Skip to content

Commit

Permalink
Merge branch 'master' into grad_check_v1
Browse files Browse the repository at this point in the history
  • Loading branch information
@amturati
amturati committed Sep 11, 2020
2 parents 9964219 + 24d099e commit 3b2480f
Show file tree
Hide file tree
Showing 305 changed files with 10,232 additions and 7,644 deletions.
8 changes: 5 additions & 3 deletions .bazelrc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -162,21 +162,18 @@ build --host_java_toolchain=//third_party/toolchains/java:tf_java_toolchain
# environment variable "TF_MKL_ROOT" every time before build.
build:mkl --define=build_with_mkl=true --define=enable_mkl=true
build:mkl --define=tensorflow_mkldnn_contraction_kernel=0
build:mkl --define=build_with_mkl_dnn_v1_only=true
build:mkl -c opt

# config to build OneDNN backend with a user specified threadpool.
build:mkl_threadpool --define=build_with_mkl=true --define=enable_mkl=true
build:mkl_threadpool --define=tensorflow_mkldnn_contraction_kernel=0
build:mkl_threadpool --define=build_with_mkl_dnn_v1_only=true
build:mkl_threadpool --define=build_with_mkl_opensource=true
build:mkl_threadpool --define=build_with_mkldnn_threadpool=true
build:mkl_threadpool -c opt

# Config setting to build with oneDNN and without the binary blob
build:mkl_opensource_only --define=build_with_mkl=true --define=enable_mkl=true
build:mkl_opensource_only --define=tensorflow_mkldnn_contraction_kernel=0
build:mkl_opensource_only --define=build_with_mkl_dnn_v1_only=true
build:mkl_opensource_only --define=build_with_mkl_opensource=true
build:mkl_opensource_only -c opt

Expand Down Expand Up @@ -336,6 +333,11 @@ build:windows --host_copt=-DWIN32_LEAN_AND_MEAN
build:windows --copt=-DNOGDI
build:windows --host_copt=-DNOGDI

# MSVC (Windows): Standards-conformant preprocessor mode
# See https://docs.microsoft.com/en-us/cpp/preprocessor/preprocessor-experimental-overview
build:windows --copt=/experimental:preprocessor
build:windows --host_copt=/experimental:preprocessor

# Misc build options we need for windows.
build:windows --linkopt=/DEBUG
build:windows --host_linkopt=/DEBUG
Expand Down
30 changes: 5 additions & 25 deletions tensorflow/c/eager/BUILD
View file
Original file line number Diff line number Diff line change
Expand Up @@ -103,9 +103,13 @@ filegroup(
"c_api_unified_experimental.h",
"c_api_unified_experimental_internal.h",
"dlpack.h",
"gradients.h",
"gradients_internal.h",
"immediate_execution_context.h",
"immediate_execution_operation.h",
"immediate_execution_tensor_handle.h",
"mnist_gradients_testutil.h",
"tape.h",
"tfe_cancellation_manager_internal.h",
"tfe_executor_internal.h",
"tfe_monitoring_internal.h",
Expand Down Expand Up @@ -164,31 +168,6 @@ cc_library(
],
)

cc_library(
name = "gradients",
srcs = [
"gradients.cc",
"gradients_internal.h",
],
hdrs = [
"gradients.h",
],
visibility = [
"//tensorflow:internal",
],
deps = [
":abstract_context",
":abstract_operation",
":abstract_tensor_handle",
":c_api_unified_internal",
":tape",
"//tensorflow/core/common_runtime/eager:attr_builder",
"//tensorflow/core/lib/llvm_rtti",
"@com_google_absl//absl/container:flat_hash_map",
"@com_google_absl//absl/strings",
],
)

cc_library(
name = "gradients_internal",
srcs = [
Expand Down Expand Up @@ -757,6 +736,7 @@ tf_cuda_cc_test(
# TODO(b/136478427): Figure out how to correctly shut the server down
args = ["--heap_check=local"],
tags = [
"no_oss", # b/168323632
"no_windows",
"noasan", # leaks gRPC server instances
],
Expand Down
10 changes: 7 additions & 3 deletions tensorflow/c/eager/mnist_gradients_testutil.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#ifndef TENSORFLOW_C_EAGER_MNIST_GRADIENTS_TESTUTIL_H_
#define TENSORFLOW_C_EAGER_MNIST_GRADIENTS_TESTUTIL_H_
#include <memory>

