From cffc6b2e2e9435c7b3db589303c5e70217c51d0e Mon Sep 17 00:00:00 2001 From: shenwei41 Date: Wed, 3 Feb 2021 23:19:27 +0800 Subject: [PATCH] update include file --- inc/external/acl/acl_base.h | 1 + inc/external/acl/error_codes/ge_error_codes.h | 14 ++ inc/external/acl/error_codes/rt_error_codes.h | 3 +- inc/external/runtime/rt_error_codes.h | 3 +- third_party/fwkacllib/inc/ops/batch_ops.h | 15 +- third_party/fwkacllib/inc/ops/data_flow_ops.h | 47 +++- .../inc/ops/elewise_calculation_ops.h | 80 +++++++ third_party/fwkacllib/inc/ops/hcom_ops.h | 16 ++ third_party/fwkacllib/inc/ops/image_ops.h | 60 ++++- third_party/fwkacllib/inc/ops/linalg_ops.h | 84 +++---- third_party/fwkacllib/inc/ops/list_ops.h | 2 +- third_party/fwkacllib/inc/ops/math_ops.h | 68 ++++++ .../inc/ops/matrix_calculation_ops.h | 21 ++ .../fwkacllib/inc/ops/nn_batch_norm_ops.h | 104 +++++++++ .../fwkacllib/inc/ops/nn_calculation_ops.h | 16 +- third_party/fwkacllib/inc/ops/nn_detect_ops.h | 47 +++- third_party/fwkacllib/inc/ops/nn_norm_ops.h | 41 ++++ third_party/fwkacllib/inc/ops/nn_ops.h | 29 ++- .../fwkacllib/inc/ops/nonlinear_fuc_ops.h | 44 +++- third_party/fwkacllib/inc/ops/reduce_ops.h | 155 ++++++++++++- third_party/fwkacllib/inc/ops/selection_ops.h | 70 +++++- third_party/fwkacllib/inc/ops/sparse_ops.h | 6 +- third_party/fwkacllib/inc/ops/spectral_ops.h | 96 ++++++++ .../fwkacllib/inc/ops/split_combination_ops.h | 18 +- third_party/fwkacllib/inc/ops/string_ops.h | 2 +- third_party/fwkacllib/inc/runtime/config.h | 19 ++ third_party/fwkacllib/inc/runtime/dev.h | 5 + third_party/fwkacllib/inc/runtime/kernel.h | 35 +++ third_party/fwkacllib/inc/runtime/rt_model.h | 14 ++ third_party/fwkacllib/inc/toolchain/slog.h | 30 +-- .../inc/toolchain/tuning_tool/tune_api.h | 209 ++++++++++++------ 31 files changed, 1188 insertions(+), 166 deletions(-) diff --git a/inc/external/acl/acl_base.h b/inc/external/acl/acl_base.h index b3111860..839bd597 100644 --- a/inc/external/acl/acl_base.h +++ b/inc/external/acl/acl_base.h @@ -111,6 +111,7 @@ static const int ACL_ERROR_DUMP_NOT_RUN = 100045; static const int ACL_ERROR_PROF_REPEAT_SUBSCRIBE = 148046; static const int ACL_ERROR_PROF_API_CONFLICT = 148047; static const int ACL_ERROR_INVALID_MAX_OPQUEUE_NUM_CONFIG = 148048; +static const int ACL_ERROR_INVALID_OPP_PATH = 148049; static const int ACL_ERROR_BAD_ALLOC = 200000; static const int ACL_ERROR_API_NOT_SUPPORT = 200001; diff --git a/inc/external/acl/error_codes/ge_error_codes.h b/inc/external/acl/error_codes/ge_error_codes.h index 041fc7ae..b477a18c 100644 --- a/inc/external/acl/error_codes/ge_error_codes.h +++ b/inc/external/acl/error_codes/ge_error_codes.h @@ -17,6 +17,20 @@ #ifndef INC_EXTERNAL_GE_GE_ERROR_CODES_H_ #define INC_EXTERNAL_GE_GE_ERROR_CODES_H_ +#if defined(_MSC_VER) +#ifdef FUNC_VISIBILITY +#define GE_FUNC_VISIBILITY _declspec(dllexport) +#else +#define GE_FUNC_VISIBILITY +#endif +#else +#ifdef FUNC_VISIBILITY +#define GE_FUNC_VISIBILITY __attribute__((visibility("default"))) +#else +#define GE_FUNC_VISIBILITY +#endif +#endif + #include #ifdef __cplusplus diff --git a/inc/external/acl/error_codes/rt_error_codes.h b/inc/external/acl/error_codes/rt_error_codes.h index d2373525..2109fb79 100644 --- a/inc/external/acl/error_codes/rt_error_codes.h +++ b/inc/external/acl/error_codes/rt_error_codes.h @@ -92,7 +92,8 @@ static const int32_t ACL_ERROR_RT_LABEL_CONTEXT = 507031; // label not static const int32_t ACL_ERROR_RT_PROGRAM_USE_OUT = 507032; // program register num use out static const int32_t ACL_ERROR_RT_DEV_SETUP_ERROR = 507033; // device setup error -static const int32_t ACL_ERROR_RT_DRV_INTERNAL_ERROR = 507899; // drv internal error +static const int32_t ACL_ERROR_RT_DRV_INTERNAL_ERROR = 507899; // drv internal error +static const int32_t ACL_ERROR_RT_AICPU_INTERNAL_ERROR = 507900; // aicpu internal error #ifdef __cplusplus } diff --git a/inc/external/runtime/rt_error_codes.h b/inc/external/runtime/rt_error_codes.h index d2373525..2109fb79 100644 --- a/inc/external/runtime/rt_error_codes.h +++ b/inc/external/runtime/rt_error_codes.h @@ -92,7 +92,8 @@ static const int32_t ACL_ERROR_RT_LABEL_CONTEXT = 507031; // label not static const int32_t ACL_ERROR_RT_PROGRAM_USE_OUT = 507032; // program register num use out static const int32_t ACL_ERROR_RT_DEV_SETUP_ERROR = 507033; // device setup error -static const int32_t ACL_ERROR_RT_DRV_INTERNAL_ERROR = 507899; // drv internal error +static const int32_t ACL_ERROR_RT_DRV_INTERNAL_ERROR = 507899; // drv internal error +static const int32_t ACL_ERROR_RT_AICPU_INTERNAL_ERROR = 507900; // aicpu internal error #ifdef __cplusplus } diff --git a/third_party/fwkacllib/inc/ops/batch_ops.h b/third_party/fwkacllib/inc/ops/batch_ops.h index a4786cd3..181bf694 100644 --- a/third_party/fwkacllib/inc/ops/batch_ops.h +++ b/third_party/fwkacllib/inc/ops/batch_ops.h @@ -107,11 +107,13 @@ across multiple sessions . \n REG_OP(Unbatch) .INPUT(x_tensor, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \ - DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE})) + DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .INPUT(index, TensorType({DT_INT64})) .INPUT(id, TensorType({DT_INT64})) .OUTPUT(y_tensor, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \ - DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE})) + DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .REQUIRED_ATTR(timeout_micros, Int) .ATTR(container, String, "") .ATTR(shared_name, String, "") @@ -146,13 +148,16 @@ across multiple sessions . \n REG_OP(UnbatchGrad) .INPUT(x_input, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \ - DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE})) + DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .INPUT(index, TensorType({DT_INT64})) .INPUT(grad, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \ - DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE})) + DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .INPUT(id, TensorType({DT_INT64})) .OUTPUT(y_grad, TensorType({DT_INT8, DT_UINT8, DT_INT16, DT_UINT16, \ - DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE})) + DT_INT32, DT_INT64, DT_BOOL, DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .ATTR(container, String, "") .ATTR(shared_name, String, "") .OP_END_FACTORY_REG(UnbatchGrad) diff --git a/third_party/fwkacllib/inc/ops/data_flow_ops.h b/third_party/fwkacllib/inc/ops/data_flow_ops.h index 45303828..0043c027 100644 --- a/third_party/fwkacllib/inc/ops/data_flow_ops.h +++ b/third_party/fwkacllib/inc/ops/data_flow_ops.h @@ -1430,6 +1430,24 @@ REG_OP(OrderedMapClear) .ATTR(shared_name, String, "") .OP_END_FACTORY_REG(OrderedMapClear) +/** +*@brief FakeQueue, support tf api FixedLengthRecordReader. \n + +*@par Inputs: +*Including: +* @li resource: A Tensor of type DT_RESOURCE. + +*@par Outputs: +*handle: A Tensor of type DT_STRING ref. \n + +*@par Third-party framework compatibility +*Compatible with the TensorFlow operator FakeQueue. +*/ +REG_OP(FakeQueue) + .INPUT(resource, TensorType({DT_RESOURCE})) + .OUTPUT(handle, TensorType({DT_STRING})) + .OP_END_FACTORY_REG(FakeQueue) + /** *@brief Returns the number of incomplete elements in the underlying container. \n @@ -2258,6 +2276,7 @@ REG_OP(LruCache) .ATTR(shared_name, String, "LruCache") .ATTR(cache_size, Int, 100000) .ATTR(load_factor, Float, 1) + .REQUIRED_ATTR(dtype, Type) .OP_END_FACTORY_REG(LruCache) /** @@ -2277,9 +2296,9 @@ REG_OP(CacheAdd) .INPUT(cache, TensorType({DT_RESOURCE})) .INPUT(ids, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) .OUTPUT(swap_in_id, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) - .OUTPUT(swap_in_idx, TensorType({DT_INT64})) + .OUTPUT(swap_in_idx, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) .OUTPUT(swap_out_id, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) - .OUTPUT(swap_out_idx, TensorType({DT_INT64})) + .OUTPUT(swap_out_idx, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) .OP_END_FACTORY_REG(CacheAdd) /** @@ -2295,9 +2314,31 @@ REG_OP(CacheAdd) REG_OP(CacheRemoteIndexToLocal) .INPUT(cache, TensorType({DT_RESOURCE})) .INPUT(ids, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) - .OUTPUT(local_idx, TensorType({DT_INT64})) + .OUTPUT(local_idx, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) .OP_END_FACTORY_REG(CacheRemoteIndexToLocal) +/** +*@brief CacheAllToLocalIndex, get id in cache +*@par Inputs: +*cache: resource data +*local_idx: id in cache. +*@par Restrictions: +*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use. +*/ +REG_OP(CacheAllIndexToLocal) + .INPUT(cache, TensorType({DT_RESOURCE})) + .OUTPUT(local_idx, TensorType({DT_INT64, DT_INT32, DT_UINT64, DT_UINT32})) + .REQUIRED_ATTR(dtype, Type) + .OP_END_FACTORY_REG(CacheAllIndexToLocal) + +REG_OP(DynamicGetNext) + .INPUT(x, TensorType::ALL()) + .DYNAMIC_OUTPUT(y, TensorType::ALL()) + .ATTR(output_types, ListType, {}) + .ATTR(output_shapes, ListListInt, {{}, {}}) + .ATTR(_dynamic_graph_execute_mode, String, "lazy_recompile") + .ATTR(_getnext_inputs_shape_range, String, "") + .OP_END_FACTORY_REG(DynamicGetNext) } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_DATA_FLOW_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/elewise_calculation_ops.h b/third_party/fwkacllib/inc/ops/elewise_calculation_ops.h index e65c7027..9f981d12 100644 --- a/third_party/fwkacllib/inc/ops/elewise_calculation_ops.h +++ b/third_party/fwkacllib/inc/ops/elewise_calculation_ops.h @@ -3627,6 +3627,35 @@ REG_OP(Lerp) .