feat(dnn/check_non_finite): add batch check_non_finite

GitOrigin-RevId: e108133282cb2c9129292715ae6eab1e396cd0bc

dnn/include/megdnn/oprs/general.h

@@ -1345,22 +1345,23 @@ protected:
  */
 class CheckNonFinite : public OperatorBase {
     DEF_OPR_PARAM(Empty);
-    DEF_OPR_IMPL(CheckNonFinite, OperatorBase, 1, 1);
+    DEF_OPR_IMPL(CheckNonFinite, OperatorBase, -1, 1);
+    size_t m_size = 0;
 
 public:
     virtual size_t get_workspace_in_bytes(
-            const TensorLayout& src, const TensorLayout& dst) = 0;
+            const TensorNDArray& srcs, const TensorLayout& dst) = 0;
 
-    void deduce_layout(const TensorLayout& src, TensorLayout& dst);
+    void deduce_layout(const TensorLayoutArray& srcs, TensorLayout& dst);
 
     virtual void exec(
-            _megdnn_tensor_in src, _megdnn_tensor_out dst,
+            _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
             _megdnn_workspace workspace) = 0;
 
 protected:
     void check_exec(
-            const TensorLayout& src, const TensorLayout& dst,
-            size_t workspace_in_bytes);
+            const TensorNDArray& srcs, const TensorND& dst, size_t workspace_in_bytes);
+    virtual size_t _get_workspace_in_bytes() = 0;
 };
 
 /*!

dnn/src/common/check_non_finite.cpp

@@ -15,16 +15,15 @@
 namespace megdnn {
 
 void CheckNonFinite::check_exec(
-        const TensorLayout& src, const TensorLayout& dst, size_t workspace_in_bytes) {
-    megdnn_assert_contiguous(src);
-    megdnn_assert_contiguous(dst);
-    megdnn_assert(src.ndim == 1);
-    megdnn_assert(src.dtype == dtype::Float32());
-    auto required_workspace_in_bytes = get_workspace_in_bytes(src, dst);
+        const TensorNDArray& srcs, const TensorND& dst, size_t workspace_in_bytes) {
+    megdnn_assert_contiguous(dst.layout);
+    megdnn_assert(srcs.size() > 0);
+    megdnn_assert(srcs.begin()->layout.dtype == dtype::Float32());
+    auto required_workspace_in_bytes = _get_workspace_in_bytes();
     megdnn_assert(workspace_in_bytes >= required_workspace_in_bytes);
 }
 
-void CheckNonFinite::deduce_layout(const TensorLayout&, TensorLayout& dst) {
+void CheckNonFinite::deduce_layout(const TensorLayoutArray&, TensorLayout& dst) {
     dst.shape[0] = 1;
     dst.ndim = 1;
     dst.dtype = dtype::Int32();

dnn/src/common/reduce_helper.h

@@ -156,21 +156,35 @@ struct MaxOp<src_ctype, dst_ctype, dt_float32> {
             : INIT(wtype(DTypeTrait<wtype>::min())), src(src), dst(dst), B(B) {}
 };
 
-template <typename src_ctype, typename dst_ctype, typename wtype_>
+template <typename src_ctype, typename index_ctype, typename dst_ctype, typename wtype_>
 struct CheckNonFiniteOp {
     typedef wtype_ wtype;
     const wtype INIT;
 
-    RefPtr src;
+    RefPtr* srcs;
+    RefPtr srcs_total_nr_elems;
     RefPtr dst;
     const size_t B;
 
-    wtype read(uint32_t idx) { return !std::isfinite(src.ptr<src_ctype>()[idx]); }
+    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; }
-    MEGDNN_HOST MEGDNN_DEVICE
-    CheckNonFiniteOp(const RefPtr& src, const RefPtr& dst, size_t B)
-            : INIT(wtype(0)), src(src), dst(dst), B(B) {}
+    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);

dnn/src/common/reduce_helper_device.h

@@ -185,28 +185,41 @@ struct MaxOp<src_ctype, dst_ctype, dt_float32> {
             : INIT(wtype(DTypeTrait<wtype>::min())), src(src), dst(dst), B(B) {}
 };
 
