[GPU] Add INT32/UINT32 to available input data types when load type is aligned in GetJitLoad (#9300) (#9300)

- Modify fusibility checking to allow sub/div eltwise fusing for other primitives
- Modify dump checking code to use node name in exec graph

inference-engine/thirdparty/clDNN/api/intel_gpu/runtime/debug_configuration.hpp

@@ -5,6 +5,7 @@
 #pragma once
 #include <string>
 #include <mutex>
+#include <vector>
 
 #ifdef GPU_DEBUG_CONFIG
 #define GPU_DEBUG_IF(cond) if (cond)
@@ -32,12 +33,13 @@ public:
     std::string dump_graphs;        // Dump optimized graph
     std::string dump_sources;       // Dump opencl sources
     std::string dump_layers_path;   // Enable dumping intermediate buffers and set the dest path
-    std::string dump_layers;        // Dump intermediate buffers of specified layers only, separated by space
+    std::vector<std::string> dump_layers;        // Dump intermediate buffers of specified layers only
     std::string dry_run_path;       // Dry run and serialize execution graph into the specified path
     int dump_layers_dst_only;       // Dump only output of layers
     int dump_layers_limit_batch;    // Limit the size of batch to dump
     int base_batch_for_memory_estimation; // Base batch size to be used in memory estimation
     static const debug_configuration *get_instance();
+    bool is_dumped_layer(const std::string& layerName) const;
 };
 
 }  // namespace cldnn

inference-engine/thirdparty/clDNN/kernel_selector/core/common/jitter.cpp

@@ -1848,7 +1848,7 @@ std::string FusedOpsCodeGenerator::GetJitLoad(const FusedOpsConfiguration& conf,
             std::string vs = vec_size > 1 ? toCodeString(vec_size)  : "";
             std::string block_read;
 
-            if (input_dt == Datatype::F32) {
+            if (input_dt == Datatype::F32 || input_dt == Datatype::INT32 || input_dt == Datatype::UINT32) {
                 block_read = CastToType(" intel_sub_group_block_read" + vs + "("
                                         + "(const __global uint*)(" + GetInputPtrName(input_id) + " + " + index_func_call_vec + "))",
                                         input_dt, vec_size);

inference-engine/thirdparty/clDNN/runtime/debug_configuration.cpp

@@ -137,7 +137,6 @@ debug_configuration::debug_configuration()
         , dump_graphs(std::string())
         , dump_sources(std::string())
         , dump_layers_path(std::string())
-        , dump_layers(std::string())
         , dump_layers_dst_only(0)
         , dry_run_path(std::string())
         , disable_onednn(0)
@@ -151,20 +150,27 @@ debug_configuration::debug_configuration()
     get_gpu_debug_env_var("DumpGraphs", dump_graphs);
     get_gpu_debug_env_var("DumpSources", dump_sources);
     get_gpu_debug_env_var("DumpLayersPath", dump_layers_path);
-    get_gpu_debug_env_var("DumpLayers", dump_layers);
     get_gpu_debug_env_var("DumpLayersDstOnly", dump_layers_dst_only);
     get_gpu_debug_env_var("DumpLayersLimitBatch", dump_layers_limit_batch);
     get_gpu_debug_env_var("DisableOnednn", disable_onednn);
     get_gpu_debug_env_var("DryRunPath", dry_run_path);
     get_gpu_debug_env_var("BaseBatchForMemEstimation", base_batch_for_memory_estimation);
+    std::string dump_layers_str;
+    get_gpu_debug_env_var("DumpLayers", dump_layers_str);
 
     if (help > 0) {
         print_help_messages();
         exit(0);
     }
 
-    if (dump_layers.length() > 0)
+    if (dump_layers_str.length() > 0) {
-        dump_layers = " " + dump_layers + " "; // Insert delimiter for easier parsing when used
+        dump_layers_str = " " + dump_layers_str + " "; // Insert delimiter for easier parsing when used
+        std::stringstream ss(dump_layers_str);
+        std::string layer;
+        while (ss >> layer) {
+            dump_layers.push_back(layer);
+        }
+    }
 #endif
 }
 
@@ -180,4 +186,16 @@ const debug_configuration *debug_configuration::get_instance() {
     return nullptr;
 #endif
 }
+
+bool debug_configuration::is_dumped_layer(const std::string& layerName) const {
+#ifdef GPU_DEBUG_CONFIG
+    if (dump_layers.empty()) return true;
+    auto iter = std::find_if(dump_layers.begin(), dump_layers.end(), [&](const std::string& dl){
+        return (layerName.find(dl) != std::string::npos);
+    });
+    return (iter != dump_layers.end());
+#else
+    return false;
+#endif
+}
 } // namespace cldnn

inference-engine/thirdparty/clDNN/src/graph_optimizer/prepare_primitive_fusing.cpp

