Lab 11#

Introduction to SIMD#

SIMD makes a program faster by executing the same instruction on multiple data at the same time. In this lab, we will use Intel Intrinsics to implement simple programs.

Part 1: Vector addition#

In this part, you will implement a vector addition program using SIMD. Please “translate” naive_add() to simd_add().
You may use the following intrinsics, search in the Intel Intrinsics Guide:

• _mm_loadu_si128
• _mm_storeu_si128
• _mm_add_epi32

Try to tell the difference of the following “load” intrisics:

• _mm_load_si128
• _mm_loadu_si128
• _mm_load_pd
• _mm_load1_pd

Part 2: Matrix multiplication#

In this part, you will implement a matrix multiplication program using SIMD. Please “translate” naive_matmul() to simd_matmul().
You may use the following intrinsics:

• _mm_setzero_ps
• _mm_set1_ps
• _mm_loadu_ps
• _mm_add_ps
• _mm_mul_ps
• _mm_storeu_ps

Explain why this makes the program faster.

Part 3: Loop unrolling#

Read Wikipedia and try to understand the concept of loop unrolling:

• Loop unrolling

Implement loop_unroll_matmul() and loop_unroll_simd_matmul(), explain the performance boost they brought.

Part 4: Compiler optimization#

Run make test, explain why -O3 makes the program much faster.
For checkup: Put this piece of code into godbolt.org , compile them with a risc-v compiler, and tell the difference between -O0 and -O3.

int a = 0;

void modify(int j) { a += j; }

int main() { for (int i = 0; i < 1000; i++) { a += 1; }

for (int i = 0; i < 1000; i++) {
    a += i;
}

return a;

}

Suting Chen <chenst AT shanghaitech.edu.cn>

Last modified: 2024-04-24