Star860
1729
代码
代码
Issues162
Pull Requests291
Actions
项目讨论
Wiki
分析
Star860
1729
已合并
【ACLNN】sacle、zeroPoint to contiguous #2714
张喻翔创建于 3月14日
【ACLNN】sacle、zeroPoint to contiguous #2714
已合并
张喻翔创建于 3月14日
从zhangyuxiang0119/ops-nnop/90/fakequant合入到cann/ops-nn9.0.0
共 4 个文件变更+101-55
Mquant/fake_quant_affine_cachemask/op_host/op_api/aclnn_fake_quant_per_channel_affine_cachemask.cpp+9-27
@@ -15,20 +15,12 @@
15 15 
16#include "aclnn_fake_quant_per_channel_affine_cachemask.h"16#include "aclnn_fake_quant_per_channel_affine_cachemask.h"
17#include "fake_quant_affine_cachemask.h"17#include "fake_quant_affine_cachemask.h"
18-#include "aclnn_kernels/cast.h"
19-#include "aclnn_kernels/contiguous.h"
20#include "aclnn_kernels/transpose.h"18#include "aclnn_kernels/transpose.h"
21-#include "aclnn_kernels/common/op_error_check.h"
22-#include "aclnn/aclnn_base.h"
23-#include "opdev/common_types.h"
24-#include "opdev/data_type_utils.h"
25-#include "opdev/format_utils.h"
26-#include "opdev/op_dfx.h"
27-#include "opdev/op_executor.h"
28-#include "opdev/op_log.h"
29#include "opdev/tensor_view_utils.h"19#include "opdev/tensor_view_utils.h"
20+#include "fake_quant_common.h"
30 21 
31using namespace op;22using namespace op;
23+using namespace FakeQuantCommon;
32#ifdef __cplusplus24#ifdef __cplusplus
33extern "C" {25extern "C" {
34#endif26#endif
@@ -195,30 +187,20 @@ aclnnStatus aclnnFakeQuantPerChannelAffineCachemaskGetWorkspaceSize(
195 return ACLNN_SUCCESS;187 return ACLNN_SUCCESS;
196 }188 }
197 189 
198- auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());190+ const auto& [selfContiguous, scaleContiguous, zeroPointContiguous] = GetContiguousInput(self, scale, zeroPoint, uniqueExecutor.get());
199- CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);191+ CHECK_RET(selfContiguous != nullptr && scaleContiguous!= nullptr && zeroPointContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
200 192 
201- auto promoteType = op::PromoteType(self->GetDataType(), scale->GetDataType());193+ aclTensor* fakeQuantOut = nullptr, *fakeQuantMask = nullptr;
202- // 将输入self的数据类型转换成隐式数据类型，根据具体算子语义按需调用
203- auto selfCasted = l0op::Cast(selfContiguous, promoteType, uniqueExecutor.get());
204- CHECK_RET(selfCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
205- 
206- // 将输入other的数据类型转换成隐式数据类型，根据具体算子语义按需调用
207- auto scaleCasted = l0op::Cast(scale, promoteType, uniqueExecutor.get());
208- CHECK_RET(scaleCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
209- 
210- aclTensor* fakeQuantOut = nullptr;
211- aclTensor* fakeQuantMask = nullptr;
212 if (axis != 0) {194 if (axis != 0) {
213- aclIntArray* axes = GetTransposeArray(selfCasted, axis, uniqueExecutor.get());195+ aclIntArray* axes = GetTransposeArray(selfContiguous, axis, uniqueExecutor.get());
214 CHECK_RET(axes != nullptr, ACLNN_ERR_INNER_NULLPTR);196 CHECK_RET(axes != nullptr, ACLNN_ERR_INNER_NULLPTR);
215 197 
216 // 对self进行transpose198 // 对self进行transpose
217- auto selfTranspose = const_cast<aclTensor*>(l0op::Transpose(selfCasted, axes, uniqueExecutor.get()));199+ auto selfTranspose = const_cast<aclTensor*>(l0op::Transpose(selfContiguous, axes, uniqueExecutor.get()));
