Skip to content
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

Jump to bottom

feat: simplify operations with sign #336

Draft
wants to merge 4 commits into
base: main
Choose a base branch
from
Draft

feat: simplify operations with sign #336

wants to merge 4 commits into from

Conversation

avik-pal
Copy link
Collaborator

@avik-pal avik-pal commented Feb 11, 2025 •
edited
Loading

Trying to simplify

module @reactant_gradient attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<8x4xf32>, %arg1: tensor<16xf32>) -> tensor<8x4xf32> {
    %cst = stablehlo.constant dense<0.000000e+00> : tensor<f32>
    %cst_0 = stablehlo.constant dense<0.000000e+00> : tensor<32x4x8xf32>
    %0 = stablehlo.broadcast_in_dim %arg0, dims = [2, 1] : (tensor<8x4xf32>) -> tensor<16x4x8xf32>
    %1 = stablehlo.broadcast_in_dim %arg1, dims = [0] : (tensor<16xf32>) -> tensor<16x4x8xf32>
    %2 = stablehlo.multiply %0, %1 : tensor<16x4x8xf32>
    %3 = stablehlo.sine %2 : tensor<16x4x8xf32>
    %4 = stablehlo.cosine %2 : tensor<16x4x8xf32>
    %5 = stablehlo.concatenate %3, %4, dim = 0 : (tensor<16x4x8xf32>, tensor<16x4x8xf32>) -> tensor<32x4x8xf32>
    %6 = stablehlo.abs %5 : tensor<32x4x8xf32>
    %7 = stablehlo.add %6, %6 : tensor<32x4x8xf32>
    %8 = stablehlo.compare  GE, %5, %cst_0 : (tensor<32x4x8xf32>, tensor<32x4x8xf32>) -> tensor<32x4x8xi1>
    %9 = stablehlo.negate %7 : tensor<32x4x8xf32>
    %10 = stablehlo.select %8, %7, %9 : tensor<32x4x8xi1>, tensor<32x4x8xf32>
    %11 = stablehlo.slice %10 [0:16, 0:4, 0:8] : (tensor<32x4x8xf32>) -> tensor<16x4x8xf32>
    %12 = stablehlo.slice %10 [16:32, 0:4, 0:8] : (tensor<32x4x8xf32>) -> tensor<16x4x8xf32>
    %13 = stablehlo.negate %3 : tensor<16x4x8xf32>
    %14 = stablehlo.multiply %12, %13 : tensor<16x4x8xf32>
    %15 = stablehlo.multiply %11, %4 : tensor<16x4x8xf32>
    %16 = stablehlo.add %14, %15 : tensor<16x4x8xf32>
    %17 = stablehlo.multiply %16, %1 : tensor<16x4x8xf32>
    %18 = stablehlo.reduce(%17 init: %cst) applies stablehlo.add across dimensions = [0] : (tensor<16x4x8xf32>, tensor<f32>) -> tensor<4x8xf32>
    %19 = stablehlo.reshape %18 : (tensor<4x8xf32>) -> tensor<8x4xf32>
    return %19 : tensor<8x4xf32>
  }
}

most of the internal operations are actually no-ops (this is from a chunk inside a transformer so kind of a common pattern). After these passes

module @reactant_gradient attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64} {
  func.func @main(%arg0: tensor<8x4xf32>, %arg1: tensor<16xf32>) -> tensor<8x4xf32> {
    %cst = stablehlo.constant dense<0.000000e+00> : tensor<f32>
    %0 = stablehlo.broadcast_in_dim %arg0, dims = [2, 1] : (tensor<8x4xf32>) -> tensor<16x4x8xf32>
    %1 = stablehlo.broadcast_in_dim %arg1, dims = [0] : (tensor<16xf32>) -> tensor<16x4x8xf32>
    %2 = stablehlo.multiply %0, %1 : tensor<16x4x8xf32>
    %3 = stablehlo.sine %2 : tensor<16x4x8xf32>
    %4 = stablehlo.cosine %2 : tensor<16x4x8xf32>
    %5 = stablehlo.concatenate %3, %4, dim = 0 : (tensor<16x4x8xf32>, tensor<16x4x8xf32>) -> tensor<32x4x8xf32>
    %6 = stablehlo.multiply %5, %5 : tensor<32x4x8xf32>
    %7 = stablehlo.negate %6 : tensor<32x4x8xf32>
    %8 = stablehlo.slice %7 [0:16, 0:4, 0:8] : (tensor<32x4x8xf32>) -> tensor<16x4x8xf32>
    %9 = stablehlo.slice %7 [16:32, 0:4, 0:8] : (tensor<32x4x8xf32>) -> tensor<16x4x8xf32>
    %10 = stablehlo.multiply %9, %3 : tensor<16x4x8xf32>
    %11 = stablehlo.negate %10 : tensor<16x4x8xf32>
    %12 = stablehlo.multiply %8, %4 : tensor<16x4x8xf32>
    %13 = stablehlo.add %11, %12 : tensor<16x4x8xf32>
    %14 = stablehlo.multiply %13, %1 : tensor<16x4x8xf32>
    %15 = stablehlo.reduce(%14 init: %cst) applies stablehlo.add across dimensions = [0] : (tensor<16x4x8xf32>, tensor<f32>) -> tensor<4x8xf32>
    %16 = stablehlo.reshape %15 : (tensor<4x8xf32>) -> tensor<8x4xf32>
    return %16 : tensor<8x4xf32>
  }
}

@avik-pal
Copy link
Collaborator Author

Some of these need to be qualified using no_nan

wsmoses
wsmoses reviewed Feb 12, 2025
View reviewed changes
src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp
// (select (x > 0) z (neg z)) -> (mul (sign x) z)
// (select (x >= 0) z (neg z)) -> (mul (sign x) z)
// (select (x > 0) (neg z) z) -> (mul (sign x) (neg z))
// (select (x >= 0) (neg z) z) -> (mul (sign x) (neg z))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

is this actually simpler/faster?

Copy link
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This one mostly enables the other optimizations.

Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

a select is usually faster than a mul, so this in isolation makes things locally worse. Is it feasible for the downstream operations to work on select style forms?

wsmoses
wsmoses reviewed Feb 12, 2025
View reviewed changes
src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp
};

// (mul (sign x) (abs x)) -> x
// (mul (abs x) (sign x)) -> x
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this seems reasonable, can you split to do this individually?

wsmoses
wsmoses reviewed Feb 12, 2025
View reviewed changes
src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp
};

// (mul (neg x) (neg y)) -> (mul x y)
// (mul (neg x) y) -> (neg (mul x y))
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

top one is always good, the one negation variations we should separate since there's a separate question of whether we want to propagate them up or down (e.g. if we had mul (neg x), constant) we'd want to do mul x (-constant)

wsmoses
wsmoses reviewed Feb 12, 2025
View reviewed changes
src/enzyme_ad/jax/Passes/EnzymeHLOOpt.cpp

// This pattern only does partially the following. We rely on transforming the op to a
// pattern which further uses the above pattern.
// (mul (sign x) (add (abs x) (abs x))) -> (mul x x)
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

longer term i feel like this merits a broader sign analysis (alongside perhaps a transpose analysis)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@wsmoses wsmoses wsmoses left review comments

Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@avik-pal @wsmoses