* Copyright (c) 2025 Huawei Technologies Co., Ltd.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
* \file test_reduce_copy.cpp
* \brief Unit test for ReduceCopy pass.
*/
#include <vector>
#include <string>
#include "gtest/gtest.h"
#include "tilefwk/data_type.h"
#include "tilefwk/tilefwk_op.h"
#include "tilefwk/platform.h"
#include "interface/function/function.h"
#include "passes/tile_graph_pass/graph_partition/reduce_copy.h"
#include "tilefwk/tilefwk.h"
#include "interface/inner/tilefwk.h"
#include "passes/pass_mgr/pass_manager.h"
#include "interface/configs/config_manager.h"
#include "computational_graph_builder.h"
namespace npu {
namespace tile_fwk {
class ReduceCopyTest : public testing::Test {
public:
static void SetUpTestCase() {}
static void TearDownTestCase() {}
void SetUp() override
{
Program::GetInstance().Reset();
config::Reset();
config::SetHostOption(COMPILE_STAGE, CS_EXECUTE_GRAPH);
config::SetHostConfig(KEY_STRATEGY, "ReduceCopyTestStrategy");
Platform::Instance().GetSoc().SetNPUArch(NPUArch::DAV_3510);
}
void TearDown() override
{
Platform::Instance().GetSoc().SetNPUArch(NPUArch::DAV_UNKNOWN);
}
};
void BuildMatmulAddBranch(
ComputationalGraphBuilder& G, int brId, std::vector<std::string>& incasts, std::vector<std::string>& outcasts)
{
std::vector<int64_t> tileShape{16, 16};
std::string br = std::to_string(brId);
std::vector<std::string> tensorNames{"tRA" + br, "tRB" + br, "tL1A" + br, "tL1B" + br,
"tA" + br, "tB" + br, "tC" + br, "tUB" + br};
std::vector<Opcode> opCodes{Opcode::OP_VIEW, Opcode::OP_VIEW, Opcode::OP_L1_TO_L0A,
Opcode::OP_L1_TO_L0B, Opcode::OP_A_MUL_B, Opcode::OP_CONVERT};
std::vector<std::vector<std::string>> ioperands{{"tRA" + br}, {"tRB" + br}, {"tL1A" + br},
{"tL1B" + br}, {"tA" + br, "tB" + br}, {"tC" + br}};
std::vector<std::vector<std::string>> ooperands{{"tL1A" + br}, {"tL1B" + br}, {"tA" + br},
{"tB" + br}, {"tC" + br}, {"tUB" + br}};
std::vector<std::string> opNames{"view" + br, "view2" + br, "toA" + br, "toB" + br, "matmul" + br, "convert" + br};
EXPECT_EQ(G.AddTensors(DataType::DT_FP32, tileShape, tensorNames), true);
EXPECT_EQ(G.AddOps(opCodes, ioperands, ooperands, opNames, true), true);
for (auto aicOp : opNames) {
G.GetOp(aicOp)->SetAttr(OpAttributeKey::isCube, true);
}
incasts.push_back("tRA" + br);
incasts.push_back("tRB" + br);
const int Num50 = 50;
for (auto opName : opNames) {
G.GetOp(opName)->UpdateSubgraphID(brId);
G.GetOp(opName)->UpdateLatency(Num50);
}
const int Num2 = 2;
for (int k = 0; k < Num2; k++) {
std::string brv1 = std::to_string(brId + 1 + k);
std::vector<std::string> tensorNamesV1{"add1" + brv1, "add2" + brv1, "out" + brv1};
std::vector<Opcode> opCodesV1{Opcode::OP_ADDS, Opcode::OP_ADDS, Opcode::OP_ASSEMBLE};
std::vector<std::vector<std::string>> ioperandsV1{
{"tUB" + br},
{"add1" + brv1},
{"add2" + brv1},
};
std::vector<std::vector<std::string>> ooperandsV1{
{"add1" + brv1},
{"add2" + brv1},
{"out" + brv1},
};
std::vector<std::string> opNamesV1{"add1" + brv1, "add2" + brv1, "assemble" + brv1};
EXPECT_EQ(G.AddTensors(DataType::DT_FP32, tileShape, tensorNamesV1), true);
EXPECT_EQ(G.AddOps(opCodesV1, ioperandsV1, ooperandsV1, opNamesV1, true), true);
outcasts.push_back("out" + brv1);
for (auto opName : opNamesV1) {
G.GetOp(opName)->UpdateSubgraphID(brId + 1 + k);
G.GetOp(opName)->UpdateLatency(Num50);
