Skip to content

CI测试不review replace Mkldnn to Onednn in compute_propagate_scales_onednn_pass [fluid_ops] #75827

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

Closed
co63oc wants to merge 1 commit into from
Closed

CI测试不review replace Mkldnn to Onednn in compute_propagate_scales_onednn_pass [fluid_ops] #75827

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
30 changes: 15 additions & 15 deletions paddle/fluid/framework/ir/onednn/compute_propagate_scales_onednn_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@

namespace paddle::framework::ir {

void ComputePropagateScalesMkldnnPass::GetTensorFromVector(
void ComputePropagateScalesOnednnPass::GetTensorFromVector(
const std::vector<float>& data_v, phi::DenseTensor* tensor) const {
const int size = static_cast<int>(data_v.size());
auto* data = tensor->mutable_data<float>({size}, phi::CPUPlace());
Expand All @@ -32,7 +32,7 @@ void ComputePropagateScalesMkldnnPass::GetTensorFromVector(
}
}

void ComputePropagateScalesMkldnnPass::GetQuantInfo(
void ComputePropagateScalesOnednnPass::GetQuantInfo(
ir::Graph* graph, StringPairMap* var_quant_scales) const {
std::unordered_map<std::string, std::vector<float>> info_map{};
GetInfoFromTheTmpOp(graph, "has_quant_info", "var_quant_scales", &info_map);
Expand All @@ -45,7 +45,7 @@ void ComputePropagateScalesMkldnnPass::GetQuantInfo(
}
}

std::vector<float> ComputePropagateScalesMkldnnPass::GetScales(
std::vector<float> ComputePropagateScalesOnednnPass::GetScales(
phi::DenseTensor* tensor, int axis) const {
PADDLE_ENFORCE_LT(axis,
2,
Expand Down Expand Up @@ -89,7 +89,7 @@ std::vector<float> ComputePropagateScalesMkldnnPass::GetScales(
return scales;
}

void ComputePropagateScalesMkldnnPass::ComputeVarScales(
void ComputePropagateScalesOnednnPass::ComputeVarScales(
ir::Graph* graph,
Scope* scope,
const std::unordered_set<std::string>& ops,
Expand Down Expand Up @@ -135,7 +135,7 @@ void ComputePropagateScalesMkldnnPass::ComputeVarScales(
}
}

void ComputePropagateScalesMkldnnPass::ComputeSingleGruWeightScales(
void ComputePropagateScalesOnednnPass::ComputeSingleGruWeightScales(
Scope* scope,
const std::string& wx_var_name,
const std::string& wh_var_name,
Expand Down Expand Up @@ -199,7 +199,7 @@ void ComputePropagateScalesMkldnnPass::ComputeSingleGruWeightScales(
GetTensorFromVector(scale_ur, tensor);
}

void ComputePropagateScalesMkldnnPass::ComputeGruWeightScales(
void ComputePropagateScalesOnednnPass::ComputeGruWeightScales(
ir::Graph* graph,
Scope* scope,
const std::string& wx_name,
Expand Down Expand Up @@ -234,7 +234,7 @@ void ComputePropagateScalesMkldnnPass::ComputeGruWeightScales(
}
}

void ComputePropagateScalesMkldnnPass::ComputeSingleLstmWeightScales(
void ComputePropagateScalesOnednnPass::ComputeSingleLstmWeightScales(
Scope* scope,
const std::string& wx_var_name,
const std::string& wh_var_name,
Expand Down Expand Up @@ -277,7 +277,7 @@ void ComputePropagateScalesMkldnnPass::ComputeSingleLstmWeightScales(
GetTensorFromVector(scale, tensor);
}

void ComputePropagateScalesMkldnnPass::ComputeLstmWeightScales(
void ComputePropagateScalesOnednnPass::ComputeLstmWeightScales(
ir::Graph* graph,
Scope* scope,
const std::string& wx_name,
Expand Down Expand Up @@ -313,7 +313,7 @@ void ComputePropagateScalesMkldnnPass::ComputeLstmWeightScales(
}
}

