dnn: Add a Vulkan based backend #12703

wzw-intel · 2018-09-30T08:52:42Z

This commit adds a new backend "DNN_BACKEND_VKCOM" and a
new target "DNN_TARGET_VULKAN". VKCOM means vulkan based
computation library.

This backend uses Vulkan API and SPIR-V shaders to do
the inference computation for layers. The layer types
that implemented in DNN_BACKEND_VKCOM include:
Conv, Concat, ReLU, LRN, PriorBox, Softmax, MaxPooling,
AvePooling, Permute

This is just a beginning work for Vulkan in OpenCV DNN,
more layer types will be supported and performance
tuning is on the way.

Build OpenCV DNN with Vulkan support

Due to usage of dynamic load mechanism, there is no compile-time dependency on Vulkan Loader Library or vendor driver. You just need add "-DWITH_VULKAN=ON" into cmake options to enable the DNN Vukand support.

Enable DNN Vulkan runtime support

Install Vulkan Loader library
You have 2 choices to install Vulkan Loader library:
1.1 Install the library from Vulkan SDK
Download Vulkan SDK from https://vulkan.lunarg.com/sdk/home#sdk/downloadConfirm/1.1.85.0/linux/vulkansdk-linux-x86_64-1.1.85.0.tar.gz
'mkdir ~/vulkan_sdk'
'tar zxf vulkansdk-linux-x86_64-1.1.85.0.tar.gz -C ~/vulkan_sdk'
Your library is at "$HOME/vulkan_sdk/1.x.x.x/x86_64/libvulkan.so.1"

1.2 Install the library by using apt1.
'sudo apt-get install libvulkan1'
Your library will be loacated at "/usr/lib/x86_64-linux-gnu/libvulkan.so.1"
Install vendor driver (For Intel and AMD GPU)
Also 2 choices:
2.1 Compiling and installing Mesa
Ref https://www.mesa3d.org/install.html for HOWTO.
Additionally, add option "--with-vulkan-drivers=intel" when doing configure to build intel Vulkan driver.
"--with-vulkan-drivers=radeon" for AMD GPU.

2.2 Install the library from apt
'sudo apt-get install mesa-vulkan-drivers'
Install Vulkan utils (optional)
If you have installed Vulkan SDK follow 1.1, you already got the utils.
'source $HOME/vulkan_sdk/1.x.x.x/setup-env.sh' to add the utils path into $PATH
You can also install the vulkan utils by 'sudo apt-get install vulkan-utils'
Verify your Vulkan installation (optional)
run 'vulkaninfo' and if you can see output string like "GPU id : 0 (Intel(R) xxxxxx", that means Vulkan works.

Run test

Before run test, if you installed Vulkan Loader library as 1.1,
you need 'export OPENCV_VULKAN_RUNTIME=$HOME/vulkan_sdk/1.x.x.x/x86_64/libvulkan.so.1'
run '$OPENCV_BUILD_DIR/bin/opencv_test_dnn'
The results with "VKCOM/VULKAN" suffix are for our backend.

How to patch GLSL shaders

The GLSL shader source code are in $OPENCV_SOURCE_DIR/module/dnn/src/vkcom/shader/, with suffix ".comp".

Once you modify the shader source code or add new shaders, go into $OPENCV_SOURCE_DIR/module/dnn/src/vkcom/shader/ and run 'python spirv_generator.py' to compile them into SPIR-V format and attach the binary into .cpp
files. And rebuild the DNN module.

In order to run this script successfully, you need tool "glslangValidator" which is provied by Vulkan SDK.

Signed-off-by: Wu Zhiwen zhiwen.wu@intel.com

force_builders=Custom
docker_image:Custom=ubuntu-vulkan:16.04
buildworker:Custom=linux-4

vpisarev · 2018-10-01T11:24:56Z

@wzw-intel, as I see, the patch includes quite a bit of binary blobs, which are nearly impossible to support by people different from the author. Is it possible to replace those blobs with the source code?

wzw-intel · 2018-10-02T01:55:57Z

@vpisarev, thanks for your comments.

Shader source code are in the dnn/src/vkcom/shader directory and with suffix ".comp", but the vkcom backend will directly use the bianry format of shader instead of compiling the .comp on the runtime.