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) .OP_END_FACTORY_REG(Lerp) +/** +*@brief Returns the num value of abs(x1-x2) > atol+rtol*abs(x2) element-wise. \n + +* +*@par Inputs: +*@li x1: A tensor. Must be one of the following types: float32, int32, uint8, int8, float16 +*@li x2: A tensor of the same type as "x1". +* +*@par Attributes: +* atol: Defaults to "1e-05". +* rtol: Defaults to "1e-03". +* +*@par Outputs: +* num: A tensor of type int32. +* diff: A tensor of type float16. +* +*@par Restrictions: +*Warning: THIS FUNCTION IS EXPERIMENTAL. Please do not use. +* +*/ +REG_OP(DataCompare) + .INPUT(x1, TensorType({ DT_FLOAT16, DT_FLOAT,DT_INT8, DT_UINT8, DT_INT32 })) + .INPUT(x2, TensorType({ DT_FLOAT16, DT_FLOAT,DT_INT8, DT_UINT8, DT_INT32 })) + .OUTPUT(num, TensorType({DT_FLOAT})) + .OUTPUT(diff, TensorType({DT_FLOAT16})) + .ATTR(atol, Float, 1e-5) + .ATTR(rtol, Float, 1e-3) + .OP_END_FACTORY_REG(DataCompare) + /** *@brief Hardmax(element in input, axis) = 1 if the element is the first maximum value along the specified axis, 0 *otherwise The input does not need to explicitly be a 2D vector.The "axis" attribute indicates the dimension along @@ -3650,6 +3679,57 @@ REG_OP(HardMax) .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(axis, Int, -1) .OP_END_FACTORY_REG(HardMax) + +/** +* @brief Computes the dot product (inner product) of two tensors. This function does not broadcast. + +* @par Inputs: +* Two inputs, including: +* @li input_x: A Tensor. the first tensor must be 1d. \n +* @li input_y: A Tensor. the second tensor must be 1d. \n + +* @par Outputs: +* @li output: A Tensor. Result of the two inputs, must be 1d. \n + +* @par Third-party framework compatibility +* Compatible with the Pytorch dot operator. \n +*/ +REG_OP(Dot) + .INPUT(input_x, TensorType({DT_FLOAT, DT_FLOAT16, DT_UINT8, DT_INT8, DT_INT32})) + .INPUT(input_y, TensorType({DT_FLOAT, DT_FLOAT16, DT_UINT8, DT_INT8, DT_INT32})) + .OUTPUT(output, TensorType({DT_FLOAT, DT_FLOAT16, DT_UINT8, DT_INT8, DT_INT32})) + .OP_END_FACTORY_REG(Dot) + +/** +*@brief Returns a new tensor with boolean elements representing \n +*if each element of input is “close” to the corresponding element of other \n + +*@par Inputs: +*Two inputs, including: +* @li x1: A tensor. Must be one of the following types: +* float16, float32, int32. \n +* @li x2: A tensor with the same type and shape of x1's. \n + +*@par Attributes: +*@li rtol: An optional float.Defaults to 1e-05. \n +*@li atol: An optional float.Defaults to 1e-08. \n +*@li equal_nan: An optional bool.Defaults to false. \n + +*@par Outputs: +*y: A Tensor bool with the same shape of x1's. \n + +*@par Third-party framework compatibility +*Compatible with the Pytorch operator isclose. \n +*/ +REG_OP(IsClose) + .INPUT(x1, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .INPUT(x2, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .OUTPUT(y, TensorType({DT_BOOL})) + .ATTR(rtol, Float, 1e-05) + .ATTR(atol, Float, 1e-08) + .ATTR(equal_nan, Bool, false) + .OP_END_FACTORY_REG(IsClose) + } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_ELEWISE_CALCULATION_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/hcom_ops.h b/third_party/fwkacllib/inc/ops/hcom_ops.h index cb9fbe22..f4ded0cd 100644 --- a/third_party/fwkacllib/inc/ops/hcom_ops.h +++ b/third_party/fwkacllib/inc/ops/hcom_ops.h @@ -238,6 +238,15 @@ REG_OP(HcomRemoteRead) .REQUIRED_ATTR(dtype, Type) .OP_END_FACTORY_REG(HcomRemoteRead) +/** + * @brief Performs Remote Ref Read of input tensors + * @par Inputs: + * remote: A tensor. describing the remote memory address to read: u64 remoteId, u64 addrRemote, u64 length + * cache_var: The local base address + * local_offset: Skip step length + * @par Outputs: + * cache_var: The local base address + */ REG_OP(HcomRemoteRefRead) .INPUT(remote, TensorType({DT_UINT64})) .INPUT(cache_var, TensorType({DT_UINT64})) @@ -258,6 +267,13 @@ REG_OP(HcomRemoteWrite) .INPUT(local, TensorType::ALL()) .OP_END_FACTORY_REG(HcomRemoteWrite) +/** + * @brief Performs Remote Write of input tensors + * @par Inputs: + * remote: A tensor. describing the remote memory address to write: u64 remoteId, u64 addrRemote, u64 length + * @par Inputs: + * local: A Tensor. whose value is length / size_of(Type) + */ REG_OP(HcomRemoteScatterWrite) .INPUT(remote, TensorType({DT_INT64, DT_UINT64})) .INPUT(local, TensorType::ALL()) diff --git a/third_party/fwkacllib/inc/ops/image_ops.h b/third_party/fwkacllib/inc/ops/image_ops.h index d7f60346..4703705b 100644 --- a/third_party/fwkacllib/inc/ops/image_ops.h +++ b/third_party/fwkacllib/inc/ops/image_ops.h @@ -652,6 +652,62 @@ REG_OP(RGBToHSV) /** *@brief Generate a single randomly distorted bounding box for an image . \n +*@par Inputs: +*Input images must be a 4-D tensor. Inputs include: +*@li image_size: 1-D, containing [height, width, channels]. +*@li bounding_boxes: 3-D with shape [batch, N, 4] describing the N bounding +boxes associated with the image. \n + +*@par Attributes: +*@li seed: If either seed or seed2 are set to non-zero, the random number +generator is seeded by the given seed. Otherwise, it is seeded by a random seed. +*@li seed2: A second seed to avoid seed collision. +*@li min_object_covered: The cropped area of the image must contain at least +this fraction of any bounding box supplied. The value of this parameter should +be non-negative. In the case of 0, the cropped area does not need to overlap +any of the bounding boxes supplied . +*@li aspect_ratio_range: The cropped area of the image must have an aspect +ratio = width / height within this range. +*@li max_attempts: Number of attempts at generating a cropped region of the +image of the specified constraints. After max_attempts failures, return the +entire image. +*@li use_image_if_no_bounding_boxes: Controls behavior if no bounding boxes +supplied. If true, assume an implicit bounding box covering the whole input. +If false, raise an error . \n + +*@par Outputs: +*@li begin: 1-D, containing [offset_height, offset_width, 0]. +*@li size: 1-D, containing [target_height, target_width, -1]. +*@li bboxes: 3-D with shape [1, 1, 4] containing the distorted bounding box . \n + +*@attention Constraints: +*Input images can be of different types but output images are always float . \n + +*@par Third-party framework compatibility +*Compatible with tensorflow SampleDistortedBoundingBox operator. +*/ + +REG_OP(SampleDistortedBoundingBox) + .INPUT(image_size, TensorType({ DT_UINT8, DT_INT8, DT_INT16, \ + DT_INT32, DT_INT64 })) + .INPUT(bounding_boxes, TensorType({ DT_FLOAT })) + .OUTPUT(begin, TensorType({ DT_UINT8, DT_INT8, DT_INT16, \ + DT_INT32, DT_INT64 })) + .OUTPUT(size, TensorType({ DT_UINT8, DT_INT8, DT_INT16, \ + DT_INT32, DT_INT64 })) + .OUTPUT(bboxes, TensorType({ DT_FLOAT })) + .ATTR(seed, Int, 0) + .ATTR(seed2, Int, 0) + .ATTR(min_object_covered, Float, 0.1f) + .ATTR(aspect_ratio_range, ListFloat, { 0.75f, 1.33f }) + .ATTR(area_range, ListFloat, { 0.05f, 1.0f }) + .ATTR(max_attempts, Int, 100) + .ATTR(use_image_if_no_bounding_boxes, Bool, false) + .OP_END_FACTORY_REG(SampleDistortedBoundingBox) + +/** +*@brief Generate a single randomly distorted bounding box for an image . \n + *@par Inputs: *Input images must be a 4-D tensor. Inputs include: *@li image_size: 1-D, containing [height, width, channels]. @@ -1424,11 +1480,11 @@ REG_OP(Resize) *@par Attributes: *@li channels: An optional int. Defaults to 0. Number of color channels for the decoded image. -*@li ratio: An optional int. Defaults to 1. Downscaling ratio. +*@li ratio: An optional int. Defaults to 1. Downscaling ratio. *@li fancy_upscaling: An optional bool. Defaults to True. If true use a slower but nicer upscaling of the chroma planes *@li try_recover_truncated: An optional bool. Defaults to False. If true try to recover an image from truncated input. *@li acceptable_fraction: An optional float. Defaults to 1. The minimum required fraction of lines before a truncated input is accepted. -*@li dct_method: An optional string. Defaults to "". string specifying a hint about the algorithm used for decompression. \n +*@li dct_method: An optional string. Defaults to "". string specifying a hint about the algorithm used for decompression. \n *@par Outputs: *image: A Tensor dtype of uint8. diff --git a/third_party/fwkacllib/inc/ops/linalg_ops.h b/third_party/fwkacllib/inc/ops/linalg_ops.h index d8f45c5d..330fef2e 100644 --- a/third_party/fwkacllib/inc/ops/linalg_ops.h +++ b/third_party/fwkacllib/inc/ops/linalg_ops.h @@ -61,8 +61,8 @@ REG_OP(CholeskyGrad) *@par Inputs: *The input x has to be symmetric and positive definite.Inputs include: -*x:A Tensor. Must be one of the following types: double, float32. Shape -is [..., M, M] . \n +*x:A Tensor. Must be one of the following types: double, float32, float16, +complex64, complex128. Shape is [..., M, M] . \n *@par Outputs: *y:A Tensor. Has the same type as x . \n @@ -76,8 +76,10 @@ form square matrices. */ REG_OP(Cholesky) - .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE, \ + DT_FLOAT16, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, \ + DT_FLOAT16, DT_COMPLEX64, DT_COMPLEX128})) .OP_END_FACTORY_REG(Cholesky) /** @@ -87,8 +89,8 @@ of one or more square matrices . \n *@par Inputs: *The input x is a tensor of shape [N, M, M] whose inner-most 2 dimensions form square matrices. Inputs include: -*x:A Tensor. Must be one of the following types: double, float32. Shape is -[..., M, M] . \n +*x:A Tensor. Must be one of the following types: double, float32, +complex64, complex128. Shape is [..., M, M] . \n *@par Outputs: *@li y:A Tensor. Has the same type as x. @@ -103,9 +105,9 @@ form square matrices. \n */ REG_OP(LogMatrixDeterminant) - .