Skip to content

Commit

Permalink
Added C matrix indexing based on a generalized hilbert space filling …
Browse files Browse the repository at this point in the history 
…curve traversal and made it the default option (#37)

* added SFC tools

* added initial SFC impl for FC fwd

* updated libxsmm

* fixed memory leak

* fixed compilation of mlp_example
  • Loading branch information
@alheinecke
alheinecke authored Jul 29, 2024
1 parent 9caa6de commit 0423291
Show file tree
Hide file tree
Showing 6 changed files with 403 additions and 41 deletions.
3 changes: 3 additions & 0 deletions include/libxsmm_dnn_fc.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -70,6 +70,9 @@ typedef struct libxsmm_dnn_fc_fwd_config {
libxsmm_meltwfunction_unary fwd_colbcast_load_kernel;
libxsmm_meltwfunction_unary fwd_store_kernel;
libxsmm_meltwfunction_unary fwd_vnniT_kernel;
libxsmm_blasint fwd_sfc;
libxsmm_blasint fwd_sfc_map_tsize;
unsigned char* fwd_sfc_map;
} libxsmm_dnn_fc_fwd_config;

typedef struct libxsmm_dnn_fc_bwd_config {
Expand Down
2 changes: 1 addition & 1 deletion libxsmm
Submodule libxsmm updated 6 files
+1 −0 include/libxsmm_macros.h
+5 −1 src/generator_gemm.c
+8 −8 src/generator_gemm_avx2_microkernel.c
+5 −1 src/generator_gemm_common.c
+28 −14 src/generator_gemm_sse_avx_avx2_avx512.c
+1 −1 version.txt
160 changes: 126 additions & 34 deletions src/libxsmm_dnn_fc.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
/* Kirill Voronin, Alexander Heinecke (Intel Corp.)
******************************************************************************/
#include <libxsmm_dnn_fc.h>
#include "libxsmm_dnn_sfc_tools.h"

#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wcast-qual"
Expand All @@ -35,6 +36,8 @@ LIBXSMM_API libxsmm_dnn_fc_fwd_config setup_libxsmm_dnn_fc_fwd(libxsmm_blasint N
libxsmm_gemm_ext_unary_argops l_argops;
libxsmm_gemm_ext_binary_postops l_postops;

const char *const l_env_sfc = getenv("LIBXSMM_DNN_DISABLE_SFC");

/* init libxsmm structs */
memset( &l_shape, 0, sizeof(libxsmm_gemm_shape) );
memset( &l_brconfig, 0, sizeof(libxsmm_gemm_batch_reduce_config) );
Expand All @@ -61,7 +64,24 @@ LIBXSMM_API libxsmm_dnn_fc_fwd_config setup_libxsmm_dnn_fc_fwd(libxsmm_blasint N
res.fwd_col_teams = 1;
res.fwd_row_teams = 1;

if ( l_is_aarch64 == 0 ) {
/* check if we should use SFC */
if ( 0 == l_env_sfc ) {
res.fwd_sfc = 1;
} else {
if ( atoi(l_env_sfc) != 0 ) {
res.fwd_sfc = 0;
} else {
res.fwd_sfc = 1;
}
}
if ( res.N % res.bn != 0 ) {
res.fwd_sfc = 0;
}
if ( res.K % res.bk != 0 ) {
res.fwd_sfc = 0;
}

if ( (l_is_aarch64 == 0) && (res.fwd_sfc == 0) ) {
if (threads == 16) {
res.fwd_bf = 1;
res.fwd_2d_blocking = 1;
Expand Down Expand Up @@ -234,7 +254,7 @@ LIBXSMM_API libxsmm_dnn_fc_fwd_config setup_libxsmm_dnn_fc_fwd(libxsmm_blasint N
res.fwd_bf--;
}
}
} else {
} else if ( res.fwd_sfc == 0 ) {
if (threads == 16) {
res.fwd_bf = 1;
res.fwd_2d_blocking = 1;
Expand All @@ -258,6 +278,11 @@ LIBXSMM_API libxsmm_dnn_fc_fwd_config setup_libxsmm_dnn_fc_fwd(libxsmm_blasint N
res.fwd_col_teams = 1;
res.fwd_row_teams = 1;
}
} else {
#if 0
printf("using SFC-based access logic\n");
#endif
res.fwd_sfc_map_tsize = libxsmm_dnn_sfc_create_map(&(res.fwd_sfc_map), res.K/res.bk, res.N/res.bn );
}