The reason is Vulkan driver don't cover the compiling from glsl source code to SPIR-V binary. If we want this functionality, 3 choices:

On the build time, use tool "glslangValidator", provided by Vulkan SDK, to make this compilation. This method introduce dependency on tool "glslangValidate". In this PR, to avoid this denpendency, I provide the pre-built binaries in the source code.
Use 3rd party helper library, like libshaderc (https://github.com/google/shaderc/tree/master/libshaderc). This introduce additional dependency.
Copy source code of 3rd party library into vkcom, so that we have no dependency on additional library and can build the shader on the runtime.

What is your opinion?

This commit adds a new backend "DNN_BACKEND_VKCOM" and a new target "DNN_TARGET_VULKAN". VKCOM means vulkan based computation library. This backend uses Vulkan API and SPIR-V shaders to do the inference computation for layers. The layer types that implemented in DNN_BACKEND_VKCOM include: Conv, Concat, ReLU, LRN, PriorBox, Softmax, MaxPooling, AvePooling, Permute This is just a beginning work for Vulkan in OpenCV DNN, more layer types will be supported and performance tuning is on the way. Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

jljusten · 2018-10-02T03:32:11Z

@wzw-intel: I don't work on opencv, so this might not be accepted by the project. But, maybe you could update opencv's cmake to attempt to detect if glslang is available. If it is available, it can enable vulkan support and use glslang to compile the spir-v source files. If it is not available, then the vulkan support could be disabled.

vpisarev · 2018-10-04T12:17:55Z

@wzw-intel, we discussed it a little bit.

More general questions first.

what will be OpenCV DNN runtime dependencies once this patch is integrated and OpenCV DNN is built with Vulkan support?
what will be OpenCV DNN compile-time dependencies if user just wants to build OpenCV DNN as-is without touching .comp files?
what will be OpenCV DNN compile-time dependencies if user wants to patch .comp files and build updated OpenCV DNN? More specifically, if you change .comp files on your machine, how do you generate those .cpp files with machine code?

For example, here is the ideal situation with OpenCL acceleration:

OpenCL runtime is the optional dependency. If OpenCL runtime is not detected, we do not use OpenCL kernels but simply switch to CPU code paths instead.
Neither OpenCL SDK nor even OpenCL runtime are needed to build OpenCV with OpenCL support.
If users patch .cl files inside OpenCV, they are automatically processed and the updated kernel source code is automatically included into OpenCV.

wzw-intel · 2018-10-05T13:07:07Z

@vpisarev
Thanks for your explanation for the "ideal situation with OpenCL acceleration".
I will cook a new commit like what the OpenCL acceleration do.

alalek · 2018-10-05T13:24:52Z

cook a new commit

I believe, at first, we should evaluate required tools/SDKs to build/run Vulkan code.
Could you provide "how to" guide from the beginning what we should download / install to build & run this DNN patch? Perhaps for Windows-based systems; or Linux if it doesn't require Kernel or Mesa complete rebuild. What H/W is required?

OpenCV source code tree has:

bundled OpenCL headers: https://github.com/opencv/opencv/tree/3.4.3/3rdparty/include/opencl
auto-generated OpenCL wrappers: headers and the loader

wzw-intel · 2018-10-07T13:39:41Z

Hi @alalek, I pushed an update to fix can't find vulkan package issue.

You can follow below steps to test the vulkan DNN patch (take Ubuntu 16.04 as an example):

Install Vulkan SDK (driver loader)
Download Vulkan SDK from https://vulkan.lunarg.com/sdk/home#sdk/downloadConfirm/1.1.85.0/linux/vulkansdk-linux-x86_64-1.1.85.0.tar.gz
mkdir ~/vulkan_sdk
tar zxf vulkansdk-linux-x86_64-1.1.85.0.tar.gz -C ~/vulkan_sdk
Install mesa vulkan driver
sudo apt-get install mesa-vulkan-drivers
Build OpenCV with DNN vulkan backend and run the test
source ~/vulkan_sdk/1.1.85.0/setup-env.sh
cd $OPENCV_DIR
mkdir build && cd build
cmake -DWITH_VULKAN=ON ..
make -j8
./bin/opencv_test_dnn --gtest_filter= *VkCom*

As you can see, this version has dependency on Vulkan runtime (libvulkan.so) for both compile-time and run-time, and can't fallback to CPU path if can't detect Vulkan. I will fix these issues in next update by using dynamic load method.

