[libc-commits] [libc] 91ee672 - [libc] Use nearest_integer instructions to improve expf performance.

[Tue Ly via libc-commits]

Author: Tue Ly
Date: 2022-07-26T09:11:27-04:00
New Revision: 91ee67206289ec80c3c02f063040fa4638be0d63

URL: https://github.com/llvm/llvm-project/commit/91ee67206289ec80c3c02f063040fa4638be0d63
DIFF: https://github.com/llvm/llvm-project/commit/91ee67206289ec80c3c02f063040fa4638be0d63.diff

LOG: [libc] Use nearest_integer instructions to improve expf performance.

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

Added: 
    

Modified: 
    libc/docs/math.rst
    libc/src/math/generic/CMakeLists.txt
    libc/src/math/generic/expf.cpp

Removed: 
    


################################################################################
diff  --git a/libc/docs/math.rst b/libc/docs/math.rst
index 6480a981acb02..42d09617d6bb7 100644
--- a/libc/docs/math.rst
+++ b/libc/docs/math.rst
@@ -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           |
 +--------------+-----------+-------------------+-----------+-------------------+-------------------------------------+------------+-------------------------+--------------+---------------+

diff  --git a/libc/src/math/generic/CMakeLists.txt b/libc/src/math/generic/CMakeLists.txt
index 7b6a378b4616c..2658b9691d221 100644
--- a/libc/src/math/generic/CMakeLists.txt
+++ b/libc/src/math/generic/CMakeLists.txt
@@ -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

diff  --git a/libc/src/math/generic/expf.cpp b/libc/src/math/generic/expf.cpp
index 948b101b755eb..1003699cbf5f1 100644
--- a/libc/src/math/generic/expf.cpp
+++ b/libc/src/math/generic/expf.cpp
@@ -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];