[clang] [X86][AVX10.2] Add comments for the avx10_2convertintrin.h file (PR #120766)

Mikołaj Piróg via cfe-commits cfe-commits at lists.llvm.org
Thu Jan 23 14:27:46 PST 2025


https://github.com/mikolaj-pirog updated https://github.com/llvm/llvm-project/pull/120766

>From 18dfaec1d43fe38dc6b3fe754153426e5c974f0f Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Miko=C5=82aj=20Pir=C3=B3g?= <mikolajpirog at gmail.com>
Date: Thu, 23 Jan 2025 23:27:17 +0100
Subject: [PATCH] rename

---
 clang/lib/Headers/avx10_2convertintrin.h | 2943 +++++++++++++++++++++-
 1 file changed, 2879 insertions(+), 64 deletions(-)

diff --git a/clang/lib/Headers/avx10_2convertintrin.h b/clang/lib/Headers/avx10_2convertintrin.h
index c67a5b890f1957..82d434c18df131 100644
--- a/clang/lib/Headers/avx10_2convertintrin.h
+++ b/clang/lib/Headers/avx10_2convertintrin.h
@@ -24,24 +24,157 @@
   __attribute__((__always_inline__, __nodebug__, __target__("avx10.2-256"),    \
                  __min_vector_width__(256)))
 
