/*
*
* Copyright (c) 2017 Texas Instruments Incorporated
*
* All rights reserved not granted herein.
*
* Limited License.
*
* Texas Instruments Incorporated grants a world-wide, royalty-free, non-exclusive
* license under copyrights and patents it now or hereafter owns or controls to make,
* have made, use, import, offer to sell and sell ("Utilize") this software subject to the
* terms herein. With respect to the foregoing patent license, such license is granted
* solely to the extent that any such patent is necessary to Utilize the software alone.
* The patent license shall not apply to any combinations which include this software,
* other than combinations with devices manufactured by or for TI ("TI Devices").
* No hardware patent is licensed hereunder.
*
* Redistributions must preserve existing copyright notices and reproduce this license
* (including the above copyright notice and the disclaimer and (if applicable) source
* code license limitations below) in the documentation and/or other materials provided
* with the distribution
*
* Redistribution and use in binary form, without modification, are permitted provided
* that the following conditions are met:
*
* * No reverse engineering, decompilation, or disassembly of this software is
* permitted with respect to any software provided in binary form.
*
* * any redistribution and use are licensed by TI for use only with TI Devices.
*
* * Nothing shall obligate TI to provide you with source code for the software
* licensed and provided to you in object code.
*
* If software source code is provided to you, modification and redistribution of the
* source code are permitted provided that the following conditions are met:
*
* * any redistribution and use of the source code, including any resulting derivative
* works, are licensed by TI for use only with TI Devices.
*
* * any redistribution and use of any object code compiled from the source code
* and any resulting derivative works, are licensed by TI for use only with TI Devices.
*
* Neither the name of Texas Instruments Incorporated nor the names of its suppliers
*
* may be used to endorse or promote products derived from this software without
* specific prior written permission.
*
* DISCLAIMER.
*
* THIS SOFTWARE IS PROVIDED BY TI AND TI'S LICENSORS "AS IS" AND ANY EXPRESS
* OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
* OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
* IN NO EVENT SHALL TI AND TI'S LICENSORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
* BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
* DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
* OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
* OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
* OF THE POSSIBILITY OF SUCH DAMAGE.
*
*/
#include <utils/draw2d/include/draw2d.h>
#include <utils/perf_stats/include/app_perf_stats.h>
#include "itidl_ti.h"
#include "app_common.h"
#include "app_test.h"
#include "mult_yolov5_post_copy.h"
#include "math.h"
/*
* define this macro to enable TIDL intermediate layer traces on target
*/
#undef APP_TIDL_TRACE_DUMP
/*
* This is the size of trace buffer allocated in host memory and
* shared with target.
*/
#define TIVX_TIDL_TRACE_DATA_SIZE (128 * 1024 * 1024)
extern const char imgnet_labels[1001][256];
static const char *tensor_num_str[] = { "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "10", "11", "12", "13", "14", "15"};
#define APP_DEBUG
#ifdef APP_DEBUG
#define APP_PRINTF(f_, ...) printf((f_), ##VA_ARGS)
#else
#define APP_PRINTF(f_, ...)
#endif
typedef struct {
/* config options */
char tidl_config_file_path[APP_MAX_FILE_PATH];
char tidl_network_file_path[APP_MAX_FILE_PATH];
char input_file_path[APP_MAX_FILE_PATH];
char input_file_list[APP_MAX_FILE_PATH];
char output_file_path[APP_MAX_FILE_PATH];
char ti_logo_file_path[APP_MAX_FILE_PATH];
uint32_t num_input_tensors;
uint32_t num_output_tensors;
/* Input image params */
vx_df_image df_image;
void *data_ptr;
tivx_utils_bmp_image_params_t imgParams;
vx_uint32 img_width;
vx_uint32 img_height;
vx_uint32 img_stride;
sTIDL_IOBufDesc_t ioBufDesc;
uint8_t *pInPlanes;
uint8_t *pOutPlanes;
uint8_t *temp;
char save_type;
uint8_t *pDisplayBuf888;
uint16_t *pDisplayBuf565;
/* OpenVX references */
vx_context context;
vx_graph graph;
vx_kernel kernel;
vx_node tidl_node;
vx_user_data_object config;
vx_user_data_object network;
vx_user_data_object createParams;
vx_user_data_object inArgs;
vx_user_data_object outArgs;
vx_user_data_object traceData;
vx_tensor input_tensors[APP_MAX_TENSORS];
vx_tensor output_tensors[APP_MAX_TENSORS];
vx_graph disp_graph;
vx_node disp_node;
vx_image disp_image;
vx_user_data_object disp_params_obj;
tivx_display_params_t disp_params;
vx_rectangle_t disp_rect;
vx_imagepatch_addressing_t image_addr;
vx_size in_tensor_data_type[APP_MAX_TENSORS];
vx_uint32 display_option;
vx_uint32 delay_in_msecs;
vx_uint32 num_iterations;
Draw2D_Handle pHndl;
uint32_t is_interactive;
vx_int32 test_mode;
vx_int32 test_case;
tivx_task task;
uint32_t stop_task;
uint32_t stop_task_done;
app_perf_point_t total_perf;
app_perf_point_t fileio_perf;
app_perf_point_t draw_perf;
ttevxYoloV5PostProcParams yolov5params;
ttevxYoloV5Detections detections;
TIDL_outArgs out_args;
} AppObj;
AppObj gAppObj;
static int app_parse_cmd_line_args(AppObj *obj, int argc, char *argv[]);
static vx_status app_init(AppObj *obj);
static void app_deinit(AppObj *obj);
static vx_status app_create_graph(AppObj *obj);
static vx_status app_verify_graph(AppObj *obj);
static vx_status app_run_graph(AppObj *obj);
static vx_status app_run_graph_interactive(AppObj *obj);
static void app_delete_graph(AppObj *obj);
static vx_user_data_object readConfig(AppObj *obj, vx_context context, char *config_file, uint32_t *num_input_tensors, uint32_t *num_output_tensors);
static vx_user_data_object readNetwork(vx_context context, char *network_file);
static vx_user_data_object setCreateParams(vx_context context);
static vx_user_data_object setInArgs(vx_context context);
static vx_user_data_object setOutArgs(vx_context context);
static void createInputTensors(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *input_tensors);
static void createOutputTensors(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *output_tensors);
static vx_status readInput(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *input_tensors, char *input_file);
static void displayOutput(AppObj *obj, vx_user_data_object config, vx_tensor *output_tensors, char *output_file);
static vx_size getTensorDataType(vx_int32 tidl_type);
static const char* vxTypeToStr(vx_enum type);
static vx_size vxTypeSize(vx_enum type);
// #define APP_WRITE_PRE_PROC_OUTPUT;
#ifdef APP_WRITE_PRE_PROC_OUTPUT
static vx_status writePreProcOutput(char* file_name, vx_tensor output);
#endif
static uint32_t num_params;
static uint32_t max_params;
static const char* vxTypeToStr(vx_enum type)
{
switch(type)
{
case VX_TYPE_INT8: return "VX_TYPE_INT8";
case VX_TYPE_UINT8: return "VX_TYPE_UINT8";
case VX_TYPE_INT16: return "VX_TYPE_INT16";
case VX_TYPE_UINT16: return "VX_TYPE_UINT16";
default: return "VX_TYPE_UNKNOWN";
}
}
static vx_size vxTypeSize(vx_enum type)
{
switch(type)
{
case VX_TYPE_INT8:
case VX_TYPE_UINT8:
return sizeof(vx_uint8);
case VX_TYPE_INT16:
case VX_TYPE_UINT16:
return sizeof(vx_uint16);
default:
return 1;
}
}
int app_tidl_main(int argc, char* argv[])
{
int status = 0;
AppObj *obj = &gAppObj;
printf("hello: %s\n", argv[0]);
gAppObj.save_type = '0';
status = app_parse_cmd_line_args(obj, argc, argv);
if(status == VX_SUCCESS)
{
status = app_init(obj);
for(int i = 0;i<5;i++)
{
obj->out_args.scale[i] = obj->ioBufDesc.outTensorScale[i];
}
}
if(status == VX_SUCCESS)
{
status = app_create_graph(obj);
}
if(status == VX_SUCCESS)
{
status = app_verify_graph(obj);
}
if(status == VX_SUCCESS)
{
if(obj->is_interactive)
{
status = app_run_graph_interactive(obj);
}
else
{
status = app_run_graph(obj);
if(status == VX_FAILURE)
{
printf("Error processing graph!\n");
}
}
}
app_delete_graph(obj);
app_deinit(obj);
if(obj->test_mode == 1)
{
if((test_result == vx_false_e) || (status != VX_SUCCESS))
{
printf("\n\nTEST FAILED\n\n");
print_new_checksum_structs();
status = (status == VX_SUCCESS) ? VX_FAILURE : status;
}
else
{
printf("\n\nTEST PASSED\n\n");
}
}
return status;
}
static int app_init(AppObj *obj)
{
int status = 0;
uint32_t num_input_tensors = 0;
uint32_t num_output_tensors = 0;
APP_PRINTF("app_tidl: Init ... \n");
obj->context = vxCreateContext();
APP_ASSERT_VALID_REF(obj->context);
/* Create a vx_array object and read the config data*/
obj->config = readConfig(obj, obj->context, &obj->tidl_config_file_path[0], &num_input_tensors, &num_output_tensors);
APP_ASSERT_VALID_REF(obj->config)
/* Save a copy of number of input/output tensors required as per config */
obj->num_input_tensors = num_input_tensors;
obj->num_output_tensors = num_output_tensors;
/* Create a vx_tensor object and read the network data */
obj->network = readNetwork(obj->context, &obj->tidl_network_file_path[0]);
APP_ASSERT_VALID_REF(obj->network)
obj->createParams = setCreateParams(obj->context);
APP_ASSERT_VALID_REF(obj->createParams)
obj->inArgs = setInArgs(obj->context);
APP_ASSERT_VALID_REF(obj->inArgs)
obj->outArgs = setOutArgs(obj->context);
APP_ASSERT_VALID_REF(obj->outArgs)
#ifdef APP_TIDL_TRACE_DUMP
obj->traceData = vxCreateUserDataObject(obj->context, "TIDL_traceData", TIVX_TIDL_TRACE_DATA_SIZE, NULL);
APP_ASSERT_VALID_REF(obj->traceData)
#endif
obj->kernel = tivxAddKernelTIDL(obj->context, num_input_tensors, num_output_tensors);
APP_ASSERT_VALID_REF(obj->kernel)
if ((vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1)) && (obj->display_option == 1))
{
obj->disp_image = vxCreateImage(obj->context, DISPLAY_WIDTH, DISPLAY_HEIGHT, VX_DF_IMAGE_RGB);
APP_ASSERT_VALID_REF(obj->disp_image)
obj->image_addr.dim_x = DISPLAY_WIDTH;
obj->image_addr.dim_y = DISPLAY_HEIGHT;
obj->image_addr.stride_x = 3; /* RGB */
obj->image_addr.stride_y = DISPLAY_WIDTH * 3;
obj->image_addr.scale_x = VX_SCALE_UNITY;
obj->image_addr.scale_y = VX_SCALE_UNITY;
obj->image_addr.step_x = 1;
obj->image_addr.step_y = 1;
obj->disp_rect.start_x = 0;
obj->disp_rect.start_y = 0;
obj->disp_rect.end_x = DISPLAY_WIDTH;
obj->disp_rect.end_y = DISPLAY_HEIGHT;
memset(&obj->disp_params, 0, sizeof(tivx_display_params_t));
obj->disp_params.opMode = TIVX_KERNEL_DISPLAY_BUFFER_COPY_MODE;
obj->disp_params.pipeId = 0;
obj->disp_params.outWidth = DISPLAY_WIDTH;
obj->disp_params.outHeight = DISPLAY_HEIGHT;
obj->disp_params.posX = (1920-DISPLAY_WIDTH)/2;
obj->disp_params.posY = (1080-DISPLAY_HEIGHT)/2;
obj->disp_params_obj = vxCreateUserDataObject(obj->context, "tivx_display_params_t", sizeof(tivx_display_params_t), &obj->disp_params);
APP_ASSERT_VALID_REF(obj->disp_params_obj)
tivxHwaLoadKernels(obj->context);
}
tivxTIDLLoadKernels(obj->context);
obj->temp = tivxMemAlloc(MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3 * 3, TIVX_MEM_EXTERNAL);
obj->pInPlanes = tivxMemAlloc(MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3, TIVX_MEM_EXTERNAL);
if(obj->pInPlanes == NULL) {
printf("app_tidl: ERROR: Unable to allocate memory for inPlanes, size = %d\n", MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3);
status = -1;
}
