1.介绍
在上篇中，介绍了ARM的Neon，本篇主要介绍Neon intrinsics的函数用法，也就是assembly之前的用法。NEON指令是从Armv7架构开始引入的SIMD指令，其共有16个128位寄存器。发展到最新的Arm64架构，其寄存器数量增加到32个，但是其长度仍然为最大128位，因此操作上并没有发生显著的变化。对于这样的寄存器，因为可以同时存储并处理多组数据，称之为向量寄存器。Intrinsics是使用C语言的方式对NEON寄存器进行操作，因为相比于传统的使用纯汇编语言，具有可读性强，开发速度快等优势。如果需要在代码中调用NEON Intrinsics函数，需要加入头文件"arm_neon.h"。关于neon的所有函数，可以参考官网的：ARM NEON intrinsics reference 这里将网上的Neon intrinsics的函数做个总结：
1.1 指令的分类
-
正常指令：生成大小相同且类型通常与操作数向量相同的结果向量
-
长指令：对双字节向量操作数执行运算，生成四字向量的结果，所生成的元素一般是操作数元素宽度的两倍
-
宽指令：一个双字向量操作数和一个四字向量操作数执行运算，生成四字向量结果，所生成的元素和第一个操作数的元素是第二个操作数元素宽度的两倍
-
窄指令：四字向量操作数执行运算，并生成双字向量结果，所生成的元素一般是操作数元素宽度的一半
-
饱和指令：当超过数据类型指定的范围则自动限制在该范围内
示例1:
- int16x8_t vqaddq_s16 (int16x8_t, int16x8_t)
- int16x4_t vqadd_s16 (int16x4_t, int16x4_t)
- 第一个字母'v'指明是vector向量指令，也就是NEON指令；
- 第二个字母'q'指明是饱和指令，即后续的加法结果会自动饱和；
- 第三个字段'add'指明是加法指令；
- 第四个字段'q'指明操作寄存器宽度，为'q'时操作QWORD, 为128位；未指明时操作寄存器为DWORD，为64位；
- 第五个字段's16'指明操作的基本单元为有符号16位整数，其最大表示范围为-32768 ~ 32767；
- 形参和返回值类型约定与C语言一致。
其它可能用到的助记符包括:
- l 长指令，数据扩展
- w 宽指令，数据对齐
- n 窄指令, 数据压缩
关于所有函数的分类，请参考博客：https://blog.csdn.net/hemmingway/article/details/44828303
1.2 数据类型
NEON Intrinsics内置的整数数据类型主要包括以下几种:
- (u)int8x8_t;
- (u)int8x16_t;
- (u)int16x4_t;
- (u)int16x8_t;
- (u)int32x2_t;
- (u)int32x4_t;
- (u)int64x1_t;
其中，第一个数字代表的是数据类型宽度为8/16/32/64位，第二个数字代表的是一个寄存器中该类型数据的数量。如int16x8_t代表16位有符号数，寄存器中共有8个数据。
2.Syntax
2.1 Arithmetic
- add: vaddq_f32 or vaddq_f64 ( sum = v1 + v2 )
float32x4_t v1 = { 1.0, 2.0, 3.0, 4.0 }, v2 = { 1.0, 1.0, 1.0, 1.0 };
float32x4_t sum = vaddq_f32(v1, v2);
// => sum = { 2.0, 3.0, 4.0, 5.0 }
- multiply: vmulq_f32 or vmulq_f64 ( sum = v1 + v2 )
float32x4_t v1 = { 1.0, 2.0, 3.0, 4.0 }, v2 = { 1.0, 1.0, 1.0, 1.0 };
float32x4_t prod = vmulq_f32(v1, v2);
// => prod = { 1.0, 2.0, 3.0, 4.0 }
- multiply and accumulate: vmlaq_f32 ( sum = v3 + v1 * v2 )
float32x4_t v1 = { 1.0, 2.0, 3.0, 4.0 }, v2 = { 2.0, 2.0, 2.0, 2.0 }, v3 = { 3.0, 3.0, 3.0, 3.0 };
float32x4_t acc = vmlaq_f32(v3, v1, v2); // acc = v3 + v1 * v2
// => acc = { 5.0, 7.0, 9.0, 11.0 }
- multiply by a scalar: vmulq_n_f32 or vmulq_n_f64 ( prod = V1 * a)
float32x4_t v = { 1.0, 2.0, 3.0, 4.0 };
float32_t s = 3.0;
float32x4_t prod = vmulq_n_f32(v, s);
// => prod = { 3.0, 6.0, 9.0, 12.0 }
- multiply by a scalar and accumulate: vmlaq_n_f32 or vmlaq_n_f64 ( acc = v1*v2 + s)
float32x4_t v1 = { 1.0, 2.0, 3.0, 4.0 }, v2 = { 1.0, 1.0, 1.0, 1.0 };
float32_t s = 3.0;
float32x4_t acc = vmlaq_n_f32(v1, v2, s);
// => acc = { 4.0, 5.0, 6.0, 7.0 }
- invert (needed for division): vrecpeq_f32 or vrecpeq_f64 ( reciprocal = 1 / v)
float32x4_t v = { 1.0, 2.0, 3.0, 4.0 };
float32x4_t reciprocal = vrecpeq_f32(v);
// => reciprocal = { 0.998046875, 0.499023438, 0.333007813, 0.249511719 }
float32x4_t v = { 1.0, 2.0, 3.0, 4.0 };
float32x4_t reciprocal = vrecpeq_f32(v);
float32x4_t inverse = vmulq_f32(vrecpsq_f32(v, reciprocal), reciprocal);