+// clang-format off
+
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed
+///    single-precision (32-bit) floating-point elements to a 128-bit vector
+///    containing FP16 elements.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF i < 4
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 	ELSE
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 4])
+/// 	FI
+///
+/// ENDFOR
+///
+/// dst[MAX:127] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __A
+///    A 128-bit vector of [4 x float].
+/// \param __B
+///    A 128-bit vector of [4 x float].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Lower 4 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvtx2ps_ph(__m128 __A,
                                                                __m128 __B) {
   return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
       (__v4sf)__A, (__v4sf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)(-1));
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 128-bit vector
+/// containing FP16 elements. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		IF i < 4
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 4])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:127] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __W
+///    A 128-bit vector of [8 x fp16].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [4 x float].
+/// \param __B
+///    A 128-bit vector of [4 x float].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128
 _mm_mask_cvtx2ps_ph(__m128h __W, __mmask8 __U, __m128 __A, __m128 __B) {
   return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
       (__v4sf)__A, (__v4sf)__B, (__v8hf)__W, (__mmask8)__U);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 128-bit vector
+/// containing FP16 elements. Zeroing mask \a __U is used to determine if given
+/// element should be zeroed instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		IF i < 4
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 4])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:127] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [4 x float].
+/// \param __B
+///    A 128-bit vector of [4 x float].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then zero is taken instead.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtx2ps_ph(__mmask8 __U, __m128 __A, __m128 __B) {
   return (__m128h)__builtin_ia32_vcvt2ps2phx128_mask(
       (__v4sf)__A, (__v4sf)__B, (__v8hf)_mm_setzero_ph(), (__mmask8)__U);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF i < 8
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 	ELSE
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvtx2ps_ph(__m256 __A,
                                                                   __m256 __B) {
   return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
@@ -49,6 +182,44 @@ static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvtx2ps_ph(__m256 __A,
       _MM_FROUND_CUR_DIRECTION);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __W
+///    A 256-bit vector of [16 x fp16].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
 _mm256_mask_cvtx2ps_ph(__m256h __W, __mmask16 __U, __m256 __A, __m256 __B) {
   return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
@@ -56,6 +227,42 @@ _mm256_mask_cvtx2ps_ph(__m256h __W, __mmask16 __U, __m256 __A, __m256 __B) {
       _MM_FROUND_CUR_DIRECTION);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements. Zeroing mask \a __U is used to determine if given
+/// element should be zeroed instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then zero is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtx2ps_ph(__mmask16 __U, __m256 __A, __m256 __B) {
   return (__m256h)__builtin_ia32_vcvt2ps2phx256_mask(
@@ -63,32 +270,239 @@ _mm256_maskz_cvtx2ps_ph(__mmask16 __U, __m256 __A, __m256 __B) {
       _MM_FROUND_CUR_DIRECTION);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements. Rounding mode \a __R needs to be provided.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF i < 8
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 	ELSE
+/// 		dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \param __R
+///    Rounding mode. Valid inputs are: _MM_FROUND_CUR_DIRECTION or
+///    result of bitwise or of _MM_FROUND_NO_EXC with at most one of the following:
+///    _MM_FROUND_TO_NEAREST_INT, _MM_FROUND_TO_NEG_INF, _MM_FROUND_TO_POS_INF,
+///    _MM_FROUND_TO_ZERO.
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 #define _mm256_cvtx_round2ps_ph(A, B, R)                                       \
   ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask(                                \
       (__v8sf)(A), (__v8sf)(B), (__v16hf)_mm256_undefined_ph(),                \
       (__mmask16)(-1), (const int)(R)))
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead. Rounding mode \a __R needs to
+/// be provided.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __W
+///    A 256-bit vector of [16 x fp16].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \param __R
+///    Rounding mode. Valid inputs are: _MM_FROUND_CUR_DIRECTION or
+///    result of bitwise or of _MM_FROUND_NO_EXC with at most one of the following:
+///    _MM_FROUND_TO_NEAREST_INT, _MM_FROUND_TO_NEG_INF, _MM_FROUND_TO_POS_INF,
+///    _MM_FROUND_TO_ZERO.
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 #define _mm256_mask_cvtx_round2ps_ph(W, U, A, B, R)                            \
   ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask(                                \
       (__v8sf)(A), (__v8sf)(B), (__v16hf)(W), (__mmask16)(U), (const int)(R)))
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed
+/// single-precision (32-bit) floating-point elements to a 256-bit vector
+/// containing FP16 elements. Zeroing mask \a __U is used to determine if given
+/// element should be zeroed instead. Rounding mode \a __R needs to be provided.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__B.fp32[i])
+/// 		ELSE
+/// 			dst.fp16[i] := convert_fp32_to_fp16(__A.fp32[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PS2PHX instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [8 x float].
+/// \param __B
+///    A 256-bit vector of [8 x float].
+/// \param __R
+///    Rounding mode. Valid inputs are: _MM_FROUND_CUR_DIRECTION or
+///    result of bitwise or of _MM_FROUND_NO_EXC with at most one of the following:
+///    _MM_FROUND_TO_NEAREST_INT, _MM_FROUND_TO_NEG_INF, _MM_FROUND_TO_POS_INF,
+///    _MM_FROUND_TO_ZERO.
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Lower elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then zero is taken instead.
 #define _mm256_maskz_cvtx_round2ps_ph(U, A, B, R)                              \
   ((__m256h)__builtin_ia32_vcvt2ps2phx256_mask(                                \
       (__v8sf)(A), (__v8sf)(B), (__v16hf)(_mm256_setzero_ph()),                \
       (__mmask16)(U), (const int)(R)))
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtbiasph_bf8(__m128i __A,
-                                                                  __m128h __B) {
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    sum of elements from \a __A and \a __B; higher order elements are zeroed.
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_cvtbiasph_bf8(__m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtbiasph_bf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Zeroing mask \a __U is used to determine if given element should be zeroed
+///    instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtbiasph_bf8(__mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_128_mask(
@@ -96,6 +510,29 @@ _mm_maskz_cvtbiasph_bf8(__mmask8 __U, __m128i __A, __m128h __B) {
       (__mmask8)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Elements correspond to the
+///    sum of elements from \a __A and \a __B.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtbiasph_bf8(__m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
@@ -103,12 +540,78 @@ _mm256_cvtbiasph_bf8(__m256i __A, __m256h __B) {
       (__mmask16)-1);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiasph_bf8(
     __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
       (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Merging mask \a __U is used to determine if given element should be taken
+///    from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set,
+///    then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtbiasph_bf8(__mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8_256_mask(
@@ -116,18 +619,108 @@ _mm256_maskz_cvtbiasph_bf8(__mmask16 __U, __m256i __A, __m256h __B) {
       (__mmask16)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Results are saturated.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Lower elements correspond to the
+///    sum of elements from \a __A and \a __B; higher order elements are zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtbiassph_bf8(__m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+/// Results are saturated. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtbiassph_bf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Results are saturated. Zeroing mask \a __U is used to determine if given
+///    element should be zeroed instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8S instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtbiassph_bf8(__mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_128_mask(
@@ -135,6 +728,30 @@ _mm_maskz_cvtbiassph_bf8(__mmask8 __U, __m128i __A, __m128h __B) {
       (__mmask8)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Results are saturated.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Elements correspond to the
+///    sum of elements from \a __A and \a __B.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtbiassph_bf8(__m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
@@ -142,12 +759,78 @@ _mm256_cvtbiassph_bf8(__m256i __A, __m256h __B) {
       (__mmask16)-1);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+/// Results are saturated. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiassph_bf8(
     __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
       (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E5M2.
+///    Results are saturated. Merging mask \a __U is used to determine if given
+///    element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2BF8S instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x bf8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set,
+///    then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtbiassph_bf8(__mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2bf8s_256_mask(
@@ -155,18 +838,107 @@ _mm256_maskz_cvtbiassph_bf8(__mmask16 __U, __m256i __A, __m256h __B) {
       (__mmask16)__U);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtbiasph_hf8(__m128i __A,
-                                                                  __m128h __B) {
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Lower elements correspond to the
+///    sum of elements from \a __A and \a __B; higher order elements are zeroed.
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_cvtbiasph_hf8(__m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtbiasph_hf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///    Zeroing mask \a __U is used to determine if given element should be zeroed
+///    instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtbiasph_hf8(__mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_128_mask(
@@ -174,6 +946,29 @@ _mm_maskz_cvtbiasph_hf8(__mmask8 __U, __m128i __A, __m128h __B) {
       (__mmask8)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := add_fp16_int8_convert_to_hf8(__A.fp16[i], __B.int8[2 * i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x half].
+/// \param __B
+///    A 256-bit vector of [16 x i16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Elements correspond to the
+///    sum of elements from \a __A and \a __B.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtbiasph_hf8(__m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
@@ -181,12 +976,78 @@ _mm256_cvtbiasph_hf8(__m256i __A, __m256h __B) {
       (__mmask16)-1);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiasph_hf8(
     __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
       (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3 
+///    Merging mask \a __U is used to determine if given element should be taken
+///    from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_fp16_int8_convert_to_hf8(__A.fp16[i], __B.int8[2 * i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8 instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x half].
+/// \param __B
+///    A 256-bit vector of [16 x i16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set,
+///    then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtbiasph_hf8(__mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8_256_mask(
@@ -194,18 +1055,108 @@ _mm256_maskz_cvtbiasph_hf8(__mmask16 __U, __m256i __A, __m256h __B) {
       (__mmask16)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///    Results are saturated.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S`instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Lower elements correspond to the
+///    sum of elements from \a __A and \a __B; higher order elements are zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_cvtbiassph_hf8(__m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+/// Results are saturated. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtbiassph_hf8(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
       (__v16qi)__A, (__v8hf)__B, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Add two 128-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///    Results are saturated. Zeroing mask \a __U is used to determine if given
+///    element should be zeroed instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Lower elements correspond to the sum of
+///    elements from \a __A and \a __B; higher order elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtbiassph_hf8(__mmask8 __U, __m128i __A, __m128h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_128_mask(
@@ -213,6 +1164,30 @@ _mm_maskz_cvtbiassph_hf8(__mmask8 __U, __m128i __A, __m128h __B) {
       (__mmask8)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///    Results are saturated.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Elements correspond to the
+///    sum of elements from \a __A and \a __B.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_cvtbiassph_hf8(__m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
@@ -220,12 +1195,78 @@ _mm256_cvtbiassph_hf8(__m256i __A, __m256h __B) {
       (__mmask16)-1);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements and 8-bit integers stored in the lower half of
+/// packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+/// Results are saturated. Merging mask \a __U is used to determine if given
+/// element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+/// 	ELSE
+/// 		dst.fp8[i] := _W[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_mask_cvtbiassph_hf8(
     __m128i __W, __mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
       (__v32qi)__A, (__v16hf)__B, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Add two 256-bit vectors, \a __A and \a __B, containing packed FP16
+///    floating-point elements and 8-bit integers stored in the lower half of
+///    packed 16-bit integers, respectively. Results are converted to FP8 E4M3.
+///    Results are saturated. Merging mask \a __U is used to determine if given
+///    element should be taken from \a __W instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+///	 	dst.fp8[i] := add_convert_fp16_to_fp8_bias(__A.fp16[i], __B.int8[2i])
+///	 ELSE
+///	 	dst.fp8[i] := 0
+///	 FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTBIASPH2HF8S instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x int16].
+/// \returns
+///    A 128-bit vector of [16 x hf8]. Elements correspond to the sum of
+///    elements from \a __A and \a __B. If corresponding mask bit is not set,
+///    then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtbiassph_hf8(__mmask16 __U, __m256i __A, __m256h __B) {
   return (__m128i)__builtin_ia32_vcvtbiasph2hf8s_256_mask(
@@ -233,17 +1274,119 @@ _mm256_maskz_cvtbiassph_hf8(__mmask16 __U, __m256i __A, __m256h __B) {
       (__mmask16)__U);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF i < 8
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvt2ph_bf8(__m128h __A,
-                                                               __m128h __B) {
-  return (__m128i)__builtin_ia32_vcvt2ph2bf8_128((__v8hf)(__A), (__v8hf)(__B));
-}
-
+                                                                  __m128h __B) {
+  return (__m128i)__builtin_ia32_vcvt2ph2bf8_128((__v8hf)(__A),
+                                                   (__v8hf)(__B));
+}
+
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvt2ph_bf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
       (__mmask16)__U, (__v16qi)_mm_cvt2ph_bf8(__A, __B), (__v16qi)__W);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvt2ph_bf8(__mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
@@ -251,18 +1394,119 @@ _mm_maskz_cvt2ph_bf8(__mmask16 __U, __m128h __A, __m128h __B) {
       (__v16qi)(__m128i)_mm_setzero_si128());
 }
 