obj->pOutPlanes = tivxMemAlloc(MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3, TIVX_MEM_EXTERNAL);
if(obj->pOutPlanes == NULL) {
printf("app_tidl: ERROR: Unable to allocate memory for outPlanes, size = %d\n", MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3);
status = -1;
}
obj->pDisplayBuf888 = tivxMemAlloc(DISPLAY_WIDTH * DISPLAY_HEIGHT * 3, TIVX_MEM_EXTERNAL);
if(obj->pDisplayBuf888 == NULL) {
printf("app_tidl: ERROR: Unable to allocate memory for displayBuf888, size = %d\n", DISPLAY_WIDTH * DISPLAY_HEIGHT * 3);
status = -1;
}
obj->pDisplayBuf565 = tivxMemAlloc(DISPLAY_WIDTH * DISPLAY_HEIGHT * sizeof(uint16_t), TIVX_MEM_EXTERNAL);
if(obj->pDisplayBuf565 == NULL) {
printf("app_tidl: ERROR: Unable to allocate memory for displayBuf565, size = %ld\n", DISPLAY_WIDTH * DISPLAY_HEIGHT * sizeof(uint16_t));
status = -1;
}
{
Draw2D_BufInfo sBufInfo;
Draw2D_LinePrm sLinePrm;
Draw2D_FontPrm sTop5Prm;
char banner_file[APP_MAX_FILE_PATH];
snprintf(banner_file, APP_MAX_FILE_PATH, "%s/ti_logo.bmp", obj->ti_logo_file_path);
Draw2D_create(&obj->pHndl);
if(obj->pHndl != NULL)
{
sBufInfo.bufWidth = DISPLAY_WIDTH;
sBufInfo.bufHeight = DISPLAY_HEIGHT;
sBufInfo.bufPitch[0] = DISPLAY_WIDTH * 2;
sBufInfo.dataFormat = DRAW2D_DF_BGR16_565;
sBufInfo.transperentColor = 0;
sBufInfo.transperentColorFormat = DRAW2D_DF_BGR16_565;
sBufInfo.bufAddr[0] = (uint8_t *)obj->pDisplayBuf565;
Draw2D_setBufInfo(obj->pHndl, &sBufInfo);
Draw2D_clearBuf(obj->pHndl);
Draw2D_insertBmp(obj->pHndl, banner_file, 0, 0);
sLinePrm.lineColor = RGB888_TO_RGB565(255, 255, 255);
sLinePrm.lineSize = 3;
sLinePrm.lineColorFormat = DRAW2D_DF_BGR16_565;
/* Draw a vertial line */
Draw2D_drawLine(obj->pHndl, DISPLAY_WIDTH/2, 0, DISPLAY_WIDTH/2, DISPLAY_HEIGHT, &sLinePrm);
/* green color heading */
Draw2D_setFontColor(RGB888_TO_RGB565(0, 255, 0), RGB888_TO_RGB565(0, 0, 0), RGB888_TO_RGB565(0, 0, 0));
sTop5Prm.fontIdx = 0;
Draw2D_drawString(obj->pHndl, (DISPLAY_WIDTH/2) + 40, 140, "Top 5 classes", &sTop5Prm);
Draw2D_resetFontColor();
}
}
appPerfPointSetName(&obj->total_perf , "TOTAL");
appPerfPointSetName(&obj->draw_perf , "DRAW");
appPerfPointSetName(&obj->fileio_perf, "FILEIO");
APP_PRINTF("app_tidl: Init ... Done.\n");
return status;
}
static void app_deinit(AppObj *obj)
{
APP_PRINTF("app_tidl: De-init ... \n");
tivxTIDLUnLoadKernels(obj->context);
if (vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1) && (obj->display_option == 1))
{
tivxHwaUnLoadKernels(obj->context);
}
vxRemoveKernel(obj->kernel);
vxReleaseContext(&obj->context);
tivxMemFree(obj->temp, MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3 * 3, TIVX_MEM_EXTERNAL);
tivxMemFree(obj->pInPlanes, MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3, TIVX_MEM_EXTERNAL);
tivxMemFree(obj->pOutPlanes, MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3, TIVX_MEM_EXTERNAL);
tivxMemFree(obj->pDisplayBuf888, DISPLAY_WIDTH * DISPLAY_HEIGHT * 3, TIVX_MEM_EXTERNAL);
tivxMemFree(obj->pDisplayBuf565 , DISPLAY_WIDTH * DISPLAY_HEIGHT * sizeof(uint16_t), TIVX_MEM_EXTERNAL);
{
Draw2D_delete(obj->pHndl);
}
APP_PRINTF("app_tidl: De-init ... Done.\n");
}
static void app_delete_graph(AppObj *obj)
{
uint32_t id;
APP_PRINTF("app_tidl: Delete ... \n");
if (vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1) && (obj->display_option == 1))
{
vxReleaseNode(&obj->disp_node);
vxReleaseGraph(&obj->disp_graph);
}
#ifdef APP_TIVX_LOG_RT_ENABLE
tivxLogRtTraceExportToFile("app_tidl.bin");
tivxLogRtTraceDisable(obj->graph);
#endif
vxReleaseNode(&obj->tidl_node);
vxReleaseGraph(&obj->graph);
vxReleaseUserDataObject(&obj->config);
vxReleaseUserDataObject(&obj->network);
vxReleaseUserDataObject(&obj->createParams);
vxReleaseUserDataObject(&obj->inArgs);
vxReleaseUserDataObject(&obj->outArgs);
#ifdef APP_TIDL_TRACE_DUMP
vxReleaseUserDataObject(&obj->traceData);
#endif
for(id = 0; id < obj->num_input_tensors; id++)
{
vxReleaseTensor(&obj->input_tensors[id]);
}
for(id = 0; id < obj->num_output_tensors; id++)
{
vxReleaseTensor(&obj->output_tensors[id]);
}
if ((vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1)) && (obj->display_option == 1))
{
vxReleaseImage(&obj->disp_image);
vxReleaseUserDataObject(&obj->disp_params_obj);
}
APP_PRINTF("app_tidl: Delete ... Done.\n");
}
static void app_show_usage(int argc, char* argv[])
{
printf("\n");
printf(" TIDL Demo - (c) Texas Instruments 2018\n");
printf(" ========================================================\n");
printf("\n");
printf(" Usage,\n");
printf(" %s --cfg <config file>\n", argv[0]);
printf("\n");
}
static void app_set_cfg_default(AppObj *obj)
{
snprintf(obj->tidl_config_file_path,APP_MAX_FILE_PATH, "test_data/tivx/tidl_models/mobilenetv1/config.bin");
snprintf(obj->tidl_network_file_path,APP_MAX_FILE_PATH, "test_data/tivx/tidl_models/mobilenetv1/network.bin");
snprintf(obj->input_file_path,APP_MAX_FILE_PATH, "test_data/psdkra/app_tidl");
snprintf(obj->input_file_list,APP_MAX_FILE_PATH, "test_data/psdkra/app_tidl/names.txt");
snprintf(obj->output_file_path,APP_MAX_FILE_PATH, "app_tidl_out");
snprintf(obj->ti_logo_file_path,APP_MAX_FILE_PATH, "test_data/tivx/tidl_models/");
obj->display_option = 1;
obj->delay_in_msecs = 0;
obj->num_iterations = 1;
obj->is_interactive = 0;
obj->test_mode = 0;
}
static int app_parse_cfg_file(AppObj *obj, char *cfg_file_name)
{
FILE *fp = fopen(cfg_file_name, "r");
char line_str[1024];
char *token;
if(fp==NULL)
{
printf("app_tidl: ERROR: Unable to open config file [%s]\n", cfg_file_name);
return -1;
}
while(fgets(line_str, sizeof(line_str), fp)!=NULL)
{
char s[]=" \t";
if (strchr(line_str, '#'))
{
continue;
}
/* get the first token */
token = strtok(line_str, s);
if(token != NULL)
{
if(strcmp(token, "tidl_config_file_path")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->tidl_config_file_path, token);
}
}
else
if(strcmp(token, "tidl_network_file_path")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->tidl_network_file_path, token);
/* for testing if relevant */
if(strstr(obj->tidl_network_file_path, "u16") != NULL)
{
obj->test_case = 1;
}
else
{
obj->test_case = 0;
}
}
}
else
if(strcmp(token, "input_file_path")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->input_file_path, token);
}
}
else
if(strcmp(token, "ti_logo_file_path")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->ti_logo_file_path, token);
}
}
else
if(strcmp(token, "input_file_list")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->input_file_list, token);
}
}
else
if(strcmp(token, "output_file_path")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
strcpy(obj->output_file_path, token);
}
}
else
if(strcmp(token, "display_option")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
obj->display_option = atoi(token);
if(obj->display_option > 1)
obj->display_option = 1;
}
}
else
if(strcmp(token, "delay")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
obj->delay_in_msecs = atoi(token);
if(obj->delay_in_msecs > 2000)
obj->delay_in_msecs = 2000;
}
}
else
if(strcmp(token, "num_iterations")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
obj->num_iterations = atoi(token);
if(obj->num_iterations == 0)
obj->num_iterations = 1;
}
}
else
if(strcmp(token, "is_interactive")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
obj->is_interactive = atoi(token);
if(obj->is_interactive > 1)
obj->is_interactive = 1;
}
}
else
if(strcmp(token, "test_mode")==0)
{
token = strtok(NULL, s);
if(token != NULL)
{
token[strlen(token)-1]=0;
obj->test_mode = atoi(token);
}
}
}
if (obj->test_mode == 1)
{
obj->is_interactive = 0;
/* display_option must be set to 1 in order for the checksums
to come out correctly */
#ifndef x86_64
printf("Turning display option on ... \n");
obj->display_option = 1;
#else
printf("Turning display option off ... \n");
obj->display_option = 0;
#endif
}
}
fclose(fp);
return 0;
}
static int app_parse_cmd_line_args(AppObj *obj, int argc, char *argv[])
{
int i;
vx_bool set_test_mode = vx_false_e;
app_set_cfg_default(obj);
if(argc==1)
{
app_show_usage(argc, argv);
return -1;
}
for(i=0; i<argc; i++)
{
if(strcmp(argv[i], "--cfg")==0)
{
i++;
if(i>=argc)
{
app_show_usage(argc, argv);
}
app_parse_cfg_file(obj, argv[i]);
}
else
if(strcmp(argv[i], "--help")==0)
{
app_show_usage(argc, argv);
return -1;
}
else
if(strcmp(argv[i], "--test")==0)
{
set_test_mode = vx_true_e;
}
}
if (set_test_mode == vx_true_e)
{
obj->test_mode = 1;
obj->is_interactive = 0;
obj->display_option = 1;
obj->delay_in_msecs = 100;
}
#ifdef x86_64
printf("Turning display option off ... \n");
obj->display_option = 0;
obj->is_interactive = 0;
#endif
return 0;
}
static void initParam(vx_reference params[], uint32_t _max_params)
{
num_params = 0;
max_params = _max_params;
}
static void addParam(vx_reference params[], vx_reference obj)
{
APP_ASSERT(num_params <= max_params);
params[num_params] = obj;
num_params++;
}
static vx_status app_create_graph(AppObj *obj)
{
vx_status status = VX_SUCCESS;
vx_reference params[APP_TIDL_MAX_PARAMS];
uint32_t i;
APP_PRINTF("app_tidl: Creating graph ... \n");
/* Create OpenVx Graph */
obj->graph = vxCreateGraph(obj->context);
APP_ASSERT_VALID_REF(obj->graph)
vxSetReferenceName((vx_reference)obj->graph, "app_tidl_graph");
/* Create array of input tensors */
createInputTensors(obj, obj->context, obj->config, obj->input_tensors);
/* Create array of output tensors */
createOutputTensors(obj, obj->context, obj->config, obj->output_tensors);
/* Initialize param array */
initParam(params, APP_TIDL_MAX_PARAMS);
/* The 1st param MUST be config */
addParam(params, (vx_reference)obj->config);
/* The 2nd param MUST be network */
addParam(params, (vx_reference)obj->network);
/* The 3rd param MUST be create params */
addParam(params, (vx_reference)obj->createParams);
/* The 4th param MUST be inArgs */
addParam(params, (vx_reference)obj->inArgs);
/* The 5th param MUST be outArgs */
addParam(params, (vx_reference)obj->outArgs);
#ifdef APP_TIDL_TRACE_DUMP
addParam(params, (vx_reference)obj->traceData);
#else
/* The 6th param MUST be NULL if trace data dump is not enabled */
addParam(params, (vx_reference)NULL);
#endif
/* Create TIDL Node */
obj->tidl_node = tivxTIDLNode(obj->graph, obj->kernel, params, obj->input_tensors, obj->output_tensors);
APP_ASSERT_VALID_REF(obj->tidl_node)
vxSetNodeTarget(obj->tidl_node, VX_TARGET_STRING, TIVX_TARGET_DSP_C7_1);
if ((vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1)) && (obj->display_option == 1))
{
/* Create OpenVx Graph */
obj->disp_graph = vxCreateGraph(obj->context);
APP_ASSERT_VALID_REF(obj->disp_graph)
vxSetReferenceName((vx_reference)obj->disp_graph, "Display");
obj->disp_node = tivxDisplayNode(obj->disp_graph, obj->disp_params_obj, obj->disp_image);
APP_ASSERT_VALID_REF(obj->disp_node)
vxSetNodeTarget(obj->disp_node, VX_TARGET_STRING, TIVX_TARGET_DISPLAY1);
}
/* Set names for diferent OpenVX objects */
vxSetReferenceName((vx_reference)obj->config, "Config");
vxSetReferenceName((vx_reference)obj->network, "Network");
vxSetReferenceName((vx_reference)obj->createParams, "CreateParams");
vxSetReferenceName((vx_reference)obj->inArgs, "InArgs");
vxSetReferenceName((vx_reference)obj->outArgs, "OutArgs");
for(i = 0; i < obj->num_input_tensors; i++) {
char tensor_name[] = "InputTensor_";
char ref_name[64];
snprintf(ref_name, 64, "%s%s", tensor_name, tensor_num_str[i]);
