Skip to content

[feat](benchmark): replace load_param with load_param_mem #6286

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
nihui merged 18 commits into from
Jan 8, 2026
+175 −54
Merged

[feat](benchmark): replace load_param with load_param_mem #6286

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
18 commits
Select commit Hold shift + click to select a range
a555f51
[feat](benchmark): replace load_param with load_param_mem
0130w Aug 26, 2025
f4999de
[docs](benchmark): update benchmark docs
0130w Aug 26, 2025
1643c32
Merge branch 'master' into dev
0130w Aug 26, 2025
41745c8
fix
0130w Aug 31, 2025
aae6ad8
Merge branch 'master' into dev
0130w Nov 30, 2025
74ae9d1
refactor(benchmark): remove model registry and optimize param loading
0130w Nov 30, 2025
086e73c
fix(benchmark): correct README param description
0130w Nov 30, 2025
41f463d
refactor(benchmark): rename header file
0130w Dec 1, 2025
1116143
Merge branch 'master' into dev
nihui Jan 6, 2026
b1fbd2e
wip
nihui Jan 6, 2026
49b6e6c
Update cmake/ncnn_generate_param_header.cmake
nihui Jan 8, 2026
c6bf3cc
Update benchmark/benchncnn.cpp
nihui Jan 8, 2026
b181924
Update cmake/ncnn_add_param.cmake
nihui Jan 8, 2026
8132739
Update benchncnn.cpp
nihui Jan 8, 2026
77e57d7
w
nihui Jan 8, 2026
8293ab9
Revert "w"
nihui Jan 8, 2026
b5b6558
Merge branch 'master' into dev
nihui Jan 8, 2026
85399e8
update doc
nihui Jan 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
49 changes: 49 additions & 0 deletions benchmark/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,3 +13,52 @@ endif()

# add benchncnn to a virtual project group
set_property(TARGET benchncnn PROPERTY FOLDER "benchmark")

# generate model data
include(${CMAKE_CURRENT_SOURCE_DIR}/../cmake/ncnn_generate_model_data_header.cmake)

set(embedded_model_param_includes "")

set(PARAM_FILES
"${CMAKE_CURRENT_SOURCE_DIR}/squeezenet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/squeezenet_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_v2.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_v3.param"
"${CMAKE_CURRENT_SOURCE_DIR}/shufflenet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/shufflenet_v2.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mnasnet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/proxylessnasnet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/efficientnet_b0.param"
"${CMAKE_CURRENT_SOURCE_DIR}/efficientnetv2_b0.param"
"${CMAKE_CURRENT_SOURCE_DIR}/regnety_400m.param"
"${CMAKE_CURRENT_SOURCE_DIR}/blazeface.param"
"${CMAKE_CURRENT_SOURCE_DIR}/googlenet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/googlenet_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/resnet18.param"
"${CMAKE_CURRENT_SOURCE_DIR}/resnet18_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/alexnet.param"
"${CMAKE_CURRENT_SOURCE_DIR}/vgg16.param"
"${CMAKE_CURRENT_SOURCE_DIR}/vgg16_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/resnet50.param"
"${CMAKE_CURRENT_SOURCE_DIR}/resnet50_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/squeezenet_ssd.param"
"${CMAKE_CURRENT_SOURCE_DIR}/squeezenet_ssd_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_ssd.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_ssd_int8.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenet_yolo.param"
"${CMAKE_CURRENT_SOURCE_DIR}/mobilenetv2_yolov3.param"
"${CMAKE_CURRENT_SOURCE_DIR}/yolov4-tiny.param"
"${CMAKE_CURRENT_SOURCE_DIR}/nanodet_m.param"
"${CMAKE_CURRENT_SOURCE_DIR}/yolo-fastest-1.1.param"
"${CMAKE_CURRENT_SOURCE_DIR}/yolo-fastestv2.param"
"${CMAKE_CURRENT_SOURCE_DIR}/vision_transformer.param"
"${CMAKE_CURRENT_SOURCE_DIR}/FastestDet.param"
)

foreach(PARAM_FILE ${PARAM_FILES})
ncnn_convert_param_file(${PARAM_FILE})
endforeach()

file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/embedded_model_params.h "${embedded_model_param_includes}")
96 changes: 46 additions & 50 deletions benchmark/benchncnn.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,8 @@
#include "net.h"
#include "gpu.h"

