payton/nt9856x
Template
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
e8f932ecb1
nt9856x/code/hdal/samples/alg_pvdcnn_sample/ai2_pvd_sample.c
payton 543d2da268 1.一次S530代码提交
2023-03-28 15:07:53 +08:00

1676 lines
55 KiB
C
Executable File
Raw Blame History

/**
@brief Source file of vendor ai net sample code.
@file ai_net_with_buf.c
@ingroup ai_net_sample
@note Nothing.
Copyright Novatek Microelectronics Corp. 2020. All rights reserved.
*/
/*-----------------------------------------------------------------------------*/
/* Including Files */
/*-----------------------------------------------------------------------------*/
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <string.h>
#include <signal.h>
#include "hdal.h"
#include "hd_type.h"
#include "hd_debug.h"
#include "vendor_ai.h"
#include "vendor_ai_cpu/vendor_ai_cpu.h"
#include "vendor_ai_cpu_postproc.h"
#include "vendor_ai_util.h"
#include <arm_neon.h>
#include <sys/time.h>
#include "limit_fdet_lib.h"
#include "pvdcnn_lib.h"
// platform dependent
#if defined(__LINUX)
#include <pthread.h> //for pthread API
#define MAIN(argc, argv) int main(int argc, char** argv)
#define GETCHAR() getchar()
#else
#include <FreeRTOS_POSIX.h>
#include <FreeRTOS_POSIX/pthread.h> //for pthread API
#include <kwrap/util.h> //for sleep API
#define sleep(x) vos_util_delay_ms(1000*(x))
#define msleep(x) vos_util_delay_ms(x)
#define usleep(x) vos_util_delay_us(x)
#include <kwrap/examsys.h> //for MAIN(), GETCHAR() API
#define MAIN(argc, argv) EXAMFUNC_ENTRY(ai_net_with_buf, argc, argv)
#endif
///////////////////////////////////////////////////////////////////////////////
#define VENDOR_AI_CFG 0x000f0000 //vendor ai config
#define AI_RGB_BUFSIZE(w, h) (ALIGN_CEIL_4((w) * HD_VIDEO_PXLFMT_BPP(HD_VIDEO_PXLFMT_RGB888_PLANAR) / 8) * (h))
#define LIMIT_PVD 1
#define max_distance_mode 0
#define debuginf 1
#define MAX_FRAME_WIDTH 2880 //1920
#define MAX_FRAME_HEIGHT 1620 //1080
#define SCALE_BUF_SIZE (MAX_FRAME_WIDTH * MAX_FRAME_HEIGHT * 3)
#define SAVE_SCALE 0
///////////////////////////////////////////////////////////////////////////////
BOOL is_net_exit = FALSE;
/*-----------------------------------------------------------------------------*/
/* Type Definitions */
/*-----------------------------------------------------------------------------*/
typedef struct _MEM_PARM {
UINT32 pa;
UINT32 va;
UINT32 size;
UINT32 blk;
} MEM_PARM;
typedef struct _NET_PROC {
CHAR model_filename[256];
INT32 binsize;
int job_method;
int job_wait_ms;
int buf_method;
MEM_PARM proc_mem;
UINT32 proc_id;
MEM_PARM io_mem;
} NET_PROC;
typedef struct _NET_IN {
CHAR input_filename[256];
UINT32 w;
UINT32 h;
UINT32 c;
UINT32 loff;
UINT32 fmt;
VENDOR_AI_BUF src_img;
MEM_PARM input_mem;
} NET_IN;
/*-----------------------------------------------------------------------------*/
/* Global Functions */
/*-----------------------------------------------------------------------------*/
BOOL need_ise_resize(NET_IN nn_in, INT32 *size)
{
if (nn_in.w < (UINT32)size[0] || nn_in.h < (UINT32)size[1]) {
return TRUE;
} else if (nn_in.w > MAX_FRAME_WIDTH || nn_in.h > MAX_FRAME_HEIGHT) {
return TRUE;
} else {
return FALSE;
}
}
static HD_RESULT model_hdal_mem_get(MEM_PARM *mem_parm, UINT32 size)
{
HD_RESULT ret = HD_OK;
UINT32 pa = 0;
void *va = NULL;
HD_COMMON_MEM_VB_BLK blk;
/*ret = hd_common_mem_init(NULL);
if (HD_OK != ret) {
printf("hd_common_mem_init err: %d\r\n", ret);
return ret;
}
// Allocate parameter buffer
if (mem_parm->va != 0) {
printf("err: mem has already been inited\r\n");
return -1;
}*/
blk = hd_common_mem_get_block(HD_COMMON_MEM_USER_DEFINIED_POOL, size, DDR_ID0); //HD_COMMON_MEM_USER_DEFINIED_POOL + id HD_COMMON_MEM_CNN_POOL
if (HD_COMMON_MEM_VB_INVALID_BLK == blk) {
printf("hd_common_mem_get_block fail\r\n");
return HD_ERR_NG;
}
pa = hd_common_mem_blk2pa(blk);
if (pa == 0) {
printf("not get buffer, pa=%08x\r\n", (int)pa);
return HD_ERR_NOMEM;
}
va = hd_common_mem_mmap(HD_COMMON_MEM_MEM_TYPE_CACHE, pa, size);
/* Release buffer */
if (va == 0) {
ret = hd_common_mem_munmap(va, size);
if (ret != HD_OK) {
printf("mem unmap fail\r\n");
return ret;
}
}
mem_parm->pa = pa;
mem_parm->va = (UINT32)va;
mem_parm->size = size;
mem_parm->blk = blk;
return HD_OK;
}
static HD_RESULT model_hdal_mem_release(MEM_PARM *mem_parm)
{
HD_RESULT ret = HD_OK;
/* Release in buffer */
if (mem_parm->va) {
ret = hd_common_mem_munmap((void *)mem_parm->va, mem_parm->size);
if (ret != HD_OK) {
printf("mem_uninit : (g_mem.va)hd_common_mem_munmap fail.\r\n");
return ret;
}
}
ret = hd_common_mem_release_block(mem_parm->blk);
if (ret != HD_OK) {
printf("mem_uninit : (g_mem.pa)hd_common_mem_release_block fail.\r\n");
return ret;
}
mem_parm->pa = 0;
mem_parm->va = 0;
mem_parm->size = 0;
mem_parm->blk = (UINT32) - 1;
return HD_OK;
}
static HD_RESULT hdal_mem_alloc(MEM_PARM *mem_parm, CHAR *name, UINT32 size)
{
HD_RESULT ret = HD_OK;
UINT32 pa = 0;
void *va = NULL;
//alloc private pool
ret = hd_common_mem_alloc(name, &pa, (void **)&va, size, DDR_ID0);
if (ret != HD_OK) {
return ret;
}
mem_parm->pa = pa;
mem_parm->va = (UINT32)va;
mem_parm->size = size;
mem_parm->blk = (UINT32) - 1;
return HD_OK;
}
static HD_RESULT hdal_mem_free(MEM_PARM *mem_parm)
{
HD_RESULT ret = HD_OK;
//free private pool
ret = hd_common_mem_free(mem_parm->pa, (void *)mem_parm->va);
if (ret != HD_OK) {
return ret;
}
//printf("1free pa = 0x%x, va = 0x%x\r\n", mem_parm->pa, mem_parm->va);
mem_parm->pa = 0;
mem_parm->va = 0;
mem_parm->size = 0;
mem_parm->blk = (UINT32) - 1;
//printf("2free pa = 0x%x, va = 0x%x\r\n", mem_parm->pa, mem_parm->va);
return HD_OK;
}
static INT32 _getsize_model(char *filename)
{
FILE *bin_fd;
UINT32 bin_size = 0;
bin_fd = fopen(filename, "rb");
if (!bin_fd) {
printf("get bin(%s) size fail\n", filename);
return (-1);
}
fseek(bin_fd, 0, SEEK_END);
bin_size = ftell(bin_fd);
fseek(bin_fd, 0, SEEK_SET);
fclose(bin_fd);
return bin_size;
}
static INT32 load_bin_file_to_mem(const CHAR *filename, MEM_PARM *mem_parm)
{
FILE *fd;
INT32 size = 0;
fd = fopen(filename, "rb");
if (!fd) {
printf("cannot read %s\r\n", filename);
return -1;
}
fseek(fd, 0, SEEK_END);
size = ftell(fd);
fseek(fd, 0, SEEK_SET);
INT32 ww = (INT32)fread((VOID *)mem_parm->va, 1, size, fd);
//printf("size: %ld, ww:%ld\n", size, ww);
if (size < 0) {
printf("getting %s size failed\r\n", filename);
//} else if ((INT32)fread((VOID *)mem_parm->va, 1, size, fd) != size) {
} else if (ww != size) {
printf("read size < %ld\r\n", size);
size = -1;
}
//mem_parm->size = size;
if (fd) {
fclose(fd);
}
return size;
}
/*-----------------------------------------------------------------------------*/
