commit 689c6ee3604ab34197812b13ad3ccefa4ca60418
author Ruomei Yan <ruomei.yan@arm.com> Thu Apr 25 17:48:41 2019 +0100
committer Ruomei Yan <ruomei.yan@arm.com> Tue Apr 30 11:47:59 2019 +0100
tree 50bd22e26d87c58f84f84fef69042ec5017274bb
parent 7f00441df4e9020b12e1e0900ad394df8540f8a7

IVGCVSW-2773 Integrate new CL tuner into ArmNN

!armnn:1037
Change-Id: Ice9c0fba405859d97aa848daa67c609166f2a8dc
Signed-off-by: Ruomei Yan <ruomei.yan@arm.com>

DriverOptions.cpp

diff --git a/DriverOptions.cpp b/DriverOptions.cpp
index cd4b6bf..6615e51 100644
--- a/DriverOptions.cpp
+++ b/DriverOptions.cpp
@@ -32,6 +32,7 @@
     : m_Backends({computeDevice})
     , m_VerboseLogging(false)
     , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters)
+    , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid)
     , m_EnableGpuProfiling(false)
     , m_fp16Enabled(fp16Enabled)
 {
@@ -41,6 +42,7 @@
     : m_Backends(backends)
     , m_VerboseLogging(false)
     , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters)
+    , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid)
     , m_EnableGpuProfiling(false)
     , m_fp16Enabled(fp16Enabled)
 {
@@ -49,6 +51,7 @@
 DriverOptions::DriverOptions(int argc, char** argv)
     : m_VerboseLogging(false)
     , m_ClTunedParametersMode(armnn::IGpuAccTunedParameters::Mode::UseTunedParameters)
+    , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid)
     , m_EnableGpuProfiling(false)
     , m_fp16Enabled(false)
 {
@@ -56,6 +59,7 @@
 
     std::string unsupportedOperationsAsString;
     std::string clTunedParametersModeAsString;
+    std::string clTuningLevelAsString;
 
     po::options_description optionsDesc("Options");
     optionsDesc.add_options()
@@ -89,6 +93,13 @@
          "If 'UpdateTunedParameters', will also find the optimum parameters when preparing new networks and update "
          "the file accordingly.")
 
+        ("cl-tuning-level,o",
+         po::value<std::string>(&clTuningLevelAsString)->default_value("rapid"),
+         "exhaustive: all lws values are tested "
+         "normal: reduced number of lws values but enough to still have the performance really close to the "
+         "exhaustive approach "
+         "rapid: only 3 lws values should be tested for each kernel ")
+
         ("gpu-profiling,p",
          po::bool_switch(&m_EnableGpuProfiling),
          "Turns GPU profiling on")
@@ -166,6 +177,24 @@
             ALOGW("Requested unknown cl-tuned-parameters-mode '%s'. Defaulting to UseTunedParameters",
                 clTunedParametersModeAsString.c_str());
         }
+
+        if (clTuningLevelAsString == "exhaustive")
+        {
+            m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Exhaustive;
+        }
+        else if (clTuningLevelAsString == "normal")
+        {
+            m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Normal;
+        }
+        else if (clTuningLevelAsString == "rapid")
+        {
+            m_ClTuningLevel = armnn::IGpuAccTunedParameters::TuningLevel::Rapid;
+        }
+        else
+        {
+            ALOGW("Requested unknown cl-tuner-mode '%s'. Defaulting to rapid",
+                  clTuningLevelAsString.c_str());
+        }
     }
 }