vxSetReferenceName((vx_reference)obj->input_tensors[i], ref_name);
}
for(i = 0; i < obj->num_output_tensors; i++) {
char tensor_name[] = "OutputTensor_";
char ref_name[64];
snprintf(ref_name, 64, "%s%s", tensor_name, tensor_num_str[i]);
vxSetReferenceName((vx_reference)obj->output_tensors[i], ref_name);
}
vxSetReferenceName((vx_reference)obj->kernel, "TIDLKernel");
vxSetReferenceName((vx_reference)obj->tidl_node, "TIDLNode");
if (vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1) && (obj->display_option == 1))
{
vxSetReferenceName((vx_reference)obj->disp_params_obj, "DisplayParams");
vxSetReferenceName((vx_reference)obj->disp_node, "DisplayNode");
}
APP_PRINTF("app_tidl: Creating graph ... Done.\n");
return status;
}
static void app_run_task(void *app_var)
{
AppObj *obj = (AppObj *)app_var;
vx_status status = VX_SUCCESS;
appPerfStatsCpuLoadResetAll();
while(!obj->stop_task)
{
status = app_run_graph(obj);
if(status == VX_FAILURE)
{
printf("Error processing graph!\n");
obj->stop_task = 1;
}
}
obj->stop_task_done = 1;
}
static int32_t app_run_task_create(AppObj *obj)
{
tivx_task_create_params_t params;
int32_t status;
tivxTaskSetDefaultCreateParams(¶ms);
params.task_main = app_run_task;
params.app_var = obj;
obj->stop_task_done = 0;
obj->stop_task = 0;
status = tivxTaskCreate(&obj->task, ¶ms);
return status;
}
static void app_run_task_delete(AppObj *obj)
{
while(obj->stop_task_done==0)
{
tivxTaskWaitMsecs(100);
}
tivxTaskDelete(&obj->task);
}
static char menu[] = {
"\n"
"\n ================================="
"\n Demo : TIDL Object Classification"
"\n ================================="
"\n"
"\n p: Print performance statistics"
"\n"
"\n x: Exit"
"\n"
"\n Enter Choice: "
};
static vx_status app_run_graph_interactive(AppObj *obj)
{
vx_status status = VX_SUCCESS;
uint32_t done = 0;
char ch;
FILE *fp;
app_perf_point_t *perf_arr[1];
status = app_run_task_create(obj);
if(status != VX_SUCCESS)
{
printf("app_tidl: ERROR: Unable to create task\n");
}
else
{
appPerfStatsResetAll();
while((!done) && (status == VX_SUCCESS))
{
printf(menu);
ch = getchar();
printf("\n");
switch(ch)
{
case 'p':
appPerfStatsPrintAll();
if(status == VX_SUCCESS)
{
status = tivx_utils_graph_perf_print(obj->graph);
}
if(status == VX_SUCCESS)
{
status = tivx_utils_graph_perf_print(obj->disp_graph);
}
appPerfPointPrint(&obj->fileio_perf);
appPerfPointPrint(&obj->draw_perf);
appPerfPointPrint(&obj->total_perf);
printf("\n");
appPerfPointPrintFPS(&obj->total_perf);
printf("\n");
break;
case 'e':
perf_arr[0] = &obj->total_perf;
fp = appPerfStatsExportOpenFile(".", "dl_demos_app_tidl");
if((NULL != fp) && (status == VX_SUCCESS))
{
appPerfStatsExportAll(fp, perf_arr, 1);
status = tivx_utils_graph_perf_export(fp, obj->graph);
appPerfStatsExportCloseFile(fp);
appPerfStatsResetAll();
}
else
{
printf("fp is null\n");
}
break;
case 'x':
obj->stop_task = 1;
done = 1;
break;
}
}
app_run_task_delete(obj);
}
return status;
}
static vx_status app_verify_graph(AppObj *obj)
{
vx_status status = VX_SUCCESS;
APP_PRINTF("app_tidl: Verifying graph ... \n");
/* Verify the TIDL Graph */
status = vxVerifyGraph(obj->graph);
if(status != VX_SUCCESS)
{
printf("app_tidl: ERROR: Verifying graph ... Failed !!!\n");
return status;
}
APP_PRINTF("app_tidl: Verifying graph ... Done.\n");
if ((vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1)) && (obj->display_option == 1))
{
APP_PRINTF("app_tidl: Verifying display graph ... \n");
/* Verify the TIDL Graph */
status = vxVerifyGraph(obj->disp_graph);
if(status != VX_SUCCESS)
{
printf("app_tidl: ERROR: Verifying display graph ... Failed !!!\n");
return status;
}
APP_PRINTF("app_tidl: Verifying display graph ... Done.\n");
}
/* wait a while for prints to flush */
tivxTaskWaitMsecs(100);
#if 1
if(status == VX_SUCCESS)
{
status = tivxExportGraphToDot(obj->graph,".", "vx_app_tidl");
APP_ASSERT(status==VX_SUCCESS);
}
#endif
#ifdef APP_TIVX_LOG_RT_ENABLE
tivxLogRtTraceEnable(obj->graph);
#endif
return status;
}
static void saveOutput(AppObj *obj, vx_tensor *output_tensors, vx_tensor *input_tensors, char *fileName)
{
uint8_t color[14][3]={{0,0,0},{255,0,0},{0,255,0},{0,0,255},{255,255,0},
{255,0,255},{0,255,255},{127,0,0},{0,127,0},{0,0,127},
{127,127,0},{127,0,127},{0,127,127},{255,255,255},};
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS], input_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc;
ioBufDesc = &obj->ioBufDesc;
uint8_t *color_mask = obj->temp;
uint8_t *gray_mask = obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3;
uint8_t *image_mask= obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3 * 2;
int32_t width, height, outPadL, outPadT, inPadL, inPadT;//, padB, padR
char output_file[APP_MAX_FILE_PATH];
char file_name_temp[APP_MAX_FILE_PATH];
char file_name_gmask[APP_MAX_FILE_PATH];
char file_name_cmask[APP_MAX_FILE_PATH];
char file_name_imask[APP_MAX_FILE_PATH];
strcpy(file_name_temp, fileName);
file_name_temp[strlen(fileName)-4] = '\0';
//yolov5参数
vx_int32 head_idx, h_idx, w_idx, anchor_idx;
vx_int32 yolov5_class_stride = 3 + 5 + 8 + 6; //cls confxywh points slot
vx_int32 bbox_num = 0;
{
uint64_t cur_time;
cur_time = tivxPlatformGetTimeInUsecs();
/*测试端到端解码接口_依赖tda4结构体*/
// if (0) {
// getMult_yolov5e2eResult(&obj->ioBufDesc, obj->output_tensors, &obj->out_args, &obj->yolov5params, &obj->detections);
// cur_time = tivxPlatformGetTimeInUsecs() - cur_time;
// printf("端到端车位解码耗时: %d us \n", (int)cur_time);
// int num = obj->detections.num_bbox;
// ttevxYoloV5BBox num0 = obj->detections.buffer[0];
// ttevxYoloV5BBox num1 = obj->detections.buffer[1];
// printf("检测到%d车位 \n",(int)num);
// //将车位检测结果保存成二进制文件
// vx_char new_name[256];
// snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%s%s%s%s", obj->output_file_path,"/psd/",file_name_temp,"_", "psd", ".bin");
// FILE *fp = fopen(new_name, "wb");
// if(NULL == fp)
// {
// printf("Unable to open file %s for writing!\n", new_name);
// }
// for(int i = 0; i< num ; i++)
// {
// ttevxYoloV5BBox num0 = obj->detections.buffer[i];
// fwrite(&num0, 4, 20, fp);
// }
// fclose(fp);
// }
/*测试端到端解码接口_不依赖tda4结构体*/
if (1) {
initMult_yolov5params(&obj->yolov5params);
/* 兜底：确保解码使用的输入尺寸有效 */
obj->yolov5params.inWidth[0] = obj->ioBufDesc.inWidth[0];
obj->yolov5params.inHeight[0] = obj->ioBufDesc.inHeight[0];
// 添加调试：打印所有输出张量的维度信息
// printf("=== 开始处理YOLOv5检测头 ===\n");
// for(int zh_idx = 0; zh_idx < 5; zh_idx++) // 查看所有5个输出张量
// {
// vx_size tensor_dims;
// vx_size tensor_sizes[APP_MAX_TENSOR_DIMS];
// vxQueryTensor(output_tensors[zh_idx], VX_TENSOR_NUMBER_OF_DIMS, &tensor_dims, sizeof(vx_size));
// vxQueryTensor(output_tensors[zh_idx], (vx_enum)VX_TENSOR_DIMS, tensor_sizes, APP_MAX_TENSOR_DIMS * sizeof(vx_size));
// printf("输出张量[%d] - ioBufDesc维度: outWidth=%d, outHeight=%d, outNumChannels=%d\n",
// zh_idx, obj->ioBufDesc.outWidth[zh_idx], obj->ioBufDesc.outHeight[zh_idx], obj->ioBufDesc.outNumChannels[zh_idx]);
// printf("输出张量[%d] - tensor维度: num_dims=%zu, sizes=[%zu, %zu, %zu, %zu]\n",
// zh_idx, tensor_dims, tensor_sizes[0], tensor_sizes[1], tensor_sizes[2], tensor_sizes[3]);
// }
// printf("=== 结束维度信息输出 ===\n\n");
// return;
// for(int zh_idx = 0; zh_idx < 3; zh_idx++)//lyp
vx_map_id map_id_output_arr[APP_MAX_TENSOR_DIMS] = {0};
int output_mapped[APP_MAX_TENSOR_DIMS] = {0};
for(int zh_idx = 1; zh_idx < 4; zh_idx++)
{
int dst_idx = zh_idx - 1; /* 将检测头1,2,3映射到0,1,2 */
obj->yolov5params.scale[dst_idx] = obj->out_args.scale[zh_idx];
obj->yolov5params.outWidth[dst_idx] = obj->ioBufDesc.outWidth[zh_idx];
obj->yolov5params.outHeight[dst_idx] = obj->ioBufDesc.outHeight[zh_idx];
obj->yolov5params.inWidth[dst_idx] = obj->ioBufDesc.inWidth[zh_idx];
obj->yolov5params.inHeight[dst_idx] = obj->ioBufDesc.inHeight[zh_idx];
obj->yolov5params.outNumChannels[dst_idx] = obj->ioBufDesc.outNumChannels[zh_idx];
obj->yolov5params.outChannelPitch[dst_idx] = obj->ioBufDesc.outChannelPitch[zh_idx];
obj->yolov5params.outPadL[dst_idx] = obj->ioBufDesc.outPadL[zh_idx];
obj->yolov5params.outPadT[dst_idx] = obj->ioBufDesc.outPadT[zh_idx];
obj->yolov5params.outPadR[dst_idx] = obj->ioBufDesc.outPadR[zh_idx];
obj->yolov5params.outPadB[dst_idx] = obj->ioBufDesc.outPadB[zh_idx];
obj->yolov5params.outDataTypeVX[dst_idx] = (int)getTensorDataType(obj->ioBufDesc.outElementType[zh_idx]);
printf("det_head[%d->%d] type(VX)=%d scale=%f outW=%d outH=%d outC=%d pitch=%d\n",
zh_idx, dst_idx,
obj->yolov5params.outDataTypeVX[dst_idx],
obj->yolov5params.scale[dst_idx],
obj->yolov5params.outWidth[dst_idx],
obj->yolov5params.outHeight[dst_idx],
obj->yolov5params.outNumChannels[dst_idx],
obj->yolov5params.outChannelPitch[dst_idx]);
printf("det_head[%d->%d] outElementType(TIDL)=%d vxType=%s\n",
zh_idx, dst_idx,
obj->ioBufDesc.outElementType[zh_idx],
vxTypeToStr((vx_enum)obj->yolov5params.outDataTypeVX[dst_idx]));
void *output_buffer;
vx_map_id map_id_output;
vx_size output_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
output_sizes[0] = ioBufDesc->outWidth[zh_idx] + ioBufDesc->outPadL[zh_idx] + ioBufDesc->outPadR[zh_idx];
output_sizes[1] = ioBufDesc->outHeight[zh_idx] + ioBufDesc->outPadT[zh_idx] + ioBufDesc->outPadB[zh_idx];
output_sizes[2] = ioBufDesc->outNumChannels[zh_idx];
vx_size out_elem_size = vxTypeSize((vx_enum)obj->yolov5params.outDataTypeVX[dst_idx]);
output_strides[0] = out_elem_size;
output_strides[1] = output_sizes[0] * output_strides[0];
output_strides[2] = output_sizes[1] * output_strides[1];
vx_size tensor_dims;
vxQueryTensor(output_tensors[zh_idx], VX_TENSOR_NUMBER_OF_DIMS, &tensor_dims, sizeof(vx_size));
/* 获取tensor不同dim占用的内存大小 */
tivxMapTensorPatch(output_tensors[zh_idx], tensor_dims, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
status = vxGetStatus((vx_reference)output_tensors[zh_idx]);
if(VX_SUCCESS == status)
{
obj->yolov5params.output_tensors[dst_idx] = output_buffer;
map_id_output_arr[dst_idx] = map_id_output;
output_mapped[dst_idx] = 1;
}
}
printf("端到端车位解码耗时 \n");
getMult_yolov5e2eResult(&obj->yolov5params, &obj->detections);
for(int dst_idx = 0; dst_idx < 3; dst_idx++)
{
if(output_mapped[dst_idx])
{
tivxUnmapTensorPatch(output_tensors[dst_idx + 1], map_id_output_arr[dst_idx]);
}
}
cur_time = tivxPlatformGetTimeInUsecs() - cur_time;
printf("端到端车位解码耗时: %d us \n", (int)cur_time);
int num = obj->detections.num_bbox;
ttevxYoloV5BBox num0 = obj->detections.buffer[0];
ttevxYoloV5BBox num1 = obj->detections.buffer[1];
printf("检测到%d车位 \n",(int)num);
//将车位检测结果保存成二进制文件
vx_char new_name[256];
snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%s%s%s%s", obj->output_file_path,"/psd/",file_name_temp,"_", "psd", ".bin");
FILE *fp = fopen(new_name, "wb");
if(NULL == fp)
{
printf("Unable to open file %s for writing!\n", new_name);
}
for(int i = 0; i< num ; i++)
{
ttevxYoloV5BBox num0 = obj->detections.buffer[i];
fwrite(&num0, 4, 20, fp);
}
fclose(fp);
}
}
//处理分割分支
// for(int32_t id = 3; id < 5; id++) //for mulTask_yolov5_seg lyp
// {
int32_t ids_to_process[] = {0, 4};
int32_t num_ids = 2;
for(int32_t idx = 0; idx < num_ids; idx++)
{
int32_t id = ids_to_process[idx];
snprintf(file_name_cmask, APP_MAX_FILE_PATH-1, "%s%s%s%d%s", "/img/", file_name_temp, "_cmask", id, ".bmp");
snprintf(file_name_imask, APP_MAX_FILE_PATH-1, "%s%s%s%d%s", "/img/",file_name_temp, "_imask", id, ".bmp");