/* Input Functions */
/*-----------------------------------------------------------------------------*/
///////////////////////////////////////////////////////////////////////////////
static HD_RESULT input_open(NET_IN *p_nn_in)
{
HD_RESULT ret = HD_OK;
p_nn_in->src_img.size = p_nn_in->w * p_nn_in->h * 3 / 2;
//ret = hdal_mem_alloc(&p_nn_in->input_mem, "ai_in_buf", p_nn_in->src_img.size);
ret = hdal_mem_alloc(&p_nn_in->input_mem, "ai_in_buf", AI_RGB_BUFSIZE(p_nn_in->w, p_nn_in->h));
if (ret != HD_OK) {
printf("alloc ai_in_buf fail\r\n");
return HD_ERR_FAIL;
}
UINT32 file_len = load_bin_file_to_mem(p_nn_in->input_filename, &(p_nn_in->input_mem));
if (file_len < 0) {
printf("load buf(%s) fail\r\n", p_nn_in->input_filename);
return HD_ERR_NG;
}
hd_common_mem_flush_cache((VOID *)p_nn_in->input_mem.va, file_len);
p_nn_in->src_img.width = p_nn_in->w;
p_nn_in->src_img.height = p_nn_in->h;
p_nn_in->src_img.channel = p_nn_in->c;
p_nn_in->src_img.line_ofs = p_nn_in->loff;
p_nn_in->src_img.fmt = HD_VIDEO_PXLFMT_YUV420;
//p_nn_in->src_img.fmt = HD_VIDEO_PXLFMT_Y8; //bin<69>n<EFBFBD><6E>
p_nn_in->src_img.pa = p_nn_in->input_mem.pa;
p_nn_in->src_img.va = p_nn_in->input_mem.va;
p_nn_in->src_img.sign = MAKEFOURCC('A', 'B', 'U', 'F');
//p_nn_in->src_img.size = p_nn_in->loff * p_nn_in->h * 3 / 2;
return ret;
}
static HD_RESULT input_close(NET_IN *p_nn_in)
{
HD_RESULT ret = HD_OK;
if (p_nn_in->src_img.pa && p_nn_in->src_img.va) {
MEM_PARM input_mem;
input_mem.pa = p_nn_in->src_img.pa;
input_mem.va = p_nn_in->src_img.va;
hdal_mem_free(&input_mem);
}
return ret;
}
/*-----------------------------------------------------------------------------*/
/* Network Functions */
/*-----------------------------------------------------------------------------*/
static HD_RESULT network_alloc_io_buf(NET_PROC *p_net)
{
HD_RESULT ret = HD_OK;
VENDOR_AI_NET_CFG_WORKBUF wbuf = {0};
printf("network_alloc_io_buf, procid: %ld\n", p_net->proc_id);
ret = vendor_ai_net_get(p_net->proc_id, VENDOR_AI_NET_PARAM_CFG_WORKBUF, &wbuf);
printf("after vendor_ai_net_get\n");
if (ret != HD_OK) {
printf("proc_id(%lu) get VENDOR_AI_NET_PARAM_CFG_WORKBUF fail\r\n", p_net->proc_id);
return HD_ERR_FAIL;
}
ret = hdal_mem_alloc(&p_net->io_mem, "ai_io_buf", wbuf.size);
if (ret != HD_OK) {
printf("proc_id(%lu) alloc ai_io_buf fail\r\n", p_net->proc_id);
return HD_ERR_FAIL;
}
wbuf.pa = p_net->io_mem.pa;
wbuf.va = p_net->io_mem.va;
wbuf.size = p_net->io_mem.size;
ret = vendor_ai_net_set(p_net->proc_id, VENDOR_AI_NET_PARAM_CFG_WORKBUF, &wbuf);
if (ret != HD_OK) {
printf("proc_id(%lu) set VENDOR_AI_NET_PARAM_CFG_WORKBUF fail\r\n", p_net->proc_id);
return HD_ERR_FAIL;
}
//printf("alloc_io_buf: work buf, pa = %#lx, va = %#lx, size = %lu\r\n", wbuf.pa, wbuf.va, wbuf.size);
//printf("alloc_io_buf: work buf size = %lu\r\n", wbuf.size);
return ret;
}
static HD_RESULT network_free_io_buf(NET_PROC *p_net)
{
HD_RESULT ret = HD_OK;
if (p_net->io_mem.pa && p_net->io_mem.va) {
hdal_mem_free(&p_net->io_mem);
}
return ret;
}
static HD_RESULT network_open(NET_PROC *p_net, INT32 cls)
{
HD_RESULT ret = HD_OK;
UINT32 loadsize = 0;
// set buf opt
/*printf("set buf opt\n");
VENDOR_AI_NET_CFG_BUF_OPT cfg_buf_opt = {0};
cfg_buf_opt.method = VENDOR_AI_NET_BUF_OPT_NONE;
cfg_buf_opt.ddr_id = DDR_ID0;
vendor_ai_net_set(proc_id, VENDOR_AI_NET_PARAM_CFG_BUF_OPT, &cfg_buf_opt);
// set job option
VENDOR_AI_NET_CFG_JOB_OPT cfg_job_opt = {0};
cfg_job_opt.method = VENDOR_AI_NET_JOB_OPT_LINEAR_O1;
cfg_job_opt.wait_ms = 0;
cfg_job_opt.schd_parm = VENDOR_AI_FAIR_CORE_ALL; //FAIR dispatch to ALL core
vendor_ai_net_set(proc_id, VENDOR_AI_NET_PARAM_CFG_JOB_OPT, &cfg_job_opt);*/
VENDOR_AI_NET_CFG_BUF_OPT cfg_buf_opt = {0};
cfg_buf_opt.method = VENDOR_AI_NET_BUF_OPT_NONE;//p_net->buf_method;
cfg_buf_opt.ddr_id = DDR_ID0;
vendor_ai_net_set(p_net->proc_id, VENDOR_AI_NET_PARAM_CFG_BUF_OPT, &cfg_buf_opt);
// set job opt
VENDOR_AI_NET_CFG_JOB_OPT cfg_job_opt = {0};
cfg_job_opt.method = VENDOR_AI_NET_JOB_OPT_LINEAR_O1;//p_net->job_method;//0
cfg_job_opt.wait_ms = 0;//p_net->job_wait_ms;//-1
cfg_job_opt.schd_parm = VENDOR_AI_FAIR_CORE_ALL; //FAIR dispatch to ALL core
//#define VENDOR_AI_FAIR_CORE_ALL 0x000000ff ///< fair: all core
//#define VENDOR_AI_FAIR_CORE(ch) ((ch) & 0xff) ///< fair: select core (default 0)
vendor_ai_net_set(p_net->proc_id, VENDOR_AI_NET_PARAM_CFG_JOB_OPT, &cfg_job_opt);
if (strlen(p_net->model_filename) == 0) {
printf("proc_id(%u) model is null\r\n", p_net->proc_id);
return 0;
}
if (cls > 0) {
/*
MEM_PARM input_mem_parm;
UINT32 input_mem_size = (UINT32)p_net->binsize;
ret = hdal_mem_alloc(&input_mem_parm, "cls_model_mem", input_mem_size);
if (HD_OK != ret) {
printf("cls model mem allocate!!\n");
return 0;
}*/
//p_net->proc_mem.va = buf->va;
//p_net->proc_mem.pa = buf->pa;
//p_net->proc_mem.size = buf->size;
//p_net->proc_mem.blk = buf->blk;
printf("before hdal_mem_alloc cls\n");
ret = hdal_mem_alloc(&(p_net->proc_mem), "cls_model_mem", p_net->binsize);
printf("after hdal_mem_alloc cls\n");
loadsize = load_bin_file_to_mem(p_net->model_filename, &(p_net->proc_mem));
//printf("line328 load_bin_size: %d\r\n", loadsize);
printf("loadsize: %ld, binsize: %ld\r\n", loadsize, p_net->binsize);
printf("model load : %s\r\n", p_net->model_filename);
printf("proc_id(%u)\r\n", p_net->proc_id);
if (loadsize <= 0) {
printf("proc_id(%u) model load fail: %s\r\n", p_net->proc_id, p_net->model_filename);
return 0;
}
printf("cls:%ld, va:%ld, size:%ld\n", cls, p_net->proc_mem.va, p_net->proc_mem.size);
VENDOR_AIS_FLOW_MEM_PARM version_check;
version_check.va = p_net->proc_mem.va;
version_check.pa = p_net->proc_mem.pa;
version_check.size = p_net->proc_mem.size;
ret = limit_fdet_version_check(&version_check);
p_net->proc_mem.va = version_check.va;
p_net->proc_mem.pa = version_check.pa;
p_net->proc_mem.size = version_check.size;
if (ret != HD_OK) {
printf("ERR: limit_fdet version check fail (%d)!!\r\n", ret);
return ret;
}
} else {
// printf("before model_1hdal_mem_get\n");
ret = model_hdal_mem_get(&p_net->proc_mem, p_net->binsize);
// printf("after model_1hdal_mem_get\n");
//load file
//p_net->proc_mem.va = buf->va;
//p_net->proc_mem.pa = buf->pa;
//p_net->proc_mem.size = buf->size;
//p_net->proc_mem.blk = buf->blk;
loadsize = load_bin_file_to_mem(p_net->model_filename, &(p_net->proc_mem));
//printf("line328 load_bin_size: %d\r\n", loadsize);
printf("loadsize: %ld, binsize: %ld\r\n", loadsize, p_net->binsize);
printf("model load : %s\r\n", p_net->model_filename);
printf("proc_id(%u)\r\n", p_net->proc_id);
if (loadsize <= 0) {
printf("proc_id(%u) model load fail: %s\r\n", p_net->proc_id, p_net->model_filename);
