Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2020 Arm Limited. |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #include "arm_compute/core/NEON/kernels/NEQLSTMLayerNormalizationKernel.h" |
| 25 | |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 26 | #include "arm_compute/core/Helpers.h" |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 27 | #include "arm_compute/core/TensorInfo.h" |
| 28 | #include "arm_compute/core/Utils.h" |
| 29 | #include "arm_compute/core/Validate.h" |
| 30 | #include "arm_compute/core/Window.h" |
| 31 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame^] | 32 | #include "src/core/CPP/Validate.h" |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 33 | #include "src/core/NEON/NEFixedPoint.h" |
| 34 | #include "src/core/NEON/NEMath.h" |
| 35 | #include "src/core/NEON/NESymm.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame^] | 36 | #include "src/core/helpers/AutoConfiguration.h" |
| 37 | #include "src/core/helpers/WindowHelpers.h" |
Georgios Pinitas | ddb93bb | 2020-10-02 16:38:59 +0100 | [diff] [blame] | 38 | |
| 39 | #include "src/core/NEON/kernels/detail/NEActivationFunctionDetail.h" |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 40 | |
| 41 | #include <map> |
| 42 | |
| 43 | namespace arm_compute |
| 44 | { |
| 45 | namespace |
| 46 | { |
| 47 | inline std::pair<int64_t, int64_t> compute_mean_variance(int64_t sum, int64_t sum_sq, uint32_t num_input) |
| 48 | { |
| 49 | const auto temp = static_cast<int64_t>(0x100000) / num_input; |
| 50 | const auto mean = sum * 1024 / static_cast<int64_t>(num_input); |
| 51 | const int64_t variance = ((sum_sq * temp) - (mean * mean)) / 0x100000; |
| 52 | |
| 53 | return std::make_pair(mean, variance); |
| 54 | } |
| 55 | |
| 56 | inline int64x2x2_t mul_add(const int32x4_t &a, const int32x4_t &b, const int32x4_t &bias) |
| 57 | { |
| 58 | using namespace wrapper; |
| 59 | const int64x2_t a_low = vmovl(vgetlow(a)); |
| 60 | const int64x2_t a_high = vmovl(vgethigh(a)); |
| 61 | const int64x2_t b_low = vmovl(vgetlow(b)); |
| 62 | const int64x2_t b_high = vmovl(vgethigh(b)); |
| 63 | |
| 64 | const int64_t a_0 = vgetlane(a_low, 0); |
| 65 | const int64_t a_1 = vgetlane(a_low, 1); |
| 66 | const int64_t a_2 = vgetlane(a_high, 0); |
| 67 | const int64_t a_3 = vgetlane(a_high, 1); |
| 68 | |
| 69 | const int64_t b_0 = vgetlane(b_low, 0); |
| 70 | const int64_t b_1 = vgetlane(b_low, 1); |
| 71 | const int64_t b_2 = vgetlane(b_high, 0); |
| 72 | const int64_t b_3 = vgetlane(b_high, 1); |
| 73 | |
| 74 | int64x2x2_t result; |
| 75 | const int64x2_t result_0{ a_0 * b_0, a_1 * b_1 }; |
| 76 | const int64x2_t result_1{ a_2 * b_2, a_3 * b_3 }; |
| 77 | result.val[0] = vadd(vmovl(vgetlow(bias)), result_0); |
| 78 | result.val[1] = vadd(vmovl(vgethigh(bias)), result_1); |
| 79 | |
| 80 | return result; |
| 81 | } |
| 82 | } // namespace |
| 83 | |
| 84 | void NEQLSTMLayerNormalizationKernel::configure(const ITensor *input, ITensor *output, const ITensor *weight, const ITensor *bias) |
| 85 | { |
Sang-Hoon Park | 9230e27 | 2020-04-18 00:46:34 +0100 | [diff] [blame] | 86 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weight, bias, output); |
| 87 | ARM_COMPUTE_ERROR_ON(input == output); |
| 88 | ARM_COMPUTE_ERROR_THROW_ON(validate(input->info(), output->info(), weight->info(), bias->info())); |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 89 | |
| 90 | static const std::map<DataType, ComputeFuncType> fn_map = |
| 91 | { |
| 92 | { DataType::QSYMM16, std::mem_fn(&NEQLSTMLayerNormalizationKernel::compute_qsymm16) }, |
| 93 | }; |
| 94 | |
| 95 | _input = input; |
