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

Llama2 model Operator/Layer level instance extraction #353

Open
wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Llama2 model Operator/Layer level instance extraction #353

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
01f83f5
Llama2 model Operator/Layer level instance extraction
Jul 31, 2024
3033e0f
Llama2 model Operator/Layer level instance extraction
Aug 1, 2024
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
408 changes: 408 additions & 0 deletions examples/BuddyNext/README.md
View file
Open in desktop

Large diffs are not rendered by default.

431 changes: 431 additions & 0 deletions examples/BuddyNext/makefile
View file
Open in desktop

Large diffs are not rendered by default.

83 changes: 83 additions & 0 deletions examples/BuddyNext/next-fc.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,83 @@
// 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: -arith-expand \
// RUN: -eliminate-empty-tensors \
// RUN: -empty-tensor-to-alloc-tensor \
// RUN: -one-shot-bufferize \
// RUN: -convert-linalg-to-affine-loops \
// RUN: -affine-loop-fusion \
// RUN: -lower-affine \
// RUN: -func-bufferize \
// RUN: -arith-bufferize \
// RUN: -tensor-bufferize \
// RUN: -buffer-deallocation \
// RUN: -finalizing-bufferize \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// 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-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-cpu-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 \
// RUN: | FileCheck %s

func.func private @rtclock() -> f64

func.func @kernel_fc_layer(%arg0 : tensor<1x40x4096xf32>, %arg1 : tensor<4096x4096xf32>, %arg2 : tensor<4096x4096xf32>, %arg3 : tensor<1x40x4096xf32>) {
%t_start = call @rtclock() : () -> f64

%cst_0 = arith.constant dense<0.0> : tensor<40x4096xf32>
%cst_1 = arith.constant dense<0.0> : tensor<40x4096xf32>

%41 = tosa.mul %arg0, %arg3 {shift = 0 : i8} : (tensor<1x40x4096xf32>, tensor<1x40x4096xf32>) -> tensor<1x40x4096xf32>
%42 = "tosa.const"() <{value = dense<[1, 0]> : tensor<2xi32>}> : () -> tensor<2xi32>
%43 = tosa.transpose %arg1, %42 : (tensor<4096x4096xf32>, tensor<2xi32>) -> tensor<4096x4096xf32>
%44 = tosa.reshape %41 {new_shape = array<i64: 40, 4096>} : (tensor<1x40x4096xf32>) -> tensor<40x4096xf32>
%45 = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%44, %43 : tensor<40x4096xf32>, tensor<4096x4096xf32>) outs(%cst_0 : tensor<40x4096xf32>) -> tensor<40x4096xf32>
%46 = tosa.reshape %45 {new_shape = array<i64: 1, 40, 4096>} : (tensor<40x4096xf32>) -> tensor<1x40x4096xf32>

%47 = "tosa.const"() <{value = dense<[1, 0]> : tensor<2xi32>}> : () -> tensor<2xi32>
%48 = tosa.transpose %arg2, %47 : (tensor<4096x4096xf32>, tensor<2xi32>) -> tensor<4096x4096xf32>
%49 = tosa.reshape %41 {new_shape = array<i64: 40, 4096>} : (tensor<1x40x4096xf32>) -> tensor<40x4096xf32>
%50 = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%49, %48 : tensor<40x4096xf32>, tensor<4096x4096xf32>) outs(%cst_1 : tensor<40x4096xf32>) -> tensor<40x4096xf32>
%51 = tosa.reshape %50 {new_shape = array<i64: 1, 40, 4096>} : (tensor<40x4096xf32>) -> tensor<1x40x4096xf32>

