Star732
1029
代码
代码
Issues57
Pull Requests177
Actions
项目讨论
Wiki
分析
Star732
1029
已合并
refactor: unify transto5hd implementations #5146
leozjr创建于 21 天前
refactor: unify transto5hd implementations #5146
已合并
leozjr创建于 21 天前
从已删除 :fix/transto5hd-inline合入到cann/asc-devkitmaster
共 14 个文件变更+590-179
Mdocs/zh/api/README.md+1-0
@@ -1373,6 +1373,7 @@
1373 - [asc_mmad（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_deprecated.md)1373 - [asc_mmad（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_deprecated.md)
1374 - [asc_mmad_mx（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_mx_deprecated.md)1374 - [asc_mmad_mx（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_mx_deprecated.md)
1375 - [asc_mmad_sparse（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_sparse_deprecated.md)1375 - [asc_mmad_sparse（废弃）](SIMD-API/c_api/deprecated_interface/asc_mmad_sparse_deprecated.md)
1376+ - [asc_transto5hd_sync（废弃）](SIMD-API/c_api/deprecated_interface/asc_transto5hd_sync_deprecated.md)
1376 1377 
1377 - [Tensor API](SIMD-API/tensor_api/tensor_api.md)1378 - [Tensor API](SIMD-API/tensor_api/tensor_api.md)
1378 - [Tensor API列表](SIMD-API/tensor_api/tensor_api_list.md)1379 - [Tensor API列表](SIMD-API/tensor_api/tensor_api_list.md)
Mdocs/zh/api/SIMD-API/c_api/c_api.md+1-0
@@ -788,3 +788,4 @@ C API文档按最细粒度公共头文件归类。除试验接口外，整体使
788|[asc_mmad(废弃)](deprecated_interface/asc_mmad_deprecated.md)|矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|788|[asc_mmad(废弃)](deprecated_interface/asc_mmad_deprecated.md)|矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|
789|[asc_mmad_mx(废弃)](deprecated_interface/asc_mmad_mx_deprecated.md)|带有量化系数的矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|789|[asc_mmad_mx(废弃)](deprecated_interface/asc_mmad_mx_deprecated.md)|带有量化系数的矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|
790|[asc_mmad_sparse(废弃)](deprecated_interface/asc_mmad_sparse_deprecated.md)|稀疏矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|790|[asc_mmad_sparse(废弃)](deprecated_interface/asc_mmad_sparse_deprecated.md)|稀疏矩阵乘加计算接口。|`c_api/cube_compute/cube_compute.h`|
791+| [asc_transto5hd_sync（废弃）](deprecated_interface/asc_transto5hd_sync_deprecated.md) | 数据格式转换接口。 |
Adocs/zh/api/SIMD-API/c_api/deprecated_interface/asc_transto5hd_sync_deprecated.md+175-0
@@ -0,0 +1,175 @@
1+# asc_transto5hd_sync（废弃）
2+ 
3+## 产品支持情况
4+ 
5+<!-- npu="950" id1 -->
6+- Ascend 950PR/Ascend 950DT：支持
7+<!-- end id1 -->
8+<!-- npu="A3" id2 -->
9+- Atlas A3 训练系列产品/Atlas A3 推理系列产品：支持
10+<!-- end id2 -->
11+<!-- npu="910b" id3 -->
12+- Atlas A2 训练系列产品/Atlas A2 推理系列产品：支持
13+<!-- end id3 -->
14+<!-- npu="310b" id4 -->
15+- Atlas 200I/500 A2 推理产品：不支持
16+<!-- end id4 -->
17+<!-- npu="310p" id5 -->
18+- Atlas 推理系列产品AI Core：不支持
19+<!-- end id5 -->
20+<!-- npu="310p" id6 -->
21+- Atlas 推理系列产品Vector Core：不支持
22+<!-- end id6 -->
23+<!-- npu="910" id7 -->
24+- Atlas 训练系列产品：不支持
25+<!-- end id7 -->
26+ 
27+## 功能说明
28+ 
29+**`asc_transto5hd_b8_sync`、`asc_transto5hd_b16_sync`和`asc_transto5hd_b32_sync`接口已废弃。请使用[asc_transto5hd](../vector_compute/vector_permute_sel/asc_transto5hd.md)接口和[asc_sync](../sync/asc_sync.md)接口替代。**
30+ 
31+数据格式转换，一般用于将NCHW格式转换成NC1HWC0格式。特别的，也可以用于二维矩阵数据块的转置。本接口单次Repeat内可处理512Byte的数据（16个DataBlock），根据数据类型不同，支持不同shape的矩阵转置（比如数据类型为half时，单次Repeat可完成16*16大小的矩阵转置），同时还可以支持多次Repeat操作。
32+ 
33+单次repeat内转换规则如下：
34+ 
35+- 当输入数据类型位宽为16位时，每个DataBlock中包含16个数，指令内部会循环16次，每次循环都会分别从指定的16个DataBlock中的对应位置取值，组成一个新的DataBlock单元放入目的地址中。如下图所示，图中的srcList[0]-srcList[15]代表源操作数的16个DataBlock。![](../figures/asc_transto5hd_b16.png "图1输入数据类型位宽为16位时的转换规则")
36+ 
37+- 当数据类型位宽为32位时，每个DataBlock包含8个数，指令内部会循环8次，每次循环都会分别从指定的16个DataBlock中的对应位置取值，组成2个新的DataBlock放入目的地址中。如下图所示：![](../figures/asc_transto5hd_b32.png "图2 输入数据类型位宽为32位时的转换规则")
38+ 
39+- 当数据类型位宽为8位时，每个DataBlock包含32个数，指令内部会循环16次，每次循环都会分别从指定的16个DataBlock中的对应位置取值，组成半个DataBlock放入目的地址中，读取和存放是在DataBlock的高半部还是低半部由参数srcHighHalf和dstHighHalf决定。如下图所示：![](../figures/asc_transto5hd_b8.png "图3 输入数据类型位宽为8位时的转换规则")
40+ 
41+基于以上的转换规则，使用该接口进行NC1HWC0格式转换或者矩阵转置。NC1HWC0格式转换相对复杂，这里给出其具体的转换方法：
42+ 
43+NCHW格式转换成NC1HWC0格式时，如果是数据类型的位宽为32位或者16位，则C0=16；如果数据类型的位宽为8位，则C0=32。下图以C0=16为例进行介绍：![](../figures/asc_transto5hd_NCHW2NC1HWC0.png)
44+ 
45+## 函数原型
46+ 
47+- 同步计算
48+```cpp
49+__aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
50+ uint16_t dst_stride, uint16_t src_stride)
51+__aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
52+ uint16_t dst_stride, uint16_t src_stride)
53+__aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
54+ uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
55+```
56+ 
57+## 参数说明
58+ 
59+**表1** 参数说明
60+ 
61+|参数名|输入/输出|描述|
62+| ------------ | ------------ | ------------ |
63+|dst|输出|目的地址寄存器。|
64+|src|输入|源地址寄存器。|
65+|repeat|输入|重复迭代次数， repeat∈[0,255]。<br>注意事项：<br>&bull;当repeat为1时，目的操作数/源操作数的有效起始位置为dst/src序列输入的起始位置加上dst_stride/src_stride。如果要让目的操作数/源操作数的有效起始位置为dst/src序列输入的起始位置，需要将dst_stride/src_stride置为0。<br>&bull;当repeat大于1时，第一次repeat中目的操作数/源操作数的有效起始位置为dst/src序列输入的起始位置，第二次会加上dst_stride/src_stride。以此类推。|