#if 0
Expand Down Expand Up @@ -1757,8 +1782,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_f32_flat( libxsmm_dnn_fc_fwd_config cfg
if (BF > 1) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ifm1 == 0 ) {
if ( cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS || cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS_RELU || cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS_RELU_WITH_MASK ) {
gemm_param_ext.op.tertiary = &blocks;
Expand Down Expand Up @@ -1797,8 +1826,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_f32_flat( libxsmm_dnn_fc_fwd_config cfg
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) { /* bias only */
gemm_param_ext.op.tertiary = &blocks;
gemm_param_ext.a.primary = (void*)&LIBXSMM_VLA_ACCESS(4, filter, 0, 0, ofm1, 0, cfg.bc, nBlocksOFm, cfg.bk);
Expand Down Expand Up @@ -1976,8 +2009,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_f32_packed( libxsmm_dnn_fc_fwd_config c
if (BF > 1) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ifm1 == 0 ) {
if ( cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS || cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS_RELU || cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS_RELU_WITH_MASK ) {
gemm_param_ext.op.tertiary = &blocks;
Expand Down Expand Up @@ -2016,8 +2053,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_f32_packed( libxsmm_dnn_fc_fwd_config c
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( cfg.fuse_type == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) { /* bias only */
gemm_param_ext.op.tertiary = &blocks;
gemm_param_ext.a.primary = (void*)&LIBXSMM_VLA_ACCESS(4, filter, ofm1, 0, 0, 0, nBlocksIFm, cfg.bc, cfg.bk);
Expand Down Expand Up @@ -2207,8 +2248,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_flat( libxsmm_dnn_fc_fwd_config cf
if ((BF > 1) || (cfg.K % 32 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -2244,8 +2289,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_flat( libxsmm_dnn_fc_fwd_config cf
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -2418,8 +2467,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_packed( libxsmm_dnn_fc_fwd_config
if ((BF > 1) || (cfg.K % 32 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -2455,8 +2508,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_packed( libxsmm_dnn_fc_fwd_config
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -2631,8 +2688,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_format( libxsmm_dnn_fc_fwd
if ((BF > 1) || (cfg.K % 32 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -2668,8 +2729,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_format( libxsmm_dnn_fc_fwd
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -2844,8 +2909,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_iact_vnniT_format( libxsmm
if ((BF > 1) || (cfg.K % 32 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -2881,8 +2950,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_iact_vnniT_format( libxsmm
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -3060,8 +3133,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_iact_vnniT_oact_vnniT_form
if ((BF > 1) || (cfg.K % 32 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -3097,8 +3174,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf16_wt_vnni_iact_vnniT_oact_vnniT_form
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -3289,8 +3370,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf8_vnni_format( libxsmm_dnn_fc_fwd_con
if ((BF > 1) || (cfg.K % 64 != 0)) {
for ( ifm1 = 0; ifm1 < BF; ++ifm1 ) {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
/* Initialize libxsmm_blasintermediate f32 tensor */
if ( ifm1 == 0 ) {
if ( (cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS ) {
Expand Down Expand Up @@ -3326,8 +3411,12 @@ LIBXSMM_API void libxsmm_dnn_fc_fwd_exec_bf8_vnni_format( libxsmm_dnn_fc_fwd_con
}
} else {
for ( mb1ofm1 = thr_begin; mb1ofm1 < thr_end; ++mb1ofm1 ) {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
if (cfg.fwd_sfc == 1) {
libxsmm_dnn_sfc_get_coord(&ofm1, &mb1, cfg.fwd_sfc_map, mb1ofm1, cfg.fwd_sfc_map_tsize);
} else {
mb1 = mb1ofm1%nBlocksMB;
ofm1 = mb1ofm1/nBlocksMB;
}
if ( ((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) == LIBXSMM_DNN_FC_ELTW_FUSE_BIAS) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU) ||
((cfg.fuse_type & LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) == LIBXSMM_DNN_FC_ELTW_FUSE_RELU_WITH_MASK) ) {
Expand Down Expand Up @@ -4625,6 +4714,9 @@ LIBXSMM_API void libxsmm_dnn_fc_bwd_exec_bf8( libxsmm_dnn_fc_bwd_config cfg, co

LIBXSMM_API void destroy_libxsmm_dnn_fc_fwd(libxsmm_dnn_fc_fwd_config* cfg) {
libxsmm_barrier_destroy(cfg->barrier);
if (cfg->fwd_sfc != 0) {
libxsmm_free(cfg->fwd_sfc_map);
}
}

LIBXSMM_API void destroy_libxsmm_dnn_fc_bwd(libxsmm_dnn_fc_bwd_config* cfg) {
Expand Down
Loading

0 comments on commit 0423291

Please sign in to comment.