Automatic commit by thirdparty_tool: update fp16 to commit 98b0a46bce…

…017382a6351a19577ec43a715b6835.

Used commit of the fp16 repository: https://github.com/Maratyszcza/FP16//commits/98b0a46bce017382a6351a19577ec43a715b6835

src/inline-thirdparty/fp16/fp16.h

@@ -4,8 +4,10 @@
 
 #include <fp16/fp16.h>
 
-#if defined(PSIMD_H)
-#include <fp16/psimd.h>
-#endif
-
 #endif /* FP16_H */
+
+// This file is part of the fp16 inline third-party dependency of YugabyteDB.
+// Git repo: https://github.com/Maratyszcza/FP16/
+// Git commit: 98b0a46bce017382a6351a19577ec43a715b6835
+//
+// See also src/inline-thirdparty/README.md.

src/inline-thirdparty/fp16/fp16/bitcasts.h

@@ -12,7 +12,7 @@
 	#include <immintrin.h>
 #endif
 
-#if defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#if defined(_MSC_VER) && !defined(__clang__) && (defined(_M_ARM) || defined(_M_ARM64))
 	#include <intrin.h>
 #endif
 
@@ -22,9 +22,9 @@ static inline float fp32_from_bits(uint32_t w) {
 	return as_float(w);
 #elif defined(__CUDA_ARCH__)
 	return __uint_as_float((unsigned int) w);
-#elif defined(__INTEL_COMPILER)
+#elif defined(__INTEL_COMPILER) || defined(_MSC_VER) && (_MSC_VER >= 1932) && (defined(_M_IX86) || defined(_M_X64))
 	return _castu32_f32(w);
-#elif defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#elif defined(_MSC_VER) && !defined(__clang__) && (defined(_M_ARM) || defined(_M_ARM64))
 	return _CopyFloatFromInt32((__int32) w);
 #else
 	union {
@@ -40,9 +40,9 @@ static inline uint32_t fp32_to_bits(float f) {
 	return as_uint(f);
 #elif defined(__CUDA_ARCH__)
 	return (uint32_t) __float_as_uint(f);
-#elif defined(__INTEL_COMPILER)
+#elif defined(__INTEL_COMPILER) || defined(_MSC_VER) && (_MSC_VER >= 1932) && (defined(_M_IX86) || defined(_M_X64))
 	return _castf32_u32(f);
-#elif defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#elif defined(_MSC_VER) && !defined(__clang__) && (defined(_M_ARM) || defined(_M_ARM64))
 	return (uint32_t) _CopyInt32FromFloat(f);
 #else
 	union {
@@ -58,9 +58,9 @@ static inline double fp64_from_bits(uint64_t w) {
 	return as_double(w);
 #elif defined(__CUDA_ARCH__)
 	return __longlong_as_double((long long) w);
-#elif defined(__INTEL_COMPILER)
+#elif defined(__INTEL_COMPILER) || defined(_MSC_VER) && (_MSC_VER >= 1932) && (defined(_M_IX86) || defined(_M_X64))
 	return _castu64_f64(w);
-#elif defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#elif defined(_MSC_VER) && !defined(__clang__) && (defined(_M_ARM) || defined(_M_ARM64))
 	return _CopyDoubleFromInt64((__int64) w);
 #else
 	union {
@@ -76,9 +76,9 @@ static inline uint64_t fp64_to_bits(double f) {
 	return as_ulong(f);
 #elif defined(__CUDA_ARCH__)
 	return (uint64_t) __double_as_longlong(f);
-#elif defined(__INTEL_COMPILER)
+#elif defined(__INTEL_COMPILER) || defined(_MSC_VER) && (_MSC_VER >= 1932) && (defined(_M_IX86) || defined(_M_X64))
 	return _castf64_u64(f);
-#elif defined(_MSC_VER) && (defined(_M_ARM) || defined(_M_ARM64))
+#elif defined(_MSC_VER) && !defined(__clang__) && (defined(_M_ARM) || defined(_M_ARM64))
 	return (uint64_t) _CopyInt64FromDouble(f);
 #else
 	union {
@@ -90,3 +90,9 @@ static inline uint64_t fp64_to_bits(double f) {
 }
 
 #endif /* FP16_BITCASTS_H */
+
+// This file is part of the fp16 inline third-party dependency of YugabyteDB.
+// Git repo: https://github.com/Maratyszcza/FP16/
+// Git commit: 98b0a46bce017382a6351a19577ec43a715b6835
+//
+// See also src/inline-thirdparty/README.md.

src/inline-thirdparty/fp16/fp16/fp16.h

@@ -10,11 +10,18 @@
 	#include <math.h>
 #endif
 
-#ifdef _MSC_VER
+#include <fp16/bitcasts.h>
+#include <fp16/macros.h>
+
+#if defined(_MSC_VER)
 	#include <intrin.h>
 #endif
-
-#include <fp16/bitcasts.h>
+#if defined(__F16C__) && FP16_USE_NATIVE_CONVERSION && !FP16_USE_FLOAT16_TYPE && !FP16_USE_FP16_TYPE
+	#include <immintrin.h>
+#endif
+#if (defined(__aarch64__) || defined(_M_ARM64)) && FP16_USE_NATIVE_CONVERSION && !FP16_USE_FLOAT16_TYPE && !FP16_USE_FP16_TYPE
+	#include <arm_neon.h>
+#endif
 
