paddlepaddle--paddle/python/paddle/tensorrt/impls/math.py

# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy as np
import tensorrt as trt

from paddle.tensorrt.converter_utils import (
    add_1D_constant_layer,
    add_cast_reduce_layer,
    add_constant_layer,
    add_elementwise_layer,
    add_reduce_layer,
    broadcast,
    cast_tensor,
    fill_constant_layer,
    get_axes_for_reduce_op,
    get_axis_length,
    get_input_constant_value,
    get_shape_tensor_element,
    set_layer_name,
    trt_cast,
    trt_concat,
    trt_equal,
    trt_expand,
    trt_max,
    trt_reshape,
    trt_shape,
)
from paddle.tensorrt.register import converter_registry


@converter_registry.register("pd_op.add")
@converter_registry.register("pd_op.add_")
def add_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.SUM
    )


@converter_registry.register("pd_op.scale")
def scale_converter(network, paddle_op, inputs):
    x = inputs[0]
    bias = paddle_op.attrs().get("bias", 0.0)
    bias_after_scale = paddle_op.attrs().get("bias_after_scale", True)

    is_int = x.dtype == trt.DataType.INT32
    if is_int:
        bias_tensor = add_1D_constant_layer(
            network,
            int(bias + 0.5) if bias > 0 else int(bias - 0.5),
            name=[paddle_op.name(), "bias_tensor"],
        )
    else:
        bias_tensor = add_1D_constant_layer(
            network,
            bias,
            dtype=np.float32,
            name=[paddle_op.name(), "bias_tensor"],
        )
    is_bias_0 = bias == 0
    bias_shapes = [1] * len(x.shape)
    bias_shapes_tensor = add_1D_constant_layer(
        network, bias_shapes, name=[paddle_op.name(), "bias_shapes_tensor"]
    )
    reshape_layer_bias = network.add_shuffle(bias_tensor)
    reshape_layer_bias.set_input(1, bias_shapes_tensor)
    set_layer_name(reshape_layer_bias, paddle_op)

    scale = get_input_constant_value(paddle_op, inputs, 1)
    if scale is not None:
        scale = scale[0]
        has_scale_tensor = False
        if is_int:
            scale_tensor = add_1D_constant_layer(
                network,
                int(scale + 0.5 if scale > 0 else scale - 0.5),
                name=[paddle_op.name(), "scale_tensor"],
            )
        else:
            scale_tensor = add_1D_constant_layer(
                network,
                scale,
                dtype=np.float32,
                name=[paddle_op.name(), "scale_tensor"],
            )
        is_scale_1 = scale == 1
    else:
        has_scale_tensor = True
        scale_tensor = inputs[1]
        is_scale_1 = False
    scale_shapes = [1] * len(x.shape)
    scale_shapes_tensor = add_1D_constant_layer(
        network, scale_shapes, name=[paddle_op.name(), "scale_shapes_tensor"]
    )
    reshape_layer_scale = network.add_shuffle(scale_tensor)
    reshape_layer_scale.set_input(1, scale_shapes_tensor)
    set_layer_name(reshape_layer_scale, paddle_op)

    # Initialize the layer variable to ensure it's defined in all branches
    layer = None

    if not has_scale_tensor and is_scale_1 and is_bias_0:
        layer = network.add_identity(x)
        set_layer_name(layer, paddle_op)
    else:
        if bias_after_scale:
            if not is_scale_1:
                layer = network.add_elementwise(
                    x,
                    reshape_layer_scale.get_output(0),
                    trt.ElementWiseOperation.PROD,
                )
                set_layer_name(layer, paddle_op)
                x = layer.get_output(0)

            if not is_bias_0:
                layer = network.add_elementwise(
                    x,
                    reshape_layer_bias.get_output(0),
                    trt.ElementWiseOperation.SUM,
                )
                set_layer_name(layer, paddle_op)

        else:
            if not is_bias_0:
                layer = network.add_elementwise(
                    x,
                    reshape_layer_bias.get_output(0),
                    trt.ElementWiseOperation.SUM,
                )
                set_layer_name(layer, paddle_op)
                x = layer.get_output(0)
            if not is_scale_1:
                layer = network.add_elementwise(
                    x,
                    reshape_layer_scale.get_output(0),
                    trt.ElementWiseOperation.PROD,
                )
                set_layer_name(layer, paddle_op)

    return layer.get_output(0)


