From 6b82e05e88c614b468053bc79c6bbecffcf4698d Mon Sep 17 00:00:00 2001 From: guopeian Date: Fri, 2 Jul 2021 15:48:01 +0800 Subject: [PATCH] fix --- ge/single_op/task/op_task.cc | 78 +++++++++++++++++++++++------------- ge/single_op/task/op_task.h | 32 +++++++++------ 2 files changed, 69 insertions(+), 41 deletions(-) diff --git a/ge/single_op/task/op_task.cc b/ge/single_op/task/op_task.cc index 2a5a4d48..5345f0f7 100755 --- a/ge/single_op/task/op_task.cc +++ b/ge/single_op/task/op_task.cc @@ -993,34 +993,56 @@ Status AiCpuCCTask::InitForSummaryAndCopy() { return SUCCESS; } -Status AiCpuCCTask::SetMemCopyTask(const domi::KernelExDef &kernel_def) { - if (kernel_def.args_size() > sizeof(STR_FWK_OP_KERNEL)) { - GELOGE(ACL_ERROR_GE_PARAM_INVALID, "[Check][Size]sizeof STR_FWK_OP_KERNEL is: %lu, but args_size is: %d", - sizeof(STR_FWK_OP_KERNEL), kernel_def.args_size()); - REPORT_INNER_ERROR("E19999", "[sizeof STR_FWK_OP_KERNEL is: %lu, but args_size is: %d", - sizeof(STR_FWK_OP_KERNEL), kernel_def.args_size()); - return ACL_ERROR_GE_PARAM_INVALID; - } - GE_CHK_RT_RET(rtMalloc(©_workspace_buf_, kernel_def.task_info_size(), RT_MEMORY_HBM)); - GE_CHK_RT_RET(rtMemcpy(copy_workspace_buf_, kernel_def.task_info_size(), - kernel_def.task_info().data(), kernel_def.task_info_size(), RT_MEMCPY_HOST_TO_DEVICE)); - - STR_FWK_OP_KERNEL aicpu_task = {0}; - auto sec_ret = memcpy_s(&aicpu_task, sizeof(STR_FWK_OP_KERNEL), - kernel_def.args().data(), kernel_def.args().size()); - if (sec_ret != EOK) { - GELOGE(ACL_ERROR_GE_MEMORY_OPERATE_FAILED, "[Update][TaskArgs] failed, ret: %d", sec_ret); - REPORT_INNER_ERROR("E19999", "update STR_FWK_OP_KERNEL args failed because memcpy_s return %d.", sec_ret); - return ACL_ERROR_GE_MEMORY_OPERATE_FAILED; - } - - aicpu_task.fwkKernelBase.fwk_kernel.inputOutputAddr = reinterpret_cast(copy_ioaddr_dev_); - aicpu_task.fwkKernelBase.fwk_kernel.workspaceBaseAddr = reinterpret_cast(copy_workspace_buf_); - aicpu_task.fwkKernelBase.fwk_kernel.extInfoAddr = 0; - aicpu_task.fwkKernelBase.fwk_kernel.extInfoLen = 0; - - GE_CHK_RT_RET(rtMemcpy(copy_task_args_buf_, sizeof(STR_FWK_OP_KERNEL), - &aicpu_task, sizeof(STR_FWK_OP_KERNEL), RT_MEMCPY_HOST_TO_DEVICE)); +Status AiCpuCCTask::SetMemCopyTask(const domi::KernelDef &kernel_def) { + auto &memcpy_args = kernel_def.args(); + memcpy_args_size_ = kernel_def.args_size(); + memcpy_so_name_ = kernel_def.so_name(); + memcpy_kernel_name_ = kernel_def.kernel_name(); + GE_IF_BOOL_EXEC(memcpy_args.size() != memcpy_args_size_, + REPORT_INNER_ERROR("E19999", "MemCopy task def args.size=%zu, but args_size=%u not equal.", + memcpy_args.size(), memcpy_args_size_); + GELOGE(FAILED, "[Check][Size]MemCopy task def args.size=%zu, but args_size=%u not equal.", + memcpy_args.size(), memcpy_args_size_); + return FAILED;); + GE_IF_BOOL_EXEC(memcpy_args_size_ < sizeof(aicpu::AicpuParamHead), + REPORT_INNER_ERROR("E19999", + "Task def args_size=%u is less than aicpu param head len=%zu.", + memcpy_args_size_, sizeof(aicpu::AicpuParamHead)); + GELOGE(FAILED, + "[Check][Size] Task def args_size=%u is less than aicpu param head len=%zu.", + memcpy_args_size_, sizeof(aicpu::AicpuParamHead)); + return FAILED;); + + memcpy_args_.reset(new(std::nothrow) uint8_t[memcpy_args_size_]()); + GE_IF_BOOL_EXEC(memcpy_args_ == nullptr, + REPORT_INNER_ERROR("E19999", "new memory failed