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

[tf-frontend] add new parten for layer_norm and l2_norm #317

Closed
wants to merge 5 commits into from
Closed

[tf-frontend] add new parten for layer_norm and l2_norm #317

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
97e3afe
add new parten for layer_norm and l2_norm
heromapwrd Jun 7, 2024
10be141
[tf-frontend] fix bug
heromapwrd Jun 7, 2024
ccb799f
[tf-frontend] fix bug
heromapwrd Jun 7, 2024
fdf29d0
[tf-frontend] fix bug
heromapwrd Jun 8, 2024
2d36d8a
[tf-frontend] fix bug
heromapwrd Jun 8, 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
2 changes: 2 additions & 0 deletions frontends/tf-frontend/tf_mlir_ext/numerical/numerical_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -64,6 +64,7 @@ def get_config(config: str):
"layer_norm_without_gamma": 6,
"layer_norm_without_beta": 6,
"layer_norm_multi_dim": 6,
"layer_norm_multi_dim_v2": 6,
"layer_norm_swap_add": 6,
"layer_norm_swap_mul": 6,
"layer_norm_swap_squarediff": 5,
Expand All @@ -80,6 +81,7 @@ def get_config(config: str):
"l2_norm_V2": 3,
"l2_norm_V2_swap_mul": 3,
"l2_norm_V3": 6,
"l2_norm_couple_with_batchmatmulv2": 5,
"onehot_case0": 6,
},
"black_list": [
Expand Down
77 changes: 58 additions & 19 deletions frontends/tf-frontend/tf_mlir_ext/numerical/rewrite_to_custom_call.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -291,20 +291,38 @@ func.func @layer_norm_multi_dim(%arg0: tensor<2x8x4xf32>) -> tensor<2x8x4xf32> {
%9 = "tf.AddV2"(%6, %8) {device = ""} : (tensor<2x8x4xf32>, tensor<2x8x4xf32>) -> tensor<2x8x4xf32>
func.return %9 : tensor<2x8x4xf32>
}
// CHECK-LABLE: func.func @layer_norm_multi_dim(%arg0: tensor<2x8x4xf32>) -> tensor<2x8x4xf32> {
// CHECK-LABLE: %cst = "tf.Const"() <{value = dense<1.000000e+00> : tensor<4xf32>}> : () -> tensor<4xf32>
// CHECK-LABLE: %cst_0 = "tf.Const"() <{value = dense<0.000000e+00> : tensor<4xf32>}> : () -> tensor<4xf32>
// CHECK-LABLE: %cst_1 = "tf.Const"() <{value = dense<[16, 4]> : tensor<2xi64>}> : () -> tensor<2xi64>
// CHECK-LABLE: %cst_2 = "tf.Const"() <{value = dense<[2, 8, 4]> : tensor<3xi64>}> : () -> tensor<3xi64>
// CHECK-LABLE: %cst_3 = "tf.Const"() <{value = dense<[[[1.000000e-01, 2.000000e-01, 3.000000e-01, 4.000000e-01]], [[5.000000e-01, 6.000000e-01, 0.699999988, 8.000000e-01]]]> : tensor<
// CHECK-LABLE: %cst_4 = "tf.Const"() <{value = dense<[[[0.00999999977, 2.000000e-02, 3.000000e-02, 4.000000e-02]], [[5.000000e-02, 6.000000e-02, 7.000000e-02, 8.000000e-02]]]> : tenso
// CHECK-LABLE: %0 = "tf.Reshape"(%arg0, %cst_1) : (tensor<2x8x4xf32>, tensor<2xi64>) -> tensor<16x4xf32>
// CHECK-LABLE: %1 = mhlo.custom_call @byteir.layer_norm(%0, %cst, %cst_0) {backend_config = "", byteir_attrs = {axis = [1], epsilon = 9.9999999747524271E-7 : f64}} : (tensor<16x4xf32>
// CHECK-LABLE: %2 = "tf.Reshape"(%1, %cst_2) : (tensor<16x4xf32>, tensor<3xi64>) -> tensor<2x8x4xf32>
// CHECK-LABLE: %3 = "tf.Mul"(%2, %cst_3) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-LABLE: %4 = "tf.Add"(%3, %cst_4) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-LABLE: return %4 : tensor<2x8x4xf32>
// CHECK-LABLE: }
// CHECK-LABEL: func.func @layer_norm_multi_dim(%arg0: tensor<2x8x4xf32>) -> tensor<2x8x4xf32> {
// CHECK: %0 = "tf.Reshape"(%arg0, %cst_1) : (tensor<2x8x4xf32>, tensor<2xi64>) -> tensor<16x4xf32>
// CHECK-NEXT: %1 = mhlo.custom_call @byteir.layer_norm(%0, %cst, %cst_0) {backend_config = "", byteir_attrs = {axis = [1], epsilon = 9.9999999747524271E-7 : f64}} : (tensor<16x4xf32>
// CHECK-NEXT: %2 = "tf.Reshape"(%1, %cst_2) : (tensor<16x4xf32>, tensor<3xi64>) -> tensor<2x8x4xf32>
// CHECK-NEXT: %3 = "tf.Mul"(%2, %cst_3) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-NEXT: %4 = "tf.Add"(%3, %cst_4) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-NEXT: return %4 : tensor<2x8x4xf32>

