diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/assign_cpu_kernel.cc b/mindspore/ccsrc/backend/kernel_compiler/cpu/assign_cpu_kernel.cc index 4d791e19e39..2a6f4edf071 100644 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/assign_cpu_kernel.cc +++ b/mindspore/ccsrc/backend/kernel_compiler/cpu/assign_cpu_kernel.cc @@ -30,7 +30,6 @@ void AssignCPUKernel::InitKernel(const CNodePtr &kernel_node) { MS_EXCEPTION_IF_NULL(kernel_node); auto input_x_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0); auto input_y_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 1); - if (input_x_shape.size() != input_y_shape.size()) MS_LOG(EXCEPTION) << "x y must be same shape"; for (size_t i = 0; i < input_x_shape.size(); ++i) { if (input_x_shape[i] != input_y_shape[i]) { diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.cc b/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.cc deleted file mode 100644 index 4788a075fac..00000000000 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.cc +++ /dev/null @@ -1,113 +0,0 @@ -/** - * Copyright 2020 Huawei Technologies Co., Ltd - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -#include "backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.h" -#include -#include "runtime/device/cpu/cpu_device_address.h" - -namespace mindspore { -namespace kernel { -template -void Compress(HashmapEntry *entry_p, const size_t &length, T entry) { - T i = (entry + 1) % length, off = 1; - for (; !entry_p[i].IsEmpty(); i = (i + 1) % length, off++) { - if (entry_p[i].tag > off) { - entry_p[entry].key = entry_p[i].key; - entry_p[entry].value = entry_p[i].value; - entry_p[entry].step = entry_p[i].step; - entry_p[entry].tag = entry_p[i].tag - off; - entry_p[i].SetEmpty(); - off = 0; - entry = i; - } - } -} - -void CacheSwapHashmapCPUKernel::InitKernel(const CNodePtr &kernel_node) { - auto hashmap_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0); - auto emb_idx_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 1); - - if (hashmap_shape.size() != 2) { - MS_LOG(EXCEPTION) << "Dimension of HashMap must be 2, (n, 4)"; - } - - for (size_t i = 0; i < emb_idx_shape.size(); ++i) { - batch_size_ *= emb_idx_shape[i]; - } - - hashmap_length_ = hashmap_shape[0]; - if (hashmap_length_ <= 0) { - MS_LOG(EXCEPTION) << "Hashmap length must > 0"; - } - dtype_ = AnfAlgo::GetPrevNodeOutputInferDataType(kernel_node, 0); -} - -bool CacheSwapHashmapCPUKernel::Launch(const std::vector &inputs, - const std::vector & /*workspace*/, - const std::vector &outputs) { - if (dtype_ == kNumberTypeInt32) { - LaunchKernel(inputs, outputs); - } else if (dtype_ == kNumberTypeInt64) { - LaunchKernel(inputs, outputs); - } else { - MS_LOG(ERROR) << "Only support int32, int64"; - return false; - } - return true; -} - -template -void CacheSwapHashmapCPUKernel::LaunchKernel(const std::vector &inputs, - const std::vector &outputs) { - HashmapEntry *hashmap = reinterpret_cast *>(inputs[0]->addr); - auto miss_emb_idx = reinterpret_cast(inputs[1]->addr); - step_ = *reinterpret_cast(inputs[2]->addr); - auto swap_cache_idx = reinterpret_cast(outputs[0]->addr); - auto old_emb_idx = reinterpret_cast(outputs[1]->addr); - - for (size_t i = 0; i < batch_size_; ++i) { - if (miss_emb_idx[i] < 0) { - swap_cache_idx[i] = -1; - old_emb_idx[i] = -1; - } else { - T emb_idx = miss_emb_idx[i]; - T entry = HashFunc(emb_idx, hashmap_length_); - T tag_count = 1; - while (!hashmap[entry].IsEmpty()) { - entry = (entry + 1) % hashmap_length_; - tag_count++; - } - - hashmap[entry].key = emb_idx; - hashmap[entry].step = step_; - hashmap[entry].tag = tag_count; - - T tmp_entry = (entry + 1) % hashmap_length_; - - while (hashmap[tmp_entry].IsEmpty() || hashmap[tmp_entry].IsUsing(step_)) { - tmp_entry = (tmp_entry + 