#include "../benchmark/embedded_model_params.h"

#ifndef NCNN_SIMPLESTL
#include <vector>
#endif
Expand Down Expand Up @@ -46,12 +48,12 @@ static ncnn::VkAllocator* g_blob_vkallocator = 0;
static ncnn::VkAllocator* g_staging_vkallocator = 0;
#endif // NCNN_VULKAN

void benchmark(const char* comment, const std::vector<ncnn::Mat>& _in, const ncnn::Option& opt, bool fixed_path = true)
void benchmark(const char* comment, const std::vector<ncnn::Mat>& _in, const ncnn::Option& opt, const char* model_param_data = NULL)
{
// Skip if int8 model name and using GPU
if (opt.use_vulkan_compute && strstr(comment, "int8") != NULL)
{
if (!fixed_path)
if (!model_param_data)
fprintf(stderr, "%20s skipped (int8+GPU not supported)\n", comment);
return;
}
Expand All @@ -78,17 +80,11 @@ void benchmark(const char* comment, const std::vector<ncnn::Mat>& _in, const ncn
}
#endif // NCNN_VULKAN

#ifdef __EMSCRIPTEN__
#define MODEL_DIR "/working/"
#else
#define MODEL_DIR ""
#endif
const char* model_param = NULL;

if (fixed_path)
if (model_param_data)
{
char parampath[256];
sprintf(parampath, MODEL_DIR "%s.param", comment);
net.load_param(parampath);
net.load_param_mem(model_param_data);
}
else
{
Expand Down Expand Up @@ -173,11 +169,11 @@ void benchmark(const char* comment, const std::vector<ncnn::Mat>& _in, const ncn
fprintf(stderr, "%20s min = %7.2f max = %7.2f avg = %7.2f\n", comment, time_min, time_max, time_avg);
}

void benchmark(const char* comment, const ncnn::Mat& _in, const ncnn::Option& opt, bool fixed_path = true)
void benchmark(const char* comment, const ncnn::Mat& _in, const ncnn::Option& opt, const char* model_param_data = NULL)
{
std::vector<ncnn::Mat> inputs;
inputs.push_back(_in);
return benchmark(comment, inputs, opt, fixed_path);
return benchmark(comment, inputs, opt, model_param_data);
}

void show_usage()
Expand Down Expand Up @@ -387,82 +383,82 @@ int main(int argc, char** argv)
if (model != 0)
{
// run user defined benchmark
benchmark(model, inputs, opt, false);
benchmark(model, inputs, opt);
}
else
{
// run default cases
benchmark("squeezenet", ncnn::Mat(227, 227, 3), opt);
benchmark("squeezenet", ncnn::Mat(227, 227, 3), opt, squeezenet_param_data);

benchmark("squeezenet_int8", ncnn::Mat(227, 227, 3), opt);
benchmark("squeezenet_int8", ncnn::Mat(227, 227, 3), opt, squeezenet_int8_param_data);

benchmark("mobilenet", ncnn::Mat(224, 224, 3), opt);
benchmark("mobilenet", ncnn::Mat(224, 224, 3), opt, mobilenet_param_data);

benchmark("mobilenet_int8", ncnn::Mat(224, 224, 3), opt);
benchmark("mobilenet_int8", ncnn::Mat(224, 224, 3), opt, mobilenet_int8_param_data);

benchmark("mobilenet_v2", ncnn::Mat(224, 224, 3), opt);
benchmark("mobilenet_v2", ncnn::Mat(224, 224, 3), opt, mobilenet_v2_param_data);

// benchmark("mobilenet_v2_int8", ncnn::Mat(224, 224, 3), opt);
// benchmark("mobilenet_v2_int8", ncnn::Mat(224, 224, 3), opt, mobilenet_v2_int8_param_data);

benchmark("mobilenet_v3", ncnn::Mat(224, 224, 3), opt);
benchmark("mobilenet_v3", ncnn::Mat(224, 224, 3), opt, mobilenet_v3_param_data);

benchmark("shufflenet", ncnn::Mat(224, 224, 3), opt);
benchmark("shufflenet", ncnn::Mat(224, 224, 3), opt, shufflenet_param_data);

