来自 NNIE_Mapper 的 .wk 文件在运行时 运行 时无效
.wk file from NNIE_Mapper not valid when running for runtime
首先，一些背景：
有 2 个 ML 模型：
- 有 50 个输出节点
- 有 5 个输出节点
两者都首先使用 https://github.com/dwayo-gh/AER-CNN-KERAS 进行训练，然后使用 https://github.com/uhfband/keras2caffe.git.
转换为 caffemodel
具有 50 个输出节点的模型的故事：
在 Windows10 环境下使用 RuyiStudio，我将 prototxt + 权重转换为 .wk 文件。生成的 .wk 文件的大小为 594112 字节。
然后我将 HiSVP SDK 的 "sample_runtime" 示例修改为 运行 这个新的 .wk 文件（来源如下）。功能模拟效果很好！它给出了输入文件的前 5 个推理结果。
具有 5 个输出节点的模型的故事：
再次使用RuyiStudio，在Win10环境下，我将文件转换为.wk，更改了运行时间源中的路径。生成的 .wk 文件的大小为 329776 字节。以下是我遇到的错误：
Thu Mar 12 11:08:01 2020 I Try to open [nnie_sim.ini] in execute workspace...
Thu Mar 12 11:08:01 2020 W open [nnie_sim.ini] in execute workspace failure
Thu Mar 12 11:08:01 2020 I Try to open [nnie_sim.ini] in Env NNIE_SIM_INI_DIR...
Thu Mar 12 11:08:01 2020 I getenv(NNIE_SIM_INI_DIR): C:\Users\admin\Downloads\HiSVP_PC_V1.2.2.2\software\hisilicon ...
Thu Mar 12 11:08:01 2020 I open [nnie_sim.ini] success(path: C:/Users/admin/Downloads/HiSVP_PC_V1.2.2.2/software/hisilicon/nnie_sim.ini)
Thu Mar 12 11:08:01 2020 I ============= nnie_sim.ini setting ==============
Thu Mar 12 11:08:01 2020 I [LOG_LEVEL] DBG
Thu Mar 12 11:08:01 2020 I =================================================
====================start to run====================
=======================File: ..\Release.wav.json.bin==========================
I[HiRT] Runtime version: 1.2.2.7
build sha1: 580506f9
compile date: Aug 9 2019 17:24:44
Runtime WK Length: 329776
I[hirt] Model size: 1 Connector size: 0
E[hirt] invalid wk file, total u32Size(329776) should greater than nnie wk u32Size(329776)
E[hirt] parseNnieWK error
E[hirt] model is loading
E[hirt] addModel error
E[hirt] parseModelAndConnector error
E[hirt] load group in addGroup by sync error
E[hirt] AddGroup fail
HI_SVPRT_RUNTIME_LoadModelGroup error
SAMPLE_RUNTIME_LoadModelGroup_Aernet failed!
[Total], TIME SPEND: 1503ms
==========run success==========
main.c
#include <stdio.h>
#include <stdlib.h>
#include "string.h"
#ifdef ON_BOARD
#include "mpi_sys.h"
#include "mpi_vb.h"
#else
#include "hi_comm_svp.h"
#include "hi_nnie.h"
#include "mpi_nnie.h"
#include "hi_type.h"
#endif
#include "math.h"
#include "sample_log.h"
#include "sample_save_blob.h"
#include "sample_data_utils.h"
#include "sample_runtime_classify.h"
static HI_VOID SAMPLE_RUNTIME_Usage(const char *pchPrgName)
{
SAMPLE_LOG_INFO("Usage : %s <filepath>", pchPrgName);
}
#ifdef ON_BOARD
static HI_S32 SAMPLE_COMM_SVP_SysInit(HI_VOID)
{
HI_MPI_SYS_Exit();
HI_MPI_VB_Exit();
VB_CONFIG_S struVbConf = {0};
struVbConf.u32MaxPoolCnt = 2;
struVbConf.astCommPool[1].u64BlkSize = 768 * 576 * 2;
struVbConf.astCommPool[1].u32BlkCnt = 1;
HI_S32 s32Ret = HI_MPI_VB_SetConfig((const VB_CONFIG_S *)&struVbConf);
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_SetConfig error");
return HI_FAILURE;
}
s32Ret = HI_MPI_VB_Init();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_Init error");
return HI_FAILURE;
}
s32Ret = HI_MPI_SYS_Init();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_SYS_Init error");
return HI_FAILURE;
}
return s32Ret;
}
static HI_S32 SAMPLE_COMM_SVP_SysExit(HI_VOID)
{
HI_S32 s32Ret = HI_MPI_SYS_Exit();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_SYS_Exit error");
return HI_FAILURE;
}
s32Ret = HI_MPI_VB_Exit();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_Exit error");
return HI_FAILURE;
}
return HI_SUCCESS;
}
#endif
HI_S32 HandleSampleNet(const HI_CHAR* filename)
{
HI_S32 ret = HI_FAILURE;
SAMPLE_LOG_INFO("\n\n=======================File: %s==========================\n\n", filename);
ret = SAMPLE_AerNet(filename);
return ret;
}
HI_S32 main(HI_S32 argc, HI_CHAR* argv[])
{
if ((argc != 2)) {
SAMPLE_RUNTIME_Usage(argv[0]);
return HI_FAILURE;
}
if (!strncmp(argv[1], "-h", 2)) {
SAMPLE_RUNTIME_Usage(argv[0]);
return HI_SUCCESS;
}
const HI_CHAR *pcSrcFile = argv[1];
SAMPLE_LOG_INFO("\n\n====================start to run====================\n");
#ifdef ON_BOARD
HI_S32 s32Ret = SAMPLE_COMM_SVP_SysInit();
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_COMM_SVP_SysInit fail");
s32Ret = HandleSampleNet(pcSrcFile);
#else
HI_S32 s32Ret = HandleSampleNet(pcSrcFile);
#endif
SAMPLE_CHK_PRINTF((s32Ret == HI_SUCCESS), "\n==========run success==========\n");
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "\n==========run fail==========\n");
#ifdef ON_BOARD
s32Ret = SAMPLE_COMM_SVP_SysExit();
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_COMM_SVP_SysExit fail");
#endif
return s32Ret;
}
aernet.modelgroup
name: "single_aernet"
priority: 0
max_tmpbuf_size_mb: 4096
align_size: 32
input {
name: "data"
}
model {
name: "aernet"
bottom: {name: "data"}
top: {name: "dense_2"}
}
sample_runtime_classify.c
#include "sample_runtime_classify.h"
#include <stdio.h>
#include <stdlib.h>
#ifdef ON_BOARD
#include "mpi_sys.h"
#include "mpi_vb.h"
#else
#include "hi_comm_svp.h"
#include "hi_nnie.h"
#include "mpi_nnie.h"
#endif
#include "hi_runtime_api.h"
#include "string.h"
#include "sample_log.h"
#include "sample_runtime_define.h"
#include "sample_memory_ops.h"
#include "sample_data_utils.h"
#include "sample_model_classification.h"
static const HI_U32 ALEXNET_OUTPUT_TOP_FIVE = 5;
const HI_U32 ALEXNET_DEBUG_PRINT_NUM = 10;
static const HI_U32 AERNET_OUTPUT_TOP_FIVE = 5;
const HI_U32 AERNET_DEBUG_PRINT_NUM = 10;
static BlobInfo aer_stSrcBlobInfo = {
"", "data", "",
HI_RUNTIME_BLOB_TYPE_S32,
{ 1, 1, 128, 300 },
HI_TRUE, ALIGN_16
};
static BlobInfo aer_stDstBlobInfo = {
"aernet", "dense_2", "",
HI_RUNTIME_BLOB_TYPE_VEC_S32,
{ 1, 50, 1, 1 },
HI_FALSE, ALIGN_16
};
HI_S32 SAMPLE_RUNTIME_LoadModelGroup_Aernet(const HI_CHAR *pcModelFileAer, HI_RUNTIME_WK_INFO_S *pstWkInfo,
