Star697
1005
代码
代码
Issues90
Pull Requests186
Actions
项目讨论
Wiki
分析
Star697
1005
已合并
【C API】feat: 新增asc_ld_dev并支持duplicate highest模式 #5039
guojianyang创建于 16 天前
【C API】feat: 新增asc_ld_dev并支持duplicate highest模式 #5039
已合并
guojianyang创建于 16 天前
从Joey_guo/asc-devkitdcache合入到cann/asc-devkitmaster
共 18 个文件变更+942-97
Mdocs/zh/api/README.md+3-0
@@ -1156,6 +1156,8 @@
1156 - [asc_pack（废弃）](SIMD-API/c_api/reg/data_compress/asc_pack_deprecated.md)1156 - [asc_pack（废弃）](SIMD-API/c_api/reg/data_compress/asc_pack_deprecated.md)
1157 - [数据填充](SIMD-API/c_api/reg/data_fill/data_fill.md)1157 - [数据填充](SIMD-API/c_api/reg/data_fill/data_fill.md)
1158 - [asc_duplicate](SIMD-API/c_api/reg/data_fill/asc_duplicate.md)1158 - [asc_duplicate](SIMD-API/c_api/reg/data_fill/asc_duplicate.md)
1159+ - [asc_duplicate_highest](SIMD-API/c_api/reg/data_fill/asc_duplicate_highest.md)
1160+ - [asc_duplicate_highest_merge](SIMD-API/c_api/reg/data_fill/asc_duplicate_highest_merge.md)
1159 - [asc_duplicate_merge](SIMD-API/c_api/reg/data_fill/asc_duplicate_merge.md)1161 - [asc_duplicate_merge](SIMD-API/c_api/reg/data_fill/asc_duplicate_merge.md)
1160 - [asc_duplicate_scalar](SIMD-API/c_api/reg/data_fill/asc_duplicate_scalar.md)1162 - [asc_duplicate_scalar](SIMD-API/c_api/reg/data_fill/asc_duplicate_scalar.md)
1161 - [asc_duplicate_scalar_merge](SIMD-API/c_api/reg/data_fill/asc_duplicate_scalar_merge.md)1163 - [asc_duplicate_scalar_merge](SIMD-API/c_api/reg/data_fill/asc_duplicate_scalar_merge.md)
@@ -1250,6 +1252,7 @@
1250 - [asc_ffs](SIMD-API/c_api/scalar_compute/asc_ffs.md)1252 - [asc_ffs](SIMD-API/c_api/scalar_compute/asc_ffs.md)
1251 - [asc_ffz](SIMD-API/c_api/scalar_compute/asc_ffz.md)1253 - [asc_ffz](SIMD-API/c_api/scalar_compute/asc_ffz.md)
1252 - [asc_float2int32](SIMD-API/c_api/scalar_compute/asc_float2int32.md)1254 - [asc_float2int32](SIMD-API/c_api/scalar_compute/asc_float2int32.md)
1255+ - [asc_load_dev](SIMD-API/c_api/scalar_compute/asc_load_dev.md)
1253 - [asc_popc](SIMD-API/c_api/scalar_compute/asc_popc.md)1256 - [asc_popc](SIMD-API/c_api/scalar_compute/asc_popc.md)
1254 - [asc_set_nthbit](SIMD-API/c_api/scalar_compute/asc_set_nthbit.md)1257 - [asc_set_nthbit](SIMD-API/c_api/scalar_compute/asc_set_nthbit.md)
1255 - [asc_sflbits](SIMD-API/c_api/scalar_compute/asc_sflbits.md)1258 - [asc_sflbits](SIMD-API/c_api/scalar_compute/asc_sflbits.md)
Mdocs/zh/api/SIMD-API/c_api/c_api.md+3-0
@@ -201,6 +201,7 @@ C API文档目录，整体使用时可以引入asc_simd.h，C API列表如下：
201| [asc_ffs](scalar_compute/asc_ffs.md) | FindFirstSet接口，输入数据的二进制表示中从最低位向最高位查找第一个值为1的位，并返回其位置，如果没找到则返回-1。 |201| [asc_ffs](scalar_compute/asc_ffs.md) | FindFirstSet接口，输入数据的二进制表示中从最低位向最高位查找第一个值为1的位，并返回其位置，如果没找到则返回-1。 |
202| [asc_ffz](scalar_compute/asc_ffz.md) | 获取一个uint64_t类型数字的二进制表示中从最低有效位开始的第一个0出现的位置，如果没找到则返回-1。 |202| [asc_ffz](scalar_compute/asc_ffz.md) | 获取一个uint64_t类型数字的二进制表示中从最低有效位开始的第一个0出现的位置，如果没找到则返回-1。 |
203| [asc_float2int32](scalar_compute/asc_float2int32.md) | 将float类型转化为int32_t类型，并支持多种舍入模式。 |203| [asc_float2int32](scalar_compute/asc_float2int32.md) | 将float类型转化为int32_t类型，并支持多种舍入模式。 |
204+| [asc_load_dev](scalar_compute/asc_load_dev.md) | 不经过DCache直接从GM地址读取整型数据。 |
204| [asc_popc](scalar_compute/asc_popc.md) | 获取一个uint64_t类型数字的二进制中1的个数。 |205| [asc_popc](scalar_compute/asc_popc.md) | 获取一个uint64_t类型数字的二进制中1的个数。 |
205| [asc_set_nthbit](scalar_compute/asc_set_nthbit.md) | 计算一个uint64_t类型数字的指定二进制位置为1，其余位保持不变。 |206| [asc_set_nthbit](scalar_compute/asc_set_nthbit.md) | 计算一个uint64_t类型数字的指定二进制位置为1，其余位保持不变。 |
206| [asc_sflbits](scalar_compute/asc_sflbits.md) | 计算一个int64_t类型数字的二进制中，从最高数值位开始与符号位相同的连续比特位的个数。 |207| [asc_sflbits](scalar_compute/asc_sflbits.md) | 计算一个int64_t类型数字的二进制中，从最高数值位开始与符号位相同的连续比特位的个数。 |
@@ -533,6 +534,8 @@ Reg矢量计算类API，单独使用时可以引入reg_vector.h，此类API列
533| [asc_unpack](reg/data_compress/asc_unpack.md) | 矢量解包操作。 |534| [asc_unpack](reg/data_compress/asc_unpack.md) | 矢量解包操作。 |
534| [asc_unsqueeze](reg/data_compress/asc_unsqueeze.md) | 根据mask进行解压缩，将生成的数据输出到dst。 |535| [asc_unsqueeze](reg/data_compress/asc_unsqueeze.md) | 根据mask进行解压缩，将生成的数据输出到dst。 |
535| [asc_duplicate](reg/data_fill/asc_duplicate.md) | 根据mask将源操作数src的最低位元素填充到目的操作数dst。 |536| [asc_duplicate](reg/data_fill/asc_duplicate.md) | 根据mask将源操作数src的最低位元素填充到目的操作数dst。 |
537+| [asc_duplicate_highest](reg/data_fill/asc_duplicate_highest.md) | 根据mask将源操作数src的最高位元素填充到目的操作数dst，dst中未被mask筛选的元素置为0。 |
538+| [asc_duplicate_highest_merge](reg/data_fill/asc_duplicate_highest_merge.md) | 根据mask将源操作数src的最高位元素填充到目的操作数dst，dst中未被mask筛选的元素保留原值。 |
536| [asc_duplicate_merge](reg/data_fill/asc_duplicate_merge.md) | 根据mask将源操作数src的最低位元素填充到目的操作数dst，dst中未被mask筛选的元素保留原值。 |539| [asc_duplicate_merge](reg/data_fill/asc_duplicate_merge.md) | 根据mask将源操作数src的最低位元素填充到目的操作数dst，dst中未被mask筛选的元素保留原值。 |
537| [asc_duplicate_scalar](reg/data_fill/asc_duplicate_scalar.md) | 根据mask将value填充到目的操作数dst。 |540| [asc_duplicate_scalar](reg/data_fill/asc_duplicate_scalar.md) | 根据mask将value填充到目的操作数dst。 |
538| [asc_duplicate_scalar_merge](reg/data_fill/asc_duplicate_scalar_merge.md) | 根据mask将value填充到目的操作数dst，dst中未被mask筛选的元素保留原值。 |541| [asc_duplicate_scalar_merge](reg/data_fill/asc_duplicate_scalar_merge.md) | 根据mask将value填充到目的操作数dst，dst中未被mask筛选的元素保留原值。 |
Adocs/zh/api/SIMD-API/c_api/reg/data_fill/asc_duplicate_highest.md+84-0
