2 // Copyright (c) 2016-2019 Intel Corporation
4 // Licensed under the Apache License, Version 2.0 (the "License");
5 // you may not use this file except in compliance with the License.
6 // You may obtain a copy of the License at
8 // http://www.apache.org/licenses/LICENSE-2.0
10 // Unless required by applicable law or agreed to in writing, software
11 // distributed under the License is distributed on an "AS IS" BASIS,
12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 // See the License for the specific language governing permissions and
14 // limitations under the License.
17 ///////////////////////////////////////////////////////////////////////////////////////////////////
29 /// @addtogroup cpp_api C++ API
32 /// @addtogroup cpp_memory Memory description and management
35 constexpr size_t float_type_mask = 0x80;
36 constexpr size_t uint_type_mask = 0x40;
37 constexpr size_t bin_type_mask = 0x20;
39 /// @brief Possible data types could be stored in memory.
40 enum class data_types : size_t {
41 bin = sizeof(int32_t) | bin_type_mask,
42 u8 = sizeof(uint8_t) | uint_type_mask,
44 f16 = sizeof(int16_t) | float_type_mask,
45 f32 = sizeof(float) | float_type_mask,
46 i32 = sizeof(int32_t),
50 class optional_data_type {
51 // Must be the same as the undrelying type of `data_types`.
52 using storage_type = size_t;
54 // Implicitly assumes that this value is not used in the `data_types`.
55 static constexpr auto non_specified_type =
56 std::numeric_limits<storage_type>::max();
60 : storage(non_specified_type) {}
62 explicit optional_data_type(data_types type)
63 : storage(static_cast<storage_type>(type)) {}
65 operator bool() const { return storage != non_specified_type; }
67 // Similarly to std::optional does *not* verify that the object has the type
68 // set. Unlike it, though, returns the value instead of pointer/reference.
69 data_types operator*() const { return static_cast<data_types>(storage); }
71 optional_data_type& operator=(const data_types new_type) {
72 storage = static_cast<storage_type>(new_type);
80 /// Converts C++ type to @ref data_types .
82 struct type_to_data_type;
83 #ifndef DOXYGEN_SHOULD_SKIP_THIS
85 struct type_to_data_type<int8_t> { static const data_types value = data_types::i8; };
87 struct type_to_data_type<uint8_t> { static const data_types value = data_types::u8; };
89 struct type_to_data_type<int32_t> { static const data_types value = data_types::i32; };
91 struct type_to_data_type<int64_t> { static const data_types value = data_types::i64; };
93 struct type_to_data_type<half_t> { static const data_types value = data_types::f16; };
95 struct type_to_data_type<float> { static const data_types value = data_types::f32; };
98 /// Converts @ref data_types to C++ type.
99 template <data_types Data_Type>
100 struct data_type_to_type;
101 #ifndef DOXYGEN_SHOULD_SKIP_THIS
103 struct data_type_to_type<data_types::bin> { typedef uint32_t type; };
105 struct data_type_to_type<data_types::i8> { typedef int8_t type; };
107 struct data_type_to_type<data_types::i32> { typedef int32_t type; };
109 struct data_type_to_type<data_types::i64> { typedef int64_t type; };
111 struct data_type_to_type<data_types::f16> { typedef half_t type; };
113 struct data_type_to_type<data_types::f32> { typedef float type; };
116 /// Helper class to identify key properties for data_types.
117 struct data_type_traits {
118 static size_t size_of(data_types data_type) {
119 return (static_cast<uint32_t>(data_type) & ~(float_type_mask | uint_type_mask | bin_type_mask));
122 static bool is_floating_point(data_types data_type) {
123 return (static_cast<uint32_t>(data_type) & float_type_mask) != 0;
126 static size_t align_of(data_types data_type) {
128 case data_types::bin:
129 return alignof(data_type_to_type<data_types::bin>::type);
131 return alignof(data_type_to_type<data_types::i8>::type);
132 case data_types::i32:
133 return alignof(data_type_to_type<data_types::i32>::type);
134 case data_types::i64:
135 return alignof(data_type_to_type<data_types::i64>::type);
136 case data_types::f16:
137 return alignof(data_type_to_type<data_types::f16>::type);
138 case data_types::f32:
139 return alignof(data_type_to_type<data_types::f32>::type);
145 static std::string name(data_types data_type) {
151 case data_types::i32:
153 case data_types::i64:
155 case data_types::f16:
157 case data_types::f32:
161 return std::string("invalid data type: " + std::to_string(static_cast<int>(data_type)));
164 template <typename T>
165 static T max(data_types data_type) {
168 return static_cast<T>(std::numeric_limits<int8_t>::max());
170 return static_cast<T>(std::numeric_limits<uint8_t>::max());
171 case data_types::i32:
172 return static_cast<T>(std::numeric_limits<int32_t>::max());
173 case data_types::i64:
174 return static_cast<T>(std::numeric_limits<int64_t>::max());
175 case data_types::f16:
176 return static_cast<T>(65504);
177 case data_types::f32:
178 return static_cast<T>(std::numeric_limits<float>::max());
181 return static_cast<T>(0);
184 template <typename T>
185 static T min(data_types data_type) {
188 return static_cast<T>(std::numeric_limits<int8_t>::lowest());
190 return static_cast<T>(std::numeric_limits<uint8_t>::lowest());
191 case data_types::i32:
192 return static_cast<T>(std::numeric_limits<int32_t>::lowest());
193 case data_types::i64:
194 return static_cast<T>(std::numeric_limits<int64_t>::lowest());
195 case data_types::f16:
196 return static_cast<T>(-65504);
197 case data_types::f32:
198 return static_cast<T>(std::numeric_limits<float>::lowest());
201 return static_cast<T>(0);
206 /// Helper function to check if C++ type matches @p data_type.
