OpenVML RVV

OpenVML 简介

OpenVML 是一套面向向量数据的数学函数库，提供浮点向量初等函数、二元运算、归约运算以及整数向量运算等能力。它的接口以批量数组为输入输出，例如 vsSqrt、vsAdd、vsSum、vdExp、vs8Dot 等，可用于一次处理大量连续元素。

在机器人、感知算法、信号处理、机器学习前后处理等场景中，常见的数据流会包含大量逐元素运算，例如绝对值、平方根、指数、对数、三角函数、向量加减乘以及整型量化数据处理。相比在业务代码中手写标量循环，使用向量数学库可以把这些基础算子交给经过平台优化的实现，从而获得更稳定的吞吐表现。

RVV 加速

K3 平台搭载的 SpacemiT X100 处理器支持 RISC-V Vector Extension（RVV 1.0）。openvml-spacemit 安装在 /opt/openvml-spacemit，面向 K3 平台提供 OpenVML 算子实现。

项目	openvml-spacemit
安装方式	sudo apt install openvml-spacemit
头文件路径	/opt/openvml-spacemit/include
库文件路径	/opt/openvml-spacemit/lib
主要头文件	openvml.h
主要库文件	libopenvml.so
典型用途	向量数学函数、浮点逐元素运算、整数向量运算与归约运算

使用示例

软硬件环境

• SpacemiT RISCV64 X100 CPU（2.4 GHz）
• Bianbu 4.0.1 操作系统
• 内存：32 GB

安装必要依赖

sudo apt update
sudo apt install openvml-spacemit

测试代码

目录结构：

openvml_check/
├── CMakeLists.txt
└── openvml_check.cpp

CMakeLists.txt：

cmake_minimum_required(VERSION 3.16)
project(openvml_simple_check LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

set(OPENVML_PREFIX "/opt/openvml-spacemit" CACHE PATH "OpenVML install prefix")

find_path(OPENVML_INCLUDE_DIR
  NAMES openvml.h
  PATHS "${OPENVML_PREFIX}/include"
  NO_DEFAULT_PATH)

find_library(OPENVML_LIBRARY
  NAMES openvml
  PATHS "${OPENVML_PREFIX}/lib"
  NO_DEFAULT_PATH)

if(NOT OPENVML_INCLUDE_DIR)
  message(FATAL_ERROR "openvml.h not found under ${OPENVML_PREFIX}/include")
endif()

if(NOT OPENVML_LIBRARY)
  message(FATAL_ERROR "libopenvml not found under ${OPENVML_PREFIX}/lib")
endif()

get_filename_component(OPENVML_LIBRARY_DIR "${OPENVML_LIBRARY}" DIRECTORY)

add_executable(openvml_check openvml_check.cpp)
target_include_directories(openvml_check PRIVATE "${OPENVML_INCLUDE_DIR}")
target_compile_options(openvml_check PRIVATE -O3 -DNDEBUG)
target_link_libraries(openvml_check PRIVATE "${OPENVML_LIBRARY}")

if(OPENVML_LIBRARY_DIR)
  set_target_properties(openvml_check PROPERTIES
    BUILD_RPATH "${OPENVML_LIBRARY_DIR}"
    INSTALL_RPATH "${OPENVML_LIBRARY_DIR}")
endif()

message(STATUS "OPENVML_PREFIX: ${OPENVML_PREFIX}")
message(STATUS "OPENVML_INCLUDE_DIR: ${OPENVML_INCLUDE_DIR}")
message(STATUS "OPENVML_LIBRARY: ${OPENVML_LIBRARY}")

openvml_check.cpp：

#include <openvml.h>

#include <algorithm>
#include <cmath>
#include <iomanip>
#include <iostream>
#include <numeric>
#include <vector>

double max_abs_error(const std::vector<float>& x, const std::vector<float>& y) {
  double err = 0.0;
  for (size_t i = 0; i < x.size(); ++i) {
    err = std::max(err, std::abs(static_cast<double>(x[i]) - static_cast<double>(y[i])));
  }
  return err;
}

