mindspore 2.3.0__cp39-none-any.whl → 2.3.0rc2__cp39-none-any.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Potentially problematic release.


This version of mindspore might be problematic. Click here for more details.

Files changed (423) hide show
  1. mindspore/.commit_id +1 -1
  2. mindspore/Third_Party_Open_Source_Software_Notice +0 -1512
  3. mindspore/__init__.py +1 -2
  4. mindspore/_c_dataengine.cpython-39-aarch64-linux-gnu.so +0 -0
  5. mindspore/_c_expression.cpython-39-aarch64-linux-gnu.so +0 -0
  6. mindspore/_c_mindrecord.cpython-39-aarch64-linux-gnu.so +0 -0
  7. mindspore/_checkparam.py +25 -5
  8. mindspore/_extends/graph_kernel/model/graph_parallel.py +1 -1
  9. mindspore/_extends/parse/__init__.py +2 -2
  10. mindspore/_extends/parse/compile_config.py +0 -29
  11. mindspore/_extends/parse/namespace.py +2 -2
  12. mindspore/_extends/parse/parser.py +5 -21
  13. mindspore/_extends/parse/resources.py +7 -5
  14. mindspore/_extends/parse/standard_method.py +59 -40
  15. mindspore/_mindspore_offline_debug.cpython-39-aarch64-linux-gnu.so +0 -0
  16. mindspore/amp.py +5 -26
  17. mindspore/bin/cache_admin +0 -0
  18. mindspore/bin/cache_server +0 -0
  19. mindspore/boost/adasum.py +1 -1
  20. mindspore/boost/base.py +1 -1
  21. mindspore/boost/boost_cell_wrapper.py +1 -1
  22. mindspore/boost/grad_freeze.py +2 -2
  23. mindspore/boost/less_batch_normalization.py +6 -9
  24. mindspore/common/__init__.py +1 -8
  25. mindspore/common/_register_for_tensor.py +9 -8
  26. mindspore/common/api.py +65 -275
  27. mindspore/common/dtype.py +4 -8
  28. mindspore/common/dump.py +5 -2
  29. mindspore/common/jit_config.py +1 -1
  30. mindspore/common/lazy_inline.py +2 -14
  31. mindspore/common/parameter.py +15 -14
  32. mindspore/common/recompute.py +5 -20
  33. mindspore/common/sparse_tensor.py +6 -21
  34. mindspore/common/tensor.py +52 -100
  35. mindspore/communication/__init__.py +11 -6
  36. mindspore/communication/management.py +94 -92
  37. mindspore/context.py +18 -180
  38. mindspore/dataset/engine/datasets.py +46 -69
  39. mindspore/dataset/engine/datasets_user_defined.py +53 -72
  40. mindspore/dataset/engine/datasets_vision.py +2 -2
  41. mindspore/dataset/engine/queue.py +38 -56
  42. mindspore/dataset/engine/validators.py +5 -11
  43. mindspore/dataset/vision/__init__.py +5 -5
  44. mindspore/dataset/vision/c_transforms.py +5 -5
  45. mindspore/dataset/vision/py_transforms_util.py +1 -1
  46. mindspore/dataset/vision/transforms.py +46 -591
  47. mindspore/dataset/vision/utils.py +1 -121
  48. mindspore/dataset/vision/validators.py +3 -9
  49. mindspore/hal/__init__.py +1 -7
  50. mindspore/hal/device.py +1 -1
  51. mindspore/include/api/model.h +0 -3
  52. mindspore/include/dataset/vision.h +2 -54
  53. mindspore/include/mindapi/base/types.h +0 -1
  54. mindspore/lib/libdnnl.so.2 +0 -0
  55. mindspore/lib/libmindspore.so +0 -0
  56. mindspore/lib/libmindspore_backend.so +0 -0
  57. mindspore/lib/libmindspore_common.so +0 -0
  58. mindspore/lib/libmindspore_core.so +0 -0
  59. mindspore/lib/libmindspore_glog.so.0 +0 -0
  60. mindspore/lib/libmindspore_gpr.so.15 +0 -0
  61. mindspore/lib/libmindspore_grpc++.so.1 +0 -0
  62. mindspore/lib/libmindspore_grpc.so.15 +0 -0
  63. mindspore/lib/libmindspore_shared_lib.so +0 -0
  64. mindspore/lib/libmpi_adapter.so +0 -0
  65. mindspore/lib/libmpi_collective.so +0 -0
  66. mindspore/lib/libnnacl.so +0 -0
  67. mindspore/lib/libopencv_core.so.4.5 +0 -0
  68. mindspore/lib/libps_cache.so +0 -0
  69. mindspore/lib/plugin/ascend/custom_aicore_ops/op_impl/ai_core/tbe/config/ascend310p/aic-ascend310p-ops-info.json +0 -35
  70. mindspore/lib/plugin/ascend/custom_aicore_ops/op_impl/ai_core/tbe/custom_aicore_ops_impl/kv_cache_mgr.py +0 -2
  71. mindspore/lib/plugin/ascend/custom_aicore_ops/op_impl/vector_core/tbe/custom_aicore_ops_impl/kv_cache_mgr.py +0 -2
  72. mindspore/lib/plugin/ascend/custom_aicpu_ops/op_impl/cpu/aicpu_kernel/impl/libcust_cpu_kernels.so +0 -0
  73. mindspore/lib/plugin/ascend/custom_aicpu_ops/op_impl/cpu/config/cust_aicpu_kernel.json +0 -72
  74. mindspore/lib/plugin/ascend/custom_aicpu_ops/op_proto/libcust_op_proto.so +0 -0
  75. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_api/include/{aclnn_all_finite.h → aclnn_add_custom.h} +11 -9
  76. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_api/include/aclnn_decoder_kv_cache.h +1 -1
  77. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_api/include/aclnn_prompt_kv_cache.h +1 -1
  78. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_api/lib/libcust_opapi.so +0 -0
  79. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/config/ascend310p/aic-ascend310p-ops-info.json +12 -184
  80. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/config/ascend910/aic-ascend910-ops-info.json +15 -7
  81. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/config/ascend910b/aic-ascend910b-ops-info.json +15 -7
  82. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/add_custom.cpp +81 -0
  83. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/add_custom.py +134 -0
  84. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/decoder_kv_cache.py +31 -77
  85. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/prompt_kv_cache.py +31 -77
  86. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/op_tiling/lib/linux/aarch64/libcust_opmaster_rt2.0.so +0 -0
  87. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/op_tiling/liboptiling.so +0 -0
  88. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_proto/inc/op_proto.h +5 -4
  89. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_proto/lib/linux/aarch64/libcust_opsproto_rt2.0.so +0 -0