66+|dst_stride|输入|相邻迭代间，目的操作数相同DataBlock地址Stride，单位：DataBlock。<br>相邻迭代间相同DataBlock的地址步长参数的详细说明请参考[repeatStride](../general_description_and_constraints.md#repeatStride)。|
67+|src_stride|输入|相邻迭代间，源操作数相同DataBlock地址Stride，单位：DataBlock。<br>相邻迭代间相同DataBlock的地址步长参数的详细说明请参考[repeatStride](../general_description_and_constraints.md#repeatStride)。|
68+|dst_high_half|输入|指定每个dst地址中的数据存储到DataBlock的高半部还是低半部，该配置只配置int8_t/uint8_t的数据类型。<br>支持的数据类型为bool，有以下两种取值：<br>&bull; true:表示存储于DataBlock的高半部。<br>&bull; false:表示存储于DataBlock的低半部。|
69+|src_high_half|输入|指定每个src地址中的数据存储到DataBlock的高半部还是低半部，该配置只配置int8_t/uint8_t的数据类型。<br>支持的数据类型为bool，有以下两种取值：<br>&bull; true:表示存储于DataBlock的高半部。<br>&bull;false:表示存储于DataBlock的低半部。|
70+ 
71+## 返回值说明
72+ 
73+无
74+ 
75+## 流水类型
76+ 
77+PIPE_V
78+ 
79+## 约束说明
80+ 
81+- 操作数地址对齐约束请参考[存储单元说明](../general_description_and_constraints.md#存储单元说明)。
82+- 操作数地址重叠约束请参考[通用地址重叠约束](../../general_description_and_constraints.md#通用地址重叠约束)。
83+- 进行NCHW格式到NC1HWC0格式的转换时，一般用法是将src/dst中的每个元素配置为每个HW平面的起点。
84+- 为了性能更优，数据类型位宽为8位时建议先固定dst_high_half/src_high_half，再HW方向Repeat后，再改变dst_high_half/src_high_half。
85+ 
86+## 调用示例
87+ 
88+```cpp
89+// 示例一：
90+constexpr uint64_t total_length = 256; // total_length指参与计算的数据长度
91+__ubuf__ half src[total_length];
92+__ubuf__ half dst[total_length];
93+const uint32_t STEP = 16;
94+// src_list和dst_list是16个DataBlock地址的数组
95+__ubuf__ half* src_list[16] = {
96+ (__ubuf__ half*)(src + 0 * STEP),
97+ (__ubuf__ half*)(src + 1 * STEP),
98+ (__ubuf__ half*)(src + 2 * STEP),
99+ (__ubuf__ half*)(src + 3 * STEP),
100+ (__ubuf__ half*)(src + 4 * STEP),
101+ (__ubuf__ half*)(src + 5 * STEP),
102+ (__ubuf__ half*)(src + 6 * STEP),
103+ (__ubuf__ half*)(src + 7 * STEP),
104+ (__ubuf__ half*)(src + 8 * STEP),
105+ (__ubuf__ half*)(src + 9 * STEP),
106+ (__ubuf__ half*)(src + 10 * STEP),
107+ (__ubuf__ half*)(src + 11 * STEP),
108+ (__ubuf__ half*)(src + 12 * STEP),
109+ (__ubuf__ half*)(src + 13 * STEP),
110+ (__ubuf__ half*)(src + 14 * STEP),
111+ (__ubuf__ half*)(src + 15 * STEP)
112+};
113+__ubuf__ half* dst_list[16] = {
114+ (__ubuf__ half*)(dst + 0 * STEP),
115+ (__ubuf__ half*)(dst + 1 * STEP),
116+ (__ubuf__ half*)(dst + 2 * STEP),
117+ (__ubuf__ half*)(dst + 3 * STEP),
118+ (__ubuf__ half*)(dst + 4 * STEP),
119+ (__ubuf__ half*)(dst + 5 * STEP),
120+ (__ubuf__ half*)(dst + 6 * STEP),
121+ (__ubuf__ half*)(dst + 7 * STEP),
122+ (__ubuf__ half*)(dst + 8 * STEP),
123+ (__ubuf__ half*)(dst + 9 * STEP),
124+ (__ubuf__ half*)(dst + 10 * STEP),
125+ (__ubuf__ half*)(dst + 11 * STEP),
126+ (__ubuf__ half*)(dst + 12 * STEP),
127+ (__ubuf__ half*)(dst + 13 * STEP),
128+ (__ubuf__ half*)(dst + 14 * STEP),
129+ (__ubuf__ half*)(dst + 15 * STEP)
130+};
131+const int32_t VA_REG_ARRAY_LEN = 8;
132+uint8_t repeat = 1;
133+uint16_t dst_stride = 0;
134+uint16_t src_stride = 0;
135+ 
136+// asc_set_va_reg接口要求前8个和后8个地址序列与地址寄存器分别关联
137+asc_set_va_reg(VA0, dst_list);
138+asc_set_va_reg(VA1, dst_list + VA_REG_ARRAY_LEN);
139+asc_set_va_reg(VA2, src_list);
140+asc_set_va_reg(VA3, src_list + VA_REG_ARRAY_LEN);
141+ 
142+asc_transto5hd_b16(VA0, VA2, repeat, dst_stride, src_stride);
143+ 
144+// 示例二：
145+constexpr uint64_t total_length = 2 * 32 * 16 * 16; // total_length指参与计算的数据长度
146+__ubuf__ half src[total_length];
147+__ubuf__ half dst[total_length];
148+const uint32_t DST_STEP = 16;
149+const uint32_t SRC_STEP = 16 * 16;
150+const uint32_t OFFSET = 16 * 16 * 16;
151+const int32_t VA_REG_ARRAY_LEN = 8;
152+uint8_t repeat = 16;
153+uint16_t dst_stride = 16;
154+uint16_t src_stride = 1;
155+ 
156+// src_list和dst_list定义在循环外部
157+__ubuf__ half* src_list[16];
158+__ubuf__ half* dst_list[16];
159+ 
160+for (int j = 0; j < 4; ++j) {
161+ // 在循环内更新地址列表
162+ for (int i = 0; i < 16; ++i) {
163+ src_list[i] = (__ubuf__ half*)(src + OFFSET * j + i * SRC_STEP);
164+ dst_list[i] = (__ubuf__ half*)(dst + OFFSET * j + i * DST_STEP);
165+ }
166+ 
167+ // 每次循环都要设置寄存器并执行转换
168+ asc_set_va_reg(VA0, dst_list);
169+ asc_set_va_reg(VA1, dst_list + VA_REG_ARRAY_LEN);
170+ asc_set_va_reg(VA2, src_list);
171+ asc_set_va_reg(VA3, src_list + VA_REG_ARRAY_LEN);
172+ 
173+ asc_transto5hd_b16(VA0, VA2, repeat, dst_stride, src_stride);
174+}
175+```
Mdocs/zh/api/SIMD-API/c_api/vector_compute/vector_permute_sel/asc_transto5hd.md+2-10
@@ -52,16 +52,6 @@ __aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t
52 uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)52 uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
53```53```
54 54 
55-- 同步计算
56-```cpp
57-__aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
58- uint16_t dst_stride, uint16_t src_stride)
59-__aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
60- uint16_t dst_stride, uint16_t src_stride)
61-__aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat,
62- uint16_t dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)
63-```
64- 
65## 参数说明55## 参数说明
66 56 
67**表1** 参数说明57**表1** 参数说明
@@ -88,6 +78,8 @@ PIPE_V
88 78 
