Address solution proposal

src/common/snippets/docs/debug_capabilities/parameters_dump.md

@@ -14,13 +14,13 @@ Examples:
 
 Output example:
 ```csv
-name,in_type,out_type,in_shapes,out_shapes,M,N,K,m_block,n_block,k_block
-i8;i8;f32,i32,1 128 16 64;1 16 64 128;1 16 64 128,1 16 128 128,128,128,64,32,-2,-2
-u8;i8,i32,1 16 128 128;1 16 128 64,1 16 128 64,128,64,128,32,-2,-2
-i8;i8;f32,i32,2 68 6 92;2 6 92 68;2 6 92 68,2 6 68 68,68,68,92,32,-2,-2
-u8;i8,i32,2 6 68 68;2 6 68 92,2 6 68 92,68,92,68,32,-2,-2
-i8;i8;f32,i32,? ? ? 100;? ? 100 128;? ? 100 128,? ? ? 128,-1,128,100,32,-2,-2
-u8;i8,i32,? ? ? 128;? ? 128 100,? ? ? 100,-1,100,128,32,-2,-2
-i8;i8;f32,i32,? ? ? ?;? ? ? ?;? ? ? ?,? ? ? ?,-1,-1,-1,32,-2,-2
-u8;i8,i32,? ? ? ?;? ? ? ?,? ? ? ?,-1,-1,-1,32,-2,-2
+name,in_type,out_type,in_shapes,out_shapes,M,N,K,m_block,n_block,k_block,acc_max_time,avg_max_time
+MatMul_437,i8;i8;f32,i32,1 128 16 64;1 16 64 128;1 16 64 128,1 16 128 128,128,128,64,32,-2,-2,157430,19678
+MatMul_451,u8;i8,i32,1 16 128 128;1 16 128 64,1 16 128 64,128,64,128,32,-2,-2,96222,12027
+MatMul_6069,i8;i8;f32,i32,2 68 6 92;2 6 92 68;2 6 92 68,2 6 68 68,68,68,92,32,-2,-2,1734,866
+MatMul_6083,u8;i8,i32,2 6 68 68;2 6 68 92,2 6 68 92,68,92,68,32,-2,-2,1929,1929
+MatMul_12488,i8;i8;f32,i32,? ? ? 100;? ? 100 128;? ? 100 128,? ? ? 128,-1,128,100,32,-2,-2,43105,43105
+MatMul_12502,u8;i8,i32,? ? ? 128;? ? 128 100,? ? ? 100,-1,100,128,32,-2,-2,3328,2801
+MatMul_18789,i8;i8;f32,i32,? ? ? ?;? ? ? ?;? ? ? ?,? ? ? ?,-1,-1,-1,32,-2,-2,6135,6135
+MatMul_18803,u8;i8,i32,? ? ? ?;? ? ? ?,? ? ? ?,-1,-1,-1,32,-2,-2,2742,1371
 ```

src/common/snippets/include/snippets/lowered/pass/brgemm_debug_params.hpp

@@ -2,9 +2,9 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
-#pragma once
+#ifdef SNIPPETS_DEBUG_CAPS
 
-#include <fstream>
+#pragma once
 
 #include "snippets/itt.hpp"
 #include "snippets/lowered/loop_manager.hpp"
@@ -51,78 +51,90 @@ class BrgemmDebugParams : public snippets::lowered::pass::RangedPass, public Brg
              snippets::lowered::LinearIR::constExprIt begin,
              snippets::lowered::LinearIR::constExprIt end) override final {  // NOLINT
         OV_ITT_SCOPED_TASK(ov::pass::itt::domains::SnippetsTransform, "Snippets::BrgemmDebugParams")
-#ifdef SNIPPETS_DEBUG_CAPS
+        static size_t seq_number = 0;
         for (auto expr_it = begin; expr_it != end; expr_it++) {
             const auto& brgemm_expr = *expr_it;
             const auto brgemm = ov::as_type_ptr<BRGEMM_TYPE>(brgemm_expr->get_node());
-            if (brgemm) {
-                dump_params_to_csv(brgemm_expr, linear_ir);
-            }
+            if (!brgemm)
+                continue;
+            // Collect brgemm parameters
+            auto params = collect_params(brgemm_expr, linear_ir);
+            const auto& perf_count_begin = std::make_shared<snippets::op::PerfCountBegin>();
+            perf_count_begin->set_friendly_name(std::string("PerfCount_Begin_") + std::to_string(seq_number) +
+                                                "_DebugParams");
+            const auto empty_inputs = std::vector<PortConnectorPtr>{};
+            linear_ir.insert_node(perf_count_begin, empty_inputs, expr_it->get()->get_loop_ids(), false, expr_it);
+
+            const auto& perf_count_end = std::make_shared<snippets::op::PerfCountEnd>(perf_count_begin->output(0));
+            perf_count_end->set_friendly_name(std::string("PerfCount_End_") + std::to_string(seq_number) +
+                                              "_DebugParams");
+            // Attach brgemm parameters to PerfCountEnd node
+            auto csv_path = linear_ir.get_config().debug_config.dumpParams.csv_path;
+            perf_count_end->get_rt_info()["brgemm_params"] = params;
+            perf_count_end->get_rt_info()["brgemm_params_csv_path"] = csv_path;
+            linear_ir.insert_node(perf_count_end, empty_inputs, expr_it->get()->get_loop_ids(), false, next(expr_it));
+            seq_number++;
         }
-#endif  // SNIPPETS_DEBUG_CAPS
         return false;
     }
 
