Closed
fix: harden。Relu6Grad and SGD validation and test references #13276
raoliang_sacAdd on 21 days agoClosed at 18 days ago
fix: harden。Relu6Grad and SGD validation and test references #13276
Closed
Total: 21 files changed+432-120
| @@ -2,14 +2,14 @@ | |||
| 2 | 2 | ||
| 3 | ## 产品支持情况 | 3 | ## 产品支持情况 |
| 4 | 4 | ||
| 5 | -|产品 | 是否支持 | | 5 | +| 产品 | 是否支持 | |
| 6 | -|:-------------------------|:----------:| | 6 | +| :--- | :---: | |
| 7 | -| <term>Ascend 950PR&950DT系列产品</term> | √ | | 7 | +| <term>Ascend 950PR&950DT系列产品</term> | √ | |
| 8 | -| <term>Atlas A3系列产品</term> | √ | | 8 | +| <term>Atlas A3系列产品</term> | √ | |
| 9 | -| <term>Atlas A2系列产品</term> | √ | | 9 | +| <term>Atlas A2系列产品</term> | √ | |
| 10 | -| <term>Atlas 200I/500 A2推理产品</term> | × | | 10 | +| <term>Atlas 200I/500 A2推理产品</term> | × | |
| 11 | -| <term>Atlas推理系列产品</term> | × | | 11 | +| <term>Atlas推理系列产品</term> | × | |
| 12 | -| <term>Atlas训练系列产品</term> | × | | 12 | +| <term>Atlas训练系列产品</term> | × | |
| 13 | 13 | ||
| 14 | ## 功能说明 | 14 | ## 功能说明 |
| 15 | 15 | ||
| @@ -26,49 +26,17 @@ | |||
| 26 | $$ | 26 | $$ |
| 27 | 27 | ||
| 28 | 其中: | 28 | 其中: |
| 29 | + | ||
| 29 | - gradients:对应Relu6操作的反向传播梯度。 | 30 | - gradients:对应Relu6操作的反向传播梯度。 |
| 30 | - features:作为输入传递给对应Relu6操作的特征,或其输出,使用任一者产生相同结果。 | 31 | - features:作为输入传递给对应Relu6操作的特征,或其输出,使用任一者产生相同结果。 |
| 31 | 32 | ||
| 32 | ## 参数说明 | 33 | ## 参数说明 |
| 33 | 34 | ||
| 34 | -<table style="undefined;table-layout: fixed; width: 820px"><colgroup> | 35 | +| 参数名 | 输入/输出/属性 | 描述 | 数据类型 | 数据格式 | |
| 35 | - <col style="width: 100px"> | 36 | +| --- | --- | --- | --- | --- | |
| 36 | - <col style="width: 150px"> | 37 | +| gradients | 输入 | 传递给对应Relu6操作的反向传播梯度,公式中的gradients。 | FLOAT16、FLOAT、BFLOAT16 | ND | |
| 37 | - <col style="width: 190px"> | 38 | +| features | 输入 | 作为输入传递给对应Relu6操作的特征,或其输出,使用任一者产生相同结果,公式中的features。与gradients的shape关系见约束说明。 | FLOAT16、FLOAT、BFLOAT16 | ND | |
| 38 | - <col style="width: 260px"> | 39 | +| backprops | 输出 | 公式中的输出张量,shape由输入shape按各产品的广播约束确定,见约束说明。 | FLOAT16、FLOAT、BFLOAT16 | ND | |
| 39 | - <col style="width: 120px"> | ||
| 40 | - </colgroup> | ||
| 41 | - <thead> | ||
| 42 | - <tr> | ||
| 43 | - <th>参数名</th> | ||
| 44 | - <th>输入/输出/属性</th> | ||
| 45 | - <th>描述</th> | ||
| 46 | - <th>数据类型</th> | ||
| 47 | - <th>数据格式</th> | ||
| 48 | - </tr></thead> | ||
| 49 | - <tbody> | ||
| 50 | - <tr> | ||
| 51 | - <td>gradients</td> | ||
| 52 | - <td>输入</td> | ||
| 53 | - <td>传递给对应Relu6操作的反向传播梯度,公式中的gradients。</td> | ||
| 54 | - <td>FLOAT16、FLOAT、BFLOAT16</td> | ||
| 55 | - <td>ND</td> | ||
| 56 | - </tr> | ||
| 57 | - <tr> | ||
| 58 | - <td>features</td> | ||
| 59 | - <td>输入</td> | ||
| 60 | - <td>作为输入传递给对应Relu6操作的特征,或其输出,使用任一者产生相同结果,公式中的features。与gradients的shape关系见约束说明。</td> | ||
| 61 | - <td>FLOAT16、FLOAT、BFLOAT16</td> | ||
| 62 | - <td>ND</td> | ||
| 63 | - </tr> | ||
| 64 | - <tr> | ||
| 65 | - <td>backprops</td> | ||
| 66 | - <td>输出</td> | ||
| 67 | - <td>公式中的输出张量,shape由输入shape按各产品的广播约束确定,见约束说明。</td> | ||
| 68 | - <td>FLOAT16、FLOAT、BFLOAT16</td> | ||
| 69 | - <td>ND</td> | ||
| 70 | - </tr> | ||
| 71 | - </tbody></table> | ||
| 72 | 40 | ||
| 73 | ## 约束说明 | 41 | ## 约束说明 |
| 74 | 42 | ||
| @@ -124,12 +92,12 @@ In0/In1 ─CopyInBrc─ Cast(->fp32) ─+ | |||
| 124 | 把fp16/bf16整体提升到fp32做cmp/and/sel,末端用`CAST_MODE_RINT` | 92 | 把fp16/bf16整体提升到fp32做cmp/and/sel,末端用`CAST_MODE_RINT` |
| 125 | (round-to-nearest-even)回退到原dtype。原因: | 93 | (round-to-nearest-even)回退到原dtype。原因: |
| 126 | 94 | ||
| 127 | -1. 与DSL `relu6_grad.py`在fp16vcmpsel不可用时fallback到fp32的行为一致; | 95 | +1. 比较和选择操作统一使用fp32中间精度; |
| 128 | 2. 上下界0.0和6.0在fp16/bf16中均能精确表达,但cmp走fp32路径更稳健, | 96 | 2. 上下界0.0和6.0在fp16/bf16中均能精确表达,但cmp走fp32路径更稳健, |
| 129 | 且dy是NaN/Inf时fp32中间不会被半精度范围截断。 | 97 | 且dy是NaN/Inf时fp32中间不会被半精度范围截断。 |
| 130 | 98 | ||
| 131 | ## 调用说明 | 99 | ## 调用说明 |
| 132 | 100 | ||
| 133 | -| 调用方式 | 样例代码 | 说明 | | 101 | +| 调用方式 | 样例代码 | 说明 | |
| 134 | -| ---------------- | --------------------------- |-----------------------------------------------| | 102 | +| --- | --- | --- | |
| 135 | -| GE图模式 | [test_geir_relu6_grad.cpp](examples/arch35/test_geir_relu6_grad.cpp) | 通过[算子IR](op_graph/relu6_grad_proto.h)构图方式调用Relu6Grad算子。 | | 103 | +| GE图模式 | [test_geir_relu6_grad.cpp](examples/arch35/test_geir_relu6_grad.cpp) | 通过[算子IR](op_graph/relu6_grad_proto.h)构图方式调用Relu6Grad算子。 | |
| @@ -18,8 +18,15 @@ ge::graphStatus InferDataType4Relu6Grad(gert::InferDataTypeContext* context) | |||
| 18 | return ge::GRAPH_FAILED; | 18 | return ge::GRAPH_FAILED; |
| 19 | } | 19 | } |
| 20 | const auto dtype = context->GetInputDataType(0); | 20 | const auto dtype = context->GetInputDataType(0); |
| 21 | + OP_CHECK_IF(dtype != ge::DT_FLOAT16 && dtype != ge::DT_FLOAT && dtype != ge::DT_BF16, | ||
| 22 | + OP_LOGE_FOR_INVALID_DTYPE(context->GetNodeName(), "gradients", Ops::Base::ToString(dtype).c_str(), | ||
| 23 | + "FLOAT16, FLOAT, BFLOAT16"), | ||
| 24 | + return ge::GRAPH_FAILED); | ||
| 21 | OP_CHECK_IF(dtype != context->GetInputDataType(1), | 25 | OP_CHECK_IF(dtype != context->GetInputDataType(1), |
| 22 | - OP_LOGE(context->GetNodeName(), "gradients and features must have the same dtype."), | 26 | + OP_LOGE_FOR_INVALID_DTYPES_WITH_REASON( |
| 27 | + context->GetNodeName(), "gradients, features", | ||
| 28 | + (Ops::Base::ToString(dtype) + ", " + Ops::Base::ToString(context->GetInputDataType(1))).c_str(), | ||
| 29 | + "gradients and features must have the same dtype"), | ||
| 23 | return ge::GRAPH_FAILED); | 30 | return ge::GRAPH_FAILED); |
| 24 | return context->SetOutputDataType(0, dtype); | 31 | return context->SetOutputDataType(0, dtype); |
| 25 | } | 32 | } |
| @@ -12,6 +12,7 @@ | |||
| 12 | * \file relu6_grad_infershape.cpp | 12 | * \file relu6_grad_infershape.cpp |
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | +#include <algorithm> | ||
| 15 | #include "op_host/infershape_broadcast_util.h" | 16 | #include "op_host/infershape_broadcast_util.h" |
| 16 | #include "register/op_impl_registry.h" | 17 | #include "register/op_impl_registry.h" |
| 17 | #include "log/log.h" | 18 | #include "log/log.h" |
| @@ -21,7 +22,33 @@ namespace ops { | |||
| 21 | 22 | ||
| 22 | static ge::graphStatus InferShape4Relu6Grad(gert::InferShapeContext* context) | 23 | static ge::graphStatus InferShape4Relu6Grad(gert::InferShapeContext* context) |
| 23 | { | 24 | { |
| 25 | + if (context == nullptr) { | ||
| 26 | + return ge::GRAPH_FAILED; | ||
| 27 | + } | ||
| 24 | OP_LOGD(context->GetNodeName(), "Begin to do InferShape4Relu6Grad in ops-nn"); | 28 | OP_LOGD(context->GetNodeName(), "Begin to do InferShape4Relu6Grad in ops-nn"); |
| 29 | + constexpr size_t MAX_INPUT_DIM_NUM = 8; | ||
| 30 | + const char* inputNames[] = {"gradients", "features"}; | ||
| 31 | + for (size_t input = 0; input < 2; ++input) { | ||
| 32 | + const auto* shape = context->GetInputShape(input); | ||
| 33 | + OP_CHECK_NULL_WITH_CONTEXT(context, shape); | ||
| 34 | + OP_CHECK_IF(shape->GetDimNum() > MAX_INPUT_DIM_NUM, | ||
| 35 | + OP_LOGE_FOR_INVALID_SHAPEDIM(context->GetNodeName(), inputNames[input], | ||
| 36 | + std::to_string(shape->GetDimNum()).c_str(), "0 ~ 8 dims"), | ||
| 37 | + return ge::GRAPH_FAILED); | ||
| 38 | + } | ||
| 39 | + const auto* gradients = context->GetInputShape(0); | ||
| 40 | + const auto* features = context->GetInputShape(1); | ||
| 41 | + const size_t commonRank = std::min(gradients->GetDimNum(), features->GetDimNum()); | ||
| 42 | + for (size_t offset = 1; offset <= commonRank; ++offset) { | ||
| 43 | + const int64_t gradDim = gradients->GetDim(gradients->GetDimNum() - offset); | ||