func.func @layer_norm_multi_dim_v2(%arg0: tensor<2x8x4xf32>) -> tensor<2x8x4xf32> {
%cst = "tf.Const"() <{value = dense<2> : tensor<1xi32>}> : () -> tensor<1xi32>
%cst_1 = "tf.Const"() <{value = dense<9.99999997E-7> : tensor<f32>}> : () -> tensor<f32>
%cst_2 = "tf.Const"() <{value = dense<[[[0.1, 0.2, 0.3, 0.4]], [[0.5, 0.6, 0.7, 0.8]]]> : tensor<2x1x4xf32>}> : () -> tensor<2x1x4xf32>
%cst_3 = "tf.Const"() <{value = dense<[[[0.01, 0.02, 0.03, 0.04]], [[0.05, 0.06, 0.07, 0.08]]]> : tensor<2x1x4xf32>}> : () -> tensor<2x1x4xf32>
%0 = "tf.Mean"(%arg0, %cst) <{keep_dims = true}> {device = ""} : (tensor<2x8x4xf32>, tensor<1xi32>) -> tensor<2x8x1xf32>
%1 = "tf.SquaredDifference"(%arg0, %0) {device = ""} : (tensor<2x8x4xf32>, tensor<2x8x1xf32>) -> tensor<2x8x4xf32>
%2 = "tf.Mean"(%1, %cst) <{keep_dims = true}> {device = ""} : (tensor<2x8x4xf32>, tensor<1xi32>) -> tensor<2x8x1xf32>
%3 = "tf.AddV2"(%2, %cst_1) {device = ""} : (tensor<2x8x1xf32>, tensor<f32>) -> tensor<2x8x1xf32>
%4 = "tf.Rsqrt"(%3) {device = ""} : (tensor<2x8x1xf32>) -> tensor<2x8x1xf32>
%5 = "tf.Mul"(%4, %cst_2) {device = ""} : (tensor<2x8x1xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
%6 = "tf.Mul"(%arg0, %5) {device = ""} : (tensor<2x8x4xf32>, tensor<2x8x4xf32>) -> tensor<2x8x4xf32>
%7 = "tf.Mul"(%5, %0) {device = ""} : (tensor<2x8x4xf32>, tensor<2x8x1xf32>) -> tensor<2x8x4xf32>
%8 = "tf.Sub"(%cst_3, %7) {device = ""} : (tensor<2x1x4xf32>, tensor<2x8x4xf32>) -> tensor<2x8x4xf32>
%9 = "tf.AddV2"(%6, %8) {device = ""} : (tensor<2x8x4xf32>, tensor<2x8x4xf32>) -> tensor<2x8x4xf32>
func.return %9 : tensor<2x8x4xf32>
}
// CHECK-LABEL: func.func @layer_norm_multi_dim_v2(%arg0: tensor<2x8x4xf32>) -> tensor<2x8x4xf32> {
// CHECK: %0 = "tf.Reshape"(%arg0, %cst_1) : (tensor<2x8x4xf32>, tensor<2xi64>) -> tensor<16x4xf32>
// CHECK-NEXT: %1 = mhlo.custom_call @byteir.layer_norm(%0, %cst, %cst_0) {backend_config = "", byteir_attrs = {axis = [1], epsilon = 9.9999999747524271E-7 : f64}} : (tensor<16x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<16x4xf32>
// CHECK-NEXT: %2 = "tf.Reshape"(%1, %cst_2) : (tensor<16x4xf32>, tensor<3xi64>) -> tensor<2x8x4xf32>
// CHECK-NEXT: %3 = "tf.Mul"(%2, %cst_3) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-NEXT: %4 = "tf.Add"(%3, %cst_4) : (tensor<2x8x4xf32>, tensor<2x1x4xf32>) -> tensor<2x8x4xf32>
// CHECK-NEXT: return %4 : tensor<2x8x4xf32>