void ComputePropagateScalesMkldnnPass::ComputeWeightScales(
void ComputePropagateScalesOnednnPass::ComputeWeightScales(
ir::Graph* graph, Scope* scope, StringPairMap* var_quant_scales) const {
ComputeVarScales(graph,
scope,
Expand All @@ -334,7 +334,7 @@ void ComputePropagateScalesMkldnnPass::ComputeWeightScales(
ComputeLstmWeightScales(graph, scope, "WeightX", "WeightH", var_quant_scales);
}

void ComputePropagateScalesMkldnnPass::UpdateScaleOpInOutScales(
void ComputePropagateScalesOnednnPass::UpdateScaleOpInOutScales(
Node* op_node,
const std::string& input_name,
const std::string& output_name,
Expand Down Expand Up @@ -376,7 +376,7 @@ void ComputePropagateScalesMkldnnPass::UpdateScaleOpInOutScales(
var_quant_scales->insert(std::make_pair(name, new_pair));
}

std::unordered_set<std::string> ComputePropagateScalesMkldnnPass::UpdateScales(
std::unordered_set<std::string> ComputePropagateScalesOnednnPass::UpdateScales(
ir::Graph* graph,
StringPairMap* var_quant_scales,
const std::unordered_set<std::string>& scale_immutable_ops) const {
Expand Down Expand Up @@ -432,7 +432,7 @@ std::unordered_set<std::string> ComputePropagateScalesMkldnnPass::UpdateScales(
}
return waiting_for_scale;
}
void ComputePropagateScalesMkldnnPass::UpdateReluOutputScales(
void ComputePropagateScalesOnednnPass::UpdateReluOutputScales(
ir::Graph* graph, StringPairMap* var_quant_scales) const {
for (auto* op_node :
ir::TopologyVariantSort(*graph, static_cast<ir::SortKind>(0))) {
Expand Down Expand Up @@ -467,7 +467,7 @@ void ComputePropagateScalesMkldnnPass::UpdateReluOutputScales(
}
}

void ComputePropagateScalesMkldnnPass::PropagateScales(
void ComputePropagateScalesOnednnPass::PropagateScales(
ir::Graph* graph,
StringPairMap* var_quant_scales,
const std::unordered_set<std::string>& scale_immutable_ops) const {
Expand All @@ -484,7 +484,7 @@ void ComputePropagateScalesMkldnnPass::PropagateScales(
}
}

void ComputePropagateScalesMkldnnPass::ApplyImpl(ir::Graph* graph) const {
void ComputePropagateScalesOnednnPass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Convert paddle model to onednn quantized model.";
const std::string pattern_name = "compute_propagate_scales_onednn_pass";
FusePassBase::Init(pattern_name, graph);
Expand Down Expand Up @@ -517,7 +517,7 @@ void ComputePropagateScalesMkldnnPass::ApplyImpl(ir::Graph* graph) const {
} // namespace paddle::framework::ir

REGISTER_PASS(compute_propagate_scales_onednn_pass,
paddle::framework::ir::ComputePropagateScalesMkldnnPass);
paddle::framework::ir::ComputePropagateScalesOnednnPass);

REGISTER_PASS_CAPABILITY(compute_propagate_scales_onednn_pass)
.AddCombination(
Expand Down
8 changes: 4 additions & 4 deletions paddle/fluid/framework/ir/onednn/compute_propagate_scales_onednn_pass.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,13 +23,13 @@ namespace paddle {
namespace framework {
namespace ir {

class ComputePropagateScalesMkldnnPass : public FusePassBase {
class ComputePropagateScalesOnednnPass : public FusePassBase {
public:
ComputePropagateScalesMkldnnPass() = default;
virtual ~ComputePropagateScalesMkldnnPass() {}
ComputePropagateScalesOnednnPass() = default;
virtual ~ComputePropagateScalesOnednnPass() {}

