@@ -0,0 +1,33 @@ | |||
from ..core._imperative_rt import TensorSanityCheckImpl | |||
from ..core._imperative_rt.imperative import sync | |||
class TensorSanityCheck: | |||
r"""An object that checks whether the input tensors of each operator have changed before and after the operation. | |||
Examples: | |||
.. testcode:: | |||
from megengine import tensor | |||
from megengine.utils.tensor_sanity_check import TensorSanityCheck | |||
with TensorSanityCheck() as checker: | |||
a = tensor([1, 2]) | |||
b = tensor([3, 4]) | |||
c = a + b | |||
print(c.numpy()) | |||
.. testoutput:: | |||
[4 6] | |||
""" | |||
def __init__(self): | |||
self.impl = TensorSanityCheckImpl() | |||
def __enter__(self): | |||
sync() | |||
self.impl.enable() | |||
return self | |||
def __exit__(self, val, type, trace): | |||
sync() | |||
self.impl.disable() |
@@ -23,6 +23,7 @@ | |||
#include "megbrain/comp_node.h" | |||
#include "megbrain/imperative/blob_manager.h" | |||
#include "megbrain/imperative/profiler.h" | |||
#include "megbrain/imperative/tensor_sanity_check.h" | |||
#include "megbrain/serialization/helper.h" | |||
#if MGB_ENABLE_OPR_MM | |||
@@ -225,6 +226,19 @@ void init_utils(py::module m) { | |||
}, | |||
py::arg("path") = std::optional<std::string>()); | |||
using mgb::imperative::TensorSanityCheck; | |||
py::class_<TensorSanityCheck>(m, "TensorSanityCheckImpl") | |||
.def(py::init<>()) | |||
.def("enable", | |||
[](TensorSanityCheck& checker) -> TensorSanityCheck& { | |||
checker.enable(); | |||
return checker; | |||
}) | |||
.def("disable", | |||
[](TensorSanityCheck& checker) { | |||
checker.disable(); | |||
}); | |||
#if MGB_ENABLE_OPR_MM | |||
m.def("create_mm_server", &create_zmqrpc_server, py::arg("addr"), | |||
py::arg("port") = 0); | |||
@@ -110,9 +110,20 @@ SmallVector<TensorPtr> apply_on_physical_tensor( | |||
return out; | |||
} | |||
SmallVector<LogicalTensorDesc> infer_output_attrs( | |||
const OpDef& def, | |||
const SmallVector<TensorPtr>& inputs) { | |||
SmallVector<LogicalTensorDesc> out; | |||
for (size_t i = 0; i < 2; ++ i) { | |||
out.push_back({TensorLayout(), inputs[0]->comp_node()}); | |||
} | |||
return out; | |||
} | |||
OP_TRAIT_REG(CondTake, CondTake, opr::CondTake) | |||
.apply_on_var_node(apply_on_var_node) | |||
.apply_on_physical_tensor(apply_on_physical_tensor) | |||
.infer_output_attrs(infer_output_attrs) | |||
.fallback(); | |||
} // namespace | |||
@@ -0,0 +1,130 @@ | |||
/** | |||
* \file src/core/impl/imperative/tensor_sanity_check.cpp | |||
* | |||
* This file is part of MegBrain, a deep learning framework developed by Megvii. | |||
* | |||
* \copyright Copyright (c) 2014-2019 Megvii Inc. All rights reserved. | |||
* | |||
*/ | |||
#include "megbrain/imperative/tensor_sanity_check.h" | |||
#include "./op_trait.h" | |||
namespace mgb { | |||
namespace imperative { | |||
TensorChecksumCalc::ChecksumResult TensorChecksumCalc::calc(TensorPtr ptr) { | |||
auto&& dt = ptr->dev_tensor(); | |||
if (!dt.layout().total_nr_elems()) { | |||
static ChecksumResult empty_checksum; | |||
return empty_checksum; | |||
} | |||
auto span = dt.layout().span(); | |||
megdnn::TensorND tensor; | |||
tensor.raw_ptr = dt.raw_ptr() + span.low_byte; | |||
tensor.layout.init_contiguous_stride({span.dist_byte()}); | |||
tensor.layout.dtype = dtype::Byte(); | |||
DeviceTensorStorage* workspace; | |||
{ | |||
MGB_LOCK_GUARD(m_workspace_mtx); | |||
workspace = &m_workspace[std::this_thread::get_id()] | |||
.storage[ptr->comp_node()]; | |||
} | |||
auto comp_node = ptr->comp_node(); | |||
comp_node.activate(); | |||
auto opr = opr::intl::get_megdnn_global_opr<megdnn::Checksum>(comp_node); | |||
auto workspace_reqsize = opr->get_workspace_in_bytes(tensor.layout); | |||
workspace->comp_node(ptr->comp_node()).ensure_size(workspace_reqsize); | |||
megdnn::Workspace mwk; | |||
if (workspace_reqsize) | |||
mwk = {workspace->ptr(), workspace_reqsize}; | |||
return opr->exec(tensor, mwk); | |||