@@ -0,0 +1,84 @@
1+# asc_duplicate_highest
M
Mmunanhw15 天前

新增文件要同步修改api目录下的README.md文件和C API列表

likedislike
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：不支持
10+<!-- end id2 -->
11+<!-- npu="910b" id3 -->
12+- Atlas A2 训练系列产品/Atlas A2 推理系列产品：不支持
13+<!-- end id3 -->
14+<!-- npu="910" id4 -->
15+- Atlas 训练系列产品：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="310b" id7 -->
24+- Atlas 200I/500 A2 推理产品：不支持
25+<!-- end id7 -->
26+ 
27+## 功能说明
28+ 
29+将src的最高位元素广播到dst中被mask筛选的位置，dst中未被mask筛选的位置被置为0。
30+ 
31+## 函数原型
32+ 
33+```cpp
34+__simd_callee__ inline void asc_duplicate_highest(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
35+__simd_callee__ inline void asc_duplicate_highest(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
36+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
37+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
38+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
39+__simd_callee__ inline void asc_duplicate_highest(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
40+__simd_callee__ inline void asc_duplicate_highest(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
41+__simd_callee__ inline void asc_duplicate_highest(vector_half& dst, vector_half src, vector_bool mask)
42+__simd_callee__ inline void asc_duplicate_highest(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
43+__simd_callee__ inline void asc_duplicate_highest(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
44+__simd_callee__ inline void asc_duplicate_highest(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
45+__simd_callee__ inline void asc_duplicate_highest(vector_float& dst, vector_float src, vector_bool mask)
46+```
47+ 
48+## 参数说明
49+ 
50+**表1** 参数说明
51+ 
52+| 参数名 | 输入/输出 | 描述 |
53+| --- | --- | --- |
54+| dst | 输出 | 目的操作数（矢量数据寄存器）。 |
55+| src | 输入 | 源操作数（矢量数据寄存器）。 |
56+| mask | 输入 | 源操作数掩码（掩码寄存器），用于指示在计算过程中哪些元素参与计算。对应位置为1时参与计算，为0时不参与计算。mask未筛选的元素在输出中置零。 |
57+ 
58+矢量数据寄存器和掩码寄存器的详细说明请参见[reg数据类型定义](../reg_data_types/data_type_definition.md)。
59+ 
60+## 返回值说明
61+ 
62+无
63+ 
64+## 约束说明
65+ 
66+无
67+ 
68+## 调用示例
69+ 
70+```cpp
71+__simd_vf__ inline void duplicate_highest_vf(__ubuf__ half* dst_addr, __ubuf__ half* src_addr,
72+ uint32_t count, uint16_t one_repeat_size, uint16_t repeat_time)
73+{
74+ vector_half dst;
75+ vector_half src;
76+ vector_bool mask;
77+ for (uint16_t i = 0; i < repeat_time; ++i) {
78+ mask = asc_update_mask_b16(count);
79+ asc_loadalign(src, src_addr + i * one_repeat_size);
80+ asc_duplicate_highest(dst, src, mask);
81+ asc_storealign(dst_addr + i * one_repeat_size, dst, mask);
82+ }
83+}
84+```
Adocs/zh/api/SIMD-API/c_api/reg/data_fill/asc_duplicate_highest_merge.md+85-0
@@ -0,0 +1,85 @@
1+# asc_duplicate_highest_merge
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：不支持
10+<!-- end id2 -->
11+<!-- npu="910b" id3 -->
12+- Atlas A2 训练系列产品/Atlas A2 推理系列产品：不支持
13+<!-- end id3 -->
14+<!-- npu="910" id4 -->
15+- Atlas 训练系列产品：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="310b" id7 -->
24+- Atlas 200I/500 A2 推理产品：不支持
25+<!-- end id7 -->
26+ 
27+## 功能说明
28+ 
29+merge模式下，将src的最高位元素广播到dst中被mask筛选的位置，dst中未被mask筛选的元素保留原值。
30+ 
31+## 函数原型
32+ 
33+```cpp
34+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
35+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
36+__simd_callee__ inline void asc_duplicate_highest_merge(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
37+__simd_callee__ inline void asc_duplicate_highest_merge(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
38+__simd_callee__ inline void asc_duplicate_highest_merge(vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
39+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
40+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
41+__simd_callee__ inline void asc_duplicate_highest_merge(vector_half& dst, vector_half src, vector_bool mask)
42+__simd_callee__ inline void asc_duplicate_highest_merge(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
43+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
44+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
45+__simd_callee__ inline void asc_duplicate_highest_merge(vector_float& dst, vector_float src, vector_bool mask)
46+```
47+ 
48+## 参数说明
49+ 
50+**表1** 参数说明
51+ 
52+| 参数名 | 输入/输出 | 描述 |
53+| --- | --- | --- |
54+| dst | 输入/输出 | 目的操作数（矢量数据寄存器）。未被mask筛选的元素保留原值。 |
55+| src | 输入 | 源操作数（矢量数据寄存器）。 |
56+| mask | 输入 | 源操作数掩码（掩码寄存器），用于指示在计算过程中哪些元素参与计算。对应位置为1时参与计算，为0时不参与计算。 |
57+ 
58+矢量数据寄存器和掩码寄存器的详细说明请参见[reg数据类型定义](../reg_data_types/data_type_definition.md)。
59+ 
60+## 返回值说明
61+ 
62+无
63+ 
64+## 约束说明
65+ 
66+无
67+ 
68+## 调用示例
69+ 
70+```cpp
71+__simd_vf__ inline void duplicate_highest_merge_vf(__ubuf__ half* dst_addr, __ubuf__ half* src_addr,
72+ uint32_t count, uint16_t one_repeat_size, uint16_t repeat_time)
73+{
74+ vector_half dst;
75+ vector_half src;
76+ vector_bool mask;
77+ for (uint16_t i = 0; i < repeat_time; ++i) {
78+ mask = asc_update_mask_b16(count);
79+ asc_loadalign(dst, dst_addr + i * one_repeat_size);
80+ asc_loadalign(src, src_addr + i * one_repeat_size);
81+ asc_duplicate_highest_merge(dst, src, mask);
82+ asc_storealign(dst_addr + i * one_repeat_size, dst, mask);
83+ }
84+}
85+```
Mdocs/zh/api/SIMD-API/c_api/reg/data_fill/data_fill.md+4-0
