Merge pull request #442 from abergeron/fixes

Mixed fixes

Author: nouiz

pygpu/_elemwise.pyx

@@ -49,6 +49,9 @@ cdef class arg:
         memset(&self.a, 0, sizeof(gpuelemwise_arg))
 
     def __init__(self, name, type, read=False, write=False, scalar=False):
+        # Make sure to clear previous storage
+        # __init__ may be called more than once
+        free(self.a.name)
         self.a.name = strdup(to_bytes(name))
         if self.a.name is NULL:
             raise MemoryError
@@ -63,6 +66,9 @@ cdef class arg:
         if self.a.flags == 0:
             raise ValueError('no flags specified for arg %s' % (name,))
 
+    def __dealloc__(self):
+        free(self.a.name)
+
     property name:
         def __get__(self):
             return self.a.name.decode('ascii')

pygpu/elemwise.py

@@ -4,7 +4,8 @@
 from . import gpuarray
 from ._elemwise import GpuElemwise, arg
 
-__all__ = ['GpuElemwise', 'elemwise1', 'elemwise2', 'ielemwise2', 'compare']
+__all__ = ['GpuElemwise', 'arg', 'as_argument',
+           'elemwise1', 'elemwise2', 'ielemwise2', 'compare']
 
 
 def _dtype(o):

pygpu/gpuarray.pyx

@@ -2455,7 +2455,7 @@ cdef class GpuKernel:
                     raise ValueError, "nd mismatch for gs (int)"
                 gs[0] = py_gs
             elif isinstance(py_gs, (list, tuple)):
-                if len(py_gs) < 3:
+                if len(py_gs) > 3:
                     raise ValueError, "gs is not of length 3 or less"
                 if len(py_ls) != nd:
                     raise ValueError, "nd mismatch for gs (tuple)"

src/gpuarray_buffer_opencl.c

@@ -58,7 +58,7 @@ static int cl_get_device_count(unsigned int platform, unsigned int* devcount) {
   cl_platform_id p;
   cl_uint numd;
   cl_int err;
-  unsigned int platcount;
+  unsigned int platcount = 0;
 
   /* This will load the library if needed */
   GA_CHECK(cl_get_platform_count(&platcount));

src/gpuarray_collectives_cuda_nccl.c

@@ -180,7 +180,6 @@ static inline ncclDataType_t convert_data_type(int typecode) {
   case GA_LONG: return ncclInt64;
   case GA_ULONG: return ncclUint64;
   case GA_HALF: return ncclHalf;
-  case GA_FLOAT16: return ncclHalf;
   }
   return nccl_NUM_TYPES;
 }
@@ -237,8 +236,9 @@ static inline int check_restrictions(gpudata *src, size_t offsrc,
 static int reduce(gpudata *src, size_t offsrc, gpudata *dest, size_t offdest,
                   size_t count, int typecode, int opcode, int root,
                   gpucomm *comm) {
-  ncclRedOp_t op;
-  ncclDataType_t datatype;
+  // need dummy init so that compiler shuts up
+  ncclRedOp_t op = nccl_NUM_OPS;
+  ncclDataType_t datatype = nccl_NUM_TYPES;
   gpudata *dst = NULL;
   int rank = 0;
   cuda_context *ctx;
@@ -286,8 +286,9 @@ static int reduce(gpudata *src, size_t offsrc, gpudata *dest, size_t offdest,
 static int all_reduce(gpudata *src, size_t offsrc, gpudata *dest,
                       size_t offdest, size_t count, int typecode, int opcode,
                       gpucomm *comm) {
-  ncclRedOp_t op;
-  ncclDataType_t datatype;
+  // need dummy init so that compiler shuts up
+  ncclRedOp_t op = nccl_NUM_OPS;
+  ncclDataType_t datatype = nccl_NUM_TYPES;
   cuda_context *ctx;
 
   ASSERT_BUF(src);
@@ -323,8 +324,9 @@ static int all_reduce(gpudata *src, size_t offsrc, gpudata *dest,
 static int reduce_scatter(gpudata *src, size_t offsrc, gpudata *dest,
                           size_t offdest, size_t count, int typecode,
                           int opcode, gpucomm *comm) {
-  ncclRedOp_t op;
-  ncclDataType_t datatype;
+  // need dummy init so that compiler shuts up
+  ncclRedOp_t op = nccl_NUM_OPS;
+  ncclDataType_t datatype = nccl_NUM_TYPES;
   int ndev = 0;
   size_t resc_size;
   cuda_context *ctx;
@@ -368,7 +370,8 @@ static int reduce_scatter(gpudata *src, size_t offsrc, gpudata *dest,
  */
 static int broadcast(gpudata *array, size_t offset, size_t count, int typecode,
                      int root, gpucomm *comm) {
-  ncclDataType_t datatype;
+  // need dummy init so that compiler shuts up
+  ncclDataType_t datatype = nccl_NUM_TYPES;
   int rank = 0;
   cuda_context *ctx;
 
@@ -407,7 +410,8 @@ static int broadcast(gpudata *array, size_t offset, size_t count, int typecode,
 static int all_gather(gpudata *src, size_t offsrc, gpudata *dest,
                       size_t offdest, size_t count, int typecode,
                       gpucomm *comm) {
-  ncclDataType_t datatype;
+  // need dummy init so that compiler shuts up
+  ncclDataType_t datatype = nccl_NUM_TYPES;
   int ndev = 0;
   size_t resc_size;
   cuda_context *ctx;