HackMD
2017q1 Homework3 (software-pipelining)
contributed by <illusion030>
When Prefetching Works, When It Doesn’t, and Why
-
ICC 和 GCC 裡只有簡單的 SW prefetching，所以我們必須手動加入 SW prefetching
-
出現兩個問題
- 對於如何最佳的使用 prefetching 的文件很少
- 對於 SW prefetching 和 HW prefetching 之間互動的複雜度並沒有很了解
開發環境
Architecture: x86_64
CPU 作業模式： 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 4
On-line CPU(s) list: 0-3
每核心執行緒數： 2
每通訊端核心數： 2
Socket(s): 1
NUMA 節點： 1
供應商識別號： GenuineIntel
CPU 家族： 6
型號： 69
Model name: Intel(R) Core(TM) i5-4200U CPU @ 1.60GHz
製程： 1
CPU MHz： 1599.920
CPU max MHz: 2600.0000
CPU min MHz: 800.0000
BogoMIPS: 4589.33
虛擬： VT-x
L1d 快取： 32K
L1i 快取： 32K
L2 快取： 256K
L3 快取： 3072K
NUMA node0 CPU(s): 0-3
- 我的電腦的 CPU 為Intel Core i5-4200U
Architecture / Microarchitecture
Microarchitecture Haswell
Platform Shark Bay
Processor core Haswell
Core stepping C0 (SR170)
Manufacturing process 0.022 micron
Data width 64 bit
The number of CPU cores 2
The number of threads 4
Floating Point Unit Integrated
Level 1 cache size 2 x 32 KB 8-way set associative instruction caches
2 x 32 KB 8-way set associative data caches
Level 2 cache size 2 x 256 KB 8-way set associative caches
Level 3 cache size 3 MB 12-way set associative shared cache
Physical memory 16 GB
Multiprocessing Uniprocessor
Features MMX instructions
SSE / Streaming SIMD Extensions
SSE2 / Streaming SIMD Extensions 2
SSE3 / Streaming SIMD Extensions 3
SSSE3 / Supplemental Streaming SIMD Extensions 3
SSE4 / SSE4.1 + SSE4.2 / Streaming SIMD Extensions 4
AES / Advanced Encryption Standard instructions
AVX / Advanced Vector Extensions
AVX2 / Advanced Vector Extensions 2.0
BMI / BMI1 + BMI2 / Bit Manipulation instructions
F16C / 16-bit Floating-Point conversion instructions
FMA3 / 3-operand Fused Multiply-Add instructions
EM64T / Extended Memory 64 technology / Intel 64
NX / XD / Execute disable bit
HT / Hyper-Threading technology
TBT 2.0 / Turbo Boost technology 2.0
VT-x / Virtualization technology
Low power features Enhanced SpeedStep technology
- Microarchitecture 為 Haswell
illusion030@illusion030-X550LD:~$ sudo rdmsr -a 0x1A4
0
0
0
0
- 4個 processor 的 4個 HW prefetcher 都是 enable 的狀態
開發紀錄
- fork prefetcher 並 clone 至電腦執行看看
$ git clone https://github.com/illusion030/prefetcher
$ cd prefetcher
$ make
$ ./main
- 執行結果
