faac: use fast math to save power (#1091)

* makefile: do not pad empty extras, flash_eraseall in sysupgrade cleans entire partition

* package/faac: update to edc0246

Update faac from 64fc541 to edc0246

Hash change: 64fc54142912088eecabaa159339768eca30f3d6 -> edc02462e529b23539d932681fc0f96120e4708e

Changelog:

  33558e6: CI: add native MSVC build and fix Windows compatibility (#66)
  081d81d: remove dead code (#65)
  0622a3a: optimize FFT reordering by combining real and imaginary passes (#68)
  a5abf5e: optimize FFT performance (#70)
  07fb0b4: ci: update GitHub Actions and add Dependabot configuration (#74)
  0a01417: Bump the actions group with 2 updates (#75)
  d3a9c22: remove redundant defines (#76)
  4f74d18: make faacgui usable again (#77)
  c5c142d: report library version on version mismatch
  407db12: maintain library SONAME version 0.0.0
  ab58250: Merge branch 'master' of https://github.com/knik0/faac
  c79fd73: build both static and shared library by default
  779a803: properly declare symbol visibility as hidden
  ad7e4f0: refactor quantizer for portable C and extensible SIMD dispatch (#71)
  edc0246: optimize memory allocation and unify MDCT implementation (#73)

* faac: use fast math to save power

---------

Co-authored-by: Paul Philippov <paul@themactep.com>
Co-authored-by: Paul Philippov <themactep@gmail.com>

Author: nschimme

package/faac/0002-fast-math-approx.patch

@@ -0,0 +1,67 @@
+ libfaac/faac_real.h | 31 +++++++++++++++++++++++++++++--
+ libfaac/quantize.c  |  2 +-
+ 2 files changed, 30 insertions(+), 3 deletions(-)
+
+diff --git a/libfaac/faac_real.h b/libfaac/faac_real.h
+index 4d004d7..a2e76fd 100644
+--- a/libfaac/faac_real.h
++++ b/libfaac/faac_real.h
+@@ -25,6 +25,33 @@
+ #endif
+ 
+ #include <math.h>
++#include <stdint.h>
++
++static inline float fast_log2(float x)
++{
++    union { float f; uint32_t i; } vx = { x };
++    float y = (float)vx.i;
++    y *= 1.1920928955078125e-7f;
++    return y - 126.94269504f;
++}
++
++static inline float fast_log10(float x)
++{
++    return fast_log2(x) * 0.3010299956639812f;
++}
++
++static inline float fast_pow2(float x)
++{
++    if (x < -126.0f) return 0.0f;
++    if (x > 128.0f) return 3e38f;
++    union { uint32_t i; float f; } v = { (uint32_t)((x + 126.94269504f) * 8388608.0f) };
++    return v.f;
++}
++
++static inline float fast_pow(float x, float y)
++{
++    return fast_pow2(y * fast_log2(x));
++}
+ 
+ #ifdef FAAC_PRECISION_SINGLE
+ typedef float faac_real;
+@@ -32,8 +59,8 @@ typedef float faac_real;
+ #define FAAC_COS cosf
+ #define FAAC_SQRT sqrtf
+ #define FAAC_FABS fabsf
+-#define FAAC_LOG10 log10f
+-#define FAAC_POW powf
++#define FAAC_LOG10  fast_log10
++#define FAAC_POW    fast_pow
+ #define FAAC_ASIN asinf
+ #define FAAC_LRINT lrintf
+ #define FAAC_FLOOR floorf
+diff --git a/libfaac/quantize.c b/libfaac/quantize.c
+index 27d3a79..0f67c9e 100644
+--- a/libfaac/quantize.c
++++ b/libfaac/quantize.c
+@@ -171,7 +171,7 @@ static void qlevel(CoderInfo * __restrict coderInfo,
+                   )
+ {
+     int sb;
+-#if !defined(__clang__) && defined(__GNUC__) && (GCC_VERSION >= 40600)
++#if !defined(__clang__) && defined(__GNUC__) && (GCC_VERSION >= 40600) && false
+     /* 2^0.25 (1.50515 dB) step from AAC specs */
+     static const faac_real sfstep = 1.0 / FAAC_LOG10(FAAC_SQRT(FAAC_SQRT(2.0)));
+ #else
+

package/faac/faac.mk

@@ -11,6 +11,8 @@ FAAC_DEPENDENCIES = host-pkgconf
 FAAC_INSTALL_STAGING = YES
 FAAC_INSTALL_TARGET = YES
 
+FAAC_CONF_ENV += CFLAGS="$(TARGET_CFLAGS) -ffast-math"
+
 FAAC_CONF_OPTS = \
 	$(if $(BR2_PACKAGE_FAAC_ENABLE_DRM),-Ddrm=true,-Ddrm=false) \
 	-Dfloating-point=single