Star1738
1647
Code
Code
Issues474
Pull Requests551
Actions
Discussion
Wiki
Analysis
Star1738
1647
Merged
feat(zero_op): aclnnInplaceZero支持UINT32/UINT64数据类型 #7016
fitZepHYrAdd on 7 days ago
feat(zero_op): aclnnInplaceZero支持UINT32/UINT64数据类型 #7016
Merged
fitZepHYrAdd on 7 days ago
Merged fromVoyageZhou/ops-math_forkimplace_zero_uint32_uint64_supportintocann/ops-mathmaster
Total: 4 files changed+57-16
Mmath/zero_op/docs/aclnnInplaceZero.md+1-1
@@ -124,7 +124,7 @@ aclnnStatus aclnnInplaceZero(
124- <term>Atlas推理系列产品</term>、<term>Atlas训练系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64。124- <term>Atlas推理系列产品</term>、<term>Atlas训练系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64。
125<!-- end id7 -->125<!-- end id7 -->
126<!-- npu="A3,910b" id8 -->126<!-- npu="A3,910b" id8 -->
127-- <term>Atlas A2系列产品</term>、<term>Atlas A3系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64、BFLOAT16。127+- <term>Atlas A2系列产品</term>、<term>Atlas A3系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64、BFLOAT16、UINT32、UINT64。
128<!-- end id8 -->128<!-- end id8 -->
129<!-- npu="950" id9 -->129<!-- npu="950" id9 -->
130- <term>Ascend 950PR&950DT系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64、BFLOAT16、FLOAT8_E5M2、FLOAT8_E4M3FN、HIFLOAT8、FLOAT4_E1M2、FLOAT4_E2M1。130- <term>Ascend 950PR&950DT系列产品</term>：数据类型支持INT8、INT32、INT64、UINT8、FLOAT16、FLOAT32、BOOL、DOUBLE、INT16、UINT16、COMPLEX128、COMPLEX64、BFLOAT16、FLOAT8_E5M2、FLOAT8_E4M3FN、HIFLOAT8、FLOAT4_E1M2、FLOAT4_E2M1。
Mmath/zero_op/docs/aclnnInplaceZero_en.md+1-1
@@ -126,7 +126,7 @@ aclnnStatus aclnnInplaceZero(
126- <term>Atlas inference products</term>, <term>Atlas training products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64.126- <term>Atlas inference products</term>, <term>Atlas training products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64.
127<!-- end id7 -->127<!-- end id7 -->
128<!-- npu="A3,910b" id8 -->128<!-- npu="A3,910b" id8 -->
129-- <term>Atlas A2 products</term>, <term>Atlas A3 products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64, BFLOAT16.129+- <term>Atlas A2 products</term>, <term>Atlas A3 products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64, BFLOAT16, UINT32, UINT64.
130<!-- end id8 -->130<!-- end id8 -->
131<!-- npu="950" id9 -->131<!-- npu="950" id9 -->
132- <term>Ascend 950PR&950DT products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64, BFLOAT16, FLOAT8_E5M2, FLOAT8_E4M3FN, HIFLOAT8, FLOAT4_E1M2, FLOAT4_E2M1.132- <term>Ascend 950PR&950DT products</term>: Supports the following data types: INT8, INT32, INT64, UINT8, FLOAT16, FLOAT32, BOOL, DOUBLE, INT16, UINT16, COMPLEX128, COMPLEX64, BFLOAT16, FLOAT8_E5M2, FLOAT8_E4M3FN, HIFLOAT8, FLOAT4_E1M2, FLOAT4_E2M1.