G.GetOp(opName)->SetAttr(OpAttributeKey::isCube, false);
}
}
}
void BuildMatmulAddsGraph(ComputationalGraphBuilder& G)
{
std::vector<std::string> incasts;
std::vector<std::string> outcasts;
const int Num3 = 3;
BuildMatmulAddBranch(G, 0, incasts, outcasts);
BuildMatmulAddBranch(G, Num3, incasts, outcasts);
Function* function = G.GetFunction();
const int Num6 = 6;
function->SetTotalSubGraphCount(Num6);
EXPECT_EQ(G.SetInCast(incasts), true);
EXPECT_EQ(G.SetOutCast(outcasts), true);
}
TEST_F(ReduceCopyTest, TestCase0)
{
ComputationalGraphBuilder G;
BuildMatmulAddsGraph(G);
Function* function = G.GetFunction();
ASSERT_NE(function, nullptr);
ReduceCopyMerge merger;
function->paramConfigs_.autoMixPartition = 1;
EXPECT_EQ(merger.RunOnFunction(*function), SUCCESS);
const int Num2 = 2;
EXPECT_EQ(function->GetTotalSubGraphCount(), Num2);
}
void BuildConnectMatmul(
ComputationalGraphBuilder& G, int brId, std::vector<std::string>& incasts, std::vector<std::string>& outcasts)
{
std::vector<int64_t> tileShape{16, 16};
const int Num50 = 50;
std::string br = std::to_string(brId);
std::vector<std::string> tensorNames{"tRA" + br, "tRB" + br, "tL1A" + br, "tL1B" + br,
"tA" + br, "tB" + br, "tC" + br, "tGM" + br};
std::vector<Opcode> opCodes{Opcode::OP_VIEW, Opcode::OP_VIEW, Opcode::OP_L1_TO_L0A,
Opcode::OP_L1_TO_L0B, Opcode::OP_A_MUL_B, Opcode::OP_ASSEMBLE};
std::vector<std::vector<std::string>> ioperands{{"tRA" + br}, {"tRB" + br}, {"tL1A" + br},
{"tL1B" + br}, {"tA" + br, "tB" + br}, {"tC" + br}};
std::vector<std::vector<std::string>> ooperands{{"tL1A" + br}, {"tL1B" + br}, {"tA" + br},
{"tB" + br}, {"tC" + br}, {"tGM" + br}};
std::vector<std::string> opNames{"view" + br, "view2" + br, "toA" + br, "toB" + br, "matmul" + br, "convert" + br};
EXPECT_EQ(G.AddTensors(DataType::DT_FP32, tileShape, tensorNames), true);
EXPECT_EQ(G.AddOps(opCodes, ioperands, ooperands, opNames, true), true);
incasts.push_back("tRA" + br);
incasts.push_back("tRB" + br);
for (auto opName : opNames) {
G.GetOp(opName)->UpdateSubgraphID(brId);
G.GetOp(opName)->UpdateLatency(Num50);
G.GetOp(opName)->SetAttr(OpAttributeKey::isCube, true);
}
std::string br2 = std::to_string(brId + 1);
std::vector<std::string> tensorNames2{"tRB" + br2, "tL1A" + br2, "tL1B" + br2, "tA" + br2,
"tB" + br2, "tC" + br2, "tGM" + br2};
std::vector<Opcode> opCodes2{Opcode::OP_VIEW, Opcode::OP_VIEW, Opcode::OP_L1_TO_L0A,
Opcode::OP_L1_TO_L0B, Opcode::OP_A_MUL_B, Opcode::OP_ASSEMBLE};
std::vector<std::vector<std::string>> ioperands2{
{"tGM" + br}, {"tRB" + br2}, {"tL1A" + br2}, {"tL1B" + br2}, {"tA" + br2, "tB" + br2}, {"tC" + br2}};
std::vector<std::vector<std::string>> ooperands2{{"tL1A" + br2}, {"tL1B" + br2}, {"tA" + br2},
{"tB" + br2}, {"tC" + br2}, {"tGM" + br2}};
std::vector<std::string> opNames2{"view" + br2, "view2" + br2, "toA" + br2,
"toB" + br2, "matmul" + br2, "convert" + br2};
EXPECT_EQ(G.AddTensors(DataType::DT_FP32, tileShape, tensorNames2), true);
EXPECT_EQ(G.AddOps(opCodes, ioperands2, ooperands2, opNames2, true), true);
incasts.push_back("tRB" + br2);
outcasts.push_back("tGM" + br2);
for (auto opName : opNames2) {
G.GetOp(opName)->UpdateSubgraphID(brId + 1);
G.GetOp(opName)->UpdateLatency(Num50);
}
}
void BuildConnectVector(
ComputationalGraphBuilder& G, int brId, std::vector<std::string>& incasts, std::vector<std::string>& outcasts)
