* 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_prod_operation.cpp
* \brief
*/
#include "test_operation.h"
using namespace tile_fwk::test_operation;
namespace {
const unsigned IDX_DIM0 = 0;
const unsigned IDX_DIM1 = 1;
const unsigned IDX_DIM2 = 2;
const unsigned IDX_DIM3 = 3;
struct ProdOpFuncArgs : public OpFuncArgs {
ProdOpFuncArgs(
std::vector<int64_t> viewShape, const std::vector<int64_t> tileShape, std::vector<int64_t> dims,
const bool keepDim)
: viewShape_(viewShape), tileShape_(tileShape), dims_(dims), keepDim_(keepDim)
{}
std::vector<int64_t> viewShape_;
std::vector<int64_t> tileShape_;
std::vector<int64_t> dims_;
bool keepDim_;
};
struct ProdOpMetadata {
explicit ProdOpMetadata(const OpFunc& opFunc, const nlohmann::json& test_data)
: opFunc_(opFunc), test_data_(test_data)
{}
OpFunc opFunc_;
nlohmann::json test_data_;
};
void AdjustTileShapeForReduce(const int dim, const Tensor& result, std::vector<int64_t> tileshape)
{
tileshape.erase(tileshape.begin() + dim);
const int alignNum = BLOCK_SIZE / BytesOf(result.GetStorage()->tensor->datatype);
tileshape[tileshape.size() - 1] = (tileshape[tileshape.size() - 1] + alignNum - 1) / alignNum * alignNum;
TileShape::Current().SetVecTile(tileshape);
}
void ProdOperationExeFunc(const std::vector<Tensor>& inputs, std::vector<Tensor>& outputs, const OpFuncArgs* opArgs)
{
auto args = static_cast<const ProdOpFuncArgs*>(opArgs);
FUNCTION("main", {inputs[0]}, {outputs[0]})
{
SymbolicScalar firstDim = inputs[0].GetShape()[0];
SymbolicScalar secondDim = inputs[0].GetShape()[1];
int dim = args->dims_[0];
bool keepDim = args->keepDim_;
if (dim < 0) {
dim = static_cast<int>(inputs[0].GetShape().size()) + dim;
}
SymbolicScalar viewShape[] = {args->viewShape_[0], args->viewShape_[1]};
viewShape[dim] = 0;
const int batch = CeilDiv(inputs[0].GetShape()[1 - dim], viewShape[1 - dim]);
LOOP("LOOP_L0_bIdx", FunctionType::DYNAMIC_LOOP, bIdx, LoopRange(batch))
{
auto viewTensor = View(
inputs[0], {viewShape[0] == 0 ? firstDim : viewShape[0], viewShape[1] == 0 ? secondDim : viewShape[1]},
{viewShape[0] == 0 ? firstDim : std::min(firstDim - bIdx * viewShape[0], viewShape[0]),
viewShape[1] == 0 ? secondDim : std::min(secondDim - bIdx * viewShape[1], viewShape[1])},
{bIdx * viewShape[0], bIdx * viewShape[1]});
TileShape::Current().SetVecTile(args->tileShape_);
std::vector<SymbolicScalar> offset = {bIdx * viewShape[0], bIdx * viewShape[1]};
auto res = Prod(viewTensor, args->dims_[0], keepDim);
if (!keepDim) {
offset.erase(offset.begin() + dim);
AdjustTileShapeForReduce(dim, res, args->tileShape_);
}
Assemble(res, offset, outputs[0]);
}
}
}
void Prod3DOperationExeFunc(const std::vector<Tensor>& inputs, std::vector<Tensor>& outputs, const OpFuncArgs* opArgs)
{
auto args = static_cast<const ProdOpFuncArgs*>(opArgs);
FUNCTION("main", {inputs[0]}, {outputs[0]})
{
SymbolicScalar firstDim = inputs[0].GetShape()[0];
SymbolicScalar secondDim = inputs[0].GetShape()[1];
SymbolicScalar lastDim = inputs[0].GetShape()[2];
int dim = args->dims_[0];
bool keepDim = args->keepDim_;
if (dim < 0) {
dim = static_cast<int>(inputs[0].GetShape().size()) + dim;
}
SymbolicScalar viewShape[] = {args->viewShape_[0], args->viewShape_[1], args->viewShape_[2]};
int loops[] = {
CeilDiv(inputs[0].GetShape()[0], viewShape[0]), CeilDiv(inputs[0].GetShape()[1], viewShape[1]),
CeilDiv(inputs[0].GetShape()[2], viewShape[2])};
viewShape[dim] = 0;
loops[dim] = 1;
if (dim == IDX_DIM0) {
viewShape[IDX_DIM1] = 0;
loops[IDX_DIM1] = 1;
} else {
viewShape[IDX_DIM0] = 0;
loops[IDX_DIM0] = 1;
}
LOOP("LOOP_L0_bIdx", FunctionType::DYNAMIC_LOOP, bIdx, LoopRange(loops[IDX_DIM0]))
{
LOOP("LOOP_L1_bIdx", FunctionType::DYNAMIC_LOOP, sIdx, LoopRange(loops[IDX_DIM1]))
{
LOOP("LOOP_L2_bIdx", FunctionType::DYNAMIC_LOOP, nIdx, LoopRange(loops[IDX_DIM2]))
