1a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// Copyright 2015 The Gemmlowp Authors. All Rights Reserved. 27b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 37b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// Licensed under the Apache License, Version 2.0 (the "License"); 47b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// you may not use this file except in compliance with the License. 57b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// You may obtain a copy of the License at 67b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 77b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// http://www.apache.org/licenses/LICENSE-2.0 87b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 97b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// Unless required by applicable law or agreed to in writing, software 107b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// distributed under the License is distributed on an "AS IS" BASIS, 117b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 127b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// See the License for the specific language governing permissions and 137b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// limitations under the License. 147b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 157b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// output_stages.h: public definitions of the output stages that can 167b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// be assembled into an output pipeline, to control how internal 177b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 32-bit accumulators are transformed to obtain the final uint8 187b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// result matrix entries. 197b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 207b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang#ifndef GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_ 217b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang#define GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_ 227b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 237b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang#include <tuple> 247b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 257b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang#include "../internal/common.h" 267b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 277b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangnamespace gemmlowp { 287b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 297b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This output stage takes int32 values and returns still int32 values, 307b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// but "quantized down" to the uint8 scale; in other words, its output 317b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// is typically what one would then clamp to [0..255] and cast to uint8 327b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// (see OutputStageSaturatingCastToUint8). 337b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 347b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This "quantization down" process depends on 3 parameters, 357b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// result_offset, result_mult_int, result_shift, 367b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// and the result is: 377b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// ((input + result_offset) * result_mult_int + rounding) >> result_shift 387b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// where 397b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1)); 407b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageQuantizeDownInt32ToUint8Scale { 417b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_offset; 427b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_mult_int; 437b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_shift; 447b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang}; 457b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 467b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This output stage takes int32 values and returns still int32 values, 477b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// but "quantized down" to the uint8 scale; in other words, its output 487b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// is typically what one would then clamp to [0..255] and cast to uint8 497b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// (see OutputStageSaturatingCastToUint8). 507b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 517b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This "quantization down" process depends on 3 parameters, 527b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// result_offset, result_mult_int, result_shift, 537b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// and the result is: 547b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// ((input + result_offset) * result_mult_int + rounding) >> result_shift 557b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// where 567b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1)); 577b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// 587b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// Difference from OutputStageQuantizeDownInt32ToUint8Scale here is that each 597b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// row or column of the output (depending on tShape) has its own result_offset 607b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// and result_mult_int numbers. 617b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangtemplate <VectorShape tShape> 627b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageQuantizeDownInt32ToUint8ScalePC { 637b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang VectorMap<const std::int32_t, tShape> result_offset; 647b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang VectorMap<const std::int32_t, tShape> result_mult_int; 657b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_shift; 667b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang}; 677b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 68a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// This output stage takes int32 values and returns still int32 values, 697d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// but "quantized down" to a difference scale; for example, in a pipeline 707d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// that outputs uint8 values in [0..255], the output of this stage would be 717d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// int32 values ready to be clamped to [0..255] and casted to uint8 72a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// (see OutputStageSaturatingCastToUint8). 73a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// 74a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// This "quantization down" process depends on 3 parameters, 75a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// result_offset, result_fixedpoint_multiplier, result_shift, 76a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// and the result is: 77a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// ((FixedPointMul(input, result_fixedpoint_multiplier) + 78a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// rounding) >> result_shift) + result_offset_after_shift 79a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// where 80a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1)); 81a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// and where FixedPointMul(x, y) is the nearest integer to the following 82a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// mathematical expression, evaluated without overflow or intermediate 83a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// rounding: 84a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// (x * y) / 2^31 85a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// In practice, it is expected that FixedPointMul will be implemented 86a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// using hardware "rounding doubling int32 multiply high" instructions, 87a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// such as VQRDMULH on ARM. See in fixedpoint.h the generic function, 88a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// SaturatingRoundingDoublingHighMul. 89a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// 90a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// Notice that the other difference from 91a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// OutputStageQuantizeDownInt32ToUint8Scale is that the result offset 92a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// is applied after the multiplier and shift, not before. This ensures 93a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// that no matter what the multiplier and shift are, the result offset 94a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// is effectively integral: offsetting the final result by an integer. 95a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// The motivation for this is to faithfully support quantization schemes 96a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// where the formula linking quantized values to the real mathematical 97a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// values that they represent, is of the form 98a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// 99a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// real_value = scale * (quantized_value - zero_point) 100a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// 101a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// where scale is a real number (represented in quantized form by 102a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// result_fixedpoint_multiplier and result_shift) and zero_point 103a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// is an integer telling which quantized value correspond to the 104a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// real value 0, and is represented here by (the opposite of) 105a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// result_offset_after_shift. 