return 0;
}
VENDOR_AIS_FLOW_MEM_PARM version_pvd_check;
version_pvd_check.va = p_net->proc_mem.va;
version_pvd_check.pa = p_net->proc_mem.pa;
version_pvd_check.size = p_net->proc_mem.size;
ret = pvdcnn_version_check(&version_pvd_check);
p_net->proc_mem.va = version_pvd_check.va;
p_net->proc_mem.pa = version_pvd_check.pa;
p_net->proc_mem.size = version_pvd_check.size;
if (ret != HD_OK) {
printf("ERR: limit_fdet version check fail (%d)!!\r\n", ret);
return ret;
}
printf("clsornot:%ld, va:%ld, size:%ld\n", cls, p_net->proc_mem.va, p_net->proc_mem.size);
}
// set model
vendor_ai_net_set(p_net->proc_id, VENDOR_AI_NET_PARAM_CFG_MODEL, (VENDOR_AI_NET_CFG_MODEL *)&p_net->proc_mem);
// open
printf("before vendor_ai_net_open\n");
vendor_ai_net_open(p_net->proc_id);
printf("after vendor_ai_net_open\n");
if ((ret = network_alloc_io_buf(p_net)) != HD_OK) {
return ret;
}
printf("after network_alloc_io_buf\n\n");
return ret;
}
static HD_RESULT network_close(NET_PROC *p_net)
{
HD_RESULT ret = HD_OK;
if ((ret = network_free_io_buf(p_net)) != HD_OK) {
return ret;
}
// close
ret = vendor_ai_net_close(p_net->proc_id);
model_hdal_mem_release(&p_net->proc_mem);
return ret;
}
static HD_RESULT network_get_ai_inputlayer_info(INT32 proc_id)
{
HD_RESULT ret = HD_OK;
VENDOR_AI_BUF p_inbuf = {0};
VENDOR_AI_BUF p_outbuf = {0};
// get layer0 in buf
ret = vendor_ai_net_get(proc_id, VENDOR_AI_NET_PARAM_IN(0, 0), &p_inbuf);
if (HD_OK != ret) {
printf("proc_id(%u) get layer0 inbuf fail !!\n", proc_id);
return ret;
}
// get layer0 in buf
ret = vendor_ai_net_get(proc_id, VENDOR_AI_NET_PARAM_OUT(0, 0), &p_outbuf);
if (HD_OK != ret) {
printf("proc_id(%u) get layer0 outbuf fail !!\n", proc_id);
return ret;
}
/*
printf("dump layer0 info:\n");
printf(" channel(%lu)\n", p_inbuf.channel);
printf(" fmt(0x%lx)\n", p_inbuf.fmt);
printf(" width(%lu)\n", p_outbuf.width);
printf(" height(%lu)\n", p_outbuf.height);
printf(" channel(%lu)\n", p_outbuf.channel);
printf(" batch_num(%lu)\n", p_outbuf.batch_num);
printf(" fmt(0x%lx)\n", p_outbuf.fmt);
printf("\n");
*/
return ret;
}
static HD_RESULT network_get_ai_outlayer_list(UINT32 proc_id, UINT32 *outlayer_num, UINT32 *outlayer_path_list)
{
HD_RESULT ret = HD_OK;
VENDOR_AI_NET_INFO net_info = {0};
// get output layer number
ret = vendor_ai_net_get(proc_id, VENDOR_AI_NET_PARAM_INFO, &net_info);
if (HD_OK != ret) {
printf("proc_id(%lu) get info fail !!\n", proc_id);
return ret;
}
*outlayer_num = net_info.out_buf_cnt;
// get path_list
ret = vendor_ai_net_get(proc_id, VENDOR_AI_NET_PARAM_OUT_PATH_LIST, outlayer_path_list);
if (HD_OK != ret) {
printf("proc_id(%u) get outlayer_path_list fail !!\n", proc_id);
}
return ret;
}
static HD_RESULT network_get_ai_outlayer_by_path_id(INT32 proc_id, UINT32 path_id, VENDOR_AI_BUF *p_outbuf)
{
HD_RESULT ret = HD_OK;
// get out buf by path_id
ret = vendor_ai_net_get(proc_id, path_id, p_outbuf);
if (HD_OK != ret) {
printf("proc_id(%u) get AI_OUTBUF fail !!\n", proc_id);
return ret;
}
/*
INT8 bitdepth = HD_VIDEO_PXLFMT_BITS(p_outbuf->fmt);
INT8 sign_bits = HD_VIDEO_PXLFMT_SIGN(p_outbuf->fmt);
INT8 int_bits = HD_VIDEO_PXLFMT_INT(p_outbuf->fmt);
INT8 frac_bits = HD_VIDEO_PXLFMT_FRAC(p_outbuf->fmt);
printf("dump ai_outbuf:\n");
printf(" name (%s)\n", p_outbuf->name);
printf(" pa (0x%lx)\n", p_outbuf->pa);
printf(" va (0x%lx)\n", p_outbuf->va);
printf(" size (%lu)\n", p_outbuf->size);
printf(" width (%lu)\n", p_outbuf->width);
printf(" height (%lu)\n", p_outbuf->height);
printf(" channel (%lu)\n", p_outbuf->channel);
printf(" layout (%s)\n", p_outbuf->layout);
printf(" bitdepth (%d)\n", bitdepth);
printf(" sign_bits (%d)\n", sign_bits);
printf(" int_bits (%d)\n", int_bits);
printf(" frac_bits (%d)\n", frac_bits);
printf(" scale_ratio (%f)\n", p_outbuf->scale_ratio);
printf("\n");
*/
return ret;
}
#if 0
static HD_RESULT network_get_ai_outlayer_by_name(INT32 proc_id, VENDOR_AI_BUF *p_outbuf, char *p_buf_name)
{
HD_RESULT ret = HD_OK;
VENDOR_AI_BUF_NAME buf_name = {0};
sprintf(buf_name.name, p_buf_name);
// get out_path by name
ret = vendor_ai_net_get(proc_id, VENDOR_AI_NET_PARAM_OUT_PATH_BY_NAME, &buf_name);
if (HD_OK != ret) {
printf("proc_id(%u) get OUT_PATH_BY_NAME fail !!\n", proc_id);
return ret;
}
// get out buf by path_id
ret = vendor_ai_net_get(proc_id, buf_name.path_id, p_outbuf);
if (HD_OK != ret) {
printf("proc_id(%u) get AI_OUTBUF fail !!\n", proc_id);
return ret;
}
/*INT8 bitdepth = HD_VIDEO_PXLFMT_BITS(p_outbuf->fmt);
INT8 sign_bits = HD_VIDEO_PXLFMT_SIGN(p_outbuf->fmt);
INT8 int_bits = HD_VIDEO_PXLFMT_INT(p_outbuf->fmt);
INT8 frac_bits = HD_VIDEO_PXLFMT_FRAC(p_outbuf->fmt);
printf("dump ai_outbuf:\n");
printf(" name (%s)\n", p_outbuf->name);
printf(" pa (0x%lx)\n", p_outbuf->pa);
printf(" va (0x%lx)\n", p_outbuf->va);
printf(" size (%lu)\n", p_outbuf->size);
printf(" width (%lu)\n", p_outbuf->width);
printf(" height (%lu)\n", p_outbuf->height);
printf(" channel (%lu)\n", p_outbuf->channel);
printf(" layout (%s)\n", p_outbuf->layout);
printf(" bitdepth (%d)\n", bitdepth);
printf(" sign_bits (%d)\n", sign_bits);
printf(" int_bits (%d)\n", int_bits);
printf(" frac_bits (%d)\n", frac_bits);
printf(" scale_ratio (%f)\n", p_outbuf->scale_ratio);
printf("\n");*/
return ret;
}
#endif
VOID pd_gfx_img_to_vendor(VENDOR_AI_BUF *out_img, HD_GFX_IMG_BUF *in_img, UINT32 va)
{
//VENDOR_AIS_IMG_PARM* p_src_img;
out_img->fmt = in_img->format;
out_img->channel = 2;
out_img->pa = in_img->p_phy_addr[0];
out_img->va = va;
out_img->line_ofs = in_img->lineoffset[0];
out_img->width = in_img->dim.w;
out_img->height = in_img->dim.h;
out_img->sign = MAKEFOURCC('A', 'B', 'U', 'F');
out_img->size = in_img->dim.w * in_img->dim.h * 3 / 2;
//return p_src_img;
}
HD_RESULT pdcnn_crop_img(HD_GFX_IMG_BUF *dst_img, VENDOR_AI_BUF *src_img, HD_GFX_SCALE_QUALITY method, PVD_IRECT *roi)
{
HD_RESULT ret;
HD_GFX_SCALE param;
memset(&param, 0, sizeof(HD_GFX_SCALE));
UINT32 dif = 0;
param.src_img.dim.w = src_img->width;
param.src_img.dim.h = src_img->height;
param.src_img.format = HD_VIDEO_PXLFMT_YUV420;
param.src_img.p_phy_addr[0] = src_img->pa;
param.src_img.p_phy_addr[1] = src_img->pa + src_img->width * src_img->height;
param.src_img.lineoffset[0] = src_img->line_ofs;
param.src_img.lineoffset[1] = src_img->line_ofs;
param.dst_img.dim.w = dst_img->dim.w;
param.dst_img.dim.h = dst_img->dim.h;
param.dst_img.format = HD_VIDEO_PXLFMT_YUV420;
param.dst_img.p_phy_addr[0] = dst_img->p_phy_addr[0];
param.dst_img.p_phy_addr[1] = dst_img->p_phy_addr[0] + dst_img->dim.w * dst_img->dim.h;
param.dst_img.lineoffset[0] = dst_img->dim.w;
param.dst_img.lineoffset[1] = dst_img->dim.w;