In order to build dnn with Vulkan support, need installing Vulkan SDK and setting environment variable "VULKAN_SDK" and add "-DWITH_VULKAN=ON" to cmake command. You can download Vulkan SDK from: https://vulkan.lunarg.com/sdk/home#linux For how to install, see https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html https://vulkan.lunarg.com/doc/sdk/latest/windows/getting_started.html https://vulkan.lunarg.com/doc/sdk/latest/mac/getting_started.html respectively for linux, windows and mac. To run the vulkan backend, also need installing mesa driver. On Ubuntu, use this command 'sudo apt-get install mesa-vulkan-drivers' To test, use command '$BUILD_DIR/bin/opencv_test_dnn --gtest_filter=*VkCom*' Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

No compile-time dependency on Vulkan library. If Vulkan runtime is unavailable, fallback to CPU path. Use environment "OPENCL_VULKAN_RUNTIME" to specify path to your own vulkan runtime library. Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

…ders The SPIR-V shaders are in format of text-based 32-bit hexadecimal numbers, and inserted into .cpp files as unsigned int32 array.

wzw-intel · 2018-10-09T05:38:51Z

Hi @vpisarev @alalek
I pushed the commits that fixed the dependency issue and shader source code patching issue.
I also update the PR message to add "HOWTO".
Please check it out. Any comments are appriciated.

alalek · 2018-10-10T08:47:46Z

cmake/OpenCVDetectVulkan.cmake

@@ -0,0 +1,17 @@
+set(VULKAN_INCLUDE_DIRS "${OpenCV_SOURCE_DIR}/modules/dnn/src/vkcom/" CACHE PATH "Vulkan include directory")


Please move 3rdparty headers (original from Khronos) into "/3rdparty/include/vkcom".

alalek · 2018-10-10T08:49:27Z

cmake/OpenCVDetectVulkan.cmake

+
+set(HAVE_VULKAN 1)
+add_definitions(-DVK_NO_PROTOTYPES)
+include_directories(${VULKAN_INCLUDE_DIRS})


Perhaps we should not add this globally.

In dnn/CMakeLists.txt:

if(HAVE_VULKAN) add_definitions(-DVK_NO_PROTOTYPES) include_directories(${VULKAN_INCLUDE_DIRS}) endif()

alalek · 2018-10-10T08:51:18Z

modules/dnn/src/vkcom/shader/concat_spv.cpp

+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at https://opencv.org/license.html.
+//
+// Copyright (C) 2017, Intel Corporation, all rights reserved.


Please fix copyright year.

… fixes Vulkan header files are copied from https://github.com/KhronosGroup/Vulkan-Docs/tree/master/include/vulkan to 3rdparty/include Fix the Copyright declaration issue. Refine OpenCVDetectVulkan.cmake

wzw-intel · 2018-10-11T00:12:48Z

Hi @alalek, I fixed the issues you mentioned in your review comments, plase check.

vpisarev · 2018-10-13T08:13:13Z

@wzw-intel, big thanks! Excellent job! @dkurt, could you please review it as well?

dkurt · 2018-10-15T05:13:05Z

modules/dnn/test/test_vkcom_layer.cpp

+using namespace cv::dnn;
+using namespace testing;
+
+static void test(Mat& input, Net& net, Backend backendId, Target targetId, bool skipCheck = false)


I think it's better to add Vulkan target tests to existing ones. Please see
test_backends.cpp (whole networks)
test_layers.cpp
test_halide_layers.cpp (there are not only Halide tests actually)

Also fixed some test failures. - Don't use bool variable as uniform for shader - Fix dispathed group number beyond max issue - Bypass "group > 1" convolution. This should be support in future.

wzw-intel · 2018-10-18T07:27:36Z

Hi @dkurt,
I updated the patch set for adding vulkan test case into existing ones and fixed some test issues. Could you please help to review? Thanks

dkurt · 2018-10-18T08:06:52Z

@wzw-intel, is there a way to choose Vulkan device because now it fallbacks to default CPU implementation for me.

wzw-intel · 2018-10-18T08:20:06Z

@wzw-intel, is there a way to choose Vulkan device because now it fallbacks to default CPU implementation for me.

