Added versioning to ADD/SUB + some rework of the existing code.
diff --git a/tensorflow/lite/kernels/add.cc b/tensorflow/lite/kernels/add.cc
index 731c2fb..eb53b7c 100644
--- a/tensorflow/lite/kernels/add.cc
+++ b/tensorflow/lite/kernels/add.cc
@@ -100,19 +100,26 @@
// 8bit -> 8bit general quantized path, with general rescalings
// as well as, 16bit -> 16bit with general rescalings
- bool pot_scale_16bit = false;
+ bool pot_scale_16bit = true;
bool input1_scale_is_pot = false;
bool input2_scale_is_pot = false;
bool output_scale_is_pot = false;
- int input1_scale_log2_rounded;
- int input2_scale_log2_rounded;
- int output_scale_log2_rounded;
+ int input1_scale_log2_rounded{0};
+ int input2_scale_log2_rounded{0};
+ int output_scale_log2_rounded{0};
if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
output->type == kTfLiteInt16) {
- // Check that param scale is POT
+ // In case of 16-bit, there are two implementation:
+ // the scale parameter is a general number
+ // the scale parameter is POT and
+ // zero_point is zero for inputs/output.
+ pot_scale_16bit = (input1->params.zero_point == 0) &&
+ (input2->params.zero_point == 0) &&
+ (output->params.zero_point == 0);
+
input1_scale_is_pot =
CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);
@@ -122,14 +129,14 @@
output_scale_is_pot =
CheckedLog2(output->params.scale, &output_scale_log2_rounded);
- pot_scale_16bit = input1_scale_log2_rounded && input2_scale_log2_rounded &&
- output_scale_log2_rounded;
+ pot_scale_16bit &=
+ input1_scale_is_pot && input2_scale_is_pot && output_scale_is_pot;
}
data->pot_scale_16bit = pot_scale_16bit;
if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8 ||
- pot_scale_16bit) {
+ !pot_scale_16bit) {
// 8bit -> 8bit general quantized path, with general rescalings
// as well as, 16bit -> 16bit with general rescalings
data->input1_offset = -input1->params.zero_point;
@@ -139,7 +146,7 @@
// The shift is set to 15 for 16-bit and 20 in case of 8-bit, accordingly.
// In case of 16-bit we have 65535 << 15 which is less than 1 << 31,
// therefore the addition will still fit in a 32 bit accumulator.
- data->left_shift = pot_scale_16bit ? 15 : 20;
+ data->left_shift = !pot_scale_16bit ? 15 : 20;
const double twice_max_input_scale =
2 * std::max(input1->params.scale, input2->params.scale);
const double real_input1_multiplier =
@@ -252,7 +259,7 @@
const TfLiteTensor* input2,
TfLiteTensor* output) {
if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8 ||
- data->pot_scale_16bit) {
+ !data->pot_scale_16bit) {
tflite::ArithmeticParams op_params;
op_params.left_shift = data->left_shift;
op_params.input1_offset = data->input1_offset;
diff --git a/tensorflow/lite/kernels/register.cc b/tensorflow/lite/kernels/register.cc
index 626c092..7b7b60a 100644
--- a/tensorflow/lite/kernels/register.cc
+++ b/tensorflow/lite/kernels/register.cc
@@ -88,7 +88,7 @@
/* max_version */ 2);
AddBuiltin(BuiltinOperator_ADD, Register_ADD(),
/* min_version */ 1,
- /* max_version */ 2);
+ /* max_version */ 4);
AddBuiltin(BuiltinOperator_SPACE_TO_BATCH_ND, Register_SPACE_TO_BATCH_ND(),
/* min_version */ 1,
/* max_version */ 3);
@@ -139,7 +139,7 @@
AddBuiltin(BuiltinOperator_DIV, Register_DIV());
AddBuiltin(BuiltinOperator_SUB, Register_SUB(),
/* min_version */ 1,
- /* max_version */ 3);
+ /* max_version */ 5);
AddBuiltin(BuiltinOperator_SPLIT, Register_SPLIT(), /* min_version */ 1,
/* max_version */ 3);
AddBuiltin(BuiltinOperator_SPLIT_V, Register_SPLIT_V(),
diff --git a/tensorflow/lite/kernels/sub.cc b/tensorflow/lite/kernels/sub.cc
index 5845815..2c126c6 100644
--- a/tensorflow/lite/kernels/sub.cc
+++ b/tensorflow/lite/kernels/sub.cc
@@ -225,19 +225,26 @@
// 8bit -> 8bit general quantized path, with general rescalings
// as well as, 16bit -> 16bit with general rescalings
- bool pot_scale_16bit = false;
+ bool pot_scale_16bit = true;
bool input1_scale_is_pot = false;
bool input2_scale_is_pot = false;
bool output_scale_is_pot = false;
- int input1_scale_log2_rounded;
- int input2_scale_log2_rounded;
- int output_scale_log2_rounded;
+ int input1_scale_log2_rounded{0};
+ int input2_scale_log2_rounded{0};
+ int output_scale_log2_rounded{0};
if (input1->type == kTfLiteInt16 && input2->type == kTfLiteInt16 &&
output->type == kTfLiteInt16) {
- // Check that param scale is POT
+ // In case of 16-bit, there are two implementation:
+ // the scale parameter is a general number
+ // the scale parameter is POT and
+ // zero_point is zero for inputs/output.
