feat: support lowering custom fp types #596
Draft
+208
−70
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
wsmoses
reviewed
Jan 22, 2025
src/TracedRArray.jl
Outdated
| @@ -507,7 +509,7 @@ end | |||
| # we need to override the outer copy method to make sure we never fall back to scalar | |||
| # iteration (see, e.g., CUDA.jl#145) | |||
| function Broadcast.copy(bc::Broadcasted{<:AbstractReactantArrayStyle}) | |||
| fn = if bc.f isa Type && bc.f <: ReactantPrimitive | |||
| fn = if bc.f isa Type && is_reactant_primitive(bc.f) | |||
There was a problem hiding this comment.
We should proably be more careful with this. For example, Ops.add won't necessarily do what one wants
There was a problem hiding this comment.
For custom interface, we probably need:
is_reactant_primitiveprimitive_typeDenseElementsAttribute- a way to map mlir_type to julia type
There was a problem hiding this comment.
The core idea is that we can allow support for external packages like Floats8.jl , https://github.com/JuliaMath/DoubleFloats.jl, etc. without having to pull them into Reactant
avik-pal
force-pushed
the
ap/new_primitives
branch
from
92cc29d to
89ef696
Compare
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
using Float8s, Reactant
Reactant.to_reactant_primitive_type(::Type{Float8_4}) = Reactant.F8E4M3FNUZ
x = rand(Float32, 10, 3) .|> Float8_4
x_ra = Reactant.to_rarray(x)
@code_hlo .+(x_ra, x_ra)module {
func.func @main(%arg0: tensor<3x10xf8E4M3FN>) -> tensor<3x10xf8E4M3FN> {
%0 = stablehlo.add %arg0, %arg0 : tensor<3x10xf8E4M3FN>
return %0 : tensor<3x10xf8E4M3FN>
}
} |
|
noice |
avik-pal
force-pushed
the
ap/new_primitives
branch
from
d68cda9 to
e60e7cf
Compare
julia> @code_hlo optimize=false sum(x_ra)
module {
func.func private @identity_broadcast_scalar(%arg0: tensor<f8E4M3FN>) -> tensor<f8E4M3FN> {
return %arg0 : tensor<f8E4M3FN>
}
func.func @main(%arg0: tensor<3x10xf8E4M3FN>) -> (tensor<f8E4M3FN>, tensor<3x10xf8E4M3FN>) {
%0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<3x10xf8E4M3FN>) -> tensor<10x3xf8E4M3FN>
%cst = stablehlo.constant dense<0.000000e+00> : tensor<f16>
%1 = stablehlo.convert %cst : (tensor<f16>) -> tensor<f8E4M3FN>
%2 = enzyme.batch @identity_broadcast_scalar(%0) {batch_shape = array<i64: 10, 3>} : (tensor<10x3xf8E4M3FN>) -> tensor<10x3xf8E4M3FN>
%3 = stablehlo.convert %2 : tensor<10x3xf8E4M3FN>
%4 = stablehlo.reduce(%3 init: %1) applies stablehlo.add across dimensions = [0, 1] : (tensor<10x3xf8E4M3FN>, tensor<f8E4M3FN>) -> tensor<f8E4M3FN>
%5 = stablehlo.transpose %0, dims = [1, 0] : (tensor<10x3xf8E4M3FN>) -> tensor<3x10xf8E4M3FN>
return %4, %5 : tensor<f8E4M3FN>, tensor<3x10xf8E4M3FN>
}
}
julia> @code_hlo optimize=false .+(x_ra, 1)
module {
func.func private @"+_broadcast_scalar"(%arg0: tensor<f8E4M3FN>, %arg1: tensor<i64>) -> (tensor<f8E4M3FN>, tensor<f8E4M3FN>, tensor<i64>) {
%0 = stablehlo.convert %arg1 : (tensor<i64>) -> tensor<f8E4M3FN>
%1 = stablehlo.add %arg0, %0 : tensor<f8E4M3FN>
return %1, %arg0, %arg1 : tensor<f8E4M3FN>, tensor<f8E4M3FN>, tensor<i64>
}
func.func @main(%arg0: tensor<3x10xf8E4M3FN>) -> (tensor<3x10xf8E4M3FN>, tensor<3x10xf8E4M3FN>) {
%0 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<3x10xf8E4M3FN>) -> tensor<10x3xf8E4M3FN>
%c = stablehlo.constant dense<1> : tensor<10x3xi64>
%1:3 = enzyme.batch @"+_broadcast_scalar"(%0, %c) {batch_shape = array<i64: 10, 3>} : (tensor<10x3xf8E4M3FN>, tensor<10x3xi64>) -> (tensor<10x3xf8E4M3FN>, tensor<10x3xf8E4M3FN>, tensor<10x3xi64>)
%2 = stablehlo.convert %1#0 : tensor<10x3xf8E4M3FN>
%3 = stablehlo.transpose %2, dims = [1, 0] : (tensor<10x3xf8E4M3FN>) -> tensor<3x10xf8E4M3FN>
%4 = stablehlo.transpose %1#1, dims = [1, 0] : (tensor<10x3xf8E4M3FN>) -> tensor<3x10xf8E4M3FN>
return %3, %4 : tensor<3x10xf8E4M3FN>, tensor<3x10xf8E4M3FN>
}
}But post optimization stablehlo seems to crash |
|
I also removed some of the primitives from ReactantPrimitive since they are not part of the StableHLO spec. I can restore them if we need them. Specifically
|
|
Open an issue for that on Enzyme-JaX |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.