+
Skip to content

Always attempting allocation reuse during inference #673

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Apr 7, 2023
Merged

Always attempting allocation reuse during inference #673

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7aadcf6
#671 reusing allocations during inference
coreylowman Apr 7, 2023
4498309
Speeding up batchnorm2d inference
coreylowman Apr 7, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion benches/batchnorm2d.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,11 @@ fn main() {
loop {
let img: Tensor<InputShape, Dtype, _> = dev.sample_normal();

let start = Instant::now();
let _ = m.forward(img.clone());
dev.synchronize();
let infer_dur = start.elapsed();

let start = Instant::now();
let out = m.forward_mut(img.traced(grads));
let loss = out.square().mean();
Expand All @@ -36,6 +41,6 @@ fn main() {
grads = loss.backward();
dev.synchronize();
let bwd_dur = start.elapsed();
println!("fwd={:?} bwd={:?}", fwd_dur, bwd_dur);
println!("infer={infer_dur:?}, fwd={fwd_dur:?} bwd={bwd_dur:?}");
}
}
8 changes: 7 additions & 1 deletion benches/conv2d.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,11 @@ fn main() {
loop {
let img: Tensor<InputShape, Dtype, _> = dev.sample_normal();

let start = Instant::now();
let _ = m.forward(img.clone());
dev.synchronize();
let infer_dur = start.elapsed();

let start = Instant::now();
let out = m.forward_mut(img.leaky_traced());
let loss = out.square().mean();
Expand All @@ -38,7 +43,8 @@ fn main() {
let _ = loss.backward();
dev.synchronize();
let bwd_dur = start.elapsed();
println!("fwd={:?} bwd={:?}", fwd_dur, bwd_dur);

println!("infer={infer_dur:?}, fwd={fwd_dur:?} bwd={bwd_dur:?}");
}
}

Expand Down
13 changes: 10 additions & 3 deletions benches/softmax.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,17 +23,24 @@ fn main() {

loop {
let img: Tensor<InputShape, Dtype, _> = dev.sample_normal();
let grads = Gradients::leaky();

let start = Instant::now();
let y = img.traced(grads).softmax::<Ax>();
let _ = img.softmax::<Ax>();
dev.synchronize();
let infer_dur = start.elapsed();

let img: Tensor<InputShape, Dtype, _> = dev.sample_normal();

let start = Instant::now();
let y = img.leaky_traced().softmax::<Ax>();
dev.synchronize();
let fwd_dur = start.elapsed();

let start = Instant::now();
let _ = y.sum().backward();
dev.synchronize();
let bwd_dur = start.elapsed();
println!("fwd={:?} bwd={:?}", fwd_dur, bwd_dur);

println!("infer={infer_dur:?}, fwd={fwd_dur:?} bwd={bwd_dur:?}");
}
}
8 changes: 7 additions & 1 deletion benches/sum.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,11 @@ fn main() {
let img: Tensor<InputShape, Dtype, _> = dev.sample_normal();
let grads = Gradients::leaky();

let start = Instant::now();
let _ = img.clone().sum::<(), _>();
dev.synchronize();
let infer_dur = start.elapsed();

let start = Instant::now();
let y = img.traced(grads).sum();
dev.synchronize();
Expand All @@ -33,6 +38,7 @@ fn main() {
let _ = y.backward();
dev.synchronize();
let bwd_dur = start.elapsed();
println!("fwd={:?} bwd={:?}", fwd_dur, bwd_dur);

println!("infer={infer_dur:?}, fwd={fwd_dur:?} bwd={bwd_dur:?}");
}
}
7 changes: 4 additions & 3 deletions src/nn/batchnorm2d.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -79,12 +79,13 @@ where
let shape = *x.shape();

// statistics for normalizing
let std = (var.clone().try_add(epsilon)?).try_sqrt()?;
let std = var.clone().try_add(epsilon)?.try_sqrt()?;

let scale = scale.clone().try_div(std)?.try_broadcast_like(&shape)?;

// normalize & affine
let x = x.try_sub(mean.clone().try_broadcast_like(&shape)?)?;
let x = x.try_div(std.try_broadcast_like(&shape)?)?;
let x = x.try_mul(scale.clone().try_broadcast_like(&shape)?)?;
let x = x.try_mul(scale)?;
x.try_add(bias.clone().try_broadcast_like(&shape)?)
}

Expand Down
4 changes: 2 additions & 2 deletions src/tensor_ops/utilities/ops.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ pub(crate) fn try_unary_op<
let (inp, mut tape) = inp.split_tape();
let inp_ghost = inp.ghost();
let dev = inp.device.clone();
if D::BACKWARD_WITHOUT_DATA {
if !T::OWNS_TAPE || D::BACKWARD_WITHOUT_DATA {
let out = inp_ghost.dev.forward(op.clone(), Err(inp))?;
let out_ghost = out.ghost();
tape.add_backward_op(move |grads| {
Expand Down Expand Up @@ -105,7 +105,7 @@ pub(crate) fn try_binary_op<
let lhs_ghost = lhs.ghost();
let rhs_ghost = rhs.ghost();
let mut tape = ltape.merge(rtape);
if D::BACKWARD_WITHOUT_DATA {
if !LhsTape::OWNS_TAPE || D::BACKWARD_WITHOUT_DATA {
let out = lhs_ghost.dev.forward(op, Err(lhs), Err(rhs))?;
let out_ghost = out.ghost();
tape.add_backward_op(move |grads| {
Expand Down
点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载