param.dst_region.x = 0;
param.dst_region.y = 0;
param.dst_region.w = dst_img->dim.w;
param.dst_region.h = dst_img->dim.h;
param.src_region.w = ALIGN_CEIL_4((UINT32)(roi->x2 - roi->x1 + 0.5));
param.src_region.h = ALIGN_CEIL_4((UINT32)(roi->y2 - roi->y1 + 0.5));
if (param.src_region.w < 16) {
dif = 16 - param.src_region.w;
param.src_region.w = 16;
param.src_region.x = ALIGN_CEIL_4((UINT32)(roi->x1 - dif / 2));
} else {
param.src_region.x = ALIGN_CEIL_4((UINT32)(roi->x1 + 0.5));
}
if (param.src_region.h < 16) {
dif = 16 - param.src_region.h;
param.src_region.h = 16;
param.src_region.y = ALIGN_CEIL_4((UINT32)(roi->y1 - dif / 2));
} else {
param.src_region.y = ALIGN_CEIL_4((UINT32)(roi->y1 + 0.5));
}
param.quality = method;
ret = hd_gfx_scale(&param);
if (ret != HD_OK) {
printf("hd_gfx_scale fail=%d\n", ret);
}
return ret;
}
// static BOOL pdcnn_check_overlap(PVDCNN_RESULT *p_box1, PVDCNN_RESULT *p_box2, FLOAT ratio, INT32 method)
// {
// FLOAT left_x, right_x, up_y, down_y;
// FLOAT inter_value, union_value, min_value;
// // caculate intersect position
// left_x = MAX(p_box1->x1, p_box2->x1);
// right_x = MIN(p_box1->x2, p_box2->x2);
// if (left_x >= right_x) return FALSE;
// up_y = MAX(p_box1->y1, p_box2->y1);
// down_y = MIN(p_box1->y2, p_box2->y2);
// if (up_y >= down_y) return FALSE;
// // caculate Intersection o Union
// inter_value = (right_x - left_x) * (down_y - up_y);
// if (1 == method)
// {
// union_value = (p_box1->x2 - p_box1->x1) * (p_box1->y2 - p_box1->y1) + (p_box2->x2 - p_box2->x1) * (p_box2->y2 - p_box2->y1) - inter_value;
// if (inter_value >= union_value * ratio) return TRUE;
// else return FALSE;
// }
// else if (2 == method)
// {
// min_value = MIN((p_box1->x2 - p_box1->x1) * (p_box1->y2 - p_box1->y1), (p_box2->x2 - p_box2->x1) * (p_box2->y2 - p_box2->y1));
// if (inter_value >= min_value * ratio) return TRUE;
// else return FALSE;
// }
// else
// {
// printf("Error: Only support IOU and MIN method!\r\n");
// return FALSE;
// }
// }
VOID pdcnn_quick_sort(PVDCNN_RESULT *persons, INT32 left, INT32 right)
{
if (left >= right) {
return;
}
INT32 i = left;
INT32 j = right;
PVDCNN_RESULT key = persons[left];
while (i < j) {
while (i < j && key.score >= persons[j].score) {
j--;
}
persons[i] = persons[j];
while (i < j && key.score <= persons[i].score) {
i++;
}
persons[j] = persons[i];
}
persons[i] = key;
pdcnn_quick_sort(persons, left, i - 1);
pdcnn_quick_sort(persons, i + 1, right);
}
HD_RESULT ai_crop_img(HD_GFX_IMG_BUF *dst_img, VENDOR_AI_BUF src_img, HD_GFX_SCALE_QUALITY method)
{
HD_RESULT ret;
HD_GFX_SCALE param;
memset(&param, 0, sizeof(HD_GFX_SCALE));
param.src_img.dim.w = src_img.width;
param.src_img.dim.h = src_img.height;
param.src_img.format = HD_VIDEO_PXLFMT_YUV420;
param.src_img.p_phy_addr[0] = src_img.pa;
param.src_img.p_phy_addr[1] = src_img.pa + src_img.width * src_img.height;
param.src_img.lineoffset[0] = src_img.line_ofs;
param.src_img.lineoffset[1] = src_img.line_ofs;
param.dst_img.dim.w = dst_img->dim.w;
param.dst_img.dim.h = dst_img->dim.h;
param.dst_img.format = HD_VIDEO_PXLFMT_YUV420;
param.dst_img.p_phy_addr[0] = dst_img->p_phy_addr[0];
param.dst_img.p_phy_addr[1] = dst_img->p_phy_addr[0] + dst_img->dim.w * dst_img->dim.h;
param.dst_img.lineoffset[0] = dst_img->dim.w;
param.dst_img.lineoffset[1] = dst_img->dim.w;
param.src_region.x = 0;//fd_result->x_min;
param.src_region.y = 0;//fd_result->y_min;
param.src_region.w = src_img.width;//fd_result->w;
param.src_region.h = src_img.height;//fd_result->h;
param.dst_region.x = 0;
param.dst_region.y = 0;
param.dst_region.w = dst_img->dim.w;
param.dst_region.h = dst_img->dim.h;
param.quality = method;
ret = hd_gfx_scale(&param);
if (ret != HD_OK) {
printf("hd_gfx_scale fail=%d\n", ret);
}
return ret;
}
/*-----------------------------------------------------------------------------*/
/* Interface Functions */
/*-----------------------------------------------------------------------------*/
static VOID *nn_thread_api(VOID *arg)
{
HD_RESULT ret;
HD_GFX_IMG_BUF gfx_img = {0};
VENDOR_AI_BUF p_src_img;
//NET_IN* p_nn_in = (NET_IN*)arg;
HD_COMMON_MEM_INIT_CONFIG mem_cfg = {0};
static struct timeval tstart, tend;
UINT32 proc_id = 0;
// config extend engine plugin, process scheduler
/*typedef enum {
VENDOR_AI_PROC_SCHD_FAIR = 0, ///< overlapping with fair core (default)
VENDOR_AI_PROC_SCHD_CAPACITY = 1, ///< overlapping with max rate (TODO)
VENDOR_AI_PROC_SCHD_FIFO = 2, ///< first in first out
} VENDOR_AI_PROC_SCHD; */
// get vendor ai version
VENDOR_AI_NET_CFG_IMPL_VERSION ai_version;
ret = vendor_ai_cfg_get(VENDOR_AI_CFG_IMPL_VERSION, &ai_version);
if (ret != HD_OK) {
printf("get vendor ai sdk fail=%d\n", ret);
goto exit_thread;
}
// get max proc_id number and each proc_id status
VENDOR_AI_NET_CFG_PROC_COUNT proc_max_count;
ret = vendor_ai_cfg_get(VENDOR_AI_CFG_PROC_COUNT, &proc_max_count);
if (ret != HD_OK) {
printf("get max proc_id number fail=%d\n", ret);
goto exit_thread;
}
for (UINT32 i = 0; i < proc_max_count.max_proc_count; i++) {
UINT32 state = 0; //1=ready, 2=open, 3=start
vendor_ai_net_get(i, VENDOR_AI_NET_PARAM_STATE, &state);
}
NET_PROC net_info;
net_info.proc_id = proc_id;
#if PRUNE37PVD
sprintf(net_info.model_filename, "/mnt/sd/CNNLib/para/pvdcnn/nvt_model_prune37.bin");
#else
sprintf(net_info.model_filename, "/mnt/sd/CNNLib/para/pvdcnn/nvt_model_origin.bin");
#endif
net_info.binsize = _getsize_model(net_info.model_filename);
//net_info.job_method = VENDOR_AI_NET_JOB_OPT_LINEAR_O1;
//net_info.job_wait_ms = 0;
//net_info.buf_method = VENDOR_AI_NET_BUF_OPT_NONE;
//job=1
net_info.job_method = VENDOR_AI_NET_JOB_OPT_GRAPH_O1;
net_info.job_wait_ms = 0;
net_info.buf_method = VENDOR_AI_NET_BUF_OPT_SHRINK_O2;
if (net_info.binsize <= 0) {
printf("proc_id(%u) model is not exist?\r\n", proc_id);
goto exit_thread;
}
//printf("proc_id(%u) set net_info: model-file(%s), binsize=%d\r\n", proc_id, net_info.model_filename, net_info.binsize);
// config hdal common pool (in)
mem_cfg.pool_info[0].type = HD_COMMON_MEM_USER_DEFINIED_POOL;
mem_cfg.pool_info[0].blk_size = net_info.binsize;
mem_cfg.pool_info[0].blk_cnt = 1;
mem_cfg.pool_info[0].ddr_id = DDR_ID0;
// scale img buf
mem_cfg.pool_info[1].type = HD_COMMON_MEM_USER_DEFINIED_POOL;
mem_cfg.pool_info[1].blk_size = SCALE_BUF_SIZE;
mem_cfg.pool_info[1].blk_cnt = 1;
mem_cfg.pool_info[1].ddr_id = DDR_ID0;
//----------------------------------------------------------------------------------------------------------------------
#if LIMIT_PVD
LIMIT_FDET_PARAM limit_fdet_params = { 0 };
// limit_fdet_mem = { 0 };
limit_fdet_params.run_id = 1;
NET_PROC net_info1;