%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64

%tensor_unranked = tensor.cast %51 : tensor<1x40x4096xf32> to tensor<*xf32>

// CHECK: Unranked Memref base@ = {{.*}} rank = 3 offset = 0 sizes = [1, 40, 4096] strides = [163840, 4096, 1] data =
// CHECK-NEXT: [
// CHECK-SAME: [
// CHECK-SAME: [49152{{(, 49152)*}}],

call @printMemrefF32(%tensor_unranked) : (tensor<*xf32>) -> ()
vector.print %time : f64

return
}

func.func @main() {
%input_tensor_1 = arith.constant dense<3.0> : tensor<1x40x4096xf32>
%input_tensor_2 = arith.constant dense<2.0> : tensor<4096x4096xf32>
%input_tensor_3 = arith.constant dense<1.0> : tensor<4096x4096xf32>
%input_tensor_4 = arith.constant dense<4.0> : tensor<1x40x4096xf32>

call @kernel_fc_layer(%input_tensor_1, %input_tensor_2, %input_tensor_3, %input_tensor_4) : (tensor<1x40x4096xf32>, tensor<4096x4096xf32>, tensor<4096x4096xf32>, tensor<1x40x4096xf32>) -> ()

return
}

func.func private @printMemrefF32(%ptr : tensor<*xf32>)
104 changes: 104 additions & 0 deletions examples/BuddyNext/next-ffn.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,104 @@
// 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: -arith-expand \
// RUN: -eliminate-empty-tensors \
// RUN: -empty-tensor-to-alloc-tensor \
// RUN: -one-shot-bufferize \
// RUN: -convert-linalg-to-affine-loops \
// RUN: -affine-loop-fusion \
// RUN: -lower-affine \
// RUN: -func-bufferize \
// RUN: -arith-bufferize \
// RUN: -tensor-bufferize \
// RUN: -buffer-deallocation \
// RUN: -finalizing-bufferize \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// 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-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-cpu-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 \
// RUN: | FileCheck %s

#map = affine_map<(d0, d1, d2) -> (d1)>
#map1 = affine_map<(d0, d1, d2) -> (d0, d2)>
#map2 = affine_map<(d0, d1, d2) -> (d0, d1)>
#map3 = affine_map<(d0, d1) -> (d0, d1)>
#map4 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
#map5 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
#map6 = affine_map<(d0, d1, d2) -> (d0, 0, d1, d2)>
#map7 = affine_map<(d0, d1) -> (0, d0, d1)>

func.func private @rtclock() -> f64

func.func @kernel_ffn(%arg0: tensor<1x40x4096xf32>, %arg9: tensor<4096xf32>, %arg10: tensor<11008x4096xf32>, %arg11: tensor<11008x4096xf32>, %arg12: tensor<4096x11008xf32>) {
%t_start = call @rtclock() : () -> f64