| 96 | _output = output; |
| 97 | _weight = weight; |
| 98 | _bias = bias; |
| 99 | _fn = fn_map.at(_input->info()->data_type()); |
| 100 | |
| 101 | auto_init_if_empty(*_output->info(), *_input->info()); |
Sang-Hoon Park | 9230e27 | 2020-04-18 00:46:34 +0100 | [diff] [blame] | 102 | _output->info()->set_quantization_info(compute_output_qinfo()); |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 103 | |
| 104 | const UniformQuantizationInfo wq_info = _weight->info()->quantization_info().uniform(); |
| 105 | const Status s = quantization::calculate_quantized_multiplier(wq_info.scale, &_output_multiplier, &_output_shift); |
| 106 | _output_shift *= -1; |
| 107 | |
| 108 | if(!bool(s)) |
| 109 | { |
| 110 | _output_multiplier = 0; |
| 111 | _output_shift = 0; |
| 112 | } |
| 113 | |
| 114 | Window win = configure_window(output); |
| 115 | INEKernel::configure(win); |
| 116 | } |
| 117 | |
| 118 | Window NEQLSTMLayerNormalizationKernel::configure_window(ITensor *target) |
| 119 | { |
| 120 | Window window = calculate_max_window(*target->info(), Steps()); |
| 121 | Coordinates coord; |
| 122 | coord.set_num_dimensions(target->info()->num_dimensions()); |
| 123 | target->info()->set_valid_region(ValidRegion(coord, target->info()->tensor_shape())); |
| 124 | |
| 125 | _window_start_x = static_cast<int32_t>(window.x().start()); |
| 126 | _window_end_x = static_cast<int32_t>(window.x().end()); |
| 127 | _window_step_x = static_cast<int32_t>(vector_size_byte) / _output->info()->element_size(); |
| 128 | |
| 129 | // input and output windows will iterator over y-axis, while execute_window will handler x-axis. |
| 130 | _inout_window = window; |
| 131 | _inout_window.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 132 | |
| 133 | // weight and bias cannot iterator along y-axis since they are 1D. |
| 134 | _weight_window = _inout_window; |
| 135 | _weight_window.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 136 | |
| 137 | return window; |
| 138 | } |
| 139 | |
| 140 | Status NEQLSTMLayerNormalizationKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *weight, const ITensorInfo *bias) |
| 141 | { |
| 142 | ARM_COMPUTE_UNUSED(output, bias, weight, input); |
| 143 | |
| 144 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, weight, bias, output); |
| 145 | |
| 146 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QSYMM16); |
| 147 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weight, 1, DataType::QSYMM16); |
| 148 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(bias, 1, DataType::S32); |
| 149 | |
| 150 | ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > max_input_dimension); |
| 151 | ARM_COMPUTE_RETURN_ERROR_ON(weight->num_dimensions() > max_weight_dimension); |
| 152 | ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > max_bias_dimension); |
| 153 | |
| 154 | ARM_COMPUTE_RETURN_ERROR_ON(input->tensor_shape().x() != weight->tensor_shape().x()); |
| 155 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(weight, bias); |
| 156 | |
| 157 | if(output->total_size() != 0) |
| 158 | { |
| 159 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 160 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 161 | } |
| 162 | |
| 163 | return Status{}; |
| 164 | } |
| 165 | |
| 166 | void NEQLSTMLayerNormalizationKernel::run(const Window &window, const ThreadInfo &info) |
| 167 | { |
| 168 | ARM_COMPUTE_UNUSED(window, info); |
| 169 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 170 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
| 171 | ARM_COMPUTE_ERROR_ON_MSG(!_fn, "internal function is not defined for computation"); |
| 172 | |