int main() {
  const VML_INT n = 8;
  std::vector<float> x{0.25f, 1.0f, 2.25f, 4.0f, 6.25f, 9.0f, 12.25f, 16.0f};
  std::vector<float> y{1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 7.0f, 8.0f};
  std::vector<float> sqrt_out(n);
  std::vector<float> add_out(n);
  std::vector<float> sqrt_ref(n);
  std::vector<float> add_ref(n);
  float sum_out = 0.0f;

  vsSqrt(n, x.data(), sqrt_out.data());
  vsAdd(n, x.data(), y.data(), add_out.data());
  vsSum(n, x.data(), &sum_out);

  for (VML_INT i = 0; i < n; ++i) {
    sqrt_ref[static_cast<size_t>(i)] = std::sqrt(x[static_cast<size_t>(i)]);
    add_ref[static_cast<size_t>(i)] = x[static_cast<size_t>(i)] + y[static_cast<size_t>(i)];
  }
  const float sum_ref = std::accumulate(x.begin(), x.end(), 0.0f);

  const double sqrt_err = max_abs_error(sqrt_out, sqrt_ref);
  const double add_err = max_abs_error(add_out, add_ref);
  const double sum_err = std::abs(static_cast<double>(sum_out) - static_cast<double>(sum_ref));

  std::cout << std::fixed << std::setprecision(6);
  std::cout << "OpenVML config: " << openvml_get_config() << '\n';
  std::cout << "vsSqrt max_abs_error=" << sqrt_err << '\n';
  std::cout << "vsAdd max_abs_error=" << add_err << '\n';
  std::cout << "vsSum abs_error=" << sum_err << '\n';

  if (sqrt_err > 1e-5 || add_err > 1e-5 || sum_err > 1e-5) {
    std::cerr << "OpenVML result: FAIL - result mismatch\n";
    return 1;
  }

  std::cout << "OpenVML result: PASS - linked OpenVML and computed correctly.\n";
  return 0;
}

编译与运行

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
./build/openvml_check

终端输出：

也可以显式指定安装路径：

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DOPENVML_PREFIX=/opt/openvml-spacemit
cmake --build build
LD_LIBRARY_PATH=/opt/openvml-spacemit/lib:$LD_LIBRARY_PATH ./build/openvml_check

更多性能测试数据

• 使用 taskset -c 绑定 CPU，脚本同时设置 OMP_NUM_THREADS。
• vector_size 为 1048576，iterations 为 50，warmup 为 10。
• avg_ms 表示单次算子调用的平均耗时，mevals_per_second 表示每秒百万次元素计算吞吐，checksum 用于确认输出参与计算且结果稳定。

