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

[CPU] fixed SIGSEGV when moving unaligned data in Quantize node (Issue 51255) (#5262)

Browse Source
This commit is contained in:
Anton Voronov 2021-04-29 11:41:33 +03:00 committed by GitHub
parent 2fb6b46e5a
commit 18fdc4b694
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 26 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

37
inference-engine/src/mkldnn_plugin/nodes/mkldnn_quantize_node.cpp
View File

@ -1014,21 +1014,36 @@ void MKLDNNQuantizeNode::init() {
bool quantizationOnly = true;
for (int i = 0; i < cropLow.size(); i++) {
float il = inputLowData[isInputLowBroadcasted ? 0 : i];
// WA: Depending on the hardware, the compiler may enable data movement optimizations that require working with aligned data.
// If we receive unaligned data, then working with them can lead to segfault. To avoid this situation, we added a function
// that pushes data into aligned memory.
auto prepareAlignedData = [](float *newData, const float *oldData, int dataSize) {
memcpy(newData, oldData, dataSize * sizeof(float));
};
cropLow[i] = il;
std::vector<float> inputLowDataAligned(inputLowAxisSize);
prepareAlignedData(&inputLowDataAligned[0], inputLowData, inputLowDataAligned.size());
std::vector<float> inputHighDataAligned(inputHighAxisSize);
prepareAlignedData(&inputHighDataAligned[0], inputHighData, inputHighDataAligned.size());
std::vector<float> outputLowDataAligned(outputLowAxisSize);
prepareAlignedData(&outputLowDataAligned[0], outputLowData, outputLowDataAligned.size());
std::vector<float> outputHighDataAligned(outputHighAxisSize);
prepareAlignedData(&outputHighDataAligned[0], outputHighData, outputHighDataAligned.size());
for (int i = 0; i < cropLow.size(); i++) {
cropLow[i] = inputLowDataAligned[isInputLowBroadcasted ? 0 : i];
}
for (int i = 0; i < cropHigh.size(); i++) {
float ih = inputHighData[isInputHighBroadcasted ? 0 : i];
cropHigh[i] = ih;
cropHigh[i] = inputHighDataAligned[isInputHighBroadcasted ? 0 : i];
}
for (int i = 0; i < inputScale.size(); i++) {
float il = inputLowData[isInputLowBroadcasted ? 0 : i];
float ih = inputHighData[isInputHighBroadcasted ? 0 : i];
float il = inputLowDataAligned[isInputLowBroadcasted ? 0 : i];
float ih = inputHighDataAligned[isInputHighBroadcasted ? 0 : i];
#if defined(VALIDATE_QUANTIZATION_RANGES)
if ((il == ih && levels != 2) || il > ih || std::isnan(il) || std::isnan(ih) || std::isinf(il) || std::isinf(ih)) {
@ -1042,8 +1057,8 @@ void MKLDNNQuantizeNode::init() {
}
for (int i = 0; i < outputScale.size(); i++) {
float ol = outputLowData[isOutputLowBroadcasted ? 0 : i];
float oh = outputHighData[isOutputHighBroadcasted ? 0 : i];
float ol = outputLowDataAligned[isOutputLowBroadcasted ? 0 : i];
float oh = outputHighDataAligned[isOutputHighBroadcasted ? 0 : i];
#if defined(VALIDATE_QUANTIZATION_RANGES)
if (std::isnan(ol) || std::isnan(oh) || std::isinf(ol) || std::isinf(oh)) {
@ -1059,7 +1074,7 @@ void MKLDNNQuantizeNode::init() {
}
for (int i = 0; i < outputShift.size(); i++) {
float ol = outputLowData[isOutputLowBroadcasted ? 0 : i];
float ol = outputLowDataAligned[isOutputLowBroadcasted ? 0 : i];
outputShift[i] = ol;