use runmat_accelerate_api::GpuTensorHandle;
use runmat_builtins::{ComplexTensor, Tensor, Value};
use runmat_macros::runtime_builtin;
use crate::builtins::common::broadcast::BroadcastPlan;
use crate::builtins::common::spec::{
BroadcastSemantics, BuiltinFusionSpec, BuiltinGpuSpec, ConstantStrategy, FusionError,
FusionExprContext, FusionKernelTemplate, GpuOpKind, ProviderHook, ReductionNaN,
ResidencyPolicy, ScalarType, ShapeRequirements,
};
use crate::builtins::common::{gpu_helpers, tensor};
use crate::builtins::math::type_resolvers::numeric_binary_type;
use crate::{build_runtime_error, BuiltinResult, RuntimeError};
#[runmat_macros::register_gpu_spec(builtin_path = "crate::builtins::math::rounding::rem")]
pub const GPU_SPEC: BuiltinGpuSpec = BuiltinGpuSpec {
name: "rem",
op_kind: GpuOpKind::Elementwise,
supported_precisions: &[ScalarType::F32, ScalarType::F64],
broadcast: BroadcastSemantics::Matlab,
provider_hooks: &[
ProviderHook::Binary {
name: "elem_div",
commutative: false,
},
ProviderHook::Unary { name: "unary_fix" },
ProviderHook::Binary {
name: "elem_mul",
commutative: false,
},
ProviderHook::Binary {
name: "elem_sub",
commutative: false,
},
],
constant_strategy: ConstantStrategy::InlineLiteral,
residency: ResidencyPolicy::NewHandle,
nan_mode: ReductionNaN::Include,
two_pass_threshold: None,
workgroup_size: None,
accepts_nan_mode: false,
notes:
"Providers can compose rem from elem_div → unary_fix → elem_mul → elem_sub. Kernels fall back to host when any hook is missing or shapes differ.",
};
#[runmat_macros::register_fusion_spec(builtin_path = "crate::builtins::math::rounding::rem")]
pub const FUSION_SPEC: BuiltinFusionSpec = BuiltinFusionSpec {
name: "rem",
shape: ShapeRequirements::BroadcastCompatible,
constant_strategy: ConstantStrategy::InlineLiteral,
elementwise: Some(FusionKernelTemplate {
scalar_precisions: &[ScalarType::F32, ScalarType::F64],
wgsl_body: |ctx: &FusionExprContext| {
let a = ctx
.inputs
.first()
.ok_or(FusionError::MissingInput(0))?;
let b = ctx.inputs.get(1).ok_or(FusionError::MissingInput(1))?;
Ok(format!("{a} - {b} * trunc({a} / {b})"))
},
}),
reduction: None,
emits_nan: true,
notes: "Fusion expands rem as trunc(a / b) followed by a - b * q; providers may override with specialised kernels.",
};
const BUILTIN_NAME: &str = "rem";
fn builtin_error(message: impl Into<String>) -> RuntimeError {
build_runtime_error(message)
.with_builtin(BUILTIN_NAME)
.build()
}
#[runtime_builtin(
name = "rem",
category = "math/rounding",
summary = "MATLAB-compatible remainder a - b .* fix(a./b) with support for complex values and broadcasting.",
keywords = "rem,remainder,truncate,gpu",
accel = "binary",
type_resolver(numeric_binary_type),
builtin_path = "crate::builtins::math::rounding::rem"
)]
async fn rem_builtin(lhs: Value, rhs: Value) -> BuiltinResult<Value> {
match (lhs, rhs) {
(Value::GpuTensor(a), Value::GpuTensor(b)) => rem_gpu_pair(a, b).await,
(Value::GpuTensor(a), other) => {
let gathered = gpu_helpers::gather_tensor_async(&a).await?;
rem_host(Value::Tensor(gathered), other)
}
(other, Value::GpuTensor(b)) => {
