Message ID | 20240725202522.276182-1-remi@remlab.net |
---|---|
State | New |
Headers | show |
Series | [FFmpeg-devel,1/6] lavu/riscv: implement floating point clips | expand |
Context | Check | Description |
---|---|---|
andriy/make_x86 | success | Make finished |
andriy/make_fate_x86 | success | Make fate finished |
Le 25 juillet 2024 23:25:15 GMT+03:00, "Rémi Denis-Courmont" <remi@remlab.net> a écrit : >Unlike x86, fmin/fmax are single instructions, not function calls. They >are much much faster than doing a comparison, then branching based on its >results. With this, audiodsp.vector_clipf gets almost twice as fast, and >a properly unrollled version of it gets 4-5x faster, on SiFive-U74. >This is only the low-hanging fruit: FFMIN and FFMAX are presumably >affected as well. > >This likely applies to other instruction sets with native IEEE floats, >especially those lacking a conditional select instruction. In fact, the same problem occurs on Armv8, and it gets even worse on Armv7 where FFMIN and FFMAX incur calls to fcmp*(). I am not sure if this really works on anything but x86. Only way to make FFMIN behave as well as math.h functions seems to be enabling -funsafe-math-optimizations -ffinite-math-only. >--- > libavutil/riscv/intmath.h | 19 +++++++++++++++++++ > 1 file changed, 19 insertions(+) > >diff --git a/libavutil/riscv/intmath.h b/libavutil/riscv/intmath.h >index 3e7ab864c5..24f165eef1 100644 >--- a/libavutil/riscv/intmath.h >+++ b/libavutil/riscv/intmath.h >@@ -22,6 +22,7 @@ > #define AVUTIL_RISCV_INTMATH_H > > #include <stdint.h> >+#include <math.h> > > #include "config.h" > #include "libavutil/attributes.h" >@@ -72,6 +73,24 @@ static av_always_inline av_const int av_clip_intp2_rvi(int a, int p) > return b; > } > >+#if defined (__riscv_f) || defined (__riscv_zfinx) >+#define av_clipf av_clipf_rvf >+static av_always_inline av_const float av_clipf_rvf(float a, float min, >+ float max) >+{ >+ return fminf(fmaxf(a, min), max); >+} >+#endif >+ >+#if defined (__riscv_d) || defined (__riscv_zdinx) >+#define av_clipd av_clipd_rvd >+static av_always_inline av_const float av_clipd_rvd(double a, double min, >+ double max) >+{ >+ return fmin(fmax(a, min), max); >+} >+#endif >+ > #if defined (__GNUC__) || defined (__clang__) > static inline av_const int ff_ctz_rv(int x) > {
diff --git a/libavutil/riscv/intmath.h b/libavutil/riscv/intmath.h index 3e7ab864c5..24f165eef1 100644 --- a/libavutil/riscv/intmath.h +++ b/libavutil/riscv/intmath.h @@ -22,6 +22,7 @@ #define AVUTIL_RISCV_INTMATH_H #include <stdint.h> +#include <math.h> #include "config.h" #include "libavutil/attributes.h" @@ -72,6 +73,24 @@ static av_always_inline av_const int av_clip_intp2_rvi(int a, int p) return b; } +#if defined (__riscv_f) || defined (__riscv_zfinx) +#define av_clipf av_clipf_rvf +static av_always_inline av_const float av_clipf_rvf(float a, float min, + float max) +{ + return fminf(fmaxf(a, min), max); +} +#endif + +#if defined (__riscv_d) || defined (__riscv_zdinx) +#define av_clipd av_clipd_rvd +static av_always_inline av_const float av_clipd_rvd(double a, double min, + double max) +{ + return fmin(fmax(a, min), max); +} +#endif + #if defined (__GNUC__) || defined (__clang__) static inline av_const int ff_ctz_rv(int x) {