snprintf(file_name_gmask, APP_MAX_FILE_PATH-1, "%s%s%s%d%s", "/img/", file_name_temp, "_gmask", id, ".bmp");
output_sizes[0] = ioBufDesc->outWidth[id] + obj->ioBufDesc.outPadL[id] + obj->ioBufDesc.outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + obj->ioBufDesc.outPadT[id] + obj->ioBufDesc.outPadB[id];
output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
input_sizes[2] = ioBufDesc->inNumChannels[0];
width = ioBufDesc->outWidth[id];
height = ioBufDesc->outHeight[id];
outPadL = obj->ioBufDesc.outPadL[id];
outPadT = obj->ioBufDesc.outPadT[id];
inPadL = obj->ioBufDesc.inPadL[0];
inPadT = obj->ioBufDesc.inPadT[0];
status = vxGetStatus((vx_reference)output_tensors[id]);
if(VX_SUCCESS == status)
{
void *output_buffer, *input_buffer;
vx_map_id map_id_output, map_id_input;
vx_size output_strides[APP_MAX_TENSOR_DIMS], input_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS], input_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
output_strides[0] = 1;
output_strides[1] = output_sizes[0];
output_strides[2] = output_sizes[1] * output_strides[1];
input_strides[0] = 1;
input_strides[1] = input_sizes[0];
input_strides[2] = input_sizes[1] * input_strides[1];
tivxMapTensorPatch(output_tensors[id], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
{
uint8_t *pOut;
uint8_t *pIn;
pOut = (uint8_t *)output_buffer;
pIn = (uint8_t *)input_buffer;
// printf("模型没加argmax \n");
for(int32_t i = 0; i < height; i++){
for(int32_t j = 0; j < width; j++){
uint8_t seg_type = pOut[output_sizes[0]*(outPadT + i) + outPadL + j];
// if(seg_type>15)
// {
// printf("seg_type 大于15，出错");
// }
uint8_t img_value_R = pIn[0 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
uint8_t img_value_G = pIn[1 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
uint8_t img_value_B = pIn[2 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
int32_t color_chosen = seg_type < 14 ? seg_type : 13;
color_mask[width * 3 * i + j * 3 + 0] = color[color_chosen][0];
color_mask[width * 3 * i + j * 3 + 1] = color[color_chosen][1];
color_mask[width * 3 * i + j * 3 + 2] = color[color_chosen][2];
gray_mask[width * 3 * i + j * 3 + 0] = seg_type;
gray_mask[width * 3 * i + j * 3 + 1] = seg_type;
gray_mask[width * 3 * i + j * 3 + 2] = seg_type;
if(seg_type > 0){
image_mask[width * 3 * i + j * 3 + 0] = img_value_R/2 + color[color_chosen][0]/2;
image_mask[width * 3 * i + j * 3 + 1] = img_value_G/2 + color[color_chosen][1]/2;
image_mask[width * 3 * i + j * 3 + 2] = img_value_B/2 + color[color_chosen][2]/2;
}else{
image_mask[width * 3 * i + j * 3 + 0] = img_value_R;
image_mask[width * 3 * i + j * 3 + 1] = img_value_G;
image_mask[width * 3 * i + j * 3 + 2] = img_value_B;
}
}
}
// if(obj->save_type == '0' || obj->save_type == '1'){
// snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_cmask);
// printf("save to %s!!!\n", output_file);
// tivx_utils_bmp_write(output_file, color_mask, width, height, (width * 3), obj->df_image);
// }
if(obj->save_type == '0' || obj->save_type == '3'){
snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_gmask);
printf("save to %s!!!\n", output_file);
tivx_utils_bmp_write(output_file, gray_mask, width, height, (width * 3), obj->df_image);
}
if(obj->save_type == '0' || obj->save_type == '3'){
snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_imask);
printf("save to %s!!!\n", output_file);
tivx_utils_bmp_write(output_file, image_mask, width, height, (width * 3), obj->df_image);
}
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
tivxUnmapTensorPatch(input_tensors[0], map_id_input);
}
}
}
static void saveOutput_tensor(AppObj *obj, vx_tensor *output_tensors, vx_tensor *input_tensors, char *fileName)
{
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc;
ioBufDesc = &obj->ioBufDesc;
uint8_t *color_mask = obj->temp;
uint8_t *gray_mask = obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3;
uint8_t *image_mask= obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3 * 2;
int32_t width, height, outPadL, outPadT, inPadL, inPadT;//, padB, padR
char output_file[APP_MAX_FILE_PATH];
char file_name_temp[APP_MAX_FILE_PATH];
char file_name_tensor[APP_MAX_FILE_PATH];
// char file_name_cmask[APP_MAX_FILE_PATH];
// char file_name_imask[APP_MAX_FILE_PATH];
strcpy(file_name_temp, fileName);
file_name_temp[strlen(fileName)-4] = '\0';
//处理分割分支 lyp
for(int32_t id = 4; id < 5; id++) //for mulTask_yolov5_seg
{
// int32_t ids_to_process[] = {0, 4};
// int32_t num_ids = 2;
// for(int32_t idx = 0; idx < num_ids; idx++)
// {
// int32_t id = ids_to_process[idx];
// snprintf(file_name_tensor, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_tensor", id, ".bmp");
// snprintf(file_name_imask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_imask", id, ".bmp");
// snprintf(file_name_gmask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_gmask", id, ".bmp");
output_sizes[0] = ioBufDesc->outWidth[id] + obj->ioBufDesc.outPadL[id] + obj->ioBufDesc.outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + obj->ioBufDesc.outPadT[id] + obj->ioBufDesc.outPadB[id];
output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
// input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
// input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
// input_sizes[2] = ioBufDesc->inNumChannels[0];
width = ioBufDesc->outWidth[id];
height = ioBufDesc->outHeight[id];
outPadL = obj->ioBufDesc.outPadL[id];
outPadT = obj->ioBufDesc.outPadT[id];
// inPadL = obj->ioBufDesc.inPadL[0];
// inPadT = obj->ioBufDesc.inPadT[0];
status = vxGetStatus((vx_reference)output_tensors[id]);
if(VX_SUCCESS == status)
{
void *output_buffer;
vx_map_id map_id_output, map_id_input;
vx_size output_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
// input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
output_strides[0] = 1;
output_strides[1] = output_sizes[0];
output_strides[2] = output_sizes[1] * output_strides[1];
// input_strides[0] = 1;
// input_strides[1] = input_sizes[0];
// input_strides[2] = input_sizes[1] * input_strides[1];
vx_char new_name[256];
// snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%d%s", obj->output_file_path, "/_tensor", id, ".bin");
snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%s%s%d%s", obj->output_file_path,"/bin/",file_name_temp,"_", id, ".bin");
FILE *fp = fopen(new_name, "wb");
if(NULL == fp)
{
printf("Unable to open file %s for writing!\n", new_name);
break;
}
tivxMapTensorPatch(output_tensors[id], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
// tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
{
uint8_t *pOut, *pIn;
pOut = (uint8_t *)output_buffer;
// pIn = (uint8_t *)input_buffer;
// printf("模型没加argmax \n");
for(int32_t c = 0; c < ioBufDesc->outNumChannels[id]; c++){
for(int32_t i = 0; i < height; i++){
uint8_t *pOut_1 = pOut + c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL;
uint8_t zh = pOut[c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL];
fwrite(pOut_1, 1, ioBufDesc->outWidth[id], fp);
}
}
fclose(fp);
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
// tivxUnmapTensorPatch(input_tensors[0], map_id_input);
}
}
}
static void saveOutput_16bit(AppObj *obj, vx_tensor *output_tensors, vx_tensor *input_tensors, char *fileName)
{
uint8_t color[14][3]={{0,0,0},{255,0,0},{0,255,0},{0,0,255},{255,255,0},
{255,0,255},{0,255,255},{127,0,0},{0,127,0},{0,0,127},
{127,127,0},{127,0,127},{0,127,127},{127,127,127},};
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS], input_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc;
ioBufDesc = &obj->ioBufDesc;
uint8_t *color_mask = obj->temp;
uint8_t *gray_mask = obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3;
uint8_t *image_mask= obj->temp + MAX_IMG_WIDTH * MAX_IMG_HEIGHT * 3 * 2;
int32_t width, height, outPadL, outPadT, inPadL, inPadT;//, padB, padR
char output_file[APP_MAX_FILE_PATH];
char file_name_temp[APP_MAX_FILE_PATH];
char file_name_gmask[APP_MAX_FILE_PATH];
char file_name_cmask[APP_MAX_FILE_PATH];
char file_name_imask[APP_MAX_FILE_PATH];
strcpy(file_name_temp, fileName);
file_name_temp[strlen(fileName)-4] = '\0';
//yolov5参数
vx_int32 head_idx, h_idx, w_idx, anchor_idx;
vx_int32 yolov5_class_stride = 3 + 5 + 8 + 6; //cls confxywh points slot
ttevxYoloV5BBox *bbox_obj_arr = vxCreateUserDataObject(obj->context, "", sizeof(ttevxYoloV5BBox), NULL);
ttevxYoloV5PostProcParams *yolov5params = vxCreateUserDataObject(obj->context, "", sizeof(ttevxYoloV5PostProcParams), NULL);
vx_int32 bbox_num = 0;
TIDL_outArgs* out_args = obj->outArgs;
ttevxYoloV5Detections *detections = vxCreateUserDataObject(obj->context, "", sizeof(ttevxYoloV5Detections), NULL);
//init
{
for(int i = 0;i<5;i++)
{
out_args->scale[i] = obj->ioBufDesc.outTensorScale[i];
}
yolov5params->num_anchors = TTE_YOLOV5_NUM_ANCHOR;
yolov5params->num_classes = TTE_YOLOV5_NUM_CLASSES;
yolov5params->conf_thresh = 0.45;
yolov5params->nms_thresh = 0.25;
int i, j, k;
for (i = 0; i < TTE_YOLOV5_NUM_HEAD; ++i) {
for (j = 0; j < TTE_YOLOV5_NUM_ANCHOR; ++j) {
for (k = 0; k < 2; ++k) {
yolov5params->anchors[i][j][k] = ANCHORS[i][j][k];
}
}
}
}
// for(int32_t head_idx = 2; head_idx < obj->num_output_tensors; head_idx++) // for mulTask_yolov5_det
// {
// output_sizes[0] = ioBufDesc->outWidth[head_idx] + obj->ioBufDesc.outPadL[head_idx] + obj->ioBufDesc.outPadR[head_idx];
// output_sizes[1] = ioBufDesc->outHeight[head_idx] + obj->ioBufDesc.outPadT[head_idx] + obj->ioBufDesc.outPadB[head_idx];
// output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
// input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
// input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
// input_sizes[2] = ioBufDesc->inNumChannels[0];
// width = ioBufDesc->outWidth[head_idx];
// height = ioBufDesc->outHeight[head_idx];
// outPadL = obj->ioBufDesc.outPadL[head_idx];
// outPadT = obj->ioBufDesc.outPadT[head_idx];
// inPadL = obj->ioBufDesc.inPadL[0];
// inPadT = obj->ioBufDesc.inPadT[0];
// //获取内存块
// void *output_buffer, *input_buffer;
// vx_map_id map_id_output, map_id_input;
// vx_size output_strides[APP_MAX_TENSOR_DIMS], input_strides[APP_MAX_TENSOR_DIMS];
// vx_size output_start[APP_MAX_TENSOR_DIMS], input_start[APP_MAX_TENSOR_DIMS];
// output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
// input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
// output_strides[0] = 1;
// output_strides[1] = output_sizes[0];
// output_strides[2] = output_sizes[1] * output_strides[1];
// input_strides[0] = 1;
// input_strides[1] = input_sizes[0];
// input_strides[2] = input_sizes[1] * input_strides[1];
// tivxMapTensorPatch(output_tensors[head_idx], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
// tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
// //
// status = vxGetStatus((vx_reference)output_tensors[head_idx]);
// if(VX_SUCCESS == status)
// {
// /* 初始时将指针定位到第一个有效内存位置 */
// uint8_t *pOut, *pIn;
// pOut = (uint8_t *)output_buffer;
// pIn = (uint8_t *)input_buffer;
// vx_uint8 *raw_data = (vx_uint8*)(pOut) + (output_sizes[0] * outPadT + outPadL);
// /* TIDL中内存实际是CHW排布，但是以WHC记录，所以查找内存时首先定位正确的通道，再进行WH的偏移 */
// for (h_idx = 0; h_idx < ioBufDesc->outHeight[head_idx]; ++h_idx) {
// for (w_idx = 0; w_idx < ioBufDesc->outWidth[head_idx]; ++w_idx) {
// for (anchor_idx = 0; anchor_idx < yolov5params->num_anchors; ++anchor_idx) {
// /* 在内存片中对应[h,w]的线性位置 */
