Skip to content

[QNN:Bugfix] Fix StridedSlice NC4HW4 layout and improve QNNPerf robustness #4583

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
lanling-47 wants to merge 1 commit into
base: master
Choose a base branch
from
+88 −16
Open

[QNN:Bugfix] Fix StridedSlice NC4HW4 layout and improve QNNPerf robustness #4583

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
34 changes: 31 additions & 3 deletions source/backend/qnn/backend/QNNPerf.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,13 +16,30 @@ QNNPerf::QNNPerf(const QNN_INTERFACE_VER_TYPE * qnnInterface) {
mQnnInterface = qnnInterface;

QnnDevice_Infrastructure_t deviceInfra = nullptr;
CALL_QNN(mQnnInterface->deviceGetInfrastructure(&deviceInfra));
auto infrastructureStatus = mQnnInterface->deviceGetInfrastructure(&deviceInfra);
if (infrastructureStatus != QNN_SUCCESS || deviceInfra == nullptr) {
MNN_PRINT("MNN_QNN: Failed to get HTP perf infrastructure, error:%lu infra=%p\n",
(unsigned long)infrastructureStatus,
deviceInfra);
return;
}
QnnHtpDevice_Infrastructure_t *htpInfra = static_cast<QnnHtpDevice_Infrastructure_t *>(deviceInfra);
if (htpInfra->infraType != QNN_HTP_DEVICE_INFRASTRUCTURE_TYPE_PERF || htpInfra->perfInfra.createPowerConfigId == nullptr) {
MNN_PRINT("MNN_QNN: HTP perf infrastructure is unavailable. infraType=%d createPowerConfigId=%p\n",
htpInfra->infraType,
htpInfra->perfInfra.createPowerConfigId);
return;
}
mPerfInfra = htpInfra->perfInfra;

uint32_t deviceId = 0;
uint32_t coreId = 0;
CALL_QNN(mPerfInfra.createPowerConfigId(deviceId, coreId, &mPowerConfigId));
auto powerConfigStatus = mPerfInfra.createPowerConfigId(deviceId, coreId, &mPowerConfigId);
if (powerConfigStatus != QNN_SUCCESS) {
MNN_PRINT("MNN_QNN: Failed to create HTP power config id, error:%lu\n", (unsigned long)powerConfigStatus);
mPowerConfigId = 0;
return;
}

mPowerConfigBurst = {
.option = QNN_HTP_PERF_INFRASTRUCTURE_POWER_CONFIGOPTION_DCVS_V3,
Expand Down Expand Up @@ -73,7 +90,9 @@ QNNPerf::QNNPerf(const QNN_INTERFACE_VER_TYPE * qnnInterface) {

// destory power config
QNNPerf::~QNNPerf() {
CALL_QNN(mPerfInfra.destroyPowerConfigId(mPowerConfigId));
if (mPerfInfra.destroyPowerConfigId != nullptr && mPowerConfigId != 0) {
CALL_QNN(mPerfInfra.destroyPowerConfigId(mPowerConfigId));
}
}


Expand All @@ -83,6 +102,9 @@ void QNNPerf::setRpcLatencyAndPolling() {
::memset(&rpcControlLatency, 0, sizeof(rpcControlLatency));
rpcControlLatency.option = QNN_HTP_PERF_INFRASTRUCTURE_POWER_CONFIGOPTION_RPC_CONTROL_LATENCY;
rpcControlLatency.rpcControlLatencyConfig = 100; // use rpc control latency recommended 100 us, refer hexagon sdk
if (mPerfInfra.setPowerConfig == nullptr || mPowerConfigId == 0) {
return;
}
const QnnHtpPerfInfrastructure_PowerConfig_t *powerConfigs1[] = {&rpcControlLatency, NULL};

