Just to complete, added function to create CUDA Arrays directly from …

…GPUArrays, maybe in future Texturea Objects and Surfeces Objects could be supported ..
inducer · Oct 30, 2015 · 6b0cfc4 · 6b0cfc4
1 parent 0797e93
commit 6b0cfc4
Show file tree

Hide file tree

Showing 2 changed files with 72 additions and 4 deletions.
diff --git a/pycuda/driver.py b/pycuda/driver.py
@@ -792,6 +792,74 @@ def np_to_array(nparray, order, allowSurfaceBind=False):
       copy2D(aligned=True)
       return cudaArray
 
+def gpuarray_to_array(gpuarray, order, allowSurfaceBind=False):
+
+    case = order in ["C","F"]
+    if not case:
+      raise LogicError("order must be either F or C")
+
+    dimension = len(gpuarray.shape)
+    if dimension == 2:
+      if order == "C": stride = 0
+      if order == "F": stride = -1
+      h, w = gpuarray.shape
+      d = 1
+      if allowSurfaceBind:
+        descrArr = ArrayDescriptor3D()
+        descrArr.width = int(w)
+        descrArr.height = int(h)
+        descrArr.depth = int(d)
+      else:
+        descrArr = ArrayDescriptor()
+        descrArr.width = int(w)
+        descrArr.height = int(h)
+    elif dimension == 3:
+        if order == "C": stride = 1
+        if order == "F": stride = 1
+        d, h, w = gpuarray.shape
+        descrArr = ArrayDescriptor3D()
+        descrArr.width = int(w)
+        descrArr.height = int(h)
+        descrArr.depth = int(d)
+    else:
+        raise LogicError("CUDArray dimensions 2 and 3 supported in CUDA at the moment ... ")
+
+    if gpuarray.dtype == np.complex64:
+      descrArr.format = array_format.SIGNED_INT32 # Reading data as int2 (hi=re,lo=im) structure
+      descrArr.num_channels = 2
+    elif gpuarray.dtype == np.float64:
+      descrArr.format = array_format.SIGNED_INT32 # Reading data as int2 (hi,lo) structure
+      descrArr.num_channels = 2
+    elif gpuarray.dtype == np.complex128:
+      descrArr.format = array_format.SIGNED_INT32 # Reading data as int4 (re=(hi,lo),im=(hi,lo)) structure
+      descrArr.num_channels = 4
+    else:
+      descrArr.format = dtype_to_array_format(gpuarray.dtype)
+      descrArr.num_channels = 1
+
+    if allowSurfaceBind:
+      if dimension==2:  descrArr.flags |= array3d_flags.ARRAY3D_LAYERED
+      descrArr.flags |= array3d_flags.SURFACE_LDST
+
+    cudaArray = Array(descrArr)
+    if allowSurfaceBind or dimension==3:
+      copy3D = Memcpy3D()
+      copy3D.set_src_device(gpuarray.ptr)
+      copy3D.set_dst_array(cudaArray)
+      copy3D.width_in_bytes = copy3D.src_pitch = gpuarray.strides[stride]
+      copy3D.src_height = copy3D.height = int(h)
+      copy3D.depth = int(d)
+      copy3D()
+      return cudaArray
+    else:
+      copy2D = Memcpy2D()
+      copy2D.set_src_device(gpuarray.ptr)
+      copy2D.set_dst_array(cudaArray)
+      copy2D.width_in_bytes = copy2D.src_pitch = gpuarray.strides[stride]
+      copy2D.src_height = copy2D.height = int(h)
+      copy2D(aligned=True)
+      return cudaArray
+
 def make_multichannel_2d_array(ndarray, order):
     """Channel count has to be the first dimension of the C{ndarray}."""
 

diff --git a/test/test_driver.py b/test/test_driver.py
@@ -442,8 +442,8 @@ def test_3d_fp_surfaces(self):
             cuBlock = (npoints,npoints,npoints)
           cuGrid   = (npoints//cuBlock[0]+1*(npoints % cuBlock[0] != 0 ),npoints//cuBlock[1]+1*(npoints % cuBlock[1] != 0 ),npoints//cuBlock[2]+1*(npoints % cuBlock[1] != 0 ))
           copy_texture.prepare('Pi')#,texrefs=[mtx_tex])
-          A_cpu = np.zeros([npoints,npoints,npoints],order=orden,dtype=prec) # To initialize surface with zeros
-          cudaArray = drv.np_to_array(A_cpu,orden,allowSurfaceBind=True)
+          A_gpu2 = gpuarray.zeros_like(A_gpu) # To initialize surface with zeros
+          cudaArray = drv.gpuarray_to_array(A_gpu2,orden,allowSurfaceBind=True)
           A_cpu = A_gpu.get() # To remember original array
           mtx_tex.set_array(cudaArray)
           copy_texture.prepared_call(cuGrid,cuBlock,A_gpu.gpudata, np.int32(0)) # Write random array
@@ -497,8 +497,8 @@ def test_2d_fp_surfaces(self):
             cuBlock = (npoints,npoints,1)
           cuGrid   = (npoints//cuBlock[0]+1*(npoints % cuBlock[0] != 0 ),npoints//cuBlock[1]+1*(npoints % cuBlock[1] != 0 ),1)
           copy_texture.prepare('Pi')#,texrefs=[mtx_tex])
-          A_cpu = np.zeros([npoints,npoints],order=orden,dtype=prec) # To initialize surface with zeros
-          cudaArray = drv.np_to_array(A_cpu,orden,allowSurfaceBind=True)
+          A_gpu2 = gpuarray.zeros_like(A_gpu) # To initialize surface with zeros
+          cudaArray = drv.gpuarray_to_array(A_gpu2,orden,allowSurfaceBind=True)
           A_cpu = A_gpu.get() # To remember original array
           mtx_tex.set_array(cudaArray)
           copy_texture.prepared_call(cuGrid,cuBlock,A_gpu.gpudata, np.int32(0)) # Write random array