chundoong-lab-ta/SamsungDS22/submissions/HW2/km.hero.lee/mat_mul.cpp

#include "mat_mul.h"

#include <cstdlib>
#include <cstdio>
#include <pthread.h>
#define MIN(x, y) (((x) < (y)) ? (x) : (y))
#define B_SIZE 32

static float *A, *B, *C;
static int M, N, K;
static int num_threads;

static void* mat_mul_thread(void *data) {

/*
 * Non-optimized
 */
#ifdef _O0
    for (int i = 0; i < M; ++i) {
        for (int j = 0; j < N; ++j) {
            for (int k = 0; k < K; ++k) {
                C[i * N + j] += A[i * K + k] * B[k * N + j];
            }
        }
    }
/*
 * Array dimension change
 * i-j-k -> i-k-j
 */
#elif _O1
    for (int i = 0; i < M; ++i) {
        for (int k = 0; k < K; ++k) {
            for (int j = 0; j < N; ++j) {
                register float tmp = A[i * K + k];
                C[i * N + j] += tmp * B[k * N + j];
            }
        }
    }
/*
 * Blocked matrix multiplication
 * Best block size: 256 bytes
 */
#elif _O2
    for (int ii = 0; ii < M; ii+=B_SIZE)
        for (int kk = 0; kk < K; kk+=B_SIZE)
            for (int jj = 0; jj < N; jj+=B_SIZE)
                for (int i=ii; i<MIN(ii+B_SIZE, M); i++)
                    for (int k=kk; k<MIN(kk+B_SIZE, K); k++) {
                        register float tmp = A[i * K + k];
                        for (int j=jj; j<MIN(jj+B_SIZE, N); j++)
                            C[i * N + j] += tmp * B[k * N + j];
                    }

/*
 * Eliminate redundant blocked dimension
 */
#elif _O3_BASE
    for (int ii = 0; ii < M; ii+=B_SIZE)
        for (int kk = 0; kk < K; kk+=B_SIZE)
            for (int jj = 0; jj < N; jj+=B_SIZE)
                for (int i=ii; i<MIN(ii+B_SIZE, M); i++)
                    for (int k=kk; k<MIN(kk+B_SIZE, K); k++) {
                        register float tmp = A[i * K + k];
                        for (int j=jj; j<MIN(jj+B_SIZE, N); j++)
                            C[i * N + j] += tmp * B[k * N + j];
                    }
#elif _O3_DEL_I
    for (int kk = 0; kk < K; kk+=B_SIZE)
        for (int jj = 0; jj < N; jj+=B_SIZE)
            for (int i=0; i<M; i++)
                for (int k=kk; k<MIN(kk+B_SIZE, K); k++) {
                    register float tmp = A[i * K + k];
                    for (int j=jj; j<MIN(jj+B_SIZE, N); j++)
                        C[i * N + j] += tmp * B[k * N + j];
                }
#elif _O3_DEL_J
    for (int ii = 0; ii < M; ii+=B_SIZE)
        for (int kk = 0; kk < K; kk+=B_SIZE)
            for (int i=ii; i<MIN(ii+B_SIZE, M); i++)
                for (int k=kk; k<MIN(kk+B_SIZE, K); k++) {
                    register float tmp = A[i * K + k];
                    for (int j=0; j<N; j++)
                        C[i * N + j] += tmp * B[k * N + j];
                }
#elif _O3_DEL_K
    for (int ii = 0; ii < M; ii+=B_SIZE)
        for (int jj = 0; jj < N; jj+=B_SIZE)
            for (int i=ii; i<MIN(ii+B_SIZE, M); i++)
                for (int k=0; k<K; k++) {
                    register float tmp = A[i * K + k];
                    for (int j=jj; j<MIN(jj+B_SIZE, N); j++)
                        C[i * N + j] += tmp * B[k * N + j];
                }
#elif _O3_DEL_IJ
    for (int kk = 0; kk < K; kk+=B_SIZE)
        for (int i=0; i<M; i++)
            for (int k=kk; k<MIN(kk+B_SIZE, K); k++) {
                register float tmp = A[i * K + k];
                for (int j=0; j<N; j++)
                    C[i * N + j] += tmp * B[k * N + j];
            }
/*
 * Using pthread library
 * Based on _O3_DEL_IJ
 */
#elif _O4
    int tid = * (int *) data;
    int elemPerTh = M / num_threads;
    int start = tid * elemPerTh;
    int end = MIN(M, ((tid + 1) * elemPerTh));

    register float tmp;
    for (int kk = 0; kk < K; kk += B_SIZE)
        for (int i = start; i < end; i++)
            for (int k = kk; k < MIN(kk + B_SIZE, K); k++) {
                tmp = A[i * K + k];
                for (int j = 0; j < N; j++)
                    C[i * N + j] += tmp * B[k * N + j];
            }

/*
 * Using SIMD instruction
 */
#elif _OAVX
    for (int i = 0; i < M; ++i) {
        for (int k = 0; k < K; ++k) {
            __m512 a0;
           a0 = _mm512_load_ps(A + (i * K + k));
            for (int j = 0; j < N; j+=B_SIZE) {

                __m512 b0;
                __m512 s0;

                b0 = _mm512_load_ps(B + (k * N + j));
                s0 = _mm512_load_ps(C + (i * N + j));
                s0 = _mm512_fmadd_ps(a0, b0, s0);
                //               _mm512_store_ps((C + (i * N + j)), s0);
            }
        }
    }
#endif

    int tid = * (int *) data;
    int elemPerTh = (M + num_threads - 1) / num_threads;
    int start = tid * elemPerTh;
    int end = MIN(M, ((tid + 1) * elemPerTh));


    register float tmp;
    for (int kk = 0; kk < K; kk += B_SIZE)
        for (int i = start; i < end; i++)
            for (int k = kk; k < MIN(kk + B_SIZE, K); k++) {
                tmp = A[i * K + k];
                for (int j = 0; j < N; j++)
                    C[i * N + j] += tmp * B[k * N + j];
            }

    return NULL;
}

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K, int _num_threads) {
    A = _A, B = _B, C = _C;
    M = _M, N = _N, K = _K;
    num_threads = _num_threads;

    pthread_t * threads = new pthread_t[num_threads];
    int tid[500];

    for (int i=0; i<num_threads; i++) {
        tid[i] = i;
        pthread_create(&threads[i], NULL, mat_mul_thread, tid + i);
    }

    for (int i = 0; i < num_threads; i++) {
        pthread_join(threads[i], NULL);
    }
}