Star661
250
代码
代码
Issues24
Pull Requests96
Actions
项目讨论
Wiki
分析
Star661
250
已合并
【PR】: revert reorganize tensor layout utilities and update codegen #1898
Phl17创建于 14 天前
【PR】: revert reorganize tensor layout utilities and update codegen #1898
已合并
Phl17创建于 14 天前
从qq_40517592/graph-autofusionrevert_0831合入到cann/graph-autofusiondevelop
共 16 个文件变更+61-157
Rautofuse/common/tensor_layout_utils.cpp→autofuse/optimize/tensor_layout_utils.cpp+3-2
@@ -15,8 +15,9 @@
15#include "ascgen_log.h"15#include "ascgen_log.h"
16#include "ascir_ops.h"16#include "ascir_ops.h"
17#include "common_utils.h"17#include "common_utils.h"
18+#include "schedule_utils.h"
atomgit-bot
atomgit-botatomgit-bot14 天前

🔴 Critical

changed line: autofuse/optimize/tensor_layout_utils.cpp 第 18 行新增的 include "schedule_utils.h" 缺少预处理指令前缀 #（diff 中为 +include "schedule_utils.h"，工作区文件内容经 read_file 与 grep 双重确认均为 include "schedule_utils.h"，而同一文件第 11~17 行的 #include 均正常带 #）。

affected behavior/contract: 该行位于文件作用域，include 会被当作未声明标识符解析，紧随其后的字符串字面量构成非法声明，整个翻译单元出现语法错误。

failure mode: tensor_layout_utils.cpp 无法通过编译，optimize 库构建失败；由于 AnalyzeLoadDiscontinuity/TensorLayoutUtils 被 un_alignment_strategy.cpp、load_to_nddma_template.cpp、nddma_template.cpp 等生产代码引用，将连带导致依赖该库的整个 autofuse 组件（以及涉及本文件的回归测试）编译失败，属 P0 级构建故障。

suggested fix: 将行首补上 # 恢复为 #include "schedule_utils.h"（该头文件存在于 autofuse/optimize/schedule_utils.h，修复后可用；若本文件实际并不需要它，也可直接删除此行）。

建议：把第 18 行改为 #include "schedule_utils.h"（或确认本文件未使用 schedule_utils 中的符号时直接删除该 include）。

  当前建议代码无改动
likedislike
不准确？
18 19 
19-namespace ascgen_utils {20+namespace optimize {
20af::Status TensorLayoutUtils::AnalyzeLoadDiscontinuity(const af::AscTensorAttr &attr, DiscontinuityInfo &info) {21af::Status TensorLayoutUtils::AnalyzeLoadDiscontinuity(const af::AscTensorAttr &attr, DiscontinuityInfo &info) {
21 const auto &axis = attr.axis;22 const auto &axis = attr.axis;
22 const auto &repeats = attr.repeats;23 const auto &repeats = attr.repeats;
@@ -63,4 +64,4 @@ af::Status TensorLayoutUtils::AnalyzeLoadDiscontinuity(const af::AscTensorAttr &
63 64 
64 return af::SUCCESS;65 return af::SUCCESS;
65}66}
66-} // namespace ascgen_utils67+} // namespace optimize
Rautofuse/common/tensor_layout_utils.h→autofuse/optimize/tensor_layout_utils.h+2-2
@@ -13,7 +13,7 @@
13 13 
14#include "ascendc_ir/ascendc_ir_core/ascendc_ir_def.h"14#include "ascendc_ir/ascendc_ir_core/ascendc_ir_def.h"
15 15 
16-namespace ascgen_utils {16+namespace optimize {
17struct DiscontinuityInfo {17struct DiscontinuityInfo {
18 bool is_tail_axis_discontinuous{false}; // 最后一根有效轴不连续18 bool is_tail_axis_discontinuous{false}; // 最后一根有效轴不连续
19 bool has_multiple_discontinuities{false}; // 多根轴不连续场景19 bool has_multiple_discontinuities{false}; // 多根轴不连续场景
@@ -22,6 +22,6 @@ class TensorLayoutUtils {
22 public:22 public:
23 static af::Status AnalyzeLoadDiscontinuity(const af::AscTensorAttr &attr, DiscontinuityInfo &info);23 static af::Status AnalyzeLoadDiscontinuity(const af::AscTensorAttr &attr, DiscontinuityInfo &info);
24};24};
25-} // namespace ascgen_utils25+} // namespace optimize
26 26 
27#endif // OPTIMIZE_TENSOR_LAYOUT_UTILS_H_27#endif // OPTIMIZE_TENSOR_LAYOUT_UTILS_H_
Dautofuse/tests/ut/common/test_tensor_layout_utils.cpp+0-94
@@ -1,94 +0,0 @@
1-/**
2- * Copyright (c) 2025 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 <gtest/gtest.h>
12- 
13-#include <vector>
14- 
15-#include "tensor_layout_utils.h"
16- 
17-namespace ascgen_utils {
18-namespace {
19-af::AscTensorAttr CreateAttr(const std::vector<int64_t> &axis, const std::vector<int64_t> &repeats,
20- const std::vector<int64_t> &strides, const std::vector<int64_t> &vectorized_axis) {
21- af::AscTensorAttr attr;
22- for (const auto value : axis) {
23- attr.axis.emplace_back(value);
24- }
25- for (const auto value : repeats) {
26- attr.repeats.emplace_back(af::Symbol(value));
27- }
28- for (const auto value : strides) {
29- attr.strides.emplace_back(af::Symbol(value));
30- }
31- attr.vectorized_axis = vectorized_axis;
32- return attr;
33-}
34-} // namespace
35- 
36-TEST(TensorLayoutUtilsTest, AnalyzeContinuousLayout) {
