提交 01158763 编写于 作者: M mindspore-ci-bot 提交者: Gitee

!4197 Add AdamApplyOneAssign fusion pass

Merge pull request !4197 from YuJianfeng/adam_assign
......@@ -125,6 +125,10 @@ void AddAscendIRFusionRulesPass(PassManager *ir_fusion_pm) {
ir_fusion_pm->AddPass(std::make_shared<LambNextMVRuleCond4>());
ir_fusion_pm->AddPass(std::make_shared<LambNextRightRule>());
ir_fusion_pm->AddPass(std::make_shared<LambUpdateWithLrV2>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneAssignCond1Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneAssignCond2Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneAssignCond3Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneAssignCond4Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond1Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond2Fusion>());
ir_fusion_pm->AddPass(std::make_shared<AdamApplyOneCond3Fusion>());
......
......@@ -15,30 +15,9 @@
*/
#include "backend/optimizer/ascend/ir_fusion/adam_apply_one_fusion.h"
#include "backend/optimizer/common/helper.h"
#include "backend/session/anf_runtime_algorithm.h"
namespace mindspore {
namespace opt {
AnfNodePtr AdamApplyOneFusion::CreateAdamApplyOneNode(const FuncGraphPtr &func_graph, const EquivPtr &equiv) const {
MS_EXCEPTION_IF_NULL(func_graph);
MS_EXCEPTION_IF_NULL(equiv);
auto prim = std::make_shared<Primitive>(kAdamApplyOneOpName);
std::vector<AnfNodePtr> new_node_inputs = {NewValueNode(prim)};
for (const auto &input_var : input_vars_) {
auto input_node = utils::cast<AnfNodePtr>((*equiv)[input_var]);
MS_EXCEPTION_IF_NULL(input_node);
new_node_inputs.push_back(input_node);
}
for (const auto &mul_x_input_var : mul_x_input_vars_) {
auto mul_x_input_node = utils::cast<AnfNodePtr>((*equiv)[mul_x_input_var]);
MS_EXCEPTION_IF_NULL(mul_x_input_node);
new_node_inputs.push_back(mul_x_input_node);
}
auto add2_y_node = utils::cast<AnfNodePtr>((*equiv)[add2_y_]);
MS_EXCEPTION_IF_NULL(add2_y_node);
new_node_inputs.push_back(add2_y_node);
auto new_node = func_graph->NewCNode(new_node_inputs);
return new_node;
}
const BaseRef AdamApplyOneFusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
......@@ -104,16 +83,152 @@ const BaseRef AdamApplyOneCond4Fusion::DefinePattern() const {
return VectorRef({prim::kPrimSub, input_vars_[3], VectorRef({prim::kPrimMul, true_div0, input_vars_[4]})});
}
const BaseRef AdamApplyOneAssignFusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
VectorRef mul2 = VectorRef({prim::kPrimMul, mul_x_input_vars_[2], input_vars_[1]});
VectorRef mul3 = VectorRef({prim::kPrimMul, mul_x_input_vars_[3], VectorRef({prim::kPrimSquare, input_vars_[0]})});
VectorRef add1 = VectorRef({add1_var_, mul2, mul3});
VectorRef sqrt0 = VectorRef({prim_sqrt, add1});
VectorRef mul1 = VectorRef({prim::kPrimMul, mul_x_input_vars_[1], input_vars_[0]});