207 template <typename T>
208 bool data_type_match(data_types data_type) {
209 return data_type == type_to_data_type<T>::value;
212 /// Helper function to get both data_types and format::type in a single, unique value. Useable in 'case' statement.
213 constexpr auto fuse(data_types dt, cldnn::format::type fmt) -> decltype(static_cast<std::underlying_type<data_types>::type>(dt) |
214 static_cast<std::underlying_type<format::type>::type>(fmt)) {
215 using dt_type = std::underlying_type<data_types>::type;
216 using fmt_type = std::underlying_type<cldnn::format::type>::type;
217 using fmt_narrow_type = int16_t;
219 return static_cast<fmt_type>(fmt) <= std::numeric_limits<fmt_narrow_type>::max() &&
220 static_cast<dt_type>(dt) <= (std::numeric_limits<dt_type>::max() >> (sizeof(fmt_narrow_type) * 8))
221 ? (static_cast<dt_type>(dt) << (sizeof(fmt_narrow_type) * 8)) |
222 (static_cast<fmt_type>(fmt) >= 0 ? static_cast<fmt_narrow_type>(fmt) : static_cast<fmt_narrow_type>(-1))
223 : throw std::invalid_argument("data_type and/or format values are too big to be fused into single value");
226 /// @brief Represents data padding information.
228 /// @brief Filling value for padding area.
229 float filling_value() const { return _filling_value; }
231 /// @brief Gets lower padding sizes. For spatials, it means size of left (X) and top (Y) padding.
232 /// @return Tensor with padding for top/left/lower bounds of data.
233 tensor lower_size() const { return _lower_size; }
235 /// @brief Gets upper padding sizes. For spatials, it means size of right (X) and bottom (Y) padding.
236 /// @return Tensor with padding for bottom/right/upper bounds of data.
237 tensor upper_size() const { return _upper_size; }
240 /// @param lower_sizes Top-left padding sizes. See @ref tensor::tensor(const std::vector<value_type>&, value_type) for details.
241 /// @param upper_sizes Bottom-right padding sizes. See @ref tensor::tensor(const std::vector<value_type>&, value_type) for details.
242 /// @param filling_value Filling value for padding area.
243 padding(const std::vector<tensor::value_type>& lower_sizes, const std::vector<tensor::value_type>& upper_sizes, float filling_value = 0.0f)
244 : _lower_size(to_abs(lower_sizes), 0), _upper_size(to_abs(upper_sizes), 0), _filling_value(filling_value) {}
246 /// @brief Constrcuts symmetric padding.
247 /// @param sizes Top-left and bottom-right padding sizes. See @ref tensor::tensor(const std::vector<value_type>&, value_type) for details.
248 /// @param filling_value Filling value for padding area.
249 explicit padding(const std::vector<tensor::value_type>& sizes, float filling_value = 0.0f)
250 : padding(sizes, sizes, filling_value) {}
252 /// @brief Constructs "zero-sized" padding.
253 padding() : padding({0, 0, 0, 0}, 0) {}
255 /// @brief Returns true if padding size is not zero.