89- 操作数地址对齐约束请参考[存储单元说明](../../general_description_and_constraints.md#存储单元说明)。79- 操作数地址对齐约束请参考[存储单元说明](../../general_description_and_constraints.md#存储单元说明)。
90- 操作数地址重叠约束请参考[通用地址重叠约束](../../general_description_and_constraints.md#通用地址重叠约束)。80- 操作数地址重叠约束请参考[通用地址重叠约束](../../general_description_and_constraints.md#通用地址重叠约束)。
81+- `dst`和`src`不能传入相同的地址寄存器起始位置，即目的地址寄存器对和源地址寄存器对不能重叠。
82+- 为获得更优性能，建议`dst`、`src`、`dst_high_half`和`src_high_half`传入编译期常量，例如直接传入VA0、VA2、true或false，或传入由这些常量初始化的局部变量。`dst`和`src`仅支持VA0、VA2、VA4、VA6，且不能相同，配置方式请参考[asc_set_va_reg](./asc_set_va_reg.md)；每个地址寄存器存储8个地址，接口会将传入的起始寄存器与相邻寄存器组成一组，处理16个地址。传入运行时变量时，编译器会保留地址寄存器或布尔参数的分支判断，影响性能。
91- 进行NCHW格式到NC1HWC0格式的转换时，一般用法是将src/dst中的每个元素配置为每个HW平面的起点。83- 进行NCHW格式到NC1HWC0格式的转换时，一般用法是将src/dst中的每个元素配置为每个HW平面的起点。
92- 为了性能更优，数据类型位宽为8位时建议先固定dst_high_half/src_high_half，再HW方向Repeat后，再改变dst_high_half/src_high_half。84- 为了性能更优，数据类型位宽为8位时建议先固定dst_high_half/src_high_half，再HW方向Repeat后，再改变dst_high_half/src_high_half。
93 85 
Mdocs/zh/api/appendix/deprecated_interface.md+3-0
@@ -480,3 +480,6 @@
480 480 
481 使用`uint8_t unit_flag`入参的`asc_mmad_sparse()`接口已废弃，请使用[asc_mmad_sparse](../SIMD-API/c_api/cube_compute/asc_mmad_sparse.md)接口。`asc_mmad_sparse_sync()`接口已废弃，请使用[asc_mmad_sparse](../SIMD-API/c_api/cube_compute/asc_mmad_sparse.md)接口和[asc_sync](../SIMD-API/c_api/sync/asc_sync.md)接口。481 使用`uint8_t unit_flag`入参的`asc_mmad_sparse()`接口已废弃，请使用[asc_mmad_sparse](../SIMD-API/c_api/cube_compute/asc_mmad_sparse.md)接口。`asc_mmad_sparse_sync()`接口已废弃，请使用[asc_mmad_sparse](../SIMD-API/c_api/cube_compute/asc_mmad_sparse.md)接口和[asc_sync](../SIMD-API/c_api/sync/asc_sync.md)接口。
482<!-- end id10 -->482<!-- end id10 -->
483+- [asc_transto5hd_sync（废弃）](../SIMD-API/c_api/deprecated_interface/asc_transto5hd_sync_deprecated.md)接口
484+ 
485+ 此接口已废弃。请使用[asc_transto5hd](../SIMD-API/c_api/vector_compute/vector_permute_sel/asc_transto5hd.md)接口和[asc_sync](../SIMD-API/c_api/sync/asc_sync.md)接口。
Mimpl/c_api/instr_impl/npu_arch_2201/vector_compute_impl.h+39-24
@@ -5308,35 +5308,50 @@ __aicore__ inline void asc_transpose_sync(__ubuf__ uint16_t* dst, __ubuf__ uint1
5308 asc_transpose_sync_impl(dst, src);5308 asc_transpose_sync_impl(dst, src);
5309}5309}
5310 5310 
5311-// __aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5311+// ==========asc_transto5hd(b8/b16/b32)==========
5312-// uint16_t src_stride, bool dst_high_half, bool src_high_half)5312+__aicore__ inline void asc_transto5hd_b8(
5313-#define asc_transto5hd_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \5313+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
5314- asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))5314+ bool src_high_half)
5315+{
5316+ asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
5317+}
5315 5318 
5316-// __aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5319+[[deprecated("NOTICE: asc_transto5hd_b8_sync is deprecated. "
5317-// uint16_t src_stride, bool dst_high_half, bool src_high_half)5320+ "Please use asc_transto5hd_b8 and asc_sync instead.")]]
5318-#define asc_transto5hd_b8_sync(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \5321+__aicore__ inline void asc_transto5hd_b8_sync(
5319- asc_transto5hd_b8_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))5322+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
5323+ bool src_high_half)
5324+{
5325+ asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
5326+}
5320 5327 
5321-// __aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5328+__aicore__ inline void asc_transto5hd_b16(
5322-// uint16_t src_stride)5329+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5323-#define asc_transto5hd_b16(dst, src, repeat, dst_stride, src_stride) \5330+{
5324- asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride))5331+ asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);
5332+}
5325 5333 
5326-// __aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5334+[[deprecated("NOTICE: asc_transto5hd_b16_sync is deprecated. "
5327-// uint16_t src_stride)5335+ "Please use asc_transto5hd_b16 and asc_sync instead.")]]
5328-#define asc_transto5hd_b16_sync(dst, src, repeat, dst_stride, src_stride) \5336+__aicore__ inline void asc_transto5hd_b16_sync(
5329- asc_transto5hd_b16_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))5337+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5338+{
5339+ asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride);
5340+}
5330 5341 
5331-// __aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5342+__aicore__ inline void asc_transto5hd_b32(
5332-// uint16_t src_stride)5343+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5333-#define asc_transto5hd_b32(dst, src, repeat, dst_stride, src_stride) \5344+{
5334- asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride))5345+ asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);
5346+}
5335 5347 
5336-// __aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,5348+[[deprecated("NOTICE: asc_transto5hd_b32_sync is deprecated. "
5337-// uint16_t src_stride)5349+ "Please use asc_transto5hd_b32 and asc_sync instead.")]]
