XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <stdbool.h> |
| 12 | #include <stddef.h> |
| 13 | #include <stdint.h> |
| 14 | |
| 15 | #include <pthreadpool.h> |
| 16 | |
| 17 | #ifdef __cplusplus |
| 18 | extern "C" { |
| 19 | #endif |
| 20 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 21 | /// The number of bytes XNNPACK may read beyond array bounds. |
| 22 | /// The caller must allocate at this this many extra bytes after the tensor data passed to XNNPACK. |
| 23 | /// |
| 24 | /// Note: XNNPACK reads, but never writes beyond array bounds. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 25 | #define XNN_EXTRA_BYTES 16 |
| 26 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 27 | /// Maximum number of dimensions in tensor shape. |
Marat Dukhan | fc2b96e | 2019-12-03 12:04:04 -0800 | [diff] [blame] | 28 | #define XNN_MAX_TENSOR_DIMS 6 |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 29 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 30 | /// The convolution operator represents a depthwise convolution, and use HWGo layout for filters. |
Marat Dukhan | dd69f0b | 2019-10-04 19:40:03 -0700 | [diff] [blame] | 31 | #define XNN_FLAG_DEPTHWISE_CONVOLUTION 0x00000001 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 32 | |
Marat Dukhan | c4f0ff9 | 2019-12-03 14:59:08 -0800 | [diff] [blame] | 33 | /// Assume transposed weights in a fully connected operator. |
| 34 | #define XNN_FLAG_TRANSPOSE_WEIGHTS 0x00000001 |
| 35 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 36 | /// The operator assumes NHWC layout for the input, regardless of the output layout. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 37 | #define XNN_FLAG_INPUT_NHWC 0x00000002 |
| 38 | |
Marat Dukhan | 8440fde | 2019-10-24 12:46:13 -0700 | [diff] [blame] | 39 | /// Match "SAME" padding in TensorFlow. Exact padding values are computed dynamically depending on input size. |
| 40 | #define XNN_FLAG_TENSORFLOW_SAME_PADDING 0x00000004 |
| 41 | |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 42 | /// Match behaviour of TensorFlow 1.x. |
| 43 | #define XNN_FLAG_TENSORFLOW_LEGACY_MODE 0x00000004 |
| 44 | |
| 45 | /// Align corners of input and output images in resize operations. |
| 46 | #define XNN_FLAG_ALIGN_CORNERS 0x00000008 |
| 47 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 48 | /// Status code for any XNNPACK function call. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 49 | enum xnn_status { |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 50 | /// The call succeeded, and all output arguments now contain valid data. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 51 | xnn_status_success = 0, |
| 52 | xnn_status_uninitialized = 1, |
| 53 | xnn_status_invalid_parameter = 2, |
| 54 | xnn_status_invalid_state = 3, |
| 55 | xnn_status_unsupported_parameter = 4, |
| 56 | xnn_status_unsupported_hardware = 5, |
| 57 | xnn_status_out_of_memory = 6, |
| 58 | }; |
| 59 | |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 60 | struct xnn_allocator { |
| 61 | /// User-specified pointer that will be passed as-is to all functions in this structure. |
| 62 | void* context; |
| 63 | /// Pointer to a function to be called for general memory allocation. |
| 64 | /// |
| 65 | /// @param context - The user-specified pointer from xnn_allocator structure. |
| 66 | /// @param size - The size of the memory block to allocate, in bytes. |
| 67 | /// |
| 68 | /// @returns Pointer to the allocated memory block of at least @ref size bytes. |
| 69 | /// If allocation fails, the function must return NULL. |
| 70 | void* (*allocate)(void* context, size_t size); |
| 71 | /// Pointer to a function to be called for general memory re-allocation, i.e. to increase or shrink a previously |
| 72 | /// allocated memory block. The content of the old memory block is copied to the new memory block. |
| 73 | /// |
| 74 | /// @param context - The user-specified pointer from xnn_allocator structure. |
| 75 | /// @param pointer - Pointer to a memory block allocated by @ref allocate or @ref reallocate functions. Can be NULL. |
| 76 | /// If the pointer is NULL, the @ref reallocate call is equivalent to an @ref allocate call. |
| 77 | /// @param size - The new size of the memory block to allocate, in bytes. |
| 78 | /// |
| 79 | /// @returns Pointer to the newly allocated memory block of at least @ref size bytes with the content of the previous |
| 80 | /// memory block. |
| 81 | /// If allocation fails, the function must return NULL, but must not release the previous memory block. |
| 82 | void* (*reallocate)(void* context, void* pointer, size_t size); |
| 83 | /// Pointer to a function to be called for general memory de-allocation. |
| 84 | /// |
| 85 | /// @param context - The user-specified pointer from xnn_allocator structure. |
| 86 | /// @param pointer - Pointer to a memory block allocated by @ref allocate or @ref reallocate functions. Can be NULL. |
| 87 | /// If the pointer is NULL, the @ref deallocate call is a no-op. |
| 88 | void (*deallocate)(void* context, void* pointer); |
| 89 | /// Pointer to a function to be called for aligned memory allocation. |
| 90 | /// |
| 91 | /// @param context - The user-specified pointer from xnn_allocator structure. |
| 92 | /// @param alignment - The alignment of the memory block to allocate, in bytes. Alignment is always a power-of-2. |
| 93 | /// @param size - The size of the memory block to allocate, in bytes. |
| 94 | /// |
| 95 | /// @returns Pointer to the allocated memory block of at least @ref size bytes. |
| 96 | /// If allocation fails, the function must return NULL. |
| 97 | void* (*aligned_allocate)(void* context, size_t alignment, size_t size); |
| 98 | /// Pointer to a function to be called for aligned memory de-allocation. |
| 99 | /// |
| 100 | /// @param context - The user-specified pointer from xnn_allocator structure. |
| 101 | /// @param pointer - Pointer to a memory block allocated by @ref aligned_allocate function. Can be NULL. |
| 102 | /// If the pointer is NULL, the @ref aligned_deallocate call is a no-op. |
| 103 | void (*aligned_deallocate)(void* context, void* pointer); |
| 104 | }; |
| 105 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 106 | /// Initialize XNNPACK library. |
| 107 | /// |
| 108 | /// XNNPACK must be successfully initialized before use. |
| 109 | /// During initialization, XNNPACK populates internal structures depending on host processor. It can be time-consuming. |
| 110 | /// |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 111 | /// @param[in] allocator - structure with function pointers to be use for memory allocation and de-allocation. |
| 112 | /// If this argument is NULL, system-provided memory management functions (e.g. malloc/free) |
| 113 | /// will be used. |
| 114 | /// |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 115 | /// @retval xnn_status_success - XNNPACK is succesfully initialized and ready to use. |
| 116 | /// @retval xnn_status_out_of_memory - initialization failed due to out-of-memory condition. |
| 117 | /// @retval xnn_status_unsupported_hardware - initialization failed because the host processor does not satisfy the |
| 118 | /// minimum hardware requirements for XNNPACK. E.g. this may happen on x86 |
| 119 | /// processors without SSE2 extension, or on 32-bit ARM processors without |
| 120 | /// the NEON SIMD extension. |
Marat Dukhan | 04f03be | 2019-11-19 12:36:47 -0800 | [diff] [blame] | 121 | enum xnn_status xnn_initialize(const struct xnn_allocator* allocator); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 122 | |
Marat Dukhan | 5609a08 | 2019-10-07 10:56:58 -0700 | [diff] [blame] | 123 | /// Deinitialize XNNPACK library. |
| 124 | /// |
| 125 | /// To avoid memory and resource leaks, users must call xnn_deinitialize once for each successful xnn_initialize call. |
| 126 | /// |
| 127 | /// @retval xnn_status_success - deinitialization call succeeded. |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 128 | enum xnn_status xnn_deinitialize(void); |
| 129 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 130 | /// Subgraph is an abstract representation of a neural network model. |
| 131 | /// Subgraph objects are used to define Values (tensors) and Nodes (operators) comprising the model. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 132 | typedef struct xnn_subgraph* xnn_subgraph_t; |
