refactor reduce into separate traits (#1798)

Author: louisfd

crates/burn-jit/src/kernel/reduce/base.rs

@@ -1,78 +1,15 @@
-use burn_cube::dialect::{Item, Scope, Variable};
-
 #[cfg(feature = "autotune")]
 use crate::kernel::reduce::reduce_dim_autotune;
 use crate::{element::JitElement, tensor::JitTensor, JitRuntime};
 
-use super::{reduce_dim_naive, reduce_dim_shared, ArgMax, ArgMin, MeanDim, ProdDim, SumDim};
-
-/// Specifies the reduce dim algorithm in use
-pub trait ReduceDimAlgorithm<E: JitElement>: Send + Sync + 'static {
-    /// The reduction accumulator
-    type Accumulator: Copy;
-
-    /// Initialization for naive algorithm
-    fn initialize_naive(
-        scope: &mut Scope,
-        input_item: Item,
-        output_item: Item,
-    ) -> Self::Accumulator;
-
-    /// Inner loop for naive algorithm
-    fn inner_loop_naive(
-        scope: &mut Scope,
-        accumulator: Self::Accumulator,
-        current_value: Variable,
-        i: Variable,
-    );
-
-    /// Assignation for naive algorithm
-    fn assign_naive(
-        scope: &mut Scope,
-        output: Variable,
-        accumulator: Self::Accumulator,
-        shape_reduce_dim: Variable,
-    );
+use super::{
+    naive::{base::ReduceDimNaive, shader::reduce_dim_naive},
+    shared::{base::ReduceDimShared, shader::reduce_dim_shared},
+};
 
-    /// Initialization for shared algorithm
-    fn initialize_shared(
-        scope: &mut Scope,
-        shared_memory_size: u32,
-        write_position: Variable,
-        input_item: Item,
-    ) -> Self::Accumulator;
-
-    /// How to write to shared memory
-    fn write_to_shared(
-        scope: &mut Scope,
-        shared_memory: Self::Accumulator,
-        write_position: Variable,
-        value: Self::Accumulator,
-    );
-
-    /// How to read from input in shared algorithm
-    fn read_from_input(
-        scope: &mut Scope,
-        input: Variable,
-        read_position: Variable,
-        i: Variable,
-    ) -> Self::Accumulator;
-
-    /// How to read from shared memory
-    fn read_from_shared(
-        scope: &mut Scope,
-        shared_memory: Self::Accumulator,
-        read_position: Variable,
-    ) -> Self::Accumulator;
-
-    /// How to assign from shared memory
-    fn assign_shared(
-        scope: &mut Scope,
-        shared_memory: Self::Accumulator,
-        output: Variable,
-        write_position: Variable,
-        shape_reduce_dim: Variable,
-    );
+pub(crate) trait ReduceDimAlgorithm<E: JitElement>:
+    ReduceDimNaive<E> + ReduceDimShared<E>
+{
 }
 
 /// Creates an empty output tensor with reduce output shape
@@ -116,7 +53,10 @@ impl Default for ReduceStrategy {
 }
 
 macro_rules! reduce_operation {
-    ($name:ident, $ops:ty) => {
+    ($name:ident, $ops:ident) => {
+        pub(crate) struct $ops;
+        impl<E: JitElement> ReduceDimAlgorithm<E> for $ops {}
+
         /// Executes the reduce operation with the given strategy.
         pub fn $name<R: JitRuntime, EI: JitElement, EO: JitElement, const D: usize>(
             tensor: JitTensor<R, EI, D>,
@@ -143,5 +83,5 @@ macro_rules! reduce_operation {
 reduce_operation!(sum_dim, SumDim);
 reduce_operation!(mean_dim, MeanDim);
 reduce_operation!(prod_dim, ProdDim);
-reduce_operation!(argmin, ArgMin);
-reduce_operation!(argmax, ArgMax);
+reduce_operation!(argmin, Argmin);
+reduce_operation!(argmax, Argmax);

crates/burn-jit/src/kernel/reduce/mod.rs

@@ -1,23 +1,11 @@
-mod argmax_dim;
-mod argmin_dim;
 mod base;
-mod mean_dim;
-mod naive_reduce_shader;
+mod naive;
 mod prod;
-mod prod_dim;
-mod shared_reduce_shader;
+mod shared;
 mod sum;
-mod sum_dim;
 mod tune;
 