5338-#define asc_transto5hd_b32_sync(dst, src, repeat, dst_stride, src_stride) \5350+__aicore__ inline void asc_transto5hd_b32_sync(
5339- asc_transto5hd_b32_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))5351+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
5352+{
5353+ asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride);
5354+}
5340 5355 
5341// ==========asc_duplicate(half/int16_t/uint16_t/bfloat16_t/float/int32_t/uint32_t)==========5356// ==========asc_duplicate(half/int16_t/uint16_t/bfloat16_t/float/int32_t/uint32_t)==========
5342__aicore__ inline void asc_duplicate(5357__aicore__ inline void asc_duplicate(
Mimpl/c_api/instr_impl/npu_arch_2201/vector_compute_impl/asc_transto5hd_impl.h+153-45
@@ -20,57 +20,165 @@
20 20 
21#include "impl/c_api/instr_impl/npu_arch_2201/utils_impl/utils_impl.h"21#include "impl/c_api/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+#define ASC_TRANSTO5HD_B8_CALL(dst_va, src_va, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
24-// uint16_t src_stride, bool dst_high_half, bool src_high_half)24+ do { \
25-#define asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \25+ if (!(dst_high_half)) { \
26- do { \26+ if (!(src_high_half)) { \
27- if ASC_IS_AIV { \27+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, false, false); \
28- scatter_vnchwconv_b8( \28+ } else { \
29- (dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \29+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, false, true); \
30- } \30+ } \
31+ } else if (!(src_high_half)) { \
32+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, true, false); \
33+ } else { \
34+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, true, true); \
35+ } \
31 } while (0)36 } while (0)
32 37 
33-// __aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t38+__aicore__ inline void asc_transto5hd_b8_impl(
34-// dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)39+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
35-#define asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \40+ bool src_high_half)
36- do { \41+{
37- asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \42+ if ASC_IS_AIV {
38- asc_sync_post_process(); \43+ if (dst == ub_addr8_t::VA0) {
39- } while (0)44+ if (src == ub_addr8_t::VA2) {
45+ ASC_TRANSTO5HD_B8_CALL(VA0, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
46+ } else if (src == ub_addr8_t::VA4) {
47+ ASC_TRANSTO5HD_B8_CALL(VA0, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
48+ } else if (src == ub_addr8_t::VA6) {
49+ ASC_TRANSTO5HD_B8_CALL(VA0, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
50+ }
51+ } else if (dst == ub_addr8_t::VA2) {
52+ if (src == ub_addr8_t::VA0) {
53+ ASC_TRANSTO5HD_B8_CALL(VA2, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
54+ } else if (src == ub_addr8_t::VA4) {
55+ ASC_TRANSTO5HD_B8_CALL(VA2, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
56+ } else if (src == ub_addr8_t::VA6) {
57+ ASC_TRANSTO5HD_B8_CALL(VA2, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
58+ }
59+ } else if (dst == ub_addr8_t::VA4) {
60+ if (src == ub_addr8_t::VA0) {
61+ ASC_TRANSTO5HD_B8_CALL(VA4, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
62+ } else if (src == ub_addr8_t::VA2) {
63+ ASC_TRANSTO5HD_B8_CALL(VA4, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
64+ } else if (src == ub_addr8_t::VA6) {
65+ ASC_TRANSTO5HD_B8_CALL(VA4, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
66+ }
67+ } else if (dst == ub_addr8_t::VA6) {
68+ if (src == ub_addr8_t::VA0) {
69+ ASC_TRANSTO5HD_B8_CALL(VA6, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
70+ } else if (src == ub_addr8_t::VA2) {
71+ ASC_TRANSTO5HD_B8_CALL(VA6, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
72+ } else if (src == ub_addr8_t::VA4) {
73+ ASC_TRANSTO5HD_B8_CALL(VA6, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
74+ }
75+ }
76+ }
77+}
40 78 
41-// __aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,79+#undef ASC_TRANSTO5HD_B8_CALL
42-// uint16_t src_stride)
43-#define asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride) \
44- do { \
45- if ASC_IS_AIV { \
46- scatter_vnchwconv_b16((dst), (src), (repeat), (dst_stride), (src_stride)); \
47- } \
48- } while (0)
49 80 
50-// __aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t81+__aicore__ inline void asc_transto5hd_b8_sync_impl(
51-// dst_stride, uint16_t src_stride)82+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
52-#define asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride) \83+ bool src_high_half)
53- do { \84+{
54- asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \85+ asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
55- asc_sync_post_process(); \86+ asc_sync_post_process();
56- } while (0)87+}
57 88 
58-// __aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,89+__aicore__ inline void asc_transto5hd_b16_impl(
59-// uint16_t src_stride)90+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
60-#define asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride) \91+{
61- do { \92+ if ASC_IS_AIV {
62- if ASC_IS_AIV { \93+ if (dst == ub_addr8_t::VA0) {
63- scatter_vnchwconv_b32((dst), (src), (repeat), (dst_stride), (src_stride)); \94+ if (src == ub_addr8_t::VA2) {
64- } \95+ scatter_vnchwconv_b16(VA0, VA2, repeat, dst_stride, src_stride);
65- } while (0)96+ } else if (src == ub_addr8_t::VA4) {
97+ scatter_vnchwconv_b16(VA0, VA4, repeat, dst_stride, src_stride);
98+ } else if (src == ub_addr8_t::VA6) {
99+ scatter_vnchwconv_b16(VA0, VA6, repeat, dst_stride, src_stride);
100+ }
101+ } else if (dst == ub_addr8_t::VA2) {
102+ if (src == ub_addr8_t::VA0) {
103+ scatter_vnchwconv_b16(VA2, VA0, repeat, dst_stride, src_stride);
104+ } else if (src == ub_addr8_t::VA4) {
105+ scatter_vnchwconv_b16(VA2, VA4, repeat, dst_stride, src_stride);
106+ } else if (src == ub_addr8_t::VA6) {
107+ scatter_vnchwconv_b16(VA2, VA6, repeat, dst_stride, src_stride);
108+ }
109+ } else if (dst == ub_addr8_t::VA4) {
110+ if (src == ub_addr8_t::VA0) {
111+ scatter_vnchwconv_b16(VA4, VA0, repeat, dst_stride, src_stride);
112+ } else if (src == ub_addr8_t::VA2) {
113+ scatter_vnchwconv_b16(VA4, VA2, repeat, dst_stride, src_stride);
114+ } else if (src == ub_addr8_t::VA6) {
115+ scatter_vnchwconv_b16(VA4, VA6, repeat, dst_stride, src_stride);
116+ }
117+ } else if (dst == ub_addr8_t::VA6) {
118+ if (src == ub_addr8_t::VA0) {
119+ scatter_vnchwconv_b16(VA6, VA0, repeat, dst_stride, src_stride);
