It seems *sometimes* with aggressive optimizations the combination
`psub(padd(a, b), b)` trick to force rounding is compiled away. Here
we replace with inline assembly to prevent this (I tried `volatile`,
but that leads to additional loads from memory).
Also fixed an edge case for large inputs `a` where adding `b` bumps
the value up a power of two and ends up rounding away more than
just the fractional part. If we are over `2^digits` then just return
the input. This edge case was missed in the test since the test was
comparing approximate equality, which was still satisfied. Adding
a strict equality option catches it.
In SSE, by adding/subtracting 2^MantissaBits, we force rounding according to the
current rounding mode.
For NEON, we use the provided intrinsics for rint/floor/ceil if
available (armv8).
Related to #1969.
The original will saturate if the input does not fit into an integer
type. Here we fix this, returning the input if it doesn't have
enough precision to have a fractional part.
Also added `pceil` for NEON.
Fixes#1969.
The previous implementations produced garbage values if the exponent did
not fit within the exponent bits. See #2131 for a complete discussion,
and !375 for other possible implementations.
Here we implement the 4-factor version. See `pldexp_impl` in
`GenericPacketMathFunctions.h` for a full description.
The SSE `pcmp*` methods were moved down since `pcmp_le<Packet4i>`
requires `por`.
Left as a "TODO" is to delegate to a faster version if we know the
exponent does fit within the exponent bits.
Fixes#2131.
I ran some testing (comparing to `std::pow(double(x), double(y)))` for `x` in the set of all (positive) floats in the interval `[std::sqrt(std::numeric_limits<float>::min()), std::sqrt(std::numeric_limits<float>::max())]`, and `y` in `{2, sqrt(2), -sqrt(2)}` I get the following error statistics:
```
max_rel_error = 8.34405e-07
rms_rel_error = 2.76654e-07
```
If I widen the range to all normal float I see lower accuracy for arguments where the result is subnormal, e.g. for `y = sqrt(2)`:
```
max_rel_error = 0.666667
rms = 6.8727e-05
count = 1335165689
argmax = 2.56049e-32, 2.10195e-45 != 1.4013e-45
```
which seems reasonable, since these results are subnormals with only couple of significant bits left.
The `half_float` test was failing with `-mcpu=cortex-a55` (native `__fp16`) due
to a bad NaN bit-pattern comparison (in the case of casting a float to `__fp16`,
the signaling `NaN` is quieted). There was also an inconsistency between
`numeric_limits<half>::quiet_NaN()` and `NumTraits::quiet_NaN()`. Here we
correct the inconsistency and compare NaNs according to the IEEE 754
definition.
Also modified the `bfloat16_float` test to match.
Tested with `cortex-a53` and `cortex-a55`.
This fixes some gcc warnings such as:
```
Eigen/src/Core/GenericPacketMath.h:655:63: warning: implicit conversion turns floating-point number into bool: 'typename __gnu_cxx::__enable_if<__is_integer<bool>::__value, double>::__type' (aka 'double') to 'bool' [-Wimplicit-conversion-floating-point-to-bool]
Packet psqrt(const Packet& a) { EIGEN_USING_STD(sqrt); return sqrt(a); }
```
Details:
- Added `scalar_sqrt_op<bool>` (`-Wimplicit-conversion-floating-point-to-bool`).
- Added `scalar_square_op<bool>` and `scalar_cube_op<bool>`
specializations (`-Wint-in-bool-context`)
- Deprecated above specialized ops for bool.
- Modified `cxx11_tensor_block_eval` to specialize generator for
booleans (`-Wint-in-bool-context`) and to use `abs` instead of `square` to
avoid deprecated bool ops.
This change also contains a few minor cleanups:
1. Remove packet op pnot, which is not needed for anything other than pcmp_le_or_nan,
which can be done in other ways.
2. Remove the "HasInsert" enum, which is no longer needed since we removed the
corresponding packet ops.
3. Add faster pselect op for Packet4i when SSE4.1 is supported.
Among other things, this makes the fast transposeInPlace() method available for Matrix<bool>.
Run on ************** (72 X 2994 MHz CPUs); 2020-05-09T10:51:02.372347913-07:00
CPU: Intel Skylake Xeon with HyperThreading (36 cores) dL1:32KB dL2:1024KB dL3:24MB
Benchmark Time(ns) CPU(ns) Iterations
-----------------------------------------------------------------------
BM_TransposeInPlace<float>/4 9.77 9.77 71670320
BM_TransposeInPlace<float>/8 21.9 21.9 31929525
BM_TransposeInPlace<float>/16 66.6 66.6 10000000
BM_TransposeInPlace<float>/32 243 243 2879561
BM_TransposeInPlace<float>/59 844 844 829767
BM_TransposeInPlace<float>/64 933 933 750567
BM_TransposeInPlace<float>/128 3944 3945 177405
BM_TransposeInPlace<float>/256 16853 16853 41457
BM_TransposeInPlace<float>/512 204952 204968 3448
BM_TransposeInPlace<float>/1k 1053889 1053861 664
BM_TransposeInPlace<bool>/4 14.4 14.4 48637301
BM_TransposeInPlace<bool>/8 36.0 36.0 19370222
BM_TransposeInPlace<bool>/16 31.5 31.5 22178902
BM_TransposeInPlace<bool>/32 111 111 6272048
BM_TransposeInPlace<bool>/59 626 626 1000000
BM_TransposeInPlace<bool>/64 428 428 1632689
BM_TransposeInPlace<bool>/128 1677 1677 417377
BM_TransposeInPlace<bool>/256 7126 7126 96264
BM_TransposeInPlace<bool>/512 29021 29024 24165
BM_TransposeInPlace<bool>/1k 116321 116330 6068
This will allow us to define multiple packet types backed by the same vector type, e.g., __m128i.