//UINT32 proc_id1 = 1;
net_info1.proc_id = limit_fdet_params.run_id;
sprintf(net_info1.model_filename, "/mnt/sd/CNNLib/para/public/nvt_model.bin");
net_info1.binsize = _getsize_model(net_info1.model_filename);
//net_info.job_method = VENDOR_AI_NET_JOB_OPT_LINEAR_O1;
//net_info.job_wait_ms = 0;
//net_info.buf_method = VENDOR_AI_NET_BUF_OPT_NONE;
//job=1
net_info1.job_method = VENDOR_AI_NET_JOB_OPT_GRAPH_O1;
net_info1.job_wait_ms = 0;
net_info1.buf_method = VENDOR_AI_NET_BUF_OPT_SHRINK_O2;
if (net_info1.binsize <= 0) {
printf("limit_fdet_params.run_id(%u) cls model is not exist?\r\n", limit_fdet_params.run_id);
goto exit_thread;
}
// config hdal common pool (in)
mem_cfg.pool_info[2].type = HD_COMMON_MEM_USER_DEFINIED_POOL;
mem_cfg.pool_info[2].blk_size = net_info1.binsize;
mem_cfg.pool_info[2].blk_cnt = 1;
mem_cfg.pool_info[2].ddr_id = DDR_ID0;
#endif
ret = hd_common_mem_init(&mem_cfg);
if (HD_OK != ret) {
printf("hd_common_mem_init err: %d\r\n", ret);
goto exit_thread;
}
INT32 cls = -1000;
if ((ret = network_open(&net_info, cls)) != HD_OK) {
printf("proc_id(%u nn open fail !!\n", proc_id);
goto exit_thread;
}
printf("load pvd model\n");
// open network(model)
#if LIMIT_PVD
cls = 1000;
if ((ret = network_open(&net_info1, cls)) != HD_OK) {
printf("limit_fdet_params.run_id(%u cls nn open fail !!\n", limit_fdet_params.run_id);
goto exit_thread;
}
printf("load cls model\n");
#endif
for (UINT32 i = 0; i < proc_max_count.max_proc_count; i++) {
UINT32 state = 0; //1=ready, 2=open, 3=start
vendor_ai_net_get(i, VENDOR_AI_NET_PARAM_STATE, &state);
}
// start network proc_id
ret = vendor_ai_net_start(proc_id);
if (HD_OK != ret) {
printf("proc_id(%u) nn start fail !!\n", proc_id);
goto exit_thread;
}
#if LIMIT_PVD
// start network limit_fdet_params.run_id
ret = vendor_ai_net_start(limit_fdet_params.run_id);
if (HD_OK != ret) {
printf("limit_fdet_params.run_id(%u) nn start fail !!\n", limit_fdet_params.run_id);
goto exit_thread;
}
#endif
//get nn layer 0 info
network_get_ai_inputlayer_info(proc_id);
//get output layer info list
CHAR out_layer_name[2][128] = {"layer115-conv", "layer125-conv"};
UINT32 outlayer_num = 0;
UINT32 outlayer_path_list[256] = {0};
UINT32 outlayer_path_list_tmp[256] = {0};
network_get_ai_outlayer_list(proc_id, &outlayer_num, outlayer_path_list_tmp);
memcpy(outlayer_path_list, outlayer_path_list_tmp, sizeof(UINT32) * outlayer_num);
for (UINT32 i = 0; i < outlayer_num; i++) {
VENDOR_AI_BUF layer_tmp;
network_get_ai_outlayer_by_path_id(proc_id, outlayer_path_list_tmp[i], &layer_tmp);
//printf("%d %hd %s\r\n", i, outlayer_path_list_tmp[i], layer_tmp.name);
for (UINT32 k = 0; k < outlayer_num; k++) {
if (strncmp(out_layer_name[k], layer_tmp.name, strlen(out_layer_name[k])) == 0) {
outlayer_path_list[k] = outlayer_path_list_tmp[i];
//printf(" --> %d %s \r\n", k, conf_loc_layer_name[k]);
break;
}
}
}
INT32 max_map[2] = {0};
VENDOR_AI_BUF *layer_buffer = (VENDOR_AI_BUF *)malloc(outlayer_num * sizeof(VENDOR_AI_BUF));
FLOAT **out_layer_float = (FLOAT **)malloc(outlayer_num * sizeof(FLOAT));
for (UINT32 i = 0; i < outlayer_num; i++) {
network_get_ai_outlayer_by_path_id(proc_id, outlayer_path_list[i], &(layer_buffer[i]));
/*
INT8 bitdepth = HD_VIDEO_PXLFMT_BITS(layer_buffer[i].fmt);
INT8 sign_bits = HD_VIDEO_PXLFMT_SIGN(layer_buffer[i].fmt);
INT8 int_bits = HD_VIDEO_PXLFMT_INT(layer_buffer[i].fmt);
INT8 frac_bits = HD_VIDEO_PXLFMT_FRAC(layer_buffer[i].fmt);
printf(" layer (%s)\n", layer_buffer[i].name);
printf(" pa (0x%lx)\n", layer_buffer[i].pa);
printf(" va (0x%lx)\n", layer_buffer[i].va);
printf(" size (%lu)\n", layer_buffer[i].size);
printf(" width (%lu)\n", layer_buffer[i].width);
printf(" height (%lu)\n", layer_buffer[i].height);
printf(" channel (%lu)\n", layer_buffer[i].channel);
printf(" layout (%s)\n", layer_buffer[i].layout);
printf(" bitdepth (%d)\n", bitdepth);
printf(" sign_bits (%d)\n", sign_bits);
printf(" int_bits (%d)\n", int_bits);
printf(" frac_bits (%d)\n", frac_bits);
printf(" scale_ratio (%f)\n", layer_buffer[i].scale_ratio);
*/
UINT32 length = layer_buffer[i].width * layer_buffer[i].height * layer_buffer[i].channel * layer_buffer[i].batch_num;
out_layer_float[i] = (FLOAT *) malloc(sizeof(FLOAT) * length);
if (i == outlayer_num - 1) {
max_map[0] = (INT32)layer_buffer[i].width;
max_map[1] = (INT32)layer_buffer[i].height;
}
}
// PVD detection out layer include params! default
PVD_LAYER_PARAM PVD_params;
PVD_params.nms_threshold = 0.5;
PVD_params.confindence_threshold = 0.38;
PVD_params.pd_limit_param.limit_module = 2;
PVD_params.vd_limit_param.limit_module = 2;
PVD_params.pd_limit_param.sm_thr_num = 2;
PVD_params.vd_limit_param.sm_thr_num = 2;
PVD_params.cls_score_thresh = 2;//0.986;
char paramsfile[256] = "/mnt/sd/CNNLib/para/pvdcnn/params.txt";
pvdcnn_getparam(&PVD_params, max_map, paramsfile);
printf("%f, %f, %f, %d %d %d %d\n", PVD_params.nms_threshold, PVD_params.confindence_threshold, PVD_params.cls_score_thresh,
PVD_params.pd_limit_param.limit_module, PVD_params.vd_limit_param.limit_module,
PVD_params.pd_limit_param.sm_thr_num, PVD_params.vd_limit_param.sm_thr_num);
#if(LIMIT_PVD)
FLOAT cls_det_threshold = 0.98;
FLOAT cls_threshold = 0.7;
INT32 cls_debug_print = 0;
char para_file[256];
snprintf(para_file, 256, "/mnt/sd/CNNLib/para/public/para.txt");
limit_fdet_read_para(&cls_det_threshold, &cls_threshold, &cls_debug_print, para_file);
printf("det_threshold:%f, cls:thresh: %f, debug_print:%d\n", cls_det_threshold, cls_threshold, cls_debug_print);
limit_fdet_params.cls_det_threshold = cls_det_threshold;
limit_fdet_params.cls_threshold = cls_threshold;
limit_fdet_params.cls_debug_print = cls_debug_print;
#endif
INT8 pvd_light_night = 9; // >0 light; <0 night
pdcnn_para_init(&PVD_params, pvd_light_night);
vdcnn_para_init(&PVD_params, pvd_light_night);
if (pvd_light_night < 0) { //night
PVD_params.cls_score_thresh = PVD_params.cls_score_thresh - 0.2;
printf("PVD is in night model now\n");
}
PVD_LAYER_MEM PVD_det_mem;
PVD_det_mem.swap_data = (FLOAT *)malloc(PVD_params.size_swap_data);
PVD_det_mem.class_score = (FLOAT *)malloc(PVD_params.size_class_score);
PVD_det_mem.predicts_init = (PVD_PREDICT *)malloc(PVD_params.size_predicts_init);
PVD_det_mem.predicts_initVD = (PVD_PREDICT *)malloc(PVD_params.size_predicts_init);
// PVD_det_mem.final_pred = (PVDCNN_RESULT*)malloc(size_final_pred);
VENDOR_AIS_FLOW_MEM_PARM tmp;
MEM_PARM input_mem_parm2;
ret = hdal_mem_alloc(&input_mem_parm2, "input_mem_parm", PVD_params.size_final_pred);
if (HD_OK != ret) {
printf("PVD_det_mem.final_pred_mem allocate!!\n");
return 0;
}
tmp.pa = input_mem_parm2.pa;
tmp.va = input_mem_parm2.va;
