已合并
aclnnReplicationPad1dBackward在大paddings场景路由aicpu配置（9.1.0） #3871
sunhao_hw创建于 7月7日
aclnnReplicationPad1dBackward在大paddings场景路由aicpu配置（9.1.0） #3871
已合并
共 1 个文件变更+57-49
| @@ -16,6 +16,8 @@ | |||
| 16 | 16 | ||
| 17 | 17 | ||
| 18 | 18 | ||
| 19 | |||
| 20 | |||
| 19 | 21 | ||
| 20 | using namespace op; | 22 | using namespace op; |
| 21 | 23 | ||
| @@ -23,13 +25,14 @@ extern "C" { | |||
| 23 | 25 | ||
| 24 | 26 | ||
| 25 | static const string REPLICATION_MODE = "edge"; | 27 | static const string REPLICATION_MODE = "edge"; |
| 28 | static const int64_t PADDING_FP32_MAX = 7200; | ||
| 26 | // 根据API定义，需要列出所能支持的所有dtype | 29 | // 根据API定义，需要列出所能支持的所有dtype |
| 27 | static const std::initializer_list<op::DataType> dtypeSupportList = { | 30 | static const std::initializer_list<op::DataType> dtypeSupportList = { |
| 28 | op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_BF16, | 31 | op::DataType::DT_FLOAT, op::DataType::DT_FLOAT16, op::DataType::DT_BF16, |
| 29 | op::DataType::DT_DOUBLE, op::DataType::DT_COMPLEX64, op::DataType::DT_COMPLEX128}; | 32 | op::DataType::DT_DOUBLE, op::DataType::DT_COMPLEX64, op::DataType::DT_COMPLEX128}; |
| 30 | 33 | ||
| 31 | inline static bool CheckNotNull(const aclTensor *gradOutput, const aclTensor *self, const aclIntArray *padding, | 34 | inline static bool CheckNotNull(const aclTensor* gradOutput, const aclTensor* self, const aclIntArray* padding, |
| 32 | const aclTensor *gradInput) | 35 | const aclTensor* gradInput) |
| 33 | { | 36 | { |
| 34 | OP_CHECK_NULL(gradOutput, return false); | 37 | OP_CHECK_NULL(gradOutput, return false); |
| 35 | OP_CHECK_NULL(self, return false); | 38 | OP_CHECK_NULL(self, return false); |
| @@ -38,7 +41,7 @@ inline static bool CheckNotNull(const aclTensor *gradOutput, const aclTensor *se | |||
| 38 | return true; | 41 | return true; |
| 39 | } | 42 | } |
| 40 | 43 | ||
| 41 | inline static bool CheckDtypeValid(const aclTensor *gradOutput, const aclTensor *self, const aclTensor *gradInput) | 44 | inline static bool CheckDtypeValid(const aclTensor* gradOutput, const aclTensor* self, const aclTensor* gradInput) |
| 42 | { | 45 | { |
| 43 | // 检查gradOutput的数据类型是否在支持列表内 | 46 | // 检查gradOutput的数据类型是否在支持列表内 |
| 44 | OP_CHECK_DTYPE_NOT_SUPPORT(gradOutput, dtypeSupportList, return false); | 47 | OP_CHECK_DTYPE_NOT_SUPPORT(gradOutput, dtypeSupportList, return false); |
| @@ -55,7 +58,7 @@ inline static bool CheckDtypeValid(const aclTensor *gradOutput, const aclTensor | |||
| 55 | return true; | 58 | return true; |
| 56 | } | 59 | } |
| 57 | 60 | ||
| 58 | inline static bool CheckFormat(const aclTensor *gradOutput, const aclTensor *self, const aclTensor *gradInput) | 61 | inline static bool CheckFormat(const aclTensor* gradOutput, const aclTensor* self, const aclTensor* gradInput) |
| 59 | { | 62 | { |
| 60 | // 如果输入格式是私有格式，记录日志，直接报错 | 63 | // 如果输入格式是私有格式，记录日志，直接报错 |
| 61 | if (op::IsPrivateFormat(gradOutput->GetStorageFormat()) || op::IsPrivateFormat(self->GetStorageFormat()) || | 64 | if (op::IsPrivateFormat(gradOutput->GetStorageFormat()) || op::IsPrivateFormat(self->GetStorageFormat()) || |
