pub struct Avx2(/* private fields */);Expand description
AVX2: 256-bit registers as two 128-bit lanes; only constructible when the CPU has AVX2
Implementations§
Trait Implementations§
impl Copy for Avx2
Source§impl Simd for Avx2
impl Simd for Avx2
type V = __m256i
fn zero(self) -> __m256i
fn splat8(self, v: u8) -> __m256i
fn splat16(self, v: u16) -> __m256i
fn splat32(self, v: u32) -> __m256i
fn splat64(self, v: u64) -> __m256i
fn splatf(self, v: f32) -> __m256i
fn store(self, b: &mut [u8], v: __m256i)
Source§fn store_px(self, b: &mut [u8], v: __m256i)
fn store_px(self, b: &mut [u8], v: __m256i)
pixel lanes saturated to bytes (packuswb), 4 * PX bytes
Source§fn load_raw(self, b: &[u8]) -> __m256i
fn load_raw(self, b: &[u8]) -> __m256i
4 * PX bytes as they are (“raw” pixels: SSE2 the low 8 bytes, AVX2 the low 16)
fn store_raw(self, b: &mut [u8], v: __m256i)
Source§fn load_lane32(self, b: &[u8]) -> __m256i
fn load_lane32(self, b: &[u8]) -> __m256i
4 bytes into the low dword of every 128-bit lane (BYTES / 4 bytes), the rest 0
fn store_i32(self, b: &mut [i32], v: __m256i)
fn to_i32(self, v: __m256i) -> i32
fn add8(self, a: __m256i, b: __m256i) -> __m256i
fn sub8(self, a: __m256i, b: __m256i) -> __m256i
fn addus8(self, a: __m256i, b: __m256i) -> __m256i
fn subus8(self, a: __m256i, b: __m256i) -> __m256i
fn maxu8(self, a: __m256i, b: __m256i) -> __m256i
fn minu8(self, a: __m256i, b: __m256i) -> __m256i
fn cmpeq8(self, a: __m256i, b: __m256i) -> __m256i
fn avgu8(self, a: __m256i, b: __m256i) -> __m256i
fn unpacklo8(self, a: __m256i, b: __m256i) -> __m256i
fn unpackhi8(self, a: __m256i, b: __m256i) -> __m256i
fn packus16(self, a: __m256i, b: __m256i) -> __m256i
fn add16(self, a: __m256i, b: __m256i) -> __m256i
fn sub16(self, a: __m256i, b: __m256i) -> __m256i
fn subus16(self, a: __m256i, b: __m256i) -> __m256i
fn mullo16(self, a: __m256i, b: __m256i) -> __m256i
fn mulhiu16(self, a: __m256i, b: __m256i) -> __m256i
fn mins16(self, a: __m256i, b: __m256i) -> __m256i
fn maxs16(self, a: __m256i, b: __m256i) -> __m256i
fn cmpeq16(self, a: __m256i, b: __m256i) -> __m256i
fn cmpgts16(self, a: __m256i, b: __m256i) -> __m256i
fn srl16<const N: i32>(self, a: __m256i) -> __m256i
fn sll16<const N: i32>(self, a: __m256i) -> __m256i
Source§fn shuf16<const IMM: i32>(self, a: __m256i) -> __m256i
fn shuf16<const IMM: i32>(self, a: __m256i) -> __m256i
pshuflw + pshufhw with the same immediate (per 64-bit half)
fn unpacklo16(self, a: __m256i, b: __m256i) -> __m256i
fn unpackhi16(self, a: __m256i, b: __m256i) -> __m256i
fn packs32(self, a: __m256i, b: __m256i) -> __m256i
Source§fn madd16(self, a: __m256i, b: __m256i) -> __m256i
fn madd16(self, a: __m256i, b: __m256i) -> __m256i
pmaddwd: i32 lanes = a0 * b0 + a1 * b1 of the signed 16-bit pairs
fn add32(self, a: __m256i, b: __m256i) -> __m256i
fn sub32(self, a: __m256i, b: __m256i) -> __m256i
fn cmpeq32(self, a: __m256i, b: __m256i) -> __m256i
fn cmpgts32(self, a: __m256i, b: __m256i) -> __m256i
Source§fn movemask32(self, a: __m256i) -> u32
fn movemask32(self, a: __m256i) -> u32
bit i = top bit of dword i (BYTES / 4 bits)
Source§fn shufps<const IMM: i32>(self, a: __m256i, b: __m256i) -> __m256i
fn shufps<const IMM: i32>(self, a: __m256i, b: __m256i) -> __m256i