@@ -2,6 +2,10 @@
2 2 
3- **[asc_duplicate](asc_duplicate.md)**3- **[asc_duplicate](asc_duplicate.md)**
4 4 
5+- **[asc_duplicate_highest](asc_duplicate_highest.md)**
6+ 
7+- **[asc_duplicate_highest_merge](asc_duplicate_highest_merge.md)**
8+ 
5- **[asc_duplicate_merge](asc_duplicate_merge.md)**9- **[asc_duplicate_merge](asc_duplicate_merge.md)**
6 10 
7- **[asc_duplicate_scalar](asc_duplicate_scalar.md)**11- **[asc_duplicate_scalar](asc_duplicate_scalar.md)**
Adocs/zh/api/SIMD-API/c_api/scalar_compute/asc_load_dev.md+77-0
@@ -0,0 +1,77 @@
1+# asc_load_dev
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：不支持
10+<!-- end id2 -->
11+<!-- npu="910b" id3 -->
12+- Atlas A2 训练系列产品/Atlas A2 推理系列产品：不支持
13+<!-- end id3 -->
14+<!-- npu="910" id4 -->
15+- Atlas 训练系列产品：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="310b" id7 -->
24+- Atlas 200I/500 A2 推理产品：不支持
25+<!-- end id7 -->
26+ 
27+## 功能说明
28+ 
29+不经过DCache直接从GM地址读取整型数据。
30+当多核操作GM地址且数据无法对齐到Cache Line时，经过DCache读写可能引入Cache Line粒度的数据覆盖。此时，可使用本接口绕过DCache读取GM数据。
31+ 
32+## 函数原型
33+ 
34+```cpp
35+__aicore__ inline int8_t asc_load_dev(__gm__ int8_t* addr)
36+ 
37+__aicore__ inline uint8_t asc_load_dev(__gm__ uint8_t* addr)
38+ 
39+__aicore__ inline int16_t asc_load_dev(__gm__ int16_t* addr)
40+ 
41+__aicore__ inline uint16_t asc_load_dev(__gm__ uint16_t* addr)
42+ 
43+__aicore__ inline int32_t asc_load_dev(__gm__ int32_t* addr)
44+ 
45+__aicore__ inline uint32_t asc_load_dev(__gm__ uint32_t* addr)
46+ 
47+__aicore__ inline int64_t asc_load_dev(__gm__ int64_t* addr)
48+ 
49+__aicore__ inline uint64_t asc_load_dev(__gm__ uint64_t* addr)
50+```
51+ 
52+## 参数说明
53+ 
54+**表1** 参数说明
55+ 
56+| 参数名 | 输入/输出 | 描述 |
57+| --- | --- | --- |
58+| addr | 输入 | 源GM地址。支持的数据类型为`int8_t`、`uint8_t`、`int16_t`、`uint16_t`、`int32_t`、`uint32_t`、`int64_t`、`uint64_t`。 |
59+ 
60+## 返回值说明
61+ 
62+从GM读取的数据，返回值的数据类型与`addr`指向的数据类型一致。
63+ 
64+## 流水类型
65+ 
66+PIPE_S
67+ 
68+## 约束说明
69+ 
70+仅支持整型数据，不支持浮点类型。
71+ 
72+## 调用示例
73+ 
74+```cpp
75+// addr是外部输入的GM地址，类型为__gm__ int32_t*。
76+int32_t value = asc_load_dev(addr);
77+```
Mdocs/zh/api/SIMD-API/c_api/scalar_compute/scalar_compute.md+2-0
@@ -10,6 +10,8 @@
10 10 
11- **[asc_float2int32](asc_float2int32.md)**11- **[asc_float2int32](asc_float2int32.md)**
12 12 
13+- **[asc_load_dev](asc_load_dev.md)**
14+ 
13- **[asc_popc](asc_popc.md)**15- **[asc_popc](asc_popc.md)**
14 16 
15- **[asc_set_nthbit](asc_set_nthbit.md)**17- **[asc_set_nthbit](asc_set_nthbit.md)**
Mimpl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl.h+17-0
@@ -23,6 +23,7 @@
23#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_zero_bits_cnt_impl.h"23#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_zero_bits_cnt_impl.h"
24#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_ffs_impl.h"24#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_ffs_impl.h"
25#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_store_dev_impl.h"25#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_store_dev_impl.h"
26+#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_load_dev_impl.h"
26#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_float2int32_impl.h"27#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_float2int32_impl.h"
27#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_sflbits_impl.h"28#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_sflbits_impl.h"
28#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_ffz_impl.h"29#include "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_ffz_impl.h"
@@ -65,6 +66,22 @@ __aicore__ inline void asc_store_dev(__gm__ int64_t* addr, int64_t value) { asc_
65 66 
66__aicore__ inline void asc_store_dev(__gm__ uint64_t* addr, uint64_t value) { asc_store_dev_impl(addr, value); }67__aicore__ inline void asc_store_dev(__gm__ uint64_t* addr, uint64_t value) { asc_store_dev_impl(addr, value); }
67 68 
69+__aicore__ inline int8_t asc_load_dev(__gm__ int8_t* addr) { return asc_load_dev_impl(addr); }
70+ 
71+__aicore__ inline uint8_t asc_load_dev(__gm__ uint8_t* addr) { return asc_load_dev_impl(addr); }
72+ 
73+__aicore__ inline int16_t asc_load_dev(__gm__ int16_t* addr) { return asc_load_dev_impl(addr); }
74+ 
75+__aicore__ inline uint16_t asc_load_dev(__gm__ uint16_t* addr) { return asc_load_dev_impl(addr); }
76+ 
77+__aicore__ inline int32_t asc_load_dev(__gm__ int32_t* addr) { return asc_load_dev_impl(addr); }
78+ 
79+__aicore__ inline uint32_t asc_load_dev(__gm__ uint32_t* addr) { return asc_load_dev_impl(addr); }
80+ 
81+__aicore__ inline int64_t asc_load_dev(__gm__ int64_t* addr) { return asc_load_dev_impl(addr); }
82+ 
83+__aicore__ inline uint64_t asc_load_dev(__gm__ uint64_t* addr) { return asc_load_dev_impl(addr); }
84+ 
68// ==========asc_float2int32 (rd/ru/rn/rna)==========85// ==========asc_float2int32 (rd/ru/rn/rna)==========
69__aicore__ inline int32_t asc_float2int32_rd(float value) { return asc_float2int32_rd_impl(value); }86__aicore__ inline int32_t asc_float2int32_rd(float value) { return asc_float2int32_rd_impl(value); }
70 87 
Aimpl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_load_dev_impl.h+69-0
@@ -0,0 +1,69 @@
1+/**
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
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.