VectorRef mul0 = VectorRef({prim::kPrimMul, mul_x_input_vars_[0], input_vars_[2]});
VectorRef add0 = VectorRef({add0_var_, mul0, mul1});
VectorRef true_div0 = VectorRef({prim_real_div, add0, VectorRef({prim::kPrimTensorAdd, sqrt0, add2_y_})});
VectorRef sub0 = VectorRef({sub0_var_, input_vars_[3], VectorRef({prim::kPrimMul, input_vars_[4], true_div0})});
VectorRef assign0 = VectorRef({prim::kPrimAssign, input_vars_[3], sub0});
VectorRef depend0 = VectorRef({prim::kPrimDepend, sub0, assign0});
VectorRef assign1 = VectorRef({prim::kPrimAssign, input_vars_[2], add0});
VectorRef depend1 = VectorRef({prim::kPrimDepend, depend0, assign1});
VectorRef assign2 = VectorRef({prim::kPrimAssign, input_vars_[1], add1});
return VectorRef({prim::kPrimDepend, depend1, assign2});
}
const BaseRef AdamApplyOneAssignCond1Fusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
VectorRef mul2 = VectorRef({prim::kPrimMul, mul_x_input_vars_[2], input_vars_[1]});
VectorRef mul3 = VectorRef({prim::kPrimMul, mul_x_input_vars_[3], VectorRef({prim::kPrimSquare, input_vars_[0]})});
VectorRef add1 = VectorRef({add1_var_, mul2, mul3});
VectorRef sqrt0 = VectorRef({prim_sqrt, add1});
VectorRef mul1 = VectorRef({prim::kPrimMul, mul_x_input_vars_[1], input_vars_[0]});
VectorRef mul0 = VectorRef({prim::kPrimMul, mul_x_input_vars_[0], input_vars_[2]});
VectorRef add0 = VectorRef({add0_var_, mul0, mul1});
VectorRef true_div0 = VectorRef({prim_real_div, add0, VectorRef({prim::kPrimTensorAdd, add2_y_, sqrt0})});
VectorRef sub0 = VectorRef({sub0_var_, input_vars_[3], VectorRef({prim::kPrimMul, input_vars_[4], true_div0})});
VectorRef assign0 = VectorRef({prim::kPrimAssign, input_vars_[3], sub0});
VectorRef depend0 = VectorRef({prim::kPrimDepend, sub0, assign0});
VectorRef assign1 = VectorRef({prim::kPrimAssign, input_vars_[2], add0});
VectorRef depend1 = VectorRef({prim::kPrimDepend, depend0, assign1});
VectorRef assign2 = VectorRef({prim::kPrimAssign, input_vars_[1], add1});
return VectorRef({prim::kPrimDepend, depend1, assign2});
}
const BaseRef AdamApplyOneAssignCond2Fusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
VectorRef mul2 = VectorRef({prim::kPrimMul, mul_x_input_vars_[2], input_vars_[1]});
VectorRef mul3 = VectorRef({prim::kPrimMul, VectorRef({prim::kPrimSquare, input_vars_[0]}), mul_x_input_vars_[3]});
VectorRef add1 = VectorRef({add1_var_, mul2, mul3});
VectorRef sqrt0 = VectorRef({prim_sqrt, add1});
VectorRef mul1 = VectorRef({prim::kPrimMul, mul_x_input_vars_[1], input_vars_[0]});
VectorRef mul0 = VectorRef({prim::kPrimMul, mul_x_input_vars_[0], input_vars_[2]});
VectorRef add0 = VectorRef({add0_var_, mul0, mul1});
VectorRef true_div0 = VectorRef({prim_real_div, add0, VectorRef({prim::kPrimTensorAdd, sqrt0, add2_y_})});
VectorRef sub0 = VectorRef({sub0_var_, input_vars_[3], VectorRef({prim::kPrimMul, true_div0, input_vars_[4]})});