-static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_cvt2ph_bf8(__m256h __A,
-                                                                  __m256h __B) {
-  return (__m256i)__builtin_ia32_vcvt2ph2bf8_256((__v16hf)(__A),
-                                                 (__v16hf)(__B));
-}
-
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF i < 16 
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
-_mm256_mask_cvt2ph_bf8(__m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
+_mm256_cvt2ph_bf8(__m256h __A, __m256h __B) {
+  return (__m256i)__builtin_ia32_vcvt2ph2bf8_256((__v16hf)(__A),
+                                                   (__v16hf)(__B));
+}
+
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __W
+///    A 256-bit vector of [32 x fp8].
+/// \param __U
+///    A 32-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
+static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvt2ph_bf8(
+    __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
       (__mmask16)__U, (__v32qi)_mm256_cvt2ph_bf8(__A, __B), (__v32qi)__W);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := 0
+/// 	ELSE
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8 instruction.
+///
+/// \param __U
+///    A 32-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    zero is taken instead.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvt2ph_bf8(__mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
@@ -270,17 +1514,120 @@ _mm256_maskz_cvt2ph_bf8(__mmask32 __U, __m256h __A, __m256h __B) {
       (__v32qi)(__m256i)_mm256_setzero_si256());
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvts2ph_bf8(__m128h __A,
-                                                                __m128h __B) {
-  return (__m128i)__builtin_ia32_vcvt2ph2bf8s_128((__v8hf)(__A), (__v8hf)(__B));
-}
-
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+/// Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF i < 8
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_cvts2ph_bf8(__m128h __A, __m128h __B) {
+  return (__m128i)__builtin_ia32_vcvt2ph2bf8s_128((__v8hf)(__A),
+                                                    (__v8hf)(__B));
+}
+
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation}
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvts2ph_bf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
       (__mmask16)__U, (__v16qi)_mm_cvts2ph_bf8(__A, __B), (__v16qi)__W);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E5M2 FP8 elements.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvts2ph_bf8(__mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
@@ -288,18 +1635,120 @@ _mm_maskz_cvts2ph_bf8(__mmask16 __U, __m128h __A, __m128h __B) {
       (__v16qi)(__m128i)_mm_setzero_si128());
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+/// Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF i < 16 
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_cvts2ph_bf8(__m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_vcvt2ph2bf8s_256((__v16hf)(__A),
-                                                  (__v16hf)(__B));
-}
-
-static __inline__ __m256i __DEFAULT_FN_ATTRS256
-_mm256_mask_cvts2ph_bf8(__m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
+                                                    (__v16hf)(__B));
+}
+
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __W
+///    A 256-bit vector of [32 x fp8].
+/// \param __U
+///    A 32-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
+static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvts2ph_bf8(
+    __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
       (__mmask16)__U, (__v32qi)_mm256_cvts2ph_bf8(__A, __B), (__v32qi)__W);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E5M2 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be zeroed
+/// instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := 0
+/// 	ELSE
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2BF8S instruction.
+///
+/// \param __U
+///    A 32-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    zero is taken instead.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvts2ph_bf8(__mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
@@ -307,17 +1756,119 @@ _mm256_maskz_cvts2ph_bf8(__mmask32 __U, __m256h __A, __m256h __B) {
       (__v32qi)(__m256i)_mm256_setzero_si256());
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF i < 8
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvt2ph_hf8(__m128h __A,
-                                                               __m128h __B) {
-  return (__m128i)__builtin_ia32_vcvt2ph2hf8_128((__v8hf)(__A), (__v8hf)(__B));
-}
-
+                                                                  __m128h __B) {
+  return (__m128i)__builtin_ia32_vcvt2ph2hf8_128((__v8hf)(__A),
+                                                   (__v8hf)(__B));
+}
+
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvt2ph_hf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
       (__mmask16)__U, (__v16qi)_mm_cvt2ph_hf8(__A, __B), (__v16qi)__W);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvt2ph_hf8(__mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
@@ -325,18 +1876,119 @@ _mm_maskz_cvt2ph_hf8(__mmask16 __U, __m128h __A, __m128h __B) {
       (__v16qi)(__m128i)_mm_setzero_si128());
 }
 
