Merged
DV121 DV122适配:ops-math 部分算子新增ascend960dt/960pr平台支持 #7033
DV121 DV122适配:ops-math 部分算子新增ascend960dt/960pr平台支持 #7033
Merged
huairuchenAdd on 15 days ago
Total: 39 files changed+2353-25
@@ -44,6 +44,7 @@ public:
44 this->AICore().AddConfig("ascend950");44 this->AICore().AddConfig("ascend950");
45 this->AICore().AddConfig("ascend350");45 this->AICore().AddConfig("ascend350");
46 this->AICore().AddConfig("ascend960dt");46 this->AICore().AddConfig("ascend960dt");
47+ this->AICore().AddConfig("ascend960pr");
47 }48 }
48};49};
49 50 
@@ -58,7 +58,9 @@ ge::graphStatus ChunkCatTiling::GetPlatformInfo()
58 OP_CHECK_IF(ubSize_ == 0, OP_LOGE(context_, "ubSize is 0"), return ge::GRAPH_FAILED);58 OP_CHECK_IF(ubSize_ == 0, OP_LOGE(context_, "ubSize is 0"), return ge::GRAPH_FAILED);
59 sysWorkspaceSize_ = ascendcPlatform.GetLibApiWorkSpaceSize();59 sysWorkspaceSize_ = ascendcPlatform.GetLibApiWorkSpaceSize();
60 isRegBase = Ops::Base::IsRegbaseSocVersion(context_);60 isRegBase = Ops::Base::IsRegbaseSocVersion(context_);
61- isAscend960dt_ = ascendcPlatform.GetCurNpuArch() == NpuArch::DAV_9201;61+ // Ascend960 (DAV_9201/9202) 共用同一套 tiling 适配逻辑
62+ isAscend960dt_ = ascendcPlatform.GetCurNpuArch() == NpuArch::DAV_9201 ||
63+ ascendcPlatform.GetCurNpuArch() == NpuArch::DAV_9202;
62 return ge::GRAPH_SUCCESS;64 return ge::GRAPH_SUCCESS;
63}65}
64 66 
@@ -0,0 +1,292 @@
1+{
2+ "op_type": "ChunkCat",
3+ "op_list": [
4+ {
5+ "bin_filename": "ChunkCat_fp16_fp16",
6+ "inputs": [
7+ [
8+ {
9+ "name": "x",
10+ "index": 0,
11+ "dtype": "float16",
12+ "format": "ND",
13+ "paramType": "dynamic",
14+ "shape": [
15+ -2
16+ ]
17+ }
18+ ]
19+ ],
20+ "outputs": [
21+ {
22+ "name": "y",
23+ "index": 0,
24+ "dtype": "float16",
25+ "format": "ND",
26+ "paramType": "required",
27+ "shape": [
28+ -2
29+ ]
30+ }
31+ ],
32+ "attrs": [
33+ {
34+ "name": "dim",
35+ "dtype": "int",
36+ "value": null
37+ },
38+ {
39+ "name": "num_chunks",
40+ "dtype": "int",
41+ "value": null
42+ }
43+ ]
44+ },
45+ {
46+ "bin_filename": "ChunkCat_fp16_bf16",
47+ "inputs": [
48+ [
49+ {
50+ "name": "x",
51+ "index": 0,
52+ "dtype": "float16",
53+ "format": "ND",
54+ "paramType": "dynamic",
55+ "shape": [
56+ -2
57+ ]
58+ }
59+ ]
60+ ],
61+ "outputs": [
62+ {
63+ "name": "y",
64+ "index": 0,
65+ "dtype": "bfloat16",
66+ "format": "ND",
67+ "paramType": "required",
68+ "shape": [
69+ -2
70+ ]
71+ }
72+ ],
73+ "attrs": [
74+ {
75+ "name": "dim",
76+ "dtype": "int",
77+ "value": null
78+ },
79+ {
80+ "name": "num_chunks",
81+ "dtype": "int",
82+ "value": null
83+ }
84+ ]
85+ },
86+ {
87+ "bin_filename": "ChunkCat_fp16_fp32",
88+ "inputs": [
89+ [
90+ {
91+ "name": "x",
92+ "index": 0,
93+ "dtype": "float16",
94+ "format": "ND",
95+ "paramType": "dynamic",
96+ "shape": [
97+ -2
98+ ]
99+ }
100+ ]
101+ ],
102+ "outputs": [
103+ {
104+ "name": "y",
105+ "index": 0,
106+ "dtype": "float32",
107+ "format": "ND",
108+ "paramType": "required",
109+ "shape": [
110+ -2
111+ ]
112+ }
113+ ],
114+ "attrs": [
115+ {
116+ "name": "dim",
117+ "dtype": "int",
118+ "value": null
119+ },
120+ {
121+ "name": "num_chunks",
122+ "dtype": "int",
123+ "value": null
124+ }
125+ ]
126+ },
127+ {
128+ "bin_filename": "ChunkCat_bf16_fp16",
129+ "inputs": [
130+ [
131+ {
132+ "name": "x",
133+ "index": 0,
134+ "dtype": "bfloat16",
135+ "format": "ND",
136+ "paramType": "dynamic",
137+ "shape": [
138+ -2
139+ ]
140+ }
141+ ]
142+ ],
143+ "outputs": [
144+ {
145+ "name": "y",
146+ "index": 0,
147+ "dtype": "float16",
148+ "format": "ND",
149+ "paramType": "required",
150+ "shape": [
151+ -2
152+ ]
153+ }
154+ ],
155+ "attrs": [
156+ {
157+ "name": "dim",
158+ "dtype": "int",
159+ "value": null
160+ },
161+ {
162+ "name": "num_chunks",
163+ "dtype": "int",
164+ "value": null
165+ }
166+ ]
167+ },
168+ {
169+ "bin_filename": "ChunkCat_bf16_bf16",
170+ "inputs": [
171+ [
172+ {
173+ "name": "x",
174+ "index": 0,
175+ "dtype": "bfloat16",
176+ "format": "ND",
177+ "paramType": "dynamic",
178+ "shape": [
179+ -2
180+ ]
181+ }
182+ ]
183+ ],
184+ "outputs": [
185+ {
186+ "name": "y",
187+ "index": 0,
188+ "dtype": "bfloat16",
189+ "format": "ND",
190+ "paramType": "required",
191+ "shape": [
192+ -2
193+ ]
194+ }
195+ ],
196+ "attrs": [
197+ {
198+ "name": "dim",
199+ "dtype": "int",
200+ "value": null
201+ },
202+ {
203+ "name": "num_chunks",
204+ "dtype": "int",
205+ "value": null
206+ }
207+ ]
208+ },
209+ {
210+ "bin_filename": "ChunkCat_bf16_fp32",
211+ "inputs": [
212+ [
213+ {
214+ "name": "x",
215+ "index": 0,
216+ "dtype": "bfloat16",
217+ "format": "ND",
218+ "paramType": "dynamic",
219+ "shape": [
220+ -2
221+ ]
222+ }
223+ ]
224+ ],
225+ "outputs": [
226+ {
227+ "name": "y",
228+ "index": 0,
229+ "dtype": "float32",
230+ "format": "ND",
231+ "paramType": "required",
232+ "shape": [
233+ -2
234+ ]
235+ }
236+ ],
237+ "attrs": [
238+ {
239+ "name": "dim",
240+ "dtype": "int",
241+ "value": null
242+ },
243+ {
244+ "name": "num_chunks",
245+ "dtype": "int",
246+ "value": null
247+ }
248+ ]
249+ },
250+ {
251+ "bin_filename": "ChunkCat_fp32_fp32",
252+ "inputs": [
253+ [
254+ {
255+ "name": "x",
256+ "index": 0,
257+ "dtype": "float32",
258+ "format": "ND",
259+ "paramType": "dynamic",
260+ "shape": [
261+ -2
262+ ]
263+ }
264+ ]
265+ ],
266+ "outputs": [
267+ {
268+ "name": "y",
269+ "index": 0,
270+ "dtype": "float32",
271+ "format": "ND",
272+ "paramType": "required",
273+ "shape": [
274+ -2
275+ ]
276+ }
277+ ],
278+ "attrs": [
279+ {
280+ "name": "dim",
281+ "dtype": "int",
282+ "value": null
283+ },
284+ {
285+ "name": "num_chunks",
286+ "dtype": "int",
287+ "value": null
288+ }
289+ ]
290+ }
291+ ]
292+}
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/built-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[ChunkCat]
13+default=0
@@ -12,7 +12,7 @@
12 * \file chunk_cat.cpp12 * \file chunk_cat.cpp
13 * \brief13 * \brief
14 */14 */
15-#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 920115+#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
16#include "chunk_cat_arch35.h"16#include "chunk_cat_arch35.h"
17#else17#else
18#include "chunk_cat.h"18#include "chunk_cat.h"
@@ -24,7 +24,7 @@ extern "C" __global__ __aicore__ void chunk_cat(GM_ADDR x, GM_ADDR y, GM_ADDR wo
24 AscendC::TPipe pipe;24 AscendC::TPipe pipe;
25 REGISTER_TILING_DEFAULT(ChunkCatTilingData);25 REGISTER_TILING_DEFAULT(ChunkCatTilingData);
26 GET_TILING_DATA_WITH_STRUCT(ChunkCatTilingData, tilingData, tiling);26 GET_TILING_DATA_WITH_STRUCT(ChunkCatTilingData, tilingData, tiling);
27-#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 920127+#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
28 ChunkCatArch35<DTYPE_X, DTYPE_Y> op(&pipe);28 ChunkCatArch35<DTYPE_X, DTYPE_Y> op(&pipe);
29 op.Init(x, y, tilingData);29 op.Init(x, y, tilingData);
30 op.Process();30 op.Process();
@@ -371,7 +371,7 @@ private:
371 uint16_t rowLoop1 = 0;371 uint16_t rowLoop1 = 0;
372 uint16_t rowLoop2 = 0;372 uint16_t rowLoop2 = 0;
373 // 部分pad情况下pad部分373 // 部分pad情况下pad部分
374-#if __NPU_ARCH__ == 9201374+#if __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