Mmath/zero_op/op_api/aclnn_zero.cpp+27-4
@@ -38,7 +38,7 @@ static const std::initializer_list<op::DataType> DTYPE_SUPPORT_LIST = {
38 op::DataType::DT_INT8, op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_UINT8,38 op::DataType::DT_INT8, op::DataType::DT_INT32, op::DataType::DT_INT64, op::DataType::DT_UINT8,
39 op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, op::DataType::DT_BOOL, op::DataType::DT_DOUBLE,39 op::DataType::DT_FLOAT16, op::DataType::DT_FLOAT, op::DataType::DT_BOOL, op::DataType::DT_DOUBLE,
40 op::DataType::DT_INT16, op::DataType::DT_UINT16, op::DataType::DT_COMPLEX128, op::DataType::DT_COMPLEX64,40 op::DataType::DT_INT16, op::DataType::DT_UINT16, op::DataType::DT_COMPLEX128, op::DataType::DT_COMPLEX64,
41- op::DataType::DT_BF16};41+ op::DataType::DT_BF16, op::DataType::DT_UINT32, op::DataType::DT_UINT64};
42 42 
43static const std::initializer_list<op::DataType> DTYPE_SUPPORT_950_LIST = {43static const std::initializer_list<op::DataType> DTYPE_SUPPORT_950_LIST = {
44 op::DataType::DT_INT8, op::DataType::DT_INT32, op::DataType::DT_INT64,44 op::DataType::DT_INT8, op::DataType::DT_INT32, op::DataType::DT_INT64,
@@ -88,6 +88,17 @@ static aclnnStatus CheckParams(const aclTensor* self)
88 return ACLNN_SUCCESS;88 return ACLNN_SUCCESS;
89}89}
90 90 
91+static void ViewDataType(const aclTensor* input, const op::DataType dtype)
92+{
93+ if (input == nullptr) {
94+ OP_LOGE(ACLNN_ERR_INNER_NULLPTR, "view failed: input tensor is null.");
95+ return;
96+ }
97+ auto tmpTensor = const_cast<aclTensor*>(input);
98+ tmpTensor->SetDataType(dtype);
99+ input = tmpTensor;
100+}
101+ 
91aclnnStatus aclnnInplaceZeroGetWorkspaceSize(aclTensor* selfRef, uint64_t* workspaceSize, aclOpExecutor** executor)102aclnnStatus aclnnInplaceZeroGetWorkspaceSize(aclTensor* selfRef, uint64_t* workspaceSize, aclOpExecutor** executor)
92{103{
93 L2_DFX_PHASE_1(aclnnInplaceZero, DFX_IN(selfRef), DFX_OUT(selfRef));104 L2_DFX_PHASE_1(aclnnInplaceZero, DFX_IN(selfRef), DFX_OUT(selfRef));
@@ -96,7 +107,7 @@ aclnnStatus aclnnInplaceZeroGetWorkspaceSize(aclTensor* selfRef, uint64_t* works
96 auto uniqueExecutor = CREATE_EXECUTOR();107 auto uniqueExecutor = CREATE_EXECUTOR();
97 CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);108 CHECK_RET(uniqueExecutor.get() != nullptr, ACLNN_ERR_INNER_CREATE_EXECUTOR);
98 109 
99- // 固定写法，参数检查110+ // 固定写法，参数检查（包含空指针与dtype校验）
100 auto ret = CheckParams(selfRef);111 auto ret = CheckParams(selfRef);
101 CHECK_RET(ret == ACLNN_SUCCESS, ret);112 CHECK_RET(ret == ACLNN_SUCCESS, ret);
102 113 
@@ -107,8 +118,18 @@ aclnnStatus aclnnInplaceZeroGetWorkspaceSize(aclTensor* selfRef, uint64_t* works
107 return ACLNN_SUCCESS;118 return ACLNN_SUCCESS;
108 }119 }
109 120 
121+ // uint32/uint64 复用 int32/int64 的 ZerosLike 实现，通过视图修改类型，不改变入参 selfRef 本身
122+ const aclTensor* self = selfRef;
123+ auto selfDtype = self->GetDataType();
124+ if (selfDtype == op::DataType::DT_UINT64 || selfDtype == op::DataType::DT_UINT32) {
125+ self = uniqueExecutor->CreateView(selfRef, selfRef->GetViewShape(), selfRef->GetStorageShape(),
126+ selfRef->GetViewStrides(), selfRef->GetViewOffset());
127+ CHECK_RET(self != nullptr, ACLNN_ERR_INNER_NULLPTR);
128+ ViewDataType(self, selfDtype == op::DataType::DT_UINT64 ? op::DataType::DT_INT64 : op::DataType::DT_INT32);
129+ }
130+ 
110 // 固定写法，将输入self转换成连续的tensor131 // 固定写法，将输入self转换成连续的tensor
111- auto selfContiguous = l0op::Contiguous(selfRef, uniqueExecutor.get());132+ auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());
112 CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);133 CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