| 133 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 134 | /// Create a empty Subgraph object. |
| 135 | /// |
| 136 | /// @param external_value_ids - number of Value IDs to reserve for communication with external graph representation. |
| 137 | /// The Subgraph object would avoid creating internal Value IDs in the |
| 138 | /// [0, reserved_value_ids-1] range. |
| 139 | /// @param flags - binary features of the subgraph. No supported flags are currently defined. |
| 140 | /// @param subgraph_out - pointer to the variable that will be initialized with a handle to the Subgraph object upon |
| 141 | /// successful return. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 142 | enum xnn_status xnn_create_subgraph( |
| 143 | uint32_t external_value_ids, |
| 144 | uint32_t flags, |
| 145 | xnn_subgraph_t* subgraph_out); |
| 146 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 147 | /// Destroy a Subgraph object, as well as Values, and Nodes associated with the subgraph. |
| 148 | /// |
| 149 | /// @param subgraph - the Subgraph object to destroy. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 150 | enum xnn_status xnn_delete_subgraph( |
| 151 | xnn_subgraph_t subgraph); |
| 152 | |
| 153 | #define XNN_VALUE_FLAG_EXTERNAL_INPUT 0x00000001 |
| 154 | #define XNN_VALUE_FLAG_EXTERNAL_OUTPUT 0x00000002 |
| 155 | |
| 156 | #define XNN_INVALID_VALUE_ID UINT32_MAX |
| 157 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 158 | /// Type of elements in a Value object. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 159 | enum xnn_datatype { |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 160 | /// Invalid data type. Valid Values never have this datatype. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 161 | xnn_datatype_invalid = 0, |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 162 | /// IEEE754 single-precision floating-point. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 163 | xnn_datatype_fp32 = 1, |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 164 | /// IEEE754 half-precision floating-point. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 165 | xnn_datatype_fp16 = 2, |
| 166 | }; |
| 167 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 168 | /// Define a tensor-type Value and add it to a Subgraph. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 169 | /// |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 170 | /// @param subgraph - a Subgraph object that will own the created Value. |
| 171 | /// @param datatype - type of the tensor elements. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 172 | /// @param num_dims - number of dimensions in the shape. |
| 173 | /// @param dims - pointer to an array of @a num_dims shape dimensions. If num_dims is 0, this pointer can be NULL. |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 174 | /// XNNPACK does not keep any pointers to this array after the function returns. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 175 | /// @param data - pointer to static data used for tensor initialization. If the tensor is not statically initialized, |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 176 | /// this pointer must be is NULL. If non-NULL, the life-time of the static data must exceed the life-time |
| 177 | /// of the Subgraph object, and of any Runtime objects created from the Subgraph. |
| 178 | /// @param external_id - external ID for the Value. The ID must be within the range of reversed Value IDs specified on |
| 179 | /// the Subgraph creation. If the external ID is XNN_INVALID_VALUE_ID, an internal ID will be |
| 180 | /// created for the Value. |
| 181 | /// @param flags - binary features of the Value. Supported values are any combination of XNN_VALUE_FLAG_EXTERNAL_INPUT |
| 182 | /// and XNN_VALUE_FLAG_EXTERNAL_OUTPUT. |
| 183 | /// @param id_out - pointer to the variable that will be initialized with the Value ID upon successful return. If a |
| 184 | /// valid @a external_id was provided, the variable will be initialized with the @a external_id value. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 185 | enum xnn_status xnn_define_tensor_value( |
| 186 | xnn_subgraph_t subgraph, |
| 187 | enum xnn_datatype datatype, |
| 188 | size_t num_dims, |
| 189 | const size_t* dims, |
| 190 | const void* data, |
| 191 | uint32_t external_id, |
| 192 | uint32_t flags, |
| 193 | uint32_t* id_out); |
| 194 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 195 | /// Define a 2D Convolution Node and add it to a Subgraph. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 196 | /// |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 197 | /// @param subgraph - a Subgraph object that will own the created Node. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 198 | /// @param input_padding_top - implicit zero-padding above 2D input data. |
| 199 | /// @param input_padding_right - implicit zero-padding to the right of 2D input data. |
| 200 | /// @param input_padding_bottom - implicit zero-padding below 2D input data. |
| 201 | /// @param input_padding_left - implicit zero-padding to the left of 2D input data. |
| 202 | /// @param kernel_height - kernel (filter) height. |
| 203 | /// @param kernel_width - kernel (filter) width. |
| 204 | /// @param subsampling_height - height of subsampling region for convolution output (convolution height stride). |
| 205 | /// @param subsampling_width - width of subsampling region for convolution output (convolution width stride). |
| 206 | /// @param dilation_height - dilation of kernel elements along the height dimension. |
| 207 | /// @param dilation_width - dilation of kernel elements along the width dimension. |
| 208 | /// @param groups - number of convolution groups. |
| 209 | /// @param group_input_channels - number of input channels per group. |
| 210 | /// @param group_output_channels - number of output channels per group. |
| 211 | /// @param output_min - lower bound for clipping output values. |
| 212 | /// @param output_max - upper bound for clipping output values. |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 213 | /// @param input_id - Value ID for the input tensor. The input tensor must be a 4D tensor defined in the @a subgraph |
| 214 | /// with [N, IH, IW, groups * group_input_channels] dimensions |
| 215 | /// @param filter_id - Value ID for the filter tensor. The filter tensor must ge a 4D tensor defined in the @a subgraph |
| 216 | /// with [groups * group_output_channels, kernel_height, kernel_width, group_input_channels] |
| 217 | /// dimensions. |
| 218 | /// @param bias_id - Value ID for the bias tensor. The bias tensor must be a 1D tensor defined in the @a subgraph with |
| 219 | /// [groups * group_output_channels] dimensions. |
| 220 | /// @param output_id - Value ID for the output tensor. The output tensor must be a 4D tensor defined in the @a subgraph |
| 221 | /// with [N, OH, OW, groups * group_output_channels] dimensions. |
| 222 | /// @param flags - binary features of the 2D Convolution Node. The only currently supported values is |
| 223 | /// XNN_FLAG_TENSORFLOW_SAME_PADDING. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 224 | enum xnn_status xnn_define_convolution_2d( |
| 225 | xnn_subgraph_t subgraph, |
| 226 | uint32_t input_padding_top, |
| 227 | uint32_t input_padding_right, |
| 228 | uint32_t input_padding_bottom, |
| 229 | uint32_t input_padding_left, |
| 230 | uint32_t kernel_height, |
| 231 | uint32_t kernel_width, |
| 232 | uint32_t subsampling_height, |
| 233 | uint32_t subsampling_width, |
| 234 | uint32_t dilation_height, |
| 235 | uint32_t dilation_width, |
| 236 | uint32_t groups, |
| 237 | size_t group_input_channels, |
| 238 | size_t group_output_channels, |
| 239 | float output_min, |
| 240 | float output_max, |
| 241 | uint32_t input_id, |
| 242 | uint32_t filter_id, |
| 243 | uint32_t bias_id, |
| 244 | uint32_t output_id, |
| 245 | uint32_t flags); |
| 246 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 247 | /// Define a 2D Depthwise Convolution Node and add it to a Subgraph. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 248 | /// |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 249 | /// @param subgraph - a Subgraph object that will own the created Node. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 250 | /// @param input_padding_top - implicit zero-padding above 2D input data. |