| 44 | + const int64_t featDim = features->GetDim(features->GetDimNum() - offset); | ||
| 45 | + OP_CHECK_IF(gradDim >= 0 && featDim >= 0 && gradDim != featDim && gradDim != 1 && featDim != 1, | ||
| 46 | + OP_LOGE_FOR_INVALID_SHAPES_WITH_REASON( | ||
| 47 | + context->GetNodeName(), "gradients, features", | ||
| 48 | + Ops::Base::ToString(std::vector<const gert::Shape*>{gradients, features}).c_str(), | ||
| 49 | + "aligned dimensions must be equal or one of them must be 1"), | ||
| 50 | + return ge::GRAPH_FAILED); | ||
| 51 | + } | ||
| 25 | return Ops::Base::InferShape4Broadcast(context); | 52 | return Ops::Base::InferShape4Broadcast(context); |
| 26 | } | 53 | } |
| 27 | 54 | ||
| @@ -60,9 +60,8 @@ struct Relu6Grad { | |||
| 60 | }; | 60 | }; |
| 61 | 61 | ||
| 62 | // Promoted path for half / bfloat16_t: cast inputs to PromoteT (float) for | 62 | // Promoted path for half / bfloat16_t: cast inputs to PromoteT (float) for |
| 63 | -// cmp/and/sel to (a) match the DSL fallback in relu6_grad.py that lifts to | 63 | +// cmp/and/sel to keep boundary comparison against 0 and 6 stable across |
| 64 | -// fp32 when fp16 vcmpsel is unavailable, (b) keep boundary comparison | 64 | +// half/bf16. Result is cast back to T with |
| 65 | -// against 0 and 6 stable across half/bf16. Result is cast back to T with | ||
| 66 | // round-to-nearest-even. | 65 | // round-to-nearest-even. |
| 67 | template <typename T, typename PromoteT> | 66 | template <typename T, typename PromoteT> |
| 68 | struct Relu6GradFloatCast { | 67 | struct Relu6GradFloatCast { |
| @@ -0,0 +1,7 @@ | |||
| 1 | +testcase_name,network_name,op_name,input_shapes,input_dtypes,input_formats,output_shapes,output_dtypes,output_formats,input_ori_shapes,input_ori_formats,output_ori_shapes,output_ori_formats,attributes,input_data_ranges,precision_tolerances,absolute_precision,output_inplace_indexes,output_shape_unknown_indexes,is_enabled,remark,soc_series,priority,dump_file_prefix,manual_input_binaries,manual_golden_binaries | ||
| 2 | +Relu6Grad_st_fp32_tail,UNKNOWN,Relu6Grad,"((33,), (33,))","('float32', 'float32')","('ND', 'ND')","((33,),)","('float32',)","('ND',)","((33,), (33,))","('ND', 'ND')","((33,),)","('ND',)",{},"((-2, 2), (-1, 7))",,1e-08,(),(),True,equal shapes and unaligned tail,,0,,(),() | ||
| 3 | +Relu6Grad_st_fp16_zero_broadcast,UNKNOWN,Relu6Grad,"((1, 3), (2, 1))","('float16', 'float16')","('ND', 'ND')","((2, 3),)","('float16',)","('ND',)","((1, 3), (2, 1))","('ND', 'ND')","((2, 3),)","('ND',)",{},"((-2, 2), (0, 0))",,1e-08,(),(),True,bidirectional broadcast and lower endpoint,,0,,(),() | ||
| 4 | +Relu6Grad_st_bf16_six,UNKNOWN,Relu6Grad,"((2, 3), (2, 3))","('bfloat16', 'bfloat16')","('ND', 'ND')","((2, 3),)","('bfloat16',)","('ND',)","((2, 3), (2, 3))","('ND', 'ND')","((2, 3),)","('ND',)",{},"((-2, 2), (6, 6))",,1e-08,(),(),True,upper endpoint,,0,,(),() | ||
| 5 | +Relu6Grad_st_bf16_rank8,UNKNOWN,Relu6Grad,"((1, 1, 1, 1, 1, 1, 2, 1), (1, 1, 1, 1, 1, 1, 1, 3))","('bfloat16', 'bfloat16')","('ND', 'ND')","((1, 1, 1, 1, 1, 1, 2, 3),)","('bfloat16',)","('ND',)","((1, 1, 1, 1, 1, 1, 2, 1), (1, 1, 1, 1, 1, 1, 1, 3))","('ND', 'ND')","((1, 1, 1, 1, 1, 1, 2, 3),)","('ND',)",{},"((2, 2), (4, 4))",,1e-08,(),(),True,maximum rank and broadcast with nonzero output,,0,,(),() | ||
| 6 | +Relu6Grad_st_fp16_feature_one,UNKNOWN,Relu6Grad,"((17,), (1,))","('float16', 'float16')","('ND', 'ND')","((17,),)","('float16',)","('ND',)","((17,), (1,))","('ND', 'ND')","((17,),)","('ND',)",{},"((-2, 2), (4, 4))",,1e-08,(),(),True,single-element feature broadcast,,0,,(),() | ||
| 7 | +Relu6Grad_st_fp32_inf_outside,UNKNOWN,Relu6Grad,"((3,), (3,))","('float32', 'float32')","('ND', 'ND')","((3,),)","('float32',)","('ND',)","((3,), (3,))","('ND', 'ND')","((3,),)","('ND',)",{},"((float('inf'), float('inf')), (-1, -1))",,1e-08,(),(),True,Inf gradient outside interval selects zero,,0,,(),() | ||
| @@ -45,3 +45,10 @@ TEST(Relu6GradGraphInfer, RejectsMismatchedInputsWithoutLegacyVerifier) | |||
| 45 | CheckDtype(ge::DT_FLOAT16, ge::DT_FLOAT, ge::GRAPH_FAILED); | 45 | CheckDtype(ge::DT_FLOAT16, ge::DT_FLOAT, ge::GRAPH_FAILED); |
| 46 | CheckDtype(ge::DT_BF16, ge::DT_FLOAT16, ge::GRAPH_FAILED); | 46 | CheckDtype(ge::DT_BF16, ge::DT_FLOAT16, ge::GRAPH_FAILED); |
| 47 | } | 47 | } |
| 48 | + | ||
| 49 | +TEST(Relu6GradGraphInfer, RejectsUnsupportedMatchingDataTypes) | ||
| 50 | +{ | ||
| 51 | + CheckDtype(ge::DT_INT32, ge::DT_INT32, ge::GRAPH_FAILED); | ||
| 52 | + CheckDtype(ge::DT_BOOL, ge::DT_BOOL, ge::GRAPH_FAILED); | ||
| 53 | + CheckDtype(ge::DT_DOUBLE, ge::DT_DOUBLE, ge::GRAPH_FAILED); | ||
| 54 | +} | ||
| @@ -27,6 +27,23 @@ protected: | |||
| 27 | static void TearDownTestCase() { std::cout << "relu6grad TearDown" << std::endl; } | 27 | static void TearDownTestCase() { std::cout << "relu6grad TearDown" << std::endl; } |
| 28 | }; | 28 | }; |
| 29 | 29 | ||
| 30 | +TEST_F(relu6grad, relu6grad_infer_shape_rank9_rejected_for_each_input) | ||
| 31 | +{ | ||
| 32 | + auto inferShapeFunc = gert::OpImplRegistry::GetInstance().GetOpImpl("Relu6Grad")->infer_shape; | ||
| 33 | + gert::StorageShape rank9 = {{1, 1, 1, 1, 1, 1, 1, 1, 1}, {1, 1, 1, 1, 1, 1, 1, 1, 1}}; | ||
| 34 | + gert::StorageShape scalar = {{1}, {1}}; | ||
| 35 | + for (size_t input = 0; input < 2; ++input) { | ||
| 36 | + gert::StorageShape output = {{}, {}}; | ||
| 37 | + auto holder = gert::InferShapeContextFaker() | ||
| 38 | + .NodeIoNum(2, 1) | ||
| 39 | + .IrInstanceNum({1, 1}) | ||
| 40 | + .InputShapes({input == 0 ? &rank9 : &scalar, input == 1 ? &rank9 : &scalar}) | ||
| 41 | + .OutputShapes({&output}) | ||
| 42 | + .Build(); | ||
| 43 | + EXPECT_EQ(inferShapeFunc(holder.GetContext<gert::InferShapeContext>()), ge::GRAPH_FAILED); | ||
| 44 | + } | ||
| 45 | +} | ||
| 46 | + | ||
| 30 | TEST_F(relu6grad, relu6grad_infer_shape_same_fp32) | 47 | TEST_F(relu6grad, relu6grad_infer_shape_same_fp32) |
| 31 | { | 48 | { |
| 32 | auto inferShapeFunc = gert::OpImplRegistry::GetInstance().GetOpImpl("Relu6Grad")->infer_shape; | 49 | auto inferShapeFunc = gert::OpImplRegistry::GetInstance().GetOpImpl("Relu6Grad")->infer_shape; |
| @@ -3,9 +3,9 @@ | |||
| 3 | ## 产品支持情况 | 3 | ## 产品支持情况 |
| 4 | 4 | ||
| 5 | | 产品 | 是否支持 | | 5 | | 产品 | 是否支持 | |
| 6 | -| :----------------------------------------- | :------: | | 6 | +| :--- | :---: | |
| 7 | | <term>Ascend 950PR&950DT系列产品</term> | √ | | 7 | | <term>Ascend 950PR&950DT系列产品</term> | √ | |
| 8 | -| <term>Atlas A3系列产品</term>| √ | | 8 | +| <term>Atlas A3系列产品</term> | √ | |
| 9 | | <term>Atlas A2系列产品</term> | √ | | 9 | | <term>Atlas A2系列产品</term> | √ | |
| 10 | | <term>Atlas 200I/500 A2推理产品</term> | √ | | 10 | | <term>Atlas 200I/500 A2推理产品</term> | √ | |
| 11 | | <term>Atlas推理系列产品</term> | √ | | 11 | | <term>Atlas推理系列产品</term> | √ | |
| @@ -61,7 +61,7 @@ | |||
| 61 | | accum | 输入 | 动量累积量,必须为可写Tensor;shape、数据类型和数据格式与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | | 61 | | accum | 输入 | 动量累积量,必须为可写Tensor;shape、数据类型和数据格式与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | |
| 62 | | momentum | 输入 | 动量因子,元素个数必须为1,数据类型与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | ND | | 62 | | momentum | 输入 | 动量因子,元素个数必须为1,数据类型与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | ND | |
| 63 | | stat | 输入 | 逐元素首步标记,取值1表示不施加阻尼;必须为可写Tensor,shape、数据类型和数据格式与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | | 63 | | stat | 输入 | 逐元素首步标记,取值1表示不施加阻尼;必须为可写Tensor,shape、数据类型和数据格式与`parameters`一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | |
| 64 | -| parameters | 输出 | 更新后的权重,与输入`parameters`共享内存,shape、数据类型和数据格式保持一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | | 64 | +| parameters | 输出 | 更新后的权重,shape、数据类型和数据格式与输入`parameters`保持一致。 | FLOAT、FLOAT16、BFLOAT16 | NC1HWC0、NDC1HWC0、ND、FRACTAL_Z、FRACTAL_Z_3D | |
| 65 | | dampening | 可选属性 | 动量阻尼系数,默认值为0.0;当`nesterov=true`时必须为0。 | FLOAT | - | | 65 | | dampening | 可选属性 | 动量阻尼系数,默认值为0.0;当`nesterov=true`时必须为0。 | FLOAT | - | |
| 66 | | weight_decay | 可选属性 | 权重衰减系数,默认值为0.0,取值必须大于或等于0。 | FLOAT | - | | 66 | | weight_decay | 可选属性 | 权重衰减系数,默认值为0.0,取值必须大于或等于0。 | FLOAT | - | |
| 67 | | nesterov | 可选属性 | 是否启用Nesterov动量,默认值为false。 | BOOL | - | | 67 | | nesterov | 可选属性 | 是否启用Nesterov动量,默认值为false。 | BOOL | - | |
| @@ -69,11 +69,12 @@ | |||
| 69 | ## 约束说明 | 69 | ## 约束说明 |
| 70 | 70 | ||
| 71 | - 图上仅声明一个`parameters`输出,但`parameters`、`accum`和`stat`均可能被原地更新, | 71 | - 图上仅声明一个`parameters`输出,但`parameters`、`accum`和`stat`均可能被原地更新, |
| 72 | - 调用方必须将这三个输入作为可写Tensor。 | 72 | + 调用方必须将这三个输入作为可写Tensor。权重更新值应从`parameters`输出读取。 |
| 73 | - `momentum=0`仅阻止`accum`和`stat`回写,不跳过权重更新中的动量乘法;当`accum`含Inf时,`0×Inf`产生的NaN会传播到权重结果。 | 73 | - `momentum=0`仅阻止`accum`和`stat`回写,不跳过权重更新中的动量乘法;当`accum`含Inf时,`0×Inf`产生的NaN会传播到权重结果。 |
| 74 | - <term>Ascend 950PR&950DT系列产品</term>: | 74 | - <term>Ascend 950PR&950DT系列产品</term>: |
| 75 | - 数据类型:各输入、输出支持FLOAT、FLOAT16、BFLOAT16。 | 75 | - 数据类型:各输入、输出支持FLOAT、FLOAT16、BFLOAT16。 |
| 76 | - 数据格式:静态shape和动态shape的各输入、输出均为ND。 | 76 | - 数据格式:静态shape和动态shape的各输入、输出均为ND。 |
| 77 | + - GE图输入按ND连续布局提供数据;转置、切片等非连续视图需由调用方或框架先转换为连续Tensor。 | ||
| 77 | - 动态shape能力:支持动态shape和动态rank。 | 78 | - 动态shape能力:支持动态shape和动态rank。 |
| 78 | - 输入shape约束:`parameters`的rank为1~8,不支持空Tensor。 | 79 | - 输入shape约束:`parameters`的rank为1~8,不支持空Tensor。 |
| 79 | - 中间计算精度:使用FLOAT进行中间计算,FLOAT16、BFLOAT16结果按就近偶数舍入转换回输入数据类型。 | 80 | - 中间计算精度:使用FLOAT进行中间计算,FLOAT16、BFLOAT16结果按就近偶数舍入转换回输入数据类型。 |
| @@ -54,7 +54,7 @@ using std::string; | |||
| 54 | using std::vector; | 54 | using std::vector; |
| 55 | #define ADD_INPUT(intputIndex, intputName, intputDtype, inputShape, inputValue) \ | 55 | #define ADD_INPUT(intputIndex, intputName, intputDtype, inputShape, inputValue) \ |
| 56 | vector<int64_t> placeholder##intputIndex##_shape = inputShape; \ | 56 | vector<int64_t> placeholder##intputIndex##_shape = inputShape; \ |
| 57 | - auto placeholder##intputIndex = op::Data("placeholder" + std::to_string(intputIndex)) \ | 57 | + auto placeholder##intputIndex = op::Data(("placeholder" + std::to_string(intputIndex)).c_str()) \ |
| 58 | .set_attr_index(intputIndex - 1); \ | 58 | .set_attr_index(intputIndex - 1); \ |
| 59 | TensorDesc placeholder##intputIndex##_desc = TensorDesc(ge::Shape(placeholder##intputIndex##_shape), FORMAT_ND, \ | 59 | TensorDesc placeholder##intputIndex##_desc = TensorDesc(ge::Shape(placeholder##intputIndex##_shape), FORMAT_ND, \ |
| 60 | intputDtype); \ | 60 | intputDtype); \ |
| @@ -78,7 +78,7 @@ using std::vector; | |||
| 78 | 78 | ||
| 79 | #define ADD_CONST_INPUT(intputIndex, intputName, intputDtype, inputShape) \ | 79 | #define ADD_CONST_INPUT(intputIndex, intputName, intputDtype, inputShape) \ |
| 80 | vector<int64_t> placeholder##intputIndex##_shape = inputShape; \ | 80 | vector<int64_t> placeholder##intputIndex##_shape = inputShape; \ |
| 81 | - auto placeholder##intputIndex = op::Const("placeholder" + std::to_string(intputIndex)); \ | 81 | + auto placeholder##intputIndex = op::Const(("placeholder" + std::to_string(intputIndex)).c_str()); \ |
| 82 | TensorDesc placeholder##intputIndex##_desc = TensorDesc(ge::Shape(placeholder##intputIndex##_shape), FORMAT_ND, \ | 82 | TensorDesc placeholder##intputIndex##_desc = TensorDesc(ge::Shape(placeholder##intputIndex##_shape), FORMAT_ND, \ |
| 83 | intputDtype); \ | 83 | intputDtype); \ |
| 84 | placeholder##intputIndex##_desc.SetPlacement(ge::kPlacementHost); \ | 84 | placeholder##intputIndex##_desc.SetPlacement(ge::kPlacementHost); \ |
| @@ -328,7 +328,13 @@ int main(int argc, char* argv[]) | |||
| 328 | printf("%s - INFO - [XIR]: Session add ir compute graph to ir session success\n", GetTime().c_str()); | 328 | printf("%s - INFO - [XIR]: Session add ir compute graph to ir session success\n", GetTime().c_str()); |
| 329 | printf("%s - INFO - [XIR]: dump graph to txt\n", GetTime().c_str()); | 329 | printf("%s - INFO - [XIR]: dump graph to txt\n", GetTime().c_str()); |
| 330 | std::string file_path = "./dump"; | 330 | std::string file_path = "./dump"; |
| 331 | - aclgrphDumpGraph(graph, file_path.c_str(), file_path.length()); | 331 | + ret = aclgrphDumpGraph(graph, file_path.c_str(), file_path.length()); |
| 332 | + if (ret != SUCCESS) { | ||
| 333 | + printf("%s - ERROR - [XIR]: Dump graph failed\n", GetTime().c_str()); | ||
| 334 | + delete session; | ||
| 335 | + GEFinalize(); | ||
| 336 | + return FAILED; | ||
| 337 | + } | ||
| 332 | printf("%s - INFO - [XIR]: Start to run ir compute graph\n", GetTime().c_str()); | 338 | printf("%s - INFO - [XIR]: Start to run ir compute graph\n", GetTime().c_str()); |
| 333 | std::vector<ge::Tensor> output; | 339 | std::vector<ge::Tensor> output; |
| 334 | ret = session->RunGraph(graph_id, input, output); | 340 | ret = session->RunGraph(graph_id, input, output); |
| @@ -22,13 +22,29 @@ namespace { | |||
| 22 | constexpr size_t PARAMETERS_INDEX = 0; | 22 | constexpr size_t PARAMETERS_INDEX = 0; |
| 23 | } // namespace | 23 | } // namespace |
| 24 | 24 | ||
| 25 | -static ge::graphStatus InferDataTypeForSgd(gert::InferDataTypeContext* context) | 25 | +ge::graphStatus InferDataTypeForSgd(gert::InferDataTypeContext* context) |
| 26 | { | 26 | { |
| 27 | + if (context == nullptr) { | ||
| 28 | + return ge::GRAPH_FAILED; | ||
| 29 | + } | ||
| 27 | OP_LOGD(context, "InferDataTypeForSgd begin."); | 30 | OP_LOGD(context, "InferDataTypeForSgd begin."); |
| 28 | - // accum / stat 不是图输出(靠覆写输入 GM 原地回写),故此处只有一路。 | 31 | + const auto dtype = context->GetInputDataType(PARAMETERS_INDEX); |
| 29 | - context->SetOutputDataType(PARAMETERS_INDEX, context->GetInputDataType(PARAMETERS_INDEX)); | 32 | + OP_CHECK_IF(dtype != ge::DT_FLOAT16 && dtype != ge::DT_FLOAT && dtype != ge::DT_BF16, |
| 33 | + OP_LOGE_FOR_INVALID_DTYPE(context->GetNodeName(), "parameters", Ops::Base::ToString(dtype).c_str(), | ||
| 34 | + "FLOAT16, FLOAT, BFLOAT16"), | ||
| 35 | + return ge::GRAPH_FAILED); | ||
| 36 | + const char* inputNames[] = {"parameters", "gradient", "learning_rate", "accum", "momentum", "stat"}; | ||
| 37 | + for (size_t input = 1; input < 6; ++input) { | ||
| 38 | + const auto inputDtype = context->GetInputDataType(input); | ||
| 39 | + OP_CHECK_IF(inputDtype != dtype, | ||
| 40 | + OP_LOGE_FOR_INVALID_DTYPES_WITH_REASON( | ||
| 41 | + context->GetNodeName(), (std::string("parameters, ") + inputNames[input]).c_str(), | ||
| 42 | + (Ops::Base::ToString(dtype) + ", " + Ops::Base::ToString(inputDtype)).c_str(), | ||
| 43 | + "all inputs must have the same dtype as parameters"), | ||
| 44 | + return ge::GRAPH_FAILED); | ||
| 45 | + } | ||
| 30 | OP_LOGD(context, "InferDataTypeForSgd end."); | 46 | OP_LOGD(context, "InferDataTypeForSgd end."); |
| 31 | - return ge::GRAPH_SUCCESS; | 47 | + return context->SetOutputDataType(PARAMETERS_INDEX, dtype); |
| 32 | } | 48 | } |
| 33 | 49 | ||
| 34 | IMPL_OP(SGD).InferDataType(InferDataTypeForSgd); | 50 | IMPL_OP(SGD).InferDataType(InferDataTypeForSgd); |
| @@ -13,6 +13,7 @@ | |||
| 13 | * \brief | 13 | * \brief |
| 14 | */ | 14 | */ |
| 15 | #include "sgd_tiling.h" | 15 | #include "sgd_tiling.h" |
| 16 | +#include <limits> | ||
| 16 | #include <graph/utils/type_utils.h> | 17 | #include <graph/utils/type_utils.h> |
| 17 | #include "error_util.h" | 18 | #include "error_util.h" |
