Merge branch 'main' into cmake-arch-cleanup

Author: allnes

app/Graph/build.cpp

@@ -101,7 +101,8 @@ void build_graph_linear(it_lab_ai::Graph& graph, it_lab_ai::Tensor& input,
                   << '\n';
       }
       auto pool_layer =
-          std::make_shared<it_lab_ai::PoolingLayer>(shape, pooltype);
+          LayerFactory::createPoolingLayer(pooltype, shape, options);
+
       layers.push_back(pool_layer);
       layerpostop.push_back(false);
       if (comments) std::cout << "PoolingLayer added to layers." << '\n';
@@ -408,8 +409,8 @@ ParseResult parse_json_model(RuntimeOptions options,
               << '\n';
         }
       } else if (layer_type == "GlobalAveragePool") {
-        auto pool_layer = std::make_shared<it_lab_ai::PoolingLayer>(
-            it_lab_ai::Shape({0, 0}), "average");
+        auto pool_layer = LayerFactory::createPoolingLayer(
+            "average", it_lab_ai::Shape({0, 0}), options);
         layer = pool_layer;
         if (comments) {
           std::cout << "GlobalAveragePool layer added (will use input spatial "
@@ -470,30 +471,9 @@ ParseResult parse_json_model(RuntimeOptions options,
           }
         }
 
-        auto pool_layer =
-            std::make_shared<it_lab_ai::PoolingLayer>(shape, pooltype);
-
-        try {
-          if (strides[0] != 2 || strides[1] != 2) {
-            pool_layer->setStrides(strides[0], strides[1]);
-          }
-
-          if (pads[0] != 0 || pads[1] != 0 || pads[2] != 0 || pads[3] != 0) {
-            pool_layer->setPads(pads[0], pads[1], pads[2], pads[3]);
-          }
+        auto pool_layer = LayerFactory::createPoolingLayer(
+            pooltype, shape, options, strides, pads, dilations, ceil_mode);
 
-          if (dilations[0] != 1 || dilations[1] != 1) {
-            pool_layer->setDilations(dilations[0], dilations[1]);
-          }
-
-          pool_layer->setCeilMode(ceil_mode);
-
-        } catch (const std::exception& e) {
-          if (comments) {
-            std::cout << "Warning: Some pooling parameters not supported: "
-                      << e.what() << '\n';
-          }
-        }
         layer = pool_layer;
       } else if (layer_type.find("Flatten") != std::string::npos) {
         int axis = 1;

app/Graph/build.hpp

@@ -34,6 +34,7 @@
 #include "layers/TransposeLayer.hpp"
 #include "layers_oneDNN/ConvLayer.hpp"
 #include "layers_oneDNN/EWLayer.hpp"
+#include "layers_oneDNN/PoolingLayer.hpp"
 
 extern std::unordered_map<std::string, std::string> model_paths;
 
@@ -99,6 +100,19 @@ class LayerFactory {
     return std::make_shared<ConvolutionalLayer>(step, pads, dilations, kernel,
                                                 bias, group, useLegacyImpl);
   }
+
+  static std::shared_ptr<Layer> createPoolingLayer(
+      const std::string& PoolType, const Shape& shape,
+      const RuntimeOptions& options, const Shape& strides = {2, 2},
+      const Shape& pads = {0, 0, 0, 0}, const Shape& dilations = {1, 1},
+      bool ceil_mode = false) {
+    if (options.backend == Backend::kOneDnn) {
+      return std::make_shared<PoolingLayerOneDnn>(
+          shape, strides, pads, dilations, ceil_mode, PoolType);
+    }
+    return std::make_shared<PoolingLayer>(shape, strides, pads, dilations,
+                                          ceil_mode, PoolType);
+  }
 };
 
 }  // namespace it_lab_ai

include/layers/Shape.hpp

@@ -5,6 +5,7 @@
 #include <iostream>
 #include <numeric>
 #include <ostream>
+#include <sstream>
 #include <stdexcept>
 #include <vector>
 