1) % hashmap_length_; - } - - swap_cache_idx[i] = hashmap[tmp_entry].value; - old_emb_idx[i] = hashmap[tmp_entry].key; - hashmap[entry].value = swap_cache_idx[i]; - hashmap[tmp_entry].SetEmpty(); - Compress(hashmap, hashmap_length_, tmp_entry); - } - } -} -} // namespace kernel -} // namespace mindspore diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.h b/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.h deleted file mode 100644 index d92acf6b059..00000000000 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/cache_swap_hashmap_cpu_kernel.h +++ /dev/null @@ -1,87 +0,0 @@ -/** - * Copyright 2020 Huawei Technologies Co., Ltd - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_CACHE_SWAP_HASHMAP_CPU_KERNEL_H_ -#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_CACHE_SWAP_HASHMAP_CPU_KERNEL_H_ - -#include -#include -#include -#include "backend/kernel_compiler/cpu/cpu_kernel.h" -#include "backend/kernel_compiler/cpu/cpu_kernel_factory.h" -#include "backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.h" - -namespace mindspore { -namespace kernel { -class CacheSwapHashmapCPUKernel : public CPUKernel { - public: - CacheSwapHashmapCPUKernel() = default; - ~CacheSwapHashmapCPUKernel() override = default; - - void InitKernel(const CNodePtr &kernel_node) override; - - bool Launch(const std::vector &inputs, const std::vector &workspace, - const std::vector &outputs) override; - - template - void LaunchKernel(const std::vector &inputs, const std::vector &outputs); - - private: - size_t batch_size_{1}; - size_t hashmap_length_{1}; - int64_t step_{0}; - - TypeId dtype_{kTypeUnknown}; -}; - -MS_REG_CPU_KERNEL(CacheSwapHashmap, - KernelAttr() - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32), - CacheSwapHashmapCPUKernel); - -MS_REG_CPU_KERNEL(CacheSwapHashmap, - KernelAttr() - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64), - CacheSwapHashmapCPUKernel); - -MS_REG_CPU_KERNEL(CacheSwapHashmap, - KernelAttr() - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64), - CacheSwapHashmapCPUKernel); - -MS_REG_CPU_KERNEL(CacheSwapHashmap, - KernelAttr() - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32), - CacheSwapHashmapCPUKernel); -} // namespace kernel -} // namespace mindspore - -#endif // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_CACHE_SWAP_HASHMAP_CPU_KERNEL_H_ diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.cc b/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.cc deleted file mode 100644 index c34180d96d9..00000000000 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.cc +++ /dev/null @@ -1,108 +0,0 @@ -/** - * Copyright 2020 Huawei Technologies Co., Ltd - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ - -#include "backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.h" -#include -#include "runtime/device/cpu/cpu_device_address.h" - -namespace mindspore { -namespace kernel { -void SearchCacheIdxCPUKernel::InitKernel(const CNodePtr &kernel_node) { - auto hashmap_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 0); - auto emb_idx_shape = AnfAlgo::GetPrevNodeOutputInferShape(kernel_node, 1); - - if (hashmap_shape.size() != 2) { - MS_LOG(EXCEPTION) << "Dimension of HashMap must be 2, (n, 4)"; - } - - for (size_t i = 0; i < emb_idx_shape.size(); ++i) { - batch_size_ *= emb_idx_shape[i]; - } - - hashmap_length_ = hashmap_shape[0]; - if (hashmap_length_ <= 0) { - MS_LOG(EXCEPTION) << "Hashmap length must > 0"; - } - dtype_ = AnfAlgo::GetPrevNodeOutputInferDataType(kernel_node, 0); -} - -bool SearchCacheIdxCPUKernel::Launch(const std::vector &inputs, - const std::vector & /*workspace*/, - const std::vector &outputs) { - if (dtype_ == kNumberTypeInt32) { - LaunchKernel(inputs, outputs); - } else if (dtype_ == kNumberTypeInt64) { - LaunchKernel(inputs, outputs); - } else { - MS_LOG(ERROR) << "Only support int32, int64"; - return false; - } - return true; -} - -template -void SearchCacheIdxCPUKernel::LaunchKernel(const std::vector &inputs, - const std::vector &outputs) { - HashmapEntry *hashmap = reinterpret_cast *>(inputs[0]->addr); - auto input_indices = reinterpret_cast(inputs[1]->addr); - step_ = *reinterpret_cast(inputs[2]->addr); - emb_max_num = *reinterpret_cast(inputs[3]->addr); - cache_max_num = *reinterpret_cast(inputs[4]->addr); - auto output_cache_idx = reinterpret_cast(outputs[0]->addr); - auto output_miss_idx = reinterpret_cast(outputs[1]->addr); - auto output_miss_emb_idx = reinterpret_cast(outputs[2]->addr); - - float total_count = 0; - int count_size = 0; - float hit_count = 0; - for (size_t i = 0; i < batch_size_; ++i) { - if (input_indices[i] == emb_max_num) { - output_miss_idx[i] = -1; - output_cache_idx[i] = cache_max_num; - output_miss_emb_idx[i] = -1; - continue; - } - - T key = input_indices[i]; - T tmp_entry = HashFunc(key, hashmap_length_); - - int count = 1; - count_size += 1; - while ((!hashmap[tmp_entry].IsEmpty() && !hashmap[tmp_entry].IsKey(key))) { - tmp_entry = (tmp_entry + 1) % hashmap_length_; - count += 1; - } - - total_count += count; - if (hashmap[tmp_entry].IsEmpty()) { - output_miss_idx[i] = i; - output_miss_emb_idx[i] = key; - output_cache_idx[i] = -1; - } else { - hit_count += 1; - output_miss_idx[i] = -1; - output_cache_idx[i] = hashmap[tmp_entry].value; - hashmap[tmp_entry].step = step_; - output_miss_emb_idx[i] = -1; - } - } - if (count_size != 0) { - MS_LOG(INFO) << "avg search count: " << total_count / count_size; - MS_LOG(INFO) << "cache hit rate: " << hit_count / count_size; - } -} -} // namespace kernel -} // namespace mindspore diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.h b/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.h deleted file mode 100644 index 54413404245..00000000000 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/search_cache_idx_cpu_kernel.h +++ /dev/null @@ -1,138 +0,0 @@ -/** - * Copyright 2020 Huawei Technologies Co., Ltd - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_SEARCH_CACHE_IDX_CPU_KERNEL_H_ -#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_SEARCH_CACHE_IDX_CPU_KERNEL_H_ - -#include -#include -#include -#include -#include "backend/kernel_compiler/cpu/cpu_kernel.h" -#include "backend/kernel_compiler/cpu/cpu_kernel_factory.h" - -#define NULLTAG 0 - -namespace mindspore { -namespace kernel { -template -struct HashmapEntry { - T key; - T value; - T step; - T tag; - - bool IsEmpty() { - if (this->tag == NULLTAG) - return true; - else - return false; - } - - bool IsUsing(const T &train_step) { - if (this->step >= (train_step - 1)) - return true; - else - return false; - } - - bool IsKey(const T &emb_idx) { - if (this->key == emb_idx) - return true; - else - return false; - } - - void SetEmpty() { this->tag = NULLTAG; } -}; - -template -T HashFunc(const T &key, const size_t &m) { - return (T)(((0.6180339 * key) - floor(0.6180339 * key)) * m); -} - -class SearchCacheIdxCPUKernel : public CPUKernel { - public: - SearchCacheIdxCPUKernel() = default; - ~SearchCacheIdxCPUKernel() override = default; - - void InitKernel(const CNodePtr &kernel_node) override; - - bool Launch(const std::vector &inputs, const std::vector &workspace, - const std::vector &outputs) override; - - template - void LaunchKernel(const std::vector &inputs, const std::vector &outputs); - - private: - size_t batch_size_{1}; - size_t hashmap_length_{1}; - size_t step_{0}; - int64_t emb_max_num = 999999999; - int64_t cache_max_num = 999999999; - TypeId