37- const auto attr = CreateAttr({0, 1}, {2, 4}, {4, 1}, {0, 1});
38- DiscontinuityInfo info;
39- 
40- EXPECT_EQ(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
41- EXPECT_FALSE(info.is_tail_axis_discontinuous);
42- EXPECT_FALSE(info.has_multiple_discontinuities);
43-}
44- 
45-TEST(TensorLayoutUtilsTest, AnalyzeSingleDiscontinuity) {
46- const auto attr = CreateAttr({0, 1}, {2, 4}, {8, 1}, {0, 1});
47- DiscontinuityInfo info;
48- 
49- EXPECT_EQ(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
50- EXPECT_FALSE(info.is_tail_axis_discontinuous);
51- EXPECT_FALSE(info.has_multiple_discontinuities);
52-}
53- 
54-TEST(TensorLayoutUtilsTest, AnalyzeMultipleDiscontinuities) {
55- const auto attr = CreateAttr({0, 1, 2}, {2, 2, 4}, {16, 3, 1}, {0, 1, 2});
56- DiscontinuityInfo info;
57- 
58- EXPECT_EQ(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
59- EXPECT_FALSE(info.is_tail_axis_discontinuous);
60- EXPECT_TRUE(info.has_multiple_discontinuities);
61-}
62- 
63-TEST(TensorLayoutUtilsTest, AnalyzeTailAxisDiscontinuity) {
64- const auto attr = CreateAttr({0, 1}, {2, 4}, {4, 2}, {1});
65- DiscontinuityInfo info;
66- 
67- EXPECT_EQ(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
68- EXPECT_TRUE(info.is_tail_axis_discontinuous);
69- EXPECT_FALSE(info.has_multiple_discontinuities);
70-}
71- 
72-TEST(TensorLayoutUtilsTest, IgnoreBroadcastAxisWhenAnalyzingDiscontinuity) {
73- const auto attr = CreateAttr({0, 1, 2}, {2, 3, 4}, {12, 0, 1}, {0, 1, 2});
74- DiscontinuityInfo info;
75- 
76- EXPECT_EQ(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
77- EXPECT_FALSE(info.is_tail_axis_discontinuous);
78- EXPECT_FALSE(info.has_multiple_discontinuities);
79-}
80- 
81-TEST(TensorLayoutUtilsTest, ReturnFailureWhenVectorizedAxisIsMissing) {
82- const auto attr = CreateAttr({0, 1}, {2, 4}, {4, 1}, {2});
83- DiscontinuityInfo info;
84- 
85- EXPECT_NE(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
86-}
87- 
88-TEST(TensorLayoutUtilsTest, ReturnFailureWhenLayoutAttributeLengthsAreInsufficient) {
89- const auto attr = CreateAttr({0, 1}, {2}, {4}, {0, 1});
90- DiscontinuityInfo info;
91- 
92- EXPECT_NE(TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info), af::SUCCESS);
93-}
94-} // namespace ascgen_utils
Mautofuse/tests/v35/st/backend_e2e_v2/floortoint_float_test/floortoint_float_backend_generate.cpp+1-1
@@ -49,6 +49,6 @@ TEST_F(TestBackendFloortointFloatE2e, FloortointFloatE2eCodegen) {
49 EXPECT_NE(kernel.find("AscendC::Cast(local_3[0], local_2[0], AscendC::RoundMode::CAST_FLOOR, "49 EXPECT_NE(kernel.find("AscendC::Cast(local_3[0], local_2[0], AscendC::RoundMode::CAST_FLOOR, "
50 "local_2_actual_size);"),50 "local_2_actual_size);"),
51 std::string::npos);51 std::string::npos);
52- EXPECT_NE(kernel.find("DataCopyPadExtend<int32_t, AscendC::PaddingMode::Compact>"), std::string::npos);52+ EXPECT_NE(kernel.find("DataCopyPadExtend<int32_t, AscendC::PaddingMode::Normal>"), std::string::npos);
53 });53 });
54}54}
Mautofuse/tests/v35/st/backend_e2e_v2/indirect_load_store_test/indirect_load_store_backend_generator.cpp+11-6
@@ -4174,10 +4174,10 @@ constexpr int64_t kOutputDim1 = 16;
4174constexpr int64_t kDim2 = 5;4174constexpr int64_t kDim2 = 5;
4175#endif4175#endif
4176constexpr int64_t kInputStride0 = IL_INPUT_STRIDE0;4176constexpr int64_t kInputStride0 = IL_INPUT_STRIDE0;
4177-constexpr int64_t kInputStride1 = (!kExpectSimt && !kExpectSk && IL_INPUT_STRIDE1 == 10) ? 5 : IL_INPUT_STRIDE1;4177+constexpr int64_t kInputStride1 = IL_INPUT_STRIDE1;
4178constexpr int64_t kInputStride2 = IL_INPUT_STRIDE2;4178constexpr int64_t kInputStride2 = IL_INPUT_STRIDE2;
4179constexpr int64_t kIndexStride0 = IL_INDEX_STRIDE0;4179constexpr int64_t kIndexStride0 = IL_INDEX_STRIDE0;
4180-constexpr int64_t kIndexStride1 = (!kExpectSimt && !kExpectSk && IL_INDEX_STRIDE1 == 10) ? 5 : IL_INDEX_STRIDE1;4180+constexpr int64_t kIndexStride1 = IL_INDEX_STRIDE1;
4181constexpr int64_t kIndexStride2 = IL_INDEX_STRIDE2;4181constexpr int64_t kIndexStride2 = IL_INDEX_STRIDE2;
4182#ifndef IL_INDEX_SELECT_CASE4182#ifndef IL_INDEX_SELECT_CASE