0 1 2 3
4 5 6 7
8 9 10 11
12 13 14 15
0 4 8 12
1 5 9 13
2 6 10 14
3 7 11 15
sse prefetch: 67599 us
sse: 131430 us
naive: 253113 us
- 看一下程式碼，在 main.c 中前半段為測試結果是否正確，後半段測出 sse prefetch、sse、naive 個別的時間
SSE
- 參考 Streaming SIMD Extensions 2 Instructions 了解 SSE2 的 function 的運作方式
- 因為不懂 unpack 的運作方式所以將 I0-I4 及 T0-T4 用 function 印出來看看
void print128_num(__m128i var)
{
uint16_t *val = (uint16_t*) &var;
printf(" %i %i %i %i %i %i %i %i \n",
val[0], val[1], val[2], val[3], val[4], val[5],
val[6], val[7]);
}
0 1 2 3
4 5 6 7
8 9 10 11
12 13 14 15
After load
I0= 0 0 1 0 2 0 3 0
I1= 4 0 5 0 6 0 7 0
I2= 8 0 9 0 10 0 11 0
I3= 12 0 13 0 14 0 15 0
After first unpack
I0= 0 0 1 0 2 0 3 0
I1= 4 0 5 0 6 0 7 0
I2= 8 0 9 0 10 0 11 0
I3= 12 0 13 0 14 0 15 0
T0= 0 0 4 0 1 0 5 0
T1= 8 0 12 0 9 0 13 0
T2= 2 0 6 0 3 0 7 0
T3= 10 0 14 0 11 0 15 0
After second unpack
I0= 0 0 4 0 8 0 12 0
I1= 1 0 5 0 9 0 13 0
I2= 2 0 6 0 10 0 14 0
I3= 3 0 7 0 11 0 15 0
T0= 0 0 4 0 1 0 5 0
T1= 8 0 12 0 9 0 13 0
T2= 2 0 6 0 3 0 7 0
T3= 10 0 14 0 11 0 15 0
0 4 8 12
1 5 9 13
2 6 10 14
3 7 11 15
- _mm_unpacklo_epi32 是將兩個參數的 lower bits 以 32 bits 為單位輪流排序並 return
- _mm_unpackhi_epi32 是將兩個參數的 higher bits 以 32 bits 為單位輪流排序並 return
- _mm_unpacklo_epi64 是將兩個參數的 lower bits 以 64 bits 為單位輪流排序並 return
- _mm_unpackhi_epi64 是將兩個參數的 higher bits 以 64 bits 為單位輪流排序並 return
- 藉此完成了矩陣轉置
Prefetch
- _mm_prefetch(char * p , int i ) 會將 address p 裡的數值加載到更靠近 processor 的位置
- Hint 有四種值
- 使用 prefetch 後時間有明顯的縮短
sse prefetch: 68962 us
- 將 PFDIST 從 8 改成 4 看看
sse prefetch: 75842 us
時間差了約 7000 us
- 將 Hint 改為 T0、T1、T2 三者沒有很大的差別，但改為 NTA 後時間大幅增加
sse prefetch: 128963 us
- 自己對 _mm_prefetch 的理解，在運算時可以同時 prefetch 之後會用到的資料，減少下一次運算需要載入資料的時間，所以當 PFDIST = 8、Hint = NTA 時，先載入了下兩次要用的資料到 L1 cache 並標示為 Non-temporal，但是下一次運算不會立即用到，所以有可能在下一次運算時又把他取代了
- 將 PFDIST = 4、Hint = NTA，因為立刻被用到所以時間減少了
./sse_prefetch
sse_prefetch: 70545 us
Object-Oriented Programming
- 參考你所不知道的 C 語言：物件導向程式設計篇進行封裝
- 將變數及 transpose 的 function 包成一個 object
struct object {
int *src;
int *dst;
int w;
int h;
func_t transpose;
};
- 用三個 init 來讓 transpose 指向不同的 transpose function
void init_object_naive(Object **self)
{
...
(*self)->transpose = naive_transpose;
}
void init_object_sse(Object **self)
{
...
(*self)->transpose = sse_transpose;
}
void init_object_sse_prefetch(Object **self)
{
...
