Add remaining CheckGpuDelegateCompatibility() logic
Now the function has all the gpu delegate compatibility logic of model_builder.cc
PiperOrigin-RevId: 381770247
Change-Id: Ie560c934a402eb4fb190e40a0af120658a31fb30
diff --git a/tensorflow/lite/tools/versioning/gpu_compatibility.cc b/tensorflow/lite/tools/versioning/gpu_compatibility.cc
index badca8f..a67f68f 100644
--- a/tensorflow/lite/tools/versioning/gpu_compatibility.cc
+++ b/tensorflow/lite/tools/versioning/gpu_compatibility.cc
@@ -14,6 +14,8 @@
==============================================================================*/
#include "tensorflow/lite/tools/versioning/gpu_compatibility.h"
+#include <string>
+
#include "absl/status/status.h"
#include "absl/strings/str_cat.h"
#include "tensorflow/core/platform/logging.h"
@@ -25,6 +27,13 @@
namespace {
+const std::string GetOpName(const OpSignature& op_sig) {
+ if (op_sig.op == tflite::BuiltinOperator_CUSTOM) {
+ return op_sig.custom_name;
+ }
+ return tflite::EnumNamesBuiltinOperator()[op_sig.op];
+}
+
// Helper functions from
// tensorflow/lite/delegates/gpu/common/model_builder_helper.cc
@@ -44,6 +53,16 @@
return absl::OkStatus();
}
+template <typename ParamsT>
+absl::Status RetrieveCustomInitialData(const OpSignature& op_sig,
+ const ParamsT** tf_options) {
+ *tf_options = static_cast<const ParamsT*>(op_sig.custom_initial_data);
+ if (!*tf_options) {
+ return absl::InternalError("Unable to retrieve custom_initial_data.");
+ }
+ return absl::OkStatus();
+}
+
absl::Status IsActivationSupported(TfLiteFusedActivation fused_activation) {
switch (fused_activation) {
case kTfLiteActNone:
@@ -129,6 +148,28 @@
return absl::OkStatus();
}
+// Checks if the given OpSignature has required number of inputs and outputs for
+// convolution operators. The number of input should be either 2 runtime inputs
+// or 1 runtime and 1 constant input. The number of output should be one.
+absl::Status CheckConvoultionInputOutput(const OpSignature& op_sig) {
+ const int runtime_inputs = GetNumberOfRuntimeInputs(op_sig);
+ if (runtime_inputs > 2) {
+ return absl::InternalError(
+ absl::StrCat("Expected 1 or 2 input tensor(s), but node has ",
+ runtime_inputs, " runtime inputs."));
+ }
+ const int runtime_outputs = op_sig.outputs.size();
+ if (runtime_outputs != 1) {
+ return absl::InternalError(
+ absl::StrCat("Expected 1 output tensor(s), but node has ",
+ runtime_outputs, " runtime outputs."));
+ }
+ if (runtime_inputs == 1) {
+ RETURN_IF_ERROR(CheckTensorIsAvailable(op_sig, 1));
+ }
+ return absl::OkStatus();
+}
+
absl::Status CheckStrides(int strides_h, int strides_w) {
if (strides_h <= 0 || strides_w <= 0) {
return absl::InvalidArgumentError(
@@ -170,11 +211,26 @@
return absl::OkStatus();
}
+// Checks if the axes tensor at the given index is a integer32 constant tensor.
