IPU C++ intrinsics

Functions that target single IPU instructions. 

#include <ipu_memory_intrinsics>

These intrinsic functions target single IPU instructions and may be used in C++ IPU code. Each function is named after the instruction it targets. Float instructions that have a type (for example, f16v2) often omit this prefix in their function name. For example, use cmpeq() to target any of the cmpeq instructions, such as f16v2cmpeq.

Refer to the Tile Vertex Instruction Set Architecture for Mk2 IPUs for more detailed information on the instructions targeted by these intrinsics.

Defines

IPU_INTRINSICS_INCLUDED

Variables

template<typename T0, typename T1>
static constexpr auto both_integral = (std::is_integral<T0>::value && std::is_integral<T1>::value)
namespace ipu

IPU intrinsic functions.

Functions

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned andc(T0 src0, T1 src1)

Targets the andc instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value, can be a 12-bit constant.

Returns

The bitwise logical and of src0 and the negated value of src1 of type unsigned.

inline float andc(float src0, float src1)

Targets the andc instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

The bitwise logical and of src0 and the negated value of src1 of type float.

inline float2 andc(float2 src0, float2 src1)

Targets the andc64 instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The bitwise logical and of src0 and the negated value of src1 of type float2.

inline unsigned bitrev8(unsigned src)

Targets the bitrev8 instruction.

Parameters

src – A value of type unsigned.

Returns

A result of type unsigned that is equivalent to the value of src with the bit order of each byte reversed.

inline unsigned cms(unsigned src)

Targets the cms instruction.

Parameters

src – A value of type unsigned.

Returns

The number of higher order bits in src that match the sign bit (bit 31), as an unsigned.

inline float2 roll32(float2 src0, float2 src1)

Targets the roll32 instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The result of a SIMD roll permutation on the 4 32-bit float values across src0 and src1, as a float2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 1 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned roll8l(T0 src0, T1 src1)

Targets the roll8l instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The result of a SIMD roll-left permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 6 | 5 | 4 | 3 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned roll8r(T0 src0, T1 src1)

Targets the roll8r instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The result of a SIMD roll-right permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 4 | 3 | 2 | 1 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned shuf8x8hi(T0 src0, T1 src1)

Targets the shuf8x8hi instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The upper word of a SIMD shuffle permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 7 | 3 | 6 | 2 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned shuf8x8lo(T0 src0, T1 src1)

Targets the shuf8x8lo instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The lower word of a SIMD shuffle permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 5 | 1 | 4 | 0 |

inline float2 sort4x32hi(float2 src0, float2 src1)

Targets the sort4x32hi instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The result of a SIMD sort permutation on the 4 32-bit float values across src0 and src1, as a float2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 3 | 1 |

inline float2 sort4x32lo(float2 src0, float2 src1)

Targets the sort4x32lo instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The result of a SIMD sort permutation on the 4 32-bit float values across src0 and src1, as a float2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 0 |

inline unsigned sort8(unsigned src)

Targets the sort8 instruction.

Parameters

src – A value of type unsigned.

Returns

The result of a SIMD sort8 permutation on the 4 8-bit values in src, as an unsigned. src -> Result | 3 | 2 | 1 | 0 | | 3 | 1 | 2 | 0 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned sort8x8hi(T0 src0, T1 src1)

Targets the sort8x8hi instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The upper word of the result of a SIMD sort8 permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 7 | 5 | 3 | 1 |

template<typename T0, typename T1, typename = std::enable_if_t<both_integral<T0, T1>>>
inline unsigned sort8x8lo(T0 src0, T1 src1)

Targets the sort8x8lo instruction.

Parameters

  • src0 – An integer value.

  • src1 – An integer value.

Returns

The lower word of the result of a SIMD sort8 permutation on the 8 8-bit values across src0 and src1, as an unsigned. src0 src1 -> Result | 7 | 6 | 5 | 4 | | 3 | 2 | 1 | 0 | | 6 | 4 | 2 | 0 |

inline unsigned swap8(unsigned src)

Targets the sort8 instruction.

Parameters

src – A value of type unsigned.