VectorRef assign0 = VectorRef({prim::kPrimAssign, input_vars_[3], sub0});
VectorRef depend0 = VectorRef({prim::kPrimDepend, sub0, assign0});
VectorRef assign1 = VectorRef({prim::kPrimAssign, input_vars_[2], add0});
VectorRef depend1 = VectorRef({prim::kPrimDepend, depend0, assign1});
VectorRef assign2 = VectorRef({prim::kPrimAssign, input_vars_[1], add1});
return VectorRef({prim::kPrimDepend, depend1, assign2});
}
const BaseRef AdamApplyOneAssignCond3Fusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
VectorRef mul2 = VectorRef({prim::kPrimMul, mul_x_input_vars_[2], input_vars_[1]});
VectorRef mul3 = VectorRef({prim::kPrimMul, mul_x_input_vars_[3], VectorRef({prim::kPrimSquare, input_vars_[0]})});
VectorRef add1 = VectorRef({add1_var_, mul2, mul3});
VectorRef sqrt0 = VectorRef({prim_sqrt, add1});
VectorRef mul1 = VectorRef({prim::kPrimMul, mul_x_input_vars_[1], input_vars_[0]});
VectorRef mul0 = VectorRef({prim::kPrimMul, mul_x_input_vars_[0], input_vars_[2]});
VectorRef add0 = VectorRef({add0_var_, mul0, mul1});
VectorRef true_div0 = VectorRef({prim_real_div, add0, VectorRef({prim::kPrimTensorAdd, sqrt0, add2_y_})});
VectorRef sub0 = VectorRef({sub0_var_, input_vars_[3], VectorRef({prim::kPrimMul, true_div0, input_vars_[4]})});
VectorRef assign0 = VectorRef({prim::kPrimAssign, input_vars_[3], sub0});
VectorRef depend0 = VectorRef({prim::kPrimDepend, sub0, assign0});
VectorRef assign1 = VectorRef({prim::kPrimAssign, input_vars_[2], add0});
VectorRef depend1 = VectorRef({prim::kPrimDepend, depend0, assign1});
VectorRef assign2 = VectorRef({prim::kPrimAssign, input_vars_[1], add1});
return VectorRef({prim::kPrimDepend, depend1, assign2});
}
const BaseRef AdamApplyOneAssignCond4Fusion::DefinePattern() const {
const auto prim_sqrt = std::make_shared<Primitive>(kSqrtOpName);
const auto prim_real_div = std::make_shared<Primitive>(kRealDivOpName);
VectorRef mul2 = VectorRef({prim::kPrimMul, mul_x_input_vars_[2], input_vars_[1]});
VectorRef mul3 = VectorRef({prim::kPrimMul, mul_x_input_vars_[3], VectorRef({prim::kPrimSquare, input_vars_[0]})});
VectorRef add1 = VectorRef({add1_var_, mul2, mul3});
VectorRef sqrt0 = VectorRef({prim_sqrt, add1});
VectorRef mul1 = VectorRef({prim::kPrimMul, mul_x_input_vars_[1], input_vars_[0]});
VectorRef mul0 = VectorRef({prim::kPrimMul, mul_x_input_vars_[0], input_vars_[2]});
VectorRef add0 = VectorRef({add0_var_, mul0, mul1});
VectorRef true_div0 = VectorRef({prim_real_div, add0, VectorRef({prim::kPrimTensorAdd, add2_y_, sqrt0})});
VectorRef sub0 = VectorRef({sub0_var_, input_vars_[3], VectorRef({prim::kPrimMul, true_div0, input_vars_[4]})});
VectorRef assign0 = VectorRef({prim::kPrimAssign, input_vars_[3], sub0});
VectorRef depend0 = VectorRef({prim::kPrimDepend, sub0, assign0});
VectorRef assign1 = VectorRef({prim::kPrimAssign, input_vars_[2], add0});