-#ifdef SNIPPETS_DEBUG_CAPS
-    void dump_params_to_csv(const ov::snippets::lowered::ExpressionPtr& brgemm_expr,
-                            const snippets::lowered::LinearIR& linear_ir) {
+private:
+    std::string collect_params(const ov::snippets::lowered::ExpressionPtr& brgemm_expr,
+                               const snippets::lowered::LinearIR& linear_ir) {
         auto debug_config = linear_ir.get_config().debug_config;
         const auto brgemm = ov::as_type_ptr<BRGEMM_TYPE>(brgemm_expr->get_node());
         OPENVINO_ASSERT(brgemm, "Brgemm is nullptr!");
-        auto csv_path = linear_ir.get_config().debug_config.dumpParams.csv_path;
         std::stringstream ss;
-        ss << brgemm_expr->get_node()->get_friendly_name() << ",";
+        ss << brgemm_expr->get_node()->get_friendly_name() << ',';
         for (size_t i = 0; i < brgemm->get_input_size(); ++i) {
             ss << brgemm->get_input_element_type(i);
             if (i != brgemm->get_input_size() - 1) {
-                ss << ";";
+                ss << ';';
             }
         }
-        ss << ",";
+        ss << ',';
         for (size_t i = 0; i < brgemm->get_output_size(); ++i) {
             ss << brgemm->get_output_element_type(i);
             if (i != brgemm->get_output_size() - 1) {
-                ss << ";";
+                ss << ';';
             }
         }
-        ss << ",";
+        ss << ',';
         for (const auto& input : brgemm->inputs()) {
             const auto& shape = input.get_partial_shape();
             for (const auto& dim : shape) {
-                ss << dim << " ";
+                ss << dim << ' ';
             }
             ss.seekp(-1, ss.cur);
-            ss << ";";
+            ss << ';';
         }
         ss.seekp(-1, ss.cur);
-        ss << ",";
+        ss << ',';
         for (const auto& output : brgemm->outputs()) {
             const auto& shape = output.get_partial_shape();
             for (const auto& dim : shape) {
-                ss << dim << " ";
+                ss << dim << ' ';
             }
             ss.seekp(-1, ss.cur);
-            ss << ";";
+            ss << ';';
         }
         ss.seekp(-1, ss.cur);
-        ss << ",";
+        ss << ',';
         int64_t M, N, K;
         std::tie(M, N, K) = get_brgemm_dimensions(brgemm_expr);
-        ss << M << "," << N << "," << K << ",";
+        ss << M << ',' << N << ',' << K << ',';
         int64_t m_block = brgemm_expr->get_input_port_descriptor(0)->get_subtensor().front();
         int64_t n_block = brgemm_expr->get_input_port_descriptor(1)->get_subtensor().back();
         int64_t k_block = brgemm_expr->get_input_port_descriptor(0)->get_subtensor().back();
-        ss << m_block << "," << n_block << "," << k_block;
-        ss << '\n';
-        std::ofstream csv_file(csv_path, std::ios_base::app);
-        if (csv_file.seekp(0, std::ios_base::end).tellp() == 0) {
-            csv_file << "name,in_type,out_type,in_shapes,out_shapes,M,N,K,m_block,n_block,k_block\n";
-        }
-        csv_file << ss.str();
+        ss << m_block << ',' << n_block << ',' << k_block;
+        ss << ',';
+        return ss.str();
     }
-#endif  // SNIPPETS_DEBUG_CAPS
 };
+
 } // namespace pass
 } // namespace lowered
 } // namespace snippets
-} // namespace ov
+} // namespace ov
+
+#endif  // SNIPPETS_DEBUG_CAPS

src/common/snippets/include/snippets/op/perf_count.hpp

@@ -74,20 +74,25 @@ class PerfCountEnd : public PerfCountEndBase {
 public:
     OPENVINO_OP("PerfCountEnd", "SnippetsOpset", PerfCountEndBase);
     PerfCountEnd(const Output<Node>& pc_begin);
-    PerfCountEnd() = default;
-    ~PerfCountEnd() {
-        output_perf_count();
-    }
+    PerfCountEnd();
+    ~PerfCountEnd();
+
     void output_perf_count();
     std::shared_ptr<Node> clone_with_new_inputs(const OutputVector& inputs) const override;
 
     void init_pc_begin();
     void set_accumulated_time();
 
+    void dump_brgemm_params_csv_file();
+
 private:
     ov::threading::ThreadLocal<uint64_t> accumulation;
     ov::threading::ThreadLocal<uint32_t> iteration;
     std::shared_ptr<PerfCountBegin> m_pc_begin = nullptr;
+
+    static std::string brgemm_csv_path;
+    static std::map<std::string, std::string> m_debug_params_map;
+    static size_t nodes_count;
 };
 