| 18 | #include "register/op_def_registry.h" | 19 | #include "register/op_def_registry.h" |
| @@ -59,7 +60,7 @@ ge::graphStatus SgdRegbaseTiling::GetAttr() | |||
| 59 | nesterov_ = (nesterovAttr != nullptr) ? *nesterovAttr : false; | 60 | nesterov_ = (nesterovAttr != nullptr) ? *nesterovAttr : false; |
| 60 | 61 | ||
| 61 | // 对齐 ascend910b 的两条属性语义校验(canndev nn_training_ops.cc:1932-1955)。 | 62 | // 对齐 ascend910b 的两条属性语义校验(canndev nn_training_ops.cc:1932-1955)。 |
| 62 | - // 走结构化上报宏(R6),非裸 OP_LOGE。 | 63 | + // 属性非法时使用结构化错误上报。 |
| 63 | OP_CHECK_IF(nesterov_ && dampening_ != 0.0f, | 64 | OP_CHECK_IF(nesterov_ && dampening_ != 0.0f, |
| 64 | OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(tilingContext_->GetNodeName(), "dampening", | 65 | OP_LOGE_FOR_INVALID_VALUE_WITH_REASON(tilingContext_->GetNodeName(), "dampening", |
| 65 | std::to_string(dampening_).c_str(), | 66 | std::to_string(dampening_).c_str(), |
| @@ -71,7 +72,7 @@ ge::graphStatus SgdRegbaseTiling::GetAttr() | |||
| 71 | "attr weight_decay must be more than or equal to 0"), | 72 | "attr weight_decay must be more than or equal to 0"), |
| 72 | return ge::GRAPH_FAILED); | 73 | return ge::GRAPH_FAILED); |
| 73 | 74 | ||
| 74 | - // 属性"是否为 0"固化为编译期分支(spec numerical_stability.skip_zero_branches): | 75 | + // 属性"是否为 0"固化为编译期分支: |
| 75 | // 条件不成立时【真正跳过】该子图,而不是乘 0 —— 0 * inf = NaN 会污染结果。 | 76 | // 条件不成立时【真正跳过】该子图,而不是乘 0 —— 0 * inf = NaN 会污染结果。 |
| 76 | useNesterovKey_ = nesterov_ ? 1U : 0U; | 77 | useNesterovKey_ = nesterov_ ? 1U : 0U; |
| 77 | hasWeightDecayKey_ = (weightDecay_ != 0.0f) ? 1U : 0U; | 78 | hasWeightDecayKey_ = (weightDecay_ != 0.0f) ? 1U : 0U; |
| @@ -91,13 +92,22 @@ ge::graphStatus SgdRegbaseTiling::CheckRank(const gert::Shape& input0Shape) | |||
| 91 | 92 | ||
| 92 | ge::graphStatus SgdRegbaseTiling::CheckNotEmpty(const gert::Shape& input0Shape) | 93 | ge::graphStatus SgdRegbaseTiling::CheckNotEmpty(const gert::Shape& input0Shape) |
| 93 | { | 94 | { |
| 94 | - // 空 Tensor(任意一轴或多轴为 0)按 spec error_codes 归 null_input,拒绝为非法。 | 95 | + int64_t elementCount = 1; |
| 95 | - // 本算子无"空进空出"语义 —— accum / stat 的原地回写在 numel == 0 下无定义。 | 96 | + for (size_t axis = 0; axis < input0Shape.GetDimNum(); ++axis) { |
| 96 | - OP_CHECK_IF(input0Shape.GetShapeSize() == 0, | 97 | + const int64_t dim = input0Shape.GetDim(axis); |
| 97 | - OP_LOGE_FOR_INVALID_SHAPESIZE_WITH_REASON(tilingContext_->GetNodeName(), "parameters", | 98 | + OP_CHECK_IF( |
| 98 | - std::to_string(input0Shape.GetShapeSize()).c_str(), | 99 | + dim == 0, |
| 99 | - "empty tensor (any axis being 0) is not supported by SGD"), | 100 | + OP_LOGE_FOR_INVALID_SHAPESIZE_WITH_REASON(tilingContext_->GetNodeName(), "parameters", "0", |
| 100 | - return ge::GRAPH_FAILED); | 101 | + "empty tensor (any axis being 0) is not supported by SGD"), |
| 102 | + return ge::GRAPH_FAILED); | ||
| 103 | + // Tiling receives concrete runtime dimensions, unlike InferShape. | ||
| 104 | + OP_CHECK_IF(dim < 0 || elementCount > std::numeric_limits<int64_t>::max() / dim, | ||
| 105 | + OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON( | ||
| 106 | + tilingContext_->GetNodeName(), "parameters", Ops::Base::ToString(input0Shape).c_str(), | ||
| 107 | + "runtime dimensions must be positive and their product must fit in int64"), | ||
| 108 | + return ge::GRAPH_FAILED); | ||
| 109 | + elementCount *= dim; | ||
| 110 | + } | ||
| 101 | return ge::GRAPH_SUCCESS; | 111 | return ge::GRAPH_SUCCESS; |
| 102 | } | 112 | } |
| 103 | 113 | ||
| @@ -120,7 +130,6 @@ ge::graphStatus SgdRegbaseTiling::CheckSameShape(int32_t inputIdx, const gert::S | |||
| 120 | auto inputShape = tilingContext_->GetInputShape(inputIdx); | 130 | auto inputShape = tilingContext_->GetInputShape(inputIdx); |
| 121 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext_, inputShape); | 131 | OP_CHECK_NULL_WITH_CONTEXT(tilingContext_, inputShape); |
| 122 | // 严格相等,不做广播 —— "可广播但不相等"(如 [2,3] vs [1,3])同样判非法, | 132 | // 严格相等,不做广播 —— "可广播但不相等"(如 [2,3] vs [1,3])同样判非法, |
| 123 | - // 对齐 spec 的 shape_mismatch 口径。 | ||
| 124 | if (inputShape->GetStorageShape() != input0Shape) { | 133 | if (inputShape->GetStorageShape() != input0Shape) { |
| 125 | return ge::GRAPH_FAILED; | 134 | return ge::GRAPH_FAILED; |
| 126 | } | 135 | } |
| @@ -149,7 +158,7 @@ ge::graphStatus SgdRegbaseTiling::CheckShapeAndType() | |||
| 149 | 158 | ||
| 150 | // 检查全部输入的shape与dtype;不能仅依赖parameters的rank校验。 | 159 | // 检查全部输入的shape与dtype;不能仅依赖parameters的rank校验。 |
| 151 | // 内层 CheckRank 已用 OP_LOGE_FOR_INVALID_SHAPEDIM 结构化上报,此处只作调用链留痕, | 160 | // 内层 CheckRank 已用 OP_LOGE_FOR_INVALID_SHAPEDIM 结构化上报,此处只作调用链留痕, |
| 152 | - // 不重复打 ERROR(R6 要求的是"校验必须走结构化宏",不是"每层都再报一次")。 | 161 | + // 不重复上报同一错误。 |
| 153 | OP_CHECK_IF(CheckRank(inputStorageShape) != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "rank check failed"), | 162 | OP_CHECK_IF(CheckRank(inputStorageShape) != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "rank check failed"), |
| 154 | return ge::GRAPH_FAILED); | 163 | return ge::GRAPH_FAILED); |
| 155 | OP_CHECK_IF(CheckNotEmpty(inputStorageShape) != ge::GRAPH_SUCCESS, | 164 | OP_CHECK_IF(CheckNotEmpty(inputStorageShape) != ge::GRAPH_SUCCESS, |
| @@ -288,7 +297,6 @@ ge::graphStatus SgdRegbaseTiling::RunTiling() | |||
| 288 | } | 297 | } |
| 289 | // GetAttr / CheckShapeAndType 内部对非法输入已用 OP_LOGE_FOR_INVALID_* 结构化上报, | 298 | // GetAttr / CheckShapeAndType 内部对非法输入已用 OP_LOGE_FOR_INVALID_* 结构化上报, |
| 290 | // 此处只作调用链留痕、不重复打 ERROR —— 与上面 CheckRank / CheckNotEmpty 的处理一致。 | 299 | // 此处只作调用链留痕、不重复打 ERROR —— 与上面 CheckRank / CheckNotEmpty 的处理一致。 |
| 291 | - // R6 要求的是"校验必须走结构化宏",不是"每层都再报一次"。 | ||
| 292 | OP_CHECK_IF(GetAttr() != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "Get attr failed."), return ge::GRAPH_FAILED); | 300 | OP_CHECK_IF(GetAttr() != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "Get attr failed."), return ge::GRAPH_FAILED); |
| 293 | OP_CHECK_IF(CheckShapeAndType() != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "Shape and dtype check failed."), | 301 | OP_CHECK_IF(CheckShapeAndType() != ge::GRAPH_SUCCESS, OP_LOGD(tilingContext_, "Shape and dtype check failed."), |
| 294 | return ge::GRAPH_FAILED); | 302 | return ge::GRAPH_FAILED); |
| @@ -10,7 +10,7 @@ | |||
| 10 | 10 | ||
| 11 | /*! | 11 | /*! |
| 12 | * \file sgd_def.cpp | 12 | * \file sgd_def.cpp |
| 13 | - * \brief SGD OpDef —— 仅适配 Ascend950(arch35 / DAV_3510 / regbase) | 13 | + * \brief SGD OpDef(arch35 / regbase) |
| 14 | * | 14 | * |
| 15 | * 支持面对齐 910B/910C 基线(canndev aic-ascend910b-ops-info.ini 的 [SGD] 段), | 15 | * 支持面对齐 910B/910C 基线(canndev aic-ascend910b-ops-info.ini 的 [SGD] 段), |
| 16 | * 唯一收窄项是 format:ascend910b 大张量支持 NC1HWC0 / NDC1HWC0 / ND / FRACTAL_Z / | 16 | * 唯一收窄项是 format:ascend910b 大张量支持 NC1HWC0 / NDC1HWC0 / ND / FRACTAL_Z / |
| @@ -19,9 +19,7 @@ | |||
| 19 | * apply_centered_rms_prop)的 def.cpp 一律只声明 ge::FORMAT_ND,无一例声明私有 | 19 | * apply_centered_rms_prop)的 def.cpp 一律只声明 ge::FORMAT_ND,无一例声明私有 |
| 20 | * format。当前支持范围及产品差异见算子 README。 | 20 | * format。当前支持范围及产品差异见算子 README。 |
| 21 | * | 21 | * |
| 22 | - * R2(ascend950 不碰 ascend910b)天然满足:ops-nn 仓内不存在 SGD 算子(optim/fused_sgd 是语义 | 22 | + * 本文件通过 AddConfig 注册 ascend950 与 ascend350 配置,不修改其他产品的配置。 |
| 23 | - * 不同的另一个算子,且无 arch35 目录),本算子为净新增,只 AddConfig("ascend950"), | ||
| 24 | - * 不触碰任何 ascend910b 配置;apply_momentum 另有 ascend350 等配置,不代表本算子的注册范围。 | ||
| 25 | * | 23 | * |
| 26 | * accum / stat 【不声明为 Output】—— 与 ascend910b 形态一致(canndev proto | 24 | * accum / stat 【不声明为 Output】—— 与 ascend910b 形态一致(canndev proto |
| 27 | * nn_training_ops.h:1431 亦只有一个 OUTPUT),两者靠覆写输入 GM 原地回写。 | 25 | * nn_training_ops.h:1431 亦只有一个 OUTPUT),两者靠覆写输入 GM 原地回写。 |