375 uint32_t colLen0 = 0;375 uint32_t colLen0 = 0;
376 uint32_t colLen1 = 0;376 uint32_t colLen1 = 0;
377 uint16_t colLoop1 = 0;377 uint16_t colLoop1 = 0;
@@ -404,7 +404,7 @@ private:
404 }404 }
405 }405 }
406 uint32_t rowNum = GetVRegSize() / sizeof(T2) / srcLen;406 uint32_t rowNum = GetVRegSize() / sizeof(T2) / srcLen;
407-#if __NPU_ARCH__ == 9201407+#if __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
408 if constexpr (IS_SCATTER) {408 if constexpr (IS_SCATTER) {
409 if constexpr (std::is_same_v<T2, bfloat16_t> || std::is_same_v<T2, half>) {409 if constexpr (std::is_same_v<T2, bfloat16_t> || std::is_same_v<T2, half>) {
410 uint32_t rowNumSafe = srcLen < UINT16_LIMIT ? (UINT16_LIMIT - srcLen) / rowStride + 1 : rowNum;410 uint32_t rowNumSafe = srcLen < UINT16_LIMIT ? (UINT16_LIMIT - srcLen) / rowStride + 1 : rowNum;
@@ -150,7 +150,7 @@ public:
150#if __CCE_AICORE__ == 310150#if __CCE_AICORE__ == 310
151 AscendC::DataCopyPad<T1, PaddingMode::Compact>(srcLocal_[localOffset], srcGlobal_[gmOffset], copyParams,151 AscendC::DataCopyPad<T1, PaddingMode::Compact>(srcLocal_[localOffset], srcGlobal_[gmOffset], copyParams,
152 padParams);152 padParams);
153-#elif __NPU_ARCH__ == 9201153+#elif __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
154 AscendC::DataCopyPad<T1, PaddingMode::Compact>(srcLocal_[localOffset], srcGlobal_[gmOffset], copyParams,154 AscendC::DataCopyPad<T1, PaddingMode::Compact>(srcLocal_[localOffset], srcGlobal_[gmOffset], copyParams,
155 padParams, true);155 padParams, true);
156#else156#else
@@ -207,7 +207,7 @@ public:
207 // 1.3.1207 // 1.3.1
208 blockLen = (remainderCol - tensorInfo.startOffset) * sizeof(T1);208 blockLen = (remainderCol - tensorInfo.startOffset) * sizeof(T1);
209 int64_t localOffsetPart = 0;209 int64_t localOffsetPart = 0;
210-#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 9201210+#if __CCE_AICORE__ == 310 || __NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202
211 localOffsetPart = localOffset + GetAlign((tensorInfo.chunkRow - ubLoopInfo.rowStart) * tensorInfo.splitCol,211 localOffsetPart = localOffset + GetAlign((tensorInfo.chunkRow - ubLoopInfo.rowStart) * tensorInfo.splitCol,
212 srcEleUbBlock_);212 srcEleUbBlock_);
213#else213#else
@@ -55,7 +55,9 @@ bool IsSecondLastAxisAiCoreSupport(op::DataType dataType, int64_t size, int64_t
55 return false;55 return false;
56 }56 }
57 57 
58- const bool isArch35 = GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510;58+ // Ascend960 (DAV_9201/9202) 与 Ascend950 共用 arch35 UB 规格
59+ const auto curArch = GetCurrentPlatformInfo().GetCurNpuArch();
60+ const bool isArch35 = curArch == NpuArch::DAV_3510 || curArch == NpuArch::DAV_9201 || curArch == NpuArch::DAV_9202;
59 const int64_t ubSize = isArch35 ? ARCH35_UB_SIZE : ARCH22_UB_SIZE;61 const int64_t ubSize = isArch35 ? ARCH35_UB_SIZE : ARCH22_UB_SIZE;
60 int64_t ubSizeT2 = 0;62 int64_t ubSizeT2 = 0;
61 if (dataType == DataType::DT_FLOAT16 || dataType == DataType::DT_BF16) {63 if (dataType == DataType::DT_FLOAT16 || dataType == DataType::DT_BF16) {
@@ -0,0 +1,167 @@
1+{
2+ "op_type": "UnfoldGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "UnfoldGrad_250a5853",
6+ "inputs": [
7+ {
8+ "name": "grad_out",
9+ "index": 0,
10+ "dtype": "float16",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ]
16+ },
17+ {
18+ "name": "input_sizes",
19+ "index": 1,
20+ "dtype": "int64",
21+ "format": "ND",
22+ "paramType": "required",
23+ "shape": [
24+ -2
25+ ]
26+ }
27+ ],
28+ "outputs": [
29+ {
30+ "name": "grad_in",
31+ "index": 0,
32+ "dtype": "float16",
33+ "format": "ND",
34+ "paramType": "required",
35+ "shape": [
36+ -2
37+ ]
38+ }
39+ ],
40+ "attrs": [
41+ {
42+ "name": "dim",
43+ "dtype": "int",
44+ "value": null
45+ },
46+ {
47+ "name": "size",
48+ "dtype": "int",
49+ "value": null
50+ },
51+ {
52+ "name": "step",
53+ "dtype": "int",
54+ "value": null
55+ }
56+ ]
57+ },
58+ {
59+ "bin_filename": "UnfoldGrad_0f47731e",
60+ "inputs": [
61+ {
62+ "name": "grad_out",
63+ "index": 0,
64+ "dtype": "bfloat16",
65+ "format": "ND",
66+ "paramType": "required",
67+ "shape": [
68+ -2
69+ ]
70+ },
71+ {
72+ "name": "input_sizes",
73+ "index": 1,
74+ "dtype": "int64",
75+ "format": "ND",
76+ "paramType": "required",
77+ "shape": [
78+ -2
79+ ]
80+ }
81+ ],
82+ "outputs": [
83+ {
84+ "name": "grad_in",
85+ "index": 0,
86+ "dtype": "bfloat16",
87+ "format": "ND",
88+ "paramType": "required",
89+ "shape": [
90+ -2
91+ ]
92+ }
93+ ],
94+ "attrs": [
95+ {
96+ "name": "dim",
97+ "dtype": "int",
98+ "value": null
99+ },
100+ {
101+ "name": "size",
102+ "dtype": "int",
103+ "value": null
104+ },
105+ {
106+ "name": "step",
107+ "dtype": "int",
108+ "value": null
109+ }
110+ ]
111+ },
112+ {
113+ "bin_filename": "UnfoldGrad_071cdc8b",
114+ "inputs": [
115+ {
116+ "name": "grad_out",
117+ "index": 0,
118+ "dtype": "float32",
119+ "format": "ND",
120+ "paramType": "required",
121+ "shape": [
122+ -2
123+ ]
124+ },
125+ {
126+ "name": "input_sizes",
127+ "index": 1,
128+ "dtype": "int64",
129+ "format": "ND",
130+ "paramType": "required",
131+ "shape": [
132+ -2
133+ ]
134+ }
135+ ],
136+ "outputs": [
137+ {
138+ "name": "grad_in",
139+ "index": 0,
140+ "dtype": "float32",
141+ "format": "ND",
142+ "paramType": "required",
143+ "shape": [
144+ -2
145+ ]
146+ }
147+ ],
148+ "attrs": [
149+ {
150+ "name": "dim",
151+ "dtype": "int",
152+ "value": null
153+ },
154+ {
155+ "name": "size",
156+ "dtype": "int",
157+ "value": null
158+ },
159+ {
160+ "name": "step",
161+ "dtype": "int",
162+ "value": null
163+ }
164+ ]
165+ }
166+ ]
167+}
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[UnfoldGrad]
13+simplified_key_mode=0
@@ -0,0 +1,167 @@
1+{
2+ "op_type": "UnfoldGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "UnfoldGrad_250a5853",
6+ "inputs": [
7+ {
8+ "name": "grad_out",
9+ "index": 0,
10+ "dtype": "float16",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ]
16+ },
17+ {
18+ "name": "input_sizes",
19+ "index": 1,
20+ "dtype": "int64",
21+ "format": "ND",
22+ "paramType": "required",
23+ "shape": [
24+ -2
25+ ]
26+ }
27+ ],
28+ "outputs": [
29+ {
30+ "name": "grad_in",
31+ "index": 0,
32+ "dtype": "float16",
33+ "format": "ND",
34+ "paramType": "required",
35+ "shape": [
36+ -2
37+ ]
38+ }
39+ ],
40+ "attrs": [
41+ {
42+ "name": "dim",
43+ "dtype": "int",
44+ "value": null
45+ },
46+ {
47+ "name": "size",
48+ "dtype": "int",
49+ "value": null
50+ },
51+ {
52+ "name": "step",
53+ "dtype": "int",
54+ "value": null
55+ }
56+ ]
57+ },
58+ {
59+ "bin_filename": "UnfoldGrad_0f47731e",
60+ "inputs": [
61+ {
62+ "name": "grad_out",
63+ "index": 0,
64+ "dtype": "bfloat16",
65+ "format": "ND",
66+ "paramType": "required",
67+ "shape": [
68+ -2
69+ ]
70+ },
71+ {
72+ "name": "input_sizes",
73+ "index": 1,
74+ "dtype": "int64",
75+ "format": "ND",
76+ "paramType": "required",
77+ "shape": [
78+ -2
79+ ]
80+ }
81+ ],
82+ "outputs": [
83+ {
84+ "name": "grad_in",
85+ "index": 0,
86+ "dtype": "bfloat16",
87+ "format": "ND",
88+ "paramType": "required",
89+ "shape": [
90+ -2
91+ ]
92+ }
93+ ],
94+ "attrs": [
95+ {