+ pot_scale_16bit = (input1->params.zero_point == 0) &&
+ (input2->params.zero_point == 0) &&
+ (output->params.zero_point == 0);
+
input1_scale_is_pot =
CheckedLog2(input1->params.scale, &input1_scale_log2_rounded);
@@ -247,14 +254,14 @@
output_scale_is_pot =
CheckedLog2(output->params.scale, &output_scale_log2_rounded);
- pot_scale_16bit = input1_scale_log2_rounded && input2_scale_log2_rounded &&
- output_scale_log2_rounded;
+ pot_scale_16bit &=
+ input1_scale_is_pot && input2_scale_is_pot && output_scale_is_pot;
}
data->pot_scale_16bit = pot_scale_16bit;
if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt8 ||
- pot_scale_16bit) {
+ !pot_scale_16bit) {
TF_LITE_ENSURE_OK(context, PrepareGeneralSubOp(context, input1, input2,
output, params, data, -1));
} else if (output->type == kTfLiteInt16) {
@@ -348,7 +355,7 @@
} else {
TF_LITE_SUB(reference_integer_ops, Add, int8_t);
}
- } else if (data->pot_scale_16bit) {
+ } else if (!data->pot_scale_16bit) {
if (need_broadcast) {
TF_LITE_SUB(reference_ops, BroadcastAdd4DSlow, int16_t);
} else {
diff --git a/tensorflow/lite/toco/tflite/op_version.cc b/tensorflow/lite/toco/tflite/op_version.cc
index bbec4f9..7b7cd40 100644
--- a/tensorflow/lite/toco/tflite/op_version.cc
+++ b/tensorflow/lite/toco/tflite/op_version.cc
@@ -49,11 +49,16 @@
{{OperatorType::kDepthwiseConv, 3}, "1.14.0"},
{{OperatorType::kAdd, 1}, "1.5.0"},
{{OperatorType::kAdd, 2}, "1.14.0"},
+ {{OperatorType::kAdd, 3}, "1.15.0"},
+ {{OperatorType::kAdd, 4}, kPendingReleaseOpVersion},
{{OperatorType::kAddN, 1}, "1.14.0"},
{{OperatorType::kSpaceToBatchND, 1}, "1.6.0"},
{{OperatorType::kSpaceToBatchND, 2}, "1.14.0"},
{{OperatorType::kSub, 1}, "1.6.0"},
{{OperatorType::kSub, 2}, "1.14.0"},
+ {{OperatorType::kSub, 3}, "1.15.0"},
+ {{OperatorType::kSub, 4}, "1.15.0"},
+ {{OperatorType::kSub, 5}, kPendingReleaseOpVersion},
{{OperatorType::kDiv, 1}, "1.6.0"},
{{OperatorType::kBatchToSpaceND, 1}, "1.6.0"},
{{OperatorType::kBatchToSpaceND, 2}, "1.14.0"},
diff --git a/tensorflow/lite/tools/versioning/BUILD b/tensorflow/lite/tools/versioning/BUILD
index 1ba221d..23f3a45 100644
--- a/tensorflow/lite/tools/versioning/BUILD
+++ b/tensorflow/lite/tools/versioning/BUILD
@@ -22,6 +22,7 @@
"//tensorflow/core:tflite_portable_logging",
"//tensorflow/lite:minimal_logging",
"//tensorflow/lite/kernels/internal:compatibility",
+ "//tensorflow/lite/kernels/internal:quantization_util",
"//tensorflow/lite/schema:schema_fbs",
"//tensorflow/lite/schema:schema_fbs_with_mutable",
"@com_google_absl//absl/memory",
diff --git a/tensorflow/lite/tools/versioning/op_version.cc b/tensorflow/lite/tools/versioning/op_version.cc
index 8598c9c..a4288ed 100644
--- a/tensorflow/lite/tools/versioning/op_version.cc
+++ b/tensorflow/lite/tools/versioning/op_version.cc
@@ -24,6 +24,7 @@
#include "absl/strings/str_split.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/lite/kernels/internal/compatibility.h"
+#include "tensorflow/lite/kernels/internal/quantization_util.h"
namespace tflite {
namespace {
@@ -359,7 +360,29 @@
}
return 1;
+ case BuiltinOperator_ADD:
+ if (op_sig.input_types.at(0) == TensorType_INT16 &&