// int32_t obj_hw_idx = h_idx * output_sizes[0] + w_idx;
// /* 找到正确conf内存通道位置，并做hw偏移 */
// int32_t obj_conf_idx = (anchor_idx * yolov5_class_stride + 4) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// /* 获得conf并将conf缩放回输入尺度 */
// float conf = (float)raw_data[obj_conf_idx] / out_args->scale[head_idx];
// if (conf >= yolov5params->conf_thresh) {
// /* 计算xywh和cls内存位置并转化为输出值 */
// int32_t obj_x_idx = (anchor_idx * yolov5_class_stride + 0) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// int32_t obj_y_idx = (anchor_idx * yolov5_class_stride + 1) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// int32_t obj_w_idx = (anchor_idx * yolov5_class_stride + 2) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// int32_t obj_h_idx = (anchor_idx * yolov5_class_stride + 3) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// float x = (float)raw_data[obj_x_idx] / out_args->scale[head_idx];
// x = (x * 2.0 - 0.5 + w_idx) * (float)(ioBufDesc->inWidth[0]) / (float)(ioBufDesc->outWidth[head_idx]);
// float y = (float)raw_data[obj_y_idx] / out_args->scale[head_idx];
// y = (y * 2.0 - 0.5 + h_idx) * (float)(ioBufDesc->inHeight[0]) / (float)(ioBufDesc->outHeight[head_idx]);
// float w = (float)raw_data[obj_w_idx] / out_args->scale[head_idx];
// w = (w * 2.0) * (w * 2.0) * yolov5params->anchors[head_idx][anchor_idx][0];
// float h = (float)raw_data[obj_h_idx] / out_args->scale[head_idx];
// h = (h * 2.0) * (h * 2.0) * yolov5params->anchors[head_idx][anchor_idx][1];
// int32_t cls = 0;
// float cls_prob = 0.0;
// /* 遍历查找类别内存，找到最大的置信度进行赋值 */
// {
// int32_t obj_cls_idx;
// int32_t max_prob_idx;
// float p;
// for (max_prob_idx = 0; max_prob_idx < yolov5params->num_classes; ++max_prob_idx) {
// obj_cls_idx = (anchor_idx * yolov5_class_stride + 5 + 8 + 5 + max_prob_idx) * ioBufDesc->outChannelPitch[head_idx] + obj_hw_idx;
// p = (float)raw_data[obj_cls_idx] / out_args->scale[head_idx];
// if (p > cls_prob) {
// cls_prob = p;
// cls = max_prob_idx;
// }
// }
// }
// /* 输出概率 = 存在目标概率 * 类别置信概率 */
// bbox_obj_arr[bbox_num].confidence = conf * cls_prob;
// if (bbox_obj_arr[bbox_num].confidence >= yolov5params->conf_thresh) {
// bbox_obj_arr[bbox_num].x1 = (int)(x - w / 2 + 0.5);
// bbox_obj_arr[bbox_num].y1 = (int)(y - h / 2 + 0.5);
// bbox_obj_arr[bbox_num].x2 = (int)(x + w / 2 + 0.5);
// bbox_obj_arr[bbox_num].y2 = (int)(y + h / 2 + 0.5);
// bbox_obj_arr[bbox_num].cls = cls;
// bbox_num++;
// }
// }
// }
// }
// }
// }
// }
// //NMS
// bbox_num = NMS(bbox_obj_arr, bbox_num, yolov5params->nms_thresh);
// detections->num_bbox = bbox_num;
//处理分割分支
for(int32_t id = 4; id < 5; id++) //for mulTask_yolov5_seg
{
snprintf(file_name_cmask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_cmask", id, ".bmp");
snprintf(file_name_imask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_imask", id, ".bmp");
snprintf(file_name_gmask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_gmask", id, ".bmp");
output_sizes[0] = ioBufDesc->outWidth[id] + obj->ioBufDesc.outPadL[id] + obj->ioBufDesc.outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + obj->ioBufDesc.outPadT[id] + obj->ioBufDesc.outPadB[id];
output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
input_sizes[2] = ioBufDesc->inNumChannels[0];
width = ioBufDesc->outWidth[id];
height = ioBufDesc->outHeight[id];
outPadL = obj->ioBufDesc.outPadL[id];
outPadT = obj->ioBufDesc.outPadT[id];
inPadL = obj->ioBufDesc.inPadL[0];
inPadT = obj->ioBufDesc.inPadT[0];
status = vxGetStatus((vx_reference)output_tensors[id]);
if(VX_SUCCESS == status)
{
void *output_buffer, *input_buffer;
vx_map_id map_id_output, map_id_input;
vx_size output_strides[APP_MAX_TENSOR_DIMS], input_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS], input_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
output_strides[0] = 1;
output_strides[1] = output_sizes[0];
output_strides[2] = output_sizes[1] * output_strides[1];
input_strides[0] = 1;
input_strides[1] = input_sizes[0];
input_strides[2] = input_sizes[1] * input_strides[1];
tivxMapTensorPatch(output_tensors[id], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
{
uint8_t *pOut, *pIn;
pOut = (uint16_t *)output_buffer;
pIn = (uint8_t *)input_buffer;
// printf("模型没加argmax \n");
for(int32_t i = 0; i < height; i++){
for(int32_t j = 0; j < width; j++){
uint8_t seg_type = (uint8_t)((uint16_t)pOut[output_sizes[0]*(outPadT + i) + outPadL + j]);
// if(seg_type>15)
// {
// printf("seg_type 大于15，出错");
// }
uint8_t img_value_R = pIn[0 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
uint8_t img_value_G = pIn[1 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
uint8_t img_value_B = pIn[2 * input_strides[2] + input_strides[1]*(inPadT + i) + inPadL + j];
int32_t color_chosen = seg_type < 14 ? seg_type : 13;
color_mask[width * 3 * i + j * 3 + 0] = color[color_chosen][0];
color_mask[width * 3 * i + j * 3 + 1] = color[color_chosen][1];
color_mask[width * 3 * i + j * 3 + 2] = color[color_chosen][2];
gray_mask[width * 3 * i + j * 3 + 0] = seg_type;
gray_mask[width * 3 * i + j * 3 + 1] = seg_type;
gray_mask[width * 3 * i + j * 3 + 2] = seg_type;
if(seg_type > 0){
image_mask[width * 3 * i + j * 3 + 0] = img_value_R/2 + color[color_chosen][0]/2;
image_mask[width * 3 * i + j * 3 + 1] = img_value_G/2 + color[color_chosen][1]/2;
image_mask[width * 3 * i + j * 3 + 2] = img_value_B/2 + color[color_chosen][2]/2;
}else{
image_mask[width * 3 * i + j * 3 + 0] = img_value_R;
image_mask[width * 3 * i + j * 3 + 1] = img_value_G;
image_mask[width * 3 * i + j * 3 + 2] = img_value_B;
}
}
}
// if(obj->save_type == '0' || obj->save_type == '1'){
// snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_cmask);
// printf("save to %s!!!\n", output_file);
// tivx_utils_bmp_write(output_file, color_mask, width, height, (width * 3), obj->df_image);
// }
// if(obj->save_type == '0' || obj->save_type == '2'){
// snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_gmask);
// printf("save to %s!!!\n", output_file);
// tivx_utils_bmp_write(output_file, gray_mask, width, height, (width * 3), obj->df_image);
// }
// if(obj->save_type == '0' || obj->save_type == '3'){
// snprintf(output_file, APP_MAX_FILE_PATH-1, "%s/%s", obj->output_file_path, file_name_imask);
// printf("save to %s!!!\n", output_file);
// tivx_utils_bmp_write(output_file, image_mask, width, height, (width * 3), obj->df_image);
// }
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
tivxUnmapTensorPatch(input_tensors[0], map_id_input);
}
}
}
static void saveOutput_tensor_16bit(AppObj *obj, vx_tensor *output_tensors, vx_tensor *input_tensors, char *fileName)
{
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc;
ioBufDesc = &obj->ioBufDesc;
int32_t width, height, outPadL, outPadT, inPadL, inPadT;//, padB, padR
char output_file[APP_MAX_FILE_PATH];
char file_name_temp[APP_MAX_FILE_PATH];
char file_name_tensor[APP_MAX_FILE_PATH];
// char file_name_cmask[APP_MAX_FILE_PATH];
// char file_name_imask[APP_MAX_FILE_PATH];
strcpy(file_name_temp, fileName);
file_name_temp[strlen(fileName)-4] = '\0';
//处理分割分支
for(int32_t id = 4; id < 5; id++) //for mulTask_yolov5_seg
{
// snprintf(file_name_tensor, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_tensor", id, ".bmp");
// snprintf(file_name_imask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_imask", id, ".bmp");
// snprintf(file_name_gmask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_gmask", id, ".bmp");
output_sizes[0] = ioBufDesc->outWidth[id] + obj->ioBufDesc.outPadL[id] + obj->ioBufDesc.outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + obj->ioBufDesc.outPadT[id] + obj->ioBufDesc.outPadB[id];
output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
// input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
// input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
// input_sizes[2] = ioBufDesc->inNumChannels[0];
width = ioBufDesc->outWidth[id];
height = ioBufDesc->outHeight[id];
outPadL = obj->ioBufDesc.outPadL[id];
outPadT = obj->ioBufDesc.outPadT[id];
// inPadL = obj->ioBufDesc.inPadL[0];
// inPadT = obj->ioBufDesc.inPadT[0];
status = vxGetStatus((vx_reference)output_tensors[id]);
if(VX_SUCCESS == status)
{
void *output_buffer;
vx_map_id map_id_output, map_id_input;
vx_size output_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
// input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
output_strides[0] = 1;
output_strides[1] = output_sizes[0];
output_strides[2] = output_sizes[1] * output_strides[1];
// input_strides[0] = 1;
// input_strides[1] = input_sizes[0];
// input_strides[2] = input_sizes[1] * input_strides[1];
vx_char new_name[256];
// snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%d%s", obj->output_file_path, "/_tensor", id, ".bin");
snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%s%s%d%s", obj->output_file_path,"/bin/",file_name_temp,"_", id, ".bin");
FILE *fp = fopen(new_name, "wb");
if(NULL == fp)
{
printf("Unable to open file %s for writing!\n", new_name);
break;
}
tivxMapTensorPatch(output_tensors[id], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
// tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
{
uint16_t *pOut;
pOut = (uint16_t *)output_buffer;
// pIn = (uint8_t *)input_buffer;
// printf("模型没加argmax \n");
for(int32_t c = 0; c < ioBufDesc->outNumChannels[id]; c++){
for(int32_t i = 0; i < height; i++){
uint16_t *pOut_1 = pOut + c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL;
uint16_t zh = pOut[c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL];
fwrite(pOut_1, 2, ioBufDesc->outWidth[id], fp);
}
}
fclose(fp);
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
// tivxUnmapTensorPatch(input_tensors[0], map_id_input);
}
}
}
static void saveOutput_tensor_32bit(AppObj *obj, vx_tensor *output_tensors, vx_tensor *input_tensors, char *fileName)
{
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc;
ioBufDesc = &obj->ioBufDesc;
int32_t width, height, outPadL, outPadT, inPadL, inPadT;//, padB, padR
char output_file[APP_MAX_FILE_PATH];
char file_name_temp[APP_MAX_FILE_PATH];
char file_name_tensor[APP_MAX_FILE_PATH];
// char file_name_cmask[APP_MAX_FILE_PATH];
// char file_name_imask[APP_MAX_FILE_PATH];
strcpy(file_name_temp, fileName);
file_name_temp[strlen(fileName)-4] = '\0';
//处理分割分支
for(int32_t id = 4; id < 5; id++) //for mulTask_yolov5_seg
{
// snprintf(file_name_tensor, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_tensor", id, ".bmp");
// snprintf(file_name_imask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_imask", id, ".bmp");
// snprintf(file_name_gmask, APP_MAX_FILE_PATH-1, "%s%s%d%s", file_name_temp, "_gmask", id, ".bmp");
output_sizes[0] = ioBufDesc->outWidth[id] + obj->ioBufDesc.outPadL[id] + obj->ioBufDesc.outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + obj->ioBufDesc.outPadT[id] + obj->ioBufDesc.outPadB[id];
output_sizes[2] = 1;//ioBufDesc->outNumChannels[id];
// input_sizes[0] = ioBufDesc->inWidth[0] + obj->ioBufDesc.inPadL[0] + obj->ioBufDesc.inPadR[0];
// input_sizes[1] = ioBufDesc->inHeight[0] + obj->ioBufDesc.inPadT[0] + obj->ioBufDesc.inPadB[0];
// input_sizes[2] = ioBufDesc->inNumChannels[0];
width = ioBufDesc->outWidth[id];
height = ioBufDesc->outHeight[id];
outPadL = obj->ioBufDesc.outPadL[id];
outPadT = obj->ioBufDesc.outPadT[id];