@dkurt
Vulkan backend will choose the first available device. Have you installed Vulkan runtime (loader library) and Vulkan driver? You can follow the steps described in the PR comments to setup your environments.

dkurt · 2018-10-18T08:53:39Z

@wzw-intel, Thanks! It looks like accuracy tests use Vulkan backend properly. I just wanted to run performance tests and it fallbacks because of the following check. Can you take at look on perf_net.cpp?

bool isAvailable()
{
    return getContext() != NULL;
}

alalek · 2018-10-18T08:10:54Z

modules/dnn/src/vkcom/vulkan/function_list.inl

@@ -0,0 +1,301 @@
+#ifndef VK_ENTRY


Please add .hpp extension (to make happy code editors): function_list.inl.hpp

alalek · 2018-10-18T10:16:20Z

Test build dumps this:

[ RUN      ] DNNTestNetwork.AlexNet/0, where GetParam() = OCV/CPU
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateInstance(&createInfo, NULL, &kInstance)
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
Memory consumption:
    Weights(parameters): 233 Mb
    Blobs: 9 Mb
Calculation complexity: 2.33441 GFlops
[ PERFSTAT ]    (samples=1   mean=23.39   median=23.39   min=23.39   stddev=0.00 (0.0%))
[       OK ] DNNTestNetwork.AlexNet/0 (320 ms)
[ RUN      ] DNNTestNetwork.AlexNet/1, where GetParam() = OCV/OCL
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateInstance(&createInfo, NULL, &kInstance)
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
[ INFO:0] Successfully initialized OpenCL cache directory: /build/.cache/opencv_opencl_cache_x64/
[ INFO:0] Preparing OpenCL cache configuration for context: Intel_R__Corporation--Intel_R__Gen9_HD_Graphics_NEO--18_31_11200
Memory consumption:
    Weights(parameters): 233 Mb
    Blobs: 9 Mb
Calculation complexity: 2.33441 GFlops
[ PERFSTAT ]    (samples=1   mean=19.14   median=19.14   min=19.14   stddev=0.00 (0.0%))
[       OK ] DNNTestNetwork.AlexNet/1 (327 ms)
[ RUN      ] DNNTestNetwork.AlexNet/2, where GetParam() = OCV/OCL_FP16
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateInstance(&createInfo, NULL, &kInstance)
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
Memory consumption:
    Weights(parameters): 233 Mb
    Blobs: 9 Mb
Calculation complexity: 2.33441 GFlops
[ PERFSTAT ]    (samples=1   mean=13.63   median=13.63   min=13.63   stddev=0.00 (0.0%))
[       OK ] DNNTestNetwork.AlexNet/2 (392 ms)
[ RUN      ] DNNTestNetwork.AlexNet/3, where GetParam() = VKCOM/VULKAN
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateInstance(&createInfo, NULL, &kInstance)
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
[ INFO:0] VKCOM: init()
[ INFO:0] VKCOM: vkCreateDevice(kPhysicalDevice, &deviceCreateInfo, NULL, &device)
[ INFO:0] VKCOM: vkCreateCommandPool(device, &commandPoolCreateInfo, NULL, &cmd_pool)
Memory consumption:
    Weights(parameters): 233 Mb
    Blobs: 9 Mb
Calculation complexity: 2.33441 GFlops
[ PERFSTAT ]    (samples=1   mean=30.58   median=30.58   min=30.58   stddev=0.00 (0.0%))
[       OK ] DNNTestNetwork.AlexNet/3 (372 ms)

Notes:

Vulkan initialization for non-Vulkan processing should be avoided. No need to allocate Vulkan resources if they are not used.
As @dkurt mentioned, there are no extra calls in "vulkan" code path in perf_dnn tests.
Current error handling is questionable (just CV_LOG_WARNING()). For example, test_dnn fails due memory allocation error (it is limited to 2-3Gb on CI). Some fallback mechanism should be designed. In OpenCL if some kernel requires missing extension then it is able to fallback to CPU code path (see OCL_RUN()). Final result would be correct. If there is no switching fallback then C++ exception should be raised with proper cleanup of used memory.
It it is a complex task, then please replace CV_LOG_WARNING -> CV_LOG_ERROR + CV_Error() call.

modules/dnn/src/vkcom/src/vkcom.cpp

dkurt · 2018-10-24T04:57:41Z

Unfortunately I'm not able to run all the tests because of segfault:

