feat: pre-commit & update clang-fmt-asc-devkit-AtomGit

M.clang-format+52-198

	@@ -1,213 +1,67 @@
1	-# 语言: None, Cpp, Java, JavaScript, ObjC, Proto, TableGen, TextProto	1	+# Copyright (c) 2025 Huawei Technologies Co., Ltd.
2	-Language: Cpp	2	+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
3	-# BasedOnStyle: LLVM	3	+# CANN Open Software License Agreement Version 2.0 (the "License").
		4	+# Please refer to the License for details. You may not use this file except in compliance with the License.
		5	+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
		6	+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
		7	+# See LICENSE in the root of the software repository for the full text of the License.
		8	+# ============================================================================
4		9
5	-# 访问说明符(public、private等)的偏移	10	+# Usage:
		11	+# Using clang-format to reformat a file:
		12	+# clang-format -i -style=Google <file>
		13	+
		14	+BasedOnStyle: Google
		15	+ColumnLimit: 120
6	AccessModifierOffset: -4	16	AccessModifierOffset: -4
		17	+SortIncludes: false
		18	+BreakBeforeBraces: Custom
7		19
8	-# 开括号(开圆括号、开尖括号、开方括号)后的对齐: Align, DontAlign, AlwaysBreak(总是在开括号后换行)	20	+# 大括号换行配置
9	-AlignAfterOpenBracket: Align	21	+BraceWrapping:
		22	+ AfterCaseLabel: false
		23	+ AfterClass: false
		24	+ AfterControlStatement: false
		25	+ # AfterEnum: false (紧凑模式，左大括号和enum同行, enum Color {...), AfterEnum: true (左大括号和enum换行).
		26	+ AfterEnum: false
		27	+ AfterFunction: true
		28	+ AfterNamespace: false
		29	+ AfterStruct: false
		30	+ AfterUnion: false
		31	+ AfterExternBlock: false
		32	+ BeforeCatch: false
		33	+ BeforeElse: false
		34	+ IndentBraces: false
		35	+ SplitEmptyFunction: false
		36	+ SplitEmptyRecord: true
		37	+ SplitEmptyNamespace: true
10		38
11	-# 连续赋值时，对齐所有等号	39	+# Align
12	-AlignConsecutiveAssignments: false	40	+AlignAfterOpenBracket: AlwaysBreak
13	-	41	+AlignEscapedNewlines: Left
14	-# 连续声明时，对齐所有声明的变量名
15	-AlignConsecutiveDeclarations: false
16	-
17	-# 右对齐逃脱换行(使用反斜杠换行)的反斜杠
18	-AlignEscapedNewlines: Right
19	-
20	-# 水平对齐二元和三元表达式的操作数
21	AlignOperands: true	42	AlignOperands: true
22	-
23	-# 对齐连续的尾随的注释
24	AlignTrailingComments: true	43	AlignTrailingComments: true
25	-	44	+AllowAllArgumentsOnNextLine: true
26	-# 不允许函数声明的所有参数在放在下一行	45	+AllowAllParametersOfDeclarationOnNextLine: true
27	-AllowAllParametersOfDeclarationOnNextLine: false
28	-
29	-# 不允许短的块放在同一行
30	-AllowShortBlocksOnASingleLine: true
31	-
32	-# 允许短的case标签放在同一行
33	-AllowShortCaseLabelsOnASingleLine: false
34	-
35	-# 允许短的函数放在同一行: None, InlineOnly(定义在类中), Empty(空函数), Inline(定义在类中，空函数), All
36	-AllowShortFunctionsOnASingleLine: Empty
37	-
38	-# 允许短的if语句保持在同一行
39	-AllowShortIfStatementsOnASingleLine: false
40	-
41	-# 禁止将短的循环体（如单行语句）放在同一行
42	-AllowShortLoopsOnASingleLine: false
43	-
44	-# 总是在返回类型后换行: None, All, TopLevel(顶级函数，不包括在类中的函数),
45	-# AllDefinitions(所有的定义，不包括声明), TopLevelDefinitions(所有的顶级函数的定义)
46	-AlwaysBreakAfterReturnType: None
47	-
48	-# 总是在多行string字面量前换行
49	AlwaysBreakBeforeMultilineStrings: false	46	AlwaysBreakBeforeMultilineStrings: false
50	-
51	-# 总是在template声明后换行
52	AlwaysBreakTemplateDeclarations: true	47	AlwaysBreakTemplateDeclarations: true
53		48
54	-# false表示函数实参要么都在同一行，要么都各自一行
55	BinPackArguments: true	49	BinPackArguments: true
56	-
57	-# false表示所有形参要么都在同一行，要么都各自一行
58	-BinPackParameters: true
59	-
60	-# 大括号换行，只有当BreakBeforeBraces设置为Custom时才有效
61	-BraceWrapping:
62	- # class定义后面
63	- AfterClass: false
64	- # 控制语句后面
65	- AfterControlStatement: false
66	- # enum定义后面
67	- AfterEnum: false
68	- # 函数定义后面
69	- AfterFunction: true
70	- # 命名空间定义后面
71	- AfterNamespace: false
72	- # struct定义后面
73	- AfterStruct: false
74	- # union定义后面
75	- AfterUnion: false
76	- # extern之后
77	- AfterExternBlock: false
78	- # catch之前
79	- BeforeCatch: false
80	- # else之前
81	- BeforeElse: false
82	- # 缩进大括号
83	- IndentBraces: false
84	- # 分离空函数
85	- SplitEmptyFunction: false
86	- # 分离空语句
87	- SplitEmptyRecord: false
88	- # 分离空命名空间
89	- SplitEmptyNamespace: false
90	-
91	-# 在二元运算符前换行: None(在操作符后换行), NonAssignment(在非赋值的操作符前换行), All(在操作符前换行)
92	-BreakBeforeBinaryOperators: NonAssignment
93	-
94	-# 在大括号前换行: Attach(始终将大括号附加到周围的上下文), Linux(除函数、命名空间和类定义，与Attach类似),
95	-# Mozilla(除枚举、函数、记录定义，与Attach类似), Stroustrup(除函数定义、catch、else，与Attach类似),
96	-# Allman(总是在大括号前换行), GNU(总是在大括号前换行，并对于控制语句的大括号增加额外的缩进), WebKit(在函数前换行), Custom
97	-# 注：这里认为语句块也属于函数
98	-BreakBeforeBraces: Custom
99	-
100	-# 在三元运算符前换行
101	BreakBeforeTernaryOperators: false	50	BreakBeforeTernaryOperators: false
102	-
103	-# 在构造函数的初始化列表的冒号后换行
104	-BreakConstructorInitializers: AfterColon
105	-
106	-#BreakInheritanceList: AfterColon
107	-
108	-BreakStringLiterals: false
109	-
110	-# 每行字符的限制，0表示没有限制
111	-ColumnLimit: 120
112	-
113	-CompactNamespaces: false
114	-
115	-# 构造函数的初始化列表要么都在同一行，要么都各自一行
116	-ConstructorInitializerAllOnOneLineOrOnePerLine: false
117	-
118	-# 构造函数的初始化列表的缩进宽度
119	-ConstructorInitializerIndentWidth: 4
120	-
121	-# 延续的行的缩进宽度
122	-ContinuationIndentWidth: 4
123	-
124	-# 去除C++11的列表初始化的大括号{后和}前的空格
125	-Cpp11BracedListStyle: true
126	-
127	-# 继承最常用的指针和引用的对齐方式
128	-DerivePointerAlignment: false
129	-
130	-# 固定命名空间注释
131	-FixNamespaceComments: true
132	-
133	-# 缩进case标签
134	-IndentCaseLabels: false
135	-
136	-IndentPPDirectives: None
137	-
138	-# 缩进宽度
139	-IndentWidth: 4
140	-
141	-# 函数返回类型换行时，缩进函数声明或函数定义的函数名
142	-IndentWrappedFunctionNames: false
143	-
144	-# 保留在块开始处的空行
145	-KeepEmptyLinesAtTheStartOfBlocks: false
146	-
147	-# 连续空行的最大数量
148	-MaxEmptyLinesToKeep: 1
149	-
150	-# 命名空间的缩进: None, Inner(缩进嵌套的命名空间中的内容), All
151	-NamespaceIndentation: None
152	-
153	-# 指针和引用的对齐: Left, Right, Middle
154	-PointerAlignment: Left
155	-
156	-# 允许重新排版注释
157	-ReflowComments: true
158	-
159	-# 允许排序#include
160	-SortIncludes: false
161	-
162	-# 允许排序 using 声明
163	-SortUsingDeclarations: false
164	-
165	-# 在C风格类型转换后添加空格
166	-SpaceAfterCStyleCast: false
167	-
168	-# 在Template 关键字后面添加空格
169	-SpaceAfterTemplateKeyword: true
170	-
171	-# 在赋值运算符之前添加空格
172	-SpaceBeforeAssignmentOperators: true
173	-
174	-# SpaceBeforeCpp11BracedList: true
175	-
176	-# SpaceBeforeCtorInitializerColon: true
177	-
178	-# SpaceBeforeInheritanceColon: true
179	-
180	-# 开圆括号之前添加一个空格: Never, ControlStatements, Always
181	-SpaceBeforeParens: ControlStatements
182	-
183	-# SpaceBeforeRangeBasedForLoopColon: true
184	-
185	-# 在空的圆括号中添加空格
186	-SpaceInEmptyParentheses: false
187	-
188	-# 在尾随的评论前添加的空格数(只适用于//)
189	SpacesBeforeTrailingComments: 1	51	SpacesBeforeTrailingComments: 1
190		52
191	-# 在尖括号的<后和>前添加空格	53	+# poniter aligment
192	-SpacesInAngles: false	54	+DerivePointerAlignment: false
		55	+PointerAlignment: Left
193		56
194	-# 在C风格类型转换的括号中添加空格	57	+# single line layout
195	-SpacesInCStyleCastParentheses: false	58	+AllowShortBlocksOnASingleLine: false
		59	+AllowShortCaseLabelsOnASingleLine: false
		60	+AllowShortFunctionsOnASingleLine: true
		61	+AllowShortIfStatementsOnASingleLine: false
		62	+AllowShortLoopsOnASingleLine: false
196		63
197	-# 在容器(ObjC和JavaScript的数组和字典等)字面量中添加空格	64	+# tab space
198	-SpacesInContainerLiterals: true	65	+UseTab: Never
199	-	66	+IndentWidth: 4
200	-# 在圆括号的(后和)前添加空格	67	+TabWidth: 4
201	-SpacesInParentheses: false
202	-
203	-# 在方括号的[后和]前添加空格，lamda表达式和未指明大小的数组的声明不受影响
204	-SpacesInSquareBrackets: false
205	-
206	-# 标准: Cpp03, Cpp11, Auto
207	-Standard: Cpp11
208	-
209	-# tab宽度
210	-TabWidth: 4
211	-
212	-# 使用tab字符: Never, ForIndentation, ForContinuationAndIndentation, Always
213	-UseTab: Never

M.gitignore+38-1

	@@ -4,4 +4,41 @@ build_out
4	output	4	output
5	__pycache__/	5	__pycache__/
6	/examples/**/build/	6	/examples/**/build/
7	-AGENTS.md	7	+AGENTS.md
		8	+oat_reports/
		9	+
		10	+oat_reports/
		11	+
		12	+oat_reports/
		13	+
		14	+oat_reports/
		15	+
		16	+oat_reports/
		17	+
		18	+oat_reports/
		19	+
		20	+oat_reports/
		21	+
		22	+oat_reports/
		23	+
		24	+oat_reports/
		25	+
		26	+oat_reports/
		27	+
		28	+oat_reports/
		29	+
		30	+log/
		31	+
		32	+log/
		33	+
		34	+log/
		35	+
		36	+oat_reports/
		37	+
		38	+oat_reports/
		39	+
		40	+oat_reports/
		41	+
		42	+oat_reports/
		43	+
		44	+log/

A.pre-commit-config.yaml+18-0

@@ -0,0 +1,18 @@
	1	+repos:
	2	+ - repo: https://github.com/pre-commit/mirrors-clang-format
	3	+ rev: v16.0.0
	4	+ hooks:
	5	+ - id: clang-format
	6	+ types_or: [c++, c]
	7	+
	8	+ # OAT compliance check (Open Source Audit Tool)
	9	+ - repo: local
	10	+ hooks:
	11	+ - id: oat-check
	12	+ name: OAT Compliance Check
	13	+ entry: bash scripts/oat_check.sh
	14	+ language: system
	15	+ pass_filenames: true
	16	+ types: [file]
	17	+ stages: [commit]
	18	+ verbose: true

Mclassify_rule.yaml+2-0

	@@ -67,6 +67,8 @@ asc-devkit:
67	- impl/c_api/instr_impl/npu_arch_3510/sync_impl/asc_get_buf_impl.h	67	- impl/c_api/instr_impl/npu_arch_3510/sync_impl/asc_get_buf_impl.h
68	- impl/c_api/instr_impl/npu_arch_2201/sync_impl.h	68	- impl/c_api/instr_impl/npu_arch_2201/sync_impl.h
69	- impl/c_api/instr_impl/npu_arch_3510/sync_impl.h	69	- impl/c_api/instr_impl/npu_arch_3510/sync_impl.h
		70	+ - impl/c_api/instr_impl/npu_arch_2201/vector_compute_impl/asc_transto5hd_impl.h
		71	+ - impl/c_api/instr_impl/npu_arch_3510/vector_compute_impl/asc_transto5hd_impl.h
70	- impl/basic_api/dav_c310/kernel_operator_determine_compute_sync_impl.h	72	- impl/basic_api/dav_c310/kernel_operator_determine_compute_sync_impl.h
71	- impl/experimental/tensor_api/algorithm/gemm_impl.h	73	- impl/experimental/tensor_api/algorithm/gemm_impl.h
72	- impl/experimental/tensor_api/atom/cube_datamove/cube_datamove_impl.h	74	- impl/experimental/tensor_api/atom/cube_datamove/cube_datamove_impl.h

Mimpl/c_api/instr_impl/npu_arch_2201/vector_compute_impl.h+18-31

	@@ -4925,42 +4925,29 @@ __aicore__ inline void asc_transpose_sync(__ubuf__ uint16_t* dst, __ubuf__ uint1
4925	asc_transpose_sync_impl(dst, src);	4925	asc_transpose_sync_impl(dst, src);
4926	}	4926	}
4927		4927
4928	-// ==========asc_transto5hd==========	4928	+// __aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
4929	-__aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4929	+#define asc_transto5hd_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
4930	- uint16_t src_stride, bool dst_high_half, bool src_high_half)	4930	+ asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))
4931	-{
4932	- asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
4933	-}
4934		4931
4935	-__aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4932	+// __aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
4936	- uint16_t src_stride, bool dst_high_half, bool src_high_half)	4933	+#define asc_transto5hd_b8_sync(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
4937	-{	4934	+ asc_transto5hd_b8_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))
4938	- asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
4939	-}
4940		4935
4941	-__aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4936	+// __aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
4942	- uint16_t src_stride)	4937	+#define asc_transto5hd_b16(dst, src, repeat, dst_stride, src_stride) \
4943	-{	4938	+ asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride))
4944	- asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);
4945	-}
4946		4939
4947	-__aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4940	+// __aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
4948	- uint16_t src_stride)	4941	+#define asc_transto5hd_b16_sync(dst, src, repeat, dst_stride, src_stride) \
4949	-{	4942	+ asc_transto5hd_b16_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))
4950	- asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride);
4951	-}
4952		4943
4953	-__aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4944	+// __aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
4954	- uint16_t src_stride)	4945	+#define asc_transto5hd_b32(dst, src, repeat, dst_stride, src_stride) \
4955	-{	4946	+ asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride))
4956	- asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);
4957	-}
4958		4947
4959	-__aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	4948	+// __aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
4960	- uint16_t src_stride)	4949	+#define asc_transto5hd_b32_sync(dst, src, repeat, dst_stride, src_stride) \
4961	-{	4950	+ asc_transto5hd_b32_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))
4962	- asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride);
4963	-}
4964		4951
4965	// ==========asc_duplicate(half/int16_t/uint16_t/bfloat16_t/float/int32_t/uint32_t)==========	4952	// ==========asc_duplicate(half/int16_t/uint16_t/bfloat16_t/float/int32_t/uint32_t)==========
4966	__aicore__ inline void asc_duplicate(__ubuf__ half* dst, half src, uint8_t repeat,	4953	__aicore__ inline void asc_duplicate(__ubuf__ half* dst, half src, uint8_t repeat,

Mimpl/c_api/instr_impl/npu_arch_2201/vector_compute_impl/asc_transto5hd_impl.h+40-40