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(sign, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(sign, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) .OP_END_FACTORY_REG(LogMatrixDeterminant) /** @@ -114,8 +116,8 @@ REG_OP(LogMatrixDeterminant) *@par Inputs: *The input x is a tensor of shape [N, M, M] whose inner-most 2 dimensions form square matrices. Inputs include: -*x:A Tensor. Must be one of the following types: double, float32. Shape is -[..., M, M] . \n +*x:A Tensor. Must be one of the following types: double, float32, complex64, +complex128. Shape is [..., M, M] . \n *@par Outputs: *y:A Tensor. Has the same type as x . \n @@ -129,8 +131,8 @@ form square matrices. */ REG_OP(MatrixDeterminant) - .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) .OP_END_FACTORY_REG(MatrixDeterminant) /** @@ -140,8 +142,7 @@ their adjoints (conjugate transposes) . \n *@par Inputs: *The input x is a tensor of shape [..., M, M] whose inner-most 2 dimensions form square matrices. Inputs include: -*x:A Tensor. Must be one of the following types: double, float. Shape is -[..., M, M] . \n +*x:A Tensor of input. Shape is [..., M, M] . \n *@par Attributes: *adjoint:An optional bool. Defaults to False.Boolean indicating whether to @@ -159,8 +160,10 @@ form square matrices. \n */ REG_OP(MatrixInverse) - .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(x, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .ATTR(adjoint, Bool, false) .OP_END_FACTORY_REG(MatrixInverse) @@ -169,8 +172,7 @@ REG_OP(MatrixInverse) *@par Inputs: *The input rhs must have the same type as matrix. Inputs include: -*@li matrix:A Tensor. Must be one of the following types: double, float. -Shape is [..., M, M]. +*@li matrix:A Tensor of input. Shape is [..., M, M]. *@li rhs:A Tensor. Must have the same type as matrix. Shape is [..., M, K] . \n *@par Attributes: @@ -189,9 +191,9 @@ dimensions form square matrices. \n */ REG_OP(MatrixSolve) - .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE})) - .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) .ATTR(adjoint, Bool, false) .OP_END_FACTORY_REG(MatrixSolve) @@ -221,8 +223,10 @@ dimensions form square matrices. \n */ REG_OP(MatrixSolveLs) - .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE})) - .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) + .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .INPUT(l2, TensorType({DT_DOUBLE})) .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) .ATTR(fast, Bool, true) @@ -234,8 +238,7 @@ matrices by backsubstitution . \n *@par Inputs: *The input rhs must have the same type as matrix. Inputs include: -*@li matrix: A Tensor. Must be one of the following types: double, float. -Shape is [..., M, M]. +*@li matrix: A Tensor. Shape is [..., M, M]. *@li rhs:A Tensor. Must have the same type as matrix. Shape is [..., M, K] . \n *@par Attributes: @@ -256,9 +259,12 @@ dimensions form square matrices. \n */ REG_OP(MatrixTriangularSolve) - .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE})) - .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(matrix, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) + .INPUT(rhs, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_FLOAT16, \ + DT_COMPLEX64, DT_COMPLEX128})) .ATTR(lower, Bool, true) .ATTR(adjoint, Bool, false) .OP_END_FACTORY_REG(MatrixTriangularSolve) @@ -268,8 +274,7 @@ REG_OP(MatrixTriangularSolve) *@par Inputs: *The input shape of x must be [..., M, N]. Inputs include: -*x:A Tensor whose shape is [..., M, N]. Must be one of the following types: -double, float . \n +*x:A Tensor whose shape is [..., M, N]. \n *@par Attributes: *full_matrices: An optional bool. Defaults to False. If true, compute @@ -289,9 +294,12 @@ dimensions form matrices of size [M, N]. \n */ REG_OP(Qr) - .INPUT(x, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE })) - .OUTPUT(q, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE })) - .OUTPUT(r, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE })) + .INPUT(x, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE, \ + DT_COMPLEX64, DT_COMPLEX128 })) + .OUTPUT(q, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE, \ + DT_COMPLEX64, DT_COMPLEX128 })) + .OUTPUT(r, TensorType({ DT_FLOAT16, DT_FLOAT, DT_DOUBLE, \ + DT_COMPLEX64, DT_COMPLEX128 })) .ATTR(full_matrices, Bool, false) .OP_END_FACTORY_REG(Qr) @@ -384,8 +392,8 @@ of the rows encoded as a list of indices in `0..M-1`. Shape is `[..., M]` . \n */ REG_OP(Lu) - .INPUT(input, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(lu, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(input, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(lu, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) .OUTPUT(p, TensorType({DT_INT32, DT_INT64})) .REQUIRED_ATTR(output_idx_type, Type) .OP_END_FACTORY_REG(Lu) @@ -404,8 +412,8 @@ y: Shape is `[..., M, M]` . \n */ REG_OP(MatrixSquareRoot) - .INPUT(input, TensorType({DT_FLOAT, DT_DOUBLE})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(input, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_COMPLEX64, DT_COMPLEX128})) .OP_END_FACTORY_REG(MatrixSquareRoot) /** diff --git a/third_party/fwkacllib/inc/ops/list_ops.h b/third_party/fwkacllib/inc/ops/list_ops.h index 292b1dbe..33270ea8 100644 --- a/third_party/fwkacllib/inc/ops/list_ops.h +++ b/third_party/fwkacllib/inc/ops/list_ops.h @@ -192,7 +192,7 @@ REG_OP(TensorListGetItem) .INPUT(element_shape, TensorType({DT_INT32})) .OUTPUT(item, TensorType({DT_FLOAT16,DT_FLOAT,DT_DOUBLE,DT_INT8, DT_INT16,DT_INT32,DT_INT64,DT_UINT8,DT_UINT16,DT_QINT8,DT_QUINT8, - DT_QINT16,DT_QUINT16,DT_QINT32,DT_BOOL,DT_RESOURCE, + DT_QINT16,DT_QUINT16,DT_QINT32,DT_BOOL, DT_STRING,DT_COMPLEX64,DT_COMPLEX128})) .ATTR(element_dtype, Type, DT_INT32) .OP_END_FACTORY_REG(TensorListGetItem) diff --git a/third_party/fwkacllib/inc/ops/math_ops.h b/third_party/fwkacllib/inc/ops/math_ops.h index 4cbcc027..50d058ba 100644 --- a/third_party/fwkacllib/inc/ops/math_ops.h +++ b/third_party/fwkacllib/inc/ops/math_ops.h @@ -222,6 +222,24 @@ REG_OP(Bucketize) .REQUIRED_ATTR(boundaries, ListFloat) .OP_END_FACTORY_REG(Bucketize) +/** +*@brief Returns a new tensor with the truncated integer values of the elements of input. \n + +*@par Inputs: +*One inputs, including: +* @li input_x: A tensor. Must be one of the following types: float16, float32, int8, uint8, int32. \n + +*@par Outputs: +*y: A tensor with the same type and shape of input_x \n + +*@par Third-party framework compatibility +*Compatible with the Pytorch operator Trunc. \n +*/ +REG_OP(Trunc) + .INPUT(input_x, TensorType({DT_FLOAT16,DT_FLOAT, DT_INT8, DT_INT32, DT_UINT8})) + .OUTPUT(output_y, TensorType({DT_FLOAT16,DT_FLOAT, DT_INT8, DT_INT32, DT_UINT8})) + .OP_END_FACTORY_REG(Trunc) + /** *@brief Computes the sum along sparse segments of a tensor . \n @@ -645,6 +663,7 @@ REG_OP(NLLLoss) .OUTPUT(y, TensorType({DT_FLOAT})) .OUTPUT(total_weight, TensorType({DT_FLOAT})) .ATTR(reduction, String, "mean") + .ATTR(ignore_index, Int, -100) .OP_END_FACTORY_REG(NLLLoss) /** @@ -674,6 +693,7 @@ REG_OP(NLLLossGrad) .INPUT(total_weight, TensorType({DT_FLOAT})) .OUTPUT(x_grad, TensorType({DT_FLOAT})) .ATTR(reduction, String, "mean") + .ATTR(ignore_index, Int, -100) .OP_END_FACTORY_REG(NLLLossGrad) /** @@ -884,6 +904,54 @@ REG_OP(LpNorm) .ATTR(keepdim, Bool, false) .ATTR(epsilon, Float, 1e-12) .OP_END_FACTORY_REG(LpNorm) + +/** +* @brief get complex. + +* @par Inputs: +* @li real: An ND tensor of type float32. double +* @li imag: An ND tensor of type float32. double \n +* +* @par Outputs: +* @li out: An ND tensor of type complex64, complex128 \n +*/ +REG_OP(Complex) + .INPUT(real, TensorType({DT_FLOAT, DT_DOUBLE})) + .INPUT(imag, TensorType({DT_FLOAT, DT_DOUBLE})) + .OUTPUT(out, TensorType({DT_COMPLEX64, DT_COMPLEX128})) + .ATTR(Tout, Type, DT_COMPLEX64) + .OP_END_FACTORY_REG(Complex) + +/** +* @brief deal complex. + +* @par Inputs: +* @li input: An ND tensor of type complex64, complex128 \n +* +* @par Outputs: +* @li output: An ND tensor of type float32. double \n +*/ +REG_OP(Imag) + .INPUT(input, TensorType({DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(output, TensorType({DT_FLOAT, DT_DOUBLE})) + .ATTR(Tout, Type, DT_FLOAT) + .OP_END_FACTORY_REG(Imag) + +/** +* @brief deal complex. + +* @par Inputs: +* @li input: An ND tensor of type complex64, complex128 \n +* +* @par Outputs: +* @li output: An ND tensor of type float32. double \n +*/ +REG_OP(Angle) + .INPUT(input, TensorType({DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(output, TensorType({DT_FLOAT, DT_DOUBLE})) + .ATTR(Tout, Type, DT_FLOAT) + .OP_END_FACTORY_REG(Angle) + } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_MATH_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/matrix_calculation_ops.h b/third_party/fwkacllib/inc/ops/matrix_calculation_ops.h index 33b596d8..6bff7f82 100644 --- a/third_party/fwkacllib/inc/ops/matrix_calculation_ops.h +++ b/third_party/fwkacllib/inc/ops/matrix_calculation_ops.h @@ -1022,6 +1022,27 @@ REG_OP(IndexAdd) .ATTR(axis, Int, 0) .OP_END_FACTORY_REG(IndexAdd) +/** +*@brief: Returns the upper triangular part of a matrix (2-D tensor) or batch of matrices input \n + +*@par Inputs: +* Two inputs, including: +*@li x: A Tensor. Must be one of the following types: +* float16, float32, double, int32, uint8, int16, int8, complex64, int64, +* qint8, quint8, qint32, uint16, complex128, uint32, uint64. +*@li diagonal:(int, optional) – the diagonal to consider。