VectorRef depend1 = VectorRef({prim::kPrimDepend, depend0, assign1});
VectorRef assign2 = VectorRef({prim::kPrimAssign, input_vars_[1], add1});
return VectorRef({prim::kPrimDepend, depend1, assign2});
}
AnfNodePtr AdamApplyOneFusion::CreateAdamApplyOneNode(const FuncGraphPtr &func_graph, const EquivPtr &equiv,
const AnfNodePtr &final_node) const {
MS_EXCEPTION_IF_NULL(func_graph);
MS_EXCEPTION_IF_NULL(equiv);
PrimitivePtr prim = nullptr;
if (AnfAlgo::CheckPrimitiveType(final_node, prim::kPrimDepend)) {
prim = std::make_shared<Primitive>(kAdamApplyOneAssignOpName);
} else {
prim = std::make_shared<Primitive>(kAdamApplyOneOpName);
}
std::vector<AnfNodePtr> new_node_inputs = {NewValueNode(prim)};
for (const auto &input_var : input_vars_) {
auto input_node = utils::cast<AnfNodePtr>((*equiv)[input_var]);
MS_EXCEPTION_IF_NULL(input_node);
new_node_inputs.push_back(input_node);
}
for (const auto &mul_x_input_var : mul_x_input_vars_) {
auto mul_x_input_node = utils::cast<AnfNodePtr>((*equiv)[mul_x_input_var]);
MS_EXCEPTION_IF_NULL(mul_x_input_node);
new_node_inputs.push_back(mul_x_input_node);
}
auto add2_y_node = utils::cast<AnfNodePtr>((*equiv)[add2_y_]);
MS_EXCEPTION_IF_NULL(add2_y_node);
new_node_inputs.push_back(add2_y_node);
auto new_node = func_graph->NewCNode(new_node_inputs);
return new_node;
}
const AnfNodePtr AdamApplyOneFusion::Process(const FuncGraphPtr &func_graph, const AnfNodePtr &node,
const EquivPtr &equiv) const {
MS_EXCEPTION_IF_NULL(func_graph);
MS_EXCEPTION_IF_NULL(node);
if (!CheckSupportDataType(node, kFloatDataTypeSet)) {
auto sub0 = node;
if (AnfAlgo::CheckPrimitiveType(node, prim::kPrimDepend)) {
auto iter_sub0 = (*equiv).find(sub0_var_);
if (iter_sub0 == (*equiv).end()) {
MS_LOG(EXCEPTION) << "The equiv map is expected to contains the sub0 var after matched.";
}
sub0 = utils::cast<AnfNodePtr>(iter_sub0->second);
}
MS_EXCEPTION_IF_NULL(sub0);
if (!CheckSupportDataType(sub0, kFloatDataTypeSet)) {
return nullptr;
}
auto new_node = CreateAdamApplyOneNode(func_graph, equiv);
auto new_node = CreateAdamApplyOneNode(func_graph, equiv, node);
MS_EXCEPTION_IF_NULL(new_node);
new_node->set_scope(node->scope());
new_node->set_scope(sub0->scope());
// Set abstract of new node
AbstractBasePtrList new_node_abstract_list;
auto iter_add0 = (*equiv).find(add0_var_);
......@@ -130,7 +245,7 @@ const AnfNodePtr AdamApplyOneFusion::Process(const FuncGraphPtr &func_graph, con
MS_EXCEPTION_IF_NULL(add1);
new_node_abstract_list.push_back(add1->abstract());
new_node_abstract_list.push_back(add0->abstract());
new_node_abstract_list.push_back(node->abstract());
new_node_abstract_list.push_back(sub0->abstract());
auto abstract_tuple = std::make_shared<abstract::AbstractTuple>(new_node_abstract_list);
new_node->set_abstract(abstract_tuple);
// Create tuple_getitem node for outputs
......
