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

[CPU][ARM] Disable default fp16 inference precision (#19445)

Browse Source
This commit is contained in:
Aleksandr Voron 2023-08-29 15:08:46 +02:00 committed by GitHub
parent f6ab1e4833
commit 84fc6fb626
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 2 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
src/plugins/intel_cpu/src/config.cpp
View File

@ -245,16 +245,10 @@ 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;
#endif else
inferencePrecision = ov::element::f32;
} else { } else {
inferencePrecision = ov::element::f32; inferencePrecision = ov::element::f32;
} }