 } // namespace op

src/common/snippets/src/lowered/pass/brgemm_debug_params.cpp

@@ -2,6 +2,8 @@
 // SPDX-License-Identifier: Apache-2.0
 //
 
+#ifdef SNIPPETS_DEBUG_CAPS
+
 #include "snippets/lowered/pass/brgemm_debug_params.hpp"
 
 #include "snippets/itt.hpp"
@@ -39,4 +41,6 @@ std::tuple<size_t, size_t, size_t> BrgemmDebugParamsBase::get_brgemm_dimensions(
 } // namespace pass
 } // namespace lowered
 } // namespace snippets
-} // namespace ov
+} // namespace ov
+
+#endif  // SNIPPETS_DEBUG_CAPS

src/common/snippets/src/op/perf_count.cpp

@@ -3,6 +3,8 @@
 //
 #ifdef SNIPPETS_DEBUG_CAPS
 
+#include <fstream>
+
 #include "snippets/op/perf_count.hpp"
 
 namespace ov {
@@ -62,9 +64,30 @@ void PerfCountBegin::set_start_time() {
 }
 
 //////////////////PerfCountEnd///////////////
-PerfCountEnd::PerfCountEnd(const Output<Node>& pc_begin) : PerfCountEndBase({pc_begin}), accumulation(0ul), iteration(0u) {
+
+size_t PerfCountEnd::nodes_count = 0;
+std::map<std::string, std::string> PerfCountEnd::m_debug_params_map;
+std::string PerfCountEnd::brgemm_csv_path;  // NOLINT
+
+PerfCountEnd::PerfCountEnd() : PerfCountEndBase() {
+    ++nodes_count;
+}
+
+PerfCountEnd::PerfCountEnd(const Output<Node>& pc_begin)
+    : PerfCountEndBase({pc_begin}),
+      accumulation(0ul),
+      iteration(0u) {
     constructor_validate_and_infer_types();
     init_pc_begin();
+    ++nodes_count;
+}
+
+PerfCountEnd::~PerfCountEnd() {
+    output_perf_count();
+    --nodes_count;
+    if (nodes_count == 0) {
+        dump_brgemm_params_csv_file();
+    }
 }
 
 std::shared_ptr<Node> PerfCountEnd::clone_with_new_inputs(const OutputVector& inputs) const {
@@ -109,6 +132,34 @@ void PerfCountEnd::output_perf_count() {
     std::cout << "max accumulated time:" << acc_max << "ns" << std::endl;
     // max avg
     std::cout << "max avg time:" << avg_max << "ns" << std::endl;
+
+    // Dump brgemm debug parameters to csv file
+    if (acc_max != 0 && avg_max != 0 && get_friendly_name().find("_DebugParams") != std::string::npos) {
+        const auto& rt_info = get_rt_info();
+        auto brgemm_params_it = rt_info.find("brgemm_params");
+        if (brgemm_params_it == rt_info.end()) {
+            return;
+        }
+        if (brgemm_csv_path.empty()) {
+            auto brgemm_csv_path_it = rt_info.find("brgemm_params_csv_path");
+            brgemm_csv_path = brgemm_csv_path_it->second.as<std::string>();
+        }
+        m_debug_params_map[get_friendly_name()] =
+            brgemm_params_it->second.as<std::string>() + std::to_string(acc_max) + ',' + std::to_string(avg_max);
+    }
+}
+
+void PerfCountEnd::dump_brgemm_params_csv_file() {
+    if (m_debug_params_map.empty()) {
+        return;
+    }
+    std::ofstream csv_file(brgemm_csv_path);
+    OPENVINO_ASSERT(csv_file.is_open(), "Failed to open csv file for brgemm debug parameters.");
+    csv_file << "name,in_type,out_type,in_shapes,out_shapes,M,N,K,m_block,n_block,k_block,acc_max_time,avg_max_time\n";
+    for (const auto& [_, params] : m_debug_params_map) {
+        csv_file << params << '\n';
+    }
+    csv_file.close();
 }
 
 } // namespace op

src/plugins/intel_cpu/src/nodes/subgraph.cpp

@@ -521,9 +521,11 @@ Subgraph::ControlFlowPasses Subgraph::getControlFlowPasses() const {
                                     ov::snippets::lowered::pass::MarkLoops,
                                     ov::intel_cpu::pass::BrgemmCPUBlocking);
 
+#ifdef SNIPPETS_DEBUG_CAPS
     SNIPPETS_REGISTER_PASS_RELATIVE(Place::After,
                                     ov::intel_cpu::pass::BrgemmCPUBlocking,
                                     ov::snippets::lowered::pass::BrgemmDebugParams<BrgemmCPU>);
+#endif  // SNIPPETS_DEBUG_CAPS
 
     SNIPPETS_REGISTER_PASS_RELATIVE(Place::After,
                                     ov::snippets::lowered::pass::InitLoops,