// FFN
%138 = tosa.reshape %arg9 {new_shape = array<i64: 1, 1, 4096>} : (tensor<4096xf32>) -> tensor<1x1x4096xf32>
%139 = tosa.mul %138, %arg0 {shift = 0 : i8} : (tensor<1x1x4096xf32>, tensor<1x40x4096xf32>) -> tensor<1x40x4096xf32>
%140 = "tosa.const"() <{value = dense<[1, 0]> : tensor<2xi32>}> : () -> tensor<2xi32>
%141 = tosa.transpose %arg10, %140 : (tensor<11008x4096xf32>, tensor<2xi32>) -> tensor<4096x11008xf32>
%142 = tosa.reshape %139 {new_shape = array<i64: 40, 4096>} : (tensor<1x40x4096xf32>) -> tensor<40x4096xf32>
%cst_24 = arith.constant dense<0.000000e+00> : tensor<40x11008xf32>
%143 = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%142, %141 : tensor<40x4096xf32>, tensor<4096x11008xf32>) outs(%cst_24 : tensor<40x11008xf32>) -> tensor<40x11008xf32>
%144 = tosa.reshape %143 {new_shape = array<i64: 1, 40, 11008>} : (tensor<40x11008xf32>) -> tensor<1x40x11008xf32>
%145 = tosa.sigmoid %144 : (tensor<1x40x11008xf32>) -> tensor<1x40x11008xf32>
%146 = tosa.mul %144, %145 {shift = 0 : i8} : (tensor<1x40x11008xf32>, tensor<1x40x11008xf32>) -> tensor<1x40x11008xf32>
%147 = "tosa.const"() <{value = dense<[1, 0]> : tensor<2xi32>}> : () -> tensor<2xi32>
%148 = tosa.transpose %arg11, %147 : (tensor<11008x4096xf32>, tensor<2xi32>) -> tensor<4096x11008xf32>
%149 = tosa.reshape %139 {new_shape = array<i64: 40, 4096>} : (tensor<1x40x4096xf32>) -> tensor<40x4096xf32>
%cst_25 = arith.constant dense<0.000000e+00> : tensor<40x11008xf32>
%150 = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%149, %148 : tensor<40x4096xf32>, tensor<4096x11008xf32>) outs(%cst_25 : tensor<40x11008xf32>) -> tensor<40x11008xf32>
%151 = tosa.reshape %150 {new_shape = array<i64: 1, 40, 11008>} : (tensor<40x11008xf32>) -> tensor<1x40x11008xf32>
%152 = tosa.mul %146, %151 {shift = 0 : i8} : (tensor<1x40x11008xf32>, tensor<1x40x11008xf32>) -> tensor<1x40x11008xf32>
%153 = "tosa.const"() <{value = dense<[1, 0]> : tensor<2xi32>}> : () -> tensor<2xi32>
%154 = tosa.transpose %arg12, %153 : (tensor<4096x11008xf32>, tensor<2xi32>) -> tensor<11008x4096xf32>
%155 = tosa.reshape %152 {new_shape = array<i64: 40, 11008>} : (tensor<1x40x11008xf32>) -> tensor<40x11008xf32>
%cst_26 = arith.constant dense<0.000000e+00> : tensor<40x4096xf32>
%156 = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%155, %154 : tensor<40x11008xf32>, tensor<11008x4096xf32>) outs(%cst_26 : tensor<40x4096xf32>) -> tensor<40x4096xf32>
%157 = tosa.reshape %156 {new_shape = array<i64: 1, 40, 4096>} : (tensor<40x4096xf32>) -> tensor<1x40x4096xf32>
%158 = tosa.add %arg0, %157 : (tensor<1x40x4096xf32>, tensor<1x40x4096xf32>) -> tensor<1x40x4096xf32>

%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64

%tensor_unranked = tensor.cast %158 : tensor<1x40x4096xf32> to tensor<*xf32>

// CHECK: Unranked Memref base@ = {{.*}} rank = 3 offset = 0 sizes = [1, 40, 4096] strides = [163840, 4096, 1] data =
// CHECK-NEXT: [
// CHECK-SAME: [
// CHECK-SAME: [461655{{(, 461655)*}}],

call @printMemrefF32(%tensor_unranked) : (tensor<*xf32>) -> ()
vector.print %time : f64

return
}

func.func @main() {
%input_tensor = arith.constant dense<0.5> : tensor<1x40x4096xf32>
%weight1 = arith.constant dense<0.1> : tensor<4096xf32>
%weight2 = arith.constant dense<0.1> : tensor<11008x4096xf32>
%weight3 = arith.constant dense<0.1> : tensor<11008x4096xf32>
%weight4 = arith.constant dense<0.1> : tensor<4096x11008xf32>

// Print timings.
call @kernel_ffn(%input_tensor, %weight1, %weight2, %weight3, %weight4) : (tensor<1x40x4096xf32>, tensor<4096xf32>, tensor<11008x4096xf32>, tensor<11008x4096xf32>, tensor<4096x11008xf32>) -> ()

return
}

func.func private @printMemrefF32(%ptr : tensor<*xf32>)
75 changes: 75 additions & 0 deletions examples/BuddyNext/next-fpowi.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,75 @@
// 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: -arith-expand \
// RUN: -eliminate-empty-tensors \
// RUN: -empty-tensor-to-alloc-tensor \
// RUN: -one-shot-bufferize \
// RUN: -convert-linalg-to-affine-loops \
// RUN: -affine-loop-fusion \
// RUN: -lower-affine \
// RUN: -func-bufferize \
// RUN: -arith-bufferize \
// RUN: -tensor-bufferize \
// RUN: -buffer-deallocation \
// RUN: -finalizing-bufferize \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// 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-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-cpu-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 \
// RUN: | FileCheck %s

module {
func.func private @rtclock() -> f64

func.func @kernel_fpowi(%arg0: tensor<1x32x40x64xf32>) {
%t_start = call @rtclock() : () -> f64

