clang  14.0.0git
Macros | Typedefs | Functions
avx512bf16intrin.h File Reference
This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Macros

#define __DEFAULT_FN_ATTRS512
 
#define __DEFAULT_FN_ATTRS   __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16")))
 

Typedefs

typedef short __m512bh __attribute__((__vector_size__(64), __aligned__(64)))
 
typedef unsigned short __bfloat16
 

Functions

static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss (__bfloat16 __A)
 Convert One BF16 Data to One Single Float Data. More...
 
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_cvtne2ps_pbh (__m512 __A, __m512 __B)
 Convert Two Packed Single Data to One Packed BF16 Data. More...
 
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtne2ps_pbh (__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B)
 Convert Two Packed Single Data to One Packed BF16 Data. More...
 
static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtne2ps_pbh (__mmask32 __U, __m512 __A, __m512 __B)
 Convert Two Packed Single Data to One Packed BF16 Data. More...
 
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_cvtneps_pbh (__m512 __A)
 Convert Packed Single Data to Packed BF16 Data. More...
 
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtneps_pbh (__m256bh __W, __mmask16 __U, __m512 __A)
 Convert Packed Single Data to Packed BF16 Data. More...
 
static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtneps_pbh (__mmask16 __U, __m512 __A)
 Convert Packed Single Data to Packed BF16 Data. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_dpbf16_ps (__m512 __D, __m512bh __A, __m512bh __B)
 Dot Product of BF16 Pairs Accumulated into Packed Single Precision. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_dpbf16_ps (__m512 __D, __mmask16 __U, __m512bh __A, __m512bh __B)
 Dot Product of BF16 Pairs Accumulated into Packed Single Precision. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_dpbf16_ps (__mmask16 __U, __m512 __D, __m512bh __A, __m512bh __B)
 Dot Product of BF16 Pairs Accumulated into Packed Single Precision. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps (__m256bh __A)
 Convert Packed BF16 Data to Packed float Data. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtpbh_ps (__mmask16 __U, __m256bh __A)
 Convert Packed BF16 Data to Packed float Data using zeroing mask. More...
 
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_cvtpbh_ps (__m512 __S, __mmask16 __U, __m256bh __A)
 Convert Packed BF16 Data to Packed float Data using merging mask. More...
 

Macro Definition Documentation

◆ __DEFAULT_FN_ATTRS

#define __DEFAULT_FN_ATTRS   __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16")))

Definition at line 23 of file avx512bf16intrin.h.

◆ __DEFAULT_FN_ATTRS512

#define __DEFAULT_FN_ATTRS512
Value:
__attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \
__min_vector_width__(512)))

Definition at line 20 of file avx512bf16intrin.h.

Typedef Documentation

◆ __attribute__

typedef short __m256bh __attribute__((__vector_size__(32), __aligned__(32)))

Definition at line 16 of file avx512bf16intrin.h.

◆ __bfloat16

typedef unsigned short __bfloat16

Definition at line 18 of file avx512bf16intrin.h.

Function Documentation

◆ _mm512_cvtne2ps_pbh()

static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_cvtne2ps_pbh ( __m512  __A,
__m512  __B 
)
static

Convert Two Packed Single Data to One Packed BF16 Data.

This intrinsic corresponds to the VCVTNE2PS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
__BA 512-bit vector of [16 x float].
Returns
A 512-bit vector of [32 x bfloat] whose lower 256 bits come from conversion of __B, and higher 256 bits come from conversion of __A.

Definition at line 53 of file avx512bf16intrin.h.

Referenced by _mm512_mask_cvtne2ps_pbh(), and _mm512_maskz_cvtne2ps_pbh().

◆ _mm512_cvtneps_pbh()

static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_cvtneps_pbh ( __m512  __A)
static

Convert Packed Single Data to Packed BF16 Data.

This intrinsic corresponds to the VCVTNEPS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
Returns
A 256-bit vector of [16 x bfloat] come from conversion of __A.

Definition at line 114 of file avx512bf16intrin.h.

References _mm256_undefined_si256().

◆ _mm512_cvtpbh_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps ( __m256bh  __A)
static

Convert Packed BF16 Data to Packed float Data.

Parameters
__AA 256-bit vector of [16 x bfloat].
Returns
A 512-bit vector of [16 x float] come from convertion of __A

Definition at line 236 of file avx512bf16intrin.h.

References _mm512_castsi512_ps(), _mm512_cvtepi16_epi32(), and _mm512_slli_epi32().

◆ _mm512_dpbf16_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_dpbf16_ps ( __m512  __D,
__m512bh  __A,
__m512bh  __B 
)
static

Dot Product of BF16 Pairs Accumulated into Packed Single Precision.

This intrinsic corresponds to the VDPBF16PS instructions.

Parameters
__AA 512-bit vector of [32 x bfloat].
__BA 512-bit vector of [32 x bfloat].
__DA 512-bit vector of [16 x float].
Returns
A 512-bit vector of [16 x float] comes from Dot Product of __A, __B and __D

Definition at line 175 of file avx512bf16intrin.h.

References __D.

Referenced by _mm512_mask_dpbf16_ps(), and _mm512_maskz_dpbf16_ps().

◆ _mm512_mask_cvtne2ps_pbh()

static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtne2ps_pbh ( __m512bh  __W,
__mmask32  __U,
__m512  __A,
__m512  __B 
)
static

Convert Two Packed Single Data to One Packed BF16 Data.

This intrinsic corresponds to the VCVTNE2PS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
__BA 512-bit vector of [16 x float].
__WA 512-bit vector of [32 x bfloat].
__UA 32-bit mask value specifying what is chosen for each element. A 1 means conversion of __A or __B. A 0 means element from __W.
Returns
A 512-bit vector of [32 x bfloat] whose lower 256 bits come from conversion of __B, and higher 256 bits come from conversion of __A.

