华为计算微信公众号
昇腾AI开发者公众号
华为计算微博
华为计算今日头条
CANN版本:8.0.RC1
现象:使用DataCopy从C1拷贝到C2报错。
我想分割bias矢量。
INFO: compile op on cpu succeed! [INFO] ATRACE(3370045,matmul_custom_cpu):2024-08-19-14:49:51.026.013 [trace_attr.c:164](tid:3370045) attr init success, timeout=0, driver version=467731. [INFO] ATRACE(3370045,matmul_custom_cpu):2024-08-19-14:49:51.026.191 [trace_recorder.c:114](tid:3370045) use root path: /root/ascend [EVENT] PROFILING(3370045,matmul_custom_cpu):2024-08-19-14:49:51.081.960 [msprof_callback_impl.cpp:334] >>> (tid:3370045) Started to register profiling ctrl callback. [EVENT] PROFILING(3370045,matmul_custom_cpu):2024-08-19-14:49:51.082.100 [msprof_callback_impl.cpp:341] >>> (tid:3370045) Started to register profiling hash id callback. [INFO] PROFILING(3370045,matmul_custom_cpu):2024-08-19-14:49:51.082.111 [prof_atls_plugin.cpp:83] >>> (tid:3370045) RegisterProfileCallback, callback type is 7 [EVENT] PROFILING(3370045,matmul_custom_cpu):2024-08-19-14:49:51.082.118 [msprof_callback_impl.cpp:348] >>> (tid:3370045) Started to register profiling enable host freq callback. [INFO] PROFILING(3370045,matmul_custom_cpu):2024-08-19-14:49:51.082.125 [prof_atls_plugin.cpp:83] >>> (tid:3370045) RegisterProfileCallback, callback type is 8 [ERROR][Core_0][/usr/local/Ascend/ascend-toolkit/8.0.RC1/aarch64-linux/tikcpp/tikcfw/inner_interface/inner_kernel_operator_data_copy_intf.cppm:203][DataCopy][3370047] unsupported dst hardware pos [ERROR][CORE_0][pid 3370047] error happened! =========
代码
__aicore__ inline void SplitBias(const LocalTensor<float>& bias1Local, const int bSplitIdx) { LocalTensor<float> bias2Local = inQueueB2.AllocTensor<float>(); DataCopyParams dataCopyParams; dataCopyParams.blockCount = 1; dataCopyParams.blockLen = 2; DataCopy(co1Local, bias1Local[bSplitIdx * biasSize / 2], dataCopyParams); inQueueC2.EnQue<float>(bias2Local); }
我要发帖子
CANN版本:8.0.RC1
现象:使用DataCopy从C1拷贝到C2报错。
我想分割bias矢量。
代码
__aicore__ inline void SplitBias(const LocalTensor<float>& bias1Local, const int bSplitIdx) { LocalTensor<float> bias2Local = inQueueB2.AllocTensor<float>(); DataCopyParams dataCopyParams; dataCopyParams.blockCount = 1; dataCopyParams.blockLen = 2; DataCopy(co1Local, bias1Local[bSplitIdx * biasSize / 2], dataCopyParams); inQueueC2.EnQue<float>(bias2Local); }