218 CHECK_RET(selfTranspose != nullptr, ACLNN_ERR_INNER_NULLPTR);200 CHECK_RET(selfTranspose != nullptr, ACLNN_ERR_INNER_NULLPTR);
219 201 
220 auto result = l0op::FakeQuantAffineCachemask(202 auto result = l0op::FakeQuantAffineCachemask(
221- selfTranspose, scaleCasted, zeroPoint, quantMin, quantMax, uniqueExecutor.get());203+ selfTranspose, scaleContiguous, zeroPointContiguous, quantMin, quantMax, uniqueExecutor.get());
222 auto fakeOut = std::get<0>(result);204 auto fakeOut = std::get<0>(result);
223 auto maskOut = std::get<1>(result);205 auto maskOut = std::get<1>(result);
224 CHECK_RET(fakeOut != nullptr && maskOut != nullptr, ACLNN_ERR_INNER_NULLPTR);206 CHECK_RET(fakeOut != nullptr && maskOut != nullptr, ACLNN_ERR_INNER_NULLPTR);
@@ -230,7 +212,7 @@ aclnnStatus aclnnFakeQuantPerChannelAffineCachemaskGetWorkspaceSize(
230 fakeQuantMask = const_cast<aclTensor*>(l0op::Transpose(maskOut, axes, uniqueExecutor.get()));212 fakeQuantMask = const_cast<aclTensor*>(l0op::Transpose(maskOut, axes, uniqueExecutor.get()));
231 } else {213 } else {
232 auto result = l0op::FakeQuantAffineCachemask(214 auto result = l0op::FakeQuantAffineCachemask(
233- selfCasted, scaleCasted, zeroPoint, quantMin, quantMax, uniqueExecutor.get());215+ selfContiguous, scaleContiguous, zeroPointContiguous, quantMin, quantMax, uniqueExecutor.get());
234 fakeQuantOut = std::get<0>(result);216 fakeQuantOut = std::get<0>(result);
235 fakeQuantMask = std::get<1>(result);217 fakeQuantMask = std::get<1>(result);
236 }218 }
Mquant/fake_quant_affine_cachemask/op_host/op_api/aclnn_fake_quant_per_tensor_affine_cachemask.cpp+10-28
@@ -15,21 +15,13 @@
15 15 
16#include "aclnn_fake_quant_per_tensor_affine_cachemask.h"16#include "aclnn_fake_quant_per_tensor_affine_cachemask.h"
17#include "fake_quant_affine_cachemask.h"17#include "fake_quant_affine_cachemask.h"
18-#include "aclnn_kernels/cast.h"
19-#include "aclnn_kernels/contiguous.h"
20#include "level0/broadcast_to.h"18#include "level0/broadcast_to.h"
21#include "level0/fill.h"19#include "level0/fill.h"
22-#include "aclnn_kernels/common/op_error_check.h"
23-#include "aclnn/aclnn_base.h"
24-#include "opdev/common_types.h"
25-#include "opdev/data_type_utils.h"
26-#include "opdev/format_utils.h"
27-#include "opdev/op_dfx.h"
28-#include "opdev/op_executor.h"
29-#include "opdev/op_log.h"
30#include "opdev/tensor_view_utils.h"20#include "opdev/tensor_view_utils.h"
21+#include "fake_quant_common.h"
31 22 
32using namespace op;23using namespace op;
24+using namespace FakeQuantCommon;
33#ifdef __cplusplus25#ifdef __cplusplus
34extern "C" {26extern "C" {
35#endif27#endif
@@ -169,38 +161,28 @@ aclnnStatus aclnnFakeQuantPerTensorAffineCachemaskGetWorkspaceSize(
169 return ACLNN_SUCCESS;161 return ACLNN_SUCCESS;
170 }162 }
171 163 
172- auto selfContiguous = l0op::Contiguous(self, uniqueExecutor.get());164+ const auto& [selfContiguous, scaleContiguous, zeroPointContiguous] = GetContiguousInput(self, scale, zeroPoint, uniqueExecutor.get());
