From fb32195562b5d3b801381ab65c563404642711ca Mon Sep 17 00:00:00 2001
From: PengC <chupeng5@huawei.com>
Date: Wed, 23 Apr 2025 11:33:34 +0800
Subject: [PATCH] =?UTF-8?q?=E7=A7=BB=E5=8A=A8=E9=83=A8=E5=88=86API?=
 =?UTF-8?q?=E7=9A=84struct=E4=BD=8D=E7=BD=AE?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .../softmax_common/softmax_common_utils.h     | 16 ---------
 impl/math/cumsum/cumsum_common_impl.h         |  5 ---
 .../layernorm/layernorm_common_basic_impl.h   | 13 --------
 .../layernormgrad/layernormgrad_common_impl.h |  5 +--
 .../normalize/normalize_common_impl.h         | 15 ---------
 .../welfordfinalize_common_impl.h             | 13 +-------
 .../antiquant/ascend_antiquant_common.h       |  9 ++---
 .../dequant/ascend_dequant_common_impl.h      |  7 +---
 impl/reduce/mean/mean_common_impl.h           |  6 +---
 impl/reduce/sum/sum_common_impl.h             |  6 +---
 impl/sort/topk/topk_common_utils.h            | 14 ++------
 lib/activation/softmax_utils.h                | 17 ++++++++++
 lib/math/cumsum_utils.h                       |  5 +++
 lib/normalization/layernorm_utils.h           | 13 ++++++++
 lib/normalization/layernormgrad.h             |  1 +
 lib/normalization/layernormgrad_utils.h       | 27 +++++++++++++++
 lib/normalization/normalize_utils.h           | 15 +++++++++
 lib/normalization/welfordfinalize.h           |  1 +
 lib/normalization/welfordfinalize_utils.h     | 33 +++++++++++++++++++
 lib/quantization/ascend_antiquant.h           |  1 +
 lib/quantization/ascend_antiquant_utils.h     | 27 +++++++++++++++
 lib/quantization/ascend_dequant.h             |  1 +
 lib/quantization/ascend_dequant_utils.h       | 26 +++++++++++++++
 lib/reduce/mean.h                             |  1 +
 lib/reduce/mean_utils.h                       | 26 +++++++++++++++
 lib/reduce/sum.h                              |  1 +
 lib/reduce/sum_utils.h                        | 26 +++++++++++++++
 lib/sort/topk.h                               |  1 +
 lib/sort/topk_utils.h                         | 33 +++++++++++++++++++
 29 files changed, 264 insertions(+), 100 deletions(-)
 create mode 100644 lib/normalization/layernormgrad_utils.h
 create mode 100644 lib/normalization/welfordfinalize_utils.h
 create mode 100644 lib/quantization/ascend_antiquant_utils.h
 create mode 100644 lib/quantization/ascend_dequant_utils.h
 create mode 100644 lib/reduce/mean_utils.h
 create mode 100644 lib/reduce/sum_utils.h
 create mode 100644 lib/sort/topk_utils.h

diff --git a/impl/activation/softmax/softmax_common/softmax_common_utils.h b/impl/activation/softmax/softmax_common/softmax_common_utils.h
index 8237582e..3314a646 100644
--- a/impl/activation/softmax/softmax_common/softmax_common_utils.h
+++ b/impl/activation/softmax/softmax_common/softmax_common_utils.h
@@ -37,15 +37,6 @@ constexpr uint32_t SOFTMAX_SPECIAL_BASICBLOCK_LEN = FLOAT_REPEAT_SIZE * SOFTMAX_
 constexpr uint32_t SOFTMAX_SUB_DIV_ROW_COLUMN_SIZE = 192;
 constexpr uint32_t SOFTMAX_FLOAT_SPECIAL_BLOCKREDUCE_LEN = DEFAULT_BLOCK_SIZE * HALF_FACTOR;
 