113 134 
114 // 调用ZerosLike算子kernel135 // 调用ZerosLike算子kernel
@@ -116,7 +137,9 @@ aclnnStatus aclnnInplaceZeroGetWorkspaceSize(aclTensor* selfRef, uint64_t* works
116 CHECK_RET(zeroOut != nullptr, ACLNN_ERR_INNER_NULLPTR);137 CHECK_RET(zeroOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
117 138 
118 // 固定写法，将计算结果拷贝到输出out上，out可能是非连续的tensor139 // 固定写法，将计算结果拷贝到输出out上，out可能是非连续的tensor
119- auto viewcopyResult = l0op::ViewCopy(zeroOut, selfRef, uniqueExecutor.get());140+ // 注意：self和selfRef指向同一个内存地址，self用int32/int64解析，selfRef用uint32/uint64解析
141+ // 但计算结果zeroOut是int32/int64类型，需要通过ViewCopy拷贝到同为int32/int64的的self上
142+ auto viewcopyResult = l0op::ViewCopy(zeroOut, self, uniqueExecutor.get());
Y
Yyue-ma7 days ago

self和selfRef指向同一个内存地址，self用int32解析，selfRef用uint32解析，所以这里用self？

likedislike
fitZepHYr
fitZepHYr
7 days ago Comments：
120 CHECK_RET(viewcopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);143 CHECK_RET(viewcopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR);
121 144 
122 // 固定写法，获取计算过程中需要使用的workspace大小145 // 固定写法，获取计算过程中需要使用的workspace大小
Mmath/zero_op/tests/ut/op_api/test_aclnn_zero.cpp+28-10
@@ -15,21 +15,15 @@
15 15 
16class l2ZeroTest : public testing::Test {16class l2ZeroTest : public testing::Test {
17protected:17protected:
18- static void SetUpTestCase()18+ static void SetUpTestCase() { std::cout << "l2ZeroTest SetUp" << std::endl; }
19- {
20- std::cout << "l2ZeroTest SetUp" << std::endl;
21- }
22 19 
23- static void TearDownTestCase()20+ static void TearDownTestCase() { std::cout << "l2ZeroTest TearDown" << std::endl; }
24- {
25- std::cout << "l2ZeroTest TearDown" << std::endl;
26- }
27};21};
28 22 
29// self的数据类型不在支持范围内23// self的数据类型不在支持范围内
30TEST_F(l2ZeroTest, l2_zero_test_unsupported_type)24TEST_F(l2ZeroTest, l2_zero_test_unsupported_type)
31{25{
32- auto selfDesc = TensorDesc({2, 3}, ACL_UINT32, ACL_FORMAT_ND);26+ auto selfDesc = TensorDesc({2, 3}, ACL_COMPLEX32, ACL_FORMAT_ND);
33 27 
34 auto ut = OP_API_UT(aclnnInplaceZero, INPUT(selfDesc), OUTPUT());28 auto ut = OP_API_UT(aclnnInplaceZero, INPUT(selfDesc), OUTPUT());
35 29 
@@ -89,6 +83,30 @@ TEST_F(l2ZeroTest, l2_zero_test_int64)
89 // ut.TestPrecision();83 // ut.TestPrecision();
90}84}
91 85 
86+// 正常路径, uint32
87+TEST_F(l2ZeroTest, l2_zero_test_uint32)
88+{
89+ auto selfDesc = TensorDesc({2, 4}, ACL_UINT32, ACL_FORMAT_ND);
90+ 
91+ auto ut = OP_API_UT(aclnnInplaceZero, INPUT(selfDesc), OUTPUT());
92+ 
93+ uint64_t workspaceSize = 0;
94+ aclnnStatus getWorkspaceResult = ut.TestGetWorkspaceSize(&workspaceSize);
95+ EXPECT_EQ(getWorkspaceResult, ACLNN_SUCCESS);
96+}
97+ 
98+// 正常路径, uint64
99+TEST_F(l2ZeroTest, l2_zero_test_uint64)
100+{
101+ auto selfDesc = TensorDesc({2, 4}, ACL_UINT64, ACL_FORMAT_ND);
102+ 
103+ auto ut = OP_API_UT(aclnnInplaceZero, INPUT(selfDesc), OUTPUT());
104+ 
105+ uint64_t workspaceSize = 0;
106+ aclnnStatus getWorkspaceResult = ut.TestGetWorkspaceSize(&workspaceSize);
107+ EXPECT_EQ(getWorkspaceResult, ACLNN_SUCCESS);
108+}
109+ 
92// 正常路径, aicore, uint8110// 正常路径, aicore, uint8
93TEST_F(l2ZeroTest, l2_zero_test_uint8)111TEST_F(l2ZeroTest, l2_zero_test_uint8)
94{112{
@@ -239,4 +257,4 @@ TEST_F(l2ZeroTest, l2_zero_test_multi_dim_double)
239 uint64_t workspaceSize = 0;257 uint64_t workspaceSize = 0;
240 aclnnStatus getWorkspaceResult = ut.TestGetWorkspaceSize(&workspaceSize);258 aclnnStatus getWorkspaceResult = ut.TestGetWorkspaceSize(&workspaceSize);
241 EXPECT_EQ(getWorkspaceResult, ACLNN_SUCCESS);259 EXPECT_EQ(getWorkspaceResult, ACLNN_SUCCESS);
242-}260+}