func.func @layer_norm_swap_add(%arg0: tensor<2x32x3xf32>) -> tensor<2x32x3xf32> {
%cst_15 = "tf.Const"() {value = dense<9.99999997E-7> : tensor<f32>} : () -> tensor<f32>
Expand Down Expand Up @@ -541,7 +559,7 @@ func.func @l2_norm_V1(%arg0: tensor<1x32x3xf32>) -> tensor<1x32x3xf32> {
// CHECK-LABEL: func.func @l2_norm_V1(%arg0: tensor<1x32x3xf32>) -> tensor<1x32x3xf32> {
// CHECK: mhlo.custom_call
// CHECK-SAME: @byteir.l2_norm
// CHECK-SAME: byteir_attrs = {axis = [2], epsilon = 9.9999999747524271E-7 : f64}
// CHECK-SAME: byteir_attrs = {axis = [2], eps_outside_sqrt = false, epsilon = 9.9999999747524271E-7 : f64}

func.func @l2_norm_V1_with_multiplyer(%arg0: tensor<2x4x8xf32>) -> tensor<2x4x8xf32> {
%cst = "tf.Const"() <{value = dense<-1> : tensor<i32>}> : () -> tensor<i32>
Expand All @@ -557,7 +575,7 @@ func.func @l2_norm_V1_with_multiplyer(%arg0: tensor<2x4x8xf32>) -> tensor<2x4x8x
}
// CHECK-LABEL: func.func @l2_norm_V1_with_multiplyer(%arg0: tensor<2x4x8xf32>) -> tensor<2x4x8xf32> {
// CHECK-NEXT: %0 = mhlo.constant dense<9.34093475> : tensor<2x4x8xf32>
// CHECK-NEXT: %1 = mhlo.custom_call @byteir.l2_norm(%arg0) {backend_config = "", byteir_attrs = {axis = [2], epsilon = 9.999999960041972E-13 : f64}} : (tensor<2x4x8xf32>) -> tensor<2x4
// CHECK-NEXT: %1 = mhlo.custom_call @byteir.l2_norm(%arg0) {backend_config = "", byteir_attrs = {axis = [2], eps_outside_sqrt = false, epsilon = 9.999999960041972E-13 : f64}} : (tensor<2x4x8xf32>) -> tensor<2x4
// CHECK-NEXT: %2 = mhlo.multiply %1, %0 : tensor<2x4x8xf32>

func.func @l2_norm_V1_swap_mul(%54: tensor<1x64xf32>) -> tensor<1x64xf32> {
Expand Down Expand Up @@ -586,7 +604,7 @@ func.func @l2_norm_V2(%1871: tensor<1x128xf16>) -> tensor<1x128xf16> {
// CHECK-LABEL: @l2_norm_V2
// CHECK: mhlo.custom_call
// CHECK-SAME: @byteir.l2_norm
// CHECK-SAME: byteir_attrs = {axis = [1], epsilon = 0.000000e+00 : f64}
// CHECK-SAME: byteir_attrs = {axis = [1], eps_outside_sqrt = false, epsilon = 0.000000e+00 : f64}

func.func @l2_norm_V2_swap_mul(%1871: tensor<1x128xf16>) -> tensor<1x128xf16> {
%cst_5 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
Expand All @@ -600,7 +618,7 @@ func.func @l2_norm_V2_swap_mul(%1871: tensor<1x128xf16>) -> tensor<1x128xf16> {
// CHECK-LABEL: @l2_norm_V2_swap_mul
// CHECK: mhlo.custom_call
// CHECK-SAME: @byteir.l2_norm
// CHECK-SAME: byteir_attrs = {axis = [1], epsilon = 0.000000e+00 : f64}
// CHECK-SAME: byteir_attrs = {axis = [1], eps_outside_sqrt = false, epsilon = 0.000000e+00 : f64}

func.func @l2_norm_V3(%15: tensor<1x100x512xf32>) -> tensor<1x100x512xf32> {
%cst_96 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
Expand All @@ -613,7 +631,28 @@ func.func @l2_norm_V3(%15: tensor<1x100x512xf32>) -> tensor<1x100x512xf32> {
// CHECK-LABEL: @l2_norm_V3
// CHECK: mhlo.custom_call
// CHECK-SAME: @byteir.l2_norm
// CHECK-SAME: byteir_attrs = {axis = [2], epsilon = 0.000000e+00 : f64}
// CHECK-SAME: byteir_attrs = {axis = [2], eps_outside_sqrt = false, epsilon = 0.000000e+00 : f64}