	@@ -20,50 +20,51 @@
20		20
21	#include "instr_impl/npu_arch_2201/utils_impl/utils_impl.h"	21	#include "instr_impl/npu_arch_2201/utils_impl/utils_impl.h"
22		22
23	-__aicore__ inline void asc_transto5hd_b8_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	23	+// __aicore__ inline void asc_transto5hd_b8_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
24	- uint16_t src_stride, bool dst_high_half, bool src_high_half)	24	+#define asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
25	-{	25	+ do { \
26	- if ASC_IS_AIV {	26	+ if ASC_IS_AIV { \
27	- scatter_vnchwconv_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);	27	+ scatter_vnchwconv_b8((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), \
28	- }	28	+ (src_high_half)); \
29	-}	29	+ } \
		30	+ } while (0)
30		31
31	-__aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	32	+// __aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
32	- uint16_t src_stride, bool dst_high_half, bool src_high_half)	33	+#define asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
33	-{	34	+ do { \
34	- asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);	35	+ asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \
35	- asc_sync_post_process();	36	+ asc_sync_post_process(); \
36	-}	37	+ } while (0)
37		38
38	-__aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	39	+// __aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
39	- uint16_t src_stride)	40	+#define asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride) \
40	-{	41	+ do { \
41	- if ASC_IS_AIV {	42	+ if ASC_IS_AIV { \
42	- scatter_vnchwconv_b16(dst, src, repeat, dst_stride, src_stride);	43	+ scatter_vnchwconv_b16((dst), (src), (repeat), (dst_stride), (src_stride)); \
43	- }	44	+ } \
44	-}	45	+ } while (0)
45		46
46	-__aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,	47	+// __aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
47	- uint16_t dst_stride, uint16_t src_stride)	48	+#define asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride) \
48	-{	49	+ do { \
49	- asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);	50	+ asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \
50	- asc_sync_post_process();	51	+ asc_sync_post_process(); \
51	-}	52	+ } while (0)
52		53
53	-__aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,	54	+// __aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
54	- uint16_t src_stride)	55	+#define asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride) \
55	-{	56	+ do { \
56	- if ASC_IS_AIV {	57	+ if ASC_IS_AIV { \
57	- scatter_vnchwconv_b32(dst, src, repeat, dst_stride, src_stride);	58	+ scatter_vnchwconv_b32((dst), (src), (repeat), (dst_stride), (src_stride)); \
58	- }	59	+ } \
59	-}	60	+ } while (0)
60		61
61	-__aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,	62	+// __aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
62	- uint16_t dst_stride, uint16_t src_stride)	63	+#define asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride) \
63	-{	64	+ do { \
64	- asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);	65	+ asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \
65	- asc_sync_post_process();	66	+ asc_sync_post_process(); \
66	-}	67	+ } while (0)
67		68
68	#endif	69	#endif
69		70
	@@ -71,4 +72,3 @@ __aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t s
71	#undef ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS	72	#undef ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS
72	#undef UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	73	#undef UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
73	#endif	74	#endif
74	-

Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl.h+19-27

	@@ -5170,37 +5170,29 @@ __aicore__ inline void asc_transpose_sync(__ubuf__ uint16_t* dst, __ubuf__ uint1
5170	asc_transpose_sync_impl(dst, src);	5170	asc_transpose_sync_impl(dst, src);
5171	}	5171	}
5172		5172
5173	-// ==========asc_transto5hd==========	5173	+// __aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
5174	-__aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)	5174	+#define asc_transto5hd_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
5175	-{	5175	+ asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))
5176	- asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
5177	-}
5178		5176
5179	-__aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)	5177	+// __aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
5180	-{	5178	+#define asc_transto5hd_b8_sync(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
5181	- asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);	5179	+ asc_transto5hd_b8_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))
5182	-}
5183		5180
5184	-__aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	5181	+// __aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5185	-{	5182	+#define asc_transto5hd_b16(dst, src, repeat, dst_stride, src_stride) \
5186	- asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);	5183	+ asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride))
5187	-}
5188		5184
5189	-__aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	5185	+// __aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5190	-{	5186	+#define asc_transto5hd_b16_sync(dst, src, repeat, dst_stride, src_stride) \
5191	- asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride);	5187	+ asc_transto5hd_b16_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))
5192	-}
5193		5188
5194	-__aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	5189	+// __aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5195	-{	5190	+#define asc_transto5hd_b32(dst, src, repeat, dst_stride, src_stride) \
5196	- asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);	5191	+ asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride))
5197	-}
5198	-
5199	-__aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5200	-{
5201	- asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride);
5202	-}
5203		5192
		5193	+// __aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
		5194	+#define asc_transto5hd_b32_sync(dst, src, repeat, dst_stride, src_stride) \
		5195	+ asc_transto5hd_b32_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))
5204		5196
5205	// ===================asc_set_va_reg=====================	5197	// ===================asc_set_va_reg=====================
5206	__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int8_t** src_array)	5198	__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int8_t** src_array)
	@@ -5242,4 +5234,4 @@ __aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ float** src_arra
5242	{	5234	{
5243	asc_set_va_reg_impl(addr, src_array);	5235	asc_set_va_reg_impl(addr, src_array);
5244	}	5236	}
5245	-#endif	5237	+#endif

Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl/asc_transto5hd_impl.h+40-33

	@@ -25,44 +25,51 @@
25		25
26	#include "instr_impl/npu_arch_3510/utils_impl/utils_impl.h"	26	#include "instr_impl/npu_arch_3510/utils_impl/utils_impl.h"
27		27
28	-__aicore__ inline void asc_transto5hd_b8_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)	28	+// __aicore__ inline void asc_transto5hd_b8_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
29	-{	29	+#define asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
30	- if ASC_IS_AIV {	30	+ do { \
31	- scatter_vnchwconv_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);	31	+ if ASC_IS_AIV { \
32	- }	32	+ scatter_vnchwconv_b8((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), \
33	-}	33	+ (src_high_half)); \
		34	+ } \
		35	+ } while (0)
34		36
35	-__aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)	37	+// __aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
36	-{	38	+#define asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
37	- asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);	39	+ do { \
38	- asc_sync_post_process();	40	+ asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \
39	-}	41	+ asc_sync_post_process(); \
		42	+ } while (0)
40		43
41	-__aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	44	+// __aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
42	-{	45	+#define asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride) \
43	- if ASC_IS_AIV {	46	+ do { \
44	- scatter_vnchwconv_b16(dst, src, repeat, dst_stride, src_stride);	47	+ if ASC_IS_AIV { \
45	- }	48	+ scatter_vnchwconv_b16((dst), (src), (repeat), (dst_stride), (src_stride)); \
46	-}	49	+ } \
		50	+ } while (0)
47		51
48	-__aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	52	+// __aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
49	-{	53	+#define asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride) \
50	- asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);	54	+ do { \
51	- asc_sync_post_process();	55	+ asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \
52	-}	56	+ asc_sync_post_process(); \
		57	+ } while (0)
53		58
54	-__aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	59	+// __aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
55	-{	60	+#define asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride) \
56	- if ASC_IS_AIV {	61	+ do { \
57	- scatter_vnchwconv_b32(dst, src, repeat, dst_stride, src_stride);	62	+ if ASC_IS_AIV { \
58	- }	63	+ scatter_vnchwconv_b32((dst), (src), (repeat), (dst_stride), (src_stride)); \
59	-}	64	+ } \
		65	+ } while (0)
60		66
61	-__aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)	67	+// __aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
62	-{	68	+#define asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride) \
63	- asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);	69	+ do { \
64	- asc_sync_post_process();	70	+ asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \
65	-}	71	+ asc_sync_post_process(); \
		72	+ } while (0)
66		73
67	#endif	74	#endif
68		75

Mimpl/experimental/tensor_api/algorithm/copy_impl.h+18-17

Mimpl/experimental/tensor_api/algorithm/mad_impl.h+18-17

Mimpl/experimental/tensor_api/arch/cube_compute/mmad/mmad_impl.h+28-30

Mimpl/experimental/tensor_api/arch/cube_compute/mmad/mmad_routing.h+21-25

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_compute/mmad/mmad_routing.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_compute/mmad/mmad_routing.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -28,34 +28,30 @@
28	namespace AscendC {	28	namespace AscendC {
29	namespace Te {	29	namespace Te {
30		30
31	-class MmadIgnore	31	+class MmadIgnore {
32	-{
33	public:	32	public:
34	- template <const MmadTrait& trait, typename ...Args>	33	+ template <const MmadTrait& trait, typename... Args>
35	- __aicore__ inline void Run(const Args&... args) {}	34	+ __aicore__ inline void Run(const Args&... args)
		35	+ {}
36	};	36	};
37		37
38	-template<Hardware dstPos, Hardware fmPos, Hardware filterPos, Hardware biasPos, uint32_t Version>	38	+template <Hardware dstPos, Hardware fmPos, Hardware filterPos, Hardware biasPos, uint32_t Version>
39	-struct MmadTensor2Tensor	39	+struct MmadTensor2Tensor {
40	-{
41	using type = MmadIgnore;	40	using type = MmadIgnore;
42	};	41	};
43		42
44	-template<>	43	+template <>
45	-struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::MAX, ArchVersion::V3510>	44	+struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::MAX, ArchVersion::V3510> {
46	-{
47	using type = Mmad3510;	45	using type = Mmad3510;
48	};	46	};
49		47
50	-template<>	48	+template <>
51	-struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::L0C, ArchVersion::V3510>	49	+struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::L0C, ArchVersion::V3510> {
52	-{
53	using type = MmadWithBias3510;	50	using type = MmadWithBias3510;
54	};	51	};
55		52
56	-template<>	53	+template <>
57	-struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::BIAS, ArchVersion::V3510>	54	+struct MmadTensor2Tensor<Hardware::L0C, Hardware::L0A, Hardware::L0B, Hardware::BIAS, ArchVersion::V3510> {
58	-{
59	using type = MmadWithBias3510;	55	using type = MmadWithBias3510;
60	};	56	};
61	} // namespace Te	57	} // namespace Te

Mimpl/experimental/tensor_api/arch/cube_compute/mmad/npu_arch_3510/instruction.h+41-25

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_compute/mmad/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_compute/mmad/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -32,15 +32,18 @@ namespace Te {
32	class MmadInstr {	32	class MmadInstr {
33	public:	33	public:
34	template <typename T, typename U, typename S, typename... Params>	34	template <typename T, typename U, typename S, typename... Params>
35	- __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const Params& ...params)	35	+ __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const Params&... params)
36	{	36	{
37	// MTE2	37	// MTE2
38	MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), params...);	38	MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), params...);
39	}	39	}
		40	+
40	private:	41	private:
41	template <typename T, typename U, typename S>	42	template <typename T, typename U, typename S>
42	- __aicore__ inline static void MmadImpl(__cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint16_t m, uint16_t k, uint16_t n,	43	+ __aicore__ inline static void MmadImpl(
43	- uint8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal) {	44	+ __cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag,
		45	+ bool disableGemv, bool cmatrixSource, bool cmatrixInitVal)
		46	+ {
44	if ASCEND_IS_AIV {	47	if ASCEND_IS_AIV {
45	return;	48	return;
46	}	49	}
	@@ -53,16 +56,21 @@ private:
53	class MmadBiasInstr {	56	class MmadBiasInstr {
54	public:	57	public:
55	template <typename T, typename U, typename S, typename V, typename... Params>	58	template <typename T, typename U, typename S, typename V, typename... Params>
56	- __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const V& bias, const Params& ...params)	59	+ __aicore__ inline static void Mmad(
		60	+ const T& dst, const U& fm, const S& filter, const V& bias, const Params&... params)
57	{	61	{
58	// MTE2	62	// MTE2
59	- MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(),	63	+ MmadImpl(
60	- reinterpret_cast<uint64_t>(bias.Data().Get()), params...);	64	+ dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), reinterpret_cast<uint64_t>(bias.Data().Get()),
		65	+ params...);
61	}	66	}
		67	+
62	private:	68	private:
63	template <typename T, typename U, typename S>	69	template <typename T, typename U, typename S>
64	- __aicore__ inline static void MmadImpl(__cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint64_t bias, uint16_t m, uint16_t k, uint16_t n,	70	+ __aicore__ inline static void MmadImpl(
65	- int8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal) {	71	+ __cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint64_t bias, uint16_t m, uint16_t k, uint16_t n,
		72	+ int8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal)
		73	+ {
66	if ASCEND_IS_AIV {	74	if ASCEND_IS_AIV {
67	return;	75	return;
68	}	76	}
	@@ -77,15 +85,18 @@ private:
77	class MmadMxInstr {	85	class MmadMxInstr {
78	public:	86	public:
79	template <typename T, typename U, typename S, typename... Params>	87	template <typename T, typename U, typename S, typename... Params>
80	- __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const Params& ...params)	88	+ __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const Params&... params)
81	{	89	{
82	// MTE2	90	// MTE2
83	MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), params...);	91	MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), params...);
84	}	92	}
		93	+
85	private:	94	private:
86	template <typename T, typename U, typename S>	95	template <typename T, typename U, typename S>
87	- __aicore__ inline static void MmadImpl(__cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint16_t m, uint16_t k, uint16_t n,	96	+ __aicore__ inline static void MmadImpl(
88	- uint8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal) {	97	+ __cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint16_t m, uint16_t k, uint16_t n, uint8_t unitFlag,
		98	+ bool disableGemv, bool cmatrixSource, bool cmatrixInitVal)
		99	+ {
89	if ASCEND_IS_AIV {	100	if ASCEND_IS_AIV {
90	return;	101	return;
91	}	102	}
	@@ -98,16 +109,21 @@ private:
98	class MmadMxBiasInstr {	109	class MmadMxBiasInstr {
99	public:	110	public:
100	template <typename T, typename U, typename S, typename V, typename... Params>	111	template <typename T, typename U, typename S, typename V, typename... Params>
101	- __aicore__ inline static void Mmad(const T& dst, const U& fm, const S& filter, const V& bias, const Params& ...params)	112	+ __aicore__ inline static void Mmad(
		113	+ const T& dst, const U& fm, const S& filter, const V& bias, const Params&... params)
102	{	114	{
103	// MTE2	115	// MTE2
104	- MmadImpl(dst.Data().Get(), fm.Data().Get(), filter.Data().Get(),	116	+ MmadImpl(
105	- reinterpret_cast<uint64_t>(bias.Data().Get()), params...);	117	+ dst.Data().Get(), fm.Data().Get(), filter.Data().Get(), reinterpret_cast<uint64_t>(bias.Data().Get()),
		118	+ params...);
106	}	119	}
		120	+
107	private:	121	private:
108	template <typename T, typename U, typename S>	122	template <typename T, typename U, typename S>
109	- __aicore__ inline static void MmadImpl(__cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint64_t bias, uint16_t m, uint16_t k, uint16_t n,	123	+ __aicore__ inline static void MmadImpl(
110	- int8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal) {	124	+ __cc__ T* dst, __ca__ U* fm, __cb__ S* filter, uint64_t bias, uint16_t m, uint16_t k, uint16_t n,
		125	+ int8_t unitFlag, bool disableGemv, bool cmatrixSource, bool cmatrixInitVal)
		126	+ {
111	if ASCEND_IS_AIV {	127	if ASCEND_IS_AIV {
112	return;	128	return;
113	}	129	}

Mimpl/experimental/tensor_api/arch/cube_compute/mmad/npu_arch_3510/mmad.h+23-20

Mimpl/experimental/tensor_api/arch/cube_compute/mmad/npu_arch_3510/mmad_with_bias.h+23-19

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/data_copy_impl.h+14-15

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/data_copy_routing.h+12-11

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1.h+1-2

	@@ -8,9 +8,8 @@
8	* See LICENSE in the root of the software repository for the full text of the License.	8	* See LICENSE in the root of the software repository for the full text of the License.
9	*/	9	*/
10		10
11	-
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2nn.h+5-3

	@@ -46,11 +46,13 @@ private:
46		46
47	auto srcRowShape = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout);	47	auto srcRowShape = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout);
48	uint32_t srcColShape = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout);	48	uint32_t srcColShape = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout);
49	- uint16_t srcBColStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout);	49	+ uint16_t srcBColStride =
50	- uint16_t dstBColStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout);	50	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout);
		51	+ uint16_t dstBColStride =
		52	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout);
51		53
52	uint16_t dnNum = 1;	54	uint16_t dnNum = 1;
53	- uint16_t nValue = srcRowShape >> 1; // use b16 for DN2NZ, so nValue = srcRowShape / 2	55	+ uint16_t nValue = srcRowShape >> 1; // use b16 for DN2NZ, so nValue = srcRowShape / 2
54	uint16_t dValue = srcColShape;	56	uint16_t dValue = srcColShape;
55	uint16_t dstNzNStride = 1;	57	uint16_t dstNzNStride = 1;
56		58

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2nz.h+4-4

	@@ -8,9 +8,8 @@
8	* See LICENSE in the root of the software repository for the full text of the License.	8	* See LICENSE in the root of the software repository for the full text of the License.
9	*/	9	*/
10		10
11	-
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -80,8 +79,9 @@ private:
80		79
81	uint8_t cacheMode = GetCacheModeFromTensor(src);	80	uint8_t cacheMode = GetCacheModeFromTensor(src);
82		81
83	- CopyGmToCbufMultiDn2nzInstr::DataCopy(dst, src, dnNum, loop2DstStride, loop3DstStride, loop4DstStride,	82	+ CopyGmToCbufMultiDn2nzInstr::DataCopy(
84	- loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, false);	83	+ dst, src, dnNum, loop2DstStride, loop3DstStride, loop4DstStride, loop1SrcStride, cacheMode, nValue, dValue,
		84	+ loop4SrcStride, false);
85	}	85	}
86	};	86	};
87		87

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2zn.h+18-15

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2zz.h+1-2

	@@ -8,9 +8,8 @@
8	* See LICENSE in the root of the software repository for the full text of the License.	8	* See LICENSE in the root of the software repository for the full text of the License.
9	*/	9	*/
10		10
11	-
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2zz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/dn2zz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2nd.h+5-7