+absl::Status CheckAxesAreInt32Const(const OpSignature& op_sig, int idx) {
+ auto axes = op_sig.inputs.at(idx);
+ if (!axes.is_const) {
+ return absl::UnimplementedError(GetOpName(op_sig) +
+ " is only supported with constant axes.");
+ }
+ if (axes.type != kTfLiteInt32) {
+ return absl::UnimplementedError(absl::StrCat(
+ GetOpName(op_sig) + " supports int32 tensor for axes. But node has ",
+ TfLiteTypeGetName(axes.type)));
+ }
+ return absl::OkStatus();
+}
+
absl::Status CheckPooling2DGpuDelegateCompatibility(const OpSignature& op_sig) {
const TfLitePoolParams* tf_options;
if (op_sig.custom_initial_data) { // custom case with indices as a second
// output
- tf_options = static_cast<TfLitePoolParams*>(op_sig.custom_initial_data);
+ RETURN_IF_ERROR(RetrieveCustomInitialData(op_sig, &tf_options));
RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
/*required_runtime_inputs=*/1,
/*required_outputs=*/2));
@@ -190,6 +246,82 @@
return IsActivationSupported(tf_options->activation);
}
+absl::Status CheckDepthwiseConvGpuDelegateCompatibility(
+ const OpSignature& op_sig) {
+ RETURN_IF_ERROR(CheckConvoultionInputOutput(op_sig));
+ const TfLiteDepthwiseConvParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ RETURN_IF_ERROR(CheckStridesAndDilation(
+ tf_options->stride_height, tf_options->stride_width,
+ tf_options->dilation_height_factor, tf_options->dilation_width_factor));
+ RETURN_IF_ERROR(IsActivationSupported(tf_options->activation));
+
+ const int depth_multiplier = tf_options->depth_multiplier;
+ const auto* input = &op_sig.inputs[0];
+ const auto* filter = &op_sig.inputs[1];
+ const auto* bias = op_sig.inputs.size() > 2 ? &op_sig.inputs[2] : nullptr;
+ const auto* output = &op_sig.outputs[0];
+ if (input->dims.size() != 4) {
+ return absl::InvalidArgumentError("input.dims.size != 4");
+ }
+ if (filter->dims.size() != 4) {
+ return absl::InvalidArgumentError("filter.dims.size != 4");
+ }
+ if (output->dims.size() != 4) {
+ return absl::InvalidArgumentError("output.dims.size != 4");
+ }
+ if (input->dims[0] != output->dims[0]) {
+ return absl::InvalidArgumentError("input.b != output.b");
+ }
+ const int input_depth = input->dims[3];
+ const int output_depth = output->dims[3];
+ if (filter->dims[3] != output_depth) {
+ return absl::InvalidArgumentError("filter.i != output.c");
+ }
+ if (output_depth != input_depth * depth_multiplier) {
+ return absl::InvalidArgumentError("output.c != input.c * depth_multiplier");
+ }
+ if (bias && bias->dims.size() != output_depth) {
+ return absl::InvalidArgumentError("bias.size != output.c");
+ }
+ if (depth_multiplier != 1 && input_depth != 1) {
+ return absl::UnimplementedError("depth_multiplier != 1 && input.c != 1");
+ }
+ return absl::OkStatus();
+}
+
+absl::Status CheckCustomOpsGpuDelegateCompatibility(const OpSignature& op_sig) {
+ if (op_sig.custom_name == "Convolution2DTransposeBias") {
+ RETURN_IF_ERROR(CheckTensorIsAvailable(op_sig, 1));
+ const TfLiteTransposeConvParams* tf_options;
+ RETURN_IF_ERROR(RetrieveCustomInitialData(op_sig, &tf_options));
+ RETURN_IF_ERROR(
+ CheckStrides(tf_options->stride_height, tf_options->stride_width));
+ return absl::OkStatus();
+ }
+ if (op_sig.custom_name == "MaxPoolingWithArgmax2D") {
+ return CheckPooling2DGpuDelegateCompatibility(op_sig);
+ }
+ if (op_sig.custom_name == "MaxUnpooling2D") {
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/2,
+ /*required_outputs=*/1));
+ const TfLitePoolParams* tf_options;
+ RETURN_IF_ERROR(RetrieveCustomInitialData(op_sig, &tf_options));
+ RETURN_IF_ERROR(CheckKernelsAndStrides(
+ tf_options->filter_height, tf_options->filter_width,
+ tf_options->stride_height, tf_options->stride_width));
+ return absl::OkStatus();
+ }
+ if (op_sig.custom_name == "Resampler") {
+ return CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/2,
+ /*required_outputs=*/1);
+ }
+ return absl::InvalidArgumentError(
+ absl::StrCat("Not supported custom op ", op_sig.custom_name));
+}
+
} // namespace
// TODO(b/189917229): Logics are copied from TFLiteOperationParser:IsSupported()
@@ -224,26 +356,14 @@
TfLiteTypeGetName(op_sig.outputs.at(0).type)));
}
- case kTfLiteBuiltinConcatenation:
- // TODO(b/189917229): Implement logic.