// inPadL = obj->ioBufDesc.inPadL[0];
// inPadT = obj->ioBufDesc.inPadT[0];
status = vxGetStatus((vx_reference)output_tensors[id]);
if(VX_SUCCESS == status)
{
void *output_buffer;
vx_map_id map_id_output, map_id_input;
vx_size output_strides[APP_MAX_TENSOR_DIMS];
vx_size output_start[APP_MAX_TENSOR_DIMS];
output_start[0] = output_start[1] = output_start[2] = output_start[3] = 0;
// input_start[0] = input_start[1] = input_start[2] = input_start[3] = 0;
output_strides[0] = 1;
output_strides[1] = output_sizes[0];
output_strides[2] = output_sizes[1] * output_strides[1];
// input_strides[0] = 1;
// input_strides[1] = input_sizes[0];
// input_strides[2] = input_sizes[1] * input_strides[1];
vx_char new_name[256];
// snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%d%s", obj->output_file_path, "/_tensor", id, ".bin");
snprintf(new_name, APP_MAX_FILE_PATH-1, "%s%s%s%s%d%s", obj->output_file_path,"/bin/",file_name_temp,"_", id, ".bin");
FILE *fp = fopen(new_name, "wb");
if(NULL == fp)
{
printf("Unable to open file %s for writing!\n", new_name);
break;
}
tivxMapTensorPatch(output_tensors[id], 3, output_start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
// tivxMapTensorPatch(input_tensors[0], 3, input_start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
{
vx_float32 *pOut;
pOut = (vx_float32 *)output_buffer;
// pIn = (uint8_t *)input_buffer;
// printf("模型没加argmax \n");
for(int32_t c = 0; c < ioBufDesc->outNumChannels[id]; c++){
for(int32_t i = 0; i < height; i++){
vx_float32 *pOut_1 = pOut + c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL;
vx_float32 zh = pOut[c * output_sizes[0] * output_sizes[1] + output_sizes[0]*(outPadT + i) + outPadL];
fwrite(pOut_1, 4, ioBufDesc->outWidth[id], fp);
}
}
fclose(fp);
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
// tivxUnmapTensorPatch(input_tensors[0], map_id_input);
}
}
}
static vx_status app_run_graph_for_one_frame(AppObj *obj, char *curFileName, vx_uint32 counter)
{
vx_status status = VX_SUCCESS;
vx_char input_file_name[APP_MAX_FILE_PATH];
vx_char output_file_name[APP_MAX_FILE_PATH];
vx_uint32 tensor_actual_checksum = 0, display_actual_checksum = 0;
appPerfPointBegin(&obj->total_perf);
snprintf(input_file_name, APP_MAX_FILE_PATH-1, "%s/%s",
obj->input_file_path,
curFileName
);
APP_PRINTF("app_tidl: Reading input file %s ... ", input_file_name);
appPerfPointBegin(&obj->fileio_perf);
/* Read input from file and poplulate the input tensors */
if(status == VX_SUCCESS)
{
status = readInput(obj, obj->context, obj->config, obj->input_tensors, &input_file_name[0]);
}
appPerfPointEnd(&obj->fileio_perf);
APP_PRINTF("Done!\n");
#ifdef APP_WRITE_PRE_PROC_OUTPUT
vx_char pre_proc_file_name[APP_MAX_FILE_PATH];
snprintf(pre_proc_file_name, APP_MAX_FILE_PATH-1, "%s/%s",
obj->output_file_path,
"pre_proc_out"
);
writePreProcOutput(pre_proc_file_name, obj->input_tensors[0]);
#endif
uint64_t cur_time;
APP_PRINTF("app_tidl: Running Graph ... ");
/* Execute the network */
if(status == VX_SUCCESS)
{
cur_time = tivxPlatformGetTimeInUsecs();
status = vxProcessGraph(obj->graph);
cur_time = tivxPlatformGetTimeInUsecs() - cur_time;
}
APP_PRINTF("Done!\n");
APP_PRINTF("20260122==============!\n");
printf("time of rmfsd: %d us \n", (int)cur_time);
if(status == VX_SUCCESS)
{
saveOutput_tensor(obj, obj->output_tensors, obj->input_tensors,curFileName);
// saveOutput_tensor_16bit(obj, obj->output_tensors, obj->input_tensors,curFileName);
}
if(status == VX_SUCCESS)
{
saveOutput(obj, obj->output_tensors, obj->input_tensors,curFileName);
appPerfPointBegin(&obj->draw_perf);
appPerfPointEnd(&obj->draw_perf);
}
tivx_utils_bmp_read_release(&obj->imgParams);
// if(status == VX_SUCCESS)
// {
// snprintf(output_file_name, APP_MAX_FILE_PATH-1, "%s/%s",
// obj->output_file_path,
// curFileName
// );
//
// saveOutput(obj, obj->output_tensors, obj->input_tensors,curFileName);
// appPerfPointBegin(&obj->draw_perf);
// /* Display the output */
// displayOutput(obj, obj->config, obj->output_tensors, &output_file_name[0]);
// appPerfPointEnd(&obj->draw_perf);
// if ((vx_true_e == tivxIsTargetEnabled(TIVX_TARGET_DISPLAY1)) && (obj->display_option == 1)) {
// /* At this point, the output is ready to copy the updated buffer */
// if(status == VX_SUCCESS)
// {
// status = vxCopyImagePatch(obj->disp_image,
// &obj->disp_rect,
// 0,
// &obj->image_addr,
// (void *)obj->pDisplayBuf888,
// VX_WRITE_ONLY,
// VX_MEMORY_TYPE_HOST
// );
// }
// APP_PRINTF("app_tidl: Running display graph ... \n");
// /* Execute the display graph */
// if(status == VX_SUCCESS)
// {
// status = vxProcessGraph(obj->disp_graph);
// }
// APP_PRINTF("app_tidl: Running display graph ... Done.\n");
// }
// /* Check that you are within the first n frames, where n is the number
// of samples in the checksums_expected */
// if ((obj->test_mode == 1) &&
// (counter < ((sizeof(checksums_expected[0])/sizeof(checksums_expected[0][0]))-1)))
// {
// /* numOutputbuf is 1 here, but for the sake of generalizing
// in the future, the loop will remain */
// if(vx_false_e == app_test_check_tensor(obj->output_tensors[0],
// getTensorDataType(obj->ioBufDesc.outElementType[0]),
// obj->ioBufDesc.outWidth[0], obj->ioBufDesc.outPadL[0],
// obj->ioBufDesc.outPadR[0], obj->ioBufDesc.outHeight[0],
// obj->ioBufDesc.outPadT[0], obj->ioBufDesc.outPadB[0],
// obj->ioBufDesc.outNumChannels[0], checksums_expected[obj->test_case*2][counter],
// &tensor_actual_checksum))
// {
// test_result = vx_false_e;
// }
// /* in case test fails and needs to change */
// populate_gatherer(0 + (obj->test_case*2), counter, tensor_actual_checksum);
//#ifndef x86_64
// if(vx_false_e == app_test_check_image(obj->disp_image, checksums_expected[1+(2*obj->test_case)][counter],
// &display_actual_checksum))
// {
// test_result = vx_false_e;
// }
// populate_gatherer(1 + (obj->test_case*2), counter, display_actual_checksum);
//#endif
// }
// }
appPerfPointEnd(&obj->total_perf);
#ifdef APP_TIDL_TRACE_DUMP
tivx_utils_tidl_trace_write(obj->traceData, curFileName);
#endif
// printf("exit_begin\n");
// exit(0);
// printf("exit_end\n");
return status;
}
static vx_status app_run_graph(AppObj *obj)
{
vx_status status = VX_SUCCESS;
vx_char curFileName[APP_MAX_FILE_PATH];
uint64_t cur_time;
uint64_t max_frames = INT64_MAX;
vx_uint32 counter = 0;
FILE* test_case_file;
uint32_t cur_iteration;
printf("network file: %s\n", obj->tidl_network_file_path);
printf("config file: %s\n", obj->tidl_config_file_path);
if (obj->test_mode == 1)
{
/* This way if the application failed at any point up to here
the test mode would return a failure */
test_result = vx_true_e;
obj->num_iterations = 1;
max_frames = (sizeof(checksums_expected[0])/sizeof(checksums_expected[0][0])) + TEST_BUFFER;
}
for(cur_iteration=0; cur_iteration<obj->num_iterations; cur_iteration++)
{
printf("Iteration %d of %d ... \n", cur_iteration, obj->num_iterations);
test_case_file = fopen(obj->input_file_list,"r");
if(test_case_file==NULL)
{
break;
}
int zh = 0;
while (fgets(curFileName, sizeof(curFileName), test_case_file) && (counter < max_frames))
{
curFileName[strlen(curFileName) - 1] = 0;
zh ++;
cur_time = tivxPlatformGetTimeInUsecs();
APP_PRINTF("Classifying input %s ...\n", curFileName);
if(status == VX_SUCCESS)
{
status = app_run_graph_for_one_frame(obj, curFileName, counter++);
}
printf("完成图像张数: %d\n", zh);
APP_PRINTF("Classifying input %s ...Done!\n", curFileName);
cur_time = tivxPlatformGetTimeInUsecs() - cur_time;
/* convert to msecs */
cur_time = cur_time;
printf("cur_time:==%ld\n",cur_time);
if(cur_time < obj->delay_in_msecs)
{
tivxTaskWaitMsecs(obj->delay_in_msecs - cur_time);
}
/* user asked to stop processing */
if(obj->stop_task || (status != VX_SUCCESS))
{
break;
}
}
fclose(test_case_file);
if(obj->stop_task || (status != VX_SUCCESS))
{
break;
}
}
printf("exit_begin\n");
exit(0);
printf("exit_end\n");
return status;
}
static vx_user_data_object readConfig(AppObj *obj, vx_context context, char *config_file, uint32_t *num_input_tensors, uint32_t *num_output_tensors)
{
vx_status status = VX_SUCCESS;
tivxTIDLJ7Params *tidlParams = NULL;
sTIDL_IOBufDesc_t *ioBufDesc = NULL;
vx_user_data_object config = NULL;
vx_uint32 capacity;
vx_map_id map_id;
FILE *fp_config;
vx_size read_count;
APP_PRINTF("app_tidl: Reading config file %s ...\n", config_file);
fp_config = fopen(config_file, "rb");
if(fp_config == NULL)
{
printf("app_tidl: ERROR: Unable to open IO config file %s \n", config_file);
return NULL;
}
fseek(fp_config, 0, SEEK_END);
capacity = ftell(fp_config);
fseek(fp_config, 0, SEEK_SET);
if( capacity != sizeof(sTIDL_IOBufDesc_t) )
{
printf("app_tidl: ERROR: Config file size (%d bytes) does not match size of sTIDL_IOBufDesc_t (%d bytes)\n", capacity, (vx_uint32)sizeof(sTIDL_IOBufDesc_t));
fclose(fp_config);
return NULL;
}
/* Create a user struct type for handling config data*/
config = vxCreateUserDataObject(context, "tivxTIDLJ7Params", sizeof(tivxTIDLJ7Params), NULL );
status = vxGetStatus((vx_reference)config);
if (VX_SUCCESS == status)
{
status = vxMapUserDataObject(config, 0, sizeof(tivxTIDLJ7Params), &map_id,
(void **)&tidlParams, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (VX_SUCCESS == status)
{
if(tidlParams == NULL)
{
printf("app_tidl: ERROR: Map of config object failed\n");
fclose(fp_config);
return NULL;
}
tivx_tidl_j7_params_init(tidlParams);
ioBufDesc = (sTIDL_IOBufDesc_t *)&tidlParams->ioBufDesc;
read_count = fread(ioBufDesc, capacity, 1, fp_config);
if(read_count != 1)
{
printf("app_tidl: ERROR: Unable to read config file\n");
}
fclose(fp_config);
memcpy(&obj->ioBufDesc, ioBufDesc, capacity);
*num_input_tensors = ioBufDesc->numInputBuf;
*num_output_tensors = ioBufDesc->numOutputBuf;
if(status == VX_SUCCESS)
{
status = vxUnmapUserDataObject(config, map_id);
}
}
else
{
fclose(fp_config);
}
}
else
{
fclose(fp_config);
}
APP_PRINTF("app_tidl: Reading config file %s ... Done. %d bytes\n", config_file, (uint32_t)capacity);
APP_PRINTF("app_tidl: Tensors, input = %d, output = %d\n", *num_input_tensors, *num_output_tensors);
return config;
}
static vx_user_data_object readNetwork(vx_context context, char *network_file)
{
vx_status status;
vx_user_data_object network;
vx_map_id map_id;
vx_uint32 capacity;
void *network_buffer = NULL;
vx_size read_count;
FILE *fp_network;
APP_PRINTF("app_tidl: Reading network file %s ... \n", network_file);
fp_network = fopen(network_file, "rb");
if(fp_network == NULL)
{
printf("app_tidl: ERROR: Unable to open network file %s \n", network_file);
return NULL;
}
fseek(fp_network, 0, SEEK_END);
capacity = ftell(fp_network);
fseek(fp_network, 0, SEEK_SET);
network = vxCreateUserDataObject(context, "TIDL_network", capacity, NULL );
status = vxGetStatus((vx_reference)network);
if (VX_SUCCESS == status)
{
status = vxMapUserDataObject(network, 0, capacity, &map_id,
(void **)&network_buffer, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (VX_SUCCESS == status)
{
if(network_buffer)
{
read_count = fread(network_buffer, capacity, 1, fp_network);
if(read_count != 1)
{
printf("app_tidl: ERROR: Unable to read network file\n");
}
fclose(fp_network);
}
else
{
printf("app_tidl: ERROR: Unable to allocate memory for reading network file of %d bytes\n", capacity);
fclose(fp_network);
}
if(status == VX_SUCCESS)
{
status = vxUnmapUserDataObject(network, map_id);
