Merged
fix(scatter_nd_update): 修复非连续回退路径的原始视图回写 #12073
陈展熹Add on Oct 16
fix(scatter_nd_update): 修复非连续回退路径的原始视图回写 #12073
Merged
Total: 3 files changed+163-5
| @@ -237,8 +237,9 @@ static bool IsSupportNonContiguous(const aclTensor* varRef, int64_t indexAxisNum | |||
| 237 | } | 237 | } |
| 238 | 238 | ||
| 239 | // 执行 ScatterNdUpdate 算子计算(公共实现) | 239 | // 执行 ScatterNdUpdate 算子计算(公共实现) |
| 240 | -static aclnnStatus ExecuteScatterNdUpdate(const aclTensor* varRef, const aclTensor* indices, const aclTensor* updates, | 240 | +static aclnnStatus ExecuteScatterNdUpdate(const aclTensor* varRefContiguous, const aclTensor* varRef, |
| 241 | - bool needViewCopy, uint64_t* workspaceSize, aclOpExecutor** executor, | 241 | + const aclTensor* indices, const aclTensor* updates, bool needViewCopy, |
| 242 | + uint64_t* workspaceSize, aclOpExecutor** executor, | ||
| 242 | UniqueExecutor& uniqueExecutor) | 243 | UniqueExecutor& uniqueExecutor) |
| 243 | { | 244 | { |
| 244 | // 将 indices 转换成连续的 tensor | 245 | // 将 indices 转换成连续的 tensor |
| @@ -250,11 +251,11 @@ static aclnnStatus ExecuteScatterNdUpdate(const aclTensor* varRef, const aclTens | |||
| 250 | CHECK_RET(updatesContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 251 | CHECK_RET(updatesContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| 251 | 252 | ||
| 252 | // 执行 L0 算子 | 253 | // 执行 L0 算子 |
| 253 | - auto scatterUpdateRes = l0op::ScatterNdUpdate(varRef, indicesContiguous, updatesContiguous, false, | 254 | + auto scatterUpdateRes = l0op::ScatterNdUpdate(varRefContiguous, indicesContiguous, updatesContiguous, false, |
| 254 | uniqueExecutor.get()); | 255 | uniqueExecutor.get()); |
| 255 | CHECK_RET(scatterUpdateRes != nullptr, ACLNN_ERR_INNER_NULLPTR); | 256 | CHECK_RET(scatterUpdateRes != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| 256 | 257 | ||
| 257 | - // 如果需要,将计算结果拷贝到输出 data 上 | 258 | + // 如果需要,按原始 varRef 的 stride 和 offset 写回连续副本的更新 |
| 258 | if (needViewCopy) { | 259 | if (needViewCopy) { |
| 259 | auto viewCopyResult = l0op::ViewCopy(scatterUpdateRes, varRef, uniqueExecutor.get()); | 260 | auto viewCopyResult = l0op::ViewCopy(scatterUpdateRes, varRef, uniqueExecutor.get()); |
| 260 | CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR); | 261 | CHECK_RET(viewCopyResult != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| @@ -303,7 +304,7 @@ static aclnnStatus ProcessContiguousCase(aclTensor* varRef, const aclTensor* ind | |||
| 303 | CHECK_RET(varRefContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); | 304 | CHECK_RET(varRefContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR); |
| 304 | 305 | ||
| 305 | bool needViewCopy = !IsContiguous(varRef); | 306 | bool needViewCopy = !IsContiguous(varRef); |
| 306 | - return ExecuteScatterNdUpdate(varRefContiguous, indices, updates, needViewCopy, workspaceSize, executor, | 307 | + return ExecuteScatterNdUpdate(varRefContiguous, varRef, indices, updates, needViewCopy, workspaceSize, executor, |
| 307 | uniqueExecutor); | 308 | uniqueExecutor); |
| 308 | } | 309 | } |
| 309 | 310 | ||
| @@ -0,0 +1,156 @@ | |||
| 1 | +/** | ||
| 2 | + * Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 3 | + * This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 4 | + * CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 5 | + * Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 6 | + * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 7 | + * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 8 | + * See LICENSE in the root of the software repository for the full text of the License. | ||
| 9 | + */ | ||
| 10 | + | ||
| 11 | +#include <dlfcn.h> | ||
| 12 | +#include <memory> | ||
| 13 | +#include <tuple> | ||
| 14 | +#include <vector> | ||
| 15 | + | ||