	@@ -8,9 +8,8 @@
8	* See LICENSE in the root of the software repository for the full text of the License.	8	* See LICENSE in the root of the software repository for the full text of the License.
9	*/	9	*/
10		10
11	-
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2nd.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2nd.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -78,9 +77,8 @@ private:
78	uint8_t leftPaddingCnt = 0;	77	uint8_t leftPaddingCnt = 0;
79	uint8_t rightPaddingCnt = padCount;	78	uint8_t rightPaddingCnt = padCount;
80		79
81	- if ((srcShapeRows == 1) \|\| (srcShapeColumns == 1)	80	+ if ((srcShapeRows == 1) \|\| (srcShapeColumns == 1) \|\|
82	- \|\| (srcStrideRows == srcShapeColumns && dstStrideRows == dstShapeColumns	81	+ (srcStrideRows == srcShapeColumns && dstStrideRows == dstShapeColumns && srcStrideRows == dstStrideRows)) {
83	- && srcStrideRows == dstStrideRows)) {
84	// compact mode	82	// compact mode
85	blockCount = 1;	83	blockCount = 1;
86	// must use srcShape, there is a scenario of small to large, using dstShape will cause src out of bound	84	// must use srcShape, there is a scenario of small to large, using dstShape will cause src out of bound
	@@ -90,8 +88,8 @@ private:
90	srcStride = 0;	88	srcStride = 0;
91	dstStride = blockLen;	89	dstStride = blockLen;
92	}	90	}
93	- CopyGmToCbufAlignV2Base::DataCopy(dst, src, blockCount, blockLen, leftPaddingCnt, rightPaddingCnt, cacheMode,	91	+ CopyGmToCbufAlignV2Base::DataCopy(
94	- srcStride, dstStride);	92	+ dst, src, blockCount, blockLen, leftPaddingCnt, rightPaddingCnt, cacheMode, srcStride, dstStride);
95	}	93	}
96	};	94	};
97	} // namespace Te	95	} // namespace Te

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2nz.h+16-15

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zn.h+16-15

	@@ -1,16 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10	-
11		10
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -31,7 +30,8 @@ namespace Te {
31	class CopyGmToCbufMultiND2Zn {	30	class CopyGmToCbufMultiND2Zn {
32	public:	31	public:
33	template <const DataCopyTrait& trait, typename T, typename U>	32	template <const DataCopyTrait& trait, typename T, typename U>
34	- __aicore__ inline static void Run(const T& dst, const U& src) {	33	+ __aicore__ inline static void Run(const T& dst, const U& src)
		34	+ {
35	DataCopyImpl<trait>(dst, src);	35	DataCopyImpl<trait>(dst, src);
36	}	36	}
37		37
	@@ -78,13 +78,14 @@ private:
78		78
79	uint8_t cacheMode = GetCacheModeFromTensor(src);	79	uint8_t cacheMode = GetCacheModeFromTensor(src);
80		80
81	- CopyGmToCbufMultiDn2nzInstr::DataCopy(dst, src, ndNum, loop2DstStride, loop3DstStride, loop4DstStride,	81	+ CopyGmToCbufMultiDn2nzInstr::DataCopy(
82	- loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, false);	82	+ dst, src, ndNum, loop2DstStride, loop3DstStride, loop4DstStride, loop1SrcStride, cacheMode, nValue, dValue,
		83	+ loop4SrcStride, false);
83	}	84	}
84	};	85	};
85		86
86	-}	87	+} // namespace Te
87	-}	88	+} // namespace AscendC
88		89
89	#endif	90	#endif
90		91

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zz.h+1-2

	@@ -8,9 +8,8 @@
8	* See LICENSE in the root of the software repository for the full text of the License.	8	* See LICENSE in the root of the software repository for the full text of the License.
9	*/	9	*/
10		10
11	-
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nd2zz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nn2nn.h+3-3

	@@ -63,9 +63,9 @@ private:
63		63
64	uint8_t leftPaddingCnt = 0;	64	uint8_t leftPaddingCnt = 0;
65	uint8_t rightPaddingCnt = 0;	65	uint8_t rightPaddingCnt = 0;
66	- CopyGmToCbufAlignV2Base::CopyGmToCbufAlignV2((__cbuf__ half*)(dst.Data().Get()),	66	+ CopyGmToCbufAlignV2Base::CopyGmToCbufAlignV2(
67	- (__gm__ half*)(src.Data().Get()), blockCount, blockLen,	67	+ (__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), blockCount, blockLen, leftPaddingCnt,
68	- leftPaddingCnt, rightPaddingCnt, cacheMode, srcStride, dstStride);	68	+ rightPaddingCnt, cacheMode, srcStride, dstStride);
69	}	69	}
70	};	70	};
71		71

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/nz2nz.h+5-6

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_gm2l1/zz2zz.h+3-3

	@@ -65,9 +65,9 @@ private:
65		65
66	uint8_t leftPaddingCnt = 0;	66	uint8_t leftPaddingCnt = 0;
67	uint8_t rightPaddingCnt = 0;	67	uint8_t rightPaddingCnt = 0;
68	- CopyGmToCbufAlignV2Base::CopyGmToCbufAlignV2((__cbuf__ half*)(dst.Data().Get()),	68	+ CopyGmToCbufAlignV2Base::CopyGmToCbufAlignV2(
69	- (__gm__ half*)(src.Data().Get()), blockCount, blockLen,	69	+ (__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), blockCount, blockLen, leftPaddingCnt,
70	- leftPaddingCnt, rightPaddingCnt, cacheMode, srcStride, dstStride);	70	+ rightPaddingCnt, cacheMode, srcStride, dstStride);
71	}	71	}
72	};	72	};
73		73

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_l12bt.h+11-11

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/data_copy_l12fb.h+11-11

Mimpl/experimental/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/instruction.h+59-43

	@@ -1,16 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10	-
11		10
12	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
13	-#warning \	12	+#warning \
14	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/data_copy/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
15	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
16	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -31,7 +30,8 @@ namespace AscendC {
31	namespace Te {	30	namespace Te {
32		31
33	template <typename T>	32	template <typename T>
34	-__aicore__ inline void SetMTE2NzPara(const T& para) {	33	+__aicore__ inline void SetMTE2NzPara(const T& para)
		34	+{
35	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {	35	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {
36	set_mte2_nz_para(para);	36	set_mte2_nz_para(para);
37	}	37	}
	@@ -40,29 +40,34 @@ __aicore__ inline void SetMTE2NzPara(const T& para) {
40	class CopyGmToCbufAlignV2Base {	40	class CopyGmToCbufAlignV2Base {
41	public:	41	public:
42	template <typename T, typename U, typename... Params>	42	template <typename T, typename U, typename... Params>
43	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params) {	43	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
		44	+ {
44	using srcType = typename U::elementType;	45	using srcType = typename U::elementType;
45	- if constexpr(sizeof(srcType) == sizeof(int8_t)) {	46	+ if constexpr (sizeof(srcType) == sizeof(int8_t)) {
46	CopyGmToCbufAlignV2((__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);	47	CopyGmToCbufAlignV2((__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);
47	} else if constexpr (sizeof(srcType) == sizeof(half)) {	48	} else if constexpr (sizeof(srcType) == sizeof(half)) {
48	CopyGmToCbufAlignV2((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);	49	CopyGmToCbufAlignV2((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);
49	} else if constexpr (sizeof(srcType) == sizeof(float)) {	50	} else if constexpr (sizeof(srcType) == sizeof(float)) {
50	CopyGmToCbufAlignV2((__cbuf__ float)(dst.Data().Get()), (__gm__ float)(src.Data().Get()), params...);	51	CopyGmToCbufAlignV2((__cbuf__ float)(dst.Data().Get()), (__gm__ float)(src.Data().Get()), params...);
51	} else if constexpr (sizeof(srcType) == sizeof(uint64_t)) {	52	} else if constexpr (sizeof(srcType) == sizeof(uint64_t)) {
52	- CopyGmToCbufAlignV2((__cbuf__ uint32_t)(dst.Data().Get()), (__gm__ uint32_t)(src.Data().Get()), params...);	53	+ CopyGmToCbufAlignV2(
53	- }	54	+ (__cbuf__ uint32_t)(dst.Data().Get()), (__gm__ uint32_t)(src.Data().Get()), params...);
		55	+ }
54	}	56	}
55		57
56	template <typename T>	58	template <typename T>
57	- __aicore__ inline static void CopyGmToCbufAlignV2(__cbuf__ T* dst, __gm__ T* src, uint32_t blockCount, uint32_t blockLen,	59	+ __aicore__ inline static void CopyGmToCbufAlignV2(
58	- uint8_t leftPaddingCnt, uint8_t rightPaddingCnt, uint8_t cacheMode, uint64_t srcStride, uint32_t dstStride) {	60	+ __cbuf__ T* dst, __gm__ T* src, uint32_t blockCount, uint32_t blockLen, uint8_t leftPaddingCnt,
		61	+ uint8_t rightPaddingCnt, uint8_t cacheMode, uint64_t srcStride, uint32_t dstStride)
		62	+ {
59	if ASCEND_IS_AIV {	63	if ASCEND_IS_AIV {
60	return;	64	return;
61	}	65	}
62		66
63	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {	67	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {
64	- copy_gm_to_cbuf_align_v2(dst, src, 0, blockCount, blockLen, leftPaddingCnt, rightPaddingCnt, true,	68	+ copy_gm_to_cbuf_align_v2(
65	- cacheMode, srcStride, dstStride);	69	+ dst, src, 0, blockCount, blockLen, leftPaddingCnt, rightPaddingCnt, true, cacheMode, srcStride,
		70	+ dstStride);
66	}	71	}
67	}	72	}
68	};	73	};
	@@ -70,10 +75,12 @@ public:
70	class CopyGmToCbufMultiNd2nzInstr {	75	class CopyGmToCbufMultiNd2nzInstr {
71	public:	76	public:
72	template <typename T, typename U, typename... Params>	77	template <typename T, typename U, typename... Params>
73	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params) {	78	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
		79	+ {
74	using srcType = typename U::elementType;	80	using srcType = typename U::elementType;
75	- if constexpr(sizeof(srcType) == sizeof(int8_t)) {	81	+ if constexpr (sizeof(srcType) == sizeof(int8_t)) {
76	- CopyGmToCbufMultiNd2nz((__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);	82	+ CopyGmToCbufMultiNd2nz(
		83	+ (__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);
77	} else if constexpr (sizeof(srcType) == sizeof(half)) {	84	} else if constexpr (sizeof(srcType) == sizeof(half)) {
78	CopyGmToCbufMultiNd2nz((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);	85	CopyGmToCbufMultiNd2nz((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);
79	} else if constexpr (sizeof(srcType) == sizeof(float)) {	86	} else if constexpr (sizeof(srcType) == sizeof(float)) {
	@@ -82,9 +89,10 @@ public:
82	}	89	}
83		90
84	template <typename T>	91	template <typename T>
85	- __aicore__ inline static void CopyGmToCbufMultiNd2nz(__cbuf__ T* dst, __gm__ T* src, uint16_t ndNum, uint16_t loop2DstStride,	92	+ __aicore__ inline static void CopyGmToCbufMultiNd2nz(
86	- uint16_t loop3DstStride, uint16_t loop4DstStride, uint64_t loop1SrcStride, uint8_t cacheMode, uint16_t nValue,	93	+ __cbuf__ T* dst, __gm__ T* src, uint16_t ndNum, uint16_t loop2DstStride, uint16_t loop3DstStride,
87	- uint32_t dValue, uint64_t loop4SrcStride, bool enableSmallC0)	94	+ uint16_t loop4DstStride, uint64_t loop1SrcStride, uint8_t cacheMode, uint16_t nValue, uint32_t dValue,
		95	+ uint64_t loop4SrcStride, bool enableSmallC0)
88	{	96	{
89	if ASCEND_IS_AIV {	97	if ASCEND_IS_AIV {
90	return;	98	return;
	@@ -93,9 +101,10 @@ public:
93	uint64_t mte2NzPara = static_cast<uint64_t>(loop4DstStride) << 48; // MTE2_NZ_PARA[63:48]	101	uint64_t mte2NzPara = static_cast<uint64_t>(loop4DstStride) << 48; // MTE2_NZ_PARA[63:48]
94	mte2NzPara \|= static_cast<uint64_t>(loop3DstStride) << 32; // MTE2_NZ_PARA[47:32]	102	mte2NzPara \|= static_cast<uint64_t>(loop3DstStride) << 32; // MTE2_NZ_PARA[47:32]
95	mte2NzPara \|= static_cast<uint64_t>(loop2DstStride) << 16; // MTE2_NZ_PARA[31:16]	103	mte2NzPara \|= static_cast<uint64_t>(loop2DstStride) << 16; // MTE2_NZ_PARA[31:16]
96	- mte2NzPara \|= static_cast<uint64_t>(ndNum); // MTE2_NZ_PARA[15:0]	104	+ mte2NzPara \|= static_cast<uint64_t>(ndNum); // MTE2_NZ_PARA[15:0]
97	- SetMTE2NzPara(mte2NzPara); // CCE: store parameters for ND2NZ DMA instructions	105	+ SetMTE2NzPara(mte2NzPara); // CCE: store parameters for ND2NZ DMA instructions
98	- copy_gm_to_cbuf_multi_nd2nz(dst, src, 0, loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, enableSmallC0);	106	+ copy_gm_to_cbuf_multi_nd2nz(
		107	+ dst, src, 0, loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, enableSmallC0);
99	}	108	}
100	}	109	}
101	};	110	};
	@@ -103,10 +112,12 @@ public:
103	class CopyGmToCbufMultiDn2nzInstr {	112	class CopyGmToCbufMultiDn2nzInstr {
104	public:	113	public:
105	template <typename T, typename U, typename... Params>	114	template <typename T, typename U, typename... Params>
106	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params) {	115	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
		116	+ {
107	using srcType = typename U::elementType;	117	using srcType = typename U::elementType;
108	- if constexpr(sizeof(srcType) == sizeof(int8_t)) {	118	+ if constexpr (sizeof(srcType) == sizeof(int8_t)) {
109	- CopyGmToCbufMultiDn2nz((__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);	119	+ CopyGmToCbufMultiDn2nz(
		120	+ (__cbuf__ uint8_t)(dst.Data().Get()), (__gm__ uint8_t)(src.Data().Get()), params...);
110	} else if constexpr (sizeof(srcType) == sizeof(half)) {	121	} else if constexpr (sizeof(srcType) == sizeof(half)) {
111	CopyGmToCbufMultiDn2nz((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);	122	CopyGmToCbufMultiDn2nz((__cbuf__ half)(dst.Data().Get()), (__gm__ half)(src.Data().Get()), params...);
112	} else if constexpr (sizeof(srcType) == sizeof(float)) {	123	} else if constexpr (sizeof(srcType) == sizeof(float)) {
	@@ -115,9 +126,10 @@ public:
115	}	126	}
116		127
117	template <typename T>	128	template <typename T>
118	- __aicore__ inline static void CopyGmToCbufMultiDn2nz(__cbuf__ T* dst, __gm__ T* src, uint16_t dnNum, uint16_t loop2DstStride,	129	+ __aicore__ inline static void CopyGmToCbufMultiDn2nz(
119	- uint16_t loop3DstStride, uint16_t loop4DstStride, uint64_t loop1SrcStride, uint8_t cacheMode, uint16_t nValue,	130	+ __cbuf__ T* dst, __gm__ T* src, uint16_t dnNum, uint16_t loop2DstStride, uint16_t loop3DstStride,
120	- uint32_t dValue, uint64_t loop4SrcStride, bool enableSmallC0)	131	+ uint16_t loop4DstStride, uint64_t loop1SrcStride, uint8_t cacheMode, uint16_t nValue, uint32_t dValue,
		132	+ uint64_t loop4SrcStride, bool enableSmallC0)
121	{	133	{
122	if ASCEND_IS_AIV {	134	if ASCEND_IS_AIV {
123	return;	135	return;
	@@ -127,9 +139,10 @@ public:
127	uint64_t mte2NzPara = static_cast<uint64_t>(loop4DstStride) << 48; // MTE2_NZ_PARA[63:48]	139	uint64_t mte2NzPara = static_cast<uint64_t>(loop4DstStride) << 48; // MTE2_NZ_PARA[63:48]
128	mte2NzPara \|= static_cast<uint64_t>(loop3DstStride) << 32; // MTE2_NZ_PARA[47:32]	140	mte2NzPara \|= static_cast<uint64_t>(loop3DstStride) << 32; // MTE2_NZ_PARA[47:32]
129	mte2NzPara \|= static_cast<uint64_t>(loop2DstStride) << 16; // MTE2_NZ_PARA[31:16]	141	mte2NzPara \|= static_cast<uint64_t>(loop2DstStride) << 16; // MTE2_NZ_PARA[31:16]
130	- mte2NzPara \|= static_cast<uint64_t>(dnNum); // MTE2_NZ_PARA[15:0]	142	+ mte2NzPara \|= static_cast<uint64_t>(dnNum); // MTE2_NZ_PARA[15:0]
131	- SetMTE2NzPara(mte2NzPara); // CCE: store parameters for DN2NZ DMA instructions	143	+ SetMTE2NzPara(mte2NzPara); // CCE: store parameters for DN2NZ DMA instructions
132	- copy_gm_to_cbuf_multi_dn2nz(dst, src, 0, loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, enableSmallC0);	144	+ copy_gm_to_cbuf_multi_dn2nz(
		145	+ dst, src, 0, loop1SrcStride, cacheMode, nValue, dValue, loop4SrcStride, enableSmallC0);
133	}	146	}
134	}	147	}
135	};	148	};
	@@ -137,14 +150,16 @@ public:
137	class CopyL12BTInstr {	150	class CopyL12BTInstr {
138	public:	151	public:
139	template <typename T, typename U, typename... Params>	152	template <typename T, typename U, typename... Params>
140	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params) {	153	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
		154	+ {
141	CopyL12BT(reinterpret_cast<uint64_t>(dst.Data().Get()), src.Data().Get(), params...);	155	CopyL12BT(reinterpret_cast<uint64_t>(dst.Data().Get()), src.Data().Get(), params...);
142	}	156	}
143		157
144	private:	158	private:
145	template <typename T>	159	template <typename T>
146	- __aicore__ inline static void CopyL12BT(uint64_t dst, __cbuf__ T* src, bool convControl, uint16_t blockCount, uint16_t blockLen,	160	+ __aicore__ inline static void CopyL12BT(
147	- uint16_t srcStride, uint16_t dstStride)	161	+ uint64_t dst, __cbuf__ T* src, bool convControl, uint16_t blockCount, uint16_t blockLen, uint16_t srcStride,
		162	+ uint16_t dstStride)
148	{	163	{
149	if ASCEND_IS_AIV {	164	if ASCEND_IS_AIV {
150	return;	165	return;
	@@ -159,14 +174,15 @@ private:
159	class CopyL12FBInstr {	174	class CopyL12FBInstr {
160	public:	175	public:
161	template <typename T, typename U, typename... Params>	176	template <typename T, typename U, typename... Params>
162	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params) {	177	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
		178	+ {
163	CopyL12FB(reinterpret_cast<uint64_t>(dst.Data().Get()), src.Data().Get(), params...);	179	CopyL12FB(reinterpret_cast<uint64_t>(dst.Data().Get()), src.Data().Get(), params...);
164	}	180	}
165		181
166	private:	182	private:
167	template <typename T>	183	template <typename T>
168	- __aicore__ inline static void CopyL12FB(uint64_t dst, __cbuf__ T* src, uint16_t blockCount, uint16_t blockLen,	184	+ __aicore__ inline static void CopyL12FB(
169	- uint16_t srcStride, uint16_t dstStride)	185	+ uint64_t dst, __cbuf__ T* src, uint16_t blockCount, uint16_t blockLen, uint16_t srcStride, uint16_t dstStride)
170	{	186	{
171	if ASCEND_IS_AIV {	187	if ASCEND_IS_AIV {
172	return;	188	return;