106a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// The motivation for such a quantization scheme, designed to 107a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// ensure that 0 is always a representable value, is that in 108a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// many applications, we need to 0-pad arrays and that can only be 109a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// done for quantized arrays if 0 is a representable value in 110a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// quantized form. In particular, convolution-like operations 111a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// are often implemented using 0-padding, or "im2col"-like 112a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// expansions that implicitly rely on 0-padding. If 0 were not 113a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// a representable value, such operations would have to pad 114a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang// using a nonzero value, introducing bias in the computation. 1157d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wangstruct OutputStageQuantizeDownInt32ByFixedPoint { 116a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang std::int32_t result_fixedpoint_multiplier; 117a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang std::int32_t result_shift; 118a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang std::int32_t result_offset_after_shift; 119a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang}; 120a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang 1217d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint is the old deprecated 1227d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// name of OutputStageQuantizeDownInt32ByFixedPoint, before we noticed that 1237d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// there really wasn't anything Uint8-specific about it. 1247d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wangusing OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint = OutputStageQuantizeDownInt32ByFixedPoint; 1257d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang 1267d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// Variant of OutputStageQuantizeDownInt32ByFixedPoint where the 'shift' 1277d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// is not necessarily just a right shift, so we can represent multipliers 1287d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// greater than 1. This takes an result_exponent parameter; when it's 1297d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// <= 0, this is equivalent to OutputStageQuantizeDownInt32ByFixedPoint 1307d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// with result_shift = -result_exponent. 1317d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// In the general case, this consists in first left-shifting by 1327d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// std::max(result_exponent, 0), before doing the same as 1337d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// OutputStageQuantizeDownInt32ByFixedPoint with 1347d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// result_shift = std::max(-result_exponent, 0). 1357d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wangstruct OutputStageScaleInt32ByFixedPointAndExponent { 1367d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang std::int32_t result_fixedpoint_multiplier; 1377d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang std::int32_t result_exponent; 1387d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang std::int32_t result_offset_after_shift; 1397d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang}; 1407d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang 1417b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This output stage takes int32 values that are expected to be already 1427b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// on the final uint8 scale, but not necessarily in the [0..255] range. 1437b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// It clamps them to the [0..255] range and returns them casted to uint8. 1447b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageSaturatingCastToUint8 {}; 1457b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 1467d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// This output stage takes int32 values that are expected to be already 1477d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// on the final int16 scale, but not necessarily in the [-32768..32767] range. 1487d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang// It clamps them to the [-32768..32767] range and returns them casted to int16. 1497d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wangstruct OutputStageSaturatingCastToInt16 {}; 1507d0d5a611e629e7c8946e6720baa6846ade9f015Miao Wang 1517b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This output stage depends on a "bias vector" that should contain int32 1527b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// entries, and be either a row-vector of the same number of columns as the 1537b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// result matrix, or a column-vector of the same number of rows as the 1547b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// result matrix. This output stage takes int32 values and adds to them 1557b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// the corresponding entry of the bias vector (broadcasted in the other 1567b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// direction to fit the matrix's shape), outputting int32 values. 1577b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangtemplate <typename VectorType> 1587b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageBiasAddition { 1597b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang VectorType bias_vector; 1607b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang}; 1617b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 1627b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This output stage clamps value between the specified min and max bounds. 1637b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// It can be used to implement "rectified linear unit" activation functions 1647b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// in neural networks. 1657b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageClamp { 1667b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t min; 1677b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t max; 1687b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang}; 1697b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 1707b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangstruct OutputStageTanh { 1717b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t real_zero_as_int32; 1727b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t real_amplitude_as_int32; 1737b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang}; 1747b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 1757b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// An output pipeline is just a std::tuple of output stages. 1767b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This function generates a standard output pipeline consisting of two stages: 1777b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// OutputStageQuantizeDownInt32ToUint8Scale, OutputStageSaturatingCastToUint8. 1787b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wanginline std::tuple<OutputStageQuantizeDownInt32ToUint8Scale, 1797b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageSaturatingCastToUint8> 1807b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao WangMakeStandardOutputPipeline(std::int32_t result_offset, 1817b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_mult_int, 1827b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang std::int32_t result_shift) { 1837b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageQuantizeDownInt32ToUint8Scale quantize_down_stage; 1847b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_offset = result_offset; 1857b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_mult_int = result_mult_int; 1867b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_shift = result_shift; 1877b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageSaturatingCastToUint8 saturating_cast_stage; 1887b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang return std::make_tuple(quantize_down_stage, saturating_cast_stage); 1897b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang} 1907b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 1917b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// An output pipeline is just a std::tuple of output stages. 1927b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// This function generates a standard output pipeline consisting of two stages: 1937b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang// OutputStageQuantizeDownInt32ToUint8ScalePC, OutputStageSaturatingCastToUint8. 1947b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wangtemplate <VectorShape tShape> 1957b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wanginline std::tuple<OutputStageQuantizeDownInt32ToUint8ScalePC<tShape>, 1967b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageSaturatingCastToUint8> 197a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao WangMakeStandardOutputPipeline( 198a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang const VectorMap<const std::int32_t, tShape>& result_offset, 199a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang const VectorMap<const std::int32_t, tShape>& result_mult_int, 200a9fd919a0080e2c3c7ed1ce451c85a4d86f2f8c1Miao Wang std::int32_t result_shift) { 2017b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageQuantizeDownInt32ToUint8ScalePC<tShape> quantize_down_stage; 2027b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_offset = result_offset; 2037b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_mult_int = result_mult_int; 2047b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang quantize_down_stage.result_shift = result_shift; 2057b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang OutputStageSaturatingCastToUint8 saturating_cast_stage; 2067b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang return std::make_tuple(quantize_down_stage, saturating_cast_stage); 2077b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang} 2087b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 2097b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang} // namespace gemmlowp 2107b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang 2117b05d573cf2e0fd3a58e98cdbfc65153a83fd6f1Miao Wang#endif // GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_ 212