18#ifndef ONERT_MICRO_EXECUTE_PAL_UTILS_H
19#define ONERT_MICRO_EXECUTE_PAL_UTILS_H
31 const uint32_t output_depth)
33 std::pair<uint32_t, uint32_t> result(0u, output_depth);
40 result.second =
static_cast<uint32_t
>(
static_cast<float>(output_depth) / 2.f);
43 result.first =
static_cast<uint32_t
>(
static_cast<float>(output_depth) / 2.f);
46 assert(
"Unsupported type");
58static const uint16_t sigmoid_table_uint16[256] = {
59 32768, 33451, 34133, 34813, 35493, 36169, 36843, 37513, 38180, 38841, 39498, 40149, 40794, 41432,
60 42064, 42688, 43304, 43912, 44511, 45102, 45683, 46255, 46817, 47369, 47911, 48443, 48964, 49475,
61 49975, 50464, 50942, 51409, 51865, 52311, 52745, 53169, 53581, 53983, 54374, 54755, 55125, 55485,
62 55834, 56174, 56503, 56823, 57133, 57433, 57724, 58007, 58280, 58544, 58800, 59048, 59288, 59519,
63 59743, 59959, 60168, 60370, 60565, 60753, 60935, 61110, 61279, 61441, 61599, 61750, 61896, 62036,
64 62172, 62302, 62428, 62549, 62666, 62778, 62886, 62990, 63090, 63186, 63279, 63368, 63454, 63536,
65 63615, 63691, 63765, 63835, 63903, 63968, 64030, 64090, 64148, 64204, 64257, 64308, 64357, 64405,
66 64450, 64494, 64536, 64576, 64614, 64652, 64687, 64721, 64754, 64786, 64816, 64845, 64873, 64900,
67 64926, 64950, 64974, 64997, 65019, 65039, 65060, 65079, 65097, 65115, 65132, 65149, 65164, 65179,
68 65194, 65208, 65221, 65234, 65246, 65258, 65269, 65280, 65291, 65301, 65310, 65319, 65328, 65337,
69 65345, 65352, 65360, 65367, 65374, 65381, 65387, 65393, 65399, 65404, 65410, 65415, 65420, 65425,
70 65429, 65433, 65438, 65442, 65445, 65449, 65453, 65456, 65459, 65462, 65465, 65468, 65471, 65474,
71 65476, 65479, 65481, 65483, 65485, 65488, 65489, 65491, 65493, 65495, 65497, 65498, 65500, 65501,
72 65503, 65504, 65505, 65507, 65508, 65509, 65510, 65511, 65512, 65513, 65514, 65515, 65516, 65517,
73 65517, 65518, 65519, 65520, 65520, 65521, 65522, 65522, 65523, 65523, 65524, 65524, 65525, 65525,
74 65526, 65526, 65526, 65527, 65527, 65528, 65528, 65528, 65529, 65529, 65529, 65529, 65530, 65530,
75 65530, 65530, 65531, 65531, 65531, 65531, 65531, 65532, 65532, 65532, 65532, 65532, 65532, 65533,
76 65533, 65533, 65533, 65533, 65533, 65533, 65533, 65534, 65534, 65534, 65534, 65534, 65534, 65534,
77 65534, 65534, 65534, 65535};
81 bool overflow = a == b && a == std::numeric_limits<std::int32_t>::min();
84 std::int64_t ab_64 = a_64 * b_64;
85 std::int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30));
86 std::int32_t ab_x2_high32 =
static_cast<std::int32_t
>((ab_64 + nudge) / (1ll << 31));
87 return overflow ? std::numeric_limits<std::int32_t>::max() : ab_x2_high32;