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.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+#if !defined(ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS)
12+#warning \
13+ "impl/c_api/instr_impl/npu_arch_3510/scalar_compute_impl/asc_load_dev_impl.h is an internal header file and must not be used directly. Functions or variables defined in this file may be removed in the future. Please use "#include "c_api/asc_simd.h"" and use public functions or variables defined in interface header files."
14+#define ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS
15+#define UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
16+#endif
17+ 
18+#ifndef IMPL_CAPI_INSTR_IMPL_NPU_ARCH_3510_SCALAR_COMPUTE_IMPL_ASC_LOAD_DEV_IMPL_H
19+#define IMPL_CAPI_INSTR_IMPL_NPU_ARCH_3510_SCALAR_COMPUTE_IMPL_ASC_LOAD_DEV_IMPL_H
20+ 
21+#include "impl/c_api/instr_impl/npu_arch_3510/utils_impl.h"
22+ 
23+constexpr int16_t ASC_C_API_LOAD_DEV_DEFAULT_OFFSET = 0;
24+ 
25+__aicore__ inline int8_t asc_load_dev_impl(__gm__ int8_t* addr)
26+{
27+ return ld_dev(reinterpret_cast<__gm__ uint8_t*>(addr), ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
28+}
29+ 
30+__aicore__ inline uint8_t asc_load_dev_impl(__gm__ uint8_t* addr)
31+{
32+ return ld_dev(addr, ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
33+}
34+ 
35+__aicore__ inline int16_t asc_load_dev_impl(__gm__ int16_t* addr)
36+{
37+ return ld_dev(reinterpret_cast<__gm__ uint16_t*>(addr), ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
38+}
39+ 
40+__aicore__ inline uint16_t asc_load_dev_impl(__gm__ uint16_t* addr)
41+{
42+ return ld_dev(addr, ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
43+}
44+ 
45+__aicore__ inline int32_t asc_load_dev_impl(__gm__ int32_t* addr)
46+{
47+ return ld_dev(reinterpret_cast<__gm__ uint32_t*>(addr), ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
48+}
49+ 
50+__aicore__ inline uint32_t asc_load_dev_impl(__gm__ uint32_t* addr)
51+{
52+ return ld_dev(addr, ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
53+}
54+ 
55+__aicore__ inline int64_t asc_load_dev_impl(__gm__ int64_t* addr)
56+{
57+ return ld_dev(addr, ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
58+}
59+ 
60+__aicore__ inline uint64_t asc_load_dev_impl(__gm__ uint64_t* addr)
61+{
62+ return ld_dev(addr, ASC_C_API_LOAD_DEV_DEFAULT_OFFSET);
63+}
64+#endif
65+ 
66+#if defined(UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC)
67+#undef ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS
68+#undef UNDEF_ASCENDC_C_API_INCLUDE_COMPILER_INTERNAL_HEADERS_ASCENDC
69+#endif
Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl.h+123-0
@@ -1584,6 +1584,67 @@ __simd_callee__ inline void asc_duplicate(vector_float& dst, vector_float src, v
1584 asc_duplicate_impl(dst, src, mask);1584 asc_duplicate_impl(dst, src, mask);
1585}1585}
1586 1586 
1587+// ==========asc_duplicate_highest(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
1588+__simd_callee__ inline void asc_duplicate_highest(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
1589+{
1590+ asc_duplicate_highest_impl(dst, src, mask);
1591+}
1592+ 
1593+__simd_callee__ inline void asc_duplicate_highest(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
1594+{
1595+ asc_duplicate_highest_impl(dst, src, mask);
1596+}
1597+ 
1598+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
1599+{
1600+ asc_duplicate_highest_impl(dst, src, mask);
1601+}
1602+ 
1603+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
1604+{
1605+ asc_duplicate_highest_impl(dst, src, mask);
1606+}
1607+ 
1608+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
1609+{
1610+ asc_duplicate_highest_impl(dst, src, mask);
1611+}
1612+ 
1613+__simd_callee__ inline void asc_duplicate_highest(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
1614+{
1615+ asc_duplicate_highest_impl(dst, src, mask);
1616+}
1617+ 
1618+__simd_callee__ inline void asc_duplicate_highest(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
1619+{
1620+ asc_duplicate_highest_impl(dst, src, mask);
1621+}
1622+ 
1623+__simd_callee__ inline void asc_duplicate_highest(vector_half& dst, vector_half src, vector_bool mask)
1624+{
1625+ asc_duplicate_highest_impl(dst, src, mask);
1626+}
1627+ 
1628+__simd_callee__ inline void asc_duplicate_highest(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
1629+{
1630+ asc_duplicate_highest_impl(dst, src, mask);
1631+}
1632+ 
1633+__simd_callee__ inline void asc_duplicate_highest(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
1634+{
1635+ asc_duplicate_highest_impl(dst, src, mask);
1636+}
1637+ 
1638+__simd_callee__ inline void asc_duplicate_highest(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
1639+{
1640+ asc_duplicate_highest_impl(dst, src, mask);
1641+}
1642+ 
1643+__simd_callee__ inline void asc_duplicate_highest(vector_float& dst, vector_float src, vector_bool mask)
1644+{
1645+ asc_duplicate_highest_impl(dst, src, mask);
1646+}
1647+ 
1587// ==========asc_duplicate_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========1648// ==========asc_duplicate_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
1588__simd_callee__ inline void asc_duplicate_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)1649__simd_callee__ inline void asc_duplicate_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
1589{1650{
@@ -1645,6 +1706,68 @@ __simd_callee__ inline void asc_duplicate_merge(vector_float& dst, vector_float
1645 asc_duplicate_merge_impl(dst, src, mask);1706 asc_duplicate_merge_impl(dst, src, mask);
1646}1707}
1647 1708 
1709+// ==========asc_duplicate_highest_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
1710+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
1711+{
1712+ asc_duplicate_highest_merge_impl(dst, src, mask);
1713+}
1714+ 
1715+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
1716+{
1717+ asc_duplicate_highest_merge_impl(dst, src, mask);
1718+}
1719+ 
1720+__simd_callee__ inline void asc_duplicate_highest_merge(
1721+ vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
1722+{
1723+ asc_duplicate_highest_merge_impl(dst, src, mask);
1724+}
1725+ 
1726+__simd_callee__ inline void asc_duplicate_highest_merge(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
1727+{
1728+ asc_duplicate_highest_merge_impl(dst, src, mask);
1729+}
1730+ 
1731+__simd_callee__ inline void asc_duplicate_highest_merge(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
1732+{
1733+ asc_duplicate_highest_merge_impl(dst, src, mask);
1734+}
1735+ 
1736+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
1737+{
1738+ asc_duplicate_highest_merge_impl(dst, src, mask);
1739+}
1740+ 
1741+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
1742+{
1743+ asc_duplicate_highest_merge_impl(dst, src, mask);
1744+}
1745+ 
1746+__simd_callee__ inline void asc_duplicate_highest_merge(vector_half& dst, vector_half src, vector_bool mask)
1747+{
1748+ asc_duplicate_highest_merge_impl(dst, src, mask);
1749+}
1750+ 
1751+__simd_callee__ inline void asc_duplicate_highest_merge(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
1752+{
1753+ asc_duplicate_highest_merge_impl(dst, src, mask);