  90. mindspore/lib/plugin/ascend/libascend_collective.so +0 -0
  91. mindspore/lib/plugin/ascend/libdvpp_utils.so +0 -0
  92. mindspore/lib/plugin/ascend/libhccl_plugin.so +0 -0
  93. mindspore/lib/plugin/ascend/liblowlatency_collective.so +0 -0
  94. mindspore/lib/plugin/ascend/libmindspore_cpu_kernels.so +0 -0
  95. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/bin/DeviceBin +0 -0
  96. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/bin/PkgInspect +0 -0
  97. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/bin/op_man +0 -0
  98. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/device/ascend910b/bin/ascend910b.bin +286 -275
  99. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/host/libasdops_cann_host.so +0 -0
  100. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/host/libasdops_host.so +0 -0
  101. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/lib/libasdops.so +0 -0
  102. mindspore/lib/plugin/ascend/ms_kernels_internal/asdops/lib/libasdops_static.a +0 -0
  103. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/add/add_impl.h +0 -1
  104. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/apply_rotary_pos_emb/apply_rotary_pos_emb_impl.h +0 -1
  105. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/asdop/asd_op_impl.h +0 -3
  106. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/backend_param.h +0 -5
  107. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/cast/cast_tiling.h +45 -1
  108. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/compare/compare_impl.h +0 -1
  109. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/flash_attention_score_impl.h +4 -8
  110. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/flash_attention_score_tiling.h +4 -11
  111. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/flash_attention_score/kernel/flash_attention_score_mix_hwsync.h +0 -18
  112. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/internal_kernel.h +0 -6
  113. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/internal_rtbackend.h +75 -1
  114. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul/kernel/matmul.h +5 -5
  115. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul/matmul_impl.h +3 -18
  116. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/pp_matmul_common_tiling.h +5 -5
  117. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/pp_matmul_info.h +2 -2
  118. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_common/tiling_data.h +3 -36
  119. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_stridedslice/kernel/matmul_stridedslice_fusion.h +2 -2
  120. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/matmul_stridedslice/matmul_stridedslice_fusion_impl.h +4 -22
  121. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/op_param.h +2 -16
  122. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/paged_attention/kernel/paged_attention_mix_hwsync.h +3 -1
  123. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/paged_attention/paged_attention_impl.h +4 -5
  124. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/paged_attention/paged_attention_tiling.h +4 -9
  125. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/attention_param.h +2 -5
  126. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/matmul_ext_param.h +0 -1
  127. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/matmul_qkv_param.h +4 -10
  128. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/sub_param.h +12 -0
  129. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/rms_norm/rms_norm_impl.h +0 -1
  130. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/sub/sub_impl.h +0 -1
  131. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/tune_repo/matmul_table.h +1 -1
  132. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/backend.h +2 -10
  133. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/elewise_utils.h +1 -5
  134. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/log/log.h +0 -1
  135. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/log/log_tiling.h +0 -17
  136. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/math.h +7 -2
  137. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libAdd_impl.so +0 -0
  138. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libSub_impl.so +0 -0
  139. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libadd_layernorm_impl.so +0 -0
  140. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libadd_rms_norm_impl.so +0 -0
  141. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libapply_rotary_pos_emb_impl.so +0 -0
  142. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libcast_impl.so +0 -0
  143. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libgelu_impl.so +0 -0
  144. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libmatmul_impl.so +0 -0
  145. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libmatmul_stridedslice_fusion_impl.so +0 -0
  146. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libms_kernels_internal.so +0 -0
  147. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libnot_equal_impl.so +0 -0
  148. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/libreshape_and_cache_impl.so +0 -0
  149. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/lib/librms_norm_impl.so +0 -0
  150. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_bf16_bnsd_full_mix.o +0 -0
  151. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_bf16_bnsd_tri_mix.o +0 -0
  152. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_bf16_bsh_full_mix.o +0 -0
  153. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_bf16_bsh_tri_mix.o +0 -0
