Automated sync from github.com/tensorflow/tensorflow (#2395) BUG=automated sync from upstream NO_CHECK_TFLITE_FILES=automated sync from upstream
diff --git a/tensorflow/lite/core/c/common.cc b/tensorflow/lite/core/c/common.cc index 5430fe2..fd7c415 100644 --- a/tensorflow/lite/core/c/common.cc +++ b/tensorflow/lite/core/c/common.cc
@@ -384,44 +384,6 @@ TfLiteDelegate TfLiteDelegateCreate() { return TfLiteDelegate{}; } -// LINT.IfChange -#ifndef TF_LITE_STATIC_MEMORY -TfLiteOpaqueDelegate* TfLiteOpaqueDelegateCreate( - const TfLiteOpaqueDelegateBuilder* opaque_delegate_builder) { - if (!opaque_delegate_builder) return nullptr; - - TfLiteDelegate* result = new TfLiteDelegate{}; - result->opaque_delegate_builder = new TfLiteOpaqueDelegateBuilder{}; - *(result->opaque_delegate_builder) = *opaque_delegate_builder; - - return reinterpret_cast<TfLiteOpaqueDelegate*>(result); -} - -void TfLiteOpaqueDelegateDelete(TfLiteOpaqueDelegate* opaque_delegate) { - if (!opaque_delegate) return; - - const TfLiteDelegate* tflite_delegate = - reinterpret_cast<const TfLiteDelegate*>(opaque_delegate); - delete tflite_delegate->opaque_delegate_builder; - delete tflite_delegate; -} -#endif // TF_LITE_STATIC_MEMORY - -void* TfLiteOpaqueDelegateGetData(const TfLiteOpaqueDelegate* delegate) { - if (!delegate) return nullptr; - - // The following cast is safe only because this code is part of the - // TF Lite runtime implementation. Apps using TF Lite should not rely on - // 'TfLiteOpaqueDelegate' and 'TfLiteDelegate' being equivalent. - const auto* tflite_delegate = - reinterpret_cast<const TfLiteDelegate*>(delegate); - - if (!tflite_delegate->opaque_delegate_builder) return tflite_delegate->data_; - - return tflite_delegate->opaque_delegate_builder->data; -} -// LINT.ThenChange(Google-internal path) - // Returns a tensor data allocation strategy. TfLiteAllocationStrategy TfLiteTensorGetAllocationStrategy( const TfLiteTensor* const t) {
diff --git a/tensorflow/lite/core/c/common.h b/tensorflow/lite/core/c/common.h index b2ceecb..4e48901 100644 --- a/tensorflow/lite/core/c/common.h +++ b/tensorflow/lite/core/c/common.h
@@ -1348,15 +1348,19 @@ TfLiteDelegate TfLiteDelegateCreate(void); /// `TfLiteOpaqueDelegateBuilder` is used for constructing -/// `TfLiteOpaqueDelegate`, see `TfLiteOpaqueDelegateCreate` below. Note: -/// This struct is not ABI stable. +/// `TfLiteOpaqueDelegate`, see `TfLiteOpaqueDelegateCreate` in c_api_opaque.h. +/// NOTE: This struct is not ABI stable. /// /// For forward source compatibility `TfLiteOpaqueDelegateBuilder` objects /// should be brace-initialized, so that all fields (including any that might be /// added in the future) get zero-initialized. The purpose of each field is /// exactly the same as with `TfLiteDelegate`. /// -/// WARNING: This is an experimental interface that is subject to change. +/// NOTE: This type is part of the TensorFlow Lite Extension APIs. +/// We reserve the right to make changes to this API in future releases, +/// potentially including non-backwards-compatible changes, on a different +/// schedule than for the other TensorFlow Lite APIs. See +/// https://www.tensorflow.org/guide/versions#separate_version_number_for_tensorflow_lite_extension_apis. typedef struct TfLiteOpaqueDelegateBuilder { /// Data that delegate needs to identify itself. This data is owned by the /// delegate. The delegate is owned in the user code, so the delegate is @@ -1393,41 +1397,23 @@ int64_t flags; } TfLiteOpaqueDelegateBuilder; -// LINT.IfChange #ifndef TF_LITE_STATIC_MEMORY -/// Creates an opaque delegate and returns its address. The opaque delegate -/// will behave according to the provided `opaque_delegate_builder`. The -/// lifetime of the objects pointed to by any of the fields within the -/// `opaque_delegate_builder` must outlive the returned -/// `TfLiteOpaqueDelegate` and any `TfLiteInterpreter`, -/// `TfLiteInterpreterOptions`, `tflite::Interpreter`, or -/// `tflite::InterpreterBuilder` that the delegate is added to. The returned -/// address should be passed to `TfLiteOpaqueDelegateDelete` for deletion. If -/// `opaque_delegate_builder` is a null pointer, then a null pointer will be -/// returned. +// See c_api_opaque.h. +// This declaration in common.h is only for backwards compatibility. +// NOTE: This function is part of the TensorFlow Lite Extension APIs, see above. TfLiteOpaqueDelegate* TfLiteOpaqueDelegateCreate( const TfLiteOpaqueDelegateBuilder* opaque_delegate_builder); -/// Deletes the provided opaque `delegate`. This function has no effect if the -/// `delegate` is a null pointer. +// See c_api_opaque.h. +// This declaration in