for Node[MemCopy], task_size[%u].", + memcpy_args_size_); + GELOGE(FAILED, "[Malloc][Memory] failed for Node[MemCopy], task_size[%u].", + memcpy_args_size_); + return FAILED;); + + errno_t sec_ret = memcpy_s(memcpy_args_.get(), memcpy_args_size_, memcpy_args.c_str(), memcpy_args.size()); + GE_IF_BOOL_EXEC(sec_ret != EOK, + REPORT_INNER_ERROR("E19999", + "memcpy_s argc_ failed for Node[MemCopy], ret: %d", sec_ret); + GELOGE(INTERNAL_ERROR, + "[Update][args] failed for Node[MemCopy], ret: %d", sec_ret); + return sec_ret;); + auto memcpy_param_head = reinterpret_cast(memcpy_args_.get()); + uint32_t memcpy_io_num = memcpy_param_head->ioAddrNum; + auto memcpy_io_addr = memcpy_args_.get() + sizeof(aicpu::AicpuParamHead); + // if has input and output, need copy to ioaddr + int cpy_ret = memcpy_s(memcpy_io_addr, memcpy_args_size_ - sizeof(aicpu::AicpuParamHead), + ©_ioaddr_dev_, sizeof(uint64_t) * memcpy_io_num); + GE_IF_BOOL_EXEC(cpy_ret != 0, + REPORT_INNER_ERROR("E19999", "Node[Memcpoy] memcpy io addr to AicpuParamHead failed," + "ret=%d, args_size=%u, io nums=%u.", + cpy_ret, memcpy_args_size_, memcpy_io_num); + GELOGE(INTERNAL_ERROR, "[Update][io_addr]Node[MemCopy] memcpy io addr to AicpuParamHead failed," + "ret=%d, args_size=%u, io nums=%u.", + cpy_ret, memcpy_args_size_, memcpy_io_num); + return INTERNAL_ERROR;); + GELOGD("Set memcpy task for node[MemCopy] successfully."); return SUCCESS; } diff --git a/ge/single_op/task/op_task.h b/ge/single_op/task/op_task.h index 2d2e6731..84eb9d00 100644 --- a/ge/single_op/task/op_task.h +++ b/ge/single_op/task/op_task.h @@ -148,6 +148,18 @@ class AiCpuBaseTask : public OpTask { std::unique_ptr aicpu_ext_handle_; void *ext_info_addr_dev_ = nullptr; vector input_is_const_; + + std::vector output_summary_; + std::vector output_summary_host_; + + void *copy_ioaddr_dev_ = nullptr; + + void *copy_input_release_flag_dev_ = nullptr; + void *copy_input_data_size_dev_ = nullptr; + void *copy_input_src_dev_ = nullptr; + void *copy_input_dst_dev_ = nullptr; + + vector out_shape_hbm_; }; class AiCpuTask : public AiCpuBaseTask { @@ -195,18 +207,6 @@ class AiCpuTask : public AiCpuBaseTask { // for copy task void *copy_task_args_buf_ = nullptr; void *copy_workspace_buf_ = nullptr; - - std::vector output_summary_; - std::vector output_summary_host_; - - void *copy_ioaddr_dev_ = nullptr; - - void *copy_input_release_flag_dev_ = nullptr; - void *copy_input_data_size_dev_ = nullptr; - void *copy_input_src_dev_ = nullptr; - void *copy_input_dst_dev_ = nullptr; - - vector out_shape_hbm_; uint64_t kernel_id_ = 0; }; @@ -225,7 +225,7 @@ class AiCpuCCTask : public AiCpuBaseTask { void SetkernelName(const std::string &kernel_Name); void SetIoAddr(uintptr_t *io_addr); size_t GetArgSize() const; - Status SetMemCopyTask(const domi::KernelExDef &kernel_def); + Status SetMemCopyTask(const domi::KernelDef &kernel_def); Status LaunchKernel(const std::vector &input_desc, const std::vector &input_buffers, std::vector &output_desc, @@ -246,6 +246,12 @@ class AiCpuCCTask : public AiCpuBaseTask { uint32_t dump_flag_ = RT_KERNEL_DEFAULT; std::string op_type_; uint64_t kernel_id_ = 0; + // host memcpy mem + std::unique_ptr memcpy_args_; + std::string memcpy_so_name_; + std::string memcpy_kernel_name_; + // args size + size_t memcpy_args_size_ = 0; }; class MemcpyAsyncTask : public OpTask {