256 explicit operator bool() const {
257 return std::any_of(_lower_size.raw.begin(), _lower_size.raw.end(), [](const tensor::value_type& el) { return el != 0; }) ||
258 std::any_of(_upper_size.raw.begin(), _upper_size.raw.end(), [](const tensor::value_type& el) { return el != 0; });
261 friend bool operator==(const padding& lhs, const padding& rhs) {
262 return lhs._lower_size == rhs._lower_size && lhs._upper_size == rhs._upper_size && lhs._filling_value == rhs._filling_value;
265 friend bool operator!=(const padding& lhs, const padding& rhs) {
266 return !(lhs == rhs);
269 friend bool operator<(const padding& lhs, const padding& rhs) {
270 if (lhs._filling_value != rhs._filling_value)
271 return (lhs._filling_value < rhs._filling_value);
272 if (lhs._lower_size != rhs._lower_size)
273 return (lhs._lower_size < rhs._lower_size);
274 return (lhs._upper_size < rhs._upper_size);
277 static padding max(padding const& lhs, padding const& rhs, float filling_value = 0.0f) {
278 auto lower = tensor::max(lhs.lower_size(), rhs.lower_size());
279 auto upper = tensor::max(lhs.upper_size(), rhs.upper_size());
280 return padding{lower.sizes(), upper.sizes(), filling_value};
284 tensor _lower_size; ///< Lower padding sizes. For spatials, it means size of left (X) and top (Y) padding.
285 tensor _upper_size; ///< Upper padding sizes. For spatials, it means size of right (X) and bottom (Y) padding.
286 // TODO: Add support for non-zero filling value (if necessary) or remove variable (if not necessary).
287 float _filling_value; ///< Filling value for an element of padding. If data type of elements is different than float it is converted
288 ///< to it using round-towards-nearest-even (for floating-point data types) or round-towards-zero (for integral
291 static std::vector<tensor::value_type> to_abs(const std::vector<tensor::value_type>& sizes) {
292 std::vector<tensor::value_type> result;
293 result.reserve(sizes.size());
294 std::transform(sizes.cbegin(), sizes.cend(), std::back_inserter(result), [](const tensor::value_type& el) { return abs(el); });
295 return result; // NRVO
299 /// @brief Describes memory layout.
300 /// @details Contains information about data stored in @ref memory.
302 /// Constructs layout based on @p data_type and @p size information described by @ref tensor
303 layout(data_types data_type, cldnn::format fmt, tensor size, padding apadding = padding())
304 : data_type(data_type), format(fmt), size(size), data_padding(apadding) {}
306 layout(const layout& other) = default;
308 layout& operator=(const layout& other) {
311 data_type = other.data_type;
312 format = other.format;
314 data_padding = other.data_padding;
318 friend bool operator==(const layout& lhs, const layout& rhs) {
319 return lhs.data_type == rhs.data_type && lhs.format == rhs.format && lhs.size == rhs.size && lhs.data_padding == rhs.data_padding;
322 friend bool operator!=(const layout& lhs, const layout& rhs) {
323 return !(lhs == rhs);
326 friend bool operator<(const layout& lhs, const layout& rhs) {
327 if (lhs.data_type != rhs.data_type)
328 return (lhs.data_type < rhs.data_type);
329 if (lhs.format != rhs.format)
330 return (lhs.format < rhs.format);
331 if (lhs.size < rhs.size)
332 return (lhs.size < rhs.size);
333 return (lhs.data_padding < rhs.data_padding);
336 /// Number of elements to be stored in this memory layout
337 size_t count() const { return size.count(); }
339 /// Layout size with padding included
340 tensor get_buffer_size() const {
341 return size.add(data_padding.lower_size()).add(data_padding.upper_size());
344 tensor get_pitches() const {
345 auto sizes = get_buffer_size().sizes(format);
347 if (format == format::byxf_af32) {
348 sizes[3] = align_to(sizes[3], 32);
351 if (format == format::byx8_f4) {
352 sizes[3] = align_to(sizes[3], 4);
353 sizes[2] = align_to(sizes[2], 8);
355 std::vector<tensor::value_type> pitches(sizes.size(), tensor::value_type(1));
356 std::partial_sum(sizes.rbegin(), sizes.rend() - 1, pitches.rbegin() + 1, std::multiplies<tensor::value_type>());
357 return {format, pitches};
360 // @brief Calculates position within buffer of the data element pointed by the provided tensor.