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/fixpipe_impl.h+19-18

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/fixpipe_routing.h+12-11

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/fixpipe_utils.h+59-44

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/fixpipe/fixpipe_utils.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/fixpipe/fixpipe_utils.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -25,7 +25,7 @@
25	#include "impl/experimental/tensor_api/arch/utils/arch_utils.h"	25	#include "impl/experimental/tensor_api/arch/utils/arch_utils.h"
26		26
27	namespace AscendC {	27	namespace AscendC {
28	-namespace Te{	28	+namespace Te {
29		29
30	constexpr uint32_t MAIN_LOOP_N_SIZE_3510 = 512;	30	constexpr uint32_t MAIN_LOOP_N_SIZE_3510 = 512;
31	constexpr uint32_t CBURST_NUM_3510 = MAIN_LOOP_N_SIZE_3510 / BLOCK_CUBE;	31	constexpr uint32_t CBURST_NUM_3510 = MAIN_LOOP_N_SIZE_3510 / BLOCK_CUBE;
	@@ -34,8 +34,7 @@ template <typename T>
34	__aicore__ inline auto AllocTempBuf(const T& calNSize)	34	__aicore__ inline auto AllocTempBuf(const T& calNSize)
35	{	35	{
36	uint64_t deqTensorTempBuf = 0;	36	uint64_t deqTensorTempBuf = 0;
37	- if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\|	37	+ if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\| CURRENT_ARCH_VERSION == ArchVersion::V2201) {
38	- CURRENT_ARCH_VERSION == ArchVersion::V2201) {
39	deqTensorTempBuf = reinterpret_cast<uint64_t>(get_imm(0));	38	deqTensorTempBuf = reinterpret_cast<uint64_t>(get_imm(0));
40	}	39	}
41	return deqTensorTempBuf;	40	return deqTensorTempBuf;
	@@ -44,8 +43,7 @@ __aicore__ inline auto AllocTempBuf(const T& calNSize)
44	template <typename T>	43	template <typename T>
45	__aicore__ inline void SetFpc(const T& deqTensorTempBuf)	44	__aicore__ inline void SetFpc(const T& deqTensorTempBuf)
46	{	45	{
47	- if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\|	46	+ if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\| CURRENT_ARCH_VERSION == ArchVersion::V2201) {
48	- CURRENT_ARCH_VERSION == ArchVersion::V2201) {
49	uint64_t deqTensorAddr = (reinterpret_cast<uint64_t>(deqTensorTempBuf) >> static_cast<uint64_t>(7)) << 8;	47	uint64_t deqTensorAddr = (reinterpret_cast<uint64_t>(deqTensorTempBuf) >> static_cast<uint64_t>(7)) << 8;
50	set_fpc(deqTensorAddr);	48	set_fpc(deqTensorAddr);
51	}	49	}
	@@ -53,8 +51,7 @@ __aicore__ inline void SetFpc(const T& deqTensorTempBuf)
53		51
54	__aicore__ inline void InsertSync()	52	__aicore__ inline void InsertSync()
55	{	53	{
56	- if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\|	54	+ if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510 \|\| CURRENT_ARCH_VERSION == ArchVersion::V2201) {
57	- CURRENT_ARCH_VERSION == ArchVersion::V2201) {
58	pipe_barrier(PIPE_FIX);	55	pipe_barrier(PIPE_FIX);
59	}	56	}
60	}	57	}
	@@ -70,10 +67,13 @@ public:
70	DataCopyL12FB3510::Run<DEFAULT_DATA_COPY_TRAIT>(dst, tileSrc);	67	DataCopyL12FB3510::Run<DEFAULT_DATA_COPY_TRAIT>(dst, tileSrc);
71	SetFpc(dstAddr);	68	SetFpc(dstAddr);
72	}	69	}
		70	+
73	private:	71	private:
74	template <typename T>	72	template <typename T>
75	- __aicore__ inline static decltype(auto) TileSrcTensor(const T& src, uint16_t calNSize, uint16_t nIterIndex) {	73	+ __aicore__ inline static decltype(auto) TileSrcTensor(const T& src, uint16_t calNSize, uint16_t nIterIndex)
76	- auto coord = MakeCoord(MakeCoord(Std::Int<0>{}, Std::Int<0>{}), MakeCoord(Std::Int<0>{}, nIterIndex * MAIN_LOOP_N_SIZE_3510));	74	+ {
		75	+ auto coord = MakeCoord(
		76	+ MakeCoord(Std::Int<0>{}, Std::Int<0>{}), MakeCoord(Std::Int<0>{}, nIterIndex * MAIN_LOOP_N_SIZE_3510));
77	auto shape = MakeShape(MakeShape(Std::Int<1>{}, Std::Int<1>{}), MakeShape(Std::Int<1>{}, calNSize));	77	auto shape = MakeShape(MakeShape(Std::Int<1>{}, Std::Int<1>{}), MakeShape(Std::Int<1>{}, calNSize));
78	return src(coord, shape);	78	return src(coord, shape);
79	}	79	}
	@@ -89,78 +89,93 @@ __aicore__ inline constexpr QuantMode_t GetFixpipeQuantPre()
89	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	89	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
90	if constexpr (trait.roundMode == RoundMode::HYBRID) {	90	if constexpr (trait.roundMode == RoundMode::HYBRID) {
91	static_assert(	91	static_assert(
92	- (Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>),	92	+ (Std::is_same_v<srcType, __cc__ float> &&
		93	+ Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>),
93	"Only when L0CType is float and output Type is hifloat8_t support RoundMode::HYBRID in Fixpipe");	94	"Only when L0CType is float and output Type is hifloat8_t support RoundMode::HYBRID in Fixpipe");
94	}	95	}
95	if constexpr (isTensor) {	96	if constexpr (isTensor) {
96	- if constexpr (Std::is_same_v<srcType, __cc__ int32_t> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {	97	+ if constexpr (
		98	+ Std::is_same_v<srcType, __cc__ int32_t> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {
97	return QuantMode_t::VDEQF16;	99	return QuantMode_t::VDEQF16;
98	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	100	+ } else if constexpr (
		101	+ Std::is_same_v<srcType, __cc__ float> &&
99	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {	102	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {
100	return QuantMode_t::VQF322B8_PRE;	103	return QuantMode_t::VQF322B8_PRE;
101	- } else if constexpr (Std::is_same_v<srcType, __cc__ int32_t> &&	104	+ } else if constexpr (
		105	+ Std::is_same_v<srcType, __cc__ int32_t> &&
102	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {	106	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {
103	return QuantMode_t::VREQ8;	107	return QuantMode_t::VREQ8;
104	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	108	+ } else if constexpr (
		109	+ Std::is_same_v<srcType, __cc__ float> &&
105	Std::is_one_of_v<dstType, __gm__ fp8_e4m3fn_t, __ubuf__ fp8_e4m3fn_t>) {	110	Std::is_one_of_v<dstType, __gm__ fp8_e4m3fn_t, __ubuf__ fp8_e4m3fn_t>) {
106	return QuantMode_t::VQF322FP8_PRE;	111	return QuantMode_t::VQF322FP8_PRE;
107	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	112	+ } else if constexpr (
		113	+ Std::is_same_v<srcType, __cc__ float> &&
108	Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>) {	114	Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>) {
109	if constexpr (trait.roundMode == RoundMode::HYBRID) {	115	if constexpr (trait.roundMode == RoundMode::HYBRID) {
110	return QuantMode_t::VQF322HIF8_PRE_HYBRID;	116	return QuantMode_t::VQF322HIF8_PRE_HYBRID;
111	} else {	117	} else {
112	return QuantMode_t::VQF322HIF8_PRE;	118	return QuantMode_t::VQF322HIF8_PRE;
113	}	119	}
114	- } else if constexpr (Std::is_same_v<srcType, __cc__ int32_t> &&	120	+ } else if constexpr (
		121	+ Std::is_same_v<srcType, __cc__ int32_t> &&
115	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {	122	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {
116	return QuantMode_t::VQS322BF16_PRE;	123	return QuantMode_t::VQS322BF16_PRE;
117	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	124	+ } else if constexpr (
118	- Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {	125	+ Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {
119	return QuantMode_t::VQF322F16_PRE;	126	return QuantMode_t::VQF322F16_PRE;
120	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	127	+ } else if constexpr (
		128	+ Std::is_same_v<srcType, __cc__ float> &&
121	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {	129	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {
122	return QuantMode_t::VQF322BF16_PRE;	130	return QuantMode_t::VQF322BF16_PRE;
123	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	131	+ } else if constexpr (
124	- Std::is_one_of_v<dstType, __gm__ float, __ubuf__ float>) {	132	+ Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ float, __ubuf__ float>) {
125	return QuantMode_t::VQF322F32_PRE;	133	return QuantMode_t::VQF322F32_PRE;
126	}	134	}
127	} else if constexpr (isScalar) {	135	} else if constexpr (isScalar) {
128	- if constexpr (Std::is_same_v<srcType, __cc__ int32_t> &&	136	+ if constexpr (
129	- Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {	137	+ Std::is_same_v<srcType, __cc__ int32_t> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {
130	return QuantMode_t::DEQF16;	138	return QuantMode_t::DEQF16;
131	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	139	+ } else if constexpr (
		140	+ Std::is_same_v<srcType, __cc__ float> &&
132	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {	141	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {
133	return QuantMode_t::QF322B8_PRE;	142	return QuantMode_t::QF322B8_PRE;
134	- } else if constexpr (Std::is_same_v<srcType, __cc__ int32_t> &&	143	+ } else if constexpr (
		144	+ Std::is_same_v<srcType, __cc__ int32_t> &&
135	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {	145	Std::is_one_of_v<dstType, __gm__ uint8_t, __gm__ int8_t, __ubuf__ uint8_t, __ubuf__ int8_t>) {
136	return QuantMode_t::REQ8;	146	return QuantMode_t::REQ8;
137	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	147	+ } else if constexpr (
		148	+ Std::is_same_v<srcType, __cc__ float> &&
138	Std::is_one_of_v<dstType, __gm__ fp8_e4m3fn_t, __ubuf__ fp8_e4m3fn_t>) {	149	Std::is_one_of_v<dstType, __gm__ fp8_e4m3fn_t, __ubuf__ fp8_e4m3fn_t>) {
139	return QuantMode_t::QF322FP8_PRE;	150	return QuantMode_t::QF322FP8_PRE;
140	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	151	+ } else if constexpr (
		152	+ Std::is_same_v<srcType, __cc__ float> &&
141	Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>) {	153	Std::is_one_of_v<dstType, __gm__ hifloat8_t, __ubuf__ hifloat8_t>) {
142	if constexpr (trait.roundMode == RoundMode::HYBRID) {	154	if constexpr (trait.roundMode == RoundMode::HYBRID) {
143	return QuantMode_t::QF322HIF8_PRE_HYBRID;	155	return QuantMode_t::QF322HIF8_PRE_HYBRID;
144	} else {	156	} else {
145	return QuantMode_t::QF322HIF8_PRE;	157	return QuantMode_t::QF322HIF8_PRE;
146	}	158	}
147	- } else if constexpr (Std::is_same_v<srcType, __cc__ int32_t> &&	159	+ } else if constexpr (
		160	+ Std::is_same_v<srcType, __cc__ int32_t> &&
148	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {	161	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {
149	return QuantMode_t::QS322BF16_PRE;	162	return QuantMode_t::QS322BF16_PRE;
150	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	163	+ } else if constexpr (
151	- Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {	164	+ Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {
152	return QuantMode_t::QF322F16_PRE;	165	return QuantMode_t::QF322F16_PRE;
153	- } else if constexpr (Std::is_same_v<srcType, __cc__ float>	166	+ } else if constexpr (
154	- && Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {	167	+ Std::is_same_v<srcType, __cc__ float> &&
		168	+ Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {
155	return QuantMode_t::QF322BF16_PRE;	169	return QuantMode_t::QF322BF16_PRE;
156	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	170	+ } else if constexpr (
157	- Std::is_one_of_v<dstType, __gm__ float, __ubuf__ float>) {	171	+ Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ float, __ubuf__ float>) {
158	return QuantMode_t::QF322F32_PRE;	172	return QuantMode_t::QF322F32_PRE;
159	}	173	}
160	} else {	174	} else {
161	if constexpr (Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {	175	if constexpr (Std::is_same_v<srcType, __cc__ float> && Std::is_one_of_v<dstType, __gm__ half, __ubuf__ half>) {
162	return QuantMode_t::F322F16;	176	return QuantMode_t::F322F16;
163	- } else if constexpr (Std::is_same_v<srcType, __cc__ float> &&	177	+ } else if constexpr (
		178	+ Std::is_same_v<srcType, __cc__ float> &&
164	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {	179	Std::is_one_of_v<dstType, __gm__ bfloat16_t, __ubuf__ bfloat16_t>) {
165	return QuantMode_t::F322BF16;	180	return QuantMode_t::F322BF16;
166	} else {	181	} else {

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_l0c2out.h+13-11

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_l0c2out/nz2dn.h+25-22

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_l0c2out/nz2nd.h+25-22

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_l0c2out/nz2nz.h+26-23

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_quant_l0c2out.h+17-15

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_quant_l0c2out/nz2dn.h+33-32

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_quant_l0c2out/nz2nd.h+38-35

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/fixpipe_quant_l0c2out/nz2nz.h+34-34

Mimpl/experimental/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/instruction.h+33-31

