added verbose flag to NaiveAlgo to control printing to stdout, stderr

elemwise.py

@@ -201,6 +201,7 @@ class RecAlgo(object):
         return self.c_src_kernel(node, nodename) + self.c_src_callkernel(node, nodename)
 
 class NaiveAlgo(object):
+    verbose = False
     def __init__(self, **kwargs):
         self.__dict__.update(kwargs)
 
@@ -287,7 +288,7 @@ class NaiveAlgo(object):
             #print >> sio, indent, "const float * i%i" % ipos, '= i%i_data', ''
         print >> sio, "}"
 
-        print sio.getvalue()
+        #print sio.getvalue()
         return sio.getvalue()
 
     def c_src_kernel_tiling(self, node, nodename):
@@ -425,7 +426,7 @@ class NaiveAlgo(object):
         n_out = len(node.outputs)
         sio = StringIO.StringIO()
 
-        if nd not in (4,):
+        if nd not in (2,):
             return sio.getvalue()
 
         # print some leading comments to make the code easier to read
@@ -620,7 +621,7 @@ class NaiveAlgo(object):
         print >> sio, "    }"
         print >> sio, "}"
 
-        print sio.getvalue()
+        #print sio.getvalue()
         return sio.getvalue()
 
     def c_src_callkernel(self, node, nodename):
@@ -672,12 +673,13 @@ class NaiveAlgo(object):
             %(output_params)s)
         {
             numEls = %(prod_dims)s;
+        """ %locals()
+        if self.verbose:
+            print >> sio, """
                 std::cerr << "calling kernel_%(scalar_op)s_%(nodename)s_%(id_self)s     w numEls" << numEls << "\\n";
             """ %locals()
-        # DEBUGPRINT
             print >> sio, 'std::cerr << ' + " << ' ' <<  ".join(['"  "']+list("dims[%i]"%di
                 for di in xrange(nd)) + ["'\\n';"])
-        # DEBUGPRINT
             for ipos in xrange(len(node.inputs)):
                 print >> sio, """
                 std::cerr << "   %(ipos)s " << 
@@ -700,7 +702,7 @@ class NaiveAlgo(object):
                         nd_collapse_size = nd_collapse_size_%(ipos)s;
                     }
                 """ %locals()
-        # DEBUGPRINT
+        if self.verbose:
             print >> sio, 'std::cerr << "  nd_collapse " << nd_collapse << " " << nd_collapse_size << "\\n";'
         for ipos in xrange(len(node.inputs)):
             print >> sio, "int local_i%(ipos)s_str[%(nd)s];"%locals()
@@ -714,7 +716,6 @@ class NaiveAlgo(object):
         for d in xrange(nd):
             print >> sio, "local_dims[%(d)s] = (%(d)s == nd_collapse) ? nd_collapse_size : dims[%(d)s];"%locals()
 
-
         def launch_Ccontiguous(nodename, id_self, scalar_op):
             kernel_call_args = ["numEls"]
             for ipos in xrange(len(node.inputs)):
@@ -739,11 +740,12 @@ class NaiveAlgo(object):
                 return 0;
                 """ %locals()
 
-        def launch_tile4():
-            if (False and nd == 4): # tiling kernel
+        def launch_nd_collapse_2(nodename, id_self, scalar_op):
+            if self.verbose:
+                print >> sio, """
+                std::cerr << "   Running tiling 2D \\n";
+                """
             print >> sio, """
-                {
-                    std::cerr << "   Running tiling 4D \\n";
             dim3 gridDim(dims[0], dims[1]);
             dim3 blockDim;
             if (0) {
@@ -782,8 +784,6 @@ class NaiveAlgo(object):
                 }                         
             }
             return 0;
-                }
-            }
                 """ %locals()
 
         def launch_General(nodename, id_self, scalar_op):
@@ -792,7 +792,7 @@ class NaiveAlgo(object):
             kernel_call_args.extend("dims[%i]"%di for di in xrange(nd))
             for ipos in xrange(len(node.inputs)):
                 kernel_call_args.append(
-                        ", ".join(["i%i_data"%ipos] + list("local_i%i_str[%i]"%(ipos, di) for di in xrange(nd)))
+                        ", ".join(["i%i_data"%ipos] + list("i%i_str[%i]"%(ipos, di) for di in xrange(nd)))
                         )
                 #strides = ", ".join("i%i_str[%i]"%(ipos, di) for di in xrange(nd))
                 #kernel_call_args.append( "%s, i%i_data" % (strides, ipos))
@@ -803,8 +803,11 @@ class NaiveAlgo(object):
                 #strides = ", ".join("o%i_str[%i]"%(ipos, di) for di in xrange(nd))
                 #kernel_call_args.append( "%s, o%i_data" % (strides, ipos))
             kernel_call_args = ", ".join(kernel_call_args)
+            if self.verbose:
                 print >> sio, """
                     std::cerr << "   Running general version \\n";
+                    """
+            print >> sio, """
                 int threads_per_block = std::min(numEls, (unsigned int)NUM_VECTOR_OP_THREADS_PER_BLOCK);
                 int n_blocks = std::min(numEls/threads_per_block + (numEls %% threads_per_block?1:0), (unsigned int)NUM_VECTOR_OP_BLOCKS);
                 kernel_%(scalar_op)s_%(nodename)s_%(id_self)s<<<n_blocks, threads_per_block>>>(%(kernel_call_args)s);
@@ -837,11 +840,13 @@ class NaiveAlgo(object):
 
 
     def c_support_code_apply(self, node, nodename):
-        return self.c_src_kernel(node, nodename) \
-                + self.c_src_kernel_Ccontiguous(node, nodename) \
-                + self.c_src_kernel_tiling(node, nodename) \
-                + self.c_src_kernel_tiling_less_registers(node, nodename) \
-                + self.c_src_callkernel(node, nodename)
+        return "".join([ 
+            self.c_src_kernel(node, nodename),
+            self.c_src_kernel_Ccontiguous(node, nodename),
+            #self.c_src_kernel_tiling(node, nodename),
+            #self.c_src_kernel_tiling_less_registers(node, nodename),
+            self.c_src_callkernel(node, nodename),
+            ])
 
     def c_code(self, node, nodename, inputs, outputs, sub):
         d = dict(sub)