wehub-resource-sync
169 lines
7.1 KiB
C++
169 lines
7.1 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 <cstdint>
|
|
#include <vector>
|
|
|
|
#include "paddle/fluid/framework/scope.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/general/constant_folding_pass.h"
|
|
#include "paddle/fluid/pir/transforms/general/dead_code_elimination_pass.h"
|
|
#include "paddle/fluid/pir/transforms/gpu/multihead_matmul_fuse_pass.h"
|
|
|
|
#include "paddle/phi/common/place.h"
|
|
#include "paddle/pir/include/core/builtin_dialect.h"
|
|
#include "paddle/pir/include/pass/pass_manager.h"
|
|
|
|
#include "paddle/phi/core/kernel_registry.h"
|
|
|
|
PD_DECLARE_KERNEL(multihead_matmul, GPU, ALL_LAYOUT);
|
|
|
|
void BuildProgram(pir::Builder &builder) { // NOLINT
|
|
paddle::dialect::FullOp matmul_1_in_1 =
|
|
builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{1, 300, 256},
|
|
0.9,
|
|
phi::DataType::FLOAT32,
|
|
phi::CPUPlace());
|
|
// The first path to matmul with scale (q).
|
|
paddle::dialect::FullOp matmul_1_in_2 =
|
|
builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{256, 256},
|
|
1.1,
|
|
phi::DataType::FLOAT32,
|
|
phi::CPUPlace());
|
|
|
|
paddle::dialect::MatmulOp matmul_1 = builder.Build<paddle::dialect::MatmulOp>(
|
|
matmul_1_in_1.out(), matmul_1_in_2.out(), false, false);
|
|
|
|
paddle::dialect::FullOp add_1_in_2 = builder.Build<paddle::dialect::FullOp>(
|
|
std::vector<int64_t>{256}, 1.5, phi::DataType::FLOAT32, phi::CPUPlace());
|
|
|
|
paddle::dialect::AddOp add_1 =
|
|
builder.Build<paddle::dialect::AddOp>(matmul_1.out(), add_1_in_2.out());
|
|
|
|
paddle::dialect::ReshapeOp reshape_1 =
|
|
builder.Build<paddle::dialect::ReshapeOp>(
|
|
add_1.out(), std::vector<int64_t>{0, 0, 8, 32});
|
|
|
|
paddle::dialect::TransposeOp transpose_1 =
|
|
builder.Build<paddle::dialect::TransposeOp>(reshape_1.out(),
|
|
std::vector<int>{0, 2, 1, 3});
|
|
|
|
paddle::dialect::ScaleOp scale_op = builder.Build<paddle::dialect::ScaleOp>(
|
|
transpose_1.out(), 0.1767766922712326, 0.0, true);
|
|
|
|
// The second path to matmul (k).
|
|
paddle::dialect::FullOp matmul_2_in_2 =
|
|
builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{256, 256},
|
|
1.1,
|
|
phi::DataType::FLOAT32,
|
|
phi::CPUPlace());
|
|
|
|
paddle::dialect::MatmulOp matmul_2 = builder.Build<paddle::dialect::MatmulOp>(
|
|
matmul_1_in_1.out(), matmul_2_in_2.out(), false, false);
|
|
|
|
paddle::dialect::FullOp add_2_in_2 = builder.Build<paddle::dialect::FullOp>(
|
|
std::vector<int64_t>{256}, 1.5, phi::DataType::FLOAT32, phi::CPUPlace());
|
|
paddle::dialect::AddOp add_op2 =
|
|
builder.Build<paddle::dialect::AddOp>(matmul_2.out(), add_2_in_2.out());
|
|
|
|
paddle::dialect::ReshapeOp reshape_2 =
|
|
builder.Build<paddle::dialect::ReshapeOp>(
|
|
add_op2.out(), std::vector<int64_t>{0, 0, 8, 32});
|
|
|
|
paddle::dialect::TransposeOp transpose_2 =
|
|
builder.Build<paddle::dialect::TransposeOp>(reshape_2.out(),
|
|
std::vector<int>{0, 2, 1, 3});
|
|
|
|
// The third path to matmul (v).