| @@ -39,10 +39,67 @@ constexpr int64_t UNKNOWN_RANK_DIM = -2; | |||
| 39 | 39 | ||
| 40 | // UNKNOWN_RANK 在 GE 下表现为 dims == {-2},须透传而非按 rank 拒绝。 | 40 | // UNKNOWN_RANK 在 GE 下表现为 dims == {-2},须透传而非按 rank 拒绝。 |
| 41 | bool IsUnknownRank(const gert::Shape* shape) { return shape->GetDimNum() == 1 && shape->GetDim(0) == UNKNOWN_RANK_DIM; } | 41 | bool IsUnknownRank(const gert::Shape* shape) { return shape->GetDimNum() == 1 && shape->GetDim(0) == UNKNOWN_RANK_DIM; } |
| 42 | + | ||
| 43 | +ge::graphStatus CheckInputShapes(gert::InferShapeContext* context, const gert::Shape* paramShape) | ||
| 44 | +{ | ||
| 45 | + const char* inputNames[] = {"parameters", "gradient", "learning_rate", "accum", "momentum", "stat"}; | ||
| 46 | + for (size_t input = 0; input < 6; ++input) { | ||
| 47 | + const auto* shape = context->GetInputShape(input); | ||
| 48 | + OP_CHECK_NULL_WITH_CONTEXT(context, shape); | ||
| 49 | + if (IsUnknownRank(shape)) { | ||
| 50 | + continue; | ||
| 51 | + } | ||
| 52 | + const bool isScalarInput = input == 2 || input == 4; | ||
| 53 | + const size_t rank = shape->GetDimNum(); | ||
| 54 | + if (!isScalarInput) { | ||
| 55 | + OP_CHECK_IF(rank < MIN_DIM_NUM || rank > MAX_DIM_NUM, | ||
| 56 | + OP_LOGE_FOR_INVALID_SHAPEDIM(context->GetNodeName(), inputNames[input], | ||
| 57 | + std::to_string(rank).c_str(), "1 ~ 8 dims"), | ||
| 58 | + return ge::GRAPH_FAILED); | ||
| 59 | + } | ||
| 60 | + for (size_t axis = 0; axis < rank; ++axis) { | ||
| 61 | + const int64_t dim = shape->GetDim(axis); | ||
| 62 | + OP_CHECK_IF(dim == 0 && !isScalarInput, | ||
| 63 | + OP_LOGE_FOR_INVALID_SHAPESIZE_WITH_REASON(context->GetNodeName(), inputNames[input], "0", | ||
| 64 | + "empty tensors are not supported"), | ||
| 65 | + return ge::GRAPH_FAILED); | ||
| 66 | + OP_CHECK_IF(isScalarInput && (dim == 0 || dim > 1), | ||
| 67 | + OP_LOGE_FOR_INVALID_SHAPE_WITH_REASON(context->GetNodeName(), inputNames[input], | ||
| 68 | + Ops::Base::ToString(*shape).c_str(), | ||
| 69 | + "must contain exactly one element"), | ||
| 70 | + return ge::GRAPH_FAILED); | ||
| 71 | + } | ||
| 72 | + // Unknown dimensions and unknown rank are resolved at runtime. Only | ||
| 73 | + // reject shape relationships that are already known to be invalid. | ||
| 74 | + if (isScalarInput || IsUnknownRank(paramShape)) { | ||
| 75 | + continue; | ||
| 76 | + } | ||
| 77 | + OP_CHECK_IF(rank != paramShape->GetDimNum(), | ||
| 78 | + OP_LOGE_FOR_INVALID_SHAPES_WITH_REASON( | ||
| 79 | + context->GetNodeName(), (std::string("parameters, ") + inputNames[input]).c_str(), | ||
| 80 | + Ops::Base::ToString(std::vector<const gert::Shape*>{paramShape, shape}).c_str(), | ||
| 81 | + "must have the same shape as parameters"), | ||
| 82 | + return ge::GRAPH_FAILED); | ||
| 83 | + for (size_t axis = 0; axis < rank; ++axis) { | ||
| 84 | + const int64_t dim = shape->GetDim(axis); | ||
| 85 | + const int64_t paramDim = paramShape->GetDim(axis); | ||
| 86 | + OP_CHECK_IF(dim >= 0 && paramDim >= 0 && dim != paramDim, | ||
| 87 | + OP_LOGE_FOR_INVALID_SHAPES_WITH_REASON( | ||
| 88 | + context->GetNodeName(), (std::string("parameters, ") + inputNames[input]).c_str(), | ||
| 89 | + Ops::Base::ToString(std::vector<const gert::Shape*>{paramShape, shape}).c_str(), | ||
| 90 | + "must have the same shape as parameters"), | ||
| 91 | + return ge::GRAPH_FAILED); | ||
| 92 | + } | ||
| 93 | + } | ||
| 94 | + return ge::GRAPH_SUCCESS; | ||
| 95 | +} | ||
| 42 | } // namespace | 96 | } // namespace |
| 43 | 97 | ||
| 44 | static ge::graphStatus InferShapeForSgd(gert::InferShapeContext* context) | 98 | static ge::graphStatus InferShapeForSgd(gert::InferShapeContext* context) |
| 45 | { | 99 | { |
| 100 | + if (context == nullptr) { | ||
| 101 | + return ge::GRAPH_FAILED; | ||
| 102 | + } | ||
| 46 | OP_LOGD(context, "InferShapeForSgd begin."); | 103 | OP_LOGD(context, "InferShapeForSgd begin."); |
| 47 | 104 | ||
| 48 | const auto* attrs = context->GetAttrs(); | 105 | const auto* attrs = context->GetAttrs(); |
| @@ -65,19 +122,11 @@ static ge::graphStatus InferShapeForSgd(gert::InferShapeContext* context) | |||
| 65 | "attr weight_decay must be more than or equal to 0"), | 122 | "attr weight_decay must be more than or equal to 0"), |
| 66 | return ge::GRAPH_FAILED); | 123 | return ge::GRAPH_FAILED); |
| 67 | 124 | ||
| 68 | - // 这里守的是「框架没把 shape 递过来」(返回 nullptr),属于内部失败,不是 parameters 取值非法, | ||
| 69 | - // 所以用裸 OP_LOGE 而非 OP_LOGE_FOR_INVALID_*:后者会走 REPORT_PREDEFINED_ERR_MSG 报 EZ0009/EZ0010, | ||
| 70 | - // 对用户提示「你的 shape 不合法」,而此刻并不知道用户给了什么。真正的 shape 取值校验在下面的 dimNum 分支。 | ||
| 71 | const gert::Shape* paramShape = context->GetInputShape(PARAMETERS_INDEX); | 125 | const gert::Shape* paramShape = context->GetInputShape(PARAMETERS_INDEX); |
| 72 | OP_CHECK_IF(paramShape == nullptr, OP_LOGE(context, "Get input shape of parameters returns nullptr."), | 126 | OP_CHECK_IF(paramShape == nullptr, OP_LOGE(context, "Get input shape of parameters returns nullptr."), |
| 73 | return ge::GRAPH_FAILED); | 127 | return ge::GRAPH_FAILED); |
| 74 | - if (!IsUnknownRank(paramShape)) { | 128 | + OP_CHECK_IF(CheckInputShapes(context, paramShape) != ge::GRAPH_SUCCESS, |
| 75 | - const size_t dimNum = paramShape->GetDimNum(); | 129 | + OP_LOGD(context, "Input shape validation failed."), return ge::GRAPH_FAILED); |
| 76 | - OP_CHECK_IF(dimNum < MIN_DIM_NUM || dimNum > MAX_DIM_NUM, | ||
| 77 | - OP_LOGE_FOR_INVALID_SHAPEDIM(context->GetNodeName(), "parameters", std::to_string(dimNum).c_str(), | ||
| 78 | - "1 ~ 8 dims"), | ||
| 79 | - return ge::GRAPH_FAILED); | ||
| 80 | - } | ||
| 81 | 130 | ||
| 82 | ge::graphStatus ret = Ops::Base::InferShape4Elewise(context); | 131 | ge::graphStatus ret = Ops::Base::InferShape4Elewise(context); |
| 83 | OP_CHECK_IF(ret == ge::GRAPH_FAILED, OP_LOGE(context, "InferShapeForSgd failed."), return ge::GRAPH_FAILED); | 132 | OP_CHECK_IF(ret == ge::GRAPH_FAILED, OP_LOGE(context, "InferShapeForSgd failed."), return ge::GRAPH_FAILED); |
| @@ -85,8 +134,6 @@ static ge::graphStatus InferShapeForSgd(gert::InferShapeContext* context) | |||
| 85 | return ret; | 134 | return ret; |
| 86 | } | 135 | } |
| 87 | 136 | ||
| 88 | -// 双挂仍然必须:漏 InferDataType 会导致 GE 侧输出 dtype 推导缺失/推错,是拷模板时的高频漏项 | 137 | +// InferDataType is registered separately in op_graph/sgd_graph_infer.cpp. |
| 89 | -// (step3 GATE 明列)。但 InferDataType 只在图场景使用,按交付件划分落在 | ||
| 90 | -// op_graph/sgd_graph_infer.cpp;此处只挂图与单算子共用的 InferShape。 | ||
| 91 | IMPL_OP_INFERSHAPE(SGD).InferShape(InferShapeForSgd); | 138 | IMPL_OP_INFERSHAPE(SGD).InferShape(InferShapeForSgd); |
| 92 | } // namespace ops | 139 | } // namespace ops |
| @@ -102,8 +102,7 @@ struct SgdDag { | |||
| 102 | using OpStatF = Bind<Vec::Cast<T, U, 0>, OpCopyInStat>; | 102 | using OpStatF = Bind<Vec::Cast<T, U, 0>, OpCopyInStat>; |
| 103 | 103 | ||
| 104 | // ── Step 1:权重衰减 grad = gradient + parameters * wd ────────────────── | 104 | // ── Step 1:权重衰减 grad = gradient + parameters * wd ────────────────── |
| 105 | - // wd == 0 时【真正跳过】而不是乘 0(spec numerical_stability.skip_zero_branches; | 105 | + // wd == 0时跳过权重衰减乘法,避免0 * inf产生NaN并污染结果。 |
| 106 | - // 且 0 * inf = NaN 会污染结果)。 | ||
| 107 | using VarWeightDecay = Placeholder::Var<T, SGD_VAR_IDX_WEIGHT_DECAY>; | 106 | using VarWeightDecay = Placeholder::Var<T, SGD_VAR_IDX_WEIGHT_DECAY>; |
| 108 | using OpParamMulWd = Bind<Vec::Muls<T>, OpParamF, VarWeightDecay>; | 107 | using OpParamMulWd = Bind<Vec::Muls<T>, OpParamF, VarWeightDecay>; |
