feat(dnn/check_non_finite): addmul scale to check_non_finite opr

GitOrigin-RevId: c35a219e52f852b37a311c6de505c6eb08e2ab8a

dnn/include/megdnn/oprs/general.h

@@ -1344,7 +1344,7 @@ protected:
  * \brief check whether input contains inf or nan value.
  */
 class CheckNonFinite : public OperatorBase {
-    DEF_OPR_PARAM(Empty);
+    DEF_OPR_PARAM(CheckNonFinite);
     DEF_OPR_IMPL(CheckNonFinite, OperatorBase, -1, 1);
     size_t m_size = 0;
 

dnn/scripts/opr_param_defs.py

@@ -1176,6 +1176,8 @@ Note: NCHW_NCHW4_WEIGHT will auto pad oc and ic, you should remove oc in later o
  )
 pdef('Fill').add_fields('float32', 'value', '0')
 
+pdef('CheckNonFinite').add_fields('float32', 'scale', '1.0')
+
 
 PADDING_MODES = [Doc('REPLICATE = 0', 'aaaaaa|abcdefgh|hhhhhhh'),
                 Doc('REFLECT = 1', 'fedcba|abcdefgh|hgfedcb'),

dnn/src/common/reduce_helper.h

@@ -156,37 +156,6 @@ struct MaxOp<src_ctype, dst_ctype, dt_float32> {
             : INIT(wtype(DTypeTrait<wtype>::min())), src(src), dst(dst), B(B) {}
 };
 
-template <typename src_ctype, typename index_ctype, typename dst_ctype, typename wtype_>
-struct CheckNonFiniteOp {
-    typedef wtype_ wtype;
-    const wtype INIT;
-
-    RefPtr* srcs;
-    RefPtr srcs_total_nr_elems;
-    RefPtr dst;
-    const size_t B;
-
-    wtype read(uint32_t idx) {
-        size_t x = idx / B;
-        size_t y = idx % B;
-        if (y < srcs_total_nr_elems.ptr<index_ctype>()[x]) {
-            RefPtr src = srcs[x];
-            return !std::isfinite(src.ptr<src_ctype>()[y]);
-        }
-        return 0;
-    }
-    void write(uint32_t idx, wtype val) { dst.ptr<dst_ctype>()[idx] = val; }
-    static wtype apply(wtype lhs, wtype rhs) { return lhs | rhs; }
-    CheckNonFiniteOp(
-            RefPtr* srcs, const RefPtr& srcs_total_nr_elems, const RefPtr& dst,
-            size_t B)
-            : INIT(wtype(0)),
-              srcs(srcs),
-              srcs_total_nr_elems(srcs_total_nr_elems),
-              dst(dst),
-              B(B) {}
-};
-
 void get_ABC(const TensorShape& shape, size_t& A, size_t& B, size_t& C, size_t axis);
 
 }  // namespace reduce

dnn/src/common/reduce_helper_device.h

@@ -194,6 +194,7 @@ struct CheckNonFiniteOp {
     index_ctype* srcs_total_nr_elems;
     dst_ctype* dst;
     const size_t B;
+    const src_ctype scale;
 
     MEGDNN_HOST MEGDNN_DEVICE wtype read(uint32_t idx) {
         size_t x = idx / B;
@@ -204,6 +205,8 @@ struct CheckNonFiniteOp {
 #else
             wtype val = std::isfinite(srcs[x][y]);
 #endif
+            if (val)
+                srcs[x][y] *= scale;
             return !val;
         }
         return 0;
@@ -214,12 +217,13 @@ struct CheckNonFiniteOp {
     }
     MEGDNN_HOST MEGDNN_DEVICE CheckNonFiniteOp(
             src_ctype** srcs, index_ctype* srcs_total_nr_elems, dst_ctype* dst,
-            size_t B)
+            size_t B, src_ctype scale)
             : INIT(wtype(0)),
               srcs(srcs),
               srcs_total_nr_elems(srcs_total_nr_elems),
               dst(dst),
-              B(B) {}
+              B(B),
+              scale(scale) {}
 };
 
