| /dports/math/libxsmm/libxsmm-1.16.3/samples/deeplearning/lstmdriver/ |
| H A D | lstmdriver_nc_kcck_bf16.c | 47 …float *wigold, *wfgold, *wogold, *wcgold, *rigold, *rfgold, *rogold, *rcgold, *bigold, *bfgold, *b… in main() local
181 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
252 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
303 rne_mask_fp32_bf16( bfgold, bfgold, K ); in main()
329 bigold, bcgold, bfgold, bogold, in main()
515 matrix_copy_f32_bf16(K, bfgold, &(b[2*K])); in main()
841 bigold, bcgold, bfgold, bogold, in main()
1134 libxsmm_free(bfgold); in main()
|
| H A D | lstmdriver_nc_ck_bf16.c | 47 …float *wigold, *wfgold, *wogold, *wcgold, *rigold, *rfgold, *rogold, *rcgold, *bigold, *bfgold, *b… in main() local
179 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
246 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
297 rne_mask_fp32_bf16( bfgold, bfgold, K ); in main()
323 bigold, bcgold, bfgold, bogold, in main()
491 matrix_copy_f32_bf16(K, bfgold, &(b[2*K])); in main()
802 bigold, bcgold, bfgold, bogold, in main()
1090 libxsmm_free(bfgold); in main()
|
| H A D | lstmdriver_nc_ck_f32.c | 42 …float *wigold, *wfgold, *wogold, *wcgold, *rigold, *rfgold, *rogold, *rcgold, *bigold, *bfgold, *b… in main() local
173 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
234 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
282 bigold, bcgold, bfgold, bogold, in main()
434 matrix_copy(K, bfgold, &(b[2*K])); in main()
719 bigold, bcgold, bfgold, bogold, in main()
1016 libxsmm_free(bfgold); in main()
|
| H A D | lstmdriver_nc_kcck_f32.c | 37 …*wcgold, *xgoldt, *rigold, *rfgold, *rogold, *rcgold, *hgoldt, *bigold, *bfgold, *bogold, *bcgold,… in main() local
175 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
297 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
301 bfgold_fb[j] = bfgold[j] + forget_bias; in main()
690 matrix_copy(K, bfgold, &(b[2*K])); in main()
1290 libxsmm_free(bfgold); in main()
|
| /dports/math/libxsmm/libxsmm-1.16.3/samples/deeplearning/grudriver/ |
| H A D | grudriver_nc_ck.c | 37 float *wigold, *wcgold, *wfgold, *rigold, *rcgold, *rfgold, *bigold, *bcgold, *bfgold; in main() local
160 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
215 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
253 bigold, bcgold, bfgold, in main()
370 matrix_copy(K, bfgold, &(b[2*K])); in main()
880 libxsmm_free(bfgold); in main()
|
| H A D | grudriver_nc_kcck.c | 37 float *wigold, *wcgold, *wfgold, *rigold, *rcgold, *rfgold, *bigold, *bcgold, *bfgold; in main() local
160 bfgold = (float*)libxsmm_aligned_malloc(K*sizeof(float), 2097152); in main()
217 LIBXSMM_MATINIT_OMP(float, 24, bfgold, 1, K, 1, 1.0); in main()
257 bigold, bcgold, bfgold, in main()
376 matrix_copy(K, bfgold, &(b[2*K])); in main()
898 libxsmm_free(bfgold); in main()
|
| /dports/math/libxsmm/libxsmm-1.16.3/samples/deeplearning/common/ |
| H A D | dnn_common.h | 2481 LIBXSMM_INLINE void lstm_fwd_copy_bias(int N, int K, float *bigold, float *bcgold, float *bfgold, f… in lstm_fwd_copy_bias() argument
2493 LIBXSMM_VLA_ACCESS(3, icfogold, j, i, l+2*K, N, 4 * K) = bfgold[l] + forget_bias; in lstm_fwd_copy_bias()
2733 float *bigold, float *bcgold, float *bfgold, float *bogold, in lstm_ref_fwd() argument
2783 lstm_fwd_copy_bias(N, K, bigold, bcgold, bfgold, bogold, forget_bias, icfogoldt, j); in lstm_ref_fwd()
|