shufps on integer registers: dwords 0, 1 from a, 2, 3 from b (per 128-bit lane)
fn srl32<const N: i32>(self, a: __m256i) -> __m256i
fn sll32<const N: i32>(self, a: __m256i) -> __m256i
fn sra32<const N: i32>(self, a: __m256i) -> __m256i
Source§fn bsll<const N: i32>(self, a: __m256i) -> __m256i
fn bsll<const N: i32>(self, a: __m256i) -> __m256i
pslldq / psrldq: byte shifts per 128-bit lane
fn bsrl<const N: i32>(self, a: __m256i) -> __m256i
fn and(self, a: __m256i, b: __m256i) -> __m256i
fn or(self, a: __m256i, b: __m256i) -> __m256i
fn xor(self, a: __m256i, b: __m256i) -> __m256i
fn cvt_i32_f32(self, a: __m256i) -> __m256i
Source§fn cvtt_f32_i32(self, a: __m256i) -> __m256i
fn cvtt_f32_i32(self, a: __m256i) -> __m256i
truncating (0x80000000 for NaN / out of range)
Source§fn cvtr_f32_i32(self, a: __m256i) -> __m256i
fn cvtr_f32_i32(self, a: __m256i) -> __m256i
rounding to nearest even (cvtps2dq)
fn addf(self, a: __m256i, b: __m256i) -> __m256i
fn subf(self, a: __m256i, b: __m256i) -> __m256i
fn mulf(self, a: __m256i, b: __m256i) -> __m256i
fn divf(self, a: __m256i, b: __m256i) -> __m256i
fn splatd(self, v: f64) -> __m256i
fn addd(self, a: __m256i, b: __m256i) -> __m256i
fn subd(self, a: __m256i, b: __m256i) -> __m256i
fn muld(self, a: __m256i, b: __m256i) -> __m256i
fn divd(self, a: __m256i, b: __m256i) -> __m256i
fn sqrtd(self, a: __m256i) -> __m256i
fn mind(self, a: __m256i, b: __m256i) -> __m256i
Source§fn cmpltd(self, a: __m256i, b: __m256i) -> __m256i
fn cmpltd(self, a: __m256i, b: __m256i) -> __m256i
cmpltpd: all ones where a < b (false when unordered)
Source§fn cvt_i32_f64(self, a: __m256i) -> __m256i
fn cvt_i32_f64(self, a: __m256i) -> __m256i
cvtdq2pd: the first BYTES / 8 i32 lanes to f64
Source§fn cvtt_f64_i32(self, a: __m256i) -> __m256i
fn cvtt_f64_i32(self, a: __m256i) -> __m256i
cvttpd2dq: f64 lanes to the first BYTES / 8 i32 lanes (truncating, 0x80000000 when
out of range), the rest 0
Source§fn store_i32_half(self, b: &mut [i32], v: __m256i)
fn store_i32_half(self, b: &mut [i32], v: __m256i)
the first BYTES / 8 i32 lanes
Source§fn sra16<const N: i32>(self, a: __m256i) -> __m256i
fn sra16<const N: i32>(self, a: __m256i) -> __m256i
psraw: arithmetic right shift of the i16 lanes
Source§fn mulhis16(self, a: __m256i, b: __m256i) -> __m256i
fn mulhis16(self, a: __m256i, b: __m256i) -> __m256i
pmulhw: high 16 bits of the signed 16-bit products
Source§fn packs16(self, a: __m256i, b: __m256i) -> __m256i
fn packs16(self, a: __m256i, b: __m256i) -> __m256i
packsswb: i16 lanes of a, then b, saturated to i8
fn store_i16(self, b: &mut [i16], v: __m256i)
Source§fn store_lanes2(self, b: &mut [u8], lo: __m256i, hi: __m256i)
fn store_lanes2(self, b: &mut [u8], lo: __m256i, hi: __m256i)
2 * BYTES bytes: per 128-bit lane k, lane k of lo then lane k of hi (SSE2: lo, hi;
AVX2: vperm2i128) – 2 x 4 pixels per lane in pixel order
Source§fn load_lanes2(self, b: &[u8]) -> (__m256i, __m256i)
fn load_lanes2(self, b: &[u8]) -> (__m256i, __m256i)
the inverse of store_lanes2: 2 * BYTES bytes -> (lo, hi)
Source§fn load_i16_lanes(self, a: &[i16], b: &[i16]) -> __m256i
fn load_i16_lanes(self, a: &[i16], b: &[i16]) -> __m256i
8 i16 per 128-bit lane from separate sources: lane 0 from a, lane 1 from b (SSE2: a)