 }  // namespace device_reduce

dnn/src/cuda/check_non_finite/opr_impl.cpp

@@ -97,7 +97,7 @@ void CheckNonFiniteImpl::exec(
                             workspace_gpu.total_size_in_bytes())),
             1, m_size * total_nr_elems_max, 1, stream,
             Op(srcs_gpu, srcs_total_nr_elems_gpu, dst.ptr<dt_int32>(),
-               total_nr_elems_max));
+               total_nr_elems_max, param().scale));
 }
 
 }  // namespace cuda

dnn/src/naive/check_non_finite/opr_impl.cpp

@@ -19,7 +19,7 @@ using namespace megdnn;
 
 #define wtype dt_int32
 
-void reduce_fwd(const TensorNDArray& srcs, wtype* dptr) {
+void reduce_fwd(const TensorNDArray& srcs, wtype* dptr, dt_float32 scale) {
     dptr[0] = 0;
     for (auto src : srcs) {
         auto sptr = src.ptr<dt_float32>();
@@ -31,6 +31,8 @@ void reduce_fwd(const TensorNDArray& srcs, wtype* dptr) {
                 return func(l, mid) | func(mid, r);
             } else {
                 auto val = std::isfinite(sptr[l]);
+                if (val)
+                    sptr[l] *= scale;
                 return static_cast<wtype>(!val);
             }
         };
@@ -47,9 +49,9 @@ void CheckNonFiniteImpl::exec(
         _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
         _megdnn_workspace workspace) {
     check_exec(srcs, dst, workspace.size);
-
+    float scale = param().scale;
     auto handle = static_cast<HandleImpl*>(this->handle());
-    MEGDNN_DISPATCH_CPU_KERN(handle, reduce_fwd(srcs, dst.ptr<dt_int32>()));
+    MEGDNN_DISPATCH_CPU_KERN(handle, reduce_fwd(srcs, dst.ptr<dt_int32>(), scale));
 }
 }  // namespace naive
 }  // namespace megdnn

imperative/python/megengine/amp/grad_scaler.py

@@ -128,16 +128,7 @@ class GradScaler:
             grad_tensors: Tensors needed to unscale grads. Should be all tensors
                 that are affected by ``target`` tensor in GradManager's backward.
         """
-        # to support tracing, _check_gradients should be applied to every grad.
-        if self._check_gradients([x.grad for x in grad_tensors]):
-            self._found_non_finite = True
-
-        if self._found_non_finite:
-            for tensor in grad_tensors:
-                if tensor is None or getattr(tensor, "grad", None) is None:
-                    continue
-                tensor.grad = None
-        else:
+        if self.growth_interval == 0:
             # use float64 for better precision
             inv_scale = Tensor(1.0 / self.scale_factor)
             for tensor in grad_tensors:
@@ -146,10 +137,19 @@ class GradScaler:
                 tensor.grad *= inv_scale
             return self
 
-    def _check_gradients(self, grad):
-        if self.growth_interval == 0:
-            return False
-        return _check_non_finite(grad)
+        # to support tracing, _check_gradients should be applied to every grad.
+        if self._check_gradients(
+            [x.grad for x in grad_tensors], 1.0 / self.scale_factor
+        ):
+            self._found_non_finite = True
+            for tensor in grad_tensors:
+                if tensor is None or getattr(tensor, "grad", None) is None:
+                    continue
+                tensor.grad = None
+        return self
+
+    def _check_gradients(self, grad, scale):
+        return _check_non_finite(grad, scale)
 
     def update(self, new_scale: float = None):
         r"""Update the scale factor according to whether encountered overflow grad.

imperative/python/megengine/functional/math.py

@@ -1183,7 +1183,7 @@ def svd(inp: Tensor, full_matrices=False, compute_uv=True) -> Tensor:
     return U, sigma, V
 
 
-def _check_non_finite(inps: Iterable[Tensor]) -> Tensor:
+def _check_non_finite(inps: Iterable[Tensor], scale=1.0) -> Tensor:
     r"""Check whether input contains infinite or nan value.
 