  154. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_fp16_bnsd_full_mix.o +0 -0
  155. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_fp16_bnsd_tri_mix.o +0 -0
  156. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_fp16_bsh_full_mix.o +0 -0
  157. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/flash_attention_score_fp16_bsh_tri_mix.o +0 -0
  158. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/paged_attention_bf16_bnsd_full_mix.o +0 -0
  159. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/paged_attention_bf16_bsh_full_mix.o +0 -0
  160. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/paged_attention_fp16_bnsd_full_mix.o +0 -0
  161. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/BSAttention/paged_attention_fp16_bsh_full_mix.o +0 -0
  162. mindspore/lib/plugin/ascend/ms_kernels_internal/lccl/lib/liblcal.so +0 -0
  163. mindspore/lib/plugin/ascend/ms_kernels_internal/lccl/lib/liblccl_wrapper.so +0 -0
  164. mindspore/lib/plugin/libmindspore_ascend.so.2 +0 -0
  165. mindspore/mindrecord/filewriter.py +2 -2
  166. mindspore/mint/__init__.py +40 -720
  167. mindspore/mint/nn/__init__.py +7 -89
  168. mindspore/mint/nn/functional.py +16 -165
  169. mindspore/mint/optim/adamw.py +16 -15
  170. mindspore/nn/__init__.py +2 -0
  171. mindspore/nn/cell.py +98 -97
  172. mindspore/nn/extend/basic.py +2 -2
  173. mindspore/nn/extend/embedding.py +1 -1
  174. mindspore/nn/extend/layer/normalization.py +5 -7
  175. mindspore/nn/generator.py +297 -0
  176. mindspore/nn/layer/activation.py +3 -4
  177. mindspore/nn/layer/basic.py +16 -79
  178. mindspore/nn/layer/conv.py +8 -17
  179. mindspore/nn/layer/embedding.py +4 -1
  180. mindspore/nn/layer/math.py +1 -1
  181. mindspore/nn/layer/normalization.py +1 -1
  182. mindspore/nn/layer/pooling.py +0 -5
  183. mindspore/nn/layer/rnn_cells.py +2 -2
  184. mindspore/nn/loss/loss.py +19 -19
  185. mindspore/nn/optim/adasum.py +1 -1
  186. mindspore/nn/optim/sgd.py +2 -3
  187. mindspore/nn/probability/distribution/exponential.py +1 -1
  188. mindspore/nn/probability/distribution/geometric.py +1 -1
  189. mindspore/nn/probability/distribution/logistic.py +1 -1
  190. mindspore/nn/wrap/cell_wrapper.py +1 -25
  191. mindspore/nn/wrap/loss_scale.py +1 -24
  192. mindspore/numpy/array_ops.py +1 -5
  193. mindspore/numpy/dtypes.py +3 -3
  194. mindspore/numpy/math_ops.py +8 -8
  195. mindspore/ops/__init__.py +1 -1
  196. mindspore/ops/_grad_experimental/grad_comm_ops.py +16 -75
  197. mindspore/ops/_vmap/vmap_array_ops.py +0 -27
  198. mindspore/ops/_vmap/vmap_math_ops.py +1 -29
  199. mindspore/ops/_vmap/vmap_nn_ops.py +18 -19
  200. mindspore/ops/auto_generate/cpp_create_prim_instance_helper.py +8 -34
  201. mindspore/ops/auto_generate/gen_arg_dtype_cast.py +9 -2
  202. mindspore/ops/auto_generate/gen_arg_handler.py +0 -26
  203. mindspore/ops/auto_generate/gen_extend_func.py +27 -603
  204. mindspore/ops/auto_generate/gen_ops_def.py +203 -993
  205. mindspore/ops/auto_generate/gen_ops_prim.py +402 -1946
  206. mindspore/ops/auto_generate/pyboost_inner_prim.py +20 -90
  207. mindspore/ops/composite/base.py +6 -3
  208. mindspore/ops/composite/math_ops.py +1 -1
  209. mindspore/ops/composite/multitype_ops/_compile_utils.py +17 -24
  210. mindspore/ops/composite/multitype_ops/_constexpr_utils.py +1 -1
  211. mindspore/ops/extend/__init__.py +3 -2
  212. mindspore/ops/extend/array_func.py +51 -10
  213. mindspore/ops/extend/nn_func.py +78 -2
  214. mindspore/ops/function/__init__.py +13 -8
  215. mindspore/ops/function/array_func.py +179 -455
  216. mindspore/ops/function/clip_func.py +1 -1
  217. mindspore/ops/function/grad/grad_func.py +3 -3
  218. mindspore/ops/function/math_func.py +103 -117
  219. mindspore/ops/function/nn_func.py +163 -275
  220. mindspore/ops/function/other_func.py +2 -2
  221. mindspore/ops/function/random_func.py +69 -202
  222. mindspore/ops/function/sparse_func.py +4 -4
  223. mindspore/ops/functional.py +327 -332
  224. mindspore/ops/operations/__init__.py +3 -13
  225. mindspore/ops/operations/_grad_ops.py +27 -3
  226. mindspore/ops/operations/_inner_ops.py +356 -53
  227. mindspore/ops/operations/_rl_inner_ops.py +2 -2
  228. mindspore/ops/operations/_tensor_array.py +8 -8
  229. mindspore/ops/operations/array_ops.py +65 -82
  230. mindspore/ops/operations/comm_ops.py +93 -784
  231. mindspore/ops/operations/custom_ops.py +28 -51
  232. mindspore/ops/operations/debug_ops.py +4 -4
  233. mindspore/ops/operations/inner_ops.py +2 -2
  234. mindspore/ops/operations/manually_defined/ops_def.py +4 -304
  235. mindspore/ops/operations/math_ops.py +50 -3
  236. mindspore/ops/operations/nn_ops.py +247 -14
  237. mindspore/ops/operations/other_ops.py +3 -3
  238. mindspore/ops/operations/random_ops.py +1 -1
  239. mindspore/ops/operations/sparse_ops.py +1 -1
  240. mindspore/ops/primitive.py +8 -9
  241. mindspore/ops/silent_check.py +5 -5
  242. mindspore/ops_generate/arg_dtype_cast.py +9 -2
  243. mindspore/ops_generate/arg_handler.py +0 -26
  244. mindspore/ops_generate/gen_aclnn_implement.py +4 -1
  245. mindspore/ops_generate/gen_ops.py +4 -26
  246. mindspore/ops_generate/gen_pyboost_func.py +12 -41
  247. mindspore/ops_generate/gen_utils.py +0 -21
  248. mindspore/ops_generate/pyboost_utils.py +2 -7
  249. mindspore/ops_generate/template.py +0 -1
  250. mindspore/parallel/_auto_parallel_context.py +1 -21
  251. mindspore/parallel/_tensor.py +5 -0
  252. mindspore/parallel/_transformer/transformer.py +1 -1
  253. mindspore/parallel/_utils.py +1 -15
  254. mindspore/parallel/algo_parameter_config.py +3 -1
  255. mindspore/parallel/checkpoint_transform.py +9 -12
  256. mindspore/parallel/cluster/process_entity/_api.py +29 -28