Returns

The result of a SIMD swap permutation on the 4 8-bit values in src, as an unsigned. src -> Result | 3 | 2 | 1 | 0 | | 2 | 3 | 0 | 1 |

inline half2 absadd(half2 src0, half2 src1)

Targets the f16v2absadd instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The result of an element-wise addition of absolute values in src0 and src1.

inline half4 absadd(half4 src0, half4 src1)

Targets the f16v4absadd instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

The result of an element-wise addition of absolute values in src0 and src1.

inline float2 absadd(float2 src0, float2 src1)

Targets the f32v2absadd instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The result of an element-wise addition of absolute values in src0 and src1.

inline float absadd(float src0, float src1)

Targets the f32absadd instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

The result of a scalar addition of absolute values src0 and src1.

inline half2 absmax(half2 src0, half2 src1)

Targets the f16v2absmax instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The element-wise maximum of absolute values in src0 and src1.

inline half4 absmax(half4 src0, half4 src1)

Targets the f16v4absmax instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

The element-wise maximum of absolute values in src0 and src1.

inline float2 absmax(float2 src0, float2 src1)

Targets the f32v2absmax instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The element-wise maximum of absolute values in src0 and src1.

inline float absmax(float src0, float src1)

Targets the f32absmax instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

The maximum of absolute values src0 and src1.

inline half2 max(half2 src0, half2 src1)

Targets the f16v2max instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The element-wise maximum of src0 and src1.

inline half4 max(half4 src0, half4 src1)

Targets the f16v4max instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

The element-wise maximum of src0 and src1.

inline float2 max(float2 src0, float2 src1)

Targets the f32v2max instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The element-wise maximum of src0 and src1.

inline float max(float src0, float src1)

Targets the f32max instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

The maximum of src0 and src1.

inline half2 maxc(half4 src)

Targets the f16v4maxc instruction.

Parameters

src – A value of type half4.

Returns

The 2x2 lateral maximum of src. The 0th element in the result vector is the maximum of src[0] and src[1], and the 1st element is the maximum of src[2] and src[3].

inline half2 min(half2 src0, half2 src1)

Targets the f16v2min instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The element-wise minimum of src0 and src1.

inline half4 min(half4 src0, half4 src1)

Targets the f16v4min instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

The element-wise minimum of src0 and src1.

inline float2 min(float2 src0, float2 src1)

Targets the f32v2min instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The element-wise minimum of src0 and src1.

inline float min(float src0, float src1)

Targets the f32min instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

The minimum of src0 and src1.

inline half2 clamp(half2 src0, half2 src1)

Targets the f16v2clamp instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The min-of-maximum result of src0 and src1, of type half2. The first element is the median value of the first element of src0 and the two elements in src1. The second element is the median of the second element of src0 and the two elements in src1.

inline half4 clamp(half4 src0, half2 src1)

Targets the f16v4clamp instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half2.

Returns

The min-of-maximum result of src0 and src1, of type half4. Each element is the median of the element in src0 at the same index, and the two values in src1.

inline float2 clamp(float2 src0, float2 src1)

Targets the f32v2clamp instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The min-of-maximum result of src0 and src1, of type float2. The first element is the median of the first element of src0 and the two elements in src1. The second element is the median of the second element of src0 and the two elements in src1.

inline float clamp(float src0, float2 src1)

Targets the f32clamp instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float2.

Returns

The median of src0 and the two elements in src1.

inline void cmac(half2 src0, half2 src1)

Targets the f16v2cmac instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

inline void cmac(half4 src0, half4 src1)

Targets the f16v4cmac instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

inline half2 exp(half2 src)

Targets the f16v2exp instruction.

Parameters

src – A value of type half2.

Returns

A vector of the results of e^X of the two elements in src.

inline float exp(float src)

Targets the f32exp instruction.

Parameters

src – A value of type float.

Returns

The result of e^{src}.

inline half2 exp2(half2 src)

Targets the f16v2exp instruction.

Parameters

src – A value of type half2.

Returns

A vector of the results of 2^X of the two elements in src.

inline float exp2(float src)

Targets the f32exp instruction.

Parameters

src – A value of type float.