#include "absl/types/span.h"
Expand Down Expand Up @@ -136,6 +138,8 @@ Status SoftmaxModel(AbstractContext* ctx,
absl::Span<AbstractTensorHandle*> outputs,
const GradientRegistry& registry);

} // internal
} // gradients
} // tensorflow
} // namespace internal
} // namespace gradients
} // namespace tensorflow

#endif // TENSORFLOW_C_EAGER_MNIST_GRADIENTS_TESTUTIL_H_
36 changes: 33 additions & 3 deletions tensorflow/c/experimental/gradients/BUILD
View file
Original file line number Diff line number Diff line change
Expand Up @@ -16,7 +16,7 @@ cc_library(
"//tensorflow/c/eager:abstract_operation",
"//tensorflow/c/eager:abstract_tensor_handle",
"//tensorflow/c/eager:c_api_unified_internal",
"//tensorflow/c/eager:gradients",
"//tensorflow/c/eager:gradients_internal",
"//tensorflow/core/lib/llvm_rtti",
],
)
Expand All @@ -34,7 +34,7 @@ cc_library(
"//tensorflow/c/eager:abstract_operation",
"//tensorflow/c/eager:abstract_tensor_handle",
"//tensorflow/c/eager:c_api_unified_internal",
"//tensorflow/c/eager:gradients",
"//tensorflow/c/eager:gradients_internal",
"//tensorflow/c/experimental/ops:array_ops",
"//tensorflow/c/experimental/ops:math_ops",
"//tensorflow/c/experimental/ops:nn_ops",
Expand All @@ -55,10 +55,40 @@ cc_library(
"//tensorflow/c/eager:abstract_operation",
"//tensorflow/c/eager:abstract_tensor_handle",
"//tensorflow/c/eager:c_api_unified_internal",
"//tensorflow/c/eager:gradients",
"//tensorflow/c/eager:gradients_internal",
"//tensorflow/c/experimental/ops:array_ops",
"//tensorflow/c/experimental/ops:math_ops",
"//tensorflow/c/experimental/ops:nn_ops",
"//tensorflow/core/lib/llvm_rtti",
],
)

cc_library(
name = "gradients",
hdrs = [
"array_grad.h",
"math_grad.h",
"nn_grad.h",
],
visibility = [
"//tensorflow:internal",
],
deps = [
":array_grad",
":math_grad",
":nn_grad",
],
)

filegroup(
name = "pywrap_required_hdrs",
srcs = [
"array_grad.h",
"math_grad.h",
"nn_grad.h",
],
visibility = [
"//tensorflow/core:__pkg__",
"//tensorflow/python:__pkg__",
],
)
5 changes: 5 additions & 0 deletions tensorflow/c/experimental/ops/BUILD
View file
Original file line number Diff line number Diff line change
Expand Up @@ -68,6 +68,11 @@ cc_library(

cc_library(
name = "ops",
hdrs = [
"array_ops.h",
"math_ops.h",
"nn_ops.h",
],
visibility = [
"//tensorflow:internal",
],
Expand Down
30 changes: 28 additions & 2 deletions tensorflow/compiler/mlir/hlo/lib/Dialect/mhlo/IR/hlo_ops.cc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1005,13 +1005,38 @@ LogicalResult ConcatenateOp::inferReturnTypes(
}
}

// If an input is unranked the output shape is unranked.
// Find the first ranked input to determine the output rank.
for (auto type : operands.getTypes()) {
auto shaped_type = type.cast<ShapedType>();
if (shaped_type.hasRank()) {
first_type = shaped_type;
break;
}
}

// If all inputs are unranked, the result must be unranked.
if (!first_type.hasRank()) {
inferredReturnTypes.push_back(UnrankedTensorType::get(out_element));
return success();
}

auto out_shape = llvm::to_vector<6>(first_type.getShape());

// Determine what the non-concatenate dimensions should be.
for (auto type : operands.getTypes()) {
auto shaped_ty = type.cast<ShapedType>();
if (!shaped_ty.hasRank()) {
continue;
}

for (auto it : llvm::enumerate(shaped_ty.getShape())) {
// If a dimension is not dynamic, the output shape should match.
if (ShapedType::isDynamic(out_shape[it.index()])) {
out_shape[it.index()] = it.value();
}
}
}

out_shape[dimension] = 0;

for (auto operand : operands.getTypes()) {
Expand Down Expand Up @@ -1362,7 +1387,8 @@ OpFoldResult OrOp::fold(ArrayRef<Attribute> operands) {
OpFoldResult XorOp::fold(ArrayRef<Attribute> operands) {
auto rType = getType().cast<ShapedType>();
if (lhs() == rhs()) {
return DenseIntElementsAttr::get(rType, 0);
Builder builder(getContext());
return builder.getZeroAttr(rType);
}