#ifdef PADDLE_WITH_TESTING
friend class ComputePropagateScalesMkldnnPassTest;
friend class ComputePropagateScalesOnednnPassTest;
#endif

protected:
Expand Down
38 changes: 19 additions & 19 deletions paddle/fluid/framework/ir/onednn/quant_dequant_onednn_pass.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,7 +22,7 @@

namespace paddle::framework::ir {

void QuantDequantMkldnnPass::MarkSkipQuantizedOps(
void QuantDequantOnednnPass::MarkSkipQuantizedOps(
ir::Graph* graph, const std::unordered_set<std::string>& skip_ops) const {
VLOG(3) << "mark skip quantized ops";
for (auto* op_node :
Expand Down Expand Up @@ -53,7 +53,7 @@ void QuantDequantMkldnnPass::MarkSkipQuantizedOps(
}
}

void QuantDequantMkldnnPass::CollectInfoFromFake(
void QuantDequantOnednnPass::CollectInfoFromFake(
ir::Graph* graph,
Scope* scope,
const std::unordered_set<std::string>& fake_dequantize_types,
Expand Down Expand Up @@ -94,7 +94,7 @@ void QuantDequantMkldnnPass::CollectInfoFromFake(
}
}

void QuantDequantMkldnnPass::CollectWeightScalesInfoFromONNXFormatDequantize(
void QuantDequantOnednnPass::CollectWeightScalesInfoFromONNXFormatDequantize(
ir::Graph* graph,
Scope* scope,
std::unordered_map<std::string, std::vector<float>>* weight_thresholds,
Expand Down Expand Up @@ -143,7 +143,7 @@ void QuantDequantMkldnnPass::CollectWeightScalesInfoFromONNXFormatDequantize(
}
}

void QuantDequantMkldnnPass::CollectInputScalesFromQuantize(
void QuantDequantOnednnPass::CollectInputScalesFromQuantize(
ir::Graph* graph,
Scope* scope,
const std::unordered_set<std::string>& fake_quantize_types,
Expand Down Expand Up @@ -203,7 +203,7 @@ void QuantDequantMkldnnPass::CollectInputScalesFromQuantize(
}
}

void QuantDequantMkldnnPass::CollectOutputScalesFromAttr(
void QuantDequantOnednnPass::CollectOutputScalesFromAttr(
ir::Graph* graph,
std::unordered_map<std::string, std::vector<float>>* var_quant_scales)
const {
Expand All @@ -230,7 +230,7 @@ void QuantDequantMkldnnPass::CollectOutputScalesFromAttr(
}
}

void QuantDequantMkldnnPass::CollectFakeQuantizeOps(
void QuantDequantOnednnPass::CollectFakeQuantizeOps(
ir::Graph* graph,
Node* op_node,
std::unordered_set<const Node*>* nodes2rm) const {
Expand Down Expand Up @@ -284,7 +284,7 @@ void QuantDequantMkldnnPass::CollectFakeQuantizeOps(
nodes2rm->insert(fake_quant_out_scale);
}

void QuantDequantMkldnnPass::CollectFakeDequantizeOps(
void QuantDequantOnednnPass::CollectFakeDequantizeOps(
ir::Graph* graph,
Node* op_node,
std::unordered_set<const Node*>* nodes2rm) const {
Expand Down Expand Up @@ -329,7 +329,7 @@ void QuantDequantMkldnnPass::CollectFakeDequantizeOps(
nodes2rm->insert(fake_dequant_out);
}

void QuantDequantMkldnnPass::CollectQuantizeDequantizeOpsFromONNXFormat(
void QuantDequantOnednnPass::CollectQuantizeDequantizeOpsFromONNXFormat(
ir::Graph* graph,
Node* op_node,
std::unordered_set<const Node*>* nodes2rm) const {
Expand Down Expand Up @@ -382,7 +382,7 @@ void QuantDequantMkldnnPass::CollectQuantizeDequantizeOpsFromONNXFormat(
nodes2rm->insert(fake_quant_out);
}

void QuantDequantMkldnnPass::RemoveFakeOps(
void QuantDequantOnednnPass::RemoveFakeOps(
ir::Graph* graph,
const std::unordered_set<std::string>& fake_quantize_types,
const std::unordered_set<std::string>& fake_dequantize_types,
Expand All @@ -409,7 +409,7 @@ void QuantDequantMkldnnPass::RemoveFakeOps(
GraphSafeRemoveNodes(graph, nodes2rm);
}