 
 /*
@@ -106,6 +113,31 @@ 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_NATIVE_CONVERSION
+	#if FP16_USE_FLOAT16_TYPE
+		union {
+			uint16_t as_bits;
+			_Float16 as_value;
+		} fp16 = { h };
+		return (float) fp16.as_value;
+	#elif FP16_USE_FP16_TYPE
+		union {
+			uint16_t as_bits;
+			__fp16 as_value;
+		} fp16 = { h };
+		return (float) fp16.as_value;
+	#else
+		#if (defined(__INTEL_COMPILER) || defined(__GNUC__)) && defined(__F16C__)
+			return _cvtsh_ss((unsigned short) h);
+		#elif defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) && defined(__AVX2__)
+			return _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128((int) (unsigned int) h)));
+		#elif defined(_M_ARM64) || defined(__aarch64__)
+			return vgetq_lane_f32(vcvt_f32_f16(vreinterpret_f16_u16(vdup_n_u16(h))), 0);
+		#else
+			#error "Archtecture- or compiler-specific implementation required"
+		#endif
+	#endif
+#else
 	/*
 	 * Extend the half-precision floating-point number to 32 bits and shift to the upper part of the 32-bit word:
 	 *      +---+-----+------------+-------------------+
@@ -211,6 +243,7 @@ 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
 }
 
 /*
@@ -221,14 +254,44 @@ 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_NATIVE_CONVERSION
+	#if FP16_USE_FLOAT16_TYPE
+		union {
+			_Float16 as_value;
+			uint16_t as_bits;
+		} fp16 = { (_Float16) f };
+		return fp16.as_bits;
+	#elif FP16_USE_FP16_TYPE
+		union {
+			__fp16 as_value;
+			uint16_t as_bits;
+		} fp16 = { (__fp16) f };
+		return fp16.as_bits;
+	#else
+		#if (defined(__INTEL_COMPILER) || defined(__GNUC__)) && defined(__F16C__)
+			return _cvtss_sh(f, _MM_FROUND_CUR_DIRECTION);
+		#elif defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) && defined(__AVX2__)
+			return (uint16_t) _mm_cvtsi128_si32(_mm_cvtps_ph(_mm_set_ss(f), _MM_FROUND_CUR_DIRECTION));
+		#elif defined(_M_ARM64) || defined(__aarch64__)
+			return vget_lane_u16(vcvt_f16_f32(vdupq_n_f32(f)), 0);
+		#else
+			#error "Archtecture- or compiler-specific implementation required"
+		#endif
+	#endif
+#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;
 #else
 	const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
 	const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
 #endif
-	float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+#if defined(_MSC_VER) && defined(_M_IX86_FP) && (_M_IX86_FP == 0) || defined(__GNUC__) && defined(__FLT_EVAL_METHOD__) && (__FLT_EVAL_METHOD__ != 0)
+	const volatile float saturated_f = fabsf(f) * scale_to_inf;
+#else
+	const float saturated_f = fabsf(f) * scale_to_inf;
+#endif
+	float base = saturated_f * scale_to_zero;
 
 	const uint32_t w = fp32_to_bits(f);
 	const uint32_t shl1_w = w + w;
@@ -244,6 +307,7 @@ 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
 }
 
 /*
@@ -449,3 +513,9 @@ static inline uint16_t fp16_alt_from_fp32_value(float f) {
 }
 
 #endif /* FP16_FP16_H */
+
+// This file is part of the fp16 inline third-party dependency of YugabyteDB.
+// Git repo: https://github.com/Maratyszcza/FP16/
+// Git commit: 98b0a46bce017382a6351a19577ec43a715b6835
+//
+// See also src/inline-thirdparty/README.md.

src/inline-thirdparty/fp16/fp16/macros.h

@@ -0,0 +1,52 @@
+#pragma once
+#ifndef FP16_MACROS_H
+#define FP16_MACROS_H
+
+#ifndef FP16_USE_NATIVE_CONVERSION
+	#if (defined(__INTEL_COMPILER) || defined(__GNUC__)) && defined(__F16C__)
+		#define FP16_USE_NATIVE_CONVERSION 1
+	#elif defined(_MSC_VER) && (defined(_M_IX86) || defined(_M_X64)) && defined(__AVX2__)
+		#define FP16_USE_NATIVE_CONVERSION 1
+	#elif defined(_MSC_VER) && defined(_M_ARM64)
+		#define FP16_USE_NATIVE_CONVERSION 1
+	#elif defined(__GNUC__) && defined(__aarch64__)
+		#define FP16_USE_NATIVE_CONVERSION 1
+	#endif
+	#if !defined(FP16_USE_NATIVE_CONVERSION)
+		#define FP16_USE_NATIVE_CONVERSION 0
+	#endif  // !defined(FP16_USE_NATIVE_CONVERSION)
+#endif  // !define(FP16_USE_NATIVE_CONVERSION)
+
+#ifndef FP16_USE_FLOAT16_TYPE
+	#if !defined(__clang__) && !defined(__INTEL_COMPILER) && defined(__GNUC__) && (__GNUC__ >= 12)
+		#if defined(__F16C__)
+			#define FP16_USE_FLOAT16_TYPE 1
+		#endif
+	#endif
+	#if !defined(FP16_USE_FLOAT16_TYPE)
+		#define FP16_USE_FLOAT16_TYPE 0
+	#endif  // !defined(FP16_USE_FLOAT16_TYPE)
+#endif  // !defined(FP16_USE_FLOAT16_TYPE)
+
+#ifndef FP16_USE_FP16_TYPE
+	#if defined(__clang__)
+		#if defined(__F16C__) || defined(__aarch64__)
+			#define FP16_USE_FP16_TYPE 1
+		#endif
+	#elif defined(__GNUC__)
+		#if 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 */
+
+// This file is part of the fp16 inline third-party dependency of YugabyteDB.
+// Git repo: https://github.com/Maratyszcza/FP16/
+// Git commit: 98b0a46bce017382a6351a19577ec43a715b6835
+//
+// See also src/inline-thirdparty/README.md.