auto lhsVal = operands[0].dyn_cast_or_null<DenseElementsAttr>();
Expand Down
9 changes: 3 additions & 6 deletions tensorflow/compiler/mlir/lite/ir/tfl_ops.td
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2447,8 +2447,7 @@ def TFL_ReluOp: TFL_Op<"relu", [
PredOpTrait<"x and y must have same element type",
TFL_TCresVTEtIsSameAsOp<0, 0>>,
NoSideEffect,
SameOperandsAndResultShape,
SameOperandsAndResultsScale]> {
SameOperandsAndResultShape]> {
let summary = "Relu operator";

let description = [{
Expand Down Expand Up @@ -2476,8 +2475,7 @@ def TFL_Relu6Op: TFL_Op<"relu6", [
PredOpTrait<"x and y must have same element type",
TFL_TCresVTEtIsSameAsOp<0, 0>>,
NoSideEffect,
SameOperandsAndResultShape,
SameOperandsAndResultsScale]> {
SameOperandsAndResultShape]> {
let summary = "Relu6 operator";

let description = [{
Expand Down Expand Up @@ -2505,8 +2503,7 @@ def TFL_Relu1Op: TFL_Op<"relu_n1_to_1", [
PredOpTrait<"x and y must have same element type",
TFL_TCresVTEtIsSameAsOp<0, 0>>,
NoSideEffect,
SameOperandsAndResultShape,
SameOperandsAndResultsScale]> {
SameOperandsAndResultShape]> {
let summary = "Relu1 operator";

let description = [{
Expand Down
9 changes: 9 additions & 0 deletions tensorflow/compiler/mlir/lite/tests/ops.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1700,6 +1700,15 @@ func @testRelu6WithQuantizedTypes(%arg0 : tensor<10x!quant.uniform<u8:f32, 1.0>>

// -----

func @testReluWithDifferentScales(%arg0 : tensor<10x!quant.uniform<u8:f32, 1.0>>) -> tensor<10x!quant.uniform<u8:f32, 4.0>> {
%0 = "tfl.relu"(%arg0) : (tensor<10x!quant.uniform<u8:f32, 1.0>>) -> tensor<10x!quant.uniform<u8:f32, 2.0>>
%1 = "tfl.relu_n1_to_1"(%0) : (tensor<10x!quant.uniform<u8:f32, 2.0>>) -> tensor<10x!quant.uniform<u8:f32, 3.0>>
%2 = "tfl.relu6"(%1) : (tensor<10x!quant.uniform<u8:f32, 3.0>>) -> tensor<10x!quant.uniform<u8:f32, 4.0>>
return %2 : tensor<10x!quant.uniform<u8:f32, 4.0>>
}

// -----

func @testEmbeddingLookup(%arg0 : tensor<?xi32>, %arg1 : tensor<?x?xf32>) -> tensor<?xf32> {
%0 = "tfl.embedding_lookup"(%arg0, %arg1) : (tensor<?xi32>,tensor<?x?xf32>) -> tensor<?xf32>
return %0 : tensor<?xf32>
Expand Down
90 changes: 90 additions & 0 deletions tensorflow/compiler/mlir/lite/tests/optimize.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,96 @@ func @fuseSubIntoConv2d(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf
// CHECK: %0 = "tfl.conv_2d"(%arg0, %arg1, %cst)
}

// CHECK-LABEL: fuseAddIntoTransposeConv
func @fuseAddIntoTransposeConv(%arg0: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
%cst = constant dense<1.5> : tensor<32xf32>
%cst_0 = constant dense<[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0]> : tensor<16xf32>
%cst_1 = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
%cst_2 = constant dense<1.0> : tensor<32x4x4x128xf32>
%cst_3 = constant dense<[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0]> : tensor<32xf32>
%0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
%1 = "tfl.add"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
return %1 : tensor<1x64x84x32xf32>