-template <typename src_ctype, typename dst_ctype, typename wtype_>
+template <typename src_ctype, typename index_ctype, typename dst_ctype, typename wtype_>
 struct CheckNonFiniteOp {
     typedef wtype_ wtype;
     const wtype INIT;
 
-    src_ctype* src;
+    src_ctype** srcs;
+    index_ctype* srcs_total_nr_elems;
     dst_ctype* dst;
     const size_t B;
 
     MEGDNN_HOST MEGDNN_DEVICE wtype read(uint32_t idx) {
+        size_t x = idx / B;
+        size_t y = idx % B;
+        if (y < srcs_total_nr_elems[x]) {
 #if defined(__CUDA_ARCH__)
-        return !isfinite(src[idx]);
+            wtype val = isfinite(srcs[x][y]);
 #else
-        return !std::isfinite(src[idx]);
+            wtype val = std::isfinite(srcs[x][y]);
 #endif
+            return !val;
+        }
+        return 0;
     }
     MEGDNN_HOST MEGDNN_DEVICE void write(uint32_t idx, wtype val) { dst[idx] = val; }
     static MEGDNN_HOST MEGDNN_DEVICE wtype apply(wtype lhs, wtype rhs) {
         return lhs | rhs;
     }
-    MEGDNN_HOST MEGDNN_DEVICE CheckNonFiniteOp(src_ctype* src, dst_ctype* dst, size_t B)
-            : INIT(wtype(0)), src(src), dst(dst), B(B) {}
+    MEGDNN_HOST MEGDNN_DEVICE CheckNonFiniteOp(
+            src_ctype** srcs, index_ctype* srcs_total_nr_elems, dst_ctype* dst,
+            size_t B)
+            : INIT(wtype(0)),
+              srcs(srcs),
+              srcs_total_nr_elems(srcs_total_nr_elems),
+              dst(dst),
+              B(B) {}
 };
 
 }  // namespace device_reduce

dnn/src/cuda/check_non_finite/kern.cu

@@ -19,7 +19,8 @@ namespace cuda {
 #define COMMA ,
 
 INST_REDUCE(
-        device_reduce::CheckNonFiniteOp<dt_float32 COMMA dt_int32 COMMA dt_int32>,
+        device_reduce::CheckNonFiniteOp<
+                dt_float32 COMMA size_t COMMA dt_int32 COMMA dt_int32>,
         false);
 
 #undef COMMA

dnn/src/cuda/check_non_finite/opr_impl.cpp

@@ -21,22 +21,83 @@ namespace megdnn {
 namespace cuda {
 
 using device_reduce::CheckNonFiniteOp;
+#define total_nr_elems_max 2048
+size_t CheckNonFiniteImpl::_get_workspace_in_bytes() {
+    // Call the _get_workspace_in_bytes to reduce the loop fetch workspace bytes
+    typedef CheckNonFiniteOp<dt_float32, size_t, dt_int32, dt_int32> Op;
+    megdnn_assert(m_size > 0);
+    WorkspaceBundle bundle(
+            nullptr, {
+                             sizeof(dt_float32*) * m_size,
+                             sizeof(size_t) * m_size,
+                     });
+    return get_reduce_workspace_in_bytes<Op>(1, m_size * total_nr_elems_max, 1) +
+           bundle.total_size_in_bytes();
+}
 
 size_t CheckNonFiniteImpl::get_workspace_in_bytes(
-        const TensorLayout& src, const TensorLayout& dst) {
-    typedef CheckNonFiniteOp<dt_float32, dt_int32, dt_int32> Op;
-    return get_reduce_workspace_in_bytes<Op>(1, src.total_nr_elems(), 1);
+        const TensorNDArray& srcs, const TensorLayout&) {
+    m_size = 0;
+    for (const auto& src : srcs) {
+        m_size += DIVUP(src.layout.total_nr_elems(), total_nr_elems_max);
+    }
+    return _get_workspace_in_bytes();
 }
 