测试结果简要说明

本次测试覆盖 81 个 OpenVML 算子，包括 float32、float64 和 int8 / int16 / int32 数据类型。

单核绑定测试结果

module	function	api	expression	dtype	input_size	output_size	eval_count	total_ms	avg_ms	mevals_per_second	checksum
unary	abs	vsAbs	y = abs(x)	float32	1048576x1	1048576x1	1048576.0000	28.0353	0.5607	1870.0962	1572878.0765
unary	recip	vsRecip	y = 1 / x	float32	1048576x1	1048576x1	1048576.0000	98.7829	1.9757	530.7475	634956.4420
unary	sqrt	vsSqrt	y = sqrt(x)	float32	1048576x1	1048576x1	1048576.0000	99.4084	1.9882	527.4083	2796397.7431
unary	sqr	vsSqr	y = x * x	float32	1048576x1	1048576x1	1048576.0000	26.6677	0.5334	1966.0032	3145759.1388
unary	pow2o3	vsPow2o3	y = x^(2/3)	float32	1048576x1	1048576x1	1048576.0000	280.8658	5.6173	186.6685	3995106.2436
unary	pow3o2	vsPow3o2	y = x^(3/2)	float32	1048576x1	1048576x1	1048576.0000	276.8267	5.5365	189.3922	26845443.5468
unary	exp	vsExp	y = exp(x)	float32	1048576x1	1048576x1	1048576.0000	82.0062	1.6401	639.3276	1149691.5115
unary	expm1	vsExpm1	y = exp(x) - 1	float32	1048576x1	1048576x1	1048576.0000	613.6415	12.2728	85.4388	101107.1236
unary	log10	vsLog10	y = log10(x)	float32	1048576x1	1048576x1	1048576.0000	589.4520	11.7890	88.9450	807500.8625
unary	ln	vsLn	y = ln(x)	float32	1048576x1	1048576x1	1048576.0000	138.1510	2.7630	379.5035	1859339.4487
unary	log1p	vsLog1p	y = log(1 + x)	float32	1048576x1	1048576x1	1048576.0000	719.0534	14.3811	72.9136	2108088.9563
unary	tanh	vsTanh	y = tanh(x)	float32	1048576x1	1048576x1	1048576.0000	1031.7560	20.6351	50.8151	-0.0011
unary	round	vsRound	y = round(x)	float32	1048576x1	1048576x1	1048576.0000	26.1761	0.5235	2002.9256	0.0001
unary	ceil	vsCeil	y = ceil(x)	float32	1048576x1	1048576x1	1048576.0000	26.2473	0.5249	1997.4920	524292.1944
unary	floor	vsFloor	y = floor(x)	float32	1048576x1	1048576x1	1048576.0000	28.1338	0.5627	1863.5518	-524290.1941
unary	sin	vsSin	y = sin(x)	float32	1048576x1	1048576x1	1048576.0000	113.4690	2.2694	462.0541	-0.0021
unary	cos	vsCos	y = cos(x)	float32	1048576x1	1048576x1	1048576.0000	113.9144	2.2783	460.2474	953007.0083
unary	tan	vsTan	y = tan(x)	float32	1048576x1	1048576x1	1048576.0000	1184.4234	23.6885	44.2653	-0.0026
unary	asin	vsAsin	y = asin(x)	float32	1048576x1	1048576x1	1048576.0000	303.8496	6.0770	172.5485	0.0006
unary	acos	vsAcos	y = acos(x)	float32	1048576x1	1048576x1	1048576.0000	396.6906	7.9338	132.1655	1647112.5339
unary	atan	vsAtan	y = atan(x)	float32	1048576x1	1048576x1	1048576.0000	289.2232	5.7845	181.2745	-0.0013
reduction	sum	vsSum	y = sum(x)	float32	1048576x1	1x1	1048576.0000	25.7794	0.5156	2033.7495	0.0344
binary	add	vsAdd	y = a + b	float32	1048576x1	1048576x1	1048576.0000	62.5870	1.2517	837.6948	1310730.4626
binary	sub	vsSub	y = a - b	float32	1048576x1	1048576x1	1048576.0000	61.2043	1.2241	856.6199	-1310730.4890
binary	mul	vsMul	y = a * b	float32	1048576x1	1048576x1	1048576.0000	61.1937	1.2239	856.7680	786439.7730
binary	pow	vsPow	y = a ^ b	float32	1048576x1	1048576x1	1048576.0000	1217.9616	24.3592	43.0463	136978240.0024
binary	powx	vsPowx	y = a ^ b	float32	1048576x1	1048576x1	1048576.0000	276.7253	5.5345	189.4615	2796397.7404