| 109 | using OpGradWithWd = Bind<Vec::Add<T>, OpGradInF, OpParamMulWd>; | 108 | using OpGradWithWd = Bind<Vec::Add<T>, OpGradInF, OpParamMulWd>; |
| @@ -11,13 +11,13 @@ | |||
| 11 | 11 | ||
| 12 | """SGD multi-path TestSpec golden (kernel + GEIR). | 12 | """SGD multi-path TestSpec golden (kernel + GEIR). |
| 13 | 13 | ||
| 14 | -The CPU true-value path is a composition of Torch tensor operators. Its | 14 | +The CPU true-value path is a composition of Torch tensor operators following |
| 15 | -arithmetic is kept identical to the legacy ``sgd_golden`` entry: | 15 | +the operator's formula: |
| 16 | 16 | ||
| 17 | grad = gradient + parameters * weight_decay (when weight_decay != 0) | 17 | grad = gradient + parameters * weight_decay (when weight_decay != 0) |
| 18 | accum_t = accum * momentum + grad (unconditional) | 18 | accum_t = accum * momentum + grad (unconditional) |
| 19 | accum_t -= grad * (1 - stat) * dampening (when dampening != 0) | 19 | accum_t -= grad * (1 - stat) * dampening (when dampening != 0) |
| 20 | - p_out = p - (grad*lr + accum_t*m*lr) (Nesterov) | 20 | + p_out = p - (grad + accum_t*m)*lr (Nesterov) |
| 21 | p - accum_t*lr (otherwise) | 21 | p - accum_t*lr (otherwise) |
| 22 | 22 | ||
| 23 | ``momentum == 0`` only masks the accum/stat writeback; it never skips the | 23 | ``momentum == 0`` only masks the accum/stat writeback; it never skips the |
| @@ -25,11 +25,10 @@ arithmetic is kept identical to the legacy ``sgd_golden`` entry: | |||
| 25 | NaN into parameters according to IEEE 754, while accum/stat retain their input | 25 | NaN into parameters according to IEEE 754, while accum/stat retain their input |
| 26 | bit patterns (including NaN payloads and negative zero). | 26 | bit patterns (including NaN payloads and negative zero). |
| 27 | 27 | ||
| 28 | -The third-party implementation deliberately differs only in the Nesterov | 28 | +The Nesterov update direction is summed before multiplication by learning rate. |
| 29 | -parenthesization: arch35 ``sgd_dag.h`` computes | 29 | +Distributing that multiplication changes IEEE 754 results for infinite learning |
| 30 | -``p - (grad + accum_t * momentum) * learning_rate`` to satisfy its TensorScalar | 30 | +rates. Both references retain this grouping and cast writeback values to the |
| 31 | -slot constraint. Cross-check must reproduce that actual device operation | 31 | +requested output dtype. |
| 32 | -order, then cast all three writeback values to the NPU input dtype. | ||
| 33 | 32 | ||
| 34 | The CPU oracle evaluates intermediate arithmetic in float64 to avoid sharing | 33 | The CPU oracle evaluates intermediate arithmetic in float64 to avoid sharing |
| 35 | float32 overflow with the device. The independent third-party composition keeps | 34 | float32 overflow with the device. The independent third-party composition keeps |
| @@ -46,12 +45,12 @@ except ImportError: # pragma: no cover - TTK normally provides ml_dtypes. | |||
| 46 | _NP_BFLOAT16 = None | 45 | _NP_BFLOAT16 = None |
| 47 | 46 | ||
| 48 | 47 | ||
| 49 | -__spec__ = {"sgd": "SgdKernelSpec"} | 48 | +__spec__ = {"sgd": "SgdKernelSpec", "SGD": "SgdKernelSpec"} |
| 50 | 49 | ||
| 51 | # Compatibility entry for existing kernel CSVs and callers. Both entries call | 50 | # Compatibility entry for existing kernel CSVs and callers. Both entries call |
| 52 | # the same _compute function; this mapping must remain until all old loaders have | 51 | # the same _compute function; this mapping must remain until all old loaders have |
| 53 | # migrated to TestSpec. | 52 | # migrated to TestSpec. |
| 54 | -__golden__ = {"kernel": {"sgd": "sgd_golden"}} | 53 | +__golden__ = {"kernel": {"sgd": "sgd_golden", "SGD": "sgd_golden"}} |
| 55 | 54 | ||
| 56 | 55 | ||
| 57 | _TOL = { | 56 | _TOL = { |
| @@ -128,8 +127,8 @@ def _compute( | |||
| 128 | """CPU reference shared by TestSpec and the legacy entry. | 127 | """CPU reference shared by TestSpec and the legacy entry. |
| 129 | 128 | ||
| 130 | Inputs and outputs are Torch tensors. This is the high-precision CPU | 129 | Inputs and outputs are Torch tensors. This is the high-precision CPU |
| 131 | - true-value path, so it keeps the legacy expanded Nesterov formula; the | 130 | + true-value path. It preserves the formula's grouping for nonfinite values; |
| 132 | - independent ``_SgdCompose`` below follows the grouped arch35 DAG order. | 131 | + the independent ``_SgdCompose`` below uses float32 intermediates. |
| 133 | 132 | ||
| 134 | Attribute defaults are ``None`` so that ``_attr`` can resolve each one | 133 | Attribute defaults are ``None`` so that ``_attr`` can resolve each one |
| 135 | from the ``attributes`` container when the caller did not pass a named | 134 | from the ``attributes`` container when the caller did not pass a named |
| @@ -168,13 +167,9 @@ def _compute( | |||
| 168 | if dampening_value != 0.0: | 167 | if dampening_value != 0.0: |
| 169 | accum_t = accum_t - grad * ((1.0 - s) * dampening_value) | 168 | accum_t = accum_t - grad * ((1.0 - s) * dampening_value) |
| 170 | 169 | ||
| 171 | - # Preserve the already-validated legacy true-value ordering. This is | 170 | + # Sum the update direction before multiplying by a possibly infinite rate. |
| 172 | - # mathematically equivalent to the grouped DAG form, but intentionally | ||
| 173 | - # independent from it for cross-checking. | ||
| 174 | if use_nesterov: | 171 | if use_nesterov: |
| 175 | - p_new = p - ( | 172 | + p_new = p - (grad + accum_t * momentum_value) * learning_rate_value |
| 176 | - grad * learning_rate_value + accum_t * momentum_value * learning_rate_value | ||
| 177 | - ) | ||
| 178 | else: | 173 | else: |
| 179 | p_new = p - accum_t * learning_rate_value | 174 | p_new = p - accum_t * learning_rate_value |
| 180 | 175 | ||
| @@ -384,8 +379,7 @@ class _SgdCompose: | |||
| 384 | dampening_term = torch.mul(grad, stat_active_damped) | 379 | dampening_term = torch.mul(grad, stat_active_damped) |
| 385 | accum_t = torch.sub(accum_t, dampening_term) | 380 | accum_t = torch.sub(accum_t, dampening_term) |
| 386 | 381 | ||
| 387 | - # Step 4: reproduce the real grouped Nesterov DAG, not the expanded | 382 | + # Step 4: sum the Nesterov update direction before scaling it. |
| 388 | - # legacy true-value expression. | ||
| 389 | if self.nesterov: | 383 | if self.nesterov: |
| 390 | accum_t_momentum = torch.mul(accum_t, momentum_value) | 384 | accum_t_momentum = torch.mul(accum_t, momentum_value) |
| 391 | nesterov_sum = torch.add(grad, accum_t_momentum) | 385 | nesterov_sum = torch.add(grad, accum_t_momentum) |
| @@ -406,7 +400,7 @@ class _SgdCompose: | |||
| 406 | 400 | ||
| 407 | 401 | ||
| 408 | class SgdKernelSpec: | 402 | class SgdKernelSpec: |
| 409 | - """Kernel and GEIR share the snake_case ``sgd`` registration.""" | 403 | + """Kernel and GEIR accept both snake_case and graph operator names.""" |
| 410 | 404 | ||
| 411 | def golden( | 405 | def golden( |
| 412 | parameters, | 406 | parameters, |
| @@ -0,0 +1,38 @@ | |||
| 1 | +# Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 2 | +# This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 3 | +# CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 4 | +# Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 5 | +# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 6 | +# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 7 | +# See LICENSE in the root of the software repository for the full text of the License. | ||
| 8 | + | ||
| 9 | +import math | ||
| 10 | +import unittest | ||
| 11 | + | ||
| 12 | +import torch | ||
| 13 | + | ||
| 14 | +from golden import _compute | ||
| 15 | + | ||