// => inverse = { 0.999996185, 0.499998093, 0.333333015, 0.249999046 }
2.2 Load
- load vector: vld1q_f32 or vld1q_f64
float values[5] = { 1.0, 2.0, 3.0, 4.0, 5.0 };
float32x4_t v = vld1q_f32(values);
// => v = { 1.0, 2.0, 3.0, 4.0 }
- load same value for all lanes: vld1q_dup_f32 or vld1q_dup_f64
float val = 3.0;
float32x4_t v = vld1q_dup_f32(&val);
// => v = { 3.0, 3.0, 3.0, 3.0 }
- set all lanes to a hardcoded value: vmovq_n_f16 or vmovq_n_f32 or vmovq_n_f64
float32x4_t v = vmovq_n_f32(1.5);
// => v = { 1.5, 1.5, 1.5, 1.5 }
2.3 Store
- store vector: vst1q_f32 or vst1q_f64
float32x4_t v = { 1.0, 2.0, 3.0, 4.0 };
float values[5] = new float[5];
vst1q_f32(values, v);
// => values = { 1.0, 2.0, 3.0, 4.0, #undef }
- store lane of array of vectors: vst4q_lane_f16 or vst4q_lane_f32 or vst4q_lane_f64 (change to vst1... / vst2... / vst3...for other array lengths)
float32x4_t v0 = { 1.0, 2.0, 3.0, 4.0 }, v1 = { 5.0, 6.0, 7.0, 8.0 }, v2 = { 9.0, 10.0, 11.0, 12.0 }, v3 = { 13.0, 14.0, 15.0, 16.0 };
float32x4x4_t u = { v0, v1, v2, v3 };
float buff[4];
vst4q_lane_f32(buff, u, 0);
// => buff = { 1.0, 5.0, 9.0, 13.0 }
2.4 Arrays
float32x4_t v0 = { 1.0, 2.0, 3.0, 4.0 }, v1 = { 5.0, 6.0, 7.0, 8.0 }, v2 = { 9.0, 10.0, 11.0, 12.0 }, v3 = { 13.0, 14.0, 15.0, 16.0 };
float32x4x4_t ary = { v0, v1, v2, v3 };
float32x4_t v = ary.val[2];
// => v = { 9.0, 10.0, 11.0, 12.0 }
2.5 Max and min
- max of two vectors, element by element:
float32x4_t v0 = { 5.0, 2.0, 3.0, 4.0 }, v1 = { 1.0, 6.0, 7.0, 8.0 };
float32x4_t v2 = vmaxq_f32(v0, v1);
// => v1 = { 5.0, 6.0, 7.0, 8.0 }
- max of vector elements, using folding maximum:
float32x4_t v0 = { 1.0, 2.0, 3.0, 4.0 };
float32x2_t maxOfHalfs = vpmax_f32(vget_low_f32(v0), vget_high_f32(v0));
float32x2_t maxOfMaxOfHalfs = vpmax_f32(maxOfHalfs, maxOfHalfs);
float maxValue = vget_lane_f32(maxOfMaxOfHalfs, 0);
// => maxValue = 4.0
- min of two vectors, element by element:
float32x4_t v0 = { 5.0, 2.0, 3.0, 4.0 }, v1 = { 1.0, 6.0, 7.0, 8.0 };
float32x4_t v2 = vminq_f32(v0, v1);
// => v1 = { 1.0, 2.0, 3.0, 4.0 }
- min of vector elements, using folding minimum:
float32x4_t v0 = { 1.0, 2.0, 3.0, 4.0 };
float32x2_t minOfHalfs = vpmin_f32(vget_low_f32(v0), vget_high_f32(v0));
float32x2_t minOfMinOfHalfs = vpmin_f32(minOfHalfs, minOfHalfs);
float minValue = vget_lane_f32(minOfMinOfHalfs, 0);
// => minValue = 1.0
2.5 conditionals
- ternary operator: use vector comparison (for example vcltq_f32 for less than comparison)
float32x4_t v1 = { 1.0, 0.0, 1.0, 0.0 }, v2 = { 0.0, 1.0, 1.0, 0.0 };
float32x4_t mask = vcltq_f32(v1, v2); // v1 < v2
float32x4_t ones = vmovq_n_f32(1.0), twos = vmovq_n_f32(2.0);
float32x4_t v3 = vbslq_f32(mask, ones, twos); // will select first if mask 0, second if mask 1
// => v3 = { 2.0, 1.0, 2.0, 2.0 }
3. Sample for openCV, c pointer, Neon
前提：图片大小 640 x 480，动作：每三行的各列相加等于当前列。例如：x（i，j） = x（i, j） +x(i - 1, j) + x(i-2, j).
- openCV的做法：其中，cv::Mat gray, src .src是来自每一帧图片(640x480 deep = 8bits)
GETTIME(&lTimeStart);
for (int col = 0; col < gray.cols; col++)
{
gray.at<uchar>(0, col) = src.at<uchar>(0, col);
}
for (int col = 0; col < gray.cols; col++)
{
gray.at<uchar>(1, col) = gray.at<uchar>(0, col) + src.at<uchar>(1, col);
}
for (int col = 0; col < gray.cols; col++)
{
gray.at<uchar>(2, col) = gray.at<uchar>(1, col) + src.at<uchar>(2, col);
}
for (int row = 3; row < gray.rows; row++)
{
for (int col = 0; col < gray.cols; col++)
{
gray.at<uchar>(row, col) = gray.at<uchar>(row - 1, col) + src.at<uchar>(row, col) - src.at<uchar>(row - 3, col);
}
}
GETTIME(&lTimeEnd);
printf("time %ldus\n",lTimeEnd - lTimeStart);
在arm-A57平台，openCV消耗的时间均值：time = 19175us
GETTIME(&lTimeStart);
unsigned char *ptr = src.ptr(0);
unsigned char *grayPtr = gray.ptr(0);
for(int col = 0; col < gray.cols; col++)
{
grayPtr[col] = ptr[col];
}
unsigned char *ptr1 = src.ptr(1);
unsigned char *grayPtr1 = gray.ptr(1);
for(int col =0; col < gray.cols; col++)
{
grayPtr1[col] = ptr[col] + ptr1[col];//34us
}
unsigned char *ptr2 = NULL;
unsigned char *grayPtr2 = NULL;
for(int row = 2; row < gray.rows; row++)
{
ptr = src.ptr(row - 2);
ptr1 = src.ptr(row -1);
ptr2 = src.ptr(row);
grayPtr2 = gray.ptr(row);
for(int col = 0; col <gray.cols; col+=16){
grayPtr2[col] = ptr[col] + ptr1[col] + ptr2[col];//11252us
}
}
GETTIME(&lTimeEnd);
printf("time %ldus\n",lTimeEnd - lTimeStart);
在arm-A57平台，C-pointer消耗的时间均值：time = 11252us
GETTIME(&lTimeStart);
unsigned char *ptr = src.ptr(0);
unsigned char *grayPtr = gray.ptr(0);
for(int col = 0; col < gray.cols; col++)
{
grayPtr[col] = ptr[col];
}
unsigned char *ptr1 = src.ptr(1);
unsigned char *grayPtr1 = gray.ptr(1);
for(int col =0; col < gray.cols; col++)
{
grayPtr1[col] = ptr[col] + ptr1[col];//34us
}
unsigned char *ptr2 = NULL;
unsigned char *grayPtr2 = NULL;
for(int row = 2; row < gray.rows; row++)
{
ptr = src.ptr(row - 2);
ptr1 = src.ptr(row -1);
ptr2 = src.ptr(row);
grayPtr2 = gray.ptr(row);
for(int col = 0; col <gray.cols; col+=16){
uint8x16_t in1,in2,in3,out;
in1 = vld1q_u8(ptr+col);
in2 = vld1q_u8(ptr1+col);
in3 = vld1q_u8(ptr2+col);
out = vaddq_u8(in1,in2);
out = vaddq_u8(in3,out);
vst1q_u8(grayPtr2+col,out);
}
}
GETTIME(&lTimeEnd);
printf("time %ldus\n",lTimeEnd - lTimeStart);
在arm-A57平台，Neon intrinscis消耗的时间均值：time = 1907us
综上，可以看到，neon相对opencv方式的性能提升快10倍。（注意，这里的加法都有溢出的情况，由于本算法特殊，所以没有做溢出处理）