\n + +*@par Outputs: +*y: A Tensor. Has the same type as "x" . \n + +*@par Third-party framework compatibility +* Compatible with the Pytorch operator Triu. +*/ +REG_OP(Triu) + .INPUT(x, TensorType::BasicType()) + .ATTR(diagonal, Int, 0) + .OUTPUT(y, TensorType::BasicType()) + .OP_END_FACTORY_REG(Triu) } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_MATRIX_CALCULATION_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/nn_batch_norm_ops.h b/third_party/fwkacllib/inc/ops/nn_batch_norm_ops.h index a35cee03..ddd70bc8 100644 --- a/third_party/fwkacllib/inc/ops/nn_batch_norm_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_batch_norm_ops.h @@ -144,6 +144,64 @@ REG_OP(BatchNorm) /** *@brief Performs batch normalization . \n +*@par Inputs: +* Five inputs, including: (NHWC, NCHW, or NC1HWC0 supported) +*@li x: A 3D or 6D Tensor of type float16 or float32, with format NDHWC or NCDHW for 4D or NDC1HWC0 for 6D. +*@li scale: A Tensor of type float32. Must be 1D if input "x" is with format NDHWC or NCDHW. Must be 6D +if input "x" is with format NDC1HWC0. Specifies the scaling factor. +*@li offset: A Tensor of type float32. Must be 3D if input "x" is with format NDHWC or NCDHW. Must be 6D +if input "x" is with format NC1HWC0. Specifies the offset. +*@li mean: A Tensor of type float32. Must be 3D if input "x" is with format NDHWC or NCDHW. Must be 6D +if input "x" is with format NC1HWC0. Specifies the mean used for inference. Must be "None" if the +operation is used for training. +*@li variance: A Tensor of type float32. Must be 3D if input "x" is with format NHWC or NCHW. Must be +5D if input "x" is with format NC1HWC0. Specifies the variance used for inference. Must be "None" +if the operation is used for training . \n + +*@par Attributes: +*@li epsilon: An optional float32, specifying the small value added to variance to avoid dividing by zero. Defaults to "0.0001". +*@li data_format: An optional string, specifying the format of "x". Defaults to "NHWC". +*@li is_training: An optional bool, specifying if the operation is used for training or inference. Defaults to "True" . \n + +*@par Outputs: +* Five outputs, including: (NHWC, NCHW, or NC1HWC0 supported) +*@li y: A 3D or 6D Tensor of type float16 or float32 for the normalized "x", with format NDHWC or NCDHW for 4D or NDC1HWC0 for 6D. +*@li batch_mean: A Tensor of type float32. Must be 3D if input "x" is with format NDHWC or NCDHW. Must be 6D +if input "x" is with format NDC1HWC0. Specifies the mean of "x". +*@li batch_variance: A Tensor of type float32. Must be 1D if input "x" is with format NDHWC or NCDHW. +Must be 6D if input "x" is with format NDC1HWC0. Specifies the variance of "x". +*@li reserve_space_1: An optional Tensor of type float32. Must be 1D if input "x" is with format NDHWC or NCDHW. +Must be 6D if input "x" is with format NDC1HWC0. Specifies the mean of "x" for gradient computation. Pass "None" to skip this output. +*@li reserve_space_2: An optional Tensor of type float32. Must be 1D if input "x" is with format NHWC or NCHW. +Must be 6D if input "x" is with format NDC1HWC0. Specifies the variance of "x" for gradient computation. Pass "None" to skip this output . \n + +*@attention Constraints: +*@li If the operation is used for inference and outputs "reserve_space_1" and "reserve_space_2" are available, +then "reserve_space_1" has the same value as "mean" and "reserve_space_2" has the same value as "variance". +*@li For Ascend 310, the result accuracy fails to reach 1‰ due to the square root instruction . \n + +*@par Third-party framework compatibility +*@li Compatible with the TensorFlow operator fused_batch_norm. +*@li Compatible with the TensorFlow operator fused_batch_norm_v2. +*/ +REG_OP(BatchNorm3D) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(scale, TensorType({DT_FLOAT})) + .INPUT(offset, TensorType({DT_FLOAT})) + .OPTIONAL_INPUT(mean, TensorType({DT_FLOAT})) + .OPTIONAL_INPUT(variance, TensorType({DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(batch_mean, TensorType({DT_FLOAT})) + .OUTPUT(batch_variance, TensorType({DT_FLOAT})) + .OUTPUT(reserve_space_1, TensorType({DT_FLOAT})) + .OUTPUT(reserve_space_2, TensorType({DT_FLOAT})) + .ATTR(epsilon, Float, 0.0001) + .ATTR(data_format, String, "NCDHW") + .ATTR(is_training, Bool, true) + .OP_END_FACTORY_REG(BatchNorm3D) +/** +*@brief Performs batch normalization . \n + *@par Inputs: * Five inputs, including: (NHWC or NCHW supported) *@li x: A 4D Tensor of type float16 or float32. @@ -242,6 +300,52 @@ REG_OP(BatchNormGrad) /** *@brief Performs the backpropagation of BatchNorm . \n +*@par Inputs: +* Five inputs, including: +*@li y_backprop: A 3D or 6D Tensor of type float16 or float32, with format NDHWC, NCDHW, or NDC1HWC0, for the gradient. +*@li x: A 3D or 6D Tensor of type float16 or float32, with format NDHWC, NCDHW, or NDC1HWC0. +*@li scale: A 3D or 6D Tensor of type float32, with format NDHWC, NCDHW, or NDC1HWC0. +*@li reserve_space_1: A 3D or 6D Tensor of type float32, with format NDHWC, NCDHW, or NC1HWC0. It is an output of BatchNorm. +*@li reserve_space_2: A 3D or 6D Tensor of type float32, with format NDHWC, NCDHW, or NC1HWC0. It is an output of BatchNorm . \n + +*@par Attributes: +*@li epsilon: An optional float32. Defaults to "0.0001". A small float number added to the variance of "x". +*@li data_format: An optional string. Defaults to "NCDHW". +*@li is_training: An optional bool. Defaults to "true". Specifies the operation is for training (default) or inference . \n + +*@par Outputs: +*@li x_backprop: A Tensor of type float16 or float32, with format NHWC, NCHW, or NC1HWC0, for the offset of "x". +*@li scale_backprop: A Tensor of type float32, with format NDHWC, NCDHW, or NDC1HWC0, for the offset of "scale". +*@li *offset_backprop: A Tensor of type float32, with format NDHWC, NCDHW, or NDC1HWC0, for the offset of "offset". +*@li *reserve_space_4: A Tensor of type float32, with shape NDHWC, NCDHW, or NDC1HWC0. Pass "None" to skip this output. +*@li *reserve_space_5: A Tensor of type float32, with shape NDHWC, NCDHW, or NDC1HWC0. Pass "None" to skip this output . \n + +*@attention Constraints: +* The preceding layer of this operator must be operator BatchNorm . \n + +*@see BatchNorm +*@par Third-party framework compatibility +* Compatible with the TensorFlow operators FusedBatchNormGradV2 and FusedBatchNorm3DGrad. +*/ +REG_OP(BatchNorm3DGrad) + .INPUT(y_backprop, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(scale, TensorType({DT_FLOAT})) + .INPUT(reserve_space_1, TensorType({DT_FLOAT})) + .INPUT(reserve_space_2, TensorType({DT_FLOAT})) + .OUTPUT(x_backprop, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(scale_backprop, TensorType({DT_FLOAT})) + .OUTPUT(offset_backprop, TensorType({DT_FLOAT})) + .OUTPUT(reserve_space_4, TensorType({DT_FLOAT})) + .OUTPUT(reserve_space_5, TensorType({DT_FLOAT})) + .ATTR(epsilon, Float, 0.0001) + .ATTR(data_format, String, "NCDHW") + .ATTR(is_training, Bool, true) + .OP_END_FACTORY_REG(BatchNorm3DGrad) + +/** +*@brief Performs the backpropagation of BatchNorm . \n + *@par Inputs: * Five inputs, including: *@li y_backprop: A 4D Tensor of type float16 or float32, with format NHWC or NCHW, for the gradient. diff --git a/third_party/fwkacllib/inc/ops/nn_calculation_ops.h b/third_party/fwkacllib/inc/ops/nn_calculation_ops.h index c848668f..53922ee6 100644 --- a/third_party/fwkacllib/inc/ops/nn_calculation_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_calculation_ops.h @@ -1059,7 +1059,7 @@ REG_OP(DeformableConv2D) *@par Attributes: * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. * @li dilations: A list of 5 integers. Specifies the dilation factor for each @@ -1119,7 +1119,7 @@ REG_OP(Conv3D) *@par Attributes: * Three attributes: * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. * @li dilations: A tuple/list of 5 integers, The dilation factor for each @@ -1167,7 +1167,7 @@ REG_OP(Conv3DBackpropInput) *@par Attributes: * Three attributes: * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. * @li dilations: A tuple/list of 5 integers, The dilation factor for each @@ -1267,7 +1267,7 @@ REG_OP(LSTM) * dimension of input. * The N, C and D dimensions must be 1. Has the same format as "x". * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. @@ -1319,7 +1319,7 @@ REG_OP(Conv3DBackpropFilter) * dimension of input. * The N, C and D dimensions must be 1. Has the same format as "x". * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. @@ -1369,7 +1369,7 @@ REG_OP(Conv3DBackpropFilterD) *@par Attributes: * Five attributes: * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li dilations: A tuple/list of 5 integers, * The dilation factor for each dimension of input. * The N, C and D dimensions must be 1. Has the same format as "x". @@ -1422,7 +1422,7 @@ REG_OP(Conv3DTranspose) * dimension of input. * The N, C and D dimensions must be 1. Has the same format as "x". * @li groups: Number of blocked connections from input channels to output - * channels. Reserved. + * channels. * @li data_format: An optional string from: "NDHWC", "NCDHW". * Defaults to "NDHWC". Specify the data format of the input and output data. * @li output_padding: The size will be added in the output shape. @@ -1624,7 +1624,7 @@ REG_OP(Conv2DTransposeD) * of the input. * @li ksize: A tuple/list of 2 integers.kernel size. *@par Attributes: - * Three attributes: + * Four attributes: * @li dilations: A tuple/list of 4 integers, The dilation factor for each dimension * of input. Defaults to [1, 1, 1, 1] * @li data_format: An optional string from: "NCHW", "NHWC". Defaults to "NCHW". Specify the data format of the input x. diff --git a/third_party/fwkacllib/inc/ops/nn_detect_ops.h b/third_party/fwkacllib/inc/ops/nn_detect_ops.h index 39b4b227..af59b4e2 100644 --- a/third_party/fwkacllib/inc/ops/nn_detect_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_detect_ops.h @@ -968,8 +968,9 @@ REG_OP(SPP) * Three inputs, including: *@li x: An NC1HWC0 tensor of type float16 or float32, describing the feature * map. -*@li rois: A tensor of type float16 or float32, with shape +*@li rois: A tensor of type float16 or float32, with 3D shape * [batch, 5, roi_max_num], describing the RIOs. +* roi_max_num must be less than or equal to 6000 and must be divided by 16. *@li