-struct SoftMaxParams {
-    uint32_t srcM{ 0 };
-    uint32_t srcK{ 0 };
-    uint32_t oriSrcM{ 0 };
-    uint32_t oriSrcK{ 0 };
-    uint32_t loopCnt{ 1 };
-    uint32_t splitMeanCnt{ 8 };
-    float alpha{ 0.9375 };
-};
 struct LastAxisShapeND {
     uint32_t m;
     uint32_t k;
@@ -67,12 +58,5 @@ struct BroadCastLastND {
     uint32_t srcK;
 };
 
-struct SoftMaxShapeInfo {
-    uint32_t srcM{ 0 };
-    uint32_t srcK{ 0 };
-    uint32_t oriSrcM{ 0 };
-    uint32_t oriSrcK{ 0 };
-};
-
 }; // namespace AscendC
 #endif // IMPL_ACTIVATION_SOFTMAX_SOFTMAX_COMMON_UTILS_H
\ No newline at end of file
diff --git a/impl/math/cumsum/cumsum_common_impl.h b/impl/math/cumsum/cumsum_common_impl.h
index 63547831..106c1385 100644
--- a/impl/math/cumsum/cumsum_common_impl.h
+++ b/impl/math/cumsum/cumsum_common_impl.h
@@ -22,11 +22,6 @@
 #include "lib/math/cumsum_utils.h"
 
 namespace AscendC {
-struct CumSumInfo {
-    uint32_t outter{0};
-    uint32_t inner{0};  // 32-byte aligned
-};
-
 __aicore__ inline TransDataTo5HDParams ExtractTransDataParam(uint8_t repeatTimes, uint32_t inner, uint16_t alignOutter, 
     uint32_t oneBlockElementNum, uint16_t dstRepStride, uint32_t srcRepStride)
 {
diff --git a/impl/normalization/layernorm/layernorm_common_basic_impl.h b/impl/normalization/layernorm/layernorm_common_basic_impl.h
index 1a76027e..b8bb9b14 100644
--- a/impl/normalization/layernorm/layernorm_common_basic_impl.h
+++ b/impl/normalization/layernorm/layernorm_common_basic_impl.h
@@ -25,19 +25,6 @@ struct LayerNormRstdTmpTensorParams {
     LocalTensor<T> varianceTmpTensor;
 };
 
-struct LayerNormPara {
-    uint32_t aLength;
-    uint32_t rLength;
-    uint32_t rLengthWithPadding;
-};
-
-struct WelfordUpdateParam {
-    uint32_t rnLength;
-    uint32_t abLength;
-    uint32_t abComputeLength;
-    float nRec;
-};
-
 template <bool isRelocate = true, bool isTransposeDst = false>
 __aicore__ inline void LayerNormReduceSumImpl(const LocalTensor<float>& dstMVTmp, const LocalTensor<float>& dst,
     const LocalTensor<float>& src, const uint32_t bsLength, const uint32_t hLength)
diff --git a/impl/normalization/layernormgrad/layernormgrad_common_impl.h b/impl/normalization/layernormgrad/layernormgrad_common_impl.h
index 57b7d47a..1460d91c 100644
--- a/impl/normalization/layernormgrad/layernormgrad_common_impl.h
+++ b/impl/normalization/layernormgrad/layernormgrad_common_impl.h
@@ -17,15 +17,12 @@
 
 #include "kernel_tensor.h"
 #include "kernel_tiling/kernel_tiling.h"
+#include "lib/normalization/layernormgrad_utils.h"
 const uint32_t LAYERNORM_GRAD_B32_BYTE_SIZE = 4;
 const uint32_t LAYERNORM_GRAD_B16_BYTE_SIZE = 2;
 
 namespace AscendC {
 
-struct LayerNormGradShapeInfo {
-    DataFormat dataFormat = DataFormat::ND;
-};
-
 struct LayerNormGradParams {
     __aicore__ LayerNormGradParams(LayerNormGradTiling &tiling, LocalTensor<float> &stackBuffer)
         : bLength(tiling.bLength),
diff --git a/impl/normalization/normalize/normalize_common_impl.h b/impl/normalization/normalize/normalize_common_impl.h
index f61ffd2f..0d4177e0 100644
--- a/impl/normalization/normalize/normalize_common_impl.h
+++ b/impl/normalization/normalize/normalize_common_impl.h
@@ -22,21 +22,6 @@
 namespace AscendC {
 const float DEFAULT_EPSILON = 1e-5;
 
