RISC-V: support RVV 0.7 in mainline RVV intrinsics

[mshabunin]

Recent T-Head GCC have ability to produce RVV 0.7.1 compatible machine code using mainline intrinsics syntax. It makes possible to have single intrinsics implementation for both 0.7 and 1.0 versions (intrin_rvv.hpp, maybe scalable too).

Build and test with RVV 1.0 (non-scalable)

Using one of the following toolchains:

• Collab toolchain: https://github.com/riscv-collab/riscv-gnu-toolchain (rvv_next branch)
• T-Head toolchain: https://github.com/T-head-Semi/xuantie-gnu-toolchain

Using one of the following qemu:

• Prebuilt T-Head qemu: https://occ.t-head.cn/community/download?id=4108987827910414336
• Collab toolchain: https://github.com/riscv-collab/riscv-gnu-toolchain (rvv_next branch)

cmake \
    -DCMAKE_TOOLCHAIN_FILE=/work/opencv/platforms/linux/riscv64-gcc.toolchain.cmake \
    -DCPU_RVV_FLAGS_ON="-march=rv64gcv" \
    -DRISCV_RVV_SCALABLE=OFF \
    ../opencv
# T-Head qemu
/opt/xuantie-qemu-20221104/bin/qemu-riscv64 \
	-L ${sysroot} \
	-cpu c908v \
	./bin/opencv_test_core
# Collab qemu
/opt/bin/qemu-riscv64 \
    -L ${sysroot} \
    -cpu rv64,v=true,vext_spec=v1.0 \
    ./bin/opencv_test_core

Build and test with RVV 0.7.1 (non-scalable)

Using one of the following toolchains:

• Build from sources (:exclamation:can not build OpenCV in this mode at this moment:exclamation:): https://github.com/T-head-Semi/xuantie-gnu-toolchain
• Prebuilt toolchain (registration needed): https://occ.t-head.cn/community/download?id=4090445921563774976

Using qemu:

• Prebuilt T-Head qemu: https://occ.t-head.cn/community/download?id=4108987827910414336

cmake \
    -DCMAKE_TOOLCHAIN_FILE=/work/opencv/platforms/linux/riscv64-gcc.toolchain.cmake \
    -DCPU_RVV_FLAGS_ON="-march=rv64gcv0p7" \
    -DRISCV_RVV_SCALABLE=OFF \
    ../opencv
# T-Head qemu
/opt/xuantie-qemu-20221104/bin/qemu-riscv64 \
	-L ${sysroot} \
        -cpu c906fdv \
	./bin/opencv_test_core

Notes:

• v_trunc implementation have modified because there is no corresponding _rtz_ intrinsics in 0.7
• v_absdiff have been reimplemented to avoid using 64-bit types for 32-bit input
• v_check_ also updated to use 32-bit internally
• CV_SIMD128_64F guard was missing in some places
• additional modification in intrin_rvv07.hpp is made to improve support on real HW (C906 / D1 SoC).

force_builders=Custom
build_image:Custom=riscv-gcc-rvv07
test_modules:Custom=core,imgproc,dnn
buildworker:Custom=linux-4
test_timeout:Custom=1200
build_contrib:Custom=OFF

[alalek]

@mshabunin I added riscv-gcc-rvv07 build image. Please verify build/test logs with your results.

[mshabunin]

Which qemu do you use? Why is it -cpu rv64,v=true,vext_spec=v1.0 when the build is for v0p7?

I only tested core and imgproc in Xuantie qemu. I will run the whole testsuite and compare results with yours.

[alalek]

Please merge build + test steps instructions for different scenarios.
Currently it is unclear from PR's description: there are 2 alternative builds and 2 alternative tests.

[mshabunin]

Updated the description, also filed an issue to the T-Head toolchain: T-head-Semi/xuantie-gnu-toolchain#5 - apparently toolchain built from sources can not compile OpenCV with RVV 0.7 support, but prebuilt one can.

Also added my Docker-based testing environment: https://github.com/mshabunin/riscv-check

[alalek]

CV_RISCV_OLD

This name is unclear.
Please be more specific when and why we use it.

[mshabunin]

Renamed to CV_RVV_THEAD_0_7 and added a comment

[Lzingnow]

Hello, I want to know if this RVV1.0-compatible opencv can be compiled using the riscv64-unknown-elf toolchain. This is because I downloaded the riscv-gnu-toolchain branch of rvv-next, but after make linux only generated the gcc compiler and no g++.It is not known what caused the problem.