}
}
else
{
fclose(fp_network);
}
}
else
{
fclose(fp_network);
}
APP_PRINTF("app_tidl: Reading network file %s ... Done. %d bytes\n", network_file, (uint32_t)capacity);
return network;
}
static vx_user_data_object setCreateParams(vx_context context)
{
vx_status status;
vx_user_data_object createParams;
vx_map_id map_id;
vx_uint32 capacity;
void *createParams_buffer = NULL;
capacity = sizeof(TIDL_CreateParams);
createParams = vxCreateUserDataObject(context, "TIDL_CreateParams", capacity, NULL );
status = vxGetStatus((vx_reference)createParams);
if (VX_SUCCESS == status)
{
status = vxMapUserDataObject(createParams, 0, capacity, &map_id,
(void **)&createParams_buffer, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (VX_SUCCESS == status)
{
if(createParams_buffer)
{
TIDL_CreateParams *prms = createParams_buffer;
TIDL_createParamsInit(prms);
prms->isInbufsPaded = 1;
prms->quantRangeExpansionFactor = 1.0;
prms->quantRangeUpdateFactor = 0.0;
#ifdef x86_64
prms->flowCtrl = 1;
#endif
#ifdef APP_TIDL_TRACE_DUMP
prms->traceLogLevel = 1;
prms->traceWriteLevel = 1;
#else
prms->traceLogLevel = 0;
prms->traceWriteLevel = 0;
#endif
}
else
{
printf("app_tidl: ERROR: Unable to allocate memory for create time params! %d bytes\n", capacity);
}
vxUnmapUserDataObject(createParams, map_id);
}
}
return createParams;
}
static vx_user_data_object setInArgs(vx_context context)
{
vx_status status;
vx_user_data_object inArgs;
vx_map_id map_id;
vx_uint32 capacity;
void *inArgs_buffer = NULL;
capacity = sizeof(TIDL_InArgs);
inArgs = vxCreateUserDataObject(context, "TIDL_InArgs", capacity, NULL );
status = vxGetStatus((vx_reference)inArgs);
if (VX_SUCCESS == status)
{
status = vxMapUserDataObject(inArgs, 0, capacity, &map_id,
(void **)&inArgs_buffer, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (VX_SUCCESS == status)
{
if(inArgs_buffer)
{
TIDL_InArgs *prms = inArgs_buffer;
prms->iVisionInArgs.size = sizeof(TIDL_InArgs);
prms->iVisionInArgs.subFrameInfo = 0;
}
else
{
printf("app_tidl: Unable to allocate memory for inArgs! %d bytes\n", capacity);
}
vxUnmapUserDataObject(inArgs, map_id);
}
}
return inArgs;
}
static vx_user_data_object setOutArgs(vx_context context)
{
vx_status status;
vx_user_data_object outArgs;
vx_map_id map_id;
vx_uint32 capacity;
void *outArgs_buffer = NULL;
capacity = sizeof(TIDL_outArgs);
outArgs = vxCreateUserDataObject(context, "TIDL_outArgs", capacity, NULL );
status = vxGetStatus((vx_reference)outArgs);
if (VX_SUCCESS == status)
{
status = vxMapUserDataObject(outArgs, 0, capacity, &map_id,
(void **)&outArgs_buffer, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST, 0);
if (VX_SUCCESS == status)
{
if(outArgs_buffer)
{
TIDL_outArgs *prms = outArgs_buffer;
prms->iVisionOutArgs.size = sizeof(TIDL_outArgs);
}
else
{
printf("app_tidl: Unable to allocate memory for outArgs! %d bytes\n", capacity);
}
vxUnmapUserDataObject(outArgs, map_id);
}
}
return outArgs;
}
static void createInputTensors(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *input_tensors)
{
vx_size input_sizes[APP_MAX_TENSOR_DIMS];
uint32_t id;
sTIDL_IOBufDesc_t *ioBufDesc = &obj->ioBufDesc;
for(id = 0; id < ioBufDesc->numInputBuf; id++) {
input_sizes[0] = ioBufDesc->inWidth[id] + ioBufDesc->inPadL[id] + ioBufDesc->inPadR[id];
input_sizes[1] = ioBufDesc->inHeight[id] + ioBufDesc->inPadT[id] + ioBufDesc->inPadB[id];
input_sizes[2] = ioBufDesc->inNumChannels[id];
vx_size data_type = getTensorDataType(ioBufDesc->inElementType[id]);
if(data_type != VX_TYPE_INVALID)
input_tensors[id] = vxCreateTensor(context, 3, input_sizes, data_type, 0);
}
return;
}
static void createOutputTensors(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *output_tensors)
{
vx_size output_sizes[APP_MAX_TENSOR_DIMS];
uint32_t id;
sTIDL_IOBufDesc_t *ioBufDesc = &obj->ioBufDesc;
for(id = 0; id < ioBufDesc->numOutputBuf; id++) {
output_sizes[0] = ioBufDesc->outWidth[id] + ioBufDesc->outPadL[id] + ioBufDesc->outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + ioBufDesc->outPadT[id] + ioBufDesc->outPadB[id];
output_sizes[2] = ioBufDesc->outNumChannels[id];
vx_size data_type = getTensorDataType(ioBufDesc->outElementType[id]);
if(data_type != VX_TYPE_INVALID)
{
output_tensors[id] = vxCreateTensor(context, 3, output_sizes, data_type, 0);
}
}
return;
}
static void resizeImage(uint8_t *inImg, uint8_t *outImg, uint32_t inWidth, uint32_t inHeight, uint32_t inStride, uint32_t outWidth, uint32_t outHeight, uint32_t outStride)
{
float xScale = (inWidth * 1.0f) / (outWidth * 1.0f);
float yScale = (inHeight * 1.0f) / (outHeight * 1.0f);
uint8_t border_constant_value = 128;
int32_t srcOffsetX = 0;
int32_t srcOffsetY = 0;
int32_t dstOffsetX = 0;
int32_t dstOffsetY = 0;
int32_t x, y, ch;
for( ch = 0; ch < 3; ch++) {
uint8_t *pIn = inImg + (ch * inStride * inHeight);
uint8_t *pOut = outImg + (ch * outStride * outHeight);
for( y = 0; y < outHeight; y++ ) {
for( x = 0; x < outWidth; x++ ) {
/* Apply scale factors to find input pixel for each output pixel */
float src_x_f = ((float)(x+dstOffsetX) + 0.5f)*xScale - 0.5f;
float src_y_f = ((float)(y+dstOffsetY) + 0.5f)*yScale - 0.5f;
float xf = floorf(src_x_f);
float yf = floorf(src_y_f);
float dx = src_x_f - xf;
float dy = src_y_f - yf;
float a[4];
int32_t src_x = (int32_t)(xf) - srcOffsetX;
int32_t src_y = (int32_t)(yf) - srcOffsetY;
uint8_t tl = 0;
uint8_t tr = 0;
uint8_t bl = 0;
uint8_t br = 0;
a[0] = (1.0f - dx) * (1.0f - dy);
a[1] = (1.0f - dx) * (dy);
a[2] = (dx)* (1.0f - dy);
a[3] = (dx)* (dy);
tl = (src_x < 0 || src_y < 0 || src_x > (inWidth-1) || src_y > (inHeight-1) ) ?
border_constant_value :
pIn[(src_y*inStride) + src_x];
tr = ((src_x+1) < 0 || src_y < 0 || (src_x+1) > (inWidth-1) || src_y > (inHeight-1) ) ?
border_constant_value :
pIn[(src_y*inStride) + src_x + 1];
bl = (src_x < 0 || (src_y+1) < 0 || src_x > (inWidth-1) || (src_y+1) > (inHeight-1) ) ?
border_constant_value :
pIn[((src_y+1)*inStride) + src_x];
br = ((src_x+1) < 0 || (src_y+1) < 0 || (src_x+1) > (inWidth-1) || (src_y+1) > (inHeight-1) ) ?
border_constant_value :
pIn[((src_y+1)*inStride) + src_x + 1];
pOut[(y*outStride) + x] = (uint8_t)(a[0]*tl + a[2]*tr + a[1]*bl + a[3]*br + 0.5f);
}
}
}
return;
}
static vx_status readInput(AppObj *obj, vx_context context, vx_user_data_object config, vx_tensor *input_tensors, char *input_file)
{
vx_status status = VX_SUCCESS;
tivx_utils_bmp_image_params_t *imgParams;
void *input_buffer = NULL;
int32_t capacity;
uint32_t id;
vx_map_id map_id_input;
vx_size start[APP_MAX_TENSOR_DIMS];
vx_size input_strides[APP_MAX_TENSOR_DIMS];
vx_size input_sizes[APP_MAX_TENSOR_DIMS];
sTIDL_IOBufDesc_t *ioBufDesc = &obj->ioBufDesc;
for(id = 0; id < ioBufDesc->numInputBuf; id++)
{
vx_size data_type = getTensorDataType(ioBufDesc->inElementType[id]);
input_sizes[0] = ioBufDesc->inWidth[id] + ioBufDesc->inPadL[id] + ioBufDesc->inPadR[id];
input_sizes[1] = ioBufDesc->inHeight[id] + ioBufDesc->inPadT[id] + ioBufDesc->inPadB[id];
input_sizes[2] = ioBufDesc->inNumChannels[id];
capacity = input_sizes[0] * input_sizes[1] * input_sizes[2];
start[0] = start[1] = start[2] = 0;
input_strides[0] = sizeof(vx_int8);
if((data_type == VX_TYPE_INT8) ||
(data_type == VX_TYPE_UINT8))
{
input_strides[0] = sizeof(vx_int8);
}
else if((data_type == VX_TYPE_INT16) ||
(data_type == VX_TYPE_UINT16))
{
input_strides[0] = sizeof(vx_int16);
}
input_strides[1] = input_sizes[0] * input_strides[0];
input_strides[2] = input_sizes[1] * input_strides[1];
APP_PRINTF(" input_sizes[0] = %d, dim = %d padL = %d padR = %d\n", (uint32_t)input_sizes[0], ioBufDesc->inWidth[id], ioBufDesc->inPadL[id], ioBufDesc->inPadR[id]);
APP_PRINTF(" input_sizes[1] = %d, dim = %d padT = %d padB = %d\n", (uint32_t)input_sizes[1], ioBufDesc->inHeight[id], ioBufDesc->inPadT[id], ioBufDesc->inPadB[id]);
APP_PRINTF(" input_sizes[2] = %d, dim = %d \n", (uint32_t)input_sizes[2], ioBufDesc->inNumChannels[id]);
if(status == VX_SUCCESS)
{
status = tivxMapTensorPatch(input_tensors[id], 3, start, input_sizes, &map_id_input, input_strides, &input_buffer, VX_WRITE_ONLY, VX_MEMORY_TYPE_HOST);
}
if (VX_SUCCESS == status)
{
vx_df_image df_image;
void *data_ptr = NULL;
vx_uint32 img_width;
vx_uint32 img_height;
vx_uint32 img_stride;
vx_int32 start_offset;
vx_uint8 *pData;
vx_uint8 *pInPlanes;
vx_uint8 *pOutPlanes;
vx_int32 i, j;
APP_PRINTF("app_tidl: Reading bmp file ...\n" );
imgParams = &obj->imgParams;
status = tivx_utils_bmp_file_read(
input_file,
vx_false_e,
imgParams);
if(status != VX_SUCCESS)
{
printf("app_tidl: Reading bmp file ... Failed !!!\n" );
}
else
{
img_width = imgParams->width;
img_height = imgParams->height;
img_stride = imgParams->stride_y;
df_image = imgParams->format;
data_ptr = imgParams->data;
APP_PRINTF("app_tidl: Reading bmp file ... Done.\n" );
APP_PRINTF("app_tidl: Image Pre processing for image of size %d x %d (pitch = %d bytes)...\n", img_width, img_height, img_stride);
APP_PRINTF("app_tidl: Deinterleaving data ...\n");
/* Save image params which are used in displayOutput function */
obj->img_width = img_width;
obj->img_height = img_height;
obj->img_stride = img_stride;
obj->data_ptr = data_ptr;
obj->df_image = df_image;
pInPlanes = (vx_uint8 *)obj->pInPlanes;
pOutPlanes = (vx_uint8 *)obj->pOutPlanes;
/* Deinterleave RGB to planar */
uint32_t chOffset = img_width * img_height;
for(i = 0; i < img_height; i ++)
{
pData = (vx_uint8 *)data_ptr + (i * img_stride);
for(j = 0; j < img_width; j ++)
{
pInPlanes[(0 * chOffset) + (i * img_width) + j] = *pData++;
pInPlanes[(1 * chOffset) + (i * img_width) + j] = *pData++;
pInPlanes[(2 * chOffset) + (i * img_width) + j] = *pData++;
}
}
APP_PRINTF("app_tidl: Resizing image ...\n");
/* Resize image to network input resolution */
//resizeImage(pInPlanes, pOutPlanes, img_width, img_height, img_width, 256, 256, 256);
resizeImage(pInPlanes, pOutPlanes, img_width, img_height, img_width, ioBufDesc->inWidth[id], ioBufDesc->inHeight[id], ioBufDesc->inWidth[id]);
APP_PRINTF("app_tidl: Rearranging data ...\n");
if((data_type == VX_TYPE_INT8) ||
(data_type == VX_TYPE_UINT8))
{
vx_uint8 *pR = NULL;
vx_uint8 *pG = NULL;
vx_uint8 *pB = NULL;
/* Reset the input buffer, this will take care of padding requirement for TIDL */
memset(input_buffer, 0, capacity);
if(ioBufDesc->inDataFormat[id] == 1) /* RGB */
{
start_offset = (0 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pR = (vx_uint8 *)input_buffer + start_offset;
start_offset = (1 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pG = (vx_uint8 *)input_buffer + start_offset;
start_offset = (2 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pB = (vx_uint8 *)input_buffer + start_offset;
}
else if(ioBufDesc->inDataFormat[id] == 0) /* BGR */
{
start_offset = (0 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pB = (vx_uint8 *)input_buffer + start_offset;
start_offset = (1 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pG = (vx_uint8 *)input_buffer + start_offset;
start_offset = (2 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pR = (vx_uint8 *)input_buffer + start_offset;
}
//chOffset = 256*256;
chOffset = ioBufDesc->inWidth[id]*ioBufDesc->inHeight[id];
if ( (NULL != pR) && (NULL != pG) && (NULL != pB) )
{
/* Write image in network required format */
for(i = 0; i < ioBufDesc->inHeight[id]; i++)
{