-struct NormalizePara {
-    uint32_t aLength;
-    uint32_t rLength;
-    uint32_t rLengthWithPadding;
-};
-
-template <typename T>
-struct NormalizeTmpTensor {
-    __aicore__ NormalizeTmpTensor(){};
-    LocalTensor<T> tempTensorA;
-    LocalTensor<T> tempTensorB;
-    LocalTensor<T> gammaTmpTensor;
-    LocalTensor<T> betaTmpTensor;
-};
-
 template <typename U, typename T>
 __aicore__ inline constexpr bool IsDtypeValid()
 {
diff --git a/impl/normalization/welfordfinalize/welfordfinalize_common_impl.h b/impl/normalization/welfordfinalize/welfordfinalize_common_impl.h
index 9a8b62dd..08d405a3 100644
--- a/impl/normalization/welfordfinalize/welfordfinalize_common_impl.h
+++ b/impl/normalization/welfordfinalize/welfordfinalize_common_impl.h
@@ -18,6 +18,7 @@
 #include "kernel_tensor.h"
 #include "kernel_pop_stack_buffer.h"
 #include "kernel_tiling/kernel_tiling.h"
+#include "lib/normalization/welfordfinalize_utils.h"
 
 namespace AscendC 
 {
@@ -25,18 +26,6 @@ constexpr uint32_t OUTPUT_SIZE = 8;
 constexpr uint32_t B32_LEN = 256 / sizeof(float);
 constexpr uint32_t OUTPUT_MASK_B32 = 254;
 
-struct WelfordFinalizePara
-{
-    uint32_t rnLength;
-    uint32_t abLength;
-    uint32_t headCount;
-    uint32_t headCountLength;
-    uint32_t tailCount;
-    uint32_t tailCountLength;
-    float abRec;
-    float rRec;
-};
-
 struct WelfordFinalizeTilingData
 {
     uint32_t computeLength;
diff --git a/impl/quantization/antiquant/ascend_antiquant_common.h b/impl/quantization/antiquant/ascend_antiquant_common.h
index ee60d037..4bf123ce 100644
--- a/impl/quantization/antiquant/ascend_antiquant_common.h
+++ b/impl/quantization/antiquant/ascend_antiquant_common.h
@@ -15,6 +15,8 @@
 #ifndef IMPL_QUANTIZATION_ANTIQUANT_ASCEND_ANTIQUANT_COMMON_H
 #define IMPL_QUANTIZATION_ANTIQUANT_ASCEND_ANTIQUANT_COMMON_H
 
+#include "lib/quantization/ascend_antiquant_utils.h"
+
 namespace AscendC {
 constexpr uint32_t ANTIQUANT_TWO = 2;
 constexpr uint32_t ANTIQUANT_FOUR = 4;
@@ -26,12 +28,5 @@ constexpr uint32_t ANTIQUANT_SINGLE_N_SIZE_FP16 = 128;
 constexpr uint32_t ANTIQUANT_MAX_K = 255;
 constexpr uint32_t MAX_K_FOR_FP16_BRCB = 4096;
 
-struct AntiQuantShapeInfo {
-    uint32_t offsetHeight{0};
-    uint32_t offsetWidth{0};
-    uint32_t scaleHeight{0};
-    uint32_t scaleWidth{0};
-};
-
 } // namespace AscendC
 #endif // IMPL_QUANTIZATION_ANTIQUANT_ASCEND_ANTIQUANT_COMMON_H
diff --git a/impl/quantization/dequant/ascend_dequant_common_impl.h b/impl/quantization/dequant/ascend_dequant_common_impl.h
index 782977dd..235ba2b7 100644
--- a/impl/quantization/dequant/ascend_dequant_common_impl.h
+++ b/impl/quantization/dequant/ascend_dequant_common_impl.h
@@ -17,17 +17,12 @@
 
