We assume familiarity with gemmlowp's low-precision uint8 computation paradigm, which is described in low-precision.md.
This document is about the possibility of further reducing precision below 8 bits.
That allows to get higher arithmetic throughput on some architectures, at the cost of decreased accuracy.
A meta note is needed here as to how this fits with the general gemmlowp design.
Less-than-8-bit computation was initially designed and implemented in gemmlowp as a drop-in replacement for regular 8bit computation, a plain optimization. The idea was that to automatically requantize 8bit operands to less-than-8bit during the O(N^2) packing stage, then take advantage of the lower bit depth during the O(N^3) compute stage. For large enough matrices, that should be worth it.
TODO(benoitjacob): update this documentation. This 'present' state just became the past (February 2017).
At the moment, this less-than-8-bit mode of gemmlowp is not much used in practice, because the implicit requantization of operands from 8bit to less-than-8bit turned out to be more expensive than initially expected, both in terms of speed and accuracy:
-
Speed: the O(N^2) requantization is only negligible compared to the O(N^3) compute kernel when the matrix size N is large enough; in practice, smaller matrix sizes turned out to be very important, making the requantization approach slower than expected.
-
Accuracy: As neural networks were optimized for size, their sensitivity to numerical accuracy increased. Then the approach of requantizing already-quantized data turned out to be more wasteful of accuracy than we could afford.
Less-than-8bit still probably has good prospects; what should be dropped is the requantization. In other words, in the future, we might have neural networkds trained right away for some bit depth lower than 8 bits. The resulting values would probably still be stored as 8 bits (unless the bit depth eventually becomes very low). Thus, no particular work would be needed in the packing stage; no overhead or loss of accuracy would be incurred anymore.
In other words: the design of less-than-8-bit kernels is probably useful in the long run; what is on the way out is requantization and packing/unpacking-stage aspects.
With that said, the rest of this page retains its old content about the present approach:
Accessing less-than-8-bit computation via the EightBitIntGemm is very simple: EightBitIntGemm takes a BitDepthSetting enum which allows to choose among a fixed set of supported bit-depth combinations.
The public/gemmlowp.h interface exposes more extensive control over quantization, by means of a BitDepthParams template parameter, which is a type parameter, carrying information about: 1. The LHS and RHS bit depth, which can be set arbitrarily and independently; 2. The 'RoundingStrategy', which is the heuristic used to choose a rounding mode, based on the accumulation size (a.k.a. the "depth" dimension of the Gemm). Details can be seen in public/bit_depth.h.
Input/output matrix data is all uint8's, ranging from 0 to 255, regardless of the BitDepth{Setting,Params}.
So the BitDepth{Setting,Params} is only an internal detail. It only means to allow gemmlowp to use lower precision internally, but the input/output data format is unaffected.
As far as the API contract goes, the only thing that the BitDepth{Setting,Params} does is to relax the accuracy requirement. With standard 8bit/8bit computation, gemmlowp is required to return the exact result as specified in low-precision.md. With lower bit depths, gemmlowp is no longer required to return an exact result.
Here we refer to the 3 stages of computation as described in design.md, namely: packing, computation kernel, unpacking.
The general idea is that at the packing stage, we requantize input (Lhs/Rhs) data to less-than-8-bit depths by scaling them, thus shrinking the range of the packed matrix entries; for instance, if the Rhs bit depth is to be 5 bits then packed Rhs matrix entries will be in the range [0 ... 31]. This then allows the GEMM kernel to use narrower accumulators without risking overflow, thus achieving higher arithmetic throughput. Finally, at the unpacking stage, it only remains to scale the result values to compensate for the scalings applied earlier.
Let us go into more detail for each of those stages:
The packing stage is where most of the work specific to the BitDepthParams takes place.
Here, we have to scale input matrix values from their original range of [0 ... 255] to the range specified by the BitDepthParams, which is [0 ... (2^N)-1] where N is the number of bits for the matrix at hand (Lhs or Rhs). For example, for a bit depth of 5 bits, we need to scale down to [0 ... 31].