4183constexpr int64_t kEffectiveInputStride0 = kInputStride0;4183constexpr int64_t kEffectiveInputStride0 = kInputStride0;
@@ -4369,8 +4369,8 @@ std::shared_ptr<af::AscGraph> CreateSubGraph() {
4369 const std::vector<af::AxisId> output_axes = {view.output_row_axis, view.output_inner_axis};4369 const std::vector<af::AxisId> output_axes = {view.output_row_axis, view.output_inner_axis};
4370 const std::vector<af::Expression> output_repeats = {view.index_rows, view.embedding_size};4370 const std::vector<af::Expression> output_repeats = {view.index_rows, view.embedding_size};
4371 const std::vector<af::Expression> output_strides = {view.embedding_size, af::ops::One};4371 const std::vector<af::Expression> output_strides = {view.embedding_size, af::ops::One};
4372- const std::vector<af::Expression> index_repeats = {view.index_rows, view.embedding_size};4372+ const std::vector<af::Expression> index_repeats = {view.index_rows, af::ops::One};
4373- const std::vector<af::Expression> index_strides = {view.embedding_size, af::ops::One};4373+ const std::vector<af::Expression> index_strides = {af::ops::One, af::ops::Zero};
4374 const std::vector<af::Expression> reduce_repeats = {view.index_rows, af::ops::One};4374 const std::vector<af::Expression> reduce_repeats = {view.index_rows, af::ops::One};
4375 const std::vector<af::Expression> reduce_strides = {af::ops::One, af::ops::Zero};4375 const std::vector<af::Expression> reduce_strides = {af::ops::One, af::ops::Zero};
4376 4376 
@@ -4388,11 +4388,16 @@ std::shared_ptr<af::AscGraph> CreateSubGraph() {
4388 index_load.x = index.y;4388 index_load.x = index.y;
4389 SetView(index_load, output_axes, index_repeats, index_strides, af::DT_INT32);4389 SetView(index_load, output_axes, index_repeats, index_strides, af::DT_INT32);
4390 4390 
4391+ af::ascir_op::Cast index_cast("index_cast");
4392+ view.graph->AddNode(index_cast);
4393+ index_cast.x = index_load.y;
4394+ SetView(index_cast, output_axes, index_repeats, index_strides, af::DT_INT64);
4395+ 
4391 af::ascir_op::Broadcast index_broadcast("index_broadcast");4396 af::ascir_op::Broadcast index_broadcast("index_broadcast");
4392 view.graph->AddNode(index_broadcast);4397 view.graph->AddNode(index_broadcast);
4393 index_broadcast.attr.api.compute_type = af::ComputeType::kComputeBroadcast;4398 index_broadcast.attr.api.compute_type = af::ComputeType::kComputeBroadcast;
4394- index_broadcast.x = index_load.y;4399+ index_broadcast.x = index_cast.y;
4395- SetView(index_broadcast, output_axes, output_repeats, output_strides, af::DT_INT32);4400+ SetView(index_broadcast, output_axes, output_repeats, output_strides, af::DT_INT64);
4396 4401 
4397 af::ascir_op::IndirectLoad indirect_load("indirect_load");4402 af::ascir_op::IndirectLoad indirect_load("indirect_load");
4398 view.graph->AddNode(indirect_load);4403 view.graph->AddNode(indirect_load);
Mautofuse/tests/v35/st/backend_e2e_v2/indirect_load_store_test/test_e2e_indirect_load_store_kernel.cpp+5-8
@@ -1682,10 +1682,10 @@ extern "C" int64_t AutofuseTiling(AutofuseTilingData *, uint32_t *, uint32_t *,
1682 1682 
1683namespace {1683namespace {
1684constexpr int32_t kInputStride0 = IL_INPUT_STRIDE0;1684constexpr int32_t kInputStride0 = IL_INPUT_STRIDE0;
1685-constexpr int32_t kInputStride1 = (!IL_EXPECT_SIMT && !IL_EXPECT_SK && IL_INPUT_STRIDE1 == 10) ? 5 : IL_INPUT_STRIDE1;1685+constexpr int32_t kInputStride1 = IL_INPUT_STRIDE1;
1686constexpr int32_t kInputStride2 = IL_INPUT_STRIDE2;1686constexpr int32_t kInputStride2 = IL_INPUT_STRIDE2;
1687constexpr int32_t kIndexStride0 = IL_INDEX_STRIDE0;1687constexpr int32_t kIndexStride0 = IL_INDEX_STRIDE0;
1688-constexpr int32_t kIndexStride1 = (!IL_EXPECT_SIMT && !IL_EXPECT_SK && IL_INDEX_STRIDE1 == 10) ? 5 : IL_INDEX_STRIDE1;1688+constexpr int32_t kIndexStride1 = IL_INDEX_STRIDE1;
1689constexpr int32_t kIndexStride2 = IL_INDEX_STRIDE2;1689constexpr int32_t kIndexStride2 = IL_INDEX_STRIDE2;
1690#ifdef IL_INDEX_SELECT_CASE1690#ifdef IL_INDEX_SELECT_CASE
1691constexpr int32_t kEffectiveInputStride0 = 138;1691constexpr int32_t kEffectiveInputStride0 = 138;