 void CheckNonFiniteImpl::exec(
-        _megdnn_tensor_in src, _megdnn_tensor_out dst, _megdnn_workspace workspace) {
-    check_exec(src.layout, dst.layout, workspace.size);
-    typedef CheckNonFiniteOp<dt_float32, dt_int32, dt_int32> Op;
+        _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
+        _megdnn_workspace workspace) {
+    check_exec(srcs, dst, workspace.size);
+    typedef CheckNonFiniteOp<dt_float32, size_t, dt_int32, dt_int32> Op;
     auto stream = cuda_stream(this->handle());
-    auto B = src.layout.total_nr_elems();
+    SmallVector<size_t> workspace_sizes{
+            sizeof(dt_float32*) * m_size,
+            sizeof(size_t) * m_size,
+    };
+    WorkspaceBundle workspace_cpu(nullptr, workspace_sizes),
+            workspace_gpu(nullptr, workspace_sizes);
+    auto total_workspace_size = workspace_cpu.total_size_in_bytes();
+    void* workspace_cpu_raw = malloc(total_workspace_size);
+    megdnn_assert_internal(workspace_cpu_raw);
+    void* workspace_gpu_raw = workspace.raw_ptr;
+    workspace_cpu = WorkspaceBundle(workspace_cpu_raw, workspace_sizes);
+    workspace_gpu = WorkspaceBundle(workspace_gpu_raw, workspace_sizes);
+
+    auto srcs_cpu = static_cast<dt_float32**>(workspace_cpu.get(0));
+    auto srcs_gpu = static_cast<dt_float32**>(workspace_gpu.get(0));
+    auto srcs_total_nr_elems_cpu = static_cast<size_t*>(workspace_cpu.get(1));
+    auto srcs_total_nr_elems_gpu = static_cast<size_t*>(workspace_gpu.get(1));
+
+    // srcs
+    // cut the tensor to a fixed length of total_nr_elems_max
+    size_t i = 0;
+    for (const auto& src : srcs) {
+        size_t src_nr_elems = src.layout.total_nr_elems();
+        size_t nr_elems = DIVUP(src_nr_elems, total_nr_elems_max);
+        for (size_t j = 0; j < nr_elems; ++j, ++i) {
+            srcs_cpu[i] = src.ptr<dt_float32>() + j * total_nr_elems_max;
+            if (j + 1 == nr_elems && src_nr_elems % total_nr_elems_max) {
+                srcs_total_nr_elems_cpu[i] = src_nr_elems % total_nr_elems_max;
+            } else {
+                srcs_total_nr_elems_cpu[i] = total_nr_elems_max;
+            }
+        }
+    }
+    for (size_t i = 0; i < workspace_cpu.nr_workspace(); ++i) {
+        cuda_check(cudaMemcpyAsync(
+                workspace_gpu.get(i), workspace_cpu.get(i), workspace_cpu.get_size(i),
+                cudaMemcpyHostToDevice, stream));
+    }
+    cuda_check(cudaStreamAddCallback(
+            stream, callback_free, static_cast<void*>(workspace_cpu_raw), 0));
+
     return run_reduce<Op, false>(
-            workspace.ptr<dt_int32>(), 1, B, 1, stream,
-            Op(src.ptr<dt_float32>(), dst.ptr<dt_int32>(), B));
+            static_cast<dt_int32*>(
+                    (void*)((char*)workspace_gpu_raw +
+                            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));
 }
 