96+ "name": "dim",
97+ "dtype": "int",
98+ "value": null
99+ },
100+ {
101+ "name": "size",
102+ "dtype": "int",
103+ "value": null
104+ },
105+ {
106+ "name": "step",
107+ "dtype": "int",
108+ "value": null
109+ }
110+ ]
111+ },
112+ {
113+ "bin_filename": "UnfoldGrad_071cdc8b",
114+ "inputs": [
115+ {
116+ "name": "grad_out",
117+ "index": 0,
118+ "dtype": "float32",
119+ "format": "ND",
120+ "paramType": "required",
121+ "shape": [
122+ -2
123+ ]
124+ },
125+ {
126+ "name": "input_sizes",
127+ "index": 1,
128+ "dtype": "int64",
129+ "format": "ND",
130+ "paramType": "required",
131+ "shape": [
132+ -2
133+ ]
134+ }
135+ ],
136+ "outputs": [
137+ {
138+ "name": "grad_in",
139+ "index": 0,
140+ "dtype": "float32",
141+ "format": "ND",
142+ "paramType": "required",
143+ "shape": [
144+ -2
145+ ]
146+ }
147+ ],
148+ "attrs": [
149+ {
150+ "name": "dim",
151+ "dtype": "int",
152+ "value": null
153+ },
154+ {
155+ "name": "size",
156+ "dtype": "int",
157+ "value": null
158+ },
159+ {
160+ "name": "step",
161+ "dtype": "int",
162+ "value": null
163+ }
164+ ]
165+ }
166+ ]
167+}
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[UnfoldGrad]
13+simplified_key_mode=0
@@ -43,6 +43,8 @@ public:
43 this->AICore().AddConfig("ascend910b");43 this->AICore().AddConfig("ascend910b");
44 this->AICore().AddConfig("ascend910_93");44 this->AICore().AddConfig("ascend910_93");
45 this->AICore().AddConfig("ascend950");45 this->AICore().AddConfig("ascend950");
46+ this->AICore().AddConfig("ascend960dt");
47+ this->AICore().AddConfig("ascend960pr");
46 }48 }
47};49};
48 50 
@@ -56,8 +56,10 @@ public:
56 this->AICore().AddConfig("ascend910b", aicore_config);56 this->AICore().AddConfig("ascend910b", aicore_config);
57 this->AICore().AddConfig("ascend910_93", aicore_config);57 this->AICore().AddConfig("ascend910_93", aicore_config);
58 this->AICore().AddConfig("ascend950", aicore_config);58 this->AICore().AddConfig("ascend950", aicore_config);
59+ this->AICore().AddConfig("ascend960dt", aicore_config);
60+ this->AICore().AddConfig("ascend960pr", aicore_config);
59 }61 }
60};62};
61 63 
62OP_ADD(AmpUpdateScale);64OP_ADD(AmpUpdateScale);
63-}65+} // namespace ops
@@ -0,0 +1,227 @@
1+{
2+ "op_type": "AmpUpdateScale",
3+ "op_list": [
4+ {
5+ "bin_filename": "AmpUpdateScale_6993687eb06b70ac5aa102ddd4bef1c4",
6+ "inputs": [
7+ {
8+ "name": "current_scale",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ]
16+ },
17+ {
18+ "name": "growth_tracker",
19+ "index": 1,
20+ "dtype": "int32",
21+ "format": "ND",
22+ "paramType": "required",
23+ "shape": [
24+ -2
25+ ]
26+ },
27+ {
28+ "name": "found_inf",
29+ "index": 2,
30+ "dtype": "float32",
31+ "format": "ND",
32+ "paramType": "required",
33+ "shape": [
34+ -2
35+ ]
36+ }
37+ ],
38+ "outputs": [
39+ {
40+ "name": "updated_scale",
41+ "index": 0,
42+ "dtype": "float32",
43+ "format": "ND",
44+ "paramType": "required",
45+ "shape": [
46+ -2
47+ ]
48+ },
49+ {
50+ "name": "updated_growth_tracker",
51+ "index": 1,
52+ "dtype": "int32",
53+ "format": "ND",
54+ "paramType": "required",
55+ "shape": [
56+ -2
57+ ]
58+ }
59+ ],
60+ "attrs": [
61+ {
62+ "name": "growth_factor",
63+ "dtype": "float",
64+ "value": 0.0
65+ },
66+ {
67+ "name": "backoff_factor",
68+ "dtype": "float",
69+ "value": 0.0
70+ },
71+ {
72+ "name": "growth_interval",
73+ "dtype": "int",
74+ "value": 0
75+ }
76+ ]
77+ },
78+ {
79+ "bin_filename": "AmpUpdateScale_float16",
80+ "inputs": [
81+ {
82+ "name": "current_scale",
83+ "index": 0,
84+ "dtype": "float16",
85+ "format": "ND",
86+ "paramType": "required",
87+ "shape": [
88+ -2
89+ ]
90+ },
91+ {
92+ "name": "growth_tracker",
93+ "index": 1,
94+ "dtype": "int32",
95+ "format": "ND",
96+ "paramType": "required",
97+ "shape": [
98+ -2
99+ ]
100+ },
101+ {
102+ "name": "found_inf",
103+ "index": 2,
104+ "dtype": "float16",
105+ "format": "ND",
106+ "paramType": "required",
107+ "shape": [
108+ -2
109+ ]
110+ }
111+ ],
112+ "outputs": [
113+ {
114+ "name": "updated_scale",
115+ "index": 0,
116+ "dtype": "float16",
117+ "format": "ND",
118+ "paramType": "required",
119+ "shape": [
120+ -2
121+ ]
122+ },
123+ {
124+ "name": "updated_growth_tracker",
125+ "index": 1,
126+ "dtype": "int32",
127+ "format": "ND",
128+ "paramType": "required",
129+ "shape": [
130+ -2
131+ ]
132+ }
133+ ],
134+ "attrs": [
135+ {
136+ "name": "growth_factor",
137+ "dtype": "float",
138+ "value": 0.0
139+ },
140+ {
141+ "name": "backoff_factor",
142+ "dtype": "float",
143+ "value": 0.0
144+ },
145+ {
146+ "name": "growth_interval",
147+ "dtype": "int",
148+ "value": 0
149+ }
150+ ]
151+ },
152+ {
153+ "bin_filename": "AmpUpdateScale_bfloat16",
154+ "inputs": [
155+ {
156+ "name": "current_scale",
157+ "index": 0,
158+ "dtype": "bfloat16",
159+ "format": "ND",
160+ "paramType": "required",
161+ "shape": [
162+ -2
163+ ]
164+ },
165+ {
166+ "name": "growth_tracker",
167+ "index": 1,
168+ "dtype": "int32",
169+ "format": "ND",
170+ "paramType": "required",
171+ "shape": [
172+ -2
173+ ]
174+ },
175+ {
176+ "name": "found_inf",
177+ "index": 2,
178+ "dtype": "bfloat16",
179+ "format": "ND",
180+ "paramType": "required",
181+ "shape": [
182+ -2
183+ ]
184+ }
185+ ],
186+ "outputs": [
187+ {
188+ "name": "updated_scale",
189+ "index": 0,
190+ "dtype": "bfloat16",
191+ "format": "ND",
192+ "paramType": "required",
193+ "shape": [
194+ -2
195+ ]
196+ },
197+ {
198+ "name": "updated_growth_tracker",
199+ "index": 1,
200+ "dtype": "int32",
201+ "format": "ND",
202+ "paramType": "required",
203+ "shape": [
204+ -2
205+ ]
206+ }
207+ ],
208+ "attrs": [
209+ {
210+ "name": "growth_factor",
211+ "dtype": "float",
212+ "value": 0.0
213+ },
214+ {
215+ "name": "backoff_factor",
216+ "dtype": "float",
217+ "value": 0.0
218+ },
219+ {
220+ "name": "growth_interval",
221+ "dtype": "int",
222+ "value": 0
223+ }
224+ ]
225+ }
226+ ]
227+}
@@ -0,0 +1,227 @@
1+{
2+ "op_type": "AmpUpdateScale",
3+ "op_list": [
4+ {
5+ "bin_filename": "AmpUpdateScale_6993687eb06b70ac5aa102ddd4bef1c4",
6+ "inputs": [
7+ {
8+ "name": "current_scale",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ]
16+ },
17+ {
18+ "name": "growth_tracker",
19+ "index": 1,
20+ "dtype": "int32",
21+ "format": "ND",
22+ "paramType": "required",
23+ "shape": [
24+ -2
25+ ]
26+ },
27+ {
28+ "name": "found_inf",
29+ "index": 2,
30+ "dtype": "float32",
31+ "format": "ND",
32+ "paramType": "required",
33+ "shape": [
34+ -2
35+ ]
36+ }
37+ ],
38+ "outputs": [
39+ {
40+ "name": "updated_scale",
41+ "index": 0,
42+ "dtype": "float32",
43+ "format": "ND",
44+ "paramType": "required",
45+ "shape": [
46+ -2
47+ ]
48+ },
49+ {
50+ "name": "updated_growth_tracker",
51+ "index": 1,
52+ "dtype": "int32",
53+ "format": "ND",
54+ "paramType": "required",
55+ "shape": [
56+ -2
57+ ]
58+ }
59+ ],
60+ "attrs": [
61+ {
62+ "name": "growth_factor",
63+ "dtype": "float",
64+ "value": 0.0
65+ },
66+ {
67+ "name": "backoff_factor",
68+ "dtype": "float",
69+ "value": 0.0
70+ },
71+ {
72+ "name": "growth_interval",
73+ "dtype": "int",
74+ "value": 0
75+ }
76+ ]
77+ },
78+ {
79+ "bin_filename": "AmpUpdateScale_float16",