HI_RUNTIME_GROUP_HANDLE *phGroupHandle)
{
HI_CHAR *pacConfig = NULL;
HI_RUNTIME_GROUP_INFO_S stGroupInfo = {0};
memset(&stGroupInfo, 0, sizeof(HI_RUNTIME_GROUP_INFO_S));
strncpy(pstWkInfo[0].acModelName, "aernet", MAX_NAME_LEN);
HI_S32 s32Ret = SAMPLE_RUNTIME_LoadModelFile(pcModelFileAer, &(pstWkInfo[0].stWKMemory));
SAMPLE_CHK_GOTO((s32Ret != HI_SUCCESS), FAIL, "SAMPLE_RUNTIME_LoadModelFile %s failed!", pcModelFileAer);
stGroupInfo.stWKsInfo.u32WKNum = 1;
stGroupInfo.stWKsInfo.pstAttrs = &(pstWkInfo[0]);
SAMPLE_RUNTIME_ReadConfig(CONFIG_DIR "aernet.modelgroup", &pacConfig);
SAMPLE_CHK_GOTO((pacConfig == NULL), FAIL, "HI_SVPRT_RUNTIME_ReadConfig error");
s32Ret = HI_SVPRT_RUNTIME_LoadModelGroup(pacConfig, &stGroupInfo, phGroupHandle);
SAMPLE_CHK_GOTO((s32Ret != HI_SUCCESS), FAIL, "HI_SVPRT_RUNTIME_LoadModelGroup error");
SAMPLE_LOG_INFO("LoadGroup succ, group handle[%p]", phGroupHandle);
SAMPLE_FREE(pacConfig);
return HI_SUCCESS;
FAIL:
SAMPLE_FREE(pacConfig);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_ForwardGroup_Aernet(const HI_CHAR *pcSrcFile, HI_RUNTIME_GROUP_HANDLE hGroupHandle)
{
HI_RUNTIME_GROUP_SRC_BLOB_ARRAY_S stGroupSrcBlob = {0};
HI_RUNTIME_GROUP_DST_BLOB_ARRAY_S stGroupDstBlob = {0};
strncpy(aer_stSrcBlobInfo.acSrcFilePath, pcSrcFile, sizeof(aer_stSrcBlobInfo.acSrcFilePath));
aer_stSrcBlobInfo.acSrcFilePath[sizeof(aer_stSrcBlobInfo.acSrcFilePath) - 1] = '[=15=]';
HI_S32 s32Ret = InitBlobs(&stGroupSrcBlob, 1, &aer_stSrcBlobInfo);
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "InitBlobs fail");
s32Ret = InitBlobs(&stGroupDstBlob, 1, &aer_stDstBlobInfo);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "InitBlobs fail");
#if PERFORMANCE_TEST
struct timespec start, end;
clock_gettime(0, &start);
#endif
s32Ret = HI_SVPRT_RUNTIME_ForwardGroupSync(hGroupHandle, &stGroupSrcBlob, &stGroupDstBlob, 0);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "HI_SVPRT_RUNTIME_ForwardGroupSync failed!");
#if PERFORMANCE_TEST
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Forward");
#endif
SAMPLE_LOG_INFO("Pic: %s", pcSrcFile);
s32Ret = SAMPLE_RUNTIME_Cnn_TopN_Output(stGroupDstBlob.pstBlobs[0].pstBlob, AERNET_OUTPUT_TOP_FIVE);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "SAMPLE_RUNTIME_Cnn_TopN_Output error");
#if DEBUG
printDebugData((HI_CHAR *)"alexnet", stDst[0].u64VirAddr, ALEXNET_DEBUG_PRINT_NUM);
s32Ret = HI_SUCCESS;
#endif
FAIL_0:
DeinitBlobs(&stGroupDstBlob);
DeinitBlobs(&stGroupSrcBlob);
return s32Ret;
}
HI_S32 SAMPLE_AerNet(HI_CHAR *pcSrcFile)
{
const HI_CHAR *pcRuntimeModelName = MODEL_AERNET_NAME;
HI_RUNTIME_GROUP_HANDLE hGroupHandle = HI_NULL;
HI_RUNTIME_WK_INFO_S astWkInfo[1];
memset(&astWkInfo[0], 0, sizeof(astWkInfo));
struct timespec start, next, end;
clock_gettime(0, &start);
HI_S32 s32Ret = HI_SVPRT_RUNTIME_Init(CPU_TASK_AFFINITY, NULL);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_1, "HI_SVPRT_RUNTIME_Init failed!");
s32Ret = SAMPLE_RUNTIME_LoadModelGroup_Aernet(pcRuntimeModelName, astWkInfo, &hGroupHandle);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_RUNTIME_LoadModelGroup_Aernet failed!");
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Load");
s32Ret = SAMPLE_RUNTIME_ForwardGroup_Aernet(pcSrcFile, hGroupHandle);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "SAMPLE_RUNTIME_ForwardGroup_Alexnet error");
clock_gettime(0, &next);
timeSpendMs(&end, &next, "Forward total");
s32Ret = HI_SVPRT_RUNTIME_UnloadModelGroup(hGroupHandle);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "HI_SVPRT_RUNTIME_UnloadModelGroup error");
FAIL_0:
s32Ret = HI_SVPRT_RUNTIME_DeInit();
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "HI_SVPRT_RUNTIME_DeInit error");
SAMPLE_FreeMem(&(astWkInfo[0].stWKMemory));
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Total");
FAIL_1:
return s32Ret;
}
sample_data_utils.c
/*
* Copyright (C) Hisilicon Technologies Co., Ltd. 2018-2019. All rights reserved.
* Description:
* Author:
* Create: 2018-05-19
*/
#include "sample_data_utils.h"
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>
#ifdef _WIN32
#include <Windows.h>
#endif
#include "math.h"
#include "sample_log.h"
#include "sample_memory_ops.h"
#include "sample_save_blob.h"
HI_S32 InitBlobs(HI_RUNTIME_GROUP_BLOB_ARRAY_S *pstGroupBlob, HI_U32 u32BlobNum, BlobInfo *pBlobInfos)
{
pstGroupBlob->u32BlobNum = u32BlobNum;
HI_RUNTIME_GROUP_BLOB_S *astGroupBlobs = (HI_RUNTIME_GROUP_BLOB_S *)malloc(
sizeof(HI_RUNTIME_GROUP_BLOB_S) * u32BlobNum);
SAMPLE_CHK_RET(astGroupBlobs == HI_NULL, HI_FAILURE, "malloc src group blobs fail");
pstGroupBlob->pstBlobs = astGroupBlobs;
HI_RUNTIME_BLOB_S *astBlobs = (HI_RUNTIME_BLOB_S *)malloc(sizeof(HI_RUNTIME_BLOB_S) * u32BlobNum);
SAMPLE_CHK_GOTO(astBlobs == HI_NULL, FREE_SRC_GRROUP_BLOBS, "malloc src blobs fail");
memset(astBlobs, 0, sizeof(HI_RUNTIME_BLOB_S) * u32BlobNum);
for (HI_U32 i = 0; i < u32BlobNum; i++) {
strncpy(astGroupBlobs[i].acOwnerName, pBlobInfos[i].acOwnerName, MAX_NAME_LEN);
astGroupBlobs[i].acOwnerName[MAX_NAME_LEN] = 0;
strncpy(astGroupBlobs[i].acBlobName, pBlobInfos[i].acBlobName, MAX_NAME_LEN);
astGroupBlobs[i].acBlobName[MAX_NAME_LEN] = 0;
astGroupBlobs[i].pstBlob = &astBlobs[i];
HI_S32 s32Ret = SAMPLE_RUNTIME_SetBlob(&astBlobs[i], pBlobInfos[i].enBlobType,
&pBlobInfos[i].stShape,
pBlobInfos[i].u32Align);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FREE_SRC_BLOBS_CONTENT,
"SAMPLE_RUNTIME_SetBlob alexnet failed!");
if (pBlobInfos[i].bIsSrc) {
s32Ret = SAMPLE_RUNTIME_ReadSrcFile(pBlobInfos[i].acSrcFilePath, &astBlobs[i]);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FREE_SRC_BLOBS_CONTENT,
"SAMPLE_RUNTIME_ReadFile failed!");
} else {
