From 7ed979a52f0af620b9287454d6a5881aebbbb2df Mon Sep 17 00:00:00 2001
From: Francois Fleuret <francois@fleuret.org>
Date: Mon, 5 Dec 2016 10:49:47 +0100
Subject: [PATCH] Now decorate function as high as possible in the class
 hierarchy, and breaks down timing by function.

---
 profiler.lua | 47 +++++++++++++++++++++++++++++------------------
 1 file changed, 29 insertions(+), 18 deletions(-)

diff --git a/profiler.lua b/profiler.lua
index 5490651..6f63861 100755
--- a/profiler.lua
+++ b/profiler.lua
@@ -46,23 +46,28 @@ function profiler.decor(model, functionsToDecorate)
       }
 
    for _, name in pairs(functionsToDecorate) do
-      model.accTime = 0
+      model.accTime = {}
 
       local functionTable = model
 
-      if not rawget(functionTable, name) then
-         functionTable = getmetatable(model)
+      -- We decorate the function where it is defined in the class
+      -- hierarchy, so we have to go up the metatables until we find
+      -- it with rawget
+
+      while functionTable and not rawget(functionTable, name) do
+         functionTable = getmetatable(functionTable)
       end
 
-      if functionTable[name] and not (functionTable.orig and functionTable.orig[name]) then
+      local nameOrig = name .. '__orig'
+
+      if functionTable[name] and not functionTable[nameOrig] then
          print('Profiler decoring ' .. functionTable.__typename .. '.' .. name)
-         functionTable.orig = functionTable.orig or {}
-         functionTable.orig[name] = functionTable[name]
+         functionTable[nameOrig] = functionTable[name]
          functionTable[name] = function(self, ...)
             local startTime = sys.clock()
-            local result = { self.orig[name](self, unpack({...})) }
+            local result = { self[nameOrig](self, unpack({...})) }
             local endTime = sys.clock()
-            self.accTime = self.accTime + endTime - startTime
+            self.accTime[name] = (self.accTime[name] or 0) + endTime - startTime
             return unpack(result)
          end
       end
@@ -80,22 +85,28 @@ end
 function profiler.print(model, nbSamples, indent)
    local indent = indent or ''
 
+   print(string.format('%s* %s', indent, model.__typename))
+
    if nbSamples then
-      print(string.format('%s%s %.02fs (%.01fmus/sample)',
-                          indent,
-                          model.__typename,
-                          model.accTime,
-                          1e6 * model.accTime / nbSamples))
+      for l, t in pairs(model.accTime) do
+         print(string.format('%s  %s %.02fs (%.01fmus/sample)',
+                             indent,
+                             l,
+                             t,
+                             1e6 * t / nbSamples))
+      end
    else
-      print(string.format('%s%s %.02fs',
-                          indent,
-                          model.__typename,
-                          model.accTime))
+      for l, t in pairs(model.accTime) do
+         print(string.format('%s  %s %.02fs',
+                             indent,
+                             l,
+                             t))
+      end
    end
 
    if torch.isTypeOf(model, nn.Container) then
       for _, m in ipairs(model.modules) do
-         profiler.print(m, nbSamples, indent .. '   ')
+         profiler.print(m, nbSamples, indent .. '  ')
       end
    end
 end
-- 
2.20.1