roi_actual_num: A optional tensor of type int32, with shape [batch, 8], specifying * the number of ROIs per batch . \n @@ -1604,6 +1605,50 @@ REG_OP(NonMaxSuppressionV7) .ATTR(max_boxes_size, Int, 0) .OP_END_FACTORY_REG(NonMaxSuppressionV7) +/** +*@brief Obtains the ROI feature matrix from the feature map list. It is a customized fused operator for mmdetection. \n + +*@par Inputs: +* Three inputs, including: +*@li features: A 5HD Tensor list of type float32 or float16. +*@li rois: ROI position. A 2D Tensor of float32 or float16 with shape (N, 5). "N" indicates the number of ROIs, +* the value "5" indicates the indexes of images where the ROIs are located, "x0", "y0", "x1", and "y1". + +*@par Attributes: +*@li finest_scale: A optional attribute of type int, specifying the scale of calculate levels of "rois". +*@li roi_scale_factor: A optional attribute of type float32, specifying the rescaling of "rois" coordinates. +*@li spatial_scale: A optional attribute of type list float32, specifying the scaling ratio of "features" +* to the original image. +*@li pooled_height: A optional attribute of type int32, specifying the H dimension. +*@li pooled_width: A optional attribute of type int32, specifying the W dimension. +*@li sample_num: An optional attribute of type int32, specifying the horizontal and vertical sampling frequency +* of each output. If this attribute is set to "0", the sampling frequency is equal to the rounded up value of "rois", +* which is a floating point number. Defaults to "0". +*@li pool_mode: An optional attribute of type string to indicate pooling mode. Defaults to "avg" . \n +*@li aligned: An optional attribute of type bool, specifying the align to corner. Defaults to true . \n + +*@par Outputs: +* output: Outputs the feature sample of each ROI position. The format is 5HD Tensor of type float32 or float16. +* The axis N is the number of input ROIs. Axes H, W, and C are consistent with the values of "pooled_height", +* "pooled_width", and "features", respectively. + +*@par Third-party framework compatibility +*Compatible with mmdetection SingleRoIExtractor operator. +*/ +REG_OP(RoiExtractor) + .DYNAMIC_INPUT(features, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(rois, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .ATTR(finest_scale, Int, 56) + .ATTR(roi_scale_factor, Float, 0) + .ATTR(spatial_scale, ListFloat, { 1.f/4, 1.f/8, 1.f/16, 1.f/32 }) + .ATTR(pooled_height, Int, 7) + .ATTR(pooled_width, Int, 7) + .ATTR(sample_num, Int, 0) + .ATTR(pool_mode, String, "avg") + .ATTR(aligned, Bool, true) + .OP_END_FACTORY_REG(RoiExtractor) + } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_NN_DETECT_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/nn_norm_ops.h b/third_party/fwkacllib/inc/ops/nn_norm_ops.h index af223552..00e2020f 100644 --- a/third_party/fwkacllib/inc/ops/nn_norm_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_norm_ops.h @@ -1233,6 +1233,47 @@ REG_OP(SigmoidCrossEntropyWithLogitsGradV2) .OUTPUT(gradient, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(reduction, String, "mean") .OP_END_FACTORY_REG(SigmoidCrossEntropyWithLogitsGradV2) +/** + * @brief Calculate the PoissonNllLoss function. + * target∼Poisson(input)loss(input,target)=input−target∗log(input)+log(target!) \n + + * @par Inputs: + * Two inputs, including: + * @li input_x: A tensor. Must be one of the following types: + * float16, float32. \n + * + * @par Inputs: + * @li target: A tensor. Must be one of the following types: + * float16, float32. \n + + * @par Attributes: + * four Attributes, including: + * @li log_input: An optional bool. Defaults to "True" \n + * + * @par Attributes: + * @li full: An optional bool. Defaults to "False" \n + * + * @par Attributes: + * @li eps: An optional float. Defaults to "1e-8" \n + * + * @par Attributes: + * @li reduction: An optional string. Defaults to "mean" \n + + * @par Outputs: + * loss: A Tensor has same element type as two inputs. \n + + * @par Third-party framework compatibility + * Compatible with the Pytorch operator PoissonNllLoss. \n + */ +REG_OP(PoissonNllLoss) + .INPUT(input_x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(target, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(loss, TensorType({DT_FLOAT16, DT_FLOAT})) + .ATTR(log_input, Bool, true) + .ATTR(full, Bool, false) + .ATTR(eps, Float, 1e-8) + .ATTR(reduction, String, "mean") + .OP_END_FACTORY_REG(PoissonNllLoss) } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_NN_NORM_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/nn_ops.h b/third_party/fwkacllib/inc/ops/nn_ops.h index 16552eee..820aa00d 100644 --- a/third_party/fwkacllib/inc/ops/nn_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_ops.h @@ -20,7 +20,34 @@ */ #ifndef OPS_BUILT_IN_OP_PROTO_INC_NN_OPS_H_ #define OPS_BUILT_IN_OP_PROTO_INC_NN_OPS_H_ - +#include "graph/operator_reg.h" #include "nn_pooling_ops.h" +namespace ge { +/** +* @brief Says whether the targets are in the top "k" predictions . \n + +* @par Inputs: +* Three inputs, including: +* @li predictions: A 2D Tensor of type float32. A "batch_size * classes" tensor. +* @li targets: A 1D Tensor of type IndexNumberType. A batch_size tensor of class ids. +* @li k: A 1D Tensor of the same type as "targets". +* Specifies the number of top elements to look at for computing precision . \n + +* @par Outputs: +* precision: A Tensor of type bool . \n + +* @attention Constraints: +* @li targets must be non-negative tensor. + +* @par Third-party framework compatibility +* @li Compatible with the TensorFlow operator InTopKV2. +*/ +REG_OP(InTopKV2) + .INPUT(predictions, TensorType({DT_FLOAT})) + .INPUT(targets, TensorType(IndexNumberType)) + .INPUT(k, TensorType({IndexNumberType})) + .OUTPUT(precision, TensorType({DT_BOOL})) + .OP_END_FACTORY_REG(InTopKV2) +}// namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_NN_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/nonlinear_fuc_ops.h b/third_party/fwkacllib/inc/ops/nonlinear_fuc_ops.h index e0897280..a225bb5f 100644 --- a/third_party/fwkacllib/inc/ops/nonlinear_fuc_ops.h +++ b/third_party/fwkacllib/inc/ops/nonlinear_fuc_ops.h @@ -223,7 +223,29 @@ REG_OP(Relu6Grad) .INPUT(features, TensorType::RealNumberType()) .OUTPUT(backprops, TensorType::RealNumberType()) .OP_END_FACTORY_REG(Relu6Grad) - +/** +*@brief Calculate the elu_grad_v2 function. +*Applies the element-wise function: +* Computes the backward for the elu: if x>0, 1; otherwise elu() + alpha . +*@par Inputs: +*One inputs, including: +* @li grads: A tensor. Must be one of the following types: +* float16, float32. +* @li activations: A tensor. Must be one of the following types: +* float16, float32. +* +*@par Outputs: +*y: A Tensor with the same type and shape of grads's. +* +*@par Attributes: +*@li alpha: scalar parameter, default value = 1.0 +*/ +REG_OP(EluGradV2) + .INPUT(grads, TensorType({DT_FLOAT, DT_FLOAT16})) + .INPUT(activations, TensorType({DT_FLOAT, DT_FLOAT16})) + .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT16})) + .ATTR(alpha, Float, 1.0) + .OP_END_FACTORY_REG(EluGradV2) /** * @brief Compute sigmoid of "x" element-wise . \n @@ -842,6 +864,26 @@ REG_OP(SoftShrinkGrad) .OUTPUT(output_y, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(lambd, Float, 0.5) .OP_END_FACTORY_REG(SoftShrinkGrad) + +/** +*@brief Calculate -ln(1+e^(-x)). \n + +*@par Inputs: +*One inputs, including: +* @li x: A tensor. Must be one of the following types: +* float16, float32. \n + +*@par Outputs: +*One outputs, including: +* @li y: A tensor with the same type and shape of x's. \n + +*@par Third-party framework compatibility +*Compatible with the Pytorch operator LogSigmoid. \n +*/ +REG_OP(LogSigmoid) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) /* "input:x" */ + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) /* "output:y" */ + .OP_END_FACTORY_REG(LogSigmoid) } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_NONLINEAR_FUC_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/reduce_ops.h b/third_party/fwkacllib/inc/ops/reduce_ops.h index 0b114134..5b97d226 100644 --- a/third_party/fwkacllib/inc/ops/reduce_ops.h +++ b/third_party/fwkacllib/inc/ops/reduce_ops.h @@ -37,7 +37,7 @@ namespace ge { *@attention Constraints: * This operator is a BatchNorm fusion operator for updating the moving * averages for training. -* This operator is used in conjunction with BNTrainingUpdate. +* This operator is used in conjunction with BNTrainingReduce. */ REG_OP(BNTrainingReduce) .