(*self)->transpose = sse_prefetch_transpose;
}
- 在 Makefile 中加上 -D 來區別不同的執行檔需要 define 的參數
naive: main.c
$(CC) $(CFLAGS) -D NAIVE -o naive main.c
sse: main.c
$(CC) $(CFLAGS) -D SSE -o sse main.c
sse_prefetch: main.c
$(CC) $(CFLAGS) -D SSE_PREFETCH -o sse_prefetch main.c
- 在 main.c 中用 ifdef 來實現不同的 init
Object *o = NULL;
#ifdef NAIVE
if(init_object_naive(&o) == -1) {
printf("Init object naive fail\n");
return -1;
}
#elif SSE
if(init_object_sse(&o) == -1) {
printf("Init object sse fail\n");
return -1;
}
#elif SSE_PREFETCH
if(init_object_sse_prefetch(&o) == -1) {
printf("Init object sse prefetch fail\n");
return -1;
}
#endif
Implement a interface
- 應該要做一個 interface，所以修改程式碼為
typedef struct {
void (*transpose)(int *src, int *dst, int w, int h);
} trans_interface;
int naive_init(trans_interface *self)
{
self->transpose = &naive_transpose;
return 0;
}
int sse_init(trans_interface *self)
{
self->transpose = &sse_transpose;
return 0;
}
int sse_prefetch_init(trans_interface *self)
{
self->transpose = &sse_prefetch_transpose;
return 0;
}
Raw counter
- 參考 kaizsv prefetch 共筆
- 從 Intel® 64 and IA-32 Architectures Developer's Manual: Vol. 3B 中的 19.4 章找到 Haswell microarchitecture 所對應的 mask
- 找了幾個想要觀察的項目
| Event Num. | Umask Value | Event Mask Mnemonic | Description |
|---|---|---|---|
| 4CH | 01H | LOAD_HIT_PRE.SW_PF | Non-SW-prefetch load dispatches that hit fill buffer allocated for S/W prefetch. |
| 4CH | 02H | LOAD_HIT_PRE.HW_PF | Non-SW-prefetch load dispatches that hit fill buffer allocated for H/W prefetch. |
| D0H | 81H | MEM_UOPS_RETIRED.ALL_LOADS | All retired load uops |
| D1H | 01H | MEM_LOAD_UOPS_RETIRED.L1_HIT | Retired load uops with L1 cache hits as data sources. |
| D1H | 02H | MEM_LOAD_UOPS_RETIRED.L2_HIT | Retired load uops with L2 cache hits |
| D1H | 04H | MEM_LOAD_UOPS_RETIRED.L3_HIT | Retired load uops with L3 cache hits as data sources. |
| D1H | 08H | MEM_LOAD_UOPS_RETIRED.L1_MISS | Retired load uops missed L1 cache as data sources. |
| D1H | 10H | MEM_LOAD_UOPS_RETIRED.L2_MISS | Retired load uops missed L2. Unknown data source excluded. |
| D1H | 20H | MEM_LOAD_UOPS_RETIRED.L3_MISS | Retired load uops missed L3. Excludes unknown data |
- 用 perf stat 觀察
Performance counter stats for './naive' (100 runs):
15,813,284 cache-misses # 97.987 % of all cache refs ( +- 0.33% ) (13.75%)
16,138,211 cache-references ( +- 0.37% ) (14.55%)
19,456,866 L1-dcache-load-misses # 3.57% of all L1-dcache hits ( +- 0.35% ) (21.86%)
545,545,436 L1-dcache-loads ( +- 0.17% ) (21.14%)
198,561,491 L1-dcache-stores ( +- 0.43% ) (19.73%)
47,086 L1-icache-load-misses ( +- 2.16% ) (13.62%)
446 r014c //LOAD_HIT_PRE.SW_PF ( +- 6.35% ) (13.54%)
324,008 r024c //LOAD_HIT_PRE.HW_PF ( +- 7.39% ) (21.03%)
509,653,725 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.45% ) (19.74%)
474,628,725 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.38% ) (18.81%)
17,779 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 9.45% ) (13.77%)
443,351 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 1.18% ) (13.68%)
15,306,768 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 0.42% ) (13.58%)
15,355,248 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 0.38% ) (13.49%)
14,982,092 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 0.53% ) (13.40%)
0.440921106 seconds time elapsed ( +- 0.31% )
Performance counter stats for './sse' (100 runs):
4,242,455 cache-misses # 88.814 % of all cache refs ( +- 0.43% ) (13.87%)
4,776,803 cache-references ( +- 0.43% ) (14.98%)