 #include "kernel_tensor.h"
 #include "kernel_pop_stack_buffer.h"
+#include "lib/quantization/ascend_dequant_utils.h"
 
 namespace AscendC {
 constexpr uint32_t FLOAT_PER_BLOCK = 8;         // 32B  = FP32(4B) * 8
 constexpr uint32_t FLOAT_PER_REPEAT = 64;       // 256B = FP32(4B) * 64
 
-struct DequantParams {
-    uint32_t m;             // outer axis length (do not need 32B align)                      in unit of element num
-    uint32_t n;             // inner axis length (must be 32B align)                          in unit of element num
-    uint32_t calCount;      // in one inner line, calCount elements do dequant calculation    in unit of element num
-};
-
 __aicore__ inline bool IsCalCountValid(const LocalTensor<int32_t>& srcTensor, uint32_t calCount)
 {
     // calCount must be in range (0, srcTensor element num]
diff --git a/impl/reduce/mean/mean_common_impl.h b/impl/reduce/mean/mean_common_impl.h
index 4215ca78..92835d4c 100644
--- a/impl/reduce/mean/mean_common_impl.h
+++ b/impl/reduce/mean/mean_common_impl.h
@@ -17,6 +17,7 @@
 
 #include "kernel_tensor.h"
 #include "kernel_operator_intf.h"
+#include "lib/reduce/mean_utils.h"
 #if ASCENDC_CPU_DEBUG
 #include "kernel_log.h"
 #include <type_traits>
@@ -25,11 +26,6 @@
 namespace AscendC {
 constexpr uint32_t HALF_NUM_PER = 128;
 constexpr uint32_t FLOAT_NUM_PER = 64;
-struct MeanParams {
-    uint32_t outter = 1;
-    uint32_t inner;  // inner = 32-byte alignment of n, inner = (n *sizeof(T) + 32 - 1) / 32 * 32 / sizeof(T)
-    uint32_t n;      // actual length of the tensor
-};
 
 template <typename T>
 __aicore__ inline void CheckParamsIsValid(const LocalTensor<T>& dstTensor, const LocalTensor<T>& srcTensor,
diff --git a/impl/reduce/sum/sum_common_impl.h b/impl/reduce/sum/sum_common_impl.h
index 0f920551..7f4dcd29 100644
--- a/impl/reduce/sum/sum_common_impl.h
+++ b/impl/reduce/sum/sum_common_impl.h
@@ -17,16 +17,12 @@
 
 #include "kernel_tensor.h"
 #include "kernel_operator_intf.h"
+#include "lib/reduce/sum_utils.h"
 #if ASCENDC_CPU_DEBUG
 #include "kernel_log.h"
 #include <type_traits>
 #endif
 namespace AscendC {
-struct SumParams {
-    uint32_t outter = 1;
-    uint32_t inner;  // inner = 32-byte alignment of n, inner = (n *sizeof(T) + 32 - 1) / 32 * 32 / sizeof(T)
-    uint32_t n;      // actual length of the tensor
-};
 
 template <typename T>
 __aicore__ inline void CheckParamsIsValid(const LocalTensor<T> &dstTensor, const LocalTensor<T> &srcTensor,
diff --git a/impl/sort/topk/topk_common_utils.h b/impl/sort/topk/topk_common_utils.h
index e999ed41..61db22b4 100644
--- a/impl/sort/topk/topk_common_utils.h
+++ b/impl/sort/topk/topk_common_utils.h
@@ -15,6 +15,8 @@
 #ifndef IMPL_SORT_TOPK_TOPK_COMMON_UTILS_H
 #define IMPL_SORT_TOPK_TOPK_COMMON_UTILS_H
 
