clDNN Documentation


Compute Library for Deep Neural Networks (clDNN) is a middle-ware software for accelerating DNN inference on Intel® HD and Iris™ Pro Graphics. This project includes CNN primitives implementations on Intel GPUs with C and C++ interfaces.

clDNN Library implements set of primitives:

  • Convolution
  • Fully connected (inner product)
  • Pooling
    • average
    • maximum
  • Normalization
    • across channel
    • within channel
    • batch
  • Activation
    • logistic
    • tanh
    • rectified linear unit (ReLU)
    • softplus (softReLU)
    • abs
    • square
    • sqrt
    • linear
  • Softmax
  • Crop
  • Deconvolution
  • Depth concatenation
  • Eltwise
  • ROI pooling
  • Simpler NMS
  • Prior box
  • Detection output

With this primitive set, user can build and execute most common image recognition, semantic segmentation and object detection networks topologies like:

  • Alexnet
  • Googlenet(v1-v3)
  • ResNet
  • VGG
  • faster-rCNN and other.

Programming Model

Intel® clDNN is graph oriented library. To execute CNN you have to build, compile graph/topology and run to get results.


  • Primitive - dnn base functionality i.e. convolution, pooling, softmax.
  • Data - special primitive type representing primitive parameters (weights and biases), inputs and outputs
  • Engine - type of accelerator that is executing network. Currently ocl engine is the only available.
  • Topology - container of primitives, data, and relations between them. Topology represents graph.
  • Program - optional step between Topology and Network. It is compiled Topology without memory allocation.
  • Network - compiled Topology with memory allocation. Ready to be executed. During compilation, buidling parameters trigger special optimizations like fusing, data reordering.

Execution Steps:

  1. Create Engine
  2. Declare or define primitives parameters (weights and biases) if needed.
  3. Create primitives. It is required to provide name for each primitive. This is a name of primitive which output will be input to current one. Name can be used before primitive definition.
  4. Create topology
  5. Add primitives to topology
  6. Build Network from topology
  7. Set Inputs data
  8. Execute Network

Graph compilation

If user choose build option optimize_data when program is being created - explicit or implicit over network creation, clDNN perform some graph optimizations as follows:

  • Stage 0: Graph initiation:
    • build nodes from primitives
    • node replacement:
      • replace each split node with series of crop nodes. Name of crop primitive will be concatenation of split + port names.
      • replace upsampling node with deconvolution node if upsampling mode is bilinear.
    • set outputs - mark nodes that are defined by user as output (blocks fusing etc) or have no users (leafs).
    • calculate processing order - using dfs on graph to establish processing order
  • Stage 1: Priorboxes:
    • priorbox is primitive that is executed during network compilation. Node is removed from a network execution.
  • Stage 2: Graph analysis:
    • mark constatns
    • mark data flow
    • mark dominators
  • Stage 3: Trimming:
    • apply backward bfs on each output to find unnecessary nodes/branches, then remove those.
  • Stage 4: Inputs and biases:
    • reorder input - format of convolution's input/output is being selected.
    • reorder biases for conv,fc and deconv nodes
  • Stage 5: Redundant reorders:
    • previous stages can provide additional reorders due to format changes per primitive. This stage removes redundant and fuses series of reorders into one.
      * Stage 6: Constant propagation:
    • prepare padding - goes thrugh all primitves and checks if its user requires padding, if so, set output padding.
    • prepare depthwise separable opt - if split param is greater than 16 and number of IFM <= 8*split in conv or deconv, this stage changes execution from multi kernels into one.
    • constant propagation - replace constant nodes, that are not outputs with data type nodes. Constant primitive is the primitive that doesn't depend on any non-constant primitive and doesn't have to be executed: priorbox, data.
  • Stage 7: Fusing:
    • buffer fusing
      • concat - if concatenation is the only user of its dependencies then remove concat node and setting proper output paddings in every dependencies.
      • crop - if crop has only one dependecy, and its users doesn't require padding, remove crop and set proper output padding in its dependecy.
      • reorder - if primitive before reorder supports different input vs output type reorder can be fused with previous node.
    • primitive fusing - right now this stage fuses activation node with previous node only, only if previous node supports activation fusing.
  • Stage 8: Compile graph:
    • at this stage using kernel selector, graph chooses the best kernel implementation for each node.
  • Stage 9: reorder weights:
    • at this stage weights are converted into format suitable for selected kernel implementation.
  • Stage 10 & 11: Redundant reorders and constant propagation:
    • check again if whole graph compilation didn't provide any redundant reorders and constants.
  • Stage 12: Compile program:
    • at this stage engine compiles cl_kernels.