  257. mindspore/parallel/cluster/process_entity/_utils.py +3 -13
  258. mindspore/parallel/cluster/run.py +16 -13
  259. mindspore/parallel/parameter_broadcast.py +2 -2
  260. mindspore/parallel/shard.py +17 -31
  261. mindspore/profiler/__init__.py +2 -3
  262. mindspore/profiler/common/util.py +2 -107
  263. mindspore/profiler/envprofiling.py +1 -1
  264. mindspore/profiler/parser/ascend_analysis/constant.py +21 -8
  265. mindspore/profiler/parser/ascend_analysis/file_manager.py +0 -82
  266. mindspore/profiler/parser/ascend_analysis/function_event.py +28 -43
  267. mindspore/profiler/parser/ascend_analysis/fwk_cann_parser.py +27 -49
  268. mindspore/profiler/parser/ascend_analysis/fwk_file_parser.py +10 -15
  269. mindspore/profiler/parser/ascend_analysis/msprof_timeline_parser.py +20 -25
  270. mindspore/profiler/parser/ascend_analysis/profiler_info_parser.py +5 -5
  271. mindspore/profiler/parser/ascend_analysis/trace_event_manager.py +1 -10
  272. mindspore/profiler/parser/ascend_hccl_generator.py +1 -4
  273. mindspore/profiler/parser/ascend_msprof_exporter.py +22 -43
  274. mindspore/profiler/parser/ascend_timeline_generator.py +5 -7
  275. mindspore/profiler/parser/minddata_parser.py +3 -72
  276. mindspore/profiler/profiling.py +59 -176
  277. mindspore/rewrite/api/node.py +1 -1
  278. mindspore/rewrite/common/namespace.py +5 -5
  279. mindspore/rewrite/parsers/assign_parser.py +0 -2
  280. mindspore/rewrite/parsers/class_def_parser.py +4 -8
  281. mindspore/run_check/_check_version.py +1 -1
  282. mindspore/scipy/fft.py +3 -1
  283. mindspore/scipy/linalg.py +3 -2
  284. mindspore/scipy/ops.py +3 -5
  285. mindspore/scipy/optimize/__init__.py +2 -2
  286. mindspore/train/__init__.py +4 -4
  287. mindspore/train/anf_ir_pb2.py +2 -8
  288. mindspore/train/callback/__init__.py +2 -5
  289. mindspore/train/callback/_backup_and_restore.py +2 -2
  290. mindspore/train/callback/_checkpoint.py +16 -104
  291. mindspore/train/callback/_landscape.py +1 -1
  292. mindspore/train/callback/_time_monitor.py +1 -1
  293. mindspore/train/data_sink.py +4 -5
  294. mindspore/train/dataset_helper.py +20 -45
  295. mindspore/train/model.py +38 -266
  296. mindspore/train/serialization.py +105 -256
  297. mindspore/train/summary/_summary_adapter.py +1 -1
  298. mindspore/version.py +1 -1
  299. {mindspore-2.3.0.dist-info → mindspore-2.3.0rc2.dist-info}/METADATA +2 -2
  300. {mindspore-2.3.0.dist-info → mindspore-2.3.0rc2.dist-info}/RECORD +303 -420
  301. mindspore/_extends/pijit/__init__.py +0 -23
  302. mindspore/_extends/pijit/pijit_func_white_list.py +0 -343
  303. mindspore/common/file_system.py +0 -48
  304. mindspore/common/generator.py +0 -260
  305. mindspore/common/no_inline.py +0 -54
  306. mindspore/common/np_dtype.py +0 -25
  307. mindspore/communication/comm_func.py +0 -1140
  308. mindspore/hal/memory.py +0 -326
  309. mindspore/lib/libavcodec.so.59 +0 -0
  310. mindspore/lib/libavdevice.so.59 +0 -0
  311. mindspore/lib/libavfilter.so.8 +0 -0
  312. mindspore/lib/libavformat.so.59 +0 -0
  313. mindspore/lib/libavutil.so.57 +0 -0
  314. mindspore/lib/libmindspore_np_dtype.so +0 -0
  315. mindspore/lib/libswresample.so.4 +0 -0
  316. mindspore/lib/libswscale.so.6 +0 -0
  317. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/all_finite.cpp +0 -326
  318. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/custom_ascendc_ops_impl/dynamic/all_finite.py +0 -180
  319. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_576ceaeef5870c451cab59af55ea46ad.json +0 -58
  320. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_576ceaeef5870c451cab59af55ea46ad.o +0 -0
  321. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_86a73ff6e28d734c96bb8d3054f7dd18.json +0 -58
  322. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_86a73ff6e28d734c96bb8d3054f7dd18.o +0 -0
  323. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_f55e0ebaad1f2f572e43677336992fa0.json +0 -58
  324. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/ascend910b/all_finite/AllFinite_f55e0ebaad1f2f572e43677336992fa0.o +0 -0
  325. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/config/ascend910b/all_finite.json +0 -109
  326. mindspore/lib/plugin/ascend/custom_ascendc_ops/op_impl/ai_core/tbe/kernel/config/ascend910b/binary_info_config.json +0 -38
  327. mindspore/lib/plugin/ascend/custom_compiler/OWNERS +0 -12
  328. mindspore/lib/plugin/ascend/custom_compiler/setup.py +0 -255
  329. mindspore/lib/plugin/ascend/custom_compiler/start.sh +0 -26
  330. mindspore/lib/plugin/ascend/custom_compiler/template.json +0 -40
  331. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/acme.h +0 -24
  332. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/acme_op.h +0 -69
  333. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/base_type.h +0 -133
  334. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/op_creator.h +0 -32
  335. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/op_param.h +0 -35
  336. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/include/tiling_info.h +0 -60
  337. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/core/kernel_register.h +0 -37
  338. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/core/platform/platform_configs.h +0 -89
  339. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/core/platform/rt_funcs.h +0 -135
  340. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/add_op.h +0 -34
  341. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/asd_backoff_base.h +0 -62
  342. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/asd_elewise_op.h +0 -33