// Power operation
%c2_i32 = arith.constant 2 : i32
%output_tensor = tensor.empty() : tensor<1x32x40x64xf32>
%result = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x32x40x64xf32>) outs(%output_tensor : tensor<1x32x40x64xf32>) {
^bb0(%in: f32, %out: f32):
%0 = math.fpowi %in, %c2_i32 : f32, i32
linalg.yield %0 : f32
} -> tensor<1x32x40x64xf32>

%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64

%tensor_unranked = tensor.cast %result : tensor<1x32x40x64xf32> to tensor<*xf32>

// CHECK: Unranked Memref base@ = {{.*}} rank = 4 offset = 0 sizes = [1, 32, 40, 64] strides = [81920, 2560, 64, 1] data =
// CHECK-NEXT: [
// CHECK-SAME: [
// CHECK-SAME: [25{{(, 25)*}}],

call @printMemrefF32(%tensor_unranked) : (tensor<*xf32>) -> ()
vector.print %time : f64

return
}

func.func @main() {
%input_tensor = arith.constant dense<5.0> : tensor<1x32x40x64xf32>

call @kernel_fpowi(%input_tensor) : (tensor<1x32x40x64xf32>) -> ()

return
}

func.func private @printMemrefF32(%ptr : tensor<*xf32>)
}
67 changes: 67 additions & 0 deletions examples/BuddyNext/next-matmul.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,67 @@
// 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: -arith-expand \
// RUN: -eliminate-empty-tensors \
// RUN: -empty-tensor-to-alloc-tensor \
// RUN: -one-shot-bufferize \
// RUN: -convert-linalg-to-affine-loops \
// RUN: -affine-loop-fusion \
// RUN: -lower-affine \
// RUN: -func-bufferize \
// RUN: -arith-bufferize \
// RUN: -tensor-bufferize \
// RUN: -buffer-deallocation \
// RUN: -finalizing-bufferize \
// RUN: -convert-vector-to-scf \
// RUN: -expand-strided-metadata \
// 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-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-cpu-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 \
// RUN: | FileCheck %s

func.func private @rtclock() -> f64

func.func @kernel_matmul(%arg0 : tensor<40x4096xf32>, %arg1 : tensor<4096x4096xf32>, %arg2 : tensor<40x4096xf32>) {
%t_start = call @rtclock() : () -> f64

%matmul_result = linalg.matmul {cast = #linalg.type_fn<cast_signed>} ins(%arg0, %arg1 : tensor<40x4096xf32>, tensor<4096x4096xf32>) outs(%arg2 : tensor<40x4096xf32>) -> tensor<40x4096xf32>

%t_end = call @rtclock() : () -> f64
%time = arith.subf %t_end, %t_start : f64

%tensor_unranked = tensor.cast %matmul_result : tensor<40x4096xf32> to tensor<*xf32>

// CHECK: Unranked Memref base@ = {{.*}} rank = 2 offset = 0 sizes = [40, 4096] strides = [4096, 1] data =
// CHECK-NEXT: [
// CHECK-SAME: [24576{{(, 24576)*}}]

call @printMemrefF32(%tensor_unranked) : (tensor<*xf32>) -> ()
vector.print %time : f64

return
}

func.func @main() {
%input_tensor_1 = arith.constant dense<3.0> : tensor<40x4096xf32>
%input_tensor_2 = arith.constant dense<2.0> : tensor<4096x4096xf32>
%output_tensor = arith.constant dense<0.0> : tensor<40x4096xf32>

call @kernel_matmul(%input_tensor_1, %input_tensor_2, %output_tensor) : (tensor<40x4096xf32>, tensor<4096x4096xf32>, tensor<40x4096xf32>) -> ()

return
}

func.func private @printMemrefF32(%ptr : tensor<*xf32>)
Loading