def callback(self, pre: tvm.relay.Expr, post: tvm.relay.Expr, node_map: tvm.ir.container.Map) -> tvm.relay.Expr: params = self.params_class(post.op.body) params.ifm.tensor = post.args[ 1] if params.reversed_operands else post.args[0] params.ifm2.tensor = post.args[ 0] if params.reversed_operands else post.args[1] channels_map = { "NHWC": 3, } if str(params.ofm.layout) not in channels_map.keys(): raise UnsupportedLayout(str(params.ofm.layout)) activation_map = {"clip": "CLIP"} if params.activation: activation = activation_map[params.activation.op.name] clip_min = int(params.activation.attrs.a_min) clip_max = int(params.activation.attrs.a_max) else: activation = "NONE" clip_min = 0 clip_max = 0 # We don't yet support activation functions that need to get legalized to LUTs. lut = relay.const([], dtype="int8") inputs = [params.ifm, params.ifm2] inputs = self.reshape_input(inputs) ethosu_binary_elementwise = ethosu_ops.ethosu_binary_elementwise( ifm=inputs[0], ifm2=inputs[1], lut=lut, operator_type=params.operator_type, ifm_scale=float(params.ifm.q_params.scale_f32), ifm_zero_point=int(params.ifm.q_params.zero_point), ifm2_scale=float(params.ifm2.q_params.scale_f32), ifm2_zero_point=int(params.ifm2.q_params.zero_point), ofm_scale=float(params.ofm.q_params.scale_f32), ofm_zero_point=int(params.ofm.q_params.zero_point), ifm_channels=params.ifm.shape[-1], ifm2_channels=params.ifm2.shape[-1], reversed_operands=params.reversed_operands, ofm_dtype=params.ofm.dtype, activation=activation, clip_min=clip_min, clip_max=clip_max, ifm_layout=str(params.ifm.layout), ifm2_layout=str(params.ifm2.layout), ofm_layout=str(params.ofm.layout), ) output = self.reshape_output(ethosu_binary_elementwise, params.ifm.shape) return output
def make_ethosu_binary_elementwise( ifm, ifm2, ifm_channels, ifm2_channels, operator_type, ofm_dtype, reversed_operands=False, activation="NONE", ifm_layout="NHWC", ifm2_layout="NHWC", ofm_layout="NHWC", rounding_mode="TFL", ): ethosu_binary_elementwise = ethosu_ops.ethosu_binary_elementwise( ifm=ifm, ifm2=ifm2, lut=relay.const([], dtype="int8"), operator_type=operator_type, ifm_scale=1, ifm_zero_point=0, ifm2_scale=1, ifm2_zero_point=0, ofm_scale=1, ofm_zero_point=0, ifm_channels=ifm_channels, ifm2_channels=ifm2_channels, reversed_operands=reversed_operands, activation=activation, ofm_dtype=ofm_dtype, clip_min=10 if activation == "CLIP" else 0, clip_max=100 if activation == "CLIP" else 0, rounding_mode=rounding_mode, ifm_layout=ifm_layout, ifm2_layout=ifm2_layout, ofm_layout=ofm_layout, ) return ethosu_binary_elementwise
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="int16"), dtype="int16") 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): reduced_op = ethosu_ops.ethosu_pooling( ifm=reduced_op, lut=lut, pooling_type="AVG", ifm_scale=float(params.ifm.q_params.scale_f32), ifm_zero_point=0, ofm_scale=float(params.ofm.q_params.scale_f32), ofm_zero_point=0, pool_shape=(filter_height, filter_width), ofm_channels=out_channels, rounding_mode="TRUNCATE", ) else: weight_scale = 1 / (filter_height * filter_width) weight_values = np.ones( [out_channels, filter_height, filter_width, in_channels]) bias = -1 * int( params.ifm.q_params.zero_point) * filter_height * filter_width scale_bias = vela_api.pack_biases( biases=np.ones([ifm_shape[-1]]) * bias, 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=0, 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, rounding_mode="NATURAL", ) # Reshape to original ofm shape if len(ofm_shape) < 4: reduced_op = relay.reshape(reduced_op, ofm_shape) return reduced_op