conv_nchw parameter updated to the one generates mobilenet benchmarks, doc typo fixed (#345)

* conv_nchw parameter updated to the one which generates mobilenet benchmarks, doc typo fixed

* removed unused variables

topi/python/topi/cuda/conv2d_nchw.py

@@ -37,12 +37,10 @@ def schedule_conv2d_small_batch(outs):
         # sheduler params
         num_thread = 8
         vthread = 2
-        out_filter = min(64, util.get_const_int(Filter.shape[0]))
-        in_filter = util.get_const_int(Filter.shape[1])
-        opart2 = max(1, out_filter//8)
-        ofactor = out_filter
-        wfactor = block_h
-        ifactor = max(8, in_filter//4)
+        opart2 = 4
+        ofactor = 64
+        wfactor = 56
+        ifactor = 8
         sfactor = max(1, ofactor//(opart2*2))
         spart = max(1, (wfactor + vthread-1) // vthread)
         block_x = tvm.thread_axis("blockIdx.x")

topi/python/topi/nn/convolution.py

@@ -7,7 +7,7 @@ from .pad import pad, _spatial2d_pad_option
 
 
 def conv2d_nchw(Input, Filter, stride, padding):
-    """Convolution operator in HWCN layout.
+    """Convolution operator in NCHW layout.
 
     Parameters
     ----------