let gathered = gpu_helpers::gather_tensor_async(&b).await?;
rem_host(other, Value::Tensor(gathered))
}
(left, right) => rem_host(left, right),
}
}
async fn rem_gpu_pair(a: GpuTensorHandle, b: GpuTensorHandle) -> BuiltinResult<Value> {
if a.device_id == b.device_id {
if let Some(provider) = runmat_accelerate_api::provider_for_handle(&a) {
if a.shape == b.shape {
if let Ok(div) = provider.elem_div(&a, &b).await {
match provider.unary_fix(&div).await {
Ok(fixed) => match provider.elem_mul(&b, &fixed).await {
Ok(mul) => match provider.elem_sub(&a, &mul).await {
Ok(out) => {
let _ = provider.free(&div);
let _ = provider.free(&fixed);
let _ = provider.free(&mul);
return Ok(gpu_helpers::resident_gpu_value(out));
}
Err(_) => {
let _ = provider.free(&mul);
let _ = provider.free(&fixed);
let _ = provider.free(&div);
}
},
Err(_) => {
let _ = provider.free(&fixed);
let _ = provider.free(&div);
}
},
Err(_) => {
let _ = provider.free(&div);
}
}
}
}
}
}
let left = gpu_helpers::gather_tensor_async(&a).await?;
let right = gpu_helpers::gather_tensor_async(&b).await?;
rem_host(Value::Tensor(left), Value::Tensor(right))
}
fn rem_host(lhs: Value, rhs: Value) -> BuiltinResult<Value> {
if let Some(result) = scalar_rem_value(&lhs, &rhs) {
return Ok(result);
}
let left = value_into_numeric_array(lhs, "rem")?;
let right = value_into_numeric_array(rhs, "rem")?;
match align_numeric_arrays(left, right)? {
NumericPair::Real(a, b) => compute_rem_real(&a, &b),
NumericPair::Complex(a, b) => compute_rem_complex(&a, &b),
}
}
fn compute_rem_real(a: &Tensor, b: &Tensor) -> BuiltinResult<Value> {
let plan = BroadcastPlan::new(&a.shape, &b.shape)
.map_err(|err| builtin_error(format!("rem: {err}")))?;
if plan.is_empty() {
let tensor = Tensor::new(Vec::new(), plan.output_shape().to_vec())
.map_err(|e| builtin_error(format!("rem: {e}")))?;
return Ok(tensor::tensor_into_value(tensor));
}
let mut result = vec![0.0f64; plan.len()];
for (out_idx, idx_a, idx_b) in plan.iter() {
result[out_idx] = rem_real_scalar(a.data[idx_a], b.data[idx_b]);
}
let tensor = Tensor::new(result, plan.output_shape().to_vec())
.map_err(|e| builtin_error(format!("rem: {e}")))?;
Ok(tensor::tensor_into_value(tensor))
}
fn compute_rem_complex(a: &ComplexTensor, b: &ComplexTensor) -> BuiltinResult<Value> {
let plan = BroadcastPlan::new(&a.shape, &b.shape)
.map_err(|err| builtin_error(format!("rem: {err}")))?;
if plan.is_empty() {
let tensor = ComplexTensor::new(Vec::new(), plan.output_shape().to_vec())
.map_err(|e| builtin_error(format!("rem: {e}")))?;
return Ok(complex_tensor_into_value(tensor));
}
let mut result = vec![(0.0f64, 0.0f64); plan.len()];
for (out_idx, idx_a, idx_b) in plan.iter() {
let (ar, ai) = a.data[idx_a];
let (br, bi) = b.data[idx_b];
result[out_idx] = rem_complex_scalar(ar, ai, br, bi);
}
let tensor = ComplexTensor::new(result, plan.output_shape().to_vec())
.map_err(|e| builtin_error(format!("rem: {e}")))?;
Ok(complex_tensor_into_value(tensor))
}
fn rem_real_scalar(a: f64, b: f64) -> f64 {
if a.is_nan() || b.is_nan() {
return f64::NAN;
}
if b == 0.0 {
return f64::NAN;
}