{
std::vector<int64_t> tileShape{16, 16};
std::string br = std::to_string(brId);
std::vector<std::string> tensorNames{"tin" + br, "tadds1" + br, "tout" + br};
std::vector<Opcode> opCodes{Opcode::OP_ADDS, Opcode::OP_ADDS};
std::vector<std::vector<std::string>> ioperands{{"tin" + br}, {"tadds1" + br}};
std::vector<std::vector<std::string>> ooperands{{"tadds1" + br}, {"tout" + br}};
std::vector<std::string> opNames{"adds1" + br, "adds2" + br};
EXPECT_EQ(G.AddTensors(DataType::DT_FP32, tileShape, tensorNames), true);
EXPECT_EQ(G.AddOps(opCodes, ioperands, ooperands, opNames, true), true);
incasts.push_back("tin" + br);
outcasts.push_back("tout" + br);
const int Num50 = 50;
G.GetOp("adds1" + br)->UpdateSubgraphID(brId);
G.GetOp("adds1" + br)->UpdateLatency(Num50);
G.GetOp("adds1" + br)->SetAttr(OpAttributeKey::isCube, false);
G.GetOp("adds2" + br)->UpdateSubgraphID(brId + 1);
G.GetOp("adds2" + br)->UpdateLatency(Num50);
G.GetOp("adds2" + br)->SetAttr(OpAttributeKey::isCube, false);
}
void BuildConnect(ComputationalGraphBuilder& G)
{
std::vector<std::string> incasts;
std::vector<std::string> outcasts;
const int Num2 = 2;
BuildConnectMatmul(G, 0, incasts, outcasts);
BuildConnectVector(G, Num2, incasts, outcasts);
Function* function = G.GetFunction();
const int Num4 = 4;
function->SetTotalSubGraphCount(Num4);
EXPECT_EQ(G.SetInCast(incasts), true);
EXPECT_EQ(G.SetOutCast(outcasts), true);
}
TEST_F(ReduceCopyTest, TestCase1)
{
ComputationalGraphBuilder G;
BuildConnect(G);
Function* function = G.GetFunction();
function->paramConfigs_.autoMixPartition = 1;
ReduceCopyMerge merger;
merger.RunOnFunction(*function);
const int Num3 = 3;
EXPECT_EQ(function->GetTotalSubGraphCount(), Num3);
}
TEST_F(ReduceCopyTest, TestCase2)
{
ComputationalGraphBuilder G;
BuildConnect(G);
Function* function = G.GetFunction();
function->paramConfigs_.autoMixPartition = 1;
G.GetOp("adds12")->scopeInfo_.cvFuseId = 0;
G.GetOp("adds22")->scopeInfo_.cvFuseId = 0;
ReduceCopyMerge merger;
merger.RunOnFunction(*function);
const int Num2 = 2;
EXPECT_EQ(function->GetTotalSubGraphCount(), Num2);
}
TEST_F(ReduceCopyTest, PreserveOriginalSubgraphId)
{
ComputationalGraphBuilder G;
BuildMatmulAddsGraph(G);
Function* function = G.GetFunction();
ASSERT_NE(function, nullptr);
auto* add11 = G.GetOp("add11");
ASSERT_NE(add11, nullptr);
const int64_t add11SubgraphId = static_cast<int64_t>(add11->GetSubgraphID());
auto* add14 = G.GetOp("add14");
ASSERT_NE(add14, nullptr);
const int64_t add14SubgraphId = static_cast<int64_t>(add14->GetSubgraphID());
function->paramConfigs_.autoMixPartition = 1;
ReduceCopyMerge merger;
EXPECT_EQ(merger.RunOnFunction(*function), SUCCESS);
int64_t preSubgraphId = -1;
ASSERT_TRUE(add11->GetAttr(OpAttributeKey::reduceCopyPreSubgraphId, preSubgraphId));
EXPECT_EQ(preSubgraphId, add11SubgraphId);
ASSERT_TRUE(add14->GetAttr(OpAttributeKey::reduceCopyPreSubgraphId, preSubgraphId));
EXPECT_EQ(preSubgraphId, add14SubgraphId);
}
TEST_F(ReduceCopyTest, TestCase3)
{
ComputationalGraphBuilder G;
BuildConnect(G);
Function* function = G.GetFunction();
function->paramConfigs_.autoMixPartition = 1;
const int largeNum = 2e7;
G.GetOp("matmul0")->UpdateLatency(largeNum);
ReduceCopyMerge merger;
merger.RunOnFunction(*function);
const int Num4 = 4;
EXPECT_EQ(function->GetTotalSubGraphCount(), Num4);
}
}
}