common.h is only for backwards compatibility. +// NOTE: This function is part of the TensorFlow Lite Extension APIs, see above. void TfLiteOpaqueDelegateDelete(TfLiteOpaqueDelegate* delegate); #endif // TF_LITE_STATIC_MEMORY -/// Returns a pointer to the data associated with the provided opaque -/// `delegate`. -/// -/// A null pointer will be returned when: -/// - The `delegate` is null. -/// - The `data` field of the `TfLiteOpaqueDelegateBuilder` used to construct -/// the `delegate` was null. -/// - Or in case of any other error. -/// - The `delegate` has been constructed via a `TfLiteOpaqueDelegateBuilder`, -/// but the `data` field of the `TfLiteOpaqueDelegateBuilder` is null. -/// -/// The data_ field of `delegate` will be returned if the -/// `opaque_delegate_builder` field is null. +// See c_api_opaque.h. +// This declaration in common.h is only for backwards compatibility. +// NOTE: This function is part of the TensorFlow Lite Extension APIs, see above. void* TfLiteOpaqueDelegateGetData(const TfLiteOpaqueDelegate* delegate); -// LINT.ThenChange(Google-internal path) /// Returns a tensor data allocation strategy. TfLiteAllocationStrategy TfLiteTensorGetAllocationStrategy(
diff --git a/tensorflow/lite/kernels/internal/portable_tensor_utils.h b/tensorflow/lite/kernels/internal/portable_tensor_utils.h index c28892c..03bfdc8 100644 --- a/tensorflow/lite/kernels/internal/portable_tensor_utils.h +++ b/tensorflow/lite/kernels/internal/portable_tensor_utils.h
@@ -259,7 +259,7 @@ const int8_t* __restrict__ matrix, const uint8_t* __restrict__ ledger, const int m_rows, const int m_cols, const int8_t* __restrict__ vectors, const float* __restrict__ scaling_factors, int n_batch, - float* __restrict__ result); + float* __restrict__ result, const float* per_channel_scale = nullptr); // Same as the above 8, 8, 8 integer matmul except for the presence of zero // point and non-accumulative.
diff --git a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc index d386203..b519d81 100644 --- a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc +++ b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.cc
@@ -294,7 +294,8 @@ void PortableSparseMatrixBatchVectorMultiplyAccumulate( const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows, const int m_cols, const int8_t* __restrict__ vectors, - const float* scaling_factors, int n_batch, float* __restrict__ result) { + const float* scaling_factors, int n_batch, float* __restrict__ result, + const float* per_channel_scale) { static const int kBlockSize = 16; TFLITE_DCHECK_EQ( // NOLINT m_cols % kBlockSize, 0); @@ -319,7 +320,11 @@ dotprod += (*row_ptr++) * (*vector_block_ptr++); } // for block } // for num_nonzero_blocks - result[batch * m_rows + row] += dotprod * batch_scaling_factor; + float scaling_factor = batch_scaling_factor; + if (per_channel_scale) { + scaling_factor *= per_channel_scale[row]; + } + result[batch * m_rows + row] += dotprod * scaling_factor; } // for row } // for batch }
diff --git a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.h b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.h index 0416db0..ab0185e 100644 --- a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.h +++ b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils.h
@@ -129,10 +129,11 @@ void SparseMatrixBatchVectorMultiplyAccumulate( const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows, const int m_cols, const int8_t* __restrict__ vectors, - const float* scaling_factors, int n_batch, float* __restrict__ result) { + const float* scaling_factors, int n_batch, float* __restrict__ result, + const float* per_channel_scale) { PortableSparseMatrixBatchVectorMultiplyAccumulate( - matrix, ledger, m_rows, m_cols, vectors, scaling_factors, n_batch, - result); + matrix, ledger, m_rows, m_cols, vectors, scaling_factors, n_batch, result, + per_channel_scale); } void MatrixBatchVectorMultiplyAccumulate(
diff --git a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h index 6c404d5..6a1f18e 100644 --- a/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h +++ b/tensorflow/lite/kernels/internal/reference/portable_tensor_utils_impl.h
@@ -99,7 +99,8 @@ void PortableSparseMatrixBatchVectorMultiplyAccumulate( const int8_t* __restrict__ matrix, const uint8_t* ledger, const int m_rows, const int m_cols, const int8_t* __restrict__ vectors, - const float* scaling_factors, int n_batch, float* __restrict__ result); + const float* scaling_factors, int n_batch, float* __restrict__ result, + const float* per_channel_scale); // Dot product of two vectors. float PortableVectorVectorDotProduct(const float* vector1, const float* vector2,