+ case kTfLiteBuiltinConcatenation: {
+ const TfLiteConcatenationParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
return absl::OkStatus();
+ }
case kTfLiteBuiltinConv2d: {
- const int runtime_inputs = GetNumberOfRuntimeInputs(op_sig);
- if (runtime_inputs > 2) {
- return absl::InternalError(
- absl::StrCat("Expected 1 or 2 input tensor(s), but node has ",
- runtime_inputs, " runtime inputs."));
- }
- const int runtime_outputs = op_sig.outputs.size();
- if (runtime_outputs != 1) {
- return absl::InternalError(
- absl::StrCat("Expected 1 output tensor(s), but node has ",
- runtime_outputs, " runtime outputs."));
- }
- if (runtime_inputs == 1) {
- RETURN_IF_ERROR(CheckTensorIsAvailable(op_sig, 1));
- }
+ RETURN_IF_ERROR(CheckConvoultionInputOutput(op_sig));
const TfLiteConvParams* tf_options;
RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
RETURN_IF_ERROR(CheckStridesAndDilation(
@@ -257,65 +377,8 @@
return CheckInputsOutputs(op_sig, /*required_runtime_inputs=*/0,
/*required_outputs=*/1);
- case kTfLiteBuiltinDepthwiseConv2d: {
- const int runtime_inputs = GetNumberOfRuntimeInputs(op_sig);
- if (runtime_inputs > 2) {
- return absl::InternalError(
- absl::StrCat("Expected 1 or 2 input tensor(s), but node has ",
- runtime_inputs, " runtime inputs."));
- }
- const int runtime_outputs = op_sig.outputs.size();
- if (runtime_outputs != 1) {
- return absl::InternalError(
- absl::StrCat("Expected 1 output tensor(s), but node has ",
- runtime_outputs, " runtime outputs."));
- }
- if (runtime_inputs == 1) {
- RETURN_IF_ERROR(CheckTensorIsAvailable(op_sig, 1));
- }
- const TfLiteDepthwiseConvParams* tf_options;
- RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
- RETURN_IF_ERROR(CheckStridesAndDilation(
- tf_options->stride_height, tf_options->stride_width,
- tf_options->dilation_height_factor,
- tf_options->dilation_width_factor));
- RETURN_IF_ERROR(IsActivationSupported(tf_options->activation));
-
- const int depth_multiplier = tf_options->depth_multiplier;
- const auto* input = &op_sig.inputs[0];
- const auto* filter = &op_sig.inputs[1];
- const auto* bias = op_sig.inputs.size() > 2 ? &op_sig.inputs[2] : nullptr;
- const auto* output = &op_sig.outputs[0];
- if (input->dims.size() != 4) {
- return absl::InvalidArgumentError("input.dims.size != 4");
- }
- if (filter->dims.size() != 4) {
- return absl::InvalidArgumentError("filter.dims.size != 4");
- }
- if (output->dims.size() != 4) {
- return absl::InvalidArgumentError("output.dims.size != 4");
- }
- if (input->dims[0] != output->dims[0]) {
- return absl::InvalidArgumentError("input.b != output.b");
- }
- const int input_depth = input->dims[3];
- const int output_depth = output->dims[3];
- if (filter->dims[3] != output_depth) {
- return absl::InvalidArgumentError("filter.i != output.c");
- }
- if (output_depth != input_depth * depth_multiplier) {
- return absl::InvalidArgumentError(
- "output.c != input.c * depth_multiplier");
- }
- if (bias && bias->dims.size() != output_depth) {
- return absl::InvalidArgumentError("bias.size != output.c");
- }
- if (depth_multiplier != 1 && input_depth != 1) {
- return absl::UnimplementedError(
- "depth_multiplier != 1 && input.c != 1");
- }
- return absl::OkStatus();
- }
+ case kTfLiteBuiltinDepthwiseConv2d:
+ return CheckDepthwiseConvGpuDelegateCompatibility(op_sig);
case kTfLiteBuiltinDepthToSpace: {
RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
@@ -426,9 +489,13 @@
case kTfLiteBuiltinMaxPool2d:
return CheckPooling2DGpuDelegateCompatibility(op_sig);
- case kTfLiteBuiltinMean:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
+ case kTfLiteBuiltinMean: {
+ RETURN_IF_ERROR(CheckInputsConstsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_const_inputs=*/1,
+ /*required_outputs=*/1));
+ return CheckAxesAreInt32Const(op_sig, 1);
+ }
case kTfLiteBuiltinMul: {
if (op_sig.inputs.size() != 2) {
@@ -466,92 +533,155 @@
case kTfLiteBuiltinPack:
return absl::OkStatus();
- case kTfLiteBuiltinReduceMax:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinReduceMin:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinReduceProd:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
case kTfLiteBuiltinQuantize:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinRelu:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinRelu6:
- // TODO(b/189917229): Implement logic.