Returns

The result of 2^{src}.

inline half2 log2(half2 src)

Targets the f16v2log2 instruction.

Parameters

src – A value of type half2.

Returns

A vector of the results of the log (base 2) of the two elements in src.

inline float log2(float src)

Targets the f32ln instruction.

Parameters

src – A value of type half2.

Returns

The result of the log (base 2) of src.

inline half2 tanh(half2 src)

Targets the f16v2tanh instruction.

Parameters

src – A value of type half2.

Returns

The result of tanh(src)`.

inline float tanh(float src)

Targets the f32tanh instruction.

Parameters

src – A value of type float.

Returns

The result of tanh(src)`.

inline half2 ln(half2 src)

Targets the f16v2ln instruction.

Parameters

src – A value of type half2.

Returns

A vector of the results of the natural log of the two elements in src.

inline float ln(float src)

Targets the f32ln instruction.

Parameters

src – A value of type half2.

Returns

The result of the natural log of src.

inline float2 axpy(float2 src0, float2 src1)

Targets the f32v2axpy instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

The single precision two-element vector res = a*src0 + src1. The scalar multiplicand a is provided by the internal state element $TAS.

inline half2 f16v2grand()

Targets the f16v2grand instruction.

Returns

Gaussian distribution, two-element half-precision random vector.

inline float2 f32v2grand()

Targets the f32v2grand instruction.

Returns

Gaussian distribution, two-element single-precision random vector.

inline half4 rmask(half4 src0, float src1)

Targets the f16v4rmask instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type float.

Returns

The result is a masked version of src0, with each element of the input being individually masked with the probability specified by the bottom 17-bits of src1:

  • if src1[16] == 1, no masking is applied;

  • if src1[16:0] == 0, the result is a zero vector;

  • otherwise each element is individually unmasked with probability src1[15:0] / 65536. PRNG is used by this instruction to generate 4 x 16-bit random values from the discrete uniform distribution.

inline float2 rmask(float2 src0, float src1)

Targets the f32v2rmask instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float.

Returns

The result is a masked version of src0, with each element of the input being individually masked with the probability specified by the bottom 17-bits of src1:

  • if src1[16] == 1, no masking is applied;

  • if src1[16:0] == 0, the result is a zero vector;

  • otherwise each element is individually unmasked with probability src1[15:0] / 65536. PRNG is used by this instruction to generate 2 x 16-bit random values from the discrete uniform distribution.

inline half2 sigm(half2 src)

Targets the f16v2sigm instruction.

Parameters

src – A value of type half2.

Returns

The result of an element-wise application of the sigmoid function on src.

inline float sigm(float src)

Targets the f32sigm instruction.

Parameters

src – A value of type float.

Returns

The result of an element-wise application of the sigmoid function on src.

inline float sum(half2 src)

Targets the f16v2sum instruction.

Parameters

src – A value of type half2.

Returns

The sum of the two elements in src as a float.

inline float2 sum(half4 src)

Targets the f16v4sum instruction.

Parameters

src – A value of type half2.

Returns

The 2x2 lateral summation of the elements in src as a float2. The first element is the sum of src[0] and src[1], the second element is the sum of src[2] and src[3].

inline half2 cmpeq(half2 src0, half2 src1)

Targets the f16v2cmpeq instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise equality test of src0 and src1. If src0[i] == src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmpeq(half4 src0, half4 src1)

Targets the f16v4cmpeq instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise equality test of src0 and src1. If src0[i] == src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmpeq(float2 src0, float2 src1)

Targets the f32v2cmpeq instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise equality test of src0 and src1. If src0[i] == src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmpeq(float src0, float src1)

Targets the f32cmpeq instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Equality test of src0 and src1. If src0 == src1 the result will be 0xffff, and 0x0000 otherwise.

inline half2 cmpge(half2 src0, half2 src1)

Targets the f16v2cmpge instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise greater-than-or-equal-to test of src0 and src1. If src0[i] >= src1[i] the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmpge(half4 src0, half4 src1)

Targets the f16v4cmpge instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise greater-than-or-equal-to test of src0 and src1. If src0[i] >= src1[i] the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmpge(float2 src0, float2 src1)

