| use core::arch::aarch64::*; |
| |
| union UnionCast { |
| // u32x4: [u32; 4], |
| f32x4: [f32; 4], |
| v: float32x4_t, |
| } |
| |
| #[inline] |
| pub const fn f32x4_from_array(f32x4: [f32; 4]) -> float32x4_t { |
| unsafe { UnionCast { f32x4 }.v } |
| } |
| |
| // #[inline] |
| // pub(crate) unsafe fn dot3_in_x(lhs: float32x4_t, rhs: float32x4_t) -> float32x4_t { |
| // let x2_y2_z2_w2 = vmulq_f32(lhs, rhs); |
| // let y2 = vdupq_laneq_f32(x2_y2_z2_w2, 1); |
| // let z2 = vdupq_laneq_f32(x2_y2_z2_w2, 2); |
| // let x2y2 = vaddq_f32(x2_y2_z2_w2, y2); |
| // vaddq_f32(x2y2, z2) |
| // } |
| |
| #[inline] |
| pub(crate) unsafe fn dot3(lhs: float32x4_t, rhs: float32x4_t) -> f32 { |
| let x2_y2_z2_w2 = vmulq_f32(lhs, rhs); |
| let x2_y2_z2 = vsetq_lane_f32(0.0, x2_y2_z2_w2, 3); |
| vaddvq_f32(x2_y2_z2) |
| // let dot = dot3_in_x(lhs, rhs); |
| // vdups_laneq_f32(dot, 0) |
| } |
| |
| #[inline] |
| pub(crate) unsafe fn dot3_into_f32x4(lhs: float32x4_t, rhs: float32x4_t) -> float32x4_t { |
| let dot = dot3(lhs, rhs); |
| vld1q_dup_f32(&dot as *const f32) |
| // let dot = dot3_in_x(lhs, rhs); |
| // vdupq_laneq_f32(dot, 0) |
| } |
| |
| #[inline] |
| pub(crate) unsafe fn dot4(lhs: float32x4_t, rhs: float32x4_t) -> f32 { |
| let x2_y2_z2_w2 = vmulq_f32(lhs, rhs); |
| // TODO: horizontal add - might perform bad? |
| vaddvq_f32(x2_y2_z2_w2) |
| } |
| |
| #[inline] |
| pub(crate) unsafe fn dot4_into_f32x4(lhs: float32x4_t, rhs: float32x4_t) -> float32x4_t { |
| let dot = dot4(lhs, rhs); |
| vld1q_dup_f32(&dot as *const f32) |
| } |