ekalda commented on a change in pull request #9576:
URL: https://github.com/apache/tvm/pull/9576#discussion_r759506677
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -831,6 +832,170 @@ def __call__(self, *args, **kwargs):
pass
+class MeanRewriter(DFPatternCallback):
+ """Convert ethosu.mean composite functions to to an equivalent
legalization:
+ - Case 1 (axis == [1, 2] and keepsdims == True):
+ ethosu_depthwise_conv2d + ethosu_binary_elementwise
+ - Case 2 (ifm qparams == ofm qparams): ethosu_pooling
+ - Case 3 (else): ethosu_depthwise_conv2d
+ """
+
+ def __init__(self):
+ super().__init__(require_type=True)
+ self.pattern = (
+ wildcard().has_attr({"Composite":
ethosu_patterns.MeanParams.composite_name})
+ )(wildcard())
+
+ def callback(
+ self, pre: tvm.relay.Expr, post: tvm.relay.Expr, node_map:
tvm.ir.container.Map
+ ) -> tvm.relay.Expr:
+ params = ethosu_patterns.MeanParams(post.op.body)
+ params.ifm.tensor = post.args[0]
+
+ ifm_shape = params.ifm.shape
+ ofm_shape = params.ofm.shape
+ lut = relay.const([], "int8")
+ axis = params.axis
+ reduced_op = params.ifm.tensor
+
+ # Enforce 4d input
+ if len(ifm_shape) < 4:
+ axis = [x + 1 for x in axis]
+ if len(ifm_shape) == 3:
+ ifm_shape = [1, params.height, params.width, ifm_shape[2]]
+ else:
+ ifm_shape = [1, params.height, params.width, 1]
+ reduced_op = relay.reshape(reduced_op, ifm_shape)
+
+ filter_height = ifm_shape[1] if 1 in axis else 1
+ filter_width = ifm_shape[2] if 2 in axis else 1
+ in_channels = out_channels = ifm_shape[-1]
+
+ # If the height is greater than max kernel height, reshape the input
+ # from [filter_height, filter_width] to [1,
(filter_height*filter_width)]
+ # only in the case the axis is [1, 2].
+ if axis == [1, 2] and filter_height > 64:
+ ifm_shape = (ifm_shape[0], 1, filter_height * filter_width,
in_channels)
+ filter_width = filter_height * filter_width
+ filter_height = 1
+ reduced_op = relay.reshape(reduced_op, ifm_shape)
+
+ if axis == [1, 2] and params.keepdims:
+ weight_scale = 1
+ weight_values = np.ones([out_channels, filter_height,
filter_width, in_channels])
+ scale_bias = vela_api.pack_biases(
+ biases=np.zeros(ifm_shape[-1]),
+ ifm_scale=params.ifm.q_params.scale_f32,
+ ifm_dtype=np.dtype(params.ifm.dtype),
+ weight_scales=np.array([weight_scale], dtype=np.float),
+ ofm_scale=params.ofm.q_params.scale_f32,
+ is_activation_tanh_or_sigmoid=False,
+ )
+
+ reduced_op = ethosu_ops.ethosu_depthwise_conv2d(
+ ifm=reduced_op,
+ weight=relay.const(weight_values, params.ifm.dtype),
+ scale_bias=relay.const(scale_bias, "uint8"),
+ lut=lut,
+ ifm_scale=float(params.ifm.q_params.scale_f32),
+ ifm_zero_point=int(params.ifm.q_params.zero_point),
+ weight_zero_point=0,
+ ofm_scale=float(params.ofm.q_params.scale_f32),
+ ofm_zero_point=int(params.ofm.q_params.zero_point),
+ kernel_shape=(filter_height, filter_width),
+ ofm_channels=out_channels,
+ ofm_dtype="int16",
+ )
+
+ n = int(filter_height * filter_width)
+ eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
+
+ scalar_tensor = relay.const(np.ones([1, 1, 1, 1], dtype="uint8"),
dtype="uint8")
+
+ reduced_op = ethosu_ops.ethosu_binary_elementwise(
+ ifm=reduced_op,
+ ifm2=scalar_tensor,
+ lut=lut,
+ operator_type="MUL",
+ ifm_scale=float(params.ofm.q_params.scale_f32),
+ ifm_zero_point=int(params.ofm.q_params.zero_point),
+ ifm2_scale=1 / (n - eps),
+ ifm2_zero_point=0,
+ ofm_scale=float(params.ofm.q_params.scale_f32),
+ ofm_zero_point=int(params.ofm.q_params.zero_point),
+ ifm_channels=out_channels,
+ ifm2_channels=out_channels,
+ reversed_operands=False,
+ ofm_dtype="int8",
+ rounding_mode="NATURAL",
+ )
+ elif (
+ params.ifm.q_params.scale_f32 == params.ofm.q_params.scale_f32
+ and params.ifm.q_params.zero_point ==
params.ofm.q_params.zero_point
+ ):
Review comment:
Yeah sounds like it is needed for the bit accuracy...
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