| @@ -64,8 +67,9 @@ inline static bool CheckFormat(const aclTensor *gradOutput, const aclTensor *sel | |||
| 64 | return false; | 67 | return false; |
| 65 | } | 68 | } |
| 66 | 69 | ||
| 67 | OP_CHECK(gradOutput->GetViewFormat() == self->GetViewFormat() && | 70 | OP_CHECK( |
| 68 | gradOutput->GetViewFormat() == gradInput->GetViewFormat(), | 71 | gradOutput->GetViewFormat() == self->GetViewFormat() && |
| 72 | gradOutput->GetViewFormat() == gradInput->GetViewFormat(), | ||
| 69 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, | 73 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, |
| 70 | "Format of input and output should be equal, gradOutput [%s], self [%s], gradInput [%s].", | 74 | "Format of input and output should be equal, gradOutput [%s], self [%s], gradInput [%s].", |
| 71 | op::ToString(gradOutput->GetViewFormat()).GetString(), op::ToString(self->GetViewFormat()).GetString(), | 75 | op::ToString(gradOutput->GetViewFormat()).GetString(), op::ToString(self->GetViewFormat()).GetString(), |
| @@ -74,8 +78,8 @@ inline static bool CheckFormat(const aclTensor *gradOutput, const aclTensor *sel | |||
| 74 | return true; | 78 | return true; |
| 75 | } | 79 | } |
| 76 | 80 | ||
| 77 | static bool CheckShape(const aclTensor *gradOutput, const aclTensor *self, const aclIntArray *padding, | 81 | static bool CheckShape(const aclTensor* gradOutput, const aclTensor* self, const aclIntArray* padding, |
| 78 | const aclTensor *gradInput) | 82 | const aclTensor* gradInput) |
| 79 | { | 83 | { |
| 80 | auto selfDimnum = self->GetViewShape().GetDimNum(); | 84 | auto selfDimnum = self->GetViewShape().GetDimNum(); |
| 81 | // self和gradInput的shape必须一致 | 85 | // self和gradInput的shape必须一致 |
| @@ -86,28 +90,25 @@ static bool CheckShape(const aclTensor *gradOutput, const aclTensor *self, const | |||
| 86 | OP_CHECK_MAX_DIM(self, 3, return false); | 90 | OP_CHECK_MAX_DIM(self, 3, return false); |
| 87 | 91 | ||
| 88 | // gradOutput, self, gradInput维度需要一致 | 92 | // gradOutput, self, gradInput维度需要一致 |
| 89 | OP_CHECK(gradOutput->GetViewShape().GetDimNum() == selfDimnum && | 93 | OP_CHECK( |
| 90 | gradInput->GetViewShape().GetDimNum() == selfDimnum, | 94 | gradOutput->GetViewShape().GetDimNum() == selfDimnum && gradInput->GetViewShape().GetDimNum() == selfDimnum, |
| 91 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "gradOutput, self, gradInput dim should be same."), | 95 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "gradOutput, self, gradInput dim should be same."), return false); |
| 92 | return false); | ||
| 93 | 96 | ||
| 94 | // padding长度为2 | 97 | // padding长度为2 |
| 95 | OP_CHECK(padding->Size() == 2, | 98 | OP_CHECK(padding->Size() == 2, |
| 96 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "padding length should be 2, but got %lu.", | 99 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "padding length should be 2, but got %lu.", padding->Size()), |
| 97 | padding->Size()), | ||
| 98 | return false); | 100 | return false); |
| 99 | 101 | ||
| 100 | // check the last dim value of gradOutput. 0, 1 are indexes | 102 | // check the last dim value of gradOutput. 0, 1 are indexes |