-static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_cvt2ph_hf8(__m256h __A,
-                                                                  __m256h __B) {
-  return (__m256i)__builtin_ia32_vcvt2ph2hf8_256((__v16hf)(__A),
-                                                 (__v16hf)(__B));
-}
-
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF i < 16 
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
-_mm256_mask_cvt2ph_hf8(__m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
+_mm256_cvt2ph_hf8(__m256h __A, __m256h __B) {
+  return (__m256i)__builtin_ia32_vcvt2ph2hf8_256((__v16hf)(__A),
+                                                   (__v16hf)(__B));
+}
+
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __W
+///    A 256-bit vector of [32 x fp8].
+/// \param __U
+///    A 32-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
+static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvt2ph_hf8(
+    __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
       (__mmask16)__U, (__v32qi)_mm256_cvt2ph_hf8(__A, __B), (__v32qi)__W);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := 0
+/// 	ELSE
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8 instruction.
+///
+/// \param __U
+///    A 32-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    zero is taken instead.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvt2ph_hf8(__mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
@@ -344,17 +1996,120 @@ _mm256_maskz_cvt2ph_hf8(__mmask32 __U, __m256h __A, __m256h __B) {
       (__v32qi)(__m256i)_mm256_setzero_si256());
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvts2ph_hf8(__m128h __A,
-                                                                __m128h __B) {
-  return (__m128i)__builtin_ia32_vcvt2ph2hf8s_128((__v8hf)(__A), (__v8hf)(__B));
-}
-
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+/// Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF i < 8
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS128
+_mm_cvts2ph_hf8(__m128h __A, __m128h __B) {
+  return (__m128i)__builtin_ia32_vcvt2ph2hf8s_128((__v8hf)(__A),
+                                                    (__v8hf)(__B));
+}
+
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvts2ph_hf8(__m128i __W, __mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
       (__mmask16)__U, (__v16qi)_mm_cvts2ph_hf8(__A, __B), (__v16qi)__W);
 }
 