1754+}
1755+ 
1756+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
1757+{
1758+ asc_duplicate_highest_merge_impl(dst, src, mask);
1759+}
1760+ 
1761+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
1762+{
1763+ asc_duplicate_highest_merge_impl(dst, src, mask);
1764+}
1765+ 
1766+__simd_callee__ inline void asc_duplicate_highest_merge(vector_float& dst, vector_float src, vector_bool mask)
1767+{
1768+ asc_duplicate_highest_merge_impl(dst, src, mask);
1769+}
1770+ 
1648// ==========asc_abs_sub(half/float)==========1771// ==========asc_abs_sub(half/float)==========
1649__simd_callee__ inline void asc_abs_sub(vector_half& dst, vector_half src0, vector_half src1, vector_bool mask)1772__simd_callee__ inline void asc_abs_sub(vector_half& dst, vector_half src0, vector_half src1, vector_bool mask)
1650{1773{
Mimpl/c_api/instr_impl/npu_arch_3510/vector_compute_impl/asc_duplicate_impl.h+175-0
@@ -102,6 +102,91 @@ __simd_callee__ inline void asc_duplicate_impl(vector_float& dst, vector_float s
102 }102 }
103}103}
104 104 
105+__simd_callee__ inline void asc_duplicate_highest_impl(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
106+{
107+ if ASC_IS_AIV {
108+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
109+ }
110+}
111+ 
112+__simd_callee__ inline void asc_duplicate_highest_impl(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
113+{
114+ if ASC_IS_AIV {
115+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
116+ }
117+}
118+ 
119+__simd_callee__ inline void asc_duplicate_highest_impl(
120+ vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
121+{
122+ if ASC_IS_AIV {
123+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
124+ }
125+}
126+ 
127+__simd_callee__ inline void asc_duplicate_highest_impl(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
128+{
129+ if ASC_IS_AIV {
130+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
131+ }
132+}
133+ 
134+__simd_callee__ inline void asc_duplicate_highest_impl(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
135+{
136+ if ASC_IS_AIV {
137+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
138+ }
139+}
140+ 
141+__simd_callee__ inline void asc_duplicate_highest_impl(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
142+{
143+ if ASC_IS_AIV {
144+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
145+ }
146+}
147+ 
148+__simd_callee__ inline void asc_duplicate_highest_impl(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
149+{
150+ if ASC_IS_AIV {
151+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
152+ }
153+}
154+ 
155+__simd_callee__ inline void asc_duplicate_highest_impl(vector_half& dst, vector_half src, vector_bool mask)
156+{
157+ if ASC_IS_AIV {
158+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
159+ }
160+}
161+ 
162+__simd_callee__ inline void asc_duplicate_highest_impl(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
163+{
164+ if ASC_IS_AIV {
165+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
166+ }
167+}
168+ 
169+__simd_callee__ inline void asc_duplicate_highest_impl(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
170+{
171+ if ASC_IS_AIV {
172+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
173+ }
174+}
175+ 
176+__simd_callee__ inline void asc_duplicate_highest_impl(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
177+{
178+ if ASC_IS_AIV {
179+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
180+ }
181+}
182+ 
183+__simd_callee__ inline void asc_duplicate_highest_impl(vector_float& dst, vector_float src, vector_bool mask)
184+{
185+ if ASC_IS_AIV {
186+ vdup(dst, src, mask, POS_HIGHEST, MODE_ZEROING);
187+ }
188+}
189+ 
105// vdup with merging mode190// vdup with merging mode
106__simd_callee__ inline void asc_duplicate_merge_impl(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)191__simd_callee__ inline void asc_duplicate_merge_impl(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
107{192{
@@ -187,4 +272,94 @@ __simd_callee__ inline void asc_duplicate_merge_impl(vector_float& dst, vector_f
187 vdup(dst, src, mask, POS_LOWEST, MODE_MERGING);272 vdup(dst, src, mask, POS_LOWEST, MODE_MERGING);
188 }273 }
189}274}
275+ 
276+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask)
277+{
278+ if ASC_IS_AIV {
279+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
280+ }
281+}
282+ 
283+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_int8_t& dst, vector_int8_t src, vector_bool mask)
284+{
285+ if ASC_IS_AIV {
286+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
287+ }
288+}
289+ 
290+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
291+ vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask)
292+{
293+ if ASC_IS_AIV {
294+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
295+ }
296+}
297+ 
298+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
299+ vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask)
300+{
301+ if ASC_IS_AIV {
302+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
303+ }
304+}
305+ 
306+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
307+ vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask)
308+{
309+ if ASC_IS_AIV {
310+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
311+ }
312+}
313+ 
314+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
315+ vector_uint16_t& dst, vector_uint16_t src, vector_bool mask)
316+{
317+ if ASC_IS_AIV {
318+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
319+ }
320+}
321+ 
322+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_int16_t& dst, vector_int16_t src, vector_bool mask)
323+{
324+ if ASC_IS_AIV {
325+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
326+ }
327+}
328+ 
329+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_half& dst, vector_half src, vector_bool mask)
330+{
331+ if ASC_IS_AIV {
332+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
333+ }
334+}
335+ 
336+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
337+ vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask)
338+{
339+ if ASC_IS_AIV {
340+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
341+ }
342+}
343+ 
344+__simd_callee__ inline void asc_duplicate_highest_merge_impl(
345+ vector_uint32_t& dst, vector_uint32_t src, vector_bool mask)
346+{
347+ if ASC_IS_AIV {
348+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
349+ }
350+}
351+ 
352+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_int32_t& dst, vector_int32_t src, vector_bool mask)
353+{
354+ if ASC_IS_AIV {
355+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
356+ }
357+}
358+ 
359+__simd_callee__ inline void asc_duplicate_highest_merge_impl(vector_float& dst, vector_float src, vector_bool mask)
360+{
361+ if ASC_IS_AIV {
362+ vdup(dst, src, mask, POS_HIGHEST, MODE_MERGING);
363+ }
364+}
190#endif365#endif
Minclude/c_api/reg_compute/reg_vector.h+54-0
@@ -389,6 +389,31 @@ __simd_callee__ inline void asc_duplicate(vector_int32_t& dst, vector_int32_t sr
389 389 
390__simd_callee__ inline void asc_duplicate(vector_float& dst, vector_float src, vector_bool mask);390__simd_callee__ inline void asc_duplicate(vector_float& dst, vector_float src, vector_bool mask);
