* Copyright (c) 2026 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_geir_eltwise.cpp
* @brief GE IR graph-mode example - Eltwise (DYNAMIC_INPUT)
*
* Three modes supported:
* mode=0 (PRODUCT): out = x0 * x1 * ... * x_{n-1}
* mode=1 (SUM): out = c0*x0 + c1*x1 + ... + c_{n-1}*x_{n-1}
* mode=2 (MAX): out = max(x0, x1, ..., x_{n-1})
*
* This example uses N=2 inputs with mode=1 (SUM, default coeffs = 1.0),
* which yields out = x0 + x1 (= 2.0 for all-ones inputs).
*/
#include <iostream>
#include <fstream>
#include <string.h>
#include <stdint.h>
#include <vector>
#include <string>
#include <map>
#include <cmath>
#include "assert.h"
#include "graph.h"
#include "types.h"
#include "tensor.h"
#include "ge_error_codes.h"
#include "ge_api_types.h"
#include "ge_api.h"
#include "array_ops.h"
#include "ge_ir_build.h"
#include "../op_graph/eltwise_proto.h"
#define FAILED -1
#define SUCCESS 0
using namespace ge;
using std::map;
using std::string;
using std::vector;
#define ADD_DYNAMIC_INPUT(inputIndex, inputDtype, inputShape, dataName) \
do { \
vector<int64_t> shape##dataName = inputShape; \
auto placeholder##dataName = op::Data("placeholder" #dataName).set_attr_index(inputIndex); \
TensorDesc desc##dataName(ge::Shape(shape##dataName), FORMAT_ND, inputDtype); \
desc##dataName.SetFormat(FORMAT_ND); \
placeholder##dataName.update_input_desc_x(desc##dataName); \
placeholder##dataName.update_output_desc_y(desc##dataName); \
Tensor tensor##dataName; \
ret = GenInputData(shape##dataName, tensor##dataName, desc##dataName, inputDtype); \
if (ret != SUCCESS) { \
printf("%s - ERROR - [XIR]: Generate input data failed\n", GetTime().c_str()); \
return FAILED; \
} \
eltwiseOp.UpdateDynamicInputDesc("x", inputIndex, desc##dataName); \
eltwiseOp.set_dynamic_input_x(inputIndex, placeholder##dataName); \
input.push_back(tensor##dataName); \
graph.AddOp(placeholder##dataName); \
inputs.push_back(placeholder##dataName); \
} while (0)
#define ADD_OUTPUT(outputName, outputDtype, outputShape) \
do { \
TensorDesc outputName##_desc = TensorDesc(ge::Shape(outputShape), FORMAT_ND, outputDtype); \
eltwiseOp.update_output_desc_##outputName(outputName##_desc); \
} while (0)
string GetTime()
{
time_t timep;
time(&timep);
char tmp[64];
strftime(tmp, sizeof(tmp), "%Y-%m-%d %H:%M:%S,000", localtime(&timep));
return tmp;
}
uint32_t GetDataTypeSize(DataType dt)
{
if (dt == ge::DT_FLOAT) {
return 4;
} else if (dt == ge::DT_FLOAT16 || dt == ge::DT_BF16) {
return 2;
}
return 4;
}
constexpr uint16_t FLOAT16_ONE_BITS = 0x3C00;
constexpr uint16_t BFLOAT16_ONE_BITS = 0x3F80;
constexpr uint32_t BF16_TO_FP32_SHIFT = 16U;
constexpr uint16_t FP16_SIGN_MASK = 0x8000U;
constexpr uint32_t FP16_EXP_MASK = 0x1FU;
constexpr uint32_t FP16_MANT_MASK = 0x3FFU;
constexpr uint32_t FP16_EXP_SHIFT = 10U;
constexpr uint32_t FP16_MANT_TO_FP32_SHIFT = 13U;
constexpr uint32_t FP16_TO_FP32_SIGN_SHIFT = 16U;
constexpr uint32_t FP32_EXP_BIAS_DELTA = 112U;
constexpr uint32_t FP32_MANT_SHIFT = 23U;
constexpr uint32_t FP32_INF_BITS = 0x7F800000U;
constexpr float FP16_DENORM_SCALE = 16777216.0f;
static float Bf16ToFp32(uint16_t bf16Bits)
{
uint32_t fp32Bits = static_cast<uint32_t>(bf16Bits) << BF16_TO_FP32_SHIFT;
