diff options
| author | Henry Cook <hcook@eecs.berkeley.edu> | 2014-11-06 17:24:39 -0800 |
|---|---|---|
| committer | Henry Cook <hcook@eecs.berkeley.edu> | 2014-11-07 16:52:51 -0800 |
| commit | d537de7deffa6036dab573ff174b7f8c8e470437 (patch) | |
| tree | ddc921eb337cda4889570f0251bdba85059a2531 /mt/bo_matmul.c | |
| parent | 5afc6b9bc2e3685220cffb3da66ad9f5f1f7b14f (diff) | |
| download | riscv-tests-d537de7deffa6036dab573ff174b7f8c8e470437.zip riscv-tests-d537de7deffa6036dab573ff174b7f8c8e470437.tar.gz riscv-tests-d537de7deffa6036dab573ff174b7f8c8e470437.tar.bz2 | |
Clean up canonical mt benchmarks and reorganize extra versions in /mt. All versions support support at least 1/2/4 threads.
Diffstat (limited to 'mt/bo_matmul.c')
| -rw-r--r-- | mt/bo_matmul.c | 98 |
1 files changed, 98 insertions, 0 deletions
diff --git a/mt/bo_matmul.c b/mt/bo_matmul.c new file mode 100644 index 0000000..2fb24ab --- /dev/null +++ b/mt/bo_matmul.c @@ -0,0 +1,98 @@ +#include "stdlib.h" + +#include "util.h" + +#include "dataset.h" +void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda, const data_t A[], const data_t B[], data_t C[] ) +{ + int i, j, k; + data_t B_trans[32*32]; + data_t acc_temp0, acc_temp1; + data_t *A_j, *B_i; + data_t *A_j_k, *B_i_k; + int z; + + //for (i = 0; i < 32; i++) { + // for (j = 0; j < 32; j++) { + // B_trans[i*lda+j] = B[i+j*lda]; + // } + //} + + if (coreid == 0) { + for (i = 0; i < 32; i++) { + B_i = B_trans+i*32; + for (z = 0; z < 32; z++) { + *(B_i+z) = B[i+z*32]; + } + for (j = 0; j < 16; j+=2) { + A_j = A+j*lda; + acc_temp0 = 0; + for (k = 0; k < 32; k+=8) { + A_j_k = A_j+k; + B_i_k = B_i+k; + acc_temp0 += *(A_j_k) * *(B_i_k); + acc_temp0 += *(A_j_k + 1) * *(B_i_k + 1); + acc_temp0 += *(A_j_k + 2) * *(B_i_k + 2); + acc_temp0 += *(A_j_k + 3) * *(B_i_k + 3); + acc_temp0 += *(A_j_k + 4) * *(B_i_k + 4); + acc_temp0 += *(A_j_k + 5) * *(B_i_k + 5); + acc_temp0 += *(A_j_k + 6) * *(B_i_k + 6); + acc_temp0 += *(A_j_k + 7) * *(B_i_k + 7); + } + A_j += 32; + + acc_temp1 = 0; + for (k = 0; k < 32; k+=8) { + acc_temp1 += *(A_j+k) * *(B_i+k); + acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1); + acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2); + acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3); + acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4); + acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5); + acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6); + acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7); + } + + C[i + j*lda] = acc_temp0; + C[i + (j+1)*lda] = acc_temp1; + } + } + } + if (coreid == 1 || ncores == 1) { + for (i = 0; i < 32; i++) { + B_i = B_trans+i*32; + for (z = 0; z < 32; z++) { + *(B_i+z) = B[i+z*32]; + } + for (j = 16; j < 32; j+=2) { + A_j = A+j*lda; + acc_temp0 = 0; + for (k = 0; k < 32; k+=8) { + acc_temp0 += *(A_j+k) * *(B_i+k); + acc_temp0 += *(A_j+k + 1) * *(B_i+k + 1); + acc_temp0 += *(A_j+k + 2) * *(B_i+k + 2); + acc_temp0 += *(A_j+k + 3) * *(B_i+k + 3); + acc_temp0 += *(A_j+k + 4) * *(B_i+k + 4); + acc_temp0 += *(A_j+k + 5) * *(B_i+k + 5); + acc_temp0 += *(A_j+k + 6) * *(B_i+k + 6); + acc_temp0 += *(A_j+k + 7) * *(B_i+k + 7); + } + A_j += 32; + + acc_temp1 = 0; + for (k = 0; k < 32; k+=8) { + acc_temp1 += *(A_j+k) * *(B_i+k); + acc_temp1 += *(A_j+k + 1) * *(B_i+k + 1); + acc_temp1 += *(A_j+k + 2) * *(B_i+k + 2); + acc_temp1 += *(A_j+k + 3) * *(B_i+k + 3); + acc_temp1 += *(A_j+k + 4) * *(B_i+k + 4); + acc_temp1 += *(A_j+k + 5) * *(B_i+k + 5); + acc_temp1 += *(A_j+k + 6) * *(B_i+k + 6); + acc_temp1 += *(A_j+k + 7) * *(B_i+k + 7); + } + C[i + j*lda] = acc_temp0; + C[i + (j+1)*lda] = acc_temp1; + } + } + } +} |