wehub-resource-sync
197 lines
7.0 KiB
C++
197 lines
7.0 KiB
C++
// Copyright (c) 2023 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.
|
|
#include <gtest/gtest.h>
|
|
#include <iostream>
|
|
|
|
#include "paddle/fluid/framework/new_executor/interpretercore.h"
|
|
#include "paddle/fluid/framework/scope.h"
|
|
#include "paddle/fluid/pir/dialect/kernel/ir/kernel_dialect.h"
|
|
#include "paddle/fluid/pir/dialect/operator/ir/control_flow_op.h"
|
|
#include "paddle/fluid/pir/dialect/operator/ir/op_dialect.h"
|
|
#include "paddle/fluid/pir/dialect/operator/ir/pd_op.h"
|
|
#include "paddle/fluid/pir/transforms/pd_op_to_kernel_pass.h"
|
|
#include "paddle/pir/include/core/builder.h"
|
|
#include "paddle/pir/include/core/builtin_op.h"
|
|
#include "paddle/pir/include/core/program.h"
|
|
#include "paddle/pir/include/dialect/control_flow/ir/cf_dialect.h"
|
|
#include "paddle/pir/include/dialect/control_flow/ir/cf_op.h"
|
|
|
|
PD_DECLARE_KERNEL(full, CPU, ALL_LAYOUT);
|
|
PD_DECLARE_KERNEL(less_than, CPU, ALL_LAYOUT);
|
|
PD_DECLARE_KERNEL(add, CPU, ALL_LAYOUT);
|
|
PD_DECLARE_KERNEL(add_grad, CPU, ALL_LAYOUT);
|
|
PD_DECLARE_KERNEL(add_n, CPU, ALL_LAYOUT);
|
|
|
|
using namespace paddle::dialect; // NOLINT
|
|
|
|
// example for while_op use
|
|
// while(i < ten) { i = i + 1;}
|
|
TEST(while_op_test, base) {
|
|
pir::IrContext* ctx = pir::IrContext::Instance();
|
|
ctx->GetOrRegisterDialect<pir::ControlFlowDialect>();
|
|
ctx->GetOrRegisterDialect<OperatorDialect>();
|
|
|
|
pir::Program program(ctx);
|
|
pir::Block* block = program.block();
|
|
pir::Builder builder(ctx, block);
|
|
|
|
auto i =
|
|
builder.Build<FullOp>(std::vector<int64_t>{1}, 1, phi::DataType::INT32)
|
|
.out();
|
|
auto ten =
|
|
builder.Build<FullOp>(std::vector<int64_t>{1}, 10, phi::DataType::INT32)
|
|
.out();
|
|
|
|
// compute condition value: i < ten
|
|
auto cond_value = builder.Build<LessThanOp>(i, ten).out();
|
|
|
|
auto while_op =
|
|
builder.Build<WhileOp>(cond_value, std::vector<pir::Value>{i, ten});
|
|
|
|
// { i = i + 1}
|
|
pir::Block& body_block = while_op.body();
|
|
auto body_i_argument = body_block.arg(0);
|
|
auto body_ten_argument = body_block.arg(1);
|
|
builder.SetInsertionPointToStart(&body_block);
|
|
auto one =
|
|
builder.Build<FullOp>(std::vector<int64_t>{1}, 1, phi::DataType::INT32)
|
|
.out();
|
|
auto new_i = builder.Build<AddOp>(body_i_argument, one).out();
|
|
|
|
// compute new condition value: new_i < new_ten
|
|
auto new_cond_value =
|
|
builder.Build<LessThanOp>(new_i, body_ten_argument).out();
|
|
|
|
builder.Build<pir::YieldOp>(
|
|
std::vector<pir::Value>{new_cond_value, new_i, body_ten_argument});
|
|
|
|
builder.SetInsertionPointAfter(while_op);
|
|
LOG(INFO) << program;
|
|
|
|
EXPECT_EQ(while_op.cond(), cond_value);
|
|
}
|
|
|
|
TEST(while_op_test, network_with_backward) {
|
|
pir::IrContext* ctx = pir::IrContext::Instance();
|
|
ctx->GetOrRegisterDialect<OperatorDialect>();
|
|
ctx->GetOrRegisterDialect<pir::ControlFlowDialect>();
|
|
ctx->GetOrRegisterDialect<paddle::dialect::KernelDialect>();
|
|
|
|
pir::Program program(ctx);
|
|
pir::Block* block = program.block();
|
|
pir::Builder builder(ctx, block);
|
|
auto i =
|
|
builder.Build<FullOp>(std::vector<int64_t>{1}, 0, phi::DataType::INT32)
|
|
.out();
|
|
auto ten =
|
|
builder.Build<FullOp>(std::vector<int64_t>{10}, 10, phi::DataType::INT32)
|
|
.out();
|
|
auto x = builder.Build<FullOp>(std::vector<int64_t>{2, 2}, 1.0f).out();
|
|