391 391 
392+// ==========asc_duplicate_highest(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
393+__simd_callee__ inline void asc_duplicate_highest(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);
W
Wwuzhaolin14 天前

注释啥的尽量添加的详细一些，能够解释当前接口是干啥的

likedislike
guojianyang
guojianyang
14 天前 评论：
394+ 
395+__simd_callee__ inline void asc_duplicate_highest(vector_int8_t& dst, vector_int8_t src, vector_bool mask);
396+ 
397+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask);
398+ 
399+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask);
400+ 
401+__simd_callee__ inline void asc_duplicate_highest(vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask);
402+ 
403+__simd_callee__ inline void asc_duplicate_highest(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask);
404+ 
405+__simd_callee__ inline void asc_duplicate_highest(vector_int16_t& dst, vector_int16_t src, vector_bool mask);
406+ 
407+__simd_callee__ inline void asc_duplicate_highest(vector_half& dst, vector_half src, vector_bool mask);
408+ 
409+__simd_callee__ inline void asc_duplicate_highest(vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask);
410+ 
411+__simd_callee__ inline void asc_duplicate_highest(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask);
412+ 
413+__simd_callee__ inline void asc_duplicate_highest(vector_int32_t& dst, vector_int32_t src, vector_bool mask);
414+ 
415+__simd_callee__ inline void asc_duplicate_highest(vector_float& dst, vector_float src, vector_bool mask);
416+ 
392// ==========asc_duplicate_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========417// ==========asc_duplicate_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
393__simd_callee__ inline void asc_duplicate_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);418__simd_callee__ inline void asc_duplicate_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);
394 419 
@@ -414,6 +439,35 @@ __simd_callee__ inline void asc_duplicate_merge(vector_int32_t& dst, vector_int3
414 439 
415__simd_callee__ inline void asc_duplicate_merge(vector_float& dst, vector_float src, vector_bool mask);440__simd_callee__ inline void asc_duplicate_merge(vector_float& dst, vector_float src, vector_bool mask);
416 441 
442+// ==========asc_duplicate_highest_merge(uint8_t/int8_t/fp8_e4m3fn_t/fp8_e5m2_t/fp8_e8m0_t/uint16_t/int16_t/half/bfloat16_t/uint32_t/int32_t/float)==========
443+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint8_t& dst, vector_uint8_t src, vector_bool mask);
444+ 
445+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int8_t& dst, vector_int8_t src, vector_bool mask);
446+ 
447+__simd_callee__ inline void asc_duplicate_highest_merge(
448+ vector_fp8_e4m3fn_t& dst, vector_fp8_e4m3fn_t src, vector_bool mask);
449+ 
450+__simd_callee__ inline void asc_duplicate_highest_merge(
451+ vector_fp8_e5m2_t& dst, vector_fp8_e5m2_t src, vector_bool mask);
452+ 
453+__simd_callee__ inline void asc_duplicate_highest_merge(
454+ vector_fp8_e8m0_t& dst, vector_fp8_e8m0_t src, vector_bool mask);
455+ 
456+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint16_t& dst, vector_uint16_t src, vector_bool mask);
457+ 
458+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int16_t& dst, vector_int16_t src, vector_bool mask);
459+ 
460+__simd_callee__ inline void asc_duplicate_highest_merge(vector_half& dst, vector_half src, vector_bool mask);
461+ 
462+__simd_callee__ inline void asc_duplicate_highest_merge(
463+ vector_bfloat16_t& dst, vector_bfloat16_t src, vector_bool mask);
464+ 
465+__simd_callee__ inline void asc_duplicate_highest_merge(vector_uint32_t& dst, vector_uint32_t src, vector_bool mask);
466+ 
467+__simd_callee__ inline void asc_duplicate_highest_merge(vector_int32_t& dst, vector_int32_t src, vector_bool mask);
468+ 
469+__simd_callee__ inline void asc_duplicate_highest_merge(vector_float& dst, vector_float src, vector_bool mask);
470+ 
417// ==========asc_abs_sub(half/float)==========471// ==========asc_abs_sub(half/float)==========
418__simd_callee__ inline void asc_abs_sub(vector_half& dst, vector_half src0, vector_half src1, vector_bool mask);472__simd_callee__ inline void asc_abs_sub(vector_half& dst, vector_half src0, vector_half src1, vector_bool mask);
419 473 
Minclude/c_api/scalar_compute/scalar_compute.h+8-0
@@ -51,6 +51,14 @@ __aicore__ inline void asc_store_dev(__gm__ int32_t* addr, int32_t value);
51__aicore__ inline void asc_store_dev(__gm__ uint32_t* addr, uint32_t value);51__aicore__ inline void asc_store_dev(__gm__ uint32_t* addr, uint32_t value);
52__aicore__ inline void asc_store_dev(__gm__ int64_t* addr, int64_t value);52__aicore__ inline void asc_store_dev(__gm__ int64_t* addr, int64_t value);
53__aicore__ inline void asc_store_dev(__gm__ uint64_t* addr, uint64_t value);53__aicore__ inline void asc_store_dev(__gm__ uint64_t* addr, uint64_t value);
54+__aicore__ inline int8_t asc_load_dev(__gm__ int8_t* addr);
55+__aicore__ inline uint8_t asc_load_dev(__gm__ uint8_t* addr);
56+__aicore__ inline int16_t asc_load_dev(__gm__ int16_t* addr);
57+__aicore__ inline uint16_t asc_load_dev(__gm__ uint16_t* addr);
58+__aicore__ inline int32_t asc_load_dev(__gm__ int32_t* addr);
59+__aicore__ inline uint32_t asc_load_dev(__gm__ uint32_t* addr);
60+__aicore__ inline int64_t asc_load_dev(__gm__ int64_t* addr);
61+__aicore__ inline uint64_t asc_load_dev(__gm__ uint64_t* addr);
H
Hhorming15 天前

返回值 跟地址数据类型保持一致把，跟store对应

likedislike
54namespace __asc_aicore {62namespace __asc_aicore {
55__aicore__ inline int32_t asc_atomic_add(__gm__ int32_t* address, int32_t val);63__aicore__ inline int32_t asc_atomic_add(__gm__ int32_t* address, int32_t val);
56__aicore__ inline uint32_t asc_atomic_add(__gm__ uint32_t* address, uint32_t val);64__aicore__ inline uint32_t asc_atomic_add(__gm__ uint32_t* address, uint32_t val);
Mtests/api/c_api/ascendc_host_header_checker/host_header_reg_compute_checker.cpp+2-0
@@ -122,6 +122,8 @@ static void test_host_c_api_reg_compute_3()
122 using ::asc_deintlv_b8;122 using ::asc_deintlv_b8;
123 using ::asc_div;123 using ::asc_div;
124 using ::asc_duplicate;124 using ::asc_duplicate;
125+ using ::asc_duplicate_highest;
126+ using ::asc_duplicate_highest_merge;
125 using ::asc_duplicate_merge;127 using ::asc_duplicate_merge;
126 using ::asc_duplicate_scalar;128 using ::asc_duplicate_scalar;
127 using ::asc_duplicate_scalar_merge;129 using ::asc_duplicate_scalar_merge;
Mtests/api/c_api/ascendc_host_header_checker/host_header_scalar_compute_checker.cpp+1-0
@@ -20,6 +20,7 @@ static void test_host_c_api_scalar_compute_0()
20 using ::asc_float2int32_rn;20 using ::asc_float2int32_rn;
21 using ::asc_float2int32_rna;21 using ::asc_float2int32_rna;
22 using ::asc_float2int32_ru;22 using ::asc_float2int32_ru;
23+ using ::asc_load_dev;
23 using ::asc_popc;24 using ::asc_popc;
24 using ::asc_set_nthbit;25 using ::asc_set_nthbit;
25 using ::asc_sflbits;26 using ::asc_sflbits;
Atests/api/c_api/npu_arch_3510/scalar_compute/test_asc_load_dev.cpp+62-0
@@ -0,0 +1,62 @@
1+/**
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
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.