//Center crop 224x224 image from 256x256 resized input
//uint32_t offset = ((16 + i) * 256) + 16;
uint32_t offset = i * ioBufDesc->inWidth[id];
for(j = 0; j < ioBufDesc->inWidth[id]; j++)
{
pR[j] = pOutPlanes[(0 * chOffset) + offset + j];
pG[j] = pOutPlanes[(1 * chOffset) + offset + j];
pB[j] = pOutPlanes[(2 * chOffset) + offset + j];
}
pR += input_sizes[0];
pG += input_sizes[0];
pB += input_sizes[0];
}
}
}
else if((data_type == VX_TYPE_INT16) ||
(data_type == VX_TYPE_UINT16))
{
vx_uint16 *pR = NULL;
vx_uint16 *pG = NULL;
vx_uint16 *pB = NULL;
/* Reset the input buffer, this will take care of padding requirement for TIDL */
memset(input_buffer, 0, capacity);
if(ioBufDesc->inDataFormat[id] == 1) /* RGB */
{
start_offset = (0 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pR = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
start_offset = (1 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pG = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
start_offset = (2 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pB = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
}
else if(ioBufDesc->inDataFormat[id] == 0) /* BGR */
{
start_offset = (0 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pB = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
start_offset = (1 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pG = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
start_offset = (2 * input_strides[2]) + (ioBufDesc->inPadT[id] * input_strides[1]) + (ioBufDesc->inPadL[id] * input_strides[0]);
pR = (vx_uint16 *)((vx_uint8 *)input_buffer + start_offset);
}
chOffset = 256*256;
chOffset = ioBufDesc->inWidth[id]*ioBufDesc->inHeight[id];
if ( (NULL != pR) && (NULL != pG) && (NULL != pB) )
{
/* Write image in network required format */
for(i = 0; i < ioBufDesc->inHeight[id]; i++)
{
//Center crop 224x224 image from 256x256 resized input
//uint32_t offset = ((16 + i) * 256) + 16;
uint32_t offset = i * ioBufDesc->inWidth[id];
for(j = 0; j < ioBufDesc->inWidth[id]; j++)
{
pR[j] = pOutPlanes[(0 * chOffset) + offset + j];
pG[j] = pOutPlanes[(1 * chOffset) + offset + j];
pB[j] = pOutPlanes[(2 * chOffset) + offset + j];
}
pR += input_sizes[0];
pG += input_sizes[0];
pB += input_sizes[0];
}
}
}
APP_PRINTF("app_tidl: Image Pre processing ... Done.\n");
}
tivxUnmapTensorPatch(input_tensors[id], map_id_input);
}
}
return status;
}
static void displayOutput(AppObj *obj, vx_user_data_object config, vx_tensor *output_tensors, char *output_file)
{
vx_status status = VX_SUCCESS;
vx_size output_sizes[APP_MAX_TENSOR_DIMS];
int32_t id, i, j;
sTIDL_IOBufDesc_t *ioBufDesc;
Draw2D_FontPrm sClassPrm;
ioBufDesc = &obj->ioBufDesc;
for(id = 0; id < ioBufDesc->numOutputBuf; id++)
{
vx_size data_type = getTensorDataType(ioBufDesc->outElementType[id]);
output_sizes[0] = ioBufDesc->outWidth[id] + ioBufDesc->outPadL[id] + ioBufDesc->outPadR[id];
output_sizes[1] = ioBufDesc->outHeight[id] + ioBufDesc->outPadT[id] + ioBufDesc->outPadB[id];
output_sizes[2] = ioBufDesc->outNumChannels[id];
if(status == VX_SUCCESS)
{
status = vxGetStatus((vx_reference)output_tensors[id]);
}
if (VX_SUCCESS == status)
{
void *output_buffer;
vx_map_id map_id_output;
vx_size output_strides[APP_MAX_TENSOR_DIMS];
vx_size start[APP_MAX_TENSOR_DIMS];
start[0] = start[1] = start[2] = start[3] = 0;
output_strides[0] = sizeof(vx_int8);
if((data_type == VX_TYPE_INT8) ||
(data_type == VX_TYPE_UINT8))
{
output_strides[0] = sizeof(vx_int8);
}
else if((data_type == VX_TYPE_INT16) ||
(data_type == VX_TYPE_UINT16))
{
output_strides[0] = sizeof(vx_int16);
}
else if(data_type == VX_TYPE_FLOAT32)
{
output_strides[0] = sizeof(vx_float32);
}
output_strides[1] = output_sizes[0] * output_strides[0];
output_strides[2] = output_sizes[1] * output_strides[1];
status = tivxMapTensorPatch(output_tensors[id], 3, start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
if(status == VX_SUCCESS)
{
vx_uint32 classid[5] = {0};
vx_uint32 label_offset;
label_offset = 0;
if(ioBufDesc->outWidth[id] == 1001)
{
label_offset = 0;
}
else if(ioBufDesc->outWidth[id] == 1000)
{
label_offset = 1;
}
APP_PRINTF("app_tidl: Finding top-5 ... \n");
if(data_type == VX_TYPE_FLOAT32)
{
float *pOut;
float score[5];
pOut = (float *)output_buffer + (ioBufDesc->outPadT[id] * output_sizes[0]) + ioBufDesc->outPadL[id];
for(i = 0; i < 5; i++)
{
score[i] = FLT_MIN;
classid[i] = (vx_uint32)-1;
for(j = 0; j < ioBufDesc->outWidth[id]; j++)
{
if(pOut[j] > score[i])
{
score[i] = pOut[j];
classid[i] = j;
}
}
if(classid[i] < ioBufDesc->outWidth[id] )
{
pOut[classid[i]] = FLT_MIN;
}
else
{
classid[i] = 0; /* invalid class ID, ideally it should not reach here */
}
}
#ifdef APP_DEBUG
APP_PRINTF("app_tidl: Image classification Top-5 results: \n");
for(i = 0; i < 5; i++)
{
APP_PRINTF("app_tidl: %s, class-id: %d, score: %f \n", (char *)&imgnet_labels[classid[i] + label_offset], classid[i], score[i]);
}
#endif
}
else if(data_type == VX_TYPE_INT16)
{
vx_int16 *pOut;
vx_int16 score[5];
pOut = (vx_int16 *)output_buffer + (ioBufDesc->outPadT[id] * output_sizes[0]) + ioBufDesc->outPadL[id];
for(i = 0; i < 5; i++)
{
score[i] = INT16_MIN;
classid[i] = (vx_uint32)-1;
for(j = 0; j < ioBufDesc->outWidth[id]; j++)
{
if(pOut[j] > score[i])
{
score[i] = pOut[j];
classid[i] = j;
}
}
if(classid[i] < ioBufDesc->outWidth[id] )
{
pOut[classid[i]] = INT16_MIN;
}
else
{
classid[i] = 0; /* invalid class ID, ideally it should not reach here */
}
}
#ifdef APP_DEBUG
APP_PRINTF("app_tidl: Image classification Top-5 results: \n");
for(i = 0; i < 5; i++)
{
APP_PRINTF("app_tidl: %s, class-id: %d, score: %d \n", (char *)&imgnet_labels[classid[i] + label_offset], classid[i], score[i]);
}
#endif
}
else if(data_type == VX_TYPE_INT8)
{
vx_int8 *pOut;
vx_int8 score[5];
pOut = (vx_int8 *)output_buffer + (ioBufDesc->outPadT[id] * output_sizes[0]) + ioBufDesc->outPadL[id];
for(i = 0; i < 5; i++)
{
score[i] = INT8_MIN;
classid[i] = (vx_uint32)-1;
for(j = 0; j < ioBufDesc->outWidth[id]; j++)
{
if(pOut[j] > score[i])
{
score[i] = pOut[j];
classid[i] = j;
}
}
if(classid[i] < ioBufDesc->outWidth[id] )
{
pOut[classid[i]] = INT8_MIN;
}
else
{
classid[i] = 0; /* invalid class ID, ideally it should not reach here */
}
}
#ifdef APP_DEBUG
APP_PRINTF("app_tidl: Image classification Top-5 results: \n");
for(i = 0; i < 5; i++)
{
APP_PRINTF("app_tidl: %s, class-id: %d, score: %d \n", (char *)&imgnet_labels[classid[i] + label_offset], classid[i], score[i]);
}
#endif
}
APP_PRINTF("app_tidl: Finding top-5 ... Done \n");
Draw2D_clearRegion(obj->pHndl, (DISPLAY_WIDTH/2) + 40, 200, 600, 300);
sClassPrm.fontIdx = 1;
for(i = 0; i < 5; i++)
{
Draw2D_drawString(obj->pHndl, (DISPLAY_WIDTH/2) + 40, 200 + (i * 40), (char *)&imgnet_labels[classid[i] + label_offset], &sClassPrm);
}
}
tivxUnmapTensorPatch(output_tensors[id], map_id_output);
}
}
{
/* Convert RGB565 to RGB888 before writing output */
for(i = 0, j = 0; i < (DISPLAY_WIDTH * DISPLAY_HEIGHT); i++)
{
uint16_t RGB_565_val = obj->pDisplayBuf565[i];
obj->pDisplayBuf888[j + 0] = (RGB_565_val & 0x1F) << 3;
obj->pDisplayBuf888[j + 1] = ((RGB_565_val >> 5) & 0x3F) << 2;
obj->pDisplayBuf888[j + 2] = ((RGB_565_val >> 11) & 0x1F) << 3;
j += 3;
}
if((obj->img_width <= (DISPLAY_WIDTH/2)) && (obj->img_height <= (DISPLAY_HEIGHT - 200)))
{
uint32_t startX = ((DISPLAY_WIDTH/2) / 2) - (obj->img_width / 2);
uint32_t startY = ((DISPLAY_HEIGHT/2) / 2) - (obj->img_height / 2);
uint32_t imgOffset = 200;
for(i = 0; i < obj->img_height; i++)
{
uint8_t *pOut = &obj->pDisplayBuf888[((imgOffset + startY + i) * DISPLAY_WIDTH * 3) + (startX * 3)];
uint8_t *pIn = obj->data_ptr + (i * obj->img_stride);
for(j = 0; j < obj->img_width; j++)
{
*pOut++ = *pIn++;
*pOut++ = *pIn++;
*pOut++ = *pIn++;
}
}
}
}
if (obj->display_option == 0)
{
tivx_utils_bmp_write(output_file, obj->pDisplayBuf888, DISPLAY_WIDTH, DISPLAY_HEIGHT, (DISPLAY_WIDTH * 3), obj->df_image);
}
/* Release the bmp buffer created in readInput() */
tivx_utils_bmp_read_release(&obj->imgParams);
APP_PRINTF("app_tidl: Showing output ... Done.\n");
}
static vx_size getTensorDataType(vx_int32 tidl_type)
{
vx_size openvx_type = VX_TYPE_INVALID;
if (tidl_type == TIDL_UnsignedChar)
{
openvx_type = VX_TYPE_UINT8;
}
else if(tidl_type == TIDL_SignedChar)
{
openvx_type = VX_TYPE_INT8;
}
else if(tidl_type == TIDL_UnsignedShort)
{
openvx_type = VX_TYPE_UINT16;
}
else if(tidl_type == TIDL_SignedShort)
{
openvx_type = VX_TYPE_INT16;
}
else if(tidl_type == TIDL_UnsignedWord)
{
openvx_type = VX_TYPE_UINT32;
}
else if(tidl_type == TIDL_SignedWord)
{
openvx_type = VX_TYPE_INT32;
}
else if(tidl_type == TIDL_SinglePrecFloat)
{
openvx_type = VX_TYPE_FLOAT32;
}
return openvx_type;
}
#ifdef APP_WRITE_PRE_PROC_OUTPUT
static vx_status writePreProcOutput(char* file_name, vx_tensor output)
{
vx_status status = VX_SUCCESS;
vx_size num_dims;
vx_enum data_type;
void *data_ptr;
vx_map_id map_id;
vx_size start[APP_MAX_TENSOR_DIMS];
vx_size tensor_strides[APP_MAX_TENSOR_DIMS];
vx_size tensor_sizes[APP_MAX_TENSOR_DIMS];
vxQueryTensor(output, VX_TENSOR_NUMBER_OF_DIMS, &num_dims, sizeof(vx_size));
if(num_dims != 3)
{
printf("Number of dims are != 3! exiting.. \n");
return VX_FAILURE;
}
vxQueryTensor(output, (vx_enum)VX_TENSOR_DIMS, tensor_sizes, 3 * sizeof(vx_size));
vxQueryTensor(output, (vx_enum)VX_TENSOR_DATA_TYPE, &data_type, sizeof(data_type));
start[0] = start[1] = start[2] = 0;
tensor_strides[0] = sizeof(vx_int8);
if((data_type == VX_TYPE_INT8) ||
(data_type == VX_TYPE_UINT8))
{
tensor_strides[0] = sizeof(vx_int8);
}
else if((data_type == VX_TYPE_INT16) ||
(data_type == VX_TYPE_UINT16))
{
tensor_strides[0] = sizeof(vx_int16);
}
else if((data_type == VX_TYPE_FLOAT32))
{
tensor_strides[0] = sizeof(vx_float32);
}
tensor_strides[1] = tensor_strides[0] * tensor_strides[0];
tensor_strides[2] = tensor_strides[1] * tensor_strides[1];
status = tivxMapTensorPatch(output, num_dims, start, tensor_sizes, &map_id, tensor_strides, &data_ptr, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);
if(VX_SUCCESS == status)
{
vx_char new_name[APP_MAX_FILE_PATH];
snprintf(new_name, APP_MAX_FILE_PATH, "%s_%dx%d.rgb", file_name, (uint32_t)tensor_sizes[0], (uint32_t)tensor_sizes[1]);
FILE *fp = fopen(new_name, "wb");
if(NULL == fp)
{
printf("Unable to open file %s for writing!\n", new_name);
return VX_FAILURE;
}
fwrite(data_ptr, 1, num_dims * tensor_strides[2], fp);
fclose(fp);
tivxUnmapTensorPatch(output, map_id);
}
return(status);
}
#endif
===============
我看完 0130 日志后，确定还有一个更关键的结构性问题，这能解释为什么修了 stride 和 map/unmap 仍没改善。
⚠ 根因：检测头索引错位（最关键）
当前模型输出中：
- 检测头是 output[1], output[2], output[3]
- 但 getMult_yolov5e2eResult() 内部固定用 head_idx=0..2
你在 main.c 里把检测头参数写入 index=1..3，
但解码函数只读 index=0..2。
这会导致：
- head_idx=0 的数据未初始化
- head_idx=1/2 读取的是错 head 的 anchors / params
- 结果就是坐标错位、整体几何错误
这也是为什么图像一直"差不多错"，不会随机炸掉，但永远不对。
✅ 我已经帮你把索引统一了
在 main.c 中做了 检测头索引重映射：
- 将 zh_idx=1..3 对应到 dst_idx=0..2
- 这样 getMult_yolov5e2eResult() 就会按 正确顺序 解码
同时日志也会打印：
det_head[1->0]
det_head[2->1]
det_head[3->2]
✅ 你明天只要做这几步（最快验证）
- 重新编译下位机程序
- 运行推理
- 看日志里是否出现 det_head[1->0]
- 再跑 draw_tensor2psd.py
- 看 _model.bmp 是否正常
为什么我确认这是关键问题？
- 模型输出 head 的尺寸完全对应 P5/P4/P3：
19×23、38×46、76×92
- 这些应该对应 head_idx=0/1/2
- 但你实际把它们放在 index=1/2/3
- 结果就是 anchors 与 head 对不上
这会导致框位置整体错误，但不会崩溃------非常符合你的现象。
如果这次改完仍不对，我会再帮你继续定位，但这个问题必须先修正。
明天你跑完日志和 _model.bmp 发给我，我们就能迅速确认是否闭环。