HI_U32 u32BlobSize = astBlobs[i].u32Stride * astBlobs[i].unShape.stWhc.u32Height *
astBlobs[i].unShape.stWhc.u32Chn *
astBlobs[i].u32Num;
memset ((HI_VOID *)(uintptr_t)(astBlobs[i].u64VirAddr), 0, u32BlobSize);
}
}
return HI_SUCCESS;
FREE_SRC_BLOBS_CONTENT:
for (HI_U32 i = 0; i < u32BlobNum; i++) {
if (astBlobs[i].u64VirAddr != 0) {
HI_RUNTIME_MEM_S stMem = {0};
stMem.u64PhyAddr = astBlobs[i].u64PhyAddr;
stMem.u64VirAddr = astBlobs[i].u64VirAddr;
SAMPLE_FreeMem(&stMem);
}
}
SAMPLE_FREE(astBlobs);
FREE_SRC_GRROUP_BLOBS:
SAMPLE_FREE(astGroupBlobs);
pstGroupBlob->u32BlobNum = 0;
pstGroupBlob->pstBlobs = HI_NULL;
return HI_FAILURE;
}
HI_VOID DeinitBlobs(HI_RUNTIME_GROUP_BLOB_ARRAY_S *pstGroupBlob)
{
if (pstGroupBlob == HI_NULL || pstGroupBlob->pstBlobs == HI_NULL) {
return;
}
HI_RUNTIME_BLOB_S *pstBlobs = pstGroupBlob->pstBlobs[0].pstBlob;
for (HI_U32 i = 0; i < pstGroupBlob->u32BlobNum; i++) {
if (pstBlobs[i].u64VirAddr != 0) {
HI_RUNTIME_MEM_S stMem = {0};
stMem.u64PhyAddr = pstBlobs[i].u64PhyAddr;
stMem.u64VirAddr = pstBlobs[i].u64VirAddr;
SAMPLE_FreeMem(&stMem);
}
}
SAMPLE_FREE(pstBlobs);
SAMPLE_FREE(pstGroupBlob->pstBlobs);
}
HI_U32 getAlignSize(HI_U32 align, HI_U32 num)
{
if (align == 0) {
align = 1;
SAMPLE_LOG_INFO("getAlignSize input Align == 0, set it to 1 to prevent zero divide");
}
return ((num + align - 1) / align * align);
}
HI_VOID SAMPLE_DATA_GetStride(HI_RUNTIME_BLOB_TYPE_E type, HI_U32 width, HI_U32 align, HI_U32 *pStride)
{
HI_U32 u32Size = 0;
if ((HI_RUNTIME_BLOB_TYPE_U8 == type) ||
(HI_RUNTIME_BLOB_TYPE_YVU420SP == type) ||
(type == HI_RUNTIME_BLOB_TYPE_YVU422SP)) {
u32Size = sizeof(HI_U8);
} else {
u32Size = sizeof(HI_U32);
}
*pStride = getAlignSize(align, width * u32Size);
return;
}
HI_U32 SAMPLE_DATA_GetBlobSize(HI_U32 stride, HI_U32 num, HI_U32 height, HI_U32 chn)
{
return num * stride * height * chn;
}
HI_VOID printDebugData(const HI_CHAR *pcName, HI_U64 u64VirAddr, HI_U32 u32PrintLine)
{
HI_U8 *pu8Tmp = HI_NULL;
SAMPLE_LOG_INFO("\n%s result print start", pcName);
pu8Tmp = (HI_U8 *)((uintptr_t)(u64VirAddr));
for (HI_U32 i = 0; i < u32PrintLine; i++) {
for (HI_U32 j = 0; j < 16; j++) { // Output 15 detection frames
printf("%02x ", pu8Tmp[i * 16 + j]);
}
SAMPLE_LOG_INFO("\n");
}
SAMPLE_LOG_INFO("%s result print end\n", pcName);
}
HI_S32 SAMPLE_RUNTIME_HiMemAlloc(HI_RUNTIME_MEM_S *pstMem, HI_BOOL bCached)
{
pstMem->u64VirAddr = 0;
HI_S32 s32Ret = SAMPLE_AllocMem(pstMem, bCached);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_AllocMem failed!");
if (bCached == HI_FALSE) {
return s32Ret;
}
s32Ret = SAMPLE_FlushCache(pstMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return s32Ret;
FAIL_0:
SAMPLE_FreeMem(pstMem);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_HiBlobAlloc(HI_RUNTIME_BLOB_S *pstBlob, HI_U32 u32BlobSize, HI_BOOL bCached)
{
HI_RUNTIME_MEM_S stMem;
memset(&stMem, 0, sizeof(stMem));
stMem.u32Size = u32BlobSize;
HI_S32 s32Ret = SAMPLE_AllocMem(&stMem, bCached);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_AllocMem failed!");
pstBlob->u64PhyAddr = stMem.u64PhyAddr;
pstBlob->u64VirAddr = stMem.u64VirAddr;
if (bCached == HI_FALSE) {
return s32Ret;
}
memset ((HI_VOID *)((uintptr_t)pstBlob->u64VirAddr), 0, u32BlobSize);
s32Ret = SAMPLE_FlushCache(&stMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return s32Ret;
FAIL_0:
(HI_VOID)SAMPLE_FreeMem(&stMem);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_LoadModelFile(const HI_CHAR *pcModelFile, HI_RUNTIME_MEM_S *pstMemInfo)
{
FILE *fp = HI_NULL;
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(pcModelFile) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, pcModelFile,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", pcModelFile);
#else
SAMPLE_CHK_RET((strlen(pcModelFile) > PATH_MAX) ||
(HI_NULL == realpath(pcModelFile, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", pcModelFile);
#endif
fp = fopen(acCanonicalPath, "rb");
SAMPLE_CHK_RET(fp == NULL, HI_FAILURE, "Open model file %s failed!", pcModelFile);
HI_S32 s32Ret = fseek(fp, 0L, SEEK_END);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_MemAlloc failed!");
HI_U32 s32RuntimeWkLen = ftell(fp);
// Divide by 16 to verify that the length of WK is correct
SAMPLE_CHK_GOTO(0 != s32RuntimeWkLen % 16, CLOSE_FILE, "Runtime WK Len %% 16 != 0 ");
SAMPLE_LOG_INFO("Runtime WK Length: %d", s32RuntimeWkLen);
SAMPLE_CHK_GOTO(0 != fseek(fp, 0L, SEEK_SET), CLOSE_FILE, "fseek fail");
pstMemInfo->u32Size = s32RuntimeWkLen;
s32Ret = SAMPLE_RUNTIME_HiMemAlloc(pstMemInfo, HI_FALSE);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_MemAlloc failed!");
s32Ret = (HI_S32)fread((HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr), s32RuntimeWkLen, 1, fp);
SAMPLE_CHK_GOTO(s32Ret != 1, FREE_MEM, "Read runtime WK failed!");
fclose(fp);
return HI_SUCCESS;
FREE_MEM:
SAMPLE_FreeMem(pstMemInfo);
CLOSE_FILE:
fclose(fp);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_SetBlob(HI_RUNTIME_BLOB_S *pstBlob,
HI_RUNTIME_BLOB_TYPE_E enType,
BlobShape *pstBlobShape,
HI_U32 u32Align)
{
pstBlob->enBlobType = enType;
pstBlob->u32Num = pstBlobShape->u32Num;
if (enType == HI_RUNTIME_BLOB_TYPE_VEC_S32) {
pstBlob->unShape.stWhc.u32Width = pstBlobShape->u32Chn;
pstBlob->unShape.stWhc.u32Chn = pstBlobShape->u32Width;
} else {
pstBlob->unShape.stWhc.u32Width = pstBlobShape->u32Width;
pstBlob->unShape.stWhc.u32Chn = pstBlobShape->u32Chn;
}
pstBlob->unShape.stWhc.u32Height = pstBlobShape->u32Height;
SAMPLE_DATA_GetStride(enType, pstBlob->unShape.stWhc.u32Width, u32Align, &(pstBlob->u32Stride));
HI_U64 u64Size = pstBlob->u32Num * pstBlob->u32Stride * pstBlob->unShape.stWhc.u32Height *
pstBlob->unShape.stWhc.u32Chn;
