fixing typos in src/plugins/intel_cpu/src/transformations (#18016)
This commit is contained in:
+1
-1
@@ -10,7 +10,7 @@
|
|||||||
* ConvertConv1DBase detects 1D Convolution / GroupConvolution and replaces
|
* ConvertConv1DBase detects 1D Convolution / GroupConvolution and replaces
|
||||||
* it with the sequence Unsqueeze - 2D Convolution / GroupConvolution - Squeeze.
|
* it with the sequence Unsqueeze - 2D Convolution / GroupConvolution - Squeeze.
|
||||||
* Unsqueeze adds the additional dimension to Convolution inputs and Squeeze
|
* Unsqueeze adds the additional dimension to Convolution inputs and Squeeze
|
||||||
* removes the additional dimention from the Convolution output.
|
* removes the additional dimension from the Convolution output.
|
||||||
*
|
*
|
||||||
* Before:
|
* Before:
|
||||||
*
|
*
|
||||||
|
|||||||
+1
-1
@@ -104,7 +104,7 @@ ov::intel_cpu::ConvertFqRnnToQuantizedRnn::ConvertFqRnnToQuantizedRnn() {
|
|||||||
const auto& cell_state = pattern_map.at(cell_state_m);
|
const auto& cell_state = pattern_map.at(cell_state_m);
|
||||||
const auto& sequence_length = pattern_map.at(sequence_length_m);
|
const auto& sequence_length = pattern_map.at(sequence_length_m);
|
||||||
|
|
||||||
// @todo prototype removal of unnecessary fq between two consequtive rnn nodes
|
// @todo prototype removal of unnecessary fq between two consecutive rnn nodes
|
||||||
auto rnn_quantized_tr = std::make_shared<op::TypeRelaxed<ngraph::opset9::LSTMSequence>>(
|
auto rnn_quantized_tr = std::make_shared<op::TypeRelaxed<ngraph::opset9::LSTMSequence>>(
|
||||||
element::TypeVector{ element::f32, element::f32, element::f32, element::f32, element::f32, element::f32, element::f32 },
|
element::TypeVector{ element::f32, element::f32, element::f32, element::f32, element::f32, element::f32, element::f32 },
|
||||||
element::TypeVector{ element::f32, element::f32, element::f32 },
|
element::TypeVector{ element::f32, element::f32, element::f32 },
|
||||||
|
|||||||
@@ -76,9 +76,9 @@ ov::intel_cpu::NonQuantizedFullyConnectedBiasFusion::NonQuantizedFullyConnectedB
|
|||||||
this->register_matcher(m, callback);
|
this->register_matcher(m, callback);
|
||||||
}
|
}
|
||||||
|
|
||||||
//CPU plugin would config LPT not to propogate dequantization scale over bias to follow ONEDNN 3.x scheme.
|
//CPU plugin would config LPT not to propagate dequantization scale over bias to follow ONEDNN 3.x scheme.
|
||||||
//It is a little tricky now to first fuse bias not DQ for pattern "FC + DQ + BIAS".
|
//It is a little tricky now to first fuse bias not DQ for pattern "FC + DQ + BIAS".
|
||||||
//todo: Will move the FullyConnnect fusing into CPU and fuse the DQ and BIAS in topology order.
|
//todo: Will move the FullyConnect fusing into CPU and fuse the DQ and BIAS in topology order.
|
||||||
ov::intel_cpu::QuantizedFullyConnectedBiasFusion::QuantizedFullyConnectedBiasFusion() {
|
ov::intel_cpu::QuantizedFullyConnectedBiasFusion::QuantizedFullyConnectedBiasFusion() {
|
||||||
MATCHER_SCOPE(FullyConnectedBiasFusion);
|
MATCHER_SCOPE(FullyConnectedBiasFusion);
|
||||||
auto input = ngraph::pattern::any_input();
|
auto input = ngraph::pattern::any_input();
|
||||||
|
|||||||
+1
-1
@@ -43,7 +43,7 @@ inline void ConvertToCPUSpecificOpset(std::shared_ptr<ngraph::Function> &nGraphF
|
|||||||
if (!ov::op::util::has_op_with_type<ngraph::op::FakeQuantize>(nGraphFunc)) {
|
if (!ov::op::util::has_op_with_type<ngraph::op::FakeQuantize>(nGraphFunc)) {
|
||||||
CPU_REGISTER_PASS_COMMON(manager, ReshapeFullyConnectedFusion);
|
CPU_REGISTER_PASS_COMMON(manager, ReshapeFullyConnectedFusion);
|
||||||
}
|
}
|
||||||
// after transformation "MoveEltwiseUpThroughDataMov" there can be Reshape sequences that should be eliminated or fused
|
// after transformation "MoveEltwiseUpThroughDataMov" there can be reshaped sequences that should be eliminated or fused
|
||||||
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ReshapeSequenceFusion);
|
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ReshapeSequenceFusion);
|
||||||
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ConstantFolding);
|
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ConstantFolding);
|
||||||
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ConvertPrecision, precisions_map {{ ngraph::element::i64, ngraph::element::i32 }});
|
CPU_REGISTER_PASS_COMMON(manager, ov::pass::ConvertPrecision, precisions_map {{ ngraph::element::i64, ngraph::element::i32 }});
|
||||||
|
|||||||
Reference in New Issue
Block a user