if !a.is_finite() && b.is_finite() {
return f64::NAN;
}
if b.is_infinite() && a.is_finite() {
return normalize_zero(a);
}
let quotient = (a / b).trunc();
if !quotient.is_finite() && b.is_finite() {
return f64::NAN;
}
let remainder = a - b * quotient;
normalize_zero(remainder)
}
fn rem_complex_scalar(ar: f64, ai: f64, br: f64, bi: f64) -> (f64, f64) {
if (ar.is_nan() || ai.is_nan()) || (br.is_nan() || bi.is_nan()) {
return (f64::NAN, f64::NAN);
}
if br == 0.0 && bi == 0.0 {
return (f64::NAN, f64::NAN);
}
if !ar.is_finite() || !ai.is_finite() {
return (f64::NAN, f64::NAN);
}
let (qr, qi) = complex_div(ar, ai, br, bi);
if !qr.is_finite() && !qi.is_finite() && br.is_finite() && bi.is_finite() {
return (f64::NAN, f64::NAN);
}
let (fr, fi) = (qr.trunc(), qi.trunc());
let (mulr, muli) = complex_mul(br, bi, fr, fi);
let (rr, ri) = (ar - mulr, ai - muli);
(normalize_zero(rr), normalize_zero(ri))
}
fn scalar_real_value(value: &Value) -> Option<f64> {
match value {
Value::Num(n) => Some(*n),
Value::Int(i) => Some(i.to_f64()),
Value::Bool(b) => Some(if *b { 1.0 } else { 0.0 }),
Value::Tensor(t) if t.data.len() == 1 => t.data.first().copied(),
Value::LogicalArray(l) if l.data.len() == 1 => Some(if l.data[0] != 0 { 1.0 } else { 0.0 }),
Value::CharArray(ca) if ca.rows * ca.cols == 1 => {
Some(ca.data.first().map(|&ch| ch as u32 as f64).unwrap_or(0.0))
}
_ => None,
}
}
fn scalar_complex_value(value: &Value) -> Option<(f64, f64)> {
match value {
Value::Complex(re, im) => Some((*re, *im)),
Value::ComplexTensor(ct) if ct.data.len() == 1 => ct.data.first().copied(),
_ => None,
}
}
fn scalar_rem_value(lhs: &Value, rhs: &Value) -> Option<Value> {
let left = scalar_complex_value(lhs).or_else(|| scalar_real_value(lhs).map(|v| (v, 0.0)))?;
let right = scalar_complex_value(rhs).or_else(|| scalar_real_value(rhs).map(|v| (v, 0.0)))?;
let (ar, ai) = left;
let (br, bi) = right;
if ai != 0.0 || bi != 0.0 {
let (re, im) = rem_complex_scalar(ar, ai, br, bi);
return Some(Value::Complex(re, im));
}
Some(Value::Num(rem_real_scalar(ar, br)))
}
fn normalize_zero(value: f64) -> f64 {
if value == -0.0 {
0.0
} else {
value
}
}
fn complex_mul(ar: f64, ai: f64, br: f64, bi: f64) -> (f64, f64) {
(ar * br - ai * bi, ar * bi + ai * br)
}
fn complex_div(ar: f64, ai: f64, br: f64, bi: f64) -> (f64, f64) {
let denom = br * br + bi * bi;
if denom == 0.0 {
return (f64::NAN, f64::NAN);
}
((ar * br + ai * bi) / denom, (ai * br - ar * bi) / denom)
}
fn complex_tensor_into_value(tensor: ComplexTensor) -> Value {
if tensor.data.len() == 1 {
let (re, im) = tensor.data[0];
Value::Complex(re, im)
} else {
Value::ComplexTensor(tensor)
}
}
fn value_into_numeric_array(value: Value, name: &str) -> BuiltinResult<NumericArray> {
match value {
Value::Complex(re, im) => {
let tensor = ComplexTensor::new(vec![(re, im)], vec![1, 1])
.map_err(|e| builtin_error(format!("{name}: {e}")))?;
Ok(NumericArray::Complex(tensor))
}
Value::ComplexTensor(ct) => Ok(NumericArray::Complex(ct)),
Value::CharArray(ca) => {
let data: Vec<f64> = ca.data.iter().map(|&ch| ch as u32 as f64).collect();