binary	atan2	vsAtan2	y = atan2(a, b)	float32	1048576x1	1048576x1	1048576.0000	245.4615	4.9092	213.5927	-122453.9317
reduction	dot	vsDot	y = dot(a, b)	float32	1048576x1	1x1	1048576.0000	56.9216	1.1384	921.0701	786433.1875
binary	offset	vsOffset	y = offset(a, b)	float32	1048576x1	1048576x1	1048576.0000	27.9806	0.5596	1873.7580	524292.1943
binary	scale	vsScale	y = scale(a, b)	float32	1048576x1	1048576x1	1048576.0000	28.5132	0.5703	1838.7553	-0.0032
pair_output	sincos	vsSinCos	sin(x), cos(x)	float32	1048576x1	1048576x2	1048576.0000	624.6710	12.4934	83.9303	953007.5056
unary	abs	vdAbs	y = abs(x)	float64	1048576x1	1048576x1	1048576.0000	88.5267	1.7705	592.2368	1572878.0829
unary	recip	vdRecip	y = 1 / x	float64	1048576x1	1048576x1	1048576.0000	365.2998	7.3060	143.5227	634956.4308
unary	sqrt	vdSqrt	y = sqrt(x)	float64	1048576x1	1048576x1	1048576.0000	1075.4431	21.5089	48.7509	2796397.7786
unary	sqr	vdSqr	y = x * x	float64	1048576x1	1048576x1	1048576.0000	66.4973	1.3299	788.4355	3145759.1658
unary	pow2o3	vdPow2o3	y = x^(2/3)	float64	1048576x1	1048576x1	1048576.0000	1987.4959	39.7499	26.3793	3995106.1786
unary	pow3o2	vdPow3o2	y = x^(3/2)	float64	1048576x1	1048576x1	1048576.0000	1986.1636	39.7233	26.3970	26845444.5869
unary	exp	vdExp	y = exp(x)	float64	1048576x1	1048576x1	1048576.0000	594.1584	11.8832	88.2404	1149691.5151
unary	expm1	vdExpm1	y = exp(x) - 1	float64	1048576x1	1048576x1	1048576.0000	1420.2793	28.4056	36.9144	101107.1265
unary	log10	vdLog10	y = log10(x)	float64	1048576x1	1048576x1	1048576.0000	1146.8622	22.9372	45.7150	807500.8742
unary	ln	vdLn	y = ln(x)	float64	1048576x1	1048576x1	1048576.0000	606.5845	12.1317	86.4328	1859339.4754
unary	log1p	vdLog1p	y = log(1 + x)	float64	1048576x1	1048576x1	1048576.0000	1842.7722	36.8554	28.4510	2108088.9783
unary	tanh	vdTanh	y = tanh(x)	float64	1048576x1	1048576x1	1048576.0000	2713.2891	54.2658	19.3230	0.0000
unary	round	vdRound	y = round(x)	float64	1048576x1	1048576x1	1048576.0000	1229.2977	24.5860	42.6494	0.0001
unary	ceil	vdCeil	y = ceil(x)	float64	1048576x1	1048576x1	1048576.0000	1228.6794	24.5736	42.6709	524290.1944
unary	floor	vdFloor	y = floor(x)	float64	1048576x1	1048576x1	1048576.0000	1229.5959	24.5919	42.6391	-524291.1942
unary	sin	vdSin	y = sin(x)	float64	1048576x1	1048576x1	1048576.0000	1655.0676	33.1014	31.6777	0.0000
unary	cos	vdCos	y = cos(x)	float64	1048576x1	1048576x1	1048576.0000	1612.7833	32.2557	32.5083	953007.5061
unary	tan	vdTan	y = tan(x)	float64	1048576x1	1048576x1	1048576.0000	2048.6185	40.9724	25.5923	0.0000
unary	asin	vdAsin	y = asin(x)	float64	1048576x1	1048576x1	1048576.0000	900.6585	18.0132	58.2116	0.0001
unary	acos	vdAcos	y = acos(x)	float64	1048576x1	1048576x1	1048576.0000	843.2233	16.8645	62.1766	1647112.5059
unary	atan	vdAtan	y = atan(x)	float64	1048576x1	1048576x1	1048576.0000	1833.6776	36.6736	28.5922	0.0000
reduction	sum	vdSum	y = sum(x)	float64	1048576x1	1x1	1048576.0000	67.1682	1.3434	780.5595	0.0000
binary	add	vdAdd	y = a + b	float64	1048576x1	1048576x1	1048576.0000	119.3994	2.3880	439.1045	1310730.4859
