discover optimization strategy for no local memory situation

[ysh329]

1. Do matrix Transpose before matrix multiplication: further optimize matrix transpose
2. Refer CPU GEMM optimization strategy (Mali GPU withou local memory) step by step: https://github.com/flame/how-to-optimize-gemm/wiki

[ysh329]

temporarily closed.

[ysh329]

According row-first, we can use vector-variable so as to decrease the times of buffer reading, such as below:

// perf up! 24% 
// 1024x1024x1024 0.458464 s 4.684079 GFLOPS
// naive float : 0.59s ; this float2: 0.45s
__kernel void mat_mult_vec2x2_continue(const int M, const int N, const int K, __global const CL_INPUT_TYPE *a, __global const CL_INPUT_TYPE *b, __global CL_INPUT_TYPE *c) {
    const int col = get_global_id(0) << 1;
    const int row = get_global_id(1) << 1;

    CL_ELEM_TYPE aa1, aa2,
                 bb1, bb2,
                 cc1 = 0,  
                 cc2 = 0;

    for (int p = 0; p < K; p+=2) {
        aa1 = *(
                   (__global CL_ELEM_TYPE *)(a + row * K + p)
               );  
        aa2 = *(
                   (__global CL_ELEM_TYPE *)(a + (row+1) * K + p)
               );  

        bb1 = *(
                   (__global CL_ELEM_TYPE *)(b + p * N + col)
               );  
        bb2 = *(
                   (__global CL_ELEM_TYPE *)(b + (p+1) * N + col)
               );  
    
        cc1.s0 += aa1.s0 * bb1.s0 + aa1.s1 * bb2.s0;
        cc1.s1 += aa1.s0 * bb1.s1 + aa1.s1 * bb2.s1;
        cc2.s0 += aa2.s0 * bb1.s0 + aa2.s1 * bb2.s0;
        cc2.s1 += aa2.s0 * bb1.s1 + aa2.s1 * bb2.s1;
    }

    c[row * N + col] = cc1.s0;        c[row * N + (col+1)] = cc1.s1;
    c[(row+1) * N + col] = cc2.s0;    c[(row+1) * N + (col+1)] = cc2.s1;

}