Definition at line 76 of file avx512bf16intrin.h.

References _mm512_cvtne2ps_pbh().

◆ _mm512_mask_cvtneps_pbh()

static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtneps_pbh ( __m256bh  __W,
__mmask16  __U,
__m512  __A 
)
static

Convert Packed Single Data to Packed BF16 Data.

This intrinsic corresponds to the VCVTNEPS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
__WA 256-bit vector of [16 x bfloat].
__UA 16-bit mask value specifying what is chosen for each element. A 1 means conversion of __A. A 0 means element from __W.
Returns
A 256-bit vector of [16 x bfloat] come from conversion of __A.

Definition at line 135 of file avx512bf16intrin.h.

◆ _mm512_mask_cvtpbh_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_cvtpbh_ps ( __m512  __S,
__mmask16  __U,
__m256bh  __A 
)
static

Convert Packed BF16 Data to Packed float Data using merging mask.

Parameters
__SA 512-bit vector of [16 x float]. Elements are copied from __S when the corresponding mask bit is not set.
__UA 16-bit mask.
__AA 256-bit vector of [16 x bfloat].
Returns
A 512-bit vector of [16 x float] come from convertion of __A

Definition at line 270 of file avx512bf16intrin.h.

References _mm512_castsi512_ps(), _mm512_cvtepi16_epi32(), and _mm512_mask_slli_epi32().

◆ _mm512_mask_dpbf16_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_dpbf16_ps ( __m512  __D,
__mmask16  __U,
__m512bh  __A,
__m512bh  __B 
)
static

Dot Product of BF16 Pairs Accumulated into Packed Single Precision.

This intrinsic corresponds to the VDPBF16PS instructions.

Parameters
__AA 512-bit vector of [32 x bfloat].
__BA 512-bit vector of [32 x bfloat].
__DA 512-bit vector of [16 x float].
__UA 16-bit mask value specifying what is chosen for each element. A 1 means __A and __B's dot product accumulated with __D. A 0 means __D.
Returns
A 512-bit vector of [16 x float] comes from Dot Product of __A, __B and __D

Definition at line 199 of file avx512bf16intrin.h.

References __D, and _mm512_dpbf16_ps().

◆ _mm512_maskz_cvtne2ps_pbh()

static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtne2ps_pbh ( __mmask32  __U,
__m512  __A,
__m512  __B 
)
static

Convert Two Packed Single Data to One Packed BF16 Data.

This intrinsic corresponds to the VCVTNE2PS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
__BA 512-bit vector of [16 x float].
__UA 32-bit mask value specifying what is chosen for each element. A 1 means conversion of __A or __B. A 0 means element is zero.
Returns
A 512-bit vector of [32 x bfloat] whose lower 256 bits come from conversion of __B, and higher 256 bits come from conversion of __A.

Definition at line 98 of file avx512bf16intrin.h.

References _mm512_cvtne2ps_pbh(), and _mm512_setzero_si512().

◆ _mm512_maskz_cvtneps_pbh()

static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtneps_pbh ( __mmask16  __U,
__m512  __A 
)
static

Convert Packed Single Data to Packed BF16 Data.

This intrinsic corresponds to the VCVTNEPS2BF16 instructions.

Parameters
__AA 512-bit vector of [16 x float].
__UA 16-bit mask value specifying what is chosen for each element. A 1 means conversion of __A. A 0 means element is zero.
Returns
A 256-bit vector of [16 x bfloat] come from conversion of __A.

Definition at line 154 of file avx512bf16intrin.h.

References _mm256_setzero_si256().

◆ _mm512_maskz_cvtpbh_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtpbh_ps ( __mmask16  __U,
__m256bh  __A 
)
static

Convert Packed BF16 Data to Packed float Data using zeroing mask.

Parameters
__UA 16-bit mask. Elements are zeroed out when the corresponding mask bit is not set.
__AA 256-bit vector of [16 x bfloat].
Returns
A 512-bit vector of [16 x float] come from convertion of __A

Definition at line 252 of file avx512bf16intrin.h.

References _mm512_castsi512_ps(), _mm512_maskz_cvtepi16_epi32(), and _mm512_slli_epi32().

◆ _mm512_maskz_dpbf16_ps()

static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_dpbf16_ps ( __mmask16  __U,
__m512  __D,
__m512bh  __A,
__m512bh  __B 
)
static

Dot Product of BF16 Pairs Accumulated into Packed Single Precision.

This intrinsic corresponds to the VDPBF16PS instructions.

Parameters
__AA 512-bit vector of [32 x bfloat].
__BA 512-bit vector of [32 x bfloat].
__DA 512-bit vector of [16 x float].
__UA 16-bit mask value specifying what is chosen for each element. A 1 means __A and __B's dot product accumulated with __D. A 0 means 0.
Returns
A 512-bit vector of [16 x float] comes from Dot Product of __A, __B and __D

Definition at line 223 of file avx512bf16intrin.h.

References __D, _mm512_dpbf16_ps(), and _mm512_setzero_si512().

◆ _mm_cvtsbh_ss()

static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss ( __bfloat16  __A)
static

Convert One BF16 Data to One Single Float Data.

This intrinsic does not correspond to a specific instruction.

Parameters
__AA bfloat data.
Returns
A float data whose sign field and exponent field keep unchanged, and fraction field is extended to 23 bits.

Definition at line 36 of file avx512bf16intrin.h.

__attribute__
short __m512bh __attribute__((__vector_size__(64), __aligned__(64)))
Definition: avx512bf16intrin.h:16