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.
8+ * See LICENSE in the root of the software repository for the full text of the License.
9+ */
10+ 
11+#include <gtest/gtest.h>
12+#include <mockcpp/mockcpp.hpp>
13+#include <type_traits>
14+#include "tests/api/c_api/stub/cce_stub.h"
15+#include "include/c_api/asc_simd.h"
16+ 
17+int64_t ld_dev(__gm__ int64_t*, int16_t) { return 0; }
18+ 
19+#define TEST_ASC_LOAD_DEV(data_type, cce_type, raw_value, expected_value) \
20+ namespace { \
21+ uint64_t ld_dev_##data_type##_stub(__gm__ cce_type* addr, int16_t offset) \
22+ { \
23+ EXPECT_EQ(offset, 0); \
24+ return static_cast<uint64_t>(raw_value); \
25+ } \
26+ } \
27+ \
28+ TEST(AscLoadDev, data_type) \
29+ { \
30+ __gm__ data_type* addr = nullptr; \
31+ static_assert(std::is_same_v<decltype(asc_load_dev(addr)), data_type>); \
32+ MOCKER_CPP(ld_dev, uint64_t(__gm__ cce_type*, int16_t)).times(1).will(invoke(ld_dev_##data_type##_stub)); \
33+ EXPECT_EQ(asc_load_dev(addr), static_cast<data_type>(expected_value)); \
34+ GlobalMockObject::verify(); \
35+ }
36+ 
37+TEST_ASC_LOAD_DEV(int8_t, uint8_t, UINT8_MAX, -1)
38+TEST_ASC_LOAD_DEV(uint8_t, uint8_t, 1, 1)
39+TEST_ASC_LOAD_DEV(int16_t, uint16_t, UINT16_MAX, -1)
40+TEST_ASC_LOAD_DEV(uint16_t, uint16_t, 2, 2)
41+TEST_ASC_LOAD_DEV(int32_t, uint32_t, UINT32_MAX, -1)
42+TEST_ASC_LOAD_DEV(uint32_t, uint32_t, 3, 3)
43+TEST_ASC_LOAD_DEV(uint64_t, uint64_t, 4, 4)
44+ 
45+namespace {
46+int64_t ld_dev_int64_t_stub(__gm__ int64_t* addr, int16_t offset)
47+{
48+ EXPECT_EQ(offset, 0);
49+ return -1;
50+}
51+} // namespace
52+ 
53+TEST(AscLoadDev, int64_t)
54+{
55+ __gm__ int64_t* addr = nullptr;
56+ static_assert(std::is_same_v<decltype(asc_load_dev(addr)), int64_t>);
57+ MOCKER_CPP(ld_dev, int64_t(__gm__ int64_t*, int16_t)).times(1).will(invoke(ld_dev_int64_t_stub));
58+ EXPECT_EQ(asc_load_dev(addr), -1);
59+ GlobalMockObject::verify();
60+}
61+ 
62+#undef TEST_ASC_LOAD_DEV
Mtests/api/c_api/npu_arch_3510/vector_compute/test_asc_duplicate.cpp+170-97
@@ -1,97 +1,170 @@
1-/**1+/**
2- * Copyright (c) 2025 Huawei Technologies Co., Ltd.2+ * Copyright (c) 2025 Huawei Technologies Co., Ltd.
3- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of3+ * 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-#include <gtest/gtest.h>11+#include <gtest/gtest.h>
12-#include <mockcpp/mockcpp.hpp>12+#include <mockcpp/mockcpp.hpp>
13-#include "tests/api/c_api/stub/cce_stub.h"13+#include "tests/api/c_api/stub/cce_stub.h"
14-#include "include/c_api/asc_simd.h"14+#include "include/c_api/asc_simd.h"
15- 15+ 
16-#define TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(class_name, c_api_name, cce_name, data_type) \16+#define TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(class_name, c_api_name, cce_name, data_type) \
17- \17+ \
18- class TestVectorCompute##class_name####data_type##CApi : public testing::Test { \18+ class TestVectorCompute##class_name####data_type##CApi : public testing::Test { \
19- protected: \19+ protected: \
20- void SetUp() {} \20+ void SetUp() {} \
21- void TearDown() {} \21+ void TearDown() {} \
22- }; \22+ }; \
23- \23+ \
24- namespace { \24+ namespace { \
25- void cce_name##_##data_type##_Stub(data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) {} \25+ void cce_name##_##data_type##_Stub(data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) {} \
26- } \26+ } \
27- \27+ \
28- TEST_F(TestVectorCompute##class_name####data_type##CApi, c_api_name##_##data_type##_Succ) \28+ TEST_F(TestVectorCompute##class_name####data_type##CApi, c_api_name##_##data_type##_Succ) \
29- { \29+ { \
30- data_type dst; \30+ data_type dst; \
31- data_type src0; \31+ data_type src0; \
32- vector_bool mask; \32+ vector_bool mask; \
33- \33+ \
34- MOCKER_CPP(cce_name, void(data_type&, data_type, vector_bool, int32_t, Literal)) \34+ MOCKER_CPP(cce_name, void(data_type&, data_type, vector_bool, int32_t, Literal)) \
35- .times(1) \35+ .times(1) \
36- .will(invoke(cce_name##_##data_type##_Stub)); \36+ .will(invoke(cce_name##_##data_type##_Stub)); \
37- \37+ \
38- c_api_name(dst, src0, mask); \38+ c_api_name(dst, src0, mask); \
39- GlobalMockObject::verify(); \39+ GlobalMockObject::verify(); \
40- }40+ }
41- 41+ 
42-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint8_t);42+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint8_t);
43-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int8_t);43+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int8_t);
44-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e4m3fn_t);44+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e4m3fn_t);
45-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e5m2_t);45+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e5m2_t);
46-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e8m0_t);46+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_fp8_e8m0_t);
47-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint16_t);47+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint16_t);
48-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int16_t);48+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int16_t);
49-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_half);49+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_half);
50-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_bfloat16_t);50+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_bfloat16_t);
51-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint32_t);51+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_uint32_t);
52-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int32_t);52+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_int32_t);
53-TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_float);53+TEST_VECTOR_COMPUTE_DUPLICATE_INSTR(Vdup, asc_duplicate, vdup, vector_float);
54- 54+ 
55-#define TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(class_name, c_api_name, cce_name, data_type) \55+#define TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(data_type) \
56- \56+ namespace { \
57- class TestVectorCompute##class_name####data_type##MergeCApi : public testing::Test { \57+ void vdup_highest_##data_type##_Stub(data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) \
58- protected: \58+ { \
59- void SetUp() {} \59+ EXPECT_EQ(pos, static_cast<int32_t>(POS_HIGHEST.value)); \
60- void TearDown() {} \60+ EXPECT_EQ(mode, MODE_ZEROING); \
61- }; \61+ } \
62- \62+ } \
63- namespace { \63+ \
64- void cce_name##_merging_##data_type##_Stub( \64+ TEST(TestVectorComputeVdupHighest, data_type) \