// CHECK: %[[SHAPE:.*]] = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
// CHECK: %[[WEIGHTS:.*]] = constant dense<1.000000e+00> : tensor<32x4x4x128xf32>
// CHECK: %[[BIAS:.*]] = constant dense<[2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00, 2.500000e+00, 3.500000e+00]> : tensor<32xf32>
// CHECK: %[[RESULT:.*]] = "tfl.transpose_conv"(%[[SHAPE]], %[[WEIGHTS]], %arg0, %[[BIAS]])
// CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseSubIntoTransposeConv
func @fuseSubIntoTransposeConv(%arg0: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
%cst = constant dense<1.5> : tensor<32xf32>
%cst_0 = constant dense<[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0]> : tensor<16xf32>
%cst_1 = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
%cst_2 = constant dense<1.0> : tensor<32x4x4x128xf32>
%cst_3 = constant dense<[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0]> : tensor<32xf32>
%0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
%1 = "tfl.sub"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
return %1 : tensor<1x64x84x32xf32>

// CHECK: %[[SHAPE:.*]] = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
// CHECK: %[[WEIGHTS:.*]] = constant dense<1.000000e+00> : tensor<32x4x4x128xf32>
// CHECK: %[[BIAS:.*]] = constant dense<[-5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01, -5.000000e-01, 5.000000e-01]> : tensor<32xf32>
// CHECK: %[[RESULT:.*]] = "tfl.transpose_conv"(%[[SHAPE]], %[[WEIGHTS]], %arg0, %[[BIAS]])
// CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseAddIntoTransposeConvNoBias
func @fuseAddIntoTransposeConvNoBias(%arg0: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
%cst = constant dense<1.5> : tensor<32xf32>
%cst_0 = constant dense<[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0]> : tensor<16xf32>
%cst_1 = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
%cst_2 = constant dense<1.0> : tensor<32x4x4x128xf32>
%cst_3 = constant unit
%0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x32xf32>
%1 = "tfl.add"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
return %1 : tensor<1x64x84x32xf32>

// CHECK: %[[SHAPE:.*]] = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
// CHECK: %[[WEIGHTS:.*]] = constant dense<1.000000e+00> : tensor<32x4x4x128xf32>
// CHECK: %[[BIAS:.*]] = constant dense<1.500000e+00> : tensor<32xf32>
// CHECK: %[[RESULT:.*]] = "tfl.transpose_conv"(%[[SHAPE]], %[[WEIGHTS]], %arg0, %[[BIAS]])
// CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseMulIntoTransposeConv
func @fuseMulIntoTransposeConv(%arg0: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
%cst = constant dense<1.5> : tensor<32xf32>
%cst_0 = constant dense<[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0]> : tensor<16xf32>
%cst_1 = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
%cst_2 = constant dense<1.0> : tensor<32x4x4x128xf32>
%cst_3 = constant dense<[1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 2.0]> : tensor<32xf32>
%0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
%1 = "tfl.mul"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
return %1 : tensor<1x64x84x32xf32>

// CHECK: %[[SHAPE:.*]] = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
// CHECK: %[[WEIGHTS:.*]] = constant dense<1.500000e+00> : tensor<32x4x4x128xf32>
// CHECK: %[[BIAS:.*]] = constant dense<[1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00, 1.500000e+00, 3.000000e+00]> : tensor<32xf32>
// CHECK: %[[RESULT:.*]] = "tfl.transpose_conv"(%[[SHAPE]], %[[WEIGHTS]], %arg0, %[[BIAS]])
// CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseMulIntoTransposeConvNoBias
func @fuseMulIntoTransposeConvNoBias(%arg0: tensor<1x32x42x128xf32>) -> tensor<1x64x84x32xf32> {
%cst = constant dense<1.5> : tensor<32xf32>
%cst_0 = constant dense<[1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0]> : tensor<16xf32>
%cst_1 = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
%cst_2 = constant dense<1.0> : tensor<32x4x4x128xf32>
%cst_3 = constant unit
%0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, none) -> tensor<1x64x84x32xf32>
%1 = "tfl.mul"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
return %1 : tensor<1x64x84x32xf32>

// CHECK: %[[SHAPE:.*]] = constant dense<[1, 64, 84, 32]> : tensor<4xi32>
// CHECK: %[[WEIGHTS:.*]] = constant dense<1.500000e+00> : tensor<32x4x4x128xf32>
// CHECK: %[[BIAS:.*]] = constant unit
// CHECK: %[[RESULT:.*]] = "tfl.transpose_conv"(%[[SHAPE]], %[[WEIGHTS]], %arg0, %[[BIAS]])
// CHECK: return %[[RESULT]]
}

// CHECK-LABEL: fuseAddIntoFollowingConv2d
func @fuseAddIntoFollowingConv2d(%arg0: tensor<256x32x32x3xf32>) -> tensor<256x30x30x16xf32> {
%cst = constant dense<1.5> : tensor<f32>
Expand Down

0 comments on commit 3b2480f

Please sign in to comment.