dtype_{kTypeUnknown}; -}; - -MS_REG_CPU_KERNEL(SearchCacheIdx, - KernelAttr() - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32), - SearchCacheIdxCPUKernel); - -MS_REG_CPU_KERNEL(SearchCacheIdx, - KernelAttr() - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64), - SearchCacheIdxCPUKernel); - -MS_REG_CPU_KERNEL(SearchCacheIdx, - KernelAttr() - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt64), - SearchCacheIdxCPUKernel); - -MS_REG_CPU_KERNEL(SearchCacheIdx, - KernelAttr() - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt32) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddInputAttr(kNumberTypeInt64) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32) - .AddOutputAttr(kNumberTypeInt32), - SearchCacheIdxCPUKernel); -} // namespace kernel -} // namespace mindspore - -#endif // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_SEARCH_CACHE_IDX_CPU_KERNEL_H_ diff --git a/mindspore/ccsrc/backend/kernel_compiler/cpu/update_cache_cpu_kernel.cc b/mindspore/ccsrc/backend/kernel_compiler/cpu/update_cache_cpu_kernel.cc index 2251d5a9b32..8936e45c843 100644 --- a/mindspore/ccsrc/backend/kernel_compiler/cpu/update_cache_cpu_kernel.cc +++ b/mindspore/ccsrc/backend/kernel_compiler/cpu/update_cache_cpu_kernel.cc @@ -82,7 +82,8 @@ void UpdateCacheCPUKernel::LaunchKernel(const std::vector &inputs, char *tmp = update + i * one_length_size; if (indices[i] * one_length_size + one_length_size <= max_size) { - int ret = memcpy_s(input_x + indices[i] * one_length_size, one_length_size, tmp, one_length_size); + int ret = + memcpy_s(input_x + indices[i] * one_length_size, max_size - indices[i] * one_length_size, tmp, one_length_size); if (ret != 0) { MS_LOG(EXCEPTION) << "memcpy_s error, errorno" << ret; } diff --git a/mindspore/ccsrc/frontend/parallel/cache_embedding/cache_embedding.cc b/mindspore/ccsrc/frontend/parallel/cache_embedding/cache_embedding.cc index 7f926985558..4e8d9cefbb7 100644 --- a/mindspore/ccsrc/frontend/parallel/cache_embedding/cache_embedding.cc +++ b/mindspore/ccsrc/frontend/parallel/cache_embedding/cache_embedding.cc @@ -145,7 +145,7 @@ void MemCopyFromHostToCache(void *hashmap_addr, void *host_addr, void *cache_add auto cache_data = static_cast(cache_addr); auto hashmap_data = static_cast *>(hashmap_addr); // default param type float - size_t param_type_size = 4; + const size_t param_type_size = 4; size_t single_col_bytes = param_type_size * col_size; for (size_t i = 0; i < hashmap_size; ++i) { if (!hashmap_data[i].IsEmpty()) { @@ -263,8 +263,6 @@ AnfNodePtr InitHashMap(const FuncGraphPtr &func_graph, const int64_t host_size, AnfNodePtr InitStep(const FuncGraphPtr &func_graph, TypeId type_id) { std::vector host_shape{1}; auto new_tensor = std::make_shared(type_id, host_shape); - auto step_data = static_cast(new_tensor->data_c()); - step_data[0] = 0; ParamInfoPtr new_param_info = std::make_shared(); std::string step_name = "cache_step"; new_param_info->set_name(step_name); diff --git a/mindspore/ccsrc/pybind_api/ir/tensor_py.cc b/mindspore/ccsrc/pybind_api/ir/tensor_py.cc index 81c592926a1..126404df34f 100644 --- a/mindspore/ccsrc/pybind_api/ir/tensor_py.cc +++ b/mindspore/ccsrc/pybind_api/ir/tensor_py.cc @@ -280,7 +280,7 @@ void MemCopyFromCacheToHost(void *hashmap_addr, void *host_addr, void *cache_add auto cache_data = static_cast(cache_addr); auto hashmap_data = static_cast *>(hashmap_addr); // default param type float - size_t param_type_size = 4; + const size_t param_type_size = 4; size_t single_col_bytes = param_type_size * col_size; for (size_t i = 0; i < hashmap_size; ++i) { if (!hashmap_data[i].IsEmpty()) { diff --git a/mindspore/ops/_op_impl/aicpu/__init__.py b/mindspore/ops/_op_impl/aicpu/__init__.py index e50cbe9584b..4c03460fa09 100644 --- a/mindspore/ops/_op_impl/aicpu/__init__.py +++ b/mindspore/ops/_op_impl/aicpu/__init__.py @@ -55,8 +55,6 @@ from .standard_normal import _standard_normal_aicpu from .gamma import _gamma_aicpu from .poisson import _poisson_aicpu from .update_cache import _update_cache_aicpu -from .search_cache_idx import _search_cache_idx_aicpu -from .cache_swap_hashmap import _cache_swap_hashmap_aicpu from .cache_swap_table import _cache_swap_table_aicpu from .uniform_int import _uniform_int_aicpu from .uniform_real import _uniform_real_aicpu diff --git a/mindspore/ops/_op_impl/aicpu/cache_swap_hashmap.py b/mindspore/ops/_op_impl/aicpu/cache_swap_hashmap.py deleted file mode 100644 index b642ac5401c..00000000000 --- a/mindspore/ops/_op_impl/aicpu/cache_swap_hashmap.py +++ /dev/null @@ -1,43 +0,0 @@ -# Copyright 2020 Huawei Technologies Co., Ltd -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================ - -"""CacheSwapHashmap op""" -from mindspore.ops.op_info_register import op_info_register, AiCPURegOp, DataType - -cache_swap_hashmap_op_info = AiCPURegOp("CacheSwapHashmap") \ - .fusion_type("OPAQUE") \ - .input(0, "hashmap", "required") \ - .input(1, "miss_emb_idx", "required") \ - .input(2, "step", "required") \ - .output(0, "swap_cache_idx", "required") \ - .output(1, "old_emb_idx", "required") \ - .dtype_format(DataType.I32_Default, DataType.I32_Default, \ - DataType.I32_Default, DataType.I32_Default, \ - DataType.I32_Default) \ - .dtype_format(DataType.I64_Default, DataType.I64_Default, \ - DataType.I32_Default, DataType.I64_Default, \ - DataType.I64_Default) \ - .dtype_format(DataType.I32_Default, DataType.I32_Default, \ - DataType.I64_Default, DataType.I32_Default, \ - DataType.I32_Default) \ - .dtype_format(DataType.I64_Default, DataType.I64_Default, \ - DataType.I64_Default, DataType.I64_Default, \ - DataType.I64_Default) \ - .get_op_info() - -@op_info_register(cache_swap_hashmap_op_info) -def _cache_swap_hashmap_aicpu(): - """CacheSwapHashmap AiCPU register""" - return diff --git a/mindspore/ops/_op_impl/aicpu/search_cache_idx.py b/mindspore/ops/_op_impl/aicpu/search_cache_idx.py deleted file mode 100644 index f83c79f3ba6..00000000000 --- a/mindspore/ops/_op_impl/aicpu/search_cache_idx.py +++ /dev/null @@ -1,51 +0,0 @@ -# Copyright 2020 Huawei Technologies Co., Ltd -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================ - -"""EmbeddingLookup op""" -from mindspore.ops.op_info_register import op_info_register, AiCPURegOp, DataType - -search_cache_idx_op_info = AiCPURegOp("SearchCacheIdx") \ - .fusion_type("OPAQUE") \ - .input(0, "hashmap", "required") \ - .input(1, "indices", "required") \ - .input(2, "step", "required") \ - .input(3, "emb_max_num", "required") \ - .input(4, "cache_max_num", "required") \ - .output(0, "cache_idx", "required") \ - .output(1, "miss_idx_1d", "required") \ - .output(2, "miss_emb_idx", "required") \ - .dtype_format(DataType.I32_Default, DataType.I32_Default, - DataType.I32_Default, DataType.I32_Default, DataType.I32_Default, - DataType.I32_Default, DataType.I32_Default, - DataType.I32_Default) \ - .dtype_format(DataType.I64_Default, DataType.I64_Default, - DataType.I32_Default, DataType.I32_Default, DataType.I32_Default, - DataType.I64_Default, DataType.I64_Default, - DataType.I64_Default) \ - .dtype_format(DataType.I32_Default, DataType.I32_Default, - DataType.I64_Default, DataType.I64_Default, DataType.I64_Default, - DataType.I32_Default, DataType.I32_Default, - DataType.I32_Default) \ - .dtype_format(DataType.I64_Default, DataType.I64_Default, - DataType.I64_Default, DataType.I64_Default, DataType.I64_Default, - DataType.I64_Default, DataType.I64_Default, - DataType.I64_Default) \ - .get_op_info() - - -@op_info_register(search_cache_idx_op_info) -def _search_cache_idx_aicpu(): - """SearchCacheIdx AiCPU register""" - return diff --git a/mindspore/ops/operations/__init__.py b/mindspore/ops/operations/__init__.py index d67b10bc52a..762428ab16d 100644 --- a/mindspore/ops/operations/__init__.py +++ b/mindspore/ops/operations/__init__.py @@ -95,8 +95,7 @@ from ._thor_ops import (CusBatchMatMul, CusCholeskyTrsm, CusFusedAbsMax1, CusImg CusMatMulCubeFraczLeftCast, Im2Col, UpdateThorGradient, Cholesky, CholeskyTrsm, DetTriangle, ProdForceSeA) from .sparse_ops import (SparseToDense, SparseTensorDenseMatmul) -from ._embedding_cache_ops import (CacheSwapHashmap, SearchCacheIdx, CacheSwapTable, UpdateCache, MapCacheIdx, - SubAndFilter, +from ._embedding_cache_ops import (CacheSwapTable, UpdateCache, MapCacheIdx, SubAndFilter, MapUniform, DynamicAssign, PadAndShift) from .quantum_ops import PQC, Evolution from .sponge_ops import (BondForce, BondEnergy, BondAtomEnergy, BondForceWithAtomEnergy, BondForceWithAtomVirial, diff --git a/mindspore/ops/operations/_embedding_cache_ops.py b/mindspore/ops/operations/_embedding_cache_ops.py index 39e872d4fa7..26ce65d95d0 100644 --- a/mindspore/ops/operations/_embedding_cache_ops.py +++ b/mindspore/ops/operations/_embedding_cache_ops.py @@ -15,7 +15,7 @@ """cache_ops""" from ..._checkparam import Validator as validator from ...common import dtype as mstype -from ..primitive import PrimitiveWithInfer, prim_attr_register, PrimitiveWithCheck +from ..primitive import prim_attr_register, PrimitiveWithCheck from .. import signature as sig @@ -30,7 +30,7 @@ class UpdateCache(PrimitiveWithCheck): - **updates** (Tensor) - The update values. Outputs: - - **out** (Tensor) - Returns a [1] Tensor, which is not usefull. + - **out** (Tensor) - Returns a [1] Tensor, which is not useful. """ __mindspore_signature__ = ( sig.make_sig('input_x', sig.sig_rw.RW_WRITE, @@ -101,92 +101,6 @@ class SubAndFilter(PrimitiveWithCheck): return input_x_dtype -class SearchCacheIdx(PrimitiveWithInfer): - """ - Search the keys of a hashmap, and return the values. - - Inputs: - - **hashmap** (Parameter) - The dim of hashmap is (n, 4), which cols represent the `key, value, step, tag`. - `key, value`: Map the indices of big table and cache table. - `step`: The resent step, when searching the key, it will be updated at the same time. - `step` can make sure the indices which are using in the last step will not be deleted in hashmap. - `tag`: We use linear probing(`h(k, i) = (h(k) + i) % m`) to solve hash conflicts. - tag is the count of linear probing times of the key. If `tag == 0`, means that the entry is empty. - The Hash Function is: - `((0.6180339 * key) - floor(0.618033 * key)) * hashmap_length`, in order to avoid data clustering. - - **indices** (Tensor) - The indices which are keys of hashmap. - - **step** (int) - The current step when searching. - - **emb_max_num** (int) - Max length of big table. - To avoid searching when `indices >= emb_max_num`, and make value = `cache_max_num`. - - **cache_max_num** (int) - Max length of cache table. - - Outputs: - - **cache_idx** (Tensor) - Result of searched value, if search missed, value = -1. - - **miss_idx** (Tensor) - The index of Tensor indices which search missed. - If search success, miss_idx[i] = -1. - - **miss_emb_idx** (Tensor) - The value of Tensor indices which search missed. - If search success, miss_emb_idx[i] = -1. - Examples: - >>> hashmap = Parameter(Tensor(np.array([[0, 0, 0, 0], - [10, 5, -5, 1], - [2, 1, -5, 1], - [15, 7, -5, 2], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [3, 3, -5, 1], - [21, 9, -5, 1]], np.int32)), name="hashmap") - >>> indices = Tensor(np.array([10, 2, 25, 5, 3], np.int32)) - >>> step = 0, emb_max_num = 25, cache_max_num = 10 - >>> ops = ops.SearchCacheIdx() - >>> cache_idx, miss_idx, miss_emb_idx = ops(hashmap, indices, step, emb_max_num, cache_max_num) - cache_idx : [5, 1, 10, -1, 3] - miss_idx : [-1, -1, -1, 3, -1] - miss_emb_idx : [-1, -1, -1, 5, -1] - hashmap after search : [[0, 0, 0, 0], - [10, 5, 0, 