 }  // namespace cuda

dnn/src/cuda/check_non_finite/opr_impl.h

@@ -18,16 +18,18 @@ namespace megdnn {
 namespace cuda {
 
 class CheckNonFiniteImpl final : public CheckNonFinite {
+    size_t _get_workspace_in_bytes() override;
+
 public:
     using CheckNonFinite::CheckNonFinite;
 
     size_t get_workspace_in_bytes(
-            const TensorLayout& src, const TensorLayout& dst) override;
+            const TensorNDArray& srcs, const TensorLayout& dst) override;
 
     bool is_thread_safe() const override { return true; }
 
     void exec(
-            _megdnn_tensor_in src, _megdnn_tensor_out dst,
+            _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
             _megdnn_workspace workspace) override;
 };
 

dnn/src/naive/check_non_finite/opr_impl.cpp

@@ -17,21 +17,25 @@
 namespace {
 using namespace megdnn;
 
-#define src_ctype dt_float32
-#define wtype     dt_int32
-
-void reduce_fwd(const src_ctype* sptr, wtype* dptr, size_t size) {
-    std::function<wtype(size_t, size_t)> func;
-    func = [&](size_t l, size_t r) -> wtype {
-        if (l + 1 < r) {
-            size_t mid = l + (r - l) / 2;
-            return func(l, mid) | func(mid, r);
-        } else {
-            return static_cast<wtype>(!std::isfinite(sptr[l]));
-        }
-    };
-
-    dptr[0] = func(0, size);
+#define wtype dt_int32
+
+void reduce_fwd(const TensorNDArray& srcs, wtype* dptr) {
+    dptr[0] = 0;
+    for (auto src : srcs) {
+        auto sptr = src.ptr<dt_float32>();
+        size_t size = src.layout.total_nr_elems();
+        std::function<wtype(wtype, wtype)> func;
+        func = [&](wtype l, wtype r) -> wtype {
+            if (l + 1 < r) {
+                wtype mid = l + (r - l) / 2;
+                return func(l, mid) | func(mid, r);
+            } else {
+                auto val = std::isfinite(sptr[l]);
+                return static_cast<wtype>(!val);
+            }
+        };
+        dptr[0] |= func(0, size);
+    }
 }
 
 }  // namespace
@@ -39,20 +43,13 @@ void reduce_fwd(const src_ctype* sptr, wtype* dptr, size_t size) {
 namespace megdnn {
 namespace naive {
 
-size_t CheckNonFiniteImpl::get_workspace_in_bytes(
-        const TensorLayout&, const TensorLayout&) {
-    return 0;
-}
-
 void CheckNonFiniteImpl::exec(
-        _megdnn_tensor_in src, _megdnn_tensor_out dst, _megdnn_workspace workspace) {
-    check_exec(src.layout, dst.layout, workspace.size);
+        _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
+        _megdnn_workspace workspace) {
+    check_exec(srcs, dst, workspace.size);
 
     auto handle = static_cast<HandleImpl*>(this->handle());
-    MEGDNN_DISPATCH_CPU_KERN(
-            handle, reduce_fwd(
-                            src.ptr<dt_float32>(), dst.ptr<dt_int32>(),
-                            src.layout.total_nr_elems()));
+    MEGDNN_DISPATCH_CPU_KERN(handle, reduce_fwd(srcs, dst.ptr<dt_int32>()));
 }
 }  // namespace naive
 }  // namespace megdnn

dnn/src/naive/check_non_finite/opr_impl.h

@@ -17,16 +17,20 @@ namespace megdnn {
 namespace naive {
 
 class CheckNonFiniteImpl final : public CheckNonFinite {
+    size_t _get_workspace_in_bytes() override { return 0; }
+
 public:
     using CheckNonFinite::CheckNonFinite;
 
     bool is_thread_safe() const override { return true; }
 
-    size_t get_workspace_in_bytes(
-            const TensorLayout& src, const TensorLayout& dst) override;
+    size_t get_workspace_in_bytes(const TensorNDArray&, const TensorLayout&) override {
+        m_size = 0;
+        return _get_workspace_in_bytes();
+    }
 
     void exec(
-            _megdnn_tensor_in src, _megdnn_tensor_out dst,
+            _megdnn_in const TensorNDArray& srcs, _megdnn_tensor_out dst,
             _megdnn_workspace workspace) override;
 };
 

dnn/test/common/opr_proxy.h

@@ -202,6 +202,27 @@ struct OprProxy<ConcatForward> {
     }
 };
 
