Add a check that new strides don't go outside allocated memory and add

some tests.

Author: abergeron

pygpu/gpuarray.pxd

@@ -73,6 +73,7 @@ cdef extern from "gpuarray/buffer.h":
     gpucontext *gpucontext_init(const char *name, int devno, int flags, int *ret)
     void gpucontext_deref(gpucontext *ctx)
     char *gpucontext_error(gpucontext *ctx, int err)
+    int gpudata_property(gpudata *ctx, int prop_id, void *res)
     int gpucontext_property(gpucontext *ctx, int prop_id, void *res)
     int gpukernel_property(gpukernel *k, int prop_id, void *res)
     gpucontext *gpudata_context(gpudata *)
@@ -99,6 +100,9 @@ cdef extern from "gpuarray/buffer.h":
     int GA_CTX_PROP_MAXGSIZE0
     int GA_CTX_PROP_MAXGSIZE1
     int GA_CTX_PROP_MAXGSIZE2
+
+    int GA_BUFFER_PROP_SIZE
+
     int GA_KERNEL_PROP_MAXLSIZE
     int GA_KERNEL_PROP_PREFLSIZE
     int GA_KERNEL_PROP_NUMARGS

pygpu/gpuarray.pyx

@@ -212,6 +212,32 @@ cdef ga_order to_ga_order(ord) except <ga_order>-2:
     else:
         raise ValueError, "Valid orders are: 'A' (any), 'C' (C), 'F' (Fortran)"
 
+cdef int strides_ok(GpuArray a, strides):
+    cdef ssize_t max_axis_offset
+    cdef size_t lower = a.ga.offset
+    cdef size_t upper = a.ga.offset
+    cdef size_t itemsize = gpuarray_get_elsize(a.ga.typecode)
+    cdef size_t size
+    cdef unsigned int i
+
+    gpudata_property(a.ga.data, GA_BUFFER_PROP_SIZE, &size)
+
+    for i in range(a.ga.nd):
+        if a.ga.dimensions[i] == 0:
+            return 1
+
+        max_axis_offset = strides[i] * (a.ga.dimensions[i] - 1)
+        if max_axis_offset > 0:
+            if upper + max_axis_offset > size:
+                return 0
+            upper += max_axis_offset
+        else:
+            if lower < -max_axis_offset:
+                return 0
+            lower += max_axis_offset
+    return (upper + itemsize) <= size
+
+
 class GpuArrayException(Exception):
     """
     Exception used for most errors related to libgpuarray.
@@ -1944,6 +1970,8 @@ cdef class GpuArray:
             cdef unsigned int i
             if len(newstrides) != self.ga.nd:
                 raise ValueError("new strides are the wrong length")
+            if not strides_ok(self,  newstrides):
+                raise ValueError("new strides go outside of allocated memory")
             for i in range(self.ga.nd):
                 self.ga.strides[i] = newstrides[i]
             array_fix_flags(self)

pygpu/tests/test_gpu_ndarray.py

@@ -8,6 +8,7 @@
 
 import numpy
 
+from nose.tools import assert_raises
 import pygpu
 from pygpu.gpuarray import GpuArray, GpuContext, GpuKernel
 
@@ -448,6 +449,32 @@ def reshape(shps, offseted, order1, order2):
     assert numpy.allclose(outc, numpy.asarray(outg))
 
 
+def test_strides():
+    yield strides_, (4, 4), 'c', 1, (4, 4)
+    yield strides_, (4, 4), 'c', 1, (4, 16)
+    yield strides_, (4, 4), 'c', 1, (16, 4)
+    yield strides_, (4, 4), 'c', 1, (16, 8)
+    yield strides_, (4, 4), 'c', 1, (16, 0)
+    yield strides_, (4, 4), 'c', -1, (-20, 4)
+    yield strides_, (4, 4), 'c', -1, (-12, 4)
+
+
+def set_strides(a, newstr):
+    a.strides = newstr
+
+
+def strides_(shp, order, sliced, newstr):
+    ac, ag = gen_gpuarray(shp, 'float32', sliced=sliced, order=order, ctx=ctx)
+    try:
+        ac.strides = newstr
+    except ValueError:
+        assert_raises(ValueError, set_strides, ag, newstr)
+        return
+    ag.strides = newstr
+    check_flags(ag, ac)
+    assert numpy.allclose(ac, numpy.asarray(ag))
+
+
 def test_transpose():
     for shp in [(2, 3), (4, 8, 9), (1, 2, 3, 4)]:
         for offseted in [True, False]: