性能优化篇（5）：NEON 优化案例——图像转换之 RGBA 到 RGB（aarch64版）
Author: stormQ
Created: Sunday, 8. December 2019 1 12:47AM
Last Modified: Wednesday, 28. October 2020 09:10PM
本文描述了如何利用 NEON 向量指令加速 RGBA 图像到 RGB 图像的转换过程，并从高速缓存的角度定量分析了 NEON 加速原理。
如何将一个大小为672x376的RGBA图像转换成相同大小的RGB图像呢？
首先实现一个最简单的转换函数simple_rgba2rgb()，完整实现如下：
#define SRC_CHANNELS 4
#define DST_CHANNELS 3
void simple_rgba2rgb(const uint8_t *rgba_img, uint8_t *rgb_img,
int height, int width)
{
const int total_pixels = height * width;
for (int i = 0; i < total_pixels; i++)
{
const int src_idx = i * SRC_CHANNELS;
const int dst_idx = i * DST_CHANNELS;
*(rgb_img + dst_idx) = *(rgba_img + src_idx);
*(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
*(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
}
}
simple_rgba2rgb()函数将一个二维的图像看作是一维的，意味着只需要一个循环就可以遍历所有的像素。这种用法的好处在于，遵循空间局部性原理以高效利用cache从而改善程序性能，而不必关心二维图像在内存中的存储顺序——是行主序还是列主序。如果用两个循环的话会比较繁琐，移植性也不好。另外，由于源图像和目的图像的大小相同但所占的字节数不同，所以采用每次遍历一个像素的方式，这样便于处理。
首先，实现一个基本的利用NEON加速图像转换的函数rgba2rgb_with_neon()，完整实现如下：
void rgba2rgb_with_neon(const uint8_t *rgba_img, uint8_t *rgb_img,
int height, int width)
{
const int total_pixels = height * width;
const int stride_pixels = 16;
for (int i = 0; i < total_pixels; i += stride_pixels)
{
const uint8_t *src = rgba_img + i * SRC_CHANNELS;
uint8_t *dst = rgb_img + i * DST_CHANNELS;
uint8x16x4_t a = vld4q_u8(src);
uint8x16x3_t b;
b.val[0] = a.val[0];
b.val[1] = a.val[1];
b.val[2] = a.val[2];
vst3q_u8(dst, b);
}
}
上述函数没有考虑图像像素个数非整除时的情形，这个放到后面进行。该函数的实现用到了两个NEON Intrinsics，分别是vld4q_u8和vst3q_u8，两者都定义在arm_neon.h文件中。
vld4q_u8的函数原型为uint8x16x4_t vld4q_u8 (const uint8_t * __a)，作用为以步长为 4 交叉地加载数据到四个连续的 128-bit 的向量寄存器。具体地：将内存地址__a、__a+4、…、__a+60处的内容分别赋值给向量寄存器Vn的lane[0]、lane[1]、…、lane[15]，将内存地址__a+1、__a+5、…、__a+61处的内容分别赋值给向量寄存器Vn+1的lane[0]、lane[1]、…、lane[15]，将内存地址__a+2、__a+6、…、__a+62处的内容分别赋值给向量寄存器Vn+2的lane[0]、lane[1]、…、lane[15]，将内存地址__a+3、__a+7、…、__a+63处的内容分别赋值给向量寄存器Vn+3的lane[0]、lane[1]、…、lane[15]。也就是说，vld4q_u8在上述函数中的作用为：将连续 16 个像素的 R 通道： R0、R1、… 、R15 加载到向量寄存器Vn，将连续 16 个像素的 G 通道： G0、G1、… 、G15 加载到向量寄存器Vn+1，将连续 16 个像素的 B 通道： B0、B1、… 、B15 加载到向量寄存器Vn+2，将连续 16 个像素的 Alpha 通道： A0、A1、… 、A15 加载到向量寄存器Vn+3。
vst3q_u8的函数原型为void vst3q_u8 (uint8_t * __a, uint8x16x3_t val)，作用为以步长为 3 交叉地存储数据到内存中。
由于每一次迭代可以同时处理 16 个连续的像素。所以，变量stride_pixels的值为 16。
由于aarch64指令集中没有ARM 指令集中的VSWP指令。因此，这里引入一个类型为uint8x16x3_t的临时向量b用于存储R、G和B通道。 g++ 带优化选项-Og编译时，语句b.val[0] = a.val[2]; b.val[1] = a.val[1]; b.val[2] = a.val[0];对应三条MOV指令。
接着用语句vst3q_u8(dst, b);将转换后的结果写入内存。这样，一次图像转换操作就完成了。
现在考虑图像像素个数非整除时的情形，完整实现如下：
void rgba2rgb_with_neon(const uint8_t *rgba_img, uint8_t *rgb_img,
int height, int width)
{
const int total_pixels = height * width;
const int stride_pixels = 16;
const int left_pixels = total_pixels % stride_pixels;
const int multiply_pixels = total_pixels - left_pixels;
for (int i = 0; i < multiply_pixels; i += stride_pixels)
{
const uint8_t *src = rgba_img + i * SRC_CHANNELS;
uint8_t *dst = rgb_img + i * DST_CHANNELS;
uint8x16x4_t a = vld4q_u8(src);
uint8x16x3_t b;
b.val[0] = a.val[0];
b.val[1] = a.val[1];
b.val[2] = a.val[2];
vst3q_u8(dst, b);
}
// handling non-multiply array lengths
for (int i = multiply_pixels; i < total_pixels; i++)
{
const int src_idx = i * SRC_CHANNELS;
const int dst_idx = i * DST_CHANNELS;
*(rgb_img + dst_idx) = *(rgba_img + src_idx);
*(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
*(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
}
}
为了验证rgba2rgb_with_neon()函数的正确性，这里引入两个函数：init()和check()。init()函数用于初始化图像，check()函数用于检查转换后的图像是否正确。完整实现如下：
#define RED 255
#define GREEN 125
#define BLUE 80
#define ALPHA 100
#define CHECK(img, height, width) \
if (check(img, height, width)) \
{ \
std::cout << "Correct:img.\n"; \
} \
else \
{ \
std::cout << "Error:img.\n"; \
}
void init(uint8_t *img, int height, int width)
{
const int total_bytes = height * width * SRC_CHANNELS;
for (int i = 0; i < total_bytes; )
{
*(img + i++) = RED;
*(img + i++) = GREEN;
*(img + i++) = BLUE;
*(img + i++) = ALPHA;
}
}
bool check(const uint8_t *img, int height, int width)
{
const int total_bytes = height * width * DST_CHANNELS;
for (int i = 0; i < total_bytes; i += DST_CHANNELS)
{
if (*(img + i) != RED || *(img + i + 1) != GREEN ||
*(img + i + 2) != BLUE)
{
return false;
}
}
return true;
}
完整的程序为main.cpp：
#include "arm_neon.h"
#include <iostream>
#define HEIGHT 376
#define WIDTH 672
#define SRC_CHANNELS 4
#define DST_CHANNELS 3
#define RED 255
#define GREEN 125
#define BLUE 80
#define ALPHA 100
#define CHECK(img, height, width) \
if (check(img, height, width)) \
{ \
std::cout << "Correct:img.\n"; \
} \
else \
{ \
std::cout << "Error:img.\n"; \
}
void init(uint8_t *img, int height, int width)
{
const int total_bytes = height * width * SRC_CHANNELS;
for (int i = 0; i < total_bytes; )
{
*(img + i++) = RED;
*(img + i++) = GREEN;
*(img + i++) = BLUE;
*(img + i++) = ALPHA;
}
}
bool check(const uint8_t *img, int height, int width)
{
const int total_bytes = height * width * DST_CHANNELS;
for (int i = 0; i < total_bytes; i += DST_CHANNELS)
{
if (*(img + i) != RED || *(img + i + 1) != GREEN ||
*(img + i + 2) != BLUE)
{
return false;
}
}
return true;
}
void simple_rgba2rgb(const uint8_t *rgba_img, uint8_t *rgb_img,
int height, int width)
{
const int total_pixels = height * width;
for (int i = 0; i < total_pixels; i++)
{
const int src_idx = i * SRC_CHANNELS;
const int dst_idx = i * DST_CHANNELS;
*(rgb_img + dst_idx) = *(rgba_img + src_idx);
*(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
*(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
}
}
void rgba2rgb_with_neon(const uint8_t *rgba_img, uint8_t *rgb_img,
int height, int width)
{
const int total_pixels = height * width;
const int stride_pixels = 16;
const int left_pixels = total_pixels % stride_pixels;
const int multiply_pixels = total_pixels - left_pixels;
for (int i = 0; i < multiply_pixels; i += stride_pixels)
{
const uint8_t *src = rgba_img + i * SRC_CHANNELS;
uint8_t *dst = rgb_img + i * DST_CHANNELS;
uint8x16x4_t a = vld4q_u8(src);
uint8x16x3_t b;
b.val[0] = a.val[0];
b.val[1] = a.val[1];
b.val[2] = a.val[2];
vst3q_u8(dst, b);
}
// handling non-multiply array lengths
for (int i = multiply_pixels; i < total_pixels; i++)
{
const int src_idx = i * SRC_CHANNELS;
const int dst_idx = i * DST_CHANNELS;
*(rgb_img + dst_idx) = *(rgba_img + src_idx);
*(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
*(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
}
}
uint8_t *rgba_img = nullptr;
uint8_t *rgb_img1 = nullptr;
uint8_t *rgb_img2 = nullptr;
int main()
{
rgba_img = new uint8_t[HEIGHT * WIDTH * SRC_CHANNELS];
rgb_img1 = new uint8_t[HEIGHT * WIDTH * DST_CHANNELS];
rgb_img2 = new uint8_t[HEIGHT * WIDTH * DST_CHANNELS];
init(rgba_img, HEIGHT, WIDTH);
simple_rgba2rgb(rgba_img, rgb_img1, HEIGHT, WIDTH);
rgba2rgb_with_neon(rgba_img, rgb_img2, HEIGHT, WIDTH);
CHECK(rgb_img1, HEIGHT, WIDTH);
CHECK(rgb_img2, HEIGHT, WIDTH);
delete [] rgb_img2;
delete [] rgb_img1;
delete [] rgba_img;
return 0;
}
编译程序（On Jetson TX2）：
g++ -std=c++11 -g -Og -o main_Og main.cpp
| 启动程序方式 | 第一次执行耗时(us) | 第二次执行耗时(us) | 第三次执行耗时(us) | 第四次执行耗时(us) | 第五次执行耗时(us) |
|---|---|---|---|---|---|
| ./main_Og |
从统计结果中可以看出，rgba2rgb_with_neon()函数的执行速度比simple_rgba2rgb()函数快 2 到 3 倍左右。
那么 NEON 版的实现为什么能达到这样一个加速呢？ 下面从 cache 的角度进行分析。
统计 cache 性能数据：
--------------------------------------------------------------------------------
I1 cache: 16384 B, 64 B, 4-way associative
D1 cache: 16384 B, 64 B, 4-way associative
LL cache: 262144 B, 64 B, 8-way associative
Command: ./main_Og
Data file: cachegrind.out.26634
Events recorded: Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
Events shown: Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
Event sort order: Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
Thresholds: 0.1 100 100 100 100 100 100 100 100
Include dirs:
User annotated: main.cpp
Auto-annotation: on
--------------------------------------------------------------------------------
Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
--------------------------------------------------------------------------------
18,330,632 1,581 1,388 2,833,111 71,198 64,190 2,014,469 42,002 41,005 PROGRAM TOTALS
--------------------------------------------------------------------------------
Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw file:function
--------------------------------------------------------------------------------
8,085,518 0 0 1,516,032 23,690 23,690 0 0 0 /home/test2/main.cpp:check(unsigned char const*, int, int)
4,548,101 3 3 758,016 15,793 15,793 758,016 11,845 11,845 /home/test2/main.cpp:simple_rgba2rgb(unsigned char const*, unsigned char*, int, int)
3,790,086 2 2 0 0 0 1,010,688 15,792 15,792 /home/test2/main.cpp:init(unsigned char*, int, int)
545,429 12 11 148,897 2,038 1,708 50,257 29 16 /build/glibc-BinVK7/glibc-2.23/elf/dl-lookup.c:_dl_lookup_symbol_x
537,013 36 36 191,329 5,065 1,024 93,645 43 24 /build/glibc-BinVK7/glibc-2.23/elf/dl-lookup.c:do_lookup_x
211,319 28 28 48,062 3,377 2,954 20,608 1,748 886 /build/glibc-BinVK7/glibc-2.23/elf/../sysdeps/aarch64/dl-machine.h:_dl_relocate_object
126,348 0 0 0 0 0 0 0 0 /home/test2/main.cpp:rgba2rgb_with_neon(unsigned char const*, unsigned char*, int, int)
108,209 8 6 37,471 330 238 14,588 25 11 /build/glibc-BinVK7/glibc-2.23/elf/dl-lookup.c:check_match
97,629 5 5 12,840 221 96 0 0 0 /build/glibc-BinVK7/glibc-2.23/string/../sysdeps/aarch64/strcmp.S:strcmp
78,960 1 1 63,168 15,793 15,793 47,376 11,845 11,845 /usr/lib/gcc/aarch64-linux-gnu/5/include/arm_neon.h:rgba2rgb_with_neon(unsigned char const*, unsigned char*, int, int)
56,685 3 3 16,081 2,184 1,017 3,209 2 0 /build/glibc-BinVK7/glibc-2.23/elf/do-rel.h:_dl_relocate_object
44,112 8 8 13,412 1,022 986 12 1 0 /build/glibc-BinVK7/glibc-2.23/elf/dl-addr.c:_dl_addr
--------------------------------------------------------------------------------
-- User-annotated source: main.cpp
--------------------------------------------------------------------------------
No information has been collected for main.cpp
--------------------------------------------------------------------------------
-- Auto-annotated source: /usr/lib/gcc/aarch64-linux-gnu/5/include/arm_neon.h
--------------------------------------------------------------------------------
Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
-- line 16574 ----------------------------------------
. . . . . . . . . return ret;
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . __extension__ static __inline uint8x16x4_t __attribute__ ((__always_inline__))
. . . . . . . . . vld4q_u8 (const uint8_t * __a)
. . . . . . . . . {
. . . . . . . . . uint8x16x4_t ret;
. . . . . . . . . __builtin_aarch64_simd_xi __o;
15,792 0 0 63,168 15,793 15,793 0 0 0 __o = __builtin_aarch64_ld4v16qi ((const __builtin_aarch64_simd_qi *) __a);
. . . . . . . . . ret.val[0] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 0);
. . . . . . . . . ret.val[1] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 1);
. . . . . . . . . ret.val[2] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 2);
. . . . . . . . . ret.val[3] = (uint8x16_t) __builtin_aarch64_get_qregxiv16qi (__o, 3);
. . . . . . . . . return ret;
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . __extension__ static __inline uint16x8x4_t __attribute__ ((__always_inline__))
-- line 16590 ----------------------------------------
-- line 23620 ----------------------------------------
. . . . . . . . . __o = __builtin_aarch64_set_qregciv2di (__o, (int64x2_t) val.val[2], 2);
. . . . . . . . . __builtin_aarch64_st3v2di ((__builtin_aarch64_simd_di *) __a, __o);
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . __extension__ static __inline void __attribute__ ((__always_inline__))
. . . . . . . . . vst3q_u8 (uint8_t * __a, uint8x16x3_t val)
. . . . . . . . . {
. . . . . . . . . __builtin_aarch64_simd_ci __o;
15,792 0 0 0 0 0 0 0 0 __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[0], 0);
15,792 1 1 0 0 0 0 0 0 __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[1], 1);
15,792 0 0 0 0 0 0 0 0 __o = __builtin_aarch64_set_qregciv16qi (__o, (int8x16_t) val.val[2], 2);
15,792 0 0 0 0 0 47,376 11,845 11,845 __builtin_aarch64_st3v16qi ((__builtin_aarch64_simd_qi *) __a, __o);
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . __extension__ static __inline void __attribute__ ((__always_inline__))
. . . . . . . . . vst3q_u16 (uint16_t * __a, uint16x8x3_t val)
. . . . . . . . . {
. . . . . . . . . __builtin_aarch64_simd_ci __o;
. . . . . . . . . __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[0], 0);
. . . . . . . . . __o = __builtin_aarch64_set_qregciv8hi (__o, (int16x8_t) val.val[1], 1);
-- line 23639 ----------------------------------------
--------------------------------------------------------------------------------
-- Auto-annotated source: /home/test2/main.cpp
--------------------------------------------------------------------------------
Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
-- line 22 ----------------------------------------
. . . . . . . . . } \
. . . . . . . . . else \
. . . . . . . . . { \
. . . . . . . . . std::cout << "Error:img.\n"; \
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . void init(uint8_t *img, int height, int width)
. . . . . . . . . {
2 0 0 0 0 0 0 0 0 const int total_bytes = height * width * SRC_CHANNELS;
758,020 2 2 0 0 0 0 0 0 for (int i = 0; i < total_bytes; )
. . . . . . . . . {
758,016 0 0 0 0 0 252,672 15,792 15,792 *(img + i++) = RED;
758,016 0 0 0 0 0 252,672 0 0 *(img + i++) = GREEN;
758,016 0 0 0 0 0 252,672 0 0 *(img + i++) = BLUE;
758,016 0 0 0 0 0 252,672 0 0 *(img + i++) = ALPHA;
. . . . . . . . . }
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . bool check(const uint8_t *img, int height, int width)
. . . . . . . . . {
4 0 0 0 0 0 0 0 0 const int total_bytes = height * width * DST_CHANNELS;
2,021,382 0 0 0 0 0 0 0 0 for (int i = 0; i < total_bytes; i += DST_CHANNELS)
. . . . . . . . . {
5,053,440 0 0 1,010,688 15,794 15,794 0 0 0 if (*(img + i) != RED || *(img + i + 1) != GREEN ||
1,010,688 0 0 505,344 7,896 7,896 0 0 0 *(img + i + 2) != BLUE)
. . . . . . . . . {
. . . . . . . . . return false;
. . . . . . . . . }
. . . . . . . . . }
4 0 0 0 0 0 0 0 0 return true;
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . void simple_rgba2rgb(const uint8_t *rgba_img, uint8_t *rgb_img,
. . . . . . . . . int height, int width)
. . . . . . . . . {
1 1 1 0 0 0 0 0 0 const int total_pixels = height * width;
1,010,692 1 1 0 0 0 0 0 0 for (int i = 0; i < total_pixels; i++)
. . . . . . . . . {
252,672 1 1 0 0 0 0 0 0 const int src_idx = i * SRC_CHANNELS;
252,672 0 0 0 0 0 0 0 0 const int dst_idx = i * DST_CHANNELS;
1,010,688 0 0 252,672 15,793 15,793 252,672 3,949 3,949 *(rgb_img + dst_idx) = *(rgba_img + src_idx);
1,010,688 0 0 252,672 0 0 252,672 3,948 3,948 *(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
1,010,688 0 0 252,672 0 0 252,672 3,948 3,948 *(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
. . . . . . . . . }
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . void rgba2rgb_with_neon(const uint8_t *rgba_img, uint8_t *rgb_img,
. . . . . . . . . int height, int width)
. . . . . . . . . {
1 0 0 0 0 0 0 0 0 const int total_pixels = height * width;
. . . . . . . . . const int stride_pixels = 16;
4 0 0 0 0 0 0 0 0 const int left_pixels = total_pixels % stride_pixels;
1 0 0 0 0 0 0 0 0 const int multiply_pixels = total_pixels - left_pixels;
. . . . . . . . .
63,171 0 0 0 0 0 0 0 0 for (int i = 0; i < multiply_pixels; i += stride_pixels)
. . . . . . . . . {
31,584 0 0 0 0 0 0 0 0 const uint8_t *src = rgba_img + i * SRC_CHANNELS;
31,584 0 0 0 0 0 0 0 0 uint8_t *dst = rgb_img + i * DST_CHANNELS;
. . . . . . . . .
. . . . . . . . . uint8x16x4_t a = vld4q_u8(src);
. . . . . . . . . uint8x16x3_t b;
. . . . . . . . . b.val[0] = a.val[0];
. . . . . . . . . b.val[1] = a.val[1];
. . . . . . . . . b.val[2] = a.val[2];
. . . . . . . . . vst3q_u8(dst, b);
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . // handling non-multiply array lengths
3 0 0 0 0 0 0 0 0 for (int i = multiply_pixels; i < total_pixels; i++)
. . . . . . . . . {
. . . . . . . . . const int src_idx = i * SRC_CHANNELS;
. . . . . . . . . const int dst_idx = i * DST_CHANNELS;
. . . . . . . . . *(rgb_img + dst_idx) = *(rgba_img + src_idx);
. . . . . . . . . *(rgb_img + dst_idx + 1) = *(rgba_img + src_idx + 1);
. . . . . . . . . *(rgb_img + dst_idx + 2) = *(rgba_img + src_idx + 2);
. . . . . . . . . }
. . . . . . . . . }
. . . . . . . . .
. . . . . . . . . uint8_t *rgba_img = nullptr;
. . . . . . . . . uint8_t *rgb_img1 = nullptr;
. . . . . . . . . uint8_t *rgb_img2 = nullptr;
. . . . . . . . .
. . . . . . . . . int main()
3 1 1 0 0 0 4 0 0 {
6 1 1 0 0 0 1 1 0 rgba_img = new uint8_t[HEIGHT * WIDTH * SRC_CHANNELS];
5 0 0 0 0 0 1 1 1 rgb_img1 = new uint8_t[HEIGHT * WIDTH * DST_CHANNELS];
3 0 0 0 0 0 1 0 0 rgb_img2 = new uint8_t[HEIGHT * WIDTH * DST_CHANNELS];
. . . . . . . . .
4 0 0 1 0 0 0 0 0 init(rgba_img, HEIGHT, WIDTH);
. . . . . . . . .
5 1 1 2 2 2 0 0 0 simple_rgba2rgb(rgba_img, rgb_img1, HEIGHT, WIDTH);
5 0 0 2 2 2 0 0 0 rgba2rgb_with_neon(rgba_img, rgb_img2, HEIGHT, WIDTH);
. . . . . . . . .
12 1 1 1 1 1 0 0 0 CHECK(rgb_img1, HEIGHT, WIDTH);
14 1 1 1 1 1 0 0 0 CHECK(rgb_img2, HEIGHT, WIDTH);
. . . . . . . . .
5 0 0 1 1 1 0 0 0 delete [] rgb_img2;
5 1 1 1 0 0 0 0 0 delete [] rgb_img1;
5 0 0 1 0 0 0 0 0 delete [] rgba_img;
. . . . . . . . . return 0;
22 3 2 9 0 0 5 0 0 }
-- line 123 ----------------------------------------
--------------------------------------------------------------------------------
The following files chosen for auto-annotation could not be found:
--------------------------------------------------------------------------------
/build/glibc-BinVK7/glibc-2.23/elf/do-rel.h
/build/glibc-BinVK7/glibc-2.23/elf/dl-lookup.c
/build/glibc-BinVK7/glibc-2.23/string/../sysdeps/aarch64/strcmp.S
/build/glibc-BinVK7/glibc-2.23/elf/dl-addr.c
/build/glibc-BinVK7/glibc-2.23/elf/../sysdeps/aarch64/dl-machine.h
--------------------------------------------------------------------------------
Ir I1mr ILmr Dr D1mr DLmr Dw D1mw DLmw
--------------------------------------------------------------------------------
91 1 1 82 78 86 90 94 96 percentage of events annotated
分析统计结果：
| 函数名称 | 内存读操作数量(Dr列) | 一级数据缓存读不命中次数(D1mr列) | 最后一级数据缓存读不命中次数(DLmr列) | 内存写操作数量(Dw列) | 一级数据缓存写不命中次数(D1mw列) | 最后一级数据缓存写不命中次数(DLmw列) |
|---|---|---|---|---|---|---|
| simple_rgba2rgb | 758,016 | 15,793 | 15,793 | 758,016 | 11,845 | 11,845 |
| rgba2rgb_with_neon | 63,168 | 15,793 | 15,793 | 47,376 | 11,845 | 11,845 |
可以看出，rgba2rgb_with_neon函数的内存读操作数量和内存写操作数量分别为simple_rgba2rgb函数的 0.0833 倍、0.0625 倍，其他的都是相同的。因此，从 cache 的角度来看，NEON 大大减少了内存读写操作数量，从而改善程序性能。
最后，能否降低 cache 不命中率从而进一步改善程序性能呢？……
下一篇：性能优化篇（6）：NEON 优化案例——低阶矩阵乘法（aarch64版）