close floating point check in popart (#665)

* close floating point check in popart

* create pipeline resource when computation created. not do it at the time of set the cache computation item

fix pixel bert detect application runtime error

* add weiming build script fix

Co-authored-by: yanwei <yw01041751@alibaba-inc.com>

Author: yanwei-gr

.github/actions/build/build_in_docker.sh

@@ -22,8 +22,8 @@ CONTAINER_NAME="halo.ci-$VER-$VARIANT"
 
 docker_run_flag=""
 cmake_flags="-DDNNL_COMPILER=gcc-10"
-check_cmds="parallel -k --plus LIT_NUM_SHARDS={##}  LIT_RUN_SHARD={#}  CUDA_VISIBLE_DEVICES={} ninja check-halo ::: {0..1}"
-check_cmds="$check_cmds && ninja FileCheck && parallel -k --plus LIT_NUM_SHARDS={##}  LIT_RUN_SHARD={#}  CUDA_VISIBLE_DEVICES={} ninja check-halo-models ::: {0..1}"
+check_cmds="ninja FileCheck && parallel -k --plus LIT_NUM_SHARDS={##}  LIT_RUN_SHARD={#}  CUDA_VISIBLE_DEVICES={} ninja check-halo ::: {0..1}"
+check_cmds="$check_cmds && parallel -k --plus LIT_NUM_SHARDS={##}  LIT_RUN_SHARD={#}  CUDA_VISIBLE_DEVICES={} ninja check-halo-models ::: {0..1}"
 
 if [[ "$VARIANT" =~ cuda ]]; then
   docker_run_flag="--runtime=nvidia"

ODLA/platforms/odla_popart/odla_compute.cc

@@ -65,13 +65,6 @@ odla_status odla_SetComputationItem(odla_computation comp, odla_item_type type,
       PopartConfig::instance()->set_load_cache(true);
       PopartConfig::instance()->set_cache_path(reinterpret_cast<char*>(value));
       PopartConfig::instance()->extract_config_from_cache();
-      _odla_computation::instance()->set_executor();
-      if (PopartConfig::instance()->execution_mode() == PARALLEL ||
-          PopartConfig::instance()->execution_mode() == PIPELINE) {
-        QManager::instance()->createQ(PopartConfig::instance()->queue_type());
-        QManager::instance()->getQ()->init(
-            PopartConfig::instance()->queue_capacity());
-      }
       break;
     default:
       std::cerr << "Unsupported property type: " << type << std::endl;
@@ -105,13 +98,13 @@ odla_status odla_CreateComputation(odla_computation* comp) {
   // Read the config file
   if (!PopartConfig::instance()->inited()) {
     PopartConfig::instance()->load_config(std::getenv("ODLA_POPART_CONFIG"));
-    _odla_computation::instance()->set_executor();
-    if (PopartConfig::instance()->execution_mode() == PARALLEL ||
-        PopartConfig::instance()->execution_mode() == PIPELINE) {
-      QManager::instance()->createQ(PopartConfig::instance()->queue_type());
-      QManager::instance()->getQ()->init(
-          PopartConfig::instance()->queue_capacity());
-    }
+  }
+  _odla_computation::instance()->set_executor();
+  if (PopartConfig::instance()->execution_mode() == PARALLEL ||
+      PopartConfig::instance()->execution_mode() == PIPELINE) {
+    QManager::instance()->createQ(PopartConfig::instance()->queue_type());
+    QManager::instance()->getQ()->init(
+        PopartConfig::instance()->queue_capacity());
   }
   return ODLA_SUCCESS;
 }

ODLA/platforms/odla_popart/odla_popart.cc

@@ -241,7 +241,7 @@ void _odla_computation::set_session_opts() {
   // session_opts_.matmulOptions["use128BitConvUnitLoad"] = "true";
   // session_opts_.matmulOptions["enableMultiStageReduce"] = "false";
   // session_opts_.matmulOptions["enableFastReduce"] = "true";
-  session_opts_.enableFloatingPointChecks = true;
+  session_opts_.enableFloatingPointChecks = false;
   session_opts_.enableStochasticRounding = false;
   session_opts_.enablePrefetchDatastreams = false; // true;
   session_opts_.enableOutlining = true;

ODLA/platforms/odla_popart/popart_config.cc

@@ -64,7 +64,6 @@ void PopartConfig::load_config(const char* file_path) {
   popart::logging::info("use default config");
   use_default();
   if (file_path != nullptr) load_from_file(file_path);
-  print();
 }
 
 void PopartConfig::parse_from_json(const json& jf) {
@@ -108,7 +107,6 @@ void PopartConfig::parse_from_json(const json& jf) {
                            element.value()[1]);
     }
   }
-
   if (jf.contains("queue_type"))
     queue_type_ = jf["queue_type"].get<std::string>();
   else
@@ -119,6 +117,7 @@ void PopartConfig::parse_from_json(const json& jf) {
   else
     queue_capacity_ = 1024 * 1024;
   if (jf.contains("debug")) debug_ = jf["debug"].get<bool>();
+  print();
 
   inited_ = true;
 }

ODLA/platforms/odla_popart/popart_config.h

@@ -108,6 +108,10 @@ class PopartConfig {
   inline bool save_model() { return save_model_; }
   inline const std::string& load_onnx_path() { return load_onnx_path_; }
   inline const std::string& save_model_path() { return save_model_path_; }
+  inline const std::string& get_default_config_string() {
+    return default_config_string_;
+    ;
+  }
   inline const int ipu_num() { return ipu_num_; }
   inline bool no_pipeline() { return pipeline_setting_.empty(); }
   inline std::string queue_type() { return queue_type_; }
@@ -116,9 +120,6 @@ class PopartConfig {
   inline bool inited() { return inited_; }
   inline std::shared_ptr<std::ifstream> get_cache_fs() { return cache_fs; }
   inline void set_cache_fs(std::shared_ptr<std::ifstream> fs) { cache_fs = fs; }
-  inline std::string get_default_config_string() {
-    return default_config_string_;
-  }
 
   inline bool load_cache() { return load_cache_; }
   inline const std::string load_cache_path() { return cache_path_; }