float out;
__builtin_memcpy(&out, &fp32Bits, sizeof(float));
return out;
}
static float Fp16ToFp32(uint16_t fp16Bits)
{
const uint32_t sign = static_cast<uint32_t>(fp16Bits & FP16_SIGN_MASK) << FP16_TO_FP32_SIGN_SHIFT;
const uint32_t exp = (fp16Bits >> FP16_EXP_SHIFT) & FP16_EXP_MASK;
const uint32_t mant = fp16Bits & FP16_MANT_MASK;
uint32_t fp32Bits;
if (exp == FP16_EXP_MASK) {
fp32Bits = sign | FP32_INF_BITS | (mant << FP16_MANT_TO_FP32_SHIFT);
} else if (exp == 0) {
if (mant == 0) {
fp32Bits = sign;
} else {
float denorm = static_cast<float>(mant) / FP16_DENORM_SCALE;
return sign ? -denorm : denorm;
}
} else {
fp32Bits = sign | ((exp + FP32_EXP_BIAS_DELTA) << FP32_MANT_SHIFT) | (mant << FP16_MANT_TO_FP32_SHIFT);
}
float out;
__builtin_memcpy(&out, &fp32Bits, sizeof(float));
return out;
}
int32_t GenInputData(vector<int64_t> shapes, Tensor& inputTensor, TensorDesc& inputTensorDesc, DataType dataType)
{
inputTensorDesc.SetRealDimCnt(shapes.size());
size_t size = 1;
for (uint32_t i = 0; i < shapes.size(); i++) {
size *= shapes[i];
}
uint32_t dataLen = size * GetDataTypeSize(dataType);
std::vector<uint8_t> pData(dataLen);
if (dataType == DT_FLOAT16) {
uint16_t* fp16Data = reinterpret_cast<uint16_t*>(pData.data());
for (size_t i = 0; i < size; ++i) {
fp16Data[i] = FLOAT16_ONE_BITS;
}
} else if (dataType == DT_BF16) {
uint16_t* bf16Data = reinterpret_cast<uint16_t*>(pData.data());
for (size_t i = 0; i < size; ++i) {
bf16Data[i] = BFLOAT16_ONE_BITS;
}
} else {
float* fp32Data = reinterpret_cast<float*>(pData.data());
for (size_t i = 0; i < size; ++i) {
fp32Data[i] = 1.0f;
}
}
inputTensor = Tensor(inputTensorDesc, pData.data(), dataLen);
return SUCCESS;
}
int CreateOppInGraph(DataType inDtype, std::vector<ge::Tensor>& input, std::vector<Operator>& inputs,
std::vector<Operator>& outputs, Graph& graph)
{
Status ret = SUCCESS;
constexpr int64_t kInputNum = 2;
auto eltwiseOp = op::Eltwise("eltwise_1").create_dynamic_input_x(kInputNum, false);
std::vector<int64_t> xShape = {4, 4, 4};
ADD_DYNAMIC_INPUT(0, inDtype, xShape, 1);
ADD_DYNAMIC_INPUT(1, inDtype, xShape, 2);
eltwiseOp.set_attr_N(kInputNum);
eltwiseOp.set_attr_mode(1);
eltwiseOp.set_attr_coeff({1.0f, 1.0f});
ADD_OUTPUT(y, inDtype, xShape);
outputs.push_back(eltwiseOp);
return SUCCESS;
}
int main(int argc, char* argv[])
{
const char* graph_name = "eltwise_ge_ir_test";
Graph graph(graph_name);
std::vector<ge::Tensor> input;
printf("%s - INFO - [XIR]: Start to initialize ge using ge global options\n", GetTime().c_str());
std::map<AscendString, AscendString> global_options = {{"ge.exec.deviceId", "0"}, {"ge.graphRunMode", "1"}};
Status ret = ge::GEInitialize(global_options);
if (ret != SUCCESS) {
printf("%s - ERROR - [XIR]: Initialize ge using ge global options failed\n", GetTime().c_str());
return FAILED;
}
printf("%s - INFO - [XIR]: Initialize ge using ge global options success\n", GetTime().c_str());
std::vector<Operator> inputs{};
std::vector<Operator> outputs{};
DataType inDtype = DT_FLOAT;
if (argc > 1) {
std::string dtypeArg(argv[1]);
if (dtypeArg == "fp16") {
inDtype = DT_FLOAT16;
} else if (dtypeArg == "bf16") {
inDtype = DT_BF16;
}
}
