@shared

A tensor declared in a @persistent kernel is scratch the launch owns. The compiler holds one in shared memory when every block writes all of it alike: each block then has its own complete copy and reads only that, so the value never leaves the block. Anything else is published to device memory, which another block can read after a grid barrier.

@shared(H, ..)

@shared asserts that each named scratch tensor is held, and fails compilation if one cannot be: some block writes only part of it, or the held scratch together would outgrow the 48 KB of static shared memory. @global vetoes holding instead.

@launch(256)
@persistent(BLOCKS)
@autotune(BLOCKS in [8])
@shared(H)
kernel k(X: tensor<f32>[1, 256], Y: tensor<f32>[1, 256]) {
  var H: tensor<f32>[1, 256]
  H[0, ctx.thread] = X[0, ctx.thread] * 2.0
  barrier()
  for i in range(0, 4) {
    let u = ctx.block.x + i * ctx.blocks.x
    if u < 32 {
      Y[0 :+ 1, u * 8 :+ 8] = H[0 :+ 1, u * 8 :+ 8] + 1.0
    }
  }
}

Every block writes all of H from X, so H is a shared tile, and its reads need no grid barrier. See the Attributes table in the spec.