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

[CPU][ARM] Fixed execution mode behavior (#19192)

Browse Source
This commit is contained in:
Gorokhov Dmitriy 2023-08-15 16:46:05 +04:00 committed by GitHub
parent 45a534d209
commit 5ff67fca40
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 14 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
src/plugins/intel_cpu/src/config.cpp
View File

@ -245,20 +245,19 @@ void Config::readProperties(const std::map<std::string, std::string> &prop, Mode
// when both execution_mode and inference_precision are specified // when both execution_mode and inference_precision are specified
if (!inferencePrecisionSetExplicitly) { if (!inferencePrecisionSetExplicitly) {
if (executionMode == ov::hint::ExecutionMode::PERFORMANCE) { if (executionMode == ov::hint::ExecutionMode::PERFORMANCE) {
inferencePrecision = ov::element::f32;
#if defined(OV_CPU_ARM_ENABLE_FP16)
//fp16 precision is used as default precision on ARM for non-convolution networks
//fp16 ACL convolution is slower than fp32
if (modelType != ModelType::CNN)
inferencePrecision = ov::element::f16;
#else
if (mayiuse(avx512_core_bf16)) if (mayiuse(avx512_core_bf16))
inferencePrecision = ov::element::bf16; inferencePrecision = ov::element::bf16;
else #endif
inferencePrecision = ov::element::f32;
} else { } else {
inferencePrecision = ov::element::f32; inferencePrecision = ov::element::f32;
} }
#if defined(OV_CPU_ARM_ENABLE_FP16)
//fp16 precision is used as default precision on ARM for non-convolution networks
//fp16 ACL convolution is slower than fp32
if (modelType != ModelType::CNN) {
inferencePrecision = ov::element::f16;
}
#endif
} }
if (!prop.empty()) if (!prop.empty())

12
src/plugins/intel_cpu/tests/functional/behavior/ov_plugin/properties.cpp
View File

@ -203,7 +203,7 @@ TEST_F(OVClassConfigTestCPU, smoke_PluginSetConfigEnableProfiling) {
const auto bf16_if_can_be_emulated = InferenceEngine::with_cpu_x86_avx512_core() ? ov::element::bf16 : ov::element::f32; const auto bf16_if_can_be_emulated = InferenceEngine::with_cpu_x86_avx512_core() ? ov::element::bf16 : ov::element::f32;
using ExpectedModeAndType = std::pair<ov::hint::ExecutionMode, ov::element::Type>; using ExpectedModeAndType = std::pair<ov::hint::ExecutionMode, ov::element::Type>;
const std::map<ov::hint::ExecutionMode, ExpectedModeAndType> exectedTypeByMode { const std::map<ov::hint::ExecutionMode, ExpectedModeAndType> expectedTypeByMode {
{ov::hint::ExecutionMode::PERFORMANCE, {ov::hint::ExecutionMode::PERFORMANCE, {ov::hint::ExecutionMode::PERFORMANCE, {ov::hint::ExecutionMode::PERFORMANCE,
expected_precision_for_performance_mode}}, expected_precision_for_performance_mode}},
{ov::hint::ExecutionMode::ACCURACY, {ov::hint::ExecutionMode::ACCURACY, {ov::hint::ExecutionMode::ACCURACY, {ov::hint::ExecutionMode::ACCURACY,
@ -223,17 +223,17 @@ TEST_F(OVClassConfigTestCPU, smoke_PluginSetConfigExecutionModeExpectCorrespondi
ASSERT_NO_THROW(execution_mode_value = ie.get_property("CPU", ov::hint::execution_mode)); ASSERT_NO_THROW(execution_mode_value = ie.get_property("CPU", ov::hint::execution_mode));
ASSERT_EQ(execution_mode_value, execution_mode_default); ASSERT_EQ(execution_mode_value, execution_mode_default);
for (const auto& m : exectedTypeByMode) { for (const auto& m : expectedTypeByMode) {
const auto execution_mode = m.first; const auto execution_mode = m.first;
const auto execution_mode_exected = m.second.first; const auto execution_mode_expected = m.second.first;
const auto inference_precision_exected = m.second.second; const auto inference_precision_expected = m.second.second;
ASSERT_NO_THROW(ie.set_property("CPU", ov::hint::execution_mode(execution_mode))); ASSERT_NO_THROW(ie.set_property("CPU", ov::hint::execution_mode(execution_mode)));
ASSERT_NO_THROW(execution_mode_value = ie.get_property("CPU", ov::hint::execution_mode)); ASSERT_NO_THROW(execution_mode_value = ie.get_property("CPU", ov::hint::execution_mode));
ASSERT_EQ(execution_mode_value, execution_mode_exected); ASSERT_EQ(execution_mode_value, execution_mode_expected);
ASSERT_NO_THROW(inference_precision_value = ie.get_property("CPU", ov::hint::inference_precision)); ASSERT_NO_THROW(inference_precision_value = ie.get_property("CPU", ov::hint::inference_precision));
ASSERT_EQ(inference_precision_value, inference_precision_exected); ASSERT_EQ(inference_precision_value, inference_precision_expected);
} }
} }