[PIR] Support translate IfOp

[zhangbo9674]

PR types

New features

PR changes

Others

Description

本 PR 用于支持将旧 Program 体系的 condition_block 算子 translate 为 Pir 的 IfOp；
示例模型为一个 if 嵌套 if 的模型结构，模型代码如下：

import paddle
paddle.enable_static()
def true_func():
    x = paddle.full(shape=[1], dtype='float32', fill_value=0.1)
    y = paddle.full(shape=[1], dtype='float32', fill_value=0.23)
    pred = paddle.less_than(x=x, y=y, name=None)
    out = paddle.static.nn.cond(pred, lambda: x + y, lambda: x * y)
    z = x * y 
    return x, z

def false_func():
    x = paddle.full(shape=[1], dtype='float32', fill_value=0.1)
    y = paddle.full(shape=[1], dtype='float32', fill_value=0.23)
    pred = paddle.less_than(x=x, y=y, name=None)
    out = paddle.static.nn.cond(pred, lambda: x + y, lambda: x * y)
    z = x * y 
    return y, z

train_program = paddle.static.Program()
startup_program = paddle.static.Program()
with paddle.static.program_guard(train_program, startup_program):
    x = paddle.full(shape=[1], dtype='float32', fill_value=0.1)
    y = paddle.full(shape=[1], dtype='float32', fill_value=0.23)
    pred = paddle.less_than(x=x, y=y, name=None)
    out1, out2 = paddle.static.nn.cond(pred, true_func, false_func)

旧 Program：

{ // block 0
    ...
    {Out=['fill_constant_1.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.1, value = 0.10000000149011612, with_quant_attr = False)
    {Out=['fill_constant_3.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.23, value = 0.23000000417232513, with_quant_attr = False)
    {Out=['less_than_0.tmp_0']} = less_than(inputs={X=['fill_constant_1.tmp_0'], Y=['fill_constant_3.tmp_0']}, axis = -1, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_6', '_generated_var_5'], Scope=['_generated_var_7']} = conditional_block(inputs={Cond=['less_than_0.tmp_0'], Input=[]}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[1], with_quant_attr = False)
    {Out=['logical_not_1.tmp_0']} = logical_not(inputs={X=['less_than_0.tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_13', '_generated_var_14'], Scope=['_generated_var_15']} = conditional_block(inputs={Cond=['logical_not_1.tmp_0'], Input=[]}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[4], with_quant_attr = False)
    {Out=['cast_2.tmp_0']} = cast(inputs={X=['less_than_0.tmp_0']}, in_dtype = 0, op_device = , op_namescope = /, op_role = 0, op_role_var = [], out_dtype = 2, use_mkldnn = False, with_quant_attr = False)
    {Out=['_generated_var_16']} = select_input(inputs={Mask=['cast_2.tmp_0'], X=['_generated_var_13', '_generated_var_5']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_17']} = select_input(inputs={Mask=['cast_2.tmp_0'], X=['_generated_var_14', '_generated_var_6']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 1
    ...
    {Out=['fill_constant_5.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.1, value = 0.10000000149011612, with_quant_attr = False)
    {Out=['fill_constant_7.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.23, value = 0.23000000417232513, with_quant_attr = False)
    {Out=['less_than_1.tmp_0']} = less_than(inputs={X=['fill_constant_5.tmp_0'], Y=['fill_constant_7.tmp_0']}, axis = -1, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_0'], Scope=['_generated_var_1']} = conditional_block(inputs={Cond=['less_than_1.tmp_0'], Input=['fill_constant_5.tmp_0', 'fill_constant_7.tmp_0']}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[2], with_quant_attr = False)
    {Out=['logical_not_0.tmp_0']} = logical_not(inputs={X=['less_than_1.tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_2'], Scope=['_generated_var_3']} = conditional_block(inputs={Cond=['logical_not_0.tmp_0'], Input=['fill_constant_5.tmp_0', 'fill_constant_7.tmp_0']}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[3], with_quant_attr = False)
    {Out=['cast_0.tmp_0']} = cast(inputs={X=['less_than_1.tmp_0']}, in_dtype = 0, op_device = , op_namescope = /, op_role = 0, op_role_var = [], out_dtype = 2, use_mkldnn = False, with_quant_attr = False)
    {Out=['_generated_var_4']} = select_input(inputs={Mask=['cast_0.tmp_0'], X=['_generated_var_2', '_generated_var_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['tmp_2']} = elementwise_mul(inputs={X=['fill_constant_5.tmp_0'], Y=['fill_constant_7.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_5']} = assign(inputs={X=['fill_constant_5.tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_6']} = assign(inputs={X=['tmp_2']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 2
   ...
    {Out=['tmp_0']} = elementwise_add(inputs={X=['fill_constant_5.tmp_0'], Y=['fill_constant_7.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_0']} = assign(inputs={X=['tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 3
    ...
    {Out=['tmp_1']} = elementwise_mul(inputs={X=['fill_constant_5.tmp_0'], Y=['fill_constant_7.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_2']} = assign(inputs={X=['tmp_1']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 4
    ...
    {Out=['fill_constant_9.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.1, value = 0.10000000149011612, with_quant_attr = False)
    {Out=['fill_constant_11.tmp_0']} = fill_constant(inputs={ShapeTensor=[], ShapeTensorList=[], ValueTensor=[]}, dtype = 5, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], place_type = -1, shape = [1], str_value = 0.23, value = 0.23000000417232513, with_quant_attr = False)
    {Out=['less_than_2.tmp_0']} = less_than(inputs={X=['fill_constant_9.tmp_0'], Y=['fill_constant_11.tmp_0']}, axis = -1, force_cpu = False, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_8'], Scope=['_generated_var_9']} = conditional_block(inputs={Cond=['less_than_2.tmp_0'], Input=['fill_constant_11.tmp_0', 'fill_constant_9.tmp_0']}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[5], with_quant_attr = False)
    {Out=['logical_not_2.tmp_0']} = logical_not(inputs={X=['less_than_2.tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_10'], Scope=['_generated_var_11']} = conditional_block(inputs={Cond=['logical_not_2.tmp_0'], Input=['fill_constant_11.tmp_0', 'fill_constant_9.tmp_0']}, is_scalar_condition = True, op_device = , op_namescope = /, op_role = 0, op_role_var = [], sub_block = block[6], with_quant_attr = False)
    {Out=['cast_1.tmp_0']} = cast(inputs={X=['less_than_2.tmp_0']}, in_dtype = 0, op_device = , op_namescope = /, op_role = 0, op_role_var = [], out_dtype = 2, use_mkldnn = False, with_quant_attr = False)
    {Out=['_generated_var_12']} = select_input(inputs={Mask=['cast_1.tmp_0'], X=['_generated_var_10', '_generated_var_8']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['tmp_5']} = elementwise_mul(inputs={X=['fill_constant_9.tmp_0'], Y=['fill_constant_11.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_13']} = assign(inputs={X=['fill_constant_11.tmp_0']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_14']} = assign(inputs={X=['tmp_5']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 5
    ...
    {Out=['tmp_3']} = elementwise_add(inputs={X=['fill_constant_9.tmp_0'], Y=['fill_constant_11.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_8']} = assign(inputs={X=['tmp_3']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}
{ // block 6
    ...
    {Out=['tmp_4']} = elementwise_mul(inputs={X=['fill_constant_9.tmp_0'], Y=['fill_constant_11.tmp_0']}, axis = -1, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
    {Out=['_generated_var_10']} = assign(inputs={X=['tmp_4']}, op_device = , op_namescope = /, op_role = 0, op_role_var = [], with_quant_attr = False)
}

