use crate::device::WgpuDevice;
use crate::device::capabilities::WORKGROUP_SIZE_1D;
use crate::device::capabilities::{DeviceCapabilities, Fp64Strategy};
use crate::device::pipeline_cache::{BindGroupLayoutSignature, create_f64_data_pipeline};
use crate::device::tensor_context::get_device_context;
use crate::error::Result;
use std::sync::Arc;
const SHADER: &str = include_str!("../shaders/special/digamma_f64.wgsl");
const DF64_CORE: &str = include_str!("../shaders/math/df64_core.wgsl");
fn shader_for_device(device: &WgpuDevice) -> Result<&'static str> {
let caps = DeviceCapabilities::from_device(device);
match caps.fp64_strategy() {
Fp64Strategy::Sovereign | Fp64Strategy::Native | Fp64Strategy::Concurrent => Ok(SHADER),
Fp64Strategy::Hybrid => {
static DF64_RESULT: std::sync::LazyLock<std::result::Result<String, Arc<str>>> =
std::sync::LazyLock::new(|| {
crate::shaders::sovereign::df64_rewrite::rewrite_f64_infix_full(SHADER)
.map(|src| format!("enable f64;\n{DF64_CORE}\n{src}"))
.map_err(|e| Arc::from(format!("digamma DF64 rewrite failed: {e}")))
});
match DF64_RESULT.as_ref() {
Ok(src) => Ok(src),
Err(msg) => Err(crate::error::BarracudaError::ShaderCompilation(Arc::clone(
msg,
))),
}
}
}
}
pub struct DigammaF64 {
device: Arc<WgpuDevice>,
}
#[expect(
dead_code,
reason = "public API — exercised by tests, available to downstream consumers"
)]
impl DigammaF64 {
pub fn new(device: Arc<WgpuDevice>) -> Result<Self> {
Ok(Self { device })
}
pub fn digamma(&self, x: &[f64]) -> Result<Vec<f64>> {
if x.is_empty() {
return Ok(vec![]);
}
self.dispatch_elementwise(x)
}
fn dispatch_elementwise(&self, x: &[f64]) -> Result<Vec<f64>> {
let n = x.len();
let ctx = get_device_context(&self.device);
let input_buf = self
.device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Digamma Input"),
contents: bytemuck::cast_slice(x),
usage: wgpu::BufferUsages::STORAGE,
});
let output_buf = ctx.acquire_pooled_output_f64(n);
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
size: u32,
_pad0: u32,
_pad1: u32,
_pad2: u32,
}
let params = Params {
size: n as u32,
_pad0: 0,
_pad1: 0,
_pad2: 0,
};
let params_buf = self.device.create_uniform_buffer("Digamma Params", ¶ms);
let layout_sig = BindGroupLayoutSignature::reduction();
let bind_group = ctx.get_or_create_bind_group(
layout_sig,
&[&input_buf, &output_buf, ¶ms_buf],
Some("Digamma BG"),
);
let shader_src = shader_for_device(&self.device)?;
let pipeline = create_f64_data_pipeline(
&self.device,
shader_src,
layout_sig,
"main",
Some("Digamma Pipeline"),
);
let workgroups = n.div_ceil(WORKGROUP_SIZE_1D as usize) as u32;
ctx.record_operation(move |encoder| {
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Digamma Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&*bind_group), &[]);
pass.dispatch_workgroups(workgroups, 1, 1);
})?;
self.device.read_buffer_f64(&output_buf, n)
}
fn digamma_cpu(&self, x: &[f64]) -> Vec<f64> {
x.iter().map(|&xi| Self::digamma_scalar(xi)).collect()
}
fn digamma_scalar(x: f64) -> f64 {
use std::f64::consts::PI;
if x <= 0.0 && x == x.floor() {
return f64::NAN;
}
let mut y = x;
let mut result = 0.0;
if y < 0.0 {
let cot_pi_y = (PI * y).cos() / (PI * y).sin();