+#include "lib/sort/topk_utils.h"
+
 #if __CCE_AICORE__ >= 200
 namespace {
 constexpr uint16_t MIN_SORT32_SIZE = 32;
@@ -46,18 +48,6 @@ constexpr uint32_t TOPK_NSMALL_INNER_LEN = 32;
 constexpr uint32_t TOPK_NORMAL_INNER_MAX_LEN = 4096;
 }  // namespace
 
-namespace AscendC {
-struct TopKInfo {
-    int32_t outter = 1;
-    int32_t inner;  // inner = 32-byte alignment of n
-    int32_t n;      // actual length of the tensor
-};
-
-enum class TopKMode {
-    TOPK_NORMAL,
-    TOPK_NSMALL,
-};
-}
 #endif
 
 #endif  // IMPL_SORT_TOPK_TOPK_COMMON_UTILS_H
\ No newline at end of file
diff --git a/lib/activation/softmax_utils.h b/lib/activation/softmax_utils.h
index 49cd6b34..afb6ebd1 100644
--- a/lib/activation/softmax_utils.h
+++ b/lib/activation/softmax_utils.h
@@ -49,5 +49,22 @@ struct SoftmaxConfig {
 
 constexpr SoftmaxConfig SOFTMAX_DEFAULT_CFG = { true, 0, 0 , SoftmaxMode::SOFTMAX_NORMAL};
 
+struct SoftMaxParams {
+    uint32_t srcM{ 0 };
+    uint32_t srcK{ 0 };
+    uint32_t oriSrcM{ 0 };
+    uint32_t oriSrcK{ 0 };
+    uint32_t loopCnt{ 1 };
+    uint32_t splitMeanCnt{ 8 };
+    float alpha{ 0.9375 };
+};
+
+struct SoftMaxShapeInfo {
+    uint32_t srcM{ 0 };
+    uint32_t srcK{ 0 };
+    uint32_t oriSrcM{ 0 };
+    uint32_t oriSrcK{ 0 };
+};
+
 }; // namespace AscendC
 #endif // LIB_ACTIVATION_SOFTMAX_UTILS_H
\ No newline at end of file
diff --git a/lib/math/cumsum_utils.h b/lib/math/cumsum_utils.h
index 39272ab2..9a19f8fa 100644
--- a/lib/math/cumsum_utils.h
+++ b/lib/math/cumsum_utils.h
@@ -23,5 +23,10 @@ struct CumSumConfig {
     bool outputLastRow{false};
 };
 
+struct CumSumInfo {
+    uint32_t outter{0};
+    uint32_t inner{0};  // 32-byte aligned
+};
+
 }; // namespace AscendC
 #endif // LIB_MATH_CUMSUM_UTILS_H
\ No newline at end of file
diff --git a/lib/normalization/layernorm_utils.h b/lib/normalization/layernorm_utils.h
index 471a0e09..a9c608f2 100644
--- a/lib/normalization/layernorm_utils.h
+++ b/lib/normalization/layernorm_utils.h
@@ -37,5 +37,18 @@ struct WelfordUpdateConfig {
 
 constexpr WelfordUpdateConfig WFUPDATE_DEFAULT_CFG = {false};
 
+struct LayerNormPara {
+    uint32_t aLength;
+    uint32_t rLength;
+    uint32_t rLengthWithPadding;
+};
+
+struct WelfordUpdateParam {
+    uint32_t rnLength;
+    uint32_t abLength;
+    uint32_t abComputeLength;
+    float nRec;
+};
+
 }; // namespace AscendC
 #endif // LIB_NORMALIZATION_LAYERNORM_UTILS_H
\ No newline at end of file
diff --git a/lib/normalization/layernormgrad.h b/lib/normalization/layernormgrad.h
index 06c572be..13f06e9e 100644
--- a/lib/normalization/layernormgrad.h
+++ b/lib/normalization/layernormgrad.h
@@ -16,6 +16,7 @@
 #define LIB_NORMALIZATION_LAYERNORMGRAD_H
 