  343. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/asd_ops.h +0 -88
  344. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/asd_pa_op.h +0 -45
  345. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/cast_op.h +0 -52
  346. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/ops/host_src/matmul_op.h +0 -95
  347. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/utils/asd_utils.h +0 -84
  348. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/acme/src/utils/comm_utils.h +0 -61
  349. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/apply_rotary_pos_emb/kernel/apply_rotary_pos_emb_fp32.h +0 -224
  350. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/and_impl.h +0 -29
  351. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/div_impl.h +0 -29
  352. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/elewise_binary_impl.h +0 -48
  353. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/elewise_binary_tiling.h +0 -25
  354. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/and_kernel.h +0 -46
  355. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/div_kernel.h +0 -46
  356. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/elewise_binary_base.h +0 -260
  357. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/elewise_binary_kernel.h +0 -35
  358. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/max_kernel.h +0 -66
  359. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/min_kernel.h +0 -66
  360. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/mul_kernel.h +0 -66
  361. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/kernel/or_kernel.h +0 -46
  362. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/max_impl.h +0 -29
  363. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/min_impl.h +0 -29
  364. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/mul_impl.h +0 -29
  365. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_binary/or_impl.h +0 -29
  366. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/abs_impl.h +0 -29
  367. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/elewise_unary_impl.h +0 -47
  368. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/elewise_unary_tiling.h +0 -24
  369. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/exp_impl.h +0 -29
  370. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/abs_kernel.h +0 -45
  371. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/elewise_unary_base.h +0 -148
  372. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/elewise_unary_kernel.h +0 -31
  373. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/exp_kernel.h +0 -45
  374. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/ln_kernel.h +0 -45
  375. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/not_kernel.h +0 -45
  376. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/reciprocal_kernel.h +0 -45
  377. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/relu_kernel.h +0 -55
  378. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/rsqrt_kernel.h +0 -45
  379. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/kernel/sqrt_kernel.h +0 -45
  380. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/ln_impl.h +0 -29
  381. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/not_impl.h +0 -29
  382. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/reciprocal_impl.h +0 -29
  383. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/relu_impl.h +0 -29
  384. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/rsqrt_impl.h +0 -29
  385. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/elewise_unary/sqrt_impl.h +0 -29
  386. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/grouped_matmul_impl.h +0 -45
  387. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/grouped_matmul_tiling.h +0 -187
  388. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/kernel/grouped_matmul.h +0 -245
  389. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/kernel/grouped_matmul_interface.h +0 -24
  390. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/kernel/grouped_matmul_utils.h +0 -111
  391. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/grouped_matmul/tiling_data.h +0 -54
  392. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/compare_param.h +0 -31
  393. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/elewise_param.h +0 -41
  394. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/param/grouped_matmul_param.h +0 -40
  395. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/profiling_util.h +0 -364
  396. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/log/log_utils.h +0 -69
  397. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/register/kernel_creator.h +0 -39
  398. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/register/kernel_registry.h +0 -114
  399. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/include/utils/utils.h +0 -98
  400. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MatMulPostFusionMixTactic/matmul_postfusion_mix.json +0 -19
  401. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MatMulPostFusionMixTactic/matmul_postfusion_mix.o +0 -0
  402. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MatMulPostFusionMixTactic/matmul_postfusion_mix_mix_aic_0.o +0 -0
  403. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MatMulPostFusionMixTactic/matmul_postfusion_mix_mix_aiv_0.o +0 -0
  404. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MultiMatMulPostFusionMixTactic/multi_matmul_postfusion_mix.json +0 -19
  405. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MultiMatMulPostFusionMixTactic/multi_matmul_postfusion_mix.o +0 -0
