Col_lin
/
mindspore2022
forked from huawei/mindspore2022
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

!19993 [GraphKernel] Speed up bert performance in ascend for graph kernel 

Merge pull request !19993 from TronZhang/bert_ascend_gk
This commit is contained in:
i-robot 2021-07-27 08:15:07 +00:00 committed by Gitee
parent d444600923 94dda0c7c7
commit bca7891a70
7 changed files with 107 additions and 84 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
mindspore/_extends/graph_kernel/model/graph_split.py
View File

@ -24,6 +24,7 @@ class GraphSplitByPattern:
"""Graph splitter"""
class ReachTable:
"""Reachable table"""
def __init__(self, size):
self.map = []
self.alive = set(range(size))
@ -61,6 +62,7 @@ class GraphSplitByPattern:
class StitchInfo:
"""StitchInfo"""
def __init__(self):
self.stitch_ops = set()
self.stitch_atomic_ops = set()
@ -441,7 +443,6 @@ class GraphSplitByPattern:
self.orig_op_map.update(self.recom_area.ori_op_map)
self.recom_area.ori_op_map.clear()
def to_subgraph(self, dom):
"""Transform area to subgraphs"""
ids = self.index_op()
@ -555,6 +556,7 @@ class GraphSplitByPattern:
while do_recompute_fuse():
self.pattern_fuse()
use_poly_reduce = True
@ -830,8 +832,17 @@ class GraphSplitAscend(GraphSplitByPattern):
REDUCE_FUSE_DEPTH = 10
def get_default_mode(self, op):
if op.prim == "MatMul" or op.prim == "BatchMatMul":
return self.Area.MODE_COMPOSITE if op.inputs[0].dtype == "float16" else self.Area.MODE_BASIC
"""Get efault mode for op"""
def _dtype_same(tensors):
dtype = tensors[0].dtype
for tensor_ in tensors:
if tensor_.dtype != dtype:
return False
return True
if op.prim == "MatMul":
if op.inputs[0].dtype == "float16" and not _dtype_same(op.inputs):
return self.Area.MODE_COMPOSITE
if op.prim in ("Tile", "BroadcastTo", "ExpandDims"):
return self.Area.MODE_COMPOSITE
return self.Area.MODE_BASIC
@ -937,7 +948,10 @@ class GraphSplitAscend(GraphSplitByPattern):
return None
fused = []
for a, _ in dom.out_relations.items():
if a.pattern == PrimLib.ELEMWISE and a.check_acyclic(dom):
if (((a.dom_op().prim == "AddN" or a.dom_op().prim == "Add" or a.dom_op().prim == "Cast")
and dom.dom_op().prim == "MatMul")
or (a.pattern == PrimLib.ELEMWISE and dom.dom_op().prim == "BatchMatMul")) \
and a.check_acyclic(dom):
fused.append(a)
return fused, False

2
mindspore/ccsrc/backend/optimizer/ascend/ascend_backend_optimization.cc
View File

@ -454,9 +454,7 @@ void AscendBackendUBFusionOptimization(const std::shared_ptr<session::KernelGrap
ub_fusion_pm->AddPass(std::make_shared<ConvSingleInFusionPass>(fusion_id_allocator));
ub_fusion_pm->AddPass(std::make_shared<BnupdateEltwiseFusionPass>(fusion_id_allocator));
ub_fusion_pm->AddPass(std::make_shared<BnupdateEltwiseEltwiseFusionPass>(fusion_id_allocator));
if (!context::GraphKernelFlags::GetInstance().IsEnableGraphKernel()) {
ub_fusion_pm->AddPass(std::make_shared<MatmulEltwiseFusionPass>(fusion_id_allocator));
}
ub_fusion_pm->AddPass(std::make_shared<ConvDoubleInFusionPass>(fusion_id_allocator));
ub_fusion_pm->AddPass(std::make_shared<ReduceEltwiseFusionPass>(fusion_id_allocator));
ub_fusion_pm->AddPass(std::make_shared<SegmentEltwiseFusionPass>(fusion_id_allocator));

9
mindspore/ccsrc/backend/optimizer/ascend/buffer_fusion/matmul_eltwise_fusion_pass.cc
View File