| 173 | _fn(*this); |
| 174 | } |
| 175 | |
Sang-Hoon Park | 9230e27 | 2020-04-18 00:46:34 +0100 | [diff] [blame] | 176 | inline QuantizationInfo NEQLSTMLayerNormalizationKernel::compute_output_qinfo() |
| 177 | { |
Sheri Zhang | 3a35398 | 2020-04-21 13:10:24 +0100 | [diff] [blame] | 178 | return QuantizationInfo(1.f / 4096); |
Sang-Hoon Park | 9230e27 | 2020-04-18 00:46:34 +0100 | [diff] [blame] | 179 | } |
| 180 | |
Sang-Hoon Park | 0d008f7 | 2020-03-13 14:56:05 +0000 | [diff] [blame] | 181 | inline std::pair<int64_t, int64_t> NEQLSTMLayerNormalizationKernel::sum_qsymm16(const int16_t *input_ptr) |
| 182 | { |
| 183 | ARM_COMPUTE_ERROR_ON(!input_ptr); |
| 184 | |
| 185 | using AccType = int64_t; |
| 186 | using InputDataType = int16_t; |
| 187 | |
| 188 | AccType sum{ 0 }; |
| 189 | AccType sum_sq{ 0 }; |
| 190 | |
| 191 | int32_t x = _window_start_x; |
| 192 | for(; x <= _window_end_x && _window_step_x <= (_window_end_x - x); x += _window_step_x) |
| 193 | { |
| 194 | using namespace wrapper; |
| 195 | const int16x8_t val = vloadq(input_ptr + x); |
| 196 | const int32x4_t val_low = vmovl(vgetlow(val)); |
| 197 | const int32x4_t val_high = vmovl(vgethigh(val)); |
| 198 | |
| 199 | #if defined(__aarch64__) |
| 200 | sum += static_cast<AccType>(vaddv(val_low)); |
| 201 | sum += static_cast<AccType>(vaddv(val_high)); |
| 202 | |
| 203 | sum_sq += static_cast<AccType>(vaddv(vmul(val_low, val_low))); |
| 204 | sum_sq += static_cast<AccType>(vaddv(vmul(val_high, val_high))); |
| 205 | #else // __aarch64__ |
| 206 | // only AArch64 supports vaddv |
| 207 | const int64x2_t pair_sum_low = vpaddl(val_low); |
| 208 | const int64x2_t pair_sum_high = vpaddl(val_high); |
| 209 | const int64x2_t pair_sum = vadd(pair_sum_low, pair_sum_high); |
| 210 | sum += vgetlane(pair_sum, 0) + vgetlane(pair_sum, 1); |
| 211 | |
| 212 | const int32x4_t square_low = vmul(val_low, val_low); |
| 213 | const int32x4_t square_high = vmul(val_high, val_high); |
| 214 | const int64x2_t pair_sum_sq_low = vpaddl(square_low); |
| 215 | const int64x2_t pair_sum_sq_high = vpaddl(square_high); |
| 216 | const int64x2_t pair_sum_sq = vadd(pair_sum_sq_low, pair_sum_sq_high); |
| 217 | sum_sq += vgetlane(pair_sum_sq, 0) + vgetlane(pair_sum_sq, 1); |
| 218 | #endif // __aarch64__ |
| 219 | } |
| 220 | |
| 221 | for(; x < _window_end_x; ++x) |
| 222 | { |
| 223 | const InputDataType val = input_ptr[x]; |
| 224 | sum += static_cast<AccType>(val); |
| 225 | sum_sq += static_cast<AccType>(val * val); |
| 226 | } |
| 227 | |
| 228 | return std::make_pair(sum, sum_sq); |
| 229 | } |
| 230 | |
| 231 | inline void NEQLSTMLayerNormalizationKernel::normalize_qasymm16(const int16_t *input_ptr, |
| 232 | int16_t *output_ptr, |
| 233 | const int16_t *weight_ptr, |
| 234 | const int32_t *bias_ptr, |
| 235 | int32_t mean, int32_t inv_std_mul, int32_t inv_std_shift) |
| 236 | { |
| 237 | using OutputDataType = int16_t; |
| 238 | |
| 239 | using namespace wrapper; |
| 240 | const int32x4_t mean_vec = vdup_n(mean, wrapper::traits::vector_128_tag{}); |
| 241 | |
| 242 | int32_t x = _window_start_x; |
| 243 | for(; x <= _window_end_x && _window_step_x <= (_window_end_x - x); x += _window_step_x) |
| 244 | { |
| 245 | const int16x8_t val = vloadq(input_ptr + x); |
| 246 | int32x4x2_t shifted; |
| 247 | shifted.val[0] = vsub(vshlq_n_s32(vmovl(vgetlow(val)), 10), mean_vec); |
| 248 | shifted.val[1] = vsub(vshlq_n_s32(vmovl(vgethigh(val)), 10), mean_vec); |
| 249 | |
| 250 | int32x4x2_t rescaled = multiply_by_quantized_multiplier_2row(shifted, inv_std_mul, inv_std_shift); |
| 251 | |
| 252 | const int16x8_t weight_val = vloadq(weight_ptr + x); |