[mshabunin]

@Lzingnow , I didn't try to build with elf toolchain. Your issue is that you should build linux target. (maybe also add --enable-linux when configuring). Also rvv_next branch is considered obsolete, instead one should use master branch of riscv-gnu-toolchain repository and update gcc submodule to 13.x manually. See the following snippet for example: https://github.com/mshabunin/riscv-check/blob/6d07d13f98cba29bf59a39b7b9de9b1ac39bd4d5/Dockerfile-collab#L34-L48

[Lzingnow]

Hello, I have successfully installed the rvv-next branch of riscv-gnu-toolchain by following the above procedure. Then, I downloaded the source code for opencv4.8. Build with the following commands:

mkdir build 

cmake .. -DCMAKE_TOOLCHAIN_FILE=../platforms/linux/riscv64-gcc.toolchain.cmake -DCPU_RVV_FLAGS_ON="-march=rv64gcv" -DRISCV_RVV_SCALABLE=OFF  ../

make -j$(nproc)

make install -j$(nproc)

At this point, it indicates that the installation was successful.

Next, I want to write c++ code that calls the opencv library to generate an executable file that can run on a CPU that supports rvv1.0. Specifically, the PULP project's ara project. Here I test the resize function, c++ code is as follows:

#include <opencv2/opencv.hpp>
#include <iostream>

int main() {

    cv::Mat originalImage = cv::Mat(100, 100, CV_8UC3);
    cv::randu(originalImage, cv::Scalar(0, 0, 0), cv::Scalar(255, 255, 255));

    cv::Mat resizedImage;
    cv::resize(originalImage, resizedImage, cv::Size(50, 50), 0, 0, cv::INTER_LINEAR);

    std::cout << "Resized Image:" << std::endl;
    for (int row = 0; row < resizedImage.rows; ++row) {
        for (int col = 0; col < resizedImage.cols; ++col) {
            cv::Vec3b pixel = resizedImage.at<cv::Vec3b>(row, col);
            std::cout << "(" << int(pixel[0]) << "," << int(pixel[1]) << "," << int(pixel[2]) << ") ";
        }
        std::cout << std::endl;
    }

    return 0;
}

The specific compilation options are as follows:

riscv64-unknown-linux-gnu-g++ -march=rv64gcv resize.cpp -o resize.o -I/home/opencv/build/install/include/opencv4 -L/home/ara/opencv/build/install/lib -lopencv_core -lopencv_imgcodecs -lopencv_imgproc

But I did not find the vector instruction after decompilation. May I ask why? The files resize.o and resize.dump are attached.

riscv64-unknown-linux-gnu-objdump -d  resize.o>resize.dump

resize.zip

[mshabunin]

Your resize.o does not use vector instructions, try to objdump libopencv_core.a (or .so).

[Lzingnow]

@mshabunin Thank you very much. I got the executable file containing the vector instructions by setting up static compilation.

But I also have a question about whether the opencv library for rvv1.0 is dependent on the linux operating system and cannot run in bare-mental environment.

This is because I once tried to compile opencv libraries using the elf toolchain, but it failed. But linx-gnu works. On the other hand, I ran the resize executable using linux-gnu statically compiled in a bare-mental environment, and it did not work as expected and reported errors.

[mshabunin]

We do not build OpenCV for bare metal environments, it is not supported officially. In general OpenCV uses C/C++ libraries, some POSIX functions and a threading library (pthreads). If your toolchain can provide these interfaces, then perhaps it might work.

[Lzingnow]

Thank you for your reply.

[Lzingnow]

@mshabunin

Hello, I would like to know if you have encountered such problems when testing with the rvv-next branch. After compiling the opencv library using the rvv-next branch, I want to test it using qemu. The steps are as follows:

1.Compile qemu: make build-qemu -j8
2.Test opencv:

/opt/bin/qemu-riscv64 \
    -L ${sysroot} \
    -cpu rv64,v=true,vext_spec=v1.0 \
    ./bin/opencv_test_core

But there was an error:

qemu-riscv64: can't apply global rv64-riscv-cpu.v=true: Property 'rv64-riscv-cpu.v' not found
make: *** [Makefile:2: all] Error 1

Then I found another qemu setup for rvv1.0 here: https://lists.riscv.org/g/tech-vector-ext/topic/85564316?p=,,,20,0,0,0::recentpostdate/sticky,,,20,2,0,85564316

3. Change the cpu option to -cpu rv64,x-v=true,vlen=128,elen=64,vext_spec=v1.0

But there is still one mistake： qemu-riscv64: Unsupported vector spec version 'v1.0'

In addition, I also tried qemu under the master branch, but still got an error: qemu-riscv64: can't apply global rv64-riscv-cpu.x-v=true: Property 'rv64-riscv-cpu.x-v' not found

I don't know how to solve this problem.

[Lzingnow]

I think I solved the problem, but instead of using qemu under the rvv-next branch, I downloaded it separately.Using qemu-7.1.0:

wget https://download.qemu.org/qemu-7.1.0-rc1.tar.xz
tar xvJf qemu-7.1.0-rc1.tar.xz
cd qemu-7.1.0-rc1
./configure --target-list=riscv64-softmmu,riscv64-linux-user --prefix=/home/dingli/qemu-7.1.0-rc1-riscv64
make -j8
make install -j8