feat(mge): add a tool which can analyze the file generated by compare_binary_iodump.py

GitOrigin-RevId: 9acab0a49f

dnn/src/cuda/cudnn_wrapper.cpp

@@ -485,7 +485,7 @@ CudnnAlgoPack::conv_bwd_data_algos() {
                                     CudnnAlgoPack::Attr>
             algos =
     { DEF_ALGO(CUDNN_CONVOLUTION_BWD_DATA_ALGO_0, false, false),
-      DEF_ALGO(CUDNN_CONVOLUTION_BWD_DATA_ALGO_1, true, false),
+      DEF_ALGO(CUDNN_CONVOLUTION_BWD_DATA_ALGO_1, true, true),
       DEF_ALGO(CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT, true, true),
       DEF_ALGO(CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING, true, true),
 #if CUDNN_MAJOR >= 5

imperative/python/megengine/tools/accuracy_shake_var_tree.py

@@ -0,0 +1,130 @@
+#!/usr/bin/env python3
+# -*-coding=utf-8-*-
+
+# This tool is used to analyze the file generated by compare_binary_iodump.py.
+# parse() can build a dependency tree with those varnodes
+# where accuracy shake occurs and show the root varnodes.
+# get_varNode()/get_dependence_list()/get_reference_list()/show_src_info()
+# are some functions which are used to query dependencies between varnodes.
+import argparse
+import os
+
+
+class varNode:
+    var_node_dict = {}
+    var_node_root_dict = {}
+
+    def __init__(self, id, dependence_list, src_info):
+        self.src_info = src_info
+        if not id in varNode.var_node_dict.keys():
+            self.id = id
+            self.dependence_list = []
+            self.reference_list = []
+        else:
+            self = varNode.var_node_dict[id]
+
+        if dependence_list:
+            self.vitrual = False
+            self.is_root = True
+        else:
+            self.vitrual = True
+            self.is_root = False
+
+        for i in dependence_list:
+            if not i in varNode.var_node_dict.keys():
+                varNode.var_node_dict[i] = varNode(i, [], "")
+
+            dv = varNode.var_node_dict[i]
+            self.dependence_list.append(dv)
+            if not dv.vitrual:
+                self.is_root = False
+            dv.reference_list.append(self)
+
+        for i in self.reference_list:
+            i.is_root = False
+            varNode.var_node_root_dict.pop[i.id]
+
+        if self.is_root:
+            varNode.var_node_root_dict[id] = self
+
+        varNode.var_node_dict[id] = self
+
+    @staticmethod
+    def get_varNode(id):
+        return varNode.var_node_dict[id]
+
+    def get_dependence_list(self):
+        return self.dependence_list
+
+    def get_reference_list(self):
+        return self.reference_list
+
+    def show_src_info(self):
+        print(self.src_info)
+
+
+def get_dependence(string, src_info):
+    start1 = "id:"
+    end1 = ","
+    e = 0
+
+    count = string.count(start1)
+    dependence_list = []
+    for x in range(0, count):
+        s = string.find(start1, e)
+        e = string.find(end1, s)
+        sub_str = string[s:e]
+        if x == 0:
+            var = sub_str
+        else:
+            dependence_list.append(sub_str)
+    varNode(var, dependence_list, src_info)
+
+
+def parse(filename):
+    with open(filename) as f:
+        varNode.var_node_dict.clear()
+        varNode.var_node_root_dict.clear()
+        line = f.readline()
+        s = ["", "", ""]
+        idx = 1
+        while line:
+            if line.find("not equal: ") != -1:
+                s[2] = line
+                src_info = s[0] + "\n" + s[1] + "\n" + s[2]
+                get_dependence(s[0], src_info)
+            else:
+                if line.find("var={id:") != -1:
+                    idx = idx ^ 1
+                    s[idx] = ""
+                s[idx] = s[idx] + line.strip()
+            line = f.readline()
+
+    return varNode.var_node_root_dict
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=(
+            "Analyze the outputs of compare_binary_iodump.py"
+            "Should save the outputs of compare_binary_iodump.py"
+            "as a file"
+        ),
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+    )
+    parser.add_argument(
+        "filename", help="file which save the outputs of compare_binary_iodump.py"
+    )
+    args = parser.parse_args()
+
+    parse(args.filename)
+
+    print("varnode root:")
+    for key, value in varNode.var_node_root_dict.items():
+        print(key)
+        print("detail info:")
+        value.show_src_info()
+
+
+if __name__ == "__main__":
+    main()

sdk/load-and-run/src/mgblar.cpp

@@ -701,25 +701,19 @@ void run_test_st(Args &env) {
     mgb::gopt::set_opr_algo_workspace_limit_inplace(vars, env.workspace_limit);
     using S = opr::mixin::AlgoChooserHelper::ExecutionPolicy::Strategy;
     S strategy = S::HEURISTIC;
+    if (env.reproducible) {
+        strategy = S::REPRODUCIBLE;
+    }
 #if MGB_ENABLE_FASTRUN
     if (env.use_full_run) {
-        if (env.reproducible) {
-            strategy = S::PROFILE | S::REPRODUCIBLE;
-        } else {
-            strategy = S::PROFILE;
-        }
+        strategy = S::PROFILE | strategy;
     } else if (env.use_fast_run) {
-        strategy = S::PROFILE | S::OPTIMIZED;
-        if (env.reproducible){
-            strategy = strategy | S::REPRODUCIBLE;
-        }
-    } else if (env.reproducible) {
-        strategy = S::HEURISTIC | S::REPRODUCIBLE;
+        strategy = S::PROFILE | S::OPTIMIZED | strategy;
+    } else {
+        strategy = S::HEURISTIC | strategy;
     }
 #else
-    if (env.reproducible) {
-        strategy = S::HEURISTIC | S::REPRODUCIBLE;
-    }
+    strategy = S::HEURISTIC | strategy;
 #endif
     mgb::gopt::modify_opr_algo_strategy_inplace(vars, strategy);
     if (!env.fast_run_cache_path.empty()) {