{
auto viewTensor = View(
inputs[0],
{viewShape[0] == 0 ? firstDim : viewShape[0], viewShape[1] == 0 ? secondDim : viewShape[1],
viewShape[2] == 0 ? lastDim : viewShape[2]},
{viewShape[0] == 0 ? firstDim : std::min(firstDim - bIdx * viewShape[0], viewShape[0]),
viewShape[1] == 0 ? secondDim : std::min(secondDim - sIdx * viewShape[1], viewShape[1]),
viewShape[2] == 0 ? lastDim : std::min(lastDim - nIdx * viewShape[2], viewShape[2])},
{bIdx * viewShape[0], sIdx * viewShape[1], nIdx * viewShape[2]});
TileShape::Current().SetVecTile(args->tileShape_);
std::vector<SymbolicScalar> offset = {
bIdx * viewShape[0], sIdx * viewShape[1], nIdx * viewShape[2]};
auto res = Prod(viewTensor, args->dims_[0], keepDim);
if (!keepDim) {
offset.erase(offset.begin() + dim);
AdjustTileShapeForReduce(dim, res, args->tileShape_);
}
Assemble(res, offset, outputs[0]);
}
}
}
}
}
void Prod4DOperationExeFunc(const std::vector<Tensor>& inputs, std::vector<Tensor>& outputs, const OpFuncArgs* opArgs)
{
auto args = static_cast<const ProdOpFuncArgs*>(opArgs);
FUNCTION("main", {inputs[0]}, {outputs[0]})
{
SymbolicScalar firstDim = inputs[0].GetShape()[0];
SymbolicScalar secondDim = inputs[0].GetShape()[1];
SymbolicScalar thirdDim = inputs[0].GetShape()[2];
SymbolicScalar lastDim = inputs[0].GetShape()[3];
int dim = args->dims_[0];
bool keepDim = args->keepDim_;
if (dim < 0) {
dim = static_cast<int>(inputs[0].GetShape().size()) + dim;
}
SymbolicScalar viewShape[] = {
args->viewShape_[0], args->viewShape_[1], args->viewShape_[2], args->viewShape_[3]};
int loops[] = {
CeilDiv(inputs[0].GetShape()[0], viewShape[0]), CeilDiv(inputs[0].GetShape()[1], viewShape[1]),
CeilDiv(inputs[0].GetShape()[2], viewShape[2]), CeilDiv(inputs[0].GetShape()[3], viewShape[3])};
viewShape[dim] = 0;
loops[dim] = 1;
LOOP("LOOP_L0_bIdx", FunctionType::DYNAMIC_LOOP, bIdx, LoopRange(loops[IDX_DIM0]))
{
LOOP("LOOP_L1_bIdx", FunctionType::DYNAMIC_LOOP, sIdx, LoopRange(loops[IDX_DIM1]))
{
LOOP("LOOP_L2_bIdx", FunctionType::DYNAMIC_LOOP, nIdx, LoopRange(loops[IDX_DIM2]))
{
LOOP("LOOP_L3_bIdx", FunctionType::DYNAMIC_LOOP, qIdx, LoopRange(loops[IDX_DIM3]))
{
std::vector<SymbolicScalar> offset = {
bIdx * viewShape[0], sIdx * viewShape[1], nIdx * viewShape[2], qIdx * viewShape[3]};
auto viewTensor = View(
inputs[0],
{viewShape[0] == 0 ? firstDim : viewShape[0], viewShape[1] == 0 ? secondDim : viewShape[1],
viewShape[2] == 0 ? thirdDim : viewShape[2], viewShape[3] == 0 ? lastDim : viewShape[3]},
{viewShape[0] == 0 ? firstDim : std::min(firstDim - bIdx * viewShape[0], viewShape[0]),
viewShape[1] == 0 ? secondDim : std::min(secondDim - sIdx * viewShape[1], viewShape[1]),
viewShape[2] == 0 ? thirdDim : std::min(thirdDim - nIdx * viewShape[2], viewShape[2]),
viewShape[3] == 0 ? lastDim : std::min(lastDim - qIdx * viewShape[3], viewShape[3])},
offset);
TileShape::Current().SetVecTile(args->tileShape_);
auto res = Prod(viewTensor, args->dims_[0], keepDim);
if (!keepDim) {
offset.erase(offset.begin() + dim);
AdjustTileShapeForReduce(dim, res, args->tileShape_);
}
Assemble(res, offset, outputs[0]);
}
}
}
}
}
}
class ProdOperationTest : public npu::tile_fwk::stest::TestSuite_STest_Ops_Aihac_param<ProdOpMetadata> {};
INSTANTIATE_TEST_SUITE_P(
TestProd, ProdOperationTest,
::testing::ValuesIn(
GetOpMetaData<ProdOpMetadata>({ProdOperationExeFunc, Prod3DOperationExeFunc, Prod4DOperationExeFunc}, "Prod")));
TEST_P(ProdOperationTest, TestProd)
{
auto test_data = GetParam().test_data_;
auto args = ProdOpFuncArgs(
GetViewShape(test_data), GetTileShape(test_data), GetValueByName<std::vector<int64_t>>(test_data, "dims"),
GetValueByName<bool>(test_data, "keepDim"));
auto testCase = CreateTestCaseDesc<ProdOpMetadata>(GetParam(), &args);
TestExecutor::runTest(testCase);
}
}
