1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

# Ref: https://gist.github.com/mizanRahman/40ba603759bfb5153189ccdc9dbbd1e4

# Disable TCP slow start on idle connections
net.ipv4.tcp_slow_start_after_idle = 0

# Increase Linux autotuning TCP buffer limits
# Set max to 16MB for 1GE and 32M (33554432) or 54M (56623104) for 10GE
# Don't set tcp_mem itself! Let the kernel scale it based on RAM.
net.core.rmem_max = 56623104
net.core.wmem_max = 56623104
net.core.rmem_default = 56623104
net.core.wmem_default = 56623104
net.core.optmem_max = 40960
net.ipv4.tcp_rmem = 4096 87380 56623104
net.ipv4.tcp_wmem = 4096 65536 56623104

# TCP Congestion Control
net.ipv4.tcp_congestion_control = bbr
net.core.default_qdisc = cake

1
2

[nfsd]
threads=128

1

options sunrpc tcp_slot_table_entries=16384

1

fs.leases-enable = 0

1

[1073511.138061] NFSD: starting 90-second grace period (net f0000000)

1
2

# Options for rpc.svcgssd.
RPCSVCGSSDOPTS="--lease-time 10 --grace-time 10"

1
2
3
4
5

[Peer]
PublicKey = <redacted>
AllowedIPs = 10.1.x.y/32
AllowedIPs = fd01::x:y/128
PersistentKeepalive = 25

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

[Interface]
PrivateKey = <redacted>
Address = 10.1.x.y/16
Address = fd01::x:y/64

[Peer]
PublicKey = <redacted>
AllowedIPs = 10.1.0.0/16  # route all VPN traffic to gateway server
AllowedIPs = fd01::/64
Endpoint = wg.ustcaigroup.xyz:51820  # gateway server is dual stack
# Endpoint = wg.ustcaigroup.xyz:51820  # IPv4
# Endpoint = wg.ustcaigroup.xyz:51820  # IPv6
PersistentKeepalive = 25

1
2

custom:
  - func: npu_add_custom(Tensor x, Tensor y) -> Tensor  # 添加的函数

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

#include <torch/csrc/autograd/custom_function.h>

#include "torch_npu/csrc/framework/utils/OpAdapter.h"
#include "torch_npu/csrc/aten/NPUNativeFunctions.h"
#include "torch_npu/csrc/aten/ops/op_api/op_api_common.h"

namespace at_npu {
  namespace native {
    using torch::autograd::Function;
    using torch::autograd::AutogradContext;

    at::Tensor NPUNativeFunctions::npu_add_custom(const at::Tensor& x, const at::Tensor& y) {
        at::Tensor result = OpPreparation::ApplyTensor(x); // 创建输出内存

        // calculate the output result of the NPU
        EXEC_NPU_CMD(aclnnAddCustom, x, y, result);
        return result;
    }
  } // namespace native
} // namespace at_npu

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40

[
    {
        "op": "AddCustom",
        "language": "cpp",
        "input_desc": [
            {
                "name": "x",
                "param_type": "required",
                "format": [
                    "ND"
                ],
                "type": [
                    "fp16"
                ]
            },
            {
                "name": "y",
                "param_type": "required",
                "format": [
                    "ND"
                ],
                "type": [
                    "fp16"
                ]
            }
        ],
        "output_desc": [
            {
                "name": "z",
                "param_type": "required",
                "format": [
                    "ND"
                ],
                "type": [
                    "fp16"
                ]
            }
        ]
    }
]

1

msopgen gen -i add_custom.json -c ai_core-Ascend910B3 -f pytorch -out . -lan cpp

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19

AddCustom
├── build.sh
├── cmake 
│   ├── config.cmake
│   ├── func.cmake
│   ├── intf.cmake
│   ├── makeself.cmake
│   └── util
├── CMakeLists.txt
├── CMakePresets.json          // 修改 ASCEND_CANN_PACKAGE_PATH
├── framework
├── op_host
│   ├── add_custom_tiling.h    // 定义 length 和 tiling 相关信息
│   ├── add_custom.cpp         // 算子 host 侧实现
│   ├── CMakeLists.txt
├── op_kernel
│   ├── CMakeLists.txt
│   ├── add_custom.cpp         // 算子 kernel 侧实现
└── scripts

1
2
3
4
5
6
7
8
9

#include "register/tilingdata_base.h"

namespace optiling {
BEGIN_TILING_DATA_DEF(AddCustomTilingData)
    TILING_DATA_FIELD_DEF(uint32_t, size);  // 定义 tensor size
END_TILING_DATA_DEF;

REGISTER_TILING_DATA_CLASS(AddCustom, AddCustomTilingData)
}

1

context->SetBlockDim(20); // 910B3 的 block_dim

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

#include "kernel_operator.h"

#ifdef __DAV_C220_VEC__

extern "C" __global__ __aicore__ void add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR workspace, GM_ADDR tiling) {
    GET_TILING_DATA(tiling_data, tiling);
    uint32_t M = tiling_data.size;  // 从 tiling_data 中获取 tensor size

    // ...
}

#else

// 重要：CANN 会尝试不同的 ccec 编译参数以推断算子的类型（VEC、CUBE、MIXED），如果不创建一个 stub 函数将会编译失败
extern "C" __global__ __aicore__ void add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z, GM_ADDR workspace, GM_ADDR tiling) {
    pip_barrier(PIPE_ALL);
}

#endif

1
2

$ bash build.sh
$ ./custom_opp_euleros_aarch64.run

1
2
3
4
5
6

import torch
import torch_npu

# ...

z = torch.npu_add_custom(x, y)  # 由于是运行时编译，第一次运行时需要等待编译

1
2
3

int mkl_serv_intel_cpu_true() {
    return 1;
}

1
2
3

export MKL_DEBUG_CPU_TYPE=5  # 兼容旧版本 MKL
export MKL_ENABLE_INSTRUCTIONS=AVX2  # 可选，指明 MKL 可以使用 AVX2
export LD_PRELOAD=/usr/local/lib/libmkl_trick.so