| 253 | const int32x4_t weight_low = vmovl(vgetlow(weight_val)); |
| 254 | const int32x4_t weight_high = vmovl(vgethigh(weight_val)); |
| 255 | |
| 256 | const int32x4_t bias_low = vloadq(bias_ptr + x); |
| 257 | const int32x4_t bias_high = vloadq(bias_ptr + 4 + x); |
| 258 | |
| 259 | int64x2x2_t result_0 = mul_add(rescaled.val[0], weight_low, bias_low); |
| 260 | int64x2x2_t result_1 = mul_add(rescaled.val[1], weight_high, bias_high); |
| 261 | |
| 262 | int32x4x2_t combined; |
| 263 | combined.val[0] = vcombine(vmovn(vrshrq_n_s64(result_0.val[0], 10)), vmovn(vrshrq_n_s64(result_0.val[1], 10))); |
| 264 | combined.val[1] = vcombine(vmovn(vrshrq_n_s64(result_1.val[0], 10)), vmovn(vrshrq_n_s64(result_1.val[1], 10))); |
| 265 | |
| 266 | int32x4x2_t out_val = multiply_by_quantized_multiplier_2row(combined, _output_multiplier, _output_shift + 12); |
| 267 | |
| 268 | vstore(output_ptr + x, vqmovn(out_val.val[0])); |
| 269 | vstore(output_ptr + x + 4, vqmovn(out_val.val[1])); |
| 270 | } |
| 271 | |
| 272 | for(; x < _window_end_x; ++x) |
| 273 | { |
| 274 | const auto val = static_cast<int32_t>(input_ptr[x]); |
| 275 | const int32_t shifted = (val << 10) - mean; |
| 276 | const int32_t rescaled = quantization::multiply_by_quantized_multiplier(shifted, inv_std_mul, inv_std_shift); |
| 277 | const int64_t weighted = rescaled * weight_ptr[x] + bias_ptr[x]; |
| 278 | const auto reverse_shifted = static_cast<int32_t>((weighted + 512) >> 10); |
| 279 | int32_t out_val = quantization::multiply_by_quantized_multiplier(reverse_shifted, _output_multiplier, _output_shift + 12); |
| 280 | out_val = utility::clamp<decltype(out_val), OutputDataType>(out_val, std::numeric_limits<OutputDataType>::min()); |
| 281 | output_ptr[x] = static_cast<OutputDataType>(out_val); |
| 282 | } |
| 283 | } |
| 284 | |
| 285 | void NEQLSTMLayerNormalizationKernel::compute_qsymm16() |
| 286 | { |
| 287 | using InputDataType = int16_t; |
| 288 | using OutputDataType = int16_t; |
| 289 | using BiasDataType = int32_t; |
| 290 | using AccType = int64_t; |
| 291 | |
| 292 | Iterator input_iterator{ _input, _inout_window }; |
| 293 | Iterator output_iterator{ _output, _inout_window }; |
| 294 | Iterator weight_iterator{ _weight, _weight_window }; |
| 295 | Iterator bias_iterator{ _bias, _weight_window }; |
| 296 | |
| 297 | const auto weight_ptr = reinterpret_cast<const InputDataType *>(weight_iterator.ptr()); |
| 298 | const auto bias_ptr = reinterpret_cast<const BiasDataType *>(bias_iterator.ptr()); |
| 299 | |
| 300 | const uint32_t column_size = _input->info()->tensor_shape()[0]; |
| 301 | |
| 302 | execute_window_loop(_inout_window, [ &, this](const Coordinates &) |
| 303 | { |
| 304 | const auto in_ptr = reinterpret_cast<const InputDataType *>(input_iterator.ptr()); |
| 305 | auto out_ptr = reinterpret_cast<OutputDataType *>(output_iterator.ptr()); |
| 306 | |
| 307 | AccType sum{ 0 }; |
| 308 | AccType sum_sq{ 0 }; |
| 309 | std::tie(sum, sum_sq) = sum_qsymm16(in_ptr); |
| 310 | |
| 311 | AccType mean{ 0 }; |
| 312 | AccType variance{ 0 }; |
| 313 | std::tie(mean, variance) = compute_mean_variance(sum, sum_sq, column_size); |
| 314 | |
| 315 | int32_t stddev_invsqrt_mul{}; |
| 316 | int32_t stddev_invsqrt_shift{}; |
| 317 | quantization::get_invsqrt_quantized_multiplier_exp(static_cast<int32_t>(variance), -1, stddev_invsqrt_mul, stddev_invsqrt_shift); |
| 318 | |
| 319 | normalize_qasymm16(in_ptr, out_ptr, weight_ptr, bias_ptr, mean, stddev_invsqrt_mul, stddev_invsqrt_shift); |
| 320 | }, |
| 321 | input_iterator, output_iterator); |
| 322 | } |
| 323 | } // namespace arm_compute |