Skip to content

Add a pass that fuses matmul and transpose operations #567

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

Open
FloatingcloudKnight wants to merge 1 commit into
base: main
Choose a base branch
from
+391 −41
Open

Add a pass that fuses matmul and transpose operations #567

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
45 changes: 28 additions & 17 deletions examples/BuddyNext/makefile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1235,54 +1235,65 @@ next-compass-run:
-shared-libs=${MLIR_RUNNER_UTILS} \
-shared-libs=${MLIR_C_RUNNER_UTILS}

tosa-matmul-transpose2-lower:
@${BUDDY_OPT} ./tosa-matmultranspose2.mlir \
next-matmul-transpose2-lower:
@${BUDDY_OPT} ./next-matmul-transpose2.mlir \
-transpose-fusion-vectorization \
-o log.mlir

tosa-matmul-transpose2-run:
@${BUDDY_OPT} ./tosa-matmultranspose2.mlir \
next-matmul-transpose2-run:
@${BUDDY_OPT} ./next-matmul-transpose2.mlir \
-pass-pipeline "builtin.module(transpose-fusion-vectorization, func.func(tosa-to-linalg-named, tosa-to-linalg, tosa-to-tensor, tosa-to-arith))" | \
${BUDDY_OPT} \
-convert-elementwise-to-linalg \
-arith-expand \
-eliminate-empty-tensors \
-convert-tensor-to-linalg \
-linalg-bufferize \
-empty-tensor-to-alloc-tensor \
-one-shot-bufferize="bufferize-function-boundaries" \
-convert-linalg-to-affine-loops \
-lower-affine \
-func-bufferize \
-arith-bufferize \
-tensor-bufferize \
-buffer-deallocation \
-finalizing-bufferize \
-affine-loop-fusion \
-affine-parallelize \
-convert-scf-to-openmp \
-convert-vector-to-scf \
-expand-strided-metadata \
-lower-affine \
-cse \
-convert-vector-to-llvm \
-memref-expand \
-arith-expand \
-convert-arith-to-llvm \
-finalize-memref-to-llvm \
-convert-scf-to-cf \
-convert-cf-to-llvm \
-convert-openmp-to-llvm \
-convert-arith-to-llvm \
-convert-math-to-llvm \
-convert-math-to-libm \
-convert-func-to-llvm \
-reconcile-unrealized-casts | \
${MLIR_CPU_RUNNER} ${OPT_FLAG} -e main -entry-point-result=void \
-shared-libs=${MLIR_RUNNER_UTILS} -shared-libs=${MLIR_C_RUNNER_UTILS}

tosa-matmul-transpose2-vec-run:
@${BUDDY_OPT} ./tosa-matmultranspose2-vec.mlir\
next-matmul-transpose2-vec-run:
@${BUDDY_OPT} ./next-matmul-transpose2-vec.mlir \
-convert-linalg-to-affine-loops \
-affine-loop-fusion \
-affine-parallelize \
-lower-affine \
-convert-scf-to-openmp \
-convert-vector-to-scf \
-expand-strided-metadata \
-lower-affine \
-cse \
-convert-vector-to-llvm \
-memref-expand \
-arith-expand \
-convert-arith-to-llvm \
-finalize-memref-to-llvm \
-finalize-memref-to-llvm \
-convert-scf-to-cf \
-convert-cf-to-llvm \
-convert-openmp-to-llvm \
-convert-arith-to-llvm \
-convert-math-to-llvm \
-convert-math-to-libm \
-convert-math-to-libm \
-convert-func-to-llvm \
-reconcile-unrealized-casts | \
${MLIR_CPU_RUNNER} ${OPT_FLAG} -e main -entry-point-result=void \
Expand Down
21 changes: 10 additions & 11 deletions examples/BuddyNext/next-matmul-transpose2-vec.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2,15 +2,13 @@
// RUN: -convert-linalg-to-affine-loops \
// RUN: -lower-affine \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// RUN: -convert-scf-to-cf \
// RUN: -convert-cf-to-llvm \
// RUN: -convert-vector-to-llvm \
// RUN: -convert-math-to-llvm \
// RUN: -convert-math-to-libm \
// RUN: -finalize-memref-to-llvm \
// RUN: -convert-arith-to-llvm \
// RUN: -convert-func-to-llvm \
// RUN: -expand-strided-metadata \
// RUN: -finalize-memref-to-llvm \
// RUN: -reconcile-unrealized-casts \
// RUN: | mlir-runner -e main -entry-point-result=void \
// RUN: -shared-libs=%mlir_runner_utils_dir/libmlir_runner_utils%shlibext \
Expand All @@ -28,21 +26,21 @@ func.func @test(%a : memref<?x?x?xf32>, %b : memref<?x?x?xf32>, %c : memref<?x?x
%vl_step = arith.constant 32 : index
%c0_f32 = arith.constant 0.000000e+00 : f32
%v0 = vector.splat %c0_f32 : vector<32xf32>
%dim = memref.dim %a, %c0 : memref<?x?x?xf32> //
%dim = memref.dim %a, %c0 : memref<?x?x?xf32> //
%dim_0 = memref.dim %a, %c1 : memref<?x?x?xf32>
%dim_1 = memref.dim %a, %c2 : memref<?x?x?xf32>
%dim_2 = memref.dim %b, %c2 : memref<?x?x?xf32>

