Closed
[Documentation|文档反馈]: foreach 八接口对「x与out的shape不满足约束」给出 161002/561002 两种码,RoundOffNumber V1 与 V2 互相矛盾(Sqrt 实测方向相反) #8009
liuhanwen8Add on  19 days agoClosed at  8 days ago
liuhanwen8
liuhanwen8
19 days ago Create

现象

foreach 族的同一条语义「x 与 out 的 shape 不满足约束」,在八个接口的资料里被归到两种不同的错误码下,而同机同构造的实测又给出第三种组合。调用方按资料做错误码分流(161002=参数非法直接改参 / 561002=内部 tiling 失败可重试)会走错路。

同一构造 x=fp32[2x3]ND|out=fp32[2x2]ND(out 比 x 小,同 dtype、同 rank)在 910B2 上逐接口实测:

接口 资料把该子项归在 实测第一段返回 是否一致
aclnnForeachReciprocal 161002 561002 不符
aclnnForeachSin 161002 561002 不符
aclnnForeachSign 161002 561002 不符
aclnnForeachSigmoid 561002 561002 一致
aclnnForeachSinh 561002 561002 一致
aclnnForeachSqrt 561002 161002 不符(方向相反)
aclnnForeachRoundOffNumber(V1) 561002 561002 一致
aclnnForeachRoundOffNumberV2 161002 161002 一致,但与 V1 矛盾

三个层面的不一致

1. 同一条语义,两种码

资料侧:Reciprocal / Sin / Sign / RoundOffNumberV2 归 161002;Sigmoid / Sinh / Sqrt / RoundOffNumber(V1) 归 561002。

2. 同一个算子,两个入口互相矛盾

aclnnForeachRoundOffNumber(V1)与 aclnnForeachRoundOffNumberV2 是同一个 l0 kernel(l0op::ForeachRoundOffNumberV2)的两个 aclnn 入口,实测分别是 561002 与 161002,各自的资料也分别写成 561002 与 161002 —— 即文档跟着实现走,但两个实现本身不一致。

3. 实现侧与资料都不对齐

八个接口的 CheckShape 失败一律映射到 ACLNN_ERR_PARAM_INVALID(161002):

// 例:foreach/foreach_sqrt/op_host/op_api/aclnn_foreach_sqrt.cpp:127-129
for (uint64_t i = 0; i < self->Size(); i++) {
    OP_CHECK_SHAPE_NOT_EQUAL((*self)[i], (*out)[i], return false);
}
...
CHECK_RET(CheckShape(self, out), ACLNN_ERR_PARAM_INVALID);   // 统一 PARAM_INVALID

但实测有 6 个接口返 561002,说明 shape 违规在 host 层没被拦住,落到了 tiling 层上报 —— 与资料写 161002 的那几个接口(Reciprocal / Sin / Sign)直接冲突。

期望

统一该子项的错误码归属,并让资料与实现一致。建议口径(与 CheckShape 的实际位置一致):

  • host 层可判的 shape 违规(成员 shape 不相等、rank > 8)→ 161002
  • 只有 tiling 层才能判的(如列表元素总数超限)→ 561002

若采纳该口径,需要同时修正:Reciprocal / Sin / Sign 的资料(改 161002→561002,或补 host 层拦截),以及 Sqrt 的实现(补 host 层 shape 校验,使其返 161002),并统一 V1/V2。

同一族此前已有多条同类 issue:#7945(aclnnForeachAsin)、#7946(aclnnForeachAtan)、#7967(aclnnForeachLerpScalar)、#7929(aclnnForeachAddScalar)。本条不重复它们覆盖的接口,只补 Reciprocal / Sin / Sign / Sigmoid / Sinh / Sqrt / RoundOffNumber / V2 这八个。

环境

  • 硬件:Ascend 910B2(Atlas A2)单卡,Health=OK
  • CANN:9.2.0(opp/version.info → Version=9.2.0,文件时间 2026-10-20)
  • 驱动:npu-smi 26.1.1
  • 仓库:cann/ops-nn master 12ad726c(2026-11-03 fetch)
  • OS:Linux aarch64;g++ -std=c++17

复现步骤

  1. 把下方代码存成 probe.cpp。
  2. 编译:
    g++ -O1 -std=c++17 -o probe probe.cpp \
        -I${ASCEND_HOME_PATH}/include -I${ASCEND_HOME_PATH}/include/aclnnop \
        -L${ASCEND_HOME_PATH}/lib64 -lascendcl -lnnopbase -lopapi -lpthread
    
  3. 逐接口跑(第 2 个参数是接口短名):
    ./probe 7 Reciprocal
    ./probe 7 Sin
    ./probe 7 Sign
    ./probe 7 Sigmoid
    ./probe 7 Sinh
    ./probe 7 Sqrt
    
  4. 各条打印第一段返回码,与上表比对。

复现脚本

实测输出(A2 / CANN 9.2.0 / 12ad726c,device 2,同构造 x=fp32[2x3]ND|out=fp32[2x2]ND):

[Reciprocal]     expect=161002  actual=561002  MISMATCH
[Sin]            expect=161002  actual=561002  MISMATCH
[Sign]           expect=161002  actual=561002  MISMATCH
[Sigmoid]        expect=561002  actual=561002  consistent
[Sinh]           expect=561002  actual=561002  consistent
[Sqrt]           expect=561002  actual=161002  MISMATCH

(RoundOffNumber 族需额外构造 roundMode,实测 V1=561002、V2=161002。)