94 assert(exponent >= 0);
95 assert(exponent <= 31);
96 const int32_t mask = int32_t((1ll << exponent) - 1);
97 const int32_t zero = int32_t(0);
98 const int32_t one = int32_t(1);
99 const int32_t remainder = x & mask;
100 const int32_t threshold = (mask >> 1) + ((x < zero ? one : zero) & one);
101 return (x >> exponent) + ((remainder > threshold ? one : zero) & one);
106 int left_shift = shift > 0 ? shift : 0;
107 int right_shift = shift > 0 ? 0 : -shift;
113 int32_t quantized_multiplier,
122 *min = params.int32_activation_min;
123 *max = params.int32_activation_max;
128 *min = params.float_activation_min;
129 *max = params.float_activation_max;
134 *min = params.int64_activation_min;
135 *max = params.int64_activation_max;
144 const int32_t num_axis,
const int32_t *axis)
151 for (
int idx = 0; idx < num_dims; ++idx)
154 bool is_axis =
false;
157 for (
int axis_idx = 0; axis_idx < num_axis; ++axis_idx)
159 if (idx == axis[axis_idx])
168 offset =
offset *
static_cast<size_t>(dims[idx]) +
static_cast<size_t>(index[idx]);
175inline bool nextIndex(
const int32_t num_dims,
const int32_t *dims, int32_t *current)
182 for (
int idx = num_dims - 1; idx >= 0; --idx)
184 int current_val = current[idx] + carry;
185 if (dims[idx] == current_val)
191 current[idx] = current_val;
203 assert(shape1.
dims(index1) == shape2.
dims(index2));
204 return shape1.
dims(index1);
213 for (
int i = 0; i < num_dims; ++i)
215 flat_size *= (i == skip_dim) ? 1 : dims_data[i];
220inline int offset(
const int32_t *dims_data,
int i0,
int i1,
int i2,
int i3)
222 return ((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3;
225inline int offset(
const int32_t *dims_data,
int i0,
int i1,
int i2,
int i3,
int i4)
227 return (((i0 * dims_data[1] + i1) * dims_data[2] + i2) * dims_data[3] + i3) * dims_data[4] + i4;
235 return min(max(x, output_activation_min), output_activation_max);
241template <
int MAX_DIM = 6>
244 size_t *compressed_input1_stride,
245 size_t *compressed_input2_stride,
size_t *compressed_output_shape)
247 size_t num_compressed_dims = 0;
248 size_t compressed_input1_shape[MAX_DIM];
249 size_t compressed_input2_shape[MAX_DIM];
250 std::fill(compressed_input1_shape, compressed_input1_shape + MAX_DIM, 1);
251 std::fill(compressed_input2_shape, compressed_input2_shape + MAX_DIM, 1);
252 std::fill(compressed_output_shape, compressed_output_shape + MAX_DIM, 1);
253 bool broadcast_input1 =
false;
254 bool broadcast_input2 =
false;
255 bool first_nonunit =
true;
263 const int32_t *input1_dims = input1_shape.
dimsData();
264 const int32_t *input2_dims = input2_shape.