SAMPLE_CHK_RET((u64Size > (HI_U32)-1), HI_FAILURE, "the blobsize is too large [%llu]", u64Size);
HI_U32 u32BlobSize = SAMPLE_DATA_GetBlobSize(pstBlob->u32Stride, pstBlobShape->u32Num, pstBlobShape->u32Height,
pstBlob->unShape.stWhc.u32Chn);
HI_S32 s32Ret = SAMPLE_RUNTIME_HiBlobAlloc(pstBlob, u32BlobSize, HI_TRUE);
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_Utils_AllocMem failed!");
return HI_SUCCESS;
}
HI_S32 SAMPLE_RUNTIME_ReadSrcFile(const HI_CHAR acFileName[MAX_FILE_NAME_LENGTH],
HI_RUNTIME_BLOB_S *pstSrcBlob)
{
HI_U32 c, h;
HI_U8 *pu8Ptr = NULL;
FILE *imgFp = NULL;
HI_U32 s32Ret = HI_FAILURE;
HI_RUNTIME_MEM_S stMem;
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(acFileName) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, acFileName,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", acFileName);
#else
SAMPLE_CHK_RET((strlen(acFileName) > PATH_MAX) ||
(HI_NULL == realpath(acFileName, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", acFileName);
#endif
pu8Ptr = (HI_U8 *)((uintptr_t)(pstSrcBlob->u64VirAddr));
imgFp = fopen(acCanonicalPath, "rb");
SAMPLE_CHK_GOTO(imgFp == NULL, FAIL_0, "open img fp error[%s]", acFileName);
for (c = 0; c < pstSrcBlob->unShape.stWhc.u32Chn; c++) {
for (h = 0; h < pstSrcBlob->unShape.stWhc.u32Height; h++) {
s32Ret = (HI_S32)fread(pu8Ptr, pstSrcBlob->unShape.stWhc.u32Width * sizeof(HI_U8), 1, imgFp);
SAMPLE_CHK_GOTO(s32Ret != 1, FAIL_0, "fread failed, (c,h)=(%d,%d)!", c, h);
pu8Ptr += pstSrcBlob->u32Stride;
}
}
fclose(imgFp);
imgFp = HI_NULL;
HI_U32 u32BlobSize = SAMPLE_DATA_GetBlobSize(pstSrcBlob->u32Stride, pstSrcBlob->u32Num,
pstSrcBlob->unShape.stWhc.u32Height,
pstSrcBlob->unShape.stWhc.u32Chn);
stMem.u64PhyAddr = pstSrcBlob->u64PhyAddr;
stMem.u64VirAddr = pstSrcBlob->u64VirAddr;
stMem.u32Size = u32BlobSize;
s32Ret = SAMPLE_FlushCache(&stMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return HI_SUCCESS;
FAIL_0:
if (imgFp != HI_NULL) {
fclose(imgFp);
}
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_ReadConfig(const HI_CHAR *pcConfigFile, HI_CHAR **acBuff)
{
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
FILE *f = NULL;
HI_CHAR *tempBuff = NULL;
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(pcConfigFile) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, pcConfigFile,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", pcConfigFile);
#else
SAMPLE_CHK_RET((strlen(pcConfigFile) > PATH_MAX) ||
(HI_NULL == realpath(pcConfigFile, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", pcConfigFile);
#endif
f = fopen(acCanonicalPath, "r");
SAMPLE_CHK_RET(f == HI_NULL, HI_FAILURE, "config file %s not exists", acCanonicalPath);
HI_S32 s32Ret = fseek(f, 0L, SEEK_END);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
HI_S64 s64FileSize = ftell(f);
SAMPLE_CHK_GOTO(-1 == s64FileSize, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
tempBuff = (HI_CHAR *)malloc(s64FileSize + 1);
SAMPLE_CHK_GOTO(tempBuff == HI_NULL, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
rewind(f);
size_t readSize = fread(tempBuff, 1, s64FileSize, f);
if (((HI_S64)readSize != s64FileSize) && !feof(f)) {
SAMPLE_FREE(tempBuff);
*acBuff = NULL;
fclose(f);
return HI_FAILURE;
}
tempBuff[readSize] = '[=16=]';
*acBuff = tempBuff;
fclose(f);
return HI_SUCCESS;
CLOSE_FILE:
fclose(f);
return HI_FAILURE;
}
sample_memory_ops.c
/*
* Copyright (C) Hisilicon Technologies Co., Ltd. 2018-2019. All rights reserved.
* Description:
* Author:
* Create: 2018-05-19
*/
#include "sample_memory_ops.h"
#include <stdlib.h>
#ifdef ON_BOARD
#include "mpi_sys.h"
#endif
#include "sample_log.h"
HI_S32 SAMPLE_AllocMem(HI_RUNTIME_MEM_S *pstMemInfo, HI_BOOL bCached)
{
// SAMPLE_CHK_RET((pstMemInfo == HI_NULL), HI_FAILURE, "param pstMemInfo is NULL");
// SAMPLE_CHK_RET((HI_NULL == pstMemInfo->u64PhyAddr), HI_FAILURE, "param u64PhyAddr is 0");
#ifdef ON_BOARD
HI_S32 s32Ret = HI_SUCCESS;
if (bCached) {
s32Ret = HI_MPI_SYS_MmzAlloc_Cached(&pstMemInfo->u64PhyAddr,
(HI_VOID **)&(pstMemInfo->u64VirAddr), NULL, HI_NULL,
pstMemInfo->u32Size);
} else {
s32Ret = HI_MPI_SYS_MmzAlloc(&pstMemInfo->u64PhyAddr, (HI_VOID **)&(pstMemInfo->u64VirAddr), NULL, HI_NULL,
pstMemInfo->u32Size);
}
#else
pstMemInfo->u64VirAddr = (HI_U32)((uintptr_t)malloc(pstMemInfo->u32Size));
HI_S32 s32Ret = pstMemInfo->u64VirAddr == 0 ? HI_FAILURE : HI_SUCCESS;
pstMemInfo->u64PhyAddr = pstMemInfo->u64VirAddr;
#endif
return s32Ret;
}
HI_S32 SAMPLE_FlushCache(HI_RUNTIME_MEM_S *pstMemInfo)
{
SAMPLE_CHK_RET((pstMemInfo->u64VirAddr == 0), HI_FAILURE, "param pu8VirAddr is NULL");
#ifdef ON_BOARD
return HI_MPI_SYS_MmzFlushCache(pstMemInfo->u64PhyAddr, (HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr),
pstMemInfo->u32Size);
#endif
return HI_SUCCESS;
}
HI_S32 SAMPLE_FreeMem(HI_RUNTIME_MEM_S *pstMemInfo)
{
SAMPLE_CHK_RET((pstMemInfo->u64VirAddr == 0), HI_FAILURE, "param pu8VirAddr is NULL");
#ifdef ON_BOARD
return HI_MPI_SYS_MmzFree (pstMemInfo->u64PhyAddr, (HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr));
#else
free ((HI_U8 *)((uintptr_t) pstMemInfo->u64VirAddr));
pstMemInfo->u64PhyAddr = 0;
pstMemInfo->u64VirAddr = 0;
#endif
return HI_SUCCESS;
}
WK 文件
有 50 个输出：https://drive.google.com/file/d/1chePxF3j3KF4KvFvhRGGUKQz7ZDipeBx/view?usp=sharing
有 5 个输出：https://drive.google.com/file/d/1chePxF3j3KF4KvFvhRGGUKQz7ZDipeBx/view?usp=sharing
问题
这是什么错误？我在哪里分配 RAM 失败？这是模拟器吗error/bug?
创建的 .wk 使用 NNIE_Mapper 生成的格式不可用于运行时模拟。
在RuyiStudio中创建类型为"Runtime"的项目，或在Linux下使用runtime_mapper生成有效的.wk并解决问题！[=13=]
希望错误消息更好，但是哦，好吧！
感谢 Telegram 上的 OpenPIC 社区！
首先，一些背景：
有 2 个 ML 模型：
- 有 50 个输出节点
- 有 5 个输出节点
两者都首先使用 https://github.com/dwayo-gh/AER-CNN-KERAS 进行训练，然后使用 https://github.com/uhfband/keras2caffe.git.
