Primitive Static Slice

Extracts a slice from a tensor using static (compile-time) indices. All indices, limits, and strides are fixed R integers.

Use nvl_dynamic_slice() instead when the start position must be computed at runtime (e.g. depends on tensor values).

Usage

nvl_static_slice(operand, start_indices, limit_indices, strides)

Arguments

operand

(tensorish)
Tensorish value of any data type.

start_indices

(integer())
Start indices (inclusive), one per dimension. Must satisfy 1 <= start_indices <= limit_indices per dimension.

limit_indices

(integer())
End indices (inclusive), one per dimension. Must satisfy limit_indices <= nv_shape(operand) per dimension.

strides

(integer())
Step sizes, one per dimension. Must be >= 1. A stride of 1 selects every element; a stride of 2 selects every other element, etc.

Value

tensorish
Has the same data type as the input and shape ceiling((limit_indices - start_indices + 1) / strides). It is ambiguous if the input is ambiguous.

Implemented Rules

• stablehlo

• backward

StableHLO

Lowers to stablehlo::hlo_slice().

See also

nvl_dynamic_slice(), nvl_scatter(), nvl_gather(), nv_subset(), [

Examples

# 1-D: extract elements 2 through 4 (limit is exclusive)
jit_eval({
  x <- nv_tensor(1:10)
  nvl_static_slice(x, start_indices = 2L, limit_indices = 5L, strides = 1L)
})
#> AnvilTensor
#>  2
#>  3
#>  4
#>  5
#> [ CPUi32{4} ] 

# 1-D: every other element using strides
jit_eval({
  x <- nv_tensor(1:10)
  nvl_static_slice(x, start_indices = 1L, limit_indices = 10L, strides = 2L)
})
#> AnvilTensor
#>  1
#>  3
#>  5
#>  7
#>  9
#> [ CPUi32{5} ] 

# 2-D: extract a submatrix (rows 1-2, columns 2-3)
jit_eval({
  x <- nv_tensor(matrix(1:12, nrow = 3, ncol = 4))
  nvl_static_slice(x,
    start_indices = c(1L, 2L),
    limit_indices = c(3L, 4L),
    strides       = c(1L, 1L)
  )
})
#> AnvilTensor
#>   4  7 10
#>   5  8 11
#>   6  9 12
#> [ CPUi32{3,3} ]