翻译之后的 Program

{
    (%0) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.1} : () -> pd_op.tensor<1xf32>
    (%1) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.23} : () -> pd_op.tensor<1xf32>
    (%2) = "pd_op.less_than" (%0, %1) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xb>
    (%3, %4) = pd_op.if (%2) -> pd_op.tensor<1xf32>, pd_op.tensor<1xf32>{
        (%5) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.1} : () -> pd_op.tensor<1xf32>
        (%6) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.23} : () -> pd_op.tensor<1xf32>
        (%7) = "pd_op.less_than" (%5, %6) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xb>
        (%8) = pd_op.if (%7) -> pd_op.tensor<1xf32>{
            (%9) = "pd_op.add" (%5, %6) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
            () = "cf.yield" (%9) {} : (pd_op.tensor<1xf32>) -> 
        } else {
            (%10) = "pd_op.multiply" (%5, %6) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
            () = "cf.yield" (%10) {} : (pd_op.tensor<1xf32>) -> 
        }
        (%11) = "pd_op.multiply" (%5, %6) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
        () = "cf.yield" (%5, %11) {} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> 
    } else {
        (%12) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.1} : () -> pd_op.tensor<1xf32>
        (%13) = "pd_op.full" () {dtype:(pd_op.DataType)float32,is_persisable:(Array)[false],place:(pd_op.Place)Place(undefined:0),shape:(pd_op.IntArray)[1],stop_gradient:(Array)[true],value:(Float)0.23} : () -> pd_op.tensor<1xf32>
        (%14) = "pd_op.less_than" (%12, %13) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xb>
        (%15) = pd_op.if (%14) -> pd_op.tensor<1xf32>{
            (%16) = "pd_op.add" (%12, %13) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
            () = "cf.yield" (%16) {} : (pd_op.tensor<1xf32>) -> 
        } else {
            (%17) = "pd_op.multiply" (%12, %13) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
            () = "cf.yield" (%17) {} : (pd_op.tensor<1xf32>) -> 
        }
        (%18) = "pd_op.multiply" (%12, %13) {is_persisable:(Array)[false],stop_gradient:(Array)[true]} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> pd_op.tensor<1xf32>
        () = "cf.yield" (%13, %18) {} : (pd_op.tensor<1xf32>, pd_op.tensor<1xf32>) -> 
    }
}

Other

Pcard-67164

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[winter-wang]

LGTM

[winter-wang]

LGTM