7,758,666 L1-dcache-load-misses # 1.78% of all L1-dcache hits ( +- 0.34% ) (22.39%)
435,901,161 L1-dcache-loads ( +- 0.30% ) (21.24%)
222,396,741 L1-dcache-stores ( +- 0.38% ) (19.23%)
43,121 L1-icache-load-misses ( +- 2.63% ) (13.69%)
373 r014c //LOAD_HIT_PRE.SW_PF ( +- 6.48% ) (13.78%)
295,849 r024c //LOAD_HIT_PRE.HW_PF ( +- 7.43% ) (21.43%)
398,878,132 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.61% ) (19.64%)
377,177,446 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.42% ) (18.47%)
8,856 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 12.83% ) (14.03%)
125,420 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 1.94% ) (13.92%)
3,771,031 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 0.84% ) (13.74%)
3,793,111 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 0.81% ) (13.59%)
3,678,870 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 0.83% ) (13.41%)
0.317429370 seconds time elapsed ( +- 0.31% )
Performance counter stats for './sse_prefetch' (100 runs):
3,978,276 cache-misses # 90.202 % of all cache refs ( +- 0.88% ) (13.96%)
4,410,407 cache-references ( +- 0.56% ) (15.38%)
7,640,737 L1-dcache-load-misses # 1.68% of all L1-dcache hits ( +- 0.73% ) (23.12%)
455,983,625 L1-dcache-loads ( +- 0.25% ) (21.90%)
228,291,008 L1-dcache-stores ( +- 0.27% ) (19.32%)
41,051 L1-icache-load-misses ( +- 2.26% ) (13.53%)
738,067 r014c //LOAD_HIT_PRE.SW_PF ( +- 3.08% ) (13.94%)
263,426 r024c //LOAD_HIT_PRE.HW_PF ( +- 8.58% ) (22.27%)
409,935,362 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.68% ) (19.97%)
405,300,498 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.54% ) (18.03%)
3,266,225 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 1.14% ) (14.18%)
20,790 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 5.32% ) (14.08%)
3,425,788 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 1.46% ) (13.88%)
47,714 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 7.27% ) (13.61%)
24,762 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 6.40% ) (13.38%)
0.260656173 seconds time elapsed ( +- 0.90% )
- LOAD_HIT_PRE.SW_PF 的部份 sse_prefetch 特別高，也代表真的有用到 SW prefetch
- LOAD_HIT_PRE.HW_PF 的部份都差不多，因為三個用到的 HW prefetch 都是全開的狀態
- 發現 sse_prefetch 在 L2 cache HIT 的情況增加許多，而 且 L2 cache MISS 減少許多，因為使用 _MM_HINT_T1 的關係，將資料 prefetch 到 L2 跟 L3 了
AVX
- 用 Intel Intrinsics Guide 幫忙找到要用的指令
- 查了許多資料查到幾個可用的函式
- 思考如何有效的轉置一個 8 * 8 的矩陣，所以再度用 function 把各個轉完的結果印出來
void print256_num(__m256i var)
{
uint16_t *val = (uint16_t*) &var;
printf(" %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d %2d\n",
val[0], val[1], val[2], val[3], val[4], val[5],
val[6], val[7], val[8], val[9], val[10], val[11],
val[12], val[13], val[14], val[15]);
}
- 最後用 _mm256_unpacklo_epi32、_mm256_unpacklo_epi32、_mm256_permute2x128_si256 完成矩陣的轉置
- 在編譯時多加 -mavx2
CFLAGS = -msse2 --std gnu99 -O0 -Wall -Wextra -mavx2
- 實際跑跑看
illusion030@illusion030-X550LD:~/Desktop/2017sysprog/prefetcher$ make run
./naive
naive: 289409 us
./sse
sse: 127841 us
./sse_prefetch
sse_prefetch: 70374 us
./avx
avx: 66426 us
- 用 AVX 的速度明顯快很多，因為是一次算 8 * 8 的矩陣，跟 SSE 一次算 4 * 4 比較，需要算的次數變少許多
- 用 perf stat 分析看看
Performance counter stats for './avx' (100 runs):
3,092,766 cache-misses # 86.521 % of all cache refs ( +- 0.54% ) (14.04%)
3,574,569 cache-references ( +- 0.63% ) (15.41%)
7,226,514 L1-dcache-load-misses # 1.86% of all L1-dcache hits ( +- 0.75% ) (23.17%)
388,074,299 L1-dcache-loads ( +- 0.23% ) (21.88%)
201,357,224 L1-dcache-stores ( +- 0.33% ) (19.41%)
45,747 L1-icache-load-misses ( +- 2.91% ) (13.78%)