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/fixpipe/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -31,66 +31,67 @@ namespace Te {
31	class CopyMatrixCcToGm3510 {	31	class CopyMatrixCcToGm3510 {
32	public:	32	public:
33	template <QuantMode_t quantPre, typename T, typename U, typename... Params>	33	template <QuantMode_t quantPre, typename T, typename U, typename... Params>
34	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params)	34	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
35	{	35	{
36	CopyMatrixCcToGm<quantPre>(dst.Data().Get(), src.Data().Get(), params...);	36	CopyMatrixCcToGm<quantPre>(dst.Data().Get(), src.Data().Get(), params...);
37	}	37	}
38		38
39	private:	39	private:
40	template <QuantMode_t quantPre, typename T, typename U>	40	template <QuantMode_t quantPre, typename T, typename U>
41	- __aicore__ inline static void CopyMatrixCcToGm(__gm__ T dst, __cc__ U src, uint32_t nSize, uint32_t mSize,	41	+ __aicore__ inline static void CopyMatrixCcToGm(
42	- uint32_t srcStride, uint32_t dstStride, uint8_t cacheMode, bool reluEn, uint8_t unitFlag, bool isChannelSplit,	42	+ __gm__ T* dst, __cc__ U* src, uint32_t nSize, uint32_t mSize, uint32_t srcStride, uint32_t dstStride,
43	- bool nz2ndEn, bool nz2dnEn)	43	+ uint8_t cacheMode, bool reluEn, uint8_t unitFlag, bool isChannelSplit, bool nz2ndEn, bool nz2dnEn)
44	{	44	{
45	if ASCEND_IS_AIV {	45	if ASCEND_IS_AIV {
46	return;	46	return;
47	}	47	}
48	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {	48	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {
49	- copy_matrix_cc_to_gm(dst, src, 0, nSize, mSize, dstStride, srcStride, cacheMode, 0, unitFlag, static_cast<uint64_t>(quantPre),	49	+ copy_matrix_cc_to_gm(
50	- reluEn, isChannelSplit, nz2ndEn, static_cast<uint64_t>(QuantMode_post::NoConv), 0, false, false, 0, false, false, false, false, false,	50	+ dst, src, 0, nSize, mSize, dstStride, srcStride, cacheMode, 0, unitFlag,
51	- nz2dnEn);	51	+ static_cast<uint64_t>(quantPre), reluEn, isChannelSplit, nz2ndEn,
		52	+ static_cast<uint64_t>(QuantMode_post::NoConv), 0, false, false, 0, false, false, false, false, false,
		53	+ nz2dnEn);
52	}	54	}
53	}	55	}
54	};	56	};
55		57
56	class CopyMatrixCcToUb3510 {	58	class CopyMatrixCcToUb3510 {
57	public:	59	public:
58	-template <QuantMode_t quantPre, typename T, typename U, typename... Params>	60	+ template <QuantMode_t quantPre, typename T, typename U, typename... Params>
59	- __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params& ...params)	61	+ __aicore__ inline static void DataCopy(const T& dst, const U& src, const Params&... params)
60	{	62	{
61	CopyMatrixCcToUb<quantPre>(dst.Data().Get(), src.Data().Get(), params...);	63	CopyMatrixCcToUb<quantPre>(dst.Data().Get(), src.Data().Get(), params...);
62	}	64	}
63		65
64	private:	66	private:
65	-
66	template <QuantMode_t quantPre, typename T, typename U>	67	template <QuantMode_t quantPre, typename T, typename U>
67	- __aicore__ inline static void CopyMatrixCcToUb(__ubuf__ T dst, __cc__ U src, uint32_t nSize, uint32_t mSize,	68	+ __aicore__ inline static void CopyMatrixCcToUb(
68	- uint32_t srcStride, uint32_t dstStride, uint8_t dualDstCtl, bool reluEn, uint8_t unitFlag, bool subBlockId,	69	+ __ubuf__ T* dst, __cc__ U* src, uint32_t nSize, uint32_t mSize, uint32_t srcStride, uint32_t dstStride,
69	- bool nz2ndEn, bool nz2dnEn)	70	+ uint8_t dualDstCtl, bool reluEn, uint8_t unitFlag, bool subBlockId, bool nz2ndEn, bool nz2dnEn)
70	{	71	{
71	if ASCEND_IS_AIV {	72	if ASCEND_IS_AIV {
72	return;	73	return;
73	}	74	}
74	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {	75	if constexpr (CURRENT_ARCH_VERSION == ArchVersion::V3510) {
75	- copy_matrix_cc_to_ub(dst, src, 0, nSize, mSize, dstStride, srcStride, dualDstCtl, subBlockId, 0, unitFlag, static_cast<uint64_t>(quantPre),	76	+ copy_matrix_cc_to_ub(
76	- reluEn, false, nz2ndEn, static_cast<uint64_t>(QuantMode_post::NoConv), 0, false, false, 0, false, false, false, false, false,	77	+ dst, src, 0, nSize, mSize, dstStride, srcStride, dualDstCtl, subBlockId, 0, unitFlag,
77	- nz2dnEn);	78	+ static_cast<uint64_t>(quantPre), reluEn, false, nz2ndEn, static_cast<uint64_t>(QuantMode_post::NoConv),
		79	+ 0, false, false, 0, false, false, false, false, false, nz2dnEn);
78	}	80	}
79	}	81	}
80	};	82	};
81		83
82	-
83	class SetRegister3510 {	84	class SetRegister3510 {
84	public:	85	public:
85	template <typename... Params>	86	template <typename... Params>
86	- __aicore__ inline static void SetRegister(const uint64_t& quant, const Params& ...params)	87	+ __aicore__ inline static void SetRegister(const uint64_t& quant, const Params&... params)
87	{	88	{
88	SetQuantPre(quant);	89	SetQuantPre(quant);
89	SetParamsToRegister<uint64_t>(params...);	90	SetParamsToRegister<uint64_t>(params...);
90	}	91	}
91		92
92	template <typename... Params>	93	template <typename... Params>
93	- __aicore__ inline static void SetRegister(const Params& ...params)	94	+ __aicore__ inline static void SetRegister(const Params&... params)
94	{	95	{
95	SetParamsToRegister<uint64_t>(params...);	96	SetParamsToRegister<uint64_t>(params...);
96	}	97	}
	@@ -122,7 +123,8 @@ private:
122	}	123	}
123		124
124	template <typename T>	125	template <typename T>
125	- __aicore__ inline static void SetParamsToRegister(uint32_t dnNum, uint32_t dstDNStride, uint32_t srcNZMatrixStride, uint32_t srcNZC0Stride)	126	+ __aicore__ inline static void SetParamsToRegister(
		127	+ uint32_t dnNum, uint32_t dstDNStride, uint32_t srcNZMatrixStride, uint32_t srcNZC0Stride)
126	{	128	{
127	if ASCEND_IS_AIV {	129	if ASCEND_IS_AIV {
128	return;	130	return;
	@@ -138,8 +140,8 @@ private:
138	}	140	}
139	};	141	};
140		142
141	-}	143	+} // namespace Te
142	-}	144	+} // namespace AscendC
143		145
144	#endif // IMPL_TENSOR_API_ARCH_CUBE_DATAMOVE_FIXPIPE_NPU_ARCH_3510_INSTRUCTION_H	146	#endif // IMPL_TENSOR_API_ARCH_CUBE_DATAMOVE_FIXPIPE_NPU_ARCH_3510_INSTRUCTION_H
145		147

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/load_data_impl.h+15-15

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/load_data_routing.h+16-19

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/instruction.h+25-21

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/instruction.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -31,15 +31,16 @@ namespace Te {
31	class LoadCbufToCa3510 {	31	class LoadCbufToCa3510 {
32	public:	32	public:
33	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>	33	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>
34	- __aicore__ inline static void LoadData(const T& dst, const U& src, const Params& ...params)	34	+ __aicore__ inline static void LoadData(const T& dst, const U& src, const Params&... params)
35	{	35	{
36	LoadCbufToCa<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);	36	LoadCbufToCa<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);
37	}	37	}
38		38
39	private:	39	private:
40	template <bool transpose, typename T>	40	template <bool transpose, typename T>
41	- __aicore__ inline static void LoadCbufToCa(__ca__ T* dst, __cbuf__ T* src, uint16_t mStartPosition,	41	+ __aicore__ inline static void LoadCbufToCa(
42	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	42	+ __ca__ T* dst, __cbuf__ T* src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		43	+ int16_t srcStride, uint16_t dstStride)
43	{	44	{
44	if ASCEND_IS_AIV {	45	if ASCEND_IS_AIV {
45	return;	46	return;
	@@ -53,15 +54,16 @@ private:
53	class LoadCbufToCaS43510 {	54	class LoadCbufToCaS43510 {
54	public:	55	public:
55	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>	56	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>
56	- __aicore__ inline static void LoadData(const T& dst, const U& src, const Params& ...params)	57	+ __aicore__ inline static void LoadData(const T& dst, const U& src, const Params&... params)
57	{	58	{
58	LoadCbufToCa<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);	59	LoadCbufToCa<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);
59	}	60	}
60		61
61	private:	62	private:
62	template <bool transpose, typename T>	63	template <bool transpose, typename T>
63	- __aicore__ inline static void LoadCbufToCa(__ca__ T* dst, __cbuf__ T* src, uint16_t mStartPosition,	64	+ __aicore__ inline static void LoadCbufToCa(
64	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	65	+ __ca__ T* dst, __cbuf__ T* src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		66	+ int16_t srcStride, uint16_t dstStride)
65	{	67	{
66	if ASCEND_IS_AIV {	68	if ASCEND_IS_AIV {
67	return;	69	return;
	@@ -75,15 +77,16 @@ private:
75	class LoadCbufToCb3510 {	77	class LoadCbufToCb3510 {
76	public:	78	public:
77	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>	79	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>
78	- __aicore__ inline static void LoadData(const T& dst, const U& src, const Params& ...params)	80	+ __aicore__ inline static void LoadData(const T& dst, const U& src, const Params&... params)
79	{	81	{
80	LoadCbufToCb<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);	82	LoadCbufToCb<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);
81	}	83	}
82		84
83	private:	85	private:
84	template <bool transpose, typename T>	86	template <bool transpose, typename T>
85	- __aicore__ inline static void LoadCbufToCb(__cb__ T* dst, __cbuf__ T* src, uint16_t mStartPosition,	87	+ __aicore__ inline static void LoadCbufToCb(
86	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	88	+ __cb__ T* dst, __cbuf__ T* src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		89	+ int16_t srcStride, uint16_t dstStride)
87	{	90	{
88	if ASCEND_IS_AIV {	91	if ASCEND_IS_AIV {
89	return;	92	return;
	@@ -97,15 +100,16 @@ private:
97	class LoadCbufToCbS43510 {	100	class LoadCbufToCbS43510 {
98	public:	101	public:
99	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>	102	template <const LoadDataTrait& trait, typename T, typename U, typename... Params>
100	- __aicore__ inline static void LoadData(const T& dst, const U& src, const Params& ...params)	103	+ __aicore__ inline static void LoadData(const T& dst, const U& src, const Params&... params)
101	{	104	{
102	LoadCbufToCb<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);	105	LoadCbufToCb<trait.transposed>(dst.Data().Get(), src.Data().Get(), params...);
103	}	106	}
104		107
105	private:	108	private:
106	template <bool transpose, typename T>	109	template <bool transpose, typename T>
107	- __aicore__ inline static void LoadCbufToCb(__cb__ T* dst, __cbuf__ T* src, uint16_t mStartPosition,	110	+ __aicore__ inline static void LoadCbufToCb(
108	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	111	+ __cb__ T* dst, __cbuf__ T* src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		112	+ int16_t srcStride, uint16_t dstStride)
109	{	113	{
110	if ASCEND_IS_AIV {	114	if ASCEND_IS_AIV {
111	return;	115	return;

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a.h+17-13

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz.h+20-15

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -30,11 +30,12 @@ namespace Te {
30	class LoadDataL12L0ANZ2NZ3510 {	30	class LoadDataL12L0ANZ2NZ3510 {
31	public:	31	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	32	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	33	+ __aicore__ inline static void Run(const T& dst, const U& src)
		34	+ {
34	LoadDataImpl<TransTrait<trait, false>, T, U>(dst, src);	35	LoadDataImpl<TransTrait<trait, false>, T, U>(dst, src);
35	}	36	}
36		37
37	-private:	38	+private:
38	template <const LoadDataTrait& trait, typename T, typename U>	39	template <const LoadDataTrait& trait, typename T, typename U>
39	__aicore__ inline static constexpr void CheckTemplate()	40	__aicore__ inline static constexpr void CheckTemplate()
40	{	41	{
	@@ -57,12 +58,16 @@ private:
57	// Nz -> Nz	58	// Nz -> Nz
58	constexpr uint32_t KHALF = 2;	59	constexpr uint32_t KHALF = 2;
59	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	60	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
60	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;	61	+ auto srcStride =
61	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;	62	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;
		63	+ auto dstStride =
		64	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;
62	if constexpr (is_b4_type<DstType>) {	65	if constexpr (is_b4_type<DstType>) {
63	- LoadCbufToCaS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	66	+ LoadCbufToCaS43510::LoadData<trait>(
		67	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
64	} else {	68	} else {
65	- LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	69	+ LoadCbufToCa3510::LoadData<trait>(
		70	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
66	}	71	}
67	}	72	}
68	};	73	};

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz_with_coord.h+19-15

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/nz2nz_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0ANZ2NZWithCoord3510 {	29	class LoadDataL12L0ANZ2NZWithCoord3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>	31	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>
33	- __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord)
		33	+ {
34	LoadDataImpl<TransTrait<trait, false>, T, U, Coord>(dst, src, coord);	34	LoadDataImpl<TransTrait<trait, false>, T, U, Coord>(dst, src, coord);
35	}	35	}
36		36
	@@ -58,12 +58,16 @@ private:
58	// Nz -> Nz	58	// Nz -> Nz
59	constexpr uint32_t KHALF = 2;	59	constexpr uint32_t KHALF = 2;
60	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	60	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
61	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;	61	+ auto srcStride =
62	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;	62	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;
		63	+ auto dstStride =
		64	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;
63	if constexpr (is_b4_type<DstType>) {	65	if constexpr (is_b4_type<DstType>) {
64	- LoadCbufToCaS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition / KHALF, mStep, kStep, srcStride, dstStride);	66	+ LoadCbufToCaS43510::LoadData<trait>(
		67	+ dst, src, mStartPosition, kStartPosition / KHALF, mStep, kStep, srcStride, dstStride);
65	} else {	68	} else {
66	- LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	69	+ LoadCbufToCa3510::LoadData<trait>(
		70	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
67	}	71	}
68	}	72	}
69	};	73	};

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nz.h+20-16

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nz.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0AZN2NZ3510 {	29	class LoadDataL12L0AZN2NZ3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	31	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);	34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);
35	}	35	}
36		36
	@@ -39,7 +39,7 @@ private:
39	__aicore__ inline static constexpr void CheckTemplate()	39	__aicore__ inline static constexpr void CheckTemplate()
40	{	40	{
41	CheckFormat::CheckNZTemplate<T>();	41	CheckFormat::CheckNZTemplate<T>();
42	- CheckFormat::CheckZNTemplate<U>();	42	+ CheckFormat::CheckZNTemplate<U>();
43	CheckDataTypeFor3510::CheckL12L0ADataType<T, U>();	43	CheckDataTypeFor3510::CheckL12L0ADataType<T, U>();
44	}	44	}
45		45
	@@ -53,13 +53,17 @@ private:
53	uint16_t mStartPosition = 0;	53	uint16_t mStartPosition = 0;
54	uint16_t kStartPosition = 0;	54	uint16_t kStartPosition = 0;
55	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	55	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
56	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) / FRACTAL_FIXED;	56	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) /
		57	+ FRACTAL_FIXED;
57	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *	58	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *
58	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout) / C0_ELEMENT<DstType>;	59	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout) /
		60	+ C0_ELEMENT<DstType>;
59	// Zn -> Nz	61	// Zn -> Nz
60	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	62	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
61	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;	63	+ auto srcStride =
62	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;	64	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;
		65	+ auto dstStride =
		66	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;
63	LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	67	LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
64	}	68	}
65	};	69	};

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nz_with_coord.h+20-15

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4.h+40-28

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0AZN2NZB8B43510 {	29	class LoadDataL12L0AZN2NZB8B43510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	31	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);	34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);
35	}	35	}
36		36
	@@ -43,9 +43,10 @@ private:
43	CheckDataTypeFor3510::CheckL12L0ADataType<T, U>();	43	CheckDataTypeFor3510::CheckL12L0ADataType<T, U>();
44	}	44	}
45		45
46	- template <const LoadDataTrait& trait, typename T, typename U>	46	+ template <const LoadDataTrait& trait, typename T, typename U>
47	- __aicore__ inline static void LoadDataImplB4(const T& dst, const U& src, uint16_t mStartPosition,	47	+ __aicore__ inline static void LoadDataImplB4(
48	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	48	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		49	+ int16_t srcStride, uint16_t dstStride)
49	{	50	{
50	using DstType = typename T::elementType;	51	using DstType = typename T::elementType;
51	auto dstLayout = dst.Layout();	52	auto dstLayout = dst.Layout();
	@@ -55,14 +56,16 @@ private:
55	mStep = M_STEP_MIN_VAL_B4;	56	mStep = M_STEP_MIN_VAL_B4;
56	for (uint16_t idx = 0; idx < mLoop; ++idx) {	57	for (uint16_t idx = 0; idx < mLoop; ++idx) {
57	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));	58	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));
58	- LoadCbufToCaS43510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	59	+ LoadCbufToCaS43510::LoadData<trait>(
		60	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
59	mStartPosition += M_STEP_MIN_VAL_B4;	61	mStartPosition += M_STEP_MIN_VAL_B4;
60	}	62	}
61	}	63	}
62		64
63	template <const LoadDataTrait& trait, typename T, typename U>	65	template <const LoadDataTrait& trait, typename T, typename U>
64	- __aicore__ inline static void LoadDataImplB8(const T& dst, const U& src, uint16_t mStartPosition,	66	+ __aicore__ inline static void LoadDataImplB8(
65	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	67	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		68	+ int16_t srcStride, uint16_t dstStride)
66	{	69	{
67	using DstType = typename T::elementType;	70	using DstType = typename T::elementType;
68	auto dstLayout = dst.Layout();	71	auto dstLayout = dst.Layout();
	@@ -72,7 +75,8 @@ private:
72	mStep = M_STEP_MIN_VAL_B8;	75	mStep = M_STEP_MIN_VAL_B8;
73	for (uint16_t idx = 0; idx < mLoop; ++idx) {	76	for (uint16_t idx = 0; idx < mLoop; ++idx) {
74	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));	77	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));
75	- LoadCbufToCa3510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	78	+ LoadCbufToCa3510::LoadData<trait>(
		79	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
76	mStartPosition += M_STEP_MIN_VAL_B8;	80	mStartPosition += M_STEP_MIN_VAL_B8;
77	}	81	}
78	}	82	}
	@@ -87,28 +91,36 @@ private:
87	uint16_t mStartPosition = 0;	91	uint16_t mStartPosition = 0;
88	uint16_t kStartPosition = 0;	92	uint16_t kStartPosition = 0;
89	auto m1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	93	auto m1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
90	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) -	94	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) -
91	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *	95	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *
92	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout);	96	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout);
93	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	97	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
94	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) / FRACTAL_FIXED;	98	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) /
		99	+ FRACTAL_FIXED;
95	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	100	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
96	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) / C0_ELEMENT<DstType>;	101	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) /
		102	+ C0_ELEMENT<DstType>;
97	// Zn -> Nz	103	// Zn -> Nz
98	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	104	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
99	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;	105	+ auto srcStride =
100	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;	106	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;
		107	+ auto dstStride =
		108	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;
101	if constexpr (is_b4_type<DstType>) {	109	if constexpr (is_b4_type<DstType>) {
102	if (m1 < FRACTAL_FIXED) {	110	if (m1 < FRACTAL_FIXED) {
103	- LoadCbufToCaS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	111	+ LoadCbufToCaS43510::LoadData<trait>(
		112	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
104	} else {	113	} else {
105	- LoadDataImplB4<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	114	+ LoadDataImplB4<trait, T, U>(
		115	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
106	}	116	}
107	} else {	117	} else {
108	if (m1 < FRACTAL_FIXED) {	118	if (m1 < FRACTAL_FIXED) {
109	- LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	119	+ LoadCbufToCa3510::LoadData<trait>(
		120	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
110	} else {	121	} else {
111	- LoadDataImplB8<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	122	+ LoadDataImplB8<trait, T, U>(
		123	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
112	}	124	}
113	}	125	}
114	}	126	}

