| /dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/gpu/ocl/ |
| H A D | simple_reorder.cpp | 83 int ndims = src_mdw.ndims(); in is_alt_faster_than_ref()
89 if (src_mdw.dims()[last] < 8) { return false; } in is_alt_faster_than_ref()
168 bool fits_3ch(const memory_desc_wrapper &src_mdw, in fits_3ch() argument
276 && src_mdw.dims()[inner_dim] == src_mdw.padded_dims()[inner_dim]) { in select_kernel()
310 if ((src_mdw.matches_one_of_tag(abcd) || src_mdw.matches_one_of_tag(acdb)) in select_kernel()
351 size_t ndims = src_mdw.ndims(); in alt_defines()
354 auto sdim = src_mdw.dims(); in alt_defines()
375 auto sdim = src_mdw.dims(); in alt_gen()
377 size_t last = src_mdw.ndims() - 1; in alt_gen()
408 conf.ndims = src_mdw.ndims(); in init_conf()
[all …]
|
| H A D | simple_concat.cpp | 65 const memory_desc_wrapper src_mdw(pd->src_md(i)); in init_conf_common() local
68 if (src_mdw.padded_dims()[pd->concat_dim()] in init_conf_common()
69 != src_mdw.dims()[pd->concat_dim()]) { in init_conf_common()
76 if (src_mdw.data_type() != dst_mdw.data_type()) in init_conf_common()
82 if (!is_same_axis_order(dst_mdw, src_mdw)) { in init_conf_common()
86 if (!src_mdw.is_dense()) return status::unimplemented; in init_conf_common()
88 const auto &src_blk = src_mdw.blocking_desc(); in init_conf_common()
92 ? src_mdw.nelems(true) in init_conf_common()
93 : (src_mdw.dims()[pd->concat_dim()] != 0) in init_conf_common()
|
| H A D | ref_layer_normalization.cpp | 31 memory_desc_wrapper src_mdw(pd->src_md()); in init_conf_common() local
35 int ndims = src_mdw.ndims(); in init_conf_common()
37 conf.data_type = src_mdw.data_type(); in init_conf_common()
41 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf_common()
52 if (src_mdw.is_blocking_desc()) { in init_conf_common()
53 c_block = src_mdw.blocking_desc() in init_conf_common()
54 .inner_blks[src_mdw.blocking_desc().inner_nblks - 1]; in init_conf_common()
60 pd->is_fwd() ? dst_mdw.md_ : src_mdw.md_); in init_conf_common()
61 auto &dims = (pd->is_fwd() ? src_mdw : dst_mdw).dims(); in init_conf_common()
64 && (c_block == 1 || src_mdw.is_dense())) { in init_conf_common()
|
| H A D | ref_reduction.cpp | 32 const memory_desc_wrapper src_mdw(pd->src_md()); in init_conf() local
35 const int ndims = src_mdw.ndims(); in init_conf()
36 const auto src_dims = src_mdw.md_->dims; in init_conf()
42 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf()
45 conf.src_type = src_mdw.data_type(); in init_conf()
49 conf.dispatch = compute_engine->create_dispatch(src_mdw.md_); in init_conf()
77 set_offsets(src_mdw, conf.off.src_off); in init_conf()
|
| H A D | gen9_reduction.cpp | 90 const memory_desc_wrapper src_mdw(pd->src_md()); in init_conf() local
93 const int ndims = src_mdw.ndims(); in init_conf()
94 const dnnl_dim_t *src_dims = src_mdw.md_->dims; in init_conf()
101 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf()
104 conf.src_type = src_mdw.data_type(); in init_conf()
108 conf.dispatch = compute_engine->create_dispatch(src_mdw.md_); in init_conf()
119 if (!(is_c_blocked_by(src_mdw, 16) || is_c_blocked_by(src_mdw, 32))) in init_conf()
124 std::tie(src_n_block_size, src_c_block_size) = get_n_c_block_sizes(src_mdw); in init_conf()
128 || src_mdw.blocking_desc().inner_nblks in init_conf()
134 if ((conf.n_block_size == 1 && src_mdw.blocking_desc().inner_nblks > 1) in init_conf()
[all …]
|
| H A D | gen9_pooling.cpp | 28 const memory_desc_wrapper src_mdw(pd->invariant_src_md()); in init_conf_common() local
42 if (!is_c_blocked_by(src_mdw, 16) && !is_c_blocked_by(src_mdw, 32) in init_conf_common()
43 && !is_c_dense(src_mdw)) in init_conf_common()
51 auto &src_blk = src_mdw.blocking_desc(); in init_conf_common()