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) @@ -45,6 +45,27 @@ REG_OP(BNTrainingReduce) .OUTPUT(square_sum, TensorType({DT_FLOAT})) .OP_END_FACTORY_REG(BNTrainingReduce) +/** +*@brief Performs reduced batch normalization . \n + +*@par Inputs: +*x: A 6D Tensor of type float16 or float32, with format NDC1HWC0 . \n + +*@par Outputs: +*@li sum: A 3D Tensor of type float32 for SUM reduced "x". +*@li square_sum: A 3D Tensor of type float32 for SUMSQ reduced "x" . \n + +*@attention Constraints: +* This operator is a BatchNorm fusion operator for updating the moving +* averages for training. +* This operator is used in conjunction with BN3DTrainingReduce. +*/ +REG_OP(BN3DTrainingReduce) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(sum, TensorType({DT_FLOAT})) + .OUTPUT(square_sum, TensorType({DT_FLOAT})) + .OP_END_FACTORY_REG(BN3DTrainingReduce) + /** *@brief Performs the backpropagation of BatchNorm . \n @@ -88,6 +109,49 @@ REG_OP(BNTrainingReduceGrad) .ATTR(epsilon, Float, 0.0001) .OP_END_FACTORY_REG(BNTrainingReduceGrad) +/** +*@brief Performs the backpropagation of BatchNorm . \n + +*@par Inputs: +* Seven inputs, including: +*@li grads: A 6D Tensor of type float16 or float32, with format NDC1HWC0, for +* the gradient. +*@li x: A 6D Tensor of type float16 or float32, with format NDC1HWC0. +*@li diff_scale: A 6D Tensor of type float32, with format NDC1HWC0, +* for the mean of "x". +*@li diff_offset: A 6D Tensor of type float32, with format NDC1HWC0, +* for the variance of "x". +*@li scale: A 6D Tensor of type float32, with format NDC1HWC0. +*@li batch_mean: A 6D Tensor of type float32, with format NDC1HWC0, +* for the mean of "x". +*@li batch_variance: A 6D Tensor of type float32, with format NDC1HWC0, +* for the variance of "x" . \n + +*@par Attributes: +*epsilon: An optional float32. Defaults to "0.0001". A small float number +* added to the variance of "x" . \n + +*@par Outputs: +*y: A Tensor of type float16 or float32, with format NDC1HWC0, for the offset +* of "x" . \n + +*@attention Constraints: +* The preceding layer of this operator must be BN3DTrainingReduceGrad . \n + +*@see BN3DTrainingReduceGrad +*/ +REG_OP(BN3DTrainingReduceGrad) + .INPUT(grads, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(diff_scale, TensorType({DT_FLOAT})) + .INPUT(diff_offset, TensorType({DT_FLOAT})) + .INPUT(scale, TensorType({DT_FLOAT})) + .INPUT(batch_mean, TensorType({DT_FLOAT})) + .INPUT(batch_variance, TensorType({DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16,DT_FLOAT})) + .ATTR(epsilon, Float, 0.0001) + .OP_END_FACTORY_REG(BN3DTrainingReduceGrad) + /** *@brief Performs reduced batch normalization . \n @@ -120,7 +184,7 @@ REG_OP(BNTrainingReduceGrad) *@attention Constraints: *@li This operator is a BatchNorm fusion operator for updating the moving averages for training. -*This operator is used in conjunction with BNTrainingReduce. +*This operator is used in conjunction with BNTrainingUpdate. *@li For Ascend 310, the result accuracy fails to reach 1‰ due to the square * root instruction. */ @@ -141,6 +205,59 @@ REG_OP(BNTrainingUpdate) .OUTPUT(batch_variance, TensorType({DT_FLOAT})) .OP_END_FACTORY_REG(BNTrainingUpdate) +/** +*@brief Performs reduced batch normalization . \n + +*@par Inputs: +* Seven inputs, including: (NDC1HWC0 supported) +*@li x: A 6D Tensor of type float16 or float32. +*@li sum: A 6D Tensor of type float32 for the output of operator +* BN3DTrainingUpdate. +*@li square_sum: A 6D Tensor of type float32 for the output of operator +* BN3DTrainingUpdate. +*@li scale: A 6D Tensor of type float32, for the scaling factor. +*@li offset: A 6D Tensor of type float32, for the scaling offset. +*@li mean: A 6D Tensor of type float32, for the updated mean. +*@li variance: A 6D Tensor of type float32, for the updated variance . \n + +*@par Attributes: +*@li epsilon: A required float32, specifying the small value added to variance +* to avoid dividing by zero. +*@li factor: A required float32, specifying the weight for updating the mean +* and variance . \n + +*@par Outputs: +* Five outputs, including: (NDC1HWC0 supported) +*@li y: A 6D Tensor of type float16 or float32, for normalized "x". +*@li mean: A 6D Tensor of type float32, for the updated mean. +*@li variance: A 6D Tensor of type float32, for the updated variance. +*@li batch_mean: A 6D Tensor of type float32, for the mean of "x". +*@li batch_variance: A 6D Tensor of type float32, for the variance of "x" . \n + +*@attention Constraints: +*@li This operator is a BatchNorm fusion operator for updating the moving +averages for training. +*This operator is used in conjunction with BN3DTrainingUpdate. +*@li For Ascend 310, the result accuracy fails to reach 1‰ due to the square +* root instruction. +*/ +REG_OP(BN3DTrainingUpdate) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(sum, TensorType({DT_FLOAT})) + .INPUT(square_sum, TensorType({DT_FLOAT})) + .INPUT(scale, TensorType({DT_FLOAT})) + .INPUT(offset, TensorType({DT_FLOAT})) + .INPUT(mean, TensorType({DT_FLOAT})) + .INPUT(variance, TensorType({DT_FLOAT})) + .REQUIRED_ATTR(factor, Float) + .REQUIRED_ATTR(epsilon, Float) + .OUTPUT(y, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(mean, TensorType({DT_FLOAT})) + .OUTPUT(variance, TensorType({DT_FLOAT})) + .OUTPUT(batch_mean, TensorType({DT_FLOAT})) + .OUTPUT(batch_variance, TensorType({DT_FLOAT})) + .OP_END_FACTORY_REG(BN3DTrainingUpdate) + /** *@brief Performs batch normalization for inference . \n @@ -284,6 +401,40 @@ REG_OP(BNTrainingUpdateGrad) .OUTPUT(diff_offset, TensorType({DT_FLOAT})) .OP_END_FACTORY_REG(BNTrainingUpdateGrad) +/** +*@brief Performs the backpropagation of BatchNorm . \n + +*@par Inputs: +* Four inputs, including: +*@li grads: A 6D Tensor of type float16 or float32, with format NDC1HWC0, +* for the gradient. +*@li x: A 6D Tensor of type float16 or float32, with format NDC1HWC0. +*@li batch_mean: A 6D Tensor of type float32, with format NDC1HWC0, +* for the mean of "x". +*@li batch_variance: A 6D Tensor of type float32, with format NDC1HWC0, +* for the variance of "x" . \n + +*@par Attributes: +*epsilon: An optional float32. Defaults to "0.0001". A small float number +* added to the variance of "x" . \n + +*@par Outputs: +*@li diff_scale: A Tensor of type float32, with format NDC1HWC0, +* for the offset of "scale". +*@li diff_offset: A Tensor of type float32, with format NDC1HWC0, +* for the offset of "offset" . \n + +*/ +REG_OP(BN3DTrainingUpdateGrad) + .INPUT(grads, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(batch_mean, TensorType({DT_FLOAT})) + .INPUT(batch_variance, TensorType({DT_FLOAT})) + .ATTR(epsilon, Float, 0.0001) + .OUTPUT(diff_scale, TensorType({DT_FLOAT})) + .OUTPUT(diff_offset, TensorType({DT_FLOAT})) + .OP_END_FACTORY_REG(BN3DTrainingUpdateGrad) + /** *@brief Performs the backpropagation of BatchNorm for inference . \n diff --git a/third_party/fwkacllib/inc/ops/selection_ops.h b/third_party/fwkacllib/inc/ops/selection_ops.h index dee9e0f7..33980d43 100644 --- a/third_party/fwkacllib/inc/ops/selection_ops.h +++ b/third_party/fwkacllib/inc/ops/selection_ops.h @@ -857,8 +857,8 @@ REG_OP(SliceDV2) * @li sorted = true * @li It's unstable sorted indices on the platform of Ascend310 -* @par Third-party framework compatibility -* @li Compatible with the TensorFlow operator TopK. +* @par Restrictions: +* Warning: THIS FUNCTION IS DEPRECATED. Please use TopKV2 instead. */ REG_OP(TopKD) .INPUT(x, TensorType::RealNumberType()) @@ -883,6 +883,44 @@ REG_OP(TopKD) * Number of top elements to look for along the last dimension (along each row * for matrices) . \n +* @par Attributes: +* @li sorted: An optional bool. Defaults to true. +* If true, the resulting "k" elements will be sorted by the values in descending +* order. +* @li dim: An optional int. Defaults to -1. For reserved use. +* @li largest: An optional bool. Defaults to true. For reserved use. \n + +* @par Outputs: +* @li values: A Tensor, specifying the sorted data. Has the same type as +* "input". +* @li indices: A Tensor of type int32, specifying the indices of sorted data . \n + +* @see TopK() +* @par Third-party framework compatibility +* @li Compatible with the TensorFlow operator TopKV2. +*/ +REG_OP(TopKV2) + .INPUT(x, TensorType::RealNumberType()) + .INPUT(k, TensorType({DT_INT32})) + .OUTPUT(values, TensorType::RealNumberType()) + .OUTPUT(indices, TensorType({DT_INT32})) + .ATTR(sorted, Bool, true) + .ATTR(dim, Int, -1) + .ATTR(largest, Bool, true) + .OP_END_FACTORY_REG(TopKV2) + +/** +* @brief Finds values and indices of the "k" largest elements for the last +* dimension . \n + +* @par Inputs: +* Two inputs, including: +* @li x: A 1D or higher tensor of type BasicType, with the last dimension +* at least "k". +* @li k: A 0D Tensor of type int32. +* Number of top elements to look for along the last dimension (along each row +* for matrices) . \n + * @par Attributes: * @li sorted: An optional bool. Defaults to true. * If true, the resulting "k" elements will be sorted by the values in descending @@ -2103,6 +2141,34 @@ REG_OP(StridedSliceV2) .OUTPUT(y, TensorType::BasicType()) .OP_END_FACTORY_REG(StridedSliceV2) +/** +*@brief Fills the elements of the input tensor with value val by selecting the indices in the order given in index. \n + +*@par Inputs: +*Three inputs, including: +* @li x: A tensor. Must be one of the following types: +* float16, float32, int32. \n +*@li assist1: A tensor. Must be one of the following types: +* float16, float32, int32. \n +*@li assist2: A tensor. Must be one of the following types: +* float16, float32, int32. \n + +* @par Attributes: +* @li dim: A required int. Used to select the dimension of this tensor. \n + +*@par Outputs: +*y: A Tensor with the same type and shape of input_x's. \n + +*@par Third-party framework compatibility +*Compatible with the Pytorch operator IndexFill. \n +*/ +REG_OP(IndexFillD) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .INPUT(assist1, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .INPUT(assist2, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_INT32})) + .REQUIRED_ATTR(dim, Int) + .OP_END_FACTORY_REG(IndexFillD) } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_SELECTION_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/sparse_ops.h b/third_party/fwkacllib/inc/ops/sparse_ops.h index 09d8ced9..a1fc9ee6 100644 --- a/third_party/fwkacllib/inc/ops/sparse_ops.h +++ b/third_party/fwkacllib/inc/ops/sparse_ops.h @@ -383,11 +383,11 @@ REG_OP(SparseFillEmptyRowsGrad) REG_OP(SparseTensorDenseMatMul) .INPUT(x1_indices, TensorType({DT_INT32, DT_INT64})) .INPUT(x1_values, TensorType({DT_FLOAT, DT_DOUBLE, DT_INT32, \ - DT_COMPLEXT64, DT_COMPLEX128, DT_FLOAT16})) + DT_COMPLEXT64, DT_COMPLEX128, DT_FLOAT16, DT_INT64})) .INPUT(x1_shape, TensorType({DT_INT64})) - .INPUT(x2, TensorType({DT_FLOAT, DT_DOUBLE, DT_INT32, DT_COMPLEXT64, \ + .INPUT(x2, TensorType({DT_FLOAT, DT_DOUBLE, DT_INT64, DT_INT32, DT_COMPLEXT64, \ DT_COMPLEX128, DT_FLOAT16})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_INT32, DT_COMPLEXT64, \ + .OUTPUT(y, TensorType({DT_FLOAT, DT_DOUBLE, DT_INT64, DT_INT32, DT_COMPLEXT64, \ DT_COMPLEX128, DT_FLOAT16})) .ATTR(adjoint_a, Bool, false) .ATTR(adjoint_b, Bool, false) diff --git a/third_party/fwkacllib/inc/ops/spectral_ops.h b/third_party/fwkacllib/inc/ops/spectral_ops.h index be3d7d00..82accc73 100644 --- a/third_party/fwkacllib/inc/ops/spectral_ops.h +++ b/third_party/fwkacllib/inc/ops/spectral_ops.h @@ -26,6 +26,24 @@ namespace ge { +/** +*@brief Computes the inverse 1-dimensional discrete Fourier transform over the +inner-most dimension of `x`. \n + +*@par Inputs: +*@li x: A Tensor. Must be the following types: complex64, complex128. \n + +*@par Outputs: +*@li y: A complex tensor of the same rank as `x`. \n + +*@par Third-party framework compatibility +* Compatible with TensorFlow IFFT operator. +*/ +REG_OP(IFFT) + .INPUT(x, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OP_END_FACTORY_REG(IFFT) + /** *@brief Real-valued fast Fourier transform . \n @@ -47,6 +65,84 @@ REG_OP(RFFT) .OUTPUT(y, TensorType({DT_COMPLEX64})) .OP_END_FACTORY_REG(RFFT) +/** +*@brief Inverse real-valued fast Fourier transform . \n + +*@par Inputs: +*@li x: A complex64 tensor. +*@li fft_length: An int32 tensor of shape [1]. The FFT length . \n + +*@par Outputs: +*@li y: A float32 tensor of the same rank as `input`. The inner-most + dimension of `input` is replaced with the `fft_length` samples of its inverse + 1D Fourier transform . \n + +*@par Third-party framework compatibility +* Compatible with TensorFlow IRFFT operator. +*/ +REG_OP(IRFFT) + .INPUT(x, TensorType({DT_COMPLEX64})) + .INPUT(fft_length, TensorType({DT_INT32})) + .OUTPUT(y, TensorType({DT_FLOAT})) + .OP_END_FACTORY_REG(IRFFT) + + +/** +*@brief 2D fast Fourier transform. \n + +*@par Inputs: +*@li x: A complex64 tensor.. + +*@par Outputs: +*@li y: A complex64 tensor of the same shape as `input`. The inner-most 2 + dimensions of `input` are replaced with their 2D Fourier transform.\n + +*@par Third-party framework compatibility +* Compatible with TensorFlow FFT2D operator. +*/ +REG_OP(FFT2D) + .INPUT(x, TensorType({DT_COMPLEX64, DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_COMPLEX64, DT_COMPLEX128})) + .OP_END_FACTORY_REG(FFT2D) + +/** +*@brief Calculate the one-dimensional discrete Fourier transform on the +innermost dimension of the input. \n + +*@par Inputs: +*@li x: A Tensor. Must be the following types: complex64, complex128. \n + +*@par Outputs: +*@li y: A complex tensor with the same shape as input. The innermost dimension +of the input is replaced by its 1-dimensional Fourier transform. \n + +*@par Third-party framework compatibility +* Compatible with TensorFlow FFT operator. +*/ +REG_OP(FFT) + .INPUT(x, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OP_END_FACTORY_REG(FFT) + +/** +*@brief Calculate the inverse 1-dimensional discrete Fourier transform on the +innermost dimension of the input. \n + +*@par Inputs: +*@li x: A Tensor. Must be the following types: complex64, complex128. \n + +*@par Outputs: +*@li y: A complex tensor with the same shape as input. The innermost dimension +of the input is replaced by its inverse two-dimensional Fourier transform. \n + +*@par Third-party framework compatibility +* Compatible with TensorFlow IFFT2D operator. +*/ +REG_OP(IFFT2D) + .INPUT(x, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OUTPUT(y, TensorType({DT_COMPLEX64,DT_COMPLEX128})) + .OP_END_FACTORY_REG(IFFT2D) + } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_SPECTRAL_OPS_H_ \ No newline at end of file diff --git a/third_party/fwkacllib/inc/ops/split_combination_ops.h b/third_party/fwkacllib/inc/ops/split_combination_ops.h index f1a93fa6..af2c37bc 100644 --- a/third_party/fwkacllib/inc/ops/split_combination_ops.h +++ b/third_party/fwkacllib/inc/ops/split_combination_ops.h @@ -62,8 +62,8 @@ REG_OP(Split) *Must be one of the following types: float16, float32, int32, int8, int16, int64, uint8, uint16, uint32, uint64 *@par Attributes: -*@li split_dim: A required int8, int16, int32, or int64. Specifies the dimension along which to split. No default value. -*@li num_split: A required int8, int16, int32, or int64. Specifies the number of output tensors. No default value . \n +*@li split_dim: A required int32. Specifies the dimension along which to split. No default value. +*@li num_split: A required int32. Specifies the number of output tensors. No default value . \n *@par Outputs: *y:Dynamic output. A list of output tensors. Has the same type and format as "x" . \n @@ -94,12 +94,12 @@ REG_OP(SplitD) *@par Inputs: * Three inputs, including: *@li x: An ND Tensor. -*Must be one of the following types: -*@li size_splits: A list of int8, int16, int32, or int64. Specifies a list containing the sizes of each output tensor along the split dimension. -*@li split_dim: An int8, int16, int32, or int64. Specifies the dimension along which to split . \n +*Must be one of the types:float16, float32, double, int64, int32, uint8, uint16, uint32, uint64, int8, int16, complex64, complex128, qint8, quint8, qint16, quint16, qint32. +*@li size_splits: Must be one of the types:int32, int64. Specifies a list containing the sizes of each output tensor along the split dimension. +*@li split_dim: Must be the following type:int32. Specifies the dimension along which to split . \n *@par Attributes: -*num_split: A required int8, int16, int32, or int64. Specifies the number of output tensors. No default value . \n +*num_split: A required int32. Specifies the number of output tensors. No default value . \n *@par Outputs: *y: Dynamic output.A list of output tensors. Has the same type and format as "x" . \n @@ -129,9 +129,9 @@ REG_OP(SplitV) *Must be one of the following types: float16, float32, int32, int8, int16, int64, uint8, uint16, uint32, uint64 *@par Attributes: -*@li size_splits: A required list of int8, int16, int32, or int64. Specifies a list containing the sizes of each output tensor along the split dimension. -*@li split_dim: A required int8, int16, int32, or int64. Specifies the dimension along which to split. No default value. -*@li num_split: A required int8, int16, int32, or int64. Specifies the number of output tensors. No default value . \n +*@li size_splits: A required list of int32. Specifies a list containing the sizes of each output tensor along the split dimension. +*@li split_dim: A required int32. Specifies the dimension along which to split. No default value. +*@li num_split: A required int32. Specifies the number of output tensors. No default value . \n *@par Outputs: *y: Dynamic output.A list of output tensors. Has the same type and format as "x" . \n diff --git a/third_party/fwkacllib/inc/ops/string_ops.h b/third_party/fwkacllib/inc/ops/string_ops.h index ec84cc83..29aec302 100644 --- a/third_party/fwkacllib/inc/ops/string_ops.h +++ b/third_party/fwkacllib/inc/ops/string_ops.h @@ -488,7 +488,7 @@ include: */ REG_OP(AsString) .INPUT(x, TensorType({DT_INT8, DT_INT16, DT_INT32, DT_INT64, DT_FLOAT, \ - DT_DOUBLE, DT_BOOL})) + DT_DOUBLE, DT_BOOL, DT_COMPLEX64, DT_COMPLEX128})) .OUTPUT(y, TensorType({DT_STRING})) .ATTR(precision, Int, -1) .ATTR(scientific, Bool, false) diff --git a/third_party/fwkacllib/inc/runtime/config.h b/third_party/fwkacllib/inc/runtime/config.h index ee104693..fed7341a 100644 --- a/third_party/fwkacllib/inc/runtime/config.h +++ b/third_party/fwkacllib/inc/runtime/config.h @@ -46,6 +46,12 @@ typedef enum tagRtChipType { CHIP_END, } rtChipType_t; +typedef enum tagRtAicpuScheType { + SCHEDULE_SOFTWARE = 0, /* Software Schedule */ + SCHEDULE_SOFTWARE_OPT, + SCHEDULE_HARDWARE, /* HWTS Schedule */ +} rtAicpuScheType; + typedef enum tagRtVersion { VER_BEGIN = 0, VER_NA = VER_BEGIN, @@ -184,6 +190,19 @@ RTS_API rtError_t rtMemGetL2Info(rtStream_t stream, void **ptr, uint32_t *size); */ RTS_API rtError_t rtGetRuntimeVersion(uint32_t *runtimeVersion); + +/** + * @ingroup + * @brief get device feature ability by device id, such as task schedule ability. + * @param [in] deviceId + * @param [in] moduleType + * @param [in] featureType + * @param [out] value + * @return RT_ERROR_NONE for ok + * @return RT_ERROR_INVALID_VALUE for error input + */ +RTS_API rtError_t rtGetDeviceCapability(int32_t deviceId, int32_t moduleType, int32_t featureType, int32_t *value); + #if defined(__cplusplus) && !defined(COMPILE_OMG_PACKAGE) } #endif diff --git a/third_party/fwkacllib/inc/runtime/dev.h b/third_party/fwkacllib/inc/runtime/dev.h index 49f6a3f6..018f4e6c 100644 --- a/third_party/fwkacllib/inc/runtime/dev.h +++ b/third_party/fwkacllib/inc/runtime/dev.h @@ -62,6 +62,11 @@ typedef enum tagRtFeatureType { FEATURE_TYPE_RSV } rtFeatureType_t; +typedef enum tagRtDeviceFeatureType { + FEATURE_TYPE_SCHE, + FEATURE_TYPE_END, +} rtDeviceFeatureType_t; + typedef enum tagMemcpyInfo { MEMCPY_INFO_SUPPORT_ZEROCOPY = 0, MEMCPY_INFO_RSV diff --git a/third_party/fwkacllib/inc/runtime/kernel.h b/third_party/fwkacllib/inc/runtime/kernel.h index dc16ca58..0ec1a163 100644 --- a/third_party/fwkacllib/inc/runtime/kernel.h +++ b/third_party/fwkacllib/inc/runtime/kernel.h @@ -191,6 +191,14 @@ typedef void (*rtCallback_t)(void *fnData); #define RT_FUSION_KERNEL_DUMPFLAG (0x04) #define RT_KERNEL_CUSTOM_AICPU (0x08) +/** + * @ingroup rt_kernel + * @brief kernel mode +**/ +#define RT_DEFAULT_KERNEL_MODE (0x00) +#define RT_NORMAL_KERNEL_MODE (0x01) +#define RT_ALL_KERNEL_MODE (0x02) + /** * @ingroup rt_kernel * @brief kernel L1 Fusion Dump bit flags @@ -207,6 +215,16 @@ typedef void (*rtCallback_t)(void *fnData); */ RTS_API rtError_t rtDevBinaryRegister(const rtDevBinary_t *bin, void **handle); +/** + * @ingroup rt_kernel + * @brief