| 101 | OP_CHECK(gradOutput->GetViewShape().GetDim(selfDimnum - 1) == | 103 | OP_CHECK(gradOutput->GetViewShape().GetDim(selfDimnum - 1) == |
| 102 | self->GetViewShape().GetDim(selfDimnum - 1) + (*padding)[0] + (*padding)[1], | 104 | self->GetViewShape().GetDim(selfDimnum - 1) + (*padding)[0] + (*padding)[1], |
| 103 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "wrong gradOutput shape."), | 105 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "wrong gradOutput shape."), return false); |
| 104 | return false); | 106 | |
| 105 | |||
| 106 | return true; | 107 | return true; |
| 107 | } | 108 | } |
| 108 | 109 | ||
| 109 | inline static aclnnStatus CheckParams(const aclTensor *gradOutput, const aclTensor *self, | 110 | inline static aclnnStatus CheckParams(const aclTensor* gradOutput, const aclTensor* self, const aclIntArray* padding, |
| 110 | const aclIntArray *padding, const aclTensor *gradInput) | 111 | const aclTensor* gradInput) |
| 111 | { | 112 | { |
| 112 | // 1. 检查参数是否为空指针 | 113 | // 1. 检查参数是否为空指针 |
| 113 | CHECK_RET(CheckNotNull(gradOutput, self, padding, gradInput), ACLNN_ERR_PARAM_NULLPTR); | 114 | CHECK_RET(CheckNotNull(gradOutput, self, padding, gradInput), ACLNN_ERR_PARAM_NULLPTR); |
| @@ -124,18 +125,18 @@ inline static aclnnStatus CheckParams(const aclTensor *gradOutput, const aclTens | |||
| 124 | return ACLNN_SUCCESS; | 125 | return ACLNN_SUCCESS; |
| 125 | } | 126 | } |
| 126 | 127 | ||
| 127 | static const aclTensor *GetPaddingTensor(int64_t dim, const aclIntArray *padding, aclOpExecutor *executor) | 128 | static const aclTensor* GetPaddingTensor(int64_t dim, const aclIntArray* padding, aclOpExecutor* executor) |
| 128 | { | 129 | { |
| 129 | FVector<int64_t, op::MAX_DIM_NUM> paddingsVector; | 130 | FVector<int64_t, op::MAX_DIM_NUM> paddingsVector; |
| 130 | // 2 is the magnification | 131 | // 2 is the magnification |
| 131 | for (size_t i = 2 * dim; i > 0; i -= 2) { | 132 | for (size_t i = 2 * dim; i > 0; i -= 2) { |
| 132 | if (i <= (size_t)padding->Size()) { | 133 | if (i <= (size_t)padding->Size()) { |
| 133 | // 2 and 1 indicate the element of padding is put into paddingsVector from the back to the front | 134 | // 2 and 1 indicate the element of padding is put into paddingsVector from the back to the front |
| 134 | paddingsVector.emplace_back((*padding)[i - 2]); | 135 | paddingsVector.emplace_back((*padding)[i - 2]); |
| 135 | paddingsVector.emplace_back((*padding)[i - 1]); | 136 | paddingsVector.emplace_back((*padding)[i - 1]); |
| 136 | } else { | 137 | } else { |
| 137 | paddingsVector.emplace_back(0); | 138 | paddingsVector.emplace_back(0); |
| 138 | paddingsVector.emplace_back(0); | 139 | paddingsVector.emplace_back(0); |
| 139 | } | 140 | } |
| 140 | } | 141 | } |
| 141 | // 2 is the magnification | 142 | // 2 is the magnification |
| @@ -144,17 +145,24 @@ static const aclTensor *GetPaddingTensor(int64_t dim, const aclIntArray *padding | |||
| 144 | return paddingsTensor; | 145 | return paddingsTensor; |
| 145 | } | 146 | } |
| 146 | 147 | ||
| 147 | static bool CheckPaddingValue(const aclIntArray *padding) { | 148 | static bool CheckPaddingValue(const aclIntArray* padding, const aclTensor* gradOutput) |