@@ -1827,13 +1827,10 @@ void InitializeData(float *input, int32_t *index, float *expected) {
1827 }1827 }
1828 }1828 }
1829 for (int32_t row = 0; row < kIndexRows; ++row) {1829 for (int32_t row = 0; row < kIndexRows; ++row) {
1830- const int32_t index_value = static_cast<int32_t>((row * 7 + 3) % kInputRows);1830+ index[row] = static_cast<int32_t>((row * 7 + 3) % kInputRows);
1831- for (int32_t col = 0; col < kEmbeddingSize; ++col) {
1832- index[row * kEmbeddingSize + col] = index_value;
1833- }
1834 float sum = 0.0F;1831 float sum = 0.0F;
1835 for (int32_t col = 0; col < kEmbeddingSize; ++col) {1832 for (int32_t col = 0; col < kEmbeddingSize; ++col) {
1836- sum += 2.0F * (input[index[row * kEmbeddingSize] * kEmbeddingSize + col] + 0.1F);1833+ sum += 2.0F * (input[index[row] * kEmbeddingSize + col] + 0.1F);
1837 }1834 }
1838 expected[row] = sum;1835 expected[row] = sum;
1839 }1836 }
@@ -1842,7 +1839,7 @@ void InitializeData(float *input, int32_t *index, float *expected) {
1842 1839 
1843TEST(E2EIndirectLoadEmbedding, GeneratedKernelMatchesReference) {1840TEST(E2EIndirectLoadEmbedding, GeneratedKernelMatchesReference) {
1844 constexpr int64_t input_count = static_cast<int64_t>(kInputRows) * kEmbeddingSize;1841 constexpr int64_t input_count = static_cast<int64_t>(kInputRows) * kEmbeddingSize;
1845- constexpr int64_t index_count = static_cast<int64_t>(kIndexRows) * kEmbeddingSize;1842+ constexpr int64_t index_count = kIndexRows;
1846 constexpr int64_t output_count = kIndexRows;1843 constexpr int64_t output_count = kIndexRows;
1847 indirect_load_test::KernelData<float, int32_t> buffers(input_count, index_count, output_count);1844 indirect_load_test::KernelData<float, int32_t> buffers(input_count, index_count, output_count);
1848 ASSERT_TRUE(buffers.IsValid());1845 ASSERT_TRUE(buffers.IsValid());
Mautofuse/tests/v35/st/backend_e2e_v2/load_where_x2_x3_is_ubscalar_store_test/load_where_x2_x3_is_ubscalar_store_backend_generate.cpp+1-1
@@ -52,6 +52,6 @@ TEST_F(TestBackendLoadWhereX2X3IsUbscalarStoreE2e, LoadWhereX2X3IsUbscalarStoreE
52 EXPECT_NE(kernel.find("Duplicate(local_blk_tensor_of_scalar_2[0], static_cast<float>(100), "52 EXPECT_NE(kernel.find("Duplicate(local_blk_tensor_of_scalar_2[0], static_cast<float>(100), "
53 "static_cast<uint64_t>(32/sizeof(float)));"),53 "static_cast<uint64_t>(32/sizeof(float)));"),
54 std::string::npos);54 std::string::npos);
55- EXPECT_NE(kernel.find("DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(global_1"), std::string::npos);55+ EXPECT_NE(kernel.find("DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(global_1"), std::string::npos);
56 });56 });
57}57}
Mautofuse/tests/v35/st/codegen/e2e_v2/load_nan_out_for_store_expect_code/test_e2e_load_nan_out_for_store_expect_kernel.cpp+2-6
@@ -1,5 +1,4 @@
1#include <gtest/gtest.h>1#include <gtest/gtest.h>
2-#include <algorithm>
3#include "tikicpulib.h"2#include "tikicpulib.h"
4 3 
5#include "autofuse_tiling_data.h"4#include "autofuse_tiling_data.h"
@@ -18,7 +17,6 @@ TEST_P(E2E_LoadNanOutForStore_Code, CalculateCorrect_Nan) {
18 float *input = (float *)AscendC::GmAlloc(test_size * sizeof(float) + 32);17 float *input = (float *)AscendC::GmAlloc(test_size * sizeof(float) + 32);
19 uint8_t *y = (uint8_t *)AscendC::GmAlloc(test_size * sizeof(uint8_t) + 32);18 uint8_t *y = (uint8_t *)AscendC::GmAlloc(test_size * sizeof(uint8_t) + 32);
20 uint8_t *expect = (uint8_t *)AscendC::GmAlloc(test_size * sizeof(uint8_t) + 32);19 uint8_t *expect = (uint8_t *)AscendC::GmAlloc(test_size * sizeof(uint8_t) + 32);
21- std::fill_n(y, test_size, uint8_t{0});
22 20 
23 // Prepare test and expect data21 // Prepare test and expect data
24 srand(1);22 srand(1);
@@ -34,8 +32,6 @@ TEST_P(E2E_LoadNanOutForStore_Code, CalculateCorrect_Nan) {
34 tiling_data.s2 = test_shape[2];32 tiling_data.s2 = test_shape[2];
35 tiling_data.tiling_key = 0;33 tiling_data.tiling_key = 0;
36 GetTiling(tiling_data);34 GetTiling(tiling_data);
37- // This generated test kernel iterates all z0 slices and does not partition them by block index.