CALL_QNN(mPerfInfra.setPowerConfig(mPowerConfigId, powerConfigs1)); // set RPC latency config on power config ID created
Expand All @@ -101,11 +123,17 @@ void QNNPerf::setPowerConfigBurst() {
#ifdef QNN_VERBOSE
MNN_PRINT("MNN QNN set burst mode\n");
#endif
if (mPerfInfra.setPowerConfig == nullptr || mPowerConfigId == 0) {
return;
}
const QnnHtpPerfInfrastructure_PowerConfig_t *powerConfigs[] = {&mPowerConfigBurst, NULL};
CALL_QNN(mPerfInfra.setPowerConfig(mPowerConfigId, powerConfigs));
}

void QNNPerf::setPowerConfigBalanced() {
if (mPerfInfra.setPowerConfig == nullptr || mPowerConfigId == 0) {
return;
}
const QnnHtpPerfInfrastructure_PowerConfig_t *powerConfigs[] = {&mPowerConfigBalanced, NULL};
CALL_QNN(mPerfInfra.setPowerConfig(mPowerConfigId, powerConfigs));
}
Expand Down
70 changes: 57 additions & 13 deletions source/backend/qnn/execution/QNNStridedSlice.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -92,15 +92,12 @@ ErrorCode QNNStridedSlice::onEncode(const std::vector<Tensor *> &inputs, const s
}
return NO_ERROR;
}
if (TensorUtils::getDescribe(inputs[0])->dimensionFormat == MNN_DATA_FORMAT_NC4HW4) {
MNN_ERROR("[QNN] Don't Support NC4HW4 stridedslice now\n");
return NOT_SUPPORT;
}
bool isNC4HW4 = (TensorUtils::getDescribe(inputs[0])->dimensionFormat == MNN_DATA_FORMAT_NC4HW4);

auto param = mOp->main_as_StridedSliceParam();
mNodeType = "StridedSlice";

// Deal with ranges.
// Deal with ranges (in MNN's NCHW logical order).
std::vector<int> beginRaw(mInputDim, 0);
std::vector<int> endRaw = inputTensor->shape();
std::vector<int> strideRaw(mInputDim, 1);
Expand All @@ -110,6 +107,30 @@ ErrorCode QNNStridedSlice::onEncode(const std::vector<Tensor *> &inputs, const s
this->computeRangesType1(inputs, beginRaw, endRaw, strideRaw);
}

// NC4HW4 layout: QNN sees NHWC, so remap ranges from NCHW → NHWC
if (isNC4HW4 && mInputDim >= 2) {
// NCHW order: [N, C, H, W, ...] → NHWC order: [N, H, W, ..., C]
// Save channel (axis=1) values
int beginC = beginRaw[1], endC = endRaw[1], strideC = strideRaw[1];
// Shift spatial dims left: axis 2..N-1 → axis 1..N-2
for (int i = 2; i < mInputDim; i++) {
beginRaw[i - 1] = beginRaw[i];
endRaw[i - 1] = endRaw[i];
strideRaw[i - 1] = strideRaw[i];
}
// Channel becomes last
beginRaw[mInputDim - 1] = beginC;
endRaw[mInputDim - 1] = endC;
strideRaw[mInputDim - 1] = strideC;

// Also remap endRaw to use NHWC shape from inputShape (already remapped above)
for (int i = 0; i < mInputDim; i++) {
if (endRaw[i] > inputShape[i]) {
endRaw[i] = inputShape[i];
}
}
}

std::vector<int> rangeData(mInputDim * 3, 0);
for (int axis = 0; axis < mInputDim; axis++) {
rangeData[3 * axis + 0] = beginRaw[axis];
Expand Down Expand Up @@ -166,21 +187,26 @@ void QNNStridedSlice::computeRangesType1(const std::vector<Tensor *> &inputs, st
auto inputTensor = inputs[0];
auto beginTensor = inputs[1];
auto endTensor = inputs[2];
auto strideTensor = inputs[4];
auto beginRawSource = beginTensor->host<int>();
auto endRawSource = endTensor->host<int>();
auto strideRawSource = strideTensor->host<int>();