| 149 | { | ||
| 148 | // padding的每一维度的数值要大于等于0 | 150 | // padding的每一维度的数值要大于等于0 |
| 149 | if ((*padding)[0] < 0 || (*padding)[1] < 0 ) { | 151 | if ((*padding)[0] < 0 || (*padding)[1] < 0) { |
| 150 | OP_LOGW("on aicore situation, padding values should be greater than 0 or equal 0."); | 152 | OP_LOGW("on aicore situation, padding values should be greater than 0 or equal 0."); |
| 151 | return false; | 153 | return false; |
| 152 | } | 154 | } |
| 155 | // fp32类型下，AtlasA2 padding最多不超过7200 | ||
| 156 | if (!IsRegBase() && gradOutput->GetDataType() == op::DataType::DT_FLOAT && | ||
| 157 | ((*padding)[0] >= PADDING_FP32_MAX || (*padding)[1] >= PADDING_FP32_MAX)) { | ||
 | |||
| 158 | OP_LOGW("on aicore situation, padding values should be less than 7200."); | ||
| 159 | return false; | ||
| 160 | } | ||
| 153 | return true; | 161 | return true; |
| 154 | } | 162 | } |
| 155 | 163 | ||
| 156 | static aclnnStatus InputPreprocess(const aclTensor *&gradOutput, const aclTensor *&self, const aclIntArray *dimArray, | 164 | static aclnnStatus InputPreprocess(const aclTensor*& gradOutput, const aclTensor*& self, const aclIntArray* dimArray, |
| 157 | int64_t dimCp, aclOpExecutor *executor) | 165 | int64_t dimCp, aclOpExecutor* executor) |
| 158 | { | 166 | { |
| 159 | // 如果非连续，需要转连续 | 167 | // 如果非连续，需要转连续 |
| 160 | gradOutput = l0op::Contiguous(gradOutput, executor); | 168 | gradOutput = l0op::Contiguous(gradOutput, executor); |
| @@ -173,11 +181,12 @@ static aclnnStatus InputPreprocess(const aclTensor *&gradOutput, const aclTensor | |||
| 173 | return ACLNN_SUCCESS; | 181 | return ACLNN_SUCCESS; |
| 174 | } | 182 | } |
| 175 | 183 | ||
| 176 | aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradOutput, const aclTensor *self, | 184 | aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor* gradOutput, const aclTensor* self, |
| 177 | const aclIntArray *padding, aclTensor *gradInput, uint64_t *workspaceSize, aclOpExecutor **executor) | 185 | const aclIntArray* padding, aclTensor* gradInput, |
| 186 | uint64_t* workspaceSize, aclOpExecutor** executor) | ||
| 178 | { | 187 | { |
| 179 | OP_CHECK_COMM_INPUT(workspaceSize, executor); | 188 | OP_CHECK_COMM_INPUT(workspaceSize, executor); |
| 180 | 189 | ||
| 181 | L2_DFX_PHASE_1(aclnnReplicationPad1dBackward, DFX_IN(gradOutput, self, padding), DFX_OUT(gradInput)); | 190 | L2_DFX_PHASE_1(aclnnReplicationPad1dBackward, DFX_IN(gradOutput, self, padding), DFX_OUT(gradInput)); |
| 182 | // 固定写法，创建OpExecutor | 191 | // 固定写法，创建OpExecutor |
| 183 | auto uniqueExecutor = CREATE_EXECUTOR(); | 192 | auto uniqueExecutor = CREATE_EXECUTOR(); |
| @@ -199,9 +208,9 @@ aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradO | |||
| 199 | if (self->GetViewShape().GetDimNum() == 3) { | 208 | if (self->GetViewShape().GetDimNum() == 3) { |
| 200 | // 1, 2 are indexes | 209 | // 1, 2 are indexes |
| 201 | if (self->GetViewShape().GetDim(1) == 0 || self->GetViewShape().GetDim(2) == 0) { | 210 | if (self->GetViewShape().GetDim(1) == 0 || self->GetViewShape().GetDim(2) == 0) { |