| 251 | /// @param input_padding_right - implicit zero-padding to the right of 2D input data. |
| 252 | /// @param input_padding_bottom - implicit zero-padding below 2D input data. |
| 253 | /// @param input_padding_left - implicit zero-padding to the left of 2D input data. |
| 254 | /// @param kernel_height - kernel (filter) height. |
| 255 | /// @param kernel_width - kernel (filter) width. |
| 256 | /// @param subsampling_height - height of subsampling region for convolution output (convolution height stride). |
| 257 | /// @param subsampling_width - width of subsampling region for convolution output (convolution width stride). |
| 258 | /// @param dilation_height - dilation of kernel elements along the height dimension. |
| 259 | /// @param dilation_width - dilation of kernel elements along the width dimension. |
| 260 | /// @param depth_multiplier - ratio of output channels to input channels. |
| 261 | /// @param input_channels - number of input channels. |
| 262 | /// @param output_min - lower bound for clipping output values. |
| 263 | /// @param output_max - upper bound for clipping output values. |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 264 | /// @param input_id - Value ID for the input tensor. The input tensor must be a 4D tensor defined in the @a subgraph |
| 265 | /// with [N, IH, IW, input_channels] dimensions |
| 266 | /// @param filter_id - Value ID for the filter tensor. The filter tensor must ge a 4D tensor defined in the @a subgraph |
| 267 | /// with [1, kernel_height, kernel_width, input_channels * depth_multiplier] dimensions. |
| 268 | /// @param bias_id - Value ID for the bias tensor. The bias tensor must be a 1D tensor defined in the @a subgraph with |
| 269 | /// [input_channels * depth_multiplier] dimensions. |
| 270 | /// @param output_id - Value ID for the output tensor. The output tensor must be a 4D tensor defined in the @a subgraph |
| 271 | /// with [N, OH, OW, input_channels * depth_multiplier] dimensions. |
| 272 | /// @param flags - binary features of the 2D Depthwise Convolution Node. The only currently supported values is |
| 273 | /// XNN_FLAG_TENSORFLOW_SAME_PADDING. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 274 | enum xnn_status xnn_define_depthwise_convolution_2d( |
| 275 | xnn_subgraph_t subgraph, |
| 276 | uint32_t input_padding_top, |
| 277 | uint32_t input_padding_right, |
| 278 | uint32_t input_padding_bottom, |
| 279 | uint32_t input_padding_left, |
| 280 | uint32_t kernel_height, |
| 281 | uint32_t kernel_width, |
| 282 | uint32_t subsampling_height, |
| 283 | uint32_t subsampling_width, |
| 284 | uint32_t dilation_height, |
| 285 | uint32_t dilation_width, |
| 286 | uint32_t depth_multiplier, |
| 287 | size_t input_channels, |
| 288 | float output_min, |
| 289 | float output_max, |
| 290 | uint32_t input_id, |
| 291 | uint32_t filter_id, |
| 292 | uint32_t bias_id, |
| 293 | uint32_t output_id, |
| 294 | uint32_t flags); |
| 295 | |
Marat Dukhan | 54dcb46 | 2020-02-10 11:06:12 -0800 | [diff] [blame] | 296 | /// Define a 2-Input Add Node and add it to a Subgraph. |
| 297 | /// |
| 298 | /// The 2-Input Add Node computes elementwise addition of two tensor inputs with numpy broadcasting rules. |
| 299 | /// |
| 300 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 301 | /// @param output_min - lower bound for clipping output values. |
| 302 | /// @param output_max - upper bound for clipping output values. |
| 303 | /// @param input1_id - Value ID for the first input tensor. The input tensor must be an N-dimensional tensor defined in |
| 304 | /// the @a subgraph with each dimension either equal to the corresponding dimension of the second |
| 305 | /// input, or equal to 1. In the latter case, the elements of the input tensor are broadcasted along |
| 306 | /// that dimension. |
| 307 | /// @param input2_id - Value ID for the second input tensor. The input tensor must be an M-dimensional tensor defined in |
| 308 | /// the @a subgraph with each dimension either equal to the corresponding dimension of the first |
| 309 | /// input, or equal to 1. In the latter case, the elements of the input tensor are broadcasted along |
| 310 | /// that dimension. |
| 311 | /// @param output_id - Value ID for the output tensor. The output tensor must be a max(N,M)-dimensional tensor defined |
| 312 | /// in the @a subgraph with each dimension equal to the maximum between the corresponding dimension |
| 313 | /// of the two inputs. |
| 314 | /// @param flags - binary features of the Add Node. No supported flags are currently defined. |
| 315 | enum xnn_status xnn_define_add2( |
| 316 | xnn_subgraph_t subgraph, |
| 317 | float output_min, |
| 318 | float output_max, |
| 319 | uint32_t input1_id, |
| 320 | uint32_t input2_id, |
| 321 | uint32_t output_id, |
| 322 | uint32_t flags); |
| 323 | |
| 324 | /// Define a 2-Input Multiply Node and add it to a Subgraph. |
| 325 | /// |
| 326 | /// The 2-Input Multiply Node computes elementwise multiplication of two tensor inputs with numpy broadcasting rules. |
| 327 | /// |
| 328 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 329 | /// @param output_min - lower bound for clipping output values. |
| 330 | /// @param output_max - upper bound for clipping output values. |
| 331 | /// @param input1_id - Value ID for the first input tensor. The input tensor must be an N-dimensional tensor defined in |
| 332 | /// the @a subgraph with each dimension either equal to the corresponding dimension of the second |
| 333 | /// input, or equal to 1. In the latter case, the elements of the input tensor are broadcasted along |
| 334 | /// that dimension. |
| 335 | /// @param input2_id - Value ID for the second input tensor. The input tensor must be an M-dimensional tensor defined in |
| 336 | /// the @a subgraph with each dimension either equal to the corresponding dimension of the first |
| 337 | /// input, or equal to 1. In the latter case, the elements of the input tensor are broadcasted along |
| 338 | /// that dimension. |
| 339 | /// @param output_id - Value ID for the output tensor. The output tensor must be a max(N,M)-dimensional tensor defined |
| 340 | /// in the @a subgraph with each dimension equal to the maximum between the corresponding dimension |
| 341 | /// of the two inputs. |
| 342 | /// @param flags - binary features of the Multiply Node. No supported flags are currently defined. |
| 343 | enum xnn_status xnn_define_multiply2( |
| 344 | xnn_subgraph_t subgraph, |
| 345 | float output_min, |
| 346 | float output_max, |
| 347 | uint32_t input1_id, |
| 348 | uint32_t input2_id, |
| 349 | uint32_t output_id, |
| 350 | uint32_t flags); |
| 351 | |
Miao Wang | 3fa1f01 | 2020-02-17 22:45:06 +0000 | [diff] [blame] | 352 | /// Define a PReLU (Parametric ReLU) Node and add it to a Subgraph. |
| 353 | /// |
| 354 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 355 | /// @param input_id - Value ID for the input tensor. The input tensor must be a 4D tensor defined in the @a subgraph |
| 356 | /// with [N, H, W, channels] dimensions |
| 357 | /// @param slope_id - Value ID for the bias tensor. The bias tensor must be a 1D tensor defined in the @a subgraph with |
| 358 | /// [channels] dimensions. |
| 359 | /// @param output_id - Value ID for the output tensor. The output tensor must be a 4D tensor defined in the @a subgraph |
| 360 | /// with [N, H, W, channels] dimensions. |
| 361 | /// @param flags - binary features of the PReLU Node. No supported flags are currently defined. |
| 362 | enum xnn_status xnn_define_prelu( |
| 363 | xnn_subgraph_t subgraph, |
| 364 | uint32_t input_id, |
| 365 | uint32_t slope_id, |
| 366 | uint32_t output_id, |
| 367 | uint32_t flags); |
| 368 | |
| 369 | /// Define a Clamp Node and add it to a Subgraph. |
| 370 | /// |
| 371 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 372 | /// @param output_min - lower bound for clipping output values. |
| 373 | /// @param output_max - upper bound for clipping output values. |
| 374 | /// @param input_id - Value ID for the input tensor. The input tensor must be defined in the @a subgraph. |