     Args:
@@ -1192,7 +1192,11 @@ def _check_non_finite(inps: Iterable[Tensor]) -> Tensor:
     Returns:
         a int32 scalar tensor, 0 for False and 1 for True.
     """
-    op = builtin.CheckNonFinite()
-    (oup,) = apply(op, *inps)
-    oup._setscalar()
-    return oup
+    op = builtin.CheckNonFinite(scale=scale)
+    oups = apply(op, *inps)
+    out = oups[-1]
+    for i in range(len(inps)):
+        inps[i]._reset(oups[i])
+
+    out._setscalar()
+    return out

imperative/python/test/unit/functional/test_math.py

@@ -191,17 +191,21 @@ def test_sum_neg_axis():
 
 def test_non_finite():
     shape = (32, 3, 32, 32)
-    data1 = np.random.random(shape).astype(np.float32)
-    data2 = np.random.random(shape).astype(np.float32)
-    rst = F.math._check_non_finite([tensor(data1), tensor(data2)])
+    data = []
+    for i in range(2):
+        data.append(np.random.random(shape).astype(np.float32))
+    tensorList = [tensor(x) for x in data]
+    rst = F.math._check_non_finite(tensorList, 0.7)
     np.testing.assert_equal(rst.numpy(), [0])
+    for i in range(len(tensorList)):
+        np.testing.assert_allclose(tensorList[i].numpy() / 0.7, data[i], rtol=1e-6)
 
-    data2[0][0][0][0] = float("inf")
-    rst = F.math._check_non_finite([tensor(data1), tensor(data2)])
+    data[1][0][0][0][0] = float("inf")
+    rst = F.math._check_non_finite([tensor(x) for x in data], 0.7)
     np.testing.assert_equal(rst.numpy(), [1])
 
-    data2[0][0][0][0] = float("nan")
-    rst = F.math._check_non_finite([tensor(data1), tensor(data2)])
+    data[1][0][0][0][0] = float("nan")
+    rst = F.math._check_non_finite([tensor(x) for x in data], 0.7)
     np.testing.assert_equal(rst.numpy(), [1])
 
 

imperative/src/impl/ops/misc.cpp

@@ -17,44 +17,62 @@ namespace mgb {
 namespace imperative {
 
 namespace check_non_finite {
-SymbolVar apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
+SymbolVarArray apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
     auto&& op = def.cast_final_safe<CheckNonFinite>();
     OperatorNodeConfig config{op.make_name()};
-    return opr::CheckNonFinite::make(inputs, {}, config);
+    return opr::CheckNonFinite::make(inputs, op.param(), config);
 }
 
 SmallVector<TensorPtr> apply_on_physical_tensor(
         const OpDef& def, const SmallVector<TensorPtr>& inputs) {
     size_t size = inputs.size();
-
-    auto dest = Tensor::make(
+    auto&& op = def.cast_final_safe<CheckNonFinite>();
+    SmallVector<TensorPtr> outputs(size + 1);
+    outputs[size] = Tensor::make(
             TensorLayout(TensorShape({1}), dtype::Int32()), inputs[0]->comp_node());
+
+    auto dest = outputs[size];
     auto cn = dest->comp_node();
     auto&& dnn_opr = opr::intl::create_megdnn_opr<megdnn::CheckNonFinite>(cn);
     size_t wk_size = 0;
     SmallVector<megdnn::TensorND> srcs(size);
+    // copy an outputs to the dnn for inplace
     for (size_t i = 0; i < size; ++i) {
-        srcs[i] = inputs[i]->dev_tensor().as_megdnn();
+        outputs[i] = Tensor::make(inputs[i]->layout(), inputs[0]->comp_node());
+        outputs[i]->dev_tensor().copy_from_fixlayout(inputs[i]->dev_tensor());
+        srcs[i] = outputs[i]->dev_tensor().as_megdnn();
     }
+    megdnn::CheckNonFinite::Param param({op.scale});
+    dnn_opr->param() = param;
     wk_size = dnn_opr->get_workspace_in_bytes(srcs, dest->layout());
     auto wk = Blob::make(cn, wk_size);
     megdnn::Workspace dnn_wk(wk->storage().get(), wk_size);
     dnn_opr->exec(srcs, dest->dev_tensor().as_megdnn(), dnn_wk);
-    return {dest};
+    return outputs;
 }
 
