use crate::device::{DeviceCapabilities, WorkloadType};
use crate::error::Result;
use crate::tensor::Tensor;
pub struct Upsample {
input: Tensor,
size: Option<(usize, usize)>,
scale_factor: Option<(f32, f32)>,
mode: UpsampleMode,
align_corners: bool,
}
#[derive(Debug, Clone, Copy)]
pub enum UpsampleMode {
Nearest,
Bilinear,
}
impl Upsample {
pub fn new(
input: Tensor,
size: Option<(usize, usize)>,
scale_factor: Option<(f32, f32)>,
mode: UpsampleMode,
align_corners: bool,
) -> Result<Self> {
let shape = input.shape();
if shape.len() != 4 {
return Err(crate::error::BarracudaError::InvalidInput {
message: format!("Upsample expects 4D tensor [B, C, H, W], got shape {shape:?}"),
});
}
if size.is_none() && scale_factor.is_none() {
return Err(crate::error::BarracudaError::InvalidInput {
message: "Either size or scale_factor must be provided".to_string(),
});
}
Ok(Self {
input,
size,
scale_factor,
mode,
align_corners,
})
}
fn wgsl_shader() -> &'static str {
{
static SHADER: std::sync::LazyLock<String> = std::sync::LazyLock::new(|| {
include_str!("../shaders/misc/upsample_f64.wgsl").to_string()
});
std::sync::LazyLock::force(&SHADER).as_str()
}
}
pub fn execute(self) -> Result<Tensor> {
let device = self.input.device();
let shape = self.input.shape();
let batch_size = shape[0];
let channels = shape[1];
let in_height = shape[2];
let in_width = shape[3];
let (out_height, out_width) = if let Some((h, w)) = self.size {
(h, w)
} else if let Some((sh, sw)) = self.scale_factor {
(
(in_height as f32 * sh) as usize,
(in_width as f32 * sw) as usize,
)
} else {
return Err(crate::error::BarracudaError::InvalidInput {
message: "Either size or scale_factor must be provided".to_string(),
});
};
let output_size = batch_size * channels * out_height * out_width;
let input_buffer = self.input.buffer();
let output_buffer = device.create_buffer_f32(output_size)?;
#[repr(C)]
#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
struct Params {
batch_size: u32,
channels: u32,
in_height: u32,
in_width: u32,
out_height: u32,
out_width: u32,
mode: u32,
align_corners: u32,
}
let params = Params {
batch_size: batch_size as u32,
channels: channels as u32,
in_height: in_height as u32,
in_width: in_width as u32,
out_height: out_height as u32,
out_width: out_width as u32,
mode: match self.mode {
UpsampleMode::Nearest => 0,
UpsampleMode::Bilinear => 1,
},
align_corners: u32::from(self.align_corners),
};
let params_buffer = device
.device
.create_buffer_init(&wgpu::util::BufferInitDescriptor {
label: Some("Upsample Params"),
contents: bytemuck::cast_slice(&[params]),
usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
});
let shader_module = device.compile_shader(Self::wgsl_shader(), Some("Upsample Shader"));
let bind_group_layout =
device
.device
.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
label: Some("Upsample 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: false },
has_dynamic_offset: false,
min_binding_size: None,
},
count: None,
},
],
});
let bind_group = device.device.create_bind_group(&wgpu::BindGroupDescriptor {
label: Some("Upsample Bind Group"),
layout: &bind_group_layout,
entries: &[
wgpu::BindGroupEntry {
binding: 0,
resource: params_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 1,
resource: input_buffer.as_entire_binding(),
},
wgpu::BindGroupEntry {
binding: 2,
resource: output_buffer.as_entire_binding(),
},
],
});
let pipeline_layout =
device
.device
.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
label: Some("Upsample Pipeline Layout"),
bind_group_layouts: &[&bind_group_layout],
immediate_size: 0,
});
let compute_pipeline =
device
.device
.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
label: Some("Upsample 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("Upsample Encoder"),
});
{
let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
label: Some("Upsample 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 (wg_x, wg_y) = caps.optimal_workgroup_size_2d(WorkloadType::Convolution);
let workgroups_x = (out_width as u32).div_ceil(wg_x);
let workgroups_y = (out_height as u32).div_ceil(wg_y);
let workgroups_z = (batch_size * channels) as u32;
compute_pass.dispatch_workgroups(workgroups_x, workgroups_y, workgroups_z);
}
device.submit_commands(Some(encoder.finish()));
Ok(Tensor::from_buffer(
output_buffer,
vec![batch_size, channels, out_height, out_width],
device.clone(),
))
}
}
#[cfg(test)]
mod tests {
use super::*;
#[tokio::test]
async fn test_upsample_basic() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..12).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![1, 1, 3, 4], device).unwrap();
let upsampled = Upsample::new(input, Some((6, 8)), None, UpsampleMode::Nearest, false)
.unwrap()
.execute()
.unwrap();
assert_eq!(upsampled.shape(), &vec![1, 1, 6, 8]);
}
#[tokio::test]
async fn test_upsample_scale_factor() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..8).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![1, 1, 2, 4], device).unwrap();
let upsampled = Upsample::new(input, None, Some((2.0, 2.0)), UpsampleMode::Bilinear, false)
.unwrap()
.execute()
.unwrap();
assert_eq!(upsampled.shape(), &vec![1, 1, 4, 8]);
}
#[tokio::test]
async fn test_upsample_invalid_shape() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_data(&[1.0, 2.0, 3.0], vec![3], device).unwrap();
assert!(Upsample::new(input, Some((10, 10)), None, UpsampleMode::Nearest, false,).is_err());
}
#[tokio::test]
async fn test_upsample_no_params() {
let device = crate::device::test_pool::get_test_device().await;
let input = Tensor::from_data(&[1.0; 12], vec![1, 1, 3, 4], device).unwrap();
assert!(Upsample::new(input, None, None, UpsampleMode::Nearest, false,).is_err());
}
#[tokio::test]
async fn test_upsample_large() {
let device = crate::device::test_pool::get_test_device().await;
let data: Vec<f32> = (0..256).map(|i| i as f32).collect();
let input = Tensor::from_data(&data, vec![1, 1, 16, 16], device).unwrap();
let upsampled = Upsample::new(input, Some((32, 32)), None, UpsampleMode::Bilinear, true)
.unwrap()
.execute()
.unwrap();
assert_eq!(upsampled.shape(), &vec![1, 1, 32, 32]);
}
}