| 375 | /// @param output_id - Value ID for the output tensor. The output tensor must be defined in the @a subgraph, and its |
| 376 | /// shape must match the shape of the input tensor. |
| 377 | /// @param flags - binary features of the Clamp Node. No supported flags are currently defined. |
| 378 | enum xnn_status xnn_define_clamp( |
| 379 | xnn_subgraph_t subgraph, |
| 380 | float output_min, |
| 381 | float output_max, |
| 382 | uint32_t input_id, |
| 383 | uint32_t output_id, |
| 384 | uint32_t flags); |
| 385 | |
| 386 | /// Define a HardSwish Node and add it to a Subgraph. |
| 387 | /// |
| 388 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 389 | /// @param input_id - Value ID for the input tensor. The input tensor must be defined in the @a subgraph. |
| 390 | /// @param output_id - Value ID for the output tensor. The output tensor must be defined in the @a subgraph, and its |
| 391 | /// shape must match the shape of the input tensor. |
| 392 | /// @param flags - binary features of the HardSwish Node. No supported flags are currently defined. |
| 393 | enum xnn_status xnn_define_hardswish( |
| 394 | xnn_subgraph_t subgraph, |
| 395 | uint32_t input_id, |
| 396 | uint32_t output_id, |
| 397 | uint32_t flags); |
| 398 | |
| 399 | /// Define a Sigmoid Node and add it to a Subgraph. |
| 400 | /// |
| 401 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 402 | /// @param input_id - Value ID for the input tensor. The input tensor must be defined in the @a subgraph. |
| 403 | /// @param output_id - Value ID for the output tensor. The output tensor must be defined in the @a subgraph, and its |
| 404 | /// shape must match the shape of the input tensor. |
| 405 | /// @param flags - binary features of the Sigmoid Node. No supported flags are currently defined. |
| 406 | enum xnn_status xnn_define_sigmoid( |
| 407 | xnn_subgraph_t subgraph, |
| 408 | uint32_t input_id, |
| 409 | uint32_t output_id, |
| 410 | uint32_t flags); |
| 411 | |
| 412 | /// Define a SoftMax Node and add it to a Subgraph. |
| 413 | /// |
| 414 | /// @param subgraph - a Subgraph object that will own the created Node. |
| 415 | /// @param input_id - Value ID for the input tensor. The input tensor must be defined in the @a subgraph, and have at |
| 416 | /// least one dimension. |
| 417 | /// @param output_id - Value ID for the output tensor. The output tensor must be defined in the @a subgraph, and its |
| 418 | /// shape must match the shape of the input tensor. |
| 419 | /// @param flags - binary features of the SoftMax Node. No supported flags are currently defined. |
| 420 | enum xnn_status xnn_define_softmax( |
| 421 | xnn_subgraph_t subgraph, |
| 422 | uint32_t input_id, |
| 423 | uint32_t output_id, |
| 424 | uint32_t flags); |
| 425 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 426 | /// Runtime is a combination of an execution plan for subgraph Nodes and a memory manager for subgraph Values. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 427 | typedef struct xnn_runtime* xnn_runtime_t; |
| 428 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 429 | /// Create a empty Runtime object from a subgraph. |
| 430 | /// |
| 431 | /// @param subgraph - a Subgraph object with all Values and Nodes that would be handled by the runtime. No Values or |
| 432 | /// Nodes can be added to the runtime once it is constructed. |
| 433 | /// @param threadpool - the thread pool to be used for parallelisation of computations in the runtime. If the thread |
| 434 | /// pool is NULL, the computation would run on the caller thread without parallelization. |
| 435 | /// @param flags - binary features of the subgraph. No supported flags are currently defined. |
| 436 | /// @param runtime_out - pointer to the variable that will be initialized with a handle to the Runtime object upon |
| 437 | /// successful return. Once constructed, the Runtime object is independent of the Subgraph object |
| 438 | /// used to create it. |
Marat Dukhan | 022c659 | 2020-02-05 18:07:41 -0800 | [diff] [blame] | 439 | enum xnn_status xnn_create_runtime_v2( |
| 440 | xnn_subgraph_t subgraph, |
| 441 | pthreadpool_t threadpool, |
| 442 | uint32_t flags, |
| 443 | xnn_runtime_t* runtime_out); |
| 444 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 445 | enum xnn_status xnn_create_runtime( |
| 446 | xnn_subgraph_t subgraph, |
| 447 | xnn_runtime_t* runtime_out); |
| 448 | |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 449 | struct xnn_external_value { |
| 450 | uint32_t id; |
| 451 | void* data; |
| 452 | }; |
| 453 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 454 | /// Setup data pointers for external inputs and outputs in a Runtime object. |
| 455 | /// |
| 456 | /// @param runtime - a Runtime object created with @ref xnn_create_runtime or @ref xnn_create_runtime_v2. |
| 457 | /// @param num_external_values - the number of external inputs and outputs specified in this call. This number must |
| 458 | /// match the number of external inputs and outputs in the runtime, i.e. all external |
| 459 | /// inputs and outputs in the runtime must be specified in one call. |
| 460 | /// @param external_values - array with location information for all external inputs and outputs in the runtime. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 461 | enum xnn_status xnn_setup_runtime( |
| 462 | xnn_runtime_t runtime, |
| 463 | size_t num_external_values, |
| 464 | const struct xnn_external_value* external_values); |
| 465 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 466 | /// Execute forward pass for all operators in the runtime. |
| 467 | /// |
| 468 | /// @param runtime - the Runtime object with the execution plan to invoke. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 469 | enum xnn_status xnn_invoke_runtime( |
| 470 | xnn_runtime_t runtime); |
| 471 | |
Marat Dukhan | dd2b588 | 2020-02-06 15:12:26 -0800 | [diff] [blame] | 472 | /// Destroy a Runtime object, as well as operators and memory associated with it. |
| 473 | /// |
| 474 | /// @param runtime - the Runtime object to destroy. |
Marat Dukhan | 1d75a54 | 2020-02-03 12:23:01 -0800 | [diff] [blame] | 475 | enum xnn_status xnn_delete_runtime( |
| 476 | xnn_runtime_t runtime); |
| 477 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 478 | typedef struct xnn_operator* xnn_operator_t; |
| 479 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 480 | enum xnn_status xnn_run_operator( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 481 | xnn_operator_t op, |
| 482 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 483 | |
| 484 | enum xnn_status xnn_delete_operator( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 485 | xnn_operator_t op); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 486 | |
| 487 | #ifndef XNN_NO_F32_OPERATORS |
| 488 | |
| 489 | enum xnn_status xnn_create_add_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 490 | size_t channels, |
| 491 | size_t a_stride, |
| 492 | size_t b_stride, |
| 493 | size_t sum_stride, |
| 494 | float sum_min, |
| 495 | float sum_max, |
| 496 | uint32_t flags, |
| 497 | xnn_operator_t* add_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 498 | |
| 499 | enum xnn_status xnn_setup_add_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 500 | xnn_operator_t add_op, |
| 501 | size_t batch_size, |
| 502 | const float* a, |
| 503 | const float* b, |
| 504 | float* sum, |
| 505 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 506 | |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 507 | enum xnn_status xnn_create_add_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 508 | float output_min, |
| 509 | float output_max, |
| 510 | uint32_t flags, |
| 511 | xnn_operator_t* add_op_out); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 512 | |
| 513 | enum xnn_status xnn_setup_add_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 514 | xnn_operator_t add_op, |
| 515 | size_t num_input1_dims, |
| 516 | const size_t* input1_shape, |
| 517 | size_t num_input2_dims, |
| 518 | const size_t* input2_shape, |
| 519 | const float* input1, |
| 520 | const float* input2, |
| 521 | float* output, |
| 522 | pthreadpool_t threadpool); |
Marat Dukhan | b1a0fc3 | 2019-12-02 19:32:02 -0800 | [diff] [blame] | 523 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 524 | enum xnn_status xnn_create_argmax_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 525 | uint32_t input_padding_top, |
