* Copyright (c) 2026 Huawei Technologies Co., Ltd. All Rights Reserved.
* This program is free software, you can redistribute it and/or modify it under the terms and conditions of
* CANN Open Software License Agreement Version 2.0 (the "License").
* Please refer to the License for details. You may not use this file except in compliance with the License.
* THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
* See LICENSE in the root of the software repository for the full text of the License.
*/
#include <string>
#include "tensorflow/core/framework/op_kernel.h"
#include "acl/acl.h"
#include "tf_adapter/common/adapter_logger.h"
#include "tf_adapter/common/common.h"
#include "tf_adapter/util/acl_channel.h"
#include "tf_adapter/util/npu_attrs.h"
namespace tensorflow {
namespace {
class OutfeedEnqueueOp : public OpKernel {
public:
explicit OutfeedEnqueueOp(OpKernelConstruction *ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("channel_name", &channel_name_));
ADP_LOG(INFO) << "OutfeedEnqueueOp built";
}
~OutfeedEnqueueOp() override {
ADP_LOG(INFO) << "OutfeedEnqueueOp has been destructed";
}
void Compute(OpKernelContext *ctx) override {
(void)ctx;
ADP_LOG(INFO) << "OutfeedEnqueueOp running";
}
bool IsExpensive() override {
return false;
}
private:
std::string channel_name_;
};
class OutfeedDequeueOp : public OpKernel {
public:
explicit OutfeedDequeueOp(OpKernelConstruction *ctx) : OpKernel(ctx) {
OP_REQUIRES_OK(ctx, ctx->GetAttr("channel_name", &channel_name_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("output_types", &out_feed_dequeue_output_types_));
OP_REQUIRES_OK(ctx, ctx->GetAttr("output_shapes", &output_shapes_));
ADP_LOG(INFO) << "Start create acl channel for out-feed dequeue op " << channel_name_;
uint32_t device_id = 0;
OP_REQUIRES_OK(ctx, GetEnvDeviceID(device_id));
device_id_ = device_id;
const size_t kDefaultCapacity = 3;
OP_REQUIRES(ctx, aclrtSetDevice(static_cast<int32_t>(device_id_)) == ACL_SUCCESS,
errors::Internal("Acl rtSetDevice failed."));
acl_handle_ = CreateAclTdtRecvChannel(device_id_, channel_name_, kDefaultCapacity);
OP_REQUIRES(ctx, acl_handle_ != nullptr, errors::Internal("Acl create receive channel failed."));
ADP_LOG(INFO) << "Succeed create acl channel for out-feed dequeue op " << channel_name_;
}
~OutfeedDequeueOp() override {
ADP_LOG(INFO) << "Start destroy acl channel for out-feed dequeue op " << channel_name_;
if (acl_handle_ != nullptr) {
if (acltdtDestroyChannel(acl_handle_) != ACL_ERROR_NONE) {
ADP_LOG(ERROR) << "Failed destroy acl channel for out-feed dequeue op " << channel_name_;
} else {
acl_handle_ = nullptr;
ADP_LOG(INFO) << "Succeed destroy acl channel for out-feed dequeue op " << channel_name_;
}
}
if (aclrtResetDevice(static_cast<int32_t>(device_id_)) != ACL_SUCCESS) {
ADP_LOG(ERROR) << "Acl rtResetDevice failed.";
}
}
void Compute(OpKernelContext *ctx) override {
ADP_LOG(INFO) << "Start compute out-feed dequeue op " << channel_name_;
CancellationManager *cm = ctx->cancellation_manager();
CancellationToken token = cm->get_cancellation_token();
bool already_cancelled = !cm->RegisterCallback(token, [this]() {
ADP_LOG(INFO) << "Start run cancellation callback of out-feed dequeue op " << channel_name_;
Status ret = StopRecvTensorByAcl(&acl_handle_, channel_name_);
if (!ret.ok()) {
ADP_LOG(ERROR) << ret.error_message();
}
});
if (TF_PREDICT_FALSE(already_cancelled)) {
ctx->SetStatus(errors::Internal("out-feed op ", channel_name_, " called after cancelled."));
return;
}
std::vector<Tensor> tensors;
ADP_LOG(INFO) << "Start recv tensors by acl out-feed dequeue op " << channel_name_;
auto status = RecvTensorByAcl(acl_handle_, tensors);
ADP_LOG(INFO) << "Start de-register callback out-feed dequeue op " << channel_name_;
(void)cm->DeregisterCallback(token);
OP_REQUIRES_OK(ctx, status);
OP_REQUIRES(ctx, !tensors.empty(), errors::OutOfRange("out-feed op ", channel_name_, " received end-of-sequence"));
OP_REQUIRES(ctx, tensors.size() == output_shapes_.size(),
errors::Internal("out-feed op ", channel_name_, " received ", tensors.size(), " tensors but expect ",
output_shapes_.size(), " tensors"));
for (int i = 0; i < ctx->num_outputs(); ++i) {
ctx->set_output(i, tensors[static_cast<size_t>(i)]);
}
}
bool IsExpensive() override {
return false;
}
private:
DataTypeVector out_feed_dequeue_output_types_;
std::vector<PartialTensorShape> output_shapes_;
std::string channel_name_;
acltdtChannelHandle *acl_handle_ = nullptr;
uint32_t device_id_ = 0U;
};
REGISTER_KERNEL_BUILDER(Name("OutfeedDequeueOp").Device(DEVICE_CPU), OutfeedDequeueOp);
REGISTER_KERNEL_BUILDER(Name("OutfeedEnqueueOp").Device(DEVICE_CPU), OutfeedEnqueueOp);
}
}