func.func @l2_norm_couple_with_batchmatmulv2(%arg0: tensor<1x8x16xf32>, %arg1: tensor<1x32x16xf32>) -> tensor<1x8x32xf32> {
%cst = "tf.Const"() <{value = dense<2> : tensor<1xi32>}> : () -> tensor<1xi32>
%cst_1 = "tf.Const"() <{value = dense<1.013280e-06> : tensor<f32>}> : () -> tensor<f32>
%0 = "tf.Square"(%arg0) {device = ""} : (tensor<1x8x16xf32>) -> tensor<1x8x16xf32>
%1 = "tf.Sum"(%0, %cst) <{keep_dims = true}> {device = ""} : (tensor<1x8x16xf32>, tensor<1xi32>) -> tensor<1x8x1xf32>
%2 = "tf.Sqrt"(%1) {device = ""} : (tensor<1x8x1xf32>) -> tensor<1x8x1xf32>
%3 = "tf.BatchMatMulV2"(%arg0, %arg1) <{adj_x = false, adj_y = true, grad_x = false, grad_y = false}> {device = ""} : (tensor<1x8x16xf32>, tensor<1x32x16xf32>) -> tensor<1x8x32xf32>
%4 = "tf.Square"(%arg1) {device = ""} : (tensor<1x32x16xf32>) -> tensor<1x32x16xf32>
%5 = "tf.Sum"(%4, %cst) <{keep_dims = true}> {device = ""} : (tensor<1x32x16xf32>, tensor<1xi32>) -> tensor<1x32x1xf32>
%6 = "tf.Sqrt"(%5) {device = ""} : (tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
%7 = "tf.BatchMatMulV2"(%2, %6) <{adj_x = false, adj_y = true, grad_x = false, grad_y = false}> {device = ""} : (tensor<1x8x1xf32>, tensor<1x32x1xf32>) -> tensor<1x8x32xf32>
%8 = "tf.AddV2"(%7, %cst_1) {device = ""} : (tensor<1x8x32xf32>, tensor<f32>) -> tensor<1x8x32xf32>
%9 = "tf.RealDiv"(%3, %8) {device = ""} : (tensor<1x8x32xf32>, tensor<1x8x32xf32>) -> tensor<1x8x32xf32>
return %9 : tensor<1x8x32xf32>
}
// CHECK-LABEL: func.func @l2_norm_couple_with_batchmatmulv2(%arg0: tensor<1x8x16xf32>, %arg1: tensor<1x32x16xf32>) -> tensor<1x8x32xf32> {
// CHECK-LABEL: %0 = mhlo.custom_call @byteir.l2_norm(%arg0) {backend_config = "", byteir_attrs = {axis = [2], eps_outside_sqrt = true, epsilon = 1.0132799843631801E-6 : f64}} : (tensor<1x8x16xf32>) -> tensor<1x8x16xf32>
// CHECK-LABEL: %1 = mhlo.custom_call @byteir.l2_norm(%arg1) {backend_config = "", byteir_attrs = {axis = [2], eps_outside_sqrt = true, epsilon = 1.0132799843631801E-6 : f64}} : (tensor<1x32x16xf32>) -> tensor<1x32x16xf32>
// CHECK-LABEL: %2 = "tf.BatchMatMulV2"(%0, %1) <{adj_x = false, adj_y = true}> : (tensor<1x8x16xf32>, tensor<1x32x16xf32>) -> tensor<1x8x32xf32>
// CHECK-LABEL: return %2 : tensor<1x8x32xf32>

func.func @dynamic_mask_stitch(%arg0: tensor<4x4xf32>, %arg1: tensor<4xi32>) -> tensor<?x4xf32> {
%cst = "tf.Const"() {value = dense<[[-0.916170597, -0.884184718, 1.60242105, -1.19678485], [0.33643803, -0.431175768, 1.71861267, 0.126368985], [-1.07191086, -1.00517535, -0.666032254, 0.776807785], [1.53380013, 0.83925873, -0.24277249, 1.53341103]]> : tensor<4x4xf32>} : () -> tensor<4x4xf32>
Expand Down
Loading
Loading