diff options
Diffstat (limited to 'verif')
-rw-r--r-- | verif/generator/tosa_arg_gen.py | 12 | ||||
-rw-r--r-- | verif/generator/tosa_test_gen.py | 24 |
2 files changed, 21 insertions, 15 deletions
diff --git a/verif/generator/tosa_arg_gen.py b/verif/generator/tosa_arg_gen.py index 193da73..35253e0 100644 --- a/verif/generator/tosa_arg_gen.py +++ b/verif/generator/tosa_arg_gen.py @@ -213,17 +213,17 @@ class TosaTensorGen: assert rank == 3 values_in_shape = testGen.makeShape(rank) + K = values_in_shape[1] # ignore max batch size if target shape is set if testGen.args.max_batch_size and not testGen.args.target_shapes: values_in_shape[0] = min(values_in_shape[0], testGen.args.max_batch_size) - W = testGen.randInt( - testGen.args.tensor_shape_range[0], testGen.args.tensor_shape_range[1] - ) - # Constrict W if one dimension is too large to keep tensor size reasonable - if max(values_in_shape) > 5000: - W = testGen.randInt(0, 16) + # Make sure W is not greater than K, as we can only write each output index + # once (having a W greater than K means that you have to repeat a K index) + W_min = min(testGen.args.tensor_shape_range[0], K) + W_max = min(testGen.args.tensor_shape_range[1], K) + W = testGen.randInt(W_min, W_max) if W_min < W_max else W_min input_shape = [values_in_shape[0], W, values_in_shape[2]] diff --git a/verif/generator/tosa_test_gen.py b/verif/generator/tosa_test_gen.py index ba10dcf..53b0b75 100644 --- a/verif/generator/tosa_test_gen.py +++ b/verif/generator/tosa_test_gen.py @@ -1771,22 +1771,28 @@ class TosaTestGen: def build_scatter(self, op, values_in, input, validator_fcns=None, error_name=None): - # Create a new indicies tensor - # here with data that doesn't exceed the dimensions of the values_in tensor - K = values_in.shape[1] # K W = input.shape[1] # W - indicies_arr = np.int32( - self.rng.integers(low=0, high=K, size=[values_in.shape[0], W]) - ) # (N, W) - indicies = self.ser.addConst(indicies_arr.shape, DType.INT32, indicies_arr) + + # Create an indices tensor here with data that doesn't exceed the + # dimension K of the values_in tensor and does NOT repeat the same K + # location as needed by the spec: + # "It is not permitted to repeat the same output index within a single + # SCATTER operation and so each output index occurs at most once." + assert K >= W + arr = [] + for n in range(values_in.shape[0]): + # Get a shuffled list of output indices and limit it to size W + arr.append(self.rng.permutation(K)[:W]) + indices_arr = np.array(arr, dtype=np.int32) # (N, W) + indices = self.ser.addConst(indices_arr.shape, DType.INT32, indices_arr) result_tens = OutputShaper.scatterOp( - self.ser, self.rng, values_in, indicies, input, error_name + self.ser, self.rng, values_in, indices, input, error_name ) # Invalidate Input/Output list for error if checks. - input_list = [values_in.name, indicies.name, input.name] + input_list = [values_in.name, indices.name, input.name] output_list = [result_tens.name] pCount, cCount = op["operands"] num_operands = pCount + cCount |