80+ "inputs": [
81+ {
82+ "name": "current_scale",
83+ "index": 0,
84+ "dtype": "float16",
85+ "format": "ND",
86+ "paramType": "required",
87+ "shape": [
88+ -2
89+ ]
90+ },
91+ {
92+ "name": "growth_tracker",
93+ "index": 1,
94+ "dtype": "int32",
95+ "format": "ND",
96+ "paramType": "required",
97+ "shape": [
98+ -2
99+ ]
100+ },
101+ {
102+ "name": "found_inf",
103+ "index": 2,
104+ "dtype": "float16",
105+ "format": "ND",
106+ "paramType": "required",
107+ "shape": [
108+ -2
109+ ]
110+ }
111+ ],
112+ "outputs": [
113+ {
114+ "name": "updated_scale",
115+ "index": 0,
116+ "dtype": "float16",
117+ "format": "ND",
118+ "paramType": "required",
119+ "shape": [
120+ -2
121+ ]
122+ },
123+ {
124+ "name": "updated_growth_tracker",
125+ "index": 1,
126+ "dtype": "int32",
127+ "format": "ND",
128+ "paramType": "required",
129+ "shape": [
130+ -2
131+ ]
132+ }
133+ ],
134+ "attrs": [
135+ {
136+ "name": "growth_factor",
137+ "dtype": "float",
138+ "value": 0.0
139+ },
140+ {
141+ "name": "backoff_factor",
142+ "dtype": "float",
143+ "value": 0.0
144+ },
145+ {
146+ "name": "growth_interval",
147+ "dtype": "int",
148+ "value": 0
149+ }
150+ ]
151+ },
152+ {
153+ "bin_filename": "AmpUpdateScale_bfloat16",
154+ "inputs": [
155+ {
156+ "name": "current_scale",
157+ "index": 0,
158+ "dtype": "bfloat16",
159+ "format": "ND",
160+ "paramType": "required",
161+ "shape": [
162+ -2
163+ ]
164+ },
165+ {
166+ "name": "growth_tracker",
167+ "index": 1,
168+ "dtype": "int32",
169+ "format": "ND",
170+ "paramType": "required",
171+ "shape": [
172+ -2
173+ ]
174+ },
175+ {
176+ "name": "found_inf",
177+ "index": 2,
178+ "dtype": "bfloat16",
179+ "format": "ND",
180+ "paramType": "required",
181+ "shape": [
182+ -2
183+ ]
184+ }
185+ ],
186+ "outputs": [
187+ {
188+ "name": "updated_scale",
189+ "index": 0,
190+ "dtype": "bfloat16",
191+ "format": "ND",
192+ "paramType": "required",
193+ "shape": [
194+ -2
195+ ]
196+ },
197+ {
198+ "name": "updated_growth_tracker",
199+ "index": 1,
200+ "dtype": "int32",
201+ "format": "ND",
202+ "paramType": "required",
203+ "shape": [
204+ -2
205+ ]
206+ }
207+ ],
208+ "attrs": [
209+ {
210+ "name": "growth_factor",
211+ "dtype": "float",
212+ "value": 0.0
213+ },
214+ {
215+ "name": "backoff_factor",
216+ "dtype": "float",
217+ "value": 0.0
218+ },
219+ {
220+ "name": "growth_interval",
221+ "dtype": "int",
222+ "value": 0
223+ }
224+ ]
225+ }
226+ ]
227+}
@@ -9,9 +9,9 @@
9# ----------------------------------------------------------------------------9# ----------------------------------------------------------------------------
10 10 
11# 设置算子定义时支持的芯片类型11# 设置算子定义时支持的芯片类型
12-set(SUPPORT_COMPUTE_UNIT "ascend910b" "ascend910_93" "ascend950")12+set(SUPPORT_COMPUTE_UNIT "ascend910b" "ascend910_93" "ascend950" "ascend960dt" "ascend960pr")
13# 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译13# 设置每种芯片类型对应的tiling文件目录,即采用op_host目录下哪个文件夹下的tiling文件编译
14-set(SUPPORT_TILING_DIR "arch22" "arch22" "arch35")14+set(SUPPORT_TILING_DIR "arch22" "arch22" "arch35" "arch35" "arch35")
15add_all_modules_sources(OPTYPE cdist_grad ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR})15add_all_modules_sources(OPTYPE cdist_grad ACLNNTYPE aclnn_exclude COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT} TILING_DIR ${SUPPORT_TILING_DIR})
16 16 
17# L0 op source is in op_api/, picked up by GLOB for the default path17# L0 op source is in op_api/, picked up by GLOB for the default path
@@ -26,6 +26,7 @@
26#include "opdev/op_dfx.h"26#include "opdev/op_dfx.h"
27#include "opdev/op_executor.h"27#include "opdev/op_executor.h"
28#include "opdev/common_types.h"28#include "opdev/common_types.h"
29+#include "opdev/platform.h"
29#include "opdev/data_type_utils.h"30#include "opdev/data_type_utils.h"
30#include "opdev/op_log.h"31#include "opdev/op_log.h"
31#include "opdev/shape_utils.h"32#include "opdev/shape_utils.h"
@@ -202,9 +203,9 @@ aclnnStatus aclnnCdistBackwardGetWorkspaceSize(const aclTensor* grad, const aclT
202 gradContiguous != nullptr && x1Contiguous != nullptr && x2Contiguous != nullptr && cdistContiguous != nullptr,203 gradContiguous != nullptr && x1Contiguous != nullptr && x2Contiguous != nullptr && cdistContiguous != nullptr,
203 ACLNN_ERR_INNER_NULLPTR);204 ACLNN_ERR_INNER_NULLPTR);
204 205 
205- SocVersion socVersion = GetCurrentPlatformInfo().GetSocVersion();206+ // BF16 先转 FP32 仅 arch22(DAV_2201,即 Ascend910B/Ascend910_93)需要,与原 SocVersion 判断等价
206 bool needCast = x1->GetDataType() == op::DataType::DT_BF16 &&207 bool needCast = x1->GetDataType() == op::DataType::DT_BF16 &&
207- (socVersion == SocVersion::ASCEND910B || socVersion == SocVersion::ASCEND910_93);208+ GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_2201;
208 209 
209 if (needCast) {210 if (needCast) {
210 gradContiguous = l0op::Cast(gradContiguous, op::DataType::DT_FLOAT, uniqueExecutor.get());211 gradContiguous = l0op::Cast(gradContiguous, op::DataType::DT_FLOAT, uniqueExecutor.get());
@@ -37,6 +37,8 @@ public:
37 .DynamicShapeSupportFlag(true)37 .DynamicShapeSupportFlag(true)
38 .ExtendCfgInfo("opFile.value", "cdist_grad_apt");38 .ExtendCfgInfo("opFile.value", "cdist_grad_apt");
39 this->AICore().AddConfig("ascend950", aicoreConfig);39 this->AICore().AddConfig("ascend950", aicoreConfig);
40+ this->AICore().AddConfig("ascend960dt", aicoreConfig);
41+ this->AICore().AddConfig("ascend960pr", aicoreConfig);
40 42 
41 OpAICoreConfig aicore910bConfig;43 OpAICoreConfig aicore910bConfig;
42 aicore910bConfig.DynamicCompileStaticFlag(true)44 aicore910bConfig.DynamicCompileStaticFlag(true)
@@ -0,0 +1,50 @@
1+{
2+ "op_type": "CdistGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "CdistGrad_6d3844989f83c1de15cb106d07138083",
6+ "inputs": [
7+ { "name": "grad", "index": 0, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
8+ { "name": "x1", "index": 1, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
9+ { "name": "x2", "index": 2, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
10+ { "name": "cdist", "index": 3, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
11+ ],
12+ "outputs": [
13+ { "name": "y", "index": 0, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
14+ ],
15+ "attrs": [
16+ { "name": "p", "dtype": "float", "value": null }
17+ ]
18+ },
19+ {
20+ "bin_filename": "CdistGrad_876161686b0bc0b329614d9b9c3b5adc",
21+ "inputs": [
22+ { "name": "grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
23+ { "name": "x1", "index": 1, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
24+ { "name": "x2", "index": 2, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
25+ { "name": "cdist", "index": 3, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
26+ ],
27+ "outputs": [
28+ { "name": "y", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
29+ ],
30+ "attrs": [
31+ { "name": "p", "dtype": "float", "value": null }
32+ ]
33+ },
34+ {
35+ "bin_filename": "CdistGrad_632f457d4787205a3df995bf22221bbe",
36+ "inputs": [
37+ { "name": "grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
38+ { "name": "x1", "index": 1, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