// Calculate the upper bound for vectorized processing
// - Subtract `vl_step` is to avoid overflow at the vectorization tail.
// - Add 1 to ensure the final loop runs when the workload length
// - Add 1 to ensure the final loop runs when the workload length
// is divisible by the vector size.
%dim_2_upbound_tmp = arith.subi %dim_2, %vl_step : index
%dim_2_upbound = arith.addi %dim_2_upbound_tmp, %c1 : index

affine.for %arg3 = %c0 to %dim {
affine.for %arg4 = %c0 to %dim_0 {
%iter_idx = scf.for %arg5 = %c0 to %dim_2_upbound
%iter_idx = scf.for %arg5 = %c0 to %dim_2_upbound
step %vl_step iter_args(%iter_init = %c0) -> (index){
%0 = vector.load %c[%arg4, %arg3, %arg5] : memref<?x?x?xf32>, vector<32xf32>
%iter_value = scf.for %arg6 = %c0 to %dim_1 step %c1 iter_args(%value_init = %0) -> (vector<32xf32>){
Expand All @@ -57,9 +55,9 @@ func.func @test(%a : memref<?x?x?xf32>, %b : memref<?x?x?xf32>, %c : memref<?x?x
scf.yield %idx_next : index
}

// Compute the tail size and Process the remaining elements
// Compute the tail size and Process the remaining elements
// using masked vector operations.
%tail_size = arith.subi %dim_1, %iter_idx : index
%tail_size = arith.subi %dim_2, %iter_idx : index
%mask = vector.create_mask %tail_size : vector<32xi1>
%0 = vector.maskedload %c[%arg4, %arg3, %iter_idx], %mask, %v0 : memref<?x?x?xf32>, vector<32xi1>, vector<32xf32> into vector<32xf32>
%iter_value = scf.for %arg6 = %c0 to %dim_1 step %c1 iter_args(%value_init = %0) -> (vector<32xf32>){
Expand All @@ -72,7 +70,7 @@ func.func @test(%a : memref<?x?x?xf32>, %b : memref<?x?x?xf32>, %c : memref<?x?x
vector.maskedstore %c[%arg4, %arg3, %iter_idx], %mask, %iter_value : memref<?x?x?xf32>, vector<32xi1>, vector<32xf32>
}
}

%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64
// Print timings.
Expand Down Expand Up @@ -111,8 +109,9 @@ func.func @main(){

%printed_m2 = memref.cast %m2 : memref<?x?x?xf32> to memref<*xf32>

// CHECK: Unranked Memref base@ = {{.*}} rank = 3 offset = 0 sizes = [40, 32, 128] strides = [4096, 128, 1] data =
// CHECK: Unranked Memref base@ = {{.*}} rank = 3 offset = 0 sizes = [40, 32, 128] strides = [4096, 128, 1] data =
// CHECK-NEXT: [
// CHECK: [
// CHECK: [240{{(, 240)*}}]
call @printMemrefF32(%printed_m2) : (memref<*xf32>) -> ()