| 16 | + | ||
| 17 | +class SgdGoldenTest(unittest.TestCase): | ||
| 18 | + def test_nesterov_infinite_learning_rate_preserves_update_sign(self): | ||
| 19 | + # The finite update direction is positive. Multiplying its two terms | ||
| 20 | + # by infinity separately would instead introduce inf + (-inf) = NaN. | ||
| 21 | + for dtype in (torch.float16, torch.bfloat16, torch.float32): | ||
| 22 | + for rate, expected in ((math.inf, -math.inf), (-math.inf, math.inf)): | ||
| 23 | + with self.subTest(dtype=dtype, learning_rate=rate): | ||
| 24 | + values = (-1.0, 10.0, rate, 1.0, -0.01, -10.0) | ||
| 25 | + inputs = [torch.tensor([value], dtype=dtype) for value in values] | ||
| 26 | + outputs = _compute(*inputs, nesterov=True) | ||
| 27 | + self.assertEqual(outputs[0].item(), expected) | ||
| 28 | + self.assertTrue(torch.isfinite(outputs[1]).all()) | ||
| 29 | + self.assertEqual(outputs[2].item(), 0.0) | ||
| 30 | + | ||
| 31 | + def test_nesterov_finite_update(self): | ||
| 32 | + inputs = [torch.tensor([value]) for value in (1.0, 2.0, 0.25, 4.0, 0.5, 1.0)] | ||
| 33 | + outputs = _compute(*inputs, nesterov=True) | ||
| 34 | + self.assertEqual([value.item() for value in outputs], [0.0, 4.0, 0.0]) | ||
| 35 | + | ||
| 36 | + | ||
| 37 | +if __name__ == "__main__": | ||
| 38 | + unittest.main() | ||
| @@ -0,0 +1,7 @@ | |||
| 1 | +testcase_name,network_name,op_name,input_shapes,input_dtypes,input_formats,output_shapes,output_dtypes,output_formats,input_ori_shapes,input_ori_formats,output_ori_shapes,output_ori_formats,attributes,input_data_ranges,precision_tolerances,absolute_precision,output_inplace_indexes,output_shape_unknown_indexes,is_enabled,remark,soc_series,priority,dump_file_prefix,manual_input_binaries,manual_golden_binaries | ||
| 2 | +SGD_st_fp32_defaults_zero_momentum,UNKNOWN,SGD,"((33,), (33,), (1,), (33,), (1,), (33,))","('float32', 'float32', 'float32', 'float32', 'float32', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((33,),)","('float32',)","('ND',)","((33,), (33,), (1,), (33,), (1,), (33,))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((33,),)","('ND',)",{},"((-2, 2), (-1, 1), (0.125, 0.125), (0.5, 0.5), (0, 0), (1, 1))",,1e-08,"(0,)",(),True,default attributes and zero momentum,,0,,(),() | ||
| 3 | +SGD_st_fp16_nesterov,UNKNOWN,SGD,"((17,), (17,), (1,), (17,), (1,), (17,))","('float16', 'float16', 'float16', 'float16', 'float16', 'float16')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((17,),)","('float16',)","('ND',)","((17,), (17,), (1,), (17,), (1,), (17,))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((17,),)","('ND',)",{'nesterov': True},"((-2, 2), (-1, 1), (0.125, 0.125), (0.5, 0.5), (0.5, 0.5), (1, 1))",,1e-08,"(0,)",(),True,Nesterov update and unaligned tail,,0,,(),() | ||
| 4 | +SGD_st_bf16_dampening_decay,UNKNOWN,SGD,"((2, 3), (2, 3), (1, 1), (2, 3), (1,), (2, 3))","('bfloat16', 'bfloat16', 'bfloat16', 'bfloat16', 'bfloat16', 'bfloat16')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((2, 3),)","('bfloat16',)","('ND',)","((2, 3), (2, 3), (1, 1), (2, 3), (1,), (2, 3))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((2, 3),)","('ND',)","{'dampening': 0.125, 'weight_decay': 0.01}","((-2, 2), (-1, 1), (0.125, 0.125), (0.5, 0.5), (0.5, 0.5), (0, 0))",,1e-08,"(0,)",(),True,dampening with weight decay and one-element rank2 rate,,0,,(),() | ||
| 5 | +SGD_st_fp32_rank8_first_step,UNKNOWN,SGD,"((1, 2, 1, 2, 1, 2, 1, 2), (1, 2, 1, 2, 1, 2, 1, 2), (1,), (1, 2, 1, 2, 1, 2, 1, 2), (1,), (1, 2, 1, 2, 1, 2, 1, 2))","('float32', 'float32', 'float32', 'float32', 'float32', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((1, 2, 1, 2, 1, 2, 1, 2),)","('float32',)","('ND',)","((1, 2, 1, 2, 1, 2, 1, 2), (1, 2, 1, 2, 1, 2, 1, 2), (1,), (1, 2, 1, 2, 1, 2, 1, 2), (1,), (1, 2, 1, 2, 1, 2, 1, 2))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((1, 2, 1, 2, 1, 2, 1, 2),)","('ND',)","{'dampening': 0.25, 'weight_decay': 0.125}","((-2, 2), (-1, 1), (0.125, 0.125), (0.5, 0.5), (0.5, 0.5), (1, 1))",,1e-08,"(0,)",(),True,maximum rank and first-step flag,,0,,(),() | ||
| 6 | +SGD_st_fp32_tiling_tail,UNKNOWN,SGD,"((3713,), (3713,), (1,), (3713,), (1,), (3713,))","('float32', 'float32', 'float32', 'float32', 'float32', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((3713,),)","('float32',)","('ND',)","((3713,), (3713,), (1,), (3713,), (1,), (3713,))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((3713,),)","('ND',)","{'dampening': 0.125, 'weight_decay': 0.01}","((-2, 2), (-1, 1), (0.125, 0.125), (0.5, 0.5), (0.5, 0.5), (0, 0))",,1e-08,"(0,)",(),True,small tile boundary regression,,0,,(),() | ||
| 7 | +SGD_st_fp16_nesterov_inf,UNKNOWN,SGD,"((3,), (3,), (1,), (3,), (1,), (3,))","('float16', 'float16', 'float16', 'float16', 'float16', 'float16')","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((3,),)","('float16',)","('ND',)","((3,), (3,), (1,), (3,), (1,), (3,))","('ND', 'ND', 'ND', 'ND', 'ND', 'ND')","((3,),)","('ND',)",{'nesterov': True},"((1, 1), (1, 1), (float('inf'), float('inf')), (-4, -4), (0.5, 0.5), (1, 1))",,1e-08,"(0,)",(),True,Nesterov direction sum before Inf learning-rate multiplication,,0,,(),() | ||
| @@ -0,0 +1,11 @@ | |||
| 1 | +# Copyright (c) 2026 Huawei Technologies Co., Ltd. | ||
| 2 | +# This program is free software, you can redistribute it and/or modify it under the terms and conditions of | ||
| 3 | +# CANN Open Software License Agreement Version 2.0 (the "License"). | ||
| 4 | +# Please refer to the License for details. You may not use this file except in compliance with the License. | ||
| 5 | +# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, | ||
| 6 | +# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. | ||
| 7 | +# See LICENSE in the root of the software repository for the full text of the License. | ||
| 8 | + | ||
| 9 | +if(UT_TEST_ALL OR OP_GRAPH_UT) | ||
| 10 | + add_modules_ut_sources(HOSTNAME ${OP_GRAPH_MODULE_NAME} MODE PRIVATE DIR ${CMAKE_CURRENT_SOURCE_DIR}) | ||
| 11 | +endif() | ||
| @@ -0,0 +1,63 @@ | |||
| 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 <gtest/gtest.h> | ||
| 12 | +#include <array> | ||
| 13 | +#include "base/context_builder/op_infer_datatype_context_builder.h" | ||
| 14 | + | ||
| 15 | +namespace ops { | ||
| 16 | +ge::graphStatus InferDataTypeForSgd(gert::InferDataTypeContext* context); | ||
| 17 | +} // namespace ops | ||
| 18 | + | ||
| 19 | +namespace { | ||
| 20 | +void CheckSgdDtype(const std::array<ge::DataType, 6>& types, ge::graphStatus expected) | ||
| 21 | +{ | ||
| 22 | + gert::OpInferDataTypeContextBuilder builder; | ||
| 23 | + builder.OpType("SGD").OpName("sgd_dtype").IONum(6, 1); | ||
| 24 | + for (size_t input = 0; input < types.size(); ++input) { | ||
| 25 | + builder.InputTensorDesc(input, types[input], ge::FORMAT_ND, ge::FORMAT_ND); | ||
| 26 | + } | ||
| 27 | + builder.OutputTensorDesc(0, ge::FORMAT_ND, ge::FORMAT_ND); | ||
| 28 | + auto holder = builder.Build(); | ||
| 29 | + ASSERT_NE(holder.GetContext(), nullptr); | ||
| 30 | + ASSERT_EQ(ops::InferDataTypeForSgd(holder.GetContext()), expected); | ||
| 31 | + if (expected == ge::GRAPH_SUCCESS) { | ||
| 32 | + EXPECT_EQ(holder.GetContext()->GetOutputDataType(0), types[0]); | ||
| 33 | + } | ||
| 34 | +} | ||
| 35 | +} // namespace | ||
| 36 | + | ||
| 37 | +TEST(SgdGraphInfer, PreservesSupportedTypes) | ||
| 38 | +{ | ||
| 39 | + for (auto dtype : {ge::DT_FLOAT16, ge::DT_FLOAT, ge::DT_BF16}) { | ||
| 40 | + std::array<ge::DataType, 6> types; | ||
| 41 | + types.fill(dtype); | ||
| 42 | + CheckSgdDtype(types, ge::GRAPH_SUCCESS); | ||