65- data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) \65+ { \
66- { \66+ data_type dst; \
67- EXPECT_EQ(pos, static_cast<int32_t>(POS_LOWEST.value)); \67+ data_type src0; \
68- EXPECT_EQ(mode, MODE_MERGING); \68+ vector_bool mask; \
69- } \69+ MOCKER_CPP(vdup, void(data_type&, data_type, vector_bool, int32_t, Literal)) \
70- } \70+ .times(1) \
71- \71+ .will(invoke(vdup_highest_##data_type##_Stub)); \
72- TEST_F(TestVectorCompute##class_name####data_type##MergeCApi, c_api_name##_##data_type##_MergeSucc) \72+ asc_duplicate_highest(dst, src0, mask); \
73- { \73+ GlobalMockObject::verify(); \
74- data_type dst; \74+ }
75- data_type src0; \75+ 
76- vector_bool mask; \76+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_uint8_t)
77- \77+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_int8_t)
78- MOCKER_CPP(cce_name, void(data_type&, data_type, vector_bool, int32_t, Literal)) \78+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_fp8_e4m3fn_t)
79- .times(1) \79+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_fp8_e5m2_t)
80- .will(invoke(cce_name##_merging_##data_type##_Stub)); \80+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_fp8_e8m0_t)
81- \81+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_uint16_t)
82- c_api_name(dst, src0, mask); \82+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_int16_t)
83- GlobalMockObject::verify(); \83+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_half)
84- }84+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_bfloat16_t)
85- 85+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_uint32_t)
86-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint8_t);86+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_int32_t)
87-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int8_t);87+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR(vector_float)
88-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e4m3fn_t);88+ 
89-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e5m2_t);89+#undef TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_INSTR
90-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e8m0_t);90+ 
91-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint16_t);91+#define TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(class_name, c_api_name, cce_name, data_type) \
92-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int16_t);92+ \
93-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_half);93+ class TestVectorCompute##class_name####data_type##MergeCApi : public testing::Test { \
94-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_bfloat16_t);94+ protected: \
95-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint32_t);95+ void SetUp() {} \
96-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int32_t);96+ void TearDown() {} \
97-TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_float);97+ }; \
98+ \
99+ namespace { \
100+ void cce_name##_merging_##data_type##_Stub( \
101+ data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) \
102+ { \
103+ EXPECT_EQ(pos, static_cast<int32_t>(POS_LOWEST.value)); \
104+ EXPECT_EQ(mode, MODE_MERGING); \
105+ } \
106+ } \
107+ \
108+ TEST_F(TestVectorCompute##class_name####data_type##MergeCApi, c_api_name##_##data_type##_MergeSucc) \
109+ { \
110+ data_type dst; \
111+ data_type src0; \
112+ vector_bool mask; \
113+ \
114+ MOCKER_CPP(cce_name, void(data_type&, data_type, vector_bool, int32_t, Literal)) \
115+ .times(1) \
116+ .will(invoke(cce_name##_merging_##data_type##_Stub)); \
117+ \
118+ c_api_name(dst, src0, mask); \
119+ GlobalMockObject::verify(); \
120+ }
121+ 
122+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint8_t);
123+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int8_t);
124+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e4m3fn_t);
125+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e5m2_t);
126+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_fp8_e8m0_t);
127+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint16_t);
128+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int16_t);
129+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_half);
130+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_bfloat16_t);
131+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_uint32_t);
132+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_int32_t);
133+TEST_VECTOR_COMPUTE_DUPLICATE_MERGE_INSTR(Vdup, asc_duplicate_merge, vdup, vector_float);
134+ 
135+#define TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(data_type) \
136+ namespace { \
137+ void vdup_highest_merge_##data_type##_Stub( \
138+ data_type& dst, data_type src0, vector_bool mask, int32_t pos, Literal mode) \
139+ { \
140+ EXPECT_EQ(pos, static_cast<int32_t>(POS_HIGHEST.value)); \
141+ EXPECT_EQ(mode, MODE_MERGING); \
142+ } \
143+ } \
144+ \
145+ TEST(TestVectorComputeVdupHighestMerge, data_type) \
146+ { \
147+ data_type dst; \
148+ data_type src0; \
149+ vector_bool mask; \
150+ MOCKER_CPP(vdup, void(data_type&, data_type, vector_bool, int32_t, Literal)) \
151+ .times(1) \
152+ .will(invoke(vdup_highest_merge_##data_type##_Stub)); \
153+ asc_duplicate_highest_merge(dst, src0, mask); \
154+ GlobalMockObject::verify(); \
155+ }
156+ 
157+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_uint8_t)
158+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_int8_t)
159+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_fp8_e4m3fn_t)
160+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_fp8_e5m2_t)
161+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_fp8_e8m0_t)
162+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_uint16_t)
163+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_int16_t)
164+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_half)
165+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_bfloat16_t)
166+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_uint32_t)
167+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_int32_t)
168+TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR(vector_float)
169+ 
170+#undef TEST_VECTOR_COMPUTE_DUPLICATE_HIGHEST_MERGE_INSTR
Mtests/api/c_api/stub/cce_stub.h+3-0
@@ -12,6 +12,9 @@
12#include <cstdint>12#include <cstdint>
13#include "stub_fun.h"13#include "stub_fun.h"
14 14 
15+// The compiler supports this overload, but it is missing from the CPU debug stub header.
16+int64_t ld_dev(__gm__ int64_t* src, int16_t offset);
17+ 
15#ifndef ULL18#ifndef ULL
16#define ULL unsigned long long19#define ULL unsigned long long
17#endif20#endif