@ -21,6 +21,7 @@
#include "debug/anf_ir_dump.h"
#include "backend/session/anf_runtime_algorithm.h"
#include "base/core_ops.h"
#include "utils/context/graph_kernel_flags.h"
#include "utils/ms_context.h"
#include "backend/optimizer/common/fusion_id_allocator.h"
@ -56,6 +57,14 @@ void MatmulEltwiseFusionPass::MatchSingleFusionPattern(const session::KernelGrap
continue;
}
auto cnode = node->cast<CNodePtr>();
if (context::GraphKernelFlags::GetInstance().IsEnableGraphKernel()) {
if (AnfAlgo::GetKernelType(cnode) == KernelType::TBE_KERNEL &&
AnfAlgo::GetFusionType(cnode) == kernel::FusionType::ELEMWISE &&
AnfAlgo::CheckPrimitiveType(cnode, prim::kPrimAddN)) {
continue;
}
}
MS_EXCEPTION_IF_NULL(cnode);
if (AnfAlgo::GetKernelType(cnode) == KernelType::TBE_KERNEL &&
AnfAlgo::GetFusionType(cnode) == kernel::FusionType::ELEMWISE &&

72
mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_cluster.cc
View File

@ -40,63 +40,63 @@ namespace {
std::vector<PrimitivePtr> GetClusterableOpList() {
std::vector<PrimitivePtr> clusterable_ops = {
prim::kPrimAbs,
prim::kPrimRound,
prim::kPrimNeg,
prim::kPrimExp,
prim::kPrimAdd,
prim::kPrimCast,
prim::kPrimMul,
prim::kPrimMinimum,
prim::kPrimMaximum,
prim::kPrimEqual,
prim::kPrimExp,
prim::kPrimInplaceAssign,
prim::kPrimLog,
prim::kPrimMaximum,
prim::kPrimMinimum,
prim::kPrimMul,
prim::kPrimNeg,
prim::kPrimPow,
prim::kPrimSub,
prim::kPrimRealDiv,
prim::kPrimReciprocal,
prim::kPrimReduceSum,
prim::kPrimReshape,
prim::kPrimRound,
prim::kPrimRsqrt,
prim::kPrimSqrt,
prim::kPrimReciprocal,
prim::kPrimSub,
prim::kPrimTanh,
prim::kPrimReshape,
prim::kPrimTranspose,
prim::kPrimRealDiv,
prim::kPrimReduceSum,
prim::kPrimEqual,
prim::kPrimAssign,
prim::kPrimInplaceAssign,
#if ENABLE_D
prim::kPrimMatMul,
prim::KPrimTransData,
prim::kPrimBatchMatMul,
#elif ENABLE_GPU
prim::kPrimSin,
prim::kPrimCos,
prim::kPrimAsin,
prim::kPrimACos,
prim::kPrimSign,
prim::kPrimReduceMax,
prim::kPrimReduceMin,
prim::kPrimGreater,
prim::kPrimLess,
prim::kPrimGreaterEqual,
prim::kPrimLessEqual,
prim::kPrimSelect,
prim::kPrimAcosh,
prim::kPrimAsin,
prim::kPrimAsinh,
prim::kPrimAssign,
prim::kPrimAtan,
prim::kPrimAtan2,
prim::kPrimExpm1,
prim::kPrimAsinh,
prim::kPrimAcosh,
prim::kPrimCos,
prim::kPrimDiv,
prim::kPrimFloorDiv,
prim::kPrimMod,
prim::kPrimFloor,
prim::kPrimFloorMod,
prim::kPrimErf,
prim::kPrimNotEqual,
prim::kPrimExpm1,
prim::kPrimFloor,
prim::kPrimFloorDiv,
prim::kPrimFloorMod,
prim::kPrimGreater,
prim::kPrimGreaterEqual,
prim::kPrimIsFinite,
prim::kPrimIsInf,
prim::kPrimIsNan,
prim::kPrimLess,
prim::kPrimLessEqual,
prim::kPrimLogicalAnd,
prim::kPrimLogicalOr,
prim::kPrimLogicalNot,
prim::kPrimIsNan,
prim::kPrimIsInf,
prim::kPrimIsFinite,
prim::kPrimMod,
prim::kPrimNotEqual,
prim::kPrimReduceMax,
prim::kPrimReduceMin,
prim::kPrimSelect,
prim::kPrimSign,
prim::kPrimSin,
#endif
};
const auto &flags = context::GraphKernelFlags::GetInstance();

41
mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_expander.cc
View File

@ -46,44 +46,43 @@ constexpr size_t kLambWeightInputIdx = 4;
std::vector<PrimitivePtr> GetExpandOps() {
std::vector<PrimitivePtr> expand_ops = {
prim::kPrimAddN,
prim::kPrimSquare,
prim::kPrimGeLUGrad,
prim::kPrimAssignAdd,
prim::kPrimLayerNorm,
prim::kPrimLayerNormGrad,
prim::kPrimExpandDims,
prim::kPrimBiasAddGrad,
prim::kPrimGeLU,
prim::kPrimSoftmax,
prim::kPrimLogSoftmax,
prim::kPrimLogSoftmaxGrad,
prim::kPrimTile,
prim::kPrimMatMul,
prim::kPrimBatchMatMul,
prim::kPrimErfc,
prim::kPrimExpandDims,
prim::kPrimGeLU,
prim::kPrimGeLUGrad,
prim::kPrimSquare,
prim::kPrimTile,
#if ENABLE_D
prim::kPrimSqrtGrad,
prim::kPrimClipByNormNoDivSum,
prim::kLambApplyOptimizerAssign,
prim::kLambApplyWeightAssign,
prim::kPrimClipByNormNoDivSum,
prim::kPrimSqrtGrad,
prim::kSoftmaxGradExt,
prim::kSquareSumV1,
prim::kFusedMulAdd,
#elif ENABLE_GPU
prim::kPrimBatchMatMul,
prim::kPrimBiasAdd,
prim::kPrimFusedAdam,
prim::kPrimFusedAdamWeightDecay,
prim::kPrimReduceMean,
prim::kPrimMaximumGrad,
prim::kPrimMinimumGrad,
prim::kPrimBiasAddGrad,
prim::kPrimDropout,
prim::kPrimDropoutGrad,
prim::kPrimFusedAdam,
prim::kPrimFusedAdamWeightDecay,
prim::kPrimMaximumGrad,
prim::kPrimMinimumGrad,
prim::kPrimLayerNorm,
prim::kPrimLayerNormGrad,
prim::kPrimLogSoftmax,
prim::kPrimLogSoftmaxGrad,
prim::kPrimMatMul,
prim::kPrimReduceMean,
prim::kPrimRelu,
prim::kPrimReluGrad,
prim::kPrimSigmoid,
prim::kPrimSigmoidGrad,
prim::kPrimSigmoidCrossEntropyWithLogits,
prim::kPrimSigmoidCrossEntropyWithLogitsGrad,
prim::kPrimSoftmax,
prim::kPrimSoftmaxCrossEntropyWithLogits,
prim::kPrimSquaredDifference,
prim::kPrimSqueeze,

2
mindspore/ccsrc/backend/optimizer/graph_kernel/graph_kernel_optimization.cc
View File

@ -61,7 +61,7 @@ PassManagerPtr GraphKernelOptimizer::PreProcess() const {
pm->AddPass(std::make_shared<CommonSubexpressionElimination>("cse1"), OptLevel_1);
// Change Assign(p, a, U) to Assign(Depend(p, U), a)
pm->AddPass(std::make_shared<SplitAssign>(), OptLevel_1);
pm->AddPass(std::make_shared<SplitAssign>(), OptLevel_1, is_gpu);
// Spread the MakeTuple input of UpdateState
pm->AddPass(std::make_shared<SpreadUpdateState>(), OptLevel_1);

3
model_zoo/official/nlp/bert/run_pretrain.py
View File

@ -62,6 +62,9 @@ def _set_bert_all_reduce_split():
context.set_auto_parallel_context(all_reduce_fusion_config=[30, 90, 150, 210, 270, 330, 390, 421])
else:
context.set_auto_parallel_context(all_reduce_fusion_config=[38, 93, 148, 203, 258, 313, 368, 397])
if device_target == 'Ascend' and enable_graph_kernel and device_num == 8:
context.set_auto_parallel_context(all_reduce_fusion_config=[
0, 1, 2, 3, 4, 5, 10, 15, 20, 25, 30, 35, 40, 50, 70, 93, 148, 203, 258, 313, 368, 397])
def _get_optimizer(args_opt, network):