void QuantDequantMkldnnPass::TransposeWeight(phi::DenseTensor* input) const {
void QuantDequantOnednnPass::TransposeWeight(phi::DenseTensor* input) const {
const auto in_dims = input->dims();
std::vector<int> out_dim_v;
std::vector<int> axis;
Expand Down Expand Up @@ -446,7 +446,7 @@ void QuantDequantMkldnnPass::TransposeWeight(phi::DenseTensor* input) const {
}
}

bool QuantDequantMkldnnPass::IsInt8Weight(
bool QuantDequantOnednnPass::IsInt8Weight(
Node* op_node, Scope* scope, const std::string& weight_name) const {
auto* op_desc = op_node->Op();
auto var_name = op_desc->Input(weight_name)[0];
Expand All @@ -466,7 +466,7 @@ bool QuantDequantMkldnnPass::IsInt8Weight(
return is_int8;
}

void QuantDequantMkldnnPass::ConvertFromINT8ToFP32(
void QuantDequantOnednnPass::ConvertFromINT8ToFP32(
const std::vector<float>& scales,
phi::DenseTensor* weight_tensor,
int8_t* int8_weight_data,
Expand Down Expand Up @@ -546,7 +546,7 @@ void QuantDequantMkldnnPass::ConvertFromINT8ToFP32(
weight_tensor->Resize(weight_dims);
}

void QuantDequantMkldnnPass::DequantizeOpWeights(
void QuantDequantOnednnPass::DequantizeOpWeights(
Node* op_node,
Scope* scope,
const std::string& weight_name,
Expand Down Expand Up @@ -581,7 +581,7 @@ void QuantDequantMkldnnPass::DequantizeOpWeights(
scales, weight_tensor, nullptr, fp32_weight_data, weight_var_name);
}

void QuantDequantMkldnnPass::DequantizeOpWeightsFromONNXFormat(
void QuantDequantOnednnPass::DequantizeOpWeightsFromONNXFormat(
Node* op_node,
Scope* scope,
const std::string& weight_name,
Expand Down Expand Up @@ -627,7 +627,7 @@ void QuantDequantMkldnnPass::DequantizeOpWeightsFromONNXFormat(
scales, weight_tensor, int8_weight_data, nullptr, weight_var_name);
}

void QuantDequantMkldnnPass::DequantizeWeights(
void QuantDequantOnednnPass::DequantizeWeights(
ir::Graph* graph,
Scope* scope,
const std::unordered_map<std::string, std::vector<float>>&
Expand Down Expand Up @@ -668,7 +668,7 @@ void QuantDequantMkldnnPass::DequantizeWeights(
}
}

void QuantDequantMkldnnPass::UpdateActivations(ir::Graph* graph) const {
void QuantDequantOnednnPass::UpdateActivations(ir::Graph* graph) const {
VLOG(3) << "update conv2d or depthwise_conv2d fused activation";
for (auto* op_node :
ir::TopologyVariantSort(*graph, static_cast<ir::SortKind>(0))) {
Expand All @@ -687,7 +687,7 @@ void QuantDequantMkldnnPass::UpdateActivations(ir::Graph* graph) const {
}
}

void QuantDequantMkldnnPass::RemoveCtrlVars(ir::Graph* graph) const {
void QuantDequantOnednnPass::RemoveCtrlVars(ir::Graph* graph) const {
VLOG(3) << "remove control flow variable";
std::unordered_set<const Node*> nodes2rm = {};
for (auto* op_node :
Expand All @@ -700,7 +700,7 @@ void QuantDequantMkldnnPass::RemoveCtrlVars(ir::Graph* graph) const {
GraphSafeRemoveNodes(graph, nodes2rm);
}

void QuantDequantMkldnnPass::ApplyImpl(ir::Graph* graph) const {
void QuantDequantOnednnPass::ApplyImpl(ir::Graph* graph) const {
VLOG(3) << "Convert paddle slim quantized model to onednn quantized model.";
const std::string pattern_name = "quant_dequant_onednn_pass";
FusePassBase::Init(pattern_name, graph);
Expand Down Expand Up @@ -759,7 +759,7 @@ void QuantDequantMkldnnPass::ApplyImpl(ir::Graph* graph) const {
} // namespace paddle::framework::ir

REGISTER_PASS(quant_dequant_onednn_pass,
paddle::framework::ir::QuantDequantMkldnnPass);
paddle::framework::ir::QuantDequantOnednnPass);