120+ } else if (src == ub_addr8_t::VA2) {
121+ scatter_vnchwconv_b16(VA6, VA2, repeat, dst_stride, src_stride);
122+ } else if (src == ub_addr8_t::VA4) {
123+ scatter_vnchwconv_b16(VA6, VA4, repeat, dst_stride, src_stride);
124+ }
125+ }
126+ }
127+}
66 128 
67-// __aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t129+__aicore__ inline void asc_transto5hd_b16_sync_impl(
68-// dst_stride, uint16_t src_stride)130+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
69-#define asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride) \131+{
70- do { \132+ asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);
71- asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \133+ asc_sync_post_process();
72- asc_sync_post_process(); \134+}
73- } while (0)135+ 
136+__aicore__ inline void asc_transto5hd_b32_impl(
137+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
138+{
139+ if ASC_IS_AIV {
140+ if (dst == ub_addr8_t::VA0) {
141+ if (src == ub_addr8_t::VA2) {
142+ scatter_vnchwconv_b32(VA0, VA2, repeat, dst_stride, src_stride);
143+ } else if (src == ub_addr8_t::VA4) {
144+ scatter_vnchwconv_b32(VA0, VA4, repeat, dst_stride, src_stride);
145+ } else if (src == ub_addr8_t::VA6) {
146+ scatter_vnchwconv_b32(VA0, VA6, repeat, dst_stride, src_stride);
147+ }
148+ } else if (dst == ub_addr8_t::VA2) {
149+ if (src == ub_addr8_t::VA0) {
150+ scatter_vnchwconv_b32(VA2, VA0, repeat, dst_stride, src_stride);
151+ } else if (src == ub_addr8_t::VA4) {
152+ scatter_vnchwconv_b32(VA2, VA4, repeat, dst_stride, src_stride);
153+ } else if (src == ub_addr8_t::VA6) {
154+ scatter_vnchwconv_b32(VA2, VA6, repeat, dst_stride, src_stride);
155+ }
156+ } else if (dst == ub_addr8_t::VA4) {
157+ if (src == ub_addr8_t::VA0) {
158+ scatter_vnchwconv_b32(VA4, VA0, repeat, dst_stride, src_stride);
159+ } else if (src == ub_addr8_t::VA2) {
160+ scatter_vnchwconv_b32(VA4, VA2, repeat, dst_stride, src_stride);
161+ } else if (src == ub_addr8_t::VA6) {
162+ scatter_vnchwconv_b32(VA4, VA6, repeat, dst_stride, src_stride);
163+ }
164+ } else if (dst == ub_addr8_t::VA6) {
165+ if (src == ub_addr8_t::VA0) {
166+ scatter_vnchwconv_b32(VA6, VA0, repeat, dst_stride, src_stride);
167+ } else if (src == ub_addr8_t::VA2) {
168+ scatter_vnchwconv_b32(VA6, VA2, repeat, dst_stride, src_stride);
169+ } else if (src == ub_addr8_t::VA4) {
170+ scatter_vnchwconv_b32(VA6, VA4, repeat, dst_stride, src_stride);
171+ }
172+ }
173+ }
174+}
175+ 
176+__aicore__ inline void asc_transto5hd_b32_sync_impl(
177+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
178+{
179+ asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);
180+ asc_sync_post_process();
181+}
74 182 
75#endif183#endif
76 184 
Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl.h+39-24
@@ -9909,35 +9909,50 @@ __aicore__ inline void asc_transpose_sync(__ubuf__ uint16_t* dst, __ubuf__ uint1
9909 asc_transpose_sync_impl(dst, src);9909 asc_transpose_sync_impl(dst, src);
9910}9910}
9911 9911 
9912-// __aicore__ inline void asc_transto5hd_b8(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9912+// ==========asc_transto5hd(b8/b16/b32)==========
9913-// uint16_t src_stride, bool dst_high_half, bool src_high_half)9913+__aicore__ inline void asc_transto5hd_b8(
9914-#define asc_transto5hd_b8(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \9914+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
9915- asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))9915+ bool src_high_half)
9916+{
9917+ asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
9918+}
9916 9919 
9917-// __aicore__ inline void asc_transto5hd_b8_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9920+[[deprecated("NOTICE: asc_transto5hd_b8_sync is deprecated. "
9918-// uint16_t src_stride, bool dst_high_half, bool src_high_half)9921+ "Please use asc_transto5hd_b8 and asc_sync instead.")]]
9919-#define asc_transto5hd_b8_sync(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \9922+__aicore__ inline void asc_transto5hd_b8_sync(
9920- asc_transto5hd_b8_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half))9923+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
9924+ bool src_high_half)
9925+{
9926+ asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
9927+}
9921 9928 
9922-// __aicore__ inline void asc_transto5hd_b16(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9929+__aicore__ inline void asc_transto5hd_b16(
9923-// uint16_t src_stride)9930+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
9924-#define asc_transto5hd_b16(dst, src, repeat, dst_stride, src_stride) \9931+{
9925- asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride))9932+ asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);
9933+}
9926 9934 
9927-// __aicore__ inline void asc_transto5hd_b16_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9935+[[deprecated("NOTICE: asc_transto5hd_b16_sync is deprecated. "
9928-// uint16_t src_stride)9936+ "Please use asc_transto5hd_b16 and asc_sync instead.")]]
9929-#define asc_transto5hd_b16_sync(dst, src, repeat, dst_stride, src_stride) \9937+__aicore__ inline void asc_transto5hd_b16_sync(
9930- asc_transto5hd_b16_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))9938+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
9939+{
9940+ asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride);
9941+}
9931 9942 
9932-// __aicore__ inline void asc_transto5hd_b32(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9943+__aicore__ inline void asc_transto5hd_b32(
9933-// uint16_t src_stride)9944+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
9934-#define asc_transto5hd_b32(dst, src, repeat, dst_stride, src_stride) \9945+{
9935- asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride))9946+ asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);
9947+}
9936 9948 
9937-// __aicore__ inline void asc_transto5hd_b32_sync(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,9949+[[deprecated("NOTICE: asc_transto5hd_b32_sync is deprecated. "
9938-// uint16_t src_stride)9950+ "Please use asc_transto5hd_b32 and asc_sync instead.")]]
9939-#define asc_transto5hd_b32_sync(dst, src, repeat, dst_stride, src_stride) \9951+__aicore__ inline void asc_transto5hd_b32_sync(
L
LLycheeeee20 天前

_sync 确认是否废弃。

likedislike
leozjr
leozjr
19 天前 评论：
9940- asc_transto5hd_b32_sync_impl((dst), (src), (repeat), (dst_stride), (src_stride))9952+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
9953+{
