[PerfModel] Protect init under FS.

(cherry picked from commit 25d4d28b068920dabb2642bb44d1f6278d19dec6)
tenstorrent · May 9, 2024 · 234f588 · 234f588
1 parent d00a010
commit 234f588
Show file tree

Hide file tree

Showing 2 changed files with 27 additions and 25 deletions.
diff --git a/pybuda/csrc/perf_model/graph.hpp b/pybuda/csrc/perf_model/graph.hpp
@@ -101,15 +101,14 @@ struct OpPerfData
         _get_execution_cycles(arch_name);
         return _theoretical_cycles;
     }
-    const balancer::OpCycleEstimates& get_op_cycle_estimates(
+    const balancer::OpCycleEstimates &get_op_cycle_estimates(
         const DeviceConfig &device_config,
         const graphlib::Graph *graph,
         bool input_queues_on_host,
         bool output_queues_on_host,
         const std::unordered_map<graphlib::Node const *, balancer::OpModel> &selected_op_models)
     {
-        _get_op_cycle_estimates(
-            device_config, graph, input_queues_on_host, output_queues_on_host, selected_op_models);
+        _get_op_cycle_estimates(device_config, graph, input_queues_on_host, output_queues_on_host, selected_op_models);
         return _op_cycle_estimates;
     }
 };
@@ -118,12 +117,13 @@ struct OpPerfCalculatedData
 {
     // BWs - ideal/actual
     std::vector<float> input_bw_needed, input_bw_got;
-    float output_bw_perc;  // the percentage of required bw we got (for worst case operand), which is also output bw%
-    float output_bw_ideal, output_bw_produced;
+    float output_bw_perc =
+        0;  // the percentage of required bw we got (for worst case operand), which is also output bw%
+    float output_bw_ideal = 0, output_bw_produced = 0;
 
     // Cycle counts, utilization
-    float utilization;
-    std::uint32_t cycle_count_actual;
+    float utilization = 0;
+    std::uint32_t cycle_count_actual = 0;
 };
 
 struct QueuePerfData
@@ -135,12 +135,12 @@ struct QueuePerfData
 
 struct QueuePerfCalculatedData
 {
-    float total_read_bw_ideal;  // ideal total BW requested by all consumers
-    float write_bw_ideal;       // ideal write BW from the producer
+    float total_read_bw_ideal = 0;  // ideal total BW requested by all consumers
+    float write_bw_ideal = 0;       // ideal write BW from the producer
 
-    float total_bw_perc;           // the percentage of requested bw that we can get from dram
-    float total_read_bw_produced;  // actual BW that can be given to the op
-    float write_bw_received;       // actual write BW from the producer
+    float total_bw_perc = 0;           // the percentage of requested bw that we can get from dram
+    float total_read_bw_produced = 0;  // actual BW that can be given to the op
+    float write_bw_received = 0;       // actual write BW from the producer
 };
 
 struct Attr

diff --git a/pybuda/csrc/perf_model/perf_model.cpp b/pybuda/csrc/perf_model/perf_model.cpp
@@ -281,7 +281,8 @@ void PerfModel::create_graphs(
         op_perf << "name, type, epoch, grid, tiles, cycles, limiter_cycles";
         for (unsigned int input_idx = 0; input_idx < c_op_max_num_inputs; ++input_idx)
         {
-            op_perf << ", " << "estimated_input_bw_" << input_idx;
+            op_perf << ", "
+                    << "estimated_input_bw_" << input_idx;
         }
         op_perf << ", estimated_output_bw_0" << std::endl;
 
@@ -306,7 +307,7 @@ void PerfModel::create_graphs(
             NodeP op = node_map.at(node);
             if (dump_op_perf)
             {
-                const balancer::OpCycleEstimates& op_cycle_estimates = 
+                const balancer::OpCycleEstimates &op_cycle_estimates =
                     op->get_perf_data()->op_perf_data.get_op_cycle_estimates(
                         device_config,
                         g,
@@ -324,8 +325,9 @@ void PerfModel::create_graphs(
 
                 for (unsigned int input_idx = 0; input_idx < c_op_max_num_inputs; ++input_idx)
                 {
-                    float input_bw = input_idx < op_cycle_estimates.input_bw_estimates.size() ?
-                                     op_cycle_estimates.input_bw_estimates[input_idx] : 0.0f;
+                    float input_bw = input_idx < op_cycle_estimates.input_bw_estimates.size()
+                                         ? op_cycle_estimates.input_bw_estimates[input_idx]
+                                         : 0.0f;
 
                     op_perf << std::to_string(input_bw) << ", ";
                 }
@@ -604,18 +606,18 @@ PerfModel::PerfModel(
     create_graphs(g, balancer_solution, input_queues_on_host, output_queues_on_host);
     SystemSpec system = SystemSpec::get_for_device(device_config);
 
-    // calculate ideal bandwidths for queues and ops
-    calculate_ideal_bws(system);
+    if (env_as<bool>("PYBUDA_PERF_SIMULATOR"))
+    {
+        // calculate ideal bandwidths for queues and ops
+        calculate_ideal_bws(system);
 
-    // calculate utilization
-    if (env_as<bool>("PYBUDA_PERF_UTIL"))
-        calculate_utilization(system);
+        // calculate utilization
+        if (env_as<bool>("PYBUDA_PERF_UTIL"))
+            calculate_utilization(system);
 
-    // Propagate BWs
-    for (auto &epoch_graph : temporal_epoch_graphs) propagate_bws(epoch_graph.get(), system);
+        // Propagate BWs
+        for (auto &epoch_graph : temporal_epoch_graphs) propagate_bws(epoch_graph.get(), system);
 
-    if (env_as<bool>("PYBUDA_PERF_SIMULATOR"))
-    {
         std::uint32_t original_microbatch = g->get_microbatch();
         if (auto sim_mb = env_as_optional<int>("PYBUDA_PERF_SIMULATOR_MICROBATCH"))
             g->set_microbatch(*sim_mb);