38- tiling_data.block_dim = 1;
39 35 
40 AscendC::SetKernelMode(KernelMode::AIV_MODE);36 AscendC::SetKernelMode(KernelMode::AIV_MODE);
41 ICPU_RUN_KF(load_nan_out_for_store, tiling_data.block_dim, (uint8_t *)input, (uint8_t *)y, nullptr,37 ICPU_RUN_KF(load_nan_out_for_store, tiling_data.block_dim, (uint8_t *)input, (uint8_t *)y, nullptr,
@@ -57,5 +53,5 @@ TEST_P(E2E_LoadNanOutForStore_Code, CalculateCorrect_Nan) {
57}53}
58 54 
59INSTANTIATE_TEST_SUITE_P(CalcWithDifferentShape, E2E_LoadNanOutForStore_Code,55INSTANTIATE_TEST_SUITE_P(CalcWithDifferentShape, E2E_LoadNanOutForStore_Code,
60- ::testing::Values(std::vector<int>{2, 8, 32}, std::vector<int>{8, 16, 32},56+ ::testing::Values(std::vector<int>{2, 8, 8}, std::vector<int>{8, 16, 16},
61- std::vector<int>{96, 16, 32}));57+ std::vector<int>{96, 16, 16}));
Mautofuse/tests/v35/ut/codegen/reg_api_call/test_codegen_load_reg_api_call.cpp+5-5
@@ -84,9 +84,9 @@ TEST(CodegenKernel, LoadRegApiCall_OneDimLoad) {
84 84 
85 std::string result;85 std::string result;
86 call_0.Generate(tpipe, vector<af::AxisId>{}, result);86 call_0.Generate(tpipe, vector<af::AxisId>{}, result);
87- EXPECT_EQ(result,87+ EXPECT_EQ(
88- std::string{88+ result,
89- "DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_0[0], local_0[0 + 0], 1, 8, 0, 0);\n"});89+ std::string{"DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_0[0], local_0[0 + 0], 1, 8, 0, 0);\n"});
90}90}
91 91 
92TEST(CodegenKernel, LoadRegApiCall_CvUbFuseUsesDtypeAwareStrides) {92TEST(CodegenKernel, LoadRegApiCall_CvUbFuseUsesDtypeAwareStrides) {
@@ -296,7 +296,7 @@ TEST(CodegenKernel, LoadRegApiCall_ThreeDimLoad) {
296 call_0.Generate(tpipe, vector<af::AxisId>{}, result);296 call_0.Generate(tpipe, vector<af::AxisId>{}, result);
297 EXPECT_EQ(297 EXPECT_EQ(
298 result,298 result,
299- std::string{"DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_0[0], local_0[0 + 0], 4, 2, 2, (2 "299+ std::string{"DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_0[0], local_0[0 + 0], 4, 2, 2, (2 "
300 "- 2), {static_cast<uint32_t>(8), "300 "- 2), {static_cast<uint32_t>(8), "
301 "static_cast<uint32_t>(1), static_cast<uint64_t>(24 * 4), static_cast<uint64_t>(8 * 4), "301 "static_cast<uint32_t>(1), static_cast<uint64_t>(24 * 4), static_cast<uint64_t>(8 * 4), "
302 "static_cast<uint64_t>(0 * 4), static_cast<uint64_t>(0 * 4)});\n"});302 "static_cast<uint64_t>(0 * 4), static_cast<uint64_t>(0 * 4)});\n"});
@@ -385,7 +385,7 @@ TEST(CodegenKernel, LoadRegApiCall_FiveDimLoad) {
385 call_0.Generate(tpipe, vector<af::AxisId>{}, result);385 call_0.Generate(tpipe, vector<af::AxisId>{}, result);
386 EXPECT_EQ(result,386 EXPECT_EQ(result,
387 std::string{"for (int outer_for_0 = 0; outer_for_0 < 8; outer_for_0++) {\nDataCopyPadExtend<float, "387 std::string{"for (int outer_for_0 = 0; outer_for_0 < 8; outer_for_0++) {\nDataCopyPadExtend<float, "
388- "AscendC::PaddingMode::Normal>(local_0[outer_for_0 * "388+ "AscendC::PaddingMode::Compact>(local_0[outer_for_0 * "
389 "32], local_0[(0 + 0 + outer_for_0 * 160)], 2, 2, 2, (2 - 2), {static_cast<uint32_t>(2), "389 "32], local_0[(0 + 0 + outer_for_0 * 160)], 2, 2, 2, (2 - 2), {static_cast<uint32_t>(2), "
390 "static_cast<uint32_t>(4), static_cast<uint64_t>(12 * 4), static_cast<uint64_t>(4 * 4), "390 "static_cast<uint32_t>(4), static_cast<uint64_t>(12 * 4), static_cast<uint64_t>(4 * 4), "
391 "static_cast<uint64_t>(32 * 4), static_cast<uint64_t>(8 * 4)});\n}\n"});391 "static_cast<uint64_t>(32 * 4), static_cast<uint64_t>(8 * 4)});\n}\n"});
Mautofuse/tests/v35/ut/codegen/reg_api_call/test_codegen_store_reg_api_call.cpp+5-5
@@ -130,13 +130,13 @@ TEST(CodegenKernel, StoreRegApiCall_TwoStoreOneOutput) {
130 EXPECT_EQ(130 EXPECT_EQ(
131 result,131 result,
132 std::string{132 std::string{
133- "DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_1[0 + 0], local_0[0], 1, 1, (16 - 1), 0);\n"});133+ "DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_1[0 + 0], local_0[0], 1, 1, (16 - 1), 0);\n"});
134 134 
135 codegen::StoreRegApiCall call_1("DataCopyPadExtend");135 codegen::StoreRegApiCall call_1("DataCopyPadExtend");
136 EXPECT_EQ(call_1.Init(store_1), 0);136 EXPECT_EQ(call_1.Init(store_1), 0);
137 call_1.inputs.push_back(&x1);137 call_1.inputs.push_back(&x1);
138 call_1.Generate(tpipe, vector<af::AxisId>{}, result);138 call_1.Generate(tpipe, vector<af::AxisId>{}, result);