......@@ -40,6 +40,7 @@ class AdamApplyOneFusion : public PatternProcessPass {
add2_y_ = std::make_shared<Var>();
add0_var_ = std::make_shared<Var>(std::make_shared<Primitive>(prim::kPrimTensorAdd->name()));
add1_var_ = std::make_shared<Var>(std::make_shared<Primitive>(prim::kPrimTensorAdd->name()));
sub0_var_ = std::make_shared<Var>(std::make_shared<Primitive>(prim::kPrimSub->name()));
}
~AdamApplyOneFusion() override = default;
......@@ -47,12 +48,14 @@ class AdamApplyOneFusion : public PatternProcessPass {
const AnfNodePtr Process(const FuncGraphPtr &, const AnfNodePtr &, const EquivPtr &) const override;
protected:
AnfNodePtr CreateAdamApplyOneNode(const FuncGraphPtr &func_graph, const EquivPtr &equiv) const;
AnfNodePtr CreateAdamApplyOneNode(const FuncGraphPtr &func_graph, const EquivPtr &equiv,
const AnfNodePtr &final_node) const;
std::vector<VarPtr> input_vars_;
std::vector<VarPtr> mul_x_input_vars_;
VarPtr add2_y_;
VarPtr add0_var_;
VarPtr add1_var_;
VarPtr sub0_var_;
};
class AdamApplyOneCond1Fusion : public AdamApplyOneFusion {
......@@ -90,6 +93,51 @@ class AdamApplyOneCond4Fusion : public AdamApplyOneFusion {
~AdamApplyOneCond4Fusion() override = default;
const BaseRef DefinePattern() const override;
};
class AdamApplyOneAssignFusion : public AdamApplyOneFusion {
public:
explicit AdamApplyOneAssignFusion(bool multigraph = true)
: AdamApplyOneFusion("adam_apply_one_assign_fusion", multigraph) {}
~AdamApplyOneAssignFusion() override = default;
const BaseRef DefinePattern() const override;
};
class AdamApplyOneAssignCond1Fusion : public AdamApplyOneFusion {
public:
explicit AdamApplyOneAssignCond1Fusion(bool multigraph = true)
: AdamApplyOneFusion("adam_apply_one_assign_cond1_fusion", multigraph) {}
~AdamApplyOneAssignCond1Fusion() override = default;
const BaseRef DefinePattern() const override;
};
class AdamApplyOneAssignCond2Fusion : public AdamApplyOneFusion {
public:
explicit AdamApplyOneAssignCond2Fusion(bool multigraph = true)
: AdamApplyOneFusion("adam_apply_one_assign_cond2_fusion", multigraph) {}
~AdamApplyOneAssignCond2Fusion() override = default;
const BaseRef DefinePattern() const override;
};
class AdamApplyOneAssignCond3Fusion : public AdamApplyOneFusion {
public:
explicit AdamApplyOneAssignCond3Fusion(bool multigraph = true)
: AdamApplyOneFusion("adam_apply_one_assign_cond3_fusion", multigraph) {}
~AdamApplyOneAssignCond3Fusion() override = default;
const BaseRef DefinePattern() const override;
};
class AdamApplyOneAssignCond4Fusion : public AdamApplyOneFusion {
public:
explicit AdamApplyOneAssignCond4Fusion(bool multigraph = true)
: AdamApplyOneFusion("adam_apply_one_assign_cond4_fusion", multigraph) {}
~AdamApplyOneAssignCond4Fusion() override = default;
const BaseRef DefinePattern() const override;
};
} // namespace opt
} // namespace mindspore
#endif // MINDSPORE_CCSRC_BACKEND_OPTIMIZER_ASCEND_IR_FUSION_ADAM_APPLY_ONE_FUSION_H_
......@@ -119,6 +119,7 @@ constexpr auto kAdamApplyOneWithDecayOpName = "AdamApplyOneWithDecay";
constexpr auto kBatchNormGradOpName = "BatchNormGrad";
constexpr auto kBNInferOpName = "BNInfer";
constexpr auto kAdamApplyOneOpName = "AdamApplyOne";
constexpr auto kAdamApplyOneAssignOpName = "AdamApplyOneAssign";
constexpr auto kResizeNearestNeighborGradOpName = "ResizeNearestNeighborGrad";
constexpr auto kFusedMulAddOpName = "FusedMulAdd";
constexpr auto kFusedMulAddNOpName = "FusedMulAddN";
......