65 set_offsets(src_mdw, off.src_off); in init_conf_common()
100 conf.is_backward ? src_mdw.md_ : dst_mdw.md_); in init_conf_common()
|
| H A D | cross_engine_reorder.cpp | 51 memory_desc_wrapper src_mdw(src_md()); in init() local
55 do_reorder_ = with_sum_ab || src_mdw != dst_mdw; in init()
129 memory_desc_wrapper src_mdw(pd()->src_md()); in execute() local
132 src_mdw.size()); in execute()
|
| H A D | xe_lp_x8s8x_1x1_convolution.cpp | 29 const memory_desc_wrapper src_mdw(src_md()); in init_conf() local
40 = src_mdw.matches_one_of_tag(nwc, nhwc, ndhwc) != format_tag::undef in init_conf()
49 conf.src_data_type = src_mdw.data_type(); in init_conf()
97 conf.src_data_type = src_mdw.data_type(); in init_conf()
134 conf.src_tag = src_mdw.format_kind() == format_kind::any in init_conf()
136 : src_mdw.matches_one_of_tag(src_tag); in init_conf()
|
| /dports/math/onednn/oneDNN-2.5.1/src/gpu/ocl/ |
| H A D | custom_reorder.cpp | 100 int ndims = src_mdw.ndims(); in is_alt_faster_than_ref()
106 if (src_mdw.dims()[last] < 8) { return false; } in is_alt_faster_than_ref()
277 bool fits_3ch(const memory_desc_wrapper &src_mdw, in fits_3ch() argument
403 && src_mdw.dims()[inner_dim] == src_mdw.padded_dims()[inner_dim]) { in select_kernel()
437 if ((src_mdw.matches_one_of_tag(abcd, acdb)) in select_kernel()
479 size_t ndims = src_mdw.ndims(); in alt_defines()
482 auto sdim = src_mdw.dims(); in alt_defines()
503 auto sdim = src_mdw.dims(); in alt_gen()
505 size_t last = src_mdw.ndims() - 1; in alt_gen()
536 conf.ndims = src_mdw.ndims(); in init_conf()
[all …]
|
| H A D | ref_layer_normalization.cpp | 33 memory_desc_wrapper src_mdw(pd->src_md()); in init_conf_common() local
37 int ndims = src_mdw.ndims(); in init_conf_common()
39 conf.data_type = src_mdw.data_type(); in init_conf_common()
43 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf_common()
54 if (src_mdw.blocking_desc().inner_nblks > 0) { in init_conf_common()
55 c_block = src_mdw.blocking_desc() in init_conf_common()
63 pd->is_fwd() ? dst_mdw.md_ : src_mdw.md_); in init_conf_common()
70 || src_mdw.is_dense())) { in init_conf_common()
108 while (src_mdw.blocking_desc().inner_nblks > 0 in init_conf_common()
131 if (src_mdw.blocking_desc().inner_nblks == 2 in init_conf_common()
[all …]
|
| H A D | simple_concat.cpp | 66 const memory_desc_wrapper src_mdw(pd->src_md(i)); in init_conf_common() local
69 if (src_mdw.padded_dims()[pd->concat_dim()] in init_conf_common()
70 != src_mdw.dims()[pd->concat_dim()]) { in init_conf_common()
77 if (src_mdw.data_type() != dst_mdw.data_type()) in init_conf_common()
83 if (!is_same_axis_order(dst_mdw, src_mdw)) { in init_conf_common()
87 if (!src_mdw.is_dense()) return status::unimplemented; in init_conf_common()
89 const auto &src_blk = src_mdw.blocking_desc(); in init_conf_common()
93 ? src_mdw.nelems(true) in init_conf_common()
94 : (src_mdw.dims()[pd->concat_dim()] != 0) in init_conf_common()
|
| H A D | ref_reduction.cpp | 32 const memory_desc_wrapper src_mdw(pd->src_md()); in init_conf() local
35 const int ndims = src_mdw.ndims(); in init_conf()
36 const auto src_dims = src_mdw.md_->dims; in init_conf()
42 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf()
45 conf.src_type = src_mdw.data_type(); in init_conf()
49 conf.dispatch = compute_engine->create_dispatch(src_mdw.md_); in init_conf()
77 set_offsets(src_mdw, conf.off.src_off); in init_conf()
|
| H A D | gen9_reduction.cpp | 90 const memory_desc_wrapper src_mdw(pd->src_md()); in init_conf() local
93 const int ndims = src_mdw.ndims(); in init_conf()
94 const dnnl_dim_t *src_dims = src_mdw.md_->dims; in init_conf()