+/// Convert two 128-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 128-bit vector containing E4M3 FP8 elements.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 16 
+/// 	IF __U[i]
+/// 		IF i < 8
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 8])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __U
+///    A 16-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \param __B
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower 8 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvts2ph_hf8(__mmask16 __U, __m128h __A, __m128h __B) {
   return (__m128i)__builtin_ia32_selectb_128(
@@ -362,18 +2117,120 @@ _mm_maskz_cvts2ph_hf8(__mmask16 __U, __m128h __A, __m128h __B) {
       (__v16qi)(__m128i)_mm_setzero_si128());
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+/// Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF i < 16 
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_cvts2ph_hf8(__m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_vcvt2ph2hf8s_256((__v16hf)(__A),
-                                                  (__v16hf)(__B));
-}
-
-static __inline__ __m256i __DEFAULT_FN_ATTRS256
-_mm256_mask_cvts2ph_hf8(__m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
+                                                    (__v16hf)(__B));
+}
+
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	ELSE
+/// 		dst.fp8[i] := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __W
+///    A 256-bit vector of [32 x fp8].
+/// \param __U
+///    A 32-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
+static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_cvts2ph_hf8(
+    __m256i __W, __mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
       (__mmask16)__U, (__v32qi)_mm256_cvts2ph_hf8(__A, __B), (__v32qi)__W);
 }
 