361 // element == { 0,0,0,0 } means first no-padding (i.e. data) element
362 size_t get_linear_offset(tensor element = tensor(0)) const {
363 auto l_padd = data_padding.lower_size();
364 auto u_padd = data_padding.upper_size();
366 if ((element.batch[0] < 0 && -element.batch[0] > l_padd.batch[0]) ||
367 (element.feature[0] < 0 && -element.feature[0] > l_padd.feature[0]) ||
368 (element.spatial[0] < 0 && -element.spatial[0] > l_padd.spatial[0]) ||
369 (element.spatial[1] < 0 && -element.spatial[1] > l_padd.spatial[1]) ||
370 (element.spatial[2] < 0 && -element.spatial[2] > l_padd.spatial[2]) ||
371 (element.spatial[3] < 0 && -element.spatial[3] > l_padd.spatial[3]) ||
372 (element.batch[0] >= size.batch[0] + u_padd.batch[0]) ||
373 (element.feature[0] >= size.feature[0] + u_padd.feature[0]) ||
374 (element.spatial[0] >= size.spatial[0] + u_padd.spatial[0]) ||
375 (element.spatial[1] >= size.spatial[1] + u_padd.spatial[1]) ||
376 (element.spatial[2] >= size.spatial[2] + u_padd.spatial[2]) ||
377 (element.spatial[3] >= size.spatial[3] + u_padd.spatial[3]))
378 throw std::invalid_argument("Requested to calculate linear offset for an element which lies outside of the buffer range.");
380 auto padded_size = size + l_padd + u_padd;
381 auto padded_element = element + l_padd;
383 return padded_size.get_linear_offset(padded_element, format);
386 /// @brief Get aligned linear size calculated as multiplication of all elements.
387 size_t get_linear_size() const {
388 auto sizes = get_buffer_size().sizes();
390 for (const auto& block : this->format.block_sizes()) {
391 auto block_axis = block.first;
392 auto block_size = block.second;
394 sizes[block_axis] = align_to(sizes[block_axis], block_size);
397 if (this->format == cldnn::format::bf8_xy16 && !(is_aligned_to(sizes[1], 8) && is_aligned_to(sizes[2] * sizes[3], 16))) {
398 sizes[3] = align_to(sizes[2] * sizes[3], 16);
400 } else if (this->format == cldnn::format::byxf_af32 && !(is_aligned_to(sizes[1], 32))) {
401 sizes[1] = align_to(sizes[1], 32);
402 } else if (this->format == cldnn::format::byx8_f4 && (!is_aligned_to(sizes[1], 4) || !is_aligned_to(sizes[2], 8))) {
403 sizes[1] = align_to(sizes[1], 4);
404 sizes[2] = align_to(sizes[2], 8);
405 } else if (this->format == cldnn::format::os_is_yx_isa8_osv8_isv4 && !(is_aligned_to(sizes[0], 8)) && !(is_aligned_to(sizes[1], 32))) {
406 sizes[0] = align_to(sizes[0], 8);
407 sizes[1] = align_to(sizes[1], 32);
408 } else if (this->format == cldnn::format::os_is_yx_isa8_osv8_isv4_swizzled_by_4 && !(is_aligned_to(sizes[0], 32)) && !(is_aligned_to(sizes[1], 32))) {
409 sizes[0] = align_to(sizes[0], 32);
410 sizes[1] = align_to(sizes[1], 32);
411 } else if (this->format == cldnn::format::is_o32_yx_isv32_swizzled_by_4 && (!is_aligned_to(sizes[1], 32) || !(is_aligned_to(sizes[0], 32)))) {
412 sizes[0] = align_to(sizes[0], 32);
413 sizes[1] = align_to(sizes[1], 32);
414 } else if (this->format == cldnn::format::os_is_y_x8_osv8_isv4 || this->format == cldnn::format::os_is_y_x8_osv8_isv4_swizzled_by_4) {
415 sizes[1] = align_to(sizes[1], 4);
416 sizes[0] = align_to(sizes[0], 8);
417 sizes[2] = align_to(sizes[2], 8);
418 } else if (this->format == cldnn::format::b_fs_yx_32fp) {
419 sizes[1] = align_to(sizes[1], 32);
420 } else if (this->format == cldnn::format::os_is_yx_osv32_isv32p) {
421 sizes[0] = align_to(sizes[0], 32);
422 sizes[1] = align_to(sizes[1], 32);
424 size_t total = std::accumulate(
427 static_cast<size_t>(1),
428 std::multiplies<size_t>());
430 return (this->data_type == data_types::bin) ? ceil_div(total, 32) : total;
433 /// Modify padding in layout
434 layout with_padding(padding const& padd) const {
436 ret.data_padding = padd;
440 /// Data type stored in @ref memory (see. @ref data_types)
441 data_types data_type;
443 /// Format stored in @ref memory (see. @ref format)
444 cldnn::format format;
446 /// The size of the @ref memory (excluding padding)
449 /// Explicit padding of the @ref memory
450 padding data_padding;
452 /// Number of bytes needed to store this layout
453 size_t bytes_count() const { return data_type_traits::size_of(data_type) * get_linear_size(); }
455 bool has_fused_format(data_types const& dt, cldnn::format const& fmt) const {
456 return (data_type == dt && format == fmt);
459 auto fused_format() const -> decltype(fuse(data_type, format)) {
460 return fuse(data_type, format);