Targets the f32v2cmpge instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise greater-than-or-equal-to test of src0 and src1. If src0[i] >= src1[i] the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmpge(float src0, float src1)

Targets the f32cmpge instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Greater-than-or-equal-to test of src0 and src1. If src0 >= src1 the result will be 0xffff, and 0x0000 otherwise.

inline half2 cmpgt(half2 src0, half2 src1)

Targets the f16v2cmpgt instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise greater-than test of src0 and src1. If src0 > src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmpgt(half4 src0, half4 src1)

Targets the f16v4cmpgt instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise greater-than test of src0 and src1. If src0 > src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmpgt(float2 src0, float2 src1)

Targets the f32v2cmpgt instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise greater-than test of src0 and src1. If src0 > src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmpgt(float src0, float src1)

Targets the f32cmpgt instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Greater-than test of src0 and src1. If src0 > src1 the result will be 0xffff, and 0x0000 otherwise.

inline half2 cmple(half2 src0, half2 src1)

Targets the f16v2cmple instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise less-than-or-equal-to test of src0 and src1. If src0 <= src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmple(half4 src0, half4 src1)

Targets the f16v4cmple instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise less-than-or-equal-to test of src0 and src1. If src0 <= src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmple(float2 src0, float2 src1)

Targets the f32v2cmple instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise less-than-or-equal-to test of src0 and src1. If src0 <= src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmple(float src0, float src1)

Targets the f32cmple instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Less-than-or-equal-to test of src0 and src1. If src0 <= src1 the result will be 0xffff, and 0x0000 otherwise.

inline half2 cmplt(half2 src0, half2 src1)

Targets the f16v2cmplt instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise less-than test of src0 and src1. If src0 < src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmplt(half4 src0, half4 src1)

Targets the f16v4cmplt instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise less-than test of src0 and src1. If src0 < src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmplt(float2 src0, float2 src1)

Targets the f32v2cmplt instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise less-than test of src0 and src1. If src0 < src1 the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmplt(float src0, float src1)

Targets the f32cmplt instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Less-than test of src0 and src1. If src0 < src1 the result will be 0xffff, and 0x0000 otherwise.

inline half2 cmpne(half2 src0, half2 src1)

Targets the f16v2cmpne instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

Element-wise inequality test of src0 and src1. If src0[i] != src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline half4 cmpne(half4 src0, half4 src1)

Targets the f16v4cmpne instruction.

Parameters

  • src0 – A value of type half4.

  • src1 – A value of type half4.

Returns

Element-wise inequality test of src0 and src1. If src0[i] != src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float2 cmpne(float2 src0, float2 src1)

Targets the f32v2cmpne instruction.

Parameters

  • src0 – A value of type float2.

  • src1 – A value of type float2.

Returns

Element-wise inequality test of src0 and src1. If src0[i] != src1[i], the result vector element at index i will be 0xffff, and 0x0000 otherwise.

inline float cmpne(float src0, float src1)

Targets the f32cmpne instruction.

Parameters

  • src0 – A value of type float.

  • src1 – A value of type float.

Returns

Inequality test of src0 and src1. If src0 != src1 the result will be 0xffff, and 0x0000 otherwise.

inline float to_float(half src)

Conversion from a half-precision float to single-precision.

Targets the f16tof32 instruction.

Parameters

src – A value of type half.

Returns

The half-precision float value src converted to single-precision.

inline float2 to_float(half2 src)

Two-element vector conversion from half-precision to single-precision.

Targets the f16v2tof32 instruction.

Parameters

src – A vector of type half2.

Returns

The vector of two half-precision float values src converted to single-precision.

inline float to_float(unsigned src)

Conversion from an unsigned integer to a single-precision float.

Targets the f32fromui32 instruction.

Parameters

src – An unsigned integer.

Returns

The unsigned integer src converted to a single-precision float value.

inline float to_float(int src)

Conversion from a signed integer to a single-precision float.

Targets the f32fromi32 instruction.

Parameters

src – A signed integer.

Returns

The signed integer src converted to a single-precision float value.

inline float to_float_su(unsigned src)

Symmetric, unbiased conversion from an unsigned integer to a single-precision float.