+/// Convert two 256-bit vectors, \a __A and \a __B, containing packed FP16
+/// floating-point elements to a 256-bit vector containing E4M3 FP8 elements.
+/// Merging mask \a __U is used to determine if given element should be zeroed
+/// instead. Resulting elements are saturated in case of overflow.
+///
+/// \code{.operation]
+/// FOR i := 0 to 32 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := 0
+/// 	ELSE
+/// 		IF i < 16 
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__B.fp16[i])
+/// 		ELSE
+/// 			dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i - 16])
+/// 		FI
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVT2PH2HF8S instruction.
+///
+/// \param __U
+///    A 32-bit zeroing mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \param __B
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 256-bit vector of [32 x fp8]. Lower 16 elements correspond to the
+///    (converted) elements from \a __B; higher order elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    zero is taken instead.
 static __inline__ __m256i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvts2ph_hf8(__mmask32 __U, __m256h __A, __m256h __B) {
   return (__m256i)__builtin_ia32_selectb_256(
@@ -381,208 +2238,1166 @@ _mm256_maskz_cvts2ph_hf8(__mmask32 __U, __m256h __A, __m256h __B) {
       (__v32qi)(__m256i)_mm256_setzero_si256());
 }
 
+/// Convert 128-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is exact.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvthf8(__m128i __A) {
   return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
       (__v16qi)__A, (__v8hf)(__m128h)_mm_undefined_ph(), (__mmask8)-1);
 }
 
-static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_mask_cvthf8(__m128h __W,
-                                                                __mmask8 __U,
-                                                                __m128i __A) {
+/// Convert 128-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is
+/// exact. Merging mask \a __U is used to determine if given element should be
+/// taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__A.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __W
+///    A 128-bit vector of [8 x fp16].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
+static __inline__ __m128h __DEFAULT_FN_ATTRS128
+_mm_mask_cvthf8(__m128h __W, __mmask8 __U, __m128i __A) {
   return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
       (__v16qi)__A, (__v8hf)(__m128h)__W, (__mmask8)__U);
 }
 
-static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_maskz_cvthf8(__mmask8 __U,
-                                                                 __m128i __A) {
+/// Convert 128-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is
+/// exact. Zeroing mask \a __U is used to determine if given element should be
+/// zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
+static __inline__ __m128h __DEFAULT_FN_ATTRS128
+_mm_maskz_cvthf8(__mmask8 __U, __m128i __A) {
   return (__m128h)__builtin_ia32_vcvthf8_2ph128_mask(
       (__v16qi)__A, (__v8hf)(__m128h)_mm_setzero_ph(), (__mmask8)__U);
 }
 
-static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvthf8(__m128i __A) {
+/// Convert 256-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is exact.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A.
+static __inline__ __m256h __DEFAULT_FN_ATTRS256
+_mm256_cvthf8(__m128i __A) {
   return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
       (__v16qi)__A, (__v16hf)(__m256h)_mm256_undefined_ph(), (__mmask16)-1);
 }
 
+/// Convert 256-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is
+/// exact. Merging mask \a __U is used to determine if given element should be
+/// taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__A.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __W
+///    A 256-bit vector of [16 x fp16].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
 _mm256_mask_cvthf8(__m256h __W, __mmask16 __U, __m128i __A) {
   return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
       (__v16qi)__A, (__v16hf)(__m256h)__W, (__mmask16)__U);
 }
 