 #include "kernel_tensor.h"
+#include "lib/normalization/layernormgrad_utils.h"
 #include "../../impl/normalization/layernormgrad/layernormgrad_common_impl.h"
 #include "kernel_tiling/kernel_tiling.h"
 
diff --git a/lib/normalization/layernormgrad_utils.h b/lib/normalization/layernormgrad_utils.h
new file mode 100644
index 00000000..b148343b
--- /dev/null
+++ b/lib/normalization/layernormgrad_utils.h
@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file layernormgrad_utils.h
+ * \brief
+ */
+#ifndef LIB_NORMALIZATION_LAYERNORMGRAD_UTILS_H
+#define LIB_NORMALIZATION_LAYERNORMGRAD_UTILS_H
+
+#include "kernel_tensor.h"
+
+namespace AscendC {
+
+struct LayerNormGradShapeInfo {
+    DataFormat dataFormat = DataFormat::ND;
+};
+
+}; // namespace AscendC
+#endif // LIB_NORMALIZATION_LAYERNORMGRAD_UTILS_H
\ No newline at end of file
diff --git a/lib/normalization/normalize_utils.h b/lib/normalization/normalize_utils.h
index 15bd1bf4..86e1710e 100644
--- a/lib/normalization/normalize_utils.h
+++ b/lib/normalization/normalize_utils.h
@@ -47,5 +47,20 @@ constexpr NormalizeConfig NLCFG_NOGAMMA = GetNormalizeConfig(false, true);
 
 constexpr NormalizeConfig NLCFG_NOOPT = GetNormalizeConfig(true, true);
 