39+ { "name": "x2", "index": 2, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
40+ { "name": "cdist", "index": 3, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
41+ ],
42+ "outputs": [
43+ { "name": "y", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
44+ ],
45+ "attrs": [
46+ { "name": "p", "dtype": "float", "value": null }
47+ ]
48+ }
49+ ]
50+}
@@ -0,0 +1,2 @@
1+[CdistGrad]
2+default=0
@@ -0,0 +1,50 @@
1+{
2+ "op_type": "CdistGrad",
3+ "op_list": [
4+ {
5+ "bin_filename": "CdistGrad_6d3844989f83c1de15cb106d07138083",
6+ "inputs": [
7+ { "name": "grad", "index": 0, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
8+ { "name": "x1", "index": 1, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
9+ { "name": "x2", "index": 2, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
10+ { "name": "cdist", "index": 3, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
11+ ],
12+ "outputs": [
13+ { "name": "y", "index": 0, "dtype": "float16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
14+ ],
15+ "attrs": [
16+ { "name": "p", "dtype": "float", "value": null }
17+ ]
18+ },
19+ {
20+ "bin_filename": "CdistGrad_876161686b0bc0b329614d9b9c3b5adc",
21+ "inputs": [
22+ { "name": "grad", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
23+ { "name": "x1", "index": 1, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
24+ { "name": "x2", "index": 2, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
25+ { "name": "cdist", "index": 3, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
26+ ],
27+ "outputs": [
28+ { "name": "y", "index": 0, "dtype": "bfloat16", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
29+ ],
30+ "attrs": [
31+ { "name": "p", "dtype": "float", "value": null }
32+ ]
33+ },
34+ {
35+ "bin_filename": "CdistGrad_632f457d4787205a3df995bf22221bbe",
36+ "inputs": [
37+ { "name": "grad", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
38+ { "name": "x1", "index": 1, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
39+ { "name": "x2", "index": 2, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" },
40+ { "name": "cdist", "index": 3, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
41+ ],
42+ "outputs": [
43+ { "name": "y", "index": 0, "dtype": "float32", "format": "ND", "paramType": "required", "shape": [-2], "format_match_mode": "FormatAgnostic" }
44+ ],
45+ "attrs": [
46+ { "name": "p", "dtype": "float", "value": null }
47+ ]
48+ }
49+ ]
50+}
@@ -0,0 +1,2 @@
1+[CdistGrad]
2+default=0
@@ -0,0 +1,212 @@
1+{
2+ "op_type": "TransformBiasRescaleQkv",
3+ "op_list": [
4+ {
5+ "bin_filename": "TransformBiasRescaleQkv_float32",
6+ "inputs": [
7+ {
8+ "name": "qkv",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ],
16+ "format_match_mode": "FormatAgnostic"
17+ },
18+ {
19+ "name": "qkv_bias",
20+ "index": 1,
21+ "dtype": "float32",
22+ "format": "ND",
23+ "paramType": "required",
24+ "shape": [
25+ -2
26+ ],
27+ "format_match_mode": "FormatAgnostic"
28+ }
29+ ],
30+ "outputs": [
31+ {
32+ "name": "q",
33+ "index": 0,
34+ "dtype": "float32",
35+ "format": "ND",
36+ "paramType": "required",
37+ "shape": [
38+ -2
39+ ],
40+ "format_match_mode": "FormatAgnostic"
41+ },
42+ {
43+ "name": "k",
44+ "index": 1,
45+ "dtype": "float32",
46+ "format": "ND",
47+ "paramType": "required",
48+ "shape": [
49+ -2
50+ ],
51+ "format_match_mode": "FormatAgnostic"
52+ },
53+ {
54+ "name": "v",
55+ "index": 2,
56+ "dtype": "float32",
57+ "format": "ND",
58+ "paramType": "required",
59+ "shape": [
60+ -2
61+ ],
62+ "format_match_mode": "FormatAgnostic"
63+ }
64+ ],
65+ "attrs": [
66+ {
67+ "name": "num_heads",
68+ "dtype": "int",
69+ "value": 0
70+ }
71+ ]
72+ },
73+ {
74+ "bin_filename": "TransformBiasRescaleQkv_float16",
75+ "inputs": [
76+ {
77+ "name": "qkv",
78+ "index": 0,
79+ "dtype": "float16",
80+ "format": "ND",
81+ "paramType": "required",
82+ "shape": [
83+ -2
84+ ],
85+ "format_match_mode": "FormatAgnostic"
86+ },
87+ {
88+ "name": "qkv_bias",
89+ "index": 1,
90+ "dtype": "float16",
91+ "format": "ND",
92+ "paramType": "required",
93+ "shape": [
94+ -2
95+ ],
96+ "format_match_mode": "FormatAgnostic"
97+ }
98+ ],
99+ "outputs": [
100+ {
101+ "name": "q",
102+ "index": 0,
103+ "dtype": "float16",
104+ "format": "ND",
105+ "paramType": "required",
106+ "shape": [
107+ -2
108+ ],
109+ "format_match_mode": "FormatAgnostic"
110+ },
111+ {
112+ "name": "k",
113+ "index": 1,
114+ "dtype": "float16",
115+ "format": "ND",
116+ "paramType": "required",
117+ "shape": [
118+ -2
119+ ],
120+ "format_match_mode": "FormatAgnostic"
121+ },
122+ {
123+ "name": "v",
124+ "index": 2,
125+ "dtype": "float16",
126+ "format": "ND",
127+ "paramType": "required",
128+ "shape": [
129+ -2
130+ ],
131+ "format_match_mode": "FormatAgnostic"
132+ }
133+ ],
134+ "attrs": [
135+ {
136+ "name": "num_heads",
137+ "dtype": "int",
138+ "value": 0
139+ }
140+ ]
141+ },
142+ {
143+ "bin_filename": "TransformBiasRescaleQkv_bfloat16",
144+ "inputs": [
145+ {
146+ "name": "qkv",
147+ "index": 0,
148+ "dtype": "bfloat16",
149+ "format": "ND",
150+ "paramType": "required",
151+ "shape": [
152+ -2
153+ ],
154+ "format_match_mode": "FormatAgnostic"
155+ },
156+ {
157+ "name": "qkv_bias",
158+ "index": 1,
159+ "dtype": "bfloat16",
160+ "format": "ND",
161+ "paramType": "required",
162+ "shape": [
163+ -2
164+ ],
165+ "format_match_mode": "FormatAgnostic"
166+ }
167+ ],
168+ "outputs": [
169+ {
170+ "name": "q",
171+ "index": 0,
172+ "dtype": "bfloat16",
173+ "format": "ND",
174+ "paramType": "required",
175+ "shape": [
176+ -2
177+ ],
178+ "format_match_mode": "FormatAgnostic"
179+ },
180+ {
181+ "name": "k",
182+ "index": 1,
183+ "dtype": "bfloat16",
184+ "format": "ND",
185+ "paramType": "required",
186+ "shape": [
187+ -2
188+ ],
189+ "format_match_mode": "FormatAgnostic"
190+ },
191+ {
192+ "name": "v",
193+ "index": 2,
194+ "dtype": "bfloat16",
195+ "format": "ND",
196+ "paramType": "required",
197+ "shape": [
198+ -2
199+ ],
200+ "format_match_mode": "FormatAgnostic"
201+ }
202+ ],
203+ "attrs": [
204+ {
205+ "name": "num_heads",
206+ "dtype": "int",
207+ "value": 0
208+ }
209+ ]
210+ }
211+ ]
212+}
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[TransformBiasRescaleQkv]
13+default=0
@@ -0,0 +1,212 @@
1+{
2+ "op_type": "TransformBiasRescaleQkv",
3+ "op_list": [
4+ {
5+ "bin_filename": "TransformBiasRescaleQkv_float32",
6+ "inputs": [
7+ {
8+ "name": "qkv",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ],
16+ "format_match_mode": "FormatAgnostic"
17+ },
18+ {
19+ "name": "qkv_bias",
20+ "index": 1,
21+ "dtype": "float32",
22+ "format": "ND",
23+ "paramType": "required",
24+ "shape": [
25+ -2
26+ ],
27+ "format_match_mode": "FormatAgnostic"
28+ }
29+ ],
30+ "outputs": [
31+ {
32+ "name": "q",
33+ "index": 0,
34+ "dtype": "float32",
35+ "format": "ND",
36+ "paramType": "required",
37+ "shape": [
38+ -2
39+ ],
40+ "format_match_mode": "FormatAgnostic"
41+ },
42+ {
43+ "name": "k",
44+ "index": 1,
45+ "dtype": "float32",
46+ "format": "ND",
47+ "paramType": "required",
48+ "shape": [
49+ -2