173- CHECK_RET(selfContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);165+ CHECK_RET(selfContiguous != nullptr && scaleContiguous!= nullptr && zeroPointContiguous != nullptr, ACLNN_ERR_INNER_NULLPTR);
174 166 
175- auto promoteType = op::PromoteType(self->GetDataType(), scale->GetDataType());167+ aclTensor* fakeQuantOut = nullptr, *fakeQuantMask = nullptr;
176- // 将输入self的数据类型转换成隐式数据类型，根据具体算子语义按需调用
177- auto selfCasted = l0op::Cast(selfContiguous, promoteType, uniqueExecutor.get());
178- CHECK_RET(selfCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
179- 
180- // 将输入other的数据类型转换成隐式数据类型，根据具体算子语义按需调用
181- auto scaleCasted = l0op::Cast(scale, promoteType, uniqueExecutor.get());
182- CHECK_RET(scaleCasted != nullptr, ACLNN_ERR_INNER_NULLPTR);
183- 
184- aclTensor* fakeQuantOut = nullptr;
185- aclTensor* fakeQuantMask = nullptr;
186 if (fakeQuantEnbled < 1.0) {168 if (fakeQuantEnbled < 1.0) {
187 // 将结果values_transpose进行transpose，转换成正确的shape169 // 将结果values_transpose进行transpose，转换成正确的shape
188- fakeQuantOut = const_cast<aclTensor*>(selfCasted);170+ fakeQuantOut = const_cast<aclTensor*>(selfContiguous);
189 fakeQuantMask = const_cast<aclTensor*>(GetOutputTensorWithValueTrue(out, uniqueExecutor.get()));171 fakeQuantMask = const_cast<aclTensor*>(GetOutputTensorWithValueTrue(out, uniqueExecutor.get()));
190 } else {172 } else {
191- int64_t tensorSize = (int64_t)(selfCasted->GetViewShape().GetDim(0));173+ int64_t tensorSize = (int64_t)(selfContiguous->GetViewShape().GetDim(0));
192 int64_t tensorShape[1] = {tensorSize};174 int64_t tensorShape[1] = {tensorSize};
193 auto expectShape = uniqueExecutor.get()->AllocIntArray(tensorShape, 1);175 auto expectShape = uniqueExecutor.get()->AllocIntArray(tensorShape, 1);
194 CHECK_RET(expectShape != nullptr, ACLNN_ERR_INNER_NULLPTR);176 CHECK_RET(expectShape != nullptr, ACLNN_ERR_INNER_NULLPTR);
195 177 
196- auto scaleBroadcast = l0op::BroadcastTo(scaleCasted, expectShape, uniqueExecutor.get());178+ auto scaleBroadcast = l0op::BroadcastTo(scaleContiguous, expectShape, uniqueExecutor.get());
197 CHECK_RET(scaleBroadcast != nullptr, ACLNN_ERR_INNER_NULLPTR);179 CHECK_RET(scaleBroadcast != nullptr, ACLNN_ERR_INNER_NULLPTR);
198 180 
199- auto zeroPointBroadcast = l0op::BroadcastTo(zeroPoint, expectShape, uniqueExecutor.get());181+ auto zeroPointBroadcast = l0op::BroadcastTo(zeroPointContiguous, expectShape, uniqueExecutor.get());
200 CHECK_RET(zeroPointBroadcast != nullptr, ACLNN_ERR_INNER_NULLPTR);182 CHECK_RET(zeroPointBroadcast != nullptr, ACLNN_ERR_INNER_NULLPTR);
201 183 
202 auto result = l0op::FakeQuantAffineCachemask(184 auto result = l0op::FakeQuantAffineCachemask(
203- selfCasted, scaleBroadcast, zeroPointBroadcast, quantMin, quantMax, uniqueExecutor.get());185+ selfContiguous, scaleBroadcast, zeroPointBroadcast, quantMin, quantMax, uniqueExecutor.get());
204 fakeQuantOut = std::get<0>(result);186 fakeQuantOut = std::get<0>(result);
205 fakeQuantMask = std::get<1>(result);187 fakeQuantMask = std::get<1>(result);
206 }188 }
Aquant/fake_quant_affine_cachemask/op_host/op_api/fake_quant_common.cpp+46-0
@@ -0,0 +1,46 @@
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+/*!