101 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf()
104 conf.src_type = src_mdw.data_type(); in init_conf()
108 conf.dispatch = compute_engine->create_dispatch(src_mdw.md_); in init_conf()
119 if (!(is_c_blocked_by(src_mdw, 16) || is_c_blocked_by(src_mdw, 32))) in init_conf()
124 std::tie(src_n_block_size, src_c_block_size) = get_n_c_block_sizes(src_mdw); in init_conf()
128 || src_mdw.blocking_desc().inner_nblks in init_conf()
134 if ((conf.n_block_size == 1 && src_mdw.blocking_desc().inner_nblks > 1) in init_conf()
[all …]
|
| H A D | ref_reorder.cpp | 35 const memory_desc_wrapper src_mdw(src_md()); in init_conf() local
38 conf.src_md_info = memory_desc_info_t::create(src_mdw); in init_conf()
49 conf.has_padding = !src_mdw.is_dense() || !dst_mdw.is_dense(); in init_conf()
50 conf.ndims = src_mdw.ndims(); in init_conf()
84 const memory_desc_wrapper src_mdw(src_md()); in init_kernel_ctx() local
|
| H A D | gen9_pooling.cpp | 28 const memory_desc_wrapper src_mdw(pd->invariant_src_md()); in init_conf_common() local
42 if (!is_c_blocked_by(src_mdw, 16) && !is_c_blocked_by(src_mdw, 32) in init_conf_common()
43 && !is_c_dense(src_mdw)) in init_conf_common()
51 auto &src_blk = src_mdw.blocking_desc(); in init_conf_common()
65 set_offsets(src_mdw, off.src_off); in init_conf_common()
100 conf.is_backward ? src_mdw.md_ : dst_mdw.md_); in init_conf_common()
|
| H A D | gen9_concat.cpp | 78 const memory_desc_wrapper src_mdw(pd->src_md(i)); in can_use_sub_group_size() local
80 && src_mdw.md_->dims[conf.concat_axis] % sub_group_size == 0; in can_use_sub_group_size()
83 && get_dim_block(src_mdw, c_idx) % sub_group_size == 0; in can_use_sub_group_size()
86 = layouts_compatible && is_dim_dense(src_mdw, c_idx); in can_use_sub_group_size()
123 const memory_desc_wrapper src_mdw(pd->src_md(i)); in init_conf() local
124 concat_axis_end += src_mdw.md_->dims[conf.concat_axis]; in init_conf()
|
| H A D | cross_engine_reorder.cpp | 51 memory_desc_wrapper src_mdw(src_md()); in init() local
55 do_reorder_ = with_sum_ab || src_mdw != dst_mdw; in init()
129 memory_desc_wrapper src_mdw(pd()->src_md()); in execute() local
132 src_mdw.size()); in execute()
|
| H A D | gen9_global_pooling.cpp | 51 const memory_desc_wrapper src_mdw(pd->invariant_src_md()); in init_conf_common() local
53 const auto &padded_src_dims = src_mdw.padded_dims(); in init_conf_common()
61 set_offsets(src_mdw, off.src_off); in init_conf_common()
70 conf.dispatch = compute_engine->create_dispatch(src_mdw.md_); in init_conf_common()
|
| /dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/gpu/ocl/rnn/ |
| H A D | rnn_reorders.cpp | 28 const memory_desc_wrapper src_mdw(src_md()); in init_conf() local
36 conf.do_reorder = src_mdw != dst_mdw; in init_conf()
37 conf.has_padding = !src_mdw.is_dense() || !dst_mdw.is_dense(); in init_conf()
38 conf.ndims = src_mdw.ndims(); in init_conf()
56 const auto &input_dims = src_mdw.dims(); in init_conf()
71 const memory_desc_wrapper src_mdw(src_md()); in init_kernel_ctx() local
74 auto input_type = src_mdw.data_type(); in init_kernel_ctx()
99 set_offsets(kernel_ctx, src_mdw, "SRC"); in init_kernel_ctx()
102 const auto &in_dims = src_mdw.dims(); in init_kernel_ctx()
108 (d < src_mdw.ndims()) ? in_dims[d] : 1); in init_kernel_ctx()
[all …]
|
| /dports/math/onednn/oneDNN-2.5.1/src/gpu/ocl/rnn/ |
| H A D | rnn_reorders.cpp | 28 const memory_desc_wrapper src_mdw(src_md()); in init_conf() local
36 conf.do_reorder = src_mdw != dst_mdw; in init_conf()
37 conf.has_padding = !src_mdw.is_dense() || !dst_mdw.is_dense(); in init_conf()
38 conf.ndims = src_mdw.ndims(); in init_conf()
56 const auto &input_dims = src_mdw.dims(); in init_conf()