  406. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MultiMatMulPostFusionMixTactic/multi_matmul_postfusion_mix_mix_aic_0.o +0 -0
  407. mindspore/lib/plugin/ascend/ms_kernels_internal/internal_kernel/op_kernels/ascend910b/MultiMatMulPostFusionMixTactic/multi_matmul_postfusion_mix_mix_aiv_0.o +0 -0
  408. mindspore/mint/linalg/__init__.py +0 -22
  409. mindspore/nn/layer/embedding_service.py +0 -531
  410. mindspore/nn/layer/embedding_service_layer.py +0 -393
  411. mindspore/ops/function/reshard_func.py +0 -102
  412. mindspore/ops/operations/_infer_ops.py +0 -19
  413. mindspore/ops/operations/reshard_ops.py +0 -53
  414. mindspore/profiler/common/process_pool.py +0 -41
  415. mindspore/profiler/common/singleton.py +0 -28
  416. mindspore/profiler/parser/ascend_integrate_generator.py +0 -42
  417. mindspore/profiler/parser/ascend_memory_generator.py +0 -185
  418. mindspore/train/callback/_cluster_monitor.py +0 -201
  419. mindspore/train/callback/_flops_collector.py +0 -238
  420. mindspore/train/callback/_mindio_ttp.py +0 -443
  421. {mindspore-2.3.0.dist-info → mindspore-2.3.0rc2.dist-info}/WHEEL +0 -0
  422. {mindspore-2.3.0.dist-info → mindspore-2.3.0rc2.dist-info}/entry_points.txt +0 -0
  423. {mindspore-2.3.0.dist-info → mindspore-2.3.0rc2.dist-info}/top_level.txt +0 -0
@@ -1,29 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ABS_IMPL_H_
17
- #define MS_KERNELS_INTERNAL_KERNEL_ABS_IMPL_H_
18
- #include "elewise_unary_impl.h"
19
- namespace mindspore {
20
- namespace internal {
21
- class AbsImpl : public ElewiseUnaryImpl {
22
- public:
23
- AbsImpl(const OpParamPtr &param) : ElewiseUnaryImpl(param) {}
24
- virtual ~AbsImpl() {}
25
- int Launch() override;
26
- };
27
- } // namespace internal
28
- } // namespace mindspore
29
- #endif // MS_KERNELS_INTERNAL_KERNEL_ABS_IMPL_H_
@@ -1,47 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ELEWISE_IMPL_H_
17
- #define MS_KERNELS_INTERNAL_KERNEL_ELEWISE_IMPL_H_
18
-
19
- #include <vector>
20
- #include "include/internal_kernel.h"
21
- #include "include/param/elewise_param.h"
22
-
23
- namespace mindspore {
24
- namespace internal {
25
- class ElewiseUnaryImpl : public InternelKernelImpl {
26
- public:
27
- ElewiseUnaryImpl(const OpParamPtr &param) : InternelKernelImpl(param) {}
28
- virtual ~ElewiseUnaryImpl() {}
29
- bool Init(const ValidateInfo &info) override;
30
- int Launch() { return -1; };
31
- int Tiling(HostRawBuf &tilingBuf) override;
32
- void SetStream(const void *stream_ptr) override;
33
- void SetDeviceTilingBuf(const DeviceRawBuf &tilingBuf) override;
34
- uint64_t GetTilingBufSize() override;
35
- std::vector<uint64_t> GetWorkSpaceSize() override;
36
- int InferShape(const std::vector<DIMS> &input_shapes, std::vector<DIMS> &output_shapes) override;
37
- virtual int32_t GetMaxUbCount(uint32_t op_dtype);
38
-
39
- protected:
40
- void *stream_ptr_ = nullptr;
41
- uint8_t *device_tiling_ = nullptr;
42
- uint32_t aligned_factor_ = 128;
43
- uint32_t ub_dtype = 0;
44
- };
45
- } // namespace internal
46
- } // namespace mindspore
47
- #endif // MS_KERNELS_INTERNAL_KERNEL_ELEWISE_IMPL_H_
@@ -1,24 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
- #ifndef MS_KERNELS_INTERNAL_ASCENDC_ELEWISE_UNARY_TILING_H_
17
- #define MS_KERNELS_INTERNAL_ASCENDC_ELEWISE_UNARY_TILING_H_
18
- #include "utils/elewise_tiling.h"
19
- namespace mindspore::internal {
20
- struct ElewiseUnaryTilingData : public ElewiseTailTilingData {
21
- uint32_t op_dtype_{0};
22
- };
23
- } // namespace mindspore::internal
24
- #endif // MS_KERNELS_INTERNAL_ASCENDC_ELEWISE_UNARY_TILING_H_
@@ -1,29 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ELEWISE_UNARY_EXP_IMPL_H_
17
- #define MS_KERNELS_INTERNAL_KERNEL_ELEWISE_UNARY_EXP_IMPL_H_
18
- #include "elewise_unary_impl.h"
19
- namespace mindspore {
20
- namespace internal {
21
- class ExpImpl : public ElewiseUnaryImpl {
22
- public:
23
- ExpImpl(const OpParamPtr &param) : ElewiseUnaryImpl(param) {}
24
- virtual ~ExpImpl() {}
25
- int Launch() override;
26
- };
27
- } // namespace internal
28
- } // namespace mindspore
29
- #endif // MS_KERNELS_INTERNAL_KERNEL_ELEWISE_UNARY_EXP_IMPL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either Absress or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_ABS_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_ABS_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class AbsFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline AbsFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Abs); }
26
- };
27
-
28
- template <typename T = half>
29
- class AbsFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline AbsFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Abs); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void abs_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 1) { // fp16
36
- AbsFp16<half> op;
37
- op.InitUnary(x, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 0) { // fp32
40
- AbsFp32<float> op;
41
- op.InitUnary(x, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_ABS_KERNEL_H_
@@ -1,148 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_ELEWISE_BASE_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_ELEWISE_BASE_H_
19
-
20
- #include "kernel_operator.h"
21
- using namespace AscendC;
22
-
23
- template <typename IN_TYPE>
24
- class ElewiseBaseWide {
25
- public:
26
- __aicore__ inline ElewiseBaseWide(){};
27
-
28
- __aicore__ inline void InitUnary(GM_ADDR in, GM_ADDR out, GM_ADDR tiling) {
29
- gm_in1 = reinterpret_cast<__gm__ IN_TYPE *>(in);
30
- gm_out = reinterpret_cast<__gm__ IN_TYPE *>(out);
31
- SetTilingInfo(tiling);
32
- }
33
-
34
- __aicore__ inline void ProcessUnary() { CalculateUnary(); }
35
-
36