139- EXPECT_EQ(result, std::string{"DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_1[0 + 1], local_0[0], "139+ EXPECT_EQ(result, std::string{"DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_1[0 + 1], local_0[0], "
140 "z0_t_size, 1, (16 - 1), 0);\n"});140 "z0_t_size, 1, (16 - 1), 0);\n"});
141}141}
142 142 
@@ -352,7 +352,7 @@ TEST(CodegenKernel, StoreRegApiCall_NeetMte3SyncMte2) {
352 EXPECT_EQ(352 EXPECT_EQ(
353 result,353 result,
354 std::string{354 std::string{
355- "DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_1[0 + 0], local_0[0], 1, 1, (16 - 1), 0);\n"355+ "DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_1[0 + 0], local_0[0], 1, 1, (16 - 1), 0);\n"
356 "auto local_0_e_mte3_2_mte2_t_0 = tpipe.AllocEventID<HardEvent::MTE3_MTE2>();\n"356 "auto local_0_e_mte3_2_mte2_t_0 = tpipe.AllocEventID<HardEvent::MTE3_MTE2>();\n"
357 "TQueSync<PIPE_MTE3, PIPE_MTE2> local_0_s_mte3_2_mte2_t_0;\n"357 "TQueSync<PIPE_MTE3, PIPE_MTE2> local_0_s_mte3_2_mte2_t_0;\n"
358 "local_0_s_mte3_2_mte2_t_0.SetFlag(local_0_e_mte3_2_mte2_t_0);\n"358 "local_0_s_mte3_2_mte2_t_0.SetFlag(local_0_e_mte3_2_mte2_t_0);\n"
@@ -449,7 +449,7 @@ TEST(CodegenKernel, StoreRegApiCall_ThreeDimStore) {
449 EXPECT_EQ(449 EXPECT_EQ(
450 result,450 result,
451 std::string{451 std::string{
452- "DataCopyPadExtend<float, AscendC::PaddingMode::Normal>(local_1[0 + 0], local_0[0], 4, 2, (2 - 2), 2, "452+ "DataCopyPadExtend<float, AscendC::PaddingMode::Compact>(local_1[0 + 0], local_0[0], 4, 2, (2 - 2), 2, "
453 "{static_cast<uint32_t>(8), static_cast<uint32_t>(1), static_cast<uint64_t>(8 * 4), static_cast<uint64_t>(24 "453 "{static_cast<uint32_t>(8), static_cast<uint32_t>(1), static_cast<uint64_t>(8 * 4), static_cast<uint64_t>(24 "
454 "* 4), static_cast<uint64_t>(0 * 4), static_cast<uint64_t>(0 * 4)});\n"});454 "* 4), static_cast<uint64_t>(0 * 4), static_cast<uint64_t>(0 * 4)});\n"});
455}455}
@@ -553,7 +553,7 @@ TEST(CodegenKernel, StoreRegApiCall_FiveDimStore) {
553 call_0.Generate(tpipe, vector<af::AxisId>{}, result);553 call_0.Generate(tpipe, vector<af::AxisId>{}, result);
554 EXPECT_EQ(result,554 EXPECT_EQ(result,
555 std::string{"for (int outer_for_0 = 0; outer_for_0 < 8; outer_for_0++) {\nDataCopyPadExtend<float, "555 std::string{"for (int outer_for_0 = 0; outer_for_0 < 8; outer_for_0++) {\nDataCopyPadExtend<float, "
556- "AscendC::PaddingMode::Normal>(local_1[(0 + 0 + "556+ "AscendC::PaddingMode::Compact>(local_1[(0 + 0 + "
557 "outer_for_0 * 160)], local_0[outer_for_0 * 32], 2, 2, (2 - 2), 2, {static_cast<uint32_t>(2), "557 "outer_for_0 * 160)], local_0[outer_for_0 * 32], 2, 2, (2 - 2), 2, {static_cast<uint32_t>(2), "
558 "static_cast<uint32_t>(4), static_cast<uint64_t>(4 * 4), static_cast<uint64_t>(12 * 4), "558 "static_cast<uint32_t>(4), static_cast<uint64_t>(4 * 4), static_cast<uint64_t>(12 * 4), "
559 "static_cast<uint64_t>(8 * 4), "559 "static_cast<uint64_t>(8 * 4), "
Mautofuse/v35/codegen/reg_api_call/reg_api_call_utils.cpp+18-17
@@ -10,7 +10,6 @@
10 10 
11#include "reg_api_call_utils.h"11#include "reg_api_call_utils.h"
12#include "graph/symbolizer/symbolic_utils.h"12#include "graph/symbolizer/symbolic_utils.h"
13-#include "tensor_layout_utils.h"
14 13 
15namespace {14namespace {
16constexpr size_t kDmaMaxLen = 2U;15constexpr size_t kDmaMaxLen = 2U;
@@ -99,21 +98,23 @@ void SetLoopModeParamsExpr(const DataCopyParams &data_copy_param, LoopModeParams
99 }98 }
100}99}
101 100 
102-std::string GetPaddingMode(const Tensor &ub_tensor) {101+// 根据ub上最后一维的对齐信息以及切分轴信息判断，是否使用Compact模式，如果能明确判断出来，stride与repeat相同且ub切分轴是首轴，则使用compact模式，否则使用normal模式
103- af::AscTensorAttr attr;102+std::string GetPaddingMode(const TPipe &tpipe, const Tensor &ub_tensor, const DataCopyParams &data_copy_param) {
104- attr.axis = ub_tensor.axis;103+ for (auto axis_pos : ub_tensor.vectorized_axis_pos) {
105- attr.repeats = ub_tensor.axis_size;104+ ascir::AxisId axis_id = ub_tensor.axis[axis_pos];
106- attr.strides = ub_tensor.axis_strides;105+ const Axis &axis = tpipe.tiler.GetAxis(axis_id);
107- attr.vectorized_axis = ub_tensor.vectorized_axis;106+ if (axis.type == ascir::Axis::Type::kAxisTypeTileInner && ub_tensor.vectorized_axis[0] != axis_id) {
108- ascgen_utils::DiscontinuityInfo info;107+ GELOGD("The TileInner axis is not the first axis, use normal mode.");
109- auto analyze_success = ascgen_utils::TensorLayoutUtils::AnalyzeLoadDiscontinuity(attr, info) == af::SUCCESS;108+ return kNormalPddingMode;
110- auto no_multiple_discontinuities = !info.has_multiple_discontinuities;109+ }
