#include <algorithm>
#include <cmath>
#include <iostream>
#include <string>
#include <vector>
#include "example_utils.hpp"
using tag = memory::format_tag;
using dt = memory::data_type;
const memory::dim N = 26,
T = 6,
C = 12,
G = 4,
L = 4,
D = 1;
memory::dims src_dims = {T, N, C};
memory::dims weights_dims = {L, D, C, G, C};
memory::dims bias_dims = {L, D, G, C};
memory::dims dst_dims = {T, N, C};
std::vector<float> src_layer_data(product(src_dims));
std::vector<float> weights_layer_data(product(weights_dims));
std::vector<float> weights_iter_data(product(weights_dims));
std::vector<float> dst_layer_data(product(dst_dims));
std::vector<float> bias_data(product(bias_dims));
std::generate(src_layer_data.begin(), src_layer_data.end(), []() {
static int i = 0;
return std::cos(i++ / 10.f);
});
std::generate(weights_layer_data.begin(), weights_layer_data.end(), []() {
static int i = 0;
return std::sin(i++ * 2.f);
});
std::generate(bias_data.begin(), bias_data.end(), []() {
static int i = 0;
return std::tanh(i++);
});
auto src_layer_md = memory::desc(src_dims, dt::f32, tag::tnc);
auto bias_md = memory::desc(bias_dims, dt::f32, tag::ldgo);
auto dst_layer_md = memory::desc(dst_dims, dt::f32, tag::tnc);
auto src_layer_mem = memory(src_layer_md, engine);
auto bias_mem = memory(bias_md, engine);
auto dst_layer_mem = memory(dst_layer_md, engine);
auto user_weights_layer_mem
= memory({weights_dims, dt::f32, tag::ldigo},
engine);
auto user_weights_iter_mem
= memory({weights_dims, dt::f32, tag::ldigo},
engine);
write_to_dnnl_memory(src_layer_data.data(), src_layer_mem);
write_to_dnnl_memory(bias_data.data(), bias_mem);
write_to_dnnl_memory(weights_layer_data.data(), user_weights_layer_mem);
write_to_dnnl_memory(weights_iter_data.data(), user_weights_iter_mem);
auto lstm_weights_layer_md = memory::desc(weights_dims, dt::f32, tag::any);
auto lstm_weights_iter_md = memory::desc(weights_dims, dt::f32, tag::any);
auto src_iter_md = memory::desc();
auto src_iter_c_md = memory::desc();
auto dst_iter_md = memory::desc();
auto dst_iter_c_md = memory::desc();
auto lstm_desc = lstm_forward::desc(prop_kind::forward_training,
rnn_direction::unidirectional_left2right, src_layer_md, src_iter_md,
src_iter_c_md, lstm_weights_layer_md, lstm_weights_iter_md, bias_md,
dst_layer_md, dst_iter_md, dst_iter_c_md);
auto lstm_pd = lstm_forward::primitive_desc(lstm_desc, engine);
auto lstm_weights_layer_mem = user_weights_layer_mem;
auto lstm_weights_iter_mem = user_weights_iter_mem;
if (lstm_pd.weights_desc() != user_weights_layer_mem.get_desc()) {
lstm_weights_layer_mem = memory(lstm_pd.weights_desc(), engine);
reorder(user_weights_layer_mem, lstm_weights_layer_mem)
.execute(engine_stream, user_weights_layer_mem,
lstm_weights_layer_mem);
}
if (lstm_pd.weights_iter_desc() != user_weights_iter_mem.get_desc()) {
lstm_weights_iter_mem = memory(lstm_pd.weights_iter_desc(), engine);
reorder(user_weights_iter_mem, lstm_weights_iter_mem)
.execute(engine_stream, user_weights_iter_mem,
lstm_weights_iter_mem);
}
auto src_iter_mem = memory(lstm_pd.src_iter_desc(), engine);
auto src_iter_c_mem = memory(lstm_pd.src_iter_c_desc(), engine);
auto weights_iter_mem = memory(lstm_pd.weights_iter_desc(), engine);
auto dst_iter_mem = memory(lstm_pd.dst_iter_desc(), engine);
auto dst_iter_c_mem = memory(lstm_pd.dst_iter_c_desc(), engine);
auto workspace_mem = memory(lstm_pd.workspace_desc(), engine);
auto lstm_prim = lstm_forward(lstm_pd);
std::unordered_map<int, memory> lstm_args;
lstm_prim.execute(engine_stream, lstm_args);
engine_stream.wait();
read_from_dnnl_memory(dst_layer_data.data(), dst_layer_mem);
}
int main(int argc, char **argv) {
return handle_example_errors(lstm_example, parse_engine_kind(argc, argv));
}
#define DNNL_ARG_DST_ITER
A special mnemonic for RNN input recurrent hidden state vector.
Definition: dnnl_types.h:2318
#define DNNL_ARG_WEIGHTS_LAYER
A special mnemonic for RNN weights applied to the layer input.
Definition: dnnl_types.h:2336
#define DNNL_ARG_WORKSPACE
Workspace tensor argument.
Definition: dnnl_types.h:2366
#define DNNL_ARG_WEIGHTS_ITER
A special mnemonic for RNN weights applied to the recurrent input.
Definition: dnnl_types.h:2342
#define DNNL_ARG_DST_ITER_C
A special mnemonic for LSTM output recurrent cell state vector.
Definition: dnnl_types.h:2324
#define DNNL_ARG_SRC_ITER_C
A special mnemonic for RNN input recurrent cell state vector.
Definition: dnnl_types.h:2301
#define DNNL_ARG_SRC_LAYER
A special mnemonic for RNN input vector.
Definition: dnnl_types.h:2286
#define DNNL_ARG_DST_LAYER
A special mnemonic for RNN output vector. An alias for DNNL_ARG_DST_0.
Definition: dnnl_types.h:2312
#define DNNL_ARG_BIAS
Bias tensor argument.
Definition: dnnl_types.h:2357
#define DNNL_ARG_SRC_ITER
A special mnemonic for RNN input recurrent hidden state vector.
Definition: dnnl_types.h:2295
oneDNN namespace
Definition: dnnl.hpp:74
An execution engine.
Definition: dnnl.hpp:869
kind
Kinds of engines.
Definition: dnnl.hpp:874
An execution stream.
Definition: dnnl.hpp:985