Targets the f32sufromui instruction.

Parameters

src – An unsigned integer.

Returns

The unsigned integer src converted to a single-precision float value via symmetric, unbiased conversion.

inline float2 to_float_su(uint2 src)

Symmetric, unbiased conversion from a vector of two unsigned integers to single-precision floats.

Targets the f32v2sufromui instruction.

Parameters

src – A vector of two unsigned integers.

Returns

The vector of two unsigned integers src converted to single-precision floats via symmetric, unbiased conversion.

inline half to_half(float src)

Conversion from a single-precision float to half-precision.

Targets the f32tof16 instruction.

Parameters

src – A value of type float.

Returns

The single-precision float value src converted to half-precision.

inline half2 to_half(float2 src)

Conversion from a vector of two single-precision floats to half-precision.

Targets the f32v2tof16 instruction.

Parameters

src – A vector of two single-precision float values.

Returns

The vector of two single-precision float values src converted to half-precision.

inline half2 to_half_su(ushort2 src)

Symmetric, unbiased conversion from a vector of two unsigned 16-bit integers to half-precision floats.

Targets the f16v2sufromui instruction.

Parameters

src – A two-element vector of unsigned 16-bit integers.

Returns

The vector of two unsigned 16-bit integers src converted to half precision floats via symmetric, unbiased conversion.

inline half4 to_half_su(ushort4 src)

Symmetric, unbiased conversion from a vector of four unsigned 16-bit integers to half-precision floats.

Targets the f16v4sufromui instruction.

Parameters

src – A four-element vector of unsigned 16-bit integers.

Returns

The vector of four unsigned 16-bit integers src converted to half precision floats via symmetric, unbiased conversion.

inline unsigned to_uint(float src)

Conversion from a single-precision float to an unsigned integer.

Targets the f32toui32 instruction.

Parameters

src – A float value.

Returns

The single-precision float src converted to an unsigned integer.

inline int to_int(float src)

Conversion from a single-precision float to a signed integer.

Targets the f32toi32 instruction.

Parameters

src – A float value.

Returns

The single-precision float src converted to a signed integer.

inline float f16tof32(half src)

Targets the f16tof32 instruction.

Parameters

src – A value of type half.

Returns

The half-precision float value src converted to single-precision.

inline float2 f16v2tof32(half2 src)

Targets the f16v2tof32 instruction.

Parameters

src – A vector of type half2.

Returns

The vector of two half-precision float values src converted to single-precision.

inline half2 f16v2sufromui(half2 src)

Targets the f16v2sufromui instruction.

Parameters

src – A two-element vector of unsigned 16-bit integers, as a variable of type half2. Note: this builtin directly targets the f16v2sufromui instruction, whose operand and result are both in the floating point register file. To achieve this conversion from a ushort2 type, see the to_half function in this header.

Returns

The vector of two unsigned 16-bit integers src converted to half precision floats via symmetric, unbiased conversion.

inline half4 f16v4sufromui(half4 src)

Targets the f16v4sufromui instruction.

Parameters

src – A four-element vector of unsigned 16-bit integers, as a variable of type half4. Note: this builtin directly targets the f16v4sufromui instruction, whose operand and result are both in the floating point register file. To achieve this conversion from a ushort4 type, see the to_half function in this header.

Returns

The vector of four unsigned 16-bit integers src converted to half precision floats via symmetric, unbiased conversion.

inline float f32fromi32(float src)

Targets the f32fromi32 instruction.

Parameters

src – A signed integer as a float variable. Note: this builtin directly targets thef32fromi32 instruction, whose operand and result are both in the floating point register file. To achieve this conversion from an int type, see the to_float function in this header.

Returns

The signed integer src converted to a single-precision float value.

inline float f32fromui32(float src)

Targets the f32fromui32 instruction.

Parameters

src – An unsigned integer as a float variable. Note: this builtin directly targets thef32fromui32 instruction, whose operand and result are both in the floating point register file. To achieve this conversion from an unsigned type, see the to_float function in this header.

Returns

The unsigned integer src converted to a single-precision float value.

inline float f32sufromui(float src)

Targets the f32sufromui instruction.