+/// Convert 256-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is
+/// exact. Zeroing mask \a __U is used to determine if given element should be
+/// zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:256] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTHF82PH instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvthf8(__mmask16 __U, __m128i __A) {
   return (__m256h)__builtin_ia32_vcvthf8_2ph256_mask(
       (__v16qi)__A, (__v16hf)(__m256h)_mm256_setzero_ph(), (__mmask16)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the (converted)
+///    elements from \a __A; upper elements are zeroed. 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtph_bf8(__m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Merging mask \a __U is used to determine if
+/// given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtph_bf8(__m128i __W, __mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Zeroing mask \a __U is used to determine if
+/// given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := 0
+/// 	ELSE
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtph_bf8(__mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtph_bf8(__m256h __A) {
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the (converted)
+///    elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS256
+_mm256_cvtph_bf8(__m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Merging mask \a __U is
+/// used to determine if given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_mask_cvtph_bf8(__m128i __W, __mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Zeroing mask \a __U is
+/// used to determine if given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8 instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then element is zeroed instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtph_bf8(__mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the (converted)
+///    elements from \a __A; upper elements are zeroed. 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtsph_bf8(__m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated. Merging mask \a __U is
+/// used to determine if given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtsph_bf8(__m128i __W, __mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated. Zeroing mask \a __U is
+/// used to determine if given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtsph_bf8(__mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtsph_bf8(__m256h __A) {
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Results are saturated.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the (converted)
+///    elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS256
+_mm256_cvtsph_bf8(__m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Results are saturated.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_mask_cvtsph_bf8(__m128i __W, __mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Results are saturated.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2BF8S instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then element is zeroed instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtsph_bf8(__mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2bf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E5M2 FP8 elements. Upper elements of
+/// resulting vector are zeroed.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the (converted)
+///    elements from \a __A; upper elements are zeroed. 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtph_hf8(__m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Merging mask \a __U is used to determine if
+/// given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtph_hf8(__m128i __W, __mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Zeroing mask \a __U is used to determine if
+/// given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtph_hf8(__mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtph_hf8(__m256h __A) {
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the (converted)
+///    elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS256
+_mm256_cvtph_hf8(__m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Merging mask \a __U is
+/// used to determine if given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_mask_cvtph_hf8(__m128i __W, __mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Zeroing mask \a __U is
+/// used to determine if given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8 instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then element is zeroed instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtph_hf8(__mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the (converted)
+///    elements from \a __A; upper elements are zeroed. 
 static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_cvtsph_hf8(__m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask8)-1);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated. Merging mask \a __U is
+/// used to determine if given element should be taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_mask_cvtsph_hf8(__m128i __W, __mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)__W, (__mmask8)__U);
 }
 
+/// Convert 128-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Upper elements of
+/// resulting vector are zeroed. Results are saturated. Zeroing mask \a __U is
+/// used to determine if given element should be zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Lower elements correspond to the
+///    (converted) elements from \a __A; upper elements are zeroed. If
+///    corresponding mask bit is not set, then element is zeroed.
 static __inline__ __m128i __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtsph_hf8(__mmask8 __U, __m128h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_128_mask(
       (__v8hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask8)__U);
 }
 
-static __inline__ __m128i __DEFAULT_FN_ATTRS256 _mm256_cvtsph_hf8(__m256h __A) {
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Results are saturated.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the (converted)
+///    elements from \a __A.
+static __inline__ __m128i __DEFAULT_FN_ATTRS256
+_mm256_cvtsph_hf8(__m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_undefined_si128(), (__mmask16)-1);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Results are saturated.
+/// Merging mask \a __U is used to determine if given element should be taken
+/// from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8 := __W.fp8[i]
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:128] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __W
+///    A 128-bit vector of [16 x fp8].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [8 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If
+///    corresponding mask bit is not set, then element from \a __W is taken instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_mask_cvtsph_hf8(__m128i __W, __mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)__W, (__mmask16)__U);
 }
 
+/// Convert 256-bit vector \a __A containing packed FP16 floating-point elements
+/// to a 128-bit vector containing E4M3 FP8 elements. Results are saturated.
+/// Zeroing mask \a __U is used to determine if given element should be zeroed
+/// instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp8[i] := convert_fp16_to_fp8(__A.fp16[i])
+/// 	ELSE
+/// 		dst.fp8[i] := 0
+/// 	FI
+/// ENDFOR
+///
+/// dst[MAX:64] := 0
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic corresponds to the \c VCVTPH2HF8S instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [16 x fp16].
+/// \returns
+///    A 128-bit vector of [16 x fp8]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set,
+///    then element is zeroed instead.
 static __inline__ __m128i __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtsph_hf8(__mmask16 __U, __m256h __A) {
   return (__m128i)__builtin_ia32_vcvtph2hf8s_256_mask(
       (__v16hf)__A, (__v16qi)(__m128i)_mm_setzero_si128(), (__mmask16)__U);
 }
 
+/// Convert 128-bit vector \a __A, containing packed FP8 E5M2 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is exact.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128 _mm_cvtbf8_ph(__m128i __A) {
   return _mm_castsi128_ph(_mm_slli_epi16(_mm_cvtepi8_epi16(__A), 8));
 }
 