printf("%s - INFO - [XIR]: Using dtype = %d\n", GetTime().c_str(), static_cast<int>(inDtype));
ret = CreateOppInGraph(inDtype, input, inputs, outputs, graph);
if (ret != SUCCESS) {
printf("%s - ERROR - [XIR]: Create op in graph failed\n", GetTime().c_str());
GEFinalize();
return FAILED;
}
if (!inputs.empty() && !outputs.empty()) {
graph.SetInputs(inputs).SetOutputs(outputs);
}
std::map<AscendString, AscendString> build_options = {};
printf("%s - INFO - [XIR]: Start to create ir session using build options\n", GetTime().c_str());
ge::Session* session = new Session(build_options);
if (session == nullptr) {
printf("%s - ERROR - [XIR]: Create ir session using build options failed\n", GetTime().c_str());
return FAILED;
}
printf("%s - INFO - [XIR]: Create ir session using build options success\n", GetTime().c_str());
printf("%s - INFO - [XIR]: Start to add compute graph to ir session\n", GetTime().c_str());
std::map<AscendString, AscendString> graph_options = {};
uint32_t graph_id = 0;
ret = session->AddGraph(graph_id, graph, graph_options);
if (ret != SUCCESS) {
printf("%s - ERROR - [XIR]: Add graph to session failed\n", GetTime().c_str());
delete session;
GEFinalize();
return FAILED;
}
printf("%s - INFO - [XIR]: Session add ir compute graph to ir session success\n", GetTime().c_str());
printf("%s - INFO - [XIR]: Start to run ir compute graph\n", GetTime().c_str());
std::vector<ge::Tensor> output;
ret = session->RunGraph(graph_id, input, output);
if (ret != SUCCESS) {
printf("%s - ERROR - [XIR]: Run graph failed\n", GetTime().c_str());
delete session;
GEFinalize();
return FAILED;
}
printf("%s - INFO - [XIR]: Session run ir compute graph success\n", GetTime().c_str());
int output_num = output.size();
printf("%s - INFO - [XIR]: Number of outputs: %d\n", GetTime().c_str(), output_num);
for (int i = 0; i < output_num; i++) {
int64_t output_shape = output[i].GetTensorDesc().GetShape().GetShapeSize();
printf("%s - INFO - [XIR]: output %d shape size = %ld, dtype = %d\n", GetTime().c_str(), i, output_shape,
static_cast<int>(output[i].GetTensorDesc().GetDataType()));
if (output_shape > 0) {
const DataType outDtype = output[i].GetTensorDesc().GetDataType();
uint8_t* outData = output[i].GetData();
int64_t printNum = (output_shape < 10) ? output_shape : 10;
const float expectedSum = 2.0f;
for (int64_t j = 0; j < printNum; j++) {
float val = 0.0f;
if (outDtype == DT_FLOAT16) {
val = Fp16ToFp32(reinterpret_cast<uint16_t*>(outData)[j]);
} else if (outDtype == DT_BF16) {
val = Bf16ToFp32(reinterpret_cast<uint16_t*>(outData)[j]);
} else {
val = reinterpret_cast<float*>(outData)[j];
}
printf(" output[%ld] = %.6f (expected = %.6f)\n", j, val, expectedSum);
}
}
}
ge::AscendString error_msg = ge::GEGetErrorMsgV2();
std::string error_str(error_msg.GetString());
if (!error_str.empty()) {
std::cout << "Error message: " << error_str << std::endl;
}
printf("%s - INFO - [XIR]: Start to finalize ir graph session\n", GetTime().c_str());
delete session;
ret = ge::GEFinalize();
if (ret != SUCCESS) {
printf("%s - ERROR - [XIR]: Finalize ir graph session failed\n", GetTime().c_str());
return FAILED;
}
printf("%s - INFO - [XIR]: Finalize ir graph session success\n", GetTime().c_str());
return SUCCESS;
}