+struct NormalizePara {
+    uint32_t aLength;
+    uint32_t rLength;
+    uint32_t rLengthWithPadding;
+};
+
+template <typename T>
+struct NormalizeTmpTensor {
+    __aicore__ NormalizeTmpTensor(){};
+    LocalTensor<T> tempTensorA;
+    LocalTensor<T> tempTensorB;
+    LocalTensor<T> gammaTmpTensor;
+    LocalTensor<T> betaTmpTensor;
+};
+
 }; // namespace AscendC
 #endif // LIB_NORMALIZATION_NORMALIZE_UTILS_H
\ No newline at end of file
diff --git a/lib/normalization/welfordfinalize.h b/lib/normalization/welfordfinalize.h
index 84e05646..d4d5c8be 100644
--- a/lib/normalization/welfordfinalize.h
+++ b/lib/normalization/welfordfinalize.h
@@ -17,6 +17,7 @@
 #ifndef LIB_NORMALIZATION_WELFORDFINALIZE_H
 #define LIB_NORMALIZATION_WELFORDFINALIZE_H
 #if __CCE_AICORE__ == 200 || __CCE_AICORE__ == 220
+#include "lib/normalization/welfordfinalize_utils.h"
 #include "../../impl/normalization/welfordfinalize/welfordfinalize_common_impl.h"
 #include "kernel_tensor.h"
 #include "kernel_tiling/kernel_tiling.h"
diff --git a/lib/normalization/welfordfinalize_utils.h b/lib/normalization/welfordfinalize_utils.h
new file mode 100644
index 00000000..61f80a83
--- /dev/null
+++ b/lib/normalization/welfordfinalize_utils.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file welfordfinalize_utils.h
+ * \brief
+ */
+#ifndef LIB_NORMALIZATION_WELFORDFINALIZE_UTILS_H
+#define LIB_NORMALIZATION_WELFORDFINALIZE_UTILS_H
+
+namespace AscendC {
+
+struct WelfordFinalizePara
+{
+    uint32_t rnLength;
+    uint32_t abLength;
+    uint32_t headCount;
+    uint32_t headCountLength;
+    uint32_t tailCount;
+    uint32_t tailCountLength;
+    float abRec;
+    float rRec;
+};
+
+}; // namespace AscendC
+#endif // LIB_NORMALIZATION_WELFORDFINALIZE_UTILS_H
\ No newline at end of file
diff --git a/lib/quantization/ascend_antiquant.h b/lib/quantization/ascend_antiquant.h
index 435367a8..0275ae1e 100644
--- a/lib/quantization/ascend_antiquant.h
+++ b/lib/quantization/ascend_antiquant.h
@@ -15,6 +15,7 @@
 #ifndef LIB_QUANTIZATION_ASCEND_ANTIQUANT_H
 #define LIB_QUANTIZATION_ASCEND_ANTIQUANT_H
 #include "kernel_tensor.h"
+#include "lib/quantization/ascend_antiquant_utils.h"
 #include "../../impl/quantization/antiquant/ascend_antiquant_impl.h"
 namespace AscendC {
 #pragma begin_pipe(V)
diff --git a/lib/quantization/ascend_antiquant_utils.h b/lib/quantization/ascend_antiquant_utils.h
new file mode 100644
index 00000000..ce21cc8f
--- /dev/null
+++ b/lib/quantization/ascend_antiquant_utils.h
@@ -0,0 +1,27 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file ascend_antiquant_utils.h
+ * \brief
+ */
+#ifndef LIB_QUANTIZATION_ASCEND_ANTIQUANT_UTILS_H
+#define LIB_QUANTIZATION_ASCEND_ANTIQUANT_UTILS_H
+
+namespace AscendC {
+struct AntiQuantShapeInfo {
+    uint32_t offsetHeight{0};
+    uint32_t offsetWidth{0};
+    uint32_t scaleHeight{0};
+    uint32_t scaleWidth{0};
+};
+
+}; // namespace AscendC
+#endif // LIB_QUANTIZATION_ASCEND_ANTIQUANT_UTILS_H
\ No newline at end of file
diff --git a/lib/quantization/ascend_dequant.h b/lib/quantization/ascend_dequant.h
index 11f7790d..1340a51e 100644
--- a/lib/quantization/ascend_dequant.h
+++ b/lib/quantization/ascend_dequant.h
@@ -14,6 +14,7 @@
  */
 #ifndef LIB_QUANTIZATION_ASCEND_DEQUANT_H
 #define LIB_QUANTIZATION_ASCEND_DEQUANT_H
+#include "lib/quantization/ascend_dequant_utils.h"
 #if defined(__CCE_AICORE__) && (__CCE_AICORE__ == 220 || __CCE_AICORE__ == 200)
 #include "kernel_tensor.h"
 #include "../../impl/quantization/dequant/ascend_dequant_common_impl.h"
diff --git a/lib/quantization/ascend_dequant_utils.h b/lib/quantization/ascend_dequant_utils.h
new file mode 100644
index 00000000..2147be7b
--- /dev/null
+++ b/lib/quantization/ascend_dequant_utils.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file ascend_dequant_utils.h
+ * \brief
+ */
+#ifndef LIB_QUANTIZATION_ASCEND_DEQUANT_UTILS_H
+#define LIB_QUANTIZATION_ASCEND_DEQUANT_UTILS_H
+
+namespace AscendC {
+struct DequantParams {
+    uint32_t m;             // outer axis length (do not need to be 32B aligned)              in unit of element num
+    uint32_t n;             // inner axis length (must be 32B aligned)                        in unit of element num
+    uint32_t calCount;      // in one inner line, calCount elements do dequant calculation    in unit of element num
+};
+
+}; // namespace AscendC
+#endif // LIB_QUANTIZATION_ASCEND_DEQUANT_UTILS_H
\ No newline at end of file
diff --git a/lib/reduce/mean.h b/lib/reduce/mean.h
index 914ce783..dbc87a13 100644
--- a/lib/reduce/mean.h
+++ b/lib/reduce/mean.h
@@ -17,6 +17,7 @@
 