result = PI.mul_add(-cot_pi_y, result);
y = 1.0 - y;
}
while y < 6.0 {
result -= 1.0 / y;
y += 1.0;
}
result + Self::digamma_asymptotic(y)
}
fn digamma_asymptotic(x: f64) -> f64 {
let inv_x = 1.0 / x;
let inv_x2 = inv_x * inv_x;
const B2: f64 = 1.0 / 12.0;
const B4: f64 = -1.0 / 120.0;
const B6: f64 = 1.0 / 252.0;
const B8: f64 = -1.0 / 240.0;
const B10: f64 = 1.0 / 132.0;
const B12: f64 = -691.0 / 32_760.0;
let mut sum = 0.5f64.mul_add(-inv_x, x.ln());
let mut term = inv_x2;
sum = B2.mul_add(-term, sum);
term *= inv_x2;
sum = B4.mul_add(-term, sum);
term *= inv_x2;
sum = B6.mul_add(-term, sum);
term *= inv_x2;
sum = B8.mul_add(-term, sum);
term *= inv_x2;
sum = B10.mul_add(-term, sum);
term *= inv_x2;
sum = B12.mul_add(-term, sum);
sum
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::device::test_pool::get_test_device_if_f64_transcendentals_available;
#[tokio::test]
async fn test_digamma_at_1() {
let Some(device) = get_test_device_if_f64_transcendentals_available().await else {
return;
};
let digamma = DigammaF64::new(device).unwrap();
let euler_mascheroni = 0.5772156649015329;
let result = digamma.digamma(&[1.0]).unwrap();
assert!(
(result[0] + euler_mascheroni).abs() < 1e-6,
"ψ(1) = {}, expected -γ = {}",
result[0],
-euler_mascheroni
);
}
#[tokio::test]
async fn test_digamma_recurrence() {
let run = |device: std::sync::Arc<crate::device::WgpuDevice>| {
let digamma = DigammaF64::new(device)?;
for x in [1.0, 2.0, 3.0, 4.5, 7.3] {
let result = digamma.digamma(&[x, x + 1.0])?;
let psi_x = result[0];
let psi_x1 = result[1];
assert!(
(psi_x1 - psi_x - 1.0 / x).abs() < 1e-6,
"ψ({}) + 1/{} = {} should equal ψ({}) = {}",
x,
x,
psi_x + 1.0 / x,
x + 1.0,
psi_x1
);
}
Ok::<(), crate::error::BarracudaError>(())
};
let Some(device) = get_test_device_if_f64_transcendentals_available().await else {
return;
};
match run(device) {
Ok(()) => {}
Err(e) if e.is_device_lost() => {
tracing::warn!("device lost in digamma recurrence, retrying");
let fresh = get_test_device_if_f64_transcendentals_available()
.await
.expect("f64 GPU unavailable on retry");
run(fresh).expect("failed on retry after device recovery");
}
Err(e) => panic!("test failed: {e}"),
}
}
#[tokio::test]
async fn test_digamma_known_values() {
let Some(device) = get_test_device_if_f64_transcendentals_available().await else {
return;
};
let digamma = DigammaF64::new(device).unwrap();
let euler_mascheroni = 0.5772156649015329;
let result = digamma.digamma(&[2.0]).unwrap();
let expected = 1.0 - euler_mascheroni;
assert!(
(result[0] - expected).abs() < 1e-6,
"ψ(2) = {}, expected {}",
result[0],
expected
);
let result = digamma.digamma(&[0.5]).unwrap();
let expected = 2.0f64.mul_add(-2.0_f64.ln(), -euler_mascheroni);
assert!(
(result[0] - expected).abs() < 1e-6,
"ψ(0.5) = {}, expected {}",
result[0],
expected
);
}
#[tokio::test]
async fn test_digamma_large_x() {
let Some(device) = get_test_device_if_f64_transcendentals_available().await else {
return;
};
let digamma = DigammaF64::new(device).unwrap();
let x = 100.0;
let result = digamma.digamma(&[x]).unwrap();
let approx = x.ln() - 0.5 / x;
assert!(
(result[0] - approx).abs() < 1e-4,
"ψ({}) = {}, asymptotic approx = {}",
x,
result[0],
approx
);
}
}