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
return absl::OkStatus();
case kTfLiteBuiltinReluN1To1:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinLeakyRelu:
- // TODO(b/189917229): Implement logic.
return absl::OkStatus();
case kTfLiteBuiltinPrelu:
return absl::OkStatus();
case kTfLiteBuiltinReshape:
- // TODO(b/189917229): Implement logic.
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
return absl::OkStatus();
- case kTfLiteBuiltinResizeBilinear:
- // TODO(b/189917229): Implement logic.
+ case kTfLiteBuiltinSlice: {
+ if (op_sig.inputs.size() < 3) {
+ return absl::UnimplementedError(
+ absl::StrCat("SLICE requires 3 inputs, but node has ",
+ op_sig.inputs.size(), " inputs."));
+ }
+ const auto& input = op_sig.inputs.at(0);
+ if (input.dims.size() != 3 && input.dims.size() != 4) {
+ return absl::UnimplementedError(absl::StrCat(
+ "SLICE supports for 3 or 4 dimensional tensors only, but node has ",
+ input.dims.size(), " dimensional tensors."));
+ }
return absl::OkStatus();
+ }
- case kTfLiteBuiltinResizeNearestNeighbor:
- // TODO(b/189917229): Implement logic.
+ case kTfLiteBuiltinSoftmax: {
+ const TfLiteSoftmaxParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ if (tf_options->beta != 1) {
+ return absl::UnimplementedError("Softmax.beta != 1 is not supported.");
+ }
return absl::OkStatus();
+ }
- case kTfLiteBuiltinSlice:
- // TODO(b/189917229): Implement logic.
+ case kTfLiteBuiltinSpaceToDepth: {
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
+ const TfLiteSpaceToDepthParams* s2d_params;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &s2d_params));
+ if (s2d_params->block_size == 1) {
+ return absl::InvalidArgumentError(
+ "SPACE_TO_DEPTH block_size = 1 is a no-op.");
+ }
+ if (s2d_params->block_size < 1) {
+ return absl::InvalidArgumentError(
+ "SPACE_TO_DEPTH block_size must be > 1.");
+ }
return absl::OkStatus();
-
- case kTfLiteBuiltinSoftmax:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
-
- case kTfLiteBuiltinSpaceToDepth:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
+ }
case kTfLiteBuiltinSplit:
- // TODO(b/189917229): Implement logic.
return absl::OkStatus();
case kTfLiteBuiltinSplitV:
- // TODO(b/189917229): Implement logic.
return absl::OkStatus();
- case kTfLiteBuiltinStridedSlice:
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
+ case kTfLiteBuiltinStridedSlice: {
+ const TfLiteStridedSliceParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ if (tf_options->ellipsis_mask) {
+ return absl::UnimplementedError(
+ "Slice does not support ellipsis_mask.");
+ }
+ if (tf_options->new_axis_mask) {
+ return absl::UnimplementedError(
+ "Slice does not support new_axis_mask.");
+ }
+ if (tf_options->shrink_axis_mask) {
+ return absl::UnimplementedError(
+ "Slice does not support shrink_axis_mask parameter. ");
+ }
- case kTfLiteBuiltinSum:
- // TODO(b/189917229): Implement logic.
+ if (op_sig.inputs.size() < 4) {
+ return absl::UnimplementedError("STRIDED_SLICE requires 4 inputs.");
+ }
+ const auto& input = op_sig.inputs.at(0);
+ if (input.dims.size() != 3 && input.dims.size() != 4) {
+ return absl::UnimplementedError(
+ "STRIDED_SLICE supports for 3 or 4 dimensional tensors only.");
+ }
return absl::OkStatus();
+ }
case kTfLiteBuiltinTile:
- // TODO(b/189917229): Implement logic.