50+ ],
51+ "format_match_mode": "FormatAgnostic"
52+ },
53+ {
54+ "name": "v",
55+ "index": 2,
56+ "dtype": "float32",
57+ "format": "ND",
58+ "paramType": "required",
59+ "shape": [
60+ -2
61+ ],
62+ "format_match_mode": "FormatAgnostic"
63+ }
64+ ],
65+ "attrs": [
66+ {
67+ "name": "num_heads",
68+ "dtype": "int",
69+ "value": 0
70+ }
71+ ]
72+ },
73+ {
74+ "bin_filename": "TransformBiasRescaleQkv_float16",
75+ "inputs": [
76+ {
77+ "name": "qkv",
78+ "index": 0,
79+ "dtype": "float16",
80+ "format": "ND",
81+ "paramType": "required",
82+ "shape": [
83+ -2
84+ ],
85+ "format_match_mode": "FormatAgnostic"
86+ },
87+ {
88+ "name": "qkv_bias",
89+ "index": 1,
90+ "dtype": "float16",
91+ "format": "ND",
92+ "paramType": "required",
93+ "shape": [
94+ -2
95+ ],
96+ "format_match_mode": "FormatAgnostic"
97+ }
98+ ],
99+ "outputs": [
100+ {
101+ "name": "q",
102+ "index": 0,
103+ "dtype": "float16",
104+ "format": "ND",
105+ "paramType": "required",
106+ "shape": [
107+ -2
108+ ],
109+ "format_match_mode": "FormatAgnostic"
110+ },
111+ {
112+ "name": "k",
113+ "index": 1,
114+ "dtype": "float16",
115+ "format": "ND",
116+ "paramType": "required",
117+ "shape": [
118+ -2
119+ ],
120+ "format_match_mode": "FormatAgnostic"
121+ },
122+ {
123+ "name": "v",
124+ "index": 2,
125+ "dtype": "float16",
126+ "format": "ND",
127+ "paramType": "required",
128+ "shape": [
129+ -2
130+ ],
131+ "format_match_mode": "FormatAgnostic"
132+ }
133+ ],
134+ "attrs": [
135+ {
136+ "name": "num_heads",
137+ "dtype": "int",
138+ "value": 0
139+ }
140+ ]
141+ },
142+ {
143+ "bin_filename": "TransformBiasRescaleQkv_bfloat16",
144+ "inputs": [
145+ {
146+ "name": "qkv",
147+ "index": 0,
148+ "dtype": "bfloat16",
149+ "format": "ND",
150+ "paramType": "required",
151+ "shape": [
152+ -2
153+ ],
154+ "format_match_mode": "FormatAgnostic"
155+ },
156+ {
157+ "name": "qkv_bias",
158+ "index": 1,
159+ "dtype": "bfloat16",
160+ "format": "ND",
161+ "paramType": "required",
162+ "shape": [
163+ -2
164+ ],
165+ "format_match_mode": "FormatAgnostic"
166+ }
167+ ],
168+ "outputs": [
169+ {
170+ "name": "q",
171+ "index": 0,
172+ "dtype": "bfloat16",
173+ "format": "ND",
174+ "paramType": "required",
175+ "shape": [
176+ -2
177+ ],
178+ "format_match_mode": "FormatAgnostic"
179+ },
180+ {
181+ "name": "k",
182+ "index": 1,
183+ "dtype": "bfloat16",
184+ "format": "ND",
185+ "paramType": "required",
186+ "shape": [
187+ -2
188+ ],
189+ "format_match_mode": "FormatAgnostic"
190+ },
191+ {
192+ "name": "v",
193+ "index": 2,
194+ "dtype": "bfloat16",
195+ "format": "ND",
196+ "paramType": "required",
197+ "shape": [
198+ -2
199+ ],
200+ "format_match_mode": "FormatAgnostic"
201+ }
202+ ],
203+ "attrs": [
204+ {
205+ "name": "num_heads",
206+ "dtype": "int",
207+ "value": 0
208+ }
209+ ]
210+ }
211+ ]
212+}
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/build-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[TransformBiasRescaleQkv]
13+default=0
@@ -61,8 +61,10 @@ public:
61 this->AICore().AddConfig("ascend910b");61 this->AICore().AddConfig("ascend910b");
62 this->AICore().AddConfig("ascend910_93");62 this->AICore().AddConfig("ascend910_93");
63 this->AICore().AddConfig("ascend950");63 this->AICore().AddConfig("ascend950");
64+ this->AICore().AddConfig("ascend960dt");
65+ this->AICore().AddConfig("ascend960pr");
64 66 
65- OpAICoreConfig config_kirin = GetKirinCoreConfig(); 67+ OpAICoreConfig config_kirin = GetKirinCoreConfig();
66 this->AICore().AddConfig("kirinx90", config_kirin);68 this->AICore().AddConfig("kirinx90", config_kirin);
67 this->AICore().AddConfig("kirin9030", config_kirin);69 this->AICore().AddConfig("kirin9030", config_kirin);
68 }70 }
@@ -112,4 +114,4 @@ private:
112};114};
113 115 
114OP_ADD(TransformBiasRescaleQkv);116OP_ADD(TransformBiasRescaleQkv);
115-} // namespace ops117+} // namespace ops
@@ -8,8 +8,8 @@
8# See LICENSE in the root of the software repository for the full text of the License.8# See LICENSE in the root of the software repository for the full text of the License.
9# ---------------------------------------------------------------------------------------------------------9# ---------------------------------------------------------------------------------------------------------
10 10 
11-set(SUPPORT_COMPUTE_UNIT "ascend950" "ascend350")11+set(SUPPORT_COMPUTE_UNIT "ascend950" "ascend350" "ascend960dt" "ascend960pr")
12-set(SUPPORT_TILING_DIR "arch35" "arch35")12+set(SUPPORT_TILING_DIR "arch35" "arch35" "arch35" "arch35")
13add_all_modules_sources(OPTYPE drop_out_v3_grad ACLNNTYPE aclnn_exclude13add_all_modules_sources(OPTYPE drop_out_v3_grad ACLNNTYPE aclnn_exclude
14 COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT}14 COMPUTE_UNIT ${SUPPORT_COMPUTE_UNIT}
15 TILING_DIR ${SUPPORT_TILING_DIR}15 TILING_DIR ${SUPPORT_TILING_DIR}
@@ -44,12 +44,13 @@ static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST = {op::DataT
44// 对外接口mask仅支持uint844// 对外接口mask仅支持uint8
45static const std::initializer_list<op::DataType> MASK_DTYPE_SUPPORT_LIST = {op::DataType::DT_UINT8};45static const std::initializer_list<op::DataType> MASK_DTYPE_SUPPORT_LIST = {op::DataType::DT_UINT8};
46 46 
47-// 本算子仅在Ascend910_95上提供kernel实现,其余产品形态直接返回错误47+// 本算子仅在 Ascend910_95/Ascend960(DAV_3510/9201/9202)上提供kernel实现,其余产品形态直接返回错误
48static inline bool CheckSocVersion()48static inline bool CheckSocVersion()
49{49{
50- if (op::GetCurrentPlatformInfo().GetCurNpuArch() != NpuArch::DAV_3510) {50+ const auto curArch = op::GetCurrentPlatformInfo().GetCurNpuArch();
51+ if (curArch != NpuArch::DAV_3510 && curArch != NpuArch::DAV_9201 && curArch != NpuArch::DAV_9202) {
51 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "aclnnDropoutV3Grad is not supported in current socversion %d.",52 OP_LOGE(ACLNN_ERR_PARAM_INVALID, "aclnnDropoutV3Grad is not supported in current socversion %d.",
52- static_cast<int>(op::GetCurrentPlatformInfo().GetCurNpuArch()));53+ static_cast<int>(curArch));
53 return false;54 return false;
54 }55 }
55 return true;56 return true;
@@ -0,0 +1,311 @@
1+{
2+ "op_type": "DropOutV3Grad",
3+ "op_list": [
4+ {
5+ "bin_filename": "DropOutV3Grad_f32_uint8",
6+ "inputs": [
7+ {
8+ "name": "grad_y",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ],
16+ "format_match_mode": "FormatAgnostic"
17+ },
18+ {
19+ "name": "mask",
20+ "index": 1,
21+ "dtype": "uint8",
22+ "format": "ND",
23+ "paramType": "required",
24+ "shape": [
25+ -2
26+ ],
27+ "format_match_mode": "FormatAgnostic"
28+ },
29+ {
30+ "name": "scale",
31+ "index": 2,
32+ "dtype": "float32",
33+ "format": "ND",
34+ "paramType": "required",
35+ "shape": [
36+ -2
37+ ],
38+ "format_match_mode": "FormatAgnostic"
39+ }
40+ ],
41+ "outputs": [
42+ {
43+ "name": "grad_x",
44+ "index": 0,
45+ "dtype": "float32",
46+ "format": "ND",
47+ "paramType": "required",
48+ "shape": [
49+ -2
50+ ],
51+ "format_match_mode": "FormatAgnostic"
52+ }
53+ ]
54+ },
55+ {
56+ "bin_filename": "DropOutV3Grad_f16_uint8",
57+ "inputs": [
58+ {
59+ "name": "grad_y",
60+ "index": 0,
61+ "dtype": "float16",
62+ "format": "ND",