| 202 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Input should not be empty."); | 211 | OP_LOGE(ACLNN_ERR_PARAM_INVALID, "Input should not be empty."); |
| 203 | return ACLNN_ERR_PARAM_INVALID; | 212 | return ACLNN_ERR_PARAM_INVALID; |
| 204 | } | 213 | } |
| 205 | } | 214 | } |
| 206 | uniqueExecutor.ReleaseTo(executor); | 215 | uniqueExecutor.ReleaseTo(executor); |
| 207 | return ACLNN_SUCCESS; | 216 | return ACLNN_SUCCESS; |
| @@ -213,14 +222,14 @@ aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradO | |||
| 213 | // 0 is index | 222 | // 0 is index |
| 214 | const int64_t appendDim[] = {0}; | 223 | const int64_t appendDim[] = {0}; |
| 215 | // 1 is the dim num to be unsqueezed | 224 | // 1 is the dim num to be unsqueezed |
| 216 | aclIntArray *dimArray = (uniqueExecutor.get())->AllocIntArray(appendDim, 1); | 225 | aclIntArray* dimArray = (uniqueExecutor.get())->AllocIntArray(appendDim, 1); |
| 217 | ret = InputPreprocess(gradOutput, self, dimArray, dimCp, uniqueExecutor.get()); | 226 | ret = InputPreprocess(gradOutput, self, dimArray, dimCp, uniqueExecutor.get()); |
| 218 | CHECK_RET(ret == ACLNN_SUCCESS, ret); | 227 | CHECK_RET(ret == ACLNN_SUCCESS, ret); |
| 219 | 228 | ||
| 220 | dim = self->GetViewShape().GetDimNum(); | 229 | dim = self->GetViewShape().GetDimNum(); |
| 221 | auto paddingsTensor = GetPaddingTensor(dim, padding, uniqueExecutor.get()); | 230 | auto paddingsTensor = GetPaddingTensor(dim, padding, uniqueExecutor.get()); |
| 222 | auto padFlag = CheckPaddingValue(padding); | 231 | auto padFlag = CheckPaddingValue(padding, gradOutput); |
| 223 | const aclTensor *pad1dbackwardResult = nullptr; | 232 | const aclTensor* pad1dbackwardResult = nullptr; |
| 224 | auto originOutDataType = gradOutput->GetDataType(); | 233 | auto originOutDataType = gradOutput->GetDataType(); |
| 225 | // cast to fp32 from fp16 or bf16 | 234 | // cast to fp32 from fp16 or bf16 |
| 226 | if (padFlag && (originOutDataType == op::DataType::DT_FLOAT16 || originOutDataType == op::DataType::DT_BF16)) { | 235 | if (padFlag && (originOutDataType == op::DataType::DT_FLOAT16 || originOutDataType == op::DataType::DT_BF16)) { |
| @@ -228,8 +237,8 @@ aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradO | |||
| 228 | OP_LOGD("[PadV3Grad] FP16 or BF16 Cast to FP32: true"); | 237 | OP_LOGD("[PadV3Grad] FP16 or BF16 Cast to FP32: true"); |
| 229 | } | 238 | } |
| 230 | 239 | ||
| 231 | pad1dbackwardResult = l0op::PadV3Grad(gradOutput, paddingsTensor, | 240 | pad1dbackwardResult = l0op::PadV3Grad(gradOutput, paddingsTensor, REPLICATION_MODE, true, padFlag, |
| 232 | REPLICATION_MODE, true, padFlag, uniqueExecutor.get()); | 241 | uniqueExecutor.get()); |
| 233 | CHECK_RET(pad1dbackwardResult != nullptr, ACLNN_ERR_INNER_NULLPTR); | 242 | CHECK_RET(pad1dbackwardResult != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| 234 | pad1dbackwardResult = l0op::SqueezeNd(pad1dbackwardResult, dimArray, uniqueExecutor.get()); | 243 | pad1dbackwardResult = l0op::SqueezeNd(pad1dbackwardResult, dimArray, uniqueExecutor.get()); |
| 235 | // 2 is dim | 244 | // 2 is dim |
| @@ -237,8 +246,8 @@ aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradO | |||
