t &fmt) {
     case NPUMGR_TENSOR_FMT_NHWC:
       layout = DATA_LAYOUT_NHWC;
       break;
-    case NPUMGR_TENSOR_FMT_TRIV2:
-      layout = DATA_LAYOUT_TRIV2;
-      break;
     default:
       break;
   }
@@ -133,6 +130,28 @@ convert_type (const npumgr_tensor_data_t &data,
   return type;
 }
 
+static gsize
+calc_data_size (const npumgr_tensor_data_t type) {
+  switch (type) {
+    case NPUMGR_TENSOR_DATA_FLOAT32:
+      return 4;
+    case NPUMGR_TENSOR_DATA_FLOAT16:
+      return 2;
+    case NPUMGR_TENSOR_DATA_INT8:
+      return 1;
+    case NPUMGR_TENSOR_DATA_UINT8:
+      return 1;
+    case NPUMGR_TENSOR_DATA_INT16:
+      return 2;
+    case NPUMGR_TENSOR_DATA_UINT16:
+      return 2;
+    case NPUMGR_TENSOR_DATA_UINT24:
+      return 3;
+    default:
+      return 0;
+  }
+}
+
 /**
  * @brief Class for triv2 npumgr network
  */
@@ -228,14 +247,20 @@ class NpumgrNetworkTriv2 {
   guint getInTensorDim (guint i, guint j) {
     return meta_->input_seg_dims[i][j];
   }
-  guint getInTensorSize (guint i) {
+  guint getInTensorSize (guint i, npumgr_tensor_fmt_t fmt,
+                         npumgr_tensor_data_t type) {
     guint size;
-    int status;
 
-    status = getNPUmodel_tensorSize (dev_, model_id_, true, i, &size);
-    if (status != 0)
-      return 0;
+    if (fmt == NPUMGR_TENSOR_FMT_NHWC || fmt == NPUMGR_TENSOR_FMT_NCHW) {
+      size = calc_data_size (type);
+      for (int j = 0; j < MAX_RANK; j++) size *= meta_->input_seg_dims[i][j];
+    } else {
+      int status;
 
+      status = getNPUmodel_tensorSize (dev_, model_id_, true, i, &size);
+      if (status != 0)
+        return 0;
+    }
     return size;
   }
   int32_t getInTensorQuantZero (guint i) { return meta_->input_seg_quant_z[i]; }
@@ -244,14 +269,20 @@ class NpumgrNetworkTriv2 {
   guint getOutTensorDim (gint i, gint j) {
     return meta_->output_seg_dims[i][j];
   }
-  guint getOutTensorSize (guint i) {
+  guint getOutTensorSize (guint i, npumgr_tensor_fmt_t fmt,
+                          npumgr_tensor_data_t type) {
     guint size;
-    int status;
 
-    status = getNPUmodel_tensorSize (dev_, model_id_, false, i, &size);
-    if (status != 0)
-      return 0;
+    if (fmt == NPUMGR_TENSOR_FMT_NHWC || fmt == NPUMGR_TENSOR_FMT_NCHW) {
+      size = calc_data_size (type);
+      for (int j = 0; j < MAX_RANK; j++) size *= meta_->output_seg_dims[i][j];
+    } else {
+      int status;
 
+      status = getNPUmodel_tensorSize (dev_, model_id_, false, i, &size);
+      if (status != 0)
+        return 0;
+    }
     return size;
   }
   int32_t getOutTensorQuantZero (gint i) {
@@ -269,7 +300,6 @@ class NpumgrNetworkTriv2 {
 
     memcpy (&in_buffers_.bufs[index], buffer->getGenericBuffer (),
             sizeof (generic_buffer));
-    in_buffers_.bufs[index].type = BUFFER_DMABUF;
     return TRUE;
   }
 
@@ -283,7 +313,6 @@ class NpumgrNetworkTriv2 {
 
     memcpy (&out_buffers_.bufs[index], buffer->getGenericBuffer (),
             sizeof (generic_buffer));
-    out_buffers_.bufs[index].type = BUFFER_DMABUF;
     return TRUE;
   }
 
@@ -1027,9 +1056,9 @@ triv2_query_input (NpumgrDevice *device, npumgr_context ctx_handle,
   for (guint rank = 0; rank < MAX_RANK; rank++) attr->strides[rank] = 1;
   attr->plane = 3;
 
-  attr->size = network->getInTensorSize (index);
-  attr->fmt = NPUMGR_TENSOR_FMT_TRIV2;
+  attr->fmt = NPUMGR_TENSOR_FMT_NHWC;
   attr->type = NPUMGR_TENSOR_DATA_UINT8;
+  attr->size = network->getInTensorSize (index, attr->fmt, attr->type);
 
   attr->quant_type = NPUMGR_TENSOR_QNT_AFFINE_ASYMM;
   attr->quant_data.affine.scale = network->getInTensorQuantScale (index);
@@ -1067,9 +1096,9 @@ triv2_query_output (NpumgrDevice *device, npumgr_context ctx_handle,
   for (guint rank = 0; rank < MAX_RANK; rank++) attr->strides[rank] = 1;
   attr->plane = 3;
 
-  attr->size = network->getOutTensorSize (index);
-  attr->fmt = NPUMGR_TENSOR_FMT_TRIV2;
+  attr->fmt = NPUMGR_TENSOR_FMT_NHWC;
   attr->type = NPUMGR_TENSOR_DATA_UINT8;
+  attr->size = network->getOutTensorSize (index, attr->fmt, attr->type);
 
   attr->quant_type = NPUMGR_TENSOR_QNT_AFFINE_ASYMM;
   attr->quant_data.affine.scale = network->getOutTensorQuantScale (index);

diff --git a/tests/apptests/npumgr/npumgr_test.cc b/tests/apptests/npumgr/npumgr_test.cc
index 8b7ddbd..9f8f670 100644 (file)
--- a/tests/apptests/npumgr/npumgr_test.cc
+++ b/tests/apptests/npumgr/npumgr_test.cc
@@ -105,8 +105,8 @@ start_npumgr_test (int fd, const string &dir) {
       goto destroy_all;
     }
 
-    /** TODO: currently, support TRIV2 (NHWC-based) format */
-    if (attr.fmt != NPUMGR_TENSOR_FMT_TRIV2) {
+    /** TODO: currently, support NHWC format only */
+    if (attr.fmt != NPUMGR_TENSOR_FMT_NHWC) {
       cerr << "Other format is not supported yet, " << attr.fmt << "\n";
       fclose (f);
       goto destroy_all;
@@ -160,8 +160,8 @@ start_npumgr_test (int fd, const string &dir) {
       goto destroy_all;
     }
 
-    /** TODO: currently, support TRIV2 (NHWC-based) format */
-    if (attr.fmt != NPUMGR_TENSOR_FMT_TRIV2) {
+    /** TODO: currently, support NHWC format only */
+    if (attr.fmt != NPUMGR_TENSOR_FMT_NHWC) {
       cerr << "Other format is not supported yet, " << attr.fmt << "\n";
       goto destroy_all;
     }