63+ "paramType": "required",
64+ "shape": [
65+ -2
66+ ],
67+ "format_match_mode": "FormatAgnostic"
68+ },
69+ {
70+ "name": "mask",
71+ "index": 1,
72+ "dtype": "uint8",
73+ "format": "ND",
74+ "paramType": "required",
75+ "shape": [
76+ -2
77+ ],
78+ "format_match_mode": "FormatAgnostic"
79+ },
80+ {
81+ "name": "scale",
82+ "index": 2,
83+ "dtype": "float32",
84+ "format": "ND",
85+ "paramType": "required",
86+ "shape": [
87+ -2
88+ ],
89+ "format_match_mode": "FormatAgnostic"
90+ }
91+ ],
92+ "outputs": [
93+ {
94+ "name": "grad_x",
95+ "index": 0,
96+ "dtype": "float16",
97+ "format": "ND",
98+ "paramType": "required",
99+ "shape": [
100+ -2
101+ ],
102+ "format_match_mode": "FormatAgnostic"
103+ }
104+ ]
105+ },
106+ {
107+ "bin_filename": "DropOutV3Grad_bf16_uint8",
108+ "inputs": [
109+ {
110+ "name": "grad_y",
111+ "index": 0,
112+ "dtype": "bfloat16",
113+ "format": "ND",
114+ "paramType": "required",
115+ "shape": [
116+ -2
117+ ],
118+ "format_match_mode": "FormatAgnostic"
119+ },
120+ {
121+ "name": "mask",
122+ "index": 1,
123+ "dtype": "uint8",
124+ "format": "ND",
125+ "paramType": "required",
126+ "shape": [
127+ -2
128+ ],
129+ "format_match_mode": "FormatAgnostic"
130+ },
131+ {
132+ "name": "scale",
133+ "index": 2,
134+ "dtype": "float32",
135+ "format": "ND",
136+ "paramType": "required",
137+ "shape": [
138+ -2
139+ ],
140+ "format_match_mode": "FormatAgnostic"
141+ }
142+ ],
143+ "outputs": [
144+ {
145+ "name": "grad_x",
146+ "index": 0,
147+ "dtype": "bfloat16",
148+ "format": "ND",
149+ "paramType": "required",
150+ "shape": [
151+ -2
152+ ],
153+ "format_match_mode": "FormatAgnostic"
154+ }
155+ ]
156+ },
157+ {
158+ "bin_filename": "DropOutV3Grad_f32_uint1",
159+ "inputs": [
160+ {
161+ "name": "grad_y",
162+ "index": 0,
163+ "dtype": "float32",
164+ "format": "ND",
165+ "paramType": "required",
166+ "shape": [
167+ -2
168+ ],
169+ "format_match_mode": "FormatAgnostic"
170+ },
171+ {
172+ "name": "mask",
173+ "index": 1,
174+ "dtype": "uint1",
175+ "format": "ND",
176+ "paramType": "required",
177+ "shape": [
178+ -2
179+ ],
180+ "format_match_mode": "FormatAgnostic"
181+ },
182+ {
183+ "name": "scale",
184+ "index": 2,
185+ "dtype": "float32",
186+ "format": "ND",
187+ "paramType": "required",
188+ "shape": [
189+ -2
190+ ],
191+ "format_match_mode": "FormatAgnostic"
192+ }
193+ ],
194+ "outputs": [
195+ {
196+ "name": "grad_x",
197+ "index": 0,
198+ "dtype": "float32",
199+ "format": "ND",
200+ "paramType": "required",
201+ "shape": [
202+ -2
203+ ],
204+ "format_match_mode": "FormatAgnostic"
205+ }
206+ ]
207+ },
208+ {
209+ "bin_filename": "DropOutV3Grad_f16_uint1",
210+ "inputs": [
211+ {
212+ "name": "grad_y",
213+ "index": 0,
214+ "dtype": "float16",
215+ "format": "ND",
216+ "paramType": "required",
217+ "shape": [
218+ -2
219+ ],
220+ "format_match_mode": "FormatAgnostic"
221+ },
222+ {
223+ "name": "mask",
224+ "index": 1,
225+ "dtype": "uint1",
226+ "format": "ND",
227+ "paramType": "required",
228+ "shape": [
229+ -2
230+ ],
231+ "format_match_mode": "FormatAgnostic"
232+ },
233+ {
234+ "name": "scale",
235+ "index": 2,
236+ "dtype": "float32",
237+ "format": "ND",
238+ "paramType": "required",
239+ "shape": [
240+ -2
241+ ],
242+ "format_match_mode": "FormatAgnostic"
243+ }
244+ ],
245+ "outputs": [
246+ {
247+ "name": "grad_x",
248+ "index": 0,
249+ "dtype": "float16",
250+ "format": "ND",
251+ "paramType": "required",
252+ "shape": [
253+ -2
254+ ],
255+ "format_match_mode": "FormatAgnostic"
256+ }
257+ ]
258+ },
259+ {
260+ "bin_filename": "DropOutV3Grad_bf16_uint1",
261+ "inputs": [
262+ {
263+ "name": "grad_y",
264+ "index": 0,
265+ "dtype": "bfloat16",
266+ "format": "ND",
267+ "paramType": "required",
268+ "shape": [
269+ -2
270+ ],
271+ "format_match_mode": "FormatAgnostic"
272+ },
273+ {
274+ "name": "mask",
275+ "index": 1,
276+ "dtype": "uint1",
277+ "format": "ND",
278+ "paramType": "required",
279+ "shape": [
280+ -2
281+ ],
282+ "format_match_mode": "FormatAgnostic"
283+ },
284+ {
285+ "name": "scale",
286+ "index": 2,
287+ "dtype": "float32",
288+ "format": "ND",
289+ "paramType": "required",
290+ "shape": [
291+ -2
292+ ],
293+ "format_match_mode": "FormatAgnostic"
294+ }
295+ ],
296+ "outputs": [
297+ {
298+ "name": "grad_x",
299+ "index": 0,
300+ "dtype": "bfloat16",
301+ "format": "ND",
302+ "paramType": "required",
303+ "shape": [
304+ -2
305+ ],
306+ "format_match_mode": "FormatAgnostic"
307+ }
308+ ]
309+ }
310+ ]
311+}
@@ -0,0 +1,311 @@
1+{
2+ "op_type": "DropOutV3Grad",
3+ "op_list": [
4+ {
5+ "bin_filename": "DropOutV3Grad_f32_uint8",
6+ "inputs": [
7+ {
8+ "name": "grad_y",
9+ "index": 0,
10+ "dtype": "float32",
11+ "format": "ND",
12+ "paramType": "required",
13+ "shape": [
14+ -2
15+ ],
16+ "format_match_mode": "FormatAgnostic"
17+ },
18+ {
19+ "name": "mask",
20+ "index": 1,
21+ "dtype": "uint8",
22+ "format": "ND",
23+ "paramType": "required",
24+ "shape": [
25+ -2
26+ ],
27+ "format_match_mode": "FormatAgnostic"
28+ },
29+ {
30+ "name": "scale",
31+ "index": 2,
32+ "dtype": "float32",
33+ "format": "ND",
34+ "paramType": "required",
35+ "shape": [
36+ -2
37+ ],
38+ "format_match_mode": "FormatAgnostic"
39+ }
40+ ],
41+ "outputs": [
42+ {
43+ "name": "grad_x",
44+ "index": 0,
45+ "dtype": "float32",
46+ "format": "ND",
47+ "paramType": "required",
48+ "shape": [
49+ -2
50+ ],
51+ "format_match_mode": "FormatAgnostic"
52+ }
53+ ]
54+ },
55+ {
56+ "bin_filename": "DropOutV3Grad_f16_uint8",
57+ "inputs": [
58+ {
59+ "name": "grad_y",
60+ "index": 0,
61+ "dtype": "float16",
62+ "format": "ND",
63+ "paramType": "required",
64+ "shape": [
65+ -2
66+ ],
67+ "format_match_mode": "FormatAgnostic"
68+ },
69+ {
70+ "name": "mask",
71+ "index": 1,
72+ "dtype": "uint8",
73+ "format": "ND",
74+ "paramType": "required",
75+ "shape": [
76+ -2
77+ ],
78+ "format_match_mode": "FormatAgnostic"
79+ },
80+ {
81+ "name": "scale",
82+ "index": 2,
83+ "dtype": "float32",
84+ "format": "ND",
85+ "paramType": "required",
86+ "shape": [
87+ -2
88+ ],
89+ "format_match_mode": "FormatAgnostic"
90+ }
91+ ],
92+ "outputs": [
93+ {
94+ "name": "grad_x",
95+ "index": 0,
96+ "dtype": "float16",
97+ "format": "ND",
98+ "paramType": "required",
99+ "shape": [
100+ -2
101+ ],
102+ "format_match_mode": "FormatAgnostic"
103+ }
104+ ]
105+ },
106+ {
107+ "bin_filename": "DropOutV3Grad_bf16_uint8",
108+ "inputs": [
109+ {
110+ "name": "grad_y",
111+ "index": 0,
112+ "dtype": "bfloat16",
113+ "format": "ND",
114+ "paramType": "required",
115+ "shape": [
116+ -2
117+ ],
118+ "format_match_mode": "FormatAgnostic"
119+ },
120+ {
121+ "name": "mask",
122+ "index": 1,
123+ "dtype": "uint8",
124+ "format": "ND",
125+ "paramType": "required",
126+ "shape": [
127+ -2
128+ ],
129+ "format_match_mode": "FormatAgnostic"
130+ },
131+ {
132+ "name": "scale",
133+ "index": 2,
134+ "dtype": "float32",
135+ "format": "ND",
136+ "paramType": "required",
137+ "shape": [
138+ -2
139+ ],
140+ "format_match_mode": "FormatAgnostic"
141+ }
142+ ],
143+ "outputs": [
144+ {
145+ "name": "grad_x",
146+ "index": 0,
147+ "dtype": "bfloat16",
