Merge pull request #446 from abergeron/check_offset

Add offset in DEBUG

Author: nouiz

pygpu/basic.py

@@ -1,10 +1,11 @@
 from string import Template
-from .gpuarray import GpuArray, GpuKernel
+from .gpuarray import GpuArray, GpuKernel, SIZE
 
 
 def _generate_kernel(ctx, cols, upper=True):
     tmpl = Template("""
-    KERNEL void extract_tri(GLOBAL_MEM ga_float *a, ga_uint N) {
+    KERNEL void extract_tri(GLOBAL_MEM ga_float *a, ga_size a_off, ga_uint N) {
+        a = (GLOBAL_MEM ga_float *)(((char *)a) + a_off);
         unsigned int idx = GID_1 * LDIM_0 * GDIM_0 +
                            GID_0 * LDIM_0 + LID_0;
         unsigned int ix = idx/${cols};
@@ -20,7 +21,7 @@ def _generate_kernel(ctx, cols, upper=True):
     else:
         le = '<'
     src = tmpl.substitute(cols=cols, le=le)
-    spec = [GpuArray, 'uint32']
+    spec = [GpuArray, SIZE, 'uint32']
     k = GpuKernel(src, "extract_tri", spec, context=ctx)
     return k
 
@@ -40,7 +41,7 @@ def triu(A, inplace=True):
         upper = True
         cols = A.shape[1]
     k = _generate_kernel(A.context, cols, upper)
-    k(A, A.shape[0] * A.shape[1], n=A.shape[0] * A.shape[1])
+    k(A, A.offset, A.shape[0] * A.shape[1], n=A.shape[0] * A.shape[1])
     return A
 
 
@@ -59,5 +60,5 @@ def tril(A, inplace=True):
         upper = False
         cols = A.shape[1]
     k = _generate_kernel(A.context, cols, upper)
-    k(A, A.shape[0] * A.shape[1], n=A.shape[0] * A.shape[1])
+    k(A, A.offset, A.shape[0] * A.shape[1], n=A.shape[0] * A.shape[1])
     return A

pygpu/gpuarray.pyx

@@ -2223,13 +2223,40 @@ cdef class GpuArray:
         def __get__(self):
             return self.ga.offset
 
+    property data:
+        """Return a pointer to the raw OpenCL buffer object.
+
+        This will fail for arrays that have an offset.
+        """
+        def __get__(self):
+            if self.context.kind != b"opencl":
+                raise TypeError("This is for OpenCL arrays.")
+            if self.offset != 0:
+                raise ValueError("This array has an offset.")
+            # This wizadry grabs the actual backend pointer since it's
+            # guarenteed to be the first element of the gpudata
+            # structure.
+            return <size_t>((<void **>self.ga.data)[0])
+
+    property base_data:
+        "Return a pointer to the backing OpenCL object."
+        def __get__(self):
+            if self.context.kind != b"opencl":
+                raise TypeError("This is for OpenCL arrays.")
+            # This wizadry grabs the actual backend pointer since it's
+            # guarenteed to be the first element of the gpudata
+            # structure.
+            return <size_t>((<void **>self.ga.data)[0])
+
     property gpudata:
         "Return a pointer to the raw backend object."
         def __get__(self):
+            if self.context.kind != b"cuda":
+                raise TypeError("This is for CUDA arrays.")
             # This wizadry grabs the actual backend pointer since it's
             # guarenteed to be the first element of the gpudata
             # structure.
-            return <size_t>((<void **>self.ga.data)[0])
+            return <size_t>((<void **>self.ga.data)[0]) + self.offset
 
     def __str__(self):
         return str(numpy.asarray(self))

src/gpuarray_array.c

@@ -118,10 +118,20 @@ int GpuArray_empty(GpuArray *a, gpucontext *ctx, int typecode,
     size *= d;
   }
 
+  /* We add a offset of 64 to all arrays in DEBUG to help catch errors. */
+#ifdef DEBUG
+  assert(SIZE_MAX - size > 64);
+  size += 64;
+#endif
+
   a->data = gpudata_alloc(ctx, size, NULL, 0, &res);
   if (a->data == NULL) return ctx->err->code;
   a->nd = nd;
+#ifdef DEBUG
+  a->offset = 64;
+#else
   a->offset = 0;
+#endif
   a->typecode = typecode;
   a->dimensions = calloc(nd, sizeof(size_t));
   a->strides = calloc(nd, sizeof(ssize_t));