OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

[JIT] log extract tool - dump NVFuser fallbacks instead of fusion groups (#73881)

Browse Source 
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73881

NVFuser fusion groups can contain nvfuser-only ops, e.g. `prim::reshape_copy`. Previously, we couldn't get a baseline performance measurement because the nvfuser-only ops would error out on nnc- and no-fusion- runs. Instead, dump the fallback graphs, after the fallbacks are corrected into runnable fallbacks.

Test Plan: Imported from OSS

Reviewed By: eellison

Differential Revision: D34698307

Pulled By: davidberard98

fbshipit-source-id: c357b2736b789bfd347afe9c83a1b610b64881e0
(cherry picked from commit 5918d826502ff75fbc22d242844ae6435dd7d22a)
This commit is contained in:
David Berard 2022-03-08 08:31:53 -08:00 committed by PyTorch MergeBot
parent 56164c07c4
commit 31b64fc3e6
2 changed files with 21 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
scripts/jit/log_extract.py
View File

@ -2,7 +2,9 @@ from contextlib import contextmanager
from torch.testing import make_tensor from torch.testing import make_tensor
from typing import Any, List, Tuple from typing import Any, List, Tuple
import argparse import argparse
import random
import torch import torch
import traceback
''' '''
Usage: Usage:
@ -52,9 +54,9 @@ def load_graph_and_inputs(ir: str) -> Tuple[Any, List[Any]]:
inputs = [] inputs = []
for inp in graph.inputs(): for inp in graph.inputs():
if isinstance(inp.type(), torch._C.FloatType): if isinstance(inp.type(), torch._C.FloatType):
inputs.append(.5) inputs.append(random.uniform(.1, 100))
elif isinstance(inp.type(), torch._C.IntType): elif isinstance(inp.type(), torch._C.IntType):
inputs.append(2) inputs.append(random.randint(1, 100))
elif isinstance(inp.type(), torch._C.TensorType): elif isinstance(inp.type(), torch._C.TensorType):
inputs.append(make_tensor_from_type(inp.type())) inputs.append(make_tensor_from_type(inp.type()))
else: else:
@ -123,10 +125,13 @@ def run_nvfuser(ir, inputs) -> float:
def test_nvfuser(graphs: List[str], baseline_fn, nvfuser_fn): def test_nvfuser(graphs: List[str], baseline_fn, nvfuser_fn):
for i, ir in enumerate(graphs): for i, ir in enumerate(graphs):
_, inputs = load_graph_and_inputs(ir) _, inputs = load_graph_and_inputs(ir)
baseline = baseline_fn(ir, inputs) try:
nvfuser = nvfuser_fn(ir, inputs) baseline = baseline_fn(ir, inputs)
improvement = (baseline / nvfuser - 1) * 100 nvfuser = nvfuser_fn(ir, inputs)
print(f" Graph {i}; baseline: {baseline:.2f} ms; nvfuser: {nvfuser:.2f} ms; improvement: {improvement:.2f}%") improvement = (baseline / nvfuser - 1) * 100
print(f" Graph {i}; baseline: {baseline:.2f} ms; nvfuser: {nvfuser:.2f} ms; improvement: {improvement:.2f}%")
except RuntimeError:
print(f" Graph {i} failed:", traceback.format_exc())
def run(): def run():

14
torch/csrc/jit/codegen/cuda/graph_fuser.cpp
View File

@ -1710,11 +1710,15 @@ void guardFusionGroups(
// c. restore conditional constant to non-constant for fallback // c. restore conditional constant to non-constant for fallback
guardFusionGroup(fusion, fusion_value_to_runtime_size); guardFusionGroup(fusion, fusion_value_to_runtime_size);
} }
}
if (GRAPH_DEBUG_ENABLED) { void dumpFusionGroups(std::shared_ptr<Graph>& g) {
GRAPH_DEBUG("Exporting all NVFuser fusions:"); DepthFirstGraphNodeIterator it(g);
for (Node* fusion : fusions) { Node* n = nullptr;
GRAPH_EXPORT("", fusion->g(attr::Subgraph)); GRAPH_DEBUG("Exporting all NVFuser fusions:");
while ((n = it.next()) != nullptr) {
if (n->kind() == prim::FallbackGraph) {
GRAPH_EXPORT("", n->g(attr::Subgraph));
} }
} }
} }
@ -2305,6 +2309,8 @@ void CudaFuseGraph(std::shared_ptr<Graph>& graph) {
revertAliasCopyOps(graph, graph->block()); revertAliasCopyOps(graph, graph->block());
GRAPH_DEBUG("revert alias_copy ops by nvfuser: ", *graph); GRAPH_DEBUG("revert alias_copy ops by nvfuser: ", *graph);
dumpFusionGroups(graph);
// After FuseGraph some common subexpressions may come back // After FuseGraph some common subexpressions may come back
EliminateCommonSubexpression(graph); EliminateCommonSubexpression(graph);
// We might have emitted a fair amount of useless shape propagating code, so // We might have emitted a fair amount of useless shape propagating code, so