Use this machanism to define packets for half and clean up the packet op implementations.
This provides a new op that matches std::rint and previous behavior of
pround. Also adds corresponding unsupported/../Tensor op.
Performance is the same as e. g. floor (tested SSE/AVX).
This change re-instates the fast rational approximation of the logistic function for float32 in Eigen (removed in 66f07efeae), but uses the more accurate approximation 1/(1+exp(-1)) ~= exp(x) below -9. The exponential is only calculated on the vectorized path if at least one element in the SIMD input vector is less than -9.
This change also contains a few improvements to speed up the original float specialization of logistic:
- Introduce EIGEN_PREDICT_{FALSE,TRUE} for __builtin_predict and use it to predict that the logistic-only path is most likely (~2-3% speedup for the common case).
- Carefully set the upper clipping point to the smallest x where the approximation evaluates to exactly 1. This saves the explicit clamping of the output (~7% speedup).
The increased accuracy for tanh comes at a cost of 10-20% depending on instruction set.
The benchmarks below repeated calls
u = v.logistic() (u = v.tanh(), respectively)
where u and v are of type Eigen::ArrayXf, have length 8k, and v contains random numbers in [-1,1].
Benchmark numbers for logistic:
Before:
Benchmark Time(ns) CPU(ns) Iterations
-----------------------------------------------------------------
SSE
BM_eigen_logistic_float 4467 4468 155835 model_time: 4827
AVX
BM_eigen_logistic_float 2347 2347 299135 model_time: 2926
AVX+FMA
BM_eigen_logistic_float 1467 1467 476143 model_time: 2926
AVX512
BM_eigen_logistic_float 805 805 858696 model_time: 1463
After:
Benchmark Time(ns) CPU(ns) Iterations
-----------------------------------------------------------------
SSE
BM_eigen_logistic_float 2589 2590 270264 model_time: 4827
AVX
BM_eigen_logistic_float 1428 1428 489265 model_time: 2926
AVX+FMA
BM_eigen_logistic_float 1059 1059 662255 model_time: 2926
AVX512
BM_eigen_logistic_float 673 673 1000000 model_time: 1463
Benchmark numbers for tanh:
Before:
Benchmark Time(ns) CPU(ns) Iterations
-----------------------------------------------------------------
SSE
BM_eigen_tanh_float 2391 2391 292624 model_time: 4242
AVX
BM_eigen_tanh_float 1256 1256 554662 model_time: 2633
AVX+FMA
BM_eigen_tanh_float 823 823 866267 model_time: 1609
AVX512
BM_eigen_tanh_float 443 443 1578999 model_time: 805
After:
Benchmark Time(ns) CPU(ns) Iterations
-----------------------------------------------------------------
SSE
BM_eigen_tanh_float 2588 2588 273531 model_time: 4242
AVX
BM_eigen_tanh_float 1536 1536 452321 model_time: 2633
AVX+FMA
BM_eigen_tanh_float 1007 1007 694681 model_time: 1609
AVX512
BM_eigen_tanh_float 471 471 1472178 model_time: 805
This also adds pset1frombits helper to Packet[24]d.
Makes round ~45% slower for SSE: 1.65µs ± 1% before vs 2.45µs ± 2% after,
stil an order of magnitude faster than scalar version: 33.8µs ± 2%.
2. Simplify handling of special cases by taking advantage of the fact that the
builtin vrsqrt approximation handles negative, zero and +inf arguments correctly.
This speeds up the SSE and AVX implementations by ~20%.
3. Make the Newton-Raphson formula used for rsqrt more numerically robust:
Before: y = y * (1.5 - x/2 * y^2)
After: y = y * (1.5 - y * (x/2) * y)
Forming y^2 can overflow for very large or very small (denormalized) values of x, while x*y ~= 1. For AVX512, this makes it possible to compute accurate results for denormal inputs down to ~1e-42 in single precision.
4. Add a faster double precision implementation for Knights Landing using the vrsqrt28 instruction and a single Newton-Raphson iteration.
Benchmark results: https://bitbucket.org/snippets/rmlarsen/5LBq9o