......@@ -217,5 +217,105 @@ TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_cond4_fusion) {
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_assign_fusion) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "before");
std::vector<int> shp{2, 32, 224, 224};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
for (size_t i = 0; i < 10; ++i) {
args_spec_list.push_back(x_abstract);
}
auto fg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
pm->AddPass(std::make_shared<opt::AdamApplyOneAssignFusion>());
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(fg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_assign_cond1_fusion) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "before_cond1");
std::vector<int> shp{2, 32, 224, 224};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
for (size_t i = 0; i < 10; ++i) {
args_spec_list.push_back(x_abstract);
}
auto fg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
pm->AddPass(std::make_shared<opt::AdamApplyOneAssignCond1Fusion>());
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(fg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_assign_cond2_fusion) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "before_cond2");
std::vector<int> shp{2, 32, 224, 224};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
for (size_t i = 0; i < 10; ++i) {
args_spec_list.push_back(x_abstract);
}
auto fg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
pm->AddPass(std::make_shared<opt::AdamApplyOneAssignCond2Fusion>());
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(fg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_assign_cond3_fusion) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "before_cond3");
std::vector<int> shp{2, 32, 224, 224};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
for (size_t i = 0; i < 10; ++i) {
args_spec_list.push_back(x_abstract);
}
auto fg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
pm->AddPass(std::make_shared<opt::AdamApplyOneAssignCond3Fusion>());
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(fg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
TEST_F(TestHWAdamApplyOneFusion, test_adam_apply_one_assign_cond4_fusion) {
FuncGraphPtr g = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "before_cond4");
std::vector<int> shp{2, 32, 224, 224};
auto x_abstract = std::make_shared<abstract::AbstractTensor>(kFloat32, shp);
AbstractBasePtrList args_spec_list;
for (size_t i = 0; i < 10; ++i) {
args_spec_list.push_back(x_abstract);
}
auto fg = GetKernelGraph(g, args_spec_list);
auto optimizer = std::make_shared<opt::GraphOptimizer>();
auto pm = std::make_shared<opt::PassManager>();
pm->AddPass(std::make_shared<opt::AdamApplyOneAssignCond4Fusion>());
optimizer->AddPassManager(pm);
FuncGraphPtr new_graph = optimizer->Optimize(fg);
FuncGraphPtr g_after = get_py_fun_.CallAndParseRet("test_adam_apply_one_assign_fusion", "after");
EXPECT_TRUE(CheckEqualGraph(g_after, new_graph));
}
} // namespace opt
} // namespace mindspore
......@@ -14,6 +14,7 @@
# ============================================================================
from mindspore.ops import Primitive
from mindspore.ops import operations as P
from mindspore.ops import functional as F
Add = P.TensorAdd()
Sub = P.Sub()
......@@ -21,9 +22,11 @@ Mul = P.Mul()
RealDiv = P.RealDiv()
Sqrt = P.Sqrt()
Square = P.Square()
Assign = P.Assign()
make_tuple = Primitive('make_tuple')
tuple_getitem = Primitive('tuple_getitem')
AdamApplyOne = Primitive('AdamApplyOne')
AdamApplyOneAssign = Primitive('AdamApplyOneAssign')
class FnDict:
......@@ -139,3 +142,138 @@ def test_adam_apply_one_fusion(tag):
return make_tuple(output)
return fns[tag]
def test_adam_apply_one_assign_fusion(tag):
fns = FnDict()
@fns
def before(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
