[TMP] MatMul experiments via benchmark_app

samples/cpp/benchmark_app/main.cpp

@@ -12,6 +12,8 @@
 
 // clang-format off
 #include "openvino/openvino.hpp"
+#include "openvino/op/constant.hpp"
+#include "openvino/op/matmul.hpp"
 #include "openvino/pass/serialize.hpp"
 
 #ifndef IN_OV_COMPONENT
@@ -406,6 +408,7 @@ int main(int argc, char* argv[]) {
                 // set to default value
                 device_config[ov::enable_profiling.name()] = FLAGS_pc;
             }
+            device_config[ov::enable_profiling.name()] = true;
             perf_counts = (device_config.at(ov::enable_profiling.name()).as<bool>()) ? true : perf_counts;
 
             auto supported_properties = core.get_property(device, ov::supported_properties);
@@ -596,7 +599,21 @@ int main(int argc, char* argv[]) {
             slog::info << "Loading model files" << slog::endl;
 
             auto startTime = Time::now();
-            auto model = core.read_model(FLAGS_m);
+            auto model = []() {
+                const size_t batch = std::atoi(std::getenv("B"));
+                const size_t M = std::atoi(std::getenv("M"));
+                const size_t K = std::atoi(std::getenv("K"));
+                const size_t N = std::atoi(std::getenv("N"));
+
+                ov::Shape data_shape{1, batch, M, K};
+                auto data = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, data_shape);
+                std::vector<float> weights_vals(K * N);
+                std::iota(weights_vals.begin(), weights_vals.end(), -1000);
+                auto weights = ov::op::v0::Constant::create(ov::element::f32, {K, N}, weights_vals);
+                auto matmul = std::make_shared<ov::op::v0::MatMul>(data, weights, false, false);
+                matmul->set_friendly_name("MatMul");
+                return std::make_shared<ov::Model>(ov::NodeVector{matmul}, ov::ParameterVector{data});
+            }();
             auto duration_ms = get_duration_ms_till_now(startTime);
             slog::info << "Read model took " << double_to_string(duration_ms) << " ms" << slog::endl;
             slog::info << "Original model I/O parameters:" << slog::endl;
@@ -1218,6 +1235,18 @@ int main(int argc, char* argv[]) {
             slog::info << "OpenVINO Runtime configuration settings were dumped to " << FLAGS_dump_config << slog::endl;
         }
 
+        const auto runtime_model = compiledModel.get_runtime_model();
+        for (const auto& op : runtime_model->get_ordered_ops()) {
+            if (op->get_friendly_name() == "MatMul") {
+                const auto& rt_info = op->get_rt_info();
+                auto exec_time = rt_info.at("execTimeMcs");
+                auto layer_type = rt_info.at("layerType");
+                auto primitive_type = rt_info.at("primitiveType");
+                std::cout << "execTimeMcs: " << exec_time.as<std::string>() << std::endl;
+                std::cout << "layerType: " << layer_type.as<std::string>() << std::endl;
+                std::cout << "primitiveType: " << primitive_type.as<std::string>() << std::endl;
+            }
+        }
         if (!FLAGS_exec_graph_path.empty()) {
             try {
                 ov::serialize(compiledModel.get_runtime_model(), FLAGS_exec_graph_path);

src/plugins/intel_cpu/src/transformations/snippets/x64/pass/lowered/brgemm_cpu_blocking.cpp

@@ -70,7 +70,10 @@ std::tuple<size_t, size_t, size_t> BrgemmCPUBlocking::get_blocking_params(const
         n_blk = get_full_dim_value();
         k_blk = get_full_dim_value();
     }
-    return std::make_tuple(m_blk, n_blk, k_blk);
+    const size_t M = std::atoi(std::getenv("M_b"));
+    const size_t K = std::atoi(std::getenv("K_b"));
+    const size_t N = std::atoi(std::getenv("N_b"));
+    return std::make_tuple(M, N, K);
 }
 
 SpecificIterationHandlers BrgemmCPUBlocking::get_k_loop_handlers(size_t work_amount, size_t block_size) const {