@converter_registry.register("pd_op.max")
def max_converter(network, paddle_op, inputs):
    input_tensor = inputs[0]
    axis = get_input_constant_value(paddle_op, inputs, 1)
    input_shape = input_tensor.shape
    keepdim = paddle_op.attrs()["keepdim"]
    if network.has_implicit_batch_dimension:
        assert axis != 0, (
            "can't reduce on axis == 0 when network has implicit batch dimension"
        )
    if len(axis) == 0:
        axis = list(range(len(input_shape)))
    for i in range(len(axis)):
        if axis[i] < 0:
            axis[i] = len(input_shape) + axis[i]
    layer = network.add_reduce(
        input_tensor,
        trt.ReduceOperation.MAX,
        axes=get_axes_for_reduce_op(axis),
        keep_dims=keepdim,
    )
    set_layer_name(layer, paddle_op)
    return layer.get_output(0)


@converter_registry.register("pd_op.divide")
def divide_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.DIV
    )


@converter_registry.register("pd_op.subtract")
def subtract_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.SUB
    )


@converter_registry.register("pd_op.multiply")
def multiply_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.PROD
    )


@converter_registry.register("pd_op.clip")
def clip_converter(network, paddle_op, inputs):
    def _get_constant_or_expand_tensor(
        value, constant_inputs, input_shape_tensor, rank, name=None
    ):
        if value is not None:
            return fill_constant_layer(
                network,
                input_shape_tensor,
                rank,
                value,
                input_tensor.dtype,
                name=name,
            )
        else:
            expanded_tensor = trt_expand(
                network, constant_inputs, 1, input_shape_tensor, rank, name=name
            )
            if expanded_tensor.dtype != input_tensor.dtype:
                expanded_tensor = cast_tensor(
                    network, expanded_tensor, input_tensor.dtype, name=name
                )
            return expanded_tensor

    input_tensor = inputs[0]
    input_shape = input_tensor.shape
    rank = len(input_shape)
    input_shape_tensor = network.add_shape(input_tensor)
    set_layer_name(input_shape_tensor, paddle_op)
    input_shape_tensor = input_shape_tensor.get_output(0)

    # handle min operation
    min_value = get_input_constant_value(paddle_op, inputs, 1)
    alpha_t = _get_constant_or_expand_tensor(
        min_value, inputs[1], input_shape_tensor, rank
    )

    # handle max operation
    max_value = get_input_constant_value(paddle_op, inputs, 2)
    beta_t = _get_constant_or_expand_tensor(
        max_value,
        inputs[2],
        input_shape_tensor,
        rank,
        name=[paddle_op.name(), 'beta_t'],
    )

    # run the clip operation
    lower_clip = trt_max(
        network, input_tensor, alpha_t, name=[paddle_op.name(), 'lower_clip']
    )
    layer = network.add_elementwise(
        lower_clip, beta_t, trt.ElementWiseOperation.MIN
    )
    set_layer_name(layer, paddle_op)
    return layer.get_output(0)


@converter_registry.register("pd_op.pow")
def pow_converter(network, paddle_op, inputs):
    from paddle.tensorrt.util import support_fp32_mix_precision

    x = inputs[0]
    factor = paddle_op.attrs()["y"]
    dims_x = x.shape
    trt_dims_y = trt.Dims([1] * len(dims_x))
    w_data = [factor]
    y = add_constant_layer(
        network, w_data, trt_dims_y, np.float32, name=[paddle_op.name(), 'y']
    )
    layer = network.add_elementwise(x, y, trt.ElementWiseOperation.POW)
    set_layer_name(layer, paddle_op)
    support_fp32_mix_precision(paddle_op.name(), layer)
    return layer.get_output(0)


@converter_registry.register("pd_op.remainder")
@converter_registry.register("pd_op.remainder_")
def remainder_converter(network, paddle_op, inputs):
    from paddle.tensorrt.util import support_fp32_mix_precision

    weight_shape = paddle_op.operands()[1].source().shape
    input_shape = inputs[0].shape

    weight_tensor = inputs[1]
    input_tensor = inputs[0]
    if type(inputs[1]) == trt.Weights:
        weight_tensor = network.add_constant(weight_shape, inputs[1])
        set_layer_name(weight_tensor, paddle_op)
        weight_tensor = weight_tensor.get_output(0)
    if type(inputs[0]) == trt.Weights:
        input_tensor = network.add_constant(input_shape, inputs[0])
        set_layer_name(input_tensor, paddle_op)
        input_tensor = input_tensor.get_output(0)

