use crate::builtins::common::shape::{dims_to_row_tensor, value_dimensions};
use crate::builtins::common::spec::{
BroadcastSemantics, BuiltinFusionSpec, BuiltinGpuSpec, ConstantStrategy, GpuOpKind,
ReductionNaN, ResidencyPolicy, ShapeRequirements,
};
use crate::builtins::common::tensor;
use crate::{build_runtime_error, RuntimeError};
use runmat_builtins::{ResolveContext, Tensor, Type, Value};
use runmat_macros::runtime_builtin;
#[runmat_macros::register_gpu_spec(builtin_path = "crate::builtins::array::introspection::size")]
pub const GPU_SPEC: BuiltinGpuSpec = BuiltinGpuSpec {
name: "size",
op_kind: GpuOpKind::Custom("metadata"),
supported_precisions: &[],
broadcast: BroadcastSemantics::None,
provider_hooks: &[],
constant_strategy: ConstantStrategy::InlineLiteral,
residency: ResidencyPolicy::GatherImmediately,
nan_mode: ReductionNaN::Include,
two_pass_threshold: None,
workgroup_size: None,
accepts_nan_mode: false,
notes:
"Reads dimension metadata from tensor handles; no kernels or provider hooks are required.",
};
#[runmat_macros::register_fusion_spec(builtin_path = "crate::builtins::array::introspection::size")]
pub const FUSION_SPEC: BuiltinFusionSpec = BuiltinFusionSpec {
name: "size",
shape: ShapeRequirements::Any,
constant_strategy: ConstantStrategy::InlineLiteral,
elementwise: None,
reduction: None,
emits_nan: false,
notes: "Metadata query; fusion planner bypasses this builtin.",
};
fn size_error(message: impl Into<String>) -> RuntimeError {
build_runtime_error(message).with_builtin("size").build()
}
fn vector_len_from_type(ty: &Type) -> Option<usize> {
let shape = match ty {
Type::Tensor { shape: Some(shape) } => Some(shape.as_slice()),
Type::Logical { shape: Some(shape) } => Some(shape.as_slice()),
_ => None,
}?;
match shape {
[Some(len)] => Some(*len),
[Some(rows), Some(cols)] if *rows == 1 => Some(*cols),
[Some(rows), Some(cols)] if *cols == 1 => Some(*rows),
_ => None,
}
}
fn normalized_rank(shape: &[Option<usize>]) -> usize {
shape.len().max(2)
}
fn size_type(args: &[Type], _context: &ResolveContext) -> Type {
let input = match args.first() {
Some(value) => value,
None => return Type::Unknown,
};
if args.len() == 1 {
let rank = match input {
Type::Tensor { shape: Some(shape) } | Type::Logical { shape: Some(shape) } => {
normalized_rank(shape)
}
Type::Num | Type::Int | Type::Bool => 2,
_ => return Type::tensor(),
};
return Type::Tensor {
shape: Some(vec![Some(1), Some(rank)]),
};
}
if let Some(len) = vector_len_from_type(&args[1]) {
return Type::Tensor {
shape: Some(vec![Some(1), Some(len)]),
};
}
if matches!(args[1], Type::Num | Type::Int | Type::Bool) {
return Type::Int;
}
Type::tensor()
}
#[runtime_builtin(
name = "size",
category = "array/introspection",
summary = "Get the dimensions of scalars, vectors, matrices, and N-D arrays.",
keywords = "size,dimensions,shape,gpu,introspection",
type_resolver(size_type),
builtin_path = "crate::builtins::array::introspection::size"
)]
async fn size_builtin(value: Value, rest: Vec<Value>) -> crate::BuiltinResult<Value> {
let dims = value_dimensions(&value).await?;
match rest.len() {
0 => dimensions_to_value(&dims),
1 => match parse_dim_selection(&rest[0])? {
DimSelection::Single(dim) => {
let extent = dimension_extent(&dims, dim);
Ok(Value::Num(extent as f64))
}
DimSelection::Multiple(dimensions) => {
let extents: Vec<usize> = dimensions
.into_iter()
.map(|dim| dimension_extent(&dims, dim))
.collect();
dimensions_to_value(&extents)