111- GELOGD("GetPaddingMode conditions: analyze_success=%d, no_multiple_discontinuities=%d", analyze_success,
112- no_multiple_discontinuities);
113- if (analyze_success && no_multiple_discontinuities) {
114- return kCompactPddingMode;
115 }110 }
116- return kNormalPddingMode;111+ if (data_copy_param.repeats.size() <= 1) {
112+ return kNormalPddingMode;
113+ }
114+ ascir::SizeExpr repeat = data_copy_param.repeats.back();
115+ ascir::SizeExpr stride = data_copy_param.ub_strides[data_copy_param.ub_strides.size() - kDmaMaxLen];
116+ bool status = af::SymbolicUtils::StaticCheckEq(repeat, stride) == af::TriBool::kTrue;
117+ return status ? kCompactPddingMode : kNormalPddingMode;
117}118}
118 119 
119std::string GenLoopModeParams(const LoopModeParams &loop_mode_param, int64_t input_dtype_size,120std::string GenLoopModeParams(const LoopModeParams &loop_mode_param, int64_t input_dtype_size,
@@ -169,7 +170,7 @@ void CreateEnhanceDmaCall(const TPipe &tpipe, const Tensor &input, const Tensor
169 LoopModeParams loop_mode_param;170 LoopModeParams loop_mode_param;
170 SetLoopModeParams(tpipe, data_copy_param, loop_mode_param, copy_in);171 SetLoopModeParams(tpipe, data_copy_param, loop_mode_param, copy_in);
171 const Tensor &ub_tensor = copy_in ? output : input;172 const Tensor &ub_tensor = copy_in ? output : input;
172- std::string padding_mode = GetPaddingMode(ub_tensor);173+ std::string padding_mode = GetPaddingMode(tpipe, ub_tensor, data_copy_param);
173 if (total_len <= kDmaMaxLen) {174 if (total_len <= kDmaMaxLen) {
174 CreateBaseDmaCall(input, output, dma_param, padding_mode, ss, copy_in);175 CreateBaseDmaCall(input, output, dma_param, padding_mode, ss, copy_in);
175 return;176 return;
@@ -348,7 +349,7 @@ Status BuildDataCopyApiParamInNormal(const TPipe &tpipe, CodegenApiParam &api_pa
348 GE_ASSERT_TRUE(CalculateDmaParams(tpipe, dst, dst, data_copy_param), "CalculateDmaParams failed");349 GE_ASSERT_TRUE(CalculateDmaParams(tpipe, dst, dst, data_copy_param), "CalculateDmaParams failed");
349 size_t total_len = data_copy_param.repeats.size();350 size_t total_len = data_copy_param.repeats.size();
350 const Tensor &ub_tensor = copy_in ? dst : src;351 const Tensor &ub_tensor = copy_in ? dst : src;
351- std::string padding_mode = GetPaddingMode(ub_tensor);352+ std::string padding_mode = GetPaddingMode(tpipe, ub_tensor, data_copy_param);
352 api_param.template_params.emplace_back(padding_mode);353 api_param.template_params.emplace_back(padding_mode);
353 354 
354 BuildDataCopyBaseParams(tpipe, data_copy_param, dma_specific_params, copy_in);355 BuildDataCopyBaseParams(tpipe, data_copy_param, dma_specific_params, copy_in);
Mautofuse/v35/codegen/reg_api_call/reg_api_call_utils.h+1-1
@@ -84,7 +84,7 @@ void SetNddmaParams(const TPipe &tpipe, const DataCopyParams &data_copy_param, N
84void SetLoopModeParams(const TPipe &tpipe, const DataCopyParams &data_copy_param, LoopModeParams &loop_mode_param,84void SetLoopModeParams(const TPipe &tpipe, const DataCopyParams &data_copy_param, LoopModeParams &loop_mode_param,
85 bool copy_in);85 bool copy_in);
86void SetLoopModeParamsExpr(const DataCopyParams &data_copy_param, LoopModeParamsExpr &loop_mode_param, bool copy_in);86void SetLoopModeParamsExpr(const DataCopyParams &data_copy_param, LoopModeParamsExpr &loop_mode_param, bool copy_in);
87-std::string GetPaddingMode(const Tensor &ub_tensor);87+std::string GetPaddingMode(const TPipe &tpipe, const Tensor &ub_tensor, const DataCopyParams &data_copy_param);
88void BuildDataCopyApiParamInCVFusion(const TPipe &tpipe, CodegenApiParam &api_param,88void BuildDataCopyApiParamInCVFusion(const TPipe &tpipe, CodegenApiParam &api_param,
89 DmaSpecificParams &dma_specific_params, const Tensor &gm, const Tensor &ub,89 DmaSpecificParams &dma_specific_params, const Tensor &gm, const Tensor &ub,
90 std::string &dtype_name, bool copy_in);90 std::string &dtype_name, bool copy_in);
Mautofuse/v35/codegen/reg_api_call/reg_broadcast_api_call.cpp+1-3
@@ -27,8 +27,6 @@ static void GenParams(const TPipe &tpipe, const Tensor &input, const Tensor &out
27 // 只保证在仅对张量尾轴做32B对齐的场景下有效，若对中间轴做了对齐，则还需要增加处理逻辑27 // 只保证在仅对张量尾轴做32B对齐的场景下有效，若对中间轴做了对齐，则还需要增加处理逻辑
28 auto vectorized_axis_size = input.vectorized_axis.size();28 auto vectorized_axis_size = input.vectorized_axis.size();
29 const char *shape_prefix = is_src ? "src_shape_" : "dst_shape_";29 const char *shape_prefix = is_src ? "src_shape_" : "dst_shape_";