[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/198, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], false, OCV/OCL)
[       OK ] Layer_Test_Halide/Deconvolution.Accuracy/198 (17 ms)
[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/199, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], false, OCV/OCL_FP16)
[       OK ] Layer_Test_Halide/Deconvolution.Accuracy/199 (28 ms)
[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/200, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], false, VKCOM/VULKAN)
[       OK ] Layer_Test_Halide/Deconvolution.Accuracy/200 (17 ms)
[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/201, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], true, OCV/OCL)
[       OK ] Layer_Test_Halide/Deconvolution.Accuracy/201 (28 ms)
[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/202, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], true, OCV/OCL_FP16)
[       OK ] Layer_Test_Halide/Deconvolution.Accuracy/202 (17 ms)
[ RUN      ] Layer_Test_Halide/Deconvolution.Accuracy/203, where GetParam() = ([6, 9, 3], 5x6, 3x1, 0x1, 2x2, [1, 1, 0, 0], true, VKCOM/VULKAN)
[ WARN:0] Vulkan check failed
Segmentation fault (core dumped)

It happens not for a specific test case but after some time.

dkurt

👍

wzw-intel · 2018-10-30T02:27:53Z

Hi @alalek @dkurt , thanks for merging the patches. I just opened another PR to track the issues you mentioned in above comments, including error handleing, vulkan backend initialization and "Segmentation fault in test" and more refinements will be added in that PR.

* dnn: Add a Vulkan based backend This commit adds a new backend "DNN_BACKEND_VKCOM" and a new target "DNN_TARGET_VULKAN". VKCOM means vulkan based computation library. This backend uses Vulkan API and SPIR-V shaders to do the inference computation for layers. The layer types that implemented in DNN_BACKEND_VKCOM include: Conv, Concat, ReLU, LRN, PriorBox, Softmax, MaxPooling, AvePooling, Permute This is just a beginning work for Vulkan in OpenCV DNN, more layer types will be supported and performance tuning is on the way. Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com> * dnn/vulkan: Add FindVulkan.cmake to detect Vulkan SDK In order to build dnn with Vulkan support, need installing Vulkan SDK and setting environment variable "VULKAN_SDK" and add "-DWITH_VULKAN=ON" to cmake command. You can download Vulkan SDK from: https://vulkan.lunarg.com/sdk/home#linux For how to install, see https://vulkan.lunarg.com/doc/sdk/latest/linux/getting_started.html https://vulkan.lunarg.com/doc/sdk/latest/windows/getting_started.html https://vulkan.lunarg.com/doc/sdk/latest/mac/getting_started.html respectively for linux, windows and mac. To run the vulkan backend, also need installing mesa driver. On Ubuntu, use this command 'sudo apt-get install mesa-vulkan-drivers' To test, use command '$BUILD_DIR/bin/opencv_test_dnn --gtest_filter=*VkCom*' Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com> * dnn/Vulkan: dynamically load Vulkan runtime No compile-time dependency on Vulkan library. If Vulkan runtime is unavailable, fallback to CPU path. Use environment "OPENCL_VULKAN_RUNTIME" to specify path to your own vulkan runtime library. Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com> * dnn/Vulkan: Add a python script to compile GLSL shaders to SPIR-V shaders The SPIR-V shaders are in format of text-based 32-bit hexadecimal numbers, and inserted into .cpp files as unsigned int32 array. * dnn/Vulkan: Put Vulkan headers into 3rdparty directory and some other fixes Vulkan header files are copied from https://github.com/KhronosGroup/Vulkan-Docs/tree/master/include/vulkan to 3rdparty/include Fix the Copyright declaration issue. Refine OpenCVDetectVulkan.cmake * dnn/Vulkan: Add vulkan backend tests into existing ones. Also fixed some test failures. - Don't use bool variable as uniform for shader - Fix dispathed group number beyond max issue - Bypass "group > 1" convolution. This should be support in future. * dnn/Vulkan: Fix multiple initialization in one thread.

wzw-intel force-pushed the vkcom branch 6 times, most recently from bc05c26 to 804c6d5 Compare September 30, 2018 13:04

wzw-intel force-pushed the vkcom branch from 804c6d5 to 0af7f33 Compare October 2, 2018 03:00

wzw-intel force-pushed the vkcom branch 3 times, most recently from 985abee to efb9e77 Compare October 9, 2018 01:08

wzw-intel added 2 commits October 9, 2018 09:11

dnn/Vulkan: dynamically load Vulkan runtime

efb9e77

No compile-time dependency on Vulkan library. If Vulkan runtime is unavailable, fallback to CPU path. Use environment "OPENCL_VULKAN_RUNTIME" to specify path to your own vulkan runtime library. Signed-off-by: Wu Zhiwen <zhiwen.wu@intel.com>

wzw-intel force-pushed the vkcom branch from d0d9b05 to fe51913 Compare October 9, 2018 05:12

dnn/Vulkan: Add a python script to compile GLSL shaders to SPIR-V sha…

fe51913

…ders The SPIR-V shaders are in format of text-based 32-bit hexadecimal numbers, and inserted into .cpp files as unsigned int32 array.