}
},
_ => Err(size_error("size: too many input arguments")),
}
}
fn dimensions_to_value(dimensions: &[usize]) -> crate::BuiltinResult<Value> {
let tensor = dims_to_row_tensor(dimensions)
.map_err(|e| size_error(format!("size: failed to build output: {e}")))?;
Ok(tensor::tensor_into_value(tensor))
}
enum DimSelection {
Single(usize),
Multiple(Vec<usize>),
}
fn parse_dim_selection(arg: &Value) -> crate::BuiltinResult<DimSelection> {
match arg {
Value::Int(_) | Value::Num(_) => {
let dim = tensor::parse_dimension(arg, "size").map_err(|e| size_error(e))?;
Ok(DimSelection::Single(dim))
}
Value::Tensor(t) => {
ensure_dim_vector(t)?;
if t.data.is_empty() {
return Err(size_error(
"size: dimension vector must contain at least one element",
));
}
let dims = t
.data
.iter()
.map(|&raw| parse_dim_scalar(raw))
.collect::<crate::BuiltinResult<Vec<_>>>()?;
Ok(DimSelection::Multiple(dims))
}
_ => Err(size_error(
"size: dimension argument must be a numeric scalar or vector",
)),
}
}
fn ensure_dim_vector(t: &Tensor) -> crate::BuiltinResult<()> {
let non_unit_dims = t.shape.iter().filter(|&&dim| dim > 1).count();
if non_unit_dims <= 1 {
Ok(())
} else {
Err(size_error(
"size: dimension vector must be a vector of positive integers",
))
}
}
fn parse_dim_scalar(raw: f64) -> crate::BuiltinResult<usize> {
if !raw.is_finite() {
return Err(size_error("size: dimension must be finite"));
}
let rounded = raw.round();
if (rounded - raw).abs() > f64::EPSILON {
return Err(size_error("size: dimension must be an integer"));
}
if rounded < 1.0 {
return Err(size_error("size: dimension must be >= 1"));
}
Ok(rounded as usize)
}
fn dimension_extent(dimensions: &[usize], dim: usize) -> usize {
dimensions.get(dim.saturating_sub(1)).copied().unwrap_or(1)
}
#[cfg(test)]
pub(crate) mod tests {
use super::*;
use crate::builtins::common::test_support;
use futures::executor::block_on;
fn size_builtin(value: Value, rest: Vec<Value>) -> crate::BuiltinResult<Value> {
block_on(super::size_builtin(value, rest))
}
use runmat_builtins::Tensor;
#[test]
fn size_type_infers_row_vector_rank() {
let out = size_type(
&[Type::Tensor {
shape: Some(vec![Some(2), Some(3), Some(4)]),
}],
&ResolveContext::new(Vec::new()),
);
assert_eq!(
out,
Type::Tensor {
shape: Some(vec![Some(1), Some(3)])
}
);
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_matrix_returns_row_vector() {
let tensor = Tensor::new(vec![1.0, 4.0, 2.0, 5.0, 3.0, 6.0], vec![2, 3]).unwrap();
let result = size_builtin(Value::Tensor(tensor), Vec::new()).expect("size");
match result {
Value::Tensor(out) => {
assert_eq!(out.shape, vec![1, 2]);
assert_eq!(out.data, vec![2.0, 3.0]);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_with_dimension_scalar_returns_extent() {
let tensor = Tensor::new(vec![1.0, 4.0, 2.0, 5.0], vec![2, 2]).unwrap();
let result = size_builtin(Value::Tensor(tensor), vec![Value::from(1.0)]).expect("size dim");
match result {
Value::Num(v) => assert_eq!(v, 2.0),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_with_dimension_vector_returns_row_vector() {
let tensor = Tensor::new(vec![0.0; 24], vec![2, 3, 4]).unwrap();
let dims_arg = Tensor::new(vec![1.0, 3.0], vec![1, 2]).unwrap();
let result = size_builtin(Value::Tensor(tensor), vec![Value::Tensor(dims_arg)])
.expect("size dims vector");
match result {
Value::Tensor(out) => {
assert_eq!(out.shape, vec![1, 2]);
assert_eq!(out.data, vec![2.0, 4.0]);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_gpu_tensor_uses_handle_shape() {