| 526 | uint32_t input_padding_right, |
| 527 | uint32_t input_padding_bottom, |
| 528 | uint32_t input_padding_left, |
| 529 | uint32_t pooling_height, |
| 530 | uint32_t pooling_width, |
| 531 | size_t channels, |
| 532 | size_t input_pixel_stride, |
| 533 | size_t output_pixel_stride, |
| 534 | float output_min, |
| 535 | float output_max, |
| 536 | uint32_t flags, |
| 537 | xnn_operator_t* argmax_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 538 | |
| 539 | enum xnn_status xnn_setup_argmax_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 540 | xnn_operator_t argmax_pooling_op, |
| 541 | size_t batch_size, |
| 542 | size_t input_height, |
| 543 | size_t input_width, |
| 544 | const float* input, |
| 545 | float* output, |
| 546 | uint32_t* index, |
| 547 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 548 | |
| 549 | enum xnn_status xnn_create_average_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 550 | uint32_t input_padding_top, |
| 551 | uint32_t input_padding_right, |
| 552 | uint32_t input_padding_bottom, |
| 553 | uint32_t input_padding_left, |
| 554 | uint32_t pooling_height, |
| 555 | uint32_t pooling_width, |
| 556 | uint32_t stride_height, |
| 557 | uint32_t stride_width, |
| 558 | size_t channels, |
| 559 | size_t input_pixel_stride, |
| 560 | size_t output_pixel_stride, |
| 561 | float output_min, |
| 562 | float output_max, |
| 563 | uint32_t flags, |
| 564 | xnn_operator_t* average_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 565 | |
| 566 | enum xnn_status xnn_setup_average_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 567 | xnn_operator_t average_pooling_op, |
| 568 | size_t batch_size, |
| 569 | size_t input_height, |
| 570 | size_t input_width, |
| 571 | const float* input, |
| 572 | float* output, |
| 573 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 574 | |
| 575 | enum xnn_status xnn_create_clamp_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 576 | size_t channels, |
| 577 | size_t input_stride, |
| 578 | size_t output_stride, |
| 579 | float output_min, |
| 580 | float output_max, |
| 581 | uint32_t flags, |
| 582 | xnn_operator_t* clamp_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 583 | |
| 584 | enum xnn_status xnn_setup_clamp_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 585 | xnn_operator_t clamp_op, |
| 586 | size_t batch_size, |
| 587 | const float* input, |
| 588 | float* output, |
| 589 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 590 | |
| 591 | enum xnn_status xnn_create_convolution2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 592 | uint32_t input_padding_top, |
| 593 | uint32_t input_padding_right, |
| 594 | uint32_t input_padding_bottom, |
| 595 | uint32_t input_padding_left, |
| 596 | uint32_t kernel_height, |
| 597 | uint32_t kernel_width, |
| 598 | uint32_t subsampling_height, |
| 599 | uint32_t subsampling_width, |
| 600 | uint32_t dilation_height, |
| 601 | uint32_t dilation_width, |
| 602 | uint32_t groups, |
| 603 | size_t group_input_channels, |
| 604 | size_t group_output_channels, |
| 605 | size_t input_pixel_stride, |
| 606 | size_t output_pixel_stride, |
| 607 | const float* kernel, |
| 608 | const float* bias, |
| 609 | float output_min, |
| 610 | float output_max, |
| 611 | uint32_t flags, |
| 612 | xnn_operator_t* convolution_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 613 | |
| 614 | enum xnn_status xnn_setup_convolution2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 615 | xnn_operator_t convolution_op, |
| 616 | size_t batch_size, |
| 617 | size_t input_height, |
| 618 | size_t input_width, |
| 619 | const float* input, |
| 620 | float* output, |
| 621 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 622 | |
| 623 | enum xnn_status xnn_create_deconvolution2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 624 | uint32_t output_padding_top, |
| 625 | uint32_t output_padding_right, |
| 626 | uint32_t output_padding_bottom, |
| 627 | uint32_t output_padding_left, |
| 628 | uint32_t kernel_height, |
| 629 | uint32_t kernel_width, |
| 630 | uint32_t stride_height, |
| 631 | uint32_t stride_width, |
| 632 | uint32_t dilation_height, |
| 633 | uint32_t dilation_width, |
| 634 | uint32_t groups, |
| 635 | size_t group_input_channels, |
| 636 | size_t group_output_channels, |
| 637 | size_t input_pixel_stride, |
| 638 | size_t output_pixel_stride, |
| 639 | const float* kernel, |
| 640 | const float* bias, |
| 641 | float output_min, |
| 642 | float output_max, |
| 643 | uint32_t flags, |
| 644 | xnn_operator_t* deconvolution_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 645 | |
| 646 | enum xnn_status xnn_setup_deconvolution2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 647 | xnn_operator_t deconvolution_op, |
| 648 | size_t batch_size, |
| 649 | size_t input_height, |
| 650 | size_t input_width, |
| 651 | uint32_t adjustment_height, |
| 652 | uint32_t adjustment_width, |
| 653 | const float* input, |
| 654 | float* output, |
| 655 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 656 | |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 657 | enum xnn_status xnn_create_divide_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 658 | float output_min, |
| 659 | float output_max, |
| 660 | uint32_t flags, |
| 661 | xnn_operator_t* divide_op_out); |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 662 | |
| 663 | enum xnn_status xnn_setup_divide_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 664 | xnn_operator_t divide_op, |
| 665 | size_t num_input1_dims, |
| 666 | const size_t* input1_shape, |
| 667 | size_t num_input2_dims, |
| 668 | const size_t* input2_shape, |
| 669 | const float* input1, |
| 670 | const float* input2, |
| 671 | float* output, |
| 672 | pthreadpool_t threadpool); |
Marat Dukhan | 6918050 | 2019-12-06 15:00:31 -0800 | [diff] [blame] | 673 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 674 | enum xnn_status xnn_create_fully_connected_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 675 | size_t input_channels, |
| 676 | size_t output_channels, |
| 677 | size_t input_stride, |
| 678 | size_t output_stride, |
| 679 | const float* kernel, |
| 680 | const float* bias, |
| 681 | float output_min, |
| 682 | float output_max, |
| 683 | uint32_t flags, |
| 684 | xnn_operator_t* fully_connected_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 685 | |
| 686 | enum xnn_status xnn_setup_fully_connected_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 687 | xnn_operator_t fully_connected_op, |
| 688 | size_t batch_size, |
| 689 | const float* input, |
| 690 | float* output, |
| 691 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 692 | |
| 693 | enum xnn_status xnn_create_global_average_pooling_nwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 694 | size_t channels, |
| 695 | size_t input_stride, |
| 696 | size_t output_stride, |
| 697 | float output_min, |
| 698 | float output_max, |
| 699 | uint32_t flags, |
| 700 | xnn_operator_t* global_average_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 701 | |
| 702 | enum xnn_status xnn_setup_global_average_pooling_nwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 703 | xnn_operator_t global_average_pooling_op, |
| 704 | size_t batch_size, |
| 705 | size_t width, |
| 706 | const float* input, |
| 707 | float* output, |
| 708 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 709 | |
| 710 | enum xnn_status xnn_create_hardswish_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 711 | size_t channels, |
| 712 | size_t input_stride, |
| 713 | size_t output_stride, |
| 714 | uint32_t flags, |
| 715 | xnn_operator_t* hardswish_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 716 | |
| 717 | enum xnn_status xnn_setup_hardswish_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 718 | xnn_operator_t hardswish_op, |
| 719 | size_t batch_size, |
| 720 | const float* input, |
| 721 | float* output, |
| 722 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 723 | |
| 724 | enum xnn_status xnn_create_max_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 725 | uint32_t input_padding_top, |
| 726 | uint32_t input_padding_right, |
| 727 | uint32_t input_padding_bottom, |