+template <>
+struct OprProxy<CheckNonFinite> {
+    static void deduce_layout(CheckNonFinite* opr, TensorLayoutArray& layouts) {
+        megdnn_assert(layouts.size() >= 2);
+        auto inp = layouts;
+        inp.pop_back();
+        opr->deduce_layout(inp, layouts.back());
+    }
+
+    static void exec(CheckNonFinite* opr, const TensorNDArray& tensors) {
+        megdnn_assert(tensors.size() >= 2);
+        auto inps = tensors;
+        inps.pop_back();
+
+        WorkspaceWrapper W(
+                opr->handle(),
+                opr->get_workspace_in_bytes(inps, tensors.back().layout));
+        opr->exec(inps, tensors.back(), W.workspace());
+    }
+};
+
 template <>
 struct OprProxy<SplitForward> : DeduceLayoutProxy<SplitForward, 0, false> {
     WorkspaceWrapper W;

dnn/test/cuda/check_non_finite.cpp

@@ -22,13 +22,16 @@ TEST_F(CUDA, CHECK_NON_FINITE_BASIC) {
     const auto nan = std::numeric_limits<float>::quiet_NaN();
     UniformFloatWithValueRNG rng(-1.0f, 1.0f, 0.1f, inf);
     checker.set_rng(0, &rng);
-    checker.execs({{512 * 16}, {1}});
+    checker.execs({{512 * 4}, {4}, {1}});
     rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 1.f, inf);
     checker.set_rng(0, &rng);
-    checker.execs({{512 * 16}, {1}});
+    checker.execs({{4}, {512 * 4}, {1}});
     rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 1.f, nan);
     checker.set_rng(0, &rng);
-    checker.execs({{512 * 16}, {1}});
+    checker.execs({{32}, {256}, {1}});
+    rng = UniformFloatWithValueRNG(-1.0f, 1.0f, 0.f, nan);
+    checker.set_rng(0, &rng);
+    checker.execs({{16}, {16}, {2}, {1}});
 }
 
 }  // namespace test

dnn/test/naive/check_non_finite.cpp

@@ -20,23 +20,28 @@ namespace test {
 TEST_F(NAIVE, CHECK_NON_FINITE_BASIC) {
     Checker<CheckNonFinite> checker(handle(), false);
     checker.exect(
-            Testcase{TensorValue({4}, dtype::Float32(), {1.1, 2.2, 3.3, 4.3}), {}},
-            Testcase{{}, TensorValue({1}, dtype::Int32(), {0})});
+            Testcase{
+                    TensorValue({4}, dtype::Float32(), {1.1, 2.2, 3.3, 4.3}),
+                    TensorValue({4}, dtype::Float32(), {1.1, 2.2, 3.3, 4.3}),
+                    {}},
+            Testcase{{}, {}, TensorValue({1}, dtype::Int32(), {0})});
     checker.exect(
             Testcase{
+                    TensorValue({4}, dtype::Float32(), {1.1, 2.2, 3.3, 4.3}),
                     TensorValue(
                             {4}, dtype::Float32(),
                             {1.1f, 2.2f, 3.3f, std::numeric_limits<float>::infinity()}),
                     {}},
-            Testcase{{}, TensorValue({1}, dtype::Int32(), {1})});
+            Testcase{{}, {}, TensorValue({1}, dtype::Int32(), {1})});
     checker.exect(
             Testcase{
+                    TensorValue({4}, dtype::Float32(), {1.1, 2.2, 3.3, 4.3}),
                     TensorValue(
                             {4}, dtype::Float32(),
                             {1.1f, 2.2f, 3.3f,
                              std::numeric_limits<float>::quiet_NaN()}),
                     {}},
-            Testcase{{}, TensorValue({1}, dtype::Int32(), {1})});
+            Testcase{{}, {}, TensorValue({1}, dtype::Int32(), {1})});
 }
 