tmp.size = input_mem_parm2.size;
PVD_det_mem.final_pred = (PVDCNN_RESULT *)tmp.va;
#if LIMIT_PVD
VENDOR_AIS_FLOW_MEM_PARM tmp2;
MEM_PARM input_mem_parm3;
ret = hdal_mem_alloc(&input_mem_parm3, "input_mem_parm", PVD_params.size_final_pred);
if (HD_OK != ret) {
printf("PVD_det_mem.final_pred_mem allocate!!\n");
return 0;
}
tmp2.pa = input_mem_parm3.pa;
tmp2.va = input_mem_parm3.va;
tmp2.size = input_mem_parm3.size;
PVDCNN_RESULT *out_pred = (PVDCNN_RESULT *)tmp2.va;
// memset(out_pred, 0, size_final_pred);
// CHAR SAVE_TXT2[256];
// FILE *fs2;
// sprintf(SAVE_TXT2, "/mnt/sd/det_results/cls_debug.txt");
// fs2 = fopen(SAVE_TXT2, "w+");
#endif
memset(PVD_det_mem.swap_data, 0, PVD_params.size_swap_data);
memset(PVD_det_mem.class_score, 0, PVD_params.size_class_score);
memset(PVD_det_mem.predicts_init, 0, PVD_params.size_predicts_init);
memset(PVD_det_mem.predicts_initVD, 0, PVD_params.size_predicts_init);
// memset(PVD_det_mem.final_pred, 0, size_final_pred);
//scale buf
MEM_PARM scale_buf = {0};
ret = model_hdal_mem_get(&scale_buf, (UINT32)SCALE_BUF_SIZE);
if (ret != HD_OK) {
printf("scale mem get fail (%d)!!\r\n", ret);
goto exit_thread;
}
NET_IN nn_in;
//batch test
UINT32 net_all_time = 0, post_all_time = 0;
UINT32 *posttime;
posttime = (UINT32 *)malloc(sizeof(posttime));
*posttime = 0;
UINT32 *nmspost;
nmspost = (UINT32 *)malloc(sizeof(nmspost));
*nmspost = 0;
INT32 img_num = 0;
INT32 instances = 0;
#if SAVE_SCALE
CHAR BMP_FILE[256];
#endif
CHAR IMG_PATH[256];
//CHAR IMG_FILE[512];
CHAR SAVE_TXT1[256];
//CHAR SAVE_TXT2[256];
//CHAR SAVE_BIN[256];
CHAR IMG_LIST[256];
CHAR list_infor[256];
CHAR *line_infor;
BOOL INPUT_STATE = TRUE;
sprintf(IMG_LIST, "/mnt/sd/jpg/pvdlist.txt");
sprintf(IMG_PATH, "/mnt/sd/jpg/PVD");
FILE *fr, *fs1;
sprintf(SAVE_TXT1, "/mnt/sd/det_results/PVDresults.txt");
fr = fopen(IMG_LIST, "r");
//fp = fopen(SAVE_BIN, "wb+");
fs1 = fopen(SAVE_TXT1, "w");
INT32 len = 0;
CHAR img_name[256] = {0};
CHAR *token;
INT32 sl = 0;
BOOL iseresize = FALSE;
//input image parameters
//nn_in.src_img.sign = MAKEFOURCC('A','B','U','F');
//nn_in.src_img.channel = 2;
//nn_in.src_img.fmt = HD_VIDEO_PXLFMT_YUV420;
//printf("nn_in.src_img_size: %d %d %d\r\n", nn_in.w, nn_in.h, nn_in.src_img.size);
INT32 numimg = 0;
char imgname[300];
if (NULL == fr) {
printf("Failed to open img_list!\r\n");
}
while (fgets(list_infor, 256, fr) != NULL) {
len = strlen(list_infor);
list_infor[len - 1] = '\0';
sl = 0;
line_infor = list_infor;
while ((token = strtok(line_infor, " ")) != NULL) {
if (sl > 2) {
break;
}
if (sl == 0) {
strcpy(img_name, token);
sprintf(nn_in.input_filename, "%s/%s", IMG_PATH, token);
sprintf(imgname, "%s", token);
// printf("%s ", token);
}
if (sl == 1) {
nn_in.w = atoi(token);
//nn_in.src_img.width = nn_in.w;
nn_in.loff = ALIGN_CEIL_4(nn_in.w);
//nn_in.loff = 8; // //bin
// nn_in.loff = nn_in.w;
//nn_in.src_img.line_ofs = nn_in.loff;
// printf("%d ",nn_in.loff);
// printf("%s ", token);
}
if (sl == 2) {
nn_in.h = atoi(token);
//nn_in.src_img.height = nn_in.h;
// printf("%s\r\n", token);
}
line_infor = NULL;
sl++;
}
numimg++;
nn_in.c = 2;
// printf("nn_in.c = %d \n",nn_in.c);
if ((ret = input_open(&nn_in)) != HD_OK) {
printf("proc_id(%u) input image open fail !!\n", proc_id);
goto exit_thread;
}
UINT32 img_width_ori = nn_in.w;
UINT32 img_height_ori = nn_in.h;
iseresize = need_ise_resize(nn_in, PVD_params.input_size);
if (debuginf) {
printf("img_name: %s, img_num: %d, w: %ld, h: %ld, isresize: %ld\n", img_name, img_num, img_width_ori, img_height_ori, iseresize);
}
if (iseresize) {
gfx_img.dim.w = PVD_params.input_size[0];
gfx_img.dim.h = PVD_params.input_size[1];
gfx_img.format = nn_in.src_img.fmt;
gfx_img.lineoffset[0] = ALIGN_CEIL_4(PVD_params.input_size[0]);
gfx_img.lineoffset[1] = ALIGN_CEIL_4(PVD_params.input_size[0]);
gfx_img.p_phy_addr[0] = scale_buf.pa;
gfx_img.p_phy_addr[1] = scale_buf.pa + PVD_params.input_size[0] * PVD_params.input_size[1];
//printf("p_phy_addr[0] = 0x%x, p_phy_addr[1] = 0x%x\r\n", gfx_img.p_phy_addr[0], gfx_img.p_phy_addr[1]);
ret = ai_crop_img(&gfx_img, nn_in.src_img, HD_GFX_SCALE_QUALITY_BILINEAR);
if (ret != HD_OK) {
printf("ai_crop_img fail=%d\n", ret);
}
gfx_img_to_vendor(&p_src_img, &gfx_img, (UINT32)scale_buf.va);
#if SAVE_SCALE
FILE *fb;
sprintf(BMP_FILE, "/mnt/sd/save_bmp/%s_scale.bin", img_name);
fb = fopen(BMP_FILE, "wb+");
fwrite((UINT32 *)scale_buf.va, sizeof(UINT32), (gfx_img.dim.h * gfx_img.dim.w + gfx_img.dim.h * gfx_img.dim.w / 2), fb);
fclose(fb);
#endif
} else {
p_src_img = nn_in.src_img;
p_src_img.width = nn_in.w;
p_src_img.height = nn_in.h;
p_src_img.channel = nn_in.c;
p_src_img.line_ofs = nn_in.loff;
p_src_img.fmt = HD_VIDEO_PXLFMT_YUV420;
p_src_img.pa = nn_in.input_mem.pa;
p_src_img.va = nn_in.input_mem.va;
p_src_img.sign = MAKEFOURCC('A', 'B', 'U', 'F');
p_src_img.size = nn_in.src_img.size;
}
// printf("input w:%ld, h:%ld\n", p_src_img.width, p_src_img.height);
// set input image
ret = vendor_ai_net_set(proc_id, VENDOR_AI_NET_PARAM_IN(0, 0), &p_src_img);
if (HD_OK != ret) {
printf("proc_id(%u) push input fail !!\n", proc_id);
continue;
}
// do net proc
gettimeofday(&tstart, NULL);
ret = vendor_ai_net_proc(proc_id);
if (HD_OK != ret) {
printf("proc_id(%u) do net proc fail !!\n", proc_id);
continue;
}
gettimeofday(&tend, NULL);
// printf("after vendor_ai_net_proc \n");
net_all_time += (tend.tv_sec - tstart.tv_sec) * 1000000 + (tend.tv_usec - tstart.tv_usec);
// do post proc
gettimeofday(&tstart, NULL);
for (UINT32 i = 0; i < outlayer_num; i++) {
hd_common_mem_flush_cache((VOID *)layer_buffer[i].va, layer_buffer[i].size);
UINT32 length = layer_buffer[i].width * layer_buffer[i].height * layer_buffer[i].channel * layer_buffer[i].batch_num;
ret = vendor_ai_cpu_util_fixed2float((VOID *)layer_buffer[i].va, layer_buffer[i].fmt, out_layer_float[i], layer_buffer[i].scale_ratio, length);
}
INT32 obj_num = pvdcnn_detectionsample_armneon(layer_buffer, &PVD_params, out_layer_float, &PVD_det_mem, 0, 1, 1, posttime, nmspost);
gettimeofday(&tend, NULL);
post_all_time += (tend.tv_sec - tstart.tv_sec) * 1000000 + (tend.tv_usec - tstart.tv_usec);
PVDCNN_RESULT *objs_info = PVD_det_mem.final_pred;
INT32 prenum = obj_num;
for (INT32 num = 0; num < obj_num; num++) {
// printf("(img_width_ori: %d img_height_ori: %d [%f %f %f %f])\r\n", img_width_ori, img_height_ori, objs_info[num].xmin, objs_info[num].ymin, objs_info[num].xmax, objs_info[num].ymax);
FLOAT x1 = objs_info[num].x1 * img_width_ori;
FLOAT y1 = objs_info[num].y1 * img_height_ori;
FLOAT x2 = objs_info[num].x2 * img_width_ori;
FLOAT y2 = objs_info[num].y2 * img_height_ori;
