use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::{BarracudaError, Result};
use crate::tensor::Tensor;
pub struct ComplexDiv {
input_a: Tensor,
input_b: Tensor,
pipeline: wgpu::ComputePipeline,
bind_group_layout: wgpu::BindGroupLayout,
}
impl ComplexDiv {
pub fn new(input_a: Tensor, input_b: Tensor) -> Result<Self> {
if input_a.shape() != input_b.shape() {
return Err(BarracudaError::Device(
"Tensors must have same shape".to_string(),
));
}
if input_a.shape().last() != Some(&2) {
return Err(BarracudaError::Device(
"Must have last dimension = 2".to_string(),
));
}
if !std::ptr::eq(input_a.device().as_ref(), input_b.device().as_ref()) {
return Err(BarracudaError::Device("Must be on same device".to_string()));
}
let device = input_a.device();
let shader = device.compile_shader(include_str!("div.wgsl"), Some("Complex Div Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Complex Div 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 pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Complex Div Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let pipeline = device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Complex Div Pipeline"),
layout: Some(&pipeline_layout),
module: &shader,
entry_point: Some("main"),
cache: None,
compilation_options: Default::default(),
});
Ok(Self {
input_a,
input_b,
pipeline,
bind_group_layout,
})
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input_a.device();
let num_elements = self.input_a.len();
let output_buffer = device.device.create_buffer(&wgpu::BufferDescriptor {
label: Some("Complex Div Output"),
size: (num_elements * std::mem::size_of::<f32>()) as u64,
usage: wgpu::BufferUsages::STORAGE | wgpu::BufferUsages::COPY_SRC,
mapped_at_creation: false,
});
let params = [num_elements as u32 / 2];
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Complex Div Params"),
contents: bytemuck::cast_slice(¶ms),
usage: wgpu::BufferUsages::UNIFORM,
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Complex Div Bind Group"),
layout: &self.bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: self.input_a.buffer().as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: self.input_b.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 mut encoder = device.create_encoder_guarded(&wgpu::CommandEncoderDescriptor {
label: Some("Complex Div Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Complex Div Compute Pass"),
timestamp_writes: None,
});
compute_pass.set_pipeline(&self.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 = ((num_elements / 2) as u32).div_ceil(optimal_wg_size);
compute_pass.dispatch_workgroups(workgroups, 1, 1);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
self.input_a.shape().to_vec(),
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_complex_div() {
let device = crate::device::test_pool::get_test_device().await;
let a = vec![1.0f32, 0.0];
let b = vec![2.0f32, 0.0];
let ta = Tensor::from_data(&a, vec![1, 2], device.clone()).unwrap();
let tb = Tensor::from_data(&b, vec![1, 2], device).unwrap();
let result = ComplexDiv::new(ta, tb)
.unwrap()
.execute()
.unwrap()
.to_vec()
.unwrap();
assert!((result[0] - 0.5).abs() < 1e-5);
assert!((result[1] - 0.0).abs() < 1e-5);
}
}