 }  // namespace test

imperative/python/megengine/amp/grad_scaler.py

@@ -128,21 +128,22 @@ class GradScaler:
             grad_tensors: Tensors needed to unscale grads. Should be all tensors
                 that are affected by ``target`` tensor in GradManager's backward.
         """
-        # use float64 for better precision
-        inv_scale = Tensor(1.0 / self.scale_factor)
-        for tensor in grad_tensors:
-            if tensor is None or getattr(tensor, "grad", None) is None:
-                continue
-            # to support tracing, _check_gradients should be applied to every grad.
-            if self._check_gradients(tensor.grad):
-                self._found_non_finite = True
-            tensor.grad *= inv_scale
+        # 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:
+            # use float64 for better precision
+            inv_scale = Tensor(1.0 / self.scale_factor)
+            for tensor in grad_tensors:
+                if tensor is None or getattr(tensor, "grad", None) is None:
+                    continue
+                tensor.grad *= inv_scale
         return self
 
     def _check_gradients(self, grad):

imperative/python/megengine/functional/math.py

@@ -9,7 +9,7 @@
 import collections
 import math
 from functools import lru_cache
-from typing import Optional, Sequence, Tuple, Union
+from typing import Iterable, Optional, Sequence, Tuple, Union
 
 from ..core import _config
 from ..core._imperative_rt.core2 import apply, dtype_promotion
@@ -1183,7 +1183,7 @@ def svd(inp: Tensor, full_matrices=False, compute_uv=True) -> Tensor:
     return U, sigma, V
 
 
-def _check_non_finite(inp: Tensor) -> Tensor:
+def _check_non_finite(inps: Iterable[Tensor]) -> Tensor:
     r"""Check whether input contains infinite or nan value.
 
     Args:
@@ -1193,6 +1193,6 @@ def _check_non_finite(inp: Tensor) -> Tensor:
         a int32 scalar tensor, 0 for False and 1 for True.
     """
     op = builtin.CheckNonFinite()
-    (oup,) = apply(op, inp.reshape(-1).astype("float32"))
+    (oup,) = apply(op, *inps)
     oup._setscalar()
     return oup

imperative/python/test/unit/amp/test_grad_scaler.py

@@ -10,21 +10,26 @@ import numpy as np
 import megengine as mge
 from megengine.amp import GradScaler
 from megengine.autodiff import GradManager
+from megengine.jit import trace
 
 
 def test_grad_scaler():
-    gm = GradManager()
-    scaler = GradScaler()
+    def f():
+        gm = GradManager()
+        scaler = GradScaler()
 
-    x = mge.tensor(1.0)
-    for _ in range(3):
-        with gm:
-            y = x + 1
-            gm.attach(y)
-            loss = y + 1
-            scaler.backward(gm, loss, unscale_grad=False)
-        np.testing.assert_equal(y.grad.numpy(), scaler.scale_factor)
+        x = mge.tensor(1.0)
+        for _ in range(3):
+            with gm:
+                y = x + 1
+                gm.attach(y)
+                loss = y + 1
+                scaler.backward(gm, loss, unscale_grad=False)
+            np.testing.assert_equal(y.grad.numpy(), scaler.scale_factor)
+            scaler.unscale(gm.attached_tensors())
+            np.testing.assert_equal(y.grad.numpy(), 1)
+        # test handle None elements
         scaler.unscale(gm.attached_tensors())
-        np.testing.assert_equal(y.grad.numpy(), 1)
-    # test handle None elements
-    scaler.unscale(gm.attached_tensors())
+
+    f()
+    trace(f)()