    lhs_val, rhs_val = broadcast(
        network,
        input_tensor,
        weight_tensor,
        "input_tensor_broadcast",
        "weight_tensor_broadcast",
        paddle_op,
    )
    is_floor_div = input_tensor.dtype != trt.DataType.INT32
    if is_floor_div:
        quotient_layer = network.add_elementwise(
            lhs_val, rhs_val, trt.ElementWiseOperation.FLOOR_DIV
        )
    else:
        quotient_layer = network.add_elementwise(
            lhs_val, rhs_val, trt.ElementWiseOperation.DIV
        )
    set_layer_name(quotient_layer, paddle_op)
    quotient = quotient_layer.get_output(0)
    support_fp32_mix_precision(paddle_op.name(), quotient_layer)

    # Multiply rhs by the quotient
    product_layer = network.add_elementwise(
        rhs_val, quotient, trt.ElementWiseOperation.PROD
    )
    set_layer_name(product_layer, paddle_op)
    product = product_layer.get_output(0)
    support_fp32_mix_precision(paddle_op.name(), product_layer)
    remainder_layer = network.add_elementwise(
        lhs_val, product, trt.ElementWiseOperation.SUB
    )
    set_layer_name(remainder_layer, paddle_op)
    remainder = remainder_layer.get_output(0)
    support_fp32_mix_precision(paddle_op.name(), remainder_layer)

    return remainder


@converter_registry.register("pd_op.min")
def min_converter(network, paddle_op, inputs):
    return add_reduce_layer(network, paddle_op, inputs, trt.ReduceOperation.MIN)


@converter_registry.register("pd_op.sum")
def sum_converter(network, paddle_op, inputs):
    return add_reduce_layer(network, paddle_op, inputs, trt.ReduceOperation.SUM)


@converter_registry.register("pd_op.mean")
def mean_converter(network, paddle_op, inputs):
    return add_reduce_layer(network, paddle_op, inputs, trt.ReduceOperation.AVG)


@converter_registry.register("pd_op.any")
def any_converter(network, paddle_op, inputs):
    return add_cast_reduce_layer(
        network, paddle_op, inputs, trt.ReduceOperation.MAX
    )


@converter_registry.register("pd_op.all")
def all_converter(network, paddle_op, inputs):
    return add_cast_reduce_layer(
        network, paddle_op, inputs, trt.ReduceOperation.MIN
    )


@converter_registry.register("pd_op.cumsum")
def cumsum_converter(network, paddle_op, inputs):
    input_tensor = inputs[0]
    dtype = input_tensor.dtype
    axis = get_input_constant_value(paddle_op, inputs, 1)[0]
    input_shape = input_tensor.shape
    rank = len(input_shape)

    if axis < 0:
        axis += rank
    axis = int(axis)

    # Obtain the number of cycles
    if input_shape[axis] > 0:
        trip_limit = add_1D_constant_layer(
            network,
            input_shape[axis],
            is_scalar=True,
            name=[paddle_op.name(), 'trip_limit'],
        )
    else:
        dynamic_shape = trt_shape(
            network, input_tensor, name=[paddle_op.name(), 'dynamic_shape']
        )
        trip_limit = get_shape_tensor_element(
            network,
            dynamic_shape,
            axis,
            True,
            name=[paddle_op.name(), 'trip_limit'],
        )

    # Obtain the slice shape
    shape_list = []
    for i in range(rank):
        if i == axis:
            shape_list.append(
                add_1D_constant_layer(
                    network, [1], name=[paddle_op.name(), f'shape_list_{i}']
                )
            )
        else:
            shape_list.append(
                get_axis_length(
                    network,
                    input_tensor,
                    i,
                    name=[paddle_op.name(), f'shape_list_{i}'],
                )
            )
    slice_shape = trt_concat(
        network, shape_list, name=[paddle_op.name(), 'slice_shape']
    )

    start = [0] * rank
    size = [1] * rank
    stride = [1] * rank
    input_sliced = network.add_slice(input_tensor, start, size, stride)
    input_sliced.set_input(2, slice_shape)
    set_layer_name(input_sliced, paddle_op)

