Refine CUDA profiler and delete the test file.

paddle/platform/cuda_profiler.h

@@ -22,23 +22,7 @@ namespace paddle {
 namespace platform {
 
 void CudaProfilerInit(std::string output_file, std::string output_mode,
-                      std::vector<std::string> config_flags) {
-  std::array<char, 128> buf;
-  std::string tmpl = "/tmp/cuda_profile_config.XXXXXX";
-  PADDLE_ENFORCE_LT(tmpl.size(), buf.size());
-  memcpy(buf.data(), tmpl.data(), tmpl.size());
-  auto result = mktemp(buf.data());
-  PADDLE_ENFORCE(strlen(result) != 0);
-  std::string config_file = result;
-
-  {
-    std::ofstream ofs(config_file, std::ios::out | std::ios::trunc);
-    PADDLE_ENFORCE(ofs.is_open(), "ofstream: ", ofs.rdstate());
-    for (const auto& line : config_flags) {
-      ofs << line << std::endl;
-    }
-  }
-
+                      std::string config_file) {
   PADDLE_ENFORCE(output_mode == "kvp" || output_mode == "csv");
   cudaOutputMode_t mode = output_mode == "csv" ? cudaCSV : cudaKeyValuePair;
   PADDLE_ENFORCE(

python/paddle/v2/fluid/profiler.py

 import paddle.v2.fluid.core as core
 from contextlib import contextmanager
+import os
 
 __all__ = ['CudaProfiler']
 
@@ -8,9 +9,20 @@ NVPROF_CONFIG = [
     "gpuendtimestamp",
     "gridsize3d",
     "threadblocksize",
+    "dynsmemperblock",
+    "stasmemperblock",
+    "regperthread",
+    "memtransfersize",
+    "memtransferdir",
+    "memtransferhostmemtype",
     "streamid",
+    "cacheconfigrequested",
+    "cacheconfigrequested",
+    "cacheconfigrequested",
     "enableonstart 0",
     "conckerneltrace",
+    "active_warps",
+    "active_warps",
 ]
 
 
@@ -30,17 +42,21 @@ def cuda_profiler(output_file, output_mode=None, config=None):
             written into this file.
         output_mode (string) : The output mode has Key-Value pair format and
             Comma separated values format. It should be 'kvp' or 'csv'.
-        config (string) : The profiler options and counters can refer to
-            "Compute Command Line Profiler User Guide".
+        config (list of string) : The profiler options and counters can refer
+            to "Compute Command Line Profiler User Guide".
     """
     if output_mode is None:
         output_mode = 'csv'
     if output_mode not in ['kvp', 'csv']:
         raise ValueError("The output mode must be 'kvp' or 'csv'.")
     config = NVPROF_CONFIG if config is None else config
-    core.nvprof_init(output_file, output_mode, config)
+    config_file = 'nvprof_config_file'
+    with open(config_file, 'wb') as fp:
+        fp.writelines(["%s\n" % item for item in config])
+    core.nvprof_init(output_file, output_mode, config_file)
     # Enables profiler collection by the active CUDA profiling tool.
     core.nvprof_start()
     yield
     # Disables profiler collection.
     core.nvprof_stop()
+    os.remove(config_file)

python/paddle/v2/fluid/tests/test_profiler.py

@@ -3,6 +3,7 @@ import numpy as np
 import paddle.v2.fluid as fluid
 import paddle.v2.fluid.profiler as profiler
 import paddle.v2.fluid.layers as layers
+import os
 
 
 class TestProfiler(unittest.TestCase):
@@ -18,10 +19,12 @@ class TestProfiler(unittest.TestCase):
         exe = fluid.Executor(place)
         exe.run(fluid.default_startup_program())
 
-        with profiler.cuda_profiler("cuda_profiler.txt", 'csv') as nvprof:
+        output_file = 'cuda_profiler.txt'
+        with profiler.cuda_profiler(output_file, 'csv') as nvprof:
             for i in range(epoc):
                 input = np.random.random(dshape).astype('float32')
                 exe.run(fluid.default_main_program(), feed={'data': input})
+        os.remove(output_file)
 
 
 if __name__ == '__main__':