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4_with_coord.h+42-27

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0a/zn2nzb8b4_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0AZN2NZB8B4WithCoord3510 {	29	class LoadDataL12L0AZN2NZB8B4WithCoord3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>	31	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>
33	- __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U, Coord>(dst, src, coord);	34	LoadDataImpl<TransTrait<trait, true>, T, U, Coord>(dst, src, coord);
35	}	35	}
36		36
	@@ -44,8 +44,9 @@ private:
44	}	44	}
45		45
46	template <const LoadDataTrait& trait, typename T, typename U>	46	template <const LoadDataTrait& trait, typename T, typename U>
47	- __aicore__ inline static void LoadDataImplB4(const T& dst, const U& src, uint16_t mStartPosition,	47	+ __aicore__ inline static void LoadDataImplB4(
48	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	48	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		49	+ int16_t srcStride, uint16_t dstStride)
49	{	50	{
50	using DstType = typename T::elementType;	51	using DstType = typename T::elementType;
51	auto dstLayout = dst.Layout();	52	auto dstLayout = dst.Layout();
	@@ -55,14 +56,16 @@ private:
55	mStep = M_STEP_MIN_VAL_B4;	56	mStep = M_STEP_MIN_VAL_B4;
56	for (uint16_t idx = 0; idx < mLoop; ++idx) {	57	for (uint16_t idx = 0; idx < mLoop; ++idx) {
57	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));	58	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));
58	- LoadCbufToCaS43510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	59	+ LoadCbufToCaS43510::LoadData<trait>(
		60	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
59	mStartPosition += M_STEP_MIN_VAL_B4;	61	mStartPosition += M_STEP_MIN_VAL_B4;
60	}	62	}
61	}	63	}
62		64
63	template <const LoadDataTrait& trait, typename T, typename U>	65	template <const LoadDataTrait& trait, typename T, typename U>
64	- __aicore__ inline static void LoadDataImplB8(const T& dst, const U& src, uint16_t mStartPosition,	66	+ __aicore__ inline static void LoadDataImplB8(
65	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	67	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		68	+ int16_t srcStride, uint16_t dstStride)
66	{	69	{
67	using DstType = typename T::elementType;	70	using DstType = typename T::elementType;
68	auto dstLayout = dst.Layout();	71	auto dstLayout = dst.Layout();
	@@ -72,7 +75,8 @@ private:
72	mStep = M_STEP_MIN_VAL_B8;	75	mStep = M_STEP_MIN_VAL_B8;
73	for (uint16_t idx = 0; idx < mLoop; ++idx) {	76	for (uint16_t idx = 0; idx < mLoop; ++idx) {
74	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));	77	auto sliceDst = dst(MakeCoord(MakeCoord(0, 0), MakeCoord(0, idx)));
75	- LoadCbufToCa3510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	78	+ LoadCbufToCa3510::LoadData<trait>(
		79	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
76	mStartPosition += M_STEP_MIN_VAL_B8;	80	mStartPosition += M_STEP_MIN_VAL_B8;
77	}	81	}
78	}	82	}
	@@ -87,28 +91,39 @@ private:
87	uint16_t mStartPosition = Std::get<1>(coord) / FRACTAL_FIXED;	91	uint16_t mStartPosition = Std::get<1>(coord) / FRACTAL_FIXED;
88	uint16_t kStartPosition = Std::get<0>(coord) / C0_ELEMENT<DstType>;	92	uint16_t kStartPosition = Std::get<0>(coord) / C0_ELEMENT<DstType>;
89	auto m1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	93	auto m1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
90	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) -	94	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) -
91	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *	95	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout) *
92	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout) - Std::get<0>(coord);	96	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(dstLayout) -
		97	+ Std::get<0>(coord);
93	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	98	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
94	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) / FRACTAL_FIXED - mStartPosition;	99	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) /
		100	+ FRACTAL_FIXED -
		101	+ mStartPosition;
95	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	102	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
96	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) / C0_ELEMENT<DstType> - kStartPosition;	103	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) /
		104	+ C0_ELEMENT<DstType> -
		105	+ kStartPosition;
97	// Zn -> Nz	106	// Zn -> Nz
98	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	107	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
99	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;	108	+ auto srcStride =
100	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;	109	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;
		110	+ auto dstStride =
		111	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(dstLayout) / STRIDE_UNIT;
101	if constexpr (is_b4_type<DstType>) {	112	if constexpr (is_b4_type<DstType>) {
102	if (m1 < FRACTAL_FIXED) {	113	if (m1 < FRACTAL_FIXED) {
103	- LoadCbufToCaS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	114	+ LoadCbufToCaS43510::LoadData<trait>(
		115	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
104	} else {	116	} else {
105	- LoadDataImplB4<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	117	+ LoadDataImplB4<trait, T, U>(
		118	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
106	}	119	}
107	} else {	120	} else {
108	if (m1 < FRACTAL_FIXED) {	121	if (m1 < FRACTAL_FIXED) {
109	- LoadCbufToCa3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	122	+ LoadCbufToCa3510::LoadData<trait>(
		123	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
110	} else {	124	} else {
111	- LoadDataImplB8<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	125	+ LoadDataImplB8<trait, T, U>(
		126	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
112	}	127	}
113	}	128	}
114	}	129	}

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b.h+17-13

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2zn.h+19-15

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0BNZ2ZN3510 {	29	class LoadDataL12L0BNZ2ZN3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	31	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);	34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);
35	}	35	}
36		36
	@@ -53,13 +53,17 @@ private:
53	auto mStartPosition = 0;	53	auto mStartPosition = 0;
54	auto kStartPosition = 0;	54	auto kStartPosition = 0;
55	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	55	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
56	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) / FRACTAL_FIXED;	56	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) /
		57	+ FRACTAL_FIXED;
57	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *	58	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *
58	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout) / C0_ELEMENT<DstType>;	59	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout) /
		60	+ C0_ELEMENT<DstType>;
59	// Nz -> Zn	61	// Nz -> Zn
60	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	62	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
61	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;	63	+ auto srcStride =
62	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;	64	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;
		65	+ auto dstStride =
		66	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;
63	LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	67	LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
64	}	68	}
65	};	69	};

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2zn_with_coord.h+20-15

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4.h+39-27

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0BNZ2ZNB8B43510 {	29	class LoadDataL12L0BNZ2ZNB8B43510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	31	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);	34	LoadDataImpl<TransTrait<trait, true>, T, U>(dst, src);
35	}	35	}
36		36
	@@ -44,8 +44,9 @@ private:
44	}	44	}
45		45
46	template <const LoadDataTrait& trait, typename T, typename U>	46	template <const LoadDataTrait& trait, typename T, typename U>
47	- __aicore__ inline static void LoadDataImplB4(const T& dst, const U& src, uint16_t mStartPosition,	47	+ __aicore__ inline static void LoadDataImplB4(
48	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	48	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		49	+ int16_t srcStride, uint16_t dstStride)
49	{	50	{
50	using DstType = typename T::elementType;	51	using DstType = typename T::elementType;
51	auto dstLayout = dst.Layout();	52	auto dstLayout = dst.Layout();
	@@ -55,14 +56,16 @@ private:
55	mStep = M_STEP_MIN_VAL_B4;	56	mStep = M_STEP_MIN_VAL_B4;
56	for (uint16_t idx = 0; idx < nLoop; ++idx) {	57	for (uint16_t idx = 0; idx < nLoop; ++idx) {
57	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));	58	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));
58	- LoadCbufToCbS43510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	59	+ LoadCbufToCbS43510::LoadData<trait>(
		60	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
59	mStartPosition += M_STEP_MIN_VAL_B4;	61	mStartPosition += M_STEP_MIN_VAL_B4;
60	}	62	}
61	}	63	}
62		64
63	template <const LoadDataTrait& trait, typename T, typename U>	65	template <const LoadDataTrait& trait, typename T, typename U>
64	- __aicore__ inline static void LoadDataImplB8(const T& dst, const U& src, uint16_t mStartPosition,	66	+ __aicore__ inline static void LoadDataImplB8(
65	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	67	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		68	+ int16_t srcStride, uint16_t dstStride)
66	{	69	{
67	using DstType = typename T::elementType;	70	using DstType = typename T::elementType;
68	auto dstLayout = dst.Layout();	71	auto dstLayout = dst.Layout();
	@@ -72,7 +75,8 @@ private:
72	mStep = M_STEP_MIN_VAL_B8;	75	mStep = M_STEP_MIN_VAL_B8;
73	for (uint16_t idx = 0; idx < nLoop; ++idx) {	76	for (uint16_t idx = 0; idx < nLoop; ++idx) {
74	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));	77	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));
75	- LoadCbufToCb3510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	78	+ LoadCbufToCb3510::LoadData<trait>(
		79	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
76	mStartPosition += M_STEP_MIN_VAL_B8;	80	mStartPosition += M_STEP_MIN_VAL_B8;
77	}	81	}
78	}	82	}
	@@ -87,28 +91,36 @@ private:
87	auto mStartPosition = 0;	91	auto mStartPosition = 0;
88	auto kStartPosition = 0;	92	auto kStartPosition = 0;
89	auto n1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	93	auto n1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
90	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) -	94	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) -
91	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *	95	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *
92	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout);	96	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout);
93	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	97	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
94	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) / FRACTAL_FIXED;	98	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) /
		99	+ FRACTAL_FIXED;
95	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	100	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
96	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) / C0_ELEMENT<DstType>;	101	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) /
		102	+ C0_ELEMENT<DstType>;
97	// Nz -> Zn	103	// Nz -> Zn
98	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	104	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
99	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;	105	+ auto srcStride =
100	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;	106	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;
		107	+ auto dstStride =
		108	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;
101	if constexpr (is_b4_type<DstType>) {	109	if constexpr (is_b4_type<DstType>) {
102	if (n1 < FRACTAL_FIXED) {	110	if (n1 < FRACTAL_FIXED) {
103	- LoadCbufToCbS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	111	+ LoadCbufToCbS43510::LoadData<trait>(
		112	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
104	} else {	113	} else {
105	- LoadDataImplB4<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	114	+ LoadDataImplB4<trait, T, U>(
		115	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
106	}	116	}
107	} else {	117	} else {
108	if (n1 < FRACTAL_FIXED) {	118	if (n1 < FRACTAL_FIXED) {
109	- LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	119	+ LoadCbufToCb3510::LoadData<trait>(
		120	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
110	} else {	121	} else {
111	- LoadDataImplB8<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	122	+ LoadDataImplB8<trait, T, U>(
		123	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
112	}	124	}
113	}	125	}
114	}	126	}

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4_with_coord.h+42-27

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/nz2znb8b4_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0BNZ2ZNB8B4WithCoord3510 {	29	class LoadDataL12L0BNZ2ZNB8B4WithCoord3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>	31	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>
33	- __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord)
		33	+ {
34	LoadDataImpl<TransTrait<trait, true>, T, U, Coord>(dst, src, coord);	34	LoadDataImpl<TransTrait<trait, true>, T, U, Coord>(dst, src, coord);
35	}	35	}
36		36
	@@ -44,8 +44,9 @@ private:
44	}	44	}
45		45
46	template <const LoadDataTrait& trait, typename T, typename U>	46	template <const LoadDataTrait& trait, typename T, typename U>
47	- __aicore__ inline static void LoadDataImplB4(const T& dst, const U& src, uint16_t mStartPosition,	47	+ __aicore__ inline static void LoadDataImplB4(
48	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	48	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		49	+ int16_t srcStride, uint16_t dstStride)
49	{	50	{
50	using DstType = typename T::elementType;	51	using DstType = typename T::elementType;
51	auto dstLayout = dst.Layout();	52	auto dstLayout = dst.Layout();
	@@ -55,14 +56,16 @@ private:
55	mStep = M_STEP_MIN_VAL_B4;	56	mStep = M_STEP_MIN_VAL_B4;
56	for (uint16_t idx = 0; idx < nLoop; ++idx) {	57	for (uint16_t idx = 0; idx < nLoop; ++idx) {
57	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));	58	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));
58	- LoadCbufToCbS43510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	59	+ LoadCbufToCbS43510::LoadData<trait>(
		60	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
59	mStartPosition += M_STEP_MIN_VAL_B4;	61	mStartPosition += M_STEP_MIN_VAL_B4;
60	}	62	}
61	}	63	}
62		64
63	template <const LoadDataTrait& trait, typename T, typename U>	65	template <const LoadDataTrait& trait, typename T, typename U>
64	- __aicore__ inline static void LoadDataImplB8(const T& dst, const U& src, uint16_t mStartPosition,	66	+ __aicore__ inline static void LoadDataImplB8(
65	- uint16_t kStartPosition, uint8_t mStep, uint8_t kStep, int16_t srcStride, uint16_t dstStride)	67	+ const T& dst, const U& src, uint16_t mStartPosition, uint16_t kStartPosition, uint8_t mStep, uint8_t kStep,
		68	+ int16_t srcStride, uint16_t dstStride)
66	{	69	{
67	using DstType = typename T::elementType;	70	using DstType = typename T::elementType;
68	auto dstLayout = dst.Layout();	71	auto dstLayout = dst.Layout();
	@@ -72,7 +75,8 @@ private:
72	mStep = M_STEP_MIN_VAL_B8;	75	mStep = M_STEP_MIN_VAL_B8;
73	for (uint16_t idx = 0; idx < nLoop; ++idx) {	76	for (uint16_t idx = 0; idx < nLoop; ++idx) {
74	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));	77	auto sliceDst = dst(MakeCoord(MakeCoord(0, idx), MakeCoord(0, 0)));
75	- LoadCbufToCb3510::LoadData<trait>(sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	78	+ LoadCbufToCb3510::LoadData<trait>(
		79	+ sliceDst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
76	mStartPosition += M_STEP_MIN_VAL_B8;	80	mStartPosition += M_STEP_MIN_VAL_B8;
77	}	81	}
78	}	82	}
	@@ -87,28 +91,39 @@ private:
87	auto mStartPosition = Std::get<0>(coord) / FRACTAL_FIXED;	91	auto mStartPosition = Std::get<0>(coord) / FRACTAL_FIXED;
88	auto kStartPosition = Std::get<1>(coord) / C0_ELEMENT<DstType>;	92	auto kStartPosition = Std::get<1>(coord) / C0_ELEMENT<DstType>;
89	auto n1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	93	auto n1 = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
90	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) -	94	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) -
91	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *	95	GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(dstLayout) *
92	- GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout) - Std::get<1>(coord);	96	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(dstLayout) -
		97	+ Std::get<1>(coord);
93	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *	98	auto mStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(srcLayout) *
94	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) / FRACTAL_FIXED - mStartPosition;	99	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::ROW, 0>(srcLayout) /
		100	+ FRACTAL_FIXED -
		101	+ mStartPosition;
95	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *	102	auto kStep = GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 1>(srcLayout) *
96	- GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) / C0_ELEMENT<DstType> - kStartPosition;	103	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::SHAPE, AttrInfo::COLUMN, 0>(srcLayout) /
		104	+ C0_ELEMENT<DstType> -
		105	+ kStartPosition;
97	// Nz -> Zn	106	// Nz -> Zn
98	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	107	constexpr uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
99	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;	108	+ auto srcStride =
100	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;	109	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::COLUMN, 1>(srcLayout) / STRIDE_UNIT;
		110	+ auto dstStride =
		111	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;
101	if constexpr (is_b4_type<DstType>) {	112	if constexpr (is_b4_type<DstType>) {
102	if (n1 < FRACTAL_FIXED) {	113	if (n1 < FRACTAL_FIXED) {
103	- LoadCbufToCbS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	114	+ LoadCbufToCbS43510::LoadData<trait>(
		115	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
104	} else {	116	} else {
105	- LoadDataImplB4<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	117	+ LoadDataImplB4<trait, T, U>(
		118	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
106	}	119	}
107	} else {	120	} else {
108	if (n1 < FRACTAL_FIXED) {	121	if (n1 < FRACTAL_FIXED) {
109	- LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	122	+ LoadCbufToCb3510::LoadData<trait>(
		123	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
110	} else {	124	} else {
111	- LoadDataImplB8<trait, T, U>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	125	+ LoadDataImplB8<trait, T, U>(
		126	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
112	}	127	}
113	}	128	}
114	}	129	}

