Merge pull request fastmachinelearning#70 from i-colbert/fix/matvec_a…

…ccumulator_range

Fix (utils): Updating matvec accumulator range calculation

src/qonnx/util/basic.py

@@ -209,16 +209,12 @@ def pad_tensor_to_multiple_of(ndarray, pad_to_dims, val=0, distr_pad=False):
 
 
 def calculate_matvec_accumulator_range(matrix: np.ndarray, vec_dt: DataType):
-    """Calculate the minimum and maximum possible result (accumulator) values
-    for a dot product x * A, given matrix A of dims (MW, MH), and vector (1, MW)
-    with datatype vec_dt. Returns (acc_min, acc_max).
-    """
-    max_weight = abs(matrix).sum(axis=0).max()
-    max_input = max(abs(vec_dt.min()), abs(vec_dt.max()))
-    max_value = max_input * max_weight
-    # If either the weight and input datatypes are signed, then the minimum
-    # value that their accumulated product can be is -max_value. Else, it's 0.
-    min_value = -max_value if (matrix.min() < 0) or vec_dt.signed() else 0
+    """Calculate the minimum and maximum possible result (accumulator) values for a dot product x * A,
+    given matrix A of dims (MW, MH), and vector (1, MW) with datatype vec_dt. Returns (acc_min, acc_max)."""
+    max_vectors = np.where(matrix > 0, vec_dt.max(), vec_dt.min())
+    min_vectors = np.where(matrix > 0, vec_dt.min(), vec_dt.max())
+    max_value = (matrix * max_vectors).sum(axis=0).max()
+    min_value = (matrix * min_vectors).sum(axis=0).min()
     return (min_value, max_value)