9954+ asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride);
9955+}
9941 9956 
9942// ===================asc_set_va_reg=====================9957// ===================asc_set_va_reg=====================
9943__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int8_t** src_array)9958__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int8_t** src_array)
Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl/asc_transto5hd_impl.h+153-45
@@ -24,57 +24,165 @@
24 24 
25#include "impl/c_api/instr_impl/npu_arch_3510/utils_impl/utils_impl.h"25#include "impl/c_api/instr_impl/npu_arch_3510/utils_impl/utils_impl.h"
26 26 
27-// __aicore__ inline void asc_transto5hd_b8_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,27+#define ASC_TRANSTO5HD_B8_CALL(dst_va, src_va, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \
28-// uint16_t src_stride, bool dst_high_half, bool src_high_half)28+ do { \
29-#define asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \29+ if (!(dst_high_half)) { \
30- do { \30+ if (!(src_high_half)) { \
31- if ASC_IS_AIV { \31+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, false, false); \
32- scatter_vnchwconv_b8( \32+ } else { \
33- (dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \33+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, false, true); \
34- } \34+ } \
35+ } else if (!(src_high_half)) { \
36+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, true, false); \
37+ } else { \
38+ scatter_vnchwconv_b8(dst_va, src_va, repeat, dst_stride, src_stride, true, true); \
39+ } \
35 } while (0)40 } while (0)
36 41 
37-// __aicore__ inline void asc_transto5hd_b8_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t42+__aicore__ inline void asc_transto5hd_b8_impl(
38-// dst_stride, uint16_t src_stride, bool dst_high_half, bool src_high_half)43+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
39-#define asc_transto5hd_b8_sync_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half) \44+ bool src_high_half)
40- do { \45+{
41- asc_transto5hd_b8_impl((dst), (src), (repeat), (dst_stride), (src_stride), (dst_high_half), (src_high_half)); \46+ if ASC_IS_AIV {
42- asc_sync_post_process(); \47+ if (dst == ub_addr8_t::VA0) {
43- } while (0)48+ if (src == ub_addr8_t::VA2) {
49+ ASC_TRANSTO5HD_B8_CALL(VA0, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
50+ } else if (src == ub_addr8_t::VA4) {
51+ ASC_TRANSTO5HD_B8_CALL(VA0, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
52+ } else if (src == ub_addr8_t::VA6) {
53+ ASC_TRANSTO5HD_B8_CALL(VA0, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
54+ }
55+ } else if (dst == ub_addr8_t::VA2) {
56+ if (src == ub_addr8_t::VA0) {
57+ ASC_TRANSTO5HD_B8_CALL(VA2, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
58+ } else if (src == ub_addr8_t::VA4) {
59+ ASC_TRANSTO5HD_B8_CALL(VA2, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
60+ } else if (src == ub_addr8_t::VA6) {
61+ ASC_TRANSTO5HD_B8_CALL(VA2, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
62+ }
63+ } else if (dst == ub_addr8_t::VA4) {
64+ if (src == ub_addr8_t::VA0) {
65+ ASC_TRANSTO5HD_B8_CALL(VA4, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
66+ } else if (src == ub_addr8_t::VA2) {
67+ ASC_TRANSTO5HD_B8_CALL(VA4, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
68+ } else if (src == ub_addr8_t::VA6) {
69+ ASC_TRANSTO5HD_B8_CALL(VA4, VA6, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
70+ }
71+ } else if (dst == ub_addr8_t::VA6) {
72+ if (src == ub_addr8_t::VA0) {
73+ ASC_TRANSTO5HD_B8_CALL(VA6, VA0, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
74+ } else if (src == ub_addr8_t::VA2) {
75+ ASC_TRANSTO5HD_B8_CALL(VA6, VA2, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
76+ } else if (src == ub_addr8_t::VA4) {
77+ ASC_TRANSTO5HD_B8_CALL(VA6, VA4, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
78+ }
79+ }
80+ }
81+}
44 82 
45-// __aicore__ inline void asc_transto5hd_b16_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,83+#undef ASC_TRANSTO5HD_B8_CALL
46-// uint16_t src_stride)
47-#define asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride) \
48- do { \
49- if ASC_IS_AIV { \
50- scatter_vnchwconv_b16((dst), (src), (repeat), (dst_stride), (src_stride)); \
51- } \
52- } while (0)
53 84 
54-// __aicore__ inline void asc_transto5hd_b16_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t85+__aicore__ inline void asc_transto5hd_b8_sync_impl(
maohp_hw
maohp_hwmaohp_hw16 天前

不再提供_sync形式的接口

likedislike
leozjr
leozjr
5 天前 评论：
55-// dst_stride, uint16_t src_stride)86+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
56-#define asc_transto5hd_b16_sync_impl(dst, src, repeat, dst_stride, src_stride) \87+ bool src_high_half)
57- do { \88+{
58- asc_transto5hd_b16_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \89+ asc_transto5hd_b8_impl(dst, src, repeat, dst_stride, src_stride, dst_high_half, src_high_half);
59- asc_sync_post_process(); \90+ asc_sync_post_process();
60- } while (0)91+}
61 92 
62-// __aicore__ inline void asc_transto5hd_b32_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride,93+__aicore__ inline void asc_transto5hd_b16_impl(
63-// uint16_t src_stride)94+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
64-#define asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride) \95+{
65- do { \96+ if ASC_IS_AIV {
66- if ASC_IS_AIV { \97+ if (dst == ub_addr8_t::VA0) {
67- scatter_vnchwconv_b32((dst), (src), (repeat), (dst_stride), (src_stride)); \98+ if (src == ub_addr8_t::VA2) {
68- } \99+ scatter_vnchwconv_b16(VA0, VA2, repeat, dst_stride, src_stride);
69- } while (0)100+ } else if (src == ub_addr8_t::VA4) {
101+ scatter_vnchwconv_b16(VA0, VA4, repeat, dst_stride, src_stride);
102+ } else if (src == ub_addr8_t::VA6) {
103+ scatter_vnchwconv_b16(VA0, VA6, repeat, dst_stride, src_stride);
104+ }
105+ } else if (dst == ub_addr8_t::VA2) {
106+ if (src == ub_addr8_t::VA0) {
107+ scatter_vnchwconv_b16(VA2, VA0, repeat, dst_stride, src_stride);
108+ } else if (src == ub_addr8_t::VA4) {
109+ scatter_vnchwconv_b16(VA2, VA4, repeat, dst_stride, src_stride);
110+ } else if (src == ub_addr8_t::VA6) {
111+ scatter_vnchwconv_b16(VA2, VA6, repeat, dst_stride, src_stride);
112+ }
113+ } else if (dst == ub_addr8_t::VA4) {
114+ if (src == ub_addr8_t::VA0) {
115+ scatter_vnchwconv_b16(VA4, VA0, repeat, dst_stride, src_stride);
116+ } else if (src == ub_addr8_t::VA2) {
117+ scatter_vnchwconv_b16(VA4, VA2, repeat, dst_stride, src_stride);