Parameters

src – An unsigned integer as a float variable. Note: this builtin directly targets thef32sufromui instruction, whose operand and result are both in the floating point register file. To achieve this conversion from an unsigned type, see the to_float_su function in this header.

Returns

The unsigned integer src converted to a single-precision float value via symmetric, unbiased conversion.

inline half f32tof16(float src)

Targets the f32tof16 instruction.

Parameters

src – A value of type float.

Returns

The single-precision float value src converted to half-precision.

inline float f32toi32(float src)

Targets the f32toi32 instruction.

Parameters

src – A float value.

Returns

The single-precision float src converted to an integer, as a float-type variable. Note: this builtin directly targets thef32toi32 instruction, whose operand and result are both in the floating point register file. To achieve this conversion to an int type, see the to_int function in this header.

inline float f32toui32(float src)

Targets the f32toui32 instruction.

Parameters

src – A float value.

Returns

The single-precision float src converted to an unsigned integer, as a float-type variable. Note: this builtin directly targets thef32toui32 instruction, whose operand and result are both in the floating point register file. To achieve this conversion to an unsigned type, see the to_uint function in this header.

inline float2 f32v2sufromui(float2 src)

Targets the f32v2sufromui instruction.

Parameters

src – A two-element vector of unsigned integers, as a variable of type type float2. Note: this builtin directly targets the f32v2sufromui instruction, whose operand and result are both in the floating point register file. To achieve this conversion from a uint2 type, see the to_float function in this header.

Returns

The vector of two unsigned integers src converted to single precision floats via symmetric, unbiased conversion.

inline half2 f32v2tof16(float2 src)

Targets the f32v2tof16 instruction.

Parameters

src – A vector of two single-precision float values.

Returns

The vector of two single-precision float values src converted to half-precision.

inline unsigned clz(int src)

Targets the clz instruction.

Parameters

src – A value of type int.

Returns

The number of higher bits in src that are zero.

inline unsigned popc(int src)

Targets the popc instruction.

Parameters

src – A value of type int.

Returns

The number of set bits in src.

inline short2 roll16(short2 src0, short2 src1)

Targets the roll16 instruction.

Parameters

  • src0 – A value of type short2.

  • src1 – A value of type short2.

Returns

The result of a SIMD roll permutation on the 4 16-bit values across src0 and src1, as a short2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 1 |

inline ushort2 roll16(ushort2 src0, ushort2 src1)

Targets the roll16 instruction.

Parameters

  • src0 – A value of type ushort2.

  • src1 – A value of type ushort2.

Returns

The result of a SIMD roll permutation on the 4 16-bit values across src0 and src1, as a ushort2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 1 |

inline half2 roll16(half2 src0, half2 src1)

Targets the roll16 instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The result of a SIMD roll permutation on the 4 16-bit values across src0 and src1, as a half2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 1 |

inline short2 sort4x16hi(short2 src0, short2 src1)

Targets the sort4x16hi instruction.

Parameters

  • src0 – A value of type short2.

  • src1 – A value of type short2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a short2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 3 | 1 |

inline ushort2 sort4x16hi(ushort2 src0, ushort2 src1)

Targets the sort4x16hi instruction.

Parameters

  • src0 – A value of type ushort2.

  • src1 – A value of type ushort2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a ushort2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 3 | 1 |

inline half2 sort4x16hi(half2 src0, half2 src1)

Targets the sort4x16hi instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a half2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 3 | 1 |

inline short2 sort4x16lo(short2 src0, short2 src1)

Targets the sort4x16lo instruction.

Parameters

  • src0 – A value of type short2.

  • src1 – A value of type short2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a short2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 0 |

inline ushort2 sort4x16lo(ushort2 src0, ushort2 src1)

Targets the sort4x16lo instruction.

Parameters

  • src0 – A value of type ushort2.

  • src1 – A value of type ushort2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a ushort2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 0 |

inline half2 sort4x16lo(half2 src0, half2 src1)

Targets the sort4x16lo instruction.

Parameters

  • src0 – A value of type half2.

  • src1 – A value of type half2.

Returns

The result of a SIMD sort permutation on the 4 16-bit values across src0 and src1, as a half2. src0 src1 -> Result | 3 | 2 | | 1 | 0 | | 2 | 0 |