} | |||
class TensorSanityCheckImpl { | |||
public: | |||
std::vector<std::tuple<OpTrait*, std::unique_ptr<ApplyOnPhysicalTensor>>> | |||
hook_list; | |||
std::unordered_map<TensorPtr, TensorChecksumCalc::ChecksumResult> | |||
tensor2chksum; // TODO: may increase device memory overhead | |||
TensorSanityCheckImpl() { | |||
m_calc = std::make_unique<TensorChecksumCalc>(); | |||
} | |||
bool check(TensorPtr p); | |||
private: | |||
std::unique_ptr<TensorChecksumCalc> m_calc; | |||
}; | |||
bool TensorSanityCheckImpl::check(TensorPtr p) { | |||
auto&& it = tensor2chksum.find(p); | |||
auto&& chksum = m_calc->calc(p); | |||
if (it == tensor2chksum.end()) { | |||
tensor2chksum[p] = chksum; | |||
return true; | |||
} | |||
return it->second == chksum; | |||
} | |||
void TensorSanityCheck::enable() { | |||
CompNode::sync_all(); | |||
OpTrait::for_each_trait([this](OpTrait& trait) { | |||
auto backup = std::make_unique<ApplyOnPhysicalTensor>( | |||
std::move(trait.apply_on_physical_tensor)); | |||
trait.apply_on_physical_tensor = [this, backup = backup.get()] ( | |||
const OpDef& def, const SmallVector<TensorPtr>& inputs) { | |||
for (auto&& i: inputs) { | |||
if (!m_checker->check(i)) { | |||
mgb_throw(TensorChecksumCalc::Error, | |||
"tensor modified before exec %s", print_op(def).c_str()); | |||
} | |||
} | |||
auto output = (*backup)(def, inputs); | |||
for (auto&& i: output) { | |||
mgb_assert(m_checker->check(i)); | |||
} | |||
for (auto&& i: inputs) { | |||
if (!m_checker->check(i)) { | |||
mgb_throw(TensorChecksumCalc::Error, | |||
"tensor modified after exec %s", print_op(def).c_str()); | |||
} | |||
} | |||
return output; | |||
}; | |||
m_checker->hook_list.push_back({&trait, std::move(backup)}); | |||
}); | |||
} | |||
void TensorSanityCheck::disable() { | |||
for (auto&& hook : m_checker->hook_list) { | |||
std::get<0>(hook)->apply_on_physical_tensor = | |||
std::move(*std::get<1>(hook)); | |||
} | |||
m_checker->tensor2chksum.clear(); | |||
m_checker->hook_list.clear(); | |||
} | |||
TensorSanityCheck::TensorSanityCheck() { | |||
m_checker = std::make_unique<TensorSanityCheckImpl>(); | |||
} | |||
TensorSanityCheck::~TensorSanityCheck () { | |||
} | |||
std::string TensorSanityCheck::print_op(const OpDef& def){ | |||
auto* opr_attr = def.try_cast_final<const OprAttr>(); | |||
if(opr_attr){ | |||
return std::string("OprAttr:") + opr_attr->type; | |||
} | |||
return def.dyn_typeinfo()->name; | |||
} | |||
} // namespace imperative | |||
} // namespace mgb |
@@ -0,0 +1,50 @@ | |||
/** | |||
* \file src/core/include/megbrain/tensor_sanity_check.h | |||
* | |||
* This file is part of MegBrain, a deep learning framework developed by Megvii. | |||
* | |||
* \copyright Copyright (c) 2014-2019 Megvii Inc. All rights reserved. | |||
* | |||
*/ | |||
#include "megbrain/comp_node_env.h" | |||
#include "megbrain/imperative/ops/opr_attr.h" | |||
#include "megbrain/imperative/op_def.h" | |||
#include "megbrain/plugin/var_sanity_check.h" | |||
#include "megbrain/opr/internal/megdnn_opr_wrapper.h" | |||
#include "megdnn/oprs/general.h" | |||
namespace mgb { | |||
namespace imperative { | |||
class TensorChecksumCalc { | |||
public: | |||
using ChecksumResult = megdnn::opr_result::Checksum; | |||
using Error = VarSanityCheckError; | |||
struct WorkspaceCache { | |||
//! var comp node to workspace | |||
CompNode::UnorderedMap<DeviceTensorStorage> storage; | |||
}; | |||
ThinHashMap<std::thread::id, WorkspaceCache> m_workspace; | |||
std::mutex m_workspace_mtx; | |||
ChecksumResult calc(TensorPtr ptr); | |||
TensorChecksumCalc() {} | |||
}; | |||
class TensorSanityCheckImpl; | |||
class TensorSanityCheck { | |||
public: | |||
TensorSanityCheck(); | |||
~TensorSanityCheck(); | |||
void enable(); | |||
void disable(); | |||
std::string print_op(const OpDef& def); | |||
private: | |||
std::unique_ptr<TensorSanityCheckImpl> m_checker; | |||
}; | |||
} // namespace imperative | |||
} // namespace mgb |