avutil/half2float: use native _Float16 if available

_Float16 support was available on arm/aarch64 for a while, and with gcc
12 was enabled on x86 as long as SSE2 is supported.

If the target arch supports f16c, gcc emits fairly efficient assembly,
taking advantage of it. This is the case on x86-64-v3 or higher.
Same goes on arm, which has native float16 support.
On x86, without f16c, it emulates it in software using sse2 instructions.

This has shown to perform rather poorly:

_Float16 full SSE2 emulation:
frame=50074 fps=848 q=-0.0 size=N/A time=00:33:22.96 bitrate=N/A speed=33.9x

_Float16 f16c accelerated (Zen2, --cpu=znver2):
frame=50636 fps=1965 q=-0.0 Lsize=N/A time=00:33:45.40 bitrate=N/A speed=78.6x

classic half2float full software implementation:
frame=49926 fps=1605 q=-0.0 Lsize=N/A time=00:33:17.00 bitrate=N/A speed=64.2x

Hence an additional check was introduced, that only enables use of
_Float16 on x86 if f16c is being utilized.

On aarch64, a similar uplift in performance is seen:

RPi4 half2float full software implementation:
frame= 6088 fps=126 q=-0.0 Lsize=N/A time=00:04:03.48 bitrate=N/A speed=5.06x

RPi4 _Float16:
frame= 6103 fps=158 q=-0.0 Lsize=N/A time=00:04:04.08 bitrate=N/A speed=6.32x

Since arm/aarch64 always natively support 16 bit floats, it can always
be considered fast there.

I'm not aware of any additional platforms that currently support
_Float16. And if there are, they should be considered non-fast until
proven fast.

5 files changed, 50 insertions, 0 deletions

diff --git a/configure b/configure
index fe94941a03..ea50c94002 100755
--- a/configure
+++ b/configure
@@ -2145,6 +2145,7 @@ ARCH_FEATURES="
     fast_64bit
     fast_clz
     fast_cmov
+    fast_float16
     local_aligned
     simd_align_16
     simd_align_32
@@ -5127,6 +5128,8 @@ elif enabled arm; then
             ;;
     esac
 
+    test_cflags -mfp16-format=ieee && add_cflags -mfp16-format=ieee
+
 elif enabled avr32; then
 
     case $cpu in
@@ -6231,6 +6234,15 @@ check_builtin sync_val_compare_and_swap "" "int *ptr; int oldval, newval; __sync
 check_builtin gmtime_r time.h "time_t *time; struct tm *tm; gmtime_r(time, tm)"
 check_builtin localtime_r time.h "time_t *time; struct tm *tm; localtime_r(time, tm)"
 
+check_builtin float16 "" "_Float16 f16var"
+if enabled float16; then
+    if enabled x86; then
+        test_cpp_condition stddef.h "defined(__F16C__)" && enable fast_float16
+    elif enabled arm || enabled aarch64; then
+        enable fast_float16
+    fi
+fi
+
 case "$custom_allocator" in
     jemalloc)
         # jemalloc by default does not use a prefix
diff --git a/libavutil/float2half.c b/libavutil/float2half.c
index c79a3abfa1..1a283956e7 100644
--- a/libavutil/float2half.c
+++ b/libavutil/float2half.c
@@ -20,6 +20,7 @@
 
 void ff_init_float2half_tables(Float2HalfTables *t)
 {
+#if !HAVE_FAST_FLOAT16
     for (int i = 0; i < 256; i++) {
         int e = i - 127;
 
@@ -50,4 +51,5 @@ void ff_init_float2half_tables(Float2HalfTables *t)
             t->shifttable[i|0x100] = 13;
         }
     }
+#endif
 }
diff --git a/libavutil/float2half.h b/libavutil/float2half.h
index 20fdc2a36b..e619046911 100644
--- a/libavutil/float2half.h
+++ b/libavutil/float2half.h
@@ -20,21 +20,37 @@
 #define AVUTIL_FLOAT2HALF_H
 
 #include <stdint.h>
+#include "intfloat.h"
+
+#include "config.h"
 
 typedef struct Float2HalfTables {
+#if HAVE_FAST_FLOAT16
+    uint8_t dummy;
+#else
     uint16_t basetable[512];
     uint8_t shifttable[512];
+#endif
 } Float2HalfTables;
 
 void ff_init_float2half_tables(Float2HalfTables *t);
 
 static inline uint16_t float2half(uint32_t f, const Float2HalfTables *t)
 {
+#if HAVE_FAST_FLOAT16
+    union {
+        _Float16 f;
+        uint16_t i;
+    } u;
+    u.f = av_int2float(f);
+    return u.i;
+#else
     uint16_t h;
 
     h = t->basetable[(f >> 23) & 0x1ff] + ((f & 0x007fffff) >> t->shifttable[(f >> 23) & 0x1ff]);
 
     return h;
+#endif
 }
 
 #endif /* AVUTIL_FLOAT2HALF_H */
diff --git a/libavutil/half2float.c b/libavutil/half2float.c
index 1967126f76..4de2180a19 100644
--- a/libavutil/half2float.c
+++ b/libavutil/half2float.c
@@ -18,6 +18,7 @@
 
 #include "libavutil/half2float.h"
 
+#if !HAVE_FAST_FLOAT16
 static uint32_t convertmantissa(uint32_t i)
 {
     int32_t m = i << 13; // Zero pad mantissa bits
@@ -33,9 +34,11 @@ static uint32_t convertmantissa(uint32_t i)
 
     return m | e; // Return combined number
 }
+#endif
 
 void ff_init_half2float_tables(Half2FloatTables *t)
 {
+#if !HAVE_FAST_FLOAT16
     t->mantissatable[0] = 0;
     for (int i = 1; i < 1024; i++)
         t->mantissatable[i] = convertmantissa(i);
@@ -60,4 +63,5 @@ void ff_init_half2float_tables(Half2FloatTables *t)
     t->offsettable[31] = 2048;
     t->offsettable[32] = 0;
     t->offsettable[63] = 2048;
+#endif
 }
diff --git a/libavutil/half2float.h b/libavutil/half2float.h
index 428a27a19f..dbd5e7150f 100644
--- a/libavutil/half2float.h
+++ b/libavutil/half2float.h
@@ -20,22 +20,38 @@
 #define AVUTIL_HALF2FLOAT_H
 
 #include <stdint.h>
+#include "intfloat.h"
+
+#include "config.h"
 
 typedef struct Half2FloatTables {
+#if HAVE_FAST_FLOAT16
+    uint8_t dummy;
+#else
     uint32_t mantissatable[3072];
     uint32_t exponenttable[64];
     uint16_t offsettable[64];
+#endif
 } Half2FloatTables;
 
 void ff_init_half2float_tables(Half2FloatTables *t);
 
 static inline uint32_t half2float(uint16_t h, const Half2FloatTables *t)
 {
+#if HAVE_FAST_FLOAT16
+    union {
+        _Float16 f;
+        uint16_t i;
+    } u;
+    u.i = h;
+    return av_float2int(u.f);
+#else
     uint32_t f;
 
     f = t->mantissatable[t->offsettable[h >> 10] + (h & 0x3ff)] + t->exponenttable[h >> 10];
 
     return f;
+#endif
 }
 
 #endif /* AVUTIL_HALF2FLOAT_H */