complete support for dumping out dense matrices for hw

scripts/util/util.py

@@ -127,9 +127,10 @@ class NumpyNPYArrayLoader:
     def __init__(self):
         pass
     def load(self, path):
-        np_byte_array = numpy.load(path)
-        np_fp_array = convert_numpy_byte_array_to_fp_array(np_byte_array)
-        coo = scipy.sparse.coo_array(np_fp_array)
+        np_array = numpy.load(path)
+        if (np_array.dtype == numpy.dtype('S16')):
+            np_array = convert_numpy_byte_array_to_fp_array(np_array)
+        coo = scipy.sparse.coo_array(np_array)
         return coo
 
 def shape_str(shape):
@@ -138,19 +139,11 @@ def shape_str(shape):
 
 # FIXME: This fixed point number of decimals may not be enough
 def array_str(array):
-    if isinstance(array[0], float):
-        return ' '.join([str(item) for item in array])
-        # return ' '.join(['{:5.5f}'.format(item) for item in array])
-
-    return ' '.join([str(item) for item in array])
+    return ' '.join([str(item) for _, item in numpy.ndenumerate(array)])
 
 
 def array_newline_str(array):
-    if isinstance(array[0], float):
-        return '\n'.join([str(item) for item in array])
-        # return '\n'.join(['{:5.5f}'.format(item) for item in array])
-
-    return '\n'.join([str(item) for item in array])
+    return '\n'.join([str(item) for _, item in numpy.ndenumerate(array)])
 
 
 # InputCacheSuiteSparse attempts to avoid reading the same tensor from disk multiple
@@ -194,7 +187,7 @@ def convert_format(self, coo, format_str):
             if format_str == "dense":
                 return coo.todense()
             elif format_str == "denseT":
-                return coo.todense().getT()
+                return numpy.transpose(coo.todense())
             if format_str == "dcsr":
                 csr = scipy.sparse.csr_matrix(coo)
                 has_row = [rc > 0 for rc in csr.getnnz(1)]
@@ -641,6 +634,52 @@ def writeout_separate_sparse_only(self, coo, dir_path, tensorname, format_str="s
             with open(filename, "w") as ofile:
                 ofile.write(array_newline_str(dcsc.data))
 
+        elif format_str == "dd01":
+            dense_dir = Path(dir_path)
+            dense_dir.mkdir(parents=True, exist_ok=True, mode=0o777)
+
+            dense = self.convert_format(coo, "dense")
+
+            if not hw:
+                filename = os.path.join(dense_dir, tensorname + "_shape.txt")
+            else:
+                filename = os.path.join(dir_path, "tensor_" + tensorname + "_mode_shape")
+            with open(filename, "w") as ofile:
+                ofile.write(array_newline_str(dense.shape))
+
+            if not hw:
+                filename = os.path.join(dense_dir, tensorname + "_vals.txt")
+            else:
+                filename = os.path.join(dir_path, "tensor_" + tensorname + "_mode_vals")
+            with open(filename, "w") as ofile:
+                ofile.write(array_newline_str(dense.data))
+
+        elif format_str == "dd10":
+            denseT_dir = Path(dir_path)
+            denseT_dir = denseT_dir.mkdir(parents=True, exist_ok=True, mode =0o777)
+
+            # here, the ndarray will be transpose, so we can get the data in column major order
+            denseT = self.convert_format(coo, "denseT")
+
+            if not hw:
+                filename = os.path.join(denseT_dir, tensorname + "_shape.txt")
+            else:
+                filename = os.path.join(dir_path, "tensor_" + tensorname + "_mode_shape")
+            with open(filename, "w") as ofile:
+                # the transposition in convert_format will change the shape of the matrix,
+                # need to transpose it back to get the correct shape here
+                ofile.write(array_newline_str(numpy.transpose(denseT).shape))
+
+            if not hw:
+                filename = os.path.join(denseT_dir, tensorname + "_vals.txt")
+            else:
+                filename = os.path.join(dir_path, "tensor_" + tensorname + "_mode_vals")
+            with open(filename, "w") as ofile:
+                ofile.write(array_newline_str(denseT.data))
+
+
+
+
 
 # UfuncInputCache attempts to avoid reading the same tensor from disk multiple
 # times in a benchmark run.