This scaling is what we call "requantization". The pedantic name matches the fact that this is actually quite nontrivial to do correctly i.e. in such a way that the result accuracy will be good enough for real-world applications. See the section below on requantization details.
Concretely, this work happens in PackingRegisterBlock::Pack(), which calls Requantize(). This is in internal/pack.h. This code can be overridden for specific architectures, see internal/pack_neon.h.
This requantization work is costly and makes packing slower. This means that, at least in our approach, less-than-8-bit computation is only interesting for large-enough, square-enough GEMMs where packing is only a small fraction of the overall cost. In cases where packing overhead is more prevalent (highly rectangular cases), less-than-8-bit is probably a waste of time as long as we treat it as an internal computation detail. What might help there, might be if we shrunk the input/output data format to lower memory bandwidth usage.
In principle, the computation kernel stage simply doesn't have to care about the bit depth at all. In fact, on architectures where we do not have specific optimized kernels for less-than-8-bit cases, we simply use our standard kernel there, and that's correct!
However, while the kernel doesn't have to know about the fact that the operands are on less than 8 bits, it can use that information to make special optimizations that would be incorrect in the general 8-bit case and become correct here thanks to the more restricted range of inputs. That's the whole point of this less-than-8-bit computation idea.
With Lhs entries guaranteed to be smaller than 2^N, and Rhs entries guaranteed to be smaller than 2^M, each product is thus guaranteed to be smaller than 2^(M+N). Thus, one may accumulate 2^(16-(M+N)) such products and still be guaranteed that such an accumulator will be smaller than 2^16, and thus can be stored as a uint16 without risking overflow.
For example, in the L7R5 case, the Lhs enties are on 7 bits (N=7) and the Rhs entries are on 5 bits (M=5), so each product fits in 12 bits and one can thus accumulate 16 ( = 2^(16-12)) such products into uint16 accumulators with no risk of overflow.
This means that a computation kernel may use uint16 accumulators for several loop iterations (16 in the above example), provided that it is allowed to assume that inputs are in such restricted range.
After this fixed number of loop iterations, the kernel must accumulate the local uint16 accumulators back into global uint32 accumulators.
On SIMD architectures with suitable uint16 arithmetic, this in principle allows to multiply arithmetic throughput by up to 2x, since twice more accumulators now fit in each SIMD vector register. This is partially offset by the cost of accumulating back into global uint32 accumulators every several loop iterations, but our experience on ARM NEON has been that we still get quite close to a 2x speedup. See internal/kernel_neon.h, specifically NEON32Kernel12x4Depth2Assuming12BitProducts.
At the unpacking stage, it only remains to scale the result values to compensate for the scaling of the inputs. This is easier because now we are expanding the range instead of shrinking it, so we don't need to worry about ways to minimize a loss of accuracy. We simply need to multiply result values by a constant fraction, rounding to nearest.
Since the inputs were scaled by factors of (2^lhs_bits - 1)/255 and (2^rhs_bits - 1)/255 respectively, the scaling of the outputs needs to be by the following factor:
255 * 255
-----------------------------------
(2^lhs_bits - 1) * (2^rhs_bits - 1)
This is done by a MultiplyByConstantFraction function, see internal/unpack.h
Here we go into more detail on the Requantize() function used at the packing stage to requantize input matrix data. See this function in internal/pack.h.
It depends on the bit depth and on a rounding mode, and requantizes an input value in [0 ... 255] to the range [0 ... (2^N)-1] specified by the bit depth N.
Naive and inaccurate ways to achieve this requantization include: 1. By shifting bits rights by (8-N) bits; 2. By multiplying by ((2^N) - 1) and dividing by 255.
Both of those are biased in some way: 1. has the wrong "derivative", since it approximates (((2^N) - 1) / 255) by ((2^N) / 256) ; 2. has bias since it effectively implements rounding towards 0.