| 728 | uint32_t input_padding_left, |
| 729 | uint32_t pooling_height, |
| 730 | uint32_t pooling_width, |
| 731 | uint32_t stride_height, |
| 732 | uint32_t stride_width, |
| 733 | uint32_t dilation_height, |
| 734 | uint32_t dilation_width, |
| 735 | size_t channels, |
| 736 | size_t input_pixel_stride, |
| 737 | size_t output_pixel_stride, |
| 738 | float output_min, |
| 739 | float output_max, |
| 740 | uint32_t flags, |
| 741 | xnn_operator_t* max_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 742 | |
| 743 | enum xnn_status xnn_setup_max_pooling2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 744 | xnn_operator_t max_pooling_op, |
| 745 | size_t batch_size, |
| 746 | size_t input_height, |
| 747 | size_t input_width, |
| 748 | const float* input, |
| 749 | float* output, |
| 750 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 751 | |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 752 | enum xnn_status xnn_create_maximum_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 753 | uint32_t flags, |
| 754 | xnn_operator_t* maximum_op_out); |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 755 | |
| 756 | enum xnn_status xnn_setup_maximum_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 757 | xnn_operator_t maximum_op, |
| 758 | size_t num_input1_dims, |
| 759 | const size_t* input1_shape, |
| 760 | size_t num_input2_dims, |
| 761 | const size_t* input2_shape, |
| 762 | const float* input1, |
| 763 | const float* input2, |
| 764 | float* output, |
| 765 | pthreadpool_t threadpool); |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 766 | |
| 767 | enum xnn_status xnn_create_minimum_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 768 | uint32_t flags, |
| 769 | xnn_operator_t* minimum_op_out); |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 770 | |
| 771 | enum xnn_status xnn_setup_minimum_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 772 | xnn_operator_t minimum_op, |
| 773 | size_t num_input1_dims, |
| 774 | const size_t* input1_shape, |
| 775 | size_t num_input2_dims, |
| 776 | const size_t* input2_shape, |
| 777 | const float* input1, |
| 778 | const float* input2, |
| 779 | float* output, |
| 780 | pthreadpool_t threadpool); |
Marat Dukhan | 79e7f84 | 2019-12-05 14:35:50 -0800 | [diff] [blame] | 781 | |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 782 | enum xnn_status xnn_create_multiply_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 783 | float output_min, |
| 784 | float output_max, |
| 785 | uint32_t flags, |
| 786 | xnn_operator_t* multiply_op_out); |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 787 | |
| 788 | enum xnn_status xnn_setup_multiply_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 789 | xnn_operator_t multiply_op, |
| 790 | size_t num_input1_dims, |
| 791 | const size_t* input1_shape, |
| 792 | size_t num_input2_dims, |
| 793 | const size_t* input2_shape, |
| 794 | const float* input1, |
| 795 | const float* input2, |
| 796 | float* output, |
| 797 | pthreadpool_t threadpool); |
Marat Dukhan | ca2733c | 2019-11-15 23:21:17 -0800 | [diff] [blame] | 798 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 799 | enum xnn_status xnn_create_prelu_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 800 | size_t channels, |
| 801 | size_t input_stride, |
| 802 | size_t output_stride, |
| 803 | const float* negative_slope, |
| 804 | float output_min, |
| 805 | float output_max, |
| 806 | uint32_t flags, |
| 807 | xnn_operator_t* prelu_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 808 | |
| 809 | enum xnn_status xnn_setup_prelu_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 810 | xnn_operator_t prelu_op, |
| 811 | size_t batch_size, |
| 812 | const float* input, |
| 813 | float* output, |
| 814 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 815 | |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 816 | enum xnn_status xnn_create_resize_bilinear2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 817 | size_t channels, |
| 818 | size_t input_pixel_stride, |
| 819 | size_t output_pixel_stride, |
| 820 | uint32_t flags, |
| 821 | xnn_operator_t* resize_op_out); |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 822 | |
| 823 | enum xnn_status xnn_setup_resize_bilinear2d_nhwc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 824 | xnn_operator_t resize_op, |
| 825 | size_t batch_size, |
| 826 | size_t input_height, |
| 827 | size_t input_width, |
| 828 | size_t output_height, |
| 829 | size_t output_width, |
| 830 | const float* input, |
| 831 | float* output, |
| 832 | pthreadpool_t threadpool); |
Marat Dukhan | 6972249 | 2019-11-11 19:55:50 -0800 | [diff] [blame] | 833 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 834 | enum xnn_status xnn_create_sigmoid_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 835 | size_t channels, |
| 836 | size_t input_stride, |
| 837 | size_t output_stride, |
| 838 | uint32_t flags, |
| 839 | xnn_operator_t* sigmoid_op_out); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 840 | |
| 841 | enum xnn_status xnn_setup_sigmoid_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 842 | xnn_operator_t sigmoid_op, |
| 843 | size_t batch_size, |
| 844 | const float* input, |
| 845 | float* output, |
| 846 | pthreadpool_t threadpool); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 847 | |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 848 | enum xnn_status xnn_create_softmax_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 849 | size_t channels, |
| 850 | size_t input_stride, |
| 851 | size_t output_stride, |
| 852 | uint32_t flags, |
| 853 | xnn_operator_t* softmax_op_out); |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 854 | |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 855 | enum xnn_status xnn_setup_softmax_nc_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 856 | xnn_operator_t softmax_op, |
| 857 | size_t batch_size, |
| 858 | const float* input, |
| 859 | float* output, |
| 860 | pthreadpool_t threadpool); |
Marat Dukhan | 1edc454 | 2020-01-27 12:40:13 -0800 | [diff] [blame] | 861 | |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 862 | enum xnn_status xnn_create_subtract_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 863 | float output_min, |
| 864 | float output_max, |
| 865 | uint32_t flags, |
| 866 | xnn_operator_t* subtract_op_out); |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 867 | |
| 868 | enum xnn_status xnn_setup_subtract_nd_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 869 | xnn_operator_t subtract_op, |
| 870 | size_t num_input1_dims, |
| 871 | const size_t* input1_shape, |
| 872 | size_t num_input2_dims, |
| 873 | const size_t* input2_shape, |
| 874 | const float* input1, |
| 875 | const float* input2, |
| 876 | float* output, |
| 877 | pthreadpool_t threadpool); |
Marat Dukhan | 05f3f6d | 2019-12-03 15:13:53 -0800 | [diff] [blame] | 878 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 879 | #ifndef XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 880 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 881 | enum xnn_status xnn_create_convolution2d_nchw_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 882 | uint32_t input_padding_top, |
| 883 | uint32_t input_padding_right, |
| 884 | uint32_t input_padding_bottom, |
| 885 | uint32_t input_padding_left, |
| 886 | uint32_t kernel_height, |
| 887 | uint32_t kernel_width, |
| 888 | uint32_t subsampling_height, |
| 889 | uint32_t subsampling_width, |
| 890 | uint32_t dilation_height, |
| 891 | uint32_t dilation_width, |
| 892 | uint32_t groups, |
| 893 | size_t group_input_channels, |
| 894 | size_t group_output_channels, |
| 895 | const float* kernel, |
| 896 | const float* bias, |
| 897 | float output_min, |
| 898 | float output_max, |
| 899 | uint32_t flags, |
| 900 | xnn_operator_t* convolution_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 901 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 902 | enum xnn_status xnn_setup_convolution2d_nchw_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 903 | xnn_operator_t convolution_op, |
| 904 | size_t batch_size, |
| 905 | size_t input_batch_stride, |
| 906 | size_t output_batch_stride, |
| 907 | size_t input_height, |
| 908 | size_t input_width, |
| 909 | const float* input, |
| 910 | float* output, |