INT32 category = objs_info[num].category;
FLOAT score = objs_info[num].score;
objs_info[num].x1 = x1;
objs_info[num].y1 = y1;
objs_info[num].x2 = x2;
objs_info[num].y2 = y2;
objs_info[num].category = category;
objs_info[num].score = score;
// printf("obj information is: (classtype: %d socre: %f [%f %f %f %f])\r\n", category, score, x1, y1, x2, y);
//fprintf(fs1, "%s %d %f %f %f %f %f\r\n", img_name, category, score, x1, y1, x2, y2);
}
// printf("")
// max_distance_mode 0;
if (max_distance_mode) {
INT32 input_size[] = { 416, 416 };
MEM_PARM scale_buf_roi = {0};
ret = hdal_mem_alloc(&scale_buf_roi, "scale_buf_roi", SCALE_BUF_SIZE);
if (HD_OK != ret) {
printf("scale_buf_roi get fail (%d)!!\r\n", ret);
goto exit_thread;
}
// printf("before ROI_MULTI scale_buf_roi\n");
// ret = model_hdal_mem_get(&scale_buf_roi, (UINT32)SCALE_BUF_SIZE);
// printf("after ROI_MULTI scale_buf_roi\n");
HD_GFX_IMG_BUF scale_img = { 0 };
scale_img.dim.w = input_size[0];
scale_img.dim.h = input_size[1];
scale_img.format = p_src_img.fmt;
scale_img.lineoffset[0] = ALIGN_CEIL_4(input_size[0]);
scale_img.lineoffset[1] = ALIGN_CEIL_4(input_size[1]);
scale_img.p_phy_addr[0] = scale_buf_roi.pa;
scale_img.p_phy_addr[1] = scale_buf_roi.pa + input_size[0] * input_size[1];
VENDOR_AI_BUF p_img = { 0 };
FLOAT roi_center_value[4][2] = {{0.3046875, 0.3125}, {0.6953125, 0.3125}, {0.3046875, 0.6875}, {0.6953125, 0.6875}}; //(cx, cy)
FLOAT roisize[2] = { PVDROI_MODE3_WIDTH, PVDROI_MODE3_HEIGHT };
INT32 roi_num = 0;
if (max_distance_mode == 1) {
roi_num = 1;
roi_center_value[0][0] = 0.5;
roi_center_value[0][1] = 0.466;
roisize[0] = PVDROI_MODE1_WIDTH;
roisize[1] = PVDROI_MODE1_HEIGHT;
} else if (max_distance_mode == 2) {
roi_num = 2;
roi_center_value[0][0] = 0.32421875;
roi_center_value[0][1] = 0.46527778;
roi_center_value[1][0] = 0.67578125;
roi_center_value[1][1] = 0.46527778;
roisize[0] = PVDROI_MODE2_WIDTH;
roisize[1] = PVDROI_MODE2_HEIGHT;
} else if (max_distance_mode == 3) {
roi_num = 4;
} else {
printf("ERR: Long distance mode only supports 1, 2 or 3, please reset!!\r\n");
exit(0);
}
PVD_IRECT roi = { 0, 0, (FLOAT)input_size[0], (FLOAT)input_size[1] };
for (INT32 n = 0; n < roi_num; n++) {
roi.x1 = (FLOAT)(roi_center_value[n][0] - (FLOAT)roisize[0] / 2.0) * img_width_ori;
roi.y1 = (FLOAT)(roi_center_value[n][1] - (FLOAT)roisize[1] / 2.0) * img_height_ori;
roi.x2 = (FLOAT)(roi_center_value[n][0] + (FLOAT)roisize[0] / 2.0) * img_width_ori;
roi.y2 = (FLOAT)(roi_center_value[n][1] + (FLOAT)roisize[1] / 2.0) * img_height_ori;
ret = pdcnn_crop_img(&scale_img, &p_src_img, HD_GFX_SCALE_QUALITY_NULL, &roi);
if (ret != HD_OK) {
printf("pdcnn_crop_img fail=%d\n", ret);
}
pd_gfx_img_to_vendor(&p_img, &scale_img, scale_buf_roi.va);
#if 0
CHAR YUVFILE[256];
FILE *fb;
sprintf(YUVFILE, "/mnt/sd/det_results/AI/yuv/%s_%09ld_%ld.bin", imgname, numimg, n);
fb = fopen(YUVFILE, "wb+");
fwrite((UINT8 *)scale_buf_roi.va, sizeof(UINT8), (416 * 416 * 3 / 2), fb);
fclose(fb);
#endif
ret = vendor_ai_net_set(proc_id, VENDOR_AI_NET_PARAM_IN(0, 0), &p_img);
if (HD_OK != ret) {
printf("proc_id(%u) push input fail !!\n", proc_id);
continue;
}
//printf("set input image\n");
// do net proc
// static struct timeval tstart, tend;
// gettimeofday(&tstart, NULL);
ret = vendor_ai_net_proc(proc_id);
if (HD_OK != ret) {
printf("proc_id(%u) do net proc fail !!\n", proc_id);
continue;
}
// gettimeofday(&tend, NULL);
//printf("after vendor_ai_net_proc \n");
// (*net_all_time) += (tend.tv_sec - tstart.tv_sec) * 1000000 + (tend.tv_usec - tstart.tv_usec);
// do post proc
// gettimeofday(&tstart, NULL);
for (UINT32 i = 0; i < outlayer_num; i++) {
hd_common_mem_flush_cache((VOID *)layer_buffer[i].va, layer_buffer[i].size);
UINT32 length = layer_buffer[i].width * layer_buffer[i].height * layer_buffer[i].channel * layer_buffer[i].batch_num;
ret = vendor_ai_cpu_util_fixed2float((VOID *)layer_buffer[i].va, layer_buffer[i].fmt, out_layer_float[i], layer_buffer[i].scale_ratio, length);
}
//printf("PVD_detection_out\n");
INT32 obj_num = pvdcnn_detectionsample_armneon(layer_buffer, &PVD_params, out_layer_float, &PVD_det_mem, prenum, 1, 1, posttime, nmspost);
// gettimeofday(&tend, NULL);
// (*post_all_time) += (tend.tv_sec - tstart.tv_sec) * 1000000 + (tend.tv_usec - tstart.tv_usec);
for (INT32 num = 0; num < obj_num; num++) {
FLOAT x1 = objs_info[prenum + num].x1 * roisize[0] * img_width_ori; //p_img.width;
FLOAT y1 = objs_info[prenum + num].y1 * roisize[1] * img_height_ori; //p_img.height;
FLOAT x2 = objs_info[prenum + num].x2 * roisize[0] * img_width_ori; //p_img.width;
FLOAT y2 = objs_info[prenum + num].y2 * roisize[1] * img_height_ori; //p_img.height;
INT32 category = objs_info[prenum + num].category;
FLOAT score = objs_info[prenum + num].score;
//printf("max_distance_mode [%f %f %f %f], roi.x1: %f, roi.y1: %f\n", x1, y1, x2, y2, roi.x1, roi.y1);
objs_info[prenum + num].x1 = x1 + roi.x1;
objs_info[prenum + num].y1 = y1 + roi.y1;
objs_info[prenum + num].x2 = x2 + roi.x1;
objs_info[prenum + num].y2 = y2 + roi.y1;
objs_info[prenum + num].category = category;
objs_info[prenum + num].score = score;
printf("max_distance_mode prenum: %ld, obj_num:%ld, obj information is: (category: %d socre: %f [%f %f %f %f])\r\n", prenum, obj_num, category, score, x1 + roi.x1, y1 + roi.y1, x2 + roi.x1, y2 + roi.y1);
}
prenum += (INT32)obj_num;
}
obj_num = pvdcnn_nonmax_suppress(objs_info, prenum, PVD_params.nms_threshold, 1, 30);
//printf("after nms, prenum: %ld, limit_para->limit_fdet: %ld\n", prenum, limit_para->limit_fdet);
hdal_mem_free(&scale_buf_roi);
}
UINT8 pvd_cls;
#if LIMIT_PVD
pvd_cls = 9;
memset(out_pred, 0, PVD_params.size_final_pred);
//printf("LIMIT_PVD\n");
limit_fdet_params.ratiow = 1.0; //nn_in.src_img.width / (FLOAT)PD_YUV_WIDTH;
limit_fdet_params.ratioh = 1.0; //nn_in.src_img.height / (FLOAT)PD_YUV_HEIGHT;
limit_fdet_params.cls_id = 0;
limit_fdet_params.pvdcls = 1;
limit_fdet_params.proc_num = obj_num;
VENDOR_AIS_FLOW_MEM_PARM input_mem;
MEM_PARM input_mem_parm;
UINT32 input_mem_size = LIMIT_FDET_WIDTH * LIMIT_FDET_WIDTH * 3 / 2;
ret = hdal_mem_alloc(&input_mem_parm, "input_mem_parm", input_mem_size);
if (HD_OK != ret) {
printf("cls input_mem allocate!!\n");
continue;
}
input_mem.pa = input_mem_parm.pa;
input_mem.va = input_mem_parm.va;
input_mem.size = input_mem_parm.size;
// fprintf(fs2, "%s \t\n", img_name);
LIMIT_FDET_MEM limit_fdet_mem = { 0 };
limit_fdet_mem.det_in_mem = tmp; //pdcnn_mem.out_result;
limit_fdet_mem.det_out_mem = tmp2;
limit_fdet_mem.input_mem = input_mem;
ret = limit_fdet_process(&limit_fdet_params, &limit_fdet_mem, &nn_in.src_img);