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
return absl::OkStatus();
case kTfLiteBuiltinTranspose:
- // TODO(b/189917229): Implement logic.
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
return absl::OkStatus();
- case kTfLiteBuiltinTransposeConv:
- // TODO(b/189917229): Implement logic.
+ case kTfLiteBuiltinTransposeConv: {
+ RETURN_IF_ERROR(CheckConvoultionInputOutput(op_sig));
+ const TfLiteTransposeConvParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ RETURN_IF_ERROR(
+ CheckStrides(tf_options->stride_height, tf_options->stride_width));
return absl::OkStatus();
+ }
+
+ case kTfLiteBuiltinResizeBilinear: {
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
+ const TfLiteResizeBilinearParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ if (tf_options->align_corners && tf_options->half_pixel_centers) {
+ return absl::InternalError(
+ "If half_pixel_centers is True, align_corners must be False.");
+ }
+ return absl::OkStatus();
+ }
+
+ case kTfLiteBuiltinResizeNearestNeighbor: {
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
+ const TfLiteResizeNearestNeighborParams* tf_options;
+ RETURN_IF_ERROR(RetrieveBuiltinData(op_sig, &tf_options));
+ return absl::OkStatus();
+ }
+
+ case kTfLiteBuiltinRelu:
+ case kTfLiteBuiltinRelu6:
+ case kTfLiteBuiltinLeakyRelu:
+ return absl::OkStatus();
+
+ case kTfLiteBuiltinReduceMax:
+ case kTfLiteBuiltinReduceMin:
+ case kTfLiteBuiltinReduceProd:
+ case kTfLiteBuiltinSum: {
+ RETURN_IF_ERROR(CheckInputsOutputs(op_sig,
+ /*required_runtime_inputs=*/1,
+ /*required_outputs=*/1));
+ return CheckAxesAreInt32Const(op_sig, 1);
+ }
case kTfLiteBuiltinPad:
case kTfLiteBuiltinMirrorPad: {
@@ -642,25 +772,8 @@
return IsActivationSupported(activation);
}
- case kTfLiteBuiltinCustom: {
- if (op_sig.custom_name == "Convolution2DTransposeBias") {
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
- }
- if (op_sig.custom_name == "MaxPoolingWithArgmax2D") {
- return CheckPooling2DGpuDelegateCompatibility(op_sig);
- }
- if (op_sig.custom_name == "MaxUnpooling2D") {
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
- }
- if (op_sig.custom_name == "Resampler") {
- // TODO(b/189917229): Implement logic.
- return absl::OkStatus();
- }
- return absl::InvalidArgumentError(
- absl::StrCat("Not supported custom op ", op_sig.custom_name));
- }
+ case kTfLiteBuiltinCustom:
+ return CheckCustomOpsGpuDelegateCompatibility(op_sig);
default:
break;
diff --git a/tensorflow/lite/tools/versioning/op_signature.cc b/tensorflow/lite/tools/versioning/op_signature.cc
index bd96ff7..59080d9 100644
--- a/tensorflow/lite/tools/versioning/op_signature.cc
+++ b/tensorflow/lite/tools/versioning/op_signature.cc
@@ -210,6 +210,7 @@
OpSignature op_sig = {
static_cast<BuiltinOperator>(registration->builtin_code)};
op_sig.builtin_data = tflite_node->builtin_data;
+ op_sig.custom_initial_data = tflite_node->custom_initial_data;
std::memset(&op_sig.ext_options, 0, sizeof(op_sig.ext_options));
op_sig.inputs =
diff --git a/tensorflow/lite/tools/versioning/op_signature.h b/tensorflow/lite/tools/versioning/op_signature.h
index ad4f804..f9ed28d 100644
--- a/tensorflow/lite/tools/versioning/op_signature.h
+++ b/tensorflow/lite/tools/versioning/op_signature.h
@@ -35,7 +35,7 @@
std::vector<OpSignatureTensorSpec> inputs;
std::vector<OpSignatureTensorSpec> outputs;
void* builtin_data;
- void* custom_initial_data;
+ const void* custom_initial_data;
std::string custom_name;
union {
struct {