| 16 | +#include "gtest/gtest.h" | ||
| 17 | +#include "../../../op_api/aclnn_scatter_nd_update.h" | ||
| 18 | +#include "aclnn_kernels/contiguous.h" | ||
| 19 | +#include "op_api_ut_common/tensor_desc.h" | ||
| 20 | +#include "opdev/op_executor.h" | ||
| 21 | +#include "opdev/platform.h" | ||
| 22 | +#include "opdev/tensor_view_utils.h" | ||
| 23 | + | ||
| 24 | +namespace { | ||
| 25 | +struct WritebackState { | ||
| 26 | + const aclTensor* original = nullptr; | ||
| 27 | + const aclTensor* contiguous = nullptr; | ||
| 28 | + const aclTensor* source = nullptr; | ||
| 29 | + const aclTensor* destination = nullptr; | ||
| 30 | + size_t contiguousCalls = 0; | ||
| 31 | + size_t viewCopyCalls = 0; | ||
| 32 | +}; | ||
| 33 | + | ||
| 34 | +thread_local WritebackState* scatterNdWritebackState = nullptr; | ||
| 35 | +} // namespace | ||
| 36 | + | ||
| 37 | +// Intercept calls only while this fixture is active. Other tests must reach the | ||
| 38 | +// original implementation, including libopapi_math.so when preloaded by CI. | ||
| 39 | +namespace l0op { | ||
| 40 | +const aclTensor* Contiguous(const aclTensor* x, aclOpExecutor* executor) | ||
| 41 | +{ | ||
| 42 | + auto* state = scatterNdWritebackState; | ||
| 43 | + if (state == nullptr) { | ||
| 44 | + using Function = const aclTensor* (*)(const aclTensor*, aclOpExecutor*); | ||
| 45 | + static const auto original = reinterpret_cast<Function>( | ||
| 46 | + dlsym(RTLD_NEXT, "_ZN4l0op10ContiguousEPK9aclTensorP13aclOpExecutor")); | ||
| 47 | + EXPECT_NE(original, nullptr); | ||
| 48 | + return original == nullptr ? nullptr : original(x, executor); | ||
| 49 | + } | ||
| 50 | + if (x == state->original) { | ||
| 51 | + ++state->contiguousCalls; | ||
| 52 | + if (!op::IsContiguous(x)) { | ||
| 53 | + state->contiguous = executor->AllocTensor(x->GetViewShape(), x->GetDataType()); | ||
| 54 | + return state->contiguous; | ||
| 55 | + } | ||
| 56 | + } | ||
| 57 | + return x; | ||
| 58 | +} | ||
| 59 | + | ||
| 60 | +const aclTensor* ViewCopy(const aclTensor* x, const aclTensor* y, aclOpExecutor* executor) | ||
| 61 | +{ | ||
| 62 | + auto* state = scatterNdWritebackState; | ||
| 63 | + if (state == nullptr) { | ||
| 64 | + using Function = const aclTensor* (*)(const aclTensor*, const aclTensor*, aclOpExecutor*); | ||
| 65 | + static const auto original = reinterpret_cast<Function>( | ||
| 66 | + dlsym(RTLD_NEXT, "_ZN4l0op8ViewCopyEPK9aclTensorS2_P13aclOpExecutor")); | ||
| 67 | + EXPECT_NE(original, nullptr); | ||
| 68 | + return original == nullptr ? nullptr : original(x, y, executor); | ||
| 69 | + } | ||
| 70 | + ++state->viewCopyCalls; | ||
| 71 | + state->source = x; | ||
| 72 | + state->destination = y; | ||
| 73 | + return x; | ||
| 74 | +} | ||
| 75 | +} // namespace l0op | ||
| 76 | + | ||
| 77 | +namespace { | ||
| 78 | +struct WritebackLayout { | ||
| 79 | + std::vector<int64_t> strides; | ||
| 80 | + bool needsWriteback; | ||
| 81 | + bool callsContiguous; | ||
| 82 | +}; | ||
| 83 | + | ||
| 84 | +using WritebackParam = std::tuple<op::SocVersion, WritebackLayout, int64_t, aclDataType>; | ||
| 85 | + | ||
| 86 | +class ScatterNdUpdateWritebackTest : public testing::TestWithParam<WritebackParam> { | ||
| 87 | +protected: | ||
| 88 | + void SetUp() override | ||
| 89 | + { | ||
| 90 | + previousState_ = scatterNdWritebackState; | ||
| 91 | + scatterNdWritebackState = &state_; | ||
| 92 | + } | ||
| 93 | + | ||
| 94 | + void TearDown() override | ||
| 95 | + { | ||
| 96 | + executor_.reset(); | ||
| 97 | + scatterNdWritebackState = previousState_; | ||
| 98 | + } | ||
| 99 | + | ||
| 100 | + WritebackState state_; | ||
| 101 | + WritebackState* previousState_ = nullptr; | ||
| 102 | + std::unique_ptr<aclOpExecutor> executor_; | ||