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

@@ -191,16 +191,17 @@ def test_sum_neg_axis():
 
 def test_non_finite():
     shape = (32, 3, 32, 32)
-    data = np.random.random(shape).astype(np.float32)
-    rst = F.math._check_non_finite(tensor(data))
+    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)])
     np.testing.assert_equal(rst.numpy(), [0])
 
-    data[0][0][0][0] = float("inf")
-    rst = F.math._check_non_finite(tensor(data))
+    data2[0][0][0][0] = float("inf")
+    rst = F.math._check_non_finite([tensor(data1), tensor(data2)])
     np.testing.assert_equal(rst.numpy(), [1])
 
-    data[0][0][0][0] = float("nan")
-    rst = F.math._check_non_finite(tensor(data))
+    data2[0][0][0][0] = float("nan")
+    rst = F.math._check_non_finite([tensor(data1), tensor(data2)])
     np.testing.assert_equal(rst.numpy(), [1])
 
 

imperative/src/impl/ops/misc.cpp

@@ -17,14 +17,56 @@ namespace mgb {
 namespace imperative {
 
 namespace check_non_finite {
-auto apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
+SymbolVar apply_on_var_node(const OpDef& def, const VarNodeArray& inputs) {
     auto&& op = def.cast_final_safe<CheckNonFinite>();
-    mgb_assert(inputs.size() == 1);
     OperatorNodeConfig config{op.make_name()};
-    return opr::CheckNonFinite::make(inputs[0], {}, config);
+    return opr::CheckNonFinite::make(inputs, {}, config);
+}
+
+SmallVector<TensorPtr> apply_on_physical_tensor(
+        const OpDef& def, const SmallVector<TensorPtr>& inputs) {
+    size_t size = inputs.size();
+
+    auto dest = Tensor::make(
+            TensorLayout(TensorShape({1}), dtype::Int32()), inputs[0]->comp_node());
+    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);
+    for (size_t i = 0; i < size; ++i) {
+        srcs[i] = inputs[i]->dev_tensor().as_megdnn();
+    }
+    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};
+}
+
+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());
+    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());
+    return dests;
+}
+std::tuple<SmallVector<MemoryDesc>, SmallVector<MemoryDesc>> infer_output_mem_desc(
+        const OpDef& def, const SmallVector<TensorPtr>& inputs_tensors,
+        const SmallVector<MemoryDesc>& inputs_mems) {
+    return {{}, {}};
 }
 OP_TRAIT_REG(CheckNonFinite, CheckNonFinite)
         .apply_on_var_node(apply_on_var_node)
+        .apply_on_physical_tensor(apply_on_physical_tensor)
+        .infer_output_attrs_fallible(infer_output_attrs_fallible)
+        .infer_output_mem_desc(infer_output_mem_desc)
         .fallback();
 }  // namespace check_non_finite
 