benchmark("shufflenet_v2", ncnn::Mat(224, 224, 3), opt);
benchmark("shufflenet_v2", ncnn::Mat(224, 224, 3), opt, shufflenet_v2_param_data);

benchmark("mnasnet", ncnn::Mat(224, 224, 3), opt);
benchmark("mnasnet", ncnn::Mat(224, 224, 3), opt, mnasnet_param_data);

benchmark("proxylessnasnet", ncnn::Mat(224, 224, 3), opt);
benchmark("proxylessnasnet", ncnn::Mat(224, 224, 3), opt, proxylessnasnet_param_data);

benchmark("efficientnet_b0", ncnn::Mat(224, 224, 3), opt);
benchmark("efficientnet_b0", ncnn::Mat(224, 224, 3), opt, efficientnet_b0_param_data);

benchmark("efficientnetv2_b0", ncnn::Mat(224, 224, 3), opt);
benchmark("efficientnetv2_b0", ncnn::Mat(224, 224, 3), opt, efficientnetv2_b0_param_data);

benchmark("regnety_400m", ncnn::Mat(224, 224, 3), opt);
benchmark("regnety_400m", ncnn::Mat(224, 224, 3), opt, regnety_400m_param_data);

benchmark("blazeface", ncnn::Mat(128, 128, 3), opt);
benchmark("blazeface", ncnn::Mat(128, 128, 3), opt, blazeface_param_data);

benchmark("googlenet", ncnn::Mat(224, 224, 3), opt);
benchmark("googlenet", ncnn::Mat(224, 224, 3), opt, googlenet_param_data);

benchmark("googlenet_int8", ncnn::Mat(224, 224, 3), opt);
benchmark("googlenet_int8", ncnn::Mat(224, 224, 3), opt, googlenet_int8_param_data);

benchmark("resnet18", ncnn::Mat(224, 224, 3), opt);
benchmark("resnet18", ncnn::Mat(224, 224, 3), opt, resnet18_param_data);

benchmark("resnet18_int8", ncnn::Mat(224, 224, 3), opt);
benchmark("resnet18_int8", ncnn::Mat(224, 224, 3), opt, resnet18_int8_param_data);

benchmark("alexnet", ncnn::Mat(227, 227, 3), opt);
benchmark("alexnet", ncnn::Mat(227, 227, 3), opt, alexnet_param_data);

benchmark("vgg16", ncnn::Mat(224, 224, 3), opt);
benchmark("vgg16", ncnn::Mat(224, 224, 3), opt, vgg16_param_data);

benchmark("vgg16_int8", ncnn::Mat(224, 224, 3), opt);
benchmark("vgg16_int8", ncnn::Mat(224, 224, 3), opt, vgg16_int8_param_data);

benchmark("resnet50", ncnn::Mat(224, 224, 3), opt);
benchmark("resnet50", ncnn::Mat(224, 224, 3), opt, resnet50_param_data);

benchmark("resnet50_int8", ncnn::Mat(224, 224, 3), opt);
benchmark("resnet50_int8", ncnn::Mat(224, 224, 3), opt, resnet50_int8_param_data);

benchmark("squeezenet_ssd", ncnn::Mat(300, 300, 3), opt);
benchmark("squeezenet_ssd", ncnn::Mat(300, 300, 3), opt, squeezenet_ssd_param_data);

benchmark("squeezenet_ssd_int8", ncnn::Mat(300, 300, 3), opt);
benchmark("squeezenet_ssd_int8", ncnn::Mat(300, 300, 3), opt, squeezenet_ssd_int8_param_data);

benchmark("mobilenet_ssd", ncnn::Mat(300, 300, 3), opt);
benchmark("mobilenet_ssd", ncnn::Mat(300, 300, 3), opt, mobilenet_ssd_param_data);

benchmark("mobilenet_ssd_int8", ncnn::Mat(300, 300, 3), opt);
benchmark("mobilenet_ssd_int8", ncnn::Mat(300, 300, 3), opt, mobilenet_ssd_int8_param_data);

benchmark("mobilenet_yolo", ncnn::Mat(416, 416, 3), opt);
benchmark("mobilenet_yolo", ncnn::Mat(416, 416, 3), opt, mobilenet_yolo_param_data);