Source§fn store_raw_lanes(self, a: &mut [u8], b: &mut [u8], v: __m256i)
fn store_raw_lanes(self, a: &mut [u8], b: &mut [u8], v: __m256i)
the low 8 bytes of every 128-bit lane to separate places: lane 0 to a, lane 1 to b
(SSE2: a)
Source§fn sadu8(self, a: __m256i, b: __m256i) -> __m256i
fn sadu8(self, a: __m256i, b: __m256i) -> __m256i
(main agent) psadbw: per 8-byte group the sum of |a - b| (u8) in its u64 lane
fn minf(self, a: __m256i, b: __m256i) -> __m256i
fn store_u16(self, b: &mut [u16], v: __m256i)
Source§fn load_u16_lo(self, b: &[u16]) -> __m256i
fn load_u16_lo(self, b: &[u16]) -> __m256i
movq: 4 u16 values into the low 64 bits (the rest 0) / the low 64 bits to 4 u16 values
fn store_u16_lo(self, b: &mut [u16], v: __m256i)
Source§fn srl16x(self, a: __m256i, n: u32) -> __m256i
fn srl16x(self, a: __m256i, n: u32) -> __m256i
psrlw with a register count: logical right shift of the 16-bit lanes by n (n > 15: 0)
fn store_f32(self, b: &mut [f32], v: __m256i)
Source§fn adds8(self, a: __m256i, b: __m256i) -> __m256i
fn adds8(self, a: __m256i, b: __m256i) -> __m256i
paddsb: signed saturating add of the i8 lanes
Source§fn subs8(self, a: __m256i, b: __m256i) -> __m256i
fn subs8(self, a: __m256i, b: __m256i) -> __m256i
psubsb: signed saturating subtract of the i8 lanes
Source§fn cmpltf(self, a: __m256i, b: __m256i) -> __m256i
fn cmpltf(self, a: __m256i, b: __m256i) -> __m256i
cmpltps: all ones where a < b (false when unordered)
Source§fn load_f32_lanes(self, a: &[f32], b: &[f32]) -> __m256i
fn load_f32_lanes(self, a: &[f32], b: &[f32]) -> __m256i
4 f32 per 128-bit lane from separate sources: lane 0 from a, lane 1 from b (SSE2: a)
Source§fn store_f32_lanes(self, a: &mut [f32], b: &mut [f32], v: __m256i)
fn store_f32_lanes(self, a: &mut [f32], b: &mut [f32], v: __m256i)
the 4 f32 of every 128-bit lane to separate places: lane 0 to a, lane 1 to b (SSE2: a)
Source§fn packssdw(self, a: __m256i, b: __m256i) -> __m256i
fn packssdw(self, a: __m256i, b: __m256i) -> __m256i
packs32 (packssdw), on AVX2 without the clamps
Source§fn packuswb_px(self, v: __m256i) -> __m256i
fn packuswb_px(self, v: __m256i) -> __m256i
pixel lanes to raw pixels: packuswb (+ vpermq 0x08 on AVX2) as in store_px, on AVX2
without the clamps; the inverse of widen_px
Source§fn ptest_raw(self, a: __m256i, b: __m256i) -> u32
fn ptest_raw(self, a: __m256i, b: __m256i) -> u32
ptest on the raw-pixel bytes (SSE2: the low 8, AVX2: the low 16): 1 when a & b == 0
(ZF), else 2 when !a & b == 0 (CF), else 0. AVX2: vptest + jz + jnc, the CF case
falls through (inline assembly: core::arch writes ptest as or-reductions, and LLVM
then often builds shuffle / reduce sequences or loses the CF test); branch on the
result, the branches thread through
Auto Trait Implementations§
impl Forget for Avx2
impl Freeze for Avx2
impl Move for Avx2
impl RefUnwindSafe for Avx2
impl Send for Avx2
impl Sync for Avx2
impl Unpin for Avx2
impl UnsafeUnpin for Avx2
impl UnwindSafe for Avx2
Blanket Implementations§
Source§impl<T> BorrowMut<T> for Twhere
T: ?Sized,
impl<T> BorrowMut<T> for Twhere
T: ?Sized,
Source§fn borrow_mut(&mut self) -> &mut T
fn borrow_mut(&mut self) -> &mut T
Mutably borrows from an owned value. Read more