// probe.cpp —— foreach 族八接口:out shape 小于 x 时的返回码
// 用法: ./probe <device_id> <ShortName>
//   ShortName ∈ {Reciprocal, Sin, Sign, Sigmoid, Sinh, Sqrt}
#include <acl/acl.h>
#include <aclnnop/aclnn_foreach_reciprocal.h>
#include <aclnnop/aclnn_foreach_sigmoid.h>
#include <aclnnop/aclnn_foreach_sign.h>
#include <aclnnop/aclnn_foreach_sin.h>
#include <aclnnop/aclnn_foreach_sinh.h>
#include <aclnnop/aclnn_foreach_sqrt.h>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <string>
#include <vector>

typedef aclnnStatus (*GwsFn)(const aclTensorList *, aclTensorList *, uint64_t *, aclOpExecutor **);

int main(int argc, char **argv)
{
    if (argc < 3) {
        printf("usage: %s <device_id> <ShortName>\n", argv[0]);
        return 2;
    }
    int dev = atoi(argv[1]);
    std::string name = argv[2];

    GwsFn fn = nullptr;
    int expect = 0;
    if (name == "Reciprocal") { fn = aclnnForeachReciprocalGetWorkspaceSize; expect = 161002; }
    else if (name == "Sin")   { fn = aclnnForeachSinGetWorkspaceSize;        expect = 161002; }
    else if (name == "Sign")  { fn = aclnnForeachSignGetWorkspaceSize;       expect = 161002; }
    else if (name == "Sigmoid") { fn = aclnnForeachSigmoidGetWorkspaceSize;  expect = 561002; }
    else if (name == "Sinh")  { fn = aclnnForeachSinhGetWorkspaceSize;       expect = 561002; }
    else if (name == "Sqrt")  { fn = aclnnForeachSqrtGetWorkspaceSize;       expect = 561002; }
    else { printf("unknown op %s\n", name.c_str()); return 2; }

    aclInit(nullptr);
    aclrtSetDevice(dev);
    aclrtStream stream = nullptr;
    aclrtCreateStream(&stream);

    // x = [2,3],out = [2,2](同 dtype / 同 rank,只有 shape 不满足)
    int64_t inDims[2] = {2, 3};
    int64_t outDims[2] = {2, 2};
    void *inDev = nullptr, *outDev = nullptr;
    aclrtMalloc(&inDev, 6 * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST);
    aclrtMalloc(&outDev, 4 * sizeof(float), ACL_MEM_MALLOC_HUGE_FIRST);
    std::vector<float> hostIn(6, 1.0f);
    aclrtMemcpy(inDev, 6 * sizeof(float), hostIn.data(), 6 * sizeof(float), ACL_MEMCPY_HOST_TO_DEVICE);

    aclTensor *xT = aclCreateTensor(inDims, 2, ACL_FLOAT, nullptr, 0, ACL_FORMAT_ND, inDims, 2, inDev);
    aclTensor *outT = aclCreateTensor(outDims, 2, ACL_FLOAT, nullptr, 0, ACL_FORMAT_ND, outDims, 2, outDev);
    aclTensorList *xList = aclCreateTensorList(&xT, 1);
    aclTensorList *outList = aclCreateTensorList(&outT, 1);

    uint64_t wsSize = 0;
    aclOpExecutor *executor = nullptr;
    int ret = fn(xList, outList, &wsSize, &executor);
    printf("[%s] expect=%d actual=%d %s\n", name.c_str(), expect, ret,
           (ret == expect) ? "consistent" : "MISMATCH");

    if (ret == 0) {
        void *ws = nullptr;
        if (wsSize > 0) { aclrtMalloc(&ws, wsSize, ACL_MEM_MALLOC_HUGE_FIRST); }
        aclrtFree(ws);
    }
    aclDestroyTensorList(xList);
    aclDestroyTensorList(outList);
    aclrtFree(inDev);
    aclrtFree(outDev);
    aclrtDestroyStream(stream);
    aclrtResetDevice(dev);
    aclFinalize();
    return (ret == expect) ? 0 : 1;
}
likedislike
liuhanwen8liuhanwen8
19 days ago Added label: documentation
yuning_chenyuning_chenCollaborators
18 days ago Assign tramp-ll as responsible person
yuning_chen
yuning_chenCollaborators
18 days ago Comments:

初步分析

感谢反馈,我们已开始核对这个问题。

现象理解:foreach 八个接口对「x 与 out 的 shape 不满足约束」给出的错误码不统一,161002 与 561002 两种都有;RoundOffNumber V1 与 V2 之间互相矛盾,Sqrt 实测方向还与资料相反。

当前判断:初步判断是各接口校验落点不一致(部分在首段 host 校验、部分落到第二段),导致对外错误码不统一,待逐接口复现确认。

涉及范围:foreach 同族八接口的 op_api shape 校验与对应 aclnn 资料的错误码表。

下一步:我们会逐个接口核对校验落点与资料声明,拉通成统一口径后再决定是改实现还是改资料。

likedislike
yuning_chen
yuning_chenCollaborators
18 days ago Comments:

您好,该问题已收到,当前问题 @tramp-ll 正在处理中

likedislike
yuning_chenyuning_chenCollaborators
11 days ago Assign east_yang as responsible person, remove original responsible person: tramp-ll
east_yangeast_yangCollaborators
11 days ago Assign tramp-ll as responsible person
east_yangeast_yangCollaborators
11 days ago Removed responsible person east_yang
tramp-ll
tramp-llCollaborators
10 days ago Comments:

当前最新代码中,只有aclnnForeachSqrt存在说明错误,已经修改md

likedislike
CANN-robotCANN-robotCollaborators
8 days ago Close issue
CANN-robotCANN-robotCollaborators
8 days ago changed custom state from 进行中 to 已完成
CANN-robotCANN-robotCollaborators
8 days ago Added label: resolved