let tensor = Tensor::new(data, vec![ca.rows, ca.cols])
.map_err(|e| builtin_error(format!("{name}: {e}")))?;
Ok(NumericArray::Real(tensor))
}
Value::String(_) | Value::StringArray(_) => Err(builtin_error(format!(
"{name}: expected numeric input, got string"
))),
Value::GpuTensor(_) => Err(builtin_error(format!(
"{name}: internal error converting GPU tensor"
))),
other => {
let tensor =
tensor::value_into_tensor_for(name, other).map_err(|err| builtin_error(err))?;
Ok(NumericArray::Real(tensor))
}
}
}
enum NumericArray {
Real(Tensor),
Complex(ComplexTensor),
}
enum NumericPair {
Real(Tensor, Tensor),
Complex(ComplexTensor, ComplexTensor),
}
fn align_numeric_arrays(lhs: NumericArray, rhs: NumericArray) -> BuiltinResult<NumericPair> {
match (lhs, rhs) {
(NumericArray::Real(a), NumericArray::Real(b)) => Ok(NumericPair::Real(a, b)),
(left, right) => {
let lc = into_complex("rem", left)?;
let rc = into_complex("rem", right)?;
Ok(NumericPair::Complex(lc, rc))
}
}
}
fn into_complex(name: &str, input: NumericArray) -> BuiltinResult<ComplexTensor> {
match input {
NumericArray::Real(t) => {
let Tensor { data, shape, .. } = t;
let complex: Vec<(f64, f64)> = data.into_iter().map(|re| (re, 0.0)).collect();
ComplexTensor::new(complex, shape).map_err(|e| builtin_error(format!("{name}: {e}")))
}
NumericArray::Complex(ct) => Ok(ct),
}
}
#[cfg(test)]
pub(crate) mod tests {
use super::*;
use crate::builtins::common::test_support;
use crate::RuntimeError;
use futures::executor::block_on;
use runmat_builtins::{
CharArray, ComplexTensor, IntValue, LogicalArray, ResolveContext, Tensor, Type,
};
fn rem_builtin(lhs: Value, rhs: Value) -> BuiltinResult<Value> {
block_on(super::rem_builtin(lhs, rhs))
}
fn assert_error_contains(error: RuntimeError, needle: &str) {
assert!(
error.message().contains(needle),
"unexpected error: {}",
error.message()
);
}
#[test]
fn rem_type_preserves_tensor_shape() {
let out = numeric_binary_type(
&[
Type::Tensor {
shape: Some(vec![Some(2), Some(3)]),
},
Type::Tensor {
shape: Some(vec![Some(2), Some(3)]),
},
],
&ResolveContext::new(Vec::new()),
);
assert_eq!(
out,
Type::Tensor {
shape: Some(vec![Some(2), Some(3)])
}
);
}
#[test]
fn rem_type_scalar_and_tensor_returns_tensor() {
let out = numeric_binary_type(
&[
Type::Num,
Type::Tensor {
shape: Some(vec![Some(4), Some(1)]),
},
],
&ResolveContext::new(Vec::new()),
);
assert_eq!(
out,
Type::Tensor {
shape: Some(vec![Some(4), Some(1)])
}
);
}
#[test]
fn rem_type_scalar_returns_num() {
let out = numeric_binary_type(&[Type::Num, Type::Int], &ResolveContext::new(Vec::new()));
assert_eq!(out, Type::Num);
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_positive_values() {
let result = rem_builtin(Value::Num(17.0), Value::Num(5.0)).expect("rem");
match result {
Value::Num(v) => assert!((v - 2.0).abs() < 1e-12),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_negative_dividend_keeps_sign() {
let result = rem_builtin(Value::Num(-7.0), Value::Num(4.0)).expect("rem");
match result {
Value::Num(v) => assert!((v + 3.0).abs() < 1e-12),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_negative_divisor_retains_dividend_sign() {
