Use hardware-accelerated __fp16 types with Clang

Maratyszcza · Jun 16, 2024 · d63ea15 · d63ea15
1 parent ae992b5
commit d63ea15
Show file tree

Hide file tree

Showing 2 changed files with 34 additions and 0 deletions.
diff --git a/include/fp16/fp16.h b/include/fp16/fp16.h
@@ -15,6 +15,7 @@
 #endif
 
 #include <fp16/bitcasts.h>
+#include <fp16/macros.h>
 
 
 /*
@@ -106,6 +107,13 @@ static inline uint32_t fp16_ieee_to_fp32_bits(uint16_t h) {
  * floating-point operations and bitcasts between integer and floating-point variables.
  */
 static inline float fp16_ieee_to_fp32_value(uint16_t h) {
+#if FP16_USE_FP16_TYPE
+	union {
+		uint16_t as_bits;
+		__fp16 as_value;
+	} fp16 = { h };
+	return (float) fp16.as_value;
+#else
 	/*
 	 * Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
 	 *      +---+-----+------------+-------------------+
@@ -211,6 +219,7 @@ static inline float fp16_ieee_to_fp32_value(uint16_t h) {
 	const uint32_t result = sign |
 		(two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
 	return fp32_from_bits(result);
+#endif
 }
 
 /*
@@ -221,6 +230,13 @@ static inline float fp16_ieee_to_fp32_value(uint16_t h) {
  * floating-point operations and bitcasts between integer and floating-point variables.
  */
 static inline uint16_t fp16_ieee_from_fp32_value(float f) {
+#if FP16_USE_FP16_TYPE
+	union {
+		__fp16 as_value;
+		uint16_t as_bits;
+	} fp16 = { (__fp16) f };
+	return fp16.as_bits;
+#else
 #if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
 	const float scale_to_inf = 0x1.0p+112f;
 	const float scale_to_zero = 0x1.0p-110f;
@@ -249,6 +265,7 @@ static inline uint16_t fp16_ieee_from_fp32_value(float f) {
 	const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
 	const uint32_t nonsign = exp_bits + mantissa_bits;
 	return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+#endif
 }
 
 /*

diff --git a/include/fp16/macros.h b/include/fp16/macros.h
@@ -0,0 +1,17 @@
+#pragma once
+#ifndef FP16_MACROS_H
+#define FP16_MACROS_H
+
+
+#ifndef FP16_USE_FP16_TYPE
+	#if defined(__clang__)
+		#if defined(__F16C__) || defined(__aarch64__)
+			#define FP16_USE_FP16_TYPE 1
+		#endif
+	#endif
+	#if !defined(FP16_USE_FP16_TYPE)
+		#define FP16_USE_FP16_TYPE 0
+	#endif  // !defined(FP16_USE_FP16_TYPE)
+#endif  // !defined(FP16_USE_FP16_TYPE)
+
+#endif /* FP16_MACROS_H */