square0 = Square(input0)
mul1 = Mul(mul1_x, input0)
mul0 = Mul(mul0_x, input2)
mul2 = Mul(mul2_x, input1)
mul3 = Mul(mul3_x, square0)
add0 = Add(mul0, mul1)
add1 = Add(mul2, mul3)
sqrt0 = Sqrt(add1)
add2 = Add(sqrt0, add2_y)
true_div0 = RealDiv(add0, add2)
mul4 = Mul(input4, true_div0)
sub0 = Sub(input3, mul4)
assign0 = Assign(input3, sub0)
depend0 = F.depend(sub0, assign0)
assign1 = Assign(input2, add0)
depend1 = F.depend(depend0, assign1)
assign2 = Assign(input1, add1)
depend2 = F.depend(depend1, assign2)
outputs = make_tuple(add1, add0, depend2)
output = tuple_getitem(outputs, 0)
return output
@fns
def before_cond1(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
square0 = Square(input0)
mul1 = Mul(mul1_x, input0)
mul0 = Mul(mul0_x, input2)
mul2 = Mul(mul2_x, input1)
mul3 = Mul(mul3_x, square0)
add0 = Add(mul0, mul1)
add1 = Add(mul2, mul3)
sqrt0 = Sqrt(add1)
add2 = Add(add2_y, sqrt0)
true_div0 = RealDiv(add0, add2)
mul4 = Mul(input4, true_div0)
sub0 = Sub(input3, mul4)
assign0 = Assign(input3, sub0)
depend0 = F.depend(sub0, assign0)
assign1 = Assign(input2, add0)
depend1 = F.depend(depend0, assign1)
assign2 = Assign(input1, add1)
depend2 = F.depend(depend1, assign2)
outputs = make_tuple(add1, add0, depend2)
output = tuple_getitem(outputs, 0)
return output
@fns
def before_cond2(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
square0 = Square(input0)
mul1 = Mul(mul1_x, input0)
mul0 = Mul(mul0_x, input2)
mul2 = Mul(mul2_x, input1)
mul3 = Mul(square0, mul3_x)
add0 = Add(mul0, mul1)
add1 = Add(mul2, mul3)
sqrt0 = Sqrt(add1)
add2 = Add(sqrt0, add2_y)
true_div0 = RealDiv(add0, add2)
mul4 = Mul(true_div0, input4)
sub0 = Sub(input3, mul4)
assign0 = Assign(input3, sub0)
depend0 = F.depend(sub0, assign0)
assign1 = Assign(input2, add0)
depend1 = F.depend(depend0, assign1)
assign2 = Assign(input1, add1)
depend2 = F.depend(depend1, assign2)
outputs = make_tuple(add1, add0, depend2)
output = tuple_getitem(outputs, 0)
return output
@fns
def before_cond3(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
square0 = Square(input0)
mul1 = Mul(mul1_x, input0)
mul0 = Mul(mul0_x, input2)
mul2 = Mul(mul2_x, input1)
mul3 = Mul(mul3_x, square0)
add0 = Add(mul0, mul1)
add1 = Add(mul2, mul3)
sqrt0 = Sqrt(add1)
add2 = Add(sqrt0, add2_y)
true_div0 = RealDiv(add0, add2)
mul4 = Mul(true_div0, input4)
sub0 = Sub(input3, mul4)
assign0 = Assign(input3, sub0)
depend0 = F.depend(sub0, assign0)
assign1 = Assign(input2, add0)
depend1 = F.depend(depend0, assign1)
assign2 = Assign(input1, add1)
depend2 = F.depend(depend1, assign2)
outputs = make_tuple(add1, add0, depend2)
output = tuple_getitem(outputs, 0)
return output
@fns
def before_cond4(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
square0 = Square(input0)
mul1 = Mul(mul1_x, input0)
mul0 = Mul(mul0_x, input2)
mul2 = Mul(mul2_x, input1)
mul3 = Mul(mul3_x, square0)
add0 = Add(mul0, mul1)
add1 = Add(mul2, mul3)
sqrt0 = Sqrt(add1)
add2 = Add(add2_y, sqrt0)
true_div0 = RealDiv(add0, add2)
mul4 = Mul(true_div0, input4)
sub0 = Sub(input3, mul4)
assign0 = Assign(input3, sub0)
depend0 = F.depend(sub0, assign0)
assign1 = Assign(input2, add0)
depend1 = F.depend(depend0, assign1)
assign2 = Assign(input1, add1)
depend2 = F.depend(depend1, assign2)
outputs = make_tuple(add1, add0, depend2)
output = tuple_getitem(outputs, 0)
return output
@fns
def after(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x, mul3_x, add2_y):
adam_apply_one_assign = AdamApplyOneAssign(input0, input1, input2, input3, input4, mul0_x, mul1_x, mul2_x,
mul3_x, add2_y)
outputs = make_tuple(tuple_getitem(adam_apply_one_assign, 0), tuple_getitem(adam_apply_one_assign, 1),
tuple_getitem(adam_apply_one_assign, 2))
output = tuple_getitem(outputs, 0)
return make_tuple(output)
return fns[tag]
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册