experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

NNAPI delegate: fix int8 delegation for activations

Browse Source 
PiperOrigin-RevId: 351388766
Change-Id: I208be2f315bad218af9b35cdff790ab4ff564974
This commit is contained in:
Lev Proleev 2021-01-12 09:41:34 -08:00 committed by TensorFlower Gardener
parent 8c55d68ba0
commit 696cb254a1
1 changed files with 23 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
tensorflow/lite/delegates/nnapi/nnapi_delegate.cc
View File

@ -277,6 +277,7 @@ bool NeedInt8Conversion(const TfLiteContext* context, int builtin_code,
case kTfLiteBuiltinGreaterEqual:
case kTfLiteBuiltinHardSwish:
case kTfLiteBuiltinL2Normalization:
case kTfLiteBuiltinLeakyRelu:
case kTfLiteBuiltinLess:
case kTfLiteBuiltinLessEqual:
case kTfLiteBuiltinLogistic:
@ -288,6 +289,7 @@ bool NeedInt8Conversion(const TfLiteContext* context, int builtin_code,
case kTfLiteBuiltinNotEqual:
case kTfLiteBuiltinPad:
case kTfLiteBuiltinPadv2:
case kTfLiteBuiltinPrelu:
case kTfLiteBuiltinReduceMax:
case kTfLiteBuiltinReduceMin:
case kTfLiteBuiltinRelu:
@ -1874,8 +1876,14 @@ bool NNAPIDelegateKernel::Validate(
} break;
case kTfLiteBuiltinSoftmax: {
ExpectOpVersion(version, 2, &val_ctx);
const auto& input = context->tensors[node->outputs->data[0]];
ExpectIsFloatOrQuant8Operator(context, node, &val_ctx);
const auto& output = context->tensors[node->outputs->data[0]];
ExpectTypeIn(output.type, {kTfLiteFloat32, kTfLiteUInt8, kTfLiteInt8},
NNAPIValidationFailureType::kUnsupportedOutputType,
"Output type should be one of kTfLiteFloat32, kTfLiteUInt8, "
"kTfLiteInt8.",
&val_ctx);
const auto& input = context->tensors[node->inputs->data[0]];
const int input_rank = input.dims->size;
Expect(input_rank <= 4,
NNAPIValidationFailureType::kUnsupportedOperandRank,
@ -2695,6 +2703,10 @@ bool NNAPIDelegateKernel::Validate(
ExpectOpVersion(version, 1, &val_ctx);
ExpectMinAndroidSdkVersion(android_sdk_version, kMinSdkVersionForNNAPI13,
&val_ctx);
const auto input_type = context->tensors[node->inputs->data[0]].type;
Expect(input_type == kTfLiteFloat32,
NNAPIValidationFailureType::kUnsupportedInputType,
"NNAPI only supports floating point input.", &val_ctx);
} break;
case kTfLiteBuiltinFill: {
ExpectOpVersion(version, 1, &val_ctx);
@ -3423,6 +3435,14 @@ TfLiteStatus NNAPIDelegateKernel::Map(
mapping_args.builder->AddNewInputConstantTensor(
ANEURALNETWORKS_TENSOR_FLOAT32, kTfLiteFloat32, alpha_tensor.dims,
alpha_value, alpha_tensor.params, &new_tensor_index);
} else if (input_type == kTfLiteInt8 &&
android_sdk_version >= kMinSdkVersionForNNAPI13) {
alpha_tensor.params.scale = builtin->alpha;
std::vector<int8_t> alpha_value = {1};
mapping_args.builder->AddNewInputConstantTensor(
ANEURALNETWORKS_TENSOR_QUANT8_ASYMM_SIGNED, kTfLiteInt8,
alpha_tensor.dims, alpha_value, alpha_tensor.params,
&new_tensor_index);
} else {
alpha_tensor.params.scale = builtin->alpha;
std::vector<uint8_t> alpha_value = {1};
@ -4321,8 +4341,10 @@ TfLiteStatus NNAPIDelegateKernel::AddOpsAndTensors(
reg->builtin_code == kTfLiteBuiltinConcatenation ||
reg->builtin_code == kTfLiteBuiltinMaximum ||
reg->builtin_code == kTfLiteBuiltinMinimum ||
reg->builtin_code == kTfLiteBuiltinLeakyRelu ||
reg->builtin_code == kTfLiteBuiltinLess ||
reg->builtin_code == kTfLiteBuiltinLessEqual ||
reg->builtin_code == kTfLiteBuiltinPrelu ||
reg->builtin_code == kTfLiteBuiltinGreater ||
reg->builtin_code == kTfLiteBuiltinGreaterEqual ||
reg->builtin_code == kTfLiteBuiltinEqual ||