@@ -39,6 +40,16 @@ class Shape {
   }
   [[nodiscard]] size_t dims() const noexcept { return dims_.size(); }
   [[nodiscard]] size_t get_index(const std::vector<size_t>& coords) const;
+  [[nodiscard]] std::string to_string() const {
+    std::stringstream ss;
+    ss << "(";
+    for (size_t i = 0; i < dims_.size(); ++i) {
+      if (i > 0) ss << ", ";
+      ss << dims_[i];
+    }
+    ss << ")";
+    return ss.str();
+  }
   bool operator==(const Shape& other) const {
     if (dims_.size() != other.dims_.size()) return false;
     for (size_t i = 0; i < dims_.size(); ++i) {

include/layers_oneDNN/PoolingLayer.hpp

@@ -0,0 +1,87 @@
+#pragma once
+
+#include <dnnl.hpp>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "layers/Layer.hpp"
+
+namespace it_lab_ai {
+
+class PoolingLayerOneDnn : public Layer {
+ public:
+  explicit PoolingLayerOneDnn(const Shape& pooling_shape,
+                              const Shape& strides = {2, 2},
+                              const Shape& pads = {0, 0, 0, 0},
+                              const Shape& dilations = {1, 1},
+                              bool ceil_mode = false,
+                              std::string pooling_type = "average")
+      : Layer(kPooling),
+        poolingShape_(pooling_shape),
+        strides_(strides),
+        pads_(pads),
+        dilations_(dilations),
+        ceil_mode_(ceil_mode),
+        poolingType_(std::move(pooling_type)),
+        engine_(std::make_unique<dnnl::engine>(dnnl::engine::kind::cpu, 0)),
+        stream_(std::make_unique<dnnl::stream>(*engine_)) {}
+
+  void run(const std::vector<Tensor>& input,
+           std::vector<Tensor>& output) override;
+
+  void setStrides(size_t h, size_t w) {
+    strides_ = {h, w};
+    initialized_ = false;
+  }
+
+  void setPads(size_t top, size_t bottom, size_t left, size_t right) {
+    pads_ = {top, bottom, left, right};
+    initialized_ = false;
+  }
+
+  void setDilations(size_t h, size_t w) {
+    dilations_ = {h, w};
+    initialized_ = false;
+  }
+
+  void setCeilMode(bool ceil_mode) {
+    ceil_mode_ = ceil_mode;
+    initialized_ = false;
+  }
+
+#ifdef ENABLE_STATISTIC_WEIGHTS
+  Tensor get_weights() override {
+    std::vector<int> v = {0};
+    Tensor a = make_tensor(v);
+    return a;
+  }
+#endif
+
+ private:
+  void initialize_onednn(const Shape& shape, Type data_type);
+  [[nodiscard]] dnnl::algorithm get_PoolType() const;
+  static void validate_input(const std::vector<Tensor>& input);
+  [[nodiscard]] static dnnl::memory::data_type get_dnnl_data_type(Type type);
+  [[nodiscard]] Shape calculate_output_shape(const Shape& input_shape) const;
+
+  Shape poolingShape_;
+  Shape strides_;
+  Shape pads_;
+  Shape dilations_;
+  bool ceil_mode_;
+  std::string poolingType_;
+
+  bool initialized_ = false;
+  Shape last_shape_;
+  Type last_type_;
+
+  std::unique_ptr<dnnl::engine> engine_;
+  std::unique_ptr<dnnl::stream> stream_;
+  std::unique_ptr<dnnl::pooling_forward> pool_prim_;
+  dnnl::memory::desc src_memory_desc_;
+  dnnl::memory::desc dst_memory_desc_;
+  Shape output_shape_;
+};
+
+}  // namespace it_lab_ai

src/layers_oneDNN/ConvLayer.cpp

@@ -369,8 +369,9 @@ void ConvLayerOneDnn::initialize_special_conv(const Shape& input_shape,
     if (has_bias) {
       bias_md = dnnl::memory::desc(
           {static_cast<dnnl::memory::dim>(bias_.get_shape()[0])}, dt,
-          dnnl::memory::format_tag::any);
+          dnnl::memory::format_tag::a);
     }
+
     dnnl::convolution_forward::primitive_desc conv_pd =
         has_bias ? dnnl::convolution_forward::primitive_desc(
                        *engine_, dnnl::prop_kind::forward_inference,
@@ -391,20 +392,21 @@ void ConvLayerOneDnn::initialize_special_conv(const Shape& input_shape,
 
     if (data_type == Type::kFloat) {
       const std::vector<float>& kernel_data = *kernel_.as<float>();
+
       size_t kh = k_shape[0];
       size_t kw = k_shape[1];
       size_t kic = k_shape[2];
       size_t koc = k_shape[3];
 
       std::vector<float> reordered(koc * kic * kh * kw);
-      size_t idx = 0;
 
