use crate::device::WgpuDevice;
use crate::device::capabilities::WORKGROUP_SIZE_COMPACT;
use crate::error::Result;
use std::sync::Arc;
use super::su3_extended::su3_extended_preamble;
const SHADER_BODY: &str = include_str!("../../shaders/lattice/hmc_leapfrog_f64.wgsl");
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct LeapfrogParams {
volume: u32,
n_links: u32,
_pad0: u32,
_pad1: u32,
dt: f64,
_padf: f64,
}
pub struct LeapfrogBuffers<'a> {
pub links_buf: &'a wgpu::Buffer,
pub momenta_buf: &'a wgpu::Buffer,
pub force_buf: &'a wgpu::Buffer,
pub rng_buf: &'a wgpu::Buffer,
}
pub struct GpuHmcLeapfrog {
device: Arc<WgpuDevice>,
n_links: u32,
kick_pipeline: wgpu::ComputePipeline,
update_pipeline: wgpu::ComputePipeline,
gen_pipeline: wgpu::ComputePipeline,
bgl: wgpu::BindGroupLayout,
}
impl GpuHmcLeapfrog {
pub fn new(device: Arc<WgpuDevice>, volume: u32) -> Result<Self> {
let n_links = volume * 4;
let src = format!("{}{}", su3_extended_preamble(), SHADER_BODY);
let module = device.compile_shader_f64(&src, Some("hmc_leapfrog"));
let bgl = device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("GpuHmcLeapfrog:bgl"),
entries: &[
uniform_bgl(0),
storage_bgl(1, false), storage_bgl(2, false), storage_bgl(3, true), storage_bgl(4, false), ],
});
let layout = device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("GpuHmcLeapfrog:layout"),
bind_group_layouts: &[&bgl],
immediate_size: 0,
});
let make_pipeline = |entry: &str, label: &str| {
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some(label),
layout: Some(&layout),
module: &module,
entry_point: Some(entry),
compilation_options: Default::default(),
cache: None,
})
};
let kick_pipeline = make_pipeline("momentum_kick", "GpuHmcLeapfrog:kick");
let update_pipeline = make_pipeline("link_update", "GpuHmcLeapfrog:update");
let gen_pipeline = make_pipeline("generate_momenta", "GpuHmcLeapfrog:gen");
Ok(Self {
device,
n_links,
kick_pipeline,
update_pipeline,
gen_pipeline,
bgl,
})
}
pub fn momentum_kick(&self, buffers: &LeapfrogBuffers<'_>, volume: u32, dt: f64) -> Result<()> {
self.dispatch(&self.kick_pipeline, buffers, volume, dt, "kick")
}
pub fn link_update(&self, buffers: &LeapfrogBuffers<'_>, volume: u32, dt: f64) -> Result<()> {
self.dispatch(&self.update_pipeline, buffers, volume, dt, "update")
}
pub fn generate_momenta(&self, buffers: &LeapfrogBuffers<'_>, volume: u32) -> Result<()> {
self.dispatch(&self.gen_pipeline, buffers, volume, 0.0, "gen")
}
fn dispatch(
&self,
pipeline: &wgpu::ComputePipeline,
buffers: &LeapfrogBuffers<'_>,
volume: u32,
dt: f64,
label: &str,
) -> Result<()> {
let params_data = LeapfrogParams {
volume,
n_links: self.n_links,
_pad0: 0,
_pad1: 0,
dt,
_padf: 0.0,
};
let params = self.device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some(&format!("GpuHmcLeapfrog:{label}:params")),
size: std::mem::size_of::<LeapfrogParams>() as u64,
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
mapped_at_creation: false,
});
self.device
.queue
.write_buffer(¶ms, 0, bytemuck::bytes_of(¶ms_data));
let bg = self
.device
.device
.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some(&format!("GpuHmcLeapfrog:{label}:bg")),
layout: &self.bgl,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: buffers.links_buf.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: buffers.momenta_buf.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: buffers.force_buf.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 4,
resource: buffers.rng_buf.as_entire_binding(),
},
],
});
let mut enc = self
.device
.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some(&format!("GpuHmcLeapfrog:{label}:enc")),
});
{
let mut pass = enc.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some(&format!("GpuHmcLeapfrog:{label}:pass")),
timestamp_writes: None,
});
pass.set_pipeline(pipeline);
pass.set_bind_group(0, Some(&bg), &[]);
pass.dispatch_workgroups(self.n_links.div_ceil(WORKGROUP_SIZE_COMPACT), 1, 1);
}
self.device.submit_commands(Some(enc.finish()));
Ok(())
}
#[must_use]
pub fn n_links(&self) -> u32 {
self.n_links
}
}
fn storage_bgl(binding: u32, read_only: bool) -> wgpu::BindGroupLayoutEntry {
wgpu::BindGroupLayoutEntry {
binding,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
}
}
fn uniform_bgl(binding: u32) -> wgpu::BindGroupLayoutEntry {
wgpu::BindGroupLayoutEntry {
binding,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
}
}
#[cfg(test)]
mod tests {
use super::*;
#[test]
fn test_leapfrog_pipeline_creation() {
let Some(device) = crate::device::test_pool::get_test_device_if_f64_gpu_available_sync()
else {
return;
};
let op = GpuHmcLeapfrog::new(device, 16).unwrap();
assert_eq!(op.n_links(), 64);
}
}