 std::tuple<SmallVector<LogicalTensorDesc>, bool> infer_output_attrs_fallible(
         const OpDef& def, const SmallVector<LogicalTensorDesc>& inputs) {
-    SmallVector<LogicalTensorDesc> dests(1);
-    dests[0].comp_node = inputs[0].comp_node;
-    dests[0].layout = TensorLayout(TensorShape({1}), dtype::Int32());
+    size_t size = inputs.size();
+    SmallVector<LogicalTensorDesc> dests(size + 1);
+    for (size_t i = 0; i < size; ++i) {
+        dests[i].comp_node = inputs[i].comp_node;
+        dests[i].layout = inputs[i].layout;
+    }
+    dests[size].comp_node = inputs[0].comp_node;
+    dests[size].layout = TensorLayout(TensorShape({1}), dtype::Int32());
     return {dests, true};
 }
 SmallVector<LogicalTensorDesc> infer_output_attrs(
         const OpDef& def, const SmallVector<TensorPtr>& inputs) {
-    SmallVector<LogicalTensorDesc> dests(1);
-    dests[0].comp_node = inputs[0]->comp_node();
-    dests[0].layout = TensorLayout(TensorShape({1}), dtype::Int32());
+    size_t size = inputs.size();
+    SmallVector<LogicalTensorDesc> dests(size + 1);
+    for (size_t i = 0; i < size; ++i) {
+        dests[i].comp_node = inputs[i]->comp_node();
+        dests[i].layout = inputs[i]->layout();
+    }
+    dests[size].comp_node = inputs[0]->comp_node();
+    dests[size].layout = TensorLayout(TensorShape({1}), dtype::Int32());
     return dests;
 }
 std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(

src/core/include/megbrain/ir/ops.td

@@ -397,7 +397,7 @@ def MagicMindRuntime: MgbHashableOp<"MagicMindRuntime"> {
 
 def CvtColor: MgbHashableOp<"CvtColor", [CvtColorParam]>;
 
-def CheckNonFinite: MgbHashableOp<"CheckNonFinite", [EmptyParam]>;
+def CheckNonFinite: MgbHashableOp<"CheckNonFinite", [CheckNonFiniteParam]>;
 
 def FastpathCopy: MgbHashableOp<"FastpathCopy">;
 

src/opr/impl/misc.cpp

@@ -487,39 +487,60 @@ CheckNonFinite::CheckNonFinite(
         const VarNodeArrayView& inp, const Param& param,
         const OperatorNodeConfig& config)
         : Super(OperatorNodeBaseCtorParam{
-                  inp[0]->owner_graph(), config, "check_non_finite", inp}) {
+                  inp[0]->owner_graph(), config, "check_non_finite", inp}),
+          m_scale(param.scale) {
     mgb_assert(!inp.empty());
+
     for (auto&& i : inp) {
         add_input({i});
+        add_output(None)
+                ->dtype(dtype::Float32())
+                .add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
     }
     add_output(None)->dtype(dtype::Int32()).add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
     cg::add_workspace_output(this);
 }
 
-SymbolVar CheckNonFinite::make(
+SymbolVarArray CheckNonFinite::make(
         const VarNodeArrayView& inp, const Param& param,
         const OperatorNodeConfig& config) {
     mgb_assert(!inp.empty());
     intl::BatchedDTypePromotion dtp{inp};
-    return SymbolVar{inp[0]}.insert_single_output_opr<CheckNonFinite>(
-            dtp.get_vars(), param, config);
+    auto outputs =
+            inp[0]->owner_graph()
+                    ->insert_opr(std::make_unique<CheckNonFinite>(inp, param, config))
+                    ->output();
+    mgb_assert(outputs.size() == inp.size() + 2);
+    SymbolVarArray ret(outputs.size() - 1);
+    for (size_t i = 0; i < ret.size(); ++i)
+        ret[i] = outputs[i];
+    return ret;
 }
 
 void CheckNonFinite::scn_do_execute() {
-    megdnn::TensorNDArray inp_arr(input().size());
-    for (size_t i = 0; i < input().size(); ++i) {
-        inp_arr[i] = input()[i]->dev_tensor().as_megdnn();
+    size_t size = input().size();
+    megdnn::TensorNDArray oup_arr(size);
+    // copy an outputs to the dnn for inplace
+    for (size_t i = 0; i < size; ++i) {
+        oup_arr[i] = output(i)
+                             ->dev_tensor()
+                             .copy_from_fixlayout(input(i)->dev_tensor())
+                             .as_megdnn();
     }
+    megdnn_opr()->param().scale = m_scale;
     megdnn_opr()->exec(
-            inp_arr, output(0)->dev_tensor().as_megdnn(),
-            intl::get_megdnn_workspace_from_var(output(1)));
+            oup_arr, output(size)->dev_tensor().as_megdnn(),
+            intl::get_megdnn_workspace_from_var(output(size + 1)));
 }
 