- __aicore__ inline void SetUnaryFunc(void (*func)(const LocalTensor<IN_TYPE> &dstLocal,
37
- const LocalTensor<IN_TYPE> &srcLocal, const int &calCount)) {
38
- elewise_unary_func_ = func;
39
- }
40
-
41
- private:
42
- __aicore__ inline void SetTilingInfo(GM_ADDR tiling) {
43
- core_idx = get_block_idx();
44
- core_num = get_block_num();
45
-
46
- avg_block_count = (uint32_t)(*((__gm__ uint32_t *)tiling + 0));
47
- avg_block_ub_num = (uint32_t)(*((__gm__ uint32_t *)tiling + 1));
48
- avg_block_ub_tail = (uint32_t)(*((__gm__ uint32_t *)tiling + 2));
49
- avg_block_ub_loop = (uint32_t)(*((__gm__ uint32_t *)tiling + 3));
50
- tail_block_count = (uint32_t)(*((__gm__ uint32_t *)tiling + 4));
51
- tail_block_ub_num = (uint32_t)(*((__gm__ uint32_t *)tiling + 5));
52
- tail_block_ub_tail = (uint32_t)(*((__gm__ uint32_t *)tiling + 6));
53
- tail_block_ub_loop = (uint32_t)(*((__gm__ uint32_t *)tiling + 7));
54
-
55
- buffer_num = (uint32_t)(*((__gm__ uint32_t *)tiling + 8));
56
- }
57
- __aicore__ inline void SetUbParam(uint32_t &ub_count, uint32_t &ub_loop, uint32_t &ub_tail) {
58
- ub_count = avg_block_ub_num;
59
- ub_loop = avg_block_ub_loop;
60
- ub_tail = avg_block_ub_tail;
61
-
62
- if (core_idx == core_num - 1) {
63
- ub_count = tail_block_ub_num;
64
- ub_loop = tail_block_ub_loop;
65
- ub_tail = tail_block_ub_tail;
66
- }
67
- }
68
-
69
- __aicore__ inline void InitUnaryInOut(uint32_t count) {
70
- pipe.InitBuffer(in1Que, buffer_num, count * sizeof(IN_TYPE));
71
- pipe.InitBuffer(outQue, buffer_num, count * sizeof(IN_TYPE));
72
- }
73
-
74
- __aicore__ inline void CopyOut(uint32_t idx, uint32_t stride, uint32_t count) {
75
- LocalTensor<IN_TYPE> out = outQue.DeQue<IN_TYPE>();
76
- DataCopy(outGm[idx * stride], out, count);
77
- outQue.FreeTensor(out);
78
- }
79
-
80
- __aicore__ inline void CopyIn1(uint32_t idx, uint32_t stride, uint32_t count) {
81
- LocalTensor<IN_TYPE> in1 = in1Que.AllocTensor<IN_TYPE>();
82
- DataCopy(in1, in1Gm[idx * stride], count);
83
- in1Que.EnQue(in1);
84
- }
85
-
86
- __aicore__ inline void CalculateUnary() {
87
- uint32_t ub_count, ub_loop, ub_tail;
88
- SetUbParam(ub_count, ub_loop, ub_tail);
89
-
90
- in1Gm.SetGlobalBuffer(gm_in1 + core_idx * avg_block_count);
91
- outGm.SetGlobalBuffer(gm_out + core_idx * avg_block_count);
92
-
93
- InitUnaryInOut(ub_count);
94
-
95
- uint32_t loop = 0;
96
- for (; loop < ub_loop - 1; loop++) {
97
- CopyIn1(loop, ub_count, ub_count);
98
- ComputeUnary(ub_count);
99
- CopyOut(loop, ub_count, ub_count);
100
- }
101
-
102
- /* for ub tail */
103
- if (ub_tail <= 0) {
104
- return;
105
- }
106
- CopyIn1(loop, ub_count, ub_tail);
107
- ComputeUnary(ub_tail);
108
- CopyOut(loop, ub_count, ub_tail);
109
- }
110
-
111
- __aicore__ inline void ComputeUnary(uint32_t count) {
112
- LocalTensor<IN_TYPE> in1 = in1Que.DeQue<IN_TYPE>();
113
- LocalTensor<IN_TYPE> out = outQue.AllocTensor<IN_TYPE>();
114
- elewise_unary_func_(out, in1, count);
115
- in1Que.FreeTensor(in1);
116
- outQue.EnQue(out);
117
- }
118
-
119
- private:
120
- void (*elewise_unary_func_)(const LocalTensor<IN_TYPE> &dstLocal, const LocalTensor<IN_TYPE> &srcLocal,
121
- const int &calCount);
122
-
123
- TPipe pipe;
124
- TQue<AscendC::QuePosition::VECIN, 1> in1Que;
125
- TQue<AscendC::QuePosition::VECOUT, 1> outQue;
126
-
127
- __gm__ IN_TYPE *__restrict__ gm_in1{nullptr};
128
- __gm__ IN_TYPE *__restrict__ gm_out{nullptr};
129
-
130
- GlobalTensor<IN_TYPE> in1Gm;
131
- GlobalTensor<IN_TYPE> outGm;
132
-
133
- uint32_t core_idx{0};
134
- uint32_t core_num{0};
135
- uint32_t buffer_num{0};
136
-
137
- uint32_t avg_block_count{0};
138
- uint32_t avg_block_ub_num{0};
139
- uint32_t avg_block_ub_tail{0};
140
- uint32_t avg_block_ub_loop{0};
141
-
142
- uint32_t tail_block_count{0};
143
- uint32_t tail_block_ub_num{0};
144
- uint32_t tail_block_ub_tail{0};
145
- uint32_t tail_block_ub_loop{0};
146
- };
147
-
148
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_ELEWISE_BASE_H_
@@ -1,31 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_KERNEL_H_
19
-
20
- void elewise_exp(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
21
- void elewise_abs(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
22
- void elewise_ln(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
23
- void elewise_not(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
24
- void elewise_relu(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
25
- void elewise_sqrt(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling, int dtype);
26
- void elewise_rsqrt(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling,
27
- int dtype);
28
- void elewise_reciprocal(uint32_t blockDim, void *l2ctrl, void *stream, uint8_t *in, uint8_t *out, uint8_t *tiling,
29
- int dtype);
30
-
31
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_KERNEL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_EXP_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_EXP_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class ExpFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline ExpFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Exp); }
26
- };
27
-
28
- template <typename T = half>
29
- class ExpFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline ExpFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Exp); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void exp_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 1) { // fp16
36
- ExpFp16<half> op;
37
- op.InitUnary(x, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 0) { // fp32
40
- ExpFp32<float> op;
41