dimsData();
265 const size_t num_common_dims = std::min(num_input1_dims, num_input2_dims);
266 for (
size_t i = 1; i <= num_common_dims; i++)
268 if (input1_dims[num_input1_dims - i] < 0 || input2_dims[num_input2_dims - i] < 0)
272 const size_t input1_dim = input1_dims[num_input1_dims - i];
273 const size_t input2_dim = input2_dims[num_input2_dims - i];
274 if (input1_dim == 0 || input2_dim == 0)
278 if (input1_dim == 1 && input2_dim == 1)
282 assert(!broadcast_input1 || !broadcast_input2);
286 if (!broadcast_input1)
288 broadcast_input1 =
true;
289 broadcast_input2 =
false;
290 num_compressed_dims++;
292 compressed_input2_shape[num_compressed_dims - 1] *= input2_dim;
293 compressed_output_shape[num_compressed_dims - 1] *= input2_dim;
295 else if (input2_dim == 1)
297 if (!broadcast_input2)
299 broadcast_input1 =
false;
300 broadcast_input2 =
true;
301 num_compressed_dims++;
303 compressed_input1_shape[num_compressed_dims - 1] *= input1_dim;
304 compressed_output_shape[num_compressed_dims - 1] *= input1_dim;
308 assert(input1_dim == input2_dim);
309 if (broadcast_input1 || broadcast_input2 || first_nonunit)
311 broadcast_input1 =
false;
312 broadcast_input2 =
false;
313 num_compressed_dims++;
315 compressed_input1_shape[num_compressed_dims - 1] *= input1_dim;
316 compressed_input2_shape[num_compressed_dims - 1] *= input1_dim;
317 compressed_output_shape[num_compressed_dims - 1] *= input1_dim;
319 first_nonunit =
false;
321 if (num_input1_dims > num_input2_dims)
323 if (!broadcast_input2)
325 num_compressed_dims++;
327 for (
size_t i = 0; i < num_input1_dims - num_input2_dims; i++)
329 if (input1_dims[i] < 0)
331 const size_t input1_dim = input1_dims[i];
336 compressed_input1_shape[num_compressed_dims - 1] *= input1_dim;
337 compressed_output_shape[num_compressed_dims - 1] *= input1_dim;
340 else if (num_input2_dims > num_input1_dims)
342 if (!broadcast_input1)
344 num_compressed_dims++;
346 for (
size_t i = 0; i < num_input2_dims - num_input1_dims; i++)
348 if (input2_dims[i] < 0)
350 const size_t input2_dim = input2_dims[i];
355 compressed_input2_shape[num_compressed_dims - 1] *= input2_dim;
356 compressed_output_shape[num_compressed_dims - 1] *= input2_dim;
359 num_compressed_dims = (num_compressed_dims > 1) ? num_compressed_dims : 1;
361 int input1_stride = 1;
362 int input2_stride = 1;
363 for (
int i = 0; i < MAX_DIM; ++i)
365 compressed_input1_stride[i] = input1_stride;
366 input1_stride *= compressed_input1_shape[i];
367 compressed_input2_stride[i] = input2_stride;
368 input2_stride *= compressed_input2_shape[i];
370 for (
int i = 0; i < MAX_DIM; ++i)
372 if (compressed_input1_shape[i] != compressed_input2_shape[i])
374 if (compressed_input1_shape[i] == 1)
376 compressed_input1_stride[i] = 0;
380 assert(compressed_input2_shape[i] == 1);
381 compressed_input2_stride[i] = 0;
int32_t dims(int i) const
int32_t dimensionsCount() const
std::pair< uint32_t, uint32_t > getUpLowerWeightTensorDepth(core::OpTrainableRankType rank, const uint32_t output_depth)
int flatSizeSkipDim(const int32_t *dims_data, int skip_dim, int num_dims)
std::int32_t saturatingRoundingDoublingHighMul(std::int32_t a, std::int32_t b)
bool nextIndex(const int32_t num_dims, const int32_t *dims, int32_t *current)
void getActivationParams(const P ¶ms, int32_t *min, int32_t *max)
int32_t multiplyByQuantizedMultiplierSmallerThanOneExp(int32_t x, int32_t quantized_multiplier, int left_shift)
bool ReduceDimensionsForBroadcast(const core::OMRuntimeShape &input1_shape, const core::OMRuntimeShape &input2_shape, size_t *compressed_input1_stride, size_t *compressed_input2_stride, size_t *compressed_output_shape)
int MatchingDim(const core::OMRuntimeShape &shape1, int index1, const core::OMRuntimeShape &shape2, int index2)
size_t reducedOutputOffset(const int32_t num_dims, const int32_t *dims, const int32_t *index, const int32_t num_axis, const int32_t *axis)
int offset(const int32_t *dims_data, int i0, int i1, int i2, int i3)
int32_t multiplyByQuantizedMultiplier(int32_t x, int32_t quantized_multiplier, int shift)
int32_t roundingDivideByPOT(int32_t x, int32_t exponent)
T activationFunctionWithMinMax(T x, T output_activation_min, T output_activation_max)