// fromType=1: inputs = [data, begin, end, axes, (stride)]
// stride is optional — when absent (4 inputs), default all strides to 1
bool hasStride = (inputs.size() >= 5);
int* strideRawSource = nullptr;
if (hasStride) {
strideRawSource = inputs[4]->host<int>();
}

auto axisTensor = inputs[3];
int sliceDim = beginTensor->length(0);
MNN_ASSERT(sliceDim == endTensor->length(0) && sliceDim == axisTensor->length(0) && sliceDim == strideTensor->length(0));

for (int i = 0; i < sliceDim; i++) {
int tempAxis = axisTensor->host<int>()[i];
tempAxis = tempAxis >= 0 ? tempAxis : (tempAxis + mInputDim);
beginRaw[tempAxis] = CLIP(beginRawSource[i], 0, inputs[0]->length(tempAxis) - 1);
endRaw[tempAxis] = CLIP(endRawSource[i], 1, inputs[0]->length(tempAxis));
strideRaw[tempAxis] = strideRawSource[i];
strideRaw[tempAxis] = hasStride ? strideRawSource[i] : 1;
}
return;
}
Expand All @@ -206,25 +232,43 @@ class QNNStridedSliceCreator : public QnnBackend::Creator {
return new QNNStridedSlice(backend, op);
}
auto param = op->main_as_StridedSliceParam();
if (nullptr == param) {
MNN_PRINT("MNN_QNN StridedSlice: param is null\n");
return nullptr;
}

MNN_PRINT("MNN_QNN StridedSlice: fromType=%d, inputs=%zu, beginMask=%d, endMask=%d, shrinkMask=%d, newAxisMask=%d, ellipsisMask=%d\n",
param->fromType(), inputs.size(), param->beginMask(), param->endMask(),
param->shrinkAxisMask(), param->newAxisMask(), param->ellipsisMask());

// <begin>, <end> and <stride> should be static.
for (int i = 1; i < inputs.size(); i++) {
MNN_ASSERT(TensorUtils::getDescribe(inputs[i])->usage == Tensor::InsideDescribe::Usage::CONSTANT);
if (TensorUtils::getDescribe(inputs[i])->usage != Tensor::InsideDescribe::Usage::CONSTANT) {
MNN_PRINT("MNN_QNN StridedSlice: input[%d] is NOT constant (usage=%d), skip\n", i, (int)TensorUtils::getDescribe(inputs[i])->usage);
return nullptr;
}
}

if (param->fromType() == 1) {
MNN_ASSERT(param->shrinkAxisMask() == 0 && param->newAxisMask() == 0 && param->ellipsisMask() == 0);
if (inputs.size() != 5) {
if (param->shrinkAxisMask() != 0 || param->newAxisMask() != 0 || param->ellipsisMask() != 0) {
MNN_PRINT("MNN_QNN StridedSlice: fromType1 unsupported masks\n");
return nullptr;
}
// fromType=1: inputs = [data, begin, end, axes, (stride)]
// stride is optional — default to 1 when absent (4 inputs)
if (inputs.size() != 4 && inputs.size() != 5) {
MNN_PRINT("MNN_QNN StridedSlice: fromType1 inputs.size=%zu, need 4 or 5\n", inputs.size());
return nullptr;
}
return new QNNStridedSlice(backend, op);
}

// [TODO] 把newAxisMask和ellipsisMask考虑在内
if (param->fromType() == 0) {
if (inputs.size() == 4 && param->newAxisMask() == 0 && param->ellipsisMask() == 0) {
return new QNNStridedSlice(backend, op);
} else {
MNN_PRINT("MNN_QNN StridedSlice: fromType0 rejected: inputs=%zu, newAxisMask=%d, ellipsisMask=%d\n",
inputs.size(), param->newAxisMask(), param->ellipsisMask());
return nullptr;
}
}
Expand Down