 void CheckNonFinite::init_output_static_infer_desc() {
     using namespace cg::static_infer;
 
     auto&& mgr = owner_graph()->static_infer_manager();
-
+    size_t size = input().size();
+    for (size_t i = 0; i < size; ++i) {
+        mgr.register_shape_infer(output(i), ShapeInferDesc::make_identity(input(i)));
+    }
     auto infer_oshp = [](TensorShape& dest, const InpVal& iv) {
         TensorLayout dst;
         dst.shape[0] = 1;
@@ -532,7 +553,7 @@ void CheckNonFinite::init_output_static_infer_desc() {
     DepVal deps;
     for (auto i : input())
         deps.push_back({i, DepType::SHAPE});
-    mgr.register_shape_infer(output(0), {SourceType::DEP, deps, infer_oshp});
+    mgr.register_shape_infer(output(size), {SourceType::DEP, deps, infer_oshp});
 
     auto infer_wk = [this](TensorShape& dest, const InpVal& inp) {
         dest.ndim = 1;
@@ -541,10 +562,11 @@ void CheckNonFinite::init_output_static_infer_desc() {
             inp_arr[i] = {NULL, {inp.val.at(i).shape(), input(0)->dtype()}};
         }
         dest.shape[0] = megdnn_opr()->get_workspace_in_bytes(
-                inp_arr, {output(0)->shape(), output(0)->dtype()});
+                inp_arr, {output(input().size() + 1)->shape(),
+                          output(input().size() + 1)->dtype()});
         return true;
     };
-    mgr.register_shape_infer(output(1), {SourceType::DEP, deps, infer_wk});
+    mgr.register_shape_infer(output(size + 1), {SourceType::DEP, deps, infer_wk});
 }
 
 void CheckNonFinite::add_input_layout_constraint() {

src/opr/impl/misc.sereg.h

@@ -56,7 +56,16 @@ struct OprMaker<opr::TopK, 2> {
 };
 
 template <>
-struct OprMaker<opr::CheckNonFinite, 0> : public OprMakerVariadic<opr::CheckNonFinite> {
+struct OprMaker<opr::CheckNonFinite, 0> {
+    using Opr = opr::CheckNonFinite;
+    using Param = Opr::Param;
+    static cg::OperatorNodeBase* make(
+            const Param& param, const cg::VarNodeArray& inputs, ComputingGraph& graph,
+            const OperatorNodeConfig& config) {
+        MGB_MARK_USED_VAR(graph);
+        auto out = Opr::make(inputs, param, config);
+        return out[0].node()->owner_opr();
+    }
 };
 
 }  // namespace serialization

src/opr/include/megbrain/opr/misc.h

@@ -183,16 +183,17 @@ public:
             const OperatorNodeConfig& config = {});
 };
 
-MGB_DEFINE_OPR_CLASS(CheckNonFinite, intl::CheckNonFiniteBase)  //{
-void scn_do_execute() override;
-void init_output_static_infer_desc() override;
-void add_input_layout_constraint() override;
+MGB_DEFINE_OPR_CLASS(CheckNonFinite, intl::CheckNonFiniteBase) // {
+    void scn_do_execute() override;
+    void init_output_static_infer_desc() override;
+    void add_input_layout_constraint() override;
+    float m_scale = 1;
 
 public:
-MGE_WIN_DECLSPEC_FUC CheckNonFinite(
+    MGE_WIN_DECLSPEC_FUC CheckNonFinite(
             const VarNodeArrayView& inp, const Param& param,
             const OperatorNodeConfig& config);
-MGE_WIN_DECLSPEC_FUC static SymbolVar make(
+    MGE_WIN_DECLSPEC_FUC static SymbolVarArray make(
             const VarNodeArrayView& inp, const Param& param = {},
             const OperatorNodeConfig& config = {});
 };

src/serialization/impl/schema.fbs

@@ -115,6 +115,7 @@ union OperatorParam {
     param.SlidingWindowTranspose = 81,
     param.Padding = 82,
     param.ShuffleRNG = 83,
+    param.CheckNonFinite = 84,
 }
 
 table Operator {