benchmark("mobilenetv2_yolov3", ncnn::Mat(352, 352, 3), opt);
benchmark("mobilenetv2_yolov3", ncnn::Mat(352, 352, 3), opt, mobilenetv2_yolov3_param_data);

benchmark("yolov4-tiny", ncnn::Mat(416, 416, 3), opt);
benchmark("yolov4-tiny", ncnn::Mat(416, 416, 3), opt, yolov4_tiny_param_data);

benchmark("nanodet_m", ncnn::Mat(320, 320, 3), opt);
benchmark("nanodet_m", ncnn::Mat(320, 320, 3), opt, nanodet_m_param_data);

benchmark("yolo-fastest-1.1", ncnn::Mat(320, 320, 3), opt);
benchmark("yolo-fastest-1.1", ncnn::Mat(320, 320, 3), opt, yolo_fastest_1_1_param_data);

benchmark("yolo-fastestv2", ncnn::Mat(352, 352, 3), opt);
benchmark("yolo-fastestv2", ncnn::Mat(352, 352, 3), opt, yolo_fastestv2_param_data);

benchmark("vision_transformer", ncnn::Mat(384, 384, 3), opt);
benchmark("vision_transformer", ncnn::Mat(384, 384, 3), opt, vision_transformer_param_data);

benchmark("FastestDet", ncnn::Mat(352, 352, 3), opt);
benchmark("FastestDet", ncnn::Mat(352, 352, 3), opt, FastestDet_param_data);
}
#if NCNN_VULKAN
delete g_blob_vkallocator;
Expand Down
41 changes: 41 additions & 0 deletions cmake/ncnn_generate_model_data_header.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,41 @@

macro(ncnn_convert_param_file PARAM_FILE)
# Use a macro to convert the contents of a single parameter file into memory content
file(READ ${PARAM_FILE} param_file_data)

# remove whitespace
string(REGEX REPLACE "\n +" "\n" param_file_data ${param_file_data})

# remove empty line
string(REGEX REPLACE "\n\n" "\n" param_file_data ${param_file_data})

# Get the file name with extension
get_filename_component(PARAM_FILE_NAME ${PARAM_FILE} NAME)
# Manually remove ".param" since NAME_WE treats ".1.param" as a multi-extension
string(REPLACE ".param" "" PARAM_FILE_NAME_WE "${PARAM_FILE_NAME}")
# Replace characters invalid in C identifiers ('.' and '-') with underscores
string(REPLACE "." "_" PARAM_FILE_NAME_WE "${PARAM_FILE_NAME_WE}")
string(REPLACE "-" "_" PARAM_FILE_NAME_WE "${PARAM_FILE_NAME_WE}")
# Check if the result is empty
if (NOT PARAM_FILE_NAME_WE)
message(FATAL_ERROR "Failed to extract valid filename from '${PARAM_FILE}'")
endif()
# Check if the extracted filename is a valid C identifier
string(REGEX MATCH "^[A-Za-z_][A-Za-z0-9_]*$" is_valid "${PARAM_FILE_NAME_WE}")
if (NOT is_valid)
message(FATAL_ERROR "Extracted filename '${PARAM_FILE_NAME_WE}' is not a valid C identifier")
endif()

# text to hex
file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/param_hex_data/${PARAM_FILE_NAME_WE}.text2hex.txt "${param_file_data}")
file(READ ${CMAKE_CURRENT_BINARY_DIR}/param_hex_data/${PARAM_FILE_NAME_WE}.text2hex.txt param_file_data_hex HEX)
string(REGEX REPLACE "([0-9a-f][0-9a-f])" "0x\\1," param_file_data_hex ${param_file_data_hex})
string(FIND "${param_file_data_hex}" "," tail_comma REVERSE)
string(SUBSTRING "${param_file_data_hex}" 0 ${tail_comma} param_file_data_hex)

# generate model param header file
file(WRITE ${CMAKE_CURRENT_BINARY_DIR}/model_param_header/${PARAM_FILE_NAME_WE}.comp.hex.h "static const char ${PARAM_FILE_NAME_WE}_param_data[] = {${param_file_data_hex},0x00};\n")

# append include line to a CMake variable for later output
string(APPEND embedded_model_param_includes "#include \"model_param_header/${PARAM_FILE_NAME_WE}.comp.hex.h\"\n")
endmacro()
Loading