test_support::with_test_provider(|provider| {
let tensor = Tensor::new(vec![0.0; 8], vec![2, 4]).unwrap();
let view = runmat_accelerate_api::HostTensorView {
data: &tensor.data,
shape: &tensor.shape,
};
let handle = provider.upload(&view).expect("upload");
let result = size_builtin(Value::GpuTensor(handle), Vec::new()).expect("size gpu");
match result {
Value::Tensor(out) => {
assert_eq!(out.shape, vec![1, 2]);
assert_eq!(out.data, vec![2.0, 4.0]);
}
other => panic!("expected tensor result, got {other:?}"),
}
});
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
#[cfg(feature = "wgpu")]
fn size_wgpu_preserves_shape_metadata() {
struct EnvGuard(Option<String>);
impl Drop for EnvGuard {
fn drop(&mut self) {
match &self.0 {
Some(prev) => std::env::set_var("RUNMAT_WGPU_FORCE_PRECISION", prev.as_str()),
None => std::env::remove_var("RUNMAT_WGPU_FORCE_PRECISION"),
}
}
}
let previous = std::env::var("RUNMAT_WGPU_FORCE_PRECISION").ok();
std::env::set_var("RUNMAT_WGPU_FORCE_PRECISION", "f32");
let _guard = EnvGuard(previous);
let _ = runmat_accelerate::backend::wgpu::provider::register_wgpu_provider(
runmat_accelerate::backend::wgpu::provider::WgpuProviderOptions::default(),
);
let tensor = Tensor::new(vec![0.0; 12], vec![3, 4]).unwrap();
let view = runmat_accelerate_api::HostTensorView {
data: &tensor.data,
shape: &tensor.shape,
};
let handle = runmat_accelerate_api::provider()
.expect("wgpu provider")
.upload(&view)
.expect("upload to device");
let result = size_builtin(Value::GpuTensor(handle), Vec::new()).expect("size");
match result {
Value::Tensor(out) => {
assert_eq!(out.shape, vec![1, 2]);
assert_eq!(out.data, vec![3.0, 4.0]);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_rejects_non_numeric_dimension() {
let err = size_builtin(Value::Num(1.0), vec![Value::from("dim")]).unwrap_err();
assert!(
err.to_string().contains("dimension argument"),
"unexpected error: {err}"
);
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_dimension_beyond_rank_returns_one() {
let tensor = Tensor::new(vec![1.0, 2.0, 3.0], vec![3, 1]).unwrap();
let result = size_builtin(Value::Tensor(tensor), vec![Value::from(5.0)]).expect("size dim");
match result {
Value::Num(v) => assert_eq!(v, 1.0),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_dimension_vector_requires_positive_integers() {
let tensor = Tensor::new(vec![0.0; 8], vec![2, 4]).unwrap();
let dims = Tensor::new(vec![1.0, 2.5], vec![1, 2]).unwrap();
let err = size_builtin(Value::Tensor(tensor), vec![Value::Tensor(dims)])
.expect_err("non-int dim");
assert!(err.to_string().contains("dimension must be an integer"));
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_dimension_vector_must_not_be_matrix() {
let tensor = Tensor::new(vec![0.0; 8], vec![2, 4]).unwrap();
let dims = Tensor::new(vec![1.0, 2.0, 3.0, 4.0], vec![2, 2]).unwrap();
let err = size_builtin(Value::Tensor(tensor), vec![Value::Tensor(dims)])
.expect_err("matrix dims");
assert!(err
.to_string()
.contains("dimension vector must be a vector"));
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn size_dimension_vector_must_not_be_empty() {
let tensor = Tensor::new(vec![0.0; 8], vec![2, 4]).unwrap();
let dims = Tensor::new(vec![], vec![1, 0]).unwrap();
let err =
size_builtin(Value::Tensor(tensor), vec![Value::Tensor(dims)]).expect_err("empty dims");
assert!(err
.to_string()
.contains("must contain at least one element"));
}
}