use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
#[repr(C)]
#[derive(Copy, Clone, Debug, bytemuck::Pod, bytemuck::Zeroable)]
struct OuterProductParams {
size_a: u32,
size_b: u32,
_pad1: u32,
_pad2: u32,
}
pub struct OuterProduct {
vec_a: Tensor,
vec_b: Tensor,
}
impl OuterProduct {
pub fn new(vec_a: Tensor, vec_b: Tensor) -> Result<Self> {
if vec_a.shape().len() != 1 {
return Err(BarracudaError::invalid_op(
"outer_product",
"First input must be 1D vector",
));
}
if vec_b.shape().len() != 1 {
return Err(BarracudaError::invalid_op(
"outer_product",
"Second input must be 1D vector",
));
}
Ok(Self { vec_a, vec_b })
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/misc/outer_product_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.vec_a.device();
let size_a = self.vec_a.len();
let size_b = self.vec_b.len();
let output_size = size_a * size_b;
let output_buffer = device.create_buffer_f32(output_size)?;
let params = OuterProductParams {
size_a: size_a as u32,
size_b: size_b as u32,
_pad1: 0,
_pad2: 0,
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("OuterProduct Params"),
contents: bytemuck::bytes_of(¶ms),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("OuterProduct Bind Group Layout"),
entries: &[
wgpu::BindGroupLayoutEntry {
binding: 0,
visibility: wgpu::ShaderStages::COMPUTE,
ty: wgpu::BindingType::Buffer {
ty: wgpu::BufferBindingType::Uniform,
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: true },
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::Storage { read_only: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("OuterProduct Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.vec_a.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: self.vec_b.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 3,
resource: output_buffer.as_entire_binding(),
},
],
});
let shader = device.compile_shader(Self::wgsl_shader(), Some("OuterProduct"));
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("OuterProduct Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("OuterProduct Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
let mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("OuterProduct Encoder"),
});
{
let mut pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("OuterProduct Pass"),
timestamp_writes: None,
});
pass.set_pipeline(&pipeline);
pass.set_bind_group(0, Some(&bind_group), &[]);
let workgroups_x = (size_b as u32).div_ceil(16);
let workgroups_y = (size_a as u32).div_ceil(16);
pass.dispatch_workgroups(workgroups_x, workgroups_y, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![size_a, size_b],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_outer_product_basic() {
let device = crate::device::test_pool::get_test_device().await;
let a = Tensor::from_vec_on(vec![1.0, 2.0, 3.0], vec![3], device.clone())
.await
.unwrap();
let b = Tensor::from_vec_on(vec![4.0, 5.0], vec![2], device.clone())
.await
.unwrap();
let result = OuterProduct::new(a, b).unwrap().execute().unwrap();
assert_eq!(result.shape(), &[3, 2]);
let output = result.to_vec().unwrap();
assert_eq!(output[0], 4.0); assert_eq!(output[1], 5.0); assert_eq!(output[2], 8.0); }
}