| 911 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 912 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 913 | enum xnn_status xnn_create_global_average_pooling_ncw_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 914 | size_t channels, |
| 915 | float output_min, |
| 916 | float output_max, |
| 917 | uint32_t flags, |
| 918 | xnn_operator_t* global_average_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 919 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 920 | enum xnn_status xnn_setup_global_average_pooling_ncw_f32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 921 | xnn_operator_t global_average_pooling_op, |
| 922 | size_t batch_size, |
| 923 | size_t width, |
| 924 | const float* input, |
| 925 | float* output, |
| 926 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 927 | |
Marat Dukhan | efc47b8 | 2019-11-18 09:25:38 -0800 | [diff] [blame] | 928 | #endif // XNN_NO_NCHW_OPERATORS |
Marat Dukhan | 8fe54e4 | 2019-10-10 14:12:59 -0700 | [diff] [blame] | 929 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 930 | #endif // XNN_NO_F32_OPERATORS |
| 931 | |
| 932 | #ifndef XNN_NO_X32_OPERATORS |
| 933 | |
| 934 | enum xnn_status xnn_create_channel_pad_nc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 935 | size_t input_channels, |
| 936 | size_t pad_before_channels, |
| 937 | size_t pad_after_channels, |
| 938 | size_t input_stride, |
| 939 | size_t output_stride, |
| 940 | const void* pad_value, |
| 941 | uint32_t flags, |
| 942 | xnn_operator_t* channel_pad_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 943 | |
| 944 | enum xnn_status xnn_setup_channel_pad_nc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 945 | xnn_operator_t channel_pad_op, |
| 946 | size_t batch_size, |
| 947 | const void* input, |
| 948 | void* output, |
| 949 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 950 | |
| 951 | enum xnn_status xnn_create_channel_shuffle_nc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 952 | size_t groups, |
| 953 | size_t group_channels, |
| 954 | size_t input_stride, |
| 955 | size_t output_stride, |
| 956 | uint32_t flags, |
| 957 | xnn_operator_t* channel_shuffle_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 958 | |
| 959 | enum xnn_status xnn_setup_channel_shuffle_nc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 960 | xnn_operator_t channel_shuffle_op, |
| 961 | size_t batch_size, |
| 962 | const void* input, |
| 963 | void* output, |
| 964 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 965 | |
| 966 | enum xnn_status xnn_create_unpooling2d_nhwc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 967 | uint32_t input_padding_top, |
| 968 | uint32_t input_padding_right, |
| 969 | uint32_t input_padding_bottom, |
| 970 | uint32_t input_padding_left, |
| 971 | uint32_t pooling_height, |
| 972 | uint32_t pooling_width, |
| 973 | size_t channels, |
| 974 | size_t input_pixel_stride, |
| 975 | size_t output_pixel_stride, |
| 976 | uint32_t flags, |
| 977 | xnn_operator_t* unpooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 978 | |
| 979 | enum xnn_status xnn_setup_unpooling2d_nhwc_x32( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 980 | xnn_operator_t unpooling_op, |
| 981 | size_t batch_size, |
| 982 | size_t input_height, |
| 983 | size_t input_width, |
| 984 | const void* input, |
| 985 | const uint32_t* index, |
| 986 | void* output, |
| 987 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 988 | |
| 989 | #endif // XNN_NO_X32_OPERATORS |
| 990 | |
| 991 | #ifndef XNN_NO_Q8_OPERATORS |
| 992 | |
| 993 | enum xnn_status xnn_create_add_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 994 | size_t channels, |
| 995 | size_t a_stride, |
| 996 | size_t b_stride, |
| 997 | size_t sum_stride, |
| 998 | uint8_t a_zero_point, |
| 999 | float a_scale, |
| 1000 | uint8_t b_zero_point, |
| 1001 | float b_scale, |
| 1002 | uint8_t sum_zero_point, |
| 1003 | float sum_scale, |
| 1004 | uint8_t sum_min, |
| 1005 | uint8_t sum_max, |
| 1006 | uint32_t flags, |
| 1007 | xnn_operator_t* add_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1008 | |
| 1009 | enum xnn_status xnn_setup_add_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1010 | xnn_operator_t add_op, |
| 1011 | size_t batch_size, |
| 1012 | const uint8_t* a, |
| 1013 | const uint8_t* b, |
| 1014 | uint8_t* sum, |
| 1015 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1016 | |
| 1017 | enum xnn_status xnn_create_average_pooling2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1018 | uint32_t input_padding_top, |
| 1019 | uint32_t input_padding_right, |
| 1020 | uint32_t input_padding_bottom, |
| 1021 | uint32_t input_padding_left, |
| 1022 | uint32_t pooling_height, |
| 1023 | uint32_t pooling_width, |
| 1024 | uint32_t stride_height, |
| 1025 | uint32_t stride_width, |
| 1026 | size_t channels, |
| 1027 | size_t input_pixel_stride, |
| 1028 | size_t output_pixel_stride, |
| 1029 | uint8_t input_zero_point, |
| 1030 | float input_scale, |
| 1031 | uint8_t output_zero_point, |
| 1032 | float output_scale, |
| 1033 | uint8_t output_min, |
| 1034 | uint8_t output_max, |
| 1035 | uint32_t flags, |
| 1036 | xnn_operator_t* average_pooling_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1037 | |
| 1038 | enum xnn_status xnn_setup_average_pooling2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1039 | xnn_operator_t average_pooling_op, |
| 1040 | size_t batch_size, |
| 1041 | size_t input_height, |
| 1042 | size_t input_width, |
| 1043 | const uint8_t* input, |
| 1044 | uint8_t* output, |
| 1045 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1046 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1047 | enum xnn_status xnn_create_convolution2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1048 | uint32_t input_padding_top, |
| 1049 | uint32_t input_padding_right, |
| 1050 | uint32_t input_padding_bottom, |
| 1051 | uint32_t input_padding_left, |
| 1052 | uint32_t kernel_height, |
| 1053 | uint32_t kernel_width, |
| 1054 | uint32_t subsampling_height, |
| 1055 | uint32_t subsampling_width, |
| 1056 | uint32_t dilation_height, |
| 1057 | uint32_t dilation_width, |
| 1058 | uint32_t groups, |
| 1059 | size_t group_input_channels, |
| 1060 | size_t group_output_channels, |
| 1061 | size_t input_pixel_stride, |
| 1062 | size_t output_pixel_stride, |
| 1063 | uint8_t input_zero_point, |
| 1064 | float input_scale, |
| 1065 | uint8_t kernel_zero_point, |
| 1066 | float kernel_scale, |
| 1067 | const uint8_t* kernel, |
| 1068 | const int32_t* bias, |
| 1069 | uint8_t output_zero_point, |
| 1070 | float output_scale, |
| 1071 | uint8_t output_min, |
| 1072 | uint8_t output_max, |
| 1073 | uint32_t flags, |
| 1074 | xnn_operator_t* convolution_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1075 | |
| 1076 | enum xnn_status xnn_setup_convolution2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1077 | xnn_operator_t convolution_op, |
| 1078 | size_t batch_size, |
| 1079 | size_t input_height, |
| 1080 | size_t input_width, |
| 1081 | const uint8_t* input, |
| 1082 | uint8_t* output, |
| 1083 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1084 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1085 | enum xnn_status xnn_create_deconvolution2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1086 | uint32_t output_padding_top, |
| 1087 | uint32_t output_padding_right, |
| 1088 | uint32_t output_padding_bottom, |
| 1089 | uint32_t output_padding_left, |
| 1090 | uint32_t kernel_height, |
| 1091 | uint32_t kernel_width, |
| 1092 | uint32_t stride_height, |
| 1093 | uint32_t stride_width, |
| 1094 | uint32_t dilation_height, |
| 1095 | uint32_t dilation_width, |
| 1096 | uint32_t groups, |
| 1097 | size_t group_input_channels, |
| 1098 | size_t group_output_channels, |
| 1099 | size_t input_pixel_stride, |
| 1100 | size_t output_pixel_stride, |
| 1101 | uint8_t input_zero_point, |
| 1102 | float input_scale, |
| 1103 | uint8_t kernel_zero_point, |
| 1104 | float kernel_scale, |
| 1105 | const uint8_t* kernel, |
| 1106 | const int32_t* bias, |
| 1107 | uint8_t output_zero_point, |
| 1108 | float output_scale, |
| 1109 | uint8_t output_min, |
| 1110 | uint8_t output_max, |
| 1111 | uint32_t flags, |
| 1112 | xnn_operator_t* deconvolution_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1113 | |