| 43 | + } | ||
| 44 | +} | ||
| 45 | + | ||
| 46 | +TEST(SgdGraphInfer, RejectsEachMismatchedInput) | ||
| 47 | +{ | ||
| 48 | + for (size_t input = 1; input < 6; ++input) { | ||
| 49 | + std::array<ge::DataType, 6> types; | ||
| 50 | + types.fill(ge::DT_FLOAT); | ||
| 51 | + types[input] = ge::DT_FLOAT16; | ||
| 52 | + CheckSgdDtype(types, ge::GRAPH_FAILED); | ||
| 53 | + } | ||
| 54 | +} | ||
| 55 | + | ||
| 56 | +TEST(SgdGraphInfer, RejectsUnsupportedMatchingTypes) | ||
| 57 | +{ | ||
| 58 | + for (auto dtype : {ge::DT_INT32, ge::DT_BOOL, ge::DT_DOUBLE}) { | ||
| 59 | + std::array<ge::DataType, 6> types; | ||
| 60 | + types.fill(dtype); | ||
| 61 | + CheckSgdDtype(types, ge::GRAPH_FAILED); | ||
| 62 | + } | ||
| 63 | +} | ||
| @@ -301,8 +301,7 @@ TEST_F(TestSgdTiling, sgd_tiling_reject_empty_tensor) | |||
| 301 | } | 301 | } |
| 302 | 302 | ||
| 303 | // ── DFX:空 tensor「全空间」枚举 ──────────────────────────────────────────── | 303 | // ── DFX:空 tensor「全空间」枚举 ──────────────────────────────────────────── |
| 304 | -// changwei-op-dev step5-verify §5.4 要求:空 tensor = **任意一轴或多轴为 0**, | 304 | +// 空Tensor包括任意一轴或多轴为0的情形,逐形态检查拒绝行为。 |
| 305 | -// 必须逐形态真跑分类,不能只验一个维度,结论要写成「仅支持哪个轴为空」。 | ||
| 306 | // 本算子的契约是【所有形态一律拒绝】(无空进空出语义),故下面每一条都断言 GRAPH_FAILED。 | 305 | // 本算子的契约是【所有形态一律拒绝】(无空进空出语义),故下面每一条都断言 GRAPH_FAILED。 |
| 307 | // 若将来放开某一轴,本组用例会立刻在该形态上变红,逼迫同步更新契约与文档。 | 306 | // 若将来放开某一轴,本组用例会立刻在该形态上变红,逼迫同步更新契约与文档。 |
| 308 | // | 307 | // |
| @@ -421,7 +420,7 @@ TEST_F(TestSgdTiling, sgd_tiling_accept_scalar_shape_any_single_element) | |||
| 421 | 420 | ||
| 422 | TEST_F(TestSgdTiling, sgd_tiling_reject_dtype_mismatch) | 421 | TEST_F(TestSgdTiling, sgd_tiling_reject_dtype_mismatch) |
| 423 | { | 422 | { |
| 424 | - // 9 个张量位必须同 dtype(parameters 为 fp32,其余为 fp16) | 423 | + // 全部输入必须同dtype(parameters为fp32,其余为fp16)。 |
| 425 | gert::StorageShape param = {{2, 4}, {2, 4}}; | 424 | gert::StorageShape param = {{2, 4}, {2, 4}}; |
| 426 | gert::StorageShape one = {{1}, {1}}; | 425 | gert::StorageShape one = {{1}, {1}}; |
| 427 | ASSERT_EQ( | 426 | ASSERT_EQ( |
| @@ -437,3 +436,26 @@ TEST_F(TestSgdTiling, sgd_tiling_reject_unsupported_dtype) | |||
| 437 | RunSgdTiling(param, param, one, param, one, param, ge::DT_INT32, ge::DT_INT32, 0.0f, 0.0f, false, nullptr), | 436 | RunSgdTiling(param, param, one, param, one, param, ge::DT_INT32, ge::DT_INT32, 0.0f, 0.0f, false, nullptr), |
| 438 | ge::GRAPH_FAILED); | 437 | ge::GRAPH_FAILED); |
| 439 | } | 438 | } |
| 439 | + | ||
| 440 | +TEST_F(TestSgdTiling, sgd_tiling_reject_shape_size_overflow) | ||
| 441 | +{ | ||
| 442 | + // Metadata-only shapes: no tensor storage is allocated for these sizes. | ||
| 443 | + uint64_t key = 0; | ||
| 444 | + EXPECT_EQ(RunSgdTilingNormal({4294967297LL, 4294967297LL}, ge::DT_FLOAT, 0.0f, 0.0f, false, &key), | ||
| 445 | + ge::GRAPH_FAILED); | ||
| 446 | +} | ||
| 447 | + | ||
| 448 | +TEST_F(TestSgdTiling, sgd_tiling_reject_overflowing_scalar_shape) | ||
| 449 | +{ | ||
| 450 | + gert::StorageShape param = {{2, 4}, {2, 4}}; | ||
| 451 | + const int64_t maxDim = std::numeric_limits<int64_t>::max(); | ||
| 452 | + // Unchecked signed multiplication can wrap this element count to one. | ||
| 453 | + gert::StorageShape invalidScalar = {{maxDim, maxDim}, {maxDim, maxDim}}; | ||
| 454 | + gert::StorageShape one = {{1}, {1}}; | ||
| 455 | + EXPECT_EQ(RunSgdTiling(param, param, invalidScalar, param, one, param, ge::DT_FLOAT, ge::DT_FLOAT, 0.0f, 0.0f, | ||
| 456 | + false, nullptr), | ||
| 457 | + ge::GRAPH_FAILED); | ||
| 458 | + EXPECT_EQ(RunSgdTiling(param, param, one, param, invalidScalar, param, ge::DT_FLOAT, ge::DT_FLOAT, 0.0f, 0.0f, | ||
| 459 | + false, nullptr), | ||
| 460 | + ge::GRAPH_FAILED); | ||
| 461 | +} | ||
| @@ -24,6 +24,7 @@ | |||
| 24 | */ | 24 | */ |
| 25 | 25 | ||
| 26 | #include <gtest/gtest.h> // NOLINT | 26 | #include <gtest/gtest.h> // NOLINT |
| 27 | +#include <array> | ||
| 27 | #include <iostream> | 28 | #include <iostream> |
| 28 | #include <limits> | 29 | #include <limits> |
| 29 | #include <vector> | 30 | #include <vector> |
| @@ -165,3 +166,72 @@ TEST_F(SGD, sgd_infershape_nonnegative_weight_decay_boundaries) | |||
| 165 | EXPECT_EQ(RunSgdInferShape(bigShape, scalarShape, 0.0f, weightDecay, false), ge::GRAPH_SUCCESS); | 166 | EXPECT_EQ(RunSgdInferShape(bigShape, scalarShape, 0.0f, weightDecay, false), ge::GRAPH_SUCCESS); |
| 166 | } | 167 | } |
| 167 | } | 168 | } |
| 169 | + | ||
| 170 | +namespace { | ||
| 171 | +using SgdShapes = std::array<gert::StorageShape, SGD_INPUT_NUM>; | ||
| 172 | + | ||
| 173 | +SgdShapes NormalSgdShapes() | ||
| 174 | +{ | ||
| 175 | + gert::StorageShape tensor = {{2, 3}, {2, 3}}; | ||
| 176 | + gert::StorageShape scalar = {{1}, {1}}; | ||
| 177 | + return {tensor, tensor, scalar, tensor, scalar, tensor}; | ||
| 178 | +} | ||
| 179 | + | ||
| 180 | +ge::graphStatus RunSgdShapes(SgdShapes& shapes) | ||
| 181 | +{ | ||
| 182 | + auto inferShapeFunc = gert::OpImplRegistry::GetInstance().GetOpImpl("SGD")->infer_shape; | ||
| 183 | + gert::StorageShape output = {{}, {}}; | ||
| 184 | + auto holder = gert::InferShapeContextFaker() | ||
| 185 | + .NodeIoNum(SGD_INPUT_NUM, SGD_OUTPUT_NUM) | ||
| 186 | + .IrInstanceNum({1, 1, 1, 1, 1, 1}) | ||
| 187 | + .InputShapes({&shapes[0], &shapes[1], &shapes[2], &shapes[3], &shapes[4], &shapes[5]}) | ||
| 188 | + .OutputShapes({&output}) | ||
| 189 | + .NodeAttrs({{"dampening", Ops::NN::AnyValue::CreateFrom<float>(0.0f)}, | ||
| 190 | + {"weight_decay", Ops::NN::AnyValue::CreateFrom<float>(0.0f)}, | ||
| 191 | + {"nesterov", Ops::NN::AnyValue::CreateFrom<bool>(false)}}) | ||
| 192 | + .Build(); | ||
| 193 | + return inferShapeFunc(holder.GetContext<gert::InferShapeContext>()); | ||
| 194 | +} | ||
| 195 | +} // namespace | ||
| 196 | + | ||
| 197 | +TEST_F(SGD, sgd_infershape_rejects_each_empty_axis) | ||
| 198 | +{ | ||
| 199 | + for (gert::StorageShape empty : {gert::StorageShape{{0, 3}, {0, 3}}, gert::StorageShape{{3, 0}, {3, 0}}}) { | ||
| 200 | + auto shapes = NormalSgdShapes(); | ||
| 201 | + for (size_t input : {0U, 1U, 3U, 5U}) { | ||
| 202 | + shapes[input] = empty; | ||
| 203 | + } | ||
| 204 | + EXPECT_EQ(RunSgdShapes(shapes), ge::GRAPH_FAILED); | ||
| 205 | + } | ||
| 206 | +} | ||
| 207 | + | ||
| 208 | +TEST_F(SGD, sgd_infershape_rejects_non_scalar_learning_rate_and_momentum) | ||
| 209 | +{ | ||
| 210 | + for (size_t input : {2U, 4U}) { | ||
| 211 | + auto shapes = NormalSgdShapes(); | ||
| 212 | + shapes[input] = {{1, 2}, {1, 2}}; | ||
| 213 | + EXPECT_EQ(RunSgdShapes(shapes), ge::GRAPH_FAILED); | ||
| 214 | + } | ||
| 215 | +} | ||
| 216 | + | ||
| 217 | +TEST_F(SGD, sgd_infershape_rejects_shape_and_rank_mismatch) | ||
| 218 | +{ | ||
| 219 | + for (size_t input : {1U, 3U, 5U}) { | ||
| 220 | + for (gert::StorageShape mismatch : {gert::StorageShape{{3, 2}, {3, 2}}, gert::StorageShape{{6}, {6}}}) { | ||
| 221 | + auto shapes = NormalSgdShapes(); | ||
| 222 | + shapes[input] = mismatch; | ||
| 223 | + EXPECT_EQ(RunSgdShapes(shapes), ge::GRAPH_FAILED); | ||
| 224 | + } | ||
| 225 | + } | ||
| 226 | +} | ||
| 227 | + | ||
| 228 | +TEST_F(SGD, sgd_infershape_preserves_partial_and_unknown_rank_inputs) | ||
| 229 | +{ | ||
| 230 | + auto shapes = NormalSgdShapes(); | ||
| 231 | + shapes[0] = {{2, -1}, {2, -1}}; | ||
| 232 | + shapes[2] = {{-1, 1}, {-1, 1}}; | ||
| 233 | + shapes[4] = {{-2}, {-2}}; | ||
| 234 | + EXPECT_EQ(RunSgdShapes(shapes), ge::GRAPH_SUCCESS); | ||
| 235 | + shapes[0] = {{-2}, {-2}}; | ||
| 236 | + EXPECT_EQ(RunSgdShapes(shapes), ge::GRAPH_SUCCESS); | ||
| 237 | +} | ||