Add smem_alloc_granularity.

inducer · Aug 13, 2011 · 280d3d9 · 280d3d9
1 parent 44007e8
commit 280d3d9
Show file tree

Hide file tree

Showing 2 changed files with 8 additions and 2 deletions.
diff --git a/doc/source/util.rst b/doc/source/util.rst
@@ -103,10 +103,14 @@ Device Metadata and Occupancy
   .. attribute:: registers
   .. attribute:: shared_memory
   .. attribute:: smem_granularity
-    
+
     The number of threads that participate in banked, simultaneous access
     to shared memory.
 
+  .. attribute:: smem_alloc_granularity
+
+    The size of the smallest possible (non-empty) shared memory allocation.
+
   .. method:: align_bytes(word_size=4)
 
     The distance between global memory base addresses that 

diff --git a/pycuda/tools.py b/pycuda/tools.py
@@ -240,8 +240,10 @@ def __init__(self, dev=None):
         self.shared_memory = dev.get_attribute(drv.device_attribute.MAX_SHARED_MEMORY_PER_BLOCK)
 
         if dev.compute_capability() >= (2,0):
+            self.smem_alloc_granularity = 128
             self.smem_granularity = 32
         else:
+            self.smem_alloc_granularity = 512
             self.smem_granularity = 16
 
         if dev.compute_capability() >= (2,0):
@@ -291,7 +293,7 @@ def __init__(self, devdata, threads, shared_mem=0, registers=0):
 
         # copied literally from occupancy calculator
         alloc_warps = _int_ceiling(threads/devdata.warp_size)
-        alloc_smem = _int_ceiling(shared_mem, devdata.smem_granularity)
+        alloc_smem = _int_ceiling(shared_mem, devdata.smem_alloc_granularity)
         if devdata.register_allocation_unit == "warp":
             alloc_regs = alloc_warps*32*registers
         elif devdata.register_allocation_unit == "block":