squiid/llvm
1
0
Fork 0
Code Issues Pull requests Activity

[libc] Use nearest_integer instructions to improve expf performance.

Browse source 
Use nearest_integer instructions to improve expf performance.

Performance tests with CORE-MATH's perf tool:

Before the patch:
```
$ ./perf.sh expf
LIBC-location: /home/lnt/experiment/llvm-project/build/projects/libc/lib/libllvmlibc.a
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH reciprocal throughput   : 9.860
System LIBC reciprocal throughput : 7.728
LIBC reciprocal throughput        : 12.363

$ ./perf.sh expf --latency
LIBC-location: /home/lnt/experiment/llvm-project/build/projects/libc/lib/libllvmlibc.a
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH latency   : 42.802
System LIBC latency : 35.941
LIBC latency        : 49.808
```

After the patch:
```
$ ./perf.sh expf
LIBC-location: /home/lnt/experiment/llvm/llvm-project/build/projects/libc/lib/libllvmlibc.a
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH reciprocal throughput   : 9.441
System LIBC reciprocal throughput : 7.382
LIBC reciprocal throughput        : 8.843

$ ./perf.sh expf --latency
LIBC-location: /home/lnt/experiment/llvm/llvm-project/build/projects/libc/lib/libllvmlibc.a
GNU libc version: 2.31
GNU libc release: stable
CORE-MATH latency   : 44.192
System LIBC latency : 37.693
LIBC latency        : 44.145
```

Reviewed By: zimmermann6

Differential Revision: https://reviews.llvm.org/D130498
This commit is contained in:
Tue Ly 2022-07-25 12:24:31 -04:00
parent 10a7c5b798
commit 91ee672062
3 changed files with 8 additions and 8 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
libc/docs/math.rst
View file

@ -199,7 +199,7 @@ Performance
+==============+===========+===================+===========+===================+=====================================+============+=========================+==============+===============+
| cosf | 37 | 32 | 73 | 72 | :math:`[0, 2\pi]` | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | |
+--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
| expf | 14 | 9 | 58 | 42 | :math:`[-10, 10]` | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA |
| expf | 9 | 7 | 44 | 38 | :math:`[-10, 10]` | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA |
+--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+
| exp2f | 25 | 8 | 81 | 37 | :math:`[-10, 10]` | Ryzen 1700 | Ubuntu 20.04 LTS x86_64 | Clang 12.0.0 | FMA |
+--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+

2
libc/src/math/generic/CMakeLists.txt
View file

@ -483,6 +483,8 @@ add_entrypoint_object(
DEPENDS
.common_constants
libc.src.__support.FPUtil.fputil
libc.src.__support.FPUtil.multiply_add
libc.src.__support.FPUtil.nearest_integer
libc.src.__support.FPUtil.polyeval
libc.include.math
COMPILE_OPTIONS

12
libc/src/math/generic/expf.cpp
View file

@ -10,9 +10,10 @@
#include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2.
#include "src/__support/FPUtil/BasicOperations.h"
#include "src/__support/FPUtil/FEnvImpl.h"
#include "src/__support/FPUtil/FMA.h"
#include "src/__support/FPUtil/FPBits.h"
#include "src/__support/FPUtil/PolyEval.h"
#include "src/__support/FPUtil/multiply_add.h"
#include "src/__support/FPUtil/nearest_integer.h"
#include "src/__support/common.h"
#include <errno.h>
@ -80,13 +81,10 @@ LLVM_LIBC_FUNCTION(float, expf, (float x)) {
// generated by Sollya.
// x_hi = (hi + mid) * 2^7 = round(x * 2^7).
// The default rounding mode for float-to-int conversion in C++ is
// round-toward-zero. To make it round-to-nearest, we add (-1)^sign(x) * 0.5
// before conversion.
int x_hi = static_cast<int>(x * 0x1.0p7f + (xbits.get_sign() ? -0.5f : 0.5f));
float kf = fputil::nearest_integer(x * 0x1.0p7f);
// Subtract (hi + mid) from x to get lo.
x -= static_cast<float>(x_hi) * 0x1.0p-7f;
double xd = static_cast<double>(x);
double xd = static_cast<double>(fputil::multiply_add(kf, -0x1.0p-7f, x));
int x_hi = static_cast<int>(kf);
x_hi += 104 << 7;
// hi = x_hi >> 7
double exp_hi = EXP_M1[x_hi >> 7];