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn.h+20-17

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0BZN2ZN3510 {	29	class LoadDataL12L0BZN2ZN3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U>	31	template <const LoadDataTrait& trait, typename T, typename U>
33	- __aicore__ inline static void Run(const T& dst, const U& src) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src)
		33	+ {
34	LoadDataImpl<TransTrait<trait, false>, T, U>(dst, src);	34	LoadDataImpl<TransTrait<trait, false>, T, U>(dst, src);
35	}	35	}
36		36
	@@ -56,13 +56,16 @@ private:
56	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout);	56	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout);
57	// Zn -> Zn	57	// Zn -> Zn
58	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	58	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
59	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;	59	+ auto srcStride =
60	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;	60	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;
		61	+ auto dstStride =
		62	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;
61	if constexpr (is_b4_type<DstType>) {	63	if constexpr (is_b4_type<DstType>) {
62	- LoadCbufToCbS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	64	+ LoadCbufToCbS43510::LoadData<trait>(
63	- }	65	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
64	- else {	66	+ } else {
65	- LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	67	+ LoadCbufToCb3510::LoadData<trait>(
		68	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
66	}	69	}
67	}	70	}
68	};	71	};

Mimpl/experimental/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn_with_coord.h+20-17

	@@ -1,15 +1,15 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/cube_datamove/load_data/npu_arch_3510/load_data_l12l0b/zn2zn_with_coord.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -27,10 +27,10 @@
27	namespace AscendC {	27	namespace AscendC {
28	namespace Te {	28	namespace Te {
29	class LoadDataL12L0BZN2ZNWithCoord3510 {	29	class LoadDataL12L0BZN2ZNWithCoord3510 {
30	-
31	public:	30	public:
32	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>	31	template <const LoadDataTrait& trait, typename T, typename U, typename Coord>
33	- __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord) {	32	+ __aicore__ inline static void Run(const T& dst, const U& src, const Coord& coord)
		33	+ {
34	LoadDataImpl<TransTrait<trait, false>, T, U, Coord>(dst, src, coord);	34	LoadDataImpl<TransTrait<trait, false>, T, U, Coord>(dst, src, coord);
35	}	35	}
36		36
	@@ -56,13 +56,16 @@ private:
56	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout);	56	auto kStep = GetEleFromLayout<decltype(dstLayout), AttrInfo::SHAPE, AttrInfo::ROW, 1>(dstLayout);
57	// Zn -> Zn	57	// Zn -> Zn
58	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;	58	uint32_t STRIDE_UNIT = C0_ELEMENT<DstType> * FRACTAL_FIXED;
59	- auto srcStride = GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;	59	+ auto srcStride =
60	- auto dstStride = GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;	60	+ GetEleFromLayout<decltype(srcLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(srcLayout) / STRIDE_UNIT;
		61	+ auto dstStride =
		62	+ GetEleFromLayout<decltype(dstLayout), AttrInfo::STRIDE, AttrInfo::ROW, 1>(dstLayout) / STRIDE_UNIT;
61	if constexpr (is_b4_type<DstType>) {	63	if constexpr (is_b4_type<DstType>) {
62	- LoadCbufToCbS43510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	64	+ LoadCbufToCbS43510::LoadData<trait>(
63	- }	65	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
64	- else {	66	+ } else {
65	- LoadCbufToCb3510::LoadData<trait>(dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);	67	+ LoadCbufToCb3510::LoadData<trait>(
		68	+ dst, src, mStartPosition, kStartPosition, mStep, kStep, srcStride, dstStride);
66	}	69	}
67	}	70	}
68	};	71	};

Mimpl/experimental/tensor_api/arch/utils/arch_utils.h+24-21

Mimpl/experimental/tensor_api/arch/utils/check_data_type_3510.h+129-118

	@@ -9,7 +9,7 @@
9	*/	9	*/
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/utils/check_data_type_3510.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/utils/check_data_type_3510.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -38,16 +38,18 @@ public:
38	using filterDataType = typename S::elementType;	38	using filterDataType = typename S::elementType;
39		39
40	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	40	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
41	- static_assert(Std::is_one_of_v<Std::tuple<dstDataType, fmDataType, filterDataType>,	41	+ static_assert(
42	- Std::tuple<__cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,	42	+ Std::is_one_of_v<
43	- Std::tuple<__cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,	43	+ Std::tuple<dstDataType, fmDataType, filterDataType>,
44	- Std::tuple<__cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,	44	+ Std::tuple<__cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,
45	- Std::tuple<__cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,	45	+ Std::tuple<__cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,
46	- Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	46	+ Std::tuple<__cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,
47	- Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	47	+ Std::tuple<__cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,
48	- Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	48	+ Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
49	- Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,	49	+ Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
50	- "The data type is not supported for L0C position.");	50	+ Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
		51	+ Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,
		52	+ "The data type is not supported for L0C position.");
51	#endif	53	#endif
52	}	54	}
53		55
	@@ -62,27 +64,29 @@ public:
62	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	64	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
63	if constexpr (biasPos == Hardware::BIAS) {	65	if constexpr (biasPos == Hardware::BIAS) {
64	static_assert(	66	static_assert(
65	- Std::is_one_of_v<Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,	67	+ Std::is_one_of_v<
66	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,	68	+ Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,
67	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,	69	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,
68	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,	70	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,
69	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,	71	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,
70	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	72	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,
71	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	73	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
72	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	74	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
73	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,	75	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
		76	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,
74	"The data type is not supported for BIAS position.");	77	"The data type is not supported for BIAS position.");
75	} else if constexpr (biasPos == Hardware::L0C) {	78	} else if constexpr (biasPos == Hardware::L0C) {
76	static_assert(	79	static_assert(
77	- Std::is_one_of_v<Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,	80	+ Std::is_one_of_v<
78	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,	81	+ Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,
79	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,	82	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e2m1_t>,
80	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,	83	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e2m1_t, __cb__ fp4x2_e1m2_t>,
81	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,	84	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e2m1_t>,
82	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	85	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp4x2_e1m2_t, __cb__ fp4x2_e1m2_t>,
83	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	86	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
84	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	87	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
85	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,	88	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
		89	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>>,
86	"The data type is not supported for L0C position.");	90	"The data type is not supported for L0C position.");
87	}	91	}
88	#endif	92	#endif
	@@ -96,17 +100,19 @@ public:
96	using filterDataType = typename S::elementType;	100	using filterDataType = typename S::elementType;
97		101
98	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	102	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
99	- static_assert(Std::is_one_of_v<Std::tuple<dstDataType, fmDataType, filterDataType>,	103	+ static_assert(
100	- Std::tuple<__cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,	104	+ Std::is_one_of_v<
101	- Std::tuple<__cc__ float, __ca__ half, __cb__ half>,	105	+ Std::tuple<dstDataType, fmDataType, filterDataType>,
102	- Std::tuple<__cc__ float, __ca__ float, __cb__ float>,	106	+ Std::tuple<__cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,
103	- Std::tuple<__cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,	107	+ Std::tuple<__cc__ float, __ca__ half, __cb__ half>,
104	- Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	108	+ Std::tuple<__cc__ float, __ca__ float, __cb__ float>,
105	- Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	109	+ Std::tuple<__cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,
106	- Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	110	+ Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
107	- Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,	111	+ Std::tuple<__cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
108	- Std::tuple< __cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,	112	+ Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
109	- "The data type is not supported for L0C position.");	113	+ Std::tuple<__cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,
		114	+ Std::tuple<__cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,
		115	+ "The data type is not supported for L0C position.");
110	#endif	116	#endif
111	}	117	}
112		118
	@@ -122,29 +128,31 @@ public:
122	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	128	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
123	if constexpr (biasPos == Hardware::BIAS) {	129	if constexpr (biasPos == Hardware::BIAS) {
124	static_assert(	130	static_assert(
125	- Std::is_one_of_v<Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,	131	+ Std::is_one_of_v<
126	- Std::tuple<__biasbuf__ int32_t, __cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,	132	+ Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,
127	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ half, __cb__ half>,	133	+ Std::tuple<__biasbuf__ int32_t, __cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,
128	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ float, __cb__ float>,	134	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ half, __cb__ half>,
129	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,	135	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ float, __cb__ float>,
130	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	136	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,
131	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	137	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
132	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	138	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
133	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,	139	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
134	- Std::tuple<__biasbuf__ float, __cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,	140	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,
		141	+ Std::tuple<__biasbuf__ float, __cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,
135	"The data type is not supported for BIAS position.");	142	"The data type is not supported for BIAS position.");
136	} else if constexpr (biasPos == Hardware::L0C) {	143	} else if constexpr (biasPos == Hardware::L0C) {
137	static_assert(	144	static_assert(
138	- Std::is_one_of_v<Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,	145	+ Std::is_one_of_v<
139	- Std::tuple<__cc__ int32_t, __cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,	146	+ Std::tuple<biasDataType, dstDataType, fmDataType, filterDataType>,
140	- Std::tuple<__cc__ float, __cc__ float, __ca__ half, __cb__ half>,	147	+ Std::tuple<__cc__ int32_t, __cc__ int32_t, __ca__ int8_t, __cb__ int8_t>,
141	- Std::tuple<__cc__ float, __cc__ float, __ca__ float, __cb__ float>,	148	+ Std::tuple<__cc__ float, __cc__ float, __ca__ half, __cb__ half>,
142	- Std::tuple<__cc__ float, __cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,	149	+ Std::tuple<__cc__ float, __cc__ float, __ca__ float, __cb__ float>,
143	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,	150	+ Std::tuple<__cc__ float, __cc__ float, __ca__ bfloat16_t, __cb__ bfloat16_t>,
144	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,	151	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e4m3fn_t>,
145	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,	152	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e4m3fn_t, __cb__ fp8_e5m2_t>,
146	- Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,	153	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e4m3fn_t>,
147	- Std::tuple<__cc__ float, __cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,	154	+ Std::tuple<__cc__ float, __cc__ float, __ca__ fp8_e5m2_t, __cb__ fp8_e5m2_t>,
		155	+ Std::tuple<__cc__ float, __cc__ float, __ca__ hifloat8_t, __cb__ hifloat8_t>>,
148	"The data type is not supported for L0C position.");	156	"The data type is not supported for L0C position.");
149	}	157	}
150	#endif	158	#endif
	@@ -157,16 +165,17 @@ public:
157	using srcDataType = typename U::elementType;	165	using srcDataType = typename U::elementType;
158		166
159	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	167	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
160	- static_assert(Std::is_one_of_v<	168	+ static_assert(
161	- Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,	169	+ Std::is_one_of_v<
162	- Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,	170	+ Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,
163	- Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,	171	+ Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,
164	- Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,	172	+ Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,
165	- Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,	173	+ Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,
166	- Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,	174	+ Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,
167	- Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,	175	+ Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,
168	- Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,	176	+ Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,
169	- "The data type is not supported.");	177	+ Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,
		178	+ "The data type is not supported.");
170	#endif	179	#endif
171	}	180	}
172		181
	@@ -177,18 +186,19 @@ public:
177	using dstDataType = typename T::elementType;	186	using dstDataType = typename T::elementType;
178		187
179	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	188	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
180	- static_assert(Std::is_one_of_v<	189	+ static_assert(
181	- Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,	190	+ Std::is_one_of_v<
182	- Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,	191	+ Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,
183	- Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,	192	+ Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,
184	- Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,	193	+ Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,
185	- Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,	194	+ Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,
186	- Std::tuple<__cbuf__ fp4x2_e1m2_t, __gm__ fp4x2_e1m2_t>,	195	+ Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,
187	- Std::tuple<__cbuf__ fp4x2_e2m1_t, __gm__ fp4x2_e2m1_t>,	196	+ Std::tuple<__cbuf__ fp4x2_e1m2_t, __gm__ fp4x2_e1m2_t>,
188	- Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,	197	+ Std::tuple<__cbuf__ fp4x2_e2m1_t, __gm__ fp4x2_e2m1_t>,
189	- Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,	198	+ Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,
190	- Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,	199	+ Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,
191	- "The data type is not supported.");	200	+ Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,
		201	+ "The data type is not supported.");
192	#endif	202	#endif
193	}	203	}
194		204
	@@ -212,17 +222,18 @@ public:
212	using dstDataType = typename T::elementType;	222	using dstDataType = typename T::elementType;
213		223
214	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	224	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
215	- static_assert(Std::is_one_of_v<	225	+ static_assert(
216	- Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,	226	+ Std::is_one_of_v<
217	- Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,	227	+ Std::tuple<dstDataType, srcDataType>, Std::tuple<__cbuf__ half, __gm__ half>,
218	- Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,	228	+ Std::tuple<__cbuf__ bfloat16_t, __gm__ bfloat16_t>, Std::tuple<__cbuf__ float, __gm__ float>,
219	- Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,	229	+ Std::tuple<__cbuf__ int8_t, __gm__ int8_t>, Std::tuple<__cbuf__ uint8_t, __gm__ uint8_t>,
220	- Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,	230	+ Std::tuple<__cbuf__ int16_t, __gm__ int16_t>, Std::tuple<__cbuf__ uint16_t, __gm__ uint16_t>,
221	- Std::tuple<__cbuf__ int64_t, __gm__ int64_t>, Std::tuple<__cbuf__ uint64_t, __gm__ uint64_t>,	231	+ Std::tuple<__cbuf__ int32_t, __gm__ int32_t>, Std::tuple<__cbuf__ uint32_t, __gm__ uint32_t>,
222	- Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,	232	+ Std::tuple<__cbuf__ int64_t, __gm__ int64_t>, Std::tuple<__cbuf__ uint64_t, __gm__ uint64_t>,
223	- Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,	233	+ Std::tuple<__cbuf__ fp8_e5m2_t, __gm__ fp8_e5m2_t>,
224	- Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,	234	+ Std::tuple<__cbuf__ fp8_e4m3fn_t, __gm__ fp8_e4m3fn_t>,
225	- "The data type is not supported.");	235	+ Std::tuple<__cbuf__ hifloat8_t, __gm__ hifloat8_t>>,
		236	+ "The data type is not supported.");
226	#endif	237	#endif
227	}	238	}
228		239
	@@ -231,8 +242,9 @@ public:
231	{	242	{
232	using ShapeRow1 = typename GetFourDimType<U, AttrInfo::SHAPE, AttrInfo::ROW, 1>::type;	243	using ShapeRow1 = typename GetFourDimType<U, AttrInfo::SHAPE, AttrInfo::ROW, 1>::type;
233	using ShapeCol1 = typename GetFourDimType<U, AttrInfo::SHAPE, AttrInfo::COLUMN, 1>::type;	244	using ShapeCol1 = typename GetFourDimType<U, AttrInfo::SHAPE, AttrInfo::COLUMN, 1>::type;
234	- static_assert(Std::is_constant<1, ShapeRow1>::value \|\| Std::is_constant<1, ShapeCol1>::value,	245	+ static_assert(
235	- "The src only support 1D tensor");	246	+ Std::is_constant<1, ShapeRow1>::value \|\| Std::is_constant<1, ShapeCol1>::value,
		247	+ "The src only support 1D tensor");
236	}	248	}
237		249
238	template <typename T, typename U>	250	template <typename T, typename U>
	@@ -243,12 +255,11 @@ public:
243		255
244	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	256	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
245	static_assert(	257	static_assert(
246	- Std::is_one_of_v<Std::tuple<dstDataType, srcDataType>,	258	+ Std::is_one_of_v<
247	- Std::tuple<__biasbuf__ float, __cbuf__ bfloat16_t>,	259	+ Std::tuple<dstDataType, srcDataType>, Std::tuple<__biasbuf__ float, __cbuf__ bfloat16_t>,
248	- Std::tuple<__biasbuf__ float, __cbuf__ half>,	260	+ Std::tuple<__biasbuf__ float, __cbuf__ half>, Std::tuple<__biasbuf__ float, __cbuf__ float>,
249	- Std::tuple<__biasbuf__ float, __cbuf__ float>,	261	+ Std::tuple<__biasbuf__ int32_t, __cbuf__ int32_t>>,
250	- Std::tuple<__biasbuf__ int32_t, __cbuf__ int32_t>>,	262	+ "The data type is not supported.");
251	- "The data type is not supported.");
252	#endif	263	#endif
253	}	264	}
254		265
	@@ -260,9 +271,7 @@ public:
260		271
261	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	272	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
262	static_assert(	273	static_assert(
263	- Std::is_same_v<	274	+ Std::is_same_v<Std::tuple<dstDataType, srcDataType>, Std::tuple<__fbuf__ uint64_t, __cbuf__ uint64_t>>,
264	- Std::tuple<dstDataType, srcDataType>,
265	- Std::tuple<__fbuf__ uint64_t, __cbuf__ uint64_t>>,
266	"The data type is not supported.");	275	"The data type is not supported.");
267	#endif	276	#endif
268	}	277	}
	@@ -274,13 +283,14 @@ public:
274	using dstType = typename T::elementType;	283	using dstType = typename T::elementType;
275	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	284	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
276	static_assert(	285	static_assert(
277	- (quantPre == QuantMode_t::NoQuant	286	+ (quantPre == QuantMode_t::NoQuant &&
278	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__gm__ float, __cc__ float>,	287	+ Std::is_one_of_v<
279	- Std::tuple<__gm__ int32_t, __cc__ int32_t>>)	288	+ Std::tuple<dstType, srcType>, Std::tuple<__gm__ float, __cc__ float>,
280	- \|\| (quantPre == QuantMode_t::F322F16	289	+ Std::tuple<__gm__ int32_t, __cc__ int32_t>>) \|\|
281	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__gm__ half, __cc__ float>>)	290	+ (quantPre == QuantMode_t::F322F16 &&
282	- \|\| (quantPre == QuantMode_t::F322BF16	291	+ Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__gm__ half, __cc__ float>>) \|\|
283	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__gm__ bfloat16_t, __cc__ float>>),	292	+ (quantPre == QuantMode_t::F322BF16 &&
		293	+ Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__gm__ bfloat16_t, __cc__ float>>),
284	"The data type is not supported.");	294	"The data type is not supported.");
285	#endif	295	#endif
286	}	296	}
	@@ -292,13 +302,14 @@ public:
292	using dstType = typename T::elementType;	302	using dstType = typename T::elementType;
293	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510	303	#if defined(__NPU_ARCH__) && __NPU_ARCH__ == 3510
294	static_assert(	304	static_assert(
295	- (quantPre == QuantMode_t::NoQuant	305	+ (quantPre == QuantMode_t::NoQuant &&
296	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ float, __cc__ float>,	306	+ Std::is_one_of_v<
297	- Std::tuple<__ubuf__ int32_t, __cc__ int32_t>>)	307	+ Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ float, __cc__ float>,
298	- \|\| (quantPre == QuantMode_t::F322F16	308	+ Std::tuple<__ubuf__ int32_t, __cc__ int32_t>>) \|\|
299	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ half, __cc__ float>>)	309	+ (quantPre == QuantMode_t::F322F16 &&
300	- \|\| (quantPre == QuantMode_t::F322BF16	310	+ Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ half, __cc__ float>>) \|\|
301	- && Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ bfloat16_t, __cc__ float>>),	311	+ (quantPre == QuantMode_t::F322BF16 &&
		312	+ Std::is_one_of_v<Std::tuple<dstType, srcType>, Std::tuple<__ubuf__ bfloat16_t, __cc__ float>>),
302	"The data type is not supported.");	313	"The data type is not supported.");
303	#endif	314	#endif
304	}	315	}

