IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

brushed the ratio compute convs to depend on the ISA, rather plain stupid #cores

Browse Source
This commit is contained in:
myshevts 2021-04-05 19:12:10 +03:00
parent 6e7109317a
commit 2003fa3f46
2 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
inference-engine/src/mkldnn_plugin/mkldnn_plugin.cpp
View File

@ -558,7 +558,7 @@ Engine::NetworkPerfStats Engine::NetworkMemBandwidthTolerance(const InferenceEng
res.ratio_compute_deconvs = total_deconvs ? static_cast<float>(compute_deconvs)/total_deconvs : 0; res.ratio_compute_deconvs = total_deconvs ? static_cast<float>(compute_deconvs)/total_deconvs : 0;
return res; return res;
} }
static bool hasAVX512();
InferenceEngine::ExecutableNetworkInternal::Ptr InferenceEngine::ExecutableNetworkInternal::Ptr
Engine::LoadExeNetworkImpl(const InferenceEngine::CNNNetwork &network, const std::map<std::string, std::string> &orig_config) { Engine::LoadExeNetworkImpl(const InferenceEngine::CNNNetwork &network, const std::map<std::string, std::string> &orig_config) {
OV_ITT_SCOPED_TASK(itt::domains::MKLDNNPlugin, "Engine::LoadExeNetworkImpl"); OV_ITT_SCOPED_TASK(itt::domains::MKLDNNPlugin, "Engine::LoadExeNetworkImpl");
@ -624,16 +624,17 @@ Engine::LoadExeNetworkImpl(const InferenceEngine::CNNNetwork &network, const std
int num_streams; int num_streams;
bool considerNonLimited = false; bool considerNonLimited = false;
if (NetworkToleranceForLowCache.maxMemTolerance > NetworkPerfStats::memThresholdNotLimited) { if (NetworkToleranceForLowCache.maxMemTolerance > NetworkPerfStats::memThresholdNotLimited) {
std::cout << "case 1.0" <<std::endl; std::cout << " case 1.0" <<std::endl;
considerNonLimited = true; considerNonLimited = true;
} }
if ((NetworkToleranceForLowCache.ratio_compute_convs == NetworkPerfStats::ALL) if ((NetworkToleranceForLowCache.ratio_compute_convs == NetworkPerfStats::ALL)
|| (NetworkToleranceForLowCache.ratio_compute_deconvs == NetworkPerfStats::ALL)) { || (NetworkToleranceForLowCache.ratio_compute_deconvs == NetworkPerfStats::ALL)) {
std::cout << "case 1.1" <<std::endl; std::cout << " case 1.1" <<std::endl;
considerNonLimited = true; considerNonLimited = true;
} }
if (num_cores >= 12 && (NetworkToleranceForLowCache.ratio_mem_limited_convs <= NetworkPerfStats::memLimitedRatioThreshold)) { if (hasAVX512() && NetworkToleranceForLowCache.maxMemTolerance > NetworkPerfStats::memThresholdAssumeLimitedAVX512
std::cout << "case 1.2" <<std::endl; && (NetworkToleranceForLowCache.ratio_mem_limited_convs <= NetworkPerfStats::memLimitedRatioThresholdAVX512)) {
std::cout << " case 1.2" <<std::endl;
considerNonLimited = true; considerNonLimited = true;
} }
if (considerNonLimited) { if (considerNonLimited) {

5
inference-engine/src/mkldnn_plugin/mkldnn_plugin.h
View File

@ -50,11 +50,12 @@ private:
float ratio_mem_limited_convs = 0; float ratio_mem_limited_convs = 0;
float ratio_compute_deconvs = 0; float ratio_compute_deconvs = 0;
// static constexpr float memComputeConvs = 0.25f;
static constexpr float memThresholdNotLimited = 1.0f; static constexpr float memThresholdNotLimited = 1.0f;
static constexpr float memThresholdAssumeLimited = 0.5f; static constexpr float memThresholdAssumeLimited = 0.5f;
static constexpr float memThresholdAssumeLimitedAVX512 = memThresholdAssumeLimited/2;
static constexpr float memThresholdUnknown = FLT_MAX; static constexpr float memThresholdUnknown = FLT_MAX;
static constexpr float memLimitedRatioThreshold = 0.05;
static constexpr float memLimitedRatioThresholdAVX512 = 0.10;
static constexpr float ALL = 1.0f; static constexpr float ALL = 1.0f;
}; };
static NetworkPerfStats NetworkMemBandwidthTolerance(const InferenceEngine::CNNNetwork &network); static NetworkPerfStats NetworkMemBandwidthTolerance(const InferenceEngine::CNNNetwork &network);