-      for (size_t oc = 0; oc < koc; oc++) {
-        for (size_t ic = 0; ic < kic; ic++) {
-          for (size_t h = 0; h < kh; h++) {
-            for (size_t w = 0; w < kw; w++) {
+      for (size_t oc = 0; oc < koc; ++oc) {
+        for (size_t ic = 0; ic < kic; ++ic) {
+          for (size_t h = 0; h < kh; ++h) {
+            for (size_t w = 0; w < kw; ++w) {
               size_t src_idx = ((h * kw + w) * kic + ic) * koc + oc;
-              reordered[idx++] = kernel_data[src_idx];
+              size_t dst_idx = ((oc * kic + ic) * kh + h) * kw + w;
+              reordered[dst_idx] = kernel_data[src_idx];
             }
           }
         }
@@ -421,14 +423,13 @@ void ConvLayerOneDnn::initialize_special_conv(const Shape& input_shape,
       size_t koc = k_shape[3];
 
       std::vector<float> reordered(koc * kic * kh * kw);
-      size_t idx = 0;
-
-      for (size_t oc = 0; oc < koc; oc++) {
-        for (size_t ic = 0; ic < kic; ic++) {
-          for (size_t h = 0; h < kh; h++) {
-            for (size_t w = 0; w < kw; w++) {
+      for (size_t oc = 0; oc < koc; ++oc) {
+        for (size_t ic = 0; ic < kic; ++ic) {
+          for (size_t h = 0; h < kh; ++h) {
+            for (size_t w = 0; w < kw; ++w) {
               size_t src_idx = ((h * kw + w) * kic + ic) * koc + oc;
-              reordered[idx++] = static_cast<float>(kernel_data_int[src_idx]);
+              size_t dst_idx = ((oc * kic + ic) * kh + h) * kw + w;
+              reordered[dst_idx] = static_cast<float>(kernel_data_int[src_idx]);
             }
           }
         }
@@ -438,6 +439,22 @@ void ConvLayerOneDnn::initialize_special_conv(const Shape& input_shape,
                   reordered.size() * sizeof(float));
     }
 
+    if (has_bias) {
+      if (data_type == Type::kFloat) {
+        const std::vector<float>& bias_data = *bias_.as<float>();
+        std::memcpy(bias_memory_.get_data_handle(), bias_data.data(),
+                    bias_data.size() * sizeof(float));
+      } else if (data_type == Type::kInt) {
+        const std::vector<int>& bias_data_int = *bias_.as<int>();
+        std::vector<float> float_bias(bias_data_int.size());
+        std::transform(bias_data_int.begin(), bias_data_int.end(),
+                       float_bias.begin(),
+                       [](int val) { return static_cast<float>(val); });
+        std::memcpy(bias_memory_.get_data_handle(), float_bias.data(),
+                    float_bias.size() * sizeof(float));
+      }
+    }
+
     conv_prim_ = std::make_unique<dnnl::convolution_forward>(conv_pd);
     initialized_ = true;
 
@@ -494,8 +511,8 @@ void ConvLayerOneDnn::run_special_conv(const std::vector<Tensor>& input,
   Shape output_shape = get_output_shape(input_shape);
 
   if (data_type == Type::kFloat) {
-    std::vector<float> output_data(dst_memory_.get_desc().get_size() /
-                                   sizeof(float));
+    size_t output_size = dst_memory_.get_desc().get_size() / sizeof(float);
+    std::vector<float> output_data(output_size);
     std::memcpy(output_data.data(), dst_memory_.get_data_handle(),
                 output_data.size() * sizeof(float));
     output[0] = make_tensor(output_data, output_shape);