    # squeeze axis
    if rank > 1:
        shape_list.pop(axis)
    new_shape = trt_concat(
        network, shape_list, name=[paddle_op.name(), 'new_shape']
    )
    squeeze_output = trt_reshape(
        network,
        input_sliced.get_output(0),
        new_shape,
        is_shape_tensor=True,
        name=[paddle_op.name(), 'squeeze_output'],
    )

    loop = network.add_loop()
    loop.add_trip_limit(trip_limit, trt.TripLimit.COUNT)

    iterator = loop.add_iterator(input_tensor, axis)
    set_layer_name(iterator, paddle_op)
    data = iterator.get_output(0)

    # create zero tensor
    zero_vec = np.array([0.0], dtype=np.float32)
    zero = add_1D_constant_layer(
        network, zero_vec, name=[paddle_op.name(), 'zero']
    )
    lhs_val, rhs_val = broadcast(
        network,
        squeeze_output,
        zero,
        "squeeze_output_broadcast",
        "zero_output_broadcast",
        paddle_op,
    )
    cast_tensor = trt_cast(
        network, rhs_val, dtype, name=[paddle_op.name(), 'cast_tensor']
    )
    zero_tensor = network.add_elementwise(
        lhs_val, cast_tensor, trt.ElementWiseOperation.PROD
    )
    set_layer_name(zero_tensor, paddle_op)
    zero_tensor = zero_tensor.get_output(0)

    # Set as scalar
    if rank == 1:
        zero_tensor = trt_reshape(
            network, zero_tensor, (), name=[paddle_op.name(), 'zero_tensor']
        )

    # Cycle and add according to the axis
    running_sum = loop.add_recurrence(zero_tensor)
    running_sum_tensor = running_sum.get_output(0)

    cur_sum = network.add_elementwise(
        data, running_sum_tensor, trt.ElementWiseOperation.SUM
    )
    set_layer_name(cur_sum, paddle_op)
    cur_sum = cur_sum.get_output(0)

    running_sum.set_input(1, cur_sum)
    set_layer_name(running_sum, paddle_op)

    reverse_flag = trt.LoopOutput.CONCATENATE
    loop_out = loop.add_loop_output(cur_sum, reverse_flag, axis)
    loop_out.set_input(1, trip_limit)
    set_layer_name(loop_out, paddle_op)

    return loop_out.get_output(0)


@converter_registry.register("pd_op.floor_divide")
def floor_divide_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.FLOOR_DIV
    )


@converter_registry.register("pd_op.log")
def log_converter(network, paddle_op, inputs):
    input_tensor = trt_cast(
        network, inputs[0], trt.float32, name=[paddle_op.name(), 'input_tensor']
    )
    layer = network.add_unary(input_tensor, trt.UnaryOperation.LOG)
    set_layer_name(layer, paddle_op)
    return layer.get_output(0)


@converter_registry.register("pd_op.elementwise_pow")
def elementwise_pow_converter(network, paddle_op, inputs):
    return add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.POW
    )


@converter_registry.register("pd_op.isnan")
def isnan_converter(network, paddle_op, inputs):
    input_tensor = inputs[0]
    equal_tensor = trt_equal(
        network,
        input_tensor,
        input_tensor,
        name=[paddle_op.name(), 'equal_tensor'],
    )
    layer = network.add_unary(equal_tensor, trt.UnaryOperation.NOT)
    set_layer_name(layer, paddle_op)
    return layer.get_output(0)


@converter_registry.register("pd_op.minimum")
def minimum_converter(network, paddle_op, inputs):
    min_layer = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.MIN
    )
    return min_layer


@converter_registry.register("pd_op.maximum")
def maximum_converter(network, paddle_op, inputs):
    max_layer = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.MAX
    )
    return max_layer


@converter_registry.register("pd_op.greater_equal")
@converter_registry.register("pd_op.greater_equal_")
def greater_equal_converter(network, paddle_op, inputs):
    greater_layer_output = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.GREATER
    )
    equal_layer_output = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.EQUAL
    )
    or_layer = add_elementwise_layer(
        network,
        paddle_op,
        [greater_layer_output, equal_layer_output],
        trt.ElementWiseOperation.OR,
    )
    return or_layer


@converter_registry.register("pd_op.less_equal")
@converter_registry.register("pd_op.less_equal_")
def less_equal_converter(network, paddle_op, inputs):
    less_layer_output = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.LESS
    )
    equal_layer_output = add_elementwise_layer(
        network, paddle_op, inputs, trt.ElementWiseOperation.EQUAL
    )
    or_layer = add_elementwise_layer(
        network,
        paddle_op,
        [less_layer_output, equal_layer_output],
        trt.ElementWiseOperation.OR,
    )
    return or_layer