+/// Convert 128-bit vector \a __A, containing packed FP8 E5M2 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is
+/// exact. Merging mask \a __U is used to determine if given element should be
+/// taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __W
+///    A 128-bit vector of [8 x fp16].
+/// \param __U
+///    A 8-bit merging mask.
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128
-_mm_mask_cvtbf8_ph(__m128h __S, __mmask8 __U, __m128i __A) {
+_mm_mask_cvtbf8_ph(__m128h __W, __mmask8 __U, __m128i __A) {
   return _mm_castsi128_ph(
-      _mm_mask_slli_epi16((__m128i)__S, __U, _mm_cvtepi8_epi16(__A), 8));
-}
-
+      _mm_mask_slli_epi16((__m128i)__W, __U, _mm_cvtepi8_epi16(__A), 8));
+}
+
+/// Convert 128-bit vector \a __A, containing packed FP8 E5M2 floating-point
+/// elements to a 128-bit vector containing FP16 elements. The conversion is
+/// exact. Zeroing mask \a __U is used to determine if given element should be
+/// zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 7
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __U
+///    A 8-bit zeroing mask.
+/// \param __A
+///    A 128-bit vector of [16 x fp8].
+/// \returns
+///    A 128-bit vector of [8 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m128h __DEFAULT_FN_ATTRS128
 _mm_maskz_cvtbf8_ph(__mmask8 __U, __m128i __A) {
   return _mm_castsi128_ph(_mm_slli_epi16(_mm_maskz_cvtepi8_epi16(__U, __A), 8));
 }
 
+/// Convert 256-bit vector \a __A, containing packed FP8 E4M3 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is exact.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15
+/// 	dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256 _mm256_cvtbf8_ph(__m128i __A) {
   return _mm256_castsi256_ph(_mm256_slli_epi16(_mm256_cvtepi8_epi16(__A), 8));
 }
 
+/// Convert 256-bit vector \a __A, containing packed FP8 E5M2 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is
+/// exact. Merging mask \a __U is used to determine if given element should be
+/// taken from \a __W instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := __W.fp16[i]
+/// 	FI
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __W
+///    A 256-bit vector of [16 x fp16].
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    element from \a __W is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
-_mm256_mask_cvtbf8_ph(__m256h __S, __mmask16 __U, __m128i __A) {
+_mm256_mask_cvtbf8_ph(__m256h __W, __mmask16 __U, __m128i __A) {
   return _mm256_castsi256_ph(
-      _mm256_mask_slli_epi16((__m256i)__S, __U, _mm256_cvtepi8_epi16(__A), 8));
-}
-
+      _mm256_mask_slli_epi16((__m256i)__W, __U, _mm256_cvtepi8_epi16(__A), 8));
+}
+
+/// Convert 256-bit vector \a __A, containing packed FP8 E5M2 floating-point
+/// elements to a 256-bit vector containing FP16 elements. The conversion is
+/// exact. Zeroing mask \a __U is used to determine if given element should be
+/// zeroed instead.
+///
+/// \code{.operation]
+/// FOR i := 0 to 15 
+/// 	IF __U[i]
+/// 		dst.fp16[i] := convert_fp8_to_fp16(__B.fp8[i])
+/// 	ELSE
+/// 		dst.fp16[i] := 0
+/// 	FI
+/// ENDFOR
+/// \endcode
+///
+/// \headerfile <immintrin.h>
+///
+/// This intrinsic does not correspond to a single instruction.
+///
+/// \param __U
+///    A 16-bit merging mask.
+/// \param __A
+///    A 256-bit vector of [32 x fp8].
+/// \returns
+///    A 256-bit vector of [16 x fp16]. Resulting elements correspond to the
+///    (converted) elements from \a __A. If corresponding mask bit is not set, then
+///    zero is taken instead.
 static __inline__ __m256h __DEFAULT_FN_ATTRS256
 _mm256_maskz_cvtbf8_ph(__mmask16 __U, __m128i __A) {
   return _mm256_castsi256_ph(
       _mm256_slli_epi16(_mm256_maskz_cvtepi8_epi16(__U, __A), 8));
 }
 
+// clang-format on
+
 #undef __DEFAULT_FN_ATTRS128
 #undef __DEFAULT_FN_ATTRS256
 



More information about the cfe-commits mailing list