@@ -940,8 +940,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program &p) {
 
             for (size_t i = 0; i < parents.size(); i++) {
                 can_fuse_parents[i] = (parents[i]->is_type<convolution>() && conv_supports_fusings(parents[i]->as<convolution>())) ||
-                                      ((prim->mode == eltwise_mode::sum || prim->mode == eltwise_mode::prod) &&
+                                      (parents[i]->is_type<binary_convolution>() && bin_conv_supports_eltw_fusings(parents[i]->as<binary_convolution>())) ||
-                                      ((parents[i]->is_type<binary_convolution>() && bin_conv_supports_eltw_fusings(parents[i]->as<binary_convolution>())) ||
                                       (parents[i]->is_type<mvn>() && mvn_supports_fusings(parents[i]->as<mvn>())) ||
                                       (parents[i]->is_type<deconvolution>()) ||
                                       (parents[i]->is_type<permute>()) ||
@@ -959,7 +958,7 @@ void prepare_primitive_fusing::fuse_simple_primitives(program &p) {
                                       (parents[i]->is_type<scatter_elements_update>()) ||
                                       (parents[i]->is_type<pooling>() && pooling_supports_fusings(parents[i]->as<pooling>())) ||
                                       (parents[i]->is_type<depth_to_space>() && dts_supports_fusings(parents[i]->as<depth_to_space>())) ||
-                                      (parents[i]->is_type<reduce>() && reduce_supports_fusings(parents[i]->as<reduce>()))));
+                                      (parents[i]->is_type<reduce>() && reduce_supports_fusings(parents[i]->as<reduce>()));
             }
 
             // Disable fusion to a node on constant path when second input is in data flow

inference-engine/thirdparty/clDNN/src/network.cpp

@@ -178,6 +178,7 @@ void dump<uint32_t>(memory::ptr mem, stream& stream, std::ofstream& file_stream)
 }
 
 static void log_memory_to_file(memory::ptr mem, stream& stream, std::string layerName) {
+    std::cout << "Dump " << layerName << std::endl;
     GPU_DEBUG_GET_INSTANCE(debug_config);
     std::string filename = layerName;
     std::replace(filename.begin(), filename.end(), '\\', '_');
@@ -208,6 +209,7 @@ static void log_memory_to_file(memory::ptr mem, stream& stream, std::string laye
     (void)layerName;
 }
 #endif
+
 /*
 Network will always have net_id = 0 when it will be cldnn internal micronetwork (created i.e by propagate_constants
 opt pass).
@@ -627,6 +629,7 @@ std::map<primitive_id, network_output> network::execute(const std::vector<event:
     return result;
 }
 
+
 void network::execute_impl(const std::vector<event::ptr>& events) {
     OV_ITT_SCOPED_TASK(itt::domains::CLDNN, "NetworkImpl::Execute");
     // Wait for previous execution completion
@@ -647,19 +650,16 @@ void network::execute_impl(const std::vector<event::ptr>& events) {
     auto surf_lock = surfaces_lock::create(get_engine().type(), in_out_mem, get_stream());
 
     set_arguments();
-
     for (auto& inst : _exec_order) {
         GPU_DEBUG_IF(debug_config->dump_layers_path.length() > 0) {
             auto& node = _program->get_node(inst->id());
-            std::string layer_name = node.id();
+            const std::string layer_name = node.id();
             GPU_DEBUG_IF(debug_config->verbose >= 2) {
                 std::cerr << get_primitive_info(inst->id()) << std::endl;
             }
 
             GPU_DEBUG_IF(debug_config->dump_layers_dst_only == 0 &&
-                            (debug_config->dump_layers.length() == 0 ||
+                            debug_config->is_dumped_layer(layer_name)) {
-                            (debug_config->dump_layers.length() != 0 && debug_config->dump_layers.find(" " + layer_name + " ") != std::string::npos))) {
-                std::cout << "Dump " << layer_name << " layer src" << std::endl;
                 for (size_t i = 0; i < get_primitive(inst->id())->dependencies().size(); i++) {
                     log_memory_to_file(get_primitive(inst->id())->dep_memory_ptr(i), get_stream(),
                                     layer_name + "_src_" + std::to_string(i));
@@ -682,10 +682,8 @@ void network::execute_impl(const std::vector<event::ptr>& events) {
         GPU_DEBUG_IF(debug_config->dump_layers_path.length() > 0) {
             get_stream().finish();
             auto& node = _program->get_node(inst->id());
-            std::string layer_name = node.id();
+            const std::string layer_name = node.id();
-            GPU_DEBUG_IF(debug_config->dump_layers.length() == 0 ||
+            GPU_DEBUG_IF(debug_config->is_dumped_layer(layer_name)) {
-                        (debug_config->dump_layers.length() != 0 && debug_config->dump_layers.find(" " + layer_name + " ") != std::string::npos)) {
-                std::cout << "Dump " << layer_name << " layer dst" << std::endl;
                 log_memory_to_file(get_primitive(inst->id())->output_memory_ptr(), get_stream(), layer_name + "_dst_0");
             }
         }