// ret = limit_pvdet_process(&limit_fdet_params, &limit_pvdet_mem, &nn_in.src_img, limit_fdet_params.score_thres);
//fprintf(fs2, "\n");
// printf("after limit_fdet_process\n");
if (ret != HD_OK) {
printf("ERR: pd limit_fdet_process fail (%ld)!\r\n", ret);
goto exit_thread;
}
//objs_info = out_pred;
obj_num = limit_fdet_params.proc_num;
hdal_mem_free(&input_mem_parm);
if (debuginf) {
printf("before cls limit %d\n", obj_num);
}
if (PVD_params.pd_limit_param.limit_module != 0)
obj_num = dynamic_limit(obj_num, &PVD_params.pd_limit_param, PVD_params.cls_score_thresh, pvd_cls,
out_pred, 1, img_width_ori, img_height_ori);
if (PVD_params.vd_limit_param.limit_module != 0)
obj_num = dynamic_limit(obj_num, &PVD_params.vd_limit_param, PVD_params.cls_score_thresh, pvd_cls,
out_pred, 2, img_width_ori, img_height_ori);
if (debuginf) {
printf("after cls limit %d\n", obj_num);
}
for (INT32 num = 0; num < obj_num; num++) {
FLOAT xmin = out_pred[num].x1;
FLOAT ymin = out_pred[num].y1;
FLOAT xmax = out_pred[num].x2;
FLOAT ymax = out_pred[num].y2;
INT32 category = out_pred[num].category;
FLOAT score = out_pred[num].score;
if (debuginf) {
printf("finally cls num: %ld obj information is: (category: %d socre: %f [%f %f %f %f])\r\n", obj_num, category, score, xmin, ymin, xmax, ymax);
}
fprintf(fs1, "%s %d %f %f %f %f %f\r\n", img_name, category, score, xmin, ymin, xmax, ymax);
}
#else //img_width_ori, img_height_ori
// printf("before limit %d\n", obj_num);
pvd_cls = -9;
if (PVD_params.pd_limit_param.limit_module != 0)
obj_num = dynamic_limit(obj_num, &PVD_params.pd_limit_param, PVD_params.cls_score_thresh, pvd_cls,
objs_info, 1, img_width_ori, img_height_ori);
if (PVD_params.vd_limit_param.limit_module != 0)
obj_num = dynamic_limit(obj_num, &PVD_params.vd_limit_param, PVD_params.cls_score_thresh, pvd_cls,
objs_info, 2, img_width_ori, img_height_ori);
// printf("after limit %d\n", obj_num);
for (INT32 num = 0; num < obj_num; num++) {
FLOAT xmin = objs_info[num].x1;
FLOAT ymin = objs_info[num].y1;
FLOAT xmax = objs_info[num].x2;
FLOAT ymax = objs_info[num].y2;
INT32 category = objs_info[num].category;
FLOAT score = objs_info[num].score;
if (debuginf) {
printf("finally num: %ld obj information is: (category: %d socre: %f [%f %f %f %f])\r\n", obj_num, category, score, xmin, ymin, xmax, ymax);
}
fprintf(fs1, "%s %d %f %f %f %f %f\r\n", img_name, category, score, xmin, ymin, xmax, ymax);
}
#endif
instances += obj_num;
// free buf every time
if ((ret = input_close(&nn_in)) != HD_OK) {
printf("proc_id(%u) input image close fail !!\n", proc_id);
}
img_num++;
if (instances == 0) {
instances = 1;
}
if (debuginf)
printf("img_num: %d, net mean_time: %d, post mean_time: %d, posttime_means: %d, posttime_eachinstance: %d, nmspost_m: %d, nmspost_ecisce: %d\n",
img_num, net_all_time / img_num, post_all_time / img_num, (*posttime) / img_num, (*posttime) / instances, (*nmspost) / img_num, (*nmspost) / instances);
}
printf("tests are done!\n");
if (INPUT_STATE == TRUE) {
printf("all test img number is: %d\r\n", img_num);
}
fclose(fs1);
//fclose(fs2);
fclose(fr);
printf("closed output stream\n");
// #if LIMIT_PVD
// fclose(fs2);
// #endif
printf("net mean_time: %d, post mean_time: %d, posttime_means: %d, posttime_eachinstance: %d, nmspost_m: %d, nmspost_ecisce: %d\n",
net_all_time / img_num, post_all_time / img_num, (*posttime) / img_num, (*posttime) / instances, (*nmspost) / img_num, (*nmspost) / instances);
free(layer_buffer);
for (UINT32 i = 0; i < outlayer_num; i++) {
free(out_layer_float[i]);
}
free(out_layer_float);
free(PVD_det_mem.swap_data);
free(PVD_det_mem.class_score);
free(PVD_det_mem.predicts_init);
free(PVD_det_mem.predicts_initVD);
free(posttime);
free(nmspost);
hdal_mem_free(&input_mem_parm2);
hdal_mem_free(&nn_in.input_mem);
hdal_mem_free(&net_info.io_mem);
hdal_mem_free(&net_info.proc_mem);
#if LIMIT_PVD
hdal_mem_free(&net_info1.io_mem);
hdal_mem_free(&net_info1.proc_mem);
#endif
exit_thread:
// stop
ret = vendor_ai_net_stop(proc_id);
if (HD_OK != ret) {
printf("proc_id(%u) nn stop fail !!\n", proc_id);
}
#if LIMIT_PVD
hdal_mem_free(&input_mem_parm3);
ret = vendor_ai_net_stop(limit_fdet_params.run_id);
if (HD_OK != ret) {
printf("proc_id(%u) nn stop fail !!\n", limit_fdet_params.run_id);
}
if ((ret = network_close(&net_info1)) != HD_OK) {
printf("proc_id(%u) nn close fail !!\n", limit_fdet_params.run_id);
}
#endif
if ((ret = input_close(&nn_in)) != HD_OK) {
printf("proc_id(%u) input image close fail !!\n", proc_id);
}
// close network modules
if ((ret = network_close(&net_info)) != HD_OK) {
printf("proc_id(%u) nn close fail !!\n", proc_id);
}
//need release scale_buf
if ((ret = model_hdal_mem_release(&scale_buf)) != HD_OK) {
//printf("ret: %ld\n", ret);
//printf("HD_ok: %ld\n", HD_OK);
printf("proc_id(%u) mem release fail !!\n", proc_id);
}
// uninit network modules
ret = vendor_ai_uninit();
if (ret != HD_OK) {
printf("vendor_ai_uninit fail=%d\n", ret);
}
return 0;
}
int main(int argc, char *argv[])
{
HD_RESULT ret;
//INT32 key;
NET_IN nn_in;
pthread_t nn_thread_id;
// init hdal
ret = hd_common_init(0);
if (ret != HD_OK) {
printf("hd_common_init fail=%d\n", ret);
goto exit;
}
// set project config for AI
hd_common_sysconfig(0, (1 << 16), 0, VENDOR_AI_CFG); //enable AI engine
ret = hd_gfx_init();
if (ret != HD_OK) {
printf("hd_gfx_init fail\r\n");
goto exit;
}
UINT32 schd = VENDOR_AI_PROC_SCHD_FAIR;
vendor_ai_cfg_set(VENDOR_AI_CFG_PLUGIN_ENGINE, vendor_ai_cpu1_get_engine());
vendor_ai_cfg_set(VENDOR_AI_CFG_PROC_SCHD, &schd);
ret = vendor_ai_init();
if (ret != HD_OK) {
printf("vendor_ai_init fail=%d\n", ret);
goto exit;
}
// create nn_proc_thread
ret = pthread_create(&nn_thread_id, NULL, nn_thread_api, (VOID *)(&nn_in));
if (ret < 0) {
printf("create encode thread failed");
goto exit;
}
/*do {
printf("usage:\n");
printf(" enter q: exit\n");
printf(" enter r: run engine\n");
key = getchar();
if (key == 'q' || key == 0x3) {
is_net_proc = FALSE;
is_net_exit = TRUE;
break;
} else if (key == 'r') {
is_net_proc = TRUE;
printf("[ai sample] nn proc done!\n");
usleep(500000);
continue;
}else if (key == 'd'){ //dkn 20210120
//enter debug menu
hd_debug_run_menu();
printf("\r\nEnter q to exit, Enter d to debug\r\n");
}
} while(1);*/
// wait encode thread destroyed
pthread_join(nn_thread_id, NULL);
exit:
ret = hd_gfx_uninit();
if (ret != HD_OK) {
printf("hd_gfx_uninit fail\r\n");
}
// uninit memory
ret = hd_common_mem_uninit();
if (ret != HD_OK) {
printf("mem fail=%d\n", ret);
}
// uninit hdal
ret = hd_common_uninit();
if (ret != HD_OK) {
printf("common fail=%d\n", ret);
}
return ret;
}
Reference in New Issue View Git Blame Copy Permalink