let result = rem_builtin(Value::Num(7.0), Value::Num(-4.0)).expect("rem");
match result {
Value::Num(v) => assert!((v - 3.0).abs() < 1e-12),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_zero_divisor_returns_nan() {
let result = rem_builtin(Value::Num(3.0), Value::Num(0.0)).expect("rem");
match result {
Value::Num(v) => assert!(v.is_nan()),
other => panic!("expected NaN result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_infinite_divisor_returns_dividend() {
let result = rem_builtin(Value::Num(4.5), Value::Num(f64::INFINITY)).expect("rem");
match result {
Value::Num(v) => assert!((v - 4.5).abs() < 1e-12),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_infinite_dividend_returns_nan() {
let result = rem_builtin(Value::Num(f64::INFINITY), Value::Num(3.0)).expect("rem");
match result {
Value::Num(v) => assert!(v.is_nan()),
other => panic!("expected NaN result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_matrix_scalar_broadcast() {
let matrix = Tensor::new(vec![4.5, 7.1, -2.3, 0.4], vec![2, 2]).unwrap();
let result = rem_builtin(Value::Tensor(matrix), Value::Num(2.0)).expect("broadcast");
match result {
Value::Tensor(t) => {
assert_eq!(t.shape, vec![2, 2]);
assert_eq!(
t.data
.iter()
.map(|v| (v * 10.0).round() / 10.0)
.collect::<Vec<_>>(),
vec![0.5, 1.1, -0.3, 0.4]
);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_complex_support() {
let lhs = ComplexTensor::new(vec![(3.0, 4.0), (-2.0, 5.0)], vec![1, 2]).unwrap();
let rhs = ComplexTensor::new(vec![(2.0, 1.0)], vec![1, 1]).unwrap();
let result =
rem_builtin(Value::ComplexTensor(lhs), Value::ComplexTensor(rhs)).expect("complex rem");
match result {
Value::ComplexTensor(t) => {
assert_eq!(t.shape, vec![1, 2]);
assert_eq!(t.data, vec![(0.0, 0.0), (0.0, 1.0)]);
}
other => panic!("expected complex tensor, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_char_array_support() {
let chars = CharArray::new("AB".chars().collect(), 1, 2).unwrap();
let result =
rem_builtin(Value::CharArray(chars.clone()), Value::Num(5.0)).expect("char rem");
match result {
Value::Tensor(t) => {
assert_eq!(t.shape, vec![1, 2]);
let expected: Vec<f64> = chars
.data
.iter()
.map(|&ch| rem_real_scalar(ch as u32 as f64, 5.0))
.collect();
assert_eq!(t.data, expected);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_logical_array_support() {
let logical = LogicalArray::new(vec![1, 0, 1, 0], vec![2, 2]).unwrap();
let value =
rem_builtin(Value::LogicalArray(logical), Value::Num(2.0)).expect("logical rem");
match value {
Value::Tensor(t) => assert_eq!(t.data, vec![1.0, 0.0, 1.0, 0.0]),
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_broadcasting_between_vectors() {
let lhs = Tensor::new(vec![1.0, -2.0], vec![2, 1]).unwrap();
let rhs = Tensor::new(vec![3.0, 4.0, 5.0], vec![1, 3]).unwrap();
let result = rem_builtin(Value::Tensor(lhs), Value::Tensor(rhs)).expect("broadcast");
match result {
Value::Tensor(t) => {
assert_eq!(t.shape, vec![2, 3]);
assert_eq!(t.data, vec![1.0, -2.0, 1.0, -2.0, 1.0, -2.0]);
}