12+ * \file fake_quant_common.cpp
13+ * \brief
14+ */
15+ 
16+#include "fake_quant_common.h"
17+ 
18+namespace FakeQuantCommon {
19+ 
20+auto nullptrInner = std::tuple<aclTensor*, aclTensor*, aclTensor*>(nullptr, nullptr, nullptr);
21+ 
22+std::tuple<const aclTensor*, const aclTensor*, const aclTensor*> GetContiguousInput(const aclTensor* self, const aclTensor* scale,
23+ const aclTensor* zeroPoint, aclOpExecutor* executor)
24+{
25+ auto selfContiguous = l0op::Contiguous(self, executor);
26+ OP_CHECK_NULL(selfContiguous, return nullptrInner);
27+ 
28+ auto promoteType = op::PromoteType(self->GetDataType(), scale->GetDataType());
29+ // 将输入self的数据类型转换成隐式数据类型，根据具体算子语义按需调用
30+ auto selfCasted = l0op::Cast(selfContiguous, promoteType, executor);
31+ OP_CHECK_NULL(selfCasted, return nullptrInner);
32+
33+ auto scaleContiguous = l0op::Contiguous(scale, executor);
34+ OP_CHECK_NULL(scaleContiguous, return nullptrInner);
35+ 
36+ // 将输入other的数据类型转换成隐式数据类型，根据具体算子语义按需调用
37+ auto scaleCasted = l0op::Cast(scaleContiguous, promoteType, executor);
38+ OP_CHECK_NULL(scaleCasted, return nullptrInner);
39+ 
40+ auto zeroPointContiguous = l0op::Contiguous(zeroPoint, executor);
41+ OP_CHECK_NULL(zeroPointContiguous, return nullptrInner);
42+ 
43+ return std::tuple<const aclTensor*, const aclTensor*, const aclTensor*>(selfCasted, scaleCasted, zeroPointContiguous);
44+}
45+ 
46+}
Aquant/fake_quant_affine_cachemask/op_host/op_api/fake_quant_common.h+36-0
@@ -0,0 +1,36 @@
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+/*!
CANN-robotCANN-robot
CANN-robotCANN-robot3月14日
代码结构与可维护性: 文件头注释中的文件名与实际文件名不一致。注释中写的是fake_quant_common.cpp，但实际文件是fake_quant_common.h。这种不一致可能导致开发者混淆，影响代码的可维护性。
问题类型: 代码结构与可维护性
文件路径: quant/fake_quant_affine_cachemask/op_host/op_api/fake_quant_common.h
行号: 11
问题代码:
/*!
 * \file fake_quant_common.cpp
 * \brief
 */
修改建议:
将注释中的文件名更正为fake_quant_common.h，确保文档与实际代码一致。
---
此评论由代码审查工具自动生成
likedislike
CANN-robotCANN-robot3月14日
代码文档不一致: 文件头注释中声明的文件名与实际文件名不一致。注释中写的是'fake_quant_common.cpp'，但实际文件是头文件'fake_quant_common.h'。这种不一致可能导致开发者在查找或理解代码时产生混淆。
问题类型: 代码文档不一致
文件路径: quant/fake_quant_affine_cachemask/op_host/op_api/fake_quant_common.h
行号: 11
问题代码:
/*!
 * \file fake_quant_common.cpp
 * \brief
 */
修改建议:
将文件头注释中的文件名修改为实际的文件名'fake_quant_common.h'，确保文档与实际代码保持一致。
---
此评论由代码审查工具自动生成
likedislike
12+ * \file fake_quant_common.cpp
13+ * \brief
14+ */
15+ 
16+#ifndef FAKE_QUANT_COMMON_H_
17+#define FAKE_QUANT_COMMON_H_
18+ 
19+#include "aclnn_kernels/contiguous.h"
20+#include "aclnn_kernels/common/op_error_check.h"
21+#include "aclnn/aclnn_base.h"
22+#include "aclnn_kernels/cast.h"
23+#include "opdev/common_types.h"
24+#include "opdev/data_type_utils.h"
25+#include "opdev/format_utils.h"
26+#include "opdev/op_dfx.h"
27+#include "opdev/op_executor.h"
28+#include "opdev/op_log.h"
29+ 
30+namespace FakeQuantCommon {
31+ 
32+std::tuple<const aclTensor*, const aclTensor*, const aclTensor*> GetContiguousInput(const aclTensor* self, const aclTensor* scale,
33+ const aclTensor* zeroPoint, aclOpExecutor* executor);
34+}
35+ 
36+#endif
CANN-robot
CANN-robotCANN-robot3月14日
代码结构与可维护性: 文件末尾缺少换行符（No newline at end of file）。虽然这不是功能性问题，但违反了许多代码规范（如POSIX标准），可能导致某些工具（如编译器、版本控制系统）处理时出现问题。
问题类型: 代码结构与可维护性
文件路径: quant/fake_quant_affine_cachemask/op_host/op_api/fake_quant_common.h
行号: 36
问题代码:
#endif
修改建议:
在#endif后添加一个换行符，确保文件以换行符结束。
---
此评论由代码审查工具自动生成
likedislike