C++ API Example MNIST network

#include <api/CPP/memory.hpp>
#include <api/CPP/topology.hpp>
#include <api/CPP/reorder.hpp>
#include <api/CPP/input_layout.hpp>
#include <api/CPP/convolution.hpp>
#include <api/CPP/data.hpp>
#include <api/CPP/pooling.hpp>
#include <api/CPP/fully_connected.hpp>
#include <api/CPP/softmax.hpp>
#include <api/CPP/engine.hpp>
#include <api/CPP/network.hpp>
#include <iostream>
using namespace cldnn;
using namespace std;
input_channels = 1,
input_size = 28,
conv1_out_channels = 20,
conv2_out_channels = 50,
conv_krnl_size = 5,
fc1_num_outs = 500,
fc2_num_outs = 10;
// Create layout with same sizes but new format.
layout create_reordering_layout(format new_format, const layout& src_layout)
return { src_layout.data_type, new_format, src_layout.size };
// Create MNIST topology
topology create_topology(const layout& in_layout, const memory& conv1_weights_mem, const memory& conv1_bias_mem )
auto data_type = in_layout.data_type;
// Create input_layout description
// "input" - is the primitive id inside topology
input_layout input("input", in_layout);
// Create topology object with 2 primitives
// 1. input layout primitive.
// 2. reorder primitive with id "reorder_input"
// input primitive for reorder (implicitly converted to primitive_id)
// output layout for reorder
create_reordering_layout(format::yxfb, in_layout))
// Create data primitive - its content should be set already.
cldnn::data conv1_weights( "conv1_weights", conv1_weights_mem );
// Add primitive to topology
// Emplace new primitive to topology
topology.add<cldnn::data>({ "conv1_bias", conv1_bias_mem });
// Emplace 2 primitives
// Convolution primitive with id "conv1"
"reorder_input", // primitive id of the convolution's input
{ conv1_weights }, // weights primitive id is taken from the object
{ "conv1_bias" } // bias primitive id
// Pooling id: "pool1"
"conv1", // Input: "conv1"
pooling_mode::max, // Pooling mode: MAX
spatial(2,2), // stride: 2
spatial(2,2) // kernel_size: 2
// Conv2 weights data is not available now, so just declare its layout
layout conv2_weights_layout(data_type, format::bfyx,{ conv2_out_channels, conv1_out_channels, conv_krnl_size, conv_krnl_size });
// Define the rest of topology.
// Input layout for conv2 weights. Data will passed by network::set_input_data()
input_layout("conv2_weights", conv2_weights_layout),
// Input layout for conv2 bias.
input_layout("conv2_bias", { data_type, format::bfyx, spatial(conv2_out_channels) }),
// Second convolution id: "conv2"
"pool1", // Input: "pool1"
{ "conv2_weights" }, // Weights: input_layout "conv2_weights"
{ "conv2_bias" } // Bias: input_layout "conv2_bias"
// Second pooling id: "pool2"
"conv2", // Input: "conv2"
pooling_mode::max, // Pooling mode: MAX
spatial(2, 2), // stride: 2
spatial(2, 2) // kernel_size: 2
// Fully connected (inner product) primitive id "fc1"
"pool2", // Input: "pool2"
"fc1_weights", // "fc1_weights" will be added to the topology later
"fc1_bias", // will be defined later
true // Use built-in Relu. Slope is set to 0 by default.
// Second FC/IP primitive id: "fc2", input: "fc1".
// Weights ("fc2_weights") and biases ("fc2_bias") will be defined later.
// Built-in Relu is disabled by default.
fully_connected("fc2", "fc1", "fc2_weights", "fc2_bias"),
// The "softmax" primitive is not an input for any other,
// so it will be automatically added to network outputs.
softmax("softmax", "fc2")
return topology;
// Copy from a vector to cldnn::memory
void copy_to_memory(memory& mem, const vector<float>& src)