| 1114 | enum xnn_status xnn_setup_deconvolution2d_nhwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1115 | xnn_operator_t deconvolution_op, |
| 1116 | size_t batch_size, |
| 1117 | size_t input_height, |
| 1118 | size_t input_width, |
| 1119 | uint32_t adjustment_height, |
| 1120 | uint32_t adjustment_width, |
| 1121 | const uint8_t* input, |
| 1122 | uint8_t* output, |
| 1123 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1124 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1125 | enum xnn_status xnn_create_fully_connected_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1126 | size_t input_channels, |
| 1127 | size_t output_channels, |
| 1128 | size_t input_stride, |
| 1129 | size_t output_stride, |
| 1130 | uint8_t input_zero_point, |
| 1131 | float input_scale, |
| 1132 | uint8_t kernel_zero_point, |
| 1133 | float kernel_scale, |
| 1134 | const uint8_t* kernel, |
| 1135 | const int32_t* bias, |
| 1136 | uint8_t output_zero_point, |
| 1137 | float output_scale, |
| 1138 | uint8_t output_min, |
| 1139 | uint8_t output_max, |
| 1140 | uint32_t flags, |
| 1141 | xnn_operator_t* fully_connected_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1142 | |
| 1143 | enum xnn_status xnn_setup_fully_connected_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1144 | xnn_operator_t fully_connected_op, |
| 1145 | size_t batch_size, |
| 1146 | const uint8_t* input, |
| 1147 | uint8_t* output, |
| 1148 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1149 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1150 | enum xnn_status xnn_create_global_average_pooling_nwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1151 | size_t channels, |
| 1152 | size_t input_stride, |
| 1153 | size_t output_stride, |
| 1154 | uint8_t input_zero_point, |
| 1155 | float input_scale, |
| 1156 | uint8_t output_zero_point, |
| 1157 | float output_scale, |
| 1158 | uint8_t output_min, |
| 1159 | uint8_t output_max, |
| 1160 | uint32_t flags, |
| 1161 | xnn_operator_t* global_average_pooling_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1162 | |
| 1163 | enum xnn_status xnn_setup_global_average_pooling_nwc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1164 | xnn_operator_t global_average_pooling_op, |
| 1165 | size_t batch_size, |
| 1166 | size_t width, |
| 1167 | const uint8_t* input, |
| 1168 | uint8_t* output, |
| 1169 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1170 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1171 | enum xnn_status xnn_create_leaky_relu_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1172 | size_t channels, |
| 1173 | size_t input_stride, |
| 1174 | size_t output_stride, |
| 1175 | float negative_slope, |
| 1176 | uint8_t input_zero_point, |
| 1177 | float input_scale, |
| 1178 | uint8_t output_zero_point, |
| 1179 | float output_scale, |
| 1180 | uint8_t output_min, |
| 1181 | uint8_t output_max, |
| 1182 | uint32_t flags, |
| 1183 | xnn_operator_t* leaky_relu_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1184 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1185 | enum xnn_status xnn_setup_leaky_relu_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1186 | xnn_operator_t leaky_relu_op, |
| 1187 | size_t batch_size, |
| 1188 | const uint8_t* input, |
| 1189 | uint8_t* output, |
| 1190 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1191 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1192 | enum xnn_status xnn_create_sigmoid_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1193 | size_t channels, |
| 1194 | size_t input_stride, |
| 1195 | size_t output_stride, |
| 1196 | uint8_t input_zero_point, |
| 1197 | float input_scale, |
| 1198 | uint8_t output_zero_point, |
| 1199 | float output_scale, |
| 1200 | uint8_t output_min, |
| 1201 | uint8_t output_max, |
| 1202 | uint32_t flags, |
| 1203 | xnn_operator_t* sigmoid_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1204 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1205 | enum xnn_status xnn_setup_sigmoid_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1206 | xnn_operator_t sigmoid_op, |
| 1207 | size_t batch_size, |
| 1208 | const uint8_t* input, |
| 1209 | uint8_t* output, |
| 1210 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1211 | |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 1212 | enum xnn_status xnn_create_softmax_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1213 | size_t channels, |
| 1214 | size_t input_stride, |
| 1215 | size_t output_stride, |
| 1216 | float input_scale, |
| 1217 | uint8_t output_zero_point, |
| 1218 | float output_scale, |
| 1219 | uint32_t flags, |
| 1220 | xnn_operator_t* softmax_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1221 | |
Marat Dukhan | fd8e689 | 2020-01-27 15:25:25 -0800 | [diff] [blame] | 1222 | enum xnn_status xnn_setup_softmax_nc_q8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1223 | xnn_operator_t softmax_op, |
| 1224 | size_t batch_size, |
| 1225 | const uint8_t* input, |
| 1226 | uint8_t* output, |
| 1227 | pthreadpool_t threadpool); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1228 | |
| 1229 | #endif // XNN_NO_Q8_OPERATORS |
| 1230 | |
| 1231 | #ifndef XNN_NO_U8_OPERATORS |
| 1232 | |
| 1233 | enum xnn_status xnn_create_clamp_nc_u8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1234 | size_t channels, |
| 1235 | size_t input_stride, |
| 1236 | size_t output_stride, |
| 1237 | uint8_t output_min, |
| 1238 | uint8_t output_max, |
| 1239 | uint32_t flags, |
| 1240 | xnn_operator_t* clamp_op_out); |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1241 | |
| 1242 | enum xnn_status xnn_setup_clamp_nc_u8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1243 | xnn_operator_t clamp_op, |
| 1244 | size_t batch_size, |
| 1245 | const uint8_t* input, |
| 1246 | uint8_t* output, |
| 1247 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1248 | |
| 1249 | enum xnn_status xnn_create_max_pooling2d_nhwc_u8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1250 | uint32_t input_padding_top, |
| 1251 | uint32_t input_padding_right, |
| 1252 | uint32_t input_padding_bottom, |
| 1253 | uint32_t input_padding_left, |
| 1254 | uint32_t pooling_height, |
| 1255 | uint32_t pooling_width, |
| 1256 | uint32_t stride_height, |
| 1257 | uint32_t stride_width, |
| 1258 | uint32_t dilation_height, |
| 1259 | uint32_t dilation_width, |
| 1260 | size_t channels, |
| 1261 | size_t input_pixel_stride, |
| 1262 | size_t output_pixel_stride, |
| 1263 | uint8_t output_min, |
| 1264 | uint8_t output_max, |
| 1265 | uint32_t flags, |
| 1266 | xnn_operator_t* max_pooling_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1267 | |
| 1268 | enum xnn_status xnn_setup_max_pooling2d_nhwc_u8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1269 | xnn_operator_t max_pooling_op, |
| 1270 | size_t batch_size, |
| 1271 | size_t input_height, |
| 1272 | size_t input_width, |
| 1273 | const uint8_t* input, |
| 1274 | uint8_t* output, |
| 1275 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1276 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1277 | #endif // XNN_NO_U8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1278 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1279 | #ifndef XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1280 | |
| 1281 | enum xnn_status xnn_create_channel_shuffle_nc_x8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1282 | size_t groups, |
| 1283 | size_t group_channels, |
| 1284 | size_t input_stride, |
| 1285 | size_t output_stride, |
| 1286 | uint32_t flags, |
| 1287 | xnn_operator_t* channel_shuffle_op_out); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1288 | |
| 1289 | enum xnn_status xnn_setup_channel_shuffle_nc_x8( |
Marat Dukhan | 03bc407 | 2020-01-28 14:52:25 -0800 | [diff] [blame] | 1290 | xnn_operator_t channel_shuffle_op, |
| 1291 | size_t batch_size, |
| 1292 | const void* input, |
| 1293 | void* output, |
| 1294 | pthreadpool_t threadpool); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1295 | |
Marat Dukhan | d620972 | 2019-10-07 12:54:25 -0700 | [diff] [blame] | 1296 | #endif // XNN_NO_X8_OPERATORS |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1297 | |
| 1298 | #ifdef __cplusplus |
| 1299 | } // extern "C" |
| 1300 | #endif |