Mimpl/experimental/tensor_api/arch/utils/check_format.h+42-39

	@@ -9,7 +9,7 @@
9	*/	9	*/
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/arch/utils/check_format.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/arch/utils/check_format.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
	@@ -38,16 +38,17 @@ public:
38	using dataType = typename T::elementType;	38	using dataType = typename T::elementType;
39	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;	39	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;
40	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;	40	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;
41	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<dataType>>>,	41	+ static_assert(
42	- "Layout->Shape->Column->ZeroDim must be 32/sizeof(dataType)!");	42	+ Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<dataType>>>,
43	- static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>,	43	+ "Layout->Shape->Column->ZeroDim must be 32/sizeof(dataType)!");
44	- "Layout->Shape->Row->ZeroDim must be 16!");	44	+ static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Row->ZeroDim must be 16!");
45		45
46	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;	46	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;
47	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;	47	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;
48	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column-ZeroDim must be 1!");	48	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column-ZeroDim must be 1!");
49	- static_assert(Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<dataType>>>,	49	+ static_assert(
50	- "Layout->Stride->Row->ZeroDim must be 32/sizeof(dataType)!");	50	+ Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<dataType>>>,
		51	+ "Layout->Stride->Row->ZeroDim must be 32/sizeof(dataType)!");
51	}	52	}
52		53
53	template <typename T>	54	template <typename T>
	@@ -56,17 +57,18 @@ public:
56	using dataType = typename T::elementType;	57	using dataType = typename T::elementType;
57	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;	58	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;
58	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;	59	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;
59	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<FRACTAL_FIXED>>,	60	+ static_assert(
60	- "Filter Layout->Shape->Column->ZeroDim must be 16!");	61	+ Std::is_same_v<ShapeColumn0, Std::Int<FRACTAL_FIXED>>, "Filter Layout->Shape->Column->ZeroDim must be 16!");
61	- static_assert(Std::is_same_v<ShapeRow0, Std::Int<C0_ELEMENT<dataType>>>,	62	+ static_assert(
62	- "Filter Layout->Shape->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	63	+ Std::is_same_v<ShapeRow0, Std::Int<C0_ELEMENT<dataType>>>,
		64	+ "Filter Layout->Shape->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
63		65
64	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;	66	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;
65	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;	67	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;
66	- static_assert(Std::is_same_v<StrideColumn0, Std::Int<C0_ELEMENT<dataType>>>,	68	+ static_assert(
67	- "Filter Layout->Stride->Column-ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	69	+ Std::is_same_v<StrideColumn0, Std::Int<C0_ELEMENT<dataType>>>,
68	- static_assert(Std::is_same_v<StrideRow0, Std::Int<1>>,	70	+ "Filter Layout->Stride->Column-ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
69	- "Filter Layout->Stride->Row->ZeroDim must be 1!");	71	+ static_assert(Std::is_same_v<StrideRow0, Std::Int<1>>, "Filter Layout->Stride->Row->ZeroDim must be 1!");
70	}	72	}
71		73
72	template <typename T>	74	template <typename T>
	@@ -75,18 +77,18 @@ public:
75	using type = typename T::elementType;	77	using type = typename T::elementType;
76	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;	78	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;
77	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;	79	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;
78	- static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>,	80	+ static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Row->ZeroDim must be 16!");
79	- "Layout->Shape->Row->ZeroDim must be 16!");	81	+ static_assert(
80	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<type>>>,	82	+ Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<type>>>,
81	- "Layout->Shape->Column->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	83	+ "Layout->Shape->Column->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
82		84
83	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;	85	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;
84	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;	86	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;
85	using StrideRow1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 1>::type;	87	using StrideRow1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 1>::type;
86	- static_assert(Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<type>>>,	88	+ static_assert(
87	- "Layout->Stride->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	89	+ Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<type>>>,
88	- static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>,	90	+ "Layout->Stride->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
89	- "Layout->Stride->Column->ZeroDim must be 1!");	91	+ static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column->ZeroDim must be 1!");
90	static_assert(	92	static_assert(
91	Std::is_same_v<StrideRow1, Std::Int<C0_ELEMENT<type> * FRACTAL_FIXED>>,	93	Std::is_same_v<StrideRow1, Std::Int<C0_ELEMENT<type> * FRACTAL_FIXED>>,
92	"Layout->Stride->Column->ZeroDimmust be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType) * 16)!");	94	"Layout->Stride->Column->ZeroDimmust be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType) * 16)!");
	@@ -111,11 +113,13 @@ public:
111	using StrideRow1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 1>::type;	113	using StrideRow1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 1>::type;
112		114
113	static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Row->ZeroDim must be 16!");	115	static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Row->ZeroDim must be 16!");
114	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<type>>>,	116	+ static_assert(
115	- "Layout->Shape->Column->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	117	+ Std::is_same_v<ShapeColumn0, Std::Int<C0_ELEMENT<type>>>,
		118	+ "Layout->Shape->Column->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
116		119
117	- static_assert(Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<type>>>,	120	+ static_assert(
118	- "Layout->Stride->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");	121	+ Std::is_same_v<StrideRow0, Std::Int<C0_ELEMENT<type>>>,
		122	+ "Layout->Stride->Row->ZeroDim must be (is_b4_type<dataType> ? 64 : 32 / sizeof(dataType))!");
119	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column->ZeroDim must be 1!");	123	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column->ZeroDim must be 1!");
120	static_assert(	124	static_assert(
121	Std::is_same_v<StrideRow1, Std::Int<C0_ELEMENT<type> * FRACTAL_FIXED>>,	125	Std::is_same_v<StrideRow1, Std::Int<C0_ELEMENT<type> * FRACTAL_FIXED>>,
	@@ -144,17 +148,18 @@ public:
144	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;	148	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;
145	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;	149	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;
146	static_assert(Std::is_same_v<ShapeRow0, Std::Int<2>>, "CopyCbufToFB Layout->Shape->Row->ZeroDim must be 2!");	150	static_assert(Std::is_same_v<ShapeRow0, Std::Int<2>>, "CopyCbufToFB Layout->Shape->Row->ZeroDim must be 2!");
147	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<1>>,	151	+ static_assert(
148	- "CopyCbufToFB Layout->Shape->Column->ZeroDim must be 1!");	152	+ Std::is_same_v<ShapeColumn0, Std::Int<1>>, "CopyCbufToFB Layout->Shape->Column->ZeroDim must be 1!");
149		153
150	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;	154	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;
151	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;	155	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;
152	using StrideColumn1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 1>::type;	156	using StrideColumn1 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 1>::type;
153	static_assert(Std::is_same_v<StrideRow0, Std::Int<1>>, "CopyCbufToFB Layout->Stride->Row->ZeroDim must be 1!");	157	static_assert(Std::is_same_v<StrideRow0, Std::Int<1>>, "CopyCbufToFB Layout->Stride->Row->ZeroDim must be 1!");
154	- static_assert(Std::is_same_v<StrideColumn0, Std::Int<0>>,	158	+ static_assert(
155	- "CopyCbufToFB Layout->Stride->Column->ZeroDim must be 0!");	159	+ Std::is_same_v<StrideColumn0, Std::Int<0>>, "CopyCbufToFB Layout->Stride->Column->ZeroDim must be 0!");
156	- static_assert(Std::is_same_v<StrideColumn1, Std::Int<MX_SCALE_K0>>,	160	+ static_assert(
157	- "CopyCbufToFB Layout->Stride->Column->OneDim must be 2!");	161	+ Std::is_same_v<StrideColumn1, Std::Int<MX_SCALE_K0>>,
		162	+ "CopyCbufToFB Layout->Stride->Column->OneDim must be 2!");
158	}	163	}
159		164
160	template <typename T>	165	template <typename T>
	@@ -162,15 +167,13 @@ public:
162	{	167	{
163	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;	168	using ShapeRow0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::ROW, 0>::type;
164	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;	169	using ShapeColumn0 = typename GetFourDimType<T, AttrInfo::SHAPE, AttrInfo::COLUMN, 0>::type;
165	- static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>,	170	+ static_assert(Std::is_same_v<ShapeRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Row->ZeroDim must be 16!");
166	- "Layout->Shape->Row->ZeroDim must be 16!");	171	+ static_assert(
167	- static_assert(Std::is_same_v<ShapeColumn0, Std::Int<FRACTAL_FIXED>>,	172	+ Std::is_same_v<ShapeColumn0, Std::Int<FRACTAL_FIXED>>, "Layout->Shape->Column->ZeroDim must be 16!");
168	- "Layout->Shape->Column->ZeroDim must be 16!");
169		173
170	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;	174	using StrideRow0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::ROW, 0>::type;
171	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;	175	using StrideColumn0 = typename GetFourDimType<T, AttrInfo::STRIDE, AttrInfo::COLUMN, 0>::type;
172	- static_assert(Std::is_same_v<StrideRow0, Std::Int<FRACTAL_FIXED>>,	176	+ static_assert(Std::is_same_v<StrideRow0, Std::Int<FRACTAL_FIXED>>, "Layout->Stride->Row->ZeroDim must be 16!");
173	- "Layout->Stride->Row->ZeroDim must be 16!");
174	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column->ZeroDim must be 1!");	177	static_assert(Std::is_same_v<StrideColumn0, Std::Int<1>>, "Layout->Stride->Column->ZeroDim must be 1!");
175	}	178	}
176		179

Mimpl/experimental/tensor_api/arch/utils/is_format.h+101-85

Mimpl/experimental/tensor_api/atom/copy_atom_impl.h+19-18

Mimpl/experimental/tensor_api/atom/copy_traits_impl.h+20-21

Mimpl/experimental/tensor_api/atom/cube_compute/cube_compute_impl.h+12-12

Mimpl/experimental/tensor_api/atom/cube_compute/mmad.h+17-16

Mimpl/experimental/tensor_api/atom/cube_datamove/copy_gm2l1.h+15-15

Mimpl/experimental/tensor_api/atom/cube_datamove/copy_l0c2out.h+35-37

Mimpl/experimental/tensor_api/atom/cube_datamove/copy_l12l0.h+15-15

Mimpl/experimental/tensor_api/atom/cube_datamove/cube_datamove_impl.h+12-12

Mimpl/experimental/tensor_api/atom/mad_atom_impl.h+19-18

Mimpl/experimental/tensor_api/atom/mad_traits_impl.h+19-20

Mimpl/experimental/tensor_api/tensor/coord_index.h+18-19

Mimpl/experimental/tensor_api/tensor/engine_impl.h+25-31

Mimpl/experimental/tensor_api/tensor/layout_definition.h+16-26

Mimpl/experimental/tensor_api/tensor/layout_dispatch.h+178-121

Mimpl/experimental/tensor_api/tensor/layout_fractal.h+60-48

Mimpl/experimental/tensor_api/tensor/layout_impl.h+22-19

Mimpl/experimental/tensor_api/tensor/layout_method.h+39-30

Mimpl/experimental/tensor_api/tensor/layout_size.h+27-22

Mimpl/experimental/tensor_api/tensor/layout_static_fractal.h+60-50

Mimpl/experimental/tensor_api/tensor/layout_struct.h+45-34

Mimpl/experimental/tensor_api/tensor/local_tensor_impl.h+75-81

Mimpl/experimental/tensor_api/tensor/pointer_adaptor_impl.h+41-36

Mimpl/experimental/tensor_api/tensor/pointer_impl.h+42-27

Mimpl/experimental/tensor_api/tensor/pointer_mem_impl.h+15-14

Mimpl/experimental/tensor_api/tensor/tuple_fractal.h+144-111

Mimpl/experimental/tensor_api/tensor/tuple_impl.h+25-24

Mimpl/experimental/tensor_api/utils/constant_impl.h+29-29

Mimpl/experimental/tensor_api/utils/extra_impl.h+15-15

Mimpl/experimental/tensor_api/utils/macro_impl.h+41-41

Mimpl/experimental/tensor_api/utils/utils_impl.h+12-12

Minclude/c_api/vector_compute/vector_compute.h+0-13

Ascripts/oat_check.sh+540-0

Mtests/api/c_api/npu_arch_2201/vector_compute/test_transto5hd.cpp+82-13

Mtests/api/c_api/npu_arch_3510/vector_compute/test_transto5hd.cpp+7-18

	@@ -1,24 +1,24 @@
1	/**	1	/**
2	-* Copyright (c) 2026 Huawei Technologies Co., Ltd.	2	+ * Copyright (c) 2026 Huawei Technologies Co., Ltd.
3	-* This program is free software, you can redistribute it and/or modify it under the terms and conditions of	3	+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
4	-* CANN Open Software License Agreement Version 2.0 (the "License").	4	+ * CANN Open Software License Agreement Version 2.0 (the "License").
5	-* Please refer to the License for details. You may not use this file except in compliance with the License.	5	+ * Please refer to the License for details. You may not use this file except in compliance with the License.
6	-* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,	6	+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
7	-* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.	7	+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
8	-* See LICENSE in the root of the software repository for the full text of the License.	8	+ * See LICENSE in the root of the software repository for the full text of the License.
9	-*/	9	+ */
10		10
11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)	11	#if !defined(ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12	-#warning \	12	+#warning \
13	"impl/tensor_api/algorithm/copy_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."	13	"impl/tensor_api/algorithm/copy_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file maybe removed in the future. Please use "#include "tensor_api/tensor.h"" and use public functions or variables defined in interface headers files."
14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS	14	#define ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC	15	#define UNDEF_ASCENDC_TENSOR_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
16	#endif	16	#endif
17		17
18	/*!	18	/*!
19	-* \file copy_impl.h	19	+ * \file copy_impl.h
20	-* \brief	20	+ * \brief
21	-*/	21	+ */
22	#ifndef IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H	22	#ifndef IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H
23	#define IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H	23	#define IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H
24		24
	@@ -28,24 +28,25 @@ namespace AscendC {
28	namespace Te {	28	namespace Te {
29		29
30	template <typename Tp, const Tp& traits, typename T, typename... Params>	30	template <typename Tp, const Tp& traits, typename T, typename... Params>
31	-__aicore__ inline void Copy(const CopyAtom<T>& atomCopy, const Params& ...params)	31	+__aicore__ inline void Copy(const CopyAtom<T>& atomCopy, const Params&... params)
32	{	32	{
33	atomCopy.template Call<traits>(params...);	33	atomCopy.template Call<traits>(params...);
34	}	34	}
35		35
36	template <typename T, typename... Params>	36	template <typename T, typename... Params>
37	-__aicore__ inline void Copy(const CopyAtom<T>& atomCopy, const Params& ...params)	37	+__aicore__ inline void Copy(const CopyAtom<T>& atomCopy, const Params&... params)
38	{	38	{
39	atomCopy.Call(params...);	39	atomCopy.Call(params...);
40	}	40	}
41		41
42	template <typename... Args>	42	template <typename... Args>
43	-__aicore__ inline auto MakeCopy(const Args& ...traits) {	43	+__aicore__ inline auto MakeCopy(const Args&... traits)
		44	+{
44	return CopyAtom<CopyTraits<Args...>>{};	45	return CopyAtom<CopyTraits<Args...>>{};
45	}	46	}
46		47
47	-}	48	+} // namespace Te
48	-}	49	+} // namespace AscendC
49		50
50	#endif // IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H	51	#endif // IMPL_TENSOR_API_ALGORITHM_COPY_IMPL_H
51		52