std::copy(src.begin(), src.end(), dst.begin());
// Execute network
int recognize_image(network& network, const memory& input_memory)
// Set/update network input
network.set_input_data("input", input_memory);
// Start network execution
// get_memory() blocks output generation completed
auto output ="softmax").get_memory();
// Get direct access to output memory
cldnn::pointer<float> out_ptr(output);
// Analyze result
auto max_element_pos = max_element(out_ptr.begin(), out_ptr.end());
return static_cast<int>(distance(out_ptr.begin(), max_element_pos));
// User-defined helpers which are out of this example scope
// //////////////////////////////////////////////////////////////
// Loads file to a vector of floats.
vector<float> load_data(const string&) { return{ 0 }; }
// Allocates memory and loads data from file.
// Memory layout is taken from file.
memory load_mem(const engine& eng, const string&) {
//return a dummy value
return memory::allocate(eng, layout{ data_types::f32, format::bfyx, { 1, 1, 1, 1 } });
// Load image, resize to [x,y] and store in a vector of floats
// in the order "bfyx".
vector<float> load_image_bfyx(const string&, int, int) { return{ 0 }; }
// //////////////////////////////////////////////////////////////
int main_func()
// Use data type: float
// Network input layout
layout in_layout(
data_type, // stored data type
format::bfyx, // data stored in order batch-channel-Y-X, where X coordinate changes first.
{1, input_channels, input_size, input_size} // batch: 1, channels: 1, Y: 28, X: 28
// Create memory for conv1 weights
layout conv1_weights_layout(data_type, format::bfyx,{ conv1_out_channels, input_channels, conv_krnl_size, conv_krnl_size });
vector<float> my_own_buffer = load_data("conv1_weights.bin");
// The conv1_weights_mem is attached to my_own_buffer, so my_own_buffer should not be changed or descroyed until network execution completion.
auto conv1_weights_mem = memory::attach(conv1_weights_layout,, my_own_buffer.size());
// Create default engine
// Create memory for conv1 bias
layout conv1_bias_layout(data_type, format::bfyx, spatial(20));
// Memory allocation requires engine
auto conv1_bias_mem = memory::allocate(engine, conv1_bias_layout);
// The memory is allocated by library, so we do not need to care about buffer lifetime.
copy_to_memory(conv1_bias_mem, load_data("conv1_bias.bin"));
// Get new topology
cldnn::topology topology = create_topology(in_layout, conv1_weights_mem, conv1_bias_mem);
// Define network data not defined in create_topology()
cldnn::data("fc1_weights", load_mem(engine, "")),
cldnn::data("fc1_bias", load_mem(engine, "")),
cldnn::data("fc2_weights", load_mem(engine, "")),
cldnn::data("fc2_bias", load_mem(engine, ""))
// Build the network. Allow implicit data optimizations.
// The "softmax" primitive is not used as an input for other primitives,
// so we do not need to explicitly select it in build_options::outputs()
// Set network data which was not known at topology creation.
network.set_input_data("conv2_weights", load_mem(engine, ""));
network.set_input_data("conv2_bias", load_mem(engine, ""));
// Allocate memory for input image.
auto input_memory = memory::allocate(engine, in_layout);
// Run network 2 times with different images.
for (auto img_name : { "one.jpg", "two.jpg" })
// Reuse image memory.
copy_to_memory(input_memory, load_image_bfyx("one.jpg", in_layout.size.spatial[0], in_layout.size.spatial[1]));
auto result = recognize_image(network, input_memory);
cout << img_name << " recognized as" << result << endl;
return 0;