1], - [2, 1, 0, 1], - [15, 7, -5, 2], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [3, 3, 0, 1], - [21, 9, -5, 1]] - """ - __mindspore_signature__ = ( - sig.make_sig('hashmap', sig.sig_rw.RW_WRITE, - dtype=sig.sig_dtype.T), - sig.make_sig('indices', dtype=sig.sig_dtype.T), - sig.make_sig('step', dtype=sig.sig_dtype.T), - sig.make_sig('emb_max_num', dtype=sig.sig_dtype.T), - sig.make_sig('cache_max_num', dtype=sig.sig_dtype.T) - ) - - @prim_attr_register - def __init__(self): - """init SearchCacheIdx""" - - self.init_prim_io_names(inputs=['hashmap', 'indices', 'step', 'emb_max_num', 'cache_max_num'], - outputs=['cache_idx', 'miss_idx', 'miss_emb_idx']) - - def infer_shape(self, hashmap_shape, indices_shape, step_shape, emb_max_num_shape, cache_max_num_shape): - - if len(hashmap_shape) != 2: - raise ValueError("The dimension of 'hashmap' in SearchCacheIdx must be 2, " - "but got %d." % len(hashmap_shape)) - out_shape = (indices_shape, indices_shape, indices_shape) - return out_shape - - def infer_dtype(self, hashmap_dtype, indices_dtype, step_dtype, emb_max_num_dtype, cache_max_num_dtype): - args = {"hashmap": hashmap_dtype, "indices": indices_dtype} - validator.check_tensors_dtypes_same_and_valid( - args, mstype.int_type, self.name) - out_dtype = (hashmap_dtype, hashmap_dtype, hashmap_dtype) - return out_dtype - - class MapUniform(PrimitiveWithCheck): """ Map a tensor by using fomula : value = key % `group_num` * `per_group_size` + key // `group_num`. @@ -227,48 +141,6 @@ class MapUniform(PrimitiveWithCheck): 'group_num', group_num_dtype, [mstype.Int], self.name) -class CacheSwapHashmap(PrimitiveWithInfer): - """ - Delete a hashmap entry,and insert a new key to hashmap, return the key and value of delete entry. - - Inputs: - - **hashmap** (Parameter) - Same to operation SearchCacheIdx. - - **miss_emb_idx** (Tensor) - The keys which are going to insert, -1 is skipped. It is the result - - **step** (int) - The current step. - - Outputs: - - **swap_cache_idx** (Tensor) - Deleted value of entry, -1 is skipped. - - **old_emb_idx** (Tensor) - Deleted key of entry, -1 is skipped. - """ - __mindspore_signature__ = ( - sig.make_sig('hashmap', sig.sig_rw.RW_WRITE, - dtype=sig.sig_dtype.T), - sig.make_sig('miss_emb_idx', dtype=sig.sig_dtype.T), - sig.make_sig('step', dtype=sig.sig_dtype.T) - ) - - @prim_attr_register - def __init__(self): - """init CacheSwapHashmap""" - - self.init_prim_io_names(inputs=['hashmap', 'miss_emb_idx', 'step'], - outputs=['swap_cache_idx', 'old_emb_idx']) - - def infer_shape(self, hashmap_shape, miss_emb_idx_shape, step_shape): - if len(hashmap_shape) != 2: - raise ValueError("The dimension of 'hashmap' in CacheSwapHashmap must be 2, " - "but got %d." % len(hashmap_shape)) - - out_shape = (miss_emb_idx_shape, miss_emb_idx_shape) - return out_shape - - def infer_dtype(self, hashmap_dtype, miss_emb_idx_dtype, step_dtype): - validator.check_tensor_dtype_valid( - "miss_emb_idx", miss_emb_idx_dtype, mstype.int_type, self.name) - out_dtype = (miss_emb_idx_dtype, miss_emb_idx_dtype) - return out_dtype - - class CacheSwapTable(PrimitiveWithCheck): """ Delete a hashmap entry,and insert a new key to hashmap, return the key and value of delete entry. @@ -396,7 +268,7 @@ class PadAndShift(PrimitiveWithCheck): Pad a tensor with -1, and shift with a length. Inputs: - - **input_x** (Tensor) - The input Tensor, which will be copyed + - **input_x** (Tensor) - The input Tensor, which will be copied to `output`. - **cum_sum_arr** (Tensor) - The last value of cum_sum_arr is the pad length of output tensor, cum_sum_arr[shift_idx] is diff --git a/tests/st/ops/cpu/test_cache_ops.py b/tests/st/ops/cpu/test_cache_ops.py index 3b30a534be9..82879fc9489 100644 --- a/tests/st/ops/cpu/test_cache_ops.py +++ b/tests/st/ops/cpu/test_cache_ops.py @@ -12,7 +12,6 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================ -import math import numpy as np import pytest @@ -26,55 +25,6 @@ context.set_context(mode=context.GRAPH_MODE, device_target='CPU', save_graphs=True) -def hash_func(key, length): - return (int)(((0.6180339 * key) - math.floor(0.6180339 * key)) * length) - - -def init_hashmap(hash_map_length): - key_np = np.array([2, 3, 10, 15, 21], np.int32) - value_np = np.array([1, 3, 5, 7, 9], np.int32) - NULLTAG = 0 - INIT_STEP = -5 - hashmap_np = np.zeros((hash_map_length, 4), np.int32) - for i, key in enumerate(key_np): - entry = hash_func(key, hash_map_length) - count = 1 - while (hashmap_np[entry, 3] != NULLTAG and hashmap_np[entry, 0] != key): - count += 1 - entry = (entry + 1) % hash_map_length - if (hashmap_np[entry, 3] == NULLTAG): - hashmap_np[entry] = [key, value_np[i], INIT_STEP, count] - - return hashmap_np - - -class SearchCacheIdxNet(nn.Cell): - def __init__(self, hashmap_np): - super().__init__() - self.ops = P.SearchCacheIdx() - self.hashmap = Parameter(Tensor(hashmap_np), name="hashmap") - self.emb_max = 25 - self.cache_max = 10 - self.step = 0 - - def construct(self, indices): - return self.ops(self.hashmap, indices, self.step, self.emb_max, self.cache_max) - - -class CacheSwapHashmapNet(nn.Cell): - def __init__(self, hashmap_np): - super().__init__() - self.net = SearchCacheIdxNet(hashmap_np) - self.ops = P.CacheSwapHashmap() - self.step = 0 - self.emb_max = 25 - self.cache_max = 10 - - def construct(self, indices): - _, _, miss_emb_idx = self.net(indices) - return self.ops(self.net.hashmap, miss_emb_idx, self.step) - - class UpdateCacheNet(nn.Cell): def __init__(self, x): super().__init__() @@ -86,72 +36,6 @@ class UpdateCacheNet(nn.Cell): return self.ops(self.x, indices, update, self.max_num) -@pytest.mark.level0 -@pytest.mark.platform_x86_cpu -@pytest.mark.env_onecard -def test_search_cache_idx(): - hashmap_np = init_hashmap(10) - indices_np = np.array([10, 2, 20, 5, 3], np.int32) - search_cache_idx = SearchCacheIdxNet(hashmap_np) - indices = Tensor(indices_np) - cache_idx, miss_idx, miss_emb_idx = search_cache_idx(indices) - - expect_cache_idx = [5, 1, -1, -1, 3] - expect_miss_idx = [-1, -1, 2, 3, -1] - expect_miss_emb_idx = [-1, -1, 20, 5, -1] - - hashmap_np_after_ops = [[0, 0, 0, 0], - [10, 5, 0, 1], - [2, 1, 0, 1], - [15, 7, -5, 2], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [3, 3, 0, 1], - [21, 9, -5, 1]] - - assert np.allclose(cache_idx.asnumpy(), - np.array(expect_cache_idx, np.int32)) - assert np.allclose(miss_idx.asnumpy(), np.array(expect_miss_idx, np.int32)) - assert np.allclose(miss_emb_idx.asnumpy(), - np.array(expect_miss_emb_idx, np.int32)) - assert np.allclose(search_cache_idx.hashmap.data.asnumpy(), - np.array(hashmap_np_after_ops, np.int32)) - - -@pytest.mark.level0 -@pytest.mark.platform_x86_cpu -@pytest.mark.env_onecard -def test_cache_swap_hashmap(): - hashmap_np = init_hashmap(10) - indices_np = np.array([10, 2, 20, 5, 3], np.int32) - net = CacheSwapHashmapNet(hashmap_np) - indices = Tensor(indices_np) - swap_cache_idx, old_emb_idx = net(indices) - - expect_swap_cache_idx = [-1, -1, 9, 7, -1] - expect_old_emb_idx = [-1, -1, 21, 15, -1] - - hashmap_np_after_ops = [[5, 7, 0, 1], - [10, 5, 0, 1], - [2, 1, 0, 1], - [20, 9, 0, 1], - [20, 9, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [0, 0, 0, 0], - [3, 3, 0, 1], - [21, 9, -5, 0]] - - assert np.allclose(swap_cache_idx.asnumpy(), - np.array(expect_swap_cache_idx, np.int32)) - assert np.allclose(old_emb_idx.asnumpy(), - np.array(expect_old_emb_idx, np.int32)) - assert np.allclose(net.net.hashmap.data.asnumpy(), - np.array(hashmap_np_after_ops, np.int32)) - - @pytest.mark.level0 @pytest.mark.platform_x86_cpu @pytest.mark.env_onecard