118+ } else if (src == ub_addr8_t::VA6) {
119+ scatter_vnchwconv_b16(VA4, VA6, repeat, dst_stride, src_stride);
120+ }
121+ } else if (dst == ub_addr8_t::VA6) {
122+ if (src == ub_addr8_t::VA0) {
123+ scatter_vnchwconv_b16(VA6, VA0, repeat, dst_stride, src_stride);
124+ } else if (src == ub_addr8_t::VA2) {
125+ scatter_vnchwconv_b16(VA6, VA2, repeat, dst_stride, src_stride);
126+ } else if (src == ub_addr8_t::VA4) {
127+ scatter_vnchwconv_b16(VA6, VA4, repeat, dst_stride, src_stride);
128+ }
129+ }
130+ }
131+}
70 132 
71-// __aicore__ inline void asc_transto5hd_b32_sync_impl(ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t133+__aicore__ inline void asc_transto5hd_b16_sync_impl(
72-// dst_stride, uint16_t src_stride)134+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
73-#define asc_transto5hd_b32_sync_impl(dst, src, repeat, dst_stride, src_stride) \135+{
74- do { \136+ asc_transto5hd_b16_impl(dst, src, repeat, dst_stride, src_stride);
75- asc_transto5hd_b32_impl((dst), (src), (repeat), (dst_stride), (src_stride)); \137+ asc_sync_post_process();
76- asc_sync_post_process(); \138+}
77- } while (0)139+ 
140+__aicore__ inline void asc_transto5hd_b32_impl(
141+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
142+{
143+ if ASC_IS_AIV {
144+ if (dst == ub_addr8_t::VA0) {
145+ if (src == ub_addr8_t::VA2) {
146+ scatter_vnchwconv_b32(VA0, VA2, repeat, dst_stride, src_stride);
147+ } else if (src == ub_addr8_t::VA4) {
148+ scatter_vnchwconv_b32(VA0, VA4, repeat, dst_stride, src_stride);
149+ } else if (src == ub_addr8_t::VA6) {
150+ scatter_vnchwconv_b32(VA0, VA6, repeat, dst_stride, src_stride);
151+ }
152+ } else if (dst == ub_addr8_t::VA2) {
153+ if (src == ub_addr8_t::VA0) {
154+ scatter_vnchwconv_b32(VA2, VA0, repeat, dst_stride, src_stride);
155+ } else if (src == ub_addr8_t::VA4) {
156+ scatter_vnchwconv_b32(VA2, VA4, repeat, dst_stride, src_stride);
157+ } else if (src == ub_addr8_t::VA6) {
158+ scatter_vnchwconv_b32(VA2, VA6, repeat, dst_stride, src_stride);
159+ }
160+ } else if (dst == ub_addr8_t::VA4) {
161+ if (src == ub_addr8_t::VA0) {
162+ scatter_vnchwconv_b32(VA4, VA0, repeat, dst_stride, src_stride);
163+ } else if (src == ub_addr8_t::VA2) {
164+ scatter_vnchwconv_b32(VA4, VA2, repeat, dst_stride, src_stride);
165+ } else if (src == ub_addr8_t::VA6) {
166+ scatter_vnchwconv_b32(VA4, VA6, repeat, dst_stride, src_stride);
167+ }
168+ } else if (dst == ub_addr8_t::VA6) {
169+ if (src == ub_addr8_t::VA0) {
170+ scatter_vnchwconv_b32(VA6, VA0, repeat, dst_stride, src_stride);
171+ } else if (src == ub_addr8_t::VA2) {
172+ scatter_vnchwconv_b32(VA6, VA2, repeat, dst_stride, src_stride);
173+ } else if (src == ub_addr8_t::VA4) {
174+ scatter_vnchwconv_b32(VA6, VA4, repeat, dst_stride, src_stride);
175+ }
176+ }
177+ }
178+}
179+ 
180+__aicore__ inline void asc_transto5hd_b32_sync_impl(
181+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
182+{
183+ asc_transto5hd_b32_impl(dst, src, repeat, dst_stride, src_stride);
184+ asc_sync_post_process();
185+}
78 186 
79#endif187#endif
80 188 
Minclude/c_api/reg_compute/compute/reg_permute_sel.h+0-16
@@ -416,22 +416,6 @@ __simd_callee__ inline vector_bool asc_unpack_upper(vector_bool src);
416 416 
417__simd_callee__ inline vector_bool asc_unpack_lower(vector_bool src);417__simd_callee__ inline vector_bool asc_unpack_lower(vector_bool src);
418 418 
419-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int8_t** src_array);
420- 
421-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ uint8_t** src_array);
422- 
423-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int16_t** src_array);
424- 
425-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ uint16_t** src_array);
426- 
427-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ half** src_array);
428- 
429-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ int32_t** src_array);
430- 
431-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ uint32_t** src_array);
432- 
433-__aicore__ inline void asc_set_va_reg(ub_addr8_t addr, __ubuf__ float** src_array);
434- 
435[[deprecated("NOTICE: asc_squeeze_with_status is deprecated. "419[[deprecated("NOTICE: asc_squeeze_with_status is deprecated. "
436 "Please use asc_squeeze_and_storeunalign instead.")]] __simd_callee__ inline void420 "Please use asc_squeeze_and_storeunalign instead.")]] __simd_callee__ inline void
437asc_squeeze_with_status(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);421asc_squeeze_with_status(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);
Minclude/c_api/vector_compute/compute/vector_convert.h+10-0
@@ -26,6 +26,16 @@
26#elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3510)26#elif defined(__NPU_ARCH__) && (__NPU_ARCH__ == 3510)
27#include "impl/c_api/instr_impl/npu_arch_3510/vector_compute_impl.h"27#include "impl/c_api/instr_impl/npu_arch_3510/vector_compute_impl.h"
28#endif28#endif
29+ 
30+// ==========asc_transto5hd(b8/b16/b32)==========
31+__aicore__ inline void asc_transto5hd_b8(
32+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride, bool dst_high_half,
33+ bool src_high_half);
34+__aicore__ inline void asc_transto5hd_b16(
35+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride);
36+__aicore__ inline void asc_transto5hd_b32(
37+ ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride);
38+ 
29__aicore__ inline void asc_bfloat162float(39__aicore__ inline void asc_bfloat162float(
30 __ubuf__ float* dst, __ubuf__ bfloat16_t* src, uint8_t repeat, uint16_t dst_block_stride, uint16_t src_block_stride,40 __ubuf__ float* dst, __ubuf__ bfloat16_t* src, uint8_t repeat, uint16_t dst_block_stride, uint16_t src_block_stride,
31 uint16_t dst_repeat_stride, uint16_t src_repeat_stride);41 uint16_t dst_repeat_stride, uint16_t src_repeat_stride);
Mtests/api/c_api/ascendc_host_header_checker/host_header_reg_compute_checker.cpp+0-1
@@ -610,7 +610,6 @@ static void test_host_c_api_reg_compute_17()
610 using ::asc_round;610 using ::asc_round;
611 using ::asc_scatter;611 using ::asc_scatter;
612 using ::asc_select;612 using ::asc_select;
613- using ::asc_set_va_reg;
614 using ::asc_shiftleft;613 using ::asc_shiftleft;
615 using ::asc_shiftleft_scalar;614 using ::asc_shiftleft_scalar;
616 using ::asc_shiftright;615 using ::asc_shiftright;
Mtests/api/c_api/npu_arch_2201/vector_compute/test_transto5hd.cpp+7-7
@@ -36,8 +36,8 @@ void scatter_vnchwconv_b8_stub(
36void scatter_vnchwconv_b16_stub(36void scatter_vnchwconv_b16_stub(