|
|
paddle::dialect::FullOp matmul_3_in_2 =
|
|
builder.Build<paddle::dialect::FullOp>(std::vector<int64_t>{256, 256},
|
|
1.1,
|
|
phi::DataType::FLOAT32,
|
|
phi::CPUPlace());
|
|
paddle::dialect::MatmulOp matmul_3 = builder.Build<paddle::dialect::MatmulOp>(
|
|
matmul_1_in_1.out(), matmul_3_in_2.out(), false, false);
|
|
|
|
paddle::dialect::FullOp add_3_in_2 = builder.Build<paddle::dialect::FullOp>(
|
|
std::vector<int64_t>{256}, 1.5, phi::DataType::FLOAT32, phi::CPUPlace());
|
|
|
|
paddle::dialect::AddOp add_3 =
|
|
builder.Build<paddle::dialect::AddOp>(matmul_3.out(), add_3_in_2.out());
|
|
|
|
paddle::dialect::ReshapeOp reshape_3 =
|
|
builder.Build<paddle::dialect::ReshapeOp>(
|
|
add_3.out(), std::vector<int64_t>{0, 0, 8, 32});
|
|
|
|
paddle::dialect::TransposeOp transpose_3 =
|
|
builder.Build<paddle::dialect::TransposeOp>(reshape_3.out(),
|
|
std::vector<int>{0, 2, 1, 3});
|
|
|
|
// softmax(qk)v
|
|
paddle::dialect::MatmulOp matmul_4 = builder.Build<paddle::dialect::MatmulOp>(
|
|
scale_op.out(), transpose_2.out(), false, true);
|
|
|
|
paddle::dialect::FullOp add_4_in_2 = builder.Build<paddle::dialect::FullOp>(
|
|
std::vector<int64_t>{1, 8, 300, 300},
|
|
1.5,
|
|
phi::DataType::FLOAT32,
|
|
phi::CPUPlace());
|
|
|
|
paddle::dialect::AddOp add_4 =
|
|
builder.Build<paddle::dialect::AddOp>(matmul_4.out(), add_4_in_2.out());
|
|
|
|
paddle::dialect::SoftmaxOp softmax_op =
|
|
builder.Build<paddle::dialect::SoftmaxOp>(add_4.out(), -1);
|
|
paddle::dialect::MatmulOp matmul_5 = builder.Build<paddle::dialect::MatmulOp>(
|
|
softmax_op.out(), transpose_3.out(), false, false);
|
|
|
|
paddle::dialect::TransposeOp transpose_4 =
|
|
builder.Build<paddle::dialect::TransposeOp>(matmul_5.out(),
|
|
std::vector<int>{0, 2, 1, 3});
|
|
|
|
paddle::dialect::ReshapeOp reshape_4 =
|
|
builder.Build<paddle::dialect::ReshapeOp>(
|
|
transpose_4.out(), std::vector<int64_t>{0, 0, 256});
|
|
|
|
builder.Build<paddle::dialect::FetchOp>(reshape_4.out(), "out", 0);
|
|
}
|
|
|
|
TEST(DrrTest, AttentionFuse) {
|
|
pir::IrContext *ctx = pir::IrContext::Instance();
|
|
ctx->GetOrRegisterDialect<paddle::dialect::OperatorDialect>();
|
|
ctx->GetOrRegisterDialect<pir::BuiltinDialect>();
|
|
pir::Program program(ctx);
|
|
pir::Builder builder = pir::Builder(ctx, program.block());
|
|
BuildProgram(builder);
|
|
EXPECT_EQ(program.block()->size(), 33u);
|
|
|
|
pir::PassManager pm(ctx);
|
|
pm.AddPass(pir::CreateMultiHeadMatmulFusePass());
|
|
std::unique_ptr<pir::Pass> constant_folding_pass =
|
|
pir::CreateConstantFoldingPass();
|
|
constant_folding_pass->Set(pir::Pass::kPlaceAttr,
|
|
new phi::Place{phi::GPUPlace{}});
|
|
constant_folding_pass->Set(pir::Pass::kParamScopeAttr,
|
|
new paddle::framework::Scope{});
|
|
pm.AddPass(std::move(constant_folding_pass));
|
|
pm.AddPass(pir::CreateDeadCodeEliminationPass());
|
|
pm.EnableIRPrinting();
|
|
|
|
PADDLE_ENFORCE_EQ(pm.Run(&program),
|
|
true,
|
|
common::errors::Unavailable("pm fail to run program"));
|
|
EXPECT_EQ(program.block()->size(), 2u);
|
|
}
|