Make sure input numpy array is aligned. (#8690)

- use `np.require` to specify that the alignment is required.
- scipy csr as well.
- validate input pointer in `ArrayInterface`.

python-package/xgboost/core.py

@@ -2174,6 +2174,7 @@ class Booster:
         )
         return _prediction_output(shape, dims, preds, False)
 
+    # pylint: disable=too-many-statements
     def inplace_predict(
         self,
         data: DataType,
@@ -2194,10 +2195,10 @@ class Booster:
 
         .. code-block:: python
 
-            booster.set_param({'predictor': 'gpu_predictor'})
+            booster.set_param({"predictor": "gpu_predictor"})
             booster.inplace_predict(cupy_array)
 
-            booster.set_param({'predictor': 'cpu_predictor})
+            booster.set_param({"predictor": "cpu_predictor"})
             booster.inplace_predict(numpy_array)
 
         .. versionadded:: 1.1.0
@@ -2310,14 +2311,16 @@ class Booster:
             )
             return _prediction_output(shape, dims, preds, False)
         if isinstance(data, scipy.sparse.csr_matrix):
-            csr = data
+            from .data import _transform_scipy_csr
+
+            data = _transform_scipy_csr(data)
             _check_call(
                 _LIB.XGBoosterPredictFromCSR(
                     self.handle,
-                    _array_interface(csr.indptr),
+                    _array_interface(data.indptr),
-                    _array_interface(csr.indices),
+                    _array_interface(data.indices),
-                    _array_interface(csr.data),
+                    _array_interface(data.data),
-                    c_bst_ulong(csr.shape[1]),
+                    c_bst_ulong(data.shape[1]),
                     args,
                     p_handle,
                     ctypes.byref(shape),

python-package/xgboost/data.py

@@ -84,6 +84,21 @@ def _array_interface(data: np.ndarray) -> bytes:
     return interface_str
 
 
+def _transform_scipy_csr(data: DataType) -> DataType:
+    from scipy.sparse import csr_matrix
+
+    indptr, _ = _ensure_np_dtype(data.indptr, data.indptr.dtype)
+    indices, _ = _ensure_np_dtype(data.indices, data.indices.dtype)
+    values, _ = _ensure_np_dtype(data.data, data.data.dtype)
+    if (
+        indptr is not data.indptr
+        or indices is not data.indices
+        or values is not data.data
+    ):
+        data = csr_matrix((values, indices, indptr), shape=data.shape)
+    return data
+
+
 def _from_scipy_csr(
     data: DataType,
     missing: FloatCompatible,
@@ -97,18 +112,14 @@ def _from_scipy_csr(
             f"length mismatch: {len(data.indices)} vs {len(data.data)}"
         )
     handle = ctypes.c_void_p()
-    args = {
+    data = _transform_scipy_csr(data)
-        "missing": float(missing),
-        "nthread": int(nthread),
-    }
-    config = bytes(json.dumps(args), "utf-8")
     _check_call(
         _LIB.XGDMatrixCreateFromCSR(
             _array_interface(data.indptr),
             _array_interface(data.indices),
             _array_interface(data.data),
             c_bst_ulong(data.shape[1]),
-            config,
+            make_jcargs(missing=float(missing), nthread=int(nthread)),
             ctypes.byref(handle),
         )
     )
@@ -157,12 +168,13 @@ def _is_numpy_array(data: DataType) -> bool:
 
 
 def _ensure_np_dtype(
-    data: DataType,
+    data: DataType, dtype: Optional[NumpyDType]
-    dtype: Optional[NumpyDType]
 ) -> Tuple[np.ndarray, Optional[NumpyDType]]:
     if data.dtype.hasobject or data.dtype in [np.float16, np.bool_]:
-        data = data.astype(np.float32, copy=False)
         dtype = np.float32
+        data = data.astype(dtype, copy=False)
+    if not data.flags.aligned:
+        data = np.require(data, requirements="A")
     return data, dtype
 
 
@@ -1213,6 +1225,7 @@ def _proxy_transform(
         data, _ = _ensure_np_dtype(data, data.dtype)
         return data, None, feature_names, feature_types
     if _is_scipy_csr(data):
+        data = _transform_scipy_csr(data)
         return data, None, feature_names, feature_types
     if _is_pandas_series(data):
         import pandas as pd

src/data/array_interface.h

@@ -1,5 +1,5 @@
-/*!
+/**
- * Copyright 2019-2021 by Contributors
+ * Copyright 2019-2023 by XGBoost Contributors
  * \file array_interface.h
  * \brief View of __array_interface__
  */
@@ -7,9 +7,10 @@
 #define XGBOOST_DATA_ARRAY_INTERFACE_H_
 
 #include <algorithm>
-#include <cinttypes>
+#include <cstdint>
 #include <map>
 #include <string>
+#include <type_traits>  // std::alignment_of
 #include <utility>
 #include <vector>
 
@@ -400,6 +401,13 @@ class ArrayInterface {
 
     data = ArrayInterfaceHandler::ExtractData(array, n);
     static_assert(allow_mask ? D == 1 : D >= 1, "Masked ndarray is not supported.");
+
+    this->DispatchCall([&](auto const *data_typed_ptr) {
+      auto ptr = reinterpret_cast<uintptr_t>(data);
+      auto alignment = std::alignment_of<std::remove_pointer_t<decltype(data_typed_ptr)>>::value;
+      CHECK_EQ(ptr % alignment, 0) << "Input pointer misalignment.";
+    });
+
     if (allow_mask) {
       common::Span<RBitField8::value_type> s_mask;
       size_t n_bits = ArrayInterfaceHandler::ExtractMask(array, &s_mask);

tests/cpp/data/test_array_interface.cc

@@ -1,10 +1,12 @@
-/*!
+/**
- * Copyright 2020-2021 by XGBoost Contributors
+ * Copyright 2020-2023 by XGBoost Contributors
  */
 #include <gtest/gtest.h>
 #include <xgboost/host_device_vector.h>
 #include "../helpers.h"
 #include "../../../src/data/array_interface.h"
+#include "dmlc/logging.h"
+#include "xgboost/json.h"
 
 namespace xgboost {
 TEST(ArrayInterface, Initialize) {
@@ -71,6 +73,14 @@ TEST(ArrayInterface, Error) {
   column["mask"]["data"] = Null{};
   common::Span<RBitField8::value_type> s_mask;
   EXPECT_THROW(ArrayInterfaceHandler::ExtractMask(column_obj, &s_mask), dmlc::Error);
+
+  get<Object>(column).erase("mask");
+  // misaligned.
+  j_data = {Json(Integer(reinterpret_cast<Integer::Int>(
+                reinterpret_cast<char const*>(storage.ConstHostPointer()) + 1))),
+            Json(Boolean(false))};
+  column["data"] = j_data;
+  EXPECT_THROW({ ArrayInterface<1> arr{column}; }, dmlc::Error);
 }
 
 TEST(ArrayInterface, GetElement) {

tests/python/test_dmatrix.py

@@ -327,7 +327,7 @@ class TestDMatrix:
         nrow = 100
         ncol = 1000
         x = rand(nrow, ncol, density=0.0005, format='csr', random_state=rng)
-        assert x.indices.max() < ncol - 1
+        assert x.indices.max() < ncol
         x.data[:] = 1
         dtrain = xgb.DMatrix(x, label=rng.binomial(1, 0.3, nrow))
         assert (dtrain.num_row(), dtrain.num_col()) == (nrow, ncol)