37 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)37 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
38{38{
39- EXPECT_EQ(ub_addr8_t::VA1, dst);39+ EXPECT_EQ(ub_addr8_t::VA2, dst);
40- EXPECT_EQ(ub_addr8_t::VA3, src);40+ EXPECT_EQ(ub_addr8_t::VA4, src);
41 EXPECT_EQ(repeat, static_cast<uint8_t>(5));41 EXPECT_EQ(repeat, static_cast<uint8_t>(5));
42 EXPECT_EQ(dst_stride, static_cast<uint16_t>(6));42 EXPECT_EQ(dst_stride, static_cast<uint16_t>(6));
43 EXPECT_EQ(src_stride, static_cast<uint16_t>(7));43 EXPECT_EQ(src_stride, static_cast<uint16_t>(7));
@@ -47,7 +47,7 @@ void scatter_vnchwconv_b32_stub(
47 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)47 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
48{48{
49 EXPECT_EQ(ub_addr8_t::VA4, dst);49 EXPECT_EQ(ub_addr8_t::VA4, dst);
50- EXPECT_EQ(ub_addr8_t::VA5, src);50+ EXPECT_EQ(ub_addr8_t::VA6, src);
51 EXPECT_EQ(repeat, static_cast<uint8_t>(8));51 EXPECT_EQ(repeat, static_cast<uint8_t>(8));
52 EXPECT_EQ(dst_stride, static_cast<uint16_t>(9));52 EXPECT_EQ(dst_stride, static_cast<uint16_t>(9));
53 EXPECT_EQ(src_stride, static_cast<uint16_t>(10));53 EXPECT_EQ(src_stride, static_cast<uint16_t>(10));
@@ -94,7 +94,7 @@ TEST_F(TestTransdataTo5HDCAPI, asc_transto5hd_b16_Succ)
94 .times(1)94 .times(1)
95 .will(invoke(scatter_vnchwconv_b16_stub));95 .will(invoke(scatter_vnchwconv_b16_stub));
96 96 
97- asc_transto5hd_b16(ub_addr8_t::VA1, ub_addr8_t::VA3, repeat, dst_stride, src_stride);97+ asc_transto5hd_b16(ub_addr8_t::VA2, ub_addr8_t::VA4, repeat, dst_stride, src_stride);
98 GlobalMockObject::verify();98 GlobalMockObject::verify();
99}99}
100 100 
@@ -107,7 +107,7 @@ TEST_F(TestTransdataTo5HDCAPI, asc_transto5hd_b16_sync_Succ)
107 .times(1)107 .times(1)
108 .will(invoke(scatter_vnchwconv_b16_stub));108 .will(invoke(scatter_vnchwconv_b16_stub));
109 109 
110- asc_transto5hd_b16_sync(ub_addr8_t::VA1, ub_addr8_t::VA3, repeat, dst_stride, src_stride);110+ asc_transto5hd_b16_sync(ub_addr8_t::VA2, ub_addr8_t::VA4, repeat, dst_stride, src_stride);
111 GlobalMockObject::verify();111 GlobalMockObject::verify();
112}112}
113 113 
@@ -120,7 +120,7 @@ TEST_F(TestTransdataTo5HDCAPI, asc_transto5hd_b32_Succ)
120 .times(1)120 .times(1)
121 .will(invoke(scatter_vnchwconv_b32_stub));121 .will(invoke(scatter_vnchwconv_b32_stub));
122 122 
123- asc_transto5hd_b32(ub_addr8_t::VA4, ub_addr8_t::VA5, repeat, dst_stride, src_stride);123+ asc_transto5hd_b32(ub_addr8_t::VA4, ub_addr8_t::VA6, repeat, dst_stride, src_stride);
124 GlobalMockObject::verify();124 GlobalMockObject::verify();
125}125}
126 126 
@@ -133,6 +133,6 @@ TEST_F(TestTransdataTo5HDCAPI, asc_transto5hd_b32_sync_Succ)
133 .times(1)133 .times(1)
134 .will(invoke(scatter_vnchwconv_b32_stub));134 .will(invoke(scatter_vnchwconv_b32_stub));
135 135 
136- asc_transto5hd_b32_sync(ub_addr8_t::VA4, ub_addr8_t::VA5, repeat, dst_stride, src_stride);136+ asc_transto5hd_b32_sync(ub_addr8_t::VA4, ub_addr8_t::VA6, repeat, dst_stride, src_stride);
137 GlobalMockObject::verify();137 GlobalMockObject::verify();
138}138}
Mtests/api/c_api/npu_arch_3510/vector_compute/test_transto5hd.cpp+7-7
@@ -36,8 +36,8 @@ void scatter_vnchwconv_b8_stub_repeat(
36void scatter_vnchwconv_b16_stub_repeat(36void scatter_vnchwconv_b16_stub_repeat(
37 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)37 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
38{38{
39- EXPECT_EQ(ub_addr8_t::VA1, dst);39+ EXPECT_EQ(ub_addr8_t::VA2, dst);
40- EXPECT_EQ(ub_addr8_t::VA3, src);40+ EXPECT_EQ(ub_addr8_t::VA4, src);
41 EXPECT_EQ(repeat, static_cast<uint8_t>(5));41 EXPECT_EQ(repeat, static_cast<uint8_t>(5));
42 EXPECT_EQ(dst_stride, static_cast<uint16_t>(6));42 EXPECT_EQ(dst_stride, static_cast<uint16_t>(6));
43 EXPECT_EQ(src_stride, static_cast<uint16_t>(7));43 EXPECT_EQ(src_stride, static_cast<uint16_t>(7));
@@ -47,7 +47,7 @@ void scatter_vnchwconv_b32_stub_repeat(
47 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)47 ub_addr8_t dst, ub_addr8_t src, uint8_t repeat, uint16_t dst_stride, uint16_t src_stride)
48{48{
49 EXPECT_EQ(ub_addr8_t::VA4, dst);49 EXPECT_EQ(ub_addr8_t::VA4, dst);
50- EXPECT_EQ(ub_addr8_t::VA5, src);50+ EXPECT_EQ(ub_addr8_t::VA6, src);
51 EXPECT_EQ(repeat, static_cast<uint8_t>(8));51 EXPECT_EQ(repeat, static_cast<uint8_t>(8));
52 EXPECT_EQ(dst_stride, static_cast<uint16_t>(9));52 EXPECT_EQ(dst_stride, static_cast<uint16_t>(9));
53 EXPECT_EQ(src_stride, static_cast<uint16_t>(10));53 EXPECT_EQ(src_stride, static_cast<uint16_t>(10));
@@ -97,7 +97,7 @@ TEST_F(TestTransto5HDCAPI, asc_transto5hd_b16_repeat_Succ)
97 .times(1)97 .times(1)
98 .will(invoke(scatter_vnchwconv_b16_stub_repeat));98 .will(invoke(scatter_vnchwconv_b16_stub_repeat));
99 99 
100- asc_transto5hd_b16(ub_addr8_t::VA1, ub_addr8_t::VA3, repeat, dst_stride, src_stride);100+ asc_transto5hd_b16(ub_addr8_t::VA2, ub_addr8_t::VA4, repeat, dst_stride, src_stride);
101 GlobalMockObject::verify();101 GlobalMockObject::verify();
102}102}
103 103 
@@ -111,7 +111,7 @@ TEST_F(TestTransto5HDCAPI, asc_transto5hd_b16_repeat_sync_Succ)
111 .times(1)111 .times(1)
112 .will(invoke(scatter_vnchwconv_b16_stub_repeat));112 .will(invoke(scatter_vnchwconv_b16_stub_repeat));
113 113 
114- asc_transto5hd_b16_sync(ub_addr8_t::VA1, ub_addr8_t::VA3, repeat, dst_stride, src_stride);114+ asc_transto5hd_b16_sync(ub_addr8_t::VA2, ub_addr8_t::VA4, repeat, dst_stride, src_stride);
115 GlobalMockObject::verify();115 GlobalMockObject::verify();
116}116}
117 117 
@@ -125,7 +125,7 @@ TEST_F(TestTransto5HDCAPI, asc_transto5hd_b32_repeat_Succ)
125 .times(1)125 .times(1)
126 .will(invoke(scatter_vnchwconv_b32_stub_repeat));126 .will(invoke(scatter_vnchwconv_b32_stub_repeat));
127 127 
128- asc_transto5hd_b32(ub_addr8_t::VA4, ub_addr8_t::VA5, repeat, dst_stride, src_stride);128+ asc_transto5hd_b32(ub_addr8_t::VA4, ub_addr8_t::VA6, repeat, dst_stride, src_stride);
129 GlobalMockObject::verify();129 GlobalMockObject::verify();
130}130}
131 131 
@@ -139,6 +139,6 @@ TEST_F(TestTransto5HDCAPI, asc_transto5hd_b32_repeat_sync_Succ)
139 .times(1)139 .times(1)
140 .will(invoke(scatter_vnchwconv_b32_stub_repeat));140 .will(invoke(scatter_vnchwconv_b32_stub_repeat));
141 141 
142- asc_transto5hd_b32_sync(ub_addr8_t::VA4, ub_addr8_t::VA5, repeat, dst_stride, src_stride);142+ asc_transto5hd_b32_sync(ub_addr8_t::VA4, ub_addr8_t::VA6, repeat, dst_stride, src_stride);
143 GlobalMockObject::verify();143 GlobalMockObject::verify();
144}144}