register device binary with all kernel + * @param [in] bin device binary description + * @param [out] handle device binary handle + * @return RT_ERROR_NONE for ok + * @return RT_ERROR_INVALID_VALUE for error input + */ +RTS_API rtError_t rtRegisterAllKernel(const rtDevBinary_t *bin, void **handle); + /** * @ingroup rt_kernel * @brief register fast memeory device binary @@ -314,6 +332,23 @@ RTS_API rtError_t rtKernelConfigDump(uint32_t kind, uint32_t dumpSizePerBlock, u RTS_API rtError_t rtKernelLaunch(const void *stubFunc, uint32_t blockDim, void *args, uint32_t argsSize, rtSmDesc_t *smDesc, rtStream_t stream); +/** + * @ingroup rt_kernel + * @brief launch kernel with handle to device + * @param [in] handle program + * @param [in] devFunc device function description. + * @param [in] blockDim block dimentions + * @param [in] args argments address for kernel function + * @param [in] argsSize argements size + * @param [in] smDesc shared memory description + * @param [in] stream associated stream + * @param [in] kernelInfo kernel info + * @return RT_ERROR_NONE for ok + * @return RT_ERROR_INVALID_VALUE for error input + */ +RTS_API rtError_t rtKernelLaunchWithHandle(void *handle, const void *devFunc, uint32_t blockDim, void *args, uint32_t argsSize, + rtSmDesc_t *smDesc, rtStream_t stream_, const void *kernelInfo); + /** * @ingroup rt_kernel * @brief launch kernel to device diff --git a/third_party/fwkacllib/inc/runtime/rt_model.h b/third_party/fwkacllib/inc/runtime/rt_model.h index 482486a8..e6d849c8 100644 --- a/third_party/fwkacllib/inc/runtime/rt_model.h +++ b/third_party/fwkacllib/inc/runtime/rt_model.h @@ -50,6 +50,7 @@ typedef enum tagModelTaskType { RT_MODEL_TASK_STREAM_LABEL_SWITCH_BY_INDEX, RT_MODEL_TASK_STREAM_LABEL_GOTO, RT_MODEL_TASK_MODEL_EXIT, + RT_MODEL_TASK_ALL_KERNEL, } rtModelTaskType_t; typedef enum tagModelStreamType { @@ -127,6 +128,18 @@ typedef struct tagKernelTaskInfo { uint16_t *argsOffset; } rtKernelTaskInfo_t; +typedef struct tagAllKernelTaskInfo { + uint16_t blockDim; + uint16_t argsCount; + uint16_t argsSize; + uint16_t reserved; + void *devfunc; + void *handle; + uint8_t *smDesc; + uint8_t *args; + uint16_t *argsOffset; +} rtAllKernelTaskInfo_t; + typedef struct tagKernelTaskInfoEx { uint32_t flags; uint32_t argsSize; @@ -251,6 +264,7 @@ typedef struct tagTaskInfo { union { rtKernelTaskInfoEx_t kernelTaskEx; rtKernelTaskInfo_t kernelTask; + rtAllKernelTaskInfo_t allKernelTask; rtEventTaskInfo_t eventTask; rtStreamSwitchTaskInfo_t streamSwitchTask; rtStreamActiveTaskInfo_t streamActiveTask; diff --git a/third_party/fwkacllib/inc/toolchain/slog.h b/third_party/fwkacllib/inc/toolchain/slog.h index ba286d02..7c4f7be2 100644 --- a/third_party/fwkacllib/inc/toolchain/slog.h +++ b/third_party/fwkacllib/inc/toolchain/slog.h @@ -120,15 +120,15 @@ typedef struct tagKV { } KeyValue; typedef enum { - APPLICATION = 0, - SYSTEM + APPLICATION = 0, + SYSTEM } ProcessType; typedef struct { - ProcessType type; - unsigned int pid; - unsigned int deviceId; - char reserved[RESERVERD_LENGTH]; + ProcessType type; + unsigned int pid; + unsigned int deviceId; + char reserved[RESERVERD_LENGTH]; } LogAttr; /** @@ -381,13 +381,13 @@ DLL_EXPORT void DlogFlush(void); * @ingroup slog * @brief Internal log interface, other modules are not allowed to call this interface */ -void DlogErrorInner(int moduleId, const char *fmt, ...) __attribute__((format(printf, 2, 3))); -void DlogWarnInner(int moduleId, const char *fmt, ...) __attribute__((format(printf, 2, 3))); -void DlogInfoInner(int moduleId, const char *fmt, ...) __attribute__((format(printf, 2, 3))); -void DlogDebugInner(int moduleId, const char *fmt, ...) __attribute__((format(printf, 2, 3))); -void DlogEventInner(int moduleId, const char *fmt, ...) __attribute__((format(printf, 2, 3))); -void DlogInner(int moduleId, int level, const char *fmt, ...) __attribute__((format(printf, 3, 4))); -void DlogWithKVInner(int moduleId, int level, KeyValue *pstKVArray, int kvNum, const char *fmt, ...) __attribute__((format(printf, 5, 6))); +void DlogErrorInner(int moduleId, const char *fmt, ...); +void DlogWarnInner(int moduleId, const char *fmt, ...); +void DlogInfoInner(int moduleId, const char *fmt, ...); +void DlogDebugInner(int moduleId, const char *fmt, ...); +void DlogEventInner(int moduleId, const char *fmt, ...); +void DlogInner(int moduleId, int level, const char *fmt, ...); +void DlogWithKVInner(int moduleId, int level, KeyValue *pstKVArray, int kvNum, const char *fmt, ...); #ifdef __cplusplus #ifndef LOG_CPP @@ -500,8 +500,8 @@ DLL_EXPORT void DlogFlushForC(void); * @ingroup slog * @brief Internal log interface, other modules are not allowed to call this interface */ -void DlogInnerForC(int moduleId, int level, const char *fmt, ...) __attribute__((format(printf, 3, 4))); -void DlogWithKVInnerForC(int moduleId, int level, KeyValue *pstKVArray, int kvNum, const char *fmt, ...) __attribute__((format(printf, 5, 6))); +void DlogInnerForC(int moduleId, int level, const char *fmt, ...); +void DlogWithKVInnerForC(int moduleId, int level, KeyValue *pstKVArray, int kvNum, const char *fmt, ...); #ifdef __cplusplus } diff --git a/third_party/fwkacllib/inc/toolchain/tuning_tool/tune_api.h b/third_party/fwkacllib/inc/toolchain/tuning_tool/tune_api.h index 6208f462..e436dafd 100644 --- a/third_party/fwkacllib/inc/toolchain/tuning_tool/tune_api.h +++ b/third_party/fwkacllib/inc/toolchain/tuning_tool/tune_api.h @@ -1,72 +1,137 @@ -/** - * @file tune_api.h - * - * Copyright (c) Huawei Technologies Co., Ltd. 2020-2020. All rights reserved.\n - * - * This program is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n - * 描述：mstune调优接口头文件 - */ -/** @defgroup mstune mstune调优接口 */ -#ifndef TUNE_API_H -#define TUNE_API_H -#include -#include -#include -#include "graph/graph.h" -#include "ge/ge_api.h" - -/** - * @ingroup mstune - * - * mstune status - */ -enum MsTuneStatus { - MSTUNE_SUCCESS, /** tune success */ - MSTUNE_FAILED, /** tune failed */ -}; - -// Option key: for train options sets -const std::string MSTUNE_SELF_KEY = "mstune"; -const std::string MSTUNE_GEINIT_KEY = "initialize"; -const std::string MSTUNE_GESESS_KEY = "session"; - -/** - * @ingroup mstune - * @par 描述: 命令行调优 - * - * @attention 无 - * @param option [IN] 调优参数 - * @param msg [OUT] 调优异常下返回信息 - * @retval #MSTUNE_SUCCESS 执行成功 - * @retval #MSTUNE_FAILED 执行失败 - * @par 依赖: - * @li tune_api.cpp：该接口所属的开发包。 - * @li tune_api.h：该接口声明所在的头文件。 - * @see 无 - * @since - */ -MsTuneStatus MsTuning(const std::map &option, std::string &msg); - -/** - * @ingroup mstune - * @par 描述: 梯度调优 - * - * @attention 无 - * @param tuningGraph [IN] 调优图 - * @param dependGraph [IN] 调优依赖图 - * @param session [IN] ge连接会话 - * @param option [IN] 参数集. 包含调优参数及ge参数 - * @retval #MSTUNE_SUCCESS 执行成功 - * @retval #MSTUNE_FAILED 执行失败 - * @par 依赖: - * @li tune_api.cpp：该接口所属的开发包。 - * @li tune_api.h：该接口声明所在的头文件。 - * @see 无 - * @since - */ -extern "C" MsTuneStatus MsTrainTuning(ge::Graph &tuningGraph, std::vector &dependGraph, - ge::Session *session, const std::map> &option); - -#endif +/** + * @file tune_api.h + * + * Copyright (c) Huawei Technologies Co., Ltd. 2020-2020. All rights reserved.\n + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n + * 描述：mstune调优接口头文件 + */ +/** @defgroup mstune mstune调优接口 */ +#ifndef TUNE_API_H +#define TUNE_API_H +#include +#include +#include +#include "graph/graph.h" +#include "ge/ge_api.h" + +/** + * @ingroup mstune + * + * mstune status + */ +enum MsTuneStatus { + MSTUNE_SUCCESS, /** tune success */ + MSTUNE_FAILED, /** tune failed */ +}; + +// Option key: for train options sets +const std::string MSTUNE_SELF_KEY = "mstune"; +const std::string MSTUNE_GEINIT_KEY = "initialize"; +const std::string MSTUNE_GESESS_KEY = "session"; + +#ifdef __cplusplus +extern "C" { +#endif + +struct RunnerInitConfig { + // onilne online + std::string profPath; + std::string parserPath; + // ncs only + std::vector devList; +}; + +struct RunnerOpInfo { + std::string opName; + uint64_t opCostTime; + uint64_t aicoreCostTime; + // gradient_split only + std::string modelName; + std::string opType; + std::vector start; + std::vector end; +}; + +struct RunnerModelInfo { + uint64_t totalCostTime; +}; + +struct RunnerRunResult { + std::vector modelInfo; + std::vector opInfo; +}; + +struct RunnerResult { + uint64_t totalCostTime; + std::map opCostTime; + std::map aicoreCostTime; +}; + +struct RunnerDataBuf { + void *ptr = nullptr; + size_t size = 0; +}; + +struct AOEBufferData { + std::shared_ptr data = nullptr; + uint64_t length; +}; + +struct RunnerConfig { + bool isProf; + uint32_t loop; + // offline only + std::vector input; + std::vector output; + std::string modelPath; + RunnerDataBuf modelData; + // online only + uint32_t devId; + std::vector> inputs; + std::vector dependGraph; // run graph (for training) +}; +#ifdef __cplusplus +} +#endif + +/** + * @ingroup mstune + * @par 描述: 命令行调优 + * + * @attention 无 + * @param option [IN] 调优参数 + * @param msg [OUT] 调优异常下返回信息 + * @retval #MSTUNE_SUCCESS 执行成功 + * @retval #MSTUNE_FAILED 执行失败 + * @par 依赖: + * @li tune_api.cpp：该接口所属的开发包。 + * @li tune_api.h：该接口声明所在的头文件。 + * @see 无 + * @since + */ +MsTuneStatus MsTuning(const std::map &option, std::string &msg); + +/** + * @ingroup mstune + * @par 描述: 梯度调优 + * + * @attention 无 + * @param tuningGraph [IN] 调优图 + * @param dependGraph [IN] 调优依赖图 + * @param session [IN] ge连接会话 + * @param option [IN] 参数集. 包含调优参数及ge参数 + * @retval #MSTUNE_SUCCESS 执行成功 + * @retval #MSTUNE_FAILED 执行失败 + * @par 依赖: + * @li tune_api.cpp：该接口所属的开发包。 + * @li tune_api.h：该接口声明所在的头文件。 + * @see 无 + * @since + */ +extern "C" MsTuneStatus MsTrainTuning(ge::Graph &tuningGraph, std::vector &dependGraph, + ge::Session *session, const std::map> &option); + +#endif