 #include "kernel_tensor.h"
 #include "kernel_operator_intf.h"
+#include "lib/reduce/mean_utils.h"
 #include "../../impl/reduce/mean/mean_common_impl.h"
 #if ASCENDC_CPU_DEBUG
 #include "kernel_log.h"
diff --git a/lib/reduce/mean_utils.h b/lib/reduce/mean_utils.h
new file mode 100644
index 00000000..925d7b6c
--- /dev/null
+++ b/lib/reduce/mean_utils.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file mean_utils.h
+ * \brief
+ */
+#ifndef LIB_REDUCE_MEAN_UTILS_H
+#define LIB_REDUCE_MEAN_UTILS_H
+
+namespace AscendC {
+struct MeanParams {
+    uint32_t outter = 1;
+    uint32_t inner;  // inner = 32-byte alignment of n, inner = (n *sizeof(T) + 32 - 1) / 32 * 32 / sizeof(T)
+    uint32_t n;      // actual length of the tensor
+};
+
+}; // namespace AscendC
+#endif // LIB_REDUCE_MEAN_UTILS_H
\ No newline at end of file
diff --git a/lib/reduce/sum.h b/lib/reduce/sum.h
index 840ba3ff..a9f877b1 100644
--- a/lib/reduce/sum.h
+++ b/lib/reduce/sum.h
@@ -17,6 +17,7 @@
 
 #include "kernel_tensor.h"
 #include "kernel_operator_intf.h"
+#include "lib/reduce/sum_utils.h"
 #include "../../impl/reduce/sum/sum_common_impl.h"
 #if ASCENDC_CPU_DEBUG
 #include "kernel_log.h"
diff --git a/lib/reduce/sum_utils.h b/lib/reduce/sum_utils.h
new file mode 100644
index 00000000..61561799
--- /dev/null
+++ b/lib/reduce/sum_utils.h
@@ -0,0 +1,26 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file sum_utils.h
+ * \brief
+ */
+#ifndef LIB_REDUCE_SUM_UTILS_H
+#define LIB_REDUCE_SUM_UTILS_H
+
+namespace AscendC {
+struct SumParams {
+    uint32_t outter = 1;
+    uint32_t inner;  // inner = 32-byte alignment of n, inner = (n *sizeof(T) + 32 - 1) / 32 * 32 / sizeof(T)
+    uint32_t n;      // actual length of the tensor
+};
+
+}; // namespace AscendC
+#endif // LIB_REDUCE_SUM_UTILS_H
\ No newline at end of file
diff --git a/lib/sort/topk.h b/lib/sort/topk.h
index d07e7c7c..11e8532d 100644
--- a/lib/sort/topk.h
+++ b/lib/sort/topk.h
@@ -15,6 +15,7 @@
 #ifndef LIB_SORT_TOPK_H
 #define LIB_SORT_TOPK_H
 
+#include "lib/sort/topk_utils.h"
 #if __CCE_AICORE__ >= 200
 
 #include "kernel_tensor.h"
diff --git a/lib/sort/topk_utils.h b/lib/sort/topk_utils.h
new file mode 100644
index 00000000..8dd20767
--- /dev/null
+++ b/lib/sort/topk_utils.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) Huawei Technologies Co., Ltd. 2025. All rights reserved.
+ * This file is a part of the CANN Open Software.
+ * Licensed under CANN Open Software License Agreement Version 1.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ */
+
+/* !
+ * \file topk_utils.h
+ * \brief
+ */
+#ifndef LIB_SORT_TOPK_UTILS_H
+#define LIB_SORT_TOPK_UTILS_H
+
+namespace AscendC {
+struct TopKInfo {
+    int32_t outter = 1;
+    int32_t inner;  // inner = 32-byte alignment of n
+    int32_t n;      // actual length of the tensor
+};
+#ifndef ASCC_ENUM_TOPKMODE
+#define ASCC_ENUM_TOPKMODE
+enum class TopKMode {
+    TOPK_NORMAL,
+    TOPK_NSMALL,
+};
+#endif
+
+}; // namespace AscendC
+#endif // LIB_SORT_TOPK_UTILS_H
\ No newline at end of file
-- 
Gitee