Expand Down
35 changes: 22 additions & 13 deletions examples/BuddyNext/next-matmul-transpose2.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,21 +1,32 @@
// RUN: buddy-opt %s \
// RUN: -pass-pipeline "builtin.module(func.func(tosa-to-linalg-named),func.func(tosa-to-linalg),func.func(tosa-to-tensor),func.func(tosa-to-arith))" \
// RUN: | buddy-opt \
// RUN: -convert-elementwise-to-linalg \
// RUN: -arith-expand \
// RUN: -eliminate-empty-tensors \
// RUN: -convert-tensor-to-linalg \
// RUN: -empty-tensor-to-alloc-tensor \
// RUN: -one-shot-bufferize="bufferize-function-boundaries" \
// RUN: -convert-linalg-to-affine-loops \
// RUN: -lower-affine \
// RUN: -affine-loop-fusion \
// RUN: -affine-parallelize \
// RUN: -convert-scf-to-openmp \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// RUN: -convert-vector-to-llvm \
// RUN: -convert-arith-to-llvm \
// RUN: -finalize-memref-to-llvm \
// RUN: -convert-scf-to-cf \
// RUN: -convert-cf-to-llvm \
// RUN: -convert-arith-to-llvm \
// RUN: -convert-func-to-llvm \
// RUN: -reconcile-unrealized-casts \
// RUN: -lower-affine \
// RUN: -cse \
// RUN: -convert-vector-to-llvm \
// RUN: -memref-expand \
// RUN: -arith-expand \
// RUN: -convert-arith-to-llvm \
// RUN: -finalize-memref-to-llvm \
// RUN: -convert-scf-to-cf \
// RUN: -convert-cf-to-llvm \
// RUN: -convert-openmp-to-llvm \
// RUN: -convert-arith-to-llvm \
// RUN: -convert-math-to-llvm \
// RUN: -convert-math-to-libm \
// RUN: -convert-func-to-llvm \
// RUN: -reconcile-unrealized-casts \
// RUN: | mlir-runner -e main -entry-point-result=void \
// RUN: -shared-libs=%mlir_runner_utils_dir/libmlir_runner_utils%shlibext \
// RUN: -shared-libs=%mlir_runner_utils_dir/libmlir_c_runner_utils%shlibext \
Expand All @@ -32,7 +43,7 @@ func.func @test(%a : tensor<1x40x32x128xf32>, %b : tensor<32x40x40xf32>) -> (ten
%3 = tosa.matmul %b, %2 : (tensor<32x40x40xf32>, tensor<32x40x128xf32>) -> tensor<32x40x128xf32>
%4 = tosa.reshape %3 {new_shape = array<i64: 1, 32, 40, 128>} : (tensor<32x40x128xf32>) -> tensor<1x32x40x128xf32>
%5 = "tosa.const"() <{value = dense<[0, 2, 1, 3]> : tensor<4xi32>}> : () -> tensor<4xi32>
%6 = tosa.transpose %4, %5 : (tensor<1x32x40x128xf32>, tensor<4xi32>) -> tensor<1x40x32x128xf32>
%6 = tosa.transpose %4, %5 : (tensor<1x32x40x128xf32>, tensor<4xi32>) -> tensor<1x40x32x128xf32>
%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64
// Print timings.
Expand All @@ -44,8 +55,6 @@ func.func @main(){

%v2 = arith.constant dense<2.0> : tensor<32x40x40xf32>
%v3 = arith.constant dense<3.0> : tensor<1x40x32x128xf32>
// %m0 = tensor.cast %v2 : tensor<32x40x40xf32> to tensor<?x?x?xf32>
// %m1 = tensor.cast %v3 : tensor<40x32x128xf32> to tensor<?x?x?xf32>

%m2 = call @test(%v3, %v2) : (tensor<1x40x32x128xf32>, tensor<32x40x40xf32>) -> (tensor<1x40x32x128xf32>)

Expand Down
1 change: 1 addition & 0 deletions midend/lib/Conversion/TransposeOptimization/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,5 +1,6 @@
add_mlir_library(TransposeOptimization
BuiltinTransposeVectorization.cpp
TransposeFusionVectorization.cpp
LINK_LIBS PUBLIC
BuddyUtils
)
Loading