-pub(crate) use argmax_dim::*;
-pub(crate) use argmin_dim::*;
 pub use base::*;
-pub(crate) use mean_dim::*;
-pub use naive_reduce_shader::*;
 pub use prod::*;
-pub(crate) use prod_dim::*;
-pub use shared_reduce_shader::*;
 pub use sum::*;
-pub(crate) use sum_dim::*;
 pub use tune::*;

crates/burn-jit/src/kernel/reduce/naive/argmax.rs

@@ -0,0 +1,49 @@
+use crate::{kernel::reduce::Argmax, JitElement};
+use burn_cube::{
+    cpa,
+    dialect::{Elem, Item, Scope, Variable},
+};
+
+use super::base::ReduceDimNaive;
+
+impl<E: JitElement> ReduceDimNaive<E> for Argmax {
+    type Accumulator = (Variable, Variable);
+
+    fn initialize_naive(
+        scope: &mut Scope,
+        input_item: Item,
+        _output_item: Item,
+    ) -> Self::Accumulator {
+        let index = scope.create_local(Elem::UInt);
+        let max = scope.create_local(input_item);
+        let max_initial =
+            Variable::ConstantScalar(E::minimum_value().to_f64().unwrap(), input_item.elem());
+        cpa!(scope, max = max_initial);
+
+        (max, index)
+    }
+
+    fn inner_loop_naive(
+        scope: &mut Scope,
+        (max, index): Self::Accumulator,
+        value: Variable,
+        i: Variable,
+    ) {
+        let condition = scope.create_local(Elem::Bool);
+        cpa!(scope, condition = value > max);
+        cpa!(scope, if(condition).then(|scope| {
+            cpa!(scope, max = value);
+            cpa!(scope, index = i);
+        }));
+    }
+
+    fn assign_naive(
+        scope: &mut Scope,
+        output: Variable,
+        (_max, index): Self::Accumulator,
+        _shape_reduce_dim: Variable,
+    ) {
+        let id = Variable::Id;
+        cpa!(scope, output[id] = index);
+    }
+}

crates/burn-jit/src/kernel/reduce/naive/argmin.rs

@@ -0,0 +1,50 @@
+use burn_cube::{
+    cpa,
+    dialect::{Elem, Item, Scope, Variable},
+};
+
+use crate::{kernel::reduce::Argmin, JitElement};
+
+use super::base::ReduceDimNaive;
+
+impl<E: JitElement> ReduceDimNaive<E> for Argmin {
+    type Accumulator = (Variable, Variable);
+
+    fn initialize_naive(
+        scope: &mut Scope,
+        input_item: Item,
+        _output_item: Item,
+    ) -> Self::Accumulator {
+        let index = scope.create_local(Elem::UInt);
+        let min = scope.create_local(input_item);
+        let min_initial =
+            Variable::ConstantScalar(E::maximum_value().to_f64().unwrap(), input_item.elem());
+        cpa!(scope, min = min_initial);
+
+        (min, index)
+    }
+
+    fn inner_loop_naive(
+        scope: &mut Scope,
+        (min, index): Self::Accumulator,
+        value: Variable,
+        i: Variable,
+    ) {
+        let condition = scope.create_local(Elem::Bool);
+        cpa!(scope, condition = value < min);
+        cpa!(scope, if(condition).then(|scope| {
+            cpa!(scope, min = value);
+            cpa!(scope, index = i);
+        }));
+    }
+
+    fn assign_naive(
+        scope: &mut Scope,
+        output: Variable,
+        (_min, index): Self::Accumulator,
+        _shape_reduce_dim: Variable,
+    ) {
+        let id = Variable::Id;
+        cpa!(scope, output[id] = index);
+    }
+}

crates/burn-jit/src/kernel/reduce/naive/base.rs

@@ -0,0 +1,32 @@
+use burn_cube::dialect::{Item, Scope, Variable};
+
+use crate::JitElement;
+
+/// Specifies the reduce dim algorithm in use
+pub trait ReduceDimNaive<E: JitElement>: Send + Sync + 'static {
+    /// The reduction accumulator
+    type Accumulator: Copy;
+
+    /// Initialization for naive algorithm
+    fn initialize_naive(
+        scope: &mut Scope,
+        input_item: Item,
+        output_item: Item,
+    ) -> Self::Accumulator;
+
+    /// Inner loop for naive algorithm
+    fn inner_loop_naive(
+        scope: &mut Scope,
+        accumulator: Self::Accumulator,
+        current_value: Variable,
+        i: Variable,
+    );
+
+    /// Assignation for naive algorithm
+    fn assign_naive(
+        scope: &mut Scope,
+        output: Variable,
+        accumulator: Self::Accumulator,
+        shape_reduce_dim: Variable,
+    );
+}

crates/burn-jit/src/kernel/reduce/naive/mean_dim.rs

@@ -0,0 +1,32 @@
+use crate::{kernel::reduce::MeanDim, JitElement};
+use burn_cube::{
+    cpa,
+    dialect::{Item, Scope, Variable},
+};
+
+use super::base::ReduceDimNaive;
+
+impl<E: JitElement> ReduceDimNaive<E> for MeanDim {
+    type Accumulator = Variable;
+
+    fn initialize_naive(scope: &mut Scope, _input_item: Item, output_item: Item) -> Variable {
+        scope.zero(output_item)
+    }
+
+    fn inner_loop_naive(scope: &mut Scope, accumulator: Variable, value: Variable, _i: Variable) {
+        cpa!(scope, accumulator += value);
+    }
+
+    fn assign_naive(
+        scope: &mut Scope,
+        output: Variable,
+        accumulator: Variable,
+        shape_reduce_dim: Variable,
+    ) {
+        let id = Variable::Id;
+        let denominator = scope.create_local(accumulator.item());
+        cpa!(scope, denominator = cast(shape_reduce_dim));
+        cpa!(scope, accumulator = accumulator / denominator);
+        cpa!(scope, output[id] = accumulator);
+    }
+}