581 r014c //LOAD_HIT_PRE.SW_PF ( +- 10.65% ) (13.95%)
326,686 r024c //LOAD_HIT_PRE.HW_PF ( +- 7.99% ) (21.75%)
357,270,167 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.79% ) (19.44%)
343,995,831 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.58% ) (17.84%)
13,222 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 12.09% ) (14.20%)
80,787 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 2.55% ) (14.01%)
951,656 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 1.21% ) (13.86%)
944,359 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 1.28% ) (13.62%)
879,774 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 1.30% ) (13.46%)
0.259288102 seconds time elapsed ( +- 0.38% )
- 每一層 cache 的 load misses 都比 sse 少很多
AVX Prefetch
- 因為 AVX 和 AVX2 的函式庫裡沒有 Prefetch 的 Intrinsic，所以先用 SSE2 的 _mm_prefetch
- prefetch 8 次
_mm_prefetch(src + (y + AVX_PFDIST + 0) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 1) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 2) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 3) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 4) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 5) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 6) * w + x, _MM_HINT_T1);
_mm_prefetch(src + (y + AVX_PFDIST + 7) * w + x, _MM_HINT_T1);
- 用 PFDIST = 16 (下兩次)、Hint = T1 跑跑看
./naive
naive: 244264 us
./sse
sse: 130113 us
./sse_prefetch
sse_prefetch: 74131 us
./avx
avx: 65895 us
./avx_prefetch
avx_prefetch: 67076 us
- 耗的時間反而提高了，用 perf stat 觀察看看
Performance counter stats for './avx_prefetch' (100 runs):
3,071,554 cache-misses # 61.280 % of all cache refs ( +- 0.52% ) (14.09%)
5,012,325 cache-references ( +- 0.47% ) (15.50%)
6,729,740 L1-dcache-load-misses # 1.68% of all L1-dcache hits ( +- 0.55% ) (23.25%)
400,788,741 L1-dcache-loads ( +- 0.18% ) (22.03%)
200,235,391 L1-dcache-stores ( +- 0.25% ) (19.95%)
64,761 L1-icache-load-misses ( +- 2.66% ) (14.43%)
191,524 r014c //LOAD_HIT_PRE.SW_PF ( +- 3.10% ) (13.55%)
550,186 r024c //LOAD_HIT_PRE.HW_PF ( +- 4.88% ) (20.59%)
385,091,102 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.73% ) (18.18%)
358,884,415 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.74% ) (17.42%)
32,042 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 1.97% ) (14.16%)
553,243 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 1.15% ) (14.16%)
1,182,803 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 1.19% ) (13.92%)
1,194,809 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 0.93% ) (13.64%)
627,029 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 0.59% ) (13.38%)
0.261470543 seconds time elapsed ( +- 0.21% )
- LOAD_HIT_PRE.SW_PF 有增加代表 _mm_prefetch 有用
- L2 cache HIT 只增加一點點，L3 cache HIT 倒是增加了許多，但是 Hint 是 T1 預期是 L2 cache HIT 會增加
- prefetch 有作用但是時間卻沒有明顯減少反而增加了，假設是 PFDIST 取的不好，來認真研究一下 PFDIST 要怎麼取
PFDIST
- 公式
- l : prefetch latency
s : 迴圈當中的最短路徑 - 參考 carolc0708 的共筆 找出 memory latency
- 先去 Intel® Memory Latency Checker v3.1a 下載 mlc 並解壓縮後執行
- 但是就算執行了 $ sudo modprobe msr 還是會出現錯誤
illusion030@illusion030-X550LD:~/Desktop/memory_latency$ sudo modprobe msr
illusion030@illusion030-X550LD:~/Desktop/memory_latency$ ./mlc
Intel(R) Memory Latency Checker - v3.1a
tried to open /dev/cpu/0/msr
Cannot access MSR module. Possibilities include not having root privileges, or MSR module not inserted (try 'modprobe msr', and refer README)
- 索性直接進入 root 模式來執行看看，結果成功了
root@illusion030-X550LD:/home/illusion030/Desktop/memory_latency# ./mlc
Intel(R) Memory Latency Checker - v3.1a
Measuring idle latencies (in ns)...