具有 50 个输出节点的模型的故事：
在 Windows10 环境下使用 RuyiStudio，我将 prototxt + 权重转换为 .wk 文件。生成的 .wk 文件的大小为 594112 字节。
然后我将 HiSVP SDK 的 "sample_runtime" 示例修改为 运行 这个新的 .wk 文件（来源如下）。功能模拟效果很好！它给出了输入文件的前 5 个推理结果。
具有 5 个输出节点的模型的故事：
再次使用RuyiStudio，在Win10环境下，我将文件转换为.wk，更改了运行时间源中的路径。生成的 .wk 文件的大小为 329776 字节。以下是我遇到的错误：
Thu Mar 12 11:08:01 2020 I Try to open [nnie_sim.ini] in execute workspace...
Thu Mar 12 11:08:01 2020 W open [nnie_sim.ini] in execute workspace failure
Thu Mar 12 11:08:01 2020 I Try to open [nnie_sim.ini] in Env NNIE_SIM_INI_DIR...
Thu Mar 12 11:08:01 2020 I getenv(NNIE_SIM_INI_DIR): C:\Users\admin\Downloads\HiSVP_PC_V1.2.2.2\software\hisilicon ...
Thu Mar 12 11:08:01 2020 I open [nnie_sim.ini] success(path: C:/Users/admin/Downloads/HiSVP_PC_V1.2.2.2/software/hisilicon/nnie_sim.ini)
Thu Mar 12 11:08:01 2020 I ============= nnie_sim.ini setting ==============
Thu Mar 12 11:08:01 2020 I [LOG_LEVEL] DBG
Thu Mar 12 11:08:01 2020 I =================================================
====================start to run====================
=======================File: ..\Release.wav.json.bin==========================
I[HiRT] Runtime version: 1.2.2.7
build sha1: 580506f9
compile date: Aug 9 2019 17:24:44
Runtime WK Length: 329776
I[hirt] Model size: 1 Connector size: 0
E[hirt] invalid wk file, total u32Size(329776) should greater than nnie wk u32Size(329776)
E[hirt] parseNnieWK error
E[hirt] model is loading
E[hirt] addModel error
E[hirt] parseModelAndConnector error
E[hirt] load group in addGroup by sync error
E[hirt] AddGroup fail
HI_SVPRT_RUNTIME_LoadModelGroup error
SAMPLE_RUNTIME_LoadModelGroup_Aernet failed!
[Total], TIME SPEND: 1503ms
==========run success==========
main.c
#include <stdio.h>
#include <stdlib.h>
#include "string.h"
#ifdef ON_BOARD
#include "mpi_sys.h"
#include "mpi_vb.h"
#else
#include "hi_comm_svp.h"
#include "hi_nnie.h"
#include "mpi_nnie.h"
#include "hi_type.h"
#endif
#include "math.h"
#include "sample_log.h"
#include "sample_save_blob.h"
#include "sample_data_utils.h"
#include "sample_runtime_classify.h"
static HI_VOID SAMPLE_RUNTIME_Usage(const char *pchPrgName)
{
SAMPLE_LOG_INFO("Usage : %s <filepath>", pchPrgName);
}
#ifdef ON_BOARD
static HI_S32 SAMPLE_COMM_SVP_SysInit(HI_VOID)
{
HI_MPI_SYS_Exit();
HI_MPI_VB_Exit();
VB_CONFIG_S struVbConf = {0};
struVbConf.u32MaxPoolCnt = 2;
struVbConf.astCommPool[1].u64BlkSize = 768 * 576 * 2;
struVbConf.astCommPool[1].u32BlkCnt = 1;
HI_S32 s32Ret = HI_MPI_VB_SetConfig((const VB_CONFIG_S *)&struVbConf);
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_SetConfig error");
return HI_FAILURE;
}
s32Ret = HI_MPI_VB_Init();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_Init error");
return HI_FAILURE;
}
s32Ret = HI_MPI_SYS_Init();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_SYS_Init error");
return HI_FAILURE;
}
return s32Ret;
}
static HI_S32 SAMPLE_COMM_SVP_SysExit(HI_VOID)
{
HI_S32 s32Ret = HI_MPI_SYS_Exit();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_SYS_Exit error");
return HI_FAILURE;
}
s32Ret = HI_MPI_VB_Exit();
if (s32Ret != HI_SUCCESS) {
SAMPLE_LOG_INFO("HI_MPI_VB_Exit error");
return HI_FAILURE;
}
return HI_SUCCESS;
}
#endif
HI_S32 HandleSampleNet(const HI_CHAR* filename)
{
HI_S32 ret = HI_FAILURE;
SAMPLE_LOG_INFO("\n\n=======================File: %s==========================\n\n", filename);
ret = SAMPLE_AerNet(filename);
return ret;
}
HI_S32 main(HI_S32 argc, HI_CHAR* argv[])
{
if ((argc != 2)) {
SAMPLE_RUNTIME_Usage(argv[0]);
return HI_FAILURE;
}
if (!strncmp(argv[1], "-h", 2)) {
SAMPLE_RUNTIME_Usage(argv[0]);
return HI_SUCCESS;
}
const HI_CHAR *pcSrcFile = argv[1];
SAMPLE_LOG_INFO("\n\n====================start to run====================\n");
#ifdef ON_BOARD
HI_S32 s32Ret = SAMPLE_COMM_SVP_SysInit();
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_COMM_SVP_SysInit fail");
s32Ret = HandleSampleNet(pcSrcFile);
#else
HI_S32 s32Ret = HandleSampleNet(pcSrcFile);
#endif
SAMPLE_CHK_PRINTF((s32Ret == HI_SUCCESS), "\n==========run success==========\n");
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "\n==========run fail==========\n");
#ifdef ON_BOARD
s32Ret = SAMPLE_COMM_SVP_SysExit();
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_COMM_SVP_SysExit fail");
#endif
return s32Ret;
}
aernet.modelgroup
name: "single_aernet"
priority: 0
max_tmpbuf_size_mb: 4096
align_size: 32
input {
name: "data"
}
model {
name: "aernet"
bottom: {name: "data"}
top: {name: "dense_2"}
}
sample_runtime_classify.c
#include "sample_runtime_classify.h"
#include <stdio.h>
#include <stdlib.h>
#ifdef ON_BOARD
#include "mpi_sys.h"
#include "mpi_vb.h"
#else
#include "hi_comm_svp.h"
#include "hi_nnie.h"
#include "mpi_nnie.h"
#endif
#include "hi_runtime_api.h"
#include "string.h"
#include "sample_log.h"
#include "sample_runtime_define.h"
#include "sample_memory_ops.h"
#include "sample_data_utils.h"
#include "sample_model_classification.h"
static const HI_U32 ALEXNET_OUTPUT_TOP_FIVE = 5;
const HI_U32 ALEXNET_DEBUG_PRINT_NUM = 10;
static const HI_U32 AERNET_OUTPUT_TOP_FIVE = 5;
const HI_U32 AERNET_DEBUG_PRINT_NUM = 10;
static BlobInfo aer_stSrcBlobInfo = {
"", "data", "",
HI_RUNTIME_BLOB_TYPE_S32,
{ 1, 1, 128, 300 },
HI_TRUE, ALIGN_16
};
static BlobInfo aer_stDstBlobInfo = {
"aernet", "dense_2", "",
HI_RUNTIME_BLOB_TYPE_VEC_S32,
{ 1, 50, 1, 1 },
HI_FALSE, ALIGN_16
};
HI_S32 SAMPLE_RUNTIME_LoadModelGroup_Aernet(const HI_CHAR *pcModelFileAer, HI_RUNTIME_WK_INFO_S *pstWkInfo,
HI_RUNTIME_GROUP_HANDLE *phGroupHandle)
{
HI_CHAR *pacConfig = NULL;
HI_RUNTIME_GROUP_INFO_S stGroupInfo = {0};
memset(&stGroupInfo, 0, sizeof(HI_RUNTIME_GROUP_INFO_S));