71 const memory_desc_wrapper src_mdw(src_md()); in init_kernel_ctx() local
74 auto input_type = src_mdw.data_type(); in init_kernel_ctx()
99 set_offsets(kernel_ctx, src_mdw, "SRC"); in init_kernel_ctx()
102 const auto &in_dims = src_mdw.dims(); in init_kernel_ctx()
108 (d < src_mdw.ndims()) ? in_dims[d] : 1); in init_kernel_ctx()
[all …]
|
| /dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/cpu/ |
| H A D | ref_reduction.cpp | 109 const memory_desc_wrapper src_mdw(pd()->src_md()); in execute_ref() local
112 const int ndims = src_mdw.ndims(); in execute_ref()
113 const auto &src_dims = src_mdw.dims(); in execute_ref()
136 const dim_t src_idle_off = src_mdw.off_v(idle_pos); in execute_ref()
141 const dim_t src_reduce_off = src_mdw.off_v(reduce_pos); in execute_ref()
|
| /dports/math/onednn/oneDNN-2.5.1/src/cpu/ |
| H A D | ref_reduction.cpp | 109 const memory_desc_wrapper src_mdw(pd()->src_md()); in execute_ref() local
112 const int ndims = src_mdw.ndims(); in execute_ref()
113 const auto &src_dims = src_mdw.dims(); in execute_ref()
136 const dim_t src_idle_off = src_mdw.off_v(idle_pos); in execute_ref()
141 const dim_t src_reduce_off = src_mdw.off_v(reduce_pos); in execute_ref()
|
| /dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/src/gpu/ |
| H A D | primitive_conf.hpp | 682 const memory_desc_wrapper src_mdw(src_md); in set_default_pool_conf() local
685 const auto &src_dims = src_mdw.dims(); in set_default_pool_conf()
688 int ndims = src_mdw.ndims(); in set_default_pool_conf()
694 conf.mb_padded = src_mdw.padded_dims()[0]; in set_default_pool_conf()
695 conf.c_padded = src_mdw.padded_dims()[1]; in set_default_pool_conf()
724 conf.src_dt = src_mdw.data_type(); in set_default_pool_conf()
741 const memory_desc_wrapper src_mdw(&src_md); in set_default_conf() local
747 int ndims = src_mdw.ndims(); in set_default_conf()
754 conf.mb = src_mdw.dims()[0]; in set_default_conf()
759 conf.iw = src_mdw.dims()[ndims - 1]; in set_default_conf()
[all …]
|
| /dports/misc/mxnet/incubator-mxnet-1.9.0/3rdparty/mkldnn/tests/gtests/ |
| H A D | test_layer_normalization.cpp | 307 const dnnl::impl::memory_desc_wrapper src_mdw(src_d.data); in check_lnorm_fwd() local
312 const auto ndims = src_mdw.ndims(); in check_lnorm_fwd()
313 const auto C = src_mdw.dims()[ndims - 1]; in check_lnorm_fwd()
356 ref_mean += src_data[src_mdw.off_l(n * C + c)]; in check_lnorm_fwd()
401 * ((float)src_data[src_mdw.off_l(n * C + c)] in check_lnorm_fwd()
443 const dnnl::impl::memory_desc_wrapper src_mdw(src_d.data); in check_lnorm_bwd() local
452 const auto ndims = src_mdw.ndims(); in check_lnorm_bwd()
453 const auto C = src_mdw.dims()[ndims - 1]; in check_lnorm_bwd()
519 ref_diff_gamma += (src_data[src_mdw.off_l(n * C + c)] in check_lnorm_bwd()
567 * (src_data[src_mdw.off_l(n * C + c)] in check_lnorm_bwd()
[all …]
|
| /dports/math/onednn/oneDNN-2.5.1/tests/gtests/ |
| H A D | test_layer_normalization.cpp | 307 const dnnl::impl::memory_desc_wrapper src_mdw(src_d.data); in check_lnorm_fwd() local
312 const auto ndims = src_mdw.ndims(); in check_lnorm_fwd()
313 const auto C = src_mdw.dims()[ndims - 1]; in check_lnorm_fwd()
356 ref_mean += src_data[src_mdw.off_l(n * C + c)]; in check_lnorm_fwd()
401 * ((float)src_data[src_mdw.off_l(n * C + c)] in check_lnorm_fwd()
443 const dnnl::impl::memory_desc_wrapper src_mdw(src_d.data); in check_lnorm_bwd() local
452 const auto ndims = src_mdw.ndims(); in check_lnorm_bwd()
453 const auto C = src_mdw.dims()[ndims - 1]; in check_lnorm_bwd()
519 ref_diff_gamma += (src_data[src_mdw.off_l(n * C + c)] in check_lnorm_bwd()
567 * (src_data[src_mdw.off_l(n * C + c)] in check_lnorm_bwd()
[all …]
|