| 237 | pad1dbackwardResult = l0op::SqueezeNd(pad1dbackwardResult, dimArray, uniqueExecutor.get()); | 246 | pad1dbackwardResult = l0op::SqueezeNd(pad1dbackwardResult, dimArray, uniqueExecutor.get()); |
| 238 | } | 247 | } |
| 239 | CHECK_RET(pad1dbackwardResult != nullptr, ACLNN_ERR_INNER_NULLPTR); | 248 | CHECK_RET(pad1dbackwardResult != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| 240 | 249 | ||
| 241 | // cast to fp16 or bf16 | 250 | // cast to fp16 or bf16 |
| 242 | if (padFlag && (originOutDataType == op::DataType::DT_FLOAT16 || originOutDataType == op::DataType::DT_BF16)) { | 251 | if (padFlag && (originOutDataType == op::DataType::DT_FLOAT16 || originOutDataType == op::DataType::DT_BF16)) { |
| 243 | pad1dbackwardResult = l0op::Cast(pad1dbackwardResult, originOutDataType, uniqueExecutor.get()); | 252 | pad1dbackwardResult = l0op::Cast(pad1dbackwardResult, originOutDataType, uniqueExecutor.get()); |
| 244 | OP_LOGD("[PadV3Grad] FP16 or BF16 Cast to FP32: true"); | 253 | OP_LOGD("[PadV3Grad] FP16 or BF16 Cast to FP32: true"); |
| @@ -252,16 +261,15 @@ aclnnStatus aclnnReplicationPad1dBackwardGetWorkspaceSize(const aclTensor *gradO | |||
| 252 | *workspaceSize = uniqueExecutor->GetWorkspaceSize(); | 261 | *workspaceSize = uniqueExecutor->GetWorkspaceSize(); |
| 253 | uniqueExecutor.ReleaseTo(executor); | 262 | uniqueExecutor.ReleaseTo(executor); |
| 254 | return ACLNN_SUCCESS; | 263 | return ACLNN_SUCCESS; |
| 255 | } | 264 | } |
| 256 | 265 | ||
| 257 | aclnnStatus aclnnReplicationPad1dBackward(void *workspace, uint64_t workspaceSize, aclOpExecutor *executor, | 266 | aclnnStatus aclnnReplicationPad1dBackward(void* workspace, uint64_t workspaceSize, aclOpExecutor* executor, |
| 258 | aclrtStream stream) | 267 | aclrtStream stream) |
| 259 | { | 268 | { |
| 260 | L2_DFX_PHASE_2(aclnnReplicationPad1dBackward); | 269 | L2_DFX_PHASE_2(aclnnReplicationPad1dBackward); |
| 261 | // 固定写法，调用框架能力，完成计算 | 270 | // 固定写法，调用框架能力，完成计算 |
| 262 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); | 271 | return CommonOpExecutorRun(workspace, workspaceSize, executor, stream); |
| 263 | } | 272 | } |
| 264 | |||
| 265 | 273 | ||
| 266 | } | 274 | } |
| 267 | #endif | 275 | #endif |
🟠 High Priority
变更后的
CheckPaddingValue（第 148-162 行）仅当gradOutput->GetDataType() == op::DataType::DT_FLOAT时检查 padding 阈值（>=7200），将 FP32 路由到 AICPU。但 FP16 和 BF16 输入在执行路径上同样会受影响：aclnnReplicationPad1dBackwardGetWorkspaceSize第 235-237 行：当padFlag == true且原始 dtype 为 FP16/BF16 时，gradOutput会被强制 Cast 为 FP32。l0op::PadV3Grad(..., padFlag=true)被调用时，gradOutput已是 FP32。PadV3Grad（padv3grad.cpp 第 283 行）中，IsPadV3GradReplicateAicoreSupport检查的是 Cast 后的 dtype（FP32），而 FP32 在REPLICATION_2D_AICORE_DTYPE_SUPPORT_LIST中，因此会走 AICORE 路径PadV3GradReplicateAiCore，以 FP32 数据 + 大 padding 执行 → 同样触发 507035 错误。触发条件：910B1(A2) 平台，FP16 或 BF16 输入，padding[0] 或 padding[1] ≥ 7200。
修复方向：
CheckPaddingValue中应将 FP16 和 BF16 也纳入检查，或检查范围覆盖所有会经 Cast 变为 FP32 后走 AICORE replication 路径的 dtype（即 FP32、FP16、BF16）。建议：将 dtype 检查从仅
DT_FLOAT扩展为同时包含DT_FLOAT16和DT_BF16，使这三种经 AICORE FP32 kernel 执行的 dtype 在大 padding 场景下均能正确路由到 AICPU。