strncpy(pstWkInfo[0].acModelName, "aernet", MAX_NAME_LEN);
HI_S32 s32Ret = SAMPLE_RUNTIME_LoadModelFile(pcModelFileAer, &(pstWkInfo[0].stWKMemory));
SAMPLE_CHK_GOTO((s32Ret != HI_SUCCESS), FAIL, "SAMPLE_RUNTIME_LoadModelFile %s failed!", pcModelFileAer);
stGroupInfo.stWKsInfo.u32WKNum = 1;
stGroupInfo.stWKsInfo.pstAttrs = &(pstWkInfo[0]);
SAMPLE_RUNTIME_ReadConfig(CONFIG_DIR "aernet.modelgroup", &pacConfig);
SAMPLE_CHK_GOTO((pacConfig == NULL), FAIL, "HI_SVPRT_RUNTIME_ReadConfig error");
s32Ret = HI_SVPRT_RUNTIME_LoadModelGroup(pacConfig, &stGroupInfo, phGroupHandle);
SAMPLE_CHK_GOTO((s32Ret != HI_SUCCESS), FAIL, "HI_SVPRT_RUNTIME_LoadModelGroup error");
SAMPLE_LOG_INFO("LoadGroup succ, group handle[%p]", phGroupHandle);
SAMPLE_FREE(pacConfig);
return HI_SUCCESS;
FAIL:
SAMPLE_FREE(pacConfig);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_ForwardGroup_Aernet(const HI_CHAR *pcSrcFile, HI_RUNTIME_GROUP_HANDLE hGroupHandle)
{
HI_RUNTIME_GROUP_SRC_BLOB_ARRAY_S stGroupSrcBlob = {0};
HI_RUNTIME_GROUP_DST_BLOB_ARRAY_S stGroupDstBlob = {0};
strncpy(aer_stSrcBlobInfo.acSrcFilePath, pcSrcFile, sizeof(aer_stSrcBlobInfo.acSrcFilePath));
aer_stSrcBlobInfo.acSrcFilePath[sizeof(aer_stSrcBlobInfo.acSrcFilePath) - 1] = '[=15=]';
HI_S32 s32Ret = InitBlobs(&stGroupSrcBlob, 1, &aer_stSrcBlobInfo);
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "InitBlobs fail");
s32Ret = InitBlobs(&stGroupDstBlob, 1, &aer_stDstBlobInfo);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "InitBlobs fail");
#if PERFORMANCE_TEST
struct timespec start, end;
clock_gettime(0, &start);
#endif
s32Ret = HI_SVPRT_RUNTIME_ForwardGroupSync(hGroupHandle, &stGroupSrcBlob, &stGroupDstBlob, 0);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "HI_SVPRT_RUNTIME_ForwardGroupSync failed!");
#if PERFORMANCE_TEST
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Forward");
#endif
SAMPLE_LOG_INFO("Pic: %s", pcSrcFile);
s32Ret = SAMPLE_RUNTIME_Cnn_TopN_Output(stGroupDstBlob.pstBlobs[0].pstBlob, AERNET_OUTPUT_TOP_FIVE);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "SAMPLE_RUNTIME_Cnn_TopN_Output error");
#if DEBUG
printDebugData((HI_CHAR *)"alexnet", stDst[0].u64VirAddr, ALEXNET_DEBUG_PRINT_NUM);
s32Ret = HI_SUCCESS;
#endif
FAIL_0:
DeinitBlobs(&stGroupDstBlob);
DeinitBlobs(&stGroupSrcBlob);
return s32Ret;
}
HI_S32 SAMPLE_AerNet(HI_CHAR *pcSrcFile)
{
const HI_CHAR *pcRuntimeModelName = MODEL_AERNET_NAME;
HI_RUNTIME_GROUP_HANDLE hGroupHandle = HI_NULL;
HI_RUNTIME_WK_INFO_S astWkInfo[1];
memset(&astWkInfo[0], 0, sizeof(astWkInfo));
struct timespec start, next, end;
clock_gettime(0, &start);
HI_S32 s32Ret = HI_SVPRT_RUNTIME_Init(CPU_TASK_AFFINITY, NULL);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_1, "HI_SVPRT_RUNTIME_Init failed!");
s32Ret = SAMPLE_RUNTIME_LoadModelGroup_Aernet(pcRuntimeModelName, astWkInfo, &hGroupHandle);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_RUNTIME_LoadModelGroup_Aernet failed!");
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Load");
s32Ret = SAMPLE_RUNTIME_ForwardGroup_Aernet(pcSrcFile, hGroupHandle);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "SAMPLE_RUNTIME_ForwardGroup_Alexnet error");
clock_gettime(0, &next);
timeSpendMs(&end, &next, "Forward total");
s32Ret = HI_SVPRT_RUNTIME_UnloadModelGroup(hGroupHandle);
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "HI_SVPRT_RUNTIME_UnloadModelGroup error");
FAIL_0:
s32Ret = HI_SVPRT_RUNTIME_DeInit();
SAMPLE_CHK_PRINTF((s32Ret != HI_SUCCESS), "HI_SVPRT_RUNTIME_DeInit error");
SAMPLE_FreeMem(&(astWkInfo[0].stWKMemory));
clock_gettime(0, &end);
timeSpendMs(&start, &end, "Total");
FAIL_1:
return s32Ret;
}
sample_data_utils.c
/*
* Copyright (C) Hisilicon Technologies Co., Ltd. 2018-2019. All rights reserved.
* Description:
* Author:
* Create: 2018-05-19
*/
#include "sample_data_utils.h"
#include <stdio.h>
#include <stdlib.h>
#include <errno.h>
#include <string.h>
#ifdef _WIN32
#include <Windows.h>
#endif
#include "math.h"
#include "sample_log.h"
#include "sample_memory_ops.h"
#include "sample_save_blob.h"
HI_S32 InitBlobs(HI_RUNTIME_GROUP_BLOB_ARRAY_S *pstGroupBlob, HI_U32 u32BlobNum, BlobInfo *pBlobInfos)
{
pstGroupBlob->u32BlobNum = u32BlobNum;
HI_RUNTIME_GROUP_BLOB_S *astGroupBlobs = (HI_RUNTIME_GROUP_BLOB_S *)malloc(
sizeof(HI_RUNTIME_GROUP_BLOB_S) * u32BlobNum);
SAMPLE_CHK_RET(astGroupBlobs == HI_NULL, HI_FAILURE, "malloc src group blobs fail");
pstGroupBlob->pstBlobs = astGroupBlobs;
HI_RUNTIME_BLOB_S *astBlobs = (HI_RUNTIME_BLOB_S *)malloc(sizeof(HI_RUNTIME_BLOB_S) * u32BlobNum);
SAMPLE_CHK_GOTO(astBlobs == HI_NULL, FREE_SRC_GRROUP_BLOBS, "malloc src blobs fail");
memset(astBlobs, 0, sizeof(HI_RUNTIME_BLOB_S) * u32BlobNum);
for (HI_U32 i = 0; i < u32BlobNum; i++) {
strncpy(astGroupBlobs[i].acOwnerName, pBlobInfos[i].acOwnerName, MAX_NAME_LEN);
astGroupBlobs[i].acOwnerName[MAX_NAME_LEN] = 0;
strncpy(astGroupBlobs[i].acBlobName, pBlobInfos[i].acBlobName, MAX_NAME_LEN);
astGroupBlobs[i].acBlobName[MAX_NAME_LEN] = 0;
astGroupBlobs[i].pstBlob = &astBlobs[i];
HI_S32 s32Ret = SAMPLE_RUNTIME_SetBlob(&astBlobs[i], pBlobInfos[i].enBlobType,
&pBlobInfos[i].stShape,
pBlobInfos[i].u32Align);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FREE_SRC_BLOBS_CONTENT,
"SAMPLE_RUNTIME_SetBlob alexnet failed!");
if (pBlobInfos[i].bIsSrc) {
s32Ret = SAMPLE_RUNTIME_ReadSrcFile(pBlobInfos[i].acSrcFilePath, &astBlobs[i]);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FREE_SRC_BLOBS_CONTENT,
"SAMPLE_RUNTIME_ReadFile failed!");
} else {
HI_U32 u32BlobSize = astBlobs[i].u32Stride * astBlobs[i].unShape.stWhc.u32Height *