alalek reviewed Oct 10, 2018

View reviewed changes

wzw-intel force-pushed the vkcom branch from c04dee2 to 1098155 Compare October 10, 2018 15:08

dnn/Vulkan: Put Vulkan headers into 3rdparty directory and some other…

1098155

… fixes Vulkan header files are copied from https://github.com/KhronosGroup/Vulkan-Docs/tree/master/include/vulkan to 3rdparty/include Fix the Copyright declaration issue. Refine OpenCVDetectVulkan.cmake

vpisarev requested a review from dkurt October 13, 2018 08:14

dkurt reviewed Oct 15, 2018

View reviewed changes

dnn/Vulkan: Add vulkan backend tests into existing ones.

99ca3d1

Also fixed some test failures. - Don't use bool variable as uniform for shader - Fix dispathed group number beyond max issue - Bypass "group > 1" convolution. This should be support in future.

alalek reviewed Oct 18, 2018

View reviewed changes

dnn/Vulkan: Fix multiple initialization in one thread.

2548367

alalek reviewed Oct 18, 2018

View reviewed changes

modules/dnn/src/vkcom/src/vkcom.cpp Show resolved Hide resolved

dkurt approved these changes Oct 29, 2018

View reviewed changes

dkurt self-assigned this Oct 29, 2018

alalek merged commit 6e3ea8b into opencv:master Oct 29, 2018

wzw-intel deleted the vkcom branch October 30, 2018 00:29

dkurt mentioned this pull request Jun 19, 2020

DNN Support matrix #17591

Open

NALLEIN mentioned this pull request Aug 10, 2020

[GSoC] GPU-enabled OpenCV.js, add webgpu-dawn support for dnn module. #18066

Open

		@@ -0,0 +1,17 @@
		set(VULKAN_INCLUDE_DIRS "${OpenCV_SOURCE_DIR}/modules/dnn/src/vkcom/" CACHE PATH "Vulkan include directory")

Uh oh!

dnn: Add a Vulkan based backend #12703

dnn: Add a Vulkan based backend #12703

Uh oh!

Conversation

wzw-intel commented Sep 30, 2018 • edited by alalek Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Build OpenCV DNN with Vulkan support

Enable DNN Vulkan runtime support

Run test

How to patch GLSL shaders

Uh oh!

vpisarev commented Oct 1, 2018

Uh oh!

wzw-intel commented Oct 2, 2018

Uh oh!

jljusten commented Oct 2, 2018

Uh oh!

vpisarev commented Oct 4, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wzw-intel commented Oct 5, 2018

Uh oh!

alalek commented Oct 5, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wzw-intel commented Oct 7, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

wzw-intel commented Oct 9, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alalek Oct 10, 2018

Choose a reason for hiding this comment

Uh oh!

alalek Oct 10, 2018

Choose a reason for hiding this comment

Uh oh!

alalek Oct 10, 2018

Choose a reason for hiding this comment

Uh oh!

wzw-intel commented Oct 11, 2018

Uh oh!

vpisarev commented Oct 13, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dkurt Oct 15, 2018

Choose a reason for hiding this comment

Uh oh!

wzw-intel commented Oct 18, 2018

Uh oh!

dkurt commented Oct 18, 2018

Uh oh!

wzw-intel commented Oct 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dkurt commented Oct 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alalek Oct 18, 2018

Choose a reason for hiding this comment

Uh oh!

alalek commented Oct 18, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

dkurt commented Oct 24, 2018

Uh oh!

dkurt left a comment

Choose a reason for hiding this comment

Uh oh!

wzw-intel commented Oct 30, 2018

Uh oh!

Uh oh!

wzw-intel commented Sep 30, 2018 •

edited by alalek

Loading

vpisarev commented Oct 4, 2018 •

edited

Loading

alalek commented Oct 5, 2018 •

edited

Loading

wzw-intel commented Oct 7, 2018 •

edited

Loading

wzw-intel commented Oct 9, 2018 •

edited

Loading

vpisarev commented Oct 13, 2018 •

edited

Loading

wzw-intel commented Oct 18, 2018 •

edited

Loading

dkurt commented Oct 18, 2018 •

edited

Loading

alalek commented Oct 18, 2018 •

edited

Loading