30- constexpr char kCompactPddingMode[] = "AscendC::PaddingMode::Compact";
31- std::string padding_mode = GetPaddingMode(output);
32 30 
33 ss << "const uint32_t " << shape_prefix << input.id << "_brc_to_" << output.id << "[" << vectorized_axis_size31 ss << "const uint32_t " << shape_prefix << input.id << "_brc_to_" << output.id << "[" << vectorized_axis_size
34 << "] = {";32 << "] = {";
@@ -67,7 +65,7 @@ static void GenParams(const TPipe &tpipe, const Tensor &input, const Tensor &out
67 ascir::AxisId axis_id = output.vectorized_axis[pos];65 ascir::AxisId axis_id = output.vectorized_axis[pos];
68 auto last_dim_size = output.vectorized_strides[pre_pos];66 auto last_dim_size = output.vectorized_strides[pre_pos];
69 if (tpipe.tiler.GetAxis(axis_id).type != ascir::Axis::Type::kAxisTypeTileInner ||67 if (tpipe.tiler.GetAxis(axis_id).type != ascir::Axis::Type::kAxisTypeTileInner ||
70- output.vectorized_axis[0] == axis_id || padding_mode == kCompactPddingMode) {68+ output.vectorized_axis[0] == axis_id) {
71 ss << tpipe.tiler.ActualSize(last_dim_size);69 ss << tpipe.tiler.ActualSize(last_dim_size);
72 } else {70 } else {
73 ss << tpipe.tiler.Size(last_dim_size);71 ss << tpipe.tiler.Size(last_dim_size);
Mautofuse/v35/optimize/template/load_to_nddma_template.cpp+2-2
@@ -42,8 +42,8 @@ af::Status LoadToNddmaTemplate::Generate(const af::AscGraph &origin_graph,
42 if (ascgen_utils::indirect_load::GetTemplateBehavior(node).uses_direct_gm_pipeline) {42 if (ascgen_utils::indirect_load::GetTemplateBehavior(node).uses_direct_gm_pipeline) {
43 continue;43 continue;
44 }44 }
45- ascgen_utils::DiscontinuityInfo info;45+ DiscontinuityInfo info;
46- GE_ASSERT_SUCCESS(ascgen_utils::TensorLayoutUtils::AnalyzeLoadDiscontinuity(node->outputs[0].attr, info),46+ GE_ASSERT_SUCCESS(TensorLayoutUtils::AnalyzeLoadDiscontinuity(node->outputs[0].attr, info),
47 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());47 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());
48 bool need_align_at_repeat1 = info.has_multiple_discontinuities && info.is_tail_axis_discontinuous;48 bool need_align_at_repeat1 = info.has_multiple_discontinuities && info.is_tail_axis_discontinuous;
49 if (IsLoadNeedAlign(node) || need_align_at_repeat1) {49 if (IsLoadNeedAlign(node) || need_align_at_repeat1) {
Mautofuse/v35/optimize/template/nddma_template.cpp+2-2
@@ -271,8 +271,8 @@ af::Status NddmaTemplate::Generate([[maybe_unused]] const af::AscGraph &origin_g
271 }271 }
272 GE_ASSERT_SUCCESS(ProcessSliceToNddma(node, is_nddma_generated_cur));272 GE_ASSERT_SUCCESS(ProcessSliceToNddma(node, is_nddma_generated_cur));
273 is_nddma_generated = is_nddma_generated || is_nddma_generated_cur;273 is_nddma_generated = is_nddma_generated || is_nddma_generated_cur;
274- ascgen_utils::DiscontinuityInfo info;274+ DiscontinuityInfo info;
275- GE_ASSERT_SUCCESS(ascgen_utils::TensorLayoutUtils::AnalyzeLoadDiscontinuity(node->outputs[0].attr, info),275+ GE_ASSERT_SUCCESS(TensorLayoutUtils::AnalyzeLoadDiscontinuity(node->outputs[0].attr, info),
276 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());276 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());
277 bool need_align_at_repeat1 = info.has_multiple_discontinuities && info.is_tail_axis_discontinuous;277 bool need_align_at_repeat1 = info.has_multiple_discontinuities && info.is_tail_axis_discontinuous;
278 if (!is_nddma_generated_cur && (IsLoadNeedAlign(node) || need_align_at_repeat1)) {278 if (!is_nddma_generated_cur && (IsLoadNeedAlign(node) || need_align_at_repeat1)) {
Mautofuse/v35/optimize/un_alignment_strategy.cpp+2-2
@@ -36,8 +36,8 @@ af::Status UnAlignmentStrategy::LoadAlignmentInferFunc(const af::AscNodePtr &nod
36 tensor_to_align_type_[&output_attr] = {AlignmentType::kNotAligned};36 tensor_to_align_type_[&output_attr] = {AlignmentType::kNotAligned};
37 return af::SUCCESS;37 return af::SUCCESS;
38 }38 }
39- ascgen_utils::DiscontinuityInfo info;39+ DiscontinuityInfo info;
40- GE_ASSERT_SUCCESS(ascgen_utils::TensorLayoutUtils::AnalyzeLoadDiscontinuity(output_attr, info),40+ GE_ASSERT_SUCCESS(TensorLayoutUtils::AnalyzeLoadDiscontinuity(output_attr, info),
41 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());41 "Failed to analyze discontinuity info for node:[%s].", node->GetNamePtr());
42 if (!info.has_multiple_discontinuities) {42 if (!info.has_multiple_discontinuities) {
43 // 走compat模式搬运43 // 走compat模式搬运