Memory node
Socket 0
0 62.3
Measuring Peak Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 19368.4
3:1 Reads-Writes : 16337.4
2:1 Reads-Writes : 16425.5
1:1 Reads-Writes : 18360.1
Stream-triad like: 17216.6
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Memory node
Socket 0
0 19441.2
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 74.88 17074.9
00002 74.87 17084.0
00008 74.79 17060.9
00015 74.34 16855.8
00050 66.22 9211.8
00100 65.24 6029.9
00200 63.82 3733.0
00300 64.71 2855.0
00400 63.50 2424.7
00500 63.97 2143.7
00700 63.72 1828.5
01000 63.42 1589.5
01300 63.21 1460.4
01700 63.10 1357.4
02500 63.07 1248.7
03500 63.06 1182.3
05000 62.99 1133.5
09000 62.96 1081.9
20000 62.99 1045.7
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 30.4
Local Socket L2->L2 HITM latency 35.7
- 邊執行程式邊跑 mlc 得到的 latency
Measuring Loaded Latencies for the system
Using all the threads from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 86.82 15573.2
00002 85.72 15515.8
00008 85.22 15528.7
00015 86.11 15460.6
00050 75.05 8977.5
00100 73.72 5871.2
00200 71.39 3587.3
00300 72.28 2730.4
00400 71.15 2291.8
00500 70.06 2041.5
00700 70.79 1712.6
01000 70.38 1474.3
01300 70.33 1347.3
01700 69.51 1259.1
02500 69.82 1146.5
03500 69.62 1083.9
05000 69.46 1037.1
09000 69.98 978.6
20000 69.15 954.5
- 算出平均為 74.029473684 (ns)
- 接下來找出迴圈跑的時間
loop time = 65344 us
avx: 65358 us
- each loop time = 65344 us / (512 * 512) = 0.249267578
- 1 us = 1000 ns
- 74.029473684 / 249.267578125 = 0.296987977
- 取上高斯之後 D >= 1，所以把 avx_prefetch 的 PFDIST 設成 8，也就是一倍，用 T1 跑跑看
naive: 247381 us
sse: 128336 us
sse_prefetch: 71009 us
avx: 66292 us
avx_prefetch: 64742 us
- 時間還是沒有加快多少，再用 perf stat 觀察看看
Performance counter stats for './avx_prefetch' (100 runs):
3,162,625 cache-misses # 79.583 % of all cache refs ( +- 0.44% ) (14.07%)
3,973,992 cache-references ( +- 0.47% ) (15.46%)
7,050,577 L1-dcache-load-misses # 1.75% of all L1-dcache hits ( +- 0.53% ) (23.21%)
403,421,197 L1-dcache-loads ( +- 0.14% ) (21.87%)
202,331,106 L1-dcache-stores ( +- 0.17% ) (19.32%)
57,234 L1-icache-load-misses ( +- 2.94% ) (13.47%)
344,731 r014c //LOAD_HIT_PRE.SW_PF ( +- 2.17% ) (13.63%)
262,287 r024c //LOAD_HIT_PRE.HW_PF ( +- 8.04% ) (22.16%)
353,934,197 r81d0 //MEM_UOPS_RETIRED.ALL_LOADS ( +- 0.68% ) (19.86%)
349,566,969 r01d1 //MEM_LOAD_UOPS_RETIRED.L1_HIT ( +- 0.42% ) (18.09%)
780,509 r02d1 //MEM_LOAD_UOPS_RETIRED.L2_HIT ( +- 0.47% ) (14.26%)
50,093 r04d1 //MEM_LOAD_UOPS_RETIRED.L3_HIT ( +- 2.24% ) (14.12%)
1,227,596 r08d1 //MEM_LOAD_UOPS_RETIRED.L1_MISS ( +- 0.72% ) (13.87%)
452,265 r10d1 //MEM_LOAD_UOPS_RETIRED.L2_MISS ( +- 0.87% ) (13.66%)
400,909 r20d1 //MEM_LOAD_UOPS_RETIRED.L3_MISS ( +- 0.91% ) (13.44%)
0.256960590 seconds time elapsed ( +- 0.18% )
-
cache hit 的部份有比較跟預期的一樣， L2 LOAD HIT 有增加
-
因為 l / s 實在太小了只有大約 0.3，沒辦法取到一個很好的 PFDIST，所以 AVX+prefetch 在這個程式上沒有很好的表現