src/opr/impl/misc.cpp

@@ -482,18 +482,74 @@ MGB_IMPL_OPR_GRAD(TopK) {
 #endif
 
 /* ================= CheckNonFinite =================  */
-namespace mgb {
-namespace opr {
-namespace intl {
-template <>
-struct MegDNNOprInitPostCtor<CheckNonFinite> {
-    static void apply(cg::OperatorNodeBase& opr) {
-        opr.output(0)->dtype(dtype::Int32());
-    }
-};
-}  // namespace intl
-}  // namespace opr
-}  // namespace mgb
 MGB_DYN_TYPE_OBJ_FINAL_IMPL(CheckNonFinite);
-MEGDNN_OPR_INIT1(CheckNonFinite, "check_non_finite")
+CheckNonFinite::CheckNonFinite(
+        const VarNodeArrayView& inp, const Param& param,
+        const OperatorNodeConfig& config)
+        : Super(OperatorNodeBaseCtorParam{
+                  inp[0]->owner_graph(), config, "check_non_finite", inp}) {
+    mgb_assert(!inp.empty());
+    for (auto&& i : inp) {
+        add_input({i});
+    }
+    add_output(None)->dtype(dtype::Int32()).add_flag(VarNode::Flag::ALLOW_EMPTY_SHAPE);
+    cg::add_workspace_output(this);
+}
+
+SymbolVar 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);
+}
+
+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();
+    }
+    megdnn_opr()->exec(
+            inp_arr, output(0)->dev_tensor().as_megdnn(),
+            intl::get_megdnn_workspace_from_var(output(1)));
+}
+
+void CheckNonFinite::init_output_static_infer_desc() {
+    using namespace cg::static_infer;
+
+    auto&& mgr = owner_graph()->static_infer_manager();
+
+    auto infer_oshp = [](TensorShape& dest, const InpVal& iv) {
+        TensorLayout dst;
+        dst.shape[0] = 1;
+        dst.ndim = 1;
+        dst.dtype = dtype::Int32();
+        dst.init_contiguous_stride();
+        dest = dst;
+        return true;
+    };
+    DepVal deps;
+    for (auto i : input())
+        deps.push_back({i, DepType::SHAPE});
+    mgr.register_shape_infer(output(0), {SourceType::DEP, deps, infer_oshp});
+
+    auto infer_wk = [this](TensorShape& dest, const InpVal& inp) {
+        dest.ndim = 1;
+        megdnn::TensorNDArray inp_arr(input().size());
+        for (size_t i = 0; i < input().size(); ++i) {
+            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()});
+        return true;
+    };
+    mgr.register_shape_infer(output(1), {SourceType::DEP, deps, infer_wk});
+}
+
+void CheckNonFinite::add_input_layout_constraint() {
+    for (auto i : input()) {
+        i->add_layout_constraint_contiguous();
+    }
+}
 // vim: syntax=cpp.doxygen foldmethod=marker foldmarker=f{{{,f}}}

src/opr/impl/misc.sereg.h

@@ -55,6 +55,10 @@ struct OprMaker<opr::TopK, 2> {
     }
 };
 
+template <>
+struct OprMaker<opr::CheckNonFinite, 0> : public OprMakerVariadic<opr::CheckNonFinite> {
+};
+
 }  // namespace serialization
 
 namespace opr {
@@ -72,7 +76,7 @@ MGB_SEREG_OPR(CumsumV1, 1);
 #if MGB_CUDA
 MGB_SEREG_OPR(NvOf, 1);
 #endif
-MGB_SEREG_OPR(CheckNonFinite, 1);
+MGB_SEREG_OPR(CheckNonFinite, 0);
 
 }  // namespace opr
 }  // namespace mgb

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

@@ -142,6 +142,8 @@ using CondTakeBase = cg::SingleCNOperatorNode<
         cg::OperatorNodeBase, mixin::MegDNNOprHolderImpl<megdnn::CondTake>>;
 using TopKBase = cg::SingleCNOperatorNode<
         cg::OperatorNodeBase, mixin::MegDNNOprHolderImpl<megdnn::TopK>>;
+using CheckNonFiniteBase = cg::SingleCNOperatorNode<
+        cg::OperatorNodeBase, mixin::MegDNNOprHolderImpl<megdnn::CheckNonFinite>>;
 }  // namespace intl
 
 /*!
@@ -181,7 +183,19 @@ public:
             const OperatorNodeConfig& config = {});
 };
 
-MGB_DEFINE_MEGDNN_OPR_WRAPPER_FWD1(CheckNonFinite);
+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;
+
+public:
+MGE_WIN_DECLSPEC_FUC CheckNonFinite(
+        const VarNodeArrayView& inp, const Param& param,
+        const OperatorNodeConfig& config);
+MGE_WIN_DECLSPEC_FUC static SymbolVar make(
+        const VarNodeArrayView& inp, const Param& param = {},
+        const OperatorNodeConfig& config = {});
+};
 
 }  // namespace opr
 }  // namespace mgb