- op.InitUnary(x, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_EXP_KERNEL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either Lnress or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_LN_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_LN_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class LnFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline LnFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Ln); }
26
- };
27
-
28
- template <typename T = half>
29
- class LnFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline LnFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Ln); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void ln_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 1) { // fp16
36
- LnFp16<half> op;
37
- op.InitUnary(x, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 0) { // fp32
40
- LnFp32<float> op;
41
- op.InitUnary(x, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_LN_KERNEL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_NOT_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_NOT_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = int16_t>
23
- class NotI16 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline NotI16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Not); }
26
- };
27
-
28
- template <typename T = uint16_t>
29
- class NotU16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline NotU16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Not); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void not_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 6) { // int16
36
- NotI16<int16_t> op;
37
- op.InitUnary(x, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 7) { // uint16
40
- NotU16<uint16_t> op;
41
- op.InitUnary(x, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_NOT_KERNEL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RECIPROCAL_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RECIPROCAL_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class ReciprocalFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline ReciprocalFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Reciprocal); }
26
- };
27
-
28
- template <typename T = half>
29
- class ReciprocalFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline ReciprocalFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Reciprocal); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void reciprocal_device(GM_ADDR x1, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 1) { // fp16
36
- ReciprocalFp16<half> op;
37
- op.InitUnary(x1, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 0) { // fp32
40
- ReciprocalFp32<float> op;
41
- op.InitUnary(x1, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RECIPROCAL_KERNEL_H_
@@ -1,55 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RELU_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RELU_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class ReluFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline ReluFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Relu); }
26
- };
27
-
28
- template <typename T = half>
29
- class ReluFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline ReluFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Relu); }
32
- };
33
-
34
- template <typename T = int32_t>
35
- class ReluI32 : public ElewiseBaseWide<T> {
36
- public:
37
- __aicore__ inline ReluI32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Relu); }
38
- };
39
-
40
- extern "C" __global__ __aicore__ void relu_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
41
- if (dtype == 1) { // fp16
42
- ReluFp16<half> op;
43
- op.InitUnary(x, y, tiling);
44
- op.ProcessUnary();
45
- } else if (dtype == 0) { // fp32
46
- ReluFp32<float> op;
47
- op.InitUnary(x, y, tiling);
48
- op.ProcessUnary();
49
- } else if (dtype == 3) { // int32
50
- ReluI32<int32_t> op;
51
- op.InitUnary(x, y, tiling);
52
- op.ProcessUnary();
53
- }
54
- }
55
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RELU_KERNEL_H_
@@ -1,45 +0,0 @@
1
- /**
2
- * Copyright 2024 Huawei Technologies Co., Ltd
3
- *
4
- * Licensed under the Apache License, Version 2.0 (the "License");
5
- * you may not use this file except in compliance with the License.
6
- * You may obtain a copy of the License at
7
- *
8
- * http://www.apache.org/licenses/LICENSE-2.0
9
- *
10
- * Unless required by applicable law or agreed to in writing, software
11
- * distributed under the License is distributed on an "AS IS" BASIS,
12
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13
- * See the License for the specific language governing permissions and
14
- * limitations under the License.
15
- */
16
-
17
- #ifndef MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RSQRT_KERNEL_H_
18
- #define MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RSQRT_KERNEL_H_
19
-
20
- #include "elewise_unary_base.h"
21
-
22
- template <typename T = float>
23
- class RsqrtFp32 : public ElewiseBaseWide<T> {
24
- public:
25
- __aicore__ inline RsqrtFp32() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Rsqrt); }
26
- };
27
-
28
- template <typename T = half>
29
- class RsqrtFp16 : public ElewiseBaseWide<T> {
30
- public:
31
- __aicore__ inline RsqrtFp16() { ElewiseBaseWide<T>::SetUnaryFunc(AscendC::Rsqrt); }
32
- };
33
-
34
- extern "C" __global__ __aicore__ void rsqrt_device(GM_ADDR x, GM_ADDR y, GM_ADDR tiling, int32_t dtype) {
35
- if (dtype == 1) { // fp16
36
- RsqrtFp16<half> op;
37
- op.InitUnary(x, y, tiling);
38
- op.ProcessUnary();
39
- } else if (dtype == 0) { // fp32
40
- RsqrtFp32<float> op;
41
- op.InitUnary(x, y, tiling);
42
- op.ProcessUnary();
43
- }
44
- }
45
- #endif // MS_KERNELS_INTERNAL_KERNEL_ASCENDC_ELEWISE_UNARY_RSQRT_KERNEL_H_