astBlobs[i].unShape.stWhc.u32Chn *
astBlobs[i].u32Num;
memset ((HI_VOID *)(uintptr_t)(astBlobs[i].u64VirAddr), 0, u32BlobSize);
}
}
return HI_SUCCESS;
FREE_SRC_BLOBS_CONTENT:
for (HI_U32 i = 0; i < u32BlobNum; i++) {
if (astBlobs[i].u64VirAddr != 0) {
HI_RUNTIME_MEM_S stMem = {0};
stMem.u64PhyAddr = astBlobs[i].u64PhyAddr;
stMem.u64VirAddr = astBlobs[i].u64VirAddr;
SAMPLE_FreeMem(&stMem);
}
}
SAMPLE_FREE(astBlobs);
FREE_SRC_GRROUP_BLOBS:
SAMPLE_FREE(astGroupBlobs);
pstGroupBlob->u32BlobNum = 0;
pstGroupBlob->pstBlobs = HI_NULL;
return HI_FAILURE;
}
HI_VOID DeinitBlobs(HI_RUNTIME_GROUP_BLOB_ARRAY_S *pstGroupBlob)
{
if (pstGroupBlob == HI_NULL || pstGroupBlob->pstBlobs == HI_NULL) {
return;
}
HI_RUNTIME_BLOB_S *pstBlobs = pstGroupBlob->pstBlobs[0].pstBlob;
for (HI_U32 i = 0; i < pstGroupBlob->u32BlobNum; i++) {
if (pstBlobs[i].u64VirAddr != 0) {
HI_RUNTIME_MEM_S stMem = {0};
stMem.u64PhyAddr = pstBlobs[i].u64PhyAddr;
stMem.u64VirAddr = pstBlobs[i].u64VirAddr;
SAMPLE_FreeMem(&stMem);
}
}
SAMPLE_FREE(pstBlobs);
SAMPLE_FREE(pstGroupBlob->pstBlobs);
}
HI_U32 getAlignSize(HI_U32 align, HI_U32 num)
{
if (align == 0) {
align = 1;
SAMPLE_LOG_INFO("getAlignSize input Align == 0, set it to 1 to prevent zero divide");
}
return ((num + align - 1) / align * align);
}
HI_VOID SAMPLE_DATA_GetStride(HI_RUNTIME_BLOB_TYPE_E type, HI_U32 width, HI_U32 align, HI_U32 *pStride)
{
HI_U32 u32Size = 0;
if ((HI_RUNTIME_BLOB_TYPE_U8 == type) ||
(HI_RUNTIME_BLOB_TYPE_YVU420SP == type) ||
(type == HI_RUNTIME_BLOB_TYPE_YVU422SP)) {
u32Size = sizeof(HI_U8);
} else {
u32Size = sizeof(HI_U32);
}
*pStride = getAlignSize(align, width * u32Size);
return;
}
HI_U32 SAMPLE_DATA_GetBlobSize(HI_U32 stride, HI_U32 num, HI_U32 height, HI_U32 chn)
{
return num * stride * height * chn;
}
HI_VOID printDebugData(const HI_CHAR *pcName, HI_U64 u64VirAddr, HI_U32 u32PrintLine)
{
HI_U8 *pu8Tmp = HI_NULL;
SAMPLE_LOG_INFO("\n%s result print start", pcName);
pu8Tmp = (HI_U8 *)((uintptr_t)(u64VirAddr));
for (HI_U32 i = 0; i < u32PrintLine; i++) {
for (HI_U32 j = 0; j < 16; j++) { // Output 15 detection frames
printf("%02x ", pu8Tmp[i * 16 + j]);
}
SAMPLE_LOG_INFO("\n");
}
SAMPLE_LOG_INFO("%s result print end\n", pcName);
}
HI_S32 SAMPLE_RUNTIME_HiMemAlloc(HI_RUNTIME_MEM_S *pstMem, HI_BOOL bCached)
{
pstMem->u64VirAddr = 0;
HI_S32 s32Ret = SAMPLE_AllocMem(pstMem, bCached);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_AllocMem failed!");
if (bCached == HI_FALSE) {
return s32Ret;
}
s32Ret = SAMPLE_FlushCache(pstMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return s32Ret;
FAIL_0:
SAMPLE_FreeMem(pstMem);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_HiBlobAlloc(HI_RUNTIME_BLOB_S *pstBlob, HI_U32 u32BlobSize, HI_BOOL bCached)
{
HI_RUNTIME_MEM_S stMem;
memset(&stMem, 0, sizeof(stMem));
stMem.u32Size = u32BlobSize;
HI_S32 s32Ret = SAMPLE_AllocMem(&stMem, bCached);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_AllocMem failed!");
pstBlob->u64PhyAddr = stMem.u64PhyAddr;
pstBlob->u64VirAddr = stMem.u64VirAddr;
if (bCached == HI_FALSE) {
return s32Ret;
}
memset ((HI_VOID *)((uintptr_t)pstBlob->u64VirAddr), 0, u32BlobSize);
s32Ret = SAMPLE_FlushCache(&stMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return s32Ret;
FAIL_0:
(HI_VOID)SAMPLE_FreeMem(&stMem);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_LoadModelFile(const HI_CHAR *pcModelFile, HI_RUNTIME_MEM_S *pstMemInfo)
{
FILE *fp = HI_NULL;
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(pcModelFile) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, pcModelFile,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", pcModelFile);
#else
SAMPLE_CHK_RET((strlen(pcModelFile) > PATH_MAX) ||
(HI_NULL == realpath(pcModelFile, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", pcModelFile);
#endif
fp = fopen(acCanonicalPath, "rb");
SAMPLE_CHK_RET(fp == NULL, HI_FAILURE, "Open model file %s failed!", pcModelFile);
HI_S32 s32Ret = fseek(fp, 0L, SEEK_END);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_MemAlloc failed!");
HI_U32 s32RuntimeWkLen = ftell(fp);
// Divide by 16 to verify that the length of WK is correct
SAMPLE_CHK_GOTO(0 != s32RuntimeWkLen % 16, CLOSE_FILE, "Runtime WK Len %% 16 != 0 ");
SAMPLE_LOG_INFO("Runtime WK Length: %d", s32RuntimeWkLen);
SAMPLE_CHK_GOTO(0 != fseek(fp, 0L, SEEK_SET), CLOSE_FILE, "fseek fail");
pstMemInfo->u32Size = s32RuntimeWkLen;
s32Ret = SAMPLE_RUNTIME_HiMemAlloc(pstMemInfo, HI_FALSE);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_MemAlloc failed!");
s32Ret = (HI_S32)fread((HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr), s32RuntimeWkLen, 1, fp);
SAMPLE_CHK_GOTO(s32Ret != 1, FREE_MEM, "Read runtime WK failed!");
fclose(fp);
return HI_SUCCESS;
FREE_MEM:
SAMPLE_FreeMem(pstMemInfo);
CLOSE_FILE:
fclose(fp);
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_SetBlob(HI_RUNTIME_BLOB_S *pstBlob,
HI_RUNTIME_BLOB_TYPE_E enType,
BlobShape *pstBlobShape,
HI_U32 u32Align)
{
pstBlob->enBlobType = enType;
pstBlob->u32Num = pstBlobShape->u32Num;
if (enType == HI_RUNTIME_BLOB_TYPE_VEC_S32) {
pstBlob->unShape.stWhc.u32Width = pstBlobShape->u32Chn;
pstBlob->unShape.stWhc.u32Chn = pstBlobShape->u32Width;
} else {
pstBlob->unShape.stWhc.u32Width = pstBlobShape->u32Width;
pstBlob->unShape.stWhc.u32Chn = pstBlobShape->u32Chn;
}
pstBlob->unShape.stWhc.u32Height = pstBlobShape->u32Height;
SAMPLE_DATA_GetStride(enType, pstBlob->unShape.stWhc.u32Width, u32Align, &(pstBlob->u32Stride));
HI_U64 u64Size = pstBlob->u32Num * pstBlob->u32Stride * pstBlob->unShape.stWhc.u32Height *