REGISTER_PASS_CAPABILITY(quant_dequant_onednn_pass)
.AddCombination(
Expand Down
6 changes: 3 additions & 3 deletions paddle/fluid/framework/ir/onednn/quant_dequant_onednn_pass.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,10 +22,10 @@ namespace paddle {
namespace framework {
namespace ir {

class QuantDequantMkldnnPass : public FusePassBase {
class QuantDequantOnednnPass : public FusePassBase {
public:
QuantDequantMkldnnPass() = default;
virtual ~QuantDequantMkldnnPass() {}
QuantDequantOnednnPass() = default;
virtual ~QuantDequantOnednnPass() {}

protected:
void ApplyImpl(ir::Graph* graph) const override;
Expand Down
18 changes: 9 additions & 9 deletions test/cpp/fluid/framework/ir/onednn/compute_propagate_scales_onednn_pass_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,10 +52,10 @@ static const std::initializer_list<std::string> conv_variable_names{
static const std::initializer_list<std::string> rnn_variable_names{
"x", "wx", "wh", "b", "h", "c"};

class ComputePropagateScalesMkldnnPassTest : public testing::Test {
class ComputePropagateScalesOnednnPassTest : public testing::Test {
public:
ComputePropagateScalesMkldnnPassTest() { // NOLINT
pass = std::make_unique<ComputePropagateScalesMkldnnPass>();
ComputePropagateScalesOnednnPassTest() { // NOLINT
pass = std::make_unique<ComputePropagateScalesOnednnPass>();
}

std::vector<float> GetScales(phi::DenseTensor* tensor, int axis) const {
Expand Down Expand Up @@ -195,7 +195,7 @@ class ComputePropagateScalesMkldnnPassTest : public testing::Test {
}

private:
std::unique_ptr<ComputePropagateScalesMkldnnPass> pass;
std::unique_ptr<ComputePropagateScalesOnednnPass> pass;
};

void SetOp(ProgramDesc* prog,
Expand Down Expand Up @@ -272,7 +272,7 @@ ProgramDesc BuildFusionLstmProgramDesc() {
return prog;
}

TEST_F(ComputePropagateScalesMkldnnPassTest, get_scales_function) {
TEST_F(ComputePropagateScalesOnednnPassTest, get_scales_function) {
const auto& values = positive_and_negative_values;
float max_val = *std::max_element(values.begin(), values.end());

Expand All @@ -287,7 +287,7 @@ TEST_F(ComputePropagateScalesMkldnnPassTest, get_scales_function) {
ASSERT_EQ(results[0], (1.f / max_val));
}

TEST_F(ComputePropagateScalesMkldnnPassTest, compute_var_scales) {
TEST_F(ComputePropagateScalesOnednnPassTest, compute_var_scales) {
auto prog = BuildConv2dProgramDesc();
const auto& values = positive_and_negative_values;
ir::Graph* graph(new ir::Graph(prog));
Expand Down Expand Up @@ -323,15 +323,15 @@ TEST_F(ComputePropagateScalesMkldnnPassTest, compute_var_scales) {
ASSERT_FLOAT_EQ(result_tensor.data<float>()[0], (1.0 / max_val));
}

TEST_F(ComputePropagateScalesMkldnnPassTest, compute_gru_weight_scales) {
TEST_F(ComputePropagateScalesOnednnPassTest, compute_gru_weight_scales) {
ComputeRnnWeightScalesTest("gru", BuildFusionGruProgramDesc(), gru_scales);
}

TEST_F(ComputePropagateScalesMkldnnPassTest, compute_lstm_weight_scales) {
TEST_F(ComputePropagateScalesOnednnPassTest, compute_lstm_weight_scales) {
ComputeRnnWeightScalesTest("lstm", BuildFusionLstmProgramDesc(), lstm_scales);
}

TEST_F(ComputePropagateScalesMkldnnPassTest, update_relu_output_scales) {
TEST_F(ComputePropagateScalesOnednnPassTest, update_relu_output_scales) {
StringPairMap var_quant_scales;
for (auto& var_name : conv_variable_names) {
phi::DenseTensor tensor;
Expand Down
Loading
Loading