CUDASIMDTypes.jl

struct BFloat16x2

A SIMD type holding 2 BFloat16 in a combined 32-bit value.

struct Float16x2

A SIMD type holding 2 Float16 in a combined 32-bit value.

struct Int16x2

A SIMD type holding 2 16-bit integers in a combined 32-bit value.

struct Int2x16

A SIMD type holding 16 2-bit integers in a combined 32-bit value.

struct Int2x4

A SIMD type holding 4 2-bit integers in a combined 8-bit value.

struct Int4x2

A SIMD type holding 2 4-bit integers in a combined 8-bit value.

struct Int4x8

A SIMD type holding 8 4-bit integers in a combined 32-bit value.

struct Int8x4

A SIMD type holding 4 8-bit integers in a combined 32-bit value.

bitifelse(cond, x, y)

Bitwise version of ifelse.

For each bit of the output, the respective bit in cond determines whether the respective bit of x or of y is selected.

d = cvt_pack_s16(a::Int32, b::Int32)
d::UInt32
d[1] = sat(b)
d[2] = sat(a)

d = cvt_pack_s8(a::Int32, b::Int32, c::UInt32)
d::UInt32
d[1] = sat(b)
d[2] = sat(a)
d[3] = c[1]
d[4] = c[2]

d = cvt_pack_s8(a::Int32, b::Int32)
d::UInt32
d[1] = sat(b)
d[2] = sat(a)
d[3] = 0
d[4] = 0

d = dp4a(a::UInt32, b::UInt32, c::Int32)
d::Int32
d = a[1] * b[1] + a[2] * b[2] + a[3] * b[3] + a[4] * b[4] + c

lop3(a, b, c, lut)

Arbitrary logical operation on 3 inputs.

Call the PTX prmt instruction. This computes a bitwise logical operation on the inputs a, b, and c.

See make_lop3_lut for creating the look-up table lut.

make_lop3_lut(f)

Create a look-up table for lop3.

muladd_sat(a, b, c)

Saturating fused multiply-add a * b + c.

Each element of the result is clamped to [0, 1]; NaN results are flushed to +0.

prmt(a, b, op)

Permute bytes bytes from a pair of inputs.

Call the PTX prmt instruction. This picks four arbitrary bytes from the input values a and b.