diff options
author | Mike Kelly <mike.kelly@arm.com> | 2020-09-30 20:22:56 +0100 |
---|---|---|
committer | mike.kelly <mike.kelly@arm.com> | 2020-09-30 19:41:56 +0000 |
commit | 7ed56dd5808d483ab8e294f36e3ecf8e8ddd84f1 (patch) | |
tree | b53d5499204280f539e17e7f66d0d4e7790863c2 /DriverOptions.cpp | |
parent | 9b088d9c275ed11ca39f8b035e02a68f681c5d45 (diff) | |
download | android-nn-driver-7ed56dd5808d483ab8e294f36e3ecf8e8ddd84f1.tar.gz |
MLCE-233 Added ability to enable fast_math through the NNAPI driver
* Added -a command line option to enable fast_math on Neon/CL when launching the NNAPI driver
* The invocation will look something like:
/vendor/bin/hw/android.hardware.neuralnetworks@1.3-service-armnn -v -c GpuAcc -n arm-armnn -a
* Enabling fast_math may cause an increase in VTS or CTS test failures due to a change in precision
Signed-off-by: Mike Kelly <mike.kelly@arm.com>
Change-Id: Ifd8931f95efd65d3ae3aca9d9d61bd381e4e24a5
Diffstat (limited to 'DriverOptions.cpp')
-rw-r--r-- | DriverOptions.cpp | 7 |
1 files changed, 7 insertions, 0 deletions
diff --git a/DriverOptions.cpp b/DriverOptions.cpp index 1f534b6e..d179d653 100644 --- a/DriverOptions.cpp +++ b/DriverOptions.cpp @@ -35,6 +35,7 @@ DriverOptions::DriverOptions(armnn::Compute computeDevice, bool fp16Enabled) , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) , m_EnableGpuProfiling(false) , m_fp16Enabled(fp16Enabled) + , m_FastMathEnabled(false) { } @@ -45,6 +46,7 @@ DriverOptions::DriverOptions(const std::vector<armnn::BackendId>& backends, bool , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) , m_EnableGpuProfiling(false) , m_fp16Enabled(fp16Enabled) + , m_FastMathEnabled(false) { } @@ -54,6 +56,7 @@ DriverOptions::DriverOptions(int argc, char** argv) , m_ClTuningLevel(armnn::IGpuAccTunedParameters::TuningLevel::Rapid) , m_EnableGpuProfiling(false) , m_fp16Enabled(false) + , m_FastMathEnabled(false) { namespace po = boost::program_options; @@ -104,6 +107,10 @@ DriverOptions::DriverOptions(int argc, char** argv) "exhaustive approach " "rapid: only 3 lws values should be tested for each kernel ") + ("fast-math,a", + po::bool_switch(&m_FastMathEnabled), + "Turns FastMath on") + ("gpu-profiling,p", po::bool_switch(&m_EnableGpuProfiling), "Turns GPU profiling on") |