binary	sub	vdSub	y = a - b	float64	1048576x1	1048576x1	1048576.0000	125.3201	2.5064	418.3592	-1310730.4857
binary	pow	vdPow	y = a ^ b	float64	1048576x1	1048576x1	1048576.0000	2072.6019	41.4520	25.2961	136978237.8088
binary	powx	vdPowx	y = a ^ b	float64	1048576x1	1048576x1	1048576.0000	1988.8939	39.7779	26.3608	2796397.7786
binary	atan2	vdAtan2	y = atan2(a, b)	float64	1048576x1	1048576x1	1048576.0000	2898.7575	57.9752	18.0866	-122453.8617
pair_output	sincos	vdSinCos	sin(x), cos(x)	float64	1048576x1	1048576x2	1048576.0000	3265.3805	65.3076	16.0560	953007.5061
integer	add	vs8Add	y = int8(a) + int8(b)	int8	1048576x1	1048576x1	1048576.0000	11.8777	0.2376	4414.0540	-10.9028
integer	sub	vs8Sub	y = int8(a) - int8(b)	int8	1048576x1	1048576x1	1048576.0000	4.3932	0.0879	11934.2022	-13.0971
integer	mul	vs8Mul	y = int8(a) * int8(b)	int8	1048576x1	1048576x1	1048576.0000	11.5218	0.2304	4550.3839	144.0038
integer	shift	vs8Shift	y = int8(a) << b	int8	1048576x1	1048576x1	1048576.0000	2.9781	0.0596	17604.5274	-24.0000
integer	scale	vs8Scale	y = int8(a) * scalar	int8	1048576x1	1048576x1	1048576.0000	2.5836	0.0517	20292.7597	-35.9999
integer	offset	vs8Offset	y = int8(a) + scalar	int8	1048576x1	1048576x1	1048576.0000	2.7626	0.0553	18978.2702	2097156.7772
integer	dot	vs8Dot	y = dot(int8(a), int8(b))	int8	1048576x1	1x1	1048576.0000	6.2316	0.1246	8413.3135	144.0000
integer	add	vs16Add	y = int16(a) + int16(b)	int16	1048576x1	1048576x1	1048576.0000	26.1217	0.5224	2007.0979	-110.9992
integer	sub	vs16Sub	y = int16(a) - int16(b)	int16	1048576x1	1048576x1	1048576.0000	23.0746	0.4615	2272.1397	-107.0000
integer	mul	vs16Mul	y = int16(a) * int16(b)	int16	1048576x1	1048576x1	1048576.0000	26.3294	0.5266	1991.2651	-31661.6291
integer	shift	vs16Shift	y = int16(a) << b	int16	1048576x1	1048576x1	1048576.0000	6.0567	0.1211	8656.2753	-217.9993
integer	scale	vs16Scale	y = int16(a) * scalar	int16	1048576x1	1048576x1	1048576.0000	5.2266	0.1045	10031.1215	-326.9989
integer	offset	vs16Offset	y = int16(a) + scalar	int16	1048576x1	1048576x1	1048576.0000	5.9744	0.1195	8775.5611	2097059.7776
integer	dot	vs16Dot	y = dot(int16(a), int16(b))	int16	1048576x1	1x1	1048576.0000	21.4046	0.4281	2449.4130	-31661.0000
integer	add	vs32Add	y = int32(a) + int32(b)	int32	1048576x1	1048576x1	1048576.0000	61.1772	1.2235	856.9985	1797469.3806
integer	sub	vs32Sub	y = int32(a) - int32(b)	int32	1048576x1	1048576x1	1048576.0000	67.6426	1.3529	775.0853	-1797681.3804
integer	mul	vs32Mul	y = int32(a) * int32(b)	int32	1048576x1	1048576x1	1048576.0000	62.0221	1.2404	845.3244	-250.9997
integer	shift	vs32Shift	y = int32(a) << b	int32	1048576x1	1048576x1	1048576.0000	28.4934	0.5699	1840.0325	-211.9998
integer	scale	vs32Scale	y = int32(a) * scalar	int32	1048576x1	1048576x1	1048576.0000	27.7604	0.5552	1888.6208	-317.9997
integer	offset	vs32Offset	y = int32(a) + scalar	int32	1048576x1	1048576x1	1048576.0000	28.3400	0.5668	1849.9929	2097062.7773
integer	dot	vs32Dot	y = dot(int32(a), int32(b))	int32	1048576x1	1x1	1048576.0000	53.6776	1.0736	976.7358	-251.0000