148+ "format": "ND",
149+ "paramType": "required",
150+ "shape": [
151+ -2
152+ ],
153+ "format_match_mode": "FormatAgnostic"
154+ }
155+ ]
156+ },
157+ {
158+ "bin_filename": "DropOutV3Grad_f32_uint1",
159+ "inputs": [
160+ {
161+ "name": "grad_y",
162+ "index": 0,
163+ "dtype": "float32",
164+ "format": "ND",
165+ "paramType": "required",
166+ "shape": [
167+ -2
168+ ],
169+ "format_match_mode": "FormatAgnostic"
170+ },
171+ {
172+ "name": "mask",
173+ "index": 1,
174+ "dtype": "uint1",
175+ "format": "ND",
176+ "paramType": "required",
177+ "shape": [
178+ -2
179+ ],
180+ "format_match_mode": "FormatAgnostic"
181+ },
182+ {
183+ "name": "scale",
184+ "index": 2,
185+ "dtype": "float32",
186+ "format": "ND",
187+ "paramType": "required",
188+ "shape": [
189+ -2
190+ ],
191+ "format_match_mode": "FormatAgnostic"
192+ }
193+ ],
194+ "outputs": [
195+ {
196+ "name": "grad_x",
197+ "index": 0,
198+ "dtype": "float32",
199+ "format": "ND",
200+ "paramType": "required",
201+ "shape": [
202+ -2
203+ ],
204+ "format_match_mode": "FormatAgnostic"
205+ }
206+ ]
207+ },
208+ {
209+ "bin_filename": "DropOutV3Grad_f16_uint1",
210+ "inputs": [
211+ {
212+ "name": "grad_y",
213+ "index": 0,
214+ "dtype": "float16",
215+ "format": "ND",
216+ "paramType": "required",
217+ "shape": [
218+ -2
219+ ],
220+ "format_match_mode": "FormatAgnostic"
221+ },
222+ {
223+ "name": "mask",
224+ "index": 1,
225+ "dtype": "uint1",
226+ "format": "ND",
227+ "paramType": "required",
228+ "shape": [
229+ -2
230+ ],
231+ "format_match_mode": "FormatAgnostic"
232+ },
233+ {
234+ "name": "scale",
235+ "index": 2,
236+ "dtype": "float32",
237+ "format": "ND",
238+ "paramType": "required",
239+ "shape": [
240+ -2
241+ ],
242+ "format_match_mode": "FormatAgnostic"
243+ }
244+ ],
245+ "outputs": [
246+ {
247+ "name": "grad_x",
248+ "index": 0,
249+ "dtype": "float16",
250+ "format": "ND",
251+ "paramType": "required",
252+ "shape": [
253+ -2
254+ ],
255+ "format_match_mode": "FormatAgnostic"
256+ }
257+ ]
258+ },
259+ {
260+ "bin_filename": "DropOutV3Grad_bf16_uint1",
261+ "inputs": [
262+ {
263+ "name": "grad_y",
264+ "index": 0,
265+ "dtype": "bfloat16",
266+ "format": "ND",
267+ "paramType": "required",
268+ "shape": [
269+ -2
270+ ],
271+ "format_match_mode": "FormatAgnostic"
272+ },
273+ {
274+ "name": "mask",
275+ "index": 1,
276+ "dtype": "uint1",
277+ "format": "ND",
278+ "paramType": "required",
279+ "shape": [
280+ -2
281+ ],
282+ "format_match_mode": "FormatAgnostic"
283+ },
284+ {
285+ "name": "scale",
286+ "index": 2,
287+ "dtype": "float32",
288+ "format": "ND",
289+ "paramType": "required",
290+ "shape": [
291+ -2
292+ ],
293+ "format_match_mode": "FormatAgnostic"
294+ }
295+ ],
296+ "outputs": [
297+ {
298+ "name": "grad_x",
299+ "index": 0,
300+ "dtype": "bfloat16",
301+ "format": "ND",
302+ "paramType": "required",
303+ "shape": [
304+ -2
305+ ],
306+ "format_match_mode": "FormatAgnostic"
307+ }
308+ ]
309+ }
310+ ]
311+}
@@ -53,6 +53,8 @@ public:
53 .ExtendCfgInfo("opFile.value", "drop_out_v3_grad");53 .ExtendCfgInfo("opFile.value", "drop_out_v3_grad");
54 this->AICore().AddConfig("ascend950", aicoreConfig);54 this->AICore().AddConfig("ascend950", aicoreConfig);
55 this->AICore().AddConfig("ascend350", aicoreConfig);55 this->AICore().AddConfig("ascend350", aicoreConfig);
56+ this->AICore().AddConfig("ascend960dt", aicoreConfig);
57+ this->AICore().AddConfig("ascend960pr", aicoreConfig);
56 }58 }
57};59};
58 60 
@@ -167,8 +167,7 @@ static const aclTensor* uniformDoublePath(const aclTensor* uniformOut, double fr
167static const aclTensor* uniformDavidPath(const aclTensor* selfRef, uint64_t seed, uint64_t offset, double from,167static const aclTensor* uniformDavidPath(const aclTensor* selfRef, uint64_t seed, uint64_t offset, double from,
168 double to, aclOpExecutor* executor)168 double to, aclOpExecutor* executor)
169{169{
170- if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510 &&170+ if (IsRegBase(GetCurrentPlatformInfo().GetCurNpuArch()) && selfRef->GetDataType() != DataType::DT_DOUBLE) {
171- selfRef->GetDataType() != DataType::DT_DOUBLE) {
172 if (!aclnnGetPytorchRandom()) {171 if (!aclnnGetPytorchRandom()) {
173 OP_LOGD("compat mode, use V3 uniform");172 OP_LOGD("compat mode, use V3 uniform");
174 int32_t uniformV3ScaleMode = 0;173 int32_t uniformV3ScaleMode = 0;
@@ -189,8 +188,7 @@ static const aclTensor* uniformTensorDavidPath(const aclTensor* selfRef, const a
189 const aclTensor* offsetTensor, uint64_t offset, double from, double to,188 const aclTensor* offsetTensor, uint64_t offset, double from, double to,
190 aclOpExecutor* executor)189 aclOpExecutor* executor)
191{190{
192- if (GetCurrentPlatformInfo().GetCurNpuArch() == NpuArch::DAV_3510 &&191+ if (IsRegBase(GetCurrentPlatformInfo().GetCurNpuArch()) && selfRef->GetDataType() != DataType::DT_DOUBLE) {
193- selfRef->GetDataType() != DataType::DT_DOUBLE) {
194 if (!aclnnGetPytorchRandom()) {192 if (!aclnnGetPytorchRandom()) {
195 OP_LOGD("compat mode, use V3 uniform");193 OP_LOGD("compat mode, use V3 uniform");
196 auto seedUint64 = l0op::Cast(seedTensor, op::DataType::DT_UINT64, executor);194 auto seedUint64 = l0op::Cast(seedTensor, op::DataType::DT_UINT64, executor);
@@ -8,7 +8,7 @@
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-# StatelessExponential only supports ascend950, ascend350 and ascend960dt ascend960pr (arch35 / SIMT).11+# StatelessExponential only supports ascend950, ascend350, ascend960dt and ascend960pr (arch35 / SIMT).
12# Kernel computation reuses sim_thread_exponential, so it must be listed as a12# Kernel computation reuses sim_thread_exponential, so it must be listed as a
13# build dependency.13# build dependency.
14set(SUPPORT_COMPUTE_UNIT "ascend950" "ascend350" "ascend960dt" "ascend960pr")14set(SUPPORT_COMPUTE_UNIT "ascend950" "ascend350" "ascend960dt" "ascend960pr")
@@ -0,0 +1,13 @@
1+; 该文件主要影响 opc 工具 编译二进制kernel时, --simplified_key_mode 选项中填写的值,格式如下所示:
2+; [某算子]
3+; default=xx
4+; ascendxx=xx
5+; 其中,default为默认mode,ascendxx为可选mode,如果不同芯片有差异化要求时,需要配置;
6+; 1)如果没有配置:非ascendC算子继续按空处理,即opc编译命令中不添加 --simplified_key_mode 选项,AscendC算子按照 simplified_key_mode=0 处理
7+; 2)如果仅有default配置:各个版本按default配置
8+; 3)如果仅有某些平台的配置,没有default配置:对应平台的按照配置的值传递,非对应平台的:非AscendC算子继续按空处理,AscendC算子按照 simplified_key_mode=0 处理
9+; 4)如果default配置和平台配置都有:对应平台的使用平台的配置,非对应的平台的以default值配置。
10+; 5)对于自定义simplified key的情况,需要在binary_simplified_key_mode.ini 文件中显式配置为None,不传入 --simplified_key_mode 选项,由opc工具和FE框架自行判断使用何种模式
11+; 6)是否是AscendC算子,由 ops/built-in/tbe/op_info_cfg/parser/ascendc_config.json 中配置的算子名字和对于的平台决定
12+[StatelessExponential]
13+default=0
@@ -62,6 +62,7 @@ static inline bool IsArch22(NpuArch arch) { return arch == NpuArch::DAV_2201; }
62 62 
63static inline bool IsArch35(NpuArch arch)63static inline bool IsArch35(NpuArch arch)
64{64{
65+ // Ascend960 (DAV_9201/9202) 与 Ascend950 共用 arch35 L0 实现路径
65 return arch == NpuArch::DAV_3510 || arch == NpuArch::DAV_9201 || arch == NpuArch::DAV_9202;66 return arch == NpuArch::DAV_3510 || arch == NpuArch::DAV_9201 || arch == NpuArch::DAV_9202;
66}67}
67 68