use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Gather {
input: Tensor,
dim: usize,
indices: Vec<u32>,
}
impl Gather {
#[must_use]
pub fn new(input: Tensor, dim: usize, indices: Vec<u32>) -> Self {
Self {
input,
dim,
indices,
}
}
fn wgsl_shader() -> &'static str {
{
static S: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/tensor/gather_f64.wgsl").to_string()
});
&S
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let shape = self.input.shape();
let dim_size = shape[self.dim];
let outer_size: usize = shape[..self.dim].iter().product();
let inner_size: usize = shape[self.dim + 1..].iter().product();
let gather_size = self.indices.len();
let output_size = outer_size * gather_size * inner_size;
let input_buffer = self.input.buffer();
let indices_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Gather Indices"),
contents: bytemuck::cast_slice(&self.indices),
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_DST,
});
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Gather Output"),
size: (output_size * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
size: u32,
dim_size: u32,
outer_size: u32,
inner_size: u32,
gather_size: u32,
}
let params = Params {
size: output_size as u32,
dim_size: dim_size as u32,
outer_size: outer_size as u32,
inner_size: inner_size as u32,
gather_size: gather_size as u32,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Gather Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Gather Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 1,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: true },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 2,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
wgpu::BindGroupLayoutEntry {
binding: 3,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Gather Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: indices_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: params_buffer.as_entire_binding(),
},
],
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Shader"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Gather Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Gather Pipeline"),
layout: Some(&pipeline_layout),
module: &shader_module,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Gather Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Gather Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&compute_pipeline);
compute_pass.set_bind_group(0, Some(&bind_group), &[]);
let caps = DeviceCapabilities::from_device(device);
let optimal_wg_size = caps.optimal_workgroup_size(WorkloadType::ElementWise);
let workgroups = (output_size as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
let output_data = crate::utils::read_buffer(device, &output_buffer, output_size)?;
let mut output_shape = shape.to_vec();
output_shape[self.dim] = gather_size;
Ok(Tensor::new(output_data, output_shape, device.clone()))
}
}
impl Tensor {
pub fn gather_wgsl(self, dim: usize, indices: Vec<u32>) -> Result<Self> {
Gather::new(self, dim, indices).execute()
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_gather_1d() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 2.0, 3.0, 4.0, 5.0];
let input = Tensor::new(data, vec![5], device);
let output = input.gather_wgsl(0, vec![0, 2, 4]).unwrap();
assert_eq!(output.shape(), &[3]);
let result = output.to_vec().unwrap();
assert_eq!(result[0], 1.0);
assert_eq!(result[1], 3.0);
assert_eq!(result[2], 5.0);
}
#[tokio::test]
async fn test_gather_2d() {
let device = crate::device::test_pool::get_test_device().await;
let data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0];
let input = Tensor::new(data, vec![3, 2], device);
let output = input.gather_wgsl(0, vec![0, 2]).unwrap();
assert_eq!(output.shape(), &[2, 2]);
let result = output.to_vec().unwrap();
assert_eq!(result[0], 1.0);
assert_eq!(result[1], 2.0);
assert_eq!(result[2], 5.0);
assert_eq!(result[3], 6.0);
}
}