other => panic!("expected tensor result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_gpu_pair_roundtrip() {
test_support::with_test_provider(|provider| {
let numer = Tensor::new(vec![-5.0, -3.0, 0.0, 1.0, 6.0, 9.0], vec![3, 2]).unwrap();
let denom = Tensor::new(vec![4.0; 6], vec![3, 2]).unwrap();
let numer_view = runmat_accelerate_api::HostTensorView {
data: &numer.data,
shape: &numer.shape,
};
let denom_view = runmat_accelerate_api::HostTensorView {
data: &denom.data,
shape: &denom.shape,
};
let numer_handle = provider.upload(&numer_view).expect("upload numer");
let denom_handle = provider.upload(&denom_view).expect("upload denom");
let result = rem_builtin(
Value::GpuTensor(numer_handle),
Value::GpuTensor(denom_handle),
)
.expect("rem");
let gathered = test_support::gather(result).expect("gather result");
assert_eq!(gathered.shape, vec![3, 2]);
assert_eq!(gathered.data, vec![-1.0, -3.0, 0.0, 1.0, 2.0, 1.0]);
});
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_int_inputs_promote() {
let result =
rem_builtin(Value::Int(IntValue::I32(-7)), Value::Int(IntValue::I32(4))).expect("rem");
match result {
Value::Num(v) => assert!((v + 3.0).abs() < 1e-12),
other => panic!("expected scalar result, got {other:?}"),
}
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
fn rem_string_input_errors() {
let err = rem_builtin(Value::from("abc"), Value::Num(3.0))
.expect_err("string inputs should error");
assert_error_contains(err, "expected numeric input");
}
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test::wasm_bindgen_test)]
#[test]
#[cfg(feature = "wgpu")]
fn rem_wgpu_matches_cpu() {
let _ = runmat_accelerate::backend::wgpu::provider::register_wgpu_provider(
runmat_accelerate::backend::wgpu::provider::WgpuProviderOptions::default(),
);
let numer = Tensor::new(vec![-5.0, -3.25, 0.0, 1.75, 6.5, 9.0], vec![3, 2]).unwrap();
let denom = Tensor::new(vec![4.0, -2.5, 3.0, 3.0, 2.0, -5.0], vec![3, 2]).unwrap();
let cpu_value =
rem_host(Value::Tensor(numer.clone()), Value::Tensor(denom.clone())).expect("cpu rem");
let provider = runmat_accelerate_api::provider().expect("wgpu provider registered");
let numer_handle = provider
.upload(&runmat_accelerate_api::HostTensorView {
data: &numer.data,
shape: &numer.shape,
})
.expect("upload numer");
let denom_handle = provider
.upload(&runmat_accelerate_api::HostTensorView {
data: &denom.data,
shape: &denom.shape,
})
.expect("upload denom");
let gpu_value = block_on(rem_gpu_pair(numer_handle, denom_handle)).expect("gpu rem");
let gpu_tensor = test_support::gather(gpu_value).expect("gather gpu result");
let cpu_tensor = match cpu_value {
Value::Tensor(t) => t,
Value::Num(n) => Tensor::new(vec![n], vec![1, 1]).expect("scalar tensor"),
other => panic!("unexpected CPU result {other:?}"),
};
assert_eq!(gpu_tensor.shape, cpu_tensor.shape);
let tol = match provider.precision() {
runmat_accelerate_api::ProviderPrecision::F64 => 1e-12,
runmat_accelerate_api::ProviderPrecision::F32 => 1e-5,
};
for (gpu, cpu) in gpu_tensor.data.iter().zip(cpu_tensor.data.iter()) {
assert!(
(gpu - cpu).abs() <= tol,
"|{gpu} - {cpu}| exceeded tolerance {tol}"
);
}
}
}