+ op_sig.output_types.at(0) == TensorType_INT16) {
+ if (op_sig.options.addsub.pot_scale_int16) {
+ return 4;
+ } else {
+ return 3;
+ }
+ }
+ if (op_sig.input_types.at(0) == TensorType_INT8) {
+ return 2;
+ }
+ return 1;
+
case BuiltinOperator_SUB:
+ if (op_sig.input_types.at(0) == TensorType_INT16 &&
+ op_sig.output_types.at(0) == TensorType_INT16) {
+ if (op_sig.options.addsub.pot_scale_int16) {
+ return 5;
+ } else {
+ return 4;
+ }
+ }
if (op_sig.options.broadcast.need_broadcast &&
op_sig.options.broadcast.num_dims > 4) {
return 3;
@@ -370,7 +393,6 @@
return 1;
case BuiltinOperator_AVERAGE_POOL_2D:
- case BuiltinOperator_ADD:
case BuiltinOperator_CONCATENATION:
case BuiltinOperator_MAX_POOL_2D:
case BuiltinOperator_PAD:
@@ -487,6 +509,53 @@
}
} break;
+ case BuiltinOperator_ADD:
+ case BuiltinOperator_SUB: {
+ op_sig.options.addsub.pot_scale_int16 = false;
+ const Tensor* input1_tensor =
+ subgraph->tensors()->Get(op->inputs()->Get(0));
+ const Tensor* input2_tensor =
+ subgraph->tensors()->Get(op->inputs()->Get(1));
+ const Tensor* output_tensor =
+ subgraph->tensors()->Get(op->outputs()->Get(0));
+ const QuantizationParameters* input1_quant =
+ input1_tensor->quantization();
+ const QuantizationParameters* input2_quant =
+ input2_tensor->quantization();
+ const QuantizationParameters* output_quant =
+ output_tensor->quantization();
+ if (input1_quant && input1_quant->scale() &&
+ input1_quant->scale()->Length() && input2_quant &&
+ input2_quant->scale() && input2_quant->scale()->Length() &&
+ output_quant && output_quant->scale() &&
+ output_quant->scale()->Length()) {
+ float input1_scale = input1_quant->scale()->Get(0);
+ float input2_scale = input2_quant->scale()->Get(0);
+ float output_scale = output_quant->scale()->Get(0);
+
+ int scale_log2_rounded = 0;
+ bool input1_scale_is_pot =
+ CheckedLog2(input1_scale, &scale_log2_rounded);
+
+ bool input2_scale_is_pot =
+ CheckedLog2(input2_scale, &scale_log2_rounded);
+
+ bool output_scale_is_pot =
+ CheckedLog2(output_scale, &scale_log2_rounded);
+
+ op_sig.options.addsub.pot_scale_int16 =
+ input1_scale_is_pot && input2_scale_is_pot && output_scale_is_pot;
+ }
+
+ if (op_code->builtin_code() == BuiltinOperator_SUB) {
+ op_sig.options.broadcast.need_broadcast =
+ !HaveSameShapes(subgraph, op, 0, 1);
+ op_sig.options.broadcast.num_dims =
+ std::max(GetNumDims(subgraph, op, 0), GetNumDims(subgraph, op, 1));
+ }
+
+ } break;
+
case BuiltinOperator_LSTM: {
auto lstm_option = op->builtin_options_as_LSTMOptions();
if (lstm_option) {
@@ -512,7 +581,6 @@
op_sig.options.space_batch.num_dims = GetNumDims(subgraph, op, 0);
} break;
- case BuiltinOperator_SUB:
case BuiltinOperator_MAXIMUM:
case BuiltinOperator_MINIMUM: {
op_sig.options.broadcast.need_broadcast =
diff --git a/tensorflow/lite/tools/versioning/op_version.h b/tensorflow/lite/tools/versioning/op_version.h
index c1931bc..bec4b67 100644
--- a/tensorflow/lite/tools/versioning/op_version.h
+++ b/tensorflow/lite/tools/versioning/op_version.h
@@ -59,6 +59,9 @@
int32_t num_dims;
bool need_broadcast;
} broadcast;
+ struct {
+ bool pot_scale_int16;
+ } addsub;
} options;
} OpSignature;