In practice, both of the above requantization functions give results that are too inaccurate in practice for the neural network that we tried (GoogLeNet).
The simplest fix is to avoid the bias in 2. by rounding-to-nearest instead of rounding towards 0. This can be achieved by doing
dst = (src * maxval + rounding_offset) / 255;
Where maxval = ((2^N) - 1) is the highest requantized value, and the rounding_offset can be set to
rounding_offset = 127
to achieve rounding-to-nearest (while the above rounding towards 0 corresponded to rounding_offset = 0).
In principle, rounding-to-nearest is unbiased and optimal in various ways.
In practice though, our input data is not random real numbers, but already-quantized 8-bit values. That means that even in the best case, there would be at most 255 different possible input values; in practice, we generally see the input values distributed non-uniformly in that range, so that a majority of input values tend to be in a much smaller range. See test/test_data.cc.
Having a large part of the input values in a very small finite set, means that the corresponding rounding errors are also in a very small finite set, which can be small enough that the mean of these rounding errors is significantly different from 0. That rounding-to-nearest is "unbiased" only means that over a sufficiently large set of input values, the bias would become arbitrarily close to 0; here, the set of input values is effectively small enough that the resulting bias is significant.
This leads to biasing the matrix product entries, resulting in an error that grows linearly with the depth dimension of the GEMM.
To address that, we can instead use probabilistic rounding. The idea is that for instance if we have to round the value 3.8 to the nearest integer, we can round it to 3 with 20% probability and to 4 with probability 80%. If that value 3.8 occurs many times, the mean requantized value will thus tend to 3.8.
This amounts to keeping the above requantization formula,
dst = (src * maxval + rounding_offset) / 255;
but now the rounding_offset is a random value in [0 .. 254].
This guarantees zero bias no matter how small the distribution of input values is.
On the other hand, the variance of the error term here is higher than with rounding-to-nearest --- one can check that it is 2x higher.
So the error term coming from the Central Limit Theorem, which grows with the square root of the accumulator depth i.e. the GEMM depth, will be 2x higher here.
Still, for large enough GEMM depth, that is better than rounding-to-nearest which has an error term growing linearly with the GEMM depth.
Thus, for fixed input values and bit depths, we expect that probabilistic rounding will give more accurate results for large enough GEMM depths, while rounding-to-nearest will be more accurate for smaller GEMM depths.
That is why use switch between these rounding modes based on GEMM depth, see ChooseRoundingMode in internal/bit_depth_util.h.
It is based on a constant, kProbabilisticRoundingThreshold, defined in internal/common.h and empirically determined. See the comment there. It would be nice to better understand the statistics here and come up with better heuristics for this switching.
We provide two PRNGs. The first is an 8-bit Xorshift. It is fast, naturally produces values ranging over an interval of width 255, which is what we need here (as opposed to an interval of width 256), and turns out, from empirical tests, to produce better results than a linear congruential generator (LCG). That's unfortunate, as a 8-bit LCG performs better (we confirmed that on various ARM devices) but we need as perfect un-biased-ness as we can get.
The second is an "add-mod" sequence generator, which generates non-random values in the sequence x = (x + 97) % 255. This generates a low-discrepancy sequence that minimizes the "clumpiness" of the random offsets (Thus, for example, quantizing a 3x3 matrix will have a maximum additive error of about 200 from the random offsets). While not random, this sequence performs well empirically for many quantizations. (For information about why 97 is a good value, see https://en.wikipedia.org/wiki/Low-discrepancy_sequence#Additive_recurrence and http://mollwollfumble.blogspot.com/2011/03/subrandom-numbers.html 97/255 = 0.38; 0.382 is the best choice. For discrete numbers, the choice must be relatively prime to the modulus. 97 is prime, so it is safely relatively prime to 255. 107 is another near-optimal choice.
The low-discrepancy sequence generator is the default.
More details and results are given in a comment on the default PRNG in internal/pack.h. Interested users can change the PRNG used by setting DefaultRoundingGenerator in bit_depth_util.h.