/**
* Block panel product
* @param The original size M
* @param Pointer to A at the current row,column
* @param num_rows_AC The number of valid rows in A,C
* @param num_acc The number of valid cols in A, and rows in B
* @param Pointer to C at the current row
*/
void gebp_opt1(const int M, const double* A, const int num_rows_AC, const int num_acc, double* C)
{
// Pack A into memory
for(int iter_col = 0; iter_col < num_acc; ++iter_col)
{
memcpy(A_pack + iter_col * num_rows_AC,
A + iter_col * M,
num_rows_AC * sizeof(double));
}
// For each slice in B,C
const int num_slice_blocks = CALC_NUM_BLOCKS(M, 2);
for(int iter_slice_block = 0; iter_slice_block < num_slice_blocks; ++iter_slice_block)
{
const int cur_slice_pos = iter_slice_block * 2;
const int num_slice = CALC_CUR_BLOCK_WIDTH(cur_slice_pos, 2, M);
// Copy into kernel memory
to_kdgemm_B_sized(num_acc, B_pack + cur_slice_pos * num_acc, B_kernel, num_acc, num_slice);
// For each row in A,C_aux
const int num_a_aux_blocks = CALC_NUM_BLOCKS(num_rows_AC, 2);
for(int iter_a_aux_block = 0; iter_a_aux_block < num_a_aux_blocks; ++iter_a_aux_block)
{
const int cur_a_aux_pos = iter_a_aux_block * 2;
const int num_a_aux = CALC_CUR_BLOCK_WIDTH(cur_a_aux_pos, 2, num_rows_AC);
// Copy into kernel memory
to_kdgemm_A_sized(num_rows_AC, A_pack + cur_a_aux_pos, A_kernel, num_a_aux, num_acc);
// Clear the value of C
clear_kdgemm_C_sized(C_kernel);
// Run kernel
kdgemm(A_kernel, B_kernel, C_kernel);
// Store results into C_aux
from_kdgemm_C_sized(num_rows_AC, C_kernel, C_aux + cur_a_aux_pos, num_a_aux, num_slice);
}
// Accumulate results from C_aux to C
for(int iter_slice_part = 0; iter_slice_part < num_slice; ++iter_slice_part)
{
for(int iter_row = 0; iter_row < num_rows_AC; ++iter_row)
{
C[iter_row + (cur_slice_pos + iter_slice_part) * M] += C_aux[iter_row + iter_slice_part * num_rows_AC];
}
}
}
}
/*
* Time the matrix multiply
*/
double time_dgemm(const double *A, const double *B, double *C)
{
double secs = -1.0;
double mflops_sec;
int num_iterations = MIN_RUNS;
while (secs < MIN_SECS) {
matrix_clear(C, DIM_M, DIM_N);
double start = omp_get_wtime();
for (int i = 0; i < num_iterations; ++i) {
kdgemm(A, B, C);
}
double finish = omp_get_wtime();
double mflops = 2.0 * num_iterations * DIM_M * DIM_N * DIM_P / 1.0e6;
secs = finish-start;
mflops_sec = mflops / secs;
num_iterations *= 2;
}
return mflops_sec;
}
/*
* Run a basic test and timing trial.
*/
int main(int argc, char** argv)
{
// Allocate space for ordinary column-major matrices
double* A = malloc(DIM_M * DIM_P * sizeof(double));
double* B = malloc(DIM_P * DIM_N * sizeof(double));
double* C = malloc(DIM_M * DIM_N * sizeof(double));
// Allocate aligned scratch space for use by the kernel
double* Ak = _mm_malloc(DIM_M * DIM_P * sizeof(double), 16);
double* Bk = _mm_malloc(DIM_P * DIM_N * sizeof(double), 16);
double* Ck = _mm_malloc(DIM_M * DIM_N * sizeof(double), 16);
// Initialize the input matrices and convert to kernel format
matrix_init(A, DIM_M, DIM_P);
matrix_init(B, DIM_P, DIM_N);
to_kdgemm_A(DIM_M, A, Ak);
to_kdgemm_B(DIM_P, B, Bk);
// Clear the kernel scratch output, run the kernel, convert to col major
matrix_clear(Ck, DIM_M, DIM_N);
kdgemm(Ak, Bk, Ck);
from_kdgemm_C(DIM_M, Ck, C);
// Check for agreement
double max_diff = check_kdgemm(A, B, C);
// Print kernel dimensions, megaflop rate, and error from check
printf("%u,%u,%u,%lg,%0.0e\n", DIM_M, DIM_P, DIM_N,
time_dgemm(Ak, Bk, Ck),
max_diff);
// Free kernel matrix space
_mm_free(Ck);
_mm_free(Bk);
_mm_free(Ak);
// Free argument matrix space
free(C);
free(B);
free(A);
return 0;
}