pstBlob->unShape.stWhc.u32Chn;
SAMPLE_CHK_RET((u64Size > (HI_U32)-1), HI_FAILURE, "the blobsize is too large [%llu]", u64Size);
HI_U32 u32BlobSize = SAMPLE_DATA_GetBlobSize(pstBlob->u32Stride, pstBlobShape->u32Num, pstBlobShape->u32Height,
pstBlob->unShape.stWhc.u32Chn);
HI_S32 s32Ret = SAMPLE_RUNTIME_HiBlobAlloc(pstBlob, u32BlobSize, HI_TRUE);
SAMPLE_CHK_RET(s32Ret != HI_SUCCESS, HI_FAILURE, "SAMPLE_Utils_AllocMem failed!");
return HI_SUCCESS;
}
HI_S32 SAMPLE_RUNTIME_ReadSrcFile(const HI_CHAR acFileName[MAX_FILE_NAME_LENGTH],
HI_RUNTIME_BLOB_S *pstSrcBlob)
{
HI_U32 c, h;
HI_U8 *pu8Ptr = NULL;
FILE *imgFp = NULL;
HI_U32 s32Ret = HI_FAILURE;
HI_RUNTIME_MEM_S stMem;
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(acFileName) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, acFileName,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", acFileName);
#else
SAMPLE_CHK_RET((strlen(acFileName) > PATH_MAX) ||
(HI_NULL == realpath(acFileName, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", acFileName);
#endif
pu8Ptr = (HI_U8 *)((uintptr_t)(pstSrcBlob->u64VirAddr));
imgFp = fopen(acCanonicalPath, "rb");
SAMPLE_CHK_GOTO(imgFp == NULL, FAIL_0, "open img fp error[%s]", acFileName);
for (c = 0; c < pstSrcBlob->unShape.stWhc.u32Chn; c++) {
for (h = 0; h < pstSrcBlob->unShape.stWhc.u32Height; h++) {
s32Ret = (HI_S32)fread(pu8Ptr, pstSrcBlob->unShape.stWhc.u32Width * sizeof(HI_U8), 1, imgFp);
SAMPLE_CHK_GOTO(s32Ret != 1, FAIL_0, "fread failed, (c,h)=(%d,%d)!", c, h);
pu8Ptr += pstSrcBlob->u32Stride;
}
}
fclose(imgFp);
imgFp = HI_NULL;
HI_U32 u32BlobSize = SAMPLE_DATA_GetBlobSize(pstSrcBlob->u32Stride, pstSrcBlob->u32Num,
pstSrcBlob->unShape.stWhc.u32Height,
pstSrcBlob->unShape.stWhc.u32Chn);
stMem.u64PhyAddr = pstSrcBlob->u64PhyAddr;
stMem.u64VirAddr = pstSrcBlob->u64VirAddr;
stMem.u32Size = u32BlobSize;
s32Ret = SAMPLE_FlushCache(&stMem);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, FAIL_0, "SAMPLE_Utils_FlushCache failed!");
return HI_SUCCESS;
FAIL_0:
if (imgFp != HI_NULL) {
fclose(imgFp);
}
return HI_FAILURE;
}
HI_S32 SAMPLE_RUNTIME_ReadConfig(const HI_CHAR *pcConfigFile, HI_CHAR **acBuff)
{
HI_CHAR acCanonicalPath[PATH_MAX + 1] = {0};
FILE *f = NULL;
HI_CHAR *tempBuff = NULL;
#ifdef _WIN32
SAMPLE_CHK_RET((strlen(pcConfigFile) > PATH_MAX) ||
HI_NULL == _fullpath(acCanonicalPath, pcConfigFile,
PATH_MAX),
HI_FAILURE, "fullpath fail %s", pcConfigFile);
#else
SAMPLE_CHK_RET((strlen(pcConfigFile) > PATH_MAX) ||
(HI_NULL == realpath(pcConfigFile, acCanonicalPath)),
HI_FAILURE, "realpath fail %s", pcConfigFile);
#endif
f = fopen(acCanonicalPath, "r");
SAMPLE_CHK_RET(f == HI_NULL, HI_FAILURE, "config file %s not exists", acCanonicalPath);
HI_S32 s32Ret = fseek(f, 0L, SEEK_END);
SAMPLE_CHK_GOTO(s32Ret != HI_SUCCESS, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
HI_S64 s64FileSize = ftell(f);
SAMPLE_CHK_GOTO(-1 == s64FileSize, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
tempBuff = (HI_CHAR *)malloc(s64FileSize + 1);
SAMPLE_CHK_GOTO(tempBuff == HI_NULL, CLOSE_FILE, "SAMPLE_RUNTIME_ReadConfig Failed");
rewind(f);
size_t readSize = fread(tempBuff, 1, s64FileSize, f);
if (((HI_S64)readSize != s64FileSize) && !feof(f)) {
SAMPLE_FREE(tempBuff);
*acBuff = NULL;
fclose(f);
return HI_FAILURE;
}
tempBuff[readSize] = '[=16=]';
*acBuff = tempBuff;
fclose(f);
return HI_SUCCESS;
CLOSE_FILE:
fclose(f);
return HI_FAILURE;
}
sample_memory_ops.c
/*
* Copyright (C) Hisilicon Technologies Co., Ltd. 2018-2019. All rights reserved.
* Description:
* Author:
* Create: 2018-05-19
*/
#include "sample_memory_ops.h"
#include <stdlib.h>
#ifdef ON_BOARD
#include "mpi_sys.h"
#endif
#include "sample_log.h"
HI_S32 SAMPLE_AllocMem(HI_RUNTIME_MEM_S *pstMemInfo, HI_BOOL bCached)
{
// SAMPLE_CHK_RET((pstMemInfo == HI_NULL), HI_FAILURE, "param pstMemInfo is NULL");
// SAMPLE_CHK_RET((HI_NULL == pstMemInfo->u64PhyAddr), HI_FAILURE, "param u64PhyAddr is 0");
#ifdef ON_BOARD
HI_S32 s32Ret = HI_SUCCESS;
if (bCached) {
s32Ret = HI_MPI_SYS_MmzAlloc_Cached(&pstMemInfo->u64PhyAddr,
(HI_VOID **)&(pstMemInfo->u64VirAddr), NULL, HI_NULL,
pstMemInfo->u32Size);
} else {
s32Ret = HI_MPI_SYS_MmzAlloc(&pstMemInfo->u64PhyAddr, (HI_VOID **)&(pstMemInfo->u64VirAddr), NULL, HI_NULL,
pstMemInfo->u32Size);
}
#else
pstMemInfo->u64VirAddr = (HI_U32)((uintptr_t)malloc(pstMemInfo->u32Size));
HI_S32 s32Ret = pstMemInfo->u64VirAddr == 0 ? HI_FAILURE : HI_SUCCESS;
pstMemInfo->u64PhyAddr = pstMemInfo->u64VirAddr;
#endif
return s32Ret;
}
HI_S32 SAMPLE_FlushCache(HI_RUNTIME_MEM_S *pstMemInfo)
{
SAMPLE_CHK_RET((pstMemInfo->u64VirAddr == 0), HI_FAILURE, "param pu8VirAddr is NULL");
#ifdef ON_BOARD
return HI_MPI_SYS_MmzFlushCache(pstMemInfo->u64PhyAddr, (HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr),
pstMemInfo->u32Size);
#endif
return HI_SUCCESS;
}
HI_S32 SAMPLE_FreeMem(HI_RUNTIME_MEM_S *pstMemInfo)
{
SAMPLE_CHK_RET((pstMemInfo->u64VirAddr == 0), HI_FAILURE, "param pu8VirAddr is NULL");
#ifdef ON_BOARD
return HI_MPI_SYS_MmzFree (pstMemInfo->u64PhyAddr, (HI_VOID *)((uintptr_t)pstMemInfo->u64VirAddr));
#else
free ((HI_U8 *)((uintptr_t) pstMemInfo->u64VirAddr));
pstMemInfo->u64PhyAddr = 0;
pstMemInfo->u64VirAddr = 0;
#endif
return HI_SUCCESS;
}
WK 文件 有 50 个输出：https://drive.google.com/file/d/1chePxF3j3KF4KvFvhRGGUKQz7ZDipeBx/view?usp=sharing
有 5 个输出：https://drive.google.com/file/d/1chePxF3j3KF4KvFvhRGGUKQz7ZDipeBx/view?usp=sharing
问题
这是什么错误？我在哪里分配 RAM 失败？这是模拟器吗error/bug?
创建的 .wk 使用 NNIE_Mapper 生成的格式不可用于运行时模拟。
在RuyiStudio中创建类型为"Runtime"的项目，或在Linux下使用runtime_mapper生成有效的.wk并解决问题！[=13=]
希望错误消息更好，但是哦，好吧！
感谢 Telegram 上的 OpenPIC 社区！