auto y = builder.Build<FullOp>(std::vector<int64_t>{2, 2}, 2.0f).out();
|
|
auto one =
|
|
builder.Build<FullOp>(std::vector<int64_t>{1}, 1, phi::DataType::INT32)
|
|
.out();
|
|
// def cond(i, x):
|
|
// return i < 10
|
|
// def body(i, x):
|
|
// return i + 1, x + y
|
|
// }
|
|
auto cond_value = builder.Build<LessThanOp>(i, ten).out();
|
|
|
|
auto [stack, inlet, outlet] = builder.Build<pir::StackCreateOp>().out();
|
|
(void)(stack);
|
|
auto while_op =
|
|
builder.Build<WhileOp>(cond_value, std::vector<pir::Value>{i, x});
|
|
|
|
// { return i + 1, x + y}
|
|
auto& body_block = while_op.body();
|
|
builder.SetInsertionPointToStart(&body_block);
|
|
|
|
auto body_i_argument = body_block.arg(0);
|
|
auto body_x_argument = body_block.arg(1);
|
|
|
|
auto new_i = builder.Build<AddOp>(body_i_argument, one).out();
|
|
auto new_x = builder.Build<AddOp>(body_x_argument, y).out();
|
|
|
|
// compute new condition value: new_i < new_ten
|
|
auto new_cond_value = builder.Build<LessThanOp>(new_i, ten).out();
|
|
|
|
builder.Build<pir::TuplePushOp>(
|
|
inlet, std::initializer_list<pir::Value>{body_x_argument});
|
|
|
|
builder.Build<pir::YieldOp>(
|
|
std::vector<pir::Value>{new_cond_value, new_i, new_x});
|
|
|
|
builder.SetInsertionPointAfter(while_op);
|
|
|
|
auto i_out = while_op->result(0);
|
|
auto x_out = while_op->result(1);
|
|
EXPECT_EQ(i_out.type(), i.type());
|
|
EXPECT_EQ(x_out.type(), x.type());
|
|
|
|
// build backward network
|
|
auto x_out_grad =
|
|
builder.Build<FullOp>(std::vector<int64_t>{2, 2}, 1.0f).out();
|
|
auto zero = builder.Build<FullOp>(std::vector<int64_t>{2, 2}, 0.0).out();
|
|
|
|
// the input of while_grad op is {x_out_grad, zero}
|
|
// the output of while_grad op is {x_grad, y_grad}
|
|
// the value {i , one, ten} is stop gradient.
|
|
|
|
auto bwd_cond = builder.Build<HasElementsOp>(stack).out();
|
|
auto while_grad = builder.Build<WhileOp>(
|
|
bwd_cond, std::vector<pir::Value>{x_out_grad, zero});
|
|
pir::Block& bwd_body_block = while_grad.body();
|
|
builder.SetInsertionPointToStart(&bwd_body_block);
|
|
auto local_x_out_grad_arg = bwd_body_block.arg(0);
|
|
auto local_y_grad_arg = bwd_body_block.arg(1);
|
|
|
|
auto pop_op = builder.Build<pir::TuplePopOp>(outlet);
|
|
auto bwd_body_x_argument = pop_op.outlet_element(0);
|
|
|
|
auto add_grad_op =
|
|
builder.Build<AddGradOp>(bwd_body_x_argument, y, local_x_out_grad_arg);
|
|
auto bwd_body_x_argument_grad = add_grad_op.x_grad();
|
|
auto local_y_grad = add_grad_op.y_grad();
|
|
|
|
// accumulate gradient
|
|
auto combine_y = builder
|
|
.Build<pir::CombineOp>(std::vector<pir::Value>{
|
|
local_y_grad, local_y_grad_arg})
|
|
.out();
|
|
auto local_next_y_grad = builder.Build<AddNOp>(combine_y).out();
|
|
|
|
auto next_bwd_cond = builder.Build<HasElementsOp>(stack).out();
|
|
builder.Build<pir::YieldOp>(std::vector<pir::Value>{
|
|
next_bwd_cond, bwd_body_x_argument_grad, local_next_y_grad});
|
|
|
|
auto x_grad = while_grad.result(0);
|
|
auto y_grad = while_grad.result(1);
|
|
|
|
EXPECT_EQ(x_grad.type(), x.type());
|
|
EXPECT_EQ(y_grad.type(), y.type());
|
|
|
|
LOG(INFO) << program;
|
|
|
|
auto place = phi::CPUPlace();
|
|
#ifdef PADDLE_WITH_CUDA
|
|
place = phi::GPUPlace(0);
|
|
#endif
|
|
auto kernel_program = pir::PdOpLowerToKernelPass(&program, place);
|
|
paddle::framework::Scope scope;
|
|
paddle::framework::InterpreterCore test_core(
|
|
place, {}, kernel_program->block(), &scope);
|
|
test_core.Run({});
|
|
}
|