| 103 | +}; | ||
| 104 | + | ||
| 105 | +TEST_P(ScatterNdUpdateWritebackTest, PreservesOriginalView) | ||
| 106 | +{ | ||
| 107 | + const auto& [soc, layout, offset, indexType] = GetParam(); | ||
| 108 | + op::SocVersionManager platform(soc); | ||
| 109 | + const auto expectedArch = soc == op::SocVersion::ASCEND950 ? NpuArch::DAV_3510 : NpuArch::DAV_2201; | ||
| 110 | + ASSERT_EQ(op::GetCurrentPlatformInfo().GetCurNpuArch(), expectedArch); | ||
| 111 | + | ||
| 112 | + auto var = TensorDesc({2, 3}, ACL_FLOAT, ACL_FORMAT_ND, layout.strides, offset, {32}).ToAclType(); | ||
| 113 | + auto indices = TensorDesc({1, 1}, indexType, ACL_FORMAT_ND).ToAclType(); | ||
| 114 | + auto updates = TensorDesc({1, 3}, ACL_FLOAT, ACL_FORMAT_ND).ToAclType(); | ||
| 115 | + ASSERT_NE(var.get(), nullptr); | ||
| 116 | + ASSERT_NE(indices.get(), nullptr); | ||
| 117 | + ASSERT_NE(updates.get(), nullptr); | ||
| 118 | + state_.original = var.get(); | ||
| 119 | + const auto originalShape = var->GetViewShape(); | ||
| 120 | + const auto originalStrides = var->GetViewStrides(); | ||
| 121 | + const auto originalStorageShape = var->GetStorageShape(); | ||
| 122 | + | ||
| 123 | + uint64_t workspaceSize = 0; | ||
| 124 | + aclOpExecutor* executor = nullptr; | ||
| 125 | + const auto status = aclnnScatterNdUpdateGetWorkspaceSize(var.get(), indices.get(), updates.get(), &workspaceSize, | ||
| 126 | + &executor); | ||
| 127 | + executor_.reset(executor); | ||
| 128 | + ASSERT_EQ(status, ACLNN_SUCCESS); | ||
| 129 | + ASSERT_NE(executor, nullptr); | ||
| 130 | + | ||
| 131 | + EXPECT_EQ(state_.contiguousCalls, layout.callsContiguous ? 1U : 0U); | ||
| 132 | + if (layout.needsWriteback) { | ||
| 133 | + ASSERT_NE(state_.contiguous, nullptr); | ||
| 134 | + ASSERT_NE(state_.contiguous, var.get()); | ||
| 135 | + EXPECT_TRUE(op::IsContiguous(state_.contiguous)); | ||
| 136 | + EXPECT_EQ(state_.viewCopyCalls, 1U); | ||
| 137 | + EXPECT_EQ(state_.source, state_.contiguous); | ||
| 138 | + EXPECT_EQ(state_.destination, var.get()); | ||
| 139 | + } else { | ||
| 140 | + EXPECT_EQ(state_.viewCopyCalls, 0U); | ||
| 141 | + EXPECT_EQ(state_.contiguous, nullptr); | ||
| 142 | + } | ||
| 143 | + EXPECT_EQ(var->GetViewShape(), originalShape); | ||
| 144 | + EXPECT_EQ(var->GetViewStrides(), originalStrides); | ||
| 145 | + EXPECT_EQ(var->GetStorageShape(), originalStorageShape); | ||
| 146 | + EXPECT_EQ(var->GetViewOffset(), offset); | ||
| 147 | +} | ||
| 148 | + | ||
| 149 | +INSTANTIATE_TEST_SUITE_P( | ||
| 150 | + PlatformsAndLayouts, ScatterNdUpdateWritebackTest, | ||
| 151 | + testing::Combine(testing::Values(op::SocVersion::ASCEND910B, op::SocVersion::ASCEND910_93, | ||
| 152 | + op::SocVersion::ASCEND950), | ||
| 153 | + testing::Values(WritebackLayout{{6, 2}, true, true}, WritebackLayout{{1, 2}, true, true}, | ||
| 154 | + WritebackLayout{{3, 1}, false, true}, WritebackLayout{{6, 1}, false, false}), | ||
| 155 | + testing::Values(int64_t{0}, int64_t{3}), testing::Values(ACL_INT32, ACL_INT64))); | ||
| 156 | +} // namespace | ||
| @@ -74,6 +74,7 @@ if(UT_TEST_ALL OR OP_API_UT) | |||
| 74 | rt2_registry | 74 | rt2_registry |
| 75 | error_manager | 75 | error_manager |
| 76 | c_sec | 76 | c_sec |
| 77 | + ${CMAKE_DL_LIBS} | ||
| 77 | ) | 78 | ) |
| 78 | ## exec ophost ut exe | 79 | ## exec ophost ut exe |
| 79 | if(ENABLE_UT_EXEC) | 80 | if(ENABLE_UT_EXEC) |
触发条件:ProcessContiguousCase 同时服务 DAV_3510 与 DAV_2201(head 中两条分支之外统一走到该函数),本次把 ViewCopy 目标改为原始 varRef 后,910B/910_93 上的非连续 fallback 也会实际执行“连续副本计算 + ViewCopy 写回非连续视图”这条新路径。PR 描述明确只做了 950PR 定向构建与 60 组回归,未执行 910B 设备测试和全量 UT。后果:arch22 上新增的这条执行路径(非连续目标拷贝、对应 tiling/拷贝 kernel)未经验证,若平台差异导致异常,回归仍会漏。建议:在 A2/A3(910B/910_93)上跑同一脚本的 fallback 用例,或补一条 910B 定向 UT/ST 后再合入。
如已修复请回复说明,我会在确认后把该意见标记为已解决。