crates/burn-jit/src/kernel/reduce/naive/mod.rs

@@ -0,0 +1,7 @@
+pub(crate) mod argmax;
+pub(crate) mod argmin;
+pub(crate) mod base;
+pub(crate) mod mean_dim;
+pub(crate) mod prod_dim;
+pub(crate) mod shader;
+pub(crate) mod sum_dim;

crates/burn-jit/src/kernel/reduce/naive/prod_dim.rs

@@ -0,0 +1,29 @@
+use crate::{kernel::reduce::ProdDim, JitElement};
+use burn_cube::{
+    cpa,
+    dialect::{Item, Scope, Variable},
+};
+
+use super::base::ReduceDimNaive;
+
+impl<E: JitElement> ReduceDimNaive<E> for ProdDim {
+    type Accumulator = Variable;
+
+    fn initialize_naive(scope: &mut Scope, _input_item: Item, output_item: Item) -> Variable {
+        scope.create_with_value(1, output_item)
+    }
+
+    fn inner_loop_naive(scope: &mut Scope, accumulator: Variable, value: Variable, _i: Variable) {
+        cpa!(scope, accumulator *= value);
+    }
+
+    fn assign_naive(
+        scope: &mut Scope,
+        output: Variable,
+        accumulator: Variable,
+        _shape_reduce_dim: Variable,
+    ) {
+        let id = Variable::Id;
+        cpa!(scope, output[id] = accumulator);
+    }
+}

crates/burn-jit/src/kernel/reduce/naive_reduce_shader.rs crates/burn-jit/src/kernel/reduce/naive/shader.rs

@@ -8,9 +8,9 @@ use std::marker::PhantomData;
 
 use crate::{element::JitElement, kernel::GpuComputeShaderPhase, tensor::JitTensor, JitRuntime};
 
-use super::ReduceDimAlgorithm;
+use super::base::ReduceDimNaive;
 
-pub(crate) struct NaiveReduceDimComputeShader<E: JitElement, RD: ReduceDimAlgorithm<E>> {
+pub(crate) struct NaiveReduceDimComputeShader<E: JitElement, RD: ReduceDimNaive<E>> {
     tensor: Variable,
     dim: usize,
     output: Variable,
@@ -20,7 +20,7 @@ pub(crate) struct NaiveReduceDimComputeShader<E: JitElement, RD: ReduceDimAlgori
 
 #[derive(new)]
 pub(crate) struct NaiveReduceDimEagerKernel<
-    RD: ReduceDimAlgorithm<EI>,
+    RD: ReduceDimNaive<EI>,
     R: JitRuntime,
     EI: JitElement,
     EO: JitElement,
@@ -32,8 +32,8 @@ pub(crate) struct NaiveReduceDimEagerKernel<
     _elem_out: PhantomData<EO>,
 }
 
-impl<RD: ReduceDimAlgorithm<EI>, R: JitRuntime, EI: JitElement, EO: JitElement>
-    GpuComputeShaderPhase for NaiveReduceDimEagerKernel<RD, R, EI, EO>
+impl<RD: ReduceDimNaive<EI>, R: JitRuntime, EI: JitElement, EO: JitElement> GpuComputeShaderPhase
+    for NaiveReduceDimEagerKernel<RD, R, EI, EO>
 {
     fn compile(&self) -> ComputeShader {
         let mut scope = Scope::root();
@@ -76,7 +76,7 @@ impl<RD: ReduceDimAlgorithm<EI>, R: JitRuntime, EI: JitElement, EO: JitElement>
     }
 }
 
-impl<E: JitElement, RD: ReduceDimAlgorithm<E>> NaiveReduceDimComputeShader<E, RD> {
+impl<E: JitElement, RD: ReduceDimNaive<E>> NaiveReduceDimComputeShader<E, RD> {
     pub(crate) fn expand(self, scope: &mut Scope) {
         let tensor = self.tensor;
         let dim: Variable = self.dim.into();
@@ -136,7 +136,7 @@ impl<E: JitElement, RD: ReduceDimAlgorithm<E>> NaiveReduceDimComputeShader<E, RD
 
 /// Executes the naive kernel for reduce dim
 pub fn reduce_dim_naive<
-    RD: ReduceDimAlgorithm<EI>,
+    RD: ReduceDimNaive<EI>,
     R: JitRuntime,
     EI: JitElement,
     EO: JitElement,