提交 81833943 编写于 作者: M mindspore-ci-bot 提交者: Gitee

!2969 Eliminate AllReduce when the input is a constant in Graph mode

Merge pull request !2969 from BowenK/fix_reduce_all
......@@ -83,6 +83,8 @@ OptimizeIRPassLib::OptimizeIRPassLib() {
reset_defer_inline_ =
MakeSubstitution(std::make_shared<ResetDeferInline>(), "reset_defer_inline", IsValueNode<FuncGraph>);
depend_value_elim_ = MakeSubstitution(std::make_shared<DependValueElim>(), "depend_value_elim", prim::kPrimDepend);
all_reduce_const_elim_ =
MakeSubstitution(std::make_shared<AllReduceConstElim>(), "reduce_all_const_elim", prim::kPrimAllReduce);
// Env Item Eliminate
env_get_item_eliminate_ =
......
......@@ -50,6 +50,7 @@ class OptimizeIRPassLib {
SubstitutionPtr check_bprop_eliminate_;
SubstitutionPtr reset_defer_inline_;
SubstitutionPtr depend_value_elim_;
SubstitutionPtr all_reduce_const_elim_;
// Env Item Eliminate
SubstitutionPtr env_get_item_eliminate_;
......
......@@ -29,6 +29,8 @@
#include "frontend/optimizer/irpass.h"
#include "frontend/optimizer/irpass/prim_eliminate.h"
#include "frontend/optimizer/optimizer.h"
#include "utils/comm_manager.h"
#include "frontend/parallel/context.h"
namespace mindspore {
namespace opt {
......@@ -203,6 +205,57 @@ class DependValueElim : public OptimizerCaller {
return nullptr;
}
};
class AllReduceConstElim : public OptimizerCaller {
public:
AnfNodePtr operator()(const OptimizerPtr &, const AnfNodePtr &node) override {
PatternNode<AnfNodePtr> x;
auto pattern = PPrimitive(prim::kPrimAllReduce, x);
// If AllReduce takes contant value as input and values across devices are all the same(ensured by parallel mode)
if (pattern.TryCapture(node) && IsVNode(x.GetNode(node)) &&
(pattern.GetFuncGraph()->has_flag(parallel::AUTO_PARALLEL) ||
pattern.GetFuncGraph()->has_flag(parallel::SEMI_AUTO_PARALLEL))) {
auto cur_func_graph = pattern.GetFuncGraph();
// If reduce operation is sum, then multiply constant by number of devices, otherwise just return the contant
auto prim_cnode = pattern.GetOriginalNode();
MS_EXCEPTION_IF_NULL(prim_cnode);
auto primitive = GetCNodePrimitive(prim_cnode);
auto reduce_op = primitive->GetAttr("op");
auto group = primitive->GetAttr("group")->ToString();
// For sum operation, multiply constant tensor by number of devices
if (reduce_op->ToString() == "sum") {
unsigned int num_of_devices;
// Get number of devices
if (!CommManager::GetInstance().GetRankSize(group, &num_of_devices)) {
MS_LOG(EXCEPTION) << "Failed to get num of devices for group [" + group + "]";
}
// Multiply constant by number of devices then return
std::vector<AnfNodePtr> mul_inputs;
auto constant_node = x.GetNode(node);
MS_EXCEPTION_IF_NULL(constant_node);
auto constant_value_node = constant_node->cast<ValueNodePtr>();
MS_EXCEPTION_IF_NULL(constant_value_node);
if (!constant_value_node->value()->isa<tensor::Tensor>()) {
MS_LOG(EXCEPTION) << "Expect the constant input for AllReduce to be a Tensor. Got " +
constant_value_node->value()->ToString();
}
auto constant_tensor = constant_value_node->value()->cast<tensor::TensorPtr>();
auto tensor_dtype = constant_tensor->Dtype();
auto num_of_device_node = NewValueNode(std::make_shared<tensor::Tensor>((int64_t)num_of_devices, tensor_dtype));
// Multiply nodes
auto mul_prim = prim::GetPythonOps("tensor_mul", "mindspore.ops.functional");
MS_EXCEPTION_IF_NULL(mul_prim);
mul_inputs.push_back(NewValueNode(mul_prim));
mul_inputs.push_back(constant_node);
mul_inputs.push_back(num_of_device_node);
return cur_func_graph->NewCNode(mul_inputs);
} else {
return x.GetNode(node);
}
}
return nullptr;
}
};
} // namespace irpass
} // namespace opt
} // namespace mindspore
......
......@@ -133,6 +133,7 @@ OptPassGroupMap GetOptPassesA(const opt::irpass::OptimizeIRPassLib &irpass) {
irpass.incorporate_env_getitem_switch_,
irpass.new_env_get_item_,
irpass.depend_value_elim_,
irpass.all_reduce_const_elim_,
});
opt::OptPassConfig a_after_grad = opt::OptPassConfig({
irpass.inline_without_move_,
......
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